Are you sure?
This action might not be possible to undo. Are you sure you want to continue?
GRIDS
A THESIS SUBMITTED TO
THE GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES
OF
MIDDLE EAST TECHNICAL UNIVERSITY
BY
BERCAN SİYAHHAN
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR
THE DEGREE OF MASTER OF SCIENCE
IN
MECHANICAL ENGINEERING
MAY 2008
Approval of the Thesis:
A TWO DIMENSIONAL EULER FLOW SOLVER ON ADAPTIVE
CARTESIAN GRIDS
submitted by BERCAN SİYAHHAN in partial fulfilment of the requirements for
the degree of Master of Science in Mechanical Engineering Department, Middle
East Technical University by,
Prof. Dr. Canan Özgen _______________
Dean, Graduate School of Natural and Applied Sciences
Prof. Dr. S. Kemal İder _______________
Head of Department, Mechanical Engineering
Prof. Dr. M. Haluk Aksel _______________
Supervisor, Mechanical Engineering Dept., METU
Assist. Prof. Dr. Cüneyt Sert _______________
CoSupervisor, Mechanical Engineering Dept., METU
Examining Committee Members:
Prof. Dr. Kahraman Albayrak _______________
Mechanical Engineering Dept., METU
Prof. Dr. M. Haluk Aksel _______________
Mechanical Engineering Dept., METU
Assist. Prof. Dr. Cüneyt Sert _______________
Mechanical Engineering Dept., METU
Prof. Dr. İsmail H. Tuncer _______________
Aerospace Engineering Dept., METU
Instructor Dr. Tahsin Çetinkaya _______________
Mechanical Engineering Dept., METU
Date: 02/05/2008
iii
I hereby declare that all information in this document has been obtained and
presented in accordance with academic rules and ethical conduct. I also declare
that, as required by these rules and conduct, I have fully cited and referenced
all material and results that are not original to this work.
Name, Last name : Bercan Siyahhan
Signature :
iv
ABSTRACT
A TWO DIMENSIONAL EULER FLOW SOLVER ON ADAPTIVE CARTESIAN
GRIDS
Siyahhan, Bercan
M.S., Department of Mechanical Engineering
Supervisor: Prof. Dr. M. Haluk Aksel
CoSupervisor: Assist. Prof. Dr. Cüneyt Sert
May 2008, 121 pages
In the thesis work, a code to solve the two dimensional compressible Euler equations
for external flows around arbitrary geometries have been developed. A Cartesian
mesh generator is incorporated to the solver. Hence the preprocessing can be
performed together with the solution within a single code. The code is written in the
C++ programming language and its object oriented capabilities have been exploited
to save memory in the data structure developed.
The Cartesian mesh is formed by dividing squares successively into its four
quadrants. The main advantage of using this type of a mesh is the ability to generate
meshes around geometries of arbitrary complexity quickly and to adapt the mesh
easily based on the solution. The main disadvantage of this method is that the
treatment of the cells that are cut by the geometry.
For the solution procedure Roe’s method as well as flux vector splitting methods are
used for the flux evaluation. The flux vector splitting schemes used are van Leer,
AUSM, AUSMD and AUSMV methods. Time discretization is performed using a
v
multistage method. To increase the accuracy least squares reconstruction is
employed.
The code is validated by performing calculations around a NACA0012 airfoil profile.
The effect of reconstruction is demonstrated by plotting the pressure coefficient on
the airfoil. The distribution obtained using reconstruction is very close to the
experimental one while there is a considerable deviation for the case without
reconstruction. Also the shock capturing capabilities of different methods have been
investigated. In addition the performance of each method is analyzed for flow around
an NLR 7301 airfoil with a flap.
Keywords: Cartesian Mesh Generation, Approximate Riemann Solver of Roe, Euler
Equations, Solution Refinement, Flux Vector Splitting
vi
ÖZ
UYARLAMALI KARTEZYEN AĞLARDA İKİ BOYUTLU EULER AKIŞ
ÇÖZÜCÜSÜ
Siyahhan, Bercan
Yüksek Lisans, Makina Mühendisliği Bölümü
Tez Yöneticisi: Prof. Dr. M. Haluk Aksel
Ortak Tez Yöneticisi: Yard. Doç. Dr. Cüneyt Sert
Mayıs 2008, 121 sayfa
Bu tez çalışması kapsamında, herhangi bir geometri etrafındaki iki boyutlu
sıkıştırılabilir dış akışları Euler denklemlerinin çözümüyle modellemek için bir kod
yazılmıştır. Çözücüye bir Kartezyen ağ çözücüsü ilave edilmiştir. Böylece hem ağ
oluşturma hem de çözüm işlemlerinin aynı kodla yapılması sağlanmıştır. Kod C++
dilinin nesneye yönelik programlama özellikleriyle veri yapısının kullandığı hafızayı
azlatma amacı güdülerek yazılmıştır.
Kartezyen hesaplama ağı karelerin art arda dörde bölünmesiyle oluşturulmaktadır.
Kartezyen ağların temel artısı değişik karmaşıklıktaki geometriler etrafında hızla ağ
oluşturulabilmesi ve çözüme göre ağın uyarlanabilmesidir. Temel eksisi ise geometri
tarafından kesilen hücrelerin ele alınmasındaki zorluktur.
Çözümde hücre akılarını bulmak için Roe’nun metodunun yanı sıra FVS metodları
da kullanılmaktadır. FVS metodlarından van Leer, AUSM, AUSMD ve AUSMV
kullanılmaktadır. Daha sonra zamanda integrasyon için çok aşamalı bir metod
kullanılmaktadır. Yöntemin doğruluğunu artırmak için hücre içinde akış değişkenleri
yeniden yapılandırılmaktadır.
vii
Yöntemin geçerliliği NACA0012 kanat profili etrafındaki akış çözülerek sınanmıştır.
Yeniden yapılandırmanın etkileri kanat üzerindeki basınç katsayılarının çizilmesiyle
vurgulanmaktadır. Yeniden yapılandırmalı dağılım deneysel dağılıma yakındır ancak
yeniden yapılandırma olmadan elde edilen sonuçlar deneyselden ciddi farklılıklar
göstermektedir. Ayrıca farklı metodların şok yakalama özellikleri de sınanmıştır.
Farklı metodların performansları NLR7301 profili etrafındaki akış çözülerek de
kıyaslanmaktadır.
Anahtar Kelimeler: Kartezyen Ağ Oluşturucusu, Roe’nun Yaklaşık Riemann
Çözücüsü, Euler Denklemleri, Yeniden Yapılandırma, Çözüm Uyarlamalı Ağ
viii
Dedicated to my family
ix
ACKNOWLEDGEMENTS
I would like to thank my supervisor Prof. Dr. Haluk Aksel, cosupervisor Asst. Prof.
Dr. Cüneyt Sert and Dr. Ruhşen Çete for their assistance, support and constructive
criticisms during my thesis work.
I would also like to thank Emre Gürdamar, Çağrı Şişman, Doruk Özdemir, Güneş
Nakiboğlu and Mehtap Çakmak with whom I had the pleasure of having fruitful
discussions.
Also I would like to express my appreciation for the company of Ender Özden and
Samet Özkan with whom I have shared the same work environment.
Last but not least I would like to express my gratitude for all individuals and
institutions that realize and cherish knowledge as something to be shared openly and
freely.
x
TABLE OF CONTENTS
ABSTRACT ................................................................................................................ iv
ÖZ ............................................................................................................................... vi
ACKNOWLEDGEMENTS ........................................................................................ ix
TABLE OF CONTENTS ............................................................................................. x
LIST OF TABLES .................................................................................................... xiv
LIST OF FIGURES ................................................................................................... xv
LIST OF SYMBOLS ................................................................................................ xix
CHAPTERS
1. INTRODUCTION.................................................................................................... 1
1.1 Overview of Computational Fluid Dynamics (CFD) ................................... 1
1.2 Main Elements of a CFD Program ............................................................... 2
1.3 Mesh Generation .......................................................................................... 3
1.3.1 Structured Meshes .................................................................................. 3
1.3.2 Unstructured Meshes .............................................................................. 4
1.3.2.1 Triangular Unstructured Meshes .................................................... 4
xi
1.3.2.2 Cartesian Meshes ............................................................................ 6
1.4 Solution Methods ......................................................................................... 7
1.4.1 Finite Difference Method ....................................................................... 7
1.4.2 Finite Element Method ........................................................................... 7
1.4.3 Finite Volume Method ........................................................................... 8
1.5 Assessment of the Solution Methods ........................................................... 9
2. DATA STRUCTURE & MESH GENERATION ................................................. 10
2.1 Data Structure............................................................................................. 10
2.1.1 Terminology for the Cartesian Cells .................................................... 10
2.1.2 Neighbour Cell Procedures .................................................................. 12
2.1.3 Information for Cut and Computational Cells ..................................... 13
2.2 Clipping Procedure .................................................................................... 16
2.2.1 Comparison of Polygon and Line Clipping Algorithms ...................... 16
2.2.2 LiangBarsky Line Clipping Algorithm ............................................... 17
2.3 Mesh Generation ........................................................................................ 21
2.3.1 Uniform Mesh Generation ................................................................... 21
2.3.2 Box Adaptation & Celltypes ............................................................... 25
2.3.2.1 Regular Cutcells .......................................................................... 28
xii
2.3.2.2 Irregular Cutcells ......................................................................... 31
2.3.2.2.1 ConcaveSplit ........................................................................... 32
2.3.2.2.2 ConvexUnited ......................................................................... 35
2.3.3 Curvature Adaptation ........................................................................... 37
3. NUMERICAL METHOD ...................................................................................... 40
3.1 Governing Equations .................................................................................. 40
3.2 One Dimensional Riemann Problem for Linear Systems .......................... 42
3.3 The Approximate Riemann Solver of Roe ................................................. 47
3.3.1 Solution Algorithm .............................................................................. 54
3.4 Flux Vector Splitting .................................................................................. 56
3.4.1 Van Leer Flux Vector Splitting ............................................................ 57
3.4.2 AUSM (LiouSteffen) Flux Vector Splitting ....................................... 58
3.4.3 AUSMD Flux Vector Splitting ............................................................ 59
3.4.4 AUSMV Flux Vector Splitting ............................................................ 61
3.5 Boundary Conditions ................................................................................. 61
3.6 Reconstruction of the Flow Variables ........................................................ 62
3.6.1 GaussGreen Reconstruction ................................................................ 63
3.6.2 Least Squares (Minimum Energy) Reconstruction .............................. 64
xiii
3.6.3 Gradient Limiting ................................................................................. 67
3.7 Temporal Discretization ............................................................................. 68
3.7.1 Explicit Time Stepping Schemes ......................................................... 68
3.7.2 Determination of the Time Step ........................................................... 70
3.8 Solution Refinement .................................................................................. 71
4. RESULTS & DISCUSSIONS ............................................................................... 72
4.1 NACA0012 ................................................................................................ 72
4.1.1 Subsonic Tests ...................................................................................... 72
4.1.2 Transonic Tests .................................................................................... 81
4.2 NLR 7301 with Flap .................................................................................. 88
5. CONCLUSIONS .................................................................................................. 106
BIBLIOGRAPHY .................................................................................................... 109
APPENDICES
A. COORDINATES OF NACA0012 ...................................................................... 113
B. COORDINATES OF NLR 7301 & FLAP .......................................................... 116
xiv
LIST OF TABLES
TABLES
Table 2. 1 Information stored for different types of cells ......................................... 15
Table 3.1 Stage coefficients and CFL numbers for the multistage method ............ 69
Table 4.1 Comparison of first and second order solutions ....................................... 78
Table 4.2 Computational characteristics of the methods for evaluating fluxes ........ 84
Table 4.3 Computational performance of the methods for NLR 7301 AoA 6˚ ........ 92
Table 4.4 Computational performance of the methods for NLR 7301 AoA 10.1˚ ... 95
Table 4.5 Computational performance of the methods for NLR 7301 AoA 13.1˚ ... 98
Table A.1 Coordinates of NACA0012 Profile ........................................................ 113
Table B.1 The coordinates of NLR 7301 ................................................................ 116
Table B.2 The coordinates of flap .......................................................................... 119
xv
LIST OF FIGURES
FIGURES
Figure 1.1 The stages of the advancing front method ................................................ 5
Figure 1.2 Vornoї diagram for the grid points ............................................................ 6
Figure 2.1 The relation between cells ....................................................................... 11
Figure 2.2 The tree structure ..................................................................................... 11
Figure 2.3 Neighbours of a cell ................................................................................ 12
Figure 2.4 The visible and invisible regions for a clipping window ........................ 18
Figure 2.5 Automatically known neighbours of a cell ............................................. 23
Figure 2.6 Neighbours associated with the parent cell’s edge neighbours ............... 24
Figure 2.7 Imaginary boxes around a three element airfoil ...................................... 26
Figure 2.8 Mesh around the slat and the airfoil ........................................................ 27
Figure 2.9 Naming convention of a cell ................................................................... 28
Figure 2.10 Examples of regular cutcells ................................................................ 30
Figure 2.11 Grouping and treatment of irregular cutcells ....................................... 32
Figure 2.12 Conversion of an irregular cutcell ........................................................ 33
xvi
Figure 2.13 Concavesplit to special subgroup of regular cutcells conversion ....... 34
Figure 2.14 Concavesplit to regular cutcells conversion ....................................... 35
Figure 2.15 Convexunited to inside cells conversion .............................................. 36
Figure 2.16 Convexunited to regular cutcells conversion ...................................... 36
Figure 2.17 The normals of two neighbouring cutcells ........................................... 38
Figure 2.18 Calculation of the normal ...................................................................... 38
Figure 3.1 Solution of a 1D Riemann problem for a set of linear equations ........... 44
Figure 3.2 Graphical interpretation of flux vector splitting ...................................... 57
Figure 3.3 Ghost right state at the solid wall ............................................................ 62
Figure 3.4 Region formed by the support set............................................................ 64
Figure 4.1 Cp vs. chord for first order solution ........................................................ 73
Figure 4.2 Cp vs. chord for second order solution ................................................... 74
Figure 4.3 A refined and nonrefined mesh around stagnation point ....................... 75
Figure 4.4 Comparison of the second order solutions .............................................. 76
Figure 4.5 Mach contours for NACA0012 subsonic flow ........................................ 77
Figure 4.6 Lift coefficient vs. angle of attack graph ................................................. 79
Figure 4.7 Residual plot for first and second order solutions ................................... 80
xvii
Figure 4.8 Pressure coefficient vs. chord graph for the AUSM methods ................. 82
Figure 4.9 Pressure coefficient vs. chord graph for Roe, van Leer and AUSMV .... 83
Figure 4.10 Residual plot for transonic flow ............................................................ 85
Figure 4.11 Effect of refinement at shock location .................................................. 86
Figure 4.12 Effect of refinement around stagnation point ........................................ 86
Figure 4.13 Mach contours for transonic flow around NACA0012 airfoil .............. 87
Figure 4.14 Pressure coefficient vs. chord for AUSMV with and without refinement
.................................................................................................................................... 88
Figure 4.15 Definition of geometry variables for NLR 7301 with flap .................... 89
Figure 4.16 Pressure coefficient vs. chord graph (AUSM methods AoA 6˚) ........... 90
Figure 4.17 Pressure coefficient vs. chord graph (AUSM, Roe, van Leer AoA 6˚) . 91
Figure 4.18 Pressure coefficient vs. chord graph (AUSM methods AoA 10.1˚) ...... 93
Figure 4.19 Pressure coefficient vs. chord graph (AUSM, Roe, van Leer AoA 10.1˚)
.................................................................................................................................... 94
Figure 4.20 Pressure coefficient vs. chord graph (AUSM methods AoA 13.1˚) ...... 96
Figure 4.21 Pressure coefficient vs. chord graph (AUSM, Roe, van Leer AoA 13.1˚)
.................................................................................................................................... 97
Figure 4.22 Effect of solution refinement around stagnation region of NLR 7301.. 99
Figure 4.23 Effect of solution refinement around the flap ....................................... 99
xviii
Figure 4.24 Pressure coefficient vs. chord for van Leer AoA 6˚ ............................ 100
Figure 4.25 Pressure coefficient vs. chord for van Leer AoA 10.1˚ ....................... 101
Figure 4.26 Pressure coefficient vs. chord for van Leer AoA 13.1˚ ....................... 102
Figure 4.27 Flow field for NLR 7301 with flap AoA 6˚ ........................................ 103
Figure 4.28 Flow field for NLR 7301 with flap AoA 10.1˚ ................................... 103
Figure 4.29 Flow field for NLR 7301 with flap AoA 13.1˚ ................................... 104
xix
LIST OF SYMBOLS
U: vector of conserved variables
ˆ
U :vector of rotated conserved variables
F(U): flux vector function
ˆ ˆ
( ) F U : rotated flux vector function
F: flux vector
dV: volume element
dA: surface area element
n
: face normal
t: time
A
cell
: area of a cell
s ∆ : length of cell face
ρ : density
u : velocity in the xdirection
v : velocity in the ydirection
t
h : total enthalpy
p : pressure
ρ¯ : Roe averaged density
u¯ : Roe averaged velocity in the x
direction
v¯ : Roe averaged velocity in the y
direction
t
h
¯
: Roe averaged total enthalpy
p¯ : Roe averaged pressure
t
e : total energy
ˆ u : normal velocity
ˆ v : tangential velocity
γ : specific heat ratio
a: speed of sound
V : velocity
a¯ : Roe averaged speed of sound
V
¯
: Roe averaged velocity
α : angle of attack
T : temperature
M : Mach number
R : gas constant
T : transformation matrix
A: Jacobian of the flux vector
V : vector of characteristic variables
λ : eigenvalue of the Jacobian matrix
( ) Av u :averaging function
Φ: gradient limiter
R: residual function
β : stage coefficients
σ : CFL number
Ψ: convective spectral radius
S : length of projection
Subscripts
c : value at centroid with th
R : right state variable
L : left state variable
stag : stagnation condition
inf : free stream
x : derivative in x component
y: derivative in x component
Superscripts
~ : Roe averaged quantity
ˆ : components in normaltangential
coordinates
T : transpose
 : averaged quantity
max : maximum value
min: minimum value
x: component in the xdirection
y: component in the ydirection
1
CHAPTER 1
INTRODUCTION
In this thesis, numerical methods are employed to resolve external flows around
arbitrary geometries. The thesis work consists of generating a Cartesian mesh around
the input geometry, solving the Euler equations numerically with the finite volume
method for the desired flow quantities and obtaining the necessary aerodynamic
forces to verify the solution. To understand the motivation for choosing the methods
followed, a brief introduction must be made to mesh generation and solution
techniques employed in the field of computational fluid dynamics (CFD).
1.1 Overview of Computational Fluid Dynamics (CFD)
Computational fluid dynamics is a branch of fluid dynamics which employs
numerical methods for solving the equations of fluid flow which are impossible to
solve analytically due to their complex nature. Hence CFD is the third approach for
solving a flow problem alongside analytical and experimental methods.
It can be said that emergence of CFD as a tool for solving flow problems has taken
place only in the mid 1900s so it is a relatively new field of study compared to
analytical and experimental methods. Since the publication of Isaac Newton’s
Principia in 1687, theoretical and experimental methods, often combined, were the
only tools for analyzing fluid flows of differing nature (1). With the development of
digital computers, however, CFD has become an indispensible tool in the field of
fluid dynamics, arguably with the pioneering work of Kopal who in 1947 formed
tables for the supersonic flow over sharp cones after solving the governing equations
numerically. In the 1950s and 1960s CFD, has been applied to reentry problems and
2
has established its place among the three essential methods for analyzing a fluid flow
problem (1).
1.2 Main Elements of a CFD Program
Any CFD program mainly consists of three parts; the preprocessor, the solver and
the postprocessor (2) .
The preprocessor gathers the inputs to the problem being analyzed, the fluid
properties, boundary conditions, geometry around which the fluid flow will be
investigated. Also the grid is generated around the input geometry by the pre
processor.
The solver part of the program uses numerical techniques to resolve the flow around
the input geometry. This is accomplished by converting the governing partial
differential methods to algebraic set of equations in the three conventional finite
difference, finite element and finite volume methods.
The postprocessor is the part where the solution obtained is visualized. Contours of
desired scalars and the vector fields are displayed within the domain of interest. Also
nondimensional variables and physical quantities of interest are calculated so as to
verify the solution and to judge whether the mesh generated and the schemes used
are appropriate for the problem at hand.
The program developed in the scope of the thesis work contains a preprocessor
which gathers the inputs of the problem and generates a Cartesian mesh in the
computational domain and a solver which uses the finite volume method for the
solution of the compressible Euler equations in the conserved form. The mesh
generation and different solution methods will be discussed in some detail below. For
the postprocessor a commercially available program Tecplot and a freely distributed
program Mayavi have been used.
3
1.3 Mesh Generation
Mesh generation is the most time consuming and the most crucial step in obtaining a
solution to a fluid flow problem. The generated mesh should comply with
requirements of the solution schemes to be used in order to get accurate results.
In general, there are two types of meshes; the structured and the unstructured meshes.
Each one of these two types of meshes can then be classified further according to the
method used to construct them.
1.3.1 Structured Meshes
A mesh is considered to be structured if the organization of the grid points and the
form of the grid cells depend on a general mathematical rule instead of their position
(3). In other words, the grid points can be arranged as a regular array as (i, j, k) with
the neighbouring point being known as (i, j, k+1) (4). Hence the connectivity of the
grid points to form the cells is implicitly implied by the general mathematical rule.
The structured grids can be classified further according to the technique used to
generate them as algebraic, elliptic and hyperbolic grids. The algebraic grids are
formed simply by transforming cells in a Cartesian computational domain to a
physical domain to obtain a boundary conforming mesh, the elliptic grids are
obtained as a result of solving an elliptic differential equation to obtain a mesh that
satisfies the Laplace equations while for the hyperbolic grids are obtained by solving
hyperbolic equations (4).
The advantages of structured meshes are that they are boundary conforming so by
adjusting the grid spacing, meshes for viscous solutions can be obtained. They
require less data storage since the connectivity information is implicitly known.
One disadvantage of the structured meshes is that they require more computation
power, since the governing partial differential equations or the algebraic rules must
be solved and computations must be performed for the transformations from the
4
computational to physical spaces. However the most important disadvantage of the
structured meshes is that they cannot easily be applied to complex geometries that
are formed by more than one body that have small clearances or are not well
streamlined.
1.3.2 Unstructured Meshes
If the grid points cannot be arranged as a regular array and additional information is
required for the connectivity of the grid points then the mesh is unstructured. The
unstructured meshes, though they can comprise of elements of any shape, are formed
of triangular or rectangular elements generally in the twodimensional case.
The advantages of the unstructured meshes are that they can be applied to geometries
of arbitrary complexity with ease, they are more suitable to automatic meshing, and
no transformation between a computational and a physical domain will be needed.
While the disadvantage of these meshes is that they require more complex data
structures for storing the connectivity of grid points and the neighbour cell
information.
1.3.2.1 Triangular Unstructured Meshes
The most widely used methods for forming triangular unstructured meshes are the
advancing front and the Delaunay triangulation methods.
In the advancing front method, the points are added to the ones that represent the
geometry as the mesh is created. The added points are connected to already existing
points to form triangles. Then, the points that are available for triangle formation
termed as the front are connected to newly added points to update the front until the
whole domain is meshed (4). In this approach, extensive search algorithms must be
employed when adding new grid points to ensure that the newly formed cells adjust
to the existing ones. Also, the closing stage of this approach, where the front folds
5
over itself, introduces some difficulties (3). An example of the advancing front
method is illustrated in Figure 1.1.
Figure 1.1 The stages of the advancing front method
The other and may be the most popular triangular meshing procedure, Delaunay
triangulation, requires points to be created before the meshing process and is closely
linked to the concept of Vornoї diagrams. The Vornoї diagram is obtained by
tessellating the domain into Vornoї regions which are regions that are closer to one
particular point within the domain than the others. This is illustrated in Figure 1.2.
Hence they can be obtained by the perpendicular bisectors of the lines between
neighbouring points. Once the diagram is formed, any points that have a Vornoї edge
between them are connected to form the triangular mesh. The circumcircles of the
triangular elements do not contain any other point in the domain and the resulting
mesh is unique for the point distribution. This method is suitable for mesh adaptation
since for a newly added point, the Vornoї diagram is updated locally and new
triangles are formed.
6
Figure 1.2 Vornoї diagram for the grid points
1.3.2.2 Cartesian Meshes
The Cartesian mesh is obtained by using the quadtree approach in two dimensions.
In the quadtree approach, a square containing the whole domain is divided into its
four quadrants to form the mesh in a tree structure. Hence, the procedure followed is
much simpler than the methods for generating triangular unstructured meshes. Also,
in the computation stage, the flux vectors won’t have to be rotated for the greater
portion of the cells. The method is very suitable for adaptation, since for the
adaptation, the cell has to be simply divided into its four quadrants.
If the quadtree approach is followed strictly, then the surface of the geometry cannot
be represented accurately hence the cutcells are used to alleviate this problem which
introduces some complexity to the approach. The orientation of the geometry may
cause formation of very small cutcells, which is generally undesirable.
The code developed uses the quadtree approach to form a Cartesian mesh for the
solution. This approach was chosen to mesh arbitrary geometries of differing
complexities in an automatic fashion. Also the approach is simpler than triangular
mesh generation techniques and it will promote simpler calculations in the solution
stage.
7
1.4 Solution Methods
The partial differential equations governing the flow of a fluid must be discretised in
order to convert them into algebraic equations and get a numerical solution. The
most widely used methods for the discretisation of the governing equations are the
finite difference, finite element and finite volume methods.
1.4.1 Finite Difference Method
The finite difference method is based on the differential form of the governing
system of partial differential equations. The derivatives of the conserved quantities
appearing in these equations are approximated by differences of these unknown
quantities at neighbouring grid points. These differences are derived from the Taylor
series and their order of accuracy is dependent on the number of points used in the
difference equation. Once the difference equations are written in terms of unknown
quantities, these algebraic equations can be solved numerically with the boundary
conditions to obtain the value of the conserved quantities at each grid point.
The major setback of this method is that it requires a regular structured mesh in order
to express the derivatives accurately (5). Hence, the method is not very suitable for
unstructured meshes or cases where the geometry is complex.
1.4.2 Finite Element Method
The finite element method which was initially developed for structural problems
from 1940s through 1960s, divides the computational domain into triangular or
rectangular subdomains and represents the variation of the unknown quantities in
terms of piecewise continuous functions. Then the unknown coefficients of these
functions are found by establishing algebraic equations through satisfying the
governing equations in a weighted residual form over each element (6).
8
The finite element method, with its mathematical background being functional
analysis, has been investigated by mathematicians since its development for
engineering purposes, and is a very rigorous method with specific conditions for
existence and convergence criteria and exactly derived error bounds.
Even though this method is very well established in structural mechanics, it is still in
the stages of evolution for complex flow problems such as compressible flows
governed by Euler or NavierStokes equations (1).
1.4.3 Finite Volume Method
The finite volume method which was first devised as a special form of the finite
difference method in the 1970s uses the integral form of the governing equations of a
fluid flow. There is a wide range of methods for discretising the convective and
diffusive terms of the governing integral equations as well as the definition of the
control volumes for which the governing equations are satisfied. More specifically it
is possible to store the flow variables at the centroid of the cells for a cell centred
approach, or at the vertices for a cell vertex formulation. Another possibility is to
have a staggered grid where the scalar quantities and different vector quantities can
be stored on overlaid cells as in the SemiImplicit Method for PressureLinked
Equations (SIMPLE) algorithm (2). Hence the method has a considerable inherent
flexibility (5). Also, since the method is directly based on the governing integral
equations, the basic concepts of the different schemes are more comprehensible than
the finite element method (2). This property of the finite volume method promotes
development of numerical schemes based on physics of the flow. Approximate
Riemann methods, which are used for evaluating the convective terms is one
example. One other advantage of the method is with the direct discretisation of the
governing integral equations, conservation of mass, momentum and energy will be
guaranteed at a discrete level (5).
9
In the code developed, cell centred finite volume method has been used with of
Roe’s approximate Riemann method for the evaluation of the convective fluxes
through the cell surfaces as well as the flux vector splitting schemes.
1.5 Assessment of the Solution Methods
Theoretically, a numerical solution will tend to the exact solution if infinite number
of cells is used. Since this is not reasonable, a numerical scheme must possess the
conservativeness, boundedness and transportiveness properties (2). The
conservativeness property states that flux leaving through a certain face of a cell
must enter the adjacent cell. The boundedness of a solution guarantees that numerical
errors introduced by the discretisation of the governing equations will not prevent
convergence. This property is related to the stability of the discretisation scheme
used. The stability of a scheme can be analyzed by the discrete perturbation analysis
or the von Neumann stability analysis (7). Finally, the transportiveness of numerical
scheme deals with the directionality of the solution. For example, the solution of a
flow dominated by convection will depend more on the upwind direction, hence
upwind schemes will be more appropriate; whereas for a flow that is diffusive, a non
directional central method will be more appropriate.
10
CHAPTER 2
DATA STRUCTURE & MESH GENERATION
2.1 Data Structure
In the code developed, a Cartesian unstructured mesh is generated around the input
geometry. Since an unstructured mesh is used, the first step in the mesh generation is
to construct an appropriate data structure that will store the necessary data for each
cell in the domain. However before describing each variable that is stored, the
terminology adopted must be explained.
2.1.1 Terminology for the Cartesian Cells
The mesh is obtained by dividing squares successively, starting from a single large
square, by connecting the midpoints of opposing edges until the desired resolution is
obtained. This process results in a tree structure. The large square comprising the
outer boundary of the domain is termed as the root cell. A cell (square) is called the
parent of its four quadrant cells which form as a result of division and the four
quadrant cells are called the children of the parent cell in turn. The children cells are
numbered according to the quadrant they occupy. A cell without any children is
called a leaf cell. The leaf cells can also be called the computational cells since the
actual calculations for the field variables are performed for these cells. These
concepts are illustrated below in Figure 2.1.
11
Figure 2.1 The relation between cells
Among the computational cells, some are cut by geometry lines and these are called
cutcells. The other cells mainly serve the purpose of traversing the tree. Also there is
a level concept associated with each one of the cells. The level of the four children of
a cell is one higher than that of their parent. The root cell is assigned a level of zero.
The tree structure associated with the structure shown in Figure 2.1 is illustrated in
Figure 2.2 to exemplify some of the concepts defined above.
Figure 2.2 The tree structure
child 1
child 1 child 2 child 4 child 3
child 3
child 1 child 2 child 4 child 3
child 2 child 4
root level 0
level 1
level 2
leaf cells
child1 child 2
child 3 child 4
leaf cell parent
12
2.1.2 Neighbour Cell Procedures
As was mentioned before for unstructured meshes the neighbours must also be stored
since for the calculation of the fluxes through a face, the neighbour along that edge
must be known. Also the neighbours will be used to determine the variation of the
conserved variables within the cells.
A cell will have four neighbours along its edges called the edge neighbours. These
neighbours are named east, west, north and south according to the edge they are
associated with. Also computational cells will have four neighbours through their
vertices named as vertex neighbours; northeast, southeast, northwest, southwest. The
general rule for the mesh is that a cell can have neighbours of levels that are one
degree different than the cell. This ensures the grid smoothness and also simplifies
flux calculations through the faces. If a neighbour is at the same level or one lower
level, then it will be recognized as the neighbour along the corresponding edge, but if
it is at one higher level the neighbour’s parent will be recognized instead of having
two neighbours along the edge. Figure 2.3 illustrates the concepts related to
neighbouring relations and also the possible configurations that might occur for a
cell.
Figure 2.3 Neighbours of a cell
no northeast
north neighbour
east neighbour
west neighbour
south neighbour
southeast neighbour
northwest neighbour
southwest neighbour
S1 S2 SE
y
x
13
In Figure 2.3 the east, west and southwest neighbours of the cell are at the same level
while the north and northwest neighbours are at one lower level. These cells are
recognized as the neighbours themselves meanwhile the parent of the cells marked as
S1, S2 and SE are recognized as the south and southeast neighbours respectively.
Since the north neighbour is at a lower level the north and northeast neighbours are
identical in which case the cell is considered to have no northeast neighbour, hence a
vertex neighbour of a cell should be distinct if it is to be stored.
2.1.3 Information for Cut and Computational Cells
As it is defined earlier there are three types of cells; the parent cells for traversing the
tree, the computational leaf cells and the cutcells, which are special types of
computational cells. The amount of data stored for each cell depends on its type. The
information stored for a parent cell is shared for all types of cells, while
computational cells will have additional information and for cutcells further
additional information will be stored.
For each cell in the domain, the x and y coordinates of the centre and the level of the
cell are stored. Also there are a total of thirteen pointers pointing to other cells in the
domain. Of these thirteen pointers four point to the children of the cell, if the cell is a
leaf cell then these pointers are null pointers. Eight of the remaining nine pointers
point to the neighbours of the cell. Of these eight pointers four point to edge
neighbours, the remaining four point to the vertex neighbours and they are allocated
only for computational cells. The remaining one of the thirteen pointers points to the
parent of the cell. This pointer is null only for the root cell.
Besides the thirteen pointers aforementioned there are nine more pointers which are
allocated only for computational cells. Four of these pointers store the conserved
field variables for the continuity, momentum and energy equations. Of the remaining
five pointers two are storing the gradient and the curl of the velocity vector. One
pointer stores information dictating the cell to be refined at a stage of mesh
generation or solution adaptation. One stores coefficients that will be used in solution
14
reconstruction and the last one stores the type of the cell; whether the cell resides
totally outside or inside the geometry or whether it is cut by it.
Finally there is a set of seven pointers which are allocated for the cutcells. Three of
these seven pointers store the x and y coordinates of the centroid and area of the cell
while two store the x and y coordinates of the end lines of the intersections of
geometry lines with the cell edges. Of the remaining two pointers one stores which
edges are cut by geometry lines while the last one stores information on which
portion of the cell’s edges flux passes through. The table below shows a summary of
the information stored for different types of cells.
15
Table 2. 1 Information stored for different types of cells
xcent:
ycent:
level:
child1:
child2:
child3:
child4:
eneigh:
wneigh:
nneigh:
sneigh:
parent:
ro:
rouvel:
rovvel:
roen:
gradient:
curl:
recons:
refine:
celltype:
neneigh:
seneigh:
nwneigh:
swneigh:
centx:
centy:
area:
xcut:
ycut:
edgecut:
fluxcut:
pointer to first quadrant child
pointer to second quadrant child
pointer to third quadrant child
pointer to southeast neighbour
density associated to the cell
product of density and velocity in the x direction associated to the cell
ALL CELLS
product of density and velocity in the y direction associated to the cell
product of density and total energy associated to the cell
pointer to gradient of velocity
pointer to fourth quadrant child
pointer to east neighbour
pointer to west neighbour
pointer to north neighbour
pointer to south neighbour
pointer to parent
x coordinate of center of square containing the cell
y coordinate of center of square containing the cell
level of the cell
pointer to the y coordinate of cut location on edge
pointer to edges cut byb the geometry
pointer to the information regarding flux
COMPUTATIONAL CELLS
CUTCELLS
pointer to northwest neighbour
pointer to southwest neighbour
pointer to x coordinate of the centroid of the cell
pointer to y coordinate of the centroid of the cell
pointer to the area of the cell
pointer to the x coordinate of cut location on edge
pointer to curl of velocity
pointer toarray containing geometry based reconstruction coefficients
pointer to flag for refinement
pointer to the cell type (inside, outside, cutcell)
pointer to northeast neighbour
16
2.2 Clipping Procedure
Before describing the mesh generation process in detail, the procedure for
determining the location of the points cut by the input geometry for the cutcells
must be explained.
Clipping is the process of determining the portion of a geometrical shape that resides
within a given clipping window. For the code developed, the input geometry, which
is a polygon comprising of connected line segments, is to be clipped and the clipping
windows are the computational cells within the domain. The two options for the
clipping procedure are the polygon clipping, if the input geometry is conceived as a
polygon, line clipping if the input geometry is conceived as a collection of individual
line segments.
2.2.1 Comparison of Polygon and Line Clipping Algorithms
When the polygon and line clipping algorithms are compared, in general, it is seen
that the simplest polygon clipping algorithm, SutherlandHodgeman (8), is not
capable of handling concave geometries effectively. However there are other
polygon clipping algorithms that can handle arbitrary geometries such as self
intersecting geometries and geometries with internal holes. Since these cases are not
in the scope of the current study, the application of these algorithms would bring
unnecessary complications. Hence, the line clipping algorithms are preferred to
polygon clipping algorithms.
After this general comparison, a few of the line clipping algorithms were compared
amongst themselves. First the LiangBarsky (9) line clipping algorithm is compared
with the most classical algorithm, SutherlandCohen (10) algorithm. In the work of
Liang and Barsky (9), the performance of their algorithm was tested along with the
SutherlandCohen line clipping algorithm with 4 sets of data containing 1000
randomly generated line segments. To reduce the effects of random variation, each
line segment was clipped 1000 times resulting in a total of 4 million clippings. It was
17
reported that an improvement of about 36 percent in the execution time was
obtained.
Even though Liang and Barsky claim that their algorithm is more efficient than
Sutherland Cohen, other researchers postulate that experimental analysis is
inadequate to measure the performance of a clipping algorithm. Instead it is
suggested that a measurement based on operation count should be used (11). A study
conducted by J. D. Day (12) takes both the experimental and operations count view
into consideration and reveals that NichollLeeNicholl (11), SobkowPospisilYang
(13) algorithms are more efficient in both aspects compared to the LiangBarsky
algorithm. However the algorithms mentioned are considerably more complex, in
fact the latter algorithm classifies lines into 81 different cases.
Since the clipping procedure will be carried out once at the mesh generation step for
all the computational cells and for the cells formed as a result of solution adaptation,
its effect on overall computation time will not be very significant. Hence, the
simplest algorithm that is the easiest to implement has been chosen. Thus the Liang
Barsky line clipping algorithm has been employed to determine the location of the
points that are cut by the input geometry lines.
2.2.2 LiangBarsky Line Clipping Algorithm
The LiangBarsky line clipping algorithm (9) tests the extensions of given line
segments against each one of the four boundaries of a rectangular clipping window
independently, to determine which if any portion of the line segment is inside the
clipping window. A line segment with end points V
0
and V
1
which happens to reside
wholly within a clipping window is shown in Figure 2.4. The line obtained from the
segment and its extensions will be tested with each one of the lines comprising the
boundary of the clipping window to determine which portion of the line is visible.
After each test the invisible part of the line will be discarded. After all the boundaries
are tested, the portion of the line segment that is inside the clipping window will be
obtained.
18
Figure 2.4 The visible and invisible regions for a clipping window
The algorithm makes use of the parametric form of the line segments with end points
V
0
(x
0
, y
0
) and V
1
(x
1
, y
1
) as given in equations (2.1) to (2.4).
0
x x x t = +∆ ⋅ (2.1)
0
y y y t = +∆ ⋅ (2.2)
1 0
x x x ∆ = − (2.3)
1 0
y y y ∆ = − (2.4)
Adopting this representation t=0 corresponds to the first end point V
0
and t=1
corresponds to the second end point V
1
. The orientation of the line is considered to
be from point V
0
to V
1
.
Once the line is parameterised, the clipping window can be visualized as the
combination of four boundary lines each having a visible and an invisible region as
visible
visible
invisible
invisible
invisible
x
right
x
left
y
bottom
y
top
line segment
V
0
V
1
extension of
line segment
19
in Figure 2.4. Ultimately, the intersection of all the visible sides of each boundary
line would give visible region for the whole clipping window.
The visible region for the clipping window can be represented in terms of four
inequalities using the parametric relations (2.1) to (2.4).
0 right
x t x x ∆ ⋅ ≤ − (2.5)
0 left
x t x x −∆ ⋅ ≤ − (2.6)
0 top
y t y y ∆ ⋅ ≤ − (2.7)
0 bottom
y t y y −∆ ⋅ ≤ − (2.8)
Finally these relationships can be generalized to determine the parametric values
corresponding to the segment of the line that resides inside the clipping window.
i i
p t q ⋅ ≤ 1,.., 4 i = (2.9)
where
1
p x = ∆
1 0 right
q x x = −
2
p x = −∆
2 0 left
q x x = −
3
p y = ∆
3 0 top
q y y = −
4
p y = −∆
4 0 bottom
q y y = −
It is observed that when p
i
is positive, the orientation of the extended line is from the
visible side of a particular boundary line to its invisible side. This will result in an
upper limit for the visible portion of the line.
20
i
i
q
t
p
≤ (2.10)
When p
i
is negative, the orientation of the line is from the invisible side to the visible
side, hence a lower bound for the extended line is expected.
i
i
q
t
p
≥ (2.11)
Since the portion of the line segment and not the extended line that resides in the
clipping window is sought another condition that the bound obtained for t should be
between 0 and 1 must also be implemented. Thus for the case when p
i
is different
than zero; the problem of determining the values of t for which a certain portion of
the line segment is within the clipping window becomes a minmax problem in the
form as in
0
t t ≥ (2.12)
1
t t ≤ (2.13)
where
{ }
{ }
0
1
max  0, 1, 2, 3, 4 , 0
min  0, 1, 2, 3, 4 , 1
i
i
i
i
i
i
q
t p i
p
q
t p i
p
⎛ ⎞ ⎧ ⎫
= < =
⎜ ⎟ ⎨ ⎬
⎜ ⎟
⎩ ⎭ ⎝ ⎠
⎛ ⎞ ⎧ ⎫
= > =
⎜ ⎟ ⎨ ⎬
⎜ ⎟
⎩ ⎭ ⎝ ⎠
The t
0
and t
1
values obtained as a result of this minmax problem should satisfy the
following equation
0 1 0 1
t t t t t ≤ ≤ ⇒ ≤ (2.14)
21
Finally, the case when p
i
is zero should be considered. The case p
i
equals to zero
corresponds to a vertical or a horizontal line. For this case a relationship independent
of t is obtained as
0
i
q ≤ (2.15)
The validity of this statement is solely dependent on the value of q
i
. If q
i
is positive,
the line segment is visible for that boundary line else it is invisible. This is referred to
as the trivial reject case.
2.3 Mesh Generation
The mesh generation is performed in three main steps. The first one is the uniform
mesh generation where a mesh of uniform sized elements of the desired level is
obtained. The second one is the box adaptation in which cells around an imaginary
box (rectangle) of given dimensions around the input bodies are refined till they
reach a desired size. The type of the cells; cutcell, inside, outside, are determined at
this stage. The final step is to refine the cells around the places where there are large
gradients on the bodies. The procedure followed is similar to the one outlined in (14).
2.3.1 Uniform Mesh Generation
An initial uniform mesh is formed in the solution domain which will be refined in the
later stages to form the mesh on which the solution is to be performed. The reason
for forming this initial mesh is to have a prescribed resolution in the outer boundaries
of the domain.
Uniform mesh generation is fairly straightforward; starting from the root cell,
refinement is performed when the cell level is less than the desired level. If a cell is
to be refined, first the level of the edge and vertex neighbours must be investigated.
If a neighbour is at a lower level, then that neighbour must be refined prior to the
refinement of the cell. In this way, the one level rule is preserved in the process of
22
mesh generation. The one level rule states that the difference in the levels of two
neighbouring cells cannot exceed one. This rule enforces a smooth mesh and also
eases the calculations for the fluxes. When a cell is to be refined, it is simply divided
into its four quadrants which are its children. Then the level of the children can be set
to be one more than that of their parent and their centres can be calculated from the
centre coordinate of their parent’s, the overall size of the domain, and its level as
(2.16) to (2.19)
( / 2) ( / 2)
,
2 2
c parent c parent l l
d d
x x y y = + = + (2.16)
( / 2) ( / 2)
,
2 2
c parent c parent l l
d d
x x y y = − = + (2.17)
( / 2) ( / 2)
,
2 2
c parent c parent l l
d d
x x y y = − = − (2.18)
( / 2) ( / 2)
,
2 2
c parent c parent l l
d d
x x y y = + = − (2.19)
for the 1
st
quadrant to 4
th
quadrant children, respectively. Where d is the domain size
and l is the level of the cell.
Once this information is obtained, the neighbours of each cell must also be
determined. The procedure followed is similar to the one outlined by De Zeeuw (14).
For any cell two edge neighbours and one vertex neighbour will be known
automatically since they will be the children of the cell’s parent. The other two edge
neighbours and one vertex neighbour will be associated with the parent’s edge
neighbours while the last vertex neighbour will be related to a neighbour of a
neighbour of the parent cell. Thus, similar operations will be performed for each
group of cells.
The neighbour search can be explained in detail for a first quadrant child
exemplifying the process for the other quadrant children. For a first quadrant child,
23
the first group of neighbours, ones known automatically, will be the west, south and
southwest neighbours. They are the second, fourth and the third quadrant children of
the cell’s parent and they are at the same level. This can be depicted as in Figure 2.5.
Figure 2.5 Automatically known neighbours of a cell
For the second group of neighbours, associated with the parent’s neighbours, the
children of the parent’s neighbours must be evaluated. For a first quadrant child these
are the east and north neighbours of the parent specifically. Due to the one level rule
there are three possibilities for these neighbours of the parent; they might be leaf
cells hence they are at a lower level than the cell, they might have leaf cell children
at the same level as the cell and they might have children with leaf cell children at
one level more than the cell. These cases are illustrated in Figure 2.6 for the east
neighbour of the parent.
southwest
neighbour
west
neighbour
cell
south
neighbour
24
Figure 2.6 Neighbours associated with the parent cell’s edge neighbours
In Case 1, the east neighbour of the cell is directly the east neighbour of the parent
and there is no southeast neighbour due to the fact that this cell does not have any
children. In Case 2, the east neighbour is the second quadrant child of the parent’s
east neighbour while the southeast neighbour is the third quadrant child of the
parent’s east neighbour. In Case 3, there are two actual east neighbours; the second
and third quadrant children of the parent’s east neighbour’s second quadrant child.
Since it is not desirable to complicate the data structure by storing the east
neighbours separately, their parent, the second quadrant child of the cell’s parent’s
east neighbour, is stored as the east neighbour. During the calculations it is always
checked to see if a neighbour has children to identify the cells that are marked as E1
and E2 in Figure 2.6. Following the same convention, the third quadrant child of the
cell’s parent’s east neighbour is stored as the southeast neighbour even though the
second quadrant child of this cell stored is the only actual southeast neighbour.
Hence, it can be said that, as a general rule, the cell stored as the neighbour will be
either at a lower level than the cell or it will be at the same level.
cell cell
cell
E
E
SE
E1
E2
SE
Case 1: East neighbour of parent is at a
lower level
Case 2: East neighbour of parent has
leaf children cells
Case 3: East neighbour of parent has
children with leaf children cells
25
With similar reasoning, the north neighbour of a first quadrant child will be its
parent’s north neighbour when this cell is a leaf cell, in which case there will be no
distinct northwest neighbour. When the cell’s parent’s north neighbour has children,
the fourth and the third quadrant children of this cell will be stored as the north and
the northwest neighbours of the cell regardless of them having children.
The only remaining neighbour for the first quadrant child is the northeast neighbour.
This neighbour is the north neighbour of the parent’s east neighbour if this cell has
no children otherwise it is the third quadrant child of this cell. The same search is
also performed through the north neighbour of the parent in case its east neighbour
resides within an input geometry. In this case, this neighbour is the east neighbour of
the parent’s north neighbour if this cell has no children; otherwise it is the third
quadrant child of this cell.
With the same line of reasoning, the rules for determining each neighbour of any
quadrant child can be devised.
2.3.2 Box Adaptation & Celltypes
The next step in mesh generation is the box adaptation. Box adaptation will allow the
mesh to have a desired resolution in regions around the input geometries. This step
consists of two stages for input geometries that are comprised of more than one body.
The first step is to refine cells that are in contact with an imaginary rectangular box
around the collection of the bodies, while the second step is to refine the cells that are
in contact with rectangular boxes around individual bodies until the input size
criterion for each body is satisfied. The rectangular box around each one of the
bodies spans from the minimum to maximum coordinate of a specific body in both of
the Cartesian directions which will be used in the second stage of the box adaptation.
Similarly a rectangular box around the collection of the bodies can be formed. This
box is then stretched by an input proportion. The latter described imaginary boxes for
a three element airfoil configuration can be depicted as in the figure below.
26
Figure 2.7 Imaginary boxes around a three element airfoil
For the first stage of box adaptation, the outer box depicted in Figure 2.7 is used.
Any cell that is in contact with the box is refined until the resulting leaf cells satisfy
the largest one of the input refinement criterion for each body. The input refinement
criteria for each body denote the maximum size of a cell as a proportion of the
maximum dimension of that body. In the first stage of box adaptation, this largest
criterion is taken to be the proportion of the maximum dimension of the collection of
bodies. The maximum dimensions are the larger of the differences of the maximum
and minimum x or y coordinates of a body or the collection of all bodies.
In the second stage, the cells that are in contact with the boxes around each body is
refined until the given refinement criterion for that body is satisfied. The second
stage allows the mesh to have differing resolutions around each body. Figure 2.8
illustrates how a mesh of different resolutions around the slat and the airfoil is
obtained by applying the second stage of box adaptation.
27
Figure 2.8 Mesh around the slat and the airfoil
After the box adaptation, the next step in mesh generation is curvature adaptation.
But before the curvature adaptation, the cells that are cut by the geometry lines must
be determined. All the cells in the mesh will be marked as one of the types outside,
inside or cutcell. For the cutcells, information regarding the edges cut by the
geometry and the portions of the edges through which fluxes pass must be stored.
To determine the cell types, first the relative location of each cell with respect to the
large box around the collection of the bodies used for box adaptation is employed.
The cells that are in contact with this box have the possibility of being inside or
outside the bodies or they might be cut by one of the bodies. Line clipping discussed
earlier is performed to determine the exact configuration. Through line clipping, the
line portions that cut the cell will be determined if the cell is cut. A cell that is cut by
only one geometry line is a regular cutcell and it can be handled easily. However a
cell cut by more than one line must be converted to either a regular cutcell, an
outside or an inside cell by means of modifying the body.
28
2.3.2.1 Regular Cutcells
Regular cutcells are the ones that are cut by only one geometry line portion. For
these cells, the coordinates of the end points of the cutting geometry line will be
known after the line clipping process. The only remaining thing is to determine on
which edges these points lay and through which portion of each edge the flux passes
through. In other words through which portion of the edge the cell exchanges
information with another neighbour cell. Before going any further, the naming
convention of the edges and the corners can be defined as in Figure 2.9.
Figure 2.9 Naming convention of a cell
The first step is to detect a special subgroup of regular cutcells, the ones with the
geometry line lying on one of the edges of the cell. If the x coordinates of both of the
cutting end points are equal then there is a chance that the line comprising of these
points lays on one of the east or the west edges. Hence if the x coordinate of one of
the cutting end points is equal to the x coordinate of corner 0, then it means that the
line lays on the east face, in which case the edge cut is stored as east and no
information is stored for the flux. If the x coordinate of one of the cutting end points
is equal to the x coordinate of corner 2 then it means that the line lays on the west
face, in which case the edge cut is stored as west and no information is stored for the
flux once again. A similar check is also performed on the y coordinates of the cutting
east edge
north edge
west edge
south edge
corner 0
corner 1
corner 2
corner 3
29
edge; if they are equal to each other and to one of the y coordinates of corner 0 or
corner 2, then the edge cut is stored as north or south respectively and no flux
information is stored. Thus for the special subgroup of regular cutcells only one
edge is marked as cut while no information on the flux is stored.
Once the subgroup of special cells is detected the information for the other regular
cutcells can be determined by edge based tests. Starting from the east edge and
going counterclockwise, it is checked whether any of the edges is cut by one of the
end points of the geometry line or not. For the east face, it is checked whether the x
coordinates of the end points of the line are equal to the x coordinate of corner 0. If
one of them is equal then the east edge is marked as being cut, making corner 0,
corner 3 and any point in between along the east edge a candidate for the location of
the cutting end point. To determine the flux information, a point along the east edge
whose location relative to the cutting end point is known must be tested to see if it is
inside the body or outside it. The insideoutside test employed is based on the simple
point location test described in (15). The two possible points are corner 0 and corner
3. If the cutting end point is not on corner 0 this point is tested. If it is outside, the
flux for the east edge is marked as larger indicating that the flux passes through the
portion of the east edge with y coordinates greater than the one of the cutting end
point. Otherwise smaller is stored for the east face flux indicating that the flux passes
through the portion of the east edge with y coordinates less than the one of the
cutting end point. If the cutting end point is on corner 0 then corner 3 is tested and if
it is outside smaller is stored if it is inside larger is stored. If the east edge is not cut
then the centre of the edge is put to the insideoutside test, and if this point is outside
any of the bodies then the flux information is stored as all, indicating that the flux
passes through the whole face and otherwise it is stored as none, indicating that no
flux passes through the face.
Once the configuration of the east face is determined it is checked whether any of the
cutting end points lay on the north face. If a point is detected to be on the north face
the possible locations for it are either corner 1 or anywhere along the north edge
between corner 0 and corner 1. For the insideoutside test corner 0 is used since it is
already known that the cutting edge cannot be on this point. If corner 0 is outside,
30
then the flux is stored as larger, meaning that it passes through the portion of the
edge with x coordinates greater than the cutting edge otherwise smaller is stored
indicating that flux passes through the portion with x coordinates less than that of the
cutting edge. If the north edge is not cut then the centre of the edge is tested to see if
the flux passes through the whole face or it doesn’t pass at all.
After the north face the west face is examined. If this edge is cut, corner 1 is sent to
the insideoutside test as the only point along west edge on which the cutting end
point may not lay. The flux information is stored as larger if this point is outside
otherwise it is stored as smaller. If the edge is not cut, the centre of the edge is tested
to see if the flux passes through the face.
Finally, if the south face is cut either one of the corner 2 and corner 3 can be tested to
fill the flux information accordingly. If the edge is not cut, the centre of the edge is
tested to see if the flux passes through the face. In Figure 2.10, exemplifying cases
for regular cutcells are illustrated.
Figure 2.10 Examples of regular cutcells
In Figure 2.10, the scribbled parts represent the portion of the cell that is occupied by
a body. Case 1 is an example of the special subgroup of regular cells. In this
particular example the only cut edge is stored as the south edge and no flux
information is stored. In Case 2 the cut edges are detected as the east and west edges.
For the east face, the flux information is stored as larger even though no flux passes
through the face. The north flux information is all while for the west face it is larger
Case 1 Case 2 Case 3
31
and for the south face it is none. In Case 3, the edges cut are the north and west
edges. The flux information for the east and the south faces are all, it is larger for the
north face and smaller for the west face.
2.3.2.2 Irregular Cutcells
The cutcells that are cut by more than one geometry line are termed as irregular cut
cells. To convert these cells into regular cutcells, refinement is the first measure
taken. The cells that are cut by more than two geometry lines and those that are cut
by two lines that do not have a common end point are refined until their children are
cells of acceptable types or until they are refined to the maximum allowable level in
the mesh. It can safely be said that there will be no cells cut by more than two
geometry lines as a result of this refinement process, since the maximum level a cell
can reach will make it much smaller than the dimensions of the body. The cells that
are cut by two lines that do not have a common end point will eventually become
regular cutcells if they are refined unconditionally. But sometimes they will require
very extensive refinement reaching levels much higher than the maximum allowable
one especially in the presence of very acute angles. These cases along with the cells
that are cut by two lines that have a common end point are treated separately.
In contrast to the other types, cells that are cut by two lines that have a common end
point are not refined unconditionally until they reach the maximum allowable level.
Instead they are refined, only when their distinct end points lay on the same face of
the cell as in Case 2 and Case 5 of Figure 2.11. The reason for this is that since the
bodies are represented as lines with specified end points, there would be a clustering
of highly refined cells around each of the specified points if these cells were to be
refined unconditionally. Those cells that are refined, the ones that have their distinct
end points on the same edges, are tested to see if they are concave or convex. If they
are convex, then they are simply converted to inside cells as in Case 2 of Figure 2.11.
At the end of this refinement process, the cells that were not converted are classified
into two groups; convexunited and concavesplit and treated accordingly. The first
32
group is formed of cells that are cut by two lines without a common end point with a
single convex area (Case 3 of Figure 2.11), and the cells that are cut by two lines
with a common end point with the distinct end points on separate edges forming a
single convex cell area (Case 4 of Figure 2.11). The second group of cells are the
ones that are cut by two lines without a common end point with two separate convex
cell areas (Case 5 of Figure 2.11), cells that are cut by two lines with a common end
point forming a single concave area (Case 6 of Figure 2.11). The grouping and the
treatment of the irregular cutcells can be depicted as in Figure 2.11.
Figure 2.11 Grouping and treatment of irregular cutcells
2.3.2.2.1 ConcaveSplit
The concavesplit cells are converted to regular cutcells by making modifications on
the body. The location of the end points on the edges and the direction towards
which the lines converge is of essential importance in the conversion process. The
basic idea is to trim the body towards the direction in which the cutting lines
Case 1 Refine unconditionally
till converted
Case 2 Refine till maxlevel
then convert to inside
Case 3 Refine till maxlevel
then treat as convexunited
Case 4 Treat as convex
united without refinement
Case 5 Refine till maxlevel
then treat as concavesplit
Case 6 Treat as convex
united without refinement
33
converge. Hence the geometry line cutting the cell will be the connection of the end
points in the diverging direction. The direction towards which the lines converge can
be determined by placing one coordinate of the end points of a line to the equation of
the other line to obtain the corresponding other coordinate. Then the line converges
towards the direction in which the difference between the calculated coordinate and
the coordinate of the end point is smaller. This can be represented more clearly
graphically as in Figure 2.12.
Figure 2.12 Conversion of an irregular cutcell
The criterion for the above conversion is expressed mathematically in the following
equation.
2 0 2 1
0 3 2 2 1 3 2 2
2 3 2 3
(  ) (  )
 (  )  (  )
(  ) (  )
x x x x
y y y y y y y y
x x x x
⎡ ⎤ ⎡ ⎤
+ > +
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
(2.20)
This trimming process will result in obtaining a member of the special subgroup of
regular cutcells in some cases, while regular cutcells are obtained in some cases.
The special subgroup will be obtained for cells that are cut by lines with a common
end point having their distinct end points on the same edge, cells having two pairs of
end points on the same edge and cells that have one pair of end points on the same
edge while converging towards the end points that are on different edges. These
cases and their conversion are illustrated in Figure 2.13.
x
0
,y
0
x
1
,y
1
x
2
,y
2
x
3
,y
3
x
0
,y
0
x
2
,y
2
34
Figure 2.13 Concavesplit to special subgroup of regular cutcells conversion
In Case 1 of Figure 2.13, since there is a common end point, the remaining end
points will be connected to obtain a cell that is cut on its east edge. In Case 2, there
are two pairs of end points that are on the same edge. In this case since the lines
diverge towards the north face, this face will be marked as being cut. In Case 3, there
is one pair of end points on the same face. Since the lines diverge towards this face,
the west face is marked as being cut.
The cells that convert into regular cutcells as in Figure 2.14 are the ones that have
common cutting end points with the other end points on different edges, cells with all
cutting end points on different edges, and the cells that have one pair of end points on
different edges converging towards a pair of end points that are on the same edge.
Case 1 Case 2
Case 3
35
Figure 2.14 Concavesplit to regular cutcells conversion
In Case 1 of Figure 2.14, there is a common end point and the other end points are
connected automatically to form a regular cutcell. In Case 2, all the end points are
on distinct edges. Since the distance between the ones on the west and the south
edges is greater than the distance between the others the cut locations on these edges
are connected to form a regular cutcell. In Case 3, since the lines diverge towards
the end points that are on different edges in contrast to Case 3 of Figure 2.13, a
regular cutcell is formed by the connection of the cut location on these edges.
2.3.2.2.2 ConvexUnited
The procedure followed for the convexunited cells are similar to the ones for
concavesplit cells. Once again the end points towards which the cutting lines
diverge are connected to form the new cutting edge. In the case of the convexunited
cells conversion is made to either an inside cell or to a regular cutcell. The cells that
have two pairs of cutting end points on the same faces, and the cells that have a pair
of end points on the same edge with the lines converging towards the end points that
are on different edges are converted to inside cells. These cases are illustrated in
Figure 2.15.
Case 1
Case 2
Case 3
36
Figure 2.15 Convexunited to inside cells conversion
Case 1 and Case 2 of Figure 2.15 are analogous to those presented in Figure 2.13.
With the difference that they convert to inside cells, and there is no special subgroup
of inside cells.
The cells that are converted to regular cutcells are the ones that have common
cutting end points with the other end points on different edges, cells with all cutting
end points on different edges, and the cells that have one pair of end points on
different edges converging towards a pair of end points that are on the same edge.
These cases are illustrated in Figure 2.16.
Figure 2.16 Convexunited to regular cutcells conversion
Case 1
Case 2
Case 3
Case 1 Case 2
37
The Case 1, Case 2 and Case 3 of Figure 2.16 are analogous to those in Figure 2.14.
The difference between them is the portions of the cells that are marked as inside and
outside are opposite in these cases.
Since each irregular cell is converted individually into a cutcell, at the end of the
process some cutcells that are not part of any body will be obtained. Hence, at the
end of the conversion process, the mesh has to be traversed to find these unconnected
cells and convert them into either outside cells, if they were processed as concave
split cells, or into inside cells if they were treated as convexunited cells.
2.3.3 Curvature Adaptation
The last step in the mesh generation process is the curvature adaptation. The purpose
of the curvature adaptation is to ensure that regions of a body that have high
curvatures are resolved enough to be represented accurately and also since these
regions will be associated with high gradients.
In curvature adaptation, the cutting edges of neighbouring cutcells are evaluated to
decide on whether to refine a cell or not. So firstly neighbouring cutcells should be
detected then curvature must be quantified to develop a criterion for refinement. In
order to quantify the curvature, the angle between the normals of the cutting edges
pointing towards the fluid domain is used. If this quantity exceeds a threshold value
then the cell is refined. This can be illustrated as in Figure 2.17.
38
Figure 2.17 The normals of two neighbouring cutcells
The normal for a cutting edge of a cell can be determined through the flux
information stored for a cut edge.
Figure 2.18 Calculation of the normal
In Case 1 of Figure 2.18, the first edge that is cut is the east edge and the flux
information associated is larger. In this case, the normal is given by the following
equation.
0 1 0 1
 (  ) (  ) n yi xj y y i x x j = ∆ + ∆ = +
(2.21)
∆x
∆y
∆x
∆y
x
0
,y
0
x
1
,y
1
x
0
,y
0
x
1
,y
1
n
n
Case 1
Case 2
θ: The angle between normals
39
In Case 2, the first edge cut is east as in the previous case but this time the associated
flux information is smaller making the normal;
0 1 0 1
 (  )  (  ) n yi xj y y i x x j = ∆ ∆ =
(2.22)
Hence, it is possible to express each normal pointing to the fluid domain using the
first edge cut and the flux information associated with it for the neighbouring cells.
Then, a relation for the angle between the normals can be obtained by their dot
multiplication as
1 2 1 2
1 2 1 2 1 2 1 2
1 2
( )
   cos cos
  
x x y y
x x y y
n n n n
n n n n n n n n
n n
θ θ
+
= + = ⇒ =

(2.23)
Then a cell is to be refined, if this θ value is greater than a threshold value.
1 2 1 2
1 2
( )
cos cos
  
x x y y
th
n n n n
n n
θ θ
+
= <
(2.24)
40
CHAPTER 3
NUMERICAL METHOD
3.1 Governing Equations
In the code developed, solution to the compressible Euler equations in the conserved
form is sought. Hence, the governing equations can be represented as
0 dV ndA
t
∂
+ =
∂
∫∫∫ ∫∫
U F

(3.1)
where, U and Fare the vectors of conserved variables and the flux respectively,
while n
is the normal to the differential area element dA . This equation can be
descritised in 2D as
1
0
edges
cell
n s
t A
∂
+ ∆ =
∂
∑
U
F
 (3.2)
n
being the face normal, s ∆ the edge length and
cell
A the cell area. The vector of
conserved variables and the flux vector can be written as
2
2
cos sin
cos sin cos
sin cos sin
cos sin
t t t
u v
u u uv p
n
v v uv p
e uh vh
ρ ρ θ ρ θ
ρ ρ θ ρ θ θ
ρ ρ θ ρ θ θ
ρ ρ θ ρ θ
+ ⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥
+ +
⎢ ⎥ ⎢ ⎥
= =
⎢ ⎥ ⎢ ⎥ + +
⎢ ⎥ ⎢ ⎥
+
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
U F
 (3.3)
In this equation θ is the angle between the normal direction of the edge and the
Cartesian xdirection. The variables , , , ,
t
u v p e ρ and
t
h are the density, velocity in
41
the x and y direction, the pressure, the total energy and the total enthalpy
respectively.
In the solution method used, determination of the fluxes through the faces of the cells
require a pseudo one dimensional form of the flux vector represented in Equation
(3.3). This form will be equivalent to the Euler equations written in terms of the
normal and tangential velocities. At this point, the rotational invariance of the Euler
equations (16) can be employed as
1 1
ˆ ˆ
( ) ( ) n
− −
= = F T F TU T F U
 (3.4)
The transformation matrix and its inverse; T and
1
T , appearing in Equation (3.4)
can be written explicitly as
1
1 0 0 0 1 0 0 0
0 cos sin 0 0 cos sin 0
0 sin cos 0 0 sin cos 0
0 0 0 1 0 0 0 1
θ θ θ θ
θ θ θ θ
−
⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥
−
⎢ ⎥ ⎢ ⎥
= =
⎢ ⎥ ⎢ ⎥ −
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
T T (3.5)
Hence, the conserved quantities and the flux vector in terms of normal and
tangential velocities can be given
2
ˆ
ˆ ˆ
ˆ ˆ ˆ
( )
ˆ ˆ ˆ
ˆ
t t
u
u u p
v uv
e uh
ρ ρ
ρ ρ
ρ ρ
ρ ρ
⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥
+
⎢ ⎥ ⎢ ⎥
= =
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
U F U (3.6)
In Equation (3.6) ˆ u and ˆ v are the normal and tangential velocities, while
ˆ
U and
ˆ
Fare
the vectors of conserved variables and the flux, in terms of the normal and tangential
velocities. The normal and the tangential velocities can be expressed in terms of the
Cartesian velocities as
ˆ ˆ cos sin , sin cos u u v v u v θ θ θ θ = + = − + (3.7)
42
Using this transformation, the governing equations can be expressed as
1
1
ˆ ˆ
0
edges
cell
s
t A
−
∂
+ ∆ =
∂
∑
U
T F(U) (3.8)
The general solution strategy is to estimate the flux in terms of the normal and
tangential directions, rotate them to Cartesian coordinates with the inverse
transformation matrix for each face, and then use an appropriate method for time
integration.
3.2 One Dimensional Riemann Problem for Linear Systems
Before discussing the solution algorithm for the Euler equations, the one dimensional
Riemann problem for a linear system of equations should be discussed (17) (16). The
model equation for the one dimensional linear system is represented by Equation
(3.9). This equation is linear if the Jacobian of the flux vector is comprised of
constant elements, that is
0
t x t x t x
∂ ∂ ∂ ∂ ∂ ∂ ∂
+ = + = + =
∂ ∂ ∂ ∂ ∂ ∂ ∂
U F U F U U U
A
U
(3.9)
where Ais the Jacobian matrix. The initial conditions for the model equation are
given as
0
0
0
( , 0)
0
x
x
x
< ⎧
=
⎨
>
⎩
L
R
U
U
U
(3.10)
where subscripts L and R refer to the left and right states respectively. The first step
in the solution process is to diagonalize the Jacobian matrix using the Q matrix
appearing in the following equation. Each column of this matrix is a right
eigenvector of the Jacobian matrix and the rows of its inverse are the left
eigenvectors of the Jacobian matrix (18). The elements of the resulting diagonal
matrix Λare the eigenvalues of the Jacobian matrix.
43
−1
Q AQ= Λ (3.11)
Using this diagonalization and multiplying Equation (3.9) by Q
1
, the linear set of
equations can be converted to a set of independent equations in terms of the new
characteristic variable vector V .
0
t x t x
∂ ∂ ∂ ∂
+ = + =
∂ ∂ ∂ ∂
1 1 1
U U V V
Q Q AQQ Λ (3.12)
The relation between the characteristic variables and the conserved quantities in
Equation (3.12) can be defined as
∂ = ∂
1
Q U V (3.13)
Since this is a linear system and the elements of Q
1
are constant, Equation (3.13)
can
be written in an alternative form as
= ⇒ ∆ = ∆
1 1
Q U V Q U V (3.14)
where;
∆ = − ∆ = −
R L R L
V V V U U U
After the diagonalization, the governing equations obtained in terms of the
characteristic variables, λ ’s, in Equation (3.12) can be written explicitly as
1 1 1
0 . . 0
. 0 . . .
. . . . . 0
. . . 0 .
0 . . 0
n n n
t x
v v
v v
λ
λ
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ + =
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
(3.15)
44
The equations represented by Equation (3.15) are n independent equations for the n
characteristic variables. In other words, n independent scalar equations are obtained
for n characteristic variables, ' s v , (17) as
1,...,
i i
i
v v
i n
t x
λ
∂ ∂
+ =
∂ ∂
(3.16)
For the solution of each one of the equations in Equation (3.16), the definition of the
total derivative is employed as
d d
d d
i i i
v v v x
t t x t
∂ ∂
= +
∂ ∂
(3.17)
Hence with the application of the condition presented in Equation (3.18), a straight
line along which the characteristic variable remains constant and the governing
equation is satisfied, is obtained.
0
d d
0 along line
d d
i
i i
v x
x x t
t t
λ λ = ⇒ = = + (3.18)
The solution obtained in terms of the characteristic lines can be represented
graphically as in Figure 3.1.
Figure 3.1 Solution of a 1D Riemann problem for a set of linear equations
x=λ
1
t
x=λ
n
t
45
In Figure 3.1, when one of the characteristic lines is crossed the respective value of
the characteristic variable changes conforming to the given initial conditions. The
form in which the solution is presented is of crucial importance in the development
of numerical methods while it doesn’t make a difference in the analytical sense.
Hence the solution can be presented in several formats (17). First representation is in
terms of the characteristic variables as
1
2 1 2
1
... / ...
... / ...
. .
. .
. .
... ... /
n
n
n
x t
x t
x
t
x t
λ λ
λ λ λ
λ λ
= −∆ − −∆ < < < ⎧
⎪
+ ∆ = −∆ − −∆ < < < <
⎪
⎪
⎪ ⎛ ⎞
=
⎨
⎜ ⎟
⎝ ⎠
⎪
⎪
⎪
+ ∆ + + ∆ = < < <
⎪
⎩
L R 1 n
L 1 R n
L 1 n R
V V v v
V v V v v
V
V v v V
(3.19)
The difference vectors appearing in Equation (3.19) can be written explicitly as
1
1
0 0
. 0 .
. . .
...
. .
0 . .
0 0
n
i
n
v
v v
v
v
∆ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
∆ = ∆ = ∆ = ⇒ ∆ + + ∆ = ∆
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
∆
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
∆
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
1 i n
v v v V (3.20)
It is possible to express the solution in terms of the conserved variables as well by
employing the relation between the characteristic variables and the conserved
variables given in Equation (3.14). Here it should be remembered that each column
of matrix Q is an eigenvector of the Jacobian matrix. This will promote the relation
presented in the following equation
i
v ∆ = ∆
i i
Q v r (3.21)
where r
i
is the i
th
eigenvector of the Jacobian matrix.
46
Hence by multiplying Equation (3.19) by the matrix Q, the solution in terms of the
conserved variables is obtained as
1 1
1 2 2 1 2
1 1
... / ...
... / ...
. .
. .
. .
... ... /
n n
n n
n n
v v x t
v v v x t
x
t
v v x t
λ λ
λ λ λ
λ λ
= − ∆ − − ∆ < < < ⎧
⎪
+ ∆ = − ∆ − − ∆ < < < <
⎪
⎪
⎪ ⎛ ⎞
=
⎨
⎜ ⎟
⎝ ⎠
⎪
⎪
⎪
+ ∆ + + ∆ = < < <
⎪
⎩
L R 1 n
L 1 R n
L 1 n R
U U r r
U r U r r
U
U r r U
(3.22)
At this point, it should be noted that for the numerical method that is used for the
solution of the Euler equations, the flux at a face of a certain cell will be of interest
rather than the value of the conserved variable. It should also be noted that the flux
vector appearing in the differential model equation is equivalent to the flux term that
would appear in its integral formulation. Also the fluxes of each face will be
investigated separately so the / can be considered to be zero for the respective face
of the cell. Taking these points into consideration one can derive expressions for the
fluxes as a solution.
( )
1 1 1
1 1 2 2 2 1 2
1 1 1
... 0 ...
... 0 ...
. .
0
. .
. .
... ... 0
n n n
n n n
n n n
v v
v v v
v v
λ λ λ λ
λ λ λ λ λ λ
λ λ λ λ
= − ∆ − − ∆ < < < ⎧
⎪
+ ∆ = − ∆ − − ∆ < < < <
⎪
⎪
⎪
=
⎨
⎪
⎪
⎪
+ ∆ + + ∆ = < < <
⎪
⎩
L R 1 n
L 1 R n
L 1 n R
AU AU r r
AU r AU r r
F
AU r r AU
(3.23)
Equation (3.23) has been obtained by exploiting the fact that
i
λ =
i i
Ar r .
Even though Equation (3.23) is a perfectly valid representation of the solution, it is
desirable to present the solution in a more compact form as in one of the following
equivalent forms:
47
1
1
(0)
(0)
n
i i
i
n
i i
i
v
v
λ
λ
−
=
+
=
= + ∆
= − ∆
∑
∑
L i
R i
F AU r
F AU r
(3.24)
where;
min(0, )
max(0, )
i i
i i
λ λ
λ λ
−
+
=
=
Since the representation is dependent on the sign of the eigenvalues in Equation
(3.24), a final form for the solution can be obtained by averaging the equivalent
solutions appearing in Equation (3.24).
[ ]
1
1
( ) ( )
2
1 1
(0) ( )
2 2
n
i i
i
v λ
=
+
= + − ∆
∑
L R
L R i
F U F U
F A U U r
.
(3.25)
3.3 The Approximate Riemann Solver of Roe
The flux term in the governing Euler equations in the integral form (3.1) can be
estimated by considering the xsplit form of the two dimensional Euler equations in
the differential form as
0
t x y
∂ ∂ ∂
+ + =
∂ ∂ ∂
U F G
(3.26)
where;
1 1 1
2
2 2 2
2
3 3 3
4 4 4
t t t
u v u f g
u u p uv u f g
v uv v p u f g
e uh vh u f g
ρ ρ ρ
ρ ρ ρ
ρ ρ ρ
ρ ρ ρ
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
+
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
= = = = = =
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ +
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦
U F G
48
As can be observed the flux vector in Equation (3.26) is identical to the one
appearing in the transformed form of the governing equations given by Equation
(3.6). The theory of the Riemann problem for one dimensional linear systems will be
employed to estimate this flux.
To obtain the Jacobian for the xsplit Euler equations, the flux vector must be written
in terms of the conserved variables as
( )
( )
2
2 2 2
3 2 2
4
1 1 1
2 3
1
2 3
3 2 2 4 2 4 2
2 2
1 1 1 1
1
2 2
1
2 2
u
u u u
u
u u u
u u
u
u u u u u u u
u u u u
γ
γ
⎡ ⎤
⎢ ⎥
⎛ ⎞
⎢ ⎥
+ − − −
⎜ ⎟
⎢ ⎥
⎝ ⎠
⎢ ⎥
=
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎛ ⎞
+ − − −
⎢ ⎥ ⎜ ⎟
⎢ ⎥
⎝ ⎠ ⎣ ⎦
F (3.27)
In Equation (3.27) γ is the specific heat ratio. Noting that the Jacobian matrix is
specified in the following form
1 1 1 2 1 3 1 4
2 1 2 2 2 3 2 4
3 1 3 2 3 3 3 4
4 1 4 2 4 3 4 4
/ / / /
/ / / /
/ / / /
/ / / /
f u f u f u f u
f u f u f u f u
f u f u f u f u
f u f u f u f u
∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ⎡ ⎤
⎢ ⎥
∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
∂
⎢ ⎥
= =
⎢ ⎥ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
⎢ ⎥
∂ ∂ ∂ ∂ ∂ ∂ ∂ ∂
⎢ ⎥
⎣ ⎦
F
A
U
(3.28)
it can be obtained by equations (3.27) and (3.28) in terms of the flow variables.
( ) ( ) ( )
( ) ( ) ( )
2 2
2 2
0 1 0 0
1 3 1 1
0
2 1 1
t
t t
h u a u v
uv v u
u h a h u uv u
γ γ γ γ
γ γ γ γ
⎡ ⎤
⎢ ⎥
− − − − − −
⎢ ⎥
=
⎢ ⎥ −
⎢ ⎥
⎡ ⎤ − − − − − −
⎢ ⎥
⎣ ⎦ ⎣ ⎦
A (3.29)
In Equation (3.29) a is the speed of sound. As can be observed from Equation (3.29),
the Jacobian matrix does not have constant elements hence the Euler equations are
49
not linear. For the approximate Riemann solvers this Jacobian matrix is replaced by
one that has constant elements since for a nonlinear system the characteristic lines of
the same family will not be parallel introducing shock waves and expansion fans into
the solution (17), (16).
In Roe’s approximate Riemann solver (19), the Jacobian matrix is expressed in terms
of an intermediate state defined from the right and left states.
First, it is observed that both the flux vector F and the vector of conserved quantities
U are quadratic in terms of a new vector W defined as
1
2
3
4
1
t
w
u w
v w
h w
ρ
⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
= =
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
W (3.30)
Another useful vector, W, is formed by taking the arithmetic mean of the W vector
for the right and left states.
1
2
3
4
2
2
2
2
L R
L L R R
L L R R
L tL R tR
w
u u
w
w
v v
w
h h
ρ ρ
ρ ρ
ρ ρ
ρ ρ
⎡ ⎤
+
⎢ ⎥
⎢ ⎥
⎢ ⎥ ⎡ ⎤
+
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
= =
⎢ ⎥ ⎢ ⎥
+
⎢ ⎥ ⎢ ⎥
⎢ ⎥
⎣ ⎦ ⎢ ⎥
⎢ ⎥
+
⎢ ⎥
⎢ ⎥
⎣ ⎦
W (3.31)
It is now possible to express the difference between the right and the left states for
the conserved quantities in terms of the new vector defined as
∆ = ∆ U B W (3.32)
where;
50
1
2 1
3 1
4 1
2 3
2 0 0 0
0 0
0 0
1 1
w
w w
w w
w w
w w
γ γ
γ γ γ γ
⎡ ⎤
⎢ ⎥
⎢ ⎥
= ⎢ ⎥
⎢ ⎥
− −
⎢ ⎥
⎢ ⎥
⎣ ⎦
B
It is also possible to express the differential flux vector in terms of the vector W in
the same manner as
∆ = ∆ F C W (3.33)
where;
2 1
4 2 3 1
3 2
4 2
0 0
1 1 1 1
0 0
0 0
w w
w w w w
w w
w w
γ γ γ γ
γ γ γ γ
⎡ ⎤
⎢ ⎥
− + − −
⎢ ⎥
−
= ⎢ ⎥
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎣ ⎦
C
Using Equations (3.32) and (3.33) it is possible to establish a direct relation between
the difference in the flux vector and the vector of conserved quantities in the form
∆ = ∆
1
CB U F (3.34)
Using the homogeneity property of the Euler equations given by the following
equation (16) one can recognize the product CB
1
as the Jacobian matrix.
∂
=
∂
A
F
F U
U
(3.35)
In order to find the eigenvalues of the Jacobian matrix, the following equation must
be solved.
51
( ) ( ) det 0 det 0 λ λ − = ⇒ − =
1
CB I C B (3.36)
Dividing Equation (3.36) by
1
w , it is possible to find the eigenvalues of the Jacobian
in terms of the new variables defined as
2
1
L L R R
L R
u u w
u
w
ρ ρ
ρ ρ
+
= =
+
¯ (3.37)
3
1
L L R R
L R
v v w
v
w
ρ ρ
ρ ρ
+
= =
+
¯ (3.38)
4
1
L tL R tR
t
L R
h h w
h
w
ρ ρ
ρ ρ
+
= =
+
¯
(3.39)
( ) ( )
2 2
1 1
1 1
2 2
t t
a h V e V γ γ γ
⎛ ⎞ ⎛ ⎞
= − − = − −
⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠
¯ ¯ ¯
¯ ¯ (3.40)
L R
ρ ρ ρ = ¯ (3.41)
1 2 3 4
u a u u a λ λ λ λ = − = = = +
¯ ¯ ¯ ¯
¯ ¯ ¯ ¯ ¯ (3.42)
The corresponding right eigenvectors can also be expressed in terms of the new Roe
averaged quantities as
2 3 4
2
1
1 1
0
0
1
1
2
t t
u
u a u a
v
v v
v V h ua h ua
⎡ ⎤
⎡ ⎤ ⎡ ⎤
⎡ ⎤
⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥
− + ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥
= = = =
⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥
⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥
⎢ ⎥
− +
⎢ ⎥ ⎢ ⎥ ⎣ ⎦
⎣ ⎦ ⎣ ⎦
⎢ ⎥
⎣ ⎦
1
r r r r
¯
¯ ¯ ¯ ¯
¯
¯ ¯
¯ ¯ ¯
¯ ¯ ¯ ¯ ¯
(3.43)
52
With the right eigenvectors defined as in Equation (3.43), it is possible to determine
the characteristic variables which can also be termed as the wave strengths by using
Equation (3.14).
1 1
2 2
3 3
2
4 4
1
1 1
0
0
1
1
2
t t
v u
u
u a u a
v u
v
v v
v u
v u v V h ua h ua
⎡ ⎤
∆ ∆ ⎡ ⎤ ⎡ ⎤
⎢ ⎥
⎢ ⎥ ⎢ ⎥
− + ⎢ ⎥
∆ ∆
⎢ ⎥ ⎢ ⎥
=
⎢ ⎥
⎢ ⎥ ⎢ ⎥ ∆ ∆
⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥
∆ ∆ − + ⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
⎢ ⎥
⎣ ⎦
¯
¯ ¯ ¯ ¯
¯
¯ ¯
¯ ¯ ¯
¯ ¯ ¯ ¯ ¯
(3.44)
Due to the proportionality between the coefficients of the first and third relations in
Equation (3.44), it is possible to eliminate the third equation by using the first one to
obtain the following relation.
3 3 1
v u v u ∆ = ∆ − ∆ ¯ (3.45)
Then a set of three equations is obtained for three wave strengths as
4 1
1 2
2
2 2 4 1 3
1 1 1
1
2
t t
v u
u a u a u v u
v u v u v u
h ua h ua V
⎡ ⎤
⎢ ⎥
∆ ∆ ⎡ ⎤ ⎡ ⎤
⎢ ⎥
⎢ ⎥ ⎢ ⎥
+ − ∆ = ∆
⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥
⎢ ⎥ ⎢ ⎥ ∆ ∆ + ∆ − ∆
⎣ ⎦ ⎣ ⎦
+ −
⎢ ⎥
⎣ ⎦
¯ ¯ ¯ ¯ ¯
¯ ¯
¯ ¯ ¯
¯ ¯ ¯ ¯
(3.46)
It is possible to solve this system using Gauss elimination to obtain the wave
strengths as
( )
( )
2
2 1 2 4 3 1 2
1
t
v u h u u u u v u v u
a
γ −
⎡ ⎤
∆ = ∆ − + ∆ − ∆ + ∆ − ∆
⎣ ⎦
¯
¯ ¯ ¯ ¯
¯
(3.47)
( )
1 1 2 2
1
2
v u a u u a v
a
∆ = + ∆ − ∆ − ∆ ⎡ ⎤
⎣ ⎦
¯ ¯ ¯
¯
(3.48)
4 1 1 2
( ) v u v v ∆ = ∆ − ∆ +∆ (3.49)
53
Defining an operator given by the following equation
( )
2
( ) pq p q q p O ∆ = ∆ + ∆ + ∆ ¯ ¯ (3.50)
neglecting higher order terms and placing the values of the conserved quantities, the
wave strengths can be expressed in terms of the flow variables as
( )
( )
( )
2 2
2 2
2 2
1
1
t t t
t t t
v h u u u u e e v v v v
a
v h e u u v v e
a
γ
ρ ρ ρ ρ ρ ρ ρ ρ
γ
ρ ρ ρ ρ
−
⎡ ⎤
∆ = ∆ − + ∆ + ∆ − ∆ − ∆ + ∆ + ∆ − ∆
⎣ ⎦
−
⎡ ⎤
⇒ ∆ = ∆ − + ∆ + ∆ − ∆
⎣ ⎦
¯
¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯
¯
¯
¯ ¯ ¯ ¯ ¯ ¯
¯
(3.51)
The terms appearing in Equation (3.51) must be treated individually to simplify the
expression. For the treatment of the first term in the brackets, Equation (3.40) can be
employed as
( )
( )
2 2 2
2 2
1 1
1 2 1 2
t t
a a a
h e V V ρ ρ ρ
γ γ γ γ
⎛ ⎞
∆ − = ∆ + − − = ∆
⎜ ⎟
⎜ ⎟
− −
⎝ ⎠
¯ ¯ ¯
¯ ¯ ¯
¯ (3.52)
The last term in the brackets must also be treated individually with the aid of the
relation
2
1
1 2
t
p V
e
γ ρ
= +
−
(3.53)
To obtain
( )
( )
2
2 2
1 1
1 2 1 1
t
p p a
e u v u u v v
ρ
ρ ρ ρ ρ
γ ρ γ γ γ
⎡ ⎤ ⎛ ⎞ ∆ ∆
∆ = ∆ + ∆ + = − + ∆ + ∆
⎢ ⎥ ⎜ ⎟
− − −
⎝ ⎠ ⎣ ⎦
¯
¯ ¯ ¯ ¯ ¯ ¯ (3.54)
Substituting Equations (3.52) and (3.54) into Equation (3.51), a simplified equation
for the wave strength can be obtained. Then applying the operator defined in
54
Equation (3.50) and substituting Equation (3.57) into Equations (3.48) and (3.49) all
the wave strengths can be found as
( )
1 2
1
2
v p a u
a
ρ ∆ = ∆ − ∆ ¯ ¯
¯
(3.55)
2 2
p
v
a
ρ
∆
∆ = ∆ −
¯
(3.56)
3
v v ρ ∆ = ∆ ¯
(3.57)
( )
4 2
1
2
v p a u
a
ρ ∆ = ∆ + ∆ ¯ ¯
¯
(3.58)
In summary, for the approximate Riemann solver of Roe, first the Jacobian matrix
for the Euler equations, which are non linear in their true essence, have been replaced
by another matrix with constant elements to approximate the governing equations as
a linear system. Thereafter the eigenvalues and eigenvectors of this approximate
Jacobian matrix are found by following the procedure outlined in (19). Then these
are used to find the wave strengths, characteristic variables, and fluxes through a face
for which the right and left states are known according to Equation (3.25).
3.3.1 Solution Algorithm
The solution algorithm adopted is similar to the one outlined in (16). The algorithm
discussed here is for the determination of the flux through a certain face of the cell
under investigation. Once the fluxes through all the faces are determined, an
appropriate method for time integration, to be discussed later, is employed to update
the values of the conserved variables.
(i) First the conserved variables of the left and right states are obtained by
reconstruction which will be explained later on.
55
(ii) The primitive variables, [ ]
T
t
u v P h ρ for the right and left states are
determined from the conserved variables [ ]
T
t
u v e ρ ρ ρ ρ
simultaneously with the transformation to normal and tangential coordinates.
The flux which will be used in Equation (3.25), is calculated from the left and
right state variables.
cos sin
ˆ
u v
u
ρ θ ρ θ
ρ
+
= (3.59)
sin cos
ˆ
u v
v
ρ θ ρ θ
ρ
− +
= (3.60)
( )
2 2
( ) ( )
1
t
u v
p e
ρ ρ
γ ρ
ρ ρ
⎡ ⎤
= − − −
⎢ ⎥
⎣ ⎦
(3.61)
( )
1
t t
h e p ρ
ρ
= + (3.62)
(iii) The Roe averaged quantities for the transformed primitive variables and the
eigenvalues are calculated from Equations (3.37) through (3.42).
(iv) The wave strengths are calculated from Equations (3.55) through (3.58). The
differences are taken as the right minus the left state variables in contrast to
Roe’s derivation in (19) to be consistent with Equation (3.25).
(v) The right eigenvectors are calculated using the transformed Roe averaged
quantities with the formulation provided in Equation (3.43).
(vi) The flux is calculated from Equation (3.25).
(vii) The flux obtained, which is normal to the face, is transformed back to the
Cartesian coordinates using the inverse transformation matrix.
ˆ
=
1
T F F (3.63)
where T
1
is given in Equation (3.5).
56
3.4 Flux Vector Splitting
In addition to using Roe’s approximate Riemann solver to evaluate the flux term
appearing in Equation (3.8), the flux vector splitting methods such as van Leer (20),
AUSM (21) and its derivatives AUSMD, AUSMV (22) have also been employed.
Considering Equation (3.9) it is possible to split the flux vector as
( ) ( ) ( ) = +
+ 
F U F U F U (3.64)
The negative and positive components of the flux vector,
+
F and

F , can be defined
by splitting the Jacobian matrix as follows
( ) , ( )
− −
= =
+ +
F U A U F U A U (3.65)
,
− −
= =
+ + 1 1
A QΛ Q A QΛ Q (3.66)
The diagonal matrices
+
Λ and
−
Λ can be defined as
1 1
0 . . 0 0 . . 0
0 . . 0 . .
, . . . . . .
. . 0 . . 0
0 . . 0 0 . . 0
m m
λ λ
λ λ
+ −
−
+ −
⎡ ⎤ ⎡ ⎤
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
= =
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
+
Λ Λ (3.67)
where the positive and negative eigenvalues can be defined as in (23) presented as in
the following equation.
( ) ( )
1 1
,
2 2
λ λ λ λ λ λ
+ −
= + = − (3.68)
When the Euler equations are written in normal and tangential coordinates as in
Equation (3.8) the positive flux component will be due to the cell itself while the
57
negative component will be calculated from the neighbour as illustrated in Figure
3.2.
Figure 3.2 Graphical interpretation of flux vector splitting
In Figure 3.2 the left state is the cell under investigation while the right state is the
neighbour. The positive eigenvalues of the left state affects the face hence the
positive component of the flux is calculated from the left state while the negative
component of the flux is calculated from the right state since the negative
eigenvalues of this state affects the face. Further details can be found in (16).
3.4.1 Van Leer Flux Vector Splitting
The van Leer flux vector splitting algorithm (20) makes use of the flux represented in
terms of the density, tangential velocity, Mach number,
ˆ
M , and the speed of sound
as
( )
2 2
2
2 2 2
ˆ
1
ˆ
ˆ
ˆ ˆ ˆ
( ) ,
ˆ
ˆ
1
ˆ ˆ
ˆ
2 1
aM
a M
u
M
a
aMv
a
aM a M v
ρ
ρ
γ
ρ
ρ
γ
⎡ ⎤
⎢ ⎥
⎛ ⎞
⎢ ⎥
+
⎜ ⎟
⎢ ⎥
⎝ ⎠
⎢ ⎥
= =
⎢ ⎥
⎢ ⎥
⎡ ⎤
⎢ ⎥
+ +
⎢ ⎥
⎢ ⎥
−
⎢ ⎥
⎣ ⎦ ⎣ ⎦
F U (3.69)
Left Right
n
t
58
The flux can then be split into its positive and negative components as presented in
the following equations.
( )
2
2
2
2
2
1
2 1
ˆ
1
2
1
ˆ ˆ
1
ˆ
4
2 1 1
ˆ
ˆ 1
1 2 2
a
M
a M
v
a
M v
γ
γ
ρ
γ
γ
⎡ ⎤
⎢ ⎥
− ⎛ ⎞
⎢ ⎥
+
⎜ ⎟
⎢ ⎥
⎝ ⎠
= +
⎢ ⎥
⎢ ⎥
⎢ ⎥
− ⎛ ⎞
⎢ ⎥ + +
⎜ ⎟
− ⎢ ⎥
⎝ ⎠
⎣ ⎦
+
F (3.70)
( )
2
2
2
2
2
1
2 1
ˆ
1
2
1
ˆ ˆ
1
ˆ
4
2 1 1
ˆ
ˆ 1
1 2 2
a
M
a M
v
a
M v
γ
γ
ρ
γ
γ
−
⎡ ⎤
⎢ ⎥
− ⎛ ⎞
⎢ ⎥
−
⎜ ⎟
⎢ ⎥
⎝ ⎠
= − −
⎢ ⎥
⎢ ⎥
⎢ ⎥
− ⎛ ⎞
⎢ ⎥ − +
⎜ ⎟
− ⎢ ⎥
⎝ ⎠
⎣ ⎦
F (3.71)
3.4.2 AUSM (LiouSteffen) Flux Vector Splitting
The AUSM scheme (21) splits the pressure and the advection terms appearing in the
normal momentum flux as
( ) ( )
1/ 2 1/ 2 1/ 2
1
ˆ ˆ ˆ ˆ
( )
2
L R R L
M M
⎡ ⎤
= + − − +
⎣ ⎦
F U φ φ φ φ p (3.72)
In equation (3.72),
1/ 2
ˆ
M is the interface Mach number,
1/ 2
p is the vector of interface
pressure and φ is the vector of flux defined through the following equations.
1/ 2
ˆ ˆ ˆ
L R
M M M
+ −
= + (3.73)
59
1/ 2
1/ 2 1/ 2
0
,
0
0
L R
p
p p p
+ −
⎡ ⎤
⎢ ⎥
⎢ ⎥
= = +
⎢ ⎥
⎢ ⎥
⎣ ⎦
p (3.74)
ˆ
ˆ
t
a
ua
va
h a
ρ
ρ
ρ
ρ
⎡ ⎤
⎢ ⎥
⎢ ⎥
=
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎣ ⎦
φ (3.75)
The split Mach numbers and the split pressures can be defined as in the following
equations to complete the algorithm.
( )
( )
2
1
ˆ ˆ
1 1
4
ˆ
1
ˆ ˆ ˆ
1
2
M M
M
M M M
±
⎧
± ± ≤
⎪
⎪
=
⎨
⎪
± >
⎪
⎩
(3.76)
ˆ ˆ
2 1
ˆ
1
ˆ
1
ˆ
M M
p pM
M
M
± ±
⎧
± − ≤
⎪
= ⋅
⎨
> ⎪
⎩
(3.77)
3.4.3 AUSMD Flux Vector Splitting
AUSMD flux vector splitting (22) is a derivative of the original AUSM scheme. As
in the AUSM scheme the advection and pressure terms are split separately with the
difference that instead of estimating the normal velocity at the interface the mass flux
is estimated. Also the velocity splitting is modified in the AUSMD method as in the
equations to follow.
( ) ( ) ( ) ( )
1/ 2
1/ 2 1/ 2
1
ˆ ˆ
ˆ ˆ ( )
2
L R R L
u u ρ ρ
⎡ ⎤
= + − − +
⎣ ⎦
F U Γ Γ Γ Γ p (3.78)
( )
1/ 2
ˆ ˆ ˆ
L L R R
u u u ρ ρ ρ
+ −
= + (3.79)
60
( )
( )
2
max
max
max
max
ˆ ˆ ˆ
ˆ 1
4 2
ˆ
ˆ ˆ
ˆ
2
L L L
L L L
L
L L
L
u u u a
u a
a
u
u u
u a
κ κ
+
⎧ ⎡ ⎤
+ +
⎪ + − ≤ ⎢ ⎥
⎪
⎢ ⎥
⎣ ⎦
=
⎨
⎪
+
>
⎪
⎩
(3.80)
( )
( )
2
max
max
max
max
ˆ ˆ ˆ
ˆ 1
4 2
ˆ
ˆ ˆ
ˆ
2
R R R
R R R
R
R R
R
u u u a
u a
a
u
u u
u a
κ κ
−
⎧ ⎡ ⎤
− −
⎪ − + − ≤ ⎢ ⎥
⎪
⎢ ⎥
⎣ ⎦
=
⎨
⎪
−
>
⎪
⎩
(3.81)
max
2( / ) 2( / )
, , max( , )
( / ) ( / ) ( / ) ( / )
L R
L R L R
L R L R
p p
a a a
p p p p
ρ ρ
κ κ
ρ ρ ρ ρ
= = =
+ +
(3.82)
Naturally the vector of flux is modified as well as the split pressure as
1
ˆ
ˆ
t
u
v
h
⎡ ⎤
⎢ ⎥
⎢ ⎥
=
⎢ ⎥
⎢ ⎥
⎢ ⎥
⎣ ⎦
Γ (3.83)
1/ 2 L R
p p p
+ −
= + (3.84)
( )
2
max
max 2
max max
max
ˆ ˆ
ˆ 2
4
ˆ ˆ
ˆ
ˆ 2
L L
L L
L
L L
L L
L
u a
u
p u a
a a
p
u u
p u a
u
+
⎧
+ ⎛ ⎞
− ≤ ⎪
⎜ ⎟
⎪
⎝ ⎠
=
⎨
+
⎪
>
⎪
⎩
(3.85)
( )
2
max
max 2
max max
max
ˆ ˆ
ˆ 2
4
ˆ ˆ
ˆ
ˆ 2
R R
R R
R
R R
R R
R
u a
u
p u a
a a
p
u u
p u a
u
−
⎧
− ⎛ ⎞
+ ≤ ⎪
⎜ ⎟
⎪
⎝ ⎠
=
⎨
−
⎪
>
⎪
⎩
(3.86)
61
3.4.4 AUSMV Flux Vector Splitting
Another derivative of the AUSM scheme is the AUSMV flux vector splitting scheme
(22). Similar to the AUSMD scheme the mass flux at the interface of two cells are
estimated as in Equations (3.78) and (3.79) in the AUSMV scheme except for the
normal momentum. The normal momentum is estimated at the interface wholly as
( ) ( ) ( )
2
1/ 2
ˆ ˆ ˆ ˆ ˆ
L R
L R
u u u u u ρ ρ ρ
+ −
= + (3.87)
The definitions of the split velocities and the pressures differs from the AUSMD
scheme as can be observed from the following equations
( )
2
ˆ
ˆ
4
ˆ
ˆ ˆ
ˆ
2
u a
u a
a
u
u u
u a
±
⎧
±
± ≤ ⎪
⎪
=
⎨
±
⎪
>
⎪
⎩
(3.88)
ˆ 1
ˆ 2
ˆ
1
ˆ
ˆ
u
u a
a a
p pu
u a
u
± ±
⎧ ⎛ ⎞
± − ≤
⎜ ⎟ ⎪
⎪
⎝ ⎠
= ⋅
⎨
⎪
>
⎪
⎩
(3.89)
3.5 Boundary Conditions
The boundary conditions encountered for the external flows considered in the scope
of the thesis work are the solid wall and the farfield boundary conditions. For the
numerical method adopted, the implementation of either type of the boundary
conditions is equivalent to determining the correct right state variables since the left
state variables will always be dictated by the cell under investigation.
For the solid wall, the physical boundary condition is that there is no flux into the
body. In the inviscid case considered, this means that the velocity at the face can be
in the tangential direction.
62
At this point, one of two methods can be adopted for the determination of the right
state variables (14). One is to directly take the flux through the face as zero and to
calculate a right flux solely based on pressure that is equal to the pressure calculated
for the left state. The reason for this equality is that under steadystate conditions, the
pressure force exerted by the fluid will be countered with an equal force. The second
method is to construct a ghost right state for which the tangential velocity, pressure
and total enthalpy are equal to the left state. The direction of the tangential velocity is
the same as the left state as well as the magnitude. For the normal velocity, the
magnitude is equal to that of the left state but the direction is opposite to it. The ghost
right state is depicted for an example cutcell in Figure 3.3.
Figure 3.3 Ghost right state at the solid wall
The second method ensures zero flux through the face as well but it also takes into
cases where there is a normal velocity component in the presence of high curvatures
as stated in (14). In the code developed, the second method has been adopted for the
solid wall treatment.
For the farfield boundary conditions the right state is calculated from the conditions
received as input and it is irrespective of the left state.
3.6 Reconstruction of the Flow Variables
For the calculation of the fluxes through a face, primitive flow variables have to be
estimated on both sides of that face, the right and left states. The simplest approach
v
R
u
R
v
L
u
L h
tL h
tR
= h
tL
p
L
p
R
= p
L
63
would be to take the values from the cell under investigation as the left state and to
take the values from the respective neighbour as the right state. This approach being
first order does not produce very accurate results. Hence it is desirable to reconstruct
the flow variables within the cell to obtain a solution of second order. That is the
variation of the flow variables within the cell must be taken into account even though
in finite volume methods the flow variables representative of the whole cell is stored
at the cell centroid for higher order accuracy.
Three reconstruction schemes have been sighted in the literature. In two of these
three schemes, the gradient of the flow variable is estimated at the centre of the cell
while the other one makes use of finite differences to estimate the value of a flow
variable at certain points within the cell. One of the reconstruction schemes based on
the gradient at the centre of the cell is the GreenGauss reconstruction, the other one
is the Least Squares (Minimum Energy) reconstruction. The general design criteria
for these reconstruction schemes are discussed in (24) and (25). The details of the
linear reconstruction based on these schemes can be found in (26), (27) and (14)
while the k
th
order exact schemes are discussed in (28). The third scheme based on
the finite difference method can be found in (29).
3.6.1 GaussGreen Reconstruction
The GaussGreen reconstruction is based mainly on the gradient theorem. For a
quantity u related to the cell this can be formulized as
1
u und
A
Ω Ω
∇ =
∫
/ (3.90)
This integral is evaluated for a region Ω which is bounded by the lines that connect
the centroids of the edge and vertex neighbours (support set) of the cell. The integral
can be evaluated numerically using the trapezoidal rule given as
sup
1
0
1 1
( )
2
N
i i i i
i
cell
u u u n
A
+
=
∇ = +
∑
/ (3.91)
64
The boundary of the region Ω is shown for two representative cases for a Cartesian
mesh in Figure 3.4.
Figure 3.4 Region formed by the support set
3.6.2 Least Squares (Minimum Energy) Reconstruction
The Least Squares reconstruction is basically the inverse of the averaging function
for a quantity related to a cell. The average of a quantity u within a cell can be
defined with the averaging function presented as
( )
cell
A
cell
udA
Av u u
A
= =
∫
(3.92)
A k
th
order polynomial u
k
is constructed to represent the quantity u within the cell
such that the average of the polynomial will be equal to that of the cell and the
polynomial will represent quantity u exactly if u is a polynomial of order k or less. In
general, the constructed polynomial can be obtained as
( )
k m n m n
mn
m n k
u u x y Av x y α
+ ≤
⎡ ⎤ = + −
⎣ ⎦
∑
(3.93)
where
mn
α are arbitrary coefficients.
65
The form given in Equation (3.93) will ensure that the average of the polynomial is
equal to that of the cell. For a linear case, this reconstruction takes the form presented
in the following equation
[ ] [ ]
1
10 01
( , )
c c
u x y u x x y y α α = + − + − (3.94)
In Equation (3.94), the centroidal coordinates of the cell appear since by definition
the averaging function applied to the polynomials x and y within the cell gives the
centroid coordinates of the cell.
The reconstruction problem simplifies once the cell average values are thought to be
the values of the flow variables stored at the centroid of the cell (24). Then one could
get the expression of the following form for the linear reconstruction.
[ ] [ ]
1
( , )
c xc c yc c
u x y u u x x u y y = + − + − (3.95)
In Equation (3.95), the only unknowns are u
xc
and u
yc
, which are the components of
the gradient of quantity u at the centroid of the cell.
Once the linear polynomial is obtained for the quantity u, the sum of the squares of
the averages of the difference between the value of u obtained from constructed
polynomial and its actual value is minimized within the region made up of the cells
that form the support set to find the unknowns u
xc
and u
yc
. The support set is defined
in the same way as the GreenGauss reconstruction shown in Figure 3.4. This
constitutes the least squares problem for the reconstruction.
sup
2
1
1
( )
N
i
i
S Av u u
=
⎡ ⎤ = −
⎣ ⎦
∑
(3.96)
0 , 0
xc yc
S S
u u
∂ ∂
= =
∂ ∂
(3.97)
66
Once the differentiations given in Equation (3.97) are performed, the expressions
given in the following equations are obtained.
sup
1
1
1
( )
2 ( ) 0
N
i
i
i
xc xc
Av u u
S
Av u u
u u
=
⎧ ⎫
⎡ ⎤ ∂ −
∂ ⎪ ⎪
⎣ ⎦
= − ⋅ =
⎨ ⎬
∂ ∂
⎪ ⎪
⎩ ⎭
∑
(3.98)
sup
1
1
1
( )
2 ( ) 0
N
i
i
i
yc yc
Av u u
S
Av u u
u u
=
⎧ ⎫
⎡ ⎤ ∂ −
∂ ⎪ ⎪
⎣ ⎦
= − ⋅ =
⎨ ⎬
∂ ∂
⎪ ⎪
⎩ ⎭
∑
(3.99)
Then the average term over the region constructed by the cells in the support set is
evaluated.
1
1
( ) ( ) ( )
i c i i i xc c i yc c i
i i i i i
Av u u u dA u dA u x x dA u y y dA
A
⎡ ⎤
− = − + − + −
⎢ ⎥
⎣ ⎦
∫ ∫ ∫ ∫
(3.100)
Once the integrals in Equation (3.100) are evaluated and the expressions simplified,
it is possible to arrive at the following equation
1
( ) ( ) ( ) ( )
i c ci xc ci c yc ci c
Av u u u u u x x u y y − = − + − + − (3.101)
When the expression presented in Equation (3.101) for each cell in the support set is
placed in the summations given in Equations (3.98) and (3.99), two equations for the
two unknowns can be obtained as
( )
sup
1
( ) ( ) ( ) 0
N
c ci xc ci c yc ci c ci c
i
xc
S
u u u x x u y y x x
u
=
∂
⎡ ⎤ = − + − + − − =
⎣ ⎦
∂
∑
(3.102)
( )
sup
1
( ) ( ) ( ) 0
N
c ci xc ci c yc ci c ci c
i
yc
S
u u u x x u y y y y
u
=
∂
⎡ ⎤ = − + − + − − =
⎣ ⎦
∂
∑
(3.103)
67
Equations (3.102) and (3.103) can be presented in a matrix form as
( )
( )
( )( )
( )( )
sup sup sup
sup sup sup
2
1 1 1
2
1 1 1
( ) ( )
( ) ( )
N N N
ci c ci c ci c ci c ci c
xc i i i
N N N
yc
ci c ci c ci c ci c ci c
i i i
x x x x y y u u x x
u
u
x x y y y y u u y y
= = =
= = =
⎡ ⎤ ⎡ ⎤
− − − − −
⎢ ⎥ ⎢ ⎥
⎡ ⎤
⎢ ⎥ ⎢ ⎥
=
⎢ ⎥
⎢ ⎥ ⎢ ⎥
⎣ ⎦
⎢ ⎥ ⎢ ⎥ − − − − −
⎢ ⎥ ⎢ ⎥
⎣ ⎦ ⎣ ⎦
∑ ∑ ∑
∑ ∑ ∑
(3.104)
This matrix can be solved to obtain the gradients at the centre of the cell from which
the value of quantity u can be obtained at any location within the cell. The least
squares method is employed for the reconstruction in the code developed.
3.6.3 Gradient Limiting
It is undesirable to introduce new maxima points within the region formed by the
support cells and the cell for which the gradient of flow variables are being
constructed. Hence, the gradients are multiplied by a limiter to avoid introduction of
new maxima points. The linear reconstruction takes the following form with the
application of this condition.
1
( ) ( )
c xc c yc c
u u u x x u y y ⎡ ⎤ = +Φ − + −
⎣ ⎦
(3.105)
The value of the limiter Φ is between 0 and 1. To determine the exact value of the
limiter, the maximum and the minimum of the values of quantity u are found at the
cell centroids.
max
min
sup
max( , )
min( , ) 1,...,
c ic
c ic
u u u
u u u i N
=
= =
(3.106)
Then the values of the reconstructed polynomial are evaluated at the corners of the
cell since these locations will have the maximum and minimum values within the cell
for a linear reconstruction. The value of quantity u which deviates the most from the
centroidal value will be marked as
1
cor
u serving as a basis for determining the value
68
of the limiter with the implementation of the formulation given in the following
equation
max
1
1
min
1
1
1
min 1, 0
min 1, 0
1 0
c
cor c
cor c
c
cor c
c cor
cor c
u u
u u
u u
u u
u u
u u
u u
⎧ ⎛ ⎞ −
− >
⎪ ⎜ ⎟
−
⎝ ⎠
⎪
⎪
⎛ ⎞ − ⎪
Φ = − <
⎨ ⎜ ⎟
−
⎝ ⎠
⎪
⎪
− =
⎪
⎪
⎩
(3.107)
3.7 Temporal Discretization
Once the flux terms in the governing Equation (3.8) is obtained, then the temporal
discretization has to be performed to update the values of the conserved quantities. A
multistage time stepping method is employed for temporal discretization. A detailed
discussion of the multistage methods can be found in (30).
3.7.1 Explicit Time Stepping Schemes
To apply the multistage time stepping method, an alternative representation of the
governing Equation (3.8) is employed as
1
0
cell
t A
∂
+ =
∂
U
R (3.108)
In Equation (3.108), the flux terms are presented in a more compact form as residual
vector, R, and its value can be determined from the known values of the conserved
quantities at a given stage. Hence, Equation (3.108) can be treated as an ordinary
differential equation. This equation can be discretized in mstages as described in
(31) to obtain the following equation.
69
1
cell
m
cell
t
A
t
A
β
β
=
∆
= −
∆
= = −
n
0
0
1 0
n+1 m m1
0
U U
U U R
U U U R
.
(3.109)
In Equation (3.109), β ’s are stage coefficients while t ∆ is the time step. The
residuals are evaluated from the values of the conserved variables from the previous
stage. Hence, this is an explicit formulation. In contrast to classical RungeKutta
methods only the values of the zeroth solution and the last residual have to be stored
in this formulation (31). It is possible to tune the value of β for stability and to
obtain larger time steps.
Table 3.1 Stage coefficients and CFL numbers for the multistage method
stages 3 4 5 3 4 5
σ 1.5 2.0 2.5 0.69 0.92 1.15
β
1
0.1481 0.0833 0.0533 0.1918 0.1084 0.0695
β
2
0.4000 0.2069 0.1263 0.4929 0.2602 0.1602
β
3
1.0000 0.4265 0.2375 1.0000 0.5052 0.2898
β
4
1.0000 0.4414 1.0000 0.5060
β
5
1.0000 1.0000
firstorder scheme secondorder scheme
Table 3.1 taken from (31) shows the optimised values of the stage coefficients and
CFL numbers for three, four and five stage methods of first and second order
accuracy. The CFL number is a factor that governs how large the time step can be
with respect to the cell area. In the code developed, three and four stage second order
schemes are used.
70
3.7.2 Determination of the Time Step
In Equation (3.109), the value of the time step is not determined. The value of the
time step, if taken to be too large, will introduce stability problems. If this value is
taken to be too small then the convergence will be very slow and unnecessary
amount of computations will be performed.
The mesh generated for an arbitrary geometry will contain cells that have different
sizes and since the time step is dependent on the cell size, every cell will have a
different maximum time step. As a result, a different time step will be used for each
cell since a steady state solution is sought.
The time step is calculated in terms of the area of the cell under investigation as in
the following formulation given in (31).
x y
cell
t
A
σ ∆
=
Ψ + Ψ
(3.110)
where σ is the CFL number and Ψ
is the spectral radius.
The spectral radii Ψthat appear in Equation (3.110) can be calculated using the
length of the edge projections of the cell in both xy directions,
x
S and
y
S .
( )
1
2
x x
edges
u a S Ψ = +
∑
(3.111)
( )
1
2
y y
edges
v a S Ψ = +
∑
(3.112)
The time step obtained thus is multiplied by a safety factor in some cases to promote
stable solutions. If Equation (3.110) is investigated, this equation is in one sense a
division of a characteristic length by a characteristic speed. The characteristic speed
is the maximum speed with which information is carried for a cell. Hence, the time
71
step is obtained by constricting the distance information travels from a cell within the
field.
3.8 Solution Refinement
Once a certain level of convergence is achieved, the cells are refined according to the
gradient of the solution. The purpose of this refinement is to provide enough
resolution at the places where there might be discontinuities due to shocks or there
are large gradients in the solution (e.g. around the stagnation point).
The solution refinement criterion used is based on the one outlined in (14). This
method is based on both the curl of the velocity, for detecting shear layers, and the
gradient of the velocity, for detecting shocks. The curl and gradient of the velocity is
multiplied by some characteristic length of the cell as
( ) ( )
3 3
2 2
, =
cell cell
A V A V τ ζ = ∇⋅ ∇×
(3.113)
The standard deviations of the quantities defined in Equation (3.113) are calculated
and the cells exceeding the standard deviation are refined.
2
1
2
1
Refine cell j
Refine cell j
cell
cell
n
i
i
j
cell
n
i
i
j
cell
n
n
τ
ζ
τ
τ σ
ζ
ζ σ
=
=
⎧
⎛ ⎞
⎪
⎜ ⎟
⎪
⎝ ⎠
> =
⎪
⎪
⎨
⎪
⎛ ⎞
⎪
⎜ ⎟
⎪ ⎝ ⎠
> =
⎪
⎩
∑
∑
(3.114)
72
CHAPTER 4
RESULTS &DISCUSSIONS
To test the validity of the code developed, the flow around a NACA0012 airfoil and
an NLR7301 with a flap has been solved for different conditions. Besides
demonstrating the characteristic features of the flows, different methods for handling
the flux terms in the governing equations have been compared.
4.1 NACA0012
The NACA0012 tests have been performed for a subsonic case and a transonic case.
In the subsonic tests the aim is to show the importance of reconstruction to obtain
second order solutions and to discuss the role of first order solution on the stability.
The transonic tests have been performed to demonstrate the shock capturing
capabilities of the various methods used to evaluate the flux terms. Also they are
compared in terms of computational efficiency.
4.1.1 Subsonic Tests
The farfield boundary condition of the first case for the subsonic tests is a free
stream Mach number of 0.6 and an angle of attack of 0˚. The experimental data for
this case can be found in (32) which were extracted from (33). For this case a total of
six runs have been performed. All the runs are performed using Roe’s method. The
convergence criterion for all of the runs is that the standard deviation of the residual
calculated for the continuity equation should drop five levels on a log scale. Two of
the six runs are first order solutions. One of these solutions is with solution
refinement while the other is not. Two runs are performed second order with and
73
without solution refinement, while the last two are run first order until the residuals
drop four levels then they switch to second order.
The pressure coefficient vs. the chord graph for the first order solutions are presented
in Figure 4.1.
Figure 4.1 Cp vs. chord for first order solution
As can be seen from Figure 4.1, the first order solution without solution refinement
performs poorly. It is not possible to catch the peak of the pressure coefficient with
this solution. However, solution refinement introduces a considerable improvement
74
to the solution. But even with solution refinement the peak of the pressure coefficient
graph is not captured accurately.
The pressure coefficient vs. chord graph can also be illustrated for the second order
solution as in Figure 4.2.
Figure 4.2 Cp vs. chord for second order solution
As can be seen from Figure 4.2, the second order solution provides a much more
accurate solution than the first order method as expected. The peak of the pressure
coefficient is captured accurately for both of the solutions with and without solution
75
refinement although the solution with solution refinement is slightly better on this
aspect.
The effect of solution refinement can be illustrated by examining the differences
between a refined and nonrefined mesh.
Figure 4.3 A refined and nonrefined mesh around stagnation point
In Figure 4.3, the mesh on the right belonging to a solution with refinement has a
considerably denser mesh around the stagnation point than the one without
refinement. This is an expected result since there are large gradients around this
region.
Finally the results obtained with the solution where first order is used until some
convergence is obtained and then it is switched to second order, can be compared
with the fully second order solution.
76
Figure 4.4 Comparison of the second order solutions
In Figure 4.4, it is almost impossible to distinguish between the two solutions. The
same situation also prevails for a solution with refinement. Hence it can be said that
in terms of accuracy there is no difference between the two solutions.
The Mach contours for the second order solution can be depicted as in Figure 4.5.
77
Figure 4.5 Mach contours for NACA0012 subsonic flow
After comparing the first and second order methods in terms of accuracy, their
computational performances can also be compared as in Table 4.1.
78
Table 4.1 Comparison of first and second order solutions
Method Number of Iterations Time
First order without refienement 1938 45sec
First order with refienement 2691 2min 37sec
Second order without refinement 4325 11min 23sec
Second order with refinement 7490 1hr 1min 36sec
Switch to second order without refinement 3236 6min 57sec
Switch to second order with refinement 4425 25min 24sec
Naturally the first order solution converges much faster than the second order
methods, but it would not be reasonable to obtain a solution with this method due to
poor accuracy. As observed from Table 4.1, starting the solutions first order then
switching to second order after some level of convergence, decreases the
computation time considerably compared to full second order solution. Also as
illustrated in Figure 4.4, there is almost no difference between these solutions.
The calculations for the second case of subsonic flow around the NACA0012 airfoil
were performed with a Mach number of 0.4167. By changing the angle of attack
values a lift coefficient vs. angle of attack graph was constructed. To obtain the
solution Roe’s method was used and solution refinement wass not employed.
Since the pressure is the main factor determining the lift force at high speeds, it is
expected to get a good agreement with the experimental data extracted from (34)
until flow separation, a viscosity governed phenomena, is encountered. This
expectation is realized as can be seen from Figure 4.6 within the region where there
is no separation.
79
Figure 4.6 Lift coefficient vs. angle of attack graph
To obtain the results presented in Figure 4.6, first order calculations are performed
until some level of convergence is obtained then the solution is switched to second
order. As discussed earlier this method decreases computation time while there is no
loss in terms of accuracy.
It was also attempted to use a fully second order solution to obtain the results. It was
possible to get solutions within the range of angle of attacks from 6˚ to 8˚. However
for larger angle of attack values than 8˚, divergence in the calculations was observed.
Using the first order solution then switching to second order provided results for
angle of attack values of 10˚, 12˚ and 16˚ also. The results obtained for 12˚ and 16˚
80
are not presented in Figure 4.6 since they are related to a flow with separation and
this case must be treated with a viscous solver. However this fact still indicates that
the first order solution is more stable than the second order one. When the residual
plots for first and second order solutions are investigated, an indication of this fact
can be seen. In Figure 4.7, it is observed that the residuals decrease smoothly for the
first order solution while for the second order one there are a lot of scattering and
oscillations as the residuals decrease.
Figure 4.7 Residual plot for first and second order solutions
By employing the first order method until some level of convergence is obtained, a
first order accurate initialization of the field variables is obtained which eases the
81
convergence of the second order solution. This procedure will be used by default for
the rest of the calculations that will be presented.
4.1.2 Transonic Tests
The farfield conditions for the transonic flow tests around the NACA0012 profile
are a free stream Mach number of 0.85 and an angle of attack of 1˚. For this flow
shocks on the upper and lower surface of the airfoil is expected, the one on the upper
surface being stronger. Each one of the five methods implemented for the evaluation
of the flux terms in the governing equations has been employed in the solution to
compare these methods. The most crucial point in the tests performed is the location
of the shocks. The transonic flow is a more challenging case than the subsonic flow.
Hence the performance of different methods to evaluate the flux terms is compared
for this case. The calculations have been performed both with and without solution
refinement. To verify the computational solution, AGARD data (35) has been
extracted from (14).
To compare the accuracy of the different methods for evaluating the fluxes,
calculations with solution refinement have been performed. First the results obtained
from the AUSM and its derivative methods are compared amongst themselves. The
pressure coefficient vs. the chord graph for these methods is illustrated in Figure 4.8.
82
Figure 4.8 Pressure coefficient vs. chord graph for the AUSM methods
It is observed from Figure 4.8 that the AUSMD and AUSMV methods capture the
location of the weaker shock on the lower surface while the AUSM predicts this
location slightly downstream. On the upper surface none of the methods were able to
capture the shock location exactly. The AUSMD and the AUSMV methods provide a
closer estimation of the shock location on this surface than the AUSM method. All of
the methods exhibit oscillations around the shock location, the ones on the upper
surface being more marked. This is due to the performance of the limiter in the
calculation of the fluxes.
83
The Roe’s method and the van Leer method can be compared with one of the
representative AUSM methods. The AUSMV method providing the most accurate
results among the AUSM methods is chosen as the representative AUSM method
and the pressure coefficient vs. chord graph for these methods is illustrated in Figure
4.9.
Figure 4.9 Pressure coefficient vs. chord graph for Roe, van Leer and AUSMV
It is seen that both the Roe’s method and the van Leer’s method estimate the shock
location at a slightly upstream position on the upper surface while there is good
agreement on the lower surface. The oscillations around the shock locations are also
84
inherent in these methods. In terms of accuracy AUSMD, AUSMV and the van
Leer’s methods provide the best results for this particular test case.
Besides the accuracy of the methods, their convergence characteristics should also be
examined. The convergence criterion is that the residual for the continuity equation
calculated in the first step should drop five levels on a log scale. None of the methods
were able to provide a five level drop in the residual for solution adaptive
calculations. Hence to compare the convergence characteristics calculations without
solution refinement have been performed. The calculations were performed with four
different CFL safety coefficients; 1.0, 0.5, 0.2 and 0.1, four each one of the five
methods. In Table 4.2, the results for the solution with the maximum CFL safety
factor providing a convergent solution has been presented.
Table 4.2 Computational characteristics of the methods for evaluating fluxes
Method Convergence Number of Iterations Calculation Time CFL Safety Coefficient
AUSM converges 5555 9min 1sec 1.0
AUSMD oscillates 100000 3hr 35min 4sec 0.1
AUSMV converges 9646 16min 19sec 0.5
Roe converges 5207 9min 15sec 1.0
van Leer diverges   
The AUSM and the Roe’s methods converged with a CFL safety coefficient of 1.0
while the AUSMV method converged with a CFL safety factor of 0.5. For the
AUSMD method with a CFL safety factor of 0.1, the residual of the continuity
equation oscillates around 8x10
4
corresponding to a residual drop of about four
levels after about 60000 steps till the maximum number of iterations of 100000. The
van Leer method diverges when solution refinement is not applied implying a
relation between the method and the mesh size. The residual plots for the convergent
solutions can be depicted as in Figure 4.10.
85
Figure 4.10 Residual plot for transonic flow
It is also possible to show the effect of solution adaptation on the mesh generated
initially. The solution refinement has been performed every time the residuals
dropped to one twentieth of the initial residual which is updated just before solution
adaptation is performed. The effect of mesh generation in the critical regions around
the airfoil is illustrated in Figure 4.11.
86
Figure 4.11 Effect of refinement at shock location
As can be seen from Figure 4.11, the mesh is refined around the location of the shock
since there are large gradients in the flow variables at this location. Another region
where large gradients are observed is at the nose of the airfoil around the stagnation
point. The effect of solution refinement in this region is illustrated in Figure 4.12.
Figure 4.12 Effect of refinement around stagnation point
The effect of solution adaptation can be illustrated by plotting the Mach contours for
a refined and nonrefined solution.
87
Figure 4.13 Mach contours for transonic flow around NACA0012 airfoil
In Figure 4.13, the Mach contours obtained with the AUSMV method are displayed.
As can be observed for the solution without refinement, the shocks are smeared. The
pressure coefficient vs. chord graphs of these solutions can also be displayed. In
Figure 4.14, it is seen that the solution obtained without refinement is able to capture
both the pressure distribution on the airfoil and the shock locations with refinement
introducing a slight improvement on both aspects. However in Figure 4.13, it is seen
that the overall flow field is not represented accurately without solution refinement.
Hence it can be said that if the main concern is obtaining the forces on the airfoil, the
solution without refinement will provide a good estimation in a considerably shorter
amount of time.
88
Figure 4.14 Pressure coefficient vs. chord for AUSMV with and without refinement
4.2 NLR 7301 with Flap
The NLR 7301 tests have been performed for a gap of 2.6% percent of chord length,
an overlap of 5.3% of chord length and a flap angle of 20˚. The definition of these
variables can be shown as in Figure 4.15.
89
Figure 4.15 Definition of geometry variables for NLR 7301 with flap
Tests were performed at three different angle of attack values; 6˚, 10.1˚ and 13.1˚.
The free stream Mach number for all the cases was 0.185. The five methods; Roe,
van Leer, AUSM, AUSMD, AUSMV, have been employed in the calculations. To
verify the results computed, data obtained from (36) has been used.
The pressure coefficient vs. chord graph for angle of attack of 6˚ is presented for the
AUSM methods in Figure 4.16. All of the AUSM methods are able to capture
pressure distribution with good accuracy for the most part. But in terms of capturing
the peak of the pressure coefficient on the upper surface of the main airfoil, AUSM
method is superior to the others. The AUSMD method performs better than the
AUSMV method on this aspect.
90
Figure 4.16 Pressure coefficient vs. chord graph (AUSM methods AoA 6˚)
The pressure coefficient vs. chord graph can also be constructed for the Roe’s, van
Leer’s and AUSM methods. In Figure 4.17, it is seen that the AUSM method
provides better results than both the Roe’s and van Leer’s methods. The solution by
Roe’s method is comparable to the other AUSM methods. In terms of capturing the
peak of the pressure coefficient on the upper surface of the main element, Roe’s
method performs similar to the other method while van Leer provides the poorest
result in terms of accuracy.
91
Figure 4.17 Pressure coefficient vs. chord graph (AUSM, Roe, van Leer AoA 6˚)
The computational performances of each method can be tabulated to obtain a
comparison.
92
Table 4.3 Computational performance of the methods for NLR 7301 AoA 6˚
Method Convergence Number of Iterations Time Lift Coefficient
AUSM converges 7072 41min 47sec 2.51
AUSMD oscillates 100000 10hr 21min 49sec 2.46
AUSMV oscillates 100000 10hr 14min 0sec 2.45
Roe oscillates 100000 10hr 56min 15 sec 2.42
van Leer converges 4172 22min 53sec 2.35
It is observed from Table 4.3 that the van Leer’s method converges faster than the
AUSM method. The other methods; AUSMD, AUSMV and Roe, do not converge,
that is the residuals in the continuity equation do not drop five levels but they
fluctuate until the maximum number of iterations is reached. Also it can be seen that
the highest lift coefficient is found with the AUSM method. This is an expected
result as the AUSM method captures the peak of the pressure distribution on the
upper surface of the main airfoil better than the other methods. The AUSMD,
AUSMV and Roe’s methods provide a similar result on this aspect while van Leer’s
method gives the smallest estimate as it was the worst in capturing the peak.
The pressure coefficient vs. chord graph for the angle of attack value of 10.1 can be
presented for the AUSM methods. From Figure 4.18, it is seen that the performance
of the AUSM method is better than both the AUSMD and the AUSMV methods. The
AUSMV method starts to deviate from the AGARD data which is especially
apparent on the upper surface of the flap. The peak of the pressure coefficient is
captured accurately with the AUSM method.
93
Figure 4.18 Pressure coefficient vs. chord graph (AUSM methods AoA 10.1˚)
The pressure coefficient vs. the chord graph can be presented for the AUSM, Roe
and van Leer methods as in Figure 4.19.
94
Figure 4.19 Pressure coefficient vs. chord graph (AUSM, Roe, van Leer AoA 10.1˚)
As can be seen from Figure 4.19, AUSM method provides a better solution. It is also
observed that the Roe’s method provides a better estimation of the peak in the
pressure distribution on the upper surface of the main element than the AUSMD and
AUSMV methods.
The computational performance of the methods can also be compared from Table
4.4.
95
Table 4.4 Computational performance of the methods for NLR 7301 AoA 10.1˚
Method Convergence Number of Iterations Time Lift Coefficient
AUSM converges 11567 1hr 10min 43sec 2.98
AUSMD oscillates 100000 10hr 21min 5sec 2.76
AUSMV oscillates 100000 10hr 13min 2sec 2.75
Roe converges 78924 8hr 15min 47 sec 2.86
van Leer converges 4175 22min 50sec 2.83
It is observed from Table 4.4 that as the angle of attack is increased Roe’s method
converges like the AUSM and the van Leer methods. The performances of the van
Leer, AUSMD and AUSMV methods essentially remain the same computationally.
The Roe’s method being able to capture the pressure peak better for this case,
estimates a relatively higher lift coefficient than the previous case. The AUSM
method still being the best one to capture the peak provides the highest estimate of
the lift coefficient.
The pressure coefficient vs. the chord graph can be constructed for the angle of
attack of 13.1˚ for the AUSM methods. From Figure 4.20, it is seen that the AUSMV
method fails at this angle of attack. Once again the AUSM method provides a
considerably better estimate of the pressure peak on the upper surface of the main
element.
96
Figure 4.20 Pressure coefficient vs. chord graph (AUSM methods AoA 13.1˚)
In Figure 4.21 the pressure coefficient vs. chord graph for the AUSM, Roe and van
Leer methods are depicted. The AUSM method providing the best of the solutions,
the Roe’s method provides a better solution than van Leer’s method. As in the case
of angle of attack of 10.1˚, the Roe’s method gives a better solution than the
AUSMD and AUSMV method which fails.
97
Figure 4.21 Pressure coefficient vs. chord graph (AUSM, Roe, van Leer AoA 13.1˚)
The computational performance of the methods for the case of angle of attack of
13.1˚ can be compared with the aid of Table 4.5.
98
Table 4.5 Computational performance of the methods for NLR 7301 AoA 13.1˚
Method Convergence Number of Iterations Time Lift Coefficient
AUSM converges 12182 1hr 14min 5sec 3.29
AUSMD oscillates 100000 10hr 20min 30sec 3.08
AUSMV oscillates 100000 10hr 13min 5sec 3.03
Roe converges 87222 9hr 11min 9sec 3.17
van Leer converges 4396 24min 12sec 3.16
The computational performance of the methods is pretty similar to the case of angle
of attack of 10.1˚ with the exception that AUSMV method fails totally. It can be
concluded that AUSM method provides the best results in all the cases considered.
Roe’s and van Leer’s methods do improve as the angle of attack increases which
may be due to the fact that as the angle of attack increases so does the velocities
around the airfoil, which makes the problem shift closer to the compressible range.
The performance of the AUSMD method remains constant while the AUSMV
method fails as the angle of attack is increased.
Solution adaptive calculations have also been performed for the van Leer method to
improve the results. The effect of solution refinement on the mesh can be shown as
in Figure 4.22 and Figure 4.23.
99
Figure 4.22 Effect of solution refinement around stagnation region of NLR 7301
Figure 4.23 Effect of solution refinement around the flap
100
The pressure coefficient vs. the chord graphs for the solution with and without
refinement can be constructed for a comparison at 6˚.
Figure 4.24 Pressure coefficient vs. chord for van Leer AoA 6˚
In Figure 4.24, it is seen that the adapted solution performs better at capturing the
peak of the pressure distribution on the upper surface of the main element. In fact the
solution is improved around this region as a whole. The estimate of the lift
coefficient increases from 2.35 to 2.49. This is much closer to the one predicted by
the AUSM method.
101
The pressure coefficient vs. the chord graph for a solution with and without
refinement for angle of attack of 10.1˚ is presented in Figure 4.25.
Figure 4.25 Pressure coefficient vs. chord for van Leer AoA 10.1˚
There is an overall improvement of the solution with the introduction of refinement,
but this time there is a considerable over estimation of the peak of the pressure
coefficient on the upper surface of the main element. In fact the estimation for the lift
coefficient becomes 3.05 which is greater than predicted by the AUSM method.
Lastly, the pressure coefficient vs. the chord graph for a solution with and without
refinement for angle of attack of 13.1˚ is presented in Figure 4.26.
102
Figure 4.26 Pressure coefficient vs. chord for van Leer AoA 13.1˚
Once again the overall solution does improve, but there is an over prediction of the
pressure coefficient peak on the upper surface of the main element.
Figure 4.27, Figure 4.28 and Figure 4.29 illustrate the flow fields obtained for angle
of attack values of 6˚, 10.1˚, 13.1˚ respectively.
103
Figure 4.27 Flow field for NLR 7301 with flap AoA 6˚
Figure 4.28 Flow field for NLR 7301 with flap AoA 10.1˚
104
Figure 4.29 Flow field for NLR 7301 with flap AoA 13.1˚
As the angle of attack increases, the acceleration of the flow on the upper surface of
the main element increases also. This makes the streamlines more compact, and also
the pressure drop on the upper surface of the main element increases. In addition,
with the increasing angle of attack the region of high pressure on the lower surface of
the main element becomes larger. This promotes higher lift forces, which is
consistent with the computations. An important feature of the NLR 7301 airfoil is the
region between its lower surface and the flap. There were no vortex formations in
this region for the flows considered. If there were a vortex formation in this region,
due to the pressure drop associated with it, there would be decrease in the lift force
accompanied by an increase in the drag force since the orientation of the face normal
105
of the main element is towards increasing the lift and decreasing the drag in this
region. Hence viscous calculations are of great importance in this region because of a
possibility of separation.
106
CHAPTER 5
CONCLUSIONS
In the subsonic calculations, the difference between the accuracy of a first order and
a second order formulation has been illustrated. The first order method is inadequate
in capturing the flow characteristics such as the peak of the pressure distribution for
the relatively simple problem of subsonic flow. There is however a considerable
improvement of the solution with the introduction of solution refinement. This is due
to the fact that with solution refinement, the mesh around the regions of high
gradients, the stagnation region in case of NACA0012 subsonic flow, is refined and
the accuracy in this region increases as the peak of the pressure distribution can be
captured more accurately.
Even though solution refinement did introduce an improvement on the first order
solution, it still wasn’t accurate enough. The second order method performed much
better than a first order solution with refinement. When solution refinement was
performed for a second order formulation, there was little improvement in the
solution for NACA0012 subsonic flow, since the nonadapted solution performed
reasonably well, to capture the peak of the solution refinement.
The second order method was superior to the first order one in terms of accuracy, but
it did call for a much increased computation time. To resolve this issue, first order
calculations would be performed until some level of convergence is obtained then the
solution would switch to a second order formulation. This resulted in up to more than
50% decrease in the computation time while there was no virtual loss in terms of
accuracy. Also as the angle of attack was varied to construct a lift coefficient vs.
chord graph it was seen that at certain angles it was not possible to get a solution
with fully second order calculations. This and the fact that second order solutions
107
provided residual plots with considerable scattering and oscillations indicated that
the first order formulation is more stable than the second order one. Hence by
starting with first order calculations, the field is initialized more accurately which
promotes the second order method to converge as it is sensitive to the initial
conditions.
The transonic flow around the NACA0012 profile was a more challenging case than
the subsonic flow, providing grounds to test the performance of different methods for
evaluating the flux terms. The main challenge was to catch the shock locations,
especially the stronger one on the upper surface of the airfoil. The AUSMD,
AUSMV and van Leer’s methods were the ones that showed the best performance on
this aspect.
For the transonic flow solution refinement is a must to get an accurate representation
of the flow field since it is already known that there will be large gradients at shock
locations in addition to the stagnation region. If solution refinement is not used
smeared shocks are obtained. It was however possible to get an accurate pressure
distribution on the airfoil as the mesh elements in this region are small in size. Hence
just to analyze the forces on the airfoil a solution without refinement can be
employed to get quick results.
It was not possible to compare the computational performance of the methods with a
solution with refinement, so calculations were also performed without refinement.
For the solutions without refinement it was seen that, the Roe’s method and the
AUSM method converged easily with a CFL safety coefficient of 1. The AUSMV
method converged with a CFL safety coefficient of 0.5 while the van Leer’s method
diverged.
Another important feature of the solutions for the transonic flow was the oscillations
seen at the shock locations, the one on the upper surface of the airfoil with the
stronger shock being more apparent. This is due to the poor performance of the
limiters used in the calculations. A limiter that won’t hamper the convergence while
preventing oscillations should be implemented to improve the solution.
108
The NLR 7301 with a flap was also a challenging case to test the performance of the
various methods implemented for evaluating the flux terms. The AUSM method was
superior in terms accuracy and it was second to van Leer’s method in terms of
computational performance. The van Leer’s method with the poorest performance for
the case with an angle of attack of 6˚ did improve with the application of solution
refinement.
The performance of both the Roe’s and van Leer’s methods improved for the cases
with higher angle of attacks. This might be due to the flow shifting closer to the
compressible range with the increasing angle of attacks promoting higher velocities
in the field. The performance of the AUSMD method was consistent throughout the
calculations while the AUSMV method failed for large angle of attack values.
From the pressure coefficient vs. chord graphs, it was observed that the estimation of
the peak pressure on the upper surface of the main element was crucial in the
prediction of the lift coefficient. van Leer’s method with solution refinement and
AUSM method performed well on this aspect, while AUSMD, AUSMV, and Roe’s
method performed worse with similar results at an angle of attack 6˚. For the angle of
attack values of 10.1˚ and 13.1˚, AUSM was the superior method once again while
the performance of Roe’s method improved considerably. The van Leer method with
solution refinement provided over estimations of the pressure peak.
109
BIBLIOGRAPHY
1. Wendt, John F, et al. Computational Fluid Dynamics An Introduction. Berlin :
Springer_Verlag, 1996.
2. Versteeg, H K and Malalasekera, W. An Itroduction to Computational Fluid
Dynamics The Finite Volume Method. Malaysia : Prentice Hall, 1995.
3. Liseikin, Vladimir D. Grid Generation Methods. Berlin : SpringerVerlag, 1999.
4. Thompson, Joe F, Soni, Bharat K and Wheatherill, Nigel P. Handbook of Grid
Generation : CRC Press, 1998.
5. Hirsch, Charles. Numerical Computation of Internal and External Flows. New
York : John Wiley & Sons, 1988.
6. Reddy, J N. An Introduction to the Finite Element Method. Singapore : McGraw
Hill, 1993.
7. Hoffmann, Klaus A and Chiang, Steve T. Computational Fluid Dynamics for
Engineers. Wichita : Enginnering Education System, 1995.
8. Reentrant Polygon Clipping. Sutherland, I E and Hodgeman, G W. January
1974, Communications of the ACM, pp. 3242.
9. A New Concept and Method for Line Clipping. Liang, Y D and Barsky, B A. 1,
January 1984, ACM Transactions on Graphics, Vol. 3, pp. 122.
10. Sproull, R F and Sutherland, I E. A Clipping Divider. Washington D. C. :
Thompson Books, 1968.
110
11. An Efficent Algorithm for 2D Line Clipping: Its Development and Analysis.
Nicholl, Tina M, Lee, D T and Nicholl, Robin A. July 1987, Computer Graphics,
pp. 253262.
12. A New Two Dimensional Line Clipping Algorithm for Small Windows. Day, J D.
1992, Computer Graphic Forums, pp. 241245.
13. A Fast TwoDimensional Line Clipping Algoritm via Line Encoding. Sobkow, M
S, Pospisil, P and Yang, Y H. 1987, Computer Graphics, pp. 459467.
14. De Zeeuw, Darren L. A QuadTree Based AdaptivelyRefined CartesianGrid
Algorithm for the Solution of The Euler Equations. Michigan : The University of
Michigan, 1993. PhD thesis.
15. Preparata, P Franco and Shamos, Michael Ian. Computational Geometry an
Introduction. New York : SpringerVerlag, 1985.
16. Toro, E F. Riemann Solvers and Numerical Methods for Fluid Dynamics.
Berlin : SpringerVerlag, 1999.
17. Laney, Culbert B. Computational Gas Dynamics. Cambridge : Cambridge
University Press, 1998.
18. Greenberg, Michael D. Advanced Engineering Mathematics. New Jersey :
Prentice Hall, 1998.
19. Approximate Riemann Solvers, Parametrized Vectors, and Difference Schemes.
Roe, P L. 1981, Journal of Computational Physics, Vol. 43, pp. 357372.
20. Flux Vector Splitting for the Euler Equations. van Leer, B. s.l. : SpringerVerlag,
1982. Proceedings of the 8th International Conference on Numerical Methods in
Fluid Dynamics.
111
21. A New Flux Splitting Scheme. Liou, M S and Steffen, C J. s.l. : Journal of
Computational Physics, 1993, Vol. 107, pp. 2339.
22. An Accurate and Robust Vector Splitting Scheme for Shock and Contact
Discontinuities. Wada, Yasuhiro and Liou, M S. 3, s.l. : Siam J. Sci. Comput.,
1997, Vol. 18, pp. 633657.
23. Flux Vector Splitting of the Inviscid Gasdynamic Equations with Applications to
Finite Difference Methods. Steger, J L and Warming, R F. s.l. : Journal of
Computational Physics, 1981, Vol. 40, pp. 263293.
24. Barth, Timothy J. Simplified Numarical Methods for Gasdynamic Systems on
Triangulated Domains. Stanford : Stanford University, 1998. PhD thesis.
25. Barth, Timothy J and Frederickson, Paul O. Higher Order Solution of the
Euler Equations. Reno : AIAA paper AIAA900013, 1990. 28th Aerospace
Sciences Meeting.
26. Barth, Timothy J and Jespersen, Dennis C. The Design and Application of
Upwind Schemes on Unstructured Meshes. Reno : AIAA paper AIAA890366,
1989. 27th Aerospace Sciences Meeting.
27. Coirier, William John. An Adaptively Refined, Cartesian, CellBased Scheme
for the Euler and Navier Stokes Equations. Michigan : The University of Michigan,
1994. PhD thesis.
28. Barth, Timothy J. Recent Developments in High Order KExact Reconstruction
on Unstructured Meshes. Reno : AIAA paper AIAA930668, 1993. 31st Aerospace
Sciences Meeting.
29. Gerris: A TreeBased Adaptive Solver for the Incompressible Euler Equations in
Complex Geometries. Popinet, Stéphane. 2003, Journal of Computational Physics,
Vol. 190, pp. 572600.
112
30. Lambert, J D. Computational Methods in Ordinary Differential Methods.
London : Wiley, 1973.
31. Blazek, J. Computational Fluid Dynamics: Principles and Applications. St
Augustin : Elsevier, 2001.
32. Amick, J L. Comparison of the Experimental Pressure Distribution on an
NACA0012 Profile at High Speeds with that Calculated by the Relaxation Method.
s.l. : National Advisory Committee for Aeronautics, 1950. Technical note.
33. Özdemir, Enver Doruk. Implementation of Rotation into a 2D Euler Solver.
Ankara : METU, 2005. Masters thesis.
34. Abbott, Ira H and von Doenhoff, Albert E. Theory of Wing Sections,Including
a Summary of Airfoil Data. New York : Dover Publications, 1959.
35. AGARD Subcommittee C. Test Cases for Inviscid Flow Field Methods. 1986 :
AGARD Advisory Report 211.
36. Advisory Group for Aerospace Research& Development. A Selection of
Experimental Test Cases for the Validation of CFD Codes . s.l. : AGARD Advisory
Report No 303, 1994.
113
APPENDIX A
COORDINATES OF NACA0012
Table A.1 Coordinates of NACA0012 Profile
x y
0.999753 0.001295
0.999013 0.001398
0.997781 0.001571
0.996057 0.001812
0.993844 0.002120
0.991144 0.002496
0.987958 0.002937
0.984292 0.003443
0.980147 0.004012
0.975528 0.004642
0.970440 0.005333
0.964888 0.006082
0.958877 0.006887
0.952414 0.007746
0.945503 0.008658
0.938153 0.009619
0.930371 0.010628
0.922164 0.011681
0.913540 0.012778
0.904508 0.013914
0.895078 0.015088
0.885257 0.016297
0.875056 0.017539
0.864484 0.018809
0.853553 0.020107
0.842274 0.021429
0.830656 0.022773
0.818712 0.024135
114
Table A.1 Coordinates of NACA0012 Profile (continued)
0.806454 0.025514
0.793893 0.026905
0.781042 0.028307
0.767913 0.029717
0.754521 0.031131
0.740877 0.032547
0.726995 0.033962
0.712890 0.035374
0.698574 0.036778
0.684062 0.038172
0.669369 0.039553
0.654508 0.040917
0.639496 0.042263
0.624345 0.043585
0.609072 0.044882
0.593691 0.046149
0.578217 0.047383
0.562667 0.048581
0.547054 0.049739
0.531395 0.050854
0.515705 0.051923
0.500000 0.052940
0.484295 0.053904
0.468605 0.054810
0.452946 0.055655
0.437333 0.056436
0.421783 0.057148
0.406309 0.057789
0.390928 0.058355
0.375655 0.058844
0.360504 0.059251
0.345492 0.059575
0.330631 0.059812
0.315938 0.059960
0.301426 0.060017
0.287110 0.059980
0.273005 0.059848
0.259123 0.059619
0.245479 0.059292
0.232087 0.058866
115
Table A.1 Coordinates of NACA0012 Profile (continued)
0.218958 0.058340
0.206107 0.057714
0.193546 0.056987
0.181288 0.056160
0.169344 0.055233
0.157726 0.054207
0.146447 0.053083
0.135516 0.051863
0.124944 0.050547
0.114743 0.049138
0.104922 0.047638
0.095492 0.046049
0.086460 0.044374
0.077836 0.042615
0.069629 0.040776
0.061847 0.038859
0.054497 0.036867
0.047586 0.034803
0.041123 0.032671
0.035112 0.030473
0.029560 0.028213
0.024472 0.025893
0.019853 0.023517
0.015708 0.021088
0.012042 0.018607
0.008856 0.016078
0.006156 0.013503
0.003943 0.010884
0.002219 0.008223
0.000987 0.005521
0.000000 0.000000
116
APPENDIX B
COORDINATES OF NLR 7301 & FLAP
Table B.1 The coordinates of NLR 7301
x y
0.94360 0.01499
0.94267 0.01519
0.93988 0.01580
0.93524 0.01690
0.92878 0.01830
0.92051 0.02008
0.91047 0.02241
0.89870 0.02512
0.88524 0.02819
0.87015 0.03162
0.85349 0.03536
0.83532 0.03941
0.81572 0.04368
0.79476 0.04814
0.77253 0.05273
0.74911 0.05735
0.72459 0.06191
0.69908 0.06628
0.67267 0.07038
0.64547 0.07410
0.61758 0.07737
0.58912 0.08015
0.56019 0.08248
0.53092 0.08438
0.50141 0.08590
0.47178 0.08704
0.44216 0.08781
0.41265 0.08826
0.38337 0.08838
0.35444 0.08818
0.32598 0.08768
117
Table B.1 The coordinates of NLR 7301 (continued)
0.29809 0.08687
0.27098 0.08577
0.24448 0.08440
0.21897 0.08276
0.19445 0.08087
0.17103 0.07874
0.14880 0.07638
0.12784 0.07381
0.11786 0.07245
0.10824 0.07103
0.09897 0.06958
0.09007 0.06807
0.08155 0.06650
0.07341 0.06488
0.06567 0.06317
0.05832 0.06137
0.05139 0.05943
0.04487 0.05731
0.03877 0.05506
0.03309 0.05268
0.02786 0.05009
0.02305 0.04717
0.01870 0.04373
0.01479 0.03978
0.01133 0.03544
0.00832 0.03076
0.00577 0.02582
0.00368 0.02064
0.00206 0.01516
0.00089 0.00959
0.00019 0.00422
0.00004 0.00080
0.00019 0.00552
0.00089 0.00992
0.00206 0.01417
0.00368 0.01804
0.00577 0.02173
0.00832 0.02521
0.01133 0.02851
0.01479 0.03158
0.01870 0.03445
118
Table B.1 The coordinates of NLR 7301 (continued)
0.02305 0.03717
0.02786 0.03972
0.03310 0.04215
0.03877 0.04443
0.04487 0.04661
0.05139 0.04873
0.05832 0.05073
0.06567 0.05263
0.07341 0.05446
0.08155 0.05622
0.09007 0.05791
0.09897 0.05954
0.10824 0.06109
0.11786 0.06259
0.12784 0.06402
0.14880 0.06672
0.17103 0.06917
0.19445 0.07132
0.21897 0.07317
0.24448 0.07470
0.27089 0.07585
0.29809 0.07666
0.32598 0.07704
0.35444 0.07701
0.38337 0.07652
0.41265 0.07550
0.44216 0.07385
0.47178 0.07144
0.50141 0.06819
0.53092 0.06417
0.56019 0.05959
0.58912 0.05459
0.61758 0.04921
0.64547 0.04355
0.67267 0.03750
0.69908 0.03080
0.72459 0.02362
0.74911 0.01597
0.77253 0.00824
0.79476 0.00080
0.81572 0.00585
119
Table B.1 The coordinates of NLR 7301 (continued)
0.83532 0.01115
0.85349 0.01472
0.87015 0.01702
0.88524 0.01818
0.89870 0.01843
0.91047 0.01807
0.92051 0.01733
0.92878 0.01645
0.93524 0.01530
0.93988 0.01460
0.94267 0.01423
0.94360 0.01410
Table B.2 The coordinates of flap
x y
1.205092 0.102133
1.204014 0.101411
1.200754 0.099373
1.195420 0.095931
1.188106 0.091056
1.178790 0.085249
1.167564 0.078726
1.154588 0.071662
1.140013 0.064282
1.124077 0.056694
1.106992 0.049135
1.098093 0.045428
1.088996 0.041798
1.079729 0.038286
1.070330 0.034929
1.060832 0.031728
1.051260 0.028701
1.041673 0.025861
1.032094 0.023205
1.022570 0.020749
120
Table B.2 The coordinates of flap (continued)
1.013137 0.018508
1.003821 0.016532
0.994643 0.014820
0.985653 0.013410
0.976874 0.012269
0.968366 0.011407
0.960148 0.010799
0.952264 0.010473
0.944743 0.010407
0.937624 0.010583
0.930955 0.010922
0.924783 0.011390
0.919139 0.011985
0.914029 0.012722
0.909442 0.013659
0.905415 0.014812
0.901946 0.016156
0.899047 0.017719
0.896720 0.019522
0.895009 0.021474
0.893925 0.023517
0.893589 0.025374
0.893925 0.027199
0.894976 0.028901
0.896887 0.030054
0.899614 0.030749
0.903046 0.031232
0.907107 0.031741
0.911738 0.032320
0.916920 0.032993
0.922644 0.033757
0.928889 0.034582
0.935621 0.035469
0.942821 0.036418
0.950452 0.037440
0.958495 0.038526
0.966901 0.039670
0.975646 0.040906
0.984677 0.042203
0.993974 0.043586
121
Table B.2 The coordinates of flap (continued)
1.003479 0.045045
1.013157 0.046609
1.022976 0.048257
1.032883 0.050011
1.042829 0.051865
1.052791 0.053820
1.062706 0.055875
1.072539 0.058038
1.082262 0.060300
1.091824 0.062652
1.101185 0.065091
1.119166 0.070188
1.135943 0.075465
1.151267 0.080787
1.164948 0.085915
1.176820 0.090662
1.186673 0.094993
1.194459 0.098572
1.200104 0.101159
1.203556 0.102670
1.204699 0.103214
Approval of the Thesis: A TWO DIMENSIONAL EULER FLOW SOLVER ON ADAPTIVE CARTESIAN GRIDS submitted by BERCAN SİYAHHAN in partial fulfilment of the requirements for the degree of Master of Science in Mechanical Engineering Department, Middle East Technical University by,
Prof. Dr. Canan Özgen Dean, Graduate School of Natural and Applied Sciences Prof. Dr. S. Kemal İder Head of Department, Mechanical Engineering Prof. Dr. M. Haluk Aksel Supervisor, Mechanical Engineering Dept., METU Assist. Prof. Dr. Cüneyt Sert CoSupervisor, Mechanical Engineering Dept., METU
_______________ _______________ _______________ _______________
Examining Committee Members: Prof. Dr. Kahraman Albayrak Mechanical Engineering Dept., METU Prof. Dr. M. Haluk Aksel Mechanical Engineering Dept., METU Assist. Prof. Dr. Cüneyt Sert Mechanical Engineering Dept., METU Prof. Dr. İsmail H. Tuncer Aerospace Engineering Dept., METU Instructor Dr. Tahsin Çetinkaya Mechanical Engineering Dept., METU Date: _______________ _______________ _______________ _______________ _______________
02/05/2008
I hereby declare that all information in this document has been obtained and presented in accordance with academic rules and ethical conduct. I also declare that, as required by these rules and conduct, I have fully cited and referenced all material and results that are not original to this work.
Name, Last name : Bercan Siyahhan
Signature :
iii
The main advantage of using this type of a mesh is the ability to generate meshes around geometries of arbitrary complexity quickly and to adapt the mesh easily based on the solution.S. M. A Cartesian mesh generator is incorporated to the solver. The Cartesian mesh is formed by dividing squares successively into its four quadrants.. For the solution procedure Roe’s method as well as flux vector splitting methods are used for the flux evaluation. Department of Mechanical Engineering Supervisor: Prof. Haluk Aksel CoSupervisor: Assist. Dr. Dr. AUSMD and AUSMV methods. Bercan M. The code is written in the C++ programming language and its object oriented capabilities have been exploited to save memory in the data structure developed. The main disadvantage of this method is that the treatment of the cells that are cut by the geometry. Hence the preprocessing can be performed together with the solution within a single code. Time discretization is performed using a iv . The flux vector splitting schemes used are van Leer. Prof. Cüneyt Sert May 2008.ABSTRACT A TWO DIMENSIONAL EULER FLOW SOLVER ON ADAPTIVE CARTESIAN GRIDS Siyahhan. AUSM. a code to solve the two dimensional compressible Euler equations for external flows around arbitrary geometries have been developed. 121 pages In the thesis work.
The distribution obtained using reconstruction is very close to the experimental one while there is a considerable deviation for the case without reconstruction. Euler Equations. Flux Vector Splitting v . Solution Refinement. The effect of reconstruction is demonstrated by plotting the pressure coefficient on the airfoil. The code is validated by performing calculations around a NACA0012 airfoil profile. Approximate Riemann Solver of Roe. In addition the performance of each method is analyzed for flow around an NLR 7301 airfoil with a flap. To increase the accuracy least squares reconstruction is employed. Also the shock capturing capabilities of different methods have been investigated.multistage method. Keywords: Cartesian Mesh Generation.
Daha sonra zamanda integrasyon için çok aşamalı bir metod kullanılmaktadır. Dr. vi . Çözücüye bir Kartezyen ağ çözücüsü ilave edilmiştir. AUSM. Kartezyen hesaplama ağı karelerin art arda dörde bölünmesiyle oluşturulmaktadır. Çözümde hücre akılarını bulmak için Roe’nun metodunun yanı sıra FVS metodları da kullanılmaktadır. Makina Mühendisliği Bölümü Tez Yöneticisi: Prof. Böylece hem ağ oluşturma hem de çözüm işlemlerinin aynı kodla yapılması sağlanmıştır. Kod C++ dilinin nesneye yönelik programlama özellikleriyle veri yapısının kullandığı hafızayı azlatma amacı güdülerek yazılmıştır. Bercan Yüksek Lisans. Cüneyt Sert Mayıs 2008. Haluk Aksel Ortak Tez Yöneticisi: Yard. Dr. Kartezyen ağların temel artısı değişik karmaşıklıktaki geometriler etrafında hızla ağ oluşturulabilmesi ve çözüme göre ağın uyarlanabilmesidir. FVS metodlarından van Leer.ÖZ UYARLAMALI KARTEZYEN AĞLARDA İKİ BOYUTLU EULER AKIŞ ÇÖZÜCÜSÜ Siyahhan. 121 sayfa Bu tez çalışması kapsamında. Yöntemin doğruluğunu artırmak için hücre içinde akış değişkenleri yeniden yapılandırılmaktadır. Temel eksisi ise geometri tarafından kesilen hücrelerin ele alınmasındaki zorluktur. Doç. M. AUSMD ve AUSMV kullanılmaktadır. herhangi bir geometri etrafındaki iki boyutlu sıkıştırılabilir dış akışları Euler denklemlerinin çözümüyle modellemek için bir kod yazılmıştır.
Çözüm Uyarlamalı Ağ vii . Yeniden yapılandırmanın etkileri kanat üzerindeki basınç katsayılarının çizilmesiyle vurgulanmaktadır. Yeniden yapılandırmalı dağılım deneysel dağılıma yakındır ancak yeniden yapılandırma olmadan elde edilen sonuçlar deneyselden ciddi farklılıklar göstermektedir. Euler Denklemleri.Yöntemin geçerliliği NACA0012 kanat profili etrafındaki akış çözülerek sınanmıştır. Anahtar Kelimeler: Kartezyen Ağ Oluşturucusu. Ayrıca farklı metodların şok yakalama özellikleri de sınanmıştır. Yeniden Yapılandırma. Farklı metodların performansları NLR7301 profili etrafındaki akış çözülerek de kıyaslanmaktadır. Roe’nun Yaklaşık Riemann Çözücüsü.
Dedicated to my family viii .
Çağrı Şişman. Prof. I would also like to thank Emre Gürdamar. cosupervisor Asst. Also I would like to express my appreciation for the company of Ender Özden and Samet Özkan with whom I have shared the same work environment. Cüneyt Sert and Dr. support and constructive criticisms during my thesis work. ix . Haluk Aksel. Dr. Last but not least I would like to express my gratitude for all individuals and institutions that realize and cherish knowledge as something to be shared openly and freely.ACKNOWLEDGEMENTS I would like to thank my supervisor Prof. Güneş Nakiboğlu and Mehtap Çakmak with whom I had the pleasure of having fruitful discussions. Dr. Doruk Özdemir. Ruhşen Çete for their assistance.
.............................................................................................................................................................................................................TABLE OF CONTENTS ABSTRACT ... xiv LIST OF FIGURES ....................................................................................................... INTRODUCTION..........................................................................................1 1................................................................................................................ 1 Main Elements of a CFD Program ....... 2 Mesh Generation .................... 3 Structured Meshes .................. 4 Triangular Unstructured Meshes ............................................2 1................... 3 Unstructured Meshes.......... 4 x 1......................................................... x LIST OF TABLES ...............................................................................1 .................................................3 Overview of Computational Fluid Dynamics (CFD) ............................................................................................................................................ xix CHAPTERS 1.........2.........3.......3.......................................3.. iv ÖZ ................... ix TABLE OF CONTENTS ............................................................................. vi ACKNOWLEDGEMENTS ............. xv LIST OF SYMBOLS ................1 1..............................................................2 1................. 1 1.
......1...............4....... 17 2...1....2 2.....1 2................................................................................................................. 16 LiangBarsky Line Clipping Algorithm ..................2................2......3.........2 2...............................1 1....3 Mesh Generation ......................... 25 Regular Cutcells ..............................................2......................... 10 Neighbour Cell Procedures ......2 2.............................3.................. 7 Finite Difference Method ......... 16 Comparison of Polygon and Line Clipping Algorithms ........ 8 1..........................................................1 2............. 12 Information for Cut and Computational Cells .............................................. 10 2..........2............................. 7 Finite Volume Method ............... DATA STRUCTURE & MESH GENERATION .........................................................3....................... 21 Uniform Mesh Generation ......................................4......................................................4 Cartesian Meshes ..........3.............1 .... 7 Finite Element Method.. 9 2.....................................2 1...... 13 2....................2 Clipping Procedure ............1 2........................................................................................................5 Assessment of the Solution Methods . 21 Box Adaptation & Celltypes ............... 28 xi 2...1 Data Structure..................................................................................................3 1...........................................1.......................3 2.............................................4...... 6 Solution Methods ............1.....2 1........ 10 Terminology for the Cartesian Cells .......
................................................ 42 The Approximate Riemann Solver of Roe .............. 64 xii 3.......................1 3..............................1 3.............2 3............. 56 Van Leer Flux Vector Splitting ..... 47 Solution Algorithm ...................2 ConvexUnited ................................................... 61 Reconstruction of the Flow Variables . 40 3.4....5 3................................2 ..........................3.........3 Governing Equations..........3...... 57 AUSM (LiouSteffen) Flux Vector Splitting ...............2. 37 3........ 31 2................................2.......................2....................3..........2 Irregular Cutcells ............3...4................................ 54 3.......................... 62 GaussGreen Reconstruction..................................2......................3...... 40 One Dimensional Riemann Problem for Linear Systems .......4 3.................................................................................... 61 3.................. 58 AUSMD Flux Vector Splitting ........................4 Flux Vector Splitting ........................6.............6....................................3 3...............2.........................1 3.......................................... NUMERICAL METHOD ....... 35 2..................................... 63 Least Squares (Minimum Energy) Reconstruction ..................................................................1 ConcaveSplit ..........................................................6 Boundary Conditions ............................. 59 AUSMV Flux Vector Splitting ..............2 3..4...2..........................3 Curvature Adaptation ........................................................................1 3.................. 32 2.............4..
............ 109 APPENDICES A..............2 4............................ 113 B................................................ 72 Subsonic Tests..........7 Gradient Limiting ...............................8 Solution Refinement ...................................................................... CONCLUSIONS .6..................................................................................... 116 xiii .............................................................................................................................................2 NLR 7301 with Flap ..................................7............... 88 5........... 72 Transonic Tests ......................3............... RESULTS & DISCUSSIONS ............7...........................1............................................ 72 4....1 3............... 70 3.......2 3.......................................................................... 106 BIBLIOGRAPHY ....................1 4. 81 4.............................. COORDINATES OF NACA0012 .............................................................................. 68 Explicit Time Stepping Schemes ........................................... 67 Temporal Discretization .............. 71 4...........................................................3 3............................................1 NACA0012 ........................ 68 Determination of the Time Step ........................ COORDINATES OF NLR 7301 & FLAP ............................................................1...........
.Information stored for different types of cells ..........1.............. 98 Table A............ 78 Table 4..2.........The coordinates of NLR 7301.2... 15 Table 3.......................................................... 84 Table 4...The coordinates of flap ............................1........LIST OF TABLES TABLES Table 2...........1.Computational performance of the methods for NLR 7301 AoA 10.... 1.......Comparison of first and second order solutions .........Computational characteristics of the methods for evaluating fluxes ..........1........ 69 Table 4.............Computational performance of the methods for NLR 7301 AoA 6˚ ........................Stage coefficients and CFL numbers for the multistage method ..............Coordinates of NACA0012 Profile..........1˚ . 116 Table B..................1˚ .... 92 Table 4.............4.... 119 xiv ...... 113 Table B.... 95 Table 4.............Computational performance of the methods for NLR 7301 AoA 13.5..3.....
... 5 Figure 1...................Examples of regular cutcells ..... 27 Figure 2....8.....................1...................6..........................Imaginary boxes around a three element airfoil .............................................................Grouping and treatment of irregular cutcells ........10.......... 6 Figure 2................The tree structure .Vornoї diagram for the grid points ....................................2............. 18 Figure 2..... 11 Figure 2.........LIST OF FIGURES FIGURES Figure 1................ 24 Figure 2............................................1...Automatically known neighbours of a cell ..........Neighbours associated with the parent cell’s edge neighbours ............................. 23 Figure 2............... 28 Figure 2....The stages of the advancing front method ........... 30 Figure 2..............7...................Mesh around the slat and the airfoil ..... 12 Figure 2........Conversion of an irregular cutcell ..........Neighbours of a cell ................... 32 Figure 2.............................................................................5.........................The visible and invisible regions for a clipping window ........ 33 xv .3......................................................The relation between cells ..................12.......................................11....................4..... 11 Figure 2...........................2................... 26 Figure 2.....9.....Naming convention of a cell ...........
.....13......Cp vs........ 44 Figure 3.................3......... 36 Figure 2...........Figure 2.....16................................3..Graphical interpretation of flux vector splitting ......................... 76 Figure 4.... 34 Figure 2...The normals of two neighbouring cutcells.. chord for second order solution ............. 38 Figure 2.....17..2......... 36 Figure 2.........15.... 62 Figure 3.1..............Cp vs.....Concavesplit to special subgroup of regular cutcells conversion . 75 Figure 4.......................................Lift coefficient vs...........................Convexunited to inside cells conversion.......................... angle of attack graph.5................................ 57 Figure 3.........Convexunited to regular cutcells conversion.. 73 Figure 4...................A refined and nonrefined mesh around stagnation point .18......................................Residual plot for first and second order solutions ...................6....................Mach contours for NACA0012 subsonic flow ...Solution of a 1D Riemann problem for a set of linear equations ............ 35 Figure 2......................1................. 38 Figure 3................. 74 Figure 4...Calculation of the normal ..............................4................Comparison of the second order solutions ...... 77 Figure 4....................Region formed by the support set......14..........................Concavesplit to regular cutcells conversion .4.................. 80 xvi .....................2............... chord for first order solution ..................... 79 Figure 4........... 64 Figure 4.....................Ghost right state at the solid wall ...........................7............
........ Roe............. 93 Figure 4................. 82 Figure 4...1˚) ...................17.. 86 Figure 4..... 96 Figure 4............................... chord graph (AUSM methods AoA 6˚)...1˚)....Mach contours for transonic flow around NACA0012 airfoil ...............Pressure coefficient vs......... 90 Figure 4.9..........................Pressure coefficient vs.....Definition of geometry variables for NLR 7301 with flap.15... chord graph (AUSM methods AoA 13....................... Roe.. 87 Figure 4. 85 Figure 4......................... chord graph for Roe............................22.16............ chord graph (AUSM....................11................................. 99 xvii ..................10....8......21..........Effect of solution refinement around stagnation region of NLR 7301.................................... van Leer AoA 13.... 83 Figure 4...... Roe.......................Pressure coefficient vs..... chord graph (AUSM............... 86 Figure 4.... van Leer AoA 6˚) ...1˚) ....... 88 Figure 4.......18...... chord for AUSMV with and without refinement .........................20. van Leer and AUSMV .. 94 Figure 4................ chord graph (AUSM methods AoA 10....Effect of solution refinement around the flap .......19..13............Effect of refinement at shock location ............. chord graph (AUSM...Residual plot for transonic flow ...................Pressure coefficient vs..Pressure coefficient vs.. chord graph for the AUSM methods .................................. 89 Figure 4..........Pressure coefficient vs................ 99 Figure 4..........23........Figure 4. 97 Figure 4...12..Pressure coefficient vs..................Pressure coefficient vs...Effect of refinement around stagnation point .........Pressure coefficient vs. van Leer AoA 10....1˚)............14.......... 91 Figure 4......
...1˚ ... 104 xviii .............1˚ . 101 Figure 4....Flow field for NLR 7301 with flap AoA 13...................... 102 Figure 4....... chord for van Leer AoA 6˚ ........25.26.......... chord for van Leer AoA 10....24......Flow field for NLR 7301 with flap AoA 6˚ . 103 Figure 4........... chord for van Leer AoA 13...27.............29................Pressure coefficient vs...Figure 4.Pressure coefficient vs......................1˚ ..Flow field for NLR 7301 with flap AoA 10.... 103 Figure 4..................................1˚ ..28... 100 Figure 4.....Pressure coefficient vs..
LIST OF SYMBOLS U: vector of conserved variables α : angle of attack T : temperature M : Mach number R : gas constant T : transformation matrix A : Jacobian of the flux vector V : vector of characteristic variables λ : eigenvalue of the Jacobian matrix Av (u ) :averaging function ˆ U :vector of rotated conserved variables F(U): flux vector function ˆ ˆ F (U ) : rotated flux vector function F : flux vector dV: volume element dA: surface area element n : face normal t: time Acell: area of a cell ∆s : length of cell face ρ : density Φ : gradient limiter R : residual function β : stage coefficients u : velocity in the xdirection v : velocity in the ydirection ht : total enthalpy p : pressure ρ : Roe averaged density u : Roe averaged velocity in the xdirection v : Roe averaged velocity in the ydirection σ : CFL number Ψ : convective spectral radius S : length of projection Subscripts c : value at centroid with th R : right state variable L : left state variable stag : stagnation condition inf : free stream x : derivative in x component y: derivative in x component Superscripts ~ : Roe averaged quantity ˆ : components in normaltangential coordinates T : transpose .: averaged quantity max : maximum value min: minimum value x: component in the xdirection y: component in the ydirection ht : Roe averaged total enthalpy p : Roe averaged pressure et : total energy ˆ u : normal velocity ˆ v : tangential velocity γ : specific heat ratio a : speed of sound V : velocity a : Roe averaged speed of sound V : Roe averaged velocity xix .
theoretical and experimental methods. often combined.1 Overview of Computational Fluid Dynamics (CFD) Computational fluid dynamics is a branch of fluid dynamics which employs numerical methods for solving the equations of fluid flow which are impossible to solve analytically due to their complex nature.CHAPTER 1 INTRODUCTION In this thesis. It can be said that emergence of CFD as a tool for solving flow problems has taken place only in the mid 1900s so it is a relatively new field of study compared to analytical and experimental methods. 1. arguably with the pioneering work of Kopal who in 1947 formed tables for the supersonic flow over sharp cones after solving the governing equations numerically. numerical methods are employed to resolve external flows around arbitrary geometries. With the development of digital computers. were the only tools for analyzing fluid flows of differing nature (1). solving the Euler equations numerically with the finite volume method for the desired flow quantities and obtaining the necessary aerodynamic forces to verify the solution. Hence CFD is the third approach for solving a flow problem alongside analytical and experimental methods. In the 1950s and 1960s CFD. To understand the motivation for choosing the methods followed. CFD has become an indispensible tool in the field of fluid dynamics. The thesis work consists of generating a Cartesian mesh around the input geometry. a brief introduction must be made to mesh generation and solution techniques employed in the field of computational fluid dynamics (CFD). has been applied to reentry problems and 1 . Since the publication of Isaac Newton’s Principia in 1687. however.
2 . This is accomplished by converting the governing partial differential methods to algebraic set of equations in the three conventional finite difference.2 Main Elements of a CFD Program Any CFD program mainly consists of three parts. the preprocessor. Contours of desired scalars and the vector fields are displayed within the domain of interest.has established its place among the three essential methods for analyzing a fluid flow problem (1). For the postprocessor a commercially available program Tecplot and a freely distributed program Mayavi have been used. The mesh generation and different solution methods will be discussed in some detail below. The postprocessor is the part where the solution obtained is visualized. finite element and finite volume methods. The preprocessor gathers the inputs to the problem being analyzed. 1. The program developed in the scope of the thesis work contains a preprocessor which gathers the inputs of the problem and generates a Cartesian mesh in the computational domain and a solver which uses the finite volume method for the solution of the compressible Euler equations in the conserved form. geometry around which the fluid flow will be investigated. boundary conditions. Also nondimensional variables and physical quantities of interest are calculated so as to verify the solution and to judge whether the mesh generated and the schemes used are appropriate for the problem at hand. Also the grid is generated around the input geometry by the preprocessor. The solver part of the program uses numerical techniques to resolve the flow around the input geometry. the solver and the postprocessor (2) . the fluid properties.
j. the elliptic grids are obtained as a result of solving an elliptic differential equation to obtain a mesh that satisfies the Laplace equations while for the hyperbolic grids are obtained by solving hyperbolic equations (4). The structured grids can be classified further according to the technique used to generate them as algebraic. The algebraic grids are formed simply by transforming cells in a Cartesian computational domain to a physical domain to obtain a boundary conforming mesh.3 Mesh Generation Mesh generation is the most time consuming and the most crucial step in obtaining a solution to a fluid flow problem.1 Structured Meshes A mesh is considered to be structured if the organization of the grid points and the form of the grid cells depend on a general mathematical rule instead of their position (3). They require less data storage since the connectivity information is implicitly known. In other words. j. Hence the connectivity of the grid points to form the cells is implicitly implied by the general mathematical rule. the grid points can be arranged as a regular array as (i. Each one of these two types of meshes can then be classified further according to the method used to construct them. k) with the neighbouring point being known as (i. meshes for viscous solutions can be obtained.1. there are two types of meshes. The advantages of structured meshes are that they are boundary conforming so by adjusting the grid spacing. since the governing partial differential equations or the algebraic rules must be solved and computations must be performed for the transformations from the 3 .3. k+1) (4). the structured and the unstructured meshes. 1. The generated mesh should comply with requirements of the solution schemes to be used in order to get accurate results. One disadvantage of the structured meshes is that they require more computation power. elliptic and hyperbolic grids. In general.
Then. In this approach. 1. they are more suitable to automatic meshing. where the front folds 4 .2 Unstructured Meshes If the grid points cannot be arranged as a regular array and additional information is required for the connectivity of the grid points then the mesh is unstructured. the points that are available for triangle formation termed as the front are connected to newly added points to update the front until the whole domain is meshed (4). The unstructured meshes. In the advancing front method. are formed of triangular or rectangular elements generally in the twodimensional case. 1.1 Triangular Unstructured Meshes The most widely used methods for forming triangular unstructured meshes are the advancing front and the Delaunay triangulation methods.2. the closing stage of this approach. though they can comprise of elements of any shape. and no transformation between a computational and a physical domain will be needed.computational to physical spaces.3. extensive search algorithms must be employed when adding new grid points to ensure that the newly formed cells adjust to the existing ones. While the disadvantage of these meshes is that they require more complex data structures for storing the connectivity of grid points and the neighbour cell information. Also. the points are added to the ones that represent the geometry as the mesh is created. The advantages of the unstructured meshes are that they can be applied to geometries of arbitrary complexity with ease. The added points are connected to already existing points to form triangles. However the most important disadvantage of the structured meshes is that they cannot easily be applied to complex geometries that are formed by more than one body that have small clearances or are not well streamlined.3.
over itself.The stages of the advancing front method The other and may be the most popular triangular meshing procedure. An example of the advancing front method is illustrated in Figure 1.2. introduces some difficulties (3).1. 5 .1. Once the diagram is formed. Hence they can be obtained by the perpendicular bisectors of the lines between neighbouring points. This is illustrated in Figure 1. The circumcircles of the triangular elements do not contain any other point in the domain and the resulting mesh is unique for the point distribution. the Vornoї diagram is updated locally and new triangles are formed. requires points to be created before the meshing process and is closely linked to the concept of Vornoї diagrams. Delaunay triangulation. Figure 1. This method is suitable for mesh adaptation since for a newly added point. The Vornoї diagram is obtained by tessellating the domain into Vornoї regions which are regions that are closer to one particular point within the domain than the others. any points that have a Vornoї edge between them are connected to form the triangular mesh.
If the quadtree approach is followed strictly.3.Vornoї diagram for the grid points 1.Figure 1.2. then the surface of the geometry cannot be represented accurately hence the cutcells are used to alleviate this problem which introduces some complexity to the approach.2 Cartesian Meshes The Cartesian mesh is obtained by using the quadtree approach in two dimensions. which is generally undesirable. 6 . in the computation stage. the procedure followed is much simpler than the methods for generating triangular unstructured meshes. In the quadtree approach. Also the approach is simpler than triangular mesh generation techniques and it will promote simpler calculations in the solution stage. since for the adaptation. the flux vectors won’t have to be rotated for the greater portion of the cells. This approach was chosen to mesh arbitrary geometries of differing complexities in an automatic fashion. The code developed uses the quadtree approach to form a Cartesian mesh for the solution. Hence. the cell has to be simply divided into its four quadrants. a square containing the whole domain is divided into its four quadrants to form the mesh in a tree structure. The orientation of the geometry may cause formation of very small cutcells. The method is very suitable for adaptation.2. Also.
4.1 Finite Difference Method The finite difference method is based on the differential form of the governing system of partial differential equations. Then the unknown coefficients of these functions are found by establishing algebraic equations through satisfying the governing equations in a weighted residual form over each element (6). Hence. The most widely used methods for the discretisation of the governing equations are the finite difference.1.4 Solution Methods The partial differential equations governing the flow of a fluid must be discretised in order to convert them into algebraic equations and get a numerical solution. These differences are derived from the Taylor series and their order of accuracy is dependent on the number of points used in the difference equation. The derivatives of the conserved quantities appearing in these equations are approximated by differences of these unknown quantities at neighbouring grid points. finite element and finite volume methods. the method is not very suitable for unstructured meshes or cases where the geometry is complex. 1.4. Once the difference equations are written in terms of unknown quantities. 7 . divides the computational domain into triangular or rectangular subdomains and represents the variation of the unknown quantities in terms of piecewise continuous functions. 1.2 Finite Element Method The finite element method which was initially developed for structural problems from 1940s through 1960s. these algebraic equations can be solved numerically with the boundary conditions to obtain the value of the conserved quantities at each grid point. The major setback of this method is that it requires a regular structured mesh in order to express the derivatives accurately (5).
the basic concepts of the different schemes are more comprehensible than the finite element method (2). conservation of mass. 1. 8 .3 Finite Volume Method The finite volume method which was first devised as a special form of the finite difference method in the 1970s uses the integral form of the governing equations of a fluid flow. or at the vertices for a cell vertex formulation. Even though this method is very well established in structural mechanics. This property of the finite volume method promotes development of numerical schemes based on physics of the flow. which are used for evaluating the convective terms is one example. since the method is directly based on the governing integral equations. Also.The finite element method. There is a wide range of methods for discretising the convective and diffusive terms of the governing integral equations as well as the definition of the control volumes for which the governing equations are satisfied. Hence the method has a considerable inherent flexibility (5). Approximate Riemann methods. it is still in the stages of evolution for complex flow problems such as compressible flows governed by Euler or NavierStokes equations (1). Another possibility is to have a staggered grid where the scalar quantities and different vector quantities can be stored on overlaid cells as in the SemiImplicit Method for PressureLinked Equations (SIMPLE) algorithm (2). has been investigated by mathematicians since its development for engineering purposes. One other advantage of the method is with the direct discretisation of the governing integral equations.4. with its mathematical background being functional analysis. momentum and energy will be guaranteed at a discrete level (5). and is a very rigorous method with specific conditions for existence and convergence criteria and exactly derived error bounds. More specifically it is possible to store the flow variables at the centroid of the cells for a cell centred approach.
a numerical scheme must possess the conservativeness. a nondirectional central method will be more appropriate.In the code developed. 9 . 1. Since this is not reasonable. a numerical solution will tend to the exact solution if infinite number of cells is used. For example. boundedness and transportiveness properties (2). The boundedness of a solution guarantees that numerical errors introduced by the discretisation of the governing equations will not prevent convergence. whereas for a flow that is diffusive. The conservativeness property states that flux leaving through a certain face of a cell must enter the adjacent cell. the transportiveness of numerical scheme deals with the directionality of the solution. hence upwind schemes will be more appropriate.5 Assessment of the Solution Methods Theoretically. cell centred finite volume method has been used with of Roe’s approximate Riemann method for the evaluation of the convective fluxes through the cell surfaces as well as the flux vector splitting schemes. This property is related to the stability of the discretisation scheme used. the solution of a flow dominated by convection will depend more on the upwind direction. The stability of a scheme can be analyzed by the discrete perturbation analysis or the von Neumann stability analysis (7). Finally.
starting from a single large square. the terminology adopted must be explained. A cell (square) is called the parent of its four quadrant cells which form as a result of division and the four quadrant cells are called the children of the parent cell in turn. The children cells are numbered according to the quadrant they occupy. the first step in the mesh generation is to construct an appropriate data structure that will store the necessary data for each cell in the domain. A cell without any children is called a leaf cell. The leaf cells can also be called the computational cells since the actual calculations for the field variables are performed for these cells. However before describing each variable that is stored. 10 . by connecting the midpoints of opposing edges until the desired resolution is obtained. The large square comprising the outer boundary of the domain is termed as the root cell. a Cartesian unstructured mesh is generated around the input geometry. Since an unstructured mesh is used. This process results in a tree structure.1. 2.1 Data Structure In the code developed. These concepts are illustrated below in Figure 2.1 Terminology for the Cartesian Cells The mesh is obtained by dividing squares successively.CHAPTER 2 DATA STRUCTURE & MESH GENERATION 2.1.
The root cell is assigned a level of zero.The tree structure 11 . The other cells mainly serve the purpose of traversing the tree.2 to exemplify some of the concepts defined above. The tree structure associated with the structure shown in Figure 2.The relation between cells parent Among the computational cells.1. some are cut by geometry lines and these are called cutcells. The level of the four children of a cell is one higher than that of their parent. root level 0 child 1 child 1 child 2 child 3 child 2 child 4 child 1 child 3 child 2 child 3 child 4 child 4 level 1 level 2 leaf cells Figure 2.2.1 is illustrated in Figure 2. Also there is a level concept associated with each one of the cells.child 2 child1 child 3 child 4 leaf cell Figure 2.
southeast. A cell will have four neighbours along its edges called the edge neighbours. but if it is at one higher level the neighbour’s parent will be recognized instead of having two neighbours along the edge. northwest. then it will be recognized as the neighbour along the corresponding edge. southwest. y x no northeast north neighbour northwest neighbour west neighbour S1 S2 SE southwest neighbour south neighbour east neighbour southeast neighbour Figure 2. the neighbour along that edge must be known.3 illustrates the concepts related to neighbouring relations and also the possible configurations that might occur for a cell. The general rule for the mesh is that a cell can have neighbours of levels that are one degree different than the cell. north and south according to the edge they are associated with. Also the neighbours will be used to determine the variation of the conserved variables within the cells. These neighbours are named east. This ensures the grid smoothness and also simplifies flux calculations through the faces. If a neighbour is at the same level or one lower level. Figure 2. Also computational cells will have four neighbours through their vertices named as vertex neighbours.Neighbours of a cell 12 .2. northeast.3.2 Neighbour Cell Procedures As was mentioned before for unstructured meshes the neighbours must also be stored since for the calculation of the fluxes through a face.1. west.
Since the north neighbour is at a lower level the north and northeast neighbours are identical in which case the cell is considered to have no northeast neighbour.3 Information for Cut and Computational Cells As it is defined earlier there are three types of cells. These cells are recognized as the neighbours themselves meanwhile the parent of the cells marked as S1. the x and y coordinates of the centre and the level of the cell are stored. Besides the thirteen pointers aforementioned there are nine more pointers which are allocated only for computational cells. Also there are a total of thirteen pointers pointing to other cells in the domain. This pointer is null only for the root cell. The information stored for a parent cell is shared for all types of cells. the computational leaf cells and the cutcells. The remaining one of the thirteen pointers points to the parent of the cell. the remaining four point to the vertex neighbours and they are allocated only for computational cells. One stores coefficients that will be used in solution 13 . Of the remaining five pointers two are storing the gradient and the curl of the velocity vector. 2. Of these eight pointers four point to edge neighbours.3 the east. which are special types of computational cells. if the cell is a leaf cell then these pointers are null pointers. momentum and energy equations. west and southwest neighbours of the cell are at the same level while the north and northwest neighbours are at one lower level. S2 and SE are recognized as the south and southeast neighbours respectively. hence a vertex neighbour of a cell should be distinct if it is to be stored. One pointer stores information dictating the cell to be refined at a stage of mesh generation or solution adaptation. The amount of data stored for each cell depends on its type. For each cell in the domain.1. Eight of the remaining nine pointers point to the neighbours of the cell. while computational cells will have additional information and for cutcells further additional information will be stored. the parent cells for traversing the tree.In Figure 2. Four of these pointers store the conserved field variables for the continuity. Of these thirteen pointers four point to the children of the cell.
Of the remaining two pointers one stores which edges are cut by geometry lines while the last one stores information on which portion of the cell’s edges flux passes through. Finally there is a set of seven pointers which are allocated for the cutcells. 14 .reconstruction and the last one stores the type of the cell. whether the cell resides totally outside or inside the geometry or whether it is cut by it. Three of these seven pointers store the x and y coordinates of the centroid and area of the cell while two store the x and y coordinates of the end lines of the intersections of geometry lines with the cell edges. The table below shows a summary of the information stored for different types of cells.
1. cutcell) neneigh: pointer to northeast neighbour seneigh: pointer to southeast neighbour nwneigh: pointer to northwest neighbour swneigh: pointer to southwest neighbour CUTCELLS centx: pointer to x coordinate of the centroid of the cell centy: pointer to y coordinate of the centroid of the cell area: pointer to the area of the cell xcut: pointer to the x coordinate of cut location on edge ycut: pointer to the y coordinate of cut location on edge edgecut: pointer to edges cut byb the geometry fluxcut: pointer to the information regarding flux 15 . outside.Table 2.Information stored for different types of cells ALL CELLS xcent: x coordinate of center of square containing the cell ycent: y coordinate of center of square containing the cell level: level of the cell child1: pointer to first quadrant child child2: pointer to second quadrant child child3: pointer to third quadrant child child4: pointer to fourth quadrant child eneigh: pointer to east neighbour wneigh: pointer to west neighbour nneigh: pointer to north neighbour sneigh: pointer to south neighbour parent: pointer to parent COMPUTATIONAL CELLS ro: density associated to the cell rouvel: product of density and velocity in the x direction associated to the cell rovvel: product of density and velocity in the y direction associated to the cell roen: product of density and total energy associated to the cell gradient: pointer to gradient of velocity curl: pointer to curl of velocity recons: pointer toarray containing geometry based reconstruction coefficients refine: pointer to flag for refinement celltype: pointer to the cell type (inside.
Hence. line clipping if the input geometry is conceived as a collection of individual line segments. 2. a few of the line clipping algorithms were compared amongst themselves. SutherlandHodgeman (8). SutherlandCohen (10) algorithm. the performance of their algorithm was tested along with the SutherlandCohen line clipping algorithm with 4 sets of data containing 1000 randomly generated line segments.1 Comparison of Polygon and Line Clipping Algorithms When the polygon and line clipping algorithms are compared. it is seen that the simplest polygon clipping algorithm. The two options for the clipping procedure are the polygon clipping.2. the application of these algorithms would bring unnecessary complications. It was 16 . Clipping is the process of determining the portion of a geometrical shape that resides within a given clipping window. the line clipping algorithms are preferred to polygon clipping algorithms. the procedure for determining the location of the points cut by the input geometry for the cutcells must be explained.2 Clipping Procedure Before describing the mesh generation process in detail. However there are other polygon clipping algorithms that can handle arbitrary geometries such as self intersecting geometries and geometries with internal holes.2. For the code developed. To reduce the effects of random variation. in general. if the input geometry is conceived as a polygon. which is a polygon comprising of connected line segments. After this general comparison. each line segment was clipped 1000 times resulting in a total of 4 million clippings. is not capable of handling concave geometries effectively. is to be clipped and the clipping windows are the computational cells within the domain. Since these cases are not in the scope of the current study. First the LiangBarsky (9) line clipping algorithm is compared with the most classical algorithm. In the work of Liang and Barsky (9). the input geometry.
the simplest algorithm that is the easiest to implement has been chosen.2. Hence. However the algorithms mentioned are considerably more complex. After each test the invisible part of the line will be discarded. Instead it is suggested that a measurement based on operation count should be used (11). the portion of the line segment that is inside the clipping window will be obtained. Thus the LiangBarsky line clipping algorithm has been employed to determine the location of the points that are cut by the input geometry lines. its effect on overall computation time will not be very significant. A study conducted by J.reported that an improvement of about 36 percent in the execution time was obtained. The line obtained from the segment and its extensions will be tested with each one of the lines comprising the boundary of the clipping window to determine which portion of the line is visible.2 LiangBarsky Line Clipping Algorithm The LiangBarsky line clipping algorithm (9) tests the extensions of given line segments against each one of the four boundaries of a rectangular clipping window independently. to determine which if any portion of the line segment is inside the clipping window. SobkowPospisilYang (13) algorithms are more efficient in both aspects compared to the LiangBarsky algorithm. Even though Liang and Barsky claim that their algorithm is more efficient than Sutherland Cohen. Day (12) takes both the experimental and operations count view into consideration and reveals that NichollLeeNicholl (11). After all the boundaries are tested. A line segment with end points V0 and V1 which happens to reside wholly within a clipping window is shown in Figure 2. in fact the latter algorithm classifies lines into 81 different cases.4. 2. 17 . Since the clipping procedure will be carried out once at the mesh generation step for all the computational cells and for the cells formed as a result of solution adaptation. other researchers postulate that experimental analysis is inadequate to measure the performance of a clipping algorithm. D.
invisible ytop visible invisible line segment V1 V0 visible ybottom xleft xright extension of invisible line segment Figure 2.4) Adopting this representation t=0 corresponds to the first end point V0 and t=1 corresponds to the second end point V1. the clipping window can be visualized as the combination of four boundary lines each having a visible and an invisible region as 18 .2) (2. y1) as given in equations (2.The visible and invisible regions for a clipping window The algorithm makes use of the parametric form of the line segments with end points V0 (x0. Once the line is parameterised.3) (2.1) to (2. The orientation of the line is considered to be from point V0 to V1.1) (2.4. y0) and V1 (x1. x = x0 + ∆x ⋅ t y = y0 + ∆y ⋅ t ∆x = x1 − x0 ∆y = y1 − y0 (2.4).
4..5) (2.1) to (2..7) −∆y ⋅ t ≤ y0 − ybottom (2.4).9) p1 = ∆x p2 = −∆x p3 = ∆y p4 = −∆y q1 = xright − x0 q2 = x0 − xleft q3 = ytop − y0 q4 = y0 − ybottom It is observed that when pi is positive. the orientation of the extended line is from the visible side of a particular boundary line to its invisible side.. 4 (2.6) (2.in Figure 2. The visible region for the clipping window can be represented in terms of four inequalities using the parametric relations (2.8) Finally these relationships can be generalized to determine the parametric values corresponding to the segment of the line that resides inside the clipping window. Ultimately. This will result in an upper limit for the visible portion of the line. the intersection of all the visible sides of each boundary line would give visible region for the whole clipping window. pi ⋅ t ≤ qi where i = 1. 19 . ∆x ⋅ t ≤ xright − x0 −∆x ⋅ t ≤ x0 − xleft ∆y ⋅ t ≤ ytop − y0 (2.
3. i = 1. Thus for the case when pi is different than zero. 4⎬ . 2. qi pi t≥ (2.10) When pi is negative. {0} ⎟ ⎜ p ⎟ ⎭ ⎝⎩ i ⎠ ⎛⎧q ⎞ ⎫ t1 = min ⎜ ⎨ i  pi > 0. {1} ⎟ ⎜ p ⎟ ⎭ ⎝⎩ i ⎠ The t0 and t1 values obtained as a result of this minmax problem should satisfy the following equation t0 ≤ t ≤ t1 ⇒ t0 ≤ t1 (2.3. i = 1. hence a lower bound for the extended line is expected. 2. the problem of determining the values of t for which a certain portion of the line segment is within the clipping window becomes a minmax problem in the form as in t ≥ t0 t ≤ t1 where (2.t≤ qi pi (2.14) 20 . the orientation of the line is from the invisible side to the visible side.13) ⎛⎧q ⎞ ⎫ t0 = max ⎜ ⎨ i  pi < 0. 4⎬ .11) Since the portion of the line segment and not the extended line that resides in the clipping window is sought another condition that the bound obtained for t should be between 0 and 1 must also be implemented.12) (2.
refinement is performed when the cell level is less than the desired level. The final step is to refine the cells around the places where there are large gradients on the bodies.Finally. If a cell is to be refined. inside. are determined at this stage. Uniform mesh generation is fairly straightforward.3. then that neighbour must be refined prior to the refinement of the cell. the case when pi is zero should be considered. the line segment is visible for that boundary line else it is invisible. cutcell. In this way. the one level rule is preserved in the process of 21 . The type of the cells. The first one is the uniform mesh generation where a mesh of uniform sized elements of the desired level is obtained. If qi is positive. first the level of the edge and vertex neighbours must be investigated. outside.15) The validity of this statement is solely dependent on the value of qi. The procedure followed is similar to the one outlined in (14). If a neighbour is at a lower level. This is referred to as the trivial reject case. For this case a relationship independent of t is obtained as 0 ≤ qi (2.1 Uniform Mesh Generation An initial uniform mesh is formed in the solution domain which will be refined in the later stages to form the mesh on which the solution is to be performed. 2.3 Mesh Generation The mesh generation is performed in three main steps. The case pi equals to zero corresponds to a vertical or a horizontal line. The second one is the box adaptation in which cells around an imaginary box (rectangle) of given dimensions around the input bodies are refined till they reach a desired size. starting from the root cell. 2. The reason for forming this initial mesh is to have a prescribed resolution in the outer boundaries of the domain.
The neighbour search can be explained in detail for a first quadrant child exemplifying the process for the other quadrant children. This rule enforces a smooth mesh and also eases the calculations for the fluxes.mesh generation. the neighbours of each cell must also be determined. yc = y parent + (2. When a cell is to be refined. Then the level of the children can be set to be one more than that of their parent and their centres can be calculated from the centre coordinate of their parent’s.16) xc = x parent − .19) for the 1st quadrant to 4th quadrant children. Where d is the domain size and l is the level of the cell.17) xc = x parent − . yc = y parent − (2. it is simply divided into its four quadrants which are its children. respectively. yc = y parent − (2. The other two edge neighbours and one vertex neighbour will be associated with the parent’s edge neighbours while the last vertex neighbour will be related to a neighbour of a neighbour of the parent cell. yc = y parent + (2. similar operations will be performed for each group of cells. Thus.19) ( d / 2) 2l ( d / 2) 2l ( d / 2) 2l ( d / 2) 2l ( d / 2) 2l ( d / 2) 2l ( d / 2) 2l ( d / 2) 2l xc = x parent + . the overall size of the domain. The one level rule states that the difference in the levels of two neighbouring cells cannot exceed one. The procedure followed is similar to the one outlined by De Zeeuw (14). For any cell two edge neighbours and one vertex neighbour will be known automatically since they will be the children of the cell’s parent. For a first quadrant child. 22 . Once this information is obtained. and its level as (2.16) to (2.18) xc = x parent + .
they might have leaf cell children at the same level as the cell and they might have children with leaf cell children at one level more than the cell. the children of the parent’s neighbours must be evaluated.the first group of neighbours. For a first quadrant child these are the east and north neighbours of the parent specifically. Due to the one level rule there are three possibilities for these neighbours of the parent. south and southwest neighbours. They are the second. they might be leaf cells hence they are at a lower level than the cell.5. These cases are illustrated in Figure 2. ones known automatically.6 for the east neighbour of the parent.5. associated with the parent’s neighbours. 23 . west neighbour cell south southwest neighbour neighbour Figure 2.Automatically known neighbours of a cell For the second group of neighbours. fourth and the third quadrant children of the cell’s parent and they are at the same level. This can be depicted as in Figure 2. will be the west.
In Case 3. Hence. it can be said that. the cell stored as the neighbour will be either at a lower level than the cell or it will be at the same level. 24 . During the calculations it is always checked to see if a neighbour has children to identify the cells that are marked as E1 and E2 in Figure 2. the third quadrant child of the cell’s parent’s east neighbour is stored as the southeast neighbour even though the second quadrant child of this cell stored is the only actual southeast neighbour. is stored as the east neighbour. the east neighbour is the second quadrant child of the parent’s east neighbour while the southeast neighbour is the third quadrant child of the parent’s east neighbour. the second and third quadrant children of the parent’s east neighbour’s second quadrant child. Following the same convention. In Case 2.6.cell E cell E SE Case 1: East neighbour of parent is at a lower level cell Case 2: East neighbour of parent has leaf children cells E1 E2 SE Case 3: East neighbour of parent has children with leaf children cells Figure 2. their parent. the second quadrant child of the cell’s parent’s east neighbour.6. as a general rule. the east neighbour of the cell is directly the east neighbour of the parent and there is no southeast neighbour due to the fact that this cell does not have any children. there are two actual east neighbours.Neighbours associated with the parent cell’s edge neighbours In Case 1. Since it is not desirable to complicate the data structure by storing the east neighbours separately.
25 . With the same line of reasoning. The latter described imaginary boxes for a three element airfoil configuration can be depicted as in the figure below. the fourth and the third quadrant children of this cell will be stored as the north and the northwest neighbours of the cell regardless of them having children. Box adaptation will allow the mesh to have a desired resolution in regions around the input geometries. The same search is also performed through the north neighbour of the parent in case its east neighbour resides within an input geometry.With similar reasoning. This neighbour is the north neighbour of the parent’s east neighbour if this cell has no children otherwise it is the third quadrant child of this cell.2 Box Adaptation & Celltypes The next step in mesh generation is the box adaptation. In this case. When the cell’s parent’s north neighbour has children. Similarly a rectangular box around the collection of the bodies can be formed. otherwise it is the third quadrant child of this cell. the north neighbour of a first quadrant child will be its parent’s north neighbour when this cell is a leaf cell. while the second step is to refine the cells that are in contact with rectangular boxes around individual bodies until the input size criterion for each body is satisfied. This step consists of two stages for input geometries that are comprised of more than one body. in which case there will be no distinct northwest neighbour. this neighbour is the east neighbour of the parent’s north neighbour if this cell has no children. The rectangular box around each one of the bodies spans from the minimum to maximum coordinate of a specific body in both of the Cartesian directions which will be used in the second stage of the box adaptation. 2. The first step is to refine cells that are in contact with an imaginary rectangular box around the collection of the bodies. This box is then stretched by an input proportion. The only remaining neighbour for the first quadrant child is the northeast neighbour.3. the rules for determining each neighbour of any quadrant child can be devised.
In the second stage. The input refinement criteria for each body denote the maximum size of a cell as a proportion of the maximum dimension of that body.8 illustrates how a mesh of different resolutions around the slat and the airfoil is obtained by applying the second stage of box adaptation. this largest criterion is taken to be the proportion of the maximum dimension of the collection of bodies. Figure 2. the cells that are in contact with the boxes around each body is refined until the given refinement criterion for that body is satisfied. Any cell that is in contact with the box is refined until the resulting leaf cells satisfy the largest one of the input refinement criterion for each body.Imaginary boxes around a three element airfoil For the first stage of box adaptation. 26 . The second stage allows the mesh to have differing resolutions around each body.Figure 2. The maximum dimensions are the larger of the differences of the maximum and minimum x or y coordinates of a body or the collection of all bodies. In the first stage of box adaptation. the outer box depicted in Figure 2.7 is used.7.
the next step in mesh generation is curvature adaptation. information regarding the edges cut by the geometry and the portions of the edges through which fluxes pass must be stored. the cells that are cut by the geometry lines must be determined.8. an outside or an inside cell by means of modifying the body. To determine the cell types. However a cell cut by more than one line must be converted to either a regular cutcell. But before the curvature adaptation. All the cells in the mesh will be marked as one of the types outside. The cells that are in contact with this box have the possibility of being inside or outside the bodies or they might be cut by one of the bodies.Figure 2. inside or cutcell.Mesh around the slat and the airfoil After the box adaptation. Through line clipping. A cell that is cut by only one geometry line is a regular cutcell and it can be handled easily. first the relative location of each cell with respect to the large box around the collection of the bodies used for box adaptation is employed. 27 . For the cutcells. Line clipping discussed earlier is performed to determine the exact configuration. the line portions that cut the cell will be determined if the cell is cut.
corner 1 north edge corner 0 west edge east edge corner 2 south edge corner 3 Figure 2.9. the ones with the geometry line lying on one of the edges of the cell.Naming convention of a cell The first step is to detect a special subgroup of regular cutcells. Before going any further. the naming convention of the edges and the corners can be defined as in Figure 2. In other words through which portion of the edge the cell exchanges information with another neighbour cell.1 Regular Cutcells Regular cutcells are the ones that are cut by only one geometry line portion. A similar check is also performed on the y coordinates of the cutting 28 . If the x coordinate of one of the cutting end points is equal to the x coordinate of corner 2 then it means that the line lays on the west face.9. the coordinates of the end points of the cutting geometry line will be known after the line clipping process. Hence if the x coordinate of one of the cutting end points is equal to the x coordinate of corner 0. then it means that the line lays on the east face.2.2. The only remaining thing is to determine on which edges these points lay and through which portion of each edge the flux passes through. in which case the edge cut is stored as east and no information is stored for the flux. For these cells. in which case the edge cut is stored as west and no information is stored for the flux once again. If the x coordinates of both of the cutting end points are equal then there is a chance that the line comprising of these points lays on one of the east or the west edges.3.
a point along the east edge whose location relative to the cutting end point is known must be tested to see if it is inside the body or outside it. The insideoutside test employed is based on the simple point location test described in (15). then the edge cut is stored as north or south respectively and no flux information is stored. corner 3 and any point in between along the east edge a candidate for the location of the cutting end point. If a point is detected to be on the north face the possible locations for it are either corner 1 or anywhere along the north edge between corner 0 and corner 1. If the east edge is not cut then the centre of the edge is put to the insideoutside test. If it is outside. indicating that the flux passes through the whole face and otherwise it is stored as none. If corner 0 is outside. To determine the flux information. The two possible points are corner 0 and corner 3. 29 . If the cutting end point is on corner 0 then corner 3 is tested and if it is outside smaller is stored if it is inside larger is stored.edge. For the insideoutside test corner 0 is used since it is already known that the cutting edge cannot be on this point. Otherwise smaller is stored for the east face flux indicating that the flux passes through the portion of the east edge with y coordinates less than the one of the cutting end point. If the cutting end point is not on corner 0 this point is tested. and if this point is outside any of the bodies then the flux information is stored as all. making corner 0. the flux for the east edge is marked as larger indicating that the flux passes through the portion of the east edge with y coordinates greater than the one of the cutting end point. it is checked whether the x coordinates of the end points of the line are equal to the x coordinate of corner 0. Once the configuration of the east face is determined it is checked whether any of the cutting end points lay on the north face. Once the subgroup of special cells is detected the information for the other regular cutcells can be determined by edge based tests. indicating that no flux passes through the face. If one of them is equal then the east edge is marked as being cut. if they are equal to each other and to one of the y coordinates of corner 0 or corner 2. Starting from the east edge and going counterclockwise. it is checked whether any of the edges is cut by one of the end points of the geometry line or not. For the east face. Thus for the special subgroup of regular cutcells only one edge is marked as cut while no information on the flux is stored.
corner 1 is sent to the insideoutside test as the only point along west edge on which the cutting end point may not lay. Case 1 is an example of the special subgroup of regular cells. Finally.10. After the north face the west face is examined. if the south face is cut either one of the corner 2 and corner 3 can be tested to fill the flux information accordingly. If the edge is not cut. If this edge is cut.10. The flux information is stored as larger if this point is outside otherwise it is stored as smaller. Case 1 Case 2 Case 3 Figure 2.10. In Case 2 the cut edges are detected as the east and west edges. In this particular example the only cut edge is stored as the south edge and no flux information is stored. exemplifying cases for regular cutcells are illustrated. the centre of the edge is tested to see if the flux passes through the face. If the north edge is not cut then the centre of the edge is tested to see if the flux passes through the whole face or it doesn’t pass at all. meaning that it passes through the portion of the edge with x coordinates greater than the cutting edge otherwise smaller is stored indicating that flux passes through the portion with x coordinates less than that of the cutting edge.Examples of regular cutcells In Figure 2. the centre of the edge is tested to see if the flux passes through the face.then the flux is stored as larger. the flux information is stored as larger even though no flux passes through the face. If the edge is not cut. the scribbled parts represent the portion of the cell that is occupied by a body. In Figure 2. The north flux information is all while for the west face it is larger 30 . For the east face.
the ones that have their distinct end points on the same edges. The first 31 . The cells that are cut by more than two geometry lines and those that are cut by two lines that do not have a common end point are refined until their children are cells of acceptable types or until they are refined to the maximum allowable level in the mesh. then they are simply converted to inside cells as in Case 2 of Figure 2. The flux information for the east and the south faces are all. At the end of this refinement process. The cells that are cut by two lines that do not have a common end point will eventually become regular cutcells if they are refined unconditionally. the edges cut are the north and west edges. cells that are cut by two lines that have a common end point are not refined unconditionally until they reach the maximum allowable level. In Case 3. 2. refinement is the first measure taken. To convert these cells into regular cutcells. only when their distinct end points lay on the same face of the cell as in Case 2 and Case 5 of Figure 2. since the maximum level a cell can reach will make it much smaller than the dimensions of the body. The reason for this is that since the bodies are represented as lines with specified end points.11. But sometimes they will require very extensive refinement reaching levels much higher than the maximum allowable one especially in the presence of very acute angles. are tested to see if they are concave or convex. It can safely be said that there will be no cells cut by more than two geometry lines as a result of this refinement process.2. there would be a clustering of highly refined cells around each of the specified points if these cells were to be refined unconditionally. If they are convex. Instead they are refined.11.and for the south face it is none. the cells that were not converted are classified into two groups. it is larger for the north face and smaller for the west face.3. These cases along with the cells that are cut by two lines that have a common end point are treated separately. In contrast to the other types.2 Irregular Cutcells The cutcells that are cut by more than one geometry line are termed as irregular cutcells. Those cells that are refined. convexunited and concavesplit and treated accordingly.
and the cells that are cut by two lines with a common end point with the distinct end points on separate edges forming a single convex cell area (Case 4 of Figure 2.11).Grouping and treatment of irregular cutcells 2.11).Treat as convexunited without refinement Figure 2.2.1 ConcaveSplit The concavesplit cells are converted to regular cutcells by making modifications on the body.2.11.Refine till maxlevel then convert to inside Case 3.Refine till maxlevel then treat as convexunited Case 4. The second group of cells are the ones that are cut by two lines without a common end point with two separate convex cell areas (Case 5 of Figure 2.Treat as convexunited without refinement Case 5. The basic idea is to trim the body towards the direction in which the cutting lines 32 .11).11). cells that are cut by two lines with a common end point forming a single concave area (Case 6 of Figure 2.11. The location of the end points on the edges and the direction towards which the lines converge is of essential importance in the conversion process.Refine till maxlevel then treat as concavesplit Case 6. The grouping and the treatment of the irregular cutcells can be depicted as in Figure 2. Case 1.Refine unconditionally till converted Case 2.group is formed of cells that are cut by two lines without a common end point with a single convex area (Case 3 of Figure 2.3.
Hence the geometry line cutting the cell will be the connection of the end points in the diverging direction. x0.y2 ) + y2 ⎥ ⎣ ( x2 .y2 x2.y2 ) + y2 ⎥ > y1 .y0 x2.12.converge.Conversion of an irregular cutcell The criterion for the above conversion is expressed mathematically in the following equation. The special subgroup will be obtained for cells that are cut by lines with a common end point having their distinct end points on the same edge. The direction towards which the lines converge can be determined by placing one coordinate of the end points of a line to the equation of the other line to obtain the corresponding other coordinate.x ) ⎤ y0 .⎢ 2 0 ( y3 .y3 x0.x3 ) ⎦ ⎣ ( x2 .13.⎢ 2 1 ( y3 . ⎡ (x . while regular cutcells are obtained in some cases.y0 x1.x ) ⎤ ⎡ (x . This can be represented more clearly graphically as in Figure 2. 33 .x3 ) ⎦ (2.y1 x3. Then the line converges towards the direction in which the difference between the calculated coordinate and the coordinate of the end point is smaller.12.20) This trimming process will result in obtaining a member of the special subgroup of regular cutcells in some cases.y2 Figure 2. These cases and their conversion are illustrated in Figure 2. cells having two pairs of end points on the same edge and cells that have one pair of end points on the same edge while converging towards the end points that are on different edges.
The cells that convert into regular cutcells as in Figure 2. In this case since the lines diverge towards the north face.13. Since the lines diverge towards this face. there are two pairs of end points that are on the same edge. the west face is marked as being cut.Concavesplit to special subgroup of regular cutcells conversion In Case 1 of Figure 2. In Case 2.14 are the ones that have common cutting end points with the other end points on different edges. and the cells that have one pair of end points on different edges converging towards a pair of end points that are on the same edge. this face will be marked as being cut. since there is a common end point. there is one pair of end points on the same face. the remaining end points will be connected to obtain a cell that is cut on its east edge. 34 .Case 1 Case 2 Case 3 Figure 2. In Case 3.13. cells with all cutting end points on different edges.
35 . there is a common end point and the other end points are connected automatically to form a regular cutcell. and the cells that have a pair of end points on the same edge with the lines converging towards the end points that are on different edges are converted to inside cells.3. In Case 3.14.15. since the lines diverge towards the end points that are on different edges in contrast to Case 3 of Figure 2. 2.Case 1 Case 2 Case 3 Figure 2.13.2. These cases are illustrated in Figure 2. a regular cutcell is formed by the connection of the cut location on these edges. The cells that have two pairs of cutting end points on the same faces.2.14. Since the distance between the ones on the west and the south edges is greater than the distance between the others the cut locations on these edges are connected to form a regular cutcell. all the end points are on distinct edges. Once again the end points towards which the cutting lines diverge are connected to form the new cutting edge. In Case 2.2 ConvexUnited The procedure followed for the convexunited cells are similar to the ones for concavesplit cells. In the case of the convexunited cells conversion is made to either an inside cell or to a regular cutcell.Concavesplit to regular cutcells conversion In Case 1 of Figure 2.
16. The cells that are converted to regular cutcells are the ones that have common cutting end points with the other end points on different edges. and the cells that have one pair of end points on different edges converging towards a pair of end points that are on the same edge. Case 1 Case 2 Case 3 Figure 2.13. With the difference that they convert to inside cells.Convexunited to regular cutcells conversion 36 .15. cells with all cutting end points on different edges.15 are analogous to those presented in Figure 2.16. and there is no special subgroup of inside cells.Case 1 Case 2 Figure 2.Convexunited to inside cells conversion Case 1 and Case 2 of Figure 2. These cases are illustrated in Figure 2.
the mesh has to be traversed to find these unconnected cells and convert them into either outside cells. If this quantity exceeds a threshold value then the cell is refined. The purpose of the curvature adaptation is to ensure that regions of a body that have high curvatures are resolved enough to be represented accurately and also since these regions will be associated with high gradients. the angle between the normals of the cutting edges pointing towards the fluid domain is used. In curvature adaptation. at the end of the conversion process. if they were processed as concavesplit cells.17.14. The difference between them is the portions of the cells that are marked as inside and outside are opposite in these cases. So firstly neighbouring cutcells should be detected then curvature must be quantified to develop a criterion for refinement. Hence. In order to quantify the curvature. the cutting edges of neighbouring cutcells are evaluated to decide on whether to refine a cell or not. Since each irregular cell is converted individually into a cutcell.3 Curvature Adaptation The last step in the mesh generation process is the curvature adaptation.The Case 1. Case 2 and Case 3 of Figure 2.16 are analogous to those in Figure 2. This can be illustrated as in Figure 2. 2.3. or into inside cells if they were treated as convexunited cells. at the end of the process some cutcells that are not part of any body will be obtained. 37 .
y1 )i + ( x0 .y0 x1. n = ∆yi + ∆xj = ( y0 .Calculation of the normal In Case 1 of Figure 2.y1 ∆x Case 1 ∆y ∆y x1.21) 38 .18. the first edge that is cut is the east edge and the flux information associated is larger.The normals of two neighbouring cutcells The normal for a cutting edge of a cell can be determined through the flux information stored for a cut edge.θ: The angle between normals Figure 2.y0 Case 2 n Figure 2.y1 ∆x x0. the normal is given by the following equation. In this case.x1 ) j (2.17.18. n x0.
Then.23) Then a cell is to be refined.In Case 2.x1 ) j (2. if this θ value is greater than a threshold value.24) 39 . it is possible to express each normal pointing to the fluid domain using the first edge cut and the flux information associated with it for the neighbouring cells.( x0 . the first edge cut is east as in the previous case but this time the associated flux information is smaller making the normal.y1 )i . cos θ = (n1x n2 x + n1 y n2 y )  n1  n2  < cos θth (2.22) Hence. n = ∆yi .∆xj = ( y0 . a relation for the angle between the normals can be obtained by their dot multiplication as n1 in2 = n1x n2 x + n1 y n2 y = n1  n2  cos θ ⇒ cos θ = (n1x n2 x + n1 y n2 y )  n1  n2  (2.
v.3) In this equation θ is the angle between the normal direction of the edge and the Cartesian xdirection. The vector of conserved variables and the flux vector can be written as ⎡ ρ ⎤ ⎢ ρu ⎥ U=⎢ ⎥ ⎢ ρv ⎥ ⎢ ⎥ ⎢ ρ et ⎥ ⎣ ⎦ ρ u cos θ + ρ v sin θ ⎡ ⎤ ⎢ ρ u 2 cos θ + ρ uv sin θ + p cos θ ⎥ ⎥ F in = ⎢ 2 ⎢ ρ v sin θ + ρ uv cos θ + p sin θ ⎥ ⎢ ⎥ ρ uht cos θ + ρ vht sin θ ⎢ ⎥ ⎣ ⎦ (3. solution to the compressible Euler equations in the conserved form is sought. et and ht are the density. This equation can be descritised in 2D as ∂U 1 + ∂t Acell edges ∑ F i n ∆s = 0 (3. The variables ρ . ∆ s the edge length and Acell the cell area. while n is the normal to the differential area element dA . Hence.1) where. u. p. the governing equations can be represented as ∂ UdV + ∂t ∫∫∫ ∫∫ FindA = 0 (3.2) n being the face normal. U and F are the vectors of conserved variables and the flux respectively.CHAPTER 3 NUMERICAL METHOD 3. velocity in 40 .1 Governing Equations In the code developed.
the conserved quantities and the flux vector in terms of normal and tangential velocities can be given ⎡ ρ ⎤ ⎢ ˆ⎥ ˆ = ⎢ ρu ⎥ U ⎢ ρv ⎥ ˆ ⎢ ⎥ ⎢ ρ et ⎥ ⎣ ⎦ ˆ ⎡ ρu ⎤ ⎢ ρu 2 + p⎥ ˆ ˆ ˆ ⎥ F(U) = ⎢ ⎢ ρ uv ⎥ ˆˆ ⎢ ⎥ ⎢ ρ uht ⎥ ⎣ ˆ ⎦ (3.6) u and v are the normal and tangential velocities. At this point. the rotational invariance of the Euler equations (16) can be employed as ˆ ˆ F i n = T−1F (TU) = T −1F ( U) (3. In the solution method used. the total energy and the total enthalpy respectively.5) Hence.7) 41 . determination of the fluxes through the faces of the cells require a pseudo one dimensional form of the flux vector represented in Equation (3.4) can be written explicitly as 0 ⎡1 ⎢0 cos θ T=⎢ ⎢0 − sin θ ⎢ 0 ⎣0 0 sin θ cos θ 0 0⎤ 0⎥ ⎥ 0⎥ ⎥ 1⎦ 0 ⎡1 ⎢ 0 cos θ T−1 = ⎢ ⎢ 0 sin θ ⎢ 0 ⎣0 0 − sin θ cos θ 0 0⎤ 0⎥ ⎥ 0⎥ ⎥ 1⎦ (3. T and T 1 .4) The transformation matrix and its inverse.6) ˆ ˆ ˆ ˆ In Equation (3. This form will be equivalent to the Euler equations written in terms of the normal and tangential velocities. The normal and the tangential velocities can be expressed in terms of the Cartesian velocities as ˆ u = u cos θ + v sin θ .the x and y direction. appearing in Equation (3. ˆ v = −u sin θ + v cos θ (3. in terms of the normal and tangential velocities.3). the pressure. while U and F are the vectors of conserved variables and the flux.
Using this transformation, the governing equations can be expressed as
∂U 1 + ∂t Acell
edges
∑T
−1
ˆ ˆ F(U)∆s = 0
(3.8)
The general solution strategy is to estimate the flux in terms of the normal and tangential directions, rotate them to Cartesian coordinates with the inverse transformation matrix for each face, and then use an appropriate method for time integration.
3.2 One Dimensional Riemann Problem for Linear Systems
Before discussing the solution algorithm for the Euler equations, the one dimensional Riemann problem for a linear system of equations should be discussed (17) (16). The model equation for the one dimensional linear system is represented by Equation (3.9). This equation is linear if the Jacobian of the flux vector is comprised of constant elements, that is
∂U ∂F ∂U ∂F ∂U ∂U ∂U + = + = +A =0 ∂t ∂x ∂t ∂U ∂x ∂t ∂x
(3.9)
where A is the Jacobian matrix. The initial conditions for the model equation are given as
⎧U U( x, 0) = ⎨ L ⎩U R
x0 < 0 x0 > 0
(3.10)
where subscripts L and R refer to the left and right states respectively. The first step in the solution process is to diagonalize the Jacobian matrix using the Q matrix appearing in the following equation. Each column of this matrix is a right eigenvector of the Jacobian matrix and the rows of its inverse are the left eigenvectors of the Jacobian matrix (18). The elements of the resulting diagonal matrix Λ are the eigenvalues of the Jacobian matrix. 42
Q−1 AQ = Λ
(3.11)
Using this diagonalization and multiplying Equation (3.9) by Q1, the linear set of equations can be converted to a set of independent equations in terms of the new characteristic variable vector V .
∂U ∂U ∂V ∂V + Q 1 AQQ 1 = +Λ =0 ∂t ∂x ∂t ∂x
Q 1
(3.12)
The relation between the characteristic variables and the conserved quantities in Equation (3.12) can be defined as
Q1∂U = ∂V
(3.13)
Since this is a linear system and the elements of Q1 are constant, Equation (3.13) can be written in an alternative form as
Q1U = V ⇒ Q1∆U = ∆V
where;
(3.14)
∆V = VR − VL
∆U = UR − UL
After the diagonalization, the governing equations obtained in terms of the characteristic variables, λ ’s, in Equation (3.12) can be written explicitly as
⎡ v1 ⎤ ⎡λ1 0 . ⎢ . ⎥ ⎢0 . ⎢ ⎥ ⎢ ⎢ . ⎥ +⎢ . . ⎢ ⎥ ⎢ ⎢.⎥ ⎢. ⎢vn ⎥ ⎢ 0 . . ⎣ ⎦t ⎣
0 ⎤ ⎡ v1 ⎤ . ⎥⎢ . ⎥ ⎥⎢ ⎥ . ⎥⎢ . ⎥ =0 ⎥⎢ ⎥ . 0 ⎥⎢ . ⎥ 0 λn ⎥ ⎢vn ⎥ x ⎦⎣ ⎦ .
(3.15)
43
The equations represented by Equation (3.15) are n independent equations for the n characteristic variables. In other words, n independent scalar equations are obtained for n characteristic variables, v 's , (17) as
∂vi ∂v + λi i ∂t ∂x
i = 1,..., n
(3.16)
For the solution of each one of the equations in Equation (3.16), the definition of the total derivative is employed as
dvi ∂vi ∂vi dx = + dt ∂t ∂x dt
(3.17)
Hence with the application of the condition presented in Equation (3.18), a straight line along which the characteristic variable remains constant and the governing equation is satisfied, is obtained.
dv dx = λi ⇒ i = 0 along line x = x0 + λi t dt dt
(3.18)
The solution obtained in terms of the characteristic lines can be represented graphically as in Figure 3.1.
x=λ1t
x=λnt
Figure 3.1 Solution of a 1D Riemann problem for a set of linear equations
44
.In Figure 3... < λn < x / t The difference vectors appearing in Equation (3. ⎝t⎠ ⎪ ⎪ . − ∆vn ⎪V + ∆v = V − ∆v − . ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ...14). .. . (3.19) λ1 < .. < λn . Hence the solution can be presented in several formats (17). This will promote the relation presented in the following equation Q∆vi = ri ∆vi where ri is the ith eigenvector of the Jacobian matrix. ⎥ ∆v1 = ⎢ ⎥ ∆v i = ⎢ ⎥ ∆v n = ⎢ ⎥ ⇒ ∆v1 + .. ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ 0 ⎥ ⎢ 0 ⎥ ⎢ ∆vn ⎥ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ (3.. ⎥ ⎢ .. < λn λ1 < x / t < λ2 < . (3.21) 45 . ⎪ ⎪ VL + ∆v1 + . ⎥ ⎢ . ⎥ ⎢ 0 ⎥ ⎢ .. The form in which the solution is presented is of crucial importance in the development of numerical methods while it doesn’t make a difference in the analytical sense. + ∆vn = ∆V ⎢ .. + ∆v n = VR ⎩ x / t < λ1 < . when one of the characteristic lines is crossed the respective value of the characteristic variable changes conforming to the given initial conditions. First representation is in terms of the characteristic variables as ⎧ VL = VR − ∆v1 − . ⎪ ⎛x⎞ ⎪ V⎜ ⎟ = ⎨ . Here it should be remembered that each column of matrix Q is an eigenvector of the Jacobian matrix.1.20) It is possible to express the solution in terms of the conserved variables as well by employing the relation between the characteristic variables and the conserved variables given in Equation (3.19) can be written explicitly as ⎡ 0 ⎤ ⎡ ∆v1 ⎤ ⎡ 0 ⎤ ⎢ . ⎥ ⎢ . ⎥ ⎢ .. − ∆v 1 R 2 n ⎪ L .. ⎥ ⎢ ∆vi ⎥ ⎢ 0 ⎥ ⎢ .
. AU L = AU R − r1λ1∆v1 − . + rn ∆vn = U R ⎪ ⎩ x / t < λ1 < . It should also be noted that the flux vector appearing in the differential model equation is equivalent to the flux term that would appear in its integral formulation.. − r λ ∆v L 1 1 1 R 2 2 2 n n n ⎪ ⎪ . . < λn < 0 (3... ⎝t⎠ ⎪ ⎪ .. − r ∆v R 2 2 n n ⎪ L 1 1 ⎪ .23) is a perfectly valid representation of the solution. ⎪ AU L + r1λ1∆v1 + .. it should be noted that for the numerical method that is used for the solution of the Euler equations.. Also the fluxes of each face will be investigated separately so the / can be considered to be zero for the respective face of the cell..Hence by multiplying Equation (3... ⎛ x⎞ ⎪ U⎜ ⎟ = ⎨ . Even though Equation (3.. λ1 < .19) by the matrix Q.... ⎪ U L + r1∆v1 + .. < λn λ1 < 0 < λ2 < .. − rn λn ∆vn ⎧ ⎪ AU + r λ ∆v = AU − r λ ∆v − . it is desirable to present the solution in a more compact form as in one of the following equivalent forms: 46 . .. ⎪ F ( 0) = ⎨ ... the solution in terms of the conserved variables is obtained as U L = U R − r1∆v1 − . < λn < x / t At this point. Taking these points into consideration one can derive expressions for the fluxes as a solution. < λn .23) has been obtained by exploiting the fact that Ari = λiri .22) λ1 < . < λn . ⎪ ⎪ . + rn λn ∆vn = AU R ⎪ ⎩ 0 < λ1 < . the flux at a face of a certain cell will be of interest rather than the value of the conserved variable.. .. (3. < λn λ1 < x / t < λ2 < ..23) Equation (3. ... − rn ∆vn ⎧ ⎪U + r ∆v = U − r ∆v − .
1) can be estimated by considering the xsplit form of the two dimensional Euler equations in the differential form as ∂U ∂F ∂G + + =0 ∂t ∂x ∂y where.24) where. λi ) λi+ = max(0.25) 3. λi ) Since the representation is dependent on the sign of the eigenvalues in Equation (3.24).26) 47 . λi− = min(0.3 The Approximate Riemann Solver of Roe The flux term in the governing Euler equations in the integral form (3. a final form for the solution can be obtained by averaging the equivalent solutions appearing in Equation (3.24). 1 1 n A ( U L + U R ) − ∑ ri λi ∆vi 2 2 i =1 1 [F ( U L ) + F ( U R ) ] 2 F (0) = (3. ⎡ u1 ⎤ ⎡ ρ ⎤ ⎡ f1 ⎤ ⎡ ρ u ⎤ ⎡ g1 ⎤ ⎡ ρ v ⎤ ⎢u ⎥ ⎢ ρ u ⎥ ⎢ f ⎥ ⎢ ρu 2 + p ⎥ ⎢ ⎥ ⎢ ⎥ ⎥ G = ⎢ g 2 ⎥ = ⎢ ρ uv ⎥ ⎢ 2⎥ = ⎢ ⎥ F = ⎢ 2⎥ = ⎢ U= ⎢ u3 ⎥ ⎢ ρ v ⎥ ⎢ f3 ⎥ ⎢ ρ uv ⎥ ⎢ g3 ⎥ ⎢ ρ v 2 + p ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢u4 ⎥ ⎢ ρ et ⎥ ⎢ f 4 ⎥ ⎢ ρ uht ⎥ ⎢ g 4 ⎥ ⎢ ρ vht ⎥ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ⎣ ⎣ ⎦ ⎣ ⎦ ⎦ (3.F(0) = AUL + ∑ ri λi− ∆vi i =1 n n F(0) = AUR − ∑ r λ ∆vi i =1 + i i (3.
Noting that the Jacobian matrix is specified in the following form ⎡ ∂f1 / ∂u1 ∂f1 / ∂u2 ⎢ ∂f / ∂u ∂f / ∂u ∂F 1 2 2 =A=⎢ 2 ⎢ ∂f 3 / ∂u1 ∂f 3 / ∂u2 ∂U ⎢ ⎢ ∂f 4 / ∂u1 ∂f 4 / ∂u2 ⎣ ∂f1 / ∂u3 ∂f 2 / ∂u3 ∂f 3 / ∂u3 ∂f 4 / ∂u3 ∂f1 / ∂u4 ⎤ ∂f 2 / ∂u4 ⎥ ⎥ ∂f 3 / ∂u4 ⎥ ⎥ ∂f 4 / ∂u4 ⎥ ⎦ (3.28) in terms of the flow variables.28) it can be obtained by equations (3.29) a is the speed of sound.As can be observed the flux vector in Equation (3. 0 ⎡ ⎢ γ −1 h − u2 − a2 ( ) t A=⎢ ⎢ −uv ⎢ 2 ⎢ u ⎡( γ − 2 ) ht − a ⎤ ⎦ ⎣ ⎣ 1 (3 − γ ) u v ht − ( γ − 1) u 2 0 ⎤ γ − 1⎥ ⎥ 0 ⎥ ⎥ − ( γ − 1) uv γ u ⎥ ⎦ 0 (1 − γ ) v u (3.29). the flux vector must be written in terms of the conserved variables as u2 ⎡ ⎤ ⎢ ⎥ 2 2 2 ⎢ u2 + ( γ − 1) ⎛ u − u2 − u3 ⎞ ⎥ ⎜ 4 ⎟ ⎢ u1 2u1 2u1 ⎠ ⎥ ⎝ ⎢ ⎥ F=⎢ u2u3 ⎥ ⎢ ⎥ u1 ⎢ ⎥ 2 3 ⎢ u 2u4 ⎛ u2u4 u2 u3 u2 ⎞ ⎥ + ( γ − 1) ⎜ − 2 − 2 ⎟⎥ ⎢ 2u1 2u1 ⎠ ⎥ ⎢ u1 ⎝ u1 ⎣ ⎦ (3.27) γ is the specific heat ratio. As can be observed from Equation (3.26) is identical to the one appearing in the transformed form of the governing equations given by Equation (3.27) and (3. To obtain the Jacobian for the xsplit Euler equations. The theory of the Riemann problem for one dimensional linear systems will be employed to estimate this flux.6).27) In Equation (3. the Jacobian matrix does not have constant elements hence the Euler equations are 48 .29) In Equation (3.
not linear. ⎡ ρL + ρR ⎤ ⎢ ⎥ 2 ⎢ ⎥ ⎡ w1 ⎤ ⎢ ρ L u L + ρ R u R ⎥ ⎥ ⎢w ⎥ ⎢ 2 ⎥ ⎢ 2⎥ = ⎢ W= ⎢ w3 ⎥ ⎢ ρ v + ρ v ⎥ L L R R ⎥ ⎢ ⎥ ⎢ ⎢ w4 ⎥ ⎢ 2 ⎣ ⎦ ⎥ ⎢ ρ h + ρ h ⎥ R tR ⎥ ⎢ L tL ⎢ ⎥ 2 ⎣ ⎦ (3. In Roe’s approximate Riemann solver (19). (16). the Jacobian matrix is expressed in terms of an intermediate state defined from the right and left states. For the approximate Riemann solvers this Jacobian matrix is replaced by one that has constant elements since for a nonlinear system the characteristic lines of the same family will not be parallel introducing shock waves and expansion fans into the solution (17). First. is formed by taking the arithmetic mean of the W vector for the right and left states. it is observed that both the flux vector F and the vector of conserved quantities U are quadratic in terms of a new vector W defined as ⎡1⎤ ⎡ w1 ⎤ ⎢w ⎥ ⎢u ⎥ ⎢ 2⎥ = ρ ⎢ ⎥ W= ⎢v⎥ ⎢ w3 ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ w4 ⎥ ⎢ ht ⎥ ⎣ ⎦ ⎣ ⎦ (3.31) It is now possible to express the difference between the right and the left states for the conserved quantities in terms of the new vector defined as ∆U = B ∆W (3.32) where. 49 .30) Another useful vector. W .
⎡ w2 ⎢γ −1 ⎢ w4 C=⎢ γ ⎢ 0 ⎢ ⎢ 0 ⎣ w1 γ +1 γ w3 w4 w2 − 0 γ −1 γ w2 0 w3 0 γ −1 ⎤ ⎥ w1 ⎥ γ ⎥ 0 ⎥ ⎥ w2 ⎥ ⎦ Using Equations (3.⎡ 2w1 ⎢w ⎢ 2 B = ⎢ w3 ⎢ ⎢ w4 ⎢ γ ⎣ 0 w1 0 0 0 w1 γ −1 w2 γ γ −1 w γ 3 0⎤ 0⎥ ⎥ 0⎥ ⎥ w1 ⎥ γ ⎥ ⎦ It is also possible to express the differential flux vector in terms of the vector W in the same manner as ∆F = C∆W (3.34) Using the homogeneity property of the Euler equations given by the following equation (16) one can recognize the product CB1 as the Jacobian matrix.32) and (3.33) where. 50 . the following equation must be solved. ∂F U ∂U A F= (3.33) it is possible to establish a direct relation between the difference in the flux vector and the vector of conserved quantities in the form CB 1 ∆U = ∆F (3.35) In order to find the eigenvalues of the Jacobian matrix.
36) Dividing Equation (3.41) (3.det ( CB1 − λ I ) = 0 ⇒ det ( C − λ B ) = 0 (3.39) a= (γ − 1) ⎛ ht − ⎜ ⎝ 1 2⎞ 1 ⎞ ⎛ V ⎟ = γ ( γ − 1) ⎜ et − V 2 ⎟ 2 ⎠ 2 ⎠ ⎝ (3. it is possible to find the eigenvalues of the Jacobian in terms of the new variables defined as w2 = w1 u= ρ L uL + ρ R uR ρL + ρR ρ L vL + ρ R vR ρL + ρR ρ L htL + ρ R htR ρL + ρR (3.42) The corresponding right eigenvectors can also be expressed in terms of the new Roe averaged quantities as ⎡ 1 ⎤ ⎡ 1 ⎤ ⎡ 1 ⎤ ⎡0⎤ ⎢ u ⎥ ⎢ u−a ⎥ ⎢ u+a ⎥ ⎢0⎥ ⎥ ⎥ r =⎢ ⎢ ⎥ r1 = ⎢ ⎢ v ⎥ r3 = ⎢ ⎥ r4 = ⎢ ⎢ v ⎥ 2 ⎢ v ⎥ ⎢1 ⎥ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢1V 2 ⎥ ⎢ ht − ua ⎥ ⎢ ht + ua ⎥ ⎣v ⎦ ⎣ ⎦ ⎣ ⎦ ⎢2 ⎥ ⎣ ⎦ (3.40) ρ = ρL ρR λ1 = u − a λ2 = λ3 = u λ4 = u + a (3.43) 51 .38) ht = w4 = w1 (3.37) v= w3 = w1 (3.36) by w1 .
49) .44).44) Due to the proportionality between the coefficients of the first and third relations in Equation (3.14).46) It is possible to solve this system using Gauss elimination to obtain the wave strengths as ∆v2 = γ −1 ⎡ ∆u1 ( ht − u 2 ) + u ∆u2 − ∆u4 + v ( ∆u3 − v ∆u1 ) ⎤ 2 ⎣ ⎦ a 1 ⎡( u + a ) ∆u1 − ∆u2 − a∆v2 ⎤ ⎦ 2a ⎣ (3. ⎡ 1 ⎢ ⎢ u−a ⎢ ⎢ v ⎢ h − ua ⎢ t ⎣ 1 u v 1 2 V 2 1 ⎤ ⎡ ∆v1 ⎤ ⎡ ∆u1 ⎤ 0 ⎥ 0 u + a ⎥ ⎢ ∆v2 ⎥ ⎢ ∆u2 ⎥ ⎥=⎢ ⎥ ⎥⎢ v ⎥ ⎢ ∆v3 ⎥ ⎢ ∆u3 ⎥ 1 ⎢ ⎥ ⎢ ⎥ v ht + ua ⎥ ⎢ ∆v4 ⎥ ⎢ ∆u4 ⎥ ⎦ ⎣ ⎦ ⎥⎣ ⎦ (3.45) (3.With the right eigenvectors defined as in Equation (3. it is possible to eliminate the third equation by using the first one to obtain the following relation. ∆v3 = ∆u3 − v∆u1 Then a set of three equations is obtained for three wave strengths as ⎡ ⎢ 1 1 ⎢ u−a ⎢ u+a ⎢ ⎢ ht + ua ht − ua ⎣ ⎤ 1 ⎥ ⎡ ∆v4 ⎤ ⎡ ∆u1 ⎤ ⎥⎢ ⎥=⎢ ⎥ u ⎥ ⎢ ∆v1 ⎥ ⎢ ∆u 2 ⎥ 1 2 ⎥ ⎢ ∆v2 ⎥ ⎢ ∆u4 + v 2 ∆u1 − v ∆u3 ⎥ ⎦ ⎣ ⎦ V ⎥⎣ 2 ⎦ (3.43).47) ∆v1 = (3.48) ∆v4 = ∆u1 − (∆v1 + ∆v2 ) 52 (3. it is possible to determine the characteristic variables which can also be termed as the wave strengths by using Equation (3.
a simplified equation for the wave strength can be obtained.53) To obtain ρ∆et = ρ ⎢ ⎡ 1 ⎤ ∆p ⎛ p⎞ 1 a 2 ∆ρ ∆ ⎜ ⎟ + ∆ ( u 2 + v2 )⎥ = − + ρ u∆u + ρ v∆v (3.52) and (3. Equation (3. Then applying the operator defined in 53 .51).54) ⎣ γ −1 ⎝ ρ ⎠ 2 ⎦ γ − 1 γ ( γ − 1) Substituting Equations (3.51) The terms appearing in Equation (3.54) into Equation (3.51) must be treated individually to simplify the expression.50) neglecting higher order terms and placing the values of the conserved quantities.Defining an operator given by the following equation ∆( pq) = p∆q + q∆p + O ( ∆ 2 ) (3. the wave strengths can be expressed in terms of the flow variables as ∆v2 = γ −1 ⎡ ∆ρ ( ht − u 2 ) + u 2 ∆ρ + ρ u ∆u − ρ∆et − et ∆ρ + v ( ρ∆v + v∆ρ − v∆ρ ) ⎤ 2 a ⎣ ⎦ a γ −1 ⇒ ∆v2 = 2 ⎡ ∆ρ ( ht − et ) + ρ u ∆u + ρ v ∆v − ρ∆et ⎤ ⎣ ⎦ (3.52) The last term in the brackets must also be treated individually with the aid of the relation 1 p V2 + γ −1 ρ 2 et = (3.40) can be employed as ⎛ a2 1 a2 1 ⎞ a2 ∆ρ ht − et = ∆ρ ⎜ + V2 − − V 2 ⎟ = ∆ρ ⎜ γ −1 2 γ ( γ − 1) 2 ⎟ γ ⎝ ⎠ ( ) (3. For the treatment of the first term in the brackets.
(i) First the conserved variables of the left and right states are obtained by reconstruction which will be explained later on.58) In summary.56) ∆v3 = ρ∆v 1 ( ∆p + ρ a ∆u ) 2a 2 (3. 3.1 Solution Algorithm The solution algorithm adopted is similar to the one outlined in (16). Then these are used to find the wave strengths. is employed to update the values of the conserved variables. which are non linear in their true essence.55) ∆v2 = ∆ρ − (3.57) into Equations (3.3. 54 .50) and substituting Equation (3. an appropriate method for time integration. Thereafter the eigenvalues and eigenvectors of this approximate Jacobian matrix are found by following the procedure outlined in (19).25).49) all the wave strengths can be found as 1 ( ∆ p − ρ a ∆u ) 2a 2 ∆p a2 ∆v1 = (3.57) ∆v4 = (3. The algorithm discussed here is for the determination of the flux through a certain face of the cell under investigation.48) and (3. have been replaced by another matrix with constant elements to approximate the governing equations as a linear system.Equation (3. for the approximate Riemann solver of Roe. and fluxes through a face for which the right and left states are known according to Equation (3. characteristic variables. to be discussed later. Once the fluxes through all the faces are determined. first the Jacobian matrix for the Euler equations.
is transformed back to the Cartesian coordinates using the inverse transformation matrix. The differences are taken as the right minus the left state variables in contrast to Roe’s derivation in (19) to be consistent with Equation (3.25). ˆ T 1 F = F (3. The wave strengths are calculated from Equations (3.37) through (3. (v) (vi) (vii) The right eigenvectors are calculated using the transformed Roe averaged quantities with the formulation provided in Equation (3.60) ⎡ ( ρ u ) 2 ( ρ v) 2 ⎤ p = ( γ − 1) ⎢ ρ et − − ρ ρ ⎥ ⎣ ⎦ 1 (3.42). 55 .63) where T1 is given in Equation (3.(ii) The primitive variables. ˆ u= ρ u cos θ + ρ v sin θ ρ (3.25). The flux which will be used in Equation (3. The flux is calculated from Equation (3.5). The flux obtained.59) ˆ v= − ρ u sin θ + ρ v cos θ ρ (3.55) through (3.61) ht = ρ ( ρ et + p ) (3.58).25). [ ρ u v P ht ] for the right and left states are T determined from the conserved variables [ρ ρ u ρ v ρ et ] T simultaneously with the transformation to normal and tangential coordinates. is calculated from the left and right state variables. which is normal to the face.62) (iii) (iv) The Roe averaged quantities for the transformed primitive variables and the eigenvalues are calculated from Equations (3.43).
AUSMV (22) have also been employed.4 Flux Vector Splitting In addition to using Roe’s approximate Riemann solver to evaluate the flux term appearing in Equation (3. .8) the positive flux component will be due to the cell itself while the 56 .67) where the positive and negative eigenvalues can be defined as in (23) presented as in the following equation. .65) (3. 0 (3. ⎡λ1− ⎢ ⎢0 − Λ =⎢ . F − ( U) = A − U A − = QΛ −Q1 (3. the flux vector splitting methods such as van Leer (20). .3. ⎢ ⎢ .8). . 0 . AUSM (21) and its derivatives AUSMD. . can be defined by splitting the Jacobian matrix as follows F + ( U) = A + U A + = QΛ +Q1 . ⎥ −⎥ 0 λm ⎦ . 1 (λ + λ ) 2 1 (λ − λ 2 λ+ = .( U ) (3. F + and F .66) The diagonal matrices Λ + and Λ − can be defined as ⎡λ1+ ⎢ ⎢0 + Λ =⎢ . λ− = ) (3. Considering Equation (3. ⎤ ⎥ ⎥ ⎥ ⎥ .68) When the Euler equations are written in normal and tangential coordinates as in Equation (3. . 0 . . . ⎢0 ⎣ 0 . ⎥ +⎥ 0 λm ⎦ .9) it is possible to split the flux vector as F ( U ) = F + ( U) + F .. . . 0 . ⎢ ⎢ . ⎤ ⎥ ⎥ ⎥ ⎥ . ⎢0 ⎣ 0 . .64) The negative and positive components of the flux vector.
3.Graphical interpretation of flux vector splitting In Figure 3. M= a (3. The positive eigenvalues of the left state affects the face hence the positive component of the flux is calculated from the left state while the negative component of the flux is calculated from the right state since the negative eigenvalues of this state affects the face. Further details can be found in (16). Left Right t n Figure 3.negative component will be calculated from the neighbour as illustrated in Figure 3.4.2. M . Mach number.2.1 Van Leer Flux Vector Splitting The van Leer flux vector splitting algorithm (20) makes use of the flux represented in ˆ terms of the density.2 the left state is the cell under investigation while the right state is the neighbour. tangential velocity. and the speed of sound as ˆ ⎡ ⎤ ρ aM ⎢ ⎥ 1⎞ ⎢ ⎥ 2⎛ ˆ 2 ρa ⎜ M + ⎟ ⎢ ⎥ γ⎠ ⎝ ˆ ˆ ⎥ F(U) = ⎢ ˆˆ ⎢ ⎥ ρ aMv ⎢ ⎥ ⎡1 2 ˆ 2 2 a2 ⎤ ⎥ ⎢ ˆ ˆ ⎢ ρ aM ⎢ 2 a M + v + γ − 1 ⎥ ⎥ ⎢ ⎣ ⎦⎥ ⎣ ⎦ ˆ ˆ u .69) ( ) 57 .
The flux can then be split into its positive and negative components as presented in the following equations.2 AUSM (LiouSteffen) Flux Vector Splitting The AUSM scheme (21) splits the pressure and the advection terms appearing in the normal momentum flux as 1 ˆ ˆ ˆ ˆ F ( U ) = ⎡ M 1/ 2 ( φ L + φ R ) − M 1/ 2 ( φ R − φ L ) ⎤ + p1/ 2 ⎦ 2⎣ (3. 1 ⎡ ⎤ ⎢ ⎥ 2a ⎛ γ − 1 ˆ ⎞ ⎢ ⎥ M + 1⎟ ⎥ 2 ⎢ γ ⎜ 2 ⎝ ⎠ 1 ˆ ˆ F+ = ρ a 1 + M ⎢ ⎥ ˆ v 4 ⎢ ⎥ 2 ⎢ 2a 2 ⎛ γ − 1 ⎞ 1 ⎥ ˆ ⎢ 2 ⎜ ˆ M + 1⎟ + v 2 ⎥ ⎢γ −1 ⎝ 2 ⎠ 2 ⎥ ⎣ ⎦ ( ) (3.72).4. M1/ 2 is the interface Mach number.70) 1 ⎡ ⎤ ⎢ ⎥ 2a ⎛ γ − 1 ˆ ⎞ ⎢ ⎥ M − 1⎟ ⎜ ⎥ 2 ⎢ γ ⎝ 2 ⎠ 1 ˆ ˆ F− = − ρ a 1 − M ⎢ ⎥ ˆ v 4 ⎢ ⎥ 2 ⎢ 2a 2 ⎛ γ − 1 ⎞ 1 ⎥ ˆ ⎢ 2 ⎜ ˆ M − 1⎟ + v 2 ⎥ ⎢γ −1 ⎝ 2 ⎠ 2 ⎥ ⎣ ⎦ ( ) (3. ˆ ˆ+ ˆ− M 1/ 2 = M L + M R (3. p1/ 2 is the vector of interface pressure and φ is the vector of flux defined through the following equations.73) 58 .72) ˆ In equation (3.71) 3.
p1/ 2 ⎡ 0 ⎤ ⎢p ⎥ + − = ⎢ 1/ 2 ⎥ .4. As in the AUSM scheme the advection and pressure terms are split separately with the difference that instead of estimating the normal velocity at the interface the mass flux is estimated.76) ˆ M >1 ˆ ⎧±2 − M ⎪ ˆ p = pM ⋅ ⎨ 1 ⎪ ˆ ⎩ M ± ± ˆ M ≤1 ˆ M >1 (3.3 AUSMD Flux Vector Splitting AUSMD flux vector splitting (22) is a derivative of the original AUSM scheme.77) 3.78) ˆ ˆ+ ˆ− ( ρ u )1/ 2 = uL ρ L + uR ρ R (3. p1/ 2 = pL + pR ⎢ 0 ⎥ ⎢ ⎥ ⎣ 0 ⎦ ⎡ ρa ⎤ ⎢ ρ ua ⎥ ˆ ⎥ φ=⎢ ˆ ⎢ ρ va ⎥ ⎢ ⎥ ⎢ ρ ht a ⎥ ⎣ ⎦ (3. 1 ˆ ˆ ˆ ˆ F ( U ) = ⎡( ρ u )1/ 2 ( Γ L + Γ R ) − ( ρ u )1/ 2 ( Γ R − Γ L ) ⎤ + p1/ 2 ⎦ 2⎣ (3. Also the velocity splitting is modified in the AUSMD method as in the equations to follow.75) The split Mach numbers and the split pressures can be defined as in the following equations to complete the algorithm.79) 59 .74) (3. ⎧ 1 ˆ ⎪± M ± 1 ˆ± =⎪ 4 M ⎨ ⎪1 M ± M ˆ ˆ ⎪2 ⎩ ( ) 2 ˆ M ≤1 ( ) (3.
80) ˆ uL > amax ˆ uR ≤ amax (3. amax = max( aL .81) ˆ uR > amax κL = 2( p / ρ ) L 2( p / ρ ) R . aR ) (3.84) ⎧ ( uL + amax )2 ⎛ ˆ ˆ uL ⎞ ⎪ pL ⎜2− ⎟ 2 amax ⎠ 4amax ⎪ + ⎝ pL = ⎨ ˆ ˆ u + uL ⎪ pL L ⎪ ˆ 2uL ⎩ ⎧ ( uR − amax )2 ⎛ ˆ ˆ uR ⎞ ⎪ pR ⎜2+ ⎟ 2 amax ⎠ 4amax ⎪ − ⎝ pR = ⎨ ˆ ˆ u − uR ⎪ pR R ⎪ ˆ 2uR ⎩ ˆ uL ≤ amax (3.83) (3.82) ( p / ρ )L + ( p / ρ )R ( p / ρ )L + ( p / ρ )R Naturally the vector of flux is modified as well as the split pressure as ⎡1⎤ ⎢u ⎥ ˆ Γ=⎢ ⎥ ⎢v⎥ ˆ ⎢ ⎥ ⎢ ht ⎥ ⎣ ⎦ + − p1/ 2 = pL + pR (3.⎧ ⎡ ( u + a )2 ⎤ ˆ ˆ ˆ uL + uL max ⎪κ L ⎢ L ⎥ + (1 − κ L ) 2 ⎪ ⎣ 4amax ⎥ ⎦ ˆ+ uL = ⎨ ⎢ ⎪ ˆ ˆ uL + uL ⎪ 2 ⎩ ⎧ ⎡ ( u − a )2 ⎤ ˆ ˆ ˆ u − uR ⎪κ R ⎢ − R max ⎥ + (1 − κ R ) R 4amax 2 ⎪ ⎣ ⎥ ⎦ ˆ− uR = ⎨ ⎢ ⎪ ˆ ˆ uR − uR ⎪ 2 ⎩ ˆ uL ≤ amax (3.86) ˆ uR > amax 60 .85) ˆ uL > amax ˆ uR ≤ amax (3. κR = .
78) and (3. the physical boundary condition is that there is no flux into the body. For the solid wall. 61 . The normal momentum is estimated at the interface wholly as ˆ ( ρu ) 2 1/ 2 ˆ+ ˆ ˆ− ˆ = uL ( ρ u ) L + uR ( ρ u ) R (3.3.4. Similar to the AUSMD scheme the mass flux at the interface of two cells are estimated as in Equations (3.89) ˆ u >a 3.5 Boundary Conditions The boundary conditions encountered for the external flows considered in the scope of the thesis work are the solid wall and the farfield boundary conditions.87) The definitions of the split velocities and the pressures differs from the AUSMD scheme as can be observed from the following equations ⎧ ( u ± a )2 ˆ ⎪± ⎪ 4a ˆ u± = ⎨ ˆ ˆ ⎪ u± u ⎪ 2 ⎩ ˆ u ≤a (3.79) in the AUSMV scheme except for the normal momentum. this means that the velocity at the face can be in the tangential direction. the implementation of either type of the boundary conditions is equivalent to determining the correct right state variables since the left state variables will always be dictated by the cell under investigation. For the numerical method adopted.4 AUSMV Flux Vector Splitting Another derivative of the AUSM scheme is the AUSMV flux vector splitting scheme (22). In the inviscid case considered.88) ˆ u >a ˆ u⎞ ⎧1 ⎛ ⎪ a ⎜ ±2 − a ⎟ ⎪ ⎠ ˆ p ± = pu ± ⋅ ⎨ ⎝ 1 ⎪ ⎪ ˆ u ⎩ ˆ u ≤a (3.
3. In the code developed. the right and left states. The second method is to construct a ghost right state for which the tangential velocity. For the normal velocity.Ghost right state at the solid wall The second method ensures zero flux through the face as well but it also takes into cases where there is a normal velocity component in the presence of high curvatures as stated in (14).3. One is to directly take the flux through the face as zero and to calculate a right flux solely based on pressure that is equal to the pressure calculated for the left state. The ghost right state is depicted for an example cutcell in Figure 3. vR vL htL pL uL uR htR= htL pR= pL Figure 3. primitive flow variables have to be estimated on both sides of that face. The simplest approach 62 .At this point. pressure and total enthalpy are equal to the left state. the pressure force exerted by the fluid will be countered with an equal force. the magnitude is equal to that of the left state but the direction is opposite to it. the second method has been adopted for the solid wall treatment. For the farfield boundary conditions the right state is calculated from the conditions received as input and it is irrespective of the left state.6 Reconstruction of the Flow Variables For the calculation of the fluxes through a face. 3. The reason for this equality is that under steadystate conditions. one of two methods can be adopted for the determination of the right state variables (14). The direction of the tangential velocity is the same as the left state as well as the magnitude.
3. One of the reconstruction schemes based on the gradient at the centre of the cell is the GreenGauss reconstruction.90) This integral is evaluated for a region Ω which is bounded by the lines that connect the centroids of the edge and vertex neighbours (support set) of the cell. the other one is the Least Squares (Minimum Energy) reconstruction.would be to take the values from the cell under investigation as the left state and to take the values from the respective neighbour as the right state. The integral can be evaluated numerically using the trapezoidal rule given as 1 ∇u = Acell ∑ 2 (u + u i =0 i Nsup 1 i +1 )ni i (3. Three reconstruction schemes have been sighted in the literature.91) 63 . This approach being first order does not produce very accurate results. the gradient of the flow variable is estimated at the centre of the cell while the other one makes use of finite differences to estimate the value of a flow variable at certain points within the cell. The general design criteria for these reconstruction schemes are discussed in (24) and (25). The details of the linear reconstruction based on these schemes can be found in (26). For a quantity u related to the cell this can be formulized as 1 ∫ und AΩ Ω ∇u = (3.6. Hence it is desirable to reconstruct the flow variables within the cell to obtain a solution of second order. The third scheme based on the finite difference method can be found in (29). In two of these three schemes. (27) and (14) while the kth order exact schemes are discussed in (28). That is the variation of the flow variables within the cell must be taken into account even though in finite volume methods the flow variables representative of the whole cell is stored at the cell centroid for higher order accuracy.1 GaussGreen Reconstruction The GaussGreen reconstruction is based mainly on the gradient theorem.
the constructed polynomial can be obtained as uk = u + m+ n≤ k ∑α mn ⎡ x m y n − Av( x m y n ) ⎤ ⎣ ⎦ (3.6.The boundary of the region Ω is shown for two representative cases for a Cartesian mesh in Figure 3.93) where α mn are arbitrary coefficients.4.4. 64 .2 Least Squares (Minimum Energy) Reconstruction The Least Squares reconstruction is basically the inverse of the averaging function for a quantity related to a cell. In general. The average of a quantity u within a cell can be defined with the averaging function presented as Av(u ) = Acell ∫ udA Acell =u (3. Figure 3.92) A kth order polynomial uk is constructed to represent the quantity u within the cell such that the average of the polynomial will be equal to that of the cell and the polynomial will represent quantity u exactly if u is a polynomial of order k or less.Region formed by the support set 3.
The form given in Equation (3.93) will ensure that the average of the polynomial is equal to that of the cell. For a linear case, this reconstruction takes the form presented in the following equation
u1 ( x, y ) = u + α10 [ x − xc ] + α 01 [ y − yc ]
(3.94)
In Equation (3.94), the centroidal coordinates of the cell appear since by definition the averaging function applied to the polynomials x and y within the cell gives the centroid coordinates of the cell. The reconstruction problem simplifies once the cell average values are thought to be the values of the flow variables stored at the centroid of the cell (24). Then one could get the expression of the following form for the linear reconstruction.
u1 ( x, y ) = uc + u xc [ x − xc ] + u yc [ y − yc ]
(3.95)
In Equation (3.95), the only unknowns are uxc and uyc, which are the components of the gradient of quantity u at the centroid of the cell. Once the linear polynomial is obtained for the quantity u, the sum of the squares of the averages of the difference between the value of u obtained from constructed polynomial and its actual value is minimized within the region made up of the cells that form the support set to find the unknowns uxc and uyc. The support set is defined in the same way as the GreenGauss reconstruction shown in Figure 3.4. This constitutes the least squares problem for the reconstruction.
Nsup i =1
S = ∑ ⎡ Av(u1 − ui ) ⎤ ⎣ ⎦
2
(3.96)
∂S =0 ∂u xc
,
∂S =0 ∂u yc
(3.97)
65
Once the differentiations given in Equation (3.97) are performed, the expressions given in the following equations are obtained.
N sup ⎧ ⎫ ⎡ ∂ ⎣ Av (u1 − ui ) ⎤ ⎪ ∂S ⎪ ⎦ =0 1 = ∑ 2 ⎨ Av (u − ui ) ⋅ ⎬ ∂u xc i =1 ⎪ ∂u xc ⎪ ⎩ ⎭
(3.98)
N sup ⎧ ∂ ⎡ Av (u1 − ui ) ⎤ ⎫ ∂S ⎪ ⎦⎪ = 0 1 = ∑ 2 ⎨ Av (u − ui ) ⋅ ⎣ ⎬ ∂u yc i =1 ⎪ ∂u yc ⎪ ⎩ ⎭
(3.99)
Then the average term over the region constructed by the cells in the support set is evaluated. ⎤ 1 ⎡ ⎢ ∫ uc dAi − ∫ ui dAi + u xc ∫ ( x − xc )dAi + u yc ∫ ( y − yc )dAi ⎥ (3.100) Ai ⎣ i i i i ⎦
Av(u1 − ui ) =
Once the integrals in Equation (3.100) are evaluated and the expressions simplified, it is possible to arrive at the following equation
Av(u1 − ui ) = (uc − uci ) + u xc ( xci − xc ) + u yc ( yci − yc )
(3.101)
When the expression presented in Equation (3.101) for each cell in the support set is placed in the summations given in Equations (3.98) and (3.99), two equations for the two unknowns can be obtained as
sup ∂S = ∑ ⎡(uc − uci ) + u xc ( xci − xc ) + u yc ( yci − yc ) ⎤ ( xci − xc ) = 0 ⎦ ∂u xc i =1 ⎣
N
(3.102)
sup ∂S = ∑ ⎡(uc − uci ) + uxc ( xci − xc ) + u yc ( yci − yc ) ⎤ ( yci − yc ) = 0 ⎦ ∂u yc i =1 ⎣
N
(3.103)
66
Equations (3.102) and (3.103) can be presented in a matrix form as
Nsup ⎡ ⎢ ∑(xci − xc )2 i =1 ⎢ Nsup ⎢ ⎢∑( xci − xc ) ( yci − yc ) ⎢ i=1 ⎣
⎤ ⎡ Nsup ⎤ ∑( xci − xc ) ( yci − yc )⎥ ⎡uxc ⎤ ⎢ ∑( uci − uc )( xci − xc ) ⎥ i =1 ⎥ ⎢ i=1 ⎥ (3.104) ⎢u ⎥ = ⎢ Nsup Nsup ⎥ ⎣ yc ⎦ ⎥ ⎥ ⎢∑( uci − uc )( yci − yc ) ⎥ ( yci − yc )2 ∑ ⎥ ⎢ i=1 ⎥ i =1 ⎦ ⎣ ⎦
Nsup
This matrix can be solved to obtain the gradients at the centre of the cell from which the value of quantity u can be obtained at any location within the cell. The least squares method is employed for the reconstruction in the code developed.
3.6.3 Gradient Limiting
It is undesirable to introduce new maxima points within the region formed by the support cells and the cell for which the gradient of flow variables are being constructed. Hence, the gradients are multiplied by a limiter to avoid introduction of new maxima points. The linear reconstruction takes the following form with the application of this condition.
u1 = uc + Φ ⎡uxc ( x − xc ) + u yc ( y − yc ) ⎤ ⎣ ⎦
(3.105)
The value of the limiter Φ is between 0 and 1. To determine the exact value of the limiter, the maximum and the minimum of the values of quantity u are found at the cell centroids. u max = max(uc , uic ) u min = min(uc , uic ) i = 1,..., N sup
(3.106)
Then the values of the reconstructed polynomial are evaluated at the corners of the cell since these locations will have the maximum and minimum values within the cell for a linear reconstruction. The value of quantity u which deviates the most from the
1 centroidal value will be marked as ucor serving as a basis for determining the value
67
Hence. A detailed discussion of the multistage methods can be found in (30).7 Temporal Discretization Once the flux terms in the governing Equation (3. and its value can be determined from the known values of the conserved quantities at a given stage.108) can be treated as an ordinary differential equation.8) is obtained.108).7. then the temporal discretization has to be performed to update the values of the conserved quantities.of the limiter with the implementation of the formulation given in the following equation ⎧ ⎛ u max − uc ⎞ 1 ⎪min ⎜1. R . ⎟ ucor − uc < 0 1 ⎝ uc − ucor ⎠ ⎪ 1 ⎪ 1 ucor − uc = 0 ⎪ ⎪ ⎩ (3. 3. the flux terms are presented in a more compact form as residual vector. 68 .107) 3. 1 ⎟ ucor − uc > 0 ⎝ ucor − uc ⎠ ⎪ ⎪ ⎛ uc − u min ⎞ 1 ⎪ Φ = ⎨ min ⎜1. Equation (3.108) In Equation (3. This equation can be discretized in mstages as described in (31) to obtain the following equation. an alternative representation of the governing Equation (3.8) is employed as ∂U 1 + R=0 ∂t Acell (3.1 Explicit Time Stepping Schemes To apply the multistage time stepping method. A multistage time stepping method is employed for temporal discretization.
1263 0.1084 0.0000 stages σ β1 β2 β3 β4 β5 Table 3.0000 0.0533 0.15 0.4000 1. 69 .2375 0. The CFL number is a factor that governs how large the time step can be with respect to the cell area.109). three and four stage second order schemes are used.5060 1. In contrast to classical RungeKutta methods only the values of the zeroth solution and the last residual have to be stored in this formulation (31).2898 0.0000 0.69 0. Hence.0833 0.Stage coefficients and CFL numbers for the multistage method firstorder scheme 3 4 5 1.4265 1. Table 3. this is an explicit formulation.0 2.2602 0. It is possible to tune the value of β for stability and to obtain larger time steps.5052 1.0000 secondorder scheme 3 4 5 0.0695 0.4414 1.109) U n+1 = U m = U 0 − β m In Equation (3.0000 0. The residuals are evaluated from the values of the conserved variables from the previous stage. four and five stage methods of first and second order accuracy.5 0. β ’s are stage coefficients while ∆t is the time step.1.5 2.U0 = Un U1 = U 0 − β1 ∆t 0 R Acell ∆t m1 R Acell (3.2069 0.1 taken from (31) shows the optimised values of the stage coefficients and CFL numbers for three.4929 1.1481 0.1918 0.1602 0.92 1.0000 0. In the code developed.
110) where σ is the CFL number and Ψ is the spectral radius.111) Ψy = (3. As a result. the value of the time step is not determined. The characteristic speed is the maximum speed with which information is carried for a cell. the time 70 . If this value is taken to be too small then the convergence will be very slow and unnecessary amount of computations will be performed. The time step is calculated in terms of the area of the cell under investigation as in the following formulation given in (31). The mesh generated for an arbitrary geometry will contain cells that have different sizes and since the time step is dependent on the cell size. The spectral radii Ψ that appear in Equation (3. ∆t σ = x Acell Ψ + Ψ y (3.2 Determination of the Time Step In Equation (3.109).110) can be calculated using the length of the edge projections of the cell in both xy directions. S x and S y . if taken to be too large.3. The value of the time step. every cell will have a different maximum time step. 1 ( u + a) ∑ S x 2 edges 1 ( v + a) ∑ S y 2 edges Ψx = (3.7. this equation is in one sense a division of a characteristic length by a characteristic speed. will introduce stability problems. a different time step will be used for each cell since a steady state solution is sought.110) is investigated.112) The time step obtained thus is multiplied by a safety factor in some cases to promote stable solutions. If Equation (3. Hence.
114) Refine cell j ⎛ ⎞ ⎜ ∑ζ i ⎟ ⎝ i =1 ⎠ ncell 71 .113) are calculated and the cells exceeding the standard deviation are refined. This method is based on both the curl of the velocity. The curl and gradient of the velocity is multiplied by some characteristic length of the cell as τ = Acell 2 ( ∇ ⋅ V ) . for detecting shear layers. and the gradient of the velocity.step is obtained by constricting the distance information travels from a cell within the field. The solution refinement criterion used is based on the one outlined in (14).g. for detecting shocks.113) The standard deviations of the quantities defined in Equation (3. ⎧ ⎪ ⎪ ⎪ τ j > στ = ⎪ ⎨ ⎪ ⎪ ⎪ζ > σ = ζ ⎪ j ⎩ 2 ⎛ ncell ⎞ ⎜ ∑τ i ⎟ ⎝ i =1 ⎠ ncell ncell Refine cell j 2 (3. 3. around the stagnation point). The purpose of this refinement is to provide enough resolution at the places where there might be discontinuities due to shocks or there are large gradients in the solution (e. ζ = Acell 2 ( ∇ × V ) 3 3 (3. the cells are refined according to the gradient of the solution.8 Solution Refinement Once a certain level of convergence is achieved.
For this case a total of six runs have been performed.1 Subsonic Tests The farfield boundary condition of the first case for the subsonic tests is a free stream Mach number of 0. the flow around a NACA0012 airfoil and an NLR7301 with a flap has been solved for different conditions. 4. One of these solutions is with solution refinement while the other is not. Besides demonstrating the characteristic features of the flows. Also they are compared in terms of computational efficiency. different methods for handling the flux terms in the governing equations have been compared. The experimental data for this case can be found in (32) which were extracted from (33). Two runs are performed second order with and 72 . 4. In the subsonic tests the aim is to show the importance of reconstruction to obtain second order solutions and to discuss the role of first order solution on the stability.6 and an angle of attack of 0˚.1. Two of the six runs are first order solutions. All the runs are performed using Roe’s method.1 NACA0012 The NACA0012 tests have been performed for a subsonic case and a transonic case. The convergence criterion for all of the runs is that the standard deviation of the residual calculated for the continuity equation should drop five levels on a log scale.CHAPTER 4 RESULTS &DISCUSSIONS To test the validity of the code developed. The transonic tests have been performed to demonstrate the shock capturing capabilities of the various methods used to evaluate the flux terms.
1. Figure 4.1. It is not possible to catch the peak of the pressure coefficient with this solution. The pressure coefficient vs. However. chord for first order solution As can be seen from Figure 4.1.Cp vs. while the last two are run first order until the residuals drop four levels then they switch to second order. the first order solution without solution refinement performs poorly. solution refinement introduces a considerable improvement 73 . the chord graph for the first order solutions are presented in Figure 4.without solution refinement.
Cp vs. Figure 4.2. the second order solution provides a much more accurate solution than the first order method as expected. chord for second order solution As can be seen from Figure 4. The pressure coefficient vs.to the solution. The peak of the pressure coefficient is captured accurately for both of the solutions with and without solution 74 . But even with solution refinement the peak of the pressure coefficient graph is not captured accurately.2.2. chord graph can also be illustrated for the second order solution as in Figure 4.
75 . Figure 4. can be compared with the fully second order solution. This is an expected result since there are large gradients around this region. the mesh on the right belonging to a solution with refinement has a considerably denser mesh around the stagnation point than the one without refinement.3. The effect of solution refinement can be illustrated by examining the differences between a refined and nonrefined mesh.A refined and nonrefined mesh around stagnation point In Figure 4. Finally the results obtained with the solution where first order is used until some convergence is obtained and then it is switched to second order.3.refinement although the solution with solution refinement is slightly better on this aspect.
Comparison of the second order solutions In Figure 4.4.Figure 4. it is almost impossible to distinguish between the two solutions. The same situation also prevails for a solution with refinement.4. The Mach contours for the second order solution can be depicted as in Figure 4. 76 . Hence it can be said that in terms of accuracy there is no difference between the two solutions.5.
Mach contours for NACA0012 subsonic flow After comparing the first and second order methods in terms of accuracy. 77 .Figure 4.5.1. their computational performances can also be compared as in Table 4.
Since the pressure is the main factor determining the lift force at high speeds. decreases the computation time considerably compared to full second order solution. angle of attack graph was constructed. 78 . is encountered. Also as illustrated in Figure 4. a viscosity governed phenomena. To obtain the solution Roe’s method was used and solution refinement wass not employed. there is almost no difference between these solutions.6 within the region where there is no separation. The calculations for the second case of subsonic flow around the NACA0012 airfoil were performed with a Mach number of 0.1. As observed from Table 4.1.4167.4. This expectation is realized as can be seen from Figure 4. By changing the angle of attack values a lift coefficient vs. starting the solutions first order then switching to second order after some level of convergence.Comparison of first and second order solutions Method First order without refienement First order with refienement Second order without refinement Second order with refinement Switch to second order without refinement Switch to second order with refinement Number of Iterations Time 1938 45sec 2691 2min 37sec 4325 11min 23sec 7490 1hr 1min 36sec 3236 6min 57sec 4425 25min 24sec Naturally the first order solution converges much faster than the second order methods. it is expected to get a good agreement with the experimental data extracted from (34) until flow separation. but it would not be reasonable to obtain a solution with this method due to poor accuracy.Table 4.
6. first order calculations are performed until some level of convergence is obtained then the solution is switched to second order. divergence in the calculations was observed. As discussed earlier this method decreases computation time while there is no loss in terms of accuracy. The results obtained for 12˚ and 16˚ 79 . However for larger angle of attack values than 8˚.Lift coefficient vs. It was possible to get solutions within the range of angle of attacks from 6˚ to 8˚. Using the first order solution then switching to second order provided results for angle of attack values of 10˚.6. It was also attempted to use a fully second order solution to obtain the results.Figure 4. 12˚ and 16˚ also. angle of attack graph To obtain the results presented in Figure 4.
6 since they are related to a flow with separation and this case must be treated with a viscous solver.Residual plot for first and second order solutions By employing the first order method until some level of convergence is obtained. When the residual plots for first and second order solutions are investigated. In Figure 4. an indication of this fact can be seen.are not presented in Figure 4. it is observed that the residuals decrease smoothly for the first order solution while for the second order one there are a lot of scattering and oscillations as the residuals decrease. However this fact still indicates that the first order solution is more stable than the second order one.7.7. a first order accurate initialization of the field variables is obtained which eases the 80 . Figure 4.
The transonic flow is a more challenging case than the subsonic flow. To verify the computational solution.1. the one on the upper surface being stronger.85 and an angle of attack of 1˚. To compare the accuracy of the different methods for evaluating the fluxes. 4. 81 . Hence the performance of different methods to evaluate the flux terms is compared for this case.2 Transonic Tests The farfield conditions for the transonic flow tests around the NACA0012 profile are a free stream Mach number of 0. For this flow shocks on the upper and lower surface of the airfoil is expected. The calculations have been performed both with and without solution refinement. Each one of the five methods implemented for the evaluation of the flux terms in the governing equations has been employed in the solution to compare these methods. This procedure will be used by default for the rest of the calculations that will be presented. the chord graph for these methods is illustrated in Figure 4. First the results obtained from the AUSM and its derivative methods are compared amongst themselves. The pressure coefficient vs. AGARD data (35) has been extracted from (14).convergence of the second order solution. The most crucial point in the tests performed is the location of the shocks. calculations with solution refinement have been performed.8.
The AUSMD and the AUSMV methods provide a closer estimation of the shock location on this surface than the AUSM method. All of the methods exhibit oscillations around the shock location. the ones on the upper surface being more marked. This is due to the performance of the limiter in the calculation of the fluxes.8. chord graph for the AUSM methods It is observed from Figure 4.Pressure coefficient vs.8 that the AUSMD and AUSMV methods capture the location of the weaker shock on the lower surface while the AUSM predicts this location slightly downstream.Figure 4. 82 . On the upper surface none of the methods were able to capture the shock location exactly.
The oscillations around the shock locations are also 83 .Pressure coefficient vs. Figure 4.9. van Leer and AUSMV It is seen that both the Roe’s method and the van Leer’s method estimate the shock location at a slightly upstream position on the upper surface while there is good agreement on the lower surface. chord graph for these methods is illustrated in Figure 4. chord graph for Roe. The AUSMV method providing the most accurate results among the AUSM methods is chosen as the representative AUSM method and the pressure coefficient vs.The Roe’s method and the van Leer method can be compared with one of the representative AUSM methods.9.
None of the methods were able to provide a five level drop in the residual for solution adaptive calculations. The calculations were performed with four different CFL safety coefficients.Computational characteristics of the methods for evaluating fluxes Method Convergence Number of Iterations Calculation Time CFL Safety Coefficient AUSM converges 5555 9min 1sec 1.0.1.2 and 0. In terms of accuracy AUSMD. four each one of the five methods. Table 4. In Table 4. 84 .0 van Leer diverges  The AUSM and the Roe’s methods converged with a CFL safety coefficient of 1. Besides the accuracy of the methods. the residual of the continuity equation oscillates around 8x104 corresponding to a residual drop of about four levels after about 60000 steps till the maximum number of iterations of 100000. The convergence criterion is that the residual for the continuity equation calculated in the first step should drop five levels on a log scale.1 AUSMV converges 9646 16min 19sec 0.5.inherent in these methods.0 AUSMD oscillates 100000 3hr 35min 4sec 0. AUSMV and the van Leer’s methods provide the best results for this particular test case.0 while the AUSMV method converged with a CFL safety factor of 0.2.5. their convergence characteristics should also be examined. Hence to compare the convergence characteristics calculations without solution refinement have been performed.1.10. The van Leer method diverges when solution refinement is not applied implying a relation between the method and the mesh size. 1. 0. The residual plots for the convergent solutions can be depicted as in Figure 4.5 Roe converges 5207 9min 15sec 1. 0.2. the results for the solution with the maximum CFL safety factor providing a convergent solution has been presented. For the AUSMD method with a CFL safety factor of 0.
85 . The effect of mesh generation in the critical regions around the airfoil is illustrated in Figure 4.11.Figure 4.Residual plot for transonic flow It is also possible to show the effect of solution adaptation on the mesh generated initially.10. The solution refinement has been performed every time the residuals dropped to one twentieth of the initial residual which is updated just before solution adaptation is performed.
Figure 4.11.Effect of refinement at shock location As can be seen from Figure 4.11. the mesh is refined around the location of the shock since there are large gradients in the flow variables at this location.12. Another region where large gradients are observed is at the nose of the airfoil around the stagnation point.12. The effect of solution refinement in this region is illustrated in Figure 4. 86 .Effect of refinement around stagnation point The effect of solution adaptation can be illustrated by plotting the Mach contours for a refined and nonrefined solution. Figure 4.
Hence it can be said that if the main concern is obtaining the forces on the airfoil. the solution without refinement will provide a good estimation in a considerably shorter amount of time. The pressure coefficient vs.14. 87 . chord graphs of these solutions can also be displayed. However in Figure 4. it is seen that the overall flow field is not represented accurately without solution refinement. it is seen that the solution obtained without refinement is able to capture both the pressure distribution on the airfoil and the shock locations with refinement introducing a slight improvement on both aspects. the Mach contours obtained with the AUSMV method are displayed.13. the shocks are smeared. In Figure 4.13. As can be observed for the solution without refinement.Mach contours for transonic flow around NACA0012 airfoil In Figure 4.Figure 4.13.
The definition of these variables can be shown as in Figure 4. 88 .15. chord for AUSMV with and without refinement 4.3% of chord length and a flap angle of 20˚.2 NLR 7301 with Flap The NLR 7301 tests have been performed for a gap of 2. an overlap of 5.14.Figure 4.6% percent of chord length.Pressure coefficient vs.
Figure 4.15 Definition of geometry variables for NLR 7301 with flap
Tests were performed at three different angle of attack values; 6˚, 10.1˚ and 13.1˚. The free stream Mach number for all the cases was 0.185. The five methods; Roe, van Leer, AUSM, AUSMD, AUSMV, have been employed in the calculations. To verify the results computed, data obtained from (36) has been used. The pressure coefficient vs. chord graph for angle of attack of 6˚ is presented for the AUSM methods in Figure 4.16. All of the AUSM methods are able to capture pressure distribution with good accuracy for the most part. But in terms of capturing the peak of the pressure coefficient on the upper surface of the main airfoil, AUSM method is superior to the others. The AUSMD method performs better than the AUSMV method on this aspect.
89
Figure 4.16 Pressure coefficient vs. chord graph (AUSM methods AoA 6˚)
The pressure coefficient vs. chord graph can also be constructed for the Roe’s, van Leer’s and AUSM methods. In Figure 4.17, it is seen that the AUSM method provides better results than both the Roe’s and van Leer’s methods. The solution by Roe’s method is comparable to the other AUSM methods. In terms of capturing the peak of the pressure coefficient on the upper surface of the main element, Roe’s method performs similar to the other method while van Leer provides the poorest result in terms of accuracy.
90
Figure 4.17 Pressure coefficient vs. chord graph (AUSM, Roe, van Leer AoA 6˚)
The computational performances of each method can be tabulated to obtain a comparison.
91
46 2. do not converge. it is seen that the performance of the AUSM method is better than both the AUSMD and the AUSMV methods. The other methods.42 2.3.45 2. The AUSMD. 92 .51 2.Computational performance of the methods for NLR 7301 AoA 6˚ Method Convergence Number of Iterations Time AUSM converges 7072 41min 47sec AUSMD oscillates 100000 10hr 21min 49sec AUSMV oscillates 100000 10hr 14min 0sec Roe oscillates 100000 10hr 56min 15 sec van Leer converges 4172 22min 53sec Lift Coefficient 2.35 It is observed from Table 4. The pressure coefficient vs. AUSMD. that is the residuals in the continuity equation do not drop five levels but they fluctuate until the maximum number of iterations is reached. The AUSMV method starts to deviate from the AGARD data which is especially apparent on the upper surface of the flap. This is an expected result as the AUSM method captures the peak of the pressure distribution on the upper surface of the main airfoil better than the other methods.1 can be presented for the AUSM methods. From Figure 4. The peak of the pressure coefficient is captured accurately with the AUSM method. Also it can be seen that the highest lift coefficient is found with the AUSM method.3 that the van Leer’s method converges faster than the AUSM method. chord graph for the angle of attack value of 10. AUSMV and Roe.18.Table 4. AUSMV and Roe’s methods provide a similar result on this aspect while van Leer’s method gives the smallest estimate as it was the worst in capturing the peak.
18.19. chord graph (AUSM methods AoA 10. Roe and van Leer methods as in Figure 4. 93 .Figure 4. the chord graph can be presented for the AUSM.Pressure coefficient vs.1˚) The pressure coefficient vs.
19. chord graph (AUSM. Roe.Pressure coefficient vs.4. van Leer AoA 10.Figure 4. The computational performance of the methods can also be compared from Table 4.19. AUSM method provides a better solution. It is also observed that the Roe’s method provides a better estimation of the peak in the pressure distribution on the upper surface of the main element than the AUSMD and AUSMV methods. 94 .1˚) As can be seen from Figure 4.
98 2. AUSMD and AUSMV methods essentially remain the same computationally.1˚ Method Convergence Number of Iterations Time AUSM converges 11567 1hr 10min 43sec AUSMD oscillates 100000 10hr 21min 5sec AUSMV oscillates 100000 10hr 13min 2sec Roe converges 78924 8hr 15min 47 sec van Leer converges 4175 22min 50sec Lift Coefficient 2. estimates a relatively higher lift coefficient than the previous case. The performances of the van Leer. The AUSM method still being the best one to capture the peak provides the highest estimate of the lift coefficient. From Figure 4.83 It is observed from Table 4. Once again the AUSM method provides a considerably better estimate of the pressure peak on the upper surface of the main element.75 2.86 2.4.Computational performance of the methods for NLR 7301 AoA 10.20.76 2.1˚ for the AUSM methods.4 that as the angle of attack is increased Roe’s method converges like the AUSM and the van Leer methods. 95 .Table 4. the chord graph can be constructed for the angle of attack of 13. The pressure coefficient vs. it is seen that the AUSMV method fails at this angle of attack. The Roe’s method being able to capture the pressure peak better for this case.
Figure 4. As in the case of angle of attack of 10. the Roe’s method gives a better solution than the AUSMD and AUSMV method which fails.1˚. the Roe’s method provides a better solution than van Leer’s method. chord graph (AUSM methods AoA 13.20.21 the pressure coefficient vs. 96 .Pressure coefficient vs. chord graph for the AUSM.1˚) In Figure 4. The AUSM method providing the best of the solutions. Roe and van Leer methods are depicted.
97 .5.Pressure coefficient vs. chord graph (AUSM.1˚ can be compared with the aid of Table 4.21. van Leer AoA 13. Roe.Figure 4.1˚) The computational performance of the methods for the case of angle of attack of 13.
5. which makes the problem shift closer to the compressible range.08 3. Solution adaptive calculations have also been performed for the van Leer method to improve the results.16 The computational performance of the methods is pretty similar to the case of angle of attack of 10. 98 .03 3. The performance of the AUSMD method remains constant while the AUSMV method fails as the angle of attack is increased.17 3.29 3.Computational performance of the methods for NLR 7301 AoA 13. It can be concluded that AUSM method provides the best results in all the cases considered.22 and Figure 4.1˚ with the exception that AUSMV method fails totally.Table 4. Roe’s and van Leer’s methods do improve as the angle of attack increases which may be due to the fact that as the angle of attack increases so does the velocities around the airfoil.1˚ Method Convergence Number of Iterations Time AUSM converges 12182 1hr 14min 5sec AUSMD oscillates 100000 10hr 20min 30sec AUSMV oscillates 100000 10hr 13min 5sec Roe converges 87222 9hr 11min 9sec van Leer converges 4396 24min 12sec Lift Coefficient 3. The effect of solution refinement on the mesh can be shown as in Figure 4.23.
22.Figure 4.23.Effect of solution refinement around stagnation region of NLR 7301 Figure 4.Effect of solution refinement around the flap 99 .
The pressure coefficient vs. This is much closer to the one predicted by the AUSM method.35 to 2. chord for van Leer AoA 6˚ In Figure 4.49. In fact the solution is improved around this region as a whole. The estimate of the lift coefficient increases from 2.24.Pressure coefficient vs. Figure 4. it is seen that the adapted solution performs better at capturing the peak of the pressure distribution on the upper surface of the main element. 100 . the chord graphs for the solution with and without refinement can be constructed for a comparison at 6˚.24.
but this time there is a considerable over estimation of the peak of the pressure coefficient on the upper surface of the main element. 101 . the chord graph for a solution with and without refinement for angle of attack of 10.25.1˚ There is an overall improvement of the solution with the introduction of refinement.26. the pressure coefficient vs.Pressure coefficient vs. chord for van Leer AoA 10. Lastly.25. Figure 4. In fact the estimation for the lift coefficient becomes 3.1˚ is presented in Figure 4.1˚ is presented in Figure 4.The pressure coefficient vs.05 which is greater than predicted by the AUSM method. the chord graph for a solution with and without refinement for angle of attack of 13.
but there is an over prediction of the pressure coefficient peak on the upper surface of the main element.27.1˚.Pressure coefficient vs.1˚ respectively. Figure 4.26.1˚ Once again the overall solution does improve.28 and Figure 4. 102 . 13. Figure 4.Figure 4. chord for van Leer AoA 13. 10.29 illustrate the flow fields obtained for angle of attack values of 6˚.
Flow field for NLR 7301 with flap AoA 10.28.1˚ 103 .Flow field for NLR 7301 with flap AoA 6˚ Figure 4.Figure 4.27.
and also the pressure drop on the upper surface of the main element increases. An important feature of the NLR 7301 airfoil is the region between its lower surface and the flap. There were no vortex formations in this region for the flows considered. the acceleration of the flow on the upper surface of the main element increases also.1˚ As the angle of attack increases.29. with the increasing angle of attack the region of high pressure on the lower surface of the main element becomes larger. This promotes higher lift forces.Flow field for NLR 7301 with flap AoA 13. In addition. This makes the streamlines more compact. due to the pressure drop associated with it. which is consistent with the computations.Figure 4. If there were a vortex formation in this region. there would be decrease in the lift force accompanied by an increase in the drag force since the orientation of the face normal 104 .
of the main element is towards increasing the lift and decreasing the drag in this region. Hence viscous calculations are of great importance in this region because of a possibility of separation. 105 .
The second order method performed much better than a first order solution with refinement. first order calculations would be performed until some level of convergence is obtained then the solution would switch to a second order formulation. the mesh around the regions of high gradients. This resulted in up to more than 50% decrease in the computation time while there was no virtual loss in terms of accuracy. since the nonadapted solution performed reasonably well. There is however a considerable improvement of the solution with the introduction of solution refinement. chord graph it was seen that at certain angles it was not possible to get a solution with fully second order calculations. To resolve this issue. This and the fact that second order solutions 106 . is refined and the accuracy in this region increases as the peak of the pressure distribution can be captured more accurately. there was little improvement in the solution for NACA0012 subsonic flow. the difference between the accuracy of a first order and a second order formulation has been illustrated. the stagnation region in case of NACA0012 subsonic flow. When solution refinement was performed for a second order formulation. Even though solution refinement did introduce an improvement on the first order solution. to capture the peak of the solution refinement. it still wasn’t accurate enough. This is due to the fact that with solution refinement. The first order method is inadequate in capturing the flow characteristics such as the peak of the pressure distribution for the relatively simple problem of subsonic flow. The second order method was superior to the first order one in terms of accuracy. Also as the angle of attack was varied to construct a lift coefficient vs.CHAPTER 5 CONCLUSIONS In the subsonic calculations. but it did call for a much increased computation time.
For the solutions without refinement it was seen that. It was however possible to get an accurate pressure distribution on the airfoil as the mesh elements in this region are small in size. the Roe’s method and the AUSM method converged easily with a CFL safety coefficient of 1. The transonic flow around the NACA0012 profile was a more challenging case than the subsonic flow. so calculations were also performed without refinement.provided residual plots with considerable scattering and oscillations indicated that the first order formulation is more stable than the second order one.5 while the van Leer’s method diverged. Hence by starting with first order calculations. especially the stronger one on the upper surface of the airfoil. The main challenge was to catch the shock locations. Another important feature of the solutions for the transonic flow was the oscillations seen at the shock locations. This is due to the poor performance of the limiters used in the calculations. providing grounds to test the performance of different methods for evaluating the flux terms. the field is initialized more accurately which promotes the second order method to converge as it is sensitive to the initial conditions. The AUSMD. A limiter that won’t hamper the convergence while preventing oscillations should be implemented to improve the solution. 107 . the one on the upper surface of the airfoil with the stronger shock being more apparent. If solution refinement is not used smeared shocks are obtained. Hence just to analyze the forces on the airfoil a solution without refinement can be employed to get quick results. AUSMV and van Leer’s methods were the ones that showed the best performance on this aspect. It was not possible to compare the computational performance of the methods with a solution with refinement. For the transonic flow solution refinement is a must to get an accurate representation of the flow field since it is already known that there will be large gradients at shock locations in addition to the stagnation region. The AUSMV method converged with a CFL safety coefficient of 0.
van Leer’s method with solution refinement and AUSM method performed well on this aspect. and Roe’s method performed worse with similar results at an angle of attack 6˚. The performance of both the Roe’s and van Leer’s methods improved for the cases with higher angle of attacks. This might be due to the flow shifting closer to the compressible range with the increasing angle of attacks promoting higher velocities in the field. From the pressure coefficient vs.1˚. while AUSMD.The NLR 7301 with a flap was also a challenging case to test the performance of the various methods implemented for evaluating the flux terms. The van Leer’s method with the poorest performance for the case with an angle of attack of 6˚ did improve with the application of solution refinement. 108 . AUSM was the superior method once again while the performance of Roe’s method improved considerably. The van Leer method with solution refinement provided over estimations of the pressure peak. it was observed that the estimation of the peak pressure on the upper surface of the main element was crucial in the prediction of the lift coefficient. The AUSM method was superior in terms accuracy and it was second to van Leer’s method in terms of computational performance. chord graphs. The performance of the AUSMD method was consistent throughout the calculations while the AUSMV method failed for large angle of attack values. AUSMV.1˚ and 13. For the angle of attack values of 10.
Bharat K and Wheatherill. 1988. : Thompson Books. C. Wichita : Enginnering Education System. Communications of the ACM. H K and Malalasekera. Charles. Thompson. 3. Versteeg. Steve T. Klaus A and Chiang. W. A Clipping Divider. Reddy. Hoffmann. Wendt. 4. Grid Generation Methods. Berlin : SpringerVerlag. Malaysia : Prentice Hall. J N. 3. January 1974. Numerical Computation of Internal and External Flows. 5. 1993. pp. 1. New York : John Wiley & Sons.BIBLIOGRAPHY 1. Washington D. Joe F. Reentrant Polygon Clipping. 1999. 1998. Liang. 3242. R F and Sutherland. 8. Handbook of Grid Generation : CRC Press. 7. Vladimir D. I E and Hodgeman. et al. 122. 2. Vol. A New Concept and Method for Line Clipping. Berlin : Springer_Verlag. 1996. G W. I E. Computational Fluid Dynamics for Engineers. B A. 6. Nigel P. ACM Transactions on Graphics. Sutherland. Computational Fluid Dynamics An Introduction. Soni. 9. 1968. Singapore : McGrawHill. Y D and Barsky. 1995. 1995. An Itroduction to Computational Fluid Dynamics The Finite Volume Method. 10. John F. An Introduction to the Finite Element Method. January 1984. Liseikin. Hirsch. 109 . Sproull. pp.
Computer Graphic Forums. Lee. 12. Approximate Riemann Solvers. pp. Day. Cambridge : Cambridge University Press. E F. Culbert B. Laney. Berlin : SpringerVerlag. Sobkow. pp. Pospisil. Roe. 15. Computer Graphics. 20. P and Yang. : SpringerVerlag. A Fast TwoDimensional Line Clipping Algoritm via Line Encoding. Proceedings of the 8th International Conference on Numerical Methods in Fluid Dynamics. s. Vol. Computational Geometry an Introduction. Toro. Y H. B. Nicholl. 1998. M S. 1993. New York : SpringerVerlag. A QuadTree Based AdaptivelyRefined CartesianGrid Algorithm for the Solution of The Euler Equations. July 1987. pp. New Jersey : Prentice Hall. 18. 43. 19. Riemann Solvers and Numerical Methods for Fluid Dynamics. pp. 1981. Preparata. Computer Graphics. 1992. Michigan : The University of Michigan. 241245.11. PhD thesis. Flux Vector Splitting for the Euler Equations. 1987. Michael D. Advanced Engineering Mathematics. 110 . Journal of Computational Physics. 357372. P Franco and Shamos. 253262. 14.l. A New Two Dimensional Line Clipping Algorithm for Small Windows. Greenberg. D T and Nicholl. P L. Darren L. 1982. Tina M. 459467. 1999. and Difference Schemes. 13. J D. Michael Ian. 1985. van Leer. An Efficent Algorithm for 2D Line Clipping: Its Development and Analysis. 1998. Robin A. 17. Parametrized Vectors. De Zeeuw. 16. Computational Gas Dynamics.
Barth. 23. s. 22. 27. C J. 1997. Dennis C. 26. 1998. Flux Vector Splitting of the Inviscid Gasdynamic Equations with Applications to Finite Difference Methods. PhD thesis. A New Flux Splitting Scheme. M S. pp. 263293. Vol. 28th Aerospace Sciences Meeting. 633657. Recent Developments in High Order KExact Reconstruction on Unstructured Meshes. 2339. PhD thesis. Journal of Computational Physics. Higher Order Solution of the Euler Equations. pp. Liou. 40. Cartesian. 31st Aerospace Sciences Meeting. pp. Gerris: A TreeBased Adaptive Solver for the Incompressible Euler Equations in Complex Geometries. 1990. Timothy J. 29. Simplified Numarical Methods for Gasdynamic Systems on Triangulated Domains. Timothy J and Frederickson. Vol. William John. An Accurate and Robust Vector Splitting Scheme for Shock and Contact Discontinuities. Barth. Comput. Reno : AIAA paper AIAA900013. Barth. Sci.l. Timothy J and Jespersen. 28. 190. : Journal of Computational Physics. 25. 111 . Timothy J. Barth. Reno : AIAA paper AIAA930668.21. 18. An Adaptively Refined. Stéphane. The Design and Application of Upwind Schemes on Unstructured Meshes. 1994. Popinet. R F. 2003.. Vol. Reno : AIAA paper AIAA890366. Michigan : The University of Michigan. 107. : Siam J.l. 24. s. 1993. Coirier. 1993. pp.l. Steger. Yasuhiro and Liou. 1989. s. M S and Steffen. 27th Aerospace Sciences Meeting. Vol. J L and Warming. Paul O. CellBased Scheme for the Euler and Navier Stokes Equations. : Journal of Computational Physics. Stanford : Stanford University. 1981. 572600. Wada. 3.
l. Advisory Group for Aerospace Research& Development. Özdemir. 34. 2001.l. 32. Amick. 2005. 1950. A Selection of Experimental Test Cases for the Validation of CFD Codes . 1959. 31. 33. Comparison of the Experimental Pressure Distribution on an NACA0012 Profile at High Speeds with that Calculated by the Relaxation Method. Technical note. Implementation of Rotation into a 2D Euler Solver. Computational Fluid Dynamics: Principles and Applications. J L. 1986 : AGARD Advisory Report 211. J D. J. Masters thesis. Abbott. Albert E. : National Advisory Committee for Aeronautics. 1994. London : Wiley. 112 . 35. Theory of Wing Sections. s. New York : Dover Publications. Lambert. Ankara : METU. AGARD Subcommittee C.Including a Summary of Airfoil Data.30. St Augustin : Elsevier. Ira H and von Doenhoff. : AGARD Advisory Report No 303. s. Computational Methods in Ordinary Differential Methods. Blazek. Test Cases for Inviscid Flow Field Methods. 1973. 36. Enver Doruk.
009619 0.964888 0.975528 0.022773 0.945503 0.895078 0.002496 0.922164 0.017539 0.952414 0.987958 0.020107 0.002120 0.016297 0.938153 0.970440 0.018809 0.003443 0.993844 0.002937 0.010628 0.001812 0.984292 0.001398 0.996057 0.904508 0.818712 y 0.012778 0.958877 0.Coordinates of NACA0012 Profile x 0.999753 0.015088 0.930371 0.991144 0.006887 0.004642 0.004012 0.006082 0.013914 0.875056 0.005333 0.913540 0.024135 113 .008658 0.021429 0.830656 0.001571 0.999013 0.853553 0.1.842274 0.980147 0.864484 0.011681 0.APPENDIX A COORDINATES OF NACA0012 Table A.997781 0.007746 0.001295 0.885257 0.
057789 0.Table A.301426 0.726995 0.515705 0.035374 0.287110 0.390928 0.345492 0.039553 0.050854 0.259123 0.051923 0.043585 0.684062 0.033962 0.639496 0.055655 0.484295 0.609072 0.562667 0.375655 0.059848 0.421783 0.Coordinates of NACA0012 Profile (continued) 0.058355 0.046149 0.056436 0.028307 0.1.026905 0.059619 0.047383 0.232087 0.654508 0.273005 0.740877 0.049739 0.057148 0.698574 0.058866 114 .031131 0.330631 0.624345 0.593691 0.059251 0.042263 0.044882 0.040917 0.060017 0.437333 0.058844 0.531395 0.059980 0.712890 0.406309 0.052940 0.781042 0.059960 0.767913 0.500000 0.059575 0.054810 0.793893 0.578217 0.245479 0.032547 0.754521 0.029717 0.025514 0.452946 0.048581 0.053904 0.360504 0.315938 0.468605 0.547054 0.059292 0.806454 0.059812 0.669369 0.036778 0.038172 0.
035112 0.056160 0.146447 0.023517 0.169344 0.029560 0.006156 0.Table A.104922 0.135516 0.193546 0.054207 0.030473 0.008856 0.028213 0.057714 0.206107 0.015708 0.040776 0.058340 0.013503 0.003943 0.021088 0.047586 0.019853 0.024472 0.181288 0.005521 0.077836 0.044374 0.012042 0.086460 0.000000 0.Coordinates of NACA0012 Profile (continued) 0.051863 0.041123 0.218958 0.018607 0.124944 0.042615 0.016078 0.002219 0.038859 0.000987 0.047638 0.055233 0.046049 0.050547 0.157726 0.1.010884 0.053083 0.114743 0.025893 0.061847 0.000000 115 .032671 0.008223 0.069629 0.049138 0.034803 0.054497 0.056987 0.095492 0.036867 0.
38337 0.01580 0.56019 0.08826 0.89870 0.85349 0.06191 0.93988 0.07038 0.53092 0.04814 0.93524 0.88524 0.87015 0.03941 0.08590 0.08015 0.47178 0.The coordinates of NLR 7301 x 0.35444 0.02241 0.01499 0.50141 0.83532 0.32598 y 0.08781 0.01830 0.08818 0.61758 0.02819 0.91047 0.74911 0.06628 0.08438 0.58912 0.03162 0.1.04368 0.44216 0.64547 0.81572 0.08838 0.08768 116 .05735 0.94267 0.01519 0.07410 0.05273 0.79476 0.APPENDIX B COORDINATES OF NLR 7301 & FLAP Table B.69908 0.02512 0.67267 0.02008 0.01690 0.72459 0.08248 0.08704 0.94360 0.03536 0.92051 0.41265 0.92878 0.77253 0.07737 0.
04487 0.00089 0.00368 0.06650 0.06807 0.01479 0.07874 0.08276 0.00206 0.02521 0.00422 0.03877 0.01870 0.01804 0.06958 0.05268 0.02173 0.03544 0.1.05506 0.24448 0.00089 0.02305 0.12784 0.01516 0.06137 0.01417 0.05139 0.01870 0.02851 0.03076 0.08087 0.00080 0.17103 0.00832 0.02786 0.21897 0.02582 0.00004 0.00577 0.05832 0.03309 0.00959 0.01479 0.04717 0.00368 0.05009 0.10824 0.06567 0.29809 0.07103 0.00019 0.02064 0.08440 0.27098 0.00019 0.14880 0.08687 0.00577 0.The coordinates of NLR 7301 (continued) 0.09897 0.00832 0.11786 0.03158 0.19445 0.04373 0.07341 0.01133 0.01133 0.03445 117 .05731 0.05943 0.06488 0.07381 0.Table B.08155 0.06317 0.09007 0.03978 0.07638 0.07245 0.08577 0.00206 0.00992 0.00552 0.
05139 0.05832 0.38337 0.01597 0.61758 0.17103 0.14880 0.04487 0.00080 0.07550 0.06259 0.04215 0.07144 0.11786 0.09007 0.06417 0.05954 0.03750 0.05959 0.02786 0.00585 118 .08155 0.06402 0.56019 0.04661 0.07341 0.53092 0.69908 0.04355 0.00824 0.47178 0.41265 0.50141 0.04873 0.81572 0.06819 0.10824 0.79476 0.04921 0.77253 0.06109 0.06917 0.05073 0.02362 0.07666 0.03080 0.07585 0.05459 0.12784 0.27089 0.06567 0.04443 0.03972 0.07652 0.05263 0.64547 0.03717 0.29809 0.05622 0.07385 0.1.67267 0.02305 0.07704 0.05791 0.05446 0.58912 0.Table B.07317 0.72459 0.24448 0.03310 0.06672 0.The coordinates of NLR 7301 (continued) 0.44216 0.09897 0.07132 0.32598 0.07470 0.74911 0.03877 0.21897 0.19445 0.35444 0.07701 0.
92051 0.020749 119 .87015 0.83532 0.93524 0.088996 1.045428 0.022570 y 0.071662 0.01410 Table B.The coordinates of flap x 1.200754 1.023205 0.01460 0.025861 0.01818 0.034929 0.056694 0.85349 0.154588 1.106992 1.188106 1.195420 1.205092 1.94360 0.01472 0.178790 1.038286 0.028701 0.Table B.032094 1.01115 0.101411 0.102133 0.1.091056 0.01645 0.041673 1.079729 1.167564 1.085249 0.88524 0.93988 0.92878 0.041798 0.095931 0.051260 1.140013 1.01423 0.01843 0.124077 1.064282 0.070330 1.049135 0.2.91047 0.098093 1.89870 0.01702 0.The coordinates of NLR 7301 (continued) 0.01530 0.060832 1.94267 0.099373 0.031728 0.078726 0.204014 1.01807 0.01733 0.
030749 0.985653 0.966901 0.038526 0.894976 0.014812 0.2.993974 0.930955 0.907107 0.011985 0.911738 0.011407 0.958495 0.896720 0.039670 0.032320 0.935621 0.010922 0.893925 0.899614 0.034582 0.895009 0.031741 0.914029 0.012722 0.028901 0.Table B.960148 0.893925 0.040906 0.018508 0.924783 0.937624 0.893589 0.017719 0.010407 0.016532 0.036418 0.025374 0.919139 0.916920 0.032993 0.928889 0.984677 0.942821 0.994643 0.035469 0.003821 0.903046 0.011390 0.033757 0.014820 0.The coordinates of flap (continued) 1.010799 0.944743 0.023517 0.010473 0.031232 0.013137 1.042203 0.010583 0.013659 0.030054 0.968366 0.016156 0.021474 0.043586 120 .896887 0.922644 0.012269 0.905415 0.950452 0.909442 0.013410 0.901946 0.027199 0.976874 0.019522 0.037440 0.975646 0.899047 0.952264 0.
194459 1.101185 1.085915 0.060300 0.022976 1.082262 1.090662 0.053820 0.072539 1.003479 1.164948 1.Table B.058038 0.032883 1.The coordinates of flap (continued) 1.200104 1.048257 0.042829 1.052791 1.075465 0.062706 1.070188 0.094993 0.102670 0.051865 0.065091 0.203556 1.2.055875 0.151267 1.045045 0.103214 121 .046609 0.091824 1.176820 1.101159 0.013157 1.050011 0.119166 1.135943 1.186673 1.080787 0.098572 0.062652 0.204699 0.
This action might not be possible to undo. Are you sure you want to continue?
Use one of your book credits to continue reading from where you left off, or restart the preview.