Computational Fluid Dynamics 2006 Proceedings of The Fourth International Conference On Computational Fluid Dynamics ICCFD4 Ghent Belgium 10 14 J

Computational Fluid Dynamics 2006
“This page left intentionally blank.”

Herman Deconinck · Erik Dick
Editors
Computational
Fluid Dynamics 2006
Proceedings of the
Fourth International Conference
on Computational Fluid Dynamics, ICCFD4,
Ghent, Belgium, 10–14 July 2006
123
Editors
Prof. Dr. Herman Deconinck Prof. Dr. Erik Dick
von Karman Institute Ghent University
for Fluid Dynamics Fac. Engineering
Waterloosesteenweg 72 Dept. Flow, Heat &
1640 Sint-Genesius-Rode Combustion Mechanics
Belgium Sint-Pietersnieuwstraat 41
deconinck@vki.ac.be 9000 Gent
Belgium
ISBN 978-3-540-92778-5 e-ISBN 978-3-540-92779-2

DOI 10.1007/978-3-540-92779-2
Springer Dordrecht Heidelberg London New York
Library of Congress Control Number: 2009921821
c Springer-Verlag Berlin Heidelberg 2009

This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting,
reproduction on microfilm or in any other way, and storage in data banks. Duplication of this publication
or parts thereof is permitted only under the provisions of the German Copyright Law of September 9,
1965, in its current version, and permission for use must always be obtained from Springer. Violations
are liable to prosecution under the German Copyright Law.
The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant protective
laws and regulations and therefore free for general use.
Cover design: WMXDesign GmbH, Heidelberg
Printed on acid-free paper
Springer is part of Springer Science+Business Media (www.springer.com)

To our families
Preface
This book contains the proceedings of the Fourth International Conference

on Computational Fluid Dynamics (ICCFD4), held in Gent, Belgium from
July 10 through 16, 2006. The ICCFD conference series is an outcome of the
merger of two important streams of conferences in Computational Fluid Dy-
namics: International Conference on Numerical Methods in Fluid Dynamics,
ICNMFD (since 1996) and International Symposium on Computational Fluid
Dynamics, ISCFD (since 1985). In 1998 it was decided to join the two and
ICCFD emerged as a biannual meeting, held in Kyoto in 2000, Sydney in
2002, Toronto in 2004 and Gent in 2006. Thus, the ICCFD series became
the leading international conference series for scientists, mathematicians and
engineers interested in the computation of fluid flow.
The 4th edition of the conference has attracted 200 participants from all
over the world; 270 abstracts were received, of which 135 were selected in a
careful peer review process by the executive committee (C. H. Bruneau, J.-J.
Chattot, D. Kwak, N. Satofuka, D.W. Zingg , E. Dick and H. Deconinck) for
oral presentation and a further 21 for poster presentation.
The papers contained in these proceedings provide an excellent snapshot
of the field of Computational Fluid Dynamics as of 2006. Invited keynote
lectures by renowned researchers are included, with contributions in the field
of discretization schemes, high-end computing and engineering challenges, and
two-phase flow. These keynote contributions are complemented by 137 regular
papers on the most diverse aspects of CFD:
- Innovative algorithm development for flow simulation, optimisation and con-
trol: higher-order methods (DG, FV, FE and RDmethods), iterative methods
and multigrid, solution adaptive mesh techniques, error estimation and con-
trol, parallel algorithms.
- Innovative modeling of flow physics in the area of compressible and incom-
pressible flows: hypersonic and reacting flows, two-phase flows, turbulence
(LES, DES, DNS, and transition), vortex dynamics, boundary layer stability,
multi-scale physics, magnetohydrodynamics.
Preface VII
- advanced applications using the above mentioned innovative technology, and

multidisciplinary applications including aero-elasticity and aero-acoustics.
Thanks are due to our sponsors NASA, the FWO Research Foundation
Flanders and the European Union through the EUA4X Marie Curie project.
In particular, the generous grant from NASA is a key factor in the success of
this conference series and the publication of these Proceedings.
We also would like to thank the staff and PhD students of the von Karman
Institute and the Department of flow, heat and combustion mechanics of the
University of Gent, for the help they provided toward the success of this
conference.
Sint-Genesius-Rode, Belgium Herman Deconinck

von Karman Institute for Fluid Dynamics
Ghent, Belgium Erik Dick

Ghent University
September 2006 Conference Chair

Contents
Part I Invited Speakers
Two new techniques for generating exactly incompressible

approximate velocities
Bernardo Cockburn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Role of High-End Computing in Meeting NASA’s Science
and Engineering Challenges
Rupak Biswas, Eugene L. Tu, William R. Van Dalsem . . . . . . . . . . . . . . . . 14
Recent Advances of Multi-phase Flow Computation with the
Adaptive Soroban-grid Cubic Interpolated Propagation (CIP)
Method
Takashi Yabe, Youichi Ogata, Kenji Takizawa . . . . . . . . . . . . . . . . . . . . . . . 29
Part II Schemes
On the Computation of Steady-State Compressible Flows

Using a DG Method
Hong Luo, Joseph D. Baum, Rainald Löhner . . . . . . . . . . . . . . . . . . . . . . . . 47
Space-Time Discontinuous Galerkin Method for Large
Amplitude Nonlinear Water Waves
Yan Xu, Jaap J.W. van der Vegt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
A discontinuous Galerkin method with Hancock-type time
integration for hyperbolic systems with stiff relaxation source
terms
Yoshifumi Suzuki, Bram van Leer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Contents IX
Very High Order, Non-Oscillatory Fluctuation Distribution

Schemes
M.E.Hubbard, N.Z.Mebrate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
High-order residual distribution : discontinuity capturing

crosswind dissipation and diffusion
N. Villedieu-Ligout, M. Ricchiuto, H. Deconinck . . . . . . . . . . . . . . . . . . . . . 71
High-Order Fluctuation-Splitting Schemes for Advection-
Diffusion Equations
Hiroaki Nishikawa and Philip Roe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Construction of Higher Order Residual Distribution Schemes
Rémi Abgrall and Cédric Tavé . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Stable and convergent residual distribution for time-
dependent conservation laws
Mario Ricchiuto, Rémi Abgrall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
An ALE Formulation of the Multidimensional Residual
Distribution Scheme for Computations on Moving Meshes
Jirı́ Dobes, Herman Deconinck . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Solution of the steady Euler equations using Fluctuation
Splitting schemes on quadrilateral elements
D. T. Rubino, P. De Palma, G. Pascazio, M. Napolitano . . . . . . . . . . . . . 101
A Residual-Based Compact Scheme for All-Speed Flows on
Unstructured Grids
Christophe Corre, Alberto Beccantini, Thibaud Kloczko . . . . . . . . . . . . . . . 107
Vorticity Preserving Scheme for Unsteady Compressible
Flows
Fabrice Falissard, Alain Lerat, Jacques Sidès . . . . . . . . . . . . . . . . . . . . . . . . 113
Extension of the SD Method to Viscous Flow on Unstructured
Grids
Z. J. Wang, Yuzhi Sun, C. Liang, Yen, Liu . . . . . . . . . . . . . . . . . . . . . . . . . 119
Strictly Stable High Order Difference Methods for the
Compressible Euler and Navier-Stokes Equations
Bernhard Müller . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
Uniform Flow Preserving Property of High Order Upwind

Finite Difference Schemes on Generalized Coordinate System
Taku Nonomura, Nobuyuki Iizuka, Kozo Fujii . . . . . . . . . . . . . . . . . . . . . . . . 131
X Contents
Implementation of an Enhanced Flux Formulation for

Unsteady Navier-Stokes Solutions
G. Xia, S. Sardeshmukh, V. Sankaran, C. L. Merkle . . . . . . . . . . . . . . . . . . 137
Computation of Eigenspaces of Hyperbolic Systems

S.K. Godunov, O.B. Feodoritova, V.T. Zhukov . . . . . . . . . . . . . . . . . . . . . . . 143
A Proposed Cure to the Carbuncle Phenomenon
Farzad Ismail, Philip L. Roe, Hiroaki Nishikawa . . . . . . . . . . . . . . . . . . . . . 149
The High Order WLSQR Scheme and its Applications in

Turbomachinery
Jivrı́ Fürst . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Building Better (Weighted) ENO Methods
William J. Rider . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
Discontinuity Diagnosis Essentially Non-Oscillatory Schemes
Yun-Feng Liu, Jian-Ping Wang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
Third Order Reconstruction on Unstructured Highly
Irregular 3D Meshes
J. Rokicki, R. Wieteska . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
An Intrinsically Multi-Dimensional Acoustics Convection
Upstream Resolution Algorithm for the Euler Equations
Joe Iannelli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
Multi-dimensional Limiting Process for Two- and
Three-dimensional Flow Physics Analyses
Sung-Hwan Yoon, Chongam Kim and Kyu-Hong Kim . . . . . . . . . . . . . . . . . 185
A Multidimensional Kinetic Upwind Method for Euler
Equations
Keshav S Malagi, P S Kulkarni, S M Deshpande . . . . . . . . . . . . . . . . . . . . . 191
High Resolution Quantum Kinetic Beam Schemes and Its
Applications to Ideal Quantum Gas Dynamical Flows
Y. H. Shi, J. Y. Yang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
Semi-GLS Stabilization of FEM Applied to Incompressible
Flows with Higher Reynolds Numbers
Pavel Burda, Jaroslav Novotný, Jakub vSı́stek . . . . . . . . . . . . . . . . . . . . . . . 203
Finite volume box scheme for a certain class of nonlinear
conservation laws in mixed form
H. Beaugendre, A. Ern . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
Contents XI
Numerical study of the Colocated Clustered Finite Volume

Scheme
O. Touazi, E. Chénier, R. Eymard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
Arbitrary High Order Finite Volume Schemes on

Unstructured Meshes
Michael Dumbser, Claus-Dieter Munz . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
Part III Algorithms
A high scalability parallel algebraic multigrid solver

T. Saad, M. Darwish . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
Jacobian-Free Newton-Krylov Methods: Issues and Solutions
David W. Zingg, Todd T. Chisholm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Non-stationary two-stage relaxation based on the principle of
aggregation multi-grid
R. Haelterman, J. Vierendeels, D. Van Heule . . . . . . . . . . . . . . . . . . . . . . . . 243
The efficient and accurate solution of porous media flow
problems with strongly discontinuous coefficients
Y.C. Lee, P.H. Gaskell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
Stability of Pressure-Correction Algorithms for Low-Speed
Reacting and Non-Reacting Flow Simulations
Pieter Rauwoens, Krista Nerinckx, Jan Vierendeels, Bart Merci . . . . . . . 255
A simple hybrid well-balanced method for a 2D viscous
shallow water model
F. Marche . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
A kinetic energy-preserving P1 iso P2/P1 finite-element
method for computing unsteady incompressible flows
David Vanden-Abeele, Deryl Snyder, Yves Detandt and Gérard Degrez . . 267
Study on the segregation algorithms of the incompressible
Navier-Stokes equations using P1P1/P2P1 finite element
formulation
Myung H. Cho, Hyoung G. Choi, Jung Y. Yoo . . . . . . . . . . . . . . . . . . . . . . . 273
A Mach-uniform algorithm: coupled versus segregated
approach
Krista Nerinckx, Jan Vierendeels, Erik Dick . . . . . . . . . . . . . . . . . . . . . . . . . 279
XII Contents
Crank-Nicolson Scheme for Solving Low Mach Number

Unsteady Viscous Flows Using an Implicit Preconditioned
Dual Time Stepping Technique
D. Vigneron, G. Deliége, J.-A. Essers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Heated Wake by Deferred Corrected ULTRA

Salem Bouhairie, Vincent H. Chu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
Convergence Acceleration for Euler Equation based on SPR
Dohyung Lee, Hyungmin Kang and Dongho Lee . . . . . . . . . . . . . . . . . . . . . . 297
Acceleration of Unsteady Incompressible Flow Calculation

Using Extrapolation Methods
Kenjiro Shimano, Shun Yonezu, Yoshiteru Enomoto . . . . . . . . . . . . . . . . . . 303
Improved Numerical Simulations of Incompressible Flows
Based on Viscous/Inviscid Interaction Procedures
M. Hafez, A. Shatalov, M. Nakajima . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Part IV AMR - Adaptive mesh techniques
A Parallel Unstructured Overset Mesh Technique for

Unsteady Flow Simulations
Mun Seung Jung, Oh Joon Kwon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
A Parallel Overset Adaptive Cartesian/Prism Grid Method
for Moving Boundary Flows
Ravishekar Kannan, Z.J. Wang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 323
Navier-Stokes Simulation of Local Winds Over the Earth’s
Topography
Neal M. Chaderjian, Jasim U. Ahmad, Marc G. Kramer, Terry L. Holst 329
Grid-adaptive Simulations of Relativistic Flows
R. Keppens, Z. Meliani . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
Solution of Laminar Combusting Flows Using a Parallel
Implicit Adaptive Mesh Refinement Algorithm
Scott A. Northrup, Clinton P. T. Groth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341
Towards Direct Numerical Simulation of a Diffusion
Flame-Shock Interaction with an AMR Algorithm
G. Billet, J. Ryan and M. Borrel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
Adaptive Multigrid Solutions of Thin Film Flows over
Topography
Y.C. Lee, H.M. Thompson, P.H. Gaskell . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
Contents XIII
A meshless solver for computing viscous flows on Cartesian

like grids
Munikrishna N., Karthikeyan N., Balakrishnan N. . . . . . . . . . . . . . . . . . . . 359
Immersed boundary technique for compressible flow

simulations on semi-structured grids
M. D. de Tullio, P. De Palma, G. Iaccarino, G. Pascazio, M. Napolitano365
Incompressible Flow Simulations Using Virtual Boundary
Method with New Direct Forcing Terms Estimation
Hidetoshi Nishida, Kazuhiro Sasao . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
A Hybrid Building-Block and Gridless Method for Computing
Shock Waves
Hong Luo, Joseph D. Baum, Rainald Löhner . . . . . . . . . . . . . . . . . . . . . . . . 377
A Residual estimator based adaptation strategy for
compressible flows
Ganesh N., Nikhil V Shende, Balakrishnan N. . . . . . . . . . . . . . . . . . . . . . . . 383
Anisotropic solution-adaptive technique applied to simulations
of steady and unsteady compressible flows
Jerzy Majewski, Aristotelis N. Athanasiadis . . . . . . . . . . . . . . . . . . . . . . . . . 389
Simulation of Flow around Wing Sections by Building-Cube
Method
Kazuhiro Nakahashi, Yota Sakurai . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
Robust Mesh Deformation using the Linear Elasticity
Equations
Richard P. Dwight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
Part V DNS and LES
Identification and role of coherent structures in two-

dimensional turbulence
Ch.-H. Bruneau, P. Fischer, H. Kellay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
Temporal Evolution of Dominant Flow Structures in
Turbulent Channel Flow
Giancarlo Alfonsi, Leonardo Primavera . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
LES of passive scalar in compressible mixing layers
C. Le Ribault . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
Buoyancy Effect on Turbulence using Blocks
Vincent H. Chu, Wihel Altai, Camilo E. Pinilla . . . . . . . . . . . . . . . . . . . . . . 427
XIV Contents
Direct numerical simulation of Taylor-Couette flows in the

fully turbulent regime
Yves Detandt, Mikhail Krivilyov, Yacine Salhi, David Vanden Abeele,
Jan Fransaer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
Flux Limiting Schemes for Implicit Large Eddy Simulation of

Synthetic Jets
Sanjay Patel, Dimitris Drikakis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Implicit Large Eddy Simulation of a Flow around a Cylindrical
Body
Satoko Komurasaki, Kunio Kuwahara . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
LES study of the impact of the wake structures on the
aerodynamics of a simplified ICE2 train subjected to a side
wind
Hassan Hemida, Sinivsa Krajnović . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Use of Immersed Boundary Technique in a Cartesian LES
solver to study wake flows
J. Bodart, R. Giammanco, P. Rambaud, C. Benocci . . . . . . . . . . . . . . . . . . 457
Stochastic generation of velocity fluctuation for turbulent
inflow and initial condition
M. Fathali, M. Klein, T. Broeckhoven, C. Lacor, M. Baelmans . . . . . . . . . 463
Study on Numerical and Modelling Error in LES of a Channel
Flow Using Explicit Filtering
Tellervo T. Brandt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 469
Wall Boundary Conditions for Variational Multiscale
Large-Eddy Simulations
S. J. Hulshoff, E. A. Munts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 475
The Sampling Based Dynamic Procedure for Numerical
Discretization Enhancement
Dieter Fauconnier, Chris De Langhe, Erik Dick . . . . . . . . . . . . . . . . . . . . . . 481
Application of a Unique Eddy–Viscosity Model for a Hybrid
LES–RANS Method
Michael Breuer, Benoit Jaffrézic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
A locally superconvergent scheme for the simulation of
turbulent flows in complex geometries
M.V. Salvetti, S. Camarri, B. Koobus, A. Dervieux . . . . . . . . . . . . . . . . . . 493
URANS Analysis of Flow-Induced Cavity Resonances
M. Mesbah, W. Desmet, M. Baelmans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 499
Contents XV
Revisiting URANS computations of the flow behind a

backward-facing step using second moment closures
A. Fadai-Ghotbi, R. Manceau, J. Borée . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
Contribution of Turbulence Equation Terms to the Shear

Stress Balance
Dragan Kovzulović, Thomas Röber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 511
Part VI Optimisation and Flow Control
Optimum Shape Design for Unsteady Three-Dimensional

Viscous Flows
S. Nadarajah, A. Jameson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
Comparison of Exact and Approximate Discrete Adjoint for
Aerodynamic Shape Optimization
Giampietro Carpentieri, Michel J.L van Tooren, Barry Koren . . . . . . . . . 525
Adjoint Sensitivity Computations for an Embedded-Boundary
Cartesian Mesh Method and CAD Geometry
Marian Nemec, Michael J. Aftosmis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
Optimum shape design through multilevel gradient-based
method using Bézier parametrisation
M. Martinelli, F. Beux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
High-Fidelity Multi-criteria Aero-structural Optimisation
using Hierarchical Parallel Evolutionary Algorithms
L. F.González, L. Damp, J. Périaux, K. Srinivas . . . . . . . . . . . . . . . . . . . . 543
Flow Control Optimization Using Neural Networks and
Genetic Algorithms
Raymond P. LeBeau, Jr., Narendra K. Beliganur, Thomas Hauser . . . . . 549
Shape Optimization for Dense Gas Flows in Turbine Cascades
Pietro Marco Congedo, Paola Cinnella, Christophe Corre . . . . . . . . . . . . . 555
Control of Flow Past a Stalled NACA0015Airfoil
Meiliang Mao, Xiaogang Deng, Jianqiang Chen . . . . . . . . . . . . . . . . . . . . . . 561
Active control of a transitional flow over a backward-facing
step
E. Creusé, A. Giovannini, I. Mortazavi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 567
Synthetic Jet Actuator Modeling for Flow Control
Applications
M. Ferlauto, R. Marsilio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 573
XVI Contents
Numerical Study of Transonic Drag Reduction for Flow Past

Airfoils Using Active Flow Control
Jose Vadillo, Ramesh K. Agarwal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
Part VII Two-phase and multimaterial flows
Level Set based Finite Element Method of Bubble-in-Liquid

Simulation
Hyoung G. Choi, Jung Y. Yoo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 589
Level Set method for Curvature-driven Flows in Microfluidics

Paul Vigneaux . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 595
Mesh-Based Microstructure Representation Algorithm for
Simulating Pore-scale Transport Phenomena in Porous Media
May-Fun Liou, Issac Greber . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 601
The Numerical Simulation of Liquid Sloshing in Microgravity
Roel Luppes, Joop A. Helder, Arthur E.P. Veldman . . . . . . . . . . . . . . . . . . 607
Numerical Prediction of Interfacial Instability
Robert Nourgaliev, Meng-Sing Liou, Theo Theofanous . . . . . . . . . . . . . . . . 613
A Lattice Boltzmann based Single-Phase Model: Surface
Tension and Wetting
Xiu Qing Xing, David Lee Butler, Chun Yang . . . . . . . . . . . . . . . . . . . . . . . 619
Hyperbolicity, Discontinuities, and Numerics of Two-Fluid
Models
Meng-Sing Liou, Loc Nguyen, Chih-Hao Chang, Sveta Sushchikh,
Robert Nourgaliev, Theo Theofanous . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625
Computation of Multiphase Mixture Flows using RoeM and
AUSMPW+ Schemes
Seung-Won Ihm, Kyung Rok Lee, Chongam Kim, and Kyu Hong Kim . . 631
A relaxation method for the Kapila model
C. Berthon, B. Braconnier, J. Claudel, B. Nkonga . . . . . . . . . . . . . . . . . . . 637
Direct numerical simulation of bubbly Taylor-Couette flow
Thomas Nierhaus, Jean-Franccois Thomas, Yves Detandt, David
Vanden Abeele . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Numerical simulation of unsteady flow inside an impulsively
started liquid drop
M. Krivilyov, J. Fransaer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 649
Contents XVII
Accurate and Efficient Solution of 2D Steady Water Flows

with Surface Waves and Turbulence
Jeroen Wackers, Barry Koren . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655
Numerical analysis of bubble migration in thermocapillary

flows of an open cylindrical container
Hiroaki Ohira, Satoshi Matsumoto, Takashi Mashiko, Shinichi Yoda,
Yasuhiro Kamotani . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 662
Part VIII Hypersonic and supersonic flow
Numerical investigation of an effusion cooled thermal

protection material
Volker Hannemann . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 671
Calculation of Transport Properties for Entry into the
Martian Atmosphere
Johannes Baumgart, Tobias Leicht, Thierry Magin, Paolo Barbante,
Pietro Rini, Gérard Degrez, Roger Grundmann . . . . . . . . . . . . . . . . . . . . . . 677
Numerical simulation of supersonic flow around a double
ramp configuration and correlation with experiment
Fedorchenko I.A., Fedorova N.N., Kharlamova Yu.V., Gaisbauer U.,
and Kraemer E. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683
Effects of Mach number on the combustion zone length for a
RAMAC configuration at sub-detonative mode
Tarek Bengherbia, Yufeng Yao, Pascal Bauer . . . . . . . . . . . . . . . . . . . . . . . . 689
A Finite Element/Finite Volume Mixed Solver
He Lixing, Zhang Laiping, Zhang Hanxin . . . . . . . . . . . . . . . . . . . . . . . . . . . 695
Numerical Analysis for Magnetic Control of Heat-Transfer
and Pressure in Hypersonic Shock Wave Interference Flows
Daisuke Tsubakino, Yoshiteru Tanaka, Kozo Fujii . . . . . . . . . . . . . . . . . . . . 701
Numerical Simulation of a Flat-Plate Hypersonic Shock Layer
Perturbed by External Acoustic Waves
A.N.Kudryavtsev, S.G.Mironov, T.V.Poplavskaya, I.S.Tsyryulnikov . . . . 707
Numerical Simulation of Supersonic Turbulent Flows over
Backward–Facing Steps
N.N. Fedorova, I.A. Bedarev . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
XVIII Contents
Unified Flow Solver Combining Boltzmann and Continuum

Models for Simulations of Gas Flows for the Entire Range of
Knudsen Numbers
V. V. Aristov, A. A. Frolova, S. A. Zabelok, V. I. Kolobov,
R. R. Arslanbekov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
Adaptive Boltzmann/Navier-Stokes Hybrid Method for
Multi-Scale Gas Flow Simulation
Koji Morinishi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
Computing Simulation of Hypersonic Magneto-Fluid-

Dynamics Interaction
J.S. Shang . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 731
Numerical Simulation of R-M instability at High Mach
Numbers
Fu Dexun, Ma Yanwen, Liang Xian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 738
Part IX Multiphysics and interdisciplinary coupled

flow problems
Boundary conditions by low-order modelling

Marcelo Buffoni, Haysam Telib, Angelo Iollo . . . . . . . . . . . . . . . . . . . . . . . . 747
Multi Scale Numerical Simulation of Dispersed Reacting
Flow, with application to Chemical Vapor Deposition of
Alumina
Andrey A. Markov . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 753
Computational Analysis of Flow through a Multiple Nozzle
Driven Laser Cavity and Diffuser
M. A. Sriram, N. K. S. Rajan, P. S. Kulkarni . . . . . . . . . . . . . . . . . . . . . . . 759
MHD Analysis of Force Acting on Dipole Magnetic Field in
Magnetized Plasma Flow
Hiroyuki Nishida, Hiroyuki Ogawa, Yoshifumi Inatani . . . . . . . . . . . . . . . . 765
Performance of High Order Filter Methods for a Richtmyer-
Meshkov Instability
B. Sjögreen, H.C.Yee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 771
Unsteady Flow Simulation of High Speed Turbopumps
Cetin C. Kiris, Dochan Kwak, William Chan, Jeffrey A. Housman . . . . . 777
Use of the Gaussian Moment Closure for the Modelling of
Continuum and Micron-Scale Flows with Moving Boundaries
J. G. McDonald, J. S. Sachdev, C. P. T. Groth . . . . . . . . . . . . . . . . . . . . . 783
Contents XIX
Development of “MATIS-SC” for High Speed Steam Flow

with Non-equilibrium Condensation
Ryo Morita, Fumio Inada . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 789
Numerical Method for Near-critical Fluids of Arbitrary

Material
Satoru Yamamoto and Atsushi Ito . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795
Flow Physics and Stokes’ Theorem in Wind Turbine
Aerodynamics
Sven Schmitz, Jean-Jacques Chattot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 801
The Role of Separation Bubble on an Airfoil at Low Reynolds
Numbers
Yusuke Nakae, Tatsuo Motohashi, Satoko Komurasaki, Kunio Kuwahara 807
Effect of the thermal boundary conditions and physical
properties variation on transient natural convection of high
Prandtl number fluids
O. Younis, J. Pallares, F. X. Grau . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 813
Aeroacoustic simulation of the flow in a Helmholtz resonator
Laurent Georges, Grégoire Winckelmans, Stéphane Caro, Philippe
Geuzaine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 819
Explicit low dispersive and low dissipative non-centered finite
differences and filters
Berland J., Bogey C., Bailly C. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825
A New Principle of Non-Reflecting and its Application to
Hyperbolic Conservation Laws
Ching Y. Loh, Philip C.E. Jorgenson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 831
Numerical Solution of the Linearized Euler Equations Using
Compact Schemes
Kris Van den Abeele, Jan Ramboer, Ghader Ghorbaniasl, Chris Lacor . . 837
Transonic Aeroelastic Computations of a Delta Wing
Configuration with High Fidelity Equations
Hiroshi Terashima, Kenji Ono . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 843
Transonic Buffet over Symmetric Airfoils
Alexander Kuz’min, Alexey Shilkin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 849
Aerodynamic Performance of a Deforming Elastic Body in
Supersonic Flow
Tomohisa Hashimoto, Koji Morinishi, Nobuyuki Satofuka . . . . . . . . . . . . . 855
XX Contents
CAD-Centric Framework for Aero-mechanical Optimization -

Counter-Rotating Fan Design
Hiromasa Kato, Stéphane Pierret, Rajan Filomeno Coelho . . . . . . . . . . . . 861
Aeroelastic Solutions for Viscous Flows using the Time

Accurate and Non-Linear Frequency Domain Methods
F. Kachra, S. Nadarajah, C. Tatossian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 867
Role of Interpolation in Airflow Induced Vibration in Hard
Disk Drive Enclosures
S. Ali, M. Damodaran . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 873
Computation of fluid structure interaction with application
to three-Dimensional combustion system
Zinedine Khatir . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 879
Numerical Computations of Unsteady Aerodynamics of
Projectiles using an Unstructured Technique
Jubaraj Sahu . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 886
Part I
Invited Speakers
Two new techniques for generating exactly
incompressible approximate velocities
Bernardo Cockburn
School of Mathematics, University of Minnesota, Minneapolis, MN 55455, USA

cockburn@math.umn.edu
1 Introduction
We consider the problem of generating exactly incompressible finite element
approximations of the velocity, that is, of approximations which are pointwise
divergence-free inside each element and whose normal component is continu-
ous across inter-element boundaries. This problem has remained open since
the beginning of the finite element simulation of incompressible fluid flow and
has only been recently solved by means of two techniques: The first is a post-
processing technique and the second a hybridization procedure. This paper is
a short review of these techniques.
The post-processing technique is devised specifically for discontinuous
Galerkin (DG) methods and is based in the fact, typical of these methods, that
the divergence-free constraint is imposed through a Galerkin weak formulation
on each element. The post-processing, which can be defined in an element-
by-element fashion, provides an exactly divergence-free approximate velocity
even though the method uses completely discontinuous elements. The tech-
nique can be applied in two or three space dimensions and works for velocity
approximations of any degree. This post-processing technique was introduced
in [2] in the framework of fluid flow in porous medium and, independently,
in [10] in the framework of the incompressible Navier-Stokes equations. The
application of this approach to incompressible elastic materials was carried
out in [11].
The hybridization technique is devised as an implementation tool for finite
element methods (including classical mixed and DG methods) which employ
divergence-free finite dimensional subspaces. It is well known that the main
difficulty with this type of methods is the construction of finite dimensional
spaces of divergence-free functions. Indeed, the construction is, in practice,
impossible due to the many inter-element continuity constraints that need to
be imposed. By using the hybridization technique, those constraints are im-
posed only on the approximate velocity, not on the test functions, by means
of a suitably chosen set of equations. Moreover, the test functions are taken
H. Deconinck, E. Dick (eds.), Computational Fluid Dynamics 2006,

DOI 10.1007/978-3-540-92779-2 1, c Springer-Verlag Berlin Heidelberg 2009
4 Bernardo Cockburn
in a finite dimensional spaces of completely discontinuous velocities which

renders the implementation of the method extremely easy. This hybridiza-
tion technique was introduced in [4] for DG methods and then in [6] and [7]
for a classical mixed method. For an overview of the recent applications of
hybridization, the reader is referred to [8].
The organization of this paper is as follows. In section 2, we describe
the post-processing technique for a local DG method and in section 3, the hy-
bridization of a classical mixed method. Finally, we end in section 3 with some
comments concerning the possibility of applying a hybridization procedure to
DG methods.
For the sake of simplicity, we restrict ourselves to the following model
problem of incompressible fluid flow,
−∆u + ∇p =f on Ω, (1a)
∇ · u =0 on Ω, (1b)
u =g on ∂Ω. (1c)
Here, Ω is a bounded polyhedral domain. The external body force f is taken

to be in L2 (Ω)3 and the Dirichlet data g in H 1/2 (Ω)3 . As usual, we assume
that hg , ni∂Ω = 0, whereR n denotes the outward unit normal to ∂Ω; we use
the notation hg , hiF := F g · h dγ.
2 Post-processing of DG methods
In this section, we describe the post-processing technique for obtaining exactly
divergence-free approximate velocities from the approximation provided by a
local DG method. The material of this section is contained in [10].
2.1 The DG method
To introduce the DG method, we begin by rewriting the equations (1) as
σ = ν∇u in Ω, (2a)
−∇ · σ + ∇p = f in Ω, (2b)
∇·u=0 in Ω, (2c)
u=g on ∂Ω. (2d)
Given any triangulation Th of Ω made of tetrahedra K, the approximation to

(σ, u, p), (σh , uh , ph ), is taken in the space Σh × Vh × Qh where
Σh = { τ ∈ L2 (Ω)3×3 : τ |K ∈ Pk (K)3×3 , K ∈ Th },
Vh = { v ∈ L2 (Ω)3 : v|K ∈ Pk (K)3 , K ∈ Th ; },
Qh = { q ∈ L2 (Ω) : q|K ∈ Pk−1 (K), K ∈ Th ; (q, 1)Ω = 0 },
Exactly incompressible approximate velocities 5
and Pk (K) is the space of polynomials of degree at most k > 0 defined on K.

It is defined by requiring that, for all functions (τ, v, q) ∈ Σh × Vh × Qh ,
uσh , τ · ni∂Ωh ,
(σh , τ )Ωh = −ν (uh , ∇ · τ )Ωh + ν hb (3a)
(σh , ∇v)Ωh − hb σh , v ⊗ ni∂Ωh
− (ph , ∇ · v)Ωh + hb
ph , v · ni∂Ωh = (f , v)Ωh , (3b)
p
− (uh , ∇q)Ωh + hb uh · n , qi∂Ωh = 0, (3c)
where u b σh , σ b ph are the so-called numerical traces which have to be

bh , pbh , and u
suitably chosen. Here have used the notation
X Z X Z
(σ, τ )Ωh := σ : τ dx, hσ , v ⊗ niΩh := σ : (v ⊗ nK ) dγ,
K∈Th K K∈Th ∂K
X Z X Z
(u, v)Ωh := u · v dx, (u, τ · n)Ωh := u · τ · nK dx,
K∈Th K K∈Th K
X Z X Z
(p, q)Ωh := p q dx, hp , v · niΩh := p v · nK dx,
K∈Th K K∈Th K
where nK stands for the outward unit normal to ∂K.

For a definition of the numerical traces that render this DG method not
only well defined but optimally convergent, we refer the reader to [10]. We
thus consider that the definition of the method is complete.
Next, we show how to post-process the approximate solution (σh , uh , ph )
to obtain a new, exactly divergence-free velocity approximation uh .
2.2 The post-processing
We take uh in the space Vh and define it as follows. On the tetrahedron

K ∈ Th , uh is the element of Pk (K)3 determined by
huh · nK , ϕiF = hbup · nK , ϕiF ∀ϕ ∈ Pk (F ), for any face F ⊂ ∂K, (4a)

(uh , ∇ϕ)K = (u, ∇ϕ)K ∀ϕ ∈ Pk−1 (K), (4b)
(uh , φ)K = (u, φ)K ∀φ ∈ Φk (K), (4c)
where
Φk (K) = {φ ∈ Pk (K)3 : ∇ · φ = 0, φ · nK |∂K = 0}.
Notice that if in the conditions (4a), we replace ub p by u, the function uh is
nothing but the well known BDM projection of u; see [3]. As a consequence,
the function uh is well defined. Let us show that it is exactly divergence-free
provided the triangulation Th is conforming, that is, provided if two tetrahedra
have an intersection whose area is different from zero, such intersection is a
full face of each of them.
6 Bernardo Cockburn
Since the triangulation is conforming and since, by the conditions (4a), the
normal component across inter-element boundaries of uh is uniquely defined,
the velocity uh belongs to H(div, Ω) and we can write
(∇ · uh , q)Ω = (∇ · uh , q)Ωh = −(uh , ∇q)Ωh + huh · n , qi∂Ωh ,
and, if q is in the space Qh ,
uph · n , qi∂Ωh
(∇ · uh , q)Ω = − (uh , ∇q)Ωh + hb by (4a) and (4b),
= 0,
since uh satisfies equation (3c). This implies that ∇ · uh ≡ 0 on Ω, as claimed.
This shows that it is possible to obtain a exactly divergence-free velocity uh
from the completely discontinuous velocity uh and the numerical trace u b ph in
an element-by-element fashion.
A natural question is if the new velocity uh converges to the exact solution
with at least the same order of convergence than that of the original velocity
uh ? It is not difficult to see that their difference e := uh − uh depends only
on the normal component of the difference between the uh and the numerical
trace ub ph . Indeed, from the equations (4), we see that
he · nK , ϕiF = h(bup − uh ) · nK , ϕiF ∀ϕ ∈ Pk (F ), for any face F ⊂ ∂K,
(e, ∇ϕ)K = 0 ∀ϕ ∈ Pk−1 (K),
(e, φ)K = 0 ∀φ ∈ Φk (K).
For most DG methods, the difference (b up − uh ) · nK on ∂K depends solely on
the jump of the normal component of uh and the jump of ph . If the numerical
traces of the DG method are suitably chosen, one expects the jumps to be
small, we might actually have that uh and uh do converge with the same
b p at the interior of the domain was
order. In [10], where the numerical trace u
chosen to be the arithmetic average of the two traces of uh , it was proven
that this is indeed the case. Moreover, numerical experiments showed that uh
provided a better approximation than uh since its L2 -error was smaller by a
fixed constant. In a forthcoming paper, we show that this result also holds
when the local space for the pressure on the element K is taken to be Pk (K)
instead of Pk−1 (K) and the numerical trace u b p involves jumps of the pressure
ph .
The results in this section can be extended in a number of ways. First,
notice that any other DG approximation of the equations (3a) and (3b) could
have been used; see the discussion in [9]. Notice also that the finite element
spaces can be different, provided that a suitable post-processing is available.
For example, we could use cubes instead of tetrahedra, or a Raviart-Thomas
projection instead of the BDM projection; see [3].
2.3 Applications
A simple application arises in the passive transport of a given substance by a
fluid might be modeled by the equation
∂
c + ∇ · ( u c ) = 0,
∂t
where c is the concentration of the substance and u is the velocity given by
the Stokes system (1). When discretizing this equation, it is very convenient
to use uh since it is very well known that using an exactly divergence-free
velocity approximation gives better results.
A less evident application concerns the discretization of the steady-state
Navier-Stokes equations
−∆u + u · ∇u + ∇p =f on Ω,
∇ · u =0 on Ω,
u =g on ∂Ω.
In [10] it was pointed out that DG methods for the the Navier-Stokes equa-
tions that are energy-stable are not locally conservative and, vice versa, that
DG methods that are locally conservative are not energy-stable unless the
approximate velocity is exactly divergence-free. This difficulty was overcome
without having to resort to working with finite dimensional spaces of exactly
divergence-free velocities as follows.
It is known that the solution of the Navier-Stokes equations can be ob-
tained by solving a sequence of Oseen problems
−∆u + w · ∇u + ∇p =f on Ω,
∇ · u =0 on Ω,
u =g on ∂Ω.
Indeed, if we define Φ(w) as the exact solution u of the above Oseen equations,
it can be proven that the sequence {u(i) }i≥0 , where u(0) is arbitrary and
u(i+1) = Φ(u(i) ), converges to the exact solution of the Navier-Stokes equation
provided f /ν 2 is small enough. A discrete version of this argument provides
the construction of a DG method that is locally conservative and energy-
stable at the same time. Indeed, let us define Φh (w) as the post-processed
DG approximate velocity uh of the above Oseen equations. Then the sequence
(i) (0) (i+1) (i)
{uh }i≥0 , where uh is arbitrary and uh = Φh (uh ) does converges to
an approximation to the exact solution of the Navier-Stokes equation. Such
approximation has optimal approximation properties, as proven in [10]. There
are no other DG methods for the Navier-Stokes equation that are locally
conservative and energy-stable.
3 Hybridization of mixed methods

In this section, we describe a hybridization procedure that allows an easy
implementation of a vorticity-velocity mixed method for the Stokes equations
which uses finite dimensional spaces of divergence-free velocities. We begin
8 Bernardo Cockburn
by introducing the numerical method. We then present its hybridization and

show that it allows us to bypass the construction of the spaces of divergence-
free velocities. Moreover, we show that it results in a formulation whose only
unknowns are the tangential velocity and the pressure on the mesh faces. The
material of this section is contained in the papers [6, 7] and in the review
about hybridization techniques [8].
3.1 The mixed method
The mixed method we consider has been studied in [12, 14]. To define it, we
begin by rewriting the equations (1) in terms of the vorticity ω = ∇ × u.
Since
−∆u = ∇ × ∇ × u − ∇∇ · u,
the Stokes system (1) can be expressed as
ω−∇×u=0 on Ω, (5a)
∇ × ω + ∇p = f on Ω, (5b)
∇·u=0 on Ω, (5c)
u| = g| on ∂Ω, (5d)
u·n=g·n on ∂Ω, (5e)
where g| := g − (g · n) n is the tangential component of g. Now, if we set
W =H(curl, Ω),
V =H(div, Ω),
V(b) ={v ∈ V : ∇ · v = 0 and v · n|∂Ω = b}.
for any b ∈ H −1/2 (∂Ω), then (ω, u) is the only element of W × V(g · n)
satisfying
(ω, τ )Ω − (u, ∇ × τ )Ω = hg| , τ i∂Ω for all τ ∈ W,

(v, ∇ × ω)Ω = (f , v)Ω for all v ∈ V(0).
We use this vorticity-velocity formulation to define the mixed method.

Thus, given a triangulation Th made of tetrahedra K, we seek an approxima-
tion to (ω, u), (ωh , uh ), in the set Wh × Vh (gn,h ) where gn,h stands for the
L2 (∂Ω)-orthogonal projection of the normal component of the boundary data
g · n onto the space {vh · n|∂Ω : vh ∈ Vh }, and
Wh = {w ∈ W : w|K ∈ W (K) for all K ∈ T},

Vh = {v ∈ V : v|K ∈ V (K) for all K ∈ T},
Vh (b) = V(b) ∩ Vh .
The local spaces are given by

W (K) = Pk (K)3 ⊕ Sk+1 (K),
V (K) = {v ∈ Pk (K)3 : ∇ · v = 0},
where S` (K) is the set of all vector functions p` (x) whose components are
homogeneous polynomials of degree ` satisfying p` (x) · x = 0; see [13].
The approximation (ωh , uh ) is defined as the only function in the space
Wh × Vh (g · n) such that
(ωh , τ )Ω − (uh , ∇ × τ )Ω = hg| , τ i∂Ω for all τ ∈ Wh , (6a)

(v, ∇ × ωh )Ω = (f , v)Ω for all v ∈ Vh (0). (6b)
This completes the definition of the mixed method. Notice that, since the
velocities in the space Vh are divergence-free, the pressure does not appear in
its formulation.
Next, we show how to avoid having to construct the spaces Vh by using a
hybridization procedure.
3.2 The hybridized mixed method
Thus, instead of working with the space Vh , we consider the space

Vh = {v : v|K ∈ V (K) and ∇ · (v|K ) = 0 for all K ∈ T}
and set, accordingly,
Qh = {q ∈ L2 (F) : p = [[v · n]] for some v ∈ Vh }.
Here F is the set of all faces of the tetrahedra K ∈ Th and [[v · n]] is the jump
of the normal component of v which at the face F ∈ F is defined by
(
v·n if F ⊂ ∂Ω,
[[v · n]] := + + − −
v ·n +v ·n otherwise.
We assume that, when the face F does not lie on the boundary of Ω, we
have that F = ∂K + ∩ ∂K − , that v ± is the trace of v|K ± and that n± is the
outward unit normal to ∂K ± .
We define the approximation (ωh , uh , ph ) as the only function of the space
Wh × Vh × Qh satisfying
(ωh , τh )Ω − (uh , ∇ × τh )Ω = hg| , τh i∂Ω for all τh ∈ Wh ,

(vh , ∇ × ωh )Ω + hph , [[vh · n]]iF = (f , vh )Ω for all vh ∈ Vh ,
hqh , [[uh · n]]iF = hg · n , qh i∂Ω for all qh ∈ Qh .
Notice that, by the definition of the space of pressures Qh and the definition of
the normal component of the Dirichlet data gn,h , the last equation implies that
uh belongs to the space Vh (gn,h ). Moreover, since Vh ⊂ Vh (0), we see that
10 Bernardo Cockburn
(ωh , uh ) given by this method satisfies the equations (6). As a consequence, it

coincides with the approximate solution (ωh , uh ) of the original method (6).
Notice that even though we can now take the velocity test functions in
the bigger space Vh , we do obtain the very same approximate velocity and
vorticity. The price to pay is the introduction of new equations enforcing
the continuity constraints on the velocity as well as the introduction of a new
unknown, namely, the pressure on the faces F, which is the Lagrange multiplier
associated to the constraints.
This increase of the number of degrees of freedom in the velocity space,
of the number of equations and of the number of unknowns could seem to
be a setback. However, when a similar hybridization procedure is applied to
mixed methods for second-order elliptic equations, it is well known that the
original unknowns can be eliminated from the equations. This results in a
simpler equation for the Lagrange multiplier only; see [1, 5]. In our case, for
something similar to happen, we have to apply a hybridization procedure to
the vorticity. Next, we show how to do that.
Instead of working with the space Wh , we use the space
Wh = {w : w|K ∈ W (K) for all K ∈ T},
and set, accordingly,
Mh = {µ ∈ L2 (F \ ∂Ω)2 : µ|F\∂Ω = [[n × τ ]] for some τ ∈ Wh }.
We now define the new approximation (ωh , uh , λh , ph ) as the only function
in Wh × Vh × Mh × Qh satisfying
(ωh , τh )Ω − (uh , ∇ × τh )Ωh − hλh , [[n × τh ]]iF\∂Ω = hg| , n × τh i∂Ω , (7a)

(vh , ∇ × ωh )Ωh + hph , [[vh · n]]iF = (f , vh )Ω , (7b)
hqh , [[uh · n]]iF = hg · n , qh i∂Ω , (7c)
hµh , [[n × ωh ]]iF\∂Ω = 0, (7d)
for all (τh , vh , µh , qh ) ∈ Wh × Vh × Mh × Qh .

Notice that the function λh is an approximation to the tangential compo-
nent of the velocity on the interior faces. Multiplying the equation (5a) by a
test function τ and integrating on the element K, we obtain
(w, τ )K − (u, ∇ × τ )K − hu , n × τ i∂K = 0,
which, taking into account the boundary conditions, can be rewritten as
(w, τ )K − (u, ∇ × τ )K − hu| , n × τ i∂K\∂Ω = hg| , n × τ i∂K∩∂Ω .
If we add over all the tetrahedra K ∈ Th , we get the equation
(w, τ )Ω − (u, ∇ × τ )Ωh − hu| , [[n × τ ]]iF\∂Ω = hg| , n × τ i∂Ω ,
and a simple glance to equation (7a) is enough to convince us that the assertion
is justified.
The method (7) has a unique solution, as shown in [7], and, as the reader
might be expecting, the function (ωh , uh ) coincides with the approximate
solution (ωh , uh ) of the original method (6). (This is why this method is
called the hybridized version of the method (6).) Let us prove this claim.
Notice that, by equation (7c), the definition of gn,h and the definition of the
pressure space Qh , uh belongs to the space Vh (gn,h ). Similarly, by equation
(7d) and the definition of the tangential velocity space Mh , ωh belongs to
Wh . Finally, if we now take the test functions appearing in the equations (7a)
and (7b) in Wh × Vh (0) ⊂ Wh × Vh , they become identical to the equations
(6). This proves the claim.
Working with the hybridized version of the mixed method allows us to
deal only with vorticity and velocity test functions in the spaces Wh and Vh ,
respectively, which are certainly very easy to construct. Moreover, the fact
that these two spaces are sets of completely discontinuous functions, suggests
that it is reasonable to expect that the original variables ωh and uh could
be eliminated from the equations, as wanted. The resulting equations would
then characterize the two Lagrange multipliers, namely, the pressure ph on
the faces and the tangential velocity λh on the interior faces. Next, we give
such a characterization.
3.3 A characterization of the Lagrange multipliers
To state the main result of this section, we need to introduce some notation.
We define (w(λ), u(λ)) ∈ Wh × Vh and (W(p), u(p)) ∈ Wh × Vh element by
element as follows:
(w(λ), τ )K − (u(λ), ∇ × τ )K = hλ , n × τ i∂K , for all τ ∈ W (K),

(v, ∇ × w(λ))K = 0, for all v ∈ V (K),
(W(p), τ )K − (u(p), ∇ × τ )K = 0, for all τ ∈ W (K),
(v, ∇ × W(p))K = −hp , v · ni∂K , for all v ∈ V (K).
In addition, we define (w(f ), u(f )) and (w(g| ), u(g| )) in Wh × Vh by
(w(f ), τ )K − (u(f ), ∇ × τ )K = 0, for all τ ∈ W (K),

(v, ∇ × w(f ))K = (f , v)K , for all v ∈ V (K),
(w(g| ), τ )K − (u(g| ), ∇ × τ )K = hg| , n × τ i∂K∩∂Ω , for all τ ∈ W (K),
(v, ∇ × w(g| ))K = 0, for all v ∈ V (K).
We also define the bilinear forms

a(λ, µ) = (w(λ), w(µ))Ω ,
c(p, q) = (W(p), W(q))Ω ,
X
b(µ, p) = − (u(µ), ∇ × W(p))K ,
K∈T
12 Bernardo Cockburn
and the functionals

`1 (µ) = (f , u(µ))Ω − hg| , w(µ)i∂Ω
`2 (q) = (f , u(q))Ω + (gn , q)∂Ω − hg| , W(q) , ∂Ωi.
With this notation, we have the following result.
Theorem 1 ([7]). The Lagrange multiplier (λh , ph ) ∈ Mh × Qh of the hy-

bridized mixed method (7) is the unique solution of
a(λh , µ) + b(µ, ph ) = `1 (µ), for all µ ∈ Mh and (8a)

b(λh , q) − c(ph , q) = `2 (q), for all q ∈ Qh . (8b)
Moreover, the approximate solution (ωh , uh ) of the original mixed method (6)
can be determined locally as follows:
ωh = w(λh ) + W(ph ) + w(g| ) + w(f ), (9a)

uh = u(λh ) + u(ph ) + u(g| ) + u(f ). (9b)
The relevance of this result is that it solves the impasse generated by

the impossibility of the actual construction of the spaces of divergence-free
approximate velocities. Moreover, it shows that the approximate solution can
be obtained by solving a problem for the tangential velocity and pressure only.
Since these variable lie on the set of faces F, the number of degrees of freedom
we have to solve for is significantly smaller than that of the original variables.
It is important to emphasize that, if the Lagrange multiplier spaces Qh
and Mh were as difficult to construct as the space of divergence-free velocities,
the hybridization under consideration would have been completely pointless.
However, such a construction is extremely easy, as shown in [6, 7], where
more details about the implementation of this saddle-point problem can also
be found.
4 Hybridization of DG methods
We have seen how to post-process a DG method which uses completely dis-
continuous approximate velocities to obtain an exactly divergence-free ve-
locity. We have also shown how to hybridize a mixed method using exactly
divergence-free velocity spaces to bypass the impossible task of constructing
those spaces. One then wonders if hybridization could also be used on DG
methods that use exactly divergence-free velocity spaces [10, 9] and, in the
affirmative case, if a result similar to the theorem of section 3 holds?
In [4], it was shown how to hybridize a DG method based on a vorticity-
velocity formulation. However, unlike the hybridization carried out for the
mixed method, only the Lagrange multiplier of the pressure was introduced
and only the vorticity could be eliminated from the equations. A similar hy-
bridization procedure could be applied to DG methods not based in such
formulation, like the ones briefly considered in [10] and later further explored
in [9]. However, also for these DG methods, a result similar to the theorem of
section 3 seems difficult to obtain.
On the other hand, there is a wide class of DG methods than can be hy-
bridized just as the mixed method considered in section 3 was. This constitutes
the subject of a forthcoming paper.
References
1. D. N. Arnold and F. Brezzi, Mixed and nonconforming finite element meth-
ods: implementation, postprocessing and error estimates, RAIRO Modél. Math.
Anal. Numér., 19 (1985), pp. 7–32.
2. P. Bastian and B. Rivière, Superconvergence and H(div) projection for dis-
continuous Galerkin methods, Internat. J. Numer. Methods Fluids, 42 (2003),
pp. 1043–1057.
3. F. Brezzi and M. Fortin, Mixed and Hybrid finite element methods, Springer
Verlag, 1991.
4. J. Carrero, B. Cockburn, and D. Schötzau, Hybridized, globally
divergence-free LDG methods. Part I: The Stokes problem, Math. Comp., 75
(2006), pp. 533–563.
5. B. Cockburn and J. Gopalakrishnan, A characterization of hybridized mixed
methods for second order elliptic problems, SIAM J. Numer. Anal., 42 (2004),
pp. 283–301.
6. , Incompressible finite elements via hybridization. Part I: The Stokes sys-
tem in two space dimensions, SIAM J. Numer. Anal., 43 (2005), pp. 1627–1650.
7. , Incompressible finite elements via hybridization. Part II: The Stokes sys-
tem in three space dimensions, SIAM J. Numer. Anal., 43 (2005), pp. 1651–1672.
8. , New hybridization techniques, GAMM-Mitt., 2 (2005), pp. 154–183.
9. B. Cockburn, G. Kanschat, and D. Schötzau, A note on discontinuous
Galerkin divergence-free solutions of the Navier-Stokes equations. Submitted.
10. , A locally conservative LDG method for the incompressible Navier-Stokes
equations, Math. Comp., 74 (2005), pp. 1067–1095.
11. B. Cockburn, D. Schötzau, and J. Wang, Discontinuous Galerkin methods
for incompressible elastic materials, Comput. Methods Appl. Mech. Engrg., 195
(2006), pp. 3184–3204. C. Dawson, Ed.
12. V. Girault and P. A. Raviart, Finite element approximations of the Navier-
Stokes equations, Springer-Verlag, New York, 1986.
13. J.-C. Nédélec, Mixed finite elements in R3 , Numer. Math., 35 (1980), pp. 315–
341.
14. , Éléments finis mixtes incompressibles pour l’équation de Stokes dans R3 ,
Numer. Math., 39 (1982), pp. 97–112.
Role of High-End Computing in Meeting
NASA’s Science and Engineering Challenges
Rupak Biswas, Eugene L. Tu, and William R. Van Dalsem
NASA Ames Research Center, Moffett Field, CA 94035, USA

{Rupak.Biswas, Eugene.L.Tu, William.R.VanDalsem}@nasa.gov
Abstract
High-end computing (HEC) has always played a major role in meeting the
modeling and simulation needs of various NASA missions. Two years ago,
NASA was on the verge of dramatically enhancing its HEC capability and
capacity by significantly increasing its computational and storage resources.
With the 10,240-processor Columbia supercomputer in production since Oc-
tober 2004, HEC is having an even greater impact within the Agency and
beyond. Advanced science and engineering simulations in space exploration,
Shuttle operations, Earth sciences, and fundamental aeronautics research are
occurring on Columbia, demonstrating its ability to accelerate NASA’s ex-
ploration vision. This paper describes how the integrated production environ-
ment fostered at the NASA Advanced Supercomputing (NAS) facility at Ames
Research Center is reducing design cycle times, accelerating scientific discov-
ery, achieving rapid parametric analyses of multiple scenarios, and enhancing
safety for several NASA missions. We focus on Columbia’s impact on two key
engineering and science disciplines: aerospace, and climate/weather. We also
discuss future mission challenges and plans for NASA’s next-generation HEC
environment.
1 Introduction
Over the years, high-end computing (HEC) has played a major role in meet-
ing the modeling and simulation needs of various NASA missions. Two years
ago, having projected its near-term and future high-fidelity computational
requirements, NASA was on the verge of dramatically increasing its HEC
capability and capacity [4]. With NASA’s 10,240-processor supercomputer,
Columbia, in production since October 2004, HEC is having an even greater
impact within the Agency and extending to partner institutions. Significant
cutting-edge science and engineering simulations in the areas of space ex-
ploration, Shuttle operations, Earth sciences, and fundamental aeronautics

High-End Computing at NASA 15
research are occurring regularly on Columbia, demonstrating its ability to ac-

celerate NASA’s vision for space exploration [6, 7]. This paper describes how
the integrated production environment fostered at the NASA Advanced Su-
percomputing (NAS) facility located at Ames Research Center is being used
to design future aerospace vehicles, conduct parametric analysis for safe op-
eration of the Shuttle, accelerate scientific discovery, and enhance crew safety
during the life cycle of NASA missions. Columbia’s impact is illustrated us-
ing two of the agency’s key engineering and science disciplines: aerospace and
climate/weather.
In aerospace, computed results are presented in three areas: debris trans-
port analysis for the Space Shuttle Launch Vehicle (SSLV); flowliner analysis
for the Space Shuttle Main Engine (SSME); and risk assessment of ascent
abort scenarios for proposed Crew Exploration Vehicle (CEV) designs. Among
NASA’s applications in climate and weather modeling are next-generation
global ocean models that resolve eddies and other narrow current systems,
and atmospheric modeling and prediction of hurricane tracks for early warn-
ing. Columbia is also having a significant impact on NASA’s numerous space
and exploration applications, such as the development of the Crew Launch Ve-
hicle (CLV), and risk assessment throughout the entire mission cycle—from
ground operations, vehicle launch, and return to Earth.
The role of the Columbia supercomputer (currently ranked the fourth
fastest system in the world [24], at 62 TFlop/s peak performance) in advanc-
ing the science and technologies related to the above topics are illustrated
through various data analysis methods. Users of Columbia are also supported
by the NAS facility’s integrated HEC environment. In addition to system
analysts, experts in code parallelization and performance optimization, high-
fidelity modeling and simulation, high-speed networking, and data analysis
and visualization exploit the power of Columbia to enhance NASA’s compu-
tational capability.
As with other federal agencies and the commercial sector, NASA’s future
mission challenges require even more powerful computing systems. At present,
the Agency is planning its next-generation HEC system, which will augment
Columbia and provide even more high-performance computing resources. De-
velopment of this future environment includes anticipated storage and archive
requirements for a balanced system, application performance enhancement
tools, and faster wide area networking technologies.
2 Aerospace Applications
2.1 Space Shuttle Launch Vehicle Debris Transport Analysis
After the STS-107 incident in February 2003, the Columbia Accident Inves-
tigation Board (CAIB) requested that CFD researchers conduct a detailed
debris transport analysis to provide insight into the actual mechanism of de-
bris shedding from the bi-pod ramp region of the SSLV. The analysis would
16 Biswas, Tu, and Van Dalsem
also furnish input for foam velocity and density for impact testing to deter-
mine damage to the orbiter’s wing leading edge [18].
Subsequently, researchers at NASA Ames developed a CFD process for
determining the aerodynamic characteristics of debris shedding during the
SSLV ascent phase. In 2005, the role of Columbia’s predecessor system in
the accident investigation [4] was reported. Since that time, a complete de-
bris scenario has been conducted on Columbia, which focused on predicting
the aerodynamic characteristics of potential debris sources, such as insulating
foam and ice. This computational analysis was critical to NASA’s Return-to-
Flight (RTF) effort, where scientists performed six-degree-of-freedom (6-DOF)
foam debris transport analyses and visualization to forecast Shuttle damage,
and for damage assessment and repair recommendations during the successful
Discovery flight in July 2005. HEC and CFD will continue to play a key role
for remaining Shuttle flights; high-fidelity simulations being integral to launch
risk analysis and component redesign.
For future flights, debris analysis has been directed to an assortment of
shapes (typically thin and conical) that can potentially be shed from the
external tank (ET) foam. The debris sources and their aerodynamic charac-
teristics are put into the debris transport code, which calculates trajectory
information to assess the potential damage or risk by a specific debris source
to a specific structural component, such as the ET foam impacting the or-
biter wing. For this interactive process to be effective, the debris transport
analysis must be done rapidly. A single trajectory calculation requires 30–60
CPU-hours on the Columbia system, which provides enough throughput to
rapidly and efficiently run hundreds of trajectories in a day, using only a frac-
tion of the computational resources. A system with the power of Columbia is
absolutely essential to run the typically hundreds of thousands of trajectories
analyzed over the entire vehicle for each iteration. Results in Fig. 1 show that
the average drag for the oscillating trajectory of an idealized frustum and the
tumbling trajectory of a highly asymmetric debris piece are similar. Also note
that farther downstream the debris travels before impact, the greater is the
impact kinetic energy, as the aerodynamic drag is constantly increasing the
relative velocity between the debris and the orbiter.
This is not the case when considering the crossrange behavior (see Fig. 2).
The dynamically stable oscillating frustum generates almost no crossrange,
as the lift force oscillates first in one direction and then in the other, with
little net effect. In order to provide a representative distribution, researchers
used several shapes to develop the crossrange constraints. These include real
digitized shapes, idealized frustums, ideal frustums with the center of mass
offset, and slightly asymmetric shapes such as elliptical frustums with the
small diameter slightly offset from the larger. The crossrange envelopes show
a zero-lift trajectory emanating from the ET flange region. The intersection
of this cone with the SSLV indicates that the fuselage and wing of the orbiter
have potential for debris impacts from this flange location, along with regions
of the left solid rocket booster.
Fig. 1. Comparison of drag (left) and kinetic energy (right) for various debris shapes
released at Mach 2.5. Symbols show unconstrained 6-DOF trajectories compared
with trajectories using a nominal drag model based on the ensemble-averages.
Fig. 2. Crossrange envelope superimposed upon the computed ballistic zero-lift

trajectory. A statistical distribution of the crossrange within the envelope can be
used for risk analysis.
The emphasis of this work is the development of an efficient process for

modeling debris beyond ET insulating foam, including the ET liquid oxygen
(LOX) frost ramps, insulating cork on the solid-rocket boosters, frost and ice
on the ET acreage regions, and ice that can form on the ET feedline brackets.
The flexibility of the modeling and simulation capability and the computing
resources provided by Columbia allows the dynamic behavior of these diverse
debris sources to be analyzed in a systematic and timely manner [16].
2.2 Space Shuttle Main Engine Flowliner Analysis
In May 2002, numerous cracks were found in the SSME #1 flowliner; specifi-
cally, at the gimbal joint in the liquid hydrogen (LH2 ) feedline flowliner. Since
then, repairs have been made to existing cracks on all orbiters. Long-term sci-
entific investigations continue, because the root cause of the original cracks
was not conclusively established and remaining Shuttle flights are involved
until 2010.
High-fidelity computations have been conducted on the Columbia super-
computer to analyze the SSME LH2 feedline flowliner [12]. Numerous com-
putational models were used to characterize the unsteady flow features in the
turbopump, including the Low-Pressure-Fuel-Turbopump (LPFTP) inducer,
the orbiter manifold, and an experimental hot fire test article representing
the manifold. Findings show that unsteady flow stemming from the LPFTP
inducer is one of the major contributors to high-frequency cyclic loading that
results in fatigue damage to the flowliners.
The flow fields for the orbiter manifold and the hot fire test article were
computed and analyzed on Columbia using the INS3D incompressible Navier-
Stokes flow solver [10, 11, 13] The first computational model included only
the LPFTP inducer; by studying it, scientists were able to compare unsteady
pressure values against existing data. To resolve the complex geometry in
relative motion, an overset grid methodology [8] was employed, containing
57 overlapping zones with 26.1 million grid points. The second computational
grid system, consisting of 264 overset grids with 65.9 million grid points, added
the flowliner geometry and is shown in Fig. 3. The flowliner component alone
contained 212 grids and 41 million points.
Fig. 3. Computational model for LPFTP inducer and the LH2 flowliner: Grid (left),
and computed results illustrating unsteady interaction of the flow in the bellows
cavity and the back flow from the inducer (right).
To accelerate the grid generation process, scripts were developed to au-

tomatically create grids for each type of component. The size of the simula-
tion is large, requiring parallel processing to obtain solutions with reasonable
turnaround times. Two parallel programming paradigms were leveraged in the
INS3D code: Multi-Level Parallelism (MLP) [23] and hybrid MPI+OpenMP.
Performance for the two programming models was similar; however, only the
MPI+OpenMP implementation can be executed on multiple nodes. Multi-
node computations on Columbia showed that point-to-point implementation
Fig. 4. INS3D performance on multiple nodes of Columbia.
of the MPI communication performs more efficiently than the master-worker

version [5] (see Fig. 4).
Results of the CFD calculations confirmed the presence of back flow caused
by the LPFTP inducer. The region of reverse flow extended far enough up-
stream to interfere with both flowliners in the gimbal joint. Computed results
for the test article were verified by correlation with pressure measurements,
and confirmed a strong unsteady interaction between this back flow and the
secondary flow in the bellows cavity through the flowliner slots. It was ob-
served that a swirl on the duct side of the downstream flowliner is stronger
than on the same side of the upstream flowliner, causing significantly stronger
unsteady interactions through the downstream slots. This turbopump applica-
tion using INS3D currently exhibits some of the best scalability performance
on the Columbia system.
2.3 Crew Exploration Vehicle Abort Risk Assessment
Researchers are running high-fidelity CFD codes on Columbia to guide the

designs of future space vehicles, including the CEV, and are building realistic
models and developing new technologies to simulate flight risks for these new
spacecraft. Risks and performance issues during both the ascent and entry-
descent-landing phases are being carefully analyzed. The CEV will replace
the Shuttle in 2010, and transport a maximum of six crew members to and
from the International Space Station and up to four astronauts to and from
the moon.
The CEV design includes a Launch Abort System (LAS) for crew escape,
similar to that used in the Apollo capsule. Several computational modeling
and simulation tools for analyzing abort scenarios have recently been devel-
oped and enhanced for use on Columbia. Under the simulation assisted risk
assessment (SARA) project, NASA researchers have developed a probabilistic
risk assessment (PRA) approach and demonstrated how risk analysis can be
applied to launch abort using the Apollo configuration [14]. A PRA identifies
the best level of fidelity for modeling critical failure modes associated with
launch abort. Columbia is then used to conduct higher-fidelity modeling on
specific failure modes. Two failure modes examined so far include booster
explosion and those caused by re-contact with the booster during separation.
Analysis of the booster failure mode using Apollo data showed a possible
catastrophic failure, leading to detonation of the propellant and creating blast
wave overpressures that could fatally damage the LAS (see Fig. 5). As the risk
model was being developed, it became apparent that the booster type and the
nature of the failure it was likely to encounter, determined the environments
Fig. 5. Surface pressures (left) and flowfield (right) for blast wave propagating
through wake of maneuvering LAS at t = 41.6, 72.0, and 85.5 msec (top to bottom).
under which the LAS must operate to ensure a successful abort. The process
for characterizing this interaction must be carefully modeled and simulated.
One particular weakness found in an engineering-level model was the effect
of headwind as the CEV ascends. To account for these effects in the risk
analysis, high-fidelity blast wave models were built and simulated on Columbia
using the Overflow Navier-Stokes code [15]. Results showed that headwinds
significantly affect the nature and magnitude of the shock wave as it impacts
an escaping CEV. This means that the warning time required to initiate the
abort sequence is also affected. Additional work in high-fidelity simulations is
being done to help engineers generate requirements for the LAS while taking
headwind into consideration.
Another failure mode dependent on high-fidelity simulation involves the
ability of the LAS to achieve clean separation of the CEV from the booster
stack in the event of impending catastrophic failure. Simply put, the CEV
must not scrape or re-contact the booster stack. This failure mode is par-
ticularly demanding because it involves complex proximity aerodynamics—
modeling transonic flow as well as the complex flow at the small gap between
the CEV and the booster stack at separation. Both Navier-Stokes simulations
using Overflow, and Euler simulations using FlowCart [1], were conducted,
and their results validated against transonic wind tunnel and abort flight test
data from the Apollo era [4].
All these cases are computationally expensive to simulate. The complexity
of the geometry and the flow-field required about 30 million grid points, which
enabled good scalable performance up to 250 processors. About 20 cases were
computed using Overflow at various ascent trajectories and separation thrust
levels. Each case required approximately 20,000 CPU-hours on Columbia, in-
cluding the computation of the initial steady-state solution. All failure modes
benefited immensely from the HEC resources at the NAS facility. These tools
and processes will likely be applied to analyze the actual LAS design, and to
further understand the CEV failure modes and their impact on the vehicle’s
survivability.
3 Climate and Weather Applications
3.1 Global Ocean Modeling
To increase understanding and predictive capability for the ocean’s role in fu-
ture climate change scenarios, NASA has initiated a project called Estimating
the Circulation and Climate of the Ocean, Phase II (ECCO2): High-resolution
global-ocean and sea-ice data synthesis [17]. The goal is to produce increas-
ingly accurate combinations of all available global-scale ocean and sea-ice data
at resolutions that begin to resolve ocean eddies and other narrow current sys-
tems, which transport heat, carbon, and other properties in the ocean. These
data syntheses are used to quantify the role of the oceans in the Earth’s
carbon cycle, understand recent changes in the polar oceans, and monitor
time-evolving term balances within and between different components of the
Earth system. This work aims to harness NASA’s computational resources
such as Columbia, advances in CFD and software engineering, and the ability
to solve massive control problems.
The most challenging numerical experiment undertaken to date is a near-
global simulation with 1/16 ◦ horizontal grid spacing (approximately 6 km at
the Equator and 1 km at high latitudes). The number of surface grid cells
is about 25 million and the configuration has 50 vertical levels, bringing the
total number of cells to just over 1.25 billion. Each of the 3D fields that
describe the simulation domain and its time-evolving state requires 10 GB
of storage. This configuration has been integrated on the 2,048-CPU sub-
cluster of Columbia [9] (see Sec. 4). This workload is distributed evenly over
1,920 processors, so that each CPU is responsible for simulating about 586,000
grid cells (equivalent to a surface region roughly 210×210 km2 ). Decomposing
the workload over this many processors yields a setup that, with extensive
diagnostics and analysis options included, uses about 870 MB of main memory
per processor. With a timestep of two minutes, this performance allows a year
of simulation to be completed in less than ten days.
To investigate solution convergence as horizontal resolution is increased,
ECCO2 researchers are conducting a series of numerical simulations at 1/4 ◦ ,
1/8 ◦ , and 1/16 ◦ resolutions. Figure 6 shows significant changes in solution
with varying resolution. Each plot captures the change in simulated sea-
surface height due to eddy activity over a single month. Variation with reso-
lution occur in regions where eddies are prevalent (such as the Gulf Stream,
the Kuroshio, the Agulhas, the Drake Passage, and the Antarctic Circumpo-
lar Current). For example, in the Gulf Stream, the area where the sea-surface
height changes vigorously increases with higher resolution. Key behaviors,
such as how tightly waters stick to the coast, or how far energetic eddies
penetrate the ocean interior, also change significantly between resolutions.
Fig. 6. Gulf Stream region sea-surface height difference plots for one month at
resolutions of 1/4 ◦ , 1/8 ◦ , and 1/16 ◦ (left to right). Color scale -0.125m to 0.125m.
Performance on Columbia shows that it is well suited for addressing these

questions. The ECCO2 code achieves about 722 MFlop/s per CPU when run-
ning on 1,920 processors; this is 14 percent of the per-CPU Linpack bench-
mark performance achieved on Columbia [24]. The code consists of predomi-
nantly BLAS1 operations and cannot exploit the level of cache reuse that Lin-
pack achieves. The scaling across multiple nodes is encouraging and suggests
that configurations spanning eight or more 512-processor Altix systems—that
would therefore support 1/20 ◦ and higher resolutions—are within reach.
3.2 Atmospheric Modeling and Hurricane Prediction
The NASA Finite Volume General Circulation Model (fvGCM) is a unified

numerical weather prediction (NWP) and climate model that could run on
daily, monthly, decadal, and century time-scales. The model was originally
designed for climate studies at a coarse resolution of about 2.5◦ , but has been
running at much finer resolution on the Columbia supercomputer to answer
the following question for NASA’s mission in hurricane research [19]: How can
weather/hurricane forecasts be improved and made more reliable over longer
periods of time using computer modeling?
Hurricane forecasts pose challenges for general circulation models (GCMs),
the most important being horizontal grid spacing. With the unique computing
resources of Columbia, the model horizontal resolution was rapidly increased
to 1/4 ◦ in 2004 [2], and 1/8 ◦ in early 2005. Recently, researchers have tested
a 1/12 ◦ resolution version, which is the first global weather model with a
single-digit resolution (9 km at the Equator). A five-day forecast of total
precipitable water with the 1/12 ◦ degree fvGCM (see Fig. 7) shows fine scale
weather events in the tropical area, which brings researchers to overcoming
the fundamental barrier between global and mesoscale models [20].
Fig. 7. Five-day forecasts of total precipitable water initialized in September 2004

with the 1/12 o fvGCM, giving a grid spacing of 9 km at the Equator.
During the 2004 hurricane season, the 1/4 ◦ model, which doubled the
resolution adopted by most global models in operational NWP centers at that
time, was running in real-time and providing remarkable landfall predictions
up to five days in advance for major hurricane such as Charley, Frances, Ivan,
and Jeanne [2]. Moreover, the model proved capable of resolving problems such
as erratic track, abrupt recurvature, and intense extratropical transition. In
the 2005 hurricane season, new research focused on the validation of the 1/8 ◦
fvGCM performance on hurricane forecasts, while the real-time 1/4 ◦ forecasts
provided a baseline for comparisons. Being a global mesoscale-resolving model,
the 1/8 ◦ resolution was the first to simulate mesoscale vortices (such as the
Catalina Eddy and the Hawaiian Lee Vortex), which were generated by the
interaction of the large-scale flows with better resolved surface forcing.
The 2005 Atlantic hurricane season was the most active in recorded history.
There were 28 tropical storms and 15 hurricanes, four of which were Category
5. Accurate forecasts of these storms was a significant challenge to global
and mesoscale modelers. It is well known that GCMs’ insufficient resolutions
undermine intensity predictions. Using the power of Columbia, NASA re-
searchers demonstrated that this limitation could be overcome by performing
six 5-day forecasts of hurricane Katrina [21] with the 1/8 ◦ fvGCM, showing
promising intensity forecasts with small errors in center pressure of only ±12
hPa. Notable improvement in Katrina’s intensity forecasts occurred when the
grid spacing decreased from 1/4 ◦ to 1/8 ◦ , at which the near-eye wind distri-
bution and the radius of maximum wind could be resolved more realistically.
While the mesoscale-resolving fvGCM has produced very promising results
for the past two years, significant potential for further advancement is still
ahead.
4 Columbia Description
When the Columbia supercomputer became fully operational in October 2004,

it increased NASA’s computing capability ten-fold and helped revitalize the
HEC effort within the U.S. Constructed in just four months, it has enabled
scientists and engineers to perform important, breakthrough simulations in
several arenas. Performing at a peak speed of 62 TFlop/s, Columbia has
demonstrated its capability to support and accelerate the space agency’s key
missions and vision for exploration [6, 7].
Columbia is a 10,240-processor SGI Altix constellation composed of twenty
512-CPU nodes, twelve of which are model 3700, and the remaining eight are
double-density 3700Bx2. Each node is a single-system-image (SSI) system,
with 2 GB of shared memory per processor (1 TB per node). It uses SGI’s
NUMAflex global shared-memory architecture that allows access to all data
directly and efficiently, without having to move them through I/O or net-
working bottlenecks. Within each node, the processors are interconnected via
SGI’s proprietary NUMAlink fabric. The 3700 utilize NUMAlink3 with a peak
bandwidth of 3.2 GB/s, while the Bx2 have NUMAlink4 where the bandwidth
is doubled to 6.4 GB/s.
The 20 Columbia nodes are connected by Voltaire InfiniBand fabric, as
well as via 10- and 1-gigabit Ethernet connections. Four of the Bx2 nodes
are tightly linked with NUMAlink4 (as well as the other fabrics) to form a
2,048-processor shared memory environment. Each processor in the 2,048-
CPU subsystem is a 64-bit Intel Itanium2, running at 1.6 GHz, with 9 MB of
level 3 cache and a peak performance of 6.4 GFlop/s. One other Bx2 node is
equipped with these same processors. The remaining 15 nodes have Itanium2
processors running at 1.5 GHz, with 6 MB of level 3 cache, and a peak per-
formance of 6.0 GFlop/s. Columbia is attached to more than 1 PB of online
RAID storage through a Fibre Channel switch.
Each 512-processor Altix node has several salient features that make it
particularly well-suited for executing large-scale compute and data-intensive
applications that are interesting to NASA. For example, its less than 1 mi-
crosecond latency to memory significantly reduces the communication over-
head. Typical problems are physics-based simulations involving a discretized
grid of the physical domain that is partitioned across multiple processors ex-
changing boundary data at every time step. Columbia also was the first system
(in November 2004) to exceed 1 TB/s memory bisection bandwidth on the
STREAM benchmark [22]. With global shared memory and cache coherency,
the nodes enable application programmers to use simpler and more efficient
programming models than message passing. Problems requiring adaptive grid-
ding and dynamic load balancing are much easier to program, control, and
solve on Columbia, leveraging shared memory paradigms such as OpenMP
and MLP [23]. In addition, the large shared memory of 1 TB per node allows
bigger problems to remain resident on the system.
The development and operating environment on Columbia features a 64-
processor SGI Altix front end, a Linux-based operating system, Altair PBS
Professional job scheduler, Intel Fortran/C/C++ compiler, and SGI ProPack
software.
5 Concluding Remarks and Future Vision
Over the last few decades, simulation methodologies have generally advanced
along with computational technologies. Advanced tools have been developed
to the point that many daily engineering and science problems can now be
routinely computed; however, this is still done mostly using geometrically and
or physically simplified or truncated models. Some of the physical models, such
as those for turbulence and transition, and for high-temperature real gas, have
not been advanced much more than what was available in the 1970s or early
’80s.
To realize the full benefit of HEC, more inclusive modeling of geometry
and physics is needed. Attempts to solve these problems have been made with
some qualitative success. However, predictive capability is still very limited

and prediction with accurate physics is yet to be accomplished; this will require
inclusion of not only fluid dynamic quantities but other factors such as thermal
loading, structural properties, and control. These computations will require
not only larger computing resources but also increased storage capacity and
sophisticated data management technologies.
Many of Columbia’s scientific and engineering users have stated that the
system has allowed them to successfully complete investigations they never
allowed themselves to dream of previously. Now, these users are envision-
ing what they can accomplish when even more powerful computing systems
are available. NASA and the HEC community are working on developing
petaflops-scale computers that can execute at rates more than 1015 operations
per second. For example, the DARPA High Productivity Computer Systems
(HPCS) program is working with several US computer vendors to provide a
new generation of economically viable supercomputer systems before the end
of this decade. The National Science Foundation (NSF) has a plan to provide
computational resources to the general scientific community that can sustain
petaflop performance by 2010. NASA is also planning its next-generation su-
percomputer system to meet the ever-increasing demand for computational
resources required for a wide range of Agency-specific scientific discoveries
and engineering applications.
For example, with NASA’s next-generation system, scientists envision a
launch simulation model designed to treat the entire launch environment un-
til the vehicle has cleared the launch tower. The model would integrate 6-DOF
multiple-body motion, debris transport and impact, propulsion system vibra-
tion and exhaust, acoustics due to exhaust, fuel accumulation in the exhaust
plume, exhaust chemistry including fuel burning, thermal stress on the vehicle
structure, and weather at the launch site. This very complex digital launch
model would integrate data from propulsion simulation, meso-scale weather
prediction, and experiment. Utilizing state-of-the-art flow simulation tools
and petaflops-scale computing systems, researchers can attempt to compute
a complete high-fidelity aerodynamic simulation with a realistic turnaround
time—within a few days rather than several weeks.
In aerospace design, the most productive aspect of HEC applications has
been to predict relative change among design variations. To push the limit
of operation and to try bold new ideas, more predictive capabilities will be
needed for complicated physical phenomena. Without accurate prediction, the
capability impacts of HEC can be limited to the current level, even if more
advanced facilities become available. To make these advances, high-fidelity
computations using HEC facilities will be absolutely critical despite all the
excitement about inexpensive PC clusters and distributed grid computing.
In Earth sciences, as resolutions increase significantly, the horizontal scales
become smaller and the hydrostatic assumption is no longer valid. Non-
hydrostatic dynamics, including eddy-resolving oceans, cloud-resolving atmo-
sphere, and land models coupled with chemical and biological components,
must therefore be included when running climate and weather applications.

New schemes are also needed to better represent physical processes at less
than 10-km resolutions. New grid systems (e.g. geodesic or icosahedral grids)
are required due to inefficiencies of non-uniform latitude-longitude grids at
ultra-high resolutions or convergence/stability problems at the poles.
A final note is related to human resources. Although modeling and sim-
ulation technology has advanced remarkably, many challenging cases require
experts in computational physics. Computer science can automate a good
portion of the simulation processes, thus saving a large amount of the human
effort. However, blind application of tools without understanding capabilities
and limitations of the methods involved could lead to catastrophic engineer-
ing results. As in many other engineering and science disciplines, modeling
and simulation researchers and practitioners need to understand physics and
the engineering systems being simulated. Experts who are willing to think
through the flow physics in addition to software engineering, must still be
developed for future generations.
References
1. M.J. Aftosmis, M.J. Berger, and G. Adomavicius, A parallel multilevel method
for adaptively refined Cartesian grids with embedded boundaries, 38th AIAA
Aerospace Sciences Mtg., AIAA Paper 2000-0808 (Reno, NV, 2000).
2. R. Atlas, O. Reale, B.-W. Shen, S.-J. Lin, J.-D. Chern, W. Putman, T. Lee,
K.-S. Yeh, M. Bosilovich, and J. Radakovich, Hurricane forecasting with the
high-resolution NASA finite volume general circulation model, Geophys. Res.
Lett. 32(3), 2005, L03807, doi:10.1029/2004GL021513.
3. F.R. Bailey, High-end computing challenges in aerospace design and engineering,
3rd Intl. Conf. on CFD (Toronto, Canada, 2004), pp. 13–26.
4. B.L. Barrier and O.C. Pendergraft Jr., Transonic aerodynamic characteristics
of a powered wind-tunnel model of the Apollo Launch Escape Vehicle during
separation, NASA TM-X 1336, 1967.
5. R. Biswas, M.J. Djomehri, R. Hood, H. Jin, C. Kiris, and S. Saini, An
application-based performance characterization of the Columbia supercluster,
SC05 (Seattle, WA, 2005).
6. R. Biswas, D. Kwak, C. Kiris, and S. Lawrence, Impact of the Columbia super-
computer on NASA space and exploration missions, 2nd Intl. Conf. on Space
Mission Challenges for Information Technology (Pasadena, CA, 2006), pp. 51–
58.
7. W. Brooks, M.J. Aftosmis, B. Biegel, R. Biswas, R. Ciotti, K. Freeman, C.
Henze, T. Hinke, H. Jin, and W. Thigpen, Impact of the Columbia super-
computer on NASA science and engineering applications, 7th Intl. Wkshp. on
Distributed Computing (Kharagpur, India, 2005), Springer LNCS 3741, 2005,
pp. 293-305.
8. P.G. Buning, D.C. Jespersen, T.H. Pulliam, W.M. Chan, J.P. Slotnick, S.E.
Krist, and K.J. Renze, Overflow user’s manual—version 1.8g, Technical Report,
NASA Langley Research Center (Hampton, VA, 1999).
9. C. Hill, D. Menemenlis, R. Ciotti, and C. Henze, Investigating solution con-

vergence in a global ocean model using a 2048-processor cluster of distributed
shared memory machines, J. Scientific Programming, submitted.
10. C. Kiris and D. Kwak, Numerical solution of incompressible Navier-Stokes equa-
tions using a fractional-step approach, Computers & Fluids 30(7), 2001, pp. 829–
851.
11. C. Kiris, D. Kwak, and W.M. Chan, Parallel unsteady turbopump simulations
for liquid rocket engines, SC2000 (Dallas, TX, 2000).
12. C. Kiris, D. Kwak, W.M. Chan, and J. Housman, High-fidelity simulations for
unsteady flow through turbopumps and flowliners, 44th AIAA Aerospace Sci-
ences Mtg., AIAA Paper 2006-0089 (Reno, NV, 2006).
13. D. Kwak, J.L. Chang, S.P. Shanks, and S. Chakravarthy, An incompressible
Navier-Stokes flow solver in three-dimensional curvilinear coordinate systems
using primitive variables, AIAA J. 24(3), 1996, pp. 390–396.
14. S. Lawrence, D.L. Mathias, G. Klopfer, S. Pandya, M.E. Olsen, J. Onufer, T.
Holst, and K. Gee, Simulation assisted risk assessment, 44th AIAA Aerospace
Sciences Mtg., AIAA Paper 2006-0090 (Reno, NV, 2006).
15. S. Lawrence, D.L. Mathias, K. Gee, and M.E. Olsen, Simulation assisted risk
assessment: Blast overpressure modeling, 8th Intl. Conf. on Probabilistic Safety
Assessment and Management, Paper 0197 (New Orleans, LA, 2006).
16. D.J. Mavriplis, M.J. Aftosmis, and M.J. Berger, High resolution aerospace ap-
plications using the NASA Columbia supercomputer, SC05 (Seattle, WA, 2005).
17. D. Menemenlis, C. Hill, A. Adcroft, J.-M. Campin, B. Cheng, R. Ciotti, I.
Fukimori, P. Heimbach, C. Henze, A. Köhl, T. Lee, D. Stammer, J. Taft, and
J. Zhang, NASA supercomputer improves prospects for ocean climate research,
Eos Trans. AGU 86(9), 2005, pp. 89,95–96.
18. S.M. Murman, M.J. Aftosmis, and S.E. Rogers, Characterization of Space Shut-
tle ascent debris aerodynamics using CFD methods, 43rd AIAA Aerospace Sci-
ences Mtg., AIAA Paper 2005-1223 (Reno, NV, 2005).
19. NASA’s Role in Hurricane Research, http://www.nasa.gov/pdf/147432main−
hurr− fact− sheet.pdf.
20. B.-W. Shen, R. Atlas, J.-D. Chern, O. Reale, S.-J. Lin, T. Lee, and J. Chang,
The 0.125 degree finite-volume general circulation model on the NASA Columbia
supercomputer: Preliminary simulations of mesoscale vortices, Geophys. Res.
Lett. 33(5), 2006, L05801, doi:10.1029/2005GL024594.
21. B.-W. Shen, R. Atlas, O. Reale, S.-J. Lin, J.-D. Chern, J. Chang, C. Henze, and
J.-L. Li, Hurricane forecasts with a global mesoscale-resolving model: Prelimi-
nary results with Hurricane Katrina (2005), Geophys. Res. Lett. 33(13), 2006,
L13813, doi:10.1029/2006GL026143.
22. STREAM Benchmark, http://www.streambench.org/.
23. J.R. Taft, Achieving 60 Gflop/s on the production CFD code Overflow-MLP,
Parallel Computing 27(4), 2001, pp. 521–536.
24. Top500 Supercomputer Sites, http://www.top500.org/.
Recent Advances of Multi-phase Flow
Computation with the Adaptive Soroban-grid
Cubic Interpolated Propagation (CIP) Method
Takashi Yabe1 , Youichi Ogata2 and Kenji Takizawa3

1
Tokyo Institute of Technology yabe@mech.titech.ac.jp
2
Tokyo Institute of Technology yogata@mech.titech.ac.jp
3
National Maritime Research Institute ktaki@nmri.go.jp
1 Introduction
Recent developments of simulation technology have enabled the combined

analysis of solid, liquid and gas. However, it is still challenging to establish
simultaneous treatment of materials undergoing phase state transition. This
is because the conventional combined analysis relies on the separate treatment
of each phase.
The CIP(Cubic Interpolated Propagation) method[2, 4, 5, 11] that gives
quite accurate and less-diffusive results is known as a powerful multi-phase
flow solver. It is a kind of semi-Lagrangian scheme and has the third-order ac-
curacy in time and space, and has been extended to treat incompressible flow
in the framework of compressible fluid. The CIP method uses the primitive
Euler method to solve all the fluid-like equations with a separate treatment
of advection part and non-advection part as in the usual semi-Lagrangian so-
lution. Therefore, the formulation into a simultaneous solution of incompress-
ible and compressible fluid is easily obtained. The advection part is solved
by the CIP method and non-advection part is calculated by the CIP-CUP
method[7] which solves the Poisson equation for pressure to treat compress-
ible and incompressible flows together. Although this method is similar to the
ICE(Implicit Continuous Eulerian)[1] method, the CIP-CUP algorithm works
robustly even with a density ratio larger than 1000, while the ICE is limited
to weak discontinuity because the continuity of ∇p/ρ is not guaranteed in the
ICE . The CIP-CUP algorithm simplifies the pressure equation, gives very
stable solutions and greatly improves the capability to treat multi-phase flow.
Although the CIP method had the drawback that the exact mass conserva-
tion could not be maintained because the CIP method used non-conservative
form, it had been overcome by the new version of the CIP method, CIP-CSL
method that guarantees the exact mass conservation even in the framework
of a semi-Lagrangian scheme[8, 9].

30 Takashi Yabe, Youichi Ogata and Kenji Takizawa
We have succeeded in multi-phase flow and fluid-structure interaction sim-

ulations even in the Cartesian system, for example, laser processing, milk-
crown, slamming of a vessel and so on. One of interesting applications is
skimmer[10]. It is the fluid-structure interaction and requires accurate com-
putation of pressure on the surface of the moving body. Using the accurate
profile of pressure represented by cubic polynomial inside a grid cell, the force
acting on the rigid body can be accurately estimated even if the rigid body has
a structure in a scale smaller than grid cell. The CIP method is able to define
such subgrid-scale structure, and the modified C-CUP method that includes
the constraint from the rigid body using density function makes it possible
to accurately solve fluid-structure interaction, and thus simulation accurately
replicated the experiments.
This paper reviews such recent progress of multi-phase flow computation
with the adaptive Soroban-grid CIP method. In Section 2, the CIP method
is reviewed and the new grid system ”Soroban grid” is brought forward. The
formulation of the CIP method applied to general hydrodynamics is presented
in Section 3. Section 3 discusses a unified procedure C-CUP method with
Soroban grid to compute both compressible and incompressible fluids, and
several applications solved by the method will be presented. The paper ends
with a short summery in Section 3.
2 CIP Method and Soroban Grid
2.1 CIP Method
For simplicity, CIP formulation in one-dimension is reviewed. We here explain

the strategy of the CIP method by using an advection equation as Eq.(1).
∂f ∂f
+u =0 (1)
∂t ∂x
The CIP method uses value f as well as spatial derivative ∂f /∂x (≡ g) to
make up cubic-interpolation between two grid points. That is why the basic
equation of g is also needed. If we differentiate Eq.(1) with spatial variable x,
we get
∂g ∂g ∂u
+u =− g (2)
∂t ∂x ∂x
In the simplest case where the velocity u is constant, Eq.(2) coincides with
Eq.(1) because of ∂u/∂x = 0. Using this equation, we can trace the time
evolution of f and g on the basis of Eqs.(1) and (2).
If two values of f and g are given at two grid points, the profile between
these points can be interpolated by the cubic polynomial F (x) = ax3 + bx2 +
cx + d. Thus, the profile at the n + 1 step can be obtained by shifting the
profile by u∆t so that
Multi-Phase Flow Computation with Soroban grid CIP Method 31
fin+1 = F (xi − ui ∆t) = ai ξ 3 + bi ξ 2 + gin ξ + fin

gin+1 = dF (xi − ui ∆t) /dx = 3ai ξ 2 + 2bi ξ + gin
gi + giup 2 (fi − fiup )
ai = + (3)
D2 D3
3 (fiup − fi ) 2gi + giup
bi = −
D2 D
where we define ξ ≡ −ui ∆t. Here, D = −∆x, iup = i − 1 for u > 0 and
D = ∆x, iup = i + 1 for u < 0. The case that velocity u is not constant,
in other words, the nonlinear case like hydrodynamics will be explained in
Section 3.
For constant velocity, the scheme becomes exactly conservative. Even in
variable velocity field, we can obtain the exactly conservative scheme[8, 9].
Interestingly, the CIP is proven to be the third-order in time and space easily
by Taylor expanding[11, 6] even in non-uniform grid system in one-dimension.
However, the application of the method to multi-dimensions is not possible
in general. This is the reason why we used the coordinate-transformation
in the previous paper[12]. However, it would be useful to consider a special
mesh system that keeps the higher-order accuracy. Next, the new grid system
”Soroban grid ” that can realize it is proposed.
2.2 Soroban Grid System
In some cases, the use of curvilinear system is sometimes inevitable to repre-

sent the detailed flow structure like boundary layer around complex bodies.
When the CIP scheme is applied to curvilinear system, however, the third-
order accuracy that was achieved in uniform mesh is severely deteriorated and
is degraded to the first-order in the deformed mesh[6]. Therefore the degra-
dation of accuracy by introducing curvilinear coordinate would cancel the
advantage of the CIP method that has originally the third order accuracy in
time and space.
In order to resolve such difficulty, the CIP method has recently been up-
graded to include the adaptive grid that guarantees the high-order accuracy
as well as robustness. The new grid is named Soroban grid[6].
The schematic of the Soroban grid is shown in Fig.1. The grid system con-
sists of the straight lines and grid points moving along these lines like abacus-
Soroban in Japanese. Planes also move in parallel in three-dimension. The
length of each line and the number of grid points in each line can be differ-
ent. An example to locate Soroban grid around a circle is shown in Fig.2. For
convenience, Finite-element-like connectivity is shown just for visualization of
Soroban grid. It can be seen that Soroban grid concentrated on a boundary
makes it possible to almost completely represent a circle.
In order to understand how to solve the advection equations with the
Soroban grid, let us consider a part of the Soroban grid shown in Fig.1(Right)
on a two dimensional plane for simplicity, where the vertical mesh (y-direction)
is the straight line, while the grid points move along each line. If the upstream
departure point T is given as (ξ, η)=(xi − u∆t, yj − v∆t), at first one pair of
lines satisfying xi1 < ξ < xi1+1 is searched. We should remind that xi can be
far from xi1 and xi1+1 so that large CFL(u∆t/∆x, v∆t/∆y) computation is
possible.
i1 i1+1
j1+1 j2+1
T( , )
T
A(i1, )
x j1
A(i1+1, )
x j2
x
Fig. 1. (Left) Soroban grid arrangement. (Right) Close view on a two-dimensional
plane.
Fig. 2. Soroban grid around a circle. Finite element mesh is used just for visualiza-
tion of Soroban grid
Next, two pairs of points satisfying yj1 < η < yj1+1 and yj2 < η < yj2+1
are searched along two lines at x = xi1 and x = xi1+1 , respectively. Therefore,
the interpolation can be briefly described as follows. At first, one-dimensional

CIP is applied to the vertical straight line(y-direction) giving Ai1,η and
Ai1+1,η , and then T is given by one-dimensional CIP along the straight line
connecting Ai1,η and Ai1+1,η in the x-direction. We have to pay attention to
spatial derivatives ∂f /∂x (≡ gx ) and ∂f /∂y (≡ gy ). As for y-direction, since f
and gy can be transported along lines, A and only ∂A/∂y (= ∂y A) are readily
obtained by CIP. However, we need some method to estimate ∂A/∂x (= ∂x A)
from gx to make up cubic interpolation function along the straight line con-
necting Ai1,η and Ai1+1,η in the x-direction.
In the previous paper[3], we proposed to use the first-order scheme because
the derivative in the direction perpendicular to the propagating direction is
not sensitive and hence can be estimated only roughly by the linear interpo-
lation. Such a splitting scheme was at first proposed in Ref.[3] which we call
the ”Type-M” scheme.
Although the Type-M scheme is sufficient for many applications, a little
more accurate scheme is possible at the price of memory requirement. This
scheme was proposed by Aoki[13] and we call it ”Type-C” scheme. In this
scheme, independent variables are f, gx , gy and ∂xy f (= ∂x gy = ∂y gx ) in two
dimensions. For example, instead of using linear interpolation for gx in y-
direction, the one-dimensional CIP scheme is applied to the advection of (∂x f )
and ∂y (∂x f ). The same can be said of x-direction, that is, the one-dimensional
CIP scheme is applied to the advection of (∂y f ) and ∂x (∂y f ).
Since this scheme uses only the one-dimensional CIP method without co-
ordinate transformation, it is able to keep the third order accuracy in time
and space even in such a deformed mesh[6].
The application of the Soroban grid to non-advection phase that will in-
clude finite difference is also straight forward, and the Poisson equation for
pressure in the C-CUP method can be also solved with the Soroban grid.
Therefore, we can solve multi-phase flow by the Soroban-grid CIP method
like Cartesian grid system.
2.3 Moving Adaptive Soroban Grid
Here, we shall discuss the possibility to use the Soroban grid as for the adaptive
grid to the moving body. We must remind that all the discussion can be
straightforwardly applied to three dimensions. At first, let us consider one-
dimensional case. For the simplest choice of the monitoring function to the
variation, we use the following quantity:
∂f 2 12 ∂2f
M (x, t) ≡ 1 + α +β (4)
∂x ∂x2
where two parameters α and β can be chosen depending on problems we treat.
Therefore monitoring function M becomes large for larger gradient region.
Since the Soroban grid is straight in one-direction, it is much easier to generate
the adaptive grid points along the line. The reorganization of the grid point
Fig. 3. Accumulated monitoring function I (x) is divided into equal pieces. The x
boundary of each piece gives the grid point.
move by line
move by point
Fig. 4. The straight lines can move in horizontal direction and grid points move
along the straight lines.
is easily performed by accumulated monitoring function I (x) as shown in

Fig.3. If we divide the accumulated function into equal pieces and get the
x coordinate to get such integration. We can see that the size of grid ∆x
becomes small where monitoring function M is large. Therefore, grid points
are concentrated on where spatial gradient is large like boundary, and this
procedure is repeated in every time step.
In the two-dimensional mesh shown in Fig.4, mesh moving is performed
as follows.
1. Calculate M (y, t) along each line
2. Generate the points along each line.
3. Calculate the average M (x, t) from all the points along each line.
4. Move the lines.
Three dimension is also straightforward, that is, the monitoring function
M (z, t) along each plane is calculated, and ”Move the planes” is also added.
The possibility to use the Soroban grid as for the adaptive grid to the
moving body has already been discussed in Ref.[6]. It is based upon the mon-
itoring function to the variation, and local mesh refinement can be achieved.
The drawback was that the Soroban grid thus generated would not be smooth
because the grid size would be automatically determined by the monitoring
function and thus the ratio of grid size would become extremely large. One way
to avoid such drawback is to smooth the spatial derivatives in the monitoring
function, or to use distance function to a body surface.
The CIP method is suitable for this mesh system because it uses only two
stencils to make up cubic-interpolation function, and the calculation of large
CFL(> 10) at locally refined mesh is easily performed. Mesh generation and
searching of upstream departure point are very simple and almost mesh-free
treatment is possible.
3 A Semi-Lagrangian Approach to Hydrodynamics

Equations
3.1 Basic Equations
We use the following set of hydrodynamics equations;

∂f
+ (u · ∇) f = S (5)
∂t
Here, f = (ρ, u, T ), S = (−ρ∆ · u + Qm , −∇p/ρ + Qu , −PT H ∇ · u/ρCv + QE ),
where ρ is the density, u the velocity, p the pressure, and T the temperature,
Qm represents the mass source term, Qu represents viscosity(and artificial
viscosity), elastic stress tensor, surface tension, etc., and QE represents vis-
cous heating, thermal conduction, and heat source. Cv is the specific heat for
constant volume and we define PT H ≡ T (∂p/∂T )ρ , which is derived from the
firat principle of thermodynamics and the Helmholtz free energy.
3.2 The Fractional Step Approach
The underlying physics included in the above equations for continuum dy-
namics is complex and may include processes that have different time scales
of variation. It is expedient to separate the solution procedure into several frac-
tional steps. For example, Eq.(5) can be separated into the advection phase,
∂f n ∗
+ (u · ∇) f = 0 (f , ∂χ f ) → (f , ∂χ f ) (6)
∂t
and the non-advection phase.
∂f
f ∗ → f n+1

=S (7)
∂t
The value and the spatial gradient in the advection phase are solved by the
CIP method. The time evolution of the spatial gradient in the non-advection
phase ∂χ f ∗ → ∂χ f n+1 is calculated according to the spatial derivative of

Eq.(7)
∂ (∂χ f )
= ∂χ S (8)
∂t
Usually, it is not easy to get a finite-difference form of ∂χ S. However, using
Eq.(7), this term can be estimated as
n+1 ∗ n+1 ∗
(∂χ f ) − (∂χ f ) (δχ f ) − (δχ f )
= (9)
∆t ∆t
where δχ f represents a centered finite-difference form of ∂χ f in the χ direction.
Thus the time evolution of ∂χ f is estimated by the time evolution of f already
given by Eq.(7).
The solution of Eq.(7) can be obtained by a simple finite difference like
f n+1 = f ∗ + S∆t. However, we have to use a more sophisticated method in
order to deal with compressible and incompressible fluids, or multi-phase flow
simulutaneously. Next section, the method and its application to Soroban grid
will be discussed.
4 A Pressure Based Algorithm in a Primitive Euler

Scheme
4.1 CCUP Method
Yabe and Wang[7] adopted the primitive Euler form instead of the conserva-
tive form to construct the pressure equation.
The original CCUP method[7] was proposed only for a special equation of
state, but here we use

∂p ∂p
∆p = ∆ρ + ∆T (10)
∂ρ T ∂T ρ
where ∆p means the pressure change pn+1 − p∗ during one time step and * is
the profile after advection. From this relation, once ∆ρ and ∆T are predicted,
∆p can be predicted based on Eq.(10); ∂p/∂ρ, ∂p/∂T are given by the EOS.
Since the non-advection terms are separated from the advection terms by
the CIP, we can concentrate on the non-advection terms related to sound
waves, which are the primary cause of the difficulty posed by the large sound
speed of liquid, and hence ρ, T are simply given by
∆ρ = −ρ∗ ∇ · un+1 ∆t ρ∗ Cv ∆T = −PT H ∇ · un+1 ∆t (11)
un+1 in this equation is given by an equation of motion as
∇pn+1
∆u = − ∆t (12)
ρ∗
Since ∆u = un+1 − u∗ , Eqs.(10)-(12) lead to a pressure equation[7, 11]

pn+1 − p∗ ∇ · u∗

1 n+1
∇· ∗
∇p = 2
+ (13)
ρ P
∆t2 ρ∗ C 2 + T H ∆t
s ρCv T
n+1
Then substituting the given p into Eq.(12), we obtain the velocity un+1
n+1
and then the density ρ from Eq.(11). From this pressure, density can be
solved in terms of pressure. Equation (13) has many important features. This
equation shows that, at sharp discontinuities, n · (∇p/ρ) is continuous. Since
∇p/ρ is the acceleration, it is essential that this term be continuous since the
density changes by several orders of magnitude at the boundary between liquid
and gas. In this case, the denominator of ∇p/ρ changes by several orders,
and the pressure gradient must be calculated accurately enough to ensure
continuous change of acceleration. It has been proved in many applications of
Eq.(13) that it works robustly even with a density ratio larger than 1000 and
enables us to treat both compressible and incompressible fluids.
xin 1 xin
N
y j +1
dS n
yj
xi 1 dS w xi dS e xi +1
dV
dS s
y j 1
xis 1 S xis
Fig. 5. C-CUP method in Soroban grid
When Eq.(13) is solved with Soroban grid, the adequate formulation can
be derived by the integration of Eq.(13) in the shaded square (≡ Ω) in Fig.5
as
pn+1 − p∗ ∇ · u∗
Z Z Z
1 n+1
∇· ∗
∇p dV = 2
dV + dV (14)
Ω ρ P
Ω ∆t2 ρ∗ C 2 + T H Ω ∆t
s ρCv T
According to the Gauss’s theorem, Eq.(14) can be changed into

pn+1 − p∗
Z Z Z
1 1
∗
∇p n+1
dS =
PT2 H
dV + u∗ dS (15)
∂Ω ρ Ω ∆t 2 ∗ 2
ρ C + ∆t ∂Ω
s ρCv T
where dV = (xi+1 − xi−1 ) (yi+1 − yi−1 ) /4, dSe = −dSw = (yi+1 − yi−1 ) /2,
dSn = −dSs = (xi+1 − xi−1 ) /2, and ∂Ω means the integration along the cell
boundary in two dimension. It is not to mention that three dimension is also
straightforward using Eq.(15)
Although density at points (N, S), and velocity at cell boundaries (n, s, w, e)
in Fig.5 are needed in the formulation of Eq.(15) with Soroban grid, they can
be interpolated by the CIP method. The methods to solve this poisson equa-
tion for pressure have been established like SOR, Bi-CGSTAB and so on, thus,
CCUP method can be used even with Soroban grid as well[14].
4.2 The Applications of CCUP Method with Soroban Grid
Karmann Vortex
As one of the most famous examples, Soroban grid is applied to two-

dimensional Karmann vortex. The Reynolds number Re ≡ U d/ν is 100, where
d is the diameter of a cylinder(=1cm), and ν is the water coefficient of kine-
matic viscosity. A cylinder moves from right to left with constant velocity.
Figure 6 shows the time evolution of adaptive Soroban grid (left col-
umn) and vorticity (right column) generated by a moving cylinder at t =
10, 20, 40, 190 second. Since this assumes incompressible flow, the absolute of
vorticity |η| = |∂v/∂x − ∂u/∂y| is selected as monitor function. Therefore, it
can be seen that Soroban grid are concentrated on vortices.
In addition, the number of grid points can be changed at each time step
depending on the situation of phenomena as shown in Fig.6. The Soroban grid
follows the movement of vortex. In this system, the grid does not move but
always new grid points are created at each time step because such additional
computation is negligibly small. We do not have to prepare more memory
than needed, and it leads to faster calculations.
Skimmer
Figure 7 shows another application to the skimmer in three dimensions. Each

small ball corresponds to each Soroban grid, and ball size is proportional to
density. The grid is adapted to the moving solid disk and water surface[15, 16].
Fig. 6. Adaptive Soroban grid in vortex shedding from a cylinder at t =

10, 20, 40, 190 second. Number of grid points increases at first then stays constant
when the number of vortices in the computational domain becomes constant.
Container
The last application is the interaction between a ship (a container) and stokes
waves. The length between profile of a container is set to be 283.8[m]. The
wavelenghth λ is also 283.8[m], which is the same as a container length, and
wave amplitude is 8.0[m]. The calculation domain is 800[m]×400[m]×200[m],
and 10 [m/sec] fixed surge motion is set. The number of Soroban grid points
is changed at each time step from 100, 000 to 250, 000.
Figure 8 shows the motion of a container by stoke waves, and Figure 9
shows the bottom of a container with Soroban grid.
Fig. 7. Grid points used in the skimmer simulation. Vertical plane shows the grid ar-
rangement. For visualization, grid points in three dimensions are depicted by spheres
whose size depends on the density.
5 Conclusion
We have reviewed various families of the CIP method to solve the problem
of multi-phase flow, and proposed recent advance of multi-phase flow com-
putation with Soroban grid CIP method. Although the conventional C-CUP
method in Cartesian grid has already been established and applied to a lot of
multi-phase flow simulations like laser-processing, skimmer, milkcrown and so
on that include solid, liquid and gas as well as phase transition, Soroban grid
CIP method is able to open new avenue for significantly improved scheme to
solve multi-phase flow simulations. We confirmed that the CIP can realize the
third-order accuracy both in time and space even in such arbitrary mesh. The
merit of the Soroban grid can be summerized as
1. It is easy to generate, and load to generate grid is very low.
2. Searching the upstream point is easy and can be done by one step by
refering to the index.
3. The original high accuracy of CIP can be kept.
4. CFL can be large.
Soroban grid is being applied to other fields like plasma simulation
with Vlasov-Boltzmann equations and the propagation of electromagnetic
wave with Maxwell’s equations[17] and so on, and more applications to the
hyperbolic-type equations will be discussed in future papers.
Fig. 8. The motion of a container interacting with stoke waves

Fig. 9. The bottom of a container with Soroban grid
References
1. Harlow,F.H. and Amsden,A.A: Numerical Simulation of Almost Incompressible
Flow, J.Comput.Phys., 3, 80–93 (1968)
2. Takewaki, H., Nishiguchi,A. and Yabe, T: The Cubic-Interpolated
Pseudo-Particle (CIP) Method for Solving Hyperbolic-Type Equations,
J.Comput.Phys., 61, 261–268 (1985)
3. Takewaki, H. and Yabe, T: Cubic-Interpolated Pseudo Particle (CIP) Method
Application to Nonlinear or Multi-Dimensional Problems, J.Comput.Phys., 70,
355–372(1987)
4. Yabe, T., and Aoki, T.: A Universal Solver for Hyperbolic Equations by Cubic-
Polynomial Interpolation, Comput. Phys. Commun., 66, 219–232 (1991)
5. Yabe, T., Ishikawa, T., Wang, P. Y., Aoki, T., Kadota, Y., and Ikeda, F.: A
Universal Solver for Hyperbolic Equations by Cubic-Polynomial Interpolation
II. Two- and Three- Dimensional Solvers, Comput. Phys. Commun., 66, 233–
242 (1991)
6. Yabe, T., Mizoe, H., Takizawa, K., Moriki, H., Im, H. N., and Ogata, Y.:
Higher-Order Schemes with CIP Method and Adaptive Soroban Grid Towards
Mesh-Free Scheme, J.Comput.Phys., 194, 57–77 (2004)
7. Yabe, T., and Wang, P. Y.: Unified Numerical Procedure for Compressible and
Incompressible Fluid, J.Phys.Soc.Japan., 60, 2105–2108 (1991)
8. Yabe, T., Tanaka, R., Nakamura, T., and Xiao, F.: An Exactly Conservative
Semi-Lagrangian Scheme (CIP-CSL) in One Dimension, Mon.Wea.Rev., 129,
332–344 (2001)
9. Nakamura, T., Tanaka, R., Yabe, T., and Takizawa, K.: Exactly Conservative
Semi-Lagrangian Scheme for Multi-Dimensional Hyperbolic Equations with Di-
rectional Splitting Technique, J.Comput.Phys., 174, 171–207 (2001)
10. Takizawa, K., Yabe, T., Chino, M., Kawai, T, Wataji, K, Hoshino, H., and
Watanabe, T.: Simulation and Experiment on Swimming Fish and Skimmer by
CIP Method, Computers & Structures, 83, 397–408 (2005)
11. Yabe, T., Xiao, F., and Utsumi, T.: Constrained Interpolation Profile Method
for Multiphase Analysis, J.Comput.Phys., 169, 556–593(2001)
12. Wang, P. Y., T, Yabe, T, Aoki.: A General Hyperbolic Solver - the CIP Method
- Applied to Curvilinear Coordinate, J.Phys.Soc.Jpn., 62, 1865–1871(1993)
13. Aoki, T.: Multi-dimensional advection of CIP(cubic-interpolate propagation),
CFD.J, 4, 279–291(1995)
14. Takizawa, K., Yabe, T., Tsugawa, Y., Tezduyar, T.E., Mizoe, H. : Computation
of free-surface flows and fluid-object interactions with the CIP method based
on adaptive meshless soroban grids , Comput.Mech, in print(2006)
15. Yabe, T., Ogata, Y., Takizawa, K., Kawai, T., Segawa, A., Sakurai, K. : The
next generation CIP as a conservative semi-Lagrangian solver for solid, liquid
and gas, J.Comput.Appl.Math. 149, 267–277 (2002)
16. Yabe, T., Takizawa, K., Chino, M., Imai, M., Chu, C.C.,: Challenge of CIP as
a Universal Solver for Solid, Liquid and Gas, Int. J. Numer. Meth. Fluids. 47,
655–676 (2005)
17. Ogata, Y.,Yabe, T.,Odagaki, K.: An Accurate Numerical Scheme for Maxwell
Equation with CIP-Method of Characteristics, Comm.Comp.Phys, 1, 308–
331(2006)
Part II
Schemes
On the Computation of Steady-State
Compressible Flows Using a DG Method
Hong Luo1 , Joseph D. Baum1 and Rainald Löhner2

1
SAIC, 1710 SAIC Dr., MS 2-6-9, McLean, VA 22102, USA, hong.luo@saic.com
2
George Mason University, Fairfax VA 22030, USA, rlohner@gmu.edu
1 Introduction
Most efforts in the development of the discontinuous Galerkin methods

(DGM) in computational fluid dynamics are primarily focused on the time
accurate compressible Euler and Navier-Stokes equations. Its accuracy, ef-
ficiency, capability, robustness for steady state flow problems are relatively
unexplored. In order for DGM to become a viable, attractive, probably even
better alternative to the more traditional, more elaborate, well established
finite volume methods (FVM), and finite element methods (FEM) for steady
state computations, the following three issues have to be addressed: 1) Lack
of efficient flow solver for steady state computations: Most efforts in the de-
velopment of the discontinuous Galerkin methods are primarily focused on
the spatial discretization. The temporal discretization methods have lagged
far behind. Usually, explicit temporal discretizations such as multi-stage TVD
(Total Variation Diminishing) Runge-Kutta schemes are used to advance the
solution in time. For large-scale simulations and especially for high-order solu-
tions, the rate of convergence slows down dramatically, resulting in inefficient
solution techniques to steady state solutions. 2) Implementation of boundary
conditions for curved solid walls: It is widely recognized and accepted that
DGM solutions are more sensitive to the error arising at curved boundaries
than those obtained with FVM of the same order of accuracy. The common
remedy to this problem is to use higher-order geometrical approximation. Un-
fortunately, curved element meshes are associated with extra computational
costs. 3) Flux and slope limiters: It is well known that the nonphysical os-
cillations exist in the vicinity of discontinuities. Classical techniques of flux
limiting in FVM are not directly applicable for high order DGFEM because of
the presence of volume terms in the formulation. Therefore, the slope limiter
is not integrated in the computation of the residual, but effectively acts as a
post-processing filter. Such a filter is easily integrated in an explicit method,
but not into an implicit method.

48 Hong Luo, Joseph D. Baum and Rainald Löhner
Computation of compressible steady state flows using a high-order discon-

tinuous Galerkin finite element method is presented in this paper. An accurate
representation of the boundary normal based on the definition of the geome-
tries is used for imposing solid wall boundary conditions for curved geometries.
Particular attention is given to the impact and importance of slope limiters
on the solution accuracy. A recently developed, fast, low-storage p-multigrid
(p=polynomial degree) is used for obtaining steady state solutions to the gov-
erning compressible Euler equations. The method is applied to compute a
variety of compressible flow problems for a wide range of flow conditions in
both 2D and 3D configurations. The numerical results indicate that with the
progress made in the development of the DGM, the DG method provides a
viable, attractive, competitive and probably even better alternative to the
finite volume method for computing compressible flows.
2 Numerical Method
2.1 p-Multigrid Discontinuous Galerkin Method
The governing compressible Euler equation is discretized using the standard

DG method on computational domain of triangles in 2D and tetrahedra in 3D.
The interface flux function is computed using the HLLC approximate Riemann
solver. The time integration is performed using a p-multigrid method [1]. A
distinct feature of this p-multigrid method is to use an explicit smoother on
the higher level approximations (p > 0) and an implicit smoother on the
lowest level approximation (p = 0), resulting in a fast, low storage method
that can be efficiently used to accelerate the convergence to a steady state
solution. Furthermore, this p-multigrid method can be naturally applied to
compute the flows with discontinuities, where a monotonic limiting procedure
is usually required for discontinuous Galerkin methods.
2.2 Curved Wall Boundary Conditions
It is recognized that DG methods are far more sensitive to errors arising

at curved boundaries than those obtained with FVM of the same order of
accuracy. Bassi et al. [2] have shown that an accurate boundary representa-
tion is necessary to maintain the formal order of the DG methods and to
avoid spurious production of entropy on the boundary. A common solution to
this problem is to use higher-order geometrical approximation. Unfortunately,
curved element meshes are associated with extra computational expenses. In a
novel approach, suggested by Krivodonova et al. [3], the elements adjacent to
the solid wall boundaries remain straight-sided elements. However, an accu-
rate representation of the boundary normals is used to define a ghost state at
quadrature points. Once the ghost state values are determined at integration
DG Computation of Compressible Flows 49
points, the numerical fluxes are computed exactly as usual straight-sided ele-
ments. Using only straight-sided elements instead of boundary fitted elements
represents a huge simplification of the code implementation and tremendous
saving in both storage requirements and computing costs. In our implementa-
tion, the normals at the boundary integration points are computed using the
local true surface normal based on the analytically defined boundary geome-
tries, which are handily available in our geometry definition file.
2.3 Monotonicity Limiter
It has been well known that just like DGM are more sensitive to the treatment
and implementation of slip boundary conditions at curved boundaries than
those obtained with FVM of the same order of accuracy, DGM are also much
more sensitive to the treatment and implementation of the slope limiters than
their FV counterparts. Slope limiters frequently identify regions near smooth
extrema as requiring limiting, and this typically results in a reduction of the
optimal high-order convergence rate and degenerates solution accuracy. To ad-
dress this concern, the limiters are applied only where they are really needed.
This is accomplished using the so-called discontinuity detector [4], which is
helpful to distinguish regions where solutions are smooth and discontinuous.
Then, the limiting is only used near discontinuities and high-order accuracy
can be preserved in smooth regions. Unfortunately, almost all discontinuity
detectors, will mistakenly identify local high gradient smooth extrema such
as leading edge regions as shock regions due to the high comparable gradient
there. For aerodynamic applications, the active limiters at the leading edge of
an airfoil will pollute the solution in the flow field and ultimately destroy the
high order accuracy solution. To address this concern, a physics-based shock
detector [5] is introduced by taking into consideration the two facts that the
normal flow before a shock is supersonic and normal flow after the shock is
subsonic and a shock wave can only be a compressive wave. Our experience in-
dicates that this shock detector can effectively eliminate high gradient smooth
regions, and performs better than alternative indicators for the steady-state
flow problems considered here.
3 Computational Results
A. Subsonic Flow past a sphere
The first example is a subsonic flow past a sphere at a Mach number of
M∞ =0.5. This test case is chosen to verify the implementation of boundary
conditions for curved geometries for the DG method, and assess the order
of accuracy of the discontinuous Galerkin method and performance of the
p-multigrid. The computational experiments have been performed on three
successively refined grids in order to obtain quantitative measurement of the
order of accuracy for different order DG methods, Fig.1 shows the medium
-1.5 -1.5 0 0
-0.5 -0.5
-2 -2
-1 -1
-2.5 -2.5 -1.5 -1.5
Log(Error-L2)
-2 -2
-3 -3
Log(resi)
-2.5 -2.5
-3.5 -3.5
-3 -3
-4 -4 -3.5 -3.5
-4 -4
-4.5 DG(P0) (Slope=0.67) -4.5
DG(P1) (Slope=1.82) -4.5 TVDRK(P2) -4.5
DG(P2) (Slope=2.71) p-Multigrid(P2)
-5 -5 -5 -5
-0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0 100 200 300 400 500 600 700 800 900 1000
Log(cell-size) CPU (Second)
Fig. 1. The medium unstructured surface meshes used for computing subsonic flow
past a sphere at M∞ = 0.5, computed Mach number contours obtained by DG(P2)
on the medium mesh, accuracy summary for DG(P0), DG(P1) and DG(P2) methods,
and comparison of convergence history versus CPU time between TVBRK and p-
multigrid methods.
mesh used in the computation, the computed Mach number contours in the
flow field obtained by DG(P2) solution on the medium mesh, the spatial accu-
racy details of each DG method for this numerical experiment, and the com-
parison of convergence histories versus CPU time for the DG(P2) solution on
the coarse mesh between 3-stage TVBRK method and p-multigrid method.
The numerical results indicate that the discontinuous Galerkin method ex-
hibits a O(hp+1 ) order of convergence on smooth solutions, and show that
our simplified implementation of boundary conditions for curved geometries
well conserves the formal order of the DG method. The p-multigrid method
is orders of magnitude faster than its explicit TVDRK counterpart.
B. Transonic Flow past a NACA0012 airfoil

The second example is transonic flow past a NACA0012 airfoil at a Mach
number of 0.80, and an angle of attack 1.25◦ . This test case is chosen pri-
marily to demonstrate the impact and importance of limiting on the accuracy
of the DG method for flows with shock waves. Fig. 2 shows the mesh used
in the computation, the computed Mach number contours in the flow field
DG Computation of Compressible Flows 51
1.4 1.4 0.04 0.04

0.035 0.035
1.2 1.2
0.03 0.03
1 1 0.025 0.025
Entropy Production
0.02 0.02
Mach Number
0.8 0.8
0.015 0.015
0.6 0.6 0.01 0.01

0.005 0.005
0.4 0.4 0 0
-0.005 -0.005
0.2 Limiter everywhere 0.2 Limiter everywhere
Limiter with shock detection -0.01 Limiter with shock detection -0.01
Limiter only shock regions Limiter only shock regions
0 0 -0.015 -0.015
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
X-coordinates X-coordinates
Fig. 2. Unstructured triangular mesh (nelem=1,999, npoin=1,048, nboun=97) used

for computing transonic flow past a NACA0012 airfoil, computed Mach number
contours in the flow field by DG(2) with Barth-Jespersen limiter everywhere, in
the region identified by Krivodonova’s shock detector, in the region identified by
the physics-based shock detector, respectively, and comparison of computed Mach
number and entropy production distributions on the surface of airfoil among these
three computations.
with Barth-Jespersen limiter applied everywhere, in the regions identified by

Krivodonova’s shock detector, and in the region identified by our physics-
based shock detector, respectively, and the comparison of the Mach number
and entropy production on the surface of the airfoil among these three com-
putations for the DG(P2) solution. As expected, application of the limiter

everywhere completely destroys high order accuracy offered by the DG(P2)
method, as any limiters tend to degenerate accuracy when used in the smooth
regions of the solution. When the limiter is only applied to the regions iden-
tified by Krivodonova’s shock detector, the solution improves dramatically
in the smooth regions. However, this shock detector mistakenly identifies the
leading edge region as shock regions, and consequently contaminates the flow
solution in the smooth region. Our physics-based shock detector evidently
does not identify the leading edge region as discontinuities, which in turn
enables DG(P2) to produce an accurate solution on a fairly coarse mesh.
4 Conclusions
Computation of compressible steady state flows using a high-order discontin-
uous Galerkin finite element method is presented in this paper. All three areas
required in the practical application of DG methods: efficient flow solver, im-
plementation of boundary conditions for curved geometries, and limiting for
discontinuities have been discussed. The developed DG method provides a vi-
able, attractive, and competitive alternative to the finite volume method for
CFD applications. Future work will explore application of this method for the
solution of the Navier-Stokes equations.
References
1. Luo, H., Baum, J.D., Löhner, R.: A p-multigrid Discontinuous Galerkin Method
for the Euler Equations on Unstructured Grids. Journal of Computational
Physics, Vol. 211, No. 2, pp. 767-783 (2006)
2. Bassi, F. and Rebay, S.: High-Order Accurate Discontinuous Finite Element
Solution of the 2D Euler Equations. Journal of Computational Physics, Vol.
138, pp. 251-285 (1997)
3. Krivodonova, L. and Berger, M.: High-order Implementation of Solid Wall
Boundary Conditions in Curved Geometries. Journal of Computational Physics,
Vol. 211, No. 2, pp. 492-512 (2006)
4. Krivodonova, L., Xin, J., Remacle, J.F., Chevaugeon, N., and Flaherty, J.E.:
Shock Detection and Limiting with Discontinuous Galerkin Methods for Hyper-
bolic Conservation Laws. Applied Numerical Mathematics, Vol. 48, pp. 323-338
(2004)
5. Luo, H., Baum, J.D., Löhner, R.: A physics-based shock detector for Discontin-
uous Galerkin Method. In preparation (2006)
Space-Time Discontinuous Galerkin Method
for Large Amplitude Nonlinear Water Waves
Yan Xu and Jaap J.W. van der Vegt
Department of Applied Mathematics, University of Twente, P.O. Box 217, 7500

AE, Enschede, The Netherlands. {y.xu,j.j.w.vandervegt}@math.utwente.nl
Summary.
A space-time discontinuous Galerkin (DG) finite element method for nonlinear
water waves in an inviscid and incompressible fluid is presented. The space-time DG
method results in a conservative numerical discretization on time dependent deform-
ing meshes which follow the free surface evolution. The dispersion and dissipation
errors of the scheme are investigated and the algorithm is demonstrated with the
simulation of waves generated by a wave maker.
Key words: space-time discontinuous Galerkin method, nonlinear water waves
1 Introduction
Large amplitude nonlinear water waves are frequently modeled by considering
the fluid as inviscid, incompressible and irrotational. This makes it possible to
describe the flow field with a potential function φ, which satisfies the Laplace
equation. In addition, the potential function must satisfy a kinematic and
dynamic boundary condition at the free surface:
∂ζ ∂φ
+ ∇ s φ · ∇s ζ − =0 (1)
∂t ∂z
∂φ 1
+ ∇φ · ∇φ + z = 0, (2)
∂t 2
∂ ∂ ∂ T ∂ ∂ T
with ζ the wave height, t time, and ∇ = ( ∂x , ∂y , ∂z ) , ∇s = ( ∂x , ∂y )
defined with respect to a Cartesian coordinate system (x, y, z), where z is the
coordinate direction pointing upward from the flat water surface. At a wave
maker a prescribed normal velocity is imposed and at other solid surfaces a
zero normal velocity. The equations are made dimensionless with an average
water depth and the gravitational constant.
The numerical solution of large amplitude nonlinear water waves is non-
trivial since the free surface boundary conditions must be imposed at the
actual free surface, which must be determined as part of the solution process.

54 Yan Xu and Jaap J.W. van der Vegt
The free surface experiences large deformations during the wave evolution, in
particular due to wave interactions. An additional complication is that the
mathematical model has no inherent dissipation, which makes it difficult to
ensure sufficient stability and robustness of the numerical scheme.
In this paper we will present a new space-time discontinuous Galerkin
finite element algorithm to compute large amplitude nonlinear waves. The al-
gorithm uses basis functions which are discontinuous both in space and time,
and discretizes the equations directly in four dimensional space, with time as
the fourth dimension. This approach combines the benefits of discontinuous
Galerkin methods, for instance their suitability for hp-mesh adaptation and
parallel computing, with an accurate representation of the time-dependent
boundary using deforming elements. The space-time DG algorithm is closely
related to the so called arbitrary Lagrangian Eulerian technique, which pro-
vides more flexibility in the mesh deformation algorithm, see e.g. [2], but the
space-time DG method ensures that the discretization remains conservative on
deforming meshes. The space-time discontinuous Galerkin method discussed
in this paper is an extension to nonlinear waves of the spatial discretization
discussed in [1] for linear waves using the space-time techniques presented in
[2].
2 Space-time discontinuous Galerkin formulation
In this section we summarize the space-time discontinuous Galerkin finite

element discretization.
We consider the space-time domain Eh ⊂ R4 split into space-time slabs
Enh with the free-surface boundary ΓSn for the time intervals (tn , tn+1 ) with
n = 0, 1, · · · . We introduce a finite element tessellation Thn with space-time
elements K on Enh and define the finite element spaces Vpn and Σpn associated
with the tessellation Thn as:
Vpn :={v ∈ L2 (Enh ) | v|K ∈ Pp (K), ∀K ∈ Thn },

Σpn :={σ ∈ [L2 (Enh )]3 σ|K ∈ [Pp (K)]3 , ∀K ∈ Thn },

with L2 (Enh ) the space of Lebesgue square integrable functions on Enh and
Pp (K) the space of polynomials of degree p on K. The finite element space
Wpn associated with the free surface is defined as
Wpn :={v ∈ L2 (ΓSn ) | v|FSn ∈ Pp (FSn ), ∀FSn ⊂ ΓSn },
where L2 (ΓSn ) is the space of Lebesgue square integrable functions on ΓSn .

Next, we define some trace operators to manipulate the numerical fluxes
in the discontinuous Galerkin formulation. For v ∈ Vpn we define the average
hvi and jump [[v ]] operators of v at an internal face F ∈ FIn as follows:
Space-Time DG Method for Large Amplitude Nonlinear Water Waves 55
1
hvi := (vL + vR ) , [[v ]] := vL n̄L + vR n̄R , (3)
2
with vL := v|∂KL and vR := v|∂KR , and KL , KR the elements connected to
the face F ∈ FIn with outward space normal vectors n̄L and n̄R , respectively.
For q ∈ Σpn we similarly define qL , qR , hqi and [[q ]].
The novel ingredient in the finite element formulation discussed in this
paper is the incorporation of the kinematic condition at the free surface (1)
as a natural boundary condition in the finite element formulation. The space-
time finite element discretization then will automatically account for the mesh
movement necessary to follow the free surface waves. In order to accomplish
this we need to establish a relation between the function f = ζ − z, describing
the free surface, the wave height ζ and the space-time normal vector n. A
straightforward calculation using (1) shows that
∇f ( ∂f , ∇f )T 1 ∂ζ ∂ζ ∂ζ
n = (nt , n̄)T = = ∂t = ( , , , −1)T , (4)
|∇f | |∇f | |∇(ζ − z)| ∂t ∂x ∂y
∂ ∂ ∂ ∂ T
with ∇ the space-time nabla operator defined as ∇ = ( ∂t , ∂x , ∂y , ∂z ) . We
can rewrite the kinematic condition (1) as:
1 ∂f ∇f
+ ∇φ · = 0, (5)
|∇f | ∂t |∇f |
which implies using (4) that the space component of the normal velocity at
the free surface ΓS is equal to
−1 ∂ζ
n̄ · ∇φ = . (6)
|∇(ζ − z)| ∂t
This relation can be used directly in the DG formulation of the Laplace equa-
tion, which is defined as:
Find a (φh , ζh ) ∈ Vpn × Wpn , such that for all (v1 , v2 ) ∈ Vpn × Wpn ,
∂ζh
ζh+ − ζh−
Bh (φh , v1 ) + ( ∂t
, v1 )ΓSn + ( , v+ ) + = Lh (v1 ), (7)
|∇h (ζh − z)| |∇h (ζh+ − z)| 1 ΓS (tn )
∂φh − +
( , v2 )ΓSn + (ζh , v2 )ΓSn + (φ+h − φh , v2 )ΓS (t+
n)
(8)
∂t
1 X X
+ ( (∇h φh − RF ([[φh ]])) · (∇h φh − RF ([[φh ]])), v2 )ΓSn = 0,
2
F∈FI
n nF∈FI
± ±
where ΓS (t+
n ) = lim ΓS (tn + ), φh = lim φh (tn ± ), ζh = lim ζh (tn ± ), with
↓0 ↓0 ↓0
φ− −
h and ζh the known potential and wave height at t = tn . The operators
Bh : Vp × Vpn → R and Lh : Vpn → R are defined as:
n
Z Z

Bh (u, v) = ∇h u · ∇h vdx − [[u]] · ∇h v + [[v ]] · ∇h u ds
En
h Γ0n
X Z
+ (ηF + nf ) RF ([[u ]]) · RF ([[v ]])dx, (9)
F∈FIn En
h
Z
Lh (v) = vgN ds,
n
ΓN
3
with the local lifting operator RF : L2 (F) → Σpn :

Z Z
RF (q) · σdx = q · hσi ds, ∀σ ∈ Σpn . (10)
En
h F
The resulting nonlinear equations are solved with a Newton method with
special attention given to the discretization of the free surface terms. During
each time step and Newton iteration the mesh is adjusted to accommodate for
the free surface motion. For more details about the scheme and the numerical
implementation, we refer the reader to [3].
3 Fourier analysis of discrete scheme

In this section, we conduct a Fourier analysis of the two-dimensional space-
time discretization for the linear water wave equations as in [1]. This will
provide us with information on the dissipation and dispersion error in the
wave motion due to the numerical discretization.
We assume a two-dimensional domain with a uniform mesh with Nx × Nz
coordinates xj = j∆x, zm = m∆z and periodic boundary conditions in the
x-direction. For the linear water wave equations, the space-time DG scheme
(7)–(8) reduces to the following form
∂ζh
Bh (φh , v1 ) + ( , v1 )ΓSn + (ζh+ − ζh− , v1+ )ΓS (t+
n)
= Lh (v1 ), (11)
∂t
∂φh − +
( , v2 )ΓSn + (φ+
h − φh , v2 )ΓS (t+
n)
+ (ζh , v2 t)ΓSn = 0. (12)
∂t
We use the Fourier ansatz for the coefficients φbj and ζbj in the space-time DG
discretization at time level tn = n∆t:
φbj (z, tn ) = λn exp(ikj∆x)φbF (z), ζbj (tn ) = λn exp(ikj∆x)ζbF (13)
with√λn = exp(−iωn∆t) the amplification factor, k the wavenumber and

i = −1.
If we introduce (13) into the space-time discretization for the linear case
(11)–(12), then we obtain the following generalized eigenvalue problem:
   bF     bF 
F 0 φD 0 φD
M   0
λ P   φbF = Y φbF (14)
  
S   S 
0H S ζbF 0V 0 ζbF
Space-Time DG Method for Large Amplitude Nonlinear Water Waves 57
-1.7E
0.8 7 0.8 -0 6
E-0 5.0E-07
5.0
5.0E-07
6
0.6
E-0 0.6
2.0 -5.6E-07
5.0E-07
! "#
! "#
6 -5.6E-07
E-0
0.4 2.0 0.4
-5.6E-07
0.2 0.2 -5.6E-07

E-0
7 -0 6
5.0 7E
-1 .
0.05 0.1 0.15 0.2 0.25 0.05 0.1 0.15 0.2 0.25
h h
Fig. 1. Dispersion (left) and dissipation (right) for the space-time DG method with
quadratic polynomial basis functions.
Np (Nz −1) bF
with φbF
D ∈R , φS ∈ RNp , ζbF ∈ RNs , where Np = (p + 1)3 , Ns = (p +
2
1) and p the order of the polynomials. The suffix S refers to the coefficients in
elements connected to the free surface and D to elements not connected to the
free surface. The matrices P, Y ∈ RNp ×Ns , H, V ∈ RNs ×Np and S ∈ RNs ×Ns
are related to the free surface. The matrix MF ∈ RNp Nz ×Np Nz is a Hermitian
positive definite block-tridiagonal matrix.
The eigenvalue λ is computed with MATLAB for a wide range of ∆t values
and k ∈ (0, 2π]. For all cases, the modulus of λ is always less than or equal to
one, hence the numerical discretization is unconditionally stable.
We also use Fourier analysis to compute the dispersion and dissipation
error of the numerical scheme by comparing the frequency and dissipation of
the discrete modes with the exact harmonic wave solution. The results for
quadratic polynomial basis functions are shown in Fig. 1.
4 Numerical example
As an illustration of the numerical scheme, we consider nonlinear free-surface

waves generated by a wave maker at x = 0 with time-harmonic frequency
ωw = 2 in a domain Ω = [0, 4] × [−1, 0]. Homogeneous Neumann boundary
conditions are assumed at the bottom z = −1 and at the end of the domain at
x = 4. The initial free-surface height and velocity potential are zero and the
computational mesh is constantly updated to follow the free surface motion.
The wave profiles in the domain at T = 7.5 and T = 20 are presented in Figs.
2. The wave profiles compare well for different meshes and different polynomial
order indicating that the wave motion is computed accurately.
0.05 0.1
STDG, P2, 80x10, T=7.5 STDG, P2, 80x10, T=20

STDG, P1, 120x10, T=7.5 STDG, P1, 120x10, T=20
Free surface elevation
Free surface elevation

0.05
-0.05 -0.05
0 1 2 3 4 0 1 2 3 4
x x
Fig. 2. Wave profile at T = 7.5 (left) and T = 20 (right) generated by a wave maker
at x = 0 for polynomial basis functions of degree p = 1 and 2 using the space-time
DG method.
5 Concluding remarks
In this paper we have presented a space-time discontinuous Galerkin method

for nonlinear water waves. This technique results in a higher order accurate
conservative numerical scheme on time dependent deforming meshes which
are necessary to follow the free surface evolution. A Fourier analysis is given
for the linear water wave equations indicating that the discretization is un-
conditionally stable and shows that the dispersion and dissipation errors of
the scheme are minimal. Numerical examples for nonlinear free-surface waves
show that the space-time DG method can accurately compute the nonlinear
waves generated by a wave maker.
References
1. J.J.W. van der Vegt, S.K. Tomar, Discontinuous Galerkin method for linear
free-surface gravity waves, J. Sci. Comput. 22-23 (2005) 531-567.
2. J.J.W. van der Vegt and H. van der Ven, Space-time discontinuous Galerkin
finite element method with dynamic grid motion for inviscid compressible flows,
Part I. General formulation, J. Comput. Phys., 182 (2002), 546-585.
3. J.J.W. van der Vegt and Y. Xu, Space-time discontinuous Galerkin method for
nonlinear water waves, preprint for J. Comput. Phys., 2006.
A discontinuous Galerkin method with
Hancock-type time integration for hyperbolic
systems with stiff relaxation source terms
Yoshifumi Suzuki1 and Bram van Leer2

1
The University of Michigan, Ann Arbor, MI 48109, USA fumifumi@umich.edu
2
The University of Michigan, Ann Arbor, MI 48109, USA bram@umich.edu
A new discretization method for hyperbolic systems with stiff relaxation

source terms (hyperbolic-relaxation equations) is introduced. The method
is based on Huynh’s “upwind moment scheme” for hyperbolic conservation
laws with implicit treatment of the source term. A Von Neumann analysis
shows superiority in both stability and accuracy of the resulting fully discrete
scheme over the method-of-line based semi-discrete schemes, and numerical
experiments confirm the analysis. Our goal is developing a unified numerical
method for simulating a continuum and transitional flow.
1 Introduction
In recent years, the need for robust high-order (more than second-order ac-
curate) discretizations for high-fidelity CFD on unstructured grids has been
widely recognized. To be attractive, a high-order method must not only be ac-
curate, but efficient as well, thus reducing the total CPU time needed to yield
a given accuracy. In this paper we propose a combination of two approaches
toward robust, accurate, and efficient schemes for advection-dominated flows
on unstructured grids, one at the PDE level and the other at the discretization
level.
The first approach is replacing everybody’s favorite Navier-Stokes (NS)
equations by a larger set of first-order hyperbolic-relaxation PDEs which con-
tains the NS equations. (N.B.: here “first-order” refers to the order of the
PDEs.) In comparison with second-order PDE models such as the NS equa-
tions, first-order PDE models offer many numerical advantages. For example,
the latter require smaller discrete stencils, reduce communications in par-
allel processing, can replace global stiffness from diffusion terms with local
stiffness from source terms, and yield the best accuracy on non-smooth, adap-
tively refined grids. We will therefore consider the one-dimensional hyperbolic-
relaxation form

60 Yoshifumi Suzuki and Bram van Leer
1
∂t u + ∂x f (u) = s(u), (1)

where u is the vector of conserved variables, f is the flux of u, s is the source
term, and is the relaxation time. This is the form of moment closures of the
Boltzmann equation, where the source term describes departure from local
thermodynamic equilibrium. In the near equilibrium ( 1), that is, when
the system is stiff, the system formally reduces to a system of second-order
equations,
∂t U + ∂x F(U) = ∂x (D∂x U), (2)
where D is a tensor of diffusion coefficients, with eigenvalues proportional to
.
The second approach is to adopt a discretization method which can pre-
serve compactness in high-order methods. In standard finite-volume methods,
higher-order accuracy relies on piecewise-polynomial reconstruction, which
requires extended stencils. For instance, stencils for the quadratic reconstruc-
tion (third-order) on tetrahedral grids include 50 to 70 cells [2]. Discontinuous
Galerkin (DG) methods overcome the issue of reconstruction by using extra
equations for updating the polynomial representation of state variables. For
a comprehensive literatures review, see Cockburn and Shu [1]. Currently, the
most successful DG methods are semi-discretizations combined with TVD
Runge-Kutta (RK) ODE solvers [1], denoted as RKm-DG(k) where m is the
order of the RK method and k is the degree of polynomial of basis function.
The method proposed here is based on the “upwind moment scheme”
recently developed by Huynh for hyperbolic conservation laws [4]. The solu-
tion representation is only piecewise linear; higher-order extensions will be
presented elsewhere. The two key characteristics of this method are: 1) cell
variables are updated over a half time step without any interactions with
neighboring cells (Hancock’s observation [8]); 2) the gradient of each flow
variable evolves by an independent equation (DG representation). The result-
ing scheme looks promising in comparison to the currently popular methods
based on a semi-discretization with method-of-line (M.O.L.) time integration.
It is a fully-discrete, one step method with one intermediate update step
needed for computing the volume integral of the fluxes. It requires solving
a Riemann problem twice at each cell interface but achieves third-order ac-
curacy in time and space. Interestingly, the upwind moment scheme for a
linear advection equation reduces to Van Leer’s “scheme III” [7], which is a
DG spatial discretization with an exact shift operator for time evolution. It
was shown that the method is linearly stable up to CFL number 1 with an
upwind flux, whereas for the RK2DG(1) (second-order) the limit is 1/3 and
for the RK3DG(2) (third-order) it is 1/5 [1]. Here we extend the upwind mo-
ment method to a simple hyperbolic-relaxation system, apply a Von Neumann
analysis to it, and confirm its results by numerical experiments.
A discontinuous Galerkin method with Hancock-type time integration 61
2 Numerical method for first-order PDEs

When discretizing hyperbolic-relaxation equations, Eq.(1), source terms have
to be treated implicitly to ensure stability in the stiff regime ( 1). Con-
versely, advection terms are treated explicitly due to the complexity of the
flux evaluation. Thus, it is expected that the stability of a method is solely
constrained by an advective CFL condition. This can indeed be realized; how-
ever, the simple implicit treatment of source terms does not always guarantee
higher-order accuracy in the stiff regime. In our method, the time integration
of source terms is computed by a two-point Gauss-Radau quadrature which
is L-stable if advection terms are omitted. The same quadrature points are
used for the volume integral of the flux whereas Huynh’s original upwind mo-
ment method uses three-point Gauss-Lobatto quadrature. Because the source
terms do not contain derivatives, the method is point-implicit, meaning that
the implicitness is local, with no need for information from neighboring cells.
For brevity, we only consider the one-dimensional case with uniform grid;
f (u) and s(u) are assumed to be linear. Let h be the cell width and Ij =
[xj−1/2 , xj+1/2 ] be the domain of cell j. The general DG method is obtained
by multiplying Eq.(1) by a test function v(x), integrating over the interval Ij ,
and applying integration by parts on the flux term. To derive the fully-discrete
method, integrate again in time over T n = [tn , tn+1 ]:
Z tn+1 Z xj+1/2

u(x, t)v(x)dx
=− f (u(x, t))v(x) dt
Ij tn Tn xj−1/2
ZZ ZZ
1
+ f (u(x, t))vx (x)dxdt + s(u(x, t))v(x)dxdt. (3)
Ij ×T n Ij ×T n
Note that Eq.(3) is still exact. To approximate the boundary and volume
integrals on the right-hand side, we use Legendre polynomials for both test
and basis functions. Thus, the basis of v(x) be {φ0 (x), φ1 (x)} where φ0 (x) =
1, Rφ1 (x) = (x − xj )/h, and u(x, t) ≈ ū(t) + ∆u(t)(x R − xj )/h where ū(t) ≡
1 12
h Ij u(x, t)dx is the cell average and ∆u(t) ≡ h2 Ij (x − xj )u(x, t)dx is the
undivided gradient of u. At the boundary, the midpoint rule is used for the
time integration of the interface flux, thus the upwind flux f ∗ (uL , uR ) ≡
A+ uL + A− uR with A(u) ≡ ∂f /∂u is evaluated based on the state variables
at the half-time level,
Z tn+k
1 n+k/2 ∗ n+k/2 n+k/2
f (u(xj+1/2 , t))dt ≈ fj+1/2 ≡ fj+1/2 (uj+1/2,L , uj+1/2,R ), (4)
k∆t tn
n+k/2
where ∆t = tn+1 − tn . The cell-interface values, uj+1/2,L/R are obtained by
a Taylor-series expansion of u(x, t) in space and time using the Lax-Wendroff
procedure (replacing time derivative by spatial derivative); the final forms are

n+k/2 1 k∆t n k∆t n+k/2
uj+1/2,L = ūnj + I− A(ūnj ) ∆uj + s(uj+1/2,L ),
2 h 2
(5)
n+k/2 1 k∆t n k∆t n+k/2
uj+1/2,R = ūnj+1 − I+ A(ūnj+1 ) ∆uj+1 + s(uj+1/2,R ).
2 h 2
Note that implicit character is caused by the source terms. The first update
formulas to compute the intermediate values at the Radau point (k = 1/3)
are found, after some algebra, to be
n+1/3 ∆t h n+1/6 n+1/6

i ∆t
n+1/3
ūj = ūnj − fj+1/2 − fj−1/2 + s(ūj ), (6)
3h 3
n+1/3 n ∆t h
n+1/6 n+1/6 n+1/6
i ∆t n+1/3
∆uj = ∆uj − 6 fj+1/2 + fj−1/2 − 2f (ūj ) + s(∆uj ),
3h 3
where the volume integral of the flux is approximated using the trapezoidal
n+1/6 n+1/3
rule, ūj = (ūnj + ūj )/2. The final update formulas are given by
∆t h n+1/2 n+1/2
i ∆t h
n+1/3
i
ūn+1
j = ūnj − fj+1/2 − fj−1/2 + 3s(ūj ) + s(ūn+1
j ) ,
h 4
n+1 n 6∆t h n+1/2 n+1/2 n+1/2
i
∆uj = ∆uj − fj+1/2 + fj−1/2 − 2f (ūj ) (7)
h
∆t h n+1/3 n+1
i
+ 3s(∆uj ) + s(∆uj ) ,
4
where the two-point Radau-Gauss quadrature is adopted for time integration
n+1/2 n+1/3
of the source terms, and also for the intermediate values ūj = (3ūj +
n+1
ūj )/4. Note that this source-term treatment would be only second-order
accurate if nonlinear problems were considered since s(u)j = s(ūj ) + O(h2 ).
This is also true for the volume integral of the flux.
3 Von Neumann analysis and grid convergence study
At first, a dimensionless linear advection equation ∂t u + r∂x u = 0, |r| ≤ 1 is

considered. Here, the normalization is rather uncommon. The advection speed
is normalized by the implicitly defined “frozen” wave speed (= 1). The motiva-
tion of this will be clear once hyperbolic-relaxation equations are considered. A
Von Neumann analysis is employed to show the order of accuracy and the dom-
inant numerical dissipation/dispersion errors for low-frequence modes. Recall
that, for a linear advection equation, both the upwind moment scheme and the
proposed method without source terms are identical to Van Leer’s scheme III
[7]. The new method (HaDG(1)) is compared with three other methods: semi-
discrete higher(second)-order Godunov with second-order RK (RK2HR2),
RK2DG(1), and RK3DG(1). Making the analysis more general, a q-scheme is
q
employed as flux function: fj+1/2 (uL , uR ) = r(uL +uR )/2−|q|(uR −uL )/2, in-
cluding the upwind (q = r), Rusanov (q = 1), and Lax-Friedrichs (q = h/∆t)
A discontinuous Galerkin method with Hancock-type time integration 63
fluxes respectively. The unity value in the Rusanov flux comes from the dimen-
sionless frozen wave speed. For DG methods, the amplification factor becomes
a 2 × 2 matrix even though a scalar equation is considered. Thus two eigen-
values exist, but only the one accurate for low frequencies is described here.
Recalling that the eigenvalue of the exact advection operator is −irk, where
k is the wave number, we only list the numerical error part of the eigenvalues
in the long wave-length limit:

1 hq i
λRK2HR2 = c3 (rν)2 + k 3 + 9c4 − (rν)3 k 4 + . . . ,
2 r

r
λRK2DG(1) = c3 (rν)2 k 3 + c4 − 9(rν)3 k 4 + . . . , (8)
q

r
λRK3DG(1) = c4 + 3(rν)3 k 4 + . . . ,
q
[(rν)2 − 1]2

λHaDG(1) = c4 3rν(qν − 1) + k4 + . . . ,
q/r − rν
where c3 = −irh2 /6, c4 = −rh3 /72. The CFL number ν is defined by
the dimensionless frozen wave speed instead of the advection speed r, thus
ν ≡ 1∆t/h. These results show that the leading error of both HaDG(1) and
RK3DG(1) is fourth-order whereas the other two methods are third-order.
Thus DG(1) with M.O.L. can be a third-order method if third-order RK is
used, however, it requires three flux calculations at each cell face whereas
HaDG(1) requires two to achieve the same order. Note the vanishing of the
error in HaDG(1) for qν = rν = 1. Regarding stability, it was shown that
both RK2DG(1) and RK3DG(2) with the LxF flux are linearly unstable [6].
The stability limit of each method with upwind and Rusanov fluxes is as fol-
lows: RK2HR2, qν ≤ 1; RK2DG(1), qν ≤ 1/3; RK3DG(1), qν ≤ 0.409 [1].
For HaDG(1) we heuristically found that the linear stability limit is different
for different fluxes: HaDG(1) with upwind, rν ≤ 1; HaDG(1) with Rusanov,
ν ≤ 1/3.
Next, consider a 2 × 2 linear system of hyperbolic-relaxation equations;
u = [u, v]T , f = [v, u]T , and s = [0, ru − v]T in Eq.(1). This system has
“frozen” wave speeds ±1 when relaxation is weak ( 1); when the relaxation
dominates ( 1), it reduces to advection-diffusion equation, ∂t u + r∂x u ≈
(1 − r2 )∂xx u, with an “equilibrium” wave speed of r. For stability, |r| ≤ 1.
Previously, it was shown that the upwind flux applied to the model equations
reduces to the Rusanov flux when 1 [3]. A Von Neumann analysis in the
long wave-length limit shows the dominant dispersion/dissipation errors are
given by Eq.(8) with q = 1 where λexact = −irk − (1 − r2 )k 2 in this case. To
confirm the analysis, consider an initial-value problem on a periodic domain
with a harmonic initial condition u0 = v0 ≡ cos(2πx), x ∈ [0, 1]; the other
parameters r, are chosen so that the reduced equation becomes an advection-
dominated advection-diffusion equation (r = 1/2, = 10−5 ). The new method
is compared with an HR2 and a DG(1) method, both incorporating the IMEX-
RK method [5]. The CFL conditions based on the unit wave speed are set as
νHaDG(1) , νRK2DG(1) = 0.3, and νRK2HR2 = 0.8. Figure 1 shows that the new
method is the least dissipative and dispersive of all, whereas the RK2HR2
method produces a completely inaccurate solution. Figure 2 shows a grid
convergence study of the solution at the final time tend = 300. Third-order
convergence is observed for the new method whereas RK2HR2 and RK2DG(1)
show second-order convergence in the L2 -norm.
0
10
L2−norm of conserved variable, L2(u)

1 :HaDG(1), N=20
:RK2DG(1), N=20
0.8 −1
:RK2HR2, N=40
Conserved variable, u
10
0.6 :Exact Solution 2
0.4
−2
0.2 10
2
0
−0.2
−3
:HaDG(1)
−0.4 10 :RK2DG(1)
:RK2HR2
−0.6 3
−0.8 −4
10 1 2 3
−1 10 10 10
0 0.2 0.4 0.6 0.8 1
X Degrees of freedom
Fig. 1. Numerical solutions at the Fig. 2. Grid convergence study in the

final time tend = 300 in the near- near-equilibrium limit. Third-order
equilibrium limit. The new method is convergence is observed for the new
the most accurate of all. method.
References
1. B. Cockburn and C. -W. Shu. Runge-Kutta discontinuous Galerkin methods for
convection-dominated problems. J. Sci. Comput., 16(3):173–261, 2001.
2. M. Delanaye. Quadratic reconstruction finite volume schemes on 3D arbitrary
unstructured polyhedral grids. AIAA Paper 1999-3259, 1999.
3. J.A.F. Hittinger, Y. Suzuki, and B. van Leer. Investigation of the discontinuous
Galerkin method for first-order PDE approaches to CFD. AIAA Paper 2005-
4989, 2005.
4. H. T. Huynh. An upwind moment scheme for conservation laws. In ICCFD 3,
Toronto, Canada, 2004.
5. L. Pareschi1 and G. Russo. Implicit-explicit Runge-Kutta schemes and appli-
cations to hyperbolic systems with relaxation. J. Sci. Comput., 25(1):129–155,
2005.
6. W. J. Rider and R. B. Lowrie. The use of classical Lax-Friedrichs Riemann
solvers with discontinuous Galerkin methods. Int. J. Numer. Meth. Fluids,
40:479–486, 2002.
7. B. van Leer. Towards the ultimate conservative difference scheme. IV. A new
approach to numerical convection. J. Comput. Phys., 23:276–299, 1977.
8. B. van Leer. Upwind and high-resolution methods for compressible flow: From
donor cell to residual-distribution schemes. Commum. Comput. Phys., 1(2):192–
206, 2006.
Very High Order, Non-Oscillatory Fluctuation
Distribution Schemes
M.E.Hubbard1 and N.Z.Mebrate1
School of Computing, University of Leeds, Leeds, LS2 9JT, U.K.

meh@comp.leeds.ac.uk, nmebrate@comp.leeds.ac.uk
1 Introduction
Fluctuation distribution schemes for approximating multidimensional systems
of conservation laws have developed to a stage where they can be used reliably
to produce accurate simulations of complex steady state fluid flow phenomena
using unstructured meshes [6]. These methods are often required to avoid
producing unphysical, numerically induced, oscillations in the flow field, and
this has so far restricted them to second order accuracy.
More recent research has led to higher order methods (for both steady state
and time-dependent flows) which use the fact that a k th order method can be
derived by evaluating the fluctuation exactly with respect to a (k −1)th degree
polynomial representation of the dependent variable and then distributing it
in a linearity preserving manner [4]. So far, two successful approaches have
been proposed for constructing this high order interpolant within each mesh
cell [4, 5]. A third alternative will be presented here.
As they stand, none of these three approaches can guarantee the absence
of spurious oscillations from the flow without the application of an additional
smoothing stage. This paper will briefly describe a technique which can be
combined with any of the above procedures to provide schemes which are both
higher than second order accurate and free of spurious oscillations. Brief re-
sults will be shown to demonstrate its effectiveness in approximating the scalar
advection equation on two-dimensional, unstructured, triangular meshes.
2 Fluctuation Splitting
Consider the two-dimensional scalar conservation law given by
ut + fx + gy = 0 or ut + λ · ∇u = 0 (1)
on a domain Ω, with u(x, y, t) = g(x, y, t) imposed on the inflow part of the
T
∂f ∂g
boundary ∂Ω. λ = ∂u , ∂u defines the advection velocity associated with

66 M.E.Hubbard and N.Z.Mebrate
the conservation law (1). This equation has an associated fluctuation, assumed
here to be calculated over a triangular mesh cell 4 and given by
ZZ I
φ = − λ · ∇u dΩ = u λ · dn , (2)
4 ∂4
in which n represents the inward pointing normal to the cell boundary. It

will be assumed throughout this paper that u has a continuous piecewise
polynomial representation and the integration in (2) is carried out exactly
(though this isn’t always necessary [2]). Given this, a simple forward Euler
discretisation of the time derivative leads to an iterative update of the nodal
solution values which is generally written [7] as
∆t X j
un+1
i = uni + αi φj , (3)
Si
j∈∪4i
where ∆t is the time-step, Si is the area of the median dual cell corresponding
to node i, αij is the distribution coefficient which indicates the appropriate
proportion of the fluctuation φj to be sent from cell j to node i, and ∪4i
represents the
P set of cells with vertices at node i. Conservation is assured
as long as i∈4j αij = 1 , ∀j, where 4j represents the set of nodes at the
vertices of cell j.
2.1 A Second Order Positive Scheme
The basis of most positive fluctuation distribution schemes is the N scheme

[7]. Its derivation relies on the fluctuation being evaluated exactly, as
1X X
φ = − ui λ̂ · ni = − ui ki , (4)
2
i∈4 i∈4
where the symbol ˆ indicates an appropriately linearised quantity and ni is

the inward unit normal to the ith edge (opposite the ith vertex) multiplied
by the length of that edge. The dependent variable u is assumed to vary in
a continuous piecewise linear manner, with the unknowns stored at the mesh
nodes. Now, for each triangle it is always possible to locate a pair of vertices
i2 and i3 for which ki2 ki3 ≥ 0. The fluctuation can then be written
φLO = ki2 (ui1 − ui2 ) + ki3 (ui1 − ui3 ) = φLO LO
i2 + φi3 . (5)
When the choice of i2 and i3 is not unique, i.e. ki = 0 for some vertex i,
this node can be paired with either of the other vertices. This leads to an
alternative formulation of the N scheme, given by
Si1 ui1 → Si1 ui1 + ∆t ki−2 (ui1 − ui2 ) + ∆t ki−3 (ui1 − ui3 )
Si2 ui2 → Si2 ui2 + ∆t ki+2 (ui1 − ui2 )
Si3 ui3 → Si3 ui3 + ∆t ki+3 (ui1 − ui3 ) , (6)
Very High Order, Non-Oscillatory Fluctuation Distribution Schemes 67
in which · ± denotes the positive/negative part of the quantity. This scheme

is locally (and hence globally) positive, so the iteration given by (3) is condi-
tionally stable, the appropriate restriction on the time-step being
Si
∆t ≤ P . (7)
1 j
j∈∪4i max 0, 2 λ̂ · ni
The contribution made by cell j to node i by the N scheme can be written

as (φji )N = (αij )N φj , where φj is the fluctuation in cell j (see (6)). The
contributions due to the PSI scheme, the most commonly used of the second
order non-oscillatory fluctuation distribution schemes, can then be defined by
[(αij )N ]+
(φji )P SI = P j N +
φj = (αij )P SI φj . (8)
k∈4j [(αk ) ]
This scheme can easily be shown to be conservative, positive for the time-step
given by (7), and linearity preserving, i.e. (αij )P SI is bounded so the order
of accuracy of the steady state scheme is one higher than the degree of the
polynomial used to represent u (in this case second order) [1, 3].
3 Higher Order Positive Schemes

The first stage in the creation of higher order fluctuation distribution schemes
is the construction of a higher order representation of the dependent variable
u in (2). In this work a continuous piecewise quadratic interpolant of the
nodal data is sought, though higher degree polynomials are also possible. In
previous work, quadratics have been produced in each mesh cell by either
(1) reconstructing ∇u at the mesh vertices and combining them with the
nodal values of u to satisfy the necessary degrees of freedom [5], or (2) storing
and updating values of u at additional nodes at the midpoints of each cell
edge [4]. A third approach is proposed here, which uses the values of u at
additional mesh nodes beyond the given cell to construct the local polynomial:
for a quadratic, three additional nodes are required and these are chosen to
be the vertices of the three neighbouring triangles opposite the given cell’s
edge (except for boundary cells, which are treated as special cases, and some
extreme mesh topologies, which do not occur here). This does not immediately
produce a continuous representation, which is imposed here by averaging the
two interpolants at each edge.
Unfortunately, it is easy to show that the fluctuations which result from
any of these three interpolants cannot lead to a locally positive scheme if they
are distributed locally within the mesh cell (or subcell in the case method (2))
[8]. It is, for example, possible to have a non-zero fluctuation in a (sub)cell for
which ui1 = ui2 = ui3 . The possibility of distributing the fluctuation farther
afield is currently under investigation but has yet to yield a scheme of practical
value.
An alternative approach is considered here which overcomes this obstacle

by modifying the high order interpolant. Let ū(x) be the linear interpolant
of the values of u at the vertices of a given triangular (sub)cell and u(x) =
ū(x) + δu(x) be a higher order representation of the data within that triangle.
If the high order correction δu(x) on each (sub)cell edge i1 i2 is limited to give
δu0 (x) along that edge which satisfies
|δu0i1 i2 (x)| ≤ C|ui1 − ui2 | ∀ x = µ xi1 + (1 − µ) xi2 , 0 ≤ µ ≤ 1 ,(9)
for some finite constant C ≥ 0 then, subject to an appropriate restriction on
the time-step in (3), it is possible to distribute the fluctuation (2) due to the
modified interpolant u0 (x) = ū(x) + δu0 (x) to the vertices of the specified
(sub)cell in a locally positive manner [8].
For a quadratic representation of u and linearly varying λ only the edge
midpoints are required by the quadrature for exact evaluation of the fluctua-
tion, so the limiting on edge i1 i2 is carried out by using
u i1 + u i2
u0i1 i2 = + αi1 i2 (ui1 − ui2 ) (10)
2
as the limited solution value at the edge midpoint, where

ui1 i2 − (ui1 + ui2 )/2
αi1 i2 = max −C, min C, . (11)
u i 1 − u i2
A value of C = 0.25 is chosen here. This is the largest value that guarantees
that the limited interpolant along each edge is monotonic (if it is chosen to
be of the form δu0 (x) = C 0 δu(x) where C 0 ∈ [0, 1]). Also, larger values of C
tend to reduce the rate of convergence of the iteration (3) to the steady state.
The general case requires limiting at additional quadrature points.
As a result of the above procedure, the limited high order fluctuation can
be written
I
φLIM = u0 λ · n dΓ = Ki2 (ui1 − ui2 ) + Ki3 (ui1 − ui3 ) (12)
∂4
where i1 , i2 and i3 are chosen to be precisely those vertices designated by the

N scheme (5). Explicit, bounded expressions can easily be found for Ki2 and
Ki3 [8]. The formulation of the N scheme given in (6) can then be applied
directly to this higher order fluctuation, i.e.
Si1 ui1 → Si1 ui1 + ∆t Ki−2 (ui1 − ui2 ) + ∆t Ki−3 (ui1 − ui3 )
Si2 ui2 → Si2 ui2 + ∆t Ki+2 (ui1 − ui2 )
Si3 ui3 → Si3 ui3 + ∆t Ki+3 (ui1 − ui3 ) . (13)
This scheme is clearly locally positive for a small enough time-step, the
limit on which is approximately inversely proportional to C. The distribu-
∗
tion coefficients of the resulting N-like (N ∗ ) scheme take the form (φji )N =
∗
(αij )N φLIM
j , and these can be limited in precisely the manner which created
the PSI scheme (8) by imposing linearity preservation on the N scheme, i.e.
Very High Order, Non-Oscillatory Fluctuation Distribution Schemes 69
1
0.8
0.9
0.6 0.8
0.7
0.4
0.6
0.2 0.5
0.4
0 0.3
−1
−0.8 1
−0.6 0.2
−0.4
−0.2
0 0.5 0.1
0.2
0.4
0.6
0.8 0 0
1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
1
0.8
0.9
0.6 0.8
0.7
0.4
0.6
0.2 0.5
0.4
0 0.3
−1
−0.8 1
−0.6 0.2
−0.4
−0.2
0 0.5 0.1
0.2
0.4
0.6
0.8 0 0
1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Fig. 1. The new scheme applied to circular advection of a square wave (top) and a
cosine-squared profile (bottom).
∗
∗ [(αij )N ]+ ∗
(φji )P SI = P j N∗ +
φLIM
j = (αij )P SI φLIM
j . (14)
k∈4j [(αk ) ]
As with the PSI scheme, the limiting procedure will never increase the magni-
tude of the distribution coefficients, so the positivity condition for the scheme
(13) is actually stronger than necessary.
4 Results
The two-dimensional scalar advection equation (1) is approximated, over the
domain [−1, 1] × [0, 1] with λ = (y, −x)T and u(x, y, 0) = g(x) for x ∈ [−1, 0]
and y = 0 (where g(x) is chosen to take a variety of forms to demonstrate the
properties of the schemes) [8].
Figure 1 shows the results obtained from the positive, high order scheme
derived from extending the stencil. Replacing g(x) with much smoother func-
tion, and carrying out a series of experiments on successively refined meshes
suggests an order of accuracy of 2.36 in the L1 norm and 2.23 in the L∞ norm.
Very similar results are obtained by applying the same limiting procedure to
the submesh reconstruction and gradient recovery schemes [8]. In all cases the
oscillations are removed completely and the results are significantly better
than those of the PSI scheme. All of the experiments have converged to their
steady states to machine accuracy.
5 Conclusions
A new fluctuation distribution scheme has been presented which is demon-
strated to give higher than second order accuracy at the steady state for
the scalar advection equation without introducing any spurious oscillations.
The procedure used to impose positivity on the high order scheme is gener-
ally applicable, but is here combined with an approach which reconstructs a
quadratic interpolant within each mesh cell by extending the stencil to include
the neighbouring cells’ nodes, averaging across cell edges to give continuity
(and hence conservation). This approach has also been successfully applied to
the two-dimensional inviscid Burgers’ equation, but it remains to extend it to
three-dimensional problems, higher than third order accuracy (both concep-
tually straightforward) and nonlinear systems of equations (not so). Ongoing
research has shown that this limiting procedure can also be used to con-
struct a fully consistent, positive, high order fluctuation distribution scheme
for time-dependent situations.
References
1. Abgrall, R.: Toward the ultimate conservative scheme: Following the quest. J.
Comput. Phys., 167, 277–315 (2001)
2. Abgrall, R., Barth, T.J.: Residual distribution schemes for conservation laws via
adaptive quadrature. SIAM J. Sci. Comput., 24, 732–769 (2002)
3. Abgrall, R., Mezine, M.: Construction of second order accurate monotone and
stable residual distributive schemes: the unsteady case. J. Comput. Phys., 188,
16–55 (2003)
4. Abgrall, R., Roe, P.L.: High order fluctuation schemes on triangular meshes. J.
Sci. Comput., 19, 3–36 (2003)
5. Caraeni, D., Fuchs, L.: Compact third-order multidimensional upwind scheme
for Navier-Stokes simulations. Theor. Comp. Fluid Dyn., 15, 373–401 (2002)
6. Deconinck, H., Sermeus, K., Abgrall, R.: Status of multidimensional upwind
residual distribution schemes and applications in aeronautics. In: AIAA paper
2000–2328 (2000)
7. Deconinck, H., Struijs, R., Bourgois, G., Roe, P.L.: High resolution shock cap-
turing cell vertex advection schemes for unstructured grids. In: Computational
Fluid Dynamics, VKI-LS 1994-05. Von Karman Institute, Brussels (1994)
8. Hubbard, M.E.: Non-oscillatory third order fluctuation splitting schemes for
steady scalar conservation laws. Submitted to J. Comput. Phys.
High-order residual distribution : discontinuity
capturing crosswind dissipation and diffusion
N. Villedieu-Ligout1 , M. Ricchiuto2 , and H. Deconinck1

1
von Karman Institute for Fluid Dynamics
nadege.villedieu@vki.ac.be, deconinck@vki.ac.be
2
INRIA Futurs, project ScAlApplix and MAB, Université de Bordeaux I
Mario.Ricchiuto@inria.fr
1 Generalities and notation
We review a class of compact methods to approximate steady solutions to

∂u
+ ∇ · F(u) = ∇ · (ν∇u) ∀(x, y) ∈ Ω (1)
∂t
on τh , an unstructured triangulation of the domain Ω. We make use of stan-
dard Lagrangian P k elements, that is the solution is approximated by a con-
tinuous piecewise k-th order polynomial. In every triangle T ∈ τh we construct
the sub-triangulation composed by k 2 triangles shown on figure 1. We denote
by Ts the generic sub-element.
g h
a b c
k 1
Fig. 1. P Lagrangian triangles with P conformal sub-triangulation
On τh , the discrete unknown uh is expressed as the following combination of

k-th Lagrangian polynomial basis functions :
X
uh (x, y) = ui ψi (x, y) (2)
i∈τh
where ui = u(xi , yi ) and ψi the i-th basis function. We call ϕi the piecewise
linear basis function associated to i ∈ T defined on the P 1 conformal sub-
triangulation (see figure 1). Clearly, in the P 1 case ϕi = ψi .

72 N. Villedieu-Ligout, M. Ricchiuto, and H. Deconinck
We consider schemes evolving the nodal values of the solution as :

X
un+1
i = uni + δi ΦTi s (3)
Ts ,i∈Ts
where δi is an iteration parameter. In the hyperbolic case ν = 0, the schemes

we consider are a particular case of the fluctuation or residual distribution
(RD) schemes of [5]. InZ particular, on every sub-element Ts we have
X
Ts
Φj = Φ = Ts
∇ · Fh (uh ) dx dy (4)
j∈Ts Ts
The paper is divided in two parts. In the first one we show the general
construction of k + 1-th order schemes for hyperbolic scalar conservation laws.
The extension to (1) is considered in the second part.
2 Quasi non-oscillatory RD for hyperbolic problems

First consider (1) in the hyperbolic case ν = 0. We review some RD schemes
satisfying all the following conditions:
Accuracy Condition to get k + 1-th order schemes is that (see [5] for details)
ΦTj s = O(hk+2 ) (5)

For the k-th degree polynomial approximation (2) we get Φ = O(h Ts k+2
),
hence the accuracy condition is also expressed by ΦTj s = O(ΦTs )
Upwinding In Ts , let ni be the inward normal to the edge facing node i
scaled by the length of the edge. Upwind schemes are the ones for which
1 ∂F(u∗ )
ki ≤ 0 ⇒ ΦTi s = 0 , with ki = · ni (6)
2 ∂u
with u∗ an arbitrary average of uh over Ts . Upwinding has a stabilizing
effect (see [5] for the analysis).
Monotonicity The rigorous definition of monotonicity for RD schemes re-
sorts to the theory of positive coefficients, see [5, 1] for details. In this
paper we will define a scheme as being monotone if, in practical com-
putations, it gives a non-oscillatory approximations of discontinuities. In
particular, we are interested in schemes for which, across a discontinuity,
ΦTj s × ΦM M
j ε0, for some first order monotone splitting Φj .
In the following subsections we review some definitions for the ΦTj s s.
2.1 Linear schemes
In this paper we make use of the following two upwind linear schemes
LDA scheme is the upwind scheme defined by
X
ΦTi s = ΦLDA
i = βiLDA ΦTs , βiLDA = ki+ / kj+ (7)
j∈Ts
Since βiLDA is uniformly bounded (w.r.t. mesh size h and solution uh ), the
LDA scheme respects the accuracy condition (5)
High-order monotone RDS and extension to advection-diffusion 73
N scheme is the upwind scheme defined by

X −1 X
ΦTi s = ΦN +
kj+ kj+ uj − ΦTs (8)

i = ki (ui − uin ) , uin =
j∈Ts j∈Ts
The N scheme is monotone (in the sense described in section 2) and first
order. On can easily show that the N is obtained by adding to the LDA
scheme a crosswind (shock capturing) dissipation term [1] :
X −1 X
ΦN
i = Φi
LDA
+ dNi , dN
i = kj+ ki+ kj+ (ui − uj ) (9)
j∈T j∈T
2.2 Nonlinear schemes

To combine high order of accuracy and monotonicity, we must define use a
nonlinear splitting. There are two ways of doing this :
Blending the LDA and N splittings in a way guaranteeing that the N scheme
is recovered only across shocks. For example :
|ΦTs |
ΦTi s = ΦB
i = θ ΦN
i + (1 − θ) ΦLDA
i , θ = P N
(10)
j∈Ts |Φj |
Due to (9), this is equivalent to add to the LDA scheme a residual shock
capturing crosswind dissipation term :
ΦB
i = Φi
LDA
+ θdN i (11)
Limiting the distribution coefficient of the N scheme
X N,+(PSI scheme) :
N,+
ΦTi s = ΦPSI
i = βi
PSI Ts
Φ , β P SI
i = βi / βj (12)
j∈Ts
Ts
with βjN = ΦNi /Φi . The PSI scheme verifies both the monotonicity re-
quirement (ΦTi s × ΦN PSI
i ε0), and the accuracy condition (5) (βi bounded).
2.3 Numerical examples

We show examples involving smooth and non-smooth solutions. First, on the
domain [−1, 1] × [0, 1], we consider the steady rotation (F = λu, λ = (y , −x))
of the inlet profile u0 = sin(16πx), defined on the boundary x ∈ [−1, 0], y = 0.
On the left on figure 2, we plot the outlet (x ∈ [0, 1], y = 0) data computed
by the second, third and fourth order PSI schemes. All the computations
have been run with the same number of degrees of freedom. The improvement
in the resolution of the high frequency profile brought by the higher order
polynomial representation is evident.
As a second example, the right picture on the same figure shows the grid
convergence rates obtained on P 3 elements for the constant advection (F =
λu, λ = (0 , 1)) of cos(πx) on the square [0, 1]2 . The high order schemes
(including the nonlinear ones) yield the expected fourth order of accuracy.
Lastly, figure 3 shows the results obtained on the Burger’s equation (F =
(u2 , 2u)/2) on the square [0, 1]2 with boundary conditions u = 1.5 − 2x for
y = 0. On the left we report the contours of the solution obtained with the
0
PSI(P2)
PSI (P3)
1.2 Exact
PSI (P1)
1 −5 b
0.8 1
0.6
−10
0.4
ln(error(u))
0.2
u
0 −15 a
-0.2
c
-0.4 1
−20
d
-0.6
e
-0.8
f
-1 −25
−5.5 −5 −4.5 −4 −3.5 −3 −2.5 −2
ln(h)
0 0.2 0.4 0.6 0.8
x
Fig. 2. Rotation of sin(16πx) and grid convergence (constant advection of cos(πx))
PSI (P3) schemes, while on the right a cut at y = 0.75 of the PSI (P2) and
PSI (P3) solutions. Some oscillations are present in very few mesh points after
the shock, however, their amplitude is small (below 15% local value of u).
3 Very high order RD and diffusion terms

We now consider the issue of including the diffusive terms into the discretiza-
tion. As shown in [4, 1], the main problem is to properly take into account the
relative magnitude of transport and diffusion terms, measured by the Peclet
number P e = hkλk/ν, with λ a local reference wave speed (flux Jacobian).
We briefly review the approach of [1].
The idea is to use RD only in advection dominated regions, to take ad-
vantage of its shock capturing. In diffusion dominated regions, on the other
1
p
1.6
0.9
1.4
PSI (P3)
0.8 Max : 1.75 1.2 PSI (P2)
Min : -0.76
0.7 1
0.6 0.8
0.5 0.6
u
y
0.4
0.4
0.2
0.3
0
0.2
-0.2
0.1
-0.4
0
0 0.5 1 0.7 0.8 0.9 1
x x
Fig. 3. Burgers equation : PSI (P3) solution contours and cut at y = 0.75 of the
PSI (P2) ans PSI (P3) solutions
High-order monotone RDS and extension to advection-diffusion 75
hand, Galerkin and stabilized Galerkin schemes perform very well. The prob-
lem is to build Petrov-Galerkin (PG) discretizations consistent with a given
RD scheme, and to combine the two discretizations to obtain uniformly (w.r.t.
h and P e) accurate approximations.
The solution proposed in [1] is the following. Given a RD scheme with dis-
tribution coefficients βjTs , such that ΦTj s = βjTs ΦTj s , build continuous piecewise
polynomialZ test functions respecting the consistency conditions
1
ωj dx dy = βjTs ∀ j ∈ T and ∀Ts ⊂ T (13)
|Ts |
Ts
There are several ways of choosing these functions, however, in a PG context,

the most natural way to do it is to define them as perturbations of some basis
functions. Here we consider the case in which the ωj s are defined as
ωj |Ts = ϕj + (3βjTs − 1)S Ts (14)
with ϕj piecewise linear basis functions on the conformal P 1 sub triangulation,
and S Ts are locally defined bubble functions. These perturbations are such that
ωj respects (13) (see [1] for details). Consider now the compact scheme
ΦC
i
zZ }| {
ΦTi s = ϕi λ · ∇u dx dy +(βiTs ΦTs − ΦC
h
i )+
Ts
Z Z
ν∇uh · ∇ϕi dx dy + (3βjTs − 1) ν∇uh · ∇S Ts dx dy (15)
Ts Ts
In the last definition, the first line represents the RD scheme3 , while the second
line contains the PG discretization of the diffusive terms. As pointed out in [1],
(15) does not introduce any coupling between the discrete advection operator
and diffusion operators. Defining the following discrete Peclet number
|ΦTs |
P eh = P R
| ν∇uh · ∇ϕi dx dy|
j∈Ts Ts
one can instead use the hybrid discretization :

Z
Ts
Φi = Φi + ν∇uh · ∇ϕi dx dy+
C
Ts
Z
ξ(P e h
)(βiTs ΦTs − ΦC
i ) + ξ(P e h
)(3βjTs − 1) ν∇uh · ∇S Ts dx dy (16)
Ts
with ξ(·) continuous, and such that limx→0 ξ(x) = 0 and limx→∞ ξ(x) = 1.
Scheme (16) reduces to a Galerkin type high order approximation in diffusion
3
the central finite element contribution ΦC
i is added and subtracted for reason
which will be soon clear
−6 −8
−10
−8
b
−12
−10
aa
c
−14
bb b
ln(error(u))
−12 cc
d
−16 a
−14
aa b
−18
−16
−20
h
q ln(error(u) = 3.18 ln(h) − 0.63
−18
Numerical error of u with LDA* −22 i
ln(error(u)) = 2.13 ln(h) −2.99
k
Numerical error of u with LDA
−20 −24
−5.5 −5 −4.5 −4 −3.5 −3 −2.5 −2 −6 −5.5 −5 −4.5 −4 −3.5 −3 −2.5 −2
Pe 0.3125 ln(h) 5 e f ln(error(u)) g
Fig. 4. Advection-diffusion : grid convergence on P 2 (left) and P 3 (right) elements
dominated regions, while the RD discretization is recovered in advection dom-

inated solutions. The relative magnitude of these phenomena are measured by
a residual based monitor, given by the discrete Peclet number P eh .
The effect of the introduction of the P eh scaling is shown on a practical
problem. We take λ = (0, 1), and the following boundary conditions :
p
1 − 1 + 16π 2 ν 2
u(x, 0) = − cos(2xπ) , u(x, 1) = − cos(2xπ) exp( )
p p 2ν
1 − 1 + 16π 2 ν 2 1 − 1 + 16π 2 ν 2
u(0, y) = − exp(y ) , u(1, y) = − exp(y )
2ν 2ν
We solve the problem on P 2 and P 3 elements. To enhance the effects of the
P eh scaling, we take ν = 10−2 in the P 2 case, and ν = 10− 3 in the P 3 one.
Figure 4 shows the convergence rates measured for a smooth problem (see
[1] for details on the set up). We use the LDA scheme with formulations (15)
and (16) (ξ = min(1, P eh )). Without a proper coupling between the advective
and diffusive operators there is an evident loss of accuracy. Conversely, scheme
(16) clearly yields optimal convergence rates.
References
1. Ricchiuto, M., Villedieu, N., Abgrall, R., Deconinck, H.: High-order residual
distribution schemes : discontinuity capturing crosswind dissipation and exten-
sion to advection-diffusion, VKI LS on Higher Order Discretization Methods for
Computational Physics, Von Karman Institute for Fluid Dynamics (2005)
2. Abgrall, R. : Very high order residual distribution methods, VKI LS-CFD, (2005)
3. De Palma, P. : Implicit third-order-accurate residal distribution schemes for
unsteady hyperbolic problems, VKI LS -CFD (2005)
4. Nishikawa, H. : Higher order discretization of diffusion terms in residual distri-
bution methods, VKI LS-CFD, (2005)
5. Abgrall, R., Roe, P.L.: High order fluctuation schemes on triangular meshes. J.
Sci. Comput., 19(3), 3–36 (2003)
High-Order Fluctuation-Splitting Schemes for
Advection-Diffusion Equations
Hiroaki Nishikawa and Philip Roe∗
Department of Aerospace Engineering,

University of Michigan, Ann Arbor, MI 48109-2140
hiroakin@engin.umich.edu, philroe@engin.umich.edu
1 Introduction
Motivated by the failure of current finite-volume(FV) codes in accurately

predicting hypersonic heat transfer rate on unstructured grids[1], we have
been studying schemes based on multidimensional upwinding, or fluctuation-
splitting(FS), as a radical alternative. This scheme is based on nodal variables
and cell residuals (fluctuations): the latter drive the change of the former in a
multidimensional fashion. The focus has been on its high-order extension with
the incorporation of viscous terms, and primarily in the simplified context of
two-dimensional advection-diffusion problems,
ut + a ux + b uy = ν (uxx + uyy ). (1)
In previous work, we found that a problem arises in regions where advection
and diffusion effects are equally important (such a region always exists in the
middle of a boundary layer): schemes for advection and diffusion cannot simply
be added, or the scheme reduces to only 1st-order accuracy [2]. To avoid this
and achieve high-order, we proposed to write the equation (1) as a first-order
system(FOS) introducing gradients as new variables, and developed uniformly
high-order FS schemes for the advection-diffusion equation. This approach
was then followed also in developing high-order FS diffusion schemes on P2
elements [3]. In this paper, we give a more detailed account for this strategy
based on a conventional design principle for high-order methods, and show
that the FOS approach has several advantages over a naive method.
2 Fluctuation-Splitting Schemes
Consider solving conservation laws of the form
ut + fx + gy = 0 (2)
∗
This work was supported under NASA URETI award NCC 3989 funded by NASA
Marshall and Glenn centers and by the USAF.

78 Hiroaki Nishikawa and Philip Roe
in a domain divided irregularly into a set of triangles {T }. Storing solutions

at nodes {j}, we begin by computing for all triangles T ∈ {T } the fluctuation
φT which is Za numerical approximation of the flux balance over the element
φT = − (fx + gy ) dxdy (3)
T
evaluated by quadrature. If this vanishes, we have the best possible solution
and take no action, but if it does not, we proceed on to distribute it, in a way
that reflects multidimensional physics (upwind for advection; isotropically for
diffusion), to the nodes. As a result, we obtain the following semi-discrete
equation at each node.
duj 1 X
= φTi ∀j ∈ {j} (4)
dt Mj
T ∈{Tj }
where {Tj } is a set of triangles that share node j, Mj is the median dual
cell area, and φTj = βjT φT : βjT is the distribution coefficient that assigns the
fraction of the fluctuation sent to node j in triangle T . If βjT is bounded, zero
fluctuation implies no updates in the solution: the scheme preserves polyno-
mial solution ( depends on the quadrature used) on arbitrary unstructured
grids; this is called the residual property and is part of the reason for the
reduced mesh sensitivity of FS schemes. In this work, we are interested only
in steady-state solutions and so we march in time simply using the forward
Euler time integration until solutions do not change.
For the advection-diffusion equation where (f, g) = (au − νux , bu − νuy ),
the fluctuation φT is the sum of two parts, advection φTa and diffusion φTd .
The former should be distributed with upwinding while the latter should be
distributed isotropically. But they should not be distributed separately, or
the residual property is lost and more importantly the accuracy of the scheme
reduces to 1st-order [2]. In this work, we therefore consider only the total
fluctuation, and distribute it with a combined upwind-isotropic distribution
coefficient proposed in [2] which becomes upwind (the LDA scheme) in the
advection limit and isotropic in the diffusion limit. This distribution coefficient
is bounded, and therefore the residual property is guaranteed: the accuracy
of the FS scheme is now determined by the accuracy of the fluctuation [4].
3 Second-Order Fluctuations
For conservation
Z laws, the fluctuation
I can be evaluated as a contour integral
− (fx + gy ) dxdy = − (f dy − g dx) (5)
T ∂T
This is convenient because only the solution variation along the element
boundary is relevant to the accuracy of Rthe fluctuation. Consider a single
2
term over an edge, say from node 1 to 2, 1 f dy. For second-order accuracy,
it suffices to use the trapezoidal rule
Z 2
f1 + f2
f dy = (y2 − y1 ) + O(h3 ) (6)
1 2
Fluctuation-Splitting Schemes for Advection-Diffusion 79
(u x) 2 , (u y) 2
2
um
T m
T
1
ui , pi , qi 3 (u x) 1 , (u y) 1
Fig. 1. DGE approach Fig. 2. FOS approach Fig. 3. High-Order
This formula is 2nd-order accurate provided the nodal fluxes f1 and f2 are
also second-order accurate. Note that for the advection-diffusion equation the
flux now involves the solution gradient ux which is not available at nodes and
must be evaluated with sufficient accuracy. In other words, we must follow
the principle that the quadrature formula and the fluxes at the quadrature
points must be of comparable accuracy, in the same spirit of other high-order
methods [5, 6, 7]. This has not been stressed so much in FS methods for
hyperbolic problems, where the solution values are stored at the quadrature
points, and the condition is automatically met, but becomes important for
problems such as advection-diffusion. The task becomes one of evaluating the
diffusive fluxes (gradients) with sufficient accuracy at the quadrature points.
In this paper we compare, theoretically and experimentally, two approaches.
One of these is interpolation, and the other is solution of a first-order system
having the gradient values as unknowns. The first of these is explicit, but the
second is implicit.
One approach is to directly obtain the gradients by reconstruction, for
example by a least-squares(LS) quadratic reconstruction which is 2nd-order
on general triangular grids. Then, the fluctuation will have a residual property,
preserving exact linear solutions. Note that although the diffusion part is exact
for quadratic functions but it is only 2nd-order because both the quadrature
and the recovered gradients are 2nd-order (being exact for quadratic does not
imply 3rd-order accuracy for terms involving 2nd-order derivatives!). This is
probably the simplest approach to achieve 2nd-order accuracy, but requires
the stencil to be enlarged (Figure 1), even more so for higher-order as at least
cubic reconstruction is needed beyond O(h2 ). In the rest of the paper, we refer
to this approach as the direct gradient evaluation(DGE) approach.
An alternative is the FOS approach in which we compute the nodal gra-
dients as additional unknowns, solving the equivalent first-order system
ut + fx + gy = 0, p − ux = 0, q − uy = 0 (7)
where (f, g) = (au − νp, bu − νq). Fluctuations are defined for the addi-
tional equations with piecewise linear variation of u, p, q, and then distributed
isotropically to update p, q; this minimizes those fluctuations in an L2 norm.
Values of p and q computed this way are used to evaluate the diffusive part of
the flux. This allows us to keep the stencil compact (Figure 2). The resulting
scheme has the residual property that exact linear solutions are preserved.
Observe that this FOS approach does not require any solution reconstruction
and the scheme remains compact.
Incidentally, it is possible and interesting to interpret the FOS approach
as a reconstruction method. In fact, the method is solving a globally coupled
system for p and q in an iterative manner. Interestingly enough, if we invoke
‘mass-lumping’ to remove the coupling, we end up with the familiar Green-
Gauss formula at every node. A similar observation is made in [8] where a
global system is derived by the Galerkin method. Certainly, the mass-lumping
yields a simple explicit formula for gradients, but in FS schemes we iterate
toward a steady-state anyway, and so we may as well iterate for gradients
along the way. Naturally, without mass-lumping, gradients are more accurate.
4 High-Order Fluctuations
For higher-order accuracy, we introduce a virtual node, m, at the midpoint
over the edge, and evaluate the fluctuation by Simpson’s rule (Figure 3),
Z 2
f1 + 4fm + f2
f dy = (y2 − y1 ) + O(h5 ) (8)
1 6
This integrates cubic polynomials exactly and therefore can be 4th-order accu-
rate. To match the accuracy, we now need 4th-order accurate fluxes at nodes
including the midpoint. This requires not only gradients at all points for the
diffusive flux but also the midpoint value um for the advective part of fm ; all
with 4th or at least 3rd-order accuracy to give a high-order scheme.
To estimate um with high-order, we follow Caraeni and Fuchs[9]. First
recover the gradients ((ux )i , (uy )i ) at nodes, and then evaluate um by the
Hermite cubic interpolation along the edge (Figure 3).
u1 + u2 1
um = − {(us )2 − (us )1 } (9)
2 8
where (us )i ≡ (ux )i (x2 − x1 ) + (uy )i (y2 − y1 ). This formula is 3rd-order with
quadratic gradient reconstruction and 4th-order with cubic reconstruction.
This completes the advective part of the high-order fluctuation.
For the diffusive part, the DGE approach requires cubic/quartic recon-
struction to ensure 3rd/4th-order accuracy. The reconstruction stencil will
then become very large, but can be reduced somewhat by utilizing the mid-
point values um already recovered for the advective part. For example, a node
shared by 4 triangles has now 12 neighbor nodes (instead of 4) which enables
cubic reconstruction. But this means that we must prepare two reconstruction
algorithms: one for estimating um in the advective flux; the other for evalu-
ating the diffusive flux. Also, we would need to store the midpoint values.
In contrast, the FOS approach can be extended to high-order without
such elaboration. All we need are high-order accurate gradient variables (p, q)
at nodes including the midpoints. The midpoint values can be estimated in
exactly the same way as um . Then, with midpoint values available, we can
Fluctuation-Splitting Schemes for Advection-Diffusion 81
define high-order fluctuations for the slope equations with piecewise quadratic
variation of u, p, q, and distribute these to compute p, q at the original
nodes [3]. Clearly, this scheme has the residual property: preserving exact
quadratic/cubic solutions with quadratic/cubic gradient recovery. Observe
now that essentially we need only a single algorithm that estimates midpoint
values via gradient recovery and the Hermite interpolation over edges, and we
only need to call this for 3 times to estimate um , pm and qm . We point out
also that this scheme can be made even simpler by substituting (9) into (8) to
eliminate the midpoint values, resulting an element-wise high-order correction
scheme [3], i.e. no need to store the midpoint values in practice.
5 Results
We show results for the advection-diffusion
√ equation with the exact solution
u = − cos(πη)exp(0.5ξ(1 − 1 + 4π 2 ν 2 )/ν) where ξ = ax + by, η = bx − ay,
ν = 0.1 and (a, b) = (7, 4) in the square domain [0, 1] × [0, 1]. This is a
case where the advective and the dffusive terms are equally important and a
scheme obtained by adding the Galerkin scheme to an advection scheme indeed
loses its accuracy. We tested the DGE schemes with quadratic/cubic/quartic
LS gradient reconstruction, and the FOS schemes with no reconstruction
and quadratic/cubic LS reconstruction, using the inverse distance-squared
weighting for all reconstructions[10]. We determined the order of convergence
through a series of computations using three different unstructured grids hav-
ing 441, 1681, and 6561 nodes. Figures 4 and 5 show the results (◦: 2nd; ∗: 3rd;
: 4th). First, 2nd, 3rd, 4th-order accuracy for both types of schemes were
confirmed. Second, the convergence of the 3rd-order schemes looks somewhat
better than expected. This could be because of the combination of the 4th-
order quadrature formula and 3rd-order fluxes. Finally, and remarkably, the
FOS schemes generate much lower error levels than the DGE schemes (note
that the figures are equally scaled); this may be because the gradients in the
flux are more accurate in the FOS approach as mentioned in Secion 2.
6 Concluding Remarks
The design principle of high-order methods equally applies to FS schemes: the
accuracy is determined by the quadrature formula and the fluxes at quadra-
ture points, and this is particularly important for the diffusive flux. With this
in mind, we compared two different strategies (FOS and DGE) for solving the
advection-diffusion equation, and conclude that the FOS approach demon-
strate several advantages over the DGE approach. First, the FOS schemes are
much more compact and accurate than the DGE schemes: no gradient recovery
required for 2nd-order, and quadratic/cubic reconstruction for 3rd/4th-order.
Second, high-order FOS schemes require only a single additional algorithm
that estimates midpoint values via gradient recovery (and this can be done
without storing the midpoint values[3]) whereas high-order DGE schemes add
-2 -2
2.0
-3 -3
Log(error)
Log(error) 1.9
-4 3.4 -4
3.3
-5 -5
3.9 3.8
-2 -1 0 -2 -1 0
Log(h) Log(h)
Fig. 4. DGE schemes Fig. 5. FOS schemes
much more complications to coding. Also, the FOS approach provides a solid
base for developing P2 -schemes (midpint values are computed as unknowns);
preliminary studies are reported in [3]. Finally, we remark that in the FOS
approach, only the gradients in the diffusive flux need to be stored (not gradi-
ents of all variables). In the Navier-Stokes equations, this means that we need
to store 3 viscous stresses and 2 heat fluxes in 2D, and 6 viscous stresses and
3 heat fluxes in 3D.
References
1. High Energy Flow Solver Synthesis (LAURA Benchmarks: Cylinder) at
http://hefss.larc.nasa.gov/
2. H. Nishikawa and P. L. Roe, On High-Order Fluctuation-Splitting Schemes for
Navier-Stokes Equations, in Third International Conference in CFD, Toronto,
2004.
3. H. Nishikawa, Higher-Order Discretization of Diffusion Terms in Residual-
Distribution Methods, in 34th VKI CFD Lecture Series Very-High Order Dis-
cretization Methods, 2005-01.
4. R. Abgrall, JCP, 167, pp. 277-315, 2001.
5. T. J. Barth and P. O. Frederickson, AIAA Paper 90-0013, 1990.
6. B. Cockburn and C. W. Shu, JCP, 141, pp. 199-224, 1998.
7. Y. Liu and M. Vinokur and Z. J. Wang, IJNMF, 216, pp.780-801, 2006.
8. R. Löhner, CFD Via Unstructured Grids: Trends and Applications, in Frontiers
of Computational Fluid Dynamics, John Wiley & Son Ltd., pp. 117-133, 1994.
9. D. Caraeni and L. Fuchs, Theoretical and Computational Fluid Dynamics, 15,
pp.373-401, 2002.
10. T. J. Barth, Aspects of Unstructured Grids and Finite-Volume Solvers for
the Euler and Navier-Stokes Equations, in Unstructured Grid Methods for
Advection-Dominated Flows, pp. 6-1-6-11, Neuilly Sur Seine France, 1992.
Construction of Higher Order Residual
Distribution Schemes
Rémi Abgrall1,2 and Cédric Tavé1,2

1
IMB, Université Bordeaux 1, 351 cours de la libération, 33400 Talence, France.
2
INRIA Futurs, projet ScAlApplix, Parc Club Orsay Université ZAC des vignes 4
rue Jacques Monod - Bât G 91893 Orsay Cedex France.
Summary. We present the construction of Residual Distribution (RD) schemes of

order higher than 2. TheRD schemes have several advantages: robustness, a com-
pact stencil and the residual property. Going to higher order can be easily done by
increasing the polynomial approximation in the cell. We focus here on a systematic
procedure to build higher order schemes from first order monotone ones. Assuming
we are given a low order monotone finite volume (rewritten in the RD framework),
or RD discretization, we show how to automatically generate a high order monotone
scheme. The formal order of accuracy is tuned by the degree of the local polynomial
interpolation of the numerical solution. This construction is done in three steps:
given a polynomial interpolation, compute the total residual and the first order RD
scheme, limit and stabilize (if needed). This is illustrated by an extension of the
Lax-Friedrichs scheme with P k approximation of the solution on triangles.
Key words: residual distribution, high order, Lax-Friedrichs scheme
1 Design principle and construction
1.1 Design Principle
We consider the numerical approximation of an hyperbolic system of conser-

vation laws of the form
(
∇·F =0 on Ω ⊂ Rn
(1)
boundary conditions on ∂Ω
and Th , a non-overlapping conformal triangulation of Ω. If E is a generic cell,
we define the cell-residual by
Z
ΦE = F · ndl (2)
∂E
A RD scheme is defined when are described

84 Rémi Abgrall and Cédric Tavé
• the evaluation of the cell-residual ΦEX

,
• a nodal distribution of the residual ΦE E
i = Φ : this is the conservation
i∈E
constraint,
ΦE
X
• a solution method for the problem i = 0.
E,i∈E
In all what follows, we assume that the solution is represented by a pth order
Lagrange continuous interpolant : ΦE is evaluated by an accurate enough
quadrature formula.
1.2 Design requirements to build a higher order scheme
Definition: A numerical RD scheme is said to be High Order Preserving

(HOP) when it satisfies the following condition:
ΦE → 0 ⇒ ΦE
i → 0.
In practical examples, we consider the distribution coefficients {βiE }i∈E , such

that βiE = ΦE E
i /Φ . A convenient way to formuate the HOP condition is to
ask for bounded distribution coefficients (βi )i∈E .
We recall here a result which is a direct extension of a result found in [1, 6].
To ensure that the scheme is pth order accurate, it has to verify the following
conditions:
• the mesh is regular,
• the approximation of the solution is pth order accurate for smooth solution,
• the cell residual is of O(hp+1 ) when we interpolate a smooth solution of
(1),
• the scheme must converge to a steady state.
2 Building a high order RD scheme from a first order

scheme
One of the most successful monotone first order RD scheme is the upwind N-
PSI scheme. However, many other finite difference and finite volume schemes
can be used as a building block for designing hih order residual distribution
scheme.
Here, we will use the Lax-Friedrichs’ scheme. First, this scheme has a
good property: it is monotone and it has an extremly low computational cost.
Moreover, for the purpose of demonstration, it is a very dissipative scheme
and we show that very good results can be obtained any way.
The elementary residuals are defined as
1 E
ΦE
X
i = {Φ + α (ui − uj )} (3)
d
i6=i
Construction of Higher Order RD Schemes 85
where d is the number of degrees of freedom in the cell (d = 3 for P 1 triangles,

d = 6 for P 2 triangles,...). This satisfies the conservation constraint.
2.1 Limitation step
Now, the idea is to find a mapping F that transform a low order distribution
{βiL }i∈E into a high order one {βiH }i∈E (we forget the super script E since
there is no ambiguity).
F : {βiL }i∈E → {βiH }i∈E
where the new coefficients must satisfy:

X
• βiH = 1 (conservation),
i∈E
• βiL .βiH ≥ 0 (monicity preserving),
• βiH uniformly bounded (HOP condition)
A possible definition of F is an extension of what Struijs has done (other
variant are given in [6] :
max(0, βiL )
βiH = F(βiL ) = X
max(0, βiL )
i
2.2 Stabilisation step
Finally, we now have to solve the non linear problem

ΦE
X
i =0 (4)
E,i∈E
To find the matrix structure of the approximation, let us rewrite the RD

schemes like this: ΦE
X
i = cij (u i − uj ), where cij ≥ 0 for positive scheme.
i6=j
We have to solve the problem :
AU = b (5)
X
where aii = cij and aij = −cij ( V(i) denoting the set of nodes that are
j∈V(i)
connected to i by an edge).
For monotone schemes, since cij ≥ 0, we clearly have:
(
aii ≥ 0
aij ≤ 0
Unfortunately, we cannot prove that A is strictly diagonal dominant and

most of the time the problem (5) is ill posed. Using an iterative procedure
to solve (5), we cannot reach a steady state in general, and this condition is
essential for accuracy.
We notice that we have this problem for non upwind schemes only. In fact,
the structure of the matrix A is not good when there exists a node i that does
not receive any part of the residual of the surrounding cells. In others words,
the scheme can locally be downwind, which is known to be unstable. The fact
that it does not blow up is related to the positivity of the scheme. Details can
be found in [1].
The solution presented here is an extension of what has been proposed in
[1]. The idea is to modify the scheme with a SUPG like term. Denoting by
βiH Φ the basic high order scheme, and by Φ∗i the stabilized one, we set:
Z
Φ∗i = βiH Φ + hθ (λ∇ψi )(λ∇u)
E
This additional term has the following properties (k being the order of the
polynomial approximation):
• it is streamline dissipative,
• it keeps the conservation property,
• it is O(hk+1 ),
The sensor may depend on the structure of the solution (the ideal case would
be θ ' 1 in the smooth part and θ ' 0 at the discontinuities. The quality of the
results is in practice unsensitive to the actual θ. The theoretical drawback of
this proecedure is that we lose the theorical positivity of the scheme. However,
in practical applications, no oscillation are observed.
3 Numerical Results
In this section, we present three different computations: two for scalar prob-
lems, one for the Euler equations. All the computations are done on triangular
unstructured grids. We denote by LF the low order Lax-Friedrichs scheme,
LLF the limited LF and LsLF the limited-stabilised LF .
• Figure 1: the scalar problem is the pure advection of a sin squared. On
the left, one can see that the LLF is unable to converge to a steady state.
With the LsLF we can reach the zero machine. The effect on the isolines
of the solution can be seen on the middle (LLF ) and the right (LsLF ).
• Figure 2: shows the advantage of going higher order for convection prob-
lems. The problem is a solid rotation of a sin. We compare here the use of
P 1 (left) and P 2 (right) interpolant (with the same number of points). We
can see the smooth and oscillatory signal is significally better represented
on the right.
Construction of Higher Order RD Schemes 87
• Figure 3: shows the effect of the limitation-stabilisation procedure. The

problem is a shock against a blunt body at Mach 2. The left part of the
figure is the result given by LF , the right one is the same computation
with the LsLF
• Figure 4: is a 4 states Riemann problem on a coarse grid (40x40) and
on a finer one (200x200). This is onfiguration number 3 of [5]. The Com-
putations are done with the 3rd LsLF scheme. As expected no spurious
oscillation are observed.
-2
-4
-6
-8
residual
-10
-12
-14
-16
stabilised scheme
without stabilisation
-18
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000
iterations
Fig. 1. Effect of the Stabilisation term on a constant advection: conver-

gence history (left), result without stabilisation (middle), result with stabilisation
(right)
LFLS_P1.desc LFLS_P2.desc
Fig. 2. Results on solid rotation: 2nd order scheme (left), 3nd order scheme
(right)
4 Future Work
The results of the LsLF scheme are very satisfactory. Without any tuning
parameters, it is able to handle difficult problems remaining stable and con-
vergent. Future extensions are planned: unsteady problems, 3D computations
and viscous flows.
choc_2D_O1.desc choc_2D_O2.desc
Fig. 3. Axisymmetric shock at Mach 2: Mach number, 1st order scheme (left),
2nd order scheme (right)
Fig. 4. 4-states Riemann problem: Density, 40x40 grid (left) and 200x200 grid
(right)
References
1. R. Abgrall: Essentially non oscillatory Residual Distribution schemes for hyper-
bolic problems, J. Comput. Phys., 2006, vol. 214
2. R. Abgrall, M. Mezine: Construction of second-order accurate monotone and sta-
ble residual distribution schemes for steady problems, JCP 195 (2004) 474-507
3. M. Ricchiuto, N. Villedieu, R. Abgrall and H. Deconinck: High-order residual
ditribution schemes: discontinuity capturing crosswind dissipation and extension
to advection diffusion, J. Comput. Applied math, 2006, accepted
4. R. Struijs: A multi-dimensional upwind discretization method for the Euler equa-
tions on unstructured grids, PhD, Von Karman Institute for Fluid Dynamics,
1994
5. A. Kurganov and E. Tadmor, Numer Meth in PDEs, vol 18, 2002.
6. R.Abgrall and P.L. Roe, High-order fluctuation schemes on triangular meshes, J.
Sci. Comput, vol 19, 2003.
Stable and convergent residual distribution for
time-dependent conservation laws
Mario Ricchiuto1 and Rémi Abgrall1,2

1
INRIA Futurs project Scalapplix and MAB Université de Bordeaux I
2
Institut Universitaire de France
Mario.Ricchiuto@inria.fr, Remi.Abgrall@math.u-bordeaux1.fr
1 Generalities
We consider the discretization of the time dependent hyperbolic problem
∂u
+ ∇ · F(u) = 0 on Ω × [0, tf ] ⊂ R2 × R+ (1)
∂t
on unstructured grids. We present residual distribution (RD) schemes which
(i) give non-oscillatory solutions, (ii) are second order accurate by construc-
tion, and (iii) lead to well-posed algebraic problems, that is, they ultimately
lead to linear systems A x = y, with A invertible. How to construct nonlinear
RD satisfying (i) and (ii) is known for some time [3]. However, it is the satis-
faction of (iii) that ensures that a (unique) discrete solution exists, and that
second order of accuracy is actually obtained in practice (convergence).
1.1 Residual distribution for time dependent problems

An abstract framework to picture the basics of RD is the following. Given Th ,
unstructured triangulation of Ω, and given (un , un−1 , . . . , u0 ), solution in the
mesh points at times (tn , tn−1 , . . . , 0), first discretize the time derivative :
p q
X αi X
δun+1−i + θj ∇ · Fn+1−j = 0
i=0
∆t n+1−i j=0
where δuk = uk − uk−1 , Fn+1−j = F(un+1−j ), and with ∆tk = tk − tk−1

the (variable) time step. The αj and θj coefficients may be associated to a
multistage method, as well as quadrature weights, in a space-time framework
[1, 7]. The only unknown being un+1 , we recast the problem as
α0
M(un+1 ) = S(un , un−1 , . . .), M(un+1 ) = un+1 + θ0 ∇ · Fn+1 (2)
∆tn+1
The nodal values of un+1 are computed through the following simple steps.
1. ∀ triangles T ∈ Th compute the element residual φT = T (Mh − Sh ),
R
where Mh and Sh are discrete

P approximations in space of M and S
2. define a splitting : φT = j∈T φTj (j ∈ T being the nodes of T )

90 Mario Ricchiuto and Rémi Abgrall,
n+1
3. compute
X the values ui by solving the (nonlinear) algebraic system
φi = 0, ∀ i ∈ Th
T
(3)
T |i∈T
System (3) is solved by means of an iterative procedure. The issue is how to

define the φTj s such that (3) is well posed, i.e. it admits a unique solution.
Since we are interested in numerical solutions with a non-oscillatory char-
acter, the discretization must have some degree of non linearity. For smooth
problems, most nonlinear RD are known to suffer from a lack of iterative con-
vergence. This is the symptom of a subtle instability, and it limits the overall
accuracy, due to the poor approximation of (3). For one class of schemes this
issue is analyzed in [2]. We proceed along the same lines.
2 Remarks on the iterative convergence of nonlinear RD

First we recall a procedure allowing to define the φTj s in a way ensuring by con-
struction that the scheme is formally second order accurate and monotonicity
preserving. At the basis of the construction are the following conditions3 :
Accuracy For a r-th degree polynomial approximation of the unknown (and
of the fluxes) a r + 1 order accurate scheme is obtained if [2, 8]
φj = βj φ , βj uniformly bounded
Monotonicity Given a linear monotonicity preserving first order scheme de-
fined by split residuals φM
j , we look for splittings verifying
φj = λj φM
j , λj ε0
We need to satisfy both conditions. One way to do this is to construct split
residuals φ∗j by applying to the φM
j s a uniformly bounded and sign preserving
nonlinear mapping, such as (in the scalar case, see [3] for systems) :
max(0, βiM ) φMj
φ∗i = βi∗ φ with βi∗ = P M
, and βjM = (4)
max(0, βj ) φ
j∈T
Due to boundedness of the βj∗ s defined by (4), and due to the properties of
the mapping, the resulting scheme is formally second order accurate, and it
does have a strong monotonicity preserving (viz. L∞ -stable) character [2, 3].
The whole procedure leads to a nonlinear algebraic system of type (3).
Let us suppose to be approximating a smooth solution, such that we can
linearize both the nonlinear part of (1) (or equivalently (2)), and also (3).
To this linearized version of (3) we associate the matrix engendered by the
component of the algebraic problem corresponding to the (linearized) operator
M(·) in (2). Let us denote by M∗h this matrix. Our (linearized) system reads
M∗h un+1 = Bh∗ (5)
where un+1 contains the unknown nodal values, and in Bh∗ we have dumped
whatever does not depend on un+1 .
3 T
for clarity, we drop the super-script
Stable and convergent RD for unsteady problems 91
We make the following remark. Denote by Mh the matrix obtained from

the first order monotone scheme, without the application of the mapping. Most
monotone schemes used as a basis for our construction are L∞ -stable and also
also L2 -stable : they can be shown to be dissipative (in an energy sense). As
a consequence, Mh is an irreducibly diagonally dominant invertible M-matrix
[5]. Thus, the existence of a unique numerical solution is guaranteed.
What can we say of M∗h ? By construction we know that for some λij ε0
∗
Mh ij = λij Mh ij (6)
∗ ∗
While ensuring that Mh ii ε0 ∀ i and Mh ij ≤ 0 ∀ i, j with i 6= j, this does
not guarantee the invertiblility of M∗h , unless we have some more information
on
ofits diagonal
P entries.
This is where the trouble comes from. Even though
| Mh ii | − j | Mh ij |ε0 ∀ i, we have no guarantee at all that
X X
λii | Mh ii | − λij | Mh ij | = λii Mh ii + λij Mh ij ε0
j j
It is even possible that λii = 0, for some i : the mapping, in general, weak-
ens the diagonally dominant character of Mh , eventually leading to a ill-
conditioned system matrix M∗h .
Another way to see it is the following. The whole construction is based on
∗
the constraint φM j × βj φε0. Upwinding is not included in the process : it is
likely that the application of the mapping might lead, locally in an element,
to a down-wind discretization, known to have poor stability.
Even so, the code never blows-up due to the L∞ −stable character of the
scheme. However, if a numerical output is obtained, the iterative convergence
in the solution of (3) is often poor, and the result might be affected by spurious
modes not identified/dumped (lack of dissipation/lack of uniqueness).
Let us now go back to the full nonlinear case. As mentioned before, the
convergence problems are relevant mainly when approximating smooth solu-
tions. When dealing with discontinuities the method has enough numerical
viscosity to converge relatively well, both in the inner iterations and with
mesh refinement. An heuristic justification of this fact is the following [2].
In elements containing a singularity, the element residual φT and the first
order monotone residuals φM j scale according to the same power of the mesh
size h. In particular, in two dimensions, simple arguments lead to
φT ≈ hk∆FkT and φM
j ≈ hk∆ukT
being k∆FkT and k∆ukT reference values for the norms of flux and variable
differences over T . The two scaling are easily obtained from the definition of
φT , and from the one of positive first order dissipation terms. If T contains a
singularity, what we can say is that k∆FkT and k∆ukT are bounded, which
j = O(h) and φ = O(h). Since φj = O(φ ), then φj /φj = O(1).
∗ ∗
leads to φM T T M
This means that, across a discontinuity, the mapping is likely to preserve more
the algebraic structure of the system obtained with the low order method.
3 Convergent schemes: a possible solution

We want to improve the properties of the matrix Mh∗ in (5). An idea comes
from the observation that our problem is not far away from the one encoun-
tered when discretizing (1) with a pure Galerkin scheme. The simplest way
to improve things could be then to add a streamline dissipation (SD) term.
For years this has been successfully used to stabilize Galerkin discretizations.
A recent analysis focusing on the time dependent case is given in [4].
Forgetting for the moment aboutZ shocks, we rewrite our discretization as
∂F(uh )
φi = φ∗i + φsi = βi∗ φT + h CT · ∇ϕi Mh − Sh dx

(7)
T ∂u
where the superscript s stands for stabilization, CT is a positive definite ma-
trix, and ϕi is the finite element (FE) shape function of node i. Numerical
experiments show that this modification solves the problem, the term added
introducing dissipation4 . Another way to see it is that the SD term introduces
some kind of upwind bias. The choice of CT seems irrelevant from the point of
view of the numerical results. This confirms that the algebraic nature of the
problem (full L2 /entropy-stability not required, even though still desirable...).
Since the evaluation of φsi in (7) is quite expensive, and since this term,
which destroys the monotonicity preserving character of the discretization, is
only needed in smooth areas, we modify the stabilization term as follows :
1 ∂F(u∗ )
φsi = βis φT , βis = Θ(uh , h)Ki , Ki = · ni (8)
2 ∂u
with θ(uh ) a solution monitor ensuring that the extra term is only active
in smooth regions, and u∗ an arbitrary average of u over T . The following
remarks can be made:
• φsi is a rough approximation of the SD term (7) (exact for linear problems)
• φsi still introduces some kind of upwind bias
This qualitatively explains the reason why this fix works : for smooth prob-
lems, the upwind bias improves the structure of M∗h in (5). The mechanism is
roughly the same guaranteeing the stability of the SD-Galerkin FE scheme.
In the case of nonlinear RD, it is rather difficult to formalize this with en-
ergy/entropy estimates, even though an analysis similar to the one made in
[4] is possible (see also [2]).
Lastly, different definitions can be used for θ(uh ). Here, we take the sim-
plest possible [2] :
h |T | kukT im
Θ = C0T min 1, , C0T positive definite (9)
kφh k
Provided that C0T is of O(h−1 ), across singularities, this definition leads to
φsi = O(hm ) × O(φT ) [2]. Thanks to this, the amplitude of any oscillations
appear in the solution decreases as some power of h.
4
even though, to be rigorous, we should evaluate the extra term in entropy variables
Stable and convergent RD for unsteady problems 93
c c c
a a a
b b b
line01 line01
101.662 101.662
a c
a
99.2033 99.2033
96.7448 96.7448
p
p
94.2863 94.2863
b b
a
91.8278 91.8278
0 0.25 0.5 0.75 1 0 0.25 0.5 0.75 1
b
x x
Fig. 1. Vortex advection. Top (left to right): Exact, LLF and LLFs (h = 1/40).
Bottom (left to right): LLFs (h = 1/80), cut along centerline for LLF and LLFs
4 Computational results
We discuss some results obtained solving the Euler equations with the nonlin-
ear discretization constructed starting from the Lax-Frederich’s scheme com-
bined with Crank-Nicholson time integration :
|T | δun+1 ψin + ψin+1
Z
1 X
φLF
i = + , ψ i = ∇·F dx+αT (ui −uj )(10)
3 ∆tn+1 2 3
T j∈T
We call LLF scheme the one obtained by applying (4) to (10). The scheme
obtained by adding the stabilization term (8) is referred to as LLFs.
We first consider a smooth problem, consisting of the advection of a vortex
in inviscid flow (see [6] for details). On fig. 1 we report, on the top row, the
exact solution (pressure contours), the solution of the LLF scheme and of the
LLFs scheme on a coarse grid (h = 1/40). On the bottom row we report
the solution of the LLFs scheme on the finer mesh (h = 1/80), and the 1D
pressure distributions through the vortex core. The top row pictures clearly
show the spurious modes not dumped by the LLF scheme, and the effectiveness
of stabilization in suppressing these modes. The contour plot on the bottom
row, as well as the 1D line plots, confirm this observation and demonstrate
the truly second order of convergence of the stabilized scheme (as shown by
the reduction in the L∞ pressure error).
Then on fig. 2 we report the results on the well known problem of the
Mach 3 wind tunnel with a forward facing step [10]. The result show the
monotonicity preserving character of the scheme, and the effectiveness of the
definition of the solution monitor (9).
5.50353
4.30635
3.10917
1.91199
0.714813
0 0.75 1.5 2.25 3
Fig. 2. Mach 3 forward facing step. Density contours (left) and distribution at
y = 0.5 (right). Results at time t = 4 computed with the LLFs scheme
5 Conclusions
We discussed the construction of non-oscillatory RD schemes which are stable
and give genuine second order of accuracy in practical applications. The ap-
proach proposed opens the way to a new class of schemes which need very few
matrix operations. Hence, they are more efficient, while retaining the advan-
tages of the RD approach. In particular, the extension to arbitrary accuracy
is quite natural and it is under way.
References
1. Abgrall, R., Andrianov, N., Mezine, M.: Towards very high-order accurate
schemes for unsteady convection problems on unstructured meshes. Int. J. Nu-
mer. Meth. Fluids, 47, 679–691 (2005)
2. Abgrall, R.: Essentially non-oscillatory Residual Distribution schemes for hy-
perbolic problems. J. Comput. Phys., 214, 773–808 (2006)
3. Abgrall, R., Mezine, M.: Construction of second-order accurate monotone and
stable Residual Distribution schemes for steady flow problems. J. Comput.
Phys., 188, 16–55 (2003)
4. Bochev, P.B., Gunzburger, M.D., Shadid, J.N.: Stability of the SUPG finite
element method for transient advection-diffusion problems. Comp. Meth. Appl.
Mech. Engrg., 193, 2301–2323 (2004)
5. Berman, A., Plemmons, R.J.: Nonnegative Matrices in the Mathematical Sci-
ences. Academic Press, 1979
6. Dobevs, J., Deconinck, H.: Second Order Blended Multidimensional Residual
Distribution Scheme for Steady and Unsteady Computations. J. Comput. Appl.
Math., accepted (2006)
7. Ricchiuto, M., Abgrall, R., Deconinck, H.: Construction of very high order
Residual Distribution schemes for unsteady advection: preliminary results. VKI
LS 2003-05 (2005)
8. Ricchiuto, M., Abgrall, R., Deconinck, H.: Application of conservative
residual distribution schemes to the solution of the shallow water equa-
tions on unstructured meshes. J. Comput. Phys., (2006). Available online.
doi:10.1016/j.jcp.2006.06.024.
9. Ricchiuto, M., Csı́k, Á., Deconinck, H.: Residual Distribution for general time
dependent conservation laws. J. Comput. Phys., 209, 249–289 (2005)
10. Woodward, P.R., Colella, P.: The numerical simulation of two-dimensional flows
with strong shocks. J. Comput. Phys., 54, 115–173 (1984)
An ALE Formulation of the Multidimensional
Residual Distribution Scheme for
Computations on Moving Meshes∗
Jirı́ Dobes1,2 and Herman Deconinck2

1
Department of Technical Mathematics, Faculty of Mechanical Engineering,
Czech Technical University, Karlovo Nám. 13, CZ-121 35 Prague, Czech Republic
Jiri.Dobes@fs.cvut.cz
2
Von Karman Institute, Waterloosesteenweg 72, B-1640 Sint–Genesius–Rode,
Belgium dobes@vki.ac.be, deconinck@vki.ac.be
1 Introduction
The Residual Distribution (RD) schemes for unsteady computations are sub-
ject of on-going intensive research [1]. Until recently, the RD schemes for
unsteady transonic flow simulations had modest success due to difficulties to
deal with large time-steps. As the solution, a new Bx scheme was introduced
in [3]. The scheme is based on a blend of the non-oscillatory first order lumped
mass matrix formulation of the N scheme and the second order full mass ma-
trix formulation of the LDA scheme using a simple pressure switch. In smooth
parts of the flow linear stability is sufficient and the LDA scheme is active. In
shocks, the N scheme is activated to prevent oscillations in the solution. The
blending parameter is smooth, giving superior iterative convergence proper-
ties and scaled such that second order of accuracy is achieved. The topic of
this paper is the extension of the Bx scheme for the computation on moving
meshes with application to transonic flutter computations.
2 ALE Formulation
We define the ALE mapping which for each t ∈ I associates a point Y of ref-
erence configuration Ω0 to a point x on the current domain configuration Ωt ,
At : Ω0 ⊂ Rd → Ωt ⊂ Rd , x(Y, t) = At (Y). The ALE mapping At is chosen
sufficiently smooth and invertible with nonzero determinant of Jacobian JAt .
A domain velocity w(x, t) is defined as the time derivative of x for constant
Y. We start from the conservative ALE formulation of the Euler equations in
d spatial dimensions
∗
This research was partially supported by Research plan of MSMT no. 6840770003.

96 Jirı́ Dobes and Herman Deconinck

1 ∂JAt u
+ ∇x · [f (u) − uw] = 0 (+I.C. and B.C.) (1)
JAt ∂t Y
where u = (ρ, ρvi , E)T is the vector of conserved variables and f (u) the well
known vector of flux functions. The system is closed with the equation for a
perfect gas. The following equality, called geometrical conservation law, will
be used later

1 ∂JAt
∇x · w = . (2)
JA ∂t t Y
3 Numerical Scheme
The problem is solved on mesh τ h consisting of simplex elements {E}. The
class of unsteady RD schemes can be formulated as a relaxation procedure in
dual time. The solution at time level n + 1 is obtained for m → ∞
X
un+1,m+1
i = un+1,m
i − σi φE
i , (3)
E∈Di
where σi is a relaxation coefficient and φE i is the nodal contribution from

element E given by the particular scheme.
The RD schemes operate on the quasilinear form of equation (1) rewritten
as follows, with use of (2)

1 ∂JAt u 1 ∂JAt ∂f
+ A · ∇ x u − u = 0, A= − wI. (4)
JA ∂t
t Y JA ∂t
t Y ∂u
The last term in the left equation is the geometrical source term. Note that
the Jacobian A includes also the mesh velocity. We define an upwind matrix
associated with node i, kiE = A · ni /d, where ni is the normal perpendicular
to the face opposite to node i, scaled by its surface. As usual in this context,
symbol ki+ denotes matrix ki with all the negative eigenvalues set to zero.
We begin with the first order (mass lumped) N scheme [1] applied to
equation (4), with three points backward (3BDF) time integration and with
the geometrical source term discretized on the dual mesh. The scheme is given
by (3) with
αn+1 µ(E n+1 )un+1 + αn µ(E n )uni + αn−1 µ(E n−1 )un−1
φE,N
i = i i
+
d+1
n+1 αn+1 µ(E n+1 ) + αn µ(E n ) + αn−1 µ(E n−1 )
∆t ki+ (un+1 − ūn+1

i − uin ) ,
d+1
(5)
where µ(E) is the volume of element E and

P n+1
j∈E uj
X X
ū n+1
= , uin = −N kj− uj , N =( ki+ )−1 , (6)
d+1
j∈E i∈E
A Residual Distribution scheme for Computations on Moving Meshes 97
with coefficients
1 + 2τ τ2 tn+1 − tn
αn+1 = , αn = −1 − τ, αn−1 = , τ= n . (7)
1+τ 1+τ t − tn−1
Each of the terms of (5) corresponds to the respective term given in (4). This
scheme is non-oscillatory, but only first order accurate.
We next consider the linear second order LDA scheme, derived from the
analogy with the Petrov-Galerkin (PG) finite element method (FEM) for
scalar problems. The solution and the mesh velocity are approximated by the
linear Galerkin shape functions ψi from the current domain configuration [4].
The equation is multiplied by the PG test function ϕE i = ψi +βi
LDA
−1/(d+1)
from the current domain configuration, with distribution coefficient βiLDA =
ki+ N , then it is integrated over the domain. The time derivative is replaced
by the 3BDF scheme. The scheme is again written as (3), with the element
contribution
X
φE,LDA
i = (αn+1 µ(E n+1 )un+1
j + αn µ(E n )unj + αn−1 µ(E n−1 )ujn−1 )m̃E
ij +
j∈E
 n+1
X X
βiLDA kj uj  −(αn+1 µ(E n+1 )+αn µ(E n )+αn−1 µ(E n−1 )) un+1
j m̃E
ij ,
j j∈E
(8)
where m̃E
R
ij = E ϕi ψj dx/µ(E) is the element contribution to the mass matrix
divided by the volume of E. The scheme for scalar problems is extended to
the system of equations simply by replacing the distribution coefficient by
the distribution matrix βjLDA with formally the same definition. The element
contribution to the mass matrix becomes

E 1 1 + δij 1 LDA
m̃ij = − I + βj . (9)
d+1 d+2 d+1
The upwind matrices ki are evaluated in a suitably chosen mean state to
assure conservativity of the scheme [7]. Note, that unlike some other ALE FE
schemes, this scheme does not contain a time derivative of the mass matrix.
Both N and LDA schemes respect the discrete Geometric Conservation Law
by construction.
Finally, the nonlinear Bx scheme takes advantage of the monotone behav-
ior of the N scheme in the shocks and the high accuracy of the LDA with
the mass matrix in the smooth regions of flow. The nodal contribution of the
Bx scheme is given by φBx N
i = θφi + (1 − θ)φi
LDA
. The element-wise blending
coefficient is defined as [3] θ = min(1, sc h), sc = ( ∂p
2 +
∂t + ∇x p · v) /δpv , where
v is the velocity vector of the flow, p is the static pressure, h the diameter of
the element and δpv is a product of the characteristic pressure and velocity
in the domain. The blending coefficient θ is O(h) in the smooth parts of the
flow, giving second order of accuracy, and O(1) in shocks, effectively switching
back to the non-oscillatory N scheme.
4 Numerical Results
For the comparison, the cell centered finite volume (FV) scheme with linear
reconstruction, Roe’s Riemann solver and Barth’s limiter is considered [2].
The time is discretized with the 3BDF scheme and the mesh movement is
treated using the Scheme A [6].
The first test case involves the well known test case AGARD CT 5 of the
NACA 0012 airfoil, which is sinusoidally pitching around its a quarter chord.
The free stream Mach number is 0.755. The airfoil performs a pitching motion
α = 2.51◦ sin(2kt) + 0.016◦ , k = ωc/2u∞ = 0.0814, where c is the chord, u∞
is the free-stream velocity and ω the frequency. The problem is solved on
an unstructured mesh consisting of 5711 nodes and 11153 elements with 206
nodes around the airfoil.
0.4
0.3
cl
0.2
0.1
! -2 -1 1
-0.1
-0.2
-0.3
-0 4
Fig. 1. NACA 0012 test case. Continuous line is the Bx scheme, discountinous line
is the FV scheme. Left and middle: isolines of the pressure at time t = 116. Right:
dependence of the lift coefficient on the angle of incidence.
The solution at time t = 116 is plotted in Fig. 1. The FV solution is more

dissipative, as one can notice above the profile, where the RD isolines are more
crisp and running straight into the shock. The RD scheme gives a higher peak
of the lift in comparison with the FV scheme, which also points to the higher
accuracy of the Bx scheme.
Test-cases two and three involve coupling with an elastic model considering
a linear anisotropic continuum. The governing equations are solved by means
of a standard finite element method for the displacement with bi-quadratic
Lagrangian elements. Newmark method is used for the time discretization.
The similar method is used for the mesh movement and a simple sub-iteration
approach couples the problems together.
As the second test, the supersonic panel flutter is considered [8]. An elastic
panel with infinite aspect ratio is clamped on both edges. Its upper side is
exposed to the supersonic airstream, while the lower side resides in the still air
with the same pressure as on the upper side. The panel has length L = 0.5 m, a
uniform thickness h = 1.25 × 10−3 m, Young modulus E = 7.728 × 1010 N/m2 ,
Poisson ration ν = 0.33 and the density ρs = 2710 kg/m3 . The plane strain
A Residual Distribution scheme for Computations on Moving Meshes 99
LDA FV
1.5
FV fine Bx
0.5 FV
0.008
!z’
0.006
0.004 0
0.002
0 -0.5
!y
-0.002
-0.004
-1
-0.006 N
LDA
-0.008
0 0.02 0.04 0.06 0.08 0.1
-0.01 t
0 0.05 0.1 0.15 0.2 0.25 0.3
t
Fig. 2. Left: Panel flutter problem. Dependence of the integral of deflection on

time for different numerical schemes. Ma ∞ = 2.2. Right: AGARD 445.6 wing.
assumption was used. The flow conditions are given by p∞ = 25714 Pa and
ρ∞ = 0.4 kg/m3 . The critical Mach number Ma cr ∞ that is, the lowest free
stream Mach number for which an unstable aero-elastic mode of the panel
appears, is given in reference [8]. Using a theoretical method the authors
get Ma cr cr
∞ ≈ 2.27 and using their numerical scheme Ma ∞ ≈ 2.23, what they
consider an “excellent agreement”.
The elastic panel is discretized with 60 × 2 elements. The computational
domain is formed by a half-circle of radius R = 5. The mesh consists of
3451 nodes and 6722 triangular elements, giving 50 elements along the panel.
Reference FV computation is performed on a regular quadrilateral mesh of
300 × 100 elements with 100 elements along the panel (referred as “fine”).
The integral of the deflection of the panel for Ma = 2.2 is plotted in figure 2.
The neutral response was correctly reproduced. Although the LDA scheme is
only linearly stable, it is able to capture weak shock waves in a non-oscillating
manner. The nonlinear Bx scheme gives similar results as the LDA scheme,
which are very different from the first order N scheme. Finally, the transonic
flutter of the AGARD 455.6 wing (“solid model”) is considered [5]. The elastic
wing was discretized using 350 tri-quadratic elements. The CFD mesh consists
of 22k nodes and 118k tetrahedral elements. The neutral response regime was
chosen, which is characterized by the flutter speed index of 0.5214 and the free
stream Mach number Ma ∞ = 0.92 with flow medium Freon-12. One period
was divided in 120 time-steps. The integral of the wing velocity is plotted in
Fig. 2. For the FV scheme, a small negative damping is observed, while by

the LDA scheme the neutral response was correctly reproduced. The difference
between the measured oscillation period and the computed period is 0.66 %
for the LDA scheme and 2.08 % for the FV scheme. We judge this result more
accurate than one can expect, given the uncertainties in the wing model.
5 Conclusions
A novel ALE formulation of the residual distribution schemes for the compu-
tation on moving meshes is presented. The nonlinear Bx scheme consists of a
blend of the LDA and N schemes. The LDA scheme is formulated with finite
element type mass matrix. The N scheme is a simple extension of the steady
version of the scheme. Results are compared with a FV method. The method
is validated on forced oscillating airfoil, where it shows clearly better accuracy
than the FV scheme, and on two typical cases of aeroelastic problems.
References
1. Abgrall R., Mezine M.: Construction of second order accurate monotone and
stable residual distribution schemes for unsteady flow problems. JCP, 188:16–
55 (2003).
2. Barth T.J.: Aspects of unstructured grids and finite-volume solvers for the Euler
and Navier-Stokes equations. In: 25th CFD VKI LS (1994).
3. Dobevs J., Deconinck H.: Second order blended multidimensional residual dis-
tribution scheme for steady and unsteady computations. J. of Comp. and Appl.
Math. (JCAM) (2005), accepted for publication.
4. Dobevs J., Deconinck H.: Extension of linear second order residual distribution
schemes for computations on moving meshes. In: J. Pvrı́hoda, K. Kozel (editors),
Proceedings of “Topical Problems of Fluid Mechanics 2006”, 27–30, IT CAS CZ
(2006), ISBN 80-85918-98-6.
5. E. Carson Yates J.: AGARD standard aeroelastic configurations for dynamic
response. Technical Memorandum 100492, NASA (1987).
6. Koobus B., Farhat C.: Second-order time-accurate and geometrically conserva-
tive implicit schemes for flow computations on unstructured dynamic meshes.
Comput. Methods Appl. Mech. Engrg., 170(1–2):103–129 (1999).
7. Michler C., Sterck H.D., Deconinck H.: An arbitrary Lagrangian Eulerian formu-
lation for residual distribution schemes on moving grids. Computers and Fluids,
32(1):59–71 (2003).
8. S. Piperno, Farhat C.: Partitioned procedures for the transient solution of cou-
pled aeroelastic problems part II: Energy transfer analysis and three-dimensional
applications. Comput. Meths. Appl. Mech. Engrg., 190(24):3147–3170 (2001).
Solution of the steady Euler equations using
Fluctuation Splitting schemes on quadrilateral
elements
D. T. Rubino1 , P. De Palma2 , G. Pascazio2 , and M. Napolitano2

1
GE Nuovo Pignone, via Matteucci 2, 50127, Firenze, Italy
2
DIMeG, Politecnico di Bari, via Re David 200, 70125, Bari, Italy
dantetommaso.rubino@ge.com, depalma@poliba.it, pascazio@poliba.it,
napolita@poliba.it
1 Introduction
Fluctuation Splitting (FS) schemes, also known as Residual Distribution
schemes, are a well-established technique for solving hyperbolic problems on
triangular meshes, see [1] and the references therein. As such, they are not
adequate for solving high Reynolds number viscous flows, where an accu-
rate resolution of the boundary layer requires high aspect ratio quadrilateral
cells. Therefore, there has been a renewed interest in the application of this
methodology on quadrilateral meshes. In [2] the authors presented a thor-
ough analysis of the properties of FS schemes when applied to quadrilateral
elements, with reference to the solution of the linear advection and advection-
diffusion equations. Using a multidimensional Fourier analysis, it is shown
that linearity preserving (LP) schemes lack dissipation for the high wavenum-
ber modes, which may hamper convergence to the steady-state solution, or
even lead to instability problems. A remedy is employed, especially needed
for convection dominated problems, namely, adding an SUPG based artificial
dissipation term to the LP signals, which is effective in stabilizing high wave-
number modes, even if it does not ensure monotonicity. This work extends the
approach of [2] to the solution of the Euler equations and applies it with suc-
cess to well documented subsonic, transonic, and supersonic two-dimensional
inviscid flows.
2 Numerical method
The Euler equations in quasi-linear form are written as:
∂Q ∂Q ∂Q ∂Q
+ A · ∇Q = +A +B = 0, (1)
∂t ∂t ∂x ∂y

102 D. T. Rubino, P. De Palma, G. Pascazio, and M. Napolitano
where Q is the vector of the conservative variables and A = ∂F/∂Q and

B = ∂G/∂Q are the Jacobian matrices of the fluxes in the x and y directions,
respectively. Multidimensional upwinding is obtained by defining the upwind
(or inflow ) matrices Kj as:
1 1
Kj = A · nj =

A nxj + B nyj ; (2)
2 2
due to the hyperbolic nature of the system, they have a complete set of real
eigenvalues and eigenvectors, and can be expressed as:
K = RΛL, (3)
where the columns of R contain the right eigenvectors, Λ is the diagonal
matrix of eigenvalues and L = R−1 . The matrices K+ and K− are defined as:
K+ = R Λ+ L , K− = R Λ− L , (4)
+ −
where Λ and Λ contain the positive and negative eigenvalues, respectively,
Λ± = 1/2 (Λ ± |Λ|). The fluctuation over an element is computed as:
Z Z I
E
Φ = ∇· F dΩ = A · ∇Q dΩ = F · n d` . (5)
ΩE ΩE ∂ΩE
Portions of the fluctuations (signals) are distributed to the nodes according

to the distribution matrices Γ, and the signal sent from element E to node i
is defined as:
ΦE E E
i = Γi Φ . (6)
Finally, the semidiscrete equation is obtained collecting the signals sent to
each node from the surrounding elements:
dQ ∆t X E E
=− Γi Φ . (7)
dt |Ωi |
i∈∆i
Conservation is guaranteed provided that the fluctuation is computed by the

contour integration of the fluxes, namely, the last term of equation (5). The
following three FS schemes are of interest for the present work. The matrix N
scheme is used in its conservative formulation [3]:
 −1  
X X
+
ΦN
i = Ki (Qi − Qc ) , with Qc =  K+j
  K+j Qj − Φ
E
.(8)
j∈E j∈E
The matrix LDA scheme is defined as:

 −1
X
ΦLDA
i = ΓLDA
i ΦE = K+ i
 K+j
 ΦE . (9)
j∈E
The matrix LW scheme is defined as:

FS on quadrilateral elements 103

1 ∆t
ΦLW
i = ΓLW
i ΦE = I+ Ki ΦE . (10)
4 2 |ΩE |
In the formulas above, the inflow matrices are evaluated in their linearized
state, computed using the arithmetic averages of the primitive variables. The
last two schemes are linearity preserving (LP) [4] and thus second-order ac-
curate for pure advection. To ensure monotonicity, the non-linear flux redis-
tribution (FR) technique developed and presented in detail in [5] has been
employed. Similarly to the FCT approach, this technique uses the solution
obtained with a low-order monotone scheme (the N scheme in this case) to
limit the high order solution.
The use of a properly scaled SUPG-based artificial dissipation term suitable
to stabilize high wavenumber modes has been investigated. This term is added
to the standard LP signal as:
Φ?,E
i = ΦE AD
i + Φi . (11)
It is noteworthy that a similar approach has been employed by Abgrall and
Marpeau [6], to improve the convergence in the iterative procedure of the
limited N scheme. Similarly, the artificial dissipation term is here formulated
as:
Z
∂Ni ∂Ni
ΦAD
i = τ (h) A + B ∇· F dΩ , with τ (h) = h . (12)
ΩE ∂x ∂y
The numerical computation of this term can be simplified greatly, avoiding the
numerical integration at each step. This is achieved using linearized Jacobians
and taking the fluctuation out of the integral (as if ∇· F was constant over
the element E), as follows:
Z
∂Ni ∂Ni
ΦAD
i = h A + B ∇· F dΩ
ΩE ∂x ∂y
ΦE
Z
∂Ni ∂Ni
=h A +B dΩ
ΩE ∂x ∂y |ΩE |
Z Z
h ∂Ni ∂Ni
= Â dΩ + B̂ dΩ ΦE . (13)
|ΩE | ΩE ∂x ΩE ∂y
This approach is very advantageous from the computational point of view

since the integrals of the shape function gradients are computed once at the
beginning of the simulation and then stored in global arrays. Moreover, con-
servation is guaranteed by construction due to the properties of the shape
functions and the error introduced using averaged Jacobians and fluctuation
is not relevant, since this is only an artificial dissipation term.
3 Results
The first test-case considered is the well-known supersonic jet interaction test-
case. It consists in the interaction between two horizontal, supersonic jets,
which are suddenly brought into contact. The upper stream is a high Mach
number, low pressure flow, while the lower stream is a lower Mach number and
higher pressure flow. The problem is solved on the square domain [0, 1]×[0, 1],
with the prescribed inlet boundary conditions:
M = 4, v = 0 , p = 0.25 , ρ = γ/2 for x = 0, y ∈ [1/2, 1] ,
M = 2.4 , v = 0 , p = 1 , ρ=γ for x = 0, y ∈ [0, 1/2] .
Computations have been performed on an 80 × 80 Cartesian grid and on an
unstructured grid of triangles with spacing h = 1/64 on boundaries. Figure 1
(a) quads: N (b) quads: LDA -N (c) triangles: LDA -N
Fig. 1. Jets interaction test-case: density contours (∆ρ = 0.05).
shows the density iso-lines obtained using the monotone N scheme on the
quadrilateral grid and the limited LDA scheme on both the quadrilateral grid
and the triangular one. The results show that the limited scheme resolves all
waves, namely, the shock, contact wave and expansion fan, much better than
the monotone N scheme; however, the residual is not fully converged because
of the limiter.
Then, the steady subsonic flow inside a channel with a cosine-shaped bump
on the bottom wall has been considered [7]. The channel has dimensions
[0, 4] × [0, 1] with a straight upper wall and a cosine shaped lower wall with
coordinates:
0 ≤ x ≤ 1 y = 0,
1 ≤ x ≤ 3 y = (1 − cos[(x − 1) π])/10 ,
3 ≤ x ≤ 4 y = 0.
The isentropic exit Mach number is equal to M∞ = 0.5. At inlet points, the
total pressure, total enthalpy and flow angle are imposed, whereas the static
FS on quadrilateral elements 105
pressure is imposed at the outlet points. Computations have been performed

on a 128×32 structured grid of quadrilateral elements and, for comparison, on
an unstructured triangular grid with 9977 elements, using the LW scheme. The
L1 norm of the entropy error, Σ = [p/ργ − (p/ργ )inlet ] /(p/ργ )inlet , is reported
in table 1, together with its maximum and minimum values. It can be seen
that the LW scheme on quadrilaterals is far less dissipative (less production
of spurious entropy, especially near the bottom wall) than the same scheme
on triangles. Accordingly, some entropy oscillations occur, which appear to
be slightly reduced using the SUPG based dissipation term. As a final test
scheme mesh type Σmax (×10−5 ) Σmin (×10−5 ) kΣk1 (×10−5 )

LW triangular 24.24 -27.47 4.172
LW quadrilateral 9.072 -15.28 2.821
LW + AD quadrilateral 1.180 -7.242 2.719
Table 1. Spurious entropy production: error maximum and minimum levels and L1
norm.
to validate the method, the transonic inviscid flow through a cascade of gas
turbine rotor blades (VKI LS–59) has been considered [8]. The blade profile
and the grid used are shown in figure 2a, together with a close-up view of the
leading and trailing edges. The grid is an hybrid multiblock one. An “O” type
structured block (561 × 17 elements) surrounds the blade profile and an un-
structured block made up of 5929 triangles covers the outer domain. Thus, the
(a) geometry and grid (b) transonic flow: LDA + N

scheme
Fig. 2. VKI LS-59 rotor blade test case.

total number of elements is 15466. Periodic boundary condition are prescribed

at the lower and upper boundaries; the total pressure and enthalpy and the
flow angle (equal to π/6) are imposed at the subsonic inlet boundary points,
whereas the static pressure corresponding to the prescribed value of the outlet
isentropic Mach number (M2,is = 1) is imposed at the outlet ones. It is note-
worthy that this is a very severe test-case for an inviscid calculation. Indeed,
due to the rounded shape of the trailing edge, the flow experiences a sudden
expansion (with local Mach numbers well beyond one) just before reaching
the stagnation point. This is a destabilizing factor which could hamper the
convergence to steady state, unless a sufficient numerical diffusion is intro-
duced by the method. The low outlet pressure causes a stronger expansion
and acceleration in the cascade. This leads to the formation on the suction
side of two transonic bubbles, one quite week halfway and the another one,
with a considerably stronger shock, at the trailing edge. Another transonic
bubble is present at the trailing edge on the pressure side, due to the rounded
shape of the blade profile. Figure 2b shows the results obtained using the non
linear LDA + N scheme. Entropy oscillations occur at the suction side and at
the trailing edge (c), but they do not grow unboundedly, even if they affect the
other variables, for instance the Mach number and the density. Still, the weak
shock on the trailing edge is well-resolved, as the stronger one downstream,
which interacts with the flow instability. It is noteworthy that the addition
of the artificial dissipation term to the LP scheme does not help removing
entropy oscillations in the solution.
In conclusion, an FS scheme has been applied with success to the solution
of the two-dimensional Euler equations, using a hybrid rectangular-triangular
grid. Future work will demonstrate the validity of such an approach for the
Navier–Stokes equations, for which a high aspect-ratio rectangular grid is a
must in the near-body viscous region.
References
1. H. Deconinck and R. Abgrall: VKI 34th CFD Lecture Series, Higher order dis-
cretization methods (2006).
2. P. De Palma, G. Pascazio, D. T. Rubino, M. Napolitano: J. Comput. Phys., in
press, (2006).
3. A. Csı́k, M. Ricchiuto, H. Deconinck: J. Comput. Phys., 179, 286 (2002).
4. R. Struijs, H. Deconinck, P.L. Roe: VKI LS 1991-01 (1991).
5. P. De Palma, G. Pascazio, G. Rossiello, M. Napolitano: J. Comput. Phys., 208,
1 (2005).
6. R. Abgrall and F. Marpeau: J. Scient. Comput., in press (2006).
7. M. Napolitano, A. Bonfiglioli, P. Cinnella, P. De Palma, G. Pascazio, in Frontiers
of computational fluid dynamics 2002, World Scientific, pp. 113-127 (2002).
8. R. Kiock, F. Lethaus, N. C. Baines, C. H. Sieverding: ASME J. Eng. Gas Tur-
bines Power, 108, 277 (1986).
A Residual-Based Compact Scheme
for All-Speed Flows on Unstructured Grids
Christophe Corre1 , Alberto Beccantini2 , and Thibaud Kloczko2

1
SINUMEF Lab. ENSAM, 151 Bd de l’Hopital, 75013 Paris - FRANCE
corre@paris.ensam.fr
2
CEA Saclay /DM2S / LTMF 91191 Gif-sur-Yvette Cedex - FRANCE
beccantini/kloczko@cea.fr
1 Context and objectives

A residual-based compact (RBC) scheme has been developed in [1]-[3] for
computing compressible flows on structured grids; 2nd and 3rd-order versions
of the scheme were successfully applied to inviscid/viscous steady/unsteady
flows, with good shock capturing properties. Building on these results, sev-
eral extensions are currently explored. First, a very high-order RBC approach
has been developed in a finite difference framework [4],[5] : a supercompact
7th -order scheme was derived, approximating the d-dimensional Euler equa-
tions on a 5d -point stencil without the linear algebra commonly required by
compact schemes. Next, since the RBC scheme design relies on the knowl-
edge of the flux Jacobian matrices eigenvalues and eigenvectors when applied
to a hyperbolic system, a low-Mach number version can be readily derived
for the preconditioned system P −1 wt + fx + gy = 0, using the eigensystems
of the preconditioned flux Jacobian matrices. Alternatively, a RBC scheme
can also be derived for the artificial compressibility system. Finally, the RBC
scheme’s compactness makes its finite volume extension to unstructured grids
attractive since compressible flows can be computed without resorting to any
reconstruction technique nor limitation process : this paper is precisely de-
voted to the presentation of such an unstructured grid version of the RBC
scheme for all-speed flows.
2 Scheme formulation
2.1 Space discretization
Let us consider the three-dimensional system of conservation laws :

∂w
+∇·F =0 (1)
∂t

108 Christophe Corre, Alberto Beccantini, and Thibaud Kloczko
where w = (ρ, ρu, ρv, ρw, ρE)T is the vector of conserved variables and F is
the physical flux tensor. The finite volume approximation of (1) on a general
unstructured grid reads :
dwi 1 X
=− Hi,k |Γi,k | (2)
dt |Ωi |
k
with wi the cell-averaged state in cell i of volume |Ωi | identified here with the
cell-center value (since only second-order accurate methods will be considered)
and Hi,k the numerical flux approximating F · n through face k of cell i, with
area |Γi,k |.
Inviscid case
Let us first assume F = F E (w) where F E contains the physical fluxes defining
the Euler equations. The numerical flux of a conventional upwind scheme will
typically read Hi,k = HE (wi,k
L R
, wi,k ; ni,k ) where HE depends on the specific
scheme retained for approximating F E (Roe, AUSM+, HLLC . . .) and wi,k L
,
R
wi,k denote the reconstructed states on the left and right sides of the interface
L/R
Γi,k , with unit normal vector ni,k pointing outward cell i. States wi,k are
computed as [6] :
L R
wi,k = wi + φi ∇wi · rik , wi,k = wo(i,k) + φo(i,k) ∇wo(i,k) · ro(i,k)k
where index o(i, k) refers to the cell sharing the interface Γi,k with cell i, rik
(resp. ro(i,k)k ) denotes the vector extending from the cell center i (resp. o(i, k))
to the center of the interface Γi,k . The gradient ∇wi is evaluated at the cell
center i using a least-square formula applied on a prescribed spatial support
and φi is a slope limiter, computed according to Barth and Jespersen’s formula
as modified by Venkatakrishnan [7]. Alternatively, the numerical flux of the
second-order RBC scheme reads :
1
Hi,k = (F E (win ) · ni,k + F E (wo(i,k) n
) · ni,k ) − di,k (3)
2
The dissipation flux vector di,k added to the centered evaluation of the normal
flux is based on a discrete integral form of the steady residual r = ∇ · F E
computed at the center of face Γi,k :
−−−−−−→
||Ci Co(i,k) ||
Z
1
di,k = Φi,k r dΩ (4)
2 |Ωi,k | Ωi,k
where Ci and Co(i,k) denote the cell centers on each side of the interface Γi,k
and Φi,k is a matrix coefficient of order O(1) designed so as to ensure the
dissipation of the scheme in the multi-D case (see [1] for more details). For a
2D problem, let us introduce f E , g E the physical fluxes associated with the
first and second space direction and AE , B E their respective Jacobian matrix.
Let us denote :
A RBC Scheme for All-Speed Flows on Unstructured Grids 109
(l) −1
(J⊥ )i,k = AE E
i,k (nx )i,k + Bi,k (ny )i,k = (T⊥ )i,k · Diag((λ⊥ )i,k ) · (T⊥ )i,k
(l) −1
(Jk )i,k = AE E
i,k (tx )i,k + Bi,k (ty )i,k = (Tk )i,k · Diag((λk )i,k ) · (Tk )i,k
where t is a unit vector tangent to face Γi,k . Matrix Φ has the same eigenvec-
(l)
tors as J⊥ , that is Φi,k = (T⊥ )i,k · Diag(φi,k ) · (T⊥−1 )i,k and its eigenvalues are
defined from the eigenvalues of J⊥ and Jk by :
(l)
(l) (l) |Γi,k | |(λ⊥ )i,k |
φi,k = sgn((λ⊥ )i,k ) min(1, −−−−−−→ (m(J )) ) (5)
||Ci Co(i,k) || k i,k
(l)
where m(Jk ) = minl (|λk |). The eigenvalues and eigenvectors at face Γi,k
are typically evaluated using a Roe-averaged state. In practice, the volume
integral over the cell Ωi,k enclosing interface Γi,k , which is used to compute
the residual at the face, is expressed as a surface integral over ∂Ωi,k . The
cell Ωi,k is built on the cell centers Ci , Co(i,k) and the vertices of face Γi,k ,
1 2
denoted Ni,k , Ni,k , so that the integral on ∂Ωi,k is easily computed applying
the trapezoidal rule on each segment of this closed contour and using on one
hand directly available cell-center states and on the other hand node values
estimated from area-weighted averages over the neighboring cell centers. Note
that, since this integral is multiplied by the distance between the cell-centers
located on each side of the face, such a first-order evaluation of the node values
is sufficient to ensure a second-order dissipation.
Viscous case
When solving the Navier-Stokes equations, F in (1) takes the form F E (w) −
F V (w, ∇w) where F V contains the viscous fluxes. The extension of a standard
upwind scheme to the Navier-Stokes equations is straightforward since the
inviscid and viscous numerical fluxes remain decoupled; the full numerical
flux takes the form :
1
Hi,k = HE (wi,k
L R
, wi,k ; ni,k ) − (F⊥V (wNi,k
1 , ∇wN 1 ) + F
V
2 , ∇wN 2 ))
⊥ (wNi,k
2 i,k i,k
where F⊥V = F V ·n. The node values are simply evaluated using area-weighted
averages over the neighboring cell centers while the node gradients are com-
puted with a least-square approach. The RBC numerical flux retains an ex-
pression similar to (3), with a purely centered approximation for both the
inviscid and viscous contribution completed by a residual-based dissipation
now built on a residual that includes the inviscid and viscous flux balance :
1 1
Hi,k = ((F⊥E )ni + (F⊥E )no(i,k) ) − ((F⊥V )Ni,k
1 + (F⊥V )Ni,k
2 ) − di,k (6)
2 2
where F⊥E = F E · n and di,k is still given by (4) with Φ unchanged w.r.t. the
inviscid case but with r = ∇ · (F E − F V ). Since the residual contains both
inviscid and viscous fluxes, evaluating the dissipation flux requires to compute
viscous fluxes and hence gradients of w not only at the grid nodes, which is
needed anyway for the consistent centered flux, but also at cell-centers, which
is done using again a least-square formula.
2.2 Low-Mach unsteady case
The extension of the previous schemes to low-Mach unsteady flows is per-

formed within a dual-time framework and relies on a low-Mach precondition-
ing technique to ensure proper convergence and well-scaled numerical dissi-
pation when M → 0. Thus, a time-accurate solution of (1) for all-speed flows
is obtained as a steady solution of the following evolution problem :
∂w ∂w
Pc−1 + + ∇ · FE = ∇ · FV , (7)
∂τ ∂t
where τ is a dual-time and Pc a preconditioning matrix of the form proposed
by Turkel in [8]. Problem (7) is driven to steady-state with respect to τ using
a first-order approximation for the dual time-derivative (which will vanish at
steady-state anyway), a second-order implicit approximation for the physical
time-derivative and a numerical flux H to approximate the normal physical
flux (F E − F V ) · n; the resulting finite-volume scheme reads :
n,m
∆win,m 32 (wi − win ) − 21 ∆win−1 1 X n,m
(Pc−1 )n,m + + Hi,k |Γi,k | = 0,(8)
i
∆τin,m | ∆t } |Ωi |
{z k
T(win,m )
where m is the pseudo-iteration (on dual-time) counter, n is the time step

counter, ∆wn,m = wn,m+1 − wn,m , ∆wn−1 = wn − wn−1 . Pseudo-time march-
ing leads to a steady solution wn+1 = wn,m+1 = wn,m approximating system
(1) at second-order in time. As far as space accuracy is concerned, the numer-
ical fluxes associated with standard upwind schemes remain unchanged with
respect to the steady case - when Pc = Id - since space and time integration
are fully decoupled. When low-Mach preconditioning is on, these numerical
fluxes must be modified so as to take into account the eigenvalues of Pc AE ,
Pc B E instead of AE , B E : the extension of the HLLC scheme proposed in
[9] will be used in the present work. The RBC numerical flux is still formally
given by (6)-(4) but with r now equal to the steady-state residual with respect
to τ , that is r = wt + ∇ · (F E − F V ). This means the RBC dissipation flux
must be computed using an estimate of T(wn,m ) at the center of face Γi,k ,
obtained from an arithmetic average of this same quantity computed at cell
centers i and o(i, k). Moreover, the use of low-Mach preconditioning means
the eigensystems used in computing the dissipation matrix Φi,k are those of
the preconditioned Jacobian matrices. A key point in making the dual time
approach efficient is to treat (8) in an implicit way, so that the steady solution
is obtained after a limited number of pseudo-iterations by solving :
∆win,m 3 ∆win,m 1 X (i)
n,m
(Pc−1 )n,m
i n,m + + ∆Hi,k |Γi,k | = −Rn,m
i , (9)
∆τi 2 ∆t |Ωi |
k
A RBC Scheme for All-Speed Flows on Unstructured Grids 111
where the numerical flux formula H(i) retained in the implicit stage is differ-
ent from the one used in R since it does not affect the accuracy of the method
provided convergence on m is reached. In the present study, the implicit treat-
ment is the matrix-free method described in [10], which preserves a reduced
cost per iteration even when low-Mach matrix preconditioning is used.
3 Numerical results
A few results demonstrating the RB approach offers an interesting alternative
to existing treatments on unstructured grids are now provided. When coupled
with the proposed matrix-free implicit stage and applied to the transonic flow
over a staggered biplane, the second-order RBC scheme converges to steady-
state faster than the second-order limited AUSM+ scheme and yields a sharp
capturing of the shock appearing in the divergent of the nozzle geometry
created by the two airfoils (see Fig. 1). The RBC results compare also well
with those given by the low-Mach HLLC scheme when computing the low-
subsonic flow over a multi-element airfoil (see Fig. 2). When applied to the
computation of the unsteady low-subsonic laminar flow over the NACA0015
airfoil at high incidence, the RBC scheme yields average values of 0.70 and
0.322 respectively for lift and drag coefficients (see Fig. 3), in good agreement
with C̄L = 0.67, C̄D = 0.304 found in [11] using a structured incompressible
flow solver. Current developments of the RBC approach on unstructured grid
include extension to 3D flows as well as to general equations of state, allowing
to deal for instance with dense gas flows [12].
0
10
10-1
-2
10 AUSM+
RBC
-3
10
Residual
-4
10
-5
10
10-6
-7
10
10-8
1000 2000 3000 4000
Iterations
Fig. 1. Flow at M∞ = 0.7 over a staggered biplane. Convergence history for AUSM+
and RBC (CF L = 106 ); RBC pressure contours on a 17170-elements triangular grid.
5
HLLC (slat)
HLLC (main body)
4 HLLC (flap)
RBC
3
-Cp
2
-1
-0.2 0 0.2 0.4 0.6 0.8 1 1.2

x
Fig. 2. Flow at M∞ = 0.1 over a multi-element airfoil. Zoom on the 7564-elements

triangular grid; wall pressure coefficient using low-Mach versions of HLLC and RBC.
Fig. 3. Flow at M∞ = 0.2, Re∞,c = 1000, α = 16◦ over a NACA0015 airfoil. Lift
and drag evolution using RBC in a 13966-elements triangular grid; instantaneous
streamlines at t = 50.
References
1. Lerat, A., Corre, C.: Comp. & Fluids, 31,639–661 (2002)
2. Lerat, A., Corre, C.: J. Comput. Phys., 170, 642–675 (2001)
3. Corre, C., Hanss, G., Lerat, A.: Comp. & Fluids, 34, 561–580 (2005)
4. Lerat A., Corre, C.:34th VKI Lectures Series on CFD (2005)
5. Corre, C., Lerat, A.: In CFD 2004, Groth & Zingg (eds), Springer (2006)
6. Barth, T.J., Jespersen, D.C.: AIAA Paper 89-0366 (1989)
7. Venkatakrishnan, V.: Comp. & Fluids, 18, 191–204 (1990)
8. Turkel, E. : Applied Numerical Mathematics, 12, 257–284 (1993)
9. Luo, H. Baum, J.D.,Löhner, R. : AIAA J., 43, 1160–1166 (2005)
10. Beccantini, A. et al : In CFD 2004, Groth & Zingg (eds), Springer (2006)
11. Fang, K.C. : PhD thesis, U. of Cincinnati (2000)
12. Congedo, P.M., Cinnella, P., Corre, C. : AIAA J. (2006), accepted.
Vorticity Preserving Scheme
for Unsteady Compressible Flows
Fabrice Falissard1 , Alain Lerat2 , and Jacques Sidès3

1
ONERA/DSNA during this work, now at SINUMEF Laboratory
fabrice.falissard@paris.ensam.fr
2
ENSAM/SINUMEF, bd de l’Hopital, 75013 Paris, France. ONERA Consultant.
alain.lerat@paris.ensam.fr
3
ONERA/DSNA, BP72, F-92322 Chatillon, France
jacques.sides@onera.fr
Summary. Taking advantage of the notion of vorticity preserving schemes intro-

duced by Morton and Roe for acoustics, and on the residual-based schemes family
proposed by Lerat and Corre, an implicit second order accurate residual-based vor-
ticity preserving scheme is presented and applied to blade vortex interaction.
1 Introduction
Reducing the numerical diffusion of vortices is a key point in many computa-

tional fluid dynamics problems such as the simulation of aircraft trailing vor-
tices, rotor-stator interaction of turbo-shaft engines, some aeroacoustic prob-
lems and weather forecasting. One way to reduce the numerical diffusion of
vorticity is to enhance the order of accuracy of the discretisation scheme. This
has motivated the development of multiple high-order approximation meth-
ods. For finite volume or finite different schemes, this may result in larger
stencils and loss of robustness making these schemes difficult to apply to com-
plex flows problems. Another point of view has been proposed by Morton and
Roe through the concept of vorticity preserving scheme in pure acoustics [1].
In the acoustic model, any initial vorticity field is time-invariant and Morton
and Roe found that the Lax-Wendroff-Ni (LWN) scheme [2], also known as
the rotated Richtmyer scheme, preserves exactly a discrete analogue of the
vorticity. In the present work, we extend the concept of vorticity preserving
schemes to acoustics problems including advection and show that in this more
general case, the LWN scheme does not treat well the vorticity. To remove
this difficulty, we identify a scheme preserving vorticity for the linearised Eu-
ler equations, in the family of residual-based schemes introduced by Lerat
and Corre [3]. Applications to the full compressible Euler equations show the
capability of the new scheme to preserve numerical dissipation of the vorticity.

114 Falissard, Lerat and Sidès
2 Vorticity preserving and LWN scheme

Consider an initial-value problem for the hyperbolic system of conservation
laws :
wt + fx + gy = 0 (1)
where t is the time, x and y are space coordinates, w = w(x, y, t) is the state
vector, f = f (w) and g = g(w) are the flux components. Pure acoustics, lin-
earised and full Euler equations can be cast in form (1) and combine acoustic
waves and vorticity phenomena. For each model, the equation for the vorticity
ω, i. e. the rotational of the velocity field, can be obtained from the velocity
equations. On the discrete level, various approximations ω̃ of the vorticity can
be defined. Here we will use the “compact vorticity” :

δ1 µ2 δ2 µ1
ω̃j+ 12 ,k+ 12 = v− u (2)
δx δy j+ 1 ,k+ 1
2 2
here defined on a Cartesian mesh xj = jδx, yk = kδy, using classical difference

and average operators in the x-direction :
1
(δ1 q)j+ 21 ,k = qj+1,k − qj,k (µ1 q)j+ 12 ,k =
(qj+1,k + qj,k ) (3)
2
and similar operators δ2 and µ2 in the y-direction. For pure acoustics, the
exact vorticity equation reduces to :
ωt = 0 . (4)
Thus numerical schemes should see the vorticity field as steady. Both experi-
ence and theory show that this is not the common behavior : most of schemes
dissipate vorticity endlessly. Nevertheless, Morton and Roe [1] proved that
the LWN scheme preserves the vorticity, i. e.
n+1 n
ω̃j+ 1
,k+ 1
= ω̃j+ 1
,k+ 1 (5)
2 2 2 2
where the superscript refers to time.

Unfortunately, this does not extend to acoustics with advection. To show this,
let us first define vorticity preserving for the Euler equations.
A numerical scheme is said to be “vorticity preserving” if its discrete trans-

port equation for vorticity is non-dissipative.
Consider now a linearisation of the Euler equation with respect to a uniform

flow of velocity u0 , v0 . The exact vorticity satisfies the transport equation :
ωt + u0 ωx + v0 ωy = 0 (6)
By applying the discrete rotational to the momentum equations of the LWN
scheme, we can obtain the following equation for the compact vorticity :
Vorticity Preserving Scheme for Unsteady Compressible Flows 115
ω̃ n+1 − ω̃ n

δ1 µ2 δ2 µ1 n
+ u0 . + v0 . µ1 µ2 ω̃ =
∆t δx δy j+ 12 ,k+ 12
(7)
δ 2 µ2 δ 2 µ2

∆t δ1 µ1 δ2 µ2
u20 1 22 . + 2u0 v0 . + v02 2 21 . ω̃ n
2 δx δxδy δy j+ 1 ,k+ 1
2 2
Similarly to the LWN scheme for the state vector w, this discrete equation
for ω̃ includes some dissipation, so that the LWN is not vorticity preserving,
except for pure acoustics (u0 = v0 = 0).
3 Residual-Based scheme preserving Vorticity (RBV)

A residual-based scheme is a scheme which can be expressed in terms of con-
sistent approximations of the residual only. The residual is the left-hand side
of the system of conservation laws, i. e. :
r = wt + fx + gy (8)
Such scheme have been considered by Lerat and Corre [3] to obtain high-order
compact approximations (Residual-Based Compact schemes). In the present
work, we have studied a residual based-scheme of second-order accuracy shar-
ing some feature with the LWN scheme. For system (1), this scheme can be
defined as :
n+1
(Λr̃)j,k = 0 (9)
with :

δx δ1 Φ1 µ2 δy δ2 Φ2 µ1


 Λ = µ1 µ2 . − .− .


 2 δx 2 δy
 n+1
∇w δ1 µ2 δ2 µ1

 n+1
r̃j+ 1 1 = µ1 µ2 + f+ g
2 ,k+ 2 ∆t δx δy

 j+ 12 ,k+ 21
 n+1 n+1 n n−1


 ∇w 3w − 4w + w
=


∆t j,k 2∆t

j,k
where r̃ is the discrete residual and Λ a difference operator composed of a

centered approximation, plus a dissipative part involving numerical viscosity
matrices Φ1 , and Φ2 (see [3]). The time differencing is fully implicit (A-stable)
and second-order accurate. The scheme is resolved iteratively using a classical
dual-time technique. After some computation, the discrete transport equation
of the vorticity for the linearised Euler equations reads :
n+1
∇ δ1 µ2 δ2 µ1 n+1
µ1 µ2 µ1 µ2 . + u0 . + v0 . ω̃ = [EΦ r̃]j+ 1 ,k+ 1 (10)
∆t δx δy j+ 1 ,k+ 1
2 2
2 2
where EΦ is a difference operator involving the dissipative contribution in Λ.

The left-hand side of (10) is a non-dissipative approximation to (6). The right-
hand side depends linearly on the residual r̃. For r̃ = 0 solution of (9), the RHS
vanishes and the scheme is vorticity preserving. Scheme (9) will be refered to as
the RBV scheme. Note that the improvement in vorticity transport provided
by RBV over LWN is mainly due to the use of the complete residual (i. e.
including wt ) in the numerical dissipation of the scheme.
4 Numerical applications
Scheme (9) has been formulated on curvilinear meshes in a finite-volume
framework and applied to various compressible flows problems using the non
linear Euler equations on smooth and irregular meshes. For applications on
non uniform grids, a tenth-order filter [4] is added to the dual-time subitera-
tion scheme.
4.1 Vortex advection
An isentropic vortex of strength Γ = 5 and radius core r = 1 is added to a

uniform flow at velocity u∞ = 1/2, v∞ = 0. The corresponding freestream
Mach number is M∞ = 0.4225. The computational domain is the square
[−5, 5]2 with periodic boundary conditions. An irregular 50 × 50 Cartesian
mesh is used, which makes 10 points in the vortex core. The time step is
taken as ∆t = δx/2u∞ (an optimal choice owing to theoretical study), which
gives CF L = 1.95. Numerical results are shown on Fig. 1 at time t = 100.
At this time, the vortex has covered a distance of 50, i. e. it has crossed 5
times the computational domain. The numerical vortex shows no diffusion
with respect to the exact solution.
(a) (b) (c)
Fig. 1. Vorticity field at t = 100 on an irregular Cartesian mesh. (a) Exact solution,
(b) RBV, (c) RBV with filtering
Vorticity Preserving Scheme for Unsteady Compressible Flows 117
4.2 Head-on parallel blade-vortex interaction
The RBV scheme has has been applied to the case of unsteady head-on inter-
action between a Scully vortex and a NACA 0012 airfoil, at M∞ = 0.5 and
zero incidence, for which experimental results are available [5]. This model
problem is representative of the parallel interaction between a helicopter ro-
tor blade and the vortex emitted at the tip of a preceding blade, which is at
the origin of the blade vortex interaction (BVI) noise. Figure 2 shows that
the use of the RBV scheme leads to grid-convergence for this complex highly
unsteady problem. The RBV scheme preserves the vortex in its displacement
on a long distance before and during the interaction with the profile (see
Fig. 3). Note that supersonic conditions are reached locally when the vortex
impinges the airfoil. Numerical tests have shown that the RBV scheme gives a
converged solution on a coarser grid than with classical schemes of same order
of accuracy. The vortex advection with very little dissipation is necessary to
capture the acoustics in the near field, as illustrated on Fig. 4.
2 2
1
0
1
-1
CP
CP
-2
-3
0
-4
-5
-1 -6
3.3 3.4 3.5 3.6 3.7 3.8 3.3 3.4 3.5 3.6 3.7 3.8
t t
(a) upper-side (b) lower-side
Fig. 2. Time history of surface-pressure coefficient at x/c = 0.02 during blade-

vortex interaction. Experiment [5] and RBV+F10 solutions on meshes with 2 ( ),
4 ( ), 8 ( )) and 16 ( ) points in the vortex core.
References
1. K. W. Morton and P. L. Roe: Vorticity-Preserving Lax-Wendroff-Type Scheme
for the System Wave Equation, SIAM J. Sci. Comput., 23:170–192, 2001.
2. R. H. Ni: A Multiple-Grid Scheme for Solving the Euler Equations, AIAA Jour-
nal, 20:1565–1571, 1982.
3. A. Lerat and C. Corre: A Residual-Based Compact Scheme for the Compressible
Navier-Stokes Equations. J. Comput. Phys., 170:642–675, 2001.
4. M. R. Visbal and D. V. Gaitonde: On the Use of High-Order Finite-Difference
Schemes on Curvilinear and Deforming Meshes. J. Comput. Phys., 181:155–185,
2002.
5. S. Lee and D. Bershader: Head-On Parallel Blade-Vortex Interaction, AIAA
Journal, 32:16–22, 1994.
(a)
(b)
(c)
Fig. 3. Grid convergence of the vorticity field at successive times. Points in vortex
core : (a) 4 pts, (b) 8 pts, (c) 16 pts.
(a) t = 3.75 (b) t = 4.00 (c) t = 4.65
Fig. 4. Divergence of the velocity field during the blade-vortex interaction

Extension of the SD Method to Viscous Flow
on Unstructured Grids
Z. J. Wang1 , Yuzhi Sun1 , C. Liang1 and Yen, Liu2

1
Department of Aerospace Engineering, Iowa State University, 2271 Howe Hall,
Ames, IA50011, USA. †
2
NASA Ames Research Center, Moffett Field, CA94035, USA.
Summary. The Spectral Difference (SD) method has been developed recently by
[1] for the wave equations on unstructured triangular grids and further developed
by [2] for 2D Euler equations. In this paper, the SD method is extended to solve vis-
cous flow governed by compressible Navier-Stokes equations in both 2D and 3D, for
unstructured triangular and hexahedral grids, respectively. Some numerical results
are presented to demonstrate its capability.
1 Introduction
High order methods on unstructured grids have received considerable atten-

tion lately because of the potential to handle complex geometry with higher
than 2nd order accuracy. For hyperbolic conservation laws, the numerical
methods should also be locally conservative. Conservative high order methods
include perhaps the leader, discontinuous Galerkin (DG) method [3], multi-
domain spectral method [4], spectral volume (SV) [5] and spectral difference
(SD) methods [1, 2]. The SD method appears to be the easiest to implement
because it does not involve any surface or volume integrals. In the paper, we
describe how to extend the SD method to the Navier-Stokes equations on
triangular and hexahedral grids. Some related work has been presented in [6].
2 Formulation of the Spectral Difference Method
Two sets of nodal points within the computational element or cell are defined.
One set is named the solution points where the solution unknowns are placed.
The other set is called the flux points, where fluxes in all coordinate directions
are located. The solution and flux points for triangular and quadrilateral cells
†
Correspondence can be addressed to the first author. Tel: +1 515 294 1614; Fax:
+1 515 294 3262; Email: zjw@iastate.edu

120 Z. J. Wang, Yuzhi Sun, C. Liang and Yen, Liu
are illustrated in Figure 1. The placement of solution and flux points within
a hexahedral cell is very similar to Figure 1a. If the solution points support a
Lagrangian interpolation of degree p, the flux points usually support a degree
p+1 Lagrangian polynomial.
(a) Second order (b) Third order
Fig. 1. Placement of solution () and flux (◦) points for triangular and quadrilateral
elements
Given the solutions at the solution points, the computation of inviscid

fluxes is described in [1, 2], and will not be repeated here. The viscous flux is
a function of both the conserved variables and their gradients. Therefore the
key is how to compute the solution gradients at the flux points. Two different
approaches have been tested, and both gave similar results. One approach is
similar to one presented in [4], and the other follows the local DG idea [7].
The algorithm to compute the solution gradients takes the following steps:
1. Compute the state variables at the flux points using the Lagrangian poly-
nomial reconstructed from the variables at the solution points;
2. For flux points located along cell interfaces, average the multiple solutions
computed from different cells;
3. Find the degree p+1 interpolation polynomial from the variables at the
flux points, and compute the gradients at the solution points;
4. Interpolate the gradients from the solution points to the flux points. For
flux points located along cell interfaces, average the multiple gradients
computed from different cells;
In the local-DG like approach, Steps 2 and 4 are slightly different. Instead
of taking the average at the flux points along a cell interface, we take the
solution and gradient from alternate side of the interface. For example, if we
take the solution from the right side of the interface, we then take the gradient
from the left side of the interface.
The formulation for a hexahedral cell (hex) is similar to the multi-domain
spectral method [4]. All hexes are transformed from the physical domain (x, y,
Extension of the SD Method to Viscous Flow on Unstructured Grids 121
z) to a computational standard element [0,1]x[0,1]x[0,1], (ξ, η, ς). The govern-

ing equations are also transformed accordingly. With the placement of solution
and flux points shown in Figure 1b, all operations are one-dimensional.
3 Numerical Results
3.1 Accuracy Study
To demonstrate the performance of the SD method, we first solve the 2D

scalar convection-diffusion equation on a set of irregular triangular grids
ut + c(ux + uy ) − µ(uxx + uyy ) = 0, c = 1, µ = 0.0001 (1)
The computational domain is (x, y) ∈ (−1, 1) × (−1, 1) and the initial
condition is given u(x, y, 0) = sin(π(x + y) and periodic boundary conditions
are applied. The numerical order of accuracy can be evaluated based on errors
on the set of meshes and the exact solution u(x, y, t) = e−2πµt sin(π(x + y −
2ct)). As shown in Table 1, 2nd order accuracy is obtained for the 2nd order
SD scheme. Numerical examples for higher than 2nd-order SD schemes showed
that these higher-order SD schemes are not convergent on the finer grids, i.e.,
the errors on the finer grids do not decay at the expected rates. These tests
may indicate that the current formulation has a weak instability. Further
investigations will be carried out to find the cause and maybe a remedy for
this problem. The performance of the SD method on the hexahedral grids,
Table 1. L1 and L∞ errors and orders of accuracy for the convection-diffusion

equation on irregular triangular meshes at t = 1.0
Order No. of cells L1 error L1 order L∞ error L∞ order

10x10x2 6.40e-02 — 1.98e-01 —
2 20x20x2 1.70e-02 1.92 6.24e-02 1.67
40x40x2 4.27e-03 1.99 1.59e-02 1.97
80x80x2 1.08e-03 1.99 4.14e-03 1.94
however, is fully satisfactory. The SD method was successfully implemented

for the Navier-Stokes equations, and a grid refinement accuracy study was
performed using the Couette flow problem, which has an exact solution. The
results of this study are presented in Table 2. Note that high-order accuracy
was demonstrated successfully.
3.2 Unsteady Flow over a Cylinder
Unsteady viscous flow over a circular cylinder was selected to demonstrate the
SD method for time-accurate viscous flow computations. Both a triangular
grid and a hexahedral grid were used. Since the flow is two-dimensional, the
hex grid only has one cell in the span-wise direction. Both grids are shown in
Figure 2. The triangular grid has 3,516 cells, while the hex grid has 590 cells.
For both grids, the cylinder wall is approximated with piecewise quadratic
curves or patches. The inflow Mach number is 0.2.
Table 2. Grid refinement study for Couette flow
Order Grid L1 error L1 order L∞ error L∞ order L2 error L2 order

2x2x1 8.72e-05 — 3.06e-04 — 1.26e-04 —
3 4x4x1 9.34e-06 3.22 3.96e-05 2.95 1.37e-05 3.20
8x8x1 1.14e-06 3.03 5.10e-06 2.96 1.56e-06 3.13
16x16x1 2.08e-07 2.46 7.05e-07 2.86 2.44e-07 2.68
2x2x1 2.35e-05 — 6.49e-05 — 3.29e-05 —
4 4x4x1 1.84e-06 3.68 4.79e-06 3.76 2.40e-06 3.78
8x8x1 1.70e-07 3.43 3.80e-07 3.66 1.98e-07 3.60
16x16x1 1.77e-08 3.26 3.73e-08 3.35 1.90e-08 3.38
2x2x1 4.69e-07 — 2.00e-06 — 7.19e-07 —
5 4x4x1 1.29e-08 5.19 7.35e-08 4.77 2.03e-08 5.14
8x8x1 3.28e-10 5.30 2.44e-09 4.91 5.10e-10 5.32
16x16x1 1.28e-11 4.68 8.30e-11 4.88 1.63e-11 4.97
(a) Triangular (b) Hexahedral
Fig. 2. Triangular and hexahedral grids around a cylinder
Many experimental and computational studies have been carried out to

study the unsteady flow around a cylinder at various Reynolds numbers [8].
In the present study, we selected to run two Reynolds numbers, Reynolds of
75 on the triangular grid, and 160 on the hex grid. In order to trigger the flow
unsteadiness, the initial condition has small upward component. The flow on
both grids quickly developed periodic vortex shedding.
Extension of the SD Method to Viscous Flow on Unstructured Grids 123
The cl and cd histories computed with the triangular grid are shown in
Figure 3a. Based on the history, the vortex shedding frequency can be com-
puted, and then the Strouhal number. The computed Strouhal number from
the 2nd order SD simulation is 0.156, comparing with the experimentally
measured 0.148. The entropy contours are shown in Figure 3b. The Karman
vortex street is captured well. On the hex grid, both the 3rd and 4th order
(a) cl and cd histories (b) Entropy contours
Fig. 3. Flow over a cylinder using the second order SD scheme on the triangular
grid
SD schemes were employed in the simulation. The computed Strouhal num-

ber is 0.194 using the 3rd order SD scheme, and 0.192 using the 4th order
SD scheme, indicating a very close agreement between the two simulations. In
comparison, the experimentally measured Strouhal number is 0.186 [8]. The
computed vorticity magnitude using the 4th order scheme are displayed in
Figure 4, again clearly showing the unsteady vortices.
Fig. 4. Vorticity magnitude contours for flow over a cylinder computed with the
4th order SD scheme on the hex grid
4 Conclusions
In this paper, the spectral difference method is extended to the Navier-Stokes
equations on both triangular and hex grids. Two approaches were tested in
the computation of viscous fluxes, and both give similar results. Although the
2nd-order SD scheme is convergent, higher order SD schemes appear to have a
weak instability, and further investigations will be carried out to find the cause
and a remedy. The extension to viscous flow on hex grids is fully satisfactory.
High-order accuracy was numerically demonstrated using the Couette flow.
The SD schemes are then used to compute unsteady flow over a circular
cylinder, and the von Kamnan vortex street was captured, and the computed
Strouhal number agrees well with experimentally measured value. Currently
we are developing an efficient implicit/multigrid solution algorithm.
5 Acknowledgements
The study was partially funded by Rockwell Scientific/DARPA under contract

W911NF-04-C-0102, DOE grant DE-FG02-05ER25677 and AFOSR grant
FA9550-06-1-0146. The views and conclusions contained herein are those of
the authors and should not be interpreted as necessarily representing the of-
ficial policies or endorsements, either expressed or implied, of DARPA, DOE,
AFOSR, or the U.S. Government.
References
1. Y. Liu, M. Vinokur, M., and Z.J. Wang: Proceedings of the ICCFD-3, July 12-16,
2004, Toronto, Canada, Springer, 2004.
2. Z.J. Wang and Y. Liu: AIAA paper 2005-5112, 2005.
3. B. Cockburn and C.-W. Shu: J. Comput. Phys. 141, 199 - 224 (1998).
4. D.A. Kopriva: J. of Comput. Phys. 143, 125-158 (1998).
5. Z.J. Wang, L. Zhang and Y. Liu: J. Comput. Phys. 194, 716-741 (2004).
6. G. May and A. Jameson: AIAA Paper No. 2006-304, 2006.
7. B. Cockburn and C. -W. Shu: SIAM J. Numer. Anal. 35, 2440-2463 (1998).
8. C.H.K. Williamson: J. Fluid Mech. 206, 579 (1989).
Strictly Stable High Order Difference Methods
for the Compressible Euler and Navier-Stokes
Equations
Bernhard Müller
Department of Energy and Process Engineering, Norwegian University of Science

and Technology, NO–7491 Trondheim, Norway bernhard.muller@ntnu.no
1 Introduction
High order finite difference methods have been constructed to be strictly stable
for linear hyperbolic and parabolic problems. The methods employ high order
central approximations in the interior and special boundary stencils to satisfy
the summation by parts (SBP) property leading to discrete energy estimates
[1][2][3][4][5].
Instead of standard central high order difference approximations, disper-
sion relation preserving schemes can be used in the interior to yield SBP
operators [6]. With entropy splitting, i.e. splitting of the nonlinear Euler and
Navier-Stokes equations into a conservative portion and a symmetric non-
conservative portion, generalized energy estimates can be derived to yield
strict stability [7][8].
Here, the perturbation form of the nonlinear Euler and Navier-Stokes equa-
tions has been employed to minimize cancellation errors at low Mach numbers
and to reduce to the linearized equations, if the nonlinear terms are neglected
[9][10][11]. The SBP operator with sixth order in the interior and third order
near the boundaries by B. Strand [2] is employed to approximate the spatial
first derivatives in the nonlinear Euler and Navier-Stokes equations. Second
derivatives are computed by applying the first order difference operator twice.
The explicit fourth order Runge-Kutta method is used for time integration.
High wave number oscillations are damped by a sixth order explicit filter.
Sound propagation of the Kirchhoff vortex, of rocket-launch noise before lift-
off and of point sources in the atmosphere has been simulated by solving the
nonlinear and linearized Euler equations [12][13][14][15][11][16][6]. The numer-
ical approach has recently been extended to the Navier-Stokes equations [17].
High accuracy is required for correctly computing not only sound propagation
over long distances and times but also sound generation due to viscous and
nonlinear effects.

126 Bernhard Müller
2 Flow Equations in Perturbation Form

The perturbation formulation is used to minimize cancellation errors when
discretizing the Euler and Navier-Stokes equations for compressible low Mach
number flow [9]. The 2D compressible Navier-Stokes equations in conservative
form can be expressed in perturbation form as [10]
U0 t + F0 x + G0 y = 0 , (1)
where F0 = Fc 0 − Fv 0 and G0 = Gc 0 − Gv 0 denote the differences of the
inviscid and viscous flux vectors in the x- and y-directions, respectively, with
respect to F(U0 ) and G(U0 ), where U0 = (ρ∞ , 0, 0, (ρE)∞ )T and subscript
∞ denotes freestream values. The subscripts in (1) and subsequently denote
derivatives. We assume perfect gas. The conservative perturbation variables
U0 and the inviscid flux vectors are defined by
 0
(ρu)0 (ρv)0
    
ρ
 (ρu)0   (ρu)0 u0 + p0 0 0
U0 =  0  , G0 =  (ρu)0 v0
  
 (ρv)0  , F =  (ρv)0 u0  (ρv) v + p0
  
 
0 0 0 0 0
(ρE) (ρ∞ H∞ + (ρH) )u (ρ∞ H∞ + (ρH) )v
The perturbation variables (superscript 0 ) are defined by ρ0 = ρ − ρ∞ , (ρu)0 =
0
ρu, (ρE)0 = ρE − (ρE)∞ , (ρH)0 = (ρE)0 + p0 , u0 = ρ(ρu)∞ +ρ
0 0
0 , p = (γ − 1)[(ρE) −
1 0 0
2 ((ρu) ·u )] . t is time, and x and y are the Cartesian coordinates, respectively.
ρ denotes the density, u and v the x- and y-direction velocities, E the specific
total energy, p the pressure, H the total enthalpy, γ = 1.4 the ratio of specific
heats for air, P r = 0.72 the Prandtl number. The viscous flux vectors Fv 0
and Gv 0 are the same as for the standard conservative form, except for using
the temperature perturbation T 0 instead of temperature T for the heat flux
terms. The perturbation formulation with respect to uniform flow or general
base flow is somewhat more elaborate [10]. General geometries are treated by
a coordinate transformation x = x(ξ, η), y = y(ξ, η) [17].
3 High Order Difference Methods

3.1 Approach for Convection Diffusion Equation
The energy method shows us how to construct difference operators such that
high order and strict stability including the boundary conditions are guaran-
teed. As a model equation, we consider the convection diffusion equation with
homogeneous boundary conditions
ut + aux = buxx , 0 ≤ x ≤ 1, (2)
u(x, 0) = f (x) , 0 ≤ x ≤ 1, (3)
u(0, t) = u(1, t) = 0 , 0 ≤ t. (4)
where a and b are assumed to be constant and positive.
Strictly Stable High Order Difference Methods 127
Let vj = vj (t) denote the approximation of the exact solution of the con-
vection diffusion equation at grid point xj = jh with the grid spacing h = N1 .
We use the notation v = [v0 , v1 , ..., vN ]T . We define the discrete scalar product
and norm
(u, v)h = huT Hv , ||u||2h = (u, u)h , (5)
where H is a diagonal and positive definite matrix.
A difference operator Q satisfies the SBP property, if [2]
(u, Qv)h = uN vN − u0 v0 − (Qu, v)h . (6)
Thus, the discrete energy estimate for vt + aQv = bQQv becomes [17]
d
||v(t)||2h ≤ av02 + 2b[vN (Qv)N − v0 (Qv)0 ] . (7)
dt
The boundary conditions can be weakly imposed by the simultaneous approx-
imation term (SAT) [3] to get strict stability, i.e. continuous dependence on
the initial and boundary data and up to O(∆x) the same growth rate as the
continuous convection diffusion equation [5][18].
3.2 Approach for Euler and Navier-Stokes Equations
The transformed 2D compressible Euler and Navier-Stokes equations in per-

turbation form are solved on an annulus. ξ and η are the radial and circum-
ferential coordinates, respectively. The ξ-derivatives in the transformed Euler
and Navier-Stokes equations and in the metric terms are discretized by the
3-6 SBP operator by B. Strand, which is third order accurate near the bound-
aries (at the boundary and the 5 adjacent grid points) and corresponds to the
(6)
standard sixth order central difference operator Qξ in the interior [2]. Due
to the periodic boundary conditions in the η-direction, the standard sixth or-
(6)
der central difference operator Qη is applied in the η-direction. The viscous
terms are discretized by first approximating the first ξ and η derivatives of
u0 , v 0 and T 0 , by B. Strand’s 3-6 SBP operator and the standard sixth order
central difference operator, respectively. After the flux vectors F̂0 and Ĝ0 are
computed at all grid points, F̂0ξ and Ĝ0η are approximated by employing the
difference operators once more. However, this approximation of the second
derivatives leads to a wide stencil and does not damp the high wave number
π
modes k = ∆x . The high order difference approximations derived for uxx [5]
and (bux )x with a variable coefficient b = b(x, t) [18] alleviate those problems,
cf. Fig. 1 where ξ denotes the nondimensional wave number.
The classical fourth order explicit Runge-Kutta method is used for time
integration. The Euler and Navier-Stokes equations are not only solved in
the interior but also at the boundaries. The ingoing characteristic variables
and isothermal no-slip wall boundary conditions are imposed at ξ = 1 for the
Euler and Navier-Stokes equations, respectively. At the farfield boundary ξ =
ξmax , the first approximation of the Engquist-Majda nonreflecting boundary

conditions is imposed.
Spurious high wave number oscillations are suppressed by a sixth order
explicit filter by modifying the numerical solution at the completion of a full
time step of the Runge-Kutta method [17]. Although the sixth order explicit
filter was found to be less accurate than a characteristic based explicit filter
[11], the former has been used here because of its lower complexity.
6
−h2 D (!)
0
0 0.5 1 1.5 2 2.5 3
! = 2" # h
Fig. 1. Damping of modes with nondimensional wave number ξ = k∆x of the exact
second derivative (-), the standard central fourth order difference operator for the
first derivative D1 applied twice (- -), the standard central fourth order difference
operator for the second derivative D2 (-·), the fourth order difference operator Q
derived for the self-adjoint form (·) and the fourth order operator L based on finite
elements with mass lumping (x) [18].
4 Results
To verify the 2D Navier-Stokes solver for computational aeroacoustics, the

flow around a circular cylinder generating aeolian tones has been selected.
The freestream Mach numbers are M∞ = 0.1 and M∞ = 0.2, respectively.
We consider a low Reynolds number (Re∞ = 150 based on the cylinder di-
ameter), where the Kármán vortex street is observed. Each time a vortex is
shed from the cylinder, a sound wave is emitted. The frequency of the vortex
shedding and the periodic variation of lift and drag are in excellent agreement
with a reference solution [17]. At the farfield, the first approximation of the
Engquist-Majda absorbing boundary conditions proved to be sufficiently non-
reflective., cf. Fig. 2. Whereas the unsteady flow could be correctly predicted
even with single precision using the perturbation formulation of the Navier-
Stokes equations, double precision was required to correctly simulate sound
Strictly Stable High Order Difference Methods 129
propagation [17]. With the present approach, the direct numerical simulation
of aeolian tones at higher Reynolds numbers can be attacked.
! p = [−0.1:0.005:0.1]*M2.5 −4 ! p = [−0.1:0.005:0.1]*M2.5 −4
x 10 x 10
Acoustic pressure w.r.t. pmean, Re=150, M=0.1, 513 x 294 grid Acoustic pressure w.r.t. pmean, Re=150, M=0.2, 513 x 294 grid 6
100 1 100
4
80 80
60 0.5 60
2
40 40
0 0
20 20
0 0
y
y
−2
−0.5
−20 −20
−4
−40 −40
−1
−60 −60 −6
−80 −80
−1.5 −8
−100 −100
−100 −50 0 50 100 −100 −50 0 50 100
x x
p̃0
Fig. 2. Nondimensional instantaneous fluctuation pressure ρ∞ c 2
= [−0.1 : 0.005 :
∞
2.5
0.1]M∞ for M∞ = 0.1 (left) and M∞ = 0.2 (right) at Re∞ = 150.
5 Conclusions
Since strict stability is taken into account in the construction of summation
by parts (SBP) difference operators, those strictly stable high order difference
methods for the Euler and Navier-Stokes equations have been providing reli-
able numerical tools for sound generation and sound propagation.
Acknowledgements
The present work was performed while the author was associated with the
Division of Scientific Computing, Department of Information Technology, Up-
psala University, Sweden. The parallelization of the Navier-Stokes code by my
former colleague Jarmo Rantakokko is gratefully acknowledged.
References
1. B. Gustafsson, H.-O. Kreiss, J. Oliger: Time Dependent Problems and Differ-
ence Methods. John Wiley & Sons, New York (1995).
2. B. Strand: Summation by Parts for Finite Difference Approximations for d/dx.
J. Comput. Physics 110, 47–67 (1994).
3. M.H. Carpenter , D. Gottlieb, S. Abarbanel: Time-Stable Boundary Conditions
for Finite-Difference Schemes Solving Hyperbolic Systems: Methodology and
Application to High-Order Compact Schemes. J. Comput. Physics 111, 220-
236 (1994).
4. J. Nordström, M.H. Carpenter: Boundary and Interface Conditions for High-

Order Finite-Difference Methods Applied to the Euler and Navier-Stokes Equa-
tions. J. Comput. Physics 148, 621–645 (1999).
5. K. Mattsson, J. Nordström: Finite Difference Approximations of Second Deriva-
tives on Summation by Parts Form, J. Comput. Physics 199, 503–540 (2004).
6. S. Johansson: High Order Difference Approximations for the Lin-
earized Euler Equations. Licentiate Thesis 2004-006, Department
of Information Technology, Uppsala University, Sweden (2004),
http://www.it.uu.se/research/publications/lic/2004-006/.
7. A. Harten: On the Symmetric Form of Systems of Conservation Laws with
Entropy. J. Comput. Physics, 49, 151–164 (1983).
8. E. Tadmor, W. Zhong: Entropy Stable Approximations of Navier-Stokes Equa-
tions with no Artificial Numerical Viscosity. J. Hyperbolic Differential Equa-
tions 3 (3), 529–559 (2006).
9. J. Sesterhenn, B. Müller, H. Thomann: On the Cancellation Problem in Cal-
culating Compressible Low Mach Number Flows. J. Comput. Physics 151,
597–615 (1999).
10. B. Müller: Computation of Compressible Low Mach Number Flow. Habilitation
Thesis, ETH Zürich (1996).
11. B. Müller, S. Johansson: Strictly Stable High Order Difference Approxima-
tions for Low Mach Number Computational Aeroacoustics. In: Proceedings
of 4th ECCOMAS Congress 2004, Jyväskylä, Finland, 24-28 July 2004, 19
pages, edited by P. Neittaanmäki, T. Rossi, K. Majava., O. Pironneau,
http://user.it.uu.se/∼bernd/eccomas2004 bm sj.pdf.
12. B. Müller: High Order Difference Method for Low Mach Number Aeroacoustics.
In: Proceedings of ECCOMAS CFD Conference 2001, Swansea, 4-7 Sept. 2001,
20 pages, http://user.it.uu.se/ ∼bernd/eccomas cfd2001.pdf.
13. B. Müller, H.C. Yee: High Order Numerical Simulation of Sound Generated
by the Kirchhoff Vortex. Computing and Visualization in Science 4, 197–204,
(2002).
14. B. Müller, H. C. Yee: Entropy Splitting for High Order Numerical Simulation
of Vortex Sound at Low Mach Numbers. J. Scientific Computing 17 (1-4),
181–190 (2002).
15. B. Müller, S. Johansson: Strictly Stable High Order Difference Approximations
for Computational Aeroacoustics. Comptes Rendus Mecanique 333 (9), 699–
705 (2005).
16. C. Müller: High Order Accurate Numerical Solution of the Linearized Euler
Equations for Sound Propagation in the Atmosphere, Master’s Thesis, Uppsala
University (2004).
17. B. Müller: High Order Numerical Simulation of Aeolian Tones. Computers &
Fluids 37, 450–462 (2008).
18. K. Kormann, M. Kronbichler: High Order Finite Difference Approximations for
Parabolic and Hyperbolic-Parabolic Problems with Variable Coefficients. Project
Report, Division of Scientific Computing, Department of Information Technol-
ogy, Uppsala University, Sweden (2006).
Uniform Flow Preserving Property of High
Order Upwind Finite Difference Schemes on
Generalized Coordinate System
Taku Nonomura1 , Nobuyuki Iizuka2 , and Kozo Fujii3

1
Univ. of Tokyo, Japan nonomura@flab.eng.isas.jaxa.jp
2
Japan Aerospace Exploration Agency, Japan iizuka.nobuyuki@jaxa.jp
3
ISAS/Japan Aerospace Exploration Agency, Japan fujii@flab.eng.isas.jaxa
1 Introduction
Recently, interests of many scientists and engineers are shifting from Reynolds-
averaged Navier-Stokes simulations to direct numerical simulations, large-
eddy simulations and aero-acoustic simulations. A major problem for these
simulations is that they are computationally very expensive because they re-
quire very high spatial resolution to guarantee accuracy of their results.
An approach to solve this problem is to use high-order schemes with less
grid points. According to Balsara[1], resolution of eighth-order scheme is four
times as high as second-order schemes in each dimension while it costs only
three times as much as second-order one. Thus, for three dimensional turbu-
lence problem, computational costs can be saved by factor of 44 /3(=83) with
eighth order scheme. Moreover, according to Shu[2], when high-order scheme
is implemented in multi-dimensional problem, finite difference schemes, that
can be constructed by dimension by dimension procedure, costs ten times as
little as finite element methods or finite volume methods which need multi
dimensional reconstruction.
Therefore various high order finite difference schemes were recently proposed.
These schemes include compact scheme proposed by Lele[3] and adapted to
generalized coordinate by Gaitonde and Visbal[4], weighted essential no os-
cillatory (WENO) scheme proposed by Jiang et al[5], and weighted compact
non-linear schemes (WCNS) proposed by Deng and Zhang[6, 7, 8]. WENO or
WCNS scheme has been developed to solve flow-fields including discontinuity
such as shock wave without numerical oscillation.
When finite difference schemes are adapted to generalized coordinate form,
capability of preserving uniform flow is important because noise from un-
preserved uniform flow hides very small oscillation, such as turbulent flow
structure or aero-acoustic wave. So far, for two-dimensional problem, preser-
vation of uniform flow has been achieved by using same stencils for both flow

132 Taku Nonomura, Nobuyuki Iizuka, and Kozo Fujii
computation and evaluation of metrics for transformation of coordinate. For

three-dimensional problems, preservation of uniform flow has been achieved
by evaluating metrics of transformation of coordinate based on finite volume
method concepts. However, present approach based on finite volume method
concepts do not suit for high order schemes.
Gaitonde and Visbal[4] proposed to use same stencil for both flow and met-
rics computations by adopting the idea of conservative form of metric terms
proposed by Thomas et al to preserve uniform flow for compact scheme on
three-dimensional generalized coordinate. On the other hands, most of re-
searches of WENO and WCNS schemes almost have been demonstrated on
only Cartesian coordinate system. Therefore uniform preserving properties of
these schemes on generalized coordinate have not been studied.
Our interest of research is evaluation of noise from rocket engine plume, whose
main noise source is turbulent flow including shock waves, around complicated
bodies. On computing such flow-fields, high order non-linear schemes such as
WENO or WCNS on the generalized coordinate system are necessary to save
the computational costs. In this study, uniform preserving and vortex preserv-
ing properties of WENO and WCNS are investigated.
2 WENO and WCNS
WENO schemes use rth-order numerical flux f̃j+ 12 defined implicitly as fol-
lowing expression.

∂f 1
= f̃j+ 12 − f̃j− 12 + O(∆xr ) (1)
∂x j ∆x
It should be noted that this numerical flux is different from the physical
flux f̂j+1/2 . The numerical flux is constructed by weighted function of fluxes
+
of computational nodes. Moreover numerical flux is divided into f̂j+1/2 and
−
f̂j+1/2 . Weighted upwind-stencil is used for each divided flux. Thus in this
formulation, it is difficult to use same stencil for both flow computation and
metrics evaluation. Detailed procedure is explained in the reference of Deng
et al[6].
In this study, an explicit version of WCNS is investigated because it is more
efficient than the compact version[6]. This WCNS scheme uses rth-order phys-
ical flux defined explicitly by the following expression.
f̂j+ 12 = f (u(xj+ 12 )) + O(∆xr−1 ) (2)
This physical flux can be constructed by interpolation of flow variables such
as MUSCL procedure instead of finite difference scheme[6]. However, as is
physical flux, simple difference scheme achieves only second order as follows.
Uniform Flow Preserving Property of WENO and WCNS 133

∂f 1
= f̂j+k+ 12 − f̂j−k− 21 + O(∆x2 ) (3)
∂x j ∆x
Where the above expression is same as second order central difference on
twice finer computational grid. Thus to achieve higher order accuracy, sth-
order staggered finite difference schemes written as following expression is
necessary.

∂f 1 X
= bk f̂j+k+ 12 − f̂j−k− 21 + O(∆xr ) + O(∆xs ) (4)
∂x j ∆x
Since this high-order staggered finite difference scheme is linear scheme, it
seems that idea of Gaitonde and Visbal[4] can be used to preserve uniform
flow. Actual metric treatment to preserve uniform flow is explained as fol-
lows. In WCNS, cell-interface metric and high-order staggered finite differ-
ence schemes are used. This cell-interface metric is high-order-interpolated
from cell-node metric. If conservative form of metric terms is computed by
same cell-interface interpolation and same high-order staggered finite differ-
ence schemes, uniform flow seems to be preserved.
3 Metric Computation for Each Case
Uniform flow preserving properties and vortex preserving properties of four

schemes are computed. WENO-G is a WENO scheme computed in general-
ized coordinate system using conservative form of metric evaluated by compact
schemes, because it is difficult to use the same stencil both the flow compu-
tation and the metrics evaluation.
WENO-C is computed in Cartesian coordinate system as the following ex-
pression with evaluating all terms numerically.
∂Q
∂t + ξx ∂E ∂E ∂E ∂F ∂F
∂ξ + ηx ∂η + ζx ∂ζ + ξy ∂ξ + ηy ∂η
∂F ∂G ∂G ∂G (5)
+ζy ∂ζ + ξz ∂ξ + ηz ∂η + ζz ∂ζ = 0
Computation using the non-conservative form written as above should pre-
serve uniform flow. However its computational cost is three times as expensive
as WENO-G.
WCNS-G1 is computed in generalized coordinate system using the conserva-
tive form of metric evaluated by the same interpolation and the same finite
difference schemes. The metric evaluation should preserve the uniform flow.
WCNS-G2 is computed in generalized coordinate system using conservative
form of metric evaluated by compact schemes. The scheme will not preserve
the uniform flow because the stencil of the metric evaluation is different from
that of the flow-computation.
4 Uniform Flow and Vortex Preservation Tests

Uniform flow and vortex preservation test has been carried out for each
scheme. The computational grid used for uniform flow preservation test
is the same wavy grid as the one used in the research of Gaitonde and
Visbal[4](Fig.1). Mach number of the uniform flow in x-direction u is 0.5.
Thus y-direction velocity v and z-direction velocity w are expected to keep
machine-zero. TVD-Runge-Kutta scheme[9, 10] is used for time integration.
Flow-fields after 100 steps are examined. Results are shown in table 1. The
present result demonstrated that WENO-C and WCNS-G1 preserve uniform
flow while WENO-G and WCNS-G2 can not.
Then vortex preserving test is carried out for the above four schemes and
MUSCL schemes for reference. The computational grid used here is also al-
most same as the one used in the research of Gaitonde and Visbal[4](Fig.2). An
isentropic vortex is given where exact solution should keep its vortex strength.
TVD-Runge-Kutta scheme is used for time integration. Flow-fields after 1000
steps are examined. Pressure distribution of the results are shown in Fig. 3.
Vortex center computed by WENO-G is moved and the corresponding suc-
tion peak is much weakened. Vortex center of WENO-C keeps its position,
but its suction peak is weakened as MUSCL. On the other hand, WCNS-G1
and WCNS-G2 keep vortex position and their suction peak.
Table 1. Results of uniform flow preserving test.(n is number of the grid points)
P 2 1P 2
max(v) max(w) n1 w n w
WENO-G 8.75E-02 0.474989 9.98E-04 3.08E-02
WENO-C 6.82E-15 7.46E-15 3.92E-30 4.04E-30
WCNS-G1 1.75E-14 1.75E-14 8.00E-30 8.00E-30
WCNS-G2 1.13E-03 1.13E-03 7.49E-08 7.49E-08
5 Conclusions
Uniform flow and vortex preserving properties on wavy grids of four high-
order sschemes have been investigated while idea of Gaitonde and Visbal is
applied to WCNS scheme. The present results show that WCNS computed
in generalized coordinate system using conservative form of metric evaluated
by the same interpolation and the same finite difference schemes as flow com-
putation, has very good properties for both uniform flow and vortex flow
preservation.
Uniform Flow Preserving Property of WENO and WCNS 135
Fig. 1. Grid for unform flow preserving test (21x21x21, every 2 line is shown)
Fig. 2. Grid for vortex preserving test (41x41x21, every 2 line is shown)
Fig. 3. Results of vortex preserving test on the wavy grid

References
1. Balsara, D. and Shu, C.W.: Journal of Computational Physics, 160, pp. 405-452
(2000)
2. Shu, C.W.: International Journal of Computational Fluid Dynamics (2003)
3. Lele, S. K.: , Journal of Computational Physics, 103, pp.16-44 (1992)
4. Gaitonde, D. V. and Visbal, M. R.: AIAA Paper 99-0557 (1999)
5. Jiang, G.-S. and Shu, C.-W.: Journal of Computational Physics, 126, pp. 200-
212 (1996)
6. Deng, X. G. and Zhang, H.: , Journal of Computational Physics, 165, pp. 22-44
(2000)
7. Deng, X. G. and Mao, M.: , Computational Fluid Dybnamics Journal, 13, pp.
173-180 (2004)
8. Deng, X. G., Liu, X. and Zhang, H.: , AIAA paper, 2005-5246, (2005)
9. Shu C.-W. and Osher, S.: , Journal of Computational Physics, 77, pp. 439-471
(1988)
10. Shu C.-W. and Osher, S.: , Journal of Computational Physics, 83, pp.32-78
(1989)
Implementation of an Enhanced Flux
Formulation for Unsteady Navier-Stokes
Solutions
G. Xia, S. Sardeshmukh, V. Sankaran and C. L. Merkle
Purdue University, West Lafayette, IN 47907, USA venke@purdue.edu
1 Introduction
Preconditioned time-marching CFD methods have become established as an

accurate and efficient framework for all Mach numbers [1]. However, unsteady
solution efficiency and accuracy suffer when the combination of low Mach
numbers and high Strouhal numbers is encountered, especially in the con-
text of high-fidelity turbulent and/or acoustics problems [1]. To counter this
difficulty, an improved discrete formulation that is valid for all steady and
unsteady regimes was proposed in an earlier article [2] for structured central-
differenced algorithms. In this article, we further extend the modified approach
to unstructured finite-volume formulations with implicit relaxation and test it
for accurately resolving turbulence dynamics using a second-order discretiza-
tion framework.
Implicit upwind methods on unstructured grids usually adopt point- or
line-relaxation procedures for the solution of the linear system. However, we
find that the enhanced scheme does not guarantee diagonal dominance that
is necessary for the stability of these schemes. Consequently, we retain the
traditional flux-difference formulation on the left-hand side and employ the
modified formulation only on the right-hand side. The inconsistent formula-
tion thereby retains the accuracy benefits of the improved formulation while,
at the same time, preserving diagonal dominance for stability. We employ
von Neumann stability analysis to confirm the overall stability properties of
the scheme and then demonstrate its application for the propagation of Tay-
lor vortices. Finally, we demonstrate the capability of the scheme to capture
turbulence dynamics of the decay of isotropic turbulence.

138 G. Xia, S. Sardeshmukh, V. Sankaran and C. L. Merkle
2 Computational Formulation
2.1 Preconditioned Equations
For unsteady flows, the governing equations are written in the so-called dual-
time form [1].
∂Qp ∂Q ∂Ei ∂Vi
Γp + + =H+ (1)
∂τ ∂t ∂xi ∂xi
where the τ represents the pseudo-time derivative, introduced purely for nu-
merical purposes. The terms comprising this pseudo-time term are:
ρ0p
   
p 0 ρT
Qp =  uj  Γp =  uj ρ0p ρδjk u j ρT 
T h0 ρ0p − (1 − ρhp ) ρuk ρhT + h0 ρT
In the above matrix, h0 is the total enthalpy and the term ρ0p represents
the scaled pseudo-property that is responsible for proper conditioning of the
system. The definition of this term is given by:
1 ρT (1 − ρhp )
ρ0p = 2
− (2)
Vp ρhT
where:
k1 V
Vp = min max V, , k2 V · Str , c (3)
Remin
Remin is the minimum cell-Reynolds number, Str is the physical Strouhal
number and k1 and k2 are scaling constants.
2.2 Traditional Flux Formulation
The traditional flux-difference formulation for an upwind finite-volume scheme

takes the following form for the interface flux:
1 1
Ẽ = (EL + ER ) − Γp |Γp−1 Ap |(QpR − QpL ) (4)
2 2
With appropriate definitions of the left and right states, the above scheme
can be viewed as first-, second- or higher-order accurate. Clearly, the second
term in the above equation which represents dissipation effects is a function
of the preconditioning matrix. As shown in previous research [1], this term
becomes overly dissipative for low Mach numbers and high Strouhal numbers.
Specifically, the scheme adds too much dissipation in the momentum and
energy equations, while adding the appropriate amount needed for stability
in the continuity equation.
Enhanced Flux Formulation for Unsteady Solutions 139
2.3 Enhanced Flux Formulation
The enhanced flux formulation consists in replacing the offending matrix dis-
sipation term with one that possesses the correct scaling properties. The cor-
responding flux becomes:
1 α
Ẽ = (EL + ER ) − Γp |Λ|(QpR − QpL ) (5)
2 2
where Λ = diag(σ[Γp−1 Ap ], uδkk , u). In other words, the spectral radius of
the preconditioned Jacobian is used only in the dissipation terms of the con-
tinuity equation, while the standard convective velocity scale is used in the
momentum and energy equations. Thus, the offending dissipation terms in
the momentum and energy equations are effectively removed in this formula-
tion, while the proper definition is retained in the continuity equation. The
enhanced formulation thus insures optimal accuracy for steady, viscous or
unsteady computations.
2.4 Linear Solution Procedure
The modified dissipation does not guarantee diagonal dominance as required

by Gauss-Siedel methods. Indeed, von Neumann stability results (not shown
here for brevity) indicate that the scheme is unconditionally unstable in cer-
tain Mach-Strouhal regimes. Consequently, we retain the traditional matrix
dissipation form on the left-hand side, and employ the enhanced dissipation
on the right side alone. A further aspect of this term is the choice of the α pa-
rameter. Again, stability theory indicates that the scheme is stable for α ≤ 1.
In our current results, we employ α = 0.5.
3 Results
3.1 Taylor Problem
To test the artificial dissipation control, we show results from the numerical
simulation of the Taylor problem. The Taylor problem represents an analytical
solution of the incompressible Navier-Stokes equations for an infinite array of
counter-rotating vortices (see Fig. 1), whose amplitudes decay with time:
1
u = − cos x·sin y·e−2νt , v = sin x·cos y·e−2νt , p = − (cos 2x+cos 2y)·e−4νt (6)
4
Figure 2 shows the pseudo-time convergence for several sequential physical
time-steps using different schemes. It is evident that the scheme without any
preconditioning scaling or only inviscid scaling performs relatively poorly. In
particular, the inviscid choice appears to be very stiff for this choice of time-
scale (CF Lu = 0.1) and performs worse when the time-scale is made smaller.
The standard matrix dissipation scheme and the modified dissipation scheme
both perform very well in comparison, although it is worthwhile pointing out

that the modified scheme performs slightly worse, a result of the inconsistent
linearization used to insure diagonal dominance.
Fig. 1. Taylor vortex flowfield with periodic boundaries in all directions.
Fig. 2. Pseudo-time convergence for several time-steps of various numerical schemes.
Accuracy of the different schemes is shown in Fig. 3 for two different

Reynolds numbers, infinity and Re = 10. The physical time-step in each case is
varied so that the viscous time-scale is controlled for the low Reynolds number
case. In both cases, it is clear that the modified dissipation scheme provides
Enhanced Flux Formulation for Unsteady Solutions 141
the best comparison with the exact solution and out-performs the traditional
matrix dissipation scheme by a significant margin. It is also noteworthy that
the no-preconditioning as well as the inviscid-preconditioning cases are both
very inaccurate, in part because of convergence issues and in part because of
the implications on the sizing of the dissipation terms.
Fig. 3. Comparisons of kinetic energy decay of various schemes for Re = ∞ and

Re = 10.
3.2 Decay of Isotropic Turbulence
As a second example to test the capabilities of the improved unsteady formu-

lation, we consider the decay of homogeneous turbulence downstream of a grid
in uniform flow [3]. The computational results are obtained by the addition of
a DES turbulence model to the equations [4, 5]. The results are interpreted as
the time evolution of spatially homogeneous turbulence in the time interval
determined by the streamwise distance and the mean flow velocity. The com-
putations are initialized using experimental data at a given non-dimensional
time and the predictions at a later time are then compared with the experi-
ments. A uniform 64 x 64 x 64 grid is used in the numerical simulation with
a mean Mach number of 0.001 and a Reynolds number of 34,000. We note
that, since no walls are present, the DES calculation represents a pure LES
solution in this case.
Figure 4 shows comparisons of the DES predictions with the experimental
data for different schemes (left) and for different values of the CDES parame-
ter. It is evident that the modified dissipation scheme provides the best predic-
tion and essentially preserves the accuracy of all the ”represented” modes in-
dicated in the initial data. In contrast, the standard matrix dissipation scheme
as well as the other preconditioning choices indicate that the inertial scales
are strongly damped. The calibration of the CDES parameter, shown only for
the enhanced scheme, indicates that the optimum choice of this parameter is
close to 0.5 for this scheme, indicating that the choice of the model parameters
are intrinsically tied to the accuracy of the underlying scheme.
Fig. 4. Comparison of DES predictions with experiments. Left side shows results
for different numerical schemes, right-side shows results for different values of CDES .
4 Summary
A modified dissipation scheme that promises uniform accuracy at all Mach

numbers, Reynolds numbers and Strouhal numbers is presented. In order to
preserve diagonal dominance of the implicit linear operator, we retain the
traditional flux-difference treatment on the left-hand-side and implement the
improved scheme only on the right-hand side. The method is tested with
exact results of the Taylor problem and for the case of decay of homogeneous
turbulence. Comparisons with the more traditional schemes suggest that the
modified scheme is capable of improved accuracy and efficiency, especially for
unsteady problems.
References
1. S. Venkateswaran and C. L. Merkle, ’Efficiency and Accuracy Issues in Navier-
Stokes Computations’, AIAA Paper 2000-2251, 2000.
2. Sankaran, V. and Merkle, C. L., ’Artificial Dissipation Control for Unsteady
Computations’, AIAA Paper 2003-3695, 2003.
3. Comte-Bellot, G., and Corrsin, S., ’Simple Eulerian Time Correlation of Full- and
Narrow-Band Velocity Signals in Grid-Generated Isotropic Turbulence’, Journal
of Fluid Mechanics, Vol. 48, pp. 273-337, 1971.
4. Strelets, M., ’Detached Eddy Simulation of Massively Separated Flows’, AIAA
Paper 2001-0879, 2001.
5. Xia, G., Sankaran, V., Li, D., and Merkle, C. L., ’Modeling of Turbulent Mixing
Layer dynamics in Ultra-High Pressure Flows’, AIAA Paper 2006-3729, 2006.
Computation of Eigenspaces of Hyperbolic
Systems
S.K. Godunov1 , O.B. Feodoritova2 and V.T. Zhukov2

1
Sobolev Institute of Mathematics, Siberian Branch of Russian Academy of
Sciences, Pr. Koptuga, 4, 630090, Novosibirck, Russia. Godunov@math.nsc.ru
2
Keldysh Institute of Applied Mathematics, Russian Academy of Sciences,
Miusskaya sq. 4, 125047, Moscow, Russia. Feodor@kiam.ru , Zhukov@kiam.ru
Summary. A method for calculating principal invariant subspaces of symmetric

hyperbolic systems arising in fluid dynamic, electromagnetic and elasticity prob-
lems is presented. Such subspaces correspond to a certain set of minimal nonzero
eigenvalues and contain sufficiently smooth functions, low-frequency modes. Diffi-
culties related to the presence of an infinite-dimensional null space in the differential
operator are successfully handled. The efficiency of the algorithm is demonstrated
for acoustics equations.
Key words: symmetric hyperbolic systems, method for calculating invariant sub-
spaces, calculation of eigenvalues and eigenfunctions
1 Introduction
Eigenvalue problems for differential equations have numerous practical appli-

cations and are of interest for the analysis and design of discrete schemes.
We discuss here an approach [1], [2] to the solution of eigenvalue problems
for skew-symmetric differential operators related to symmetric hyperbolic sys-
tems that arise in a wide range of applications. For such differential operators,
we represent an approach for calculating discrete approximations of the main
invariant subspaces that contain smooth functions, low-frequency modes. We
give here the main framework and illustrate it by a numerical example.
2 Problem Statement
Consider a linear hyperbolic system in a d-dimensional space Rd

d
∂u X ∂u
A + Bk = 0. (1)
∂t ∂xk
k=1

144 S.K. Godunov, O.B. Feodoritova and V.T. Zhukov
Here u = (u1 , . . . , um ) is the state vector and A, Bk are m × m Hermitian

matrices such that A = A(x1 , . . . , xd ) = A∗ > 0 and Bk = Bk∗ = const.
We are interested in computation for the system (1) a certain group of
minimal (nonzero) eigenvalues and a basis in the corresponding invariant sub-
space. The problem reads: find u 6= 0 and λ such that
Du = λu, (x1 , . . . , xd ) ∈ Ω ⊂ Rd , 0 < |λ| ≤ r1 , Φu|∂Ω = 0. (2)
Here D is a differential operator of the form:
d
X ∂
D=− A−1 Bk , (3)
∂xk
k=1
and the boundary conditions Φu imposed on ∂Ω are assumed to be conserva-

tive; i.e., it holds that
Z X d

ξi Bi u, u dS = 0, (4)
∂Ω i=1
where (ξ1 , . . . , ξd ) is the outer normal to ∂Ω.

The upper bound r1 in (2) implicitly determines the dimension of the
desired subspace. This can be a rather approximate bound, which can be
refined in the solution process.
We introduce the energy inner product [u, v]A andR the energy norm kukA
in the space of vector functions L2 (Ω): [u, v]A = (Au, v)dΩ, kuk2A =
Ω
[u, u]A . It can be shown that D is a skew-symmetric operator in the energy
inner product: D∗ = −D. In what follows, we use this property.
We assume that the characteristic equation of the hyperbolic system (1)
has a few zero roots and hence the system has vertical characteristics. This
implies the spectrum of D contains zero eigenvalue of infinite multiplicity.
The existence of vertical characteristics for the system (1) is ensured by spe-
cial properties of the constant matrices Bk . These properties hold in many
problems, particularly, in the case of the Maxwell and acoustics equations,
which are considered below as examples.
1. The Maxwell equations in R3
∂(E) ∂(µH)
= rot H, = −rot E
∂t ∂t
are a hyperbolic system of type (1) for the state vector u = (Ex , Ey , Ez , Hx ,
Hy , Hz ). The characteristic equation of this system has six real roots: ±c, ±c,
√
0 and 0, where c = 1/ µ is the speed of light.
2. The acoustics equations in R2 for a state vector u = (U, p) = (u, v, p)
∂U 1 ∂p
ρ0 + grad p = 0 , + div U = 0 (5)
∂t ρ0 c20 ∂t
are also a hyperbolic system. The characteristic equation of this system has
three real roots: ±c0 and 0, where c0 is the speed of sound.
Computation of Eigenspaces of Hyperbolic Systems 145
For the Maxwell and the acoustics equations, the relations on the vertical
characteristics have, respectively, the form
∂ ∂ ∂
div(E) = 0, div(µH) = 0 and rot U = 0.
∂t ∂t ∂t
To avoid the difficulties related to the null space of the operator D, the
system (1) is often reduced to the second-order equations, which have no ver-
tical characteristic. For instance, the acoustics equations (5) can be reduced
to the wave equation. However, this transition may cause difficulties in set-
ting up boundary conditions, writing a scheme, etc., in the case of variable
coefficients describing the medium. Note that hyperbolic formulation is more
natural for application of theory and numerical methods and is of great inter-
est for problems in magnetohydrodynamics.
3 Computations
The problem of calculating invariant subspaces for differential operators is
not widespread in computational practice, especially for nonself-adjoint oper-
ators. The results presented below are obtained in numerical experiments. We
try to start directly from the first-order system (1) and develop a technique
that makes it possible to eliminate the influence of the zero frequency on the
numerical process. For simplicity we consider a model problem for the system
(5), though the numerical method discussed below is also applicable to the
more general cases. The model problem is the acoustics equations (5) with
the constant coefficients c0 = 1 and ρ0 = 1 in Ω = (0, π)2 with the condition
p = 0 imposed on the boundary ∂Ω.
For the differential operator D corresponding to this boundary value prob-
lem and acting on a function u = (u, v, p) according to the rule Du =
−(px , py , ux + uy ), (x, y) ∈ Ω, we find a grid approximation of the invariant
subspace SD associated with the eigenvalues in the interval 0 < |λ| ≤ 4. From
this problem it is easy to find analytically all eigenvalues and eigenfunctions
[1], and one may verify dim(SD ) = 16.
To construct discrete approximations, we introduce a square grid with
a mesh size h. For simplicity of the presentation the approximations of D
are described for a strictly interior grid cell. We introduce the difference op-
erators D1 and D2 . The operator D2 is obtained by approximating all the
spatial derivatives at each cell by central divided differences. On the grids in-
troduced above, D2 inherits skew-symmetry inherent in D and approximates
D to O(h2 ) accuracy. The operator D1 is the upwind difference operator that
approximate D to O(h) accuracy. Both operators are supplemented by con-
servative boundary conditions.
Let us briefly describe the spectra of D, D2 , D1 for the model problem. In
the complex plane, the spectrum of D consists of the point (0, 0) of infinite
multiplicity and points on the √ imaginary axis. For this example the exact
eigenvalues λ are simply ±i m2 + n2 , 0 < m, n ∈ N and among them
exactly 16 values belong to the interval 0 < |λ| ≤ 4, but only 10 of them are
distinct. Remark also the zero eigenvalue is separated from the other ones.
The spectrum of D2 consists of the point (0, 0) of high multiplicity and
points on the imaginary axis. On the 32×32 grid, the spectrum consists of 3072
eigenvalues (counting multiplicities). There are 1122 eigenvalues in 0 < |λ| ≤ 4
of which only 96 are nonzero. The distinct eigenvalues form 9 conjugate pairs,
which are given in Table 1. The table also lists their multiplicities and the
number of pairs that approximate the eigenvalues of D.
Table 1. The eigenvalues of the operators D and D2 in 0 < |λ| ≤ 4
Spectrum D Spectrum D2 Multiplicity Number

(number of pairs) of proper pairs
±0.9984i 4 0
±1.4142i ±1.4119i 4 1
±1.9872i 4 0
±2.2361i ±2.2239i 8 2
±2.8284i ±2.8103i 4 1
±2.9568i 4 0
±3.1623i ±3.1208i 8 2
±3.6056i ±3.5625i 8 2
±3.8980i 4 0
The remaining eigenvalues are of grid origin. They are associated with os-
cillating grid modes. Moreover, such modes are also present in the eigenspaces
corresponding to the eigenvalues of D2 in 0 < |λ| ≤ 4. It is practically im-
possible to distinguish the desired approximations of the smooth part of the
spectrum due to presence of such spurious solutions among the eigenfunctions
of D2 . These spurious modes have no relation to the differential problem; actu-
ally, the presence of such modes speaks against D2 . However, it is the operator
D2 that inherits the skew-symmetry of D.
The spectrum of D1 has a rather complex structure. Figure 2a shows the
entire spectrum of D1 for the 32 × 32 grid. On the whole, the spectrum of D1
does not approximate that of D. In contrast to D, the operator D1 has a lot
of real eigenvalues ( O(Nh2 ) ) and part of then are located in the vicinity of
the origin. However, the spectrum of D1 includes some points that correspond
to the eigenvalues of D in 0 < |λ| ≤ 4. It is seen from Figure 2b that each
point in the spectrum of D marked by a circle is associated with a point in
the spectrum of D1 that has the same ordinate up to several digits. When the
grid is refined, these points in the spectrum of D1 approach the corresponding
points in the spectrum of D.
Let us also remark that the operator D1 can be used in an explicit time-
integration scheme with stability restriction τ = O(h) but D2 is an unpractical
operator in such computations.
Computation of Eigenspaces of Hyperbolic Systems 147
15 a 4 b
x
3 x
10
x
2 x
5 x
1
0 0 xxxxxxxxxxxxxxx x x xxxxxx x x
-1
-5 x
-2
x
-10 x
-3 x
x
-15 -4
-40 -30 -20 -10 0 -0.4 -0.2 0
Fig. 1. a) The eigenspectrum of D1 , b) 10 points (o) of the eigenspectrum of D

and their approximating points (x) of Sp(D1 )
Our aim is to construct an orthogonal basis YD2 such that span(YD2 )

is O(h2 ) approximation of the desired eigenspace SD . Firstly, a basis YD1
such that span(YD2 ) = SD + O(h) is computed by an iterative procedure [1].
Assume that a fairly arbitrary, though reasonable set of frequencies {ωj , j =
1, ..., J} is given in the prescribed interval 0 < |λ| ≤ 4. A set of vector pairs
{Yω } is also given as initial rough approximation to YD1 . For instance, it is
possible to choose a repeated set of collinear vector pairs. For each frequency
ω = ωj and the corresponding vector pair Yω the solution Zω (t) of the problem

dZω cos (ωt) −sin (ωt)
= D1 Zω + Yω , Z(0) = 0 (6)
dt sin (ωt) cos (ωt)
is found by a time-integration scheme up to tω = 2π/ω. This stage ensures
the convergence to the eigenvalues ±iλ with the modulus λ closest to ω [2].
The next iteration of the basis YD1 is obtained by a collection of all new
vectors Yω = tω Zω (tω ). Then YD1 is extended using the operator D2 , and the
new frequencies {ωj } are computed. If the prescribed accuracy specified by
(∗)
the norm of D1 -residual D1 YD1 − YD1 [YD1 D1 YD1 ] is not attained, then this
iteration step is repeated. Otherwise, the basis YD2 is calculated by treatment
of the basis YD1 and its D2 –residual.
For this example we find an orthonormal basis in the eigenspace of dimen-
sion M = 16, starting with the initial set of frequencies Ω0 = {ω1 = 1, ω2 =
2, ω3 = 3, ω4 = 4}. The calculations are performed for a sequence of square
grids with the sizes 16 × 16, 32 × 32 and 64 × 64. A measure of the error is the
angle between the constructed subspace and the subspace span(Yex ) , where
the basis Yex is obtained by the restriction of a basis in the subspace SD to
a grid and by the subsequent orthonormalization. Table 2 characterizes the
accuracy of the computed eigenspaces span(YD2 ) and span(YD1 ). It lists the
sines of the angles between exact and approximate subspaces for the dimen-
sions M = 16 and M = 6. In the latter case, the subspace contains the first 6
eigenfunctions (which are the smoothest) in the 16-dimensional subspace. We
observe that the error converges to 0 at the rate O(h2 ) and O(h) for YD2 and
YD1 respectively as the grid is uniformly refined. In addition Figure 2 shows

the convergence rate O(h2 ) for the five conjugate eigenpairs of D2 with the
distinct modulus.
Table 2. The accuracy of eigenspace computations
16-dimensional eigenspace 6-dimensional eigenspace

Grid sin hYex , YD2 i sin hYex , YD1 i sin hYex , YD2 i sin hYex , YD1 i
16 × 16 0.16e − 01 0.75e − 01 0.78e − 02 0.53e − 01
32 × 32 0.39e − 02 0.37e − 01 0.19e − 02 0.26e − 01
64 × 64 0.96e − 03 0.18e − 01 0.48e − 03 0.13e − 01
first
-1
X second
10 Eigenpairs:
X third
-2 X
fourth
10
fifth
X
2
-3
10 X
1
X
-4
10
64 128 192 256

1/h
Fig. 2. Convergence plot for the first five conjugate eigenpairs
This research is supported by RFBR and NWO under grant 047.016.003,

and Division of Mathematical Sciences of RAS under Program vs 3.1 and vs4.
References
1. Godunov, S.K., Feodoritova, O.B. and Zhukov, V.T.: A method for computing
invariant subspaces of symmetric hyperbolic systems. Comput. Maths and Math.
Phys., 46(6), 971–982 (2006), Pleiades Publishing Inc., 2006
2. Godunov, S.K. and Selivanova, S.V.: Experiments Involving Resonance for Spec-
tral Analysis of Skew-Symmetric Operators. Sib. Zh. Vychisl. Mat., 9(2), 123–
136 (2006)
A Proposed Cure to the Carbuncle
Phenomenon
Farzad Ismail1 , Philip L. Roe2 , and Hiroaki Nishikawa3

1
Dept. of Aerospace Engineering, University of Michigan jaymzz@umich.edu
2
Dept. of Aerospace Engineering, University of Michigan philroe@umich.edu
3
Dept. of Aerospace Engineering, University of Michigan hiroakin@umich.edu
A new finite volume methodology is introduced to combat the carbuncle. The

method features a more accurate treatment of entropy in the flux formulation
at the cost of a small computational overhead. This new flux function is tested
on a hypersonic flow past a circular cylinder on both structured quadrilateral
and unstructured triangular grids, producing encouraging results.
1 Introduction
Numerical shock prediction is a very important aspect of computing aerody-
namic flows, and shock capturing finite-volume methods are commonly used
to predict shocks in various situations with considerable success. However, it
seems that most shock capturing methods fall short in predicting very strong
shocks, which is a crucial element in designing hypersonic vehicles. Except for
a few notoriously diffusive schemes ([21] and the simplest version of Harten-
Lax-van Leer ([7]), most schemes4 exhibit some form of anomaly when pre-
dicting strong shocks. The commonest of these is the carbuncle phenomenon,
produced when computing a hypersonic flow past a blunt body such as a cir-
cular cylinder. Instead of having a smooth bow shock profile upstream of the
cylinder, the solution features a pair of oblique shocks ahead of the stagna-
tion region (Fig 2). Such a solution is actually a true solution of the Euler
equations, and can even be produced experimentally [2]. Many have proposed
cures to the carbuncle problem [15], [14], [12], [13], [11], [3], [20] but none are
universally accepted, and most papers begin by criticizing previous work.
The most common view follows Quirk [16] in supposing that the carbuncle
manifests a two-dimensional numerical instability of the Euler equations, often
4
There are proposals to adopt a hybrid of very dissipative and less dissipative
fluxes, deploying the former near the shock and the latter away from shock but
the basis of the switch is somewhat ad hoc. Furthermore, it is not clear how any
switch would work for complex problems like shock-boundary layer interactions
or shock-contact interactions.

150 Farzad Ismail, Philip L. Roe, and Hiroaki Nishikawa
2
3
Mach
20
2
18
1
16
1
14
YCoord
YCoord
12
0 0
10
8
-1
6
-1
4
-2
2
-3 0
-2
-4
-4 -3 -2 -1 0 1 2 3 4 -3 -2 -1 0 1 2
XCoord XCoord
Fig. 1. Quadrilateral grids on a circular Fig. 2. Results of the original Roe flux
cylinder with 80 x 160 cells. with M=20. Contours of Mach number
and velocity vectors.
supposed to be an odd-even decoupling. However, Roe et al [17] note that the

phenomenon occurs at all frequencies and that a version can be found in
one dimension. If the physical viscosity is included as in the Navier-Stokes
equations, the tendency to form a carbuncle is reduced, but it disappears
only at very low Reynolds number [14]. Nor does it help to include the real
gas effects that accompany very strong shocks in the real world [4], [5], [6].
Here, we propose to combat the carbuncle phenomenon by strongly en-
forcing entropy stability. First, we will describe some motivating discoveries.
2 The Root of the Carbuncle ?
A simple setting for the carbuncle phenomenon, proposed by [3], is a steady

one-dimensional shock on a rectangular two-dimensional grid. Such a car-
buncle evolves in three universal but distinct stages: “pimples”, “bleeding”
and “carbuncle” [17]. The “pimples” are an initial instability largely confined
to the vicinity of the shock, whereas the “bleeding” sees these instabilities
propagated downstream as layers of alternately high and low velocity. After
sufficient amplification the low velocities develop regions of reversed flow that
break out ahead of the shock to form the “carbuncle”. The second and third
stages weakly satisfy the Euler equations but are only observed experimen-
tally in some artificial setup. Accordingly, we seek to prevent the instability
at the pimple stage, by improving the basic process of shock-capturing.
Since anomalies appear in both entropy and vorticity, Ismail [9] investi-
gated preventing the carbuncle by controlling either vorticity or entropy. He
found vorticity control to be ineffective, so we focus here on control of entropy.
Also, even in one dimension, very strong shocks can be unstable both in the
sense of entropy [1], and in spontaneous relocation [18], which suggests a link.
A Proposed Cure to the Carbuncle Phenomenon 151
3 The Entropy-Stable (ES) Flux

Entropy stability can be studied in one dimension by applying a finite volume
method to the conservation laws ∂t u + ∂x f(u) = 0. Semi-discrete entropy con-
servation requires [22, 1] the numerical interface flux f∗ to satisfy [v]T f∗ = [ρu]
γ−s
where v = ( γ−1 − 12 ρp (u2 ), ρu ρ T 5
p , − p ) are the entropy variables and the square
bracket denotes a difference operator. The following entropy-conservative flux
[18] is explicit and numerically well-formed.
 
ρ̂û
f∗ =  ρ̂û2 + pˆ1  = fC (1)
ρ̂ûĤ
The quantities (ˆ·) are averaged quantities at the flux-interface satisfying

z¯2 z¯3
ρ̂ = z¯1 z3ln , û = , pˆ1 =
z¯1 z¯1
γ + 1 z3ln γ − 1 z¯3 1 γ pˆ2
pˆ2 = + , Ĥ = û2 + (2)
2γ z1ln 2γ z¯1 2 γ − 1 ρ̂
ln
where¯and represent arithmetic and logarithmic6 means [18] and
√
r r
ρ ρ
z1 = , z2 = u, z3 = ρp (3)
p p
To ensure that entropy is generated with the correct sign, we add upwind
terms to give the entropy-stable flux (one for which [v]f∗ ≤ 0)
1
f∗ (uL , uR ) = fC (uL , uR ) − R̂|Λ̂Ŝ|R̂T [v] (4)
2
where R and Λ denote the right eigenvectors and the diagonalized eigenvalues
ρ (γ−1)ρ ρ
of the Euler equations. The scaling factor S = diag( 2γ , γ , 2γ ) relates to
−1 T
the differential identity [1], R du = SR dv. The new flux function coin-
cides to second order accuracy with the original Roe-flux7 , but is additionally
constrained to capture exactly pure contact discontinuities of any strength.
This property, which guarantees accuracy in boundary layers, has sometimes
been thought to induce carbuncles. To enforce the contact-capturing property
1
[18], the averaged speed of sound must be evaluated from â = ( γρ̂pˆ1 ) 2 and the
density averaged using a logarithmic mean. Based on these restrictions and
for computational economy, it was proposed in [9] that the averages in the
dissipative flux are exactly the averages in the entropy-conserving flux.
5
This is the only choice of entropy variables that can be used in the Navier Stokes
equations [8] and s = ln p − γln ρ is the physical entropy.
6
The logarithmic mean is here defined as L(x, y) = (x − y)/(ln x − ln y), and has
an efficient series representation if x ' y.
7
Recall that the original Roe-flux [19] is f∗ = f̄ − 21 R̂|Λ̂|R̂−1 [u]
2
Mach
20 0.75
18
1
16
14
YCoord
12
0.5
0
P
10
8
6 0.25
-1
4
0
-2 0
-3 -2 -1 0 1 2 -2 -1.75 -1.5 -1.25

XCoord XCoord
Fig. 3. The ES flux at M=20, on the Fig. 4. Spurious overshoots of pressure

grid of Figures 1 and 2. along the centerline.
For conciseness, we have only presented the flux-function in one dimension.

However, because it is a finite volume method, the extension to arbitrary grids
in higher dimensions is straightforward (for details see [9]).
4 Numerical Results
The test case is the steady-state flow past a two dimensional cylinder. Various
grids and Mach numbers have been employed; a typical structured quadrilat-
eral grid is shown in Figure 1. In all cases reported here, we used a first
order explicit method with ν = 0.2 and the code was run until the residual
is of O(10)−8 . Second-order results were reported in [9]. All of our results for
structured grids indicate that results from the entropy-stable flux (for exam-
ple Figure 3) are quite free of the carbuncle phenomenon. However, the shock
profile is slightly broader with the introduction of a few intermediate cells
(Figure 4) and this may be the price we have to pay.
Previous studies [3] suggest that the carbuncle can be made to appear if the
cell aspect is increased, with the short dimension parallel to the shock. This
is consistent with a view that one needs numerical damping in the tangential
direction. However, our entropy-stable computations were not sensitive to this
aspect of the grid (compare Figures 5 and 6 with Figure 3).
Also, the profile around the shock exhibits spurious oscillations. This is
because our flux function guarantees only the correct sign of entropy pro-
duction but not necessarily the correct amount. The actual production is of
order δ 2 whereas the required production is of order δ 3 . To achieve mono-
tonicity, a flux function must generate ‘enough’ entropy production across
a shock. Flux functions that have this property are called entropy-consistent
(EC) fluxes [18]. We therefore modify the eigenvalues of equation (4) such that
|Λ̂| = |Λ̂| + α|[Λ]| [10]. For the choice of α = 0.2, the entropy-consistent flux
A Proposed Cure to the Carbuncle Phenomenon 153
0.6 0.6
0.4 0.4
0.2 0.2
YCoord
YCoord
0 0
-0.2 -0.2
-0.4 -0.4
-0.6 -0.6
-3 -2.5 -2 -1.5 -1 -3 -2.5 -2 -1.5 -1

XCoord XCoord
Fig. 5. ES flux with 20 x 200 cells. Fig. 6. ES flux with 20 x 600 cells.
2
Mach
20 0.75
18
1
16
14
YCoord
12
0.5
0
P
10
6 0.25
-1
4
2
0
-2 0
-3 -2 -1 0 1 2 -2 -1.75 -1.5 -1.25

XCoord XCoord
Fig. 7. EC-flux at M=20. Fig. 8. Pressure profile of EC-flux.
produces monotone, carbuncle-free solutions on quadrilateral grids (Figures

7-8).
However, the carbuncle is not yet vanquished. On unstructured triangular
grids, such as Figure 9, we discovered that it is produced even by the entropy-
consistent flux (although in less severe form than the regular flux). It seems
to be reduced (rather surprisingly, and at the expense of increasing spurious
overshoots) if we choose smaller values of α and is least when α → 0. This
strongly suggests that entropy stability is not the only cause of carbuncles.
One possibility may lie in the nature of the finite-volume method itself, with
its reliance on pairwise interactions between cells, and therefore merely one-
dimensional physics. This may be responsible for the shock being very sharp
but poorly aligned. In fact, a better solution can be obtained by creating a
better alignment of the grid with the shock, although we do not have suf-
ficient space to show these results. Analysis in [18] offers the possibility of
multidimensional extensions, that should be less dependent on the grid.
Fig. 9. An unstructured grid. Fig. 10. ES flux on the grid of Fig. 9.
References
1. Barth, T. In Num. Meth. for Gasdyn. Syst. On Unstructured Meshes. 1999.
2. Bogdonoff, S., and Vas, I. J. Aero. Sci 26 (1959), 584.
3. Dumbser, M., Moschetta, J., and Gressier, J. JCP 197 (2004), 647.
4. Edwards, J. In CFD Conference (2001), no. A01-31131, AIAA Conference.
5. Gnoffo, P. Ann. Rev. Fluid Mech. 31 (1999), 459–494.
6. Gnoffo, P., and White, A. No. 2004-2371, AIAA Thermophysics Conference.
7. Harten, A., Lax, P.D. and van Leer, B. SIAM Review (1983)
8. Hughes, T. Comp. Meth. App. Mech. Eng. 54 (1986).
9. Ismail, F. PhD thesis, The University of Michigan, 2006.
10. Ismail, F., and Roe, P. in preparation for JCP .
11. Kim, S.-s., Kim, C., Rho, O.-H., and Hong, S. JCP 186 (2003), 342.
12. Lin, H.-C. JCP 117 (1995), 20–27.
13. Liou, M. JCP 160 (2000), 623–648.
14. Pandolfi, M., and d’Ambrosio. JCP 166 (2001), 271–301.
15. Peery, K., and Imlay, S. No. 88-2924, AIAA Conference.
16. Quirk, J. Int. J. Num. Meth. Fluids 18 (1994), 555–574.
17. Roe, Nishikawa, Ismail, and Scalabrin. No. 2005-4872, AIAA Conference.
18. Roe, P. to be published in JCP .
19. Roe, P. J. Comp. Physics 43 (1981), 357–372.
20. Sanders, R., Morano, E., and Druguet, M. JCP 54 (1994).
21. Steger, J., and Warming, R. JCP 40 (1981), 263–293.
22. Tadmor, E. In Acta Numerica. 2002.
The High Order WLSQR Scheme and its
Applications in Turbomachinery
Jivrı́ Fürst
Faculty of Mech. Engineering, Czech Technical University in Prague,

Karlovo námvestı́ 13, 121 35 Praha 2, Czech Republic
Jiri.Furst@fs.cvut.cz
1 Introduction
This article describes the development of a high order numerical method for
the solution of compressible transonic flows. The discretisation in space is
based on the standard finite volume method of Godunov’s type. A higher
order of accuracy is achieved by the piecewise polynomial interpolation similar
to the ENO or weighted ENO method (see e.g. [8]).
The here mentioned method is developed with the aim to simplify the
implementation of the reconstruction procedure especially for the case of un-
structured meshes. The reconstruction procedure uses single stencil and com-
putes the interpolation polynomial by minimizing the weighted interpolation
error over the cells in this stencil.
The complete finite volume scheme equipped with the piecewise linear
reconstruction has been successfully used for the solution many transonic flow
problems (see e.g. [4, 3]). This article presents the extension of the method to
piecewise parabolic case with the application for the solution of flows through
turbine cascades.
The flow is described by the set of Euler or Navier-Stokes equations in
conservative form
Wt + F (W )x + G(W )y = F v (W )x + Gv (W )y + S(W ), (1)
T
where W = [ρ, ρu, ρv, e] is the vector of conservative variables, F (W ) and
G(W ) are the inviscid fluxes, F v (W ) and Gv (W ) are the viscous fluxes (F v =
Gv = 0 for the case of Euler equations) and S(W ) is a source term [2].
The equations equipped with proper boundary conditions are solved nu-
merically using an unstructured mesh and a finite volume scheme with all
unknowns located at cell centers. The fluxes through the cell interfaces are
approximated by the Gauss quadrature with the physical fluxes replaced by
the numerical ones

156 Jivrı́ Fürst
Z J
X
(F (W ), G(W )) · dS ≈ ωq F AU SM P W + (Wijq
L R
, Wijq , Sijq ). (2)
Ci ∩Cj q=1
L/R
Here Wijq denotes the values of the vector of unknowns interpolated to the
Gauss point q of the interface Ci ∩Cj from the left cell (superscript L) or from
the right cell (superscript R). The resulting finite volume scheme for inviscid
case can be then written in semi-discrete form as
J
dWi XX
|Ci | =− ωq F AU SM P W + (Wijq
L R
, Wijq , Sijq ). (3)
dt
j∈Ni q=1
Here Ci is the i-th cell, Wi = Ci W (x, t)dx, and Ni = {j : dim(Ci ∩ Cj ) = 1}.

R
L
The basic first order scheme can be obtained by setting J = 1, Wijq = Wi ,
R
and Wijq = Wj .
2 The WLSQR Interpolation

However the basic first order scheme posses very good mathematical proper-
ties, it is well known, that it is very diffusive. Therefore people prefer to use
higher order schemes, especially for the viscous flow calculations. The higher
order scheme can be constructed within this framework simply by improving
the interpolation of W L/R . There exists several methods for the construction
of a stable interpolation, the most known are the limited least squares of Barth
[1], the ENO/WENO schemes [8], or the TVD schemes [7].
The use of limiters as in TVD or Barth’s scheme usually cut the order of
accuracy near extrema and may also hamper the convergence to steady state.
On the other hand, the implementation of ENO/WENO schemes is relatively
complicated for unstructured meshes. Therefore a novel reconstruction proce-
dure has been introduced in [3] where the interpolation polynomial Pi (x; φ)
for the cell Ci and the component φ of W is constructed by minimizing the
weighted interpolation error 1
" Z !#2
X
err := wij P̃ (x; φ) dx − |Cj |φj (4)
j∈Mi Cj
with respect to the conservativity constraint

Z
Pi (x; φ) dx = |Cj |φj . (5)
Cj
The weights wij are chosen in such a way, that w is big whenever the solution
is smooth and w is small when the solution is discontinuous. The single stencil
Mi is selected according to the order of the polynomial P .
1
Herefrom comes the name of the method - the Weighted Least Square Recon-
struction.
The WLSQR Scheme 157
2.1 The Second Order Scheme
The formally second order scheme can be obtained by using linear polynomials
Pi . For this case, the choice of Mi := M1i = {j : Ci ∩Cj 6= ∅} (i.e. cells touching
Ci at least by a vertex) has been tested together with
v
u h−r
wij = t , (6)
u
φi −φj p

h + hq
with h being the distance between cell centers of Ci and Cj and p = 4, q = −3,
and r = 3. The analysis of simplified cases has been carried out in [4] showing
the stability of WLSQR interpolation for special discontinuous data.
2.2 The Third Order Scheme
The scheme can be extended to formally third order of accuracy by using

quadratic polynomials Pi . It is also necessary to widen the stencil to Mi :=
M2i = M1i ∪ {j : Cj ∩ M1i 6= ∅} (i.e. the stencil is extended by the cells touching
M1i ). Although there are no analytical results for quadratic reconstruction, the
same definition of w has been used successfully.
2.3 Numerical experiments with the WLSQR scheme
The piecewise linear and piecewise quadratic reconstruction were numerically

analyzed for the case of scalar linear and nonlinear initial value problem in
[4]. Next, the more complicated case of the transonic flow in a 2D channel was
examined in [4]. However, the order of accuracy was impaired by the polygonal
approximation of the boundary. Therefore another numerical experiment is
carried out in this article. In order to be able to compare the numerical solution
to the exact one, the Ringleb’s flow [5] was chosen. The domain was bounded
by the axis x, by two streamlines defined by k = 0.5 and k = 0.8 and by the
outlet part defined by q = 0.4 (see [5] for the definition of k and q).
The numerical solution was obtained using three levels of unstructured
meshes, the coarse one with 1004 triangles, the middle one with 3936 triangles
and the fine one with 15780 triangles. Moreover, three different variants WL-
SQR interpolations were used: I1B1 with piecewise linear interpolation and
boundary approximated by linear segments, I2B1 with piecewise quadratic
interpolation with boundary approximated by linear segments, and I2B2 with
piecewise quadratic interpolation with boundary approximated by second or-
der parabolas. For the I1B1 case the midpoint rule was used for the evaluation
of the flux integrals whereas for the I2B1 and I2B2 cases the two-point Gauss
integration formula was employed. Figure 1 shows the isolines of Mach number
obtained with the I2B2 variant of the scheme on the coarse mesh. Moreover, it
compares the distribution of the entropy along the left wall for three variants
of the scheme and three meshes.
-0.3346 I1B1 coarse

I2B1 coarse
-0.3348 I2B2 coarse
I1B1 middle
-0.335 I2B1 middle
I2B2 middle
-0.3352 I1B1 fine
I2B1 fine
I2B2 fine
-0.3354
Entropy
Exact solution
-0.3356
-0.3358
-0.336
-0.3362
-0.3364
0 1 2 3
CoordinateY
Fig. 1. Isolines of Mach number and distribution of entropy along the left wall for
Ringleb’s flow problem.
It follows from this test case, that for the scheme with a piecewise quadratic
reconstruction the boundary has to be approximated also with higher order
of accuracy, otherwise the numerical results can be worse than the results of
low-cost second order method.
3 Applications in Turbomachinery
The above mentioned numerical method has been applied for the solution
of transonic flows in 2D turbine cascades. The compressible viscous flow is
described by the set of Favre averaged Navier-Stokes equations (RANS) cou-
pled with the TNT k − ω model of turbulence (see [9]). The transonic flow
through a 2D turbine cascade was solved using a hybrid mesh with quadri-
laterals around the profile, in the mixing region behind the outlet edge and
at the outlet boundary. The remaining part of the domain was filled up with
triangles. The total number of elements was 24087 with the y1+ < 1.
Figure 2 shows the isolines of the Mach number the detail of isolines of
entropy near the outlet edge obtained with the help of second and third order
method for the flow characterized by the outlet Mach number M2i = 0.906 and
Reynolds number Re = 848000. The isolines of entropy document clearly the
difference between those two results - the second order scheme gives stationary
solution whereas the wake is unsteady for the third order solution.
3.1 Notes on the Outlet Condition
The proper choice and implementation of the outlet boundary condition plays
an important role in the calculation of transonic flows in cascades. It is obvi-
ous, that uniform outlet pressure gives the non-physical reflection of the shock
waves at the boundary. The implementation of the so-called non-reflecting
boundary conditions can be complicated for the case of unstructured meshes.
Therefore a simpler outlet condition with given average static pressure at the
The WLSQR Scheme 159
Fig. 2. Isolines of Mach number (above) and entropy (below) in 2D turbine cascade,
second and third order solution.
outlet was chosen in this case. Although this boundary condition admits non-
unique solution, it was used successfully in many calculations with structured
meshes (see e.g. [6]). On the other hand, this boundary condition sometimes
fails for the case of unstructured meshes with triangles at the outlet boundary.
See fig. 3 for the non-physical reflection at the outlet. However, the numerical
experiments show that the non-reflective behavior can be recovered by using
one or two layers of quadrilaterals at the outlet.
(a) Triangles at the outlet (b) Quadrilaterals at the outlet
Fig. 3. Non-physical shock reflection and non-reflective behavior at the outlet con-
dition, (M2i = 1.162).
Conclusion
The article describes briefly the weighted least-square reconstruction proce-
dure and extends the previous work by the polynomial approximation of the
boundary. The resulting is successfully applied for the solution of turbulent
transonic flow through turbine cascades.
Problems with the outlet condition with given mean pressure are described
and a temporary solution is proposed.
Acknowledgments. Partial support of the project No. 201/05/0005 of the

Grant Agency of the Czech Republic, Research Plan MSM No. 6840770010 is
acknowledged.
References
1. T. J. Barth and D. C. Jesperson: The design and application of upwind schemes
on unstructured meshes. AIAA Paper 89–0366, AIAA, Jan 1989.
2. M. Feistauer, J. Felcman, and I. Stravskraba: Mathematical and Computational
Methods for Compressible Flow. Numerical Mathematics and Scientific Compu-
tation. Oxford University Press, 2003. ISBN 0-19-850588-4.
3. Jivrı́ Fürst and Karel Kozel. Second and third order weighted ENO scheme
on unstructured meshes. In F. Benkhaldoun and R. Vilsmeier, editors, Finite
Volumes for Complex Applications. Problems and Perspectives. Hermes, July
2002.
4. Jivrı́ Fürst. A finite volume scheme with weighted least square reconstruction.
In S. Raghay F. Benkhaldoun, D. Ouazar, editor, Finite Volumes for Complex
Applications IV, pages 345–354. Hermes Science, 2005. ISBN 1-905209-48-7.
5. G. Giocchia. Exact solutions to transonic and supersonic flows. Technical Report
AR-211, AGARD, 1985.
6. Jan Halama, Tony Arts, and Jaroslav Fovrt. Numerical solution of steady and
unsteady transonic flow in turbine cascades and stages. Computers and Fluids,
33:729–740, 2004.
7. Ami Harten. High resolution schemes for hyperbolic conservation laws. Journal
of Computational Physics, 49:357–393, 1983.
8. Changquink Hu and Chi-Wang Shu. Weighted essentially non-oscillatory
schemes on triangular meshes. Journal of Computational Physics, 150:97–127,
1999.
9. J. C. Kok: Resolving the dependence on free stream values for the k-omega
turbulence model. Technical Report NLR-TP-99295, NLR, 1999.
Building Better (Weighted) ENO Methods
William J. Rider1
Applied Physics Division, Los Alamos National Laboratory, Los Alamos NM 87545
rider@lanl.gov
1 Background on ENO and WENO methods
The development of essentially non-oscillatory (ENO) methods was a natural

extension of the original high-resolution methods to completely self-adaptive
numerical stencils [3]. These methods were developed to overcome the intrinsic
limitations of earlier methods (TVD [2]), which include numerical accuracy
of O h(1+m/N ) where m is the order of the high-order scheme and N is

the order of the norm for measuring error. In particular, this relation implies
first-order accuracy at extrema. With ENO’s basic success came a number
of pathologies that came with the freedom associated with completely adap-
tive stencils. Weighted ENO (WENO) methods were introduced to eliminate
these problems by retaining the basic stencil adaptivity that makes ENO well-
behaved at discontinuities, but providing much better stability, efficiency and
accuracy where the solution is smooth [6, 5]. This is achieved by combining
stencils in a weighted way to achieve better stability along with higher ac-
curacy. The key to the method’s accuracy and self-adaptive nature are the
smoothness detectors that provide the stencil selection. Each of the stencils
that could result from the original ENO stencil selection is used in the final
method through a weighting procedure where the weights are selected to pro-
vide higher accuracy (2m − 1 where m is the order the individual stencils).
The well-developed method used a smoothness detector based on the integral
of the total variation for the stencil polynomial’s derivatives. This method has
achieved great success in a broad spectrum of applications. This method is
extended to high-order methods up to 11th order [1], and I have implemented
a 13th order version. A major problem with the method is the substantial
complexity and cost of the smoothness detectors as the order of the method
increases.

162 William J. Rider
2 The Comparison Principle

In the design of high-resolution methods the general principle can be encap-
sulated through a pithy quote by Scott Adams of the Dilbert comic strip as
his character, Dogbert, Logically all things are created by a combination of
simpler, less capable components. The simpler components are linear schemes
arising from a constant choice of stencil, and the issue is how to combine these
elements to achieve non-oscillatory, but high accuracy results. The basic issue
that hampers the current generation of ENO methods is the hierarchical na-
ture of the stencil selection. All methods are built upon the choice of stencils
order-by-order either directly as in the original ENO methods, or in WENO
methods designed to emulate the ENO procedure.
Here, I introduce a different procedure based on stencil selection only at
high-order and determined through comparison with a TVD method. By com-
parison I mean that the final ENO stencil will be the high-order stencil closest
to the TVD stencil. This results is far simpler, but as we show, no less effective
methods. In fact, the methods may be more effective due to simpler, and a
more straightforward implementation. These methods also allow for a consid-
erable degree of flexibility through the choice of TVD method used for com-
parison. TVD methods are quite well characterized in terms of their numerical
properties [8]. For example, a very dissipative ENO method might utilize a
monotone scheme such as first-order upwind as the comparison scheme. An-
other example of use a compressive TVD stencil such as the superbee limiter
would produce an ENO method with similar character.
The flexibility does not end with the usual family of TVD schemes [8]. One
could introduce another high-order stencil (or optimized stencil) and utilize
this with a TVD limiter as the basis of comparison and produce an ENO
scheme. This high-order stencil could also be chosen from among the high-
order ENO stencil considered as the candidate stencils. Were the desirable
stencil to be TVD itself, the procedure will preferentially chose that stencil.
3 Implementation
To implement these methods one needs a set of functions that will allow us to
choose the stencil closest to the TVD comparison method. These are based on
the standard functions such as the minmod function that returns the argument
with the smallest magnitude unless the arguments differ in sign,
1
minmod (a, b) = (sign (a) + sign (b)) (|a + b| − |a − b|) , (1)
4
which returns a, b or 0. The second function is used with the hierarchical ENO
methods that returns the argument with the minimum magnitude,
1
xminmod (a, b) = (sign (a + b)) (|a + b| − |a − b|) , (2)
2
Building Better (Weighted) ENO Methods 163
which returns either a or b. These two functions can define two functions that
are used to provide a bounding functionality needed for comparison ENO
methods, the median function that returns the argument that is bounded by
the other two,
median (a, b, c) = a + minmod (b − a, c − a) , (3)
which returns a, b or c. If one argument is TVD, it will return the TVD
argument or that closest to it. The other option is to return the argument
that is closest to another
xmedian (a, b, c) = a + xminmod (b − a, c − a) , (4)
in this case the function returns either b or c depending which is closest to a.
These functions can be used to orchestrate a “playoff system” of sorts
where comparison with the TVD stencil is used to determine stencil fitness.
Take the example of a fourth-order stencil with four different alternative fourth
order stencils Sk , and a TVD stencil for comparison, ST V D ,
Sa = median (ST V D , S1 , S2 ) , Sb = median (ST V D , S1 , S2 ) ,
S = median (ST V D , Sa , Sb )
A weighted ENO scheme would proceed in a similar fashion although all
the stencil will contribute to the final one. The smoothness detectors will us
a difference between the TVD and high-order stencil that then will be used
to compute the weights just as the standard WENO methods.
1 w̄k
ISk = 2 ; wk = ,
(Sk − ST V D ) IS 2 +
where w̄k are the weights that would produce 2m − 1 stencil with m being
the order of the candidate stencil Sk . The weights are then renormalized
so their sum is one. Truncation error analysis confirms that this technique
produces methods with the same accuracy as the original WENO methods.
The difference in the efficiency is easy to see through the complexity and
number of operations to implement the smoothness detectors.
is(0) = 22658.D0*f(ic-d)**2 - 208501.D0*f(ic-d)*f(ic-c)
& + 482963.D0*f(ic-c)**2 + 364863.D0*f(ic-d)*f(ic-b)
& - 1704396.D0*f(ic-c)*f(ic-b) + 1521393.D0*f(ic-b)**2
& - 288007.D0*f(ic-d)*f(ic-a)
& + 1358458.D0*f(ic-c)*f(ic-a)
& - 2462076.D0*f(ic-b)*f(ic-a) + 1020563.D0*f(ic-a)**2
& + (86329.D0*f(ic-d) - 411487.D0*f(ic-c)
& + 758823.D0*f(ic-b) - 649501.D0*f(ic-a))*f(ic)
& + 107918.D0*f(ic)**2
This is the ninth-order WENO smoothness detector for a single (the most
upwind) stencil in Fortran [1]. On the other hand, consider the new WENO
method’s smoothness detector for all five stencils.
is(0) = (f5(0) - ftvd)**2

is(1) = (f5(1) - ftvd)**2
is(2) = (f5(2) - ftvd)**2
is(3) = (f5(3) - ftvd)**2
is(4) = (f5(4) - ftvd)**2
The much greater simplicity of the new method is easy to see.
4 Analysis
The analysis of the methods insofar as accuracy is concerned is simple. All
the methods trivially produce the order of accuracy of the stencils supplied.
In particular, the median function will produce the order of accuracy of two
of its arguments irregardless of the nature of the third argument. Thus if two
of the arguments to the median function are of a certain order, the median
function will return a result of at least that order. The xmedian function will
also provide the order of accuracy necessary with ENO schemes. The difficult
and unsolved task is to determine whether the method produces a result that
is T V D + O hm+1 . The original ENO methods did not necessarily provide
proven results of this nature. This has led to the conjecture that ENO schemes
are total variation bounded instead. I believe that the comparison framework
of ENO methods leads to a form that can be more straightforwardly provide
the structure for a subsequent proof. This is because of the strong connection
of the method to a method that is provably TVD. I will focus on the semi-
discrete three-point case for a positive velocity then,
∂uj
= −Cj−1/2 (uj − uj−1 ) , (5)
∂t
where Cj−1/2 ε0 and is usually a nonlinear function for high-order TVD meth-
ods (Cj−1/2 = 1/∆x for first-order upwind) [4]. An example of high-order
TVD method would produce Cj−1/2 = 1/∆x + Qj+1/2 − Qj−1/2 where the
TVD requirement would place restrictions on the form Qj−1/2 might take.
The comparison based ENO schemes would then be built upon these schemes.
It seems plausible that the choice of the stencil that is closest to the TVD
scheme would produce a total variation behavior desired with ENO schemes,
T V D + O hm+1 .

5 Results
Space is limited, but a simple explanatory result can be offered. I will solve
the motion of a complex waveform containing both smooth and discontinuous
data using a simple scalar wave equation [7] with a standard and comparison-
based ENO methods. We will compare both the results and the total variation
behavior using 200 cells.
Building Better (Weighted) ENO Methods 165
1.1
1 ENO
ENO!C
Exact
0
!0.1
!0.995 0 0.995
x/t
Fig. 1. Computation of a waveform for one period using a 6th order standard
ENO methods and the 6th order comparison ENO scheme. The standard ENO
(ENO) produces an error of 1.70 × 10−2 , the comparison ENO (ENO-C) produces
an error of 1.21 × 10−2 .
I can also examine the L1 error in the solution for this case. The error with
the comparison method is smaller than the classic ENO method and its total
variation behavior is better. Furthermore, the total variation behavior indi-
C
8
ENO
ENO!C
TV
7
0 1 2
time
Fig. 2. Computation of a waveform for one period using a 6th order standard
ENO methods and the 6th order comparison ENO scheme. The total variation is
compared. The total variation change is smaller with the ENO-C method, and does
not have the large excursion late in the calculation.
cates that the method is acting just as the classic ENO method. Although we
do not present the weighted version of the ENO method here similar results
are obtained. While the total variation behavior of these weighted schemes is
worse than the classic WENO method, the errors are smaller and the oscilla-
tions produced are less severe.
Conclusion
I have introduced a methodology to develop ENO method that is different than
the previously available ENO techniques. The approach employs a comparison
principle where the high-order stencil; is selected on the basis of its “closeness”
to a TVD (or monotone) stencil. This approach can lead to a large degree of
flexibility in the choice of method properties. The results produced by these
methods are promising and generally less oscillatory and more accurate than
the standard ENO methods.
References
1. D. Balsara and C.-W. Shu. Monotonicity preserving weighted essentially non-
oscillatory schemes with increasingly high order of accuracy. Journal of Compu-
tational Physics, 160(2):405–452, 2000.
2. A. Harten. High resolution schemes for hyperbolic conservation laws. Journal
of Computational Physics, 49:357–393, 1983.
3. A. Harten, B. Engquist, S. Osher, and S. Chakravarthy. Uniformly high or-
der accurate essentially non-oscillatory schemes, III. Journal of Computational
Physics, 71:231–303, 1987.
4. A. Jameson and P. D. Lax. Conditions for the construction of multipoint to-
tal variation diminishing schemes. Applied Numerical Mathematics, 2:335–345.,
1986.
5. G.-S. Jiang and C.-W. Shu. Efficient implementation of weighted ENO schemes.
Journal of Computational Physics, 126:202–228, 1996.
6. X.-D. Liu, S. Osher, and T. Chan. Weighted essentially non-oscillatory schemes.
7. A. Suresh and H. T. Huynh. Accurate monotonicity-preserving schemes with
Runge-Kutta time stepping. Journal of Computational Physics, 136:83–99, 1997.
8. P. K. Sweby. High resolution schemes using flux limiters for hyperbolic conser-
vation laws. SIAM Journal on Numerical Analysis, 21:995–1011, 1984.
Discontinuity Diagnosis Essentially
Non-Oscillatory Schemes
Yun-Feng Liu1 and Jian-Ping Wang2

1
Institute of Mechanics, Chinese Academy of Sciences, Beijing 100080, China
yfliu@pku.org.cn
2
Department of Mechanical and Aerospace Technology, College of Engineering,
Peking University, Beijing 100871, China wangjp@pku.edu.cn
1 Introduction
Essentially non-oscillatory (ENO) schemes were first introduced by Harten,

Engquist, Osher, and Chakravarthy [1] in the form of cell averages. Later, Shu
and Osher [2] [3] developed ENO schemes applying the adaptive stencil idea to
the numerical fluxes and using TVD Runge-Kutta type time discretizations.
Flux-version ENO schemes of Shu and Osher [3] are uniformly high order
accurate right up to the shock wave, simpler to program and very robust to
use. The stencil is designed to adapt in the vicinity of discontinuities to yield
a one-sided interpolation if that becomes necessary. This gives an essentially
non-oscillatory shock transition while maintaining an uniformly high order
accuracy. However, there is also a drawback. ENO schemes have the inherent
mechanism to compare discontinuities to chose the ”smoothest” stencil, but
they lack the mechanism to detect the discontinuities. In other words, when
they approach two discontinuities at both sides, ENO schemes do not know
there are two discontinuities and will choose the weaker discontinuity. This
is the main reason why oscillation sometimes occurs and amplifies for ENO
schemes with accuracy higher than three-order.
In this paper, we introduce a discontinuity diagnosis mechanism into the
flux-version ENO schemes of Shu and Osher [3] and propose the discontinuity
diagnosis essentially non-oscillatory (DDENO) schemes. The essence of this
discontinuity diagnosis mechanism is to find out the discontinuities according
to the information coming form neighboring stencils. DDENO schemes can
detect the discontinuities at both sides automatically and cease to choose the
weaker discontinuity. Thus essentially non-oscillatory property is achieved by
this diagnosis mechanism around discontinuities and higher order accuracy is
obtained by the upstream central schemes at the smooth regions.

168 Yun-Feng Liu and Jian-Ping Wang
2 Derivation of DDENO schemes

In this section, we use the flux-version ENO schemes as the basis to formulate
DDENO schemes. We choose the one-dimensional scalar conservation laws as
an example:
L(u) = ut . (1)
Let us discretize the space into uniform intervals of size ∆x. Let j be an in-
teger, and let xj = j∆x denote cell centers and xj+1/2 denote cell boundaries.
Then we take the conservative schemes
1 ˆ
L(u)j = − (fj+1/2 − fˆj−1/2 ), (2)
∆x
where fˆj+1/2 is a higher-order numerical flux at the xj+1/2 cell boundary.
We can actually assume f (u)x ≥ 0 for all u in the range of our interest. For
a general flux, we can split it into two parts either globally or locally. Here we
will only describe how fˆj+1/2
+
is computed on the basis of DDENO schemes.
For simplicity, we will drop the ”+” sign in the superscript. The formulas for
the negative part of the split flux (with respect to xj+1/2 ) are similar and
will not be shown.
As we well known, the three-order ENO schemes work very well in almost
all physical computations. It chooses one ”smoothest” stencil from three can-
didate stencils and only uses the chosen stencil to approximate the numerical
flux fˆj+1/2 at the cell boundary of xj+1/2 . Thus we use ENO-3 to construct
the DDENO schemes. In the following parts, we use the phrase of ”base sten-
cil” to represent the ”smoothest stencil of ENO-3” for the sake of simplicity.
(1) Compute the divided difference table of f (u) up to rth desirable order.
(2) Construct the base stencil for each node. Choose the smoothest 3-point
stencil for each node as the base stencil by comaparing the divided differences
like ENO-3 schemes. Let us denote the base stencil at node j by S(j):
j j j
S(j) = (kmin , kmin + 1, kmax ), (3)
j j
where kmin and kmax are the left and right nodes of S(j). The base stencil is
chosen for each node, not for each cell boundary.
(3) Start ENO-3 schemes to approximate the numerical flux fˆj+1/2 at
the cell boundary of xj+1/2 . Let us denote the ENO-3 stencil at xj+1/2 by
S(j + 1/2):
S(j + 1/2) = (kmin , kmin + 1, kmax ). (4)
The discontinuity diagnosis mechanism is introduced into ENO-3 schemes
here. Firstly, DDENO schemes analyzes the information coming from the left
neighboring base stencil of node kmin . It compares the present stencil S(j +
1/2) with the base stencil S(kmin ) of node kmin . Secondly, it analyzes the
information coming from the base stencil S(kmax ) of the right adjacent node
kmax .
Discontinuity Diagnosis Essentially Non-Oscillatory Schemes 169
(i) If
S(kmin ) = S(j + 1/2) = (kmin , kmin + 1, kmax ), (5)
it means that there is a discontinuity between the nodes kmin − 1 and kmin .
Otherwise, it means that the left region of the present stencil S(j + 1/2) is
smooth and the present stencil of ENO-3 can extend at least one point to the
left.
(ii) If
S(kmax ) = S(j + 1/2) = (kmin , kmin + 1, kmax ), (6)
it means that there is a discontinuity between the nodes kmax and kmax + 1.
Otherwise, it means that the right region of S(j + 1/2) is smooth and it can
add at least one point to the right.
(iii) If
S(kmin ) = S(kmax ) = S(j + 1/2) = (kmin , kmin + 1, kmax ), (7)
it means that the present stencil S(j + 1/2) is surrounded by two discontinu-
ities at both sides and DDENO ceases to increase points and just uses ENO-3
to approximate the numerical flux at xj+1/2 .
(iv) Otherwise, if
kmin (S(kmin )) < kmin and kmax (S(kmax )) > kmax , (8)
DDENO extend ENO-3 to ENO-5.
(4) Inductively, DDENO repeats step (3) to increase the ENO-3 to higher
order accuracy in smooth regions.
We remark that in the logical operation (iv), we use the logical opera-
tor ”and”. It is a very strong restriction and we do not need to know which
nodes are chosen by S(j + 1/2). If we use the less strict logical operator ”or”,
we should care about the condition of the present stencil. In some severe
positions, the numerical flux is one-sidedly interpolated or extrapolated and
high-order ENO schemes will induce oscillation. Therefore, the logical opera-
tor ”or” should be carefully used. We do not suggest use higher order one-side
extrapolation to approximate the numerical fluxes near discontinuities.
3 Numerical results
We consider here the classical Riemann problems for one-dimensional Euler

system of gas dynamics for a polytropic gas. The time discretization was
performed by third-order TVD Runge-Kutta-type methods developed by Shu
and Osher [2]. All three examples were run with a CFL number of 0.6 and
γ = 1.4. To solve the ordinary differential equation
du
= L(u), (9)
dt
where L(u) is a discretization of the spatial operator, each physical flux

was firstly split into there fluxes by Steger-Warming [5] flux vector splitting
method with three eigenvalues of u − a, u, and u + a, where a is the sound
velocity. Then Steger-Warming fluxes were further performed by the global
Lax-Friedrichs (LF) splitting.
EXAMPLE 1. This is the well known Sod’s problem [6]. The initial data
are (ρL , uL , PL ) = (1, 0, 1), (ρR , uR , PR ) = (0.125, 0, 0.1). We use ENO-3 as
the base stencil. We find that if equation (8) is satisfied, we can use ninth-order
ENO schemes directly and do not need to repeat step (3) and (4). Interpolation
of ninth-order accuracy is not necessary in physical simulations. We just want
to examine the discontinuity diagnosis mechanism. Therefore, the accuracy of
DDENO in this example is either three-order near discontinuities or ninth-
order at smooth regions. The numerical results are presented in Fig. 1. By
comparison, we can see that the shock wave and contact surface of DDENO
are steeper than that of ENO-3. Also notice that the corners of rarefaction
waves are better resolved.
1 1
0.8 0.8
0.6 0.6
0.4 0.4
0.2 0.2
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 1. DDENO-3-9 and ENO-3, Sod’s problem, 100 points, t=0.2 (solid lines are
DDENO-3-9 and dashed lines are ENO-3): (left) pressure; (right) density.
EXAMPLE 2. This is the Riemann problem proposed by Lax [7]. The

initial data are (ρL , uL , PL ) = (0.445, 0.698, 3.528), (ρR , uR , PR ) = (0.5, 0,
0.571). As Sod’s problem, we also use ENO-3 as the base stencil and ENO-9
at the smooth region. The results are shown in Fig. 2. Lax’s problem is a
tough test case for non-characteristic-based schemes of order at least three.
Oscillations can easily appear for such schemes [4]. From Fig. 2 we can see
that the density of DDENO is more steeper than that of eno-3.
EXAMPLE 3. Shu and Osher [3] presented a problem of a moving Mach
3 shock wave interacting with sine waves. This problem is a good model for
the kinds of interactions that occur in simulations of compressible turbulence.
The initial data are specified by
Discontinuity Diagnosis Essentially Non-Oscillatory Schemes 171
1.4
1.2
0.8
0.6
0.4
0 0.2 0.4 0.6 0.8 1
Fig. 2. DDENO-3-9 and ENO-3, Lax’s problem, density, 100 points, t=0.13 (solid
line is DDENO-3-9 and dashed line is ENO-3).
0 2 4 6 8 10
Fig. 3. DDENO-5-9 and ENO-5, Shu’s problem, density, 200 points, t=1.8 (solid
line is DDENO-5-9 and dashed line is ENO-5).

(3.857, 2.629, 10.333), x < 1.0
(ρ, u, P ) = (10)
(1 + ε sin(5x), 0, 1), xε1.0
where ε = 0.2. The problem is run with 200 points in the interval [0, 10]. The
results are plotted in Fig. 3. We mention above that the logical operator ”or”
in logical equation (8) should be used carefully. This problem looks severe
but it is not difficult for ENO schemes to deal with because sine waves are
periodically continuous functions. In this example, we use ENO-5 as the base
stencil because the fluxes are one-side smooth. Then we use logical operator
”and” to extend ENO-5 to ENO-9 at more smooth regions. Thus fifth-order
one-sided interpolation or extrapolation is performed in this example. There-
fore, the accuracy of the results in this example is fifth-order or ninth-order.
From Fig. 3 we can see the expected improvements in resolution with higher
orders.
4 Concluding remarks
In this paper, we introduce a discontinuity diagnosis mechanism into the flux-
version ENO schemes and propose discontinuity diagnosis essentially non-
oscillatory (DDENO) schemes. DDENO schemes use 3-point smoothest ENO
stencil as the base stencil and detect discontinuities by comparing the present
base stencil with neighboring base stencils. This mechanism prevents high-
order ENO schemes from choosing the weaker discontinuities when it encoun-
ters two or more than two discontinuities at both sides. DDENO schemes
are higher order accurate at smooth regions and three-order (the accuracy of
the ENO base stencil) at discontinuities. Numerical experiments demonstrate
that DDENO schemes work well in the examples and achieve higher order
accuracy at smooth regions.
Acknowledgments
The authors gratefully acknowledge Professor Chi-Wang Shu for helpful dis-
cussions and suggestions about flux-version ENO schemes. We also thank Dr.
Hong-Wei Liu at The Hong Kong University of Science and Technology for
helpful discussions.
References
1. Harten, A., Engquist, B., Osher, S., Chakravarthy, S.J.: Uniformly high order
essentially non-oscillatory schemes, III, J.Comput. Phys. 71, 231, (1987)
2. Shu, C.W., Osher, S.: Efficient implementation of essentially non-oscillatory
shock capturing schemes, J.Comput. Phys. 77, 439, (1988)
3. Shu, C.W., Osher, S.: Efficient implementation of essentially non-oscillatory
shock capturing schemes II, J.Comput. Phys. 83, 32, (1989)
4. Jiang, G.S., Shu, C.W.: Efficient implementation of weighted ENO schemes,
J.Comput. Phys. 126, 202, (1996)
5. Steger, J.L., Warming, R.F: Flux vector splitting of the inviscid gasdynamic
equations with application to finite-difference methods, J.Comput. Phys. 40,
263, (1981)
6. Sod, G.A.: A survey of finite difference methods for systems of nonlinear hyper-
bolic conservation laws, J.Comput. Phys. 27, 1, (1978)
7. Lax, P.D.: Weak solutions of nonlinear hyperbolic equations and their numerical
computation, Comm. Pure Appl. Math. 7, 159 (1954)
Third Order Reconstruction on Unstructured
Highly Irregular 3D Meshes
J. Rokicki1 and R. Wieteska1
Institute Of Aeronautics and Applied Mechanics - Warsaw University of

Technology, Nowowiejska 24, 00-665 Warsaw, Poland
[jack,wietek]@meil.pw.edu.pl
The paper discusses finite volume WENO reconstruction applied to simula-

tion of 3D compressible Euler flows on unstructured tetrahedral meshes. The
main point of interest is the case when the mesh becomes highly irregular or
stretched. Special transformation is used in such cases to preserve the third–
order accuracy of reconstruction. This transformation has to be locally applied
to each stencil, on which the reconstruction is sought. This procedure is veri-
fied first on a sequence of 3D meshes in a unit cube and subsequently for the
3D flow in a channel with smooth sinusoidal bump.
1 Introduction
Finite Volume method provides straightforward conservative discretisation for

Euler equations. In its basic form the control volumes coincide with grid cells,
while the unknowns are located at the cell centers (cell-centered approach). In
order to calculate fluxes at the cell walls, the solution is reconstructed within
each cell using information from its immediate neighbourhood.
The order of spatial discretisation depends on two factors: (i) the order
of reconstruction within the cell, (ii) the order of integration formula used at
the cell walls to calculate fluxes.
Higher order reconstruction coupled with WENO (Weighted Essentially
Non Oscillatory method) was already considered in [1],[3],[3], [4], mainly for
very regular meshes.
In the present paper we investigate in more detail the possibility to ob-
tain higher (third) order reconstruction of the solution on irregular/stretched
triangular/tetrahedral meshes.

174 J. Rokicki and R. Wieteska
A
!h
B C
Fig. 1. Typical stencil used in the 3rd order reconstruction procedure.
2 Third–order reconstruction
For third–order reconstruction within the cell Ωh the information is collected
from all neighbouring cells (see Fig. 1 for 2D illustration). The function values
ϕp = ϕ(rp ) are located at cell centres rp (p = 0, . . . , m).
We seek the reconstruction function in the form:
1 T
ϕ (r) = ϕ0 + GT (r − r0 ) + (r − r0 ) H (r − r0 ) (1)
2
where G and H stand for approximated gradient vector and Hessian matrix
respectively (at the point r0 ). To obtain third order accuracy, G and H have
to be estimated with second and first order accuracy, respectively.
To achieve this goal both G and H are expressed as as linear combinations
of the function increments:
m
X m
X
G= Gp wp (ϕp − ϕ0 ) , H= Hp wp (ϕp − ϕ0 ) (2)
p=1 p=1
where vector coefficients Gp and matrix coefficients Hp = HTp are unknown,

while weights wp = krp0 k−1 , wp = krp0 k−2 are chosen such, that Gp and Hp
become ”dimensionless” (rp0 = rp − r0 ).
These unknown coefficients can be found by expanding ϕp − ϕ0 in Taylor
series and by substituting it into (2). The resulting equation system contains
n = N +N ·(N +1)/2 equations and is in general underdetermined (N denotes
the space dimension, n < m). This system can be solved in a least–square sense
minimising the solution norm.
In order to apply WENO procedure, the reconstruction on biased stencils is
also needed, therefore on the same stencil gradient and Hessian are calculated
at cells ”A”, ”B” and ”C” (see Fig. 1).
A typical WENO finite volume scheme can be constructed as follows (for
scalar function ϕ):
- For each cell, gradient and Hessian are calculated using central and 4 (in
3D) biased stencils.
Third Order Reconstruction on Unstructured Highly Irregular 3D Meshes 175
- The resulting reconstruction function is defined by (1) with gradient and

Hessian expressed as a weighted sum of all contributions (thus limiting
oscillations).
This procedure, works well and delivers prescribed order of accuracy only
if the cloud of points is sufficiently regular. This problem will be illustrated
in the next Section.
3 Third–order reconstruction on irregular meshes

We consider now the linear transformation (in 3D):
 
1 00
rη =  0 1 0 · r η < 1. (3)
0 0η
which stretches the mesh (or cloud of points) in one direction. For the purpose
of numerical experiment we take consecutive transformed meshes (for η =
20 , 2−1 , 2−2 , . . . ) and calculate the reconstructed ϕ (r) (1). The approximation
error is obtained by subtracting the exact value of the benchmark function:
ϕ? (r) = sin x2 + 2y 2 + z 2 + xy + 2yz + 4

(4)
A sequence of 3D isotropic unstructured meshes (each characterised by the
average cell size h) is taken to estimate the convergence characteristics of the
scheme.
(a) original method (b) method including transformation
Fig. 2. kLk1 extrapolation error (stretching factor η = 1/2, 1/4, . . . , 1/64).
The effect of stretching is at once visible, when one examines Table 1

presenting the percentage of mesh cells, in which the algorithm described in
Section 2 is unable to solve the corresponding linear systems.
Table 1. Percentage of mesh cells for which gradient and Hessian could not be
calculated using the standard procedure of Section 2
η Failure ratio [%]

1/2 0.0
1/4 0.1
1/8 2.5
1/16 14.7
1/32 51.6
1/64 90.7
1/128 95.6
Figure 2(a) presents additionally the corresponding kLk1 extrapolation

error. It is clear that not only third–order of accuracy is lost but also that the
absolute level of error increases.
The failure of the high–order algorithm introduced in Section 2 can be
attributed to two factors:
• the condition number of the underdetermined linear system grows roughly
as η −3 (the condition number denotes the ratio of the largest and smallest
modulus of all singular values)
• the weights introduced in formulas (2) do not reflect the directional infor-
mation present in the cloud of points (the weights should rather be tensors
than scalars).
Both problems can be alleviated by applying local transformation to the
cloud of points included in the stencil (the use of tensor weights in (2) might
prove difficult).
This transformation is defined, with the help of a matrix:
m m
1
rpc · rTpc ,
P P
M = M [r0 , . . . , rm ] = rpc = rp − rc , rc = m+1 rp (5)
p=0 p=0
which characterises the anisotropy of the cloud. In particular M ≈ I for

sufficiently regular cloud (e.g., when symmetric vertices are located on the
unit sphere).
The matrix M is strictly positive when the cloud has a full space dimension
and in such case new matrix M−1/2 is well defined. Numerically M−1/2 is
easily evaluated by the stable variant of the Newton algorithm [5] (p. 360).
The regularising transformation is defined now by:
r?pc = M−1/2 · rpc (6)
The new cloud becomes fully isotropic since the new matrix:
M? = M [r?0 , . . . , r?m ] = I (7)
Third Order Reconstruction on Unstructured Highly Irregular 3D Meshes 177
On this transformed cloud both the gradient G? and the Hessian H? can
be evaluated using the procedures introduced in Section 2.
The original gradient and Hessian can be recovered now as:
G = M−1/2 · G? H = M−1/2 · H? · M−1/2 (8)

This two–step procedure proved to be successful even for highly stretched
meshes (see Fig. 2(b)).
In this case both gradient and Hessian were correctly evaluated at every
cell within the mesh.
4 Numerical results
Fig. 3. The grid used in numerical test with a local stretch ratio η ≈ 24.
0.62
0.62
0.72 0.55
0.53
0.82
0.50 0.88 0.50
0.
0.45 42
Fig. 4. Mach isolines in the 3D channel. Inlet Mach number 0.5 (dashed lines -
isotropic mesh, solid lines - stretched mesh of Fig 3).
The performance of the full method was investigated for the compressible
flow in 3D channel with sinusoidal bump (see Fig. 3). The mesh consisted of
21915 isotropic tetrahedral cells. To illustrate the robustness of the method
presented in both Sections 2 and 3, the fully subsonic flow (Ma=0.5 at in-
let) was calculated on the original mesh as well as on the mesh with highly
stretched cells (stretch ratio reaching 24).
The comparison of Mach isolines is presented in Fig. 4 - showing almost
perfect coincidence of both simulations. On the other hand the method of
Section 1 alone is unable for the stretched mesh to deliver reasonable accuracy.
5 Conclusions
In the paper the cell-centred Finite-Volume WENO method, was extended to

third-order for unstructured tetrahedral (triangular) meshes.
It was shown that this algorithm is successful for highly distorted (stretched)
meshes only if the additional preprocessing step is implemented (local trans-
formation of the computational stencil).
The performance of the reconstruction was numerically investigated for
a prescribed benchmark function on a sequence of meshes. This proved the
recovery of the third-order accuracy also for highly stretched meshes.
The full algorithm was successfully tested for subsonic flow in the 3D
channel with sinusoidal bump both for regular and stretched meshes.
Further work is necessary to prove the robustness of nonlinear weighting
procedure when applied for transonic flows.
References
1. Carl F., Ollivier-Gooch: Quasi-ENO Schemes for Unstructured Meshes Based on
Unlimited Data-Dependent Least-Squares Reconstruction. Journal of Computa-
tional Physics, 133, 6–17, (1997)
2. Shi J., Hu Ch., Shu Ch-W.: A Technique of Treating Negative Weights in WENO
Schemes. Journal of Computational Physics, 175, 108–127, (2002).
3. Hu Ch., Shu Ch-W.: Weighted Essentially Non-oscillatory Schemes on Triangular
Meshes. Journal of Computational Physics, 150, 108–127, (2002).
4. Henrick A.K., Aslam T.D., Powers J.M.: Mapped weighted essentially non-
oscillatory schemes: Achieving optimal order near critical points. Journal of
Computational Physics, 207, 542–567, (2005).
5. Schatzman M.: Numerical Analysis - a mathematical introduction. Clarendon
Press - Oxford, (2002).
6. Wieteska R., Majewski J., Rokicki J.: Investigation of WENO Schemes For 3D
Unstructured Grids. Computer Assisted Mechanics and Engineering Sciences,
(2006), (in print).
An Intrinsically Multi-Dimensional
Acoustics Convection Upstream Resolution
Algorithm for the Euler Equations
Joe Iannelli1
Centre for Aeronautics, School of Engineering and Mathematical Sciences

The City University of London, Northampton Square, London EC1V 0HB, UK
g.iannelli@city.ac.uk
1 Introduction
This paper documents the performance of an implicit Acoustics - Convection

Galerkin (ACG) algorithm for the Euler Equations. This algorithm induces
a consistent multi-dimensional upwinding in the continuum, directly at the
partial differential system level, before the spatial discretization, within a
characteristics-inspired system [1, 2, 3, 4], which is then discretized via a
Galerkin finite element process.
The procedure generalizes the Streamline Upwind Petrov Galerkin (SUPG)
method [5, 6], for ACG automatically generates Petrov-Galerkin test func-
tions with streamline bias. ACG, however, does not require the SUPG pre-
multiplication of the Euler system by the transpose of the Euler flux Jacobian.
Automatically, ACG incorporates an acoustics-convection stability as well as
shock-capturing expression and induces but minimal artificial diffusion. Pro-
gressing from the results in [1, 2, 3, 4], this paper presents the performance
of the algorithm with GMRES and body adapted grids.
The resulting discrete equations are solved via an implicit Runge-Kutta
algorithm [3] coupled with GMRES [7], preconditioned by GMRES itself. The
ACG-GMRES algorithm has rapidly computed the solution of an inviscid gas
dynamic supersonic flow featuring a type-I interaction of two oblique shocks,
with computational results that reflect the exact solution.
This paper consists of 6 sections. After the introductory remarks in Section
1, Section 2 summarizes the multi-dimensional non-discrete upstream-bias
approximation for the Euler equations. The finite element Galerkin spatial
discretization is delineated in Section 3, and the Runge-Kutta time integration
algorithm is outlined in Section 4. The computational results are presented in
Section 5, with concluding remarks in Section 6.

180 Joe Iannelli
2 Non-Discrete Upstream-Bias Approximation

With implied summation on repeated subscript indices, the Euler equations
are abridged as the non-linear hyperbolic system
∂q ∂fj (q)
+ =0 (1)
∂t ∂xj
For two-dimensional formulations, 1 ≤ j ≤ 2, and with R denoting the real-
number field, the independent variable (x, t), x ≡ (x1 , x2 ), in (1) varies in the
domain D ≡ Ω × [to , tf ], [to , tf ] ⊂ R+ , Ω ⊂ R2 . The dependent variable q
and flux-vector components fj = fjq + fjp are defined as
     


 ρ 




 m j





 0 



 
 
 
 
 

m

 
 
 
 
 


m    j   j 
 1
 
 m1  
 
 pδ1 


 



 ρ 







q p
q≡ , fj (q) ≡ m j
, fj (q) ≡ j
(2)
 m 2
  m 2
  pδ 2

ρ

 
 
 
 
 


 
 
 
 
 

     
m

 
 
 
 
 

   j   

 E  
 (E + p) 
 
 0  

 
  ρ
 
 
 

with straightforward generalization to 3-D flows. In the array q, the variables

ρ, m1 , m2 , E, respectively denote static density, volume-specific linear mo-
mentum components and total energy. The static pressure p and speed of
sound c result from the usual perfect-gas expressions. The Eulerian flow ve-
locity u, with cartesian components uj , 1 ≤ j ≤ 2, is then defined as u ≡ m/ρ,
with Mach number M ≡ kuk/ρ.
The acoustics-convection formulation induces a consistent multi - dimen-
sional upstream-bias directly in the continuum, at the partial-differential equa-
tion level, before the eventual discretization on a prescribed grid, by way of
the characteristics-bias system [1, 2, 3, 4]

∂q ∂ ∂q
− εψt ai +
∂t ∂xi ∂t
∂fjq ∂fjp
" !#
∂fj ∂ N N N
∂q
− εψ c αai aj + α ai aj + ai + ai δ = 0 (3)
∂xj ∂xi ∂xj ∂xj ∂xj
which introduces a streamline upwinding on the time derivative of q and where

the non-negative variables ε, ψt ≤ ψ, ψ, α, αN and δ, each with magnitude of
order not exceeding one, respectively indicate a reference length, upstream-
bias controllers [1], streamwise and crossflow acoustic upstream functions and
pressure-gradient upstream function. The direction cosines ai , aNi , 1 ≤ i ≤ 2,
correspond to the components of unit vectors a, aN respectively in the local
streamline and crossflow directions. This system results from a decomposition
of the hyperbolic flux divergence ∂fj /∂xj into components that physically
Acoustics-Convection Galerkin Multidimensional Upwind Algorithm 181
represent acoustic and convection propagation [3, 4]; in such a decomposition,

fjq and fjp respectively denote the convection and pressure flux components.
When ψt = ψ, this system is Galilean invariant, a property that ensures
that the modeled law of mass conservation, second law of Newtonian me-
chanics and first law of thermodynamics remain independent of the state of
constant motion of the underlying Cartesian reference
frame. In this
system,
∂q ∂fjq ∂fjp N N N ∂q
the expressions cαai aj ∂xj + ai ∂xj + ai δ ∂xj and cα ai aj ∂xj determine
the upstream biases within respectively the streamline and crossflow wave
propagation regions. The upstream-bias functions α = α(M ), δ = δ(M ), and
αN = αN (M ) are then directly and exactly determined in terms of the local
Mach number [1, 2, 3, 4].
3 Finite Element Galerkin Weak Statement

Since the acoustics-convection system (3) is developed independently and be-
fore any discretization, a genuinely multi-dimensional upstream-bias discrete
approximation for the governing equations (1) on arbitrary grids directly re-
sults from a classic centered discretization of this system on the prescribed
grid. To this end a Galerkin finite element method [1, 2, 3, 4] is employed to
discretize this system in space.
The discrete solution q h , subject to prescribed boundary conditions, is thus
h
sought within a finite dimensional subspace H0 (Ω h ) ⊆ H0 (Ω) of dimension
h
N , for all wh ∈ H1 (Ω h ), where superscript “h” signifies spatial discrete ap-
proximation, Ω h denotes the discrete computational domain, and N indicates
the number of computational nodes. The corresponding finite element weak
statement associated with (3) is
( h )
h ∂fjh
Z
h h h h ∂w ∂q
w + ε ψ ai + dΩ +
Ωh ∂xi ∂t ∂xj
 
h ∂q h ph
∂f
Z
∂w h h h j
εh ψ h ch αh ahi ahj + αN aN i aj
N
+ ahi (δ h − 1)  dΩ = 0(4)
Ωh ∂xi ∂xj ∂xj
In this form, this statement is recognized as a Galilean invariant, generalized

SUPG integral statement for the Euler equations [1, 2]. This formulation auto-
matically provides a stability and shock capturing expression, which remains
anisotropic and consistent with characteristic wave propagation.
The discrete solution q h and flux fjh over Ω h at each time t are expressed,
[1, 2, 3, 4], as group linear combinations of time-dependent nodal values and
trial functions. The complete integration with respect to x transforms (4)
into a system of continuum-time ordinary differential equations (ODE) for
determining at each time level t the unknown nodal values q h (x` , t), 1 ≤ ` ≤
N.
182 Joe Iannelli
4 Implicit Runge Kutta Time Integration

The finite element equations (4), along with appropriate boundary equations
and conditions, can be abridged as the non-linear ODE system
dQ(t)
M = F (t, Q(t)) (5)
dt
where the array Q contains the nodal values of q h , M denotes the “mass
matrix” that couples the time derivatives of Q, M dQ(t)dt indicates the corre-
sponding coupling in (4), and F (t, Q(t)) represents the remaining terms in
(4). The numerical time integration of (5) in this study takes place through a
two-stage diagonally implicit Runge-Kutta algorithm (IRK2) [1, 2, 3], which
has been proven absolutely non-linearly energy stable. The eventual non-linear
algebraic equations corresponding to this algorithm are solved by way of New-
ton’s method via matrix-free linear system solutions generated by the GMRES
procedure, [7] as preconditioned by GMRES itself.
The Acoustics-Convection Galerkin method has generated accurate essentially

non - oscillatory results that reflect an available exact solution for a supersonic
gas dynamic flow with two interacting oblique shocks.
Owing to its consistent multi-dimensional upwinding, ACG has managed
to update the upstream directions continuously as the computational solution
evolved, without any filtering or freezing, while maintaining a high rate of
convergence of the residual norm to machine zero. The calculations proceeded
with a prescribed constant maximum Courant number and corresponding ∆t
as
∆t Cmax (∆`)e
Cmax = max{kuk+c, kuk−c, c} ⇒ ∆t = (6)
(∆`)e max{kuk + c, kuk − c, c}
where (∆`)e denotes the minimum element size within the grid. The bench-
mark flow was calculated on an adapted body-fitted grid, consisting of 40
Lagrange bilinear elements in the transverse and longitudinal directions, for
a total of 1600 elements, 1681 nodes and 6724 degrees of freedom.
The clustering of elements towards the wall has generated a grid that
features a large spread of element sizes, which has increased the stiffness of
the system of fully discrete equations. Nevertheless, the implicit implementa-
tion in conjunction with the preconditioned GMRES rapidly determined the
computational solution at each time step.
The benchmark in this paper involves a type-I interaction of two oblique
shocks. The supersonic inlet corresponds to a free-stream Mach number M∞ =
2.40, hence the inlet boundary conditions constrain density ρ, longitudinal and
transversal linear momentum components m1 and m2 and total energy E. The
Acoustics-Convection Galerkin Multidimensional Upwind Algorithm 183
solid wall specification only extends to x = 0.4; beyond this station, the upper
and lower boundaries correspond to outlets, to allow the reflected shocks to
cross the computational boundaries with minimal further reflection. The total
outlet remains supersonic. At the solid upper and lower walls, the inviscid-
flow wall-tangency boundary condition is enforced via surface integrals [1, 3].
The final steady state is computationally achieved with ψ = 0.4 by advancing
the solution in time with a Courant number equal to 100.
The wall-tangency boundary conditions on the upper and lower boundary
walls induce emergence of two separate oblique shocks of different strengths,
shocks that propagate toward each other, interact, and reflect away from each
other, towards the outlet as displayed in Figure 1
M3 = 2.28
M∞ = 2.40 M4,5 = 2.08
M2 = 2.20
Fig. 1. M∞ = 2.40 Shock-on-Shock Interaction, Mach Number Contours
In comparison to the solution in previous articles [2, 3], this solution re-
mains essentially non-oscillatory even at the compression corners, which are
two geometric singularities. This performance is achieved by simply increas-
ing ψ to 2, but only locally over the four elements closest to each singularity.
Overall, the Mach-number distribution presents an essentially non-oscillatory
solution with crisply calculated incident and reflected shocks. In particular,
the algorithm allows the reflected shocks to cross the outflow boundary essen-
tially unperturbed. Significantly, this computational solution mirrors the avail-
able exact solution, with four juxtaposed plateaus connected by four oblique
shocks. The calculated Mach numbers in the plateaus downstream of the four
shocks are M2 = 2.20, M3 = 2.28, M4 ' M5 = 2.08; the shock inclination
angles are θ2 = 28.53o , θ3 = −26.90o , θ4 = 29.52 and θ5 = −30.15, computed
results that coincide with the corresponding exact values.
The ACG-GMRES algorithm has rapidly computed the solution at each
time station and quickly achieved a steady state, with a reduction of the
equation-system residual norm to 1 × 10−14 , hence machine zero, achieved in
about 45 IRK cycles “C” at a constant maximum Courant number of 100.
The recorded essentially monotone decrease of the residual corresponds to an
exponential convergence rate.
184 Joe Iannelli
The Acoustics-Convection Galerkin formulation induces a consistent multi-
dimensional upstream-bias directly in the continuum, at the differential equa-
tion level, before the spatial discretization. The discrete equations are obtained
by way of a classical Galerkin finite element discretization of the formulation
characteristics-bias system.
These time-dependent discrete equations are then integrated in time via
an implicit Runge-Kutta procedure coupled with GMRES, with analytically
determined Jacobians. This procedure remains non-linearly stable and affords
rapid convergence to steady states, with continuously updated upstream di-
rections and maximum Courant number of order 100.
References
1. Iannelli J., Characteristics Finite Element Methods in Computational Fluid Dy-
namics, Springer Verlag, 2006.
2. Iannelli J., “SUPG Formulation of the Acoustics-Convection Upstream Resolu-
tion Algorithm for Compressible Flows”, 44th Aerospace Sciences Meeting and
Exhibit, 2006, Reno, NV; AIAA 2006-1087.
3. Iannelli J., “Finite element and implicit Runge-Kutta implementation of an
acoustics-convection upstream resolution algorithm for the time-dependent two-
dimensional Euler equations”, Int. J. Numer. Meth. Fluids 2005; 49:1261-1286.
4. Iannelli J., “Derivation and characteristics analysis of an acoustics-convection
uptream resolution algorithm for the two-dimensional Euler and Navier-Stokes
equations”, Int. J. Numer. Meth. Fluids 2005; 49:1233-1260.
5. Hughes T.J.R., Recent progress in the development and understanding of SUPG
methods with special reference to the compressible Euler and Navier-Stokes
equations. Int. J. Numer. Meth. Fluids 1987; 7:11-21.
6. Aliabadi S.K., Tezduyar T.E., Parallel Fluid Dynamics Copmutations in
Aerospace Applications. Int. J. Numer. Meth. Engr. 1995; 21:783-805.
7. Saad Y., Schultz M.H., “GMRES: A Generalized Minimal Residual Algorithm
for Solving Nonsymmetric Linear Systems”, SIAM J. Sci. Stat. Comput. 1986;
7(3):856-869.
Multi-dimensional Limiting Process for Two-
and Three-dimensional Flow Physics Analyses
Sung-Hwan Yoon, Chongam Kim∗ and Kyu-Hong Kim
Department of Aerospace Engineering, Seoul National University, Seoul, Korea

151-742 chongam@snu.ac.kr
In this paper, we extend the idea of MLP to three-dimensional space and

present the multi-dimensional limiting process for three-dimensional flow. The
basic idea of the multi-dimensional limiting condition is that the vertex values
interpolated at a grid point should be within the maximum and minimum
cell-average values of neighboring cells for the monotonic distribution. By
applying the MLP (Multi-dimensional Limiting Process), we can obtain non-
oscillatory shock profile even in multiple dimension, which results in very good
convergence characteristic. Furthermore, while maintaining the oscillation-free
attribute, we can achieve an enhancement of solution accuracy by combining
the MLP with high-order polynomial interpolation.
1 Introduction
Since the late 1970s, numerous ways to control oscillations have been studied
and several limiting concepts have been proposed. Most representatives would
be TVB, ENO/WENO and TVD[1]. However, most oscillation-free schemes
have been based on the mathematical analysis of one-dimensional convection
equation. Although they may work very well in many cases, it is often insuf-
ficient or almost impossible to control oscillations near shock discontinuity in
multiple dimension. For that reason, it is obvious that more practical oscilla-
tion control method for multi-dimensional applications is needed.
Kim and Kim[2] extended the one-dimensional monotonic condition to
two-dimensional flow and presented the two-dimensional limiting condition
successfully. With the limiting condition, a multi-dimensional limiting process
(MLP) is proposed which gives more accurate results for the two-dimensional
Euler and Navier-Stokes equations. It is this approach which prompts the
work of the present research. As a continuous study on MLP, we extend the
idea of MLP to three-dimensional space and also consider three-dimensional
MLP approach with high-order polynomial interpolation.

186 Sung-Hwan Yoon, Chongam Kim and Kyu-Hong Kim
2 Multi-dimensional Limiting Process (MLP)

2.1 Multi-dimensional Limiting Condition
The present research essentially extends the MUSCL approach[3] with TVD
limiter. One-dimensional limiting condition using TVD constraint can be writ-
ten as follows.[1]
0 ≤ φ(r) ≤ min(2r, 2). (1)
The extension of Eq.(1) in a dimensional splitting manner may be in-
sufficient to prevent oscillations in multi-dimensional flow. The dimensional
splitting extension does not possess any information on property distribution
at cell vertex points, which would be essential when property gradient is not
aligned with grid lines. Thus, as an extended condition including the missing
information, Kim and Kim[2] proposed multi-dimensional limiting condition
for two-dimensional flow.
In order to realize the limiting condition in three-dimensional situation,
the values at vertex points are required to satisfy the following condition.
Φ̄min max
p,q,r ≤ Φi+p/2,j+q/2,k+r/2 ≤ Φ̄p,q,r , (2)
where Φi+p/2,j+q/2,k+r/2 is a vertex point value and Φ̄min max
p,q,r and Φ̄p,q,r are the
minimum and maximum cell-averaged values among neighboring candidates,
respectively. To derive the multi-dimensional limiting function from Eq.(2),
we need to express the vertex point values in terms of variations at the cell-
interfaces.
After that, we determine each variation to satisfy the multi-dimensional
limiting condition using the limiting coefficient α. The coefficient α possesses
the information of multi-dimensionally distributed physical property. Thus,
with the help of the coefficient α, we can formulate the multi-dimensional
limiting function. Here, we will give only the final formulation of MLP for
three-dimensional flow and refer to [4] for most details of derivation procedure.
2.2 The Formulation of MLP for Three-dimensional Flow
With the multi-dimensional limiting function, a new family of limiting process

in multi-dimensional flow can be developed. For three-dimensional flow,
1 ξ
ΦL
i+1/2,j,k = Φi,j,k + max(0, min(αL , αL rL,i,j,k , βL ))∆Φi−1/2,j,k , (3)
2
1 ξ
ΦR
i+1/2,j,k = Φi,j,k − max(0, min(αR , αR rR,i+1,j,k , βR ))∆Φi+3/2,j,k , (4)
2
where α is the multi-dimensional restriction coefficient which determines the
baseline region of MLP and β is the coefficient evaluated by various kinds of
interpolations. Since the evaluations of interpolated values are independent of
a numerical flux, MLP can be combined with any numerical flux. Values of
MLP for Two- and Three-dimensional Flow Physics Analyses 187
αL,R and βL,R in Eq.(3) and Eq.(4) can be summarized as follows. Along the
ξ-direction, if ∆Φpξ ≥ 0,
 
ξ max
2max 1, r L,i,j,k Φp,q,r − Φi,j,k 
αL = g   , (5)

∆Φqη ∆Φrζ

1 + ∆Φp + ∆Φp ∆Φi+1/2,j,k
ξ ξ i,j,k
 
ξ min
2max 1, 1/r R,i+1,j,k Φp,q,r − Φi+1,j,k 
αR = g   , (6)

∆Φqη ∆Φrζ

1 + ∆Φp + ∆Φp ∆Φi+3/2,j,k
ξ ξ i+1,j,k
ξ ∆Φ
i+1/2,j,k ξ ∆Φ
i+1/2,j,k
where rL,i,j,k = ∆Φi−1/2,j,k , rR,i+1,j,k = ∆Φi+3/2,j,k , g(x) = max(1, min(2, x)).
Along the η- and ζ-direction, the left and right values at the cell-interfaces
can be formulated in the same way.
Combining Eq.(5) and Eq.(6) with β in the form of third order and fifth
order polynomial, we finally obtain three-dimensional versions of MLP3 and
MLP5, respectively. For more detailed explanation, see Ref.[2].
MLP with using 3rd order polynomial (MLP3):

ξ ξ
1 + 2rL,i,j,k 1 + 2rR,i+1,j,k
βL = , βR = . (7)
3 3
MLP with using 5th order polynomial (MLP5):
ξ ξ ξ ξ
−2/rL,i−1,j,k + 11 + 24rL,i,j,k − 3rL,i,j,k rL,i+1,j,k
βL = ,
30
ξ ξ ξ ξ
−2/rR,i+2,j,k + 11 + 24rR,i+1,j,k − 3rR,i+1,j,k rR,i,j,k
βR = . (8)
30
In addition, we can consider various high-order interpolation methods to con-
struct coefficient β which may affect solution accuracy greatly. One challenge-
able candidate is WENO scheme, namely MLP combined with WENO scheme.
Since an oscillation control mechanism is already included in WENO itself,
direct combination of MLP with WENO is more robust than MLP only ap-
proach. Solution accuracy of straightforward combination, however, is notice-
ably decreased by excessive restrictions to control oscillation. Nevertheless,
it is expected that, if we adjust the way of combination, MLP can prevent
some defects of WENO scheme in multi-dimensional flow with little loss of
accuracy.
3 Numerical Results
3.1 Three-dimensional Supersonic Corner Flow
Firstly, we consider a box-type inlet supersonic corner flow corresponding to

irregular reflection in order to examine the capturing capability of complex
shock structure with TVD MUSCL limiters and MLP. Since this test is focused
on the comparison of monotonic shock capturing, viscous fluxes are excluded.
The freestream Mach number is 3.0 and 6.0 and the deflection angle of each
corner is 10, 15, 20 deg., respectively.
Figure 1 shows the comparison of density contours on lower wall and error
history, which corresponds to Mach number 3.0 and inclination angle of 10 de-
gree. The conventional van leer limiter shows slight oscillations, but MLP-van
leer limiter prevents any oscillation. Although difference between the solu-
tions in contour seems to be tiny, that makes considerable difference in error
history. As shown in error history, the van leer limiter case is never converged
due to slight oscillation. On the other hand, MLP-van leer limiter and MLP5
case show smooth convergences.
Fig. 1. Comparison of density contours on lower wall and error history
Fig. 2. Numerical schlieren : M=6.0 and θw = 10 (left), M=6.0 and θw = 20 (right)

MLP for Two- and Three-dimensional Flow Physics Analyses 189
We also verify the complex shock structure capturing ability of MLP5 with
various Mach number and inclination angle of wedge.
Figure 2 is the numerical schlieren image using MLP5. The formulation
of the triple-point and accurate prediction of reflected shock wave, slip line
and spiral curl are important from the standpoint of numerical analysis. The
numerical results indicate that MLP is one of the sound numerical method
for complex shock structure capturing in multiple dimension.
3.2 Three-dimensional Viscous Shock Tube Problem
The two-dimensional viscous shock tube problem was initially proposed by

Daru and Tenaud[5] and has been successfully used for the validation of high-
resolution numerical methods. We extend it to three-dimensional problem
with the same condition. The Reynolds number is 200 with constant viscosity
and the initial state is given as follows.
(ρ, u, v, w, p)L = (120, 0, 0, 0, 120/γ) and (ρ, u, v, w, p)R = (1.2, 0, 0, 0, 1.2/γ)
Viscous fluxes are calculated by 4th order spatial accuracy, and 3rd order TVD
Runge-Kutta method is used for time integration.
At symmetric boundary, the profile of flow field is same as two-dimensional
case in which sensitive separated vortex is developed. On the other hand, near
the corner of the walls, there are spiral curls which can not be observed in
two-dimensional problem, but in three-dimensional supersonic corner flow.
(see figure 3)
Fig. 3. Comparison of density contours at yz-plane (x=0.8725)
Moreover, the curls interact with a pillar-vortex near the corner of the
walls, which leads to much more complicated flow structure. In figure 4, we
can see the interaction of curls and pillar-vortex tip, and the profile is very
sensitive to numerical scheme. The result of M-AUSMPW+ with MLP5 is
clearer than that of AUSMPW+ with van leer limiter, which means that we
can achieve accuracy enhancement by using M-AUSMPW+ with MLP5.
Fig. 4. Comparison of density contours at xz-plane (y=0.08)
The multi-dimensional limiting process (MLP) for three-dimensional flow is

developed by combining the multi-dimensional limiting function with high-
order polynomial interpolation. This method turns out to have several de-
sirable characteristics such as non-oscillatory behavior across a discontinuity
even in multi-dimensional flow, robust convergence. In addition, high-order
interpolation can be easily incorporated.
The most distinguishable property of MLP is the achievement of oscillation-
free profiles in two- and three-dimensional flow and, as a result, exhibits a good
convergence characteristic. Through several test cases, it is verified that MLP
can control numerical oscillations in multiple dimension very effectively. At
the same time, MLP provides substantial accuracy improvement, compared
with conventional TVD MUSCL approach.
As an on-going work, the combination of MLP-WENO scheme is being
developed. It is expected that improved results can be obtained by using the
own smoothness indicator of WENO for combination of the two schemes.
References
1. P. K. Sweby :High Resolution Schemes Using Flux Limiters for Hyperbolic Con-
servation Laws. SIAM J. on Num. Analysis, Vol. 21(5), 995–1011 (1984)
2. K. H. Kim and C. Kim : Accurate, Efficient and Monotonic Numerical Methods
for Multi-dimensional Compressible Flows, Part II Multi-dimensional Limiting
Process. J. of Comput. Physics, Vol. 208(2), 570–615 (2005)
3. B. van Leer : Toward the ultimate conservative difference scheme. J. of Comput.
Physics, Vol. 32, 101–136 (1979)
4. S. Yoon, C. Kim, and K. Kim : Multi-Dimensional Limiting Process for Three
Dimensional Compressible Flows. 17th AIAA Computational Fluid Dynamics
Conference, Toronto, Ontario, June 6-9 (2005)
5. V. Daru, C. Tenaud : Evaluation of TVD high resolution schemes for unsteady
viscous shocked flows. Computers and Fluids, 30 89–113 (2001)
A Multidimensional Kinetic Upwind Method
for Euler Equations
Keshav S Malagi1 , P S Kulkarni1 , and S M Deshpande2

1
CFD Centre & JATP, Dept. of Aerospace Engg. Indian Institute of Science,
Bangalore, India ksmalagi@gmail.com, psk@aero.iisc.ernet.in
2
Engg. Mechanics Unit, Jawaharlal Nehru Centre for Advanced Scientific
Research, Bangalore, India smd@jncasr.ac.in
The study of directional derivative lead to the development of a rotationally

invariant kinetic upwind method (KUMARI)3 which avoids dimension by di-
mension splitting. The method is upwind and rotationally invariant and hence
truly multidimensional or multidirectional upwind scheme. The extension of
KUMARI to second order is as well presented.
1 Symmetry of differential operator and its discrete

approximation
Deshpande [1] has studied the question of symmetry of differential operators

and their discrete approximations. In kinetic schemes [4] one often deals with
differential operator
∂f ∂f
D (f ) = v1 + v2 (1)
∂x ∂y
which arises in Boltzmann equation. It has been discussed in [1, 3] that D(f )
is invariant under 2D rotation group O(2) and that though the discrete differ-
ential operator DD(f ) obtained using least squares approximation for deriva-
tives is invariant under O(2) the upwind approximation using least squares is
not. However DD(f ) obtained using least squares upwind approximation is
invariant under a subgroup E of O (2)
E = {R(0), R(π/2), R(π), R(3π/2)} (2)
Thus it was concluded that upwinding along coordinate directions based on
stencil division is symmetry breaking! [1, 3]. Hence the central question for
construction of high fidelity numerical scheme [3] for multidimensional Euler
equations is: how do we construct an upwind approximation to D(f ) defined
3
Kinetic Upwind Method Avec Rotational Invariance, Avec in french means with

192 Keshav S Malagi, P S Kulkarni, and S M Deshpande
by eqn.(1) which is also rotationally invariant i.e., no symmetry breaking is

allowed. The new algorithm called KUMARI has been constructed precisely
[2], to satisfy rotational invariance as well as upwinding. The KUMARI makes
use of the fact that directional derivative of the flux vector is infact a Fourier
series with three terms [3]. If Q = GXı̂ + GY ̂ is the flux vector in 2-D and `ˆ
is the unit vector making an angle θ with x-axis and s is the distance along `ˆ
as shown in Fig.(1.a) then the directional derivative of Q along θ is given by
[2],
∂ ˆ = A0 + A1 cos θ + B1 sin θ + A2 cos 2θ + B2 sin 2θ
(Q `) (3)
∂s
with co-efficients as defined below,

1 ∂ ∂
A0 = (GX) + (GY ) , A1 = B1 = 0,
2 ∂x ∂y

1 ∂ ∂ 1 ∂ ∂
B2 = (GX) + (GY ) , A2 = (GX) − (GY )
2 ∂y ∂x 2 ∂x ∂y
From the above equation it is clear that the term A0 contains the divergence
Y’ Y !
Y
l^ C1
X’
^
l
+ s
Pi s C1 P0 ! X
P0 ! X
Pi
(a) A cloud of points (b) stencil division along

around point P0 rotated coordinate frame
Fig. 1.
of flux vector Q. Thus the Fourier series expansion of directional derivative

contains the information of divergence of flux vector Q. Hence if we can find a
consistent and rotationally invariant approximation to A0 , then we can devise
a rotationally invariant scheme for Euler equations. This is very central to the
development of new algorithm KUMARI. For simplicity of explanation let us
introduce the following notations,
∂
E = Q `ˆ = special flux and ζ(θ) = (E) = directional derivative (4)
∂s
From the above discussion it is clear that we can represent directional deriva-
tive by Fourier series. Let us approximate ζ(θ) by Fourier series (say of three
terms),
A Multidimensional Kinetic Upwind Method for Euler Equations 193
ζ(θ) = A0 + A1 cos θ + B1 sin θ (5)

Suppose we have m values of ζ(θj ) for (j = 1, ..., m), where m is the num-
ber of directions which can be arbitrary. Obviously if m > 3, the problem
becomes overdetermined, thus suggesting the use of least squares. The least
squares principle is then used to estimate the co-efficients of the Fourier series
so approximated by fitting the approximation to the data (in this case the
directional derivatives, ζ(θ1 ), ζ(θ2 ), . . . ζ(θm )). The method yields,
 P P    P 
Pm P cos2θj P sin θj A0 P ζ(θj )
 cos θj
P cos θj cos θj sin θj  A1  =  P ζ(θj ) cos θj  (6)
sin2 θj
P P
sin θj sin θj cos θj B1 ζ(θj ) sin θj
We can easily find A0 from the above set of equations. Further using the
simple Euler time stepping yields the update formula,
U n+1 = U n − 2∆tA0 (7)
2 Special Flux Vector Splitting (SFVS) and evaluation

of directional derivative
In the previous section we had assumed that ζ(θ) is given. Note that ζ(θ)
ˆ of component of flux vector Q along `.
is a derivative along any direction `, ˆ
Since the Euler equations are hyperbolic in nature we need to evaluate ζ(θ) in
an upwind way along any direction θ. This has been achieved by developing
special flux vector splitting(SFVS) [3, 2] method which splits the special flux
E into E + and E − and using the least squares princilple to evaluate derivative.
Note that the special flux E,
E = Q `ˆ = GX cos θ + GY sin θ (8)
contains the flux components GX and GY which are along x and y directions
while we need to split the flux along `.ˆ Hence the name special flux vector
splitting (SFVS) is used. The [3] can be referred for the detailed derivation
and the special split flux expressions. Using SFVS the directional derivative
becomes,
∂E ∂E + ∂E −
ζ(θ) = = + (9)
∂s ∂s ∂s
To compute the derivatives of split fluxes in the above equation we can again
use the least squares approximation of derivatives, as in standard LSKUM,
along with special split fluxes. The upwinding is achieved by stencil division,
as in LSKUM, along each direction (Fig.(1.b)). The required least squares
formulae for spatial derivatives in eqn.(9) are given in [3]. It is thus clear
that the split derivatives in the expression for the directional derivative, ζ(θj )
can be determined in an upwind way by using the above formula along each
direction θj , for j = 1, . . . . . . , m, m being number of directions which can be
+ −
arbitrary. It is important to note that ζ(θj ) = ∂E ∂E
∂s + ∂s is determined for
each `ˆ or equivalently each θj . Full stencil into two halves is required but the
substencils keep changing. The union of two substencils is always the same
full stencil. It is interesting to observe that we are using least squares twice.
First it is used in generating the data and then in fitting Fourier series to the
data so generated.
3 q-KUMARI
Even though KUMARI gives good accuracy, it is only first order accurate.
It is highly desirable to get more accurate results by reducing the dissipa-
tion further by making it higher order accurate. It can be made higher order
accurate by evaluating ζ(θ) to a higher order which will result in less dissipa-
tion and more accurate results. This is achieved by using defect correction [1]
along with q-variables [7] (also called as entropy variables). The higher order
accurate formula for ζ(θ) is given by,
P P P P 
2
 ∆Ẽi+ ∆x0i ∆y 0 i − ∆Ẽi+ ∆yi0 ∆x0 i ∆y 0 i 
ζ (2) (θj ) = 2
∆x0 2i ∆y 0 2i − ( ∆x0 i ∆y 0 i )
P P P
 
Pi C1+
P P P P 
2
 ∆Ẽi− ∆x0i ∆y 0 i − ∆Ẽi− ∆yi0 ∆x0 i ∆y 0 i 
+ 2
∆x0 2i ∆y 0 2i − ( ∆x0 i ∆y 0 i )
P P P
 
Pi C1−
where, ∆Ẽi± (q) = E ± (q̃i ) − E ± (q̃0 ). The q̃ are given by,

∆xi ∆yi ∆xi ∆yi
q˜i = qi − qxi − qyi and q˜0 = q0 − qx0 − qy0 (10)
2 2 2 2
where q derivatives are evaluated using least squares with full stencil. The
method to evaluate A0 essentially remains the same.
4 Results and Discussion
The method has been successfully tested on a standard shock reflection prob-
lem with M = 2.9, θ = 110 over a domain with [0 ≤ x ≤ 3] and [0 ≤ y ≤ 1].
We can see from the plot of pressure Fig.(2.a) at y=0.35 along x-axis that the
shock is captured crisply for KUMARI method as compared to LSKUM. The
better residue fall Fig.(2.b) is obtained for KUMARI method, signifying bet-
ter convergence. Studies on sensitivity of the method to varying aspect ratio
AR, have been carried out and is found to be less sensitive and are reported
in [2, 5]. The q-KUMARI has been applied to standard test cases of flow
past NACA 0012 airfoil for supersonic, subsonic and transonic flow regimes.
A Multidimensional Kinetic Upwind Method for Euler Equations 195
In supersonic test case an uniform supersonic flow at M∞ =1.2 at an angle

of attack α = 00 is considered. The pressure contours Fig.(2.c) shows that
the shocks have been captured correctly. A subcritical flow of M∞ =0.63 and
angle of attack α = 20 is considered. The pressure contours have been given
in Fig.(2.d). The transonic test case with M∞ =0.85 and α = 10 is considered
as a critical test case. The pressure contours Fig.(2.e) show that the shocks on
upper and lower surfaces are captured correctly. Further discussion and com-
parison of the overall co-efficients with standard results for airfoil test cases
have been given in [3]. The 2D Riemann problem [6] has highly discontinuous
unsteady flow and is suggested for testing the truly multidimensional schemes.
In the present case we consider the third test case of Lax [6] which has Kelvin-
Helmholtz instability. The density isoplots have been shown in Fig.(2.f). We
can see that all the features have been captured very well. The 2D blast wave
test case is an unsteady test case used for simulating explosion. A cloud of
points is obtained using a cartesian grid (201×201) to demonstrate the ability
of the scheme to capture circular symmetry well even on a cartesian grid. The
pressure contours Fig.(2.g) & Fig.(2.h) have been given for both q-KUMARI
and LSKUM respectively. We can notice more structures in case of q-LSKUM
along the diagonal or along the axes. It is interesting to note that the circular
symmetry is captured well with both the methods.
References
1. S M Deshpande: ”Meshless method, accuracy, symmetry breaking, upwinding
and LSKUM”, Fluid mechanics report Report No 2003 FM 1, Department of
Aerospce engineering, Indian Institute of Science, Bangalore.
2. Keshav S Malagi, P S. Kulkarni, S M Deshpande: ”Rotationally invariant kinetic
upwind method(KUMARI)”, Fluid mechanics report Report No 2004 FM 23,
Department of Aerospce engineering, Indian Institute of Science, Bangalore.
3. Keshav Shrinivas Malagi: ”Rotationally invariant kinetic upwind
method(KUMARI)”, MSc thesis submitted, Department of Aerospce en-
gineering, Indian Institute of Science, Bangalore.
4. Mandal J C and S M. Deshpande: ”Kinetic Flux Vector Splitting for Euler
Equations”, Computers and Fluids, Vol 23. No. 2, pp.447-478, 1994.
5. Keshav S. Malagi, P S. Kulkarni, S M. Deshpande: ”Rotationally invariant ki-
netic upwind method (KUMARI)”, Proceedings of ”8th Annual AeSI CFD Sym-
posium”, Aug-2005, NAL, Bangalore India.
6. P. D. Lax and X D. Liu: ”Solution of two dimensional Riemann problem of
gas Dynamics by positive schemes”, SIAM J Sci. Comput. Vol. 19, No.2, pp.
319-340,1998.
7. S M. Deshpande, K Anandhanarayanan, C. Praveen and V. Ramesh: ”Theory
and applications of 3-D LSKUM based on entropy variables”, Int. J for Num.
Methods in Fluids: vol40, Nos 1-2, pp.47-62, 2002.
0
4
KUMARI
3.5
KUMARI LSKUM
LSKUM -5
3
RESIDUE
pressure
EXACT SOLUTION
2.5
-10
2
1.5
-15
1
0 1 2 3 0 5000 10000 15000
X N
(a)comparison of pressure at y = 0.35 (b)convergence history for

for shockreflection test case shockreflection test case
(c)Pressure contours for supersonic (d)Pressure contours for subsonic

flow past airfoil using q-KUMARI flow past airfoil using q-KUMARI
1
0.8
0.6
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
(e)Pressure contours for transonic (f)Density contours for 2D Riemann

flow past airfoil using q-KUMARI test case using q-KUMARI
(g)Pressure contours for blast wave (h)Pressure contours for blast wave
test case with q-KUMARI test case with q-LSKUM
Fig. 2.
High Resolution Quantum Kinetic Beam
Schemes and Its Applications to Ideal
Quantum Gas Dynamical Flows
Y. H. Shi1 and J. Y. Yang2
Institute of Applied Mechanics, National Taiwan University, 106 Taipei, TAIWAN

f92543072@ntu.edu.tw
Extensive theoretical and computational methods have been developed to

solve the Euler equations of classical gas dynamics. Among the various meth-
ods such as the exact and approximate Riemann solvers [5, 12], most are
constructed based on the hyperbolic conservation laws of the Euler equations.
Some kinetic methods have been devised in the past to solve the Euler equa-
tions of classical gas dynamics [8, 10, 2, 9, 14, 1, 7]. The beam scheme is
presented by [11] for solving the equilibrium limit of the classical Boltzmann
equation. The derivation was based on the local thermodynamic equilibrium
Maxwell-Boltzmann distribution. This scheme has been extended and gen-
eralized to ideal quantum gas dynamics by Yang & Shi [15] based on the
equilibrium limit solution of the quantum Boltzmann equation [6] namely,
the Bose-Einstein and Fermi-Dirac distributions. This work is an extended
work based on Yang & Shi [15]. The kinetic beam scheme is generalize to gen-
eral coordinate system and various high-order methods such as TVD, ENO
and WENO were implemented into the basic first order beam scheme.
1 Quantum Boltzmann Equation
Following [6, 13], we consider the quantum Boltzmann equation

∂ p δf
( + · ∇x − ∇U (x, t) · ∇p )f (p, x, t) = ( )coll. (1)
∂t m δt
The conservation laws are obtained by multiplying Eq. 1 by 1, p, or p2 /2m,
and then integrating the resulting equations over all p. The integrals of the
collision terms in all three cases vanish automatically and we have the dif-
ferential conservation laws for the conserved macroscopic quantities, i.e., the
particle number density n(x, t), the momentum density, J = mj, and the
energy density, (x, t) as follows:

198 Y. H. Shi and J. Y. Yang
∂n(x, t)
+ ∇x · j(x, t) = 0 (2)
∂t Z
∂mj(x, t) dp p
+ ∇x · p f (p, x, t) = 0 (3)
∂t h3 m
dp p p2
Z
∂(x, t)
+ ∇x · f (p, x, t) = 0. (4)
∂t h3 m 2m
Here the number density, number density flux, and energy density are given,
respectively, by
dp p2
Z Z Z
dp dp p
n(x, t) = f, j(x, t) = f, (x, t) = f. (5)
h3 h3 m h3 2m
One can obtain these macroscopic quantities in closed form in terms of the
Bose or Fermi functions in the local equilibrium limit. The effect of the ex-
ternally applied field U (x, t) is not considered. We first consider the following
local equilibrium quantum distribution in two space dimensions,
2
(0) +(py −muy )2 ]/2mkB T (x,y,t)
f2 (px , py , x, y, t) = [z −1 e[(px −mux ) + θ]−1 (6)
where θ = +1 denotes the Fermi-Dirac statistics and θ = −1 the Bose-
Einstein statistics, z(x, y, t) = eµ(x,y,t)/kB T (x,y,t) is the fugacity, ux (x, y, t),
and uy (x, y, t) are the mean velocity components. Then the number density
n(x, y, t) is given by
Z ∞
dpx dpy (0) h1 (z)
n(x, y, t) = 2
f2 (px , py , x, y, t) = (7)
−∞ h λ2
the momentum j(x, y, t),
Z ∞
dpx dpy px (0)
jx (x, y, t) = f (px , py , x, y, t) = n(x, y, t)ux (x, y, t) (8)
−∞ h2 m 2
Z ∞
dpx dpy py (0)
jy (x, y, t) = f (px , py , x, y, t) = n(x, y, t)uy (x, y, t) (9)
−∞ h2 m 2
and the energy density (x, y, t),
Z ∞
dpx dpy (p2x + p2y ) (0) h2 (z) 1
(x, y, t) = 2
f2 = + n(u2x + u2y ) (10)
−∞ h 2m βλ 2
q
βh2
where λ = 2πm is the thermal wavelength and β = 1/kB T (x, y, t).
In the above, the function hν (z) is equal to gν (z) which denotes the Bose
function of order ν when we consider the boson gas. The Bose function is
defined by
Z ∞ ∞
1 xν−1 X zl
gν (z) ≡ dx −1 x = (11)
Γ (ν) 0 z e −1 lν
l=0
And in the case of Fermi gas,hν (z) is equal to fν (z),the Fermi function, defined
by
High Resolution Quantum Kinetic Beam Schemes 199
∞ ∞
xν−1 X (−z)l
Z
1
fν (z) ≡ dx = − ν . (12)
Γ (ν) 0 z −1 ex + 1 l
l=0
In Bose-Einstein gas,the fugacity could not greater than one or less than zero,
but in Fermi-Dirac gas we do not have such limits.
2 Quantum Kinetic Beam Scheme
The main idea of the beam scheme is to approximate the local equilibrium dis-
tribution function into several delta functions. The distribution in two spatial
(0)
dimensions, f2 , is approximate by
f2 (p, x, t) ∼
(0)
= qi,j (px , py )
= ai,j δ(px − px0 , py − py0 )
−
+ bi,j δ(px − p+ x0 , py − py0 ) + bi,j δ(px − px0 , py − py0 )
−
+ ci,j δ(px − px0 , py − p+ y0 ) + ci,j δ(px − px0 , py − py0 ) (13)
where p±
x0 = px0 ± ∆px and p±
= py0 ± ∆py . The unknown parameters
y0
a, b, c, px0 , py0 , ∆px , and ∆py in the beam approximation can be found in
[15,16] and they are given by
2πm 4 g22 (z) 1 2πm g22 (z)
a=( )[g1 (z) − ], b = c = ( )
β 3 g3 (z) 3 β g3 (z)
s
3m g3 (z)
∆px = ∆py = px = mux , pyo = muy
2β g2 (z) o
The conservative quantities carried by each beam in cell (i, j) are Qσ,i,j =
(Rσ , Mσ , Nσ , Eσ )i,j , with
Z
dpx dpy
Rσ,i,j = cσ,i,j δ(px − p̄x,σ , py − p̄y,σ ) (14)
h2
Z
dpx dpy px
Mσ,i,j = cσ,i,j δ(px − p̄x,σ , py − p0y ) (15)
h2 m
Z
dpx dpy py
Nσ,i,j = 2
cσ,i,j δ(px − p0x , py − p̄y,σ ) (16)
h m
p2x + p2y
Z
dpx dpy
Eσ,i = cσ,i,j δ(px − p̄x,σ , py − p̄y,σ ) (17)
h2 2m
where p̄x,σ = p0x , for σ = 1, 4, 5, p̄x,2 = p0x −∆px , and p̄x,3 = p0x +∆px ; p̄y,σ =
p0y , for σ = 1, 2, 3, p̄y,4 = p0y − ∆py , and p̄y,5 = p0y + ∆py ; and cσ,i,j = ai,j ,
for σ = 1, and cσ,i,j = bi,j , if σ = 2, 3, 4, 5. The overall conservative quantities
P5
of gases in cell (i, j) are the summation of every beam ie. Qi,j = σ=1 Qσ,i,j .
The final 2D governing equations in generalized coordinates (ξ, η) based on
the beam splitting method can be expressed as:
∂t Q̂σ + ∂ξ (F̂σ+ + F̂σ− ) + ∂η (Ĝ+ −

σ + Ĝσ ) = 0, (18)
For explicit methods in multiple space dimensions, the Strang-type dimen-
sional splitting was employed and the integrating scheme can be expressed in
terms of operators as
Qn+2 n
σ,i,j = Lξ (∆t)Lη (∆t)Lη (∆t)Lξ (∆t)Qσ,i,j . (19)
3 Implementation of High Resolution Schemes

We adopt the TVD (total variation diminishing) [3] method and ENO(essentially
non-oscillatory) interpolation method developed by [4].
N 1 n n
Fσ,i+1/2 = (F + Fσ,i+1 + φσ,i+1/2 ) (20)
2 σ,i
φσ,i+ 21 = σ(āσ,i+ 12 )(gσ,i + gσ,i+1 ) + ωσ̄(āσ,i+ 21 )(dσ,i + dσ,i+1 )

−ψ(āσ,i+ 21 + γσ,i+ 21 + ωδσ,i+ 12 )4+ Qnσ,i (21)
The limiter functions gσ,i and dσ,i are defined as
gσ,i = m(∆+ Qnσ,i − ϑm̄(∆− ∆+ Qnσ,i , ∆+ ∆+ Qnσ,i ),
∆− Qnσ,i + ϑm̄(∆− ∆− Qnσ,i , ∆+ ∆− Qnσ,i )) (22)
m̄(∆− ∆− Qnσ,i , ∆+ ∆− Qnσ,i ), if |∆− Qnσ,i | ≤ |∆+ Qnσ,i |

dσ,i = (23)
m̄(∆− ∆+ Qnσ,i , ∆+ ∆+ Qnσ,i ), if |∆− Qnσ,i | > |∆+ Qnσ,i |

s min(|a|, |b|), if sign(a) = sign(b) = s
m(a, b) = (24)
0, otherwise

a, if |a| ≤ |b|
m̄(a, b) = (25)
b, if |a| > |b|
a and b are arbitrary real numbers.
āσ,i+ 12 = ā(Qnσ,i , Qnσ,i+1 )
 n n
 Fσ,i+1 − Fσ,i , if Qnσ,i 6= Qnσ,i+1
= Qnσ,i+1 − Qnσ,i (26)
 n
Vσ,i , if Qnσ,i = Qnσ,i+1

|z|, if |z| ≥
ψ(z) = (27)
(z 2 + 2 )/2, if |z| <
σ(z) = (ψ(z) − λz 2 )/2 (28)
2 3 2
(λ |z| − 3λ|z| + 2|z|)/6, if |∆− vin | ≤ |∆+ vin |
σ̄(z) = 2 3 (29)
(λ |z| − |z|)/6, if |∆− vin | > |∆+ vin |
σ(āσ,i+ 12 )(gσ,i+1 − gσ,i )/∆+ Qnσ,i , if ∆+ Qnσ,i 6= 0

γσ,i+ 12 = (30)
0, otherwise
High Resolution Quantum Kinetic Beam Schemes 201
Fig. 1. Number density isolines of quantum gas with the same fugacity in state 1
Fig. 1. Number density isolines of quantum gas with the same fugacity in state 1:
:Bose-Einstein gas(left),
Bose-Einstein Fermi-Dirac
gas(left), Fermi-Diracgas(right)
gas(right)
n
if ∆+ Qnσ,i 6=

[Chou97] Chou,
δσ,i+ 12 S.
σ̄(ā Baganoff,
= Y. & σ,i+
1 )(dσ,i+1
2
D.−(1997)
dσ,i )/∆Kinetic
+ Qσ,i , flux-vector 0
splitting for
(31)the
Navier-Stokes equations, 0, J. Comput. Phys. 130, pp.otherwise 217-230.
[Des86] Deshpande,
∆± Qnσ,i = ±(Q S. (1986)
n NASA-TM
σ,i±1 − Qσ,i )
n 01234. (32)
[Har83] Harten, A. (1983) High resolution schemes for hyperbolic conservation
where Comput.
laws,J. is a small positive
Phys., 49, real
357. number. The setting of the parameters ω and
ϑ inHarten,
[Har87] the TVD A., and ENO schemes
Engquist, B., Osher,areS. & Charkravathy, S. R. (1987) Uniformly

high order
 ω =accurate
0, ϑ = essentially
0 → T Vnon-oscillatory
D2 schemes, III, J. Comput. Phys.,
71, 231.ω = 0, ϑ = 0.5 → EN O2 (33)
ω =C.
[Hir88] Hirsch, 1, (1988)
ϑ = 0 Numerical → EN O3 Computation of Internal and External Flows,

Vol. I & II, Wiley.
[KB62] Kadanoff, L. P. & Baym, G. (1962) Quantum Statistical Mechanics, Chap.
6.4Benjamin,
Numerical New York.
Examples and Discussions
[Ohw02] Ohwada, T. (2002) On the construction of kinetic schemes, J. Comput.
Phys. 177, pp. 156-175.
The numerical examples are performed in two dimension. The unsteady quan-
[Pul80]
tumPullin, D. I.wave
gas shock (1980)pass J. byComput. Phys.
a circular 34, 231-244.
cylinder. Initially, the shock with Mach
[PX93] Prendergast,
number Ms = 2.0 K.isH.ahead& Xu, K.(1993)
of the circularNumerical
cylinder. Ahydrodynamics
simple cylindrical from gas-
grid
kinetic
systemtheory,
of 361J.×Comput.
241 was Phys. used, 109, 53. of 361 rays around the cylinder
consisting
[Rei81]
andReitz, R. D.(1981)
241 circles between J. Comput. Phys.
the cylinder 42, 108.
surface and outer boundary which is
[SP74]slightly
Sanders, R. H. &
stretched withPrendergast,
∆rmin = .007. K. H.The 1974 The possible
diameter of the relation
cylinder ofis the 1.0 3-
kiloparsec arm to explosions
and the distance between the in the galactic
origin of thenucleus,
cylinderAstrophys. J. 188,
and the outer 489.
bound-
[Tor99]
aryToro,
is 7.0.E.The
F. (1999)
fugacity Riemann Solvers and Numerical
of the Bose-Einstein gas in theMethods for Fluid
left part(state Dy-
1) is
namics, Springer.
z1 = 0.8,and in right part (state 2) is 0.744. The corresponding macroscopic
[UU33] Uehling,are
quantities E. A.(ρ1&, uUhlenbeck,
1 , 1 , T1 ) =G. E. (1933)
(1.205, Transport
0.000, phenomena
0.603, 0.749), (ρ2 , u2 , in
2 , Einstein-
T2 ) =
Bose and1.000,
(2.410, Fermi-Dirac
4.214, 1.763).gases.It I,isPhys. Rev.that
observed 43, the
552.complicated flow interac-
[Xu13]tion
Xu, K. & Prendergast,
resulting in Mach shocks, K. H.second
(1994)contact
Numerical Navier-Stokes
discontinuities, andsolutions
triple point from
gas-kinetic theory, J.For
were well captured. Comput.
the case Phys. 114, 9. gas, The initial conditions are
of Fermi-Dirac
[YS06] Yang, J. Y. & Shi, Y. H. (2006) A kinetic Beam Scheme for Ideal Quantum
Gas Dynamics, Proc. Roy. Soc. A 462, 1553-1572.
set as Ms = 2.0 and fugacity z1 = 0.8 and the flow conditions (ρ1 , u1 , 1 , T1 ) =
(0.254, 0.000, 0.127, 0.432), (ρ2 , u2 , 2 , T2 ) = (0.508, 1.000, 0.889, 1.116) The fu-
gacity in state 2 is 0.576.
References
1. Chou, S. Y. & Baganoff, D. (1997) Kinetic flux-vector splitting for the Navier-
Stokes equations, J. Comput. Phys. 130, pp. 217-230.
2. Deshpande, S. (1986) NASA-TM 01234.
3. Harten, A. (1983) High resolution schemes for hyperbolic conservation laws,J.
Comput. Phys., 49, 357.
4. Harten, A., Engquist, B., Osher, S. & Charkravathy, S. R. (1987) Uniformly
high order accurate essentially non-oscillatory schemes, III, J. Comput. Phys.,
71, 231.
5. Hirsch, C. (1988) Numerical Computation of Internal and External Flows, Vol.
I & II, Wiley.
6. Kadanoff, L. P. & Baym, G. (1962) Quantum Statistical Mechanics, Chap. 6.
Benjamin, New York.
7. Ohwada, T. (2002) On the construction of kinetic schemes, J. Comput. Phys.
177, pp. 156-175.
8. Pullin, D. I. (1980) J. Comput. Phys. 34, 231-244.
9. Prendergast, K. H. & Xu, K.(1993) Numerical hydrodynamics from gas-kinetic
theory, J. Comput. Phys. 109, 53.
10. Reitz, R. D.(1981) J. Comput. Phys. 42, 108.
11. Sanders, R. H. & Prendergast, K. H. 1974 The possible relation of the 3-
kiloparsec arm to explosions in the galactic nucleus, Astrophys. J. 188, 489.
12. Toro, E. F. (1999) Riemann Solvers and Numerical Methods for Fluid Dynam-
ics, Springer.
13. Uehling, E. A. & Uhlenbeck, G. E. (1933) Transport phenomena in Einstein-
Bose and Fermi-Dirac gases. I, Phys. Rev. 43, 552.
14. Xu, K. & Prendergast, K. H. (1994) Numerical Navier-Stokes solutions from
gas-kinetic theory, J. Comput. Phys. 114, 9.
15. Yang, J. Y. & Shi, Y. H. (2006) A kinetic Beam Scheme for Ideal Quantum
Gas Dynamics, Proc. Roy. Soc. A 462, 1553-1572.
Semi-GLS Stabilization of FEM Applied to
Incompressible Flows with Higher Reynolds
Numbers
Pavel Burda1 , Jaroslav Novotný2 , and Jakub vSı́stek1

1
Dept. of Math., Czech University of Technology, Karlovo nám. 13, CZ-12135
Praha 2, Czech Republic burda@fsik.cvut.cz, sistek@seznam.cz
2
Institute of Thermomechanics, Czech Academy of Science, Dolejvskova 5,
CZ-18200 Praha 8, Czech Republic novotny@bivoj.it.cas.cz
Summary. We deal with 2D flows of incompressible viscous fluids with higher

Reynolds numbers. Galerkin Least Square technique of stabilization of the finite
element method is modified to semi-GLS stabilization. Results of numerical exper-
iments are presented. Positive as well as negative properties of stabilization are
discussed, esp. the loss of accuracy is carefully traced, using a posteriori error esti-
mates.
Key words: incompressible fluid, finite element method, stabilization, GLS,

semi-GLS
1 Introduction
Reliable modelling of flow for high Reynolds numbers, which appear in en-
gineering practice, is still not satisfactorilly solved. In the paper we employ
stabilization technique derived from Galerkin Least Squares (GLS) method.
In our paper [3] we modified the GLS method introduced in [7]. Our semi-GLS
algorithm alows to solve flows with markably higher Reynolds numbers than
without stabilization. Unfortunately the stabilization process leads to the loss
of accuracy, as was shown in [3]. In this paper we make use of a posteriori error
estimates of the finite element solution in order to trace the loss of accuracy of
the stabilized algorithm. We present some results of numerical experiments to
show the impact of this stabilization on the solution, namely on the accuracy
of the solution.

204 Pavel Burda, Jaroslav Novotný, and Jakub vSı́stek
2 Model Problem
Let Ω be an open bounded domain in R2 filled with an incompressible viscous
fluid, and let Γ be its boundary. Isothermal flow of such fluid is governed by
the following Navier-Stokes system
∂u
+ (u · ∇)u − ν∆u + ∇p = f in Ω × [0, T ] (1)
∂t
∇ · u = 0 in Ω × [0, T ] (2)
u = g on Γg × [0, T ] (3)
−ν(∇u)n + pn = 0 on Γh × [0, T ] (4)
u = u0 in Ω, t = 0 (5)
where
• t denotes time variable
• u = (u1 , u2 )T denotes the vector of flow velocity
• p denotes the pressure divided by the density
• ν denotes the kinematic viscosity of the fluid supposed to be constant
• f denotes the density of volume forces per mass unit
• Γg and Γh are two subsets of Γ satisfying Γ = Γ g ∪ Γ h , µR1 (Γg ∩ Γh ) = 0
• n denotes an outer normal vector R to the boundary Γ with unit lenght
• g is a given function satisfying Γ g · ndΓ = 0 in the case of Γ = Γg
• u0 is a given flow field satisfying ∇ · u0 = 0
3 Stabilized Formulation
Following ideas of Hughes et al. [7], we apply the GLS stabilizing technique
with some modifications. Applying stabilization to the momentum equation
(1) and adding the continuity equation (2), we introduced in [3] the stabilized
algorithm that we call semi-GLS:
Find uh (t) ∈ Vgh , t ∈ [0, T ] and ph (t) ∈ Qh , t ∈ [0, T ] satisfying for any
t ∈ [0, T ]
BsGLS (uh , ph ; vh , ψh ) = LsGLS (vh , ψh ), ∀vh ∈ Vh , ∀ψh ∈ Qh (6)
uh − ugh ∈ Vh (7)
where
Z Z
∂uh
BsGLS (uh , ph ; vh , ψh ) ≡ · vh dΩ + (uh · ∇)uh · vh dΩ +
∂t
Z Z Ω Z Ω
ν ∇uh : ∇vh dΩ − ph ∇ · vh dΩ + ψh ∇ · uh dΩ +
Ω Ω Ω
N Z
X ∂uh
+ (uh · ∇)uh − ν∆uh + ∇ph · τ [(uh · ∇)vh − ν∆vh + ∇ψh ] dΩ ,
TK ∂t
K=1
Semi-GLS Stabilization of FEM for Flows with Higher Reynolds Numbers 205
Z N Z
X
LsGLS (vh , ψh ) ≡ f · vh dΩ + f · τ [(uh · ∇)vh − ν∆vh + ∇ψh ] dΩ .
Ω K=1 TK
Here Vh , Qh mean Hood-Taylor element spaces, τ is positive stabilization pa-

rameter (more details in [3]).
4 Measuring of Error
To review the accuracy of the semi-GLS algorithm, we use a posteriori error

estimates to evaluate the obtained error on elements. Suppose now, that the
exact solution of the problem is denoted as (u1 , u2 , p) and the approximate
solution obtained by the FEM as (u1h , u2h , ph ). The exact solution differs from
the approximate solution in the error (eu1 , eu2 , ep ) = (u1 −u1h , u2 −u2h , p−ph ).
For the solution (u1 , u2 , p) we denote
U2 (u1 , u2 , p) = k(u1 , u2 )k2H 1 (TK ) + kpk2L2 (TK ) =
Z 2 2 2 2 ! Z
2 2 ∂u 1 ∂u 1 ∂u 2 ∂u 2
= u1 + u2 + + + + dΩ + p2 dΩ.
TK ∂x1 ∂x2 ∂x1 ∂x2 TK
The following estimate of error is used (see e.g. [1])

U2 (u1 − u1h , u2 − u2h , p − ph ) ≤ E2 (u1h , u2h , ph ) (8)
where
U2 (u1 − u1h , u2 − u2h , p − ph ) = k(eu1 , eu2 )k2H 1 (TK ) + kep k2L2 (TK ) ,
h Z
E2 (u1h , u2h , ph ) = C h2K <21 (u1h , u2h , ph ) + <22 (u1h , u2h , ph ) dΩ +

TK
Z i
+ <23 (u1h , u2h , ph )dΩ
TK
where <1 , <2 , and <3 stand for the residuals

2
∂ 2 u1h

∂u1h ∂u1h ∂ u1h ∂ph
<1 (u1h , u2h , ph ) = fx1 − u1h + u2h +ν + −
∂x1 ∂x2 ∂x21 ∂x22 ∂x1
2
∂ 2 u2h

∂u2h ∂u2h ∂ u2h ∂ph
<2 (u1h , u2h , ph ) = fx2 − u1h + u2h +ν + −
∂x1 ∂x2 ∂x21 ∂x22 ∂x2
∂u1h ∂u2h
<3 (u1h , u2h , ph ) = +
∂x1 ∂x2
and constant C is determined from a numerical experiment (cf. [1]).
Our way to ‘measure’ the error on elements is to compute the error related
to the numerical solution, i.e. relative error. This is given by the ratio of
absolute norm of the solution error related to unit area of element TK
1
E2 (u1h , u2h , ph , TK )
|TK |
to the solution norm on the whole domain Ω related to unit area of Ω
1
U2 (u1h , u2h , ph , Ω)
|Ω|
i.e.
|Ω| E2 (u1h , u2h , ph , TK )

R2 (u1h , u2h , ph , TK ) = (9)
|TK | U2 (u1h , u2h , ph , Ω)
5 Numerical Results
5.1 Steady Solution of Lid Driven Cavity
We presented a series of results for steady case of the problem of ‘lid driven
cavity’ in the paper [3]. Here we show the errors on elements measured by
means of a posteriori error estimates explained in Section 4. On Fig. 1 we
show the streamlines and pressure contours obtained by semi-GLS algorithm
on the mesh 128 × 128 elements for the cavity problem with Re = 10 000.
On Fig. 2 we present, for the same data, the errors on elements done by
a posteriori estimates. We can observe that the accuracy is violated not only
at the upper corners (where singularities arise) but also at other unexpected
places.
5.2 Unsteady Solution of Flow Past NACA 0012 Airfoil
Unsteady flow past NACA 0012 airfoil was investigated in [3] as a practical
application. In [3] the results of Guermond and Quartapelle [5] were com-
pared to ours obtained by the semiGLS algorithm. To show the power of the
algorithm, we present in Fig. 3 the streamlines and pressure contours for the
problem with Reynolds number 100 000.
6 Conclusion
The developed algorithm is suitable to solve flow problems for higher Reynolds
numbers, approximately ‘double’ than the method without stabilization. In
[3], [8] a lot of numerical tests are given. They show that the stabilization has
also some negative effects and we have to be aware of them. Important tool
for testing the reliability of the numerical solution are the a posteriori error
estimates.
Acknowledgement: This work has been supported by the grant No.
1ET400760509 of the Information Society Program of the Czech Acad. Sci.
and by the State Research Project No. MSM 684 0770010.
Semi-GLS Stabilization of FEM for Flows with Higher Reynolds Numbers 207
Fig. 1. Streamlines (left) and pressure contours (right) by the semi-GLS algorithm,
Re = 10 000, mesh 128 × 128
0.8
0.6
AEE
6
5
4
0.4 3
2
1
0
0.2
0
0 0.2 0.4 0.6 0.8 1
Fig. 2. A posteriori estimates error: cavity, semi-GLS, Re= 10 000, mesh 128 × 128
Fig. 3. Streamlines (left) and pressure contours (right) by the semi-GLS algorithm,
t = 3.6s, Re = 100 000
References
1. Burda, P., Novotný, J., Sousedı́k, B.: A posteriori error estimates applied to
flow in a channel with corners. Mathematics and Computers in Simulation, 61,
375-383 (2003)
2. Burda, P., Novotný, J., vSı́stek, J.: Precise FEM solution of a corner singularity
using an adjusted mesh. Int. J. Numer. Meth. Fluids, 47, 1285–1292 (2005)
3. Burda, P., Novotný, J., vSı́stek, J.: On a modification of GLS stabilized FEM for
solving incompressible viscous flows. Int. J. Numer. Meth. Fluids, 51, 1001–1016
(2006)
4. Franca, L.P., Madureira, A.L.: Element diameter free stability parameters for
stabilized methods applied to fluids. Comput. Methods Appl. Mech. Engrg.,
105, 395–403 (1993)
5. Guermond, J.L., Quartapelle, L.: Calculation of viscous incompressible viscous
flow by an unconditionally stable projection FEM. J. Comp. Phys., 132, 12–33
(1997)
6. Hughes, T.J.R., Franca, L.P., Balestra, M.: A new finite element formulation for
computational fluid dynamics: V. Circumventing the Babuvska-Brezzi condition:
A stable Petrov-Galerkin formulation of the Stokes problem accommodating
equal-order interpolations. Comput. Methods Appl. Mech. Engrg., 59, 85-99
(1986)
7. Hughes, T.J.R., Franca, L.P., Hulbert, G.M.: A new finite element formulation
for computational fluid dynamics: VIII. The Galerkin/least-squares method for
advective-diffusive equations. Comput. Methods Appl. Mech. Engrg., 73, 173-
189 (1989)
8. vSı́stek, J.: Stabilization of finite element method for solving incompressible
viscous flows. MSc. Thesis, Czech University of Technology, Praha (2004)
Finite volume box scheme for a certain class of
nonlinear conservation laws in mixed form
H. Beaugendre1 and A. Ern2

1
MAB, Université Bordeaux 1, 351 cours de la Libération, 33405 Talence Cedex,
France, Heloise.Beaugendre@math.u-bordeaux1.fr
2
CERMICS, Ecole Nationale des Ponts et Chaussées, ParisTech, 6 et 8 avenue
Blaise Pascal, 77455 Marne la Vallée Cedex 2, France, ern@cermics.enpc.fr
1 Introduction
We develop a finite volume box scheme to approximate a certain class of non-

linear conservation laws in the form ∂t ψ(p) − ∇ · u(p, ∇p) = f . The proposed
scheme was first introduced in [1] for the Poisson equation (for which ψ(p) ≡ 0
and u(p, ∇p) = ∇p). The principles of the scheme are first to use degrees of
freedom located on mesh faces (or edges in two dimensions) for both the pri-
mal variable p and its flux u, and second to build the discrete equations by
averaging the continuous ones onto ”boxes”. In this sense the scheme can be
viewed as a finite volume scheme ensuring a local conservation property at the
mesh level. The scheme can also be viewed as a non-conforming mixed finite
element scheme [2]. The two trial spaces, that is the spaces where the discrete
solutions are sought, are the non-conforming Crouzeix-Raviart finite element
space for the primal variable p and the div-conforming Raviart-Thomas finite
element space (of lowest-order) for the flux u. The test spaces, that is the
spaces where the test functions are taken, are spanned by piecewise constant
functions. A key property of the scheme is that the flux can be locally elimi-
nated from the mixed formulation. In other words, it is possible to solve first
for the discrete primal variable in the Crouzeix-Raviart finite element space
and then to reconstruct the discrete flux. This amounts to the post-processing
of mixed finite element methods introduced in [3]. Moreover, since the recon-
structed discrete flux is in the div-conforming Raviart-Thomas finite element
space, the proposed scheme is locally conservative, which is important in many
applications. The outline of this paper is as follows. Section 2 presents the
finite volume box scheme for the above class of nonlinear conservation laws.
Applications to a specific PDE modeling subsurface flow dynamics in variably
saturated porous media, namely Richards’ equation, are proposed in Section
3. Section 4 addresses the validation of the finite volume box scheme on a
one-dimensional test case with an analytical solution to Richards’ equation.

210 H. Beaugendre and A. Ern
Results obtained on two-dimensional test cases with heterogeneous hydraulic

permeability are also discussed. Section 5 draws the conclusions.
2 The finite volume box scheme

Let Ω denote a domain in Rd with boundary ∂Ω. Let f be a square integrable
function of space and time. This section generalizes the scheme designed in
[1, 2] for the Poisson problem to unsteady nonlinear conservation laws written
in the form

 ∂t ψ(p) − ∇ · u(p, ∇p) = f , on Ω
u(p, ∇p) = −[A(p)∇p + B(p)] , on Ω (1)
p = g1 on ∂Ω1 and u · n = g2 on ∂Ω2

where A(p) is a d × d tensor, B(p) is a vector of dimension d; the components

of A(p) and B(p) are functions of p; p is called the primal variable and u
the flux. ψ is a function of p and n is the outward normal to Ω. An initial
condition for p is imposed in the form p = p0 at t = 0 in Ω. For the sake of
simplicity, homogeneous boundary conditions are considered, i.e., g1 and g2
are taken equal to zero. In addition, to avoid singularities, the initial condition
is assumed to be compatible with the boundary conditions.
The main idea when writing Problem (1) in weak form is to seek the flux
and the pressure in spaces that yield sufficient regularity to avoid integration
by parts (in space). In a steady-state setting, this amounts to seeking the pri-
1
mal variable in the Sobolev space H∂Ω 1
= {q ∈ H 1 (Ω); q = 0 on ∂Ω1 } and the
flux in the space Hdiv,∂Ω2 = {v ∈ [L2 (Ω)]d ; ∇ · v ∈ L2 (Ω); v · n = 0 on ∂Ω2 }.
In an unsteady setting, the approach can be extended by considering the fol-
lowing mixed formulation: Find functions u(t), p(t) that are smooth enough
in time and such that for every tε0,
 1

 (u(t), p(t)) ∈ Hdiv,∂Ω2 × H∂Ω 1
and
Z
 Z
u(p, ∇p) · v + [A(p)∇p + B(p)] · v = 0 , ∀v ∈ [L2 (Ω)]d ,

(2)
 Z Ω Z Ω Z

∂t ψ(p)q − q∇ · u(p, ∇p) = f q , ∀q ∈ L2 (Ω).



Ω Ω Ω
2.1 Design of the scheme

Let (Th ) be a shape-regular family of triangulations of Ω. For the sake of
simplicity, we use the terminology of the 2D case, being understood that in
the 3D case, triangles should be replaced by tetrahedra, edges by faces, etc.
Let Eh , Ei and Eb denote respectively the set of edges, internal edges and
boundary edges in Th . For an edge e ∈ Eh , let Te be the set of elements in Th
containing e. For an element T ∈ Th , let |T | be its area. For e ∈ Eh , choose
a unit normal vector ne . For a piecewise continuous function ϕ on Th , [ϕ]e
denotes the jump of ϕ across e in the direction of ne , with the convention that
a zero outer value is taken for edges contained in ∂Ω.
Finite volume box scheme for nonlinear conservation laws 211
Recall that the lowest-order div-conforming Raviart-Thomas finite element

space is defined as
0
= uh ∈ [L2 (Ω)]d ; ∀T ∈ Th , uh |T ∈ RT 0 (T ); ∀e ∈ Ei , [uh · ne ]e = 0;

RTc,∂Ω 2
∀e ∈ Eb ∩ ∂Ω2 , uh · ne = 0} ,
where RT 0 (T ) = [P 0 (T )]2 +(x, y)t P 0 (T ), and that the non-conforming Crouzeix-
Raviart finite element space is defined as
Z
Pnc,∂Ω1 = ph ∈ L (Ω); ∀T ∈ Th , ph |T ∈ P (T ); ∀e ∈ Ei , [ph ]e = 0;
1 2 1
e
Z
∀e ∈ Eb ∩ ∂Ω1 , ph = 0 ,
e
where P (T ) is the space of polynomials on T with degree ≤ 1. Let P 0 be the

1
space spanned by piecewise constant functions. Let p0h ∈ Pnc,∂Ω 1

1
be a suitable
approximation to the initial data p0 . Then, the time-marching finite volume
box scheme consists of solving sequentially in time the following problems: for
n ≥ 1,
Seek (unh , pnh ) ∈ RT 0 1
c,∂Ω2 × Pnc,∂Ω1 such that



R X Z
un · vh +
Ω h
[A(pnh )∇pnh + b(pnh )] · vh = 0, ∀vh ∈ [P 0 ]d , (3)

 T ∈Th T
ψ(pnh )qh + δt Ω qh ∇ · unh = δt Ω f n qh + Ω ψ(pn−1 )qh , ∀qh ∈ P 0 ,
R R R R
Ω h
i.e. an implicit Euler scheme is employed to approximate the second equation

in time and the test functions in both equations are piecewise constant in
space. It is important to emphasize the fact that Problem (3) is never solved
in practice. Instead, a problem of smaller size (see (5) below) is solved first,
and then the discrete flux is reconstructed using (6).
2.2 Properties of the scheme
For a function φ ∈ L2 (Ω) we denote by Πh φ its orthogonal projection onto

P 0 (Th ), i.e., Πh φ is the piecewise constant function such that
Z
1
∀T ∈ Th , Πφ φ|T = ΠT φ = φ. (4)
|T | T
The same notation is used for space-time functions.
Theorem 1. For all n ≥ 1, the solution (unh , pnh ) to (3) satisfies the following
properties:
1. Equivalence with the non-conforming approximation of the pri-
mal formulation: pnh is the solution of the problem: For n ≥ 1,
pnh ∈ Pnc,∂Ω
1

 Seek
Z 1
such that


−1
δt Πh ψ(ph ) − ψ(pn−1
n


h ) qh +
Ω Z Z (5)
X
n n n
Πh (f n )qh , 1

[A(p )∇p + B(p )] · ∇q = ∀qh ∈ Pnc,∂Ω .

h
 h h h 1


T ∈Th T Ω
2. Local flux reconstruction formula:

∀T ∈ Th , unh |T = −ΠT A(pnh )∇pnh + B(pnh ) + (6)
1
ΠT f n − δt−1 ψ(pnh ) − ψ(pn−1

h ) πT (x).
d
where πT (x) = (x1 − gT,1 , ..., xd − gT,d ), (x1 , ..., xd ) denote the cartesian
coordinates, and gT is the barycenter of T .
3. Local mass conservation

∀T ∈ Th , ΠT δt−1 ψ(pnh ) − ψ(pn−1 ) + ∇ · unh − f n = 0.

h (7)

or, in other words, ∇ · unh |T = ΠT − δt−1 ψ(pnh ) − ψ(pn−1 ) + fn .

h
Statement 3 in the previous theorem means that starting with the discrete
primal variable solving (5), the local flux reconstruction formula (6) yields a
locally conservative flux. Observe that a locally conservative reconstruction of
the discrete flux is, in general, not possible when working with more standard
P 1 conforming finite elements to approximate the primal variable.
3 An example: Richards’ equation

An important aspect of hillslope hydrology during rainfall episodes is the fact
that subsurface flows can saturate the soil in various regions near the surface
and, therefore, contribute to the production of overland flow. A widely used
model to study variably saturated flows in porous media is Richards’ equation
∂t θ(p) + ∇·u(p) = f , (8)

where θ is the volumetric water content (dimensionless), p the soil water
pressure head (L) and u(p) the Darcy flow velocity (LT−1 ) defined by
u(p) = −k(p)∇(p + z) , (9)
where k is the hydraulic conductivity, z the vertical coordinate (upwards)
and f represents the volumetric water sources or sinks in Ω. The hydraulic
conductivity is written in the form k(p) = ks kr (p) where ks is the hydraulic
conductivity at saturation and kr (p) is the relative hydraulic conductivity
tensor. Eqs. (8-9) require the knowledge of the soil water retention curve,
p 7→ θ(p), and the relative hydraulic conductivity function, p 7→ kr (p).
Problem (8) and (9) can be written in the mixed form (1) of Section 2
with ψ(p), A(p), B(p) defined as follows
ψ(p) = θ(p) A(p) = ks kr (p) B(p) = ks kr (p)ez (10)
where ez is the upward unit vector.
Finite volume box scheme for nonlinear conservation laws 213
-0.6 1
Analytical Conforming
Num Non conforming
-0.8
Error L2L2
-1
-Z (m)
0.1
-1.2
-1.4
-1.6 0.01
-4.5 -4 -3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 10 20 30 40 50 60 70 80
Pressure head (m) Number of nodes
Fig. 1.
4 Numerical test cases

Barry analytical test case. An analytical unstationary solution to the
Richards’ equation for a one-dimensional column test case is proposed in [4].
On the left of Figure 1 we compare our numerical solution to the exact solu-
tion. The analytical and numerical solutions are in very close agreement.
Results obtained with the finite volume box scheme are also compared with
those obtained using P 1 conforming finite elements. The two schemes behave
similarly: the L2 (t0 , T ); L2 (Ω) -norm of the relative error in pressure head

decreases as the reciprocal of the number of nodes. Furthermore the error is

of the same order of magnitude for both schemes.
Heterogeneous unstationary 2D test
case. In this test case we explore the ability
of the finite volume box scheme to deal with
impermeable
discontinuities in permeability. The aquifer layer
is a square (100 m × 100 m). We assume

a constant head equal to 0 m for the left
bottom corner and -4 m for the right up-
per corner. The soil consists of Sand OW for conductivity : fixed
head
which the van Genuchten relative hydraulic 5 m/h 0.05 m/h
conductivity law decreases very slowly with

Fig. 2.
the pressure [5]. Figure 3 (from left to right)
shows the soil water pressure head distribu-
tion for the initial condition, after 60 hours and at equilibrium. The evolution
of θ along the diagonal (from (0, 0) to (100, 100)) is presented in figure 4, left.
In the middle of figure 4 we display the evolution of inflow and outflow Z show-

ing the obtention of a steady state. On the right we compare max div(u)
T T
computed using the finite volume box scheme with the value computed using
P 1 conforming finite elements. We notice that the finite volume box scheme
ensures mass conservation, but not the conforming finite element scheme.
Xd3d 8.2.1 (14 May 2004) Xd3d 8.2.1 (14 May 2004)
Xd3d 8.2.1 (14 May 2004)
0 0 0
-0.2 -0.4 -0.4
-0.4 -0.8 -0.8
-0.6 -1.2 -1.2
-0.8 -1.6 -1.6
-1. -2. -2.
-1.2 -2.4 -2.4
-1.4 -2.8 -2.8
-1.6 -3.2 -3.2
-1.8 -3.6 -3.6
-2. -4. and < -4. and <
Fig. 3.
0.45 10 100
outflow box scheme
inflow 1 conformal scheme
0.4 8
0.01
6
maxdiv
0.35 1e-04
theta
Flux
initial condition 1e-06

0.3 30h 4
60h 1e-08
0.25 100h 2
350h 1e-10
steady state
0.2 0 1e-12
0 10 20 30 40 50 60 70 80 90 100 0 200 400 600 800 1000 0 200 400 600 800 1000
x time time
Fig. 4.
Conclusions
Keeping the mesh fixed, the finite volume box scheme is more expensive than
the P 1 conforming finite element scheme. However the finite volume box
scheme uses the degrees of freedom more efficiently and ensures local mass
conservation. Further work will deal with the following question: how much
is a conservative flow field needed for accurate, low-order pollutant transport
simulations?
References
1. Courbet, B. and Croisille, J.-P.: Finite volume box schemes on triangular meshes,
Math. Mod. Numer. Ana., 32(5), 631-649, June (1998)
2. Croisille, J.-P.: Finite volume box schemes and mixed methods, Math. Mod.
Numer. Ana., 31(5), 1087–1106, (2000)
3. Marini, L.D.:An inexpensive method for the evaluation of the solution of the
lowest order Raviart-Thomas mixed method, SIAM J. Num. An., 22, 493-496
(1985)
4. Ginzburg, I., Carlier, J.-P. and Kao, C.: Lattice Boltzmann approach to
Richards’ equation, CMWR Proceedings, North Carolina, USA, (2004)
5. Beaugendre, H. , Ern, A., Esclaffer, T., Gaume, E., Ginzburg, I. and Kao, C.: A
seepage face model for the interaction of shallow water tables with the ground
surface: Application of the obstacle-type method, J. of Hydrology, April (2006)
Numerical study of the Colocated Clustered
Finite Volume Scheme
O. Touazi, E. Chénier, and R. Eymard
LETEM, Université de Marne-la-Vallée, 77454 Marne-la-Vallée Cedex, France

touazi,chenier,Eymard@univ-mlv.fr
Summary. The classical formulation of the finite volume scheme for incompressible
flows in simple geometries is based on orthogonal grids with staggered meshes [1]
to avoid the growth of unphysical oscillations, also called ”spurious checkerboards
modes”. For more complex domains, the use of unstructured meshes seems to be
more judicious but it raises the problem of the definition of the staggered grids. In
case of unstructured meshes, a colocated approach is preferred. This arrangement has
become popular since the works of Rhie and Chow [3] that formulated a numerical
method ”Momentum interpolation” to avoid the apparition of unphysical pressure
oscillations. Numerous finite volume schemes have been performed with the Rhie and
Chow interpolation or its variants, as for example the CFD code Fluent [4]. This
paper presents a colocated and clustered finite volume method for unstructured grids
and some numerical tests.
1 Low Mach number approximation for an ideal gas
Let us consider an ideal gas in an enclosure with impervious walls. The di-
mensionless low Mach number equations:
∂
ρ + ∇.(ρv) = 0 (1)
∂t
∂ RaP r
(ρv) + ∇.(ρv ⊗ v) = −∇p − ρj + P r∇.τ̄¯ (2)
∂t εT
∂ γ − 1 dP
(ρθ) + ∇.(ρθv) = + ∇.(k∇θ) (3)
∂t εT γ dt
Z Z
P (t)
dx = m(Ω), p(x, t)dx = 0 (4)
Ω 1 + εT θ Ω
with τ̄¯ = µ ∇v + ∇T v − 23 (∇.v)I , ρ = 1+ε P

and εT = ∆T

Tθ T0∗ . For conve-
nience, dimensional quantities and variables associated with the initial condi-
tion at t = 0 are noted by the upper and lower scripts ∗ and 0 , respectively. The
dimensionless variables P = P ∗ /P0∗ , p = p∗ /(ρ∗0 U 2 ), ρ = ρ∗ /ρ∗0 , v = v∗ /U

216 O. Touazi, E. Chénier, and R. Eymard
T ∗ −T ∗
and θ = ∆T 0 stand for the thermodynamic pressure, the dynamic pressure,
the density, the velocity and the reduced temperature. The thermodynamic
pressure P0∗ , density ρ∗0 and absolute temperature T0∗ satisfy the ideal gas
law. The velocity and length are scaled by the thermal diffusivity α0 and the
size of the enclosure H. The scale of the temperature fluctuation is denoted
by ∆T . The dimensionless shear viscosity µ is assumed to fulfil the Suther-
3/2 1+S/T0∗
land law approximation: µ = (εT θ + 1) 1+εT θ+S/T0∗ where S = 110.5 K
is the Sutherland constant. The variation laws of the thermal conductivity k
and shear viscosity µ, as a function of the temperature, are supposed identi-
cal. The Prandtl and Rayleigh numbers are defined by P r = µ∗0 /(ρ∗0 α0∗ ) and
Ra = ρ∗0 gεT H 3 /(µ∗0 α0∗ ).
2 Numerical scheme
The discrete equations are obtained by integrating
the continuous equations over each control volume
K and by applying the Gauss theorem. To make
easier the presentation of the numerical scheme
and to take into account the steadiness of the
solutions presented in this paper, only the space
discretization for two-dimensional domains is pre-
sented hereafter (for more details, refer to paper
[5],[6]).
One of the main objectives conducing to elabo-
rate the numerical scheme was the simplicity. The Fig. 1. Notations
simplicity is first due to the choice of a colocated
scheme: the discrete unknowns are evaluated at
the same point xK , representative of the control volume K and usually called
centre (see Fig.1). Moreover, the approximation of the flux of a scalar vari-
able, normal to the edge of a control volume σKL , only depends on the two
cells K and L having this edge in common. This particular choice of normal
flux induces two requirements:
• each interface between two adjacent control volumes is orthogonal to the
line joining their centres (σKL ⊥(xK xL )),
• xK xL · nKL > 0 with nKL being the unit vector normal to the edge σKL .
These conditions are suitable for the equations of transport-diffusion such
as the energy equation but they are not sufficient to ensure a satisfactory
convergence for the Stokes problem. To obtain a good coupling between the
velocity and the pressure field, the straight line (xK xL ) must intersect at the
centre of the edge σKL at the point xKL . Therefore, triangles can be chosen as
control volumes and the circumcenters as centre of the cells. The requirement
xK xL · nKL > 0 is automatically verified once the triangulation satisfies the
Delaunay criterion.
Numerical study of the Colocated Clustered Finite Volume Scheme 217
The second constraint imposed to the numerical scheme is that it accu-

rately approximates the Navier-Stokes and energy equations but also the con-
tinuous kinetic energy balance. In particular, the work of the pressure forces at
the steady state, or under the Boussinesq approximation (ρ = const), as well
as the work of the non-linear contribution of the inertial forces must vanish.
The contributions of the Navier-Stokes and energy equations on each in-
ternal edge of the control volumes are now detailed. The calligraphic letters
concerns the continuous contributions whereas the classical uppercase letter
refers to the discrete approximation.
Mass flux: MKL = σKL ρv(x) · nKL ds(x)
R

mKL
MKL = ρKL mKL (bKL vK + bLK vL ) · nKL − λKL (pL − pK )
dKL
where mKL is the interface measure (Fig.1). The coefficients bKL and bLK
interpolate the velocity on the edge σKL :
xKL xL nKL
bKL = , bKL + bLK = 1 (5)
dKL
The density on interface σKL , ρKL , is evaluated by a linear interpolation.
To prevent from the apparition of oscillations in our colocated scheme, the
discrete mass equation is modified by introducing a stabilization term as
proposed by Brezzi and Pitkaranta [7] and Kechkar and Silvester [8] in the
framework of the finite element methods. The principle of the stabilization
method is briefly reminded. The family of control volumes are partitioned
into disjoint subsets including groups of neighbouring control volumes,
with any algorithm ensuring that two control volumes belonging to the
same subset are not too distant one from each other [5]. For any pair
of neighbouring control volumes K and L belonging to the same cluster,
λKL = λ > 0 otherwise R λKL = 0.
Pressure force: PKL = σKL pnKL ds(x)
PKL = mKL (bKL pL + bLK pK )nKL
To avoid the work of the pressure forces, the discrete pressure gradient
must be the transpose of the divergence term in discrete mass equation.
The coefficients bKL and bLK being defined by the relations 5, the expres-
sion bKL pL + bLK pK is obviously not equal to the pressure on the edge,
except when bKL = 0.5.
Non-linear inertial force: TKL = σKL (ρv(x) · nKL ) φ ds(x)
R
φK + φL
TKL = MKL
2
To satisfy the kinetic energy balance the two following conditions are
required:
1. the edge velocity must include the stabilization term ρKL m
dKL (pL −
KL
pK ),
2. an identical weight on φK and φL must be chosen: 0.5.

Notice that this last condition implies that the transport quantity φ is
rarely expressed on the edge,
R as for the pressure term.
Normal heat flux: HKL = σKL k∇θ(x) · nKL ds(x)
mKL
HKL = kKL (θL − θK ) (6)
dKL
This particularly simple expression is due to the orthogonal properties
σKL ⊥(xK xL ) that was adopted to construct the mesh. The dimensionless
conductivity kKL at the interface σKL is computed by an harmonic or
arithmetic interpolation.
Viscous force: Let v (1) and v (2) the components of the velocity, then the ap-
proximation of the viscous term for the first component of the momentum
equation ∂1 τ (11) + ∂2 τ (21) writes:
µ h i
∇.(µ∇v (1) ) + ∂1 ( ∇.v) + ∂2 (µ∂1 v (2) ) − ∂1 (µ∂2 v (2) )
3
The discrete form of the first term is similar to Eq. (6).
The crossed partial derivatives require an approximation of the gradient
Gji = ∂i v (j) [9]:
(j) 1 X mKL (j) (j) (i) (i)
GK,i = (v − vK )(xKL − xK )
mK dKL L
L∈NK
with mK the measure of theR cell K.

(1)
- Approximation of V1KL = σKL µ3 (G11 + G22 )nKL ds(x):
1 mKL µK (1) (2) (1) (1) µL (1) (2) (1) (1)

VKL = (GK,1 + GK,2 )(xKL − xK ) + (GL,1 + GL,2 )(xL − xKL )
dKL 3 3
h i
(2) (1)
- Approximation of W1KL = σKL µ G21 nKL − G22 nKL ds(x):
R
1 mKL h (2) (2) (2) (2) (2) (2)

WKL = µK GK,1 (xKL − xK ) + µL GL,1 (xL − xKL ) −
dKL
i
(2) (1) (1) (2) (1) (1)
µK GK,2 (xKL − xK ) + µL GL,2 (xL − xKL )
3 Numerical study
Spatial accuracy. The spatial accuracy of the numerical method is checked
by simulating the natural convective flow, under the Boussinesq approx-
imation, in a [0, 1]-square enclosure with isothermal rigid walls and for
P r = 1, Ra = 1 and λ = 10−3 . The chosen analytical solution is: u(x, y) =
2π cos(πy) sin(πy) sin2 (πx), v(x, y) = −2π cos(πx) sin(πx) sin2 (πy), θ(x, y) =
sin2 (πx) sin2 (πy) and p(x, y) = sin2 (πx) sin2 (πy).
The accuracy of the numerical method presented in Fig.3 is measured by the
L∞ norm (kφK − φ(xK )k∞ = maxK |φK − φ(xK )| with φ ∈ {v (j) , p}) and H1
Numerical study of the Colocated Clustered Finite Volume Scheme 219
(a) (b) (c)
Fig. 2. Meshes, (a) Squares, (b) Regular triangles (c) Random triangles
(j)
norm (kvK − v (j) (xK )k2H1 = √K mK i (Gji (xK ) − Gji (xK ))2 ) as a function
P P
of the average mesh size h = 1/ n and for three kinds of meshes consisting of
n squares (Fig. 2(a)) or regular triangles (Fig. 2(b)) or random triangles (Fig.
2(c)). For the random grid and a fixed number of cells, a series of 50 computa-
tions were performed to obtain an average values for each norm. As foreseen,
the numerical results deteriorate with the deconstruction of the mesh. Notice
that about 2% and 50% of the cell centres are located outside of their con-
trol volumes for the regular and random triangles respectively. However, the
orders of convergence for the velocity and pressure in L∞ and H1 norms are
quite independent of the cell-shape.
Low mach number comparison. The differentially heated cavity is simu-
lated and results are compared to a benchmark (Tab. 1). The physical param-
eters are described in [10]. The distribution of the cell sizes of the rectangular
grid is sinusoidal with finer mesh in the boundary layer. The triangular mesh
consists of triangles quite regular. This mesh is therefore fewer appropriate
-0.5 1.5 0.5

Square (2) 1 Square (1.5)
-1
Regular triangles (1.8) Regular triangles (1) 0
0.5
-1.5 Random triangles (1.75)
0 Random triangles (1)
log10(||e(u)||)max
log10(||e(p)||)max
-0.5
1
-2
log10(||u||)H
-0.5
Square (1)
-2.5 -1 -1 Regular triangles (1)
-1.5 Random triangles (0.9)
-3
-2 -1.5
-3.5
-2.5
-2
-4 -3
-4.5 -3.5 -2.5
-2.8 -2.6 -2.4 -2.2 -2 -1.8 -1.6 -3 -2.5 -2 -1.5 -1 -0.5 -2.8 -2.6 -2.4 -2.2 -2 -1.8 -1.6
log10(h) log10(h) log10(h)
1 (1)
kvK − v 1 (xK )k∞ kvK − v (1) (xK )k2H1 kpK − p(xK )k∞
Fig. 3. Spatial accuracy for velocity and pressure.
than rectangles to represent the solutions in the boundary layers where the
longitudinal and transversal characteristics of the flow vary on very different
scales. This remark explains the less accuracy obtained with triangles than
rectangles. It can be however noticed than the relative discrepencies to the
Vierendels results are good and less than 1.1%.
Triangular Rectangular Vierendeels

Type of non-uniform
mesh 103 584 48 400 4.2 × 106
Nu 0.04% 0.003% 8.6866

N u(y = 0.5)(h) 0.2% 0.002% 7.4593
N u(y = 0.5)(c) 0.8% 0.003% 8.6372
N umax(h) 0.1% 0.03% 20.2704
N umin(h) 0.1% 0.05% 1.0667
N umax(c) 1.1% 0.03% 15.5194
N umin(c) 1.1% 0.08% 0.7575
c = (p̄/p0 ) 0.003% 0.001% 0.924489
Table 1. Comparisons with Benchmark [10], Ra = 106 , P r = 0.71, εT = 1.2,
Sutherland law
References
1. Harlow, F., Welch J.: Numerical Calculation of Time-Dependent Viscous In-
compressible Flow of Fluid with a Free Surface. Phys. Fluids, 8, 2182–2189
(1965)
2. Patankar, S.V : Numerical Heat Transfer and Fluid Flow. MacGraw-Hill, New-
York, (1980)
3. Rhie, C.M, Chow, W.L.: Numerical Study of the Turbulent Flow Past an Airfoil
with Trailing Edge Separation. AIAA J., 21, 1523–1532 (1983)
4. Mathur, S.R, Murthy, J.Y.: A Pressure-Based Method for Unstructured Meshes.
Num. heat Transfer, part B, 31, 195–215 (1997)
5. Chénier, E., Eymard, R., Touazi, O.: Numerical results using a Colocated Finite
Volume Scheme on Unstructured Grids for Incompressible Fluid Flows. Num.
Heat Transfer, part B, 49, 259–276, (2006)
6. Touazi, O., Chénier, E., Eymard, R.: Simulation of natural convection with the
Clustered Finite Volume Scheme. Computers & Fluids, submitted (2006)
7. Brezzi, F., Pitkaranta, J.: On the Stabilization of Finite Element Approxi-
mations of the Stokes Equations. Notes Numer. Fluid Mech., Vieweg, Braun-
schweig, 10 (1984)
8. Kechkar, N., Silvester, D.: Analysis of Locally Stabilized Mixed Finite Element
Methods for the Stokes Problem. Mathematics of Computation, 58(197), 1–10,
(1992)
9. Eymard, R., Gallouët, T., Herbin, R.: A cell-centred finite volume approxima-
tion for anisotropic diffusion operators on unstructured meshes in any space
dimension. IMAJNA, accepted for publication (2005)
10. Paillère, H., Le Quéré, P., Weisman, C., Vierendeels, J., Dick, E., Braack, M.,
Dabbene, F., Beccantini, A., Studer, E., Kloczko, T., Corre, C., Heuveline,
V., Darbandi, M., Hosseinizadeh, S.F.: Modelling of natural convection flows
with large temperature differences: A benchmark problem for low Mach number
solvers.ESAIM, Math. Model. Numer. Anal., 617–621 (2005)
Arbitrary High Order Finite Volume Schemes
on Unstructured Meshes
Michael Dumbser1 and Claus-Dieter Munz1
Institut für Aerodynamik und Gasdynamik, Universität Stuttgart, Pfaffenwaldring

21, 70550 Stuttgart, Germany. michael.dumbser@iag.uni-stuttgart.de
1 Introduction
We present a non-oscillatory finite volume (FV) scheme of arbitrary accuracy

in space and time for solving nonlinear hyperbolic systems of the form
∂ ∂
up + Fpi (uq ) = 0 (1)
∂t ∂xi
on unstructured triangular and tetrahedral grids in two and three space di-
mensions. Throughout the whole article we use classical tensor index notation.
The cell averaged quantities ūp of an element T (m) are evolved in time via a
finite volume scheme of the form
n+1
Z tZ
1
ūn+1 = ūnp − (m) Fpi uGRP

p q (t) · ni dt dA. (2)
T
∂T (m) tn
The spatial and temporal accuracy of the final scheme depends merely on
the order of the argument of the numerical flux used in (2). The key point
is the reconstruction operator presented in [2] that makes use of techniques
developed originally in the discontinuous Galerkin finite element framework.
First, it uses a hierarchical orthogonal basis to perform reconstruction. Second,
reconstruction is not done in physical coordinates, but in a reference coordi-
nate system which eliminates scaling effects and thus avoids ill-conditioned
reconstruction matrices. Third, the proposed algorithm does not reconstruct
point-values but entire polynomials which can be easily evaluated and differ-
entiated at any point, which is crucial for an efficient implementation of the
ADER approach that essentially relies on spatial derivative information. To
achieve non-oscillatory properties for nonlinear hyperbolic systems, we extend
the component-wise WENO reconstruction presented in [2] to characteristic
reconstruction. Numerical fluxes at the element interfaces are then computed
solving generalized Riemann problems (GRPs), according to the ADER ap-
proach of Toro et al., [9, 10, 8]. The resulting one-step ADER finite volume

222 Michael Dumbser and Claus-Dieter Munz
scheme obtained in this way performs only one nonlinear WENO reconstruc-
tion per element and time step and can thus be implemented very efficiently
even for unstructured grids in three space dimensions. We emphasize that the
method is of uniform high order of accuracy in space and time. The surface
integral appearing in (2) is approximated using Gaussian quadrature formulae
of suitable order of accuracy.
2 Characteristic WENO Reconstruction

The computational domain Ω is discretized by conforming elements T (m)
(triangles in 2D and tetrahedrons in 3D), indexed by a unique mono-index m.
As usual for FV schemes, we need to reconstruct higher order polynomials wp
from the given cell averages ūp to achieve high order of accuracy for the spatial
discretization. We write the reconstruction polynomial for element T (m) as
(m) (m)
wpl (ξ, η, ζ) = ŵpl Ψl (ξ, η, ζ) , (3)
where the functions Ψl are the basis functions, similar as for DG finite ele-
ments, and ξ, η and ζ are the coordinates in a reference coordinate system,
aligned with the reference element TE . The reference elements are the trian-
gles and tetrahedrons with vertices (0, 0), (1, 0) and (0, 1) in 2D and (0, 0, 0),
(1, 0, 0), (0, 1, 0) and (0, 0, 1) in 3D, respectively. Index l ranges from 0 to L−1
where L is the number of reconstructed degrees of freedom and depends on
the degree M of the reconstruction polynomial.
To perform the reconstruction for element T (m) , we choose a reconstruc-
tion stencil S(m) and then transform all elements of the stencil S(m) to the
reference coordinate system associated with the central element T (m) for which
reconstruction is to be done. The central element itself is thus mapped onto
the reference element TE and all other elements of the stencil are mapped to
some neighbors of the reference element. This mapping has the advantage that
it takes out all scaling problems. Furthermore, highly stretched regular grids
are mapped via this transformation to regular unstretched grids. This could
be beneficial for further use e.g. in boundary layer computations. For more
details on the reconstruction algorithm on unstructured meshes, especially
concerning stencil selection and stencil transformation, see [2].
To obtain the coefficients ŵpl , we require conservation for the reconstruc-
tion polynomial in all elements of the stencil and obtain finally a system of
linear equations that we solve using least squares:
Akl ŵpl = ūpk . (4)
In order to produce an essentially non-oscillatory reconstruction polyno-
mial, we use a nonlinearly weighted combination of reconstruction polynomials
obtained on different stencils, in which smoother reconstructions are preferred,
see [6, 1, 5, 7]. Whereas in [2] the reconstruction is done componentwise for
linear systems, in this section we present characteristic reconstruction for
Arbitrary High Order Finite Volume Schemes on Unstructured Meshes 223
nonlinear systems: The WENO algorithm is applied onto each characteris-

tic variable of the system, depending on the direction. On the unstructured
meshes considered here we use the face normal vectors nj as characteristic
directions, where j denotes the number of the edge for triangles in 2D and
the number of the face for tetrahedrons in 3D. On each stencil i, the degrees
of freedom defining the polynomial for each characteristic direction nj are
(i,n ) nj −1 i

ĉpl j = Rpq · ŵql , (5)
n
where Rpqj is the matrix of right eigenvectors of the hyperbolic system, evalu-
(m) (k )
ated for each face normal nj at the arithmetic average 12 (ūp + ūp j ) of the
cell averages of the considered element T (m) and the associated neighbor ele-
ment T (kj ) at face number 0 ≤ j ≤ NE , where NE is the number of faces and
where we additionally define the zeroth set of characteristic variables to be
n0
equal to the conservative variables, i.e. Rpq = δpq , with the Kronecker symbol
δpq . We note that in the following equations (6) to (12) there is no automatic
summation induced over index p. First, we compute the smoothness σ of each
characteristic polynomial from the quadratic functional according to [2],
(i,nj )
σp(i,nj ) = ĉpl (i,nj )
· Σlm · ĉpm . (6)
(i,n )
The non-normalized nonlinear weights ω̃p j are then functions of the linear
(i,n )
weights λi and the previously computed oscillation indicators σp j . Since
(i,nj ) (i,n )
the final weights ωp must sum to 1, the non-normalized weights ω̃p j
must be normalized with their sum. One therefore obtains
(i,nj )
ω̃p λi
ωp(i,nj ) = ns , with ω̃p(i,nj ) = r . (7)
(s,nj ) (i,nj )
+ σp
P
ω̃p
s=1
The linear weights are chosen according to [2], putting a large linear weight
λ1 >> 1 on the central stencil and λi = 1 else. Finally, we obtain for each
characteristic variable for each characteristic direction nj a reconstruction
polynomial of degree M that is obtained from a nonlinear combination of the
reconstruction polynomials on each stencil i:
ns
(WENO,nj ) (i,nj )
X
ĉpl = ωp(i,nj ) ĉpl , ∀0 ≤ j ≤ NE . (8)
i=1
Since in our approach the final result of the reconstruction operator is one
single reconstruction polynomial defined in the whole element, we still have
to perform a selection among the various characteristic reconstruction poly-
nomials. Therefore, we transform all characteristic WENO polynomials back
to conservative variables, i.e.
(WENO,nj ) nj (WENO,nj )
ŵpl = Rpq · ĉql , ∀0 ≤ j ≤ NE , (9)
and then choose for each conservative variable the smoothest polynomial ob-
tained from all the different characteristic reconstructions. First, the oscilla-
tion indicators for the polynomials (9) are computed in (10),
(WENO,nj )
σpnj = ŵpl (WENO,nj )
· Σlm · ŵpm , (10)
then the smoothest one is selected according to an ENO-type algorithm:
( nN

n
nj 1 if σp j = min σpn0 , σpn1 , ..., σp E ,
ωp = (11)
0 else.
The final reconstruction polynomial for the conservative variables valid in the
whole element T (m) is thus given by
NE
(WENO,nj )
X
WENO
ŵpl = ωpnj ŵpl . (12)
j=0
3 Numerical Flux Computation

The fundamental idea to obtain a high order time-accurate one-step method
is to use the solution of generalized Riemann problems for numerical flux
computation. This is a natural high order extension of the original first order
method of Godunov. Whereas in the method of Godunov the initial condition
of the Riemann problems at the element interfaces was piecewise constant,
it is piecewise polynomial for generalized Riemann problems. Those polyno-
mials are generated for each element from cell averages by the reconstruction
algorithm presented in the previous section. The key point of the GRP solvers
are a temporal Taylor series expansion of the state at the interface where then
time derivatives are replaced by space derivatives using repeatedly the gov-
erning conservation law in differential form (1). This is the so-called Cauchy-
Kovalewski procedure. However, the problem is that in general neither the
state nor the derivatives are defined on the element interfaces since jumps
are admitted. The idea is now to solve a conventional homogeneous nonlinear
Riemann problem with piecewise constant states for the leading order term in
the Taylor series and then the conservation law is linearized about the leading
order solution. Subsequently, linearized homogeneous conventional Riemann
problems are solved for all space-derivatives. This strategy defines the values
of the state and the space derivatives on the element interfaces which can
then be plugged into the Cauchy-Kovalewski procedure. For more details of
the GRP solution and the ADER approach see [9, 10] and for an efficient al-
gorithm for the Cauchy-Kovalewski procedure of the Euler equations see e.g.
[4, 3]. Therefore, we can write the time-dependent argument of the flux used
(m) (k )
in (2) as function depending on the reconstructed polynomials wq and wq j
(m) (kj )
in the considered element T and in the neighbor T ,
uGRP
p (t) = uGRP
p (wq(m) , wq(kj ) , t). (13)
4 Results for the Euler Equations
Z Z Z
Y X Y X Y X
0.5 0.5 0.5
0 0 0
z
z
-0.5 0.5 -0.5 0.5 -0.5 0.5
0.5 0.5 0.5
0 0 0
0 x 0 x 0 x
y y y
-0.5 -0.5 -0.5 -0.5 -0.5 -0.5
Fig. 1. Regular tetrahedral meshes used for the convergence study.
To assess the accuracy of the proposed method, we solve the three di-
mensional system of Euler equations with source term in the computational
3
domain Ω = − 21 ; 12 with six periodic boundary conditions using a sixth

order ADER-FV scheme based on fifth order WENO reconstruction. The

mesh topology is shown in Fig. 1. The analytical solution uep for the density
ρe , the velocity components ue , ve , we and the pressure pe of the problem
is defined to be ρe (x, t) = 1 + 0.8 sin (ωt − kx), ue = ve = we = 0 and
pe (x, t) = 1 + 0.9 sin (ωt − kx), setting k = (2π, 2π, 2π) and ω = 2π. The
source term on the right hand side of the Euler equations is computed in-
serting uep into the equations such that the prescribed analytical solution is
obtained. The reconstructed numerical solution w is compared against the
exact one after t = 0.25 using different norms. The results are shown in Table
1 for different meshes, which are characterized by the number of tetrahedral
edges NG per space dimension, and confirm very well that the designed order
of accuracy of the method has been reached.
Finally, we compute the double Mach reflection problem [5, 10] on an ir-
regular triangular mesh using a third order ADER-FV scheme in order to
show the robustness and accuracy of the proposed characteristic WENO re-
construction in the presence of strong discontinuities. The problem consists in
a single shock wave propagating with Mach number 10, hitting a 30◦ ramp.
The solution exhibits strong discontinuities and slip surfaces that start rolling
when the mesh is refined. Since for the Euler equations there is no limit for
Table 1. Numerical convergence results in 3D using a 6th order ADER-FV scheme.
NG L∞ L1 L2 O L∞ O L1 O L2
16 1.0136E-04 1.0987E-05 1.9043E-05
20 3.2029E-05 4.1034E-06 7.1727E-06 5.2 4.4 4.4
24 1.1199E-05 1.3482E-06 2.3640E-06 5.8 6.1 6.1
32 1.7327E-06 2.0130E-07 3.5644E-07 6.5 6.6 6.6
Fig. 2. Results for the double Mach reflection problem obtained with a third order
1 1
ADER-FV scheme on different grids: h = 100 (top), h = 400 bottom.
this rolling process due to a lack of physical viscosity, this test case is com-
monly used to assess the numerical viscosity of a scheme. The more rolling the
less the numerical viscosity. We compute this test case on differently refined
meshes, where the mesh spacing h is the edge length at the wall boundary. At
the boundaries, we use purely one-sided reconstruction stencils and then solve
inverse generalized Riemann problems to prescribe the correct flux at the wall,
see [4]. This is done by symmetry considerations, where point-symmetry is ap-
plied for the normal velocity and axial-symmetry for all other components.
Comparing our results with those presented in [5] we note that our third
order solution shows more rolling on the finest grid, i.e. less numerical dissi-
pation. This is probably due to the following facts: in [5] a third order TVD
Runge-Kutta time integration scheme was used which implies three nonlinear
WENO reconstructions per time step with their associated limiting, whereas
in the ADER approach only one WENO reconstruction is performed per time
step. Furthermore, the third order method presented in [5] relies on an optimal
pointwise WENO reconstruction based on the combination of first order poly-
nomials. Our third order reconstruction, however, is based on a combination
of second order polynomials.
References
1. D. Balsara and C.W. Shu. Monotonicity preserving weighted essentially non-
oscillatory schemes with increasingly high order of accuracy. Journal of Com-
putational Physics, 160:405–452, 2000.
2. M. Dumbser and M. Käser. Arbitrary high order non-oscillatory finite vol-
ume schemes on unstructured meshes for linear hyperbolic systems. Journal of
Computational Physics, 2006. in press.
3. M. Dumbser and C.D. Munz. Building blocks for arbitrary high order discon-
tinuous Galerkin schemes. Journal of Scientific Computing, 27:215–230, 2006.
4. M. Dumbser. Arbitrary High Order Schemes for the Solution of Hyperbolic
Conservation Laws in Complex Domains. Shaker Verlag, Aachen, 2005.
5. C. Hu and C.W. Shu. Weighted essentially non-oscillatory schemes on triangu-
lar meshes. Journal of Computational Physics, 150:97 – 127, 1999.
6. G.-S. Jiang and C.W. Shu. Efficient implementation of weighted ENO schemes.
7. M. Käser and A. Iske. ADER schemes on adaptive triangular meshes for scalar
conservation laws. Journal of Computational Physics, 205:486–508, 2005.
8. T. Schwartzkopff, M. Dumbser, and C.D. Munz. Fast high order ADER schemes
for linear hyperbolic equations. Journal of Computational Physics, 197:532–539,
2004.
9. E.F. Toro and V. A. Titarev. Solution of the generalized Riemann problem for
advection-reaction equations. Proc. Roy. Soc. London, pages 271–281, 2002.
10. V.A. Titarev and E.F. Toro. ADER schemes for three-dimensional nonlinear
hyperbolic systems. Journal of Computational Physics, 204:715–736, 2005.
Part III
Algorithms
A high scalability parallel algebraic multigrid
solver
T. Saad1 and M. Darwish2

1
AUB, FEA and P.O. Box 11-0236, Riad El Solh Street, Beirut 1107 2020,
Lebanon tsaad@aub.edu.lb
2
darwish@aub.edu.lb
Summary. In this paper we present the performance analysis of a parallel Al-

gebraic Multigrid Solver (AMG) for a finite volume unstructured CFD code. The
multigrid solver is part of an unstructured cell-centered finite volume code. The
parallelization of the solver is based on the domain decomposition approach using
the single program multiple data paradigm. The Message passing interface Library
(MPI) is used for communication of data. An ILU(0) iterative solver is used for
smoothing the errors arising within each partition at the different grid levels, and a
multi-level synchronization across the computational domain partitions is enforced
in order to improve the performance of the parallelized Multigrid solver. A number
of Multigrid cycles (V, W and F-Cycle) and two strategies have been implemented.
Tests on different grids have been conductied for a non-linear problems on up to 28
parallel processors. Results show, that synchronization plays across partitions for
multigrid-levels plays an essential role in ensuring good scalability. Furthermore for
large partition number gathering coefficients across partitions is important to ensure
a convergence history on par with the sequential solver.
1 Introduction
Computational Fluid Dynamics (CFD) has become an essential engineering

tool in many industries (aerospace, automotive, chemical processing, power
generation, etc ...). At its basic level CFD involves the following fundamental
steps: (i) a discretization step that translates a set of highly non-linear par-
tial differential equations representing conservation principles (conservation
of momentum, mass, energy, etc) into a sparse system of linearized algebraic
equations, (ii) and a solution step to solve the system of algebraic equations it-
eratively. Usually iterative solvers are used for solving the system of equations
(inner loop) because of their lower computational requirements in memory and
cpu time [1]. Because of nonlinearities these two steps need to be performed
repeatedly [2] (outer loop) until a final converged solution is reached. The
converged solution is then analyzed for the needed engineering information.

232 T. Saad and M. Darwish
To accelerate the convergence of the outer loop effort was expanded to de-
velop more efficient algorithms [3] to treat the inter-equation coupling of the
momentum and pressure equations either in a segregated manner [4] or more
implicitly [5]. Effort to accelerate the inner loop, have led to the development
of geometric multigrid methods in the 1970’s and later in the 1980’s to alge-
braic multigrid methods (AMG) [6] that extended the main ideas of geometric
multigrid methods to purely algebraic settings, yielding more robustness and
algorithmic simplicity. This dual approach sequential has been quite success-
ful in tackling increasingly larger cfd simulations, but the reliance of many
industries on using CFD to simulate large scale engineering problems has led
to a more drastic approach to decrease simulation time, i.e parallel processing,
which enables the scaling of the CFD problem through the decomposition of
the problem into as many sub-problems as there are processors. The Paral-
lelization of CFD codes can be achieved in several ways however when the
target of the parallelization effort is clusters of interconnected processors [7]
then the Domain Decomposition approach is remains the most effective. In
Domain Decomposition the parallelization is enforced by dividing (partition-
ing) the domain of interest into a number of sub-domains or partitions (usu-
ally one for each processor). In this case inter-domain synchronization where
neighboring sub-domains swap solution information is essential in ensuring
a consistent solution. The synchronization can be performed on one extreme
only at the finest mesh, with the multigrid solver basically playing the role of
its sequential counterpart over the sub-domain or partition, or on the other
extreme the synchronization can be performed at each multigrid level, with
the multigrid solver then playing a additional role of smoother across parti-
tions but at the expense of additional communication cost between partitions
[8]. The subject of this paper is the development of a highly scalable Alge-
braic Multigrid solver. After a brief review of the different synchronization
strategies possible across partitions results for the performance of the pAMG
solver for different grid sizes and partition numbers are presented.
2 Parallel Agglomeration Strategies
There are various approaches of parallelizing the coarse grid agglomeration

schemes The simplest strategy is to decouple the agglomerated grids across
partitions. In this case agglomeration proceeds in each partition independently
and the agglomerated levels do not communicate across partitions and thus
synchronization across partitions become a non-issue. However, the benefits
of multigridding are lost in this case, as the smoothing of the multigrid is
no more applied throughout the computational domain. In the second ap-
proach denoted by Coupled Coarsening Strategy case the synchronization is
performed across partitions for any coarsened grid level. This requires that
the total number of multigrid levels across partitions be the same, and that
agglomerated shadow be consistent with their agglomerated core elements.
A high scalability parallel algebraic multigrid solver 233
In this approach the agglomeration is restricted to core elements and starts

after the partitioning step, and is performed in parallel in the different parti-
tions, this is repeated until the coarsest grid level is reached [9]. At each level,
partitions exchange information at the interface regarding the agglomerated
shadow and sender elements.
3 Coarse Grid Parallelism

The solver on the coarsest grid can limit the ultimate speedup that can be
attained in a parallel computation for two related reasons. First the linear
system at this level is generally small and the time required for communication
may be higher that the time required to solve the system on a single processor.
Second, the coarsest grid may couple all pieces of the global problem, and thus
an accurate solution at this level is important, as is the global communication
of the right hand side. If the coarse grid is small enough instead of solving
in parallel, the coarsest grid problem may factored and solved on a single
processor with the right hand side gathered and solution scattered to the
other processors. Another option would be to solve the coarses grid problem
on all processors, this redundant form of the calculation does not require
communication to distribute the results [10].
As the number of processors is increased the coarsest grid problem may
become too large, in this case it might become preferable to do a parallel
computation. However communication complexity can be reduced by solving
with only a subset of the processors. Solving redundantly with a subset of the
processors is again an option.
4 Test Problems
To test our parallelization, we choose to solve a nonlinear diffusion problem

governed by
∇ · (Γ ∇φ) = 0 (1)
Γ = φ0.1 (2)
where the various boundary conditions are shown in Figure 1. Four un-
structured meshes were used for the geometric discretization; 99,454 - 199,222
- 298,154 - 397,393 elements. The runs were performed on a cluster of ten G5
dual processor machines running at 2.0 GHz each with a memory of 512 MB
per node. The interconnection network is a 100-MBits Ethernet switch. The
runs were made until the residual was reduced to a value of 10−4 .
Fig. 1. Test problem geometry
4.1 speedup
The speedup, for a given number of processors P, is defined as the ratio of the
solution time of the sequential algorithm to the solution time on P processors.
Fig. 2. Speedup
Figure 2 shows the speedup obtained for the four meshes using up to 16
processors. For the 99,454 elements mesh, the scalability starts deteriorating
A high scalability parallel algebraic multigrid solver 235
as of 4 processors. For the 199,222 elements mesh, the behavior is linear up to

14 processors. The scalability then starts to deteriorate as of where the ratio
of the number of shadow elements to the number of core elements (shadow-to-
core ratio hereafter) exceeds 0.02. For the 298,154 elements mesh, it is noticed
that the behavior is super linear up to 16 processors at which the shadow-
to-core ratio remains almost the same but the communication cost increases.
The 397,393 elements mesh displays the same behavior as the latter mesh.
4.2 Communication and Computational Costs
The communication and computational costs are directly related to the num-
ber of shadow elements and the number of core elements since the computation
takes place over the core elements and the communication takes place over
the shadow elements respectively.
Figure 3 shows the efficiency versus the maximum shadow-to-core ra-
tio.This graph basically characterizes the relation between the maximum
shadow-to-core ratioand the efficiency of the parallel solver. For the 99,454
elements mesh, as the shadow-to-core ratio exceeds 0.008, the efficiency starts
to decrease. For the 199,222 elements mesh, the efficiency starts to degenerate
at a shadow-to-core ratio exceeding 0.02. The same happens for the other two
meshes. In general, it is noticed that when the shadow-to-core exceeds - on
average - a value of 0.02, the efficiency starts to deteriorate.
5 Conclusion
In this paper we have presented the performance of a pAMG solver used in

a Finite Volume cell-centered unstructured CFD code. The performance of
the parallel solver was compared to that of the sequential solver and it was
shown that a substantial improvement in the solution time was gained as
long as the computational load is well distributed amongst processors and the
ratio of parallel time to the computational time does not exceed 0.25 which
is equivalent to a value of 0.02 for the ratio of shadow to core elements. The
solver scalability was shown to exceed the theoretical limit in some cases,
which is encountered in practice, as the usage of the cache memory is more
efficient for small problems. The solver behaves extremely well for a wide
range of processors.
6 Acknowledgment
The authors wish to acknowledge the generous support of the Lebanese Na-
tional Council for Scientific Research (LNCSR).
Fig. 3. Relation between efficiency and shadow-to-core ratio
References
1. Bucker H.M.: Iteratively solving large sparse linear systems on parallel com-
puters. NIC Serices, John Von Neumann Institute for Computing, Julich, 10,
521-548 (2002)
2. Patankar S.V.: Numerical heat transfer and fluid flow. Washington; New York:
Hemisphere Pub. Corp.; McGraw-Hill, 1980
3. Moukalled F., Darwish M.S.: A Unified Formulation of the Segregated Class
of Algorithms for Fluid Flow at All Speed, Num. Heat Transfer, 37, 103-139
(2000)
4. Issa, R.I.: Solution of the Implicit Discretized Fluid Flow Equations by Operator
Splitting, Mechanical Engineering Report, FS/82/15, Imperial College, London,
(1982)
5. Webster R.: Iteratively solving large sparse linear systems on parallel comput-
ers. Int. J. Numer Meth. Fluids, 36, 743-773 (2001)
6. Brandt A. ’Algebraic Multigrid Therory: THe symmetric Case’ Appl. Math.
COmput, vol 19, pp.24-56, 1986.
7. V. Dolean and S. Lanteri, ”Parallel multigrid methods for the calculation of
unsteady flows on unstructured grids: algorithmic aspects and parallel perfor-
mances on clusters of PCs,” Parallel Computing, vol. 30, pp. 503-525, 4. 2004
8. Mitchell W. ’A parallel multigrid methoid using the full domain partition’,
Electron. Trans. Numer. Anal., vol. 6, pp. 224-233, 1998
9. A. Krechel and K. Stüben. Parallel algebraic multigrid based on subdomain
blocking. Parallel Computing, 27:1009-1031, 2001
10. Gropp W. D., ’Parallel computing and domain decomposition’, in T. F. Chan,
D. E. Keyes, G. A. Meurant, J. S. Scroggs, and R. G. Voigt, eds., Fifth Confer-
ence on Domain Decomposition Methods for Partial Differentia equations, pp.
349-362, SIAM, 1992.
Jacobian-Free Newton-Krylov Methods:
Issues and Solutions
David W. Zingg1 and Todd T. Chisholm2

1
University of Toronto Institute for Aerospace Studies, 4925 Dufferin St.,
Toronto, ON M3H 5T6 Canada dwz@oddjob.utias.utoronto.ca
2
University of Toronto Institute for Aerospace Studies (currently at MDA
Corporation) todd@oddjob.utias.utoronto.ca
1 Introduction
A Newton-Krylov method computes the solution of a system of nonlinear al-

gebraic equations, often arising from a discretization of a system of partial dif-
ferential equations, using an inexact-Newton method combined with a Krylov
subspace method for linear systems. Such methods are very efficient in a va-
riety of applications, as discussed in the review paper by Knoll and Keyes [1].
However, they have not gained widespread acceptance in computational fluid
dynamics (CFD), i.e. in the numerical solution of the Reynolds-averaged com-
pressible or incompressible Navier-Stokes equations. Considerable interest was
generated during the 1990’s [2, 3, 4, 5, 6, 7], primarily using the Krylov method
GMRES [8], but current interest is more limited [9, 10, 11, 12, 13, 14, 15].
One of the reasons for the limited popularity of Newton-Krylov methods
in CFD is that there are a number of subtle issues that can significantly de-
grade or prevent the convergence of a Newton-Krylov algorithm. These can be
difficult to identify and may be responsible for unexpected failures of current
Newton-Krylov algorithms in some cases. Some are neither well understood
nor widely recognized. When these issues are successfully addressed, the re-
sulting algorithm can be very efficient. For example, the Newton-Krylov ap-
proach has been used very successfully by Nemec and Zingg [9, 11] in aerody-
namic shape optimization, where algorithm speed and reliability are essential.
The purpose of this paper is to present and discuss a number of issues that
can impair the convergence and efficiency of Jacobian-free Newton-Krylov
methods as well as strategies for identifying and addressing them.
2 Jacobian-free Newton-Krylov Methods

The basic ideas behind Jacobian-free Newton-Krylov methods are straight-
forward. Newton’s method is applied to the nonlinear system of equations

238 David W. Zingg and Todd T. Chisholm
arising from a spatial discretization of the steady flow equations. At each

inexact-Newton iteration, a linear system of equations must be solved to some
tolerance. Certain iterative methods, such as GMRES, require only the prod-
uct of the system Jacobian and a series of vectors; the actual Jacobian is
never needed. This permits a matrix-free approximation to the matrix-vector
product based on a finite-difference approximation to a Frechet derivative as
follows:
R(Q + v) − R(Q)
Av ≈ (1)

where R(Q) is the residual function, Q is the current solution vector, A is
the Jacobian based on a linearization about the current solution, v is an
arbitrary vector, and is a small parameter. Generally, the conditioning of
the Jacobian is such that preconditioning is required. An incomplete lower-
upper (ILU) factorization of some approximation to the Jacobian can be an
effective preconditioner. Alternatively, another algorithm can be used as a
preconditioner; this is usually not as effective as an ILU factorization but
saves memory. Development of efficient preconditioners remains an active area
of research. Finally, given that Newton’s method is not globally convergent,
a pseudo-transient continuation method or equivalent must be implemented;
the switched-evolution-relaxation (SER) strategy of Mulder and van Leer [16]
is popular.
Unfortunately, this basic Newton-Krylov formulation can be extremely in-
efficient unless certain key parameters and strategies are correctly chosen to
maximize speed and reliability with consideration of memory use. For exam-
ple, the linear system tolerance should be selected to avoid oversolving as
much as possible. The fill level permitted in the ILU factorization is another
important parameter. The ILU(0) factorization, in which no fill is permitted,
is popular but not particularly effective. The SER strategy introduces some
parameters controlling the increase in the pseudo-time step as the residual
decreases. However, this strategy is also not optimal, as a spike in the resid-
ual can lead to unnecessarily small time steps. Other areas that affect the
efficiency of a Newton-Krylov algorithm are the approximate Jacobian used
to form the ILU factorization and the ordering of the variables. Generally, the
approximate Jacobian includes nearest-neighbour contributions only. When
an upwind spatial discretization is used, it is natural to base the approximate
Jacobian on a first-order discretization. When a centered scheme is used with
added numerical dissipation, there is some flexibility in the nearest-neighbour
Jacobian approximation. Finally, several authors have shown that a reverse-
Cuthill-McKee (RCM) reordering of the mesh nodes can improve efficiency,
especially in conjunction with an ILU factorization.
Fortunately, the parameters and strategies described in the previous para-
graph have been studied by many researchers, and, although the pseudo-
transient stage remains a challenge, good choices are available. However, there
are three subtle issues that are rarely discussed: 1) the effect of the root node
Jacobian-Free Newton-Krylov Methods: Issues and Solutions 239
in the RCM reordering, 2) misleading convergence of the linear iterations,

and 3) inaccurate Jacobian-free approximate matrix-vector products. These
can be difficult to identify and may be responsible for unexpected failures of
current Newton-Krylov algorithms in some cases. These issues will be further
discussed after a brief description of the present Newton-Krylov algorithm
and some sample results.
3 Present Algorithm and Results
The present algorithm originated with the work of Pueyo and Zingg [6], who
developed a Newton-Krylov algorithm applicable to the computation of aero-
dynamic flows on single-block structured meshes using scalar artificial dissi-
pation and the Baldwin-Lomax turbulence model. The algorithm has since
been extended to incorporate matrix dissipation, the Spalart-Allmaras one-
equation turbulence model, including trip terms for laminar-turbulent transi-
tion, and multi-block meshes.
Our algorithm is based on a Jacobian-free inexact-Newton strategy using
ILU-preconditioned GMRES, as described above. There are numerous param-
eters involved, such as the tolerance to which the inner iterations are solved,
the amount of fill permitted in the incomplete factorization, and a parameter
used in forming the approximate Jacobian upon which the ILU factorization
is based. These have been carefully chosen based on numerous experiments to
maximize efficiency while maintaining robustness, such that a fixed set of pa-
rameters can be used for a broad range of flow conditions, including complex
flows over multi-element geometries. Considerable effort has been put into op-
timizing the pseudo-transient phase. In the original algorithm of Pueyo and
Zingg [6], an approximate factorization algorithm was used initially. This has
been replaced by a Newton-Krylov pseudo-transient scheme with gradually
increasing spatially-varying time steps. In particular, it is critical to maintain
positivity of the Spalart-Allmaras turbulence model variable, since the model
can become unstable when negative values occur. The technique suggested
by Spalart and Allmaras [17] to maintain positivity involving M-matrices is
incompatible with Newton-like convergence. We use an alternative approach
based on a separate local time step for the turbulence model designed to main-
tain positivity during the pseudo-transient phase. Full details of the present
algorithm are deferred to a forthcoming paper and thesis [18].
Two examples are shown in Fig. 1. Convergence of the error in the lift co-
efficient3 is plotted as a function of both CPU time (top axis label) and equiv-
alent function evaluations, i.e. CPU time divided by the time required for one
right-hand side evaluation (bottom axis label). The present results are labelled
“NK” and, in order to provide a well-known reference, are contrasted with
the convergence of an approximate-factorization algorithm (labelled “AF”)
3
Defined as the difference between the current lift coefficient and the fully con-
verged lift coefficient.
CPU time - seconds CPU time - seconds

0 20 40 60 80 100 0 500 1000 1500 2000
1 1
AF AF
0.1 NK 0.1 NK
0.01 0.01
0.001 0.001
Cl error
Cl error
1e-04 1e-04
1e-05 1e-05
1e-06 1e-06
1e-07 1e-07
1e-08 1e-08
0 200 400 600 800 1000 1200 1400 0 2000 4000 6000 8000 10000
Equivalent RHS evaluations Equivalent RHS evaluations
(a) NACA 0012 airfoil, M∞ = 0.15, (b) Three-element configuration,

Re = 9 × 106 M∞ = 0.197, Re = 3.52 × 106
Fig. 1. Sample convergence histories
driving the identical right-hand side to zero. The approximate-factorization

algorithm is run as efficiently as possible, using the diagonal form and an
optimal local time step. Both solvers use the same spatial discretization and
hence produce identical converged solutions. The CPU times are obtained on
an Intel 2800 Pentium 4 desktop computer. For a subsonic turbulent flow
over the NACA 0012 airfoil, the lift coefficient is fully converged in roughly
40 seconds on a mesh with 17,385 nodes. The complex three-element airfoil
flowfield converges in roughly 8 minutes on a mesh with 71,868 nodes. Both
meshes have cells with aspect ratios greater than 10,000. Similar performance
is obtained for transonic flows. The equivalent function evaluations provide a
measure that is somewhat independent of hardware. For convergence to two
or more significant figures in lift coefficient, the Newton-Krylov algorithm is
much faster than the approximate-factorization algorithm in both cases.
4 Issues and Solutions

4.1 RCM Root Node
The RCM reordering minimizes the bandwidth of a sparse matrix. It is gen-

erally believed that this improves the accuracy of an ILU factorization by
reducing the fill required for a complete lower-upper factorization. While this
is a reasonable conjecture, it does not fully explain the observations. For ex-
ample, RCM has been shown to outperform algorithms designed specifically
to minimize fill. There is also no explanation for the sensitivity to the root
node used for the RCM ordering. Fig. 2 shows the dependence of the CPU
time required for convergence of the single-element airfoil case described in
the preceding section for various root nodes lying on the outer boundary of
the mesh. The “Root angle” is measured from the positive x-axis. Hence an
angle of zero corresponds to a root node directly downstream of the airfoil,
ninety degrees is above the airfoil, 180 degrees lies upstream, etc. A CPU time
Jacobian-Free Newton-Krylov Methods: Issues and Solutions 241
180
160
140
CPU time (sec)

120
100
80
60
40
20
0
-200 -150 -100 -50 0 50 100 150 200
Root angle in degrees
Fig. 2. CPU time plotted vs. far-field root node location
of zero indicates nonconvergence. The figure shows that root nodes upstream
of the airfoil can lead to varied performance of the algorithm, while root nodes
located above and below often lead to nonconvergence. It is clear that it is
best to select a root node directly downstream of the airfoil. This choice has
been used for the results shown previously.
4.2 Misleading Convergence of the Linear Iterations
Another problem arises if the magnitude of the residual of one equation (e.g.
the turbulence model) is much larger than that of the others. The exit toler-
ance for the linear solver can then be satisfied if the larger residuals are re-
duced, even though the other equations have not converged sufficiently, thus
preventing convergence of the Newton iterations. This problem can be di-
agnosed by examining the residual norms of each equation separately. The
equations can be suitably scaled such that the residuals are of the same order.
We also perform a minimum of five inner iterations per outer iteration.
4.3 Inaccurate Jacobian-Free Approximate Matrix-Vector

Products
The choice of the parameter in (1) is a balance between round-off and trun-
cation errors. Many authors use
√
m
= (2)
||v||2
where m is the value of machine zero. Problems arise when there exist large
differences in the magnitudes of the entries in Q but not in v or vice versa.
The former occurs when the variables are poorly scaled, the latter when the
residual equations are poorly scaled. Poor scaling of the variables can occur
for various reasons. In our case, the nondimensionalization of the turbulent
viscosity using the actual viscosity leads to values of the turbulence parameter
that greatly exceed the nondimensional mean-flow variables. It can then be
difficult or impossible to find a suitable value of that leads to sufficiently low
truncation error and round-off error. This difficulty is particularly insidious
because the linear iterations appear to converge. However, since the matrix-
vector products are not accurate, GMRES is converging to the solution of
a different linear problem. The problem can be diagnosed by recalculating
the linear residual after the GMRES iterations have terminated. Note that
GMRES does not directly evaluate the residual at each iteration, but instead
finds it indirectly, assuming that the matrix-vector products are accurate. If
the residual calculated subsequently differs from that reported by GMRES,
then the Jacobian-free matrix-vector products are inaccurate. The solution
is to rescale both the variables and the residual equations on the fly so that
both are well scaled. In addition, we use a value of that is two orders of
magnitude greater than that given by (2) to reduce round-off error.
5 Conclusions
A number of subtle problems that can arise when Newton-Krylov algorithms
are applied to the Reynolds-averaged Navier-Stoes equations have been dis-
cussed. Strategies have been presented for identifying and addressing these
issues.
References
1. Knoll, D.A., and Keyes, D.E., J. Comp. Phys., 193:357-397, 2004.
2. Venkatakrishnan, V., and Mavriplis, D.J., J. Comp. Phys., 105:83-91, 1993.
3. Degrez, G., and Issman, E., VKI Lecture Notes 1994-05, 1994.
4. Barth, T.J., and Linton, S.W., AIAA Paper 95-0221, 1995.
5. Nielsen, E.J., Anderson, W.K., Walters, R.W., and Keyes, D.E., AIAA Paper
95-1733, 1995.
6. Pueyo, A., and Zingg, D.W., AIAA J., 36:1991-1997, 1998.
7. Geuzaine, P., Lepot, I., Meers, F., and Essers, J.A., AIAA Paper 99-3341,1999.
8. Saad, Y., and Schultz, M.H., SIAM J. Sci. Stat. Comput., 7:856-869, 1986.
9. Nemec, M., and Zingg, D.W., AIAA J., 40:1146-1154, 2002.
10. Chisholm, T.T., and Zingg, D.W., AIAA Paper 2003-3708, 2003.
11. Nemec, M., Zingg, D.W., and Pulliam, T.H., AIAA J., 42:1057-1065, 2004.
12. Isono, S., and Zingg, D.W., AIAA Paper 2004-0433, 2004.
13. Smith, T.M., Hooper, R.W., Ober, C.C., Lorber, A.A., and Shadid, J.N., AIAA
Paper 2004-0743, 2004.
14. Luo, H., Baum, J.D., and Lohner, R., AIAA Paper 2004-1103, 2004.
15. Nichols, J.C., and Zingg, D.W., AIAA Paper 2005-5230, 2005.
16. Mulder, W.A., and van Leer, B., J. Comp. Phys., 59:232-246, 1985.
17. Spalart, P., and Allmaras, S., AIAA Paper 92-0439, 1992.
18. Chisholm, T.T., Ph.D. Thesis, University of Toronto Institute for Aerospace
Studies, 2006.
Non-stationary two-stage relaxation based on
the principle of aggregation multi-grid
R. Haelterman1 , J. Vierendeels2 and D. Van Heule1

1
Royal Military Academy, Department of Mathematics, Renaissancelaan 30,
B-1000 Brussels Belgium. Robby.Haelterman@rma.ac.be
2
Ghent University, Dept. of Flow, Heat and Combustion Mechanics,
St.-Pietersnieuwstraat 41, B-9000 Gent, Belgium. Jan.Vierendeels@UGent.be
We propose an automatization procedure to find an optimal value of the

parameter used in some relaxation schemes. The basis of the method is formed
by an aggregation-type procedure into a single aggregate and is meant to be
applicable as a patch for existing codes.
1 Introduction
Huge algebraic systems of equations are routinely solved by iterative meth-
ods. Some of these use a relaxation paramater (SOR, damped Jacobi,...),
the choice of which is sometimes awkward. Often these methods are used as a
smoother for a multi-grid method. In [1] it was shown that a V-cycle multi-grid
method using the Galerkin construction will converge whenever the smoother
converges, which depends on the relaxation parameter. We propose a proce-
dure to automatize the choice of the parameter and limit ourselves initially
to (semi-iterative) damped Jacobi and related methods as smoothers for a
multi-grid method. The idea is to add a small number of lines to an existing
code as a patch. Emphasis is put on simplicity and low cost and applicability
to general cases of matrices to guarantee convergence independently of the
input relaxation parameters (or at least to reduce its influence). We compare
different variations of the idea on the advection-diffusion equation.
2 The algorithm
Vectors will be written in boldface and matrices in capital letters. We are
interested in solving linear algebraic equations of the type
Au = f (1)
We assume that A ∈ IRn×n is non-singular; u, f ∈ IRn×1 . We will solve this
system with an iterative procedure (also called relaxation scheme). The latter
typically splits the matrix A in two parts: A = M − N , where M is non-
singular. The solution is found by iteratively applying

244 R. Haelterman, J. Vierendeels and D. Van Heule
M um+1 = N um + f (2)
starting from an initial guess uo . This will also be symbolically written as
um+1 = S(um ) (3)
2.1 Dynamic over-under-relaxation (DOUR)

We suppose that we have an undamped relaxation scheme given by (3). In its
damped guise, with the user-defined parameter ω, this becomes
um+1 = um + ω(S(um ) − um ) (4)
We propose the following predictor-corrector routine to dynamically modify
ω into an effective relaxation parameter ωe :
ũm = um + ω(S(um ) − um ) (5)
m+1 m 0 0 0 −1 0 m
u = ũ + P (R AP ) R (f − Aũ ) (6)
| {z }
κ
where
P 0 = ũm − um = ∆um (7)

R0 = (P 0 )T (8)
It results that
um+1 = um + ωe (S(um ) − um ) (9)

where ωe = ω(1 + κ). We can distinguish the following situations
• κ > 0: over-relaxation w.r.t. (4)
• κ > ω1 − 1: genuine over-relaxation
• κ ∈] − 1, 0]: under-relaxation w.r.t. (4)
• κ = −1 : um+1 = um
• κ < −1: the scheme ”goes the other way”, which is possible if the unmod-
ified scheme was divergent.
Equation (6) is the same formulation as would be used by a two-grid method
using aggregation into a single aggregate. Figure 1 illustrates the principle,
showing that the method can be interpreted as a step-length method [2].
2.2 Convergence of DOUR

Convergence of the method has been proven, based on the following theorems.
Theorem 1. The application of one modified relaxation step as defined by

(8) and (6) will return the value of ωe that minimizes the error em+1 in
the energy norm, for any underlying smoother S, as long as A is Hermitian
positive-definite.
Non-stationary two-stage relaxation (DOUR) 245
Fig. 1. Graphical representation of the DOUR-algorithm
Theorem 2. Suppose that ∃ ω ∗ ∈ IRo that allows (4) to be convergent in

the energy norm. In that case the DOUR relaxation scheme ((8)-(6)) will
converge in the energy norm for any relaxation scheme irrespective of the
input relaxation parameter ω 6= 0 as long as
• A is hermitian positive-definite (and thus induces a norm)
• a linearly independent set of n eigenvectors of B can be found
2.3 Cost
The dominant cost of an iterative scheme is the matrix-vector (M-V) multi-
plication. We will allocate one workunit (WU) to a M-V multiplication and
neglect the cost of the scalar products. We can quantify the cost of a damped
relaxation as 2 WU (predictor step). The corrector step (κ-modification) also
requires two M-V multiplications. The cost of the multi-grid defect correction
is more difficult to quantify, depending on the real cost of the transfer opera-
tors. As we are only interested in the robustness (convergence) of the scheme,
and all variants use the same two-grid defect correction, the exact cost is a
moot point and can be estimated as 3 WU.
3 Results
The equations under considerations are the two dimensional advection-diffusion
equations:
(a · ∇u) + ∇2 u = f (10)

With f a randomly chosen,but constant, source term. The advection vector
is taken as a = √110 , √310 . With = 0 we obtain the Poisson equation.
All the tests in this section use the LDA-version of the Residual Distribution
Method for the space discretization [3] on a 676 node triangular mesh, and
a combination of damped Jacobi relaxations as a smoother for a two-grid
method using smoothed aggregation. We denote by J(ωi ) a Jacobi relaxation
Fig. 2. Results for different MG schemes. ω1 = ω2 . = 0 (left) and = 1000 (right)
with relaxation parameter ωi and by J(ωi , κ) a DOUR Jacobi relaxation with

input parameter ω. DC stands for the defect correction on the coarse grid.
We draw attention to the fact that we compare with the baseline stencil that
contains two relaxation sweeps with possibly different relaxation parameters.
Strictly, these are also non-stationary iterative solvers, but with a periodicity
of 2, which can be shown to correspond to a two-stage Runge-Kutta style
Jacobi relaxation or a semi-iterative method.
Although for 6= 0 the discretization matrix A is not symmetric positive
definite, it might be interesting to see how the DOUR handles these cases.
We require the scheme to reduce the L2-norm error to 0.01 % of the initial
value. We compare the following schemes with stencils
• MG1: J(ω1 ) − DC − J(ω2 )
• MG2: J(ω1 ) − DC − J(ω2 , κ)
• MG3: J(ω1 , κ) − DC − J(ω2 , κ)
• MG4: a DOUR, not on a single relaxation step, but on the difference
between the input and output of the V-cycle: (J(ω1 ) − DC − J(ω2 )) (κ)
• MG5: (J(ω1 , κ) − DC − J(ω2 )) (κ)
• MG6: (J(ω1 ) − DC − J(ω2 , κ)) (κ)
• MG7: (J(ω1 , κ) − DC − J(ω2 , κ)) (κ)
where we try all possible combinations of ω1 and ω2 between 0 and 2 with a
step of 0.05. We take the following values of : 0,1,100,1000.
Without going into detail, we can summarize them as follows: The baseline
scheme (MG1) is very sensitive to the input parameters. The other schemes
(except for MG4) are stable over a wider range of parameters. Schemes MG5,
MG6 are very weakly dependent on the input, and MG7 is completely inde-
pendent. We noted that the computed values of κ vary only slowly after an
initial transient phase. As the computation of κ is responsible for an extra cost
of 25 to 100 % over the baseline scheme, we will try to gain some efficiency
by only computing the value once in a while. We also drop schemes MG2 and
MG3 because of their low performance and MG4 because it is not uncondi-
tionally convergent. When we analyze the evolution of the κ-parameters for
schemes MG5 and MG7 we see that after a small number of iterations their
value is almost constant. The transient phase appeared to be close to 4-5 it-
erations. For that reason, we tried to fix the values of κ after a small number
of iterations for MG5, MG6 and MG7. While for MG5 and MG6 we gained in
Non-stationary two-stage relaxation (DOUR) 247
Fig. 3. Results for MG7e vs MG1. ω1 = ω2 . = 0, 1, 100, 1000
cost, we lost a lot of the robustness. The modified MG7 (called MG7e), being
independent of the input parameters, gave very good results when we fixed
κ after 4 iterations, being better than the baseline, except for a very narrow
range of parameters where it was roughly equivalent. We show the results for
ω1 = ω2 in figure 3 As the scheme is applicable to any damped smoother
we also try it on damped Symmetric Gauss-Seidel instead of Jacobi. The
schemes with the most stable output were MG3,5 and 7. Again they would
need to have the computation of κ frozen after some iterations to make them
computational competitive with MG1 using the optimal combinations of the
relaxation parameters.
4 Conclusions
We used a simple aggregation procedure using a single aggregate to automatize
the choice of relaxation parameters. We applied the scheme to a damped
Jacobi and damped Gauss-Seidel iterative method as a smoother for multi-
grid, allowing for two different relaxation parameters during the iteration in
a periodic fashion, thus creating a semi-iterative relaxation. We found that
by varying the exact application of the method we could make the scheme
(almost) independent of the input parameters.
References
1. K. Stüben. A review of algebraic MG. Tech. rep. 69, GMD, Bonn, 1999
2. J.M. Ortega, W.C. Rheinboldt. Iterative Solution of Nonlinear Eq. in Several
Variables. Classics in Appl.Math. 30, SIAM, ISBN 0-89871-461-3, 2000
3. R. Haelterman. Aggregation Multi-grid for Residual Distribution Schemes. PR
2005-14, 2005, Von Karman Institute for FD
Table 1. Optimal results for the different multi-grid schemes on the Poisson prob-
lem.
MG1 MG2 MG3 MG4 MG5 MG6 MG7
minimal WU 112 90 154 108 99 99 143
optimal ω1 0.50 0.55 N/A 0.50 N/A 0.90 N/A
optimal ω2 1.35 N/A N/A 1.60 0.95 N/A N/A
Table 2. Optimal results for the different multi-grid schemes on the advection-
diffusion equation with = 1.

minimal WU 112 99 154 99 110 110 143
optimal ω1 0.55 0.55 N/A 0.50 N/A 0.65 N/A

minimal WU 245 360 605 261 330 319 390
optimal ω1 0.50 0.15 N/A 0.55 N/A 0.70 N/A

minimal WU 287 432 979 306 385 374 468
optimal ω1 0.50 0.10 N/A 0.55 N/A 0.70 N/A
The efficient and accurate solution of porous
media flow problems with strongly
discontinuous coefficients
Y.C. Lee1 and P.H. Gaskell2

1
The University of Leeds, LS2 9JT, Leeds, U.K. y.c.lee@leeds.ac.uk
2
The University of Leeds, LS2 9JT, Leeds, U.K. p.h.gaskell@leeds.ac.uk
1 Introduction
Governing equations featuring strongly discontinuous coefficients arise in
many physical and engineering applications. A case in point concerns flow
in porous media which in a general sense, is anisotropic and heterogeneous
in nature, and the use of simple homogeneous isotropic assumptions is in-
adequate [1, 2]. Indeed, accurate numerical solutions can only be achieved if
the discretisation employed encompasses its surrounding neighbouring cells to
take account of the full hydraulic conductivity tensors [3]. The sharp discon-
tinuous characteristics observed in these tensorial coefficients require careful
attention.
Accordingly, a multigrid strategy stands out as a clear candidate for for-
mulating an effective, fast and robust numerical solver [4]. We note however,
that the use of uniformly-spaced grids would be highly inefficient for porous
media flows in which discontinuities in hydraulic conductivity are often lo-
calised or preferentially stratified. Hence, this paper describes in addition the
implementation of automatic grid refinement as a means of further improv-
ing the computational efficiency of such a solver, without incurring loss in
accuracy.
2 Mathematical model and method of solution

Porous media flows can be described in terms of the general continuity equa-
tion and Darcy’s law, which for the case of incompressible steady-state con-
ditions, in the absence of body forces, reduce to:
− ∇ · K∇p = s. (1)
K denotes hydraulic conductivity, p the pressure and s a source.
Two different approaches were used to derive a discrete analogue for equa-
tion (1); a weighted averaging (WA), and a control-volume (CV) formulation.

250 Y.C. Lee and P.H. Gaskell
The multigrid scheme employed over the solution domain Ω uses a sequence
of progressively finer grids, Gk , where k = 0, 1, 2, . . . , kmax correspond to de-
creasing mesh size, ∆k = 0, 1, 2, . . . , kmax . On grid level k, ∆k is defined as
the number of mesh nodes in each co-ordinate direction, nk = 2k+c+1 , where c
is a constant defining the coarsest grid level. In addition, both cell- and patch-
wise adaptive grid refinement strategies are explored. The former consists of
locally refined arbitrary regions comprised of individually marked grid cells;
the latter is made up of coherent uniformly patched regions on a collection
of marked grid cells which intersect or overlaps one another. The τ -indicator
[4] is the natural choice for quantifying errors since this information is readily
available on each multigrid level. The criterion for grid refinement, τkk−1 > ε,
is defined as a user specified tolerance.
3 Results
The effectiveness of the adaptive multigrid solver is demonstrated for a se-

ries of benchmark problems [5, 6]. The computations were performed using
a multi-coloured Gauss-Seidel scheme and FMG V(2,2)-cycle [4] with coarse
grid constant, c = 2 (9 × 9 nodes in each direction). The convergence is based
on the l2 norm with the residual reduced by a factor 10−6 .
Discontinuous two-region anisotropy. For this problem a mixture of di-
agonal and non-diagonal hydraulic conductivity tensors is used, see Figure
1(a):

10 21
K1 = , and K2 = λ , (2)
01 12
over regions 1 and 2, respectively, with diffusivity coefficient, λ = 1, and
boundaries defined by :

(2 sin(y) + cos(y))λx + sin(y) in region 1,
u= (3)
ex sin(y) in region 2.
Results show that both patch- (Figure 1(b)) and cell-wise (Figure 1(c)) adap-
tivity identify areas requiring further grid refinement utilising grid levels,
5 ≤ k ≤ 7, with ε = 0.3τkk−1 ; notably, at both the sudden discontinuous
change in hydraulic conductivity in the centre of the domain and the steep
pressure gradient along the eastern boundary. Figure 1(d) shows the corre-
sponding iso-contour plot of the resultant pressure field.
The relative performance of the multigrid solver, is presented in Figure
2(a) for solutions obtained on a 257 × 257 mesh without adaptivity. Using
a fix number of multigrid iterations, the CV method is seen to yield better
convergence rates. However, as shown in Figure 2(b), when adaptivity is im-
plemented, the amount of CPU time required is slightly higher with the CV
approach – this is due its greater complexity. Implementation of adaptive local
Solution of porous media flow problems with discontinuous coefficients 251
0 0.5 1
1 1
1.0
0.5 Region 1 Region 2 0.5 0.5
0.0 0 0
(a) 0.0 0.5 1.0 (b) 0 0.5 1
0 0.5 1
0 0.5 1 1 1
1 1
0.5 0.5
0.5 0.5
0 0
0 0
(c) 0 0.5 1
(d) 0 0.5 1
Fig. 1. Discontinuous two region anisotropy: (a) problem definition; (b) patch-
wise adaptive grid structure; (c) cell-wise adaptive grid structure; (d) pressure iso-
contours.
mesh refinement demonstrates significant savings in CPU time and thus re-
sources. For example, when k = 8, there is as much as a factor of four savings
in CPU time to be achieved compared to the time expended non-adaptively.
Modified Stone’s problem [7]. Here we examine the effectiveness of
adaptive local mesh refinement for a problem exhibiting multiple sources
and sinks in a domain with anisotropic and heterogeneous structures. The
problem is depicted in Figure 3(a) for Neumann boundary conditions with
source and sink singularities, 1.0, 0.5, 0.6, −1.83, −0.27, located at co-ordinates
(3, 3),(3, 27),(23, 4),(14, 15),(27, 27) respectively, and hydraulic conductivities
given by:

10 1 1 100 0 00
KA = , KB = , KC = , KD = . (4)
01 1 100 0 1 00
Automatic adaptive grid refinement takes place on grid levels, 6 ≤ k ≤ 7, with
adaptive tolerance, ε = 0.1τkk−1 . Figures 3(b) and 3(c) shows that both adap-
tive schemes effectively identify regions requiring further refinement. Here,
the differences in mesh adaptation approach are more clearly depicted. No-
tably, cell-wise adaptive is particularly efficient at determining areas of rapid
change in which further refinement is required. Nonetheless, both approaches
-2 3
10 WA 10 Non-adaptive, WA
CV Non-adaptive, CV
-3 Patch-wise, WA
Computational time (s)

10
Patch-wise, CV
2
10 Cell-wise, WA
Errror norm
-4
10 Cell-wise, CV
-5
10 1
10
-6
10
-7 0
10 4 5 6 7 10 4 5 6 7
Grid level, k Grid level, k
(a) (b)
Fig. 2. Error and CPU times associated with (a) non-adaptive and (b) adaptive
solutions (CV and WA) for the problem specified in Figure 1(a).
provide, in general, good solution accuracy; Figure 3(d) shows the resultant
iso-contour pressure profile for this problem.
0 0.5 1
1 1
28
D
21 A
0.5 0.5
16
12
C B
5
0 0 0
(a) 0 5 12 14 19 32 (b) 0 0.5 1
0 0.5 1
0 0.5 1 1 1
1 1
0.5 0.5
0.5 0.5
0 0
(c) 0
0 0.5 1
0
(d) 0 0.5 1
Fig. 3. Generalised Stone’s problem: (a) problem definition; (b) patch-wise adaptive
grid structure; (c) cell-wise adaptive grid structure; (d) pressure iso-contours.
Solution of porous media flow problems with discontinuous coefficients 253
y
0 0.5 1
1 1
0.5 0.5
A
1/8
0 x 0 0
(a) 0 1 (b) 0 0.5 1
0 0.5 1 0 0.5 1
1 1 1 1
0.5 0.5 0.5 0.5
0 0 0 0
(c) 0 0.5 1
(d) 0 0.5 1
Fig. 4. Diagonally layered anisotropy: (a) problem definition; (b) cell-wise adaptive
grid structure; (c) pressure iso-contours; (d) streamline contours.
Diagonally layered anisotropy. A cross-bedded system that is infinite and

periodic, see Figure 4(a), is considered with periodic boundaries and hydraulic
conductivities is given by,

0.505 −0.495 20
KA = , KB = , (5)
−0.495 0.505 02
on the shaded and non-shaded regions; the anisotropy is specified over one
eighth of the total width of the computational domain; Ω = (1, 16) × (1, 16),
and thus c is taken as 3 here. Automatic adaptive grid refinement takes place
on grid levels, 5 ≤ k ≤ 6, with adaptive tolerance, ε = 0.01τkk−1 . The na-
ture of the problem highlights the shortcomings of patch-wise adaptivity in
that locally refined regions can only be defined in the direction of the global
axis. The cell-wise approach, see Figure 4(b), successfully adapts itself to the
sudden changes in hydraulic conductivity, thus minimising unnecessary com-
putational work. Figures 4(c) and 4(d) illustrate the resultant pressure and
streamline profiles; the large changes in pressure in the diagonal strip result
in a corresponding slowly moving flow.
The efficiency of cell-wise adaptivity, see Figure 5, highlights the advantage
of local mesh refinement. In general, both the WA and CV approaches perform
relatively well with the simplicity of the former resulting in slightly faster CPU
300
Non-adaptive, WA
250 Non-adaptive, CV
Computational time (s)

Cell-wise, WA
200 Cell-wise, CV
150
100
50
4 5 6 7 8
Grid level, k
Fig. 5. Diagonally layered anisotropy: CPU times associated with the problem
specified in Figure 4(a).
times. Once again, CPU time savings becomes more significant with adaptivity
as finer grid levels of increasingly smaller mesh size are used.
4 Conclusion
Both the WA and CV discretisation methods allow problems exhibiting non-
diagonal permeability or hydraulic conductivity to be solved accurately, while
the associated adaptive multigrid solver offers significant savings in terms of
CPU time and resources. Both patch- and cell-wise adaptivity is shown to be
an effective means of solution which, when coupled with an efficient multigrid
strategy, leads to an accurate, fast and robust solver that can be readily
extended to a range of flow applications, especially where localised features
are important.
References
1. Molenaar, J.: A simple cell-centred multigrid method for 3D interface problems.
Comput. Math. Applic., 31(9), 25–33 (1996).
2. Calvo, P., Lisbona, F.:A finite volume multigrid method for flow simulation on
stratified porous media on curvilinear co-ordinate systems. Int. J. Numer. Meth.
Fluids, 37, 375–397 (2001).
3. Edwards, M.G.: Simulation with a full-tensor coefficient velocity field recovered
from a diagonal-tensor solution. SPE J. 5(4), 387–393 (2000).
4. Brandt, A.: Multigrid Techniques: 1984 guide with applications to fluid dynam-
ics. GMD-Studie Nr. 85, Sankt Augustin, West Germany (1984).
5. Crumpton, P.I., Shaw, G.J., Ware, A.F.: Discretisation and multigrid solution
of elliptic equations with mixed derivative terms and strongly discontinuous
coefficients. J. Comp. Phys., 116, 343–358 (1995).
6. Durlofsky, L.J.: Numerical calculation of equivalent grid block permeability ten-
sors for heterogeneous porous media. Water Resources Research, 27(5), 699–708
(1991).
7. Stone, H.L.: Iterative solution of implicit approximations of multidimensional
partial differential equations. SIAM J. Numer. Anal., 5, 530–558 (1968).
Stability of Pressure-Correction Algorithms for
Low-Speed Reacting and Non-Reacting Flow
Simulations
Pieter Rauwoens, Krista Nerinckx1 , Jan Vierendeels and Bart Merci2
Ghent University – UGent, Dept. of Flow, Heat and Combustion Mechanics,

Sint-Pietersnieuwstraat 41 B-9000 Gent, Belgium. pieter.rauwoens@ugent.be
1
Fellow of the Flemish Institute for the Promotion of Scientific Research in the
Industry (IWT).
2
Postdoctoral Fellow of the Fund of Scientific Research - Flanders (Belgium)
(FWO-Vlaanderen).
1 Introduction
A promising technique for reliable and accurate simulations of turbulent non-
premixed flames is large-eddy simulation (LES). With this technique time-
accurate flow field predictions are imperative and efforts for improvement of
algorithms in terms of computational cost are of great value. One way of
reducing computing time is to solve the Navier-Stokes equations in their low-
Mach number formulation, and to apply a segregated solution method to the
set of equations. Such methods are called pressure-correction methods, since
the kinematic part of the pressure in low-Mach number flows acts as a variable
whose magnitude is determined by imposing a constraint on the velocity field
in a corrector step. In flows where the density remains constant in time and
space, these methods have been extensively used [1] and do not suffer from
inexpected stability problems. However, when these methods are applied to
variable density flows in the context of non-premixed combustion simulations,
instabilities arise.
2 Governing Equations
Since we are dealing with low-Mach number flows, the Navier-stokes equa-
tionspare rewritten by expanding every variable in power series of M̃∞ =
u∞ / p∞ /ρ∞ , and taking the asymptotic limit for M̃∞ going to zero1 . For
every variable the lowest order term remains in the equations, except for the
pressure, which has now two parts: a thermodynamic part p0 and a kinematic
1
The subscript ∞ denotes reference values

256 Pieter Rauwoens, Krista Nerinckx, Jan Vierendeels and Bart Merci
2
part p2 . The pressure is p = p0 + M̃∞ p2 and the low-Mach number equations
read [2]:
p0 = p0 (t), (1)
∂ρ ∂ρui
+ = 0, (2)
∂t ∂xi
∂ρuj ∂ρui uj ∂p2 1 ∂τij
+ =− + , (3)
∂t ∂xi ∂xj Re∞ ∂xi

∂T ∂T 1 ∂ ∂T γ − 1 dp0
ρ + ui = κ + , (4)
∂t ∂xi Re∞ P r∞ ∂xi ∂xi γ dt

∂u
with shear stress tensor τij = µ ∂xj ∂ui
+ ∂xji − 23 µ ∂u
∂xk δij . ρ denotes the lo-
k
cal density, ui is the component of the velocity vector in the i-direction, T

is the temperature, γ = cp /cv the specific heat ratio, cp and cv the specific
heats at constant pressure and volume respectively and κ the heat conduc-
tion coefficient. The nondimensional variables are Re∞ = ρ∞ u∞ L/µ∞ and
P r∞ = cp µ∞ /κ∞ with µ the dynamic viscosity coefficient. The zeroth-order
equation of state is p0 = ρT . Unless we are dealing with enclosed systems, the
thermodynamic pressure has a constant value in time and space.
3 Different Formulations
In this section we present the different algorithms. In order to keep the

overview, the equations (2) to (4) are simplified to their Euler-formulation
by taking the limit Re∞ → ∞. Furthermore we do not consider enclosures,
so p0 is a constant. Hence the equations to solve become:
∂ρ ∂ρui
+ = 0, (5)
∂t ∂xi
∂ρuj ∂ρui uj ∂p2
+ =− , (6)
∂t ∂xi ∂xj
∂T ∂T
+ ui = 0, (7)
∂t ∂xi
The prediction of the velocity is the same for all different formulations: the
pressure term in the momentum equation (6) is discretized at the old time
level. If we assume the convective terms to be treated explicitly, the predicted
velocity u∗j follows from
∗ n n n
(ρuj ) − (ρuj ) ∂ρui uj ∂p2
+ =− . (8)
∆t ∂xi ∂xj
This field is corrected to give the velocity at the new time level:
n+1 ∗ 0
(ρuj ) = (ρuj ) + (ρuj ) , (9)
Stability of Pressure-Correction Algorithms 257
0
where the correction for the momentum (ρuj ) is related to the correction for
the pressure p0 = pn+1 − pn by
0
(ρuj ) ∂p0
=− . (10)
∆t ∂xj
3.1 Frequently used Pressure-Correction Algorithms

Standard Pressure-Correction
This version can be considered as a direct extension of the pressure-correction
scheme for constant-density flow. The energy-equation (7) is used to calculate
the temperature T n+1 , from which the density follows: ρn+1 = p0 /T n+1 . The
velocity has to obey the continuity equation (5), written as
n+1
ρn+1 − ρn

∂ρui
=− . (11)
∂xi ∆t
Inserting (9) and (10) in (11) results in a Poisson equation for the pressure:
∗
∂ 2 p0 ρn+1 − ρn

∂ρui
− ∆t 2 = − . (12)
∂xi ∂xi ∆t
Constraint Pressure-Correction
For heat transfer applications a constraint, different than the conservation
of mass can be found. Indeed, by writing the temperature equation (4) in
conservative form, an equation for the thermodynamic pressure is obtained:

dp0 ∂ui (γ − 1) ∂ ∂T
+ γp0 = κ . (13)
dt ∂xi Re∞ P r∞ ∂xi ∂xi
When dealing with the Euler-equations in open atmosphere, this simplifies
into ∂u
∂xi = 0, requiring the velocity field to be solenoidal. This gives rise to a
i
Poisson-like pressure equation:

∗
1 ∂p0

∂ui ∂
− ∆t = 0, (14)
∂xi ∂xi ρn+1 ∂xi
The density field again follows from (7) and ρn+1 = p0 /T n+1 .
3.2 Improvements
Several arguments form the basis for improved schemes: (a) the scheme must
remain stable, even in regions with high density jumps; (b) mass must be con-
served; (c) a pressure Poisson-equation with constant coefficients is desirable;
(d) in statistically stationary low-speed non-premixed combustion applica-
tions under adiabatic circumstances, no energy-equation is solved when unity
Lewis number is assumed for all species since the static enthalpy can then
immediately be recovered from the mixture fraction. Therefore an algorithm
which uses a constraint solely based on the energy equation, is not a viable
approach.
The Standard Scheme Revised
The equation for temperature (7) is rewritten as an equation for the density:
∂ρ ∂ρ
+ ui = 0, (15)
∂t ∂xi
from which the density at the new time level ρn+1 immediately follows. The
mass conservation equation (11) is replaced by
†
ρn+1 − ρn

∂ρui
=− . (16)
∂xi ∆t
†
The dagger-symbol indicates a partly-implicit value given by (ρu) = ρn u∗ +
0
(ρu) . These definitions give rise to a pressure Poisson equation of the form
∂ρn u∗i ∂ 2 p0 ρn+1 − ρn
− ∆t 2 = − . (17)
∂xi ∂xi ∆t
The advantage of the latter equation over (12) is that, for a case with pure
convective transport of density in a region with constant velocity, (17) gives
a constant pressure field, which is physically correct.
The Constraint Correction Revised
The density field at the new time is calculated by using the continuity equation
(5), or the equation of temperature in the form (15). The constraint on the
velocity field follows from the combination of these two equations. Eliminating
the time derivative of the density, we get:
∂ρui ∂ρ
= ui (18)
∂xi ∂xi
If we evaluate the latter equation at time level n + 1, we can derive the
Poisson-like equation
∗
∂ 2 p0 ∂ρn+1 ∗ ∆t ∂p0

∂ρui
− ∆t 2 = ui − n+1 . (19)
∂xi ∂xi ∂xi ρ ∂xi
4 Results: Density jump

In a 1D channel a step in density with height (ρ1 − ρ0 ) is convected with
a constant velocity u. No diffusive effects are considered. As inlet boundary
conditions u = 1 and ρ = 1 are imposed. At the outlet the pressure is fixed to
a value p2 = 1. The thermodynamic pressure p0 equals an arbitrary strictly
positive value. The initial field for the density is the piecewise constant func-
tion:

 1 for i ∈ [1, i1 [
ρi = r for i ∈ [i1 , i2 ] . (20)
1 for i ∈ ]i2 , 100]

Stability of Pressure-Correction Algorithms 259
r is given a value of 20. The interval of the jump is [20, 40]. The equations are
solved using a 1D finite volume method on a collocated mesh. The pressure
term in the momentum equation is discretized centrally. First order upwinding
is used for the convective terms. All calculations are done with a CFL-number
of 0.9.
2.5 2.5
density/20 density/20
velocity velocity
2 2
1.5 1.5
1 1
0.5 0.5
0 0
−0.5 −0.5
10 15 20 25 30 35 40 45 50 10 15 20 25 30 35 40 45 50
2.5 2.5
density/20 density/20
velocity velocity
2 2
1.5 1.5
1 1
0.5 0.5
0 0
−0.5 −0.5
10 15 20 25 30 35 40 45 50 10 15 20 25 30 35 40 45 50
Fig. 1. Convection of a density jump with factor 20 in a straight channel after one
time step with schemes 1 to 4; 1: top left, 2: top right, 3: bottom left; 4: bottom
right.
Table 1. Properties of the different schemes

# stable constant coefficients mass conserving applicable in combustion
1 x x x
2 x
3 x x x
4 x x x
Results for the different schemes for the convection of a density jump after
one time step are given in Fig. 1. The properties of the schemes are summerized
in Table 1. The major requirement for the algorithm is to be stable. From the
study, it follows that only algorithm 4 renders stable results for high density
ratios, and is applicable in combustion simulations, making use of the mixture
fraction approach, under the conditions as stated in section 3.2. Furthermore
mass is conserved with this scheme. It has to be noted that an extra cost is
involved in this case, since the elliptic equation for the pressure, that follows
from the constraint on the velocity, does not have constant coefficients, and
therefore needs recalculation at each timestep. Scheme 3 does not suffer from
this shortcoming, but, although stable in the test case of a density jump, does
not guarantee stability in a more general case, e.g. when conductive effects
are taken into account (proof not given in this paper).
5 Conclusions
In this paper we presented four different pressure-correction algorithms for

variable-density low-Mach number flow simulations. The standard formula-
tions fail for high density ratios. Two improved schemes were presented:
the standard pressure-correction scheme in revised form and the constraint
pressure-correction in revised form. The former has a low computational cost
for solving the Poisson-equation but has an inherent danger of becoming un-
stable for too high density ratios. The latter is stable for any density ratio but
requires more computing time.
References
1. Chorin, A.J.: A numerical method for solving incompressible viscous flow prob-
lems. J. Comp. Phys. 2, 12–26 (1976)
2. Müller, B.. Low Mach number asymptotics of the Navier-Stokes equations. J.
Eng. Math., 34, 97–109 (1998)
A simple hybrid well-balanced method for a
2D viscous shallow water model
F. Marche
MAB, University Bordeaux 1, 351 cours de la libération, 33405 Talence, France

Fabien.Marche@math.u-bordeaux1.fr
1 Introduction
Viscous shallow water models are extensively used in the field of coastal engi-
neering. Recent studies by Brocchini et al. [2] have shown that computation of
2D bore propagation over a complex topography is a difficult task. Dissipation
in 2D bores crossing breakwaters leads to the generation of large scale vortic-
ity. Numerical investigations revealed that such study demands a high order
accuracy well-balanced method. Another difficulty concerns the treatment of
the shoreline. Flow properties change rapidly as the water depth vanishes, and
the location of the shoreline is always evolving. Various methods have been
proposed to cope with this, often leading to complex algorithms [9].
We introduce here an hybrid Finite-Volumes (FV) Finite-Differences (FD)
method for the solution of a 2D viscous shallow water model derived by
Marche in [7] :


 ∂t h + div (hu) = 0 ,

∂t (hu) + div (hu ⊗ u) + g h∇h = −g h∇d − a0 (h) u (1)


−a1 (h) h |u| u + 2µ div hD(u) + 2µ∇(h div u),

where u = t (u, v) is the depth-averaged velocity, h is the local water depth, d

is a parametrization of the topography and µ is the kinematic viscosity. a0 (h)
and a1 (h) stand respectively for the laminar and turbulent friction coefficients
[7]. The horizontal viscosity tensor D(u) is given as follows :
1
(D(u))ij = (∂j ui + ∂i uj ) for 1 ≤ i, j ≤ 2, (2)
2
where u1 = u, u2 = v, ∂1 = ∂x and ∂2 = ∂y . Note that eddy viscosity formu-
lations may be added [7].
In a first step, the shallow water equations with bed-slope source term are
solved [8] using a linearized Riemann solver which enables to preserve the

262 F. Marche
water depth positivity for interface values. This property appears to be fun-
damental for simulations involving moving shorelines. This solver is then com-
bined with the hydrostatic reconstruction [1]. This raises a new second order
accuracy well-balanced model which can handle bore propagation over strong
varying topography. In addition, an efficient algorithm is proposed to accu-
rately simulate wetting and drying phenomena.
In a second step, a FD discretization of the remaining source terms is per-
formed. This step enables to take into account dissipation processes, through
friction at the bottom and diffusion, without disturbing the well-suited prop-
erties inherited from the previous step.
A new model ’SU RF− W B’ based on this method is applied to waves propa-
gation situations. First, the ability of the well-balanced scheme to accurately
handle moving shorelines is highlighted through comparisons with analytical
solutions. Then, we propose a comparison between the new viscous formu-
lation and a classical formulation. Some realistic applications will be also
introduced during the communication.
2 Numerical method
2.1 Convective step
We consider first the following 1D hyperbolic system with source term :

U,t + F (U),x = S(U), (3)
!
hu

h g 0
U= , F (U) = 2 , S(U) = ,
hu hu2 + h −g h dx
2
and recall briefly the well-balanced approach used by Marche et al. [8]. A FV
discretization of this system is written as follows :
U∗i (t) − Uni (t) 1 n,−
+ Fi+ 1 − Fn,+1
i− 2
= Snc,i , (4)
∆t ∆x 2
with left and right numerical fluxes through the mesh interfaces :
Fn,−
i+ 1
= Fni+ 1 + Sni+ 1 − , Fn,+
i+ 1
= Fni+ 1 + Sni+ 1 + , (5)
2 2 2 2 2 2
Fni+ 1 = F (Ui+ 12 (0, Uni+ 1 − , Uni+ 1 + )). (6)

2 2 2
The interface value Ui+ 12 (0, Ui+ 12 − , Ui+ 12 + ) between cell i and i + 1 is ob-
tained from a VFRoe-ncv linearized Riemann solver. This solver relies on a
symmetrizing change of variable [5] which preserves the water depth posi-
tivity at least for interface values. The key point is that this interface value
is computed from values Ui+ 12 − and Ui+ 12 + issued from a combined limited
linear/hydrostatic reconstruction. More precisely, considering the cell i, we
compute first linear reconstructions Ui,r and Ui,l respectively at i + 12 − and
A simple hybrid well-balanced method for viscous flows 263
i − 12 +, using a minmod limiter. Values of Hi,l and Hi,r , where H = h + d,

are also reconstructed, and we deduce reconstructions of di,l and di,r . Then,
interface topography values di+ 21 ,j are defined with di+ 21 = max(di,r , di+1,l ).
The positivity preserving hydrostatic reconstruction of the water height is
thus defined as follows :
hi+ 12 − = max(0, hi,r + di,r − di+ 12 ), hi+ 12 + = max(0, hi+1,l + di+1,l − di+ 12 ),
and we deduce reconstructed values on each side of the interface :

hi+ 12 − hi+ 21 +
Ui+ 21 − = , Ui+ 21 + = . (7)
hi+ 12 − ui,r hi+ 12 + ui+1,l
Finally we introduce :
! !
0 0
Sni+ 1 −
= g 2 g , Si− 12 + = g 2 g , (8)
2 h 1 − h2 h − h2 1
2 i+ 2 − 2 i,r 2 i,l 2 i− 2 +
and a centered source term discretization is added to preserve consistency and
well-balancing [8] :
!
0
Sc,i = hi,l + hi,r . (9)
g (di,l − di,r )
2
The extension to the two-dimensional framework on Cartesian meshes is
straightforward.
2.2 Dissipative step
In a second step we take into account the dissipative processes through the
following system :

∂t h = 0,
∂t (hu) = −a0 (h) u − a1 (h) h |u| u + 2µ div hD(u) + 2µ∇(h div u).
Considering the values (h∗ , (hu)∗ ) issued from the previous step, this system
is solved relying on the following semi-implicit discretization :
hn+1 = h∗ ,


 ∗ n+1 ∗
h u − (hu)
= −a0 (h∗ ) un+1 − a1 (h∗ ) h∗ |u∗ | un+1 (10)

 ∆t ∗ ∗
∗ ∗
+2µ div h D(u ) + 2µ∇(h div u ).
Using a second order accuracy centered discretization of the viscous terms and
avoiding any division by h∗ , this step can clearly handle the occurrences of dry
areas which may arise through the convective step and does not modify the
steady states at rest. Finally, these two steps are combined in a Runge-kutta
time marching algorithm to achieve second order accuracy in time. The time
step ∆t for both steps is defined with ∆t = min(∆tc , ∆ts ) where ∆tc and ∆td
are respectively obtained from stability analysis for each step [8].
264 F. Marche
h* hs*
-0.15
-0.05
-0.15
-0.05
0.05
0.15
0.05
0.15
-0.1
-0.1
0.1
0.1
0
0
-1
2
Analytical solutions
Numerical results
-0.8
4
-0.6
6
-0.4
x*
t*
8
-0.2
0
10
0.2
12
Fig. 1. The Carrier and Greenspan solution. Comparison between numerical and
analytical results. Left : surface elevation versus the onshore coordinate for different
values of time. Right : shoreline motion during three periods
3 Validation
We propose here some validations relying on analytical solutions or compar-
isons with more classical models.
3.1 The Carrier and Greenspan solution
We compare here numerical results (solid lines) with the Carrier and Greenspan
1D analytical solutions [3] (in dots). This solution represents the motion of a
periodic wave traveling shoreward and being reflected out to sea, generating
a standing wave which is let run-up and run-down on a plane beach. Fig.1
highlights the ability of the model in the computation of oscillating shore-
lines. The analytical solution is used as a left inlet boundary condition and
we use ∆x = 0.01 and CF L = 0.7. The shoreline oscillations are accurately
computed even after a large number of period. We stress out that many other
well-balanced methods, highly validated in hydraulic situations, are unable to
provide such results [7].
3.2 The two-dimensional Thacker solution
We perform a comparison with the analytical solution of Thacker [10], de-

scribing free surface oscillations in a 2D parabolic basin. This case is the most
difficult for the model since it involves wetting and drying processes on a no
radially symmetric configuration. The shoreline is a circle in the (x, y) plane
and the motion is such that the center of the circle orbits the center of the
basin, while the surface remains planar with constant gradient at any given
instant. We use here ∆x = ∆y = 0.01 m and CF L = 0.7. To give an idea
of the 2D behavior of the solution, we show on Fig.2 the contour profiles for
the quantity h + d at t = 2T for both exact and approximate solution. Tiny
distortions can be observed but the planar aspect of the surface is preserved
A simple hybrid well-balanced method for viscous flows 265
Fig. 2. Thacker’s solution. Contour plots for h+d (a): exact solution, (b): numerical
results, at t = 2T .
even after 2 periods. These results are more accurate than those presented by
Hubbard and Dodd [6] with their adaptive refinement method.
3.3 The two-dimensional oblique dam-break problem
We study the evolution of a mound of water over a flat bottom which is re-
leased, generating a bore-like wave such that the front of the water propagates
with an inclination of 45◦ with respect to the boundaries of the domain. Flow
properties computed on the central cross-section orthogonal
to the propa-
gating front with the new viscous term 2µ div hD(u) + 2µ∇(h div u) are
compared to the results obtained with the classical term 4µ div hD(u) , as-
suming that the boundaries induced effects can be neglected for this section
and for a small time. Values of ∆x = ∆y = 0.01 and CF L = 0.7 are used.
h Q
0.5
0.6
0.7
0.8
0.9
0.32
0.34
0.36
0.38
0.42
0.44
0.4
1
-0.4
-0.4
-0.2
-0.2
a)
r
b)
0
r
0
0.2
0.2
0.4
0.4
Fig. 3. The 2D oblique dam-break problem. Comparison between numerical results

obtained with the new viscous term (solid lines) and the classical term (dashed lines)
for: a) the water depth h and b) the discharge Q = (h|u|) at t = 0.18 s.
266 F. Marche
The friction is neglected (α0 = α1 = 0) and µ is set to 0.1. The comparison

is reported on Fig.3 at t = 0.18 s. The differences in viscous effects are more
observable on the discharge profile, with a shift of magnitude which becomes
greater with respect to time. This behaviour is amplified as µ is increasing.
More realistic simulations are obviously needed and a comparison with some
experiments concerning the dam-break problem over a dry bed [4] is actually
under study.
References
1. Audusse, E., Bouchut, F., Bristeau, M.O., Klein, R., Perthame, B.: A fast and
stable well-balanced scheme with hydrostatic reconstruction for shallow water
flows. SIAM J.Sci.Comp., 25(6), 2050–2065 (2004)
2. Brocchini, M., Mancinelli, A., Soldini, L., Bernetti, R.: Structure-generated
macrovortices and their evolution in very shallow depths. Proc. Coastal Eng.,
772–783 (2002)
3. Carrier G.F., Greenspan H.P.: Water waves of finite amplitude on a sloping
beach, J. Fluid Mech., 4, 97–109 (1958)
4. Dressler R.F.: Comparison of theories and experiments for the hydraulic dam-
break wave, Int. Assoc. Sci. Hydraul., 38, 319–328, (1954)
5. Gallouet, T., Herard, J.M., Seguin, N.: Some approximate Godunov schemes
to compute shallow-water equations with topography. Computers and Fluids,
32, 479–513 (2003)
6. Hubbard M.E., Dodd N.: A 2D numerical model of wave run-up and overtop-
ping, Coastal Engineering, 47, 1–26 (2002)
7. Marche, F.: Derivation of a new two-dimensional viscous shallow water model
with varying topography, bottom friction and capillary effects. To appear in
European. J. Mech. B/Fluids.
8. Marche, F., Bonneton, P., Fabrie, P., Seguin, N.: Evaluation of well-balanced
bore-capturing schemes for 2D wetting and drying procedure. To appear in
Int. J. Num. Meth. Fluids.
9. Prasad, R.S., Svendsen, I.A.: Moving shoreline boundary condition for
nearshore models, Coastal Engineering, 49, 239–261 (2003)
10. Thacker WC.: Some exact solutions to the nonlinear shallow-water wave equa-
tions, J. Fluid Mech., 107, 499–508 (1981)
A kinetic energy-preserving P1 iso P2/P1
finite-element method for computing unsteady
incompressible flows
David Vanden-Abeele1 , Deryl Snyder2 , Yves Detandt3 and Gérard Degrez3

1
von Karman Institute, St. Genesius-Rode, Belgium
2
Brigham Young University, Provo, Utah, USA
3
Université Libre de Bruxelles, Brussels, Belgium
david.vanden.abeele@ulb.ac.be, deryl.snyder@byu.edu,
ydetandt@ulb.ac.be, gdegrez@ulb.ac.be
1 Introduction
Numerical experience with Direct Numerical and Large Eddy Simulations
(DNS, LES) of incompressible turbulenth flows consistently shows that cen-
tral schemes are to be preferred to stabilized methods [1, 2, 3, 4]. The turbu-
lence community has therefore put forth considerable effort to develop cen-
tral schemes that preserve the discrete kinetic energy (KE) in the inviscid
limit, which guarantees numerical stability. Designing KE preserving finite-
difference and finite-volume (FD,FV) schemes is not trivial, in particular for
higher orders and on irregular meshes [5, 6].
This contribution confirms earlier results of Sung et al. [7] who, to our best
knowledge, first demonstrated that skew-symmetric Galerkin FE methods nat-
urally preserve the KE. They thus provide a straightforward and systematic
method of creating KE preserving schemes of any order accuracy, even on
highly distorted unstructured meshes. Numerical evidence of KE preservation
is provided by implementing a P1 iso P2/P1 FE method. We also investi-
gate the behavior of schemes that do not exactly preserve the discrete kinetic
energy, namely (a) when non skew-symmetric formulations of the convective
terms are used and (b) when numerical stabilization is added.
2 Theory of kinetic energy preserving FE methods

We write the incompressible Navier-Stokes equations in dimensionless form:
∂u 1 2
RCont = ∇ · u = 0 RM om = + ∇p − ∇ u + C = 0, (1)
∂t Re
where C represents the nonlinear convective terms, written, respectively, in
conservative, advective, or skew-symmetric form:

268 A kinetic energy-preserving P1 iso P2/P1 finite-element method
1
Cc = ∇ · (uu) ; Ca = (u · ∇)u ; Cs = 2 (cc + ca ) . (2)
Although these three forms are analytically identical, they are not equivalent
when solved numerically.
We seek to obtain a FE approximation to the exact solution of Eqs. (1) in
compatible velocity and pressure spaces:
n o
Uh = span {Ni ; i = 1, . . . , n} ; Ph = span Ñj ; j = 1, . . . , m . (3)
Herein, Ni and Ñj represent basis functions with compact support. The
Galerkin FE discretization of (1) follows from the variational statement [8]
Z
qRCont (u) dΩ = 0 ∀q ∈ P, u ∈ U, (4)
Ω
Z
v · RM om (u, p) dΩ = 0 ∀v, u ∈ U; q, p ∈ P. (5)
Ω
This is, of course, valid when (v, q) are equal to the discrete solution fields (u,
p). For a skew-symmetric formulation of the advective terms, one can easily
verify that this leads to the following KE preservation statement:
Z
REn dΩ = −V ≤ 0, (6)
Ω
where
∂u2 /2 1 1
REn = + ∇· [uu2 /2] + ∇· [up] + ∇· [ (u · ∇)u] ; V = ∇u ⊗ ∇u.
∂t Re Re
3 Illustration via the P1 iso P2/P1 element
To illustrate the above KE preservation statements, we solve the unsteady in-

compressible Navier-Stokes equations on 2D meshes of P1 iso P2/P1 elements
(see Fig. 1). The velocity field is represented by means of usual P1 basis func-
tions, while the pressure field is represented in a piecewise-linear manner on
the larger P2 element—for instance Ñ1 = N1 + (Na + Nc )/2. Equations (1)
are discretized in time using the Crank-Nicolson method for the pressure and
viscous terms, while an Adams-Bashforth method is used for the convective
c
b
1
a
2
Fig. 1. Illustration of the P1 iso P2/P1 stable linear element pair. Pressure degrees
of freedom are only stored in vertices marked by circles.
A kinetic energy-preserving P1 iso P2/P1 finite-element method 269
terms. For comparison, stabilized formulations have also been implemented,

including schemes that use P1/P1 equal-order interpolation:
(1) P1 iso P2/P1 with skew-symmetric convective terms (baseline)
(2) P1 iso P2/P1 with advective formulation of the convective terms
(3) P1 iso P2/P1 with conservative formulation of the convective terms
(4) Baseline with higher order momentum artificial dissipation
(5) Skew-symmetric P1/P1 with Pressure-Stabilized Petrov-Galerkin (PSPG)
formulation of the continuity equation
(6) Skew-symmetric P1/P1 R and first-order
R pressure dissipation added to the
continuity equation: Ω Ni ∇· u dΩ + Ω τ ∇Ni · ∇p dΩ = 0, where τ is a
small time scale of O(h).
Full details on the stabilized formulations may be found in Ref. [9].
4 Results
4.1 Test case definition
The schemes introduced in the preceding section are now tested on a 2D

variant of the Taylor-Green flow test case suggested by McDermott [10]. For
a periodicity length of 2π in both the x and y−directions, the dimensionless
analytical expression of this flow is:
u(x, y, t) = U − cos (x − U t) sin (y − V t) exp (−2t/Re), (7)
v(x, y, t) = V + sin (x − U t) cos (y − V t) exp (−2t/Re). (8)
For all tests presented hereafter, ν = 0, U = π and V = π/10. The flow
therefore traverses the domain in the x−direction in two time units. Tests are
performed for meshes of 8 × 8 (resolved) and 4 × 4 (marginally resolved) P2
elements. Unless stated otherwise, the discrete time step is set to 0.001. In
Fig. 3, we show u−velocity isocontours computed using the baseline method
π/10 m/s
(periodic)
π m/s π m/s
(periodic) (periodic)
π/10 m/s
(periodic)
Fig. 2. Initial vorticity isocontours and test case definition.

6.28 6.28
3.14
y
3.14
y
3.14 6.28 3.14 6.28

x x
(a) (b)
6.28 6.28
3.14 3.14
y
3.14 6.28 3.14 6.28

x x
(c) (d)
Fig. 3. Computed u-velocity contours for the baseline scheme with the 8×8 resolved
mesh: (a) t = 0; (b) t = 0.1; (c) t = 1; (d) t = 10.
6.28 6.28
3.14 3.14
y
3.14 6.28 3.14 6.28

x x
(a) (b)
Fig. 4. Comparison of u−velocity profiles after 2 time units: (a) baseline, (b) base-
line with momentum dissipation.
A kinetic energy-preserving P1 iso P2/P1 finite-element method 271
at different instances in time. Due to dispersion errors and the nonlinear na-
ture of the equations, the solution is severely distorted after only one time
unit. Shortly thereafter, the solution becomes fully chaotic. Nevertheless, the
numerical method remains stable. Fig. 4 compares solutions obtained after
2 time units. The baseline scheme rapidly distributes kinetic energy over all
available wave numbers. Adding a small amount of artificial momentum dis-
sipation suffices to maintain the flow structures.
Fig. 5(a) shows that for ∆t = 0.001 (CFL = 0.013), the time integration
errors become so small that the time integration can be considered ‘exact’
for the 100 time units considered. In this case, KE is preserved up to four
significant digits after 100 time units. As shown in Fig. 5(b), the use of non
skew-symmetric nonlinear terms leads to nonlinear instabilities independent
of the time step size. Figs. 5(c–d) show that the addition of higher-order dissi-
pation, either to the momentum or continuity equations, leads to a damping of
the flow structures. The PSPG stabilization damps out even the bulk velocity
once chaos appears.
10 8
∆t = 0.01
∆t = 0.001 7
9
8 6
〈KE〉
〈KE〉
7 5
6 4 Baseline, ∆t = 0.001
Advective, ∆t = 0.001
Advective, ∆t = 0.0005
5 3
Conservative, ∆t = 0.001
Conservative, ∆t = 0.0005
4 2
0 20 40 60 80 100 0 5 10 15 20 25
Time Time
(a) (b)
6
5
5
4
4
〈KE〉
〈KE〉
3
3
2 Baseline
2 Baseline JST momentum diss.
JST momentum diss. PSPG pressure diss.
1 PSPG pressure diss. 1
Pressure Laplacian
Pressure Laplacian
0
0 20 40 60 80 100 0 20 40 60 80 100
Time Time
(c) (d)
Fig. 5. Time evolution of the mean kinetic energy: (a) baseline formulation, in-
fluence of the time step; (b) influence of the formulation of the nonlinear terms;
(c) influence of adding stabilization on the resolved mesh; (d) influence of adding
stabilization on the marginally resolved mesh.
5 Conclusions
The shown results clearly support our claim that Galerkin FE methods are
well-suited for making kinetic KE preserving numerical methods. Although
stable, KE preserving central schemes are found to generate rapidly generate
chaotic solutions. Whether this is desirable, or not, remains to be established.
A stability analysis for the considered flow would be most useful for this.
Acknowledgments. The authors acknowledge financial support from the

Higher Education and Scientific Research division of the Belgian French Com-
munity (ARC project 02/07-283, FRIA fellowship of Y. Detandt) and the In-
stitute for the Promotion of Innovation by Science and Technology in Flanders
(IWT, MuTEch project, SBO contract 040092).
References
1. S. Ghosal. An analysis of numerical errors in large-eddy simulations of turbu-
lence. Journal of Computational Physics, vol. 125, pp. 187:206 (1996).
2. A. G. Kravchenko and P. Moin. On the effect of numerical errors in large eddy
simulations of turbulent flows. Journal of Computational Physics, vol. 322, pp.
131:310 (1997).
3. N. Park and J.-Y. Yoo and H. Choi. Discretization errors in large eddy sim-
ulation: on the suitability of centered and upwind-biased compact difference
schemes. Journal of Computational Physics, vol. 198, pp. 580:616 (2004).
4. L. Georges, G. Winckelmans and Ph. Geuzaine. Improving shock-free com-
pressible RANS solvers for LES on unstructured meshes. JCAM, accepted for
publications, 2006.
5. Y. Morishini and T. Lund and O. Vasilyev and P. Moin. Fully conservative
higher-order finite-difference schemes for incompressible flow. Journal of Com-
putational Physics, vol. 143, pp. 90:124 (1998).
6. O. Vasilyev. High-order finite-difference schemes non non-uniform meshes with
good conservation properties. Journal of Computational Physics, vol. 157(2),
pp. 746:761 (2000).
7. J. Sung, H.G. Choi, J.Y. Yoo. Time-accurate computation of unsteady free
surface flows using an ALE-segregated equal-order FEM. Comput. Methods
Appl. Mech. Engrg., vol. 190, pp. 1425:1440 (2000).
8. A. Ern and J.L. Guermond. Theory and Practice of Finite Elements. Applied
Mathematical Series, Vol. 159. Springer-Verlag, New York, 2004.
9. D. Vanden Abeele and G. Degrez and D.O. Snyder. Parallel turbulent flow
computations using a hybrid spectral/finite-element method on Beowulf clus-
ters. Proceedings of the Third International Conference on Computational Fluid
Dynamics (ICCFD3), Springer Verlag, 2004.
10. R.J. McDermott. Toward one-dimensional turbulence subgrid closure for large
eddy simulation. PhD Thesis, University of Utah, 2005.
Study on the segregation algorithms of the
incompressible Navier-Stokes equations using
P1P1/P2P1 finite element formulation
Myung H. Cho,1 Hyoung G. Choi2 and Jung Y. Yoo3

1
School of Mechanical and Aerospace Engineering, College of Engineering, Seoul
National University, San 56-1, Shillim-Dong, Kwanak-Gu, Seoul 151-742,
Republic of Korea (1:cho7213@snu.ac.kr, 3:jyyoo@snu.ac.kr)
2
Department of Mechanical Engineering, Seoul National University of
Technology, 172 Gongreung-2-Dong, Nowon-Gu, Seoul 139-743, Republic of
Korea (hgchoi@snut.ac.kr)
1 Introduction
Since velocity and pressure fields are obtained at each separate step in a
segregated formulation, it produces a small and simple matrix at each step
than in an integrated formulation. And the memory and cost requirements of
computations are significantly reduced because the pressure equation for the
mass conservation of the Navier-Stokes equations is constructed only once if
the mesh is fixed. However, segregation formulation solves Poisson equation
of elliptic type so that it always needs a pressure boundary condition along
a boundary even when no physical information on pressure is given. In gen-
eral, open boundary problems sometimes have no information on both the
pressure and pressure gradient at the exit because velocities on downstream
boundary are unknown. Therefore, the problem of proper imposition of out-
flow boundary condition is not clearly stated, which is an important issue in
the segregation approach [1, 2].
On the other hand, an integrated finite element formulation in which the
governing equations are simultaneously treated has an advantage over a seg-
regated formulation in the sense that it can give a more robust convergence
behavior because all variables are implicitly combined especially when the flow
fields are complicated, as in strong swirl flows and impinging jet. Further an
integrated formulation needs a very small number of iterations to achieve con-
vergence. However, the application of an ILU type preconditioner to a global
matrix resulting from an integrated formulation is not so straightforward as
that to a small and simple matrix resulting from a segregated formulation [3].
Thus, the saddle-point-type matrix (SPTM) in the integrated formulation is
assembled and preconditioned every time step, so that it needs a large mem-

274 Myung H. Cho, Hyoung G. Choi and Jung Y. Yoo
ory and computing time.

Therefore, we newly proposed the P2P1 semi-segregation formulation. In order
to utilize the fact that the pressure equation is assembled and preconditioned
only once in the P1P1 segregation FEM, a fixed symmetric SPTM has been
obtained for the continuity constraint of the present P2P1 semi-segregation
FEM. The momentum equation in the P2P1 semi-segregation FEM will be
separated from the continuity equation so that the SPTM is assembled and
preconditioned only once during the whole computation as long as the mesh
does not change. We name it ‘semi-segregation formulation’ since the momen-
tum equation is separated from the continuity equation while the pressure
variable is coupled with the velocity variables in the symmetric SPTM. We
compare the P1P1 segregation finite element formulation and the P2P1 semi-
segregation finite element formulation in terms of the CPU time and accuracy
for the well-known benchmark problems.
2 Theoretical background
2.1 P1P1 segregation formulation
The P1P1 segregation formulation (P1P1) can be stated as follows:

ûi − uni 1 1 1
+ (uˆj uî,j + unj uni,j ) = − pn,i + n
(τ̂ij + τij ),j (1)
∆t 2 ρ 2ρ
u∗i − ûi 1
= pn,i (2)
∆t ρ
ρ ∗
pn+1
,jj = u (3)
∆t i,i
ui+1 − u∗i 1
i
= − pn+1 (4)
∆t ρ ,i
where ∆t is the time increment, u∗ and û are intermediate velocities, and
superscript n denotes the time level. The convection and diffusion terms are
integrated using the Crank-Nicolson method (second order accuracy). In the
fractional step method, the pressure gradient term is decoupled from those of
convection and diffusion. In this procedure, the intermediate velocity does not
necessarily satisfy the continuity equation. Hence, at the next step, the pres-
sure is obtained from the continuity constraint and the velocity is corrected
by the pressure. In the first fractional step, intermediate velocity ûi of Eq.(1)
is linearized using the successive substitution method. In the third step, we
obtain pressure fields from the Poisson equation derived from the continuity
equation. Therefore, it needs an extra boundary condition for the pressure. In
this study we use the following boundary condition for the pressure equation.
First, the pressure boundary condition at the exit can be indicated using the
traction free condition as follows:
Study on the segregation algorithms using P1P1/P2P1 element 275
n+1 n+1
σij = −pn+1 δij + τij =0 =⇒ pn+1 δij = τij = τ̂ij (5)
The pressure boundary condition at the wall can be indicated using the normal
velocity component equal to zero as follows:
∂pn+1

n+1 2 ∂u
n · ∇p = = ρn · ν∇ n − − u · ∇u (6)
∂n ∂t
In the last fractional step final velocity un+1
i satisfying the continuity equa-
tion is obtained. For more detailed explanation on the P1P1 fractional step
method, one can refer to [1].
2.2 P2P1 semi-segregation formulation
The P2P1 semi-segregation formulation (P2P1) used in the present study can
be written as follows:
ûi − uni 1 1 1
+ (uˆj uî,j + unj uni,j ) = − pn,i + n
(τ̂ij + τij ),j (7)
∆t 2 ρ 2ρ
u∗i − ûi 1
= pn,i (8)
∆t ρ
( un+1 − u∗ 1
i i
= − pn+1
∆t ρ ,i (9)
un+1
i,i = 0
In the first and second fractional step it solves the same equations as P1P1 seg-
regated formulation to satisfy the momentum equation. For time integration,
the Crank-Nicolson scheme, which is second-order accurate and uncondition-
ally stable, is to be used. In order to utilize the fact that a separate pressure
equation, which is assembled and preconditioned only once, is to be solved in
the P1P1 fractional step method, a fixed symmetric SPTM has been obtained
for the continuity constraint of the present P2P1 finite element formulation.
3 Numerical examples
3.1 Two-dimensional impinging jet
The first example we consider is the two-dimensional impingement heat trans-

fer problem with a jet that is laminar at the nozzle exit. Fig. 1 illustrates the
flow region for the impinging jet problem. The geometric arrangement is char-
acterized by the nozzle width B. A calculation is performed at ReB = 450 with
a 201 × 101 mesh. At the nozzle exit, either a constant or a parabolic velocity
profile is used. A no-slip condition is assumed at the impingement plate. In
the case of the P1P1, the pressure boundary condition is given as traction-free
as shown in Fig. 1. Table 1 represents the CPU time of the P2P1 and P1P1.
Nozzle
B=1
H=2
y
x u = 0, v = 0
Impingement plate
W=8
Fig. 1. The flow regions considered for the impinging jet problem
Table 1. Condition number for the flow past a circular cylinder

CPU Time
Method 1st step 2nd step 3rd step 4th step Total
P2P1 semi-segregation FEM 0.45 0.11 0.32 0.9
P1P1 segregation FEM 0.46 0.11 1.12 0.12 1.8
The results show that the P2P1 is about two times faster than the P1P1. In
particular, the third step of P2P1 is faster than the third and fourth step of
P1P1. Fig. 2 compares the eigenvalue distribution of the ILU (or AILU) pre-
conditioned matrix to that without the preconditioning in both the P2P1 and
P1P1. A more clustered eigenvalue distribution is obtained with precondition-
ing than without preconditioning. Fig. 3 compares the local Nusselt number
(Nu) obtained from the present computations at ReB = 450 and the result
of Cziesla et al. [4]. Pressure boundary condition of an elliptic type pressure
equation in the P1P1 should be given even when the physical information of
the pressure boundary condition is not provided. Therefore, we can confirm
that the pressure boundary conditions have an effect on the accuracy and
convergence of the solutions.
3.2 Flow past a sphere at Reynolds number 300
In the second example we solved the flow of a viscous fluid past a stationary
isolated sphere at the Reynolds number 300 (Re is based on inlet fluid velocity
and diameter of the sphere). ICEM-CFD package is used for unstructured
mesh generation. Total numbers of nodes and elements in the P2P1 mesh are
173584 and 128211, respectively. In order to find the same accuracy as P2P1,
P1P1 needs much more nodes and elements than P2P1. Total numbers of
nodes and elements used in the P1P1 are 493320 and 2910779, respectively.
Time increment ∆t = 0.2 is used. Time history of the pressure is shown in
Fig. 4. It is clear that a periodic flow pattern appears behind the sphere. The
corresponding Strouhal number of the present is 0.139. This also agrees well
Study on the segregation algorithms using P1P1/P2P1 element 277
1 1
Imaginary
Imaginary
0.75
0.5 0.5
0.25
0 0
-0.25
-0.5 -0.5
-0.75
-1 -1
4 8 12 16 20 24 28 32 36 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
Real Real
(a) Without preconditioning (P1P1) (b) ILU preconditioning (P1P1)
1 1
Imaginary
Imaginary
0.5 0.5
0 0
-0.5 -0.5
-1 -1
0 4 8 12 16 20 24 28 32 0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
Real Real
(c) Without preconditioning (P2P1) (b) ILU preconditioning (P2P1)
Fig. 2. Eigenvalue distribution
Fig. 3. Nu distribution for uniform and parabolic inlet velocity profiles at the nozzle
exit (ReB = 450)
with the value of 0.137 found by the numerical experiments of Johnson and
Patel [5].
4 Conclusion
The present P2P1 semi-segregation method has the ability to give a more
accurate solution than the existing P1P1 segregation method. In terms of
the pressure boundary condition, the P2P1 semi-segregation method is more
flexible to use because the P1P1 segregation method always needs a pressure
0.08
0.04
Present study, P1P1
Present study, P2P1
0
Pressure
-0.04
-0.08
-0.12
-0.16
-0.2
3100 3200 3300 3400 3500
Time
Fig. 4. Pressure history of the flow around the sphere at Re = 300
boundary condition along a boundary even when no physical information of

the pressure is given while only a reference pressure value needs to be given
in the case of the P2P1 semi-segregation method. This has been confirmed
from the simulation of the two-dimensional impinging jet problem. As for
the CPU time, P2P1 semi-segregation method is about two times faster than
P1P1 segregation method because CPU time of the third step in the P2P1
semi-segregation method is faster than the third and fourth steps of P1P1
segregation method. Comparing with the condition number of the two meth-
ods, the eigenvalues of the AILU preconditioned matrix (P2P1) are tightly
more clustered around (1,0) than ILU preconditioned matrix (P1P1), which
ensures the fast convergence of the preconditioned BCG method. Compared
to the P1P1 segregation method, the P2P1 semi-segregation method takes
advantage of both accuracy and computation time as shown in the numerical
examples.
References
1. Choi, H.G., Choi, H., Yoo, J.Y.: A fractional of the unsteady incompressible
Navier-Stokes equations using SUPG and linear equal-order element methods.
Comput. Methods in Appl. Mech. Engrg., 143, 333–348 1997
2. Ramaswamy, B. and Jue, T.C.: Some recent trends and developments in finite
element analysis for incompressible thermal flows. Int. J. Numer. Meth. Engrg.,
35, 671–707 1992
3. Nam, Y.S., Choi, H.G., Yoo, J.Y.: AILU preconditioning for the finite element
formulation of the incompressible Navier-Stokes equations. Comput. Methods in
Appl. Mech. Engrg., 191, 4323–4339 2002
4. Cziesla, T., Tandogan, E., Mitra, N.K.: Large-eddy simulation of heat transfer
from impinging slot jets. Numer. Heat. Trans. Part A, 32, 1–17 (1997)
5. Johnson, T.A. and Patel, V.C.: Flow past a sphere up to a Reynolds number of
300. J. Fluid. Mech., 378, 19–70 (1999)
A Mach-uniform algorithm: coupled versus
segregated approach
Krista Nerinckx, Jan Vierendeels, and Erik Dick
Ghent University, Department of Flow, Heat and Combustion Mechanics,

Sint-Pietersnieuwstraat 41, B-9000 Ghent, Belgium
Krista.Nerinckx@UGent.be, Jan.Vierendeels@UGent.be, Erik.Dick@UGent.be
(Research funded with a fellowship granted by the Flemish Institute for the
Promotion of Scientific and Technological Research in the Industry (IWT)).
1 Introduction
Historically, algorithms for CFD have been grouped into two classes: coupled
density based methods and segregated pressure based methods. Both types of
algorithms have been adapted to make them Mach-uniform, i.e. applicable for
all speeds (see f.e. the references in [1]). In this paper, we present the idea
that only a classification with regard to the solution technique (coupled versus
segregated) is valuable: three different types of algorithms are constructed,
each of them based on the same idea to reach Mach-uniformity, but applying
a different solution technique.
2 Semi-implicit system
Mach-uniform efficiency implies a good convergence rate for any level of the
Mach number. In [1] we explain how Mach-uniformity can be reached by treat-
ing the acoustic and diffusive (friction, heat conduction) terms implicitly. The
convective terms, on the other hand, can be treated explicitly. A finite volume
method for the two-dimensional Navier-Stokes equations therefore results in
the following semi-implicit system
h i
(m+1) [m]
X [m] (m+1)
ρi − ρi Vi /∆tI + ρf wf dsf = 0, (1)
h i
(m+1) [m] [m] (m+1)
X X
(ρv)i − (ρv)i Vi /∆tI + (ρvw)f dsf + pf dxf =
(m+1)
X
(τ .1̄y .n̄)f dsf − ρi gVi , similar for u, (2)
h i
(m+1) [m] [m] (m+1)
X
(ρE)i − (ρE)i Vi /∆tI + (ρH)f wf dsf =
(m+1) (m+1)
X X
(κ∇T.n̄)f dsf + (τ .w̄)f .n̄dsf − ρi vi gVi , (3)

280 Krista Nerinckx, Jan Vierendeels, and Erik Dick
with ρ the density, p the pressure, T the temperature, E the total energy, H
the total enthalpy, ḡ the gravitational acceleration vector, τ the viscous stress
tensor and κ the heat conduction coefficient. The summation runs over all
faces f of the control volume around node i. dsf is the length of the face f
and wf represents the projection of the velocity vector w̄ = u1̄x + v 1̄y on the
outward normal n̄ of the face f , i.e. wds = udy + vdx.
In the convective fluxes, the transported quantities u, v, H are upwinded
using a higher order method. Like in [2], we use for the mass flux and the
pressure at a face a blending of the high speed AUSM+ flux and a low-speed
central flux. A pressure dissipation term is added to the mass flux to prevent
pressure-velocity decoupling at low Mach numbers.
A multistage time stepping with k stages is used,
[m]
Q[0] = Qn ; ...; Q[m+1] = Q[0] + αm+1 ∆tQt ; ...; Qn+1 = Q[k] . (4)
n and n + 1 represent respectively the old and the new time level, while [m]
and [m + 1] represent stage levels. Q is the state vector and ∆t = tn+1 − tn
[m]
the time step. The slope Qt = ∂Q ∂t is determined as Qt = ∆Q[m] /∆tI =
(m+1) [m] I I
(Q − Q )/∆t . ∆t a chosen internal time step. The unknown state
Q(m+1) is determined from the semi-implicit system, and then inserted into the
multistage time stepping, Q[m+1] = Q[0] +αm+1 ω∆Q[m] , with ω = ∆t/∆tI . In
the following we will use a Fourier stability analysis to analyse the algorithms.
Notice that ∆tI determines the shape of the Fourier symbol, while ω acts as
a scaling factor for this curve.
3 Solution technique: coupled versus segregated
The way in which this semi-implicit system is solved for the updated values
(m + 1) can now be chosen: three different solution techniques are considered,
varying between a fully coupled and a fully segregated approach. In a fully
coupled method, the semi-implicit (4Nx4N)-system (1)-(3) is solved in a cou-
pled way, in each step of the multistage time stepping. As a second solution
technique, we consider the coupled pressure and temperature algorithm from
[1]. This algorithm finds its place in between the fully coupled and the fully
segregated approach. An explicit convective predictor step is followed by an
acoustic / diffusive corrector step in which a (2Nx2N)-system of correction
equations has to be solved. The third algorithm is a fully segregated pressure-
correction algorithm with pressure corrections determined from the energy
equation. As explained in [1], this fully segregated algorithm can be applied
for a perfect gas flow without heat transfer. Here, the dimension of the system
to be solved in each multistage step is (NxN) only.
A Mach-uniform algorithm: coupled versus segregated approach 281
4 Behavior in some typical flow problems

4.1 Inviscid flow
The invicid flow of a perfect gas is considered. First, we focus on low speed
flow. In the stability analysis we consider a flow in the x-direction aligned
with the grid, at M = 0.001. We use the Van Leer-κ scheme for the trans-
ported quantities (with κ = 1/3), and the multistage time stepping with four
stages and standard coefficients. The time step ∆tI is calculated from a chosen
convective CFL number, ∆tI = cf lI ∆x/u and the scaling factor ω is chosen.
Figure 1 shows the Fourier symbol. For the fully coupled algorithm
(COUP), cf lI = has to be decreased down to 0.5 to obtain a good shaped
curve. Next, the curve can be scaled up with ω = 2.7 before it crosses the
time stepping curve.
In the predictor-corrector approach, we consider three algorithms: a pres-
sure-correction algorithm with pressure corrections form the energy equation
(PE), one with pressure corrections from the continuity equation (classical
approach) (PC), and the coupled pressure and temperature correction algo-
rithm (PT). Explicit convective terms are used. Apparently, the PE algorithm
can use the highest global CFL number (i.e. cf lI ∗ ω). For the PT algorithm
the global CFL number is somewhat lower. For the PC algorithm the global
CFL number is the lowest. This confirms that, for the case of a perfect gas
without heat transfer, the corrections should by determined from the energy
equation.
3 3
2 2
1
1
–2.5 –2 –1.5 –1 –0.5

–2.5 –2 –1.5 –1 –0.5
–1
–1
–2
3 3
2
2
1
1
–2.5 –2 –1.5 –1 –0.5 0
–1 –2.5 –2 –1.5 –1 –0.5
–2
–1
Fig. 1. Fourier symbol, M = 0.001. Top left: fully coupled algorithm, cf lI = 0.5,
ω = 2.7. Top right: PE algorithm, cf lI = 0.6, ω = 2.7. Bottom left: PT algorithm,
cf lI = 0.6, ω = 2.4. Bottom right: PC algorithm, cf lI = 0.2, ω = 2.7.
As a numerical test case, we take the inviscid flow past a bump in a channel,
with an inlet Mach number of 10−5 . Figure 2 compares the convergence plots.
First, the maximum value for cf lI was determined to keep the computation
stable with ω = 1. Next, this maximum value for cf lI was kept constant and
ω was increased as much as possible. For the predictor-corrector algorithms,
the PE algorithm converges the fastest. The PC algorithm needs the most
time to converge, due to the use of the continuity equation to determine the
pressure (see [1]). The PT algorithm has a convergence rate in between the
PE and the PC algorithm. Indeed, due to the influence of the continuity
equation in the corrector step, the convergence slows down compared to the
PE algorithm. The PE algorithm has the same convergence behavior as the
fully coupled algorithm. However, the cost to solve one time step is lower
for the PE algorithm: only a (NxN)-system has to be solved per time step,
opposed to a (4Nx4N)-system for the fully coupled method. Therefore, the
more segregation is introduced, the cheaper the simulation can be done.
-11.5
-12
log ( |Res Mom| )
-12.5
-13
-13.5
-14
PC
-14.5 PE
-15 COUP PT
-15.5
100 200 300 400 500 600
Number of time steps
Fig. 2. Test case bump, Min = 10−5 . ρv-momentum residual as a function of the
number of time steps. COUP: cf lI = 0.7, ω = 2.2. PE: cf lI = 0.7, ω = 2.9. PT:
cf lI = 0.7, ω = 1.9. PC: cf lI = 0.5, ω = 1.8.
Similar tests were done for high Mach number flow. The same conclusions
hold.
4.2 Heat transfer
In [1] it was shown that the PE algorithm suffers from a diffusive time step
limit due to the conductive terms. For the coupled pressure and temperature
correction algorithm, however, such a limit is absent. The stability analysis
now confirms this. The cf lI number is chosen as the maximum value for the
Euler case and is kept constant. When the value of κ is increased, the PE
algorithm becomes unstable if κ is taken too high. For the PT algorithm,
however, κ can be taken infinitely high.
Thus, the PT algorithm is advantageous for those cases where the heat
conduction is important. What’s more, for such a case the behavior of the PT
algorithm improves. Indeed, in case of Euler equations, the convective CFL
A Mach-uniform algorithm: coupled versus segregated approach 283
number can be taken at most 1.1 when implicit convective terms are used.
However, if the diffusive terms dominate (i.e. a cell Peclet number P ec <
O(1)), the convective CFL restriction is weakened, and even can disappear
completely. This is confirmed by the Fourier stability analysis, where the CFL
number can be taken higher as the value of P ec decreases.
4.3 The role of gravity
The gravity terms in the y-momentum and the energy equation act as source
terms. Therefore, they may introduce an additional time step restriction. To
analyse this, we add gravity terms to the Euler equations, and perform the
Fourier stability analysis for a vertical flow at M = 0.001.
In the fully coupled algorithm, the gravity terms can be treated implicitly
without difficulties. For the Euler equations, i.e. g = 0, the maximum value of
the convective CFL number cf lv is 0.6 when a first order upwind discretisation
is used for the transported quantities. We take cf lv = 0.6 and increase the
value of g (which represents some non-dimensional number).
Figure 3 compares the eigenvalue plots for the algorithm without gravity
(g = 0) and with an extremely high value for g (g = 106 ). We conclude that
the fully coupled algorithm remains stable, no matter how high g is taken. No
additional time step restriction due to the gravity terms is introduced.
In a segregated predictor-corrector algorithm, however, it is no longer pos-
sible to treat the gravity terms in a fully implicit way. To get insight in the
effect of the gravity term, we analyse the eigenvalues of the predictor amplifi-
cation matrix. We find that the eigenvalues are completely determined by two
1 1
–2 0
–2 0
1 4
0 –2 0
–2
–2
–4
Fig. 3. Fourier symbol. Top: fully coupled algorithm with implicit gravity terms,
left: g = 0, right: g = 106 . Bottom: PT algorithm, left: g = 10−7 , cf lg = 0.071,
right: g = 10−5 , cf lg = 7.1.
nondimensional groups: the convective CFL number cf lv and a CFL number

2
associated with the gravity force, cf lg = g∆t
∆y . Next, we analyse the stability
of the full predictor-corrector scheme as a function of the latter parameter,
for the coupled pressure and temperature algorithm with implicit convective
terms. Figure 3 shows the eigenvalue plots for cf lv = 1 and different values
of g. For g = 10−7 (cf lg = 0.071) the eigenvalues are situated within the first
order time stepping curve. For g = 10−5 (cf lg = 7.1), however, a part of the
eigenvalue curve is situated in the right half plane.
Thus, for a segregated predictor-corrector algorithm, an additional time
step restriction appears. This is a consequence of the segregation, i.e. the cou-
pling between certain terms is broken up. This can be understood as follows.
Consider the gravity term −ρg in the right hand side of the y-momentum
equation. In a fully coupled algorithm, a perturbation of this term will pro-
voke a response of the pressure: an equilibrium is created between the gravity
force and the pressure gradient. In a segregated algorithm, however, the pres-
sure is frozen in the predictor step, and therefore cannot respond directly to
the gravity term. Consequently, a disturbance of the velocity field v occurs. In
the corrector step, the latter disturbance is transfered to the pressure field. In
this way, the pressure responds to the gravity term in an indirect way. In order
to prevent that the imbalance in the predictor step between pressure field and
gravity force generates an excessively perturbed velocity field, the time step
has to be chosen small enough. This explains the time step restriction due to
gravity in a segregated algorithm.
5 Conclusion
We presented a Mach-uniform algorithm, using a semi-implicit approach.

Three different solution techniques were considered to solve the semi-implicit
system. Naturally, the fully coupled algorithm can do every flow problem with-
out time step restrictions if all terms are treated implicitly. However, thinking
of the computational cost to solve a large system, this is not beneficial. Indeed,
the more segregation can be introduced, the cheaper the calculation can be
done. Therefore, the right algorithm has to be used for the right application.
References
1. K. Nerinckx, J. Vierendeels, and E. Dick. Mach-uniformity through the cou-
pled pressure and temperature correction algorithm. Journal of Computational
Physics, 206:597–623, 2005.
2. J. Vierendeels, B. Merci, and E. Dick. Blended AUSM+ method for all speeds
and all grid aspect ratios. AIAA Journal, 39(12):2278–2282, 2001.
Crank-Nicolson Scheme for Solving Low Mach
Number Unsteady Viscous Flows Using an
Implicit Preconditioned Dual Time Stepping
Technique
D. Vigneron, G. Deliége, and J.-A. Essers
University of Liège, Mechanical and Aerospace Department

Chemin des chevreuils 1, 4000 Liège, Belgium, Didier.Vigneron@ulg.ac.be,
Geoffrey.Deliege@ulg.ac.be, JA.Essers@ulg.ac.be
1 Introduction
The aim of this paper is to simulate low Mach number unsteady viscous flows
using a density-based finite volumes solver. This kind of solver is known to
encounter some difficulties to simulate low Mach number flows in which the
density is almost constant. Convergence fails or accuracy decreases when the
speed of sound is much greater than the fluid velocity. Local preconditioning
methods intend to solve this problem by altering the time derivative terms
of the Navier-Stokes equations in order to artificially modify the speed of
sound, improving the convergence and accuracy in the case of low Mach num-
ber steady flows [1], [4]. Is it still the case for unsteady flows for which a
dual time stepping approach is used ? Does the condition number depend on
the time step ? Does the artificial dissipation terms of numerical advective
flux functions should be adequately modified ? In this paper, a semi-implicit
Crank-Nicolson time integration scheme is used while the advective terms up-
winding is performed by the AUSM+up scheme designed by Liou [2]. At each
time step the solution is found by means of a fully implicit Newton-GMRes
dual time stepping (pseudo-transient) technique. An eigenvalues analysis will
answer to the previous questions for this special case. It will show that the
preconditioning technique designed for steady flows leads to an ill-conditioned
system for small CF Lc numbers1 . Therefore, a new unsteady preconditioning
technique will be introduced and the AUSM+up artificial dissipation func-
tion will also be modified to achieve robustness and accuracy for all Mach
and CF Lc numbers.
1
CF Lc = c ∆
∆t
must be close to unity or to the inverse of the Mach number M −1
when the time step is dictated by the speed of sound or by the fluid velocity (that
is when CF Lu = kuk∆∆t ) respectively.

286 Scheme for Solving Low Mach Number Unsteady Viscous Flows
2 Pseudo-transient Newton-GMRes solver

In this paper, the
semi-implicit
Crank-Nicolson scheme written below is chosen
s wl+1 − s wl
= θ Rhs wl+1 + (1 − θ) Rhs wl

Str l
(1)
∆t
where s and w are the conservative and primitive variables respectively while
Rhs contains the discretized advective and diffusive terms. In this paper, all
the computations are made with θ = 12 . At each time step, the following
pseudo-transient iterations (the pseudo-time is noted τ ) are performed to
calculate the new solution at time tl+1 knowning the solution at time tl ,.
sp wn+1 − sp (wn )
+ f wn+1 = −g wl

(2)
∆τ n
The new vector sp must be chosen to achieve a good preconditioning of
the system. The choice of this vector will be detailed later in section 3. The
fonctions f and g are the terms of equation (1) that depend on the unknowns
wl+1 and the previous time solution wl respectively. At each pseudo-time
step, the equation (2) is linearized. The jacobian matrices of sp and f are
noted P and J. The following system is then solved by means of a matrix-free
GMRes
iterations method.

1
P + J ∆wn = − f (wn ) + g wl −→ wn+1 = wn + ∆wn (3)
n n

∆τ n
A pseudo-CFL number can be defined as CF LnP = max (|λi |) ∆τ n /∆ in

which λi are the eigenvalues of the pseudo-transient system. The pseudo-time
step is increased during the newton iterations
thanks to the following rule
CF LnP = CF L0P kf (wn ) + g wl k/kf w0 + g wl k with CF L0P = 10.
As the residual diminishes, the pseudo-time derivative term becomes negligi-
ble. Therefore, after convergence, equation (1) is solved.
3 Local preconditioning
A local preconditioning technique is used to modify the condition number

CN = max (|λi |) / min (|λi |) of the fully implicit pseudo-transient system (2)
and keep it close to unity. To achieve this, the preconditioning matrix P must
be well chosen. The idea of the method is to modify the density thermodynam-
ical dependence on pressure. The vector sp is equal to the physical conservative
vector s except for the derivative ρp that is replaced by the parameter ρ0p that
must be evaluate to  improve convergence. The matrix Pcan be written as
ρ0p 0T ρT
∂sp 0
P= =  ρp u ρI ρT u  (4)
∂w wn 0 T
ρp H + ρ hp − 1 ρ u ρT H + ρ hT
Crank-Nicolson Scheme for Low Mach Number Viscous Flows 287
7
10
No Precond. 10
0
10
6 Steady Precond.
Unsteady Precond.
5
10
-1
4
10
10
CN
fc
103
-2
M = 1
102 10 M = 10
-1
M = 10 -2
M = 10 -3
101 -4
M = 10
-3
100 10
-1 0 1 2 3 4
10 10 10 10 10 10 10 0 10 1 10 2 10 3
CFL c CFL c
Fig. 1. Left : Condition number CN (CF Lc ) for different preconditioning techniques

for M = 10−3 . Right : The scaling function fc (M, CF Lc ) for the new unsteady
preconditioning technique (M = 10−3 ).
3.1 Eigenvalues analysis
The eigenvalues analysis of the pseudo-transient system (2) is made on its non
spatially discretized version. If the source term g wl is assumed to have no
effect on the dispersion equation and if the diffusive terms are supposed to be
negligible the system canbe written in its quasi-linearform as
∂w Str ∂w ∂w ∂w
P + S w + θ Ax + Ay + Az =0 (5)
∂τ ∆t ∂x ∂y ∂z
where the matrices S, Ax , Ay and Az are the jacobian matrices of time deriva-
tive and advective terms
of f . By introducing a Fourier mode w (τ, x, y, z) =
w0 exp kT x − i ω τ into the system, this last equation reduces to a disper-
sion equation for ω. Due to the finite number of mesh nodes, the wavelengths
are limited by the grid spacing. To take this filtering into account, one can set
kkk = φ/∆ where ∆ is a measure of the grid spacing, which is evaluated as the
minimal distance between the node and its neighbours, and φ ∈ [0, π]. Defin-
T
ing successively the wave speed λ = ω/kkk and uk = ukkkk , the eigenvalues are
found to be such that
λ1,2,3 = θ uk 1 − i CF L−1 and λ4,5 = λ± = θ uk 1 − i CF L−1

u u T± (6)
The CF Lu number based on the local velocity and the CF Lc number based
on the speed of sound are defined through the cell Strouhal number Str by
Str = Str ∆/ ∆t φ uk = θ CF L−1 −1
u and M Str = θ CF Lc . It can be seen
that the first three eigenvalues have a real part corresponding to the advective
velocity and an imaginary part corresponding to the damping brought by the
time derivatives term of the implicit Crank-Nicolson scheme. Nevertheless this
decoupling is no longer valid for the two other eigenvalues since the factors
T± depend on M , M∗ = c0 /c (where the artificial speed of sound is defined as
c02 = ρ hT / ρ hT ρ0p + ρT (1 − ρ hp ) ) and CF Lc .
Table 1. Parameter M∗ and function fc for different preconditioning techniques.

M∗ fc
No prec. 1 q
1
(1−M∗2 )2 M 2 +4 M∗2
Steady prec. min (1, max (M, M ))
(1+M∗2 )
√ M Re(√∗)+CF L−1 √
c Im( ∗)
Unsteady prec. min 1, max M 2 + CF L−2 , M
(1+M∗ )2
s
1 1 2 4 M∗2
T± = 1 + M∗2 ± (1 − M∗2 ) − 2 (7)
2 2 i M + CF L−1
c
This expression reduces to the results found be Weiss and Smith [5] in the
case of steady Euler flows for which CF Lc tends to infinity.
3.2 Preconditioning techniques

Following the idea of Turkel et al. [3], a preconditioning for unsteady flow is
now introduced. The factor T± can be kept close to unity for all Mach numbers
and CF Lc numbers
if
the
p following rule is applied

M∗ = min 1, max M 2 + CF L−2c , M (8)
The condition number obtained by this formula can be compared in Fig-
ure 1 to those obtained by the unpreconditioned or the steady preconditioned
schemes. The expressions taken by M∗ for the three preconditioning methods
are summarized in Table 1. A good preconditioning is obtained for all CF Lc
and Mach numbers. It is equivalent to the unpreconditioned scheme when
CF Lc is very low and to the steady preconditioning when CF Lc is high. Be-
tween these two extreme cases, the condition number is maintained to small
values.
3.3 Modification of the AUSM+up scheme

The preconditioning function fc used to properly scale the artificial dissipation
terms in the AUSM+up scheme (see Liou [2]) must be constructed for this
new unsteady preconditioning method. To do this, the real part of λ± ,√that
corresponds to the advective part, is calculated below. The notation ∗ is
used for the square root that intervenes
" in equation
√ (7). −1 √ #
1 2
1 M Re ∗ + CF Lc Im ∗
Re (λ± ) = θ uk 1 + M∗ 1 ±
2 M (1 + M∗2 )

1 1
1 + M∗2 1 ±

= θ uk fc
2 M
The corresponding preconditioning function fc is plotted for different Mach
and CF Lc numbers in Figure 1. For small CF Lc numbers it reduces to the un-
preconditioned scheme for which fc = 1. For large CF Lc numbers, it recovers
its value in the case of steady preconditioning.
Crank-Nicolson Scheme for Low Mach Number Viscous Flows 289
10 3
Non Linear Residual

2
No Prec.
10 Steady Prec
10 1 Unsteady Prec
10 0
10 -1
-2
10
-3
10
10 -4
-5
10
-6
10
0 15 30 45 60 75
Newton Iterations
Fig. 2. Laminar flow past a cylinder at Re = 100 and M0 = 10−3 . Left : Streamlines
at different times. Right : Dual time stepping convergence curves (CF Lc = 103 ).
4 Results
A time periodic laminar flow past a cylinder at Re = 100 and M = 10−3 is
solved using the three different preconditioning techniques with a time step
corresponding to CF Lc = M −1 = 103 (based on the smallest grid spacing). It
can be seen in Figure 2 that the new unsteady and the steady preconditioning
exhibit equivalent convergence rates while the unpreconditioned method does
not converge as previously stated by theory for such CF Lc numbers (see
Figure 1). The vortex shedding Strouhal number is found to be Strvs =
0.166. The drag coefficient CD has a mean value of 1.32 and oscillates with
an amplitude of 0.0085 whereas the lift coefficient CL oscillates around zero
with an amplitude of 0.315. These values are in agreement with numerical and
experimental values found in the literature.
The 3D laminar lid driven cavity flow test case at Re = 1000 and M0 =
10−3 is solved. Streamlines and convergence curves are shown in Figure 3.
Tests on convergence have been made by integrating the equations with a time
step that corresponds to CF Lc = 102 . As predicted by theory (see Figure 1),
the unsteady preconditioning has a better convergence rate. Furthermore, less
GMRes iterations are performed to solve the linear systems when CF LnP tends
to infinity that is when the effect of the preconditioning matrix P becomes
negligible. That illustrates the role of the scaling function fc on convergence.
Conclusion
A new unsteady preconditioning technique, taking the CF Lc number into ac-
count, has been designed for low Mach unsteady flows solved by a fully implicit
dual time stepping scheme. The scaling function of the artificial dissipation
0.4 0.4
0.2 0.2
0 0
y
y
-0.2 -0.2
-0.4 -0.4
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4

x x
10 6 60
Non Linear Residual
No Prec.
GMRes Iterations
10 4 Steady Prec.
50
Unsteady Prec.
10 2
40
0
10
10 -2 30
10 -4 20
-6
10 No Prec.
10 Steady Prec.
-8
10 Unsteady Prec.
0
0 5 10 15 20 25 0 5 10
Newton Iterations Newton Iterations
Fig. 3. 3D lid driven cavity flow (M0 = 10−3 and Re = 1000). Up left : streamlines
at t = 2. Up Right : streamlines at t = 4. Down Left : Dual time stepping convergence
curves (CF Lc = 103 ). Down Right : GMRes convergence (CF Lc = 103 ) curves.
terms in the AUSM+up scheme has been consequently modified. Simple test
cases have illustrated the improvement of convergence predicted by theory. A
better study of time and space accuracy should be performed in future works.
References
1. Y.-H. Choi and C. L. Merkle. The application of preconditioning in viscous
flows. Journal of Computational Physics, 105:207–223, 1993.
2. M.-S. Liou. A further development of the ausm+ scheme towards robust and
accurate solutions for all speeds. June 2003. AIAA Paper 2003-4116.
3. E. Turkel and Verr N. Vasta. Local preconditioners for steady and unsteady flow
applications. ESAIM: Mathematical Modelling and Numerical Analysis, 39:515–
535, 2005.
4. D. Vigneron, J.-M. Vaassen, and J.-A. Essers. An implicit finite volume method
for the solution of 3d low mach number viscous flows using a local preconditioning
technique. In Acoment 2005 Conference, Ghent, Belgium., May-June 2005.
5. J. M. Weiss and W. A. Smith. Preconditioning applied to variable and constant
density flows. AIAA Journal, 33(11):2050–2057, November 1995.
Heated Wake by Deferred Corrected ULTRA
Salem Bouhairie1 and Vincent H. Chu2

1
Department of Civil Engineering, McGill University, Montreal H3A 2K6, Canada
salem.bouhairie@elf.mcgill.ca
2
vincent.chu@mcgill.ca
Summary. The instability problem associated with DWF formulation (an implicit
implementation of ULTRA) is removed by a deferred correction procedure. With
this deferred correction, ULTRA becomes one of the most accurate flux-limiting
strategies for unsteady flow simulations. Unlike some TVD schemes, ULTRA exerts
control only occasionally when the DWF value exceeds its bounds. This deferred
corrected DWF strategy is applied to a series of numerical simulations conducted
for the advection of heat in the wake of a heated cylinder for two Reynolds numbers,
Re = 200 and 1000.
1 Introduction
A large number of flux limiters have been developed to control unphysical
oscillations associated with advection schemes ( Van Leer [5]; Gaskell and
Lau [3]; Leonard and Mokhtari [6]; Leonard and Drummond [4]). These lim-
iters may alter the accuracy of the base schemes. Therefore, their application
should be minimized to only those instances when intervention by the lim-
iters is absolutely necessary. In this paper, computations were conducted for
a realistic flow of the wake behind a heated circular cylinder. The goal was
to study the performance of a number of flux limiters on a practical problem.
Unlike some TVD (Total-Variation Diminution) schemes, the method of UL-
TRA (Universal Limiter for Tight Resolution and Accuracy) by Leonard and
Mokhtari [7] is most accurate due to its ability to optimize for the highest
order of accuracy during the strategy’s intervention. In an implicit implemen-
tation, Leonard and Mokhtari [7] introduced the DWF formulation. However,
the original formulation is numerically unstable. In the present investigation,
this numerical instability is removed by a deferred correction procedure to en-
sure diagonal dominance of the solution matrix. Calculations were carried out
using a 2D (two-dimensional) model without sub-grid viscosity3 , for Reynolds
3
The effect of the sub-grid viscosity was considered and reported in the Ph.D.
thesis by Bouhairie [2].

292 Salem Bouhairie and Vincent H. Chu
FOD n
1
a b
0.8
K
IC
QU
0.6 f g CD h
0.4
m
DWF 0.2
0 c d
S OU FOU
-0.2 e
-0.4
-0.5 0 0.5 1 1.5 2

φ~c
Fig. 1. DWF values as a function of the normalized variable φ̃C for various finite-
difference schemes. The constraints imposed by ULTRA are along the line e-c for
the SOU (second order upwind) scheme, between the line a-b for the FOD (first
order downwind) scheme and the line c-d for the FOU (first order upwind) scheme,
and along the line g-h for the CD (central difference) scheme.
number, Re = 200 and 1000. The advective process at these Reynolds num-
bers is sufficiently dominant to provide a rigorous assessment of the strategy’s
performance.
2 Downwind weighting factor
In the implicit implementation of ULTRA, DWF (Downwind Weighting Fac-

tor) is an auxiliary variable that defines the region of constraints in the DWF
diagram as shown in Figure 1. It is also the interpolation parameter for the
face value φf in a finite-volume formulation between the central node φC and
the downwind node φD :
φf = φC + DWF (φD − φC ) (1)
Leonard and Mokhtari [7] introduced the normalized variable,
φ − φU
φ̃ = , (2)
φD − φU
and expressed the DWF as a function of the normalized variable as follows:
φf − φC φ̃f − φ̃C
DWF = = (3)
φD − φC 1 − φ̃C
Heated Wake by Deferred Corrected ULTRA 293
According to this functional relation, the DWF is constant, that is DWF =

0, 1 and 0.5, for FOU (first-order upwind), FOD (first-order downwind), and
CD (central difference) schemes, respectively. For a higher order of approxima-
tion, DWF is not a constant but is a function of the normalized variable φ̃C .
For QUICK (Quadratic Upstream Interpolation for Convective Kinematics),
this function is represented by the curve m-n, while for SOU (second-order
upwind), the function is the curve e-c in the DWF diagram. The DWFs on
the faces of the finite volumes are initially computed by a base scheme such
as CD or QUICK. Modification to the DWF values is made if and only if any
of them exceed the constraints defined by the line e-c, the area a-b-c-d, and
the line g-h, as shown in Figure 1. The advantage of ULTRA is its ability to
maintain an order of accuracy close to the base scheme (e.g. CD or QUICK).
(a) CD (b) ULTRA-CD
Fig. 2. Heat wake at Re = 1000 obtained using CD (central difference) scheme (a)
without and (b) with the flux limiter ULTRA. The negative temperature associated
with the spurious oscillations (blue in part (a) of the figure) is as high as -6.58o C.
Table 1. Cross sectional maximum and minimum temperature by various discreti-

sation schemes and flux-limiting strategies at r = 20a downstream from the cylinder;
Re = 200
Advection Scheme Cr Tmax Tmin Tmax − Tmin P d

CD (Central Difference) 0.48 7.46 -5.71 13.17 +106 %
QUICK 0.48 7.65 -0.36 8.01 +13 %
ULTRA-CD 0.48 6.15 0.00 6.15 -4 %
ULTRA-QUICK 0.48 6.38 0.00 6.38 0%
TVD-CD 0.48 5.52 0.00 5.52 -13 %
TVD-QUICK 0.48 5.31 0.00 5.31 -17 %
PL (Power Law) 0.48 4.36 0.00 4.36 -32 %
20 20
Largest Tma x over time = 14.94 Largest Tma x over time = 10.36
(a) CD Smallest Tm in over time = -6.575 (b) QUICK Smallest Tmin over time = -0.728
15
14.94 C o
Average Tma x over time = 6.63294 15 Average Tma x over time = 3.75699
Average Tm in over time = -2.04733 o Tmin over time = -0.190863
Average
10.36 C
10 10
T 5 T 5
0 0
-0.73oC
-5 -5
-6.58oC
20 20
60 70 80 90 100 60 70 80 90 100
Largest Tm ax over time = 7.79 Largest Tma x over time = 8.874
time
Smallest Tmin over time = -0.0001092 time
Smallest T over time = -0.0001337
15 (c) ULTRA-CD Average Tma x over time = 3.15418 15 (d) ULTRA-QUICK min
Average T over time = 3.43111
ma x
Average Tm in over time = 0 Average Tm in over time = 0
10
7.79oC 10 8.87oC
T 5 T 5
0 0
-5 -5
20 20
60 70 80 90 100 60 70 80 90 100
Largest Tma x over time = 6.422 Largest Tma x over time = 6.257
time
Smallest Tmin over time = 0 Smallest Tmtime
Dimensionless in over time time
=0
15 Dimensionless time
Average Tma x over time = 2.84479 15 Average Tma x over time = 2.77966
(e) TVD-CD (f) TVD-QUICK
Fig. 3. Temperature maximum and minimum at a cross section 20 radii downstream
of the heated cylinder.
Table 2. Cross sectional maximum and minimum temperature by various discreti-

sation schemes and flux-limiting strategies at r = 20a downstream from the cylinder;
Re = 1000
Advection Scheme Cr Tmax Tmin Tmax − Tmin P d

CD (Central Difference) 0.48 14.94 -6.58 21.52 +143 %
QUICK 0.48 10.36 -0.73 11.09 +25 %
ULTRA-CD 0.48 7.79 0.00 7.79 -12 %
ULTRA-QUICK 0.48 8.87 0.00 8.87 0%
van Leer-CD 0.48 6.42 0.00 6.42 -28 %
van Leer-QUICK 0.48 6.26 0.00 6.26 -29 %
Minmod[8] 0.48 4.29 0.00 4.29 -52 %
SMART[3] 0.48 8.27 0.00 8.27 -7 %
PL (Power Law) 0.48 1.73 1.44 0.29 -97 %
3 Results
The computations were conducted using a refined grid. To ensure uniform

numerical accuracy for all Reynolds numbers, the dimensions of the compu-
tational mesh are selected in proportion to the boundary-layer thickness[1].
Figure 2 compares the results obtained using the CD with and without the
flux limiter, deferred corrected ULTRA. The limiter is particularly effective in
highly unstable base schemes, such as the CD scheme. Figure 3 shows the tem-
perature maxima and minima obtained by two base schemes, CD and QUICK,
and their corresponding results obtained with the flux limiter. The difference
Heated Wake by Deferred Corrected ULTRA 295
(a) ULTRA-CD (b) ULTRA-CD
f CD g CD h
c d
SOU FOU
e
(c) ULTRA-QUICK (d) ULTRA-QUICK

FOD
n b
IC K g CD h
QU
m
c d
SOU FOU
e
Fig. 4. DWF values and the areas where the order of accuracy are reduced in the
wake. Points in (a) and (c) mark the intervention by the flux limiter ULTRA. White
and pink in (b) and (d) denote the areas where the order of accuracy is first and
second order, respectively. Green denotes the default order not altered by the flux
limiter.
between the temperature maximum, Tmax , and the temperature minimum,

Tmin , is an indicator of the scheme’s ability to resolve sharp temperature
gradients in the wake. If ULTRA-QUICK were taken as the benchmark, the
percentage of the deviation from this benchmark would be
(Tmax − Tmin ) − (Tmax − Tmin )ULTRA−QUICK
Pd = × 100. (4)
(Tmax − Tmin )ULTRA−QUICK
This percentage of deviation was evaluated for the performance of a number
of flux limiters, and are given in Tables 1 and 2 for two Reynolds numbers: Re
= 200 and 1000, respectively. At Re = 200, the deviation from the benchmark
obtained by the CD scheme is an excess of P d ' 106%, while the deviation by
the PL (Power Law) scheme is the negative P d ' - 32%, due to the scheme’s
inherent damping. At the higher Reynolds number, Re = 1000, the deviation
by the CD scheme increases to P d ' 143% (the damping by the PL scheme
increases very significantly to P d ' - 97%). The oscillations associated with
the CD and QUICK schemes, and the artificial damping by the PL scheme,
get worse with increasing Reynolds number.
The results of several TVD schemes, with fixed DWF constaints, are in-
cluded in the table for comparison. The performance of the SMART by Gaskell
and Lau [3] is very good with only a 7% deviation from the benchmark.
4 Conclusion
Stability in the solution process is achieved in the present calculations of the
unsteady flow by deferred correction and by application of a flux limiter to
suppress oscillations. This combination ensures the robustness of the algo-
rithm, which is essential in the numerical simulations of advection-dominant
processes at high Reynolds number. Numerical oscillations in highly unsta-
ble base schemes, such as the CD scheme, is managed by the flux limiter to
produce acceptable performance. The advantage of ULTRA and SMART is
the minimal intervention on their base schemes, unlike the TVDs examined in
this paper. As shown in Figure 4, the order of the accuracy of the base scheme
is compromised only sparingly in small and isolated areas in the wakes.
References
1. S. Bouhairie and V. H. Chu: Two-dimensional simulation of unsteady heat
transfer from a circular cylinder in cross flow. J. Fluid Mech., (in press), 2006.
2. S. Bouhairie: Computational methods for calculating heat transfer from a circu-
lar cylinder in a cross flow. PhD thesis, McGill University, Montreal, Quebec,
Canada, 2005.
3. P. H. Gaskell and A. K. C. Lau: Curvature compensated convective transport:
SMART, a new boundedness preserving transport algorithm. Int. J. Num. Meth.
Fluids, 8:617–641, 1988.
4. B. P. Leonard and J. E. Drummond. Why you should not use ‘HYBRID’,
‘POWER-LAW’ or related exponential schemes for convective modelling - there
are much better alternatives. Int. J. Num. Meth. Fluids, 20:421–442, 1995.
5. B. Van Leer. Towards the ultimate conservative difference scheme. II. mono-
tonicity and conservation combined in a second-order scheme. J. Comp. Phys.,
14:361–370, 1974.
6. B. P. Leonard and S. Mokhtari. Beyond first-order upwinding: the ULTRA-
SHARP alternative for non-oscillatory steady-state simulation of convection. Int.
J. Num. Meth. Engng., 30:729–766, 1990.
7. B. P. Leonard and S. Mokhtari. ULTRA-SHARP nonoscillatory convection
schemes for high-speed steady multidimensional flow. Technical Report NASA
TM 102568 (ICOMP-90-12), NASA-Lewis Research Center, 1990.
8. P. L. Roe. Large scale computations in fluid mechanics, pt. 2. In S. Osher
E. Engquist and R. J. C. Sommerville, editors, Lectures in Applied Mathematics,
volume 22, pages 163–193. Amer. Mathematical Soc., Providence, RI, 1985.
Convergence Acceleration for Euler Equation
based on SPR
Dohyung Lee1 , Hyungmin Kang2 and Dongho Lee2

1
Department of Mechanical Engineering, Hanyang University, Ansan,
KyungKi-Do, Korea dohyung@hanyang.ac.kr
2
School of Mechanical and Aerospace Engineering, Seoul National University,
Seoul, Korea kangm@snu.ac.kr and donghlee@snu.ac.kr
1 Introduction
In Computational Fluid Dynamics (CFD), a variety of high order schemes

have been extensively applied to an integrated design process, however its
capacity still fall short of the full scale integrated design. Full scale integrated
design requires a large number of grid points in spatial domain, entailing sub-
stantial increase of computing time. However, it is a waste of computational
resources to use fine grid in the whole domain. A dense grid is needed only
in a rapidly changing region such as shock waves, boundary layers, etc. For
that reason, several types of adaptive methods have been implemented to im-
prove efficiency of analysis tool. In these days, numerical techniques based on
wavelets methods have emerged as one of revolutionary technologies.
Wavelet-based numerical methods can be grouped as adaptive wavelet
Galerkin methods ([1], [2]) or adaptive wavelet collocation methods ([3], [4]).
AWGM is investigated by Bacry, Mallat and Papanicolaou [1] and it is also
considered by Holmström and Walden [2]. They mentioned that nonlinear op-
erators such as multiplication are too computationally expensive in AWGM.
On the other hand, dealing nonlinearity is uncomplicated in AWCM because
all computations are performed in physical space. In AWCM researches, Holm-
ström [3] represents the algorithm of organizing sparse point representation
(SPR) with interpolating wavelet transformation. Sjögreen [4] uses multires-
olution scheme to solve compressible Euler equations.
Based on SPR method, we develop new threshold method to maintain
the spatial accuracy of conventional solver in the present research. Also, it
includes the stabilization process in constructing a SPR dataset in order to
enhance compression ratio of SPR method. In time integration step, the tiny
variations of the flow variables at excluded points are restricted by multiplying
the weighting factor. Numerical examples of Euler equations are presented to
verify the enhancement of efficiency and accuracy of the developed method.

298 Dohyung Lee, Hyungmin Kang and Dongho Lee
2 SPR method for Euler equations

The generalized coordinate transformation of two dimensional Euler equation
yields following Eq. (1).
∂Q ∂E ∂F n
= −[ + ] = −Ri,j , (1)
∂τ ∂ξ ∂η
Q 1 1
where Q = , E = [ξt Q + ξx E + ξy F ], F = [ηt Q + ηx E + ηy F ].
J J J
All properties and governing equations are non-dimensionalized. In order to
solve this system of equations by the modified SPR method, the following
processes are executed.
At first, we organize an SPR dataset by interpolating wavelet decomposi-
tion and threshold for given conservative variables, Qni,j . The set of dyadic grid
in two-dimensions can be presented as in Fig. 1. Positions of symbol are
even numbered grids and the other positions are odd numbered grid points.
By using the 4th order of interpolating polynomial, we can approximate values
at odd numbered points from values of even numbered grids as Eq. (2).
Qe ni+1,j = − 1 Qni−2,j + 9 Qni,j + 9 Qni+2,j − 1 Qni+4,j ,

16 16 16 16
1 9 9 1
e ni,j+1 = − Qni,j−2 + Qni,j + Qni,j+2 − Qni,j+4 ,
Q (2)
16 16 16 16
e ni+1,j+1 = 0.5 × (− 1 Qni−2,j−2 + 9 Qni,j + 9 Qni+2,j+2 − 1 Qni+4,j+4 )
Q
16 16 16 16
1 n 9 n 9 n 1
+0.5 × (− Qi−2,j+4 + Qi,j+2 + Qi+2,j − Qni+4,j−2 .
16 16 16 16
Wavelet coefficients of each odd numbered positions are defined as the differ-
ence between real values and approximated values.
In threshold process, the new threshold method which consists of the
Fig. 1. A dyadic grid system for two-dimensional case
modification of threshold value and the stabilization of SPR dataset is ap-

Convergence Acceleration for Euler Equation based on SPR 299
plied to maintain the spatial accuracy of conventional method and to improve

computing efficiency. Firstly, the modified threshold values is written as
0 = min((∆x)2 , ). (3)
From Eq. (3), we threshold the dataset if a wavelet coefficient is smaller than
0 and a SPR dataset is constructed. Also, we prevent the SPR data points
from switching by additional errors when organizing SPR dataset and real-
ize the stable construction of SPR dataset which lead to the enhancement of
compression ratio.
Based on a SPR dataset, flux evaluations are only performed at included
points in a SPR dataset by conventional spatial discretization schemes and
residual values are calculated at those points. At excluded points in a SPR
dataset, residual values are reconstructed from values at even numbered po-
sitions in Fig. 1 by using the same interpolating polynomial in Eq. (2)
After constructing residual distributions in the whole computing domain,
time integration is performed. If variations of flow variables, dQi,j are smaller
than the order of threshold value at excluded points in a SPR dataset, these
values can be restricted with maintaining the solutions’ accuracy and improv-
ing the convergence. The restriction is performed as Eq. (4).
e i,j × min( |dQ

e i,j |
dQ
e i,j = dQ
0.1 , 1) (4)
α∆t min( ∆x , 1)
The whole processes are executed iteratively until the given criterion is satis-
fied.
3 Numerical Test and Discussion
In order to investigate the performance of newly modified SPR method, some

numerical results are presented. To demonstrate advantages of introduced
methods, a shock-sine wave interaction problem for one-dimensional Euler
equations and a steady-state flow problem around NACA0012 airfoil for two-
dimensional Euler equations are solved according to the level of SPR method.
Firstly, we test the modified threshold value through a shock-sine wave
interaction problem. The initial conditions for shock-sine wave interaction
problem are given as Eq. (5).
(ρ, U, p) = (3.8571, 2.6294, 10.333) where 0 ≤ x ≤ 1,
(ρ, U, p) = (1 + 0.2sin(5(x − 5)), 0., 1.) where 1 ≤ x ≤ 10. (5)
We use AUSMPW+ method with 2nd order of MUSCL scheme for spatial
discretization and 4th order Runge-Kutta explicit method for time integra-
tion. In order to confirm spatial accuracy of the modified threshold value,
grid convergence tests are performed and results are shown in Fig. 2. In case
of ∆x = 0.01 and ∆x = 0.05, ∆x2 is larger than (= 10−5 ). On the other
0.004
0.0035
Conventional
0.003
Previous threshold
0.0025 Modified threshold
0.002
0.0015
Accuracy
0.001
0.0005
0.01 0.008 0.006 0.004 0.002

Grid Spacing
Fig. 2. Comparison of solution accuracy according to grid spacing
cases, ∆x2 is smaller than . From Fig. 2, the modified threshold value main-
tains the 2nd order of spatial accuracy by changing the threshold value from
to ∆x2 . Secondly, we test the stabilization technique in constructing SPR
dataset throughout NACA0012 airfoil calculations with 201 × 97 grid points.
The 2nd level SPR method with = 10−5 is applied. The angle of attack is
5◦ and the MAch number is 0.8. Remaining grid points in SPR dataset are
shown in Fig. 3. If the stabilization technique is not used, unnecessary grid
points are remaining in a SPR dataset. The remaining grid points in a SPR
dataset without the stabilization technique are about 60% of total number
of grid points. However, when the stabilization technique is used, grid points
are extracted effectively according to the flow solution and the remaining grid
points in a SPR dataset 37% of total number of grid points. Consequently,
the compression ratio of SPR method is improved about 170%.
Thirdly, we test the restriction technique of meaningless variations of
(a) Stabilization technique-Off (b) Stabilization-On
Fig. 3. Test of stabilization technique
flow variables. The residual histories are presented in Fig. 4 for subsonic and
transonic cases. In the subsonic case (a), the characteristic of convergence is
Convergence Acceleration for Euler Equation based on SPR 301
10-1
-2 Conventional Conventional
10 Restriction-Off Restriction-Off
Restriction-On Restriction-On
10-2
Residual History
Residual History
10-3
-3
10
-4
10 -4
10
-5
10 10-5
50 100 150 200 50 100 150 200
Time Time
(a)Subsonic flow (b) Transonic flow
Fig. 4. Residual histories with/without restriction technique
substantially enhanced and the total computing time decreases by half when
tiny variations are restricted. In the transonic case (b), even though comput-
ing time reduces substantially, the improvement of computing efficiency does
not reach the level of that of subsonic case. It is because the large amount
of error is generated inevitably nearby shock discontinuity, flows downstream
and contaminates a substantial portion of computational domain, which is be-
yond the order affected by restriction technique. If some way of treating these
errors induced by shock discontinuity is introduced additionally, the compu-
tational efficiency may be improved more even in transonic case.
Combining all of newly introduced methods, the modified thresholds
value, the stabilization technique and restriction technique, we check the over-
all enhancement in computing efficiency with AUSMPW+ scheme for spatial
discretization and 4th order Runge-Kutta explicit methods for time integra-
tion. The total iteration number, CPU time gain and difference of lift co-
efficient according to each case are summarized in table 1. From table 1,
Table 1. Test cases and results for the flow at NACA0012 airfoil
NACA0012 Time SPR Iteration CPU Time Lift Difference
Airfoil Integration Level Number Time Ratio Coefficient (%)
Dense Grid Conventional 0.5497
Subsonic Conventional 14912 1952.22 1.00 0.5337 2.911
(Ma=0.3, Runge-Kutta Level1 7853 793.72 2.46 0.5327 3.093
AOA=5) (201×97) Level2 7227 569.42 3.43 0.5349 2.692
Level3 8018 652.58 2.99 0.5318 3.256
Dense Grid Conventional 1.0629
Subsonic Conventional 12013 1532.98 1.00 1.0497 1.242
(Ma=0.8, Runge-Kutta Level1 8054 893.58 1.72 1.0491 1.298
AOA=5) (201×97) Level2 7737 764.08 2.01 1.0453 1.656
Level3 7861 786.34 1.95 1.0465 1.543
differences of lift coefficients have little relation with the number of multires-
olution level and change within 2nd order of accuracy. Comparing with time
ratio of 4th order explicit Runge-Kutta method, the overall computing time
substantially decreases and about 300% enhancement of efficiency is achieved
for subsonic flow case and 190% is for transonic case.
The modified SPR method is developed to improve computing efficiency of
conventional Euler solver while conserving the spatial accuracy of the solver.
The threshold value is modified so as to consider grid spacing and as a result,
it is realized that SPR method maintains the spatial accuracy of conventional
solver. Also, by the stabilization technique, unnecessary grid points are ef-
fectively cut off in a SPR dataset and the compression ratio is enhanced. In
time integration, if variations of flow variables are smaller than the threshold
value, ´, they are restricted by adopting a weighting factor and convergence
of conventional solver is improved.
The modified SPR method can be easily generalized to three-dimensional
Euler equation and more complicated Navier-Stokes equation. Also, it is com-
pletely independent of other convergence acceleration methods such as the
multigrid method. Therefore the modified SPR method combined with multi-
grid method can elevate the computing efficiency even more.
Acknowledgement. Authors express their special gratitude to professor Kyu-Hong

Kim for his aids for this research. Also, the authors gratefully acknowledge that
this research was supported by the Brain Korea-21 program for the Mechanical and
Aerospace Engineering Research at Seoul National University and by the Center of
Innovative Design Optimization Technology, Korea Science and Engineering Foun-
dation.
References
1. Bacry, E., Mallat, S., Papanicolaou, G.: A wavelet based space-time adaptive
numerical method for partial differential equation. Math. Model. Numer. Anal.,
26, 793–834 (1992)
2. Holmström, M., Walden, J.: Adaptive wavelet method for hyperbolic PDEs. J.
Sci. Comput., 13, 19–49 (1998)
3. Holmström, M.: Solving hyperbolic PDEs using interpolation wavelets. SIAM J.
Sci., Comput., 21, 405–420 (1999)
4. Sjögreen, B.: Numerical experiments with the multiresolution scheme for the
compressible Euler equations. J. Comput. Phys., 117, 251–261 (1995)
Acceleration of Unsteady Incompressible Flow
Calculation Using Extrapolation Methods
Kenjiro Shimano1 , Shun Yonezu2 , and Yoshiteru Enomoto3

1
Musashi Institute of Technology kshimano@sc.musashi-tech.ac.jp
2
Musashi Institute of Technology g0312107@sc.musashi-tech.ac.jp
3
Musashi Institute of Technology yenomoto@sc.musashi-tech.ac.jp
1 Introduction
Acceleration techniques are important in computational fluid dynamics as

scientists and engineers hope to shorten the computational time. Although the
multigrid method has been successful in achievement of quick convergence,
this approach tends to show poor performance in parallel computing with
the domain decomposition technique, especially when a small number of grid
points are assigned to one processor. Another shortcoming of the multigrid
method is complicated numerical procedures.
On the other hand, there is a class of acceleration techniques called extrap-
olation methods, in which the converged solution is predicted from a series
of intermediate solutions generated by a basic iterative solver. Shimano and
Arakawa[1] applied several extrapolation methods to a steady incompressible
Navier-Stokes solver to find that the acceleration effect hardly deteriorated
in parallel computing based on domain decomposition. Each extrapolation
method can be implemented with no modification to the original solver be-
cause the extrapolation procedures are independent of the solver. This sim-
plicity is great advantage common to all extrapolation methods.
Powerful computers nowadays allow researchers to carry out unsteady sim-
ulations like LES. As in steady problems, it is crucial to reduce the computa-
tional time for unsteady calculations. However, performance of extrapolation
methods for unsteady problems has not been reported.
In this study, as a first trial, the authors applied two extrapolation meth-
ods, TVE (Transition Vector Extrapolation) and MPE (Minimal Polynomial
Extrapolation), to an unsteady incompressible Navier-Stokes solver in order
to investigate their acceleration effects. The flow around a 2-D rod was solved
as it was considered to be appropriate for this basic research. TVE is a new
extrapolation method developed by the authors for unsteady problems while
MPE is one of existing extrapolation methods originally used for steady prob-
lems. A joint effect of TVE and MPE is discussed in the paper.

304 Kenjiro Shimano, Shun Yonezu, and Yoshiteru Enomoto
2 Extrapolation methods
Use of an implicit time-marching method is assumed. An intermediate solution
vector is identified by two indices, the time step and the number of iterations.
For example, xj,k represents the intermediate solution vector obtained after
k iterations at the j-th time step.
2.1 Minimal Polynomial Extrapolation (MPE)
MPE was proposed by Leland and Rollett[2] for linear systems. Its origin is
found in Cabay and Jackson[3] and Mevsina[4].
When a conventional iterative method based on matrix splitting is used,
the intermediate solution vector xj,k is expressed in the form
N
X
xj,k = x˜j + ai λki φi (1)
i=1
where N stands for the degree of freedom, x˜j is the converged solution at the
j-th time step, and λi and φi are the i-th eigenpair of the iteration matrix.
Although eigenpairs depend on the time step, the subscript j is omitted in
Eq.(1) for simplicity. Suppose that the eigenvalues are arranged in order of
magnitude, namely 1 ≥ |λ1 | ≥ · · · ≥ |λN |. It is assumed in MPE that the first
p eigenvalues λ1 , · · · , λp are dominant and responsible for slow convergence:
p
X
xj,k = x˜j + ai λki φi (p N ). (2)
i=1
From Eq.(2), the shift vector sj,k ≡ xj,k − xj,k−1 is expressed in the similar
form
p
X
sj,k = bi λki φi . (3)
i=1
The slowly decaying errors are removed by linear extrapolation and the con-
verged solution x˜j can be obtained from
(c0 + c1 + · · · + cp−1 )x˜j = c0 xj,k + c1 xj,k+1 + · · · + cp−1 xj,k+p−1 (4)
where {c0 , · · · , cp−1 } is a set of coefficients such that
c0 sj,k + c1 sj,k+1 + · · · + cp−1 sj,k+p−1 = 0. (5)
In practice, the coefficients {c0 , · · · , cp−1 } are calculated from the least square
problem derived from Eq.(5) because Eq.(5) itself is not determinate. As most
eigenvalues are neglected, MPE is not effective unless the slowly decaying
errors are dominant and should not therefore be applied at an early stage of
iterations.
MPE explained above was originally developed for linear systems. The
applicability of MPE to non-linear systems was theoretically given by Shimano
Acceleration of Flow Calculation Using Extrapolation Method 305
and Arakawa[1]. They recommended that p should be larger for non-linear

systems than for linear ones. For instance, they chose p = 19 for the lid-
driven square cavity flow at Re = 100 and p = 29 at Re = 1000. These values
are larger than those used in Leland and Rollett[2] who solved a Poisson’s
equation with MPE.
2.2 Transition Vector Extrapolation (TVE)
TVE predicts the converged solution at the present time step from a series of
converged solutions at previous time steps.
First, we consider a linear system. In the time-marching, an increment of
the converged solution, ∆x̃j (≡ x̃j − x̃j−1 ), is subject to sequential linear map-
ping. The behaviour of the incremental vector is determined by eigenvalues of
the mapping. When the system contains periodic fluctuations, eigenvalues of
which absolute value is equal to unity play an important role. If the number
of such eigenvalues is M , one component of the incremental vector is written
in the following form:
M
X
∆x̃j = {am sin(j∆tωm ) + bm cos(j∆tωm )} (6)
m=1
where am , bm are some coefficients and ωm is the angular velocity determined

by the imaginary part of the corresponding eigenvalue. In TVE, only q( M )
dominant angular velocity components are taken into consideration. Presum-
ably, ∆x̃j can be approximated by a linear combination of the history vectors
{∆x̃j−1 , · · · , ∆x̃j−q }, i.e.,
q
X
∆x̃j ' ci ∆x̃j−i (7)
i=1
The coefficients {c1 , · · · , cq } should satisfy the following conditions for ∀ m ≤

q:
q
X q
X
ci cos(i∆t ωm ) = 1, ci sin(i∆t ωm ) = 0. (8)
i=1 i=1
However, it is practically impossible to obtain the coefficients from Eq.(8) be-

cause precise evaluation of each angular velocity is difficult. Instead of Eq.(8),
it is assumed that the relation (7) also holds true for j − 1, i.e.,
q
X
∆x̃j−1 ' ci ∆x̃(j−1)−i . (9)
i=1
The coefficients {c1 , · · · , cq } are calculated by solving the least square problem
derived from Eq.(9) with all components of ∆x̃j taken into account.
TVE is applicable not only to linear systems but also to nonlinear ones because
terms generated by nonlinearity can be expressed in the same form as Eq.(6).
3 Application to unsteady N-S solver

The 2-dimensional flow around a square rod was calculated by the SIMPLE
algorithm. Systems of discretised equations are solved iteratively by the SOR
method. The number of grid points was 200×120 and 400×240. The Reynolds
number was 100 and 500. As this was the first test of the extrapolation meth-
ods in the unsteady N-S analysis, high order schemes were not always neces-
sary. Therefore, the 1st order backward Euler scheme was employed for time-
dependent terms and convection-diffusion terms were evaluated by the hybrid
scheme. Calculation at each time step was terminated when dimensionless
residuals of all the equations became less than 10−6 .
The solution vector predicted by TVE at j-th time step was used as xj,0 ,
the initial solution vector from which the iteration process started. Therefore,
application of TVE was followed by calls of the iterative solver at each time
step. When TVE was not applied, the converged solution at the previous time
step, x̃j−1 , was adopted as xj,0 . Extrapolation by MPE was carried out after
some iterations. Timings of MPE application are listed in Table 1. Generally,
the timing has a great influence on the acceleration effect and the optimal
timing depends on the problem solved. Those shown in Table 1 were the best
timings as long as the authors investigated by trial and error.
The extrapolation parameters, p and q, should be as small as possible be-
cause the time for extrapolation is approximately proportional to max(p3 , q 3 )
and because large values for the parameters would require a large memory
space. In this study, both parameters were equal to 5. No significant reduc-
tion in the number of iterations was observed even when values larger than 5
were used for p or q.
Figures 1 and 2 show residual histories at the 14841st time-step in calcu-
lation of the flow at Re = 500 with 200 × 120 grid points. It can be seen in
Fig.1 that a better initial guess was provided by TVE and resulted in quicker
convergence. However, the rate of the residual decay was almost the same as in
the case of no extrapolation. In addition, it is common to ”No extrapolation”
and ”TVE” that the decay of the residual slowed down when the residual
reached around 4 × 10−6 in Fig.1(a) and 2 × 10−6 in Fig.1(b), respectively. It
can be concluded that TVE hardly removes slowly decaying error components,
which are related to large eigenvalues of the iteration matrix.
MPE is effective for elimination of those error components responsible for
slow convergence. As shown in Fig.2, the combination of MPE and TVE made
Table 1. The number of iterations before MPE application
Re=100 Re=500
200 × 120 400 × 240 200 × 120 400 × 240
MPE only 25 25 25 25
TVE+MPE 15 25 15 25
Acceleration of Flow Calculation Using Extrapolation Method 307
the solution converge faster than TVE alone. Despite a jump-up in the residual
caused in the x-momentum equation, decay of the residual was considerably
quick after the MPE application.
Remarkably, the rapid residual reduction by MPE lasted longer when TVE
and MPE were jointly used. TVE worked as a preconditioner which strength-
ened the effect of MPE. This can be also observed in Fig.3 where averaged
speed-up factors are compared. At each calculation except Re = 100 with
400 × 120 grid points, the speed-up achieved by TVE+MPE is higher than
the product of those achieved by TVE and MPE separately. The speed-up by
TVE+MPE ranged approximately from 2.0 to 4.4. In calculation at Re = 100
with 400 × 120 grid points, the solution did not converge when only TVE was
applied. Further study is necessary to investigate the reason.
10-1 10-1
No extrapolation No extrapolation
10-2 TVE 10-2 TVE
10-3 10-3
Residual
Residual
10-4 10-4
10-5 10-5
10-6 10-6
10-7 10-7
20 40 60 20 40 60
Iterations Iterations
(a) x-momentum equation (b) Continuity equation
Fig. 1. Effect of TVE
10-1 10-1
MPE MPE
10-2 TVE+MPE 10-2 TVE+MPE
10-3 10-3
Residual
Residual
10-4 10-4
10-5 10-5
10-6 10-6
10-7 10-7
20 40 60 20 40 60
Iterations Iterations
(a) x-momentum equation (b) Continuity equation
Fig. 2. Effect of MPE (MPE applied after 15 iterations)

5
TVE 4.41
4 MPE
TVE+MPE
Speed-up
3
2.59 2.65
2.15
1.98 1.92
2 1.76
1.40 1.50
1.35
1.19
1
0
Re=100 Re=100 Re=500 Re=500
200x120 400x240 200x120 400x240
Fig. 3. Speed-up
Conclusions
A new extrapolation method, TVE, was proposed for unsteady flow analysis.
A joint effect of TVE and an existing extrapolation method,MPE, was inves-
tigated by applying them to an unsteady N-S solver. TVE provided a good
initial guess for the iterative solver. An acceleration effect of MPE was im-
proved by joint application of TVE because TVE worked as a preconditioner
for MPE. As a result, the speed-up achieved by the combination of TVE and
MPE ranged approximately from 2.0 to 4.4.
References
1. Shimano, K., Arakawa, C.: Incompressible Navier-Stokes Solver Using Extrap-
olation Method Suitable for Massively Parallel Computing, Comp. Mech., 23,
172–181 (1999)
2. Leland, R.W., Rollett,J.S. : A comparison of parallel extrapolation methods
In: Hackbusch, W. (ed) Parallel Algorithms for Partial Differential Equations.
Viewveg (1991)
3. Cabay, S., Jackson, W.: A Polynomial Extrapolation Method for Finding Limits
and Antilimits of Vector Sequences, SIAM J. Numer. Anal., 13-5, 734–752 (1976)
4. Mevsina, M.: Convergence Acceleration for the Iterative Solution of the Equa-
tions X=AX+f, Comp. Meth. in Appl. Mech. and Eng., 10, 165–173 (1977)
Improved Numerical Simulations of
Incompressible Flows Based on
Viscous/Inviscid Interaction Procedures
M. Hafez1 , A. Shatalov2 , and M. Nakajima3

1
University of California, Davis, USA mhafez@ucdavis.edu
2
University of California, Davis, USA
3
Kanazawa, Institute of Technology, Japan
Summary: Viscous/Inviscid Interaction procedures consist usually of cou-

pling potential flow and boundary layer calculations. In this study, the in-
teraction is modeled using Helmholtz-type velocity decomposition where the
gradient of the potential is augmented with a correction accounting for the
vorticity effects in the viscous layers. Different ways to calculate the rotational
components are discussed and methods to systematically improve the model
are studied. Numerical results are compared with standard Navier-Stokes cal-
culations to justify the present approach.
1 Introduction
Inviscid potential flows do not satisfy the no-slip boundary condition at a

solid surface and have zero drag for smooth bodies, for example cylinders
and spheres ( D’Alembert paradox ). On the other hand, standard boundary
layer theory with specified streamwise pressure gradient suffers from Gold-
stein singularities at separation points and trailing edges. Coupling potential
and boundary layer can produce reasonable solutions and remove some of the
difficulties associated with each simulation by itself. Research on interacting
boundary layers is aimed at providing engineering tools at least for some flow
simulations which do not require full Navier-Stokes calculations. On the the-
oretical side, sophisticated asymptotic triple deck theories, with a transition
layer within the boundary layer, were developed to study fundamental aspects
of these singular perturbation problems.
In the present paper, a new formulation is used to facilitate the coupling
and to avoid some limitations of boundary layer calculations. The model is
based on an additive correction to the potential flow in the viscous layer,
using Helmholtz-type velocity decomposition. The rotational components are
calculated from the momentum equations with a fictitious pressure gradient

310 M. Hafez, A. Shatalov, and M. Nakajima
term. The latter is obtained via analytical continuation of the pressure of

the potential flow region to the viscous layer, via a modified Bernoulli’s law.
The relations between this model and standard boundary layer theory on one
hand and Navier-Stokes equations on the other hand are briefly discussed.
Numerical results for steady and unsteady incompressible laminar flows over
airfoils at moderate Reynolds numbers are presented.
2 Viscous/Inviscid Interaction Procedures

Prandtl suggested to calculate the displacement thickness δ ∗ from the bound-
ary layer solution, to augment the body with δ ∗ , and to solve the potential
flow equation over the augmented body. The iterations based on this process
converge for attached flows. For separated flows, the boundary layer equations
with specified pressure gradient cannot be solved via a marching scheme. This
difficulty can be avoided if an inverse boundary layer model is adopted where
the displacement thickness at the outer edge, rather than the inviscid veloc-
ity is described. Other fixes, based on semi-inverse or simultaneous solution
strategies were developed by several authors to obtain efficient coupling. Fast
convergence, however, is required otherwise the interactive procedures are not
attractive.
Even if the singularities of the boundary layers can be removed, through
the interaction, the solution is not locally accurate. In the region near the
separation point the tangential and normal velocity components, are of the
same order while at the trailing edge the pressure is singular. Other difficulties
may still exist in certain areas, for example, in three dimensional flows over
wing tips and wing body junctions. The boundary layer assumptions allow for
great simplifications; the equations become locally of parabolic nature and the
pressure gradient across the layer is ignored. However, these simplifications
are associated with limitations and there is a trade-off between the level of
fidelity of the model and the limits of its validity.
In ref [1] , [2], the authors introduced the following velocity splitting:
u = φx + u∗ , v = φy + v ∗ (1)
Hence, the continuity equation yields a Poisson’s equation for φ:
φxx + φyy = −u∗x − vy∗ (2)
with φn = 0 at the solid surface and a far field solution at the outer boundaries.
The velocity components, u∗ and v ∗ are calculated from the momentum
equations, imposing no slip and no penetration boundary conditions at the
solid surface:
px
ut + uux + vuy = − + ν(uxx + uyy ) (3)
ρ
Simulations of Incompressible Flows Based on Viscous/Inviscid Interaction 311
py
vt + uvx + vvy = − + ν(vxx + vyy ) (4)
ρ
The correction terms to the potential field u∗ and v ∗ are calculated only
in the viscous layer and they vanish in the inviscid region.
The level of approximation of this model depends on the calculation of
the pressure. In the present work, the pressure is calculated from a modified
Bernoulli’s law:
p 1 2
+ (φ + φ2y + 2φt ) − ν(φxx + φyy ) = c(t) (5)
ρ 2 x
Substitution of Equations 1 and 5 in 3 and 4 yields:
u∗t + uu∗x + vu∗y = −φxx u∗ − φxy v ∗ + ν(u∗xx + u∗yy ) (6)
and
vt∗ + uvx∗ + vvy∗ = −φyx u∗ − φyy v ∗ + ν(vxx
∗ ∗
+ vyy ) (7)
with the boundary conditions at the solid surface:
u∗ = −φx , v ∗ = −φy (8)
Notice that the potential is not a single valued function and a cut is intro-
duced, along which the circulation Γ is calculated to guarantee the continuity
of the pressure, i.e.
Γt + hφs i Γs = νΓss (9)
where s is in the flow direction, starting from the trailing edge to the outer
boundary. For steady flow, Γ is constant along the cut. A non-reflecting
boundary condition is used in the far field exit plane for all the quantities. No-
tice also that φx and φy , like u and v, are governed by 2-D Burgers’ equations
with non homogeneous terms.
Similar, but not the same, ideas can be found in literature, see for example
ref. [3] where generalized boundary layer equations are derived via asymptotic
analysis. In ref. [4] Gordnier and Rubin used multiplicative correction to the
gradient of the potential to solve the Navier Stokes equations.
The present formulation can be viewed also as a zonal method based on
heterogeneous domain decomposition. However, continuation of the potential
and the pressure inside the viscous layer eliminates some of the interface
difficulties associated with the standard zonal methods, see ref.[5].
At any rate, a comparison with the standard pressure correction methods
where a Poisson’s equation for the pressure together with the momentum
equations are solved everywhere in the field reveals some of the merits of the
present formulation.
(a) u-velocity contours using (b) u-velocity contours using pressure

Present Formulation from Normal Momentum Equation
Fig. 1. u-velocity contours, Re = 500, α = 10◦
0.5
p
−C
−0.5
−1
0 0.2 0.4 0.6 0.8 1

x/c
(a) Surface Pressure (b) Skin Friction
Fig. 2. Surface pressure and skin friction, Re = 500, α = 10◦
3 Numerical Results
Algebraic grids are first generated around NACA 0012. The governing equa-
tions are discretized using finite volumes. Line relaxations are adopted to
solve the potential equation iteratively together with tri-diagonal solvers for
the momentum equations, marching in the flow direction.
Improvement of the numerical method in the present work includes the
solution of potential equation in a mapped plane using ADI method and using
a third order upwinding scheme for the convection terms in the momentum
equations. Details of the Numerical Methods are given in ref [7].
In figures 1 - 2 , the results for steady flow calculations at Re = 500
and α = 10◦ are shown. In figures 1(a), 1(b), the u-velocity contours are
plotted based on the present formulation and based on the method of ref.
[1] where the normal momentum equation is solved for the pressure. The
Simulations of Incompressible Flows Based on Viscous/Inviscid Interaction 313
corresponding surface pressure and skin friction distributions are given in

figures 2(a) , 2(b). The results for unsteady flow at Re= 800 and α = 20◦ are
shown next. The u-velocity contours, and the v-velocity contours are plotted
in figures 3(a), 3(b) and 4(a), 4(b) respectively. The development of the v-
velocity component at selected points are given in figures 5(a) and 5(b). The
Navier-Stokes calculations are based on SMAC method of ref [6].
For external aerodynamic problems, it is not necessary to solve Navier Stokes
equations everywhere in the field. Viscous / Inviscid interaction procedures
with potential flow in the outer region is an alternative. Using Helmholtz-
type velocity decomposition and a modified Bernoulli’s law for the pressure
in terms of the potential, leads to an attractive formulation. Preliminary nu-
merical results are encouraging for both steady and unsteady laminar incom-
pressible flow simulations. Unlike the stream function / vorticity formulation,
the present method can be extended to three dimensions in a straight forward
manner.
References
1. Hafez, M. , Shatalov A., Wahba, E.: Numerical Simulations of Incompress-
ible Aerodynamic flows using Viscous / Inviscid Interaction Procedures. Comp.
Methods in Appl. Mech. and Engrg., Vol 195, pp. 3110-3127 (2006)
2. Shatalov, A., Nakajima, M., Hafez, M.: Numerical Simulations of Unsteady Lam-
inar Incompressible Flows using Viscous / Inviscid Interaction Procedures, CFD
Journal, Vol 15, pp. 246-254 (2005)
3. Cousteix, J., Mauss, J.: Approximations of the Navier-Stokes equations for high
Reynolds number flows past a solid wall, Journal of Computational and Applied
Mathematics, Vol 166, pp. 101-122 (2004)
4. Gordnier, R.E., Rubin, S.G.: Flow Simulations using a composite velocity pro-
cedures for potential, Euler and RANS equations, Computer and Fluids, Vol 17,
pp. 85-98 (1989)
5. Quarteroni, A.: Domain Decomposition Methods for PDE’s, Oxford University
Press (1999)
6. Amsden, A., Harlow, F.: A Simplified MAC Technique for Incompressible Fluid
Flow Calcualtions, JCP, Vol 6, pp. 322-325 (1970)
7. Shatalov, A.: Ph.D. Dissertation, Univ. of Calfornia, Davis (2006)
time=5 time=5
2 2
1.5 1.5
1 1
y
y
0.5 0.5
0 0
−0.5 −0.5
−0.5 0 0.5 1 1.5 2 2.5 3 −0.5 0 0.5 1 1.5 2 2.5 3
x x
(a) present method (b) SMAC method

3rd order upwinding 3rd order upwinding
Fig. 3. u-velocity contours, Re = 800, α = 20◦
time=5
time=5
(a) present method (b) SMAC method

3rd order upwinding 3rd order upwinding
Fig. 4. v-velocity contours, Re = 800, α = 20◦
SMAC
1 Present Method (2nd order upwinding)
1
Present Method (3rd order upwinding)
0.8
0.8
0.6 0.6
v
0.4 0.4
0.2 0.2
0 0
−0.2
−0.2 0 5 10 15 20 25 30
0 2 4 6 8 10 12 14 16 18 20
time (a) time
at x = 1.1 , y = 0.0 (b) at x = 3.0 , y = 0.0
Fig. 5. Time variation of v-velocity, Re = 800, α = 20◦

Part IV
AMR - Adaptive mesh techniques

A Parallel Unstructured Overset Mesh
Technique for Unsteady Flow Simulations
Mun Seung Jung1 and Oh Joon Kwon2

1
Graduate Research Assistant chimera@kaist.ac.kr
2
Professor, Department of Aerospace Engineering, Korea Advanced Institute of
Science and Technology, 373-1 Guseong-dong, Yuseong-gu, Daejeon 305-701,
KOREA ojkwon@kaist.ac.kr
1 Introduction
In practical applied aerodynamics, solutions of unsteady time-accurate three-

dimensional flows involving multiple bodies in relative motion are frequently
required. Typical examples include the store/airframe separation from aircraft
or the rotor-fuselage interaction problem in helicopter aerodynamics. Several
research works have been previously conducted for the simulation of the com-
plicated unsteady flows involving mutual aerodynamic interference of multiple
bodies by solving the Euler or Navier-Stokes equations using either structured
or unstructured overset grid methods [1, 2, 3].
It is well known that structured overset grid methods are not grid-efficient
and require excessively large number of cells to enhance the resolution of the
solution by mesh adaptation because of the regularities required in the grid
point distribution. On the other hand, unstructured mesh topologies are based
on randomly distributed points, and thus have a great flexibility in generat-
ing cells, particularly around complex configurations. However, unstructured
mesh flow solvers require computational resources larger than those for struc-
tured grid counterparts. Therefore, it is crucial to enhance the performance of
unstructured mesh flow solvers, especially for overset mesh techniques which
require extra memory and computational overhead.
In the present study, an unstructured overset mesh method has been devel-
oped for the simulation of unsteady flow fields around multiple bodies in
relative motion, including overlapping of each other. The flow solver was par-
allelized for the efficient calculation of complex flows requiring a large number
of cells. For the validation of the present method, single and multiple stores
separating from the Eglin/Pylon configuration and the flow around a generic
missile with deploying fins were simulated. The present method coupled with
a quasi-unsteady solution-adaptive mesh refinement technique was validated
by simulating the helicopter rotor-fuselage interactional problem.

318 Mun Seung Jung and Oh Joon Kwon
2 Numerical Method
2.1 Discretization of Governing Equations
The governing unsteady Euler equations were discretized using a cell-centered

finite-volume method in conjunction with Roe’s flux-difference splitting. To
obtain second-order spatial accuracy, estimation of the state variables at each
cell face was achieved by interpolating the solution using a Taylor series ex-
pansion in the neighborhood of each cell center. An implicit time integration
algorithm based on the linearized second-order Euler backward difference was
used to advance the solution in time. Dual-time stepping was used to mini-
mize the errors involved in the linearization. The linear system of equations
was solved at each time step using a point Gauss-Seidel method.
On the solid surface of the body, the flow tangency condition was applied.
Characteristic boundary conditions were applied to the far-field boundary us-
ing the fixed and extrapolated Riemann invariants.
In order to reduce the large computational time, a parallel algorithm based on
a domain decomposition strategy was adopted. The load balancing between
processors was achieved by partitioning the global computational domain into
local subdomains using MeTiS libraries. The Message Passing Interface was
used to transfer the flow variables across the subdomain boundaries. All cal-
culations were performed on PC-based Linux clusters.
For the purpose of reducing the numerical dissipation and enhancing the spa-
tial accuracy of the solution, a solution-adaptive mesh refinement procedure
was adopted. To avoid excessive computational time required for the dynamic
mesh adaptation applied to unsteady flows, a ‘quasi-unsteady’ mesh adapta-
tion technique was adopted [4].
2.2 Unstructured Overset Mesh Technique
For overset mesh methods, a search procedure is needed for identifying the lo-
cation of the nodes on opposite overlapping blocks. For unstructured meshes,
the search should be performed for all nodes of all mesh blocks, because the
nodes and the cells are randomly distributed. Once the search process is com-
pleted for all nodes, the information is used for clarifying the node type, either
active or non-active, and for determining the weighting factors for interpola-
tion. To overcome the large overhead involved in the search process, a fast
and robust neighbor-to-neighbor search technique was implemented by uti-
lizing the property of the linear shape functions. The geometric relationship
between the beeline connecting the starting position and the target point and
the faces of the cells the beeline crosses are shown in Fig.1 (a).
Under parallel environment, this point search may develop across the subdo-
main boundary, and thus the number of data information in each processor
changes at every iteration. To handle this varying number of data efficiently, a
new data structure for parallel distributed memory machines was introduced
Parallel Unstructured Overset Mesh Technique 319
as shown in Fig.1 (b). Each node is tagged as either active or non-active based
on the distance-to-wall technique [2]. Determination of cell type, active, in-
terpolation, or non-active, for hole-cutting was made based on the number of
active nodes assigned to each tetrahedral cell.
Since transfer of the flow variables between mesh blocks was made through
interpolation, the results obtained from overset mesh methods do not satisfy
the conservation property. The local solution accuracy degrades furthermore
when the characteristic cell size changes drastically from one mesh block to
another. In order to reduce this error, the interpolation was made by consid-
ering not only the cell itself containing the interpolation receiver but also the
neighbor cells enclosing the cell.
To treat overlapped bodies in relative motion, those cells with all nodes lo-
cated underneath the solid boundary of other objects were excluded from the
calculation, while the cells with nodes partially immersed inside the body were
treated as active cells for computation. The interpolation of the flow variables
for these cells was made such that the flow tangency condition was satisfied
at the solid surface as shown in Fig.1 (c). For unsteady flow calculations, the
flow variables were assigned also for non-active cells by interpolation and ex-
trapolation from surrounding cells so that these variables can be used in the
subsequent time steps as they become active.

3.1 Store Separation
For the validation of the present method, single/triple ejector rack cases were
simulated by modelling each individual body with a separate mesh block.
The main block covers the entire flow field including wing/pylon, and the
other blocks represent the stores. The stationary main block was composed
of 404,738 tetrahedral cells and each subgrid block was composed of 249,536
tetrahedral cells. The calculations were made at a freestream Mach number
of 0.95 and an angle of attack of 0 degree. The six degree-of-freedom equa-
tions of motion were coupled to the flow equations for the prediction of the
time-accurate store trajectories. The results in Figs.2 (a)–(d) show that the
predicted trajectories compare well with the experimental data [5] and other
predictions [6] for all stores. In Figs.2 (e) and (f), the separation sequence of
single and three stores are presented at every 0.1 sec.
3.2 Generic Missile with Deploying Fins
The second validation of the present method was made for a generic missile
with deploying fins. The fins are made of a NACA0018 airfoil section and
are located at 80% from the nose of the body. The calculation was made at
320 Mun Seung Jung and Oh Joon Kwon
a freestream Mach number of 0.8 and an angle of attack of 0 degree. The

deploying velocity of the fins was a Mach number of 0.1. To model the com-
plete system, four unstructured mesh blocks were used as shown in Figs.3
(a) and (b). The stationary main block covering the main body of the mis-
sile was composed of 680,312 tetrahedral cells, and each subgrid for the fin
was composed of 314,436 tetrahedral cells. In Figs.3 (c)–(h), the mesh blocks
after hole cutting and the pressure contours at selected time levels are pre-
sented, demonstrating the usefulness of the present method for the simulation
of overlapping bodies in relative motion.
3.3 Rotor-Fuselage Configuration

The last validation of the present method was made for the rotor-fuselage
configuration tested at Georgia Tech [7]. The rotor operated at a blade tip
Mach number of 0.295 and an advancing ratio of 0.1. The blade collective
pitch angle was set to 10 degrees, and the measured thrust coefficient was
0.0091. The rotor shaft had a forward tilt of six degrees.
The initial coarse mesh for this configuration consisted of 499,301 cells in the
stationary main block covering the fuselage and 787,261 cells in the rotating
blocks for blades. After the second level mesh adaptation, the number of cells
increased to 3,032,870 and 2,331,001, respectively. In Fig.4 (a), the computa-
tional mesh on the blade and fuselage surfaces and at the fuselage symmetric
plane after the mesh adaptation is presented. It shows that the cell refinement
was mostly made along the vortical wake trajectory.
In Fig.4 (b), the perspective view of the instantaneous iso-vorticity surface is
presented. Formation of the trailing tip vortices, their migration to the down-
stream of the rotor, and the impingement of the tip vortex to the fuselage
can be confirmed in the figure. In Figs.4 (c) and (d), the streamwise distribu-
tions of the time-averaged downward and streamwise velocity components at
15% chord lengths above the fuselage are presented. It shows that the present
results compare reasonably well with other predicted results based on a mo-
mentum source modelling of the rotor [8] and the experimental data [7] for
both velocity components. The two peaks and the steep velocity gradient due
to the passage of the tip vortex at the front and rear parts of the fuselage
were well predicted.
A parallel unstructured overset mesh method has been developed for the sim-
ulation of unsteady flow fields. Applications were made to the single/triple
ejector rack, a generic missile with deploying fins, and a rotor-fuselage config-
uration. The predicted results showed reasonable agreement with experiment
and other predictions within the accuracy of the present inviscid calculation.
It was demonstrated that the present method is efficient and robust for sim-
ulating unsteady flows involving multiple bodies in relative motion.
Parallel Unstructured Overset Mesh Technique
321
Fig. 1. Node
Fig. search,
1. Node parallelization
search, parallelization and interpolation
and interpolation inside
inside the solid
the solid surface
surface
Fig. 1. Node search, parallelization and interpolation inside the solid surface
Fig. 1. Node search, parallelization and interpolation inside the solid surface
Fig. 2. Trajectories of stores and separating sequence at every 0.1 sec.

Fig.Fig.
2. Trajectories
2. Trajectoriesofofstores
stores and separatingsequence
and separating sequence at every
at every 0.1 sec.
0.1 sec.
Fig. 2. Trajectories of stores and separating sequence at every 0.1 sec.
Fig. 3. Pressure contour and hole-cutted meshes for a missile with deploying fins
at selected time levels
Fig. 3. Pressure contour and hole-cutted meshes for a missile with deploying fins
at
3.selected
Fig.Fig. Pressuretimecontour
3. Pressure levels
contourand
andhole-cutted meshesforfor
hole-cutted meshes a missile
a missile withwith deploying
deploying fins fins
at selected time
at selected levels
time levels
322
6 Mun
Mun Seung
Seung Jung and Oh
Jung and Oh Joon
Joon Kwon
Kwon
Fig. 4.
Fig. 4. Rotor-fuselage
Rotor-fuselage configuration
configuration and
andtime-averaged
time-averagedvelocity
velocitydistributions
distributions
References
[TJ95] Thoms, R.D., Jordan, J.K.: Investigation of Multiple-Body Trajectory
References
Prediction Using Time-Accurate Computational Fluid Dynamics. AIAA
paper 95-1970 (1995)
[NT00]
1. Thoms,Nakahashi, K.,Togashi,
R. D., Jordan, F., Sharov, of
J. K. : Investigation D.:Multiple-Body
Intergrid-Boundary Definition
Trajectory Predic-
tion Using Time-Accurate Computational Fluid Dynamics. AIAA paper 38,11,
Method for Overset Unstructured Grid Approach. AIAA Journal, 95-1970
(1995) 2077–2084 (2000)
[TI03]
2. Togashi,
Nakahashi, F., Ito, Y.,
K., Togashi, F.,Nakahashi,
Sharov, D. K., Obayashi, S.: Overset
: Intergrid-Boundary Unstructured
Definition Method
for Overset Unstructured Grid Approach. AIAA Journal, 38, 11,2003-3405
Grid Method for Viscous Flow Computations. AIAA paper 2077–2084
(2000) (2003)
[PK04] Park, Y.M., Kwon, O.J.: Simulation of Unsteady Rotor Flow Fields Using
3. Togashi, F., Ito, Y., NaNakahashi, K., Ohayashi, S. : Overset Unstructured Grid
Unstructured Sliding Meshes. Journal of the American Helicopter Soci-
Method for Viscous Flow Computations. AIAA paper 2003-3405 (2003)
ety,49,4,391–400 (2004)
4. Park, Y. M., Kwon, O. J. : Simulation of Unsteady Rotor Flow Fields Using
[LS95] Lijewski, L.E., Suhs, N.E.: Time Accurate Computational Fluid Dynamics
Unstructured Sliding Meshes. Journal of the American Helicopter Society, 49,
Approach to Transonic Store Separation Trajectory Prediction. Journal
4, 391–400 (2004)
of Aircraft, 31,4,886–891 (1995)
5. Lijewski,
[LP00] L. E.,Park,
Lee, S., Suhs,M.,
N. E. : Time
Cho, K.W.,Accurate Computational
Kwon. J.H.: New Fully Fluid Dynamics
Automated Ap-
Proce-
proach dure
to Transonic Store Separation
for the Prediction Trajectory Journal
of Store Trajecotry. Prediction. Journal37,6,1038–
of Aircraft, of Aircraft,
31, 4, 886–891 (1995)
1049 (2000)
6. Lee,
[LK89] S.,Liou, S.G.,Cho,
Park, M., K. W., Kwon,
Komerath, J. H. : New H.M.:
N.M., McMahon, Fully Automated Procedure for
Velocity Measurements
the Prediction of Store
of Airframe Trajectory.
Effects on a Rotor Journal of Aircraft,
in Low-Speed 37, 6,Flight.
Forward 1038–1049 (2000)
Journal of
7. Liou, S.Aircraft,
G. Komerath, N. M.,
24,4,340–348 McMahon, H. M. : Veloctity Measurements of
(1989)
Airframe
[ZR97] Effects
Zori, on a
L.A.J., Rotor in Low-Speed
Rajagopalan, Forward Flight.
R.G.: Navier-Stokes Journal of Aircraft,
Calculations Rotor-
24, 4, 340–348
Airframe(1989)
Interaction in Forward Flight. Journal of the American Heli-
8. Zori, L.copter
A. J., Rajagopalan, R. G. : Navier-Stokes
Society, 42,3,235–243 (1997) Calculations of Rotor-Airframe
Interaction in Forward Flight. Journal of the American Helicopter Society, 42,
3, 235–243 (1997)
A Parallel Overset Adaptive Cartesian/Prism
Grid Method for Moving Boundary Flows
Ravishekar Kannan1 and Z.J. Wang2

1
Iowa State University, Ames, IA 50011 kannan@iastate.edu
2
Iowa State University, Ames, IA 50011 zjw@iastate.edu
Summary. The use of overset grids in CFD started more than two decades ago,
and has achieved tremendous success in handling complex geometries[6]. In particu-
lar, overset grids have the advantage of avoiding grid remeshing when dealing with
moving boundary flow problems. In this paper, two particular unstructured grids
are advocated for moving boundary flow simulation, i.e., the use of overset adap-
tive Cartesian/prism grids. Semi-structured prism grids are generated around solid
walls. These prism grids then overlap a single adaptive Cartesian background grid.
With the adaptive Cartesian grid, the mesh resolution of the prism grid near the
outer boundary can easily match that of the oversetting Cartesian grid cells. In ad-
dition, the tree-based data structure of the Cartesian grid can be used efficiently in
hole-cutting and donor cell identification. In order to expedite the simulations, the
entire solver was made capable of running in parallel. The overset adaptive Carte-
sian/prism grid method is tested for both steady and unsteady flow computation. It
is demonstrated that moving boundary flow computations can be automated with
minimum user interferences.
1 Introduction
In this paper, we advocate the use of an overset adaptive Cartesian/prism grid

method for moving boundary flow computations. The method combines the
advantage of adaptive Cartesian/prism grid in geometry flexibility with that of
Chimera approach in tackling moving boundary flow without grid remeshing.
There are several reasons why an adaptive Cartesian grid is used for moving
boundary problems: 1. Cartesian cells are more efficient in filling space given a
certain length scale than triangular/tetrahedral cells; 2. Searching operations
can be performed very efficiently with the Octree data structure; 3. Solution
based and geometry-based grid adaptations are straightforward to carry out.
The grid generation process is as follows. Body-fitted prism grids are gener-
ated first near solid bodies to resolve viscous boundary layers. An adaptive
Cartesian grid is then generated to cover the outer domain and serve as the
background grid for bridging the ”gaps” between the prism grids. The prism

324 Ravishekar Kannan and Z.J. Wang
grids are used to generate holes in the adaptive Cartesian grid to facilitate
data communications. If the bodies move, the prism grids move with the bod-
ies, while the Cartesian grid remains stationary. After a few (tens of) time
steps, new holes are cut out of the Cartesian grids, and new donor cells are
also identified. Solution fields are interpolated from the old Cartesian grid to
the new grid using cell-wise linear reconstruction.
The paper is organized as follows. The numerical method is reviewed in the
next section. Solutions for some problems(using the current approach) are
presented in Section 3. Conclusions are summarized in Section 4.
2 Numerical Method
2.1 Generation of Overset Cartesian/Prism Mesh
Kallinderis[1] defined the term node-manifold as the list of faces confining the
node to be marched. Common sense tells us that the marching vector at any
node should not make an angle greater than 90◦ with the face normals of its
manifold. If the above criterion is violated, the tip of the marching vector
is not visible from all the faces of the manifold. So the paramount objective
here is to ensure that the marching vectors satisfy the visibility criterion. The
secondary objective is to impose orthogonality. An angle based weighting is
used to obtain the initial guess for the marching vector. The marching vector
is then refined locally in order to satisfy the secondary objective. An inverse
distance based smoothing was used to further smooth the marching vectors.
The marching step was obtained based on the local curvature. The nodes in the
current layer were smoothened. The intersections (if any) were obtained and
were corrected on a local basis. After the prism grid generation, an adaptive
Cartesian grid was generated automatically matching the grid resolution near
the outer boundaries of the prismatic grids
2.2 Finite Volume Method for Dynamic Grids
The time-dependent Reynolds-averaged Navier-Stokes equations for dynamic

grids can be expressed as
∂(QVi )
+ Σf (F i (Q) − Qvgn )f dSf = Σf Ffν (Q)dSf (1)
∂t
The conservation of mass, momentum and energy is a necessary condition for
any scheme. If we carefully examine the above equation, in order to preserve
free stream, we must have
∂Vi
= Σf vgn dSf (2)
∂t
This is the so-called Geometric Conservation Law in its semi discretized form.
A finite difference discretization of the above results in
Title Suppressed Due to Excessive Length 325
Vin+1 − Vin
= Σf vgn dSf (3)
∆t
The above equation is employed to obtain vgn . In addition, the volume swept
by a cell is the sum of volumes swept by its faces, i.e.
Vin+1 − Vin = Σf ∆Vf (4)
Comparing equations 3 and 4, we arrive at the following equation:
∆Vf
vgn = (5)
∆tdSf
2.3 Boundary Conditions
For a steady inviscid flow, the velocity components at a ghost cell for a solid
wall boundary are computed as
ughost = u − 2nx vn , vghost = v − 2ny vn (6)
where vn = unx +vny . For unsteady moving boundary problems, vn is modified
to unx + vny − vgn . For an unsteady viscous problem, the velocity components
at a ghost cell are computed as
ughost = −u + 2nx vgn , vghost = −v + 2ny vgn (7)
In the far field, a characteristic analysis based on Reimann invariants is used
to determine the values of the flow variables on the outer ghost cells. This
acts as a non-reflecting boundary for the unsteady applications. The solution
at the face center of the hole/interpolation boundary faces are obtained by
extrapolating the solution from the corresponding donor cell
3 Numerical Results
3.1 Problem 1: Turbulent flow over a moving sphere
This case was selected to validate the moving grid flow solver. The sphere
moves from right to left in quiescent air with a Mach number of 0.2. If the
reference frame is fixed on the moving sphere, the flow field should reach a
steady state after the initial transients propagate out of the solution domain.
The moving grid flow solver was first verified that the GCL was satisfied.
Then it was used to solve the moving sphere problem. The Reynolds number
of the flow was 1.1e6. The SA method was employed to close the turbulence
equations. Roughly 6.0e5 cells were used for the current simulation.
The density field created the moving sphere after a long time was compared
with that created by turbulent flow over a stationary sphere in figure 1. It is
observed that the density fields are similar. The Cp and the skin friction coeffi-
cients are plotted in figure 2. The Cp values are in good accord with the exper-
imental results[2] and the simulation results obtained by other researchers[3].
A good match could not be obtained for the skin friction coefficient. This
discrepancy was also mentioned by other researchers. This discrepancy is in
all probability due to the fact that the turbulence model enforces the effect of
turbulence throughout the computational domain. In reality the effect of tur-
bulence become predominant only after some distance. Hence the mismatch.
The Cd was in good agreement with the experimental values and is presented
in the below table
Table 1. Data on Drag; Experimental results from Aachanbach[2] and DES results
from Constantinescu[3]
Current Simulation 0.1
DES 0.084
Experiment 0.12
Fig. 1. The density distribution for moving sphere in quiescent air and flow around
a stationary sphere
1 8
Experiment
Experiment
DES 6 DES
Moving Grid Moving Grid
0
cf * sqrt(Re)
4
Cp
2
−1
0
−2 −2
0 60 120 180 0 60 120 180
Theta in degrees Theta in Degrees
Fig. 2. Comparison of Cp and Cf at Re=1.1e6
3.2 Problem 2: Thrust production using pitching+plunging airfoils
Thrust production by an oscillating NACA-0012 airfoil was studied. The air-

foil had a high aspect ratio(7.6). It moves at a constant speed U and executes
Table 2. Comparison of Thrust coefficients; Experimental results from Anderson

ho
c
St αmax Experiment Simulation
0.75 0.3 30o 0.7 0.66
0.75 0.2 30o 0.22 0.20
0.75 0.3 10o 0.31 0.30
0.75 0.2 10o 0.19 0.17
a harmonic plunging motion of amplitude ho and frequency ω while simultane-

ously undergoing a harmonic pitching operation of amplitude θo and frequency
ω. The plunging and the pitching operations have a phase difference of 90◦ .
The one third chord location was used as the pivot point.
During the beginning of the previous century, Knoller and Betz proved that
flapping wings result in production of thrust. In the mid 1920s, Birnbaum pro-
posed the use of flapping wings as substitutes to the propeller. Von Karman[4]
studied the wake structures and refined the above concept. Later experiments
by Koochesfahani (on pitching airfoils), Ohmi(1991) and Anderson[5] yielded
a gamut of information on vortex formation, optimal location of the pivot etc.
CFD predictions by Dong and Mittal(2005) were able to qualitatively predict
thrust prediction for many cases. Nevertheless, the quantitative match with
the experiments( for 3D cases) is not very encouraging.
There are two intrinsic parameters which dictate the thrust production: (1)
ω 2ho
the Strouhal number, St = 2π U and (2) the maximum Instantaneous angle
ωho
of attack, αmax = arctan( U )-θo
In general, the thrust generated is directly proportional to the Strouhal
number and increases by enhancing θo . The Reynolds number effects were neg-
ligible beyond a critical value(≈ 6000). The drag created by viscous stresses
was much smaller than the thrust by the pressure forces.Four flow cases were
simulated in the current study. The thrust obtained by these cases were com-
pared with the experimental results of Anderson[5]. The computed thrusts
concur with the experimental results and are presented in table 2. The dif-
ference between the coarse and fine grid results was less than 5/1000. The
experiment was quasi 2D: the ends of the airfoil were fitted with large circular
plates. This results in less slippage and hence higher thrust. Hence the exper-
imental thrust readings were slightly higher than the thrust obtained using
CFD. Vortices shed by the airfoil during the upstroke and downstroke are
shown in figure 3. This alignment of the vortices results in a positive x com-
ponent of velocity(behind the airfoil). This corresponds to a thrust indicative
pattern as explained by Von Karman[4].
4 Conclusions
The moving boundary method was tested on both turbulent and unsteady
problems. It produced good results for both the cases. This method combines
Fig. 3. Vortices shed by the airfoil during the upstroke and downstroke positions
the advantage of adaptive Cartesian/Prism grid in geometric flexibility with

that of Chimera approach in tackling moving boundary problems without
grid remeshing. The grid generator and the solver were coupled successfully
to enable efficient automation. Future research includes the following areas:
(1) extend the current 4 DOF solver to handle all the 6 DOF; (2) enable non
rigid body deformation; and (3) further validate with a high Reynolds number
store separation problem.
5 Acknowledgments
The authors gratefully acknowledge the support from the Air Force Office
of Scientific Research (Grant Number FA9550-04-1-0053), with Dr. Fariba
Fahroo as the Technical Monitor.
References
1. Kallinderis, Y., and Ward, S., ”Prismatic Grid Generation for 3-D Complex
Geometries,” AIAA Journal, Vol 31, No. 10, 1993, pp. 1850-1856.
2. Achenbach, E., 1974, ”Vortex shedding from the spheres”, J. Fluid Mech., 62,
pp. 209-221
3. Constantinescu, G.S., Pacheco, R. and Squires, K.D., 2002, ”Detached-Eddy
Simulation of Flow over a Sphere,” AIAA Paper 2002-0425
4. Von Karman, T., and Burgers, J.M., “General Aerodynamic Theory-Perfect Flu-
ids,” Aerodynamic Theory, Vol. 2. Julius-Springer, Berlin, 1943, p. 308.
5. Anderson, J.M, Streitlien, K., Barrett D.S, and Triantafyllou, M.S, ”Oscillating
foils of high propulsive efficiency,” JFM, Vol. 360, 1998, pp. 41-72
6. Togashi, F., Nakahashi K., Ito Y., Iwamiya T. and Shimbo Y.,”Flow Sim-
ulation of NAL Experimental Supersonic Airplane/Booster Using Overset
Grids”,Computers and Fluids, Vol 30, pp. 673-688,July 2001
Navier-Stokes Simulation of Local Winds Over
the Earth’s Topography
Neal M. Chaderjian1 , Jasim U. Ahmad2 , Marc G. Kramer3 , and Terry L.

Holst4
1
NASA Ames Research Center, Moffett Field,CA, Neal.M.Chaderjian@nasa.gov
2
NASA Ames Research Center, Moffett Field,CA, ahmad@nas.nasa.gov
3
NASA Ames Research Center, Moffett Field,CA, kramerm@FSL.ORST.EDU
4
NASA Ames Research Center, Moffett Field,CA, Terry.L.Holst@nasa.gov
Abstract
A numerical approach is described that simplifies and automates the CFD
solution process so that Earth scientists can utilize high-resolution Navier-
Stokes flow solvers as a research tool to investigate wind events on the Earth’s
surface. The current approach utilizes the OVERFLOW-2 structured overset
RANS code. A genetic algorithm is used to obtain an optimal multi-zone
overset grid system that reduces the grid size and simulation time by main-
taining high resolution over high-gradient land regions and lower resolution
over low-gradient water regions. Flow simulations are presented that include
flow separation and reattachment over mountainous terrain for coastal islands
in Alaska (USA) and British Columbia (Canada).
1 Introduction
Earth scientists often rely on local meteorological measurements and simple
linear airflow models to study how winds impact the Earth’s surface. Surface
winds directly impact the Earth’s surface by eroding and depositing dust and
soil. High speed surface winds can cause catastrophic damage to forest ecosys-
tems throughout the world . In high latitude coastal regions of North Amer-
ica (southeastern Alaska) and Europe maritime cyclonic windstorms disturb
forested ecosystems by snapping and uprooting trees and causing landslides.
[1]
Reynolds-averaged Navier-Stokes (RANS) simulations have not been read-
ily used to predict winds with micro-scale spatial resolution (grid resolutions
under 1 Kft) because these high-resolution simulations require large compu-
tational resources, and advanced Computational Fluid Dynamic (CFD) codes
often require expert knowledge to effectively use them. However, large super-
computing centers are becoming more readily available at government labs
and universities. For example, NASA’s Advanced Supercomputing (NAS) Di-
vision at Ames Research Center makes available to scientists 10,240 1.5 GHz

330 Chaderjian, Ahmad, Kramer and Holst
Intel Itanium 2 processors. This paper describes an approach to automate

much of the grid generation, flow solution, and data extraction process. The
goal is to make high-fidelity RANS simulations more readily available as a
science tool for researchers with little or no CFD experience. The solution
procedure is described in Section 2, numerical results are discussed in Section
3, and concluding remarks are made in Section 3.
2 Solution Procedure
The OVERFLOW-2 CFD RANS code [2] is used to simulate high-resolution
viscous flow over large coastal regions of Alaska and British Columbia. This
code uses overset structured grids, has a variety of steady-state and time-
accurate numerical integration algorithms, and several turbulence models. A
low Mach number preconditioner maintains solution accuracy and code effi-
ciency for low-speed flow simulations. Much of the solution process has been
automated in order to simplify code execution and provide a flow simulation
tool for scientists that are not expert users of CFD codes.
An Automated Planetary Grid (APG) code has been developed to au-
tomatically generate overset grids and OVERFLOW-2 input files. The grid
generation process begins by reading in a Digital Elevation Model (DEM)
of the Earth’s surface. The DEM is an ascii file that defines the terrain ele-
vation (z) on a uniformly spaced xy-grid. The micro-scale spatial resolution
is often on the order of tens to hundreds of feet. Since these coastal regions
involve islands surrounded by large bodies of water, improved computational
efficiency is obtained by keeping land regions at the DEM resolution and re-
ducing the resolution of water regions by a factor of two. These water regions
are flat and have small flow gradients. This is accomplished by partitioning
the terrain surface into structured overset water and land zones.
The zonal partition is established by first subdividing the terrain into small
water and land sub-grids, see Fig. 1. Water sub-grids are regions that only have
zero (sea level) elevations, while land sub-grids are regions that have at least
one non-zero elevation. The process of reassembling these sub-grids into an
optimal water and land zonal topology is based on a Genetic Algorithm (GA)
reported by Holst and Pulliam.[3] Certain rules and constraints are included in
the optimization process. For example, grid zones obtained by reassembling
sub-grids must be rectangular, a requirement of the OVERFLOW-2 code.
Moreover, when forming rectangular zones, water sub-grids can be converted
into higher resolution land sub-grids, but not visa versa. This is done to ensure
DEM resolution is maintained over all land regions. Additional constraints
include a minimum number of sub-grids in a zone, and a maximum zonal
aspect ratio to reduce the GA optimization time and avoid very small isolated
zones.
A two-objective GA optimization is performed, subject to the rules and
constraints mentioned above, that simultaneously minimizes the total number
of land zones and the total number of grid cells, including the overset zonal
Navier-Stokes Simulation of Local Winds Over the Earth’s Topography 331
overlap. A “pareto front” is the result of the optimization process, see the
example in Fig. 2. This pareto front represents the minimum number of grid
cells for a specific number of land zones. Any other combination of land and
water zones will have more surface cells and lie above the pareto front. In the
example shown in Fig. 2, the number of land zones varies from one to five
along the pareto front. Any combination of water and land zones that has
more than five land zones will result in a higher number of surface cells. The
choice of which combination of water and lands zones to use on the pareto
front depends on the requirements of the flow simulation. For example, one
may wish to have fewer land zones to simplify flow visualization or reduce the
number of zonal interfaces. In the cases presented in this study, the number
of zones that results in the least number of grid cells is chosen. This would
correspond to five land zones in the Fig. 2 example.
Fig. 1. Initial land (gray) and Fig. 2. Two-parameter GA Opti-

water (white) sub-grids. mization. Land Zone (gray); wa-
ter zone (white).
Once the surface zonal topology is determined, the APG code generates
volume grids using a hyperbolic grid generation method. User input controls
the grid attributes, e.g., initial body-normal spacing, stretching factors, and
far-field grid boundary height. The APG code also automatically generates all
OVERFLOW-2 input files and force/moment input files to monitor solution
convergence. Post-processing software is provided to extract the flow solution
at a fixed distance Above Ground Level (AGL) for data analysis and flow
visualization. The web-based AeroDB framework [4] is used to run the flow
solver and simplify the job submission and solution monitoring process.
3 Numerical Results
The OVERFLOW-2 RANS code [2] uses a Message-Passing Interface (MPI)
for parallel processing. The Pulliam-Chaussee diagonal central-difference al-
gorithm [5], together with matrix artificial dissipation, is used to integrate
the RANS equations. Steady solutions are obtained using a multi-grid tech-
nique with local time-stepping based on a constant Courant-Friedrichs-Levy
(CFL) condition. The Menter SST turbulence model [6] is used to model fully
turbulent flow.
Figure 3 shows the result of the APG zoning algorithm for the Queen
Charlotte Islands, British Columbia, Canada. This is an archipelago of more
than 150 islands with an overall length of 170 miles and width of 83 miles. The
surface grid resolution for land zones is 328 ft while the water zone resolution
is 656 ft. Throughout this paper, the water grid resolution is always twice
as coarse as land grids. The GA optimization selected 21 land grids and 15
water grids, resulting in a 25% reduction in the total grid size. Four slender
water grids 8 miles wide (using exponential stretching) are added around the
perimeter of the grid system to allow the turbulent boundary layer to grow to
sufficient height. Typical Earth boundary layers vary from 1000 ft to over 3000
ft, depending on the time of day and local weather conditions. This 40 zone
grid system consists of 262 million grid points and covers over 14,000 sq. mi.
The serial (1 CPU) APG code took 30 seconds for the GA zonal optimization,
and 50 minutes to generate all 40 volume grids on the NAS Columbia system.
Fig. 3. GA optimized zonal grid Fig. 4. OVERFLOW-2 wind

system for Queen Charlotte Is- speed at 32.8 ft AGL for Queen
lands. Dark regions (low eleva- Charlotte Islands. Dark regions
tion), light regions (high eleva- (low speed); light regions (high
tion). speed).
The wind speed computed by the OVERFLOW-2 RANS code is shown

in Fig. 4 on a surface at 32.8 ft AGL. This high-wind event has a freestream
windspeed of 75 mph from the South Easterly direction, a Reynold’s Number
(Re) of 700,000/ft, and a static temperature of 59 degrees Fahrenheit. The
APG code automatically provides all inputs and post-processing software to
Navier-Stokes Simulation of Local Winds Over the Earth’s Topography 333
utilize OVERFLOW-2’s overset interpolation capability to extract the 32.8

ft AGL surface. Figure 5(a) shows an enlarged view of a portion of the is-
land. Dark regions correspond to low wind speed while light regions are high
wind speed. Figure 5(b) shows local streamlines constrained to the extracted
surface. For the most part, the flow separates along mountain ridges and reat-
taches in the valleys between mountain ridges. The local wind speed also slows
down as the flow traverses several mountain ridges.
Depending on the flow direction, steady solutions for this 262 million grid-
point system requires 3000-5000 iterations to achieve convergence of the is-
land’s computed lift and drag coefficients, and the L2 norm of the resid-
ual. This corresponds to 24-39 wall-clock hours using 128 CPUs on the NAS
Columbia system. The data processing rate is 10−7 sec/iteration/grid-point.
(a) Wind speed shaded surface. (b) Streamlines and wind speed
Dark regions (low speed); light re- shaded surface. Dark regions (low
gions (high speed). speed); light regions (high speed).
Fig. 5. Enlarged views of the Queen Charlotte Islands, 32.8 ft AGL.
(a) Top view. (b) Side view.
Fig. 6. Wind speed shaded surface and streamlines at 27 ft AGL. Mitkof Island,
Alaska. Dark regions (low speed); light regions (high speed).
Fig. 6 shows the wind speed and flow streamlines on a surface 27 ft AGL
for Mitkof Island, Alaska, USA. The free stream conditions are identical to
the Queen Charlotte Islands, except this high-wind event comes directly from
the South. There are 4 land zones at 200 ft resolution and 3 water zones
at 400 ft resolution. As before, water zones are added around the perimeter
of the island to obtain an appropriate boundary-layer height. This island is

approximately 20 mi long and 16 mi wide. As before, high winds and flow
separation occur near mountain ridges, and low speed reattachment occurs
in the valleys. Figure 6(b) shows an enlarged side view and the approaching
velocity boundary layer profile.
4 Conclusions
Automated Planetary Grid (APG) generation software has been successfully
developed and demonstrated to simplify and automate the grid generation and
flow solution process for the OVERFLOW-2 CFD code. This software pro-
vides an automated high-resolution RANS simulation tool for Earth scientists
to investigate wind events over the Earth’s topography. Examples are given
where the APG software automatically generated multi-zone overset grid sys-
tems and flow-solver input files. Moreover, the APG software provides tools
to automatically extract solution data on a surface at a fixed distance above
ground level for scientific visualization and analysis. A robust genetic algo-
rithm efficiently determines an optimum grid-zoning partition that reduces the
total grid size and solution time by maintaining high spatial resolution over
land regions (with high flow gradients) and lower spatial resolution over water
regions (with lower flow gradients). The GA is flexible in that it is relatively
easy to include constraints and rules that provide robust control of grid quality
and zonal partition characteristics. RANS solutions using the APG software
and AeroDB web-based solution framework have been presented for coastal
islands in Alaska (USA) and British Columbia (Canada). The OVERFLOW-2
code is able to simulate complex separated flow over the Earth’s topography.
References
1. Kramer, M. G., Hansen, A. J., Taper, M., and Kissinger, E., “Abiotic Controls
on Windthrow and Forest Dynamics in a Coastal Temperate Rainforest, Kuiu
Island, Southeast Alaska,” Ecology, Vol. 82, No. 10, 2001, pp. 2749-2768.
2. Buning, P. G., Jespersen, D. C., Pulliam, T. H., Klopfer, G. H., Chan, W. M.,
Slotnick, J. P., Krist, S. E., and Renze, K. J., “Overflow User’s Manual, Version
1.8aa” NASA Langley Research Center, Hampton, VA, 2003.
3. Holst, T. L, Pulliam, T. H., “Evaluation of Genetic Algorithm Concepts Us-
ing Model Problems, Part II: Multi-Objective Optimization,” NASA/TM-2003-
212813, December 2003.
4. Chaderjian, N. M., Rogers, S. E., Aftosmis, M. J., Pandya, S. A., Ahmad, J.
U., and Tejnil, E., “Automated CFD Database Generation for a 2nd Generation
Glide-Back Booster,” AIAA Paper 2003-3788, June 2003.
5. Pulliam, T. H., and Chaussee, D. S., “A Diagonal Form of an Implicit
Approximate-Factorization Algorithm,” Journal of Computational Physics, Vol.
39, No. 2, 1981, pp. 347–363.
6. Menter, F.R., “Two-Equation Eddy-Viscosity Turbulence Models for Engineer-
ing Applications,” AIAA Journal, Vol. 32, No. 8, 1994, 1598–1605.
Grid-adaptive Simulations of Relativistic Flows
R. Keppens1,2,3 , Z. Meliani2,4
1
Centre for Plasma Astrophysics, K.U.Leuven
2
FOM-Institute for Plasma Physics Rijnhuizen, Nieuwegein
3
Sterrenkundig Instituut, Utrecht
4
Max-Planck-Institut fur Astrophysik, Garching
Summary. Motivated by observations of relativistic jets in galactic X-ray sources

and in radio galaxies, we embark on a numerical survey of relativistic jet streams.
We use second order, shock-capturing discretizations in finite volume treatments
for relativistic hydro- and magnetohydrodynamic flows in combination with adap-
tive mesh refinement (AMR). We ensure physical consistency between the primitive
(ρ, v, p, B) and conservative variables at limited linear reconstruction stages, as well
as at all AMR restriction and prolongation stages. Special relativistic hydro simu-
lations compare favorably with recent higher order implementations thanks to the
AMR strategy. Magnetized relativistic jet simulations confirm morphological differ-
ences controlled by the inlet ratio of kinetic to electromagnetic (Poynting) flux.
Astrophysical jets are observed to emanate from sources at a wide variety

of scales. In X-ray binaries, where a compact object revolves about a normal
star, radio observations show episodic ejection of relativistically flowing ‘blobs’
in a precessing twin jet. At galactic scales, relativistic, well-collimated jets up
to a million light years long are associated with Active Galactic Nuclei, where
accretion on massive black holes (masses up to 108 M ) is ultimately respon-
sible for the jet phenomenon. Synchrotron emission from these jets provides
direct evidence of their magnetization. The propagation dynamics of rela-
tivistic jets can only be assessed quantitatively by means of shock-capturing
numerical treatments at high resolutions. We present an update of our ef-
forts towards impulsively injected magnetized jet simulations, encountering
denser clouds, where the jet is characterized by its terminal Lorentz factor,
and internal plasma beta parameter.
1 Relativistic Fluid treatments

Restricting the analysis to special relativistic magnetofluids, the governing
physical laws can be expressed in flat Minkowski space-time. They incor-
porate particle conservation, conservation of energy-momentum – a second
rank tensorial quantity in space-time including stress and electromagnetic

336 R. Keppens, Z. Meliani
contributions – and the Maxwell equations. In the ideal magnetohydrody-

namic (MHD) limit, one can use the vanishing of the electric field in the
comoving fluid rest frame to eliminate it from occuring explicitly in the equa-
tions. It is customary to express the resulting set of equations in a pre-
chosen fixed Lorentzian frame of reference, splitting them in temporal and
spatial contributions. A numerically convenient conservation form is then ob-
tained, specifically ∂t U + ∂i Fi = 0, where the conserved variables are taken as
U = (D, S, τ, B) and fluxes per spatial direction are indicated with Fi . Sum-
mation over indices i ∈ 1, 2, 3 is understood for 3D cases. The first variable is
the fluid density D = Γ ρ, where ρ is the proper rest mass density (the density
in the local rest frame) indicating number density per unit volume multiplied
with particle rest mass. Due to the Lorentz contraction for a volume element,
the fluid density D in the pre-chosen Lorentz frame includes the extra Lorentz
−1/2
factor Γ = 1 − v2 where v = (v 1 , v 2 , v 3 ) is the spatial three-velocity.
Its spatial flux density in the i-th spatial direction is simply Dv i , making the
first equation
∂t D + ∂i (Dv i ) = 0. (1)
Writing all quantities in terms of the primitive variables (ρ, v, p, B), where
p is the kinetic pressure and B denotes the magnetic field three-vector, the
remaining equations are
∂t (ξ + B 2 )v j − (v · B)B j

Bj Bi

2 j i j i j i ij
+∂i (ξ + B )v v − − (v · B)(B v + v B ) + δ ptot = 0, (2)
Γ2
∂t τ + ∂i τ v i + ptot v i − (v · B)B i = 0,

(3)
j j i j i i

∂t B + ∂i B v − v B = 0, ∂i B = 0. (4)
In these equations, we introduced the total pressure ptot , containing both
kinetic and magnetic energy density, and we write the energy density τ in the
pre-chosen frame as
(v · B)2 B2
ptot = p + + , (5)
2 2Γ 2
B2 1
τ =ξ+ + (v 2 B 2 − (v · B)2 ) − p − D. (6)
2 2
In the case of an ideal equation of state with constant adiabatic index γ, the
still unspecified variable ξ becomes

γp
ξ = Γ2 ρ + ≡ Γ 2 w. (7)
γ−1
The quantity w (expression inside the brackets) represents the relativistic fluid
enthalpy which includes a rest mass energy contribution. Note in particular
that the magnetic field induction equation (4), and its vanishing spatial di-
vergence, are identical in form to their non-relativistic ideal MHD variants,
Grid-adaptive Simulations of Relativistic Flows 337
and that the relativistic hydrodynamic equations are contained as the limit of
vanishing magnetic field. We further adopted conventional scaling where the
light speed c = 1.
2 Numerical treatment in grid-adaptive settings

The equations (1-4) can be incorporated in any framework capable of adaptive
mesh refinement (AMR) computations for conservative systems. We extended
our AMRVAC code [5] to allow both special relativistic hydro- and magnetohy-
drodynamic simulations in 1D, 2D, or 3D configurations, augmenting its ear-
lier applicability to non-relativistic Euler or MHD problems. Our AMR code
can exploit OpenMP parallelism. The relativistic modules conveniently make
use of auxiliary variables, (Γ, p) for HD and (Γ, ξ) in the MHD case, which
are to be determined at any point from the conserved variables U only. Once
known, the conversion from conservative to primitive variables (ρ, v, p, B) is
straightforward. We implemented a suggested [6] Newton-Raphson strategy
to compute ξ as the zero of
1 B2 (S · B)2

γ − 1 (ξ − Γ D) 2
ξ− + B − + − τ − D = 0. (8)
γ Γ2 2 Γ2 ξ2
In this equation, one has the additional dependency Γ (S, B; ξ), as one can
write v = S + ξ −1 (S · B)B /(ξ + B 2 ). In the root finding algorithm, we first
find a bracket for the root consistent with an imposed minimal pressure p
and maximal lorentz factor Γmax . These latter quantities can be chosen in
accord with available arithmetic precision.
We use the robust Total Variation Diminishing Lax-Friedrichs discretiza-
tion, for which only the (local) maximal physical propagation speed is re-
quired. In the relativistic hydro case, this involves the sound speed cs readily
computed from c2s = γp/w, and the resulting maximal speed is for a pure
1D problem (| v | +cs )/(1 + cs | v |). For relativistic MHD applications, one
typically resorts to a root finding approach to compute the various (up to 7)
characteristic speeds. We use Laguerre’s method to compute both speed pairs
pertaining to slow and fast magneto-acoustic waves as the roots of a quartic
polynomial (see e.g. [3]). Since all 4 roots must lie in the interval ] − 1, +1[ but
can become notoriously close to each other and unity, it helps to transform [2]
to a variable µ = 1/(1 − λ) with well-seperated roots on ]0.5, +∞[. An addi-
tional Newton-Raphson polishing of the roots is implemented for those fairly
sporadic cases where the Laguerre method did not acquire the requested accu-
racy. All needed cell-center to cell-edge limited linear reconstructions to obtain
fluxes at cell interfaces are best performed on the variables (ρ, Γ v, p, B).
In the adaptive mesh strategy, conservation across the nested grid hier-
archy is ensured. A Richardson type extrapolation is used in the refinement
process where the result of a (low order) prediction for a future time step is
obtained twice: the order of coarsening and integrating is reversed from case to
case. We use a weighted average of selected conserved or auxiliary variables to

trigger automated refining or coarsening, but additional user-enforced refine-
ment criteria can be activated at will. In multi-dimensional relativistic MHD
applications, our strategy to handle ∂i B i = 0 is one where a diffusive source
term proportionate to potential local errors is added to the induction equation
only. On a Cartesian mesh, this source term for equation (4) is Cd ∆x2i ∂i (∂k B k )
with Cd ' 0.2 and it diffuses numerical divergence errors at a rate compliant
with the CFL condition for the explicit time integration method. In [5], the
same approach compared favorably to other common source term strategies for
non-relativistic MHD problems. Finally, in all necessary restriction and pro-
longation steps to interpolate spatio-temporally between different grid levels,
we can enforce either primitive or conservative variables to use in the typically
second order (limited) linear reconstructions.
3 Riemann problems and Relativistic hydro jets
We tested the code against a variety of Riemann problems available from

recent literature. In Fig. 1 the result at t = 0.4 is shown of a shock tube
problem from [1]. We start with left state (ρ, v 1 , v 2 , v 3 , p, B 1 , B 2 , B 3 ) =
(1, 0, 0, 0, 1000, 10, 7, 7) and right state (1, 0, 0, 0, 0.1, 10, 0.7, 0.7), and param-
eter γ = 5/3. We used up to 7 refinement levels, with a base level resolution
of 200. We find good agreement with the exact solution as documented in [4],
with from left to right: a left-going fast and slow rarefaction, a contact dis-
continuity, and a right-going slow and fast shock.
A demonstrative example of a relativistic hydro jet flow is shown in Fig. 2,
which considers an axisymmetric light jet, injected at Lorentz factor Γ = 7
from the lower (left) boundary. Earlier computations of such jets have been
presented by Martı́ et al. [9], and were recently used to validate an extension
of the Piecewise Parabolic Method to special relativistic flows by Mignone et
al. [10]. In our AMR simulation, we used 5 levels of (cartesian) properly nested
grids to get an effective resolution of 960×2880, on a domain spanning 15×45
jet beam radii. At time t = 80, a logarithmic plot of the rest mass density
Fig. 1. A 7 level AMR solution of a 1D relativistic MHD Riemann problem.

Grid-adaptive Simulations of Relativistic Flows 339
reproduces the main morphological features found in these jets. This picture
can be compared to Fig. 8 of Mignone et al. [10]: one finds good agreement on
the internal shocks in the jet beam, with qualitative reproduction of the fine
structure in the jet cocoon produced by shear flow instabilities at the contact
interface between jet and swept-up shocked ambient material.
Fig. 2. Resulting density structure of a Γ ' 7, beam Mach number vz /cs ' 11
relativistic hydro jet after 80 time units.
4 Magnetized Relativistic jets
Relativistic MHD simulations of impulsively injected jets typically assumed

simplified magnetic field configurations. The most straightforward extension
to MHD jets only considers pure toroidal magnetic fields in non-rotating ax-
isymmetric setting [7]. In Fig. 3, we reproduce both reference cases, which
have the numerical advantage that no measure for controlling ∂i B i is re-
quired. Both correspond to Γ ' 10 inflow and strong magnetization, and the
two cases mainly differ in the inlet ratio of material energy flux to Poynt-
ing flux. As shown in [7], the ‘nose cone’ topology then emerges when a high
Poynting flux characterizes the inlet conditions (left panel in Fig. 3). Jets with
kinetic energy dominated inflow (right panel) can develop extensive cocoons.
A more comprehensive parameter study of jet morphologies for purely toroidal
field configurations can be found in [8], along with selected cases with initially
homogeneous, pure poloidal magnetic fields, excluding rotation. We plan to
extend these results to more realistic magnetic jet configurations having both
toroidal and poloidal components, and investigate magnetized flows at higher
Fig. 3. Toroidally magnetized jets, with high (left) and low (right) Poynting flux
at the inlet.
Lorentz factors of varying plasma beta. We will apply both relativistic hy-
dro and magnetohydro simulations to the modeling of the enigmatic Gamma
Ray Burst phenomenon, where current models favor extremely relativistic
flow conditions. To this end, we will demonstrate robustness at high Lorentz
factors (order Γ ' 100) in multi-dimensional simulations.
Acknowledgement. We acknowledge financial support from the Netherlands Orga-
nization for Scientific Research, NWO grant 614.000.421, initial coding efforts by J.
Bergmans, and stimulating discussions with A. Achterberg.
References
1. D. Balsara, ApJSS 132 83 (2001).
2. J. Bergmans, et al., Lect. Not. Comp. Sc. Eng. 41 223 (2005).
3. L. Del Zanna, N. Bucciantini, and P. Londrillo, A&A 400 397 (2003).
4. B. Giacomazzo and L. Rezzolla, astro-ph gr-qc/0507102v1.
5. R. Keppens, et al., Comp. Phys. Comm. 153 317 (2003).
6. A.V. Koldoba, O.A. Kuznetsov, and G.V. Ustyugova, MNRAS 333 932 (2002).
7. S.S. Komissarov, MNRAS 308 1069 (1999).
8. T. Leismann, et al., A&A 436 503 (2005).
9. J.M. Martı́, et al., ApJ 479 151 (1997).
10. A. Mignone, T. Plewa, and G. Bodo, ApJSS 160 199 (2005).
Solution of Laminar Combusting Flows Using a
Parallel Implicit Adaptive Mesh Refinement
Algorithm
Scott A. Northrup and Clinton P. T. Groth
University of Toronto Institute for Aerospace Studies, Toronto, Ontario,Canada

northrup@utias.utoronto.ca,groth@utias.utoronto.ca
1 Introduction
Numerical methods have become an essential tool for investigating combus-
tion processes. In spite of the advances in solution algorithms and computer
hardware, the solution of combusting flows can still place severe demands on
available computational resources. In the previous work, Northrup and Groth
[1] and Gao and Groth [2] have developed a parallel adaptive mesh refinement
(AMR) algorithm that both reduces the overall problem size and the time to
calculate a solution for non-premixed laminar and turbulent combusting flows.
A preconditioned nonlinear multigrid algorithm with multi-stage semi-implicit
time marching scheme as a smoother was used to integrate the governing
partial differential equations. Although accurate solutions were obtained, for
viscous reacting flows, the approach is non-optimal and in many cases a large
number of multigrid cycles and solution residual evaluations were required to
obtain steady state solutions. A fully implicit treatment would seem neces-
sary to deal with the numerical stiffness of the reactive flow problems and
reduce the time to achieve a solution. However, the parallel implementation
of implicit methods with AMR can be challenging.
Recently, a parallel implicit AMR scheme has been proposed that combines
a block-based AMR approach with Newton’s method to solve the equations
governing two-dimensional inviscid flows of compressible gas dynamics [3]. A
Jacobian-free inexact Newton method is used to solve the system of nonlinear
equations arising from a finite-volume spatial discretization procedure and
a preconditioned generalized minimal residual (GMRES) method is used to
solve the resulting system of linear equations at each Newton step. An addi-
tive Schwarz preconditioner is used in combination with block-fill incomplete
lower-upper (BFILU) preconditioning to improve performance of the linear
iterative solver. The Schwarz preconditioning and block-based AMR readily
allow efficient scalable parallel implementations of the approach.
This work considers the extension and application of the preceding implicit
AMR scheme to the prediction of non-premixed laminar diffusion flames. De-

342 Scott A. Northrup and Clinton P. T. Groth
tails are provided concerning the mathematical models used to describe the
laminar flow combustion processes and a description is given of the parallel
implicit AMR finite-volume scheme, including details of the inexact Newton
and GMRES iterative methods, AMR strategy, and parallel implementation.
Numerical predictions are described for methane-air laminar diffusion flames.
The results demonstrate the validity of the parallel implicit AMR approach
and the efficiency of the method for resolving laminar flame structure.
2 Mathematical Modelling
The Navier-Stokes equations for a thermally perfect mixture, given by
∂ ∂
(ρ) + ∇ · (ρu) = 0 , (ρu) + ∇ · (ρuu + pI) = ∇ · τ − ρg , (1)
∂t ∂t

∂ p
[ρE] + ∇ · ρu(E + ) = ∇ · (u · τ − q) − ρg · u , (2)
∂t ρ
∂
(ρcs ) + ∇ · (ρcs u) = ∇ · (ρDs ∇cs ) + ρω˙s , (3)
∂t
are used herein to model the combustion of gaseous fuels and oxidizers. Equa-
tions (1)–(2) reflect the conservation of mass, momentum, and energy for the
reactive mixture, ρ is the mixture mass density, u is the mixture velocity, E
is the total specific energy of the mixture given by E = e + 12 |u|2 , e is the
specific internal energy, p is the mixture pressure, τ is the fluid stress tensor
for the mixture, q is the heat flux vector, and g is the acceleration due to
gravitational forces. Equation (3) is the species concentration equation for
species s, where cs is the species mass fraction, Ds is the diffusion coefficient,
and ω˙s is the time rate of change of the species concentration due to finite-
rate chemistry. It followsP from the caloric equation of state for a thermally
N
perfect mixture that e = s=1 cs hs − ρp , where hs is the species enthalpy, N
is the
PN number of species, and the ideal gas law for the mixture is given by
p = s=1 ρcs Rs T , where Rs is the species gas constant and T is the mixture
temperature.
Expressions for species and mixture thermodynamic and transport proper-
ties are required to complete the system of partial differential equations for the
reactive mixture. The empirical expressions complied by Gordon and McBride
[4] are used to specify the hs and the species specific heat, cps , entropy, ∆ss ,
viscosity, µs , and thermal conductivity, κs , as functions of temperature. Per-
fect mixture rules are used to determine the thermodynamic properties of the
reactive mixture and Wilke’s [5] and Mason and Saxena’s [6] mixture rules
are used to evaluate the mixture viscosity, µ, and thermal conductivity, κ, re-
spectively. The species diffusion coefficients, Ds are calculated by additionally
specifying a Schmidt number, Scs = µ/ρDs for each species.
Simplified one- and two-step reduced chemical reaction mechanisms are
used to model the chemical kinetics of methane-air combustion [7]. In the
one-step scheme, methane oxidation is represented as follows:
Parallel Implicit AMR for Laminar Combusting Flows 343
CH4 + 2 O2 → CO2 + 2 H2 O , (4)

whereas in the two-step approach, it is represented by
2 CH4 + 3 O2
2 CO + 4 H2 O , 2 CO + O2
2 CO2 . (5)
3 Parallel Implicit AMR Scheme

The governing equations for the reactive mixture are solved by applying a
finite-volume method in which the mixture conservation equations are inte-
grated over quadrilateral cells of a body-fitted multi-block quadrilateral mesh.
The finite-volume formulation applied to cell (i, j) can be expressed by
dUi,j 1 X
Γ =− (F · n ∆`)i,j,k + Si,j = Ri,j (U) , (6)
dt Ai,j
k
where U is the solution vector, F = (F − Fv , G − Gv ), F, G and Fv , Gv

are the inviscid and viscous fluxes respectively, S is the source term associ-
ated with the axisymmetric geometry, finite rate chemistry, and gravitational
forces, Ai,j is the area of cell (i, j), and ∆` and n are the length of the cell
face and unit vector normal to the cell face or edge, respectively. The local
preconditioning technique proposed by Weiss and Smith [8] is used to alleviate
numerical difficulties for low-Mach-number, nearly incompressible flows that
generally characterize laminar combusting flows. The inviscid component of
the numerical fluxes at the faces of each cell, (F, G) · n, are determined using
the least-squares piece-wise limited linear solution reconstruction procedure
of Barth [9] and Riemann solver based flux functions. An extension of the
approximate linearized Riemann solver of Roe [14] is used to account for mix-
ture composition. The viscous component of the cell face fluxes, (Fv , Gv ) · n,
are evaluated by employing a diamond-path reconstruction procedure [11].
A flexible block-based hierarchical data structure has been developed and
is used in conjunction with the finite-volume scheme described above to facili-
tate automatic solution-directed mesh adaptation on multi-block quadrilateral
mesh according to physics-based refinement criteria. The method allows for
anisotropic mesh refinement and is well suited to parallel implementation via
domain decomposition. Aspects of the block-based AMR algorithm for multi-
block quadrilateral mesh are described in the recent work by Sachdev et al.
[12].
For steady flows, Newton’s method is used to solve the large coupled non-
linear system of algebraic equations R(U) = 0 and thereby obtain time-
invariant solutions of Eq. (6). Starting with an initial estimate, U0 , succes-
sively improved estimates for the solution are obtained by solving Jn ∆Un =
−R(Un ) at each step, n, of the Newton method, where J = ∂R/∂U is the
residual Jacobian. The improved approximation for the solution is given by
Un+1 = Un + ∆Un . The iterative procedure is repeated until an appro-
priate norm of the solution residual is sufficiently small, i.e., ||R(Un )||2 <
||R(U0 )||2 where is some small parameter (typically, ≈ 10−8 –10−10 ).
Each step of Newton’s method requires the solution of a system of linear

equations of the form Jx = b. This system is large, sparse, and non-symmetric
and a restarted version of the GMRES algorithm [13], GMRES(m), is used for
its solution, where m is the number of steps after which the GMRES algorithm
is restarted. As discussed by Dembo et al. [14], an exact solution of the linear
system is not necessary for rapid convergence of Newton’s method and an
inexact approach is adopted here in which the inner GMRES iterations are
not fully converged at each Newton step.
Right preconditioning of the form (JM−1 )(Mx) = b is used here to
improve the performance of the GMRES method where M is the precon-
ditioning matrix. An additive Schwarz global preconditioner with variable
overlap [13, 15, 16] is used in conjunction with local BFILU preconditioners
for each sub-domain. This combination of preconditioning fits well with the
block-based AMR described above and is compatible with domain decomposi-
tion methods, readily enabling parallel implementation of the overall Newton
method. Rather efficient parallel implementations of implicit algorithms via
Schwarz preconditioning have been developed by other researchers and suc-
cessfully applied to the prediction of realistic flows [15, 16]. As the GMRES
method does not require the evaluation of the global Jacobian matrix, J, a
so-called “Jacobian-free” approach can be adopted in which matrix-vector
products JM−1 x are approximated using numerical differentiation.
In order to increase the radius of convergence and ensure global conver-
gence of the Newton method, an implicit Euler time-marching method with
switched evolution/relaxation (SER) is used as a startup procedure for the
Newton algorithm [17]. This provides a good estimate of the solution and, as
the time step is increased, Newton convergence can be achieved.
By design, the multi-block body-fitted AMR scheme and Newton algo-
rithm are well suited to parallel implementation on distributed-memory multi-
processor architectures. A parallel implementation of the block-based AMR
scheme has been developed using the MPI library. For homogeneous archi-
tectures with multiple processors all of equal speed, the self-similar nature
of the solution blocks is exploited and parallel implementation is carried out
via domain decomposition where the solution blocks are simply distributed
equally among the available processors, with more than one block permitted
on each processor. A Morton-ordered space filling curve is used to provide
nearest-neighbour ordering of blocks for more efficient load balancing [2].
4 Numerical Results
The proposed parallel implicit AMR algorithm is now applied to the solution
of an axisymmetric co-flow methane-air diffusion flame. In particular, the
solution of the steady laminar flame studied by Mohammed et al. [19] and
Day and Bell [20] is considered. The flame boundary and initial conditions
are the same as those used in the previous studies. The Mach and Reynolds
number based on the diluted methane fuel flow in the inlet are M = 0.0016
Parallel Implicit AMR for Laminar Combusting Flows 345
0.04 0.04 0.04 T (K)
2200
2160
2120
2080
2040
2000
1960
1920
1880
1840
1800
1760
1720
1680
1640
1600
1560
1520
1480
1440
1400
0.02 0.02 0.02 1360
1320
1280
1240
1200
1160
z (m)
z (m)
z (m)
1120
1080
1040
1000
960
920
880
840
800
760
720
680
640
600
560
520
480
0 0 0 440
0 0.01 0 0.01 0 0.01 400
r (m) r (m) r (m) 360

320
Initial Level - 96 Blocks(4x8) 3rd Level - 195 Blocks (4x8) 5th Level - 396 Blocks (4x8)
3072 Cells 6240 Cells 12672 Cells
Fig. 1. Comparison of predicted temperature distributions within the axisymmetric

methane-air laminar diffusion flame for three different levels of mesh refinement.
and Re = 169. Additional details concerning this case are given elsewhere
[19, 20, 1].
The predicted solution for the laminar diffusion flame obtained using the
parallel implicit AMR method is shown in Figure 1. The figure shows the
computed isotherms and flame structure obtained using a sequence of adap-
tively refined grids starting from the initial mesh (96 solution blocks with
3,072 cells) and proceeding to the final mesh after five levels of refinement
(396 blocks with 12,672 cells). A comparison of the results of Figure 1 with
those given in the previous studies [19, 20] reveals, that in spite of the inherent
simplifications used in the two-step reaction mechanism, the predicted flame
structure agrees very well with the previous work. The “wishbone” structure
of the high-temperature region is present and the computed lift-off and flame
heights are 0.05 cm and 3.3 cm, respectively, with a maximum centre-line
temperature of 2080 K. All of these values agree reasonably well with the
previously published results. The predicted value of the CO mass fraction
concentration at z = 3 cm along the centerline is cCO = 0.026 and, consid-
ering the limitations of the reduced chemistry mechanism used herein, is in
good agreement with the value of Mohammed et al. [19], who report a mass
fraction of cCO = 0.03 at this location.
Estimates of the parallel performance and scalability of the proposed al-
gorithm for the laminar diffusion flame problem are shown in Figure 2. The
figure depicts both the parallel speed-up, Sp = (t1 /tp )p, and parallel efficiency,
Ep = Sp /p, for a fixed-size problem as a function of the number of processors,
p, where tp is the processor time required to solve the problem using p proces-
sors and t1 is the time required to solve the problem using a single processor.
The performance curves are shown and compared to their idealized values for
up to 48 processors for the diffusion flame calculation on a mesh consisting of
96 4×8 cell solution blocks (3,072 cells). Although there are some inefficiencies
in the Schwarz preconditioning, it is quite evident that the speedup is nearly
linear and the efficiency remains above about 90% for up to 48 processors.
1
40
Parallel Efficiency (Ep)

Parallel Speedup (Sp)
0.8
30
0.6
20
0.4
Relative Speedup
10 Relative Efficiency 0.2
Ideal Speedup
Ideal Efficiency
0
10 20 30 40
Number of Processors (p)
Fig. 2. Parallel performance of the Newton algorithm for laminar diffusion flame
calculation on a mesh consisting of 96 4×8 cell solution blocks (3,072 cells) showing
the parallel speed-up, Sp = (t1 /tp )p, and efficiency, Ep .
5 Conclusions
A parallel implicit AMR scheme has been developed for solving laminar com-
busting flows. The combination of finite-volume discretization procedure, par-
allel block-based AMR strategy, low-Mach-number preconditioning, and New-
ton method solution procedure has resulted in a powerful computational tool
for predicting a wide range of steady laminar reactive flows.
References
1. Northrup, S. A. and Groth, C. P. T., Paper 2005–0547, AIAA, January 2005.
2. Gao, X. and Groth, C. P. T., Paper 2006-1448, AIAA, January 2006.
3. Groth, C. P. T. and Northrup, S. A., Paper 2005-5333, AIAA, June 2005.
4. Gordon, S. and McBride, B. J., Reference Publication 1311, NASA, 1994.
5. Wilke, C. R., J. Chem. Phys., Vol. 18, 1950, pp. 517–519.
6. Dixon-Lewis, G., Combustion Chemistry, edited by W. C. Gardiner, Springer-
Verlag, New York, 1984, pp. 21–126.
7. Westbrook, C. K. and Dryer, F. L., Combust. Sci. Tech., Vol. 27, 1981, pp. 31.
8. Weiss, J. M. and Smith, W. A., AIAA J., Vol. 33, No. 11, 1995, pp. 2050–2057.
9. Barth, T. J., Paper 93-0668, AIAA, January 1993.
10. Roe, P. L., J. Comput. Phys., Vol. 43, 1981, pp. 357–372.
11. Coirier, W. J. and Powell, K. G., AIAA J., Vol. 34, No. 5, May 1996, pp. 938.
12. Sachdev, J. S., Groth, C. P. T., and Gottlieb, J. J., Int. J. Comput. Fluid Dyn.,
Vol. 19, No. 2, 2005, pp. 157–175.
13. Saad, Y., Iterative Methods for Sparse Linear Systems, PWS, Boston, 1996.
14. Dembo, R. S., Eisenstat, S. C., and Steihaug, T., SIAM J. Numer. Anal.,
Vol. 19, No. 2, 1982, pp. 400–408.
15. Gropp, W. D., Kaushik, D. K., Keyes, D. E., and Smith, B. F., Parallel Com-
puting, Vol. 27, 2001, pp. 337–362.
16. Knoll, D. A. and Keyes, D. E., J. Comput. Phys., Vol. 193, 2004, pp. 357–397.
17. Mulder, W. A. and van Leer, B., J. Comput. Phys., Vol. 59, 1985, pp. 232–246.
18. Aftosmis, M. J., Berger, M. J., Murman, S. M., Paper 2004-1232, AIAA, 2004.
19. Mohammed, R. K., Tanoff, M. A., Smooke, M. D., A. M. Schaffer, A. M., and
Long, M. B., 27th Symposium on Combustion, Combustion Institute, Pitts-
burgh, 1998, pp. 693–702.
20. Day, M. S. and Bell, J. B., Combust. Theory Modelling, Vol. 4, No. 4, 2000,
pp. 535–556.
Towards Direct Numerical Simulation of a
Diffusion Flame-Shock Interaction with an
AMR Algorithm
G. Billet, J. Ryan and M. Borrel
ONERA BP 72 - 29, av. de la Division Leclerc 92322 Châtillon Cedex, FRANCE

Germain.Billet@onera.fr, Juliette.Ryan@onera.fr, Michel.Borrel@onera.fr
Key words: AMR platform, Reactive Flow Solver, Diffusion Flame

Summary. A reactive solver based on a time splitting method has been coupled
with an AMR platform. After having tested the solver potentiality, a diffusion
flame/shock interaction is presented with AMR.
1 Introduction
The simulation of unsteady reactive flows where very stiff phenomena occur
requires algorithms with specific properties. In particular, strong and rapid
fluctuations have to be captured precisely. It is therefore necessary to use a
method that is able to describe on a large range of Mach number the physical
and chemical events with good accuracy and stability if the wavelike oscil-
latory behavior is to be taken into account. In this paper we present three
simulations realized with the reactive solver MAJIC [1]. It is a second order
accuracy solver based on time splitting, MUSCL approximation and one step
time scheme on uniform Cartesian meshes.
The hyperbolic terms of the Navier-Stokes equations are generally the most
difficult to solve because of the presence of nonlinear terms. For these terms,
adaptive limiters have been proposed in [1] [2] [3] associated with AUSM+
flux splitting [4]. This approach is compared to ENO and Jameson’s schemes
in the case of freely decaying isotropic turbulence.
The parabolic terms have also to be discretized correctly. A centered
second-order scheme in space is used. In order to study the parabolic op-
erator, first the diffusion of two species and generation of acoustic waves after
one time step is studied and compared with the analytical solution. Then as
we proceed in time, a quasi-incompressible convection-diffusion flow which
couples the hyperbolic and parabolic operators is simulated.

348 G. Billet, J. Ryan and M. Borrel
Finally, in order to show that the code is able to work also in the do-
main of compressible reactive flows, we associate a source operator with the
hyperbolic-parabolic operators. A reactive supersonic flow where a H2 -air cir-
cular diffusion flame interacts with a shock is studied. For this last case,
MAJIC has been coupled with an adaptive mesh refinement (AMR) plat-
form developed at ONERA ([5],[6]). The AMR method is the one presented
by Berger, Oliger, Colella, Quirk and many other authors (see the website
http://seesar.lbl.gov/AMR/) and adapted to finite differences (see [7]).
2 Freely Decaying Isotropic turbulence

For this first case, the Euler equations are solved in order to evaluate the
scheme numerical dissipation. The simulations are performed in a cube of edge
length 2π with 643 grid points uniformly distributed. The boundary conditions
are periodic in the three directions. A quasi-incompressible flow is retained.
The initial conditions are those of case 1 in [8]. The CF L number is 0.05. Fig.
1 shows the vorticity isosurface obtained with the proposed 1-step/AUSM+ -
triad algorithm at t = 10 at a chosen value similar to the one in [8]. The
1-step/AUSM+ -triad gives a finer representation of these structures than the
multi-time steps schemes presented in [8], [2]. The isotropic Taylor micro-scale
λ, characteristic of the mean spatial extension of the velocity gradients [9],
is used to measure the resolved gradients with the numerical algorithms. At
t = 10, the respective values of λ based on the mesh size ∆x are summarized
in this table:
+ +
RK3/ENO Jameson RK3/WENO RK3/MENO 2-step/AUSM -triad 1-step/AUSM -triad centered/Smag.
5.8∆x 5.4∆x 5.0∆x 4.3∆x 4.0∆x 3.6∆x 3.3∆x
The value of λ obtained from 1-step/AUSM+ -triad is close to the smallest

value computed with a centered scheme and the Smagorinsky model adapted
to this simulation for obtaining the theoritical -5/3 slope.
3 Acoustic waves from a species diffusion

3.1 Analytical formulation
A H2 top-hat shape located in the middle of a 1-D domain diffuses in a O2

flow. At t0 = 0, temperature T and pressure p are uniform and the flow is at
rest (u = 0). At t0 , if Fick’s law is used with Lewis numbers equal to 1, the
Navier-Stokes equations reduce to
ρt = u t = 0
(ρY1 )t = −(ρY2 )t = −( CλP Y1x )x
P2
(ρe)t = −( CλP i=1 hi Yix )x
where mass fractions Y1 = YH2 and Y2 = YO2 = 1−Y1 . In the above equations,
ρ denotes the density, λ the thermal conductivity, CP the heat capacity at
Diffusion Flame-Shock interaction with an AMR algorithm 349
Fig. 1. Iso-surfaces of a constant value of vorticity at t = 10.
constant pressure, hi the enthalpy per unit mass of the ith species and e the
internal energy per unit mass. If we assume
CP = CP1 Y1 + CP2 Y2 and λ = λ1 X1 + λ2 X2 , (Xi is the mole fraction of
the ith species) with
CP i = ani T n + ... + a0i , λi = bm i T
m
+ ... + b0i (n, m) ∈ N,
we are able to obtain the pressure space variation at time t1 = t0 + δt from
the space variations of the mass fraction of H2 at time t0 :
2
(W2 −W1 )
p(x, t1 ) = p(t0 ) + (t1 − t0 ) CP [Yc1 (W 2 −W1 )+W1 ]
G(x, t0 ) with
h i
W1 W2 (λ1 −λ2 ) λ
G(x, t0 ) = λY1xx + [Y1 (W2 −W1 )+W1 ]2 − CP (CP1 − CP2 ) Y1x Y1x .
c represents the speed of sound and Wi the molecular weight of the ith species.
3.2 Analytical - numerical comparison
When the molecular weights are equal, the diffusive interface does not pro-
duce acoustic waves, but as soon as the molecular weights are different, two
acoustic waves set up and propagate in the opposite directions. At t0 , we take
Y1 = 21 (1 + tanh C(xc − |x − x0 |)), x0 = 5mm, xc = 2.5mm, T (x) = 1000K,
p(x) = 1atm., δt = 10−9 s, ∆x = 25µm. At t1 , for two initial thicknesses
δY ≈ 100µm (C = 500) and δY ≈ 1mm (C = 50) of the diffusion layer de-
fined as δY = |xY1 =0.001 − xY1 =0.999 |, we plot p(x, t1 ) − p(t0 ) and the acoustic
pressure obtained by the simulation. Fig. 2 shows the analytical and numerical
curves. When the diffusion layer is spread off on only few cells, the maximum
numerical values are perceptibly weaker. As soon as the thickness encloses a
sufficiently number of cells (δY ≥ 10∆x), the curves are close and the code
reproduces well the phenomenon. The log-log plot of the error versus grid size
dp (exact)
dp (num)
3 dp (exact)
300
dp (num)
C = 500
C = 50
200 2
1
100
dp (pa)
dp (pa)
0
0
-1
-100
-2
-200
-3
0.0024 0.0025 0.0026 0.0015 0.002 0.0025 0.003
x (m) x (m)
Fig. 2. Comparison of the exact - numerical solutions (left: δY = 100µm, center:

δY = 1mm) and accuracy order of the parabolic operator for L2 -norm.
for the parabolic operator is shown on Fig. 2. For this last result, the thick-
ness remains constant (δY ≈ 1mm) and the grid size varies. The second-order
(s ≈ 2) is obtained for log10 ∆x ≤ −2.6 (δY ≥ 40∆x).
3.3 Acoustic field from a diffusion layer in a square domain
The previous case is extended to 2D non reactive flow. A circular H2 bubble is

centered in a 1cm square domain filled with air. ∆x = ∆y = 25µm. At t0 = 0,
p(x, y) = 1atm., T (x, y) = 1000K, u(x, y) = v(x, y) = 0. The mean radius
of the bubble is 2.8mm and the interface thickness is δY ≈ 250µm. Periodic
conditions are imposed on the boundaries. Soret effect, pressure-induced dif-
fusion and binary diffusion coefficients are taken into account by using library
EGlib [10]. Fig. 3 presents at three times the acoustic field set up by the initial
conditions. At t = 10−6 s, two single waves are visible. The first wave is con-
vected to the boundaries and the second wave focuses on the domain center.
At t = 5 10−6 s, the first wave coming from the neighboring domain begins
to cross the second wave. At t = 6 10−5 s, a complex acoustic field perfectly
symmetric with regard to both diagonals is enclosed in the square. To keep
this symmetry, a circular permutation of the space split hyperbolic operators
is applied at each time step.
4 Diffusion H2 -Air flame/shock wave interaction

4.1 Initial and boundary conditions
At t0 = 0, a circular bubble filled with H2 −N2 (YH2 = 0.233) species begins to

diffuse in the ambient air. The temperature is 300K in the bubble and 1500K
in the air. The pressure is uniform and equal to 1 atm. A circular diffusion
flame begins to appear and at t = 1.6 10−4 s, this flame hits a Mach 2 planar
shock. At this time, the size of the computational domain is 10mm x 5mm A
symmetry boundary condition is applied at y = 0 and Neumann conditions at
Diffusion Flame-Shock interaction with an AMR algorithm 351
0.01 0.01 0.01

t = 1 10-6 s t = 5 10-6 s t = 6 10-5 s
0.009 0.009 0.009
press press
press 418 -117
0.008 306 0.008 334 0.008 -167
202 250 -218
97 166 -268
0.007 -7 0.007 0.007
82 -319
-112 -2 -369
0.006 -216 0.006 -86 0.006 -420
-321 -170 -470
-425
0.005 0.005 0.005
y
y
0.004 0.004 0.004
0.003 0.003 0.003
0.002 0.002 0.002
0.001 0.001 0.001
0 0 0
0 0.005 0.01 0 0.005 0.01 0 0.005 0.01
x x x
Fig. 3. Acoustic field at three different times (left: t = 10−6 s, center: t = 5 10−6 s,
right: t = 6 10−5 s).
y = 5mm. Dirichlet conditions are imposed at x = 0 and NSCBC conditions

are used at x = 10mm. A detailed chemical mechanism is introduced [11]. The
pressure-induced diffusion, Soret effect and the binary diffusion coefficients are
taken into account but Dufour effect and the bulk viscosity are neglected.
4.2 Results
For this AMR computation, the finest mesh size is ∆x = ∆y = 25µm. Two
levels of embedded grids have been used (of ratio 2). Patches are defined
using a sensor based on the gradient of all primitive variables. During the
computation, the finest time step is 10−11 s ≤ δt ≤ 5 10−10 s. The results are
presented at a time when half of the bubble has crossed the shock. The density
distribution is shown on Fig. 4 where a typical shape of the shock deformed
by the flame is visible. At the same time, the distribution of some species are
plotted on Fig. 4 (for y ≥ 0). The shock has not a great influence on YH2O , but
the distribution of the other presented species changes strongly, particularly
for YOH , YHO2 , YH2O2 . These variations are due to the heat supply of the
shock that develops temporarily an endothermal process in the flame and
modifies the kinetic of the reactions and particularly the reactions that have
the shortest characteristic times.
5 Conclusion
The MAJIC solver effectiveness has been shown with different flows. For the
first flow, the results are compared with those of high order methods. For the
second flow, numerical and analytical solutions have been compared. Circular
permutation of the split-ups of the hyperbolic operator has shown to be effec-
tive in keeping the flow symmetry. The structure of this solver is well adapted
to an AMR strategy. MAJIC integration within the AMR platform has been
validated. The domain of research is thus enlarged since we may simulate
detailed reactive phenomena where high speed flows keep close to low speed
flows. Even if some work remains to be done to optimize the coding, for exam-
ple concerning high performance parallel computation, perspectives for com-
puting complex 3D flows are quite good. The physical models can be adapted
with the level of refinement : Euler/Navier-Stokes, mono/multi-species, empir-
ical laws/detailed transport mechanisms, global/detailed chemical scheme,...
0.003 ro 0.003 H2O 0.003 OH

1.8E-02 0.200 0.028
1.6E-02 0.182 0.025
1.5E-02 0.0025 0.164 0.0025 0.023
1.3E-02 0.146 0.020
1.2E-02 0.128 0.018
1.0E-02 0.110 0.015
0.002 8.7E-03 0.002 0.092 0.002 0.012
7.1E-03 0.074 0.010
5.6E-03 0.056 0.007
y
y
4.0E-03 0.038 0.005
0.0015 0.0015
0.001 0.001 0.001
0.0005 0.0005
0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006
x x x
0.003 H 0.003 HO2 0.003 H2O2

4.5E-03 1.6E-04 3.5E-05
4.1E-03 1.5E-04 3.2E-05
0.0025 3.7E-03 0.0025 1.3E-04 0.0025 2.9E-05
3.3E-03 1.2E-04 2.5E-05
2.9E-03 1.0E-04 2.2E-05
2.5E-03 9.0E-05 1.9E-05
0.002 2.1E-03 0.002 7.6E-05 0.002 1.6E-05
1.7E-03 6.2E-05 1.3E-05
1.3E-03 4.8E-05 9.4E-06
y
y
9.0E-04 3.4E-05 6.2E-06
0.0015 0.0015 2.0E-05 0.0015 3.0E-06
0.001 0.001 0.001
0.0005 0.0005 0.0005
0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006 0.003 0.004 0.005 0.006
x x x
Fig. 4. Density field and mass fraction zooms of H2O, OH, H, HO2 and H2O2
species when the flame crosses the shock.
References
1. G. Billet ’Improvement of convective concentration fluxes in a one step reactive

flow solver’, J. Comput. Phys. 204, (2005), pp319-352.
2. G. Billet and O. Louédin ’Adaptive limiters for improving the accuracy of the
MUSCL approach for unsteady flows’, J. Comput. Phys. 170, pp161-183 (2001).
3. G. Billet and R. Abgrall ’An adaptive shock-capturing algorithm for solving
unsteady reactive flows’, Comput. & Fluids, 32, pp1473-1495, (2003).
4. M.S. Liou ’A sequel to AUSM: AUSM+ ’,J. Comput. Phys., 129, p364, (1996).
5. J.-C. Jouhaud and M. Borrel. ’A hierarchical adaptive mesh refinement method:
application to 2D flows’, 3rd ECCOMAS CFD Conf., 1996.
6. J. Ryan and M. Borrel ’Adaptive Mesh Refinement : a coupling framework for
Direct Numerical Simulation of reacting gas flow’, ICCFD, Oxford, 2004.
7. M. Borrel, J. Ryan, G. Billet ’A generalized patch AMR platform that uses cell
centered or cell vertex solvers’, to appear in Eccomas CFD 2006.
8. E. Garnier, M. Mossi, P. Sagaut, P. Comte and M. Deville ’On the use of
shock-capturing schemes for large-eddy simulation’, J. Comput. Phys. 153,
273 (1999).
9. J. Jiménez, A. A. Wray, Ph. G. Saffman and R. S. Rogallo ’The structure of
intense vorticity in isotropic turbulence’, J. Fluid Mech. 255, 65 (1993).
10. A. Ern and V. Giovangigli ’EGlib: a general-purpuse for-
tran library for multicomponent transport property evaluation’,
www.cmap.polytechnique.fr/www.eglib/home.html
11. J.A. Miller, R.E. Mitchell, M.D. Smooke and R.J. Kee, ’Toward a comprehen-
sive chemical kinetic mechanism for the oxidation of acetylene: comparison of
model predictions with results from flame and shock tube experiments’, 19th
symposium on combustion, Combustion Institute , pp181-196, (1982).
Adaptive Multigrid Solutions of Thin Film
Flows over Topography
Y.C. Lee1 , H.M. Thompson2 , and P.H. Gaskell3
The University of Leeds, LS2 9JT, Leeds, U.K. y.c.lee@leeds.ac.uk,

h.m.thompson@leeds.ac.uk, p.h.gaskell@leeds.ac.uk
1 Introduction
The behaviour of thin liquid films whether forced to spread or deposited as a
distinct pattern on the surface of a substrate, is of enormous significance to
many manufacturing and biological processes. The topic of the present study is
the flow of continuous thin liquid films over surfaces containing topographical
features. In the electronics sector (displays, printed circuits, micro-devices,
sensors etc), for example, the industrial goal is often to minimise free surface
deviations from planarity either for aesthetic reasons or to ensure predictable
product properties [1].
These three-dimensional flows present extremely challenging practical de-
sign problems since free surface disturbances caused by even small-scale to-
pography can persist over length scales several orders of magnitude greater
than the actual size of the topography. The focus here is on the numerical
modelling of such flows, an area which is still in its infancy. The majority of
previous numerical studies have modelled the flow using the long wave, lu-
brication approximation which reduces the three-dimensional Navier-Stokes
equations to more tractable two-dimensional fourth order partial differential
equations for film thickness, pressure and (for evaporative flow) solvent con-
centration. Most have solved the resultant time-dependent lubrication equa-
tions with alternating-direction implicit (ADI) algorithms using alternating
sweeps in each direction so that only a banded system of equations needs to
be solved at each time step [2].
Recently, however, a Multigrid approach with adaptive time-stepping has
been developed as a more efficient alternative to ADI schemes, having already
been applied successfully to droplet spreading flows [3] and continuous film
flow with [4] or without evaporation [5]. This paper highlights the additional
benefits of adopting an automatic local grid refinement procedure within the
Multigrid algorithm which allows fine grids to be used only where they are
needed, i.e. near the topographies themselves, and much coarser grids to be
used over the rest of the substrate. Several previous numerical studies have

354 Y.C. Lee, H.M. Thompson, and P.H. Gaskell
successfully used mesh adaptivity in a wide variety of different contexts [6,7].

Here, the adaptive grid refinement strategy employed lies within the multilevel
adaptive technique (MLAT) of Brandt [8]. This approach is readily applicable
to the existing Multigrid solver and facilitates grid adaption using local trun-
cation error estimates coupled within the overall Full Approximation Storage
(FAS) Multigrid method.
2 Problem Specification and Mathematical Formulation

Figure 1 shows a sketch of the motion of a thin liquid film of thickness H(X, Y )
over a flat substrate, containing a small circular trench topography inclined
at an angle θ to the horizontal, with a constant volumetric flow Q0 per unit
width. The liquid is assumed Newtonian and incompressible, with constant
Free surface
z
H(X,Y)
y
RT Inclined plane
S0
LP
WP
!
Fig. 1. Gravity-driven thin film flow over a small circular trench.
density ρ, viscosity µ and surface tension σ, and its motion governed by the
Navier-Stokes and continuity equations, viz:

∂U
ρ + U .∇U = − ∇P + µ ∇2 U + ρ g, (1)
∂T
∇.U = 0, (2)
where U = (U, V, W ) and P are the fluid velocity and pressure respectively,
T is time and g = g(sin θ, 0, − cos θ) is the acceleration due to gravity.
Assuming that = H0 /L0 is small, where H0 and L0 are the character-
istic film thickness and in-plane length scale respectively, yields the following
lubrication equation for non-dimensional film thickness h,
∂ h3 ∂p ∂ h3 ∂p

∂h
= −2 + , (3)
∂t ∂x 3 ∂x ∂y 3 ∂y
and pressure, p:
6 2
p = − 3 ∇2 (h + s) + 61/3 N (h + s) , (4)
β β
Adaptive Multigrid Solutions of Thin Film Flows over Topography 355
1/3
σH0
where L0 = β 3ρg sin θ and N measures the influence of gravity on free
surface shape. Topographies are defined via arctangent functions which enable
the steepness of their sides to be controlled easily and the boundary conditions
are that the flow is fully developed upstream and downstream. Further details
are given in [4].
3 Numerical Method
3.1 Spatial Discretisation
The lubrication equations (3) and (4) are solved on a square computational
domain, (x, y) ∈ Ω = (0, 1) × (0, 1), with equal, uniform grid spacings in the
x and y directions, ∆ say, leading to the following discretised equations:
1 h3 h3

∂hi,j
= 2 |i+ 12 ,j (pi+1,j − pi,j ) − |i− 12 ,j (pi,j − pi−1,j ) +
∂t ∆ 3 3
3 3

h h
| 1 (pi,j+1 − pi,j ) − | 1 (pi,j − pi,j−1 ) −
3 i,j+ 2 3 i,j− 2
3
h3

2 h
|i+ 12 ,j − |i− 12 ,j , (5)
∆ 3 3

6
pi,j + (hi+1,j + si+1,j ) + (hi−1,j + si−1,j ) + (hi,j+1 + si,j+1 ) +
∆2
√

3
(hi,j−1 + si,j−1 ) − 4(hi,j + si,j ) − 2 6N (hi,j + si,j ) = 0, (6)
3 3
for each, (i, j), in the computational domain. The terms, h3 |i± 12 ,j , h3 |i,j± 21
are the pre-factors obtained from linear interpolation between neighbouring
vertices. Time integration is performed using the standard, second-order ac-
curate Crank-Nicholson method and writing the right hand side of equation
(5) as F (hi,j , pi,j , hi±1,j , pi±1,j , hi,j±1 , pi,j±1 ) leads to the equation
∆tn+1
hn+1
i,j − F (hn+1 n+1 n+1 n+1 n+1
i,j , pi,j , hi±1,j , pi±1,j , hi,j±1 , pi,j±1 )
n+1
2
∆tn+1
= hni,j + F (hni,j , pni,j , hni±1,j , pni±1,j , hni,j±1 , pni,j±1 ), (7)
2
for which ∆tn+1 = tn+1 − tn , and the right hand sides are given in terms of
known values at the end if the nth time step, t = tn .
3.2 Adaptive Multigrid Solution Strategy
Equations (6) and (7) and are solved using a Full Approximation Storage
Multigrid algorithm within a Full Multigrid Cycle. Error reduction is per-
formed using a fixed number of pre- and post- Red-Black Gauss-Seidel New-
ton relaxations. On the coarsest grid level, the discretised equations are solved
using Newton iteration. Adaptive local mesh refinement is implemented us-
ing the Multi-Level Adaptive Technique (MLAT) first proposed by Brandt [8].
The τ -indicator is used to quantify errors since information is readily available
from the different grid levels. The difference in truncation error on successive
grids Gk and Gk−1 is approximated via a relative truncation error quantity,
τkk−1 , with large values of τkk−1 indicating regions of significant error between
successive grid levels and where correspondingly further grid refinement is
necessary.
A general approach to the discretisation at local refinement interfaces is
to conserve numerical flux at both the coarse and fine locally refined regions,
where the numerical flux across a control volume is defined via;
Z Z
F = u · n dS . (8)
S
For equation (7)
3 3
h ∂p h ∂p
u = uh = −2 , , (9)
3 ∂x 3 ∂y
and for equation (6)
6
u = up = − ∇(h + s) . (10)
β3
Further details of the adaptive approach used here are given in [9].
4 Results
The cases considered here are for the flow of thin water films of asymptotic
film thickness H0 = 100µm, viscosity 0.001Pa s, density ρ = 1000kg m−3
and surface tension σ = 0.07N m−1 down a substrate inclined at 30◦ to the
horizontal and with a constant inlet flow rate Q0 = 1.635 × 10−6 m2 s−1 [1].
These parameters yield a Capillary length Lc = 0.78mm and N = 0.122,
the latter value indicating that gravity has little influence on the free surface
shape. All results are obtained using an FMG V(4,2) cycle with a coarse grid
with 9x9 nodes in each direction and finest grid levels up to 513x513 (k = 6).
Figure 2 illustrates the flexibility of the adaptive approach by solving flow
over a square 39mm x 39mm domain containing three trenches each of char-
acteristic length 3.9mm, depth 10µm and with square, circular and a diamond
cross-sectional shapes indicated on the right hand plan view of the computa-
tional grids. The left hand side of the Figure shows the resultant free surface
Adaptive Multigrid Solutions of Thin Film Flows over Topography 357
disturbances, with the characteristic upstream capillary ridge and downstream

bow wave evident around each isolated topography [4]. It also clearly shows
Fig. 2. Flow past three small topographies.
that the greatest refinement is concentrated around the topographies them-

selves while allowing coarser 33x33 grids to be employed in the simple regions
of the flow with the smallest free surface disturbances. The efficiency of the
adaptive approach is highlighted for flow past a square trench of dimensions
0.78mm x 0.78mm x 10µm. Figure 3 shows the CPU time for (a single time
step?) needed for the adaptive Multigrid solver compared to that taken by
the non-adaptive solver. The grid levels used are k = 2 (33x33 nodes in each
direction), k = 3 (65x65),..., k = 6 (513x513) over a square 39mm x 39mm
substrate. Non-adaptive solutions are calculated on uniformly-spaced grids
whereas the adaptive ones use a global coarse 33x33 grid and refine adap-
tively if the residuals of the discretised equations on grid k are greater than
0.1τkk−1 . Figure 3 shows that the adaptive solutions are much more efficient
4
10
computational time, t(s)
3
10
2
10
adaptive
non-adaptive
1
10 2 3 4 5 6
grid level, k
Fig. 3. Comparison between CPU times for adaptive and non-adaptive solution of
flow past a square trench.
than the non-adaptive ones. For the case of a 513x513 grid (k = 6), for exam-
ple, the adaptive solution calculates a solution of equal accuracy in less than
10% of the CPU time.
5 Conclusion
Thin film and spreading flows over topographically heterogeneous substrates

are of enormous significance in a variety of biological, scientific and industrial
processes. A new Multigrid strategy with automatic mesh adaption for inves-
tigating thin film flows over arbitrary topography (peaks, trenches and their
combination) is introduced which offers substantial efficiency gains compared
to the non-adaptive approach. This offers the ability to create fine meshes
only in regions exhibiting large gradients in the free surface while allowing
coarser meshes to be used in the simpler flow regions.
The adaptive solver is an important development toward the production
of practically useful simulation tools that can solve real problems, for exam-
ple identifying the optimum secondary topography distribution or localised
heating sources needed as part of a manufacturing process to planarise free
surface flows over desirable topography [1].
References
1. Decre, M.J., Baret, J,C.: Gravity driven flows of viscous liquids over two-
dimensional topographies. J. Fluid Mech., 487, 147–166 (2003).
2. Schwartz, L.W., Eley, R.R.: Simulation of droplet motion on low energy and
heterogeneous surfaces. J. Coll. Int. Sci., 202, 173–188 (1998).
3. Gaskell, P.H., Jimack, P.K., Sellier, M., Thompson, H.M.: Efficient and accurate
time adaptive multigrid simulations of droplet spreading. Int J. Numer. Meth.
Fluids, 45, 1161–1186 (2004).
4. Gaskell, P.H., Jimack, P.K., Sellier, M., Thompson, H.M., Wilson, M.C.T.:
Gravity-driven flow of continuous thin liquid films on non-porous substrates
with topography. J. Fluid Mech., 509, 253–280 (2004).
5. Gaskell, P.H., Jimack, P.K., Sellier, M., Thompson, H.M.: Flow of evaporat-
ing, gravity-driven thin liquid films over topography. Phys. Fluids, 18, 031601
(2006).
6. Bai, D., Brandt, A.: Local Mesh Refinement Multilevel Techniques. SIAM J.
Sci. Stat. Comput., 8(2), 109–134 (1987).
7. Trangenstein, J.A., Kim, C.: Operator Splitting and Adaptive Mesh Refinement
for the Luo-Rudy I Model. J. Comput. Phy., 196, 645–79 (2004).
8. Brandt, A.: Multi-Level Adaptive Solutions to Boundary-Value Problems. Math.
Comp., 31, 333–90 (1977).
9. Lee, Y.C., Thompson, H.M., Gaskell, P.H.: An efficient adaptive multigrid algo-
rithm for predicting thin film flow on surfaces containing localised topographic
features. Computer and Fluids, to appear (2006).
A meshless solver for computing viscous flows
on Cartesian like grids
Munikrishna N.1 , Karthikeyan N.2 , and Balakrishnan N.3

1
Indian Institute of Science, Bangalore-12 kishan@aero.iisc.ernet.in
2
National Aerospace Laboratories, Bangalore-12 nkarthikeyan@css.nal.res.in
3
Indian Institute of Science, Bangalore-12 nbalak@aero.iisc.ernet.in
Summary. A meshless solver LSFD-U (Upwind-Least Squares Finite Difference)

has been evaluated for its performance on point distribution obtained from Carte-
sian like grids generated using a novel grid stitching strategy. The scheme performs
well for inviscid and laminar viscous flow computations. The procedure lacks ro-
bustness for turbulent flows, because of non-positivity of viscous discretization and
ill-conditioning of the geometric matrix involved in the LSFD-U procedure, partic-
ularly on highly stretched viscous grids.
1 Introduction
Cartesian mesh calculations provide great potential for future industrial cal-
culations. While the unstructured data based finite volume method has had
success in terms of solving inviscid flows on such grids, the performance of
the new class of meshless solvers (in particular LSFD-U), is yet to be tested.
It is needless to mention that the Cartesian grids provide a most natural way
of obtaining point distribution for meshless solvers. The present attempt is
to evaluate the performance of the meshless solver LSFD-U on a point dis-
tribution obtained from a Cartesian grid. It is remarked that the LSFD-U
procedure has already been successfully used for solving inviscid flows [1, 2].
Also, certain aspects of its viscous implementation along with positivity anal-
ysis are presented in reference [3]. Similar to Cartesian mesh calculations, one
of major hurdles in the extension of meshless solver for viscous flows, is the
non-positivity of the viscous discretization procedure.
Inspired by the success of the structured grid in computing viscous flows,
a novel grid stitching strategy has been proposed [4] for generating what we
refer to as “Cartesian like grids”. The details of this procedure are presented in
the next section. These grids have been successfully used for the computation
of laminar viscous flows [5, 6] within the framework of unstructured data
based cell center finite volume procedure. In this work, the point distribution
obtained from Cartesian like grids is used in conjunction with the LSFD-U
solver.

360 Munikrishna N., Karthikeyan N., and Balakrishnan N.
2 Cartesian like grid generation

It is not practical to simulate viscous flows using the conventional Cartesian
grid strategies employing unit aspect ratio cells. Use of high aspect ratio cells
and a positive viscous discretization procedure are the key to the success of
structured grids in solving viscous flows. Inspired by this success, recently a
novel grid stitching procedure wherein stretched Cartesian grids are employed,
has been proposed [4, 5]. While offering higher levels of automation, this pro-
cedure generates relatively smooth grids. The steps involved in generating
Cartesian like grids are: 1. background stretched Cartesian grid just enclosing
the body geometry is generated; body is superimposed 2. body cut cells are
recursively divided until a length scale associated with body geometry and
flow(say, the boundary layer thickness) are resolved. 3. Cartesian points close
to the wall are identified and moved on to the body; this completely avoids
small cut cells. 4. the grid is extended to farfield employing multiple blocks
allowing for one hanging node on a given edge falling on the block bound-
ary. This results in relatively coarse grid away from the body (as can be seen
in figure (7)), limiting the total number of cells. 5. the grid data required
for the solver is collected in a suitable format. The present strategy is also
amenable for solution adaptive refinement/derefinement. In fact, the refine-
ment or coarsening is carried out always on the background Cartesian mesh
and the point movement is effected after every adaptation step. While the
point movement strategy is critical for finite volume calculations, it may not
be significant for the LSFD-U procedure. In this work, we have used the point
movement strategy to enable comparison of results obtained using LSFD-U
and finite volume procedures.
3 LSFD-U
The compressible Reynolds averaged Navier-Stokes equation in divergence

form reads,
. NW
x
NE
x
. g .j NW NE
N
x x x .
Wx O
xE
f
.
J . x
SW
x
S
x
SE
. o NW N NE
x x
. Wx O
. xE
x
Wx x
. O
. x x x
. SW S SE
x
SW
x
S
x
SE
(a) Cluster for LSFD-U (b) Cartesian type (c) Hanging node type
Fig. 1. Cartesian grid support stencils

Viscous LSFD-U for Cartesian like grids 361
phi 1.2075 phi 1.01
1.14653 0.959
1.08556 0.908
1.02458 0.857
0.963612 0.806
0.902639 0.755
0.841666 0.704
0.780693 0.653
0.71972 0.602
0.658747 0.551
0.597774 0.5
0.536801 0.449
0.475828 0.398
0.414855 0.347
0.353882 0.296
0.292909 0.245
0.231936 0.194
0.170963 0.143
0.10999 0.092
0.0490175 0.041
(a) Point distribution (b) Equi φ contours: Variant 1a (c) variant 1a+limiter
Fig. 2. Solution of unsteady diffusion equation
Ut + (fi + fv )x + (gi + gv )y = 0, (1)

where, U is vector of Conserved variables, fi & gi are inviscid fluxes and
fv & gv represent the viscous fluxes. The inviscid flux derivatives at given
node ‘o’ are computed using linear least squares procedure [7], using cloud
of neighbours around it. Referring to Figure 1, a fictitious interface ‘J’ is
introduced at the mid-point of ray joining ‘oj’. Upwind fluxes along the co-
ordinate directions is computed at the fictitious interface. The upwind flux
difference is used in the least squares fit to estimate fxo and gyo . Suitable ‘k’
exact reconstruction procedure is employed for higher order accuracy [1].
Different possibilities for viscous flux discretization(VFD) are presented
in [3]. The positivity of these procedures as applied to discrete Laplacian
has also been studied. Quadratic least squares based procedure referred as
Variant 2(the most positive amongst the procedures considered) was used
for solving viscous flows [3]. To demonstrate the importance of the positivity
property, in this work, variant 1a [3] is employed for solving 2D unsteady
diffusion equation: φt = ∇2 φ with b.c.: φ(x, y) = x. Figure 2.a shows the
random point distribution used. Figure 2.b shows the equi φ contours obtained
from Variant 1a, wherein the loss of solution monotonicity is clearly seen. A
new viscous flux limiting strategy has been proposed in the context of finite
volume computations in the reference [8]. We have attempted to exploit the
proposed viscous flux limiter for preserving solution monotonicity. We can
(2)
∆φJ (2)
define, κ = (1) , where ∆φJ obtained from non-positive scheme (presently
∆φJ
(1) (2)
Variant 1a) and ∆φJ = φj − φo . If 0 < κ ≤ 1 then ∆φJ = ∆φJ else
(1)
∆φJ = ∆φJ . On employing this limiting strategy in conjunction with Variant
1a, the solution becomes monotonous as shown in Figure 2.c. Inspired by this
(1)
observation, ∆φJ is directly employed to obtain hessians associated with
viscous fluxes. It is referred to as Variant-1C and is employed for solving
viscous flows using point distribution obtained from Cartesian like grids.
In this work, we exploit the structure in data associated with any Carte-
sian mesh, almost all through the computational domain, except near the
wall boundary. In the regions where the structured data are available, it is
economical to employ simpler finite difference strategy, rather than using a

more complicated generalized finite difference strategy in the form of LSFD-
U. For this purpose, depending on the support stencil, the points are classified
into the following types: 1. general type, shown in Figure 1.a, and 2. Cartesian
type, shown in Figure 1.b. and 3. hanging node type as shown in Figure 1.c.
Suitable frameworks involving finite difference and 1D least squares procedure
are implemented for computing the solution gradients and flux derivatives.
Grid Configuration Np Number of Figure
Body Points Number
Grid1 NLR 7301 10805 465 7,8
Grid2: Level1 NACA 0012 6911 634 10
Grid3: Level1 NACA 0012 16920 648 13
Grid4: Level1 RAE 2822 49686 3149 16
Table 1: Grid details

4 Numerical Results
The LSFD-U solver along with point distribution obtained from Cartesian
like grids, is employed to solve inviscid and viscous flows. Numerical results
are compared in Figure 3 with available experimental or standard numerical
values. Also, the solutions are compared with the results obtained from un-
structured data based cell center finite volume solver, HIFUN-2D on the same
grids. van Leer and Roe schemes are used for interfacial convective fluxes for
inviscid and viscous flows, respectively. SGS implicit relaxation procedure [2]
is employed for convergence acceleration. Strong boundary treatment involv-
ing no slip condition and Riemann boundary condition are implemented at
the wall and far field points, respectively. Algebraic Baldwin-Lomax turbu-
lence model is implemented to compute eddy viscosity. Details of initial and
solution adaptive Cartesian like grids are shown in table 1. Details of the test
cases are described in table 2.
Case Airfoil Grid Re M∞ α 0 Validation Figure

with Number
Case1 NLR 7301 Grid1 – 0.185 6.1 Experiment [9] 7,8
9
Case2 NACA 0012 Grid2 500 0.8 10 Fortunato [10] 10,11
12
Case3 NACA 0012 Grid3 5000 0.5 0 Venkatakrishnan [11] 13,14
15
Case4 RAE 2822 Grid4 5.7 × 106 0.676 1.92 Experiment [12] 16, 17
18
Table 2: Test cases: Flow conditions.

Viscous LSFD-U for Cartesian like grids 363
The finite volume solutions compare well with benchmark results on the ini-
tial grids. Whereas the LSFD-U solver requires finer grids. Finer grids are
obtained using solution adaptive gridding strategy. The surface coefficient dis-
tribution obtained on adapted grids using LSFD-U solver compare well with
the standard results. It can be observed for laminar viscous flow solutions,
the oscillations in the wall data are reduced in case of finer grids. However,
unacceptable oscillations are seen in turbulent flow test case for both pressure
and skin friction coefficient distribution. The loss in robustness of the flow
solver on highly stretched turbulent flow grids is primarily attributed to non-
positive viscous discretization procedure and ill-conditioning of the geometric
matrix associated with the LSFD-U procedure.
5 Conclusions
The point distribution obtained from Cartesian like grids in conjunction with
meshless solver LSFD-U, is employed for simulating inviscid and viscous flows.
Partial success is achieved for turbulent flow computations involving highly
stretched viscous grids. In our view, the work presented has provided useful
direction to further research in meshless solvers as applied to viscous flows.
The present effort is towards evolving suitable grid generation and discretiza-
tion strategies for solving turbulent flows within LSFD-U framework.
Acknowledgment: The authors like to thank, VSSC, Thiruvananthapuram,

India for supporting this work under RESPOND program through the schemes
project ISRO 069.
References
1. Sridar D., Balakrishnan N., JCP, Vol. 189, pp 1-29 (2003).
2. Sridar D., Balakrishnan N., To appear in AIAA Journal.
3. Anup Ninawe, Munikrishna N. and Balakrishnan N., David Zing(ed), Compu-
tational Fluid Dynamics, Springer, 2004.
4. Partha Mondal, Munikrishna N. and Balakrishnan N., Proceedings of 8th AeSI
CFD Symposium, Bangalore 12, (2005).
5. Partha Mondal, M.Sc.(Engg) Thesis, Dept. of Aerospace Engg., IISc, Bangalore
12, (2005).
6. Partha Mondal, Munikrishna N. and Balakrishnan N., Proceedings of 18th
NCAE, IIT Kharagpur, Kharagpur, (2004).
7. Gnana Kumar A., M.E. Thesis, Dept. of Aerospace Engg., IISc, Bangalore 12,
(2002).
8. Balakrishnan N. and Munikrishna N., NSSCAPDE’05, IIT Kanpur, Kanpur
(2005).
9. van den Berg B.,, Report No. NLR TR-79009 U, Dutch Aerospace Laboratory
(NLR), (1979).
10. Fortunato B., Vinicio M.: In Proceedings of 14t h ICNMFD, Lecture Notes in
Physics, S.M.Deshpande et. al.,(ed) ( Springer-Verlag Berlin 1995), p 259.
11. Venkatakrishnan V., Computers and Fluids, Vol. 18, (2), pp 191-204 (1990).
12. Cook P.H., Report No. AR-138, AGARD, (May, 1979).
8
cart−lsfdu
cart−fv−main−airfoil
7
cart−fv−flap
experiment
6
p
−C
3
−1
−2
−0.5 0 0.5 1 1.5
x/c
(a) Grid1 (b) Grid in the slotted region (c) −Cp distribution
Inviscid flow: NLR 7301, M∞ =0.185, α=6o
1.5 1
cart−lsfdu−level1
1 Fortunato
cart−fv−level0
0.5 0.5
−Cp
Cf
−0.5
0
Fortunato
−1
cart−fv−level0
−1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 −0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
(d) Grid2 (e) −Cp distribution (f) Cf distribution

Laminar flow: NACA 0012, M∞ =0.8, Re∞ = 500, α=10o
0.6
0.4 cart−lsfdu−level0
0.1
venkatakrishnan
0.2
cart−fv−level0
−0.2
−Cp
Cf
0
−0.4
−0.6
−0.8 venkatakrishnan
cart−fv−level0
−0.1
−1
−1.2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
(g) Grid3(near Trailing edge) (h) −Cp distribution (i) Cf distribution

Laminar flow: NACA 0012, M∞ =0.5, Re∞ = 5000, α=0o
−3
1.5 x 10
6
5 cart−lsfdu−level0
1
Exeriment
0.5
3
−Cp
0
Cf
1
−0.5
0
Exeriment
−1 cart−fv−level0
−1
−1.5 −2
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
(j) Grid4 (k) −Cp distribution (l) Cf distribution

Turbulent flow: RAE 2822, M∞ =0.676, Re∞ = 5.7x106 , α=1.92o
Fig. 3. Inviscid and laminar viscous flow solutions
Immersed boundary technique for compressible
flow simulations on semi-structured grids
M. D. de Tullio1 , P. De Palma1 , G. Iaccarino2 , G. Pascazio1 , and M.

Napolitano1
1
DIMeG & CEMeC, Politecnico di Bari, via Re David 200, 70125, Bari, Italy
2
CTR, Mechanical Eng. Dept., Stanford University, Stanford, California 94305
1 Introduction
The Immersed Boundary (IB) method simplifies the grid generation process
for the simulation of flows with complex and/or moving solid boundaries by
avoiding the need for a body-fitted mesh. The IB technique was originally
developed for incompressible flows [1, 2, 3] using Cartesian grids. Recently,
some of the authors have extended the IB technique to compressible flows [4]
using the preconditioned Navier–Stokes equations, which allow one to provide
accurate and efficient solutions for a wide range of the Mach number. To date,
IB methods employ structured grids, which allow only limited control on the
distribution of the grid points in the computational domain; in fact, clustering
of grid points is needed close to solid boundaries in order to describe its ge-
ometry accurately and, since mesh lines run through the entire computational
domain, a high concentration of grid points is obtained also in regions away
from the solid walls, where flow gradients are usually small.
In order to cope with this problem, a flexible local grid refinement technique
is to be employed, increasing the mesh resolution near the body. This work
extends the method of [4] to the solution of three-dimensional flows using a
semi-structured grid obtained by coarsening a uniformly fine mesh far from
high-gradient flow regions, such as boundary layers and shocks.
2 Governing equations and numerical method
In this work, the Reynolds Averaged Navier–Stokes (RANS) equations, writ-

ten in terms of Favre mass-averaged quantities, are solved in conjunction with
the low-Reynolds number k−ω turbulence model [5]. Such equations are given
in compact form as:
∂Qv ∂Q ∂E ∂F ∂G ∂Ev ∂Fv ∂Gv
Γ + + + + − − − = D, (1)
∂τ ∂t ∂x ∂y ∂z ∂x ∂y ∂z

366 de Tullio et al.
where Q is the conservative variable vector, E, F , G and Ev , Fv , Gv indicate

the inviscid and viscous fluxes, respectively, and D is the vector of the source
terms. A pseudo-time derivative for the primitive variable vector Qv , which
is related to Q by a Jacobian matrix, has been added to the left-hand-side of
equation (1) in order to use a time marching approach for both steady and
unsteady problems. The preconditioning matrix, Γ, proposed in [6] is used to
premultiply the pseudo-time derivative in order to enhance the efficiency of
the solution procedure at all values of the Mach number. Equation (1) is dis-
cretized by an Euler implicit scheme in the pseudo-time and the physical-time
derivative is approximated by a second-order-accurate three-point backward
difference. The resulting algebraic system is diagonalized according to the pro-
cedure of Pulliam and Chaussee [7] and solved by a BiCGStab method [8].
A collocated cell-centred finite volume space discretization is used. The con-
vective terms are discretized using either an upwind flux-difference-splitting
scheme with first-, second- or third-order accuracy, or a second-order-accurate
centred scheme, for low Reynolds number flows. When computing flows with
shocks, a total variation diminishing approach is employed using the min-
mod limiter function and either a second- or a third-order accurate upwind
scheme. The viscous terms are discretized by second-order-accurate centred
differences.
2.1 Data structure and semi-structured grid generation
Local grid refinement allows for efficient clustering of cells close to the im-
mersed boundary. The basic idea was recently introduced by Durbin and Iac-
carino [9] for a finite difference discretization and extended to a finite volume
formulation by Iaccarino et al. [10]. The following description of the algorithm
refers to two dimensions, the extension to three-dimensions being straightfor-
ward. An auxiliary structured grid is employed to handle the data structure
of the semi-structured locally refined grid (shown in figure 1). The auxiliary
grid covers the whole computational domain employing the finest mesh size
on the semi-structured grid. Therefore, each cell of the semi-structured grid is
bounded by the lines passing through the vertices (i, j) and (i + ∆i, j + ∆j),
see figure 1, where the indeces i = 0, . . . , Ni and j = 0, . . . , Nj refer to the
auxiliary (finest) grid and ∆iε1, ∆jε1 depend on (i, j). Therefore, having an
auxiliary grid with Ni × Nj cells, the N < Ni × Nj cells belonging to the
semi-structured grid are defined using the two couples of indices (i, j) and
(∆i, ∆j), with a total memory requirement of 4 N integers. In addition, an
array of integers, ID(i, j), is needed to store the correspondence between the
cells of the auxiliary and semi-structured grids. All of the cells of the for-
mer grid not employed in the latter one, namely, those included in the range
[i : i + ∆i − 1] and [j : j + ∆j − 1], are tagged using the same cell number. The
total storage required for allocating ID(i, j) is, therefore, Ni × Nj integers.
The connectivity information for each cell is retrieved by querying the array
ID(i, j).
Immersed boundary compressible flow simulations 367
N1
j+!j=const
(i+2,j+2)
W1
P E1
W2
(i,j)
j=const
S1 S2
i=const i+!i=const
Fig. 1. Locally refined grid showing a cell P and its neighbors.
Within the IB method, the generation of the semi-structured grid is carried

out by firstly creating the auxiliary (fine) grid and coarsening it in the regions
away from the immersed boundary. The advantage of this approach is that
all of the cell-tagging (ray-tracing) needed for using the IB technique can be
performed on a structured grid, taking full advantage of the alignment of the
cell centers and of the grid nodes. Moreover, it is possible to define regions of
the computational domain to be refined, such as wake and shock regions.
2.2 Immersed boundary treatment
The geometry under consideration, which is described by a closed curve in

two dimensions (a closed surface in three dimensions), is overlapped onto a
Cartesian (non uniform) grid. Using the ray tracing technique, the computa-
tional cells occupied entirely by the flow are tagged as fluid cells; those whose
centres lie within the immersed body are tagged as solid cells; the remaining
ones are finally tagged as interface cells. The application of the boundary con-
ditions at the immersed surface is treated explicitly, by assigning the values of
the variables at the interface cells. At solid cells, the flow variable φ (e.g., the
velocity components and, in the case of isothermal surfaces, the temperature)
is set to its wall value. At the interface cells, the nearby wall is modeled with
an off-wall boundary condition which consists of an interpolation of the flow
variables, using the computed values of the surrounding fluid cells and the
imposed values at the wall. At each interface cell it is possible to find Nnbr
neighbouring fluid cells and Nib intersections of the faces of each cell with the
immersed boundary, and the following interpolation formula is used:
Nnbr N
ib
X αi X βj
φint = φi + φj,wall , (2)
i
q j
q
where φj,wall is the value of the flow variable to be imposed at the immersed
surface, αi = 1/di and βj = 1/dj , di and dj being the distances between the
surrounding cell centers and the interface cell center and between the wall
intersections and the interface cell center, respectively, and
N
Xnbr Nib
X
q= αi + βj . (3)
i j
It can be shown that in the one-dimensional case, this procedure coincides with
the linear interpolation scheme used in [2, 4]. The pressure gradient along the
normal to the immersed surface is set to zero by assigning the corresponding
flow field value at the interface cell. In case of an adiabatic surface, also the
temperature gradient is set to zero.
3 Results
The unsteady two-dimensional low-Mach-number flow past a heated circular

cylinder has been chosen in order to validate both the unsteady terms and
the correct implementation of the energy equations, since experimental [11]
and numerical [12] investigations indicate that the temperature field has a
significant influence on the flow pattern, expecially when the ratio between
the cylinder wall temperature Tw and the free-stream one T∞ , T ∗ = Tw /T∞
exceeds 1.1. It has been found that, for a given Re∞ , the vortex shedding
frequency, f , and thus the Strouhal number St = f D/U∞ , decreases for in-
creasing values of T ∗ . The computational domain has the inlet and outlet
0.19
0.18
0.17
0.16
St
0.15
0.14
T*=1.0 exp.
T*=1.1 exp.
0.13 T*=1.5 exp.
T*=1.8 exp.
T*=1.0 present
T*=1.1 present
0.12 T=*1.5 present
T*=1.8 present
0.11
60 80 100 120 140 160
Re
Fig. 2. Strouhal number vs Reynolds number for the unsteady flow past a heated
circular cylinder.
Immersed boundary compressible flow simulations 369
boundaries located at xi = −10D and xo = 40D, and the far-field boundaries

located at yw = ±15D, the origin of the box coinciding with the centre of
the cylinder. Standard characteristic boundary conditions have been imposed
at the inlet and outlet points, whereas free-shear wall boundary conditions
are imposed at the far-field points. For such a low Re problem, centred dif-
ferences are employed also for the advection terms. Computations have been
performed using a semi-structured mesh with 41509 cells and 293647 faces.
The grid is highly refined at the cylinder surface, in order to solve the ther-
mal boundary layer, and inside a box surrounding the cylinder and the wake,
so as to obtain a satisfactory resolution of the vortex shedding phenomenon.
The auxiliary mesh is composed of 796 × 379 cells. The physical time step
has been chosen in order to have about 500 steps per period; about 250 inner
iterations are needed to reduce the unsteady residual to 10−6 at each time
step. Figure 2 shows the computed values of the Strouhal number, St, for
Re∞ = 100, 120, 140 and T ∗ = 1.0, 1.1, 1.5, 1.8, as well as the experimental
results provided by Wang et al. [11] and Sabanca & Durst [12]: a very good
agreement is obtained; moreover the present results are comparable with those
obtained using the uniform auxiliary grid [4], having seven times more cells.
4
3
3
2
2
1
1
Y
0
Y
0
-1
-1
-2
-2
-3
-3
-4
-5 -4 -3 -2 -1 0 1 2 3 4 5
X
-4
-5 -4 -3 -2 -1 0 1 2 3 4 5
Fig. 4. Supersonic flow past a circular

Fig. 3. Local view of the grid for the cylinder at M∞ = 1.7: Mach number con-
case M∞ = 1.7. tours (∆M = 0.08).
The steady turbulent supersonic flow past a circular cylinder has been con-
sidered as a suitable test case to validate the method for compressible flows
with shocks. The flow with M∞ = 1.7 and Re∞ = 2 × 105 has been computed,
the inlet values of the turbulence kinetic energy and specific dissipation rate
2
being k/U∞ = 0.0009 and ωD/U∞ = 40, respectively. For the considered val-
ues of M∞ , a bow shock is formed upstream of the cylinder; the subsonic flow
behind the shock close to the cylinder accelerates along its surface forming
a supersonic-flow region, enveloping the subsonic recirculation region behind
the cylinder, and two symmetric tail shocks are formed at the end of the sep-
aration region. Results have been obtained using a rectangular computational
domain with dimensions [−10 D; 15 D] × [−10 D; 10 D], D being the diameter
of the cylinder centred at the origin. Standard characteristic boundary condi-
tions have been imposed at the inlet and outlet surfaces and free-shear wall
boundary conditions are imposed at the far-field boundaries. Numerical results
are obtained using the third-order-accurate upwind scheme. Simulations on a
coarse grid have been performed at first to locate the position of the shocks,
approximatively. Based on such a solution, a semi-structured grid with 75556
cells and 545699 faces has been obtained starting from a very fine auxiliary
grid with 1805 × 2159 cells and removing cells away from the shock regions. A
local view of such semi-structured grid is shown in figure 3. The Mach number
contours are given in figure 4, showing that a clear description of the shocks
and of the wake are obtained, thanks to the local grid refinement in those
regions. The computed separation angle, measured clockwise from the leading
edge, is equal to 113◦ and agrees well with the corresponding experimental
datum, 112◦ [13]. Moreover, the computed and experimental drag coefficients
are equal to 1.41 and 1.43, respectively. All numerical results agree very well
with the numerical solutions obtained using the same numerical method and
a body-fitted grid [4] and reasonably well with the experimental ones.
References
1. C. S. Peskin, J. Comput. Phys. 10 (1972) 252.
2. E. A. Fadlun, R. Verzicco, P. Orlandi, J. Mohd-Yosuf, J. Comput. Phys. 161
(2000) 35.
3. R. Mittal, G. Iaccarino, Annu. Rev. Fluid Mech. 37 (2005) 239.
4. P. D. Palma, M. de Tullio, G. Pascazio, M. Napolitano, Comput. Fluids 35
(2006) 693.
5. D. C. Wilcox, Turbulence models for CFD, 2nd Edition, DCW Industries, Inc.,
1998.
6. C. L. Merkle, in: M. Hafez, K. Oshima (Eds.), Computational Fluid Dynamics
Review 1995, John Wiley & Sons, pp. 419–436 (1995).
7. T. H. Pulliam, D. S. Chaussee, J. Comput. Phys. 39 (1981) 347.
8. H. van der Vorst, SIAM J.Sci.Statist.Comput. 13 (1992) 361.
9. P. A. Durbin, G. Iaccarino, J. Comput. Phys. 128 (2002) 110.
10. G. Iaccarino, G. Kalitzin, P. Moin, B. Khalighi, Paper AIAA-2004-0586.
11. A.-B. Wang, Z. Trávnı́ĉek, K.-C. Chia, Phys. Fluids 12 (2000) 1401.
12. M. Sabanca, F. Durst, Phys. Fluids 15 (2003) 1821.
13. V. A. Bashkin, A. V. Vaganov, I. V. Egorov, D. V. Ivanov, G. A. Ignatova,
Fluid Dynamics 37 (2002) 473.
Incompressible Flow Simulations Using Virtual
Boundary Method with New Direct Forcing
Terms Estimation
Hidetoshi Nishida1 and Kazuhiro Sasao1
Department of Mechanical and System Engineering, Kyoto Institute of Technology

Matsugasaki, Sakyo-ku, Kyoto 606-8585, Japan nishida@kit.ac.jp
In this paper, a new external forcing terms estimation in the virtual boundary
method is proposed. In the usual virtual boundary method with feedback and
direct forcing terms estimations, the unphysical oscillations near the boundary
appear in the pressure field. In order to remove these unphysical oscillations,
the new direct forcing terms are added not only near the boundary but also
inside the boundary. In the incompressible flows past a circular cylinder and
a sphere, the pressure oscillations near the boundary can be removed and
the flow characteristics on the boundary are in very good agreement with the
reference results.
1 Introduction
For the practical flow simulations, the boundary fitted coordinates (BFC) is
usually adopted. This BFC approach has the high adaptability to the body
configuration. However, in the complicated flow geometry, it is difficult to
generate the computational grid. And in the BFC approach, it is necessary
that the governing equations are transformed from the physical plane to the
computational plane, so that the transformed governing equations have more
terms than the original equations in Cartesian coordinates. Therefore, the
computational effort is larger than the Cartesian coordinates. Then, in recent
years, Cartesian grid approach is highlighted again for the numerical flow
simulations. There are mainly three ways for Cartesian grid approach, i.e.,
immersed boundary method, cut cell method, and building cube method.
In this work, we consider the virtual boundary method [1]-[4] that belongs
to the immersed boundary method. In order to satisfy the velocity condi-
tions at the (virtual) boundary points, the virtual boundary method requires
the external forcing terms added to the momentum equations. These forcing
terms are usually estimated by two ways, that is, feedback [1] and direct [3]
procedures. Also, in both forcing terms estimations the grid points added the

372 Hidetoshi Nishida and Kazuhiro Sasao
forcing terms are restricted only near the boundary. Therefore, the unphysical
oscillations near the boundary appear in the pressure field. These unphysical
oscillations give the difficulty of pressure prediction on the (virtual) boundary.
In the present work, we propose the new direct forcing terms estimation in
order to remove these unphysical oscillations.
In order to validate the present new direct forcing terms estimation, the
direct numerical simulations (DNSs) of incompressible flow past a circular
cylinder and a sphere are considered.
2 Virtual Boundary Method

In the virtual boundary method, the velocity conditions on the boundary
are satisfied by adding the forcing terms to the momentum equations. Then
the incompressible Navier-Stokes equations can be written in nondimensional
form by
∂ui
= 0, (1)
∂xi
∂ui ∂ui ∂p 1 ∂ 2 ui
+ uj =− + + Gi , (2)
∂t ∂xj ∂xi Re ∂xi 2
where Re denotes the Reynolds number defined by Re = U L/ν. The last term
in Eq.(2), Gi , denotes the additional forcing term.
2.1 Numerical Method

The incompressible Navier-Stokes equations (1) and (2) are solved by the
variable order method of lines based on the fractional step approach on the
collocated grid system [5]. The fractional step approach is written by
ui ∗ = ui n + c∆tFi , (3)
∂p
ui n+1 = ui ∗ − c∆t , (4)
∂xi
where Fi is the convective and diffusive terms, c the parameter determined
by the time integration scheme, and ∆t the time increment, respectively. In
the method of lines approach, the spatial derivatives are discretized by the
appropriate scheme, so that the partial differential equations (PDEs) in space
and time are reduced to the system of ordinary differential equations (ODEs)
in time. The resulting ODEs are integrated by the 2nd order Runge-Kutta type
time integration scheme. In the spatial discretization, the convection terms are
approximated by the variable order proper convective scheme [5], because of
the consistency of the discrete continuity equation, the conservation property,
and the variable order of spatial accuracy. On the other hand, the pressure and
diffusion terms are discretized by the modified differential quadrature (MDQ)
method [5]. In this work, the order of spatial accuracy is specified by the 2nd
order. And the pressure equation is solved by the 2nd order SOR method.
Virtual Boundary Method with New Forcing Terms Estimation 373
Fig. 1. Additional forcing terms estimation
2.2 Usual Forcing Terms Estimation
In order to estimate the additional forcing terms in the governing equation,

Gi , there are mainly two ways, that is, the feedback and direct forcing term
estimations. In the feedback forcing terms shown in Fig.1(a), the forcing terms
at n time step are defined by
Z t
Gi n = α (Ūi − Ui )n dt + β(Ūi − Ui )n , (5)
0
where Ui denotes the specified velocity on the virtual boundary point and
Ūi denotes the interpolated velocity by using values on the surrounding grid
points. α and β are the parameters of the feedback forcing approach. In this
work, these parameters are set by α = −400000 and β = −600 [2]. And in the
direct forcing terms shown in Fig.1(b), the forcing terms are defined by
n
1 ∂ 2 ui Ūin+1 − ui n

∂ui ∂p
Gi n = u j + − + , (6)
∂xj ∂xi Re ∂xi 2 ∆t
where Ūin+1 denotes the interpolated velocity by linear interpolation. In both
additional forcing terms estimations, the grid points added forcing terms are
restricted near the boundary. Then, the pressure distributions near the bound-
ary show unphysical oscillations. It is difficult to obtain the precise pressure
distributions on the boundary.
2.3 New Forcing Terms Estimation
In order to remove aforementioned unphysical oscillations near the boundary,

we propose the new additional forcing terms estimation in the direct forcing
terms estimation. In the new additional forcing terms estimation, the forcing
terms are added not only near the boundary but also in the region surrounded
by the boundary shown in Fig.2. On the grid points near the boundary, the ad-
ditional forcing terms are estimated by the same estimation as the usual direct
forcing terms estimation, Eq.(6). In the region surrounded by the boundary,
the forcing terms are determined by satisfying the relation, Ūin+1 = U , where
(a) Usual direct forcing (b) Present direct forcing

Fig. 2. Grid points added forcing terms
U is the specified velocity and U = 0 in solid media. Then, the type of pres-
sure equation can be changed automatically from Poisson equation in the fluid
region to Laplace equation in the solid region.
In order to calculate the additional forcing terms, there are two ways in
the fractional step approach. One is to calculate in Eq.(3) and another is to
calculate in Eq.(4). The present forcing terms estimation gives the reasonable
solutions in both calculations. However, in the usual direct forcing estimations,
the reasonable solutions cannot be obtained for calculating forcing terms in
Eq.(4). Then, in this work, we adopt the calculation of forcing terms in Eq.(3).
3 Validation of Present Forcing Terms Estimation

First, we consider the flow past a circular cylinder with Re = 40. Figure 3
shows the pressure distributions near the boundary. It is clear that the re-
sults of usual feedback and direct forcing arise the unphysical oscillations at
neighboring points of boundary. On the other hand, the present direct forcing
gives the smooth pressure field. The pressure coefficient along the boundary
is shown in Fig.4. The notations VBM1, VBM2 and VBM3 denote the results
obtained by the usual feedback, direct forcing and present direct forcing, re-
spectively. In comparison with the result of Dennis et al. [7], the present direct
forcing shows the very smooth profile and the very good agreement with the
reference solution. Table 1 shows the comparison of characteristic quantities,
that is, drag coefficient, length of twin vortices, and pressure difference. We
can obtain the results comparable to the reference solutions quantitatively.
Figure 5 shows the pressure distributions near the boundary with Re =
200. The smooth pressure field can be obtained by the present direct forc-
ing. The time history of drag and lift coefficients is shown in Fig.6. And the
characteristic quantities are compared in Table 2. In comparison with the ref-
erence solution, the more reasonable solution can be obtained by the present
approach.
Next, the present approach is applied to the flows past a sphere. In Fig.7
the pressure profiles, that is, iso-surfaces of ∇2 p, are shown in the cases with
Re = 300 and 500. The released vortex ring from a sphere is clearly observed.
Virtual Boundary Method with New Forcing Terms Estimation 375
(a) Usual feedback forcing (b) Usual direct forcing (c) Present direct forcing
Fig. 3. Pressure distributions near a circular cylinder (Re = 40)
Table 1. Characteristic quantities

Re = 40 Cd Lv /D PE
VBM1 1.475 2.392 0.901
VBM2 1.596 2.457 0.950
VBM3 1.506 2.323 0.853
Fornberg [6] 1.498 2.240 0.800
Dennis et al.[7] 1.522 2.345 0.826
Fig. 4. Pressure coefficient distributions
(a) Usual feedback forcing (b) Usual direct forcing (c) Present direct forcing
Fig. 5. Pressure distributions near a circular cylinder (Re = 200)
Table 2. Characteristic quantities

Re = 200 Cd Clamp St
VBM1 1.437 0.734 0.201
VBM2 1.745 0.294 0.192
VBM3 1.318 0.679 0.199
Rosenfeld [8] 1.329 0.674 0.197
Fig. 6. Time history of Cd and Cl

(a) Re = 300 (b) Re = 500
Fig. 7. Iso-surface of ∇2 p
In this paper, a new direct forcing terms estimation is proposed in the virtual
boundary method. The present approach is applied to the numerical simula-
tions of flow past a circular cylinder and a sphere. In comparison with the
results obtained by usual feedback and direct forcing terms, the present ones
show that the smooth pressure field can be obtained so that the character-
istic quantities are in more excellent agreement with the reference solutions.
Also, since the pressure in solid media can be obtained precisely, the present
approach is very fruitful for fluid-structure interaction analysis.
This work was supported in part by a Grant-in-Aid for Scientific Research
(16560146) from the Japan Society for the Promotion of Science.
References
1. Goldstein, D., Handler, R., Sirovich, L.: Modeling a No-slip Flow Boundary with
an External Force Field. J. Comput. Phys., 105, 354–366 (1993)
2. Saiki, E.M., Biringen, S.: Numerical Simulation of a Cylinder in Uniform Flow:
Application of a Virtual Boundary Method. J. Comput. Phys., 123, 456–465
(1996)
3. Fadlun, E.A., Verzicco, R., Orlandi, P., Mohd-Yusof, J.: Combined Immersed-
Boundary Finite-Difference Methods for Three-Dimensional Complex Flow Sim-
ulations. J. Comput. Phys., 161, 35–60 (2000)
4. Nishida, H.: Cartesian Grid Approach with Virtual Boundary Method and Its
Applications. Notes on Numerical Fluid Mechanics, Springer, 78, 17–30 (2001)
5. Nishida, H., Satofuka, N.: A Variable Order Method of Lines: Accuracy, Conserva-
tion and Applications. Lecture Notes in Computational Science and Engineering,
Springer, 21, 167–174 (2002)
6. Fornberg, B.: A Numerical Study of Steady Viscous Flow Past a Circular Cylin-
der. J. Fluid Mech., 98-4, 819–869 (1980)
7. Dennis, S.C.R., Chang, G.Z.: Numerical Solution for Steady Flow Past a Circular
Cylinder at Reynolds Number up to 100. J. Fluid Mech., 42-3, 471–489 (1970)
8. Rosenfeld, M.: Grid Refinement Test of Time-Periodic Flows over Bluff Bodies.
Computers Fluids, 23-5, 693–709 (1994)
A Hybrid Building-Block and Gridless Method
for Computing Shock Waves
Hong Luo1 , Joseph D. Baum1 and Rainald Löhner2

1
SAIC, 1710 SAIC Dr., MS 2-6-9, McLean, VA 22102, USA, hong.luo@saic.com
2
George Mason University, Fairfax VA 22030, USA, rlohner@gmu.edu
1 Introduction
The numerical methods for the solution of the compressible Euler and Navier-
Stokes equations can be classified by the mesh they use as structured grid
methods, unstructured grid methods, Cartesian grid methods, and gridless
methods. Each of these methods, advocated, promoted, developed, and used
by their respective supporters, has its own advantages and disadvantages. The
structured grid methods have a disadvantage in mesh generation for complex
geometries. The main advantage of the unstructured grid methods is the ease
of grid generation for complex configurations. However, the computational
costs and memory requirements are generally higher than their structured
grid counterparts. The advantages of the Cartesian grid methods include ease
of grid generation, lower computational storage requirements, and significantly
less operational count per cell. However, the main challenge in using Cartesian
methods is how to deal with arbitrary boundaries, as the grids are not body-
aligned and the cells of a Cartesian mesh near the body can extend through
surfaces of boundaries. Lately, gridless methods came into focus. This class
of methods is essentially propelled by the fact that there exists a perceived
difficulty in generating volume filling grids for complex geometries in spite
of significant progress in the theory and practice of mesh generation over the
last decade. However, global conservation of mass, momentum, and energy for
these methods is not necessarily assured. Furthermore, these gridless meth-
ods are generally slower than their mesh-based counterparts. Recently, a new
approach [1] named building-cube method has been developed for large-scale
high resolution flow computations around complex geometries. In this method,
a flow field is divided into a number of cubes (squares in 2D) of various sizes.
Each cube is a computational sub-domain with a Cartesian mesh of equal
spacing and equal number of nodes. In this approach, the wall boundary is
defined by a staircase representation in order to keep the simplicity of the
algorithm and to minimize the memory requirement per node. Unfortunately,

it is not appropriate, convincing, and practical to use such a staircase repre-

sentation to approximate the curved boundaries.
The objective of the efforts presented in this paper is to develop a fast and
efficient numerical method for time-accurate computation of the compressible
Euler equations using a hybrid building-block and gridless method. The de-
velopment of such a numerical method is strongly motivated by the need to be
able to simulate blast and shock waves for complex geometries on a personal
computer platform in a reasonable time and accuracy. The idea is to combine
the advantages of both gridless and building-block Cartesian methods in an
attempt to develop a fast, low storage method for complex geometries. This
method uses gridless method to address the boundary or interface while a
building-block Cartesian grid method is used everywhere else. This method
can be regarded as a natural extension of the so-called hybrid Cartesian and
gridless methods [2, 3]to treat non-uniform grids. The developed method has
been used to compute a number of test cases to validate the accuracy and
efficiency of the method. The numerical results obtained indicate that the use
of this hybrid method leads to a significant increase in performance over its
unstructured grid counterpart.
2 Numerical Method
The basic idea behind the present hybrid method is to combine a building-
block method and a gridless method in an attempt to develop a fast, low
storage method for time-accurate computation of the unsteady Euler equa-
tions for complex geometries. In this method, body geometries are first defined
and the desired grid sizes in the flow field are specified via a background mesh
and/or sources. A baseline block (root block) covering the computational do-
main is then refined to create eight blocks of equal size. In turn, these blocks
are refined again until a suitable block size is achieved, which is determined by
the specified number of cells in x-, y-, and z-directions in a block and the pre-
scribed grid size. Each block is a sub-domain of the flow computation, where
a uniform-spacing Cartesian mesh is used. All blocks have the same number
of Cartesian cells in all three directions, so that the local computational res-
olution is determined by the block size. In order to minimize the numerical
errors between blocks, a smoothing of the size differences among blocks is
performed, only allowing the size difference between two adjacent blocks to
two. An unstructured data structure is used to manage the blocks and com-
municate with each other. Meanwhile, a triangulation of the body geometries
is performed using the advancing front method based on the representation
of the body geometries and the prescribed grid size. The center points of tri-
angles are chosen as a set of boundary points, which will be updated using
the gridless method. Note that such chosen points have a well-defined nor-
mal direction, therefore avoiding ambiguity of normal definitions for surface
singularities such as the tip of an airfoil trailing edge. Since a majority of
computational cells is solved using the building-block Cartesian method, the
A Hybrid Building-Block and Gridless Method 379
resulting method is very efficient in terms of both computational cost and

storage requirements. Since the gridless method is used to solve computa-
tional cells in the boundary and interface, the developed method can easily
be used for complex geometries without a need to generate a computational
mesh. Note that when the number of cells in all three directions is set to
one in each block, a classical octree-based Cartesian grid will be generated.
The underlying numerical method becomes the unstructured Cartesian grid
method, which can be used to serve a good reference in assessing the efficiency
of the present hybrid method.
The governing compressible Euler equations are solved using a gridless
method for all the gridless points and a standard cell-centered finite vol-
ume method for all the Cartesian cells. The dual least-squares approximation,
a variant of the least-squares approximation and able to introduce upwind-
type discretization, is used to compute the inviscid fluxes [3]. The numerical
fluxes at the interface are computed using the HLLC approximate Riemann
solver. A second order accuracy is achieved using the reconstruction scheme.
Barth-Jesperson limiter is used to suppress oscillations in the vicinity of dis-
continuities. The time integration is performed using an explicit four-stage
Runge-Kutta time-stepping scheme to advance the solution in time.
A. Steady two-dimensional oblique shock wave
A 2D shock wave reflecting from a rigid surface ((ρ, u, v, p)(0,y,t) =
(1, 2.9, 0, 1/1.4), (ρ, u, v, p)(x,1,t) = (1.69997, 2.61934, −0.50632, 1.52819)) is
considered to validate and verify the baseline building-block Cartesian grid
method, which is the foundation of the proposed hybrid method. The com-
putational domain is a rectangle of length 4.1 and height 1. Fig. 1 shows
the unstructured boundary surface mesh generated using a specified uniform
grid size of 0.05, the computed density contours in the flow field, and the
comparison of the computed density distribution with the exact solution at
y = 0.5. When the numbers of cells in x-, y-, and z-directions are specified as
82x20x20, a single block will be generated automatically. When these numbers
are halved, eight blocks will be created. When the numbers of cells in all three
directions are set to one, 32,800 (82x20x20) blocks will then be generated.
The CPU cost between the building-block Cartesian grid method and a
unstructured grid method [4] for 500 time steps is compared in Table 1, where
the number of faces for the building-block Cartesian grid and the number of
edges for the unstructured grid are also given. The unstructured tetrahedral
grid consists of 283,359 elements, 52,148 points, and 8,276 boundary points.
Note that computing cost for the building-block Cartesian grid method and
the unstructured grid method is proportional to the number of faces and edges,
respectively, where Riemann fluxes need to be evaluated for each method
and represent the most dominate CPU consuming operations. On the one
hand, when a single block is used, the building-block Cartesian grid method
2.8 2.8
2.6 Exact solution 2.6
Computation
2.4 2.4
2.2 2.2
2 2
Density
1.8 1.8
1.6 1.6
1.4 1.4
1.2 1.2
1 1
0.8 0.8
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
X-coordinates
Fig. 1. Unstructured boundary surface mesh used for the representation of gridless
points for stationary shock reflection problem (nface=16,548,nboun=8,276), com-
puted density contours in the flow field for stationary shock reflection problem, and
comparison of the computed density profile with the exact solution at y=0.5 for
stationary shock reflection problem.
Table 1. Computing cost analysis for the building-block Cartesian grid and
unstructured-grid methods.
Method FEFLO B-Block(32,800) B-Block(8) B-Block(1)
CPU (Second) 956 197 135 132
No of faces (edges) 343,780 102,080 102,080 102,080
Speed-up (Fluxes) 1 3.37 3.37 3.37
Speed-up (Others) 1 1.44 2.10 2.15
Speed-up (Total) 1 4.85 7.08 7.24
becomes the traditional Cartesian grid method. A total speed-up factor of

6.83 is observed for this case, where a speed-up factor of two is achieved due
to the saving in both indirect addressing and less operations (named other in
the table), representing the best scenario of the building-block Cartesian grid
method. On the other hand, when 32,800 blocks are used, the building-block
Cartesian grid method becomes the traditional unstructured Cartesian grid
method. A speed-up factor of 4.85 is still observed for this case mainly due
to less number of Riemann flux evaluations, representing the worst scenario
of the building-block Cartesian grid method, although the speed-up factor
due to the indirect addressing and less operations is reduced, as expected. In
addition, a larger time step can be used in the building-block Cartesian grid
method due the larger size of cells, leading to another speed-up factor of 1.8.
As a result, the building-block Cartesian grid method is 13 times faster using
a single block and 8 times faster using 32,800 blocks than its unstructured
grid counterpart.
A Hybrid Building-Block and Gridless Method 381
B. A blast wave past a multi-buildings

As an example of application, the developed method is used for the simula-
tion of a 50kg TNT blast wave past a multi-building configuration. Figure
2 shows the automatically generated building-blocks used for the represen-
tation of Cartesian cells and unstructured boundary surface mesh used for
the representation of gridless points based on the geometric definition of the
configuration and specified grid size, respectively, and the computed pressure
contours in the flow field at four different times obtained using the hybrid
method. The number of blocks in this case is 106, the numbers of Cartesian
cells in three directions inside a block are 13. The number of Cartesian cells
and the number of gridless points are 223,666 and 13,198, respectively. The
unstructured tetrahedral grid with the same geometric definition file and spec-
ified mesh size, used for our unstructured grid method, consists of 1,559,837
elements, 273,847 points, and 16,124 boundary points. An overall speed-up
factor about five in comparison with the unstructured grid method is observed
for this computation.
4 Conclusions
A hybrid Building-block Cartesian grid and gridless method has been devel-
oped for solving the unsteady compressible Euler equations. The developed
method combines the efficiency of a Cartesian method and the flexibility of a
gridless method for the complex geometries. The numerical results obtained
indicate that the use of this hybrid method leads to a significant improvement
in performance over its unstructured grid counterpart without compromising
solution accuracy, demonstrating the great potential and benefits of this hy-
brid method for the simulation of transient shock interaction problems. An
overall speed-up factor from six to twelve and a saving in storage requirements
up to one order of magnitude for a typical 3d simulation in comparison with
the unstructured grid method are obtained.
References
1. Nakahashi, K., and Kim, L.S.: Building-Cube Method for Large-scale, High Res-
olution Flow Computations. AIAA Paper 2004-0423 (2004)
2. Kirshman, J., and Liu, F.: Gridless Boundary Condition Treatment for a Non-
Body-Conforming Mesh. Journal of Computational Physics, Vol. 201, No. 1, pp.
119-147 (2004)
3. Luo, H., Baum, J.D., Löhner, R.: A Hybrid Cartesian Grid and Gridless Method
for Compressible Flows. Journal of Computational Physics, Vol. 214, pp. 618-632
(2006)
4. Luo, H., Baum, J.D., Löhner, R.: A Fast, Matrix-free Implicit Method for Com-
pressible Flows on Unstructured Grids. Journal of Computational Physics, Vol.
146, pp. 664-690 (1998)
Fig. 2. Automatically generated building-blocks used for the representation of

Cartesian cells (top left) and unstructured boundary surface mesh used for the
representation of gridless points (top right) based on the geometric definition of the
configuration and computed pressure contours (bottom) at different times for the
simulation of a 50kg TNT blast wave past a multi-building configuration using the
hybrid method.
A Residual estimator based adaptation
strategy for compressible flows
Ganesh N.1 , Nikhil V Shende2 , and Balakrishnan N.3

1
Indian Institute of Science, Bangalore-12 ganesh@aero.iisc.ernet.in
2
Indian Institute of Science, Bangalore-12 nikvijay@aero.iisc.ernet.in
3
Indian Institute of Science, Bangalore-12 nbalak@aero.iisc.ernet.in
Summary. Conventional error indicators can be effectively employed for grid re-
finement, but provide no information on error levels in the domain, and consequently
no “termination criterion”. Error estimators, on the other hand give reasonable es-
timate of error distribution and serve to decide suitable “termination criterion”.
In this paper, we discuss the design of a new residual–based error referred to as
“<–parameter”. The relation between the estimated error and the global errors are
investigated. Consequently, the <–parameter is used to design an effective AMR
strategy, which is applied successfully to inviscid flows in two and three dimensions.
1 Introduction
Research in adaptive CFD is directed towards obtaining accurate solutions

for real–life problems in shorter turnaround times. Solution adaptive grids
are useful to achieve accurate solutions with minimal computational effort.
At the heart of these AMR algorithms are sensors that can detect regions
of higher errors, which need to be selectively refined. While the commonly
used error indicators such as the divergence and curl of velocity can precisely
detect these high gradient regions, they cannot be used to decide if the grid
resolution is sufficient to resolve the flow features. On the other hand, error es-
timators provide reasonable estimates of error levels in the domain and hence
can provide a “termination criterion” for grid adaptation.
2 Theory of the residual error estimator
In this section, we develop from first principles a new aposteriori residual–

based error estimator for steady–state fluid flow problems. The estimator de-
pends on the numerical solution to provide reasonable estimates of error and
hence is aposteriori. Consider the governing conservation equations (Euler or

384 Ganesh N., Nikhil V Shende, and Balakrishnan N.
Navier–Stokes) expressed in an integral form over a finite volume Ωi bounded

by surface Γi ,
I
1
Ii [U ] = F · dΓ = 0 (1)
Ωi Γi
The discrete approximation to the above equation reads,
δi1 [u] = 0 (2)
In Eq.(1), U represents the exact solution and F = F (U ) represents the
flux vector, while in Eq.(2) δi represents a discrete approximation to Ii and
u is the numerical solution. The fundamental idea in the design of the <–
parameter is to exploit the imbalance that would arise if the exact operator,
I operated on the numerical solution, u. We then have,
Ii [u] = Ri1 [u] (3)

However, in working with a discretised domain, an estimate of this error can
be obtained using another discrete operator δi2 approximating I.
δi2 [u] = Ri1 [u] + Ri2 [u] (4)
If R1 ∼ O(hm ) and R2 ∼ O(hn ), the necessary condition for the <–
parameter to be an estimate of the local truncation error is m < n. In the
present procedure, a linear reconstruction (m=1) is employed to obtain the
solution while a quadratic reconstruction (n=2) with a three–point Gaussian
quadrature for flux integration, is used to estimate the errors.
3 Probing the <–parameter
A ‘good’ error estimator must decrease with grid refinement, be insensitive

to grid topology and be inexpensive to compute. We investigate the ability
of the proposed <–parameter to be a valid error estimator by applying it to
isentropic flow (M∞ =0.3) past a circular cylinder. The cylinder is one unit in
diameter and the farfield is 20 diameters. Calculations are performed on four
progressively refined structured grids, starting from a 40 × 60 grid (Fig.1(a)).
The computed pressure contours shown in Fig.1(b) clearly indicates the sym-
metry of the solution. A volume averaged norm [5] is used to characterise the
error distribution. The <–parameter decreases with refinement (Fig.1(c)),
with an asymptotic slope close to 1.8, owing to the regularity of the grids
while the global error (in entropy) decrease at a rate close to 3, which is in
agreement with published results [1]. However, in regions where limiters are
active, the <–parameter is rendered inconsistent [3]. Detailed numerical ex-
periments have also shown the insensitivity of the estimator to grid topology.
Residual estimator based adaptation strategy 385
(a) Structured (b) Pressure contours (c) Error plot

grid
Fig. 1. <–parameter studies on isentropic flow past cylinder
The ultimate aim of any grid adaptation strategy is to minimise the global
error. Therefore, some of the earlier works deal with estimating the global
error itself for the purpose of adaptation. On the other hand, in this work,
we estimate the local truncation error, in the form of the <–parameter. The
global error (GE) can be shown to be composed of two parts: Transmitted
error, which is error convected by the flow from “error sources” and Cell error,
which is error generated in each cell. The <–parameter is a cell–level imbalance
and can detect “error source”, which the global error cannot. Consequently,
an adaptive strategy based on the residual would be preferred as it results
in a faster decrease in error levels, and better error control. These arguments
are verified on the isentropic flow (M∞ = 0.6, AoA= 0o ) flow past a NACA
0012 airfoil. Starting from the same initial grid, a fixed percentage of cells
are refined at each level using the global error and the <–parameter. The
final grids resulting from this selective refinement are shown in Fig.2(a) and
Fig.2(b). Error levels resulting from residual based refinement are seen to
be lower than those from resulting from the global error based refinement
(Fig.2(c)).
(a) Final grid from (b) Final grid from < (c) Global error plot
GE based refinement based refinement
Fig. 2. Global and estimator error studies on isentropic flow past airfoil
4 New adaptive strategy

The success of any AMR algorithm is critically dependent on the criterion
employed for refinement/derefinement. While error indicators are employed
for refinement, we make use of the proposed <–parameter for derefinement in
the present work.
The refinement criterion depends on the user inputs of percentage cells

to be divided and the indicator to be employed. The value of the indicator
is computed everywhere in the domain and then normalised by its maximum
value. The cells with the larger normalised values of the indicator, starting
from the cell with the maximum value are progressively flagged until the user–
defined fraction is flagged for refinement.
The derefinement criterion is based on the <–parameter. The criterion

is based on the hypothesis that at any level of the mesh, the cells that are
never refined have the lowest values of the <–parameter. We therefore define a
threshold value of the parameter, <∗ = sup(<i ) ∀i ∈ L0 where L0 defines the
‘never-refined’ set of cells (or cells at level zero), where limiters are not opera-
tional. We state here without mathematical proof that < = X(h)h, where h is
a characteristic length (square/cube root of cell area/volume) and X(h) is re-
ferred to as ‘Dissipation function’ which shows a weak dependence on h. A cell
<∗
threshold length’ h∗ is computed as, h∗ = X ∗ , where X
∗
is computed using
Richardson extrapolation. The cell is marked for derefinement if k × hl ≤ h∗
where k = 2 for isotropic derefinement. The present strategy stands apart
from existing error–indicator based AMR strategies in that it makes use of an
error–estimator based derefinement strategy that is self–evolving.
5 Numerical Simulations
The proposed adaptive strategy is applied to three different test cases, rep-
resentative of different flow physics. All simulations are performed using the
in-house code HIFUN–2D/HIFUN–3D, which are based on cell–centered finite
volume approach. Time integration is achieved using an implicit Symmetric
Gauss Seidel (SGS) procedure [4] and Venkatakrishnan’s limiter [7] is used to
enforce monotonicity. VanLeer’s scheme [6] is used to calculate the interfacial
fluxes and the divergence of velocity is employed as an error indicator, unless
otherwise mentioned.
5.1 Inviscid flow past double ellipse (M∞ =8.15, AoA= 30o )
The hypersonic flow past a double ellipse configuration [2] is a challenging

problem to test the adaptive strategy as well as the flow solver. Three levels of
Residual estimator based adaptation strategy 387
refinement–derefinement were performed starting from an initial triangulated

grid. Initial,intermediate and final meshes (2020,8326 and 19147 cells) are
shown in Fig.3. The strong bow shock as well as the weak canopy shock have
been well resolved as seen from the mach contours (Fig.4).
Fig. 3. Initial, intermediate and final meshes for flow past double ellipse
Fig. 4. Mach contours on above meshes
5.2 Inviscid transonic flow past ONERA M6 wing (M∞ =0.84,

AoA=3.06o )
In attempts to extend the strategy toward real–life applications, the transonic

flow past a ONERA M6 wing is solved. Starting from an initial grid with
48000 cells, three levels of refinement–derefinement are performed. The initial
(48000 cells), intermediate and final grids (313111 cells) are shown in Fig.5.
The λ–shock on the upper surface of the wing is distinctly clear from the mach
contours in Fig.6.
6 Conclusion
A new residual–based error estimator referred to as the <–parameter has
been developed from from first principles. The consistency of the estimator
and its relation with the global error are investigated. It is shown that the
Fig. 5. Initial, intermediate and final meshes for flow past ONERA M6 wing
Fig. 6. Mach contours on above meshes
<–parameter is valid in continuous flow regions and can detect “sources” of

error. These observations are used to design a new AMR strategy which is
indicator–based for refinement and estimator–based for derefinement, which
is successfully applied to inviscid flows in two and three dimensions. Future
work would concentrate on building a purely estimator based strategy for
shock–free and incompressible flows as well as extending the present strategy
for time–dependent flows.
References
1. Bassi F., Rebay S., JCP, Vol.138, pp 251-285, 1997
2. Desideri. et.al.(Ed), Hypersonic flow for reentry problems, Vol.2, 1991
3. Ganesh N., Nikhil V Shende, Balakrishnan N., Proc. 7th AeSI CFD Symposium,
2004
4. Nikhil V Shende, Balakrishnan N., AIAA Jl., 42(9), Sep. 2004
5. Sun M., Takayama K., Error localization in solution-adaptive grid, JCP,
Vol.190, pp 346-350, 2003
6. van Leer B.,Report No. 82-30, NASA Langley Research Centre, Virginia.
7. Venkatakrishnan, V., “Convergence to Steady State Solutions of the Euler
Equations on Unstructured Grids with Limiters”, JCP, Vol. 118, 1995, pp 120–
130.
Anisotropic solution-adaptive technique
applied to simulations of steady and unsteady
compressible flows
Jerzy Majewski1 and Aristotelis N. Athanasiadis2

1
Warsaw University of Technology, Institute of Aeronautics and Applied
Mechanics, Nowowiejska 24, 00-665 Warsaw
jmajewsk@meil.pw.edu.pl
2
Computational Electrochemistry Group, Vrije Universiteit Brussel, Pleinlaan 2,
1050 Brussel, Belgium
1 Introduction
Grid adaptation is a very powerful tool for optimizing CFD calculations. Un-
fortunately the typical isotropic adaptation used for 3D flows may result in
excessive number of elements. The reason for this is that during refinement of
the grid around, e.g. the shock wave the grid cells becomes smaller uniformly
in all directions. It means that after splitting N cells in the direction normal
to the shock wave the number of the cells will be increased to N K 3 (K de-
scribes how many new edges are created after splitting an old one - if an edge
of a cell is split into two edges then K = 2). When the anisotropic adaptation
is used the grid is refined only in the direction normal to the shock wave and
the number of the cells would be increased to N K.
2 Anisotropic Adaptation
The adaptation algorithm presented in this paper relies on grid remeshing.
The remeshing is based on the generation of a new grid in the Riemann space
in which uniform spacing of the cells is assumed. The final spacing of the new
grid is a function of the metric field used for the definition of the Riemann
space. Detailed description of the anisotropic grid generation algorithms can
be found in [4] and [2].
The metric used for the definition of the Riemann space is obtained from
the solution field by means of the presented error estimator. The estimator
provides data in a form of a metric tensor which describes the directionality of
flow features. Several possibilities exist with regard to the choice of the error

390 Jerzy Majewski and Aristotelis N. Athanasiadis
estimator. In the present case the flow solver is using linear reconstruction
within each cell. Therefore it is convenient to use an estimate of the interpo-
lation error as the error indicator ([3], [4]).
Calculation of the metric. Assume that E denotes a grid cell inside which a
function u is being interpolated and xc is the center of E. Then after dropping
terms of higher order, the interpolation error for E can be estimated [4] as:
εE ≤ max x − xc )T |H|(x − xc

(1)
x∈E
where H is a Hessian of u. The Hessian is symmetric, therefore it has real

eigenvalues and can be decomposed as follows:
H = R · Λ · R−1 (2)
The matrix R consist of columns created from right eigenvectors of H and Λ
is a diagonal matrix consisting of corresponding eigenvalues λi . Consequently
|H| can be defined as:
|H| = R · |Λ| · R−1 (3)
The interpolation error in a given direction defined by the unit vector w, is
proportional to a constant C calculated as follows:
C = h2 wT |H|w (4)
where h denotes a length of a cell in the direction of unit vector w. In the
present approach w becomes a direction of a given edge and h becomes an edge
length. The construction of an optimal grid is based on the equidistribution of
the interpolation error. This is equivalent to assuming that for every edge ei ,
the constant Ci is equal to a global C. We can introduce now a scaled metric
M:
M = C −1 |H| (5)
In the last step the eigenvalues are limited to avoid unrealistic values:
1
λ̂i = (6)
max(hmin , min(hmax , hi ))
where values of hmin and hmax are provided by the user.
The metric M is subsequently calculated for every node of the old grid
and by means of interpolation is extended to form a continuous metric field.
Reconstruction of the Hessian. The key problem in the evaluation of the

metric tensor lays in the reconstruction of the Hessian ([5]). The approach
presented in this paper is based on a very robust and efficient Green formula.
The calculations are performed in two steps:
1. Calculation of ∇u at grid nodes with the Green formula
2. Calculation of H = ∇(∇u) at grid nodes with the Green formula
Anisotropic solution-adaptive technique 391
3 Steady and Unsteady Flow Adaptation Algorithm

Steady adaptation. A typical adaptation algorithm for steady flows can be
defined in the following way:
I. Generate initial grid G0 and set k ← 0
II. Solve equations on a grid Gk in order to obtain solution S k .
III. Check the criterion to end the adaptation loop (e.g., number of grid
nodes between Gk and Gk−1 does not vary significantly or k is equal to
kmax ). If this criterion is satisfied, adaptation loop is finished.
IV. Using error estimator find metric field M k for the solution S k .
V. Generate a new grid Gk+1 in a Riemann space using metric field M k .
VI. Interpolate solution S k on the new grid Gk+1
VII. return to point II setting k ← k + 1.
Unsteady Flow Adaptation Algorithm. The main problem arising in
adaptation of unsteady flows is due to the fact that it is next to impossible
to predict in advance the motion of flow features (e.g. shock waves). Thus
the adapted grid may not be refined accordingly to the future position of
the flow feature. In particular, the shock wave leaving the refined area may
drastically deteriorate the quality of the solution or even crash the code. The
only reasonable procedure is to track flow features on the grid which is not
yet adapted. Then a new grid should be created to accommodate all future
positions of the flow features along the adaptation time step [1].
For the adaptation of the unsteady flow the time step ∆t = [t0 , tmax ] is
divided into smaller steps ∆τ . The length of ∆τ is chosen according to the flow
problem, trying to keep balance between small values (too many interpolations
which can deteriorate conservation property of the scheme) and large ones (cell
size becomes too large). For each time step ∆τ , a quasi-steady adaptation loop
is performed. In this case however the definition of the metric field used for
the generation of the new grid is different. Namely, in the adaptation loop,
the metric is calculated as an intersection of many metric fields obtained from
snapshots of the solution taken during ∆τ step.
The algorithm:
I. Generate uniform grid G0
II. Set i ← 0 and τ0 ← t0
III. Define initial solution I0 for t0
IV. Set k ← 0
V. Start adaptation loop for time period [τ0 , τ0 + ∆τ ] using G0 as the grid
G0i
1. Interpolate initial solution Ii on a new grid Gki to obtain a solution
field Sik (τ0 )
2. Solve flow problem starting from Sik (τ0 ) and store solutions Sik (τ0 +
j∆τ /(l − 1)) where j = 0, ..., l − 1
3. If k is greater or equal k max stop adaptation loop and jump to point

VI
k
4. Using error estimator find metric field Mi,j for every stored solution
k
Si (τ0 + j∆τ /(l − 1))
5. Find a metric field Mik by intersecting metric fields Mi,j k
j =
0, ..., l − 1
6. Generate a new grid Gk+1 i using metric field Mik
7. Return to point 1 setting k ← k + 1.
max
VI. Save solution Sik (τ0 + ∆τ ) as initial solution Ii+1
VII. Set i ← i + 1 and τ0 ← τ0 + ∆τ
VIII. If τ0 is less than tmax jump to point IV.
4 Numerical Results
Adaptation of the MHD steady flow in a sinusoidal 3D channel. The
first testcase is the steady MHD flow through a 3D channel with a sinusoidal
bump, with supersonic inlet flow. The wall is fully conducting, hence van-
ishing normal B-field is imposed. Inside the channel the pattern of oblique
shock waves should be resolved. The initial grid and the grid after the 3rd
adaptation are shown on the Figure 1. The corresponding field of magnitude
of magnetic induction vector is shown on Figure 2. It is worth noting that sig-
nificant improvement of the solution was obtained while the number of points
was increased by a factor less than 1.6.
a)
b)
Fig. 1. Grids used for calculation of the MHD problem inside a channel with sinu-
soidal bump: (a) initial grid (23117 nodes), (b) grid after 3rd adaptation loop (36361
nodes)
Anisotropic solution-adaptive technique 393
4.30776
4.01185
3.71595
3.42004
3.12413
2.82822
2.53231
2.23641
1.9405
1.64459
1.34868
1.05277
a)
4.30776
4.01185
3.71595
3.42004
3.12413
2.82822
2.53231
2.23641
1.9405
1.64459
1.34868
1.05277
b)
Fig. 2. The field of the magnitude of the magnetic induction vector: (a) calculated
on the initial grid (b) calculated on the grid after 3rd adaptation loop
Adaptation of the unsteady Euler flow for a 2D channel with for-

ward facing step. The next testcase is the unsteady Euler flow past a forward
facing step in a channel (Mach at inlet equal 3.0). The solution of this problem
has been presented in [7]. It represents a flow with complex unsteady shock
wave interacting with the solid walls of the channel.
The initial grid with uniform spacing h = 0.02 is presented on Figure 2a.
The adaptation time step is equal to ∆τ = 0.02, which means that 150 adap-
tations were performed before reaching global time t = 3.0. For each time step
∆τ , three levels of adaptations were done gradually increasing the number of
nodes. The grid for the final level of adaptation was created with setting a
minimum cell size to 0.0015. The number of nodes for the finest grid for time
t = 3.0 consisted of 42193 nodes. A uniform grid with equivalent resolution
would roughly contain 1300000 nodes.
The final grid after the third adaptation and the density field for time
t = 3.0 can be seen on Figure 3. The improvement of the solution obtained
by means of adaptation is clearly visible - the shock waves are thin and the
slip lines starting from triple points have been fully recovered.
The solution of the Euler equations has been obtained using 2nd order
Cell Centered FVM with WENO reconstruction.
5 Conclusions
The presented anisotropic adaptation algorithm has been successfully applied
to steady and unsteady inviscid flows. The adaptation has improved the qual-
ity of the solution and recovered details of flow which were missing in the
solution based on uniform grid. The presented algorithm is capable to locally
predict the grid areas that should be adapted.
a)
6.2
5.8
5.4
4.9
4.5
4.1
3.6
3.2
2.8
2.3
1.9
1.5
1.0
0.6
0.2
b)
Fig. 3. Results after third adaptation for t = 3.0: (a) grid (42193 nodes), (b) density
field
Acknowledgements The work was done with cooperation with Von Karman
Institute and was supported by Belgian Federal Science Policy Office.
References
1. Alauzet, F., George, P.L.,Mohammadi, B., Frey, P., Borouchaki, H.: Transient
fixed point-based unstructured mesh adaptation, Int. J. Numer. Meth. Fluids,
43, 729–745 (2003)
2. Athanasiadis, A.N.: Three-Dimensional Hybrid Grid Generation with Applica-
tion to High Reynolds Number Viscous Flows, Ph.D. thesis, Von Karman Insti-
tute, Université Libre de Bruxelles (2004)
3. Castro-Dı́az, M.J., Hecht, F., Mohammadi, B., Pironneau, O.: Anisotropic Un-
structured Mesh Adaptation for Flow Simulation, International Journal for Nu-
merical Methods in Fluids, 25, 4, 475–491 (1997)
4. George, P.L., Borouchaki,H.: Delaunay Triangulation and Meshing - Application
to Finite Element, Hermes, ISBN 2-86601-692-0, Paris (1998)
5. Formaggia, L., Perotto, S.: Anisotropic Error Estimation for Finite Element
Methods, 31st Computational Fluid Dynamics, VKI Lecture Series 2000-05
(2000)
6. Majewski, J.: An anisotropic adaptation for simulation of compressible flows,
Mathematical Modelling and Analysis, 7, 127–134 (2002)
7. Woodward, P.R., Colella, P.: The numerical simulation of two-dimensional flows
with strong shocks, J. Comput. Phys., 54, 115–173 1984
Simulation of Flow around Wing Sections by
Building-Cube Method
Kazuhiro Nakahashi1 and Yota Sakurai2

1
Department of Aerospace Engineering, Tohoku University
naka@ad.mech.tohoku.ac.jp
2
Graduate student, Department of Aerospace Engineering, Tohoku University
sakurai@ad.mech.tohoku.ac.jp
1 Introduction
Flow around an airplane includes various scales; from micro-scale near the
surface to macro-scale around the airplane. The macro-scale flow basically
determines the surface pressure distribution around an airplane. The micro-
scale flow dynamics also has a significant contribution to such as the skin
friction, the flow separations and the aerodynamic noise generation. Current
CFD has become a powerful tool to accurately simulate macro-scale flows for
designing airplanes. The micro-scale flow simulation, however, is still difficult
due to the extremely large difference in the flow characteristic length.
In order to simulate the micro-scale flows in the macro-scale, an approach
based on Cartesian grid named ’Building-Cube Method (BCM)’ [1] was devel-
oped. The main strategy of the method is to use very fine and orderly grid in
the boundary layer regions in order to resolve micro-scale flows. The objective
of this study is to apply the BCM to simulate flows around airfoils/wing in
order to evaluate the capability of the method.
2 Building-Cube Method
2.1 Basic Scheme
In the BCM, a flow field is described as an assemblage of building blocks

of cuboids, named ’Cube’ as shown in Fig.1. Each cube is a sub-domain of
the flow computation and an equally-spaced Cartesian mesh is used in it.
All cubes have the same number of grid points so as to simplify the parallel
computations. The geometrical size of each cube is determined by adapting
to the geometry and the flow features so as to fit the grid spacing to the
local flow scale. The use of the Cartesian grid in each cube allows a simple

396 Kazuhiro Nakahashi and Yota Sakurai
Fig. 1. Two-dimensional cube frame (left) around A-airfoil and an enlarged view
(right) of the Cartesian grid around a tripping wire.
treatment of complex geometries and an easy implementation of higher-order

numerical schemes.
In each cube, the compressible Navier-Stokes equations are solved by a
cell-centered, finite-volume scheme on an equally-spaced Cartesian mesh. The
inviscid numerical flux vector is computed using an approximate Riemann
solver of HLLEW (Harten-Lax-van Leer-Einfeldt-Wada) [2], with the fourth-
order compact-like MUSCL scheme [3]. The lower/upper symmetric Gauss-
Seidel (LU-SGS) implicit method is used for the time integration. A sub-
iteration scheme [4] is employed in order to keep the time accuracy.
Body boundaries is defined by a staircase representation as shown in Fig.1
(right) in order to keep the algorithm simplicity. Ghost cells are defined as
those cells that locate in the body and next to the flow field. Density and
pressure are given at the ghost cells as average values of the surrounding cells.
The velocity components for the solid wall boundary are set to be zero.
For information exchanges between adjacent cubes, three ghost cells are
added beyond the cube boundary of each side. If the sizes of two adjacent
cubes are same, the exact overlapping of six cells at cube boundary ensures
the accurate transfer of the information up to the fourth-order special ac-
curacy. The information transfer between different cube sizes is performed
by averaging and interpolation. Cube-boundary values are updated by sub-
iterations to keep the time accuracy.
2.2 2D/3D Coupling Computation
For an external flow over an infinite span wing, three-dimensionality of the

flow is significant only in the turbulent boundary layers and the wake, and
the remaining field is essentially two-dimensional. Therefore, a 2D/3D zonal
method was used by several researchers, where the 3D computation is limited
to a zone close to the surface.
The present computation also employs this approach in the framework of
the BCM. Namely, the boundary layer region is computed by using 3D cubes
Simulation of Flow around Wing Sections by Building-Cube Method 397
and the remaining field is simulated by 2D cubes as schematically shown

in Fig. 2. At the interface between 2D and 3D cubes, the data transfer is
conducted by the following definitions.
1 s 3D
Z
q 2D (x, 0, z) = q (x, y, z)dy (1)
s 0
q 3D (x, y, z) = q 2D (x, 0, z) (2)
where s denotes the spanwise extent of the computational domain. The com-
putation is started by 2D cubes, and with the development of the entire flow
field, the local 2D cubes near wall are converted to 3D cubes.
Fig. 2. Building-Cube frame for 3D computation around an airfoil with a spanwise

extent. Right figure shows the 3D-cube region.
2.3 Parallel Computations

In the Building-Cube Method, the flow solver can be parallelized by simply
distributing the cubes to the CPUs in a sequential manner. On a sheared
memory machine, the parallelization is achieved by inserting a command line
of OpenMP in front of the DO-loop of the cube computations.
In the 2D/3D mixed cube computations, however, the load balancing can-
not be achieved by the above-mentioned parallelization. Here, an approach of
grouping the cubes so as to equalize the computational load at each CPU.
This is simply achieved by a proper serialization so as 3D cubes to appear
periodically in the 2D cubes.
3 Computed Results
3.1 NACA0012 Airfoil with Trailing Edge Modifications
Two-dimensional flows around NACA0012 airfoil at relatively low Reynolds
number of 5 × 104 based on the chord length were simulated to observe the
effect of small trailing edge modification to the flows. The free stream Mach
number is 0.5 and the angle of attack is three degrees. With this Reynolds
number, the boundary layer on the airfoil is essentially laminar [5] so that the
2D computations were performed without using turbulence models.
The number of cubes is 522 and each cube has 64x64 Cartesian mesh in
it. The total number of computational cells is 522 × 64 × 64 = 2, 138, 112. The
minimum spacing in the near-wall cube is 4.58 × 10−4 to the chord length.
Outer boundary is located at 60 chord length.
Figure 3(a) shows the pressure distribution around the original airfoil.
The laminar boundary layer on the upper surface separates at mid-chord and
generates vortices that flows down along the airfoil surface. When a vortex
reaches at the trailing edge, it generate a pressure wave that propagates in all
directions from the trailing edge.
Fig. 3. Pressure distributions around NACA0012 airfoils; (a) original, (b) with a
slit, (c) with a downward plane.
In order to investigate the cause of the pressure wave generation, the trail-
ing edge was modified as shown in Fig.3(b) and (c). The modification shown
in Fig.3(b) is to add a slit of 0.01c width at 0.9 chord length. The second one
shown in Fig.3 (c) is to add a small downward plate of 0.02 chord height at
the trailing edge. These geometrical modifications are very easy because of
the Cartesian grid in each Cube.
With these small modifications near the trailing edge, the frequencies of
the vortex shedding were significantly affected as shown in Fig.3(b) and (c).
These results indicate an existence of a feedback loop between the pressure
wave at the trailing edge and the vortex shedding on the upper surface due
to the laminar separation.
Simulation of Flow around Wing Sections by Building-Cube Method 399
3.2 Aerospatiale A-Airfoil
The Aerospatiale A-airfoil at an angle of attack =13.3 deg, the maximum lift
configuration, and at Re = 2.1 × 106 and M∞ = 0.15 was chosen for the
3D computation because of the availability of experimental data as well as
computational results using RANS and LES [6]. The outer boundary of the
computational region locates at 25 chord length far from the airfoil. Total
number of cubes is 569. The minimum spacing near the airfoil surface is
1.9 × 10−4 with respect to the chord length if each cube employs 128x128
Cartesian mesh and 0.95 × 10−4 for 256x256 mesh. The tripping wire used
in the experiment at x/c=0.3 on the lower side was geometrically included in
the grid as shown in Fig. 1.
For 3D computations, 3D zone is selected similar to Mary and Sagaut
computation [7]. That is, the airfoil boundary layer on the upper side is covered
by 3D cubes but the one on the lower side is treated as 2D as shown in Fig. 2.
With this case, the number of 3D cubes is 141 and the number of 2D cubes is
428. The spanwise extent is 0.012 to the chord length and a periodic condition
was used in the spanwize direction. The computation was started by the 2D
flow field with 64x64 grid in all cubes. Then, the grid in cube is doubled to
setup the initial flowfield. The 3D field is computed by using the 2D result
as the initial condition. All 2D and 3D computations are carried out without
using any turbulence models.
Fig. 4. Comparison of 2D and 3D density distributions near the leading edge.
The 2D and 3D results are shown in Fig. 4 for instantaneous density distri-
bution near the leading edge. In the 2D result, small vortices are generated in
a short distance from the leading edge and flows downward. The 3D result also
shows the generation of small vortices. However, those vortices disappear in
a short distance and the flow becomes fully three-dimensional. This indicates
a transition process; the development of 2D waves, followed by the formation
of 3D flow structures, as was also shown by Mary and Sagaut [7].
The pressure plateau near the leading edge and the boundary layer separa-
tion on the upper surface near the trailing edge observed by experiment were
well reproduced by the present computations of 2D of 128x128 and 256x256
as well as 3D with 128x16x128 grid. In the 3D computations, however, a large
vortical structure appeared when the spanwise resolution was increased with
the same spanwize extent. The scale of the vortical structure is much larger
than the spanwise extent. More investigation is required especially for the
spanwize extent in 3D computations.
4 Conclusions
The Building-Cube Method was applied to the NACA0012 airfoil at relatively

low Reynolds number and the A-airfoil near stall condition. For NACA0012
airfoil, it was shown that a small modification of the trailing edge geometry
significantly affected to the vortex shedding frequencies, indicating a feedback
loop between the trailing edge and the laminar boundary layer. For the A-
airfoil, main flow features, such as the transition near the leading edge and
the separation near the trailing edge, were captured. These results indicate
that the present approach is useful to reproduce the micro-scale flow features.
However, it is still very expensive in CPU time for quantitative discussions of
3D micro-macro flows, and needs further study.
References
1. Nakahashi, K.: High-Density Mesh Flow Computations with Pre-/Post-Data
Compressions, AIAA 2005-4876, 17th AIAA CFD Conf. (2005).
2. Obayashi, S., Guruswamy, G. P.: Convergence Acceleration of a Navier-Stokes
Solver for Efficient Static Aeroelastic Computations, AIAA J., 33, 6, 1134–1141
(1995).
3. Yamamoto, S., Daiguji, H.: Higher-Order-Accurate Upwind Schemes for Solving
the Compressible Euler and Navier-Stokes Equations, Computer and Fluids, 22,
2/3, 259–270 (1993).
4. Matsuno, K.: Improvement and Assessment of an Arbitrary-Higher-Order Time-
Accurate Algorithm, Computer and Fluids, 22, 2/3, 311–322 (1993).
5. Elimelech,Y., et. al.: On the Onset of Transition at Low Reynolds Number Flow
Over Airfoils, AIAA 2005-5311 (2005).
6. Mellen, C. P., Frohlich, J., Rodi, W.: Lessons from LESFOIL Project on Large-
Eddy-Simulation of Flow Around an Airfoil, AIAA J., 41, 4, 573–581 (2003).
7. Mary, I., Sagaut, P.: Large Eddy Simulation of Flow Around an Airfoil, AIAA
2001-2559, 15h AIAA CFD Conf. (2001).
Robust Mesh Deformation using the Linear
Elasticity Equations
Richard P. Dwight
Institute of Aerodynamics and Flow Technology, German Aerospace Center

(DLR), D-38108 Braunschweig, Germany. richard.dwight@dlr.de
Summary. A modification is proposed to the equations of linear elasticity as used

to deform Euler and Navier-Stokes meshes. In particular it is seen that the equations
do not admit rigid body rotations as solutions, and it is shown how these solutions
may be recovered by modifying the constitutive law. The result is significantly more
robust to general deformations, and combined with incremental application gener-
ates valid meshes well beyond the point at which remeshing is required.
1 Introduction
The ability to deform meshes according to changes in the computational do-
main is a fundamental tool in CFD, and a foundation on which many other
processes are built. It is for example essential in aerodynamic optimization,
where remeshing the geometry after each design step would cause the change
in the quantity of interest to be swamped by the discretization error [2].
Achieving high-order accuracy in time is much easier when the mesh topology
remains unchanged between time steps [4], but grid deformation is commonly
used in a wide variety of other applications simply because it tends to be
considerably cheaper and more convenient than remeshing; an example is
aero-elastic coupling.
However in practice most mesh deformation algorithms suffer from ro-
bustness problems when faced with large deformations (in particular those
involving large relative motion of bodies), but poor quality input meshes
cause difficulties for small deformations too. Even worse many important
problems involve these two aspects combined; a particularly acute example
is the movement of the flaps and slats of 3d high-lift configurations, for which
few algorithms perform satisfactorily.
One method that shows considerable promise models the mesh as an elas-
tic solid using the equations of linear elasticity. It was applied in [1], together
with refinement and derefinement for time-dependent problems undergoing
substantial deformation. Stein et al. [3] have applied it with an elastic stiff-
ness varying in inverse proportion to cell volume thereby preserving quality

402 Richard P. Dwight
in boundary-layers and regions of high resolution, and in [4] this modifica-

tion was seen to be considerably more robust than several other deformation
algorithms tested. Here the equations are discretized using a Galerkin finite
element method with piecewise linear shape-functions on triangular and tetra-
hedral unstructured grids. A novel modification is developed by examining the
behaviour of the equations under rigid rotations of the mesh, and modifying
the equations to admit these solutions. The modified equations are shown
to be much more robust to large deformations. By applying the deformation
incrementally, extremely large displacements are shown for 2d and 3d Navier-
Stokes (NS) meshes.
The equations of linear elasticity govern small displacements u(x) = (u, v, w)
of an elastic solid subject to body forces and surface tractions. Using the
summation convention they may be written
∇·σ =f on Ω, σ = λTr()I + 2µ, (1)
where f is some body force, Ω is the computational domain, and σ is the
stress tensor, given in terms of the strain tensor by the constitutive relation,
where Tr is the trace, λ and µ are the Lamé constants, and are a property
of the elastic material. It is convenient to express these in terms of Young’s
modulus E and Poisson’s ratio ν as
νE E
λ= , µ= . (2)
(1 + ν)(1 − 2ν) 2(1 + ν)
E > 0 may be thought of as the stiffness of the material, where large E
indicates rigidity. Poisson’s ratio ν is a measure of how much the material
shrinks in the lateral direction as it extends in the axial direction; for physical
materials −1 < ν < 12 . The linear kinematic law = 12 ∇u + ∇uT , quantifies
the change in length and orientation of a material fibre in the elastic body.
Boundary conditions are Dirichlet, u = g on ∂Ω, completing the system.
A significant advantage of the application of these equations for mesh
deformation is that diverse features required in practice can be readily and
simply realized in a manner consistent with the model of the mesh as an
elastic body. For example for a symmetry plane it is convenient that mesh
points move, but remain within the plane, which may be implemented by
the specification of only one coordinate of the displacement. For adaptive
methods it may be desirable to deform the mesh such as to cluster mesh
points in regions of interest, which is readily possible by applying a non-zero
body force f .
3 Finite Element Discretization

The governing equations are discretized on a triangular or tetra Euler or NS
mesh using a Galerkin method based on the trial and test spaces
Robust Mesh Deformation 403
U h = uh |uh ∈ H h (Ω)n , uh = g h on ∂Ω ,

Φh = φh |φh ∈ H h (Ω)n , φh = 0 on ∂Ω ,

where H h (Ω) is a finite-dimensional function space on Ω, and n is the di-

mension of the space. The finite element problem may then be stated: find
uh ∈ U h such that ∀φh ∈ Φh
Z Z
(φh ) : σ(uh ) dΩ = φh · f dΩ. (3)
Ω Ω
It is expected that the mesh resolution will be more than adequate to resolve
features of the elastic solution, therefore H h is taken as the space of functions
linear on the elements of the grid, and then (φh ) : σ(uh ) is constant on
the elements, simplifying implementation considerably. The linear system is
solved using an ILU(m) preconditioned restarted GMRES method, with a
restart vector of 50. For high-Reynolds number NS grids, an ILU fill-in of
m = 4 is required for convergence, leading to high memory costs relative to
the flow solver.
4 Admission of Rigid-Body Rotations

While accuracy per se is not of concern, it seems reasonable to request that
the deformation equations admit rigid body motions of the mesh, which is
not true for the linearized kinematic law, which gives a non-zero strain for a
rotation. An improvement, such as the Lagrangian strain tensor,
1
∇u + ∇uT + ∇uT · ∇u ,

=
2
must necessarily be non-linear, raising the computational cost of the method
unacceptably. However, although the linear strain is non-zero for a rotation
(and for any affine transformation) it is independent of x :

cos θ sin θ cos θ − 1 0
u= − I · x, = , (4)
− sin θ cos θ 0 cos θ − 1
and therefore the stress is also independent of x, equation (1) is satisfied, and
rigid body rotations are admitted “by accident”.
However varying the stiffness of grid cells based on their size increases
the robustness of the method to large deformations considerably. But this
introduces an x dependence into the constitutive law, and rigid body rotations
are lost as solutions. To recover them, it is sufficient that σ = 0 for rotations.
Substituting the expression for above into (1) we have
σ = 2λ(cos θ − 1)I + 2µ(cos θ − 1)I = (λ + µ) [2(cos θ − 1)] I,
which may be set to zero by choosing λ + µ = 0. This is achieved by replacing
the expressions of (2) by λ = −E, µ = E. The same effect can be obtained by
setting the Poisson ratio ν to a very large value, which emphasizes that the
equations can no longer be thought of as a model of elasticity, although they
behave somewhat similarly. Effectively a defect in the kinematic relation has
been corrected by introducing a defect into the constitutive law, resulting in
160
Std Elastic - NACA
Mod Elastic - NACA
140 Std Elastic - Highlift
Mod Elastic - Highlift
Max deflection (Degrees)
120
100
80
60
40
20
0
-1 -0.5 0 0.5
Poisson Ratio !
Fig. 1. Maximum possible deflection for a single step of the standard and modi-
fied elastic models for the NACA0012 and high-lift configurations. Initial and 51◦
defected high-lift grids.
Increment 2°
101 Increment 10°
Increment 30°
Smallest angle in triangles (Degrees)
Increment 90°
100
10-1
0 1 2 3 4
Rotations (Degrees/360)
Fig. 2. Incremental rotation of a NACA0012 with steps of 2 − 90◦ .
an entirely new set of equations. Nonetheless in the following section it will

be seen that not only are rigid body rotations now admitted, but that the
scheme is mush more robust to other deformations.
5 Numerical Results
Two two-dimensional test cases are considered: a Euler grid for a NACA0012
single element aerofoil with a coarse fully unstructured Euler grid, a circular
farfield at a distance of 40 chord lengths, and the aerofoil is rotated about
Robust Mesh Deformation 405
its nose while the farfield is held stationary, see e.g. Figure 2. Also a NS grid
about a three-element high-lift configuration where the slat is deflected and
the remaining elements held fixed, see Figure 1. The robustness of the modified
equations is compared with that of the standard equations for −1 < ν < 0.5,
by determining the maximum deflection that still results in a valid grid, with
no negative volumes. In all cases E is proportional to the inverse of the cell
volume. The results are shown on the left of Figure 1; the modified constitutive
law allows a 70% greater rotation of the NACA, and a 40% greater deflection
of the flap. The grid for the greatest valid flap deflection is shown on the right
of Figure 1. Near the body the point distributions are almost unmodified due
to the much greater elastic stiffness there, a desirable property given that the
initial meshes are likely to resolve surface regions well.
Provided intermediate surface definitions exist, exceptionally large defor-
mations are possible if several steps are taken. For example for the NACA0012
by first calculating the grid for a 90◦ turn an additional 90◦ rotation was pos-
sible. Figure 2 shows the NACA rotated completely 4 times, the left-hand plot
shows the rotation angle against the smallest angle in the grid (a measure of
grid quality) for several increment sizes. The same may be done for the high-
lift case, resulting in a valid grid for a flap deflection of 155◦ , the tip of the
flap almost touching the main element, and an entire grid block compressed
into the intervening space, Figure 3. Of course a flow solution on this grid
reveals defects in the solution, notably the black triangle of low pressure at
the trailing edge of the main element. Finally Figure 4 shows the deformation
of a 3d turbulent wing-body configuration mesh, whereby the calculation was
performed in two steps, a deflection first to 45◦ and then to 90◦ . The mesh
close to the wing is almost unchanged; because of the relatively rapid grid
coarsening with distance from the surface, the algorithm is able to put the
weight of the deformation onto exclusively large cells.
Conclusions and Further Work

It has been shown that it is possible to build deformation algorithms that
allow extremely large deformations of poor quality meshes, something that is
not taken for granted in the aerospace community. It has also been seen that
the linear elastic method may be made significantly more robust with a minor
modification to the governing equations. The efficient solution of these equa-
tions requires some study; the current ILU preconditioned GMRES approach
requires excessive storage in 3d, and a Jacobi iteration converges extremely
slowly for NS grids. Since by far the dominant source of stiffness is the high
aspect-ratio cells in the boundary-layer a line-implicit algorithm combined
with multigrid would be particularly appropriate [4]. A second concern is the
quality of the resulting meshes, which is not usually an issue for small defor-
mations, but - as seen above - strongly deformed meshes are often not suitable
for flow calculations. The amount by which a grid cell is stretched during a
deformation could be quantified by examining the displacement Jacobians,
Z
Y
X
Smallest angle in quads (Degrees)
0
10
Y
X
-1
10
Z
Increment 1° Y
X
Increment 2°
Increment 5°
Increment 10°
Increment 30°
0 50 100 150
Flap Deflection (Degrees)
Fig. 3. Incremental deflection of flap to 150◦ .
Fig. 4. Deformed mesh for the DLR-F4 wing-body configuration.
and mesh refinement could be performed based on this indicator [1]. Finally
a demanding practical test case should be considered.
References
1. T. Baker and P. Cavallo. Dynamic adaption for deforming tetrahedral meshes.
37th AIAA Aerospace Sciences Meeting, Reno, NV. AIAA-1999-3253, 1999.
2. R.P. Dwight and J. Brezillon. Effect of various approximations of the discrete
adjoint on gradient-based optimization. Proceedings of the 44th AIAA Aerospace
Sciences Meeting and Exhibit, Reno NV. AIAA-2006-0690, 2006.
3. K. Stein, T. Tezduyar, and R. Benney. Mesh moving techniques for fluid-
structure interactions with large displacements. J. Appl. Mech., 70:58–63, 2003.
4. Z. Yang and D. Mavriplis. Unstructured dynamic meshes with higher-order
time integration schemes for the unstready Navier-Stokes equations. 43th AIAA
Aerospace Sciences Meeting, Jan. 10-13, Reno, NV. AIAA-2005-1222, 2005.
Part V
DNS and LES

Identification and role of coherent structures in
two-dimensional turbulence
Ch.-H. Bruneau, P. Fischer1 and H. Kellay2

1
MAB, Université Bordeaux 1, 351 Cours de la Libération, 33405 Talence, France
CNRS UMR 5466 - INRIA FUTURS Equipe MC2
bruneau@math.u-bordeaux1.fr, fischer@math.u-bordeaux1.fr
2
CPMOH, Université Bordeaux 1, 351 Cours de la Libération, 33405 Talence,
France h.kellay@cpmoh.u-bordeaux1.fr
Direct numerical simulations show the coexistence of both enstrophy cascade

and inverse energy cascade in two dimensional turbulence. A wavelet packets
filtering allows to split the whole flow into two coherent sub flows, the solid ro-
tation part of the vortices and the vorticity filaments that end in spirals inside
the vortices. Both coherent structures have their own dynamics as revealed by
the computation of the spectra and the fluxes. The computation of the energy
flux shows clearly that the first sub flow is responsible of the inverse energy
cascade from the injection scale to the large scales and the computation of
the enstrophy flux shows clearly that the second sub flow is responsible of the
direct enstrophy cascade from the injection scale to the small scales.
1 The outline of the Direct Numerical Simulation

The two-dimensional incompressible Navier-Stokes equations are solved to
simulate the turbulent flow behind arrays of cylinders. This simulation corre-
sponds to a soap film experiment in a channel where an array of cylinders is
transverse to the flow and two arrays of cylinders are located along the wall.
The injection scale is given by the diameter of the cylinders and is reinforced
by the arrays of cylinders along the wall [3]. To take into account the various
obstacles easily we are solving the two-dimensional penalised non dimensional
Navier-Stokes equations [1] written in ΩT = Ω × (0, T ) where Ω is the full
domain including the obstacles (∂Ω = ΓD ∪ΓW ∪ΓN where ΓD is the entrance
boundary ΓW are the walls and ΓN is the artificial boundary) as
1 U
∂t U + (U · ∇)U − ∆U + + ∇p = 0 in ΩT
Re K
divU = 0 in ΩT
U (0, .) = U0 in Ω (1)
U = U∞ on ΓD × (0, T )
U =0 on ΓW × (0, T )
σ(U, p) n + 12 (U · n)− (U − Uref ) = σ(Uref , pref ) n on ΓN × (0, T )

410 Ch.-H. Bruneau, P. Fischer and H. Kellay
where K = ρkΦU
µH is the non dimensional coefficient of permeability for k
the intrinsic permeability and Φ the porosity of the medium without gravity
(in practise we set K = 1020 in the fluid and K = 10−8 inside the obstacles),
Re = ρUµH is the non dimensional Reynolds number and (Uref , pref ) is a
reference flow. A Poiseuille flow is given upstream and a no-slip boundary
condition is set on the walls of the channel, the boundary condition on ΓN
allows to convey properly the vortices downstream [2].
The equations are solved in primitive variables on Cartesian staggered grids
with a multigrid algorithm coupled to a cell-by-cell relaxation procedure..
The discretization is done by a global second order scheme in time and space
including an accurate third order approximation of convection terms [4]. A
view of the flow behind the arrays of cylinders is given on figure 1.
Fig. 1. Snapshot of the vorticity field including the cylinders.
2 The filtering process
The theory concerning the wavelet packets is detailed in [6] and will not be
reminded here. The same Daubechies type wavelets are used in the current
paper to build the packets array, and the entropy criterion is used to select
the best basis. In [6], few tests have been performed in order to get the best
wavelet mother, and to determine the number of scales necessary for an effi-
cient representation of the flow.It is shown that it is necessary to go over at
least 5 scales from the finest scale corresponding to a 640 × 2560 grid to the
coarsest scales as long as one is concerned with smoothing the discontinuities.
Using an appropriate threshold as in [5] for an academic test it is possible
to separate the structures of the flow. The overall filtering process can be
summarised as follows:
1. Computation of the wavelet packets decomposition of the two components
of the velocity U = (u, v) over 5 scales.
2. Separation of the velocity fields into two sub fields: one sub field, Us =
(us , vs ), where the wavelet packet coefficients have their modulus larger
Role of coherent structures in 2D turbulence 411
than a given threshold and another one, If = (uf , vf ) where the wavelet
packet coefficients are smaller than .
3. Construction of the corresponding vorticity fields, ωs and ωf . The filtered
field ωs is then essentially formed of the solid rotation part of the vortices,
and the filtered field ωf is formed of the vorticity filaments that roll up in
spirals inside the vortices as can be seen on figure 2 on a window of the
flow behind the cylinders.
4. Computations of the physical data: energy, enstrophy, spectra and fluxes.
100
200
300
400
500
600
100 200 300 400 500 600
(a) Pressure field (b) Vorticity field
100
200
300
400
500
600
100 200 300 400 500 600
(c) Solid rotations sub field (d) Filaments sub field
Fig. 2. Snapshots at the end of the channel (kinj ≈ 20). Whole pressure and
vorticity fields (top) and filtered vorticity sub fields (bottom).
Using this efficient filtering, it is possible to separate the whole flow into two
orthogonal coherent sub flows that have their own dynamics as it is shown
below. The flow is not splitted into coherent vortices and background or noise
but into solid rotation part of the vortices and filaments that are also present
inside the vortices. The first sub field contains more than 95% of the energy
and 70% of the enstrophy while the second subfield contains less than 5% of
the energy but 30% of the enstrophy. That is why the dynamics of the flow
starting from the first subfield deviates from the dynamics of the whole flow
as an important part of the enstrophy is missing. The first sub field is strongly
related to the pressure field and can isolate the vortices while the second sub
field is a continuous field that operates the necessary link between the vortices
to yield the whole dynamics.
For the last step of the filtering process a collection of 25 windowed snapshots
thanks to a Tukey window with a small parameter is used to compute the
energy spectra and the fluxes. The time between two snapshots is long enough
for describing a significant evolution of the flow. Indeed between the 1st and
the 25th snapshots, the vortices cross several times the observation box.
The global spectrum shows clearly on figure 3 that two ranges coexist on both
sides of the injection scale which is k = 20 in this simulation and the sum of
the spectra of the two orthogonal sub fields corresponds to the spectrum of
the whole flow.
To confirm the coexistence of both cascades the energy flux ΠE and the
0
10
−5
10
−10
10
Total energy E
Filtered energy Es
Filtered energy Ef
Slope k−2
Slope k−5.5
−15
10
0 1 2
10 10 10
Fig. 3. Original and filtered (with 5 scales packets) energy spectra (kinj ≈ 20).
enstrophy flux ΠZ are computed. We recall that the energy flux is computed
from the non linear term in the Navier-Stokes equation written in Fourier
space:
Role of coherent structures in 2D turbulence 413
Z +∞
ΠE (k) = TE (k 0 )dk 0 (2)
k
where TE (k) is the nonlinear energy transfer function and is obtained by

\
angular integration of U ∗ (k).(U.∇)U
\(k). The enstrophy flux is obtained in
the same way:
Z +∞
ΠZ (k) = TZ (k 0 )dk 0 (3)
k
where TZ (k) is the enstrophy transfer function and is obtained by angular

\
integration of w ∗ (k).(U.∇)w(k).
\
Figures 4 and 5 show that a negative energy flux is obtained on the left hand
0.2
!E
!s " s
E
0.1 !f " f
E
−0.1
−0.2
−0.3
−0.4
−0.5
0 1 2
10 10 10
Fig. 4. Energy fluxes of the whole flow and of the separate structures (kinj ≈ 20).
side of the injection scale and thus an inverse energy cascade stands from this
injection scale to larger scales and the flux vanishes for larger scales. A main
positive part of the enstrophy flux is obtained on the right hand side of the
injection scale and thus a direct enstrophy cascade stands from this scale to the
smaller scales. Let us note also a negative part of the flux before the injection
scale showing that some vorticity is brought by the inverse energy cascade.
These fluxes confirm what can be seen on the spectra. The whole spectrum
is given before the injection scale by the solid rotation field and beyond the
injection scale by the filaments field, the other field being negligible in both
cases.
1000
!
Z
!s " s
Z
f"f
!Z
500
−500
0 1 2
10 10 10
Fig. 5. Enstrophy fluxes of the whole flow and of the separate structures (kinj ≈ 20).
3 Conclusions
A direct numerical simulation of soap film experiments shows the evidence of

the coexistence of an inverse energy cascade and a direct enstrophy cascade
in two dimensional turbulence on both sides of the injection scale. Besides
a wavelet packets filtering allows to separate the flow into two coherent sub
fields that have their own dynamics.
References
1. Angot, Ph., Bruneau, C.H., Fabrie, P.: A penalization method to take
into account obstacles in an incompressible flow. Num. Math., 81, 497–
520 (1999)
2. Bruneau, C.H.: Boundary conditions on artificial frontiers for incompress-
ible and compressible Navier-Stokes equations. M2AN, 34, 303–314 (2000)
3. Bruneau, C.H., Kellay, H.: Coexistence of two inertial ranges in two-
dimensional turbulence. Phys. Rev. E, 71, 046305 (2005)
4. Bruneau, C.H., Saad, M.: The 2D lid-driven cavity problem revisited.
Computers & Fluids, 35, 326–348 (2006)
5. Farge M., Kevlahan N.: Vorticity filaments in two-dimensional turbulence:
creation, stability and effect. J. Fluid Mech., 346, 49-76 (1997)
6. Fischer, P.: Multiresolution analysis for two-dimensional turbulence. Part
1: Wavelets vs Cosine packets, a comparative study. Discr. Cont. Dynam-
ical Syst. B, 5, 659–686 (2005)
Temporal Evolution of Dominant Flow
Structures in Turbulent Channel Flow
Giancarlo Alfonsi1 and Leonardo Primavera2

1
Dipartimento di Difesa del Suolo, Università della Calabria, Via P. Bucci 42b,
87036 Rende (Cosenza), Italy alfonsi@dds.unical.it
2
Dipartimento di Fisica, Università della Calabria, Via P. Bucci 33b, 87036
Rende (Cosenza), Italy lprimavera@fis.unical.it
Summary. The flow of a viscous incompressible fluid in a plane channel is simulated

numerically with the use of a computational code for the numerical integration of
the unsteady three-dimensional Navier-Stokes equations. The numerical method is
based on a mixed spectral-finite difference algorithm. The calculations in the two
homogeneous directions (the streamwise and the spanwise) are performed in Fourier
space and second-order finite differences are used in the direction orthogonal to the
solid walls. A turbulent-flow database representing the turbulent statistically steady
state of the velocity field through 10 viscous time units is assembled at a nominal
friction Reynolds number Reτ = 180 and the coherent structures of turbulence
are extracted from the fluctuating portion of the velocity field with the use of the
Proper Orthogonal Decomposition technique (POD). The temporal evolution of a
number of the most energetic POD modes is visualized, unveiling the mechanisms
of interaction between dominant flow structures in wall-bounded turbulent flows.
1 Introduction
A valuable approach for the numerical investigation of turbulent flows is the
Direct Numerical Simulation (DNS ) according to which the system of the
Navier-Stokes equations is integrated numerically with no modifications of any
kind. The critical aspect in following this approach is the accuracy of the calcu-
lations, that has to be sufficiently high to resolve the essential turbulent scales.
In turbulent shear flows a case that has been studied by some authors is that
of the plane channel. DNS calculations of turbulent channel flow have been
carried out by Kim et al. [1], Lyons et al. [2], Kasagi et al. [3], Rutledge and
Sleicher [4], Moser et al. [5]. In these works the Navier-Stokes equations in ro-
tational form are integrated by means of the fully spectral Fourier-Chebychev
numerical technique introduced in [1]. In the present work a mixed spectral-
finite difference technique for the numerical integration of the Navier-Stokes
equations is introduced and the issue is addressed of the coherent structures
of turbulence in the wall region of turbulent channel flow. The calculations

416 Giancarlo Alfonsi and Leonardo Primavera
are performed at nominal friction Reynolds number Reτ = 180. The coherent
turbulent structures are educed with the technique of the Proper Orthogonal
Decomposition (POD) and the evolution of the dominant flow structures near
the wall is followed in time.
2 Numerical Methods and Simulations

The system of the unsteady three-dimensional Navier-Stokes equations for
incompressible fluids in nondimensional conservative form is considered (i, j =
1, 2, 3):
∂ui
=0 (1)
∂xi
∂ui ∂ ∂p 1 ∂ 2 ui
+ (ui uj ) = − + (2)
∂t ∂xj ∂xi Reτ ∂xj ∂xj
where ui (u, v, w) are the velocity components in the cartesian coordinate
system xi (x, y, z). Variables and operators are nondimensionalized√ by the
channel half-height δ for lengths, wall shear velocity uτ = τw /ρ for ve-
locities, ρu2τ for pressure and δ/uτ for time, being Reτ = uτ δ/ν the friction
Reynolds number, ρ the fluid density and ν the fluid kinematic viscosity. The
fields are admitted to be periodic in the streamwise (x ) and spanwise (z )
directions and equations (1-2) are Fourier transformed accordingly. The non-
linear terms in the momentum equation are evaluated pseudospectrally, by
anti-transforming the velocities back to physical space to perform the prod-
ucts (FFTs are used). A dealiasing procedure is applied to avoid errors in
transforming the results back to Fourier space. In order to have a better spa-
tial resolution near the walls, a grid-stretching law of hyperbolic-tangent type
is introduced for the grid points along y, the direction orthogonal to the solid
walls. For time advancement a third-order Runge-Kutta algorithm is imple-
mented and time marching is accomplished with the fractional-step method
([1]).
By recalling the wall formalism, one has: x+ + 2
i = xi uτ /ν = xi /δτ , t = tuτ /ν =
+ + +
tuτ /δτ , δ = δ/δτ , u = u/uτ , Reτ = uτ δ/ν = δ/δτ = δ , where u de-
notes a mean x -velocity averaged on a x-z plane and time, δτ = ν/uτ is
the viscous length and δ/uτ the viscous time unit. The characteristic param-
eters of the simulations are the following. Computing domain: Lx = 2πδ,
Ly = 2δ, Lz = πδ; L+ + +
x = 1131, Ly = 360, Lz = 565. Computational grid:
+
Nx = 96, Ny = 129, Nz = 64. Grid spacing: ∆x+ = 11.8, ∆ycenter = 4.4,
+
∆ywall = 0.87, ∆z + = 8.8. It can be verified that there are 6 grid points in
the y direction within the viscous sublayer (y + ≤ 5). The Kolmogorov spa-
tial microscale, estimated using the criterion of the average dissipation rate
per unit mass across the width of the channel, results η + ≈ 1.8. After the
Evolution of Dominant Structures in Turbulent Channel Flow 417
insertion of appropriate initial conditions (an initial velocity profile evolving

with time) the initial transient of the flow in the channel is first simulated,
the turbulent statistically steady state is reached and then calculated for a
time t = 10δ/uτ (t+ = 1800). 20000 time steps are computed with a temporal
resolution of 4t = 5 × 10−4 δ/uτ (4t+ = 0.09). The estimated Kolmogorov
time scale results τ ≈ 1.89×10−2 δ/uτ . In Table 1 the predicted and computed
values of a number of mean-flow variables are reported and compared with
DNS data obtained by other authors (Ub is the bulk mean velocity, Reb the
related Reynolds number, Uc is the mean centerline velocity, Rec the related
Reynolds number). The predicted values of Uc /Ub and Cf reported in Ta-
ble 1 are obtained from the experimental correlations suggested by Dean [6]
−0.116 −0.25
[Uc /Ub = 1.28(2Reb ) , Cf = 0.073(2Reb ) ], while the computed skin
friction coefficient [Cf = 2τw /ρUb2 , τw = µ(∂U/∂y)wall ] is calculated using the
value of the shear stress at the wall actually obtained in the computations (a
finite difference routine is used).
Table 1. Predicted vs. computed mean-flow variables
Reτ Reb Rec Ub /uτ Uc /uτ Uc /Ub Cf

predicted
180 2800 3240 15.56 18.02 1.16 8.44 × 10−3
computed
178.74 2786 3238 15.48 17.99 1.16 8.23 × 10−3
The Proper Orthogonal Decomposition (POD) is first introduced in turbu-

lence research by Lumley [7] and is extensively presented in Sirovich [2] and
Berkooz et al. [9]. By considering an ensemble of temporal realizations of a
velocity field ui (xj , t) on a finite domain D, one wants to find the highest
mean-square correlated structure to the elements of the ensemble on average.
This corresponds to find a deterministic vector function ϕi (xj ) that maxi-
mizes the normalized inner product of the candidate structure with the field.
A necessary condition for this problem is that ϕi (xj ) is an eigenfunction, so-
lution of the eigenvalue problem and Fredholm integral equation of the first
kind:
Z Z
Rij (xl , x0k )ϕj (x0k )dx0k = hui (xl , t)uj (x0k )iϕj (x0k )dx0k = λϕi (xl ) (3)
D D
where Rij is the two-point velocity correlation tensor. To each eigenfunc-
(n)
tion ϕi (xj ) is associated a real positive eingenvalue λ(n) . Every member
of the ensemble can be reconstructed by means of the modal decomposition
P (n) R (n)
ui (xj , t) = n an (t)ϕi (xj ), with an (t) = D ui (xk , t)ϕi (xk )dxk . The con-
tribution
R of each mode to the kinetic
P energy content of the flow is given by
E = D hui (xj , t)ui (xj , t)idxj = n λ(n) , being E the kinetic energy in the
Fig. 1. Surfaces of constant streamwise fluctuation reconstructed from the first

three most energetic POD modes at t+ = 50.4 and t+ = 198 (left to right)
domain D. In the present work the POD is used for the analysis of the fluctuat-
ing portion of the velocity field. The two homogeneous directions are handled
in Fourier space, so that the optimal representation of the velocity field in the
statistical sense outlined above is sought in the direction normal to the walls.
3 Results
As a result of the decomposition, 3Ny (387) POD modes and correspondent

eigenvalues are determined for each wavenumber index pair. About 9.9% of the
energy resides in the first 5 most energetic eigenfunctions. The flow structure
formed by the first 3 most energetic modes of the decomposition - identified
in terms of surfaces of constant streamwise velocity - is a structure elongated
in the streamwise direction that propagates with time along the spanwise
direction (Figures 1, the light surfaces are positive streamwise velocity, the
dark surfaces are negative streamwise velocity). The fourth and fifth most
energetic POD modes exhibit a streamwise dependence. The flow structure
formed by the fourth and fifth most energetic modes of the decomposition is
a bean-shaped structure aligned in the streamwise direction that propagates
with time following rigorously the streamwise direction (Figure 2).
Figures 3 show the flow structure formed by the sum of the first 5 most en-
ergetic eigenfunctions of the decomposition (three x -independent structures
and two x -dependent structures) through times t+ = 136.8, t+ = 165.6,
t+ = 172.8 and t+ = 190.8, respectively. With particular reference to the
negative (dark) structure, at t+ = 136.8 a surface extending away from the
streamwise-elongated structure is visible. This phenomenon represents the
onset of the break-up of the main flow structure. The physical reason for the
disruption of the main structure is the result of the interaction between the x-
independent modes and the x-dependent turbulent strucutres. At t+ = 165.6
Evolution of Dominant Structures in Turbulent Channel Flow 419
Fig. 2. Surfaces of constant streamwise fluctuation reconstructed from the second

two most energetic POD modes at t+ = 61.2 and t+ = 302.4 (left to right)
Fig. 3. Surfaces of constant streamwise fluctuation reconstructed from the first five
most energetic POD modes at t+ = 136.8, t+ = 165.6, t+ = 172.8 and t+ = 190.8
(left to right)
the detachment of the extended surface at one end of the negative streamwise-
elongated structure occurs and, at the other end, the first manifestation of the
structure’s break-up becomes evident. At t+ = 172.8 and t+ = 190.8 the pro-
cess of disruption of the originary structure becomes more evident.
The analysis of the first 5 most energetic eigenfunctions of the decomposition
has shown the existence of two kinds of different structures, i.e. elongated
streamwise-independent modes aligned in the x direction and bean-shaped
streamwise-dependent modes. The analysis of the evolution in time of the
flow field produced by the sum of the first 5 most energetic POD modes
has revealed the existence of dominant structures elongated in the stream-
wise direction whose shape is altered due to the process of interaction with
the bean-shaped, quasi-streamwise flow modes. This process of interaction is
responsible for the gradual break-up of the x -elongated structures.
References
1. Kim, J., Moin, P., Moser, R.D.: Turbulence statistics in fully developed channel
flow at low Reynolds number. J. Fluid Mech., 177, 133–166 (1987)
2. Lyons, S.L., Hanratty, T.J., McLaughlin, J.B.:Large-scale computer simulation
of fully developed turbulent channel flow with heat transfer. Int. J. Numer.
Meth. Fluids, 13, 999–1028 (1991)
3. Kasagi, N., Tomita, Y., Kuroda, A.: Direct numerical simulation of passive
scalar field in a turbulent channel flow. ASME J. Heat Transf., 114, 598–606
(1992)
4. Rutledge, J., Sleicher, C.A.: Direct simulation of turbulent flow and heat trans-
fer in a channel. Part I: smooth walls. Int. J. Numer. Meth. Fluids, 16, 1051–
1078 (1993)
5. Moser, R.D., Kim, J., Mansour, N.N.: Direct numerical simulation of turbulent
channel flow up to Re = 590. Phys. Fluids, 11, 943–945 (1999)
6. Dean, R.B.: Reynolds number dependence of skin friction and other bulk flow
variables in two-dimensional rectangular duct flow. ASME J. Fluids Eng., 100,
215–223 (1978)
7. Lumley, J.L.: Stochastic tools in turbulence. Academic Press, (1971)
8. Sirovich, L.: Turbulence and the dynamics of coherent structures. Part I: co-
herent structures. Part II: symmetries and transformations. Part III: dynamics
and scaling. Quart. Appl. Math., 45, 561–590 (1987)
9. Berkooz, G., Holmes, P., Lumley, J.L.: The Proper Orthogonal Decomposition
in the analysis of turbulent flows. Ann. Rev. Fluid. Mech., 25, 539–575 (1993)
LES of passive scalar in compressible mixing
layers
C. Le Ribault
LMFA, ECL, CNRS UMR 5509, UCB Lyon 1

36, avenue Guy de Collongue, 69134 Ecully, FRANCE
catherine.le-ribault@ec-lyon.fr
1 Introduction
Understanding compressible mixing is crucial in aeronautics, for example for

the design of SCRAMJET combustor. A great amount of experimental stud-
ies have shown that compressibility inhibits the growth of the momentum
shear layer and supresses the turbulence intensity and Reynolds stress level.
In two-species shear layers, scalar mixing is also inhibited in the presence of
compressibility (see Clemens [1], Freund [2]) and concentration fluctuation
statistics show that fluctuations are smaller for compressible mixing layers.
Increasing the Mach number is also found to change the mixture fraction
probability density function. Experiments visualizations also reveal changes
in turbulent structures.
The convective Mach number Mc , introduced by Bogdanoff, measures the
intrinsic compressibility of a mixing layer. Large Eddy Simulation (LES) of
compressible mixing layers have already been succefully performed at convec-
tive Mach number up to 1.1 and LES was proved to be able to predict the
decrease of the growth rate and of the kinetic energy. Large Eddy Simulation
(LES) is now carried out to investigate passive scalar mixing in spatially evolv-
ing compressible mixing layers. LES of a temporary evolving mixing layer has
been performed by Sankaran at the convective Mach number of 0.62 [5]. But
this convective Mach number is not high enough to obtain important com-
pressibility effects. The ability of two subgrid models, the dynamic Smagorisky
model and the dynamic mixed model, to predict scalar mixing in compressible
shear layers at convective Mach number up to 1.1 is assessed in this study.
2 Governing equations and numerical method

The flow is governed by the Navier-Stokes equations in their compressible
form, representing mass conservation, momentum conservation and energy

422 C. Le Ribault
conservation. The Navier-Stokes equations in the LES approach are filtered.

In this paper, the top-hat filter with a filter width ∆ is used.
The filtered equations were previously described in Le Ribault [3, 4]. The
subgrid stress tensor qij = ug i uj − ũi ũj is modeled, either by the dynamic
Smagorinsky model, either by the dynamic mixed model. Similar models are
used for the subgrid terms appearing in the pressure equation.
Since the simulations are performed for constant scalar diffusivity, the
nondimensionalized filtered equations for the passive scalar can be written as:
!
∂ρξ˜ ∂ 1 ∂ ∂ ξ˜ ∂ρqkξ
+ ρu˜k ξ˜ = ρ − (1)
∂t ∂xk ReSc ∂xk ∂xk ∂xk
The standard subgrid models (dynamic Smagorinsky and dynamic mixed
models), developed for the velocity field are also applied to the passive scalar
equation.
Spatial derivatives are computed using a non uniform fourth-order com-
pact scheme of Lele. A fourth-order Runke-Kutta scheme is used for the time
integration of the convective terms. Mean hyperbolic profiles are used for the
velocity and for the scalar at the inflow. A broad-band forcing, representative
of isotropic turbulence is added to the velocity profile.
For the passive scalar equation, a flux-corrected-transport, FCT, scheme
of Zalesak is used.
3 Parameters of the problem

The convective Mach number Mc = Uc11 −U +c2 measures the intrinsic compress-
2
ibility of a mixing layer. Mixing layers at the convective Mach numbers of 0.16,
0.5 and 1.1 are simulated. At the first convective Mach number, Mc = 0.16,
the compressibility effects are negligible. The convective Mach numbers of 0.5
has been chosen to correspond to the experiments of Elliot and Samimy [2].
At this convective Mach number, the compressibility effetcs are moderate. At
the convective Mach number of 1.1, important compressibility effects are ex-
pected. The Reynolds number, based on the inflow momentum thickness δθ0
and the velocity difference, is equal to Re = ρδθ0 ∆U/µ = 300. The grid sizes
are also non-dimensionalized by the inflow momentum thickness.
Table 1 summarizes the dimensions of the computational domains and the
number of grid points for the different cases.
Only one filter size ∆ = 2h, where h is the local grid spacing, is used.
4 Momentum mixing
An experimental curve, called “Langley curve” is obtained from the compi-
lation of results corresponding to experiments with air-air shear layers. This
LES of passive scalar in compressible mixing layers 423
Table 1. Computational domains parameters

Mc Lx × Ly × Lz Nx × Ny × Nz
0.16 800 × 144 × 160 185 × 119 × 50
0.5 1000 × 144 × 80 227 × 119 × 50
1.1 2040 × 820 × 80 426 × 179 × 50
curve gives the shear momentum thickness non dimensionalized by its incom-
pressible counterpart in function of the convective Mach number. Shear layer
thickness growth rates predicted by the Langley curve has been plotted on
the figure 1 along with LES results. The spreading rate of the mixing layers
decreases with the convective Mach number in agreement with experimental
results.
1.0
0.8
(d!"/dx)/(d!"/dx)inc
0.6
0.4
Langley experimental curve
LES Smag. dyn.
LES dyn. mixed
0.2
0.0 0.5 1.0 1.5
Mc
Fig. 1. Dependence of shear layer growth rate on Mc
The figure 2 presents the dependence of the peaks streamwise and trans-
verse velocities on Mc . The results of the two turbulence models are com-
pared to experimental results. The difference between the experimental re-
sults are very important on those quantities. All the Reynolds stress compo-
nents slightly decrease with the convective Mach number. The dynamic mixed
model predicts values slightly closer to experimental results than the dynamic
Smagorinsky model but the difference between the two models predictions is
weak.
5 Scalar mixing
The figure 3 compares the similarity profiles of passive scalar turbulence in-
tensities obtained by the two subgrid models at the convective Mach numbers
of 0.16 and 1.1. The self-similarity shape is predicted by the two models. The
difference between the prediction of the two models is small. The peak is
smaller at the convective Mach number of 1.1.
The peak values of the rms of the scalar predicted by the two subgrid
models are plotted together with the experimental results of Clemens [1] in
424 C. Le Ribault
0.20 0.20
Goebels & Dutton (1991)

Elliot & Samimy (1990)
LES dyn. Smag.
0.15 0.15
LES dyn. mixed
rms(u)/Δu
rms(v)/!u
0.10 Goebels & Dutton (1991) 0.10
Elliot & Samimy (1990)

LES Smag. dyn.
LES dyn. mixed
0.05 0.05
0.00 0.00
0.00 0.50 1.00 1.50 0.0 0.5 1.0 1.5
Mc Mc
Fig. 2. Dependence of peak rms streamwise (left) and vertical (right) velocity
0.3 0.3
dyn. Smag. dyn. Smag.

dyn. mixed dyn. mixed
0.2 0.2
rms(")
rms(")
0.1 0.1
0 0
−1.5 −0.5 0.5 1.5 −1.5 −0.5 0.5 1.5
y/! y/!
Fig. 3. Comparison of the RMS of the passive scalar obtained with the two subgrid
models Mc = 0.16 (left) and Mc = 1.1 (right)
function of the convective Mach number (figure 4). The peak value decreases
with increasing compressibility, even at the convective Mach number of 0.5. It
is in agreement with the DNS results of Freund [2] for the annular mixing layer.
Clemens [1] experimental scalar fluctuation statistics show that fluctuations
are smaller at Mc = 0.82 than at Mc = 0.35.
0.25
0.20
0.15
rms(!)
0.10
Dyn. Smag
0.05 dyn. mixed
Clemens and Paul (1995)
0.00
0.00 0.50 1.00 1.50
Mc
Fig. 4. Dependence of peak passive scalar RMS on Mc

LES of passive scalar in compressible mixing layers 425
6 Probability Density Functions

Figure 5 shows the PDF of the mixture fraction obtained with the dynamic
Smagorinsky model for Mc = 0.16 and for Mc = 1.1 at different transversal
sections. The filled symbols across the tops of the plots give the mean values
associated with the PDF. Those plots shows that the most probable scalar
value marches across the mixing layer.
4
η=0.29
3.5 "=0.29 η=0.18
"=0.18 4 η=0
3 "=0.0 η=−0.16
η=−0.29
"=#0.16
2.5 "=#0.29
rms(ξ)
PDF(!)
2
1.5
0.5
0 0
0 0.5 1 0 0.5 1
! ξ
Fig. 5. PDF of the mixture fraction at various transverse locations. Mc = 0.16

(left) and Mc = 1.1 (right)
Figure 6 shows the PDF of the mixture fraction obtained with the two
subgrid models for Mc = 0.16 and for Mc = 1.1 at one section η = 0. There
is a weak trend of narrower and taller PDFs with increasing convective Mach
number. This is consistent with smaller RMS fluctuations. Large eddies will
broaden the PDFs since fluid can be transported from regions far from the
measurement location. On the contrary, a layer that is dominated by small
eddies will produce concentration PDFs that are more narrow. This argu-
ment suggests that scalar transport in compressible layer may occur through
smaller-scale eddies.
7 Isovalues
Figure 7 presents a zoom of isovorticity contours obtained with the dynamic

mixed model and the two different convective Mach numbers (Mc = 0.16
and Mc = 1.1). At low Mc , the turbulent structures appear primarily as
rollers. At high Mc , the structures are more irregular and appeared as jagged,
irregular structures. When the convective Mach number increases, the low
Mach number organized structures are lost and the size of the structures
decreases.
426 C. Le Ribault
3 3
dyn. Smag. dyn. Smag.

dyn. mixed dyn. mixed
PDF(ξ) 2 2
PDF(ξ)
1 1
0 0
0 0.5 1 0 0.5 1
ξ ξ
Fig. 6. PDF of the mixture fraction at ∇ = 0. Mc = 0.16 (left) and Mc = 1.1

(right)
Fig. 7. Isovorticity in a x − y plane. Dynamic mixed model. Mc = 0.16 (left) and

Mc = 1.1 (right)
8 Conclusion
LES of spatially developing mixing layers have been performed at three dif-
ferent convective Mach numbers. Two-subgrid models are compared: the dy-
namic Smagorinsky model and the dynamic mixed model. The principal com-
pressiblity effects such as the decrease of the RMS statistics and qualitative
changes in the PDFs and the structures shape are predicted by both models.
References
1. Clemens N.T., Paul P.H.: Scalar measurements in compressible mixing layers,
Physics of Fluids, 7, no 5 (1995)
2. Freund J.B., Moin P., Lele S.: Compressibility effects in a turbulent annular
mixing layer. Part 2. Mixing of a passive scalar , 421 (2000)
3. Le Ribault C.:Large eddy simulation of compressible mixing layers, International
Journal of Dynamics of Fluid , 1, no 1 (2005)
4. Le Ribault C., Sarkar S., Stanley S.: Large-eddy simulation of evolution of a
passive scalar in a plane jet, AIAA Jounal, 39, no 8, (2001)
5. Sankaran V., Menon S.: LES of scalar mixing in supersonic mixing layers , Pro-
ceedings of the combustion Institute, 30, (2005)
Buoyancy Effect on Turbulence using Blocks
Vincent H. Chu, Wihel Altai, and Camilo E. Pinilla
Department of Civil Engineering, McGill University, Montreal H3A 2K6, Canada

vincent.chu@mcgill.ca
Summary. Buoyancy effect on turbulence was simulated by LES and K- models
using blocks as computational elements. Three values of the Smagorinsky model
coefficient were used in the LES: Cs = 0.0375, 0.075 and 0.15. The fourfold varia-
tion in the coefficient (sixteenfold in subgrid viscosity) had produced essentially the
same results. The insensitivity of the LES results to model coefficient were in sharp
contrast to the simulations of the same flow using the K- model. The buoyancy-
extended coefficient c3 in the K- model was not a constant but dependent on
buoyancy stratification. The optimal values of the coefficient were c3 = 1.0 for
stably stratified flow and c3 ' 0.34 ∼ 0.45 for unstably stratified flow. A slight
modification to the K- model coefficient can led to significantly different prediction
by the model.
1 Introduction
Controlling unphysical oscillations is a problem that has initiated the de-
velopment of many computational strategies (Van Leer [7]; Gaskell and Lau
[4]; Leonard and Mokhtari [9]; Leonard [8]) However, few implementations of
these strategies were carried out for realistic flow computations. In this paper,
the LBM (Lagrangian block method) of Chu and Altai [1] was employed to
simulate the buoyancy effect on turbulence for two flows with opposite buoy-
ancy stratification. In one flow, turbulence was produced by Kelvin-Helmholtz
(KH) instabilities initiated by shear across a stably stratified interface (Figure
1a). In the other case, turbulence was produced by the Rayleigh-Taylor (RT)
instabilities due to unstable stratification (Figure 1b). In the case of stably
stratified flow, the mixing across the interface is bounded by the stablizing
effect of the buoyancy force. In unstably stratified flow, the mixing across the
unstable interface is unbounded as the turbulent motion is reinforced by the
positive work of the buoyancy force The LBM is employed for the sinulations
because the method is free of unphysical oscillations and relatively not diffu-
sive. It is a numerical scheme that can resolve accurately the sharp gradients
across laminar and turbulent interface [2] (see also Figure 2).

428 Vincent H. Chu, Wihel Altai, and Camilo E. Pinilla
Fig. 1. (a) Turbulence initiated by Kelvin-Helmhotz instabilities. The lighter fluid

in the upper layer moved to the right (U1 = 3 cm/s). The heavier fluid in lower layer
(in red) moved in the opposit direction to the left (U2 = -3 cm/s, g’ = 0.49 cm/s2 ).
(b) Turbulence initiated by Rayleigh-Taylor instabilities. The initial reduced gravity
across the unstably stratified interface is g’ = 9.81 cm/s2 .
2 Large Eddy Simulation
The LES (Large Eddy Simulation) of the turbulent flows were conducted using
following sub-grid scale viscosity of the Smagorinsky model:
∂ ũ 2 ∂ṽ ∂ṽ ∂ ũ 2
vSG = Cs2 ∆2 [2( ) + 2( )2 + ( + ) ] (1)
∂x ∂y ∂x ∂y
in which (ũ, ṽ) are filtered components of simulated velocity, and ∆ = [(∆x)2 +
1
(∆y)2 ] 2 is the mesh size of the computation grid. A Smagorinsky coefficient
of Cs = 0.075 was selected for the simulations shown in Figure 1. The results
obtained for the stably stratified flow have produced a Richardson number for
the asymptotic state in close agreement with the laboratory observation by
Thorpe [10]. The simulations of the RT instabilities for the unstably stratified
flow produced the following formulae for the depth of the mixing layer:
σ t2 σ t
= 0.015 2 if t < 6 ts ; = 0.090 if t > 6 ts (2)
h1 ts h1 ts
which agree with the experimental observation by Li and Chu [6]. In these
formulae for the RT instabilities, the lengthp scale and time scale are: h1 =
initial depth of the upper dense layer, ts = h1 /g 0 = time scale, and g 0 = re-
duced gravity. The depth of the mixing region σ is determined by the following
integrals of the buoyancy profile G(z):
R∞ R∞
2 (z − z)2 (dG/dz)dz
o R
z(dG/dz)dz
σ = ∞ , z = Ro ∞ (3)
o
(dG/dz)dz o
(dG/dz)dz
Buoyancy Effect on Turbulence using Blocks 429
(a) QUICKEST 1 rotation (b) QUICKEST 4 rotations
(c) LBM 1 rotation (d) LBM 4 rotations
Fig. 2. One and four rotations of an 8x8 block in a velocity field produced by solid
body rotation. The numerical diffusion associated with LBM is almost negligible.
3 K- Model Simulation

The buoyancy effect on the turbulent flow was alternatively simulated by the
buoyancy-extended K- model of Hossain and Rodi [5]:
DK ∂ ∂K
= [νK ]+P −W − (4)
Dt ∂xα ∂xα
D ∂ ∂ 2
= [ν ] + c1 [P − (1 − c3 )W ] − c2 (5)
Dt ∂xα ∂xα K K
In these model equations, the averaged work done by the buoyancy force per
unit mass of the fluid in the turbulent flow is W = g 0 w0 , in which g 0 is the
buoyancy force per unit mass of the fluid and w0 the vertical velocity fluc-
tuation in the flow. This buoyancy-extended model has an additional model
coefficient, c3 , that depends on stratification. Various attempts to find this
model coefficient has led to very different values of c3 . For stably stratified
flow, we found the value c3 = 1.0, recommended by Gibson and Launder [3]
and Hossain and Rodi [5], to be acceptable. As shown in Figure 3, it had pro-
duced an asymptotic state of the mixing layer with a Richarson number of the
asymptotic state that was consistent with Thorpe’s [10] laboratory observa-
tion. However, the simulation of the turbulent flow in the unstably stratified
Fig. 3. The minimum gradient Richardson number obtained by K- model in

the mixing region of the stably stratified flow produced by KH instability; Ri =
Min[Gz /(Uz )2 ], t0s = (U1 − U2 )/g 0 . The thin solid line delineates the results obtained
using a model coefficient of c3 = 0.4, while the thick solid line the results using the
coefficient of c3 = 1.0
Fig. 4. Depth of the mixing region in the unstably stratified flow produced by the
RT instability. The dash line delineates the results obtained by LES and laboratory
experiments of Li and Chu [6]. The thin solid line delineates the results obtained by
K- model using the model coefficient c3 = 0.4, while the thick solid line the results
obtained using c3 = 0.445 if t < 4.3 ts and c3 = 0.34 if t > 4.3 ts
Buoyancy Effect on Turbulence using Blocks 431
1.6
1.2
0.8
0.4
0.0
0 4 8 12 16
Fig. 5. Depth of the mixing region initiated by RT instabilities in the unstably

stratified flow. The LES results were obtained using three different Cs coefficents.
The lines delineate the experimental results.
flow was more difficulty because the simulation results were sensitive to the
values of the model coefficient c3 . Signficantly different results were obtained
if the value of c3 were only slightly deviated from the optimal. To fit the
simulation with the experimental data, the optimal values are c3 = 0.445 if
t < 4.3 ts and c3 = 0.34 if t > 4.3 ts (see Figure 4).
4 Conclusion
The LES of the turbulent flow was successfully conducted using one Smagorin-
sky model coefficient, Cs = 0.075, for the entire range of Richardson number.
However, the K- model simulations were not as successful. The buoyancy-
extended coefficients in the K- model was not a constant but dependent
on buoyancy stratification. To reproduce the experimental observation, the
buoyancy-extend coefficient was taken to be c3 = 1.0 for stably stratified
flows, and c3 ' 0.34 to 0.455 for unstably stratified flows. The K- model
results are sensitive to slight changes in the values of the model coefficients.
In contrast, the LES were not as sensitive to the model coefficient. As shown
in Figure 5, with the fourfold variation in the Smagorinsky model coefficient,
Cs = 0.0375, 0.075 and 0.15, the LES produced essentially the same results.
References
1. V. H. Chu and W. Altai. Simulation of shallow transverse shear flow by gener-
alized second moment method. J. Hydraulic Research, 39(6):575–582, 2001.
2. V. H. Chu. Laminar and turbulent plumes in density stratified flow. In Proc.

30th IAHR Congress, Vol. C1, Thessaloniki, Greece, pages 279–286. Intl. Assoc.
Hydraulic Research, 2003.
3. M. M. Gibson and B. E. Launder. On the calculation of horizontal non-
equilibrium turbulent flow under gravitational influence. J. Heat Transfer,
ASME, 98(C1):81–87, 1976.
4. P. H. Gaskell and A. K. C. Lau. Curvature compensated convective transport:
SMART, a new boundedness preserving transport algorithm. Int. J. Num.
Meth. Fluids, 8:617–641, 1988.
5. M. S. Hossain and W. Rodi. A turbulence model for buoyant flows and its
application to vertical buoyant jets. In Turbulent Buoyant Jets and Plumes,
HMT-Series, Vol. 6. Pergamon Press, Oxford, England, 1982.
6. X. Li and V. H. Chu. Q2d turbulence by rayleigh-taylor instabilities. In Lee &
Lam, editor, Environmental Hydraulics and sustainable Water Management
(ISBN 04 1536 546 5), pages 263–268. Taylor & Francis Group, 1993.
7. B. Van Leer. Towards the ultimate conservative difference scheme. II. mono-
tonicity and conservation combined in a second-order scheme. J. Comp. Phys.,
14:361–370, 1974.
8. B. P. Leonard. The ultimate conservative difference scheme applied to unsteady
one-dimensional advection. Comp. Meth. Appl. Mech. Engng., 88:17–74, 1991.
9. B. P. Leonard and S. Mokhtari. ULTRA-SHARP nonoscillatory convection
schemes for high-speed steady multidimensional flow. Technical Report NASA
TM 102568 (ICOMP-90-12), NASA-Lewis Research Center, 1990.
10. S. A. Thorpe. Experiments on the instability of stratified shear flow: miscible
fluids. J. Fluid Mech., 46:299–319, 1971.
Direct numerical simulation of Taylor-Couette
flows in the fully turbulent regime
Yves Detandt1,3 , Mikhail Krivilyov2 , Yacine Salhi1 , David Vanden Abeele3 ,

and Jan Fransaer2
1
Université Libre de Bruxelles, Brussels, Belgium
2
Katholieke Universiteit Leuven, Leuven, Belgium
3
von Karman Institute, St. Genesius-Rode, Belgium
ydetandt@ulb.ac.be, mikhail.krivilev@mtm.kuleuven.be,
yasalhi@ulb.ac.be, vdabeele@vki.ac.be, Jan.Fransaer@mtm.kuleuven.be
1 Introduction
In the present paper, we perform a numerical analysis of the turbulent flow

between co-axial cylinders. We consider that only the inner cylinder rotates
(angular velocity ω) while the outer one is at rest. The problem can be de-
scribed by three parameters :
ri
• the ratio of the inner to outer cylinder diameters denoted by η = ro ,
ωri (ro −ri )
• the Reynolds number based on the gap between cylinders : Re = ν ,
• the imposed periodic length in the transverse direction λz .
The reference length is the gap between cylinders denoted in the rest of the
paper by s = ro − ri .
The present paper reports the analysis for fixed geometrical parameters
and for a variable Reynolds number. The analysis focuses on the mixing prop-
erties and in particular on the influence of adding riblets to enhance turbu-
lence. No subgrid scale stress is used in these simulations which are shown to
be resolved through a comparison of mesh size with Kolmogorov length.
The geometry leads to a natural description in a cylindrical coordinate sys-

tem. The flow is periodic in the axial and in the azimuthal direction. Manna
et al. [4] applied a spectral expansion in both periodic directions but were
therefore limited to infinitely long cylinders and could not describe the influ-
ence of the top and bottom boundaries in real electro-chemical cells.
In the present paper, we combine a spectral development in one direction

together with a finite element discretisation in the two remaining directions.
The main objective of the present paper is to validate both versions on a test

434 Detandt et al.
case extensively documented in the literature, in order to use the developed

code for modeling of finite length electro-chemical cells.
2 Numerical aspects
We advance the incompressible Navier-Stokes equations in time in a 2nd order
accurate manner, by setting the following continuity and momentum residuals
to zero at each new time step n + 1:
RCn+1 = ∇· (u n+1 ) = 0, (1)
n+1 1
RM = (u n+1 − u n )/∆t + ∇p n+1/2 − ∇2 (u n+1 + u n ) (2)
2Re
+ (3/2)∇· (u n u n ) − (1/2)∇· (u n−1 u n−1 ) = 0. (3)
Herein, the pressure and viscous terms are treated in an implicit manner
using the Crank-Nicolson scheme, whereas the convective terms are treated
explicitly using the 2nd order Adams-Bashforth method.
2.1 Spatial representation
Both azimuthal and axial directions are periodic for the considered geometry.
It is then natural to represent any flow unknown q (where q represents either
the pressure p or any component of the velocity vector u) by means of a trun-
cated Fourier series in at least one of these directions. In cross planes formed
by the two remaining directions, we introduce a finite element representation
on an unstructured triangular mesh (P1 elements). In the present study, we
have performed all simulations using two different hybrid spectral/FE ap-
proaches:
Planar model: We use a cartesian coordinate system (x, y, z) and discretize
in the axial direction using a spectral expansion:
Nz /2 NNod
1 X X
q n (x = rcosθ, y = rsinθ, z) = Qnk,j Nj (x, y)eIk2πz/Lz , (4)
Nz
k=−Nz /2 j=1
where Lz represents the spanwise periodic distance.

Cylindrical model: We use a cylindrical coordinate system (z, r, θ) and dis-
cretize in the azimuthal direction using a spectral expansion:
Nθ /2 NNod
1 X X
q n (z, r, θ) = Qnk,j Nj (z, r)eIkθ , (5)
Nθ
k=−Nθ /2 j=1
since we have a natural period of 2π in the azimuthal direction.

Note that to ensure the real-valuedness of the solution, the Fourier coefficients
must obey the symmetry relation Qk = Q∗−k .
DNS of Taylor-Couette flows 435
2.2 Variational formulation of the problem
The discrete equations for the unknowns Qnk,i are obtained via the following
weighted residual formulation:
Z Z
φk Ni RCn+1 dV + τpspg φk ∇N˜ i · R n+1 dV = 0, (6)
M
Z
n+1

φk Ni RM dV = 0. (7)
Herein, ∇˜ contains only the cross-plane derivatives of the usual ∇ operator.

˜ = ( ∂/ ∂x, ∂/ ∂y, 0). Pressure stabilization
For instance, for the planar case, ∇
has been added to the Galerkin weighted residual formulation of the conti-
nuity equation, in order to suppress spurious pressure wiggles. The pressure
stabilization time scale τpspg is piecewise constant over each element:
he
τpspg = min(1, Ree /3) . (8)
2β
Herein, he represents the length of the shortest edge of the considered element
e, β stands for an estimate of the maximal velocity in the solution domain
and the element Reynolds number is defined to be Ree = he β/ν.
2.3 Pseudo-spectral time integration scheme
It can be shown that the above approach allows to decouple the 3D linear
problem (Eq. 3) to be solved at each time step into a series of decoupled
2D linear problems, one for each Fourier mode. Coupling still occurs through
the (nonlinear) convective terms, which are however treated explicitly and
therefore do not complicate the (costly) linear solution phase. For the complete
technical details of the discretization procedure outlined above, the reader is
referred to [1].
The discretized equations can be easily computed directly in terms of
the Fourier components, except for the convective terms, which couple each
Fourier mode to all others. The convective terms are therefore computed in
physical space and then transformed to Fourier space using the Fast Fourier
Transform (FFT) algorithm. The basic idea behind this pseudo-spectral solu-
tion procedure is as follows :
(1) Evaluate the convective terms in physical space, using the previously cal-
culated u n and u n−1 ;
(2) transform these terms to Fourier space using the FFT algorithm;
(3) set up the linear problems corresponding to each Fourier mode;
(4) apply appropriate boundary conditions for each Fourier mode;
(5) solve the linear systems to obtain the updated solution unknowns Qn+1 k,j
in Fourier space;
(6) apply the inverse Fourier transform to obtain the updated unknowns q n+1
in physical space.
436 Detandt et al.
For more details on the implementation and in particular the parallelization

approach taken, the reader is referred to [3].
3 Results
We use the same geometry for each computation (η = rroi = 0.875; roλ−r z
i
=
2.5). While our code is able to use fully unstructured meshes, we have choose
to use a structured mesh for this simple geometry (grid spacing is constant
in each direction). For Reynolds number 140, we use a 323 mesh (Nz = Nr =
Nθ = 32) for both versions, while for the turbulent test case (Re=3000), we
use 643 mesh in order to resolve all the turbulent length scales. Five horizontal
ribs have also been placed on the inner cylinder to evaluate the influence of
these on the mixing properties around the inner cylinder. The rib size is 1/15th
of the gap and corresponds to the experiments described in [2].
For the laminar test case at Re=140, the following integral flow quantities
Z 2π Z h Z ro
1
E= |u|2 r dr dz dθ, (9)
2hs2 (ωri )2 0 0 ri
Z 2π Z h Z ro
1
L= uθ r2 dr dz dθ, (10)
2hs3 ωri 0 0 ri
Z 2π Z h Z ro
1
E= |∇ × u|2 r dr dz dθ, (11)
2h(ωri )2 0 0 ri
are compared to literature results by Marcus [6] and Manna et al. [4] to assess
the quality of our results.
For the turbulent simulations at Re=3000, the Kolmogorov length has also
been computed to assess whether the mesh used is sufficiently fine or not :
s
ν
lKolm = p . (12)
2Sij Sij

∂ui ∂uj
Herein Sij is the strain tensor defined as Sij = 21 ∂x j
+ ∂xi
. Defining4 a
cell size as the radial distance between mesh nodes, a simulation is resolved
if lCell ≤ lKolm . In the present case, the Kolmogorov length scale is averaged
over two directions of periodicity (< lKolm >2 = 1 R 2π R Lνz √ ).
2πh 0 0
2Sij Sij
3.1 Reynolds 140
This low value of the Reynolds number, for which the flow is steady and
axisymmetric, is used to assess the performance of both versions of the solver.
The integral flow quantities are compared to numerical results presented in
Marcus [5] and in Manna et al. [4].
The Kolmogorov length scale is 3 times greater than the cell size which
indicates that our simulations are resolved on this mesh.
4
Different definition of a cell size lCell are acceptable. This explains why the cri-
terium consider acceptable ratio llKolm
Cell
should be of the order of 1 and not exactly.
DNS of Taylor-Couette flows 437
Marcus [5] Manna et al. [4] Planar Cylindrical

E 7.14 7.14 7.11 7.15
L 163 163 162 161.9
E 27 27.1 26.5 27.2
Table 1. Integral quantities for Taylor-Couette flows at Re=140
3.2 Reynolds 3000

At this value of the Reynolds number, the flow is turbulent. Large Taylor
vortices are still present and are more intense as illustrated by the Enstrophy
reported in table 2. We attempt to enhance mixing near the inner cylinder
by placing some horizontal riblets. The comparison is made in a vertical cross
plane defined by θ = 0 where we see the small vortices created by the ribs.
Even if the riblets have small effect on the global mixing, we still need to
verify by passive scalar transport if electrons deposit is enhanced around the
inner cylinder by these horizontal riblets. The Kolmogorov length is of the
same order of magnitude as the cell size which indicates that our simulations
are resolved or, at worst, slightly under-resolved.
Planar Cylindrical Cylindrical + Ribblets

E 5.37 6.03 6.16
L 154.5 165 166.3
E 165.6 172.3 159
Table 2. Integral quantities for Taylor-Couette flows at Re=3000
(a) Planar version (b) Cylindrical version (c) Cylindrical version (with ribs)
Fig. 1. Vz instantaneous contours and streamstraces at Re=3000
4 Conclusion
In the present paper we have demonstrated the performance of two different
combinations of Spectral and Finite Elements discretisations of Navier-Stokes
438 Detandt et al.
equations. Good agreement was found between both versions and with liter-
ature.
The main objective of the present paper was to investigate the influence
of putting riblets on the inner cylinder. The next step in this investigation
will be to test the influence of adding vertical riblets. Further analysis on the
transport of a passive scalar is needed to verify the most effective configuration
of riblets for electro-chemical applications.
Acknowledgments
The research was supported by the Institute for the Promotion of Innovation
by Science and Technology in Flanders (IWT, MuTEch project, SBO contract
040092) at VKI and KUL, and by the Higher Education and Scientific Re-
search division of the French Community Government at ULB (ARC project
02/07-283). Yves Detandt was supported by a FRIA fellowship. Computa-
tions are performed on the linux cluster ANIC4 financed by several contracts
(see http://anic4.ulb.ac.be).
References
1. David Vanden Abeele, Gérard Degrez, and Snyder Deryl Owen. A combined
spectral/finite elements method for the direct and large eddy simulation of tur-
bulent flows in complex, two dimensional geometries. In Proceeding CMFF 03,
pages 783–790, 2003.
2. O. Cadot, Y. Couder, A. Daerr, S. Douady, and A. Tsinober. Energy injection in
closed turbulent flows : Stirring through boundary layers versus inertial stirring.
Physical Review E, 56(1), 1997.
3. Vanden Abeele David, Degrez Gérard, and Snyder Deryl Owen. Parallel turbu-
lent flow computations using a hybrid spectral/finite element method on beowulf
clusters. In Proceeding ICCFD 3, 2004.
4. M. Manna and A. Vacca. An efficient method for the solution of the incompress-
ible Navier-Stokes equations in cylindirical geometries. Journal of Computational
Physics, 151:563–584, 1999.
5. Marcus P.S. Simulation of Taylor-Couette flow. Part 1. Numerical methods and
comparison with experiment. Jounral of Fluid Mechanics, 146:45–64, 1984.
6. Marcus P.S. Simulation of Taylor-Couette flow. Part 2. Numerical results for
wavy-vortex flow with one travelling wave. Journal of Fluid Mechanics, 146:65–
113, 1984.
Flux Limiting Schemes for Implicit Large Eddy
Simulation of Synthetic Jets
Sanjay Patel and Dimitris Drikakis
Cranfield University, Department of Aerospace Sciences, Fluid Mechanics &

Computational Science Group, Cranfield, Bedfordshire MK43 0AL, UK
s.patel@cranfield.ac.uk, d.drikakis@cranfield.ac.uk
Summary. The evolution of synthetic jets into quiescent air has been simulated
by performing three-dimensional numerical investigations using various flux limiting
high-resolution schemes in the context of Implicit Large Eddy Simulation (ILES).
The results compare well with existing experimental data from NASA Langley Re-
search Center.
1 Introduction
A synthetic jet refers to a flow created with zero net mass flux. The synthetic
jet is generated using a device consisting of an oscillating diaphragm, where
fluid is alternately pushed in and out of an orifice. The exiting fluid separates
at the edges of the orifice and rolls up to form a vortex ring which propagates
away from the orifice due to its own self induced velocity. As the diaphragm
moves away from the orifice fluid is drawn back into the the cavity which
results in an overall zero net mass flux across the orifice exit. Hence, synthetic
jets are synthesized from the working fluid and do not require complex pip-
ing in order to generate the flow. We examine the evolution of synthetic jets
into quiescent air by performing three-dimensional numerical investigations
using various flux limiting high-resolution schemes for solving the compress-
ible Navier-Stokes equations in the context of Implicit Large Eddy Simulation
(ILES) [1]. Various limiting approaches have been considered including Min-
bee; Superbee; van Albada [2] and the limiter developed by Drikakis [3]. The
computational results have been validated against experimental data obtained
from the NASA Langley workshop on CFD Validation of Synthetic Jets and
Turbulent Separation Control. Results of phase-averaged velocity fields close
to the orifice have shown a good comparison between the computational and
experimental results.

440 Sanjay Patel and Dimitris Drikakis
2 Numerical Method
A three-dimensional parallel compressible Navier-Stokes flow solver has been
used for the simulations. This section will outline the various limiting ap-
proaches considered in the context of high-resolution methods. High-resolution
methods are nonlinear, nonoscillatory methods which use the local solution
to select a technique for approximating the solution. This suggests that high-
resolution methods tend to adapt themselves to the particular circumstances
so that the solution obtained is accurate and has some physical meaning. It has
been shown by Drikakis [1] and Margolin and Rider [4] that high-resolution
methods can be used in (under-resolved with respect to grid resolution) turbu-
lent flow computations without the need to resort to a turbulence model. Lim-
iters are the general nonlinear mechanism that distinguishes modern methods
from classical linear schemes. Their role is to act as a nonlinear switch between
more than one underlying linear method thus adapting the choice of numeri-
cal method based upon the behaviour of the local solution. Limiters result in
nonlinear methods even for linear equations in order to achieve second-order
accuracy simultaneously with monotonicity. Limiters can act like dynamic,
self-adjusting models, modifying the numerical viscosity to produce a nonlin-
ear eddy viscosity [1, 4].
The high-resolution scheme employed here is the Godunov-type, characteristics-
based scheme by Eberle [5]. It is a Godunov-type method that defines the con-
servative variables along the characteristics as function of their characteristic
values. The scheme will not be presented here in detail and the author would
refer the reader to the paper by Eberle [5] for further details. The cell centered
data is interpolated to the cell faces using a MUSCL-type high-order interpo-
lation which incorporates the various flux limiters studied in this paper. This
method can be expressed as;
L 1
Ui+1/2 = Ui + [(1 − k)(Ui − Ui−1 ) + (1 + k)(Ui+1 − Ui )]. (1)
4
R 1
Ui+1/2 = Ui+1 − [(1 − k)(Ui+2 − Ui+1 ) + (1 + k)(Ui+1 − Ui )]. (2)
4
This form of the extrapolation is symmetric about the interface i + 1/2, and k
is a free parameter between -1 and 1. The MUSCL scheme as presented above
is not stable on it’s own and hence a limiting method must be incorporated.
Limiters are generally defined using the parameter r where;
Ui+1 − Ui Ui+1 − Ui
rL = , rR = . (3)
Ui − Ui−1 Ui+2 − Ui
A limiting function φ(r) is defined giving a symmetric limited MUSCL scheme
in the form;

L 1 L 1
Ui+1/2 = Ui + [(1 − k)φ(r )(Ui − Ui−1 ) + (1 + k)φ (Ui+1 − Ui )].(4)
4 rL
Flux Limiting Schemes for ILES of Synthetic Jets 441

R 1 1
Ui+1/2 = Ui+1 − [(1−k)φ(rR )(Ui+2 −Ui+1 )+(1+k)φ (Ui+1 −Ui )].(5)
4 rR
The four limiters investigated in this paper are the van Albada (VA), Minbee
(MB), Superbee (SB) and the limiter developed by Drikakis (DD) [3]. They
can be expressed as;
r(1 + r)
φV A = (6)
1 + r2
φM B = min(1, r) (7)
φSB = max(0, min(1, 2r), min(r, 2)) (8)
N
2N r 2r
φDD = 1 − 1 + 1− (9)
1 + r2 1 + r2
In all the above cases we have used k = 1/3 which gives third order accuracy
for the MUSCL reconstruction.
We have performed three-dimensional computations of a synthetic jet exiting

into quiescent air and compared different flux limiting schemes. The results
have been compared to experiments undertaken at NASA Langley. The ex-
perimental data [6] was collected using a variety of methods including Particle
Image Velocimetry (PIV), Laser Doppler Velocimetry (LDV) and Hot Wire
anemometry (HW), thus providing a range of data available to use in the
validation.
The Reynolds number based on the orifice width and average velocity
over the discharge phase of a cycle is 1150. The forcing frequency that the
diaphragm oscillates at is 444.7 Hz taken from the experimental data and the
Mach number is set at 0.04. A grid resolution of approximately 3.2 million
grid points is used for the computations presented in this paper. The flow was
allowed to develop over several cycles and phase averaged data was calculated
once the flow was regarded to be fully developed. Figure 1 shows a schematic
representation of the experimental and corresponding computational setup.
Due to the complexity of the experimental setup a more simplified approach
to model the synthetic jet which did not include the oscillating diaphragm and
cavity was adopted. Simply the unsteady nature of the synthetic jet diaphragm
was modelled as a sinusoidal blowing/suction boundary condition at the orifice
exit.
Phase averaged streamwise velocity profiles at the center of the slot were
obtained at various positions above the slot exit for each of the different lim-
iter cases (see figure 2). At 0.1mm above the orifice exit it is clear to say
that none of the three experimental techniques agree which each other and
that the velocity does not vary in a strict sinusoidal profile. This is due to
Fig. 1. Schematic representation of the experimental and computational setup
the flow having to travel through the cavity which has not been modelled in
the computational case. A phase lag in the computations can be seen on this
figure with the experimental results reaching a peak velocity earlier than the
computations. At a distance of 4mm from the orifice exit the peak velocity
in the computations compares well with the hot wire data but the PIV mea-
surements show a large increase in velocity. The various limiters compare well
especially in the suction phase, with Minbee, van Albada and the DD limiter
giving the best results in comparison to the PIV data. Again, the phase lag
in the computational results is especially noticeable in the expulsion phase.
Phase averaged results of u- and v-velocity profiles along the horizontal line
2mm above the orifice exit at phase angles of 90◦ and 270◦ are shown in figure
30 PIV data
HW data
LDV data 30
DD limiter
20 van Albada limiter DD limiter
Superbee limiter van Albada limiter
25
Minbee limiter Superbee limiter
Minbee limiter
10 PIV data
Velocity m/s
Velocity m/s
20
HW data
0 15
10
-10
5
-20
0
0 100 200 300 0 50 100 150 200 250 300 350
Phase deg Phase deg
(a) 0.1 mm (b) 4 mm
Fig. 2. Phase-averaged u-velocity at 0.1mm (a) and 4mm (b) above the center of
the orifice exit.
Flux Limiting Schemes for ILES of Synthetic Jets 443
6
DD limiter
van Albada limiter DD limiter
25 van Albada limiter
Superbee limiter 4
Minbee limiter Superbee limiter
PIV data Minbee limiter
20 PIV data
2
Velocity m/s
Velocity m/s
15
0
10
-2
5
-4
0
-6
-2 -1 0 1 2 3 4 -2 -1 0 1 2 3 4
Position across slot mm Position across slot mm
(a) U-velocity 2mm (b) V-velocity 2mm
DD limiter
-1 van Albada limiter
Superbee limiter 3
Minbee limiter
-1.5 PIV data DD limiter
2 van Albada limiter
Superbee limiter
-2 Minbee limiter
1
Velocity m/s
PIV data
Velocity m/s
-2.5
0
-3
-1
-3.5 -2
-4 -3
-2 -1 0 1 2 3 4 -2 -1 0 1 2 3 4
Position across slot mm Position across slot mm
(c) U-velocity 2mm (d) V-velocity 2mm

Fig. 3. Phase-averaged velocity profiles along the 2mm horizontal line above the
jet exit plane at 90◦ (a, b) and 270◦ (c, d)
3. The various limiters tend to over predict the peak u-velocity at phase of
90◦ with Minbee and van Albada giving the best results away from the slot
edges. In contrast Minbee, van Albada and the DD limiter under predict the
maximum and minimum velocities in the cross-stream direction whereas Su-
perbee gives an extremely close match especially in the region of the orifice
edges. At maximum suction (270◦ ) all four limiters give different results in
comparison with the PIV u-velocity data. All limiters over predict the mini-
mum velocity with van Albada and the DD limiter showing the best results
with respect to the velocity profile shape. The cross streamwise profile shows
a steep gradient across the orifice exit in the computational results whereas
the PIV shows a more smoother transition across the orifice. A comparison
of phase averaged centerline streamwise velocity at maximum expulsion and
suction are shown in figure 4. Both plots show that the velocity at the orifice
exit is over predicted by all the different limiters. No limiter manages to cap-
ture the position and magnitude of the maximum expulsion velocity. The van
DD limiter
van Albada limiter
30 Superbee limiter 10
Minbee limiter
PIV data
25
Velocity m/s
0
Velocity m/s
20
DD limiter
15 -10 van Albada limiter
Superbee limiter
Minbee limiter
10 PIV data
-20
0 2 4 6 8 0 2 4 6 8
Streamwise position mm Streamwise position
◦ ◦
(a) Phase 90 (b) Phase 270
Fig. 4. Phase-averaged centerline u-velocity at maximum expulsion and suction

Albada and DD limiters are shown to best predict the velocity at positions
further downstream at both phase angles.
4 Conclusions
The evolution of a three-dimensional synthetic jet in quiescent air has been
simulated using various limiting schemes. The results obtained have compared
relatively well with the experimental data from NASA Langley. It was found
that the various limiters investigated differed with regards to the amount
of dissipation provided by the scheme. All limiting schemes gave reasonable
results close to the orifice exit but further downstream the difference in the
schemes was much higher, with the van Albada limiter and the limiter of
DD performing best. The study of various limiters is complex with limiters
behaving differently depending on the characteristics of the flow situation. It is
clear from these results that the cavity section with the oscillating diaphragm
needs to be modelled in the computations. This should provide a much more
realistic velocity profile at the orifice exit and hence give more accurate results.
References
1. Drikakis, D: Advances in turbulent flow computations using high-resolution
methods. Progress in Aerospace Sciences. 39, 405–424 (2003).
2. Toro, E: Riemann solvers and numerical methods for fluid dynamics. Springer-
Verlag (1997)
3. Zóltak, J. and Drikakis, D: Hybrid upwind methods for the simulation of un-
steady shock-wave diffraction over a cylinder. Computer Methods in Applied
Mechanics and Engineering. 162, 165–185 (1998).
4. Margolin, L.G. and Rider, W.J: A rationale for implicit turbulence modeling.
International Journal for Numerical Methods in Fluids. 39, 821–841 (2002).
5. Eberle, A: Characteristic flux averaging approach to the solution of Euler’s equa-
tions. Computational Fluid Dynamics, VKI Lecture Series. (1987).
6. Yao, C., Chen, F.J., Neuhart, D. and Harris, J: Synthetic jet flow field database
for CFD validation. 2nd AIAA Flow Control Conference. 2218, (2004).
Implicit Large Eddy Simulation of a Flow
around a Cylindrical Body
Satoko Komurasaki1 and Kunio Kuwahara2

1
Dept. of Mathematics, College of Science and Technology, Nihon University
1-8-14 Kanda-Surugadai, Chiyoda-ku, Tokyo 101-8308, Japan.
satoko@math.cst.nihon-u.ac.jp
2
Institute of Computational Fluid Dynamics
1-22-3 Haramachi, Meguro-ku, Tokyo 152-0011, Japan. kuwahara@icfd.co.jp
Summary. Numerical simulation was carried out for a flow around a cylindrical
body. The incompressible Navier-Stokes equations were solved by the multidirec-
tional finite-difference method. For high-Reynolds-number flows, a third-order up-
wind scheme was utilized for the convective terms to stabilize the computation. As
a result of the computation, drag variation caused by different section shape of body
was captured well.
1 Introduction
Many simulations of a flow past a streamlined body have been performed.
But those of bluff body have not been done much. To resolve the boundary
layer, we use finite difference approach. Flow around a car is one of the most
important applications. But flow around a car is very complicated, grid gen-
eration is not easy. To understand the most important physical features of
these flows, we compute a flow around a cylindrical body.
2 Computational Method
The governing equations are the 3-d incompressible Navier-Stokes equations
and the equation of continuity as follows:
divu = 0, (1)
∂u 1
+ u · gradu = −gradp + 4u. (2)
∂t Re
where u, p, t and Re denote the velocity vector, pressure, time and the
Reynolds number respectively. For high-Reynolds-number flows, time-dependent
computations are required owing to the strong unsteadiness.

446 Satoko Komurasaki and Kunio Kuwahara
The numerical procedure is based on the projection method. The pressure

field is obtained by solving the following Poisson’s equation:
Dn
4p = −div(u · gradu) + (3)
δt
D = divu,
where n is the time step and δt is the time increment. Dn+1 is assumed to be
zero, but Dn is retained as a corrective term.
The equations are discretized based on the multi-directional finite-difference
method. In case of 2-dimensional computations, when structured grid points
are given, the black points in Fig.1(a) are usually used to approximate the
derivatives at the central point (system A). In the multi-directional finite-
difference method, we introduce the other 45 degrees-rotated local grid system,
and employ the black points in Fig.1(b) for approximation to the derivative
at the central point (system B). In order to improve the derivative value at
the central point, the values from both systems are used. If a ratio A:B = 2:1
is adopted, the resulting finite-difference scheme for the Laplacian, coincides
with well-known 9 point formula with forth-order accuracy. This method im-
proves the rotational invariance of the coordinate system, then those cases
where flow direction is not parallel to the grid direction are better simulated.
In 3 dimensions, three different grid systems are used. Each grid system is
obtained by rotating a perpendicular plane 45 degrees with respect to each
coordinate axis (Fig.1(c)). The multidirectional-finite-difference formula for
the Laplacian is given by:

2 1 1 1
4p ≈ (p 1 +p 2 +p 3 +p 4 −4p 0 ) + (p x1 +p x2 +p x3 +p x4 −4p 0 )
3 h2 3 2h2
The space derivatives are discretized using a three-point central difference

approximation with exception of the convective terms. For the convective
terms, a third-order upwind scheme is used to stabilize the computation. It
P2 P2 P5
Px2 Px1
Px2 P2 Px1
P3 P0 P1 P3 P0 P1
P3 P0
h P1
P4 Px3 P4 Px4
Px3 P4 Px4
P6
(a) System A. (b) System B. (c) System in 3D.
Fig. 1. Multidirectional-finite-difference scheme

Implicit Large Eddy Simulation of a Flow around a Cylindrical Body 447
r
d
Fig. 2. Section shape
has been found to be the most suitable for high-Reynolds number flow com-
putations. The second-order Crank-Nicolson implicit scheme is used for time
integration. The equations are iteratively solved at each time step by SOR
method. For Poisson’s equation, a multi-grid method is utilized.
In the computations, the flow around a square cylinder with rounded cor-
ners is solved. For a section of the cylinder we define d as its side length and
r as the radius of rounded corners (Fig.2). Computations are carried out for
several cases of r/d. An O-type grid is employed, and the grid points can be
concentrated near the body surface. The number of grid points is 257×129×81,
and the aspect ratio is one. In the spanwise direction, a periodic condition is
employed. Figure 3 shows body-fitted grid used in this computation. Global
view and local view near a corner are displayed in Fig.3 (a) and (b) respec-
tively. Fully developed 2-D flow is used as an initial condition for the 3-D
computation to save computation time.
3 Results
Three-dimensional flows around a rectangular cylinder with r/d =0, 0.021,

0.167 and 0.333, were simulated at Re = 105 . Figure 4 shows the Cd variation
with r/d. Experimental values are added by white points in this graph.
In Figs. 5 and 6, the flow field for each r/d is exhibited. An instantaneous
flow field is expressed by using contour surfaces of longitudinal component of
vorticity and pressure contour lines in Fig.5. Figure 6 shows time-averaged
(a) Global view. (b) Local view.
Fig. 3. Computational grid

Fig. 4. Drag coefficient Cd at each r/d. (Re = 105 )
flow field visualized by stream lines, pressure contour lines and shading of
lateral component of vorticity.
Also, other computations at Re = 103 to 105 were carried out at Re = 103
to 105 , where aspect ratio based on the chord length is three and the number
of grid points is 129 × 65 × 193. These results are shown in Figures 7 and
8 presenting the instantaneous flow field at r/d = 0 and 0.333 respectively.
The Reynolds number is 105 . Contour surfaces of longitudinal component of
vorticity and pressure contour lines are visualized.
4 Conclusion
Simulation of flows around a cylindrical body was carried out by using
multidirectional-finite-difference method with O-grid. As a result of the com-
putation, flow fields were clarified in each case, and the drag variation caused
by different section shape of the body was captured well. Differences of flow
characteristics by r/d were elucidated by visualizing suitably.
References
1. K. Kuwahara and S. Komurasaki, ”Simulation of High Reynolds Number Flows
Using Multidirectional Upwind Scheme”, AIAA Paper 2002-0133 (2002)
2. S. Komurasaki and K. Kuwahara, ”Implicit Large Eddy Simulation of a Subsonic
Flow around NACA0012 Airfoil”, AIAA Paper 2004-0594 (2004)
3. T. Kawamura and K. Kuwahara, ”Computation of high Reynolds number flow
around a circular cylinder with surface roughness”, AIAA Paper 84-0340 (1984)
4. H. Suito, K. Ishii and K. Kuwahara, ”Simulation of Dynamic Stall by Multi-
Directional Finite-Difference Method”, AIAA paper 95-2264 (1995)
5. T. Igarashi, ”Characteristics of the Flow around Rectangular Cylinders (The
Case of the Angle of Attack 0 Deg)”, J. of JSME(B), 50-460, 3185–3192 (1984)
(in Japanese)
Implicit Large Eddy Simulation of a Flow around a Cylindrical Body 449
(a) r/d = 0. (a) r/d = 0.
(b) r/d = 0.021. (b) r/d = 0.021.
(c) r/d = 0.167. (c) r/d = 0.167.
(d) r/d = 0.333. (d) r/d = 0.333.

Fig.5. Instantaneous flow field; Fig.6. Time-averaged flow field;
pressure contour lines and streamlines, pressure contour lines and
longitudinal component of vorticity. lateral component of vorticity.
Fig.7. Flow around a rectangular cylinder at Re = 105 , contour surfaces of

longitudinal component of vorticity, pressure contour lines.
Fig.8. Flow around a rectangular cylinder with r/d = 0.333 at Re = 105 , contour
surfaces of longitudinal component of vorticity, pressure contour lines.
LES study of the impact of the wake structures
on the aerodynamics of a simplified ICE2 train
subjected to a side wind
Hassan Hemida1 and Sinivsa Krajnović2

1
Division of Fluid Dynamics, Department of Applied Mechanics,
Chalmers University, SE-412 96 Gothenburg, SWEDEN hemida@chalmers.se
2
Division of Fluid Dynamics, Department of Applied Mechanics,
Chalmers University, SE-412 96 Gothenburg, SWEDEN sinisa@chalmers.se
1 Introduction
Until recently, experimental studies and numerical simulations of the flow

around trains in side winds have focused on measuring the main integral
quantities, such as the aerodynamic forces and moments, but not much on
understanding the flow structures. Chiu and Squire [1] experimentally found
that at low yaw angles (up to 40o ) the flow is similar to the steady slender body
flow, in which pairs of steady line vortices are emerging from the separation
lines on the lee-side face to form the wake structures. When the side-wind yaw
angle increases above 60o , the flow changes from that associated with a slender
body to unsteady vortex shedding. However, the mechanism of this transition
between these two types of flows and its effect on the aerodynamic forces
have not been investigated in detail. The present investigation focuses on
the impact of the wake structures on the aerodynamic coefficients by solving
the flow around a simplified ICE2 train subjected to a 30o side-wind yaw
angle using large-eddy simulation (LES). Standard Smagorinsky model with
a model constant Cs = 0.1 is used to model the subgrid scales. The Reynolds
number is 2 × 105 , based on the freestream velocity and the height of the
train, D. The simplified train model consists of a leading car to which an end
car dummy is attached (see Fig. 1b). The total length of the train model is
L = 3.552m and its height is D = 0.358m. The clearance between the train
and the computational domain floor is 0.0537m (0.15D). The LES results are
compared with wind tunnel test results. The measurements were made in the
subsonic Russian T sAGI T − 103 open jet tunnel. The experimental set-up
is shown in Fig. 1a. In contrast to the computational model, the wind tunnel
model has bogies and an inner gap between the leading car and the dummy
end car. The model stands on a thin elliptic plate (see Fig. 1a) to reduce the
thickness of the approaching boundary layer. Moreover, in order to keep the

452 Hassan Hemida and Sinivsa Krajnović
=0.580
=0.03
=0.07
=0.14
=0.44
=0.75
End car dummy
Support in the
Leading car End car experiment
Veiw from
front
Inner gap
Bogies
Spoiler in the
experiment Bogies in the
a) Ground board b) experiment
Fig. 1. (a) Experimental set-up, (b) simplified model.
ground clearance, the model is held by a single cylindrical support positioned

midway between the bogies. To simplify the model in the computations, the
inner gap is filled and both the bogies and the cylindrical support are removed.
In the measurements, the Reynolds number was Re = 1.4 × 106 based on the
free-stream velocity and the height of the train. In the computational domain,
the model is yawed 30o to the free-stream direction as shown in Fig. 2. The
clearance between the model and the ground board is the same as in the
experimental set-up. The model centerline is kept parallel to the inlet of the
domain with a streamwise distance of 8D to ensure the same thickness of
the approaching boundary layer. The distance between the train tail and the
exit of the computational domain is 21D. The height of the computational
domain is 5.2D. The flow enters the domain with a uniform velocity profile
Lateral wall
Sli
Exit
pw
all
a) b)
Lateral wall
Fig. 2. The computational domain.
constant in time. A convective-velocity boundary condition is implemented at

the domain exit. No-slip boundary conditions are used on the train surface.
To simulate the experimental set-up, a slip boundary condition is applied at
the lateral walls and the roof. To reduce the thickness of the boundary layer
approaching the train model, two kinds of boundary conditions are applied at
the channel floor. A slip boundary condition is applied on a part extending
5D (see Fig. 2b) from the inlet to suppress the development of a boundary
layer, while a no-slip boundary condition is used on the rest of the floor. A
homogeneous Neumann boundary condition is used for the pressure on all the
boundaries.
LES of side-wind flow over ICE2 train 453
2 Results
Numerical accuracy is investigated by performing two simulations with dif-
ferent number of nodes: 6 and 12 million nodes. The spatial resolutions of
the first cell layer at the model and floor are shown in Table 1 for the two
meshes, where u∗ is the friction velocity, n is the distance between the first
node and the train surface in the wall normal direction, ∆s is the cell width in
the streamwise direction and ∆l is the cell width in the span-wise direction.
Table 1. Spatial resolutions for the coarse and the fine mesh simulations.
y + = nu∗ /ν s+ = ∆su∗ /ν l+ = ∆lu∗ /ν

fine mean 0.5 40 90
maximum 4 250 700
coarse mean 1.5 100 300
maximum 20 450 1200
The surface pressure distributions are computed at certain cross-sections

along the length of the train where experimental data are available, as shown
in Fig. 1b. Figure 1b also shows some details from the experimental train, such
as spoiler, support, bogies and inner-car gap, which are omitted from the com-
putational model. These differences should be kept in mind when comparing
the LES results with the experimental data. Figure 3 shows the local pressure
coefficient, Cp , over the circumferential angle, θ, measured counterclockwise
around the x-axis, as shown in Fig. 1b. Although the LES results from the
coarse and fine meshes are found to be in a good agreement with the experi-
mental data on the nose of the train (x/L < 0.14), the coarse mesh simulation
was not able to obtain the minimum pressure on the top and bottom-side faces
as shown in Fig. 3. Far from the nose of the train, the resolution of the coarse
mesh is not sufficient to get results that agree with the experimental data,
especially on the bottom and top-side faces. The LES results from the fine
mesh show good agreement with the measurements on the streamwise and
top-side faces at all the cross-sections. Moreover, good correlation is obtained
between the LES results from the fine mesh and the measured data on the
lee-side face. The difference in the geometry under the train (i.e. bogies, the
spoiler and the support) between the experimental and the numerical set-ups
resulted in a difference in Cp values on the surface of the under side. The
effect of these complexities is believed to be localized since it does not in-
fluence the pressure on the other faces. The maximum suction peak and the
highest pressure are found on the upper-face leeward edge of the nose and
on the streamwise face of the nose, respectively, at two front cross-sections,
x/L = 0.03 and x/L = 0.07 (Fig. 3). The peak Cp magnitudes in these two
sections are almost twice as large as those in the subsequent cross-sections.
This suggests that the loads are accumulated in the nose region of the leading
Lee-side Bottom-side Streamwise

face face face
$ $
# #
" "
! !
!" !"
! "!! #!! $!! ! "!! #!! $!!
Angle ( ) Angle ( )
$ $
# #
" "
! !
!" !"
! "!! #!! $!! ! "!! #!! $!!
Angle ( ) Angle ( )
$ $
# #
" "
! !
!" !"
! "!! #!! $!! ! "!! #!! $!!
Angle ( ) Angle ( )
Fig. 3. Time-averaged surface pressure: (solid) fine mesh, (dashed) coarse mesh,
(symbols) experimental data.
car. In general, although the LES Reynolds number is seven times lower than
the experimental one, the LES results are in fairly good agreement with the
measurements.
Figure 4 shows the LES wake structures by means of isosurface of the in-
stantaneous pressure. The flow separates from the top-side face at the wind-
ward edge. It reattaches to the surface forming small separation bubble on
the top side face. On the other side, the flow remains attached to the bottom-
side face before it enters the wake. Both of the flows on the bottom-side and
top-side faces separate from the surface when they reach the lee-side face to
form line vortices emerging from the separation lines. These line vortices are
forming the wake flow patterns. The upper vortices are stronger and more
steady than the lower-side ones. Figure 4 shows that the first upper vortex
stretches in the wake flow all the way from the nose tip to the end of the
leading car. It grows in size but remains attached to the train surface. It is
steady for a length of about 5D from its onset at the front of the train before
it becomes unsteady and breaks up. The lower vortices are highly unsteady.
The cores of these vortices oscillate in a horizontal plan with low frequencies,
LES of side-wind flow over ICE2 train 455
where they attach and detach to and from the surface of the train with the
same frequency. The vortex shedding in the wake is mainly from the lower
vortices. Attachment and detachment of the wake vortices to the train leave
disturbances on its surface pressure which affect the aerodynamic coefficients.
Our LES resulted in wake structures are in agreement with those obtained
Upper vortex
Detached Attached
vortex vortex
Lower vortex
a) b)
Fig. 4. Isosurface of the instantaneous pressure p = −0.4P a at two different

times.View is from above the train.
Incident wind
a) b)
Fig. 5. (a) Experimental view (taken from [2]). (b) DES results (taken from [3]).
by Wu [3] using Detached-Eddy Simulation ( see Fig. 5b). On contrary to our

finding, Copley [2] reported that the lower-side vortices of the wake are mirror
images of the upper-side ones. They found that these vortices are shed from
successive points along the train length with the same vortex strength to form
a three-dimensional Von Karmen Vortex Street (see Fig. 5a ).
The time histories of the side force and the lift force coefficients, Cs and Cl ,
are shown in Figs. 6a and 6b, respectively. In order to find the frequencies
of the different motions in the flow, the power spectra of the aerodynamic
coefficients are drawn against the Strouhal number, St, as shown in Figs. 6c
and 6d. The major peak in the spectra is corresponding to the attachment
and detachment of the wake vortices to and from the surface of the train. The
second dominating frequencies are corresponding to the shedding frequencies.
0.46 0.2
0.44 0.15
0.42
0.1
0.4
0.38 0.05
0.36 0
0 10 20 30 40 0 10 20 30 40
a) t b) t
3 15
2 10
1 5
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
c) d)
Fig. 6. (a) and (b) are time history of the side force and the lift force coefficients,
Cl and Cs , respectively. (c) and (d) are the corresponding autopower spectra.
To conclude, the present work shows that LES is capable of accurately

predicting the flow around a high-speed train such as ICE2. Our LES have
provided a new picture to the wake flow that differs from the previously pub-
lished experimental picture.
Acknowledgments
This work was supported by the Swedish Agency for Innovation System (VIN-
NOVA), Bombardier Transportation, and Scania. Computer time on HPC2N
(High Performance Computing Center North), NSC (National Supercomputer
Center in Sweden), HIVE supercomputer at SWEGRID, and the HELIOS su-
percomputer provided by UNICC at Chalmers are gratefully acknowledged.
References
1. T. W. Chiu and L. C. Squire. An experimental study of the flow over a train
in a crosswind at large yaw angles up to 90o . Journal of Wind Engineering and
Industrial Aerodynamics, 45:47–74, 1992.
2. J. M. Copley. The three-dimensional flow around railway trains. Journal of Wind
Engineering and Industrial Aerodynamics, 26:21–52, 1987.
3. D. Wu. Predictive prospects of unsteady detached-eddy simulations in industrial
external aerodynamic flow simulations. Diploma thesis. matriculation number:
219949, Lehrstuhl fur Strömungslehre und Aerodynamishes Institute, Aachen,
Germany, 2004.
Use of Immersed Boundary Technique in a
Cartesian LES solver to study wake flows
J. Bodart, R. Giammanco, P. Rambaud and C. Benocci
Von Karman Institute for fluid Dynamics

Chaussee de Waterloo, 72
B-1640 Rhode-Saint-Genèse, Belgium
mioma@vki.ac.be
Summary. In the present contribution a pre-existing Large Eddy Simulation code

(MiOma) based on finite differences and Cartesian structured grids has been ex-
tended through the use of the Immersed Boundary Technique. The latter allows
the simulation of complex geometries far beyond the ones originally allowed by the
multi-domain technique present in MiOma and opens the way to study the wake
behind complex bluff bodies, using a starting point the circular cylinder at Re = 300.
Key words: LES, Immersed Boundary Technique, Wake Flow, Parallelism, Finite
Differences
Introduction
The Immersed Boundary Technique (IBT, [7]) aims to simulate efficiently

generic bodies immersed in a fluid flow using completely Cartesian structured
grids, contrary to the classical solid body fitted approach handling complex
geometries, the latter requiring usually unstructured meshes or at least the
recasting of the governing equations into the frame of curvilinear coordinates.
An interesting benefit of the IBT is as well the possibility of simulating a mov-
ing object without re-meshing, quite a noticeable feat for a simple Cartesian
solver.
The solid body will be immersed in the grid by the means of a forcing
term f IB in the momentum equation for an incompressible fluid (1) after non-
dimensionalization. Of course, any transport equation, including the energy
equation, can be solved using this approach.
∂Ui ∂Ui Uj ∂P 1 ∂ 2 Ui
+ =− + + fiIB (1)
∂t ∂xj ∂xi Re ∂xj xj
Principal advantages of the IBT includes rather inexpensive user inter-
vention for setting up a new simulation (see CAD to CFD passage in 2.1),

458 J. Bodart, R. Giammanco, P. Rambaud and C. Benocci
possibility as stated above of handling moving geometries without re-meshing,

and also the ability to directly solve coupled problem such as fluid-structures
interaction or energy transfer through a fluid/solid interface.
1 Overview of the MiOma Solver for Incompressible

Flows
The IBT can be introduced with relative ease in a pre-existing solver, thanks
to the features of the host solver. MiOma, the solver in question, belongs to a
general frame-work for LES computations recently developed at the authors’
institution[2]. Based entirely on (FLOSS 1 ), the framework uses state-of the
art software libraries and supporting tools, it is natively parallel and supports
static load-balancing and domain auto-partitioning.
The proper LES code consists of a pre-processor for grid, matrix genera-
tion and flow initialization; in a solver for the solution of the filtered Navier
Stokes equations, and in a post-processor for the extraction of the computed
statistics. The three components are strictly coupled and interact by means of
data written in NetCDF, an industry standard data format. The solver charac-
teristics include: finite differences for spatial discretization, staggered grid cell
arrangement[3], multi-domains for moderately complex geometries, grid con-
formity across the domains, fractional time step method, explicit treatment
of the viscous terms, Adam-Bashforth time stepping for the predictor step,
basic Smagorinsky and WALE2 techniques for SGS modeling with implicit
filtering provided by the finite differencing.
Starting from (1), the filtered Navier Stokes equations are obtained in (2)
a
∂U i ∂U i ∂U i U j ∂P 1 ∂2U i ∂τij
=0 ; + =− + − (2)
∂xi ∂t ∂xj ∂xi Re ∂x2j ∂xj
where P is the modified pressure including the isotropic part of the SGS
a
stress tensor τij = Ui Uj − U i U j , so that only its anisotropic part τij needs
to be modeled. As well known[2], the modified pressure P is only a numerical
expedient needed to insure the divergence free of the solution at the end of
each time-step advancement. In the MiOma formulation, the predictor step
assumes the form seen in (3), where the modified pressure is taken for its value
Pn at the previous time-step; the corrector step (4) makes use of a differential
modified pressure dP whose purpose is to insure a divergence free field at time
step n + 1 and that analytically can be expressed as a Laplacian equation as
for (5).
U˚i − U i
n
∂Pn n n
=− + Hj (U i , U j ) (3)
∆t ∂xi
1
Free/Libre and Open Source Software
2
Wall Adapting Local Eddy viscosity
Immersed Boundary Technique 459
− U˚i
n+1
Ui ∂dP
=− (4)
∆t ∂xi
1 ∂ U˚i
= ∇2 dP (5)
∆t ∂xi
Due to (4), the boundary conditions for the modified pressure dP are homo-
geneous Neumann on all the non immersed boundaries where the predicted
velocity is assumed the same as the corrected one. After (5), the modified
pressure is updated: Pn+1 = Pn + dP.
In MiOma the linear system associated to (5) is built on a cell-by-cell
basis dynamically at the initialization phase of the solver[2]: the construction
is done in a typical direction splitting approach. This characteristic will be
exploited in the future when the IBT implemented in MiOma will be extended
to pressure equation as well.
2 Overview of the IBT implemented in MiOma
One interesting feature of the IBT is the direct link between CAD3 and the
CFD environments; indeed, the solid can be directly created in a CAD soft-
ware and read by the CFD software. To take into account solid location, the
pre-processor analyzes the CAD generated geometry file and distinguishes
cells as belonging to the solid, the fluid, or the boundary in between. This
procedure can be performed at the pre-processor stage in case of absence of
relative motion of the body with respect to the reference frame: for a moving
body the procedure hereafter described should be moved inside the solver.
2.1 How to handle a general body in a CFD software
There exist three ways to treat a 2D closed surface in a three dimensional en-
vironment, such as analytical functions, NURBS 4 , or discrete representation,
through a discrete set of points and a surface triangulation.
Based on the suggestions of previous studies[6, 7, 4], the last method was
chosen, through the use of the STL file format: being a de facto standard, the
format is almost always available when exporting geometries from any CAD
software. The STL or file format is an ASCII or binary file format, containing
a list of the triangular surfaces that describe a solid model plus some auxiliary
data. The main advantages to deal with such a triangulated geometry is the
capability to build general procedures.
3
Computer Aided Design
4
Non Uniform Rational B-Splines
2.2 Static procedure (Preprocessor stage)
To define cell attributes (Figure 1), it is necessary to know whether or not

each of the cell nodes (corner), is located in the solid. This procedure has been
achieved in MiOma using built-in algorithms provided by FOSS 5 library GTS
[1]. Well known ray-triangle intersection algorithm is used in combination
with a random ray coming from the tested point. After counting the number
of intersections with all the triangles defining the closed surface, the parity
determines whether the point is in or out of the solid, as sketched in Figure
1(b).
Finally, according to the number of nodes lying in the solid (0, 6 or an
intermediate value), the cell attribute can be defined respectively as fluid,
solid, or boundary, Fig.1(a).
2.3 Dynamic procedure (Solver stage)
As explained before, each IBT can be characterized by a forcing term added to

the considered transport equation. Among the large number of existing meth-
ods, reviewed by Iaccarino and Verzicco [6], the so called direct forcing has
been implemented in MiOma. In the context of fractional step method(section
1), each velocity component (computed using a staggered arrangement) lying
inside the solid is set to zero after the predictor step stage.
Next is a boundary reconstruction procedure to take into account the in-
terface location: a linear interpolation stencil [7] is built to correct all velocities
components computed in a boundary cell (section 2.2) and outside of the solid,
Figure 1(a).
At this stage it has to be stressed that no special treatment is applied
at the immersed boundary to compute the modified Poisson equation at the
corrector step.
In(3) Out(2)
Not modified
Corrected
In(1)
Set to 0
Fluid cell
Solid cell
Boundary cell
(a) Cell attributes and related IBT approach to (b) Ray-surface: number of
the velocity components according to the stag- intersection’s parity deter-
gered arrangement mine in or out position
Fig. 1.
5
Free and Open Source Software
Immersed Boundary Technique 461
3 Investigation of the flow around a circular cylinder,

Re = 300
Several test cases introducing cylindrical geometries have been run using IBT,
including torus and sphere. Investigation of the flow around a circular cylinder
at low Reynolds number is presented here.
As described in [8], spanwise instabilities, known as B-instabilities (in op-
position to A-instabilities at lower Reynolds numbers) are expected in this
transitional range of Reynolds number. A 3D flow field, with streamwise struc-
tures in addition to the classical Von Karman street of spanwise structures
should be observed. Spanwise wavelength λy ' D has been proposed by [8]
at this Reynolds number.
The computational domain presented in Figure 2(a) has been arranged
using 200 cells in the streamwise direction, 64 cells in the spanwise direction
and 160 cells in the normal direction with a uniform area near the body. In
this area square cells (normal to the spanwise direction) are characterized
by a side length of 0.04D. Boundary conditions are uniform velocity inlet,
periodicity in the spanwise direction and outlet downstream of the body.
As presented in Figure 2(b), coherent structures are obtained in the wake of
the cylinder. In the spanwise direction, 7 couple of contra-rotative structures
are observed which is in agreement with the 6 expected according to the
instability wavelength, given the strong dependency of the spanwise length
when using periodicity conditions.
Statistics have been gathered during 21 typical shedding after the removal
of initial transients. (approximately 100 hours on 8 nodes6 of the VKI cluster),
Figure 3. Small differences are observed in the near wake, with a lower mean
velocity value and an overestimation of the fluctuation intensity, but excellent
predictions are noticed for the wake behavior further downstream the body.
4.5D
uniform area
14D 3.5D
5D 17D
(a) Computational domain (cut), spanwise (b) Q isocontour around a circular

length equal to 6D cylinder (B and W for streamwise
vorticity sign)
Fig. 2. Circular cylinder, Re = 300
6
Pentium IV , 3.4GHz
2 0.4
Spectral Simulation (Mittal & Balachandar) Spectral Simulation (Mittal & Balachandar)
MIOMA simulation (with IB) MIOMA simulation (with IB)
X/D =1.2
1 0.2
X/D=1.2
X/D =1.5 0
0
X/D=1.5
2
u’ /U∞
u/U∞
X/D =2.0 -0.2
2
-1
X/D=2.0
X/D =2.5 -0.4
-2
X/D=2.5
-0.6
X/D =3.0
-3
X/D=3.0
-0.8
-2 -1.5 -1 -0.5 0 0.5 1 1.5 2 0
Z/D (normal direction) Z/D (normal position)
(a) Mean streamwise velocity in the (b) Time averaged streamwise velocity
wake fluctuations in the wake
Fig. 3. Circular cylinder at Re = 300. (Comparison with [5])
Conclusion
A promising method has been introduced to study flows around bluff bodies
using IBT. A succesful validation at low Reynolds number around a circu-
lar cylinder is obtained. It has to be underlined how easy and fast are the
CAD and meshing stages. Future implementation will include moving and/or
deforming object, as well as mesh refinement localised near the body.
References
1. GTS development team. Gts web page. http://gts.sourceforge.net/, 2006.
2. R. Giammanco. Numerical Study of Coherent Structures within a legacy LES
code and development of a new parallel Frame Work for their computation. PhD
thesis, von Kármán Institute for Fluid Dynamics, 2005.
3. F. H. Harlow and J. E. Welsh. Numerical calculation of time-dependent vis-
cous incompressible flow of fluid with free surface. The Physics of Fluids, 8, N.
10:2182–2189, 1965.
4. G. Kalitzin and G. Iaccarino. Toward immersed boundary simulation of high
reynolds number flows. CTR-Annual Research Briefs, pages 369–379, 2003.
5. R. Mittal and S. Balachandar. On the inclusion of thre dimensional effects in
simulations of two-dimensional bluff body wake flows. Proceedings of ASME
Fluids engineering division Summer Meeting, jun 22-26 1997.
6. R. Verzicco and G. Iaccarino. Immersed boundary technique for turbulent flow
simulations. Applied mechanics review, 56:331–341, 2003.
7. R. Verzicco and G. Iaccarino. Immersed boundary technique for large eddy sim-
ulation. In Von Karman Institute, editor, VKi Lecture series on Large Eddy
Simulation and related techniques: Theory and Applications, march 2006.
8. C.H.K. Williamson. Vortex dynamics in the cylinder wake. Annual review of
Fluid Mechanics, 28:477–539, 1996.
Stochastic generation of velocity fluctuation for
turbulent inflow and initial condition
M. Fathali1 , M. Klein2 , T. Broeckhoven3 , C. Lacor4 , and M. Baelmans5

1
Department of Mechanical Engineering, K.U.Leuven, Celestijnenlaan 300 A,
3001 Heverlee, Belgium Mani.Fathali@mech.kuleuven.be
2
Department of Energy and Powerplant Technology, TU Darmstadt,
Petersenstrasse 30, 64287 Darmstadt, Germany kleinm@ekt.tu-darmstadt.de
3
Department of Mechanical Engineering, Vrije Universiteit Brussel, Pleinlaan 2,
1050 Brussel, Belgium.tim@stro.vub.ac.be
4
Department of Mechanical Engineering, Vrije Universiteit Brussel,
chris.lacor@.vub.ac.be
5
Department of Mechanical Engineering, K.U.Leuven
Tine.Baelmans@mech.kuleuven.be
Summary. Direct or large eddy simulation of a turbulent flow field is strongly

influenced by its initial or inflow boundary condition. This paper aims to present a
new stochastic approach to generate an artificial turbulent velocity field for initial
or inflow boundary condition based on digital filtering. The generated velocity field
reproduces locally desired Reynolds-stress components and integral length scales
including cross-integral length scales. The method appears to be simple, flexible and
more accurate in comparison with previously developed methods. The accuracy and
performance of the method is demonstrated by Direct Numerical Simulation (DNS)
of a homogeneous turbulent shear flow. To assess the accuracy and performance of
the method, simulation results are compared with a reference simulation.
1 Introduction
It is well known that besides the mathematical model and numerical scheme,
proper specification of boundary and initial conditions is of great importance
in the representation of a physical system. From a theoretical point of view
an adequate set of boundary and initial conditions is necessary to achieve a
unique solution of the set of partial differential equations to be solved. This
issue, is even more critical in turbulent flow simulations, using DNS or Large
Eddy Simulation (LES) approaches. Since in these approaches more flow field
structures are captured, all the resolved scales must be quantified at the inflow
or initial conditions. Moreover, due to the chaotic character of the Navier-
Stokes equations, turbulent flow field simulations are strongly influenced by
their inflow or initial conditions [5]. Therefore, any uncertainty in these data

464 M. Fathali, M. Klein, T. Broeckhoven, C. Lacor, and M. Baelmans
can remain as a source of error during the simulation. Performing separate

numerical simulations (DNS or LES) to obtain inflow or initial conditions is
very expensive. The available experimental information on the other hand,
is often insufficient to fully reconstruct the necessary data. Therefore, a cer-
tain level of approximation is needed to generate initial or inflow boundary
conditions at reasonable cost. Stochastic approaches, to generate a random
turbulent flow field for initial or inflow condition, have been extensively con-
sidered as an efficient alternative with a plausible level of approximation [1, 3].
In this research a new method is presented, which generates a turbulent ve-
locity field with exact prescribed Reynolds-stresses and integral length scales
including the cross-integral length scales between different velocity compo-
nents thus leading to a more accurate description of the flow field structures.
The required statistical information for this method are Reynolds-stresses and
integral length scales of the flow field.
2 Improved approach to generate artificial turbulence
In this section, the new methodology for generating artificial turbulent veloc-
ity fields is explained. To improve the readability of the paper, mathematical
derivations are discussed for a two-dimensional velocity field which is consis-
tent with the considered test case. In the new method, each velocity com-
ponent u and v is formed by a linear combination of different uncorrelated
random fields fij with zero mean, i.e.,
u = a11 f11 + a12 f12 ,
v = a21 f12 + a22 f22 . (1)
The presence of field f12 in u and v establishes a cross-correlation between

these velocity components. It can be shown that the correlation functions
for the velocity field (1), are related to the autocorrelation functions of the
individual random fields fij as follow:
Ruu = (a11 )2 r11 + (a12 )2 r12
Rvv = (a21 )2 r12 + (a22 )2 r22 (2)
Ruv = (a12 )(a21 ) r12
where Rij are the correlation functions for the generated velocity components
and rij are the autocorrelation functions of the initial random fields fij , e.g.
r12 = hf12 (x)f12 (x0 )i where x0 = x+r. It can be shown that filtering an uncor-
related random field generates a correlation in the filtered field closely related
to the filter function [1]. More specifically, by choosing a gaussian filter, the
correlation function has also a gaussian shape with a variance related to the
filter width [1, 2]. Considering this property, each initial random uncorrelated
Stochastic generation of turbulence 465
field, fij in equation (1), can be generated by applying a gaussian filter, with
appropriate filter width, to an arbitrary white noise field ξij . Moreover, it is
assumed that in equation (1) the cross-term coefficients aij and aji are es-
tablished with a common structure fij . However, the sign of these cross-term
coefficients depend on the sign of the respective Reynolds-stress. This leads
to:
aij = sign(hui uj i).aji = ±aji . (3)
With these assumptions the following set of equations for the generated ve-
locity field is obtained:
r2 r2
u = a11 . exp(− 2 ) ◦ ξ11 + a12 . exp(− 2 ) ◦ ξ12 ,
σ11 σ12
r2 r2
v = ±a12 . exp(− 2 ) ◦ ξ12 + a22 . exp(− 2 ) ◦ ξ22 , (4)
σ12 σ22
where ◦ denotes convolution. Correlation functions for the generated velocity
field expressed by equation (4) are:
r2 r2
r r
2 π 2 π
Ruu = (a11 σ11 ). exp(− 2 ) + (a12 σ12 ). exp(− 2 ),
2 2σ11 2 2σ12
r2 r2
r r
π π
Rvv = (a212 σ12 ). exp(− 2 ) + (a222 σ22 ). exp(− 2 ), (5)
2 2σ12 2 2σ22
r2
r
2 π
Ruv = sign(huvi)(a12 σ12 ). exp(− 2 ).
2 2σ12
Therefore, Reynolds-stresses and integral length scales can be easily calculated
by respectively setting r = 0 and integrating with respect to r in the expression
for the correlation functions (5). These provide six linear algebraic equations
for six unknown filter widths σij and coefficients aij in equation (4).
3 Application of the new approach
In this section, the proposed method is examined by generating turbulent

initial conditions for homogeneous turbulent shear flow. To this end, differ-
ent simulations, starting from various turbulent initial conditions, are com-
pared with a reference simulation. The reference simulation is started from an
isotropic turbulent velocity with top hat energy spectrum over wavenumbers
16 < k < 32. Then, various turbulent initial conditions are generated, using
extracted statistical information at a well-developed stage of the reference
flow field in the self-similar period, i.e. τ ∗ . An overview of the main proper-
ties of different initial conditions used to initialize the simulations is given in
table 1. The simplest initial condition, i.e. IC1, consists of white noise with
Table 1. Properties of different turbulent initial conditions used for different simula-
tions. The imposed statistical quantities for each initial turbulent field are specified.
Initial hui uj i hui uj i lij lij

conditions where i = j where i 6= j where i = j where i 6= j
IC1 Yes No No No
IC2 Yes Yes No No
IC3 Yes No Yes No
IC4 Yes Yes Yes Yes
prescribed turbulent kinetic energy components huui, hvvi and hwwi. Initial
condition IC2 has complete Reynolds-stresses information generated based on
Lund’s transformation [3]. Initial condition IC3 has prescribed turbulent ki-
netic energy components huui, hvvi,hwwi and normal integral length scales
luu , lvv and lww , generated by applying appropriate gaussian filters on the
velocity field of IC1. Finally, initial condition IC4 is generated based on the
new approach and satisfies both Reynolds-stresses and integral length scales
lij , including cross terms.
In the considered temporal homogeneous turbulent shear flow, the im-
posed mean flow in streamwise direction x is a linear function of the normal
coordinate z:
U (z) = Sz, (6)
where S is the mean shear rate. For all simulations the kinematic viscosity is
ν = 0.001 and S = 10.0. All the flow field parameters are in arbitrary units
[4]. The fully compressible Navier-Stokes equations are solved (DNS), only for
the fluctuating part of the velocity field −→
u (−
→x , t),
→−
−
U (→x , t) = U (z)−→
ex + −
→
u (−
→
x , t), (7)
→−
−
where U (→ x , t) is the complete velocity field and −→
ex is the unit vector in
streamwise direction x. The set of Navier-Stokes equations are solved based on
a cell-averaged finite-volume approach. Convective and viscous fluxes are dis-
1
cretized on a uniform mesh with 1003 grid cells, with grid spacing 4x/(ν/S) 2
equal to 6, using a second-order discretization. For time integration an explicit
four-stage compact-storage Runge–Kutta scheme of second-order accuracy is
used.
Figure 1 shows the evolution of the normalized Reynolds-stress compo-
nents for these simulations. Comparing the evolution of these quantities in
figure 1 reveals that the flow field started from initial condition with pre-
scribed integral length scales, i.e. IC4 and IC3 are much closer to the reference
simulation. Moreover, using the initial velocity field IC4, which is generated
based on the new procedure, results in a more accurate flow field evolution
compared to the simulation with only prescribed normal integral length scale
IC3. The difference between the latter two cases can be better observed by
Stochastic generation of turbulence 467
Table 2. Total deviation for all simulations ; q = hui ui i, u0 = huui, w0 = hwwi.

p p
huui/qIC4 = 0.1311, hvvi/qIC4 = 0.0946, hwwi/qIC4 = 0.0639, huwi/u0 wIC4 0

=
0.4093.
IC4 IC3 IC2 IC1
huui huui
total error for q
/ q 1.0000 1.0117 3.9239 3.7668
IC4
hvvi hvvi
total error for q
/ q IC4
1.0000 1.1496 1.5848 1.1524
hwwi hwwi
total error for q
/ q 1.0000 1.0383 5.7538 6.0832
IC4
huwi huwi
total error for 0
u w 0 / u0 w0 IC4
1.0000 1.6597 2.1895 2.6312
considering their respective deviation from the reference simulation. The error
is determined as:
Z τ =end q
2
errtot (φ) = φ(τ ) − φref (τ ) dτ, (8)
τ =τ ∗
where φ represents the normalized Reynolds-stress components. Table 2 lists

the total errors for these simulations. Computing total deviations from the
reference simulation during computation time confirms that the most accurate
0.48 0.33
0.46 0.32
0.31
0.44
<vv>/q
<uu>/q
0.3
0.42
0.29
0.4
0.28
0.38 0.27
0.36 0.26
5 10 15 20 5 10 15 20
St (a) St (b)
0.32 0.6
0.31 0.5
0.3
<−uw>/u’w’
0.4
<ww>/q
0.29
0.3
0.28
0.2
0.27
0.26 0.1
0.25 0
5 10 15 20 5 10 15 20
St
(c) St (d)
Fig. 1. Flow field evolution of homogeneous turbulent shear flow, initialized with
different random p (b):hvvi/q; (c):hwwi/q; (d):−huwi/u0 w0 where
fields (a):huui/q; p
0 0
q = hui ui i, u = huui and w = hwwi. (—): reference simulation, (◦): IC4 or
new approach, (square): IC3, (.): IC2, (4): IC1.
results are obtained from initial condition where all the Reynolds-stresses and
integral length scales are prescribed, i.e. IC4. Listed in table 2 are total errors
for all simulations computed based on the equation (8) and normalized with
the total error of IC4.
4 Conclusion
A new stochastic method for generating an artificial turbulent velocity field

for DNS and LES has been developed and tested. This method generates a
fluctuating velocity field that satisfies both prescribed local integral length
scales, including cross-integral length scales and Reynolds-stresses. In this
method, first by filtering individual random white noise fields, different un-
correlated random fields with prespecified gaussian autocorrelation functions
are generated. Subsequently these random fields are combined with appropri-
ate coefficients to generate the final turbulent velocity field. The application
of the proposed method is inexpensive as compared to counterparts methods
and does not require periodic directions.
To examine the application of the proposed method, different direct nu-
merical simulations of a temporally developing homogeneous turbulent shear
flow have been performed. Turbulent velocity fields used to initialize the sim-
ulations contain different levels of statistical information. Comparing the sim-
ulation results with a reference simulation revealed that the new procedure
yields a velocity field with more complete spectral content and consequently
more accurate results.
References
1. Klein, M., Sadiki, A., and Janicka, J., “A digital filter based generation of inflow
data for spatially developing direct numerical or large eddy simulations,” Journal
of Computational Physics, Vol. 186, 2003, pp. 652–665.
2. Kempf, A., Klein, M., and Janicka, J., “Efficient Generation of Initial- and In-
flow Conditions for Transient Turbulent Flows in Arbitrary Geometries,” Flow,
Turbulence and combustion, Vol. 74, No. 1, 2005, pp. 67–84.
3. Lund, T. S., Wu, X., and Squires, K. D., “Generation of turbulent inflow data
for spatially-developing boundary layer simulations,” Journal of Computational
Physics, Vol. 140, 1998, pp. 233–258.
4. Rogers, M. M. and Moin, P., “The Structure of the Vorticity Field in Homoge-
neous Turbulent Flows,” Journal of Fluid Machanics, Vol. 176, 1985, pp. 33–66.
5. Dziomba, B. and Fiedler, H. E., “Effect of initial conditions on two-dimensional
free shear layers,” Journal of Fluid Mechanics, Vol. 152, 1985, pp. 419–442.
Study on Numerical and Modelling Error in
LES of a Channel Flow Using Explicit Filtering
Tellervo T. Brandt
TKK Helsinki University of Technology, Laboratory of Aerodynamics,

P.O. BOX 4400, FI-02015 TKK, Finland, tellervo.brandt@iki.fi
1 Introduction
It is well recognised that the numerical error involved in a large eddy simu-
lation (LES) applying low-order finite-difference schemes is large when com-
pared to the effect of sub-grid scale (SGS) model [1]. However, depending on
the applied SGS model and the choice of the numerical method, the modelling
error can be of the same size as the numerical error and the two error compo-
nents may interact with each other [2, 3]. In this paper, the error components
are studied a posteriori in a fully developed turbulent channel flow between
two infinite parallel walls using explicit filtering. In Reference [4], this case was
studied a posteriori using an approach based on implicit filtering [3], and here
the two approaches are compared. Previously, explicit filtering with smooth
filters was noticed to have a large effect on the simulations [5], and here the
effect is studied a posteriori to exclude the effect of numerics.
2 Approaches to A Posteriori Testing
To separate the numerical and modelling error, the so-called grid-independent

LES or fine-grid LES result [3, 2] is applied in the present study, and the total
error is divided into numerical and modelling error and to the effect of explicit
filter as
εnumerical error = ũgrid-indep. LES − ũ (1)
εmodelling error = ufiltered DNS − ũgrid-indep. LES (2)
εfiltering = uDNS − ufiltered DNS (3)
where ũ refers to an LES solution and u to DNS solution.
In the present study, to approach the grid-independent LES, the absolute
explicit filter width is kept constant while the grid resolution is increased
as done e.g. in [2]. As suggested in [6], only the non-linear term of the

470 Tellervo T. Brandt
Navier–Stokes equations is filtered explicitly. For SGS modelling, the stan-

dard Smagorinsky model is applied. Here, the value 0.085 is used for the
model coefficient CS . In the grid-independent LES, the effect of SGS model
should be preserved despite of grid refinement, and here, this is enforced by
keeping the model length scale ∆S constant as finer grid resolutions are ap-
plied. In the first a posteriori test, ∆S is equivalent to grid spacing of the
1/3
coarsest applied grid, ∆ = (∆x ∆y ∆z) , and in the second case, it is equiv-
alent to the explicit filter width applied on the coarsest grid, ∆f , i.e. twice
the grid spacing. Van Driest damping is applied in all simulations.
If implicit filtering is applied to obtain grid-independent LES, as done for
the present test case in [4], filtering is provided only via the SGS model and
∆S is kept constant as the grid is refined [3].
3 Applied Numerical Methods and Test Case

In the present test case, the Reynolds number based on the channel half-
height and the friction velocity is Reτ = 395. For spatial discretization, the
second-order central-difference scheme is applied on a staggered grid system,
and for time integration, a third-order, three-stage Runge–Kutta method is
applied. The non-dimensional mean-pressure gradient is fixed, and the fluctu-
ating pressure is solved from a Poisson equation.
The applied resolutions of the base test case (grid 1) and the two finer
grids (grid 2, grid 3), which are applied to approximate the grid-independent
LES and filtered DNS, are given in Table 1. If DNS performed on grid 3 is
compared to the data of [6], the obtained numerical error on grid 3 is small
and does not change the conclusions of the present study.
Table 1. Applied grid resolution in wall units, ∆+ , and extent of the computational
domain, l, in streamwise (x), spanwise (y) and wall-normal (z) directions
grid ∆x+ ∆z + ∆y + lx /h lz /h ly /h
LES 1 44 23 1.0, . . . , 20 6 3 1
fine-grid LES 2 22 12 0.5, . . . , 10 6 3 1
fine-grid LES 3 15 8 0.5, . . . , 10 3 1.6 1
wall units: x+ = Reτ x, where x is scaled by the channel half-height
In grid 1 case, the applied explicit filter is a fourth-order commutative

filter [7] with the width of two grid spacings. On the finer grids, the absolute
filter width is kept the same as that of grid 1 which means that the ratio of
filter width to the grid spacing is larger that in grid 1 case. All the filters are
smooth and filtering is performed in the physical space in all three coordinate
directions. It was verified that the SGS shear stress and the energy spectra of
the simulations with different resolutions are sufficiently close to each other.
Some differences naturally occur because of numerics.
Study on Numerical and Modelling Error in LES 471
4 A Posteriori Tests
On the left-hand side of Fig. 1, we have the mean-velocity profile obtained from
simulation on grid 1 and the corresponding finer-grid cases, and in addition,
the DNS from [6] and the grid 3 case with filtering but no SGS model, which
approximates a filtered DNS, are included. The difference between grid 2 and
grid 3 results is small and thus, grid 3 result is a fair approximation to grid-
independent LES. The error components (1) evaluated using these results
are depicted on the right-hand side of the figure. In all the simulations with
filtering, the shape of the velocity profile is almost the same, and it differs from
the DNS result. Since the change is visible already in the filtered DNS, the
shape of the profile is due to filtering and not due to modelling or numerics.
The large effect of filtering is visible also on the right-hand side of the figure
where the effect of filtering is larger than the total simulation error. The
difference between grid 3 case and the filtered DNS is small which indicates
that the effect of SGS modelling is small and this is visible also in the small
modelling error. Both modelling and numerical error are visible mainly in the
mean bulk velocity, and the error components are thus nearly constants in the
logarithmic layer. In addition, the numerical and modelling error are of the
opposite sign and they partially cancel out each other.
1.5
20
1
18
0.5
16
0
! U+
14
U+
12 !0.5
LES, grid 1 !1
10 fine!grid LES, grid 2 total
fine!grid LES, grid 3 filter
8 filtered DNS !1.5 modelling
DNS numerical
6 !2
10 100 10 100
y+ y+
Fig. 1. Left: Mean-velocity profile, ∆S = 0.5∆f = ∆1 Right: Related error

components ∆U +
In the earlier study using implicit filtering, similar counteraction of the nu-
merical and modelling error was noticed, and there the numerical and mod-
elling error were approximately of the same size [4]. In the present results,
both the error components are smaller, which is probably partly due to ex-
plicit filtering, since it reduces the effect of the Smagorinsky model [5], and
partly due to different resolutions in the wall-normal direction.
In Figure 2, the corresponding plots are given for the deviatoric streamwise
Reynolds stress. Here, the effect of modelling is even smaller than in the mean-
velocity profile, and most part of the error is formed by explicit filtering.
The numerical error is again of the opposite sign to the modelling error and
thus, they partially cancel out each other. The error components are similar
for the other diagonal Reynolds stresses. In the earlier study with implicit
filtering, the obtained modelling error was larger than here, the distribution
of the numerical error somewhat different, and counteraction of the error
components was not visible there.
0 2
!1
1
!2
!3 0
! u’u’*
!u’u’*
!4
!1
!5
LES, grid 1
!6 fine!grid LES, grid 2 !2 total
!7 filtered DNS modelling
!3
DNS numerical
!8
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2 0.25
y/h y/h
Fig. 2. Left: Deviatoric Reynolds stress, ∆S = 0.5∆f = ∆1 Right: Related

error components ∆ − u0 u0∗
Previously, it was noticed that as the length scale of the Smagorinsky

model is increased, the results with explicit filtering slightly improve [5]. Here,
the a posteriori tests were repeated using a model length scale which was set
proportional to the filter width, i.e. it had twice the value when compared
to the first test case discussed above. The plots depicting the mean-velocity
profile and the involved error components are given in Fig. 3. Increasing the
model constant naturally increases the effect of modelling when compared
to the curves of Fig. 1. In the mean-velocity profile, this is visible in the
increased mean bulk velocity and increased thickness of the viscous sublayer.
The change in the viscous sublayer also slightly improves the slope of the
profile. As the model length scale is increased, the total error in the mean-
velocity profile decreases in the middle of channel and in the end of the viscous
sublayer. However, in the lower part of the logarithmic layer, it increases. On
the right-hand side of Fig. 3, we see that this is caused by the interaction
of numerical and modelling error. The increased model length scale affects
mainly the modelling error. Since the two error components are of the different
sign, this leads to decreased total error in some parts of the channel. The
effect of the increased model length scale on the numerical error is small, and
in some parts of the channel the numerical error even appears to increase
slightly when compared to Fig. 1 although the SGS model should smooth out
the resolved flow field. Similar behaviour was also visible in the study with
implicit filtering [4].
The corresponding plots for the streamwise deviatoric Reynolds stress are
given in Fig. 4. Here, the grid convergence is not as fast as with the lower model
length scale in Fig. 2. This behaviour was also noticed in simulations with
Study on Numerical and Modelling Error in LES 473
1.5
20
1
18
0.5
16
0
! U+
14
U+
12 !0.5
LES, grid 1 !1
10 fine−grid LES, grid 2 total
fine−grid LES, grid 3 filter
8 filtered DNS !1.5 modelling
DNS numerical
6 !2
10 100 10 100
y+ y+
Fig. 3. Left: Mean-velocity profile, ∆S = ∆f = 2∆1 Right: Related error

components ∆U +
implicit filtering [4]. Increasing the model length scale improves the prediction
of the peak value of the Reynolds stress, but at the same time, it widens the
distribution. This is visible also in the total error which decreases in the near-
wall region but increases after the peak of the Reynolds stress when compared
to the case with lower model constant in Fig. 2. However, changes in the total
error are mainly due to the counteraction of the numerical and modelling
error. Increasing the model length scale increases the effect of modelling and
the modelling error, and now, modelling error is larger than the numerical
error, but they are still of different sign. When compared to the case with
the smaller model length scale, the numerical error clearly increases, and the
increase is stronger than in the mean-velocity profile. Similar behaviour was
noticed in the study with implicit filtering [4]. However, there the numerical
error of the Reynolds stress increased, as it should, also when the model length
scale was reduced to 0.5∆.
0 2
!1
1
!2
!3 0
! u’u’*
!u’u’*
!4
!1
!5
LES, grid 1
!6 fine!grid LES, grid 2 !2 total
!7 filtered DNS modelling
!3
DNS numerical
!8
0 0.05 0.1 0.15 0.2 0 0.05 0.1 0.15 0.2 0.25
y/h y/h
Fig. 4. Left: Deviatoric Reynolds stress, ∆S = ∆f = 2∆1 Right: Related error

components ∆ − u0 u0∗
5 Conclusions
In this paper, a posteriori tests were performed using explicit filtering in the
channel flow. The numerical error related to the second-order scheme and the
modelling error of the standard Smagorinsky model were approximately of
the same size and of different sign. When compared to results obtained with
implicit filtering [4], the results were similar except for the effect of explicit
filtering itself. Filtering reduced both error components. In the present results,
the large effect of filtering with a smooth filter was demonstrated at fine grid
resolution, and this verifies that the strong effect of filtering obtained in [5]
was not due to numerics.
The a posteriori tests were repeated with a larger model length scale. This
showed that the slight improvement of simulation results with this model
length scale that was noticed in [5] was due to interaction of the error com-
ponents. The increased modelling error cancelled out the numerical error.
In the study with implicit filtering, some increase of numerical error with
increasing model length scale was visible in the mean-velocity profile and in
the Reynolds stresses, and this trend was noticed also here. This behaviour
suggests that in these a posteriori studies, part of the effect of modelling is
included in the numerical error as the model length scale is increased.
References
1. S. Ghosal. An analysis of numerical errors in large-eddy simulations of turbulence.
2. B. Vreman, B. Geurts, and H. Kuerten. Comparison of numerical schemes in
large-eddy simulation of the temporal mixing layer. International Journal for
Numerical Methods in Fluids, 22:297, 1996.
3. B. Geurts and J. Fröhlich. A framework for predicting accuracy limitations in
large-eddy simulation. Physics of Fluids, 14(6):L41–L44, June 2002.
4. T. Brandt. A posteriori study on modelling and numerical error in LES applying
the Smagorinsky model. In Complex Effects in LES, Lecture Notes in Computa-
tional Science and Engineering. Springer, 2006. Accepted for publication.
5. T. Brandt. Study of large eddy simulation and Smagorinsky model using explicit
filtering. In 36th AIAA Fluid Dynamics Conference and Exhibit, San Francisco,
California, June 5-8 2006. AIAA-2006-3541.
6. T. S. Lund. On the use of discrete filters for large eddy simulation. Center
for Turbulence Research, Annual Research Briefs, pages 83–95, 1997. Stanford
University.
7. O. V. Vasilyev, T. S. Lund, and P. Moin. A general class of commutative filters for
LES in complex geometries. Journal of Computational Physics, 146(1):82–104,
October 1998.
8. R. D. Moser, J. Kim, and N. N. Mansour. Direct numerical simulation of tur-
bulent channel flow up to Reτ = 590. Physics of Fluids, 11(4):943–945, April
1999.
Wall Boundary Conditions for Variational
Multiscale Large-Eddy Simulations
S. J. Hulshoff and E. A. Munts
Faculty of Aerospace Engineering, Delft University of Technology,

Kluyverweg 1, 2629 HS Delft, The Netherlands
S.J.Hulshoff@LR.TUDelft.NL
1 Introduction
Variational multiscale discretisations for turbulent flows, first introduced by

Hughes [1] have been shown to produce accurate results with relatively simple
modelling assumptions. A particularly flexible discontinuous Galerkin variant
has been developed by Collis [2]. Continuous Galerkin discretisations, how-
ever, can be advantageous in that their low implicit dissipation simplifies the
calibration of subgrid-scale models. Recent work has indicated that combining
continuous methods with weak boundary conditions can result in improved
accuracy [4], while adding some of the flexibility inherent in discontinuous
methods. We consider a continuous-Galerkin variational-multiscale method
combined with penalty-based weak boundary conditions, and describe the re-
lation of the latter to traditional wall-stress models. The performance of the
weak conditions is compared to that of hard conditions for both single and
multiscale computations of low-Mach number turbulent channel flows.
2 Variational Multiscale Discretisation
The compressible Navier-Stokes equations can be written

U,t + Fi,i (U) − Fvi,i (U) = S, (1)
where U is the vector of conservative variables, Fi (U) and Fvi (U) are the
inviscid and viscous fluxes and S is the constant source vector required for
driving channel flows simulated with periodic streamwise conditions. Intro-
ducing a trial space Yn and test space Wn , the variational formulation can
then be stated as follows: Find Y ∈ Yn such that ∀ W ∈ Wn

− (W,t , U(Y))Qn − W,i , Fi (Y) − K e ij Y,j
Qn
+ (W, (Fi (Y) − Fvi (Y)) ni )Pn + (W(tn+1 ), U(Y(tn+1 )))Ωn+1

476 S. J. Hulshoff and E. A. Munts
− (W(tn ), U(Y(tn )))Ωn = (W, S)Qn . (2)

Here (, )Ωn denotes the L2 -inner product over region Ωn , Ωn is the spatial
domain at time tn , Qn is the portion of space-time domain between time
levels tn and tn+1 , and Pn is the surface connecting the boundary of Ωn
with that of Ωn+1 . ni is the local space-time surface normal vector, and K e ij
v
are the diffusivity matrices defined by Fi (U) = (K e ij Y,j ),i . For the present
calculations, quadratic interpolation is used in space, and linear interpolation
with constant weighting functions is used in time (see [3]).
Single-scale large-eddy simulations can be performed by directly adding a
subgrid scale model to (2). To construct a variational multiscale method, (2)
can be expressed compactly as:
B(W, U) = (W, S). (3)
If we decompose the trial and test space in to large resolved scales (bar), small
resolved scales (tilde) and unresolved scales (hat), we can write the following
equations for the resolved scale ranges:
Large: B(W, U + U
e + U)
b = (W, S) (4)
Small: B(W,
f U+U e + U)
b = (W,
f S) (5)
where for example, the large-scale equation can be expanded as:
B(W, U) + B 0 (W, U, U)
e − R(W, U)
e = (W, S)
− B 0 (W, U, U)
b + R(W, U)
b + C(W, U,
e U),
b (6)
0 0
where B (W, U, U ) is the operator B(W, U) linearised about U for a linear
perturbation U0 , and C(W, U, e U)
b and R(W, U) e are generalised cross and
Reynolds stress projections onto the large scales.
It is assumed that interactions only occur between scales of like size, so
that the unresolved scale terms in the large-scale equation can be neglected.
Furthermore, it is assumed that the terms involving unresolved scales in the
small-scale equation can be replaced by a single model term. This results in:
B(W, U) + B 0 (W, U, U)
e − R(W, U)
e = (W, S), (7)
B 0 (W, e − R(W,
f U, U) e = −[B(W,
f U) f U) − (W,
f S)] + M(W,
f U).
e (8)
A simple constant-coefficient Smagorinsky model is used for M(W,
f U).
e
3 Application of boundary conditions

For the channel flows considered here, periodic conditions are used in the
spanwise and streamwise directions. This leaves only solid-wall conditions to
be applied on the upper and lower boundaries.
If primitive variables Y = {ρ, u, T }T are used for interpolation. hard no-
slip and isothermal conditions can be applied by setting the weights of the
Wall Boundary Conditions for Multiscale LES 477
momentum and energy equations on the wall to zero, and explicitly constrain-
ing the velocity and temperature variables. The value of density can be left
free to be determined by the numerical solution.
As described in reference [4], weak solid-wall conditions can be based on
the adjoint-consistent form of the penalty method introduced by Nitsche (see
[5] for details). As the near-wall flows considered here are essentially incom-
pressible, such conditions can be constructed by adding the following two
terms to the momentum components of (2):
− (2µ∗ ∇s W · n, U − g)Pn + (Wαµ∗ /h, U − g)Pn (9)
∗
Where g is the desired boundary state, µ = µ + µm is the sum of the physical
and model viscous coefficients, h is a measure of the element size and and α is
a positive parameter. In addition we explicitly constrain the normal velocity
to zero and the wall temperature to its specified value, and leave the value of
density free as in the hard condition procedure.
4 Relation to wall modelling
At higher Reynolds numbers it becomes impractical to resolve near-wall tur-

bulent structures. For large-eddy simulations, wall-layer models are often used
as an alternative. A common approach is to supplement the boundary con-
ditions with wall stresses obtained from other sources, for example equations
based on law of the wall assumptions or thin boundary-layer computations.
One of the first successful applications in this vein was by Schumann [6] who
directly related the wall stresses to velocity values at a height, h above the
wall, located in the logarithmic layer, using
< τw > w(x, h, z)
τxy (x, 0, z) = u(x, h, z), τyz (x, 0, z) = µ (10)
< u(x, h, z) > h
where the <> denotes time-averaged quantities.
The weak condition with g fixed to zero has some characteristics in com-
mon with (10). This can be demonstrated by noting W = 1 is included in
the test space of the discretisation, and considering the time-averaged x-
momentum component of (2) + (9):
((− < Fvi (Y) >)ni )Pn + (αµ̂/h < U − g >)Pn = (S)Qn . (11)
In fully-resolved simulations of channel flow, the steady body force S is bal-
anced by the mean shear stress on the solid walls. In coarse LES discretisations
employing the weak condition, the second term in (11) will on average make
up the difference between the resolved viscous shear stress and the body force,
so that the correct mean viscous stress is obtained. This effect is also produced
in discretisations using (10) and the exact value for < τw >. Like (10), the
weak boundary condition will also provide an instantaneous increase in the ef-
fective wall stress for a positive velocity excursion, although the position of the
excursion measurement is adjacent to the wall rather than in the logarithmic

layer.
In contrast to the weak condition, the hard condition does not contain
a corrective term which ensures x-momentum conservation. As the wall test
functions for momentum are set to zero, x-momentum conservation is only
realised approximately via the resolved shear stress. The lack of test functions
on the boundary also makes the addition of external wall-stress models less
convenient.
5 Results for Reτ = 180
We first consider a low Reynolds number Reτ = 180 case with a standard
level of LES resolution. For the spatial discretisation, a continuous hierarchical
basis with linear and quadratic element modes was used on a 16 × 16 × 16
mesh. The element geometries were stretched in the wall-normal direction
using a hyperbolic tangent function. The flow conditions for this case and all
subsequent cases have been scaled so that the maximum Mach number is less
than 0.3.
Figure 1 compares RMS velocity profiles from a single-scale method with
with van Driest damping (SvD) and a variational-multiscale method (VM)
with data obtained from direct numerical simulation (DNS) [7]. Results for
both hard (H) and and weak (W) solid-wall conditions are shown. For the
single-scale method, van Driest damping is required to achieve acceptable re-
sults on the current mesh. In contrast, the VM produces good results with only
a constant-coefficient model. As both results are relatively resolved, however,
the impact of weak boundary condition is minimal. In this case the first inter-
element boundary is located at y + = 5.9, which provides sufficient resolution
of the near-wall gradient when quadratic interpolation is used.
6 Results for Reτ = 590
We now consider a higher Reynolds number, Reτ = 590, with a coarser LES
resolution. The same quadratic basis as used previously is employed on a
8 × 8 × 8 mesh. No stretching is used, so that the first intra-element interface
is located at y + = 74.
Both the SvD and VM methods with hard conditions underpredict the wall
velocity gradient (figure 2, left). In the case of the VM, this results in a higher
Wall Boundary Conditions for Multiscale LES 479
3.5
SvD H 1.4 SvD H
SvD W SvD W
3 VM H VM H
VM W 1.2 VM W
DNS DNS
2.5
1
2
0.8
wrms
urms
1.5
0.6
1 0.4
0.5 0.2
0 0
0 50 100 150 200 0 50 100 150 200
+
y y+
Fig. 1. Reτ = 180: Streamwise and spanwise velocity fluctuations
30 3.5
VM H
VM W
3 DNS
25
2.5
20
Mean Velocity
2
urms
15
1.5
10
1
SvD H
5 SvD W
VM H 0.5
VM W
DNS
0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
+ +
y y
Fig. 2. Reτ = 590: Streamwise mean velocity and fluctuation
mean velocity. In the case of SvD, the coarse mesh appears to accentuate
the effect of the dissipative model on the large scales, resulting in a reduced
velocity profile. As in the previous case, the use of weak boundary conditions
does not significantly benefit the SvD results. As can be seen in figures 2 and
3, however, the VM results are strongly improved. The mean flow prediction
is far better, particularly at large y + , as is required for a good wall model.
The fluctuating velocity components are also improved at large y + .
2.5
1.4 VM H VM H
VM W VM W
DNS DNS
1.2 2
1
1.5
0.8
wrms
vrms
0.6 1
0.4
0.5
0.2
0 0
0 100 200 300 400 500 600 0 100 200 300 400 500 600
+
y y+
Fig. 3. Reτ = 590: Normal and spanwise velocity fluctuations
Concluding remarks
The combination of a continuous variational multiscale discretisation with

penalty-based weak boundary conditions appears to offer several benefits, in-
cluding improved accuracy and a potential framework for applying external
wall-stress models.
References
1. Hughes, T. J. R., Mazzei, L. and Jansen, K. E., Large Eddy Simulation and the
Variational Multiscale Method. Comput.Visual Sci. 3 47–59 (2000)
2. Collis, S. S., The DG/VMS Method for Unified Turbulence Simulation. AIAA
Paper No. 2002-3124 (2002)
3. Munts, E., Space-time multiscale methods for large-eddy simulation. Ph.D. The-
sis, Faculty of Aerospace Engineering, Delft University of Technology (2006)
4. Bazilevs Y., and Hughes, T. J. R., Weak Imposition of Dirichlet Boundary Con-
ditions in Fluid Mechanics, ICES Report 05-25, University of Texas at Austin
(2005)
5. Arnold, D. N., Brezzi, F., Cockburn, B, and Marini, L. D., Unified Analysis of
Discontinuous Galerkin Methods for Elliptic Problems. SIAM J. Numer. Anal.,
Vol 39, No 5, pp 1749-1779 (2002)
6. Schumann, U., Subgrid Scale Model for Finite Difference Simulations of Turbu-
lent Flows in Plane Channels and Annuli. J. Comp. Physics, 18 376–404 (1975)
7. Moin, P., and Kim, J., Numerical investigation of turbulent channel flow. Journal
of Fluid Mechanics, 118, 341–371 (1982)
The Sampling Based Dynamic Procedure for
Numerical Discretization Enhancement
Dieter Fauconnier, Chris De Langhe, and Erik Dick
Ghent University, Department of Flow, Heat and Combustion Mechanics,

St.-Pietersnieuwstraat 41, B-9000 Ghent, Belgium
Dieter.Fauconnier@Ugent.be, Chris.Delanghe@Ugent.be, Erik.Dick@Ugent.be
1 Introduction
Recently, a new sampling formalism for Large Eddy simulation was proposed
by Winckelmans et al. [1] and Knaepen et al. [2], which is a projection method
for Navier-Stokes equations from continuum space to a discrete space, using
a sampling operator instead of a filter operator. Since the sampling operator
is not commutative with spatial derivatives, a closure term appears which
represents the loss of information due to the projection on a discrete mesh. In
e.g. [2] a Smagorinsky model was proposed that, by relying on a generalized
dynamic procedure, succeeded in accounting for the subgrid scales. In this
paper, we investigate the ability of this sampling based dynamic procedure, in
combination with an appropriate model for the truncation error, to obtain
higher-order numerical accuracy. Two such possible models are presented.
Further, we show that Richardson extrapolation is a simplified formulation of
this procedure.
2 The Sampling Formalism
We define the sampling operator S∆1 , which operates between the continuum
space Ω ⊂ IRn and the discrete space Ω ∆1 with N number of grid points,
and spacing ∆1 . This sampling operator S∆1 is idempotent, and commutative
with the product of the non-linear terms. However, S∆1 does not commute
with spatial derivatives. We use the notation S∆1 ◦ ui = ui and S∆1 ◦ ∂ = δ.
Applying S∆1 to the continuity equation and the momentum equations gives
δui
= Π ∆1 (1)
δxi
∂ui δui δp δ 2 ui
+ uj =− + ν 2 + Σi∆1 (2)
∂t δxj δxi δxj

482 Dieter Fauconnier, Chris Delanghe, Erik Dick
The truncation errors are due to the non-commutativity of the operator S∆1
with the spatial derivatives, and have the basic form
δui ∂ui
Π ∆1 = − (3)
δxi ∂xi
!
δ 2 ui ∂ 2 ui

δui ∂ui δp ∂p
Σi∆1 = uj − + − −ν − (4)
δxj ∂xj δxi ∂xi δx2j ∂x2j
3 The generalized dynamic procedure

Projection is done of the Navier-Stokes equations from a continuum domain
Ω to a corresponding discrete domain Ω ∆1 , and to Ω ∆2 with the number
of gridpoints N2 < N1 (Ω ∆2 ⊂ Ω ∆1 ). This corresponds with the sampling
operators S∆1 and S∆2 , projecting respectively Ω → Ω ∆1 and Ω → Ω ∆2 . S∆2
also projects Ω ∆1 → Ω ∆2 , since S∆2 ◦ S∆1 = S∆2 . We introduce S∆2 ◦ ui =
u ei . We keep the notation for the discrete derivative-operator S∆2 ◦ ∂ = δ.
ei = u
Applying the operator S∆1 on the continuous set of equations leads to
δui
0 = C∆1 (ui ) + Π ∆1 = − + Π ∆1 (5)
δxi
∂ui δui δp δ 2 ui
= Ni∆1 (ui ) + Σi∆1 = −uj − + ν 2 + Σi∆1 (6)
∂t δxj δxi δxj
C∆2 and Ni∆2 are called here the continuity and Navier-Stokes operators,
respectively. Applying S∆2 to the continuous set equations gives
δe
ui
0 = C∆2 (e
u i ) + Π ∆2 = − + Π ∆2 (7)
δxi
∂e
ui δe
ui δep δ2 u
ei
= Ni∆2 (e
ui ) + Σi∆2 = −e
uj − + ν 2 + Σi∆2 (8)
∂t δxj δxi δxj
Ideally, the latter set should also be obtained by applying the sampling oper-
ator S∆2 to the first set of equations (5)-(6) and because of S∆2 ◦ S∆1 = S∆2 ,
following relation is imposed:
S∆2 ◦ C∆1 (ui ) − C∆2 (eu i ) = Π ∆ 2 − S∆ 2 ◦ Π ∆ 1 (9)
S∆2 ◦ Ni∆1 (ui ) − Ni∆2 (e
ui ) = Σi∆2 − S∆2 ◦ Σi∆1 (10)
These explicitly express the commutation errors made by the projection
Ω ∆1 → Ω ∆2 . This expression can be determined in terms of the resolved
velocity ui since S∆2 ◦ ui = u
ei = uei . Suppose models are adopted for both
truncation errors Π ∆1 and Σi∆1 with the basic forms
Π ∆1 = C π mπ,∆1 , Σi∆1 = Ciσ mσ,∆
i
1
(11)
π
and analogously for the test-level ∆2 . Then the fields C and can be ob- Ciσ
tained from relation (9)-(10).We suggest to use a least square method with
The Sampling Based Dynamic Procedure for Numerical Improvement 483
a smoothing filter, such as a local moving average or a global average. For

the latter averaging, constant fields are obtained. C π and Ciσ can be positive
or negative, dependent on the adopted discretization scheme. If excessive val-
ues jeopardize the algorithm’s stability, clipping may be necessary. Finally,
C π and Ciσ , given on the coarse grid, are interpolated to the fine grid using
a piecewise Cubic Hermite interpolation. A fully embedded test grid is applied.
δ n u ∆

Consider the Taylor series expansion of the nth discrete derivative δxn
on grid with spacing ∆, for a k th -order central discretization scheme
∆
∂nu δ n u k∂
k+n
u
+ O ∆k+2

(x) = + ck ∆ (12)
∂xn δxn ∂xk+n
2∆
∂nu δn u k ∂
k+n
u
+ O ∆k+2
e
n
(x) = n
+ ck (2∆) k+n
(13)
∂x δx ∂x
We assume now, that the leading order truncation term is an adequate model
for the complete truncation error. Discretization of this term and applying
the generalized dynamic procedure leads to a coefficient ck . Substitution in
(12) results finally in
n ∆ k+n 2∆
2∆ k+n ∆
δn u

k δ u δ u e δ u
−
e
2
∂nu δxn δxk+n δxn δxk+n
(x) ≈ (14)
k+n 2∆
∂xn ∆
δ k+n u

kδ u
−
e
2
δxk+n δxk+n
This expression is closely related to Richardson extrapolation. The Richardson
extrapolation formula is obtained for
2∆ ∆
δ k+n u
e δ k+n u ∂ k+n u
≡ ≡ (15)
δxk+n δxk+n ∂xk+n
th
which gives a (k + 2) -order accurate central scheme for the nth derivative.
For further analysis of expression (14), and the implications of approximation
(15), a Fourier analysis for a single wave is performed for the 2nd -order ac-
curate gradient. A blending factor f is introduced in order to switch between
the dynamic procedure (f = 1) and Richardson extrapolation (f = 0), and
to investigate intermediate behaviour. Modified wavenumbers of the resulting
expression are given in figure 1 for a 2nd , 4th (equals Richardson Extrapola-
tion f = 0) and 6th order central scheme, and also for the dynamic procedure
formula at values of f = 1, 12 , 13 , 14 , 15 , 10
1
. We noticed from a semilog error plot
1
that f = 5 is the value for which the dynamic procedure obtains best results.
For the lower wavenumbers the accuracy is slightly better than 6th -order, for
the higher wavenumbers it is slightly less.
1
0.9
0.8
0.7 f=1
f=1/2
0.6
f=1/3
f=1/4
!"/!max
0.5
f=1/5
f=1/10
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
!/!
max
Fig. 1. Modified wavenumbers for ∂u ∂x

: (. . .) 2nd -order central; (- . -) 4th -order
th
central; (- - -) 6 -order central; (—) Dynamic Procedure.
4 Modelling truncation errors

Instead of evaluating a truncation correction for every separate derivative in
the Navier-Stokes equations, we want to increase the accuracy of the dis-
cretization by using one model for all truncation errors together. For a 2nd -
order central difference discretization scheme for first and second order partial
derivatives, we suggest as a first model:
δ 3 ui
Π ∆1 = C π ∆x2i (16)
δx3i
!
3 3 4
δ ui 2δ p ν 2 δ ui
Σi∆1 = Ciσ uj ∆x2j + ∆xi − ∆xj (17)
δx3j δx3i 2 δx4j
This model is obtained from the leading order terms in the Taylor series ex-
pansion. The theoretical value of the factors C π and Ciσ is 61 . The major
drawback is the requirement of a broader 5-point stencil to evaluate the 3rd
and 4th derivatives. This can be very unpleasant near walls, where excentric
6-point stencils (also second order accurate) have to be constructed to main-
tain the overall accuracy. It would be more convenient if the higher order
derivatives could be reduced to maximum 2nd -order, the highest appearing
order in the physical Navier-Stokes equations.
The idea behind the reduction of the high order derivatives for modeling
purposes, is based on the observation that odd/even higher-order derivatives
appearing in the analytical Taylor series of a central finite difference approx-
imation of an odd/even-order derivative display a similar behaviour as the
derivative itself, although they do not have the same scaling. We therefore
propose the non-linear approximation
! !
δ n ui δ p ui δ q ui
ui n ≈ (18)
δxj δxpj δxqj
The Sampling Based Dynamic Procedure for Numerical Improvement 485
Code Continuity Momentum

ldc.2o60×60 2nd -order central
ldc.4o60×60 4th -order central
ldc.6o60×60 6th -order central
ldc.DRE60×60 Differential Richardson extrapolation (f = 0)
ldc.1a60×60 exact truncation with f = 15
ldc.1b60×60 exact truncation with f = 1
ldc.260×60 exact truncation no correction
ldc.360×60 no correction Smagorinsky
ldc.460×60 exact truncation Smagorinsky
Table 1. Summary of the different simulations
in which (p + q) = n. If n is even, p and q should both be even. If n is odd, then

p should be even and q should be odd, or vice versa. This reduction can be
applied repeatedly. If n = 3, and if we choose p = 1, q = 2, a Smagorinsky-like
modelling approximation appears for the convective terms in (17)
δ 3 ui uj δui δ 2 ui δui δ 2 ui

uj 3 ≈ ≈
(19)
δxj ui δxj δx2j δxj δx2j
For the truncation error of the continuity equation (16), the pressure terms
and the viscous terms in the truncation error of the momentum equations (17),
the foregoing hypothesis (18) does not lead directly to a suitable Smagorinsky-
like model. Nevertheless, we decide to use a Smagorinsky-like model only for
the convective terms in the truncation error of the momentum equations
2
∆1 σ∗ 2 δui δ ui

Σi = Ci ∆xj (20)
δxj δx2j
and so an additional approximation is made by giving all terms in the trun-
cation error model the same coefficient. Explicit corrections for pressure and
viscous terms in the momentum equations are neglected.
5 Results
The testcase is a laminar flow in a 2D driven cavity at Reynolds number
Re = 400. A pseudo-compressible code is used with a 3rd -order Runge-Kutta
method for stepping in pseudo time. Spatial discretization is 2nd -order central.
No artificial stabilisation is used in the continuity equation. Therefore, the
fields contain a minimal spurious pressure mode, that however does not affect
the velocity results. The exact truncation errors are (16)-(17). We use a 6th -
order central scheme on a 180 × 180 mesh, to generate the reference solution.
The different simulations are given in Table 1.
Using an exact truncation model in combination with the dynamic pro-
cedure, a global least-square averaging was chosen for C π en C σ . For the
! !
u v
0.025
0.02
0.02
0.015
0.015
0.01
0.01
0.005
0.005
0
0
−0.005
−0.005
−0.01
−0.01
−0.015
−0.015
−0.02
−0.02
−0.025
−0.025
−0.03
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Fig. 2. Error levels εu (x = L2 ) and εv (y = L2 ): (4) ldc.6o60×60 ; () ldc.4o60×60 ;

(◦) ldc.2o60×60 ; (O) ldc.DRE60×60 ; (×) ldc.1a60×60 ;(+) ldc.1b60×60 ; (–.–) ldc.260×60 ;
(- - -) ldc.360×60 ; (—) ldc.460×60 .
Smagorinsky-like model, we use a local least-square averaging over a 3 × 3

subdomain. Moreover, a clipping to C σ > −0.2 for excessive negative values
of C σ was necessary.
The results for the test sections x = L2 and y = L2 are given in figure 2. As ex-
pected, the dynamic procedure, with the exact truncation model (ldc.1a60×60 )
and f = 15 obtains approximately 6th -order accuracy. Results of ldc.1b60×60
with f = 1, are more or less equivalent. Differential Richardson extrapo-
lation (ldc.DRE60×60 ) obtains close to 4th -order accuracy. From simulations
ldc.260×60 , ldc.360×60 en ldc.460×60 , we can conclude first that correction of the
continuity equation is crucial for overall accuracy in the driven cavity (mainly
due to the large corner-pressure gradients of the driven lid). Secondly, we see
that the Smagorinsky model for the momentum equations leads to noticeable
improvement of accuracy, but as expected it is less performant than the exact
model.
Acknowledgement. The research of the first author was funded by a Ph.D
grant of the Institute for the Promotion of Innovation through Science and
Technology in Flanders (IWT-Vlaanderen).
References
1. G. Winckelmans, L. Bricteux, and L. Georges. The “sampling-based dynamic
procedure” for LES without filtering: validation using finite differences. In
E. Lamballais et al., editors, Ercoftac Workshop - Direct and Large-Eddy Sim-
ulations 6, pages 140–141, Poitiers, September 2005. Université de Poitiers. IN
PRESS.
2. B. Knaepen, O. Debliquy, and D. Carati. Large-eddy simulation without filter.
J. Comp. Phys., 205:98–107, 2005.
Application of a Unique Eddy–Viscosity Model
for a Hybrid LES–RANS Method
Michael Breuer and Benoit Jaffrézic
Institute of Fluid Mechanics, University of Erlangen–Nürnberg, Cauerstr. 4,

D–91058 Erlangen, Germany, breuer/bjaffrez@lstm.uni-erlangen.de
1 Introduction
Although LES is a highly promising simulation technique, it still suffers from
extremely large resources required for the resolution of the near–wall region,
especially for high–Re flows. That is the main motivation for setting up hy-
brid LES–RANS methods. Whereas RANS suits reasonably well to attached
boundary layers, requiring much less CPU–time and memory than LES, the
latter is recommended for complex large–scale flow phenomena, which RANS
often fails to predict correctly. Both characteristics are combined in hybrid
LES–RANS methods to obtain an optimal solution at low cost. Meanwhile a
variety of different hybrid concepts were proposed including DES.
In the present study a non–zonal approach based on two unique models is
preferred. The predefinition of RANS and LES regions is avoided and a grad-
ual transition between both methods takes place which weakens the problem
of setting up an appropriate coupling strategy. The new hybrid LES–RANS
approach relies on a one–equation model for the turbulent kinetic energy in
both modes. In addition to a linear eddy–viscosity model (LEVM), an explicit
algebraic Reynolds stress model (EARSM) is applied in order to account for
the Reynolds stress anisotropy. Both model variants have been tested on the
basis of the standard plane channel flow. Additionally, for the LEVM the flow
over a periodic arrangement of hills was considered.
2 Hybrid LES–RANS Technique

An important issue is the question of suitable models for such hybrid methods.
In the context of eddy–viscosity models for RANS, a two–equation model is
a natural choice since one transport equation is solved for the velocity scale
and one for the length scale. For LES zero– or one–equation models are more
obvious since the length scale is naturally given by the filter size ∆. These
facts are in contradiction to a unique model. However, if the near–wall region

488 Michael Breuer and Benoit Jaffrézic
is the main target for RANS, the equation for the length scale can be simpli-
fied to an analytical relation leading to a one–equation model for both zones.
Presently, the near–wall model by Rodi et al. [1] is applied in the viscosity–
affected near–wall layer and combined with a one–equation SGS model in the
outer LES region. The resulting unique model consists of a transport equation
for the modeled turbulent kinetic energy kmod in RANS mode and the sub-
grid scale turbulent kinetic energy ksgs in LES mode, respectively. For LES
the unknown dissipation correlation is modeled by ε = Cd k 3/2 /∆ leading to
νt = Cµ k 1/2 ∆ (Cµ = 0.05, Cd = 1.0).
As shown by Rodi et al. [1], for the near–wall region (RANS mode) the ap-
plication of the wall–normal velocity fluctuations (v 0 2 )1/2 as velocity scale is
better suited to characterize the turbulent motion than k. The length scale
lµ,v appearing in the eddy–viscosity relation νt = (v 02 )1/2 lµ,v now scales lin-
early with the wall–distance y and needs no damping function anymore, i.e.,
lµ,v = Cl,µ y. Indeed, the eddy–viscosity reduction in the vicinity of the wall is
an effect of the decreasing v 0 fluctuations rather than a viscous one and hence
naturally taken into account. The dissipation length lε,v used to define the
dissipation rate ε = (v 02 )1/2 k/lε,v also scales linearly near the wall. Only in
the immediate vicinity of the wall the distribution must be modified to yield
the correct behavior of ε ∼ y 0 as y goes to zero. In order to apply (v 0 2 )1/2 as
the velocity scale, an equation relating the wall–normal velocity fluctuations
to k = kmod was introduced [1] so that the transport equation does not have
to be modified. However, some adjustments on the model constants have been
done in order to fit more closely to high–Re flows [2].
In addition, an explicit algebraic Reynolds stress model (EARSM) de-
signed by Wallin and Johansson [3] for pure RANS applications has been
implemented in complement to the LEVM. This special EARSM was cho-
sen because of its near–wall treatment ensuring realizability of the individ-
ual stresses. The objective is to investigate whether a more advanced for-
mulation, which avoids the limitations of LEVM such as the poor predic-
tion of the normal Reynolds stresses, could be beneficial to the approach
suggested. The EARSM provides an algebraic relation for the Reynolds
stress tensor [3], which can be introduced in the momentum equation as
(ex) (ex)
u0i u0j = 23 k δij − 2 νt S ij + k aij . Here aij represents an extra anisotropy
tensor which is computed explicitly based on the normalized mean strain and
rotation tensors. Hence in this formulation the transport equation for k and
the relation for determining νt are still needed but the additional term takes
the anisotropy of the stresses appropriately into account. A further conse-
quence of applying an EARSM is on the production term in the transport
equation for k, which does not require any modeling anymore. Concerning
the interface between RANS and LES, an approach involving an interface
treatment without synthetic or reconstructed turbulence is preferred. Since
the criterion used by Rodi et al. [1] in their two–layer approach was found to
be unsuitable in the hybrid LES–RANS technique, it has been abandoned [2].
Unique Eddy–Viscosity Model for a Hybrid LES–RANS Method 489
1/2
As a remedy, the dynamic criterion y ∗ = kmod y/ν has been chosen to define
the LES–RANS interface [2]. y ∗ ≤ 60, which is also the validity restriction of
the relation given for (v 02 ) in the LEVM, means RANS mode, whereas y ∗ > 60
sets the simulation in LES mode. Alternatives are discussed in [2].
3 Results
Numerical Methodology and Test Cases
All simulations were performed with the LES code LESOCC, used for the solu-
tion of the filtered Navier–Stokes equations. This is a 3–D finite–volume solver
for arbitrary non–orthogonal and non–staggered (block–structured) grids [4].
The discretization is second–order accurate in space and time.
Two different test cases are considered. The first is the classical plane
channel flow at Reτ = 590 [6] performed on two grids. The dimensions of the
computational domain are: 2π (streamwise)× π (spanwise) × 2 (wall–normal
direction). The first grid A consists of 128 × 128 × 128 control volumes (CVs),
which represents a resolution of ∆x+ = O(30) (streamwise) and ∆z + = O(15)
(spanwise). The first grid point (half first cell) is located at y + = 0.68. The
second grid B contains 64×64×64 CVs leading to a resolution of ∆x+ = O(60)
and ∆z + = O(30). Here the first grid point is located at y + = 1.46.
A more challenging test case is the flow over periodic hills at Reb =
10, 595 [5]. The grid consists of 160×100×60 CVs in streamwise, wall–normal
(∆ycrest /h = 0.005, 1st CV height) and spanwise direction, respectively. This
grid has been originally designed for DES. The number of grid points in the
spanwise direction is moderate so that the RANS region becomes larger.
The results of the EARSM hybrid method are shown and compared with
the LEVM hybrid formulation which does not account for the anisotropy. Since
the approaches using the EARSM in both RANS and LES modes (E–RL) or
in RANS only (E–R) provide similar results, only the latter is presented in
Fig. 1. The performance of the LEVM is not studied here for the channel
flow but for the hill flow (c.f. [2]). The EARSM hybrid method is first studied
without and finally with the non–modeled production term in k–equation.
Channel Flow using EARSM and LEVM

Considering the plot of the mean streamwise velocity U + (y + ) (Fig. 1, left,
grid B), the influence of the EARSM is weak compared to LEVM. This is
especially true in the region close to the wall which is devoted to the RANS
mode and where the anisotropy is mainly located. For both grids the impact of
the models E–R and E–RL seems stronger in the LES region. Nevertheless, the
profiles of the non–scaled quantity U (y) do not show any real influence on the
LES region. Actually, the variations observed on U + (y + ) are obviously more
a consequence of the variations of the uτ values used for the normalization of
U + (uτ = 0.0539, 0.0541, 0.0525 for DNS, LEVM, and hybrid version with
EARSM in RANS mode only, i.e. E–R).
25 3
L 1.4 L
E–R E–R
20 2.5 1.2
E–Rnp E–Rnp
DNS DNS
RMS(u′ u′ )tot
RMS(v ′ v ′ )tot
2 1
15
+ 1.5 L 0.8
U
10 E–R 0.6
1 E–Rnp
DNS 0.4
5 0.5 0.2
0 0 0
1 10 100 0.1 1 10 100 0.1 1 10 100
y+ y+ y+
Fig. 1. Channel flow at Reτ = 590 on coarse grid B. Mean streamwise velocity
U + (left). Total normal Reynolds stresses (u0 u0 )tot (center) and (v 0 v 0 )tot (right). L:
LEVM hybrid method. E–R: hybrid method with EARSM in RANS mode.
Moreover, the unphysical step observed between y + ≈ 40 and 100 is larger

for the EARSM than for the LEVM. As expected, the consequences of the
EARSM implementation are more obvious on the normal Reynolds stresses.
(u0 u0 )tot is now well reproduced (Fig. 1, center) although a minor kink, which
is assumed to be related to the interface behavior, appears around y + ≈ 40.
This kink is less pronounced on grid A. In this region the normal stress is also
over–estimated but similarly to LEVM. Here the EARSM applied in both
RANS and LES regions provides slightly better results. Concerning (v 0 v 0 )tot
(Fig. 1, right), the curves obtained by the EARSM are not as good as for
(u0 u0 )tot . Nevertheless, the improvement over the LEVM hybrid variant is
significant. Again a kink, more pronounced than for (u0 u0 )tot appears. The
explanation is found in the distribution of the modeled contributions of the
stresses (not shown here). Whereas the transition between RANS and LES
looks smooth for (u0 u0 )mod , (v 0 v 0 )mod drops sharply at the interface location.
For the total Reynolds shear stress (u0 v 0 )tot it is found that each variant pre-
dicts this component similarly. The prediction of ktot is also in good agreement
in each case. The differences between the LEVM and the EARSM are minor,
although the last shows a trend to under–estimate ktot in comparison with
the former. kres is responsible for that.
With the implementation of the non–modeled production term for k (des-
ignation ‘np’ on the plots), a major difference is observed on U + (y + ) (Fig. 1,
left). As can be seen the profiles are shifted upwards leading to a good predic-
tion of the RANS near–wall region, whereas the LES logarithmic region shows
larger deviations to DNS. However, as mentioned above, the different profiles
of U (y) do not show strong deviations from each other. Thus the discrepancies
observed for U + (y + ) are still related to the values of uτ . Indeed, the value of
uτ decreases with the introduction of the non–modeled production term. The
reason of these variations between LEVM and the variants of the EARSM
has to be studied further. However, and conversely to the previous remark,
the unphysical step is weakened by this implementation. This fact is more
Unique Eddy–Viscosity Model for a Hybrid LES–RANS Method 491
3.5 3.5 3.5
Hybrid Amsch Hybrid Amsch
3 3 Ref. Solution 3 Ref. Solution
2.5 Hybrid Amsch 2.5 2.5
DES (Spalart)
2 Ref. Solution 2 2
y/h
y/h
y/h
1.5 1.5 1.5
1 1 1
0.5 0.5 0.5
0 0 0
-0.2 0 0.2 0.4 0.6 0.8 1 1.2 0.02 0.04 0.06 0.08 0.1 -0.04 -0.03 -0.02 -0.01 0
U/Ub ufuftot ufvftot
3 3
2 2
y
y
1 1
0 0
0 2 4 6 8 0 2 4 6 8
x x
Fig. 2. Periodic hill flow at Reb = 10, 595. Mean velocity U/Ub at x/h = 0.5, 2 and
6 (left), u0 u0 tot (middle) and u0 v 0 tot (right) at x/h = 2. Streamlines of the averaged
flow of the reference solution [5] (left) and the hybrid approach (right).
obvious when the value of uτ used for the normalization of each computation
is set to the DNS value (not shown here). The consequence is interesting and
this result could provide a good indication on the manner how to reduce this
unphysical step specific to the hybrid technique and also observed by other
groups. Regarding the prediction of the normal Reynolds stresses (Fig. 1, cen-
ter and right), the modeled part is reduced in comparison with the previous
hybrid EARSM (E–R). This reduction is compensated by an increase of the
resolved contribution in the RANS region for (u0 u0 ) but not for (v 0 v 0 ) and
(w0 w0 ), which leads to an under–estimation of these stresses. The Reynolds
shear stress estimation improves with the non–modeled production term. The
region close to the wall is now perfectly predicted and the magnitude of the
kink observed around the interface location is weakened. Furthermore, the
curves at this kink location are smoothened. kmod is naturally that quantity
which is mostly affected by the implementation of the production term. Here
the drop of kmod is significant. The peak value of kmod is almost halved for
E–Rnp compared to E–R. This behavior is not desired since the objective is
to obtain a mainly modeled contribution in RANS mode rather than an in-
creased resolved one. However, the compensation by kres is not large enough,
which provokes the present under–estimation of the total kinetic energy ktot .
Periodic Hill Flow using LEVM

The second, reasonably complex test case is the flow over periodic hills at
Reb = 10, 595. Additionally, a DES applying the Spalart–Allmaras (SA) model
and an LES using the Smagorinsky model have been performed. For the hy-
brid approach the interface position occurs at a distance of 6–12 RANS cells
from the wall (depending on the streamwise location) against 7–9 RANS for
DES. The separation region contains 12 RANS cells, whereas the hill crest is
covered by 6 cells.
The results (Fig. 2) of the hybrid model (LEVM) are plotted together with
the DES and the highly–resolved reference LES data [5]. The mean velocity
U/Ub is plotted at the positions x/h = 0.5, 2 and 6, which represent the
beginning and the center of the recirculation region and the flow after reat-
tachment, respectively. The mean velocity U/Ub is well reproduced by the
hybrid model at each position. However, concerning the wall shear stress dis-
tribution τw , the hybrid simulation shows at the hill crest larger deviations
(also noticed in RANS simulations) than the DES, which is in good agreement
with the reference LES. Regarding the recirculation region the LES applying
the Smagorinsky model predicts the best separation point (x/h = 0.201) with
respect to the reference LES (x/h = 0.19). Whereas the DES under–predicts
this location (x/h = 0.173) but stays in close agreement with the reference,
the hybrid simulation (x/h = 0.254) delays this point marginally. The hybrid
version (x/h = 4.751) and the DES (x/h = 5.197) show a slightly delayed
reattachment compared to the reference (x/h = 4.694). It is interesting that
the hybrid approach shows a better prediction of the reattachment location
than the DES and even the LES (x/h = 4.547). In comparison with LES and
DES, which show deviations of −3.1% and +10.7%, respectively, the hybrid
approach leads to a deviation of +1.2% only. Fig. 2 depicts plots of the aver-
aged streamlines for the hybrid technique and the reference LES. The center
of the recirculation zone is predicted similarly by the hybrid version and the
DES, but in comparison with the reference data, this location is slightly shifted
backward. However, both hybrid and DES still offer satisfactory results.
In general, the Reynolds stresses are well predicted by the hybrid approach
(Fig. 2). The profiles are overall recovered with respect to the reference LES.
Regarding the intensity, only slight under– or over–estimations are observed.
However, at the position x/h = 0.5 the hybrid technique shows an under–
estimation in the prediction of the peaks located at the vicinity of the lower
wall for v 0 v 0 tot and u0 v 0 tot . Other simulations, not mentioned here, show that
these under–estimations result from a lack of resolved scales in this region. In
conclusion, the hybrid simulation gives encouraging statistical results similar
(or in some cases better) than DES. Nevertheless, the hybrid method sug-
gested can be seen as superior in the prediction of the recirculation region.
In the next step of this study, the hill flow test case has to be considered in
combination with the EARSM formulation taking the anisotropy into account.
Acknowledgments: The project is supported by DFG (BR 1847/8–1).
References
1. Rodi, W., Mansour, N.N., Michelassi, V. (1993). J. Fluids Eng. 115: 196–205.
2. Breuer, M., Jaffrézic, B., Arora, K. (2005). EUROMECH Colloquium 469, LES
of Complex Flows, TU Dresden, Germany, Oct. 6–8, 2005.
3. Wallin, S., Johansson, A.V. (2000). J. Fluid Mech. 403: 89–132.
4. Breuer, M. (1998). Int. J. for Num. Methods in Fluids 28: 1281–1302.
5. Breuer, M. (2005). personal comm., http://www.hy.bv.tum.de/DFG-CNRS/
6. Moser, R.D., Kim, J., Mansour N.N. (1999). Phys. Fluids 11: 943–945.
A locally superconvergent scheme for the
simulation of turbulent flows in complex
geometries
M.V. Salvetti1 , S. Camarri1 , B. Koobus2 , A. Dervieux3

1
Dip. Ingegneria Aerospaziale, Università di Pisa (Italy)
mv.salvetti@ing.unipi.it, s.camarri@ing.unipi.it
2
Dep. Mathématiques, Université de Montpellier II (France)
koobus@math.univ-montp2.fr
3
INRIA Sophia Antipolis (France) Alain.Dervieux@sophia.inria.fr
Summary. A locally superconvergent scheme for the discretization of convective

fluxes in a mixed finite-element/finite-volume approach is proposed. This scheme
is second-order accurate on the whole unstructured mesh but may enjoy accuracy
up to fifth order in cartesian subregions. High-order accuracy is achieved through
a particular construction of finite-volume cells and through linear reconstruction of
the fluxes at the cell interfaces. The proposed method is applied to the simulation
of a laminar boundary layer and of the flow around a square cylinder.
1 Introduction
The present work is part of a research project aimed at developing and validat-
ing numerical codes for the simulation of complex turbulent flows of industrial
or environmental interest. In particular, we focus on such classes of flows for
which the most classically used turbulence approach in engineering applica-
tions, i.e. the Reynolds-averaged Navier-Stokes (RANS) approach, encounters
difficulties in giving accurate predictions. This is the case for flows character-
ized by massive separated regions or by significant unsteady phenomena, as,
for instance, in separated wakes behind bluff bodies. For such classes of flows
a more promising turbulence approach, which has received large attention in
the last decade, is large-eddy simulation (LES).
In our research project we mutually adapted an industrial CFD method-
ology and a LES approach [1] [2]. Our starting point was a numerical solver
for the Navier-Stokes equations in the case of compressible flows and perfect
Newtonian gases, based on a mixed finite-element/finite-volume scheme for-
mulated for unstructured grids made of tetrahedral elements. Finite elements
(P1 type) and finite volumes are used to treat the diffusive and convective

494 M.V. Salvetti, S. Camarri, B. Koobus, A. Dervieux
fluxes, respectively. This numerical technology has been extensively used for
RANS simulations of many different flows (see, for instance, [3]).
A significant part of our work concerned the adaption of the numerics
to the LES approach. Indeed, some numerical features which are positive in
a RANS simulation, such as, for instance, numerical viscosity, may become
dangerous in LES. In our opinion, the most critical point is indeed numerical
viscosity and we proposed [2] a MUSCL scheme in which stabilization is ob-
tained through numerical diffusion based on sixth-order derivatives, which has
a dissipative effect much more localized on high frequencies than the one of
stabilizations based on second-order derivatives. In this way we can reduce the
interaction between, from one hand, numerical dissipation, which damps in
priority the highest frequencies, in particular those for which the phase error
is too large and can produce oscillations, and from the other hand, the LES
closure model, which should reproduce the effects of unresolved frequencies on
the resolved ones. Moreover, a key coefficient (γ) permits to tune numerical
dissipation to the smallest amount required to stabilize the simulation.
In the present study, we push forward this concept and we investigate
the application of a locally superconvergent scheme, that is a scheme which is
second-order accurate on the whole unstructured mesh but enjoys accuracy up
to 5th order in regions where the grid is cartesian. This is achieved by applying
the linear reconstruction to flux functions rather than to flow variables.
2 Numerical method
The Navier-Stokes equations for compressible flows are discretized in space

using a mixed finite-volume/finite-element method applied to unstructured
tetrahedrizations. The adopted scheme is vertex centered and P1 Galerkin
finite elements are used to discretize the diffusive terms. A dual finite-volume
grid is obtained by building a cell Ci around each vertex i; two different ways
of constructing the finite-volume cells are considered.
In 2D, starting from an unstructured triangulation, the cells of the first
type (median cells, MC) are limited by part of the medians of the triangles
having the node i as a vertex; the extension to 3D is straightforward and is
described in [4].
The second type of cells (new generation cells, NGC) can be obtained in
3D as follows: to build the cell centered at node i, let us consider all the
neighboring nodes of i (j). For each element containing the nodes i and j,
the cell surface is given by the triangles connecting the middle of the edge
joining these two vertices, the surface center of the faces of the element having
this edge in common, and the volume center of the element. The surface
center of a given face is the center of its circumscribed circle, if the face
comprises only acute angles, otherwise it is the middle of its longest edge,
and the volume center of an element is the center of its circumscribed sphere
if the former is located inside the element, otherwise, it is the surface center
A locally superconvergent scheme for complex geometries 495
Z Z
Y Y X
6
5
5
4
4
3
Z
Z
3 2
2 1
6
1 6 0
4
6
0 4
Y
4
0 2 Y
2 2 X
2
X 4
0
6 0 0
(a) (b)
Fig. 1. New finite-volume cells in 3D: (a) division in tetrahedrons, (b) trace of NGC
on a tetrahedron resulting from the previous division.
(among those of the four tetrahedorn faces), which is closest to the center of
the circumscribed sphere. Although the NGC, as well as the MC, can be built
starting from a generic tetrahedrization, it is interesting to consider the case
of a Cartesian mesh, thus, made of rectangle parallelepipeds (thereafter called
bricks), which are cut in a particular way in tetrahedrons, following [5]. This
division splits each brick in six identical tetrahedra, each being the mirror
image of its neighbors (see Fig. 1(a)). Starting from such a tetrahedization,
the NGC cells are bricks, centered around the vertices of the mesh, as can be
seen in Fig. 1b, in which the trace of the division of an element into NGC is
shown.
The convective fluxes are then discretized in terms of fluxes through the
common boundaries shared by neighboring cells. The Roe scheme (with Turkel
preconditioning) represents the basic upwind component for the numerical
evaluation of the convective fluxes F:
F(Wi ) + F(Wj ) 1
ΦR (Wi , Wj , ) = · nij − γs P −1 |P R| (Wj − Wi )
2 2
in which ΦR (Wi , Wj ) is the numerical approximation of the flux between
the i-th and the j-th cells, Wi is the solution vector at node i, nij is the
integrated outward normal to the cell boundary and R(Wi , Wj , nij ) is the
Roe matrix. The matrix P (Wi , Wj ) is the Turkel-type preconditioning term,
introduced to avoid accuracy problems at low Mach numbers [7]. Finally, the
parameter γs , which multiplies the upwind part of the scheme, permits a
direct control of the numerical viscosity, leading to a full upwind scheme for
γs = 1 and to a centered scheme when γs = 0. As well known, the spatial
accuracy of this scheme is only first order. In our previous works, the MUSCL
linear reconstruction method (“Monotone Upwind Schemes for Conservation
Laws”, [8]) was employed to increase the order of accuracy of the Roe scheme.
The basic idea of this approach is to use in the Roe flux computation the
extrapolated values of the flow variables at the cell interface. In [2], a linear
extrapolation was proposed, based on a combination of different approximate
gradients of the variables, which leads to a numerical viscosity proportional
to sixth-order derivatives. The resulting scheme (V6) is second-order accurate

(for multi-dimensional non-linear problems on unstructured grids).
In the present work, the same kind of reconstruction is applied to the
fluxes, instead of to the flow variables. Thus, the following numerical flux is
obtained:
Fij + Fji 1
Φij = · nij − γs P −1 sign(P R)P (Fij − Fji )
2 2
where F̄ij and F̄ji are the extrapolated fluxes at the cell interface, computed
through the scheme proposed in [2]. The resulting numerical approximation
of the convective fluxes (V6NL) is 5th-order accurate for a conformal tetra-
hedrization obtained from a Cartesian grid as described previously. Indeed,
by interpolating the fluxes, the second-order limitation for MUSCL schemes
can be overcome.
Finally, the time marching algorithm is implicit, based on a backward
difference scheme and second-order accurate in time.
3 Applications and discussion

As a first validation, we wish to separate the effects of the numerical scheme
from those of the turbulence model; thus, we carried out the simulation of
a laminar boundary layer over a flat plate, which allows the behavior of the
new scheme to be appraised near a wall and in presence of stretched grids.
The inflow of the computational domain is located at a distance x = 0.1L
from the leading edge (L being the flat plate length) and the Blasius velocity
profile is imposed. The outflow is located at x = L, the lower boundary of
computational domain coincides with the flat plate, while the upper boundary
is at a distance of L. The Reynolds number, based on L, is equal to 104 . The
computational domain is discretized by always using the same number of nodes
in the streamwise, x, and in the normal, y, directions; the nodes are uniformely
spaced in x, while they are clustered near the wall following a hyperbolic
tangent law, allowing different stretching factors to be obtained. Each brick
of the Cartesian grid is then divided in conformal tetrahedrons, as described
in Sec. 2. Simulations have been carried out with different resolutions and
with different combinations of numerical ingredients, namely different finite-
volume cell types and different schemes for the convective-flux discretization
(V6 or V6NL). Fig. 2a shows the L2 norm of the error, computed with respect
to the exact Blasius solution at x = 0.82L, as a function of the resolution,
for two different grid stretching factors. In the case of a moderate aspect
ratio (AR ' 1.8), the gain in convergence offered by the NGC cells and the
V6NL scheme is noticeable, while for a higher aspect ratio (AR ' 33), at the
considered section, the differences are very small. This is probably due to the
fact that, with a high stretching factor, the discretization error of diffusive
terms, which are computed through P1 finite elements, overwhelms the one
A locally superconvergent scheme for complex geometries 497
45&26&78 34%15%67
!# 45698 34587
!"
45&2698 34%1587
456&78 345%67
:;,!,!<=,
2 ,3++-+
1#+2**,*
!%
#
!" !$
!"
:;,!,%%,
!$
!" # #
!" !"
&'()*+,-.,/-0*1 %&'()*+,-+.,/)0
Fig. 2. L2 norm of the error at (a) x = 0.82L; (b) x = 0.2L and AR ' 33. AR is
the maximum aspect ratio of the grid elements.
of convective terms. If the error at a section closer to the inflow is considered

for the same case, where the boundary layer is thinner, the most noticeable
improvement is given by the use of NGC (Fig. 2b). Note that at this latter
section, the results obtained with AR ' 1.8 are not reported, because they
are fairly inaccurate due to the inadequate number of nodes in the boundary
layer.
The second considered test-case is the flow around a square cylinder at
Re = 2.2 × 104 , which may be considered as a typical benchmark for large-
eddy simulation of bluff-body flows (see, for instance, [6]).
The computational domain and the boundary conditions are the same as
in [2]. Simulations have been carried out on two different grids. The first one
(GR1) is unstructured and made of approximately 105 nodes; it was already
used in the simulations described in [2]. The second (GR2) one is structured
and has about 4.1×105 nodes. The distance of the first node from the wall is
6×10−2 for GR1, while it is 6×10−3 for GR2; note also that the stretching of
the elements of GR2 is significant, the maximum aspect ratio being equal to
12, while the grid GR1 is rather homogeneous (see [2]). Large-eddy simula-
tions have been carried out on both grids with different combinations of the
previously described numerical ingredients; the main features of the simula-
tions are summarized in Tab. 1. In all cases the Smagorinsky model has been
used for the closure of the LES equations, the model constant being equal to
0.1 (see [2]).
The bulk flow parameters obtained in the different simulations are also re-
ported in Tab. 1, together with some reference experimental data. The overall
agreement with the experiments is significantly better when the structured
grid GR2 is used, especially for the fluctuations of the lift coefficient, due to a
better resolution of the shear layers detaching from the cylinder. The effect of
the finite-volume cell construction is also significant, NGC yielding a better
accuracy, while the difference between the two discretizations of the convec-
Table 1. Flow around a square cylinder. Main features of the simulations and
related values of the flow bulk parameters. C d is the mean drag coefficient,Cl0 the
r.m.s. of the lift coefficient and lr the length of the mean recirculation bubble.
Simulation Grid Cells Scheme γs C d Cl0 lr

S1 GR1 MC V6 0.05 1.89 0.84 1.41
S2 GR2 MC V6 0.2 2.43 1.36 1.1
S3 GR2 NGC V6 0.2 2.22 1.17 1.28
S4 GR1 MC V6NL 0.1 1.89 0.81 1.66
S5 GR1 NGC V6NL 0.1 1.82 0.65 1.84
S6 GR2 NGC V6NL 0.2 2.33 1.17 1.29
Exp. Lyn et al. (1995) – – – – 2.1 – 1.4
Exp. Bearman & Obasaju (1981) – – – – 2.28 1.2 1.385
tive fluxes are small. Note, however, that, as previously, this effect might be
masked by the influence of the discretization error of diffusive terms, and, in
this case, also by the subgrid scale model.
References
1. Camarri, S., Salvetti, M.V., Koobus, B. and Dervieux, A.: Large-eddy simulation
of a bluff-body flow on unstructured grids. Int. J. Num. Meth. Fluids, 40, 1431–
1460 (2002).
2. Camarri, S., Salvetti, M.V., Koobus, B. and Dervieux, A.: A low-diffusion
MUSCL scheme for LES on unstructured grids. Computers and Fluids, 33, 1101–
1129 (2004).
3. Koobus, B., Farhat, C. and Tran, H.: Computation of unsteady viscous flows
around moving bodies using the k-epsilon turbulence model on unstructured
dynamic grids. Comp. Meth. Appl. Mech., 190(11-12), 1441–1466 (2000).
4. Dervieux, A. and Désidéri, J.A.: Compressible flow solvers using unstructured
grids. INRIA Report, RR-1731 (1992).
5. Gourvitch, N., Rogé, G., Abalakin, I., Dervieux, A. and Kozubskaya, T.: A
tetrahedral-based superconvergent scheme for aeroacoustics. INRIA Report, RR-
5212 (2004).
6. Rodi, W., Ferziger, J.H., Breuer, M. and Pourquie, M.: Status of large eddy
simulation: results of a workshop. ASME J. Fluids Eng., 119, 248–262 (1997).
7. Turkel, E.: Review of preconditioning methods of fluid dynamics. Appl. Num.
Math., 12, 257-284 (1993).
8. van Leer, B.: Towards the ultimate conservative difference scheme, V. A second-
order sequence to Godunov’s method. J. Comp. Physics, 32, 101-136 (1979).
URANS Analysis of Flow-Induced Cavity
Resonances
M. Mesbah, W. Desmet, and M. Baelmans
K.U.Leuven, Department of Mechanical Engineering,

Celestijnenlaan 300 A, B-3001, Leuven, Belgium
Majid.Mesbah@mech.kuleuven.be
Summary. In this paper, flow oscillations in a two-dimensional rectangular cav-

ity are being studied. The unsteady compressible RANS equations are investigated
both in combination with a two-equations k − ω model and with the Reynolds Stress
model (RSM) to include the turbulence effects. The simulations are performed us-
ing Fluent 6.2 for low Reynolds flow with a range of moderate Mach numbers. Flow
features and frequency analysis are presented and discussed. The obtained results
show good agreement with Rossiter’s predictions based on experiments. Further-
more, a comparison between results obtained from URANS simulations employing
different turbulent models with the results obtained from DNS and LES simulations
are presented.
1 Introduction
The physics of cavity flow oscillation has been the subject of research for
many decades. Rossiter was the first researcher who described an acoustic
feedback mechanism [3]. The flow forms a shear layer which amplifies flow
disturbances and the subsequent scattering of these disturbances into acoustic
waves at the downstream corner. These acoustic waves propagate upstream
and excite further disturbances in the shear layer, creating a feedback loop,
which often leads to self-sustained oscillations at discrete resonant frequencies.
He also introduced a semi-empirical formula for the dimensionless frequencies
(Strouhal number, Str) of this periodic phenomenon:
fL n−γ
Str = = , n = 1, 2, ... (1)
U∞ M + 1/k
with f the frequency, L the length of the cavity, U∞ the free-stream velocity,
n the mode number, M the Mach number, and constants k and γ are 0.57
and 0.25, respectively.
Numerical and experimental studies reveal two distinct types of flow os-
cillations inside the cavity: a shear-layer and a wake mode. Shear-layer mode

500 M. Mesbah, W. Desmet, and M. Baelmans
is characterized by the acoustic feedback process described above. The wake

mode oscillation was first observed experimentally in an axisymmetric cav-
ity with incompressible flow (water) and very low speed (0.23 m/s) [4]. In
this mode, the flow is characterized by a large scale vortex shedding from the
cavity leading edge.
Several CFD approaches such as DNS [2], LES [1], RANS/DNS [8] and
RANS/LES [7] have already been used to predict this flow induced noise. It
should be noticed that, although the wake mode has been observed only in
incompressible and very low speed flows, these approaches visualized the wake
mode for two-dimensional cavities even for a compressible flow at a moderate
Mach number.
However, the applications of these approaches to complex geometries is
limited and the computational cost for high Reynolds flow is still infeasible.
The unsteady RANS turbulence models have also been used by a number
of researchers [6, 5] but these approaches tend to predict only low-frequency
discrete acoustic tones while the noise spectrum of cavity noise contains both
broadband and tonal components. On the other hand, some low cost engineer-
ing methods such as Stochastic Noise Generation and Radiation (SNGR) has
been introduced in the recent years which can capture the broadband noise.
The motivation for this research is to investigate a hybrid approach based on
URANS and SNGR methods to capture both broadband and tonal noise in
cavity flows.
In the current research, oscillations in the flow passing a two-dimensional
rectangular cavity are studied using URANS simulations. Subsequently, the
k−ω model and the Reynolds Stress model (RSM) are used to include the tur-
bulence effect. The ability of those turbulence models to predict the frequency
of oscillations as well as flow features are investigated.
The paper is organized as follows. First, the computational setup is il-
lustrated. Next, the flow features obtained from URANS simulation are dis-
cussed. Furthermore, frequencies of flow oscillation are analyzed and compared
with results obtained by other approaches. Conclusions are presented in the
last section.
2 Computational setup
Simulations are performed on a two-dimensional rectangular cavity using the

Fluent 6.2 solver.1 Figure 1 shows a schematic diagram of the computational
domain. The cavity has the length to depth ratio (L/D) of 4 and simulations
are performed for Mach numbers ranging from 0.3 to 0.7. The Reynolds num-
ber based on cavity depth (ReD ) is 1500 in all cases. The setup is similar
to the numerical setup presented in references [1, 2], where LES and DNS
simulations are performed respectively. The computations are initiated from
1
Fluent6.2 , Fluent Inc, 2005
URANS Analysis of Flow-Induced Cavity Resonances 501
a laminar flat-plate Blasius boundary layer along the wall and spanning the
cavity. The ratio L/θ0 is constant and equals 102. The parameter θ0 corre-
sponds to the momentum thickness of the boundary layer at the upstream
edge for the initial condition. Furthermore, the URANS simulations are per-
formed using the simplistic Cartesian multi-block grid. The mesh consists of
two blocks and the grid is stretched around the edges. The cavity block con-
sists of 101 × 51 mesh points and the block above the cavity has 301 × 131
mesh points.
Fig. 1. Sketch of the computational domain
3 Flow results
In order to study the physical structure of flow inside the cavity, an average
over 10 sampled data points in one period of oscillation is performed. Figures 2
(a) and (b) illustrate the time-averaged streamlines of the cavity with M=0.7
using κ-ω and RSM simulations, respectively. The flow pattern displayed in
Figure 2 (a) shows the presence of two vortices in the cavity, though one is
predominant. The streamlines above the cavity are nearly horizontal along
the mouth of the cavity. The presented flow structure for shear-layer mode is
similar to the observations which have been reported by other investigators
[7, 1, 2]. The time-averaged streamlines obtained from RSM simulations are
illustrated in Figure 2 (b) where only one large vortex in the cavity can be
seen. The mean flow pattern is similar to the case when flow converges to
the steady state due to the high dissipation [5]. It can be concluded from
comparison between Figures 2 (a) and (b) that RSM is more dissipative than
κ-ω model. Therefore, it is expected that the fluctuations captured in shear
layer by RSM have smaller amplitudes than when the κ-ω model is used.
4 Frequency analysis
The time signals of primitive variables for each simulation are recorded at
the observation point which is located at the mouth of cavity at x = 0.78L.
(a)
(b)
Fig. 2. Mean streamlines of cavity flow with L/D = 4 and M=0.7. (a) κ-ω simula-
tion, (b) RSM simulation,
Fourier analysis of the recorded signals provides the frequency spectrum where
several tonal frequencies and their harmonics become evident.
Figures 3 (a) and (b) show the spectrum of the recorded pressure for M=0.5
using RSM and k − ω models respectively. Oscillation modes corresponding
to the first and second Rossiter modes can be observed. These figures reveal
that the simulation using k − ω model is able to capture the first and second
Rossiter’s modes with the similar amplitudes, while RSM simulation captures
only the second mode.
170 170
Mode 1 Mode 2
160 160
Mode 2
150 150 Higher harmony
140
SPL (dB)
140
SPL (dB)
130 130
120 120
110 110
100 100
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Str Str
(a) (b)
Fig. 3. Power spectra of pressure at the point located on top of the cavity for
M=0.5. (a)κ-ω model (b) RSM model
The frequencies of the two most dominant peaks in the spectra for all
simulations, as well as predictions from Rossiter’s model [3], LES [1] and DNS
[2] are presented in Figure 4. The obtained results show a good agreement with
Rossiter’s predictions which are based on experiments.
URANS Analysis of Flow-Induced Cavity Resonances 503
1 First Rossiter mode

Second Rossiter mode
K−w simulations
0.9 RSM simulations
LES results
0.8 DNS resutls
0.7
Str= f.L/U
0.6
0.5
0.4
0.3
0.2
0.1
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Mach
Fig. 4. Strouhal numbers for peaks in spectra compared to Rossiter’s model
An overview of the oscillation modes that are obtained using different ap-
proaches is presented in Table 1. DNS simulations show the wake mode for
M ≥ 0.4. In this oscillatory mode, the frequency of oscillation is not changed
by increase of Mach number. LES simulations illustrate almost the same re-
sults but at the M = 0.4 both shear-layer and wake modes are observed. In
simulations using URANS, the RSM model does not reveal any oscillation
for M < 0.5. For the simulations with M ≥ 0.5 only shear-layer mode is ob-
served. The k − ω model provides an oscillatory flow field for M ≥ 0.4 and like
the RSM model the flow oscillates in the shear-layer mode but with higher
amplitude.
Table 1. Overview of flow oscillation modes using different approaches.
DNS LES RSM k−ω

M=0.3 mixed mixed non-oscillatory non-oscillatory
M=0.4 wake mixed non-oscillatory shear-layer
M=0.5 wake wake shear-layer shear-layer
M=0.6 wake wake shear-layer shear-layer
M=0.7 wake not available shear-layer shear-layer
5 Conclusions and future work

In the present studies, the unsteady Reynolds-averaged Navier-Stokes equa-
tions (URANS) was applied to simulate the flow passing over a two-dimensional
cavity. The k − ω model and the Reynolds Stress model (RSM) are imple-
mented to include the turbulence effect. The cavity has the length to depth
ratio of 4 and simulations are performed for flow with constant Reynolds
number of 1500 and Mach numbers ranging from 0.3 to 0.7.
The results reveal that both turbulent models can capture the oscillations
in the shear layer and the Strouhal numbers of the predicted dominant oscil-
lation mode compare well with Rossiter’s formula. Unlike the LES and DNS
results, only shear-layer mode was observed and no wake mode was visualized.
Furthermore, by comparing the ability of RSM and k − ω models in capturing
the flow oscillations, it can be concluded that RSM is more dissipative than
k − ω model.
Since the URANS approaches are naturally dissipative, the simulations
should be performed for the flow with high Reynolds numbers and compared
with the experimental results. This is the future work for this research.
References
1. Rubio, G., De Roeck, W., Baelmans, M., Desmet,W.: Numerical identification of
flow-induced oscillation in Rectangular Cavities using Large Eddy Simulation.
Accepted for publication in the International Journal for Numerical Methods in
Fluids (2006)
2. Rowley, C. W.: Modeling simulation and control of cavity flow oscillations., PhD
Thesis, California Institute of Technology (2002)
3. Rossiter, J. E.: Wind tunnel experiments on the flow over rectangular cavities at
subsonice and transonic speeds. Royal Aircraft Establishment, Technical Report
64037 (1964).
4. Gharib, M., Roshko, A.: The effect of flow oscillation of on cavity drag. JFM
(1987)
5. Sinha, N., Dash, S. M., Chidambaram, N., Findlay, D.: A perspective on the
simulation of cavity aeroacoustics. AIAA paper, 98-0286 (1998)
6. Henderson, J., Badcock, K., Richards, B. E.: Subsonic and transonic transitional
cavity flows. AIAA paper, 2000-1966 (2000)
7. Shieh, C. M., Morris, P.: Parallel computational aeroacoustic simulation of tur-
bulent subsonic cavity flow. AIAA paper, 2000-1914 (2000)
8. Cain, A. B., Rubio, A. D., Bortz, D. M., Banks, H. T., Smith, R.C.: Optimizing
control of open bay acoustics. AIAA paper, 2000-1928 (2000)
Revisiting URANS computations of the flow
behind a backward-facing step using
second moment closures
A. Fadai-Ghotbi, R. Manceau, and J. Borée
Laboratoire d’Études Aérodynamiques, UMR 6609 CNRS/Université de

Poitiers/ENSMA, SP2MI - Téléport 2, Bd Marie et Pierre Curie, BP 30179, 86962
Futuroscope Chasseneuil cedex, France. atabak.fadai@lea.univ-poitiers.fr
Summary. Results of Unsteady Reynolds-Averaged Navier-Stokes (URANS) com-

putations of the backstep flow at Re = 37500 are presented. The standard k − ε and
LRR models using wall-functions are compared to the Elliptic Blending Reynolds-
stress Model (EB-RSM), a recent second-moment turbulence closure which repro-
duces the non-viscous wall-blocking effect. The skin friction and reattachment length
are significantly improved by the EB-RSM model. Consistent with previous studies,
the vortex shedding can be captured, with the correct Strouhal number (St ' 0.20)
and convection velocity (0.5U0 ). However, the energy of the unsteady, resolved part
is less than 5% of the total energy and is strongly dependent on the mesh. Using
a refined mesh, surprisingly, a steady solution is obtained. It is shown that this
behaviour can be linked to the numerical errors at the step corner, which act as
perturbations that can be sufficient to excite the natural mode of the shear layer,
when the local Peclet number, comparing convection and diffusion, is high enough.
This results suggest that URANS is not appropriate to quantitatively predict the
amplitude of the large-scale structures developing in a shear-layer.
1 Introduction
Recently, a considerable interest has emerged in relatively low-cost (compared

to LES) time-dependent computations of complex flows, in different fields of
industrial activities, e.g., for thermal fatigue studies, fluid/structure interac-
tion, noise prediction, etc. In particular, the URANS methodology has become
quite popular, because of its successes in predicting the most energetic modes
and their Strouhal number, in particular for vortex shedding behind bluff
bodies (e.g., [9], [2], [10]). Although fundamental questions remain about the
general definition of the URANS filter, the methodology is straightforward: it
simply consists in solving standard RANS-model equations in a time-accurate
mode. However, the possibility of using standard RANS models in an un-
steady calculation is questionable: for instance, Carpy and Manceau [3] have
shown that linear eddy-viscosity models give a completely wrong dynamics in

506 A. Fadai-Ghotbi, R. Manceau, and J. Borée
&#!$.
,%
'()*+ /0+)*+
&#!$1
"#!$!- ! & &#!$%
"
"#!$% "#!$-%
Fig. 1. Geometry and computational domain.
a pulsed jet, due to the lack of representation of the stress–strain misalign-

ment. On the contrary, second-moment closures (SMC) are able to reproduce
the global dynamics of this flow without any case-specific modification.
In the case of separated shear layers, it was shown by Lasher & Taulbee
[12], in a backstep flow, that SMCs are able to reproduce the vortex shed-
ding, as soon as sufficiently accurate numerical methods are used. Despite
its geometrical simplicity, the backstep flow is a particularly challenging test
case for unsteady calculations, since it conjugates several fundamental mecha-
nisms: vortex shedding, interaction between the vortices and the wall, flapping
of the recirculation region at a low frequency. The Elliptic Blending Reynolds
stress model, derived in order to reproduce the nonlocal blocking effect of the
wall, is expected to improve the representation of the physics of this case,
compared to eddy-viscosity models and other second moment closures. The
aim of the present work is to evaluate the performance of this model, not only
as concerns the mean quantities, but also the quantitative prediction of the
characteristic frequencies and the amplitude of the large-scale structures.
2 Flow configuration and numerical methods
The 2D backstep flow configuration is presented in Fig. 1. The Reynolds

number is Re = 37500, based on the inlet centre velocity U0 and the step
height h, and the inlet momentum thickness Reynolds number is 5000. The
experiment [4, 5] underlines the existence of two dominant frequencies. A
high frequency is associated with the vortex shedding due to the Kelvin-
Helmholtz instability, with a Strouhal number of 0.20, based on the local
vorticity thickness and shear velocity (U0 /2) and a low frequency (St = 0.06),
characteristic of a flapping motion.
In the present work, the standard k − ε and LRR [13] models, using wall
functions, are compared with EB-RSM [14]. The latter model is based on the
elliptic relaxation theory [6], which enables the reproduction of the crucial,
nonlocal blocking effect of solid boundaries on turbulence.
Computations were performed with Code Saturne, a parallel, finite vol-
ume solver on unstructured grids, developed at EDF [1]. Space discretization
is based on a collocation of all the variables at the centre of gravity of the
cells. Velocity/pressure coupling is ensured by the SIMPLEC algorithm [8],
with a Rhie & Chow interpolation in the pressure-correction step. The Poisson
equation is solved with a conjugate gradient method. Time advancement is
Revisiting URANS computations of the backstep flow using SMCs 507
Mesh 1 2 3 4 5 6 Mesh 7 8 9 10 11 12
fr 1 1.5 2 2.5 3 4 fr 1 1.2 1.4 1.6 1.8 2
Ncell 1396 3396 5970 9842 14234 25154 Ncell 3954 5936 7920 10816 13838 17256
y1+ 150 100 75 60 50 37 y1+ 3 2.5 2.1 1.9 1.6 1.5
Table 1. Mesh characteristics. Left: high-Reynolds models (k − ε and LRR). Right:
low-Reynolds model (EB-RSM).
based on an implicit Euler scheme, and time-step convergence has been care-
fully checked. Space derivatives are approximated by a second-order central-
difference scheme (CDS) and first-order upwind-difference scheme (UDS) for
comparison. Six other meshes, numbered from 1 to 6, are used for the high-
Reynolds number models (k − ε and LRR), mesh 1 being the reference mesh
(see Tab. 1, left). Six meshes, numbered from 7 to 12, are also used for the
low-Reynolds number model (EB-RSM), mesh 7 being the reference mesh (see
Tab. 1, right). fr is the factor of refinement in each direction compared to the
reference mesh. Ncell is the number of cells and y1+ is the distance of the first
point to the wall in wall units. Boundary layer profiles are extracted from
separate calculations and applied at the inlet of the domain in order to match
the boundary layer thickness 4h upstream of the step corner observed in the
experiments.
3 Results
The Elliptic Blending Model significantly improves the skin friction coefficient
Cf (Fig. 2) and the reattachment length lr (Fig. 3) compared to the standard
k − ε and LRR models with wall functions. In Fig. 2, the friction coefficients
obtained with the EB-RSM model with the different meshes are superimposed,
showing that the results are not sensitive to a reasonable coarsening of the
near-wall mesh, which is very interesting for industrial applications. The EB-
RSM mean velocity profiles (Fig. 2) show that the boundary layer downstream
of reattachment recovers more slowly than the experimental data. This feature
is common to most of the turbulence models, SMCs as well as eddy-viscosity
models. All the computations with the k − ε model give steady solutions,
Fig. 2. Left: mean skin friction coefficient. Right: mean velocity profile. Symbols:
experiment [4]; lines: EB-RSM calculation.
Experiment k − ε LRR EB-RSM

lr /h 6.26 5.14 4.72 6.51
Fig. 3. Top left: mean reattachment length. Bottom left: isocontours of vertical re-
solved velocity with mesh 10 and EB-RSM model (dashed lines are negative values).
Right: maximum value of the ratio resolved energy/total energy as a function of the
streamwise location (EB-RSM calculation).
whatever the mesh or the space discretization used. On the contrary, the
behaviour of the SMCs is strongly dependent on numerics: using a UDS,
whatever the meshes and the model, the solution is steady; on the contrary,
using a CDS and meshes similar to those used by Lasher & Taulbee [12], the
solution is unsteady, exhibiting a single frequency. This behaviour shows the
significant influence of numerical diffusion in unsteady flows, consistent with
the conclusions drawn by Lasher & Taulbee. The correct order of magnitude
(St = 0.2) of the Strouhal number is found with both SMCs: St = 0.21 (EB-
RSM) and St = 0.16 (LRR). This frequency is associated with the vortex
shedding which can be visualized in Fig. 3. These large-scale structures are
convected downstream at 0.5U0 as reported in the experiment. The capture of
these structures by the recirculation zone, observed in the experiments, is not
reproduced: consequently, the flapping motion, which is linked to the capture
and release of the shed vortices [11] is missed.
Moreover, the resolved, large-scale structures contain a very limited part of
the total fluctuating energy: indeed, the ratio kres /ktot is less than 5%, where
kres is the kinetic energy of the resolved structures and ktot = kres + kmod is
the sum of the resolved and the modelled kinetic energies.
One could expect that refining the mesh would intensify the resolved en-
ergy, by reducing the numerical errors. On the contrary, what is obtained is a
drastic reduction of the resolved energy and eventually a steady solution for
the finest mesh (mesh 12), as seen in Fig. 3. This figure shows the stream-
wise evolution of M (x) for each mesh used in the EB-RSM calculation, where
M (x) = M axy kres (x, y)/ktot (x, y) . (The results with the LRR model, not
shown here, are qualitatively similar).
The origin of this behaviour can be traced to the excitation of the most
unstable mode of the shear layer by numerical errors at the step corner.
Indeed, it is well known, at least for a 1D convection/diffusion equation,
that central differencing can generate numerical oscillations if the Peclet
number exceeds 2 [8]. The Peclet number compares the relative magni-
tudes of convection and diffusion: in the boundary layer just upstream the
step corner, the convection and diffusion can be considered as essentially
Revisiting URANS computations of the backstep flow using SMCs 509
Fig. 4. Left: streamwise evolution of the local Peclet number at y + = 15. Right:
streamwise evolution of ∂x U at y + = 15.
in the streamwise and wall-normal directions, respectively, and can be lo-

cally approximated by U 2 /∆x and νtot U/∆y 2 , respectively. Thus, the local
Peclet number is P e = U ∆y 2 /(νtot ∆x), where νtot accounts for the total
(molecular + turbulent) viscosity.
Fig. 4 shows the streamwise evolution of the local Peclet number at
y + = 15, which is the location of the streamline passing through the in-
flexion point of the shear layer just after the step, where the velocity profile
is the most sensitive to perturbations, according to the linear stability theory.
It can be seen, comparing Fig. 3 (right) and Fig. 4 (left), that there is at
least a qualitative link between the orders of magnitude of the local Peclet
number and the energy contained in the resolved eddies. This comparison
suggests that, for coarse meshes, the numerical oscillations due to the high
local Peclet number in a region of high gradients (very close to the step, say
−0.1 < x/h < 0) act as perturbations that are sufficient to excite the natural
mode of the shear layer. Fig. 4 shows the evolution of the streamwise deriva-
tive of the long-time averaged velocity ∂x U , at y + = 15. It is seen that the
amplitude of the numerical oscillations decreases when the mesh is refined.
The Strouhal number of the large-scale structures does not depend on the
mesh, since it corresponds to the most amplified mode of the shear layer,
as predicted by the linear stability theory. However, the amplitude of these
structures is strongly dependent on the amplitude of the perturbations, which
is driven by the numerical errors. Obtaining an unsteady solution is therefore
a numerical artifact, which makes the use of URANS uncertain in such flows,
in particular as concerns the prediction of the amplitude of the large-scale
oscillations, which are of major interest for industrial applications.
Conclusion
The Elliptic Blending Model, a recent near-wall second-moment closure, has
been tested on the backstep flow at Re = 37500. The skin friction and the
reattachment length are highly improved compared to the standard k − ε
or LRR models with wall-functions, and are not sensitive to a reasonable
coarsening of the near-wall mesh. The EB-RSM is much more robust than
previous models based on Elliptic Relaxation method ([7],[15]) and can be
easily implemented in RANS codes using second-moment closures. It has been

shown that obtaining an unsteady solution, whatever the model used (high
or low-Reynolds model), is due to the amplification of numerical errors in a
region very close to the step, where the velocity gradients and local Peclet
number are high. These numerical errors act as perturbations that can be
sufficient to excite the natural mode of the shear layer. As shown by previous
studies, URANS is able to reproduce the correct Strouhal number of the
vortex shedding, since it is imposed by the mean velocity profile, according to
the linear stability theory. However, the amplitude of the vortices is strongly
dependent on numerical errors, and therefore, its prediction with the URANS
methodology does not seem to be reliable: this result suggests that future
work should be devoted to the development of hybrid RANS/LES methods.
References
1. F. Archambeau, N. Méchitoua, and M. Sakiz. Code Saturne: A finite volume
code for the computation of turbulent incompressible flows - Industrial applica-
tions. Int. J. on Finite Volume, Electronical edition: http://averoes.math.univ-
paris13.fr/html, ISSN 1634(0655), 2004.
2. G. Bosch and W. Rodi. Simulation of vortex shedding past a square cylinder
with different turbulence models. Int. J. Num. Meth. Fluids, 28:601–616, 1998.
3. S. Carpy and R. Manceau. Turbulence modelling of statistically periodic flows
: the case of the synthetic jet. In Proc. 6th Int. Symp. Engng. Turb. Modelling
and Measurements, Sardinia, Italy, 2005.
4. D. Driver and H.L. Seegmiller. Features of a reattaching turbulent shear layer
in divergent channel flow. AIAA J., 23:163–171, 1985.
5. D.M. Driver, H.L. Seegmiller, and J.G. Marvin. Time dependent behavior of a
reattaching shear layer. AIAA J., 25:914–919, 1987.
6. P. A. Durbin. Near-wall turbulence closure modeling without “damping func-
tions”. Theoret. Comput. Fluid Dynamics, 3:1–13, 1991.
7. P. A. Durbin. A Reynolds stress model for near-wall turbulence. J. Fluid Mech.,
249:465–498, 1993.
8. J. H. Ferziger and M. Perić. Comp. meth. for fluid dynamics. Springer, 1996.
9. G. Iaccarino, A. Ooi, P. Durbin, and M. Behnia. Reynolds averaged simulation
of unsteady separated flow. Int. J. Heat & Fl. Flow, 24:147–156, 2003.
10. S. Johansson, L. Davidson, and E. Olsson. Numerical simulation of vortex shed-
ding past triangular cylinders at high Reynolds number using a k−ε turbulence
model. Int. J. Num. Meth. Fluids, 16(10):859–878, 1993.
11. M. Kiya and K. Sasaki. Structure of a turbulent seperation bubble. J. Fluid
Mech., 137:83–113, 1983.
12. W.C. Lasher and D.B. Taulbee. On the computation of turbulent backstep
flow. Int. J. Heat and Fluid Flow, 13:30–40, 1992.
13. B. E. Launder, G. J. Reece, and W. Rodi. Progress in the development of a
Reynolds-stress turbulence closure. J. Fluid Mech., 68(3):537–566, 1975.
14. R. Manceau. An improved version of the Elliptic Blending Model. Application
to non-rotating and rotating channel flows. In In Proc. 4th Int. Symp. Turb.
Shear Flow Phenomena, Williamsburg, VA, USA, 2005.
15. R. Manceau and K. Hanjalić. Elliptic Blending Model: a new near-wall
Reynolds-stress turbulence closure. Phys. Fluids, 14(2):744–754, 2002.
Contribution of Turbulence Equation Terms to
the Shear Stress Balance
Dragan Kovzulović and Thomas Röber
DLR – Institute of Propulsion Technology, Linder Höhe 1, D-51147 Cologne,

Germany, dragan.kozulovic@dlr.de, thomas.roeber@dlr.de
Abstract — A modification to Menter’s SST turbulence model is proposed,

which sets up the blending functions based on local flow conditions. Specif-
ically the ratio of k-equation terms and a criterion for curvature influence
are introduced, leading to improved balancing of shear stress in complex 3D
boundary layers and simultaneously retaining the benefits of the SST model
in 2D flows. As a positive side effect, the new approach does not require the
computation of the wall distance. Besides generic test cases, an application to
a single stage transonic compressor is presented.
1 Introduction
From the modelling point of view there are mainly two benefits from Menter’s
Shear Stress Transport (SST) turbulence model [1] as compared to other two-
equation models: blending from ω- to ε-equation while moving away from the
wall and the application of the Bradshaw assumption in boundary layers. The
former reduces the dependency on the prescribed free stream turbulence val-
ues while the latter prevents the overprediction of eddy viscosity in adverse
pressure gradient flows by assuming that the shear stress does not increase
more rapidly than the turbulence kinetic energy. On the other hand, both fea-
tures strongly rely on the wall distance for determining their exertion regions.
This can pose a problem in complex 3D boundary layers of large elongation.
For example, the above mentioned regions only partly cover the turbomachin-
ery side wall boundary layers, as will be shown in section 3. Hence a criterion
is proposed which locally determines the blending functions by balancing the
terms of the k-equation. This is considered to be a more physical criterion as
compared to the wall distance, since it represents a charcteristic condition of
the wall bounded layers. Consequently, an improved balancing of shear stress
can be expected. Below the new approach is denoted as the BSS model, which
is an abbreviation for Balance of Shear Stress.

512 Dragan Kovzulović and Thomas Röber
2 Modelling Approach
The equations for turbulent kinetic energy k and dissipation rate ω of the BSS
model are formally the same as the SST model by Menter [1]:

∂ (%k) ∂ (%Ui k) ∗ ∂ ∂k
+ = P̃k − β %kω + (µ + σk µT )
| ∂t
{z } | ∂x {zi }
| {z } ∂xi ∂xi (1)
|{z}
| {z }
klocal kconv kprod kdest kdif f

∂ (%ω) ∂ (%Ui ω) 2 ∂ ∂ω
+ = 2α%Sij Sij − β%ω + (µ + σω µT ) +
∂t ∂xi ∂xi ∂xi (2)
1 ∂k ∂ω
2 (1 − F1 ) %σw2
ω ∂xi ∂xi
To avoid anomalous turbulence production in stagnation regions, the produc-
tion term of k-equation is limited:

∂Ui ∂Ui ∂Uj
P̃k = min µT + , 10β ∗ %kω (3)
∂xj ∂xj ∂xi
Furthermore, the eddy viscosity µT , containing the Bradshaw assumption, is
defined as follows:
a1 %k
µT = p with: a1 = 0.31 (4)
max a1 ω, F2 2Sij Sij
The constants are blended from the corresponding constants from the k-ε and
k-ω model, e.g. α = α1 F1 +α2 (1−F1 ). Following constants have been applied:
α1 = 5/9, β1 = 3/40, σk1 = 0.85, σω1 = 0.5, α2 = 0.44, β2 = 0.0828, σk2 = 1,
σω2 = 0.856 and β ∗ = 0.09.
Instead of using the wall distance as proposed by Menter [1] the blending
functions F1 and F2 in Eqs. (2) and (4) are determined by weighting the terms
of the k–equation (1). The following instructions are applied subsequently.
First of all, F1 is determined:

QS1 − 0.5 |kprod |
F1 = with: QS1 = max min , 4.0 , 0.5 (5)
3.5 |kconv | + |kdif f |
This approach is based on the idea that the source terms dominate near the
wall and transport terms prevail far away from wall.
Secondly, F2 is the maximum value of F1 and a function which depends
on the ratio of turbulence production to destruction:

−0.9 QS2 −0.2
F2 = max F1 , QP0.2 · 0.3
h i
|k |
with: QP = max min |kprod dest | , 1.1 , 0.9 (6)
h i
|k |
and: QS2 = max min |kconv prod |+|kdif f | , 0.5 , 0.2
QP typically exceeds 1 in the wake region of the boundary layer, and assures
F2 = 1. This is usually the region where the Bradshaw assumption is effective
and prevents the overproduction of eddy viscosity. Furthermore, the term
Contribution of Turbulence Equation Terms to the Shear Stress Balance 513
involving QS2 prevents high values of F2 in free stream regions with QP 1

(e.g. shock).
The third step accounts for the specific conditions in the viscous sublayer,
where the ω-gradient is very large and the product of k- and ω-gradient is
negative:
r
∂k ∂ω ∂ω ∂ω 106
F1 = F2 = 1 for simultaneously: < 0 and: > (7)
∂xi ∂xi ∂xi ∂xi ms
Finally, in regions of high streamline curvature considerable amplification
of turbulence can occur, hence the Bradshaw limitation of turbulence pro-
duction is not necessarily valid and is neutralized by the following equation:
√ p
10(r∗ − 1)κ Ui Ui

∗ 2Sij Sij
F2 = F2 · 1 − max ,0 for: r = p > 1(8)
βω ω 2Ωij Ωij
According to Kovzulović and Röber [3] the regions of curvature-induced turbu-
lence amplification can be identified by r∗ > 1, using the following definitions
of strain Sij and vorticity Ωij :

1 ∂Ui ∂Uj 1 ∂Ui ∂Uj
Sij = + and: Ωij = − (9)
2 ∂xj ∂xi 2 ∂xj ∂xi
Furthermore, the streamline curvature κ is determined by deriving the tangent
unit vector T in flow direction s:
s 2 2 2
dT ∂Tx ∂Ty ∂Tz
κ= = + + (10)
ds ∂s ∂s ∂s
Following transformation yields dT/ds:
 ∂T ∂T ∂T   
x x x
∂x ∂y ∂z Tx
dT  ∂T y ∂Ty ∂Ty  
U
=  ∂x ∂y ∂z  Ty  with: T= (11)
ds ∂Tz ∂Tz ∂Tz |U|
Tz
∂x ∂y ∂z
The novel approach shows negligible additional computing time as com-

pared to the standard k–ω model by Wilcox [2] or the SST model by
Menter [1], since the terms of k-equation are already known. The compu-
tational accuracy of the k-equation terms depends on the applied numeri-
cal schemes and can negatively affect the convergence behaviour of the BSS
model, due to sudden discontinuities of blending functions. For this reason the
blending function values of every cell are smoothed with the corresponding
values of adjacent cells, weighting the actual cell as strongly as all the adja-
cent cells together. In this way the convergence of the BSS model shows very
similar behaviour to the SST or k-ω model for all investigated test cases.
In the present work the BSS model has been implemented in the unsteady
RANS solver TRACE, which is the DLR’s solver for turbomachinery flows.
The RANS equations are solved using a finite volume technique, where the
convective fluxes are discretized by the TVD upwind scheme by Roe and the
3 3
y/! a) k-" y/! b) BSS
"FS =108 Hz "FS =108 Hz
2 "FS =102 Hz 2 "FS =102 Hz
1 1
0 0
0 20 40 60 µ /µ 80 0 20 40 60 µ /µ 80
T T
Fig. 1. Boundary layer predictions at different free stream conditions
0.10
y [m] Exp.
SST
0.08 BSS
k-!
0.06
T5 T6
0.04
T1 T3
0.02
0.00
0.000 0.002 0.004 0.006 0.008 0.010
2
-u’v’/Ue [-]
Fig. 2. Development of Reynolds stresses for Samuel–Joubert flow; origin shifted
diffusive fluxes are determined by a central differencing scheme. Present inves-

tigations cover only steady computations on parallelly distributed multi-block
meshes. Further information about TRACE and various validation examples
can be found in references [3] and [4].
3 Application of the BSS Turbulence Model
Figure 1 shows the predicted eddy viscosity µT profiles of a flat plate at

different free stream conditions of ω. The dependency of the k-ω model is
successfully attenuated by the BSS approach.
The development of Reynolds stresses in the adverse pressure experiment
by Samuel and Joubert [5] is shown in Fig. 2. In contrast to the k-ω model,
both approaches which apply the Bradshaw assumption (SST and BSS) pre-
dict lower Reynold stresses and are thus in a better agreement with the mea-
surements. Furthermore, the BSS approach yields results almost identical to
those of the SST model.
A comparison of the predicted and measured performance map of the 1-
stage transonic compressor of TU Darmstadt (measurements by Müller [6],
test case details in Blaha et al. [7]) is shown in Fig. 3. The applied grid consists
Contribution of Turbulence Equation Terms to the Shear Stress Balance 515
1.8 0.90
" [-] !is [-]
1.7 0.85
!is
1.6 0.80
1.5 0.75
"
Exp.
1.4 SST 0.70
BSS
1.3 0.65
14.0 14.5 15.0 15.5 16.0 16.5 17.0
m’ [kg/s]
Fig. 3. Performance map of a 1-stage transonic compressor
SST BSS
blade blade
µT/µ: 0 200 400 600 800 1000 µT/µ: 0 200 400 600 800 1000
SST BSS
lines = const. µT
lines = const. µT (cf. upper plot)
(cf. upper plot)
F1: 0 0.2 0.4 0.6 0.8 1

F1: 0 0.2 0.4 0.6 0.8 1
k-" k-!
Fig. 4. Eddy viscosity µT and blending function F1 in the casing boundary layer
of the rotor, xax /lax = const. = 0.7, ṁ = 16 kg/s
of 240 000 nodes and the dimensionless wall distance of the wall-adjacent
nodes is y + ≈ 50. Overall, the predicted pressure ratios and efficiencies for
SST and BSS model agree very well with the measurement at higher mass
flow rates. However, the SST model does not yield converged solutions when
moving towards the lower mass flow range. In contrast, the BSS model is
able to predict stable operating points closer to the stall line, although the
gap to the experimentally acquired stall line is still large. The authors hope to
considerably reduce this gap by including the effects of streamline curvature [3]
to the BSS model. Investigation of Fig. 4 reveals that the blending function F1
of the SST model only partly covers the tip vortex region, which is denoted by
high values of eddy viscosity. Hence, the switching between ω- and ε-equation
starts already in the vortex core region. On the other hand, the BSS model
adopts the blending to the turbulence structure itself, which is the reason
for slightly better predictions of the BSS model at lower mass flow rates.
Besides the side wall region, the boundary layer at the blade surface also
shows reasonable blending.
4 Conclusions
The terms of the k-equation are used for the blending of the SST turbulence
model, leading to improved balancing of shear stresses in complex 3D bound-
ary layers, as demonstrated on a simulation of a transonic compressor. In
addition, the computation of the wall distance is not required. The BSS ap-
proach shows much less dependency on the free stream values and improved
turbulence predictions in adverse pressure gradient flows as compared to the
k-ω model. Future work will focus on the refinement of blending functions,
on application to multistage turbomachines and on the interaction with the
effects of streamline curvature.
References
1. Menter, F.R., Kuntz, M., Langtry, R.: Ten years of industrial experience with the
SST turbulence model. In: Hanjalić, K. (ed) Proc. 4th. Int. Symp. on Turbulence,
Heat and Mass Transfer. Begell House, New York (2003)
2. Wilcox, D.C.: Reassessment of the scale-determining equation for advanced tur-
bulence models. AIAA Journal, 26, 1299–1310 (1988)
3. Kovzulović, D., Röber, T.: Modelling the Streamline Curvature Effects in Tur-
bomachinery Flows. Proc. of ASME Turbo Expo, Paper No. ASME–GT2006–
90265, Barcelona, Spain (2006)
4. Yang, H., Nürnberger, D., Kersken, H.-P.: Toward excellence in turbomachin-
ery computational fluid dynamics: A hybrid structured-unstructured Reynolds-
averaged Navier-Stokes solver. J. Turbomach., 128, 390–402 (2006)
5. Samuel, A.E., Joubert, P.N.: A boundary layer developing in an increasingly
adverse pressure gradient. J. Fluid Mech., 66, 481–505 (1974)
6. Müller, M.: Private Communication. Technical University Darmstadt, Depart-
ment of Gas Turbines and Aerospace Propulsion, Darmstadt (2006)
7. Blaha, C., Hennecke, D.K., Fritsch, G., Hoeger, M., Beversdorf, M.: Laser-2-
Focus Measurements in a Transonic Compressor Blisk-Rotor and Comparison
with 3D Numerical Simulations. Proc. 13th Int. Symp. on Air Breathing Engines,
Paper No. ISABE 97–7069, Chattanooga, USA (1997)
Part VI
Optimisation and Flow Control

Optimum Shape Design for Unsteady
Three-Dimensional Viscous Flows
S. Nadarajah1 and A. Jameson2

1
Department of Mechanical Engineering, McGill University, 688 Sherbrooke St.
West, # 711, Montreal, QC H3A 2S6 Canada. siva.nadarajah@mcgill.ca
2
Department of Aeronautics and Astronautics, Stanford University, California
94305 U.S.A. jameson@baboon.stanford.edu
There are numerous important engineering applications in which the flow is

inherently unsteady but periodic. Helicopter rotors in forward flight, turbo-
machinery blades and cooling fans operate in unsteady flow and are constantly
subjected to unsteady loads. Optimization techniques for unsteady flows are
clearly needed to improve their performance. As yet there have been few efforts
in this direction. One of the major reasons is the demanding computational
cost associated with the calculation of unsteady flows. Nevertheless, the de-
velopment of optimum shape design for two-dimensional unsteady flows using
the time accurate adjoint based design approach has been pursued by Nadara-
jah and Jameson [2]. This work is largely based on algorithms developed for
Aerodynamic Shape Optimization (ASO) for a steady flow environment [1, 4].
Nadarajah derived and applied the time accurate adjoint equations (both the
continuous and discrete) to the redesign of an oscillating airfoil to reduce the
time-averaged drag coefficient in an inviscid transonic flow. However, the en-
couraging results were obtained at a substantial computational expense. The
motivation of the research in this paper has been fueled both by the success
of our current capability for automatic shape optimization for unsteady flows
and the future potential of the NLFD method. The general goal of this work is
to extend the NLFD method for adjoint based design from two-dimensional [3]
to three-dimensional flows.
1 Development of the Non-Linear Frequency Domain

Adjoint Equations
The semi-discrete form of the governing equations can be written as
∂w
V + R = 0. (1)
∂t
The derivation of the NLFD method starts with the semi-discrete form of the
governing equations, and assumes that the solution w and spatial operator

520 S. Nadarajah and A. Jameson
P N2 −1 ikt
R can be represented by separate Fourier series: w = k=− N ŵk e , R=
2
P N2 −1 √
k=− N
R̂k eikt where, i = −1. Substitution of the Fourier series represen-
2
tation into equation (1) yields the unsteady residual, Rˆ∗ k = i kV ŵk + R̂k . The
nonlinearity of the unsteady residual stems from the spatial operator which
can be computed using a Fast Fourier Transform (FFT) algorithm. Consistent
with the time accurate approach, a pseudo-time derivative can be added, and
a time-stepping scheme can be employed to numerically integrate the resulting
equations; V ∂∂τ ŵk
+ Rˆ∗ k = 0. In the NLFD case, an unsteady residual exists
for each wavenumber used in the solution and the pseudo-time derivative acts
as a gradient to drive the absolute value of all of these components to zero
simultaneously.
The NLFD discrete adjoint equation is developed using a pseudo-spectral
approach similar to the one applied to the flow equations, which can be as
expressed as V ∂ψ ∂t + R(ψ) = 0, where R(ψ) is the sum of all the spatial
operators, both convective and dissipative, used in the discretized adjoint
equations. Next, we assume that the adjoint variable and spatial operator can
be expressed as a Fourier series. The derivation of the NLFD adjoint then
follows that of the NLFD flow equations. The NLFD adjoint equations are
∗ ∗
expressed as, V ∂∂τ ψ̂k [ = 0, where R(ψ)
+ R(ψ) k
[ = i kV ψ̂k + R(ψ)
k
[ .
k
2 Results
The following subsections presents results from simulations of a three-dimensio-

nal wing undergoing a change in angle of attack as a function of time. The
flight conditions (reduced frequency, ωr = 0.102, and M∞ = 0.822) are based
on Run 73 of the central transonic test case CT5 conducted by R.J Zwaan at
the NLR [5]. The first part of the results section contains a code validation
study based on a 257 × 65 × 97 structured grid. A redesign of the LANN wing
is demonstrated, in the second section.
Figures 1 and 2 illustrates the convergence of the viscous NLFD flow and
adjoint solvers for three, five, and seven time steps per period. Only 400
multigrid cycles were needed to converge the adjoint solver to the same level
of accuracy as the flow solver. All modes converge at similar rates. The lift
hysteresis is demonstrated in Fig. 2 for various number of time steps per
period. Here, as indicated in the legend, ’NLFD 1-Harmonic’ is synonymous
to three time steps per period. As the wing oscillates at a small angle of
attack, the shock wave moves back and forth about a mean location and
is closely sinusoidal and lags the wing motion. The nonlinear behavior of
unsteady viscous transonic flows is primarily due to the movement of the
shock and this is evident in Fig. 2. One harmonic is sufficient to produce
an accurate lift hysteresis; however, Fig. 3 demonstrates that at least two
harmonics are needed to capture the variation of the drag coefficient.
Optimum Shape Design for Unsteady Three-Dimensional Viscous Flows 521
3
10
NLFD−3 0−Mode
NLFD−3 1−Mode
0
2 NLFD−5 0−Mode 10
10 NLFD−5 1−Mode
NLFD−5 2−Mode
NLFD−7 0−Mode
1 NLFD−7 1−Mode
10 NLFD−7 2−Mode −1
10
NLFD−7 3−Mode
log(Residual)
log(Residual)
0
10
−2
10
−1
10
−3
−2
10
10
−3
10 −4
10
−4
10
200 400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 50 100 150 200 250 300 350 400
Multigrid Cycles Multigrid Cycles
Fig. 1. Viscous NLFD Flow Solver Fig. 2. Viscous NLFD Adjoint Solver
Convergence; M∞ = 0.822, ωr = 0.102 Convergence; M∞ = 0.822, ωr = 0.102
In Fig. 5, a validation of the surface pressure coefficient is presented. The

figure illustrates the pressure distribution at two separate span stations. In
Fig. 5a results based on both the inviscid and viscous NLFD methods are
compared to experimental data at the span location, η = 20% and angle of
attack, α = 0.59◦ . The inviscid NLFD results compare closely to the viscous
solutions, however, the location of the shock for the inviscid NLFD solution
differs by 5% of the chord. Nevertheless both solutions differ from the experi-
mental work. One possible reason for the difference could be due to the effect
of the turbulence model on the unsteady viscous flows. A second important
outcome of Fig. 5 is the required number of time steps per period or modes to
accurately resolve the pressure distribution. At the 20% span station, three
time steps per period is sufficient; however, at the 65% span location, a slight
discrepancy is observed at the shock location, thus requiring five time steps
to effectively resolve the flow field. This is further illustrated in Fig. 6 that
demonstrates the real and imaginary components of the first mode of the
pressure coefficient.
2.1 LANN wing: Redesign
This section documents the results of the redesign of the LANN Wing to re-
duce the time-averaged drag coefficient for a fixed time-averaged lift coefficient
and wing thickness ratio. The simulation was performed at a Mach number,
M∞ = 0.82, reduced frequency, ωr = 0.102, and deflection of ±0.25◦ about
the zero angle of attack. As established in the previous section, five time steps
per period will be used for the redesign of the LANN Wing. During the ini-
tial stage to compute the flow solution, the time-averaged lift and drag are
computed and used as the target lift and objective function.
Figure 7 illustrates the wing surface pressure contour and the initial and
final pressure distributions at three span locations at the 0◦ phase after 50
design cycles. The pressure contour illustrates the severe weakening of the
λ-shock system and this is further validated in the three pressure coefficient
plots at span stations 6.2%, 49.2%, and 92.3% (initial pressure-dotted line;
final design-solid). The mean angle of attack was perturbed from the initial
zero degrees to 0.344◦ to maintain the time-averaged lift coefficient at 0.348.
The time-averaged drag coefficient reduced by 5.65% from 0.0115 to 0.01085
within 50 design cycles.
Figure 8 illustrates the gradients of the objective function for various tem-
poral modes and demonstrates the need for at least five time steps per period
to capture accurate gradients. Lastly, Fig. 9 presents the convergence of the
time-averaged drag coefficient from 115 drag counts to 108.5.
3 Conclusion
Five time steps per period is sufficient to resolve the unsteady viscous tran-
sonic flow field via the NLFD approach. A redesign of the LANN Wing has
been demonstrated with a reduction of the time-averaged drag coefficient by
5.65% while maintaning the time-averaged lift coefficient constant. Accurate
gradients are achieved with five time steps per period as well. These results fur-
ther demonstrate the potential of the method to provide significant improve-
ments to more realistic problems such as helicopter rotors, turbomachinery,
and other unsteady devices operating in the transonic regime.
References
1. A. Jameson. Computational aerodynamics for aircraft design. Science, 245:361–
371, 1989.
2. S. Nadarajah and A. Jameson. Optimal control of unsteady flows using a time
accurate method. AIAA paper 2002-5436, 9th. AIAA/ISSMO Symposium on
Multidisciplinary Analysis and Optimization Conference, Atlanta, GA, Septem-
ber 2002.
3. S. Nadarajah, M. McMullen, and A. Jameson. Optimum shape design for un-
steady flows using time accurate and non-linear frequency domain methods.
AIAA paper 2003-3875, 16th. Computational Fluid Dynamics Conference, Or-
lando, FL, June 2003.
4. J. Reuther, A. Jameson, J. Farmer, L. Martinelli, and D. Saunders. Aerody-
namic shape optimization of complex aircraft configurations via an adjoint for-
mulation. AIAA paper 96-0094, 34th Aerospace Sciences Meeting and Exhibit,
Reno, Nevada, January 1996.
5. R. J. Zwaan. Data set 9, LANN wing. Pitching oscillation. Technical report.
Agard-R-702 Addendum no. 1., AGARD, Amsterdam, Netherlands, 1985.
Optimum Shape Design for Unsteady Three-Dimensional Viscous Flows 523
0.44 0.0155
NLFD 1−Harmonic
NLFD 2−Harmonic 0.015
0.43
NLFD 3−Harmonic
0.0145
0.42
Drag Coefficient
Lift Coefficient
0.014
0.41 0.0135
0.013
0.4
0.0125
0.39
0.012
0.0115
0.38
0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.75 0.8
Angle of Attack Angle of Attack
Fig. 3. Comparison of Lift Hysteresis Fig. 4. Comparison of Variation of

for Various Modes Drag Coefficient versus Angle of Attack
for Various Modes
−1
−1 NLFD 3 Viscous NLFD 3 Viscous
NLFD 5 Viscous NLFD 5 Viscous
NLFD 7 Viscous −0.8 NLFD 7 Viscous
−0.8 NLFD 3 NLFD 3
NLFD 11 NLFD 11
Exp Upper −0.6 Exp Upper
−0.6 Exp Lower Exp Lower
−0.4
−0.4
Pressure Coefficient, Cp
Pressure Coefficient, Cp
−0.2
−0.2
0
0
0.2 0.2
0.4 0.4
0.6 0.6
0.8 0.8
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
◦ ◦
(a) η = 20%, α = αmean = 0.59 (b) η = 65%, α = αmean = 0.59
Fig. 5. Comparison of Pressure Distribution between Inviscid, Viscous, and Exper-

imental Data
0 0
NLFD 3 Viscous
0.02 0.05 NLFD 3 Viscous
NLFD 5 Viscous
NLFD 5 Viscous
NLFD 7 Viscous
0.04
0.1 NLFD 7 Viscous
imag(Cp )
real(Cp )
0.15
0.06
0.2
0.08
0.25
0.1
0.3
0.12
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x/c x/c
(a) η = 65%, real (cp ) (b) η = 65%, imag (cp )
Fig. 6. Comparison of the Real and Imaginary Components of the First Harmonic
Pressure Coefficient for Various Modes
Cp = -2.0
NLR LANN WING
Mach: 0.822 Alpha: 0.344 IT: 1
CL: 0.332 CD: 0.01512 CM:-0.3034
TIMEAVE CL: 0.348 TIMEAVE CD: 0.01085
Design: 50 Residual: 0.1135E+01 Grid: 257X 65X 49
Tip Section: 92.3% Semi-Span

Cl: 0.285 Cd:-0.01131 Cm:-0.1608
Cp = -2.0 Cp = -2.0
Root Section: 6.2% Semi-Span Mid Section: 49.2% Semi-Span

Cl: 0.288 Cd: 0.04358 Cm:-0.1108 Cl: 0.367 Cd: 0.00143 Cm:-0.1368
Fig. 7. Initial and Final Pressure Distribution for Various Span Locations at Phase
= 0 deg
−3
x 10
2 0.0115
1
0.0114
0
−1 0.0113
Gradient Value
−2
Drag Coefficient
0.0112
−3
−4
0.0111
−5
−6 0.011
−7
0.0109
−8
0.0108
20 40 60 80 100 120 140 160 0 5 10 15 20 25 30 35 40 45 50
Grid Point Design Cycles
Fig. 8. Comparison of Gradient for Fig. 9. Convergence of Time-Averaged

Various Modes for the LANN Wing, Drag Coefficient
Run 73, M∞ = 0.82, ωr = 0.102
Comparison of Exact and Approximate
Discrete Adjoint for Aerodynamic Shape
Optimization
Giampietro Carpentieri1 , Michel J.L van Tooren2 , and Barry Koren3

1
Delft University of Technology G.Carpentieri@tudelft.nl
2
Delft University of Technology M.J.L.vanTooren@tudelft.nl
3
Delft University of Technology B.Koren@tudelft.nl
Summary. The effect of approximations in the discrete adjoint on constrained

shape optimization is investigated. The different approximations are compared with
the exact discrete adjoint, which is capable of producing exact gradient/sensitivity
information. The purpose of such a comparison is to understand whether or not
approximate adjoint codes can be effective for shape optimization in spite of the
error in the computed gradient.
1 The Discrete Adjoint

The discrete adjoint method is well established in the field of aerodynamic
shape optimization [7]. In order to implement the method, the residuals vector
R must be differentiated to obtain the residuals Jacobian, ∂R/∂U. The latter
must be transposed and used to solve a linear system of equations with the
adjoint variables Λ as unknown:
∂R T ∂J T
Λ= , (1)
∂U ∂U
where U is the conservative variables vector and J a functional (e.g. lift, drag
or pitching Moment coefficient) that is to be optimized. The vectors U, R
and Λ have dimensions N × nvar, with N the number of nodes in the mesh
and nvar the number of variables (for the 2D Euler equations nvar = 4).
The practical implementation of the method can be challenging due to the
differentiation of the residuals vector and the derivation of the assembly for the
left–hand side of Eq. (1), which preferably is to be performed in a matrix–free
fashion due to the large storage that otherwise would be required.
In the present work, the flow solver is based upon a median–dual discretiza-
tion and a MUSCL–type reconstruction scheme. On each edge of the mesh
the second order numerical fluxes are computed by means of Roe’s approxi-
mate Riemann solver using reconstructed primitive variables. Reconstruction

526 Giampietro Carpentieri, Michel J.L van Tooren, and Barry Koren
is performed using a least–squares gradient and a multi–dimensional type of

limiter [2]. For the purpose of deriving the transposed residuals Jacobian, the
vector of numerical fluxes, H, and the vector of reconstructed left and right
states, UL and UR , can be introduced [3]. The three vectors have dimen-
sions E × nvar, with E the number of edges in the mesh. According to the
MUSCL–type scheme the residuals vector has a dependency on the conser-
vative variables vector of the type R = R(H(UL (U), UR (U))). Therefore,
by means of the chain rule and transposition, the left–hand side of Eq. (1)
becomes:
!
∂R T ∂UL T ∂H T ∂UR T ∂H T ∂R T
Λ= + Λ, (2)
∂U ∂U ∂UL ∂U ∂UR ∂H
where ∂R/∂H is a dummy matrix that only contains −1 and 1. ∂H/∂UL
and ∂H/∂UR are diagonal matrices containing the differentiation of the
fluxes with respect to the left and the right states respectively. ∂UL /∂U and
∂UR /∂U are sparse rectangular matrices that represent the reconstruction
operator, i.e., the differentiation of the reconstructed states with respect to
the conservative variables.
In practice, none of the matrices introduced in Eq. (2) is stored but more
likely the assembly is performed on the edges similarly to what is done for
the residuals vector. Details about the edge–based assembly can be found
in [4] where Eq. (2) is derived in detail, rather than only in abstract matrix
form. Exactness of the differentiation was demonstrated showing that the
residuals Jacobian is capable of yielding quadratic convergence when employed
in Newton iterations. Quadratic convergence can only be attained when the
Jacobian is exact [2].
Approximations can be introduced in the differentiation in order to sim-
plify the implementation of the discrete adjoint. Three approximations are
considered in this work: (i) The first approximation is obtained by neglecting
the differentiation of the limiter in the reconstruction operator. The simplifi-
cation is appreciable since the limiter implemented here requires a construc-
tion phase which is quite involved compared to that of mono–dimensional
limiters [2]. (ii) The second approximation is obtained by neglecting the dif-
ferentiation of the Jacobian matrix in the Roe flux. The approximation saves
lot of human work since the differentiation is cumbersome [1]. (iii) The third
approximation is that obtained by ignoring the complete reconstruction op-
erator, which makes the implementation of the adjoint trivial.
The price to pay for the simplifications is the detrimental effect on the
accuracy of the computed gradient. For instance, for the test cases presented
below, the first and second approximations show a percentage error of 0.1–
2.5% compared to the exact adjoint code. The error increases to 10–30% for
the third approximation. However, the error on the gradient can’t be used to
judge whether or not an approximation is acceptable. It seems more reasonable
to consider the effect that the gradient has on the behavior of the optimization
process [6, 4].
Comparison of Exact and Approximate Discrete Adjoint 527
2 Numerical results
Numerical results for constrained shape optimization are presented. The pur-
pose is to minimize the drag coefficient while keeping the lift coefficient of
the airfoil constant by means of an equality constraint. Also, lower inequal-
ity constraints are enforced on the relative maximum thickness, on the nose
radius and on the trailing edge angle. Unconstrained optimization algorithms
that include the constraints as penalty terms in the objective could be used.
However, their accuracy is known to be poor and moreover they can be ill–
conditioned. Therefore, optimization algorithms capable of handling directly
equality and inequality constraints are necessary. Two algorithms have been
employed: a widely used sequential quadratic programming (SQP) algorithm
and a sequential linear programming (SLP) algorithm known as the method
of centers [8].
(a) Mesh (b) NACA64A410 (c) Exact
(d) Approx 1 (e) Approx 2 (f) Approx 3
1.4
1.4
0.06
1.2 EXACT, [ 49/19 ]
1.2
0.04 1 APPROX 1, [ 44/17 ]
1
0.02 0.8 0.8 APPROX 2, [ 86/34 ]
0 0.6 0.6 APPROX 3, [2000/17]
0.4 NACA64A410 0.4

-0.02
EXACT
EXACT 0.2 APPROX 1
APPROX 1 APPROX 2 0.2
-0.04
APPROX 2 APPROX 3
0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0
0 2 4 6 8 10 12 14
(g) Airfoils (h) Mach number (i) Objective
Fig. 1. Optimization of the NACA64A410 airfoil.

(a) RAE2822 (b) Exact (c) Approx 1
0.06
0.04
0.02
-0.02
-0.04
-0.06 EXACT
APPROX 1
APPROX 2
-0.08
0 0.2 0.4 0.6 0.8 1
(d) Approx 2 (e) Approx 3 (f) Mach number
Fig. 2. Optimization of the RAE2822 airfoil.
The flow and the adjoint equations are solved using an implicit pseudo–
time stepping scheme. During the optimization, the two solvers are converged
until the residual norm is reduced about 6 orders of magnitude. A spring
analogy is used to deform the mesh and automatic differentiation in forward
mode has been used to derive the geometric sensitivities, i.e., the derivatives
with respect to the shape parameters. The shape of the airfoil is parameterized
by means of Chebyschev polynomials. Further details can be found in [4, 5].
The computations presented below have been performed on an unstruc-
tured mesh of triangles, see Fig. 1.a, with 12161 nodes, 500 of which are on the
airfoil. The mesh has been deformed in order to be used for all airfoils. In or-
der to make sure that the mesh was capable of capturing weak shocks, a mesh
of 30092 nodes, 700 of which are on the airfoil, has been used to check the
optimization results. The shock–free pressure distributions obtained on the
first mesh after optimization have been re–obtained on the second mesh. The
SQP algorithm from the Matlab Optimization Toolbox4 has been used for the
NACA64A410 and RAE2822 cases whereas the SLP algorithm [5] has been
used for the NACA0012 case. For the two algorithms a tolerance of 10−5 for
both the objective function and the constraints has been adopted as criterion
to stop the optimization.
NACA64A410, M∞ = 0.75 and α = 0 deg angle of attack. The pressure con-
tours for the original airfoil are shown in Fig. 1.b. The airfoils optimized using
the exact and the first two approximations are shock–free, see Figs. 1.c, 1.d
4
Matlab 1984–2004,
c The Mathworks, Inc.
Comparison of Exact and Approximate Discrete Adjoint 529
and 1.e. As can be seen from Fig. 1.g the three airfoils are different. Com-
pared to the airfoil obtained from the exact adjoint, the other two airfoils
show a maximum difference in y–coordinates of 5.6 × 10−3 and 2 × 10−3 re-
spectively. However, they all satisfy the design problem accurately. The airfoil
obtained using the third approximation shows three weak shocks on the upper
side because its optimization has stalled, see Fig. 1.f. In Fig. 1.i the objec-
tive function, which is the drag divided by its initial value, is shown at each
gradient iteration and, in the legend, the ratio of function/flow evaluations to
gradient/adjoint evaluations is shown. The ratio is around 2.6 except for the
third approximation. The latter has exceeded the maximum number, 2000, of
allowed function calls.
RAE2822, M∞ = 0.75 and α = 2 deg angle of attack. The pressure contours
for the original airfoil are shown in Fig. 2.a. As before, the airfoils obtained
from the exact and the first two approximate adjoint codes are shock–free, see
Figs. 2.b, 2.c and 2.d. The geometry of the three airfoils is different, especially
for the first 40% of the chord on the lower side, see Fig. 2.f. Maximum differ-
ence in y–coordinates are of the order of 10−3 . The optimization stalled when
the third approximation was used. The airfoil still exhibits a weak shock, see
Fig. 2.e.
(a) NACA0012 (b) Exact (c) Approx 3
Fig. 3. Optimization of the NACA0012 airfoil.
NACA0012, M∞ = 0.75 and α = 2 deg angle of attack. The pressure contours

for the original airfoil are shown in Fig. 3.a. The SLP algorithm seems to be
insensitive to the approximations and identical shock–free airfoils are obtained
when using the exact and the first two approximations. The airfoils can be
said to be identical since the maximum difference between the y–coordinates
is of the order of 10−4 . Since they are identical, only one of them is shown,
see Fig. 3.b. When using the third approximations, also the SLP algorithm
stalls and the resulting airfoil exhibits a shock on the upper side, see Fig. 3.c.
3 Conclusions
It has been shown that the first two approximate adjoint codes considered
in this work can be effective for shape optimization purposes. However, the
SQP algorithm has shown to converge to different airfoils, i.e., different local
minima, for each approximation. Nevertheless, each of these airfoils satisfied
the design problem. The SLP algorithm appeared to be insensitive to the two
approximations and converged to a single airfoil.
The third approximation, which neglects the reconstruction operator, has
shown to be ineffective. For the third approximation, both the SQP and the
SLP algorithms have stalled during the optimization process, failing to con-
verge to a shock–free airfoil.
Acknowledgments: This research was supported by the Dutch Technology
Foundation STW, applied science division of NWO and the technology pro-
gram of the Dutch Ministry of Economic Affairs.
References
1. Barth, T.J.:Analysis of Implicit Local Linearization Techniques for Upwind and
TVD Algorithms. AIAA Paper 87–0595 (1987).
2. Barth, B.J.: Aspects of Unstructured Grids and Finite–Volume Solvers for the
Euler and Navier–Stokes Equations. VKI Lecture Series, 1991-06 (1991).
3. Barth, T.J., Linton S.W.: An Unstructured Mesh Newton Solver for Compress-
ible Fluid Flow and its Parallel Implementation. AIAA Paper 95–0221 (1995).
4. Carpentieri, G., van Tooren, M.J.L., Koren, B.: Improving the Efficiency of Aero-
dynamic Shape Optimization on Unstructured Meshes. AIAA Paper 2006-298
(2006).
5. Carpentieri, G., van Tooren, M.J.L., Koren, B.: Aerodynamic Shape Optimiza-
tion by Means of Sequential Linear Programming Techniques. ECCOMAS CFD
2006 Paper, to appear (2006).
6. Dwight, R.P., Brezillon, J.: Effect of Various Approximations of the Discrete
Adjoint on Gradient–Based Optimization. AIAA Paper 2006-690 (2006).
7. Giles, M.B., Pierce, N.A.: An Introduction to the Adjoint Approach to Design.
Flow, Turbulence and Combustion, 65, 393–415 (2000)
8. Vanderplaats, G.N.: Numerical Optimization Techniques for Engineering Design.
3th Edition, Vanderplaats Reasearch & Development (2001)
Adjoint Sensitivity Computations for an
Embedded-Boundary Cartesian Mesh Method
and CAD Geometry
Marian Nemec1 and Michael J. Aftosmis2

1
ELORET Corp., Moffett Field, CA 94035, USA mnemec@mail.arc.nasa.gov
2
NASA Ames Research Center, CA 94035, USA maftosmis@mail.arc.nasa.gov
Key words: shape optimization, adjoint, Cartesian mesh, cut-cells
1 Introduction
Complex geometry remains a challenging issue facing the application of ad-
joint and flow-sensitivity methods in practical engineering design. Among the
most promising approaches for complex-geometry problems is the embedded-
boundary Cartesian mesh method [1]. In this approach, the discretization of
the surface geometry is decoupled from that of the volume mesh, which enables
rapid and robust mesh generation, and ultimately, an automatic analysis of
aerodynamic performance. The purpose of this work is to extend the automa-
tion and efficiency of Cartesian methods to the computation of aerodynamic
sensitivities for shape optimization problems.
The most common way to account for the effect of boundary shape per-
turbations in the adjoint and flow-sensitivity formulations is via domain map-
ping approaches. In implementations on body-fitted structured or unstructured
meshes, this involves the use of mesh-perturbation schemes. For local shape
deformations, the extent of the mesh perturbations can be limited to just
the boundary cells [2, 3, 4]. The approach we propose here for non-body-
fitted Cartesian meshes is similar, but the boundary faces of the volume mesh
maintain their Cartesian orientation as the surface evolves. This approach
permits the computation of mesh sensitivities via a direct linearization of the
boundary-cell geometric constructors of the mesh generator.
2 Discrete Adjoints and Flow Sensitivities

Our goal is to minimize a scalar objective function J, such as drag, by adjust-
ing a design variable X using gradient-based optimization. To compute the
gradient, dJ/dX, we use a discrete formulation. Hence, a variation in X influ-
ences the computational mesh M and the flow solution Q, which satisfies the

532 M. Nemec and M. J. Aftosmis
three-dimensional Euler equations of a perfect gas. The spatial discretization

of the flow equations uses a cell-centered, second-order accurate finite volume
method with a weak imposition of boundary conditions, resulting in a system
of equations
R(X, M, Q) = 0 (1)
where M is an explicit function of the surface triangulation T: M = f [T(X)].
The gradient of the objective function J(X, M, Q) is given by
dJ ∂J ∂J ∂M ∂T ∂J dQ
= + + (2)
dX ∂X ∂M |∂T{z∂X} ∂Q dX
A
The evaluation of the term dQ/dX, referred to as the flow sensitivities, is

obtained by linearizing
 Eq. 1 
∂R dQ  ∂R ∂R ∂M ∂T 
= − +  (3)
∂Q dX  ∂X ∂M |∂T{z∂X}
B
The adjoint equation is obtained by combining Eqs. 3 and 2 and defining

the following intermediate problem
∂R T ∂J T
ψ= (4)
∂Q ∂Q
where the vector ψ represents the adjoint variables. Details of the solution
method for Eqs. 3 and 4 are given in [5]. We focus on the evaluation of the
terms A and B in Eqs. 2 and 3, which are discussed in the next section.
3 Shape Sensitivities
The flow equations are discretized on a multilevel Cartesian mesh. The mesh
consists of regular Cartesian hexahedra everywhere, except for a layer of body-
intersecting cells, or cut-cells, adjacent to the boundaries. An infinitesimal
perturbation of the boundary shape affects only the cut-cells. Consequently,
the mesh-sensitivity term ∂M/∂T, which contains the linearization of the
Cartesian-face areas and centroids, volume centroids and the wall normals
and areas with respect to the surface triangulation, is non-zero in these cells.
The crux in the evaluation of ∂M/∂T is the linearization of the geometric con-
structors that define the intersection points between the surface triangulation
and the Cartesian hexahedra.
We explain the salient steps of the linearization using the example shown
in Fig. 1, where a Cartesian hexahedron is split into two cut-cells by the
surface triangulation. We require the linearization of the intersection points
that lie on Cartesian edges, e.g., point A, and also those that lie on triangle
edges, e.g., point B. Focusing on point B, its location along the triangle edge
V0 V1 is given by
Adjoint Sensitivity Computations 533
Fig. 1. Sensitivity of face centroids (solid vectors) to perturbation of vertex V1 .
ro
Inflow
Mi
ri r
!
Outflow
Fig. 2. Problem setup (Mi = 2.25, ri = 1 and ro = 1.382)
B = V0 + s(V1 − V0 ) (5)
where s denotes the distance fraction of the face location relative to the ver-
tices V0 and V1 . The linearization of this geometric constructor is given by
∂B ∂V0 ∂V1 ∂V0 ∂s

= + s( − ) + (V1 − V0 ) (6)
∂X ∂X ∂X ∂X ∂X
A similar constructor is used for point A [6]. An example result of the lin-
earization is shown in Fig. 1 for the position sensitivity of Cartesian face
centroids. Note that the “motion” of the face centroids is constrained to the
plane of the face. An advantage of this formulation is that the dependence of
the mesh sensitivities ∂M/∂T on the surface triangulation (∂V1 /∂X in Eq. 6
and the term ∂T/∂X in Eq. 3) is determined on-the-fly for each instance of
the surface geometry. Put another way, there is no requirement for a one-to-
one triangle mapping as the surface geometry changes. This allows a flexible
interface for geometry control based on tools such as computer-aided design
(CAD).
4 Verification Studies
Supersonic Vortex Problem. We investigate the error convergence rate of
a representative objective function and its gradient on a model problem with
a known analytic solution [6]. The problem involves isentropic flow between
concentric circular arcs at supersonic conditions, as shown in Fig. 2. The ob-
jective function is the integral of pressure along the outer arc, which is similar
to the lift and drag objectives used in aerodynamic design. We compute the
-1
ρ Objective (J)
-2 dρ/dMi dJ/dMi
-2
log10(L1 Error)
log10(Error)
-3 -3
-4
Slope=2.04
-4 Slope=2.05
-5
1 1.25 1.5 1.75 2 1 1.25 1.5 1.75 2
0.5*log10(Cells) 0.5*log10(Cells)
(a) Density and density gradient (b) Objective function and gradient
Fig. 3. Error convergence (slopes are computed using the three finest meshes)
gradient and the sensitivities of the flow solution, Eq. 3, with respect to the
inlet Mach number, Mi . The problem is solved on a sequence of five nested
Cartesian meshes. Figure 3 summarizes the results. Figure 3(a) shows the er-
ror convergence rates in the L1 norm of density and its sensitivity to variations
in Mi . The error convergence rate of the objective function and its gradient
is shown in Fig. 3(b). The asymptotic convergence rate of all errors, which is
measured over the three finest meshes, is just slightly over 2. These results
verify the accuracy of the linearization and the convergence of these methods
to the continuous problem.
Shape sensitivities for the NACA 0012 airfoil. The goal of this verifica-
tion test is to study the convergence of gradients for design variables that alter
the shape of the boundary. We consider a subsonic flow over the NACA 0012
airfoil at M∞ = 0.5 and α = 1 deg. We investigate the sensitivity of lift to
the angle of attack using two equivalent approaches. First, we consider the
influence of the angle of attack via a change in the farfield boundary condi-
tions, which does not alter the relationship between the mesh and the airfoil.
We contrast this with an angle of attack change implemented via a rigid-body
rotation of the airfoil about its trailing edge within a fixed mesh. The mesh-
refinement study is performed on a sequence of five nested Cartesian meshes
for each airfoil orientation.
The results are summarized in Fig. 4. Referring to Fig. 4(b), note that the
differences in the gradients between the farfield and rigid-body rotation cases
are decreasing as the mesh is refined. Additional regression analysis of this
data indicates that the rate of convergence is first-order. This is a consequence
of the fact that the mesh perturbations are confined to only the cut-cells.
Unlike the supersonic vortex problem, a perturbation of the angle of attack
via rigid-body rotation modifies the cut-cell boundary and introduces an error
in the objective function proportional to second-order spatial discretization.
Nevertheless, the gradient values for the rigid-body rotation case, even on the
coarsest mesh of roughly 3, 200 cells, are within 1% of the fine-mesh values.
Adjoint Sensitivity Computations 535
5 Design Example: Reentry Capsule

We consider the optimization of a heat-shield shape for a reentry capsule. The
objective is to enhance the lift-to-drag ratio, L/D, of the capsule, thereby im-
proving trajectory control for landing-site selection, and reducing the reentry
load factor and heat rates. The Pro/ENGINEER R
Wildfire CAD system is
used to create the geometry model. The capsule configuration is shown in
Fig. 5(a), where the design variables are the three labeled spline points of the
heat-shield center-line. The freestream conditions are M∞ = 10 and α = 156.5
deg., measured clockwise from the positive x-axis. High-temperature effects
are approximated by the use of an “effective” ratio of specific heats, γ. We use
γ = 1.3. The initial (symmetric) capsule generates an L/D of 0.37, which is
attained using a center-of-gravity offset. The target value of L/D is set to 0.4.
The volume mesh contains roughly 665, 000 cells and we use 64 processors to
solve the flow and adjoint equations.
Convergence of the optimization problem is shown in Fig. 5(b). The target
L/D is reached within five design iterations and the L2 norm of the gradient is
reduced by roughly four orders of magnitude. The initial and final heat-shield
shapes are shown in Figs. 5(c) and 5(d), respectively. The shape modifica-
tions are relatively minor, yet the improvement in L/D is 8%. The wall-clock
time per design iteration is approximately 11 minutes. This time includes the
regeneration and triangulation of the part, as well as the flow solution and
adjoint gradient computation. We emphasize that for problems with more
design variables, the design-cycle time would remain essentially constant.
6 Conclusions
We have presented an approach for the computation of aerodynamic shape
sensitivities using a discrete formulation on Cartesian meshes with cut-cells at
the wall boundaries. The verification studies show that the convergence rate
of gradients is second-order for design variables that do not alter the bound-
ary shape, and is reduced to first-order for shape design variables. This is a
consequence of confining the mesh sensitivities to the cut-cells. The design
0.142
0.142
Farfield
Shape
0.14 0.14
dCL/dα
Farfield
CL
Shape
0.138 0.138
0.136
0.136
0 0.0001 0.0002 0.0003 0 0.0001 0.0002 0.0003

1/Cells 1/Cells
(a) Lift (b) Gradient of lift
Fig. 4. Convergence of lift and its gradient with respect to angle of attack: imple-
mented via the farfield boundary (“Farfield”) and rigid-body rotation (“Shape”)
1e-02
1e-04
1e-06
Objective Function
L2 Gradient
1e-08
1e-10
0 1 2 3 4 5
Design Iterations
(a) CAD model (b) Convergence history
(c) Initial shape (d) Optimized shape

Fig. 5. Heat-shield shape optimization (M∞ = 10, α = 156.5◦ , γ = 1.3)
example demonstrates the effectiveness of the new approach for engineering

design studies that require a fast turn-around and include CAD-based geom-
etry, complex flow, and many design variables.
Acknowledgments. The authors gratefully acknowledge Robert Haimes (MIT) for
help with CAPRI. This work was supported by the NASA Ames Research Center
contract NNA05BF35C.
References
1. Aftosmis, M. J., Berger, M. J., and Melton, J. E., “Robust and Efficient Carte-
sian Mesh Generation for Component-Based Geometry,” AIAA Journal , Vol. 36,
No. 6, 1998, pp. 952–960.
2. Anderson, W. K. and Venkatakrishnan, V., “Aerodynamic Design Optimization
on Unstructured Grids with a Continuous Adjoint Formulation,” Computers &
Fluids, Vol. 28, 1999, pp. 443–480.
3. Dadone, A. and Grossman, B., “Efficient Fluid Dynamic Design Optimization
Using Cartesian Grids,” AIAA Paper 2003–3959, 2003.
4. Lu, J., An a posteriori Error Control Framework for Adaptive Precision Op-
timization using Discontinuous Galerkin Finite Element Method , Ph.D. thesis,
MIT, 2005.
5. Nemec, M., Aftosmis, M. J., Murman, S. M., and Pulliam, T. H., “Adjoint
Formulation for an Embedded-Boundary Cartesian Method,” AIAA Paper 2005–
0877, 2005.
6. Nemec, M. and Aftosmis, M. J., “Aerodynamic Shape Optimization Using a
Cartesian Adjoint Method and CAD Geometry,” AIAA Paper 2006–3456, 2006.
Optimum shape design through multilevel
gradient-based method using Bézier
parametrisation
M. Martinelli1 and F. Beux2

1
Scuola Normale Superiore di Pisa, Piazza dei Cavalieri, 7 - 56126 Pisa, Italy
max.martinelli@sns.it
2
Scuola Normale Superiore di Pisa, Piazza dei Cavalieri, 7 - 56126 Pisa, Italy
fbeux@sns.it
1 Introduction
In the context of aerodynamic optimum shape design, the evaluation of the

cost functional requires the solution of the governing flow equations, and thus,
each optimisation iteration is characterised by a high computational cost if
problems of industrial interest are considered. Thus, the improvement of op-
timisation algorithm efficiency is a crucial point. In [1] a multilevel strategy
considering shape grid-point coordinates as design variables has been defined
to make the convergence rate of the gradient-based method almost indepen-
dent of the number of control parameters. The idea of a multilevel strategy for
shape optimisation has also been used in [2], but, in a context slightly differ-
ent. Indeed, the study was not focused on the gradient method and moreover,
a polynomial representation of the shape was considered with Bézier control
points as design variables.
In the present study, a new multilevel strategy based on the use of Bézier
control points, but in the context of a gradient-based method, is proposed.
Indeed, the present algorithm is grounded on some basic concepts yet proposed
in [2], as, for instance, the degree elevation property of the Bézier curves, but
on another hand, can be also interpreted as a multilevel strategy as defined
in [1], in which a particular prolongation operator is applied. Furthermore, a
descent direction is always obtained considering as control parameters Bézier
control points as well as shape-grid points. A first validation is proposed on a
classical 2D nozzle inverse problem for inviscid flows.

538 M. Martinelli and F. Beux
2 Shape grid-point multilevel gradient method

Among the different algorithms to solve optimal aerodynamic shape prob-
lems, gradient-based methods, which require the computation of the sensitiv-
ity derivatives with respect to the control variables γ, are often employed.
In this context, a descent direction method, based on a preconditioned
gradient, has been defined in [1]. The basic idea is to minimise the cost func-
tional alternatively on different control subspaces according to multigrid-like
cycles. More particularly, considering the ordinates of the shape grid-points as
design variables, a hierarchical parametrisation has been defined considering
different subsets of parameters extracted from the complete parametrisation,
which can be prolongated to the higher level by linear mapping. More pre-
cisely, let us consider, at level l, a linear application P (l) and its associated
matrix M (l) ∈ IRp×nl , i.e. such that for all α in IRnl P (l) (α) = M (l) α, the
following descent algorithm is then obtained:
γr+1 = γr − ωr M (l) (M (l) )T gr with ωr ∈ IR and gr ∈ IRp (1)
gr being the cost functional gradient at iteration r.
In particular, in [1] P (l) is defined by iteratively using an hermitian interpo-
lation of degree 3 on the nested parametrisation (see, also [3]).
Note that using shape grid-point coordinates as design variables is a nat-
ural choice since this parametrisation allows a direct correlation with the ex-
plicit representation of the shape in the discrete cost functional. Nevertheless,
in this case, non-smooth profiles can appear during the shape optimisation
process, and moreover, the large number of variables involved has also a nega-
tive effect on the convergence rate. As a matter of fact, the multilevel strategy
was defined exactly in order to reduce these drawbacks.
Alternatively, a polynomial representation of the shape is often used al-
lowing a more compact description with only few control parameters. In this
context, the Bézier curves, in which a shape is characterised by control points,
act as a basic tool.
3 Multilevel approach with Bézier parametrisation

A Bézier curve of degree n can be defined according to the following parametri-
sation:
Xn
S(t) = (x(t), y(t)) = Bnq (t)Sq with t ∈ [0, 1] (2)
q=0
in which {Sq = (xq , yq )}q=0,n and {Bnq (t)}q=0,n corresponds to the Bézier
control points and the Bernstein polynomials of degree n respectively.
Let us consider γ = (y0Γ , · · · , ym
Γ T
) , the ordinates of the shape grid-points,
as control variables and m + 1 parameters tk given (t0 = 0 < t1 < · · · < tm =
1). Then, it clearly appears that the ordinates of the Bézier control points,
Multilevel Bézier gradient method 539
T
i.e. α = (y0 , · · · , yn ) , are directly related with the control variables applying
(2) with y(tk ) = ykΓ for k = 0, · · · , m. More precisely, the following linear
operator from α to γ can be considered:
P : IRn+1 −→ IRm+1
(3)
α 7−→ γ = P (α) = M α in which Mij = Bnj (ti ).
Due to the linearity of P , it seems then possible to define a multilevel approach
in a similar way as done in [1]. Indeed, let us consider the following descent
direction at iteration r and level l:
j (l) (l)
d(l) (l) (l) T
r = M (M ) gr with Mij = Bnl (ti ). (4)
The only difference with respect to the algorithm defined in [1] is the choice of
the matrix M (l) , i.e. the choice of the specific subspace used in the multilevel
process. Nevertheless, it is not so obvious to describe the different levels since
the subparametrisation definition occurs in a less natural way. Indeed, the
abscissae of shape grid-points ({xΓk }k=0,m ) being given, at each level, X (l) =
(l) (l) (l) (l)
(x0 , · · · , xnl )T with nl > nl−1 and T (l) = (t0 , · · · , tm )T should be defined
such that (2) be verified, i.e.:
nl
(l) (l)
X
xΓk = x(tk ) = Bnq l (tk )x(l)
q for k = 0, · · · , m (5)
q=0
In order to define an adequate parametrisation at level l, the property

of degree elevation, which allows to increase the degree and the number of
control points of a Bézier curve, is used here. More precisely, given a Bézier
curve of degree n associated to the n + 1 control points Sq = (xq , yq )T , the
same geometrical curve can be also understood as a Bézier curve of degree
n + 1 considering a new set of n + 2 control points Sq0 = (x0q , yq0 )T . Sq0 can be
obtained from Sq as follows:

0 q q
Sq = Sq−1 + 1 − Sq for q = 0, · · · , n + 1 (6)
n+1 n+1
Let us now suppose that the parametrisation on the coarsest level has
been yet defined with X (0) and T (0) consistent with respect to (5). Then,
one can construct X (l) for l > 0 by applying successively (6) until to obtain
nl + 1 abscissae. In this way, since the distribution of the parameters t over
the Bézier curve does not change by degree elevation, the parameters tk keep
(l) (0)
unchanged on all the levels, i.e. tk = tk for k = 0, · · · , m + 1 and for all
l > 0. Furthermore, if, as for the case of the nozzle inverse problem presented
in Sec. 4, the (xΓk )k=0,m are uniformly distributed, then a consistent coarsest
level can be easily obtained by simply choosing an uniform distribution for
both X (0) and T (0) .
Note that, in [2], a multilevel strategy based on Bézier curves has been al-
ready defined for aerodynamic shape optimisation. In this approach, since the
control variables are the variations of the ordinates of the Bézier control points,
the change of level is directly obtained by degree elevation. On the contrary,

in the present formulation, even if each subparametrisation corresponds to the
ordinates of a set of Bézier control points, these control points do not explic-
itly appear in the definition of the descent direction since only the parameters
(tk )k=0,m are directly involved in (4). Thus, the knowledge of the particular
position of the Bézier control points is not required in the practical algo-
rithm implementation. Nevertheless, an explicitly dependence on these con-
trol points can be, also, reintroduced in the formulation. Indeed, it can remark
that the degree elevation formula (6) yields a linear prolongation between n
ordinates of Bézier control points to n + 1 ones through the multiplication
by the matrix Dn ∈ IR(n+1)×n defined by (Dn )ij = (1 − i)δij + iδ(i−1)j /n.
Then, PlL , the prolongation operator from level l to the finest level L of Bézier
control points, can be defined by:
(∀α ∈ IRnl ) PlL (α) = DnL · · · Dnl +2 Dnl +1 α = DL

l α
Let us also consider the operator P (L) , defined by eq. (3) with n = nL ,
which relates the finest level of Bézier control points and the shape grid-points
ordinates. Then, at a particular level l, one can directly relate the current set
of Bézier control points with the shape grid-points as done in (3) with n = nl ,
or alternatively, apply successively the operators PlL and P (L) . The two ways
should be equivalent from a theoretical view point, even if they can differ in
the implementation. Thus, the following algorithm can be used instead of eqs.
(1) and (4):
γr+1 = γr − ωr M (L) DL L T
l (Dl ) (M
(L) T
) gr (7)
In (7), even if the degree elevation explicitly appears through DL
l ,
the control
variables are still the ordinates of the shape grid-points. Nevertheless, due
to the linearity of the degree elevation procedure, it can be also envisaged
to directly use the Bézier control points as design variables in a multilevel
gradient-based approach. More precisely:
αr+1 = αr − ωr DL
l (Dl ) Gr
L T
(8)
in which Gr is the gradient with respect to α = (y0 , · · · , ynL )T at iteration r.
As soon as αr+1 is obtained, the shape grid-points can be also updated
by γr+1 = P (L) αr+1 . Then, starting from (8) and using the linearity of the
operator P (L) , the following expression is achieved for γr+1 :
γr+1 = γr − ωr M (L) DL
l (Dl ) Gr
L T
with Gr ∈ IRnL +1 (9)
But, since Gr can be related with the gradient with respect to γ = (y0Γ , · · · , ym
Γ T
)
by Gr = (M ) gr , the algorithm defined by (9) coincides with the one de-
(L) T
fined in (7).
Multilevel Bézier gradient method 541
4 Numerical experiments
A first validation of the present optimisation procedure has been done on a 2D
convergent-divergent nozzle inverse problem for inviscid subsonic flows (the
flow is modelled by the Euler equations). The corresponding cost functional is
expressed as a boundary integral which involves the flow variables through the
pressure field. This test-case is a classical one [4] already used for the multi-
level approach associated to grid-point coordinates parametrisation [1, 3]. An
initial constant-section nozzle and a target sine shape are considered while
the computational mesh is composed here of 1900 nodes with, in particular,
63 points on the upper boundary of the throat nozzle, i.e. where the shape
should be optimised.
Fig. 1 summarizes the convergence behaviour for different gradient pre-
conditioning. A reference run is obtained considering as control variables, the
ordinates of 31 shape grid-points among the 63 available. This one level strat-
egy yields a very slow convergence since the cost functional is diminished of
only about one order of magnitude. Indeed, for the same computational cost,
with a multilevel approach based on shape grid-points and a V-cycle strategy
on 4 levels (3, 7, 15 and 31 parameters), a plateau is attained after a decrease
of nine orders of magnitude. Concerning the preconditioning based on Bézier
control points as described in Sec. 3, both one level approach with 10 and 15
parameters respectively and a multilevel approach with a V-cycle strategy on
3 levels (5, 10 and 15 parameters) are considered. The convergence behaviour
of the three approaches is very similar for the decrease of the first five orders
of magnitude. For the one level approach, surprisingly the convergence curve
is better with less parameters but this behaviour inverts asymptotically. The
multilevel approach seems more advantagous only for very small values of the
cost functional with, in particular, an overlap of the plateau obtained with
the shape grid-points multilevel approach.
5 Conclusion
In the present study, a multilevel gradient-based method for optimum shape

design in aerodynamics is described. Considering the shape grid-points as
control parameters, the present approach appears as a multilevel method as
defined in [1], in which a particular prolongation operator is applied. Nev-
ertheless, it should be also interpretated as a generalisation of the original
multilevel method to Bézier control points as control variables. Thus, this for-
mulation appears also rather close to the one proposed in [2], even if, here, a
descent direction method is always considered.
The first validation, done on a simple 2D nozzle inverse problem, shows
that the new multilevel is still an interesting approach, even if, the improve-
ment due to multilevel strategy appears certainly less impressible with a set of
parametrisations based on Bézier control points. Two reasons can be adduced:
1
-1
-2 31 param.
shape grid-points
3-7-15-31 param.
log(cost functional)
-3
10 param.
Bezier control points 15 param.
-4
5-10-15 param.
-5
-6
-7
-8
-9
0 5000 10000 15000 20000 25000 30000
iterations
Fig. 1. Convergence histories.
on the one hand the multilevel is particularly well-adapted for shape grid-
points representation since, in this case, the presence of non-smooth shapes is
avoided, and on the other hand, the present test-case probably suffers from its
simplicity. Thus, the validation should be also carried out on other test-cases
in which the Bézier representation could furnish major contribution. Note
that, since the Bézier curves act as a basic tool for polynomial shape repre-
sentation, one can also envisage to extend the formulation to more complex
curves as B-splines, and also, to 3D case through, for instance, tensorial Bézier
parametrisation. Furthermore, since the present approach connects multilevel
gradient-based methods with Bézier hierarchical parametrisation originally
defined for gradient-free approaches [2], another possible development could
be to define a hybrid method, which combines favourable features of different
kinds of optimisation algorithm.
References
1. Beux, F., Dervieux, A.: A hierarchical approach for shape optimization. Engi-
neering Computations, 11, 25–48 (1994)
2. Désidéri, J.-A.: Hierarchical optimum-shape algorithms using embedded Bézier
parameterizations. Numerical Methods for Scientific Computing, Variational
Problems and Applications, Y. Kuznetsov (ed), CIMNE, Barcelona (2003)
3. Held, C., Dervieux, A.: One-shot airfoil optimisation without adjoint. Computers
and Fluids, 31/8, 1015–1049 (2002)
4. Beux F.: Shape optimization of an Euler flow in a nozzle. Notes on Numerical
Fluid Mechanics, 55, Périaux et al. eds, Vieweg Publishing, 115-131 (1994)
High-Fidelity Multi-criteria Aero-structural
Optimisation using Hierarchical Parallel
Evolutionary Algorithms
L. F. González1 , L. Damp1 , J. Périaux2 , and K. Srinivas1

1
The University of Sydney, Sydney, NSW 2006, Australia
gonzalez,lloyd.damp,ragh@aeromech.usyd.edu.au
2
CIMNE-UPC Barcelona jperiaux@gmail.com
Summary. This paper describes a parallel multi-criteria (multi-objective) evolu-

tionary algorithm for aero-structural problems. The foundations of the algorithm are
based upon traditional evolution strategies and incorporate the concepts of a multi-
objective optimisation, hierarchical topology, asynchronous evaluation and parallel
computing. The algorithm works as a black-box optimiser and has been coupled to
several aerodynamic and aircraft conceptual design solvers such as a higher order
panel method, a commercial FEA solver and an automated aerodynamic and struc-
tural mesh generation program. In this automated process the optimiser defines de-
sign variables from which the geometry, both internal and external, is generated and
then analysed by the aerodynamic solver. The pressure forces are mapped into the
structural model. Single or multiple objectives can be defined to find non-dominated
-Pareto optimal solutions or the Nash equilibrium. The parallel computing capabil-
ities in combination with hierarchical levels of aerodynamic and FEA solvers are
exploited to reduce computational expense. Applications and results on the use of
the methodology are illustrated on the aero-structural analysis of a high aspect ratio
aircraft/UAV wing.
1 Introduction and Motivation
The use of Multi-objective and Multi-disciplinary Design Optimisation (MDO)

methodologies is now being considered to solve many industrial problems [4].
The most common approach in optimisation is the use of gradient based
techniques but these have some limitations [1]. Researchers are now apply-
ing evolutionary methods for these type of problems as they have several
desirable characteristics: the ability to find global solutions, adaptability to
parallel computing, capability for difficult deceptive, multi-objective, multi-
modal problems[3].

544 L. F. González, L. Damp, J. Périaux, and K. Srinivas
2 Hierarchical Asynchronous Parallel Multi-objective

Evolutionary Algorithms (HAPMOEA)
A framework is developed using object oriented principles at the core of which
is a robust multi-criteria optimisation software tool - a Hierarchical Asyn-
chronous Parallel Multi-Objective Evolutionary Algorithm (HAPMOEA) [5].
The foundations of this algorithm lie upon traditional evolution strategies
(ESs) and incorporate the concepts of a Covariance Matrix Adaptation (CMA)
[2], multicriteria optimisation based on game strategies, a hierarchical topol-
ogy, parallel computing and asynchronous evaluation. Details regarding these
features can be found in Whitney et al. [5].
3 Medium Altitude Long Endurance (MALE) UAV

Design And Optimisation
The application of the framework is illustrated on a multidisciplinary, multi-
objective design optimisation of a MALE UAV Wing similar to the Altair
MALE Unmanned Aerial Vehicle (UAV). The two objectives are the maximi-
sation of the lift to drag ratio (L/D), and the minimisation of wing weight.
The cruise Mach number and altitude are 0.3663 and 12680 m respectively
and the wing area is set to 29.21 m2.
3.1 Design Variables
The wing geometry is represented by three aerofoil sections with sixteen vari-
ables for each section, eight variables that describe the wing planform and
eleven variables to describe the internal structure (Figure 1a and 1b). The
aerofoil geometry is represented by the combination of a mean line and thick-
ness distribution. Both lines are represented by Bézier curves with leading and
trailing edge points fixed at (0.0,0.0) and (1.0,0.0) respectively, and a variable
number of intermediate control points whose x-positions are fixed in advance
and whose y-heights form the problem unknowns. In this case 6 free control
points on the mean line and 10 free control points on the thickness distribu-
tion are taken. The wing planform design variables are indicated in Figures
1a and 1b. The optimisation is subject to various design constraints which are
imposed through several penalties details of which are omitted here.
3.2 Fitness Functions
The two fitness functions to be optimised are defined in Eqn 1 and Eqn 2:
f1 = min(1/(L/D) + penalty) (1)
f2 = min(W ingW eight + penalty) (2)
High-Fidelity Multi-criteria Aero-structural Optimisation using HAPEA 545
Fig. 1. a) Design variables: external geometry b) Design variables: internal geometry
The Evolutionary Optimiser seeks to minimise the inverse of the lift to

drag ratio. Eqn 2 relates to the structural aspects of the candidate wing. This
function has the units of mass and seeks to minimise the calculated structural
mass of the wing.
3.3 Aero-structural Analysis
An aero structural solver was developed. This solver integrates the two analy-
sis tools for FEA and CFD namely- MSC. NASTRAN and the high order panel
method- PANAIR. The entire aero-structural program is controlled through a
MatlabÂvZ script file. This allows for an easy coupling of the different required
programs as one continuous MatlabÂvZ data structure could be utilized to
define all the information passed between programs.
3.4 Implementation
The present optimisation has been carried out using two approaches; the first
one uses a traditional evolutionary method with a single population model
based upon a fine mesh aerodynamic model of each candidate wing. This
model uses between 1904 and 2992 aerodynamic panels. The second method
uses a hierarchical topology of mesh resolutions comprising three levels. The
same aerodynamic resolution as was used in the single population experiments
was used for the top level; a coarser mesh model used in an intermediate
aerodynamic level (between 1428 and 2244 aerodynamic panels), and an even
coarser mesh (between 952 and 1496 aerodynamic panels) used in the lowest
model to explore the design search space. Three P4, 2.4 GHz machines were
used in the calculation and the population size for this problem was set to 30
in all levels.
3.5 Results
The algorithm was run for 500 function evaluations and took 36 hours to com-
pute. Figure 2a shows the Pareto fronts obtained by using the two approaches.
The Pareto members selected for further investigation are also indicated. By
inspecting Figure 2a and b) it can be seen that the use of a multi-fidelity (hier-
archical) approach gives an overall lower front as compared to a single model
approach. Both approaches give results which lie to the left of the benchmark
value for the MALE wing. The algorithm progress for the hierarchical ap-
proach is shown in Figure 2b. The fitness for all levels decreases with time
and it can be seen that after 350 evaluations, the fitness of each level reaches
a steady state value indicating an optimum solution.
Fig. 2. a) Pareto Fronts comparison b) Evolution progress
Three wings from the hierarchical optimisation strategy results are taken
for further evaluation (Pareto members 3, 8 and 23) in order to illustrate the
two objective extremes and a compromise between the two results. Pareto
3 has the best aerodynamics, Pareto 23 the best structural response to the
aerodynamic loading and Pareto 8 lies between the two solutions.
Figure 3a shows a top view of the wings selected for further analysis and
Figure 3b compares the baseline aerofoils and the aerofoil sections of each
Pareto member at the root, crank and tip.
A Mach number sweep was performed at a constant angle of attack for
the selected Pareto members and the benchmark. Over the range of Mach
numbers investigated, the Pareto members have lower drag coefficient than
the benchmark wing (results not shown here). The Pareto members three
and eight operate at an almost constant drag value over the range of Mach
numbers. When the coefficient of lift is plotted versus the coefficient of drag
for the selected Pareto members and the benchmark MALE wing in Figure
4a, one can see that the performance of all the wings belonging to the Pareto
front is higher than that of the benchmark wing. The required coefficient of
lift of 0.64 is met at a lower coefficient of drag with all the selected Pareto
High-Fidelity Multi-criteria Aero-structural Optimisation using HAPEA 547
Fig. 3. a) Top views of selected Pareto members b) Selected MALE Pareto member
Root aerofoil sections
Fig. 4. a) Coefficient of Lift vs. Coefficient of Drag for selected MALE Pareto
members and Benchmark. b) MALE Pareto 3 Displacement
member wings. Figure 4b shows one example of the deflected Pareto wings.
The original wing position and internal structure is shown as a wire frame
for comparison. All the MALE wings deflect nonlinearly. A nonlinear analysis
was necessary as the deflections encountered are too large for a Linear Static
analysis and if used, would have yielded incorrect results.
A summary of each of the Pareto and benchmark wing overall character-
istics is shown in Table 3 for comparison. It can be noted that all Pareto
member wings out-performed the benchmark wing on L/D ratio with Pareto
members 8 and 23 had weights which were lower than that of the benchmark
wing while keeping the deflection below 20 percent.
4 Conclusions
This paper described the features of a hierarchical, asynchronous parallel

multi-objective evolutionary algorithm that can be used to solve aero struc-
tural design problems. The method is easily coupled, particularly adaptable to
different solvers, easily parallelised, require no gradient of the objective func-

tion(s), have been used for multi-objective optimisation and applied to solve
different aeronautical design problems. Extension of the method to include
structural model and its coupling with the flow solver has been investigated
with encouraging results.
Table 1. Summary of Displacement, Strain and Mass for Selected MALE Pareto
members and the Benchmark
Description Pareto 3 Pareto 8 Pareto 23 Benchmark
Deflection % 16 17.3 13.8 12.5
Maximum Strain 0.000396 0.00052 0.000485 0.000403
Mass (kg) 233.735 217.873 185.1 220.9
L/D 43.367 44.4 44.25 23.9
Acknowledgements
The authors acknowledge M.Sefrioui, Dassault Aviation and E. Whitney at

The University of Sydney for fruitful discussions on Hierarchical Asynchronous
EAs., S. Armfield of the University of Sydney for allowing access to the cluster
of computers. The project was supported by the Asian Office of Aerospace
Research & Development (AOARD) of the US AirforceÂ through the contract
AOARD-044078.
References
1. K. Deb. Multi-objective genetic algorithms: Problem difficulties and construction
of test problems. Evolutionary Computation, 7(3):205–230, 1999.
2. N. Hansen and A. Ostermeier. Completely derandomised self-adaption in evo-
lution strategies. In Evolutionary Computation, volume 9, pages 159–195. MIT
Press, 2001.
3. M. Sefrioui and J. Périaux. A hierarchical genetic algorithm using multiple mod-
els for optimization. In M. Schoenauer, K. Deb, G. Rudolph, X. Yao, E. Lutton,
J.J. Merelo, and H-P. Schwefel, editors, Parallel Problem Solving from Nature,
PPSN VI, pages 879–888. Springer, 2000.
4. J. Sobieske and R.T. Haftka. Multidisciplinary aerospace design optimization:
survey of recent developments. In Structural Optimization, v14, pp. 1-23. 1997.
5. E. J. Whitney, M. Sefrioui, K. Srinivas, and J. PuAvSriaux. Advances in hier-
archical, parallel evolutionary algorithms for aerodynamic shape optimisation.
JSME (Japan Society of Mechanical Engineers) International Journal, 45(1),
2002.
Flow Control Optimization Using Neural
Networks and Genetic Algorithms
Raymond P. LeBeau, Jr., Narendra K. Beliganur1 and Thomas Hauser2

1
Mechanical Engineering, University of Kentucky, Lexington, KY, USA
rplebeau@engr.uky.edu
2
Mechanical & Aerospace Engineering, Utah State University, Logan, UT, USA
thomas.hauser@usu.edu
Evolutionary algorithms have now been used as tool to optimize complex de-
sign spaces in aerospace applications, notably in the areas of Multidisciplinary
Design Optimization (MDO) [4, 2] and flow control [9]. However, in the latter
area a limiting factor has been the cost of evaluating the performance of each
tested flow control configuration. This process is conventionally accomplished
through computational fluid dynamics (CFD) simulation of the design, but
in cases that require a full, viscous Navier-Stokes solver, the computational
demands still can be high even for simple flow control applications like steady
suction or blowing. More challenging flow control applications typically re-
quire even more expensive simulations.
A possible means for accelerating the evaluation process within a search
algorithm is to replace CFD computations with an artificial neural network
(NN). The NN initially requires training and testing for the given problem,
the data for which is provided by CFD simulations. Since neural networks do
not actually solve the Navier-Stokes equations, the proposed optimum per-
formance regions determined by the GA-NN system need to be confirmed.
However, as proposed, the NN approach can replace nearly all of the CFD
computations for a majority of the generations of the GA, dramatically re-
ducing the computational cost. The NN can also be used to test GA methods
with a realistic, but far less costly fitness evaluation, leading to improved GA
design for both NN and CFD evaluations.
1 Numerical Methodology
1.1 Geometry
The flow under consideration is that of a NACA 0012 at an 18 degree angle-

of-attack and a Reynolds number of 500,000. Details on general grid indepen-
dence and numerical accuracy for this case without flow control jets have been

550 Raymond P. LeBeau, Jr., Narendra K. Beliganur and Thomas Hauser
presented in previous work [8] and are generally satisfactory for the given grid.
For the jet array GA-CFD computations, the total number of grid points is
between 140,000 and 150,000. The computations are done on the structured
two-dimensional code GHOST. GHOST is an in-house, finite volume CFD
code developed at University of Kentucky by P. G. Huang. This code has
been tested extensively and is routinely used for turbulence model validation
and 2D flow control analysis [9, 8]. The baseline case under consideration is
both fully turbulent and at a large angle of attack producing strong separa-
tion if flow control is not applied. As such, these simulations employ Menter’s
two-equation SST turbulence model.
1.2 Genetic Algorithm - CFD approach
Two real-coded genetic algorithm approaches have been applied in this project
as an objective is to determine the GA with a minimum of fitness evaluations
for cases when relatively expensive CFD computations are required to eval-
uate the fitness functions. The first GA is a Continuous Genetic Algorithm
[5] with elitism. The second GA is the EARND GA with diversity control,
which is detailed in [6, 8]. For the EARND GA, additional modifications have
been made to a standard GA to accelerate convergence: a normal distribu-
tion scheme [11] and dynamic updating of the parameter bounds to eliminate
unnecessary evaluations (computations) in un-promising areas. Finally, diver-
sity control isa dded to the selection function during the initial part (typically
20%) of the evolution to reduce local early convergence.
1.3 Genetic algorithm - Neural network approach
Fitness evaluation remains the primary cost of both these optimization ap-
proaches, as each evaluation of an individual requires a very costly CFD sim-
ulation. While the computations considered in this project are on the order
of CPU hours, not days, this is still quite expensive when thousands of simu-
lations are required. An alternative to the CFD computation of every fitness
is to use interpolation schemes; however, these must be applied with care as
a GA approach fundamentally assumes a multi-dimensional solution space
with multiple local maximum and minimum that need to be maintained, not
blurred out by poor interpolation. One option is use a non-linear approach
such as a NN. Generally, a NN consists of layers of interconnected nodes, each
node producing a non-linear function of its input [10]. Our neural network ap-
proach is based on the Pyro [1] library implemented in Python. Currently, we
are using one input layer with the optimization variables like suction strength,
suction angle, etc. Then a hidden layer connects this input layer to an output
layer of two nodes representing lift and drag. There exist no general rules on
the structure of the network and we train several different networks with an
accuaracy of three percent. Next, we select the best network with the least
amount of nodes in the hidden layer for the NN-GA approach.
Flow Control Optimization 551
1.4 Jet optimization
Three parameters for each jet are selected for the search investigation, namely
jet location Lj (measured in percent chord), suction/blowing amplitude A
(measured in term of orifice velocity relative to the inflow freestream velocity),
and suction/blowing angle θ (Fig. 1). The jet width for both suction and
blowing is fixed at 2.5% chord length based on a study by Dannenberg [3].
Note that a negative θ corresponds to suction while positive indicates blowing.
Thus, perpendicular suction is −90o and perpendicular blowing is 90o . The
previous single jet studies [7] have demonstrated that the potential flow control
due to suction tends to dominate that of blowing in this configuration if they
are of comparable magnitudes. To prevent the suction jet from dominating
the genetic search (and thereby significantly reducing the complexity of the
parameter space), the total suction amplitude is limited to be no more than
0.03, or 3% of the inlet velocity for the two jet case, while the net blowing
velocity can reach 0.2. These values correspond to jet momentum coefficients
(Cµ ) of 2.25x10−5 and 0.001, respectively, if applied to a single jet.
Re = 500,000 Control Parameters

0
α = 18 Lj A 1. Jet Location (Lj)
θ = 900 2. Amplitude (A)
s0 3. Angle (θ)
0
θ = -90
0
18
Fig. 1. Jet parameters and flow conditions for the multiple jet CFD simulations
The results of the three two-jet evolutions were collectively reviewed in

reference [9] and are shown together with the four jet results in section 3.
The genetic parameters were selected as NGeneration = 100, NPopSize =
32, NVariable = 5, and NUpdate = 8, with crossover and mutation probabili-
ties of Pc = 0.2 and Pm = 0.1, respectively. The fitness function for evaluation
of each genome was
(F itA )max = a · Cl /ClB + b · CdB /Cd , (1)
with a = 1.0 and b = 1.0, providing equal weight to lift and drag. The baseline
values ClB and CdB were determined from simulations of the base airfoil
without jets.
A similar approach has been taken for a four jet GA-CFD evolution with
two suction and two blowing jets. This evolution more the doubles the number
of parameters considered, with the angle and location for each jet allowed to
vary in the same ranges as the two jet case. The amplitude of one suction jet
is fixed as before, with the others allowed to vary. The GA used in the four
jet case is the Continuous GA with 48 individuals per generation.
2 Results
While the two jet GA-CFD results are fairly definitive, the four jet results
should be considered tenative as the Continuous GA currently has too low
a rate of mutation, resulting in too many similar values in some parameters,
and only 15 generations have been included, short of a full evolution. Still, the
results suggest that the four jet case actually is favoring placing both suction
jets and one blowing jet near the leading edge in a series, while the remaining
blowing is again located downstream.
The results for the best individual for the two and four jet cases are given
in Table 1. The optimal drag seems to be predicted very well by the NN,
whereas there are differences in the lift prediction.
2 jet GA-CFD 2 jet GA-NN 4 jet GA-CFD 4 jet GA-NN

Suction location 1 0.137 0.781 0.0500 0.050
Suction angle 1 -89.31 0.0 -89.92 -12.25
Blowing location 1 0.742 0.59 0.643 0.831
Blowing angle 1 45.47 0.0 19.03 0.0
Blowing magnitude 1 0.196 0.0 0.00896 0.014
Suction location 2 - - 0.116 0.815
Suction angle 2 - - -89.98 -36.92
Suction magnitude 2 - - 0.00207 0.00152
Blowing location 2 - - 0.659 0.784
Blowing angle 2 - - 34.15 90.0
Blowing magnitude 2 - - 0.00948 0.014
Lift 0.914 0.923 0.988 0.965
Drag 0.146 0.146 0.149 0.148
Table 1. Comparison of best individual for the two and four jet cases
To illustrate the behavior of the GA-CFD and GA-NN system better, the
fitness for the best 2,500 individuals is compared in Fig. 2.
3 Conclusion and outlook

The GA-NN system accelerates the optimization process from weeks to min-
utes but additional computational effort goes into generating the training data
sets through CFD and training the NN. Since neural networks do not actually
solve the Navier-Stokes equations and generate less accurate results, the pro-
posed optimum performance regions determined by the GA-NN system need
Flow Control Optimization 553
(a) Two jet case (b) Four jet case
Fig. 2. Comparison of the fitness function between the CFD-GA approach and the
NN-GA approach for the two and four jet cases
Fig. 3. Comparison of the lift prediction between the CFD-GA approach and the
Fig. 4. Comparison of the drag prediction between the CFD-GA approach and the
to be confirmed. The NN can also be used to test GA design with a realistic,

but far less costly fitness evaluation, leading to improved GA design for both
NN and CFD evaluations. Further work is necessary to develop guidelines for
robust NN architectures, generation of training data sets and developing an
integrated, automated NN-GA-CFD system.
References
1. D.S. Blank, D. Kumar, L. Meeden, and H. Yanco. Pyro: A python-based ver-
satile programming environment for teaching robotics. Journal of Educational
Resources in Computing, 2004.
2. S. Chen, F. Zhang, and M. Khalid. Aerodynamic optimization for a high-lift
airfoil/wing configuration. In 22nd AIAA Applied Aerodynamics Conference,
Providence, RI, number 2004-5078, August 2004.
3. R.E. Danneberg and J.A. Weiberg. Section characteristics of a 10.5-percent
thick aifoil with area suction as affected by chordwise distribution of perme-
ability. NASA Technical Note 2847, Ames Aeronautical Laboratory, Moffet
Field, CA, 1952.
4. L.F. Gonzalez, E.J. Whitney, K. Srinivas, and J. Periaux. Optimum multidis-
ciplinary and multi-objective wing design in cfd using evolutionary techniques.
In International Conference on Computational Fluid Dynamics 3, Toronto,
Canada, July 2004.
5. R.L. Haupt and S.E. Haupt. Practical Genetic Algorithms. John Wiley and
Sons, 2004.
6. L. Huang, , R.P. LeBeau, and T. Hauser. Application of genentic algorithm
to two-jet control system on naca 0012 airfoil. In International Conference on
Computational Fluid Dyanmics 3, Toronto, Canada, July 2004.
7. L. Huang, P.G. Huang, R.P. LeBeau, and T. Hauser. Numerical study of blow-
ing and suction control mechanism on naca0012 airfoil. AIAA Journal of Air-
craft, 41:1005–1013, 2004.
8. L. Huang, P.G. Huang, R.P. LeBeau, and T. Hauser. Optimization of blowing
and suction control on naca 0012 airfoil using a genentic algorithm. In 42nd
AIAA Aerospace Sciences Meeting and Exhibit, Reno, NV, number 2004-0423,
January 2004.
9. R.P. LeBeau, L. Huang, and Th. Hauser. Application of evolutionary algo-
rithms to small jet arrays for flow control. In 17th AIAA Computational Fluid
Dynamics Conference, number 2005-4859, June 2005.
10. D. Michie, D.J. Speigelhalter, and C.C. Taylor. Machine learning, neural and
statistical classification, February 17 1994.
11. A. Oyama, S Obayashi, and K. Nakahashi. Real-coded adaptive range genetic
algorithm and its application to aerodynamic design. JSME International Jour-
nal Series A, 43:124–129, 2000.
Shape Optimization for Dense Gas Flows in
Turbine Cascades
Pietro Marco Congedo1 , Paola Cinnella1 , and Christophe Corre2

1
Università di Lecce, via Monteroni, 73100, Lecce (ITALY)
pietro.congedo@unile.it, paola.cinnella@unile.it
2
Ecole Nationale Supérieure d’Arts et Métiers
christophe.corre@paris.ensam.fr
1 Introduction
In recent years, great attention has been paid to a class of fluids of the retro-
grade type (i.e. fluids that superheat when expanded), known as the Bethe–
Zel’dovich–Thompson (BZT) fluids, which exhibit in the vapor phase, above
the upper saturation curve, a region of negative
values of the Fundamental
ρ ∂a
Derivative of Gasdynamics Γ := 1 + a ∂ρ , with ρ the fluid density, a
s
the sound speed, and s the entropy. In the transonic and supersonic regimes,
this leads to nonclassical gasdynamic behaviors, such as expansion shocks and
mixed waves. Moreover, flow discontinuities with jump conditions in the vicin-
ity of the Γ = 0 contour have necessarily limited strength, producing losses
(entropy rise) one order of magnitude lower than usual [2]. The interested
reader may refer to [3] for a review of the complex dynamics of BZT fluids.
An appealing application of BZT fluids is efficiency enhancement for Or-
ganic Rankine Cycles (ORCs). ORCs’ working fluids are heavy organic com-
pounds with large heat capacities: interestingly, several of these fluids possess
BZT properties. One major source of losses in ORC turbines is wave drag,
since they usually operate in the transonic/supersonic regime: the use of a
BZT fluid could avoid shock formation and, ideally, allow isentropic turbine
expansion. Unfortunately, simply utilizing a BZT working fluid is not suffi-
cient to maximize the reduction in losses. Operating the turbine cascade at a
pressure and temperature near the thermodynamic region where BZT effects
appear is also necessary. On the other hand, this region, called the inversion
zone, has a quite limited extent. As a consequence, a reduction in the cascade
pressure ratio is required to operate the turbine entirely within this zone. Now,
it is known from thermodynamic theory that a too small pressure ratio leads
to poor global thermal cycle efficiency. Thus, the development of BZT ORCs
needs finding a reasonable tradeoff between two opposite requirements: on the
one hand, turbine expansion must happen as close as possible to the inversion

556 Pietro Marco Congedo, Paola Cinnella, and Christophe Corre
zone, on the other, turbine pressure ratio must be sufficiently high for achiev-
ing high global cycle efficiency and power output. Previous studies about
dense gas flows in turbine cascades [4] have shown that, for moderately high
turbine pressure ratios, the use of a BZT working fluid may allow efficiency
improvements up to 3% over air for properly chosen operating conditions. For
higher pressure ratios, efficiency improvements with respect to a perfect gas
progressively disappear, whereas BZT fluids remain advantageous compared
with other working fluids typically employed in Rankine cycles, such as steam
or toluene. Increasing the inlet pressure and Fundamental Derivative along
an isentrope crossing the inversion zone, efficiency tends initially to increase
and reach an optimum. At optimum conditions, the average value of Γ is less
than 1 at the wall, and the sound speed increases with dropping pressure.
This limits Mach number growth during flow expansion and reduces shock
strength. For higher inlet values of Γ , Γinl , efficiency drops again. The higher
the cascade pressure ratio, the higher the Γinl value corresponding to peak
efficiency, and the quicker efficiency drops when moving away from optimum
conditions. Note however, that efficiency improvements observed in previous
studies were simply due to the special nature of the working fluids, since the
blade shapes considered were typical gas turbine blade sections, not specifi-
cally adapted for dense gas flows. The objective of the present study is finding
optimal blade shapes for BZT ORC turbines, providing high efficiency over a
large range of operating conditions and working with high cascade pressure
ratios.
2 Governing Equations, Flow Solver and Optimizer
In the present study, dense gas flows are modeled by the compressible Euler
equations for single-phase, non-reacting flows, completed by Martin-Hou real-
istic equation of state [5]. The governing equations are discretized by a finite
volume method for structured multiblock grids, using a third-order accurate
centered spatial approximation [6]. The solution is advanced in time using a
four-stage Runge-Kutta scheme with local time-stepping and implicit residual
smoothing. The solver has been previously validated for a variety of perfect
and dense gas flow simulations.
The flow solver is coupled with a multi-objective genetic algorithm (MOGA).
Genetic algorithms (GA) have proved their interest with respect to gradient-
based methods because of their high flexibility and their ability to find global
optima of multi-modal problem. The MOGA applied in this study is the Non-
Dominated Sorting Algorithm proposed by Srinivas and Deb [7]. For single
objective problems, the algorithm uses an elitism strategy to ensure that the
best individuals survive when algorithm evolves. For multi-objective problems,
a Pareto-based genetic algorithm is applied. The MOGA has been previously
validated for the optimization of airfoil shapes in dense gas flows [8]. In order
to select a proper starting population for the genetic algorithm, the repre-
Shape Optimization for Dense Gas Flows in Turbine Cascades 557
sentation of the design space is previously investigated through a Design of

Experiment (DOE) procedure using a Sobol sequence technique.
3 Results
The MOGA optimizer, coupled with the flow solver and a structured mesh
generator is applied to shape optimization of turbine cascades. Blade shape
is parametrized using Bezier polynomials, starting from the baseline profile
VKI LS-59 [9]. Sixteen control points are imposed to parametrize the profile.
Globally the optimization problem depends on twenty-five variables. Geomet-
ric constraints are imposed on the maximum blade thickness and on trailing
edge thickness (normalized with the blade chord) that may vary within ±10%
of the corresponding non-dimensional thickness for the baseline geometry. The
turbine cascade is operated with an inlet flow angle of 30◦ and a pressure ra-
tio of 1.82 corresponding, for a perfect diatomic gas, to sonic isentropic exit
conditions. For dense gas flows, inlet thermodynamic conditions, i.e. the ther-
modynamic operation point, should be also specified. All the computations
presented in the following are performed on C-grids composed by 192 × 16
cells, with average non-dimensional height of the closest cells to the wall ap-
proximately equal to 5 × 10−3 . This mesh refinement provides a reasonable
tradeoff between accuracy and computational cost. Since the objective func-
tion chosen for optimization runs is cascade efficiency (defined as real-to-ideal
static enthalpy drop), preliminary computations have been performed, for the
baseline configuration, to investigate the sensitivity of this parameter to mesh
refinement. Results computed on finer grids composed by 382 × 32 (first cell
height 10−3 ) at several flow conditions show that the magnitude of cascade
efficiency tend to increase of about 2∇ · 3% when refining the grid, but trends
of behavior (e.g. efficiency dependency on inlet conditions) are well conserved.
Firstly, a single-objective shape optimization for a diatomic perfect gas
(PFG) has been performed in order to maximize turbine efficiency. The ini-
tial population of the genetic algorithm is selected after a preliminary DOE
over 100 individuals. After about 20 generations the mean and maximum
value of the objective function in the population reach approximately the
same value, indicating that the GA has converged to a population of almost
identical (optimal) individuals. The solution for the baseline blade is char-
acterized by an oblique shock at about 70% the chord and a second shock
attached to the trailing edge (Fig. 2a); computed efficiency is 93.2%. Figure
2b shows the solution after shape optimization: the oblique shock is almost
suppressed, the trailing edge shock is weaker with respect to the reference
case and efficiency grows to 96.2%. Figures 2c,d display Mach number and
pressure (normalized by inlet stagnation conditions) distributions at the wall.
For the optimal configuration the maximum Mach number at the wall and,
consequently, shock strength, is noticeably reduced compared to the baseline
configuration; consequently, wave drag lowers and cascade efficiency improves.
0.4 0.4
0.2 0.2
0 0
y
-0.2 -0.2
-0.4 -0.4 Baseline

Baseline Pareto 1
Best PFG Best OPT1
-0.6 -0.6 Best OPT2
0 0.2 0.4
x
0.6 0.8 1
(a) 0 0.2 0.4
x
0.6 0.8 1
(b)
Fig. 1. Optimal shapes for perfect (a) and dense (b) gas flow.
The optimal individual (Fig. 1a) has a thicker leading edge, thinner trailing
edge, and greater camber than the baseline configuration.
Then, computations are performed with the BZT fluorocarbon PP10 as
the working fluid. Three optimization runs are undertaken. In the first and
second run, a single objective function is maximized, i.e. cascade efficiency
at fixed operating conditions pinl /pc = 1.00, ρinl /ρc = 0.752, Γinl = 0.416
(optimization point OPT1) and pinl /pc = 1.10, ρinl /ρc = 1.09, Γinl = 1.91
(optimization point OPT2), respectively. Subscript c indicates critical point
values. Both operation points lay on an isentrope crossing the inversion zone,
respectively at lower and higher pressure than peak efficiency conditions. In
the third run, both objectives (i.e. efficiencies at OPT1 and at OPT2) are
simultaneously maximized. Here again, a DOE is preliminarily run in order
0.5
0.5
1
1
1.1 1.1
0.5
0.5 0.5
0.5 1.2
1
1.1 1.1
0.5
0.5 1.2
(a) (b)
1
1.4
Optimal
1.2 Baseline 0.8 Pressure
Adimensional Pressure
1
Mach number
0.6
0.8
Suction
Suction
0.6 0.4
0.4
0.2
0.2
Pressure Optimal
Baseline
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c (c) x/c (d)
Fig. 2. PFG flow, isoMach lines: a) baseline; b) optimized. c,d) Wall distributions.
Shape Optimization for Dense Gas Flows in Turbine Cascades 559
0.8 0.8
1.3
0.8
0.75 1.3 0.8
1
1
1.2
1.3
0.8
0.8
1.3 1.2
(a) (b)
1.5 2.5
Baseline Baseline
Pareto 1 Pareto 1
2
Best OPT1 Best OPT1
Fundamental Derivative
Best OPT2 Best OPT2
1 1.5
Mach number
Pressure
Suction
1
0.5 0.5
Suction
Pressure 0
0 -0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/c (c) x/c (d)
Fig. 3. DG flow, isoMach lines: a) baseline; b) optimized. c,d) Wall distributions.
to properly initialize the population of the GA. Figure 1b shows optimal indi-
viduals for operating conditions OPT1 and OPT2 and an individual selected
on the Pareto front resulting from the two-point optimization. The baseline
profile is also represented. For each configuration, a parametric study of cas-
cade efficiency at off-design conditions has been performed. In addition to
operation points OPT1 and OPT2, two test points corresponding to operat-
ing conditions pinl /pc = 1.02, ρinl /ρc = 0.813, Γinl = 0.886 (TEST1) and
pinl /pc = 1.05, ρinl /ρc = 0.944, Γinl = 1.62 (TEST2) are considered. Results
are summarized in Table 1. Note that optimal individuals provide an improve-
ment of 1∇ · 2% only over the baseline at optimization conditions. However,
improvements up to 7% are obtained at conditions TEST1 and TEST2, which
lay close to peak efficiency conditions for the baseline cascade. The best overall
results are obtained for the individual derived from the two-point optimization
(”Pareto 1”). Note that its shape is quite similar to that of the optimal indi-
vidual for perfect gas flow. Figures 3a,b show iso-Mach lines for the baseline
configuration and for individual ”Pareto 1” at operating condition TEST2.
Figures 3c,d show the Mach number and Γ distributions at the wall for the
best individuals from single-objective optimizations, the ”Pareto 1” individ-
ual and the baseline blade. Here again the mechanism leading to efficiency
improvement is a reduction in the maximum Mach number of the flow and
hence of shock strength: major gains come from a significant weakening of the
trailing edge shock for the optimized configuration. This is related to lower
values of Γ at the rear part of the suction side, causing slower growth of the
Mach number when the flow re-expands downstream of the first shock.
Table 1. Efficiencies (%) for optimal blade shapes at several operating conditions.
OPT1 TEST1 TEST2 OPT2

Γinl 0.416 0.886 1.62 1.91
Best OPT1 91.25 95.74 97.20 89.71
Best OPT2 89.67 95.45 96.08 90.22
Pareto 1 91.00 96.36 97.47 89.75
VKI LS-59 88.66 90.72 90.80 88.77
4 Final remarks
Shape optimization for flows of perfect and dense gases in turbine cascades
has been achieved by means of a multi-objective genetic algorithm. For perfect
gas flows, shape optimization allows efficiency improvements of about 3% over
the baseline configuration. For dense gas flows, proper optimization starting
from the same baseline geometry leads to efficiency gains up to 7%. Multipoint
optimization allows improving performance over a large range of thermody-
namic operating conditions. For the high cascade pressure ratio considered in
this study, BZT effects play a minor role in efficiency improvement. The use
of properly designed turbine cascades working with somewhat lower pressure
ratios could allow higher efficiency improvements due to BZT effects, opening
the door to the development of BZT turbines for Organic Rankine Cycles.
References
1. P.A. Thompson: Phys. Fluids 14, 1843 (1971)
2. M.S. Cramer, A. Kluwick: J. Fluid Mech. 142, 9 (1984)
3. M.S. Cramer: Nonclassical dynamics of classical gases. In: Nonclassical waves in
real fluids, ed by A. Kluwick (Springer, Berlin, 1991) pp 91–145
4. P. Cinnella, P.M. Congedo, D. Laforgia: Transonic flows of BZT fluids through
turbine cascades. In: Computational Fluid Dynamics 2004, ed by C. Groth and
D.W. Zingg (Springer, Berlin, 2006). In press.
5. J.J. Martin, Y.C. Hou: A.I.Ch.E. J., 1, 142 (1955)
6. P. Cinnella, P.M. Congedo: AIAA J. 43, 2458 (2005)
7. N. Srinivas, K. Deb: Evol. Comput. 2, 221 (1995)
8. P.M. Congedo, C. Corre, P. Cinnella: Airfoil shape optimization for transonic
flows of Bethe-Zel’dovich-Thompson fluids. AIAA J. (2006), accepted.
9. R. Kiock et al.: ASME J. Eng. Gas Turb. Power 108 : 810 (1996)
Numerical Simulation of Control of Flow Past
a Stalled NACA0015 Airfoil with Plasma-based
Body Forces
Meiliang Mao, Xiaogang Deng, and Jianqiang Chen
China Aerodynamics Research and Development Center, Mianyang, Sichuan,

China, 621000
1 Introduction
Plasma-based techniques exploiting electromagnetic forces for flow control are

currently of considerable interest. Particularly attractive properties stem from
the absence of moving parts and lack of mechanical inertia, which permits near
instantaneous deployment over a broad range of frequencies. Recent studies
have demonstrated the capability of the dielectric barrier discharge (DBD) to
promote boundary layer attachment on airfoil at a high angle of attack[1, 2, 8]
at atmospheric pressure.
Although the kinetics theories have been developed to analyses the phe-
nomenology to the characteristic of this plasma discharge, numerical model-
ing of high-Reynolds number flows with these kinetics theories is a daunt-
ing endeavor[3, 4] because of the unknown fundamental processes and ex-
pensive computation. The difficulties have fostered the development of a
wide spectrum of models with varying degrees of empirical elements into
the procedure[5, 6, 7]. Shyy etc.[6] have developed separate estimates for the
charge distribution and the electric field by coupling known plasma physics
parameters to qualitative experimental data and proposed a model suitable
for coupling with the fluid response. Suzen etc.[7] have described several 2-D
simulation with a consistent approach satisfying Gauss’s law.
Utilizing the NACA0015 airfoil with actuators of an asymmetric-staggered
configuration on its upper surface, Roth[8] has completed the experimental ob-
servation of the flows around the airfoil with influences of the above plasma
discharge and found that the flow attachment effects are best facilitated by
actuators near the leading edge of the airfoil, actuators toward the trailing
edge of the airfoil have relatively little additional effect. One actuator on a
given point near the leading edge of a NACA0015 airfoil has been numeri-
cally investigated in [9]. By numerical simulation, the present paper tries to
reproduce the phenomena in [8] and to investigate the relation between the
aerodynamics of the airfoil and the flow characteristics simultaneously.

562 Meiliang Mao, etc..
2 Numerical method
The flow around the airfoil is assumed to be two-dimensional.The fluid dynam-
ics is described by the full RANS equations with the k-g turbulence model,
augmented by terms representing the local forcing of the DBD device. They
may be written in non-dimensional form:
∂Q ∂FI ∂GI ∂FV ∂GV
+ + = + +S (1)
∂t ∂x ∂y ∂x ∂y
where Q = {ρ ρu ρv ρe}T , S = {0 FT avex FT avey 0}T is plasma-
based body force, and is obtained from the relation[6]:
∗
FT∗ ave = cα ϑ∗ (ρc ec ) ∆∗T e E∗ δ (2)
where ϑ∗ = 4.2 kHz is the frequency of the applied voltage, ∆∗T e = 67 µs is
an effective duty cycle to recover an effective mean force, cα = 1 is a factor
for collision efficiency, ρc = 1011 cm−3 is the charge density and E∗ is the
electric field. δ is 1 when the magnitude of the electric field is at or above
breakdown level (|E∗ b | = 3 kV mm−1 for air, the plasma exists in this region)
and zero otherwise. Here superscript * denotes dimensional quantity.
Neglecting the influence of charges on the electric potential ϕ∗ , the electric
field is obtained by the Gauss’ law for the electric potential,
∆ϕ∗ = 0 , E∗ = −∇ϕ∗ (3)
The non-dimensionalization of the FT∗ ave is accomplished through the fol-
lowing relation:
FT∗ ave Lref
FT ave = 2 (4)
ρref Uref
where Lref , ρref and Uref are dimensional reference quantities, and set to
chord length, density and velocity of incoming flow respectively.
To simplify the numerical procedure, the influence of the flow on the
plasma is neglected assuming that the low speed flow changes the thermody-
namic state quantities of the fluid trivially and the fluid moves little distance
during a cycle. Equation (3) is solved by implicit central scheme with Gauss-
Sedial technique. The convergence criteria is

∗(n)
i , j , k
X ϕ
max ϕ∗(n+1) − ϕ∗(n) ≤ 10−8 ϕ∗(n) , ϕ∗(n) = (5)

av av Ncell
i,j ,k
where Ncell is the number of total cells.

Once the distribution of the potential is determined, the distributions of
the electric field and the force FT∗ ave can be obtained accordingly.
The numerical method employed to solve the RANS equations consists
of a third-order upwind scheme with Steger–Warming flux vector splitting
Control of Flow Past a Stalled NACA0015 Airfoil 563
method for convective terms, a central scheme for viscous terms, and first-
order backward difference for time derivative term. The implicit discretization
of the spatial derivative terms is adopted to increase the stability, and then
LU–SGS technique is used to solve the algebraic equations.
3 Results
The problem considered is the flow past a NACA0015 airfoil at angles of attack
from 2◦ to 10◦ and unit Reynolds number is 1.9512 × 105 m−1 . The sizes of
the actuators and their locations on the airfoil, and the size of the airfoil are
from reference [8].
3.1 Electric fields

According to the experimental investigation conducted by Roth, the basic
case is one actuator. Since the sizes of an actuator consisted of a pair of
electrodes is very small, the basic case is simplified to a plate board with
one electrode on its each side and its thickness is neglected. The computa-
tional domain and mesh and the contour of the electric field intensity in the
plasma region are sketched in Fig. 1. The boundary conditions are the up-
per electrode ϕ = V0 (V0 is the applied voltage, V0 = 3 kV), the lower electrode
ϕ = 0, the top and the bottom boundaries ∂ϕ/∂n = 0, the left and the right
boundaries ∂ϕ/∂n = 0(n is normal vector of the boundary). The region of the
plasma(|E∗ |ε3kV mm−1 ) is clearly shown and the most intense region of the
electric field is located between the two electrodes.
When multi-actuator work, there may be interaction between their elec-
tric fields. For an extreme case, there are infinite actuators along the plate in
sequence, the left boundary and the right boundary are set to periodic con-
dition, ϕLef t = ϕRight . However, the difference of the electric field in plasma
region (|E∗ |ε3kV mm−1 ) induced by the alternation of the boundary condi-
tion at the left and the right can be neglected because this interaction is not
so severe. Therefore, the electric field in all cases can be composed with the
electric field due to one actuator by moving its spatial location, and there is
no need to be computed for the actual configuration of the actuators.
Fig. 1. the computational domain (left), mesh (center ) and contour of the electric
field intensity(right)
Fig. 2. enlarged mesh(left) and location of actuators(right)
3.2 Flow fields
A O-type mesh is employed. The enlarged meshes are shown in Fig. 2, and
the small panes labeled with number from 1 to 8 denote the locations of eight
actuators.
The simulated cases are listed as following: (1) the influence on the flow
and aerodynamics at different angles of attack caused by the plasma due
to the actuator 1; (2) the influences of flow and aerodynamics at angle of
attack 10◦ with plasmas of 0, 2, 4, 6 and 8 actuators at different locations;
(3) the influence of a single actuator at different location on the flow and
aerodynamics at angle of attack 4◦ . Limited by the length of the paper, only
a few typical results are presented here.
Figure 3 shows that the plasma at the location of pane 1 can damp the sep-
aration at leeside of the airfoil obviously. The separation at 2◦ angle-of-attack
is eliminated, the sizes of separation are reduced and the separation pattern
of the multi-vortex are eliminated or delayed at higher angles of attack.
Figure 4 presents the curve of the normal force vs iteration steps at various
angle-of-attack. The plasma at pane 1 changes the normal force from the
constant oscillation to a certain value at 4◦ angle-of-attack, and from irregular
oscillation to a constant one at 10◦ angle-of-attack.
Figure 5 presents the curves of the peak and the valley of the oscillat-
ing aerodynamics on the airfoil vs angle-of-attack with/without the plasma
at pane 1. It shows that the plasma can reduce the amplitude, and reduce
chordwise force, while increase normal force.
Figure 6 presents the flow around the airfoil at 10◦ angle-of-attack with
the plasma of six actuators at different locations. It shows that the plasma
on the adjacence of the leading edge can damp the separation more efficiently
than the trailing edge.
Fig. 3. the influence of the plasma at pane 1 on the normal separation

Control of Flow Past a Stalled NACA0015 Airfoil 565
Fig. 4. the influence of the plasma at pane 1 on the normal force
Fig. 5. the influence of the plasma at pane 1 on the amplitude of aerodynamics
Fig. 6. Comparison of the influence of different location of plasma
Neglecting the influence of net charge density, the spatial distributions of the
electric field around the actuators have been obtained by solving the potential
equation with several types of far field boundary conditions. The numerical
results show that the influences of the interaction between actuators on the
electric field in the plasma region are trivial and neglected, the plasma-based
body forces acting on the fluid in all cases with various number of actuators
can be obtained simply and conveniently.
The numerical results agree with Roth’s experimental observation that the
plasmas locating on the leading part of the airfoil are more efficient in control
of the airfoil stall than on the trailing part. The oscillation amplitude of aero-
dynamics due to the flow separation over the airfoil is damped or eliminated,
the normal force and the stability increase while the chordwise force decreases
as plasma-based body force increases.
Acknowledgment. This work is supported by the National Science Founda-

tion of China(No.10225208, No.10321002 and No.10372110).
References
1. Roth J.R., etc.: Flow RE-attachment and Acceleration by Paraelectric and
Peristaltic Electrohydrodynamic (EHD) Effects. 41st Aerospace Science Meet-
ing and Exhibit, Reno, Nevada (2003)
2. Post M., etc.: Separation Control on High Angle of Attack Airfoil Using Plasma
Actuators. AIAA Journal, Vol.42, No.11, 2177(2004)
3. G.I. Font, etc.: Plasma Discharges in Atmospheric Pressure Oxygen for Bound-
ary Layer Separation Control. 36th AIAA Plasmadynamics and Lasers Confer-
ence, Toronto, Ontario, Canada (2005)
4. T.J. Sommerer, etc.: Numerical Investigation of the Kinetics and Chemistry of
RF Glow Discharge Plasma Sustained in He, N2, O2, He/N2/O2, He/CF4/O2,
and SiH4/NH3 using a Monte Carlo-fluid hybrid Model. J. App. Phys, 71(4),
1654(1992)
5. Roth, J.R.: Aerodynamics Flow Acceleration Using Paraelectric and Peristaltic
Electrohydrodynamic Effects of a One Atmosphere Uniform Glow Discharge
Plasma. Phys. Plasmas, Vol.10, 2117(2003)
6. W.Shyy, etc.: Modeling of glow discharge-induced fluid dynamics. J. Applied
Phys., Vol.92, No.11, 6434(2002)
7. Y.B.Suzen, etc.: Numerical Simulation of Plasma Based Flow Control Applica-
toins. 35th Fluid Dynamics conference and Exhibit, Toronto, Ontario, Canada
(2005)
8. Roth, J.R.: Subsonic Plasma Aerodynamics Using Paraelectric and peristaltic
electrohydrodynamic Effects. 29th IEEE International Conference on Plasma
Science, Banff, Alberta, Canada (2002)
9. D.V. Gaitonde, etc.: Control of Flow Past a Wing Section with Plasma-based
Body Forces. 36th AIAA Plasmadynamics and Lasers Conference, Toronto,
Ontario, Canada (2005)
Active control of a transitional flow over a
backward-facing step
E. Creusé1 , A. Giovannini2 , and I. Mortazavi3

1
LAMAV, Université de Valenciennes, 59313 Valenciennes Cedex 09, France.
Equipe SIMPAF, INRIA Futurs, site de Lille. ecreuse@univ-valenciennes.fr.
2
IMFT, 118, route de Narbonne, 31062 Toulouse, France. giova@cict.fr.
3
MAB, Université Bordeaux I, 33405 Talence Cedex, France.
Equipe MC2, INRIA Futurs, site de Bordeaux. mortaz@math.u-bordeaux1.fr.
Introduction
The backward-facing step case study, exhibiting all the essential features of
separated flows, is one of the most studied benchmarks in internal channel
flows. Research on control of the flow in such a configuration is recently de-
velopped. In severeal cases the control is performed using flow through slots
implemented in the step or the lower wall of the channel. The aim of the
present work is to develop a control, based on pulsed inlet velocities on a
two-dimensional backward-facing step for an incompressible transitional flow.
This work follows a previous study [3] on the numerical investigation of the
recirculation areas obtained by an inlet pulsed velocity in a laminar flow.
Here, the impact of the control on several significant characteristics of the
flow are carefully analysed. The numerical simulations are performed using a
Vortex-In-Cell method [1]. VIC methods for viscous flows may be veiwed as an
appealing alternative to pure grid-free vortex methods in simple geometries.
In this case, the use of fast Poisson solvers enables fast velocity evaluations.
Compared with pure finite-difference methods, VIC methods offer the advan-
tage of a robust and accurate treatment of the convective part of equations
with time steps not constrained by convective CFL conditions [2]. Moreover,
the diffusion part of equations are solved by the Random Walk Algorithm
that is well adaptable to the Lagrangian particle transport [5], [6]. In this
paper, after a brief description of the flow configuration and the numerical
scheme’s behaviour and validation, the control procedure and devices are pre-
sented. Then, open-loop and closed-loop control results are studied in order to
achieve an efficient control strategy, both with intrusive (located inside flow)
and non intrusive (wall based) sensors.

568 E. Creusé, A. Giovannini, and I. Mortazavi
1 Configuration, governing equations and numerical code

Flow configuration and governing equations. The dynamics of a two-
dimensional incompressible flow over a backward-facing step is governed by
the conservation of mass ∇. u = 0 in Ω and the Vorticity Transport Equa-
tion (VTE) obtained taking the curl of Navier-Stokes equations:
∂ω 1
+ u .∇ ω = ∆ω in Ω, (1)
∂t Re
where u is the velocity vector, ω the vorticity, Ω the computational domain
of Figure 1 with Ldown = 10, Lup = 1, Hstep = 0.5 and Hup = 0.5, similar to
[6]. Here, u = (u, v) is the velocity normalized with respect to a given velocity
U , x = (x, y) the two orthogonal space directions normalized with respect to
2H where H is the physical height of the step (that’s why the computational
height is Hstep = 0.5), t the time normalized with respect to 2H/U , and p
the pressure normalized with respect to ρU 2 where ρ is the density of the
fluid. The numerical Reynolds number Re is then defined by Re = 2ρ U H/µ,
µ being the dynamic viscosity of the fluid, which corresponds to a physical
Reynolds number Reφ = ρ U H/µ.
Numerical scheme and Boundary conditions. The solution of the pre-

vious equations is approximated using a Vortex-In-Cell method [1]. The VTE
is solved in two fractional steps solving separately the advective and diffusive
terms [2]. In the first fractional step, the transport of vorticity due to con-
vection is obtained in terms of the Lagrangian displacement of a set of finite
vortex elements. The Poisson equation −∆ψ = ω with associated boundary
conditions is solved on a grid to recover the stream function field Ψ n (i, j)
at time tn . By a linear interpolation procedure, a convective velocity u nv (it)
is then associated to each finite vortex element (1 ≤ it ≤ nv n ), and so a
convective displacement given by dl nconv (it) = u nv δt, with δt = tn+1 − tn . In
the second fractional step, the solution of the diffusion equation is simulated
by the random walk displacement of the same vortex elements [4]. Then, the
location of each finite vortex element x n+1∗ (it) at time tn+1 is obtained by
adding the convective (dl nconv (it)) and diffusive (dl ndif f (it)) movements:
x n+1
∗ (it) = x n (it) + dl nconv (it) + dl ndif f (it). (2)
Γwall
B
Γin Hup
Ω
A Γout
Hstep
O C
Lup Ldown
Fig. 1. The computational domain.

Active control of a transitional flow over a backward-facing step 569
The last step of the algorithm is to generate new vortex elements on the wall
to cancel the slip velocity and to satisfy the no-slip boundary condition.The
circulation of each vortex is finally distributed on the nodes of the mesh imbed-
ded on rotationnal zones, which allows to recover the vorticity field ω n+1 (i, j)
on the grid, and to repeat the whole procedure to get the variables at the
following time steps. The boundary conditions in vortex methods are based
on stream function Ψ definition related to the vorticity. On the bottom part
of Γwall the no-slip boundary condittion yields to Ψ = 0. On Γin the uniform
velocity is defined as ∂Ψ/∂y = 1. For the top Γwall , Ψ = Ψ (B) = Hup − Hstep
and for Γout , ∂ 2 Ψ/∂x2 = 0.
Validation of the simulation. The control-off simulations are validated in

the backward-facing step geometry for Re = 500 and Re = 2000. For the first
Reynolds number corresponding to a low transitional flow, the comparisons
mainly reffer to [6]. The grid convergence is achieved for a 330 × 30 based on
the length and the height of the computational domain, with Γmin = 1E − 04
and δt = 1E −01. Here, Γmin is the elementary circulation of a vortex element
[5], [6]. The comparisons are based on the average length of the recirculation
zone and the average velocity profiles in different sections of the channel. A
detailed study on an appropriate choice of the discretization parameters for
the vortex methods is available in [5]. For Re = 2000 (a highly transitional
flow) the validations reffer to [7] based on the Strouhal number value. A
Strouhal number close to St = 0.067 confirms the results.
2 Control approaches
Frequency analysis of the uncontrolled flow. A preliminary test based

on a non-controlled flow simulation consists in a spectral analysis of ∂ 2 u/∂y 2
as a function of the time, who represents the longitudinal pressure gradient at
the wall, at 3 sensors on the bottom wall (y = 0) respectively located at x = 6,
x = 7 and x = 8 (see Figure 2). For this simulation, the average recircula-
tion area lenght is equal to Lr = 4.51 and oscillates between its minimal and
maximal values respectively equal to 2.60 and 5.89. For each of these sensors,
the dominant frequency of the signal is equal to 0.1367, what corresponds to
a Strouhal number very close to 0.067 obtained in [7].
Control Functionals and devices. In order to quantify the control effect,

we introduce the following time dependant fonctionnals :
- The recirculation area lenght Lr (t); ZZ
- The total enstrophy in the domain, defined as J1(t) = ω 2 (t) dx dy;
Ω
- The near wall recirculation area intensity, defined as :
Z C
J2(t) = ∂u/∂y(t) (∂u/∂y(t) − |∂u/∂y(t)|) dx;
0
d2udy2 at sensor located at x=6, y=0 5 Frequency analysis

x 10
150 2
100
50 1.5
Sensor signal
Energy
0
1
−50
−100 0.5
−150
−200 0
100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5
Time Frequency
x 10
150 2
100
50 1.5
Sensor signal
Energy
0
1
−50
−100 0.5
−150
−200 0
100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5
Time Frequency
x 10
150 2
100
50 1.5
Sensor signal
Energy
0
1
−50
−100 0.5
−150
−200 0
100 200 300 400 500 600 700 0 0.1 0.2 0.3 0.4 0.5
Time Frequency
2 2
Fig. 2. Signal of ∂ u/∂y (t) at non intrusive sensors on the bottom wall at x = 6
(top), x = 7 (middle) and x = 8 (bottom) and spectral analysis.
ZZ
dx dy
- The fluctuation coefficient, defined as J3(t) = (u0 (t))2 +(v 0 (t))2
;
B 1+ 2
where B is the box [0; 6] × [0; 0.5], and where u0 (t) (resp. v 0 (t)) is the variation
of u(t) (resp. v(t)) around its mean value.
The time-average functionals related to above parameters for a simulation
between t = 320 and t = 600 are noted by Lr , J1, J2 and J3. The objective of
this work is to control these time-average functionals. The main control device
is to replace the uniform inlet velocity by a pulsing inlet velocity from t = 300
to t = 600. The Ψ boundary conditions are fitted to this new inlet condition.
The pulsing frequency and amplitude choice will be defined in the next section.
Open-loop Control. This kind of control is performed by imposing uin (t) =

1.0+A cos(2 π f t) on Γin , instead of the steady boundary condition uin (t) = 1
associated to the non-controlled simulation. We set A = 0.2. For 20 low fre-
quencies uniformly distributed in the range of [0.0567; 0.4367], around the
fundamental frequency 0.1367, we plot the time-average functional values de-
fined in section 2 as a function of the frequency f (see Figure 3).
As the figure shows, we achieve an optimal value for each functional for a
frequency nearly equal to the basic fundamental frequency of the flow. The
recirculation length as well as J1 and J3 are significatively diminished, and
J2 is consequently increased. It corresponds to the fact that the recirculation
area is broken by this control strategy, and then the total enstrophy of the
flow descreases and is focussed in the near wall small advective coherent struc-
tures. For the larger values of the frequency, the efficiency of the control is
reduced, but it becomes stable with a permanent effect on the flow behaviour.
Moreover, the parametric study on the amplitude (A ∈ [0; 0.2]) shows that
Active control of a transitional flow over a backward-facing step 571
Lr J1
5 70
65
4
60
3
55
2 50
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
f f
J2 J3
200 3.2
180 3.18
3.16
160
3.14
140
3.12
120
3.1
0 0.1 0.2 0.3 0.4 0 0.1 0.2 0.3 0.4
f f
Fig. 3. Open-loop control : Functionals versus frequency. Dashed line : uncontrolled
simulation.
the efficiency of the control almost linearly increases with its value.
Closed-loop Control. A first closed-loop control approach specifies the inlet

velocity value by :
uin (t) = 1.0 + M ∂ 2 u/∂y 2 (t) S ,

n
where M = 2.5E − 03 is chosen to make vary uin (t) in the range of [0.8; 1.2] as
in the previous section, and ∂ 2 u/∂y 2 (t) Sn is the value of ∂ 2 u/∂y 2 (t) related
to a non intrusive sensor Sn located at xSn = 8 and ySn = 0.
An alternative procedure is to specify the inlet velocity value by :

!
uSi (t) − uSi
uin (t) = 1.0 + 2 A ,
umax
Si − umin
Si
where the intrusive sensor Si is located at x = 6 and y = 0.33, and where uSi ,
umin
Si and umax
Si are respectively the average, the minimum and the maximum
values of the uncontrolled velocity uSi (t) computed for the same time and
space location. We set A = 0.2, leading to an inlet velocity uin oscillating
between 0.8 and 1.2. Values of the functionals are reported on Table 1.
As Table 1 shows, closed-loop control is efficient since each obtained func-

tional value is close to the best result achieved by the open-loop control. Fur-
thermore, the control effect in the intrusive case is always more significative
than in the non intrusive one. This behaviour can be justified by the fact that
Lr J1 J2 J3
Uncontrolled case 4.52 63.53 122.09 3.190
Best value, open-loop 2.09 55.28 186.71 3.137
Closed-loop and non intrusive sensor 2.94 57.41 141.83 3.162
Closed-loop and intrusive sensor 2.38 52.64 159.95 3.150
Table 1. Functional values : uncontrolled, open-loop control, closed-loop control.
the intrusive control sensor is located in the transport dominated part of the
flow and then is not perturbated by the near-wall local effects. However the
optimal open-loop control results remain the best (except for J1) compared
to the closed-loop control values. It seems to be due to the noise free property
of uin (t) in the open-loop control.
Conclusion
In this work, both open and closed loop control methods were applied to a
backward-facing step flow, pulsing the inlet velocity conditions. The achieved
results showed that the natural fundamental frequency of the flow offers the
characteristic value to tune the pulsing frequency. The control should there-
fore be based on this value, whatever the used devices. We also observed the
efficiency of implementing the intrusive sensors. This approach improves the
closed-loop control results, reducing the near-wall noisy flow fluctuations.
References
1. Christiansen, J.P. : Numerical Simulation of Hydrodynamics by the Method of
Point Vortices, J. Comp. Phys., 13, 363–379 (1971).
2. Cottet, G.H. and Koumoutsakos, P. : Vortex methods : Theory and Practise,
Cambridge University Press (2000).
3. Creusé, E., Mortazavi, I.: Simulation of Low Reynolds Number Flow Control
over a Backward-Facing Step Using Pulsed Inlet Velocities, AMRX, 4, 133–152
(2004).
4. Ghoniem, A. F. and Gagnon, Y. : Vortex Simulation of Laminar Recirculating
Flow, J. Comp. Phys., 68, 346–377 (1987).
5. Mortazavi, I., Micheau, P. and Giovannini, A. : Etude de la convergence
numérique d’une méthode vortex pour un écoulement à grand nombre de
Reynolds dans un mélangeur, C.R. Mécanique, 330, 409–416 (2002).
6. Sethian, J.A., Ghoniem, A.F. : Validation Study of Vortex Methods, J. Comp.
Phys., 74, 283–317 (1988).
7. Wee, D., Yi, T., Annaswamy, A. and Ghoniem, A.F. : Self-sustained oscilla-
tions and vortex shedding in backward-facing step flows : Simulation and linear
instability analysis, Physics of Fluids, 16, 9, 1–12 (2004).
Synthetic Jet Actuator Modeling for Flow
Control Applications
M. Ferlauto and R. Marsilio
DIASP - Politecnico di Torino, Corso Duca degli Abruzzi,24, 10129 Torino, Italy,
michele.ferlauto@polito.it, roberto.marsilio@polito.it
Summary. A numerical model of synthetic jet actuator, suitable for flow control
simulations, is presented. The controlled flowfield is simulated by a standard CFD
method for compressible RANS equations, while flow inside the actuator is reduced
to a one-dimensional piston flow. The nonlinear matching between the two systems
ensures conservation of the mass, momentum and energy.
1 Introduction
Recent advances in flow control deals with Synthetic Jet (SJ) manipulators.
This technology has widespread potential applications in aeronautics, with
the aim of enhancing performances and safety margins against flow separa-
tion and stall. With little costs one can, for instance, modify the pressure
distribution over an airfoil at high angle of attack and enhance remarkably
the stall characteristics. The application of such devices has been investigated
for separation control [1], noise or drag reduction and lift enhancement [2],
flow vectoring [3] and more. The related results reveal the great potentials of
flow manipulation via synthetic jets, as well as the need of tools to investigate
and assess adequate control strategies.
From the CFD point of view, one has to deal with unsteady, turbulent
flows in time varying domains. Therefore, the simulation of flowfields actively
controlled via synthetic jets requires accurate computational schemes in both
time and space. The full simulation of the flowfield inside the actuator and
of the main flow is feasible in simple cases and for short integration time. As
alternative, the followed approach is to simulate the presence of actuators by a
boundary condition (BC) that imposes a periodically oscillating velocity pro-
file at the actuator orifice [4]. This modelisation has difficulties in evaluating
the correct pressure level into the actuator since it does not ensure the con-
servation of mass, momentum and energy [5], which is crucial in the suction
phase of the cycle, where pressure is the variable that determines the influx
of mass. More complicated models, based on linear governing equations, also

574 M. Ferlauto and R. Marsilio
Main Flow
Piston
Fig. 1. Actuator Model.
fail in capturing flow resonances and nonlinearities [6][7]. For the region far
from the junction between actuator and main flow, a reduced order model has
been suggested in [5] . The approach retains most of the features of the full
simulation with significant CPU savings but a large number of computational
points are still needed to resolve accurately the junction region. This fact rep-
resents a limitation for the application of the model to flow control problems,
involving long time unsteady simulations with one or multiple jet actuators.
In the present paper a nonlinear model of synthetic jet actuator is pro-
posed, based on the domain decomposition approach. The model is a reduction
of [5] suitable for flow control applications. The main flowfield is simulated
using standard CFD technique for compressible, turbulent flows. The flow into
the actuator is modeled as the one-dimensional piston flow, based on the Euler
equation. The interaction at the interface of the two flowfields is evaluated by
solving a Riemann problem. With little CPU expense compared to the BC ap-
proach, the proposed approach ensures the conservation of mass, momentum
and energy across the two systems. Simulations of some flow manipulations
are discussed and compared with available data from literature.
2 Mathematical Model
2.1 External flow
The main flowfield is simulated using a finite volume discretization of the com-
pressible Reynolds Averaged Navier-Stokes equations, second order in time
and space by using Flux-Difference Splitting and Essentially Non-Oscillatory
schemes. The one-equation model of Spalart-Allmaras is used for the turbu-
lence modeling. The computational domain is truncated to a finite extent in
Synthetic Jet Actuator Modeling for Flow Control Applications 575
all directions and appropriate boundary conditions applied to these compu-

tational boundaries. Total conditions, flow angles and boundary layer charac-
teristics are specified at inlets. Static pressure is imposed at exit boundaries.
Asymptotic free stream states are treated as non-reflecting boundary condi-
tions. This choice of boundary conditions coupled with the use of a relatively
large computational domain ensures that the results are insensitive to the
domain size.
2.2 Actuator flow model
Consider the synthetic jet device in Fig. 1 attached, for instance, beneath a
flat plate on which develops a turbulent boundary layer. The synthetic jet
is created at the slot by solving the flow generated by an oscillation piston
moving inside the duct. The motion of the piston is given by Y (t) = A0 sin(ωt)
where ω is the angular frequency of the piston and A0 is the amplitude. The
duct is characterized by is length L and width h. For flow control applications,
the details of the flow inside the actuator are unessential, while the interaction
with the external flow is of relevance. The flow into the actuator is then
modeled as the piston flow, based on the one-dimensional Euler equation. At
the interface the flow state between the two fields is evaluated by solving a
Riemann problem. Within this model, a profile varying along the actuator
orifice is regarded as a sequence of pistons moving synchronously. Similarly,
a vibrating membrane can be simulated by a series of pistons moving with
imposed amplitudes but same frequency (see Fig.1). Viscous effects inside the
piston duct can be accounted for by a distributed loss model.
The numerical solver adopted for the flow in the duct integrates the one-
dimensional, compressible, unsteady Euler equations in time varying domains.
The following transformation of the independent variables is performed
x − b(t)
ξ= , τ =t (1)
c(t) − b(t)
where b(t) and c(t) are the time varying position of the left and right boundary,
respectively. The flow governing equations are rewritten as

∂ U ∂ U ∂ξ F (U ) ∂ξ
+ + =0 (2)
∂τ J ∂ξ J ∂t J ∂x
where
   
 ρ   ρu 
∂ξ
U = ρu F (U ) = p + ρu2 , J= (3)
∂x
e (e + p)u
   
and are integrated using a finite volume method fully compliant to that used
for the external flow and having the same accuracy properties. Conventionally,
the left boundary represents the piston wall, moving with velocity b(t) = Y (t).
The right boundary (c(t) = 0) is assumed as the synthetic jet slot, where a
Riemann problem between the internal and the external flow is computed.
1 1.1
1
0.9 Freq = 20Hz
Freq = 20Hz 0.9 THE ORY
0.8 THEORY Freq = 40 Hz
Freq = 40 Hz 0.8
0.7
0.7
Ucl/Ucl-max
0.6
U/Ucl
0.6
0.5
0.4 0.5
0.3
0.4
0.2
0.1
0 0.3 0 1 2
-2 -1 0 1 2 10 10 10
Y/b X/h
Fig. 2. Synthetic jet in a quiescent air: selfsimilarity and decay rate.

a b c
T T T
1.00172 1.00172 1.00172
1.00154 1.00154 1.00154
1.00137 1.00137 1.00137
1.00119 1.00119 1.00119
1.00102 1.00102 1.00102
1.00084 1.00084 1.00084
1.00067 1.00067 1.00067
1.00049 1.00049 1.00049
1.00031 1.00031 1.00031
1.00014 1.00014 1.00014
0.999964 0.999964 0.999964
d e f
T T T
1.00172 1.00172 1.00172
1.00154 1.00154 1.00154
1.00137 1.00137 1.00137
1.00119 1.00119 1.00119
1.00102 1.00102 1.00102
1.00084 1.00084 1.00084
1.00067 1.00067 1.00067
1.00049 1.00049 1.00049
1.00031 1.00031 1.00031
1.00014 1.00014 1.00014
0.999964 0.999964 0.999964
Fig. 3. Synthetic jet in a quiescent air (f = 20Hz): temperature contours at (a)

t ◦ 0,(b) t = T /4, (c)t=T/2,(d) t = T , (e)t=5T/4,(f) t = 3T /2 .
3 Numerical Results
In this section two case of flow manipulation using synthetic jet actuator
are described: the generation of a synthetic jet in a quiescent air and the
interaction of a flat plate boundary layer and the synthetic jet.
3.1 Synthetic jet in quiescent air
The synthetic jet produced in a quiescent external flow has been extensively
studied ([8, 9]). The observation that the far field effects of a turbulent jet
can be reproduced by this type of actuators is the starting point of the whole
research on synthetic jets. At moderate frequencies, during the ejection phase,
the piston pushes the flow out of the actuator and a pair of counter rotating
Synthetic Jet Actuator Modeling for Flow Control Applications 577
0.03
0.025
0.02
0.015
Y
0.01
0.005
-0.005
1 1.1 1.2
X
0.03
0.025
0.02
0.015
Y
0.01
0.005
-0.005
1 1.1 1.2
X
Fig. 4. Turbulent Boundary layer ma-

nipulation. Entropy contours and stream-
lines during the blowing (top) and suction Fig. 5. Experimental (3D) and numerical
phase (bottom). (2D) velocity at the slot centerline during
a forcing period.
vortices form at the manipulator orifice (Fig. 3). This vortex pair travels in
the external flow by its own induced velocity. When the piston reverts its
motion, it entrains external fluid through the slot. As long as the vortices
have already travel away from the orifice, they are not effected by the motion
of the entrained fluid. During the subsequent ejection phase, a new dipole
is formed into the fluid. The self-similar characteristics of the flow are show
in Fig. (Fig2) where the transverse velocity profiles at different stations, ob-
tained numerically for different frequencies (f1 = 20Hz and f2 = 40Hz) are
compared with the theoretical solution of a turbulent two-dimensional jet.
The decay of the centerline jet velocity also reach the theoretical rate at a
distance of about 10h .
3.2 Jet in an external cross flow
The interaction of a synthetic jet with a flat plate boundary layer is stud-
ied numerically. Due to the lack of availability of experimental data in tran-
sonic, high-Reynolds-number flow regime, the simulations we made are not
suitable to validate rigorously the actuator model. The test-case, fully three-
dimensional, proposed in [7] has been reduced to a two-dimensional configura-
tion while maintaining the main flow parameters. In Fig. 4 two flow patterns
are shown, the upper concerns the blowing phase, the lower the suction phase.
Although obvious discrepancies between the 3D and 2D flow configurations,
the main flow feature on which the test-case focused, that is the time evolution
of the velocity at the actuator orifice centerline Ucl , was recovered (see Fig. 5).
The observed behaviour is due to the nonlinearities. The replacemente of the
actuator by a periodic boundary condition would lead instead to a sinusoidal
evolution in time.
4 Conclusions
The main outlines of a numerical model of synthetic jet actuator has been
presented. The model aims to overcome drawbacks related to the lack of con-
servativeness of other approaches suitable for flow control simulations. The
significant increase of computational costs of more sophisticate modelisations
is avoided. A preliminary validation has been performed. Rigorous validation
and fine tuning are expected during the experimental testings of the under-
going research project.
Acknowledgments
This work was supported by Regione Piemonte through the grant E40-2004.
References
1. Seifert, A., Darabi, A., Wygnanski, I.: Delay of airfoil stall by periodic excitation.
AIAA J., 33(4), 691-707 (1996).
2. Smith, D., Amitay, M., Kibens, V., Parekh, D., Glezer, A.: Modification of lifting
body aerodynamics using synthetic jet actuators. AIAA Paper 98-0209 (1998).
3. Guo, D., Gary, A.W.: Vectoring Control of a Primary Jet with Synthetic Jets.
AIAA paper 2001–0738 (2001).
4. Donovan, J., Kral, L., Cary, A.: Active flow control applied to an airfoil. AIAA
Paper 98-0210 (1998).
5. Yamaleev, N.K., Carpenter, M.H.: A Reduced-Order Model for Efficient Simu-
lation Synthetic Jets. NASA TM 2003-212664 (2003).
6. Gallas, Q. et al.: Lumped element modeling of piezoelectric-driven synthetic jet
actuators. AIAA paper 2002-0125 (2002).
7. AA.VV. : CFD Validation of Synthetic Jets and Turbulent Separation Control.
Langley Research Center Workshop, Virginia, USA (2004).
8. James, R.D., Jacobs, J.W., Glezer, A.: A round turbulent jet produced by an
oscillating diaphram. Phys. Fluids, 9(8) (1996).
9. Rizzetta, D.P., Visbal, M.R., Stanek, M.J.: Numerical investigation of synthetic
jet flowfields. AIAA Paper 982910 (1998).
Numerical Study of Transonic Drag Reduction
for Flow Past Airfoils Using Active Flow
Control
Jose Vadillo1 and Ramesh K. Agarwal1
Washington University, St. Louis, MO 63130, USA rka@me.wustl.edu
Summary. The objective of this paper is to evaluate, via numerical simulation, the
feasibility of weakening the shock wave(s) and reducing the pockets of supersonic
regions on airfoils/wings in transonic flight using active flow control (AFC) devices
such as synthetic jet actuators (SJA), and thereby achieving the desired changes
in aerodynamic forces and moments. Most importantly, it is shown that both the
enhancement in lift and reduction in drag can be achieved by using multiple synthetic
jet actuators suitably placed on the upper surface of the airfoil.
1 Introduction
In recent years, there has been a great deal of emphasis toward the develop-
ment of advanced aerodynamic technologies based on the fluidic modification
of aerodynamic and propulsive flow fields that can cover multiple flight condi-
tions without the need for conventional control surfaces such as flaps, spoilers
and variable wing sweep. The fluidic modification (or flow control) can be
accomplished by employing micro-surface effectors or actuators operated by
an intelligent control system. One of the promising device for AFC that has
been explored extensively is the zero net-mass-flux synthetic jet (or oscilla-
tory jet) actuator. With synthetic jet actuator, a variety of impressive flow
control results have been achieved experimentally including the vectoring of
conventional propulsive jets, modification of aerodynamic characteristics of
bluff bodies, control of lift and drag of airfoils, enhanced mixing in circular
jets, reduction of skin-friction of a flat plate boundary layer etc. For subsonic
flow past airfoils at low angles of attack, recently it was shown experimentally
by Amitay et al [1] and computationally by the authors of this paper [2] that
the pressure drag of an airfoil can be significantly reduced with a minimum
change in lift by modification of the apparent aerodynamic shape of the airfoil.
This virtual aerodynamic shape modification can be achieved by creating a
small domain adjacent to the upper airfoil surface (downstream of the leading
edge) using a synthetic jet actuator which displaces the local streamlines suffi-
ciently to modify the local pressure distribution. These results have been very

580 Jose Vadillo and Ramesh K. Agarwal
encouraging. However, although some computational and experimental work

has been done in the application of active flow control for virtual aero-shaping
of airfoil to reduce drag in subsonic flow, hardly any work has been reported
in the literature to date for active control of the strength of shock waves and
their unsteadiness, the supersonic regions and shock-induced separation due
to shock/boundary layer interaction. This problem is of great relevance for
reducing the transonic cruise drag of a transport aircraft and high speed im-
pulsive noise of a rotorcraft. To explore this possibility, the authors recently
conducted a numerical study [3] which showed the potential for modulation
of aerodynamic forces and moments in transonic flow past an airfoil using
a synthetic jet actuator. However in the study reported in [3], the goal of
reduction of transonic drag with minimum change in lift was not achieved.
This follow up study has indicated that the multiple actuators are needed to
achieve this objective. Both the computations performed by the authors and
the experiments conducted at Boeing demonstrate the potential for reducing
the transonic drag with minimum change in lift (or even enhancing the lift)
by suitably employing three synthetic jet actuators on the upper surface of
the airfoil. This paper reports the results of this numerical study.
2 Solution methodology
CFD simulations were performed by employing an extensively validated well

known multi-zone Unsteady Reynolds- Averaged Navier-Stokes (URANS) flow
solver WIND [4] using the MenterŐs two-equation Shear Stress Transport
(SST) model [5]. CFD code WIND is a product of the NPARC Alliance, a
partnership between the NASA Glenn Research Center (GRC) and the Arnold
Engineering Development Center (AEDC). WIND computes the solution of
the Euler and Navier-Stokes equations, along with supporting equation sets
for turbulent and chemically reacting flows by employing a variety of turbu-
lence models and chemistry models respectively. WIND is coded in Fortran
77, Fortran 90, and C programming languages. The governing equations are
solved in conservation form. Explicit viscous terms are computed employing
the second-order central differencing and the convection terms are discretized
using the third-order accurate Roe upwind algorithm. In all the simulation re-
sults presented here, the overall spatial accuracy of the algorithm is of second-
order. The time-implicit convection terms are computed using an approximate
factorization scheme with four-stage Runge-Kutta time-stepping. WIND uses
externally generated computational grids. The solution is executed iteratively
on this grid. For all the computations reported in this paper, the grid refine-
ment study was performed to ensure that the computed solutions are grid-
independent. The uncertainty in the computed solutions due to the numeric
(size of the computational domain, grid, numerical algorithm and boundary
conditions) was minimized.
Transonic drag reduction using active control 581
3 Results and discussion

The results are presented in two parts. In the first part of the study, the com-
putations were performed for transonic flow past a NACA64A010 airfoil using
AFC for a number of free-stream conditions and synthetic jet parameters. The
baseline conditions (employed as reference conditions) for flow without AFC
were Mach number M∞ = 0.8, chord Reynolds number Rec = 5.67e+6, and
angle of attack α = 1◦ . For this baseline condition, the foot of the shock on
the upper surface of the airfoil was determined to be at xj /c. A parametric
study was then performed by varying both the free-stream conditions and
the synthetic jet parameters to evaluate the impact of AFC on lift, drag, and
pitching moment coefficients. The range of parameters covered in the study
was: M∞ = 0.8, 0.85 and 0.9; angle of attack α = 0, 1 and 2 deg; jet Mach
number Mj = 0.1, 0.2., and 0.3; and jet width b/c = 0.0035, 0.0044, and
0.0055. In all the computations, Rec = 5.67e+6, jet frequency f = 150 Hz and
jet location xj /c= 0.53. It should be noted that the location xj /c= 0.53 for
foot of the shock (without AFC) changes for different values of M∞ and α.
Figure 1 shows the variation in lift coefficient Cl and drag coefficient Cd with
angle of attack (0, 1 and 2◦ ) for M∞ = 0.8, Rec = 5.67e+6, b/c = 0.0044, f
= 150Hz, Mj = 0.3, and xj /c= 0.53. The calculations show that Cl decreases
significantly compared to the baseline case while Cd increases slightly by using
AFC. Figure 2 shows the variation of Cl and Cd with jet width (b/c = 0.0035,
0.0044, and 0.0055) for M∞ = 0.8, Rec = 5.67e+6, α = 1◦ , f = 150Hz, Mj =
0.3, and xj /c= 0.53. These calculations show that both Cl and Cd decrease
significantly by using AFC. Figure 3 shows the variation of Cl and Cd with
jet Mach number (Mj = 0.0035, 0.0044, and 0.0055) for M∞ = 0.8, Rec =
5.67e+6, α = 1◦ , f = 150Hz, b/c =0.0044 and xj /c= 0.53. These calculations
again show that both Cl and Cd decrease significantly by using AFC compared
to the baseline case. While the decrease in Cd is a desirable goal, a significant
decrease in Cl is not desirable under cruise conditions. However, these changes
in aerodynamic coefficients may be desirable in descent flight or for chang-
ing the pitching moment coefficient. Thus it is possible to use AFC to attain
desired changes in aerodynamic forces and moments with the ultimate aim
of achieving the ’hingeless’ control. However, the goal of reducing drag with
minimal change in lift was not achieved by employing a single synthetic jet at
or near the foot of the shock of the airfoil (without AFC). Computations were
therefore performed by using multiple synthetic jets on the upper surface of
the airfoil. In the second part of this study, a large number of computations
were performed with three synthetic jets employed at three different locations
on the upper surface of the airfoil. In the following, the results obtained for
almost optimal locations of the jets are presented. These results show that it
is possible to both enhance lift and reduce drag in transonic flow by employing
suitably placed multiple synthetic jet actuators on the upper surface of the
airfoil. Computations were performed for flow past a NACA 64A010 airfoil in
transonic flow at Mach 0.8 and 0.9, angle of attack = 2.8◦ , chord Reynolds
Cl (Baseline) Cl (AF C) Cd (Baseline) Cd (AF C)

4.14e-1 4.79e-1 2.83e-2 2.55e-2
Table 1. Comparison of the computed lift and drag coefficients for the flow past a
NACA64A010 airfoil at M∞ = 0.8, Rec = 5.67e + 6 and α = 2.8◦
Cl (Baseline) Cl (AF C) Cd (Baseline) Cd (AF C)

2.29e-1 2.49e-1 8.36e-2 8.26e-2
Table 2. Comparison of the computed lift and drag coefficients for the flow past a
NACA64A010 airfoil at M∞ = 0.9, Rec = 5.67e + 6 and α = 2.8◦
number Re = 5.67 10+6, with and without AFC using three synthetic jets
on the upper surface of the airfoil; experiments for this configuration were
performed at Boeing. The three synthetic jets were installed on the upper
surface of the airfoil at x/c = 0.1, 0.53 and 0.68, c being the airfoil chord. The
three jets were identical in width b/c =0.00084 with frequency f = 220 Hz and
velocity amplitude Mj = 0.05, and operated in phase. The grid requirements
were very sensitive to the flow conditions. To maintain y + < 2, a different
computational grid was required for each flow condition M∞ = 0.8 and 0.9.
Figures 4(a) and 4(b) show the close-up view of the two grids used for M∞
= 0.8 and M∞ = 0.9 respectively. Both the grids have a C-H topology with
296 x 61 nodes; the main difference between the two grids is in the cluster-
ing near the airfoil surface. The boundary conditions for the three synthetic
jets are imposed at the airfoil surface without accounting for the cavity. It
is important however that in addition to sinusoidal velocity of the synthetic
jet, the pressure and temperature at the three jet locations (x/c = 0.1, 0.53,
and 0.68) are correctly specified. We employ the pressure and temperature
at these locations calculated for the baseline airfoil (without synthetic jets).
Figures 5(a) and 5(b) show the computed pressure coefficient Cp at M∞ =
0.8 and M∞ = 0.9 respectively. Figures 6(a) and 6(b) respectively show the
pressure drag coefficient without and with AFC, and the pressure lift coeffi-
cient without and with AFC for M∞ = 0.8, while Figures 7(a) and 7(b) show
these quantities for M∞ = 0.9. These figures demonstrate the effectiveness
of AFC using suitably placed multiple synthetic jets, in both reducing the
drag and enhancing the lift. Tables 1 and 2 respectively show the reduction
in drag and increase in lift at Mach 0.8 and 0.9 using AFC. These results are
very close to the experimental values obtained in the experiments performed
at Boeing. These calculations, for the first time, demonstrate the potential of
AFC in transonic drag reduction.
Fig. 1. Numerical results for a NACA 64A010 airfoil: M∞ = 0.8, Rec = 5.67e+6,
α = 0, 1 and 2◦ , b/c = 0.0044, f = 150Hz, Mj = 0.3, xj /c= 0.53.
α = 1◦ , b/c = 0.0035, 0.0044, and 0.0055, f = 150Hz, Mj = 0.3, xj /c= 0.53
α = 1◦ , b/c = 0.0044, f = 150Hz, Mj = 0.1, 0.2 and 0.3, xj /c= 0.53
Fig. 4. Computational domain and grids used in the simulations for flow past a
NACA 64A010 airfoil with three synthetic jet actuators.
Fig. 5. Computed pressure coefficient distributions Cp for flow past a NACA64A010

airfoil (without AFC) at M∞ = 0.8 and M∞ = 0.9, Rec = 5.67e+6 and α = 2.8◦ .
Conclusions
This paper reports the results of numerical simulations on investigating the

feasibility of controlling the shock/boundary layer (SBL) interactions and re-
gions of supersonic flow on airfoils in transonic flow by employing the active
flow control (AFC) with oscillatory zero-net-mass jet actuators. It is shown
for the first time that with multiple actuators suitably placed on the upper
surface of the airfoil, it is possible to both enhance lift and reduce drag in
transonic flow. Good agreement with experimental data is obtained.
Fig. 6. Comparison of the computed pressure drag coefficient and pressure lift
coefficient, without and with AFC for flow past a NACA 64A010 airfoil at M∞ =
0.8, Rec = 5.67e+6, α = 2.8◦ , b/c = 0.00084, f = 220 Hz, and Mj = 0.05 using three
synthetic jet actuators located at x/c = 0.1, 0.53, and 0.68 on the upper surface of
the airfoil.
Fig. 7. Comparison of the computed pressure drag coefficient and pressure lift
coefficient, without and with AFC for flow past a NACA 64A010 airfoil at M∞ =
0.9, Rec = 5.67e+6, α = 2.8◦ , b/c = 0.00084, f = 220 Hz, and Mj = 0.05 using three
synthetic jet actuators located at x/c = 0.1, 0.53, and 0.68 on the upper surface of
the airfoil.
References
1. Amitay, M., Horvath, M., Michaux, M. and Glezer, A., ”Virtual Aerodynamic
Shape Modification at Low Angles of Attack Using Synthetic Jet Actuators”,
AIAA Paper 01-2975, 2001
2. Vadillo, J.L., Agarwal, R.K., Cary A.W. and Bower, W.W., ”Numerical Study
of Virtual Aerodynamic Shape Modification of an Airfoil Using a Synthetic Jet
Actuator”, AIAA Paper 03-4158, 2003
3. Vadillo, J.L., Agarwal, R.K. and Hassan, A.A., ”Modulation of Aerodynamic
Forces and Moments at Transonic Speeds Using Active Flow Control”, AIAA
Paper 05-0486, 2005
4. Bush, R.H., ”The Production Flow Solver of the NPARC Alliance”, AIAA Paper
88-0935, 1988.
Part VII
Two-phase and multimaterial flows

Level Set based Finite Element Method of
Bubble-in-Liquid Simulation
Hyoung G. Choi1 and Jung Y. Yoo2

1
Department of Mechanical Engineering, Seoul National University of
Technology, 172 Gongreung-2-Dong, Nowon-Gu, Seoul 139-743, Republic of
Korea (hgchoi@snut.ac.kr)
2
School of Mechanical and Aerospace Engineering, College of Engineering, Seoul
National University, San 56-1, Shillim-Dong, Kwanak-Gu, Seoul 151-742,
Republic of Korea (jyyoo@snu.ac.kr)
1 Introduction
There are various numerical methods to perform a direct numerical simulation

of a bubble-in-liquid flow. Among them are level set method[1], front tracking
method, volume of fluid(VOF) and body fitted coordinate method[2]. Body
fitted coordinate method is able to treat the jump condition across the phase
interface most accurately since a body fitted structured or unstructured mesh
around a bubble interface, across which singularity occurs, is generated. How-
ever, this method has a difficulty in solving the merging/breaking case of a
bubble since the generation of a body fitted mesh of the case with time is
formidable. Front tracking method adopts a structured background mesh in
order to treat the singularity across an interface. However, it is quite compli-
cated for this method to treat the merging/breaking case of a bubble. Com-
pared with the body fitted and front tracking method, both VOF and level
set method are easy to implement and the merging/breaking problem of a
bubble is more easily treated. VOF and level set method use a background
mesh, typically structured one, and the phase interface is defined by VOF and
level set function, respectively. The physical understanding of both methods
is straightforward since the evolution of an interface is obtained by the advec-
tion of the interface with time.
However, VOF method has to perform an ad-hoc procedure[3] to find the
filling pattern of control volumes around interface, which is necessary step
for the construction of a newly defined interface at each time step. On the
other hand, the level set method does not require that procedure. Further-
more, the level set method has an advantage over other methods in that the
interface with singularity is managed by the continuous level set function,
which is differentiable, and the extension of a two-dimensional level set code
to the three-dimensional one is straightforward. On the other hand, the level

590 Hyoung G. Choi and Jung Y. Yoo
method has to additionally solve a redistancing equation of hyperbolic type,

which is essential in order to obtain an accurate interface without volume
loss/gain. This additional equation of level set method has to be solved at
each time step after the level set function is advected by the flow field. The
accurate solving of the redistancing equation is very important step because
the shape of a deformed bubble depends heavily on how to solve the redistanc-
ing equation. Since the redistancing algorithm has to solve a hyperbolic type
of equation, the typical Galerkin finite element method can not be used for
the solution of the redistancing equation. Therefore, a level set method that
can be used with an unstructured mesh needs to be newly developed for the
level set simulation of a bubble-in-liquid flow in a complex geometry such as a
MEMS device. Most of reported numerical studies of a level set method have
been performed using a finite difference or volume method with a structured
background mesh except the works of Pillapakkam and Singh[4] and Lin et
al.[5]. Pillapakkam and Singh firstly adopted the finite element discretization
for the level set simulation of a bubble-in-liquid flow on unstructured mesh.
However, they did not introduce a stabilization technique for the discretiza-
tion of the redistancing equation of hyperbolic type, while the characteristic
Galerkin method has been adopted for the hyperbolic type equations in the
study of Lin et al. In the present finite element simulation of a two-phase flow,
SUPG(Streamline Upwind Petrov-Galerkin), CSUPG(Consistent Streamline
Upwind Petrov-Galerkin) and Taylor-Galerkin method have been tested for
the stabilized solution of the redistancing equation. It has been shown that
the Taylor-Galerkin method gives a satisfactory solution. Furthermore, the
algorithms proposed in the present study for the redistancing and advection
equations are combined with the P1P1 finite element formulation[6] of the in-
compressible Navier-Stokes equations in order to solve some benchmark prob-
lems. In the following sections, the numerical method of the present level set
simulation based on finite element method is described in detail and then some
results of benchmark problems are given for the verification of the proposed
numerical method.
2 Numerical methods
2.1 Level set based finite element formulation of the

incompressible Navier-Stokes equations
Given the initial distribution of the level set function and velocity field, the
following incompressible Navier-Stokes equations are to be solved to obtain
an updated velocity and pressure field using the P1P1 splitting finite element
formulation:
Du
ρ(φ) = ∇p + ∇ · (2µ(φ)D) − σκ(φ)δ(φ)∇φ + ρ(φ)g (1)
Dt
Level Set based FEM of Bubble-in-Liquid Simulation 591
where, φ is the level set function and fluid density and viscosity, curvature
of an interface and is represented by a function of φ. The surface tension
effect has been evaluated by using the CSF(continuum surface force) method
proposed by Brackbill et al.
2.2 Finite element formulation of the advection equation
After solving the incompressible Navier-Stokes equations based on level set

formulation in section 2.1, the following advection equation of the level set
function should be solved at each time step so that the bubble/liquid interface
can be updated using the velocity field obtained from the Eq. (1).
φt + (u · ∇)φ = 0 (2)
For the discretization of the Eq. (2), the Taylor-Galerkin method has been
adopted.
2.3 Finite element formulation of the redistancing equation
The inaccurate estimation of the distribution of φ results in the incorrect

evaluation of fluid property and thus errors such as a volume gain/loss and
an inaccurate calculation of surface tension occur in the incompressible level
set simulation. Therefore, it is essential to solve the following equation such
that φ represents the shortest distance to the nearest interface.
φt + uc · ∇φ = S(φ0 )
uc = S(φ0 )(∇φ/|∇φ|) (3)
Since the Eq. (3) is a hyperbolic type, SUPG(streamline upwind Petrov-
Galerkin), CSUPG (consistent streamline upwind Petrov-Galerkin) method
and the Taylor-Galerkin method have been tested for the discretization of the
Eq. (3). The CSUPG formulation of the Eq. (3) can be written as follows:
For all admissible weighting functions w ∈ V , find φ ∈ H1 (Ω) that satisfies
the following equation.
Z

(w + p̃) φt + uc · ∇φ − S(φ0 ) dΩ = 0 (4)
where, p̃ is the perturbation weight function. In the Taylor-Galerkin method,

the following equation is derived applying the Taylor series expansion to the
Eq. (3).
h i ∆t2
φn+1 = φn + ∆tS(φ0 ) − ∆t un+1 c · ∇φ n+1
+ (5)
2
Eq. (5) is to be discretized using the Galerkin method and formulas proposed
by Lin et al.
2.4 Overall procedure for the solution of a bubble-in-liquid

simulation
The numerical procedure for the solution of a bubble-in-liquid simulation can

be summarized as follows:
(a) Given the level set function and velocity field, the incompressible Navier-
Stokes equations are to be solved in order to update the velocity field.
(b) Interfaces are updated by solving the advection Eq. (2) using the velocity
field obtained at step (a).
(c) Redistancing Eq. (5) must be solved so that that φ represents the shortest
distance to the nearest interface.
(d) Go to the next time step and then perform the steps (a) (c).
3 Results
3.1 Tank sloshing problem
Tank sloshing problem at a low filling level has been solved using the present
level set FEM and compared with the existing results. Fig. 1 shows the pres-
sure history at a point of the bottom of the tank with time. The present result
agrees well with the experiment.
3.2 Bubbles rising in a liquid
For the validation of the present FEM code for the bubble-in-liquid simulation,
a rising bubble problem in a quiescent liquid has been solved. The dimension-
less parameters introduced in the present simulation are density
√ ratio(ρg /ρl =
0.001), viscosity ratio(µg /µl = 0.01), Reynolds number(ρl R (gR)/µl ) and
Bond number(ρl R2 /σ). Non-dimensionalized terminal velocities calculated in
Fig. 1. Pressure history at a point of the bottom of the tank

Level Set based FEM of Bubble-in-Liquid Simulation 593
1.4
1.2
U
1
Case 2
0.8
0.6
0.4
0.2
0
2 4
t
Fig. 2. Bubble rising velocity with time for the case 2
the present simulation for various Reynolds and Bond numbers are compared
with those obtained by Ryskin and Leal[2], who used a body-fitted coordinate
method in order to find a shape at the equilibrium state from the stress bal-
ance. Their calculation can be considered to be the most accurate since the
curvature along the interface can be estimated most accurately by the body-
fitted coordinate. The results by Son[3] have been obtained by combining the
level set method with VOF. Fig. 2 shows the time history of the terminal
velocity for the case II. It is shown that the simulation reaches a terminal ve-
locity as time goes on. It needs to be noted that the present bubble-in-liquid
simulation method based on FEM and level set method alone gives a satisfac-
tory solution, particularly in terms of mass conservation. It should be noted
that Sussman and Puckett[7] and Son have used a CLSVOF(combined level
set set VOF) in order to circumvent volume conservation problem of level set
method based on finite difference method.
4 Conclusion
(1) The most satisfactory result has been obtained with the Taylor-Galerkin
discretization of the advection and redistancing equations compared to the
SUPG type discretization.
(2) The P1P1 splitting finite element method combined with the Taylor-
Galerkin discretization of the hyperbolic type equations solves some bench-
mark problems well and it is directly applicable to a problem with unstruc-
tured mesh.
(3) The present FEM code based on level set method gives a reasonable so-
lution for the benchmark problem of a rising bubble in a quiescent liquid
compared with the solution of Ryskin and Leal, who used a body-fitted coor-
dinate method.
References
1. Sussman, M., Smereka, P. and Osher, S.: A Level Set Approach for Computing
Solutions to Incompressible Two-Phase Flow. Comput. J. Comput. Phys.,114,
146–159 1994
2. Ryskin, G. and Leal, L.G.: Numerical solution of free-boundary problems in fluid
mechanics. J. Fluid. Mech., 148, 1–17 1984
3. Son, G.: Efficient implementation of a coupled level-set and volume-of-fluid
method for three dimensional incompressible two-phase flows. Numer. Heat.
Transfer. Part B., 43, 549–565 2003
4. Pillapakkam, S.B. and Singh P.: A level-set method for computing solutions to
viscoelastic two-phase flow. J. Comput. Phys., 174, 552–578 (2001)
5. Lin, C., Lee, H., Lee, T. and Weber, L.J.: A level set characteristic Galerkin nite
element method for free surface flows. Int J. Numer. Meth. Fluids, 49, 521–547
(2005)
6. Sung, J.Y., Choi, H.G. and Yoo, J.Y.: Time-accurate computation of unsteady
free surface flows using an ALE segregated equal-order FEM. Comput. Methods.
Appl. Engrg., 190, 1425–1440 (2000)
7. Sussman, M. and Puckett, E.G.: A coupled level set and volume-of-fluid method
for computing 3D and axisymmetric incompressible two-phase flows. J. Comput.
Phys., 162, 301–337 (2000)
Level Set method for Curvature-driven Flows
in Microfluidics
Paul Vigneaux
MAB, Université Bordeaux 1 - CNRS and INRIA Futurs - 351 Cours de la

Libération, 33405 Talence Cedex, France Paul.Vigneaux@math.u-bordeaux1.fr
Summary. A Level Set method for bifluid flows in Microfluidics is presented. After
a stability analysis which conducts to a restrictive stability condition on the compu-
tational time step due to surface tension, we introduce a splitting to shorten simula-
tion times. Numerical results of these curvature-driven flows are in good agreement
with physical experiments and are used to explore mixing processes inside micro-
droplets.
Key words: level set, microfluidics, surface tension
1 Introduction
Over the last decade, manipulations of tiny volumes of fluid have dramati-
cally improved. For instance, microchannels, with a section in the order of 100
square micrometers, are now casted in polymer or etched in silicium. Volume
of fluids are transported inside these channels. This miniaturization allows to
increase experiments in field such as biology, chemistry or pharmacology (viz.
genome sequencing, pharmacological molecules targeting). The Laboratory of
the Future (LOF) - a joint laboratory between Rhodia, Bordeaux 1 Univer-
sity and the CNRS (France) - develops experiments in microchannels in order
to analyse flows of polymers, concentrated surface-actives or emulsions. LOF
also pay a lot of attention to the mixing of these complex fluids. As far as the
Applied Mathematics Laboratory of Bordeaux (MAB) is concerned, we are in-
terested in the numerical dimension of these experiments through simulation.
This allows us in turn to compare the results with the physical studies of LOF.
In the following, we present the physical framework which inspired our

study, the mathematical model and numerical approach chosen to make the
simulations. The stability analysis of the model leads to restrictive time steps.
We thus propose a splitting method to shorten simulation times. Finally, we
present numerical results of droplets dynamics in microchannels.

596 Paul Vigneaux
2 Physical Problem
We are concerned by the simulation of immiscible bifluid flows in microchan-
nels. In such systems, the Reynolds number is lower than one. As a con-
sequence, we can use a quasi-stationary Stokes equation to model the flow.
Furthermore, we can assume that fluids are incompressible. The fact that the
Reynolds number is low could let us to believe that flows are laminar and that
the shape of the structures in the flow (e.g. droplets) is a simple one ; indeed,
this is confirmed by LOF experiments. Nevertheless, these flows are such that
surface tension forces are much stronger than inertia forces which can induce
vortices inside the structures. In addition, the geometries of our applications
are such that we have to deal with moving interfaces with topological changes.
By the way, we want to use a Level Set method [1] [3] which, in this case, is
particularly suited to easily follow interfaces.
3 Numerical Approach
Level Set formulation
It is the Eulerian nature of the Level Set method which allows to easily handle
topological changes of the interface. This method has been introduced by
Osher and Sethian in [1], for fronts propagating with curvature dependent
speed. It was later used by Sussman et al. [2] for incompressible two-phase
flows. Let us recall that the Level Set method consists in representing the
interface, Γ ⊂ Rd , as the zero level set of a continuous level set function,
φ : Ω ⊂ Rd → R. The function φ is defined everywhere in the domain Ω
whereas Γ = {x ∈ Ω / φ(x) = 0}.
The motion of the interface is obtained by solving the transport equation with
velocity field v :
φt + v.∇φ = 0 (1)
Actually, v is the desired velocity of the underlying problem on the interface
and is arbitrary elsewhere. However, it was showed that a smooth extension
of the interface velocity off the front gives better results.
Following the approach of [2], and adapting it to our microchannels prob-

lem, we use the Stokes equations as flow equations to obtain the velocity field
which convects the level set function. We consider a bifluid model with one
fluid of variable viscosity. This leads to the following model :
div(2ηDu) − ∇p = σκδ(φ)n (2)
div(u) = 0 (3)
φt + u.∇φ = 0 (4)
Level Set method for Curvature-driven Flows in Microfluidics 597
where Du = (∇u + ∇T u)/2, η is the variable viscosity such that:

η1 in fluid 1
η= (5)
η2 in fluid 2
σ is the surface tension coefficient, n is the normal to the interface, κ is
the curvature of the interface and δ(φ) is the Dirac function which allows to
localise the interface since the level set function φ is defined as the signed
distance function to the interface :

 > 0 in fluid 1
φ = 0 on the interface (6)
< 0 in fluid 2

The normal and the curvature are simply computed thanks to the level set
function :
∇φ
n= and κ = div(n) (7)
|∇φ|
It is also known, that one has to periodically reinitialize φ to a signed
distance function since, along the computation, if v is not the extension of
the interface velocity, φ does not remain a distance function. This can leads
to severe loss of mass. One way to reinitialize a level set function φ0 to a
signed distance function is to solve the so-called reinitialization equation :
φt + sign(φ0 )(|∇φ| − 1) = 0 with initial condition φ(., t = 0) = φ0 . Since
all the information propagates from the interface, no boundary condition is
needed. Level Set method well lends itself to Microfluidics because interfaces
have a near circular shape which sprawls all over the width of the channel.
By the way, reinitialization and mass correction with translation of φ can be
legitimately used and achieve good mass conservation and accurate curvature
computation.
Geometry and Boundary Conditions
All our numerical simulations are realized in the 2D case. The domain geom-
etry is composed of channels which cross each other with injections of various
fluids at various extremities, or exits of fluids, depending on the physical ex-
periments. (see Figure 1)
In addition, we must impose sliping boundary conditions (BC) on the walls
of the channels :
u.τ = αus (η) + βLs (η) ∂(u.τ )

∂nw (8)
u.nw = 0
where nw and τ are respectively the normal and the tangent to the boundary,
us is the slip velocity and Ls is the so called “slip length” coefficient both
depending a priori on η since sliping is different from one fluid to another. We
suggest the hybrid condition (8) to take into account whether the interface
touches the wall (α 6= 0 and β = 0 : non homogeneous Dirichlet BC) or not
(α = 0 and β 6= 0 : Navier BC). Classical boundary conditions are used at
the inlets and outlets.
598 Paul Vigneaux
Discretization, Solvers and Stability condition
In order to solve the Stokes equation (2), we use a finite volume discretization
on staggered cartesian grid and an Augmented Lagrangian method to take
into account the incompressibility constraint (3). If there are solid obstacles
in the flow, we use a penalization method of the obstacles [4] in order to keep
the cartesian grid. The transport equation (4) is solved through the fifth order
WENO schemes [5].
The time step ∆t must obey stability conditions due to convective and
surface tension terms. The time step associated to the convection is classicaly
∆tc = ∆x/max(v). Using an approach that allows to retrieve the stability
condition of Brackbill et al. [6] for Navier-Stokes equation, we find that the
stability constraint associated to the surface tension term in Stokes equation
is ∆tσ = ση ∆x. For our microfluidic applications, ση ∼ 1 and max(v) ∼ 10−2
which means that ∆tσ is 100 times smaller than ∆tc . Thus, since the smaller
time step is taken, the stability constraint is very restrictive. In the following,
we propose an approach to relax this stability constraint.
Relaxing the stability constraint
We observe that in microfluidics applications, the structures can have a sta-

tionary shape and their motion is essentially a translation motion (e.g. in a
straight channel). By the way, in a such situation, one does not want to be
restricted by ∆tσ . One idea consists in working in the droplet’s frame of refer-
ence using a splitting and translate the droplet fastly with the CFL condition.
The procedure is as follows :
1. Find the speed of translation of the droplet and the time step ∆tinj as-
sociated to the injection
2. Compute the shape correction of the interface with an iterative step for
which the total time is less or equal to ∆tinj with a combination of ∆tσ
time steps. The stationary shape is then obtained
3. Compute the translation of the interface by solving the transport equa-
tion over the time step ∆tinj and the resulting speed induced by speeds
computed in (2)
4. Loop (2)-(3)
4 Evolving Interfaces in Microfluidics

We present various numerical simulations which show the ability of the model
to simulate moving interfaces in microfluidics. First we consider the case of
a small unconfined droplet in a microchannel. The injection velocity is uinj
= 4 cm/s. The viscosity of the droplet is η1 = 10−3 Pa.s and the viscosity
of the continuous phase is η2 = 2.10−2 Pa.s. The surface tension coefficient σ
= 3.10−2 N/m. Figure 2 shows the total velocity field and two curves : the
Level Set method for Curvature-driven Flows in Microfluidics 599
black one is the interface of the droplet which is, in this case, stationary, the
red curve is a perfect circle superimposed on the droplet in order to show
the difference with the asymptotic shape obtained when there is no injection
(static case). Figure 3 shows, for the same simulation, the velocity field in
the droplet’s frame of reference. We can see that the asymptotic shape is
obtained since velocity is tangent at the interface, thus there is no deformation
of the droplet. A bigger droplet, with the same physical properties, is shown
in Figure 4 with the velocity field in the droplet’s frame of reference : we
observe additional vortices compared to the small droplet. These simulations
allows to study the mixing inside droplets for different flow conditions. Figure
5 shows a train of droplets which flow upward and exit the two horizontal
channels. In Figure 6, we see two droplets, arriving by the two horizontal
channels, colliding and flowing downward. These two cases of figures 5 and 6
shows the ability of the code to handle topological changes of the interface.
5 Conclusion
We have presented a Level Set method for the simulation of flows in Microflu-
idics, which appears to be well adapted to deal with such curvature-driven
flows. A stability analysis shows that surface tension induces a restrictive
condition on the computational time step. We thus introduced a splitting
method to relax this constraint and shorten simulation times. Finally we give
numerical results which show the ability of the method to handle moving in-
terfaces in Microfluidics ; the agreement with physical experiments leads us
to use this code with our collegues to explore the mixing inside microdroplets.
References
1. Osher, S., Sethian, J.: Fronts propagating with curvature-dependent speed: al-
gorithms based on Hamilton-Jacobi formulations. J. Comput. Phys., 79, 12–49
(1988)
2. Sussman, M., Smereka, P., Osher, S.: A level set approach for computing solu-
tions to incompressible two-phase flow. J. Comput. Phys., 114, 146–159 (1994)
3. Osher, S., Fedkiw, R.: Level Set Methods and Dynamic Implicit Interfaces.
Springer, New York (2003)
4. Angot, P., Bruneau, Ch.-H., Fabrie, P.: A penalization method to take into
account obstacles in incompressible viscous flows. Numer. Math., 81, 497–520
(1999)
5. Jiang, G-S., Peng, D.: Weighted ENO schemes for Hamilton-Jacobi equations.
SIAM J. Sci. Comput., 21, 2126–2143 (2000)
6. Brackbill, J. U., Kothe, D. B., Zemach, C.: A continuum method for modeling
surface tension. J. Comput. Phys., 100, 335–354 (1992)
600 Paul Vigneaux
inlet
u u
wall outlet
Fig. 1. Geometry Fig. 2. Global flow in a small droplet
Fig. 3. Flow in drop’s frame of reference Fig. 4. Confined droplet
Fig. 5. Train of droplets Fig. 6. Collision of droplets

Mesh-Based Microstructure Representation
Algorithm for Simulating Pore-scale Transport
Phenomena in Porous Media
May-Fun Liou1 and Issac Greber2

1
NASA Glenn Research Center, Cleveland, Ohio 44135, USA,
May-Fun.Liou@nasa.gov
2
Case Western Reserve University, Cleveland, Ohio 44106, USA
1 Mesh-Based Microstructure Representation Algorithm
In the conventional CFD approach, a prescribed physical domain of porous

medium with existing fluid-solid interface is mapped onto a computational do-
main, which is then geometrically discretized by mesh generation. This would
be formidable task if hundreds or thousands of pores of irregular shapes are
to be mapped. In the new approach developed here, the geometry discretiza-
tion is made easy by taking the reverse sequence of conventional steps. First,
the outline of the computational domain is selected to represent the physical
domain of interest. Next, the computational domain is discretized with a fine
enough mesh, which is generally an easy task since there are no convoluted
interfaces to conform to. Then mesh locations are chosen as seed locations of
solid material; these locations are chosen randomly, but with prescribed rules.
The solid regions are then allowed to grow, again randomly but in accordance
with prescribed rules. The rules are established so that the eventual statistical
properties of the porous region matches closely the geometric properties (such
as the porosity and the pore size) of the physical porous medium that it is
sought to simulate. The solid and fluid-filled pore regions are then assigned
the thermodynamic and transport properties of the corresponding physical
materials. As a natural byproduct of this process, the interface between solid
and pore regions always lies at the boundary between mesh cells. Thus it re-
sults in a complete conforming (body-fitted) at the solid and fluid interfaces.
Another advantage over the conventional approach is that fluid and solid ma-
trix interaction (e.g. surface reaction) can be accommodated easily since the
mesh for the solid matrix is also available. Because the mesh cells act as a
bed or base from which the porous medium is constructed, this approach is
termed a mesh-based microstructure representation algorithm (MBMRA).

602 May-Fun Liou and Issac Greber
1.1 Methodology
In general guidelines, this methodology entails three steps, namely generat-

ing the mesh, selecting seeds, and growing the solid matrix from the seeds. A
mesh generator is used to generate a bed of unstructured or structured meshes
in the computational domain. This mesh bed is used both for generating the
solid matrix and for performing the numerical computations. The mesh used
herein is triangular and quadrilateral for two-dimensional and tetrahedral and
hexahedral for three dimensional geometries. Seeds, which are initially solid
elements, are selected within the mesh bed. The seed locations are chosen
using a random number generator that has been well validated for its uni-
formity of randomness [1]. These seed cells are grown spatially into a desired
geometry, for example a fiber-like or sphere-like solid matrix as used in the
current computations. The processes of seeding and growing are iterated until
the desired pore quality and porosity are achieved.
1.2 Porosity and porous samples
Porosity is defined as the ratio of the volume of the total pore space filled by
the fluid to the total volume of the porous medium, all within the domain of
interest. Since mesh cells in the domain have been tagged as solid or fluid, the
volume of total pore space can easily be obtained numerically by finding the
sum of volumes of the cells tagged as fluid. Similarly, the volume of the solid
matrix is the sum of volumes of the cells tagged as solid. Finally, the total
medium volume is simply the sum of the volumes of all cells. Figures 1.a − 1.c
are three representative porous media obtained numerically. These three plots
are used to demonstrate that various solid microstructures can be obtained
from triangular, rectangular, and hybrid meshes respectively. They exhibit
different characteristics determined by inputs provided in the processes of
seeding and growing inside unstructured mesh beds. The blue and red colors
represent fluid and solid elements, respectively.
Fig. 1. Randomly formed solid matrix representing porous medium in 2-D domain
(a) fibrous shape formed with triangular mesh; (b) small circular shape solids; (c)
large solids in circular shape with hybrid mesh.
Mesh-Based Microstructure Representation Algorithm 603
1.3 Numerical Implementation
With the discretized geometry established, appropriate computer codes, for

example, Navier-Stokes equation solvers, or molecular dynamics codes can
all be used to produce the appropriated level of numerical accuracy and ef-
ficiency. An existing three-dimensional Navier-Stokes code (National Com-
bustion Code [2]) ,in conjunction with the heat conduction equation of solid
matter, is adopted to incorporate the methodology described in the above
sections. Note here that the solid matrix is assumed to be isotropic and ho-
mogeneous. That is, the solid matrix has the same mechanical and thermal-
physical properties everywhere, and they are independent of orientation. The
matching conditions at the fluid-solid interfaces are zero fluid velocity and
continuity of temperature and heat transfer rate.
The studied case is a 2D channel flow with porous sample located in the
middle of the channel. For numerical stability, a long entrance length is pro-
vided before flow reaches the porous section, shown in Fig. 1.c.
2 Sensitivity test
2.1 Random number distribution
Though an exact representation of materials is not the goal for MBMBA,

the resulting porous medium has to be statistically meaningful representa-
tion of it. The first sensitivity test is conducted by varying the initial value
in the working area of 624 words in adopted random number generator. A
schematic results of three initial seeds which result in three random number
distributions are displayed in Fig. 2. Corresponding to each seed, the result-
ing pressure drops across the porous media, obtained by MBMRA are plotted
on the right, where the pressure drop is non-dimensionalized by the inflow
kinetic energy. Through the good agreements in pressure drop, a macroscopic
property, Figure 3 indicates the non-dimensional axial velocity profiles vary
with random number distribution. Figure 3 shows the flow in the interstices
and the zero velocities in the solid. It also presents the non-uniform profiles
Fig. 2. Sensitivity test of pressure drop versus initial seed used in random number
generator.
Fig. 3. Heterogeneity of porous structures ( = 0.83) is displayed in terms of the

axial velocity profiles at various locations for studied random number distributions.
at exit which are shown in black line with square symbol; the effect from the
porous micro-structure persists.
2.2 Geometric types
The solid structure is approximated by two types of geometric representations,
sphere-like and fiber-like. Table 1 lists structures of different porosities. In the
Table 1. Morphologic characteristics of porous samples used

Shape porosity d(mm) No. of solid elements
sphere-like 0.85 3 650
fibrous 0.83 0.75 1705
case of sphere-like porous medium, d is the radius of the sphere, while d is the
thickness of fiber for fibrous solid phases. The fibrous structure is composed
of slender and long fiber in sheets in which the form drag is much smaller
than the circular cylinder tubes structure. Figure 4 shows the effects of solid
shape and permeability on the sphere-like and fibrous structures listed above.
2
The non-dimensional pressure gradient parameter is defined by P ∗ = L∆pµ dV ,
Fig. 4. Sensitivity test of pressure drop versus Reynolds number in various types
of solid structure.
Mesh-Based Microstructure Representation Algorithm 605
where Red is the Reynolds number based on the pore scale ρ Vµ d . It is evident
that the sphere-like structure causes higher pressure drop than the fibrous
structure.
2.3 Grid Independence Study
It is expected that the more the mesh is refined, the more closely the micro-
scale can be simulated. In the vicinity of the solid-fluid interface, there exist
large velocity and temperature gradients, and an adequate number of mesh
points is necessary to accurately resolve these variations. Also, the use of fine
grids in the flow direction may be critical when the effects of axial conduc-
tion and convection are of concern. However, the desire for finer meshes is
limited by available computing resources. Hence, grid-independence tests are
conducted before this methodology could be used for further computations.
These tests were done by varying mesh sizes while holding porosity fixed.
3 Numerical example - Forced convection with heated

walls
A two dimensional channel case, which is confined between two heated par-
allel plates, is used as a numerical example. The same geometry but without
porous sample situated inside the channel is also computed with the same
boundary conditions for comparison. It is referred to as the baseline case. In
this study, the velocity, temperature profiles and Nusselt number are deter-
mined for flow through the channel with a porous sample of Silicon Nitride.
The dimensionless temperature profile, T ∗ versus y ∗ , at various locations are
plotted in Fig. 5.a, and figure 5.b is for the baseline case. It is clear that the
thermal boundary layer gets thinner due to porous sample. Figure 6.a plots
the axial distribution of Nusselt number, N ux , along the bottom wall starting
at the inlet of the channel (x/L = 0) . Also, the enhancement of heat transfer
by the porous sample of Silicon Nitride is displayed. These numerical solu-
tions are obtained only when governing equations of both fluid and solid are
converged; the convergence history is presented in Fig. 6.b.
4 Conclusions
A robust modeling technique for porous medium transport has been proposed
for detailed quantification of how microscopic properties affect macroscopic
transport phenomena. The pore density is currently controlled by random
number generator; it can be extended to conform specified geometries. To meet
the requirement of sufficient resolution around solid structures, the adaptive
meshing technique can be incorporated to locally refine meshes in regions
that require it. Replacement of the non-slip condition with a Maxwellian slip
condition at the interface or wall may be added if the local characteristic
length (finest mesh scale) is comparable to the mean free path of the fluid [3].
Fig. 5. Non-dimensional temperature profiles in a channel flow heated from bottom

wall; (a) at locations of upstream, within and downstream from the Silicon Nitride
Fig. 5. Non-dimensional temperature profiles in a channel flow heated from bottom
porous sample; (b) baseline case at the same locations.
wall; (a) at locations of upstream, within and downstream from the Silicon Nitride
porous sample; (b) baseline case at the same locations.
Fig. 6. (a) Nusselt number distribution along the axial direction; (b) Convergence
history of residuals from solving Navier-Stokes equations for fluid and solving heat
conduction equations for solid.
Fig. 6. (a) Nusselt number distribution along the axial direction; (b) Convergence
References
history of residuals from solving Navier-Stokes equations for fluid and solving heat
conduction equations for solid.
1. Matsumoto, M. and Kurita, Y., ”Twisted GFSR generators II,” ACM Trans. on
Modeling and Computer Simulation,” 4, 254–266, (1994).
2. Chen, K.-H., Norris, A. T., Quealy A., and Liu, N.-S., ” Benchmark Test Cases
References
for the National Combustion Code,” AIAA Paper 98-3855, (1998).
3. Barber, R. W. and Emerson, D. R., ”The Influence of Knudsen Number on
[1] Matsumoto, M. andDevelopment
the Hydrodynamic Kurita, Y., ”Twisted GFSR Parallel
Length within generators II,”Micro-Channels,”
Plate ACM Trans. on
Modeling and Computer Simulation,” 4, 254–266, (1994).
Advances on Fluid Mechanics IV, Rahman et. al. (Eds.), WIT press, (2002).
[2] Chen, K.-H., Norris, A. T., Quealy A., and Liu, N.-S., ” Benchmark Test Cases
for the National Combustion Code,” AIAA Paper 98-3855, (1998).
[3] Barber, R. W. and Emerson, D. R., ”The Influence of Knudsen Number on
the Hydrodynamic Development Length within Parallel Plate Micro-Channels,”
Advances on Fluid Mechanics IV, Rahman et. al. (Eds.), WIT press, (2002).
The Numerical Simulation of Liquid Sloshing
in Microgravity
Roel Luppes1 , Joop A. Helder and Arthur E.P. Veldman
University of Groningen, IWI, P.O. Box 800, 9700 AV Groningen, The Netherlands
r.luppes@math.rug.nl
Summary. The influence of sloshing liquid on board spacecraft is investigated.

Experiments have been carried out with the satellite Sloshsat FLEVO, containing
a partially filled water tank. The experiments were supported by a computational
model for 3D incompressible free-surface flow, including capillary surface physics
and coupled solid-liquid interaction dynamics. Experiments and numerical simula-
tions are compared. The obtained frequencies in angular velocities are reasonably
comparable at various satellite manoeuvres. The damping of nutation amplitudes in
the simulations is too large, as a result of additional diffusion in the computational
model. At low rotational rates and small-scale liquid motion, capillary effects are
important for the damping of manoeuvre-induced oscillations.
Key words: Free-Surface Flows, Microgravity, Solid-Liquid Interaction, Sloshsat

FLEVO, Numerical Simulations
1 Introduction
With the increasing amount of liquid on board spacecraft, liquid management

and its influence on the overall spacecraft dynamics is becoming increasingly
important. The influence of sloshing liquid may hamper critical manoeuvres
in space, such as the docking of liquid-cargo vehicles or the pointing of obser-
vational satellites. Severe problems with sloshing liquid in space (propellant
slosh) have been reported [1, 2, 3].
Experiments have been carried out with the satellite ’Sloshsat FLEVO’ in
an orbit around earth to investigate the interaction between sloshing liquid
and spacecraft dynamics [4]. A cylindrical shaped fluid tank with hemispher-
ical ends, with a volume of 86.9 litres, is positioned inside the satellite. This
tank is partly filled with 33.5 kg distilled water, which could freely slosh during
the experiments. The experiments were controlled by means of 12 thrusters
that were fed with N2 gas. During the experiments, several relevant quanti-
ties for the Sloshsat motion were measured, such as the angular velocity, the
linear acceleration and the pressure inside the fuel tank. A description of the

608 R. Luppes, J.A. Helder and A.E.P. Veldman
Fig. 1. The Sloshsat FLEVO satellite with a schematic view inside.
technological equipment on board as well as various other Sloshsat data is

given in [5, 6, 7, 8]. A schematic view inside Sloshsat is depicted in Fig. 1.
The experiments were supported by a computational model based on the
Navier-Stokes equations for 3D incompressible free-surface flow. This model
includes capillary surface physics and coupled solid-liquid interaction dynam-
ics. By means of such a model, more insight can be gained in the relevant
physics that plays a part in the interaction between liquid sloshing and over-
all spacecraft dynamics. In the absence of gravity, capillary effects at the free
liquid surface, such as surface tension and wetting characteristics, are domi-
nating liquid motion [9]. The present CFD code Comflo is the successor of the
model that was used to support experiments on board Spacelab [10]. Comflo
is also used for maritime, industrial and offshore flow applications [11].
2 The Mathematical Model
The description of the interaction between the sloshing fluid inside Sloshsat
and the motion of Sloshsat itself requires a coupling between the mathemat-
ical models for fluid dynamics and solid-body dynamics. These models are
extensively described in [12, 13, 14].
The fluid flow inside Sloshsat can be described by means of the Navier-
Stokes equations for incompressible flow in terms of the fluid velocity u, the
pressure p and the fluid density ρ and viscosity µ. The influence of the solid-
body motion on the fluid is described through a virtual body force, which
depends on the angular velocity ω and acceleration ω̇ and the linear acceler-
ation q̇ of Sloshsat. These quantities follow from the equations for solid-body
dynamics, describing conservation of linear and angular momentum of ’dry’
Sloshsat. The equations contain additional terms representing the force and
torque that the fluid via pressure and viscous effects exerts on the tank wall,
as well as an additional external force describing the thruster induced forces
on the satellite. The equations for solid-body dynamics are rewritten by using
the total (liquid+solid) mass and moment of inertia of the coupled system. In
this way, the system can be solved iteratively in a stable manner [9, 12, 13].
The Numerical Simulation of Liquid Sloshing in Microgravity 609
continuity of
θ normal and
static tangential
contact angle stress
conservation
no slip of mass and
momentum
Fig. 2. Schematic overview of the liquid model. Dark and light shading represent
solid body and liquid, respectively.
In Fig. 2, a schematic overview of the boundary conditions is given. At the

tank wall, u = 0 is prescribed (no-slip). At the free surface, continuity of normal
and tangential stress is prescribed in terms of velocity at the free surface,
ambient pressure of air, surface tension σ of the fluid and mean curvature κ
of the free surface. For κ, spatial derivatives of the free-surface S(x, y, z, t) = 0
have to be computed and hence boundary conditions are needed at the contact
line. In this study, a static contact angle θ is prescribed, which is defined as the
angle between the normals on the solid boundary and the free surface. Hence,
the free surface intersects the solid boundary at a fixed angle. The value of
θ depends on the material properties of liquid, air and the solid boundary of
the tank inside Sloshsat and is taken 90o in this study. In reality, the contact
line probably sticks to the tank wall and therefore the contact angle may not
be constant. As an alternative, a dynamic contact angle in combination with
contact-angle hysteresis may be applied [15].
3 The Numerical Model
For the spatial derivatives of the flow model, finite-volume discretisation is ap-
plied, making use of geometry apertures describing the parts of the cell faces
that are open for flow [11, 12, 14]. To enhance the stability of the numeri-
cal scheme, interpolations are chosen carefully to ensure that the discretised
equations have the same symmetry properties as the continuous equations
[16]. For the temporal derivatives in the flow model the forward Euler method
is used, leading to a Poisson equation for the pressure that is solved iteratively
using SOR with automatically adjusted relaxation for optimal convergence be-
haviour [17]. The temporal derivatives of the solid-body model are integrated
in time using a 4th order Runge-Kutta method. This leads to a linear system
of 6 six unknowns (3 components of both q̇ and ω̇), which is solved by means
of Gaussian elimination during each Runge-Kutta step [12, 13].
When the pressure at a time level is determined, the velocity field is easily
obtained and the free surface can be displaced. This is done using an adapted
version of the volume-of-fluid (VOF) method introduced in [18]. An important
drawback of the original method is the creation of unrealistic small drops that
get separated from the bulk fluid. Moreover, free-surface apertures may be
outside their natural range. Therefore, a correction is used in terms of a local
height function that is based on the orientation of the free surface [12, 11]. An
alternative method, which is however far more difficult to implement, has been
proposed in [19]. Both methods have been compared in [11], leading to the
conclusion that for certain flow applications the second method yields slightly
better results. However, for the Sloshsat simulations in the present study, the
VOF method with local height function leads to satisfactory results.
The simulation of a Sloshsat manoeuvre is started with a lasting period
of constant rotation ω around the stable axis of maximum moment of inertia
(MOI) and q = 0, without any thruster action. In this initial period, the water
inside the tank can settle towards a certain stable configuration. Only the
discretised flow model is solved, as ω and q are constant in this period. The
final state after the initial period is referred to as t = 0. At t = 0, the manoeuvre
is started and all discretised equations are solved in sequential order.
4 Experiments and Simulations
During a period of 8 days, several experiments have been carried out with
Sloshsat in an orbit around earth. These experiments were controlled by means
of 12 thrusters that could be fired with a frequency of 30Hz. The status
of each thruster (whether or not a thruster has been fired) was recorded.
The thruster forces that operate on Sloshsat can be determined subsequently
from the measured pressure inside the fuel tank. The various experiments and
measurement equipment are extensively described in [5, 6, 7].
In Fig. 3 (left column), measurements of a spin-up around the axis of maxi-
mum MOI are compared with the results of a simulation. The rotational veloc-
ity changes from ω =(-0.004;0.104;-0.007)T towards a rotation with ωy ≈ 0.179
and nutating ωx and ωz components. The increase of ωy is almost perfect.
However, the results of ωx and ωz show a clear difference. Although the fre-
quency of the nutation in ωx and ωz is comparable, in the simulation the
amplitudes are too small. A proper calibration of the measured data may
yield a favourable correction of the small measured values of ωx and ωz . The
decrease of the amplitude is too strong in the simulation, probably as a result
of the use of first-order spatial discretisation to enhance stability. This gen-
erally results in additional non-physical (numerical) damping, depending on
the magnitude of fluid velocity. Capillary forces are of less importance for the
observed damping, as ω is relatively large in this experiment.
The results for a nutation-avoidance manoeuvre (NAM) are depicted in
Fig. 3 (middle column), where the objective of the experiment is to move
ω without incurring too much nutation. Initially, the fluid is adapted to
ω =(0.003;-0.101;0.007)T . In the ωx component (top figure), a series of thruster
pulses can be distinguished at t = 0, 70 and 140s, resulting in a rapid change of
The Numerical Simulation of Liquid Sloshing in Microgravity 611
! x [ 1/s ] ! x [ 1/s ] ! x [ 1/s ]

+0.005 +0.030 +0.600
+0.000 +0.020 +0.400
+0.010 +0.200
-0.005 +0.000
-0.010 +0.000
-0.010 -0.200
-0.020
-0.015 -0.030 -0.400
-0.020 -0.040 -0.600
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
! y [ 1/s ] ! y [ 1/s ] ! y [ 1/s ]
+0.200 -0.085 +0.200
+0.180 -0.090 +0.000
+0.160
-0.095 -0.200
+0.140
+0.120 -0.100 -0.400
+0.100 -0.105 -0.600
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
! z [ 1/s ] ! z [ 1/s ] ! z [ 1/s ]
+0.000 +0.020 +0.600
-0.005 +0.010 +0.400
-0.010 +0.000 +0.200
+0.000
-0.015 -0.010 -0.200
-0.020 -0.020 -0.400
-0.025 -0.030 -0.600
0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700 0 100 200 300 400 500 600 700
t [s]
Fig. 3. The comparison t [s]
of numerical simulations with measurements t for
[ s ] a spin-up
experiment (left column), NAM experiment (middle column) and flat-spin exper-
iment (right column). For each experiment, the rotational velocity components ωx
(upper), ωy (middle) and ωz (lower) are depicted. The simulations are represented
by the darker lines.
ωx , followed by periods of free-tumble movement. The simulation is in good

agreement with the measurements of ωx and ωz , with comparable nutation
frequencies and amplitudes. After t = 300s, ωy starts to deviate. Since ω is
rather small and the NAM only induces small scale liquid motion, the level
of numerical diffusion is small, resulting in a comparable level of damping.
Improvement may be obtained by the use of a dynamic contact-angle model,
as capillary forces are rather important when ω is small.
The comparison for a flat-spin manoeuvre is given in Fig. 3 (right column).
Initially, the fluid is adapted to ω =(-0.005;0.165;-0.011)T . Next, during 14
seconds, Sloshsat is approximately rotating around the axis of intermediate
MOI, at which the fluid largely adjusts. The thruster action continues until
t = 33s. Then a free tumble of Sloshsat commences and ω slowly moves towards
the axis of maximum MOI with a lot of nutation and large-scale fluid action.
This makes the manoeuvre very suitable for validation of the numerical model.
The obtained components of ω of the simulation are in good agreement with
measurements, except for the high level of (numerical) damping, because of
the high velocity values of the sloshing liquid. The profile of ωy shows an
interesting phenomenon. Each nutation period contains a higher frequency, as
a result of sloshing water pounding the tank wall. This secondary frequency is
present in both experiment and simulation, indicating that in the simulation
the Sloshsat motion adopts the sloshing frequency correctly.
5 Conclusions
The influence of sloshing liquid on board spacecraft has been investigated by
carrying out experiments with the satellite ’Sloshsat FLEVO’. These experi-
ments were supported by a computational model based on the Navier-Stokes
equations for 3D incompressible free-surface flow with capillary surface physics
and coupled solid-liquid interaction dynamics.
Measurements of the rotational velocity ω of Sloshsat have been com-
pared with numerical simulations. The obtained frequencies are reasonably
comparable at various satellite manoeuvres. Hence, in the numerical model,
the sloshing frequency of the water inside the tank is adopted correctly by the
satellite motion. The nutation damping in the simulations is too large in case
of large rotational velocities and large-scale liquid motion, because of addi-
tional numerical diffusion in the computational model. The use of higher-order
discretisation to reduce the level of numerical diffusion is recommended.
At low rotational rates and small-scale liquid motion, capillary effects are
important for the damping of manoeuvre-induced oscillations. A dynamic
contact-angle model probably yields better agreement between simulations
and experiments in these cases. A proper calibration of the small components
of ω may yield a favourable correction of the measured values.
References
1. Strikwerda, T.E., Ray, J.C., Haley, D.R. and O’Shaughnessy, D.J. In Guidance
and Control 2001, volume 107, pages 597–614. Am. Astron. Soc., 2001.
2. http://near.jhuapl.edu.
3. http://www.history.nasa.gov/alsj/a11/a11.landing.html.
4. http://www.nlr.nl/public/publications/pdf/f183-02.pdf.
5. Vreeburg, J.P.B. In 56th Int. Astronautical Congr. IAF, 2005. IAF-05-C1.2.09.
6. Vreeburg, J.P.B. Technical Report CR 96755 L, NLR, 1996.
7. Vreeburg, J.P.B. In 50th Int. Astronautical Congr. IAF, 1999. IAF-99-J.2.04.
8. Prins, J.J.M. In 51st Int. Astronautical Congr. IAF, 2000. IAF-00-J.2.05.
9. Vreeburg, J.P.B. and Veldman, A.E.P. In R. Monti, editor, Physics of Fluids in
Microgravity, pages 293–321. Gordon and Breach Acad. Publ., 2002.
10. Veldman, A.E.P and Vogels, M.E.S. Acta Astronautica, 11:641–649, 1984.
11. Kleefsman, K.M.T., Fekken, G., Veldman, A.E.P., Iwanowski, B. and Buchner,
B. J. Comp. Phys., 206:363–393, 2005.
12. Gerrits, J. Dynamics of Liquid-Filled Spacecraft. PhD thesis, University of
Groningen, The Netherlands, 2001.
13. Gerrits, J. and Veldman, A.E.P. J. Eng. Math., 45:21–38, 2003.
14. Luppes, R., Helder, J.A. and Veldman, A.E.P. In 56th Int. Astronautical Congr.
IAF, 2005. IAF-05-A2.2.07.
15. Van Mourik, S., Veldman, A.E.P and Dreyer, M.E. Microgr. Sci. Techn., 17-
3:91–98, 2005.
16. Verstappen, R.W.C.P. and Veldman, A.E.P. J. Comp. Phys., 187:343–368, 2003.
17. Botta, E.F.F. and Ellenbroek, M.H.M. J. Comp. Phys., 60:119–134, 1985.
18. Hirt, C.R. and Nichols, B.D. J. Comp. Phys., 39:201–225, 1981.
19. Youngs, D.L. Technical Report AWRE/44/92/35, AWRE, 1987.
Numerical Prediction of Interfacial Instability
Robert Nourgaliev1 , Meng-Sing Liou2 , and Theo Theofanous3

1
Center for Risk Studies and Safety, UC Santa Barbara robert@engr.ucsb.edu
2
NASA Glenn Research Center, Cleveland meng-sing.liou@grc.nasa.gov
3
Center for Risk Studies and Safety, UC Santa Barbara theo@engr.ucsb.edu
1 Introduction
“Free”, fluid-fluid interfaces, in the presence of body forces and/or differential

velocities are subject to deformation and breakup - processes that are princi-
pally responsible for flow regime development, and thus for the macroscopic
features in all multi-fluid systems. Body forces normal to an interface can be
stabilizing or destabilizing, differential velocities parallel to an interface are
always destabilizing, and for development of instability such driving forces
must be sufficient to overcome the surface tension force (always stabilizing).
Under unstable conditions, early growth of an interfacial disturbance is ex-
ponential in time, and the theory for understanding this regime, based on
the linearized Navier-Stokes equations, rests on firm grounds. At amplitudes
that are a significant fraction of the wavelength, this theory breaks down,
non-linear analysis becomes scarcely feasible, and numerical simulation is the
key to further progress. In this paper we develop and demonstrate the first
reliable approach to this quest.
In contrast to all previous approaches4 , our method is focused on the sin-
gular nature of real fluid-fluid interfaces - boundaries of discontinuity in ma-
terial properties and in normal stresses. In particular, the solution is obtained
by coupling the two flows on either side through a comprehensive consider-
ation of the balance of stresses all along the interface. The alternative, an
approach that “imbeds” the boundary in the solution of bulk flows through a
narrow interfacial region with variable properties, and a localized body force
to represent interfacial tension - that is, a “smearing approximation” known
as the Diffuse Interface Method (DIM) [13, 22, 2] - has been shown to be,
when shear stresses are not negligible, physically inconsistent with the basic
instability character of such flows [20].
4
Some rare and tentative attempts for a sharp treatment are noted below. The
elegant and essentially exact work with the boundary integral method is limited
to the ideal limits of inviscid [5, 15], or Stokes [14, 16] flows.

614 Robert Nourgaliev, Meng-Sing Liou, and Theo Theofanous
Fig. 1. Spurious velocity and grid convergence. a) Spurious velocity for Rayleigh-
Taylor instability. b) Spurious velocity for a static drop. c) Grid convergence
of velocity, pressure and stresses for discontinuous flow test, defined as fol-
lows. A bubble of radius R = 0.4 is suspended in the center of the com-
putational domain h1×1. The i pressure
h and
−velocity
inside and outside of the
i
bubble are set to P, Vξ , Vη (r) = P0 + PI − P0 Rr 3 ; VξI Rr 2 ; VηI Rr 3 and

h − + + i
P − σ
R
R
r ; VξI 1 − 2 µ− r
R 1− r
R ; Vη 1 − 3 µ− 1 − r
R , respectively; where
I µ I µ
− µ+
P0 = 20, P
I
= 10, Vξ
I
= 1, Vη
I
= 1, σ = 1 and µ−
= 100. Pressure and velocity
fields are shown in d).
Concerns about the fidelity of DIMs have been raised previously by Kang
et al. [6] and Nourgaliev et al. [10] but could not be resolved, while Ye et al.
[24, 21], and Lee and LeVeque [7] offered sharp interface treatments but not
in the interfacial instability context considered here and in Theofanous et al.
[19, 20].
2 The Sharp Interfacial Dynamics Simulation (SIDS)

method
The central element of our method is a numerical imbedding of finite-volume
treatment on unstructured, cut-cell grid (C2 -grid) into a Cartesian grid
(C1 -grid), both incorporated into a Structured Adaptive Mesh Refinement
Numerical Prediction of Interfacial Instability 615
Fig. 2. Rayleigh-Taylor instability. a) On measuring of growth factor in DNS. b)

Scaled growth factor vs. scaled wavenumber: comparison of simulations and theory.
For definitions of scaled growth factor and wavelength see [3]. c) Dynamics of in-
h (t)
terface disturbance and αb (t) = Atbg t2 . d) Mixing at late stages of multimode RT
instability, At = 34 , σ ? = 0.0875 and Reh = 1000.
(SAMR) computational environment [1] that includes a Level Set (LS) treat-
ment [9, 11]. The C2 -grid is dynamically generated from the LS function in
a narrow band around the interface in order to allow conservative treatment
of the fluid dynamics, and to accurately piecewise-linearly represent the in-
terface, while avoiding small-cell issues. This complex data structure will be
denoted hereafter as adaptive Cartesian Cut-Cell Mesh Refinement (aC3 MR).
This quasi-Lagrangian-type treatment of the interface is complemented, dur-
ing fluid dynamics updates of the cut-cells, by using dynamic flux treatment
at interfacial edges, taking into account the momenta of inertia of irregular
cells, producing up to the third-order accuracy, with appropriate mapping of
cell-average to cell-center variables and communication between irregular C2
and structured C1 grids. The cost of supporting the C2 -grid is minimal as it
is generated in a narrow band around the interface, which pertains only to
the finest level of SAMR hierarchy.
The dynamic coupling between the two fluids on either side of the in-
terface is accomplished through a special discretization of fluid dynamics at
the centers of interfacial edges of the cut-cells. For this we have developed a
weighted least-squares, multi-dimensional, Taylor-expansion-based reconstruc-
Fig. 3. Stability of parallel flows. a) On measurement of growth factor in DNS. b)

Growth factor vs. wavenumber. c) (n − α) stability map. d) Dynamics of interface
and vorticity field: formation of “fingers” in non-linear regime (n = 1, m = 20,
Re = 20 and S = 0, see [20] for definitions).
tion, which allows incorporation of interface jump conditions up to any desired

order of accuracy (here, up to the third-order). A sample problem involving a
rather challenging reconstruction is shown in Figures 1(c,d). We can see that
our method provides the 2nd - and 3rd -order convergence for shear stresses
and primitive variables respectively - in contrast, but as expected, the DIM is
only first order in velocity and non-convergent in stresses. In addition, and not
surprisingly, Figures 1(a,b) demonstrate that our method virtually eliminates
the “spurious velocity” problem of DIM [17].
The fluid dynamics are treated by the pseudo-compressibility (PSC) method
[8]. First and foremost this allows high order discretization that fully exploits
the high-order features of our interface treatment just discussed. In addition,
this approach: (a) allows straightforward extensions to Godunov-based com-
pressible flow solvers, and thus coverage of all flow speeds [9]; (b) eliminates
the need for additional interpolations of velocity components (all variables
are collocated), which simplifies implementation; (c) eliminates the need for
variable-coefficient Poisson equation solvers, which is especially important for
the massively parallel computations needed for such simulations; and (d) is
readily adaptable to SAMR [1], with its own advantages for problems of this
type [9, 11].
Numerical Prediction of Interfacial Instability 617
The code, MuSiC-SIDS, has been built on the basis of previous versions,
MuSiC-ua2 LS, and -aCBM, it has been run on a 40-node Xeon-based clus-
ter, and shown to exhibit the good scalability needed for massively parallel
simulations.
3 Code Testing and Prediction of Instability

The code has been extensively verified on a number of challenging test prob-
lems, covering a wide range of flow speeds and fluid properties [8, 9, 10, 11].
Here, we will present a selected set of results for two canonical interface insta-
bility problems: Rayleigh-Taylor(RT) and viscous Kelvin-Helmholtz (VKH),
with a particular focus placed on ability to predict interfacial instability,
comparing results to linear stability theory.
A sample prediction of RT instability is shown in Figure 2. For the single-
mode cases the calculations were started with an imposed interface distur-
bance of maximum amplitude 2.5% and zero velocities everywhere. In (a) we
see the approach to the exact result of linear theory, in (b) we see the excellent
agreement of the growth factors with the linear theory over the whole spec-
trum of wave-numbers, in (c) we see the transition to the non-linear regime,
and the good agreement with the two-phase mixing layer growth rates (so-
called “bubble acceleration constant”) found in experiments. Finally, in Figure
2(d) we see a snapshot (t=5) of advanced RT mixing at an interface seeded, in
the same manner as just noted, with 8 modes. We have also run calculations
to mimic (to some extent) the DIM by smearing viscosity and surface tension
over 4 computational cells around the interface (density was treated sharply).
For the conditions and timescales examined shear stresses are of secondary
importance, and the results obtained were acceptable (qualitatively the same
and typically 5% error in growth factors).
A sample VKH instability case is shown in Figure 3. Linear theory (Orr-
Sommerfeld, O-S) results were obtained by Chebyshev Collocation Method in
the manner described by Theofanous et al. [20]. Again, we started the calcu-
lations from a completely stationary flow field, but now the instability was
seeded with much lower amplitude waves (0.01%). In (a) we see the approach
of the growth factor to the O-S result, in (b) we see the essentially exact
agreement of growth factors over a range of wave-numbers, and in (c) we see
that so-obtained stability map is in excellent agreement with classical results
[23]. Finally, in Figure 3 (d) we see the growth of an interfacial wave well into
the non-linear regime. While such long fingers have been produced previously
with the DIM [18, 4], in view of DIM’s failure to represent the linear regime
(see Figure 3 (a), and [20]), we must at least question these previous non-linear
results - further it should be noted that they were obtained by preconditioning
the whole flow field with a rather advanced stage in the linear regime.
Future applications will include superposition of these elementary insta-
bilities as found for example in drop breakup at Weber numbers greater than
100 [19, 20].
Acknowledgments. This work was carried out for the Joint Science and
Technology Office, Defense Threat Reduction Agency, with the support and
cooperation of Mr. C. Fromer and Dr. J. Hannan. We are also grateful to Drs.
G. Nakafuji, F. Handler (LLNL), and R. Babarsky (NGIC) for continuing,
long term support and collaboration.
References
1. Berger, M.J., and Colella, P.: Local Adaptive Mesh Refinement for Shock Hydrodynamics.
J. Comput. Phys., 82, 67–84 (1989)
2. Brackbill, J.U., Kothe, D.B., and Zemach, C.: A continuum method for modeling surface-
tension. J. Comp. Phys., 100(2), 335–354 (1992)
3. Chandrasekhar, S.: Hydrodynamic and Hydromagnetic Stability. Dover Publications Inc.,
New York (1981)
4. Coward, A.V., Renardy, Y.Y., Renardy, M., and Richards, J.R.: Temporal Evolution of
Periodic Disturbances in Two-Layer Couette Flow. J. Comp. Phys., 132, 346–361 (1997)
5. Hou, T.Y., and Lowengrub, J.S., and Shelley, M.J.: The long-time motion of vortex sheets
with surface tension, Phys. Fluids, 9(7), 1933 (1997)
6. Kang, M., Fedkiw, R.P., and Liu, X.-D.: A Boundary Condition Capturing Method for
Multiphase Incompressible Flow. SIAM J. Sci. Comp., 15(2), 323–360 (2000)
7. Lee, L., LeVeque, R.J.: An immersed interface method for incompressible Navier-Stokes
equations. SIAM J. Sci. Comp., 25(3), 832–856 (2003)
8. Nourgaliev, R.R., Dinh, T.N., and Theofanous, T.G.: A Pseudo-Compressibility Method
for the Numerical Simulation of Incompressible Multifluid Flows. Intern. J. Mult. Flow,
30(7-8), 901–937 (2004)
9. Nourgaliev, R.R., Dinh, T.N., and Theofanous, T.G.: Adaptive Characteristics-Based
Matching for Compressible Multifluid Dynamics, J. Comp. Phys., 213(2), 500–529 (2006)
10. Nourgaliev, R.R., Dinh, T.N., and Theofanous, T.G.: Sharp Treatment of Surface Tension
and Viscous Stresses in Multifluid Dynamics, AIAA 2005-5349, 17th AIAA CFD Conference,
Toronto (2005)
11. Nourgaliev, R.R., and Theofanous, T.G.: High-Fidelity Interface Tracking in Compressible
Flows: Unlimited Anchored Adaptive Level Set, J. Comp. Phys, under review (2006)
12. Nourgaliev, R.R., Liou, M.-S., and Theofanous, T.G.: Sharp Treatment of Interface Jump
Conditions in Multifluid Dynamics. J. Comp. Phys., to be submitted (2006)
13. Peskin, C.: Flow patterns around heart valves: a digital computer method for solving the
equations of motion. PhD Thesis, Physiol., Albert Einstein Coll. Med., Univ. Microfilms.
378:72-30 (1972)
14. Pozrikidis, C.: Instability of multi-layer channel and film flows, In: Advances in Applied
Mech., Ed. H. Aref and E. van der Giessen, Elsevier, (2004)
15. Prosperetti, A.: Boundary integral methods, In: Course on Drop-Surface Interactions, Ed.
M. Rein, Springer, (2002)
16. Ryskin, G., and Leal, G.: Numerical Solution of Free Boundary Problems in Fluid Mechan-
ics. Part 1. The Numerical Technique, J. Fluid Mech. 148(1) (1984)
17. Scardovelli, R., and Zaleski, S.: Direct numerical simulations of free-surface and interfacial
flow. Annu. Rev. Fluid Mech., 31, 567–603 (1999)
18. Tauber, W., Unverdi, S.O., and Tryggvason, G., The nonlinear behavior of a sheared im-
miscible fluid interface, Phys. Fluids, 14, 8 (2002)
19. Theofanous, T.G., Nourgaliev, R.R., Li, G.J., and Dinh, T.N.: Compressible Multi-
Hydrodynamics (CMH): Breakup, Mixing, and Dispersal, of Liquids/Solids in High Speed
Flows, In: IUTAM Symp. on Computational Approaches to Disperse Multiphase Flow, Ed.
A. Prosperetti and S. Balachandar, Springer, (2006)
20. Theofanous, T.G., Nourgaliev, R.R., Wiri, S., and Liou, M-.S.: Can Diffuse Methods be
Tolerated in Numerical Predictions of Interfacial Instability? Phys. Fluids, under review
(2006)
21. Shyy, W.: Multiphase computations using sharp and continuous interface techniques for
micro-gravity applications. Comptes Rendus Mechanique, 332, 375–386 (2004)
22. Unverdi, S.O., and Tryggvason, G.: A Front-Tracking Method for Viscous, Incompressible,
Multi-Fluid Flows. J. Comp. Phys., 100, 25 (1992)
23. Yiantsios, S.G., and Higgins, B.G.: Linear-stability of plane poiseuille flow of 2 superposed
fluids, Phys. Fluids, 31, 11 (1988)
24. Ye, T., Shyy, W., and Chung, J.N.: A fixed-grid sharp-interface method for bubble dynamics
and phase change. J. Comp. Phys., 174, 781–815 (2001)
A Lattice Boltzmann based Single-Phase
Model: Surface Tension and Wetting
Xiu Qing Xing, David Lee Butler, and Chun Yang
School of Mechanical and Aerospace Engineering, Nanyang Technological

University, 50 Nanyang Avenue, Singapore, 639798
xqxing@ntu.edu.sg
Summary. A 2D single-phase free surface tracking model using the Lattice Boltz-
mann method (LBM) is developed in this study. In contrast to the conventional
multi-phase models, it is not necessary to simulate the motion of the gas phase
using this LBM single-phase algorithm, and thus improves the computational effi-
ciency without sacrificing the underlying physics. Based on Gunstensen0 s immiscible
Lattice Boltzmann algorithm, a perturbation is added to the distribution functions
of the interface cells for introducing the surface tension and wetting into the LBM
single-phase model. Implementations of the model are demonstrated for simulations
of droplet deformation under the surface tension effect and the droplet spreading on
solid surface with various wetting properties. Simulation results show that the pro-
posed model can generate a surface tension that is isotropic, and the model allows
for continuous control of contact angles by varying the mass gradient at the wall
boundary cells.
1 Introduction
The simulation of flows with free surfaces has a variety of technical appli-
cations such as flow through porous media, boiling dynamics and etching
processes. Difficulties in the simulation of these types of problems lie in mod-
eling interface dynamics and dealing with the complex boundaries. Several
LBM models to simulate the interfacial dynamics have been developed in the
past decade. Gunstensen et al.[1] proposed a color model for simulating im-
miscible binary fluids based on Lattice Gas model of Rothman and Keller[2].
D0 Ortona et al.[3] studied the surface tension and wetting properties by mod-
ifying the color redistribution step and obtained good agreement with theory.
Shan and Chen[4, 5] and Shan and Doolen[6] considered a fluid with S differ-
ent components on a regular lattice. In order to include the surface tension,
they modified the collision operator by adding the force on the kth phase ow-
ing to a pairwise interaction between different phases. Orlandini et al.[7] and
Swift et al.[8, 9] proposed a free energy approach for simulating binary fluids
and liquid-gas fluids. The surface tension is considered by incorporating a free

620 X.Q. Xing et al
energy function. He et al.[10, 11] proposed a Lattice Boltzmann scheme for

simulation of multiphase flows with nearly incompressible limits. Besides the
multiphase and multi-component models, there are a couple of free surface
models developed for simulating the moving interface between immiscible gas
and liquid fluids. One of the free surface models is proposed by Ginzburg and
Steiner[12, 13]. Another free surface algorithm for gas-liquid interface sim-
ulations was proposed by Thürey et al.[14, 15]. The main character of this
method is that the gas phase only has negligible influence on the liquid phase
and is not simulated as a fluid. It can achieve a high computational efficiency
and require a low memory.
A LBM based single-phase model has been developed in this paper. It not
only retains the advantages of Thürey0 s free surface approach but also is able
to to simulate the effect of the surface tension and wetting properties in a
simple way with high computational efficiency and high accuracy. The sur-
face tension and wetting properties are added in the reconstruction step and
the boundary conditions in a simple way similar to the color model. Thus,
compared with the color model, the LBM single-phase model does not need
the color redistribution step, while no curvature calculation and surface re-
construction are needed compared with the free surface algorithm. Therefore,
the procedure is significantly simplified.
2 The LBM Single-phase Model

Standard lattice classification D2Q9 BGK model is used in this paper. The
LBM single-phase algorithm doesn0 t include the gas phase in the simulation.
The different phases only use a single lattice, and are distinguished by flags
for each cell in the grid. Each cell might have three types: filled (or fluid)
cells, interface cells or empty cells. Tracking the free surface interface layer
consists of a few steps as follow: the computation of the interface movement,
the reconstruction of the unknown distribution functions at interfaces, the
incorporation the effect of surface tension at fluid-gas interfaces as well as
wetting at the fluid-solid interfaces, and the re-initialization of the cell types.
Details about the basic idea of the free surface model can be found in [15],
here, we only focus on how to incorporate the surface tension and wetting into
the model.
Here, the surface tension is introduced in the same fashion as the perturbation
step in Gunstensen0 s color model[1] of immiscible fluids. The basic idea is to
add an anisotropic perturbation to the distribution functions for the interface
cells. Therefore, the reconstruction step becomes:

0 eq eq
∂m
fiopp (x, t+∆t) = [fi (ρA , u)+fiopp (ρA , u)+β
cos2(θi −θn )]−fi (x, t)(1)
∂n
The
∂msurface
tension is included by adding the perturbation term Cib =
β ∂n cos2(θi − θn ). θi is the angle of the lattice direction i, ∂m
∂n is the mass
Single-Phase Model: Surface Tension and Wetting 621
gradient of the liquid whose direction is perpendicular to the interface, and

θn = tan−1 ( ∂m
∂n ) is the angle of the local mass gradient. Here β, typically of
order 0.001, is a constant chosen to set the surface tension for the model. The
dependence of surface tension on β is basically linear. Theoretical analysis of
the link between the perturbation and the surface tension can be found in
Gunstensen[1] and D0 Ortona[3]. Therefore, compared with the free surface al-
gorithm of Thürey, our model doesn0 t need curvature calculation and surface
reconstruction. Moreover, no color redistribution is needed in our model when
compared with the Gunstensen0 s color model.
In order to include specific wetting properties of the wall, a perturbation pro-
cess similar to that used for modeling the surface tension
is added to the
distribution functions. By setting different values of ∂m
∂n , different wetting

properties can be obtained at the wall. A droplet of fluid in contact with a
wall with a zero gradient of the liquid should spread on the wall completely.
On the other hand, the blob should be repelled if the mass gradient equals to
the density of the liquid, which corresponds to a perfectly non-wetting case.
Intermediate situations should lead to partial wetting cases.
3 Numerical Results and Discussions
The test cases chosen for the application of the single-phase model concerns
surface tension and wetting. The model will be applied to two cases, droplet
deformation under the effect of surface tension, and droplet spreading on a
solid wall with different wetting properties. In our simulation for surface ten-
sion and wetting, gravity is not included as the dimensionless Bond number
Bo is less than 1.0 indicting negligible effect of the gravitational force com-
pared with the surface tension force.
Test case 1: Droplet deformation due to surface tension. From a thermody-
namics viewpoint, in the absence of external forces a two-phase system, under
surface tension effect, tends to minimize its interface area. This is caught by
the LBM single-phase model developed in this study. The simulation was car-
ried out in a computational domain with 40 × 40 lattices filled with fluid. The
relaxation time τ was set as 2.0. Here the droplet initial shape is assumed as
a square, and then it undergoes deformation under the effect of surface ten-
sion, as shown in Fig.1. It can be seen that the interface is quite smooth. In
calculation, surface tension control parameter is β = 0.002, corresponding to
the lattice surface tension of 0.37 ± 0.01 . The effective radius of the circular
blob as a function of the angular direction (−180◦ to 180◦ ) was checked. The
average value of the radius is equal to 22.5 cell units, and small fluctuation is
observed, of about 0.4 cell unit, which demonstrates that our model generates
an isotropic surface tension. The effect of surface tension on droplet shape is
also studied, and the results are shown in Fig.2. When the surface tension
control parameter β is set as 0.001, it can be seen that the final shape is not
an exact circle. While, when β is set as 0.002 or 0.003, the final shape of the
622 X.Q. Xing et al
(a) (b) (c) (d)
Fig. 1. Time evolution of a droplet changing from square shape to circular shape
(a) (b) (c)
Fig. 2. Final shape with different surface tension forces
droplet is a circle. The simulation results are in accordance with the surface
physics. The higher the surface tension forces, the faster the deformation. The
computational cost is also reduced significantly when the surface tension con-
trol parameter β increases.
Test case 2: Droplet wetting. Wetting involves interactions of liquids with a
solid surface. In simulation, collision parameter was chosen as τ = 2.0 , surface
tension control parameter β = 0.002. Initial droplet shape is assumed as a cir-
cle with a radius of 15 lattice units. Then the droplet undergoes deformation
on the solid surface due to surface tension effect, and finally reaches equi-
librium shape, Fig.3 shows the final equilibrium shape of droplets in contact
with a solid surface with different wetting properties. The different wetting
results are obtained by assigning different values of the mass gradient ∂m ∂n at
the wall cell. For Figs.3(a) to (i), the corresponding mass gradients are 1.0,
0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.2 and 0.0 respectively. Contact angles correspond-
ing to Figs.3(a) to (i) are 180◦ , 160◦ , 142◦ , 125◦ , 105◦ , 90◦ , 75◦ , 40◦ , and 0.0◦
respectively. Computational costs are in the order of 104 time steps.
In Fig.4, the symbols are the contact angles obtained from the final equilib-
rium shape of droplets simulated by the model proposed in this study. The
solid line is obtained from a simple model, namely a zero mass gradient corre-
sponds to a zero contact angle and a mass gradient equal to the liquid density
corresponds to a 180◦ contact angle, and intermediate situations should lead
to an angle between 0◦ and 180◦ according to a linear rule. It can be seen
that contact angle varies almost linearly with the mass gradient of the wall
cells, which is in good agreement with the simple model.
4 Conclusion
A LBM based single-phase free surface tracking algorithm has been developed.
The proposed model has been implemented to simulate the deformation of a
droplet due to surface tension effect, the time evolution of a droplet spreading
Single-Phase Model: Surface Tension and Wetting 623
(a) (b) (c) (d) (e) (f)
(g) (h) (I)
Fig. 3. Final equilibrium shape of droplets on solid surfaces with different wetting
properties
180 x
x
x simulation results
150
theoretical value x
x
Contact Angle
120
x
90 x
x
60
x
30
0x
0 0.2 0.4 0.6 0.8 1
Mass gradient
Fig. 4. Contact angle versus mass gradient at a solid wall
on solid surface with various wetting properties. Simulation of the deforma-

tion of a droplet driven by surface tension suggests that the proposed model
can generate an isotropic surface tension effect. Further, the time evolution of
a droplet spreading on a solid surface has been demonstrated and the differ-
ent wetting properties have been successfully simulated by varying the mass
gradient at the wall boundary cells. It is found that contact angle varies al-
most linearly with the mass gradient of the wall cells and the contact angles
obtained by simulation are in agreement with the simple model proposed in
this study.
Acknowledgement. The authors are grateful to the financial support from

Agency for Science, Technology and Research of Singapore under Project No:
0421010018.
References
1. Gunstensen,A.K., Rothman,D.H.: Lattice Boltzmann Model of Immiscible Flu-
ids, Phys. Rev. A 43, 4320–4327 (1991).
624 X.Q. Xing et al
2. Rothman,D.H., Keller,J.M.: Immiscible Cellular-automation Fluids, J. Stat.

Phys 52, 1119–11270 (1988).
3. D’Ortona,U., Salin,D., Cieplak, M., Rybka,R.B., Banavar, J.R.: Two-color non-
linear Boltzmann Cellular Automata: Surface Tension and Wetting, Phys. Rev.
E 51, 3718–3728 (1995).
4. Shan, X., Chen, H.: Lattice Boltzmann Model for Simulating Flows with Mul-
tiple Phases and Components, Phys. Rev. E 47, 1815–1819 (1993).
5. Shan, X., Chen, H.: Simulation of Non-ideal Gases and Liquid-gas Phase Tran-
sitions by the Lattice Boltzmann Equation, Phys. Rev. E 49, 2941–2948 (1994).
6. Shan, X., Doolen,G.D.: Multicomponet Lattice Boltzmann Model with Inter-
particle Interaction, J. Stat. Phys. 81, 379–393 (1995).
7. Orlandini,E., Swift, M.R., Yeomans, J.M.: Lattice Boltzmann Model of Binary-
fluid Mixtures, Europhysics Lett. 32-6, 463–468 (1995).
8. Swift, M.R., Osborn, W.R., Yeomans, J.M.: Lattice Boltzmann Simulation of
Nonideal Fluids, Phys. Rev. Lett. 75-5, 830–834 (1995).
9. Swift, M.R., Orlandini,E., Osborn, W.R., Yeomans, J.M.: Lattice Boltzmann
Simulation of Liquid-gas and Binary Fluid Systems, Phys. Rev. E 54, 5041–
5052 (1996).
10. He, X., Chen, S., Zhang,R.: A Lattice Boltzmann Scheme for Incompressible
Multiphase Flow and Its Application in Simulation of Rayleigh-Taylor Insta-
bility, J. Comput. Phys. 152, 642–663 (1999).
11. He, X., Chen, S., Doolen, G.D.: A Novel Thermal Model for the Lattice Boltz-
mann Method in Incompressible Limit, J. Comput. Phys. 146, 282–300 (1998).
12. Ginzburg, I., Steiner, K.: A free-surface Lattice Boltzmann Method for Model-
ing the Filling of Expanding Cavities by Bingham Fluids, Phil. Trans. R.Soc.
Lond. A 360, 453–466 (2002).
13. Ginzburg, I., Steiner, K.: Lattice Boltzmann Model for Free Surface Flow and
Its Application to Filling Process in Xasting, J. Comput. Phys. 185, 61–99
(2003).
14. Thürey, N.: A Single-phase Free Surface Lattice Boltzmann Method, MA The-
sis, University of Erlangen-Nuremberg, Germany(2003)
15. Thürey, N., Rüde, U., Korner, C.: Interactive Free Surface Fluids with the Lat-
tice Boltzmann Method, Technical report, University of Erlangen-Nuremberg,
Germany (2005).
Hyperbolicity, Discontinuities, and Numerics
of Two-Fluid Models
Meng-Sing Liou1,2 , Loc Nguyen1 , Chih-Hao Chang1 , Sveta Sushchikh1 ,

Robert Nourgaliev1 , and Theo Theofanous1
1
CRSS, University of California, Santa Barbara, CA 93117, USA
2
NASA Glenn Research Center, Cleveland, OH 44135, USA
1 Introduction
In flow situations where multiple material phases are found at scales that can-
not be resolved by direct numerical simulation, the dynamics can be described
by effective-field, or coarse-grain models [1-3]. Here we are concerned with the
flow involving two compressible fluids (at least one of them moving at high
Mach number), the basic formulation of the so-called two-fluid model, and
the numerical issues that arise from the well-known non-hyperbolic and non-
conservative characters of this model. Of particular interest is the application
of this model to flows with steep gradients in flow quantities and material
densities, as occurring in high-speed, disperse flows [4]. These, in conjunction
with the numerical issues just noted, can cause numerical difficulties, such
as spurious oscillations and non-convergence of solutions. The “remedy” of
incorporating adequate amounts of numerical diffusion is not an option.
Other “remedies” found in past work, all for low Mach number condi-
tions, are based on the recognition that certain inviscid interactions between
the phases have been lost in the coarse-graining procedure (space-, time-, or
ensemble-averaging) and that they must be recovered through modeling. In
particular we note the concepts of virtual mass [5-7] and interfacial pressure
[8]. The latter is to represent a local difference between the pressures in the
two phases; but its origin and role remain obscure. From our present stand
point we need to address the well-posedness of the two-fluid models and how
much the solution changes by the change of parameters that control the “de-
gree” of hyperbolicity. In particular, we will examine how hyperbolicity affects
the numerical solution in the presence of shocks and contact discontinuities,
under general flow conditions (in speeds, thermodynamic and material states).
2 The two-fluid equations
We use the standard form of the two-fluid model as given in [1-3]

626 Liou, M.-S. et al.
∂Uk ∂Fk
+ = Pk + Pint vm v
k + Fk + Fk , k = gas or liquid phase (1)
∂t ∂x
where Uk = αk ρk (1, uk , Ek )T denote the conservative variables, Fk = αk ρk uk
(1, uk , Hk )T + (0, αk p, 0)T the inviscid fluxes, Pk = p(0, ∂αk /∂x, −∂αk /∂t)T
the pressure interaction between phases, and Fvk the viscous diffusion terms.
The inviscid interfacial transfer terms include effects of interfacial pressure
(Pint
k ) and virtual
vm
P mass (Fk ). αk is the volume fraction of phase k, satisfying
the constraint k αk = 1. The other variables are standard.
3 Hyperbolicity
As the interfacial terms are typically expressed as first derivatives in space
and time, they can exert a significant influence on the eigenvalues, hence the
hyperbolicity of the system. Studies of hyperbolicity of two fluid models have
been reported in the literature for both concepts (see for example [7,8]), but
they were all made for incompressible flows under the assumptions w2 << a2g ,
where w is the relative velocity and ag the speed of sound in gas phase. the
result is a simple, explicit hyperbolicity condition, however, this is not the case
in general. In what follows we will establish the conditions under which the
system becomes hyperbolic for all flow conditions and demonstrate the effects
of hyperbolicity on the numerical solutions, especially in terms of convergence
and accuracy.
3.1 Interfacial pressure
Following Stuhmiller [8] we express interfacial pressure Pint

k by
∗
Pint T
k = −δp (0, ∂αk /∂x, −∂αk /∂t) , (2)
where δp∗ = σαg αl ρg ρl w2 /(αg ρl + αl ρg ) and w = ul − ug . σ ≥ 1 is the
necessary condition for hyperbolicity under the assumption w2 << a2g [8].
In Fig.1(a) we show that this condition is violated in a way that would
appear to be cumbersome for computations, when this assumption does not
apply. On the other hand in Fig.1(b) we show that the formulation
δp∗ = Cp∗ ρg w2 , (3)
yields the much simpler and convenient to implement bound, Cp∗ ≥ 1, which
was found to be adequate for all pressure conditions (not shown).
3.2 Virtual mass
Following Drew et al. [5], we express the virtual mass term by Fvm
d =
fdvm (0, 1, uint )T , with
∂(uc − ud ) ∂uc ∂ud
fdvm = Cvm αd ρc [ + uvm
c − uvm
d ], (4)
∂t ∂x ∂x
Hyperbolicity, Discontinuities, and Numerics of Two-Fluid Models 627
Fig. 1. Effects of the interfacial pressure parameters σ and Cp∗ on hyperbolicity.
where subscripts “c”, and “d” refer to the continuous and disperse phase
respectively, and superscript “vm” is to indicate the velocities appropriate to
the substantial derivatives involved in the virtual mass formulation [5]. Also,
uint is the interfacial velocity, and Cvm is an appropriate coefficient which
we will use as a parameter in investigating the hyperbolicity of the system.
We solve for the boundary of the hyperbolic zone using the full characteristic
equation of the system (no assumptions on the magnitude of the relative
velocities), and we use Cp∗ and Cvm as parameters alone or in combination.
In Fig. 2(a) we show minimum values of Cvm required for hyperbolicity as
a function of the value of Cp∗ covering the whole space of void fractions and
correspondingly of disperse flow regimes, from bubbly to droplet, but under
the assumption w2 << a2g . Note that in certain combinations the magnitude of
the virtual mass needed for hyperbolicity appears to somewhat exceed what
is known to represent physical reality—in particular the low void fraction
(bubbly) flows. In Fig. 2(b) we show the regularizing potency of the virtual
mass term alone, as a function of Mach number referred to the relative velocity
(w). In contrast to the case just shown, the very high void fraction (droplet)
flows appear to need somewhat excessive values of the coefficient, while the
need for regularization of bubbly flows diminishes with increasing relative
velocity, to a vanishingly small quantity beyond Mach number of about 0.3.
Also it should be clear that in Fig. 2 the limit w → 0 is approached correctly
in that the virtual mass term needed for regularization also goes to zero, as
appropriate to reflect the well-posed homogeneous flow model, even though
the coefficient itself appears to remain finite.
3.3 Numerical evidences—stability and convergence
In the following we show, by example, that careful conditioning of the system,

informed by results such as those shown above, in combination with an appro-
priate numerical scheme that recognizes and takes account of the presence of
discontinuities in the interactions between phases, as is the case with AUSM+ -
up [9], can produce rather satisfactory results in predicting flow dynamics,
Fig. 2. The hyperbolicity boundary on Cvm for the virtual mass in [5].
even under pathologically extreme conditions. We implemented AUSM+ -up

for the two-fluid model in our All-Regime Multiphase Simulator (ARMS),
with which the following results were obtained.
One such example is Ransom’s Faucet problem [10]. In the case consid-
ered we let a water column, with an initially uniform volume fraction of 0.8
(surrounded by air, in an 1D, vertical channel, continuously supplied at the
inlet/top) and velocity of 10 m/s, suddenly be subject to gravitational ac-
celeration. There is no viscous interaction between the phases. The flow is
characterized by a contact discontinuity moving down with the free-fall ve-
locity and a gradually decreasing void fraction behind this front, all the way
to the inlet. We use only interfacial pressure to regularize the system, and we
observe ability to capture the contact discontinuity in terms of the degree of
conditioning applied, and convergence of solutions under grid refinement.
Sample results are shown in Fig. 3 in comparison with the analytical so-
lution. This is a low Mach number flow, and as seen above a value of Cp∗ of
about 0.5 should be an appropriate definition of the hyperbolic boundary. At
the contact discontinuity the jump in void fraction induces a very large gradi-
ent term in Pk (becoming ever larger with grid refinement), which is also the
source of non-hyperbolicity. In Figure 3(a) we can see that ARMS can pro-
duce a solution, even for the ill-posed problem, albeit with a sharp spike just
behind the moving front. By placing the system deeper into the hyperbolic
region (increasing the value of Cp∗ ), the magnitude of this spike gets smaller,
but at the expense of increased smearing of the front. On the other hand,
in Fig. 3(b) we see that for a value of Cp∗ that places the system somewhat
deeper inside the hyperbolic zone, the analytical result can be captured to any
degree of accuracy by grid refinement, with the solution remaining robust and
no indication of catastrophic instabilities found in previous attempts with this
problem. This example strongly suggests that hyperbolicity is necessary
for a unique and stable solution, while, for a conditioned problem,
the non-conservative term does not give rise to non-uniqueness or
inaccuracy in speed/location of the discontinuity.
Hyperbolicity, Discontinuities, and Numerics of Two-Fluid Models 629
Fig. 3. Faucet problem. (a): effect of hyperbolicity, and (b): grid convergence (note
the expanded scale in x to view the solution near the volume fraction front.)
Finally, we consider a 1D shock tube problem studied by Toumi [11] to fur-

ther address the issues of ill-posedness and capability of a numerical method.
There are two isothermal bubbly mixtures, with void fractions of 25% and
10%, and pressures of 200 and 100 bar, respectively on the left and right sides
of a diaphragm. The solution involves compression and rarefaction waves in
complex dynamics for which no analytical expression is known. Past results
indicate either failure to obtain a solution, or a wide variation of results when
such solutions were possible. In Fig. 4 we show sample results obtained with
either the virtual mass or interfacial pressure. We can see clearly that once
sufficiently conditioned (Cp∗ = 1, as predicted by Eq. (3)), in both cases the
solutions were robust under grid refinement, and more importantly the wave
structures are identical. The only minor difference between them lies in the
contact discontinuity situated in the middle; this is to be expected because of
the different regularization terms employed in the model, the choice being a
matter of physics.
Fig. 4. Shock tube problem [11]. (a): with interfacial pressure alone and (b): with
virtual mass alone [5].
We showed the effects of interfacial pressure and virtual mass on hyperbolicity
of the two-fluid equations, providing a complete map in an all-speed range.
In turn, the effects of hyperbolicity on stability and convergence of numerical
solutions were investigated using an accurate method. We concluded that
the concern due to the con-conservative form did not impact convergence
or uniqueness, as long as the system was conditioned adequately. On the
other hand our numerical results indicate that placing modestly inside the
hyperbolic zone may not be always sufficient for numerical regularization,
and this requires further investigation. Also more work is needed, especially on
the interfacial pressure term, to clarify whether the amount of regularization
indicated from the hyperbolicity analysis is quite consistent with physical
reality.
Acknowledgments. This work was carried out for the Joint Science and Technol-
ogy Office, Defense Threat Reduction Agency, with the support and cooperation
of Mr. C. Fromer and Dr. J. Hannan. We are also grateful to Drs. G. Nakafuji, F.
Handler (LLNL), and R. Babarsky (NGIC) for continuing, long term support and
collaboration.
References
1. Drew, D., Passman, S.L.: Theory of Multicomponent Fluids. Springer-Verlag
NY (1999)
2. Ishii, M.: Thermo-Fluid Dynamic Theory of Two-Phase Flow. Eyrolles, Paris
(1975)
3. Zhang, D.Z., Prosperetti, A.: Physics of Fluids, 6, 2956–2970 (1994)
4. Theofanous T. et al: Computational Approaches to Multiphase Flow (Ed. Bal-
achander, S. and Prosperetti, A.), Springer, 353–369 (2006)
5. Drew D. et al.: Int. J. Multiphase Flow, 5, 233–242 (1979)
6. Drew D., Lahey R.T.: Int. J. Multiphase Flow, 13, 113–121 (1987)
7. Watanabe, T., Kukita, Y.: Nuclear Eng. and Design, 135, 327–340 (1992)
8. Stuhmiller, J.H.: Int. J. Multiphase Flow, 3, 551–560 (1977)
9. Liou, M.-S.: J. Comput. Phys. 214, 137–170 (2006)
10. Ransom, V.H.: Multiphase Science and Technology, 3 (1987)
11. Toumi, I.: Nuclear Eng. and Design, 123, 147–168 (1996)
Computation of Multiphase Mixture Flows
using RoeM and AUSMPW+ Schemes
Seung-Won Ihm, Kyung Rok Lee, Chongam Kim∗ , and Kyu Hong Kim
Department of Aerospace Engineering, Seoul National University, Seoul 151-742,

Korea chongam@snu.ac.kr
Summary. RoeM[1] and AUSMPW+[2] schemes, originally developed for the ro-
bust and accurate simulation of single phase gas dynamics, are extended to the
compressible gas-liquid two-phase flows. We adopted homogeneous equilibrium
model(HEM) using mass fraction to describe different two phases. In the Eulerian-
Eulerian framework, HEM assumes dynamic and thermal equilibrium of the two
phases in the same computational mesh. We derived new pressure weighting func-
tion from the mixture equation of state(EOS) to construct the new shock disconti-
nuity sensing term(SDST), which is commonly used in RoeM and AUSMPW+ for
the stable numerical flux calculation. The developed two-phase versions of RoeM
and AUSMPW+ are applied on several air-water two-phase test problems. In spite
of the large discrepancy of material properties such as density, enthalpy, and speed
of sound, the numerical results show that both schemes provide very satisfactory
solutions.
HEM[3] with mass fraction is adopted to describe two-phase flows. As we
assume fully compressible flows, the governing equations are composed of
mixture mass, momentum, and energy conservation laws together with one
phase mass conservation law. The two-dimensional Euler equations are as
follows:
∂Q ∂E ∂F
+ + =0, (1)
∂t ∂x ∂y
T
Q = [ρm , ρm u, ρm v, ρm E, ρm Y1 ] ,
T
E = [ρm U, ρm uU + nx p, ρm vU + ny p, ρm HU, ρm Y1 U ] ,
where ρm is a mixture density, and Y1 stands for the mass fraction of gas
phase.
And the following definition of mixture density combined with each phase’s
EOS plays a role of mixture EOS[4]:

632 Seung-Won Ihm, Kyung Rok Lee, Chongam Kim, and Kyu Hong Kim
1 1
ρm = P Yi
=P Yi
. (2)
ρ̂i (p,T ) ρ̂i (p,hi )
In Eq. (2), ρ̂i means density defined on the occupied volume of i-th fluid,
while ρm is defined on the each whole computational mesh. We adopted the
stiffened fluid EOS for liquid phase, and the ideal gas EOS for gas phase[5].
Finally, the system is closed by the following assumption of dynamic and
thermal equilibrium:
p = pl = pg , T = T l = T g . (3)
2 Shock Discontinuity Sensing Term

The pressure ratio term min ppRL , ppRL at the control surface (i.e., cell inter-
face) is used both in RoeM and AUMSPW+ to sense the shock discontinuity
and control the proper numerical fluxes. For the single phase flows of gas dy-
namics, the direct use of real pressure ratio works well. In the liquid phase,
however, due to the large density and high speed of sound, pressure field varies
drastically even for non-shock region. So the proper scaling of the SDST is
required.
Following the idea of ref.[6], we start from the density ratio, and try to
change it into the ratio of pressure function. With the mixture EOS, the
density can be expressed as
∂ρm
∂p
ρm = α1 1−α1 , (4)
γp + n(p+pc )
where α1 is a volume fraction of gas phase, and γ, n, pc are constant coefficients

from the EOS. If we assume to use the common interfacial values of ∂ρ ∂p and
m
α1 , we can define the following pressure function p̄(p).

1
p̄(p) = α1 1−α1 . (5)
γp + n(p+pc )
Then, the density ratio can be changed into the ratio of p̄.
ρm,L p̄(pL ) p̄L
≈ = (6)
ρm,R p̄(pR ) p̄R

Now we can use min p̄p̄RL , p̄p̄RL as a SDST. For one dimensional normal
shock condition of M=1.5 and 2, the proposed SDST has the value of 1/O(1)
even for the near liquid phase.
RoeM and AUSMPW+ for Two-phase Flows 633
3 The RoeM scheme for two-phase flow

In the compressible two-phase flow, direct derivation of the system Jacobian
matrix using conservative variables is very difficult to obtain due to its com-
plicated form of EOS. Moreover, because of the large speed of sound at liq-
uid phase, many two-phase flow analyses require preconditioning technique,
which alters the governing system to primitive variable-based form. Thus the
Roe scheme based on primitive variables are popularly used in the two-phase
flow research area. The primitive variable-based Roe scheme in our two-phase
model can be summarized as follows.
1
EL + ER − Γe,1/2 Γe−1 Ap ,

E1/2 = (7)
2
∂Q ∂E
where, Γe = , Ap = .
∂Qp ∂Qp
And our choice of primitive variables is
T
Qp = [p, u, v, hm , Y1 ] . (8)
After some mathematical procedure, Roe scheme can be converted into
a HLLE-like form. And by introducing control functions f and g, the RoeM
scheme for compressible two-phase flow is derived and summarized as follows:
1
E1/2 = EL + ER − M̃ Ap ∆Qp + ĉ M̃ 2 − 1 Γe ∆Qp + gĉ 1 − M̃ B∆Q ,(9)

2
   
1 0
 û   ∆u − nx ∆U 
∆p    
B∆Q = (∆ρm − f 2 )   + ρ̂m 
 v̂   ∆v − n y ∆U ,
ĉm   ∆p 
∆H − Û ∆U − ρ̂m

Ĥ 
Ŷ1 ∆Y1
1 , û2 + v̂ 2 = 0

f= ,
|M̂ |h , û2 + v̂ 2 6= 0
( p̄ ,j p̄
1−min( p̄ i , i+1,j
p̄i ,j )
g= |M̂ | i+1,j , M̂ 6= 0 .
1 , M̂ = 0
It is noted that Mach number-based functions f and g are introduced to
balance damping and feeding rates, which leads to a shock-stable Roe scheme.
4 The AUSMPW+ scheme for two-phase flow

Contrary to FDS type schemes, AUSM type schemes have an advantage in
the application to complex fluid systems, because a consistent form used in
single phase can be extended. In the following, the AUSMPW+ scheme for
the two-phase flow is given.
E1/2 = M̂L+ c1/2 φL + M̂R− c1/2 φR + (PL+ pL + PR− pR ) , (10)
1) M1/2 ≥ 0 ,
M̄L+ = ML+ + MR− [(1 − ω)(1 − fR ) − fL ] ,
M̄R− = MR− ω(1 + fR ) ,
2) M1/2 < 0 ,
M̄L+ = ML+ ω(1 + fL ) ,
M̄R− = MR− + ML+ [(1 − ω)(1 − fL ) − fR ] ,
where, ω = max(ω1 , ω2 ) ,
3
ω1 = 1 − min p̄p̄RL , p̄p̄RL ,
min(p̄ ,p̄ ,p̄ ,p̄ )
ω2 = 1 − max(p̄1,L 1,R 2,L 2,R
,p̄1,R ,p̄2,L ,p̄2,R ) ,
1,L

p̄ ρm,L ρm,R
fL,R = L,R p̄s − 1 × (1 − ω2 ) × min ρm,R , ρm,L ,
T
φ = [ρm , ρm u, ρm v, ρm H, ρm Y1 ] .
Our modifications are
• M1: Introduction of new SDST

• M2: Use of Roe type speed of sound at the interface
• M3: Modification of function f to consider the density ratio of both sides
Note that the speed of sound at phase interface is lower than that of either
phases, because HEM allows mixture region. From the mixture EOS, typical
averaging of mass fraction at phase interface can show this characteristic, but
any averaging of left and right speed of sound cannot predict low speed of
sound. It is also noted that the function f is modified to stabilized the scheme
near the phase interface of large density ratio.
5 Numerical Results
5.1 Odd-Even Decoupling Test
In order to test the shock stability of the schemes, we solved two-phase mix-
ture moving shock problem on a perturbed grid. Figure 1 shows the results of
each scheme for Y1 = 0.1 mixture flow. While Roe scheme destroys the nor-
mal shock structure, both RoeM and AUSMPW+ clearly capture the shock,
showing their robust and stable behavior even in the two-phase flow region.
RoeM and AUSMPW+ for Two-phase Flows 635
Fig. 1. Quirk’s test on Y1 = 0.1 mixture flow
5.2 Liquid Shock-Phase Interface Interaction(SII)
In this challenging problem, the liquid shock of M=1.7 encounters with the
phase interface of about 1:1000 density ratio. Initial right-going liquid shock
is located at x=0.1, and phase interface is at x=0.5. Results at t=0.3ms are
shown in Fig. 2. Well known physics including shock transmission/reflection
pattern, acceleration after the shock-phase interface interaction are shown.
Fig. 2. SII: density(left), velocity(middle), pressure(right) at t=0.3 ms
5.3 2-D Shock-Bubble Interaction(SBI)
As a more practical problem, we chose liquid shock-cylindrical air bubble in-

teraction problem with 175 by 125 mesh. A moving shock of M=1.422 in liquid
hits the cylindrical gas bubble of d=2mm. This problem is the modeling of
the contribution of gas bubbles near kidney stone in lithotripsy process. Nu-
merical results show well-known flow physics including blast wave and liquid
jet formation.
Fig. 3. Numerical schlieren of SBI at every 0.1 µs
6 Conclusion
Numerical methods for simulating compressible two-phase flows with large
density ratio are presented. We extend the RoeM and AUSMPW+, which are
originally developed for single-phase gas dynamics, to two-phase flows. For
the two-phase flows using homogeneous equilibrium model(HEM) with mass
fraction, new shock discontinuity sensing term(SDST) is derived from mixture
EOS. The developed two-phase versions of RoeM and AUSMPW+ are tested
on several air-water two-phase problems. In spite of the large discrepancy of
material properties, the numerical results show a good performance of the
developed schemes. We expect the proposed SDST can be used for other nu-
merical schemes for two-phase flows. And let us mention that both the SDST
and the developed schemes have a consistent form for isothermal compressible
flows.
References
1. S. S. Kim, C. Kim, O. H. Rho, S. K. Hong : Cures for the shock instability:
Development of a shock-stable Roe scheme, J. Computational Physics, 185,
342–374 (2003)
2. K. H. Kim, C. Kim, and O. H. Rho : Methods for the Accurate Computations of
Hypersonic Flows, Part I: AUSMPW+ Scheme, J. Computational Physics, 174,
38–80 (2001)
3. S. Venkateswaran, J. W. Lindau, R. F. Kunz, C. L. Merkle : Computation of Mul-
tiphase Mixture Flows with Compressibility Effects, J. Computational Physics,
180, 54–77 (2002)
4. Robin K. S. Hankin : The Euler Equations for Multiphase Compressible Flow in
Conservation Form, J. Computational Physics, 172, 808–826 (2001)
5. H. Luo, J. D. Baum, R. Lohner : On the computation of multi-material flows
using ALE formulation, J. Computational Physics, 194, 304–328 (2004)
6. K. H. Kim, J. H. Lee, O. H. Rho : An Improvement of AUSM Schemes by In-
troducing the Pressure-Based Weight Functions, Computers and Fluids, 27(3),
311–346 (1998)
A relaxation method for the Kapila model
C. Berthon12 , B. Braconnier1,2,3 , J. Claudel3 , and B. Nkonga12

1
MAB, UMR 5466, Université Bordeaux 1, 351 cours de la libération, 33400
Talence, France.
2
INRIA Futurs, projet ScAlApplix, Domaine de Voluceau-Rocquencourt, B.P.
105, 78153 Le Chesnay Cedex, France.
3
CEA CESTA, route des Gargails, 33114 Le Barp.
braconnier.benjamin@math.u-bordeaux1.fr
Summary. The present work is devoted to the numerical approximation of a com-

pressible two-phase flow. The phases are non-miscible. We use a diffuse interface
model having a single velocity and pressure [1]. For this non-conservative system,
we propose a Godunov type method based on a pressure relaxation procedure. We
also reformulate the relaxation solver in the term of the wave propagation method
[2]. Two numerical illustrations are performed: the 1D test proves the results accu-
racy and the 2D test, the hability to deal with fine meshes.
Key words: Kapila model, relaxation scheme, non-conservative system
1 Mathematical model
We are concerned with the numerical approximation of a two phase compress-
ible flow. We assume the phases non-miscible. For the sake of simplicity, the
viscosity effects are neglected. We consider the physical model, derived in [1],
defined as:
∂t W + ∂xF + B ∂x U = 0 (1)
with:
       
α1 0 u00β 0 α1
 α1 ρ1   α1 ρ1 u  0 0 0 0 0  ρ1 
       
 α2 ρ2  , F =  α22ρ2 u  , B =  0 0 0 0
W= 0 , U =  ρ2  (2)
     
 ρu   ρu + p  0 0 0 0 0   u 
E ρHu 0000 0 ε
where α1 ∈ (0, 1) is the volume fraction for the first phase and α2 = 1 − α1
is recovered from the saturation constraint. For each phase k, we note ρk > 0
the partial density. ε > 0 is the fluid internal energy, u the velocity and p the
pressure. The mixture density, total energy and enthalpy are:

638 C. Berthon, B. Braconnier, J. Claudel, and B. Nkonga
2
X 1 E+p
ρ= αk ρk , E = ρ ε + u2 , H= (3)
2 ρ
k=1
The parameter β, associated to compaction effects, is defined as:

ρc2

1 X αk
β = α1 1 − 2 where = (4)
ρ1 c1 ρc 2 ρk c2k
k
Finally, the internal energy is given by the Stiffened gas equation of state:
p + γP ∞ 1 X αk γP ∞ X αk γk P ∞
k
ρε = where = and = (5)
γ−1 γ−1 γk − 1 γ−1 γk − 1
k k
where γk and Pk∞

are constant data of the phase k.
After a mathematical analysis, we deduce the system (1) unconditionally
hyperbolic with the eigenvalues u and u ± c.
2 Numerical approximation: 2D case

According to the finite volume approximation, the computation domain Ω
is recovered by non overlapping cells Ci (i = 1, .., N s). We denote ν(i) the
set of the neighbouring vertices of the node i. Each cell Ci is decomposed in
elementary area Cij and its frontier ∂Ci is also decomposed in X elementary
frontiers ∂Cij . We note aij the area of the sub-cell Cij and ai = aij the
j∈ν(i)
area of the cell CiZ
. Then, we define the interface normal between the cells Cij
and Cji by η ij = n dx where n is the unit normal of ∂Cij directed from
∂Cij
Ci to Cj .
The solution is updated on cells by using the solution of 1D Riemann
problems associated to the direction η ij :
n
n Wi if η < 0
∂t W + A ∂η W = 0 with W(t , η ) = (6)
Wnj if η > 0
where η is a coordinate associated to η ij . Then, the update state Wn+1 i is

defined as the integral of the exact or approximate solutions Vij of (6) over
each cells Cij with j ∈ ν(i). It can be rewritten as follow:
Z
∆t X 1
Wn+1
i = W n
i − φ ij where φ ij = (Wni − Vij (ηη )) dηη (7)
ai ∆t Cij
j∈ν(i)
Let note that in the general case of a non-conservative system, the fluctuations
φ ij and φ ji are not equal. For conservative systems, the fluctuations balance
and the scheme becomes conservative.
If we assume that we can approximate the solution of the Riemann prob-
lem (6) by constants discontinuous waves, our formulation meets the wave
A relaxation method for the Kapila model 639
propagation method [2]. Let suppose that the approximate solution satisfies
the consistency condition:
Xm
λl δ Wl = ∆FF + B̃
B ∆U (8)
l=1
where λl denotes the speed of the discontinuity, δ Wl the jump of the solution
and m the number of discontinuities. B̃B is a constant matrix standing for the
non-conservative terms.
Then the numerical scheme (7) can be reformulated as follow:
ai X
Wn+1 − Wni + C NC

i Ψij · ηij + Ψij · ηij = 0 (9)
∆t
j∈ν(i)
where:
C 1
NC 1
Ψij = F j + F i − |Ã
Aij | (∆W)ij , Ψij = B̃B ij (∆U)ij (10)
2 2
and the viscosity matrix is given by the relation:
Xm
|Ã
A|ij (∆W)ij = |λl |δ Wl (11)
l=1
NC
Ψij stands for the non-conservative contribution. For the conservative part,
C
we have the flux Ψij with its centered and dissipative terms put in evidence.
3 Relaxation based Riemann solver

After the works of Jin and Xin [3] and Liu [4], the relaxation method can be
viewed as a well established tool to approximate the solution of the compress-
ible Euler equations of gas dynamics. For the Kapila model, we propose the
following relaxation system:
1
∂t WR + ∂xF R + B R ∂x UR = R (12)
λ
T T
where WR = WT , π, a and UR = UT , π, a . π is the relaxation pa-

rameter, supposed free of thermodynamical property and tending toward the

pressure p in the time λ. In order to obtain a diffusive approximation, the
parameter a is defined by a subcharacteristic condition [4, 5]. It requires that
the modified sound speed should be at least as large as the sound speed of the
Kapila system (1). The relaxation flux, matrix and source term are given by:
     
0 u 0 0 β̃ 0 0 0 0
 α1 ρ1 u  0 0 0 0 0 0 0  0 
     
 α2 ρ2 u  0 0 0 0 0 0 0  0 
 2
    
FR =  ρu + π  , B R =  0 0 0 0 0 0 0  , R =  0  (13)
    
 (E + π)u  0 0 0 0 0 0 0  0 
     
 0   0 0 0 a2 /ρ 0 u 0  p − π
0 000 0 00u 0
An additional difficulty for the Kapila system is the term β which expression
depends on the closure (5). We want our numerical scheme free of thermody-
namical closure. In this way, we introduce the term β̃ for which we demand
two properties: it must be a Riemann invariant across shock and rarefaction
waves, the volume fraction saturation constraint must be recovered. We pro-
pose β̃ = α1 − y1 a21 /a2 where yk = αk ρk /ρ is the mass fraction of the fluid k
2
1 X yk
and the positive reals ak are such that: 2 =
a a2k
k=1
Then we rewrite the relaxation system in the form:
1 ∂FFR UR
∂U
∂t WR + A R ∂x WR = R with A R = + BR (14)
λ ∂ WR WR
∂W
With λ set to infinity, the system is unconditionally hyperbolic and linearly
degenerated. We illustrate the Riemann invariants associated to each wave
in the table (1). As the relaxation system is linearly degenerated , it is easy
−1
to determine the viscosity matrix |Ã AR | = P̃P R |Λ̃
ΛR |P̃
P R . Actually, the matrix
ΛR must contain the eigenvalues of the relaxation system and the matrix
Λ̃
P R must contains non nul vectors proportionnal to the solution jumps. After
P̃
straightforward calculations, we find:  L
L ∗ ∗ R
λ = λu−a/ρ
|Λ̃
Λ|R = Diag |λ |, |u |, ..., |u |, |λ |

u∗ = λu





 λR = λu+a/ρ
 L L 

β /ρ 1 0 0 0 0 0 β R /ρR

 L
∗a


 y1L 0 1 0 0 0 0 y1R   L
 J = H L
− u
ρL
  

 y2L 0 0 1 0 0 0 y2R  
R
where ∗a (15)
 
 y2L 0 1 0 0 0 0 y2R  J R
= H R
+ u
PR = 
P̃ R

 λL 0 u ∗ u ∗ 0 0 0 λR 
 

 ρ
aL
  

 JL 0 0 0 0 0 0 JR  L ∗ R
∆ = (u − λ )


ρL
  

 ∆L 0 0 0 0 0 0 ∆R  

aR


0 0 0 0 001 0  ∆R = (λR − u∗ ) R


ρ
Table 1. Riemann invariants for the relaxation system
λ Iλ1 Iλ2 Iλ3 Iλ4 Iλ5 Iλ6

λu u π
λu±a/ρ u ± a/ρ π ∓ au π 2 − 2a2 ε y1 = α1 ρ1 /ρ (α1 − y1 a2 /a21 )/ρ a
4 Numerical results
The first numerical test (figure 1) is proposed in [6]. We ran this experiment
with the relaxation method and the acoustic solver [7] which is based on a
A relaxation method for the Kapila model 641
Fig. 1. Shock tube problem at t = 0.01 s, domain length 1 m, γ1 = 1.667, γ2 = 1.4.

At initial time ρ1 = 3.984 kg.m−3 , ρ2 = 0.01 kg.m−3 , if x < 0.5 m, α1 = 1 − 10−8 ,
u = 27.335 m.s−1 , p = 1000 P a, if x > 0.5 m, α1 = 10−8 , u = 0 m.s−1 , p = 1 P a.
mesh 200 and 2000 points. numerical method: relaxation and acoustic solver
splitting of non-conservative terms. Both methods are in good agreement with

[6] even for the coarse mesh.
The second numerical experiment (figure 2) is proposed in [7]. Because of
the symetry, we reduced to half the computational domain. We use 300 × 600
points and performed parallel computing on 4 processors. Gravity forces and
surface tension force [8] are added with a splitting method. Compared with
[7], we note that the addition of surface tension force prevent the bubble to
divide.
Mesh decomposition T = 0.642 s T = 1.27 s T = 1.867 s
Fig. 2. Rising air bubble in water in a 2 m × 2 m domain. The computational is

1 m × 2 m (symetry x = 0) is discretised with 300 × 600 points and decomposed in
four sub-domains for parallel computing. Initial data ρair = 1.0 kg.m−3 , ρwater =
1000 kg.m−3 , u = 0 m.s−1 , P = 105 P a, gravity g = −10 m.s−2 , surface tension
coefficient σ = 0.0073, bubble radius 0.20 m, bubble center (0.0, 0.3).
5 Conclusion
A relaxation system for the Kapila model has been devised. The numerical
implementation is based on Godunov type methods feeting the propagation
waves methods. Two experiments have been run proving the accuracy and re-
liability of the method. The future work is devoted to the implicit formulation
and the addition of viscosity effects.
References
1. A.K. Kapila, R. Menikoff, D.S. Stewart: Two-Phase Modeling of Deflagration
to Detonation Transition in Granular Materials: Reduced Equations. Physics of
Fluids, 13, 3002–3024, (2001).
2. LeVeque, Randall, Shyue, Keh-Ming: Two-dimensional front tracking based on
high resolution wave propagation methods. Journal of Computational Physics,
123, Issue 2, 354–368 (1996)
3. S. Jin, Z. P. Xin: The relaxation schemes for systems of conservation laws in
arbitrary space dimensions. Comm. Pure Appl. Math. 48, 235–278, (1995).
4. T.P. Liu: Hyperbolic conservation laws with relaxation. Comm. Math.
Phys.,108, 153–175, (1987).
5. J. Whitham Linear and Nonlinear Waves. Wiley, New-York, (1974).
6. X. Y. Hu, B. C. Khoo: An interface interaction method for compressible multi-
fluids. Journal of Heat and Mass Transfert, 45, No 6 (2001), 1287–1307
7. H. Guillard, A. Murrone: A five equation reducezd model for compressible two
phase flow problems. Technival Report RR-4778, INRIA, 2003.
8. J.U. Brackbill, D.B. Kothe, C. Zemach: A Continuum Method for Modeling
Surface Tension. Journal of Computational Physics, 100, 335–354, (1992)
Direct numerical simulation of bubbly
Taylor-Couette flow
Thomas Nierhaus1 , Jean-Franccois Thomas1 , Yves Detandt2 , and David

Vanden Abeele1
1
Von Karman Institute for Fluid Dynamics, Belgium (nierhaus@vki.ac.be)
2
Université Libre de Bruxelles, Belgium (ydetandt@ulb.ac.be)
1 Introduction
Taylor-Couette flow is a typical research target for investigating the fluid me-
chanical phenomena in shear flows. It features the flow of an incompressible,
viscous fluid contained in the gap between two concentric cylinders at low
Reynolds numbers (see references [1] and [2]). The Couette apparatus has
originally been developed for measuring the viscosity of a fluid at small im-
posed angular velocities of the cylinders. A simple sketch of the test case
geometry is shown in figure 1.
Fig. 1. Sketch of a Taylor-Couette reactor (taken from [3]).
In the present study, we will concentrate on the case where the outer
cylinder (of radius ro ) is immovable and the inner cylinder (of radius ri ) is
rotating. It will be assumed that there is no axial flow. The shear stresses
on the surface of the inner rotating cylinder entrain the fluid to rotate. Cen-
trifugal forces then lead to vortical structures, their complexity depends on
the rotational speed, domain geometry and fluid properties. A wide variety of

644 Thomas Nierhaus et al.
flow regimes, depending on the Reynolds number Re = 2πΩi ri (ro − ri )/ν, is

known to exist.
• At small angular velocities Ωi of the inner cylinder, a purely azimuthal
circular Couette flow occurs.
• If the angular velocity exceeds a critical value of Rec1 = 102, the flow
becomes unstable to axisymmetric perturbations. The radial and axial
velocity components grow in time, leading to a flow pattern consisting of
superposed counter-rotating cells [1].
• When the angular velocity is further increased, a second instability occurs,
which is characterized by the superposition of an azimuthal wave over the
initial cells. This instability occurs at Rec2 = 138.4 [2].
• At Reynolds numbers above Rec3 = 1023, the azimuthal wave is modulated
by a second wave mode.
• Rotation at even higher speeds gradually induces breakup of the afore-
mentioned structures, eventually leading to fully turbulent flow regimes.
We perform Direct Numerical Simulation (DNS) of Taylor-Couette flow laden
with gas bubbles. In recent years, bubbly flows have received great attention
from numericists in order to acquire better understanding of the interactions
between the dispersed phase and the continuous carrier flow. They occur in
in a wide variety of industrially relevant situations, such as electrochemical
processes in reactors characterized by turbulent flows inside geometries similar
to the present one.
2 Numerical Method
The modelling of the physical phenomena in bubble-laden Taylor-Couette

flow is achieved using the Eulerian-Lagrangian formulation. This approach
describes the continuous liquid phase by the incompressible Navier-Stokes
equations, while the trajectories of the gas bubbles are tracked sequentially.
2.1 Primary Phase Flow
The velocity field u of the primary phase flow in the reactor is obtained by
DNS of the incompressible Navier-Stokes equations. For the present simula-
tions, we write them in a cylindrical coordinate system, which is the most
natural for flow problems in axisymmetric geometries.
∂rvr ∂vθ ∂vz
+ +r = 0, (1)
∂r ∂θ ∂z
∂rvz r ∂p ∂ ∂vz 1 ∂ 2 vz ∂ 2 vz
+ − ν( (r )+ + r )
∂t ρc ∂z ∂r ∂r r ∂θ2 ∂z 2
DNS of bubbly Taylor-Couette flow 645

∂vz ∂vz ∂vz
= − rvr + vθ + rvz , (2)
∂r ∂θ ∂z
∂rvr r ∂p ∂ 1 ∂rvr 1 ∂ 2 vr 2 ∂vθ ∂ 2 vr

+ − ν(r ( )+ 2
− +r 2 )
∂t ρc ∂r ∂r r ∂r r ∂θ r ∂θ ∂z

∂vr ∂vr ∂v r
= − rvr + vθ − vθ2 + rvz , (3)
∂r ∂θ ∂z
∂rvθ 1 ∂p ∂ 1 ∂rvθ 1 ∂ 2 vθ 2 ∂vr ∂ 2 vθ

+ − ν(r ( )+ 2
+ +r 2 )
∂t ρc ∂θ ∂r r ∂r r ∂θ r ∂θ ∂z

∂vθ ∂vθ ∂vθ
= − rvr + vθ + vr vθ + rvz . (4)
∂r ∂θ ∂z
Here, ρc is the density of the continuous fluid phase, ν is the kinematic
viscosity of the fluid and p is the pressure. In a cylindrical coordinate system,
a point is referenced by its axial (z), radial (r), and azimuthal (θ) coordinates.
The numerical framework used for the channel flow simulations is the hybrid
spectral/finite element solver SFELES [4]. The Navier-Stokes equations are
discretized in z-r-planes using P1 linear finite elements, while the azimuthal
direction is represented by means of a truncated Fourier series. The period-
icity in the azimuthal direction naturally leads itself to a spectral expansion.
The Crank-Nicholson scheme is used to integrate the pressure and viscous
terms in time while the Adams-Bashforth method is used for the convective
terms. For the finite element discretization of the continuity equation, the
Pressure-Stabilized Petrov-Galerkin (PSPG) dissipation is used. To eliminate
convective instabilities, a fourth order Laplacian dissipation term is used to
stabilize the momentum equation.
The 3D problem in physical space is by this methodology transformed into
a system of N/2 loosely coupled 2D problems in Fourier space. For reasons of
computational efficiency, the convective terms are first computed in physical
space and then transformed to Fourier space (pseudo-spectral approach). The
2D linear problems are solved sequentially on a distributed memory parallel
computer using MPI.
2.2 Lagrangian Tracking of Bubbles
We state Newton’s second law as a Lagrangian equation of motion for each

bubble according to the notation in [5]:

∂v 3 CD ρc 1 ρc Du dv
= |u − v| (u − v) + −
∂t 4 d ρd 2 ρd Dt dt
ρc Du ρc
+ + CL (u − v) × ω, (5)
ρd Dt ρd
where ρd is the density of the dispersed phase. The terms on the right
hand side are, in order of appearance, the Stokes drag force, the virtual mass
force, the force due to the fluid pressure gradient and viscous stresses and the
slip-shear lift force. In the lift force term, ω = ∇ × u denotes the vorticity
of the fluid. We neglect gravitational effects on the bubbles, because from [8]
we see that they are not playing a major role in the flow domain of interest.
Following empirical correlations for a fluid bubble, the drag coeffcient CD is
calculated according to [6]. The lift coefficient is set to CL = 0.5 throughout
our investigations.
Equation 5 and the relation v = dx/dt form a system of six ordinary dif-
ferential equations in three space dimensions, which is solved for each bubble
using the Crank-Nicholson scheme. Sequential tracking of all bubbles in the
system is performed at each time step of the Navier-Stokes solver. Momen-
tum back-coupling from the dispersed to the continuous phase is neglected,
because the ratio of momentum between the phases is very small [7]. Also no
back-coupling through the continuity equation has been included.
The Lagrangian solver adapts the geometical partitioning strategy of the
Navier-Stokes solver (see section 2.1), to which it is coupled via standardized
data interfaces. This allows to distribute the tracking of the bubbles to the
parallel partitions according to the partitioning of the computational domain.
The parallel architecture of the Lagrangian solver is realized through MPI.
3 Results
The test case geometry and flow conditions are chosen to match the experi-
mental setup from [8], where glycerol bubbles of ρd = 4.12kg/m3 were pro-
duced by cavitation in a mixture of water-glycerol of ρc = 1088kg/m3 and
ν = 1.7 · 10−5 m2 /s. The inner and outer cylinder radii are ri = 30mm and
ro = 35mm while the reactor hight is h = 110mm. There is a fixed non-slip
wall at the bottom of the reactor while the top is formed by a free surface
where the pressure is fixed. The computational grid consists of 80×220 points
in the radial and axial directions, while 64 Fourier modes are used in the
azimuthal direction. The grid is refined near the reactor walls.
At the start of the simulation, the fluid is at rest and the inner cylinder is
put in rotation. Due to shear stress, the azimuthal velocity propagates to the
outer cylinder. Counter-rotating vortices propagate from the top and bottom
to the half-height of the cylinder. Since the reactor oprates at Re > Rec3 ,
these Taylor vortices become wavy. This trend is more pronounced near the
upper free surface, because the bottom wall limits the axial flow modulations
while the upper free surface does not limit them. An average in time and over
the azimuthal angle direction of the flow field is presented in figure 2. The
wavy Taylor vortices are clearly visible in figures 2a and 2b. Once the Taylor
cells are established, 70 000 bubbles of diameter d = 0.5mm are injected into
the flow at random initial positions. The bubbles start to follow the flow
DNS of bubbly Taylor-Couette flow 647
in the azimuthal direction. Due to centrifugal and lift forces, they migrate
toward the center of the Taylor vortices. After 100 inner cylinder rotations,
the averaged volume fraction in time and azimuthal direction shown in figure
2c is obtained. The bubbles are accumulated inside the vortices as expected.
The concentration ragions are broader with increasing axial direction, because
the wavy vortices are located in the upper part of the reactor.
(a) Azimuthal velocity uθ
(b) Radial velocity ur
(c) Dispersed phase volume fraction αd
Fig. 2. Averaged flow field in time and in the azimuthal direction.
Comparing our results with the experiments described in [8], we see from
figures 3a, 3b and 3c that the bubbles naturally collapse to form bubble rings
around the cylinder. These rings strongly modify the flow field by reducing
the number of Taylor cells. It is obvious that this phenomenon cannot be
reproduced numerically by Lagrangian tracking of the bubbles (see result
in figure 3d), since it fails the dispersed entities hypothesis. Phase coupling
through the continuity equation, as it is applied in [9] and has been neglected
in the present work, could lead to a better prediction of the flow field in this
case.
(a) Re = 357 (b) Re = 459 (c) Re = 1122 (d) Simulation
Fig. 3. Experimental results from [8] compared to simulation.

Conclusion
DNS of Taylor-Couette flow laden with gas bubbles has been applied. The
main point to retain from this numerical test case is that the forces acting
on the bubbles are modeled effectively and experimental observations can
be qualitatively reproduced. The accumulation of gas in the center of the
Taylor vortices modifies the flow field by reducing the number of Taylor cells,
a phenomenon which cannot be predicted with the numerical method used.
Phase coupling through the continuity equation could lead to better results.
Acknowledgements
We acknowledge the support from the Instituut voor de aanmoediging van

innovatie door Wetenschap en Technologie in Vlaanderen (IWT) for the
MuTEch project (contract nr. SBO 040092).
References
1. Taylor, G.I.: Stability of a viscous liquid contained between two rotating cylin-
ders. Philos. Trans. R. Soc. London, A223, 289–343 (1923)
2. Andereck, C.D., Liu, S.S., Swinney, H.L.: Flow regimes in a circlular couette sys-
tem with independently rotating cylinders. J. Fluid Mech., 164, 155–183 (1986)
3. Krivilyov, M., Detandt, Y., Vanden Abeele, D., Degrez, G., Fransaer, J.: Direct
numerical simulations of Taylor-Couette flow using a hybrid spectral/finite ele-
ment approach. In: 7th National Congress on theoretical and applied Mechanics,
Mons, Belgium (2006)
4. Vanden Abeele, D., Degrez, G., Snyder, D.O.: Parallel turbulent flow computa-
tions using a hybrid spectral/finite element method on Beowulf clusters. In: 3rd
International Conference on Computational Fluid Dynamics, Toronto, Canada
(2004).
5. Maxey, M.R., Riley, J.J.: Equation of motion for a small rigid sphere in a nonuni-
form flow. Phys. Fluids, 26, 883–889 (1983).
6. Lain, S., Bröder, D., Sommerfeld, M., Göz, M.F.: Modelling hydrodynamics and
turbulence in a bubble column using the Euler-Lagrange procedure. Int. J. Mul-
tiphase Flow, 28, 1381–1407 (2002).
7. Nierhaus, T., Vanden Abeele, D., Deconinck, H., Planquart, P.: Modelling of tur-
bulent bubble-laden flow in a parallel plate electrochemical reactor. In: 13th In-
ternational Conference on Fluid Flow Technologies, Budapest, Hungary (2006).
8. Djeridi, H., Gabillet, C., Billard, J.Y.: Two-phase Couette-Taylor flow: Arrange-
ment of the dispersed phase and effects on the flow structures. Phys. Fluids,
16(1), 128–139 (2004).
9. Ferrante, A., Elghobashi, S.: On the physical mechanisms of drag reduction in a
spatially developing turbulent boundary layer laden with microbubbles. J. Fluid
Mech., 503, 345–355 (2004).
Numerical simulation of unsteady flow inside
an impulsively started liquid drop
M. Krivilyov and J. Fransaer
KU Leuven, Dept. MTM, 3001 Leuven, Belgium

mikhail.krivilev@mtm.kuleuven.be, jan.fransaer@mtm.kuleuven.be
1 Introduction
In this work, the transient flow around a liquid drop after injection in a qui-
escent fluid is calculated. This kind of fluid flow occurs in many natural phe-
nomena and technological processes. Numerical simulation of this problem is
important for industrial applications like atomization, ink-jet printing, powder
metallurgy, electrochemistry and the behavior of emulsions. The description
of such flows is essential if fluid flow is coupled with phase transformations
as is the case during the rapid solidification of atomized metal drops. The
scientific background on this subject goes back to the works of Stokes, and
Hadamard and Rybczynski and to the more recent works of [1, 3, 2]. However,
the majority of this work is devoted to the study of the steady flow around a
drop. The first goal of this article is to model and analyze the transient flow
inside a drop. The second goal is to find a formula for the characteristic time
of the internal transient flow useful for practical purposes.
2 Model
Suppose a fluid of viscosity µext and density ρext is moving unsteadily around
a spherical drop, having viscosity µint and density ρint . Then the governing
equations of the model are the Navier-Stokes equation (NSE) and the equa-
tion of motion. The dimensionless time-dependent NSE written in terms of a
stream function Ψ
E2Ψ E2Ψ

4 1 2 ∂Ψ ∂ ∂Ψ ∂ ∂
E Ψ + sin θ Re 2 2 − 2 2 = (E 2 Ψ ),(1)
2 ∂r ∂µ r sin θ ∂µ ∂r r sin θ ∂t
subject to
1 U 2
{e
vr , veθ } < M at r = 0; Ψ →− r (1 − µ2 ) as r → ∞; (2)
2 Uss

650 M. Krivilyov and J. Fransaer
vr = 0, ver = 0, vθ = veθ , τrθ = τerθ at r = 1;

Ψ = Ψ0 , Ψe = 0 at t = 0
describes the problem of unsteady fluid flow around a liquid drop. The super-
script ˜ is used for functions inside a drop and there is no superscript for the
outer flow. Here t is time, r and θ are the radial and angular coordinates, U is
∂2 1−µ2 ∂ 2
the instantaneous drop velocity, µ = cos θ, E 2 = ∂r 2 + r 2 ∂µ2 is the Stokes
operator, Reext = 2aUss /νext and Reint = 2aUss /νint are the external and in-
ternal Reynolds numbers (a – drop radius, Uss – steady-state drop velocity,
ν – kinematic viscosity), vr = r12 ∂Ψ∂µ and vθ =
√ 1 2 ∂Ψ
∂r are the radial and
r 1−µ
tangential velocities, τrθ is the shear stress. The stream function is scaled by
Uss a2 , velocity by Uss , the radial coordinate by a and time by a2 /ν.
The instantaneous velocity of the drop is found from the Newton’s equa-
dU
tion of motion: (mint + madd ) = Fgrav + Fbuoy + Fdrag + Fhist , where mint
dt
and madd are the drop and added masses, F are the forces acted on the drop.
Incorporating the history force into the generalized drag force and express-
ing the forces through parameters of the drop, the dimensionless equation of
motion in scalar form becomes:

dÛ 1 ρ̂ − 1 Reext 3 1
= − Reext Û 2 CD , (3)
dt 2 ρ̂ + 12 Fr 16 ρ̂ + 12
2
where the superscript ˆ denotes dimensionless parameters, Fr = Uss /ag is the
Froude number, CD is the drag coefficient, ρ̂ = ρint /ρext is the density ratio.
(3) is subject to Û (t = 0) = Û0 where Û0 is the dimensionless injection
velocity.
For numerical modeling, the stream function is expanded through Leg-
endre polynomials and (1) is solved using the Van Dyke’s series-truncation
technique by the method of weighted residuals. Hermite polynomials are used
as the interpolations functions and the Crank-Nicolson scheme as the integra-
tion method [2]. The nonlinear system of equations is solved by the Newton-
Raphson method. The equation of motion is integrated by a second order
Runge-Kutta algorithm. The initial regime of flow immediately after injec-
tion is characterized by a thin boundary layer near the drop surface. In order
to describe this initial phase correctly, a boundary layer transformation of
space and time coordinates is implemented in the model [1]. Such coordinate
transformation is important for drops injected with a high initial velocity U0
and can be skipped for problems with small U0 .
3 Results
In the calculations, the physical properties of both phases are considered to

be constant. The drop is assumed to be spherical, the Reynolds number varies
Unsteady flow inside a liquid drop 651
between 0.01−100 and there is no internal flow at t = 0. In order to benchmark

our code, the steady-state fluid flow around a liquid drop moving with a
constant velocity was first calculated. The flow patterns, separation point,
drag coefficient and pressure distribution were compared with literature data
[1, 2, 3] and good agreement was found for all parameters. Verification of
the time-dependent solution was done in the creeping flow limit Reext →
0, for which an analytical solution exists [4]. This validation confirmed the
correctness of the developed model.
The flow around a drop depends on three characteristic numbers: the
Reynolds number Re, viscosity ratio µ̂ = µint /µext and density ratio ρ̂ = ρint /ρext .
In this work, three test cases were calculated: (1) µ̂, ρ̂ 1 (i.e. motion of a gas
bubble in water), (2) µ̂, ρ̂ ∼ 1 (a silicon oil droplet in water), (3) µ̂, ρ̂ 1 (a
metal drop in helium during atomization). The structure of the external flow
depends on µ̂ and Re, Fig. 1. At µ̂ < 1 the external flow is almost equal to the
creeping flow solution. At µ̂ > 1 the flow outside the drop resembles either
the creeping flow solution (Re < Rec ) or is separated into the free-stream and
wake parts (Re > Rec ). The critical Reynolds number Rec is a function of µ̂
and increases when µ̂ decreases (Rec = 20 for the solid sphere and Rec → ∞
as µ̂ → 1). Interaction of the liquid drop with the external flow leads to de-
velopment of internal flow. The internal steady flow is always represented by
a Hill vortex. The Hill vortex is shifted, depending on µ̂, towards the front
critical point in comparison to the Hadamard-Rybczynski solution.
(a) µ̂ = 0.02, ρ̂ = 10−3 (b) µ̂ = 5, ρ̂ = 0.98 (c) µ̂ = 240, ρ̂ = 2.4 × 105
Fig. 1. Flow patterns at fixed Reext = 100 and different values of µ̂ and ρ̂.
If µ̂ 1, the model predicts an interesting transition inside the drop. The

streamlines around a drop injected inside a stagnant fluid are shown in Fig. 2
at Re = 100, µ̂ = 240 and ρ̂ = 2.4 × 105 . Immediately after injection, the flow
structure is similar to the inviscid potential flow and thin boundary layers are
observed in both phases near the drop surface. The width of the boundary
layer inside the drop depends on µ̂ and decrease as µ̂ increases. Then two Hill
vortices appear inside the drop, Fig. 2b-d. (Figures Fig. 2b and Fig. 2c show
only location of the vortex centers marked by small crosses, Fig. 2d gives also
the isolines). The first vortex is centered near the meridional plane and is
driven by friction acting from the free-stream flow on the drop. The second
vortex is located at the back of the drop and is driven by the return flow in the
wake behind the drop. As the flow inside the drop accelerates with time due
to action of shear stresses on the surface, the ratio of the inertial to viscous
forces grows, which leads to the decay of the small vortex, Fig. 2e-f.
(a) t = 0 (b) t = 0.05 (c) t = 0.27
(d) t = 1.5 (e) t = 61 (f) t = 430
Fig. 2. Transient flow around a liquid drop at different dimensionless times t after
injection. The simulation is performed at Re = 100, µ̂ = 240 and ρ̂ = 2.4 × 105 .
4 Analysis
The transition from the two-vortices to single-vortex regime occurs if two con-
ditions are fulfilled. Firstly, the Reynolds number should be Re > Rec in order
to achieve separation of the external flow. Secondly, the viscosity of the drop
should be µ̂ 1. Under these conditions, the external flow reaches its steady
regime in time τext ∼ 1. The total time τint required for the internal flow to
reach its steady regime is much bigger: τint τext . The internal flow is better
described by the reduced internal Reynolds number Re∗int = Reint (vint max
/Uss ),
max
where vint is the actual (time-dependent) velocity maximum inside the drop.
Then the transient two-vortices regime is observed if Re∗int . 1. Increase of
Re∗int with time leads to the decay of the small vortex. The transition from
the two-vortices to a single vortex regime can be characterized in terms of
Unsteady flow inside a liquid drop 653
1.2
0.8
int / f(µ)
0.6
vavg
0.4
0.2
Gas bubble
Oil drop
Metal drop
0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
(! / µ) t
avg
Fig. 3. Volume-averaged dimensionless internal velocity vint as a function of di-
mensionless time t, where f (µ) = µ if µ > 1; f (µ) = 1 if µ ≤ 1.
the stream function: its first derivative (∂Ψ/∂r)/(1 − µ2 )1/2 becomes nonneg-
ative at r = 1. This implies that the tangential velocity becomes vθ > 0 over
the drop surface. The results reported early in [3] are not fully confirmed in
the present study, i.e. we did not observe the internal flow break-up in the
steady-state regime.
The development of the internal flow is evaluated in terms of the bulk
avg
average internal velocity vint as a function of time. Results for all test cases
are presented in Fig. 3. The curve for the gas bubble shows a smooth increase
avg
of vint with time and the transitional regime takes τint = 0.1 (ρ̂/µ̂) dimen-
sionless time units. Transition times for the oil and metal drops are about
0.2 (ρ̂/µ̂) and 0.35 (ρ̂/µ̂). Hence, the dimensional characteristic time τint can
be approximated by the formula
ρ̂ a2
τint ∼ M , M = 0.1..0.5. (4)
µ̂ νext
The correspondence between the calculated transition time and this formula
is reasonable, taking into account that τint differs by 5 order of magnitude for
drops with different µ̂ and ρ̂. The dimensional average internal velocity in the
steady state is estimated by
 Uss , if µ̂ > 1,

avg
vint = µ̂ (5)
Uss , if µ̂ < 1.

Conclusions
The transient flow inside an impulsively started liquid drop is analyzed by
means of a numerical model. The steady flow inside a drop is always a single
Hill vortex. In case of a highly viscous drop, a transitional flow regime with
two vortices is predicted, where the smaller vortex decays with time. This
effect can be important in the rapid solidification of metal droplets. A formula
was found for the transition time of the internal flow that gives reasonable
approximation to the calculated ones.
This work was supported by the European Space Agency under the ESTEC
Contract No. 14980/01/NL/SH. We thank Dr. T. Volkmann, Dr. M. Kolbe
and Prof. D. Herlach from DLR for helpful discussions.
References
1. S.C.R. Dennis and J.D.A. Walker. Numerical solution for time-dependent flow
past an impulsively started sphere. Phys. of Fluids., 15(4):517–525, 1972.
2. D.L.R. Oliver and J.N. Chung. Flow about a fluid sphere at low to moderate
Reynolds numbers. J. Fluid Mech., 177(4):1–18, 1987.
3. V.Ya. Rivkind and G.M. Ryskin. Flow structure in motion of a spherical drop in
a fluid medium at intermediate Reynolds numbers. Fluid Mechanics, 11(1):5–12,
1976. Translated from Russian.
4. F. Sy and E.N. Lightfoot. Transient creeping flow around fluid spheres. AIChE
Journal, 17:177–191, 1971.
Accurate and Efficient Solution of 2D Steady
Water Flows with Surface Waves and
Turbulence
Jeroen Wackers1 and Barry Koren2

1
CWI, P.O. Box 94079, 1090 GB Amsterdam, The Netherlands.
Jeroen.Wackers@cwi.nl
2
CWI and Delft University of Technology, Faculty of Aerospace Engineering, P.O.
Box 5058, 2600 GB Delft, The Netherlands.
Barry.Koren@cwi.nl
Summary. A surface capturing method is developed for steady water–air flow with
gravity. Second-order accuracy is obtained with flux limiting and turbulence is mod-
eled with Menter’s model. The equations are solved efficiently with a combination
of multigrid and defect correction. Results for two test cases confirm the efficiency
and accuracy of the method.
1 Introduction
Numerical simulation of steady water flow plays a continuously increasing role
in the development of ships and offshore structures. Accurate predictions of a
ship’s wave pattern, for example, can help in reducing the ship’s drag during
the design, and thus minimise operating costs [3].
A key choice in the development of a numerical model for water flow is the
representation of the water surface. The most widely used approach is surface
fitting: the computational mesh is deformed during the computation, such that
it coincides with the water surface. Another approach is the surface capturing
technique (the volume of fluid and level set methods are well-known examples)
where the grid is fixed, but the location of the water surface on the grid can
change. These methods are more flexible than surface fitting methods; they
allow steeper waves and more complex object geometries. A disadvantage is
that, for steady flow problems, the equations resulting from classical surface
capturing methods are hard to solve efficiently. Currently, they are usually
solved by time-marching the unsteady flow equations to convergence, which
is a time-consuming process.
In previous work, we have shown that efficient solution of a surface capturing

model is possible [7]. We developed a volume-of-fluid model based on conserva-

656 Jeroen Wackers and Barry Koren
tion laws only, which is solved with classical multigrid. Very fast convergence
of the solution process was obtained. But it also became clear that the first-
order, laminar flow model was not sufficient to accurately compute realistic
water flow.
Therefore, the method is improved by making it second-order accurate
and by adding a turbulence model. The second-order model is solved with a
defect-correction procedure, that uses the efficiency of the first-order multigrid
method. Turbulence is modeled with Menter’s model [5]: a simple yet accurate
one-equation turbulence model that has already been applied in some ship-
flow solvers.
This paper briefly describes the first-order accurate discretisation and focuses
on the extension to second-order accuracy and on the inclusion of the tur-
bulence model. The flow equations are given in Sect. 2, the discretisation of
the different terms in Sect. 3. Section 4 describes the defect correction and
the special multigrid method that are used to solve the difficult two-fluid flow
equations with turbulence. Finally, the practical application of the method
is illustrated in Sect. 5 with a test case: the water flow in a channel with a
bottom bump.
2 Flow Equations
The flow equations used here are based on the Reynolds–averaged Navier–
Stokes (RANS) equations. The distinction between the water w and air a is
made by adding a mass conservation equation for the water. Turbulence is
modeled with Menter’s model. The resulting system has conservation laws
only, which makes it suitable for multigrid solution.
∂ ∂ ∂
p + %u2 +

(%uv) = ((µ + µT ) 2ux ) +
∂x ∂y ∂x
∂
((µ + µT ) (uy + vx )) (x–mom.),
∂y
∂ ∂ ∂
p + %v 2 =

(%uv) + ((µ + µT ) (uy + vx )) +
∂x ∂y ∂x
∂
((µ + µT ) 2vy ) − %g (y–mom.),
∂y
(1)
∂ ∂
(u) + (v) = 0 (tot. mass),
∂x ∂y
∂ ∂
(uα) + (vα) = 0 (water mass),
∂x ∂y
∂ ∂ ∂
(ν̃T u) + (ν̃T v) = ((ν + ν̃T ) (ν̃T )x ) +
∂x ∂y ∂x
∂
((ν + ν̃T ) (ν̃T )y ) + P − D (Menter),
∂y
Solution of Steady Water Flows with Waves and Turbulence 657
where α is the volume fraction of water, ν = µ/% and ν̃T is a scaled version
of νT [5]. P and D are turbulence production and dissipation terms, based on
velocity derivatives.
3 Discretisation
The system (1) is discretised with a cell-centered finite-volume technique
on structured, curvilinear grids. The convective and diffusive fluxes and the
source terms are discretised separately, to independently control the stability
of each part. Both first- and second-order accurate fluxes are used. The deriva-
tion of the first-order fluxes is given in [7]. These fluxes are briefly described
here, together with a description of the second-order accurate fluxes and the
turbulence source terms.
Convective fluxes The convective fluxes are computed by reconstructing

the states q = [u, v, p, α, ν̃T ]T on the left and right side of the cell faces from
the states in the cell centres. These states are then put in a flux function, that
gives the flux over the cell face. For the first-order accurate fluxes, the state
at the cell faces is taken equal to the state in the cell centres. For the second-
order accurate fluxes, the states are reconstructed with a slope limiter. We
use the κ = 13 limiter proposed by Koren [4]. The flux function itself is defined
in [7]. It is an approximate Riemann solver, based on artificial compressibility
and comparable to the flux used in [2].
Diffusive fluxes The diffusive fluxes in the momentum equations are mod-
eled with central differences. These are both stable and second-order accu-
rate, therefore they are used for both the first- and the second-order accurate
discretisation. For non-cartesian grids and for the cross-diffusion terms, the
velocity derivatives are computed with a control volume approach.
Source terms Two different source terms appear in the system (1). The
gravity force in the y–momentum equation is discretised as −%i g ∆x ∆y. The
production and dissipation terms in the turbulence equation contain both first-
order and second-order derivatives of the state variables, which are computed
with finite differences in a local cartesian coordinate system. When the grid is
smooth and locally (close to) orthogonal, then this approach is second-order
accurate. It is used for both the first-order and the second-order accurate
schemes.
4 Multigrid
For the solution of nonlinear systems, multigrid is one of the most powerful
techniques known today. Its application to fluid flow solutions is mature and
well studied [6]. The multigrid principle is to apply a simple solution technique
or smoother to the solution on the finest grid and to copies of this solution
on underlying coarser grids, to eliminate high-frequency and low-frequency
errors, respectively.
Choice of smoother When multigrid is used to solve RANS equations, it is

usually combined with time stepping. Either a time step smoother is used, or
multigrid is used to compute the individual steps in an implicit time stepper.
This procedure is robust and stable.
We are not going to use time stepping for two reasons. First, we think that
the fastest solution techniques are obtained when the steady flow equations
are solved directly with multigrid. And second, time stepping is inefficient as a
smoother for flow with gravity waves, because these flows have high Reynolds
numbers and transient waves that damp out very slowly in time. Therefore, we
solve the discretised flow equations (1) with direct multigrid and alternating
line Gauss–Seidel smoothing.
Linear multigrid The usual multigrid technique for nonlinear systems is

nonlinear multigrid. Here, discretised flow equations are constructed on the
coarser grids in exactly the same way as on the finest grid. Then these flow
equations are smoothed, using the residual in the solution on the finer grids
as a source term.
For the two-fluid RANS equations, this approach gives two problems. One,
the two-fluid flow equations cannot handle large arbitrary source terms [7].
And two, for nonlinear multigrid to work well, the flow equations on the
coarser grids must resemble the flow equations on the finest grid. Here, this is
not the case, because the volume fraction α is discontinuous and because the
Menter source terms are sensitive to the grid resolution.
We overcome these problems by switching to linear multigrid. The flow
equations on the finest grid are linearised around the current solution and
this linearised operator is copied directly to the coarser grids. These linear
coarse grid operators can be solved for any source term and they always
resemble the fine grid operator reasonably well. Thus, on the coarse grids, we
do not compute full solutions q but (small) corrections. We have chosen this
unusual combination of linear multigrid on the coarser grids with nonlinear
line smoothing on the finest grids, because the nonlinear smoother is very
robust. It can correct small unphysical solutions (ν̃T < 0 etc.) that arise from
the linear coarse grid corrections.
A disadvantage of the method is that it requires much computer memory.
Also, Galerkin operators can reduce the convergence rate of multigrid for
convective flows ([6], Sect. 5.4). However, in practice this effect was observed
but it caused no problems.
On the other hand, the method can solve flows whose features cannot
be resolved accurately on coarse grids. Another advantage is that the coarse
grid corrections are cheap. The linear line smoothing is less expensive than
nonlinear line smoothing and the computation of the linear operator takes
little extra work, as the linearisation is already needed for the line smoothing
on the finest grid.
Full multigrid As initial condition, we choose ν̃T very close to zero. Then,
when the solution process is started, the turbulence intensity has to grow.
Typically, we see the residual in the turbulence equation increase at first,
because the turbulence intensity increases. Then, when the boundary layers
have more or less developed, it starts to decrease. Experiments show that
multigrid relaxation is of little use in the first stage.
Therefore, full multigrid (FMG) is essential to our method: we find the
initial condition on the finest grid by computing solutions on the coarser grids,
from coarse to fine. Thus, all but the first computations start at the second
stage, with developed boundary layers. The time-consuming development of
turbulence only occurs on the coarsest grid.
Defect correction To compute second-order accurate solutions, we use de-

fect correction. In an iterative process, the residual of the second-order accu-
rate discretisation is computed and used as a source term in the first-order
multigrid smoother. Theoretically, this process converges to the second-order
solution. In practice, defect correction does not always converge. But when
the process is started from the first-order solution, a great improvement is ob-
tained in the first iterations and the solution is usually second-order accurate
after a few steps.
For efficiency, we perform the multigrid step in the defect correction with
the linearised operators, even on the finest grid. These operators are not up-
dated during the process, since they are stable and close enough to the con-
verged solution to give good smoothing.
5 Results
As a test case, the flow in a channel with a bottom bump is computed. Exper-
imental results for this test are given by Cahouet [1]. The flow has a Froude
number F r = 2.05 and a Reynolds number Re = 1.9 × 105 (based on inflow
water height). The top wall is modeled as a slip wall, the bottom is a no-slip
wall. The bump has a thickness of 44% of the water height. The curvilinear
grid has 128 × 512 cells and is compressed near the boundary layer and the
water surface.
In the velocity plot (Fig. 1a), we see low-velocity regions near the leading
and trailing edge of the bump and high-velocity regions above the bump and
in the air region near the top wall. The turbulent viscosity (Fig. 1b) shows
the boundary layer. Figure 1c gives a comparison of the second-order solution
with Cahouet’s measurements. Overall, the agreement is very good. The slight
difference at the back of the wave may be caused by the more developed
% # #
$'*
#$* )'(
$')
#$%
# )
"
"
"
#
$'(
&$* $'(
$'%
#
#
&$'
'
&$
& $ $
!% & % ( ) !# $ # % & !# $ # % &
!
a) !
b) c) !
Fig. 1. Cahouet test case, F r = 2.05. Speed (a), turbulent viscosity ν̃T (b), and a
comparison (c) of the second-order α = 0.5 isolines (—) with Cahouet’s experiment
(averaged wave height, - -).
) #
! !
!) !#
)$ (" $# )"%
!" !$
2/34-,156
!( 1.23,+045 !(
!# !'
!' !"
!$ !&
!& !%
! "! #! $! %! ! "! #!! #"! $!!
*+,-.+*/01 )*+,-*)./0
Fig. 2. Cahouet test case, convergence of the sum of the residual for multigrid
(FMG on four grids, left) and line Gauss–Seidel smoothing on a single grid (right).
boundary layer in Cahouet’s experiment, as Cahouet’s tunnel has a longer

inflow length than our domain.
The bottom bump solution is computed on four grids, the multigrid con-
vergence is given in Fig. 2. We see bad convergence on the first two grids.
Apparently the boundary layer on the second grid is so different from the first
grid that boundary layer development is needed on the second grid too. The
convergence on the last two grids is acceptable.
Compared with a solution on a single grid, the number of (expensive)
fine-grid iterations is reduced by a factor 12. The total computation time is
reduced by a factor 8. The current method does not converge so fast in difficult
locations, like stagnation points, but we believe this can be solved by simple
fixes: therefore, the convergence can likely be improved even further.
6 Conclusion
The computational results for the flow in a channel with a bottom bump, show
that the method is efficient. Comparison with experimental results indicates
that it is also accurate.
Acknowledgement. This research was supported by the Dutch government

through ICT project BRICKS (http://www.bsik-bricks.nl).
References
1. Cahouet, J.: Etude numérique et expérimentale du problème bidimensionnel
de la résistance de vagues non-linéaire. Technical report 185, Ecole Nationale
Supérieure de Techniques Avancées, Paris (1984)
2. Dick, E., Linden, J.: A multigrid method for steady incompressible Navier-Stokes
equations based on flux-difference splitting. Int. J. Num. Meth. Fluids 14, 1311–
1323 (1992)
3. Hino, T. (ed) CFD Workshop Tokyo 2005. National Maritime Research Insti-
tute, Tokyo (2005)
4. Koren, B.: A robust upwind discretization method for advection, diffusion and
source terms. In: Vreugdenhil, C.B., Koren, B. (eds) Numerical Methods for
Advection-Diffusion Problems. Vieweg (1993)
5. Menter, F.R.: Eddy viscosity transport equations and their relation to the k − ε
model. J. Fluids Engineering, 119, 876–884 (1997)
6. Trottenberg, U., Oosterlee, C.W., Schüller, A.: Multigrid. Academic Press
(2001)
7. Wackers, J., Koren, B.: A surface capturing method for the efficient computation
of steady water waves. To appear in J. Comp. Appl. Math.
Numerical analysis of bubble migration in
thermocapillary flows of an open cylindrical
container
Hiroaki Ohira1 , Satoshi Matsumoto1 , Takashi Mashiko1 , Shinichi Yoda1 ,

and Yasuhiro Kamotani2
1
ISS Science Project Office, Japan Aerospace Exploration Agency, 2-1-1, Sengen,
Tsukuba, Ibaraki, 305-8505 Japan ohira.hiroaki@jaxa.jp
2
Department of Mechanical Aerospace Engineering, Case Western Reserve
University, 10900 Euclid Avenue, Cleveland, Ohio, 44106-7222 U.S.A.
1 Introduction
Thermocapillary flow is driven by temperature-induced surface tension

variations along a liquid free surface. The experiments and the calculations
for the flows in open cylindrical containers were carried out both in space and
on ground. The detail phenomena have been made clear from these investi-
gations. In 1995, however, numbers of air bubbles were merged incidentally
in the cylindrical container and moved along the flows while the Oscillatory
Thermocapillary Flow Experiment 2 (OTFE-2) was carried out on the Second
United States Microgravity Laboratory (USML-2). These bubbles gradually
moved to the flow centers and finally all the distances between two bubbles
became almost the same. The behavior of such different substances in the
thermocapillary flow is considered to affect the new material processing in
space. Hence the experiments and the simulations to investigate the behav-
ior should be carried out both in space and on ground. These bubbles are
estimated to be driven by the thermocapillary flow in the container and also
by Marangoni forces caused on the bubble surfaces. Numerical investigations
have never been reported for these complex phenomena. The purpose of this
study is to develop a calculation model by applying Level Set method[1] and to
evaluate the interaction between the thermocapillary flows and these bubble
migrations.

Numerical analysis of bubble migration 663
2 Numerical methods
2.1 Governing equations
In the present calculations, mass, momentum and energy equation for the
incompressible flow were applied to the thermocapillary flows and the bubble
migrations as shown in Eqs. (1), (2) and (3). These equations were described
in nondimensional forms:
∇ · u = 0, (1)

∂u Pr 2 Pr 1
ρ + (u · ∇)u = −∇p + µ∇ u − ∇k T − κn δε , (2)
∂t Ma Ma We
∂T α 2
+ (u · ∇)T = ∇ T, (3)
∂t Ma
where u, p and T are nondimensional velocity, pressure, temperature and
average curvature, respectively. The above properties of ρ, µ and α are density,
dynamic viscosity and thermal diffusivity, respectively, which are normalized
by liquid properties. The nondimensional numbers of M a, P r and W e are
Marangoni number, Prandtl number and Weber number, respectively. ∇k T
in Eq. (2) is a temperature gradient in the tangential direction on the bubble
surfaces.
2.2 Level set method for bubble surface
Level-Set method (LSM) developed by Sussman, et al.[1] were applied to

simulate the bubble surface: The distant functions from the bubble surface to
every grid point, F , were advected along the streamline of the surrounding
fluid, as shown in Eq. (4):
∂F
+ (u · ∇)F = 0. (4)
∂t
The algorithm for solving these equations was Generalized Simplified Marker
And Cell method for Finite Element Method (GSMAC-FEM) developed by
Tanahashi et al.[2].
The discontinuous change of the physical properties on the bubble surface
was smoothed by the approximate Heaviside function Hε as follows;


 Hε (F ) = 0 F < 0
 1 F 1 πF
Hε (F ) = 1 + + sin −ε ≤ F ≤ ε (5)

 2 ε π ε

Hε (F ) = 1 F >0
where ε was the smoothing bandwidth. The approximate Delta function δε
was obtained by differentiating the Heaviside function and combined with the
surface forces as shown in Eq. (2).
664 Hiroaki Ohira et al.
Free surface
Heater 1.0
z 0.0
0.1 1.0
r
Fig. 1. 2D mesh arrangement of an open cylindrical container
Additionally, in order to reinitialize the distant function, we also solved

Eq. (6) by Cubic Interpolated Pseude-particle (CIP) scheme developed by
Yabe et al.[3] together with GSMAC-FEM to resolve the surface in fine accu-
racy;
∂F ∇F F0
+ (w · ∇)F = S(F0 ), w = S(F0 ) , S(F0 ) = p 2 , (6)
∂t |∇F | F0 + ε2
where τ was the quasi-time and F0 was the initial distant function. Since the
interface is positioned implicitly in the level set method, the total volume of a
fluid may fluctuate during the reinitialization process. The following equation
was solved at all nodes in order to preserve the initial volume.
V (t) − V0
F = Fr + , (7)
S(t)
where V (t) and S(t) were the calculated bubble volume and area, respectively.
Fr was the distant function calculated after the reinitialization and V0 was
the initial bubble volume.
3 Numerical conditions
3.1 Calculation model
Figure 1 shows the 2 dimensional mesh arrangement of the cylindrical

container. The normalized radius and the height were both 1.0 and the cylin-
drical heater whose normalized radius was 0.1 was installed in the center of
the container. The fluid region was divided into 150 × 2 × 150 elements in the
radius (r), azimuthal (θ) and height (z) direction, respectively. The bubble
radius R was 0.06 and the initial position (r0 , z0 ) was (0.7, 0.7).
M a, P r and W e number were defined as M a ≡ σT (TH − TL )R/(µl αl ),
P r ≡ νl /αl and W e ≡ ρl u20 R/σ, where TH and TL was the heater and the side
wall temperature, respectively, u0 was defined as u0 ≡ σT (TH − TL )/µl , ν was
the kinematic viscosity and σT is temperature coefficient of surface tension.
In above equations, subscript l means the liquid phase. In the present study,
two kinds of M a number were applied to the steady thermocapillary flows as
shown in Table 1. Case 1 conditions were almost similar to the experiments
and the calculations reported by Kamotani, et al.[4] to verify the present nu-
merical method for thermocapillary flows. Since the oscillation was measured
in Case 1 condition, the smaller M a number of Cases 2 and 3 were applied to
the bubble calculations. We applied two assumptions for the bubble surface
tension; the constant surface tension (Case 2) and the temperature dependent
surface tension (Case 3): We applied Eq. (2) in Case 3 for the bubble migra-
tion, while we neglected the fourth term of the right hand side in Eq. (2) in
Case 2.
Table 1. Nondimensional numbers

Case Ma Pr We
5
1 1.3 × 10 30.0 Not used
2,3 6.4 × 104 30.0 1.2 × 102
3.2 Boundary conditions
The nondimensional temperature of the heater and the side wall were
1.0 and 0.0, respectively. The top surface was assumed to be undeformable in
the present calculations. All the boundary surfaces except for the free surface
were set as no-slip conditions, while the velocities on the free surface were
given in Cartesian forms as shown in Eq.(8);
∂u ∂T ∂v ∂T
=− , =− . (8)
∂z ∂x ∂z ∂y
4 Results and discussions
4.1 Steady thermocapillary flow
Non-dimensional temperature and velocity distribution on the free sur-

face calculated in Case 1 condition are shown in Fig.2 (a) and (b), respectively.
The calculated steady state 2-dimensional results reported by Kamotani, et
al. [1] were also plotted in these figures. The present temperature and velocity
distribution agreed well with the reported distributions.
(a) (b)
Fig. 2. (a)Temperature and (b)velocity distribution in the radial direction on the

free surface
4.2 Modeling of Level set method
Figures 3(a) and 3(b) show the normal and the tangential components
of the bubble surface tension in Case 3 at 386,000 steps after the bubble
calculation was started (point 0 A0 in Fig.4). The tangential components were
calculated in the 3rd term of the right hand side of Eq.(2), while the normal
ones were calculated in the 4th term. From these figures, the forces acted on
the bubble were well calculated by the present LSM model. The gas velocity
profile observed on the bubble at the same step is also shown in Fig. 3(c). The
liquid was observed to flow from the front region of the bubble, while the gas
in the bubble was observed to have two eddy centers and it flowed along the
bubble surface. The above results show the LSM was properly modeled in the
present algorithms.
(a) (b)
(c)
Fig. 3. (a) Normal and (b) tangential components of bubble surface tension and
(c) liquid and gas velocity profile observed on the bubble. All figures were plotted
at 386,000 steps after the bubble calculation was started
1.0
Non-dimensional height
0.9
A
Case 2
B
0.8
Case 3
Initial position
0.7
(0.7, 0.7)
0.6
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Non-dimensional radius
Fig. 4. Two kinds of traces for the bubble centers
4.3 Bubble migration in thermocapillary flow
The traces of the bubble center in Cases 2 and 3 are shown in Fig.4.
The bubbles flowed along the thermocapillary flow and gradually moved to
the initial eddy center, which described as the point 0 B 0 . The eddy centers
were moved from 0 B 0 to the left hand side region when the bubble migrations
were started. These phenomena were similar in the two cases, however, the
bubble in Case 3 approached closer to the heater surface than that in Case 2
because of the Marangoni effect on the bubble. The bubble velocity in Case
3 became much smaller near the free surface and much larger near the heater
compared with the velocities in Case 2. Finally, the bubbles in the two cases
continued to flow around one flow center. From these results, it became clear
that the bubble migration and the thermocapillary flow affected each other,
and Marangoni effects on the bubbles also affected the bubble migration and
its effects could not be negligible. The movie of the bubble migration in Case
2 was presented in the conference.
Temperature and velocity profiles in the radial direction are shown in
Figs. 5(a) and 5(b), respectively. These profiles were calculated in Case 2 at the
point 0 C 0 in Fig. 4. The temperature around the bubble surface became lower
since the gas temperature of the bubble was lower than the surrounding liquid
1.0 0.10
(a) (b)
Non-dimensional temperature
Non-dimentional velocity
0.8 0.08
0.6 0.06
0.4 0.04
0.2 0.02
0.0 0.00
0.2 0.4 0.6 0.8 1.0 0.2 0.4 0.6 0.8 1.0
Non-dimensional radius Non-dimensional radius
Fig. 5. (a)Temperature and (b)velocity profiles on the point C in Fig.4

temperature. The temperature gradient near the bubble surface also became
larger than that of the other surface region. The velocity in the region between
the free surface and the bubble surface became larger because of its narrower
flow area. This larger velocity was estimated to cause the bubble acceleration
and the deformation of the thermocapillary flow.
5 Conclusions
In order to evaluate bubble migrations in thermocapillary flows in open
cylindrical containers, a calculation model was developed by applying LSM
to GSMAC-FEM and 2 dimensional calculations were performed in the mi-
crogravity conditions. The following conclusions were obtained from these
calculations:
(1) The bubble and the thermocapillary flow were affected each other and
consequently the bubble migration paths were decided,
(2) The Marangoni forces acted on the bubble surface affected the bubble mi-
gration in the thermocapillary flows and its effects could not be negligible
in the initial bubble migration.
We are also going to calculate 3 dimensional phenomena and evaluate the
interaction between more than two bubbles in the thermocapillary flows.
References
1. Kamotani, Y., Ostrach, S. and Masud, J.: Microgravity experiments and analysis
of oscillatory thermocapillary flows in cylindrical containers. Journal of Fluid
Mechanics 410, 211–233 (2000)
2. Sussman, M., Smereka, P. and Osher, S.: A level set approach for computing
solutions to incompressible two-phase flow. Journal of Computational Physics,
114, 146–159 (1994)
3. Tanahashi, T., Okanaga, H. and Sano, T.: GSMAC finite element method for
unsteady incompressible Navier-Stokes equations at high Reynolds numbers. In-
ternational Journal for Numerical Methods in Fluids, 11, 479–499 (1990)
4. Yabe, T. and Wang, P.: Unified numerical procedure for compressible and in-
compressible fluid. Journal of The Physical Society of Japan, 60, 2105–2108
(1991)
5. Ohira, H., Matsumoto, S., Mashiko, T., Yoda, S. and Kamotani, Y.: Numerical
analysis of bubble migration in thermocapillary flows of an open cylindrical
container. 25th International Symposium on Space Technology and Science, June
4–11, CD-ROM, 2006
Part VIII
Hypersonic and supersonic flow

Numerical investigation of an effusion cooled
thermal protection material
Volker Hannemann
DLR, German Aerospace Center, Institute of Aerodynamics and Flow Technology,

Bunsenstrasse 10, 37073 Goettingen Volker.Hannemann@dlr.de
Introduction
A spacecraft re-entering the earth’s atmosphere has to endure extreme heat

loads on parts of its structure as a result of the surrounding hypersonic flow
field. The design of such parts benefits from a close coupling of structural
and fluid mechanics. A good prediction of the vehicle’s surface temperature
is crucial for an optimal choice of materials to fulfill the technical require-
ments. The numerical analysis of the flow field with a CFD (Computational
Fluid Dynamics) code has to take into account the high temperature effects
occurring in hypersonics (like chemical reactions and vibrational excitation in
thermo-chemical non-equilibrium) and to provide a local heat flux in the gas
phase at a given surface temperature. The CSM (Computational Structure
Mechanics) analysis of the structure is necessary to provide the temperature
distribution and to determine the exact shape of the structure. The coupled
CFD-CSM analysis, reveals the ability of the structure’s thermal conductivity
to diminish and broaden local temperature peaks predicted by CFD alone.[1]
[2] [3]
The current investigation adds to the above scenario by including the
porosity of a thermal protection material (C/C-SiC) and its usage for an
active cooling of the structure. The future goal of a complete coupled analysis
of flow field and structure including an effusion cooled porous medium is
attacked from the CFD side. The modifications of the CFD code to enable a
coupling with a CSM simulation of a porous media are discussed. A simplified
model replacing the CSM input is introduced to check the functionality of the
interface and to study the influence of the coolant mass flow on the flow field
topology. The calculations are related to experiments conducted in the DLR
wind tunnel L3K.

672 Volker Hannemann
Target configuration
The generic geometry as depicted in figure 1 is a blunted wedge with a cooled
copper nose part and a radius of 10 mm. On the plain wedge surface the
porous C/C-SiC part is embedded in non permeable C/C-SiC followed by
isolation (KAPYROK). The plain surface has an angle of attack of −30◦ The
free stream flow conditions are: Mach number 7.6, pressure 50 Pa, temperature
491 K, velocity 3730 m/s and a composition of the gas mixture in mass frac-
tions YN 2 = 0.757, YO2 = 0.012, YN O = 0.018, YN ≤ 10−6 , YO = 0.213. These
conditions lead to a laminar and except of the boundary layer almost chemi-
cally frozen flow. Pure nitrogen is used as a coolant. The surface temperature
distribution is of mayor interest, because it is measured in the experiments by
optical techniques. The investigation is focused on the steady state behavior.
Coupling model
For given distributions of the effusion mass flux and the wall temperature
the CFD calculates the wall pressure and heat flux distributions. The CSM
provides the local effusion mass flux and wall temperature for given pressure,
heat flux and known reservoir conditions summarized in table 1. An iterative
exchange between both tools at the surface of the porous medium should
converge to a steady state solution at least with a good starting solution and
sufficient under relaxation.
Table 1. Coupling model
flow field porous medium coolant reservoir

CFD CSM
Pwall mc Pres
qwall Twall Tres
composition
The flow field analysis was performed with the DLR-TAU code [4] [3],
which is a finite volume solver for the compressible, multi species Navier Stokes
equations. The implementation of the coupling interface in TAU is done inside
the boundary condition for a viscous wall. The values of a conservative variable
on a boundary can either be iterated by closing the respective conservation
equation with an adequate flux or by setting the value directly. The given
mass flux is included in the mass balance and is assumed to enter the flow
field normal to the wall, so the tangential components of the velocity are still
zero at the wall. The momentum of the gas is set according to the mass flux,
in other words the wall normal velocity is set equal to the given mass flow
Numerical investigation of an effusion cooled thermal protection material 673
divided by the iterated local density. The given temperature determines the
inner energy of the gas at the wall. The kinetic energy of the gas due to the
mass flux is determined by one half of the mass flux squared and divided by
the iterated density. For the conditions of interest in this investigation the
kinetic part stays below 0.5% of the inner energy.
Simplified model
The simplified model assumes a given mass flux distribution and iterates an
equilibrated surface temperature. Similar to an adiabatic wall or a wall in
radiative equilibrium the heat flux within the structure is neglected except the
part transfered to the coolant. This approximation is quite useful for free flight
investigations, where no heat sink is within the vehicle, and determines an
upper limit for the predicted peaks of the surface temperature. The heat flow
into the structure (normal component) will diminish the surface temperature,
where as the tangential heat flow spreads the peaks, i.e. decreases the highest
temperatures by increasing the surrounding temperature level. In wind tunnel
test models the internal heat flow can become quite complicated (cooled model
parts, instrumentation) but still neglecting the heat flow within the structure
gives a good first impression.
The mass and momentum conservation equations are closed at the bound-
ary as described in the last section. Without a prescribed temperature the
energy equation has to be closed by an expression for the energy flux across
the wall. The energy flux from the fluid onto the wall consisting of a thermal
and a diffusion part is in this simplified model equalized by the radiation part
and a part due to the coolant flow. The radiation heat flux qrad is modeled for
the simple convex geometry under investigation by the Stefan-Boltzmann law
qrad = σT 4 with the Stefan-Boltzmann constant σ and emissivity factors of
Kapyrok = 0.95 and C/C−SiC = 0.8. At least for low wall temperatures the
4 4
back ground radiation has to be taken into account: qrad = σTwall − σTbg
with a background temperature Tbg usually set to room temperature for wind
tunnel investigations.
The part due to coolant flow can be split into one part which is transported
together with the coolant through the surface (into the adjacent control vol-
ume) and another part is added to account for the cooling of the porous media.
To model the first part it is assumed that the coolant leaves the porous media
with the wall temperature and surface pressure, so the coolant mass flux mc
enters the control volume with the local enthalpy H of the flow field. For the
second part the cooling is set equal to the difference of the coolant energy ec
before and behind the porous media times the mass flux. Together the influ-
ence of the cooling gas is modeled as: qc = mc {H − ec (Twall ) + ec (Tres } where
Tres denotes the temperature of the coolant in the reservoir before entering
the porous medium.
Results
To gain a first impression over the parameter space of interest, the investiga-
tion was conducted on an appropriate fixed grid. Except the new parameters
of the simplified model all other parameters were kept constant, which means
a 5 species air model in chemical non equilibrium with reaction rates accord-
ing to Gupta et al. [5] and a thermodynamic model of the species including
coupled electronic excitation, vibration and rotation (the last two only for
molecules). Blottner curve fits are used for the species viscosities and Eu-
cken’s correction for the thermal conductivity. A constant diffusion coefficient
is assumed for all species connected to the viscosity via a constant Schmidt
number of 0.7. The copper nose is assumed to be full catalytic at a constant
temperature of Twall = 500K, where as all other materials are assumed to be
non catalytic at radiative or radiative plus effusion equilibrium temperature.
The mass flux distribution is assumed to be constant due to the homogenous
porous media of constant thickness and the negligible small pressure gradient
along the surface (≈ 103 Pa/m) compared with the strong gradient within the
porous media (≈ 108 Pa/m).
The highest investigated coolant mass flux of mmaxc = 0.354 kg/(sm2 ) is
about 37% of the free stream value or about 8 % of the post shock value out-
side the boundary layer. In the following all effusion mass fluxes are related
to this value. In figure 1 the colors indicate the Mach number contours with-
out effusion mass flux and the black contour lines show the same contours
of the flow field with the maximal effusion mass flux. At this high effusion
mass flux the boundary layer thickens and separates generating a compres-
sion strong enough to change the shape of the leading shock. Figure 2 shows
the temperature distributions along the wedge surface for different effusion
mass flow rates. The dashed lines indicate the changing of the surface ma-
150
100
y [mm]
50
-50
-50 0 50 100 150

x [mm]
Fig. 1. Mach number contours without effusion mass flux (colors) versus Mach
number contours with effusion mass flux (black contour lines)
Numerical investigation of an effusion cooled thermal protection material 675
1200
1000
T[K]
800
600
400
0.05 0.1 0.15 0.2
x[m]
Fig. 2. Wall temperature along the wedge at different mass flow rates: red/plus
without mass flow, black/cross with 1 mmax
c , green/square with 0.6 mmaxc , or-
ange/triangle with 0.2 mc , violet/diamond with 0.1 mmax
max
c , blue/circle with
0.01 mmax
c
terial. Starting at 500 K on the cooled nose, the temperature rises on the
isolator part and goes to even higher temperatures on the first C/C-SiC part
due to higher emission abilities (radiation cooling) of the isolator. Without ef-
fusion cooling (red line/plus), the temperature stays at a high level but with
a small decrease with increasing distance due to the growing temperature
layer and a small drop when reaching the isolator at the end. At low mass
flow rates of 0.01 mmax
c (blue/circle) the temperature stays nearly at the same
level as without cooling. With a moderate coolant flow 0.1 − 0.2 mmax c (vio-
let/diamond, orange/triangle), the temperature decreases significantly on the
effusion cooled part, then rises again on the non permeable C/C-SiC part and
shows the same small drop on the final isolator part. Between 0.3 − 0.4 mmax c
boundary layer separation starts, so that at the high effusion mass flow rates
0.6 (green/square) and 1 mmax
c (black/cross) the temperature is influenced far
ahead of the effusion cooled zone. In case of the high effusion mass flow rates
the wall temperature on and behind the cooled zone decouples from the flow
field such that the wall temperature becomes dependent only on the coolant
temperature and the (background) radiation. In the temperature distributions
at moderate mass flow rates a small peak is visible at the beginning of the
cooled zone. These peaks as well as the very small cooling effect for the low
mass flow rate originate from an increase of the heat flux due to diffusion.
When blowing air as a coolant into air instead of N2 into air the concentra-
tion gradients diminish and the temperature becomes smaller as can be seen
in picture 3.
1280
1260
1240
T[K]
1220
1200
1180
0.05 0.1 0.15 0.2
x[m]
Fig. 3. Closeup view on the wall temperature at different mass flow rates: red/plus
without mass flow, blue/circle with 0.01mmax
c , green/star again 0.01mmax
c but air
instead of N2 as coolant
Concluding Remarks
A simplified one dimensional model for an effusion cooled wall was imple-
mented into the DLR-TAU code. It neglects tangential mass or heat flows
inside the porous medium as well as any heat flux leaving the porous medium
on the back side (adiabatic inner wall). The surface temperature is iterated via
equilibrating the radiation heat flux together with a simple energy gain of the
coolant versus the thermal and diffusion heat fluxes from the flow field. Un-
der these assumptions the model shows the influence of the effusion mass flow
rate on the flow field topology (separation prediction) and the effectiveness of
different coolants due to diffusion.
References
1. Mack, A., Schaefer, R., Esser, B., Guelhan, A.: Fluid Structure Interaction on a
Hypersonic Generic Body-Flap Model. Proceedings of ICCFD3, Toronto (2004)
2. Mack, A., Schaefer, R.: Fluid Structure Interaction on a Generic Body-Flap
Model in Hypersonic Flow. AIAA Journal of Spacecraft and Rockets, Vol. 42,
No.5, (2005)
3. Mack, A.: Analyse von heissen Hyperschallstroemungen um Steuerklappen mit
Fluid-Struktur Wechselwirkung. DLR-FB 2005-23 and PhD, TU Braunschweig
(2005)
4. Mack, A., Hannemann, V.: Validation of the Unstructured DLR TAU-Code for
Hypersonic Flows. AIAA 2002-3111, (2002)
5. Gupta, R.N., Yos, J.M., Thompson, R.A., Lee, K.-P.: A Review of Reaction Rates
and Thermodynamic and Transport Properties for an 11-Species Air Model for
Chemical and Thermal Nonequilibrium Calculations to 30 000 K, NASA Refer-
ence Publication 1232, (1990)
Calculation of Transport Properties
for Entry into the Martian Atmosphere
Johannes Baumgart1,2 , Tobias Leicht1 , Thierry Magin2 , Paolo Barbante2 ,

Pietro Rini2 , Gérard Degrez2 , and Roger Grundmann1
1
Institut für Luft- und Raumfahrttechnik, Technische Universität Dresden, 01062
Dresden, Germany, baumgart@tfd.mw.tu-dresden.de
2
von Karman Institute for Fluid Dynamics, 72 Chaussée de Waterloo, 1640
Rhode-Saint-Genèse, Belgium
1 Introduction
The computation of entry problems into a planetary atmosphere strongly re-
lies on the models used to describe the thermal and chemical nonequilibrium
effects in the flow, the radiation phenomenom, and the surface chemistry.
Moreover, the solution of the flow field in the continuum approach also re-
quires an accurate description of the transport properties, the main focus of
this work. The transport properties (diffusion coefficients, viscosity, thermal
conductivity) are described in terms of collision integrals obtained here from
three intermolecular potential models (Tang–Toennies, Born–Mayer and (m,6)
potentials). The Martian atmosphere is essentially made of carbon dioxide
(95.7%), plus some nitrogen (2.7%) and argon (1.6%). A CFD computation is
carried out to investigate the influence of nitrogen and argon on the convective
heat flux.
2 Collision integral
At high temperatures, the macroscopic transport properties cannot be directly
measured. However, they can be derived in kinetic theory from the Boltzmann
equation. They are thus linked to the microscopic scale by collision integrals
(l,s)
Q̄ij that represent energy-averaged cross sections Q(l) of a pair of colliding
particles i and j
Z∞
(l,s) 2 (l + 1) −E (l)
Q̄ij (T ) = h i exp E s+1 Qij dE.(1)
(s + 1)! 2l + 1 − (−1)
l
(kT )s+2 kT
0
where E is the relative kinetic energy, T the gas temperature, k the Boltz-
mann constant, the indices l and s are related to the spectral method used

678 Johannes Baumgart et al.
to solve the Boltzmann equation. The cross-section can be either measured

experimentally or computed from the deflection angle χ of the collision. In
the latter case, it reads
Z∞
(l)
1 − cosl (χ) b db,

Qij (E) = 2 π (2)
0
in which b is the impact parameter, distance between the asymptotic trajec-

tories of both particles before or after collision. The deflection angle reads
Z∞
dr
χ (E, b) = π − 2 b p . (3)
r2 1 − ϕ(r)/E − b2 /r2
rmin
where ϕ(r) is the intermolecular potential. During an interaction, the distance

r between the colliding particles varies from infinity to rmin , the distance of
closest approach. A collision is described based on the intermolecular potential
of interaction between the particles. The parameters of the potential can be
determined from experiments.
The Tang–Toennies potential currently represents the most adequate ana-
lytical potential describing neutral–neutral and ion–neutral interactions over
the entire range of separation distance between two colliding particles [4]
∞
−br
X C2n
ϕ(r) = Ae − f2n (r) , (4)
n=3
r2n
where A and b are the parameters of the repulsive part of the potential (short
range) and C2n the dispersion coefficients for the attractive part (long range).
We propose to modify the semi-empirical damping function f2n by adding an
additional parameter z
2n
X (zbr)k
f2n (r) = 1 − e−zbr . (5)
k!
k=0
This parameter yields a better description of the collision integral at high

temperatures where repulsive effects are dominant. For rare gases, a value
of z = 1.3 is found adequate for a temperature range from 102 to 105 K.
The polarization coefficients can be used to calculate the dispersion coeffi-
cients and the well location for the Tang–Toennies potential. Here it is not
recommended to determine the well location based on the (m,6) potential,
since the (m,6) potential is empirical and just approximates this region. Al-
though this approach is an estimation, the final is reasonably accurate. As
an example the Ar–Ar collision integral with different parameter sets for the
potential are shown in Fig. 1. The corresponding potentials are presented in
Fig. 2. The (m,6) potential is suitable to compute the collision integrals in the
low temperature range. The Born–Mayer potential is a reference for the high
temperature range. The collision integral at all temperature can be obtained
Transport properties for Martian Atmosphere 679
m!6
60
Born!Mayer
!"2#* (1,1) (10!20 m2) TT polarization
TT original
40 TT modified
20
0
2 3 4
10 10 10
Temperature (K)
Fig. 1. Collision integral for Ar–Ar by use of (m,6) potential from Mourits and
Rummens [6], Born–Mayer potential from Cubley and Mason [3], Tang–Toennies
potential with coefficients obtained from Patil [8], given by Tang and Toennies [9]
and a modified case of them.
6
10 m!6
potential (eV)
3 Born!Mayer
10 TT polarization
TT original
0
10 TT modified
0.01
0.005
0
!0.005
!0.01
1 2 3 4 5 6 7 8 9 10
distance (m!10)
Fig. 2. Intermolecular potential for Ar–Ar based on (m,6) potential from Mourits
and Rummens [6], Born–Mayer potential from Cubley and Mason [3], Tang–Toennies
potential with coefficients obtained from Patil [8], given by Tang and Toennies [9]
and a modified case of them.
from a smooth fit between the values obtained with both types of potential.
Another way to comptute the collision integral is to use a potential valid at
all temperatures, for instance the Tang–Toennies potential with parameters
given in the literature or by estimation of them from the polarization. The
Tang–Toennies potential based on the polarization data does not allow for a
correct description of the collision integrals at all temperatures, since it is just
an estiamtion. The Tang–Toennies potential given by Tang and Toennies [9]
fails at high temperature (more than 2000 K), due to the fact that the repul-
sive part is not well represented. Finally, resonable integrals are obtained, if
the damping function of the Tang–Toennies potential is modified according
to (5). The collision integrals are solved by using a Gauss-quadrature method
and taking into account the singularity of the integrand. Frequently, it is not
possible to retrieve the correct parameters for an intermolecular potential.
For instance, combination rules can be applied for interactions between un-
like particles.
Once the intermolecular potential is known, the collisions integrals can be
solved and the transport properties can be computed. As an example, the
total thermal conductivity is presented in Fig. 3 for different gas mixtures.
The nominal mixture of Mars atmosphere (so-called MARS15) is composed
of CO2 95.7%, N2 2.7% and Ar 1.6%. During an entry, temperature rises,
molecules dissociate, recombine, and molecules and atoms ionize. Our mixture
is composed of 15 species CO2 , N2 , Ar, O2 , C, NO, O, N, CO, Ar+ , C+ ,
N+ , NO+ , O+ , e− . Three other mixtures are also presented: an 8-species
carbondioxide mixture (CO2 , O2 , C, O, CO, C+ , O+ , e− ), a 4-species nitrogen
mixture (N2 , N, N+ , e− ) and a 3-species argon mixture (Ar, Ar+ , e− ). The
Mutation library is used to compute the thermodynamic and the transport
properties based on collision integrals [5]. The thermal conductivity of the four
mixtures are shown in Fig. 3 for a gas in local thermodynamic equilibrium. It
includes three contributions: diffusion, translation, and Eucken’s correction for
the internal energy. The value of the thermal conductivity of carbon dioxide
is close to the one of the nominal mixture.
6
MARS15
5 CO2
N
2
! (W m!1 K!1)
4
Ar
3
0
0 5000 10000 15000
Temperature (K)
Fig. 3. Total thermal conductivity for martian mixture MARS15, pure N2 and Ar
at a pressure of 10 kPa
Transport properties for Martian Atmosphere 681
3 CFD simulation
The influence of the nitrogen and argon compounds on the convective heat
flux is investigated in a simulation of a Martian entry for a gas in chemical
nonequilibrium by means of a finite-volume axisymmetric code solving the
Navier-Stokes equations [1]. In our simulation, ionization is neglected: first a
pure carbon dioxide mixture (C, O, CO, O2 and CO2 ) is chosen, then for the
Martian mixture Ar, N, N2 and NO are added to the previous mixture.
The computations are carried out for the NASA MESUR (Mars Environ-
mental Survey, also called Pathfinder) geometry and flight conditions for the
peak heat load. Chen and Candler [2] performed a Navier–Stokes calculation
with surface catalysis and a CO2 –N2 gas mixture. The boundary conditions
given therein are used. The chemistry model considered in this work is taken
from Park [7]. So far a single temperature model is applied.
The freestream condition of Chen and Candler are used for an altitude
of 41.67 km, where the temperature is 160.9 K and the density 0.2687×10−3
kg/m3 for a mixture of CO2 and N2 . The velocity of the vehicle is 6155 m/s.
At the wall a non catalytic boundary condition is considered for the chemistry.
Radiation of the wall into the farfield is accounted for with an emissitivity co-
efficient of 0.9 assuming that the flow is optically transparent. The radiation of
the wall is computed with the Stefan–Boltzmann law. The velocity boundary
condition is a no-slip wall.
The surface heat flux, here shown in Fig. 4 for the non catalytic wall, is
400
Heat flux (kW m!2)
300
200
MARS
CO2
100 Chen and Candler, Parks kinetic
Chen and Candler, McKenzies kinetic
0
0 0.2 0.4 0.6 0.8 1
Distance from stagnation point (m)
Fig. 4. Surface heat flux for CO2 and MARS mixtures and non catalytic wall
compared to [2] with to different chemistry models
qualitativly comparable to the wall temperature. The difference between both

mixtures is within about 10 kW/m2 .
Conclusion
We have proposed a modified potential of Tang–Toennies valid at all tem-
peratures. We have computed the collision integrals to obtain the transport
properties of the Martian atmosphere. The results obtained in local thermo-
dynamic equilibrium or in the CFD computation did not exhibit significant
influence of argon and nitrogen on the convective heat flux. However, it should
be taken into account that some minor species such as CN or C2 may have a
strong influence in the prediction of the radiative heat flux.
References
1. P.F. Barbante. Accurate and efficient modelling of high temperature nonequilib-
rium air flows. PhD thesis, von Karman Institute for Fluid Dynamics, Rhode-
Saint-Genèse, Belgium, 2001.
2. Y.-K. Chen and G. V. Candler. Navier-stokes solutions with surface catalysis
for martian atmospheric entry. Journal of Spacecraft and Rockets, 30(1):32 – 42,
1993.
3. S.J. Cubley and E.A. Mason. Atom-molecule and molecule-molecule potentials
and transport collision integrals for high-temperature air species. The Physics
of Fluids, 18(9):1109 – 1111, 1975.
4. E. Levin and M.J. Wright. Collision integrals for ion-neutral interactions of
nitrogen and oxygen. Journal of Thermophysics and Heat Transfer, 18(1):143–
147, 2004.
5. T. Magin. A Model for Inductive Plasma Wind Tunnels. PhD thesis, von Karman
Institute for Fluid Dynamics, Rhode-Saint-Genèse, Belgium, 2004.
6. F.M. Mourits and F.H.A. Rummens. A critical evaluation of Lennard-Jones
and Stockmayer potential parameters and of some correlation methods. Can. J.
Chem., 55:3007–3020, 1977.
7. C. Park, R.L. Jaffe, and H. Partridge. Chemical-kinetic parameters of hyperbolic
Earth entry. Journal of Thermophysics and Heat transfer, 15(1):76–90, January-
March 2001.
8. S.H. Patil. Thomas-fermi model electron density with correction boundary con-
ditions: applications to atoms and ions. Atomic Data and Nuclear Data Tables,
71:41–68, 1999.
9. K.T. Tang and J.P. Toennies. The van der waals potentials between all the rare
gas atoms from He to Rn. Journal of Chemical Physics, 118:4976–4983, 2003.
Numerical simulation of supersonic flow
around a double ramp configuration and
correlation with experiment
Fedorchenko I.A.,1, 2 Fedorova N.N.,1 Kharlamova Yu.V.,1 Gaisbauer U.,2

and Kraemer E.2
1
Khristianovich Institute of Theoretical and Applied Mechanics SB RAS,
Institutskaya, 4/1, 630090, Novosibirsk, Russia nfed@itam.nsc.ru
2
Institut fuer Aerodynamik und Gasdynamik, Pfaffenwaldring 21, 70569
Stuttgart, Germany irina.fedorchenko@iag.uni-stuttgart.de
Summary. The aim of this work is to find out optimal geometrical parameters of
a double ramp configuration in a supersonic flow in order to attain the minimum
value of pressure drag. The computations are provided by an URANS-based code
that implements several turbulence closures. The experiments were carried out at
IAG, Germany. The results of the numerical research for Mach numbers M=2.51 and
M=2.995 and for different Reynolds numbers are presented, along with a comparison
of the numerical data with the experimental ones. The obtained information allowed
to disclose three flow regimes, viz., regime with a big separation zone, with a two-
shock flow structure and a transitional regime. An optimal ramp length for each of
the flow cases has been estimated. The verifications of the computations performed
show a reasonable ability of URANS-based applications to predict the structure and
the wall pressure behavior of the detached flow.
1 Introduction
Nowadays the use of high-accuracy models such as DNS, LES and DES for
the simulation of engineering turbulent flows is confronted with high com-
puting costs. That way, advanced RANS-based models seem to be the most
feasible choice to solve the problem. This is one of the reasons for intensive
turbulence model design. Due to the semi-empirical nature of these models,
they demonstrate different levels of obtained accuracy for different applica-
tion areas. Therefore it is believed to be an urgent and important task to
verify the numerical results on a base of experimental data for a wide range
of applications.
The present work is subjected to a supersonic double ramp configuration.
From an engineering point of view the research is conditioned by a growing
interest in the development of scramjet facilities. The double ramp is a compo-
nent of a scramjet inlet used for the so-called external compression upstream

684 Fedorchenko I.A., Fedorova N.N. et al.
of the intake itself. The objectives of the investigation are experimental and
numerical studies of the aerodynamical characteristics of a scramjet intake
flow.
The problem offers a certain level of complexity since the configuration
has two compression angles resulting in a shock system. Thus, depending
on the geometry of the configuration, separation can occur. In order to get
minimal pressure drag along with the compression needed, conditions for a
short separation zone have to be attained.
Hence this investigation presents two aspects of the problem. From an
engineering point of view the parameters, structure and regimes of the flow
are studied. On a numerical point of view the ability of URANS methods to
produce a correct solution for detached turbulent flows is shown.
Part of this work is carried out in the frame of the Research Training
Center 1095/1 financed by DFG.
2 Formulation of the Problem and Solution Methods

2.1 Initial conditions
A sketch of the investigated problem and the boundary conditions are pre-
sented in Fig. 1. The calculations were conducted jointly with experimental
research [2] for three cases of flow parameters. The flow conditions of the ex-
periments are presented in Table 1. The angles of the first and second ramp
kinks are α1 = 11◦ and α2 = 9◦ , respectively. The distance d between the two
kinks is varied from 0 to 39 mm in calculations as well as in the experiment.
Upper boundary
"Simple wave" boundary conditions
Outlet
boundary
Zero second derivative conditions

for all parameters
Inlet boundary
Freestream conditions
Boundary
layer edge
Profiles of all parameters

obtained by a turbulent
boundary layer approach Wall
Non-slip boundary
conditions for velocity
Adiabatic conditions
for temperature
Fig. 1. Formulation of the boundary conditions
2.2 Numerical technique
The numerical investigation is carried out with the original 2D code developed
at ITAM SB RAS, Novosibirsk, Russia [1]. The Favre-averaged compressible
Numerical simulation of supersonic flow around a double ramp 685
Table 1. Flow parameters of the investigated problems
Case number, N Mach number Re, [1/m] δ, [mm] T0 , [K] P0 , [hPa]

1 2.540 12.7·106 5.07 288 1200
2 2.513 9.82·106 5.15 283 934
3 2.995 7.50·106 9.35 293 960
Navier-Stokes equations for ideal gas are used. Several turbulence models
for closure are implemented such as the k-ω Wilcox model [5], the BSL and
the SST models by Menter [4]. The numerical technique includes the TVD-
scheme of the Flux Vector Splitting method by van Leer with third order of
accuracy for the approximation of convective terms and second-order central
finite-difference relations for the approximation of viscous terms. The four-step
implicit finite-difference scheme of splitting according to the space directions
realized by scalar sweeps is used for the time approximation.
A regular grid with a toward-the-wall refinement is constructed. Typically
the grid consists of 100∇ · 200 nodes in y-direction and 200∇ · 400 nodes in
x-direction.
All computational results presented below have been obtained with the
aid of k-ω model unless otherwise stated.
3.1 Case 1
A comparison of the numerical and experimental wall pressure distributions

for the case of M=2.54 and Re=12.7 · 106 1/m is presented in Fig. 2. Here, the
solid lines represent the numerical solution, and the symbols correspond to the
experimental data. A good agreement for various distances d is indicated. The
numbers on the lines stand for the following distances between the kinks: 1 –
d = 0 mm, 2 – 5 mm, 3 – 8 mm, 4 – 10 mm, 5 – 12 mm, 6 – 16 mm, 7 – 20 mm,
8 – d = ∞. The obtained skin friction distributions shown in Fig. 3 allow us
to estimate the separation lengths for each case. On the base of data analysis
three flow regimes have been disclosed [3]. The first one is a coupled regime
where the separation zones appearing due to the two shocks are combined in
a common large separation. The second regime is a transitional one with a
short detached zone at the first angle and a zero skin friction coefficient area
downstream. The third regime results in the generation of two spaced shocks,
while a short separation in the vicinity of the first kink and an attached flow
at the second kink are realized. On the basis of the obtained data it can be
concluded that a ramp length of d = 16∇ · 20 mm is optimal as it guarantees
the minimal pressure drag value.
Fig. 2. Experimental and calculated wall pressure distributions for the case 1
2 8
1.5
7
1
Cf 10^3
6
3
0.5
4
1
0 5
2
-0.5
-5 0 5 10
x/delta
Fig. 3. Calculated wall skin friction coefficient distributions for the case 1
3.2 Case 2
The results of the computations for Mach number M=2.51 and a lower
Reynolds number Re = 9.82 · 106 , shown in Fig. 4 for the various distances be-
tween the ramp kinks, exibit that similar flow regimes are to be seen. The line
numbers correspond to the following distances: 1 – d = 0 mm, 2 – 9.43 mm,
3 – 12.36 mm, 4 – 18.336 mm, 5 – 24.26 mm, 6 – 30.23 mm, 7 – 34.66 mm, 8 –
39.24 mm. The calculated pressure distributions for this case also agree quite
well with the experimental data. However for the distances corresponding to
the transitional regime a slight data disagreement can be seen. It means that
in the computations the transition to the two-shock regime takes place at a
smaller d. This fact can be explained by an incorrect computational turbulent
viscosity level, which in the k-ω turbulence model is essentially influenced
by a freestream value of the specific turbulence dissipation rate ω. For the
two-shock regime the predicted pressure behind the interaction zone is a bit
higher than in the experiment. A possible reason of this disagreement is an
effect of non-stationarity of the shock wave position, which is not caught by
Numerical simulation of supersonic flow around a double ramp 687
the method. Analysis of the wall skin friction distributions (not shown here)
indicates that the optimal distance between the kinks is 11∇ · 12.4 mm, or,
in terms of d/δ, is 2.13∇ · 2.4.
3.0
3
2 6 7
2.5 8
Pw/P1
4 5
2.0
1
1.5
1.0
-5 0 5 10
x/delta
Also calculations of this case for few distances between the kinks have been
performed with BSL and SST turbulence models. The BSL and SST model
formulations allow to avoid the dependency of the results on a freestream value
of turbulence model parameters. The wall pressure distributions and skin
friction distributions indicate an excess turbulent viscosuty level that leads
to a shorter separation length prediction in comparison with both measured
data and k-ω model calculations. Due to the lack of space these data are not
presented.
3.3 Case 3
Finally similar investigations have been performed for the higher Mach num-
ber 3 (Case 3). In this case a good agreement between the experiments and
the numerical computations could also be found as presented in Fig. 5. The
optimal distance between the kinks is estimated to be in the range from 31 to
40 mm.
4 Conclusions and Future Work
The numerical simulation based on an URANS algorithm with several imple-

mented turbulence models has been carried out. An ability of the averaged
Navier-Stokes equations to predict the supersonic double-ramp flow behav-
ior is found out to be reasonable. A good agreement with the experimental
wall pressure distributions has been obtained. Three flow regimes have been
4
4
3.5 3
1 2
3
Pw/Pw1
2.5
1.5
0.5
-20 0 20 40 60 80
x, mm
discovered and the optimal ramp lengths have been estimated. An analysis
shows that with Re number decreasing the transition to the two-shock regime
takes place at a shorter distance between the kinks that agrees with classical
theory of viscous flows.
Future work will be directed toward a non-stationary effect investigation,
experimentally and numerically (with the aid of an URANS algorithm), since
the non-steadiness can essentially influence the flow parameters. [5].
References
1. Borisov, A.V., Fedorova, N.N.: Numerical simulation of turbulent flows near the
forward-facing steps. Thermophysics and Aeromechanics, 4, No. 1, 69–83 (1996)
2. Gaisbauer, U., Knauss, H., Wagner, S.: Experimental Investigation about Exter-
nal Compression of Highly Integrated Airbreathing Propulsion Systems. Basic
Research and Technologies for Two-Stage-to-Orbit Vehicles, WILEY-VCH, 347–
364 (2005)
3. Gaisbauer, U., Knauss, H., Wagner, S., Kharlamova, Yu.V., Fedorova, N.N.:
Experimental Investigation and Numerical Simulation of Supersonic Turbulent
Flows in Vicinity of a Double-Ramp Configuration. ICMAR: Proc. Pt 4 / Ed.
A.M. Kharitonov. -Novosibirsk: Publishing House Nonparel (2004)
4. Menter, F.R.: Two-equation eddy-viscosity turbulence models for engineering
applications. AIAA Journal, 32, No. 8, 1598–1605 (1994)
5. Wilcox, D.C.: Turbulent modeling for CFD. La Canada, California: DCW In-
dustries Inc. (1993)
Effects of Mach number on the combustion
zone length for a RAMAC configuration at
sub-detonative mode
Tarek Bengherbia,1,2 Yufeng Yao1 and Pascal Bauer2

1
Faculty of Engineering, Kingston University, Roehampton Vale, Friars Avenue,
London SW15 3DW, UK
2
Laboratoire de Combustion et de Détonique (CNRS), ENSMA, BP 109, 86960
Futuroscope CDX, France
Summary. Computational investigation of Mach number effects on the combustion

zone length in supersonic flow over a ram accelerator has been carried out to provide
valuable data for an on-going theoretical study of the same problem. It is found that
the combustion zone length is inversely proportional to the Mach number with a
significant reduction of about 33% at Mach 3.5 and about 44% at Mach 4, respec-
tively, while comparing to the combustion zone length at Mach 2.5. This correlation
provides some useful guidelines to define the control volume box in corresponding
theoretical study.
1 Introduction
The ram accelerator (RAMAC) is a ramjet-in-tube hypervelocity launcher

that uses chemical reactive combustion energy to drive a projectile to ultra-
high velocities. The geometry of projectile is similar to the center body su-
personic ramjet. It travels inside a tube filled with high pressure combustible
gas mixture, which burns on just behind the projectile base to provide the
ultra-thrust. One of primary interests recently is to quantify the correlation of
thrust-Mach number for the thermally choked propulsive mode. The finding
could be used to assist some theoretical approaches applying the quasi-steady
conservation equations to a control volume around the projectile (see Fig-
ure 1).
Previous investigations on propulsion modes have been carried out exper-
imentally and theoretically by Hertzberg [1]. It was found that one important
propulsive cycle is the thermally-choked sub-detonative mode, with typically
in-tube Mach number ranging from 2.5 to 4 and velocities just below the
Chapman-Jouguet (CJ) detonation speed of given propellant mixture. This
propulsive mode is initiated by a starting process that establishes a normal

690 T. Bengherbia et al.
shock on the rear of the projectile, which is then stabilized by thermally-

choking of the reactive flow at the full tube area behind the projectile [2].
As the projectile accelerates, the normal shock recedes and eventually falls off
the projectile. The flow field described here presents three key Mach numbers,
a minimum Mach number corresponding to that at which the incoming flow
will just reach sonic velocity at the throat, an immediate Mach number which
is just at the point where the released combustion heat drives the normal
shock up to the throat, and a maximum Mach number which corresponds
to the point where the normal shock falls off the projectile, or recedes com-
pletely behind the projectile. While well-established theoretical gas dynamic
relations can be used to estimate the position of the normal shock and to
determine the theoretical bounds of thermally chocked sub-detonative oper-
ation mode [3] [4], the detailed flow field has not been fully understood yet
and more work needs to be done.
Continuing our previous exercises on numerical simulation of the Ram
Accelerator [5] [9] [10], this paper aims to further extend the study by in-
vestigating the Mach number influence on the combustion zone length. The
objective is to quantify this length for typical Mach numbers and finally to
use the predictions for follow-up theoretical study using the control volume
method.
Fig. 1. Sketch of a ram accelerator in thermally choked sub-detonative propulsion

mode.
2 Physical Problem and Numerical Setup
The physical problem and computational procedure are going to be described

here in brief, focusing on some key issues as flow feature, geometry, meshing,
etc. More detailed numerical study can be found in [5]. Unless otherwise
stated, the ideal gas law will be applied and the chemical reaction is consid-
ered.
Mach number effects on the combustion zone length 691
2.1 Configuration
Figure1 illustrates a sketch of ’clean’ projectile configuration that has been

widely used in the experimental test [7]. The projectile has a bi-conical shape
with the nose cone part of an angle of 25 degree and a length of 52 mm,
and the main body with multiple integrated guide fins. The overall length of
projectile is 128 mm. The cylindrical test tube has a diameter of 38 mm.
2.2 Operating conditions
The supersonic incoming flow at the entrance of the domain is assumed to be

uniform with supersonic flow at Mach numbers ranging between 2.5 and 4.
Other flow conditions are static pressure P0 of 4.5 MPa and static temperature
T0 of 300 K. The unit Reynolds number based on the flow conditions and
characteristic length is 3.2×106 approximately. The simulation applies a single
step chemical reaction model, whereas considering the flow field with the
fuel/oxidized gas mixture at a composition of 2.8CH4 + 2O2 + 5.7N2 that was
used in the experimental test [4].
2.3 Numerical setup
We choose a computational domain of 400 mm in length along the axial di-

rection. Due to the axi-symmetrical feature of the flow field, only a quarter
of the full circumferential domain is considered, i.e. 90 degree. A multi-block
structured/unstructured mesh system is generated with in the domain.
Due to the characteristics of supersonic inflow, the flow conditions are fixed
at inlet boundary. At exit boundary, supersonic conditions are applied. The
non-slip and isothermal wall conditions are applied on both the tube and the
projectile walls; expect that part of the tube wall near the inlet and before
the oblique shock impinging point is modeled as free-slip condition in order
to reflecting the real physical flow situation during the experiments. Two side
surfaces of the computational domain are modeled as symmetric boundaries.
Careful grid convergence study has been carried out to identify a proper
baseline mesh for productive calculation. The details were presented in [6].
The complete governing equations are solved by finite volume method with
the SST turbulence model used for closure. The total energy transport model
is also applied to determine the viscous heating effects in the boundary layer,
where the kinetic energy effects become significantly important for the high
Mach number flow. The results presented here are based on a mesh with 70
points in the radial direction and a total of about 1.25 million grid points.
3 Results and Discussions

3.1 Determination of choke point
In order to establish combustion zone length and criteria for identifying the
sonic 3.1
pointDetermination
from the pressureof choke point
data, the thermal choking point is assumed to be
located in where the large amplitude oscillations in the pressure field behind
In order to establish combustion zone length and criteria for identifying the
the projectile changing the characteristics rapidly.
sonic point from the pressure data, the thermal choking point is assumed to be
Alocated
validation study
in where has been
the large performed
amplitude at in
oscillations Mach 3.5 andfield
the pressure results
behindare pre-
sentedtheinprojectile
Fig. 2. Itchanging
can bethe characteristics
seen that results rapidly.
from present numerical simulation
A validation
are compared well study
with has
the been performed
typical at Mach 3.5
experimental and results
record of theare pre-
pressure dis-
sented in Fig. 2. It can be seen that results from present numerical
tributions reported in Bauer et al. [3]. The sudden increase of pressure (i.e. simulation
are compared well with the typical experimental record of the pressure dis-
location of normal shock) agrees extremely well with the theory (also in Bauer
tributions reported in Bauer et al. [3]. The sudden increase of pressure (i.e.
et al.location
[3]) and the magnitude
of normal shock) agreesmatches
extremely the
well measurement data,
with the theory (also whereas the
in Bauer
theoryet over-predicts
al. [3]) and thethe pressure
magnitude jumpthe
matches most probablydata,
measurement due whereas
to the neglecting
the
of thetheory
viscousover-predicts the pressure
effects. The jump most between
good agreement probably these
due to three
the neglecting
data sets pro-
videsofthe
thegood
viscous effects. The
validation forgood
theagreement
proposedbetween these
criterion ofthree data setsthe
estimating pro-thermal
vides the good validation for the proposed criterion of estimating the thermal
choking location as highlighted in Fig. 2.
choking location as highlighted in Fig. 2.
25
Analytical estimation
Experimental data
20 Numerical prediction
Normalised pressure (P/P0)
15
10
Choke point
0
Projectile
-5
0 0.05 0.1 0.15 0.2
Axial coordinate, m
Fig.
Fig. 2. 2. Normalized
Normalized pressureratio
pressure ratio (P/P
(P/P 0 ) distributions along the projectile surface
0 ) distributions along the projectile surface
and at the centerline behind the projectile up to thermal choking point.
and at the centerline behind the projectile up to thermal choking point.
3.2 Mach number effects on combustion zone length

3.2 Mach number effects on combustion zone length
Further simulations have been performed at incoming Mach numbers of 2.5
and 4, respectively. Following the similar procedure as described previously,
Further simulations
Figure 3 plots the have
Mach been
numberperformed at incoming
variations starting from theMach numbers
location of the of 2.5
and 4,first normal shock,Following
respectively. which is justthe
at the point where
similar the combustion
procedure heat release
as described previously,
Figure drives the normal
3 plots the Machshocknumber
up to thevariations
throat, andstarting
ending atfrom
the location of the of the
the location
first normal shock, which is just at the point where the combustion heat release
drives the normal shock up to the throat, and ending at the location of the
second normal shock, which corresponds to the point where the normal shock
falls off the projectile body or recedes completely behind the projectile.
Mach number effects on the combustion zone length 693
4
M2.5
3.5 M3.5
M4
2.5
Mach Number
2
L1
L2
1.5 L3
0.5
0
Rear of the projectile
-0.5
0.1 0.12 0.14 0.16 0.18 0.2
Axial coordinate [m]
Fig. 3. Numerical prediction of combustion zone length at three typical Mach num-
bers of 2.5, 3.5 and 4.
The predicted combustion zone lengthes are denoted as L1, L2, and L3, in
correspondence to the incoming Mach numbers of 2.5, 3.5 and 4. The results
(Fig. 3) show that the combustion zone length is inversely proportional to
the Mach number. Taking the combustion zone length at Mach 2 (L1) as a
reference, the combustion zone lengthes at Mach 3.5 (L2) and Mach 4 has
been decreased to about (L1 − L2)/L1 = 33% and (L1 − L3)/L1 = 44%,
respectively. As an example, the combustion zone length (L1) at Mach 2.5 is
about 20% of the control volume length used in the theoretical study. Thus the
predictions from 3D simulation provides at least some qualitative comparisons
with those obtained by Bauer [9].
3.3 Surface pressure surface contours
The pressure contours around the projectile wall and fin surfaces are shown
in Fig. 4a. Because of the high pressure after normal shock, the flow exhibits
quite uniformity between the adjacent fins. Fig. 4b illustrates the pressure
distributions on the tube wall. It shows the zones of interaction of two shock
waves, one from the conical nose and another from the interaction of the
shocks over the fins located in each side.
(a) b)
Fig. 4. Contours of pressure distributions at (a) the projectile wall and fin surfaces
and (b) on the tube wall.
4 Conclusions
Numerical simulations have been carried out by solving the three-dimensional
Navier-Stokes equations, in conjunction with the chemical reactive model for
supersonic flows over a Ram Accelerator. Flows at three Mach numbers 2.5, 3.5
and 4 are considered with the sub-detonative propulsion mode applied. The
simulation reveals that the combustion zone length is inversely proportional
to the Mach numbers of 2.5, 3.5 and 4 with a significant reduction about
33% and 44% at two high Mach numbers (3.5 and 4) while referencing to the
prediction length at low Mach number 2.5, qualitatively in agreement with
other investigations.
References
1. Hertzberg A., Bruckner A.P. and Bogdanoff D.W., AIAA Journal, 26(2),
195-203 (1988).
2. Bruckner A.P. et al., J. of Propulsion and Power, 7(5), 828-836 (1991).
3. Bauer P. et al., J. Phys. IV 10, Pr11:59-67 (2000).
4. Bauer P., Knowlen C., Eur. Phys. J. Appl. Phys., 21, 233-238 (2003).
5. Bengherbia T., Yao Y.F. and Bauer P., submitted to Eur. Phys. J. Appl.
Phys. (2006).
6. Bengherbia T., Yao Y.F. and Bauer P., AIAA 2006-0558 (2006).
7. Leblanc J.E., et al., J. Phys. IV 10, Pr11:119-130 (2000).
8. Li C.P., Kailasanath K. and Oran E.S., Combustion and Flame, 108(1-2),
173-186 (1997).
9. Bauer P., Knowlen C. and Bruckner A.P., J. of Propulsion and Power,
21(5), 955-958 (2005).
10. Bauer P., Knowlen C. and Bruckner A.P., Eur. Phys. J. Appl. Phys., 29,
253-258 (2005).
A Finite Element/Finite Volume Mixed Solver
and Applications on Heat Flux Prediction∗
He Lixing, Zhang Laiping, and Zhang Hanxin
China Aerodynamics Research and Development Center, Mianyang, Sichuan,

China, 621000
Summary. A finite element/finite volume mixed solver is presented. A discontin-
uous Galerkin (DG) finite element (FE) solver is used near the boundary layers to
capture the viscous effects, whereas a finite volume (FV) solver is adopted in the
outer field to save CPU time and store memory. The mixed solver is validated with
several hypersonic cases. The computational results show good agreements with ex-
perimental data, and the comparison on CPU time and memory demonstrates the
higher efficiency than that of the pure FE solver.
Key words: Discontinuous Galerkin Finite Element Method; Finite Volume Method;
FE/FV Mixed Method
1 Introduction
There is an ever-increasing demand of simulating sophisticated flows over

more and more complex geometries in computational fluid dynamics (CFD).
The traditional structured grids faced so many challenges because of their
inherent limitation. Hence, unstructured grids, and structured/unstructured
hybrid grids recently, were paid more and more attention due to their extreme
flexibility for complex geometries. Correspondingly, the numerical methods
(including FVMs and FEMs) were developed rapidly.
For supersonic flows with strong non-linear characteristics, such as shock
waves, the traditional linear FEMs are not suitable. Therefore, many new-form
FEMs were proposed, for instance, the streamline upwind/Petrov–Galerkin
method[1], Taylor–Galerkin method[2], NND FEMs[3] and FD/FE hybrid
method[4]. However, these methods cannot be used for the hybrid grids with
hanging nodes (popular for Cartesian grids and self-adaptive grids) due to the
limitation of continuity on element boundaries. In the end of 1980s, Cockburn
and Shu[5, 6, 7, 8] proposed serial Runge–Kutta Discontinuous Galerkin meth-
ods (RKDG), which showed good performance on solving non-linear equa-
∗
This work is supported by Chinese National Foundation of Nature Science (No.
10321002 and No. 10472012).

696 He Lixing, Zhang Laiping, Zhang Hanxin
tions. Following the DG idea, we proposed a DG–FE solver on 3D arbitrary

elements, and applied it to simulating hypersonic viscous flows[9].
As mentioned in some references, FEMs have some superiority over FVMs
on solving derivative variables, because the derivatives of physical variables
can be directly solved with FE equations. However, they usually require more
CPU time and memory than FVMs. On the other hand, FVMs have not
completely overcome the difficulty of predicting the heat flux over complex
geometries, especially on unstructured grids[10]. An optimal choice is to merge
them. So in this paper, a finite element/finite volume mixed solver is presented,
and applied to heat flux prediction of hypersonic viscous flows. The DG–FE
solver is used near the boundary layers to capture the viscous effects, whereas
a FV solver is adopted in the outer field to save CPU time and memory. The
mixed solver is validated by two hypersonic cases, i.e. hypersonic flows over
blunt-cone and double-ellipsoids. The computational results, including flow
patterns and heat flux distributions, show good agreements with experimental
data, and the comparison on CPU time and memory demonstrates its higher
efficiency.
2 Finite Element/Finite Volume Mixed Solver

As mentioned above, both FEMs and FVMs
have their own advantages. Merging them
with each other will be a better choice. There
were some merging strategies in literatures.
For example, FEMs are used to discretize
the viscous terms in governing equations,
whereas FVMs are adopted to discretize the
inviscid terms. Here we adopt another merg-
ing strategy based on zone-decomposition:
the DG–FE solver runs only in the zone of Fig. 1. Merging strategy
boundary layer, whereas a FV solver runs in the outer field. Figure 1 shows
the sketch of our strategy. Therefore, the viscous effect can be simulated more
accurately. On the other hand, the CPU time and memory will be saved also.
In order to simulate boundary layers more accurately, staggered grids
should be used in these regions. However, the pure staggered tetrahedral grids
are not easy to generate for complex configurations[11]. Therefore, so many
hybrid grid techniques[11, 12, 13, 14] were proposed, and have been proven to
be more flexible and efficient than pure tetrahedral grid generator. Generally,
for three-dimensional cases, there will exist two types of elements in boundary
layers, i.e. hexahedron and prism.
For these two types of elements, the DG–FE solver[9] chooses different
basis functions, limiters and numerical integral formulae. Unlike tetrahedral
grids (the basis functions are the volume coordinates naturally and no coordi-
nate transformation is required), the arbitrary hexahedral and prismatic grids
A Finite Element/Finite Volume Mixed Solver 697
should be transformed into a normalized shape, as shown in Fig. 2 and Fig. 3.
Fig. 2. Mapping for hexahedron Fig. 3. Mapping for prism
In the local coordinate system, the governing equations can be written as:
∂U ∂ Ẽ ∂ F̃ ∂ G̃ ∂U 1˜
+( + + )/J = + ∇ · H̃ = 0 (1)
∂t ∂ξ ∂η ∂ς ∂t J
where U is the conservative vector, H̃ the flux (including the inviscid and
viscous parts) and J4 the Jacobion. In each element, supposing that:
X
uh (t, x, y, z) = uj (t)ϕj (x, y, z) (2)
j=1
in which the orthogonal basis for cubic cell in local system (Fig. 2) are
ϕ1 = 1, ϕ2 = ξ, ϕ3 = η, ϕ4 = ζ (3)
whereas the orthogonal basis for prismatic cell (Fig. 3) are
ϕ1 = N1 + N2 − N3 , ϕ2 = N2 + N3 − N1 , ϕ3 = N1 + N3 − N2 , ϕ4 = ζ (4)
where N1 , N2 and N3 are the area-coordinates in the bottom triangle.
Substituting relation (2) into equation (1), multiplying by the basis func-
tions, and then integrating in each cell, and setting the weighted-residual equal
to zero,Z we have: Z
d ˜ · H̃ϕj /Jdξdηdς = 0
uh ϕj dξdηdς + ∇ (5)
dt
K K
The integral in equation (5) can be solved with Gauss’s numerical integral
method. For the inviscid flux, Roe’s flux splitting scheme can be used here. To
diminish the unphysical oscillation near shock waves, limiters are introduced
for each type of cells (See previous work [9]). Meanwhile, a LU–SGS implicit
method is adopted to accelerate the convergence history.
On the grids away from boundary, a cell-centered finite volume solver[14]
is adopted here. For the integral form of governing equation, we can rewritten
it as following
∂U X
Vi + H̃ (U )f dSf = 0 (6)
∂t
f
where S is the surface of control volume V .

Similar to the DG–FE solver, Roe’s flux splitting scheme is used to cal-
culate the inviscid flux, and center scheme for the viscous flux. To reach sec-
ond order, the variables on cell interface are calculated with Taylor’s series.
Similarly, proper limiter is used here to guarantee the monotonicity of recon-

struction. Also, the same LU–SGS implicit method is used here for temporal
discretization. See [14] for details.
On the interface of sub-zones in which difference solvers run, the flux is
set as that solved by the DG–FE solver, and is added directly to the total flux
of cells sharing the interface in the FVM sub-zone. So the conservation of the
mixed solver is guaranteed.
3 Numerical Results
Two hypersonic cases are simulated to validate the mixed solver. The first case
is hypersonic flow over blunt-cone when angle of attack equals to 20◦ . The
computational conditions are same as those in Cleary’s experiment: M∞ =
10.6, Re = 1.1 × 105 , T∞ = 47.3K, Tw = 294.44K.
(a) Computational grid (b) Pressure contours (c) Flow patterns

Fig. 4. Computational results over blunt cone
Fig. 5. Heat flux on meridional lines Fig. 6. Heat flux on cross-sections
For this simple model, a structure grid (75×65×61, as shown in Fig. 4(a))is
used, but the data structure is transformed into unstructured format. For the
mixed solver, the DG–FE solver runs only in the inner 20 layers of grid close to
the body surface, whereas the FV solver runs on the other region. Figure 4(b)
shows the pressure contours. It can be seen that the mixed solver captures
the bow shock clearly, and there is no unphysical oscillation in the whole flow
field. In Fig. 4(c), we show the surface flow patterns obtained by the DG–FE
A Finite Element/Finite Volume Mixed Solver 699
solver and the mixed solver. They are almost same as each other, and both
of them capture the second separation clearly. The heat-flux distributions, as
well as experimental data, are shown in Fig. 5 and Fig. 6. It can be seen that
they agree with each other very well. Furthermore, we compare the CPU time
and memory for both solvers, which are listed in Table 1. It can be seen that
about 37% memory and more than 55% CPU time are saved.
Table 1. Comparison of CPU time and memory

Solver Memory CPU Time
Pure DG–FE Solver 1.61 2.22
Mixed Solver 1.00 1.00
(a) Computational grid (b) Density contours (c) Flow patterns

Fig. 7. Computational results over double-ellipsoids
(a) Symmetric plane (b) x = 78 (c) x = 120

Fig. 8. Heat flux distribution on symmetric plane and two cross-sections
The second validation case is hypersonic flow over double-ellipsoids. Hy-

brid grid(Fig. 7(a)) is used here, which includes prismatic grids near the wall,
Cartesian grids in the outer field and tetrahedral grids in the gap. Here the
DG–FE solver runs only on the prismatic grids. The computational condi-
tions are M∞ = 8.0, α = 0◦ , Re = 2.5 × 105 , T∞ = 65.217K, Tw = 300.0K.
As shown in Fig. 7(b), the density contours, the mixed solver captures the
flow structure clearly. Once again, the flow pattern on the wall is plotted in
Fig. 7(c), in which the main and second separation structure are captured
clearly. In Fig. 8, we give the heat-flux distributions in the symmetric merid-
ional lines and two cross-sections. The numerical results are very similar to
those of experimental data.
Concluding Remarks
A DG–FE/FV mixed solver is presented in this work, and applied successfully
to heat flux prediction of hypersonic viscous flows. The mixed solver is set up
based on unstructured/hybrid grids, so it is suitable for simulations of flows
over complex configurations. Numerical results demonstrate the high efficiency
and accuracy of present mixed solver.
References
1. Hughes T J R.: Recent progress in the development and understanding of SUPG
methods with special reference to the compressible Euler and Navier–Stokes
equations. Int. J. Num. Meth. Fluids, 7:1261–1275(1987)
2. Donea J.: A Taylor–Galerkin method for convective transport problems. Int. J.
Num. Meth. Eng., 20:101–120(1984)
3. Wu W.Y, Cai Q.D: A new NND finite element scheme on unstructured grid,
Acta Aerodynamica Sinica, 16(1):1–13(1998)
4. Duan Z.Y., Tong B.G., Jiang G.Q.: A finite difference/finite element mixed
method and applications on aero-thermal prediction, ACTA Aerodynamica
Sinica, 15(4):1–8(1997)
5. Cockburn B., Hou S., and Shu C.W.: TVD Runge–Kutta local projection dis-
continuous Galerkin finite element method for conservation laws IV: The mul-
tidimensional case. Math. Comp., 54:545–581(1990)
6. Cockburn B., Lin S.Y., and.Shu C.W.: TVD Runge–Kutta local projection
discontinuous Galerkin finite element method for conservation laws III: One
dimensional systems. J. Comput. Phys., 84:90–113(1989)
7. Cockburn B. and Shu C.W: TVD Runge–Kutta local projection discontinuous
Galerkin finite element method for scalar conservation laws II: General frame-
work. Math. Comp., 52:411–435(1989)
8. Cockburn B. and Shu C.W.: The P1 -RKDG method for two-dimensional Euler
equations of gas dynamics. ICASE Report No.91–32(1991)
9. He L.X., Zhang L.P., Zhang H.X.: Discontinuous Galerkin finite element method
on 3D arbitrary elements, Acta Aerodynamica Sinica, accepted(2005)
10. Peter A. Gnoffo and Jeffery A. White.: Computational aerothermodynamic
simulation issues on unstructured grids. AIAA 2004–2371(2004)
11. Weatherill NP.: Unstructured grids: procedures and applications. Handbook of
Grid Generation. CRC Press(1998)[Chapter 26]
12. Pirzadeh S.: Three-dimensional unstructured viscous grids by the advancing-
layers method. AIAA J;34(1):43–9(1996)
13. Kallinderis Y, Khawaja A, McMorris H.: Hybrid prismatic/tetrahedral grid
generation for complex geometries. AIAA J;34:291–8(1996)
14. Zhang LP, Yang YJ, Zhang HX.: Numerical simulations of 3D inviscid/viscous
flow fields on Cartesian/unstructured/prismatic hybrid grids. In: Proceedings
of the 4th Asian CFD Conference, Mianyang, Sichuan, China(2000)
Numerical Analysis for Magnetic Control of
Heat-Transfer and Pressure in Hypersonic
Shock Wave Interference Flows
Daisuke Tsubakino1 , Yoshiteru Tanaka2 , and Kozo Fujii3

1
University of Tokyo, 3-1-1, Yoshinodai, Sagamihara, Kanagawa.
tsubakino@flab.eng.isas.jaxa.jp
2
University of Tokyo, currently Toyota.
tanaka@flab.eng.isas.jaxa.jp
3
ISAS/JAXA, 3-1-1, Yoshinodai, Sagamihara, Kanagawa.
fujii@flab.eng.isas.jaxa.jp
1 Introduction
Local heat-transfer and pressure increases due to shock wave interference are
critical problems in structural designs for complicated hypersonic vehicles.
Therefore, many experiments and computational simulations have been done
to identify the critical shock interference phenomena [1, 2, 3]. Types of the
shock interference were classified by Edney in Ref. [1]. His classification shows
that some types are severe interference and the others types are not. The most
critical interference is called Type IV.
Conventionally, the thermal protection system and the aerodynamic shape
against the shock interference of the hypersonic vehicles have been designed
based on such experimental and computational data. However, such conven-
tional design approach may increase the weight and lose the aerodynamic char-
acteristics of the vehicles. The actual shock interference phenomena around
the hypersonic vehicle change variously according to its flight conditions and
is not always severe interference. Considering the actual situation, it is bet-
ter to use a dynamic flow control method which can handle various critical
situations during the actual flight.
Recently, some dynamic control methods have been developed. A magnetic
flow control is proposed as one of such control method. This method has the
characteristic that a magnetic field is applied to ionized fluid in hypersonic
flow. Its application to a blunt body in a free stream was computationally
investigated [4, 5]. According to their research, a magnetic field shifts the
bow shock to windward and friction drag and heat-transfer-rate on the body
surface are reduced. Considering that the hypersonic shock interference may
cause ionization, the magnetic flow control can be applied to the hypersonic

702 Daisuke Tsubakino, Yoshiteru Tanaka, and Kozo Fujii
shock interference problem. However, the investigation of the magnetic control

of hypersonic shock interference have hardly been done.
The purpose of this study is to obtain a useful knowledge about the mag-
netic dynamic flow control of hypersonic shock interferences by computational
simulations. As the objective flow field, the Type IV interference flow around
a blunt body is selected. Actually, there exist many parameters of a magnetic
field to be investigated such as a type, position and magnitude of the magnetic
field in the present study. Therefore, the use of computational simulations is
desirable for such parametric study because many cases with different mag-
netic field conditions can be performed easily and detailed information of the
flow fields can be obtained by the computational simulations. In this paper,
the magnetic flow control with a magnetic dipole is considered, and the ef-
fect on the shock interference phenomena by the magnetic field magnitude is
investigated.
2 Computational Methods
Interaction between the ionized fluid flow and the magnetic field induces a
total current. An electromagnetic field exerts the Lorentz force f and does
work P on this current. These values per unit volume are given by
f = ρe E + J × B , P = E · J. (1)
where ρe is the total charge density, J is the total current, E is the electric
field and B is the magnetic field. Neglecting magnetization and polarization
of the fluid, the total current is the sum of the convected charge ρe u and
conduction current j, i.e. J = ρe u + j, where u is the velocity of the fluid.
In general, electromagnetic phenomena are governed by Maxwell’s equa-
tions. However, in many aerodynamic problems, flow time scales are larger
than the reciprocal of the plasma frequency, flow velocities are much less than
the speed of light, and flow length scales are much larger than the Debye
length. Thus simplified Maxwell’s equations can be applied. Under these as-
sumptions, the convected charge ρe u is negligible as compared to the conduc-
tion current j and the first term of the Lorenz force is negligible as compared
to the second term. The conductive current density is assumed to be given by
Ohm’s law
j = σ (E + u × B) , (2)
where σ is the electrical conductivity.
Furthermore, low magnetic Reynolds number approximation is applied.
This approximation is based on the fact that the distortion of the magnetic
field by the flow can be neglected for many aerospace applications. Thus,
the imposed magnetic field is only important. Therefore, the governing two-
dimensional equations in generalized coordinate form are given as
Magnetic Control of Hypersonic Shock Wave Interference Flows 703
∂X ∂Fk ∂Fνk
+ = +S , k = 1, 2, (3)
∂t ∂ξk ∂ξk
where X is the conserved variable vector, Fi are inviscid flux vector and Fνi is
viscous flux vector and all quantities in Eq. (3) are nondimensionalized. The
source term S contains the nondimensional f and P written as
f =Q σ j×B , P =Q σ j·E , (4)
where the quantities in Eqs. (4) are nondimensionalized. The parameter Q in
Eqs. (4) is called interaction parameter and given as
σr B0 U∞ Lr
Q= 2
, (5)
ρ∞ U∞
where the U∞ and ρ∞ represent the free stream velocity and density and Lr
is the blunt diameter. σr and B0 are the reference electric conductivity and
the reference magnetic field magnitude. This parameter Q represents the ratio
between the inertia terms and the electromagnetic terms. In this study, this
parameter denotes a degree of a magnitude of the magnetic field. Then S can
be expressed as S = (0, f1 , f2 , f3 , P )T , where fi is the “ i th ” component of
the nondimensional Lorentz force f .
The computational schemes are the AUSMDV scheme [6] for the convective
terms and the LU-SGS scheme [7] for the time integration. The third order
upwind biased MUSCL TVD method [8] is also employed for the higher-order
approximation of the numerical flux. The magnetic field for the flow control
is two-dimensional magnetic dipole whose origin is set at the center of a blunt
body and whose dipole moment vector is directed to the opposite of free stream
direction. B0 in Eq. (5) is set the magnitude of the magnetic dipole at the top
of th blunt. The electric field is not used. In this paper, the calorically perfect
gas is assumed for qualitative investigation. Therefore, the fluid behind bow
shock wave is given a constant electric conductivity σr .
3 Flow conditions
Flow conditions for this analysis is based on the previous shock interference
study [3]. The Mach number of the free stream is M∞ = 8.03 and the Reynolds
number referring the diameter of the blunt body is Re = 5.15×105 . The angle
of the impinging shock is 18.11 [deg] where the extended line of the impinging
shock impinges at the front edge of the body. This condition cause the Type
IV interference. The Mach number behind impinging shock wave is 5.25. This
impingement is generated at inlet boundary by the oblique shock relation.
The isothermal wall boundary condition is employed on the body surface.
The wall temperature Tw is 294.44 [K] and the free stream temperature T∞
is 111.56 [K]. The simulations are done for Q = 0, 3, 6 and 9.
4 Results
Figure 1 shows a schematic of Type IV interference. The impinging shock
impinges on the bow shock (B. S.) and generates the triple points (T. P.).
From this tripe point, the share layer (S. L.) and the lambda shock occur.
The lambda shock generates another triple point where another share layer
occurs. Then, the supersonic jet occurs between the two shear layers and
impinges on the body surface. The impingement of jet leads extremely high
heat-transfer-rate and pressure in the localized region.
B. S.
T. P. Body surface
Supersonic jet
M∞ Jet bow shock
Impinging shock Expansion wave

S. L. S. L.
Lambda shock
T. P.
B. S.
Fig. 1. Schematic of the Type IV interference.
The time averaged temperature contours for each Q are shown in Fig.
2. Light color expresses higher value. As shown in the figure, the higher the
magnetic field becomes, the larger the length of shock standoff becomes. Due
to this effect, the triple point made by the interference between the impinging
shock and the bow shock moves downward along the impinging shock. This
means the shear layer generated by this triple point impinges lower part of
the body. As the result, the interference type turns to another type from the
Type IV over Q = 6. This type is called the Type III interference and this
interference type is less severe than Type IV.
The heat-transfer-rate and the pressure distributions on body surface are
shown in Figs. 3,4. The horizontal axis θ means angles of the body surface
measured clockwise from the horizontal direction, and p and q are pressure and
heat-transfer-rate. The subscripts “0” and “w” means the values at a stagna-
tion point without a shock impingement and on the body surface, respectively.
As expected from flow fields, the heat-transfer-rate and the pressure are dras-
tically reduced by the magnetic field. From these results, it is sure that this
control method is effective.
Conclusion
The magnetic control on the shock interference flows was computationally in-
vestigated. The magnetic field can be used to change the interference type to
Magnetic Control of Hypersonic Shock Wave Interference Flows 705
(a). Q = 0 (b). Q = 3
(c). Q = 6 (d). Q = 9
Fig. 2. Temperate contours for each Q.
12
Q0
10 Q3
8 Q6
Q9
6
q
0
-80 -60 -40 -20 0 20 40 60 80
theta
Fig. 3. Heat-transfer-rate ratio distribution on the body surface.

12
Q0
10 Q3
8 Q6
Q9
p 6
0
-80 -60 -40 -20 0 20 40 60 80
theta
Fig. 4. Pressure ratio distribution on the body surface.
reduce the local heat-transfer-rate and pressure on the body surface. Further-
more, it is expected that the changes of direction or position of a dipole may
produce more effective control.
References
1. Edney, B. E., “Anomalous Heat Transfer and Pressure Distributions on Blunt
Bodies at Hypersonic Speeds in the Presence of an Impinging Shock,” FFA Report
115, Aeronautical Research Institute of Sweden, Stockholm, Sweden, 1968.
2. Wieting, A. R. and Holden, M. S., “Experimental Shock-Wave Interference Heat-
ing on a Cylinder at Mach 6 and 8,” AIAA Journal , Vol. 27, No. 11, 1989,
pp. 1557–1565.
3. Yamamoto, S. and Kano, S., “Structure of Bow Shock and Compression Wave In-
teractions in Unsteady Hypersonic Shock/Shock Interference Flow,” AIAA Paper
96–2152, 1996.
4. Poggie, J. and Gaitonde, D. V., “Magnetic control of flow past a blunt body:
Numerical validation and exploration,” Physics of Fluid , Vol. 14, No. 5, 2002,
pp. 1720–1731.
5. Fujino, T., Funaki, I., Sugita, H., Mizuno, M., and Ishikawa, M., “Numerical
Analyses on Flow Control around Blunt Body “OREX” by Magnetic Field,”
AIAA Paper 2003–3760, 2003.
6. Wada, Y. and Liou, M. S., “A Flux Splitting Scheme with High-Resolution and
Robustness for Discontinuities,” AIAA Paper 94–0083, 1994.
7. Yoon, S. and Jameson, A., “Lower-Upper Symmetric-Gauss-Seidel Method for
the Euler and Navier-Stokes Equations,” AIAA Journal , Vol. 26, No. 9, 1988,
pp. 1025–1026.
8. Van Leer, B., “Toward the ultimate Conservative Difference Scheme. 4, A New
Approach to Numerical Convection,” Journal of Computational Physics, Vol. 23,
March 1977.
Numerical Simulation of a Flat-Plate
Hypersonic Shock Layer Perturbed
by External Acoustic Waves
A.N.Kudryavtsev, S.G.Mironov, T.V.Poplavskaya, and I.S.Tsyryulnikov
ITAM SB RAS, Novosibirsk, 4/1 Institutskaya St. E-mail:popla@itam.nsc.ru
1 Introduction
In high-velocity high-altitude flight, the thick boundary layer on the surface
of a flying vehicle merges with the bow shock wave so that, even at a large
distance from a leading edge, they may form the so-called viscous shock layer
(VSL). Like the boundary layer, the laminar VSL is unstable, and flow dis-
turbances arising and developing in this layer induce its transition to the
turbulent flow regime. However, the mechanisms governing their emergence
and development in the hypersonic VSL may differ substantially from those
investigated in supersonic flows with lower Mach numbers [1, 2]. A better
understanding of the receptivity and instability mechanisms is a necessary
condition for the development of efficient methods for controlling the laminar-
turbulent transition in a hypersonic flow around flying vehicles.
The possibilities of experimental modeling of receptivity and evolition of
disturbances in hypersonic wind tunnels are rather limited. The present paper
describes the results of direct numerical simulation of receptivity and evolution
of perturbations in a VSL on a flat plate for a very high Mach number (M∞ =
21) and a moderate Reynolds number (ReL = 1.44 × 105 ). The problem of
interaction of the VSL with free-stream acoustic disturbances of slow and fast
modes and with perturbations introduced in the vicinity of the leading edge
of the plate by means of periodic blowing and suction was studied by solving
the 2D Navier-Stokes equations. The computed results are compared with
characteristics of density fluctuations measured in experiments performed at
the same flow parameters in the T-327A hypersonic nitrogen wind tunnel
based at ITAM.
2 Numerical Formulation of the Problem

The 2D Navier-Stokes equations written as a system of conservation laws along
with the equation of state for a perfect gas are solved. The convective terms

708 A. N. Kudryavtsev et al.
are approximated by the MP5 (monotonicity-preserving, 5th order) scheme

developed in [3]. This scheme has a built-in analyzer, which can distinguish be-
tween solution discontinuities and smooth extrema points. The diffusive terms
are approximated by the 4th order central finite differences. Time stepping
is performed by the 3rd order Runge-Kutta method. The numerical method
used is described in more details in [4].
A part of the lower side of the rectangular computational domain coincides
with the plate surface. The left (inflow) boundary is located at a distance of
a few computational cells upstream from its leading edge. The height of the
computational domain is chosen such that the bow SW emanating from the
leading edge does not interact with the top boundary. The right (outflow)
boundary is moved downstream from the plate trailing edge so that the flow
in the exit cross section is fully supersonic.
The steady flow is computed first with a uniform hypersonic flow set on
the inflow and top boundaries. On the outflow boundary, the solution is ex-
trapolated from inside the computational domain. As the rarefaction effects
in the problem considered are fairly significant (the streamwise velocity for
x = 0.1 is approximately 17% of the free-stream value on the plate surface
and about 7% near the trailing edge), the boundary conditions on the plate
take into account the velocity slip and temperature jump (see [4]).
A uniform computational grid consisted of 1050 cells in the streamwise
direction and 240 cells in the crossflow direction.
In [4], the computed mean flow was compared with experimental data.
Excellent agreement of the total pressure, density, and Mach number distri-
butions was demonstrated.
In solving the problem of interaction of the viscous shock layer with natural
disturbances, the latter are introduced by setting appropriate time-dependent
boundary conditions. The natural disturbances observed in experiments con-
sist of acoustic waves propagating in the external flow. In numerical simula-
tions, the variables on the inflow boundary of the computational domain are
prescribed in the form of a superposition of the steady basic flow and a planar
monochromatic acoustic wave, which had the following form:
 0  
u ± cos θ
 v0 
 0  = A  ∓ sin θ  exp [i(kx x + ky y − ωt)] .
 
p   1  (1)
ρ0 1
Here θ is the angle of incidence of the wave, A is its amplitude, t =
t∗ c∗∞ /L∗ is the dimensionless time, c∗∞ is the free-stream sound velocity,
kx = k cos θ, ky = −k sin θ are the components of the wave vector related
to the dimensionless frequency ω = 2πf ∗ L∗ /c∗∞ by the dispersion relation
k = ω/(M∞ cos θ ± 1). The upper (lower) sign in (1) refers to the fast (slow)
acoustic wave, respectively.
The boundary conditions on the plate surface are the same as those used to
find the steady solution, except for zero temperature perturbations (because
Hypersonic Shock Layer Receptivity 709
of considerable thermal inertia of the plate). Thus, the wall temperature is

always equal to the temperature obtained in solving the steady problem. After
introduction of perturbations, the Navier-Stokes equations are integrated until
the unsteady solution reached a quasi-periodic regime.
If blowing-suction type disturbances are introduced into the VSL, they are
simulated by imposing the boundary condition for the transverse mass flow
rate on some portion of the plate surface:

∗ ∗ ∗ ∗ x − x1
ρ v |y=0 /ρ∞ c∞ = A sin π sin(ωt) (2)
x2 − x1
Here x1 = 0, 065 and x2 = 0, 08 are the boundaries of the region where
the perturbation is introduced.
3 Results
All computations have been performed at the free-stream Mach number M∞ =

21, the Reynolds number based on the plate length ReL = 1.44 × 105 , the
stagnation temperature of the incoming flow T0 = 1200 K and the flate-plate
temperature Tw = 300 K. Up to 20 processors of a multiprocessor computer
cluster have been used.
Figure1 shows the mean density field (Fig.1a) and the instantaneous fields
of density fluctuations in the case of excitation of the VSL by external acoustic
disturbances (Fig.1b). The frequency of acoustic disturbances is f ∗ = 19.2
kHz. The amplitude of the introduced perturbations is A = 0.0286. The solid
and dashed isolines in Fig.1b show the positive and negative fluctuations of
density, respectively. It is seen from the figure that there are two regions with
the maximum fluctuations of density: on the SW and on the upper edge of the
boundary layer, where the mean density changes rapidly because of an increase
in static temperature. The fluctuations on the SW and on the boundary-layer
edge are opposite in phase. Following to the linear theory of SW-disturbance
interaction 5], it was demonstrated in [4] that, under the given flow conditions,
the external acoustic disturbances passing through the SW can generate only
waves of the entropy-vortex mode.
It seems of interest to compare the computed growth rates of disturbances
on the SW with the data of the linear stability theory. The computation by
the locally parallel linear stability theory was performed taking into account
the influence of a closely located SW [6]. Figure 2 shows the growth rates of
disturbances. Two solutions (dashed lines) obtained within the framework of
the locally parallel linear stability theory are seen to be in close agreement with
the results of direct numerical simulations of development of perturbations
generated by the slow and fast modes of external acoustic disturbances (solid
lines) as well as with the experimental data for slow acoustic disturbances
(triangles).
0.2
0.15
0.1
0.05
(b)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 1. Isolines of mean density (a) and density fluctuations (b) at A = 0.0286,
f ∗ = 19.2 kHz
Fig. 2. Growth rates of disturbances on the SW
The solid lines in Fig.3 show the variation along the plate of the com-
puted amplitudes of density fluctuations on the SW for slow external acoustic
disturbances of different frequencies f ∗ = 19.2 kHz, 38.4 kHz, 50 kHz, and
80 kHz. Symbols in Fig.3 show the experimentally measured amplitudes at
the same frequencies. The computational and experimental data are in good
agreement. It can also be seen from the figure that an increase in frequency
yields a nonmonotonic variation of the amplitude of density fluctuations on
the SW along the streamwise coordinate. This phenomenon can be explained
by the interaction of external flow perturbations and vortex disturbances in
the VSL, which propagate with different streamwise wavenumbers. This in-
teraction is manifested on the SW as amplitude beatings.
Variation of the angle of incidence of external acoustic waves leads to
changes in both their streamwise phase velocity and their streamwise wavenum-
ber in the external flow. This alters the interference of external acoustic dis-
Hypersonic Shock Layer Receptivity 711
Fig. 3. Amplitudes of density fluctuations on the SW for slow acoustic waves of

different frequencies
Fig. 4. Amplitudes of density fluctuations on the SW for different angles of incidence

θ of a fast acoustic wave A = 0, 0286, f ∗ = 50 kHz
turbances and internal vortex disturbances interacting in the vicinity of the

SW and the amplitude beatings on SW intensify (Fig.4).
If the VSL is perturbed by blowing and suction near the leading edge then
the flowfields (Fig.5a) in many aspects resemble those generated by external
acoustic disturbances (Fig.5b). This provides a possibility of active control of
the evolution of disturbances in the VSL. The introduction of periodic pertur-
bations of the blowing-suction type from the flat-plate surface in the opposite
phase to external acoustic disturbances leads to significant suppression of den-
sity fluctuations on the boundary-layer edge (Fig.5c).
Acknowledgements. This work was supported by Russian Foundation
for Basic Research (Grant No. 05-08-33436). Computer resources were pro-
vided by Siberian Supercomputer Center (Novosibirsk) and Joint Supercom-
puter Center (Moscow).
0.2
0.1
(a)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.2
0.1
(b)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.2
0.1
(c)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 5. Isolines of density fluctuations for excitation by blowing/suction with f ∗ =

38.4 kHz and A = 0.06 (a) and by a slow acoustic wave with f ∗ = 38.4 kHz and
A = 0.001 (b) and by the same acoustic wave in superposition with opposite-phase
blowing/suction (c)
References
1. Ma Y., Zhong, X.: Receptivity of a supersonic boundary layer over a fflat plate.
Part 2. Receptivity to free-stream sound. J. Fluid Mech., 488, 79-121 (2003)
2. Egorov, I.V., Sudakov, V.G., Fedorov, A.V.: Numerical simulation of super-
sonic boundary-layer receptivity to acoustic disturbances. Izvestiya RAN, Mech.
Zhidk. i Gaza, 2, 43-52 (2006)
3. Suresh, A., Huynh , S.T.: Accurate monotonicity-preserving schemes with
Runge-Kutta stepping. J. Comput. Phys., 136, 83-99 (1997)
4. Kudryavtsev, A.N., Mironov, S.G., Poplavskaya, T.V., Tsyryulnikov, I.S.: Exper-
imental investigation and direct numerical simulation of evolution disturbances
in viscous shock layer on a flat plate. J. Appl. Mech. Tech. Phys., 5, 43-52(2006)
5. McKenzie, J.F., Westphal , K.O.: Interaction of linear waves with oblique shock
waves. Phys. Fluids, 11, 2350-2362(1968)
6. Maslov, A.A., Poplavskaya , T.V., Smorodsky, B.V.: Stability of a hypersonic
shock layer on a flat plate. Comptes Rendus Mecanique, 332, 875-880 (2004)
Numerical Simulation of Supersonic Turbulent
Flows over Backward–Facing Steps
N.N. Fedorova1 , and I.A. Bedarev2

1
Institute of Theoretical and Applied Mechanics, Novosibirsk, 630090, Russia
nfed@itam.nsc.ru
2
bedarev@itam.ncs.ru
Key words: Turbulence modeling, Shock waves, Base flow
Summary. The turbulent flows over 2–D backward–facing steps with different step
inclination angles at Mach numbers M = 2∇ · 5 are considered. The flows are nu-
merically simulated on the basis of the full Favre averaged Navier Stokes equations
complemented with the two–equation turbulence model. The comparison with ex-
perimental data available is performed. On the basis of computations the flow wave
picture has been constructed. Special attention is paid to the investigation of the
temperature factor influence on the structure of recirculation zone.
1 Introduction
The flow over a backward-facing step (BFS) is a classical problem in aero-

dynamics closely related to base and near–wake flows. Separated flows and
subsequent formation of shear layer are important for many engineering ap-
plications. For example, BFS is used for ignition and stabilization of the flame
in a scramjet engine. Numerous analytical, experimental and computational
studies of the flow have been performed during last decades. Nevertheless,
the problem still remains one of the CFD challenges because of complex wave
structure and vortex shedding in the tail region. Another point of interest
is turbulence model capabilities to predict the separation zone length and
properties of the shear layer at the reattachment region.
The scheme of an inviscid supersonic flow in vicinity of BFS is shown
in Fig. 1. The flow at Mach number M∞ comes to the BFS of height h
and face angle α. The first flow turn takes place in the expansion fan (EF)
emanating from the external corner and the second flow turn occurs in the tail
shock wave (SW) arising in the internal corner. Figure 2 shows the relative
pressure after the SW as a function of α for M∞ = 2, 2.9, 3.75 and 5 (curves
1 to 4, correspondingly). The solid lines show ”weak” solutions and dashed

714 N.N. Fedorova, and I.A. Bedarev
P2/P∞
M ∞ , P∞ EF 5
M 4
SW
M 1, P1 3
M 2, P2
2
h
1
0
5 15 25 35
Fig. 1. Scheme of inviscid supersonic Fig. 2. Pressure after the BFS for dif-
flow in vicinity of BFS ferent M
lines stand for non-realizable ”strong” solutions of Euler equations for SW.
Figures shows that the pressure drop occurred in EF is fully compensated
by pressure rise in SW only for small angles α. For large α, pressure after
the SW is essentially lower than that in the incoming flow. In inviscid flow
pressure recovery takes place downstream of BFS in the zone of the EF/SW
interaction resulting in SW degeneration in Mach line.
The wave picture of viscous flow is much more complex because of presence
of boundary layer on the step surface and mixing layer over the separation
region. Reflections and deflections of EF characteristics may occur inside the
BL. As a result, EF becomes non-isentropic and uncentered, the lip shock
[7] and the slip line [4] are formed in the region of EF/BL interaction. The
experiments [6] have shown that lines of constant pressure are not straight,
and the angle of flow deflection in EF is smaller than that calculated by the
inviscid theory. Static pressure after the tail SW recoveries much earlier than
the last EF characteristic meet the tail SW. Pressure distribution along the
surface is non-monotone with pronounced local maximum that was explained
by the lip shock action [5] or the action of expansion fan emanated from the
tripple point of the λ-configuration of shock wave over the separation zone
[9].
2 Experimental conditions and method of computations
The computations were carried out under the conditions of the ITAM wind
blow down tunnel (total temperature T 0 = 300∇ · 500K) and the hot – shot
facility (T 0 = 2000∇ · 3000K). The conditions of the experiments (incoming
Mach number M , unit Reynolds number Re1 , total pressure P 0 and temper-
ature T 0) simulated in the present paper are given in Table 1 .
The calculations have been performed on the basis of the full transient
Favre averaged Navier-Stokes equations and Wilcox two equation turbulence
model [8]. The original numerical algorithm [1] was used with the temporal
approximation based on four step finite-difference scheme of splitting accord-
ing spatial variables. At each fractional step the finite difference scheme was
Backward–Facing Steps 715
Table 1. Conditions of the Experiments
M Re1 , 106 /m P 0, KP a T 0, K
2 30 1980 280
2.9 38 418 265
3.75 58 1070 270
5 16.4 800 370
3 36.5 7300 2500
4 77.2 30000 2500
realized by scalar sweeps. The TVD scheme of Flux Vector Splitting by van
Leer [2] and AUSM scheme by Liou [3] of the third order of accuracy has
been used for the approximation of convective terms. The viscous terms have
been approximated with the central finite-difference relations of second order
of accuracy.
The calculation domain does not include the whole model but the part
of it contacted directly to the BFS. From left and right, the computational
domain is bounded with input and output sections. On the top free boundary,
so called ”simple wave” conditions are assigned provided all the disturbances
to come out the computation region. The ”soft” boundary conditions have
been set in the outlet section. In the input section, the incoming flow condi-
tions and the profiles of all gas dynamic and turbulent parameters obtained
from the computation of the turbulent boundary layer on the plate have been
assigned. The no-slip conditions for velocities are specified on the wall. Adia-
batic condition for temperature is used in the computation performed under
the condition of long-duration facility and constant temperature (”cold wall”)
for computations under the condition of hot shot facility. The regular compu-
tational grid condensed toward the solid surfaces has been constructed in the
calculation region. In the most computations, the grid consists of 500 points
in lengthwise direction and 200 points in transverse direction.
3 Computation results
The computations were carried out for α = 8◦ , 15◦ , 25◦ , 45◦ and 90◦ . At
α = 8◦ the flow is attached for all Mach numbers, but because of the inter-
action of the EF characteristics with viscous layer the points when pressure
reach the level determined by inviscid theory go downstream both for expan-
sion and compression corner. At α = 15◦ and 25◦ , a small separation zone
exist in vicinity of internal corner. Fig. 3 presents the computed static pres-
sure distributions (solid lines) for α = 25◦ and α = 45◦ together with inviscid
solutions (dashed lines) and experimental data (symbols). Lines 1 to 4 stand
for M = 2, 2.9, 3.75 and 5, correspondingly. Computed data show good agree-
ment with experimental points for M = 2.9 and 3.75. Figure shows that in
Fig. 3. Static pressure distributions for α = 25◦ (a) and α = 45◦ (b) BFS
vicinity of the external corner pressure first drops and then decreases tending
to an inviscid level. A separation zone may be identified by pressure ”plateau”
region. After the reattachment static pressure rises coming to an inviscid so-
lution. But then new compression waves are formed that increase pressure up
to the incoming level. These additional compression waves can clearly be seen
in Fig. 4 where the pressure contours for M = 2.9 are presented. The origin
of these compression waves is the region where the tail shock interacts with
the mixing layer.
Computations have shown that the flow in vicinity of expansion corner is
dividing into two parts. The first, low-speed, low-density near-wall part turns
on angle α and forms the rare base region. The external layer with higher
velocity and density turns on an angle less then α and forms the mixing layer
going from the inclined step face. It was shown, that after crossing the tail
shock the flow still directs to the wall. In the region of mixing layer / tail
shock interaction the secondary compression waves appear that come to the
wall and reflect from it. Further downstream these secondary compression
waves overtake the tail shock and reflect from it as rarefaction wave. The
system of secondary waves exists far downstream from the interaction zone
causing a non-uniform pressure, skin friction and heat transfer behavior.
Next series of computations was done for α = 90◦ at M = 3 and 4.
Two kinds of flow were simulated, a ”cold” flow (T 0 = 300K) and a ”hot”
flow (T 0 = 2500K). For the hot flow, several temperature wall conditions
20
0
0
40 80 120 160
Fig. 4. Static pressure contours for α = 25◦ BFS at M = 2.9
Backward–Facing Steps 717
P/P1
1.2 Cf1.5
*10
3
2
a 1.25 3 2
1 b
1.00 1 3
0.8 1 0.75 4 1
4
0.6 0.50
0.4 0.25
0.00
0.2 2,3,4 -0.25
0 -0.50
-1 1 3 5 7 x/h9 -1 1 3 5 7 x/h
9
Fig. 5. Pressure (a) and skin friction (b) distributions for α = 90◦ BFS at M = 3
were explored, namely adiabatic conditions and ”cold” wall with different
Twall
temperature factors K = Tadiabatic .
In Figure 5 the pressure (a) and skin friction coefficient (b) distributions
along the step surface are presented for M = 3. Here 1 stands for ”cold”
flow with adiabatic temperature conditions on the wall; 2, 3 and 4 denote
numerical solutions for ”hot” flow with adiabatic wall and K = 0.8 and 0.5,
correspondingly. Figure shows that in the case of ”hot” flow the size of sepa-
ration zone is smaller and the base pressure is lower that those in the ”cold”
flow. The temperature factor has no effect on the static pressure distributions
along the wall, but on the skin friction distribution the differences between
the regimes are evident. For M = 4 case there is the same effect.
Analysis of the flow temperature fields has shown that for both consid-
ered Mach numbers the decrease of temperature factor K essentially reduce
the temperature in the recirculation zone. Figure 6 present the streamlines
for ”hot flow” at M = 3 and different wall temperature conditions: adiabatic
temperature (a) and cold wall with K = 0.5 (b). For adiabatic condition the
recirculation zone has typical two-vortex configuration with high–temperature
external vertex and relative cold near–face vortex. Decreasing the wall tem-
perature leads to significant reduction of near-face vortex and lowering the
temperatures of both vortexes. Since the experiments are conducted in the
hot shot facility with run duration of about 100 ms, such temperature condi-
2 2
a b
1.5 1.5
1 1
y/ h
y/ h
0.5 0.5
0
0 x/h 1 2 0
0 x/h 1 2
Fig. 6. Streamlines at vicinity α = 90◦ BFS at M = 3 for adiabatic (a) and cold
wall K = 0.2 (b) temperature condition
tions may cause the ignition delay or suppressing.
Acknowledgements The work was supported by Russian Foundation for

Basic Research grant 05-01-00684.
References
1. Borisov, A.V. and Fedorova, N.N.: Numerical simulation of turbulent flows near
the forward-facing steps. Thermophysics and Aeromechanics, 4, 1, 69–83 (1996)
2. Van Leer, B. Flux–vector splitting for the Euler equation. Lecture Notes in Phys.
170 507–512 (1982)
3. Liou, M.S., Steffen, C.J.Jr A new flux splitting scheme J. Comp. Phys. 107
23–39 (1993)
4. Charwat, A.F. and Yakura, J.K.: An investigations of two-dimensional super-
sonic base pressures. J. Aeronautical Sci. 25, 122-128 (1958)
5. Hama, F.R.: Experimental studies of the lip shock. AIAA Jour., 6, 1086–1088
(1968)
6. Scherberg M.G., Smith H.E.: An experimental study of supersonic flow over a
rearward facing step. AIAA Jour. 5, 51–56 (1967)
7. Weinbaum, S.:Rapid Expansion of a supersonic boundary layer and its applica-
tion to the near wake. AIAA Jour., 4, 35–47 (1966)
8. Wilcox, D.C.: Turbulence modeling for CFD, DCW Ind. Inc., La Canada, Cali-
fornia (1993)
9. Zheltovodov, A.A., Shuelein, E. Kh., Horstman, C.C.: Development of the sepa-
ration uder the interaction of shock and boundary layer pertuberated by expan-
sion fan. Priikl. Mech. Tekn. Fiz.,34, N 3, 58-68 (1993)
Unified Flow Solver Combining Boltzmann and
Continuum Models for Simulations of Gas
Flows for the Entire Range of Knudsen
Numbers
V. V. Aristov1 , A. A. Frolova1 , S. A. Zabelok1 , V. I. Kolobov2 , and

R. R. Arslanbekov2
1
Dorodnicyn Computing Center of the Russian Academy of Sciences,
Vavilova str. – 40, 119991, Moscow, Russia aristov@ccas.ru,
frolova@ccas.ru, serge@ccas.ru
2
CFD Research Corporation, 215 Wynn Drive, Huntsville, Alabama 35805, USA
vik@cfdrc.com, rra@cfdrc.com
1 Introduction
We have developed a Unified Flow Solver (UFS) for simulations of gas flows
for the entire range of Knudsen numbers from rarefied to continuum regimes
[1, 2]. The UFS methodology is based on the direct numerical solution of the
Boltzmann equation for rarefied flow domains [3, 4] and the kinetic schemes
of gas dynamics for the continuum flow domains [5, 6]. This approach enables
an easy coupling of kinetic and continuum solvers because similar numerical
techniques are used for solving both the Boltzmann and continuum equations.
The UFS can separate rarefied and continuum domains and use appropriate
solvers to combine efficiency of continuum models with accuracy of kinetic
models. Domain decomposition criteria and coupling algorithms are important
part of the UFS. The computational framework of the UFS is provided by
a tree-based data structure of the GERRIS Flow Solver (GFS) [7] enabling
dynamically adaptive Cartesian grid with support of complex boundaries. The
UFS can automatically generate Cartesian mesh around embedded boundaries
defined through standard files, perform dynamic adaptation of the mesh to
the solution and geometry, detect kinetic and continuum domains and select
appropriate solvers based on continuum breakdown criteria. Parallelization
of the UFS has been performed enabling dynamic load balancing for multi-
processor systems. This paper presents solutions of several test problems for
supersonic and subsonic flow regimes to illustrate current capabilities of the
UFS for different applications. UFS extensions to multi-component mixtures
of polyatomic gases are also discussed.

720 V. V. Aristov et al.
2 Description of kinetic schemes of direct methods

Let us briefly describe the direct method of solving the Boltzmann transport
equation (BTE) developed earlier in [3, 4]. The Boltzmann equation is the
fundamental transport equation, describing the evolution of a velocity distri-
bution function f in phase space (r, ξ). Introducing Cartesian grid with nodes
ξβ in velocity space, the BTE is reduced to a set of equations in physical space
∂fβ
+ ∇r · (ξβ fβ ) = I (fβ , fβ ) .
∂t
We split the solution of this system of equations into two stages: free flow
and relaxation. For the free flow, we use an explicit finite volume numerical
scheme and for the relaxation stage the explicit finite-difference scheme is
applied. Collision integrals are considered in the symmetric form of 8-fold
integrals. The quasi-Monte Carlo procedure with the Korobov sequences is
used for evaluation of the collision integrals. We currently use uniform grid in
velocity space.
For the continuum equations, kinetic schemes are used. Generally, kinetic
schemes are preferable for developing hybrid codes since the BTE is used as
a foundation for both algorithms. Kinetic schemes for the Euler and Navier-
Stokes equations by means of the distribution function have been proposed,
e.g., in [8]. Kinetic schemes using moments of the equilibrium distribution
function in the modern variant are developed in [5, 6, 9]. Recently, we extended
our Euler-kinetic scheme for gas mixtures with chemical reactions.
We use the Gerris framework to generate dynamically adaptive Cartesian
mesh in physical space. The parallel algorithm with dynamic load balancing
among processors has been developed.
The main problem of unified methods is to separate kinetic and continuum
regions. We have used several switching criteria, one of them is as follows:
v
u 2
∂u 2 ∂v ∂w 2

+ +
u
u 2 ∂x ∂y ∂z
Kn t ∇p
SN S = + ,
ρ p u2 + v 2 + w 2
where ρ is the density, p is the pressure, u, v, w are appropriate components of
velocity (all values are given in dimensionless form), Kn is the global Knudsen
number (e.g., for a flow around a Kn = λ/R cylinder where λ is the mean
free path and R is the radius of a cylinder).
The Fig. 1 illustrates some current capabilities of the UFS. In Fig. 1,
domain decomposition into kinetic and continuum regions is presented for
supersonic gas flow around a cylinder at Mach number M = 3, for different
Knudsen numbers Kn.
We have studied the influence of the breakdown parameter on the flow
characteristics calculated by the UFS. Left plot of Fig. 2 shows the gas pres-
sure and the heat flux distributions over the cylinder surface calculated with
Unified Flow Solver Combining Boltzmann and Continuum Models ... 721
Fig. 1. Gas flow around a cylinder for M = 3 for different Kn numbers (0.5, 0.005).
The density profiles are on the left side, the computational grids with kinetic (red)
and continuum (white) domains are on the right side.
Fig. 2. Left plot: pressure and heat flux on the cylinder surface as a function of angle
for different values of the breakdown parameter S. Right plot: distribution of nor-
mal force over the cylinder surface for different values of the continuum breakdown
parameter S
Fig. 3. Left plot: drag coefficient versus Kn number. Comparison of UFS simulations
with experiment. Solid lines indicate the free molecular flow and continuum limits.
Right plot: pressure on the cylinder surface
different breakdown parameters S for supersonic gas flow around a cylinder

at M = 2, Kn = 0.1 and Kn = 2.28. In right plot of Fig. 2 the convergence of
the computations with respect to a breakdown parameter is illustrated for the
Fig. 4. Mass flow rate versus Knudsen number for a channel L/d = 21 (L is length
and d is width of the channel), P0 /P1 = 2.0 (left plot), P0 /P1 = 1.5 (right plot)
case of M = 3, Kn = 0.25. One can see that all curves converge at small S
numbers when the Boltzmann region grows. At the same time, by decreasing
the S number, the computation time increases. Therefore, for quick results
one can use larger S numbers if precision of the order of 10% is satisfactory.
3 Results of simulations
We compare solutions for a drag coefficient for a circular cylinder (hard

spheres, monatomic gas) with evidence of the experiment [10] for diatomic
gas in a wide range of Knudsen numbers. Results are presented in Fig. 3.
On the other hand the comparison is made for some parameters with
recent calculations on the basis of different DSMC techniques by Bird and
Boyd et al. (see [11]). In the last under consideration the Mach M = 10
flow of argon past a circular cylinder with diameter 0.3048m and temperature
500 K at Knudsen number Kn = 0.01. The stream velocity is 2634.1m/s,
the number density 4.247 · 1020 m−3 and the temperature 200 K. According
to computations of mentioned authors the overall drag per meter width is
40.0 N, in our computations by means of the UFS (Boltzmann+Euler) this
value is 40.3 N, therefore the difference of results is less than 1%. The pressure
distribution on the cylinder surface (see Fig. 3) is an good agreement with
results of [11].
Subsonic flows is an interesting class of processes which is recently in-
tensively investigated due to their possible applications in micro-technology.
Flows in channels at different regimes were studied by the UFS. The UFS al-
lows us to consider effects in a wide range of Knudsen numbers. In Fig. 4 the
so-called Knudsen minimum of a mass flow rate for two values of the pressure
ratio is presented for 2D case.
An example of 3D simulations on 7-processor Linux cluster is shown in
Fig. 5 for the Inflatable Reentry Vehicle Experiment (IRVE). The flow condi-
tions are for 91km altitude: Kn = 0.01 and M = 3.94 [12]. The flow is at zero
angle of attack. Streamlines and a velocity field are given. The flow recircula-
Unified Flow Solver Combining Boltzmann and Continuum Models ... 723
Fig. 5. Streamlines and a field of velocities
tion is observed in the kinetic domain behind the body. The streamlines are
not closed due to 3D effects.
In Fig. 6 distribution of parameters of different components of air with
chemical reactions for a flow at Mach number M = 2 about the cylinder
along the central streamline are shown. The kinetic-Euler scheme with chemi-
cal reactions is used. It is considered 17 chemical reactions and 5 components,
Fig. 6. Species densities, velocity, temperature and vibrational temperatures along

the central streamline
namely O2 , N2 , O, N and N O.
Acknowledgements. This research is supported by the SBIR Project F33615-

03-M-3326 and also in part by the Russian Foundation for Basic Research,
Grant No 04-01-00347 and by the Presidium of the Russian Academy of Sci-
ences, Project No 14.
References
1. Kolobov, V.I., Bayyuk, S.A., Arslanbekov, R.R., Aristov, V.V., Frolova, A.A.,
Zabelok, S.A.: Construction of a Unified Continuum/kinetic Solver for Aerody-
namic Problems, J. Spacecraft and Rockets, 42, 598 (2005)
2. Kolobov, V.I., Arslanbekov, R.R., Aristov, V.V., Frolova, A.A., Zabelok, S.A.:
Unified Flow Solver for aerospace application, AIAA Paper 2006-0988
3. Aristov., V.V.: Direct Methods for Solving the Boltzmann Equation and Study
of Non-Equilibrium Flows, Kluwer Academic Publishers, Dordrecht (2001)
4. Tcheremissine, F.G.: Conservative evaluation of Boltzmann collision integral in
discrete ordinate approximation, Computers Math. App. 35, 215 (1998)
5. Xu, K.: A Gas-kinetic BGK Scheme for the Navier-Stokes Equations and its
Connection with Artificial Dissipation and Godunov Method, J. Comp. Phys.
171, 289 (2001)
6. Ohwada, T., Fukata, S.: Simple Derivation of High-resolution Schemes for Com-
pressible Flows by Kinetic Approach, J.Comp. Phys. 211, 424 (2006)
7. http://gfs.sourceforge.net/
8. Aristov, V.V., Tcheremissine, F.G.: Solving the Euler and Navier-Stokes equa-
tions on the basis of the operator splitting of the kinetic equations, Doklady
USSR Acad. Sci. 272, 555 (1983)
9. Kolobov, V.I., Arslanbekov, R.R., Aristov, V.V., Frolova, A.A., Zabelok, S.A.:
Unified Solver for Rarefied and Continuum Flows with Adaptive Mesh and Al-
gorithm Refinement J.Comput.Phys. (submitted).
10. Maslach, G.J., Schaaf, S.A.: Cylinder Drag in the Transition from Continuum
to Free-Molecule Flow, Phys. Fluids 6, 315 (1963)
11. Lofthouse, A.J., Boyd, I.D., Wright, M.J.: AIAA Paper 2006-993
12. Moss, J.N. et al: Low-Density Aerodynamics of the Inflatable Reentry Vehicle
Experiment (IRVE), AIAA Paper 2006-1189
Adaptive Boltzmann/Navier-Stokes Hybrid
Method for Multi-Scale Gas Flow Simulation
Koji Morinishi1
Kyoto Institue of Technology Matusgasaki, Sakyo-ku, Kyoto 606-8585, Japan

morinisi@kit.ac.jp
1 Introduction
Numerical flow simulation about micro devices is one of the new frontiers of
computational fluid dynamics (CFD). It may gives essential knowledge about
gas flows around micro-electro-mechanical systems (MEMS). Since the flows
about the micro devices range from continuum to free molecule flow regimes,
numerical methods designed for automatically simulating the continuum to
free molecule, multi-scale flows are preferable.
Usual continuum methods can not predict flows in transitional and free
molecule regimes. While any kinetic approaches, such as the direct simulation
Monte Carlo (DSMC) method [1], can be theoretically applied to the multi-
scale flows, the approach is computatinally several orders of magnitude more
expensive than continuum methods, especially for flows in the continuum
regime. Thus an adaptive continuum/kinetic hybrid method, in which a kinetic
method is automatically used in the transitional and free molecule regimes
and a continuum method in the rest, is highly preferable for simulating the
multi-scale gas flows.
In the past decade, several studies [2, 3] about continuum/kinetic hybrid
approaches have been reported. Most of them are hybrid approaches using the
Navier-Stokes method and the DSMC method. The DSMC method has widely
been used for simulating high speed rarefied flows. The method, however,
becomes a poor simulation tool for the low speed flows about micro devices,
because huge sample size is required to reduce the inherent statistical scatter
of DSMC method to a level of the small changes of flow quantities in the low
speed flows [4].
In contrast to the DSMC method, a CFD method based on the kinetic
model Boltzmann equation [5, 6] is free from the statistical scatter. The nu-
merical Boltzmann solver is definitely superior to the DSMC method for
simulating the subsonic flows. Moreover a Navier-Stokes/Boltzmann hybrid
method, may be easily constructed, since the same CFD methods can be

726 Koji Morinishi
adopted for the kinetic model Boltzmann equation as well as the Navier-
Stokes equations. In this paper, a novel continuum/kinetic hybrid solver is
proposed for simulating the continuum to free molecule, multi-scale flows, by
coupling the Navier-Stokes solver and the kinetic model Boltzmann solver.
2 Kinetic Model Equation

The kinetic model Boltzmann equation in nondimensional form without any
external force may be written as follows:
∂f ∂f
+c· = ν(f0 − f ) (1)
∂t ∂x
where f is the velocity distribution function which depends on the time t, the
physical space x, and the molecular velocity c. The distribution function f0 of
the BGK model, which is the most fundamental model, is the local equilibrium
distribution function fe :
(c − u)2

n
fe = exp − (2)
(πT )3/2 T
where n is the number density, u the macroscopic flow velocity, and T the
temperature. For a higher order model equation which correctly resembles the
lower 13 moments of the Boltzmann equation [7], the distribution function f0
is given as
(c − u) · q 2(c − u)2

f0 = fe 1 + 2(1 − P r) −5 (3)
5pT T
where p is the pressure, q the heat flux vector, and P r the Prandtl number
(=2/3 for a monatomic gas).
The macroscopic flow quantities are obtained as follows:
Z
n = f dc
Z
nu = cf dc
Z (4)
3
nT = (c − u)2 f dc
2
Z
q = (c − u)(c − u)2 f dc
The pressure p is obtained from the equation of state:

p = nT (5)
All these quantities are normalized with a reference length L, a reference
number density n∞ , a reference temperature T∞ , and a reference velocity
Adaptive Boltzmann/Navier-Stokes Hybrid Method 727
C∞ . The reference velocity C∞ is the most probable molecular thermal speed

which is defined as:
p
C∞ = 2RT∞ (6)
The collision frequency ν is usually defined as:
8nT 1−s
ν= √ (7)
5 πKn
where Kn is the reference Knudsen number based on the reference length L
and the molecular mean free path λ∞ at reference state which is defined as
16µ∞
λ∞ = √ (8)
5mn∞ 2πRT∞
where µ is the viscosity coefficient, m the mass of a molecule, and R the gas
constant. The viscosity coefficient is assumed to depend on the temperature
as:
µ ∝ Ts (9)
where the Maxwell molecules correspond to the power s of 1 and the hard
sphere molecules to 0.5. In this study, the hard sphere molecules is adopted.
3 Navier-Stokes Equations
The basic equations of continuum approach is the compressible Navier-Stokes

equations which may be written in the following nondimensional form.
∂Q 1
+∇·F= ∇·R (10)
∂t Re
where Q is the conservative vector, F the convective flux, R the viscous flux,
and Re the reference Reynolds number. The conservative vector and the flux
terms are given with:
     
ρ ρu 0
Q =  u  , F =  ρuu + pI  , R =  τ  (11)
e u(e + p) u·τ −q
where e is the total energy per unit volume, which is given for a perfect gas
as:
p 1
e= + ρu2 (12)
γ−1 2
Here γ is the ratio of specific heats. The viscous stress tensor τ and the heat
flux vector q are defined with:

∂ui ∂uj 2 ∂ui
τij = µ + − δij (13)
∂xj ∂xi 3 ∂xj
728 Koji Morinishi
γ µ ∂T
qi = − (14)
γ − 1 P r ∂xi
All these quantities in the continuum approach are normalized with a
reference length L, a reference density ρ∞ (= mn∞ ), a reference temperature
T∞ , and a reference velocity U∞ . The reference velocity U∞ is defined as:
p
U∞ = RT∞ (15)
where nondimensional velocities in the continuum
√ approach are greater than
those in the kinetic approach by the ratio of 2.
4 Hybrid Approach
Domain decomposition strategy is adopted for the continuum/kinetic hybrid
approach. A computational domain is decomposed into the continuum regions
and the kinetic regions. The Navier-Stokes solver works on the points in the
continuum regions and the kinetic model Boltzmann solver in the kinetic
regions.
An upwind gridless method [8] with the third order accurate weighted es-
sentially non-oscillatory (WENO) reconstruction is used for discretizing both
the kinetic model Boltzmann equation and the compressible Navier-Stokes
equations. At the interface between the continuum and kinetic regions, infor-
mation exchange is necessary at every time step over the two layers of either
side of the interface, because the stencil of the third order accurate WENO
reconstruction is generally five. At the points incontinuum regions two layer
inside from the interface, Chapman-Enskog distribution functions for the ki-
netic model Boltzman equations are constructed as:
2
1 τij Ci Cj 1 C 5 Ci ∂T
f = fe 1 − + − (16)
ν µ T Pr T 2 T ∂xi
On the other hand, the numerical flux of the Navier-Stokes equations can
be simply evaluated over the interface, since the macroscopic flow quantities
are obtained using the equation (4) at every time step even in the kinetic
regions.
As the continuum/kinetic interface indicator, the gradient-length local
Knudsen number KnGLL
λ
KnGLL = |∇ρ| (17)
ρ
and Garcia’s parameter B [2]
∗
B = max{|τij |, |qi∗ |} (18)
∗
are used. Here τij
and qi∗ are the normalized stress tensor and the normalized
heat flus, respectively.
Adaptive Boltzmann/Navier-Stokes Hybrid Method 729
Local Kn Full Kinetic
2.0
1.5
1.0
0.5
2.5
2.5
0.5
1.0
2.0
1.5
Hybrid
Fig. 1. Contours of local Knudsen Fig. 2. Density contours obtained

number and points which belong con- with full Boltzmann solver and
tinuum and kinetic regions. Boltzmann/Navier-Stokes hybrid at
M∞ = 2.0 and Kn = 0.01.
B-function Full Kinetic
2.0
1.5
1.0
0.5
2.5
2.5
0.5
1.0
2.0
1.5
Hybrid
Fig. 3. Contours of Garcia’s parame- Fig. 4. Density contours obtained

ter B and points which belong contin- with full Boltzmann solver and
uum and kinetic regions. Boltzmann/Navier-Stokes hybrid at
M∞ = 2.0 and Kn = 0.01.
5 Numerical Results
Numerical experiments are carried by the adaptive Boltzmann/Navier-Stokes

hybrid method for supersonic and subsonic flows over a circular cylinder at
the transitional flow regime. The free stream Mach numbers are 2.0 for the
supersonic flow case and 0.1 for the subsonic flow case. Knudsen number based
on the diameter of circular cylinder is 0.1 for both the cases. The gradient-
length local Knudsen number KnGLL as well as the Garcia’s parameter B
are tested as the continuum/kinetic interface indicator. Only results for the
supersonic flow case is presented here.
730 Koji Morinishi
Figure 1 shows the gradient-length local Knudsen number contours and

points which belong to the continuum region ◦ and the kinetic region • at the
supersonic flow case. The density contours obtained with the Boltzmann/Navier-
Stokes hybrid method are compared with those of full Boltzmann method in
Fig. 2. The both results obtained for the full flow field around the circular
cylinder, while the result of the adaptive hybrid method is plotted for the
lower half flow field and the full Boltzmann method for the upper half flow
field. The comparison is fairly good.
Numerical results obatined using the continuum/kinetic interface indicator
of Garcia’s parameter B are shown in Figs. 3 and 4 for the same flow field
as Figs. 1 and 2. The comparison between the full kinetic results and the
continuum/kinetic hybrid results is again very good.
References
1. G.A. Bird. Molecular gas dynamics and the direct simulation of gas flows, Oxford
Science Publications, (1994).
2. A.L. Garcia, J.B. Bell, W.Y. Crutchfield, and B.J. Alder. Adaptive mesh and
algorithm refinement using direct simulation Monte Carlo. Journal of Computa-
tional Physics, 154, 134–155, (1999).
3. Q. Sun, I.D. Boyd, and G.V. Candler. A hybrid continuum/particle approach for
modeling subsonic, rarefied gas flows. Journal of Computational Physics, 194,
256–277, (2004).
4. E.S. Oran, C.K. Oh, and B.Z. Cybyk. Direct simulation monte carlo: Recent
advances and applications. Annual Review of Fluid Mechanics, 30, 403–442,
(1998).
5. K. Morinishi and H. Oguchi. A Computational method and its application to
analyses of rarefied gas flows. Proc. 14th International Symposium on Rarefied
Gas Dynamics, University of Tokyo Press, Vol. 1, 149–158, (1984).
6. K. Morinishi. Numerical simulation for gas microflows using Boltzmann equa-
tion. Computers & Fluids, 35, 978–985, (2006).
7. T. Abe and H. Oguchi. A hierarchy kinetic model and its applications. Proc. 10th
International Symposium on Rarefied Gas Dynamics, Progress in Astronautics
and Aeronautics, AIAA, 58 781–793, (1977).
8. K. Morinishi. Parallel computation of higher order gridless type solver. Parallel
Computational Fluid Dynamics, Elsevier, 427–434, (2003).
Computing Simulation of Hypersonic
Magneto-Fluid-Dynamics Interaction
J.S. Shang∗
Department of Mechanical and Materials Engineering

Wright State University, Dayton Ohio, USA
1 Introduction
Numerous ideas have been advocated for using electromagnetic force for high-
speed flow control [1-4]. Unfortunately, the inefficient ionization process of
flow medium becomes the major obstacle for applying electromagnetic force
for flow control [5]. In an attempt to accentuate the magneto-fluid-dynamic
(MFD) interaction, some of the research efforts apply an externally magnetic
field to intensify the MFD interaction by invoking the Lorentz force in addi-
tion to Joule heating [1,2,3]. Surzhikov and Shang [6] have also shown that
the Hall current exerts significant influence to the plasma generation via the
electron collision process. All the aforementioned computing simulations are
in a general agreement with experimental observations by Bityurin et al [4].
Therefore, the only viable plasma flow control mechanism may require the
implementation of the plasma actuator as a small perturbation and amplifies
by aerodynamic viscous-inviscid interaction [1,2].
The most effective location of an electromagnetic perturbation is the lead-
ing edge domain where the flow is laminar and the intensity of the perturba-
tion can be relatively weak, but exerts a profound influence on the entire flow
field. The classic hypersonic flow theory by Hayes and Probstein describes the
inviscid-viscous interaction over a sharp leading edge as the pressure inter-
action [7]. The magnitude of the induced pressure is well-known and can be
calculated by a single interaction parameter χ = M 3 (C/Rx)1/2 . The viscous-
inviscid interaction is strongly amplified in hypersonic flows due to the cubic
power dependence on the freestream Mach number. This control mechanism
can be actuated in microseconds to induce a higher surface pressure domain
downstream. Results from both experimental and computational efforts have
shown the chain of events constitute a very effective hypersonic flow control
∗
Emeritus Scientist of US Air Force Research Laboratory and Research Professor
of Wright State university, joseph.shang@wright.edu

732 J.S. Shang
technique. The triggering mechanism is a simple direct current surface dis-

charge [1,2].
The present computational investigation simulates the basic mechanism
of a MFD interaction by an actuated direct current discharge (DCD). The
combined volumetric Joule and convective electrode heating thickens the lo-
cal displacement thickness. The increased slope of the displacement thickness
through viscous-inviscid interaction generates a series of compression waves
and coalesces into an oblique shock. The resultant compression is equally ap-
plicable as a virtual leading edge strake or sidewall compression of high-speed
inlet. The present numerical simulation is focused on the MFD phenomenon
that is initiated by a DC surface discharge and followed by viscous-inviscid in-
teraction in a rectangular inlet. The adopted numerical procedure for solving
the MFD equation in the low Magnetic Reynolds number limit is a semi-
discrete finite-volume scheme.
2 Analysis
In most aerodynamic application of flow control using plasma actuator, the
Magnetic Reynolds number, the Rem = µm σuL is much less than unity [2].
According to the investigated flow condition, Rem has a value of 8.48 × 10−6 ;
thus the governing equations of the low Magnetic Reynolds number approxi-
mation are fully justified for the present analysis [1,2,6]:
∂ρ
+ ∇ · (ρU ) = 0 (1)
∂t
∂ρU
+ ∇ · ρU U + pI − τ = (J × B) (2)
∂t
∂ρe
+ ∇ · [(ρe + p) U − (U · τ ) − Q] = E · J (3)
∂t
The governing equations system is identical to the Navier-Stokes equations
except the non-zero source terms. While the DCD has the maximum charged
particles number density in the cathode layer is around 8.8 × 1011 /cc, the
mass fraction is less than 10−6 . It is therefore permissible to consider only
the transport properties of the weakly ionized gas and to neglect the effects
of nonequilibrium thermodynamics and chemical kinetics. In the above for-
mulation, the electrical current density, J and electric field intensity,E that
appear in the governing equation are evaluated using a three-components,
two-temperature plasma model by Surzhikov and Shang [6].
The partially weakly ionized gas model is based on the two dominant mech-
anisms in plasma - the drift velocities and diffusion (including the ambipolar
diffusion) of discharged particles. The compatible electrical field intensity, E,
of the discharge domain is obtained by satisfying the charge conservation equa-
tion [1,6]. In globally neutral plasma field, the charge conservation equation
reduces to the Poisson equation of electrical potential, ϕ, E = −∇ϕ:
Hypersonic Magneto-Fluid-Dynamics Interaction 733
e
∇2 ϕ = (ne − ni ) (4)
ε
∂ne
+ ∇ · Γe = α(E, p)|Γe | − βne n+ (5)
∂t
∂n+
+ ∇ · Γ+ = α(E, p)|Γe | − βne n+ (6)
∂t
where the flux vectors of the charged particles densities for electrons and ions
are Γe = −De ∇ne − ne µe E and Γ+ = −D+ ∇n+ + n+ µ+ E. Once the flux
vectors are known, the electrical current density of plasma can be obtained
as J = e(Γ+ − Γe ). α and β are the Townsend discharge and recombination
coefficients. De , D+ are the electron and ion diffusion coefficients, and µe , µ+
are the mobility of the electron and ion respectively [6].
Since the wave speeds of the governing equations span an extraordinary
range from the sonic speed to speed of light, a loosely coupled solving scheme
was developed for the MFD and electro-dynamics equations. Both the govern-
ing equations can be cast into the flux vector form [1,2]. The spatial dsicretiza-
tion is based on a semi-discrete finite-volume scheme [8]. In the flux-difference
splitting procedure, the flux vectors at the control surface are written as an
exact solution to the approximate Riemann Problem.
The initial values and boundary conditions for the governing equation
system are straightforward. For the fluid dynamic variables, the no-slip con-
dition for the velocity components, the vanishing pressure pressure gradient
condition, as well as the prescribed surface temperature are imposed on the
solid surface. For the electromagnetic variables, the full electric potential is
imposed on the anode in reference to the cathode. A key element of bound-
ary condition for the DCD is specified the electron number density on the
cathode for the secondary emission phenomenon. This physical requirement
is met by requiring the normal component of the electron flux at the cathode
to be proportional to the ion counterpart: Γe · n = γΓ+ · n. Where γ is the
secondary emission coefficient which depends on the material of the cathode
and intensity of the electrical field [5]. In the earlier and present investigations,
the value of 0.1 was used [1,2,6].
Taking advantage of the symmetrical property of the rectangular inlet
and the electrodes placement, two relatively sparse mesh systems are used for
numerical resolution assessment. The coarse and fine mesh systems consist of
(85 × 45 × 81) and (105 × 57 × 101) in the ξ, η, and ζ coordinates respectively.
The solutions of the two mesh system differ by a maximum of less than 4% in
the computed Pitot pressure. All the presented numerical results are derived
from the fine-mesh calculations. For convergence acceleration, three-level mesh
sequencing of the multigrid technique is applied [9]. The data processing rate
on a 400 mHz SGI Octane workstation is 61.6 × 10−6 seconds per number
of cells per number of iterations. The convergence criterion of the present
analysis is preset at a value of 3.0 × 10−7 of the normalized global residue for
the finite-volume scheme.
734 J.S. Shang
Fig. 1. Direct current discharges in hypersonic rectangular inlet
3 Discussion of Results
The present numerical simulation duplicates the experiment in a Mach 5.15
plasma channel. The flow field is characterized by a static pressure of 78.4 Pa,
a temperature of 43 K, and the streamwise velocity of 675.5 m/s. Under these
condition, the Reynolds number per meter is 2.57×105 . The surface plasma
is generated by a Universal Voltronic DC power supply that can deliver a
current of 800 mA up to 10 kV.
The relatively small rectangular inlet model has the physical dimension of
101.6×38.1×31.1mm (L×W×H). The electrodes are embedded in the phenolic
vertical sidewalls. The cathodes are located upstream of the anodes and at a
distance of 7.9 mm downstream of the sharp leading edge. All electrodes have
the identical dimensions of 31.8×6.4×1.6 mm (length×width×depth).
The DC glow discharge over the vertical side walls is displayed in Figure
1. The surface plasma after the initial breakdown is maintained by an electric
potential of 800 V and a current of 80 mA in the external electric circuit. The
maximum charged particles number density is 8.8×1011 at a total surface DC
discharge power supply of 64 Watts. The electric current density on the anode
is 21.27 mA/cm2 which put the discharge in the abnormal mode. The DCD
in fact exhibits noticeable fluctuations in the Pitot pressure data [2]. The
dominant and visible emission is over the cathode region. This distinctive
feature is the consequence of a layer of positive space charge at the cathode
with a substantial drop of electrical potential.
In the absence of an externally applied transverse magnetic field, the prin-
cipal mechanism of the electromagnetic perturbation to the flowfield is Joule
heating. The contours of the computed Joule heating in the x-z plane is de-
Fig. 2. Joule heating distribution within the inlet

Fig. 3. Comparison of Pitot pressure distributions with data
picted in Figure 2. Although 64 watts of energy is supplied the plasma gener-

ation, but only a fraction of the energy is dissipated by the Joule heating. The
computed value indicated a value of 15.5%; this result is in general agreement
with the best estimated from experimental observation and chemical kinetic
theory [5,6]. It is observed that the Joule heating is concentrated in the cath-
ode layers. The thickened boundary-layer by the volumetric Joule and con-
vective electrode heating displaces the stream away from the side walls and
induces compression waves at the leading edge of the cathode. These com-
pression waves coalesces and reinforces the oblique shocks originated from the
leading edge of the inlet.
A part of the extensive verification of numerical results with experimental
data is presented in Figure 3. The measured Pitot pressure data along the
centerline of the inlet is presented together with the computational results
either with actuated or deactivated activated surface discharge. Computed
results capture the physics of the interacting shock-shock and shock-boundary-
layer interactions. When the DCD is actuated, the induced oblique shocks
become steeper and move the shock incipient point upstream. The computed
results reach a very good agreement with the experimental observation. The
computed result underpredicts the peak Pitot pressure by 2%. The similar
discrepancy is also noticed in the comparison of Pitot pressure profiles in
cross-section planes of the inlet [2].
Figure 4 presents the surface static pressure distributions across the inlet at
three different streamwise location with and without the actuated DCD. This
presentation reveals that a complex flowfield structure exists in the rather
simple configuration. The flowfield consists of four mutually perpendicular
shocks but intersect each other at different streamwise location. Upstream to
the shock intersections, the four perpendicular shock surfaces must be sepa-
rated by connecting shocks with two triplet points each. Thus in the enveloping
shock surface contains a total of eight triplet points in the inlet entrance re-
gion. Downstream of shock intersections, the shock-shock interaction ensues.
However, the key point of MFD interaction stands out in that the DCD per-
736 J.S. Shang
Fig. 4. Static pressure behavior with/without DC discharge
turbation is amplified by the hypersonic pressure interaction to act as plasma

induced compression. At the inlet exit, x/L=1.0, the integrated static pressure
is elevated by 11.7% over that of the DCD deactivated counterpart.
Conclusion
The magneto-fluid-dynamic interaction within a rectangular, constant cross-

section area inlet has been successfully investigated by solving the interdis-
ciplinary fluid dynamic equations with a weakly ionized gas model. The nu-
merical results duplicate the global behavior of a hypersonic interacting phe-
nomenon and demonstrate an innovative and viable hypersonic flow control
mechanism. The control effectiveness is illustrated by the low power input of
7.87W/cm2 per unit electrode area for plasma generation to achieve an addi-
tional compression of 11.7% without any moving components.
Acknowledgment. The author gratefully acknowledges the support from Dr.

Fariba Fahroo, Dr. John Schmisseur, and Dr. Donald Paul of the U.S. Air Force Re-
search Laboratory. The simulating exchanges with Prof. Sergey Surzhikov of Russian
Academy of Science and my colleague Dr.James Menart are deeply appreciated.
References
1. Shang, J.S. and Surzhikov, S.T., Magneto-aerodynamic Actuator for Hypersonic
Flow Control, AIAA Journal Vol. 43, No. 8, August 2005, pp. 1633-1643.
2. Shang, J.S. Surzhikov, S.T., Kimmel, R. Gaitonde, D.V., Hayes, J.R., and
Menart, J., Mechanism of Plasma Actuators for Hypersonic Flow control,
Progress in Aerospace Sciences, Vol. 41, No. 8, Nov. 2005, pp. 642-668.
3. Macheret, S.O., Shneider, M.N., and Miles, R.B., Magnetohydrodynamic Control
of Hypersonic Flows and Scramjet Inlets Using Electron Beam Ionization, AIAA
J., Vol. 40, January 2001, pp.74-81
4. Bityurin, V.A., Bocharov, A.N., and Lineberry, J.T., Results of Experiments on

MHD Hypersonic Flow Control, AIAA 2004-2263, Portland OR, June 2004.
5. Raizer, Yu, P., Gas Discharge Physics, Springer-Verlag, Berlin, 1991, pp.52-72.
6. Surzhikov, S.T. and Shang, J.S. Two-Component Plasma Model for Two-
Dimensional Glow Discharge in Magnetic Field, J. Computational Physics, Vol.
199, 2, Sept. 2004, pp. 437-464.
7. Hayes, W.D., and Probstein, R.F., Hypersonic Flow theory, Academy Press,
1959, pp. 333-374.
8. Rumsey, C., Biedron, R., and Thomas, J., CFL3D: Its History and Some Recent
Applications, NASA TM-112861, May 1997; presented at the Gounov’s Method
for Gas Dynamics” Symposium, Ann Arbor, MI, May 1997.
9. Thomas, J.L., Diskin, B., Brandt, A., Textbook Multigrid Efficiency for Fluid
Simulation, Annual Review of Fluid Mechanics, Vol. 35, January 2003, pp.317-
340.
Numerical Simulation of R-M instability
at High Mach Numbers
Fu Dexun, Ma Yanwen, and Liang Xian
Institute of Mechanics, Chinese Academy of Sciences, Beijing 100080, China
Summary. In the present paper R-M instability problems with high Mach number
and high density ratio are studied numerically. The compressible N-S equations in
cylindrical coordinate are approximated with a new developed sixth order accurate
WGVC scheme (weighted group velocity control). Some results of interaction be-
tween a cylindrical shock and a perturbed cylindrical material interface with shock
Mach numbers Ms = 1.2, 5, 10 and Atwood numbers A = 0.818, 0.905, 0.961,
0.980 (corresponding to ρρout
in
= 10, 20, 50, 100) are presented. The effects of shock
strength, Atwood number and mode competition on characteristics of R-M instabil-
ity are discussed.
1 Introduction
Recently scientists pay more attention on the research of Richtmyer-Meshkov

(R-M) instability because it is a basic physical phenomenon present in many
important practical applications, for example, the propagation of sound boom
in turbulent atmosphere, the Inertial Confinement Fusion (ICF), the explosion
of the supernova and so on. With rapid advances in computational technol-
ogy the direct numerical simulation becomes an important tool for studying
this kind of problems. But most of the simulations of R-M instability were
restricted to planar geometry with small shock Mach number [1,2,3]. For the
case of interaction between cylindrical shock and cylindrical interface there
are some numerical results but their shock Mach numbers and Atwood num-
bers were small [4]. The Mach number was high in [5] but the Atwood number
was small (A = 0.672).
In the present paper a sixth order accurate difference scheme with weighted
group velocity control (WGVC) is developed in order to simulate the R-M
instability with high shock Mach number and high density ratio. The N-S
equations are approximated with the developed method, and the instability
resulting from the interaction between a cylindrical material interface and a
cylindrical shock wave imploding from a heavy fluid outside of the cylindrical

Numerical Simulation of R-M instability at High Mach Numbers 739
material interface onto a light fluid inside is studied numerically. Some results
for the case of shock Mach numbers M a = 1.2, 5, 10, and Atwood numbers
A = 0.818, 0.905, 0.961, 0.980 are presented. The effects of shock strength,
Atwood numbers and mode competition on R-M instability characteristics are
discussed.
2 High order accurate scheme

For correctly simulating the R-M instability problems with high Mach number
and high Atwood number it is required that the method could capture both
the strong discontinuities (shock and contact discontinuity) and the small
scale structures with sufficient precision. In this paper we developed a new
weighted group velocity control method.
Consider the simple wave equation. Define ∆x(∂f /∂x)j = Fj , Fj =
hj+1/2 − hj+1/2 where hj+1/2 is called the numerical flux. Suppose we have
the following 2kth order accurate symmetrical difference approximation on
the grid points j − k, ..., j − 1, j, j + 1, ..., j + k
(2k,0) (2k,0) (2k,0)
Fj = hj+1/2 − hj+1/2
from which after Taylor series expansion we can obtain
(2k,0) (2k,0)
hj+1/2 − hj+1/2 = ∆x(∂f /∂x)j
+ C∆x2k [∂ 2k f /∂x2k ]j+1/2 − C∆x2k [∂ 2k f /∂x2k ]j−1/2 + O(∆x2k+2 ) (1)
where C is a known constant obtained after Taylor series expansion. With the
enlarged grid points j − k − 1, ..., j, ..., j + k + 1 after discretization of the
terms in [ ] we can get the expression
1
[∂ 2k f /∂x2k ]j+1/2 = ∆x2k [δx2k fj + δx2k fj+1 ].
2
1
where δx2k fj = fj−k − C2k fj−k+1 + · · · + (−1)m C2k
m
fj−k+m + · · · + (−1)2k fj+k .
A higher 2(k + 1)th order accurate symmetrical numerical flux can be written
as
(2k+2,0) (2k,0) C
hj+1/2 = hj+1/2 − [δx2k fj + δx2k fj+1 ]. (2)
2
If the approximation ∆x2k [∂ 2k f /∂x2k ]j+1/2 = δx2k fj−m is used, the following
(2k + 1)th order accurate upwind biased numerical flux can be obtained
(2k+1,+) (2k,0)
hj+1/2 = hj+1/2 − Cδx2k fj−m . (3)
where mε0 for the case C > 0. For the case of large integer m the scheme is
strongly upwind biased, and it may not be stable. The simplest case is m = 0.
740 Fu Dexun, Ma Yanwen, and Liang Xian
As it is noted in Ref.[6], for getting high shock resolution the scheme

should be reconstructed so that the scheme is FST/MXD behind the shock
and SLW in front of the shock.. The above constructed scheme is SLW for
the case of m = 0. For making the scheme being MXD the following operator
extrapolation is used
(2k+1,±) (2k+1,±) (2k,0)
h̄j+1/2 = (1 + σ)hj+1/2 − σhj+1/2 . (4)
For improving the shock resolution the scheme is further reconstructed with
group velocity control and weighting. The final (2k)th order accurate WGVC
scheme is expressed as
(2k) (2k) (2k) (2k) (2k,+) (2k,−)
Fj = Hj+1/2 − Hj−1/2 , Hj+1/2 = Hj+1/2 + Hj+1/2 (5)
(2k+1,±) 1 ± SS(j + 1/2) (2k,±) 1 ∓ SS(j + 1/2) (2k,0)

Hj+1/2 = h̃j+1/2 + h̃j+1/2 (6)
2 2
(2k,±) (2k−l) (2k,±) (2k−l) (l,±)

h̃j+1/2 = (1 − gj+1/2 )h̄j+1/2 + gj+1/2 h̄j+1/2 (7)
(2k,0) (2k−l) (2k,0) (2k−l) (l,0)
h̃j+1/2 = (1 − gj+1/2 )hj+1/2 + gj+1/2 hj+1/2
[SS(j) + SS(j + 1)]

SS(j + 1/2) = (8)
2
(2k)
SS(j) = sign(δx◦ fj · δx2 fj ). gj+1/2 = [σ0,j+1/2 ]2k
|fj+1 − fj |
σ0,j+1/2 = (9)
|fj+1 | + |fj | + ε
Equations (4), (6) and (7) correspond to the steps of extrapolation, group
velocity control and weighting, respectively. The case of k = 3 (6th order) is
used to simulate the R-M instability problems.
3 Simulation of R-M instability

Consider the R-M instability of a perturbed cylindrical material interface be-
tween a light medium inside and a heavy medium outside. A cylindrical shock
moves from the heavy medium towards the center. The initially perturbed
interface is expressed as
R◦ = 1 + acos(k1 ϕ) + bcos(k2 ϕ) (10)
where a 6= 0, b = 0 corresponds to the single mode case; and a 6= 0, b 6= 0
to the double mode case. The general features of the R-M instability are as
follows. As the cylindrical shock moves towards the center from the heavy
medium it collides with the material interface, after that it bifurcates into
a transmitted shock which moves towards the center of the cylinder, and a
reflected rarefaction wave which propagates outward from the interface. This
is the wave bifurcation stage. When the transmitted shock reaches the center
it reflects, and thus a reflected shock is formed. This reflected wave propagates
outward and interacts with the material interface again. This interaction is
known as reshock. After that a new wave bifurcation is produced. Because the
reflected shock from the center propagates in the light medium towards the
material interface, the reflected wave after collision with the material interface
becomes a new reflected shock. Finally the spike-bubble structure is formed
after multiple reshocks. In DNS we consider the mean radius of the initial
material interface r◦ as the characteristic length, the initial shock speed vs as
the characteristic velocity, and the dimensionless time is defined as t = vs t̄/r◦
The density and the temperature are normalized by the initial density and
temperature inside the material interface.
4 Effect of Atwood number
The effect of Atwood number on the characteristics of R-M instability is dis-

cussed on the basis of computed results for the cases of single mode interface
perturbation (b = 0) with the amplitude a = 0.033, the wave number k1 = 12
(see equation (10)), the Reynolds number Re = 50000., the shock Mach num-
ber Ms = 1.2 and the Atwood number A = 0.818, 0.961, 0.980 (corresponding
density ratio is 10, 50, 100 respectively). From numerical results it can be seen
that increase of the Atwood number accelerates the propagation speed of the
shocks inside the light medium, and increases the times of reshock process, but
restrains the pressure growth at the center. Fig.1 gives the density contours
for different Atwood numbers after four times reshock. It can be seen that
the width and the length of the spike are related with the Atwood number.
The spike becomes shorter and thinner with increasing Atwood number. Vari-
ation of Atwood number also changes the characteristics of the growth rate.
From spectral analysis it can be seen that increasing the Atwood number can
accelerates the excitation of high wave number modes, and enhances mixing
between the two different media.
a. A=0.818 b. A=0.961 c. A=0.981
Fig. 1. density contours for different Atwood numbers

a. Ms=1.2 b. Ms=5.0 c. Ms=10.0

Fig. 2. Variations of pressure at center with time for single mode with and without
perturbation
5 Effect of shock strength
Generally speaking, when the incident shock is stronger, the speed of the mov-
ing interface, the speed of the transmitted shock and the reflected rarefaction
wave will be faster. In this case the reshock exists in the early stage of strong
shock acceleration. The effect of shock strength is discussed on the basis of
computed results for the case of single mode with a = 0.033, k1 = 12, b = 0,
Re = 50000., Ms = 1.2, 5.0, 10, and Atwood number A = 0.818, 0.905. The
variations of pressure at the cylinder center with time are given in Fig.2 for
different Mach numbers. From the figures it can be seen that the pressure at
the center grows rapidly with increasing Mach number for both cases with
and without the interface perturbation. For example, the maximal pressure
at the center for the case of Ms = 10 is 6000 times larger than the pressure
at t = 0. For the case of Ms = 1.2 the maximal pressure at the center is
only 4 times larger. Also we see that the interface instability restrains the
pressure growth at the center in the later time and the compression degree of
the interior medium C = rmin /r0 increases with increasing Mach number. In
our case C = 0.68, 0.11 and 0.05 for Ms = 1.2, 5 and 10 respectively.
From computed results we can see that the variation of the growth rate
with time is much different for different Mach numbers. Increase of the Mach
number greatly speeds up the non-linear development of the interface insta-
bility and changes the spike-bubble structures.
Fig.3 shows the variation of the growth rate of perturbation with time
for different shock Mach numbers. The growth rate is defined as ∂(rmax −
rmin )/(2∂t) where rmin and rmax correspond to the maximal and minimal
radius of the material interface. It can be seen that the increase of the shock
Mach number changes the characteristics of the growth rate of perturbation.
The density and vorticity contours at different time in the evolution process
of R-M instability are given in Fig.4 and 5 for the cases of Ms = 5 and
Ms = 10 respectively. For clarity the central parts of the contours are amplified
(the radius normalized by the mean radius of the initial material interface is
equal to one). From these figures together with Fig.1 we can see that the spike-
a. Ms=1.2 b. Ms=5 c. Ms=10

Fig. 3. Variation of growth rate with time for different shock Mach numbers
bubble structures after multiple reshocks are much different for different shock
Mach numbers. It means that increasing the shock Mach number speeds up
the nonlinear development of R-M instability, and greatly changes the spike-
bubble structures. Fig.6 shows the vorticity contours at different times for the
case of shock Mach numbers Ms = 5. From the figure we can see pairs of
contra-rotating vortices. In the middle stage after several times collisions of
the shocks with the interface, unstable phenomenon occurs at the center.
From spectral analysis it can be seen that with increasing shock Mach num-
ber the Kelvin-Helmoholtz instability resulting from R-M instability develop-
ment speeds up the excitation of small structures with high wave numbers. In
the later stage of development wide range of spectra are found. Finally, the
merging of perturbation waves with different scales leads to breakdown of the
spike-bobble structures, and a central merging region is formed.
The above presented numerical results with high resolution further confirm
the effectiveness of the high order accurate scheme developed in this paper.
The effect of double mode competition on the R-M instability with the
basic wave number of initial perturbation k1 = 12 is also discussed. From the
computed results we see that when the difference between the basic mode and
the subharmonic is large (for example, k1 = 12, k2 = 60) the competition
is weak in the spike-bubble structures, and merging can not be seen in a
short period of time. When the difference between the basic mode and the
g p p p
a. t=0.54 b. t=0.66 c. t=0.78

Fig. 4. Density contours at different time for Ms = 5
Fig. 5. Density contours at different time for Ms = 10
a. t=0.66 b. t=0.78 c. t=1.09

Fig. 6. Vorticity contours at different time for Ms = 5
subharmonic is small (for example, k1 = 12, k2 = 24 or 36) the smaller

structures will be swallowed up by the larger one. In fact, for given basic
mode there exists a range of subharmonic modes in which the merging of
spike-bubble structure occurs easily. Increase of the Atwood number further
accelerates the excitation of small scale structures.
Conclusions
A high order accurate scheme with weighted group velocity control is devel-
oped to simulate the R-M instability problems with high shock Mach number
and high density ratio. The effects of Atwood number, shock strength and
mode competition on the characteristics of R-M instability are discussed. The
computed results verify the effectiveness of the developed schemes.
References
1. M. Brouillette, Annual Reviews, p445-468, 2002
2. N. J. Zabusky, A. D. Kolnikov, J. Fluid Mech. Vol. 475, p147-162, 2003
3. R. H. Cohen et al., Physics of Fluids, vol. 14, No.10, 2002
4. Q. Zhang and M. J. Granham, Physical Review Letters vol. 79, No.14,1997
5. Q. Zhang and M. J. Granham, Physics of Fluids, vol. 10, No. 4, 1998
6. D. Fu and Y. Ma, JCP 134:1-15, 1997
Part IX
Multiphysics and interdisciplinary coupled

flow problems
Boundary conditions by low-order modelling
Marcelo Buffoni12 , Haysam Telib1 , and Angelo Iollo2

1
DIASP - Politecnico di Torino, 10129 Torino, Italy.
2
MAB - Université Bordeaux 1 and INRIA Projet MC2, 33405 Talence, France
1 Introduction
In this contribution we are concerned with the coupling between a full order
simulation and a reduced order model. The idea is to reduce the extent of the
simulation domain and hence the computational costs.
In a broad sense, there exist many applications where far from the bound-
ary the solution is weakly dependent on the details of the boundary geometry.
In such regions we use a reduced order model based on proper orthogonal de-
composition (POD) [1] to solve the problem. This approach allows a represen-
tation of the solution by a small number of unknowns that are the coefficients
of an appropriate Galerkin expansion. Therefore away from a narrow region
close to the boundary of interest the number of unknowns to be solved for,
is drastically reduced. Like all other approaches based on POD, a solution
database is necessary to build the Galerkin modes, therefore this method will
be useful when many computations for relatively similar cases are to be per-
formed, like for example in shape optimization.
2 Approximation of the Steklov-Poincaré operator by

POD
In order to explain the method we take a particular case. Let us consider

the Laplace equation ∆u = 0 defined inside a square Ω = [0, 1] × [0, 1]. Let
d ∈ [0, 1], Ω1 = [0, d] × [0, 1], Ω2 = [d, 1] × [0, 1] and Γ = Ω 1 ∩ Ω 2 the interface
between the two sub-domains, see fig. 1(a). On the right (uR ) boundary as well
as on the upper (uU ) and lower (uD ) boundaries we have Dirichlet conditions.
We want to solve this problem for different values of the Dirichlet data on
the left (uL ) boundary. To that end we build an appropriate solution database
for several boundary conditions on the left side. In particular, the Dirichlet
data on the left boundary is g k , 1 ≤ k ≤ N .

748 Marcelo Buffoni, Haysam Telib, and Angelo Iollo
(a) Schur (b) Schwarz
Fig. 1. Problem set-ups

(k)
Let the functions uΩ be the discrete solutions of the boundary value prob-
lem posed in Ω for different k, and let ū2 be an armonic function, restricted
PN (k)
to Ω2 , defined as follows: ū2 = 1/N k=1 uΩ . We compute a Galerkin base
PN (k) (k) (k)
of the form ϕi = k=1 bik (u2 − ū2 ) where u2 is the restriction of uΩ to
Ω2 and the coefficients bik are found as in [2]. This base gives an optimal
(k)
representation of the original data set uΩ by construction.
PM
Let us define û2 = ū2 + i=1 ai ϕi , where M is much smaller than the
number of discretization points in Ω2 . For an arbitrary Dirichlet condition on
the left boundary of Ω, we want to determine the discrete solution as a usual
discrete approximation in Ω1 and by the above defined Galerkin representation
in Ω2 .
One simple way to implement this idea is to solve the problem by a
Dirichlet-Neumann iteration. To this end, we follow the steps below:
1. solve the problem in Ω1 by any discretization method (FD, FEM, etc.), im-
posing Neumann b.c. on Γ ;
2. on interface Γ project the trace of the above solution in the subspace spanned
by the trace of the POD modes ϕi ;
3. recover û2 as the prolongation inside Ω2 of û2 on Γ by using the POD modes;
4. determine ∂ û2 /∂n on Γ ;
5. goto 1. until convergence is attained.
This is just one possible solution algorithm, corresponding to a classical
domain decomposition method (Schur complement). Another approach con-
sists in solving the problem all at once, as detailed in the following. Let us
define A1 the discretized operator acting on u1 , the restriction of the unknowns
belonging to Ω1 ; AΓ the discretized operator acting on uΓ , the unknowns be-
longing to Γ and A2 the discretized operator acting on u2 , the restriction of
the unknowns belonging to Ω2 .
The discretized non-linear problem in Ω can be written
Boundary conditions by low-order modelling 749
    
A1 B1 0 u1 f1
 B1t AΓ B2t   uΓ  =  fΓ  (1)
0 B2 A2 u2 f2
where B1 and B2 are appropriate interface matrices and f1 , fΓ , f2 take into
account the boundary conditions.
From (1) we have
A1 u1 + B1 uΓ = f1
B1t u1 + (AΓ − B2t A−1
2 B )u t −1
2 Γ = fΓ − B2 A2 f2 (2)
The matrix AΓ −B2t A−1 2 B2 is the discrete counterpart of the Steklov-Poincaré
operator for Ω2 , see [3]. Consider now the second step of the solution algorithm
proposed above. Let a ∈ RM be a vector of components a1 . . . aM and c ∈ RM
a vector of components c1 . . . cM . Posing ϕi the trace of ϕi on Γ , we have
!
XM
a = arg min uΓ − ck ϕk (3)

c∈R M
k=1
where ||·|| is the norm induced by the canonical l2 scalar product, noted by
(·, ·).
PM
Solution of (3) reduces to the solution of the linear problem i=1 ai (ϕi , ϕj ) =
PM −1
(uΓ , ϕj ), 1 ≤ j ≤ M . Therefore ai = (uΓ , Pi ), where Pi = j=1 [(ϕi , ϕj )] ϕj
is a constant vector computed once for all from the POD modes.
At this point we approximate u2 with û2 and substitute in (1). Since
PM PM
B2 û2 = B2t ū2 + i=1 ai B2t ϕi , we have B2t û2 = B2t ū2 + i=1 B2t ϕi (uΓ , Pi ).
t
PM
Finally, letting Ŝ2 = i=1 B2t ϕi Pi we obtain the approximation of (2)
B1t u1 + (AΓ − Ŝ2 )uΓ = fΓ − B2t ū2 (4)
where B2t ū2 ≡ B2t A−1
2 f2 .
Matrix Ŝ2 is the approximation of the discrete
Steklov-Poincaré operator obtained by the POD expansion. Equations (2)
can of course be solved simultaneously by a standard linear solver.
3 Numerical experiments and discussion
The models range from Laplace equation with non-linear boundary conditions
to compressible Euler equations. Since this method can be of interest for
optimal design applications, where many different geometries must be tested
to improve performance, in some cases we have explored the idea of simulating
by usual discretization methods just the region where the geometry changes,
modelling the rest by POD.
3.1 Schur complement
A second order finite differences (FD) method coupled to a fix point iteration is
used to solve the Laplace equation inside the square domain shown in fig. 1(a).
The left Dirichlet boundary condition is varied to build the needed database.
In particular, uL = sin (kπy) + y, 1 ≤ k ≤ 49. The boundary conditions on
the other sides are, referred to fig. 1(a), the following: on uU : u = 1, on
∂u
uD : u = 0 and on uR : u4 − u40 + ∂n = 0.
The domain is splitted at d = 1/3. Then, the Galerkin base is generated
over Ω2 using the previously built database. In order to check the accuracy
of the method, a boundary condition which was not included in the database
used to build the POD modes is imposed on the left boundary: uL = y 2 , and
a second order FD method is used to solve the problem in Ω1 , by a fix point
iteration. We use 6 POD modes to recover û2 inside Ω2 . Figure 2 presents the
result of the test by means of the distribution of the relative error between the
FD solution over the entire domain and the approximate Schur complement
approach.
(a) (b)
Fig. 2. Distribution of the relative error between the solution obtained by the
present method and the solution obtained by a second order FD method on the
whole domain. Ω1 (a), Ω2 (b).
3.2 Schwarz method
In the following we considere a convergent-divergent nozzle-like domain Ω. As

before we divide Ω in two subdomains, Ω1 and Ω2 , in a way that there exist
an overlap region Ωov = Ω 1 ∩ Ω 2 shared by both subdomains (see fig. 1(b)).
We solve the Laplace equation with the boundary conditions detailed below
to generate a database of k solutions with 1 ≤ k ≤ 60, this time varying
the geometry of Ω1 . The solution is obtained by means of the finite elements
Boundary conditions by low-order modelling 751
method using P1 elements on a triangular non-structured mesh [4] and a fixed

point iteration. The boundary conditions imposed are: on uL : u = 13 · y, on
∂u
uU : u = 1, on uD : u = 0 and on uR : u4 − u40 + ∂n = 0.
For the geometry of the divergent part which was not included in the
database, we determine the solution following a similar approach to the de-
scribed above but employing this time the Schwarz method, see [3]. 4 POD
modes are used to recover û2 inside Ω2 . Figure 3 shows the results obtained
for this case, again in terms of the relative error between a numerical solution
(FEM P1) over the entire domain and the approximate Schwarz method.
(a) (b)
Fig. 3. Percentage error distribution over (a) Ω1 and (b) Ω2 . The reference solution
is obtained by FEM P1 elements on a triangular mesh.
3.3 Newton method
We solve the compressible Euler equations in a convergent-divergent nozzle

on a structured mesh. The λ-scheme [5] is used to solve the equations. Total
temperature, total pressure and the flow angle are imposed at the inlet; static
pressure at the exit and impermeabilty at the walls. The complete sytem
is solved by Newton iterations. The resulting linear problems are solved by
preconditioned GMRES iterations [6]. Using this code we calculate a database
of 100 solutions varying the geometry of the divergent part of the nozzle and
we deduce a low order model for the flow angle v/u in the convergent part.
In figure 4 we show results obtained by combining the low order model with
the original solver. The divergent part of the nozzle has a geometry which
was not included in the POD-database. The two subdomains are coupled by
the least squares approximation on the overlapping domain and the boundary
condition at the interface. Again we solve these two coupled problems by a
Newton method, all at once.
(a) (b)
Fig. 4. Percentage error distribution of v/u over (a) Ω1 and (b) Ω2 .
3.4 Discussion
We presented three possible implementations of a method to reduce the extent

of the computational domain in the numerical solution of partial differential
equations. The idea of using models that take into account different physical
phenomena in different subdomains is old. Here we revisited this approach
using a low order model. Since a solution database is necessary for building the
POD modes, the application of this method is limited to those problems where
large number of computations pertinent to similar configurations is necessary,
for example in shape optimisation. Indeed the results above show that this is
a viable approach to rigourously take into account boundary variations when
using a low order model for optimisation.
References
1. J. L. Lumley, “The structure of inhomogeneous turbulent flows,” In Atmospheric
Turbulence and Radio Wave Propagation, edited by A. M. Yaglom and V. L.
Tatarski, Moscow, pp. 166–178, 1967.
2. L. Sirovich, “Turbulence and the dynamics of coherent structures. Parts I,II and
III,” Quarterly of Applied Mathematics, vol. XLV, pp. 561–590, 1987.
3. A. Quarteroni and A. Valli, Domain Decomposition Methods for Partial Dieren-
tial Equations. Oxford Science Publications, 1999.
4. O. Pironneau, F. Hecht, and A. L. Hyaric, “Freefem++, a language implemen-
tation dedicated to the finite element method.” http://www.freefem.org.
5. L. Zannetti and B. Favini, “About the numerical modelling of multidimensional
unsteady compressible flow,” Computers and Fluids, vol. 17, pp. 289–299, 1989.
6. Y. Saad, “Sparsekit: a basic tool kit for sparse matrix computations.” http:
//www-users.cs.umn.edu/~saad/software/SPARSKIT/sparskit.html.
Multi Scale Numerical Simulation of Dispersed
Reacting Flow, with application to Chemical
Vapor Deposition of Alumina
Andrey A. Markov
Institute for Problems in Mechanics of Rusian Academy of Sciences, Moscow,

119526, Vernadskii ave. 101, RUSSIA, markov@ipmnet.ru
Summary. The micro equations are coupled with the macro balance equations for
mass, momentum, and energy and written down, using only the Eulerian approach.
The relationship between gas, surface, and solid phases is numerically simulated. The
examples of kinetics include homogeneous and heterogeneous competitive chemical
reactions and chemical condensation in a gas phase, as well as homogeneous reactions
relative to high-temperature kinetics of AlCl3 decomposition in the presence of
additives.
1 Introduction
Cooperative CVD-processes as a tool of new advanced nanocomposites pro-

duction are applied. The kinetics and mechanism of such fast running cooper-
ative CVD-processes at temperatures from 1000 to 3000 K, time duration of
about one second and more, the forming product structure and properties are
under investigation. Due to particle generation and growth, a macroscopic
spatial nonuniformity is developed around the particle or films of the con-
densed product. On the macroscopic scale heat- and mass transfer and kinet-
ics of the new phase formation are strongly influenced by the nonuniformity.
The examples of kinetics include homogeneous and heterogeneous competi-
tive chemical reactions and chemical condensation in a gas phase [1] as well
as high-temperature kinetics of AlCl3 decomposition in the presence of addi-
tives [2]. The results of computation for 45 elementary homogeneous chemical
reactions between 36 species are in agreement with [2]. Method presented is
the modification of [3], where some results of computation were performed for
simplified kinetics. The physical-chemical mathematical model of such pro-
cesses including the spontaneous nucleation and particle coagulation are in
progress.

754 Andrey A. Markov
2 Formulation of the problem

The process of chemical conversion in a flow of dusted gas mixture for coop-
erative chemical reactions in a cylindrical tube is simulated.
Let us consider an arbitrary set of L simultaneous chemical reactions in-
volving n distinct chemical species. The bulk and surface reactions are written
as follows
n vol n n n
X v,0 kif ,Qvol
i
X v,00 X 0 kif ,Qi X 00
νij Cj ←→ νij Cj ; νij Cj ←→ νij Cj ; i = 1, . . . , L
vol ,−Qvol
kib kib ,−Qi
j=1 i j=1 j=1 j=1
v,0 v,00 0 00
Here νij , νij and νij , νij are appropriate stoichiometric coefficients for
j’th specie in the i’th chemical process for bulk and surface reaction respec-
vol vol
tively. The values kif , kib are the constant rates for forward and backward
vol
bulk reactions and Q denote the thermal effect. The values kif , kib , Qi are
the constant rates and thermal effects for surface reactions. Let components
Cl , ; l = 1, ..., np change the phase from gas to solid due to chemical
condensation
kl,ph
Cl −→ Pj ; l = 1, ..., np ; np ≤ n
Ql,ph
Including the phase transition, the mass rate production M˙K of specie CK for
an arbitrary set of L simultaneous surface reaction steps is presented in [3].
Let Cj and Mj are the mass fraction and molar mass of j-th specie, ρ is
gas density, JpK is the rate of condensation of specie, M˙iK
vol is the mass flux for
bulk reaction, csat

Kb is concentration of saturated vapor of intermediate product
ph
CK over particle surface of radius b, the value of kK = kK,ph , K = 1, ..., np
ph
and kK = 0, K > np .
Using the balance equations of mass and energy, we write down the micro-
level equations for the spherical condensed particle of volume Vb and surface
Sb = 4πb2 in a gas volume V as follows.
1 dCks ρs V
= M˙K − MK JpK + βkc (Ck,ex − Cks ) (1)
Sb dt
Equation (1) represents the mass conservation of reagent and product. The
right hand side includes the mass fluxes due to chemical reactions and con-
densation at the particle surface as well as mass fluxes βkc (Ck,ex − Cks ) due
to advection and diffusion via mass transfer coefficients βkc .
L np
1 d (Cb Vb ρp Ts ) X X
= Q̇ + α(Tex − Ts ); Q̇ = Ṙi Qi + Jpl Ql,ph (2)
Sb dt i=1 l=1
Equation (2) demonstrates the temperature variation at the particle surface, Q̇

is the thermal source due to surface reaction. Term α(Tex − Ts ) represents the
Multi Scale Numerical Simulation of Dispersed Reacting Flow 755
advection and conduction via heat transfer coefficient α. Here Cb is specific

heat of particle at constant pressure, ρp is particle density. The saturated
concentration csat
Kb with respect to K-specie depends on critical radius bk,cr of
a condensed particle and on saturated vapor concentration at plane surface
csat
k,∞ . These values are found using thermodynamics consideration.
np
1 dρp Vb Vb X
= (1 − ) Ml Jpl (3)
Sb dt V
l=1
Equation (3) represents the mass conservation for a condensed particle that
density and volume equal ρp and Vb = 4πb3 /3 respectively. The subscripts
ex and s are referred to the values near the particle and at the particle sur-
face respectively
The heat transfer and mass transfer coefficients α ,and β ck depend on the
coordinates of a particle. These values are found resolving the flow problem
near the particle in the region b(t) < r < r∞ :
Dρck DρCp T
= ∇r ·Dk (∇r ck ) ; = ∇r · (λ∇r T )
dt dt
where Dk is the diffusivity.
We apply the boundary conditions in the form :
r = b(t) : Ck = Cks ; T = Ts ; r = r∞ : Ck = Ck,ex ; T = Tex
As the result of solution, we come to the coefficients:

Dk ∂Ck λ ∂T
βkc = , α=
(Ck,ex − Cks ) ∂r (Tex − Ts ) ∂r
We apply the above common form of kinetics to competitive chemical reactions

and chemical condensation [1,3]. The bulk reactions are written as follows
v
k1f ,Qv
1
v
k2f ,Qv
2
C3 + C4 v←→ v C1 ; C3 + C4 v←→ v C2 (4)
k1b ,−Q1 k2b ,−Q2
k1f ,Q1
1 kph 2 k2f ,Q2 kph
C3 + C4 ←→ C1 −→ P1 C3 + C4 ←→ C2 −→ P2 (5)
k1b ,−Q1 Qph k2b ,−Q2 Qph
1 2
For macro-level equations we use nondimensional values as follows:

R0 T00 2σ 0 M 0
T 0 = T00 T, b0 = b00 b, β = 0 , Bl,0 = 0 0 l0
E0 ρp E 0 b 0
Here, σ is the energy per unit square for particle surface creation, 0 refers
to dimensional values. The dimensionless similarity parameters that charac-
terize the nonequilibrium rate of chemical-process will be described by the
Fig. 1. AlCL3 thermal decomposition ALCL3 (t)/ALCL3 (0) at x = 0.5, y = 0.95

at T = 1325K, p = 0.1 atm for three initial compositions. Lines 1 refer to
molar fractions: ALCL3 = 0.02, CO2 = 0.02, AR = 0.96; lines 2 refer to mo-
lar fractions:ALCL3 = 0.02, CO2 = 0.02, H2 = 0.96; lines 3 refer to molar
fractions:ALCL3 = 0.001, CO2 = 0.006, HCL = 0.001, H2 = 0.992
Damköhler numbers that are the ratios of timescale of forward, backward

chemical reaction and condensation to the advection transport time scale re-
spectively:
0 0 0 0 0
τf v,i = t0f v,i /tM , τbv,i = t0bv,i /tM , τf s,i = t0f s,i /tM , τbs,i = t0bs,i /tM , τp,j = t0p,j /tM ,
0 0 0 0
The particle is characterized by nondimensional parameter zkb = l0 ρ0 / b0 ρp
0
where l0 is a characteristic macro scale. These scales as well as the Reynolds
and Peclet numbers are included in the coefficients of governing equations.
The basic set of the equations is the Navier-Stokes equations supplemented

by the terms resulting from microscopic analysis together with relaxation
equations for the chemical species (see [3] in detail). The governing equations
represent the mass conservation for gas species, particle number density and
size variation, the momentum and energy conservation for carrier gas flow,
using the approximation of dusted gas flow. The state equation for perfect gas
is used to complete the macro equation set. Initial and boundary conditions
are imposed at the entrance of tube and at the wall [3]. Here the wall is
assumed to be chemically neutral.
3 Computational setup and results
The equations are numerically solved using the implicit finite–difference ap-
proximations and the splitting of the differential operator as a sum of two
terms LD and LKI . The LD (f ) is the transfer via advection and diffusion
[3]. The high accurate time resolution for stiff chemical system is coupled
Multi Scale Numerical Simulation of Dispersed Reacting Flow 757
mesoscale step mesoscale step

the values at the particle surface the heat mass transfer coefficients
n+1/2
fs ⇒ α, βic
macro scale step macro scale step
Jihom
micro- fluxes Jihet averaging Jihet
macro scale kinetics macro scale transport problem
f n+(1/2) ⇒ f n+1
Table 1. The mesoscale –macro scale interaction
Fig. 2. Gas and surface phase values against time at x = 0.5, y = 0.95, T0 =
0.2, b0 = 10−6 , n0 = 5 · 108 a01 = 0.05, a02 = 0.45, c01 = 0.01, c02 = 0.49
with macro equation solver. At the current semi-integer time level we find
n+1/2
the value fs of surface phase of unknown vector f n+1 . The micro–macro
scale interaction can be sketched as shown in table 1.
High-Temperature AlCl3 thermal decomposition in the presence of addi-
tives is simulated. We consider the model of aluminium particle growth, using
the simplified model (4),(5) of surface reactions [1]. Let C1 = Al2 O3 , C2 =
(Al2 O3 )α , C3 = OAlO, C4 = AlO. Some computations are performed for to-
tal set of bulk reactions [2]. The results presented in Fig. 1 are in agreement
with experimental and theoretical ones [2].
The results of computation refer to particle number density n0 = 104 , n0 =
5 · 108 , zkb = 5, b0 = 10−3 , b0 = 10−6 . The reference temperature and molar
0
mass are T0 = 1100K and M00 = M10 , Reynolds and Peclet numbers are equal
to 10−1 .
The parameters that were not varied in simulation are listed below. The
particle parameters, activation energy, and thermal effect are as follows:
0
ρ0p /ρ0 = 10−3 , E1fv
= E1f s s
= 15, E1b s
= 19.2, E2f s
= 22.5, E2b = 28.8, E1ph =
35, E2ph = 50; KJ/mole
Let ξj be the ration of backward to forward heterogeneous reaction rate
0 0
ξj = kjb /kjf = τf s,i /τbs,i
Fig. 3. zkb = 5, n0 = 5 · 104 ; b0 = 10−3 ; T0 = 0.6;Particle components are B1 (y)

left and B2 (y) right in sections x = 0.4, x = 0.6at time t = 2.25, 7
Fig. 2 shows the time history at x = 0.5, Z = 0.95. The temperature, mass
fractions of components in gas and surface phase are presented for low thermal
effect of reactions. The parameters are as follows: T0 = 0.2, n0 = 5 · 108 , b0 =
10−6 ,zkb = 5 initial mass fractions a01 = 0.05, a02 = 0.45, c01 = 0.1, c02 =
−1 −1
0.49, and the rate parameters, ξ1 = 0.91, ξ2 = 0.91, τs,1 = 1012 , τs,2 =
12 −1 14 −1 12
5 · 10 , τp,1 = 10 , τp,2 = 10 . We see the difference between the gas and
surface values when temperature is increasing.
In the following example the initial data at t = 0 is imposed such a way
that the evaporation for B1 and condensation for B2 takes place. However, this
condition is not sufficient in total time interval as it is found in computation.
Particle components B1 (left) and B2 (right) are presented in Fig. 3 at time
t = 2.25, 7 in sections x = 0.4, x = 0.6. the parameters are as follows: b0 =
10−3 , n0 = 5 · 108 ,zkb = 5 for initial mass fractions a01 = 0.12, a02 = 0.4, c01 =
−1
0.08, c02 = 0.4 and the rate parameters, ξ1 = 0.995, ξ2 = 0.091, τs,1 =
12 −1 12 −1 14 −1 12
10 , τs,2 = 5 · 10 , τp,1 = 10 , τp,2 = 10 .
References
1. Grigor’ev, Yu. M., Doronin, S. I., and Filimonov, I. A. Kinetics of phase transition
in cooperative CVD processes. J. Chem. Phys.,18, No. 12, 25–30. 1999.
2. Catoire, L, and Swihart, M. T. High-Temperature Kinetics of AlCl3 Decomposi-
tion in the Presence of Additives for Chemical Vapor Deposition., J. Electrchem.
Society, 149, 261–267,2002.
3. Markov, A. A. Multi Scale Numerical Technique for Strongly Coupled Two-Phase
Flow. In: Proc. NURETH11 Symposium, Avignon, October, 2005.
Computational Analysis of Flow through a
Multiple Nozzle Driven Laser Cavity and
Diffuser
M. A. Sriram1 , N. K. S. Rajan2 , and P. S. Kulkarni3

1
Department of Aerospace Engineeering, Indian Institute of Science
msriram@cgpl.iisc.ernet.in
2
nksr@cgpl.iisc.ernet.in
3
psk@aero.iisc.ernet.in
Summary. Computations have been carried out for simulating supersonic flow
through a set of converging-diverging nozzles with their expanding jets forming a
laser cavity and flow patterns through diffusers, past the cavity. A thorough numeri-
cal investigation with 3-D RANS code is carried out to capture the flow distribution
which comprises of shock patterns and multiple supersonic jet interactions. The
analysis of pressure recovery characteristics during the flow through the diffusers is
an important parameter of the simulation and is critical for the performance of the
laser device. The results of the computation have shown a close agreement with the
experimentally measured parameters as well as other established results indicating
that the flow analysis done is found to be satisfactory. The model is extended for
study of different parameters influencing the performance of the system, generating
useful data for an optimised design. The work assumes importance as it validates
the design considerations by the parametric analysis of the geometrical and other
aerodynamic parameters governing the performance of the laser system.
1 Introduction
The configuration selected for the analysis is taken out from a problem re-
alised from the field of work that involves using aerodynamically controlled
laser cavity for generation of high powered laser beams. In this arrangement,
in order to produce an active lasing medium, rapid and deep cooling of a
mixture of gases from a combustion chamber (CO2 − N2 − H2 O) is achieved
by allowing the gases to expand in a set of convergent-divergent nozzles, and
the allowing the supersonic jets to drive the cavity used as the zone for the
laser activity. This cavity would have a distribution of CO2 in non-equilibrium
condition with many quantum levels, providing a favorable condition for the
lasing activity. The bank of nozzles followed with the cavity followed by a

760 M. A. Sriram, N. K. S. Rajan, and P. S. Kulkarni
diffuser section for stable operation makes the main parts of the unit, an
aerodynamically controlled laser cavity, as being considered in the present
configuration. Fig 1. shows a 3-D view of the configuration considered.
The simulation of flow in a Gas Dynamics Laser system is done using
an industrially standard CFD tool with the code solving for 3D-RANS. The
process of the modelling is taken up in steps of increasing complexity to
ensure the numerical stability and the computational consistency. The dif-
ferent stages have provided clarity on different parameters that include the
optimal grid distribution and the optimal initial flow profiles to obtain a
stable and consistent solution. The details of the analysis are as follows.
An approach with multiple geometrical blocks that are matching to the

geometry of the unit considered at different locations, depending on its com-
plexity, is used with their mating surfaces blended so as to enable generation
of a good computational mesh. The structured mesh is preferred since the
geometry having a large complexity and the flow analysis having a need to
capture supersonic, transonic and subsonic zones over the same computational
domain. Optimisation on the number of computational nodes is made by al-
lowing for cluster of grids in critical zones, at places expected to have larger
gradients in the flow and in places with more geometrical complexity. It is ob-
served that the entire configuration required about 2 million computational
nodes while critical sections of the analysis that include the nozzles, cavity
and diffuser needing about 1.5 million to have a consistent and validated re-
sults. On a four-node parallelised computer with each having a Pentium IV
processor and 2 GB of memory, a test-case for the flow analysis of the full
configuration, it needs to have about 24 to 30 hours of computer time.
The computational Flow Analysis made during the preliminary stage of the
study includes flow analysis through a single nozzle followed by flow analysis
through a set of four nozzles and further analysis of a set of nozzles with a
set of two diffuser vanes. In its final stage the computations were taken up to
capture the flow through multiple nozzles and the diffusers. The method of
analysis, the results obtained at different stages been presented and discussed
during the sections to follow.
2 Method of Analysis
The simulation of flow is considering a system consisting of nozzles, laser cavity

and the diffusers as its main elements and an optionally extended acoustic
filter at the tail end of the diffuser as a special case, has been presented. The
complete geometry has been divided into a number of blocks. Hexahedron
cells are generated using the mesh generator in each block. Mesh clustering
has also been done to obtain the desired resolution. Grid refinement studies
have been carried out to obtain the optimum mesh size. Fig 2. gives a sample
CFD of Multiple Nozzle Driven Laser Cavity and Diffusser 761
view of close-up of the mesh generated near in cavity at the upstream of the
diffuser.
The objective of the study is to capture the flow patterns that provide
the details of the shock pattern, the supersonic jet interactions and multiple
reflected shock interactions along with the associated pressure and temper-
ature distributions in the laser cavity and the diffuser. The captured details
of the flow analysis include the pressure recovery patterns and shock interac-
tions in the diffuser, the pressure recovery being an important parameter for
the stable performance of the laser cavity. A study of grid independence was
obtained by a series of test runs with increasing computational nodes. A com-
prehensive study of different turbulent models was carried out to decide the
suitable scheme of turbulence. It was found that k − ωturbulence model was
most appropriate to capture all the features of the flow where as the laminar
model and k − ε turbulence and other models were inadequate to capture the
essential features of the flow.
3 Governing equations and the boundary conditions
For the present flow analysis the 3D RANS equations have been considered.
The Reynolds-Averaged Navier-Stokes Equations are solved for steady, single-
phase, compressible and viscous flow.
Mass conservation:
ρ1 U1 = ρ2 U2 = ṁ (1)
Momentum conservation:
p1 + ρ1 U1 2 = p2 + ρ2 U2 2 (2)
Energy conservation:
1 1
Cp T1 + U1 2 + q = Cp T2 + U2 2 (3)
2 2
The following boundary conditions are implemented:
1) No slip, impermeable and adiabatic wall conditions
2)The inlet pressure and temperature of the fluid entering the gas dynamic
Laser system is set as an assigned boundary value
3)The outlet boundary condition is to set with matching with the Ambient
Upwind scheme is used in obtaining the solution. The solution is obtained by
implicit successive iterative scheme. The solution is taken to have converged
with the diminishing RMS residuals reducing to less than of 10-4 on the
variables used.

Fig. 3 shows a result used for the validation of the computational scheme.
This indicates a consistency of the Mach number and associated pressure
distribution of a single nozzle expansion as referred to the calculated values
of isentropic case. Fig. 4 shows verification with the consistency of the results
between the computed and the analytically obtained shock angle ahead of the
diffuser plate. The capture of multiple reflecting shocks in the diffuser zone
is verification that the scheme of computation and the mesh generated are
producing the acceptable results. These cases indicate the verification of the
method adopted with the matching of the reference data to the computed
one, enabling the further analysis to be progressed satisfactorily.
Fig. 5 shows the Mach contour plot in a cross sectional plane of the main
geometry considered. The plot shows the complex shock patterns captured
well. In the upstream of the diffuser (basically used as the lasing cavity),
the diamond shock patterns are seen because of the multiple super sonic jet
interactions. However, these interactions remain as weak shocks thereby not
significantly disturbing the pressure profile for the desired lasing activity. The
static pressures and temperature predicted are found to be closely matching
with the experimental test-runs, providing a check on the adequacy of cap-
turing the physical behaviour of the configuration considered for the analysis.
Fig. 6 is a contour plot of static pressure and it shows up well the diamond
shock behaviour, in the cavity (though the pressure variations from this is
small, of the order of .005 atm.) the expected discontinuities are interpreted as
a consequence of the interactions of multiple supersonic jets from the nozzles.
It can also be seen in Fig.7 that the pressure recovery patterns along the
diffuser, providing verification for the design characteristics of the diffuser to
have met with the requirements. The pressure recovers near to 1 atm. at the
downstream of the diffuser little before the diffuser and, that is open to the
ambient, another important criteria for the design verification to ensure a
stable functioning of the aerodynamic laser cavity.
Figs. 8 and 9 show a set of flow fields captured in the computational anal-
ysis that allows making few important observations. One is that the addition
of the acoustic control device, added does not significantly the pressure re-
covery at the diffuser. It is found that though the flow in the acoustic filter is
fully subsonic and strongly recirculating, the pressure drop is not significant
to provide an adverse back pressure and thereby validating the design part
of this to be used at the tail end safely for reduction of acoustic noise of the
exhaust.
Figs. 10 is a sample plot of the diminishing errors during the computation,
with the RMS residuals falling steadily and meeting the convergence criteria.
CFD of Multiple Nozzle Driven Laser Cavity and Diffusser 763
The results of the computation have captured all the essential features of
the fluid dynamics of the aerodynamically driven laser cavity and the main
thermodynamic parameters are found to have a close match to the experi-
mentally measured parameters indicating the flow modelling is satisfactory.
Optimal mesh generation and grid resolution has been obtained during the
study. The results of the analysis have effectively helped in characterising the
design details and the other aerodynamic parameters that would be useful for
further activities in the area.
Fig 1.Geometrical Configuration Considered Fig 2.Sample Mesh Generated
Fig 4.Diamond Shock Patterns Fig 3.Mach number at the Throat

Fig 5. Mach Contour Plot Fig 6. Pressure Contour Plot
Fig 7. Pressure Recovery Plots Fig 8. Pressure Contours With

Acoustic Control Device
Fig 9. Vortex Flow in Fig 10. Residual fall in a

Acoustic Control Device test-run of the CFD analysis
MHD Analysis of Force Acting on Dipole
Magnetic Field in Magnetized Plasma Flow
Hiroyuki Nishida1 , Hiroyuki Ogawa2 , and Yoshifumi Inatani3

1
University of Tokyo, 3-1-1, Yoshinodai, Sagamihara, Kanagawa 229-8510, Japan.
nishida@isas.jaxa.jp
2
Japan Aerospace Exploration Agency, 3-1-1, Yoshinodai, Sagamihara, Kanagawa
229-8510, Japan. ogawa@isas.jaxa.jp
3
Japan Aerospace Exploration Agency, 3-1-1, Yoshinodai, Sagamihara, Kanagawa
229-8510, Japan. inatani@isas.jaxa.jp
1 Introduction
Magnetic Sail is a propulsion system that utilizes the solar wind in the in-
terplanetary space [1]. This propulsion system creates a large magnetic field
around the spacecraft by a super-conducting coil and generates thrust through
the interaction between the solar wind and the magnetic field. The interaction
produces the “magnetosphere” around the spacecraft and the magnetosphere
works as a sail for catching the solar wind. This propulsion system is suitable
for deep space exploration missions because this propulsion system can gen-
erate large thrust and does not need any propellant, and so many researches
about the propulsion system have been conducted recently.
The solar wind jets out of the Sun in all directions. The source of the solar
wind is the Sun’s hot corona. The temperature of the corona is so high that
the Sun’s gravity cannot hold on to it. The solar wind is full-ionized plasma
flow at speeds of about 400km/sec in the interplanetary space, and interacts
with the magnetic field of the Earth and then confines the magnetic field of
the Earth to the magnetosphere.
It is generally known that the solar wind is magnetized because of the
magnetic field of the Sun. The magnetic field in the solar wind is called the
Interplanetary Magnetic Field (IMF ; as shown in Fig.1). The strength and
direction of the IMF is changing in response to activities of the Sun. The IMF
affects the magnetosphere of the Earth and causes magnetic storms and the
aurora.
Similary, it is to be expected that the thrust of Magnetic Sail is affected by
the IMF, because Magnetic Sail generates thrust by its magnetosphere. It is
therefore important to understand how the IMF affects thrust and to estimate
amount of change in thrust. Although many researches about the interaction

766 Hiroyuki Nishida, Hiroyuki Ogawa, and Yoshifumi Inatani
Fig. 1. Solar Wind and Interplanetary Magnetic Field
between the geomagnetic field and the IMF have been conducted, there is no
reserch on the relation between thrust and the IMF.
In this study, the interaction between the dipole magnetic field by a coil
and magnetized plasma flow is simulated using magnetohydrodynamics, and
the force acting on the coil is examined.
2 Dungey Model
Dungey model is a widely known theoretical model explaining the configura-

tion of the geomagnetosphere interacted with the IMF [2]. Dungey proposed
that open magnetic field region was formed by the “magnetic field reconnec-
tion” in the magnetosphere and the solar wind entered into the magnetosphere
through the open magnetic field. Recently, some evidences of the magnetic
field reconnection were observed by some artificial satellites, and so Dungey
model gives good explanations to the configuration of the geomagnetic field.
Because flow field and magnetic field around Magnetic Sail are similar to those
around the Earth, Dungey model is helpful for studying Magnetic Sail. In this
section, Dungey model is outlined briefly.
2.1 Southern Interplanetary Magnetic Field
When a dipole magnetic field is confined within plasma flow, magnetic neutral
points exists in the magnetic field. The magnetic field reconnection occurs
at the magnetic neutral point. In the case that the IMF is in the southern
direction, a magnetic field neutral line surrounds the Earth on the equatorial
plane (as shown in Fig. 2). First, a magnetic field reconnection occurs at
a magnetic neutral point on the windward side of the Earth and the IMF is
MHD Analysis of Dipole Field in Magnetized Plasma Flow 767
connected with the geomagnetic field by this magnetic field reconnection. The
connected IMF is transported to the leeward of the Earth by the solar wind
and reconnects at a magnetic neutral point in the geomagnetic tail. Newly-
formed closed magnetic field goes back to the Earth and open magnetic field,
which does not connect the geomagnetic field, is transported to the leeward.
Fig. 2. Dungey model : Southern interplanetary magnetic field.
2.2 Northern Interplanetary Magnetic Field
When the IMF is in the northern direction, magnetic neutral points are on the
leeward of the Earth (as shown in Fig. 3). Magnetic reconnections occur at
these magnetic neutral points, however the magentic field reconnections are
weaker than the reconnections in the case of southern IMF. Open magentic
fields formed by the reconnections are transported to the leeward, and closed
magnetic field region is formed around the Earth.
Fig. 3. Dungey model : Northern interplanetary magnetic field.

3 Governing Equations and Numerical Method

We adopt magnetohydrodynamic equations to simulate the interaction be-
tween the magnetic field and the magnetized plasma flow. Finite electric con-
ductivity has an important role in magnetic field reconnection, and we intro-
duce a model of electric conductivity in numerical simulation: In this study,
the model of anomalous resistivity is adopted [3].
The magnetohydrodynamics equations are solved by a high-resolution
TVD Lax-Friedrich scheme. MUSCLE interpolation is used to achieve third-
order space accuracy and MINMOD limitter is used to ensure numerical sta-
bility.
4 Computational Conditions
The interaction between the dipole magnetic field and the magnetized plasma
flow is simulated in two-dimensional space. The simulation box is defined to
be a rectangular coordinate system as shown in Fig. 4. Two rigid cables with
electric currents flowing in the direction perpendicular to the X-Y plane are lo-
cated, and these currents generate a dipole magnetic field in two-dimensional
space. In this simulation, the dipole magnetic moment vector is in the Y direc-
tion. Magnetized plasma flows into the box at the beginning of the simulation.
The magnetic field of the plasma flow is in the Y or minus Y direction. The
plasma flow’s magnetic field in the Y direction corresponds to the southern
IMF in the case of the Earth, and the magnetic field in the minus Y direction
corresponds to the northward IMF.
Computational parameters are in Table. 1.
Simulation area
Y
Cables with currents
Dipole magnetic field

Magnetic field of plasma flow
Magnetized
a n plasma flow
Fig. 4. Simulation model.

MHD Analysis of Dipole Field in Magnetized Plasma Flow 769
Table 1. Computational parameters.
Plasma flow velocity 400 [km/sec]

Plasma flow temperature 20 [eV]
Plasma flow density 8.4 × 10−21 [kg/m3 ]
Magnetic flux density of plasma flow 0.5 [nT]
Magnetic flux density at the center of cables 200 [nT]
5 Simulation Results
5.1 Southern Magnetic Field of the Plasma Flow
The simulation result of flow field and magnetic field is shown in Fig. 5 in the
case of the plasma flow with the southern magnetic field (in the Y direction). It
is observed in Fig. 5 that a bow shock is formed on the windward of the dipole
magnetic field and the magnetosphere is formed around the dipole. The solar
wind flows around the magentosphere without entering into it. The flow field
varies cyclically and At the reconnection point in the magnetotail, a heated
plasma ball, called plasmoid, is ejected to the downstream at regular intervals.
This simulation result is in agreement with the Dungey model denoted in
Subsect. 2.1. Figure 6 shows the time change of the force acting on the
dipole magnetic field (in other words, two cables), here the force and time
are normalized by the characterize force F0 and the characterize time t0 ,
respectively. The force is estimated by calculating the Lorentz force between
the induced current in the magnetosphere and the currents in cables [4]. In
Fig.3, the dashed line shows the force acting on the dipole magnetic field in the
non-magnetized plasma flow for comparison. In the case of non-magnetized
plasma flow, the flow field and the force acting on the dipole are in steady-state
after enough time passing. The force in the magnetized plasma flow changes
in time periodically and is larger than the force in the non-magnetized plasma
flow.
Reconnection Plasmoid
Fig. 5. Flow field in the case of the southern

Fig. 6. Force acting on the
magnetic field.
dipole magnetic field.
5.2 Northern Magnetic Field of the Plasma Flow
The simulation result of flow field and magnetic field is shown in Fig. 7 in the
case that the plasma flow has the northern magnetic field (in the minus Y
direction). In this case, significant changes in flow field caused by the recon-
nection does not occur, and these simulation results agree with the Dungey
model denoted in Subsect. 2.2. Figure 8 shows the time change of the force
acting on the dipole magnetic field. The force in the plasma flow with northern
magnetic field is smaller than the force in the plasma flow with no magnetic
field and changes in time, but the amplitude of the change is smaller than
that in the case of plasma flow with southern magnetic field.
Reconnection
Fig. 7. Flow field in the case of the northern

Fig. 8. Force acting on the
magnetic field.
dipole magnetic field.
Conclusion
In this study, the interaction between the dipole magnetic field and the mag-
netized plasma flow was simulated in order to analyze the force acting on
the dipole magnetic field. The flow field and the force change in time. These
changes are caused by the magnetic field reconnection.
References
1. R.Zubrin, D.Andrews, “Magnetic sails and interplanetary space”, Journal of
Spacecraft and Rockets, Vol.28, No.2 (1991).
2. T.Saito, Space Science Rev. 10, 319 (1969).
3. J.Raeder, .et.al., “Global simulation of the Geospace Environment Modeling
substorm challenge event”, Journal of Geophysical Research, Vol.106, No.A1,
381-395 (2001).
4. H.Nishida, .et.al., “Two-dimensional Magnetohydrodynamic Simulation of a
Magnetic Sail”, Journal of Spacecraft and Rockets, 43, 667-672 (2006).
Performance of High Order Filter Methods for
a Richtmyer-Meshkov Instability
B. Sjögreen1 and H.C.Yee2

1
NADA, KTH, Stockholm, Sweden bjorns@nada.kth.se
2
NASA Ames Research Center, Moffett Field, CA, USA yee@nas.nasa.gov
Summary. Sixth-order compact and non-compact filter schemes that were de-
signed for multiscale Navier-Stokes, and ideal and non-ideal magnetohydrodynam-
ics (MHD) systems are employed to simulate a 2-D Rightmyer-Meshkov instability
(RMI). The suppression of this RMI in the presence of a magnetic field was investi-
gated by Samtaney (2003) and Wheatley et al. (2005). Numerical results illustrated
here exhibit behavior similar to the work of Samtaney. Due to the different amounts
and different types of numerical dissipations contained in each scheme, the struc-
tures and the growth of eddies for this chaotic-like inviscid gas dynamics RMI case
are highly grid size and scheme dependent, even with many levels of refinement.
1 Numerical Method and Objective
Methods commonly used for shock/turbulence interactions relying on switch-

ing between spectral or high order compact schemes and shock-capturing
schemes are not practical for multiscale shock interactions. Frequent switching
between these two types of schemes can create severe numerical instability.
Our highly parallelizable class of high order filter schemes [7, 5, 8, 9, 10, 12, 13]
does not rely on switching between schemes to avoid the related numerical in-
stability. They have built-in flow sensors to control the amount and types
of numerical dissipation only where needed, leaving the rest of the flow re-
gion free of numerical dissipation. Instead of solely relying on very high order
high-resolution shock-capturing methods for accuracy, the filter schemes take
advantage of the effectiveness of the nonlinear dissipation contained in good
shock-capturing schemes and standard linear filters (and/or high order linear
dissipation) as stabilizing mechanisms at locations where needed. The method
consists of two steps, a full time step of spatially high order non-dissipative
(or very low dissipative) base scheme and an adaptive multistep filter con-
sisting of the products of wavelet based flow sensors and linear and nonlinear
numerical dissipations to filter the solution. The adaptive numerical dissipa-
tion control idea is very general and can be used in conjunction with spectral,

772 B. Sjögreen and H.C.Yee
Fig. 1. Problem definition
spectral element, finite element, discontinuous Galerkin, finite volume and fi-
nite difference spatial base schemes. The type of shock-capturing scheme used
as nonlinear dissipation is very general and can be any dissipative portion of
a high resolution TVD, MUSCL, ENO, or WENO shock-capturing method.
By design, the flow sensors, spatial base schemes and linear and nonlinear dis-
sipation models are stand alone modules, and a whole class of low dissipative
high order schemes can be derived at ease. The objective of this work is to
illustrate the performance of our sixth-order low dissipation filter schemes for
a 2-D inviscid Richtmyer-Meshkov instability (RMI) problem.
2 RMI Test Problem and Numerical Results
RMI occurs when an incident shock accelerates an interface between two fluids
of different densities. This interfacial instability was theoretically predicted by
Richtmyer [3] and experimentally observed by Meshkov [2]. For the present
study, the RMI problem investigated by Samtaney [4] and Wheatley et al.
[6] as indicated in Fig. 1 has been chosen. The mathematical models are
the 2-D Euler gas dynamics equations and the ideal MHD equations. The
computational domain is −2 < x < 6 and 0 < y < 1. A planar shock at
x = −0.2 is moving (left to right) toward the density interface with an incline
angle of θ with the lower end initialized at x = 0. The density ratio across the
interface is denoted by η, and the nondimensional strength of the magnetic
field β = 2p0 /B02 , where the initial pressure in the preshocked regions is p0 = 1,
High Order Filter Methods for RM Instability 773
Fig. 2. Comparison between Euler gas dynamics and MHD for the 6th-order com-
pact spatial base scheme (left) and the 6th-order central (non-compact) spatial base
scheme (right) using a (801 × 101) grid at t = 3.33. MHD solutions shown are mirror
images of the original computations.
and B0 is the initial magnetic field. The initial magnetic field is uniform in the
(x, y) plane and perpendicular to the incident shock front. Numerical results
shown below are for M = 2, θ = 45o , η = 3, β −1 = 0 (Euler gas dynamics) and
β −1 = 0.5 (magnetic field present). The computation stops at an evolution
time t = 3.33. For this set of parameters and all studied numerical schemes,
instability occurs near t = 1.8 for the gas dynamics case but not for the MHD
case for the entire time evolution. Our numerical results exhibit behavior
similar to the study of Samtaney.
Computations by the sixth-order centered spatial compact base scheme
with the compact linear filter [1], in conjunction with a second step nonlin-
ear WENO5 filter (WENOfi) denoted by Comp66+Compfi+WENOfi using a
801 × 101 grid is shown in Fig. 2 (left) for the inviscid gas dynamics (GD)
and the ideal MHD equations. Here WENOfi means the dissipative portion
of the fifth-order WENO scheme (WENO5) in conjunction with our wavelet
flow sensor as the nonlinear filter [10, 12, 13]. The same computation using
the sixth-order central spatial (non-compact) base scheme in conjunction with
WENOfi denoted by CEN66+WENOfi is shown in Fig. 2 (right). The clas-
sical fourth-order Runge-Kutta method is used for the sixth-order compact
and non-compact filter schemes. For this low resolution grid, the accuracy
between the two filter methods is similar. See Fig. 3 for the grid refinement
study below. Computations using Comp66+WENOfi (i.e., without the linear
compact filter step) or Comp66+Compfii (i.e., without the nonlinear WENOfi
filter step) indicate spurious oscillations around shock regions.
The present study arrives at the same conclusion drawn in [7, 12] on the
behavior of compact spatial schemes for problems containing multiscale shock
interaction. High order compact schemes are methods of choice for many in-
compressible and low speed turbulent/acoustic flows due to their advantage
of requiring very low number of grid points per wavelength. In the presence
of multiscale shock interactions, however, this desired property of high order
compact base schemes seems to have diminished in both the gas dynamic and
MHD test cases that we have studied (compared with the same order of ac-
curacy of non-compact central base schemes). Also the compact spatial base
scheme requires more CPU time per time step and it is less compatible with
parallel computations than the central spatial base scheme. Consequently, the
compact spatial base scheme requires added CPU time in a parallel computer
framework.
Figure 3 shows the inviscid gas dynamics comparison between CEN66 +
WENOfi and a second-order MUSCL for four grids (801 × 101, 1601 × 201,
3201×401, 6401×801). Here, computations using a second-order MUSCL and
a second-order Runge-Kutta method are denoted by MUSCL. Not shown are
the same computation using CEN66 as the base scheme in conjunction with
the dissipative portion of the Harten-Yee scheme and the wavelet flow sensor
as the nonlinear filter (CEN66+HYfi). For similar resolution, MUSCL requires
nearly 3 times finer grid size per spatial direction than CEN66+WENOfi and
CEN66+HYfi. The eddy structures are different among the three filter meth-
ods and they are very different from the Samtaney adaptive mesh refinement
(AMR) simulation with an equivalent uniform grid of 16384 × 2048. Note that
except for WENOfi, the van Leer version of the van Albada limiter is used.
For the second-order MUSCL scheme, the limiter is applied to the primitive
variables. All methods employed the Roe’s approximate Riemann solver for
the gas dynamics case and the Gallice approximate Riemann solver for the
MHD case using our method of solving the conservative MHD system [9].
Fig. 3. Grid refinement study of the second-order MUSCL (left) and

CEN66+WENOfi (right) at t = 3.33 using (801 × 101), (1601 × 201), (3201 × 401)
and (6401 × 801) grids.
High Order Filter Methods for RM Instability 775
The present study can serve as a good example of failure of grid refinement
for unsteady chaotic-like inviscid flow. As the grid is refined (in conjunction
with different amounts and types of numerical dissipations contained in each
scheme), smaller and smaller eddies are formed which combine to affect global
flow through the energy cascade effect. Future work, including viscous effects
and extensive comparisons is in progress [14].
The efficiency, accuracy and flexibility of the present class of low dissipative
high order filter schemes rest on the fact that the multistep filter can be ap-
plied after a full time step of the Runge-Kutta method, and that a wavelet
flow sensor is employed to control the amount of numerical dissipation where
needed. The major CPU time intensive part of the shock-capturing computa-
tion is the nonlinear filter. In fact, the filter procedure requires slightly more
CPU time than the Harten-Yee and MUSCL schemes. This is due to the fact
that all sixth-order (or any order) filter schemes require only one Riemann
solve per time step per direction (independent of the time discretizations of
the base scheme step) as opposed to two Riemann solves per time step per
direction by the MUSCL and Harten-Yee schemes using a second-order Runge-
Kutta method. WENO5-RK4 requires at least twice the CPU time of all other
methods since four Riemann solves per time step per direction are required by
WENO5-RK4. RK4 stands for the classical 4th-order Runge-Kutta temporal
discretization and WENO5 is the 5th-order WENO spatial scheme.
References
1. D.V. Gaitonde and M.R. Visbal, Further Development of a Navier-Stokes Solu-
tion Procedure Based on Higher-Order Formulas, AIAA Paper 99-0557, 1999.
2. Y.Y. Meshkov, Instability of a Shock Wave Accelerated Interface Between Two
Gases, NASA Tech. Trans. NASA TT F-13074, (1970).
3. R.D. Richtmyer, Taylor Instability in Shock Acceleration of Compressible Fluids,
Commun. Pure Appl. Math. 13, (1960), 297.
4. R. Samtaney, Suppression of the Richtmyer-Meshkov Instability in the Presence
of a Magnetic Field, Physics of Fluids, 15 (2003), L53-L56.
5. B. Sjögreen and H. C. Yee, Multiresolution Wavelet Based Adaptive Numer-
ical Dissipation Control for Shock-Turbulence Computation, RIACS Technical
Report TR01.01, NASA Ames research center (Oct 2000); also, J. Scient. Com-
puting, 20, (2004), 211-255.
6. V. Wheatley, D.I. Pullin and R. Samtaney Regular Shock Refraction at an oblique
Planar Density Interface in Magnetohydrodynamics, J. Fluid Mech. 522, (2005),
179-214.
7. H.C. Yee, N.D. Sandham, N.D., and M.J. Djomehri, Low Dissipative High Order
Shock-Capturing Methods Using Characteristic-Based Filters, J. Comput. Phys.,
150 (1999) 199-238.
8. H.C.Yee and B.Sjögreen, Designing Adaptive Low Dissipative High Order

Schemes for Long-Time Integrations for Long-Time Integrations, Turbulent
Flow Computation, (Eds. D. Drikakis & B. Geurts), Kluwer Academic Pub-
lisher (2002); also RIACS Technical Report TR01-28, Dec. 2001.
9. H.C. Yee and B. Sjögreen, Efficient Low Dissipative High Order Scheme for
Multiscale MHD Flows, II: Minimization of Div(B) Numerical Error, RIACS
Technical Report TR03.10, July, 2003, NASA Ames Research Center; also, J.
Scient. Computing, (2005) DOI: 10.1007/s10915-005-9004-5.
10. H.C. Yee and B. Sjögreen, Nonlinear Filtering and Limiting in High Order Meth-
ods for Ideal and Non-Ideal MHD, Proceedings of the ICOSAHOM, July 21-25,
2004, Brown University, RI; to appear, J. Scientific Computing, 2006.
11. H.C. Yee and B. Sjögreen, Performance of Adaptive Numerical Dissipation Con-
trol in High Order Methods for the Resistive MHD Equations, Proceedings of the
HYP2004, Sept. 13-17, 2004, Osaka, Japan.
12. H.C. Yee and B. Sjögreen, Nonlinear Filtering in Compact High Order Schemes,
Proceedings of the 19th ICNSP & 7th APPTC Conference, Nara, Japan, July
11-15, 2005.
13. H.C. Yee and B. Sjögreen, Development of Low Dissipative High Order Fil-
ter Schemes for Multiscale Navier-Stokes/MHD Systems, Proceedings of the
CalSpace/UCR ASTRONUM Conference, Palm Springs, CA, March 27-30, 2006.
Expanded version submitted to J. Comput. Physics.
14. H.C. Yee and B. Sjögreen, Simulation of Richtmyer-Meshkov Instability by Sixth-
Order Filter Methods, Proceedings of the 17th International Shock Interaction
Symposium, University of Rome, Rome, Italy, Sept. 4-8, 2006.
Unsteady Flow Simulation of High Speed
Turbopumps
Cetin C. Kiris1 , Dochan Kwak2 , William Chan3 , and Jeffrey A. Housman4

1
NASA Ames Research Center, Moffett Field,CA, Cetin.C.Kiris@nasa.gov
2
NASA Ames Research Center, Moffett Field,CA, dkwak@mail.arc .nasa.gov
3
NASA Ames Research Center, Moffett Field,CA, wchan@nas.nasa.gov
4
U.C. Davis, Davis,CA, jhousman@math.ucdavis.edu
Summary. High fidelity computations were carried out to analyze the orbiter LH2
feedline flowliner. Computations were performed on the Columbia platform which is
a 10,240-processor supercluster consisting of 20 Altix nodes with 512 processors each.
Various computational models were used to characterize the unsteady flow features
in the turbopump, including the orbiter Low-Pressure-Fuel-Turbopump (LPFTP)
inducer. Unsteady flow originating from the orbiter LPFTP inducer is one of the
major contributors to the high frequency cyclic loading that results in high cycle
fatigue damage to the gimbal flowliners just upstream of the LPFTP. An incom-
pressible Navier-Stokes flow solver INS3D, based on the artificial compressibility
method, was used to compute the flow of liquid hydrogen in each test article.
1 Introduction
Computation of high-speed hydrodynamics requires high-fidelity simulation

to resolve flow features involving transient flow, cavitation, tip vortex and
multiple scales of unsteady fluctuations. One example of this type of flow
in aerospace is related to liquid-fueled rocket turbopumps. Rocket turbop-
umps operate under severe conditions at very high rotational speeds typi-
cally at thousands of rpm. For example, the Shuttle orbiter low-pressure-fuel-
turbopump creates transient flow features associated with reverse flows, tip
clearance effects, secondary flows, vortex shedding, junction flows, and cavi-
tation effects. Flow unsteadiness originating from the orbiter Low-Pressure-
Fuel-Turbopump (LPFTP) inducer is one of the major contributors to the
high frequency cyclic loading that results in high cycle fatigue damage to the
flow liners just upstream of the LPFTP. The reverse flow generated at the tip
of the inducer blades travels upstream and interacts with the bellows cavity.
A simulation procedure appropriate for this-type of high-speed hydrodynamic
problem requires a method for quantifying multi-scale and multi-phase flow as
well as an efficient high-end computing strategy. The current paper presents

778 Kiris, Kwak, Chan and Housman
a high-fidelity computational procedure for unsteady hydrodynamic problems

using a high-speed liquid-fueled rocket turbopump.
2 Solution Procedure and Computational Models
The incompressible Navier-Stokes flow solver, INS3D [2], [3], [4], based on the
artificial compressibility method [1] was used to compute the flow of liquid
hydrogen in several test geometries including the feed line and an inducer
for pumping. To obtain time-accurate solutions, the equations are iterated
to convergence in pseudo-time for each physical time step until the diver-
gence of the velocity field has been reduced below a specified tolerance value.
The total number of sub-iterations required varies depending on the problem,
the time step size and the artificial compressibility parameter used. Typically
the number ranges from 10 to 30 sub-iterations. The computational require-
ment for the model problem is large, and thus two distinct parallel processing
paradigms have been implemented. These include the Multi-Level Parallel
(MLP) [5] and the MPI/OpenMP [6] hybrid parallel programming models.
Both models contain coarse and fine grain parallelism. Coarse grain parallelism
is achieved through a UNIX fork in MLP and through explicit message pass-
ing in the MPI/OpenMP hybrid code. Fine grain parallelism is achieved us-
ing OpenMP compiler directives in both the MLP and MPI/OpenMP hybrid
codes. Computations were performed to compare the scalability between the
MLP and MPI/OpenMP hybrid versions of the INS3D code on the Columbia
system using the BX2b processors.
All computations included tip leakage effects with a radial tip clearance of
0.006 inches, a pump operating condition of a mass flow rate of 154.7 lbm/sec,
and a rotational speed of 15,761 rpm. A simplified computational model in-
cludes the LPFTP inducer with 4 long and 4 short blades, and a straight
duct, which extends 4 duct diameters upstream of the inducer. The objective
of studying this model is to compare unsteady pressure values against exper-
imental data. To resolve the complex geometry in relative motion, an overset
grid approach is employed. The geometrically complex body is decomposed
into a number of simple grid components. Connectivity between neighboring
grids is established by interpolation at each grid’s outer boundaries. Addition
of new components to the system and simulation of arbitrary relative mo-
tion between multiple bodies are achieved by establishing new connectivity
without disturbing the existing grids. This computational grid has 57 overset
zones with 26.1 million grid points. An extended computational grid system is
based on the first model with the addition of the flowliner geometry. The grid
system includes upstream and downstream slots, the overhang area between
liners and the bellows cavity. This model is very similar to the ground test
article. It consists of 264 overlapped grids with 65.9 million grid points. De-
tails of the grid system are shown in figures 1 and 2. The flowliner component
consists of an axisymmetric chamber around the external wall of the pipe,
Unsteady Flow Simulation of High Speed Turbopumps 779
and two rows of slots in the streamwise direction. Each slot is a rectangular
shaped hole with rounded corners. On the outside wall of the chamber are the
bellows, which are shaped like 10 periods of a sine wave. The bellows cavity is
connected to the duct via the overhang area and the slots. Two-dimensional
overset grids are first created for the bellows, side walls and the overhang area
of the bellows cavity. These are then revolved 360 degrees to form the volume
grids. Each slot consists of a body-conforming grid and a warped Cartesian
core grid in the middle of the hole. The flowliner component alone contains
212 grids and 41 million points.
Fig. 1. Surface grids for LPFTP inducer and the Liquid LH2 flowliner.
3 Computed Example
An initial condition of flow at rest and no inducer rotation is used to start
the computations. Then, the inducer is rotated at full speed as a start-up
procedure. Mass flow is specified at the inflow and characteristic boundary
conditions are used at the outflow. Simulations for 14 inducer rotations were
competed for model I, and 12 inducer rotations were completed for model II.
The time history of non-dimensional pressure difference from INS3D calcula-
tions (model I) at a location where experimental measurements are taken is
plotted in figure 3(a). Even though computed results have not fully converged
to periodicity in time and may still show the evidence of start-up transients,
Fig. 2. Details of the flowliner overset grid system.
the dominant 4N unsteadiness at a fixed location is shown in figure 3(a).

Figure 3(b) shows maximum and minimum non-dimensional pressure values
recorded from the experimental data. Comparisons between CFD results and
hot fire test data also show good correlation in the non-dimensional pressure
amplitudes.
(a) (b)
Fig. 3. Time history of non-dimensional pressure during one inducer rotation (model
I, 14th inducer rotation), and Min/Max values of non-dimensional pressure in hot
fire test.
The region of reverse flow extends far enough upstream to interfere with
both flowliners in the gimbal joint. Figure 4 shows the extent of the backflow
at an instantaneous time. Positive axial velocity values are colored in red and
negative axial velocity values are colored in blue. Axial velocity values are
non-dimensionalized by tip velocity. It is clearly seen that there is a strong
interference between the backflow in the duct and the flow in the bellows
cavity. The backflow velocity magnitude reaches 15-20 % of the tip velocity
magnitude in the overhang area between the liners. It should be noted that
this interaction is unsteady and backflow travels in the circumferential direc-
tion as well. Due to strong interactions in the overhang area, flow is excited
in the bellows area that results in time-dependent recirculation regions. This
Unsteady Flow Simulation of High Speed Turbopumps 781
Fig. 4. Axial velocity contours at an instantaneous time in a vertical cut plane.
observation can be seen in figure 5, where velocity vectors are plotted in the
region near the bellows cavity and overhang area. As shown in figure 5, strong
jet flow, with velocities of about 10-15 % of the inducer tip speed, penetrates
directly into the bellows cavity resulting in strong unsteady recirculation re-
gions in the cavity. The time-dependent interaction between the duct and the
bellows cavity can be one of the major contributors for high cycle loading.
Figure 5 also shows that modeling the gap in the overhang area between flow-
liners is very important. Jet-like flow in the overhang area pushes the fluid in
the bellows cavity toward the duct through the slots. Without proper model-
ing of this detailed geometry, one cannot obtain fine scale flow unsteadiness
on the liner. This transient phenomenon creates unsteady pressure loading
spectrum on the flowliner surfaces. Backflow also causes pre-swirl to occur in
the flow approaching the inducer.
Fig. 5. Velocity vectors colored by total velocity magnitude.

Conclusion
High fidelity turbopump and flowliner calculations were performed on the
Columbia supercluster by using both MLP and MPI/OpenMP hybrid imple-
mentations of the INS3D code. CFD results confirmed that there is a backflow
caused by the LPFTP inducer. The region of reverse flow extends far enough
upstream to interfere with both flowliners in the gimbal joint. Computed
results for the straight duct hot fire test configuration have been verified by
correlation with unsteady pressure measurements. CFD results confirmed that
there is a strong unsteady interaction between the backflow regions caused by
the LPFTP inducer and secondary flow regions in the bellows cavity through
the flowliner slots. It is observed that there are more significant unsteady
flow interactions through the downstream slots than those observed in the
upstream slots. Inducer rotations are simulated in order to understand the
root cause of the flowliner crack problem.
Acknowledgments. The authors are grateful to Alex Te for his help in CAD
work, Tim Sandstrom and David Ellsworth for their help in the visualization
work, and to NASA-NESC Flowliner ITA team members for helpful discus-
sions.
References
1. Chorin, A., J., “A Numerical Method for Solving Incompressible Viscous Flow
Problems,” Journal of Computational Physics,’ Vol. 2, pp. 12-26, 1967
2. Kiris, C., and Kwak, D., “Parallel Unsteady Turbopump Simulations for
Reusable Launch vehicle,” Frontiers of Computational Fluid Dynamics,’
Caughey, D.A. and Hafez, M., ed, World Scientific, 2002.
3. Kiris, C., Kwak, D., and Rogers, S., “Incompressible Navier-Stokes Solvers in
Primitive Variables and Their Applications to Steady and Unsteady Flow Simu-
lations,” Numerical Simulations of Incompressible Flows,’ Hafez, M., ed, World
Scientific, 2003.
4. Rogers, S. E., Kwak, D. and Kiris, C., “Numerical Solution of the Incompress-
ible Navier-Stokes Equations for Steady and Time-Dependent Problems,” AIAA
Journal,’ Vol. 29, No. 4, pp. 603-610, 1991.
5. Taft, J. R., “Achieving 60 Gflop/s on the Production CFD Code OVERFLOW-
MLP,” Parallel Computing,’ 27(4):521-536,2001.
6. H. Jin, M. Frumkin and J. Yan, “Automatic Generation of OpenMP Directives
and Its Application to Computational Fluid Dynamics Codes,” in the Proceeding
of the Third International Symposium on High Performance Computing,’ Tokyo,
Japan, Oct. 16-18, 2000.
Use of the Gaussian Moment Closure for the
Modelling of Continuum and Micron-Scale
Flows with Moving Boundaries
J. G. McDonald, J. S. Sachdev, and C. P. T. Groth
University of Toronto Institute for Aerospace Studies, 4925 Dufferin Street,

Toronto, Ontario, Canada, M3H 5T6, mcdonald@utias.utoronto.ca
1 Introduction
Non-equilibrium micron-scale flows, such as those encountered in the complex

micron-sized conduits of micro-electromechanical systems and flows associated
with chemical-vapour deposition, are difficult to solve using existing mathe-
matical models and numerical methods. In most cases, these flows are in the
subsonic to low-supersonic regimes and, due to their micro-geometries, have
low Reynolds numbers and remain laminar. Flow Knudsen numbers, Kn, in
the range 0.01 < Kn < 10 may be encountered, even for pressures above one
atmosphere, and, as a result, non-continuum and thermal non-equilibrium
effects can significantly influence momentum and heat transfer phenomena
in typical micro-channel flows [5, 8]. The situation is further complicated by
the fact that in many cases, micron-scale flows can involve complex evolving
boundaries which require special gridding techniques.
One approaches that is showing considerable promise for the treatment
of micron-scale flows is moment closures [10]. Moment closures provide an
extended set of hyperbolic partial differential equations (PDEs) describing
the transport of macroscopic fluid properties. In general, the solution of these
PDEs require considerably less effort than obtaining solutions using a particle
simulation method. Furthermore, the treatment of non-equilibrium flows with
a purely hyperbolic model can have additional advantages from a computa-
tional perspective. The hyperbolic moment equations involve only first-order
derivatives and can be readily solved using high-order Godunov-type finite-
volume schemes coupled with adaptive mesh refinement (AMR) techniques
[4, 1, 16]. Schemes of this type are robust, minimize numerical discretization
errors, provide accurate resolution of discontinuities, and permit systematic
application of physically realistic boundary conditions. When coupled with
AMR, they permit treatment of complex and evolving flow geometries and
the resolution of highly disparate length scales while, at the same time, op-
timizing usage of computational resources. They also have narrow stencils,

784 J. G. McDonald, J. S. Sachdev, and C. P. T. Groth
making them suitable for implementation on massively parallel computers

[2, 16]. Note that higher-than-first-order derivatives can be problematic when
using adaptive mesh refinement or meshes with embedded boundaries as irreg-
ularities in the grid can make accurate evaluation of these derivatives difficult.
This situation is avoided with moment closure methods.
2 Gaussian Moment Closure

Levermore [10] has proposed a new hierarchy of non-perturbative moment
closures with several desirable mathematical properties. These methods are
based on approximate solutions to the Boltzmann equation of kinetic theory
and provide a hyperbolic mathematical description of non-equilibrium flows.
The lowest-order closure, the 10-moment or Gaussian closure is considered
here. It consists of a set of macroscopic transport equations which, in addition
to equations for the gas density, ρ, and momentum, ρui , contain equations for a
symmetric non-equilibrium pressure tensor, Pij . An extension to the standard
Gaussian closure for monatomic gases, proposed by Hittinger [9], allows for a
treatment of the molecular rotational energy, Erot , present in diatomic gases.
This includes a two-time-scale relaxation-time approximation to the collision
term. Previous numerical studies indicate that this closure holds considerable
promise for describing non-equilibrium transport, at least for subsonic and
transonic flows in the slip and transition regimes [11].
For two space dimensions, the Gaussian closure for a diatomic gas can be
written in weak conservation form as
∂U ∂Fx ∂Fy
+ + = S, (1)
∂t ∂x ∂y
where U = [ρ, ρu, ρv, ρu2 + Pxx , ρuv + Pxy , ρv 2 + Pyy , Pzz , Erot ]T is the
vector of conserved variables, Fx and Fy are x- and y-direction components
of the flux dyad F, and S is the source vector.
3 Parallel AMR Solution Scheme
A Godunov-type, finite-volume scheme has been developed for the solution of

the two-dimensional weak conservation form of the Gaussian moment equa-
tions as given in Eq. 1, on multi-block quadrilateral meshes. The scheme al-
lows for solution-directed block-based AMR and an efficient and highly scal-
able parallel implementation has been achieved via domain decomposition.
An arbitrary Lagrangian Eulerian (ALE) treatment is used to describe situ-
ations with moving boundaries. The ordinary differential equation resulting
from spacial discretization procedure that governs the evolution of the average
solution in a computational cell of the multi-block mesh, Ū, is given by
dŪ 1 X Ū dA
=− [Fk − wk Uk ] · n̂k ∆`k − + S, (2)
dt A A dt
k
Use of the Gaussian Moment Closure with Moving Boundaries 785
where A is the cell area, w, n̂k and ∆`k are the velocity of, unit normal
to, and length of the k th cell face, while Fk and Uk are the flux dyad and
solution state at the quadrature point on the k th face. Second-order spacial
accuracy is achieved via limited linear reconstruction. Flux evaluations are
carried out using Roe’s approximate Riemann solver [14]. Mesh refinement is
carried out according to physics-based refinement criteria as well as to resolve
embedded/moving boundaries. Refer to the recent paper by McDonald and
Groth [11] for further details of the parallel AMR scheme.
4 Mesh Adjustment Algorithm

Implementation of the preceding finite-volume scheme with the mesh adjust-
ment scheme proposed recently by Sachdev and Groth [15] is now consid-
ered. The mesh adjustment scheme provides an automated treatment for fixed
and moving, non-grid-aligned boundaries embedded in a body-fitted, multi-
block mesh. Similar in nature to the Cartesian-cut-cell methods developed by
Bayyuk et al. [3] and Murman et al. [12], this scheme allows for the nodes of
an underlying body-fitted mesh to be adjusted so as to coincide with the em-
bedded boundary. By making only local alterations to the grid, this scheme
enables the solution of unsteady flows involving moving boundaries or for
steady flow problems involving stationary boundaries that are not necessarily
aligned with the mesh, while preserving the structured nature of the blocks
and avoiding the creation of small cut cells that are often generated by tradi-
tional cut-cell approaches. In addition, the mesh adjustment algorithm is fully
compatible with block-based AMR and parallel implementation via domain
decomposition used in the finite-volume solution scheme.
Mesh adjustment is carried out in several steps. Firstly, a pre-mesh-
adjustment flagging is used in order to determine which cells may require
adjustment. The first mesh-adjustment step involves identifying sharp cor-
ners on the interfaces. The cells flagged for adjustment which contain each
sharp corner are identified and their nearest nodes moved to the corners. For
the remaining cells tagged for adjustment, the nodes which are closest to the
boundary are moved to the point of intersection between the interface and
the mesh lines, this can be seen in Figure 1(b). This movement will leave
cells that are bisected diagonally by the boundary. A secondary adjustment
step is used to move the nearest node of bisected cells so that it also lies on
the boundary. This step will produce triangular cells, as seen in Figure 1(c);
these cells are simply treated as degenerate quadrilaterals with two coinci-
dent nodes. The final step in the mesh-adjustment algorithm involves use of
a ray-tracing technique to determine which of the cells are within the flow
domain and which are outside the flow domain. The resulting mesh remains
structured and does not have neighbouring cells of radically different size. The
present implementation also allows for moving embedded boundaries whose
motion can be prescribed either explicitly or through a level-set method[13].
(a) (b) (c) (d)

Fig. 1. Mesh adjustment algorithm: (a) Initial mesh and embedded boundary (thick
line), (b) result of primary adjustment, (c) result of secondary adjustment (dashed
lines indicate inactive cells), and (d) example of (i, j)-indexing on an adjusted mesh.
5 Numerical Results
Several flow problems are now considered. The first such problem is subsonic
boundary-layer flow over a flat plate. The goal in this case is to demonstrate
that the embedded boundary treatment and hyperbolic nature of the gov-
erning equations will yield smooth predictions of the frictional forces acting
on the plate; traditional cut-cell-type approaches combined with the Navier-
Stokes equations have been shown to produce large oscillations in viscous drag
predictions [6]. Next, subsonic flow past a circular cylinder is considered. This
case is considered to illustrate the scheme’s applicability to transition-regime
flows. Previous investigation using body-fitted meshes has shown good agree-
ment with experimental results [11] and it will be shown that equally good
results can be achieved with embedded boundary treatment. Finally, a channel
flow with complex moving geometry is considered to illustrate the potential
of the proposed approach for tackling such problems.
Subsonic boundary-layer flow past a flat plate is considered. For the case
of interest, the free-stream Mach number is 0.2 and Reynolds number per unit
length is 2000. For this situation, the Knudsen number was 1.5×10−4 , which
indicates that the flow is in the continuum regime. Two computational meshes
are considered: one aligned with, or at 0◦ to, the plate and a second mesh at
30◦ to the plate. Both meshes initially consisted of one 16 × 16 Cartesian
square block. This block was then subjected to six or eight mesh refinements
such that at each level of refinement, any block crossed by the boundary was
refined. Sections of the resulting meshes are shown in Figure 2.
Numerical predictions of the friction coefficient, Cf , are shown in Figure 3.
In the figure, comparisons are made to the classical boundary layer results of
Blasius. It can be seen that there is good agreement between the computed
results and Blasius’ solution. There are no oscillations present in the predicted
skin friction coefficients, even when the flat plate intersects the grid at an
angle. Furthermore, on the more refined mesh with eight levels of refinement,
the numerical solutions are nearly indistinguishable from the Blasius solution.
As a second case, subsonic flow past an circular cylinder is considered.
Experimental measurements of drag on the cylinder have been obtained by
Use of the Gaussian Moment Closure with Moving Boundaries 787
Fig. 2. (left) Section of 48128-cell mesh with flat plate embedded at 0◦ . (right) Sec-
tion of 65024-cell mesh with flat plate embedded at 30◦ .
Fig. 3. Coefficient of friction calculated for a Cartesian grid with an embedded flat
plate at 0◦ and 30◦ to the grid as compared to Blasius solution.
Coudeville et al. [7]. The Gaussian closure has previously been applied to this
case using body fitted meshes [11] for several speed ratios, S, and a wide
range of Knudsen numbers in the continuum and transition regimes. These
predictions are shown in Figure 4, where numerical results are compared to
experimental data of Coudeville and an analytical expression of Paterson [11].
In order to demonstrate that the present embedded mesh treatment can
recover these previous results with virtually equal accuracy, values for the
coefficient of drag were computed for varying Knudsen numbers for flow with
a speed ratio of 0.107. Predicted results obtained using the adjusted mesh
scheme are compared in Figure 4 with the Gaussian-closure solutions obtained
using a body fitted mesh, as well as to the experimental results. Clearly, the
agreement between body-fitted and non-aligned-mesh results is excellent.
Coirier explored Mach 0.1 flow through a branched channel containing
fourteen pin cooling fins [6]. A similar geometry has been considered here for
both continuum and non-equilibrium flow, except two rows of the pins are now
assumed to oscillate with a prescribed motion. The results of the continuum
situation can be seen in Figure 5.
Fig. 4. (left) Coefficients of drag computed using the Gaussian closure with a body-
fitted mesh compared with experimental data and an approximate solution due to
Paterson. (right) Coefficients of drag computed using the Gaussian closure with a
body-fitted mesh and a Cartesian mesh with an embedded boundary at a speed
ratio of 0.107 are compared with experimental data.
Fig. 5. Computational grid and x-direction velocity contours for flow through a
branched channel with moving pins.
Conclusions
The use of the Gaussian moment closures with embedded-mesh treatment for
continuum- and transition-regime flows has been demonstrated. The proposed
algorithm allows for accurate treatment of micron-scale flows with embedded
and possible moving boundaries, not aligned with the mesh.
References
1. M. J. Aftosmis, M. J. Berger, and J. E. Melton. AIAA J., 36(6):952–960, 1998.
2. M. J. Aftosmis, M. J. Berger, S. M. Murman. Paper 2004-1232, AIAA, 2004.
3. S. A. Bayyuk, K. G. Powell, and B. van Leer. Paper 93–3391, AIAA, 1993.
4. M. J. Berger and P. Colella. J. Comput. Phys., 82:67–84, 1989.
5. A. Beskok, G. E. Karniadakis. J. Thermophys. Heat Transfer, 8(4):647, 1994.
6. W. J. Coirier. PhD thesis, University of Michigan, 1994.
7. H. Coudeville, P. Trepaud, E.A. Brun. Proceedings of the Fourth International
Symposium on Rarefied Gas Dynamics, New York, 1965. Academic Press.
8. J. C. Harley, Y. Huang, H. B. Bau, J. N. Zemel. J. Fluid Mech., 284:257, 1995.
9. J. A. Hittinger. PhD thesis, University of Michigan, 2000.
10. C. D. Levermore. J. Stat. Phys., 83:1021–1065, 1996.
11. J. McDonald and C.P.T. Groth. Paper 2005-5035, AIAA, June 2005.
12. S. M. Murman, M. J. Aftosmis, M. J. Berger. Paper 03–1119, AIAA, 2003.
13. S. Osher and J. A. Sethian. J. Comput. Phys., 79:12–49, 1988.
14. P. L. Roe. J. Comput. Phys., 43:357–372, 1981.
15. J. S. Sachdev and C. P. T. Groth. Proceedings of the 3rd International Confer-
ence on Computational Fluid Dynamics, 109–115, 2004.
16. J. S. Sachdev, C. P. T. Groth, and J. J. Gottlieb. Int. J. CFD, 19(2):157, 2005.
Development of “MATIS-SC” for High Speed
Steam Flow with Non-equilibrium
Condensation
Ryo Morita1 and Fumio Inada2

1
Central Research Institute of Electric Power Industry (CRIEPI), 2-11-1, Iwado
Kita, Komae-shi, Tokyo, 201-8511, Japan ryo@criepi.denken.or.jp
2
Central Research Institute of Electric Power Industry (CRIEPI), 2-11-1, Iwado
Kita, Komae-shi, Tokyo, 201-8511, Japan inada@criepi.denken.or.jp
1 Introduction
FIV(Flow Induced Vibration) is the main cause of troubles at Japanese

power plants. The clarification and the removal of the cause of FIV become
main subjects for the rationalization of maintenance and management in the
plants. Though vibration, noise and erosion caused by local high speed steam
flow may occur around a steam control valve, steam turbine and other plant
structures in the steam piping like orifice, it is difficult to measure the lo-
cal state quantities (pressure, temperature and so on) of the complex 3D
supersonic flow and understand the flow characteristics in detail only from
experiments. Therefore, it is useful to combine experiments and CFD (Com-
putational Fluid Dynamics) calculations for the problems, but there are few
CFD codes that can be applied to a wide range of steam conditions and cal-
culate actual steam condensation.
In this paper, we develop the new CFD code “MATIS-SC” that can calcu-
late the state quantities accurately, be applicable to the wide range of steam
conditions and be available to non-equilibrium condensation.
2 Development of “MATIS-SC”
2.1 Calculation Model
There are two main types of steam flow calculation models. One uses ap-
proximate equation of state for steam to calculate state quantities [1, 2, 3]
(named “equation of state model” in this paper) and the other uses
steam table [4, 5, 6, 7] (named “steam table model” in this paper). These
models have both advantages and disadvantages. The equation of state model

790 Ryo Morita and Fumio Inada
can calculate non-equilibrium condensation by using condensation model (e.g.

classical condensation theory model). But an accuracy of the state quantities
calculation is not high-order because of its approximation. On the other hand,
in the steam table model, the state quantities are derived from the state ta-
ble constructed by IAPWS [8] and have high-order accuracy. But this model
can’t consider the non-equilibrium condensation because the state quantities
provided by the state table are the value of the equilibrium state.
Then, we develop a new model (named “mixed model” in this paper)
in combination with these two models that can be applied to a wide range
of steam conditions (Fig.1) including non-equilibrium condensation and can
calculate state quantities accurately (O(10−7 )).
Fig. 1. Applicable Range of MATIS-SC (Colored region) and Steam Condition in

Steam Turbine
2.2 Features of “mixed model”
As mentioned above, the mixed model we adopt is the combined model of

the equation of state model and the steam table model.
For state quantities calculation except wetness, we use the steam table
model by Senoo [6]. In this model, state quantities are derived from the state
table that is used density (ρ) and internal energy (ε) as independent table.
For wetness calculation, we use classical condensation theory model that
can consider the non-equilibrium condensation. In this model, wetness and
number density of droplet are calculated by differential equations in parallel
with N-S equations [2]. To give the state quantities in the non-equilibrium wet
steam condition, we prepare 2 state tables. One is the table for the normal
equilibrium condition (state table), and the other is the one for the super-
saturation condition (supersaturation state table). In the non-equilibrium
Development of “MATIS-SC” 791
condition, if the calculated wetness is β and equilibrium wetness derived from

steam table is βeq , the state quantities (X) is calculated with state quantities
derived from state and supersaturation state table (Xeq and Xne ) as follows;
β βeq − β
X = Xeq × + Xne × , X = p, T, etc. (1)
βeq βeq
Features of mixed model and other models are shown in Table 1.
Table 1. Comparison of steam calculation models
Model Steam Table Equation of Mixed Model

Model State Model
Non-equilobrium ×
Condideration
State Quantity Calc. IAPWS- Approximate IAPWS-

equation of
(Excluding Wetness) IF97 state IF97
Based on Based on
Calculation of Wetness IAPWS- classical classical
IF97 concideration consideration
theory theory
Precisions of State 4
Quantities Calculation
2.3 Development of MATIS-SC Code
We already have built CFD code “MATIS-C” for compressible fluid and
confirmed in previous researches [9, 10]. “MATIS-C” is 3 dimensional FDM
based code and 2nd order spatial and temporal precision. We incorporate
mixed model into “MATIS-C”, and develop “MATIS-SC” for high speed steam
with non-equilibrium condensation.
The feature of “MATIS-SC” is shown in Table 2. 2nd order TVD scheme
is adopted to avoid the numerical oscillations, and 2nd order spatial precision
and full implicit 2nd order temporal precision with newton iteration. LU-SGS
algorithm is adopted for matrix inversion.
3 Confirmation of MATIS-SC Code
A steam flow in a laval nozzle (low-pressure condition, experiments by

Binnie et al. [13]) is calculated for a confirmation of “MATIS-SC”. The com-
putational grid of the nozzle is shown in fig.2 and the distribution of wetness,
pressure, and temperature in a flow section are shown in fig.3. As conden-

sation shockwave by non-equilibrium condensation can be seen in wetness
distribution, it is found that “MATIS-SC” can calculate non-equilibrium con-
densation and discontinuity change can be seen in pressure and temperature
distribution.
Table 2. Features of “MATIS-SC”

scheme 3 dimensional FDM
Fluid Compressible Steam with non-equilibrium
condensation
Government /3D Navier-Stokes Equations
Equations / Equation of Wetness
/Equation of Number Density of Droplet
State Quantities Calculated from IAPWS-IF97 by density and internal
(except Wetness) energy
Discretization Conv.:Harten-Yee Symmetric TVD(2nd order) [11]
nd
Visc. : Central(2 order)
nd
Time Marching 2 order backward + Newton Iteration
Matric Inversion:LU-SGS[12]
Fig. 2. Computational grid of Laval nozzle for benchmark test

Development of “MATIS-SC” 793
Fig. 4 shows the comparison of pressure distributions on the center of

the sidewall. It is found that the results of “MATIS-SC” agree well with
experiments quantitatively including the position of condensation shock (the
position of discontinuity change).
Fig. 3. Distribution of wetness, pressure and temperature
Fig. 4. Comparison of wall pressure distribution

Conclusions
For accurate calculation of a high-speed steam flow with non-equilibrium
condensation in wide range, we make mixed model to combine the advantages
of the existing models and develop “MATIS-SC”. A steam flow in a laval
nozzle is calculated for the benchmark test, and the reproducibility of non-
equilibrium condensation and the validation of the code are confirmed.
References
1. Senoo, S., et al., Proc. of ASME FEDSM, 31191 (2002)
2. Ishizaka, K., et al., Proc. of 6th ISCFD, 1(1995), 479-484
3. Gunter H. Schnerr et al., Proc. of 5th ISCFD, 3 (1993), 87-92
4. Grossman, B. and Walters, R. W., AIAA Journal, 27-5 (1989), 524-531
5. Meng-Sing Liou, Bram van Leer and Jian-Shun Shuen, J. of Computational
Physics, 87(1990), 1-24
6. Senoo, S., et al., Trans. JSME Series B (in Japanese), 671-68 (2002), 2006-2013
7. Gunter H. Schnerr and Ulrich Dohrmann, AIAA Journal, 28-7 (1990), 1187-
1193
8. Release on the IAPWS Industrial Formulation 1997 for the Thermodynamic
Properties of Water and Steam, IAPWS, (1997)
9. Morita, R. et al., 2004 ASME HT/FED Conf., 56017 (2004)
10. Morita, R. et al., Trans. JSME Series B (in Japanese), 696-70 (2004), 1984-1991
11. Yee, H., NASA TM 89464, (1987)
12. Yoon, S. and Jameson, A., AIAA Journal, 26-9 (1988), 1025-1026
13. Binnie, A.M., Green, M.A. et al., Proc. Roy. Soc., A, 181, plate 3, pp. 134-154,
(1943)
Numerical Method for Near-critical Fluids of
Arbitrary Material
Satoru Yamamoto and Atsushi Ito
Dept. of Computer and Mathematical Sciences, Tohoku University,

Sendai 980-8579, Japan
yamamoto@caero.mech.tohoku.ac.jp
1 Introduction
A computational code for simulating flows of an arbitrary substance at an

arbitrary condition is being developed by our research group. In this code,
the preconditioned flux-vector splitting(PFVS) scheme fully coupled with the
database of thermal properties, PROPATH [1], is implemented. The PFVS
scheme was presented in the 2nd ICCFD conference at Sydney [2] for simulat-
ing condensate flows. This method is based on the preconditioning method de-
veloped by Weiss and Smith [3]. Also in the last ICCFD at Toronto [4] , we ap-
plied our method to near-critical fluids introducing the Peng-Robinson(P-R)
equation of state(EOS) [5]. Then, near-critical carbon-dioxide(CO2 ) flows in a
micro-channel were calculated and some anomalous properties never found in
fluids at atmospheric temperature and pressure could be captured. The main
thermal properties such as density, temperature, and speed of sound could
be calculated by this approach. However, other properties such as thermal
conductivity and viscosity should be evaluated by any other mathematical
models.
The present numerical method is based on the PFVS coupled with the
PROPATH [1] which is a program package for thermal properties of substances
permitting bulk pressure and temperature in the wide range of them. This
program has been developed by Kyushu University. Especially P-PROPATH
which is one of sub-packages is employed in this study. All of the functions for
thermal properties programmed in this sub-package can be completely linked
with the present our method as a library file, *.lib. The present method can
calculate supercritical fluids of an arbitrary substance, for examples, CO2 , N2 ,
H2 O, CH4 , and so on. This code can apply more to flows at atmospheric or
cryogenic conditions such as liquid water and liquid nitrogen. Some typical
calculated results assuming supercritical conditions and other conditions are
briefly reported here.

796 Satoru Yamamoto and Atsushi Ito
2 Preconditioned Flux-vector Splitting(PFVS)

The preconditioning method with the dual-time stepping applied to the 2D
compressible Navier-Stokes equations in curvilinear coordinates are given by
∂ Q̂ ∂ Q̂ ∂Fi ∂Fvi
∂Q/∂t + Γ ∂ Q̂/∂τ + L(Q̂) = Γp +Γ + + =0 (1)
∂t ∂τ ∂ξi ∂ξi
where Q is the vector of unknown variables in conservation form and Q =
J[ρ ρu1 ρu2 e]T . ρ is the density, ui (i = 1, 2) are the velocity vector com-
ponents, e is the total internal energy per unit volume, and J is the Jacobian
for transformation. Q̂ is the vector of unknown variables for the precondition-
ing method and Q̂ = J[p u1 u2 T ]T . p and T are the static pressure and
the static temperature. t and τ are the physical time and the pseudo-time.
Γ is the preconditioning matrix. Γp is the matrix for transformation from
the compressible Navier-Stokes equations to the incompressible Navier-Stokes
equations without the pseudo-compressibility. Fi and Fvi are the vectors of
inviscid and viscous flux.
The numerical flux (Fi )`+1/2 for Fi in Eq.(1) defined at the interface be-
tween the control volume ` and ` + 1 in each coordinate i(i = 1, 2) can be
written by a flux-vector splitting form as
(Fi )`+1/2 = (Fi+ )`+1/2 + (Fi− )`+1/2
−
= (Â+ L R
i )`+1/2 Q̂`+1/2 + (Âi )`+1/2 Q̂`+1/2 (2)
The superscripts ± indicate the sign of characteristic speeds. Fi± and Â±i are
the numerical flux vectors and the preconditioned Jacobian matrices divided
by positive or negative characteristic speeds, respectively. The detail expres-
sion in Eq.(2) has been presented in Ref. [6]. This flux-vector splitting form
is applied to the second-order LU-SGS scheme [7]. This form is extended to a
form for supercritical fluids as
Γ D∆Q̂∗m = RHS m + θL ∆tG+ (∆Q̂∗m ) (3)
∆Q̂m = ∆Q̂∗m − Γ −1 D−1 θL ∆tG− (∆Q̂m ) (4)
G+ (∆Q̂∗m ) = (Â+
1 ∆Q̂
∗m
)i−1,j + (Â+
2 ∆Q̂
∗m
)i,j−1 (5)
G− (∆Q̂m ) = (Â− m − m
1 ∆Q̂ )i+1,j + (Â2 ∆Q̂ )i,j+1 (6)
where m is the number of Newton iterations in each time step. The delta-form
fluxes in Eqs.(5) and (6) can be calculated by using Eq.(2). Then, Q̂L(R) is re-
placed by the time derivative δ Q̂L(R) . The script of each time-derivative flux in
the right hand side of Eqs.(5) and (6) denoted by i, j indicates the grid location
of the flux. D is the diagonal matrix approximated by the spectral radius of
the preconditioned Jacobian matrices and modified by the dual-time stepping
as D = Γ −1 Γp + ∆tθL [r(Âi ) + rv (∂Fvi /∂ Q̂)], where r(Âi ) and rv (∂Fvi /∂ Q̂)
indicate the sum of the maximum values of the spectral radius derived from
the Jacobian matrix Âi and ∂Fvi /∂ Q̂. In this study, these values are approx-
imated by r(Âi ) = αc max[λ(Âi )] and rv (∂Fvi /∂ Q̂) = αv (µ + κ)gii /(ρ∆ξi ),
Near-critical Fluids of Arbitrary Material 797
where αc and αv are empirical constants. λ(Âi ) means the eigenvalues of Âi .
µ and κ are the molecular viscosity coefficient and the heat-conductivity co-
efficient. gii = ∇ξi ∇ξi . RHS m is calculated by the following equation with
the Crank-Nicolson method.
RHS m = −(Qm − Qn ) − ∆tL∗ [(Q̂m + Q̂n )/2] (7)
∗
where L means the finite-difference operator of L in Eq.(1). The vector de-
rived from Γ −1 Γp is actually effective only to the continuity equation.
3 Calculations of Thermal Properties
We have ever employed one of cubic equations, that is, the Peng-Robinson
equation of state(P-R EOS) [5], (p + β)(1 − bρ) = ρRT , where β = aρ2 /(1 +
2bρ−b2 ρ2 ). The square of the sound speed is defined by dp/dρ = ρhT /(ρT (1−
ρhp ) + ρρp hT ) , where ρp , ρT , hp and hT are the partial derivatives of the
density and the enthalpy with respect to the pressure and the temperature.
The preconditioning matrix Γ is also modified by such partial derivatives.
These partial derivatives can be derived theoretically from the P-R EOS. The
P-R EOS should be accurate when the bulk conditions in fluids are at super-
critical state. But the P-R EOS may be inaccurate at subcritical conditions.
Unfortunately even if the P-R EOS is used, the coefficients of thermal conduc-
tivity and molecular viscosity should be calculated by any other mathematical
models, because the theoretical derivation from the P-R EOS is impossible.
In such a situation, we know the PROPATH database developed by
Kyushu University. A number of mathematical models for thermal proper-
ties of arbitrary substances are programmed. This database can calculate
about fifty chemical substances at arbitrary temperature and pressure in a
wide range of them, not only supercritical conditions but also atmospheric
and cryogenic conditions. For examples, the EOS for carbon-dioxide is based
on IUPAC [8], and that for water is made from the steam table presented
in IAPWS IF97 [9]. Other properties such as molecular viscosity and heat-
conductivity coefficients are also programmed according to the original paper
where the mathematical model was proposed. The partial derivatives such as
ρp , ρT , hp , and hT are also prepared in the PROPATH.
4 Numerical Examples
As numerical examples, two-dimensional Rayleigh-Bènard(R-B) convections

in supercritical CO2 and H2 O are numerical investigated. The aspect ratio
of the flow field is fixed at 9 and 217 × 25 grid points are generated for the
computational grid. The R-B convections of supercritical CO2 are first calcu-
lated. The bulk pressure is fixed at 7.3925×106 [Pa]. The top wall temperature
(a) Temperature (b) Density

Fig.1 Rayleigh-Bènard convection in supercritical CO2 (Ideal gas, Ra=8E3)
Ttop is set up to 307.21[K]. These pressure and temperature introduce a near-

critical condition at the top wall. The bottom wall temperature Tbottom is
30[K] higher than Ttop . Thus Tbottom =337.21[K] and ∆T = Tbottom − Ttop =
30. The left and right walls of the field are assumed as an adiabatic wall.
The height length of the flow field is h = 0.8mm. The first case calculated
here uses the above conditions and the equation of state for ideal gas. The
corresponding Rayleigh number is Ra = 8000, where the Rayleigh number is
defined by Ra = gαb ∆T h3 ρ2b Cpb /µb κb herein. The subscript b indicates the
bulk value, and g, α, Cp , µ, and κ are the gravity force, the coefficient of the
heat expansion, the specific heat at constant pressure, the molecular viscosity
coefficient, and the heat conductivity coefficient. Figures 1(a) and 1(b) show
the calculated temperature contours and the density contours. In this case, a
steady state solution is obtained. A typical pattern of the R-B convection is
captured. [b]
(a) t = t1 (b) t = t1 + ∆t
Fig.2 Instantaneous temperature contours(CO2 , Ra=1.3E5)
Next, the above case is calculated with the PROPATH database. The cor-
responding Rayleigh number results in Ra = 1.3 × 105 because of a higher
density near the critical point. The calculated results have a time-dependent
and periodical flow property. The Rayleigh number may be beyond the criti-
cal one. Figure 2(a) and 2(b) show the calculated instantaneous temperature
contours at a different non-dimensional time. The density near the top wall is
relatively higher than that near the bottom wall, because the top wall has a
near-critical temperature. The heaver CO2 moves downward, whereas lighter
CO2 near the bottom wall moves upward due to the buoyancy effect. [t] [t]
The Rayleigh number is increased to Ra = 3×105 by increasing the height
length of the flow field and the same CO2 case except for the height length
Near-critical Fluids of Arbitrary Material 799
(a) t = t2 (b) t = t2 + ∆t
Fig.3 Instantaneous temperature contours(CO2 , Ra=3E5)
(a) t = t3 (b) t = t3 + ∆t
Fig.4 Instantaneous temperature contours(H2 O, Ra=3E5)
(a) Liquid H2 O (b) Cryogenic vapor N2

Fig.5 Density contour-surfaces and temperature contour-lines
of the field is calculated again. Figures 3(a) and 3(b) show the calculated
instantaneous temperature contours at a different non-dimensional time. On
the other hand, the present method is further applied to an R-B convection in
H2 O near critical conditions. The Rayleigh number is fixed to the same value
with the CO2 case, that is, Ra = 3 × 105 . The corresponding bulk pressure is
22.074×106 [Pa] and the top wall temperature Ttop = 650.096[K](∆T = 60[K]).
Figures 4(a) and 4(b) show the calculated instantaneous temperature contours
at a different non-dimensional time. Unsteady flow fields are captured in both
calculated results. Although both cases set the same Rayleigh number, the
wideness of the field at a lower temperature may be different. The field in
H2 O is more widely spread relatively compared with that in CO2 .
This paper focuses on the near-critical fluids. In addition to them, another
distinctive point of the present method can be noticed if the following calcu-
lated results are added here. As a simple flow field, two-dimensional square-
cavity flows are calculated assuming a liquid of water(H2 O) and a vapor of
cryogenic nitrogen(N2 ). The bulk pressure is set up as the atmospheric value.

The corresponding Reynolds number is introduced as 100 in both cases. Fig-
ure 5(a) shows the calculated density contour-surfaces and the temperature
contour-lines in the case of the liquid water. The temperatures of the top wall
and the other wall are fixed to 274.15[K] and 313.15[K]. Since this case is
almost an incompressible flow, the present code may be unnecessary. But, the
accurate density can be calculated by the present method even if the water
is liquid and the difference of density is terribly small. Figure 5(b) shows the
same contours in the case of the vapor nitrogen at a cryogenic temperature.
The temperatures of the top wall and the other wall are fixed to 83.15[K] and
123.15[K]. N2 is vaporized at 77.35[K]. Since the temperature setting in this
case is very close value to the boiling temperature, compressibility appears in
the flow field.
Conclusion
A numerical method for simulating flows of three substances, CO2 , H2 O,

and N2 at supercritical, atmospheric and cryogenic conditions was presented.
Although some programs in the PROPATH seem not to be completed yet,
the present method is expected to apply to not only supercritical fluids but
also arbitrary gas and liquid only if the database will be completed in future.
References
1. A Program Package for Thermophysical Properties of Fluids(PROPATH),
Ver.12.1, PROPATH GROUP.
2. S. Yamamoto and B.-R. Shin, Proc. of the Second Int. Conf. on Computational
Fluid Dynamics, (2002), 112-117, Springer.
3. J.M. Weiss and W.A. Smith, AIAA J., 33(1995), pp.2050-2056.
4. S. Yamamoto, Proc. of the Third Int. Conf. on Computational Fluid Dynamics,
(2004), Springer, in press.
5. D.-Y. Peng and D.B.Robinson, Ind. Eng. Chem. Fundam, 15(1976), pp.59-64.
6. S.Yamamoto, J. Comp. Phys., 207(2005), pp.240-260.
7. S. Yamamoto, S.Kano, and H. Daiguji, Comp. and Fluids, 27(1998), pp.571-580.
8. S.Angus et al., International Thermodynamic Table of the Fluid State-3 Carbon
Dioxide, IUPAC, Vol.3, (1976).
9. Japan Society of Mechanical Engineers, 1999 JSME Steam Tables, (1999).
Flow Physics and Stokes’ Theorem in
Wind Turbine Aerodynamics
Sven Schmitz1 and Jean-Jacques Chattot2

1
University of California, Davis shschmitz@ucdavis.edu
2
University of California, Davis jjchattot@ucdavis.edu
Summary. A viscous lift theorem is derived from a momentum balance and Stokes’
theorem around one section of a wind turbine blade. The theorem is a generaliza-
tion of the classical Kutta-Zhukovsky lift theorem and is validated for 2D attached
and separated flow. The application of the viscous lift theorem within a coupled
Navier-Stokes/Vortex-Panel solver gives insight into the complex 3D aerodynamics
pertinent to wind turbines.
1 Introduction
The accurate prediction of wind turbine aerodynamics is still challenging [1].

Computational Fluid Dynamics (CFD) codes suffer from high computational
cost and rapid artificial dissipation of vortical structures in the wake of wind
turbines and rotorcrafts. Prescribed and free wake models are computation-
ally more efficient, yet fail to predict accurately separated and stalled flow
along the blade, as available two-dimensional (2D) airfoil polar data are used
directly or modified semi-empirically to account for three-dimensional (3D)
effects.
Some hybrid methods have been developed for wind turbines and rotorcrafts
[2, 3, 4] combining a near-field Navier-Stokes (NS) analysis with a far-field
representation by potential and vortex methods, thus reducing both artificial
dissipation and computational cost compared to a full domain NS analysis.
The linking quantity within hybrid methods is the spanwise bound circulation
Γ . In general, Γ inside the NS zone is determined by the Kutta-Zhukovsky
(K-Z) lift theorem from sectional blade loads [3, 4]. However, the K-Z lift the-
orem is strictly valid only for 2D inviscid and incompressible flow. A recently
developed parallelized coupled Navier-Stokes/Vortex-Panel solver (PCS) [2]
determines Γ by integrating along closed contour lines around each section of
blade according to Stokes’ theorem and without further assumptions.
This work presents the derivation of a viscous lift theorem, the validation in
2D, and the application to a parked wind turbine blade as well as a compar-

802 Sven Schmitz and Jean-Jacques Chattot
Fig. 1. Simply Connected Domain Ω - Viscous Flow
ison with measured wind tunnel data [5] and a vortex line method (VLM)
[6].
2 Methods
2.1 Viscous Lift Theorem
Let us assume 2D, steady, viscous, and incompressible high Reynolds number
flow around an airfoil . A simply connected domain Ω is defined in Figure 1
where (C) is the airfoil surface contour, (Σe) is assumed to be in the inviscid
region, and (Σw) passes through the trailing edge orthogonal to the direction
of U∞ . Applying Gauss’ divergence theorem to the gradient components of a
momentum balance in the normal direction yields
Z Z Z Z
0
L = (ρ∞ U ω) dΩ = −ρ∞ U∞ Γ − ρ∞ (u0 ω) dΩ (1)
Ω Ω
0
for the lift per unit length L after neglecting higher order viscous terms. In
Equation (1), ω is the vorticity component normal to Ω, and u0 is a perturba-
tion velocity with U = U∞ + u0 . Note that the Kutta-Zhukovsky lift theorem
is exactly recovered for irrotational flow and ω = 0. The integrand in Eqn. (1)
reads in conservation form:
∂ 1 ∂
u0 ω = (u0 v 0 ) − u0 2 − v 0 2

(2)
∂x 2 ∂y
Substituting Eqn. (2) into Eqn. (1) and applying Green’s theorem we obtain
L0 = −ρ∞ U∞ Γl = −ρ∞ U∞ (Γ + ∆Γw + ∆Γe ) (3)
Flow Physics and Stokes’ Theorem in Wind Turbine Aerodynamics 803
with
I Z Z
Γ = v ds = ω dΩ (4)
(Σw)+(Σe) Ω
Z
1
∆Γw = (u0 v 0 ) dy (5)
U∞ (Σw)
Z
1 1 02
u − v 0 2 dx + (u0 v 0 ) dy

∆Γe = (6)
U∞ (Σe) 2
In Equation (4), the circulation Γ is governed by Stokes’ theorem without

restriction. The simply connected domain Ω includes all sources of vorticity
ω in the viscous region around the airfoil.
2.2 Coupled Navier-Stokes/Vortex-Panel Solver (PCS)

The objective of the PCS method [2] is to combine the advantages of both
NS and vortex methods while maintaining computational efficiency. Navier-
Stokes analysis is mandatory in the vicinity of the blade in order to capture
complex 3D flow phenomena. The PCS method uses a commercial NS code,
CFX 5.7, for the analysis of the near-field and the VLM code [6] for the
far-field. The spanwise bound circulation Γ is the main component of the
coupling methodology. It is determined from closed contour lines such as in
Fig. 1 based on Stokes’ theorem in Eqn. (4). The circulation Γ is distributed
along a bound vortex and the trailing vorticity δΓ along a rigid wake such as
in VLM. A discrete form of the Biot-Savart law is used to calculate induced
velocities on the boundary of the NS zone. The coupling is repeated, until full
convergence is achieved.

3.1 Viscous Lift Theorem
The viscous lift theorem is validated for 2D, steady, and incompressible flow
around the S809 airfoil at Re = 106 using the K − ω SST turbulence model.
Figures 2 and 3 illustrate vorticity magnitude around the airfoil for attached
and separated flow. A uniform spacing of 0.001c was used along Contour 1 and
Contour 2 for both angles of attack. Further refinement along Contour 1 and
Contour 2 revealed an accuracy in Γ of order 10−5 . Solutions were converged
to a maximum residual of 10−6 for all flow quantities. The lift per unit length
L0 was determined by integrating the airfoil surface pressure; quantities in
Eqns. (4-6) were obtained by interpolating nodal values on Contour 1 and
Contour 2. Table 1 gives a quantitative comparison between the lift coefficient
Cl computed by pressure integration on the airfoil surface and through Eqns.
(4-6). The accuracy in Cl is of order 10−3 for all cases. Table 1 reveals that
the contribution of ∆Γw is about 3 percent of Γl for separated flow in Fig. 3.
Fig. 2. S809 (Re = 106 , α = 4.1◦ ) Fig. 3. S809 (Re = 106 , α = 20◦ )
Table 1. Contour Line Integration

L0
S809 Airfoil −Γ −∆Γw −∆Γe −Γl ρ∞ U∞
∆Cl
Contour 1 4.4762 -0.0127 0.0196 4.4861 -0.0006

α = 4.1◦ 4.4787
Contour 2 4.5072 -0.0028 0.0079 4.5123 -0.0043
Contour 1 7.5215 0.2516 0.0521 7.8252 0.0018

α = 20◦ 7.8393
Contour 2 7.5691 0.2599 0.0381 7.8671 -0.0036
3.2 3D Wind Turbine Aerodynamics
The PCS method was applied to the parked National Renewable Energy Lab-
oratory (NREL) Phase VI rotor. Results are compared with measured data
obtained in the NASA Ames wind tunnel [5] and the VLM code [6] as a stand-
alone method. The normal force coefficient CN at r/R = 0.47 is shown in Fig.
4 for PCS, VLM, and NREL data as a function of α47 . The blade setting angle
α47 is the geometric incidence of the NREL blade at r/R = 0.47. PCS results
agree very well with NREL data, while VLM shows discrepancies for separated
and stalled flow. Figure 5 shows results obtained for the bound circulation Γ
at a blade pitch angle of α47 = 23.49◦ . The rapid variation in the PCS bound
circulation near r/R = 0.40 is associated with a vortical structure that is
counter-rotating to the neighboring root vortex and attached to a region of
separated flow more outboard. Figure 5 also shows the PCS (K-Z) bound cir-
culation computed by the classical K-Z lift theorem. It is apparent that values
for Γ deviate for r/R ≥ 0.35. The very good agreement of PCS results with
NREL data in Fig. 4 for α47 = 23.49◦ supports the physical correctness of
the solution and computed PCS bound circulation. The Iso-vorticity surface
in Fig. 6 illustrates the vortical structure trailing downstream of the blade.
Such complex 3D flow phenomena cannot be attained with a vortex model
such as VLM that is based on strip theory.
Flow Physics and Stokes’ Theorem in Wind Turbine Aerodynamics 805
Fig. 4. Normal Force Coefficient CN Fig. 5. Bound Circulation Γ
Fig. 6. Iso-Vorticity Surface (U∞ = 20.1m/s, α47 = 23.49◦ )
Conclusions
A novel method of finding the spanwise bound circulation along a wind tur-
bine blade inside a NS region was proposed from a viscous lift theorem. This
method can serve as an alternative to the classical K-Z lift theorem being
employed in hybrid solvers for wind turbines and rotorcrafts. The bound cir-
culation plays a fundamental role in a coupled Navier-Stokes/Vortex-Panel
solver (PCS) for the computational analysis of the complex flow physics per-
tinent to wind turbines. Some results have been presented and compared with
NREL measurements. PCS is found to give improved performance predictions
compared to a stand-alone vortex method (VLM). The spanwise bound cir-
culation Γ gives insight into the 3D aerodynamics of wind turbines that is
not easily accessible with standard local blade loads such as thrust/torque or
normal/tangential force coefficients.
References
1. Leishman, J. G.: Challenges in Modeling the Unsteady Aerodynamics of Wind
Turbines. AIAA-2002-0037 (2002)
2. Schmitz, S., Chattot, J. J.: A Parallelized Coupled Navier-Stokes/Vortex-Panel
Solver. ASME Journal of Solar Energy Engineering, 127, 475–487 (2005)
3. Xu, G., Sankar, L. N.: Computational Study of Horizontal Axis Wind Turbines.
ASME Journal of Solar Energy Engineering, 122, 35–39 (2000)
4. Bhagwat, M., Moulton, M. A., Caradonna, F. X.: Hybrid CFD for Rotor Hover
Performance Prediction. AIAA-2006-3474 (2006)
5. Hand, M. M., Simms, D. A., Fingersh, L. J., Jager, D. W., Cotrell, J. R., Schreck,
S., Larwood, S. M.: Unsteady Aerodynamics Experiment Phase VI: Wind Tun-
nel Test Configurations and Available Data Campaigns. NREL/TP-500-29955
(2001)
6. Chattot, J. J.: Helicoidal Vortex Model for Steady and Unsteady Flows. Com-
puters and Fluids, 35, 733–741 (2006)
The Role of Separation Bubble on an Airfoil at
Low Reynolds Numbers
Yusuke Nakae1 , Tatsuo Motohashi2 , Satoko Komurasaki3 , and Kunio

Kuwahara4
1
College of Science and Technology, Nihon University, 7-24-1 Narashinodai,
Funabashi-shi, Chiba 274-8501, Japan, m0705013fd@edu.cst.nihon-u.ac.jp
2
College of Science and Technology, Nihon University, 7-24-1 Narashinodai,
Funabashi-shi, Chiba 274-8501, Japan, tmoto@aero.cst.nihon-u.ac.jp
3
College of Science and Technology, Nihon University, 1-8-14 Kanda-Surugadai,
Chiyoda-ku, Tokyo 101-8308, Japan, satoko@math.cst.nihon-u.ac.jp
4
Institute of Computational Fluid Dynamics, 1-22-3 Haramachi, Meguro-ku,
Tokyo 152-0011, Japan, kuwahara@icfd.co.jp
Numerical simulations have been made to clarify the relation between the
behavior of separation bubble and aerodynamic characteristics of an air-
foil(NACA0012) at low Reynolds numbers below 105 . Two and three dimen-
sional time-dependent Navier-Stokes equations are solved by finite-difference
approximation without using any turbulence models. And the multi-directional
finite-difference scheme was also utilized to stabilize computations. Syn-
chronous interaction of a separation bubble near the leading edge and vortical
structure generated at the trailing edge is responsible for the aerodynamic
performance of the airfoil.
1 Introduction
Numerical predictions of aerodynamic characteristics of an airfoil(NACA0012)

show good agreement with the experimental results at high Reynolds numbers
above 106 [1, 2, 3]. At the lower Reynolds numbers below 105 , however, the
flow around an airfoil drastically changes due to transition and separation of
the boundary layer[4, 5, 6, 7]. Crucial role of a laminar separation bubble is
to be examined more precisely. Our objectives are to disclose an interactive
feature of separation bubble with its ambient flow structure to predict the
aerodynamic characteristics of airfoil at low Reynolds numbers. In this study,
we approached these problems by using finite-difference approximation.

808 Yusuke Nakae et. al.
2 Computational Method
The governing equations are the unsteady incompressible Navier-Stokes equa-
tions and the equation of continuity:
divBorn − −M ayeru = 0, (1)
∂Born − −M ayeru
+ (Born − −M ayeru · grad) Born − −M ayeru
∂t
1
= −gradp + ∇p, (2)
Re
where Born − −M ayeru, p, t, Re denote the velocity vector, pressure, time,
and Reynolds number.
These equations are solved by the projection method. The pressure field is
obtained by solving Poisson’s equation derived from Navier-Stokes equations.
And we applied SOR method and the multi-grid method to solve the Poisson’s
equation as follows:
Dn
4p = −div (Born − −M ayeru · gradBorn − −M ayeru) + , (3)
δt
D = divBorn − −M ayeru, (4)
where n is the number of time step and δt is the time increment. Dn+1 is
assumed to be zero, but Dn is retained as a corrective term.
A generalized coordinate system with O type grid is employed to avoid
instability induced by concentrated grid points near the centerline of the wake.
All the space derivative terms except the convection terms are represented
by a central difference approximation. The convection terms are represented
by a third-order upwind scheme(K-K scheme¶ ). And for time integration, the
Crank-Nicolson implicit scheme is utilized.
Multi-directional finite-difference scheme

We applied the multi-directional finite-difference scheme[2] for all terms in this
computation. The multi-directional scheme of two dimensional computation
is based on two local coordinate systems as shown in Fig.1(a). The X’-Y’
coordinate system is generated by rotating the X-Y coordinate system 45
degrees. We calculate the values of physical quantities at the marked points to
improve accuracy of derivative; the ratio of the weight of the X’-Y’ coordinate
system to the X-Y coordinate system is 1:2.
In case of three dimensional computation, we can obtain three different
coordinate systems obtained by rotating a perpendicular plane 45 degrees
with respect to each coordinate axis (i.e. X’-Y’-Z, X’-Y-Z’ and X-Y’-Z’, see
Fig.1(b)). And the ratio of the weight of the X’-Y’-Z coordinate system to the
X’-Y-Z’ and X-Y’-Z’ coordinate system is 1:1:1.
¶
Kawamura-Kuwahara sheme
Separation Bubble on an Airfoil at Low Reynolds Numbers 809
Y Z Z’ Z Z’ Z
Y’ X’ Y’
X’ Y
Y’ Y
X’
X X X Y
X
(a) 2D-Grid (b) 3D-Grid

Fig. 1. 2D & 3D Grid for Multi-directional finite-difference scheme
Fig. 2. Without Multi-directional finite- Fig. 3. Utilized Multi-directional finite-

difference scheme difference scheme
Fig.2 and 3 show comparison of pressure fieldsk between utilized Multi- di-
rectional finite-difference scheme and without Multi-directional finite-difference
scheme. You can see shakes of pressure in the pressure field where Multi-
directional finite-difference scheme is not utilized, especially near the outer
boundary where grid becomes coarse(Fig.2). But, by utilizing Multi-directional
finite-difference scheme, shakes of pressure are prevented and we could get
convergence well(Fig.3).
The numerical simulation was done at the Reynolds number 10000, which is
based on the uniform velocity and the chord length. The angle of attack of
the airfoil is fixed at 10 degrees examining behavior of a separation bubble in
detail.
Fig.4 is the flow field averaged over one period. A pattern of surface oil
traces at the same flow conditions is also obtained experimentally (Fig.5).
A line near the leading edge indicates the position of flow separation and
another line near the trailing edge is so-called a re-attachment line. The region
bordered by the two lines is considered to be a laminar separation bubble. The
agreement between the numerical results (Fig.4) and the experimental one is
satisfactory.
k
Same grid, same caluculation condition
Fig. 4. Time averaged flow field (Contour Line:Pressure, Fig. 5. Separation

Shading:Vorticity, Streamlines) lines (oil flow experi-
ments)
With time development,the lift coefficient Cl and the drag coefficient Cd

fluctuate periodically(see Fig.6). Now, we define one period as an interval
between the peaks where Cl take maximum values. Furthermore we define
phase of fluctuation ϕ in a period.
Sequential change of the flow field in one period is depicted in Fig.7. Six
instantaneous phases are marked by its phase angles. The blue color region
indicates a domain with clockwise vorticity and the red region is dominated
with counter-clockwise vorticity. When the phase angles ϕ are 0 and 140 de-
grees, the lift coefficient Cl takes maximum and minimum values respectively.
The surface pressure distributions are also plotted in Fig.8 at the phase of
maximum lift and in Fig.9 at the phase of minimum lift.
At the case of maximum lift(ϕ=0◦ ), we find a fluid lump of clockwise
vorticity situated approximately 20 percent chord upstream of the trailing
edge. We notice that the lump has its origin in the laminar separation bubble
generated near the leading edge.
Fig. 6. Fluctuations of Cl and Cd

Separation Bubble on an Airfoil at Low Reynolds Numbers 811
On the other hand, at the case of minimum lift(ϕ=140◦ ), the detached fluid
lumps with clockwise vorticity just above the trailing edge induces a vortex
with counter-clockwise vortex(colored red) which produces positive pressure
distributions on the surface of suction side.
As time passed, the separated region and the reattachment point moved
toward the trailing edge. The rolling-up lump sweeps away into the wake. And
the lift coefficient gradually increases.
When the long bubble developed enough, reattachment point suddenly
moved toward the leading edge. Then a newly generated separated region
appeared near the leading edge and results in taking a maximum lift coefficient
again.
4 Conclusion
• Multi-directional finite-difference scheme is useful to keep the stability of

calculation and enhance the convergence of iteration.
• Configurations of vortices around the airfoil are responsible for perfor-
mance of the airfoil; the maximum of lift coefficient is realized when a vor-
tex with clockwise vorticity is located near the trailing edge and then the
lift coefficient becomes minimum when a counter-clockwise vortex reaches
just above the trailing edge.
References
1. K. Kuwahara and S. Komurasaki, ”Direct Simulation of a Flow around a Sub-
sonic Airfoil”, AIAA Paper 2001-2545, 2001.
2. K. Kuwahara and S. Komurasaki, ”Simulation of High Reynolds Number Flows
Using Multidirectional Upwind Scheme”, AIAA Paper 2002-0133, 2002.
3. S. Komurasaki and K. Kuwahara, ”Implicit Large Eddy Simulation of a Subsonic
Flow around NACA0012 Airfoil”, AIAA Paper 2004-594, 2004.
4. Thomas J. Mueller and Stephen M. Batill, ”Experimental Studies of Separation
on a Two-Dimensional Airfoil at Low Reynolds Numbers”, AIAA Paper 80-1440,
1980.
5. M. Brendel and Thomas J. Mueller, ”Boundary-Layer Measurements on an Air-
foil at Low Reynolds Numbers”, AIAA Paper 87-0495, 1987.
6. K. Rinoie, ”Measurements of Short Bubble and Long Bubble Formed on
NACA63-009 Airfoil”, Journal of Aeronautical and Space Sciences Japan 38,
249-257, 1990.
7. K. Rinoie and N. Takemura, ”Oscillating Behaviour of Laminar Separation Bub-
ble Formed on an Airfoil near Stall”, CEAS Aerospace Aerodynamics Research
Conference, Cambridge, 21.1-21.11, 2002.
ϕ = 0◦ Clmax ϕ = 140◦ Clmin
ϕ = 45◦ ϕ = 270◦
ϕ = 90◦ ϕ = 315◦
Fig. 7. Flow fields at each phase (the blue color corresponds to clockwise vorticity
and the red color corresponds to counter clockwise vorticity)
Fig. 8. Pressure distribution at Clmax Fig. 9. Pressure distribution at Clmin

Effect of the thermal boundary conditions and
physical properties variation on transient
natural convection of high Prandtl number
fluids
O. Younis1,2 , J. Pallares1 , and F. X. Grau1
Department of Mechanical Engineering1 , Universitat Rovira i Virgili

43007, Tarragona, Spain, jordi.pallares@urv.cat and younis.obai@urv.cat
Department of Mechanical Engineering2 , University of Khartoum
P. O. Box 321, Khartoum, Sudan
Summary. This work presents numerical analysis of effects of thermal boundary

conditions, fluid variable viscosity and wall conduction on transient laminar natural
convection of high Prandtl number (P r = 104 ) fluid (Golden Syrup) in a cubical
cavity. The simulations consider the flow in a cavity with Plexiglas walls of 1cm of
thickness, the inside dimension of the cavity is L = 20cm. The initial Rayleigh (Ra)
number is 106 , the cavity is initially full of fluid at rest and at constant temperature
(Ti = 45o C) higher than the temperature of the walls (Tw = 25o C). The time
evolution of the flow patterns, the temperature contours, the mean temperature of
the fluid and the Nusselt number (Nu) for four different cases of cooling are presented
and analyzed.
1 Introduction
Transient natural convection in cubical, rectangular and cylindrical cavities
have attracted numerous researchers due to its implications in a wide set of
technological applications and because of its geometric simplicity. Most of
the work done in the thermal storage area considers water as working fluid.
The usage of the oil - or high Prandtl number fluid - as working fluid in
the area of the thermal storage is rare. To the author’s knowledge only a
few studies have been reported. Lin and Akins [1], studied experimentally
the natural convection in cubical enclosures using different kinds of fluids
(6 ≤ P r ≤ 9000)and different sizes of cubes. These author’s found that the
inclusion of the time and/or Pr number does not improve the correlation
between Nusselt and Rayleigh number and that the use of the conventional
correlations is accurate enough for prediction purposes. Ogawa et al. [2] carried
out three-dimensional steady calculation of natural convection in a fluid with
variable viscosity. They classified the flow patterns into two main regimes

814 O. Younis, J. Pallares, and F. X. Grau
depending on the behaviour of the upper boundary layer; the stagnant lid
regime and the whole layer regime. They concluded that the Nusselt number
of the top surface is highly dependent on the viscosity in the whole layer
regime. Davaille and Jaupart [3] reported experimental results of transient
natural convection at high Ra numbers with large viscosity variation in a
tank with insulated bottom wall and cold top wall. They studied the effect of
the viscosity in the unstable top boundary layer on the onset of instabilities.
They introduced a new viscous temperature scale to compute the heat transfer
rate at the cold top wall of the tank and claimed that this scale is sufficient
to account for the characteristics of convection.
The general objective of the present study is to develop and to validate
a simulation tool able to predict the cooling rates of oil in a tank. This tool
has to take into account the variation of the oil viscosity with temperature,
the influence of the conductivity and the thickness of the walls of the tank to
give detailed information of the cooling rates of the oil under different realistic
thermal boundary conditions, as those present in real systems. As a first step
towards this objective, we considered the transient laminar cooling of a fluid
with P r = 104 in a cubical cavity.
2 Physical and mathematical models

The case under consideration is the three dimensional unsteady natural con-
vection of high Prandtl number fluid (Golden Syrup), with viscosity highly
dependent on temperature, in a cubical cavity. The cavity is made of Plexiglas
with dimension of L = 20cm. Initially, the fluid in the cavity is considered to
be at rest, at constant temperature (Ti = 45o C) and the temperature of the
six walls is set to constant value (Tw = 25o C) through the cooling process.
We have also studied the effect of insulating the bottom wall of the cavity.
The physical properties of the golden syrup according to Davaille and Jaupart
kg 2
−4
[3] are: ρ = 1.438 · 103 m3 , β = 4.33 · 10 K −1 and α = 1.21 · 10−7 ms . The
viscosity-temperature relation is given by:

1
µ = µo exp (1)
AT 2 + BT + C
Where: µo = 4.485 · 10−8 P a s,A = −7.5907 · 10−7 ,B = 3.8968 · 10−4 and
C = 4.0130 · 10−2 .
The natural convection is governed by the three-dimensional unsteady

Navier - Stokes equations and the energy equation along with the Boussinesq
approximation. The governing equations in non-dimensional form in Cartesian
coordinates system can be written as:
∂u∗i
=0 (2)
∂x∗i
∂u∗i ∂(u∗i u∗j )

+ =
∂t∗ ∂x∗j
" !#
∂P ∗ ∂ ∂u∗i ∂u∗j
− ∗ + P r(T ∗ ) + + δi1 Rao P ro T ∗ (3)
∂xi ∂x∗j ∂x∗j ∂x∗i
∂T ∗ ∂(u∗j T ∗ ) ∂2T ∗
+ = (4)
∂t∗ ∂x∗j ∂x∗j ∂x∗j
α
The length and velocity scales are L and L , respectively. The non dimensional
∗ T −Tw
temperature is defined as T = Ti −Tw .
The Prandtl and Rayleigh number are defined as:
βg(Ti − Tw )L3 νo
Rao = = 106 , P ro = = 104
νo α α
According to Kays and Crawford [4] , the flow considered can be classified
as laminar. The Prandtl number that appears in the diffusion term in equation
(3) is calculated using equation (1) . The Prandtl and Rayleigh numbers (P ro
and Rao ) in the buoyancy term are based on kinematic viscosity evaluated at
the reference temperature To = (Tw + Ti ) /2.
The associated initial and boundary conditions are:
u∗ = v ∗ = w∗ = Ti∗ = 0 when t = 0
Velocity:
u∗ = v ∗ = w∗ = 0 when t > 0
The different thermal boundary conditions used and the cases considered are
summarised in table (1).
Case Viscosity Thermal boundary conditions Case tag

∗
1 Constant TBC1 (Tallwalls = −1) CV1
∗ ∗
2 Constant TBC2 (Twalls = −1, at bottom wall ∂T ∗
∂x ∗
= 0) CV2
∗
3 Variable TBC2 (Twalls = −1, at bottom wall ∂T
∂x ∗ = 0) VV2
∗
4 Variable TBC1 (Tallwalls = −1) VVWC1
Table 1. Studied cases
Case (4) considers the heat conduction through the 1cm thick Plexiglass
walls, kkfwall
luid
= 0.47 .
3 Numerical methods
The set of governing equations and boundary conditions have been solved
numerically with the finite voulme, second order accuracy 3DINAMICS code.
This code has been validated and tested by Pallares et al [5] , the complete
mathematical formulation can be found in Cuesta [6] .
The grid and time-step dependence study carried out shows that a mesh of
403 control volumes in the fluid and a time-step of 10−7 are accurate enough.
When the wall conduction is considered, the mesh in the solid walls contains
4 control volumes in the direction perpendicular to the wall, resulting in a
mesh of 483 control volumes.
4 Results and discussion

The flow has four planes of symmetry, X − Y plane at Z = 0.5, X − Z plane
at Y = 0.5 and the vertical diagonal planes of the cavity. The presented flow
fields are plotted in the X − Y plane at Z = 0.5.
At the very beginning stage of cooling, the flow near the lateral walls is
descending while it is ascending in the center because of continuity. As a result,
toroidal rolling motion with its axis aligned with the vertical symmetry axis
of the cubical cavity is formed. The thermal boundary layer is confined near
the walls due to the high Pr number. In this early stage, the flow patterns of
all cases are identical except for VVWC1.
As time evolves, the heat transfer thorough the bottom wall (when it is
active) has a small contribution to the flow since it produces a stable stratified
region. This region forces the location of vortex center to move towards the
top wall in comparison with the cases in which the bottom wall is insulated.
The heat transfer through the top wall produces an unstable stratified region,
forcing the fluid at the top to descend. The contribution of the top and bottom
wall produces a secondary flow region near the center of the top wall.
Figures (1(a), 1(b), 1(c) and 1(d))show the velocity vector and the temper-
ature isotherms at t = 1800s. It can be seen that for CV1, the flow velocities
in the lower half of the cube are considerably reduced where a stagnant layer
forms. The convection is confined in the upper half of the cube (figure 1(a)).
Instability effects starts to become evident in CV2 as indicated by the de-
scending plume in figure (1(b)). The flow velocities in the upper half of figure
(1(c)) are considerably reduced pushing the secondary flow towards the cen-
ter of the cavity. No big changes in the lower half are observed due to the
combined effects of TBC and variable viscosity. Figure (1(d)) shows the de-
formation of the secondary flow region and the reduction of the extension of
the hot core.
The time evolution of the volume averaged temperature of the fluid is
plotted in figure (2(a)). As expected, CV1 achieved the minimum temperature
due to the TBC1 and constant viscosity. In the early begging of cooling, the
averaged temperature of CV2 and VV2 are the same. However after t = 600s,
the viscosity variation slightly affects the averaged temperature because of
the presence of the temperature, and viscosity, gradients within the fluid. For
the last two mentioned cases (CV2 and VV2), the difference of the averaged
-0.1
-0.1
-0.2
-0.2
-0.3
-0
.3
7
-0.
-0.4 -0.6
-0.4 -0.5
-0.5
-0.5
-0.6
-0.7 -0.6
-0.8 -0.7
-0.8
-0.9
-0
-1
.9
-1
(a) CV1 (b) CV2
-0.1
-0.2
-0.3
wall
-0.4
-0.1
-0.5
-0.6
-0.2
-0.7 -0.3
-0.8 -0.4 -0.5
-0
-0.6 -0.8
-1
-0.7 -0.9
.9
-1
(c) VV2 (d) VVWC1
Fig. 1. Velocity vectors and temperature contours at t = 1800s

3
10
0 CV1
CV1 CV2
CV2 VV2
Volume averaged temperature(NON−DIM)
VV2 VVWC1
VVWC1
−0.1
−0.2 2
10
−0.3
Nu
−0.4
1
10
−0.5
−0.6
0
−0.7 10
0 500 1000 1500 2000 2 3
10 10
Time (s) Time (s)
∗
(a) Time evolution of T (b) Time evolution of N u
Fig. 2. Averaged quantities
temperature is negligible (at t = 1800s the difference is only 1.4%). The

difference of the averaged temperature of CV1 and CV2 is mainly attributed
to the TBC. The low value of the thermal conductivity of the walls has a great
impact on to the cooling process, as can be deduced from the low value of the
slope of the time evolution of the averaged temperature of VVWC1 as shown

in figure (2(a)) . The fluid cools down very slowly due to the convection delay.
The averaged non-dimensional heat transfer rate is shown in figure (2(b)).
Variable viscosity has no effect on heat transfer rate as shown by the profiles
of CV2 and VV2. At t > 600s, the heat transfer rates of CV1,CV2 and VV2
are not significantly different. For the VVWC1 case, the heat transfer rate
starts from 0 (due to the wall conductivity)and starts to grow as time evolves
in contrast to other cases, where it starts form its maximum (due to the step
change in the temperature) and then decreases.
Conclusions
Transient natural convection of high Pr (104 ) number fluid (Golden syrup) in

a cubical cavity has been studied numerically to determine the combined and
separate effects of thermal boundary conditions, variable viscosity and wall
conductivity. In the very beginnings, the flow consists of a toroidal rolling
motion that constitutes the main flow. Secondary flow starts to develop near
the top wall as time evolves. The insulation of the bottom wall results in
increasing the averaged temperature by 6% after 30 minutes of the initiation
of the cooling process. The instantaneous flow field is significantly affected by
the effect of the variable viscosity. However this effect is damped out in the
averaged quantities (e.g. averaged temperature and heat transfer rate). The
consideration of the wall conductivity adds an additional resistance to the
heat transfer and delays the convection.
References
1. Lin, Y. S. , Akins, R. G.,ASME conference, July1983.
2. Ogawa, M., Schubert, G. , Zebib, A., J. Fluid Mechanics, 233, 299–328 (1991).
3. Davaille, A. , Jaupart, C. , 253, 141–166 (1993).
4. Kays, W. M. , Crawford ,M. E. , (1993).
5. Pallares, J. , Grau, F. X. , Giralt, F. , Int. J. Heat Mass Transfer, 42, 753–769
(1999).
6. Cuesta, I. ,Ph.D. thesis, Universitat Rovira i Virgili, Tarragona, Spain, (1993).
Aeroacoustic simulation of the flow in a
Helmholtz resonator
Laurent Georges1 , Grégoire Winckelmans2 , Stéphane Caro3 and Philippe

Geuzaine1
1
CENAERO, CFD-Multiphysics Group, 30 avenue Jean Mermoz, 6041 Gosselies,
Belgium : laurent.georges@cenaero.be, philippe.geuzaine@cenaero.be
2
Université catholique de Louvain (UCL), Department of mechanical engineering,
1348 Louvain-la-Neuve, Belgium
3
Free Field Technologies, 1 rue Emile Francqui, 1435 Mont-Saint-Guibert,
Belgium
Summary. This paper reports on the application of a computational methodology

for the simulation of aeroacoustics problems. An acoustic analogy is adopted and a
in-house three-dimensional unstructured flow solver is coupled to the Actran/LA
commercial finite element solver that uses a variational formulation of the Lighthill
analogy. Numerical investigations are performed to study the noise radiated by a
Helmholtz resonator placed in a duct.
1 Introduction
After a lot of work has been performed to reduce the noise generated by vi-
brating structures, aerodynamic noise sources are becoming more and more
important in the transportation industry. Nowadays computational aeroacous-
tics (CAA) emerges as a viable solution to reduce the number of prototypes
and tests, and thus the global design cost and duration. Aeroacoustics is gov-
erned by the compressible Navier-Stokes equations. Since a direct numerical
simulation (DNS) of these equations is impractical for engineering problems,
it thus calls for an alternative approach. The approach followed in this paper
consists in using an acoustic analogy, as first proposed by Lighthill [5]. Acous-
tic analogies rest on the assumption that noise generation and propagation are
decoupled, that is, flow generated noise does not impact the internal dynamics
of the flow. In practice, using an acoustic analogy is a two-step procedure. In
the first step, an unsteady computational fluid dynamics (CFD) analysis is
used to compute aerodynamic sources. The second step consists in computing
the propagation and radiation of these aerodynamic sources.

820 L. Georges, G. Winckelmans, S. Caro, P. Geuzaine
2 Unsteady CFD on unstructured meshes

In order to perform the high Reynolds number flows analysis, CENAERO has
developed a parallel implicit solver for three-dimensional compressible flows
on unstructured tetrahedral meshes. The method uses an edge-based hybrid
finite volume and finite element discretization. It blends an upwind scheme
for the convection fluxes based on Roe’s approximate Riemann solver and
a piecewise linear reconstruction of the flow variables in each control vol-
ume, with a P1 finite element Galerkin approximation of the diffusive fluxes.
This second-order accurate numerical scheme, which is representative of most
numerical schemes used on unstructured meshes, was designed for Euler and
Reynolds averaged Navier-Stokes (RANS) simulations, and therefore performs
well for this type of approximations. Since large eddy simulations (LES) have
proved to perform better than RANS for complex phenomena like aerody-
namic noise prediction, this standard discretization on unstructured meshes
has been modified to perform equally well for LES applications.
The main issue with standard CFD schemes is the effect of the numerical
dissipation introduced to stabilize the discretization of the convection term.
It is well-known that this dissipation competes strongly with the effect of
the subgrid scale (SGS) model. An easy way to circumvent this problem is to
resort to central schemes. Unfortunately, central schemes are in general unsta-
ble as the Reynolds number is increased. This behavior is due to the spurious
discrete kinetic energy injection of those schemes. At low Reynolds numbers,
this injection can be counterbalanced by the diffusion term. As a consequence,
the simulation remains stable but results can be unphysical. Nevertheless, the
simulation becomes rapidly unstable at high Reynolds numbers. A possible
way to reach stability is to ensure that the central scheme conserves the dis-
crete kinetic energy. In this paper, the central scheme developed in Georges
et al. [4, 3] for compressible shock-free flows is adopted and further validated
for aeroacoustics applications.
3 Computational aeroacoustics methodology
Within an acoustic analogy framework, a popular approach to the propaga-

tion and radiation of the aerodynamic sources is to rely on explicit integral
methods, amongst which the most famous is the Ffowcs-Williams and Hawk-
ings equation. There are however several limitations to such techniques since
they are practically limited to pure exterior radiation problems and they can
hardly be used in interior problems (e.g. in ducts). Furthermore, they require
compressible CFD simulations, although the vortex structures could mostly
be computed with incompressible CFD well enough.
These limitations have pushed Free Field Technologies to implement an
alternate method in the finite element code Actran/LA. The method is
based on the variational formulation of the Lighthill equation, is designed to
Aeroacoustic simulation of the flow in a Helmholtz resonator 821
be used for exterior or interior problems with or without liners, and has been
shown to possess the potential to handle industrial problems, see for example
Ref. [1]. The theory [2] behind the formulation used in Actran/LA is briefly
summarized hereafter.
Starting from the mass and momentum conservation equations, it is possi-
ble to derive Lighthill’s equation without any assumption, as in the beginning
of the original paper [5]. Then, some classical assumptions, valid only in the
case of a low Mach number and a high Reynolds number flow where isentropic
assumptions are reasonable from an acoustic point of view, lead to dramatic
simplifications. The final equation is a true wave equation whose right-hand
side term is the simplified Lighthill’s tensor
∂ 2 ρa ∂ 2 ρa ∂ 2 Tij
2
− a20 = , (1)
∂t ∂xi ∂xi ∂xi ∂xj
with
Tij ' ρ0 vi vj . (2)
In the frequency domain and adapted to a finite element framework, the for-
mulation becomes [2],
Z !
∂ ρ̃a ∂δρa 1 ∂ T̃ij ∂δρa
− k 2 ρ̃a δρa + + 2 dΩ = 0 ∀ δρa . (3)
Ω ∂xi ∂xi a0 ∂xj ∂xi
The only missing quantity is the source term, represented by the divergence
of the Lighthill tensor in the frequency domain. This quantity is computed in
the time domain by the CFD solver and a Fourier transform is then used
before the field is passed to Actran/LA. The Fourier transform is per-
formed using a dedicated package which comes with the standard distribution
of Actran/LA.
The whole computational aeroacoustics process can be summarized as fol-
lows:
1. The user produces a file with the node coordinates of the subset of the
acoustic mesh where the sources must to be accounted for;
2. The CFD solver reads this file; then, during the unsteady CFD simulation,
at each time step, it computes and writes the divergence of the Lighthill
tensor at all these nodes locations;
3. The user does the Fourier transform of the result, using filters if needed;
4. The user launches the Actran/LA simulation, which gives a direct access
to all acoustic fields in the finite and infinite elements, including some
energy indications.
The files produced for Actran/LA by the CFD solver are written in the
Hierarchical Data Format (HDF), an open-source format used to store named
matrices and developed by NCSA. It is an OS-independent binary format.
The Scientific DataSets of HDF version 4 is used.
4 Analysis of the flow in a Helmholtz resonator

To illustrate and validate the accuracy of the computational aeroacoustic
methodology, the simulation of noise radiated by a Helmholtz resonator in
a duct is considered (see Fig. 1). This testcase has been studied experimen-
tally in the reverberation chamber of the acoustic laboratory at Behr [1]. The
experimental results show that the resonance frequency, which is indepen-
dent of the inlet flow velocity, is equal to 358 Hz which is consistent with the
analytical formula given in Ref. [1].
The isotropic tetrahedral mesh used for the CFD computations is shown
in Fig. 1(a). It contains 580,918 nodes and 3,357,181 tetrahedra and is refined
in the neck region of the resonator (see Fig. 1(c)). The LES is performed
with the WALE SGS model and Reichardt’s wall function. The spatial dis-
cretization of the convective term is based on the kinetic-energy conserving
scheme supplemented by a small amount of high-order artificial dissipation in
form of an hyperdiffusion (see Refs. [4, 3]), while the time-integration uses a
mass matrix and is performed by the second-order accurate three-point back-
ward difference scheme with a time-step equal to 2 × 10−5 s. Figures 2(a) and
2(b) present respectively snapshots of the instantaneous pressure and vorticity
fields.
(a) CFD mesh (b) CA mesh
(c) Close-up for the CFD mesh (d) Close-up for the CA mesh
Fig. 1. CFD and CA meshes for the flow in a Helmholtz resonator

Aeroacoustic simulation of the flow in a Helmholtz resonator 823
(a) Pressure field (b) Vorticity field
Fig. 2. CFD analysis of the flow in a Helmholtz resonator
The mesh used for the computational acoustic (CA) simulations is shown in
Figs. 1(b) and 1(d). It contains 79,436 nodes and 472,786 tetrahedra. The data
generated by the CFD computation and used for the acoustic simulation allows
a frequency resolution of 10 Hz. Figure 3 compares the numerical solution with
available experimental data. Both the frequency, computed at 353 Hz, and the
amplitude of the resonance are well reproduced.
70
Numerical simulation
Experiment
60
50
40
Lw (dB)
30
20
10
0
0 200 400 600 800 1000 1200 1400
Frequency (Hz)
Fig. 3. CA analysis of the flow in a Helmholtz resonator
Conclusions
A computational aeroacoustics methodology has been presented. It is based
on the Lighthill acoustic analogy and couples a in-house three-dimensional
unstructured flow solver to the Actran/LA commercial finite element solver.
The study of the noise radiated by a Helmholtz resonator in a duct reveals

that the proposed computational aeroacoustics methodology is able to predict
accurately the frequency and the amplitude of the resonance.
Acknowledgment
The first and last authors acknowledge the support by the Walloon Region and
the European funds ERDF and ESF under contract N◦ EP1A122030000102.
References
1. S. Caro, P. Ploumhans, F. Brotz, M. Schrumpf, F. Mendonca, and A. Read.
Aeroacoustic simulation of the noise radiated by an Helmholtz resonator placed
in a duct. AIAA paper 2005-3067, May 2005.
2. S. Caro, P. Ploumhans, and X. Gallez. Implementation of Lighthill’s acoustic
analogy in a finite/infinite elements framework. AIAA paper 2004-2891, May
2004.
3. L. Georges, K. Hillewaert, G. Winckelmans, and P. Geuzaine. Development and
study of discretization schemes on sample unstructured grid topologies for large
eddy simulation. In 36th AIAA Fluid Dynamics Conference and Exhibit, San
Francisco, California, June 2006. AIAA-2006-3059.
4. L. Georges, G. S. Winckelmans, and P. Geuzaine. Improving shock-free com-
pressible RANS solvers for LES on unstructured meshes. Journal of Computa-
tional and Applied Mathematics, In press, 2006.
5. M.J. Lighthill. On sound generated aerodynamically. Proc. Roy. Soc. (London),
A 211, 1952.
Explicit low dispersive and low dissipative
non-centered finite differences and filters
Berland J., Bogey C. and Bailly C.
LMFA, UMR CNRS 5509

École Centrale de Lyon
36 avenue Guy de Collongue
69134 Ecully FRANCE
julien.berland@ec-lyon.fr
1 Introduction
Since the earliest stages in computational aeroacoustics (CAA), the need for
highly accurate schemes has been recognized [4]. To meet the stringent accu-
racy requirements of CAA, low dispersive, low dissipative and large spectral
bandwidth numerical methods have been designed by optimizing their disper-
sion and dissipation properties in the Fourier space for low wave numbers.
Available optimized finite differences are for instance the explicit Dispersion-
Relation-Preserving scheme of Tam & Webb [3] or the compact schemes of
Lele [2]. Bogey & Bailly [1] also recently developed explicit finite differences
and selective filters accurate for waves down to four points per wavelength.
These examples deal only with centered schemes designed for uniform mesh
grids, but numerical tools with similar properties are also required for some
local specific features of the computational domain.
The implementation of free-field boundary conditions as well as wall
boundary conditions commonly involves small-stencil schemes which locally
lower the order of accuracy. However, to avoid extra damping and loss of ac-
curacy, it may be recommended to use non-centered optimized schemes with
high accuracy properties. The development of appropriate explicit boundary
filters is in particular of crucial importance and must be performed with at-
tention because standard explicit non-centered filters are unstable [5]. The
boundary filters must therefore be optimized in the Fourier space to form
stable algorithms. With this in mind, non-centered seven- and eleven-point
explicit finite differences and selective filters, with properties optimized in the
wave number space, are proposed in the present work. They are designed to
allow the accurate resolution of waves with at least five points per wavelength.
In the present paper, low dispersive and low dissipative seven- and eleven-
point non-centered boundary schemes are designed in section 2. An acoustic
reflection test problem is resolved in section 3 to highlight the properties of

826 Berland J., Bogey C. and Bailly C.
the developed methods with respect to the standard ones. Concluding remarks
are finally drawn in section 4.
2 Low dispersive and low dissipative algorithms

2.1 Finite differences
On a uniform grid (xi ), the spatial derivative of a function f can be approxi-

mated using a non-centered finite difference scheme by
Q
∂f 1 X
= aj f (xi + j∆x) (1)
∂x i ∆x
j=−P
where ∆x is the mesh spacing and aj are the stencil coefficients. The scheme
uses P points left to xi and Q points right to xi , and it will referred to as
FDP Q . Applying Fourier transform in space to (1) yields the effective wave
number k ∗ of the scheme
Q
X
k ∗ ∆x = −i aj eijk∆x (2)
j=−P
In the present work, following Tam & Webb [3] for instance, coefficients aj
are determined to minimize the dispersion and the dissipation provided by
the scheme. First, to ensure a minimum order of accuracy, the terms of the
Taylor series of (1) are canceled up to the fourth order. Coefficients aj are
then determined to minimize the integral error defined by
Z π/2 h i d(k∆x)
(1 − α)|k∆x − Re(k ∗ ∆x)| + α|Im(k ∗ ∆x)| (3)
π/16 k∆x
where the parameter α is chosen between 0 and 1 depending on the scheme.
Fourth-order seven- and eleven-point non-centered finite-difference schemes
FDP Q have been designed in this way. Coefficients are given in table 1.
2.2 Selective filters
Non-centered selective filtering of a function f can be processed on a uniform

grid (xi ) so that
Q
X
f d (xi ) = f (xi ) − σ dj f (xi + j∆x) (4)
j=−P
where f d is the filtered function, ∆x is the mesh spacing, dj are the coeffi-
cients of the filter, and σ is a constant taken between 0 and 1 defining the
filtering strength. The algorithm is denoted SFP Q . To design the filter and to
investigate its properties, σ is chosen to be equal to 1, its maximum value.
Explicit non-centered finite differences and filters 827
Consider now the Fourier transform in space of (4) to compute the transfer
function of the filtering:
Q
X
G(k∆x) = 1 − dj eijk∆x (5)
j=−P
The filter is made of second order by imposing G(k) = 1 + O(k∆x2 ) with

a Taylor expansion as k∆x → 0. Regularization is enforced with G(π) = 0.
The coefficients dj are then optimized in the spectral space by minimizing the
following integral error
Z π/2 h i d(k∆x)
(1 − α)|1 − G(k∆x)| + α|ϕG (k∆x)| (6)
π/16 k∆x
where ϕG is the argument of the frequency response function and the constant
α is taken between 0 and 1 depending on the filters. In this work, second-
order seven- and eleven-point optimized non-centered selective filters have
been designed. The present optimization failed in providing completely off-
centered filters involving seven and eleven points. However a second-order
four-point filter, denoted by SF03 , has been designed by putting emphasis only
on the amplitude accuracy during optimization. Moreover, the optimization
yields an eleven-point filter SF19 introducing excessive dissipation over a large
range of wavelengths. A seven-point filter SF15 will be therefore used instead
of SF19 . Coefficients of the various optimized filters are given in table 2.
2.3 Properties in the Fourier space – Limits of accuracy
Criteria of accuracy based on half a percent error, |k∆x − Re(k ∗ ∆x)|/π <
∗
5 × 10−3 and |1 − eIm(k ∆x) | < 5 × 10−3 , are introduced to illustrate, re-
spectively, the phase and the dissipation errors. The accuracy limits are given
in table 3 for the different eleven-point finite difference schemes. The non-
centered schemes FD46 and FD37 have dispersion limits close to four points per
wavelength. The schemes FD28 , FD19 and FD010 , are more asymmetric and are
consequently more dispersive: their dispersion limits are about five points per
wavelength. As for the dissipation limits, they are about five points per wave-
length for the schemes FD46 , FD37 , FD28 and FD19 , ranging from λ/∆x = 4.30
for the FD37 scheme up to λ/∆x = 6.87 for the FD19 scheme. The disper-
sion limit of the off-centered finite differences FD010 is about ten points per
wavelength. This scheme is nevertheless used at only one mesh point and the
overall accuracy of the eleven-point non-centered schemes is thus set to five
points per wavelength.
The filter properties are also reported in table 3. Limits of accuracy in
amplitude and in phase are estimated using the criteria 1 − |G(k∆x)| < 5 ×
10−3 and |ϕG (k∆x)|/π < 5 × 10−3 . The dissipation limits of the filters SF46 ,
SF37 and SF28 range from four to six points per wavelength. The filter SF15 is
more dissipative with a limit of fourteen points per wavelength, but it is used
only at one mesh point. The accuracy limit in phase of the filter SF46 , whose
stencil is very weakly asymmetric, is close to two points per wavelength. For
the filter SF37 , the limit of accuracy in phase is about about five points per
wavelength, while it is close to seven points per wavelength for the schemes
SF28 and SF15 . The overall accuracy limit may nevertheless be set to five
points per wavelength as the non-centered schemes are implemented on a
small number of mesh points.
3 Test problem
A linear acoustic problem is considered by the solving the dimensionless equa-
tion:
∂u ∂p ∂p ∂u
+ = 0, + =0 (7)
∂t ∂x ∂t ∂x
on a domain extending from x = −50 to x = 50 and discretized by a
uniform grid with ∆x = 1. Wall boundary conditions are enforced at the
boundaries of the computation domain, i.e. ∂x p = 0 at x = ±50. Numer-
ically, ∂x p is set to zero at the boundaries and ∂x u is computed using the
interior points. The initialn disturbance at to= 0 is defined by u = 0 and
2
p = cos [2πx/(a∆x)] exp − ln(2) [x/(b∆x)] , with a = 8 and b = 12. The
perturbation is thus a wave with a dominant component discretized by eight
points per wavelength.
The system of equations (7) is solved up to time t = 200. At this time
position, each part of the initial disturbance, the left-running and the right-
running parts merge after two reflections into a wave identical to the ini-
tial perturbation. Time integration is performed using an optimized six-stage
Runge-Kutta algorithm [1], and the eleven-point optimized finite differences
and selective filtering of Bogey & Bailly [1] are used for centered spatial deriva-
tives. The CFL number is set to 0.8, i.e. ∆t = 0.8, and filtering strength σ is
set to 0.8.
The problem is first solved with centered schemes at the boundaries. The
centered finite differences and selective filters are the standard schemes whose
coefficients are computed by maximizing the order of accuracy. The 11-point
boundary schemes are then implemented to assess their accuracy.
Computed solutions are plotted in Figure 1. Figure 1.a shows the solutions
obtained using only centered finite differences and filters. The waves are seen
to be strongly damped. The solution obtained using the non-centered finite
differences and filters at the boundaries are presented in Figure 1.b. The
computed perturbation is observed to be in good agreement with the exact
solution, and the gain in accuracy is significant compared to the solution
determined with centered schemes at the boundaries.
Explicit non-centered finite differences and filters 829
(a) (b)
1 1
0.5 0.5
0 0
p
p
!0.5 !0.5
!1 !1
!50 !25 0 25 50 !50 !25 0 25 50
x/"x x/"x
Fig. 1. Computed pressure for the reflection test problem. The gray stripes represent
the wall boundaries. Use of: (a) centered finite differences and selective filters. (b)
non-centered optimized schemes. , computed solution; ◦, exact solution.
Table 1. Coefficients of the non-centered optimized 7- and 11-point finite differences.

FD24 FD15 FD06 FD46 FD37 FD28 FD19 FD010
a−4 0.016756
a−3 -0.117478 -0.013277
a−2 0.048264 0.411035 0.115976 0.057982
a−1 -0.488256 -0.212933 -1.130287 -0.617479 -0.536135 -0.180022
a0 -0.366016 -1.060320 -2.225834 0.341436 -0.274114 -0.264090 -1.237551 -2.391602
a1 1.048005 2.078926 4.827780 0.556397 1.086209 0.917446 2.484732 5.832490
a2 -0.289326 -1.287179 -5.001388 -0.082526 -0.402952 -0.169688 -1.810321 -7.650218
a3 0.050392 0.685176 3.911104 0.003566 0.131067 -0.029716 1.112990 7.907811
a4 -0.003065 -0.245321 -2.115267 0.001173 -0.028155 0.029682 -0.481087 -5.922599
a5 0.041651 0.718883 -0.000072 0.002596 -0.005222 0.126599 3.071037
a6 -0.115276 0.000000 0.000129 -0.000119 -0.015511 -1.014957
a7 0.000000 -0.000119 0.000025 0.170022
a8 -0.000020 0.000156 0.002820
a9 -0.000007 -0.004791
a10 -0.000013
4 Conclusion
Explicit schemes have been proposed for local features of the computational
domain. In particular, non-centered finite differences and selective filters for
boundary conditions have been designed by minimizing their phase and dis-
sipation errors. These schemes are highly accurate for waves with at least
five points per wavelength. For reflection test problems, they provided results
showing a significant improvement of precision with respect to the centered
schemes.
Table 2. Coefficients of the non-centered optimized 7- and 11-point selective filters.

SF24 SF15 SF46 SF37 SF28 SF03
d−4 0.008391 0.320882
d−3 -0.047403 -0.000055 -0.465
d−2 0.032649 0.121439 0.042125 0.052524 0.179118
d−1 -0.143340 -0.085777 -0.200063 -0.173103 -0.206299 -0.035
d0 0.273321 0.277628 0.240069 0.299616 0.353528
d1 -0.294622 -0.356848 -0.207269 -0.276544 -0.348142
d2 0.186712 0.223119 0.122263 0.131223 0.181482
d3 -0.062038 -0.057347 -0.047121 -0.023425 0.009441
d4 0.007318 -0.000747 0.009015 0.013938 -0.077675
d5 -0.000027 0.001856 -0.024565 0.044887
d6 -0.001177 0.013098 -0.009972
d7 -0.002309 0.000113
d8 0.000113
Table 3. Accuracy limits in dissipation and in phase of the eleven-point non-

centered finite differences (top), and dissipation and phase accuracy limits of
the eleven-point non-centered selective filters (bottom), in points per wavelength
λ/∆x = 2π/(k∆x).
FD46 FD37 FD28 FD19 FD010
Phase error |k∆x−Re(k∗ ∆x)|/π<5×10−3 3.77 3.82 5.15 4.60 5.59
∗
Dissipation |1−eIm(k ∆x) |<5×10−3 4.46 4.30 6.04 6.87 10.12
SF46 SF37 SF28 SF15

Dissipation 1−|G(k∆x)|<5×10−3 4.34 5.91 4.80 14.82
Phase error |ϕG (k∆x)|/π<5×10−3 2.12 5.51 7.33 7.25
References
1. Bogey C., Bailly C.: A family of low dispersive and low dissipative explicit
schemes for flow noise and noise computations, J. Comput. Phys., 194, 194–
214 (2003).
2. Lele S.K.: Compact finite difference schemes with spectral-like resolution, J.
Comput. Phys., 103, 16–42 (1992).
3. Tam C.K.W., Webb J.C.: Dispersion-Relation-Preserving finite difference
schemes for computational acoustics, J. Comput. Phys., 107, 262–281 (1993).
4. Tam C.K.W.: Computational aeroacoustics: issues and methods, AIAA J.,
33(10), 1788–1796 (1995).
5. Visbal M.R., Gaitonde D.V.: Very high-order spatially implicit schemes for com-
putational acoustics on curvilinear meshes, J. Comput. Acous., 9(4), 1259–1286
(2001).
A New Principle of Non-Reflecting and its
Application to Hyperbolic Conservation Laws
Ching Y. Loh1 and Philip C.E. Jorgenson2

1
NASA Glenn Research Center, Cleveland, Ohio 44135, USA,
Ching.Y.Loh@grc.nasa.gov
2
Taitech Inc.
1 A first principle of non-reflecting

Non-reflecting boundary conditions (NRBCs) play a key role in fluid dynamics
computations and still remain a challenging topic in the current research areas.
The characteristics-based NRBC is one of the most important non-reflecting
treatments [2], [3]. In the present paper, however, instead of following the
classical wisdom of signal propagation along characteristics, the non-reflecting
concept is revisited from a different perspective and a new but simple principle
of non-reflecting is proposed, followed by a generic NRBC.
A first principle of non-reflecting
Recall a general observation that spurious reflection can only be generated
at the domain boundaries, no reflection ever occurs at an interior point.
Theoretically, the observation can be justified as given below. Consider the
Euler equations:
∂V ∂V ∂V ∂V
+ Ã + B̃ + C̃ = Q̃, (1)
∂t ∂x ∂y ∂z
where V = (ρ, u, v, p)T is the primitive flow variable vector. Equation (1) can
be locally linearized in a small neighborhood of an interior point O by setting
the jacobians Ã, B̃, C̃ to their values at O. The solution of the linearized
Euler equations (1) is now the superposition of a homogeneous solution and
a nonhomogeneous solution.
The nonhomogeneous solution of (1) with zero initial conditions is fur-
nished by the classical Duhamel’s principle for first order PDE system ([1],
p.204). The homogeneous solution is the solution of (1) with Q̃ = 0 and
the given initial conditions. By using the Cauchy’s method of the Fourier
integral (see [1], pp.210-211), the homogeneous solution is decomposed into
a sum of the Fourier integrals of plane waves. A plane wave in the form of
V = Ṽeiθ can be easily solved by substituting it in (1) (with
√ Q̃ = 0). Here
θ = θ(x, t) = x · k − ωt is the phase of the simple wave, i = −1, x = (x, y, z).
The homogeneous part of (1) then becomes:

832 Ching Y. Loh and Philip C.E. Jorgenson
∂ Ṽ ∂ Ṽ ∂ Ṽ ∂ Ṽ
( + Ã + B̃ + C̃ ) + i(K̃ − ωI)Ṽ = 0, (2)
∂t ∂x ∂y ∂z
where I is the 5 × 5 identity matrix. For any given wave number vector k =
(kx , ky , kz ), the matrix K̃ = kx Ã + ky B̃ + kz C̃. From the hyperbolicity of (1),
real ωs (eigenvalues) exist such that det|K̃ − ωI| = 0 (dispersion relation),
where the amplitude vector Ṽ = Ṽ(k) may be determined as an eigenvector
of K̃. For each plane wave (Fourier mode), both its amplitude Ṽ and phase
θ remain intact when passing through this interior point O, no distortion or
reflection occurs.
Based on such observations and the theoretical justification, a first princi-
ple of non-reflecting is proposed:
A boundary of the solution domain is said to be non-reflecting if
every point on the boundary surface can be considered as an interior
point.
There are two basic requirements defining an interior point:
1. the point must be physically located in the domain interior;
2. the consistency requirement that the governing PDE system is the only
condition to be satisfied at this point. Other conditions, if any, must com-
ply with the governing PDE system.
In order that all the boundary points be interior points, the solution domain
must be extended to include ghost cells. Then the consistency requirement
leads to:
The continuity criterion - a generic NRBC
Assume that in the domain exterior (i.e. ghost cells), the flow is also gov-
erned by another Euler equation system. If the two Euler equation systems
are identical at all the points on the boundary, the consistency requirement is
satisfied and these boundary points are interior points of the extended domain
(Fig.1a) and hence non-reflective. The continuity criterion leads naturally
to such non-reflecting, it is stated and proved as follows:
At each point on the boundary, if the flow variables V and their
spatial derivatives, namely, Vx , Vy , Vz , are continuous; or, V is
C 1 continuous in space across the boundary, the boundary is non-
reflecting .
Proof : Consider the Euler equations (1) for both domain interior and
exterior. Note that the jacobian matrices Ã, B̃, and C̃ are directly functions
of V (details may be found in [3]). As a result of the C 1 continuity of V, all
the terms
∂V ∂V ∂V ∂V
∂t , ∂x , ∂y , ∂z , Ã, B̃, C̃, and Q̃
in (1) are identical across the boundary surface. Here, continuity of ∂V ∂t is
inferred from (1). Equation (1) now may be considered as a single governing
equation system in the extended domain. The consistency requirement is sat-
isfied at all the boundary points. Hence there is no reflection at the boundary.
Non-reflecting principle 833
2 Implementation of the absorbing NRBC

The C 1 continuity criterion is a generic NRBC. It is here combined with
the conservation laws in discretized form and yields a desirable absorbing
NRBC - simple but robust, and truly multi-dimensional. After discretization,
the boundary surface S is also divided into small boundary elements, ∆S,
according to the grid. Let CV be a small control volume which contains ∆S
as its internal interface or cell surface. Note that CV may or may not be a grid
cell. Then, with any prescribed physical boundary conditions at the ghost cell
center, ∆S is already non-reflective, no additional numerical implementation
is needed. The mechanism works as follows:
1. At the reconstruction stage in the finite volume scheme, a linear and hence
locally C 1 continuous vector function V is rebuilt within the CV . As ∆S
is part of CV , the local C 1 continuity of the flow variables V across ∆S
provides the non-reflecting property.
2. Then possible inconsistent flow data between the imposed boundary con-
ditions and the domain interior is treated as a ‘numerical jump’ around
the boundary, captured, and absorbed by the shock-capturing scheme.
As a result, a matched layer (ML) or absorbing layer is automatically
generated around the boundary.
Here are some general guidelines in numerical implementation:
1. Choose an appropriate control volume, CV , at the boundary and ensure
that the boundary surface element, ∆S, is included within the CV .
2. In the reconstruction stage, in order to keep V (or the conservative flow
variables U U = (ρ, ρu, ρv, ρe)T ) C 1 continuous within the entire control
volume CV , the usual dimension by dimension reconstruction of gradients
of V (or U) is inappropriate. Instead, truly multi-dimensional unsplit
reconstruction procedure is recommended. One of such procedure in 2-D
space is outlined below.
At the reconstruction stage, U are already updated at the four cell centers
O, D, E and F . Let u denote a component of U, by simple finite difference, a
linear equation for the gradients (ux , uy ) is obtained:
(xD − xO )ux + (yD − yO )uy = uD − uO ,
Here, the subscripts O and D denote the corresponding cell centers. Sim-
ilarly, two more such linear equations are available from the u data at E and
O as well as at F and O. Generally, any combination of two of such equations
yields a set of spatial gradients for u. There are three sets of such gradi-
(1) (1) (2) (2) (3) (3)
ents, namely,
q (ux , uy ), (ux , uy ) and (ux , uy ). Their moduli or norms
(i) (i)
mi = (ux )2 + (uy )2 , i = 1, 2, 3, are calculated. Then, a flux limiter is ap-
plied to all these spatial gradients to achieve a single, unified set of gradients
(i) (i)
for u. When the minmod limiter is employed, the gradients (ux , uy ) corre-
sponds to mi = min(m1 , m2 , m3 ) is chosen. Similarly, other flux limiters may
be used.
a b
C
boundary pt. now, a new
interior pt. F E
R L L
O R
interior pt.
extended domain
!S
L
A boundary point becomes an A R B
interior pt., but consistency boundary
requirement must be observed ghost cell
surface D cntr.
ghost cell
Fig. 1. a: A boundary point becomes an interior point in the extended domain;

b: An upwind scheme with unstructured triangular grid; ‘R’ and ‘L’ represent the
right and left states for the Riemann solvers, they are obtained by extrapolating
flow data linearly from the current cell center O and the neighboring cell centers D,
E and F . The prism cylinder based on ∆ABC is the space-time control volume. ∆S
practically coincides with the side AB.
Numerical implementations with other grid cell shapes, e.g. quadrilateral,

are similar and they can be consistently extended to 3-D tetrahedral or hexa-
hedral grids. This absorbing NRBC is flexible and valid with practically any
cell shapes, structured or unstructured, and flow of any speed, supersonic or
subsonic.
3 Discussions on the NRBCs

Theoretically, the first principle of non-reflecting offers a perfect generic
NRBC. However, after discretization, the center O of ∆S (Fig. 2) is employed
to represent the entire ∆S. Consider the Fourier mode in a plane wave, i.e.,
eiθ = ei(k·x−ωt) . Here, θ is the phase and θ = const. stands for a wavefront (or
a characteristic surface, see [1] or [3]). The spatial phase error is closely related
to the spurious reflections and is less tolerable. Assume that time t is held un-
changed, the spatial phase error can be easily analysed: Let x = (x, y, z) be
the position vector of any point on ∆S and xO the position vector of the
center O. Then, the spatial phase error ∆θ due to replacing x by xO is:
∆θ = k · x − k · xO = k · (x − xO )
When k is normal to ∆S, ∆θ = 0. Otherwise, reflections may occur. Reme-
dies include: 1. change the orientation of the boundary according to k (wave
direction) or 2. apply a buffer zone.
4 Numerical examples
Figure 3 shows propagation of an acoustic pulse together with an entropy
wave and vortex in a square domain. Comparison with exact solution is also
Non-reflecting principle 835
boundary wave No.
surface
x k
x
O
0
boundary
a b surface
element ! S
wave front (characteristic surface) c
Fig. 2. Wave propagation across a boundary; a: normal to the boundary, b: oblique
to the boundary, phase error may occur; c: sketch of x − x0 and k.
shown. The problem in Fig. 4 is a 2-D radial Riemann problem in an oval

domain. Initially, a circular region with high pressure and density exists at
the domain center. The same radial Riemann problem is solved in Fig. 5 but
with a different domain. They show how a shock or contact is absorbed at the
boundary. A 3-D radial Riemann problem in a sphereical domain is demon-
strated in Fig. 6. Initially, a cubic region with high pressure/high density is
located at the center of the sphere. All flows move at 45o with M = 0.2 in the
x − z plane. It is shown how the shock and contact develop and are absorbed.
t=60
Fig. 3. Propagation of acoustic pulse, entropy and votex, a:isopycnics, b:comparison

to exact solution. A CE/SE scheme is employed.
rho contact
shock
contact
t=300 t=500 t=700
Fig. 4. 2-D radial Riemann problem in an oval domain. Density contours show how
a shock is almost perfectly absorbed by the boundary. The mean flow is stationary
(M = 0). A Godunov scheme is used.
Fig. 5. The same Riemann problem in a peanut shape domain.
Fig. 6. 3-D Radial Riemann problem in a sphere, initial high pressure region is a
cube. One million unstructured hexahedral cells with the CE/SE scheme are used.
Concluding Remarks
In the present paper, the problem of non-reflecting is revisited from a different

perspective and a generic NRBC is then derived. The plane waves are treated
as a physical entity and remain intact across a non-reflecting boundary. The
principle applies more generally to hyperbolic systems other than the Euler
equations.
When combined with hyperbolic conservation laws in a finite volume
scheme, the absorbing NRBC is simple and robust. Limitation and remedy
are also discussed.
References
1. R. Courant, and D. Hilbert “ Methods of Mathematical Physics”, Vol.II, John
Wiley & Sons, London, 1962.
2. B. Engquist and A. Majda “Absorbing boundary conditions for the numerical
simulation of waves”, Math. Comp., Vol. 31, pp. 629-651 (1977).
3. C. Hirsch “Numerical Computation of Internal and External Flows” Vol. I and
II, John Wiley & Sons (1993).
4. F.Q. Hu “On Absorbing Boundary Conditions for Linearized Euler Equations
by a Perfectly Matched Layer”, J. Comput. Phys., Vol.129, 201-219 (1996).
5. C.Y. Loh and P.C.E. Jorgenson, AIAA paper 2005-4716, 2005.
Numerical Solution of the Linearized Euler
Equations Using Compact Schemes
Kris Van den Abeele1,3 , Jan Ramboer2 , Ghader Ghorbaniasl1 , Chris Lacor1
1
Vrije Universiteit Brussel, Dept. of Mechanical Engineering, Fluid Dynamics and
Thermodynamics Research Group, Pleinlaan 2, 1050 Brussels, Belgium
kvdabeel@vub.ac.be, ghader@stro.vub.ac.be, chris.lacor@vub.ac.be
2
Cardijnstraat 11A, 1980 Eppegem, Belgium jan.ramboer@gmail.com
3
FWO Research Engineer
Summary. For computational aeroacoustics (CAA), schemes with low dissipation

and dispersion errors are a necessity. This is because of the small amplitude of the
acoustic waves and because of the large distances over which they have to travel.
Also, non-reflecting boundary conditions play a very important role, since spurious
reflections could significantly reduce the quality of the solution. To damp out small
spurious waves that might occur, Artificial Selective Damping (ASD) can be used.
In this paper, the performance for CAA of a fourth order central compact finite
volume scheme, in combination with ASD and two different sets of boundary con-
ditions, is investigated.
1 Numerical Algorithm
1.1 Governing Equations
The linearized Euler equations (LEE) are given by the following general form:
∂U
+ ∇ · F (U) = 0 (1)
∂t
where the unknowns are U = [ρ 0 , v 0 , p 0 ]T and the fluxes are given by:
ρ0 v 0 + v0 ρ 0
 
0
F (U) =  v0 ⊗ v 0 + pρ0 I  (2)
v0 p 0 + γp0 v 0
ρ0 , v0 and p0 are the mass density, velocity and pressure of the mean flow,
while ρ 0 , v 0 and p 0 are small perturbations superimposed on this mean flow.

838 K. Van den Abeele, J. Ramboer, G. Ghorbaniasl, C. Lacor
1.2 Finite Volume Formulation
Spatial discretization of (1) for the 2D case on a Cartesian mesh gives:

dŪij 1 h i 1 h i
=− F̃i+ 12 ,j − F̃i− 21 ,j − F̃i,j+ 12 − F̃i,j− 21 (3)
dt ∆x ∆y
where the Ūij are the unknowns of the problem, averaged over the cell with
indices (i, j). ∆x and ∆y are the width and height of the cells. F̃i+ 21 ,j is
the fluxe averaged over the cell face with indices (i + 12 , j), which has to be
approximated. In the present work, the mass density, velocity and pressure
perturbations averaged over a cell face are approximated with a fourth-order
accurate implicit compact scheme as proposed by Lacor et al. [1]:
1 1 3
Ūi+1,j + Ūi,j + O ∆x4

Ũi+ 32 ,j + Ũi+ 12 ,j + Ũi− 12 ,j = (4)
4 4 4
Since this scheme is central, the diffusion error is uniformly zero. An analysis
of the dispersion error can be found in [1]. The tridiagonal systems defined by
(4) can be solved with the Thomas algorithm. After the face-averaged values
are determined, the fluxes can be calculated using (2).
1.3 Artificial Selective Damping
To stabilize the central scheme and to dissipate spurious high-frequency waves,

artificial selective damping (ASD) as proposed by Tam [2] was implemented.
The following term is added to the right hand side of (3):
k=3
1 X
Dij = − dk (Ui+k,j + Ui,j+k ) (5)
Rem
k=−3
Rem is the mesh Reynolds number, which determines the amplitude of the
damping term. The coefficients dk = d−k are chosen such that only the high
frequency waves are affected by this term.
1.4 Time Integration
For time integration, an N -stage low-storage Runge-Kutta algorithm is used:
U0ij = Unij
Ukij = U0ij + ∆t k−1
N −k+1 Rij 1≤k≤N (6)
Un+1
ij = UNij
In (6), Rkij corresponds to the right hand side of (3), evaluated at stage k. ∆t
is the time step and n is the time index. The N -stage algorithm is N -th order
accurate in time for linear problems.
Solution of the Linearized Euler Equations Using Compact Schemes 839
2 Non-reflecting Boundary Conditions

2.1 Characteristic Solution Boundary Conditions
If one assumes that the solution is a function only of the normal coordinate and
time near the boundaries, (1) reduces to a system in one spatial dimension.
An eigenvalue analysis of this system reveals the following characteristic wave
solutions, with accompanying wave speeds:
0
w1 = 21 ρp0 c − 1n · v 0 v0 · 1n − c
0
w2 = ρ 0 − pc2 v0 · 1n
(7)
w3 = v 0− (1n · v 0 ) 1n v0 · 1n
0
w4 = 21 ρp0 c + 1n · v 0 v0 · 1n + c
In (7) c is the speed of sound and n denotes the direction normal to the wall.
Expressing that there are no incoming waves provides a number of physical
boundary conditions. The remaining boundary conditions are determined by
extrapolation (numerical boundary conditions). For instance, at a subsonic
inlet w2 , w3 and w4 should all be zero. The fourth boundary condition is
obtained by extrapolating e.g. the pressure perturbation. These boundary
conditions will be referred to as the CSBC boundary conditions.
2.2 Radiation and Outflow Boundary Conditions

The radiation boundary condition of Tam [2] is given by:
1 ∂U ∂U U 5
+ + = 0 + O r− 2
V ∂t
∂r 2r
q
2 2
V = c M0 (1v · 1r ) + 1 − M0 |1v × 1r | (8)
In (8), r is the distance to the sound source, M0 is the mean flow Mach
number and 1r and 1v are the unit vectors
in the r-direction and the direction
of v0 . Notice the error term O r−5/2 . It shows that the error caused by
the boundary conditions will diminish with increasing distance between the
boundary and the source. The radiation boundary conditions are suitable
for a boundary where only acoustic disturbances exit the domain. If entropy
and vorticity waves also pass through the boundary, an outflow boundary
condition, the formulation of which can be found in [2], must be used. These
boundary conditions will be referred to as the ROBC boundary conditions.
3 Numerical Results
3.1 2D Linearized Euler - Grid Convergence Tests
As a first testcase, the advection of the three characteristic wave solutions of
the 2D LEE’s, namely the entropy, vorticity and acoustic wave, was computed.
Grid convergence tests were performed for each

√ wave. The initial solution was
a 2D Gaussian profile with a half width of 0, 004 for each case. The calcu-
lation parameters for these tests are included in Table 1. Figure 1 shows the
Table 1. Parameters of 2D LE calculations.

Physical ρ0 1, 0 p0 1, 0 ux0 0, 5c = 0, 592
Parameters γ 1, 4 c 1, 183 uy0 0, 0
Mesh xmin −0, 25 ymin −0, 5 ∆x = ∆y 0, 02/0, 01/0, 005
Parameters xmax 0, 75 ymax −0, 5 Nx = Ny 50/100/200
Other Parameters ∆t 10−4 Rem 10, 0 CFL σmax 0, 1775
L2 -error on the pressure perturbation versus the mesh size at three different
moments in time, for the acoustic pulse. At t = 0.2, the wave has not yet
reached the boundary, at t = 0.5 it is going through the boundary and at
t = 1.0 it has left the domain. From the figure it can be seen that without
ASD the scheme is fourth order accurate when the wave is inside the do-
main. With ASD, this fourth order accuracy is lost. When the wave reaches
the boundary, neither of the boundary conditions preserves the fourth order
accuracy. The error is the largest for the CSBC boundary conditions. It is
caused by the fact that the tangential velocity component is forced to zero,
which in general is not true for the exact solution. For the ROBC boundary
conditions, the error is caused by the term O r−5/2 . However, this term can
be reduced by increasing the distance from the source to the boundary.
The graphs for the entropy and the vorticity wave are not shown here. For
these waves, the advantage of ASD is more apparent, since the error after the
waves have left the domain is smaller because the reflected high frequency
waves are damped. For the entropy wave, both boundary conditions conserve
the fourth order accuracy. For the vortex however, only the ROBC boundary
conditions give good results. It can thus be concluded that the latter are more
suited. For the remaining testcases, ROBC boundary conditions are used.
3.2 2D Linearized Euler - Dipole and Quadrupole Sound Sources
For these calculations, source terms were added to the right hand side of
(1). The source terms corresponding to dipole and quadrupole sources can be
found in Bailly and Juvé [4]. They result in the sound fields in Figure 2.
3.3 Rotor Noise
The last testcase was the rotor noise testcase as proposed by Tam in [3]. A
non-uniform mesh was used for this, the parameters of which can be found
Solution of the Linearized Euler Equations Using Compact Schemes 841
Fig. 1. Convergence test acoustic pulse with CSBC (left) and ROBC (right).
Fig. 2. Dipole (left) and quadrupole (right) sound pressure field at t = 640∆t.
– – : values from −0.011 to 0.001 with step 0.001. — : values from 0.001 to 0.011
with step 0.001.
in Table 2. The directivities of the sound field radiated by an open rotor

with dimensionless angular velocity Ω of 0.85 and 1.15 are shown in Figure
3, along with the analytical solutions. There is a good agreement with the
latter in both cases. In Figure 4, the sound fields for the ducted rotor are
Table 2. Parameters of rotor noise calculations.

Ω m xmin xmax rmin rmax ∆xnear = ∆rnear ∆xf ar = ∆rf ar
0.85 8 −17, 05 17, 05 0 17.01 0.02 0.08
1.15 8 −17, 03 17, 03 0 17.03 0.02 0.05
plotted. The noise is not radiated outside the duct in the subsonic case, since
the frequency is below the cut-off limit. In the supersonic case, the noise is
radiated outside the duct.
Fig. 3. Directivity of Rotor Sound Field for Ω = 0, 85 (left) and Ω = 1, 15 (right).
Fig. 4. Sound Pressure for Ducted Rotor at Ω = 0, 85 (left) and Ω = 1, 15 (right).
Conclusions
A code for the solution of the 2D LEE’s and the rotor noise equations using a
fourth order accurate compact finite volume scheme was developed. Two dif-
ferent sets of boundary conditions were tested, of which the ROBC’s produced
the best results. The code was tested for the advection of characteristic waves,
for dipole and quadrupole sound sources and for the sound field radiated by
an open and a ducted rotor. Good results were achieved for all testcases.
References
1. Lacor, C., Smirnov, S., Baelmans, M.: A finite volume formulation of compact
central schemes on arbitrary structured grids. J. Comput. Phys., 198, 535–566
(2004)
2. Tam, C. K. W.: Numerical Methods in Computational Aeroacoustics. AIAA
Short Course Computational Aeroacoustics: Methods and Applications, Lahaina,
Hawaii, 10-11 June 2000.
3. Third Computational Aeroacoustics (CAA) Workshop on Benchmark Problems
(2000), NASA CP 2000-209790.
4. Bailly, C., Juvé, D.: Numerical Solution of Acoustic Propagation Problems Using
Linearized Euler Equations AIAA Journal, Vol. 38, No. 1, 2000, pp. 22-29.
Transonic Aeroelastic Computations of a Delta
Wing Configuration with High Fidelity
Equations
Hiroshi Terashima and Kenji Ono
RIKEN (The Institute of Physical and Chemical Research), 2-1 Hirosawa, Wako,
Saitama, 351-0198, Japan, htera@riken.jp / keno@riken.jp
1 Introduction
With the progress of Computational Fluid Dynamics (CFD) technologies,

Navier-Stokes equations have been applied to many transonic aeroelastic anal-
yses such as flutter simulations for the accurate prediction of nonlinear aero-
dynamics. Importance of the nonlinear aerodynamics such as the shock wave
motion has been fully recognized through those numerical analyses and also
experimental studies. In the transonic flutter simulations, only the linear
structural methods like modal analysis are often used for the computation
of structural deformation and are considered to be enough approaches for
the accurate flutter predictions. Nonlinearities in structure, therefore, have
not been almost discussed in the transonic flutter simulations compared with
those in aerodynamics. Recently, Gordnier [2] and Attar and Gordnier [3] nu-
merically investigated the large amplitude Limit Cycle Oscillation (LCO) of a
cropped delta wing of Schairer and Hand [1] and indicated the importance of
the geometrical nonlinearities in structure for capturing the LCO with large
amplitudes. And also, in the low subsonic flow regime, importance of modeling
the nonlinearities of both the fluid and structure for the accurate prediction
of LCO was presented by the numerical and experimental study of Sheta et
al..[4]
Although the importance of nonlinearities in structure has been indicated
through those LCO computations, the numerical analysis using the nonlin-
ear structural models coupled with high fidelity aerodynamic models is still
limited to quite few examples. Therefore discussions on the use and choice of
the structural models for the wing responses with large deformation have not
yet been made sufficiently. Comprehensive comparisons of various structural
models on a aeroelastic wing with large deformation are required. This paper
presents some comparisons of various structural models on aeroelastic compu-
tations of a transonic delta wing with large deformations. Effects of structural
models are investigated on the flutter prediction at first. Comparison between

844 Hiroshi Terashima and Kenji Ono
the linear and nonlinear structural models are performed on wing responses
with large deformation. In the comparison, influences of the number of selected
mode in linear structural analysis and assumption of strain-displacement rela-
tionships in nonlinear structural analysis are discussed. Finally, a criterion for
the use and choice of structural models will be set on aeroelastic computations
of a transonic delta wing with large deformations.
2 Outline of Numerical Methods

The governing equations for the flow fields are the unsteady compressible
Navier-Stokes equations written in the generalized coordinate system. The
time derivative term is discretized by the three-point backward differencing
and the LU-SGS implicit method is used for the time integration. The numeri-
cal fluxes are evaluated by the SHUS scheme, which is a family of AUSM-type
schemes. The higher-order spatial accuracy is achieved by the MUSCL with
the primitive variables interpolation. The Baldwin-Lomax model is used for
the estimation of the turbulent viscosity.
The linear and nonlinear equations of motion are used for the computation
of structural deformations. The damping term is ignored and a materially
linear elastic model is used since the strains are assumed to be small. In this
study, the following four cases are considered for the comparison of structural
models.
1. Modal analysis (6 or 12 modes) used in most flutter computations
2. Full linear analysis (all the modes considered)
3. Nonlinear analysis (linear part of tangent stiffness only considered)
4. Full nonlinear analysis (nonlinear part of tangent stiffness included)
The governing equations are discretized by the finite elements and the total-
Lagrangian formulation is used in the nonlinear analysis. The Mixed Inter-
polation of Tensorial Components (MITC) shell element considering finite
rotation increments [5] which has the superior convergence characteristics is
applied. The time integration is accomplished by the Newmark β method.
The loosely coupling approach is applied for the coupling between the
fluid and structural equations. In this study, Constant Volume Tetrahedron
(CVT) transformation method proposed by Goura et al. [6] is employed for
the displacement data exchanges from structural grid points to the fluid grid
points. The CVT defines a tetrahedron from one fluid grid point and one
triangulated structural grid. The motion of the fluid grid is derived from
the known motion of the structural grid by conserving the volume of each
tetrahedron. The CVT can handle not only out-of-plane displacements but
also in-plane displacements which is important for the nonlinear structural
computations. Actually, the structural grid is first interpolated onto a finer
grid using B-spline interpolation before the CVT is applied. This processing
can generate well smooth fluid surface grid. The two-dimensional mapping
Transonic Aeroelastic Computations of a Delta Wing 845
algorithm presented by Bhardwaj et al. [7] is used for the transformation of

aerodynamic forces from the fluid grid points to the structural grid points. The
required cost of the data exchanges was about 5% of the fluid computations.
Computational flow conditions and structural parameters are taken directly

from the experiment of Schairer and Hand and the computations of Gordnier.
Freestream Mach numbers are around 0.87 and the corresponding Reynolds
numbers are set between 2.5 × 106 and 5.0 × 106 . Dynamic pressures are
varied from 2.29(psi) through 5.46(psi). Geometries of the delta wing model
are shown in Fig. 1. The wing model was made of the steel plate with the
uniform thickness of 0.035(in.).
The eigenvalue analysis was first performed for the understanding of the
wing vibration characteristics. The computed first three structural modes were
the first bending, the first torsion, and the second bending. The present results
agree well with that computed by Schairer and Hand.
The aeroelastic computations are started from the corresponding steady
flow fields calculated at the angle of attack, 0(deg). At the steady flow fields,
the shock wave does not appear on the surface although in the transonic flow
regime. The nondimensional small velocity, 0.0001 for the first mode is given to
the structural equation as the initial perturbation. The nondimensional time
step size is ∆t = 0.001 which was proved to be sufficient for the estimation of
the required wing responses in preliminary studies and corresponds to about
12,000 iterations per cycle.
First a comparison of flutter dynamic pressure using each structural model
is made to assess the ability of flutter prediction. Although not shown in this
paper, each time response indicates that flutter dynamic pressure exists be-
tween dynamic pressure of 2.29 and 2.58 where the experimental data shows
a flutter onset at about 2.4. This indicates the wing responses near the flutter
onset is insensitive to the structural models. Computation using the linear
structural model coupled with Navier-Stokes equations is suitable for the pre-
diction of flutter onset in accuracy and efficiency.
Figure 2 shows the time responses of the wingtip trailing-edge using sev-
eral structural models at the dynamic pressure of 2.58. It is clearly seen that
each linear structural model produces the limit cycle oscillations with well
large amplitudes which are about 60 times the wing thickness. The time re-
sponses using the linear structural models indicate almost the same limit cycle
oscillation in the amplitude and frequency. Therefore, this result means that
the number of selected structural mode does not influence the wing responses
even with the large amplitude. On the other hand, the time response of the
nonlinear structural model demonstrates the limit cycle oscillation with rel-
atively smaller amplitude. The structural deformation is actually suppressed
due to the geometrical nonlinearity. As mentioned later, the amplitude and
Fig. 1. Geometries of the delta

Fig. 2. Time responses using various struc-
wing model
tural models at dynamic pressure of 2.58(psi)
Fig. 3. Flow fields and deformed shape using linear and nonlinear structural model
at dynamic pressure of 2.78(psi)
frequency using the nonlinear structural model is closer to that observed in

the experiment.
The reasons for generating and sustaining the limit cycle oscillation in
the linear structural model are the appearance of wingtip vortex during the
oscillation in fluid and the unphysical constant stiffness in structure. Figure
3 shows the flow fields and deformed shapes using the linear and nonlinear
structural model at a dynamic pressure of 2.78. It can be observed that the
large vortex is generated on the surface with the quite large deformation in
the linear structural model. Such a vortex does not appear in the nonlinear
structural model. of structure.
In Fig. 4 the amplitudes of the limit cycle oscillation are shown for
three other computations and the present computation of nonlinear struc-
tural model coupled with Navier-Stokes equations. Also included are the ex-
perimental results of Schairer and Hand. For the range of dynamic pressure
below 3.88, there are little difference between the present computation and
the others with the nonlinear structural models. This result means the present
approach including the coupling techniques is valid for the computation of the
Transonic Aeroelastic Computations of a Delta Wing 847
wing responses with large deformation. Although all the computations with
the nonlinear structural models could not capture the experimental trend
accurately, note that the published available data on aeroelastic wing with
large deformation is only this one. For the range of larger dynamic pressure
above 4.0, differences appear apparently between the results with the non-
linear structural models. The reason for these differences has not yet been
understood clearly.
Fig. 4. Comparison of present nonlinear Fig. 5. Comprehensive comparison of

results with published other results structural models
In Fig. 5 comparisons of various structural models is made in the ampli-

tudes. As mentioned previously, it is shown that the results using the linear
structural models have quite larger amplitudes than the experimental data
and those using the nonlinear structural models. It is concluded that the lin-
ear structural models should not be applied for the computations of wing
responses with divergence tendency except for the flutter prediction. On the
other hand, the amplitudes computed using the nonlinear structural models
increase mildly with dynamic pressures and the tendency in the amplitudes
for dynamic pressures is closer to the experimental data.
As a comparison in the nonlinear structural models, effect of the nonlinear
part of tangent stiffness is investigated. The amplitudes indicate that the re-
sults with the nonlinear part of tangent stiffness are almost the same as those
only with the linear part of tangent stiffness. Therefore, the nonlinear part of
tangent stiffness does not have effects on the wing responses over the present
range of deformation. And also, this fact might suggest that the differences
between the present results and the other results for the range of dynamic
pressure above 4.0 (psi) shown in Fig. 4 do not come from the assumption on
strain-displacement relationships in nonlinear structural analysis. The turbu-
lence modeling and grid quality in aerodynamic analysis might be important
for the wing responses over such a range of deformation.
As criteria for the structural models on the aeroelastic computations of

wing response with large deformation, this study indicates that the nonlinear
stiffness is indispensable for the reasonable wing responses with large deforma-
tion, but higher-order terms in stress-displacement relationships do not have
much effect over the present range of deformation.
Conclusions
Comparisons of linear and nonlinear structural models coupled with Navier-
Stokes equations have been conducted on aeroelastic computations of a delta
wing with large deformation in the transonic flows. Some criteria on the struc-
tural models are made for the better understanding on the computation of
aeroelastic wing responses. The linear structural models are suitable for the
accurate predictions of transonic flutter boundary. There are little differences
between the linear and nonlinear structural models on the decision of flutter
boundary. Discussions on the wing responses with large deformation after a
flutter onset should be made by the nonlinear structural models. The higher-
order terms in the strain-displacement relationships dose not influence the
wing responses over the present range of deformation.
Acknowledgement. The author gratefully acknowledges the help of profes-

sor Hirohisa Noguchi of Keio University in developing the nonlinear finite
element structural solver. This study was supported by Industrial Technology
Research Grant Program in ’04 from New Energy and Industrial Technology
Development Organization (NEDO) in Japan.
References
1. Schairer, E., Hand, L.: AIAA 97-2217 (1997)
2. Gordnier, R. E.: Journal of Aircraft., 40, No. 6, 1206–1207 (2003)
3. Attar, P. J., Gordnier, R. E.: AIAA 2005-1915 (2005)
4. Sheta, E. F., Harrand, V. J., Thompson, D. E., Strganac, T. W.: AIAA 2000-1399
(2000)
5. Noguchi, H., Hisada, T.: Computer & Structures, 47, No. 4/5, 699–710 (1993)
6. Goura, G. S. L., Badcock, K. J., Woodgate, M. A., Richards, B. E.: The Aero-
nautical Journal, 105, 215–221 (2001)
7. Bhardwaj, M. K., Kapania, R. K., Eichenbach, E., Guruswamy, G. P.: AIAA
Journal, 36, No. 12, 2179–2186 (1998)
Transonic Buffet over Symmetric Airfoils
Alexander Kuz’min1 and Alexey Shilkin2

1
St Petersburg State University, Institute of Mathematics and Mechanics,
28 University Ave., 198504 St Petersburg, Russia
alexander.kuzmin@pobox.spbu.ru
2
University of Dortmund, BCI, Chair of Fluid Separation Processes,
70 Emil-Figge-Strasse, D-44227 Dortmund, Germany
a.schilkin@bci.uni-dortmund.de
Summary. Supercritical airfoils typically have small curvature of the upper surface
as compared with that of a conventional airfoil. At the same time, physical phe-
nomena associated with transonic flow over flattened surfaces have not been well
understood. For instance, the inviscid steady flow over such surfaces may exhibit
an unexpected bifurcation and intricate instability of its entire structure at certain
freestream conditions [1, 2, 3, 4]. In this paper, we examine the viscous transonic
flow. A numerical study confirms the bifurcation and structural instability of the
flow past a flattened airfoil. Apart from the bifurcation, computations demonstrate
periodic oscillations (buffet) caused by the shock-induced boundary layer separation.
1 Formulation of the Problem and a Numerical Method
We consider two-dimensional flow past the symmetric airfoil

p
y(x) = ±0.06 1 − (2x − 1)4 (1 − x16 )2 , 0 ≤ x ≤ 1. (1)
A computational domain is bounded by airfoil (1) and a C-type far-field
boundary Γ = Γ1 ∪ Γ2 ∪ Γ3 ∪ Γ4 placed at 15 to 18 chord lengths from the
airfoil. The angle of attack α, Mach number M∞ < 1, and static temperature
T∞ are given on the inflow part Γ1 of Γ , while the static pressure p∞ is given
on the outflow part Γ4 . The pressure is related to the density and tempera-
ture by the equation of state with the specific heats ratio of 1.4. The free-slip
condition is used on the upper and lower segments Γ2 ∪ Γ3 which are parallel
to the freestream velocity vector. The no-slip condition is prescribed on the
airfoil. Initial data are either a uniform flow determined by the freestream
conditions or a steady flow obtained previously for other values of M∞ and α.
Unsteady solutions of the Reynolds-averaged Navier-Stokes equations were
obtained with a finite volume solver in which the equations are discretized in
space on unstructured meshes using the Roe scheme, a MUSCL reconstruc-
tion, and Van Albada type limiters. The time derivatives are approximated

850 Alexander Kuz’min and Alexey Shilkin
with the second-order backward Euler scheme. We used the SST−k − ω tur-
bulence model which predicts the buffet onset quite reasonably [5].
Computations were performed primarily on a mesh of about 2 × 105 grid
points which were clustered in the regions of strong gradients of the solution.
The high accuracy of the method was confirmed by application to the buffet
problem for the 18% circular-arc airfoil and comparison with results available
in the literature [6].
2 Transonic Flow Bifurcation and Buffet at α = 0

For all the considered Mach numbers 0.823 ≤ M∞ ≤ 0.844, and T∞ = 250 K,
p∞ = 29800 N·m−2 computations showed self-sustained oscillations of shock
waves and boundary layers at airfoil’s aft. In addition, for
0.8296 ≤ M∞ ≤ 0.8375 (2)
computations revealed the non-uniqueness of transonic flow past airfoil (1).
The shaded domains in Fig. 1 indicate the margins of oscillations of the lift
coefficient CL at various M∞ or the freestream velocities U∞ . Each domain
corresponds to the flow field with two, three or four local supersonic regions
as sketched in Fig. 1. The realization of a definite flow field depends on the
initial data and the time history of freestream conditions.
Figure 2 demonstrates instantaneous Mach number contours obtained by
solving the problem at M∞ = 0.824 and then increasing M∞ step-by-step to
0.832. In this case, the mean flow is symmetric with respect to the x−axis
and involves two pairs of local supersonic regions. Figure 3 displays the de-
pendence of CL on time which looks similar to the one obtained for other
thick symmetric airfoils in [6, 7].
For M∞ = 0.832 the replacement of the vanishing angle of attack by
α = 0.1◦ followed by the reset α = 0 entails an abrupt transition to the
asymmetric flow with two local supersonic regions on the lower side of the
airfoil and a single supersonic region on the upper side (see Fig. 4). Restruc-
turing the flow field is caused by the instability associated with the coales-
cence/rupture of local supersonic regions [4]. The obtained asymmetric flow
persists when M∞ varies in the interval (2). The Reynolds number based on
airfoil’s chord and the midvalue of interval (2) is 5.4 × 106 .
At M∞ ≥ 0.8368 there exists transonic flow with a single supersonic region
on each side of the airfoil (see Fig. 5 and the right domain in Fig. 1). The
amplitudes of oscillations at 0.8368 ≤ M∞ ≤ 0.840 are reduced as compared
with those for the quasi-symmetric flow at smaller values of M∞ . This is ac-
counted for by the expansion of the supersonic regions and, hence, weakening
the feedback via the subsonic region under the increasing freestream Mach
number.
The bifurcation interval (2) is contracted by about a quarter and shifted
to the greater values of M∞ as compared with the one obtained for inviscid
Transonic Buffet over Symmetric Airfoils 851
flow [4]. The error in calculation of the endpoints of interval (2) under the
fixed model of turbulence is estimated by ±0.0004 with test computations on
refined meshes.
For the quasi-symmetric flows, the frequency of oscillations drops from 95
to 76 Hz with M∞ increasing from 0.823 to 0.844. For the asymmetric flows
involving three local supersonic regions, the frequency of oscillations decreases
from 89 to 78 Hz with M∞ increasing within the interval (2).
At a double freestream pressure p∞ and corresponding Re = 10.8 × 106 ,
the amplitudes of oscillations increase, whereas the bifurcation interval slightly
shrinks (see Fig. 6). At the end of this interval, an abrupt drop of CL for the
asymmetric flow is caused by instability of the shock foot F residing on the
rapidly thickening boundary layer.
At a halved p∞ and Re = 2.7 × 106 , the oscillations of CL slightly weaken.
The frequency of oscillations also insignificantly depends on Re. This fact,
along with the flow bifurcation, makes the flattened airfoil (1) markedly dif-
ferent from the 18% circular-arc airfoil.
Conclusion
Turbulent transonic flow past airfoil (1) is non-unique at α = 0 in the narrow

range (2) of freestream Mach numbers. In this range, each obtained flow pre-
serves its main structure in spite of the self-induced oscillations of the bound-
ary layers at airfoil’s aft. A slight unsteadiness of the freestream conditions is
expected to trigger structural transitions enhancing the flow fluctuations.
References
1. Jameson, A.: Airfoils admitting non–unique solutions of the Euler equations.
AIAA Paper, No. 91-1625, 1–13 (1991)
2. Hafez, M., Guo, W.: Nonuniqueness of transonic flows. Acta Mechanica, 138,
177–184 (1999)
3. Caughey, D. A.: Stability of unsteady flow past airfoils exhibiting transonic non-
uniqueness. CFD Journal, 13, No. 3 (60), 427 – 438 (2004)
4. Kuz’min, A.: Instability and bifurcation of transonic flow over airfoils. AIAA
Paper, No. 2005-4800, 1–8 (2005)
5. Geissler, W., Koch, S.: Adaptive airfoil. In : Sobieczky, H. (ed) Proc. IUTAM
Symposium Transsonicum IV. Kluwer, Dordrecht, 303–310 (2003)
6. Geissler, W., Ruiz-Calavera, L.P.: Transition and turbulence modelling for dy-
namic stall and buffet. In: 4th Internat. Symp. on Engineering Turbulence Mod-
elling and Measurements, Corsica, France (1999)
7. Hafez, M., Guo, W.: Some anomalies of numerical simulation of shock waves.
Part II: effect of artificial and real viscosity. Computers and Fluids, 28, 721–739
(1999)
Fig. 1. Lift coefficient CL versus M∞ for the oscillating transonic flow past airfoil
(1) at α = 0, Re = 5.4 × 106
Fig. 2. Instantaneous Mach number contours in the flow with two pairs of local
supersonic regions at α = 0, Re = 5.4 × 106 , M∞ = 0.832
Transonic Buffet over Symmetric Airfoils 853
Fig. 3. Oscillations of the lift coefficient for the transonic flow shown in Fig. 2
Fig. 4. Instantaneous Mach number contours in the asymmetric flow with three
local supersonic regions at α = 0, Re = 5.4 × 106 , M∞ = 0.832
Fig. 5. Instantaneous Mach number contours at α = 0, Re = 5.4 × 106 , M∞ =

0.839
Fig. 6. Lift coefficient CL versus M∞ for the oscillating transonic flow past airfoil
(1) at α = 0, Re = 10.8 × 106 . The lower shaded domain is dropped in order
to avoid excessive complications of the diagram.
Aerodynamic Performance of a Deforming
Elastic Body in Supersonic Flow
Tomohisa Hashimoto1 , Koji Morinishi2 and Nobuyuki Satofuka3

1
Venture Laboratory, Kyoto Institute of Technology, Matsugasaki, Sakyo-ku,
Kyoto 606-8585 Japan hasimoto@kit.ac.jp
2
Department of Mechanical and System Engineering, Kyoto Institute of
Technology, Matsugasaki, Sakyo-ku, Kyoto 606-8585 Japan
morinishi@kit.ac.jp
3
The University of Shiga Prefecture, Hassaka-cho, Hikone-shi, Shiga 522-8533
Japan satofuka.n@office.usp.ac.jp
1 Introduction
Numerical simulations of fluid-structure interaction problem are one of the

hot topics in today’s Computational Fluid Dynamics (CFD). It is an inter-
esting problem that if an elastic body can greatly deform in response to the
fluid dynamic force, what shape does it change into? What is the relation
between the shape and its drag force like? For example, a parachute functions
as equipment for the control of rate of descent. It is expanded by the fluid dy-
namic force and becomes the designated shape to increase the air resistance.
Such a system of utilizing the fluid dynamic force is desirable from the point
of view of saving energy. However we need to understand the every-changing
flow field around a moving and deforming body.
The purpose of this study is to design the body shape by using the fluid
dynamic force in supersonic flow, assuming that the elasticity can be changed
from place to place. As a first step in this attempt, we developed a numerical
method for simulating a comparatively deformable elastic body under the fluid
dynamic force in incompressible flow [1]. To treat the interactions between a
largely deformable elastic body and surrounding flow field, we proposed an
elastic body model. In this model, the elastic body, whose inside is filled with
fluid, moves and deforms according to the difference between the inside and
outside pressures and the shear stress acting on the surface of the body. There
are also the interactive effects of restoring force and damping force of the body.
Numerical simulations of compressible unsteady supersonic flow around a
deforming elastic body in response to the fluid dynamic force were carried
out under the constraint of fixed volume and center of gravity. The relation
between the shape and its drag force was investigated.

856 Tomohisa Hashimoto, Koji Morinishi and Nobuyuki Satofuka
2 An Elastic Body Model
m
k
i
i-1/2
c !l i+1/2
i-1 i+1
Fig. 1. An elastic body model
The equations of motion based on the wave equation with the term of
damping effect are formulated for the shell made of elastic film. For example,
a balloon that has been filled with fluid. The film is composed of material
particles connected with elastic springs and dampers as shown in Fig. 1, so
that the film can expand and contract with the effect of damped vibration
in response to the fluid dynamic force. In Fig. 1, the segment shown in the
shaded portion of the film corresponds to a material particle. The notations
m, c, k and ∆l are the mass of a material particle, the damping coefficient,
the spring coefficient and the length of segment respectively. In this model, it
is assumed that the grid points are distributed on the circular film at regular
intervals, and the position corresponds to that of a material particle. The
equations of motion in the system of a material particle are given by
d2 xi
m = pi + τi + fi r + fi d + fi p (i = 1, · · · , N ) (1)
dt2
pi = −pout n∆l, τi = Tn∆l (2)
r
fi = k(xi+1 − 2xi + xi−1 ) (3)
fi d = c(ui+1 − 2ui + ui−1 ) (4)
where N is the number of material particles.
The external force consists of the fluid dynamic force, the elastic body force
and the artificial force. In the fluid dynamic force, the pressure pi and shear
stress τi given in Eq. (2) act on the film, where pout is the outside pressure,
T is the viscous stress tensor and n is the unit normal outward vector to the
film. These forces are obtained by solving the Navier-Stokes equations. In the
elastic body force, the restoring force fi r and damping force fi d given in Eq.
(3) and Eq. (4) are defined, where x and u are the position vector and velocity
vector of a material particle respectively.
Aerodynamic Performance of a Deforming Elastic Body 857
Furthermore, suppose that the elastic body deforms under the constraint
for fixed volume and center of gravity. Note that in the case of 2D the volume
is area.
In order to keep the area constant, the following additional force with
adjustable pressure pin is introduced inside the body.
fi p = pin n∆l (5)
The pressure can be determined by solving the following equation through
sub-iterations,
pν+1 ν
in = pin + α(S0 − S), α > 0 (6)
where S0 is the initial area and α is a constant parameter for control of the
area.
In order to fix the center of gravity, the additional force is given to offset
all forces acting on the body as follows:
The equation (1) is rewritten as the general form
d2 xi
m = fi (7)
dt2
by substituting fi as
fi = pi + τi + fi r + fi d + fi p (8)
Consider now summation in Eq. (7),
X d2 xi X
m 2 = fi (9)
dt
the equation of motion for the center of gravity is derived as
d2 x̄
M =F (10)
dt2
with the relations
P
X X mxi
M= m, F = fi , x̄i = P (11)
m
where M , F and x̄ are the total mass, the resultant force and the position
vector at the center of gravity respectively. The restoring force and damping
force automatically satisfy
X X
f r ≡ 0, f d≡0 (12)
By using the resulting Eq. (10) together with Eq. (7), finally the equation of
motion for fixed center of gravity is obtained as
d2 (xi − x̄) F 1
m = fi − m = fi − F (13)
dt2 M N
3 Numerical Method
Numerical simulations are carried out by coupling the compressible Navier-
Stokes equations together with the equations of motion given in Eq. (13). The
unsteady compressible flow field around a deforming body is solved by using
sub-iterations with pseudo time at every physical time step. In the numerical
method, a central difference scheme with artificial dissipation [2] is used for the
spatial discretization. An iterative scheme in the time integration is the lower-
upper symmetric Gauss-Seidel implicit relaxation method [3]. The overset grid
technique [4] is adopted to remesh the grid easily in treating the deforming
body.
4 Flow around a Deforming Elastic Body under Fluid

Dynamic Force
In the present simulations, an elastic body of a circular film begins deforming
in response to the fluid dynamic force under the constraint of fixed volume
and center of gravity.
The overset grid is composed of an irregular Cartesian grid, which covers
the whole computational domain, and an overlapped O-type grid around the
body. The number of grid points is 129 × 129 for the Cartesian grid and
129 × 33 for the overlapped grid. The overlapped grid is remeshed according
to the motion of the body. Both the physical and pseudo time increments are
0.01.
The elastic body starts deforming from the steady flow around a circular
cylinder obtained at a free stream Mach number of 2.0 and Reynolds number
of 1.0 × 103 . Figure 2 shows Mach number contours for the steady flow. The
adjustable parameters of the elastic body at the initial state are set as k = 30,
50 and c = 100. Note that the elastic body stays stationary in the absence of
the fluid dynamic force. For the two different spring coefficients, the relation
between the shape and the drag force of the elastic body is investigated. The
results are shown in Figs. 3-5 for k = 30 and Figs. 6-8 for k = 50. Figure 3 and
Figure 6 show Mach number contours at the time when the flow field becomes
a steady state. The elastic body changes to a crescent moon-like shape for
k = 30, and a hamburger bun-like shape for k = 50, which is about 2 times
and 1.3 times longer respectively. The time histories of the shape are shown in
Fig. 4 and Fig. 7. As for k = 30, it is observed that the position of shock wave
moves upstream as the elastic body becomes a slender shape, and then moves
downstream after beginning to return from the most slender shape with about
3 times longer at about t = 30. On the other hand, in the case of k = 50, the
defromation of the elastic body is suppressed at about t = 10 due to a larger
spring coefficient. The time histories of the drag coefficient are shown in Fig.
5 and Fig. 8. In the case of k = 30, the pressure drag drastically increases
by 3.2 to 4.6 of the peak at about t = 30. After that it decreases and finally
Aerodynamic Performance of a Deforming Elastic Body 859
reaches 2.3. The total drag is considered to be the pressure drag because the
friction drag is almost negligible. In the case of k = 50, while the friction drag
slightly decreases, the pressure drag gradually increases until about t = 10
and remains 2.0, so that the total drag increases by 0.6.
Fig. 2. Mach number contours for a circu- Fig. 3. Mach number contours (k = 30)
lar cylinder
1.5 6.0
1.0 5.0
0.5 4.0
Y 0.0 Cd 3.0 Total drag
-0.5 2.0
Pressure drag
-1.0 1.0
Friction drag
-1.5 0.0
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 0 150 300
X Nondimensional time
Fig. 4. Histories of body shape (k = 30) Fig. 5. Histories of drag coefficient (k =

30)
1.5
1.0
0.5
Y 0.0
-0.5
-1.0
-1.5
-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
X
Fig. 6. Mach number contours (k = 50) Fig. 7. Histories of body shape (k = 50)
6.0
5.0
4.0
Cd 3.0
Total drag
2.0
Pressure drag
1.0
Friction drag
0.0
0 25 50
Nondimensional time
Fig. 8. Histories of drag coefficient (k = 50)
Conclusion
The relationship between the shape and its drag force of the elastic body
was investigated. As the elastic body becomes a slender shape, the total drag
increases, where the friction drag is almost negligible in comparison to the
pressure drag. A considerable difference is observed in the time histories of
pressure drag for different spring coefficients. In this paper the spring coeffi-
cient is assumed to be constant over the elastic body, however it is desirable
to control the drag force by changing the spring coefficient from place to place
so that the designated shape can be made under the fluid dynamic force.
References
1. Hashimoto, T., Morinishi, K., Satofuka, N., Numerical simulation of unsteady
flow around deformable elastic body under fluid dynamic force, ECCOMAS 2004,
Paper 364, 2004, pp.1-19
2. Jameson, A. and Baker T.J., Solution of the Euler equations for complex con-
figurations, AIAA Paper 83-1929,1983
3. Yoon, S. and Kwak, D., Multigrid convergence of an LU scheme, Frontiers of
CFD, 1994, pp.319-338
4. Steger, J.L. and Benek, J.A., On the use of composite grid schemes in computa-
tional aerodynamics, Computer Methods in Applied Mechanics and Engineering,
64, 1987, pp.301-320
CAD-Centric Framework for Aero-mechanical
Optimization - Counter-Rotating Fan Design
Hiromasa Kato, Stéphane Pierret, and Rajan Filomeno Coelho
CENAERO, avenue Jean Mermoz 30, Gosselies, Belgium

hiromasa.kato@cenaero.be
1 Introduction
Traditionally, the three dimensional blade shape is very often designed by ex-
perienced designers by iteratively and manually modifying the blade shape,
taking into account several results coming from aerodynamic computations
and structural mechanics computations (static and dynamic), among others.
In recent years, progress has been made in the development of automatic
optimization packages capable of optimizing complex shapes using advanced
CFD solver and optimization algorithms [4, 6]. However they are not yet
commonly used in industrial shape optimization. A more intensive use still
requires progress in the field of automatic shape optimization. Some of the
issues are 1) the large number of design variables typical of industrial prob-
lems, 2) the efficient and robust shape parameterization, and 3) the ability to
easily couple multiple simulation tools with minimum labor. The first two are
numerical efficiency issues, while the third one is often overlooked, but just as
important if automatic shape optimization is to be incorporated in a realistic
industrial environment. Emphasized here is the role of CAD systems as an
integral element of optimization framework. A major advantage of modern
CAD systems is that design intent can be incorporated into the CAD model
as a master model, which is a set of parameters that control features within
the model. However the simulation codes often lack access to the feature tree
and parameters defined in the model due to file translations, which results
in loss of information and robustness. In addition, manual intervention that
may be required to heal the model results in a procedure that is not general
enough and automated to be acceptable in industrial cases. It is therefore ad-
vantageous to integrate the CAD-system directly inside the simulation chain,
and an object-oriented framework was developed that can interact with the
underlying CAD systems transparently, based on the CAPRI [2] middleware
which in turn modifies the shape design variables, regenerates the CAD model,
and provides a native geometry representation that can be used for analyses.

862 Hiromasa Kato, Stéphane Pierret, and Rajan Filomeno Coelho
2 The Optimizer
The optimization algorithm is based on the use of genetic algorithms because
they provide a very robust method responding to the criteria mentioned above.
One drawback of GAs is that they suffer from a slow convergence because of
its stochastic nature. One approach to accelerate the optimization process is
to use an approximate model. Ref. [3] presents an overview of several similar
optimization techniques. The blade design algorithm is therefore organized
with the following 5 steps :
1. Build a database using a design of experiments procedure (DOE).
2. Build an approximate model using the DOE points, i.e. construct an ana-
lytical relation between the design variables and the simulation responses.
3. Perform a GA optimization to find the optimum using the approximate
model to evaluate the objective functions and constraints.
4. Perform accurate simulations to evaluate and verify the real objective and
constraint values. This new simulation result is added to the database. The
database is therefore always improved with new design points, leading to
an improved approximate model.
5. Go to step 2 until the maximum number of design cycles specified by the
user is reached.
In this research, the radial basis function interpolation technique is used to
build an approximate model, mainly because of its robustness in providing a
more accurate approximate model. Moreover it allows constructing a global
approximate model which is valid for the entire design space. This is an im-
portant aspect for the application of the method to multiple objective op-
timization techniques based on the Pareto front concept, which may need
information from the whole search space. For industrial applications, the com-
putational cost of one optimization iteration mainly depends on the cost of
the simulation employed. In general the cost for building the approximate
model and running the genetic algorithm is from a few seconds to a few
minutes, depending on the number of training examples and the number of
input and output variables. More details about this method can be found in
Refs. [7, 6, 8].
3 The Flow Solver
For flow computations the TRAF code developed by the University of Florence
is used in this research [1]. The code can compute the flow in linear and annular
cascades as well as the flow in stationary and rotating blade passages with
tip clearance. The RANS equations are solved using a Runge-Kutta scheme
in conjunction with accelerating techniques : local time stepping, residual
smoothing and Full-Approximation-Storage (FAS) multigrid. These equations
are discretized using finite volumes and a cell-centered scheme with artificial
CAD-Centric Framework for Aero-mechanical Optimization 863
dissipation. A very low level of artificial viscosity is ensured by eigenvalue

scaling. The eddy-viscosity hypothesis is used to account for the effect of
turbulence. The algebraic turbulence model based on the two-layer mixing
length model of Baldwin and Lomax (1978) is used to model the turbulence.
4 The Counter-Rotating Fan Test Case

The preliminary design of the fan was performed in [5]. The multi-stage com-
putations are performed using the mixing plane approach. An H-grid type
is used to mesh both rotor blade passage geometries. The dimensions of the
mesh are 162 × 65 × 57 for the first rotor and 145 × 65 × 57 for the second
rotor. The first cell size at the wall is 10−4 times the axial chord along the
blade walls, while it is 5.0 10−4 times the spanwise length along the endwalls.
This leads to a y + of the order of 1 along the blade walls and of the order of
5 along the endwalls. The CFD flow solver is run for 500 cycles, leading to a
decrease of the residuals of 5 orders of magnitude. The Reynolds number is
5 ∗ 106 based on the blade chord length of the rotor 1 hub section.
The inlet total temperature and total pressure are imposed as inlet bound-
ary conditions together with axial flows. At the outlet boundary, the static
pressure is imposed at the hub and the exit profile of static pressure is ob-
tained by the radial equilibrium assumption. Three computations with three
different outlet static pressure values are performed. These static pressure val-
ues are imposed as the ratio between the outlet static pressure to the inlet
total pressure. The values are 2.2, 2.4 and 2.6. The peak adiabatic efficiency
calculated by the 3D viscous turbulent flow solver is 84.5% and the peak total
pressure ratio is 3.16.
The structural mechanics computations are performed with the Samcef
software. The first 20 dynamic vibration modes are computed for both the
first and second rotors. The maximum von Mises stresses are 478 MPa and
207 MPa in rotor 1 and 2 respectively.
The three-dimensional geometry is parameterized using 5 two-dimensional
sections equally distributed from hub to tip. The two-dimensional sections are
parameterized using B-Spline curves. 35 design parameters are used for each
rotor blade which define the modifications of the blade geometry with respect
to the original blade geometry. Therefore, 70 design variables are used to
parameterize the counter-rotating fan geometry.
The Objective Function
The aim of this optimization is to find the optimal geometry for 3 different
aerodynamic operating conditions : one at choked mass flow, one near effi-
ciency peak and one near the stall flow. The three points are analyzed at the
same blade rotational speed, which is the counter-rotating fan design speed.
These 3 objectives are transformed into a single objective by summing the
individual penalties. Lower limits on the choked mass flow and the pressure
ratio are imposed as aerodynamic constraints. Structural constraints are also
imposed. An upper limit on the static stress due to centrifugal forces is im-
posed at 450MPa. Then a set of constraints are imposed on the frequencies
of the blade vibration modes. The rotational frequency of the first fan blade
is 72.765 Hz while this frequency is 65Hz for the second fan blade. These fre-
quencies are the first excitation frequencies and are called the 1N excitation
frequencies. The other excitation frequencies are the multiples of these 1N
frequencies. The excitation frequencies controlled during the optimization are
1N, 2N, 3N, 4N, 6N, 8N, and 10N respectively. These frequencies are used in
the constraints as forbidden ranges.
CAD-free Optimization
A database with 70 geometries and 70 computed performances has first been

generated. As 3 CFD computations are performed per geometry, this leads to
210 three-dimensional CFD computations. Among these 70 geometries only 14
were computed successfully to build the initial database. This initial database
contains the design variables, the objectives and the constraints (aerodynamic
and mechanical) as well as a switch indicating the success or failure of the
simulation. Several criteria are used to check the validity or success of the
simulation along with the convergence level of the finite volume residuals.
The convergence history is shown in figure 1. After 36 design cycles the
objective function has already been significantly improved. One design cycle
requires three detailed CFD simulations at the three operating points. The
performance curve is shown in figure 2. The efficiency is improved by 2.5 %
along the whole performance curve. Figures 3 and 4 highlights the geometry
variations for the mid-span section of the rotors 1 and 2. The adiabatic
efficiency curves are shown in Figure 2. The three points starting from the
choked operating curve extremity are the 3 operating points used for the
optimization. For the optimized geometry the choked mass flow is increased
Fig. 1. Convergence history Fig. 2. Operating curve

CAD-Centric Framework for Aero-mechanical Optimization 865
Fig. 3. Mid-span geometry - rotor 1 Fig. 4. Mid-span geometry - rotor 2
from 403 kg/s to 406.8 kg/s. The peak adiabatic efficiency point is increased
by 1.5% from 84.7 to 86.7%. The static stresses on the optimized geometries
are 483 MPa for the rotor 1 and 250 MPa for the second rotor.
CAD-Based Optimization
The original counter-rotating fan geometry has been imported into CATIA
V5. For this proof-of-concept test case only the rotor 1 tip section was modi-
fied during the optimization, leading to 7 design variables. Then a DOE with
21 geometries has been run, among which 14 geometries were computed suc-
cessfully. The objective is to maximize the sum of the stage efficiency at 3
operating conditions. A lower bound constraint is imposed on the pressure ra-
tio and choked mass flow. For this test case a constraint on the static stresses is
imposed while no constraint is imposed on the dynamic vibration frequencies.
The convergence history of the optimization process presented in figure
5 shows a rapid convergence to a better geometry in 14 design cycles. The
efficiency operating curve presented in figure 2 shows an efficiency improve-
ment of 0.5% and also a choked mass flow increase. This is a rather important
improvement considering the number of design variables. The geometry mod-
ifications are presented in figure 6.
This paper demonstrates the capabilities of a multi-disciplinary design system

for turbomachinery blades. It is based on a fast optimizer capable of finding
a good improvement in a reduced number of function evaluations as well as
being capable of dealing with large number of design variables and constraints
. The elapsed time to reach the optimum is further reduced by both the
parallelization of the initial database generation and by parallelization of the
objective computations. This resulted in a turn around time of about 2 weeks
to optimize a complex counter-rotating geometry.
Fig. 5. Convergence history (CAD-based) Fig. 6. Tip geometry - rotor 1
Acknowledgment.
This research was supported by the Walloon Region and the ERDF European
funds under contract no EP1A122030000102.
References
1. A. Arnone, M. Marconcini, and A. Scotti Del Greco. Numerical investigation of
three-dimensional clocking effects in a low pressure turbine. ASME 2003-GT-
38414, 2003.
2. R. Haimes and G. Follen. Computational analysis programming interface. Pro-
ceedings of the 6th International Conference on Numerical Grid Generation in
Computational Field Simulations, University of Greenwich, June 1998.
3. P. Hajela. Nongradient methods in multidisciplinary design optimization - status
and potential. Journal of Aircraft, pages 255–265, January-February 1999.
4. Yongsheng Lian. Multi-objective optimisation using coupled response surface
model and evolutionary algorithm. AIAA paper 2004-4323, 2004.
5. A. Merchant, J L. Kerrebrock, and A H Epstein. Compressors with aspirated
flow control and counter-rotation. AIAA-2004-2514.
6. S. Pierret. Multi-objective and multi-disciplinary optimization of three-
dimensional turbomachinery blades. 6th World Congresses of Structural and
Multidisciplinary Optimization, 30 May - 03 June 2005, Rio de Janeiro, Brazil.
7. S. Pierret. Multi-objective optimization of three-dimensional turbomachinery
blades. Int. Conf. on Computational Methods for Coupled Problems in Science
and Engineering, 25-28 May 2005, Santorini Island, Greece.
8. S. Pierret, H. Kato, R. Coelho, and A. Merchant. Multi-objective and multi-
disciplinary shape optimization. Evolutionary and Deterministic Methods for
Design, Optimization and Control with Applications to Industrial and Societal
Problems 2005.
Aeroelastic Solutions for Viscous Flows using
the Time Accurate and Non-Linear Frequency
Domain Methods
F. Kachra, S. Nadarajah, and C. Tatossian
Department of Mechanical Engineering, McGill University, 688 Sherbrooke St.

West, # 711, Montreal, QC H3A 2S6 Canada. siva.nadarajah@mcgill.ca
1 Introduction
Most computationally intensive problems in aerodynamics and acoustics are

inherently unsteady. The ability to accurately model the time evolution of
the flow can be computationally demanding, especially when dealing with 3D
problems. As a result, there has been much effort focused on the develop-
ment of efficient and practical alternatives to the study of unsteady periodic
problems.
Examples of engineering applications where the flow is unsteady yet pe-
riodic include flows resulting from helicopter rotor motion, turbomachinery
blades and aerodynamic flutter. Linearized frequency domain and determin-
istic stress methods [1] are examples of periodic methods that are widely used
in industry. Unfortunately, the inability of these methods to accurately model
the solution becomes evident for systems that contain strong nonlinearities.
The Harmonic Balance technique, a pseudo-spectral approach initially pro-
posed by Hall [3] and later modified by McMullen [6], demonstarted that a
temporal resolution of only three time steps per period are needed using the
NLFD method versus the 36 with a backward difference formulation of the
time derivative [2] to accurately model an oscillating airfoil in inviscid flow.
In order to further validate the NLFD method’s ability to efficiently com-
pute solutions to unsteady periodic problems subject to strong nonlinearities,
it will be used to compute flutter boundaries at different speed indices Vf
in the transonic regime. Time accurate solutions by Alonso[2] and Willcox[7]
will serve as our numerical benchmark when determining the overall compu-
tational savings one can achieve by using the NLFD method.

868 F. Kachra, S. Nadarajah, and C. Tatossian
2 Non-Linear Frequency Domain Equations and

Aeroelastic Model
The semi-discrete form of the governing equations can be written as
∂w
V + R = 0. (1)
∂t
The derivation of the NLFD method starts with the semi-discrete form of the
governing equations, and assumes that the solution w and spatial operator
P N2 −1 ikt
R can be represented by separate Fourier series: w = k=− N ŵk e , R=
2
P N2 −1 √
k=− N
R̂k eikt where, i = −1. Substitution of the Fourier series represen-
2
tation into equation (1) yields the unsteady residual, Rˆ∗ k = i kV ŵk + R̂k . The
nonlinearity of the unsteady residual stems from the spatial operator which
can be computed using a Fast Fourier Transform (FFT) algorithm. Consistent
with the time accurate approach, a pseudo-time derivative can be added, and
a time-stepping scheme can be employed to numerically integrate the resulting
equations; V ∂∂τ ŵk
+ Rˆ∗ k = 0. In the NLFD case, an unsteady residual exists
for each wavenumber used in the solution and the pseudo-time derivative acts
as a gradient to drive the absolute value of all of these components to zero
simultaneously.
The governing coupled second-order structural equations of motion can be
written as
mḧ + Sα α̈ + Kh h = −L
Sα ḧ + Iα α̈ + Kα α = Mea
L is the lift (positive up) and Mea is the moment about the elastic axis (pos-
itive nose up). Sα , Iα , Kh , and Kα are the static imbalance, moment of
inertia, bending stiffness and torsional stiffness respectively. h is the plunging
coordinate (positive down) and α is the angle of attack (clockwise positive).
Given the periodic nature of the problem, both the plunging displacement
h and angle of attack α can be written in spectral forms. After taking advan-
tage of the orthogonality of the Fourier terms and substituting the spectral
representations into our structural equations yields a separate equation for
each wavenumber k in the solution
(Kh − mk 2 )ĥk − (Sα k 2 )α̂k = −L̂k

−(Sα k 2 )ĥk + (Kα − Iα k 2 )α̂k = M̂eak .
After solving the simple system above for ĥk and α̂k for each wavenumber we
can use an inverse FFT to determine our displacements at k equally spaced
points in our period. The displacement velocities ḣ and α̇, are computed using
a trigonometric interpolation of the displacements followed by a backward
difference of adjacent points.
Time Accurate and Non-Linear Frequency Domain Methods 869
3 Results
ωc
The reduced frequency and speed index are defined as kc = 2V ∞
and Vf =
V∞
bwα µ respectively. Structural constants used are similar to that of Isogai[4]:
√
xα = 1.8, rα2 = 3.48, a = −2.0, wh = 100 rad rad

sec , wα = 100 sec , and µ = 60.
All numerical simulations are performed using a structured 512x64 mesh of
a NACA64A010 airfoil section. Figure 1 illustrates the convergence of the
viscous NLFD flow solver for various number of time steps per period. The
five time steps per period case seems to exhibit a slightly higher convergence
rate. All modes for each case converge at the same rate.
In Fig. 2 a comparison of the lift hysteresis loop for various cases against
the experimental data are shown. As the airfoil oscillates at a small angle
of attack, the shock wave moves back and forth about a mean location and
is closely sinusoidal and lags the airfoil motion. This lag is evident in the
lift hysteresis loop where the maximum lift does not occur at the maximum
angle of attack. The nonlinear behavior of inviscid unsteady transonic flows
is primarily due to the movement of the shock and this is evident in Fig. 2.
However, the width of the loop is not captured as accurately. The addition
of the viscous terms, provides a better capture of the phase lag, providing
a more accurate lift coefficient at the maximum angle of attack and results
provide a better match to the experimental data over the entire range of angle
of attack. Figure 3 shows the variation of the drag coefficient versus angle of
attack. The one harmonic case does not provide an accurate representation
of the variation of drag versus angle of attack. At least two harmonics are
required to accurately capture the variation and at least 4 is required to get
a converged loop. In the case of pitching moment as shown in Fig. 4 at least
three harmonics are needed. Figure 5 illustrates the drag versus lift coefficient
for various number of modes.
Figures 6 through 9 illustrates the plunge and angle of attack displace-
ments as well as their respective first derivatives for various number of modes.
The solutions represents the second period in the simulation for a neutral
response at the freestream Mach number of 0.825. The figures illustrate that
one mode is sufficient to accurately compute the displacements. Nevertheless,
the close-up views for each figure determines that with modes provides for
higher accuracy.
Figure 10 demonstrates a damped response at a Mach number of 0.825 and
reduced frequency of 0.1932. The amplitude of both the plunge and rotational
displacements reduce over a number of periods. In Fig. 11 a neutral response is
illustrated. Figure 12 illustrates an unstable response at a reduced frequency
of 0.1916. It was observed that the viscous simulations had a tendency to
diverge rapidly as the reduced frequency was increased above the neutral
response compared to the inviscid simulations[5].
Engineers are primarily concerned with periodic flow solutions only once
they have reached steady state. Unfortunately, when using a time accurate
approach, most of the computational cost is spent resolving the initial decay
of transients. The NLFD code is able to resolve the solution in as few as five
periods, resulting in a substantial savings. The number of periods needed to
reach a periodic steady state solution may vary for the time accurate method.
However, it remains clear that the NLFD method still offers tremendous sav-
ings due to the reduced number of samples per period needed to resolve the
periodic steady state solution.
An important distinction in the NLFD results (refer to Figs. 10 through
12) is the discontinuity between the first and second periods in the NLFD
solution. It is important for the reader to understand that we are solving for
the periodic steady state solution immediately in the NLFD case, bypass-
ing all the transients present in the time accurate solution. At steady state,
the unsteady motion of the airfoil prescribes non-zero loads at zero angle of
attack which, when input in the structural equations, would yield non-zero
displacements.
Conclusion
The viscous NLFD approach is able to produce accurate lift hysteresis loops
with just three time steps per period or one harmonic mode in addition to the
zeroth mode. Aeroelastic solutions were computed at M∞ = 0.825. Stable,
neutral, and unstable responses were obtained with just three time steps per
period. The viscous NLFD approach provides a cost effective aeroelastic so-
lution when compared to time accurate methods that require at least 24 time
steps per period. Hence, we can conclude that the NLFD method is indeed an
efficient and accurate alternative to the study of unsteady periodic problems.
References
1. J. J. Adamczyk. Model equation for simulating flows in multistage turboma-
chinery. Technical report, NASA Technical Memorandum 86869., Nov 1984.
2. J. J. Alonso and A. Jameson. Fully-implicit time-marching aeroelastic solu-
tions. AIAA paper 94-0056, AIAA 32nd Aerospace Sciences Meeting and Ex-
hibit, Reno, NV, January 1994.
3. K. C. Hall, J. P. Thomas, and W. S. Clark. Computation of unsteady non-
linear flows in cascades using a harmonic balance technique. Technical report,
9th International Symposium on Unsteady Aerodynamics, Aeroacoustics and
Aeroelasticity of Turbomachines, Lyon, France, September 2000.
4. K. Isogai. On the transonic-dip mechanism of flutter of a sweptback wing. AIAA
Journal, 17(7):793–795, July 1979.
5. F. Kachra and S. Nadarajah. Aeroelastic solutions using the time accurate and
non-linear frequency domain methods. AIAA, 44th Aerospace Sc. Meeting, 2006.
6. Matthew McMullen. The Application of Non-Linear Frequency Domain Methods
to the Euler and Navier-Stokes Equations. Ph.d. dissertation, Department of
Aeronautics and Astronautics, Stanford University, Stanford, CA, March 2003.
7. K. Willcox and J. Peraire. Aeroelastic computations in the time domain using
unstructured meshes. Int. J. for Num. Meth. in Engineering, 40:2413–2431,
1997.
Time Accurate and Non-Linear Frequency Domain Methods 871
0.1
NLFD−3 0−Mode
0 0.08
10 NLFD−3 1−Mode
NLFD−5 0−Mode 0.06
−2
10
NLFD−5 1−Mode 0.04
NLFD−5 2−Mode
Lift Coefficient
−4
log(Residual)
10 0.02
NLFD−7 0−Mode
0
10
−6 NLFD−7 1−Mode
Time Accurate, 23 Time Steps
NLFD−7 2−Mode −0.02
NLFD 1−Harmonic Inviscid
−8
10 NLFD−7 3−Mode −0.04
Experimental
−10 −0.06 NLFD 1−Harmonic Viscous
10
NLFD 2−Harmonic Viscous
−0.08 NLFD 3−Harmonic Viscous
−12 NLFD 4−Harmonic Viscous
10
−0.1
200 400 600 800 1000 1200 1400 1600 1800 2000 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Multigrid Cycles Angle of Attack
Fig. 1. NLFD Flow Solver Conver- Fig. 2. Comparison of Lift Hysteresis

gence for Various Time Steps per Pe- to NACA 64A010 CT6 Experimental
riod Data −3
−3
x 10 x 10
9.6
NLFD 2−Harmonic Viscous 3
9.4
9.2
NLFD 5−Harmonic Viscous 2
1
Drag Coefficient
Lift Coefficient
8.8
8.6 NLFD 2−Harmonic Viscous

0
8.4
−1
8 −2
7.8
−3
7.6
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Angle of Attack Angle of Attack
Fig. 3. Comparison of Variation of Fig. 4. Comparison of Variation of

Drag Coefficient versus Angle of At- Pitching Moment Coefficient versus
tack for Various Time Steps per Pe- Angle of Attack for Various Time
riod Steps per Period
−3
x 10
9.6
0.025 1 Harmonic
2 Harmonics
NLFD 3−Harmonic Viscous 3 Harmonics
0.02
0.015
9
0.01
Drag Coefficient
8.8
0.005
8.6 0.022
h/b
0
0.021
8.4
−0.005 0.02
8.2 0.019
−0.01
0.018
8
−0.015
0.017
7.8
−0.02 0.016
7.6 19.8 20 20.2 20.4

−0.025
−0.08 −0.06 −0.04 −0.02 0 0.02 0.04 0.06 0.08 18 20 22 24 26 28 30 32

Lift Coefficient Dimensional Time
Fig. 5. Comparison of Variation of Fig. 6. Non-Dimensional Plunging

Lift Coefficient versus Drag Coeffi- Displacement for Various Time Steps
cient for Various Time Steps per Pe- per Period
riod
−3
x 10
5 −4
0.015 x 10
1 Harmonic 4
2 Harmonics 4
3 Harmonics 2
0.01 0
3
−2
2 −4
0.005 −6
Plung Speed
1 −8
Angle of Attack
−10
0 −12
0
0.0135
22.2 22.4 22.6 22.8
0.013 −1
−0.005 0.0125
−2
0.012
1 Harmonic
0.0115 −3
2 Harmonics
−0.01
0.011 3 Harmonics
−4
20.2 20.4 20.6 20.8 21
−0.015 −5
18 20 22 24 26 28 30 32 18 20 22 24 26 28 30 32
Dimensional Time Dimensional Time
Fig. 7. Angle of Attack Displacement Fig. 8. Non-Dimensional Plung

for Various Time Steps per Period Speed for Various Time Steps per Pe-
riod
−3
x 10
6
1 Harmonic 0.025
h/b
!
2 Harmonics
4 0.02
3 Harmonics
0.015
AOA and Plung Displacements
2 0.01
Rotational Speed
−4
x 10
20 0.005
0
15 0
10 −0.005
−2
5 −0.01
−0.015
0
−4
−0.02
−5
22 22.5 23
−0.025
−6
16 18 20 22 24 26 28 30 32 34 0 50 100 150 200 250 300
Fig. 9. Rotational Speed for Various Fig. 10. Stable Response, M∞ =

Time Steps per Period 0.825, ωr = 0.1932.
0.03
0.03
0.02
0.02
0.01
0.01
0 0
−0.01
−0.01
−0.02
−0.02 h/b
h/b
!
! −0.03
−0.03
0 20 40 60 80 100 120 140 0 20 40 60 80 100
Fig. 11. Neutral Response, M∞ = Fig. 12. Unstable Response, M∞ =

0.825, ωr = 0.1921. 0.825, ωr = 0.1916.
Role of Interpolation in Airflow Induced
Vibration in Hard Disk Drive Enclosures
S. Ali1 and M. Damodaran2

1
Singapore-MIT Alliance, Nanyang Technological University 50 Nanyang Avenue,
Singapore 639798, shidrati@ntu.edu.sg
2
School of Mechanical and Aerospace Engineering, Nanyang Technological
University 50 Nanyang Avenue, Singapore 639798, mdamodaran@ntu.edu.sg
Summary. Fluid-structure interaction is vital for the study of airflow induced vi-
brations, a cause of hard disk drive failure. With the loose fluid-structure coupling,
there is a need for an accurate exchange of information at the fluid-structure in-
terface. This work looks at two variations of the Shepard’s interpolation method to
assess their accuracy in such an exchange and their application in simulating the
airflow induced vibration on the actuator arm of a small one-inch form factor.
1 Introduction
As airflow induced vibrations is one of the primary causes of hard disk drive
(HDD) failure, it is important to understand the airflow characteristics and
fluid-structure interactions for the optimal design of the HDD. While re-
searchers [10] have performed experiments on simple rotating disks within a
casing, it still remains a challenge to perform experiments on the actual hard
disk drives. To this end, simulation is an efficient and economical tool in the
early design stages of the HDD. Coupling a Computational Fluid Dynamics
(CFD) code and a computational structural dynamics (CSD)code to address
the fluid-structure interaction problem, as studied by [8, 2], assumes that the
accuracy of the information exchanged depends on the accuracy of the inter-
polation algorithm used. In the present study, Shepard’s interpolation method
[7] is used in conjunction with several neighbouring point selection techniques
to evaluate the accuracy of these techniques in exchanging the information
for the airflow induced vibration studies for a one-inch HDD.
2 Models and Algorithms

Based on the studies conducted by Goh et. al. [6], the airflow within the one-
inch HDD as shown in Figure 1(a) can be modeled by the incompressible

874 S. Ali and M. Damodaran
circumradius, R c
circumcenter, x 0 circumradius, R c
circumcenter, x 0
Z X
Y
(a) (b) (c) (d)
Fig. 1. (a) HDD enclosure configuration and construction of the (b) local region of
influence via (c) circumsphere and (d) circumcirle methods.
Navier-Stokes equations with no-slip boundary conditions while the struc-

tural deformation is assumed to be elastic. Only pressure forces are imposed
on the structural domain at the interface since the viscous forces on the struc-
ture are assumed to be much smaller than the pressure forces. The current
fluid-structure interaction considers only a one-way coupling from the CFD
(Fluent [3]) to the CSD solver (Abaqus [1]) due to the very small deflections
of the structures with the exchange of information facilitated by Shepard’s
interpolation method [7].
Shepard’s interpolation method [7], also known as the inverse distance
weighted method, can be expressed as F (x) = N
P PN
k=1 wk (x)fk / k=1 wk (x) for
a given scattered data, f (xk ) and weights, wk (x) = (dk ) −n
= kx − xk k−n
2 ,
where usually n = 2, which takes into consideration the influence of all of the
data. An improvement to the weighting functions could be found by selecting
nearby data points, thus limiting the influence of the scattered data to a local
subset of data points [7, 4] where the weights should be dependent on a local
radius, Rw centered at x as shown in Figure 1(b).
In the present study, two techniques for obtaining Rw and the variations of
the weighting function based on the information on the non-matching fluid and
structural meshes are proposed. The techniques assume that the meshes are
tetrahedral meshes and estimate the associated circumspheres and circumcir-
cles. Similar methods could be derived for hexagonal or wedge elements. The
first technique constructs a circumsphere for each element of the coarser mesh
as shown in Figure 1(b) and the second technique constructs a circumcircle for
each surface element of the coarser mesh as shown in Figure 1(c). The choice
of the coarser mesh is to ensure that the local radius obtained will be large
enough to include the nodes of both meshes. In the construction of circum-
sphere, the circumcenter, x0 , and the circumradius Rc could be obtained using
the equations given in the MathWorld website [9] utilizing the four nodes of
the tetrahedral elements. For the circumcircle, x0 is found by the intersection
of the three perpendicular bisectors of the edges of each triangular face. The
circumradius, Rc , is then obtained as Rc = kx0 − xi k2 where i denotes one
of the three nodal points of the triangular face element. Having obtained the
circumcenter and circumradius for each element, the new weights for Shep-
Interpolation for Airflow Induced Vibration 875
ard’s interpolation for any node, x, within the radius Rc of each coarse mesh
element can be established as follows:

 G : dk ≤ ε
w(x) = (dk )−n : ε < dk ≤ Rc (1)
0 : dk > R c

for some big number G and small tolerance, ε. The first criterion for Equation
1 is selected to avoid discontuinity of w(x) t xk .
The HDD enclosure in Figure 1 consists of a single disk rotating at 3600 rpm,
a latch, and an actuator arm with sliders. The material used for the actuator
arm is aluminium which has a density of 2700 kg/m3, Young’s Modulus of 70
GPa and a Poisson ratio of 0.33.
In order to compare the performance of the various weighting functions,
two test functions have been adapted from Gantovnik et. al. [5], i.e.
f1 (x) = 1 − 2/3(|x1 − 0.001| + |x2 − 0.02| + |x3 − 0.003|) (2)
 3
 X
2/3(x + x + x ) , xi ≤ 0.03

1 2 3



f2 (x) = i=1 (3)
 X3
−2/3(x + x + x ) + 0.05 , xi > 0.03



 1 2 3
i=1
For assessing accuracy, the approximation error is characterized by the maxi-

mum relative error, emax , and the mean relative error, ē. The errors of Shep-
ard’s original method [7], Franke-Nielson’s method [4], the circumcircle and
circumsphere methods are tabulated in Table 1. Based on Table 1, the cir-
cumcircle and circumsphere techniques seem to out-perform Shepard’s and
Franke-Nielson methods in their interpolation results when comparing the
mean relative error for the given test functions.
Table 1. Interpolation errors for the test functions (%)

Test functions Errors Shepard Franke-Nielson Circumsphere Circumcircle
1 emax 0.8 0.4 0.05 0.04
ē 0.3 0.1 0.005 0.005
2 emax 26.4 28.3 33.3 33.3
ē 6.6 2.0 0.2 0.2
Modal analysis has been performed on the actuator arm to compute its
natural frequencies prior to the fluid-structure interaction. Figure 2(a) and
(b) show the first bending and sway modes contributing to the on-track and
Fig. 2. First (a) bending (mode 1 = 2.066kHz) and (b) sway modes (mode 5 =
17.917kHz) and the initial flow characteristics (c) Gauge static pressure and (d)
Velocity Vectors in the vicinity of the sliders
off-track displacements respectively. The on-track displacement is the displace-

ment along the plane in Figure 1(a) (in green) while the off-track displacement
is the displacement perpendicular to the plane.
First the steady flow field corresponding to a fixed disk rotational speed in
the HDD enclosure are computed and used as the initial flow conditions for the
study of the unsteady flow field associated with fluid structure interactions.
Figure 2(c) and (d) show that the highest and lowest pressures are located
on these slider surfaces as well as the formation of several recirculation flow
regions in the vicinity of the sliders. Random fluctuations are then added to
the steady velocity field to evaluate the response of the CFD and CSD models.
A global time step size of 50 µs based on flow physics consideration is used to
advance the solution to model the unsteady flow, utilizing Shepard’s method,
the circumcircle and circumsphere techniques to transfer information from the
CFD model to the CSD model.
The computed time variation of the computed on-track and off-track dis-
placements of the lower slider based on the three interpolation methods for the
first 10 ms are shown in Figure 3(a) and (b) respectively. Figure 3 shows that
the displacement amplitudes from the circumcircle and circumsphere tech-
niques are much larger than those from the Shepard’s method. The off-track
displacement amplitudes of the circumcircle and circumsphere techniques are
similar as observed in Figure 3(a) while the on-track displacement amplitudes
of the two techniques differ by a factor of ten as shown in Figure 3(b). The
larger displacement amplitudes could be attributed to the ability of these two
techniques in capturing the actual pressure distribution on the actuator arm.
Table 2. Percentage mean relative errors of pressure

Time (ms) 1 2 3 4 5 6 7 8 9 10
Circumsphere 2.92 3.16 2.49 9.22 5.75 4.53 2.73 2.64 2.45 3.94
Circumcircle 1.06 1.14 1.08 4.50 1.12 2.09 1.11 1.14 1.07 1.85
Interpolation for Airflow Induced Vibration 877
−9 −10
10 10 −150
on−track
off−track
−10 −11 −160
10 10
−170
−11 −12
10 10
displacement (m)
displacement (m)
circumcircle
Decibel [dB]
circumsphere −180
−12 −13
10 10 Shepard
−190
circumcircle
−13 −14
10 circumsphere 10
Shepard −200
−14 −15
10 10
−210
−15 −16
10 10 −220
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 0 5 10 15 20 25
time (s) x 10
−3
time (s) −3 frequency [kHz]
x 10
(a) on-track (b) off-track (c) power spectrum
Fig. 3. (a)On-track and (b)Off-track displacement amplitudes (in meters) of the

lower slider and (c) the corresponding vibration power spectrum using circumcircle
technique.
Based on the displacement amplitudes of Figure 3(a) and (b), the accu-
racy of the interpolation results affects the on-track displacements more than
the off-track displacements. The mean relative errors shown in Table 2 clearly
indicate that the results from the circumcircle technique are relatively more
accurate than those from the circumsphere technique. This difference in accu-
racy could be attributed to the size of the coarser mesh elements in relation
to the thickness of the actuator arm. The inaccuracy of the interpolation re-
sults of circumsphere technique arises from the inclusion of pressure values
from both the upper and lower surfaces of the arm as opposed to those of
the circumcircle technique due to the size of the coarse mesh spanning both
surfaces. However, the mean relative interpolation errors shown in Table 2
for both techniques based on pressures from the nearest neighbouring fluid
nodal points are less than ten percent of the actual pressure values of these
neighbouring points. These figures give some assurance on the accuracy of the
two techniques in capturing the pressure distribution.
Using Discrete Fourier transformation on the displacements determined
from Abaqus for the circumcircle technique, the frequency analysis is plotted
in Figure 3(c). Though offsetted by about 30 dB, the off-track and on-track
power spectrum patterns are similar with several peaks. The first peak could
be found at 2 kHz which is close to the natural frequency of the actuator as
shown in Figure 2(a) and similarly, the peak at 18 kHz is close to the natural
frequency of the fifth mode as shown in Figure 2(b). These close proximities
to the natural frequencies are likely to lead to the problem of vibrational
resonance and affect the integrity of data being written or retrieved by the
actuator arm. However, this analysis is still incomplete due to the lack of
experimental data validation and the fact that the slider to disk gap used for
the present study is around 9 µm which is much larger than the nano-sized
gap of the actual HDD assembly.
Conclusions
The role of interpolation methods in ensuring the accuracy of the information
passed between the CFD and CSD solver is explored in this paper and it is
found that the displacement amplitudes are sensitive to the difference in in-
terpolation results. The two techniques greatly improve on the interpolation
results of the Shepard’s method by enforcing selection rules for the local sub-
set of the data points with the circumcircle technique observed to be more
accurate. As fluid-structure interactions are applied to smaller structures, the
accuracy of the interpolation methods used in the loose-coupling approach
play an increasingly important role in ensuring the accuracy of the results.
Acknowledgements. The authors wish to acknowledge Mr. Elson Goh and

Dr Q.Y. Ng of Seagate Technology International for their valuable assistance.
Part of this work was also based on the work done by Miss Leela Krishna
Coumar, an exchange student from Matmeca, France during her summer in-
ternship at School of Mechanical and Aerospace Engineering, Nanyang Tech-
nological University. This work has been funded partially by Seagate Tech-
nology International.
References
1. ABAQUS Inc.: Abaqus documentation, Version 6.5.4 (2005).
2. Damodaran, M., Ali, S. and Dayanandan, S.: Computational Fluid Dynamics in
the Analysis and Design of Engineered Systems. In: Caughey, D.A. and Hafez,
M.A. (ed) Frontiers of Computational Fluid Dynamics 2006, World Scientific,
37-48, (2005).
3. FLUENT Inc.: Flow simulation software user guide, Fluent Version 6.2.16
(2005).
4. Franke, R., and Nielson, G.: Smooth Interpolation of Large Sets of Scattered
Data. International J. Numerical Methods Engineering, 15, 1691-1704 (1980).
5. Gantovnik, V.B., Gurdal, Z., and Watson, L.T.: Linear Shepard Interpolation
for High Dimensional Piecewise Smooth Functions. AIAA 2004-4486, Proceed-
ings of 10th AIAA/ISSMO Multidiscplinary Analysis and Optimization Con-
ference, Albany, New York, 30 Aug - 1 Sep (2004).
6. Goh, H.J., Damodaran, M. and Ng, Q.Y.: Modeling Airflow and Particle Tra-
jectories near the Head/Disk Interface Region of a Small Form Factor Hard Disk
Drive Enclosure. WTC2005-63630, ASME-STLE World Tribology Conference
III, Washington DC, US, 12- 16 Sep (2005).
7. Shepard D.: A Two-Dimensional Interpolation Function for Irregularly Spaced
Data. Proceedings of the 1968 23rd ACM national conference, 517-524 (1968).
8. Tatewaki, M., Tsuda, N., and Maruyama, T.: A Numerical Simulation of Un-
steady Airflow in HDDs. Fujitsu Sci. Tech. Journal, 37, 2, 227-235 (2001).
9. Weisstein. E.W.: Circumsphere. In: MathWorld–A Wolfram Web Resource.,
CRC Press LLC (1999). http://mathworld.wolfram.com/Circumsphere.html
10. Wu, S.C. and Chen, Y.M.: Phase-Averaged Method Applied to Periodic Flow
between Shrouded Co-Rotating Disks. International Journal of Rotating Ma-
chinery, 8, 2, 413-421 (2002).
Computation of Fluid Structure Interaction
with Application to Three-Dimensional
Combustion System
Zinedine Khatir
School of Mechanical and Aerospace Engineering, Queen’s University Belfast

Ashby Building, Stanmillis Road, Belfast BT9 5AG, Northern Ireland, UK.
z.khatir@qub.ac.uk /zkhatir@hotmail.com
Summary. The computation of many fluid-structure interaction (FSI) problems

requires solving simultaneously the coupled fluid and structure equations with an
appropriate set of interface boundary conditions. When loosely-coupled partitioned
procedures are considered in realistic situations where the fluid and structure sub-
problems have non-matching meshes, it is necessary to transfer the fluid pressure and
stress fields at the fluid-structure interface into a structural load, and to convert the
structural motion to the fluid system. For that purpose, an interpolation method-
ology based on volume spline is applied for transferring data between the fluid and
structure grids. Results are presented with application to combustion system.
Key words: Fluid Structure Interaction, CFD-CSD Coupling, 3-Dimensional Mod-

elling, Non-Matching Meshes, Combustion System.
1 Introduction
The numerical simulation of FSI problems occurs in many engineering and sci-
entific applications, ranging from airfoil oscillations [1] or aero-hydrodynamics,
haemodynamcis [2] to blade flutter analysis in turbomachinery [3] and power
generation when designing gas turbine combustors [4, 5] for instance. Un-
like tightly-coupled approaches where the fluid and structural equations are
solved in a single computational domain, loosely-coupled partitioned proce-
dures [6, 7] use independent techniques for the fluid and structure subdomains
and exchange data along the fluid-structure interface. The loosely-coupled ap-
proach can thus take the full advantage of existing, well developed and tested
codes for fluid and structural analysis.
In this article, we are interested in enforcing the interface boundary con-
dition by transferring the fluid loads to the structure, while displacement is
transferred back to the fluid from the structure [8]. Various methods have been
developed for the problem of data transfer with non-matching meshes [9]. A

880 Zinedine Khatir
detailed review of data transfer methods is given in [10], which can be clas-
sified as methods suitable for the transfer of: (i) load only, (ii) displacement
only and (iii) both load and displacement. The volume spline interpolation
method by Hounjet et al [11] used for the data transfer in the present paper
belongs to the third class. After a brief description of the method, the accuracy
of the interpolation is assessed against simple 2D and 3D test cases. We will
then apply the methodology for the application of FSI in three-dimensional
combustion system, and the numerical results show good agreement with ex-
perimental data.
2 Fluid-Structure Coupling
Principle. The evaluation of the structure load induced by the fluid and
the displacement of the CFD grid induced my the structural motion needs
to satisfy the requirement of conservation of work and accuracy. The princi-
ple of virtual work is usually used to ensure conservation of energy. For this
purpose, a linear transformation is sought. If the displacement ∆xa of the
CFD grid is expressed by means of the structural displacement ∆xs using the
transformation matrix [G] such that:
∆xf = [G] ∆xs (1)
then the requirement for conservation leads to a corresponding matrix for the
transformation of the fluid and structural load forces ff and fs
fsT ∆xs = ffT ∆xf = ffT [G] ∆xs and fs = [G]T ff . (2)
Volume Spline Interpolation. Volume spline interpolation is a very simple
method to interpolate non-uniformly spaced data. The method belongs to the
category of scattered interpolation methods using radial basis functions [12,
13] and is well suited for transferring both load and displacement. Consider
a set of unordered support points xi , i = 1 . . . N with functions values {fi ∈
R}, = 1 . . . N , and the interpolation function defined as:
N
X
f (x) = α0 + αi Ri (x) (3)
i=1
with Ri (x) =k x − xi k. The coefficients αi , i = 0 . . . N are computed from

PN
the conditions f (xi ) = fi together with the side condition i=1 αi = 0.
Let us recall that the Laplacian of a scalar function ϕ(R) with R =k x k
2 ∂ϕ
obeys in 3D the relation ∆ϕ = R1 ∂R ∂
2 (R ∂R ). For ϕ(R) = R we have ∆ϕ = R
2
and thus ∆∆ϕ = 0. Accordingly, the volume spline interpolation function

f (x) verifies the biharmonic equation everywhere except at the support points.
Therefore, the interpolation function f (x) will not possess new local extrema
outside the support points [14]. This important property makes the volume
spline method suitable, as it will reduce the risk of the occurrence of grid
folding in the deformed mesh.
3D Fluid Structure Interaction Computation 881
The transformation matrix [G] as defined in Eq. (1) is computed as follows.

Let us first define [G]∗ = [A][C]−1 where matrices [A] and [C] are constructed
as:
 
0 1 1 ... 1  a a a

 1 E11s s s 1 E11 E12 . . . E1N s
E 12 . . . E 1N s

 1 E21a a a
 s s s E22 . . . E2N s

[C] =  1 E21 E22 . . . E2N s  and [A] =  . . (4)
   
. . .
 .. .. .. . . ..   .. .. .. .. .. 
. . . . .  a a a
s s s 1 E a
N 1 E N 2 a . . . EN N a s
1 EN s 1 EN s 2 . . . EN s N s
where N a and N s are the number of fluid and structural elements respectively.
The functions E a and E s are given by
s
p
Elm = (xl − xm )2 + (yl − ym )2 + (zl − zm )2
p (5)
a
Elm = (xal − xm )2 + (yla − ym )2 + (zla − zm )2
with (x, y, z) and (xa , y a , z a ) are the coordinates of the structural points and
fluid points respectively. The coupling matrix [G] is then obtained by deleting
the first column of [G]∗ .
3 Numerical Results
2D Simple Test. The interface between CFD and CSD consists of the unit
circle centered at the origin. As seen in Fig. 1-a the two meshes are non-
matching. The aerodynamic input field points is obtained from 30 discretized
points on the unit circle, whereas the structural input field is taken from 23
discretized points on the circle with radius 1.05. After that, the following
deformation is applied to the aerodynamic input points (xi , yi ) to obtain the
new position (x∗i , yi∗ ):
For i=1 to 30:
∗ ∗
θ = 2π
30 × (i − 1); xi = r cos θ = xi ; yi = r sin θ = yi ;
∗ 3
If xi > 0.8: xi = xi − 0.8 ∗ xi
End
The final solution is depicted in Fig. 1-b after interpolation. The deforma-
tion of the structural mesh is of very good quality. The interpolated solution
is compared with the exact solution. The maximum obtained error is around
0.07 which gives second order precision for this test E ' (δx)2 .
3D Simple Test. Here a random pressure field is applied onto the unstruc-
tured CFD mesh as in Fig. 2. This pressure field is then interpolated onto
the structured CSD mesh. As shown in Fig. 2, the input random pressure
field and the interpolated pressure field on the CSD mesh compare well. The
maximum obtained error is E ' 0.008 which leads to a third order precision
with this kind of discretization.
882 Zinedine Khatir
Fig. 1. Two circular concentric non-matching meshes, (dots) CFD mesh and
(square) CSD mesh configuration: (a) Before deformation and (b) After deformation.
Combustion System Problem. Many combustors exhibit combustion in-

stabilities which constitute significant risks for project developments [15]. The
prediction of combustion stability in a specified burner is the focus of either
experimental [16] or numerical investigations [17]. Of crucial importance for
the operation of the engine combustors is its structural integrity. This may
cause the liner to vibrate excessively and could be dangerous. It is thus impor-
tant to understand transient combustion an its coupling with wall vibration
in typical combustion chamber. A typical experimental setup together with
computational domain used for the numerical simulation is shown if Fig.3
A transient FSI calculation is implemented to assessed volume spline
method for data transfer. The LES solver AVBP (i.e. www.cerfacs.fr/cfd/
CFDWeb.html) simulates the full compressible multi-species Navier-Stokes
equations and provides the pressure loads needed for the FEM structural code
solver CalculiX (i.e. www.calculix.de). The interface non-matching meshes is
shown in Fig.4. The LES data are transferred onto the CSD grid every 0.15 ms
with a total run of 21 ms. Fig. 5 display the computed and measured velocity
(a) (b)
Fig. 2. (a) CFD and (b) CSD non-matching meshes alongside pressure fluctuation.
Fig. 3. (Left) Experimental setup and (Right) Computational domain [5, 17].
Fig. 4. Fluid and structure meshes with non-matching interfaces.
Fig. 5. Experiment ( ), and calcu- Fig. 6. Time evolution of the sum of

lated (− − −) velocity on the liner the CFD ( ) CSD (− − −) loads on
[5, 17]. the vibrating liner.
on the liner. Results compare well. The time evolution of the loads is plotted
in Fig. 6. CFD and CSD loads after transfer coincide. Fig. 7 demonstrate
the energy conservation during the simulation. A mean error of 1.26 × 10−5
and 9.01 × 10−4 is found for the time evolution of the 1st and 2nd moment
respectively. This proves again an accurate data transfer between the 2 non-
matching grids as plotted in Fig. 8.
Conclusion
A volume spline method for the data transfer in transient FSI computation
with non-matching meshes has been presented. Results show good agreement
with analytical and experimental results. Further investigations are still being
Fig. 7. Time evolution of 1st (Top) and 2nd (Bottom) Moment (w.r.t axis x=0.895
m) and relative errors.
884 Zinedine Khatir
Fig. 8. Surface plot of CFD (Left) and CSD (Right) loads at time t=0.485s.
undertaken to provide a better understanding of FSI in combustion systems.
Acknowledgement. This work is carried out within the EC Framework

Marie Curie RTN project FLUISTCOM-Fluid-Structure Interaction for Com-
bustion Systems through contract number MRTN-CT-2003-504183. The au-
thor wishes to thank A.X. Sengissen and P. Mauro from CERFACS - France,
R. Huls and A.K. Pozarlik from Univ. of Twente - Netherlands their constant
help and assistance. He also expresses his gratitude to Dr R. Cooper and Dr
J. Watterson for their help and comments on the paper. The invaluable help
of N. Forsythe at Queen’s Univ. Belfast is greatly appreciated.
References
1. Farhat C., On the three-field formulation & solution of nonlinear
Fluid/Structure Interaction problems, ECCOMAS Course - Advanced Compu-
tational Methods for Fluid-Structure Interaction, Ibiza, Spain, May 3–7, 2006
2. Quarteroni A., Fluid-Structure Interaction: Two Applications and Some
RemarksECCOMAS Course - Advanced Computational Methods for Fluid-
Structure Interaction, Ibiza, Spain, May 3–7 2006
3. Sadeghi M. and Liu F., Coupled fluid-structure simulation for turbomachinery
blade rows, 43rd AIAA Aerospace Sciences Meeting and Exhibit, AIAA 2005–
0018, Reno, NV., 2005
4. Sengissen A.X. , Poinsot T.J. , Van Kampen J.F. and Kok J.B.W, Response of
swirled non-premixed burner to fuel flow rate modulation, Complex Effects in
Large Eddy Simulations, Limassol, Cyprus, September 2005
5. Huls R., Acousto-Elastic Interaction in Combustion Chambers, PhD thesis,
Univ. of Twente, Netherlands, ISBN: 90-365-2336-2, 2006
6. Piperno S., Farhat C., Partioned procedures for the transient solution of coupled
aeroelastic problems - Part II: Energy transfer analysis and three-dimensional
applications, Comput. Methods Appl. Mech. Engng., 190, 3147–3170, 2001
7. Felippa C.A., Park K.C., Farhat C., Partioned analysis of coupled mechanical
systems, Comput. Methods Appl. Mech. Engng., 190, 3247–3270, 2001
8. C. Farhat, M. Lesoinne, P. Le Tallec P., Comput. Methods Appl. Mech. Engng.,
157, 95–114, 1998
9. Jaiman R.K., Jiao X., Geubelle P.H., Loth E., Assessment of conservative load
transfer for fluid-solid interface with non-matching meshes, Int. J. Numer.
Meth. Engng, 64, 2014–2038, 2005
10. Forsythe N., A partitioned approach to fluid-structure interaction for artificial
heart valves, PhD thesis, Queen’s Univ. Belfast, Northern Ireland - UK, to be
submitted, 2006
11. Hounjet M.H.L., Meijer J.J., Evaluation of elastomechanical and aerodynamic

data transfer methods for non-planar configurations in computational aeroelas-
tic analysis, NLR-TP-95690-U, National Aerospace Laboratory NLR, Amster-
dam, The Netherlands, 1995
12. Franke R., Scattered data interpolation: tests of some methods, Math. Comp.,
38(157), 181–200, 1982
13. Schaback R., Wendland H., Characterization and construction of radial basis
functions, in Multivariate Approximation and Applications, Cambridge Univer-
sity Press, N. Dyn, D. Leviatan, D. Levin and A. Pinkus Editors, 2001
14. Cordero-Gracia M., Ribpollés P., Valero E., Gómez M., A volume spline in-
terpolation tool for elastomechanical and aerodynamic data transfer problems,
Proceedings of the IASTED International Conference: Applied Simulation and
Modelling, Rhodes, Greece, June, 265 – 269, 2004
15. Poinsot T., Selle L., LES and acoustic analysis of combustion instability in
gas turbines, PLENARY LECTURE - ECCOMAS Computational Combustion
Symposium, Lisbonne, Portugal, 2005
16. Pozarlik A.K., Kok J.B.W., The interaction of combustion pressure oscilla-
tion and liner vibrations, 13th International Congress on Sound and Vibration
(ICSV13), Vienna, Autria, July 2-6, 2006
17. Sengissen A.X., Simulation aux grandes échelles des instabilités de combustion:
vers le couplage fluid/structure, PhD thesis, TH/CFD/06/30, INP Toulouse,
France, 2006
Numerical Computations of Unsteady
Aerodynamics of Projectiles using an
Unstructured Technique
Jubaraj Sahu
U.S. Army Research Laboratory AMSRD-ARL-WM-BC Aberdeen Proving

Ground, MD 21005-5066, USA, Jubaraj.sahu@us.army.mil
Summary. This paper describes a new multidisciplinary computational study un-

dertaken to compute the flight trajectories and the free flight aerodynamics of pro-
jectiles. Actual flight trajectories are computed using an advanced coupled compu-
tational fluid dynamics (CFD)/rigid body dynamics (RBD) technique. In addition,
our goal is to be able to perform time-accurate multidisciplinary-coupled CFD/RBD
computations for complex guided projectiles with control maneuvers using control
surfaces such as fins/canards. An advanced time-accurate Navier- Stokes compu-
tational technique has been used in CFD to compute the unsteady aerodynamics
associated with the free flight of a finned projectile at supersonic speeds. Computed
positions and orientations of the projectile have been compared with actual data
measured from free flight tests and are found to be generally in good agreement.
Unsteady numerical results obtained from the coupled method show the flow field,
the aerodynamic forces and moments, and the flight trajectories of the projectile. In
addition, computed results have been presented for another complex configuration
with control maneuver using an unstructured Chimera method.
Key words: Unsteady CFD, Multidisciplinary computations.
1 Introduction
Understanding the aerodynamics of projectiles, rockets, and missiles is criti-

cal to the design of stable configurations and contributes significantly to the
overall performance of weapon systems [1-3]. The prediction of aerodynamic
coefficients for these weapon systems is essential in assessing the performance
of new designs. Numerical simulations have the potential of greatly reducing
design costs while providing a detailed understanding of the complex aero-
dynamics associated with each change. Recently, we have made progress in
coupling computational fluid dynamics and flight dynamics to perform re-
quired multidisciplinary simulations for moving body problems. This involves
real-time multidisciplinary-coupled computational fluid dynamics/rigid body
aerodynamics computations for the entire flight trajectory of a complex guided

Unsteady Aerodynamics of Projectiles using an Unstructured Technique 887
projectile system. It can lead to accurate determination of aerodynamics, crit-

ical to the low-cost development of new advanced guided projectiles, rockets,
missiles, and smart munitions.
Improved computer technology and state-of-the-art numerical procedures
now enable solutions to complex, 3-D problems associated with projectile and
missile aerodynamics. In particular our recent focus has been directed at the
development and application of advanced predictive capabilities to compute
unsteady projectile aerodynamics, especially during and after control maneu-
vers. During these maneuvers [4-5], very limited data is available and there
is a lack of knowledge and understanding of the detailed aerodynamics. Ac-
curate numerical modeling of the unsteady aerodynamics has been found to
be challenging and has required the use of time-accurate solutions techniques.
The present work is focused on the coupling of CFD and rigid body dynam-
ics (RBD) techniques for simultaneous prediction of the unsteady free- flight
aerodynamics and the flight trajectory of projectiles. In addition, our goal
is to be able to perform time-accurate multidisciplinary-coupled CFD/RBD
computations for complex guided projectiles with control maneuvers using
control surfaces such as fins/canards.
Multidisciplinary computations can provide detailed fluid dynamic under-
standing of the unsteady aerodynamics processes involving the maneuver-
ing flight of modern guided weapon systems. The advanced CFD capability
used here solves the unsteady Navier-Stokes equations, incorporates unsteady
boundary conditions and a special coupling procedure. The present research
is a big step forward in that it allows vNvirtual fly-outÓ of projectiles on
the supercomputers, and allows numerical prediction of the actual fight paths
of a projectile and all the associated unsteady free flight aerodynamics using
coupled CFD/RBD techniques in an integrated manner.
2 Numerical Procedure
A real-time accurate approach is used in the present work; however, time-

accurate computations require much greater computer resources. The real-
time accurate approach also requires that the six-degrees-of-freedom body
dynamics be computed at each repetition of the fluid flow solver. In three-
dimensional space, a rigid object has six degrees of freedom: three translations
and three rotations. The six-degree-of-freedom code computes linear and an-
gular velocities as well as the orientation of the missile, which are used as
input to the computational fluid dynamics code. In turn, the aerodynamic
forces and moments obtained from the flow solver are used to solve the 6-dof
body dynamics before moving on to the next time step. This procedure allows
one to perform real-time multidisciplinary-coupled computational fluid dy-
namics/rigid body aerodynamics computations for the partial or entire flight
trajectory of a complex guided projectile system.
The CFD capability used here solves the Navier-Stokes equations [6-8] and
incorporates advanced boundary conditions and grid motion capabilities. The
888 Jubaraj Sahu
present numerical study is a big step forward and a direct extension of that
research which now includes numerical simulation of the actual fight paths
of the projectile using coupled CFD/RBD techniques using real-time accu-
rate approach. The complete set of 3-D time-dependent Navier-Stokes equa-
tions is solved in a time-accurate manner for simulations of actual flights. A
commercially available code, CFD++ [6-9], is used for the time-accurate un-
steady CFD simulations. The basic numerical framework in the code contains
unified-grid, unified-physics, and unified-computing features. The user is re-
ferred to these references for details of the basic numerical framework. The
3-D time-dependent Reynolds-averaged Navier-Stokes (RANS) equations are
solved using the finite volume method [7]:
Z I Z
∂
W + [F − G]dA = HdV (1)
∂t V V
where W is the vector of conservative variables, F and G are the inviscid and
viscous flux vectors, respectively, H is the vector of source terms, V is the cell
volume, and A is the surface area of the cell face.
Second-order discretization was used for the flow variables and the tur-
bulent viscosity equation. The turbulence closure is based on topology-
parameter-free formulations. Two- equation [6] and higher order hybrid RANS/
LES [10,11] turbulence models were used for the computation of turbulent
flows. The hybrid RANS/LES approach based on Limited Numerical Scales
(LNS) is well suited to the simulation of unsteady flows and contains no addi-
tional empirical constants beyond those appearing in the original RANS and
LES sub-grid models. With this method a regular RANS-type grid is used
except in isolated flow regions where denser, LES-type mesh is used to re-
solve critical unsteady flow features. The hybrid model transitions smoothly
between an LES calculation and a cubic k − ε model, depending on grid fine-
ness. These models are ideally suited to unstructured book-keeping and mas-
sively parallel processing due to their independence from constraints related
to the placement of boundaries and/or zonal interfaces. For computations of
unsteady flow fields that are of interest here, dual time-stepping as described
below was used to achieve the desired time-accuracy [12]. In addition, the
projectile in the coupled CFD/RBD simulation along with its grid is actually
moved as it flies downrange.
Grid velocity is assigned to each mesh point. This general capability can be
tailored for many specific situations. For example, the grid point velocities can
be specified to correspond to a spinning projectile. In this case, the grid speeds
are assigned as if the grid is attached to the projectile and spinning with it.
Similarly, to account for rigid body dynamics, the grid point velocities can
be set as if the grid is attached to the rigid body with six degrees of freedom
(6 DOF). For the rigid body dynamics, the coupling refers to the interaction
between the aerodynamic forces/moments and the dynamic response of the
projectile/body to these forces and moments. The forces and moments are
computed every CFD time step and transferred to a 6DOF module which
computes the body’s response to the forces and moments. The response is
converted into translational and rotational accelerations that are integrated
to obtain translational and rotational velocities and integrated once more to
obtain linear position and angular orientation. The 6DOF rigid body dynamics
module uses quaternions to define the angular orientations. However, these
are easily translated into Euler angles. From the dynamic response, the grid
point locations and grid point velocities are set. Typically, we begin with
a computation performed in ”steady state mode” with the grid velocities
prescribed to account only for the translational motion component of the
complete set of initial conditions. At this stage we also impose the angular
orientations from the initial conditions. The complete set of initial conditions
includes both translational and rotational velocity components along with
initial position and angular orientation.
3 Results
Time-accurate unsteady numerical computations were performed using Navier-
Stokes and coupled 6-DOF methods to predict the flow field and aerodynamic
coefficients, and the flight paths of a finned projectile at supersonic speeds. In
all cases, full 3D computations were performed and no symmetry was used.
The supersonic projectile modeled in this study is an ogive-cylinder-finned
configuration (see Figure 1). The length of the projectile is 121 mm and the
diameter is 13mm. The ogive nose is 98.6 mm long and the afterbody has a
22.3 mm, 2.5◦ boat-tail. Four fins are located on the back end of the projectile.
Each fin is 22.3 mm long and 10.16 mm thick. An unstructured computational
mesh was generated for this projectile. In general, most of the grid points are
clustered in the boundary-layer and afterbody fin regions. The total number
of grid points is about 4 million for the full grid.
Here, our primary interest is in the development and application of coupled
CFD and RBD techniques for accurate simulation of the free flight aerody-
namics and flight dynamics of the projectile in supersonic flight. The first
step was to obtain the steady state results for this projectile at a given ini-
tial supersonic velocity. Also imposed were the angular orientations at this
stage. Corresponding converged steady state solution was then used as the
starting condition along with the other initial conditions for the computa-
tion of coupled CFD/RBD runs. Numerical computations have been made for
these cases at an initial velocity of 1034 m/s. The simulations were started
a small distance away from the muzzle. The corresponding initial angle of
attack was, α = 4.9◦ and initial spin rate was 2500 rad/s. Figure 2 shows
Fig. 1. Finned Configuration

890 Jubaraj Sahu
Fig. 2. Computed pressure contours Fig. 3. Euler pitch angle vs. x-distance
the computed pressure contours at a given time or at a given location in the

trajectory. It clearly shows the orientation of the body at that instant in time
and the resulting asymmetric flow field due to the body at angle of attack.
The orientation of the projectile of course changes from one instant in time
to another as the projectile flies down range. Figure 3 shows the variation
of the Euler pitch angle with distance traveled. As seen in this figure, both
the amplitude and frequency in the Euler angle variation are predicted very
well by the computed results and match extremely well with the data from
the flight tests. One can also clearly see that the amplitude damps out as the
projectile flies down range i.e. with the increasing x-distance. As shown in
figure 4, similar behavior is observed with the Euler yaw angle and it damps
out with the increasing x-distance. Computed results again compare very well
with the measured data from the flight tests.
The time histories of the pitch and yaw angles are often customarily pre-
sented as a motion plot where the pitch angle is plotted versus the yaw angle
during the flight of the projectile. It represents the path traversed by the nose
of the projectile during the flight trajectory (looking forward from the back
of the projectile). Such motion plots are shown in Figure 5. This figure shows
the comparison of the motion plots obtained both from the numerical simu-
Fig. 4. Euler yaw angle vs. x-distance

Fig. 5. Motion plot (a) computation, (b) flight test
Fig. 6. Unstructured Chimera mesh in the nose region
lations and the 6DOF analysis of the flight results from ARFDAS, software
commonly used for this purpose. Computed results match very well with the
experimental flight test results. The unsteady simulations took thousands of
hours of CPU time on a Xeon PC cluster system running with either 32 to 64
processors.
Another case considered in the study is a complex canard-controlled finned
projectile. Here, the control maneuver is achieved by the two horizontal ca-
nards in the nose section (Figs. 6-8). Unstructured Chimera overlapping grids
were used (see Fig. 6) and solutions have been obtained for several canard
892 Jubaraj Sahu
Fig. 8. Computed surface pressure con-

tours, M = 3.0, α = 0◦
Fig. 7. Computed pressure contours,
M = 3.0, α = 0◦
deflection cases. Figure 7 shows the computed pressure contours at M = 3.0

and α = 0◦ for a canard deflection of 20 deg. Although not shown here, this
produces lift that can be used to obtain increased range. A typical result is
shown in Figure 8 for the canard deflection of 20◦ .
Conclusions
This paper describes a new coupled CFD/RBD computational study un-

dertaken to determine the free flight aerodynamics of projectiles both with
and without control maneuvers. A three- dimensional unsteady Navier Stokes
solver is employed to compute the time-accurate aerodynamics associated with
the free flight of a finned projectile at supersonic velocities. Computed posi-
tions and orientations of the projectile have been compared with actual data
measured from free flight tests and are found to be generally in good agree-
ment. In addition, computed results have been presented for another complex
configuration with control maneuver using an unstructured Chimera method.
This work demonstrates a coupled method to accurately predict the time-
accurate unsteady aerodynamics and the flight trajectories of projectiles at
various speeds. The present CFD/RBD simulations clearly show the capabil-
ity of the coupled approach and form the basis for future multidisciplinary,
time-dependent computations of advanced maneuvering munitions.
References
1. Sahu, J., K. R. Heavey, and E. N. Ferry, ”Computational Fluid Dynamics for
Multiple Projectile Configurations.” Proceedings of the 3rd Overset Composite
Grid and Solution Technology Symposium, Los Alamos, NM, October 1996.
2. Sahu, J., K. R. Heavey, and C. J. Nietubicz, ”Time-Dependent Navier-Stokes
Computations for Submunitions in Relative Motion.” 6th International Sym-
posium on Computational Fluid Dynamics, Lake Tahoe, NV, September 1995.
3. Meakin, R. L., ”Computations of the Unsteady Flow About a Generic
Wing/Pylon/Finned-Store Configuration.” AIAA 92-4568-CP, August 1992.
4. Smith B. L., and A. Glezer, ”The Formation and Evolution of Synthetic Jets.”
Journal of Physics of Fluids, vol. 10, No. 9, September 1998.
5. Davis, S. A. and Glezer, A., ”The Manipulation of Large- and Small-Scales
in Coaxial Jets using Synthetic Jet Actuators”, AIAA Paper No. 2000-0403,
January 2000.
6. Batten, P., U. Goldberg and S. Chakravarthy, ”Sub-grid Turbulence Modeling
for Unsteady Flow with Acoustic Resonance”, AIAA Paper 00-0473, 38th AIAA
Aerospace Sciences Meeting, Reno, NV, January 2000.
7. Peroomian, O., S. Chakravarthy, and U. Goldberg, ”A ’Grid-Transparent’
Methodology for CFD.” AIAA Paper 97-07245, 1997.
8. Peroomian, O., S. Chakravarthy, S. Palaniswamy, and U. Goldberg, ”Conver-
gence Acceleration for Unified-Grid Formulation Using Preconditioned Implicit
Relaxation.” AIAA Paper 98-0116, 1998.
9. Goldberg, U. C., O. Peroomian, and S. Chakravarthy, ”A Wall-Distance-Free
K-E Model With Enhanced Near-Wall Treatment.” ASME Journal of Fluids
Engineering, Vol. 120, pp. 457-462, 1998.
10. Avancha, R. and Pletcher, R. H., ”Large Eddy Simulation of the Turbulent Flow
Past a Backward Facing Step”, AIAA paper No. 2000-0542, January 2000.
11. Arunajatesan, S. and Sinha, N., ”Towards Hybrid LES-RANS Computations
of Cavity Flowfields”, AIAA Paper No. 2000-0401, January 2000.
12. Sahu, J., ”Unsteady CFD Modeling of Aerodynamic Flow Control over a Spin-
ning Body with Synthetic Jet.” AIAA Paper 2004-0747, Reno, NV, 5-8 January
2004.

Computational Fluid Dynamics 2006 Proceedings of The Fourth International Conference On Computational Fluid Dynamics ICCFD4 Ghent Belgium 10 14 J

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Computational Fluid Dynamics 2006 Proceedings of The Fourth International Conference On Computational Fluid Dynamics ICCFD4 Ghent Belgium 10 14 J

Uploaded by

Copyright:

Available Formats

Computational Fluid Dynamics 2006

“This page left intentionally blank.”

ISBN 978-3-540-92778-5 e-ISBN 978-3-540-92779-2

Library of Congress Control Number: 2009921821

c Springer-Verlag Berlin Heidelberg 2009

Cover design: WMXDesign GmbH, Heidelberg

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)

This book contains the proceedings of the Fourth International Conference

- advanced applications using the above mentioned innovative technology, and

Sint-Genesius-Rode, Belgium Herman Deconinck

Ghent, Belgium Erik Dick

September 2006 Conference Chair

Part I Invited Speakers

Two new techniques for generating exactly incompressible

On the Computation of Steady-State Compressible Flows

Very High Order, Non-Oscillatory Fluctuation Distribution

High-order residual distribution : discontinuity capturing

Uniform Flow Preserving Property of High Order Upwind

Implementation of an Enhanced Flux Formulation for

Computation of Eigenspaces of Hyperbolic Systems

The High Order WLSQR Scheme and its Applications in

Numerical study of the Colocated Clustered Finite Volume

Arbitrary High Order Finite Volume Schemes on

Part III Algorithms

A high scalability parallel algebraic multigrid solver

Crank-Nicolson Scheme for Solving Low Mach Number

Heated Wake by Deferred Corrected ULTRA

Acceleration of Unsteady Incompressible Flow Calculation

Part IV AMR - Adaptive mesh techniques

A Parallel Unstructured Overset Mesh Technique for

A meshless solver for computing viscous flows on Cartesian

Immersed boundary technique for compressible flow

Part V DNS and LES

Identification and role of coherent structures in two-

Direct numerical simulation of Taylor-Couette flows in the

Flux Limiting Schemes for Implicit Large Eddy Simulation of

Revisiting URANS computations of the flow behind a

Contribution of Turbulence Equation Terms to the Shear

Part VI Optimisation and Flow Control

Optimum Shape Design for Unsteady Three-Dimensional

Numerical Study of Transonic Drag Reduction for Flow Past

Part VII Two-phase and multimaterial flows

Level Set based Finite Element Method of Bubble-in-Liquid

Level Set method for Curvature-driven Flows in Microfluidics

Accurate and Efficient Solution of 2D Steady Water Flows

Numerical analysis of bubble migration in thermocapillary

Part VIII Hypersonic and supersonic flow

Numerical investigation of an effusion cooled thermal

Unified Flow Solver Combining Boltzmann and Continuum

Computing Simulation of Hypersonic Magneto-Fluid-

Part IX Multiphysics and interdisciplinary coupled

Boundary conditions by low-order modelling

Development of “MATIS-SC” for High Speed Steam Flow

Numerical Method for Near-critical Fluids of Arbitrary

CAD-Centric Framework for Aero-mechanical Optimization -

Aeroelastic Solutions for Viscous Flows using the Time

School of Mathematics, University of Minnesota, Minneapolis, MN 55455, USA

H. Deconinck, E. Dick (eds.), Computational Fluid Dynamics 2006,

in a finite dimensional spaces of completely discontinuous velocities which

Here, Ω is a bounded polyhedral domain. The external body force f is taken

2.1 The DG method

To introduce the DG method, we begin by rewriting the equations (1) as