This action might not be possible to undo. Are you sure you want to continue?

2009-005

Numerical Methods for the Navier–

Stokes Equations applied to Turbulent

Flow and to Multi-Phase Flow

MARTIN KRONBICHLER

UPPSALA UNIVERSITY

Department of Information Technology

Numerical Methods for the Navier–

Stokes Equations applied to Turbulent

Flow and to Multi-Phase Flow

BY

MARTIN KRONBICHLER

December 2009

DIVISION OF SCIENTIFIC COMPUTING

DEPARTMENT OF INFORMATION TECHNOLOGY

UPPSALA UNIVERSITY

UPPSALA

SWEDEN

Dissertation for the degree of Licentiate of Philosophy in Scientiﬁc Computing with

specialization in Numerical Analysis

at Uppsala University 2009

Numerical Methods for the Navier–

Stokes Equations applied to Turbulent

Flow and to Multi-Phase Flow

Martin Kronbichler

martin.kronbichler@it.uu.se

Division of Scientiﬁc Computing

Department of Information Technology

Uppsala University

Box 337

SE-751 05 Uppsala

Sweden

http://www.it.uu.se/

c Martin Kronbichler 2009

ISSN 1404-5117

Printed by the Department of Information Technology, Uppsala University, Sweden

Abstract

This thesis discusses the numerical approximation of ﬂow problems, in particular the large eddy

simulation of turbulent ﬂow and the simulation of laminar immiscible two-phase ﬂow. The com-

putations for both applications are performed with a coupled solution approach of the Navier–

Stokes equations discretized with the ﬁnite element method.

Firstly, a newimplementation strategy for large eddy simulation of turbulent ﬂowis discussed. The

approach is based on the variational multiscale method, where scale ranges are separated by vari-

ational projection. The method uses a standard Navier–Stokes model for representing the coarser

of the resolved scales, and adds a subgrid viscosity model to the smaller of the resolved scales.

The scale separation within the space of resolved scales is implemented in a purely algebraic way

based on a plain aggregation algebraic multigrid restriction operator. A Fourier analysis underlines

the importance of projective scale separations and that the proposed model does not affect consis-

tency of the numerical scheme. Numerical examples show that the method provides better results

than other state-of-the-art methods for computations at low resolutions.

Secondly, a method for modeling laminar two-phase ﬂow problems in the vicinity of contact lines

is proposed. This formulation combines the advantages of a level set model and of a phase ﬁeld

model: Motion of contact lines and imposition of contact angles are handled like for a phase ﬁeld

method, but the computational costs are similar to the ones of a level set implementation. The

model is realized by formulating the Cahn–Hilliard equation as an extension of a level set model.

The phase-ﬁeld speciﬁc terms are only active in the vicinity of contact lines. Moreover, various as-

pects of a conservative level set method discretized with ﬁnite elements regarding the accuracy of

force balance and prediction in jump of pressure between the inside and outside of a circular bub-

ble are tested systematically. It is observed that the errors in velocity are mostly due to inaccura-

cies in curvature evaluation, whereas the errors in pressure prediction mainly come fromthe ﬁnite

width of the interface. The error in both velocity and pressure decreases with increasing number

of mesh points.

List of Papers

This thesis is based on the following papers, which are referred to in the text by

their Roman numerals.

I V. Gravemeier, M. W. Gee, M. Kronbichler, and W. A. Wall (2009) An alge-

braic variational multiscale–multigrid method for large eddy simulation of

turbulent ﬂow. Accepted for publication in Computer Methods in Applied

Mechanics and Engineering. doi:10.1016/j.cma.2009.05.017.

Contributions: Initial implementation of the method together with the ﬁrst and

second author. Preparation of a draft of the manuscript, in particular section 2.

Ideas were developed in close cooperation between the authors.

II V. Gravemeier, M. Kronbichler, M. W. Gee, and W. A. Wall (2009) An

algebraic variational multiscale–multigrid method for large-eddy

simulation: generalized-α time integration, Fourier analysis and

application to turbulent ﬂow past a square-section cylinder. Submitted to

Computers and Fluids.

Contributions: Ideas and analysis in section 3. Preparation of a draft of the

manuscript and ﬁnal work in close cooperation between the authors.

III M. Kronbichler and G. Kreiss (2008) A Hybrid Level–Set–Cahn–Hilliard

Model for Two-Phase Flow. In Proceedings of the 1st European Conference

on Microﬂuidics, Bologna, Italy.

Contributions: Implementation of the method and computations. Preparation of

the manuscript. Ideas were developed in discussion with the co-author.

IV S. Zahedi, M. Kronbichler, and G. Kreiss (2009) Evaluation of inaccuracies

in two-phase ﬂow simulations with level set methods using ﬁnite element

discretizations. Technical Report 2009-026, Department of Information

Technology, Uppsala University, 2009. Submitted to Computers and Fluids.

Contributions: Implementation of and experiments with the coupled two-phase

ﬂow solver. First draft of the manuscript. Ideas were developed by close

cooperation between the ﬁrst and the second author in discussion with the third

author.

Reprints were made with permission from the publishers.

Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Numerical Methods for the Navier–Stokes Equations . . . . . . . . . . . . . . . . 3

2.1 Weak form and spatial discretization . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2.2 Stability of spatial discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.3 Time discretization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.4 Numerical linear algebra for coupled discretizations . . . . . . . . . . . . . 6

3 Computational methods for turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3.1 Basics on turbulence and large eddy simulation . . . . . . . . . . . . . . . . . 7

3.2 A multiscale approach to large eddy simulation . . . . . . . . . . . . . . . . . 9

4 Simulation of Two-Phase Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4.1 Overview of numerical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

4.2 Level set models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4.3 Phase ﬁeld models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

5 Summary of Papers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.1 Paper I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.2 Paper II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

5.3 Paper III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5.4 Paper IV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

6 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1. Introduction

The numerical simulation of ﬂuid dynamics is one of the main ﬁelds in compu-

tational mathematics. Simulationis today bothanalternative anda complement

to experiments in many engineering disciplines as it helps in predicting the be-

havior of ﬂuids. Moreover, simulation permits to systematically improve the de-

sign and material properties of products by means of optimization algorithms.

Such a procedure would be tremendously expensive if it had to be done by pro-

totyping and other experimental means. Simulation is thus a tools that renders

technical processes more effective.

One interesting area of research in ﬂuid dynamics is the behavior of turbulent

incompressible ﬂow. Turbulent ﬂows often occur in engineering applications,

for example, in the air ﬂow around a vehicle, the water ﬂow in turbines of hy-

droelectric power plants as well as in water networks and oil pipelines with high

throughput. For instance, relevant issues are the determination and the control

of aerodynamic drag as well as efﬁciency of engines.

Another area of active research is multiphase ﬂow, a setting that involves two

or more different ﬂuids. There is a wide range of applications that involve multi-

ple ﬂuids and where numerical simulation is extensively used. An exceptionally

challenging problem is computational combustion, where different ﬂuids and

ﬂame fronts need to be represented. Often, the ﬂows occurring in combustion

are turbulent. But there are also a lot of multi-phase settings where the ﬂowﬁeld

is laminar and thus more regular. One such application is intravenous therapy in

medicine. The objective of numerical simulations is to understand how bubbles

consisting of oleaginous substances are transported and resolved in the circu-

lation of blood. These processes are often strongly driven by surface tension, a

force that strives after keeping different ﬂuids separated. A similar application is

the modeling of the absorption of air in the blood circulation in human lungs.

Multiphase ﬂow is a phenomenon that is particularly relevant also in subsur-

face ﬂow. In this area, it is often assumed that the individual ﬂuids are well-

separated, that is, the ﬂuids do not mix and there are no additional particles

resolved in the ﬂuids. This thesis contributes to the simulation of such lami-

nar well-separated two-phase ﬂows. For subsurface ﬂows, an additional difﬁ-

culty arises, namely the interaction of systems consisting of different ﬂuids with

solids, called capillarity. This interaction takes place on so-called contact lines,

locations where the interface (separating two ﬂuids) meets a solid. The prop-

erties of the solid and the composition of the ﬂuid give rise to a phenomenon

called wetting, indicating which ﬂuid preferably is in contact with the solid. Ap-

plications of this kind are the modeling of groundwater basins and oil extraction

1

from subsurface reservoirs driven by water injection. The pores of rock where

the ﬂuid moves are often of the order of a few millimeters or less, whereas the

actual simulation region can extend over thousands of square kilometers. In or-

der to tackle problems with such an enormously wide range of scales, simpliﬁca-

tions need to be applied. One example is to performsmall-scale ﬂowsimulations

with detailed rock structures in order to determine the permeability of various

stone conﬁgurations. Such effective parameters are then used as input to a large

scale simulation, usually based on Darcy’s law [17]. Other applications of multi-

ple ﬂuids in contact with solids are found in industrial processes like sintering,

where mixtures of different ﬂuids and their solidiﬁcation are tracked; lab-on-a-

chip processes, which form the basic component for microsystem design in the

biomedical industry; and inkjet printing.

This thesis consists of ﬁve chapters. First, we discuss the equations that de-

scribe the ﬂow of individual ﬂuids, the Navier–Stokes equations. Next, an intro-

duction to turbulence is given, which is followed by a presentation of numerical

models for the simulation of laminar two-phase ﬂow. Chapter 5 gives a short

summary of the papers included in this thesis, and Chapter 6 identiﬁes possible

directions of future research.

2

2. Numerical Methods for the

Navier–Stokes Equations

The motion of ﬂuids at low and medium velocities in a connected computa-

tional domain Ω⊂R

d

(spatial dimension d =2, 3) is given by the incompressible

Navier–Stokes equations [35], [62], [73],

ρ(u

t

+u· ∇u) −∇·

_

2µε(u)

_

+∇p =ρf, (2.1)

∇· u=0. (2.2)

The vector udenotes the d components of ﬂuid velocity and p denotes the ﬂuid’s

(dynamic) pressure. The rank-2 tensor ε(u) =

1

2

_

∇u+(∇u)

T

_

represents the vis-

cous stress tensor for a Newtonian ﬂuid. The ﬂuid density is denoted by ρ, and

the ﬂuid’s (dynamic) viscosity by µ. In single-phase ﬂuids, density is constant

due to the incompressibility assumption. The termf speciﬁes external forces act-

ing on the ﬂuid.

The system of Navier–Stokes equations (2.1)–(2.2) is completed by a diver-

gence-free initial velocity ﬁeld,

u(·, 0) =u

0

, with ∇· u

0

=0, (2.3)

and application-speciﬁc boundary conditions (cf. Gresho and Sani [32] for a dis-

cussionof various possibilities). Inthis thesis, we consider ﬂowsubject to no-slip

boundary conditions

u=0 on Γ

g

⊂∂Ω, (2.4)

and natural (inﬂow/outﬂow) boundary conditions

n· ε(u) +pn=h on Γ

h

⊂∂Ω, (2.5)

where n denotes the unit outer normal on the domain boundary Γ

h

and h is

some external traction boundary force. The boundary subdomains Γ

g

and Γ

h

are assumed to be non-overlapping and to span the whole boundary ∂Ω.

For the numerical approximation of the equations of ﬂuid ﬂows, there are var-

ious methods available. The most popular ones are the ﬁnite volume method,

the ﬁnite difference method, and the ﬁnite element method. A comparative in-

troduction of these methods can be found in Quartapelle [74].

The algorithms developed in this thesis are based on the ﬁnite element

method (FEM). There is a vast literature discussing various aspects of the

ﬁnite element discretization of system (2.1)–(2.2). Mathematical aspects are

3

discussed, e.g., in Glowinski [28] and Gunzburger [35]. The book by Gresho

and Sani [32] focuses on practical issues for implementation such as boundary

conditions, discretization schemes, and stabilization. The work by Elman,

Silvester & Wathen [23] and Turek [88] put special emphasis on algorithmic

matters like efﬁcient numerical linear algebra for the ﬂow equations.

2.1 Weak form and spatial discretization

The Navier–Stokes equations (2.1)–(2.2) depend on both space and time. Most

discretization approaches split the approximation into separate approximations

for spatial and temporal part.

For setting up the spatial FE approximation, the ﬁrst step is to rewrite the sys-

tem of equations (2.1)–(2.2) in variational form. To this end, we deﬁne the space

of admissible velocity solutions at a certain time instant by V

u

=

_

u ∈ H

1

(Ω); u =

0 on Γ

w

_

, and the space of admissible pressure solutions by V

p

=L

2

(Ω). The vari-

ational problem corresponding to (2.1)–(2.2) is to ﬁnd, at each instant in time, a

pair

_

u(·, t ), p(·, t )

_

∈ V

u

×V

p

such that

_

v, ρ(u

t

+u· ∇u)

_

Ω

+

_

ε(v), µε(u)

_

Ω

−

_

∇· v, p

_

Ω

−

_

q, ∇· u

_

Ω

=

_

v, f

_

Ω

+

_

v, h

_

Γ

h

(2.6)

holds for all test functions

_

v, q

_

∈ V

u

×V

p

. The form

_

·, ·

_

Ω

denotes the standard

L

2

inner product.

The spatial discretization of system (2.6) replaces the (inﬁnite-dimensional)

solution function spaces V

u

and V

p

by ﬁnite-dimensional subspaces V

h

u

and V

h

p

.

Then, only the projectionof the Navier–Stokes equations ontoﬁnite dimensional

test function spaces V

h

u

×V

h

p

is considered, resulting in numerical approxima-

tions u

h

and p

h

. The discretizations considered in this thesis use a decompo-

sition of the computational domain into small subsets of characteristic size h,

called elements. We mainly focus on quadrilateral elements in 2D and hexahe-

dral (brick) elements in3D. The basis functions that spanV

u

and V

p

are chosento

be piecewise polynomials within the elements, and continuous over the whole

domain Ω. For the representation of the velocity u, equally shaped basis func-

tions are used for each spatial component. For instance, we will denote a ﬁnite

element approximationwithquadratic basis functions for velocity and linear ba-

sis functions for pressure on the elements by Q

d

2

Q

1

. On each element, the basis

functions for u are deﬁned as a tensor product of Lagrangian interpolation poly-

nomials of degree two in each coordinate direction [43], and the pressure basis

functions as tensor product of linear Lagrangian interpolants. Globally, the ba-

sis functions are nonzero in exactly one node, and zero on all the others. The

support of the basis functions is conﬁned to a patch of elements around the re-

spective nodes.

4

2.2 Stability of spatial discretization

For the spatial discretization to be stable, a compatibility condition between the

approximation of velocity and pressure must be satisﬁed, namely

inf

q∈V

h

p

sup

v∈V

h

u

_

q, ∇· v

_

Ω

_

_

q

_

_

L

2

_

_

v

_

_

H

1

≥β>0, (2.7)

where the constant βdoes not dependonthe meshsize parameter h. This condi-

tion was independently derived by Ladyzhenskaya [58], Babuška [4], and Brezzi

[12] and is called LBB condition or inf–sup condition. The condition ensures that

the divergence matrix discretizing the term

_

q, ∇· u

_

Ω

is of full rank, and hence

the discretized system has a unique solution for velocity and pressure.

The element combinationQ

d

2

Q

1

, the so-calledTaylor–Hoodelement pair, sat-

isﬁes the LBB condition [82], whereas the element combination Q

d

1

Q

1

does not.

We refer to Donea & Huerta [18] and Gresho & Sani [32] for an extensive dis-

cussion of LBB-stable elements. However, an element pair that does not sat-

isfy the LBB condition (2.7) can yet be modiﬁed so that stability of discretiza-

tion is obtained. Papers I and II investigate ﬂow problems with the element pair

Q

d

1

Q

1

. For stabilization, an artiﬁcial diffusion on the pressure space is intro-

duced in a systematic way, so that consistency is preserved. This approach [38],

[45], [84] modiﬁes the test functions in the discrete variational form (2.6), re-

sulting in a Petrov–Galerkin type scheme. Similar stabilization approaches have

been developed to avoid unphysical oscillations in velocity that otherwise ap-

pear in convection-dominated regimes for the standard central difference ap-

proximations introduced by the Galerkin FEM [18], [31], [32]. We refer to Braack

et al. [10] for a review of various commonly used stabilization techniques.

2.3 Time discretization

For discretization in time, it is usual to use some standard procedure developed

for ordinary differential equations and differential–algebraic equations, see

Donea & Huerta [18] and Hairer et al. [36], [37] for derivation and analysis of

numerical schemes. For schemes that treat both convection and diffusion

explicitly, linear stability theory sets a restrictive limit on the time step ∆t in

terms of the spatial mesh size ∆x, namely

∆t

max

∝

∆x

u

∞

+

ρ∆x

2

µ

,

see, e.g., the discussion in [52]. For applications where the viscosity is not too

small compared to the velocity, (semi-)implicit time stepping schemes are often

preferred in order to avoid the diffusive stability constraint. Popular schemes

are variants of the one-step theta method [51] and the BDF-2 method. For a

5

coupled solution approach, stability requires implicit time discretization of the

(differential–algebraic) terms

_

∇· v, p

_

Ω

and

_

q, ∇· u

_

Ω

. These terms enforce the

divergence-free condition on the velocity in Lagrangian multiplier form, see,

e.g., Gunzburger [35] and Lions [62].

Harlow and Welch [39] proposed to implement time propagation of

theNavier–Stokes equations by fractional time stepping strategy, and which was

then recast as a projection scheme by Chorin [15] and Temam[83]. This concept

has been mathematically justiﬁed for the ﬁnite element method by Guermond

and Quartapelle [33], [34]. Projection schemes ﬁrst advance velocities subject

to the momentum equation with pressure extrapolated from the old time step.

The resulting velocity is in general not divergence-free, so a pressure Poisson

equation is solved with forcing given by the divergence of the medium-step

velocity in a second step, in order to correct the velocity. Projection schemes

are to date the most popular variant for the numerical approximation of the

Navier–Stokes equations [74]. The subsystems resulting from fractional time

stepping are of easier structure than the original saddle-point system. However,

each of the subproblems needs to be equipped with suitable boundary

conditions for well-posedness, which are usually not the physical ones. This can

yield a non-consistent approximation close to the boundary. We refer to Gresho

& Sani [32] and Quartapelle [74] for discussion, and to the articles [13] and [52]

for recent developments.

The work presented in this thesis applies an alternative approach where the

system remains coupled. Coupled schemes solve for pressure and velocity as

one algebraic system, with boundary conditions set according to physical rea-

soning. However, the algebraic system is in general more challenging because of

its saddle point nature, and puts high demands on numerical linear algebra.

2.4 Numerical linear algebra for coupled discretizations

The introduction of spatial and temporal discretization schemes for the

Navier–Stokes system (2.6) results in a nonlinear system to be solved. For

resolving the non-linearity, ﬁxed point methods or Newton-type iterations are

usually applied, ﬁnally yielding a system of linear equations. The Lagrangian

basis functions are localized, so the ﬁnal coefﬁcient matrix is sparse. For the

solution of the saddle-point Navier–Stokes system, different techniques have

been proposed. Efﬁcient schemes employ multigrid methods with suitably

deﬁned smoothing schemes (see Trottenberg, Oosterlee & Schüller [85] and

John [48]), or use preconditioners based on the Schur complement of the block

2 ×2 matrix system deﬁned by velocity and pressure matrices (cf. [55], [53],

[22], see also the books by Turek [88] and Elman, Silvester & Wathen [23]).

Papers I and II use an algebraic multigrid scheme for the coupled system, and

the numerical method used in Papers III and IV is based on a preconditioner

constructed from a Schur complement.

6

3. Computational methods for

turbulence

This chapter discusses basics of the simulation of turbulent ﬂow for a single-

phase ﬂuid, which is the context of Papers I and II. The modeling equations

are the Navier–Stokes equations (2.1)–(2.2), where the viscous stresses are small

compared to the inertial stresses. The ratio between inertial and viscous stresses

is measured by the so-called Reynolds number,

Re =

ρl u

µ

, (3.1)

where l is a reference length and u a reference velocity. A situation where inertial

stresses dominate is characterized by high Reynolds numbers.

3.1 Basics on turbulence and large eddy simulation

A physical characterization of ﬂow is to distinguish vortex structures by their

size relative to the ﬂow domain. A ﬂow ﬁeld consists of vortices of different size.

These vortices are instationary and ﬂuctuating — a phenomenon called turbu-

lence. Fig. 3.1 depicts a typical pressure ﬁeld ina turbulent ﬂow. Vortex structures

of different sizes as well as the irregular pattern of the ﬂow can clearly be seen.

For ﬂow at high Reynolds numbers, instabilities lead to a breakup of larger

structures into smaller ones. This induces an energy transport fromlarger scales

to smaller scales. The breakup process continues to create smaller and smaller

structures until the scale is sufﬁciently small for viscous dissipation to become

dominant, see Dubois et al. [19], Pope [73], and Sagaut [77]. In the Navier–Stokes

equations, the interaction between different scales is driven by the nonlinear

convective term u· ∇u.

The transfer of energy that takes place inthe model of ﬂuidmotionis schemat-

ically depicted as a Kolmogorov scale spectrumin Figure 3.2. The diagramshows

the region of active scales in a ﬂow, ranging fromthe energy-containing scales l

0

(scale range of external forcing f and h) to the regime where viscosity becomes

dominant at a scale level η, the so-called Kolmogorov scale [73]. The Kolmogorov

scale η depends on the Reynolds number based on the production length scale

l

0

as

η

l

0

∝Re

−3/4

l

0

. (3.2)

7

Figure 3.1: Pressure iso-surfaces for turbulent ﬂow around a square obstacle (cf. Paper

II).

Kolmogorov [56] noticed a self-similarity between the eddies in the inertial

regime, as opposed to large-scale eddies, which depend mainly on geometry.

Kolmogorov formulated the hypothesis that the energy content at scales in the

inertial subrange, i.e., the range between energy production and viscous ﬂow, is

proportional to κ

−5/3

, where κ is the scale length.

All the scales in incompressible ﬂows are coupled with one another. Hence, a

numerical method approximating the Navier–Stokes equations (2.1)–(2.2) needs

to resolve all the scales in the ﬂow in order to represent any scale correctly. A nu-

merical approach that aims at resolving all necessary scales is called a direct nu-

merical simulation (DNS) [63]. For three-dimensional instationary ﬂow, a DNS

requires to resolve a scale range according to (3.2) in each spatial direction as

well as in the temporal dimension, which amounts in a number of degrees of

freedom that is proportional to Re

3

, often even with a rather large proportional-

ity constant. Many technically relevant ﬂows are describedby Reynolds numbers

within the range 10

4

−10

6

, so the requirements in computational effort exceed

the capabilities of state-of-the-art schemes for a DNS by far, necessitating other

techniques.

Since a full resolution of all ﬂow features is not feasible, at least a fraction of

these features needs to be modeled. The so-called large eddy simulation (LES) is

one such modeling strategy. It aims at resolving the larger scales (large eddies),

and to model the effect of unresolved scales onthe large scales, see Pope [73] and

Sagaut [77]. This idea is depictedinFig. 3.2. The large scales inthe range between

8

Figure 3.2: Kolmogorov energy spectrum in a double logarithmic plot. The wave length

κ

0

denotes the resolution limit of the discretization. All the scales beyond that level will

not be captured by the numerics and their effect needs to be modeled in order to cor-

rectly describe the physics of turbulence.

l

0

and κ

0

are resolved by the numerical discretization, whereas all scales beyond

κ

0

are not resolved and need to be included by some modeling.

Asimple, yet very powerful approachis the Smagorinsky model [80]. It is based

on the observation that the insufﬁcient resolution within a LES leads to under-

resolved viscous effects. Hence, Smagorinsky introduced an artiﬁcial subgrid or

eddy viscosity µ

T

with the aims of representing the effect of viscosity on the level

of the computational mesh size. The eddy viscosity is deﬁned through a non-

linear coefﬁcient, based on the magnitude of the viscous stress tensor ε(u), the

square of the mesh size, and a proportionality constant [77].

3.2 A multiscale approach to large eddy simulation

The deﬁnition of a LES needs to distinguish between resolved and unresolved

scales in order to formulate the subgrid viscosity term. Traditional methods use

low-pass ﬁlters to extract the large-scale information from the Navier–Stokes

equations and introduce the subgrid viscosity [77]. However, this spatial ﬁlter-

ing is problematic close to domain boundaries. These problems can be traced

back to so-called commutation errors [7], [20]. Moreover, ﬁnding appropriate

boundary conditions for the averaged large scales is to date an open problem.

9

An alternative approach to LES is based on the variational multiscale

method [44], introduced by Hughes, Mazzei, and Jansen [46]. The ideas are

closely related to the concept of dynamic multilevel methods, described in

Dubois, Jauberteau & Temam [19], and the multiscale turbulence analysis in

Collis [16]. In this framework, ﬁlters are replaced by variational projections

onto appropriate subspaces. The space for admissible velocity solutions is

decomposed into

V

u

=V

3h

u

⊗V

δh

u

⊗

¯

V

u

, (3.3)

where V

3h

u

is a space associated with mesh size 3h, representing large resolved

scales, V

δh

u

is the space spanned by V

h

u

except V

3h

u

, representing small resolved

scales, and ﬁnally, the space of unresolved scales

¯

V

u

. We choose a resolution of

3h, however, different deﬁnitions of large resolved scales are common as well,

see the reviewarticle by Gravemeier [30]. Applying an equivalent decomposition

for pressure, we get (following Harten’s notation [40])

u=u

3h

+δu

h

. ¸¸ .

u

h

+ˆ u, p =p

h

+ ˆ p. (3.4)

Let us denote by

B

NS

_

v, q; u, p

_

=

_

v, ρ(u

t

+u· ∇u)

_

Ω

+

_

ε(v), µε(u)

_

Ω

−

_

∇· v, p

_

Ω

−

_

q, ∇· u

_

Ω

the form that contains all the left-hand-side terms of the weak form of the

Navier–Stokes equations (2.6). The aim of any numerical method is to ﬁnd

numerical approximations u

h

, p

h

in the resolved subspace V

h

u

⊗V

p

according to

the direct sum decomposition (3.3). A systematic way of doing so is to insert the

decomposition (3.4) into the weak form (2.6), and to restrict the test functions

to the space of resolved variables. Then, the nonlinear convective term is

linearized and resorted in terms of resolved and unresolved quantities [46].

Hence, we search for

_

u

h

, p

h

_

, such that

B

NS

_

v

h

, q

h

; u

h

, p

h

_

=

_

v

h

, f

_

Ω

+

_

v

h

, h

_

Γ

h

−B

1

NS

_

v

h

, q

h

; u

h

; ˆ u, ˆ p

_

−B

2

NS

_

v

h

, ˆ u

_

(3.5)

holds for all test functions

_

v

h

, q

h

_

∈ V

h

u

×V

h

p

. The termB

1

NS

_

v

h

, q

h

; u

h

; ˆ u, ˆ p

_

con-

tains linear terms in( ˆ u, ˆ p) andthe linearizedconvectionaroundu

h

, andthe term

B

2

NS

_

v

h

, ˆ u

_

contains the quadratic convection term. For the numerical approxi-

mation to be computable, these two terms need to be substituted by some mod-

eling that eliminates the unresolved variables.

The variational multiscale large eddy simulation with a subgrid viscosity

model constructs the approximation as

B

1

NS

_

v

3h

, q

3h

; u

h

; ˆ u, ˆ p

_

+B

2

NS

_

v

3h

, ˆ u

_

≈0,

B

1

NS

_

δv

h

, δq

h

; u

h

; ˆ u, ˆ p

_

+B

2

NS

_

δv

h

, ˆ u

_

≈

_

ε

_

δv

h

_

, 2ν

T

ε

_

δu

h

_

_

Ω

.

(3.6)

10

Note that the subgrid viscosity is only applied to the smaller of the resolved

scales, as discussed, e.g., in the review articles by Gravemeier [30] and John [49].

A physical explanation for application of eddy viscosity only on small resolved

scales is that energy transport occurs mainly on neighboring scale groups, so

that the effect of a not adequately resolved viscosity is best modeled on the ﬁne

resolved scales [66].

The implementation of a method according to the framework (3.5) with ap-

proximation (3.6) needs to distinguish between coarse and ﬁne resolved scales.

Standard methods for doing this are based on different polynomial orders for

the FE basis functions, or two different grid levels as used, e.g., in geometric

multigrid. Such methods have been presented by John & Kaya [50] and Grave-

meier [29]. Papers I and II discuss an implementation where this separation is

performed on the algebraic equation system by using routines from algebraic

multigrid.

11

4. Simulation of Two-Phase Flow

We consider the dynamics of two immiscible incompressible ﬂuids separated by

an interface Γ as displayed in Fig. 4.1. By Ω

1

we denote the region occupied by

the ﬁrst ﬂuid, and by Ω

2

the region occupied by the second. We assume that the

ﬂow is laminar, i.e., the Reynolds number is low for the considerations in this

chapter.

Ω

1

Ω

2

Γ

∂Ω

Figure 4.1: Schematic viewof a two-phase ﬂowproblem. Fluid 1 occupies region Ω

1

, and

ﬂuid 2 occupies Ω

2

. The interface Γ separates the two ﬂuids.

The motion of each ﬂuid is described by the Navier–Stokes equations (2.1)–

(2.2) using the variables u

i

and p

i

, i =1, 2. Extra conditions are necessary to de-

scribe the behavior on the interface Γ. A widespread formulation is to pose the

two-phase ﬂow problem on the whole domain Ω=Ω

1

∪Γ∪Ω

2

with variables u

and p that take the respective velocity and pressure values in the individual do-

mains. Generally, the ﬂuid densities ρ

i

and viscosities µ

i

are different for the two

ﬂuids. Hence, the material parameters ρ and µ in the global momentum equa-

tion have a discontinuity on the interface. In addition, interfacial tension enters

the formulation as a forcing of the form

f

st

=σκnδ

Γ

. (4.1)

Here, σ is a constant specifying the relative strength of the surface tension, κ

denotes the interface curvature, nthe directionnormal to the interface (pointing

into Ω

2

), and δ

Γ

is a delta function that localizes the force to the interface. The

interface Γ is transported by the local ﬂuid motion u|

Γ

.

13

The variables in a single-domain formulation comprise kinks and discontinu-

ities due to changes in material parameters and forces: The discontinuous vis-

cosity introduces a discontinuity in the symmetric gradient ε(u). Similarly, due

to the distributional character of the delta function δ

Γ

there is a jump in pressure

proportional to σκ. Any change in curvature tangential to the interface gives rise

to a discontinuity in the pressure derivative normal to the interface as well as the

viscous stress tensor. The conditions describing the irregularities in velocity and

pressure over the interface Γ are called jump conditions, see Edwards, Brenner &

Wasan [21] and Lee & LeVeque [59].

Because of different ﬂuid densities, gravitational forces are often essential for

the dynamics of two-phase ﬂow. The gravity force is modeled as

f

g

=ρge

g

, (4.2)

where g is the gravitational acceleration and e

g

denotes the direction of gravity.

The total forcing in the momentum equation (2.1) for two-phase ﬂow is hence

given by f =f

st

+f

g

.

4.1 Overview of numerical models

The two main challenges in the numerical simulation of two-phase ﬂow are the

representation of the interface as it evolves in time, and the evaluation of the

distributional force term (4.1). Several discrete approaches have been proposed

during the last decades. There are two main strategies to couple the interface

evolution problem to the Navier–Stokes equations discretized on a ﬁxed grid as

discussed in Chapter 2.

The ﬁrst strategy is to explicitly track the interface position by introducing an

additional discrete set of marker points or marker elements that describe the

interface. These methods are called front tracking and were introduced in dif-

ferent variants by Peskin [71] and Glimm et al. [27], and reviewed in the articles

[78], [86], and [89]. A special class of front tracking schemes are the immersed

boundary method by Peskin (see Peskin’s [72] reviewpaper) and a more accurate

variant, the immersed interface method, by LeVeque and co-workers [59], [60],

[61]. The evolution of the interface in front tracking schemes is accomplished by

Lagrangian advection of the marker points, and the surface tension force is gen-

erally evaluated based on a discrete approximation of the delta function around

the interface. Normal vectors and curvatures can be calculated using the in-

terface points by straight-forward geometric identities. Explicit descriptions are

very powerful because they allow to include general models of the interface, like

elastic forces in immersed boundary and interface implementation [72]. How-

ever, the methods are quite difﬁcult to implement because the marker points

need to interact with the ﬂuid grid in order to describe advection and interface

forces. Straight-forward implementations are also prone to unphysical changes

of the area/volume of the respective ﬂuid. To retain an accurate interface repre-

14

sentation, the marker points need to be redistributed when the interface is de-

formed.

The second strategy is front-capturing, where the interface is implicitly de-

ﬁned. Historically, the ﬁrst scheme of this kind was the marker-and-cell method

proposed by Harlow and Welch [39], and can be considered to be a volume-

tracking method. Here, one ﬂuid is colored by marker particles whose location

is advected with the ﬂow ﬁeld. The position of the interface can then be recon-

structed from this particle ﬁeld. An extension of this approach is to replace the

marker particles by a marker function. This approach has been pursued by Noh

& Woodward [65] and Hirt & Nichols [42], who introduced the volume-of-ﬂuid

(VOF) method. VOF includes anadditional variable that stores the fractionof the

ﬁrst ﬂuid on the total ﬂuid for each cell. From the volume fractions, the position

of the interface can be reconstructed, e.g., by deﬁning line segments. For details

of the reconstruction, including higher order schemes, we refer to Scardovelli

& Zaleski [78] and Sethian [79], and references therein. The advection of the

interface is usually implemented by increasing or decreasing the volume frac-

tion depending on the composition of neighboring cells and the velocity ﬁeld.

To put surface tension force in VOF into practice, several different approaches

are common. One is to use discrete delta functions in combination with the re-

constructed interface, and another employs the continuous surface tension ap-

proach by Brackbill, Kothe, and Zemach [11], see also the review by Tryggvason

et al. [86]. The advantage of VOF methods is the volume conservation and a nat-

ural mechanismfor breakup and fusion of bubbles. However, the reconstruction

of the interface and the interface motion are considerably more complicated to

implement here than for front tracking methods.

While VOF methods operate on the discrete level, a continuous front cap-

turing framework is provided by the level set method, introduced by Osher &

Sethian [70] and ﬁrst applied to two-phase incompressible ﬂow by Sussman,

Smereka &Osher [81]. Yet another methodis providedby the phase ﬁeldmethod,

formulated by Jacqmin [47] for the simulation of two-phase ﬂow. The methods

discussed in Paper III and IV are based on the latter two approaches, which are

discussed in detail in the following two sections.

4.2 Level set models

The basic idea of level set methods is to deﬁne the interface implicitly as the

zero contour line of a function φthat is deﬁned in Ω. Level set methods allowfor

a straight-forward evaluation of normal vectors by setting n(φ) = ∇φ/|∇φ|, and

curvature, κ(φ) =∇· n(φ). Both these quantities can be computed locally fromφ

on the same computational mesh as the equations of ﬂuid ﬂow. The mechanism

for moving the interface in level set methods is advection of the function φ with

local ﬂuid velocity, i.e.,

φ

t

+u· ∇φ=0. (4.3)

15

This is an additional partial differential equation coupled to the Navier–Stokes

equations. The advection equation (4.3) can be solved with standard tools for

hyperbolic transport equations, like upwind ﬁnite difference/volume schemes

or stabilized ﬁnite element methods. The advection can be implemented so that

the volume is conserved, since the velocity is divergence-free, i.e., ∇· u=0.

d

d

d

d

d

d

d

interface φ=0

d

ds

φ

)

Ω

1

Ω

2

Ω

2

φ<0 φ>0 φ<0

(a) Signed distance function

¯

φ

A

interface φ=0

d

ds

Ω

2

Ω

1

Ω

2

φ<0 φ>0 φ<0

(b) Smoothed color function

Figure 4.2: Representing an interface by level set functions φ.

The most popular choice for a level set functionis the signeddistance function

φ, depicted in Fig. 4.2(a). We refer to the books by Sethian [79] as well as Osher

and Fedkiw [69] for an extensive discussion of these methods. A signed distance

function encodes the distance of a particular point to the interface, and different

signs identify the two ﬂuids. A signed distance function has the property |∇φ| =

1 almost everywhere. A beneﬁt of this choice is that the integral in the surface

tension force can be simpliﬁed to a one-dimensional discrete delta function as

_

Γ

δ

_

x−y(s)

_

ds ≈δ

h

_

φ(x)

_

,

where x denotes an arbitrary point in the computational domain and y(s) is the

image of a parametrization of the interface, see, e.g., [79]. However, it has been

shown by Engquist, Tornberg, and Tsai [24] that such an approximation can lead

to O(1) errors in the force unless the width h of the delta function is varied. Re-

cent implementations therefore reconstruct the interface position from φ and

apply a d-dimensional discrete delta function that does not show these consis-

tency problems. Since the information fromφ is only needed in a region around

the interface, so-called narrow-band level set implementations are widely used.

These methods calculate φ only on a few grid cells around the interface, often in

combination with high-resolution grids. We refer to Adalsteinsson and Sethian

[1] for details. To start a level set calculation, an initial proﬁle φ(·, 0) needs to

generated given a certain interface position. Moreover, the ﬂow ﬁeld as well as

inaccuracies in the numerical scheme deform the signed distance function in

the course of the simulation. Therefore, algorithms have been proposed to (re-

16

)initialize the signed distance function. These algorithms enforce the property

|∇φ| =1, for example, by solving the partial differential equation

φ

τ

+S

_

φ

0

__

|∇φ| −1

_

=0, (4.4)

to steady state, where S(φ

0

) is a sign function with value 1 in the ﬁrst ﬂuid and

value −1 in the other (see, e.g. [69]). However, discretizations of the reinitializa-

tion scheme do not preserve mass, so that the overall level set implementation

based on signed distance functions may suffer fromunphysical volume changes

in the ﬂuid phases.

Another choice is to use a smoothed color/indicator function as depicted in

Fig. 4.2(b). Such a function is (almost) constantly 1 or −1 away from the inter-

faces, with positive value in one region and negative value in the other. Around

the interface, there is a transition region where the function smoothly switches

from one value to the other. This function φ can be calculated from the reinitial-

ization equation (cf. Olsson and co-workers [67], [68])

φ

τ

+∇·

_

n

_

1−φ

2

__

−∇·

_

n∇φ· n

_

=0. (4.5)

In this equation, diffusion in direction normal to the interface is balanced by

a compressive ﬂux. The parameter is typically chosen to be of similar size as

the computational mesh. In one spatial dimension with an interface located at

x =α, the steady-state value of (4.5) is given by

φ(x) =tanh

_

x −α

_

. (4.6)

Since the reinitialization equation (4.5) is posed as a conservation law, a conser-

vative implementation of the method is possible [68]. A smoothed color func-

tion also avoids the introduction of an additional discrete delta function and the

problems associated with that procedure, since it approximately holds

∇φ≈2nδ

Γ

. (4.7)

The factor two comes fromthe fact that the functionswitches from−1 to +1. This

scheme for evaluating the interface force corresponds to the continuous surface

tension model proposed by Brackbill, Kothe, and Zemach [11]. A disadvantage

of the formulation with a smoothed color function is that the evaluation of cur-

vature is potentially less accurate because it is calculated from a steep proﬁle.

This generally increases resolution requirements compared to the signed dis-

tance function approach. Another drawback of the smoothed indicator function

is that the function needs to be deﬁned on the whole domain in order to con-

serve volume. This can be alleviated by the use of adaptive meshes that have low

resolution away from the interface, though. Moreover, the discrete delta func-

tion underlying (4.6) has support on the whole domain Ω, which necessitates

evaluating the discrete version of the force (4.1) everywhere.

17

4.3 Phase ﬁeld models

All the methods discussed above are based on the mathematical model (4.1) for

interfacial tension, assuming a continuous PDE model for ﬂuid dynamics that

does not take processes on the level of atoms and molecules into account. The

numerical approximations aim at discretely reproducing the effect of a delta

function, and evaluating the curvature to determine the strength of the surface

tension. A different approach takes chemical considerations of the interface into

account. The so-called van der Waals hypothesis [90] states that immiscible ﬂu-

ids actually mix on molecular level, forming a diffuse interface. The proﬁle of the

interface is given by a balance of random molecular motion and molecular at-

traction, as devised by Cahn and Hilliard [14]. The considerations are based on

the free energy of a molecule, given by

f (C) =βΨ(C) +

α

2

|∇C|

2

, (4.8)

where C denotes the concentration and Ψ(C) = (C +1)

2

(C −1)

2

. Similar to the

smoothed color function (4.6), the function seeks to attain the two equilibrium

values C = ±1, which is driven by the double-well potential Ψ(C). The gradient

term ∇C on the other hand enforces a smooth transition. The system seeks to

minimize the energy (4.8). This process is described by the Cahn–Hilliard equa-

tion

C

t

=∇·

_

M∇ψ

_

=∇·

_

M∇

_

βΨ

(C) −α∇

2

C

__

=0, (4.9)

with the chemical potential ψand the mobility factor M.

The phase ﬁeld method, formulated by Anderson et al. [2] and Jacqmin [47],

includes transport of the interface by convection, i.e., a term u· ∇C, and formu-

lates an interface force

κnδ

Γ

≈−C∇ψ+∇P, (4.10)

where P =Cψ. Equation (4.10) can be rewritten as

κnδ

Γ

≈ψ∇C,

which is the same form as for the continuous surface tension model, that is,

equation (4.7) multiplied by curvature κ. Boyer [9] extended the phase ﬁeld

model to describe motion of ﬂuids with different densities and viscosities.

The phase ﬁeld model can be augmented in order to include contact forces of

the interface with solid walls [47], which results in ﬂux-boundary conditions on

the chemical potential ψ. This approach allows for setting static contact angles

that correspond to physical wetting properties. The direct inclusion of contact

line dynamics is different from the other approaches in computational multi-

phase dynamics, where additional mechanisms as slip-velocities and near-wall

interface diffusion need to be added, cf. [76], [92].

18

A considerable challenge in the simulation of phase ﬁeld models are the high

resolution requirements. Even though the interface width α/β can be chosen

considerably larger than molecular dynamics would suggest (see Villanueva and

Amberg [91]), accurate results require sophisticated numerical approximations

and ﬁne computational meshes. Furthermore, the equations are nonlinear and

contain fourth order derivatives. Several attempts have been made to make this

approach more competitive, see for example the work by Kay and Welford [54].

19

5. Summary of Papers

5.1 Paper I

Paper I proposes a particular implementationfor the scale separationinthe vari-

ational multiscale large eddy simulation. Our method uses a plain aggregation

scheme that is part of an aggregation-based algebraic multigrid implementation

[26], [87]. Unlike for geometric multigrid approaches that have been used ear-

lier, no grid hierarchy is needed in our implementation. This makes the method

also applicable for discretizations based on fully unstructured grids. The subgrid

viscosity term is implemented within a parallel ﬂuid ﬂow solver based on linear

interpolation ﬁnite elements and uses Trilinos solver algorithms [41]. The focus

of the method is the improvement of LES at low resolutions.

The method is tested for some standard sample problems for turbulent ﬂow.

A grid reﬁnement study is performed for a turbulent channel ﬂow with DNS

data as reference [64]. This study shows that our implementation of the varia-

tional multiscale LES model provides better results than other state-of-the-art

LES schemes.

5.2 Paper II

Paper II extends and analyzes the method presented in Paper I. In this paper it

is proposed to use a second-order accurate generalized-alpha time integration

scheme (instead of the Crank–Nicholson scheme in Paper I). The generalized-

alpha scheme is a variant of the one-step theta schemes that includes approxi-

mations of the ﬁrst temporal derivative as variables and enables control of the

dissipation for high frequencies, which is favorable in low-resolution computa-

tions.

The analysis considers details of the subgrid viscosity implementation based

on plain aggregation scale separation, showing that the additional model

does not affect consistency of the overall Navier–Stokes scheme. Moreover,

a clear difference between projective and non-projective scale separations

is identiﬁed, as was previously observed in numerical computations [29].

Projective implementations retain low frequencies, whereas non-projective

implementations also damp low frequencies. We apply our method to turbulent

ﬂow around a square cylinder, where the results are better than those of other

LES schemes at low resolutions, even though a high sensitivity of the results on

stabilization parameters is observed.

21

5.3 Paper III

The paper is based on the observation that level set methods enable efﬁcient

implementations of two-phase ﬂow physics at relatively coarse meshes because

they are based on simple advection schemes together with a reinitialization.

However, a standard level set method fails in predicting phenomena relevant in

wetting where the ﬂuid-ﬂuid interface aligns at some oblique angle with solid

walls. Such a simulation is, however, possible with a phase ﬁeld model.

As the discussion in Section 4.3 suggests, the conservative level set method

[68] and the Cahn–Hilliard model have a similar equation structure. Therefore,

the paper formulates the Cahn–Hilliard-speciﬁc terms as an extension of the

level set model, but restricts that additional phase description to a region close

to contact points. A continuous switch function is used in order to retain a con-

sistent model. The model is discretized with the ﬁnite element method and im-

plemented based on the deal.II library [5]. A test on a channel ﬂow driven by

inﬂow illustrates the beneﬁcial effects of the equation coupling.

5.4 Paper IV

This paper analyzes inaccuracies of the numerical implementation of the con-

servative level set model [68]. A circular bubble at rest is considered as a model

problem. In this case the force due to surface tension should ideally be balanced

by a jump in pressure. The study is performed by applying a complete Navier–

Stokes/two-phase ﬂow solver to the problem and recording spurious velocities,

see, e.g. [3], [25].

We observe that the velocity errors can be traced back to inaccuracies in the

numerically calculated curvature that enter the continuous surface tension

model. If the exact curvature is prescribed, there are no imbalances in the

scheme, in case equal order elements are chosen for approximating pressure

and the level set function. Choosing the element order of the level set function

higher than the one for pressure introduces an additional imbalance. However,

the use of higher order elements for the level set function has beneﬁcial impact

on the accuracy of curvature, so that the numerical model is still more accurate.

Moreover, we study the effect of the continuous representation of surface

tension on the prediction of the jump in pressure between the inside and the

outside of the bubble. This showed that the variation of the curvature along the

ﬁnite width of the interface has the largest impact on pressure accuracy.

We perform the tests using two different ﬂow solvers, in order to further eval-

uate the accuracy of the force balance for ﬁnite element discretizations. The re-

sults for a fully coupled implicit method and a decoupled (projection) solver

are similar, which shows that choice of the solver does not affect the balance

of forces.

22

6. Future Work

The method proposed in Paper III aims at providing phase-ﬁeld-like features

without having to use all its terms everywhere. This is done so that the phase

ﬁeld method is reduced to a level set formulation in regions far away from con-

tact lines. It will be necessary to perform an in-depth convergence study of level

set models and phase ﬁeld models on several test cases in two-phase ﬂow dy-

namics. In particular, the resolution requirements of the two methods should be

quantiﬁed in order to identify the precise beneﬁts of the hybrid method. In our

hybrid model, the equations have been coupled by a smoothly varying switch

function. The inﬂuence of the switch function needs to be considered in an en-

ergy estimate in order to ensure stability of the coupling of the equations, which

has up to date only been veriﬁed numerically. Since the numerical implementa-

tion of the equations is based on a system of two equations to avoid the direct

appearance of fourth derivatives, the analysis needs to be performed in a way to

correctly reﬂect the discretized state of the equations.

We also aim at using our models for larger problems, especially three-dimen-

sional simulations. For this purpose, parallel implementations of the ﬂow solver

and the coupled level-set/Navier–Stokes systemare necessary. Within the deal.II

software that was used in Papers III and IV, it is possible to use parallel assembly

and linear algebra routines. For a Stokes problemcoupled to an advection equa-

tion, a parallel version for up to about 50 processors has already been imple-

mented by the author of this thesis.

1

The objective is to apply this implementa-

tion framework to the coupled Navier–Stokes/level-set system and to eventually

extend it to a massively parallel implementation for even larger systems.

The hybrid level-set/phase-ﬁeld method is a potential tool for improving the

simulation of multi-phase ﬂowin porous media [17]. A ﬁrst step in that direction

is to apply the new model to more complicated geometries, like, e.g., channels

with oscillating walls or ﬂow around small obstacles. Small scale results are a

helpful tool in the simulation of subsurface ﬂow like the prediction of the ﬂow

in oil and gas reservoirs. The imposition of a pressure gradient on a small-scale

material conﬁguration induces a ﬂow ﬁeld, which can be used to determine val-

ues for the permeability of that conﬁguration. The permeability tensor is one of

the main input parameter in coarse-scale simulations based on Darcy’s law.

In Paper IV, a study regarding error sources in the discrete approximation with

the conservative level set implementation is presented. These results point out

directions where the methodneeds tobe improved. One suchdirectionconcerns

1

See the deal.II tutorial program http://www.dealii.org/developer/doxygen/deal.II/step_32.html

23

the accuracy in the curvature calculations. A particularly easy approach is to use

higher order ﬁnite elements for representing the level set function. When keep-

ing the interface width constant, increasing the order of approximation consid-

erably increases resolution of the interface, and thus, accuracy of normals and

curvature that rely on the interface representation. This improvement should

be advantageous for the overall accuracy, even though a small imbalance in the

force representation is introduced. Such a strategy increases the costs for the

level set portion of the algorithm if the same mesh is used as for the Navier–

Stokes equations, though. An alternative would be to coarsen the level set mesh

in regions away from the interface. This is reasonable since in these remote re-

gions, only the momentum and continuity balances need to be represented ac-

curately, while the level set equation does not contain any additional informa-

tion. Similarly, the interface region can be resolved by more mesh points than

there are used for the Navier–Stokes equations. However, this requires a special

implementation for the evaluation of the surface tension force in the ﬁnite ele-

ment algorithms because the variables are related to different meshes.

An alternative hybrid approach that might be taken into account is to not mix

the level set formulationwiththe phase ﬁeldformulationas has beendone inPa-

per III, but to use the phase ﬁeld model as a micro-scale input to a macro-scale

formulation based on a level set description of two-phase ﬂow. A heterogeneous

multiscale approach has been proposed by Ren and E [75] for modeling the in-

teraction of macro-scale ﬂuid ﬂow by micro-scale information from molecular

dynamics. The response on molecular level provides information about the ex-

pected motion of contact lines during one time step. This coupled model be-

haves differently than large-scale level set model, which would assume zero ve-

locity. It is conceivable that a similar interaction could be applied for a micro-

scale simulation done with the phase ﬁeld model. The difﬁculty of such a micro-

macro coupling is to ﬁnd appropriate transfer operations from the micro scales

to the macro scales. Suitable boundary conditions for the Cahn–Hilliard sub-

problem given the solution of a macro-scale level set model are needed, as well

as a shear velocity for the macro-scale Navier–Stokes equations, based on the

results from a micro model.

24

Acknowledgments

I am very grateful to my adviser Gunilla Kreiss for introducing me to this inter-

esting subject and for stimulating discussions on various aspects of multiphase

ﬂow. My thanks go to Volker Gravemeier who acquainted me with the subject of

variational multiscale large eddy simulation for turbulent ﬂow and stimulated

ideas. I acknowledge discussions with Wolfgang Bangerth on implementation

aspects of ﬁnite element programming. Finally, I would like to thank Katharina

Kormann and Eddie Wadbro for countless discussions on computational math-

ematics as well as proof-reading manuscripts of my articles and this thesis.

This work was supported by the Graduate School in Mathematics and Com-

puting (FMB).

25

Bibliography

[1] D. Adalsteinsson and J. A. Sethian. A Fast Level Set Method for Propagating Inter-

faces. J. Comput. Phys., 118(2):269–277, 1995.

[2] D. M. Anderson, G. B. McFadden, and A. A. Wheeler. Diffuse-Interface Methods in

Fluid Mechanics. Annu. Rev. Fluid Mech., 30:139–165, 1998.

[3] E. Aulisa, S. Manservisi, and R. Scardovelli. A novel representation of the surface

tension force for two-phase ﬂowwith reduced spurious currents. Comput. Methods

Appl. Mech. Engrg., 195:6239–6257, 2006.

[4] I. Babuška. Error-bounds for ﬁnite element method. Numer. Math., 16(4):322–333,

1971.

[5] W. Bangerth, R. Hartmann, and Kanschat G. deal.II — a General Purpose Object

OrientedFinite Element Library. ACMTrans. Math. Softw., 33(4):article no. 24, 2007.

[6] Y. Bazilevs, V. M. Calo, J. A. Cottrell, T. J. R. Hughes, A. Reali, and G. Scovazzi. Varia-

tional multiscale residual-based turbulence modeling for large eddy simulation of

incompressible ﬂows. Comput. Methods Appl. Mech. Engrg., 197:173–201, 2007.

[7] L. Berselli, C. Grisanti, and V. John. On Communtation Errors in the Derivation

of Space Averaged Navier–Stokes Equations. Technical Report 12, Dipartimento di

Matematica Applicata, Università di Pisa, 2004.

[8] R. P. Beyer andR. J. Leveque. Analysis of a one-dimensional model for the immersed

boundary method. SIAM J. Numer. Anal., 29(2):332–364, 1992.

[9] F. Boyer. Atheoretical and numerical model for the study of incompressible mixture

ﬂows. Comput. & Fluids, 31:41–68, 2002.

[10] M. Braack, E. Burman, V. John, and G. Lube. Stabilized ﬁnite element methods for

the generalized Oseen problem. Comput. Methods Appl. Mech. Engrg., 196:853–866,

2007.

[11] J. U. Brackbill, D. B. Kothe, and C. Zemach. A Continuum Method for Modeling

Surface Tension. J. Comput. Phys., 100:335–354, 1992.

[12] F. Brezzi. On the existence, uniqueness and approximation of saddle-point prob-

lems arising from Lagrangian multipliers. Rev. Française Automat. Informat.

Recherche Opérationelle Sér. Rouge, 8(R-2):129–151, 1974.

[13] D. L. Brown, R. Cortez, and M. L. Minion. Accurate Projection Methods for the In-

compressible Navier–Stokes Equations. J. Comput. Phys., 168(2):464–499, 2001.

27

[14] J. W. Cahn and J. E. Hilliard. Free Energy of a NonuniformSystem. I. Interfacial Free

Energy. J. Chem. Phys., 28(2):258–267, 1958.

[15] A. J. Chorin. Numerical Solution of the Navier–Stokes Equations. Math. Comput.,

22:745–762, 1968.

[16] S. S. Collis. Monitoring unresolved scales in multiscale turbulence modeling. Phys.

Fluids, 13:1800–1806, 2001.

[17] C. Crowe. Multiphase Flow Handbook. CRS Press, Boca Raton, 2006.

[18] J. Donea and A. Huerta. Finite Element Methods for Flow Problems. J. Wiley & Sons,

Chichester, 2003.

[19] T. Dubois, F. Jauberteau, and R. Temam. Dynamic Multilevel Methods and the Nu-

merical Simulation of Turbulence. Cambridge University Press, Cambridge, 1999.

[20] A. Dunca, V. John, and W. J. Layton. The Commutation Error of the Space Averaged

Navier–Stokes Equations on a Bounded Domain. In G. P. Galdi, J. G. Heywood, and

R. Rannacher, editors, Contributions to Current Challenges in Mathematical Fluid

Mechanics, volume 3, pages 53–78. Birkhäuser Verlag, Basel, 2004.

[21] D. A. Edwards, H. Brenner, and D. T. Wasan. Interfacial Transport Processes and

Rheology. Butterworth–Heinemann, Stoneham, 1991.

[22] H. Elman, D. Silvester, and A. Wathen. Performance and analysis of saddle point

preconditioners for the discrete steady-state Navier–Stokes equations. Numer.

Math., 90(4):665–688, 2002.

[23] H. Elman, D. Silvester, and A. Wathen. Finite Elements and Fast Iterative Solvers

with Applications in Incompressible Fluid Dynamics. Oxford Science Publications,

Oxford, 2005.

[24] B. Engquist, A.-K. Tornberg, and R. Tsai. Discretization of Dirac delta functions in

level set methods. J. Comput. Phys., 207:28–51, 2005.

[25] M. M. Francois, S. J. Cummins, E. D. Dendy, D. B. Kothe, J. M. Sicilian, and M. W.

Williams. A balanced-force algorithm for continuous and sharp interface surface

tension models within a volume tracking framework. J. Comput. Phys., 213:141–

173, 2006.

[26] M. W. Gee, C. M. Siefert, J. J. Hu, R. S. Tuminaro, and M. G. Sala. ML 5.0 Smoothed

Aggregation User’s Guide. Technical Report 2006-2649, Sandia National Laborato-

ries, 2006.

[27] J. Glimm, J. Grove, B. Lendquist, O. A. McBryan, and G. Tryggvason. The bifurcation

of tracked scalar waves. SIAM J. Sci. Stat. Comput., 9(1):61–79, 1988.

[28] R. Glowinski. Finite element methods for incompressible viscous ﬂow. In J. L. Lions

and P. G. Ciarlet, editors, Numerical Methods for Fluids (3), Handbook of Numerical

Analysis, volume 9, pages 3–1176. North-Holland, Amsterdam, 2003.

28

[29] V. Gravemeier. Scale-separating operators for variational multiscale large eddy sim-

ulation of turbulent ﬂows. J. Comput. Phys., 212:400–435, 2006.

[30] V. Gravemeier. The Variational Multiscale Method for Laminar and Turbulent Flow.

Arch. Comput. Meth. Engrg., 13(2):249–324, 2006.

[31] P. M. Gresho and R. L. Sani. Incompressible Flow and the Finite Element Method,

volume one: Advection–Diffusion. John Wiley & Sons, Chichester, 2000.

[32] P. M. Gresho and R. L. Sani. Incompressible Flow and the Finite Element Method,

volume two: Isothermal Laminar Flow. John Wiley & Sons, Chichester, 2000.

[33] J. L. Guermond and L. Quartapelle. Calculation of Incompressible Viscous Flow by

an Unconditionally Stable Projection FEM. J. Comput. Phys., 132:12–33, 1997.

[34] J. L. Guermond and L. Quartapelle. A projection FEM for variable density incom-

pressible ﬂows. J. Comput. Phys., 165(1):167–188, 2000.

[35] M. D. Gunzburger. Finite Element Methods for Viscous Incompressible Flows. Aca-

demic Press, Boston, 1989.

[36] E. Hairer, S. P. Nørsett, and G. Wanner. Solving Ordinary Differential Equations I.

Nonstiff Problems. Springer-Verlag, Berlin, 2nd edition, 1993.

[37] E. Hairer and G. Wanner. Solving Ordinary Differential Equations II. Stiff and

Differential-Algebraic Problems. Springer-Verlag, Berlin, 1991.

[38] P. Hansbo and A. Szepessy. A velocity–pressure streamline diffusion FEM. Comput.

Methods Appl. Mech. Engrg., 84(2):175–192, 1990.

[39] F. H. Harlow and J. E. Welch. Numerical Calculation of Time-Dependent Viscous

Incompressible Flowof FluidwithFree Surface. Phys. Fluids, 8(12):2182–2189, 1965.

[40] A. Harten. Multiresolution Representation of Data: A General Framework. SIAM J.

Numer. Anal., 33(3):1205–1256, 1996.

[41] M. A. Heroux, R. A. Bartlett, V. E. Howle, R. J. Hoekstra, J. J. Hu, T. G. Kolda, R. B.

Lehoucq, K. R. Long, R. P. Pawlowski, E. T. Phipps, A. G. Salinger, H. K. Thornquist,

R. S. Tuminaro, J. M. Willenbring, Williams A., and K. S. Stanley. An overview of the

Trilinos project. ACM Trans. Math. Softw., 31(3):397–423, 2005.

[42] C. W. Hirt and B. D. Nichols. Volume of ﬂuid (VOF) method for the dynamics of free

boundaries. J. Comput. Phys., 39(1):201–225, 1981.

[43] T. J. R. Hughes. The ﬁnite element method: linear static and dynamic ﬁnite element

analysis. Dover, Mineola, NY, 2000.

[44] T. J. R. Hughes, G. R. Feijóo, L. Mazzei, and J.-B. Quincy. The variational multiscale

method – a paradigm for computational mechanics. Comput. Methods Appl. Mech.

Engrg., 166:3–24, 1998.

29

[45] T. J. R. Hughes, L. P. Franca, and M. Balestra. A new ﬁnite element formulation

for computational ﬂuid dynamics: V. Circumventing the Babuska–Brezzi condition:

A stable Petrov–Galerkin formulation of the Stokes problem accomodating equal-

order interpolations. Comput. Methods Appl. Mech. Engrg., 59:85–99, 1986.

[46] T. J. R. Hughes, L. Mazzei, and K. E. Jansen. Large Eddy Simulation and the varia-

tional multiscale method. Comput. Visual. Sci., 3:47–59, 2000.

[47] D. Jacqmin. Calculation of two-phase Navier–Stokes ﬂows using phase-ﬁeld mod-

eling. J. Comput. Phys., 155(1):96–127, 1999.

[48] V. John. On the efﬁciency of linearization schemes and coupled multigrid methods

in the simulation of a 3Dﬂowaround a cylinder. Int. J. Numer. Meth. Fluids, 50:845–

862, 2006.

[49] V. John. On large eddy simulation and variational multiscale methods in the nu-

merical simulation of turbulent incompressible ﬂows. Appl. Math., 50(4):321–353,

2008.

[50] V. John and S. Kaya. A ﬁnite element variational multiscale method for the Navier–

Stokes equations. SIAM J. Sci. Comput., 26(5):1485–1503, 2005.

[51] V. John, G. Matthies, and J. Rang. A comparison of time-discretization/lineariza-

tionapproaches for the incompressible Navier–Stokes equations. Comput. Methods

Appl. Mech. Engrg., 195(44-47):5995–6010, 2006.

[52] H. JohnstonandJ.-G. Liu. Accurate, stable andefﬁcient Navier–Stokes solvers based

on explicit treatment of the pressure term. J. Comput. Phys., 199(1):221–259, 2004.

[53] D. A. Kay, D. Loghin, and A. Wathen. A preconditioner for the steady-state Navier–

Stokes equations. SIAM J. Sci. Comput., 24(1):237–256, 2002.

[54] D. A. Kay and R. R. Welford. Efﬁcient Numerical Solution of Cahn–Hilliard–Navier–

Stokes Fluids in 2D. SIAM J. Sci. Comput., 29(6):2241–2257, 2007.

[55] A. Klawonn. An Optimal Preconditioner for a Class of Saddle Point Problems with a

Penalty Term. SIAM J. Sci. Comput., 19(2):540–552, 1998.

[56] A. N. Kolmogorov. The Local Structure of Turbulence in Incompressible Flows (in

Russian). Proceedings of the USSR Academy of Sciences, 30:299–303, 1941.

[57] D. B. Kothe, W. J. Rider, S. J. Mosso, J. S. Brock, and J. I. Hochstein. Volume tracking

of interfaces having surface tension in two and three dimensions. In Proceedings

of the 34th Aerospace Sciences Meeting and Exhibit, Technical Report AIAA 96-0859,

Reno, NV, 1996.

[58] O. A. Ladyzhenskaya. The mathematical theory of viscous incompressible ﬂow. Gor-

don and Breach Science, New York, 1969.

[59] L. Lee and R. LeVeque. An immersed interface method for incompressible Navier–

Stokes equations. SIAM J. Sci. Comput., 25(3):832–856, 2003.

30

[60] R. J. LeVeque and Z. Li. The immersed interface method for elliptic equations with

discontinuous coefﬁcients and singular sources. SIAM J. Numer. Anal., 31(4):1019–

1044, 1994.

[61] R. J. LeVeque and Z. Li. Immersed interface method for Stokes ﬂow with elastic

boundaries or surface tension. SIAM J. Sci. Comput., 18(3):709–735, 1997.

[62] P.-L. Lions. Mathematical Topics in Fluid Mechanics, volume 1: Incompressible

Models. Oxford University Press, Oxford, 1996.

[63] P. Moin and K. Mahesh. Direct numerical simulation: a tool in turbulence research.

Annu. Rev. Fluid Mech., 30:539–578, 1998.

[64] R. D. Moser, J. Kim, and N. N. Mansour. Direct numerical simulation of turbulent

channel ﬂow up to Re

τ

=590. Phys. Fluids, 11(4):943–945, 1999.

[65] W. Noh and P. Woodwards. SLIC (simple line interface calculation). In Proceedings

of the 5th International Conference on Numerical Methods in Fluid Dynamics, pages

330–340, Enschede, 1976. Springer-Verlag.

[66] A. A. Oberai, V. Gravemeier, and G. C. Burton. Transfer of Energy in the Variational

Multiscale Formulation of LES. Technical report, Center for Turbulence Research,

Stanford, Summer Program 2004.

[67] E. Olsson and G. Kreiss. A conservative level set method for two phase ﬂow. J.

Comput. Phys., 210:225–246, 2005.

[68] E. Olsson, G. Kreiss, and S. Zahedi. A conservative level set method for two phase

ﬂow II. J. Comput. Phys., 225:785–807, 2007.

[69] S. Osher and R. P. Fedkiw. Level Set Methods and Dynamic Implicit Surfaces. Applied

Mathematical Sciences. Springer, New York, 2003.

[70] S. Osher and J. A. Sethian. Fronts Propagating with Curvature-Dependent Speed:

Algorithms Based on Hamilton–Jacobi Formulations. J. Comput. Phys., 79(1):12–

49, 1988.

[71] C. S. Peskin. Numerical Analysis of Blood Flow in the Heart. J. Comput. Phys.,

25(3):220–252, 1977.

[72] C. S. Peskin. The immersed boundary method. Acta Numerica, 11:479–517, 2002.

[73] S. B. Pope. Turbulent Flows. Cambridge University Press, Cambridge, 2000.

[74] L. Quartapelle. Numerical solution of the incompressible Navier–Stokes equations,

volume 113 of International Series of Numerical Mathematics. Birkhäuser-Verlag,

Basel, 1993.

[75] W. Ren and W. E. Heterogeneous multiscale method for the modeling of complex

ﬂuids and micro-ﬂuidics. J. Comput. Phys., 204(1):1–26, 2005.

[76] M. Renardy, Y. Renardy, and J. Li. Numerical Simulation of Moving Contact Line

Problems Using a Volume-of-Fluid Method. J. Comput. Phys., 171(1):243–263, 2001.

31

[77] P. Sagaut. Large Eddy Simulation for Incompressible Flows. Springer-Verlag, Berlin,

3rd edition, 2006.

[78] R. Scardovelli and S. Zaleski. Direct Numerical Simulation of Free-Surface and In-

terfacial Flow. Ann. Rev. Fluid Mech., 31:567–603, 1999.

[79] J. A. Sethian. Level Set Methods and Fast Marching Methods. Evolving Interfaces

in Computational Geometry, Fluid Mechanics, Computer Vision, and Materials Sci-

ence. Cambridge University Press, 2000.

[80] J. Smagorinksy. General Circulation Experiments with the Primitive Equations. I.

The Basic Experiment. Monthly Weather Rev., 91(3):99–164, 1963.

[81] M. Sussman, P. Smereka, and S. Osher. A level set method for computing solutions

to incompressible two-phase ﬂow. J. Comput. Phys., 114(1):146–159, 1994.

[82] C. Taylor and P. Hood. A numerical solution of the Navier–Stokes equations using

the ﬁnite element technique. Comput. Fluids, 1(1):73–100, 1973.

[83] R. Teman. Une méthode d’approximation de la solution des équations de Navier–

Stokes. Bull. Soc. Math. France, 96:115–152, 1968.

[84] T. E. Tezduyar. Stabilized Finite Element Formulations for Incompressible Flow

Computations. Adv. Appl. Mech., 28:1–44, 1992.

[85] U. Trottenberg, C. Oosterlee, and A. Schüller. Multigrid. Elsevier Academic Press,

London, 2001.

[86] G. Tryggvason, B. Bunner, A. Esmaeeli, D. Juric, N. Al-Rawahi, W. Tauber, J. Han,

S. Nas, and Y.-J. Jan. A Front Tracking Method for the Computations of Multiphase

Flow. J. Comput. Phys., 169(2):708–759, 2001.

[87] R. Tuminaro and C. Tong. Parallel smoothed aggregation multigrid: aggregation

strategies on massively parallel machines. In J. Donnelley, editor, Super Computing

2000 Proceedings, 2000.

[88] S. Turek. Efﬁcient Solvers for Incompressible Flow Problems: An Algorithmic and

Computational Approach. Springer, Berlin, 1999.

[89] S. O. Unverdi and G. Tryggvason. AFront Tracking Method for Viscous, Incompress-

ible, Multi-ﬂuid Flows. J. Comput. Phys., 100(1):25–37, 1992.

[90] J. D. van der Waals. The thermodynamic theory of capillarity under the hypothesis

of continuous variation of density. J. Stat. Phys. (originally published in Dutch in

Verhandel. Konink. Akad. Weten. Amsterdam, 1893), 20(2):200–244, 1979.

[91] W. Villanueva and G. Amberg. Some generic capillary-driven ﬂows. Int. J. Multi-

phase Flow, 32:1072–1086, 2006.

[92] S. Zahedi, K. Gustavsson, and G. Kreiss. A conservative level set method for contact

line dynamics. J. Comput. Phys., 228(17):6361–6375, 2009.

32

Recent licentiate theses from the Department of Information Technology

2009-004 Katharina Kormann: Numerical Methods for Quantum Molecular Dynamics

2009-003 Marta L´ arusd´ ottir: Listen to Your Users - The Effect of Usability Evaluation on

Software Development Practice

2009-002 Elina Eriksson: Making Sense of Usability - Organizational Change and Sense-

making when Introducing User-Centred Systems Design in Public Authorities

2009-001 JoakimEriksson: Detailed Simulation of Heterogeneous Wireless Sensor Networks

2008-003 Andreas Hellander: Numerical Simulation of Well Stirred Biochemical Reaction

Networks Governed by the Master Equation

2008-002 Ioana Rodhe: Query Authentication and Data Conﬁdentiality in Wireless Sensor

Networks

2008-001 Mattias Wiggberg: Unwinding Processes in Computer Science Student Projects

2007-006 Bj¨ orn Halvarsson: Interaction Analysis and Control of Bioreactors for Nitrogen

Removal

2007-005 Mahen Jayawardena: Parallel Algorithms and Implementations for Genetic Anal-

ysis of Quantitative Traits

2007-004 Olof Rensfelt: Tools and Methods for Evaluation of Overlay Networks

2007-003 Thabotharan Kathiravelu: Towards Content Distribution in Opportunistic Net-

works

2007-002 Jonas Boustedt: Students Working with a Large Software System: Experiences and

Understandings

2007-001 Manivasakan Sabesan: Querying Mediated Web Services

2006-012 Stefan Blomkvist: User-Centred Design and Agile Development of IT Systems

2006-011

˚

Asa Cajander: Values and Perspectives Affecting IT Systems Development and

Usability Work

Department of Information Technology, Uppsala University, Sweden

Numerical Methods for the Navier– Stokes Equations applied to Turbulent Flow and to Multi-Phase Flow

BY

M ARTIN K RONBICHLER December 2009

D IVISION OF S CIENTIFIC C OMPUTING D EPARTMENT OF I NFORMATION T ECHNOLOGY U PPSALA U NIVERSITY U PPSALA S WEDEN

Dissertation for the degree of Licentiate of Philosophy in Scientiﬁc Computing with specialization in Numerical Analysis at Uppsala University 2009

Numerical Methods for the Navier– Stokes Equations applied to Turbulent Flow and to Multi-Phase Flow

Martin Kronbichler

martin.kronbichler@it.uu.se

Division of Scientiﬁc Computing Department of Information Technology Uppsala University Box 337 SE-751 05 Uppsala Sweden

http://www.it.uu.se/

c Martin Kronbichler 2009 ISSN 1404-5117 Printed by the Department of Information Technology, Uppsala University, Sweden

in particular the large eddy simulation of turbulent ﬂow and the simulation of laminar immiscible two-phase ﬂow. The computations for both applications are performed with a coupled solution approach of the Navier– Stokes equations discretized with the ﬁnite element method. where scale ranges are separated by variational projection. This formulation combines the advantages of a level set model and of a phase ﬁeld model: Motion of contact lines and imposition of contact angles are handled like for a phase ﬁeld method. The error in both velocity and pressure decreases with increasing number of mesh points.Abstract This thesis discusses the numerical approximation of ﬂow problems. The phase-ﬁeld speciﬁc terms are only active in the vicinity of contact lines. but the computational costs are similar to the ones of a level set implementation. A Fourier analysis underlines the importance of projective scale separations and that the proposed model does not affect consistency of the numerical scheme. The method uses a standard Navier–Stokes model for representing the coarser of the resolved scales. a method for modeling laminar two-phase ﬂow problems in the vicinity of contact lines is proposed. whereas the errors in pressure prediction mainly come from the ﬁnite width of the interface. Numerical examples show that the method provides better results than other state-of-the-art methods for computations at low resolutions. . It is observed that the errors in velocity are mostly due to inaccuracies in curvature evaluation. The model is realized by formulating the Cahn–Hilliard equation as an extension of a level set model. Firstly. a new implementation strategy for large eddy simulation of turbulent ﬂow is discussed. The approach is based on the variational multiscale method. The scale separation within the space of resolved scales is implemented in a purely algebraic way based on a plain aggregation algebraic multigrid restriction operator. Secondly. Moreover. various aspects of a conservative level set method discretized with ﬁnite elements regarding the accuracy of force balance and prediction in jump of pressure between the inside and outside of a circular bubble are tested systematically. and adds a subgrid viscosity model to the smaller of the resolved scales.

.

Gravemeier. Zahedi. III M. Kronbichler.2009. and W. Contributions: Ideas and analysis in section 3. M. Submitted to Computers and Fluids. I V. Preparation of a draft of the manuscript and ﬁnal work in close cooperation between the authors. Kronbichler. II V. Ideas were developed by close cooperation between the ﬁrst and the second author in discussion with the third author. 2009. Kreiss (2008) A Hybrid Level–Set–Cahn–Hilliard Model for Two-Phase Flow. Preparation of the manuscript. Ideas were developed in discussion with the co-author. W.1016/j. Submitted to Computers and Fluids. A. in particular section 2. .List of Papers This thesis is based on the following papers. Contributions: Implementation of and experiments with the coupled two-phase ﬂow solver. and G. Gravemeier. Kreiss (2009) Evaluation of inaccuracies in two-phase ﬂow simulations with level set methods using ﬁnite element discretizations. Contributions: Implementation of the method and computations. Department of Information Technology.cma. Preparation of a draft of the manuscript. Italy. Accepted for publication in Computer Methods in Applied Mechanics and Engineering. M. Bologna. Contributions: Initial implementation of the method together with the ﬁrst and second author. Technical Report 2009-026.05.017. Gee. In Proceedings of the 1st European Conference on Microﬂuidics. Ideas were developed in close cooperation between the authors. Wall (2009) An algebraic variational multiscale–multigrid method for large-eddy simulation: generalized-α time integration. Kronbichler. M. Wall (2009) An algebraic variational multiscale–multigrid method for large eddy simulation of turbulent ﬂow. A. W. Kronbichler and G. doi:10. Uppsala University. and W. First draft of the manuscript. M. Gee. which are referred to in the text by their Roman numerals. IV S. M. Reprints were made with permission from the publishers. Fourier analysis and application to turbulent ﬂow past a square-section cylinder.

.

. . . . . . . . .2 Level set models . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . 6 Future Work . . . . . . . . . 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . . . . . . .1 Paper I . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Summary of Papers . . . .4 Paper IV . . . . . . 2. . . . . . . . . . . . . . . . . .3 Paper III . . . . . . . . . . . . . . . . . . 5. . . . . . . . . . . . . . . . . . . . . . . . 3 Computational methods for turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 A multiscale approach to large eddy simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4. . . .1 Overview of numerical models . . . . . . . . . . . . .Contents 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 3 4 5 5 6 7 7 9 13 14 15 18 21 21 21 22 22 23 27 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Basics on turbulence and large eddy simulation . . . . . Bibliography . . . . . 5. . . . . . . . . . . . . . . .2 Paper II . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Stability of spatial discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4. . . . . . . . . . . . . . . . .1 Weak form and spatial discretization . . . . . . . . . . . 2 Numerical Methods for the Navier–Stokes Equations . . . . . . . . . . . . . . . .3 Time discretization . . . . 5. 3. . .3 Phase ﬁeld models . . . . . . . . . . 2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Simulation of Two-Phase Flow .4 Numerical linear algebra for coupled discretizations . . . . . . . . . . . . . . . . .

.

These processes are often strongly driven by surface tension. where different ﬂuids and ﬂame fronts need to be represented. Simulation is today both an alternative and a complement to experiments in many engineering disciplines as it helps in predicting the behavior of ﬂuids. For instance. a force that strives after keeping different ﬂuids separated. A similar application is the modeling of the absorption of air in the blood circulation in human lungs. Turbulent ﬂows often occur in engineering applications. One interesting area of research in ﬂuid dynamics is the behavior of turbulent incompressible ﬂow. Simulation is thus a tools that renders technical processes more effective. Introduction The numerical simulation of ﬂuid dynamics is one of the main ﬁelds in computational mathematics. This interaction takes place on so-called contact lines. Such a procedure would be tremendously expensive if it had to be done by prototyping and other experimental means. For subsurface ﬂows. the ﬂuids do not mix and there are no additional particles resolved in the ﬂuids. An exceptionally challenging problem is computational combustion. One such application is intravenous therapy in medicine. namely the interaction of systems consisting of different ﬂuids with solids. This thesis contributes to the simulation of such laminar well-separated two-phase ﬂows. In this area. The properties of the solid and the composition of the ﬂuid give rise to a phenomenon called wetting. an additional difﬁculty arises. There is a wide range of applications that involve multiple ﬂuids and where numerical simulation is extensively used. the water ﬂow in turbines of hydroelectric power plants as well as in water networks and oil pipelines with high throughput. that is. relevant issues are the determination and the control of aerodynamic drag as well as efﬁciency of engines. But there are also a lot of multi-phase settings where the ﬂow ﬁeld is laminar and thus more regular. it is often assumed that the individual ﬂuids are wellseparated. the ﬂows occurring in combustion are turbulent. Multiphase ﬂow is a phenomenon that is particularly relevant also in subsurface ﬂow. Another area of active research is multiphase ﬂow. called capillarity. simulation permits to systematically improve the design and material properties of products by means of optimization algorithms. The objective of numerical simulations is to understand how bubbles consisting of oleaginous substances are transported and resolved in the circulation of blood. Often. in the air ﬂow around a vehicle.1. a setting that involves two or more different ﬂuids. Applications of this kind are the modeling of groundwater basins and oil extraction 1 . locations where the interface (separating two ﬂuids) meets a solid. Moreover. indicating which ﬂuid preferably is in contact with the solid. for example.

where mixtures of different ﬂuids and their solidiﬁcation are tracked. and inkjet printing. One example is to perform small-scale ﬂow simulations with detailed rock structures in order to determine the permeability of various stone conﬁgurations. and Chapter 6 identiﬁes possible directions of future research. usually based on Darcy’s law [17]. Chapter 5 gives a short summary of the papers included in this thesis. Next. the Navier–Stokes equations. The pores of rock where the ﬂuid moves are often of the order of a few millimeters or less. simpliﬁcations need to be applied. lab-on-achip processes. we discuss the equations that describe the ﬂow of individual ﬂuids. This thesis consists of ﬁve chapters. First. an introduction to turbulence is given. Other applications of multiple ﬂuids in contact with solids are found in industrial processes like sintering. In order to tackle problems with such an enormously wide range of scales. 2 . whereas the actual simulation region can extend over thousands of square kilometers. Such effective parameters are then used as input to a large scale simulation. which form the basic component for microsystem design in the biomedical industry.from subsurface reservoirs driven by water injection. which is followed by a presentation of numerical models for the simulation of laminar two-phase ﬂow.

1)–(2.5) where n denotes the unit outer normal on the domain boundary Γh and h is some external traction boundary force. Numerical Methods for the Navier–Stokes Equations The motion of ﬂuids at low and medium velocities in a connected computational domain Ω ⊂ Rd (spatial dimension d = 2. density is constant due to the incompressibility assumption. There is a vast literature discussing various aspects of the ﬁnite element discretization of system (2.2) is completed by a divergence-free initial velocity ﬁeld. 0) = u0 . The algorithms developed in this thesis are based on the ﬁnite element method (FEM).3) and application-speciﬁc boundary conditions (cf.1)–(2. u(·.1) (2. The rank-2 tensor ε(u) = 2 ∇u + (∇u)T represents the viscous stress tensor for a Newtonian ﬂuid. [62].4) and natural (inﬂow/outﬂow) boundary conditions n · ε(u) + pn = h on Γh ⊂ ∂Ω. The boundary subdomains Γg and Γh are assumed to be non-overlapping and to span the whole boundary ∂Ω.2. A comparative introduction of these methods can be found in Quartapelle [74]. The ﬂuid density is denoted by ρ. (2. ρ (ut + u · ∇u) − ∇ · 2µε(u) + ∇p = ρf. (2. we consider ﬂow subject to no-slip boundary conditions u = 0 on Γg ⊂ ∂Ω. the ﬁnite difference method. The term f speciﬁes external forces acting on the ﬂuid. In this thesis. [73]. ∇ · u = 0. (2. there are various methods available. In single-phase ﬂuids.2). For the numerical approximation of the equations of ﬂuid ﬂows. The most popular ones are the ﬁnite volume method. Mathematical aspects are 3 .2) The vector u denotes the d components of ﬂuid velocity and p denotes the ﬂuid’s 1 (dynamic) pressure. 3) is given by the incompressible Navier–Stokes equations [35]. with ∇ · u0 = 0. Gresho and Sani [32] for a discussion of various possibilities). and the ﬁnite element method. (2. and the ﬂuid’s (dynamic) viscosity by µ. The system of Navier–Stokes equations (2.

1)–(2. The variational problem corresponding to (2.6) denotes the standard Ω = v. only the projection of the Navier–Stokes equations onto ﬁnite dimensional test function spaces Vuh × Vph is considered. The support of the basis functions is conﬁned to a patch of elements around the respective nodes. p Ω− q. e. q ∈ Vu × Vp . and the space of admissible pressure solutions by Vp = L 2 (Ω). t ) ∈ Vu × Vp such that v. the ﬁrst step is to rewrite the system of equations (2. equally shaped basis functions are used for each spatial component. The form ·. For instance. resulting in numerical approxima- (2. at each instant in time. we will denote a ﬁnite element approximation with quadratic basis functions for velocity and linear bad sis functions for pressure on the elements by Q2 Q1 . Then. and the pressure basis functions as tensor product of linear Lagrangian interpolants. p(·. The discretizations considered in this thesis use a decomposition of the computational domain into small subsets of characteristic size h. For the representation of the velocity u. The book by Gresho and Sani [32] focuses on practical issues for implementation such as boundary conditions.2) is to ﬁnd. The spatial discretization of system (2.discussed. t )..1 Weak form and spatial discretization The Navier–Stokes equations (2. On each element. To this end. ∇ · u holds for all test functions v. For setting up the spatial FE approximation. Globally. h Γh tions uh and p h . 4 .6) replaces the (inﬁnite-dimensional) solution function spaces Vu and Vp by ﬁnite-dimensional subspaces Vuh and Vph . discretization schemes. Silvester & Wathen [23] and Turek [88] put special emphasis on algorithmic matters like efﬁcient numerical linear algebra for the ﬂow equations. the basis functions for u are deﬁned as a tensor product of Lagrangian interpolation polynomials of degree two in each coordinate direction [43]. and zero on all the others. ρ (ut + u · ∇u) Ω+ ε(v). we deﬁne the space of admissible velocity solutions at a certain time instant by Vu = u ∈ H 1 (Ω). µε(u) Ω− ∇ · v.1)–(2. f Ω+ v. and continuous over the whole domain Ω. called elements.g. and stabilization. The basis functions that span Vu and Vp are chosen to be piecewise polynomials within the elements.2) in variational form. · Ω L 2 inner product. The work by Elman. u = 0 on Γw . a pair u(·. 2. in Glowinski [28] and Gunzburger [35]. We mainly focus on quadrilateral elements in 2D and hexahedral (brick) elements in 3D.1)–(2. the basis functions are nonzero in exactly one node. Most discretization approaches split the approximation into separate approximations for spatial and temporal part.2) depend on both space and time.

namely ∆t max ∝ ∆x ρ∆x 2 + . (semi-)implicit time stepping schemes are often preferred in order to avoid the diffusive stability constraint. Popular schemes are variants of the one-step theta method [51] and the BDF-2 method. a compatibility condition between the approximation of velocity and pressure must be satisﬁed. u ∞ µ see. However. Papers I and II investigate ﬂow problems with the element pair d Q1 Q1 . We refer to Donea & Huerta [18] and Gresho & Sani [32] for an extensive discussion of LBB-stable elements.7) where the constant β does not depend on the mesh size parameter h.. [10] for a review of various commonly used stabilization techniques. ∇ · v q L2 Ω H1 q∈Vph v ≥ β > 0. [36]. resulting in a Petrov–Galerkin type scheme. (2. This condition was independently derived by Ladyzhenskaya [58]. an element pair that does not satisfy the LBB condition (2. [32]. For a 5 . whereas the element combination Q1 Q1 does not. an artiﬁcial diffusion on the pressure space is introduced in a systematic way. ∇ · u Ω is of full rank. [37] for derivation and analysis of numerical schemes. see Donea & Huerta [18] and Hairer et al. it is usual to use some standard procedure developed for ordinary differential equations and differential–algebraic equations. For applications where the viscosity is not too small compared to the velocity. For stabilization. Babuška [4]. and Brezzi [12] and is called LBB condition or inf–sup condition. [45].2 Stability of spatial discretization For the spatial discretization to be stable.6). [84] modiﬁes the test functions in the discrete variational form (2. 2. The condition ensures that the divergence matrix discretizing the term q. Similar stabilization approaches have been developed to avoid unphysical oscillations in velocity that otherwise appear in convection-dominated regimes for the standard central difference approximations introduced by the Galerkin FEM [18]. satd isﬁes the LBB condition [82]. so that consistency is preserved. This approach [38]. d The element combination Q2 Q1 . e. [31].2.3 Time discretization For discretization in time. the so-called Taylor–Hood element pair. the discussion in [52].7) can yet be modiﬁed so that stability of discretization is obtained. namely inf sup h v∈Vu q. linear stability theory sets a restrictive limit on the time step ∆t in terms of the spatial mesh size ∆x.g. and hence the discretized system has a unique solution for velocity and pressure. For schemes that treat both convection and diffusion explicitly. We refer to Braack et al.

Efﬁcient schemes employ multigrid methods with suitably deﬁned smoothing schemes (see Trottenberg. These terms enforce the divergence-free condition on the velocity in Lagrangian multiplier form.g. so the ﬁnal coefﬁcient matrix is sparse. with boundary conditions set according to physical reasoning. We refer to Gresho & Sani [32] and Quartapelle [74] for discussion. This concept has been mathematically justiﬁed for the ﬁnite element method by Guermond and Quartapelle [33]. Projection schemes are to date the most popular variant for the numerical approximation of the Navier–Stokes equations [74]. Projection schemes ﬁrst advance velocities subject to the momentum equation with pressure extrapolated from the old time step. stability requires implicit time discretization of the (differential–algebraic) terms ∇ · v. which are usually not the physical ones. The Lagrangian basis functions are localized.. For resolving the non-linearity. [53]. different techniques have been proposed. ﬁnally yielding a system of linear equations.4 Numerical linear algebra for coupled discretizations The introduction of spatial and temporal discretization schemes for the Navier–Stokes system (2. see. so a pressure Poisson equation is solved with forcing given by the divergence of the medium-step velocity in a second step. However. [55]. in order to correct the velocity. The subsystems resulting from fractional time stepping are of easier structure than the original saddle-point system. and which was then recast as a projection scheme by Chorin [15] and Temam [83]. Harlow and Welch [39] proposed to implement time propagation of theNavier–Stokes equations by fractional time stepping strategy.6) results in a nonlinear system to be solved. However. This can yield a non-consistent approximation close to the boundary. [22]. the algebraic system is in general more challenging because of its saddle point nature. p Ω and q. 6 . The resulting velocity is in general not divergence-free. The work presented in this thesis applies an alternative approach where the system remains coupled. Silvester & Wathen [23]). For the solution of the saddle-point Navier–Stokes system. Gunzburger [35] and Lions [62]. or use preconditioners based on the Schur complement of the block 2 × 2 matrix system deﬁned by velocity and pressure matrices (cf. e. Coupled schemes solve for pressure and velocity as one algebraic system. ∇ · u Ω .coupled solution approach. and puts high demands on numerical linear algebra. and the numerical method used in Papers III and IV is based on a preconditioner constructed from a Schur complement. 2. see also the books by Turek [88] and Elman. Papers I and II use an algebraic multigrid scheme for the coupled system. each of the subproblems needs to be equipped with suitable boundary conditions for well-posedness. Oosterlee & Schüller [85] and John [48]). and to the articles [13] and [52] for recent developments. [34]. ﬁxed point methods or Newton-type iterations are usually applied.

For ﬂow at high Reynolds numbers. The diagram shows the region of active scales in a ﬂow.1 Basics on turbulence and large eddy simulation A physical characterization of ﬂow is to distinguish vortex structures by their size relative to the ﬂow domain.2. and Sagaut [77].1) where l is a reference length and u a reference velocity. instabilities lead to a breakup of larger structures into smaller ones. The modeling equations are the Navier–Stokes equations (2. 3.1 depicts a typical pressure ﬁeld in a turbulent ﬂow. which is the context of Papers I and II. Vortex structures of different sizes as well as the irregular pattern of the ﬂow can clearly be seen. (3. ranging from the energy-containing scales l 0 (scale range of external forcing f and h) to the regime where viscosity becomes dominant at a scale level η.1)–(2. A ﬂow ﬁeld consists of vortices of different size. Computational methods for turbulence This chapter discusses basics of the simulation of turbulent ﬂow for a singlephase ﬂuid. the interaction between different scales is driven by the nonlinear convective term u · ∇u. the so-called Kolmogorov scale [73]. A situation where inertial stresses dominate is characterized by high Reynolds numbers.2) l0 l0 7 . These vortices are instationary and ﬂuctuating — a phenomenon called turbulence. In the Navier–Stokes equations. Re = ρl u .2). The transfer of energy that takes place in the model of ﬂuid motion is schematically depicted as a Kolmogorov scale spectrum in Figure 3. where the viscous stresses are small compared to the inertial stresses. µ (3. 3. The Kolmogorov scale η depends on the Reynolds number based on the production length scale l 0 as η ∝ Re−3/4 . see Dubois et al. The breakup process continues to create smaller and smaller structures until the scale is sufﬁciently small for viscous dissipation to become dominant. This induces an energy transport from larger scales to smaller scales. Pope [73]. Fig. [19].3. The ratio between inertial and viscous stresses is measured by the so-called Reynolds number.

1)–(2.2) needs to resolve all the scales in the ﬂow in order to represent any scale correctly. which amounts in a number of degrees of freedom that is proportional to Re3 . Kolmogorov [56] noticed a self-similarity between the eddies in the inertial regime. where κ is the scale length. Kolmogorov formulated the hypothesis that the energy content at scales in the inertial subrange. A numerical approach that aims at resolving all necessary scales is called a direct numerical simulation (DNS) [63]. a DNS requires to resolve a scale range according to (3. 3. is proportional to κ−5/3 . as opposed to large-scale eddies. The large scales in the range between 8 . All the scales in incompressible ﬂows are coupled with one another. the range between energy production and viscous ﬂow. Many technically relevant ﬂows are described by Reynolds numbers within the range 104 − 106 . This idea is depicted in Fig.Figure 3.2) in each spatial direction as well as in the temporal dimension. a numerical method approximating the Navier–Stokes equations (2. The so-called large eddy simulation (LES) is one such modeling strategy. It aims at resolving the larger scales (large eddies).2. i. Hence. see Pope [73] and Sagaut [77]. often even with a rather large proportionality constant. Since a full resolution of all ﬂow features is not feasible. which depend mainly on geometry. Paper II).1: Pressure iso-surfaces for turbulent ﬂow around a square obstacle (cf. For three-dimensional instationary ﬂow. at least a fraction of these features needs to be modeled.e.. so the requirements in computational effort exceed the capabilities of state-of-the-art schemes for a DNS by far. necessitating other techniques. and to model the effect of unresolved scales on the large scales.

3. and a proportionality constant [77]. whereas all scales beyond κ0 are not resolved and need to be included by some modeling. A simple. the square of the mesh size. The eddy viscosity is deﬁned through a nonlinear coefﬁcient. Smagorinsky introduced an artiﬁcial subgrid or eddy viscosity µT with the aims of representing the effect of viscosity on the level of the computational mesh size.2: Kolmogorov energy spectrum in a double logarithmic plot. It is based on the observation that the insufﬁcient resolution within a LES leads to underresolved viscous effects. The wave length κ0 denotes the resolution limit of the discretization. ﬁnding appropriate boundary conditions for the averaged large scales is to date an open problem. [20]. this spatial ﬁltering is problematic close to domain boundaries.Figure 3.2 A multiscale approach to large eddy simulation The deﬁnition of a LES needs to distinguish between resolved and unresolved scales in order to formulate the subgrid viscosity term. yet very powerful approach is the Smagorinsky model [80]. These problems can be traced back to so-called commutation errors [7]. 9 . Moreover. l 0 and κ0 are resolved by the numerical discretization. Hence. based on the magnitude of the viscous stress tensor ε(u). However. All the scales beyond that level will not be captured by the numerics and their effect needs to be modeled in order to correctly describe the physics of turbulence. Traditional methods use low-pass ﬁlters to extract the large-scale information from the Navier–Stokes equations and introduce the subgrid viscosity [77].

6). In this framework. and to restrict the test functions to the space of resolved variables. p h . p Ω− q.4) into the weak form (2. and Jansen [46]. uh . u. u ≈ 0.6) . (3. 10 . uh . ∇ · u Ω the form that contains all the left-hand-side terms of the weak form of the Navier–Stokes equations (2. q. u. The aim of any numerical method is to ﬁnd numerical approximations uh . Mazzei. For the numerical approximation to be computable. u. and ﬁnally. h (3. Hence. representing large resolved scales. q h ∈ Vuh ×Vph . p h in the resolved subspace Vuh ⊗ Vp according to the direct sum decomposition (3. Jauberteau & Temam [19]. the nonlinear convective term is linearized and resorted in terms of resolved and unresolved quantities [46]. We choose a resolution of 3h. and the multiscale turbulence analysis in Collis [16]. different deﬁnitions of large resolved scales are common as well.An alternative approach to LES is based on the variational multiscale method [44]. we get (following Harten’s notation [40]) ˆ u = u3h + δuh +u. uh . q h . u. Applying an equivalent decomposition for pressure. the space of unresolved scales Vu .3). p − BNS vh . u (3. A systematic way of doing so is to insert the decomposition (3. u. uh . Vuδh is the space spanned by Vuh except Vu3h . p h = vh . q h . (3.6). however. The variational multiscale large eddy simulation with a subgrid viscosity model constructs the approximation as 1 2 ˆ ˆ ˆ BNS v3h . u contains the quadratic convection term. these two terms need to be substituted by some modeling that eliminates the unresolved variables. q 3h . 1 2 ˆ ˆ ˆ BNS δvh . The term BNS vh . p) and the linearized convection around uh . p + B NS v3h . representing small resolved scales. q h . u ≈ ε δvh . δq h . see the review article by Gravemeier [30]. uh Let us denote by BNS v.3) where Vu3h is a space associated with mesh size 3h. ρ (ut + u · ∇u) Ω+ ˆ p = p h + p. ﬁlters are replaced by variational projections onto appropriate subspaces. f Ω + vh . p + B NS δvh . p = v. introduced by Hughes. 2νT ε δuh Ω Γh 1 2 ˆ ˆ ˆ − BNS vh .4) ε(v). The space for admissible velocity solutions is decomposed into Vu = Vu3h ⊗ Vuδh ⊗ Vu . The ideas are closely related to the concept of dynamic multilevel methods. and the term 2 ˆ BNS vh . µε(u) Ω− ∇ · v. uh . p conˆ ˆ tains linear terms in (u. such that BNS vh . we search for uh . Then.5) 1 ˆ ˆ holds for all test functions vh . described in Dubois.

Standard methods for doing this are based on different polynomial orders for the FE basis functions. Papers I and II discuss an implementation where this separation is performed on the algebraic equation system by using routines from algebraic multigrid. e. in geometric multigrid..6) needs to distinguish between coarse and ﬁne resolved scales..g. so that the effect of a not adequately resolved viscosity is best modeled on the ﬁne resolved scales [66]. or two different grid levels as used. as discussed.Note that the subgrid viscosity is only applied to the smaller of the resolved scales. 11 . e. in the review articles by Gravemeier [30] and John [49].g. The implementation of a method according to the framework (3.5) with approximation (3. A physical explanation for application of eddy viscosity only on small resolved scales is that energy transport occurs mainly on neighboring scale groups. Such methods have been presented by John & Kaya [50] and Gravemeier [29].

.

e. A widespread formulation is to pose the two-phase ﬂow problem on the whole domain Ω = Ω1 ∪ Γ ∪ Ω2 with variables u and p that take the respective velocity and pressure values in the individual domains. the Reynolds number is low for the considerations in this chapter. Simulation of Two-Phase Flow We consider the dynamics of two immiscible incompressible ﬂuids separated by an interface Γ as displayed in Fig.. Extra conditions are necessary to describe the behavior on the interface Γ.1.2) using the variables ui and p i . and δΓ is a delta function that localizes the force to the interface.1)– (2. (4.4. n the direction normal to the interface (pointing into Ω2 ).1: Schematic view of a two-phase ﬂow problem. Ω1 Γ ∂Ω Ω2 Figure 4. and ﬂuid 2 occupies Ω2 . We assume that the ﬂow is laminar. Fluid 1 occupies region Ω1 . In addition. The interface Γ separates the two ﬂuids. σ is a constant specifying the relative strength of the surface tension.1) Here. the material parameters ρ and µ in the global momentum equation have a discontinuity on the interface. 4. i. Generally. the ﬂuid densities ρ i and viscosities µi are different for the two ﬂuids. By Ω1 we denote the region occupied by the ﬁrst ﬂuid. 2. Hence. The motion of each ﬂuid is described by the Navier–Stokes equations (2. interfacial tension enters the formulation as a forcing of the form fst = σκn δΓ . and by Ω2 the region occupied by the second. The interface Γ is transported by the local ﬂuid motion u|Γ . κ denotes the interface curvature. 13 . i = 1.

like elastic forces in immersed boundary and interface implementation [72]. and the evaluation of the distributional force term (4.The variables in a single-domain formulation comprise kinks and discontinuities due to changes in material parameters and forces: The discontinuous viscosity introduces a discontinuity in the symmetric gradient ε(u). The gravity force is modeled as fg = ρg eg . Similarly.2) where g is the gravitational acceleration and eg denotes the direction of gravity. the immersed interface method. the methods are quite difﬁcult to implement because the marker points need to interact with the ﬂuid grid in order to describe advection and interface forces. There are two main strategies to couple the interface evolution problem to the Navier–Stokes equations discretized on a ﬁxed grid as discussed in Chapter 2. gravitational forces are often essential for the dynamics of two-phase ﬂow. Normal vectors and curvatures can be calculated using the interface points by straight-forward geometric identities. The evolution of the interface in front tracking schemes is accomplished by Lagrangian advection of the marker points. by LeVeque and co-workers [59]. These methods are called front tracking and were introduced in different variants by Peskin [71] and Glimm et al. see Edwards. [86]. To retain an accurate interface repre14 .1 Overview of numerical models The two main challenges in the numerical simulation of two-phase ﬂow are the representation of the interface as it evolves in time. [27].1). and the surface tension force is generally evaluated based on a discrete approximation of the delta function around the interface. The total forcing in the momentum equation (2. [60]. A special class of front tracking schemes are the immersed boundary method by Peskin (see Peskin’s [72] review paper) and a more accurate variant. Brenner & Wasan [21] and Lee & LeVeque [59]. and [89]. (4. due to the distributional character of the delta function δΓ there is a jump in pressure proportional to σκ. Several discrete approaches have been proposed during the last decades. Explicit descriptions are very powerful because they allow to include general models of the interface.1) for two-phase ﬂow is hence given by f = fst + fg . and reviewed in the articles [78]. 4. The ﬁrst strategy is to explicitly track the interface position by introducing an additional discrete set of marker points or marker elements that describe the interface. The conditions describing the irregularities in velocity and pressure over the interface Γ are called jump conditions. Because of different ﬂuid densities. Any change in curvature tangential to the interface gives rise to a discontinuity in the pressure derivative normal to the interface as well as the viscous stress tensor. However. [61]. Straight-forward implementations are also prone to unphysical changes of the area/volume of the respective ﬂuid.

i. The advantage of VOF methods is the volume conservation and a natural mechanism for breakup and fusion of bubbles. Both these quantities can be computed locally from φ on the same computational mesh as the equations of ﬂuid ﬂow. The second strategy is front-capturing. VOF includes an additional variable that stores the fraction of the ﬁrst ﬂuid on the total ﬂuid for each cell. To put surface tension force in VOF into practice. and another employs the continuous surface tension approach by Brackbill. and curvature. Smereka & Osher [81]. the position of the interface can be reconstructed. Level set methods allow for a straight-forward evaluation of normal vectors by setting n(φ) = ∇φ/|∇φ|. formulated by Jacqmin [47] for the simulation of two-phase ﬂow. κ(φ) = ∇ · n(φ). and references therein. and Zemach [11]. one ﬂuid is colored by marker particles whose location is advected with the ﬂow ﬁeld. several different approaches are common. we refer to Scardovelli & Zaleski [78] and Sethian [79].. [86]. From the volume fractions. Kothe. introduced by Osher & Sethian [70] and ﬁrst applied to two-phase incompressible ﬂow by Sussman. where the interface is implicitly deﬁned. The mechanism for moving the interface in level set methods is advection of the function φ with local ﬂuid velocity. a continuous front capturing framework is provided by the level set method. (4.2 Level set models The basic idea of level set methods is to deﬁne the interface implicitly as the zero contour line of a function φ that is deﬁned in Ω. e. 4. Yet another method is provided by the phase ﬁeld method. by deﬁning line segments. including higher order schemes. Here. and can be considered to be a volumetracking method.e. The position of the interface can then be reconstructed from this particle ﬁeld. An extension of this approach is to replace the marker particles by a marker function. the reconstruction of the interface and the interface motion are considerably more complicated to implement here than for front tracking methods. the ﬁrst scheme of this kind was the marker-and-cell method proposed by Harlow and Welch [39]. For details of the reconstruction. The advection of the interface is usually implemented by increasing or decreasing the volume fraction depending on the composition of neighboring cells and the velocity ﬁeld. the marker points need to be redistributed when the interface is deformed. which are discussed in detail in the following two sections.. φt + u · ∇φ = 0.3) 15 . This approach has been pursued by Noh & Woodward [65] and Hirt & Nichols [42]. The methods discussed in Paper III and IV are based on the latter two approaches.g. One is to use discrete delta functions in combination with the reconstructed interface. Historically.sentation. see also the review by Tryggvason et al. While VOF methods operate on the discrete level. However. who introduced the volume-of-ﬂuid (VOF) method.

e. often in combination with high-resolution grids. Tornberg.2: Representing an interface by level set functions φ.g. like upwind ﬁnite difference/volume schemes or stabilized ﬁnite element methods. To start a level set calculation. However. A signed distance function encodes the distance of a particular point to the interface. Moreover.3) can be solved with standard tools for hyperbolic transport equations. algorithms have been proposed to (re16 . 0) needs to generated given a certain interface position. an initial proﬁle φ(·. [79]. ∇ · u = 0. the ﬂow ﬁeld as well as inaccuracies in the numerical scheme deform the signed distance function in the course of the simulation. A beneﬁt of this choice is that the integral in the surface tension force can be simpliﬁed to a one-dimensional discrete delta function as Γ δ x − y(s) d s ≈ δh φ(x) . These methods calculate φ only on a few grid cells around the interface. Since the information from φ is only needed in a region around the interface. depicted in Fig. The most popular choice for a level set function is the signed distance function φ. where x denotes an arbitrary point in the computational domain and y(s) is the image of a parametrization of the interface.. so-called narrow-band level set implementations are widely used. The advection can be implemented so that the volume is conserved. i.2(a). We refer to Adalsteinsson and Sethian [1] for details. since the velocity is divergence-free. and different signs identify the two ﬂuids. Recent implementations therefore reconstruct the interface position from φ and apply a d -dimensional discrete delta function that does not show these consistency problems. e. 4. and Tsai [24] that such an approximation can lead to O (1) errors in the force unless the width h of the delta function is varied. it has been shown by Engquist. see.This is an additional partial differential equation coupled to the Navier–Stokes equations. We refer to the books by Sethian [79] as well as Osher and Fedkiw [69] for an extensive discussion of these methods. The advection equation (4. A signed distance function has the property |∇φ| = 1 almost everywhere. Therefore.. φ ) d φ>0d φ<0 φ<0 d Ω2 Ω2 d Ω1 d s d d interface φ = 0 d d (a) Signed distance function φ A φ<0 φ>0 φ<0 Ω2 s d d Ω1 interface φ = 0 Ω2 (b) Smoothed color function Figure 4.

[68]) φτ + ∇ · n 1 − φ2 − ∇ · n ∇φ · n = 0. e. (4. so that the overall level set implementation based on signed distance functions may suffer from unphysical volume changes in the ﬂuid phases. (4. The parameter is typically chosen to be of similar size as the computational mesh. for example. where S(φ0 ) is a sign function with value 1 in the ﬁrst ﬂuid and value −1 in the other (see.2(b).6) φ(x) = tanh Since the reinitialization equation (4.5) is given by x −α .7) The factor two comes from the fact that the function switches from −1 to +1. However. 4. A disadvantage of the formulation with a smoothed color function is that the evaluation of curvature is potentially less accurate because it is calculated from a steep proﬁle.5) is posed as a conservation law. (4. since it approximately holds ∇φ ≈ 2nδΓ . there is a transition region where the function smoothly switches from one value to the other. diffusion in direction normal to the interface is balanced by a compressive ﬂux. the steady-state value of (4. A smoothed color function also avoids the introduction of an additional discrete delta function and the problems associated with that procedure. (4.4) to steady state. which necessitates evaluating the discrete version of the force (4. and Zemach [11]. In one spatial dimension with an interface located at x = α. Kothe. Another choice is to use a smoothed color/indicator function as depicted in Fig.5) In this equation. [69]).6) has support on the whole domain Ω. Another drawback of the smoothed indicator function is that the function needs to be deﬁned on the whole domain in order to conserve volume. with positive value in one region and negative value in the other. a conservative implementation of the method is possible [68]. Such a function is (almost) constantly 1 or −1 away from the interfaces. This scheme for evaluating the interface force corresponds to the continuous surface tension model proposed by Brackbill. Moreover. Around the interface. though. the discrete delta function underlying (4. These algorithms enforce the property |∇φ| = 1. This function φ can be calculated from the reinitialization equation (cf. by solving the partial differential equation φτ + S φ0 |∇φ| − 1 = 0. discretizations of the reinitialization scheme do not preserve mass.g. Olsson and co-workers [67]. This generally increases resolution requirements compared to the signed distance function approach.)initialize the signed distance function.1) everywhere. This can be alleviated by the use of adaptive meshes that have low resolution away from the interface. 17 .

assuming a continuous PDE model for ﬂuid dynamics that does not take processes on the level of atoms and molecules into account. The considerations are based on the free energy of a molecule. The direct inclusion of contact line dynamics is different from the other approaches in computational multiphase dynamics. The system seeks to minimize the energy (4. cf.4. The so-called van der Waals hypothesis [90] states that immiscible ﬂuids actually mix on molecular level. [76]. which results in ﬂux-boundary conditions on the chemical potential ψ. The proﬁle of the interface is given by a balance of random molecular motion and molecular attraction. formulated by Anderson et al. Boyer [9] extended the phase ﬁeld model to describe motion of ﬂuids with different densities and viscosities. which is the same form as for the continuous surface tension model. i. The phase ﬁeld model can be augmented in order to include contact forces of the interface with solid walls [47]. Similar to the smoothed color function (4. the function seeks to attain the two equilibrium values C = ±1.10) can be rewritten as κnδΓ ≈ ψ∇C .6).1) for interfacial tension.e. Equation (4. a term u · ∇C . forming a diffuse interface. and formulates an interface force κnδΓ ≈ −C ∇ψ + ∇P.10) where P = C ψ. given by α f (C ) = βΨ(C ) + |∇C |2 . where additional mechanisms as slip-velocities and near-wall interface diffusion need to be added. The phase ﬁeld method.7) multiplied by curvature κ. [92].3 Phase ﬁeld models All the methods discussed above are based on the mathematical model (4. The numerical approximations aim at discretely reproducing the effect of a delta function. (4. as devised by Cahn and Hilliard [14]. This process is described by the Cahn–Hilliard equation C t = ∇ · M ∇ψ = ∇ · M ∇ βΨ (C ) − α∇2C = 0.9) with the chemical potential ψ and the mobility factor M . and evaluating the curvature to determine the strength of the surface tension. (4. includes transport of the interface by convection. A different approach takes chemical considerations of the interface into account. 18 .. equation (4. The gradient term ∇C on the other hand enforces a smooth transition. (4.8). which is driven by the double-well potential Ψ(C ). [2] and Jacqmin [47].8) 2 where C denotes the concentration and Ψ(C ) = (C + 1)2 (C − 1)2 . This approach allows for setting static contact angles that correspond to physical wetting properties. that is.

accurate results require sophisticated numerical approximations and ﬁne computational meshes. Even though the interface width α/β can be chosen considerably larger than molecular dynamics would suggest (see Villanueva and Amberg [91]).A considerable challenge in the simulation of phase ﬁeld models are the high resolution requirements. the equations are nonlinear and contain fourth order derivatives. see for example the work by Kay and Welford [54]. 19 . Furthermore. Several attempts have been made to make this approach more competitive.

.

Our method uses a plain aggregation scheme that is part of an aggregation-based algebraic multigrid implementation [26].2 Paper II Paper II extends and analyzes the method presented in Paper I. Summary of Papers 5. In this paper it is proposed to use a second-order accurate generalized-alpha time integration scheme (instead of the Crank–Nicholson scheme in Paper I). Projective implementations retain low frequencies.1 Paper I Paper I proposes a particular implementation for the scale separation in the variational multiscale large eddy simulation. even though a high sensitivity of the results on stabilization parameters is observed. Moreover. The generalizedalpha scheme is a variant of the one-step theta schemes that includes approximations of the ﬁrst temporal derivative as variables and enables control of the dissipation for high frequencies. The method is tested for some standard sample problems for turbulent ﬂow. a clear difference between projective and non-projective scale separations is identiﬁed. [87]. This makes the method also applicable for discretizations based on fully unstructured grids. no grid hierarchy is needed in our implementation. This study shows that our implementation of the variational multiscale LES model provides better results than other state-of-the-art LES schemes. whereas non-projective implementations also damp low frequencies. as was previously observed in numerical computations [29]. We apply our method to turbulent ﬂow around a square cylinder. 5. where the results are better than those of other LES schemes at low resolutions. The analysis considers details of the subgrid viscosity implementation based on plain aggregation scale separation. A grid reﬁnement study is performed for a turbulent channel ﬂow with DNS data as reference [64]. showing that the additional model does not affect consistency of the overall Navier–Stokes scheme. Unlike for geometric multigrid approaches that have been used earlier. The focus of the method is the improvement of LES at low resolutions. 21 . The subgrid viscosity term is implemented within a parallel ﬂuid ﬂow solver based on linear interpolation ﬁnite elements and uses Trilinos solver algorithms [41].5. which is favorable in low-resolution computations.

The study is performed by applying a complete Navier– Stokes/two-phase ﬂow solver to the problem and recording spurious velocities. Such a simulation is. the paper formulates the Cahn–Hilliard-speciﬁc terms as an extension of the level set model. [3]. The model is discretized with the ﬁnite element method and implemented based on the deal.II library [5]. see. As the discussion in Section 4. the conservative level set method [68] and the Cahn–Hilliard model have a similar equation structure. In this case the force due to surface tension should ideally be balanced by a jump in pressure. in order to further evaluate the accuracy of the force balance for ﬁnite element discretizations. A continuous switch function is used in order to retain a consistent model. there are no imbalances in the scheme. 5. 22 . However. A circular bubble at rest is considered as a model problem. the use of higher order elements for the level set function has beneﬁcial impact on the accuracy of curvature. We perform the tests using two different ﬂow solvers. However. A test on a channel ﬂow driven by inﬂow illustrates the beneﬁcial effects of the equation coupling. e.4 Paper IV This paper analyzes inaccuracies of the numerical implementation of the conservative level set model [68]. a standard level set method fails in predicting phenomena relevant in wetting where the ﬂuid-ﬂuid interface aligns at some oblique angle with solid walls. so that the numerical model is still more accurate. in case equal order elements are chosen for approximating pressure and the level set function. however. Therefore. We observe that the velocity errors can be traced back to inaccuracies in the numerically calculated curvature that enter the continuous surface tension model. but restricts that additional phase description to a region close to contact points.3 suggests. The results for a fully coupled implicit method and a decoupled (projection) solver are similar. Moreover. Choosing the element order of the level set function higher than the one for pressure introduces an additional imbalance. If the exact curvature is prescribed. which shows that choice of the solver does not affect the balance of forces. [25].3 Paper III The paper is based on the observation that level set methods enable efﬁcient implementations of two-phase ﬂow physics at relatively coarse meshes because they are based on simple advection schemes together with a reinitialization. This showed that the variation of the curvature along the ﬁnite width of the interface has the largest impact on pressure accuracy.g. we study the effect of the continuous representation of surface tension on the prediction of the jump in pressure between the inside and the outside of the bubble. possible with a phase ﬁeld model.5.

6. the resolution requirements of the two methods should be quantiﬁed in order to identify the precise beneﬁts of the hybrid method. like. which can be used to determine values for the permeability of that conﬁguration. a parallel version for up to about 50 processors has already been implemented by the author of this thesis. It will be necessary to perform an in-depth convergence study of level set models and phase ﬁeld models on several test cases in two-phase ﬂow dynamics. We also aim at using our models for larger problems. Future Work The method proposed in Paper III aims at providing phase-ﬁeld-like features without having to use all its terms everywhere. This is done so that the phase ﬁeld method is reduced to a level set formulation in regions far away from contact lines. The hybrid level-set/phase-ﬁeld method is a potential tool for improving the simulation of multi-phase ﬂow in porous media [17]. a study regarding error sources in the discrete approximation with the conservative level set implementation is presented.II/step_32. parallel implementations of the ﬂow solver and the coupled level-set/Navier–Stokes system are necessary. The permeability tensor is one of the main input parameter in coarse-scale simulations based on Darcy’s law. A ﬁrst step in that direction is to apply the new model to more complicated geometries. In Paper IV. the analysis needs to be performed in a way to correctly reﬂect the discretized state of the equations.dealii. One such direction concerns 1 See the deal.g.org/developer/doxygen/deal. it is possible to use parallel assembly and linear algebra routines. Small scale results are a helpful tool in the simulation of subsurface ﬂow like the prediction of the ﬂow in oil and gas reservoirs. the equations have been coupled by a smoothly varying switch function.II tutorial program http://www. which has up to date only been veriﬁed numerically. For a Stokes problem coupled to an advection equation. These results point out directions where the method needs to be improved.II software that was used in Papers III and IV. Within the deal. In particular. channels with oscillating walls or ﬂow around small obstacles. e. The imposition of a pressure gradient on a small-scale material conﬁguration induces a ﬂow ﬁeld.html 23 .1 The objective is to apply this implementation framework to the coupled Navier–Stokes/level-set system and to eventually extend it to a massively parallel implementation for even larger systems. In our hybrid model. For this purpose. The inﬂuence of the switch function needs to be considered in an energy estimate in order to ensure stability of the coupling of the equations.. Since the numerical implementation of the equations is based on a system of two equations to avoid the direct appearance of fourth derivatives. especially three-dimensional simulations.

though. Similarly. An alternative would be to coarsen the level set mesh in regions away from the interface. When keeping the interface width constant. increasing the order of approximation considerably increases resolution of the interface. This coupled model behaves differently than large-scale level set model. A heterogeneous multiscale approach has been proposed by Ren and E [75] for modeling the interaction of macro-scale ﬂuid ﬂow by micro-scale information from molecular dynamics. A particularly easy approach is to use higher order ﬁnite elements for representing the level set function. only the momentum and continuity balances need to be represented accurately.the accuracy in the curvature calculations. However. accuracy of normals and curvature that rely on the interface representation. this requires a special implementation for the evaluation of the surface tension force in the ﬁnite element algorithms because the variables are related to different meshes. 24 . It is conceivable that a similar interaction could be applied for a microscale simulation done with the phase ﬁeld model. The response on molecular level provides information about the expected motion of contact lines during one time step. This is reasonable since in these remote regions. which would assume zero velocity. Such a strategy increases the costs for the level set portion of the algorithm if the same mesh is used as for the Navier– Stokes equations. The difﬁculty of such a micromacro coupling is to ﬁnd appropriate transfer operations from the micro scales to the macro scales. and thus. while the level set equation does not contain any additional information. the interface region can be resolved by more mesh points than there are used for the Navier–Stokes equations. Suitable boundary conditions for the Cahn–Hilliard subproblem given the solution of a macro-scale level set model are needed. but to use the phase ﬁeld model as a micro-scale input to a macro-scale formulation based on a level set description of two-phase ﬂow. as well as a shear velocity for the macro-scale Navier–Stokes equations. even though a small imbalance in the force representation is introduced. This improvement should be advantageous for the overall accuracy. An alternative hybrid approach that might be taken into account is to not mix the level set formulation with the phase ﬁeld formulation as has been done in Paper III. based on the results from a micro model.

This work was supported by the Graduate School in Mathematics and Computing (FMB).Acknowledgments I am very grateful to my adviser Gunilla Kreiss for introducing me to this interesting subject and for stimulating discussions on various aspects of multiphase ﬂow. I would like to thank Katharina Kormann and Eddie Wadbro for countless discussions on computational mathematics as well as proof-reading manuscripts of my articles and this thesis. I acknowledge discussions with Wolfgang Bangerth on implementation aspects of ﬁnite element programming. My thanks go to Volker Gravemeier who acquainted me with the subject of variational multiscale large eddy simulation for turbulent ﬂow and stimulated ideas. 25 . Finally.

.

J. Cortez. U. V.. Analysis of a one-dimensional model for the immersed . uniqueness and approximation of saddle-point problems arising from Lagrangian multipliers. 1992. 1974. C. 2002. 30:139–165. Anderson. Diffuse-Interface Methods in Fluid Mechanics. Sethian. & Fluids. J. Mech. A. Kothe. [2] D. Fluid Mech. Accurate Projection Methods for the Incompressible Navier–Stokes Equations. 2007. J.. 27 . Brezzi. Informat. Lube. P Beyer and R. G. Dipartimento di Matematica Applicata. Babuška. 195:6239–6257. 2004. and C. Brackbill. A theoretical and numerical model for the study of incompressible mixture ﬂows. A Continuum Method for Modeling Surface Tension.. Numer. [4] I.. Numer. Comput. Phys. A novel representation of the surface tension force for two-phase ﬂow with reduced spurious currents. and G. Aulisa. Bazilevs. A. Methods Appl. Berselli. Leveque. Math. Technical Report 12. and G. deal. and R. E. [3] E. Error-bounds for ﬁnite element method. Hughes. L. J. On the existence. Rouge.. Calo. R.. 16(4):322–333. 24. 2001. Stabilized ﬁnite element methods for the generalized Oseen problem. S. J. boundary method. 29(2):332–364. Comput. Anal. [7] L. A. and M. Phys. Scardovelli. Engrg. B. 168(2):464–499.II — a General Purpose Object Oriented Finite Element Library. John. On Communtation Errors in the Derivation of Space Averaged Navier–Stokes Equations.. Braack. 31:41–68. Softw. 1995. Manservisi. 1992. Bangerth. Reali. and A. Mech. Recherche Opérationelle Sér. M. Variational multiscale residual-based turbulence modeling for large eddy simulation of incompressible ﬂows. Comput. Scovazzi. 33(4):article no. R. A. [6] Y. ACM Trans. Mech. Burman. [9] F.. Minion. Phys. Methods Appl. Cottrell. 118(2):269–277. R. Hartmann. Comput. Engrg. Rev. A Fast Level Set Method for Propagating Interfaces. Annu. [5] W. 196:853–866.Bibliography [1] D.. Française Automat.. B. 1971. [12] F. Rev. Brown. Comput. Wheeler. Adalsteinsson and J. Zemach. Comput. [10] M. V. J. D. 1998. Comput. 197:173–201. Università di Pisa. 100:335–354. SIAM J. 8(R-2):129–151. Math. 2006. and Kanschat G. [8] R. T. Boyer. 2007. M. [13] D. L. Methods Appl. and V. John. Engrg. 2007. Grisanti. McFadden. [11] J.

and D. Technical Report 2006-2649. J. Numerical Solution of the Navier–Stokes Equations. Elman. Wathen. [21] D. and R. J. Glowinski. Tornberg. The Commutation Error of the Space Averaged Navier–Stokes Equations on a Bounded Domain. Cummins. Wiley & Sons. In G.. Elman. R. J. Interfacial Transport Processes and Rheology. W. 2003. Stoneham. B.. Sicilian. Hu. 2002. 2005. M. A. E. M. Numerical Methods for Fluids (3). 213:141– 173. [16] S. Dendy. In J. Phys. pages 3–1176. and . S. Kothe. Silvester. [23] H.0 Smoothed Aggregation User’s Guide. 2006. 28(2):258–267. Birkhäuser Verlag. volume 3. W. Francois. Finite Elements and Fast Iterative Solvers with Applications in Incompressible Fluid Dynamics.. E. Numer. and W. Hilliard. ML 5. 90(4):665–688. Layton. [17] C. Comput. S. J. Comput. Tsai. and A. The bifurcation of tracked scalar waves. Oxford. [28] R. L. Interfacial Free Energy. 28 . and R. John. Wasan.. Handbook of Numerical . Sala. Math. editors. I. [19] T. 207:28–51. Ciarlet. [15] A. Cambridge University Press. 1991. Oxford Science Publications. Cambridge. Heywood. Glimm. SIAM J. G. 1988. Crowe. Multiphase Flow Handbook. Finite Element Methods for Flow Problems. and G. Chichester. Cahn and J. 9(1):61–79.. J. Dubois. H. [24] B. 1999. Stat. O. [22] H. editors. Phys. Wathen. 2006. Grove. S. Fluids. Williams. A. Siefert. Finite element methods for incompressible viscous ﬂow. 2001. J. Comput. D. R. McBryan. Donea and A. Dunca. J. [18] J. Tryggvason. 2006. 1958. J. pages 53–78. Monitoring unresolved scales in multiscale turbulence modeling. Brenner. Contributions to Current Challenges in Mathematical Fluid Mechanics. CRS Press. Rannacher. 2005. Lendquist. D. Basel. [20] A.. Jauberteau. D. Huerta. A. Tuminaro. Math. and A. C. Amsterdam. 22:745–762. Free Energy of a Nonuniform System. J. F. Butterworth–Heinemann. Silvester. Comput.[14] J. Sci. 1968. Chem. D. J. G. B. Discretization of Dirac delta functions in level set methods. Engquist. volume 9. 2004. Analysis. Edwards. [25] M. [26] M. Temam. 13:1800–1806. Collis. Lions and P G. Performance and analysis of saddle point preconditioners for the discrete steady-state Navier–Stokes equations. 2003. and M. M. Boca Raton. Phys.-K. Dynamic Multilevel Methods and the Numerical Simulation of Turbulence. and M. W. J. V. Sandia National Laboratories. T. J. Phys. Gee. A balanced-force algorithm for continuous and sharp interface surface tension models within a volume tracking framework. P Galdi. [27] J. North-Holland. Chorin.

Mech. 2000. 8(12):2182–2189. [41] M. Long. R. R. G. A. Comput. and K. Chichester. A. An overview of the Trilinos project. Incompressible Flow and the Finite Element Method. 33(3):1205–1256. R. Engrg. R. 1991. John Wiley & Sons. Welch. K. Harlow and J. 132:12–33. 1981. Quartapelle. Hughes. 84(2):175–192. Phipps. Springer-Verlag. E. Wanner. 1998. Phys.. Phys. . [42] C. and G. Chichester. Hughes. L. 2006. Meth. Mech. Quartapelle. Nonstiff Problems. J. 39(1):201–225. Solving Ordinary Differential Equations I. B. Szepessy. R. Comput. . [36] E. 1997. S. A. Tuminaro. Hairer. H. . 1993.. J. J. [37] E. Phys.. Stanley. [40] A. Gunzburger.. volume one: Advection–Diffusion. Comput. Math. NY. volume two: Isothermal Laminar Flow. Kolda. 212:400–435. Thornquist. S. Mineola.. Berlin. Arch. 2000. Methods Appl. D. L. and J. Berlin. 165(1):167–188. Wanner. Anal. 2005. . ACM Trans.. L. Gravemeier. 1990. A projection FEM for variable density incompressible ﬂows. [32] P M. E. R. J. Fluids. M. Lehoucq.. 1989. Sani. Numer. Solving Ordinary Differential Equations II. 31(3):397–423. J. 13(2):249–324. 166:3–24. 2000. 1965.[29] V. A velocity–pressure streamline diffusion FEM. [33] J. The ﬁnite element method: linear static and dynamic ﬁnite element analysis. Dover. J. [39] F. J. Comput. Gresho and R. Volume of ﬂuid (VOF) method for the dynamics of free boundaries. Phys. Harten. Gravemeier. Feijóo. V. SIAM J. G. John Wiley & Sons. Softw. S. The Variational Multiscale Method for Laminar and Turbulent Flow. Finite Element Methods for Viscous Incompressible Flows. [30] V. Quincy. Comput.-B. [43] T. Howle. K. R. [44] T. Phys. R. Springer-Verlag. Methods Appl.. Heroux. 2nd edition. Willenbring. 2000. D. T. Hu. [38] P Hansbo and A. Incompressible Flow and the Finite Element Method. Academic Press. Mazzei. Guermond and L. Calculation of Incompressible Viscous Flow by an Unconditionally Stable Projection FEM. Stiff and Differential-Algebraic Problems. G. . T. 29 . [34] J. P Pawlowski. Engrg. L. [31] P M. J. Hairer and G. H. E. Engrg.. Gresho and R. Bartlett. Scale-separating operators for variational multiscale large eddy simulation of turbulent ﬂows. P Nørsett.. Multiresolution Representation of Data: A General Framework. Hirt and B. Comput. Salinger. The variational multiscale method – a paradigm for computational mechanics. 2006. W. 1996. Nichols. [35] M. Hoekstra. L. Sani. Numerical Calculation of Time-Dependent Viscous Incompressible Flow of Fluid with Free Surface. R. J. Comput. J. Guermond and L. Boston. Williams A.

An Optimal Preconditioner for a Class of Saddle Point Problems with a Penalty Term. Liu. Fluids. John. Reno. Mech. Math. 2006. J. 2008. Sci.-G. Matthies. Engrg.. Rang. E. 199(1):221–259. [53] D. Technical Report AIAA 96-0859. 50(4):321–353. SIAM J. stable and efﬁcient Navier–Stokes solvers based on explicit treatment of the pressure term. Comput. Phys. 1941. J. Comput.. [46] T. [59] L. J. 2005. A preconditioner for the steady-state Navier– Stokes equations.. 2004. J. Engrg. 50:845– 862. I. NV. G. J. [48] V. Jacqmin. Welford. SIAM J. [54] D. [50] V. 24(1):237–256. [49] V. [57] D. 30 . Gordon and Breach Science. SIAM J. Ladyzhenskaya. J. A. Meth. 2000. Balestra. B. Comput. John. L. Efﬁcient Numerical Solution of Cahn–Hilliard–Navier– Stokes Fluids in 2D. Kay. Comput. 1986. Int. S. A comparison of time-discretization/linearization approaches for the incompressible Navier–Stokes equations. Brock. Sci. 2002. Sci. Kothe. SIAM J. L. In Proceedings of the 34th Aerospace Sciences Meeting and Exhibit. A. 1998. On large eddy simulation and variational multiscale methods in the numerical simulation of turbulent incompressible ﬂows. for computational ﬂuid dynamics: V. John. and J. 19(2):540–552. A ﬁnite element variational multiscale method for the Navier– Stokes equations. 2003. 59:85–99. Johnston and J. R. and K. 2007. Proceedings of the USSR Academy of Sciences. D. The Local Structure of Turbulence in Incompressible Flows (in Russian). John and S. Loghin. 195(44-47):5995–6010. 29(6):2241–2257. Klawonn.. Mazzei. [56] A. Comput. Comput.. 1999. Calculation of two-phase Navier–Stokes ﬂows using phase-ﬁeld modeling. Wathen. Mosso. [55] A.. Comput. Kolmogorov. On the efﬁciency of linearization schemes and coupled multigrid methods in the simulation of a 3D ﬂow around a cylinder. [52] H. New York. Mech. Sci.. Numer. Comput. An immersed interface method for incompressible Navier– Stokes equations. Hughes. Comput. Rider. Methods Appl. Visual. 2006. [51] V. SIAM J. Sci.. Lee and R. 1969. 30:299–303.. [58] O. The mathematical theory of viscous incompressible ﬂow. and M. and A.. Volume tracking of interfaces having surface tension in two and three dimensions. Methods Appl. S. Hughes. N. Circumventing the Babuska–Brezzi condition: A stable Petrov–Galerkin formulation of the Stokes problem accomodating equalorder interpolations. LeVeque. W. J. Kay and R.. P Franca. Comput. Kaya. J. and J. 25(3):832–856. Hochstein. Phys. A. 155(1):96–127. R. 3:47–59. 26(5):1485–1503.[45] T. Large Eddy Simulation and the variational multiscale method. A new ﬁnite element formulation . [47] D. Accurate. 1996. Sci. R. Jansen. Appl.

Stanford. pages 330–340. Technical report. 2005. Renardy. Enschede. Direct numerical simulation: a tool in turbulence research. Osher and J. 2001. Turbulent Flows. Fronts Propagating with Curvature-Dependent Speed: Algorithms Based on Hamilton–Jacobi Formulations. 1999. 31 . Mahesh. D. Noh and P Woodwards. 1988. Numer.. Oxford University Press. Li. 25(3):220–252. [75] W. Phys. [61] R. Pope. Immersed interface method for Stokes ﬂow with elastic boundaries or surface tension.. 1996. J. Basel.. 2000. [73] S. LeVeque and Z. S. Springer-Verlag. New York. 1977. LeVeque and Z. Phys. 171(1):243–263. [65] W. 79(1):12– 49. and N. Oxford. 1994. The immersed interface method for elliptic equations with discontinuous coefﬁcients and singular sources. Phys. Olsson. Phys. Cambridge. 1993.. and J. Oberai.. A. 210:225–246. Comput. SIAM J. 2002. [70] S. . 2003.. 30:539–578. Quartapelle. N. 1976. Mansour. Sethian. Comput. Comput. C. 1997.-L. Annu. Comput. 31(4):1019– 1044. Transfer of Energy in the Variational Multiscale Formulation of LES. A conservative level set method for two phase ﬂow. Phys. [69] S. Kim. [76] M. SLIC (simple line interface calculation). Comput. Numerical Simulation of Moving Contact Line Problems Using a Volume-of-Fluid Method. A. Rev... Direct numerical simulation of turbulent channel ﬂow up to Reτ = 590. [62] P . 225:785–807. [64] R. J. A conservative level set method for two phase ﬂow II. 204(1):1–26. S. Anal. E. Zahedi. V. G. J. B. [67] E. P Fedkiw. Level Set Methods and Dynamic Implicit Surfaces. J. J. Gravemeier. J. Sci. Li. Olsson and G. volume 1: Incompressible Models. The immersed boundary method. volume 113 of International Series of Numerical Mathematics. SIAM J. Peskin. Kreiss. [74] L. Phys. 11(4):943–945. 1998. [68] E. J. Renardy. 2007. Comput. Fluid Mech. Li. Mathematical Topics in Fluid Mechanics. Fluids. Summer Program 2004. J. Burton. [63] P Moin and K. Mathematical Sciences. Cambridge University Press. Acta Numerica. and G. Moser. Peskin. Phys. Springer. Osher and R. [72] C.[60] R. of the 5th International Conference on Numerical Methods in Fluid Dynamics. 2005. Numerical Analysis of Blood Flow in the Heart. [71] C. Ren and W. Numerical solution of the incompressible Navier–Stokes equations. Lions. J. Kreiss. [66] A. 11:479–517. 18(3):709–735. In Proceedings . Y. and S. Center for Turbulence Research. Heterogeneous multiscale method for the modeling of complex ﬂuids and micro-ﬂuidics.. Comput. Applied . Birkhäuser-Verlag.

Schüller. Springer. [82] C. [80] J. Teman. [86] G.. Some generic capillary-driven ﬂows. [79] J.. Fluid Mech. N.. 2000. Comput. Efﬁcient Solvers for Incompressible Flow Problems: An Algorithmic and Computational Approach. 2006. Smagorinksy. Jan. E. 1979. Multiphase Flow. 28:1–44. Unverdi and G. Super Computing 2000 Proceedings. Evolving Interfaces in Computational Geometry. D. and A. A. Parallel smoothed aggregation multigrid: aggregation strategies on massively parallel machines. Springer-Verlag. Bull. Direct Numerical Simulation of Free-Surface and Interfacial Flow.. In J. The thermodynamic theory of capillarity under the hypothesis of continuous variation of density. Al-Rawahi. 2009. and S. A level set method for computing solutions . Han. 31:567–603.. Stabilized Finite Element Formulations for Incompressible Flow Computations. Computer Vision. Phys. 114(1):146–159. J. Math. Comput. 20(2):200–244. 1994.[77] P Sagaut. Akad. and G. J. 169(2):708–759. Juric. O. Multigrid. [81] M. Esmaeeli. 100(1):25–37. 1973. Sethian. [89] S. 32:1072–1086. Comput. [83] R. Amsterdam. Sussman. Tryggvason. Soc.. Tezduyar. 1968. [90] J. Scardovelli and S. A. S. D. Villanueva and G. 2000. Mech. the ﬁnite element technique. 1(1):73–100. Tong. A conservative level set method for contact line dynamics. Gustavsson. and Y. [88] S. Weten. Phys. Cambridge University Press. 1992. General Circulation Experiments with the Primitive Equations. Appl. Taylor and P Hood. 1992. Turek. 96:115–152. J. 228(17):6361–6375. Une méthode d’approximation de la solution des équations de Navier– Stokes. . B. Level Set Methods and Fast Marching Methods. 1893). Phys. Rev. 2006. van der Waals. Osher. W. Large Eddy Simulation for Incompressible Flows. Stat. Kreiss. France. (originally published in Dutch in Verhandel. [92] S. C. Phys. Nas. J. Tauber. J. Zaleski. Fluids. Phys. I.. K. Elsevier Academic Press. Int. Fluid Mechanics. Multi-ﬂuid Flows. Donnelley. Tryggvason. and Materials Science. Monthly Weather Rev. 3rd edition. 91(3):99–164. editor. 1999. The Basic Experiment. [87] R. 32 . Berlin. J. [91] W. Tuminaro and C.-J. to incompressible two-phase ﬂow. P Smereka. [78] R. 2001. A Front Tracking Method for the Computations of Multiphase Flow. Oosterlee. [84] T. J. Zahedi. A Front Tracking Method for Viscous. Incompressible. Ann. 1963. Trottenberg. Comput. Adv. Amberg. [85] U. Konink. 1999. Bunner. 2001. London. Berlin. A numerical solution of the Navier–Stokes equations using . Comput.

.

Organizational Change and Sensemaking when Introducing User-Centred Systems Design in Public Authorities Joakim Eriksson: Detailed Simulation of Heterogeneous Wireless Sensor Networks Andreas Hellander: Numerical Simulation of Well Stirred Biochemical Reaction Networks Governed by the Master Equation Ioana Rodhe: Query Authentication and Data Conﬁdentiality in Wireless Sensor Networks Mattias Wiggberg: Unwinding Processes in Computer Science Student Projects Bj¨ rn Halvarsson: Interaction Analysis and Control of Bioreactors for Nitrogen o Removal Mahen Jayawardena: Parallel Algorithms and Implementations for Genetic Analysis of Quantitative Traits Olof Rensfelt: Tools and Methods for Evaluation of Overlay Networks Thabotharan Kathiravelu: Towards Content Distribution in Opportunistic Networks Jonas Boustedt: Students Working with a Large Software System: Experiences and Understandings Manivasakan Sabesan: Querying Mediated Web Services Stefan Blomkvist: User-Centred Design and Agile Development of IT Systems ˚ Asa Cajander: Values and Perspectives Affecting IT Systems Development and Usability Work Department of Information Technology.Recent licentiate theses from the Department of Information Technology 2009-004 2009-003 2009-002 2009-001 2008-003 2008-002 2008-001 2007-006 2007-005 2007-004 2007-003 2007-002 2007-001 2006-012 2006-011 Katharina Kormann: Numerical Methods for Quantum Molecular Dynamics Marta L´ rusd´ ttir: Listen to Your Users .The Effect of Usability Evaluation on a o Software Development Practice Elina Eriksson: Making Sense of Usability . Uppsala University. Sweden .

- Hybrid CFD-CAA simulation of a heated and unheated Cylinder in turbulent cross ows
- AIAA-2013-2567-charest
- BulkVisc
- Large Eddy Simulation of Channel Flow using Wall Functions
- ANALELE UNIV BUC 2001
- Computational Methods for Multi Phase Flow
- Shliomis_Ferrohydrodynamics Testing a Third Magnetization Equation_ 2001
- Mesh DiscretIzation
- 9783540768326-c2.pdf
- Smith Woodruff Re Normalization Group Analysis of Turbulence Ann Rev Fluid Mech 1998
- Physics of Fluids
- Smoothed particle idrodynamics
- Lecture notes on GDynamics
- CFD report
- fluidsToc
- Me2204 Fmm 2011 ANNA UNIVERSITY Apr May Qp
- Using Moody Diagram
- 92675983-Rheology-Book.pdf
- Notes
- Aerodynamics II Notes
- Interfacial Mixing in Viscous Pipe Flows
- Numerical Simulations - Examples and Applications in Computational Fluid Dynamics
- Saint Ill An 2010 the Dilute Rheology of Swimming Suspensions-A Simple Kinetic Model
- Viscosity Conversion Guide
- fl3-navier-stokes
- Modiﬁed Chaplygin Gas Cosmology with Bulk Viscosity 1401.8002
- Nash Notes on Fluids
- Hadamard 1911 English Translation
- 20100001729_2010001521
- 142lett1
- 2009-005

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd