Modeling of Heat Transfer in Two-Phase

Flow Using the Level-Set Method

Magnus Aashammer
Gjennestad

Master of Science in Physics and Mathematics

Submission date: June 2013
Supervisor:
Jon Andreas Stvneng, IFY
Co-supervisor:
Svend Tollak Munkejord, SINTEF Energi AS

Norwegian University of Science and Technology

Department of Physics

The cover photo is by Nir Schneider and it is licensed under the Creative
Commons Attribution 2.0 Generic license.
An electronic version of this document (navigable and in color) should be
available at http://mastersthesis.aashammer.net/thesis.pdf.

Abstract
The implementation of a two-phase flow model, developed at SINTEF Energy Research and relying on the level-set method, was extended through
the discretization and implementation of an advection-diffusion equation
for temperature and a Boussinesq coupling between the temperature and
velocity fields. In two-phase flow, both the continuum surface force method
and the ghost-fluid method was employed for handling jumps at the interface. Results from simulated cases indicated that the implementation for
both single- and two-phase flow with the ghost-fluid method was correct,
with second- and first-order convergence, respectively.
Also, a model for phase transition was implemented to allow for vaporization and condensation mass transport between the phases. Results from
simulated one-dimensional cases indicated that the implementation of this
model was correct in one dimension, with first-order convergence. These
results from one dimension and the qualitatively correct results from two
dimensions gave reason to believe that the implementation was correct
also in two dimensions.
Through the introduction of heat-transport physics in the implementation of the two-phase flow model, this implementation has been developed
in direction of performing more detailed simulations that are relevant for
natural gas liquefaction processes.

Sammendrag
Implementasjonen av en modell for tofasestrmning, som er utviklet ved
SINTEF Energi og basert p
a level set-metoden, ble utvidet gjennom
a
diskretisere og implementere en adveksjons-diffusjonsligning for temperaturen og en Boussinesq-kobling mellom temperatur- og hastighetsfeltene.
For tofasestrmning ble b
ade continuum surface force-metoden og ghost
fluid-metoden brukt for
a h
andtere grenseflatehopp. Resultatene fra simuleringene som ble utfrt tydet p
a at implementasjonen for b
ade en- og
tofasestrmning med ghost fluid-metoden var korrekt, med henholdsvis
andre- og frsteordens konvergens.
I tillegg ble en modell for faseovergang implementert for
a kunne ta
hyde for massetransport som flge av fordamping og kondensering mellom fasene. Simuleringsresultater fra endimensjonale problemer tydet p
a
at implementasjonen var rett i en dimensjon, med frsteordens konvergens.
Disse resultatene og de kvalitativt rette resultatene fra to dimensjoner ga
grunn til
a tro at implementasjonen var korrekt ogs
a i to dimensjoner.
Gjennom
a inkludere varmetransportfysikk i implementasjonen av tofasemodellen har denne blitt utviklet i retning av kunne utfre mer detaljerte simuleringer som er relevante for flytendegjring av naturgass.

iii

Preface
This work will be submitted as a thesis for the degree Master of Science
in Physics and Mathematics at The Norwegian University of Science and
Technology.
I have had the privilege of conducting this work under the supervision
of Chief Scientist Svend Tollak Munkejord at SINTEF Energy Research,
whose experience, insight, advice and feedback have all been indispensable
to me, and Associate Professor Jon Andreas Stvneng at The Department
of Physics at The Norwegian University of Science and Technology, who
has been very helpful and has always found the time to answer my questions. I am most grateful to both of them.
Through the supervision by Svend Tollak Munkejord, this thesis is part
of the Enabling Low-Emission LNG Systems project at SINTEF Energy
Research and The Norwegian University of Science and Technology. I want
to acknowledge contributions from Statoil, GDF SUEZ E&P Norge and
the Petromaks program of the Research Council of Norway (193062/S60).
I wish to thank PhD candidate Karl Yngve Lerv
ag for good advice and
for lending me his Python scripts for parsing .tec-files and I must thank
my office mate MSc Eskil Aursand for productive discussions, comments
and coffee breaks and for challenging me at explaining concepts I may not
have understood as well as I initially thought. Also, I am very grateful to
my brother Morten and my dear Ida for proofreading and, especially to
Ida, for all encouragement and support.
Of course, in spite of all assistance, the responsibility for any errors in
this work, both typographic and more severe, is mine.

Magnus Aashammer Gjennestad

Trondheim, June 2013

Nomenclature
Latin letters
~a
cp
d
d0
Eabs
Ess ()
E1
f~b
f~s
f~sfd
~g
H

H
h
k
L
m
~
N
n
n

p
Pr
~q
Ra

Intermediate vector field, see (3.2)

Specific heat capacity at constant pressure
Signed Euclidean distance-function, see (3.12)
Euclidean distance-function, see (3.13)
Absolute error, see (3.165)
Upper bound on simulated t , see (3.170)
Scaled 1-norm error, see (3.166)
Body force density
Singular interfacial force density
Interfacial force density
Gravitational acceleration
Heaviside function
Smeared Heaviside function, see (3.57)
Specific enthalpy
Interface curvature
Fluid domain length
Mass flux density
Number of grid cells
Order of convergence, see (3.167), and time step
number
Unit normal vector
Pressure
Prandtl number, see (4.2)
Heat flux density
Rayleigh number, see (4.1) and (4.37)

m s2
J kg1 K1
m
m

Nm3
Nm3
Nm3
m s2

J kg1
m1
m
2
1
kg m s
1
1
1
Pa
1
2
Wm
1

vii

S
T
t
t
~u
u
V
v
w
~
~x
x

y
z

Smeared sign function, see (3.17)

Temperature
Time
Unit tangential vector
Velocity
x-component of velocity
Viscous stress tensor
y-component of velocity
Interface velocity
Position vector
Unit vector in x-direction
Unit vector in y-direction
Unit vector in z-direction

1
K
s
1
m s1
m s1
Pa
m s1
m s1
m
1
1
1

Dummy section
Greek letters

viii

Thermal diffusivity
Coefficient of thermal expansion
Interface between phases
Dirac and Kronecker delta function
Interface location parameter
Cell division parameter
Threshold cell division parameter
Thermal conductivity
Dynamic viscosity
Kinematic viscosity
Second viscosity
Density
Coefficient of interfacial tension
Pseudo time and generic integration variable
Level-set function

m2 s1
K1

1
1
W m1 K1
Pa s
m2 s1
Pa s
kg m3
N m1
m

Stream function
Fluid domain
Boundary of fluid domain

m2 s1

Norms and Operators

[]

J (, )
i
n

t
kk1
kk2
()T

Difference operator (across interface)

Divergence operator
Gradient operator
Laplacian operator
Curl operator
Binary differential operator, see (4.22)
Partial derivative operator with respect to spatial coordinate i
Partial derivative operator, where the derivative
is taken in the direction of some normal vector
n

Partial derivative operator with respect to time

1-norm
2-norm (Euclidean norm)
Transpose

Abbreviations
BC
CD
CFL
CSF
ENO

Boundary condition
Central differences
CourantFriedrichsLewy
Continuum surface force
Essentially non-oscillatory

ix

FUP
GFM
LNG
MAC
ODE
PDE
RK
SI
SSP
WENO

First-order upwind
Ghost-fluid method
Liquefied natural gas
Marker-and-cell
Ordinary differential equation
Partial differential equation
RungeKutta
International system of units
Strong-stability-preserving
Weighted essentially non-oscillatory

Contents
Abstract

Sammendrag

iii

Preface

Nomenclature

vii

Contents

xi

1 Introduction
1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Survey of Thesis . . . . . . . . . . . . . . . . . . . . . . . .

1
2
3

2 The
2.1
2.2
2.3
2.4
2.5

Mathematical Model for the Two-Phase Flow Physics

The NavierStokes Equations . . . . . . . . . . . . . . .
The Advection-Diffusion Equation for Temperature . . .
The Boussinesq Approximation . . . . . . . . . . . . . .
A Model for Phase Transition . . . . . . . . . . . . . . .
The Jump Conditions . . . . . . . . . . . . . . . . . . .
2.5.1 Two-Phase Immiscible Flow . . . . . . . . . . . .
2.5.2 Two-phase Flow with Phase Transition . . . . .
2.6 Cartesian vs. Cylindrical Coordinates . . . . . . . . . .
2.7 Boundary Conditions . . . . . . . . . . . . . . . . . . . .
2.7.1 Velocity Boundary Conditions . . . . . . . . . . .
2.7.2 Pressure Boundary Conditions . . . . . . . . . .
2.7.3 Temperature Boundary Conditions . . . . . . . .
2.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

5
5
7
8
10
12
12
13
14
15
16
16
17
17

xi

3 Numerical Methods
3.1 The Projection Method . . . . . . . . . . . . . . . . . . .
3.1.1 Solving the Poisson Equation . . . . . . . . . . . .
3.2 The Level-Set Method . . . . . . . . . . . . . . . . . . . .
3.2.1 Reinitialization . . . . . . . . . . . . . . . . . . . .
3.2.2 Velocity Extrapolation . . . . . . . . . . . . . . . .
3.3 Time Integration with Strong-Stability-Preserving Runge
Kutta Methods . . . . . . . . . . . . . . . . . . . . . . . .
3.4 Spatial Discretization . . . . . . . . . . . . . . . . . . . .
3.4.1 Discretization of the Advective Operator . . . . . .
3.4.2 Interface-Capturing Methods . . . . . . . . . . . .
3.5 Numerical Strategy for Handling Phase Transition . . . .
3.5.1 Computing the Mass Flux Density from the Temperature Field . . . . . . . . . . . . . . . . . . . .
3.5.2 Enforcing the Interface Conditions on the Velocity
3.5.3 Computing the Interface Velocity . . . . . . . . . .
3.6 Boundary Conditions . . . . . . . . . . . . . . . . . . . . .
3.6.1 Velocity Boundary Conditions . . . . . . . . . . . .
3.6.2 Level-set Function Boundary Conditions . . . . . .
3.6.3 Pressure Boundary Conditions . . . . . . . . . . .
3.6.4 Temperature Boundary Conditions . . . . . . . . .
3.7 Time-Step Restrictions . . . . . . . . . . . . . . . . . . . .
3.7.1 Single-Phase Simulations . . . . . . . . . . . . . .
3.7.2 Two-Phase Simulations . . . . . . . . . . . . . . .
3.8 Estimation of Error and Convergence Order . . . . . . . .
3.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Numerical Experiments
4.1 De Vahl Davis Benchmark Case . . . . .
4.1.1 Numerical Results . . . . . . . . .
4.2 Two-Phase Heat Conduction . . . . . . .
4.2.1 Analytical Solution . . . . . . . . .
4.2.2 Numerical Results . . . . . . . . .
4.3 Wang et al.s Two-Layer Convection Case
4.3.1 Analytical Solution . . . . . . . . .
4.3.2 Numerical Results . . . . . . . . .

xii

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.

.
.
.
.
.

19
19
21
22
24
24

.
.
.
.
.

25
26
27
33
40

.
.
.
.
.
.
.
.
.
.
.
.
.

41
43
44
44
44
46
46
47
48
48
49
51
52

.
.
.
.
.
.
.
.

53
53
54
59
60
61
66
67
73

4.4
4.5

4.6

4.7
4.8
4.9

Lava Lamp . . . . . . . . . . . . . . .
4.4.1 Numerical Results . . . . . . .
Vaporization with Uniform Mass Flux
4.5.1 Analytical Solution . . . . . . .
4.5.2 Numerical Results . . . . . . .
Vaporization with Physical Mass Flux
4.6.1 Analytical Solution . . . . . . .
4.6.2 Numerical Results . . . . . . .
Vaporization of a Drop . . . . . . . . .
4.7.1 Numerical Results . . . . . . .
Boiling Film . . . . . . . . . . . . . . .
4.8.1 Numerical Results . . . . . . .
Summary . . . . . . . . . . . . . . . .

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.

76
77
80
80
81
85
87
88
92
93
96
97
99

5 Conclusions and Suggestions for Further Work

103
5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.2 Suggestions for Further Work . . . . . . . . . . . . . . . . . 104
Bibliography

105

xiii

1 Introduction
During the past few years, liquefied natural gas (LNG) has become an increasingly important product of the Norwegian petroleum industry. Since
the start of extraction in 2007, the Snhvit field alone has yielded 2.0 1010
standard m3 oil equivalents of natural gas, 4.65 109 standard m3 oil equivalents of these in 2012 [Norwegian Petroleum Directorate, 2013]. Due
to Snhvits remote location, natural gas transport by pipeline all the
way to the markets was deemed unfeasible as the pipelines would have to
be too expensive. Instead, the well stream from Snhvit is transported
through a 145 km pipeline to the processing facility at Melkya where it
is cooled down to become LNG, which is more efficiently transported by
ship [Gisvold, 2004].
The liquefaction process is performed in heat exchangers that can liquefy
11 106 kg of natural gas per day [Gisvold, 2004]. A large amount of the
greenhouse gas emissions from production, transport and combustion of
LNG is incurred at the liquefaction stage [Tamura et al., 2001]. Therefore
there is a great potential for enabling a more energy-efficient and less
emission-intense LNG life cycle by optimizing the liquefaction stage. Any
reduction in energy consumption in the processing would also be beneficial
from a financial point of view.
Even though some optimizations can be performed with more conventional engineering approaches, a better and more detailed understanding
of the underlying physical behavior of the fluid streams involved in the gas
liquefaction is essential. To contribute to such an understanding, SINTEF
Energy Research and The Norwegian University of Science and Technology are running the five-year (2009-2014) research program Enabling
Low-Emission LNG Systems [Lvseth, 2013] with industry partners Statoil and GDF SUEZ E&P Norge.
According to Kunugi [2012], there seems to be a general consensus

Still, according to Gisvold [2004], the longest of its kind in the world.

1 Introduction
that direct numerical simulation is one of the most promising approaches
towards clarifying heat transfer characteristics and condensation/boiling
phenomena and discussing their mechanisms. Therefore this thesis aims
to describe and implement the discretization of a transport equation for
temperature into an existing implementation of a two-phase flow model,
developed at SINTEF Energy Research, and thus enable the numerical
study of heat transfer in two-phase flow. Also, the ambition is to describe
and implement a phase-transition model to enable detailed numerical studies of the liquefaction process. Performing such physically realistic LNG
simulations, however, is beyond the scope of the thesis.
The flow model relies on the level-set method to keep track of the location of the interface between the phases under consideration. The level-set
method was first described by Osher and Sethian [1988] and has since been
used in two-phase flow models; both compressible and incompressible [Osher and Fedkiw, 2003]. A review of some of the early advances in the use
of this method to study boiling phenomena is given by Dhir [2001]. In this
thesis, however, we consider a more recent level-set-based phase-transition
model similar to that proposed by Gibou et al. [2007].

1.1 Notation
In this section, we comment on some important points regarding the notation used in this work.
Vector quantities are distinguished from scalars using arrows. The
exception is unit vectors who wear hats. The velocity vector field is
therefore denoted by ~u, whereas the unit vector in the x-direction is
x
and the scalar pressure field is p.
We will use derivative notations in full, e.g.
forms interchangeably.

x ,

and short, e.g. x ,

Superscripts + and most often indicate that the superscripted

quantity is associated with one of the two phases. For instance,
is the density of Phase 1 and + is the density of Phase 2.
When we refer to field values, e.g. of the velocity field, at discrete
temporal or spatial points, we use indices enclosed by parenthesis.

1.2 Survey of Thesis

Subscripted indices are used for spatial indices and superscripted
indices to indicate time step numbers.
We will sometimes use the repeated index summing convention, also
known as Einsteins summing convention, where repeated indices
denote an implicit summation over these indices. For example, in
this notation, i ui is equivalent to ~u.
We will apply the gradient operator to vector fields like ~u with
the understanding that the gradient operator is to be applied in a
component-by-component fashion, producing a second-order tensor
(the Jacobian matrix).
In analogy to the previous point we will also apply the Laplacian
operator 2 to vector fields with the understanding that the operator
is to be applied to each component.
A subscripted g indicates that the subscripted quantity is a ghost
value.

1.2 Survey of Thesis

This thesis is divided into five chapters. Chapter 1 is the introduction that
you have been reading now. In Chapter 2, we present the mathematical
model for the two-phase flow physics whose most significant parts are
the NavierStokes equations, the temperature transport equation and the
phase-transition model. In Chapter 3, we describe the numerical methods
used to solve the model from Chapter 2. To test the implementation and
the methods in Chapter 3, a few test cases were simulated. These test cases
and their results are presented and discussed in Chapter 4. In Chapter 5,
we make some concluding remarks and a few suggestions for further work
with heat-transfer modeling using the SINTEF-developed implementation
of the two-phase flow model.

2 The Mathematical Model for the

Two-Phase Flow Physics
The fluid and heat flow physics that is to be simulated is formulated as
a mathematical model for two-phase immiscible and incompressible fluid
flow and a transport equation for the temperature. The model consists
of a set of partial differential equations (PDEs). They are equations of
motion and describe the time development of the velocity ~u, pressure p and
temperature T fields from a known initial configuration subject to a set
of known boundary conditions. In this chapter, this model is presented.
In addition, we present a model for phase transition. From the temperature field T we calculate the vaporization and/or condensation mass flux
density across the interface and use this to find the correct fluid velocities
at the interface and calculate the velocity of the interface itself.

2.1 The NavierStokes Equations

From the conservation of mass, one of the most fundamental principles
in classical physics, one can derive the continuity equation for a fluids
density ,

+ (~u) = 0.
(2.1)
t
In (2.1), ~u is the fluid velocity. The interested reader is referred to Landau
and Lifshitz [1987, 1] for a derivation of this equation.
We will assume incompressible flow throughout this work, that is we
will assume that
~u = 0.
(2.2)
The continuity equation for the fluid density (2.1) then becomes

+ ~u = 0.
t

(2.3)

2 The Mathematical Model for the Two-Phase Flow Physics

This is an advection equation for the density, indicating that we can have a
spatially varying density under the incompressible flow assumption (2.2),
but that time evolution of the density field is restricted to advection with
the fluid flow. A control surface containing a constant number of fluid
particles may not expand or be compressed. Here we will assume that
is uniform within each phase.
From the principle of conservation of momentum, one can derive the
balance equation for the i-component of a fluids momentum


ui
ui
p
Vik

+ uk
=
,
(2.4)
+ (fb )i +
t
xk
xi
xk
where p is the pressure, (fb )i is the i-component of the body force density
and V is the viscous stress tensor. The derivation of this equation can be
found in Landau and Lifshitz [1987, 15].
According to Landau and Lifshitz [1987, 15], the general form of the
viscous stress tensor V for an isotropic fluid and small velocity gradients
is


ui
uk
2 u`
u`
Vik =
+
ik
+ ik
,
(2.5)
xk
xi
3 x`
x`
where is the dynamic viscosity and is the second viscosity. Since Vik
is a linear combination of derivatives i uk , the fluids we consider may be
classified as Newtonian fluids [White, 2008, Section 4.3].
Under assumption of incompressible flow, i ui = 0 (2.2), the balance
equations for all momentum components (2.4) reduce to




~u

+ (~u ) ~u = p + {~u} + {~u}T + f~b ,

(2.6)
t
in vector notation. This is commonly known as the incompressible momentum balance equation. The form of (2.6) is that of a continuity equation
for the momentum, with the gradient
 of the pressure
 field, the body force
T
density and the viscous term {~u} + {~u} as source terms. It is
worth mentioning that the form of the viscous term, under the assumption
of incompressible flow and constant and uniform viscosity, is identical to

Body forces is the collective name given to all external forces, such as gravity or
electric forces, that are acting on the fluid. In the case where we consider only
gravitational external forces, f~b = ~g .

2.2 The Advection-Diffusion Equation for Temperature

a diffusion term for the momentum and that the process of viscous action
thus can be viewed as diffusion of momentum.
Equations (2.2) and (2.6) together are often referred to as the incompressible NavierStokes equations.
When considering two-phase immiscible flows, we need to model the
interface between the two fluids and the interfacial forces acting on it.
According to Hansen [2005], the interface can be modeled as a material
membrane that flows with the fluids. We will refer to the interface as the
set of coordinates
= {~x | ~x is on the interface} .

(2.7)

The interface separates the fluid domain in two regions + and .

The interfacial forces are modeled by a singular force
Z
~
fs (~x, t) =
f~sfd (~s, t) (~x ~xi {~s}) d~s,
(2.8)

where f~sfd is a force density, is the Dirac delta function and ~xi is a
parametrization of the interface. In this thesis, f~sfd is modeled by
f~sfd = k
n,

(2.9)

where is the coefficient of interfacial tension, k is the local interface

curvature and n
is a unit normal vector to the interface, pointing into + .
The singular interfacial force is added as another source term in the
momentum balance equation (2.6), so that the equation becomes




~u

+ (~u ) ~u = p + {~u} + {~u}T + f~s + f~b . (2.10)

t

2.2 The Advection-Diffusion Equation for

Temperature
In order to model heat transfer, we need a heat-transport equation. In
this section, we rely on an argument from Landau and Lifshitz [1987] that
reduces the general equation for heat transfer to a transport equation for
the temperature and we present this equation.

2 The Mathematical Model for the Two-Phase Flow Physics

According to Landau and Lifshitz [1987, 50], we can neglect the temperature dependence of the fluids viscosity , thermal conductivity and
specific heat capacity cp if the temperature differences in a fluid are small.
We may also suppose that the density is independent of temperature in
the equations of motion and regard the fluid as incompressible in the sense
that i ui = 0 if the fluid velocities are small with respect to the speed of
sound. With these assumptions, and neglecting viscous heat dissipation,
the general equation for heat transfer reduces to an advection-diffusion
equation for temperature
T
1
+ (~u ) T =
(T ) .
t
cp

(2.11)

This equation is derived from the general equation for heat transfer by
Landau and Lifshitz [1987, 50] and more directly from considering a list
of selected physical effects by Griebel et al. [1998, Section 9.2].
The source term on the right hand side of (2.11) is the diffusion term.
It models direct molecular transfer of energy from areas with high temperature to areas with low temperature by a heat flux proportional to the
temperature difference. This term is independent of the fluid velocity ~u
and may lead to heat transfer even when the fluid is macroscopically at
rest.
Within each phase, we will assume that is uniform and we can rewrite
(2.11) as
T
+ (~u ) T = 2 T,
(2.12)
t
where the thermal diffusivity [Griebel et al., 1998, Section 9.2] is
=

.
cp

(2.13)

2.3 The Boussinesq Approximation

Time integration of the temperature transport equation (2.11) alongside
the NavierStokes equations gives a temperature field that evolves in time
subject to diffusion and advection with the fluid flow. What it does not
give is any kind of mechanism through which the temperature can couple
to the velocity field and thus be a driving force of the flow. Physically

2.3 The Boussinesq Approximation

such a coupling would happen through compression and expansion of the
fluid as it is heated and cooled. With a few notable exceptions, fluid
volumes containing a constant number of fluid particles expand, become
less dense and rise when heated, and contract, become more dense and
sink when cooled. In our incompressible flow model, however, we do not
allow such volumes to expand or contract. Therefore we have to couple the temperature to the velocity field by some other means than a
temperature-dependent density, if we want to model temperature-driven
flows. In constructing such a coupling, we will follow a widely used strategy and employ the Boussinesq approximation. Based on Griebel et al.
[1998, Section 9.1], we state it as follows.
The fluid density is constant and uniform within each phase, except
in the buoyancy part of the body force density f~b in the Navier
Stokes equations, where it has a linear dependence on temperature.
All other fluid properties are constant and uniform, though they also
may be discontinuous at the interface.
The linear relationship between density and temperature in the buoyancy
part of f~b is often expressed as
(T ) = (1 {T T }) ,

(2.14)

where and T are a reference density and a reference temperature,

respectively, such that (T ) = , and is the coefficient of thermal
expansion.
When we consider gravitational body forces only, the body force density
becomes f~b = (T )~g . The body force term we will use in the NavierStokes
equations with the Boussinesq coupling is thus
f~b = ~g (1 {T T }) .

(2.15)

Since ~g appears only in the body force term in the NavierStokes equations, we can transfer the temperature dependence of the density in this
term to the gravitational acceleration and still expect the same solutions
for the ~u and p fields. In other words, we can replace the buoyancy force
density (T ) ~g with ~g (T ). The temperature-dependent gravitational
acceleration now becomes
~g (T ) = ~g (1 {T T }) ,

(2.16)

2 The Mathematical Model for the Two-Phase Flow Physics

where ~g is the usual gravitational acceleration constant. This means
that using the Boussinesq approximation presented here is equivalent to
introducing a temperature-dependent gravity, where the temperature dependence is given by (2.16).

2.4 A Model for Phase Transition

Consider now the case where we have two phases, liquid and gas, of the
same substance and there is mass transport, i.e. vaporization and/or condensation, between the two phases. To simulate this situation, we will
use a model very similar to that proposed by Gibou et al. [2007]. In this
section, we present this model.
The general idea is to impose that the interface temperature must be
equal to the saturation temperature of the liquid Tsat . Then we use any
resulting interfacial jump in the component of the conductive heat flux
density ~q = T that is normal to the interface to find the mass flux
density across the interface that is consistent with this jump. Specifically,
when more heat flows into a point on the interface than flows out, we
assume that this heat has been absorbed in the vaporization of some of
the liquid around that point, and vice versa when more heat flows out
than in.
To formalize this idea, we first define the interface velocity w
~ and the
mass flux density across the interface m.
~ The latter is always normal to
the interface, so we can write it in terms of the scalar quantity m as
m
~ = m
n.

(2.17)

We will adopt the convention that the liquid phase occupies and the
gaseous phase occupies + . Since n
is defined as pointing into + , local
mass flux from to + , vaporization, is characterized by a positive m.
Applying mass conservation across the interface gives the following expression for the mass flux density m in terms of the fluid velocities on each
side of the interface and the interface velocity,


m = + ~u+ n
w
~ n
,


= ~u n
w
~ n
.

10

(2.18)
(2.19)

2.4 A Model for Phase Transition

Solving for w
~ n
gives an expression for the normal component of the
interface velocity. Assuming that the tangential component of w
~ is equal
to the tangential component of ~u, we get
w
~ = ~u

m
n
.

(2.20)

From these equations we see that when there is no mass flux m = 0, the
velocity field is continuous at the interface and w
~ is equal to the fluid
velocity at the interface.
Eliminating w
~ from (2.19) and (2.20) gives


~u n
~u n
=m

1
1

+


.

(2.21)

Assuming that ~u t is continuous at the interface, we get

+

~u ~u = m

1
1


n
.

(2.22)

As the interface condition for temperature, we demand that the temperature in both phases should be equal to the saturation temperature,
T + = T = Tsat .

(2.23)

Now, we equate the difference in the normal component of the conductive

heat flux on each side of the interface with the heat needed to obtain a
mass flux density m,




mh = + T + n
T n
.

(2.24)

h is here the specific enthalpy difference (latent heat) associated with

the phase change.
The strategy for modeling phase transition is now to evolve the temperature field T in time with (2.11), subject to the interface condition (2.23).
Equation (2.24) is used to find the mass flux density m from the temperature field, which is again used to find the interface velocity w
~ (2.20) and
enforce the interfacial conditions on ~u (2.22).

11

2 The Mathematical Model for the Two-Phase Flow Physics

2.5 The Jump Conditions

When considering two-phase flow, we will assume that all fluid properties
such as the viscosity , density , thermal conductivity and so on,
are constant and uniform within each phase. They may, however, have
different values in one phase compared to the other. We say that they
jump at the interface and denote their jumps as
[] = + ,

(2.25)

for the viscosity and so on for other fluid and flow properties.
So far, the presented equations have been valid in any number of spatial dimensions, depending on the definitions of the vector fields and the
spatial derivative operators. In the following, however, we will restrict our
attention to two spatial dimensions.
In order to treat the interface sharply, with the method we will later call
the ghost-fluid method (GFM), we need to be able to compute the jumps
in various quantities. How we do that will depend on whether we include
the phase-transition model or not. In the following two subsections we
will first state the required jumps for two-phase immiscible flow and then
state those for two-phase flow with phase transition.

2.5.1 Two-Phase Immiscible Flow

The required jump conditions for treating the interface sharply in the case
of two-phase immiscible flow are
[~u] = 0,
[p] = 2 [] n
~u n
+ f~sfd ~n,



[~u] = [] (
n ~u n
) n
n
+ n
~u t n
t




n
~u t tn
+ t ~u t tt


t f~sfd tn
,
[p] = 0,
[T ] = 0,


[T ] = t[] T t .

12

(2.26)
(2.27)

(2.28)
(2.29)
(2.30)
(2.31)

2.5 The Jump Conditions

The jump conditions (2.26) to (2.29) were derived in Lerv
ag [2008]. The
jumps in temperature [T ] and heat flux [T ] at the interface are derived
in this section.
According to Landau and Lifshitz [1987, 50], the temperature is continuous at any interface or boundary. The temperature jump then quite
simply becomes
[T ] = 0.

(2.32)

From the temperature jump (2.32), it immediately follows that

[T ] t = 0.

(2.33)

If this were not the case, the temperature could vary independently along
each side of an interface and thus violate (2.32).
Also according to Landau and Lifshitz [1987, 50], the local conductive
heat flux out of one phase must be equal to the local conductive heat flux
into the other. This amounts to demanding that the component of the
heat flux normal to the interface is continuous at the interface,
[T ] n
= 0.

(2.34)

To get an expression for the jump in T at the interface, we decompose

it into normal and tangential components,
[T ] = n
[T ] n
+ t[T ] t.

(2.35)

Using (2.34) and (2.33), we get

[T ] = t[T ] t,


= t[] T t .

(2.36)
(2.37)

This means that the conductive heat flux tangential to an interface may
be discontinuous if the phases have different thermal conductivities.

2.5.2 Two-phase Flow with Phase Transition

When modeling phase transition, it is necessary to modify the jump conditions with respect to those stated in the previous section.

13

2 The Mathematical Model for the Two-Phase Flow Physics

Due to the mass flux across the interface, the velocity is no longer
continuous there and the jump condition (2.26) is no longer valid. Instead,
we replace it with an expression derived from mass conservation at the
interface (2.22).
Because of the heat absorbed or released by the phase transition, the
conductive heat flux is not continuous at the interface either, and we
replace the interface condition (2.31) with (2.24).
As a jump condition for pressure, we use (13) from Gibou et al. [2007].
In this equation, the jump contribution from the viscous stress has changed
with respect to (2.27) because we can no longer assume that ~u is continuous
at the interface and there is an additional term to account for the pressure
jump from the acceleration of the mass that crosses the interface. The
jump conditions on T and p are the same as in the previous section.
Stated together, the jump conditions with phase transition are
 
1
[~u] = m
n
,
(2.38)

 
2 1
~
,
(2.39)
[p] = [
n ~u n
] + fsfd ~n m

[p] = 0,
(2.40)
[T ] = 0,
[T n
] = mh.

(2.41)
(2.42)

Note that we have not stated a jump condition for ~u. This is because,
according to Gibou et al. [2007], it is not clear how to treat the jump in
viscous stress sharply when modeling phase transitions and we will use
the continuum surface force (CSF) method rather than the GFM when
computing the viscous stress in these cases.

2.6 Cartesian vs. Cylindrical Coordinates

On regular rectangular grids, we will use the standard Cartesian definitions
of the divergence, gradient and Laplacian operators in two-dimensions
[Kreyszig, 2006, Appendix 3.4],
~u =

14

u v
+
,
x y

(2.43)

2.7 Boundary Conditions

2
2
~u = x

+ 2
2
x
y
T
T
T = x

+ y ,
x
y
2T
2T

2 T =
+
.
x2
y 2
2

2
2
u + y
+ 2
2
x
y



v,

(2.44)
(2.45)
(2.46)

However, when we do three-dimensional simulations under the assumption of rotational symmetry around the y-axis, we use a different grid and
different definitions of the above operators. We now let x be the radial
coordinate and u be the radial component of the velocity and use the
cylindrical-coordinate operators and an axisymmetric grid. With the assumption of rotational symmetry around the y-axis, these operators are
[Kreyszig, 2006, Appendix 3.4]
1
v
(xu) +
,
x x
y
 2

 2


1
2

2
2
~u = x

+
u + y
+
v,
x2 x2 y 2
x2 y 2
T
T
T = x

+ y ,
x
y
2
1 T
2T
T
+
+
.
2 T =
x2
x x
y 2

~u =

(2.47)
(2.48)
(2.49)
(2.50)

Note that because we have assumed rotational symmetry, there is no

contribution from the derivatives with respect to the azimuthal angle in
(2.47) to (2.50) and our three-dimensional problem is reduced to a twodimensional one.

2.7 Boundary Conditions

In order to evolve the ~u, p and T fields in time, we need well-posed boundary conditions (BCs) for them. What these are depend, of course, on what
kind of physical of properties the boundaries have. In this section, we
present the different BCs used in this thesis.

15

2 The Mathematical Model for the Two-Phase Flow Physics

2.7.1 Velocity Boundary Conditions

No-Slip
When we demand that a fluid must be at rest at a boundary, we apply
the no-slip BC, and set
~u = 0
(2.51)
at that boundary.
Free-Slip
When a boundary is considered a solid surface that cannot supply shear
force, we impose that shear stress is zero at the boundary and that there
is no flow normal to it. This is expressed mathematically as
~u n
= 0,
n (~u t) = 0.

(2.52)
(2.53)

Outflow
At a boundary where the fluids can flow freely out of and into the domain,
we impose that the derivative in the normal direction of both velocity
components is zero,
n ~u = 0.
(2.54)

2.7.2 Pressure Boundary Conditions

Dirichlet
With Dirichlet BCs for the pressure, we set
p=0

(2.55)

along the boundary.

Neumann
We demand that the normal component of the pressure gradient is zero
at the boundary,
n p = 0.
(2.56)

16

2.8 Summary
This choice of BC is not physically founded, but merely an artifact from
the projection method. For further details, see Griebel et al. [1998, Section
3.2.3]. With this BC on all boundaries, p is only determined up to the
addition of a constant. Therefore we must also supply the value of p at
given point.

2.7.3 Temperature Boundary Conditions

Dirichlet
When a boundary can be considered an infinitely large reservoir that can
absorb or supply an infinite amount of heat at a constant temperature,
we set the temperature equal to a constant value,
T = Ts ,

(2.57)

where s {west, east, north, south}, at that boundary.

Neumann
If we wish to model a boundary with a constant heat flux density q, we
set
n T = q
(2.58)
at that boundary. With the convention that we take the derivative n
along the outward normal to the boundary, heat flux into the fluid domain
is associated with a positive q. A thermally insulating boundary is a
special case of the Neumann BC with q = 0.

2.8 Summary
In this chapter, the mathematical model for the two-phase flow physics was
presented. The most significant parts of the model were the incompressible
NavierStokes equations (2.2) and (2.10) for the velocity ~u and pressure p
fields (Section 2.1) and the advection-diffusion equation (2.11) for temperature T (Section 2.2). The body force term in the NavierStokes equations
was given a temperature dependence (2.15) according to Boussinesq approximation (Section 2.3) to enable simulations with temperature-driven
flows and a phase-transition model was presented (Section 2.4) to allow

17

2 The Mathematical Model for the Two-Phase Flow Physics

for vaporization and condensation mass transfer between phases. Jump
conditions that are required to treat the interfaces sharply (Section 2.5)
and the boundary conditions used in this work were given (Section 2.7).

18

3 Numerical Methods
In the previous chapter, the mathematical model for the two-phase flow
physics was presented. In this chapter, we present the methods used to
numerically solve it. The general scheme is similar to those used by e.g.
Kang et al. [2000] and Gibou et al. [2007] and can be summarized as
follows.
We will use explicit RungeKutta (RK) methods to integrate the equations of motion and thus approximate the velocity, pressure and temperature fields discrete at points in time. This can be done if we are able
to compute the time derivatives of the velocity and temperature fields at
every time step (and RK stage) and the pressure at the next. To this avail
we will make use of a variant of Chorins projection method. This method
yields explicit expressions for the required time derivatives and a Poisson
equation for the pressure. To compute these time derivatives and solve the
Poisson equation, we also need schemes to approximate spatial derivative
operators. We will use the weighted essentially non-oscillatory (WENO)
scheme for advective operators and central differencing (CD) schemes for
Laplacian, gradient and divergence operators. This again requires knowledge of the interface location at each time step and to keep track of that,
we employ the level-set method. Also, when computing approximations
to the spatial derivatives, we need to treat discontinuities at the interface.
For this we use the GFM, the CSF method or a combination of them,
depending on the case at hand.

3.1 The Projection Method

In order to approximate the time evolution of the velocity and pressure
fields with our explicit RK schemes, we need explicit expressions or solvable equations instead of NavierStokes equations (2.2) and (2.10).
In this thesis, we use a variant of the Chorins projection method [Griebel
et al., 1998, Section 3.2.1]. This method gives an explicit expression for

19

3 Numerical Methods
the time derivative of the velocity field t ~u(n) at time step n and a Poisson equation the pressure p(n+1) at the next time step (or time stage for
higher-order RK methods). It is thus explicit in all fields except the pressure where it is implicit. In this section, we present the projection method
by giving a brief derivation of it.
We start our derivation by defining an intermediate vector field ~a(n) .
This contains the entire right hand side of the momentum balance equation
(2.10) when solved for t ~u at time step n, except the contribution from
the gradient of the pressure,
t ~u(n) = ~a(n)

p(n+1)
.

(3.1)

Explicitly stated then, ~a(n) is

n
oT 
o n


(n)
(n)
(n)
(n)
(n)
+ ~u
~a = ~u ~u +
~u
(n)
+ f~b + f~s(n) .

(3.2)

This equation for ~a(n) can be directly evaluated, with proper discretization
of the spatial differential operators, since the right hand side contains
known fields only. However, in order to use (3.1) to get t ~u(n) , we must
find the as of yet unknown pressure p(n+1) . To do this, we discretize the
time derivative in (3.1) using the forward Euler method and get
~u(n+1) ~u(n)
p(n+1)
= ~a(n)
.
t

(3.3)

The t is to be understood as the time-step length from time step n to

n + 1. We do not yet know the new velocity field ~u(n+1) , but we can
demand that it should satisfy (2.2) and be divergence free and then let
p(n+1) be whichever pressure field that makes ~u(n+1) = 0. Applying
to both sides of (3.3), and demanding that ~u(n+1) = 0, we get
!
~u(n)
p(n+1)
= ~a(n) +
.
(3.4)

t
This is a Poisson equation for the pressure p(n+1) , with a known right-hand
side. According to Hansen [2005], it can be solved with the appropriate
boundary conditions.

20

3.1 The Projection Method

In this derivation of Chorins projection method we have used the pressure to enforce the divergence-freeness of the velocity field. We must
therefore expect that the pressure field will adapt such as to make (3.1)
give any t ~u(n) that will make the velocity field divergence free in the
next time step. This behavior of the pressure can be seen, from a physical point of view, as an infinite speed of sound. Information about local
changes in the velocity field must be instantaneously propagated to the
entire domain by the pressure to give a divergence free velocity field. As a
consequence, the pressure we calculate in this incompressible flow model
will not necessarily correspond to a thermodynamic state variable.

3.1.1 Solving the Poisson Equation

The Poisson equation (3.4) is discretized as described by Kang et al. [2000].
Solving the discretized equation for p(n+1) can be done by solving a linear
system of equations corresponding to a symmetric matrix. In this thesis,
the implementations of the conjugate gradient method with incomplete
Cholesky preconditioning and the biconjugate gradient stabilized method
with incomplete LU preconditioning from the PETSc package [Balay et al.,
2013] were used. The tolerance of the solvers was set to 109 for all cases.
The Compatibility Condition
If we integrate both sides of the Poisson equation (3.4) over the entire
fluid domain ,
!
!
Z
Z
p(n+1)
~u(n)
(n)

d~x =
~a +
d~x,
(3.5)

and apply the divergence theorem of Gauss [Kreyszig, 2006, Section 10.7],
we get
!
Z
Z
~u(n)
p(n+1)
(n)
~a +
n
d~s =
n
d~s.
(3.6)

When imposing Neumann BCs on the pressure, we set p(n+1) n

= 0
(n)
along the boundary. Remembering that we also always set ~a n
= 0
along the boundary, we see that we must have
Z
~u(n) n
d~s = 0.
(3.7)

21

3 Numerical Methods
to satisfy (3.5) when we impose the Neumann BCs. Therefore, (3.7) is
our condition for compatibility between the right hand side of the Poisson
equation (3.4) and the Neumann BCs on p. According to Griebel et al.
[1998, Section 3.2.1], compatibility is necessary for the Poisson equation
(3.4) to have a solution (determined up to the addition of a constant).
Looking more closely at (3.7), we see that we have compatibility whenever the net flow of fluid across the boundaries and into the domain is zero.
Therefore, as long as there is no phase transition, we can follow Griebel
et al. [1998] and Kang et al. [2000] and say that (3.7) is satisfied because
of the BCs we impose on ~u.
When we do have a phase transition, however, these arguments are no
longer valid. The easiest way to ensure that the Poisson equation (3.4) still
has a solution is to resort to Dirichlet BCs for the pressure and therefore
this is what we will do in this thesis.

3.2 The Level-Set Method

When employing finite-difference approximations to the spatial derivative
operators in (3.1), (3.2) and (3.4), it is necessary to know the location of
the interface between the two phases under consideration. To this avail we
need a method to keep track of the interface location at each time step.
Several approaches, such as the volume of fluid method [Scardovelli and
Zaleski, 1999], the level-set method [Osher and Fedkiw, 2003] and front
tracking methods [Tryggvason et al., 2001], can be used. In this work, we
use the level-set method and in this section it is briefly presented.
With the level-set method, the interface is implicitly defined as the
zero isocontour of the scalar level-set function ,
(t) = {~x | (~x, t) = 0} .

(3.8)

The level-set function thus divides the fluid domain in two parts, +
and , where
+ (t) = {~x | (~x, t) > 0} ,

(t) = {~x | (~x, t) < 0} .

(3.9)
(3.10)

We refer to the fluid occupying the region as Phase 1 and the fluid
occupying + as Phase 2.

22

3.2 The Level-Set Method

The temporal evolution of the level-set function is given by taking the
total time derivative on both sides of (~x, t) = 0,

+w
~ = 0,
t

(3.11)

where w
~ is the velocity of the interface. In cases without phase transition,
w
~ is the fluid velocity at the interface. Then (3.11) can be viewed as an
advection equation for and the temporal evolution of the interface is
advection with the velocity field ~u.
There are many scalar functions that could serve as suitable level-set
functions. We will, however, stick to a particularly useful and widely used
choice, the signed distance-function
(
d0 (~x) if ~x
,
(3.12)
d (~x) =
d0 (~x)
if ~x +
where d0 (~x) is the Euclidean distance from the point ~x to its closest point
on the interface,



~x ~x0 .
(3.13)
d0 (~x) = min
2
0
~
x

The advantages of the level-set method are many. Two of them are the
relative ease with which it can be implemented and its ability to handle
topological changes such as merging and pinching of drops without any
special treatment. Another advantage is the straightforward definition of
the normal vector n
to the interface, pointing from into + [Lerv
ag,
2008, Gibou et al., 2007],

n
=
,
(3.14)
kk2
and the interface curvature k,

k = n
.

(3.15)

One of the most significant drawbacks of the level-set method is that it

does not conserve mass [Tanguy et al., 2007, Gibou et al., 2007]. That is,
the global volume fractions of phases 1 and 2 may change during the time
evolution, without this being a result of physical mass transport between
the phases. This problem is especially pronounced when the grid is coarse
and/or the interface exhibits high curvature.

23

3 Numerical Methods

3.2.1 Reinitialization
In the discretization of some of the spatial derivative operators, we will
rely on being a signed distance-function. However, accumulating numerical errors in the time integration of (3.11) and a non-uniform velocity
field w
~ will cause to gradually lose this property during the course of
a simulation. To mend this, we employ the reinitialization procedure by
Sussmann et al. [1994] at regular time-step intervals. With this procedure,
we take our current level-set function 0 and iterate the reinitialization
equation,

+ S (0 ) (kk2 1) = 0,
(3.16)

towards steady-state in pseudo time . This eventually produces the

steady state solution . In (3.16) S is a smeared sign function,
S () = q

2 + (2x)2

(3.17)

Since is the steady-state solution to (3.16), it must satisfy k k2 = 1.

According to Osher and Fedkiw [2003, Section 2.4] this is a property only
signed distance-functions possess. Thus, by replacing 0 with , the
level-set function has regained its signed distance-function property.

3.2.2 Velocity Extrapolation

As showed by Zhao et al. [1996], another measure we can take to make
keep its signed distance-function property upon time integration of (3.11)
is to define w,
~ in points not on the interface, as being equal to the interface
velocity at the closest interface point. In practice this is achieved by
extrapolating w
~ such that it is constant in directions normally away from
the interface before computing t in (3.11). The extrapolation is done
by solving the PDE
w
~
+ S () n
w
~ =0
(3.18)

to steady-state in pseudo time . Again, S () is the smeared sign function

defined by (3.17). Since the steady-state solution w
~ must satisfy n
w
~ =

0, we see that w
~ is indeed constant in direction normal to the interface.

24

3.3 Time Integration with Strong-Stability-Preserving RungeKutta

Methods

3.3 Time Integration with

Strong-Stability-Preserving RungeKutta
Methods
The equations of motion considered in this thesis, the extrapolation PDEs
and the level-set reinitialization equation are spatially discretized such
that the time evolution of their solutions can be approximated by solving
initial value problems in the form

= f (, t) ; t 0,
t

(0) = 0 .

(3.19)

Here, (t) is a placeholder for an arbitrary smooth function of time and

0 is its known initial value.
The solution to (3.19) can be approximated at discrete points in time t(n)
with e.g. RK methods. We let (n) approximate at time t(n) . According
to Ketcheson and Robinson [2005], any irreducible s-stage RK method can
be uniquely specified by its Butcher table,
c1 a11 . . . a1s
..
.
..
..
. .. ,
.
.
cs as1 . . . ass
b1 . . . bs
where
ci =

s
X
j=1

aij , i {1, ..., s} .

(3.20)

The coefficients in this table then specify the RK method as

{i}

(n)

+ t

(n+1) = (n) + t

s
X
j=1
s
X



aij f {j} , t(n) + cj t ,


bj f {j} , t(n) + cj t .

i {1, .., s} , (3.21)

(3.22)

j=1

In these equations, {i} denotes the value of at the ith RK stage.

25

3 Numerical Methods
If aij = 0 for j i, the RK method is explicit. In this work, only
explicit RK methods with the strong-stability-preserving (SSP) property
are used. Strong-stability-preserving methods are higher-order RK methods that preserve the stability of the forward Euler method. This property
is obtained by ensuring that the RK method can be written as a convex
combination for Forward Euler steps. For a more complete discussion
of RK methods and the SSP property and its practical importance, the
reader is referred to S
uli and Mayers [2006, Section 12.5] and Ketcheson
and Robinson [2005] respectively.
In this thesis, the forward Euler method with Butcher table
0
1

the two-stage second-order method (SSP-RK 2()2),

0
1

1
,
1/2 1/2

and the four-stage third-order method (SSP-RK 3()4),

0
1/2 1/2
1 1/2 1/2
,
1/2 1/6 1/6 1/6
1/6 1/6 1/6 1/2
were used. All equations of motion (2.11), (3.1) and the level-set equation
(3.11), were integrated with the SSP-RK 2()2 method and all extrapolation equations and the level-set reinitialization equation (3.16) were
integrated with SSP-RK 3()4 where otherwise is not explicitly stated.

3.4 Spatial Discretization

We use equations (2.11), (3.1), (3.2), (3.4) and (3.11) to approximate
the pressure and the time derivatives of the level-set function and of the

Strong-stability-preserving methods are also known as total variation diminishing

(TVD) methods in the literature [Ketcheson and Robinson, 2005].

26

3.4 Spatial Discretization

velocity and temperature fields. Also, we have PDEs that we solve to
steady-state to do extrapolation and reinitialization. However, in order
to do so, we need schemes to estimate the spatial derivative operators
in these equations. Most of these schemes can be found in Kang et al.
[2000], Gibou et al. [2007] and the references therein. In this section we
discuss some of the most important and non-standard ones in detail. For
simplicity, we restrict our discussion to rectangular grids.
Throughout this thesis, we work on a staggered marker-and-cell (MAC)
grid, where scalar fields such as the pressure p and the temperature T
are defined at the cell centers and vector fields such as the velocity ~u are
defined at the cell edges. This is illustrated in Figure 3.1.

3.4.1 Discretization of the Advective Operator

When evaluating (3.2) and the right-hand sides of the level-set advection equation (3.11) and the advection-diffusion equation for temperature
(2.11), it is necessary to discretize the advective operator (~u ). In this
work, the weighted essentially non-oscillatory (WENO) scheme is used
and in this section we present it and motivate its use in disfavor of the
simpler first-order central differences (CD), first-order upwind (FUP) and
third-order essentially non-oscillatory (ENO) schemes.
For simplicity and concreteness, we will consider the convective operator
in one dimension and use temperature as the example field. That is, we
will consider the discretization of
T
u
,
(3.23)
x
since the extension to multiple dimensions can be done in a dimension
by dimension fashion. We will also assume uniformly spaced grid nodes.
This section follows roughly the arguments from Osher and Fedkiw [2003,
Chapter 3].
The Central Differences Scheme
The CD scheme is a natural first naive choice of discretizing the spatial
derivative in (3.23),


T(i+1) T(i1)
T

.
(3.24)
x (i)
2x

27

3 Numerical Methods

(i, j + 1/2)
Grid cell

(i + 1/2, j)

(i, j)

Boundary

Figure 3.1: Schematic illustration of the MAC grid. The solid lines mark the cell edges.
Scalar values are stored at the cell centers (i, j), marked by circles. Vector quantities
have their x-components stored at the vertical cell edges (i + 1/2, j), marked by squares,
and their y-components at the horizontal cell edges (i, j + 1/2), marked with diamonds.
Boundaries lie along cell edges and boundary conditions are applied by setting the field
values in the gray ghost cells.

28

3.4 Spatial Discretization

If the velocity field is interpolated to the cell centers by simple averaging
and multiplied with (3.24), we obtain a discretization of (3.23). Unfortunately, however, this scheme will become unstable fail or at best produce
spurious oscillations in T , as described by Osher and Fedkiw [2003, Section
3.2] and Lerv
ag [2008].
The First-Order Upwind Scheme
There are ways to construct better behaving schemes than CD, e.g. following the principle of upwind differencing. If u(i) > 0, then values of T
move from left to right around (i), advected with the fluid flow. Since information about the temperature flows from left to right, it is then natural
to use values located predominantly to the left of (i) when determining
what value of T that will be advected to (i) in the next time step, that
is, when approximating x T at (i). We name this approximation to x T ,
using values of T predominantly to the left, Tx . Analogously, if u(i) 0
we use values that lie predominantly to the right of (i) to approximate
x T at (i) and name the approximation Tx+ .
One possible choice for the approximations Tx and Tx+ are the backward

+
D(i) T and forward D(i)
T differences respectively. This results in the FUP
scheme,
(


T
T(i)
+
if u(i) 0
(Tx+ )(i) = D(i)
T = (i+1)
T
x
.
(3.25)

T
T
x (i)
(Tx )(i) = D T = (i) (i1) if u(i) > 0
(i)

This scheme is first-order accurate and produces much better results than
CD, according to Osher and Fedkiw [2003, Section 3.2].
The Essentially Non-Oscillatory Scheme
A scheme of higher than first-order accuracy would be an obvious improvement over the FUP scheme. The ENO scheme accomplishes increased accuracy by calculating the approximations Tx and Tx+ in a more elaborate
fashion compared to FUP. The general idea is to use divided differences
to construct a Newton interpolation polynomial Q to T and then compute the derivative of this polynomial. See e.g. Kreyszig [2006, Section
19.3] for an introduction to Newton interpolation polynomials and divided
differences.

29

3 Numerical Methods
The first four divided differences, here indexed 0 to 3, can be expressed
as
0
T
D(i)

= T(i) ,

1
T
D(i+1/2)

2
T
D(i)

3
D(i+1/2)
T

(3.26)

0
0 T
D(i+1)
T D(i)

,
x
1
1
D(i+1/2)
T D(i1/2)
T
2x
2
2 T
D(i+1)
T D(i)
3x

(3.27)
,

(3.28)
(3.29)

Note that the zeroth and second divided differences are defined at the cell
centers and that the first and third divided differences are defined at the

1
cell edges. We notice also that D(i1/2)
T = D(i)
T , that is, the first divided
difference at (i 1/2), is the backward difference approximation to x T
+
1
at (i) and that D(i+1/2)
T = D(i)
T is the forward difference approximation
to x T at (i).
From the divided differences, we can construct a Newton interpolation
polynomial Q approximating T ,
Q (x) = Q0 (x) + Q1 (x) + Q2 (x) + Q3 (x) ,

(3.30)

where Qi is the ith-order contribution. The derivative of Q is

Q0 (x) = Q01 (x) + Q02 (x) + Q03 (x) ,

(3.31)

where we note that the constant zeroth-order contribution has vanished.

The first-order contribution Q1 is




1
Q1 (x) = D(k+1/2)
T x x(i) ,
(3.32)
and the derivative is
1
Q01 (x) = D(k+1/2)
T.

(3.33)

We set k = i 1 if u(i) > 0 and k = i if u(i) 0. Observe now that if

we ignore the higher order contributions to Q0 , we recover the first-order
FUP scheme. To get the desired third-order accuracy, however, we must
include the contributions Q02 and Q03 .

30

3.4 Spatial Discretization

2 T and
When calculating Q2 we can choose to make use of either of D(k)
2
D(k+1)
T . Wanting to construct the smoothest interpolation polynomial,
we opt for the one with the smallest absolute value and define

C2 =
` =

(
2 T
D(k)

2 T | < |D 2
if |D(k)
(k+1) T |

2
D(k+1)
T otherwise
(
2 T | < |D 2
k 1 if |D(k)
(k+1) T |

otherwise

(3.34)
(3.35)

Then
Q2 (x) = C2 x x(k)

x x(k+1) ,

(3.36)

and differentiating at x(i) we get

Q02 x(i) = C2 (2 {i k} 1) x.

(3.37)

The third order contribution Q3 is found from a similar argument. We

let
(
3
3
3
D(`+3/2)
T if |D(`+3/2)
T | < |D(`+1/2)
T|
.
(3.38)
C3 =
3
D(`+1/2) T otherwise
Then
Q3 (x) = C3 x x(`)

x x(`+1)

x x(`+2) ,

(3.39)

and differentiating at x(i) we get





Q03 x(i) = C3 3 {i `}2 6 {i `} + 2 (x)2 .

(3.40)

Using this procedure to find Q0 and evaluating it at x(i) to approximate

x T at (i) is the ENO method.
The Weighted Essentially Non-Oscillatory Scheme
It turns out that the ENO scheme produces one of three possible approximations Tx or one of another three approximations Tx+ , depending on the

31

3 Numerical Methods
relative magnitudes of the divided differences and the sign of u(i) . Defin
ing j = D(i3+j)
T, j {1, 2, 3, 4, 5} the three possible approximations

Tx are
1 72 113

+
,
3
6
6
2 53 4
+ ,
= +
6
6
3
3 54 5
=
+
.
3
6
6

Tx1 =

(3.41)

Tx2

(3.42)

Tx3

(3.43)

If we instead define j = D(i+3j)

T, j {1, 2, 3, 4, 5}, (3.41) to (3.43)
give the three possible approximations Tx+ .
While the ENO method chooses one of the three possible approximations, the WENO method uses a convex linear combination of them and
thus achieves fifth-order accuracy in P
smooth regions. Introducing the
3
weights 1 , 2 , 3 [0, 1] such that
j=1 j = 1, the WENO scheme

choice of Tx is
Tx = 1 Tx1 + 2 Tx2 + 3 Tx3 .
(3.44)

In the rest of this section we will consider this choice of Tx , noting that the
following formulas apply also for the choice of Tx+ using the appropriate
definition of j .
In Jiang and Shu [1996], it was shown that the WENO scheme produces
fifth-order accurate results in smooth regions if the weights are chosen as
1 = 0.1, 2 = 0.6, and 3 = 0.3. Unfortunately, this choice of weights
can be very inaccurate in regions where T is not smooth. The key to
mending this is to choose the weights so that they are close to the optimal
weights when T is smooth around (i) and Tx1 , Tx2 and Tx3 all correspond to
smooth interpolation polynomials, and, when T is not smooth around (i),
to give less weight to those of Tx1 , Tx2 and Tx3 that correspond to the least
smooth interpolation polynomials. This results in third order-accuracy in
non-smooth regions.
The non-smoothness measures associated with each of the approximations Tx1 , Tx2 and Tx3 are
Z1 =

32

13
1
(1 22 + 3 )2 + (1 42 + 33 )2 ,
12
4

(3.45)

3.4 Spatial Discretization

Z2 =
Z3 =

13
(2 23 + 4 )2 +
12
13
(3 24 + 5 )2 +
12

1
(2 4 )2 ,
4
1
(33 44 + 5 )2 .
4

(3.46)
(3.47)

Next, we define weights that are small when the corresponding measure
of non-smoothness is large,
1 =
2 =
3 =

0.1
,
(Z1 + )2
0.6
,
(Z2 + )2
0.3
,
(Z3 + )2

(3.48)
(3.49)
(3.50)

where  = 106 to make sure that we never attempt division by zero.

These weights are normalized and chosen as the weights in (3.44),
1 =
2 =
3 =

1
,
1 + 2 + 3
2
,
1 + 2 + 3
3
.
1 + 2 + 3

(3.51)
(3.52)
(3.53)

The WENO scheme is thus preferable to the ENO scheme and, by extension, to the other schemes considered in this section because of its
fifth-order accuracy in smooth regions. It also has another advantage over
the ENO scheme in that a smaller computational effort is needed for each
approximation to x T . This was demonstrated e.g. by Lerv
ag [2008] in
his Table 3.1.

3.4.2 Interface-Capturing Methods

In the spatial discretizations of the Laplacian operator (viscous term) in
(3.2), the Laplacian (diffusive term) in (2.11), the Laplacian in (3.4) and
the gradient operator in the projection equation (3.1) for two-phase simulations, we need a way to handle interfacial discontinuities in the fields
we apply these operators to and/or their spatial derivatives.

33

3 Numerical Methods
In this section, we present two strategies for handling the interfacial
discontinuities in the diffusive term in the advection-diffusion equation
(2.11). These are the continuum surface force (CSF) method and the
ghost-fluid method (GFM). The GFM scheme is discussed twice, once
for two-phase immiscible flow and once for two-phase flow with phase
transition.
For a presentation of these methods in the context of the viscous term
in (3.2), the Poisson equation (3.4) and the projection equation (3.1), the
interested reader is referred to Kang et al. [2000] and Osher and Fedkiw
[2003].
The Continuum Surface Force Method
The CSF method handles discontinuities by smearing out the discontinuous properties around the interface in a controlled way.
Consider the discretization of the term


1
T

.
(3.54)
cp x
x
The discretization of the y-term will be completely analogous to this. We
discretize it at the grid point (i, j) using a standard CD scheme


1
cp x


(i,j)

2 (i,j)

T(i+1,j) T(i,j)
x(i)

T(i,j) T(i1,j)
x(i1)

x(i) + x(i1)

, (3.55)

where x(i) is the distance between grid points (i, j) and (i + 1, j) and
is smeared out in the vicinity of the interface. The smearing of is given
by
() ,
() = + (+ )H

(3.56)

() refers to a smeared Heaviside function, which guarantees that

where H
the thermal diffusivity is smooth,

0
  if < 

() = 1 + + 1 sin
if   .
(3.57)
H
2
2
2


1
if  <
() is illustrated in Figure 3.2.
The smoothness of H

34

3.4 Spatial Discretization

1.5

H()

1.0

0.5

0.0

0.5
2.0

1.5

1.0

0.5

0.0
/

0.5

1.0

1.5

2.0

Figure 3.2: A plot of the smeared Heaviside function (3.57).

The Ghost-Fluid Method for Immiscible Two-Phase Flow

It would be advantageous to have a way of sharply discretizing terms like
(3.54) without numerical smearing of . For this purpose, we will use
the GFM. Again, we will consider only the discretization of the x-term,
as discretization of the y-term will be completely analogous to this. We
discretize the derivative in (3.54) at the grid point (i, j) follows,



(i,j)

(x T )(i+1/2,j) (x T )(i1/2,j)
x(i) + x(i1)

(3.58)

Now, if (i1,j) , (i,j) and (i+1,j) are either all positive or all negative,
the discretization of the numerator above is straightforward. If all are
positive, then


T(i+1,j) T(i,j)
T

+
,
(3.59)
x (i+1/2,j)
x(i)


T(i,j) T(i1,j)
T
+
.
(3.60)

x (i1/2,j)
x(i1)

35

3 Numerical Methods

x(i1)

(1 )x(i1)

(i 1, j)

(i, j)
x(i1)

Figure 3.3: An illustration of the sub-cell resolution of the interface location. The
dashed line shows the interface position.

For the case where (i1,j) , (i,j) and (i+1,j) are all negative, we replace
+ with in the above equations,


T(i+1,j) T(i,j)
T

,
(3.61)
x (i+1/2,j)
x(i)


T(i,j) T(i1,j)
T

.
(3.62)

x (i1/2,j)
x(i1)
If (i1,j) and/or (i+1,j) have a different sign from (i,j) , we must take
the presence of an interface into account.
Consider first the case where there is an interface located between the
grid points (i1, j) and (i, j), that is when (i1,j) and (i,j) have opposite
signs. Assume also that (i1,j) 0 and (i,j) > 0. Relying on the
distance-function property of the level-set function, we will assume that
the interface is located at

where is given by

xi = x(i1) + x(i1) ,

(3.63)



(i1,j) 
 
.
= 
(i1,j)  + (i,j) 

(3.64)

In this way, we resolve the position of the interface on a scale smaller than
the cell sizes, see Figure 3.3.

36

3.4 Spatial Discretization

To estimate the temperature at the interface Ti , we discretize the jump
condition (2.31) at the interface as follows and call it Ji .
Ji +

T(i,j) Ti
Ti T(i1,j)

.
(1 ) x(i1)
x(i1)

(3.65)

The jump condition itself is interpolated to the interface as

Ji (1 ) J(i1,j) + J(i,j) .

(3.66)

Equation (3.65) can be solved for the interface temperature,

Ti

+ T(i,j) + T(i1,j) (1 ) Ji (1 ) x(i1)

,
+ + (1 )

(3.67)

and this expression for Ti can be used to discretize x T at (i 1/2, j),



T(i,j) Ti
T
+

,
(3.68)
x (i1/2,j)
(1 )x(i1)
0

T(i,j) T(i1,j)
Ji
+ 0 .
x(i1)

(3.69)

In these equations we have defined the effective thermal conductivity,

0 =

+
.
+ + (1 )

(3.70)

By a similar argument one finds, for the case where (i1,j) > 0 and
(i,j) 0,


T(i,j) Ti
T

,
x (i1/2,j)
(1 )x(i1)
00

T(i,j) T(i1,j)
Ji
+ 00 + ,
x(i1)

(3.71)
(3.72)

where the effective thermal conductivity is

00 =

+
.
+ + (1 )

(3.73)

37

3 Numerical Methods
Now we can insert the appropriate one of (3.69) and (3.72) for the discretization of x T at (i 1/2, j) to evaluate (3.58) in the case where
there is an interface between (i 1, j), and (i, j).
Next, consider the situations where there is an interface located between
(i, j) and (i + 1, j). Then we can interpolate the jump condition to the
interface as
Ji (1 )J(i+1,j) + J(i,j) ,
(3.74)
where is now given by


(i+1,j) 
.
 
= 
(i+1,j)  + (i,j) 
For the case where (i,j) > 0 and (i+1,j) 0, we use


Ti T(i,j)
T

+
,
x (i+1/2,j)
(1 )x(i)
0

T(i+1,j) T(i,j)
Ji
+ 0 ,
x(i)

(3.75)

(3.76)
(3.77)

in the evaluation of (3.58). And for the case where (i,j) 0 and (i+1,j) >
0, we use


Ti T(i,j)
T

,
(3.78)
x (i+1/2,j)
(1 )x(i)
00

T(i+1,j) T(i,j)
Ji
+ 00 + ,
x(i)

(3.79)

in (3.58).
The Ghost-Fluid Method for Two-Phase Flow with Phase Transition
When simulating two-phase flow with phase transition, we replace the interface condition (2.31) on the temperatures with the condition (2.23);
that the interface temperature should be equal to the saturation temperature Tsat . This leads to a slightly different GFM treatment of the Laplacian
operator in (2.11), which we will proceed to describe in this section. This
new scheme is similar to that described by Gibou et al. [2002] for treating
Stefan problems.

38

3.4 Spatial Discretization

We choose the discretization of the Laplacian operator like in the immiscible two-phase flow case, according to (3.58). Again, if (i1,j) , (i,j)
and (i+1,j) are all positive, we plug (3.59) and (3.60) into (3.58) and,
as before, if (i1,j) , (i,j) and (i+1,j) are all negative we use (3.61) and
(3.62).
However, when (i1,j) and (i,j) do not have the same sign, we need to
approximate x T at (i 1/2, j) in a way that enforces the condition that
the interface temperature is equal to Tsat . First, define the cell division
parameter ,


(i,j) 
,
 
= 
(3.80)
(i1,j)  + (i,j) 
and consider the case when (i,j) > 0. Then, the straightforward way to
approximate x T at (i 1/2, j) is


T(i,j) Tsat
T

+
.
(3.81)
x (i1/2,j)
x(i1)
According to Gibou et al. [2002], this scheme becomes unstable when is
small, because we end up subtracting two numbers of similar magnitude
and dividing the result by a small number. Therefore, we use it only when
> c , for some specified constant c h0, 1i. To get good approximations
to x T , we should choose c reasonably small.
If instead c and (i,j) and (i+1,j) have the same sign, we choose


T(i+1,j) Tsat
T
+

,
(3.82)
x (i1/2,j)
x(i1) + x(i)
and thus avoid the problem that occurs when is small. From this equation it is evident that the cost of choosing a finite c to keep this scheme
stable is a smearing out of the approximated x T at the interface.
When c and (i,j) and (i+1,j) do not have the same sign, we give
up and set


T
0.
(3.83)

x (i1/2,j)
According to Gibou et al. [2002], this is acceptable since we now have two
interfaces close together, both have temperatures Tsat and the temperature
derivative must therefore be small.

39

3 Numerical Methods
Similarly, we need to account for the interface condition also when (i,j)
and (i+1,j) have different signs. Assume for now that (i,j) > 0 and define


(i,j) 
 
.
= 
(3.84)
(i,j)  + (i+1,j) 
When > c , we approximate


Tsat T(i,j)
T

,
+
x (i+1/2,j)
x(i)

(3.85)

and when c and (i1,j) and (i,j) have the same sign, we use


Tsat T(i1,j)
T

+
.
(3.86)
x (i+1/2,j)
x(i) + x(i1,j)
Again, when c and (i1,j) and (i,j) have different signs we give up
and set


T

0.
(3.87)
x (i+1/2,j)
In the case where (i,j) 0, we use the same procedure, but replace +
with in the equations above.

3.5 Numerical Strategy for Handling Phase

Transition
The phase-transition model was presented in Section 2.4. In this section,
we give our strategy for handling it numerically. This strategy is similar
to that by Gibou et al. [2007] and can be roughly divided into three parts.
First, we have to calculate the mass flux density m from the temperature
field T . Second, we need to enforce the interface condition (2.38) on the
velocities and third, we must calculate the interface velocity in a way that
is consistent with mass conservation.
Note that in this work, we do not use the Boussinesq coupling when
simulating phase-transition cases. In these cases, the temperature may
drive the flow by initiating mass transfer between the phases, but not
through a temperature-dependent buoyancy force.

40

3.5 Numerical Strategy for Handling Phase Transition

3.5.1 Computing the Mass Flux Density from the Temperature

Field
Our first challenge when incorporating phase transitions in our model is
to accurately and robustly approximate n
T in cell centers on both sides
of the interface so that we can use (2.42) to get the mass flux density m.
We will use the following scheme to obtain approximations [Tx , Ty ] to
T at every cell center and then project them onto the interface normal
vector n
and extrapolate the projected values across the interface in a
constant fashion to compute m. This is similar to the scheme presented
by Gibou et al. [2002] for Stefan problems.
If |(i1,j) | < |(i+1,j) |, the point (i 1, j) lies closer to the interface
than (i + 1, j). Then, if (i1,j) and (i,j) are either both positive or both
negative, we discretize according to


T(i,j) T(i1,j)
T
Tx =
.
(3.88)
x (i,j)
x(i1)
On the other hand, if (i1,j) and (i,j) have different signs, there is an
interface between (i 1, j) and (i, j) and we need to use that the interface
temperature is equal to the saturation temperature (2.23), when approximating x T . Let


(i,j) 
.
 
(3.89)
= 
(i1,j)  + (i,j) 
Then, if > c for some specified constant c h0, 1i, we use the approximation


T(i,j) Tsat
T
Tx =
.
(3.90)
x (i,j)
x(i1)
By the same arguments as in Section 3.4.2, this scheme becomes unstable
when is small. This is precisely why we use it only when > c . If
instead c and (i,j) and (i+1,j) have the same sign, we choose


T(i+1,j) Tsat
T
Tx =
.
(3.91)
x (i,j)
x(i1) + x(i)
In the case where c and (i,j) has a different sign than (i+1,j) , we
set


T
Tx = 0.
(3.92)
x (i,j)

41

3 Numerical Methods
If instead |(i1,j) | |(i+1,j) |, then the point (i + 1, j) lies closer to the
interface than (i 1, j) and we follow an analogous procedure to approximate x T at (i, j). If (i+1,j) and (i,j) are both positive or both negative,
we discretize according to


T(i+1,j) T(i,j)
T
Tx =
.
(3.93)
x (i,j)
x(i)
On the other hand, if (i,j) and (i+1,j) have different signs, there is an
interface between (i, j) and (i + 1, j) and we need to enforce the interface
condition (2.23). Let


(i,j) 
 
.
(3.94)
= 
(i,j)  + (i+1,j) 
Then, if > c for the specified constant c h0, 1i, we use the approximation


Tsat T(i,j)
T
Tx =
.
(3.95)
x (i,j)
x(i)
If instead c and (i,j) and (i1,j) have the same sign, we choose


T
x


(i,j)

Tx =

Tsat T(i1,j)
.
x(i) + x(i1)

(3.96)

In the case where c and (i,j) has a different sign than (i1,j) , we
give up and set


T
Tx = 0.
(3.97)
x (i,j)
The scheme for approximating y T is completely analogous to that presented for x T and will not be discussed in further detail.
Assume now that we have our approximations Tx and Ty in a band
around the interface. Then we compute approximations Tn to n
T by
taking
Tn = n
[Tx , Ty ].
(3.98)
To compute the jump (2.42) we need to extrapolate Tn from Phase 1
into Phase 2 and vice versa to have liquid values Tn and gas values Tn+

42

3.5 Numerical Strategy for Handling Phase Transition

at each grid point. We perform the extrapolation by employing the PDE
approach presented by Aslam [2003] and solve the equations
Tn+

Tn

= {1 H ()} n
Tn+ ,
= H () n
Tn ,

(3.99)
(3.100)

to steady-state in pseudo time . In these equations, H is the Heaviside

function.
Having extrapolated Tn+ into and Tn into + , we can reliably calculate m according to the following approximation to (2.42),
m

(+ Tn+ Tn )
.
h

(3.101)

3.5.2 Enforcing the Interface Conditions on the Velocity

In order to enforce the jump in fluid velocity at the interface (2.38), we
define the liquid and gaseous velocities ~u and ~u+ point by point in terms
of the physical fluid velocities ~u in their respective phases and the ghost
values ~u
u+
g and ~
g,
~u+
~u

(
~u
=
~u+
g
(

~ug
=
~u

if > 0
,
if 0

(3.102)

if > 0
.
if 0

(3.103)

The values of the scalar level-set function is obtained at the velocity points
by linear interpolation. The ghost values are defined as
 
1
+
~u
=
~
u

m
n
,
(3.104)
g

 
1
+

~ug = ~u + m
n
.
(3.105)

To maintain the discontinuity in ~u as the interface moves across velocity

points, we evolve ~u+ and ~u separately in time. Specifically, we use (~u+ )(n)

43

3 Numerical Methods
and (~u )(n) to calculate (~a+ )(n) and (~a )(n) , respectively, with (3.2). Then
we approximate the corresponding time derivatives of ~u+ and ~u with
(3.1). The pressure is obtained by solving the Poisson equation (3.4),
where the right hand side is discretized using values from (~u )(n) and
(~a )(n) for Phase 1 points and (~u+ )(n) and (~a+ )(n) for points in Phase 2.
Then, after each time step (or stage), we update the physical velocities
~u(n+1) point by point according to
(
(~u+ )(n+1) if (n+1) > 0
~u(n+1) =
.
(3.106)
(~u )(n+1) if (n+1) 0

3.5.3 Computing the Interface Velocity

When there was no mass flux across the interface, the interface velocity
w
~ was equal to the fluid velocity at the interface. When there is mass
flux across the interface, however, we must account for this and use (2.20)
when computing w.
~ We arrived at this expression in Section 2.4 by imposing mass conservation at the interface. This interface velocity w
~ is used
in the level-set advection equation (3.11) in cases with phase transition.
Note that although we in (2.20) use the liquid (Phase 1) velocities ~u to
calculate w,
~ we could, in principle, use either of ~u and ~u+ .

3.6 Boundary Conditions

This section presents the implementation of the boundary conditions (BCs)
described in Section 2.7. The stated formulas are based on those given by
Griebel et al. [1998] and are applied in the beginning of each time step (or
stage). The fields are defined on an imax jmax rectangular MAC grid and
there is a band of ghost cells extending three cells into the far side of each
of the four boundaries; west, east, south and north. It is the field values
in these ghost cells and on their edges we set when we apply the BCs.

3.6.1 Velocity Boundary Conditions

No-Slip
With the no-slip BC, both the normal and the tangential component of
the velocity are to be zero. At a specific boundary, we set the normal

44

3.6 Boundary Conditions

component with the appropriate one of
u(0,j) = 0,
u(imax ,j) = 0,
v(i,0) = 0,
v(i,jmax ) = 0,

j {1, . . . , jmax } (West),

(3.107)

i {1, . . . , imax } (South),

(3.109)

j {1, . . . , jmax } (East),

(3.108)

i {1, . . . , imax } (North).

(3.110)

The tangential velocity components do not lie directly on the boundary,

so we enforce the no-slip condition by setting the ghost values so that
the tangential velocities become zero when linearly interpolated to the
boundary. That is, we set
v(0,j) = v(1,j) ,

j {1, . . . , jmax }

(West),

(3.111)

i {1, . . . , imax }

v(imax +1,j) = v(imax ,j) ,

j {1, . . . , jmax }

(East),

(3.112)

(South),

(3.113)

u(i,jmax +1) = u(i,jmax ) ,

i {1, . . . , imax }

(North).

(3.114)

u(i,0) = u(i,1) ,

Free-Slip
With the free-slip BCs, the normal components of the velocities are zero
and we can set them with (3.107) to (3.110). The normal derivative of
the tangential velocity components are to be zero and we achieve this by
mirroring their values across the boundary,
j {1, . . . , jmax } , ` {1, 2, 3}

(West),
(3.115)

v(imax +`,j) = v(imax +1`,j) , j {1, . . . , jmax } , ` {1, 2, 3}

(East),
(3.116)

i {1, . . . , imax } , ` {1, 2, 3}

(South),
(3.117)

i {1, . . . , imax } , ` {1, 2, 3}

(North).
(3.118)

v(1`,j) = v(`,j) ,

u(i,1`) = u(i,`) ,
u(i,jmax +`) = u(i,jmax +1`) ,

Outflow
With the outflow BC, the normal derivatives of both the normal and
tangential velocity components are to be zero. We accomplish this by

45

3 Numerical Methods
mirroring both the normal,
u(0,j) = u(1,j) ,
u(imax +1,j) = u(imax ,j) ,
v(i,0) = v(i,1) ,
v(i,jmax +1) = v(i,jmax ) ,

j {1, . . . , jmax } (West),

(3.119)

i {1, . . . , imax } (South),

(3.121)

j {1, . . . , jmax } (East),

(3.120)

i {1, . . . , imax } (North),

(3.122)

and the tangential velocity components,

j {1, . . . , jmax }

v(0,j) = v(1,j) ,

j {1, . . . , jmax }

v(imax +1,j) = v(imax ,j) ,

i {1, . . . , imax }

u(i,0) = u(i,1) ,

i {1, . . . , imax }

u(i,jmax +1) = u(i,jmax ) ,

(West),

(3.123)

(East),

(3.124)

(South),

(3.125)

(North),

(3.126)

at the boundaries in question.

3.6.2 Level-set Function Boundary Conditions

As in Lerv
ag [2008], the level-set function is mirrored at all boundaries,
(1`,j) = (`,j)
(imax +`,j) = (imax +1`,j)
(i,1`) = (i,`)
(i,jmax +`) = (i,jmax +1`)

j {1, . . . , jmax } , ` {1, 2, 3}

(West)
(3.127)

j {1, . . . , jmax } , ` {1, 2, 3}

(East)
(3.128)

i {1, . . . , imax } , ` {1, 2, 3}

(South)
(3.129)

i {1, . . . , imax } , ` {1, 2, 3}

(North)
(3.130)

This results in an interface that is always normal to the boundary.

3.6.3 Pressure Boundary Conditions

Dirichlet
With this BC we set the pressure equal to a constant value along the
boundaries in question. Since we are free to choose the zero-level of the

46

3.6 Boundary Conditions

pressure, we set
p(0,j) = 0
p(imax +1,j) = 0
p(i,0) = 0

j {1, . . . , jmax }

(West)

(3.131)

(East)

(3.132)

i {1, . . . , imax }

(South)

(3.133)

(North)

(3.134)

j {1, . . . , jmax }

p(i,jmax +1) = 0

i {1, . . . , imax }

Neumann
Due to the choice of projection method we can in most cases set the
normal component of the pressure gradient equal to zero at the boundaries
[Griebel et al., 1998, Section 3.2.10],
p(0,j) = p(1,j) ,
p(imax +1,j) = p(imax ,j) ,
p(i,0) = p(i,1) ,
p(i,jmax +1) = p(i,jmax ) ,

j {1, . . . , jmax }

(West),

(3.135)

j {1, . . . , jmax }

(East),

(3.136)

(South),

(3.137)

i {1, . . . , imax }

(North).

(3.138)

i {1, . . . , imax }

3.6.4 Temperature Boundary Conditions

Dirichlet
With Dirichlet BCs, we want to set the temperature at the west, east,
south or north boundary equal to the constant values Twest , Teast , Tsouth or
Tnorth , respectively. However, as the temperatures are not located directly
on the boundaries, we resort to setting them such that they have the
desired value at the boundary when linearly interpolated. We set
T(0,j) = 2Twest T(1,j) ,

j {1, . . . , jmax }

(West),

(3.139)

T(imax +1,j) = 2Teast T(imax ,j) ,

j {1, . . . , jmax } (East),

(3.140)

T(i,jmax +1) = 2Tnorth T(i,jmax ) ,

i {1, . . . , imax } (North). (3.142)

T(i,0) = 2Tsouth T(i,1) ,

i {1, . . . , imax } (South),

(3.141)

Neumann
With Neumann BCs, we want to impose heat flux densities qwest , qeast ,
qsouth or qnorth into the fluid domain from the west, east, south or north

47

3 Numerical Methods
boundaries respectively. We set
qwest
T(0,j) = T(1,j) +
x(0) ,

j {1, . . . , jmax } (West),

(3.143)

T(imax +1,j)

qeast
= T(imax ,j) +
x(imax ) ,

j {1, . . . , jmax } (East),

(3.144)

T(i,0)

qsouth
= T(i,1) +
y(0) ,

i {1, . . . , imax } (South),

(3.145)

T(i,jmax +1) = T(i,jmax ) +

qnorth
y(jmax ) ,

i {1, . . . , imax } (North).

(3.146)

3.7 Time-Step Restrictions

To ensure stability of the SSP RK schemes, we use standard Courant
FriedrichsLewy (CFL) criteria to calculate the safe length of each time
step. These are limited by a safety factor, the CFL number CCFL h0, 1i.
In this work, all simulations were run with CCFL = 0.5.
As Griebel et al. [1998], we calculate a safe time step for integration
of the velocities (t)~u and a safe time step for the temperatures (t)T
separately and let the time-step length be
t = min {(t)~u , (t)T } .

(3.147)

The manner in which we calculate (t)~u and (t)T depends on whether

we have a single- or two-phase case and, if we have a two-phase case,
whether we model phase transition or not. Before we consider these different cases, we define


xmin = min x(i) ,
(3.148)
i


ymin = min y(j) .
(3.149)
j

3.7.1 Single-Phase Simulations

The stability criterion for ~u in single-phase incompressible flow states that
no fluid particle may travel longer than across one grid cell in the time

48

3.7 Time-Step Restrictions

interval t [Griebel et al., 1998, Section 3.2.4]. This is ensured by selecting
a t satisfying


x(i)
t Cx = min
,
(3.150)
i,j
|u(i,j) |


y(j)
t Cy = min
.
(3.151)
i,j
|v(i,j) |
In cases of viscous flow, we also need to have
(
!)1
1
1
1
t C =
min
.

2 +

2
i,j
2
x(i)
y(j)

(3.152)

We ensure that t satisfies all of (3.150) to (3.152) by setting

(t)~u = CCFL min (Cx , Cy , C ) .

(3.153)

Looking to Griebel et al. [1998, Section 9.4], we need to make t satisfy

!
1
1
(3.154)
2t min

2 +

2 1
i,j
x(i)
y(j)
to ensure stability of the temperature integration. In this work, we accomplish that by setting
(t)T = CCFL

(min {xmin , ymin })2

.
4

(3.155)

3.7.2 Two-Phase Simulations

For two-phase flows, we continue in the spirit of Kang et al. [2000] and
demand that t must satisfy
q

t 
Cc + C + (Cc + C )2 + 4Cg2 + 4Cs2 1,
(3.156)
2
where

Cc = max
i,j

|u(i,j) |
x(i)


+ max
i,j

|v(i,j) |
y(j)


,

(3.157)

49

3 Numerical Methods

= 2 max + ,
s

Cs =
s
Cg =

max
i,j

maxi,j

1
x(i)


k(i,j)

max (+ , ) (min {x

2 +

1
y(j)

2,
min , ymin })

|gy |
|gx |
+
.
xmin ymin

2

(3.158)

(3.159)
(3.160)

Note that in (3.159), Kang et al. [2000] use min (+ , ). However, this
was judged overly conservative and replaced with max (+ , ) by Lerv
ag
[2008]. We then let
(t)~u =

2CCFL
Cc + C +

(3.161)

(Cc + C )2 + 4Cg2 + 4Cs2

In a case where we do not model phase transition, we let (t)T be an

extension of its single-phase counterpart (3.153),
(t)T = CCFL

(min {xmin , ymin })2

.
4 max (+ , )

(3.162)

When we do model phase transition then, according to Gibou et al.

[2002], we must demand that t satisfies
!


2
2

+

1 (3.163)
t max + ,
mini (x(i) )2
minj (y(j) )2
to guarantee stability of the temperature. In this thesis > c , so it is
safe to set
(c min {xmin , ymin })2
.
(3.164)
(t)T = CCFL
4 max (+ , )
If we compare this time-step restriction with (3.162), we see that it is
stricter by a factor c2 . We want to choose c reasonably small to get good
approximations to the spatial derivatives of T , e.g. in (3.81), but the price
for this is safe time steps that are short. In this thesis we have chosen
c = 0.25 to avoid very short time steps.

50

3.8 Estimation of Error and Convergence Order

3.8 Estimation of Error and Convergence Order

In order to quantify the error in an obtained solution with respect to some
reference, we will use one of several possible measures of error.
In the case where the solution a and reference aref are real numbers,
a, aref R, we simply define the error Eabs as the absolute difference
Eabs = |a aref | .

(3.165)

When we have vectors ~a, ~aref Rn , we define the scaled 1-norm E1 ,

E1 =

k~a ~aref k1
,
N

(3.166)

as a measure of error. For a formal definition of norms and their properties,

see e.g. S
uli and Mayers [2006].
We will assume that the error as a function of grid size E (N ) can be
written as a polynomial expansion in 1/N . For a sufficiently fine grid,
that is for large enough values of N , the error of a solution is dominated
by the leading term in this polynomial expansion,
E (N )

C
.
Nn

(3.167)

The number n is what we will refer to as the convergence order.

Taking the logarithm on both sides of (3.167) we get
log10 {E (N )} log10 (C) n log10 (N ) ,

(3.168)

so plotting log10 {E (N )} against log10 (N ) should give a straight line with

slope n. This is one way of estimating the convergence order.
By dividing the error in a solution for one grid size E (N1 ) by the error
in the solution for another E (N2 ) we get an expression that we can solve
for n,


E(N1 )
log10 E(N
2)
  .
n
(3.169)
N2
log10 N
1
Another way of estimating the convergence order is then to evaluate
(3.169) for pairs of N .

51

3 Numerical Methods
To have a quantitative notion of how close a simulated solution is to
steady-state, we define Ess of the field as


(n)
(n`)
maxi,j (i,j) (i,j)
Ess () =
,
(3.170)
t(n) t(n`)
where ` 1 and n is the final time step in the simulation. If the solution
converges to steady-state, then Ess should provide an upper bound for the
simulated t at the end of the simulation.

3.9 Summary
We have presented the numerical methods used to solve the two-phase
flow model. Chorins projection method (Section 3.1) was employed to
get an explicit expression (3.1) for the velocity derivatives t ~u(n) and a
Poisson equation (3.4) for the pressure p(n+1) . The level-set method, used
to keep track of the interface location through the time integration of
the level-set equation (3.11), was presented (Section 3.2) and the finitedifference approximations to the spatial derivative operators (Section 3.4),
and thus to e.g. the temperature derivative t T (n) , were described. We also
described a strategy for handling the phase-transition model numerically
(Section 3.5), by calculating the mass flux density m from the temperature
field using (3.101) and use m to enforce the interface condition on the
velocity field and find the velocity of the interface. The implementation
of the boundary conditions (Section 3.6) was presented and to ensure
stability of the RungeKutta schemes (Section 3.3), we gave expressions
for the maximum length of safe time steps (Section 3.7).

52

4 Numerical Experiments
In this chapter, we describe the eight test cases for which simulations were
run in this thesis. The main purpose of most of these numerical experiments is to verify the methods and their implementation by comparing
results from simulation against analytical or reliable numerical results.
Some of the cases are therefore maybe not so interesting in themselves,
but they may reveal interesting features of the method or errors in the
implementation.

4.1 De Vahl Davis Benchmark Case

Before we can hope to get correct numerical results for simulations of heat
transport in two-phase flow, we should confirm that we are able to obtain
reasonable results for single-phase simulations. To this avail, a widely
used benchmark case was run and the results were compared with results
from de Vahl Davis [1983].
In this case, as illustrated in Figure 4.1, we consider a two-dimensional,
fluid-filled, square cavity of width and height Lx = Ly = 1.0 m. We impose no-slip boundary conditions on the walls. There is a constant high
temperature Twest = 294 K on the western wall, a constant low temperature Teast = 292 K on the eastern wall and thermally insulating north
and south walls. Initially the fluid is at rest and the temperature in the
fluid domain is uniform and equal to Teast . The gravitational acceleration
vector ~g is set to 9.81 m s2 along the y-axis.
The simulations were run with Prandtl number Pr = 0.71 and two
different Rayleigh numbers, Ra = 103 (Case A) and Ra = 106 (Case B),
corresponding to the fluid parameters in Table 4.1. They were run to
t = 5 104 s, where steady-state was verified to have been established.

At the time of writing de Vahl Davis [1983] had 186 article citations according to
Web of Knowledge.

53

4 Numerical Experiments
y

Ly

Twest

Teast

x
Lx
Figure 4.1: Schematic illustration of the de Vahl Davis benchmark case configuration.

The Ra and Pr numbers are

Ra =
Pr

k~g k2 (Twest Teast )L3x

,

(4.1)
(4.2)

for these cases.

4.1.1 Numerical Results

Contour plots of the temperatures obtained on 301 301 grids at t =
5 104 s for both Case A and Case B are shown in Figure 4.2. These
contour plots can be qualitatively compared with Figures 4a and 4d from
de Vahl Davis [1983]. It was verified that the magnitude of Ess did not
exceed 107 in the appropriate SI base units for the T and ~u fields. Looking
at the vertical velocity distribution along the horizontal center lines for
the two cases, we see that it is more sharply peaked at the hot and the
cold walls for Case B than for Case A. This is consistent with the higher
Ra of Case B, which says that ratio of buoyancy forces to viscous forces
are higher for Case B.
A quantitative comparison with the benchmark case was done with respect to the maximum horizontal velocity component umax along the ver-

54

4.1 De Vahl Davis Benchmark Case

294.0

0.8

293.6

0.6

293.2

0.4

292.8

0.2

292.4

y [m]

T [K]

1.0

0.0
0.0

0.2

0.4

0.6

0.8

1.0

292.0

x [m]

(a)
294.0

0.8

293.6

0.6

293.2

0.4

292.8

0.2

292.4

y [m]

T [K]

1.0

0.0
0.0

0.2

0.4

0.6

0.8

1.0

292.0

x [m]

(b)
Figure 4.2: Contour plot of the temperature in the de Vahl Davis bench mark case
for (a) Case A, Ra = 103 and (b) Case B, Ra = 106 . The black arrows illustrate the
velocity field.

55

4 Numerical Experiments

Table 4.1: Fluid properties used in the de Vahl Davis benchmark case.

Unit

Case A

Case B

Pr
Ra

1
1

0.71
103

0.71
106

cp

kg m3
Pa s
J kg1 K1
K1
W K1 m1
K

1.2041
1.7400 105
1.012 103
1.4990 108
2.4801 102
293.0

1.2041
1.7400 105
1.012 103
1.4990 105
2.4801 102
293.0

tical center line of the fluid domain. In the benchmark solution, this
quantity is 7.4268 105 m s1 for Case A and 1.3154 103 m s1 for Case
B.
In Case A, on the 301 301 grid, umax was 7.4279 105 m s1 . This
value of umax was obtained, as in de Vahl Davis [1983], by doing a fourth
order polynomial interpolation of u around the maximum point on the vertical center line and then finding the maximum of the interpolation polynomial. The obtained value corresponded to an error of approximately
0.01% with respect to the benchmark solution. de Vahl Davis [1983] estimates the error in this benchmark solution to be 0.1%. The error in our
solution with respect to the benchmark is then smaller than the estimated
error of the benchmark itself. In that sense, our 301 301 solution is as
close to the benchmark as one could reasonably hope to get. Therefore we
conclude that there was agreement between our solution on the 301 301
grid and the benchmark solution.
Similarly, the simulated umax was 1.3206 103 m s1 for Case B on the
301 301 grid, corresponding to an error with respect to the benchmark
of 0.4%. This is again smaller than the estimated error of the benchmark
solution of 1% and therefore in agreement with the benchmark solution.
The order of convergence was investigated for both cases A and B, with
the 301 301 solutions as reference solutions. The resulting errors Eabs
and orders of convergence n, estimated using (3.169) from page 51, are

56

4.1 De Vahl Davis Benchmark Case

Table 4.2: Convergence table based on the maximum horizontal velocity component
umax along the vertical center line for the de Vahl Davis benchmark Case A with Ra =
103 .

N
-

umax
m s1

Eabs
m s1

n
-

21
31
41
51
61
81

7.454 1005
7.440 1005
7.435 1005
7.432 1005
7.431 1005
7.430 1005

2.65 1007
1.23 1007
6.95 1008
4.48 1008
3.08 1008
1.70 1008

1.98
2.04
2.01
2.09
2.09

Table 4.3: Convergence table based on the maximum horizontal velocity component
umax along the vertical center line for the de Vahl Davis benchmark Case B with Ra =
106 .

N
-

umax
m s1

Eabs
m s1

n
-

21
31
41
51
61
81

1.462 1003
1.431 1003
1.384 1003
1.359 1003
1.346 1003
1.334 1003

1.41 1004
1.11 1004
6.34 1005
3.80 1005
2.49 1005
1.31 1005

0.62
2.00
2.35
2.35
2.26

57

4 Numerical Experiments

6.4
6.6

log10 (Eabs )

6.8
7.0
7.2
7.4
7.6
7.8
1.3

1.4

1.5

1.6
1.7
log10 (N )

1.8

1.9

2.0

1.8

1.9

2.0

(a)
3.4
3.6

log10 (Eabs )

3.8
4.0
4.2
4.4
4.6
4.8
5.0
1.3

1.4

1.5

1.6
1.7
log10 (N )

(b)
Figure 4.3: Log-log plot of the error Eabs in umax against the linear grid size N in
the de Vahl Davis bench mark case for (a) Case A, Ra = 103 , where the solid line
represents a first order polynomial fit and has a slope of approximately 2.03 and (b)
Case B, Ra = 106 , where the solid line represents a first order polynomial fit to the five
rightmost points and has a slope of approximately 2.24.

58

4.2 Two-Phase Heat Conduction

shown in Table 4.2 for Case A and in Table 4.3 for Case B. These tables
show a steady decrease of the Eabs upon grid refinement and an estimated
convergence order that is consistent with the expected value of 2. This
steady decrease in error is visualized in Figures 4.3a and 4.3b.
From the results of the simulations of the de Vahl Davis benchmark
case we can conclude that the implementation is able to reproduce the
benchmark results from de Vahl Davis [1983] on the 301 301 grid. Also,
a convergence study indicates that the order of convergence is consistent
with the expected value of 2. This gives reason to believe that the current
implementation of the discretization of the advection-diffusion equation
and the Boussinesq coupling to the NavierStokes equations for singlephase flow cases is correct.

4.2 Two-Phase Heat Conduction

A natural first test of the two-phase CSF and GFM discretizations of the
diffusion term in the advection-diffusion equation (2.11) on page 8 is to
solve this in the absence of convection. In other words, to solve the heat
equation. For simplicity we also restrict ourselves to one dimension. In
this case (2.11) reduces to


d
dT

= 0,
(4.3)
dy
dy
in steady-state.
We consider a cavity of height L containing two phases, Phase 2 on top of
Phase 1 and a horizontal boundary between them at y = L. The relevant
fluid properties are given in Table 4.4. In order to have no convection, the
fluid is initially at rest and the gravitational vector ~g is set to zero. The
temperature is initially uniform and equal to 290 K.
This situation was simulated with two sets of boundary conditions, Case
A and Case B. First it was run with constant-temperature boundary conditions (Dirichlet), Tnorth = 285 K on the north and Tsouth = 295 K on the
south wall and insulating western and eastern walls (Case A) and second
with constant heat flux densities qsouth = qnorth = 1.0 W m2 (Neumann)
and again insulating eastern and western walls (Case B). Due to our convention from Section 2.7.3, we have heat flowing into the domain through
the southern wall and out through the northern wall.

59

4 Numerical Experiments

Table 4.4: Relevant fluid properties used in the two-phase heat conduction test case.

cp

Unit

Phase 1

Phase 2

kg m3
J kg1 K1
W K1 m1

2.0
100.0
0.04

1.0
100.0
0.01

Although the case was simulated on a two-dimensional grid, the boundary conditions and initial configuration should lead to an x-independent
solution and the temperature field in steady-state along any constant-x
line should be governed by (4.3). The case was run to t = 1 104 s, where
steady-state was verified to have been obtained with the GFM. The results
were then compared to an analytical solution.

4.2.1 Analytical Solution

One of the appealing aspects of (4.3) is that it can quite easily be solved
analytically. By differentiation and insertion it is easy to confirm that the
following linear temperature profile is a solution to (4.3),
(
T (y) = C0 + C1 y if y
.
(4.4)
T (y) =
T + (y) = C0+ + C1+ y if y +
We assume that the interface is located at a known location L, where
[0, 1].
To determine the unknown coefficients of (4.4), we need to use the
boundary conditions and impose continuity of the temperature field and
continuity of the heat flux. From the Dirichlet boundary conditions in
Case A, that is constant and known temperatures Tsouth and Tnorth at
y = 0 and y = L respectively, we have
T (0) = Tsouth ,
+

T (L) = Tnorth ,

(4.6)

(4.7)

T (L) = T (L) ,



+
dT 
+ dT 

=
,
dy L
dy L

60

(4.5)

(4.8)

4.2 Two-Phase Heat Conduction

which determines the four coefficients C0 ,C1 ,C0+ and C1+ as
C1 =

T
Tsouth
 north
,

L + {1 } +

(4.9)

C0 = Tsouth ,

C1+ =
C ,
+ 1
C0+ = Tnorth C1+ L.

(4.10)
(4.11)
(4.12)

For the Neumann boundary conditions in Case B, with constant heat

flux density q = |qsouth | = |qnorth | at the top and bottom walls, continuity
of heat flux and continuity of temperature, we have


dT 
q =
(4.13)
dy 0

+
+ dT 
q =
(4.14)
dy 
L

T (L) = T + (L) ,



+
dT 
+ dT 

=
dy L
dy L

(4.15)
(4.16)

which determines the three coefficients C1 ,C0+ and C1+ as

q
,

q
= +,

C1 =
C1+

C0+ = C0 + Lq

(4.17)
(4.18)


1
1

+


.

(4.19)

C0 , again equal to T (0), is undetermined by these boundary conditions.

4.2.2 Numerical Results

Since there is no gravity, there should be no advection and the fluid velocity should remain zero everywhere at all times. This was also found to be
true for all simulations run. Thus the advective term in (2.11) was zero
in the whole domain in all simulations.

61

4 Numerical Experiments
The analytical steady-state solution for the temperature profile is piecewise linear with a kink at the boundary, both for Case A and Case B. Thus
the diffusive term in (2.11) should also be zero in the whole domain. In
the absence of advection, we see from (2.11) that a diffusive term that is
zero in the entire domain is necessary to get a steady-state solution. If it
is not zero then t T 6= 0.
Continuum Surface Force Method
The temperature profiles obtained with the CSF method for Case A are
shown in Figure 4.4a for different grid sizes. These simulations did reach
steady-state in the sense that Ess (T ) was zero. This is consistent with the
linear temperature profile obtained from the simulations.
However, as is apparent from Figure 4.4a, the temperature profile is
not in agreement with the analytical solution and grid refinement did
not seem to produce better results. Also, the spatial derivative of the
temperature profile is physically unacceptable. To see why, consider the
heat flux dT /dy at y = 0 and at y = L. Four times as much heat flows
in through the southern wall at y = 0 compared to the amount of heat
that flows out through the northern wall at y = L. So there is a net heat
flow into the domain, yet the domain temperatures do not change. This
is a clear violation of energy conservation.
This leads us to conclude that the CSF discretization of the diffusive
term in the advection-diffusion equation failed to reproduce the analytical
solution in the two-phase heat conduction case with Dirichlet boundary
conditions.
The temperature profile obtained with the CSF method in Case B is
shown in Figure 4.4b for different grid sizes. With Neumann boundary conditions, the analytical steady-state temperature profile is only
determined up to the addition of a constant. Specifically, the constant
C0 = T (0) is undetermined by the boundary conditions. For the sake of
comparison, this constant is here taken to be the calculated south wall
temperature of the 10 80 simulation.
In contrast to Case A with CSF, these simulations did not reach steadystate. That is, t T averaged in the whole domain did not approach zero
during the time integration, but seemed to approach a constant, negative
value instead. Looking at (2.11), we see that a local negative curvature in

62

4.2 Two-Phase Heat Conduction

296
294

T [K]

292
290
288
286
284
0.0

Analytical
1080
1020
1010
0.2

0.4

0.6

0.8

1.0

0.6

0.8

1.0

y [m]

(a)
120
110

T [K]

100
90
80
70
60
50
0.0

Analytical
1080
1020
1010
0.2

0.4
y [m]

(b)
Figure 4.4: Comparison between the analytical solution to the one-dimensional heat
equation and the numerical solution with different grid sizes and the CSF method for
(a) constant temperature (Dirichlet) boundary conditions (Case A) and (b) constant
heat flux (Neumann) boundary conditions (Case B).

63

4 Numerical Experiments
the temperature field is equivalent to a temperature drain. The drain of
temperature is then consistent with the negative curvature of the temperature profile in Figure 4.4b, but still unphysical and not in agreement with
the analytical solution. To see why it is unphysical, we again consider the
heat flux at the boundaries. The heat flux into the domain at the south
wall and out of the domain at the north wall are in this case equal, as
posed by the boundary conditions, yet the average domain temperature
does not approach a constant value. This is a clear violation of energy
conservation. The CSF method seems to introduce an artificial curvature
in T that drains temperature from the inner domain.
This leads us to conclude that the CSF discretization of the diffusive
term in the advection-diffusion equation failed to reproduce the analytical
solution in the two-phase heat conduction case with Neumann boundary
conditions.
Ghost-Fluid Method
The temperature profile obtained with the GFM in Case A is shown in Figure 4.5a for different grid sizes. These simulations did reach steady-state,
Ess (T ) was zero, consistent with the obtained piecewise linear temperature
profiles. For all grid sizes E1 with respect to the analytical solution was
zero. The GFM method was thus able to obtain the exact solution of the
one dimensional heat equation in Case A . This leads us to conclude that
the GFM discretization of the diffusive term in the advection-diffusion
equation produced the correct solution to the one dimensional heat equation with Dirichlet boundary conditions.
The temperature profile obtained with the GFM in Case B is shown
in Figure 4.5b for different grid sizes. These simulations did also reach
steady-state, Ess (T ) was zero, consistent with the obtained piecewise linear
temperature profiles. For all grid sizes E1 was approximately 4 1015 K,
so the correct temperature profiles were obtained to machine precision for

It may seem strange and even suspicious that the E1 turned out to be zero and not
a small number on order of magnitude with the machine precision. The reason for
it was probably that, for this particular case and choice of grid point locations,
the analytical solution evaluated in the grid points consisted of rational numbers
that could be exactly represented within the machine precision. Alternatively, the
difference in the numerical solution to the analytical solution was small enough to
cause underflow in the double precision data type used to store it.

64

4.2 Two-Phase Heat Conduction

296
294

T [K]

292
290
288
286
284
0.0

Analytical
1080
1020
1010
0.2

0.4

0.6

0.8

1.0

0.6

0.8

1.0

y [m]

(a)
310
300

T [K]

290
280
270
260
250
240
0.0

Analytical
1080
1020
1010
0.2

0.4
y [m]

(b)
Figure 4.5: Comparison between the analytical solution to the one-dimensional heat
equation and numerical solutions with different grid sizes and the GFM method for (a)
Case A, constant temperature (Dirichlet) boundary conditions and (b) Case B, constant
heat flux (Neumann) boundary conditions.

65

4 Numerical Experiments
y

Phase 2

Ly

(1 ) Ly

q
Phase 1

Ly

Lx
Figure 4.6: Schematic illustration of the two-layer configuration in Wang et al.s convection case.

all grid sizes. This leads us to conclude that the GFM discretization of
the diffusive term in the advection-diffusion equation produced the correct
solution to the one-dimensional heat equation with Neumann boundary
conditions.
This test case shows the advantage of treating the interface sharply with
the GFM instead of with the CSF method when solving the heat equation.
The GFM is able to obtain temperature profiles with an error limited by
machine precision regardless of grid size while those produced by the CSF
method are, at best, questionable.

4.3 Wang et al.s Two-Layer Convection Case

As a test for temperature driven two-phase flow, the case presented by
Wang et al. [1991] was simulated. A two-dimensional, rectangular cavity
of length Lx , height Ly and aspect ratio A = Lx /Ly contains two phases,
Phase 2 on top of Phase 1 in a two-layer configuration. A constant and
uniform heat flux density qwest = q points into the domain on the western
wall and a heat flux density qeast = q points out of the domain on the
eastern wall, as illustrated in Figure 4.6. The northern and southern walls
are insulating.
The size of the domain used in the numerical simulation was Lx =

66

4.3 Wang et al.s Two-Layer Convection Case

Table 4.5: Fluid parameters used in the numerical simulation of Wang et al.s two-layer
convection case.

cp

Unit

Phase 1

Phase 2

kg m3
Pa s
J kg1 K1
W K1 m1
K1
K

1.5
2.0 105
120.0
0.013
1 104
290

1.0
1.0 105
100.0
0.010
1 103
290

0.8 m and Ly = 0.1 m, with the interface located at the constant-y line
defined by y = Ly = 0.05 m. The initial temperature was set to 290 K
in the entire inner domain and the gravitational acceleration vector ~g
was set to 9.81 m s2 along the y-axis. The heat flux density q used was
1.0 W m2 . The fluid properties are given in Table 4.5.
This case was run to t = 103 s with the GFM and compared to the
analytical solution by Wang et al. [1991].

4.3.1 Analytical Solution

In the limiting case where A , Wang et al. [1991] showed that one
can use the steady-state and the parallel flow assumptions to obtain an
analytical solution to the NavierStokes equations (2.2) and (2.10) and
advection-diffusion equation (2.11) that is valid near the middle of the
cavity. Here, we will present this solution and a derivation of the expressions needed to verify it. Thus we elaborate on some of the points in the
derivation in more detail than what is done in Wang et al. [1991].
The Partial Differential Equations
We start the derivation by introducing the stream function ,

,
y

v = .
x

u =

(4.20)
(4.21)

67

4 Numerical Experiments
Inserting this definition of into the incompressible flow assumption (2.2),
we see that (2.2) is automatically satisfied through the definition of .
Further, to simplify our notation a bit and to be more consistent with
Wang et al. [1991], we introduce the binary differential operator J . This
operator is defined for two scalar functions a and b as
J (a, b) =

a b
b a

.
y x y x

(4.22)

Consider now the advection-diffusion equation (2.12) on page 8, which

is valid within each phase. In steady-state, that is when t T = 0, we can
restate it in terms of J and as
J (, T ) = 2 T.

(4.23)

Next, we consider the momentum balance equation (2.6) on page 6, also

valid within each phase. In steady-state, t ~u = 0, and (2.6) becomes
1
(~u ) ~u = p + 2 ~u + ~g .

(4.24)

Applying on both sides of (4.24), we get [Landau and Lifshitz, 1987,

15]
(~u { ~u}) + 2 ( ~u) + ~g = 0.
(4.25)
Considering the first term, we employ standard vector calculus relations
[Rottmann, 2003] and (2.2) to rewrite it as
(~u { ~u}) = ({ ~u} ) ~u (~u ) { ~u}

+ ~u ( { ~u}) { ~u} ( ~u) , (4.26)

= ({ ~u} ) ~u (~u ) { ~u} ,

(4.27)

(~u ) { ~u} ({ ~u} ) ~u = 2 { u} + ~g .

(4.28)

and then (4.25) becomes

We can easily rewrite ~u in terms of the stream function ,

~u = 2
z.

68

(4.29)

4.3 Wang et al.s Two-Layer Convection Case

Now, using that ~g = g y and (4.29) before dividing by z, we can recast
(4.28) as an equation for





g

2
2 = 2 2 +
,
y x
x y
x

(4.30)

which is equivalent to



g
.
J , 2 = 2 2 +
x

(4.31)

Finally, employing the Boussinesq approximation (2.16) on page 9, we get

J , 2

g (1 {T T }) ,
= 2 2 +
x


T
= 2 2 g
.
x

We now have two coupled PDEs,

give their solutions different names
(
+
=

(
T+
T =
T

(4.32)
(4.33)

(4.23) and (4.33), for and T and we

in the two phases,
if (x, y) +
,
if (x, y)

(4.34)

if (x, y) +
.
if (x, y)

(4.35)

In order to use expressions for and T that are notationally easily

comparable to those from Wang et al. [1991], we introduce the relative
fluid properties

+
+
+
,

=
,

=
,

(4.36)

and the dimensionless Rayleigh and Prandtl numbers

g + qL4y
,
+ + +
g qL4y
= ,

+
,
+

= .

Ra + =

Pr + =

(4.37)

Ra

Pr

(4.38)

69

4 Numerical Experiments
We also introduce dimensionless quantities x
, y, u
, and T,
x
=

1
1
= Ly ~u, = 1 , T = T.
x, y =
y, ~u
Ly
Ly

qLy

and obtain dimensionless PDEs for and T,

(
)





2
2

J ,
= Pr
Ra
,
x

)
(




+

T
+
2
+
+
2
+
2
+

Ra
,
J ,
=
Pr

x



= 2 T ,
J , T


J + , T+
=
2 T+ .

(4.39)

(4.40)
(4.41)
(4.42)
(4.43)

The spatial derivatives in these equations are now, of course, taken with
respect to the dimensionless coordinates x
and y.
The Boundary and Interface Conditions
We pose no-slip BCs on the velocity field. Thus,

= 0,
y

v =
= 0,
x

u
=

(4.44)
(4.45)

at the boundaries.
Since is continuous, otherwise the velocity field would not be welldefined, must have the same constant value along the entire boundary.

Then, since the velocity is invariant upon the addition of a constant to ,

we are free to choose the zero-level of such that

= 0
at the boundary.

70

(4.46)

4.3 Wang et al.s Two-Layer Convection Case

At the interface, the velocity field must be continuous and well-defined.
Also, the shear stress must be continuous. Assuming that the interface is
stationary and located at y = , we get the following conditions on
= + ,
+

=
,
y
y

+
=
= 0,
x

2 +
2
=

,
y2
y2

(4.47)
(4.48)
(4.49)
(4.50)

at y = .
Since the north and south walls are insulating and there is constant and
uniform heat flux density through the east and west walls,
T
y
T+

y
T
x

T +

= 0,

at y = 0,

(4.51)

= 0,

at y = 1,

(4.52)

= 1,

at x
= A/2,

(4.53)

= 1,

at x
= A/2.

(4.54)

At the interface, we must have continuity of the temperature and the

heat flux density. Thus,
T = T+ ,
T
T+
=

,
y
y

(4.55)
(4.56)

at y = .
The Solution
In the parallel flow approximation [Wang et al., 1991], we assume that
= (
y) ,

(4.57)

71

4 Numerical Experiments
+ = + (
y) ,

T
= C0 + C1 x
+ + (
y) ,
+
+
+

T
= C +C x
+ (
y) .
0

(4.58)
(4.59)
(4.60)

From the form of these equations, we see that we must have C0 = C0+ =
C0 and C1 = C1+ = C1 to have continuity of temperature at the interface.
In the parallel flow approximation, the PDEs for and T reduce to the
following ordinary differential equations (ODEs)
d4
d
y4
d4 +
d
y4
d2
d
y2
d2 +
d
y2

= Ra C1 ,

(4.61)

Ra + C1 ,

(4.62)

=
=

d
C1 ,
dy
1 d+
C1 .

dy

(4.63)
(4.64)

By differentiation and insertion it can be confirmed that

Ra C1 2
=
y (
y ) y + + ,
(4.65)
24


Ra C1

+ =
{1 y}2 ( y)
{2 y } + + ,
(4.66)
24

 5

Ra (C1 )2
y 5
y3 3
=
2
24
5
3
 4

3
3
4
y
y
+

,
(4.67)
4
3
(
Ra (C1 )2 {1 y}5 {1 }5
+
=
24

5
!
{1 y}4 {1 }4
{1 y}3 {1 }3
(1 )2
+ +
3
4
!)
{1 y}3 {1 }3
(1 )
,
(4.68)
3

72

4.3 Wang et al.s Two-Layer Convection Case

where
=

{1 }3 2 (4
+ 5 {1 })
,
2 ({1 } +
)

(4.69)

(4.70)

{1 }2 (4 {1 } + 5
)

2 {1 } ({1 } +
)

are solutions to these ODEs with the posed boundary conditions. Using
the definition of (4.20), we get expressions for the velocity


Ra C1  3
4
y + 3
y 2 2
y + ,
24 

Ra
C1
u+ (
y) =
(1 y)2 (
y )
24

+ 3
y 2 + 2
y {2 + } 1 2



+
{2 y } +
.

u (
y) =

(4.71)

(4.72)

In order to determine the coefficient C1 , we must demand that the sum

of the advective and conductive heat transfer across any constant-x line is
equal to the heat transfer through the side walls. This yields the integral
equation
Z
Z

1 + +

u
d
y+
u
d
y C1 ( +
{1 }) = 1.
(4.73)

0
This equation was solved numerically for C1 using Simpsons method [S
uli
and Mayers, 2006, Section 7.5] to evaluate the integrals and the bisection
method [S
uli and Mayers, 2006, Section 1.6] to find the appropriate root
of the resulting polynomial. The coefficient C0 , equal to the dimensionless temperature at the interface, is undetermined by the posed boundary
conditions.

4.3.2 Numerical Results

The analytical solution for this case is, strictly speaking, valid only when
A . However, Oueslati et al. [2012] claimed that using A = 8, as we
do, was sufficient to accurately simulate the situation considered in the

73

4 Numerical Experiments
0.12

0.10

Analytical
101100
5150
2120

y [m]

0.08

0.06

0.04

0.02

0.00
0.03

0.02

0.01

0.00
u [m s1 ]

0.01

0.02

0.03

Figure 4.7: Comparison of selected numerically obtained velocity profiles along the
vertical center line in Wang et al.s two-layer convection case with the analytical solution.
It is evident that the numerical solutions approach the analytical solution upon grid
refinement.

steady-state analytical solution. It was verified that Ess (u) did not exceed
107 m s1 in magnitude.
The resulting horizontal velocity component along the vertical center
line of the cavity is shown in Figure 4.7 for a selection of grid sizes. We
observe a buoyancy-driven convection cell in the top layer (Phase 2), similar to that in the de Vahl Davis case, where the fluid moves clockwise,
and we see a viscosity-driven cell in the bottom (Phase 1), where the fluid
moves anti-clockwise. To plotting accuracy, there is good agreement between the analytical solution and the numerical results for the 101 100
grid. Furthermore, the numerical solutions appear to approach the analytical solution when the grid is refined.

74

4.3 Wang et al.s Two-Layer Convection Case

Table 4.6: Convergence table based on the maximum horizontal velocity component
umax along the vertical center line for Wang et al.s two-layer convection case.

Ny
-

umax
m s1

Eabs
m s1

n
-

20
30
40
50
60
80
100

2.487 1002
2.442 1002
2.421 1002
2.405 1002
2.395 1002
2.386 1002
2.380 1002

1.31 1003
8.66 1004
6.49 1004
4.96 1004
3.97 1004
3.02 1004
2.43 1004

1.03
1.01
1.20
1.21
0.95
0.97

To do a quantitative comparison between the analytical solution and

the numerical results, we consider the maximum of the horizontal velocity
component umax along the vertical center line. As in the de Vahl Daviscase, we locate the umax from the numerical results by doing a fourth-order
polynomial interpolation. In the analytical solution, umax was found to
be 2.3557 102 m s1 . A convergence table based on umax for different
grid sizes, with the analytical value as reference, is shown in Table 4.6.
According to the estimated orders of convergence n from this table, the
convergence to the analytical solution upon grid refinement is of first order.
This is expected since the GFM is a first order method. The convergence
to the analytical solution is also illustrated in Figure 4.8.
From this test case we can conclude that our implementation produced
results in agreement with the analytical steady-state solution for a case
involving temperature-driven two-phase flow. Also, the performed convergence study showed the expected first-order convergence. This gives
reason to believe that the current implementation of the discretization
of the advection-diffusion equation and the Boussinesq coupling to the
NavierStokes equations is correct also for two-phase cases where the interface is treated with the GFM.

75

4 Numerical Experiments
2.8
2.9

log10 (Eabs )

3.0
3.1
3.2
3.3
3.4
3.5
3.6
3.7
1.3

1.4

1.5

1.6
1.7
log10 (Ny )

1.8

1.9

2.0

Figure 4.8: Log-log plot of the error Eabs in umax against the linear grid size Ny in Wang
et al.s two-layer convection case. The solid line represents a first-order polynomial fit
and has a slope of approximately 1.06.

4.4 Lava Lamp

In this case we consider a situation which is very similar to that found in
commercial lava lamps. In these decorative contraptions, two immiscible
fluids are contained in a closed vessel. At room temperature, one of the
fluids (Phase 1) is denser than the other (Phase 2), but with a larger coefficient of thermal expansion. For visual effect, Phase 1 is usually opaque
and dyed with some color while Phase 2 is more transparent. As the fluids
are heated from below, Phase 1 eventually becomes less dense than Phase
2 and, under the right conditions, it forms a rising blob . The blob is
subsequently cooled by the surrounding Phase 1 and the container walls.
Eventually, it drops and the cycle starts over, creating the mesmerizing
and almost hypnotic effect that we are used to seeing in lava lamps.
Recently, Gy
ure and J
anosi [2009] conducted an experimental study
on silicone oil in a saline solution to describe the mechanisms of lava
lamp convection. Under specific conditions, they observed blob exchange

The author is uncertain as to whether the term blob has a well-defined scientific
meaning or not. Still, it is used in the lava lamp literature [Gy
ure and J
anosi, 2009]
and therefore also in this work.

76

4.4 Lava Lamp

Table 4.7: Fluid parameters used in the lava lamp case.

cp

Unit

Phase 1

Phase 2

kg m3
Pa s
J kg1 K1
W m1 K1
K1
K

1080
0.05
1591
0.146
470 106
293

1074
1.08 103
3993
0.596
280 106
293

oscillations in which a blob of Phase 1 would rise from the bottom of the
vessel, attach itself to the top, be cooled and sink down, with regular time
periods. We will use the similar parameters to those of Gy
ure and Janosi
[2009], as far as our model permits it, and try to replicate the oscillating
behavior in our simulations.
The simulation was run on an axisymmetric 75 168 grid. With this
grid, the domain is a cylinder. The eastern wall represents its entire outer
face and the western wall its central axis of symmetry. The height of
the cylinder was Ly = 0.28 m and the radius was Lx = 0.125 m. The
coefficient of surface tension was set
to = 0.04 N m1 , the gravitational
acceleration to ~g = 9.81 m s2 y and the rest of the relevant fluid
parameters are given in Table 4.7.
We imposed no-slip BCs for the velocity and Dirichlet for the temperature at all walls, except the western which, due to symmetry, needed to
have free-slip and be thermally insulating. On the north wall we set the
temperature to Tnorth = 312 K, on the south wall we set Tsouth = 332 K
and on the eastern we set Teast = 321.9 K. The fluids were initially at
rest with a uniform temperature of 321.9 K. The initial interface was a
spherical shell of radius 0.0829 m, centered in the south east corner.

4.4.1 Numerical Results

The simulated interface location is shown in Figure 4.9 at different times.
From this figure, we can observe the aforementioned blob exchange oscillations. According to the experimental measurements by Gy
ure and

77

4 Numerical Experiments
Janosi [2009], Figure 9, we should get that the time period of the oscillations is between 9.2 102 s and 1.8 103 s. The period of the simulated
oscillations was found to be 1.4 103 s, and thus in good agreement with
the experimental results.
However, in spite of having similar time periods, we see, e.g. from comparing Figure 4.9 with Figure 4 from Gy
ure and Janosi [2009], that the
simulated blob oscillations do not look the same as those in the experiments. The reasons for the deviations are probably many, as there are
many aspects of lava lamp convection that are not accounted for in our
model.
For instance, our model assumes uniform viscosities in both phases.
However, according to Gy
ure and Janosi [2009], the viscosity of the oil
(Phase 1) is almost exponentially decreasing with increasing temperature.
This strongly temperature-dependent viscosity is not accounted for by our
model.
Also, we choose constant temperature boundary conditions as the driving force for the lava lamp convection. As e.g. the temperature plot Figure
5 from Gy
ure and J
anosi [2009] shows, this might not be realistic. The
boundary conditions used here were chosen to guarantee fluid domain
temperatures in the region around where Phase 1 and Phase 2 are equally
dense.
The boundary conditions that we use for the level-set function give a
contact angle between the interface and the boundary walls that is always
90 . As we see from Figure 4 in Gy
ure and Janosi [2009], the oil (Phase
1) wets the walls of the vessel slightly in the experiments and the 90
contact angle from the simulations is therefore not correct. The boundary
conditions on the level-set function are chosen the way they are, and not
with the aid of a force balance equation like (61.12) from Landau and
Lifshitz [1987], because we have not included any material properties of
domain walls in our model.
From this test case, we can conclude that our two-phase flow model
reproduced the experimentally observed blob exchange oscillations and
their time periods. This gives reason to believe that the implementation
was correct also on axisymmetric grids. Still, the simulated behavior can
not be said to be in full agreement with the experimental results and
neither can we expect it to be, since many of the physical aspects of lava
lamp convection that were presented by Gy
ure and Janosi [2009] are not

78

0.25

0.25

0.20

0.20
y [m]

y [m]

4.4 Lava Lamp

0.15

0.15

0.10

0.10

0.05

0.05

0.00

0.10 0.05 0.00 0.05

x [m]

0.00

0.10

0.25

0.25

0.20

0.20

0.15

0.15

0.10

0.10

0.05

0.05

0.00

0.10 0.05 0.00 0.05

x [m]

(c)

0.10

(b)

y [m]

y [m]

(a)

0.10 0.05 0.00 0.05

x [m]

0.10

0.00

0.10 0.05 0.00 0.05

x [m]

0.10

(d)

Figure 4.9: The interface location in the lava lamp case at (a) t = 500 s, where the
increased temperature in Phase 1 causes its shape to deviate slightly from a the initial
half-sphere, (b) at t = 766 s, where a rising blob has formed, (c) at t = 775 s, where the
rising blob attaches itself to the top of the vessel, and (d) at t = 1214 s, where the blob
has been cooled and sinks back down.

79

4 Numerical Experiments
accounted for in our model.

4.5 Vaporization with Uniform Mass Flux

As a first test of the phase-transition model and its ability to preserve the
discontinuities in ~u and p at the interface, a one-dimensional case with
uniform mass flux density m was simulated. We emphasize that in this
particular case m was set to a given value, initially 2.0 kg m2 s1 , and
not calculated from the temperature field.
We consider a fluid domain with length L = 2.0 m. In the initial configuration, we have a liquid film in the middle of the domain (Phase 1), with
interfaces to the gaseous phase (Phase 2) at x = 0.5 m and at x = 1.5 m.
We initialize the liquid film to be at rest, we
set ~u = 0 m s1 , and we
initialize the gas velocities as ~u+ = 1.0 m s1 n
, so that they satisfy the
jump condition (2.38) on page 14. The relevant fluid properties are given
in Table 4.8.
On both the eastern and western boundaries, we impose outflow BCs on
the velocities and Dirichlet BCs on the pressure. On the northern and the
southern boundaries we pose free-slip BCs on the velocity and Neumann
BCs on the pressure. Even though the simulations were performed on a
two-dimensional 21 21 grid, the initial and boundary conditions make
this an effective one-dimensional problem and the simulated results will
be independent of y.
The simulation was run for two slightly different cases A and B. Case A
was run from t = 0 s to t = 0.15 s with m = 2.0 kg m2 s1 for the entire
run and Case B from t = 0 s to t = 8.0 105 s where m was increased to
2.1 kg m2 s1 after t = 5.0 105 s. To simplify the analysis of the time
stepping, the forward Euler method with a constant time-step length of
t = 2 105 s was used. It was confirmed that this time-step length was
consistent with the CFL criteria from Section 3.7.

4.5.1 Analytical Solution

Let us consider Case A first. Due to the initial configuration, the time
derivatives of ~u+ and ~u are zero at all

times and the velocities will remain

1
+
1
~u = 0 m s and ~u = 1.0 m s
n
. However, because there is mass
transport from the liquid to the gaseous phase, the liquid film will shrink

80

4.5 Vaporization with Uniform Mass Flux

Table 4.8: Fluid parameters used in the vaporization with uniform mass flux case.

Unit

Phase 1

Phase 2

kg m3
Pa s

2.0
1.0 105

1.0
1.0 105

and the interfaces

will move from their initial positions with velocity w
~=


1.0 m s1 n
.
Since the velocity field is uniform on both sides of the interface and
there is no interface curvature, the only contribution to the pressure jump
(2.39) from page 14 is m2 [1/] and thus [p] = 2.0 Pa. This means that
the pressure in the liquid will be constant and equal to 2.0 Pa and the
pressure in the gas will be constant and equal to 0.0 Pa.
Case B will develop similarly to Case A until m is increased at t =
5.0 105 s. After
this time, the gas velocities at the interface will have
to be 1.05 m s1 n
to satisfy the jump condition (2.38). Because the
fluids are incompressible, the change in velocity at the interface should
be instantly propagated

through the entire domain. Therefore we should
+
1
have ~u = 1.05 m s
n
in the entire gas phase, not just at the interface,
immediately after m has increased.

4.5.2 Numerical Results

The simulated velocity and pressure fields from Case A at t = 0.0 s and at
t = 0.15 s are shown, together with the analytical solution, in Figure 4.10.
We observe that the implementation shows correct behavior [Gibou
et al., 2007] in that it has preserved the sharp jumps in ~u and p at the
interface even as the grid nodes change character from liquid to gas.
Further, we observe that there is agreement between the results from
the simulation and the exact analytical results, to plotting accuracy. More
quantitatively, the error E1 in the pressure field was 2.33 1010 Pa and
E1 in the velocity field 5.32 1014 m s1 with respect to the analytical
solution at t = 0.15 s. The implementation was therefore able, even on
this coarse 21 21 grid, to produce results whose accuracy was limited by
machine precision and the tolerance in the Poisson solver.

81

4 Numerical Experiments

1.5
1.0

t = 0.00 s
t = 0.15 s

u [m s1 ]

0.5
0.0
0.5
1.0
1.5
0.0

0.5

1.0
x [m]

1.5

2.0

1.5

2.0

(a)
3.0
2.5

t = 0.00 s
t = 0.15 s

p [Pa]

2.0
1.5
1.0
0.5
0.0
0.5
0.0

0.5

1.0
x [m]

(b)
Figure 4.10: Plot of (a) the velocity profile and (b) the pressure field in the initial
configuration t = 0.0 s and at t = 0.15 s in the vaporization with uniform mass flux
Case A. The analytical solution is shown as solid lines.

82

4.5 Vaporization with Uniform Mass Flux

The velocity fields from Case B at t = 6.0 105 s, at t = 8.0 105 s
and the analytical solution are shown in Figure 4.11a. From this figure, we
immediately see that the simulated velocity field at t = 6.0 105 s does
not agree with the analytical solution. This must, however, be expected
due to the explicit method. The velocity profile at t = 6.0 105 s was
calculated with m = 2.0 kg m2 s1 from the previous time step, at t =
4.0 105 s. The new value m = 2.1 kg m2 s1 was not used until t =
6.0 105 s to calculate the velocity at t = 8.0 105 s. Therefore, the
velocity at t = 8.0 105 s is the first to be consistent with the new m.
The error E1 in the velocity field at t = 8.0 105 s was 4.38 106 m s1 .
The pressure field at t = 8.0 105 s is shown in Figure 4.11b. Its large
numerical value with respect to the pressure fields in Figure 4.11a alone
is enough to make it seem strange and rouse suspicion. Still, it may be
explained as follows. When m changes from one time step to the next and
we set the ghost values according to (3.104) and (3.105) from page 43, then
the gas ghost velocity ~u
g in the liquid has increased with respect to the
gas velocity on the gas side and thus we no longer have ~u divergence-free.
The pressure field must then adapt to restore the divergence-freeness, as
we demand in the projection step (3.1), and the result is what we see in
Figure 4.11b.
This speculation can be supported by evaluating x p/ anywhere on the
interior of the gas phase west of the film. Doing so, we get approximately
2500 m s2 . This is consistent with the t u needed here to get a uniform velocity

field and thus

enforce the divergence-freeness of ~u, namely
0.05 m s1 / 2.0 105 s = 2500 m s2 .
The strange pressure field at t = 8.0 105 s is therefore probably an
artifact of the incompressible model. Its strange shape and large numerical
values are necessary to mediate the information about the new velocity
jump at the interface, caused by the new mass flux value, to the entire
domain so that the velocity field keeps its divergence-freeness.
We should also mention that this Case B, with a large, instantaneous
jump in m is not physically realistic. If it were, m would be determined
from the spatial derivatives of the temperature field. These should be
continuous in time and space and should therefore not produce such large
jumps. Still, we must expect m to fluctuate and information about these
fluctuations must be instantaneously propagated to the entire domain to
keep the velocity field divergence free. This somewhat exaggerated exam-

83

4 Numerical Experiments

1.5
1.0

t = 6.00 1005 s
t = 8.00 1005 s

u [m s1 ]

0.5
0.0
0.5
1.0
1.5
0.0

0.5

1.0
x [m]

1.5

2.0

1.5

2.0

(a)
1600
1400
1200

p [Pa]

1000
800
600
400
200
0
0.0

t = 8.00 1005 s
0.5

1.0
x [m]

(b)
Figure 4.11: Plot of (a) the velocity field at t = 6.0 105 s and at t = 8.0 105 s and
(b) the pressure field at t = 8.0 105 s in the vaporization with uniform mass flux Case
B.

84

4.6 Vaporization with Physical Mass Flux

Table 4.9: Fluid parameters used in the numerical simulation of vaporization with
physical mass flux case.

cp

Unit

Phase 1

Phase 2

kg m3

1.0
1.0 105
1.0
1.0

0.2
1.0 105
5.0
1.0

Pa s
J kg1 K1
W K1 m1

ple indicates that the implementation is robust enough to handle this.

From this test case we can conclude that the current implementation
succeeded in preserving the discontinuities in ~u and p, even when grid
nodes changed character from liquid to gas, in the case of uniform mass
flux density m. It also succeeded in propagating an instantaneous change
in the velocity field at the interface to the entire domain instantaneously as
is necessary in the incompressible model. However, this reminds us of the
fact that the pressure in the incompressible model is not a thermodynamic
state variable, but merely a field that we choose to be whatever is needed
to keep the velocity field divergence-free.

4.6 Vaporization with Physical Mass Flux

To test the phase-transition model with a physical mass flux density m,
calculated from the temperature profile, another simple one-dimensional
case was simulated. This case was chosen partly because of its simplicity,
but mostly because it has an analytical solution for the temperature and
velocity profiles and for the interface location to which the results from
the simulations could be compared.
We consider a one-dimensional fluid domain of length L = 1.0 m, occupied by a liquid (Phase 1) and a gaseous phase (Phase 2) of the same substance. The two phases are separated by an interface located at x = (t),
as shown in Figure 4.12. The specific enthalpy difference associated with
the phase transition is h = 1.0 103 J kg1 . The other relevant fluid
parameters are given in Table 4.9.
The northern, southern and western walls have free-slip BCs and the

85

4 Numerical Experiments

Phase 2

Phase 1

Twest

Tsat

x
L

Figure 4.12: Schematic drawing of the initial fluid configuration in the vaporization
with physical mass flux case. Phase 1 is the liquid and Phase 2 is the gaseous phase.
The east wall is held at the saturation temperature Tsat and the west wall is held at
Twest > Tsat .

eastern wall has an outflow condition. When solving the Poisson equation
for the pressure, we impose Neumann BCs on the pressure at all walls
except the eastern, where we demand p = 0 Pa. As boundary conditions
for temperature, we hold the temperature on the western wall constant,
T (0, t) = Twest = 0.1185 K, and the temperature on the eastern wall constant and equal to the saturation temperature, T (Lx , t) = Tsat = 0 K .
Contrary to the other cases, we do not start the simulation at 0 s, but
rather at t = 0.25 s. The initial interface location is thus (0.25 s) = 0.5 m.
Both the temperature and the velocity profiles were initialized according
to the analytical solution. The simulations were run from t = 0.25 s to
t = 0.6 s.
As in the vaporization case with uniform mass flux, the initial and
boundary conditions made this an effective one-dimensional problem and
the numerical results should be independent of y even though the simula

In spite of having performed most of the work on this thesis behind a computer rather
than in the laboratory, the author is aware that a saturation temperature of 0 K is
rare in real physical systems. This temperature was chosen to have results that
would be comparable to those from Gibou et al. [2007]. Note also that the velocity
and pressure fields do not depend on the zero-level of the temperature, so this could
easily have been chosen differently.

86

4.6 Vaporization with Physical Mass Flux

tions were performed on a two-dimensional grid.

4.6.1 Analytical Solution

The analytical solution to this case was given, in part, by Gibou et al.
[2007]. Here it is presented in some more detail.
As before, we give our solution different names in the two phases,
(
T + (x, t) if x +
T (x, t) =
,
(4.74)
T (x, t) if x
(
u+ (x, t) if x +
u(x, t) =
.
(4.75)
u (x, t) if x
Due to the boundary and interface conditions, we can set the temperature
in the entire liquid phase equal to Tsat at all times,
T (x, t) = Tsat ,

(4.76)

and the velocity in the gaseous phase to zero,

u+ (x, t) = 0.

(4.77)

Thus the advection-diffusion equation for temperature (2.11) from page 8

in the gaseous phase reduces to
t T + = + 2 T + .

(4.78)

It can be confirmed by straightforward differentiation and insertion that

(4.78) has a general solution


x
+

T (x, t) = C1 + C2 erf
,
(4.79)
2 + t
where erf is the Gauss error function, defined in terms of an integral of
the Gauss function,
Z
2
2
erf () =
e d.
(4.80)
0

87

4 Numerical Experiments
To determine the constants C1 and C2 , we consider the boundary conditions. The boundary condition at the western wall, T (0, t) = Twest , gives
C1 = Twest .

(4.81)

Next, we choose our C2 in terms of a new constant C3 , related by

C2 =

Tsat Twest
.
erf (C3 )

Then, the interface condition T ((t), t) = Tsat is satisfied when

(t) = 2C3 + t.

(4.82)

(4.83)

To determine C3 , we demand that the velocity of the interface w = u+

m/+ evaluated at the interface is equal to t (t). Using that u+ = 0 and
(2.42) on page 14, this results in the transcendental equation
2

C3 e(C3 ) erf (C3 ) =

c+
p (Tsat Twest )

.
h

(4.84)

This equation was numerically solved for C3 using the bisection method
[S
uli and Mayers, 2006, Section 1.6]. To find u (x, t) we solve the jump
condition (2.38) from page 14, that is
 
1
+

u u =m
,
(4.85)

for u . Again using that u+ = 0 and (2.42), in addition to (4.83), we get

 
2
1
+

u (x, t) =
C2 e(C3 ) .
(4.86)
h + t

4.6.2 Numerical Results

This case was run on five different grid sizes 10 10, 20 10, 40 10,
60 10 and 80 10. As Gibou et al. [2007], we use the GFM method for
all fluid properties, except for the viscous term where we resort to CSF.
The obtained temperature profiles on the 4010 gird is shown in Figure
4.13 at the initial time t = 0.25 s, the final time t = 0.6 s and at t = 0.425 s.

88

4.6 Vaporization with Physical Mass Flux

0.12

t = 0.25 s
t = 0.425 s
t = 0.6 s

0.10

T [K]

0.08
0.06
0.04
0.02
0.00
0.0

0.2

0.4

0.6

0.8

1.0

x [m]
Figure 4.13: The temperature profiles obtained on the 40 10 grid at different times
in the vaporization with physical mass flux case. The analytical solution at each time
is drawn as a solid line.

At each time the analytical solution is also shown. From inspecting Figure
4.13, we see that the jump in x T at the interface appears to decrease in
time, so we expect the jump in velocity to decrease too, in accordance
with (2.42) and (2.38) from page 14. We can also conclude that there is
agreement between the simulation results and the analytical solution to
plotting accuracy.
The velocity profiles obtained on the same 40 10 grid at the same
points in time are shown in Figure 4.14. Again the analytical solution at
each point in time is shown for comparison. From inspecting this figure, we
see that the velocity jump at the interface does indeed decrease with time,
as expected from the temperature profiles in Figure 4.13. Also, the jump
in velocity appears a little overestimated with respect to the analytical
solution, especially at t = 0.25 s. This is probably due to overestimation
of + x T + and may thus be attributed to the smearing of x T around
the interface that we commented on in Section 3.4.2. Overall, we conclude
that there is good agreement between the simulated and analytical velocity
profiles to plotting accuracy.

89

4 Numerical Experiments

0.8

u [m s1 ]

0.6

t = 0.25 s
t = 0.425 s
t = 0.6 s

0.4

0.2

0.0
0.0

0.2

0.4

0.6

0.8

1.0

x [m]
Figure 4.14: The velocity profiles obtained on the 40 10 grid at different times in the
vaporization with physical mass flux case. The analytical solution at each time is drawn
as a solid line.

The interface location obtained from simulations on a selection of grid

sizes is plotted against time t in Figure 4.15. The corresponding analytical
solution is also shown for comparison. It is clear that (t) is somewhat
overestimated, especially on the coarse 10 10 and 20 10 grids. This
is consistent with overestimation of + x T + . In any event, it is evident
that the simulation results approached the analytical solution upon grid
refinement and that there is agreement, to plotting accuracy, between the
simulated results on the 80 10 grid and the analytical solution.
To quantify the convergence order, a convergence study based on Eabs
in (0.6 s) with respect to the analytical solution 0.7746 m was performed.
The results are given in Table 4.10 and in Figure 4.16. These are consistent
with the first-order convergence rate we would expect from the first-order
GFM treatment of the interface conditions.
From this case, we can conclude that there is good agreement between
the numerical results and the analytical solution to the temperature and
velocity profiles and the interface location. Also, we demonstrated the
expected first-order convergence when refining the spatial grid. This indi-

90

4.6 Vaporization with Physical Mass Flux

0.80

Analytical
8010
2010
1010

0.75

[m]

0.70
0.65
0.60
0.55
0.50
0.25

0.30

0.35

0.40

0.45
t [s]

0.50

0.55

0.60

0.65

Figure 4.15: A plot of the interface location against time t obtained on a selection of
grid sizes in the vaporization with physical mass flux case. The analytical solution is
shown as a solid line.

Table 4.10: Convergence table based on the error Eabs , with respect to the analytical
solution, in the interface location at t = 0.6 s in the vaporization case with physical
mass flux.

Nx
-

(0.6 s)
m

Eabs
m

n
-

10
20
40
60
80

7.803 1001
7.776 1001
7.761 1001
7.756 1001
7.753 1001

5.70 1003
2.99 1003
1.49 1003
1.00 1003
7.49 1004

0.93
1.01
0.97
1.02

91

4 Numerical Experiments
2.2

log10 (Eabs )

2.4
2.6
2.8
3.0
3.2
1.0

1.2

1.4
1.6
log10 (Nx )

1.8

2.0

Figure 4.16: Log-log plot of the error Eabs in the interface location (0.6 s) in the vaporization case with physical mass flux. The solid line represents a first order polynomial
fit and has a slope of approximately 0.98.

cates that the implementation of the phase transition model is correct in

one dimension.

4.7 Vaporization of a Drop

To test the phase-transition model and its ability to preserve the discontinuities in ~u and p at the interface also in two dimensions, a case with a
vaporizing drop was simulated. In contrast to the earlier cases with phase
transition, we now have the possibility of a non-zero interface curvature
and thus surface-tension forces come into play.
In this case we consider a two-dimensional fluid domain with Lx = Ly =
0.08 m. Initially, we have a circular liquid drop (Phase 1) of radius 0.02 m
located in the center of the domain, surrounded by a gaseous phase (Phase
2). The fluids are initialized to be at rest and temperature on the entire
domain is initially set equal to the saturation temperature Tsat = 0 K.
The surface tension coefficient between the two phases is = 0.1 N m1
and there is no gravity. The specific enthalpy difference associated with
the phase change is h = 1.0 103 J kg1 . Table 4.11 contains the rest
of the relevant fluid parameters.

92

4.7 Vaporization of a Drop

Table 4.11: Fluid parameters used in the vaporizing drop case

cp

Unit

Phase 1

Phase 2

kg m3
Pa s
J kg1 K1
W K1 m1

200.0
0.1
400.0
40.0

5.0
0.005
200.0
1.0

To have a supply of thermal energy to drive the vaporization, we set

Dirichlet BCs, T = 10 K, for the temperature on all boundaries. We wish
to allow the gas from the vaporizing drop to escape freely, so all boundaries
carry outflow conditions for the velocity. Dirichlet boundary conditions
were imposed for the pressure at all boundaries.
For this case, we expect that heat will flow from the boundaries into
the fluid domain and increase its temperature. Eventually, the gas temperatures around the drop will be larger than the saturation temperature,
vaporization will be initiated and the drop will gradually shrink. For symmetry reasons and due to surface tension, the drop should keep its circular
shape and position in the center of the domain as it shrinks.
The simulation was run on a 101 101 grid to t = 1.0 s.

4.7.1 Numerical Results

The interface location at t = 0 s and at t = 1.0 s is shown in Figure 4.17a.
As expected, the drop has shrunk and it has kept its circular shape and
its position in the center of the domain.
The velocity field at t = 1.0 s is illustrated in Figure 4.17b. As expected,
the velocities points normally away from the interface at the interface and
they decrease gradually in magnitude as we move away the drop.
We observe also, and perhaps more unexpectedly, that the velocity field
seems to bend towards the constant-x and the constant-y lines that go
through the center of the drop. We might have expected a cylindrically
symmetric velocity field with the magnitude of the velocity depending on
the distance from the interface only. Still, the obtained velocity field is
consistent with the Dirichlet BCs that were set of the pressure, which

93

4 Numerical Experiments
0.08
0.07
0.06

s
0.00

y [m]

0.05
0.04

0
1.0

0.03

0.02
0.01
0.00
0.00

0.01

0.02

0.03

0.04
x [m]

0.05

0.06

0.07

0.08

0.05

0.06

0.07

0.08

(a)
0.08
0.07
0.06

y [m]

0.05
0.04
0.03
0.02
0.01
0.00
0.00

0.01

0.02

0.03

0.04
x [m]

(b)
Figure 4.17: Plots of (a) the initial t = 0 s and final t = 1.0 s interface locations and (b)
the velocity field at t = 1.0 s in the vaporizing drop case.

94

4.7 Vaporization of a Drop

10

6
4

T [K]

2
0
0.08
0.07
0.06
0.05
0.00 0.01
0.04 ]
0.02 0.03
0.03 y [m
0.04
0.02
x [m] 0.05 0.06
0.01
0.07 0.08 0.00

(a)

6
4
3

p [Pa]

2
1
0
0.08
0.07
0.06
0.05
0.00 0.01
0.04 ]
0.02 0.03
0.03 y [m
0.04 0.05
0.02
x [m]
0.01
0.06 0.07
0.00
0.08

(b)
Figure 4.18: Plots of (a) the temperature and (b) the pressure fields at t = 1.0 s in
the vaporizing drop case. The gaseous field values are indicated by green dots and the
liquid values by blue dots.

95

4 Numerical Experiments
do not have cylindrical symmetry. Because of the these BCs, p will
tend to point along the direction normal to closest boundary rather than
radially away from the drop and this will make the fluid tend to flow in
this direction.
The temperature field at t = 1.0 s is plotted in Figure 4.18a. This plot
shows what we would qualitatively expect, a drop with uniform temperature Tsat and gas temperatures that decrease gradually when moving from
the boundary to the interface.
Figure 4.18b shows the pressure at t = 1.0 s. As expected, it exhibits
an interfacial jump that, for this case, can be mostly attributed to the
interfacial tension. We observe that the interfacial pressure jumps at the
points directly north, south, east and west of the drop center are somewhat higher compared to those located north-east, north-west, south-east
and south-west. The reason for this is probably that the magnitude of
temperature gradient in the gas phase is larger around these points resulting in a bigger mass flux. This is natural since the distance to the hot
boundary is smaller for these points.
From this test case we can conclude that the our implementation of the
phase-transition model is able to cope with interfacial jumps also in two
dimensions. We have also seen that we must take care when imposing
Dirichlet BCs on the pressure as they may not always be consistent with
the physical situation we want to model.

4.8 Boiling Film

Finally, we turn our attention to a case with thin film boiling in two
dimensions. In this situation, a thin layer of vapor (Phase 2) covers the
south boundary and the rest of the fluid domain is occupied by a liquid
(Phase 1). The south wall is heated to a temperature above the saturation
temperature of the liquid and this causes the liquid to evaporate at the
interface and increase the volume of the film. At the appropriate wall
heating, rising vapor bubbles should form due to gravity and the density
differences between the liquid and the vapor.
We consider a fluid domain with equal width and height, Lx = Ly =
0.08 m. The fluids are initially at rest and have a uniform temperature of

96

4.8 Boiling Film

Table 4.12: Fluid parameters used in the boiling film case.

cp

Unit

Phase 1

Phase 2

kg m3
Pa s
J kg1 K1
W K1 m1

200.0
0.1
400.0
40.0

5.0
0.005
200.0
1.0

0 K. The interface location is initialized according to the parametrization




Lx
2(x Lx /2)
y(x) =
4 + cos
.
(4.87)
128
Lx
The specific enthalpy difference associated with the phase change in
3
1
this case is h

= 1.0 10 J kg , the gravitational acceleration is ~g1=
1
9.81 m s
y and the coefficient of surface tension is = 0.1 N m .
The rest of the relevant fluid parameters are given in Table 4.12.
We pose no-slip BCs on the south wall, free-slip on the east and west
walls and outflow BCs on the north wall. The north wall has a constant
temperature equal to the saturation temperature of the liquid, Tsat = 0 K,
and the south wall a constant temperature of 0.1 K. The east and west
walls are thermally insulating. We pose Neumann BCs on the pressure on
all walls except the north where we use the Dirichlet BC.
The simulation was run on a 71 71 grid.

4.8.1 Numerical Results

The temporal development of the interface is shown in Figure 4.19. We
see that the film increases in volume from t = 0.3 s to t = 0.4 s as the
liquid at the interface evaporates. The vapor then gathers in the middle
of the domain, forming the beginnings of a bubble. At t = 0.5 s the bubble
has begun to rise. In the final frame, at t = 0.6 s, the bubble has obtained
a mushroom shape as its side lobes are pulled down by the vortices on
either side of the stem. These vortices are illustrated in Figure 4.20.
According to Gibou et al. [2007], the qualitatively correct behavior for
film boiling in two dimensions is, perhaps somewhat counter-intuitively,

97

0.08

0.08

0.07

0.07

0.06

0.06

0.05

0.05
y [m]

y [m]

4 Numerical Experiments

0.04

0.04

0.03

0.03

0.02

0.02

0.01

0.01

0.00
0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
x [m]

0.00
0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
x [m]

(b)

0.08

0.08

0.07

0.07

0.06

0.06

0.05

0.05
y [m]

y [m]

(a)

0.04

0.04

0.03

0.03

0.02

0.02

0.01

0.01

0.00
0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
x [m]

0.00
0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
x [m]

(c)

(d)

Figure 4.19: The interface location in the boiling film case at (a) t = 0.3 s, (b) t = 0.4 s,
(c) t = 0.5 s and (d) t = 0.6 s.

98

4.9 Summary
0.08
0.07
0.06

y [m]

0.05
0.04
0.03
0.02
0.01
0.00
0.00

0.01

0.02

0.03

0.04
x [m]

0.05

0.06

0.07

0.08

Figure 4.20: Illustration of the velocity field at t = 0.6 s in the the boiling film case..
Vortices can be observed on either side of the stem.

that the stem of the rising bubble should not pinch off. In two dimensions, the curvature of the stem is negligible and therefore so are the
surface tension effects and thus the stem persists and continues to fill the
rising bubble with vapor. Gibou et al. [2007] observed this behavior in
their boiling film simulations on fine grids. It was only on coarser grids,
that could not resolve the stem of the rising bubbles, that pinch-off was
observed. This explains why bubble pinch-off was not observed in the
simulations run here.

4.9 Summary
In this chapter, we have presented eight test cases and discussed results
from their simulations. Here, we briefly summarize the conclusions from
these discussions.

99

4 Numerical Experiments
De Vahl Davis Benchmark Case
From the results of the simulations of the de Vahl Davis benchmark case we
concluded that the implementation was able to reproduce the benchmark
results on the 301 301 grid. Also, a convergence study indicated that the
order of convergence was consistent with the expected value of 2. This gave
reason to believe that the current implementation of the discretization
of the advection-diffusion equation for temperature and the Boussinesq
coupling to the NavierStokes equations for single-phase flow cases was
correct.
Two-Phase Heat Conduction
In this case we saw that the CSF discretization of the diffusive term in
the advection-diffusion equation for temperature failed to reproduce the
analytical solution both with Dirichlet and Neumann boundary conditions.
The GFM discretization, however, produced solutions in agreement with
the analytical for both Dirichlet and Neumann boundary conditions. This
test case thus showed the advantage of treating the interface sharply with
the GFM instead of with the CSF method when solving the heat equation.
Wang et al.s Two-Layer Convection Case
From this test case we concluded that our implementation produced results in agreement with the analytical steady-state solution for a case
involving temperature-driven two-phase flow. Also, the performed convergence study showed the expected first-order convergence. This gave
reason to believe that the current implementation of the discretization
of the advection-diffusion equation and the Boussinesq coupling to the
NavierStokes equations was correct also for two-phase flow cases where
the interface is treated with the GFM.
Lava Lamp
Here we saw that our two-phase model reproduced the experimentally observed blob exchange oscillations and their time periods. This gave reason
to believe that the implementation was correct also on axisymmetric grids.
Still, the simulated behavior could not be said to be in full agreement with

100

4.9 Summary
the experimental results and neither could we expect it to be, since many
of the physical aspects of lava lamp convection are not accounted for in
our model.
Vaporization with Uniform Mass Flux
From this test case we could conclude that the current implementation
succeeded in preserving the discontinuities in the velocity and pressure
fields, even when grid nodes changed character from liquid to gas, in one
dimension and in the case of uniform mass flux density. It also succeeded
in propagating an instantaneous change in the velocity field at the interface
to the entire domain instantaneously as is necessary in the incompressible
model.
Vaporization with Physical Mass Flux
Here we observed that there was good agreement between the numerical
results and the analytical solution to the temperature and velocity profiles
and the interface location. Also, we demonstrated the expected first-order
convergence when refining the spatial grid.
Vaporization of a Drop
From this test case we could conclude that the our implementation of the
phase-transition model was able to cope with interfacial jumps also in two
dimensions.
Boiling Film
In the boiling film test case we saw a boiling film that produced a rising
bubble. The results were in qualitative agreement with those from Gibou
et al. [2007].

101

5 Conclusions and Suggestions for

Further Work
5.1 Conclusions
In this thesis, we have presented a two-phase flow model, based on the incompressible NavierStokes equations. We have also presented the numerical methods used in solving this model, with emphasis on non-standard
methods such as the level-set method for tracking the interface, the ghostfluid method for treating the interface sharply and the weighted essentially
non-oscillatory scheme for discretizing advective terms in the governing
equations.
The main contributions of this work have been to extend the implementation of the flow model through the discretization and implementation
of the advection-diffusion equation for temperature and a Boussinesq coupling between the temperature and velocity fields. In two-phase flow, both
the continuum surface force method and the ghost-fluid method were employed for handling jumps at the interface. Results from simulated cases
indicated that implementation for both single- and two-phase flow with
ghost-fluid method was correct, with second- and first-order convergence,
respectively. Also, a model for phase transition has been presented, discussed and implemented. This implementation allowed for vaporization
and condensation mass transport between the phases. Results from simulated one-dimensional cases with phase transition indicated that the implementation of this model was correct, with first-order convergence. The
results from one dimension and the qualitatively correct results from two
dimensions gave reason to believe that the implementation was correct
also in two dimensions.
Through the introduction of heat-transport physics in the implementation of the two-phase flow model, it has been developed in direction of
performing more detailed simulations that are relevant for two-phase heat

103

5 Conclusions and Suggestions for Further Work

exchangers and hence for the natural gas liquefaction processes.

5.2 Suggestions for Further Work

As the results from the two-phase heat conduction case suggested, the
continuum surface force (CSF) method did not perform very well at discretizing the diffusion term in the advection-diffusion equation for temperature. Still, we rely on this method when treating the viscous term in
the NavierStokes equations when simulating cases with phase transition.
Therefore, an effort could be made to assess the effect of the CSF treatment of viscosity, or if one can implement a ghost-fluid-method (GFM)
treatment, for this type of simulations.
The chosen discretization of the diffusive term in the advection-diffusion
equation with phase transition led to a strict time-step restriction. An implicit treatment of the advection-diffusion equation could be attempted to
avoid this restriction and run stable simulations with a smaller cell division threshold c and thus longer time steps. Another benefit of reducing
c is that we might get better-quality approximations to the temperature
gradient at the interface.
One could also look at a way of posing boundary conditions on the levelset function, e.g. using equation (61.12) from Landau and Lifshitz [1987],
to get a realistic contact angle between the interface and the domain walls.
This might result in more accurate simulations in situations like the lava
lamp case.
In this thesis, we have put more emphasis on verifying the implementation of the model (testing if the numerical method and its implementation
solves the model equations) and less on validating the correctness of the
model with respect to the real world (testing if the model accurately predicts experimental results). To this end, more detailed comparisons of
simulated results with experiments should be performed.

104

Bibliography
T. D. Aslam. A partial differential equation approach to multidimensional
extrapolation. Journal of Computational Physics, 193:349355, 2003.
S. Balay, J. Brown, K. Buschelman, W. D. Gropp, D. Kaushik, M. G.
Knepley, L. C. McInnes, B. F. Smith, and H. Zhang. PETSc web page,
June 2013. URL http://www.mcs.anl.gov/petsc.
G. de Vahl Davis. Natural convection of air in a square cavity: A bench
mark numerical simulation. International Journal for Numerical Methods in Fluids, 3(3):249264, 1983.
V. K. Dhir. Numerical simulations of pool-boiling heat transfer. AIChE
Journal, 47(4):813834, 2001.
F. Gibou, R. Fedkiw, L. Cheng, and M. Kang. A second-order-accurate
symmetric discretization of the Poisson equation on irregular domains.
Journal of Computational Physics, 176:205227, 2002.
F. Gibou, L. Chen, D. Nguyen, and S. Banerjee. A level set based sharp
interface method for the multiphase incompressible NavierStokes equations with phase change. Journal of Computational Physics, 222:536
555, 2007.
M. Gisvold. Eventyret p
a Melkya. Gemini Forskingsnytt fra NTNU og
SINTEF, June 2004.
M. Griebel, T. Dornseifer, and T. Neunhoeffer. Numerical Simulation in
Fluid Dynamics. SIAM, 1998.
B. Gy
ure and I. M. J
anosi. Basics of lava lamp convection. Physical
Review E, 80(4):046307, 2009.

105

Bibliography
E. B. Hansen. Numerical Simulation of Droplet Dynamics in the Presence
of an Electric Field. PhD thesis, Norwegian University of Science and
Technology, November 2005.
G. S. Jiang and C. W. Shu. Efficient implementation of weighted ENO
schemes. Journal of Computational Physics, 126(1):202228, 1996.
M. Kang, R. Fedkiw, and X. Liu. A boundary condition capturing method
for multiphase incompressible flow. Journal of Scientific Computing, 15:
323360, 2000.
D. I. Ketcheson and A. C. Robinson. On the practical importance of
the SSP property for RungeKutta time integrators for some common
Godunov-type schemes. International Journal for Numerical Methods
in Fluids, 48(3):271303, 2005.
E. Kreyszig. Advanced Engineering Mathematics. Wiley, 9th edition, 2006.
T. Kunugi. Brief review of latest direct numerical simulation on pool and
film boiling. Nuclear Engineering and Technology, 44(8):847854, 2012.
L. D. Landau and E. M. Lifshitz. Fluid Mechanics, volume 6 of A Course
on Theoretical Physics. Elsevier, 2nd edition, 1987.
K. Y. Lerv
ag. Simulation of two-phase flow with varying surface tension.
Masters thesis, The Norwegian University of Science and Technology,
June 2008.
S.

W. Lvseth.
Enabling low-emission LNG systems web
page, June 2013.
URL http://www.sintef.no/Projectweb/
Low-Emission-LNG-Systems.

Norwegian Petroleum Directorate. Factpages, May 2013. URL http:

//factpages.ndp.no.
S. Osher and R. Fedkiw. Level Set Methods and Dynamic Implicit Surfaces.
Springer, 2003.
S. Osher and J. A. Sethian. Fronts propagating with curvature-dependent
speed: Algorithms based on HamiltonJacobi formulations. Journal of
Computational Physics, 79:1249, 1988.

106

Bibliography
F. S. Oueslati, R. Bennacer, H. Sammouda, and M. El Ganaoui. Analytical and numerical solutions for natural convection in a shallow cavity
filled with two immiscible fluids: Shear stress action. Numerical Heat
Transfer; Part A: Applications, 62(8):605623, 2012.
K. Rottmann. Matematisk Formelsamling. Spektrum forlag, 2003.
R. Scardovelli and S. Zaleski. Direct numerical simulation of free-surface
and interfacial flow. Annual Review of Fluid Mechanics, 31:567603,
1999.
M. Sussmann, P. Smereka, and S. Osher. A level set approach for computing solutions to incompressible two-phase flow. Journal of Computational Physics, 114:149159, 1994.
E. S
uli and D. Mayers. An Introduction to Numerical Analysis. Cambridge
University Press, 2006.
I. Tamura, T. Tanaka, T. Kagajo, S. Kuwabara, T. Yoshioka, T. Nagata,
K. Kurahashi, and H. Ishitani. Life cycle CO2 analysis of LNG and city
gas. Applied Energy, 68(3):301319, 2001.
S. Tanguy, T. Menard, and A. Berlemont. A level set method for vaporizing two-phase flows. Journal of Computational Physics, 221(2):837853,
2007.
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. Journal of Computational
Physics, 169(2):708759, 2001.
C. H. Wang, M. Sen, and P. Vasseur. Analytical investigation of Benard
Marangoni convection heat transfer in a shallow cavity filled with two
immiscible fluids. Applied Scientific Research, 48(1):3553, 1991.
F. M. White. Fluid Mechanics. McGrawHill, sixth edition, 2008.
H. K. Zhao, T. Chan, B. Merriman, and S. Osher. A variational level set
approach to multiphase motion. Journal of Computational Physics, 127
(1):179195, 1996.

107