Analysis of Kinetic Models and Macroscopic Continuum

Equations for Rarefied Gas Dynamics

by
Yingsong Zheng
B. Eng., University of Science and Technology of China, 1996
M. Eng., University of Science and Technology of China, 1999

A Dissertation Submitted in Partial Fulfillment of the Requirements for the Degree of

DOCTOR OF PHILOSOPHY
in the
Department of Mechanical Engineering

We accept this dissertation as conforming

to the required standard

OYINGSONG ZHENG, 2004

University of Victoria

All rights reserved. This thesis may not be reproduced in whole or in part, by photocopy
or other means, without the express written permission of the author
Supervisor: Dr. Henning Struchtrup

Abstract

The Boltzmann equation is the basic equation to describe rarefied gas flows. Some
kinetic models with simple expressions for the collision term have been proposed to
reduce the mathematical complexity of the Boltzmann equation. All macroscopic
continuum equations can be derived from the Boltzmann equation or kinetic models
through the Chapman-Enskog method, Grad's moment method, etc.

This thesis is divided into three parts. In the first part, existing kinetic models (BGK
model, ES-BGK model, v(C) -BGK model, S model, and Liu model), and two newly
proposed v(C)-ES-BGK type kinetic models are described and compared, based on
properties that need to be satisfied for a kinetic model. In the new models a meaningful
expression for the collision frequency is used, while the important properties for a kinetic
model are retained at the same time.

In the second part of this work, the kinetic models (BGK, ES-BGK, v(C) -BGK, and
two new kinetic models) are tested numerically for one-dimensional shock waves and
one-dimensional Couette flow. The numerical scheme used here is based on Mieussens's
discrete velocity model (DVM). Computational results from the kinetic models are
compared to results obtained from the Direct Simulation Monte Carlo method (DSMC).
It is found that for hard sphere molecules the results obtained from the two new kinetic
models are very similar, and located in between the results from the ES-BGK and the
v(C)-BGK models, while for Maxwell molecules the two new kinetic models are
identical to the ES-BGK model. For one-dimensional shock waves, results from the new
kinetic model II fit best with results from DSMC; while for one-dimensional Couette
flow, the ES-BGK model is suggested.

Also in the second part of the work, a modified numerical scheme is developed from
Mieussens's original DVM. The basic idea is to use a linearized expression of the
 ...
111

reference distribution function, instead of its exact expression, in the numerical scheme.
Results from the modified scheme are very similar to the results from the original scheme
for almost all done tests, while 20-40 percent of the computational time can be saved.

In the third part, several sets of macroscopic continuum equations are examined for
one-dimensional steady state Couette flow. For not too large Knudsen numbers
(Knc=O.l) in the transition regime, it is found that the original and slightly linearized
regularized 13 moment equations give better results than Grad's original 13 moment
equations, which, however, give better results than the Burnett equations, while the
Navier-Stokes-Fourier equations give the worst results, which is in agreement with the
expectation. For large Knudsen number situations (Kn>O.l), it turns out that all
macroscopic continuum equations tested fail in the accurate description of flows, while
the Grad's 13 moment equations can still give better results than the Burnett equations.
Examiners:
Table of contents

Abstract
Table of contents v
Nomenclature viii
List of Tables xiii
List of Figures xiv
Acknowledgements xvi
1 Introduction 1
1.1 Background 1
1.2 Motivation and objective 7
2 Kinetic models 9
2.1 Introduction 9
2.2 Motivation of kinetic models and expression of collision frequency 10
2.3 Existing kinetic models 11
2.4 Two new v ( C )-ES-BGK type kinetic models 17
2.4.1 New kinetic model I 17
2.4.2 New kinetic model I1 23
2.5 Conclusion 27
3 Numerical work on one-dimensional shock waves 28
3.1 Introduction of shock wave 28
3.2 Numerical scheme 29
3.2.1 Introduction of explicit scheme 29
3.2.2 Collision frequency 31
3.2.3 Reference distribution f,(,; (original) 33

3.2.4 Reference distribution f r:f,i, (linearized)

3.2.5 Boundary conditions and initial guess
3.2.6 Time step, space grid and velocity grid
3.3 Test examples
3.4 Results and discussion
 3.4.1 Some important notes on dealing with results
3.4.2 Convergence of results
3.4.3 Discussion
3.4.4 Comparison with DSMC
3.5 Conclusion
4 Numerical work on one-dimensional Couette flow
4.1 Introduction of Couette flow
4.2 Numerical scheme
4.2.1 Introduction of explicit scheme
4.2.2 Reference distribution fr>,i,j (original)

4.2.3 Reference distribution f r>,i,j (linearized) 79

4.2.4 Boundary conditions and initial guess 82

4.3 Test examples 84
4.4 Results and discussion 86
4.4.1 Some important notes on dealing with results 86
4.4.2 Convergence of results 87
4.4.3 Discussion 95
4.4.4 Comparison with DSMC 110
4.5 Conclusion 132
5 Comparison of NSF, Burnett, Grad13 and R13s equations to kinetic models 134
5.1 Introduction 134
5.2 NSF and Burnett equations for BGK and ES-BGK models 135
5.2.1 NSF and Burnett equations in three dimensions 135
5.2.2 Linear stability analysis of the Burnett equations for the ES-BGK model
146
5.2.3 NSF and Burnett equations for one-dimensional Couette flow at steady state
150
5.3 Grad13 and R13s equations for BGK and ES-BGK models 152
5.3.1 Grad13 and R 13s equations in three dimensions 152
5.3.2 Grad 13 and R 13s equations for one-dimensional Couette flow at steady state
157
 vii
5.4 The Knudsen layer 162
5.5 Description about the test method 163
5.6 Test examples 163
5.7 Results and discussion 164
5.7.1 Some important notes on dealing with data 164
5.7.2 Results and discussion of oiiand qi 169

5.7.3 Results and discussion of third-degree and fourth-degree moments 180

5.8 Conclusion 190
6 Conclusions and outlook 192
6.1 Conclusions 192
6.2 Outlook 195
References 197
Appendix A: How diagrams of N-R algorithm and computational programs 204
Appendix B: Expressions of functions and Jacobians in the N-R algorithm for all kinetic
models 207
Appendix C: Linearized dimensionless reference distribution Fr>,i,, 223

Appendix D: Average relative errors in comparisons 236

Appendix E: Moments computed from the reference distribution f,, in the ES-BGK
model 245
 viii
Nomenclature

Symbols
a Coefficient in reference distribution function
A A function used in new kinetic models in Sections 2.3.1 and 3.2.2
b A number used in the ES-BGK model and new kinetic models to adjust the
Prandtl number
B A function used in new kinetic models in Sections 2.3.1 and 3.2.2
BB 1, BB2.. . Some function used in Appendix B and C
Microscopic particle velocity
Peculiar velocity
Mass distribution function
Numerical flux (only in Sections 3.2.1 and 4.2.1), or dimensionless reference
distribution
Relative speed of the colliding particles
Number of total position nodes
Unit matrix
Number of discrete velocities
Jacobian
Boltzmann constant, or positive real wave number ( only in Section 5.2.2)
Knudsen number
Mean free path
Domain width or Characteristic geometry length
Mass of one microscopic particle, or a parameter defined as muk= p<uk,
Order of potential
Avogadro's number
Pressure
Prandtl number
Heat flux
Distance between two particles
R Gas constant, or a parameter defined as Ru = p,,,, - 7 R T a ,
S Collision term

g,,,, &, intermediate expression for collision term in Section 2.2

t Time
T Temperature
u Macroscopic flow velocity
x-y-z Cartesian coordinates
X 1-X2-X3 Cartesian coordinates

Greek
Coefficient in reference distribution function
Coefficient in reference distribution function
Delta function

Parameter used in moment equations, defined as A = p , - 15- p Z

P
Difference of a parameter a
time step
Parameter used to determine Prandtl number in new kinetic models
Angle of collision (only in Eq. (1.2)), or a scaling parameter (in Section 5.3.1)
Inverse matrix of R in the ES-BGK model and the new kinetic model I
Dimensionless peculiar velocity
Angle of collision
Thermal conductivity
A complex number in the expression of plane wave in Section 5.2.2 (A = RR + Ri)
A matrix used in the ES-BGK model and the new kinetic models
Viscosity
Collision frequency
Po(n) A constant used in the expression of collision frequency
p Mass density, or with some subscript to mean high degree moment
pe Density of internal energy
a Scattering factor (only in Eq. (1.2)), or trace-free pressure
z Collision time
@ Power interaction potentials between particles
ly Flux limiter function
p Some macroscopic parameter
@ A function used in the Burnett equations
x Production of entropy
Y Polynomial of peculiar velocity
w Power index in the power law of viscosity as a function of temperature
ny1(2) A coefficient in the Liu model
SZ A function defined in Eq. (5.29)
J A coefficient in Eq. (5.31)
Y A vector used in Appendix E
M Inverse matrix of II (used in Appendix E)
E Spectral matrix of E (used in Appendix E)
II Modal matrix of matrix E (used in Appendix E)

Subscript
BGK BGKmodel
D Downstream side of shock wave
E Equilibrium
ES ES-BGK model
i Index of position node, or index of coordinates (only used in ui , p, of,p, , p,),
R, and m, )

ini Initial state

j Index of discrete velocity, or index of coordinates (only used in p, o, , p, , p, ,

R, and m,,)

j ,j ,j Index of discrete velocity at x, y, or z direction

 xi
k Iteration number (only in the N-R algorithm in the Appendix B), or index of
coordinates
L Liu model
m Kinetic model
M Maxwellian
Min Minimum value
Max Maximum value
NI New kinetic model I
NII New kinetic model 11
P Plates in Couette flow
P1 Plate one in Couette flow
ref Reference
S S model
U Upstream side of shock wave
&Y,Z One set index of coordinates
Y V ( C )-BGK model
0 Values at some reference state
<> Trace-free

Superscript
B Burnett equations
G13 Grad 13
K Kinetic model
M Moment method
n Index of time step
(n) nth order term in the C-E expansion
NSF Navier-Stokes-Fourier equations
R13I original R13
R 13II slightly linearized R 13
1 Particle 1
f
State after collision
 xii
-
Coefficient small compared to 1.0
A Dimensionless value
- Functions computed in the N-R algorithm, or amplitude of the plane wave (only
in Section 5.2.2)

Acronyms
C-E Chapman - Enskog
DSMC Direct simulation Monte Carlo method
DVM Discrete velocity method
G13 Grad 13 moment equations
MM Maxwellian molecules
N-R Newton - Raphson algorithm
NSF Navier-Stokes-Fourier equations
lDSW One-dimensional shock wave problem
lDCF One-dimensional Couette flow problem
R13 Regularized 13 moment equations
R13s Two types (original and slightly linearized) regularized 13 moment
equations
Hard sphere molecules
 xiii
List of Tables

Table 2.1 : Comparison of kinetic models 16

Table 2.2: Several n and corresponding b values for the new kinetic models 21
Table 3.1: Situations of numerical tests of kinetic models and DSMC for lDSW 41
Table 3.2: Parameters in the numerical tests of kinetic models for lDSW 42
Table 3.3: Some common parameters in the numerical tests of DSMC for 1DSW 42
Table 3.4: Total time step (seconds) for kinetic models in the computatiolr of cases 3.1-
3.3 49
Table 4.1 : Situations of numerical tests of kinetic models and DSMC for lDCF 85
Table 4.2: Some common parameters in the numerical tests of DSMC for lDCF 85
Table 4.3: Parameters in the numerical tests of kinetic models for lDCF 86
Table 4.4: Total time step (unit: seconds) for kinetic models in the computation of cases
4.1-4.12 88
Table 5.1: Burnett coefficients lZJl to D6 and 0, to e5 for different power index n and

Table 5.2: Situations of numerical tests for lDCF in Chapter 5 164

Table 5.3: Parameters in the numerical tests of kinetic models for lDCF in Chapter 5
163
Table D. 1: Average relative errors of kinetic models compared to DMSC in lDSW 236
Table D.2: Average relative errors of kinetic models compared to DMSC in 1DCF. Part I
237
Table D.3: Average relative errors of kinetic models compared to DMSC in 1DCF. Part I1
238
Table D.4: Average relative errors of kinetic models compared to DMSC in 1DCF. Part
III 239
Table D.5: Average relative errors of macroscopic continuum equations compared to
kinetic models. Part I 240
Table D.6: Average relative errors of macroscopic continuum equations compared to
kinetic models. Part 11 24 1
 xiv
Table D.7: Average relative errors of third-degree and fourth-degree moments of moment
equations compared to kinetic models. Part I 242
Table D.8: Average relative errors of third-degree and fourth-degree moments of moment
equations compared to kinetic models. Part 11 243
Table D.9: Average relative errors of third-degree and fourth degree moments of moment
equations compared to kinetic models. Part 111 244
 xv
List of Figures

Figure 3.1: Density profiles at five certain iterations (BGK model, case3.3: Ma=6.0) 46
Figure 3.2: Density profiles from BGK and ES-BGK models (case3.2: Ma=3.0) 47
Figure 3.3: Results to show the use of linearized fr@ and the effect of step of space grid
(Ma=1.5, ES-BGK model) 54
Figure 3.4: Results to show the use of linearized fr@ and the effect of step of space grid

(Ma=3.0, V(C)-BGK model) 55

Figure 3.5: Results to show the use of linearized fr@, and the effect of step of space grid
(Ma=6.0, new kinetic model 11) 56
Figure 3.6: Results to show the effect of bounds and step of velocity grid (Ma=1.5, new
kinetic model I) 57
Figure 3.7: Comparison of kinetic models with DMSC (case 3.4: Ma=1.5) 59-63
Figure 3.8: Comparison of kinetic models with DMSC (case 3.5: Ma=3.0) 64-68
Figure 3.9: Comparison of kinetic models with DMSC (case 3.6: Ma=6.0) 69-73
Figure 4.1: Profiles of some parameters at five certain iterations (BGK model, case 4.18:
Kn=l .O, hard sphere molecules, 300.0 m/s plate velocity) 89-94
Figure 4.2: Results to show the use of linearized fr@ and the effect of step of velocity
grid (new kinetic model 11, situation Sa: Kn=0.025, hard sphere molecules,
300.0 m/s plate velocity) 100-104
Figure 4.3: Results to show the effect of bounds of velocity grid and the effect of step of
space grid (ES-BGK model, situation Sc: Kn=O.l, hard sphere molecules,
300.0 m/s plate velocity) 105-109
Figure 4.4: Comparison of kinetic models with DSMC (case 4.14: Kn=0.025, Maxwell
molecules, 300.0 m/s plate velocity) 112-116
Figure 4.5: Comparison of kinetic models with DSMC (case 4.6: Kn=0.5, Maxwell
molecules, 300.0 m/s plate velocity) 117-121
Figure 4.6: Comparison of kinetic models with DSMC (case 4.9: Kn=0.5, hard sphere
molecules, 600.0 m/s plate velocity) 122-126
 xvi
Figure 4.7: Comparison of kinetic models with DSMC (case 4.12: Kn=0.5, Maxwell
molecules, 1000.0 m/s plate velocity) 127-131
Figure 5.1: Comparison of original and computed trace-free pressure and heat flux (BGK
model, case 5.2: KnS. 1,300.0 m/s plate velocity) 167-168
Figure 5.2: Results of trace-free pressure and heat flux from the BGK model and its
corresponding sets of macroscopic equations (case 5.1: Kn=O.O25, 300.0 m/s
plate velocity) 172-173
Figure 5.3: Results of trace-free pressure and heat flux from the BGK model and its
corresponding sets of macroscopic equations (case 5.7: Kn=O.l, 600.0 m/s
plate velocity) 174-175
Figure 5.4: Results of trace-free pressure and heat flux from the ES-BGK model and its
corresponding sets of macroscopic equations (case 5.6: Kn=0.5, 1000.0 m/s
plate velocity) 176-177
Figure 5.5: Results of trace-free pressure and heat flux from the ES-BGK model and its
corresponding sets of macroscopic equations (case 5.4: Kn=l.O, 300.0 m/s
plate velocity) 178-179
Figure 5.6: Results of third-degree and fourth-degree moments from the BGK model and
its corresponding sets of moment equations (Case 5.1: Kn=0.025, 300.0 m/s
plate velocity) 181-185
Figure 5.7: Results of third-degree and fourth-degree moments from the ES-BGK model
and its corresponding sets of moment equations (Case 5.6: Kn=0.5, 1000.0
m/s plate velocity) 185-190
Figure A. 1: Flow diagram of N-R algorithm used for reference distribution 204
Figure A.2: Flow diagram of program for one-dimensional shock waves 205
Figure A.3: Flow diagram of program for one-dimensional Couette flow 206
 xvii .

Acknowledgements

First of all, I would like to thank my advisor, Dr. Henning Struchtrup, for his support,
which comes in many ways. His enthusiasm and real excitement with all aspects of
applied science is infectious, which made me imagine trying things that previously I
thought myself incapable of doing. The research study has been fascinating.

I would also like to thank everyone who provided valuable assistance by sharing his or
her own expertise. In particular, I wish to thank: Dr. Luc Mieussens for lots of helpful
discussion about his discrete velocity model; Dr. Yongcai Liu for many helpful
discussions, especially about numerical calculations; Toby Thatcher for discussion about
Mathematica; Dr. Sadik Dost for helping during the topic change of my Ph.D study;
Adam Schuetze for providing the DSMC computations. Other people I would like to
thank include, but absolutely not only, Ozan Peksoy, Maurice Bond, Mehmet Yildiz,
MacMurray Whale (previous advisor), Ray Brougham, Rong Zheng.

Finally, I am grateful to my family and friends, especially my wife Yingzi Zhang, for
their support and encouragements.
Chapter 1 Introduction

1.1 Background
The subject of rarefied gas dynamics can be conveniently defined as the study of gas
flows in which the average value of the distance between two subsequent collisions of a
molecule (the so-called mean free path) is not negligible in comparison with a typical
length of the flow structure being considered. This situation is encountered in the areas of
aerodynamics, environmental problems, aerosol reactors, micromachines, vacuum
industry, etc [I].

In the microscopic theory' of rarefied gas dynamics, the state variable is the
distribution function. In this thesis the term "distribution function" always refers to the
mass distribution function, which is the mass of one microscopic particle m times the
number distribution function (or phase density) that is often used in kinetic theory. The
distribution function f (x,c,t) specifies the density of microscopic particles with velocity
c at time t and position x. The particles, which can be thought of as idealized atoms,
move freely in space unless they undergo collisions. The corresponding evolution off is
described by the Boltzmann equation [l, 3-51, which, when external forces are omitted, is
written as

Here, the first term on the left hand side describes the local change off with time, the
second term on the left hand side is the convective change off, and the term on the right
hand side describes the change off due to collisions among particles.

For the Boltzmann equation, the expression of the collision term S(f) is written as
[GI
S(f) = j(f'f"-fi')crgsinOdOd~dcl, (1.2)

' In some references (e.8. [2]), it is said that the Boltzmann equation describes the gas on a mesoscopic
level, instead of a microscopic level.
 2
where the superscript 1 denotes parameters for particle I, which is the collision partner of
the particle considered, the superscript ' denotes parameters for the state after collision,
k
g = -cll is the relative speed of the colliding particles, o is the scattering factor, and

E and 8 are the angles of collision.

The Boltzmann equation is a nonlinear integro-differential equation, and difficult to

handle. Therefore, some alternative, simpler expressions have been proposed to replace
the Boltzmann collision term. These are known as collision models, and any Boltzmann-
like equation where the Boltzmann collision integral is replaced by a collision model is
called a model equation or a kinetic model [3, 41. Kinetic models are discussed in detail
in Chapter 2.

In macroscopic continuum theory of rarefied gas dynamics, the state of gas is

described by macroscopic parameters (or say moments), such as mass density p ,
macroscopic flow velocity u, temperature T, and so on, which depend on position x and
time t. These quantities can be recovered from the distribution f by taking velocity
averages of the corresponding microscopic parameters such as

1
puk= f C~C,C,~C,p, = If c 2 c i c j d c ,
where pe is the density of internal energy, R = is the gas constant, k is Boltmann's

constant, C = e -u is the peculiar velocity (therefore, g = le - c1 = 1 IC - C' ( , dc = dC ), p

is the hydrostatic pressure, pg is the pressure tensor, ouis the trace-free part of pressure
tensor (bracket in subscript denotes the symmetric and trace-free part of a tensor, such as,
1
C,C, = CiCj --C2dU), pi is the heat flux, pUkis a third-degree moment, and p, is a
3
 3
fourth-degree moment. Eq. (t.3.c) gives the definition of temperature from the ideal gas
law.

A key non-dimensional parameter to describe rarefaction effects in rarefied gas

dynamics is the local Knudsen number Kn = x, which is defined as the ratio of the

local mean free path of particles I over a characteristic geometric length L or a length
over which a large variation of a macroscopic quantity may take place [7, 81. The usual
continuum equations (such as the Navier-Stokes-Fourier equations) are only applicable
for Kn<<1.

c
Multiplying the Boltzmann equation (Eqs. (1.1-1.2)) successively by 1, ci , and -,
2
then integrating over c, and utilizing the conservation laws of mass, momentum and
energy

yields the conservation laws for mass, momentum and energy at the macroscopic level,

Note that this set of equations, which should be satisfied by any set of macroscopic
continuum equations, is not closed, unless additional equations for o O ,and q i are given.

The traditional way to obtain expressions of oijand qi from kinetic theory is the
Chapman-Enskog (C-E) method [3, 5, 91, which is a perturbation method that seeks the
distribution function f in the form of an expression in powers of the small parameter Kn,

In this method, a factor ,$& is inserted in front of the collision term, that is to say, the

Boltzmann equation (or a kinetic model equation) to be considered is

(which is the same as the Boltzmann equation (or a kinetic model equation) in
dimensionless form). substituting the expression off in Eq. (1.6) into Eqs. (1.7, 1.2), and
equating the coefficients of the same power of Kn yields a set of equations for f '"I. By
solving this set of equations, an asymptotic solution of the distribution function f is
obtained (note that Kn=l is set in the ultimate expression).

The distribution obtained from the zeroth order expansion is the Maxwellian
distribution f, ,

The corresponding set of macroscopic equations are the Euler equations where
Oii = qi =o.

The first order expansion yields the Navier-Stokes-Fourier (NSF) equations with

where ,u is the viscosity, and K is the thermal conductivity.

Burnett and super-Burnett equations are obtained from the second and third order
expansion respectively [lo-131. Both sets of equations suffer from instabilities and
 5
sometimes give unphysical behavior in steady state processes. Attempts to improve these
equations mostly are non-satisfactory [lo, 14, 151.

The moment method is another traditional way to obtain macroscopic continuum

theory from kinetic theory [5,6,9, 16, 171. In this method, the state of gas is described by
a set of moments, which include but will not be limited to p , ui (or pui ), T , pg and qi .
Multiplication of the Boltzmann equation (or a kinetic model equation) by a set of N
polynomials of the peculiar velocity YA (A=l,. .., N) and subsequent integration over
velocity space yields the N moment equations to determine the N moments
pA= f y A
fdc. These equations do not form a closed system for the N moments being

considered, since some higher degree moments and production terms (moments of the
collision term) are contained in the equations. A closure assumption is required that
allows to relate the additional moments and the production terms to the variables, which
can be achieved if the distribution function f defined in the closure assumption only
depends on those N moments. There are two basic ways to find the distribution function
for the closure, Grad's method [9, 161, and Levermore's method [17- 19).

There are two major points of criticism against Grad's method. One is that Grad's
systems fail to describe smooth shock structures for Mach numbers above a critical value
(such as 1.65 for the 13 moment set) [20]. Another one is that it is very difficult to
develop criteria for the choice of moments that must be considered, for the equations are
not related a priori to the Kn number as a smallness parameter [15].

For the Levermore moment method, the advantage is that every system derived is
always hyperbolic (not like Grad's systems, which might loss their hyperbolic property
depending on the values of the fields, such as heat flux, stress, etc. 1211) and possess a
locally dissipated entropy [22]. The disadvantage is that the systems derived generally
have non-convex domains of definition, and the equilibrium stated is typically located on
the boundary of the defined domain with singular fluxes [23].
 6
The C-E method and the moment method are completely independent of each other,
since they were derived from different premises. However, from a method of Maxwellian
iteration, which is essentially equivalent to a C-E expansion of the moment equations, the
NSF equations and the Burnett equations can be derived from certain sets of Grad's
moment equations [14,24,25].

Recently, a new set of moment equations, the Regularized 13 moment (R13)

equations, has been developed in [26], based on a regularization of Grad's 13 moment
equations. It has been shown that the equations are linearly stable for all wavelengths and
frequencies, lead to smooth shock structures for all Mach numbers [26, 271, and the
description of Knudsen boundary layers from these equations are better than the
description from the Burnett equations [lo].

A similar set of equations, which is called the slightly linearized R13 equations, has
been developed from a new method [15]. This new method is based on accounting orders
in the Kn number, and provides a direct link between Grad type equations and the Kn
number. All sets of equations derived from this method, the Euler equations and the NSF
equations in zeroth and first order, the slightly linearized Grad 13 equations in second
order, the slightly linearized R13 equations in third order, and so on, are stable for
disturbances of all wavelengths and frequencies. Therefore, we can say that this new
method provides a common umbrella for sets of equations that up to now were thought to
stem from very different arguments [15].

The main aim of the Direct Simulation of Monte Carlo (DSMC) method is to calculate
practical flows through the use of the collision mechanics (considered on a probabilistic
rather than a deterministic basis, which cause the noise phenomenon in results of DSMC)
of model molecules, which number is some hundred of thousand or million, but still is
extremely small in comparison with the number of molecules that would be present in the
real gas flow in most applications. For each of simulated molecules, the space
coordinates and velocity components are stored in memory and are modified with time as
the molecules are simultaneously followed through representative collisions and
 7
boundary interactions in the simulated region of space. The calculation is unsteady and
the steady solutions are obtained as asymptotic limits of unsteady solutions. At or near
continuum gas flow (low Knudsen number) situations, the DSMC calculation is very
complicated and time consuming [I].

1.2 Motivation and objective

Some kinetic models with simple expressions for the collision term (such as BGK
model [3,28], ES-BGK model [ l l , 29-33], v(C)-BGK model [34,35], S model [31,36,
371, and Liu model [38-401) have been proposed to reduce the mathematical complexity
of the Boltzmann equation. Up to now, physical meaningful expressions for the collision
frequency as function of microscopic velocity have not been applied in any existing
kinetic model. One aim of this work is to propose and examine some new kinetic models,
in which a meaningful expression for collision frequency is used, while retaining the
important properties for a kinetic model, such as transport coefficients and Prandtl
number, at the same time.

In order to compare kinetic models and to see which one can be suggested to use for
future work, numerical tests need be done, and these form the second part of this work.
Numerical tests are performed for one-dimensional shock waves (IDSW) and one-
dimensional Couette flow (1DCF). The numerical scheme used is Mieussens's Discrete
Velocity Model @VM) [34, 41-44] with some necessary modifications. The main
advantage of Mieussens's DVM is that the distribution function remains always positive,
and that conservation laws and dissipation of entropy are ensured in the discrete
equations. Another advantage is that computational time is saved, since the number of
discrete velocities does not need to be too large. These features are achieved since the
reference distribution function frq is not discretized directly, but determined by the
minimum entropy principle (therefore, the values of coefficients in the discrete reference
distribution are not identical to those in the continuous reference distribution). Further, in
order to save computational time and reduce the complexity of the program, a modified
numerical scheme is developed. The basic idea is to use a linearized expression of the
reference distribution, instead of its exact expression, in the numerical scheme. For
 8
comparison, results from the Direct Simulation Monte Carlo (DSMC) calculations are
used, since the DSMC method is equivalent to solving the Boltzmann equation for a
monatomic gas undergoing binary collisions. Indeed, both can be derived from the
Liouville equation [3, 6,45,46], and the classical Bird's DSMC software is available in
the public domain [47,48].

In the scope of macroscopic continuum theory of rarefied gas dynamics, there are the
NSF equations and the Burnett equations from the C-E expansion method, as well as the
Grad13 equations, the original R13 equations, and the slightly linearized R 13 equations,
which are obtained by the moment method or related methods. In the third part of this
work, the above sets of macroscopic equations derived from kinetic models (BGK and
ES-BGK) will be compared with the kinetic models themselves. The idea is that a set of
macroscopic equations can be considered to be better than another one, when its results
are closer to the results from kinetic models. Numerical tests are performed for one
dimensional Couette flow at steady state, and the Maxwell molecules.

Since boundary conditions for the extended sets of macroscopic equations (Burnett,
G13, R13s) are still in development, and non-mature [15, 26, 491, a testing method has
been developed that does not require knowledge of the boundary conditions. In the test,
trace-free pressure and heat flux are computed from the NSF, the Burnett, the Grad13 and
the R13s equations, where values of the moments in the respective expressions are
chosen to be the values obtained from results of the kinetic models. The computed trace-
free pressure and heat flux are then compared to the values from the kinetic models
directly to check the quality of the computed results. Third-degree and fourth-degree
moments will be tested similarly (for Grad13 and R13s only).
Chapter 2 Kinetic models

2.1 Introduction
The Boltzmann equation is a nonlinear integro-differential equation, and is very
difficult to handle. Therefore, some alternative, simpler expressions have been proposed
for the collision term, which are known as collision models, and any transport equation
where the Boltzmann collision integral S(f ) (Eq. (1.2)) is replaced by a collision model
S, (f) is called a model equation or a kinetic model [3,4].

A collision model S,(f) needs to retain the main properties satisfied by the
Boltzmann collision integral [3,4, 17,34,50]. These are:
It guarantees the conservation of mass, momentum and energy, which is Eq. (1.4).
The production of entropy 2 is always positive (H-theorem)

x=-kiln f . S , ~ C Z O . (2.1)

In equilibrium, S, (f,) = 0, and therefore, the equilibrium distribution fE is

equal to the Maxwellian distribution f, (Eq. 1.8).
Yield the right transport coefficients viscosity y , thermal conductivity K , and
5R
Prandtl number Pr = -- at the hydrodynamic limit. The Prandtl number Pr is
2 K
close to 213 for all physically meaningful collision factors o ,calculated by means
2
of the C-E method from the Boltzmann equation [9]. Pr E - for ideal monatomic
3
gas is also found in experiments [511.
The collision term S, ( f ) depends on the peculiar velocity C = c - u , and not the
microscopic velocity c, since the Boltzmann equation is invariant under Galilean
transformation.
Predict positive distributionf at any situation for the kinetic model.

It is difficult to prove the last issue strictly (The author has not seen any literature
related to the prooj). What can be done is to apply the model to some real situations, such
 10
as shock waves, Couette flow, etc. If the distribution f is not positive for some range of
velocity c in the problem, then we can say that the model does not satisfy this property, at
least in that specific situation.

2.2 Motivation of kinetic equation and expression of collision frequency

For motivation of a kinetic equation, we simplify the Boltzmann collision term Eq.
(1.2) in three steps [35,52].

Step 1: Because of the collisions, the distribution will tend to a distribution fm for
which the Boltzmann collision term vanishes (means equilibrium). Thus In f m must be a

linear combination of the collisional invariants 1, ci and CZ. The distribution f 'f " in

the collision term Eq. (1.2) refer to the velocities after the collision, which means they
may be replaced by fif: ,

S(f) fm= I(fmT

fm"-f')ogsin8d8d&dc1 . (2.2)

Step 2: Since In f ,must be a linear combination of the collisional invariants, we may

replace fif: by fmf:,

~ m + ~ , , , = f m ~ f ~ ~ g s i n 8 d 8 d ~ d ~ ' - f ~ f ' o g ~ i n i 9 d(2.3)

8d~dc'.

Step 3: The difference between the two integrals may be neglected. This last step leads
to the collision term of a kinetic model
gm-+ Sm= v(fm - f ), (2.4)
where the collision frequency is identified as
v(xi,t,ci)= If~ogsin8d8d&dc1. (2.5)

One should not rely on the above equation as the exact value for collision frequency V

because all the assumptions in steps 1-3 will lead to errors. However, it is possible to
 11
choose the function v so as to fit the transport coefficients (viscosity and thermal
conductivity) to measurements.

When power interaction potentials between particles are applied, to simulate the real
situation, [5, 6, 341 4 - r-'"-" (n > 3 , where 4 is the interaction potential, r is the
distance between particles, and n gives the order of the potential, n=5 represents Maxwell
molecules, and n + 00 describes hard sphere molecules), one can show [6] that

ag sin tlie = gn-~-ls,dso (2.6)

where so = so (8).

Combining Eqs. (2.5-2.6), considering f,,, as the Maxwellian f,, and after some
manipulation, the expression of collision frequency in spherical coordinates can be
reduced to

,
where Qo(n) = s ds d& is a constant which depends on the interaction potential, but is

6 = - and
L
q =-
b
are
independent on the macroscopic gas properties.
m J2RT
dimensionless peculiar velocities.

2.3 Existing kinetic models

Several kinetic models have been proposed and developed in the past. Among them,
BGK model [3, 281, ES-BGK model [11, 29-33], v(C) -BGK model 134, 3512, S model
[31,36,37], and Liu model [38-401 are commonly used. These models will be described
briefly, and compared in this section. Table 2.1 shows their basic properties, such as
whether the models fulfill the requirements 1-6 in the above section.

This model was also discussed in 13,531, but explicit form of the collision frequency was not discussed
there. In [54-561, this model is discussed only for the hard sphere molecules (denoted as rigid-sphere
molecular model).
 12
The collision term S,( f ) in all above kinetic models can be written as

sm(f )=-vCf - f r @ ) , (2.8)

where v is the coIlision frequency, which is a constant with respect to velocity C in the
above models except for the v(C)-BGK model, and fr@ is a specific reference
distribution function, which depends on the model.

The BGK model is the simplest kinetic model, in which fr@ = f,. This model is

widely used for theoretical considerations, but it gives P e l . All models described below
were introduced to correct this failure.

The ES-BGK model (ellipsoidal statistical BGK) is also known as Gaussian-BGK

model, in which fr@ = f a , and

Here, the matrix 1 is defined as

where b is a parameter that serves to adjust the Prandtl number, is the unit matrix,

and E is the inverse of the tensor X .b must be in the interval [-I/ 2,1] to ensure that R ,
is positive definite. This expression for the reference distribution f, is defined by the
following ten conditions (the first five conditions are the conservation laws),
I(f,-f)dc=o, Ici(f, - f ) / d e = ~ , Jc2(f, -f)/de=~,

~ ~ < ~ ~ ~ , ( f , - f ~ c = v ( b - l ) o. ,

From Eqs. (2.9-2.1 1, 1.3), one can get

If& =P , Ic,f,dc=0, Ic,c,f,dc = pAu = (1-b)p6, +bpu, (2.12)

which will be used in the numerical work in Chapters 3 and 4.

 13
Only recently, Andries et al. succeeded in proving the validity of the H-theorem for
the ES-BGK model [32,33], which revived the interest in this model.

The Burnett equations for the ES-BGK model with power interaction potentials have
been proposed and examined by Zheng and Struchtrup in [ll]. The ES-BGK Burnett
equations are found to be identical to the Burnett equations for the Boltzmann equation
only in the case of Maxwell molecules, while the Burnett coefficients exhibit some
differences for other interaction types (e.g. hard sphere molecules). The linear stability of
the ES-BGK Burnett equations is also discussed in [l 11 and the main results are given in
Section 5.2.

The v(C) -BGK model is also called as BGK model with velocity-dependent collision
frequency, in which fM = f, ,and

fr = a.exp(-rc2 + y i ~ i ) . (2.13)
Here, the coefficients a, r >O, y are chosen so as to guarantee the conservation of mass,
momentum and energy. In general situations, the explicit theoretical expressions of a, I?,
y cannot be given, and only numerical values are obtained from these five constraints.

At the small Kn number situations, the distributionf and the reference distribution f,

should be close to Maxwellian f, ,then approximate explicit expressions of a, T , y can

be obtained through the first order C-E method. In these situations, one finds [34,35]
 C
where q = -is a dimensionless velocity3.The transport coefficients and the Prandtl
m
number can be obtained from this model as

In principle, the collision frequency v(q)can be assumed to be any function with two
unknown coefficients, which are determined by experimental values of viscosity ,U and
thermal conductivity K (or one experimental parameter and the condition P1=2/3). Some
expressions of ~ ( qcan
) be found in [34]. Here, only one expression, which will be used
in the numerical test in Chapters 3 and 4, is listed.
P=~(I.O+~'), (2.16)
with two coefficients a = 0.0268351 and y = 14.2724, where P is the dimensionless
collision frequency, which is defined as

However, expressions of v(C) based on the Boltzmann collision term (Eq. (2.7)) do not
give the proper Pr number [34].

The S model was proposed by E. M. Shakhov [31,36,37], in which f,, = fs,and

 The Liu model was proposed by G. Liu [38-401, in which fW = f, ,and

Here 0i2)(2)is a coefficient, which is found in [12,38].

Since heat flux is a vector, and the range of the peculiar velocity is (-m, =J), f, and
f, will become negative for large values of the peculiar velocity. The further result is
that the distribution function f from these two kinetic models would be negative for
some range of peculiar velocity.
 Table 2.1: Comparison of kinetic models
Kinetic models BGK model I ES-BGK model v(C) -BGK model I S model Liu model New model I New model I1
in collision term fNI (Eqs. (2.209 fNII (Eqs. (2%
fmf
S,(f) (1.8)) 1 (2.18)) (2.19)) 2.23)) ( 2.43)) -
Collision frequency V Velocity Velocity Velocity dependent - I Velocity Velocity Velocity dependent 1 Velocity
-
Related references -
1. Conservation of
mass, momentum and
energy
2. H-theorem Proved Proved Only proved Only proved in
in the near the near local
equilibrium situations situations
equilibrium situations
situations -
3. In equilibrium Yes
f =f&, -
+
4. Viscosity p,
Thermal conductivity
+
K ,and Prandtl number
Pr
F'F1 .o pr =x-b -
5. Galilean invariance
6. Positiveness off (To
Satisfied
Satisfied
Satisfied
Satisfied
Satisfied
Satisfied
Satisfied
Possibly not
Satisfied
Not true in
Satisfied
Satisfied
Satisfied
Satisfied
-
the author's best some
knowledge)
Remark Simplest fw is a local In order to obtain
situations I
1. Satisfy the requirements of realistic
-
model; fn/ anisotropic Pr=U3, the collision frequency and correct
is a local Gaussian; Can expression of transport coefficients
isotropic simplify to the velocity-dependent 2. Can simplify to the ES-BGK model
Gaussian BGK model frequency does not and the V(C)-BGK model at certain
when b = o . met with the physics. conditions
A
 17
2.4 Two new v(C) -ES-BGK type kinetic models
In the existing kinetic models, the collision frequency v is assumed to be an average
value, which is not dependent on the velocity C, except for the v(C) -BGK model. For
real gases, however, the collision frequency is a function of the velocity C. The C-
dependence of v has an important influence on the results at large Kn situations (e.g.
Kn 2 0. l), and thus v(C) in kinetic models should be close to the value predicted from
the Boltzmann equation. However, when the realistic collision frequency is used in the
v(C) -BGK model, the Prandtl number is not 213, but very close to unity [34].

Therefore, we propose a new type kinetic model, in which the realistic collision
frequency can be used, while the transport coefficients are predicted correctly (including
Pr = 213). The idea is to combine the anisotropic Gaussian of the ES-BGK model and
the velocity-dependent collision frequency of the v(C) -BGK model to develop a new
kinetic model, named as ES-BGK model with velocity-dependent collision frequency, or
v(C)-ES-BGK model. There are two ways to create the reference distribution frq

following this idea, which will be referred to as new kinetic model I and 11.

2.4.1 New kinetic model I

In this new kinetic model, the collision term S, ( f ) is

s, ( f 1= -v(cXf - f N , ) 9

Here, the matrix sijis the same expression as in the ES-BGK model. The coefficients a,
r (>O), y are chosen so as to guarantee the conservation of mass, momentum and
energy.

In general situations, explicit expressions of a, r (>O), y cannot be given, and only

numerical values can be obtained from the five conservation laws Eq. (1.4). At small
 18
Knudsen number situations, approximate explicit expressions of a, T , y can be given
through the first order C-E method, and this is described in the following.

At small Kn number situations, any distribution is close to a Maxwellian distribution,

so that (when we only keep the zero and first order terms)
6, bo,
r=i-f., E, =---
2RT RT pRT9
6, 6, bo,,
I-&..= -- r---
A

, yi = pi ,
" RT RT pRT

Here, the values of the undetermined coefficients 2 , f and fi are small compared to 1.
The value of parameter b must lie in the interval [-0.5, 11, which is the same range as b in
the ES-BGK model, to ensure the matrix E, is positive definite.

After performing the first order C-E expansion, one finds

This approximate solution fulfills the five conservation laws for any distribution f,, ,

thus, the coefficients A , f and pi cannot be determined from these conditions. However,
the distribution f must reproduce the first five moments, which are density, velocity and
pressure (Eq.(1.3)). It follows
a = f=o,
fi =L" ~&(~-?)CZdc,
3pT ax, v 2RT 2

At last, for small Kn numbers, one finds

 .+--I-(C' " fM
3pT ax, v
c2 ) ,,, ]
-+- 5~1
2pRT 2RT 2

.f cic, a h +_

v RT ax,, cT iax,
a T~( Rc 2
T2

After computing pressure tensor and heat flux vector from the above first order C-E
expansion, one finds

Therefore, the transport coefficients and the Prandtl number obtained from this kinetic
model are

where ,A is defined as

Combining Eqs. (2.7, 2.17, 2.26.a), one finds, after some manipulation, the following
expressions for the new kinetic model I,
where

For Maxwell molecules where n = 5 , A and B are constants4,

A(n = 5) = 0.9375 , B(n = 5,q) = 1.77245.

-
For hard sphere molecules where n = ,one finds [12,34]

where A(n = -) = 0.308855, B(n = -,q = 0.0) = 3.0, and erf(q) is the error function,
which is defined as

For other values of n, the integration in Eq. (2.7) needs to be done numerically.

For ideal gases, it is well known that the viscosity ,U is a function of temperature,
which can be described as [9]

The computation uses a Mathematica program [57].

where p0,the viscosity at the reference temperature To,will be used to determine Q0(n),
and w is a positive number of order 1. From Eqs. (2.7,2.26.a, 2.32), we can see that

For real monatomic gas, the n value locates in between five and infinity, and the w value
locates in between 0.5 and 1.0. That is to say, Maxwell molecules and hard sphere
molecules are the two limiting situations of molecular models for real monatomic gas.

Table 2.2 several n and corresponding b and Pr values for the new kinetic models
b for Pr=2/3 b=-0.5
Pr JR-2/31
213
5 1 1 .o -0.50 2/3=0.6667 0.000
6 0.9 1.00839 -0.5126 0.6723 0.008
231 .0/31.0 0.81 1.01327 -0.5 199 0.6755 0.013
10 0.72 1.01572
13 0.67 1.01615 -0.5242 0.6774 0.016
20 0.61 1 .01566 -0.5235 0.6771 0.016
w 0.50 1.0126 -0.5189 0.6751 0.013
Notes: The computation of A (Eq. : .27) uses a Mathematica program.

Table 2.2 gives numerical results of the Prandtl number, Eq. (2.26.c), for several n
values. It is seen that in order to obtain Pr=213, the b value should be a little smaller than
-0.5, which is the low limit of this new kinetic model to ensure the matrix cij is positive
definite. Therefore in the application of this new kinetic model, b is chosen as -0.5, and
Pr will be a number which very near to 213. For Maxwell molecules, b = -0.5 will obtain
Pr=2/3. Therefore, our requirement, which is that realistic collision frequencies are used
and the transport coefficients are predicted correctly in one kinetic model, is met.

Besides this, some other expression can be found in the literature, such as the Sutherland's law [58].
 When v(C) is not constant, and b=O (which implies that E~ = ~4), the new kinetic
model I is identical to the v(C) -BGK model. When v(C) is constant (therefore
Ti = Ti = 0.0 from Eq. (2.23)), this new kinetic model is identical to the ES-BGK model.

At last, let us consider the H-theorem for this new kinetic model. Since the explicit
theoretical expression of a, r (>O), y cannot be given in the general situations, the H-
theorem is hard to prove in general6. When small Kn number situations are considered, it
is found that the H-theorem is indeed satisfied. In these specific situations, from Eqs.
(2.22,2.24), one gets

Utilizing the conservation law Eq. (1.4), and after some manipulation, the production of
entropy to the first order is

6
Several ways have been attempted, such as to utilize the method used to prove the H-theorem of the ES-
BGK model, or the method for the V(C)-BGKmodel, or even the combination of these two methods, but,
so far, all attempts failed.
therefore,

Utilizing Eq. (2.26), one finally gets

2.3.2 New kinetic model II

This time, the collision term S, (f) is chosen as
 24
Here, v(C) is the collision frequency, and the coefficients a, T, (I?, = Tji 2 0), yi are
chosen so as to satisfy the following 10 equations, which are given by analogy with the

ES-BGK model (Eq. (2.1 1) and v = for the ES-BGK model)

(b- l)P

In general situations, explicit expressions for a, rij,yi cannot be given, and only
numerical values can be obtained. For small Kn number situations, approximate explicit
expressions of a, I?, y, can be obtained through the first order C-E method, and further
the H-theorem can be proved in these situations.

At small Kn number situations, any distribution is close to a Maxwellian distribution,

so that

Here, the values of undetermined coefficients 2 , cj and fi are small, compared to 1.

After performing the first order C-E expansion, one finds

This approximate solution fulfills the 10 constraint equations (Eq. (2.39)) for any

distibution fN,/, since oU= - 2 , u h for small Kn number situations. Thus, the
axj>

coefficients 2, and fi. cannot be determined from these conditions. However, the
 25
distribution f must reproduce the first 5 moments, which are density, velocity and
pressure (Eq.(1.3)). It follows

If we use the first 10 moments, which are density, velocity, and whole pressure tensor, to
determine the expression of coefficients, we get

Then we cannot express trace-free pressure tensor and viscosity as a function of collision
frequency, which will give trouble to relate Pr and v(q).

Since p, is only requested to be a trace free tensor from the above equation (2.42), the

choice of f', is not unique. For an easy treatment, the expression of Fu would be chosen
to be as simple as possible. However, choosing pu = 0 reproduces the v(c)-BGK

model. Therefore, here we choose the second simplest choice of $, which is a linear
function of the trace free tensor o,.Since the unit of f', should be the same as the unit
1
of -,the expression of
RT
Pg would be r,,. = - b
2pRT
o,. The parameter b is introduced to

adjust the Prandtl number. The range of b is-1 l b 50.5, which is proved in the
following paragraph, in order to make matrix Fj positive definite, which is a little
different from the range of b value in the ES-BGK model.

Since the tensor I', must be positive definite, there is some limitation for the value of
b. Based on the idea given in [33], one can find the range of b. To simplify the
description, here we only show the derivation for the matrices in a diagonal basis
(matrices in a non-diagonal basis can be transformed into matrices in a diagonal basis).
. -a, bo, - (1 +b)S,
-bp,
" 2RT 2pRT 2RT 2pRT

1
2pRT
0
0
0
(l+b)P-be22
0
0
0
(l+b)P--be33 1
1
6pRT

---6pRT
-
( l + b X ~ , lP+2 2 + P33)--3bPIl

0)
0
P22
0
0

o ] l+-l 6pRT
0 + b ) [ :0 P,3
0
0
(l+bXP,,+ P,, + P33)-3bP,
0

o ] (+-l6pRT
0 + b ) [ :0
0
PI,
0
0
(l+bXPll+P22 + P33 )-3bP33

:]
I
0 P33 0 PI, 0 P22

In order to make Tij positive definite for any situation, and since p,, , p,, and p3, are
positive, 1- 2b 2 0 and 1+ b 2 0 need be satisfied. That is to say, - 1I b I 0.5.

At last, for small Kn numbers, one finds

Go + Ci- ",fM-(-c2
3pT ax, v 2RT
-
5)C2dc]
2
c . a r (-
+>- c2 i)]
- 9

T ax, 2RT 2

which is the same expression as f in the new kinetic model I (Eq. (2.24)). Therefore, the
transport coefficients can be obtained from this model as

where A is defined as Eq. (2.27).

 It can be seen from Eq. (2.44) and Table 2.2 that Pr=2/3 can be satisfied for any power
index n in this new kinetic model, for - 15 b 5 0.5 this time in order to make the matrix
q, positive definite. Therefore, our requirement, which is that realistic collision
frequencies are used and the transport coefficients are predicted correctly in one kinetic
model, is met.

It is easy to find that the new kinetic model 11 can be simplified into the ES-BGK
model when v(C) is set constant, while the v(C)-BGK model is recovered for
r, = rs,.

Let us consider the H-theorem for this new kinetic model II. Since the explicit
theoretical expression of a, TV (>O), y cannot be given in general situations, the H-
theorem is hard to prove in general situations7. When small Kn number situations are
considered, we find that the H-theorem is indeed satisfied, which proof is the same as the
proof for the new kinetic model I.

2.4 Conclusion
Several existing kinetic models (BGK model, ES-BGK model, v(C) -BGK model, S
model, and Liu model) have been compared, based on properties that need to be satisfied
for a kinetic model. The main disadvantage of these existing kinetic models is that a
meaningful expression for collision frequency (Eq. (2.7)) and Pr 213 can not be
reached at the same time. In order to overcome this shortcoming, two new v(c)-ES-
BGK type kinetic models have been proposed here, which both can be simplified to the
ES-BGK model and the v(c)-BGK model. The next step will be to apply these kinetic
models in numerical computations, to see which new kinetic model is better, and whether
the new kinetic models can give better results than existing kinetic models or not. This
will be the task of Chapters 3 and 4.

7
Several ways have been attempted, such as the method used to prove the H-theorem of the ES-BGK
model, or the method for the V(C)-BGK model, or even the combination of these two methods, but, so
far, all attempts failed.
Chapter 3 Numerical work on one-dimensional shock waves

3.1 Introduction of shock wave

A shock wave is a thin transitive area propagating with supersonic speed in which
there is a sharp change of moments, in particular a speed change from supersonic to
subsonic and increases in density and temperature. Shock waves arise at explosions,
detonations, supersonic movements of bodies, and so on [59].

The problem we consider in this chapter is the one-dimensional shock wave problem
(1DSW) at steady state, in which the changes of moments are along the x direction in an
x-y-z Cartesian frame (or the X1 direction in an XI-X2-X3 Cartesian frame), and the
boundary conditions are the equilibrium states at upstream and downstream sides.

In the lDSW, the macroscopic flow velocity u, + 0 , and u, = u, = 0 , therefore for

trace-free pressure tensor elements, we have
CT,, = -20,~ = -20,~ . (3.1)

From the conservation law Eq (1.5) it is realized that the fluxes of mass, momentum,
and energy, which are pu,, pc: +p,,, and 1 . 5 +0.5pu:
~ ~ ~ +p,,u, + q , , must be
constant in the whole domain at the steady state of 1DSW. Since p,, = p and q, = 0 for
equilibrium state, these conditions, known as the Rankine - Hugoniot relations [9, 601, at
this situation can be written as
Puuu = P D ~ D 7

2
PUU; + Pa = P D ~+D
Po
3
2 . 5 + ~0 . 5 ~p U u~ ~= 2.5pDuD
~ +0.5pDuD,
where the subscript U denotes the equilibrium state at upstream side, and the subscript D
denotes the equilibrium state at downstream side.

The Mach number is defined as

 5
where a = -=is the sound speed at the upstream equilibrium state..
3

3.2 Numerical scheme

The numerical scheme we use is based on Mieussens's Discrete Velocity Model
(DVM). We will briefly recall the main ideas of this method and give some remarks in
this section and Section 4.2, and some flow diagrams of computational programs are
shown in appendix A. For a complete description, please refer to [34,41-44].

3.2.1 Introduction of explicit scheme

Any kinetic model equation for lDSW can be written as

where the collision frequency v and the reference distribution function fr6 are different
for different kinetic models.

The finite volume method is used to discretize the above equation. The space variable
x is discretized on a uniform grid defined by nodes (centers of finite volumes)
L
xi = (i - 1)Ax + x, , i = 1,. .. I , Ax = - ,L is the domain width, 1 is the total number of
I
position points. The microscopic velocities are discretized as (for any position point and
time step t , ,they are the same)
 30
Here for lDSW, since the macroscopic velocity has a non-zero value only in x direction,
c t and c) are chosen symmetrically, and are the same set (that is c,,, = -c,, ,

CyrJ,+l,n = 0 , c,,, = -c,,,,, , c,,, -

- -c,,, ), while c j and c j - u,, are not symmetric,

because the macroscopic flow velocity u , varies

~ with position. The total number of

discrete velocities at one position is J,J2J, . In the following description a dense notation
is used where
f(xi,tn,Ci',c~,C:.)=f (xi,tn,cj)=fi:h.j2.j3= f i r j .
Since the microscopic velocities are discrete, not continuous, then macroscopic quantities
(i.e. the moments), which are continuous integrals off, now must be replaced by discrete
sums on the velocity grid. For instance the density must be computed according to

The discretized kinetic model equation based on the above finite volume scheme is

where At, is the time step, and the numerical fluxes are defined as

which is a Helen Yee's second-order flux expression [41, 61, 621 with the flux limiter
function

';+is" 2 j = minm~d(f~;~
- fi!1.j9 - fi;j 9 fiZ2.j 1.
- fL,j

The definition of the minmod function is

a ab>0and(a(<(bl
b ab > 0 andla1 > # ,and minmod(a,b, c )= minmod(minmod(a,b), c ) .
0 ab<O

When distributions at near boundary positions, such as f,:;', f$ , f,?&and f:?, are

, fdlj , f L , , f A,, , where

wanted, information about boundary conditions (such as f_?,
 31
nodes i = -1,0,1+1,1+2 are ghost cells beyond boundaries) are needed from the above
discretized equations (Eqs. (3.5-3.7)). Since the upstream and downstream flows are at
equilibrium states which do not change with time, f_l,, = f ( j = f o , j and

fL.j = f L 2 . j = f i + l , j . How to obtain f,,, and f,,,,, will be discussed in section 3.2.5.
An iterative technique is applied to obtain the solution at steady state, which results are
what we are interested in.

3.2.2 Collision frequency

The discrete dimensionless collision frequency, corresponding to Eq. (2.17) for the
continuous situation, is defined as

The expression of Q:, is different for various kinetic models. In particular,

vf:GK,i,,
= 1'0 (3.9)
for the BGK model, and

1
for the ES-BGK model. Recall that - - Ib S 1 and that b = -0.5 to get P1=2/3. For the
2
v(c)-BGK model, and the new kinetic models I and 11, Q:, are not constant.

For the V ( C )-BGK model, O;,i,j is a function with two unknown coefficients (such as
Eq. (2.16) for the continuous situation), which are obtained from experimental viscosity
and thermal conductivity, or one experimental transport coefficient and condition
2
Pr = - . There are several expressions for Q(q)in [34], while in this work we only focus
3

We have
for any scalar function ~ ( 7 ) The
. left hand side in the above equation is an integral in
spherical coordinates, while the right hand side in the above equation is an integral in
Cartesian coordinates. Using this, it follows from Eq. (2.15) that

Therefore the two conditions used for the discretized v(c)-BGK model to get the two
coefficients are

where Aqtj = A?#'j,A<j2A771j3 = . These two conditions are solved by the

Y:
( 2 ~ ~ "
Newton-Raphson (N-R) algorithm [63] (also known as the steepest descent technique
[64], or as Newton method with a backtracking linesearch [65]) in this work, the flow
chart for this method to obtain two coefficients is similar to Figure A. 1 in Appendix A.

For the new kinetic models I and 11, from Eq. (2.28-2.31), we have the following
expression for discrete situation,

with

In the derivation, Eq. (3.12) need be used.

For Maxwell molecules where n = 5 ,

and for hard sphere molecules where n = w ,

with A(n = m) = 0.308855, B(n = m,qtyj = 0.0) = 3.0, and erf(#,) is the error function.
After some numerical calculation, it is found that Pr 213 is indeed satisfied in the
discrete situation as well as the continuous situation for the new kinetic models and all
test cases we did.

3.2.3 Reference distribution f$,i,, (original)

The main feature of Mieussens's DVM is that the reference distribution function fr@

is not discretized directly, but determined by the minimum entropy principle (Therefore,
the values of coefficients in the discrete reference distribution are not identical to them in
continuous situation, such as y",,,, + 0.0 in the following Eq. (3.16) for the BGK model,
while it is zero in the continuous situation). Coefficients in the reference distribution
fr'f,i,j of all kinetic models (e.g. r in Eq. (2.13)) are obtained through the N-R
algorithm in this work. A nonlinear system is easier and more robust to be solved when
magnitudes of the equations are the same, therefore dimensionless quantities are used in
the program. Expressions of Jacobians used in the N-R algorithm are given in Appendix
B. The dimensionless reference distribution F
f: ,i,j is defined such that
 For the BGK model, the dimensionless reference distribution is

and the conditions to determine the three coefficients rGK,i

and 3/iGKVi)
are the
conservation laws Eq. (1.4), which is given in the discrete case as

For the ES-BGK model, the dimensionless reference distribution is

F&,i,j = a L , .exP(-r&,n.i(5'~jl Y -'%,yy.i(($j2 P +(n:j3 Y )+ ~ L , i ~ ; j , ) 9

(3.18)

and the conditions to determine the four coefficients are the discrete form of Eq. (2.12),

J
Here we have used the abbreviations pk,,,, = (c:, )2 f,&j,j A ~ and
j=1

J
p;GK,yy,i
=
j=1
Y f:GK,i,jAc, which are pressure tensor components computed with the
C(C;,~
reference distribution function of the BGK model. Since the macroscopic flow velocity
u:, varies with position and time step, the discrete velocity set ( c , ~
and ClSi)has a lack

of symmetry, so that piGK,,, is not exactly the same as piGK,yy,i

or pln (Note:
computational results show that their values are very close to each other). Using Eq.
(3.19) as conditions to determine coefficients can ensure the discrete collision operator is
zero (means at equilibrium state) if and only if fi: = fiGK.. [42].
 35
For the V(C)-BGK model, the dimensionless reference distribution is

F "Y.J J. = ay.1n . e x p ( - ~ ; , i ( q : j p + y;l,i$jl 1 (3.20)

and the conditions to determine the three coefficients are the discrete form of Eq. (1.4),

J - f")
~ P : ~ ( ? # ~'""
Y pi" '"
(f".

j=1

For the new kinetic model I, the dimensionless reference distribution is

F;I,i.j = a t ~ , i.eXP(-'iI,i('&.n,i(~:j~ Y +'&,y.y,i(($j2Y +h:j3 )

Y ) ) + G ~ , i q : j ~9 (3.22)
and the conditions to determine the three coefficients are the discrete form of Eq. (1.4)

For the new kinetic model 11, the dimensionless reference distribution is

F&,i,j =aL.i.exP(-'~~~.n.i(q:~~Y-'i~~,p,i((q:j~Y + ( ~ : j ~ ) ~ ) + y " , , i ' $ j , 1

9 (3.24)
and the conditions to determine the four coefficients are the discrete form of Eq. (2.39),

(f;II.i,j - A:') Ac = 0 , c g y q :
J
, b&,i, j - .fi;) Ac=O,
i g ; j '91 'dl
J=I pi" j=1 pi"

k$:j(q:j1 Y f i j - Ac = b ~ , x . r , -i p ; , i )
j=1 pi" 2 ~ ;
3.2.4 Reference distribution fr>,i, (linearized)
As we see from the above section, equations to determine coefficients in the reference
distribution fr>,i*j form a nonlinear system, and the N-R algorithm (or some other non-
linear solver) is needed to solve the equations. This makes computation program complex
and requires a considerable amount of computational time.

In order to simplify the program and save computational time, and simultaneously get
almost the same results, the author proposes to use linearized reference
distributions f$,i,j, instead of the original non-linear functions fr>,i, in the program. If
this is done, the unknown coefficients follow from a linear system of equations, which
can be solved much faster. The reference distributions are expanded around the
Maxwellian distribution, and the expressions of linearized dimensionless reference
distributions are given in the following, while the corresponding linear equations can be
found in Appendix C.

For the BGK model, the linearized dimensionless reference distribution is

F&K,i,j = eXP(- b:j Y k 1 L . i +a2kK,i ~:j~'+~~iGK

($j, i )2 ). (3.26)

For the ES-BGK model, the linearized dimensionless reference distribution is

F&i,j = (V2j)2). (a1k.i+a'L,i $jI + a3;s.i ($jI P + a 4 k , (($j2 P h;j3P)),(3.27)

+

and this choice makes results of the ES-BGK model with linearized fr!f,i,j and b = 0

identical to results of the BGK model with linearized fr>,i,j.

For the V ( C ) -BGK model, the linearized dimensionless reference distribution is

F".
YJ+J. = e ~ ~ ( - ( q : ~ ~ ) . ( a+a2;,77:,
l> +a3",(77i'fj)2). (3.28)

For the new kinetic model I, the linearized dimensionless reference distribution is
 For the new kinetic model XI, the linearized dimensionless reference distribution is

~ l =~
e x p~
( - ( q, $ )~
( a ~,~ .~
~+ a 2 L , i ~ ; ~+a3~,i(q:j,
, )2 +a4b,i((7;j2 )2 +($j3 )2)). (3.30)

In the above, the reference distributions are expanded around the Maxwellian
distribution, which is an isotropic Gaussian distribution. For the ES-BGK model, and the
new kinetic models, in which an anisotropic Gaussian distribution is used in the original
reference distribution, there is another way to build their linearized reference distribution,
which is expanding their original reference distribution around an anisotropic Gaussian
distribution. This choice is supposed to give better results than the first one, for the
isotropic Gaussian distribution in the expression of first choice (Eq. (3.27)) is in fact a
linear approximate expression of the anisotropic Gaussian distribution in Eq. (3.3 1). If we
consider the discrete reference distribution at limiting situation (bounds of velocity grid
approach to infinite and step of velocity grid approaches to zero), then we can recover the
continuous reference distribution Eq. (2.9) from discrete reference distribution Eq. (3.18)
or Eq. (3.31), but can not from Eq. (3.27). This is another reason why the second type
linearized reference distribution is supposed to be better than the first type.

Following this idea, the second type linearized reference distribution for the ES-BGK
model is

where

which are obtained from Eq. (3.19), are the discrete values of the dimensionless
coefficient E~RT in the ES-BGK model. Because of the lack of symmetry of the discrete

velocity set ( c , ~and C:i), ~ i ~

is not~exactly
, ~the ,same
~ as piGK,,, or pi" [42],
 38
results from this for b = 0 are not exactly the same as (but very close to) results from the
BGK model with linearized reference distribution.

The second type linearized reference distribution for the new kinetic model I from the
above idea is

The second type linearized reference distribution for the new kinetic model II from the
above idea is

3.2.5 Boundary conditions and initial guess

As described in the last paragraph of Section 3.2.1, the boundary conditions used for
lDSW are fOvj and which are the discrete equilibrium distribution (same
f,+l,j,

expression as the reference distribution in the BGK model, which is Eqs. (3.15-3.16)) at
the upstream and downstream sides. First, fo+j is obtained from the upstream values of
density, velocity and temperature, which are given parameters, through the same idea as
how to obtain the f;G,,i,j. Then the three constant fluxes of mass, momentum, and

EF, = -x
1
2 j=1
, the upstream side are computed from fo,j . Then we compute
c: c2fo, j ~ c at

the downstream mass flux (or say momentum) MI+, = M o , momentum flux
MF1+,= MF, , and energy flux EF,+, = EFo ,which correspond to the Rankine-Hugoniot
relations Eq. (3.2). Last step is to compute the discrete equilibrium distribution at the
 39
downstream side fI+,,j from the above three moments. In the program, the dimensionless

distribution FI+,,j
(which is not the same as Eq. (3.16)) is introduced,

FI+I,j = fI+l,j'c = 'I+, "xP(-~I+I (VI+I,~Y YI+IVI+I,~ 1 9

(3.34)
and the conditions to determine the three coefficients are

The same procedure was also used by Mieussens in his work [66]. Note that this
approach to get fI+,,j guarantees conservation of fluxes at the discrete level, while

downstream densityp,,,, velocity u,+, and temperature TI+, obtained from this way
would be a little different from their values from the Rankine- Hugoniot relations at the
continuous situation directly.

The shock wave is computed through time-stepping into steady state, and in the
program, the initial guess &:j is a step function, which is built by fOsj and with the
f,+l,j

jump occurring in the middle of the computational domain.

3.2.6 Time step, space grid and velocity grid

Since an explicit scheme is used, the time step At, is limited by the CFL condition
[4 1,431
a
At, =

where 0 < a < 1.0. In the program of this work, a = 0.99 is used. This restriction does
not need to be applied if the implicit scheme of Mieussens's DVM [34, 41-4419 is
adopted. From Eq. (3.36), we know when the step of space grid Ax decreases, or bounds
of velocity grid increases, the time step will decrease. If the time to reach steady state for
- --

The implicit scheme used by L. Mieussens in fact is not a fully implicit scheme, for the collision
frequency used is still the value computed from previous time step.
 40
certain kinetic model at certain flow situation is thought as a fixed number, then decrease
Ax or increase bounds of velocity grid means more iteration steps are needed to reach
steady state, and at the same time more computational time is needed for each iteration.
Therefore, in the computation, large Ax and small bounds of velocity grid are preferred.
For different kinetic models at the same flow situation, max (v:,), and further the time
i.i

step At, ,would be different.

Since we wish to resolve the flow structure at the microscopic level, the step of space
grid Ax should be much smaller than the mean free path (compute estimated value from
boundary or initial conditions), which is the standard for choosing the number of position
nodes in this work.

We follow Mieussens choicelo for bounds and step of velocity grid, by choosing

c,,, 5 Min(u,,, -4@), c,, 2 Max(u,, + 4 m ) , Ac, i ~ i n ( m ) .

If the values at the boundaries of velocity and temperature are only known before the
computation, then i = 0,Z + 1 in Eq. (3.37). If there were some results for the same
problem from another computation (such as the NSF equations) before your
computation, then i = 0,1, ...,I,I +1 in Eq. (3.37) is suggested.

The above rules for time step, step of space grid, bounds and step of velocity grid are
applied in the numerical tests in this chapter and Chapter 4. In order to study the
influence of step of space grid, bounds and step of velocity grid on computational results,
some tests with smaller step of space grid, larger bounds of velocity grid and smaller step
of velocity grid have been done for comparison.

lo In fact, Mieussens did not follow this choice strictly in his work (e.g. test case 1 in [42]), and explains
this by saying "I tried a coarser velocity grid, an since the results where good as compared to DSMC, I
stopped there." [66].
 41
3.3 Test examples
One situation for weak shock wave (Ma=1.5), one for medium shock wave (Ma=3.0),
and one for strong shock wave (Ma=6.0) have been tested, which are shown in Table 3.1.
Table 3.2 shows the parameters used in the numerical tests of kinetic models. Cases 3.1-
3.3 in Table 3.2 corresponds to the three situations in Table 3.1, and cases 3.4-3.7 are
tested in order to see the effect of step of space grid, and cases 3.8-3.9 are tested in order
to see the effect of bounds and step of velocity grid. In the tests, width of finite volume
cell, bounds and number of discrete velocities are chosen following the description in
Section 3.2.6. For each case in Table 3.2, the BGK model, the ES-BGK model with
b = -0.5 , the v(c)-BGK model with Eq. (3.1 1) as collision frequency, the new kinetic
model I with b = -0.5, and the new kinetic model II with P1=2/3 are tested. For each
kinetic model, both original and linearized frH are applied in cases 3.1-3.7, while only

original frH is applied in cases 3.8-3.9. Cases 3.8-3.9 are done much later than cases 3.1-
3.7, and in order to save computational time, smaller number of iterations is used. The
discussion in Section 3.4.2 shows the reason why we can choose smaller iteration number
than the value in cases 3.1-3.6 in computation. Table 3.3 shows some common
parameters used in the numerical tests of DSMC 1671.

Table 3.1: Situations of numerical tests of kinetic models and DSMC for lDSW
Situation Mach Number Upstream velocity (m)
Sa 1.5 1116.25
Sb 3.O 2232.5
Sc 6.0 4465.0

There are some common parameters in the numerical tests, which are: hard sphere
molecules; material is Helium; upstream temperature is 160.0 K, upstream number
density n is 2.889E21 l/m3, viscosity at the reference temperature p, is 1.86E-5 kglm-s;

reference temperature To is 273.0 K [68]; Boltzmann's constant k is 1.381E-23 J/K;

Avogadro's number N is 6.022E23 llmol [60]; molecular mass of Helium m is 4.003
 42
glmol [69]. Mean free path at upstream side I is 1.287 mm, based on Cercignani's
definition [ 1 , 31,67,70]

Table 3.2: Parameters in the numerical tests of kinetic models for lDSW
Case Situa
tion
Domain
width (m)
Running
Iteration I Num
ber of I 1
I Number of
velocities I Bounds of velocities (mls)
Xl direction I X2 (or X3) 1

Note 1: mean free path use its value at upstream.

Table 3.3: Some common parameters in the numerical tests of DSMC for lDSW
Parameter Meaning Value
FNUM Number of real molecules represented by each simulated molecule 1.4445El5
MNM Maximum number of simulator molecules 5E6
DTM 6 ) Time step over which the movement and collision steps are uncoupled 1E-7
NIS Number of time steps between samples 4
NSP Number of samples between file updates 20
NPS Number of updates to reach steady flow 100
NPT Number of file updates to completion 20000
SP(1,l) (m) Reference molecular diameter 2.19E-10
SP(3,l) Viscosity-temperatureindex 0.5
Sp(4.1) The reciprocal of the VSS scattering parameter 1.O
SP(5,l) (kg) Molecular mass 6.65E-27
MNC Maximum number of cells 400
MNSC Maximum number of sub cells 4000
SXB (m) x-coordinates of the domain limit at upstream side -0.02
DXB x-coordinates of the domain limit at downstream side 0.02

3.4 Results and discussion

3.4.1 Some important notes on dealing with results

Since the explicit scheme is used, and the time step of each iteration is very small (its
magnitude is l.D-7 seconds from the computational results), the difference between two

" Here, we omit the coefficient &A. 1.25.

 43
adjacent iterations will be very small. Therefore, it is not too informative to show the
convergence of results from the difference of macroscopic parameters between two
adjacent iterations. Another issue is that it is the first time for us to run such kind of
program, and we want to know the long running behavior of program, such as stability.
Therefore, what we do is to let the program run a large number of iterations for each
computation, and output the values of macroscopic parameters five times during each
computation which are chosen as the iterations after 115, 215, 315, 415 and 515 of the
whole iteration number, and then check the convergence from these outputs.

From the results, if one compares data based on an fixed position, the values of
macroscopic parameters are found to change slowly, and there is some movement of the
whole shock wave structure; but if one compares data after some shifting such that the
point, where the density has the average value between its values at upstream and
downstream, is fixed in the comparison, the values of macroscopic parameters are found
to remain after some iteration, which means convergence is reached. That is to say, the
whole shock wave structure is moving slowly along the x direction during the
computation, while the shape remains unchanged after some time. For larger Mach
numbers, this computational phenomenon becomes more apparent. As an example,
Figure 3.1 shows density profiles at different iterations in case 3.3 computed with the
BGK model, before and after the shifting.

Another issue found from the results is that even if the initial guess is the same for all
kinetic models, the final location of the shock wave structure within the computational
domain from computation using different kinetic models will not be the same. As an
example, Figure 3.2 shows the final density profiles from the BGK model and the ES-
BGK model in case 3.2 before and after the shifting.

The above two issues happen because there is no fixed coordinate label for the shock
profile inherent to the problem [l, 341. In order to make a fair comparison, all data will be
analyzed only after the shifting, for the two reasons mentioned above. After shifting, the
origin point x = 0 in the new Cartesian frame is some special point, which would be the
 44
point where the density takes on the arithmetic mean of the downstream and upstream
values [I], or where the derivative of density takes on the maximum value [I], or where
is defined by the following relation [34]

I ' ( p ( 4 - P" )dr = 1:(P, - pG))dr 9

(3.39)

where subscript U means the upstream state, subscript D means downstream state, and
the point x * (which is called "equal area point" in [34]) is the new origin point. The
above first and third ways for shifting has been applied by the author, and the results are
found to be very similar.

Since the step of space grid Ax for same Mach number in different computation case
would be different (such as for Ma=3.0, Ax in case 3.2 is not same value as it in case
3.5.), in order to do comparison, linear interpolation technology has been used to make
Ax in each case for same Mach number is the same as the value used in the DSMC
computation. Another reason using linear interpolation to add some nodes in results is to
make the density at the new origin point meet the shifting condition in the above
paragraph better.

In this chapter, all graphs and data are based on the first way of shifting, which means
the new origin point is where the density is the mean value of its values at upstream and
downstream, and the linear interpolation, which makes Ax in each case for same Mach
number is the same as the value used in the DSMC computation.

In the discussion, the average relative error of some macroscopic parameters is often
used, which is defined as

where g> means the macroscopic parameter considered (in lDSW, it would be density p ,
velocity u, , temperature T , pressure p , trace-free pressure tensor element a,, and a,,
 45
and heat flux q,'2), p1 means the values of the macroscopic parameter considered at
situation 1, p2 means the values of the macroscopic parameter considered at situation 2,
subscript i denotes the position point, I,,, means the total number of position points used
in comparison (which might be not same as the number of cells in the DMSC, for the
shifting makes the common domain part of two compared situations smaller than the
original domain size), is the average relative error, which means the average value of
the absolute value of relative errors of the macroscopic parameter considered at two
situations for the position points considered.

Since relative error becomes meaningless when a quantity approaches zero, and the
values of o l l , o,,, and q1 near or at upstream and downstream states are near zero (e.g.
see graphs (e-g) in Figures 3.7-3.9), then how to choose a domain to compute the average
relative error of oll, o,, , and q, is hard to decide. Therefore, the average relative error
of o , , , o,,, and q, are not considered here, and when average relative error is
mentioned in the description, we always mean the average relative error for p , u, , T and
p ,not specific one parameter.

During the analysis of the computational results, two steps have been done for each
kind of analysis (such as convergence and stability, effect of cell width, effect of
linearized frq, effect of bounds and step of velocity grid, comparison among kinetic
models and DSMC). The first step is to graph the output of one macroscopic parameter
from different situations in one figure. The second step is to compute the average relative
errors of macroscopic parameters. From these two steps, we can discuss how much
difference the results obtained from two different situations13.

l2 For IDSW, other basic moments are equal to zero (such as u 2 , u3 q 2 , q3), or are related to these basic
moments (such as 033 = 02,,p33 = pZZ= p + 022 , pl, = p + O,, ).
l3 Graphs and data from these two steps are too many to be shown in this thesis. If interested in graphs and
data which are not shown, contact the author (or his supervisor) for them
 - 15 -10 -5 0 5 10 15
Dimensionless domain width x / l

(a) Before shifting

-20 -15 -10 -5 0 5 10

Dimensionless domain width xll

(b) After Shifting

Figure 3.1: Density profiles at five certain iterations (BGK model, case 3.3: Ma=6.0) (Triangle for
Iteration=16000, Star for Iteration=32000,Diamond for Iteration=48000,Box for Iteration=64000, Triangle
filled for Iteration=80000)
 -15 -10 -5 0 5 10 15
Dimensionless domain width x l l

(a) Before shifting

-20 -15 -10 -5 0 5 10

Dimensionless domain width x / l

(b) After Shifting

Figure 3.2: Density profiles from BGK and ES-BGK models (case3.2: Ma=3.0) (Triangle for BGK,
Triangle filled for ES-BGK)
 48
3.4.2 Convergence of results
The two steps mentioned in the last paragraph of section 3.4.1 have been done to
check convergence of results and stability of the program. In the first step, one
macroscopic parameter after five certain numbers of iterations have been graphed in one
figure to see whether the profiles are overlapped or not. This step has been done for all
computations, and these graphs (e.g. graph (b) in Figure 3.1) show the results are
converged. Then the average relative errors of macroscopic parameters among values
after five certain iterations are computed, where the situation 1 in Eq. (3.39) now is the
four intermediate iterations, while the situation 2 is the final iteration. The results show
that those values of average relative errors are smaller than 0.01 for almost all cases. Note
that this small difference of macroscopic parameters is obtained from two iterations,
which distance is at least 4800 iterations. From the above two steps, it can be said that the
computational results are converged, and the stability of the numerical scheme is good.
Since computational results have been converged already at the iteration, which is 115 or
2 6 of final iteration, we do not need to compute so many iterations, when some proper
criterion for convergence is developed and applied in the future. To find a proper
criterion is a very useful task to save the computational time, meanwhile get the results
converged.

Here the author would like to give a very simple estimation for time to reach the
converged result. The BGK model for a homogeneous situation is used in the estimation,
which can be written as

where the collision frequency v and the reference distribution function f,,, are not
function of time. Transform terms and do integration from initial situation ( t = O and
f = f,) to the situation at time t, one gets
 49
1 -.
where the mean free time z =-= fl If we say the converged result is reached when
v P

-fm =lo4, then the estimated time to reach converged results is t = 9 . 2 .~

f,,, - fB,

For the IDSW, let us compute the mean free time at upstream state, where the
viscosity is computed from Eq. (2.32),

If we compute the mean free time at downstream state, the value is smaller than the
above number. Therefore the magnitude of time to get converged result for the lDSW
from the above simple estimation is 10" s. The magnitude of total time step is
loe2 - s (refer to Table 3.4, which shows the total time step (unit: seconds) for
kinetic models in the computation of cases 3.1-3.3), which is much larger than the above
estimated time, for all our computational results. Therefore, we can say that our
computational results are the converged results from the above comparison.

Table 3.4: Total time step (seconds) for kinetic models in the computation of cases 3.1-
3.3
Kinetic model BGK ES-BGK V(C)-BGK New I New II
Case and situation
3.1 Sa 7.97D-3 8.23D-3 4.33D-3 7.471)-3 7.47D-3
3.2 Sb 4.851)-3 5.12D-3 2.01D-3 4.36D-3 4.36D-3
3.3 SC 2.55D-3 1.36D-3 8.59D-4 1.13D-3 1.13D-3

Similar to the kinetic models, the DSMC program is run for a large number of
iterations for each computation (parameter NPT in Table 3.2), in order to make sure the
results are converged, and to study the long running behavior of program, and to
eliminate the inherent stochastic noise of DSMC. Macroscopic parameters are output
several times in the course of each computation. After similar analysis to kinetic models
has been done, the results are realized to be converged, and the program is stable.
 50
3.4.3 Discussion
Since the steady state is what we are interested in, we need to establish whether
converged computational results are indeed results at steady state. In this part of the
analysis, three fluxes (mass flux pu,, momentum flux put + p,, , and energy flux
1S p u , + 0 . 5 ~ :+ p, ,u, + q, , which all should be constant in the whole domain at the
steady state of 1DSW) can be used as a criterion. From the results, this criterion is indeed
satisfied, see graphs (e, f, h-j) in Figures 3.7-3.9.

From the computational results in Table 3.4, it is shown that the v(c)-BGK model
requires the smallest time step At, (is only about 50% of At, of other kinetic models)
among the tested kinetic models, and At, for other kinetic models are similar.

Results from the new kinetic model I and the new kinetic model 11 are similar (see
Figures 3.7-3.9), and if the structure of program and computational time for one iteration
is considered, the new kinetic model 11 is better than the new kinetic model I. The
computational time for one iteration of the new kinetic model 11 is about 80% of the
computational time needed for one iteration of the new kinetic model I. The reason is that
when the reference distribution is solved (which occupy a large portion of computational
time), the new kinetic model I needs results from the ES-BGK model, while the new
kinetic model IT does not, therefore the new kinetic model I need to perform the N-R
algorithm to find coefficients of reference distributions three times, while the new kinetic
model II only need to perform the N-R algorithm two times.

Let us consider the use of the linearized reference distribution frH discussed in
Section 3.2.4. From Figures 3.3-3.5, which are examples to show temperature and heat
flux profiles from kinetic models (the new kinetic model I is omitted, for its results are
similar to results of the new kinetic model II) with original and linearized f , at Ma=1.5,
3.0 and 6.0 situations, and data obtained from the two steps mentioned in the last
paragraph of section 3.4.1 (when the average relative error is computed, use the results
 51
with original fr@ as the situation 2 in Eq. (3.39)), the following conclusions can be
drawn (If it is not pointed out for which kinetic model, then it means the statement is
applicable for all kinetic models).

When the computational time for each time step is considered, using the linearized
frd can save 20-35% time compared to using the original fr@ since coefficients in

the linearized frd can be solved directly from a linear system, while the coefficients
in the original frd need be solved through some iterative method, such as the N-R
algorithm used in this work.

When the Mach number increases, the difference between results obtained by using
the linearized fr@ and results obtained by using the original fr@ also increases.

For Ma=1.5 and 3.0, the average relative errors for all macroscopic parameters, all
kinetic models and both types linearized frd are smaller than 0.01, except for the

V(C)-BGK model in cases 3.5.

For the BGK model, the average relative errors for all macroscopic parameters are
smaller than 0.01 for all test situations.

For the ES-BGK model, average relative errors for the second type linearized f M are

much smaller than the values for the first type linearized f r @ , while both types are
acceptable. Normally, the former values would be only one-tenth of the later ones,
which are still smaller than 0.025 for Ma=6.0.

For the v(c)-BGK model, the average relative errors would be 0.05 or higher, which
are not acceptable, for Ma=3.0 and 6.0. (the reason for these large errors is still being
investigated).
 52
For the new kinetic model I with hard sphere molecules, using the first type linearized
frM in cases 3.1-3.6 gives divergent results (NaNQ appear in the output of the

FORTRAN computation program), while using the second type linearized f r,f yields
convergent results in cases 3.1-3.2, 3.4-3.5 (situation Sa and Sb) (average relative
errors are smaller than 0.01) and some divergent results in cases 3.3 and 3.6 (situation
Sc). Reasons why diverged results are obtained is still being investigated.

For the new kinetic model 11, the second type linearized frM does not give much

better results than the first type linearized frM . The average relative errors are smaller
than 0.030 for Ma=6.0.

Let us consider the effect of the step of space grid on computational results, based on
the comparison of results from cases 3.4-3.6 with results from cases 3.1-3.3, and the
comparison of results from case 3.7 with results from case 3.5 (situation sb14). Average
relative errors are smaller than 0.01 for weak shock waves (situation Sa), but would be
large as 0.05 for medium and strong shock waves when results from cases 3.5-3.6 are
compared with results from cases 3.2-3.3 (situation Sb and Sc). Average relative errors
are smaller than 0.01 for medium shock waves when results from case 3.7 are compared
to results from case 3.5. When the Mach number of lDSW increases, the thickness of
shock wave decreases first, and reach some minimum value at about Ma=3.0-5.0, then
increases [I, 9, 261. Based on our computation, choosing the step of space grid as one
half or one third of mean free path at upstream side is not accurate enough for medium or
strong shock waves, while one sixth of the mean free path would be enough. A smaller
step of the space grid will increase the computational time for each time step and
decrease the value of time step (which means more iterations need to be done to get
converged results). It is not suggested, based on the author's experience (e.g. compute the
new kinetic model I in case 3.6 with 40000 iterations need about one month, which is
almost the limit running time for each computational job, on "Minerva", which is a

l4 For the v(c)-BGK model in case 3.7, only run Iteration=24802, the stop reason is the values of
coefficientsof collision frequency can not be gotten correctly from the N-R algorithm.
 53
cluster of workstations provided by the University of Victoria, and is used for the
computation), to do some further tests using smaller step of space grid than it used in case
3.6, before some ways are applied to decrease the computational time, such as an
effective criterion to check the convergence, implicit numerical scheme, parallel
computing, higher performance workstation, etc.

Let us consider the effect of bounds and step of velocity grid on computational results,
based on the comparison of results from cases 3.8-3.9 with results from cases 3.1.
Average relative errors are smaller than 0.005, which tell us that the use of Eq. (3.37) as
standard for choice of bounds and step of velocity grid for lDSW is acceptable.

As some example graphs to show the use of linearized f, , effect of step of space
grid, effect of bounds and step velocity grid, Figure 3.3 shows the profiles of T and q,
(heat flux along the shock wave direction) for Ma=1.5 (cases 3.1 and 3.4) with the ES-
BGK model; Figure 3.4 shows the profiles of T and q, for Ma=3.0 (cases 3.2 and 3.5)
with the v(c)-BGK model; Figure 3.5 shows the profiles of T and q, for Ma=6.0 (cases
3.3 and 3.6) with the new kinetic model I.;Figure 3.6 shows the profiles of T and q, for
Ma=1.5 (cases 3.1,3.8 and 3.9) with the new kinetic model I.
 -15 -10 -5 0 5 10 15
Dimensionlessdomain width x / l

-15 -10 -5 0 5 10 15
Dimensionless domain width x / l
(b)

Figure 3.3: Results to show the use of linearized fief and the effect of step of space grid (Ma=1.5, ES-
BGK model) (Triangle for the original f r in
~ case 3.1, Triangle filled for the original fmf in case 3.4,
Diamond for the first type linearized f M in case 3.1, Diamond filled for the second type linearized frH in
case 3.1)
 Dimensionless domain width x / l

-15 -10 -5 0 5 10 15
Dimensionless domain width x/l
(b)

Figure 3.4: Results to show the use of linearized f M and the effect of step of space grid (Ma=3.0, v(c)-
BGK model) (Triangle for the original frd in case 3.2, Triangle filled for the original fp6 in case 3.5,
Diamond for the linearized fWf in case 3.2)
 -15 -10 -5 0 5 10 15
Dimensionlessdomain width xll
(a)

-15 -10 -5 0 5 10 15
Dimensionless domain width xll

(b)

Figure 3.5: Results to show the use of linearized fief and the effect of step of space grid (Ma=6.0, new
kinetic model 11) (Triangle for the original fief in case 3.3, Triangle filled for the original f,ef in case 3.6,
Diamond for the linearized fM in case 3.3)
 -15 -10 -5 0 5 10 15
Dimensionlessdomain width x/l
(a)

-15 -10 -5 0 5 10 15
Dimensionlessdomain width x / l
(b)
Figure 3.6: Results to show the effect of bounds and step of velocity grid (Ma=1.5, new kinetic model I)
(Triangle for case 3.1, Triangle filled for case 3.8, Diamond for case 3.9)
 58
3.4.4 Comparison with DSMC
Graphs and data to show the comparison of kinetic models and DSMC are obtained
from the two steps mentioned in the last paragraph of Section 3.4.1. Table D.1 in
Appendix D shows average relative errors of macroscopic parameters between kinetic
models and DSMC in cases 3.4-3.6. Figures 3.7-3.9 show profiles of macroscopic
parameters from kinetic models (cases 3.4-3.6) and DSMC at Ma=1.5, 3.0 and 6.0. The
following points are realized from these graphs and data.

For hard sphere molecules, results from the new kinetic models locate in between
results from the ES-BGK model and results from the v(c)-BGK model. It is found
that the shape of shock waves from the V(C)-BGK model is sharper than the shape of
shock waves from the DSMC, while they are both sharper than results from the ES-
BGK model. The above two points indicate that results from the new kinetic models
are closer to results of DSMC than results from the BGK model, the ES-BGK model
and the ;(c)-BGK model.

In the temperature profiles, graph (c) in Figure 3.9, for Ma=6.0 from the ES-BGK
model, the new kinetic models and the DSMC exhibit the overshoot phenomenon,
while results from the BGK model and the v(c)-BGK model do not have this
phenomenon.

The biggest difference between results from various kinetic models and results from
DSMC always happens at the first half part of the shock wave structure.

When the degree of moment becomes higher, the difference between moment
obtained from kinetic models and moments from DSMC will be larger. For example,
the difference of a,,, a
, and q, between kinetic models and DSMC from Figures
3.7-3.9 are larger than the difference of other parameters ( p , u,, T ,and p ).
 -15 -10 -5 0 5 10 15
Dimensionlessdomain width x / l

-15 -10 -5 0 5 10 15
Dimensionlessdomain width x/l

(b)
Figure 3.7: Comparison of kinetic models with DSMC (case 3.4: Ma=1.5) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v@)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 - 10 -5 0 5 10 15
Dimensionlessdomain width x/l

(c)

- 15 -10 -5 0 5 10 15
Dimensionless domain width x / l

(dl
Figure 3.7: Comparison of kinetic models with DSMC (case 3.4: Ma=1.5) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 -10 -5 0 5 10 15
Dimensionlessdomain width x l l

-
1 !

-15
, I

-10 -5
r

0 5
r I

Dimensionlessdomain width x / l
7 r P r

10
2 ,

15
2 1 , :
(0
Figure 3.7: Comparison of kinetic models with DSMC (case 3.4: Ma=1.5) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 -10 -5 0 5 10 15
Dimensionless domain width x / l

-15 -10 -5 0 5 10 15
Dimensionlessdomain width xll

Figure 3.7: Comparison of kinetic models with DSMC (case 3.4: Ma=1.5) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 30.27

-15 -10 -5 0 5 10 15
Dimensionless domain width x/l

-
N

'
E
-.
31170
tj:
+,
2
+
0
31165

m
f 31160
s
8
7 31155
f?
n,
m
31150
3
rl
W
:,31145
0
u
a,
I2
31140
-15 -10 -5 0 5 10 15
Dimensionless domain width x/l

0)
Figure 3.7: Comparison of kinetic models with DSMC (case 3.4: Ma=1.5) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the V ( ~ ) - B G Kmodel, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 - 10 -5 0 5 10 15
Dimensionlessdomain width x / l

-15 -10 -5 0 5 10 15
Dimensionless domain width x / l

(b)
Figure 3.8: Comparison of kinetic models with DSMC (case 3.5: Ma=3.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 -10 -5 0 5 10 15
Dimensionless domain width x / l

(c)

-15 -10 -5 0 5 10 15
Dimensionless domain width x / l

Figure 3.8: Comparison of kinetic models with DSMC (case 3.5: Ma=3.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 -10 -5 0 5 10 15
Dimensionless domain width x / l

-1 5 -10 -5 0 5 10 15
Dimensionlessdomain width x/l

(0
Figure 3.8: Comparison of kinetic models with DSMC (case 3.5: Ma=3.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic,model I, Star for the new kinetic model II, Line without symbols for the DSMC)
 Dimensionless domain width x/l

(g)

-15 - 10 -5 0 5 10 15
Dimensionless domain width x/l

01)
Figure 3.8: Comparison of kinetic models with DSMC (case 3.5: Ma=3.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 -10 -5 0 5 10 15
Dimensionlessdomain width x / l

-15 -10 -5 0 5 10 15
Dimensionless domain width x/l

Figure 3.8: Comparison of kinetic models with DSMC (case 3.5: Ma=3.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 - 10 -5 0 5 10 15
Dimensionlessdomain width x / l

-15 -10 -5 0 5 10 15
Dimensionlessdomain width x/l

(b)
Figure 3.9: Comparison of kinetic models with DSMC (case 3.6: Ma=6.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 70

-15 - 10 -5 0 5 10 15
Dimensionless domain width x/l

-15 -10 -5 0 5 10 15
Dimensionless domain width x / l

(4
Figure 3.9: Comparison of kinetic models with DSMC (case 3.6: Ma=6.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 -10 -5 0 5 10 15
Dimensionless domain width x l l

(0
Figure 3.9: Comparison of kinetic models with DSMC (case 3.6: Ma=6.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model II, Line without symbols for the DSMC)
 -15 - 10 -5 0 5 10 15
Dimensionless domain width x / l

-15 -10 -5 0 5 10 15
Dimensionlessdomain width x/l

Figure 3.9: Comparison of kinetic models with DSMC (case 3.6: Ma=6.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 -15 -10 -5 0 5 10 15
Dimensionless domain width x / l

-15 -10 -5 0 5 10 15
Dimensionlessdomain width x / l

Figure 3.9: Comparison of kinetic models with DSMC (case 3.6: Ma=6.0) (Triangle for the BGK model,
Triangle filled for the ES-BGK model, Diamond for the v(c)-BGK model, Diamond filled for the new
kinetic model I, Star for the new kinetic model 11, Line without symbols for the DSMC)
 74
3.5 Conclusion
For the one-dimensional steady state shock wave problem, the following conclusions
can be drawn from the results and discussion presented above.

(1) The explicit numerical scheme, based on Mieussens's DVM with necessary
modification, works well and the program is stable for running.

(2) The second type linearized frd for the ES-BGK model and two new kinetic

models is better than the first type linearized frd for them. The linearized f , can be

used to replace the original f,, to save computational time and get almost identical
results at the same time. Exceptions are the following situations, in which no (or not
good) results could be obtained: the v(c)-BGK model for medium and strong shock
waves, the first type linearized fw of the new kinetic model I, and the second type
linearized fief of the new kinetic model I for strong shock waves.

(3) The choice of bounds and step of velocity grid based on Eq. (3.37) is accurate
enough, while the step of space grid need be smaller than about one sixth of the mean
free path at upstream side for strong shock wave situations from our computation.

(4) Through the comparison of kinetic models with DSMC, it is seen that the new
kinetic models give similar results, which are closer to DSMC than the other kinetic
models. The new kinetic model I1 is suggested for the simulation of shock wave problems
since its program structure is simpler than the program structure of the new kinetic model
I, and the Prandtl number can be exactly equal to 213, instead of near to 213 in the new
kinetic model I.
Chapter 4 Numerical work on one-dimensional Couette flow

4.1 Introduction of Couette flow

In the one-dimensional Couette flow (1DCF) problem considered in this chapter, there
are two parallel infinite-size plates, one plate is fixed, while the other plate is moving
with a certain velocity at x direction in a x-y-z Cartesian frame (or the X1 direction in a
XI-X2-X3 Cartesian frame), and the direction perpendicular to the plates is the y
direction (or the X2 direction). The temperatures of the two plates are the same, denoted
as T,. Initially the gas between the two plates is uniform, which temperature is same as
T' ,the density is pin,. What we are interested is the flow at steady state.

Some main characteristic dimensionless numbers of lDCF are the Mach number, the
Reynolds number and the Knudsen number, which are defined as

where the mean free path 1 is defined as Eq. (3.38), up, is the speed of the moving plate,

a,, = , / f R T , is the sound speed at plate temperature, L is the distance between the two

plates, and phi is the viscosity of gas at the initial state.

From the conservation law Eq. (IS), one obtains for 1DCF at steady state after some
manipulations,
 76
which means that a,,, p,, (= p + a,,), and q, + a,,u, are constant in the whole domain at
steady state, and this will be used to check whether the computed results are results at
steady state or not ,see Section 4.4.3.
One interesting observation is that from the NSF theory Eq. (1.9) follows that
NSF
oF=CTF=q, =o.
The computational results from the kinetic models and the DSMC shown below (graphs
(g, h, j) in Figures 4.4-4.7) show that the above relations are not satisfied for larger
Knudsen numbers, and only approximately satisfied for small Knudsen numbers. This
shows that the NSF equations, which do not account for any rarefaction effects, are not
sufficient to describe rarefied gas flow. The suitability for the NSF equations and other,
higher order, macroscopic equations to describe Couette flow will be discussed in
Chapter 5.

4.2 Numerical scheme

Since the idea and procedure of numerical scheme here is similar to the idea and
procedure of numerical scheme in Chapter 3, we only describe the difference of the
program owing to the different test situation. Therefore readers are suggested to go
through Section 3.2 before reading this section.

4.2.1 Introduction of explicit scheme

Any kinetic model equation for lDCF we consider can be expressed as

The space variable y is discretized on a uniform grid defined by nodes (centers of

L
finite volumes) y, = (i - 0.5)Ay, i = 1,. .I , Ay = - , I is the number of total position
I
points. In the following description a dense notation is used, where
f (yi,tn,c:' ,c; ,c:')= h;,.

The discretized kinetic model equation based on a finite volume scheme is

where the numerical fluxes are defined as

and the CFL condition for the time step At, becomes
a
At" =

with 0.0 < a < 1.0.

4.2.2 Reference distribution fr>,i,j (original)

The density profile at steady state is not the same as in the initial uniform flow state,
therefore u , , # 0.0 need to be applied in the process of computation so that the flow can
move along the y direction and reach steady state from the initial guess. u , , = 0.0 can be
used as a criterion to see whether the final results are results at steady state or not. It is
not a pure one-dimensional, but a two-dimensional, problem when reference distribution
fr>,i,j is solved, which is an important difference between the programs for lDCF and
1DSW. As before for lDSW, a dimensionless reference distribution (defined as Eq.
(3.15)) and the N-R algorithm are used. Expressions of Jacobians used in the N-R
algorithm for lDCF are given in Appendix B.

For the BGK model, the dimensionless reference distribution is

and the conditions to determine the four coefficients ( , riGK,i,r",,,,,i and Y",,,y.i )
are the conservation laws Eq. (1.4), which give in the discrete case
 For the ES-BGK model, the dimensionless reference distribution is
F,&, =&., .exp(-~.,,~($,, Y - r h . . , ( ~ iY~-r*,.zzi(Vij,
~ Y -G.,.i%,G + G.x,i%j, + G.y.i%z (4.9)
and the conditions to determine the seven coefficients are the discrete form of Eq. (2.12),

For the V(C)-BGK model, the dimensionless reference distribution is

F;i, = a;,i * eXP(-';,i (V:j Y+ c,.t,iV:jl + c.y,iV:j2 (4.11)

and the conditions to determine the four coefficients are the discrete form of Eq. (1.4),

For the new kinetic model I, the dimensionless reference distribution is

n
FNl,i, j -- an ~ ~ , i
exp(- T;,,~kk,,, (qtj1)2 + r.&,u,i
(tltj2 Y +'k.n.i h;j3 I+ fi.x.iq&,
)2 + +C.xy.i7):jlGj2 + J.iq;j2 1'
(4.13)
 79
and the conditions to determine the four coefficients are the discrete form of Eq. (1.4)

For the new kinetic model 11, the dimensionless reference distribution is
-
FNii,i. j -a h . i '

ex& rill,m,i (7;jI y - r i I l , n , i (7;j2 Y - r&i,.zz,i Y

(72j3 - %I ,.xy,i'?? jl 7;j2 + 't%i,x,i7~jl fifld';j2 Y
(4.15)
and the conditions to determine the seven coefficients are the discrete form of Eq. (2.39),

4.2.3 Reference distribution fr:,i, (linearized)

Only expressions of linearized dimensionless reference distribution functions are
given here, while corresponding linear equations can be found in Appendix C.

For the BGK model, the linearized dimensionless reference distribution is

FiGK,i,j
= exp(- (vi?j[al;GK,i+ a2;GK,i + a 3 ~ ,v : ji2 + a 4 ~ K .( i~ : j P )+ (4.17)
 80
-BGK model, the linearized dimensionless reference distribution is
For the V(C)

For the ES-BGK model and two new kinetic models, as in lDSW, there are two
choices to build the linearized reference distribution. One is to expand the reference
distribution around the Maxwellian distribution, and another one is to expand the
reference distribution around an anisotropic Gaussian distribution. These two choices
have been tested in this work to see which one gives the more similar results to the results
with original reference distribution.

For the ES-BGK model, the linearized dimensionless reference distribution from the
first choice is

(4.19)
and this choice makes the results of the ES-BGK model with linearized f j f i , j and b = 0

identical to results of the BGK model with linearized f : f , i , j .

For the new kinetic model I, the linearized dimensionless reference distribution from
the first choice is

For the new kinetic model 11, the linearized dimensionless reference distribution from
the first choice is
 81
For the ES-BGK model, the linearized dimensionless reference distribution from the
second choice is
FA,,, = ~ X P ( - E G . , ~ )2 -EL,,,
(~;j, h:j2 )2 - - ~ ~ . z , i Y - G , x y , i G l G, ).
h:j3

(al;, +a2is,i , ~ + a 4 k i (7Ljn)2 + G J(c2

+ ~ 3 &11;,
2
)2 + 6 s . i (?G3) +a7L,i rl:~VLj2 1'
(4.22)
where EL, are the discrete values of the dimensionless coefficient E,RT in the ES-BGK
model. After some manipulation, the expression for EL, can be obtained as

For the new kinetic model I, the linearized dimensionless reference distribution from
the second choice is

For the new kinetic model 11, the linearized dimensionless reference distribution from
the second choice is
 82
4.2.4 Boundary conditions and initial guess
For lDCF, the distribution at boundaries f_l, ( = f:j), and fL,,j (= f,",2,j ) changes

with global iteration. Maxwell's boundary conditions [I, 70]15 and a classical ghost cell
technique are used here. At the plate 1 (we can do a similar analysis for plate 2), the
distribution is

where

x;(c; ,c,h,c:)=
12
f,l,(cj ,-c, ,c,
j3). (4.27)

a,,is the accommodation coefficient, which ranges from 0 to 1, wherea,, = 0 means a

"polished" boundary, and a,, =1.0 means thermalizing boundary. The accomodation
coefficient and chosen as a constant in all iterations, f;,, is a Maxwellian at plate
temperature, where the density is chosen so as to have no accumulation of particles at the
walls. How to determine f;, will be described below.

the macroscopic velocity of plate 1, which both are constants in the whole iteration.
= Fp,,j,which means in the whole
Because these two values are constants, therefore, FPnlvj
computation, their values only need be computed one time. The three conditions to
determine the three coefficients in Eq. (4.28) are

l5 There are some more advanced boundary conditions available [I]. Since our aim is to test kinetic models,
we use this simple boundary condition.
 After F,,, are obtained,

where p,",, the density of the thermalized particles, is not a constant during the process,
therefore f,",,,is not a constant in the whole computation. The condition to determine pi,

and further f& as follows, which means no particle of fluid will go through the
boundary.

If &, = 0.0, then from Eq. (4.26), we get

which indeed satisfies the condition Eq. (4.3 1).

If a,, # 0.0 ,then from Eqs. (4.26,4.30,4.3 I), we get

J
where ~ e ~ is a~ constant
~ ~ during
, the
, computation. After p;, is obtained, f:l.B,,,j
,.cp >o

and f,ljwill be also obtained.

There are several choices for the initial guess of the distribution for lDCF, such as,
equilibrium distribution at one boundary, step function of the two distributions at
boundaries, average value of the two distributions at boundaries, distribution
corresponding to the average value of macroscopic parameters at boundaries, etc. We did
 84
some tests, and realized that the choice of initial guess, which only changes results of the
beginning several hundred iterations, does not change the final results of our
computations. We prefer to choose the last choice, for it is the guess nearest to the final
result for large Knudsen numbers and/or large plate velocity situations.

Since the total density should be conserved for IDCF, f,:are modified by a factor to
guarantee this condition every certain iteration (such as, every 100 iteration in this work).
What is done in this work is the same as in [34], which is that modified distributions
equal to the original computed distributions times a factor, and this factor equals to the
initial total mass divided by the current computed total mass.

4.3 Test examples

In the tests, several Knudsen numbers (Kn=0.025,0.1,0.5 and 1.0) have been chosen
to represent the weak, medium and strong rarefaction effect, the two limiting situations
for the molecular model, hard sphere molecules and Maxwell molecules, and three values
of plate speed, 300.0m/s, 600.0ds and 1000.0ds, are considered. Altogether, there are
twelve different test situations. Table 4.1 shows situations of numerical tests of kinetic
models and DSMC, and Table 4.2 gives the parameters, with the same meaning as in
Table 3.2, in the numerical tests of DSMC [7 11.

There are some common parameters in the numerical tests, which are, the material is
argon, the temperatures of the two plates are both 273.0 K, the velocity of plate 1 is zero,
the velociy of plate 2 is chosen as indicated in Table 4.1, the initial number density is
1AE2O l/m3, the viscosity at the reference temperature is 1.9552E-5kg/m-s,the reference
temperature is 273.0 K [71], Boltzmann's constant is 1.381E-23 J/K, Avogadro's
Number is 6.022E23 llmol, the molecular mass of argon is 39.95 glmol [69]. Mean free
path from Cercignani's definition, Eq. (3.38), is 8.833 mm.

Table 4.3 shows parameters used in the numerical tests of kinetic models. Direct
output macroscopic parameters are density p , velocity u, and u,, temperature T ,
 85
pressure p,, , p, and p,, , and heat flux q, and q , ,which are the same as the DSMC
results [69]. The test kinetic models are the BGK model, the ES-BGK model with
b = -0.5, the v(c)-BGK model, the new kinetic model I with b = -0.5, the new kinetic

model I1 with Pr = 21 3. For the reference distribution function fr6 ,what are tested is not

only the original expression (refer to Section 4.2.2), but also the two linearized
expressions (refer to Section 4.2.3). In cases 4.1-4.12 of Table 4.3, which corresponds to
the twelve basic situations in Table 4.1, all choices of kinetic model and reference
distribution are tested. In cases 4.13-4.24 of Table 4.3, which aim is to test the influence
of step of space grid (4.17-4.20), bounds of velocity grid (4.21-4.22) and step of velocity
grid (4.13-4.16, 4.23-4.24) on the computational results, all kinetic models with the
original reference distribution are tested.

Table 4.2: Some common parameters in the numerical tests of DSMC for lDCF
Parameter 1 Value 1 Parameter 1 Value
PNUM 1 5.6E14 1 SP(1,l) 1 3.7758E-10 for hard sphere; 4.80573-10 for Maxwell
molecules
MNM 2.5E5 SP(3,l) 0.5 for hard sphere; 1.0 for Maxwell molecules
DTM 3.125E-7 SP(4,l) 1.0 for hard sphere; 0.4672 for Maxwell molecules
NIS 4 SWJ) 6.63E-26
NSP 1600 MNC 2000 for Kn4.025; 500 for KwO.1; 100 for Kn=0.5; 50 for
Kn=l .O
NPS 50 MNSC 20000 for Krk0.025; 5000 for Kn=O.l; 1000 for Kn=0.5; 500
for Kn=1.O
NPT 100
 86
Table 4.3: Parameters in the numerical test of kinetic models for IDCF
Case Situ Number Ratio of Iteration number Number Bound of velocities (rnls)
atio of cells cell width of X l direction X2 (or X3)
n over mean velocities

4.4 Results and Discussion

4.4.1 Some important notes on dealing with results

In order to get converged results and do further analyses, similar steps as mentioned in
Section 3.4.1 for shock waves have been done for Couette flow, such as, graphs, which
show the output of one macroscopic parameter from different situations in one figure, and
data (e.g. average relative error) are obtained for discussion of each kind of analysis.
Here, only differences from Section 3.4.1 will be pointed out. No shifting need be done
before discussing the results since the boundaries are fixed. Since the relative error
becomes meaningless when a quantity approaches zero, average relative errors of
velocity u, , heat flux q, and q,, for which it is hard to choose a domain in which
relative error is meaningful, have not been computed in analyses. The direct output
parameters for pressure tensor are p , p,, , p, and p,, (= o,,), not o,,and o,,,which are
 87
computed from the output values of p , p,, and pZ2 with four digits accuracyI6.
Therefore, it is not surprising to see that average relative errors for a,, and a,, in the
following analyses are a somewhat higher than average relative errors for p , u,, T , p
and a,, at the same situation. One suggestion for future computation is to compute the
values of a,,and o,,directly from the distribution function in the program.

Another thing that needs to be mentioned is that if the step of space grid is not the
same for the two compared situations, some "smoothed" data will be derived from the
original data with the smaller step of space grid and used in the comparison. The
smoothing procedure means that the new value in one finite volume cell is the average of
the original values in that finite volume cell size (one new cell covers several original
cells). Of course, this smoothing procedure will introduce some numerical error, for
profiles of macroscopic parameters are not exact linear in one new finite volume cell.
Therefore, the computed average relative errors from this method cannot be used as a
very precise criterion, but a little rough criterion, in the comparison. This type of
situation, where curves on different step of space grid are compared, is met when the
effect of cell width is considered, and when the results from kinetic models are compared
to DSMC simulations.

4.4.2 Convergence of results

At the first stage of analysis (convergence of results), what we do is to compare results
among the five certain iterations (which are 115, 215, 315, 415 and 515 of total iteration).
Average relative errors of macroscopic parameters among these five certain iterations are
computed. It is found that those average values are smaller than 0.01 for almost all cases.
This shows that running so many iterations (at least 4800), values of macroscopic
parameters change are very small; therefore the computational results are converged. As
an example, Figure 4.1 shows profiles of some macroscopic parameters at five certain
iterations for case 4.18 when the BGK model is used.

l6 When the numerical experiments were all done, it was not realized that a higher accuracy would be
better. Since the computations take so long, there was no time to repeat them.
 88
The simple estimate for time to reach converged results in Section 3.4.2 of Chapter 3
is t = 9.22, where 2 is the mean free time.

For the lDCF, let us compute the mean free time using initial state, where the
viscosity is computed from Eq. (2.32) (since T =To = 273.0 K here, there is no difference
of z for hard sphere or Maxwell molecules),

1.9552x lo-'
-P -
r--- - = 3.7 x 10" (s) .
p nkT 1 . 4 ~ 1 x01.381~
~~ x 273.0
Therefore the magnitude of time to reach converged results for the lDCF from the above
simple estimation is 10"s. The magnitude of total time step in computation is lo-' - 10-'s
(smallest value is about 4 . 7 8 ~ s for cases 4.24), which is much larger than the above
estimated time, for all our computational results. Table 4.4 shows the total time step
(unit: seconds) for kinetic models in cases 4.1-4.12. Therefore, we can say that our
computational results are the converged results from the above comparison. This is the
reason why the iteration number for cases, computed a later time than others, is 24000,
instead of 40000 or 60000 (see Table 4.3).

Similar tests have been done on the DSMC computation, and results are converged.

Table 4.4: Total time step (seconds) for kinetic models in cases 4.1-4.12
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(b)
Figure 4.1: Profiles of some parameters at five certain iterations (BGK model, case 4.18: Kn=l.O, hard
sphere molecules, 300.0 m/s plate velocity) (Triangle for iteration=8000, Star for iteration=16000,
Diamond for iteration=24000, Box for iteration=32000, Triangle filled for iteration=40000)
 90

0
0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

(c)

0.2 0.4 0.6 0.8

DimensionlessDomain width y / L
(dl
Figure 4.1: Profiles of some parameters at five certain iterations (BGK model, case 4.18: Kn=l .O, hard
sphere molecules, 300.0 mls plate velocity) (Triangle for iteration=8000, Star for iteration=16000,
Diamond for iteration=24000, Box for iteration=32000,Triangle filled for iteration=40000)
 0.2 0.4 0.6 0.8
Dimensionless Domain width ylL
(4

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L

Figure 4.1: Profiles of some parameters at five certain iterations (BGK model, case 4.18: Kn=l .O, hard
sphere molecules, 300.0 mls plate velocity) (Triangle for iteration=8000, Star for iteration=16000,
Diamond for iteration=24000, Box for iteration=32000,Triangle filled for iteration=40000)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y / L
(g)

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L

O
Figure 4.1: Profiles of some parameters at five certain iterations (BGK model, case 4.18: Kn=l.O, hard
sphere molecules, 300.0 mls plate velocity) (Triangle for iteration=8000, Star for iteration=16000,
Diamond for iteration=24000, Box for iteration=32000,Triangle filled for iteration=40000)
 I

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y l L
0)
Figure 4.1: Profiles of some parameters at five certain iterations (BGK model, Case 4.18: Kn=l.O, hard
sphere molecules, 300.0 m/s plate velocity) (Triangle for iteration=8000,Star for iteration=16000,
Diamond for iteration=24000, Box for iteration=32000,Triangle filled for iteration=40000)
 Dimensionless Domain width ylL
Q

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L

(1)

Figure 4.1: Profiles of some parameters at five certain iterations (BGK model, case 4.18: Kn=l .O, hard
sphere molecules, 300.0 m/s plate velocity) (Triangle for iteration=8000, Star for iteration=16000,
Diamond for iteration=24000, Box.foriteration=32000, Triangle filled for iteration=40000)
 95
4.4.3 Discussion
From Eq. (4.2), it is realized that four parameters ( u, , p , , o12
, up,, + q, ) should be
constant (and u, = 0.0) in the whole domain at the steady state of IDCF. Therefore, we
can use these parameters to see whether the converged results are results at the steady
state or not. At the steady state, the velocity perpendicular to the plates u, should be very
small compared to the characteristic velocity m.Therefore the dimensionless velocity
YJRT will be considered. After some analysis, it is found that the average values of

vfi1 for all positions is smaller than lo4 in all test cases, and p,, q2and

u p l 2 + q, are indeed almost constant in the whole domain, and vary near the boundaries
(this jump, which is slight violation of the conservation laws due to numerical
inaccuracy, can be reduced when the step of space grid becomes smaller, which example
refer to graphs (g,j) in Figure 4.3)), which means the computational results are results at
the steady state of Couette flow. Graphs (c, f, i, 1) in Figure 4.1 can be used as examples
here.

When the computational time for each iteration is considered, the observations are
similar to 1DSW. The new kinetic model I needs the longest time, followed by the new
kinetic model 11, the v(c)-BGK model and the ES-BGK model, and the BGK model
need the shortest time. The ratio of computational time for each iteration among the new
kinetic model I, the ES-BGK model and the BGK model is about 3:2:1.

From the computational results Table 4.4, it is shown that the v(c)-BGK model has
the smallest time step At, (For situations Sa-Sd, it is only about 50% of At,, of other
kinetic models) among tested kinetic models, and At, for other kinetic models are
similar.

Results from the new kinetic model I and the new kinetic model I1 are very similar
(see Figures 4.4-4.7), and if the structure of program and computational time for one
 96
iteration is considered, the new kinetic model U is better than the new kinetic model I.
The computational time for one iteration of the new kinetic model I1 is something like
80% of the computational time need for one iteration of the new kinetic model I, the
reason is that when the reference distribution is solved (which takes a large portion of
computational time), the new kinetic model I needs results from the ES-BGK model,
while the new kinetic model I1 need not, therefore the new kinetic model I needs to
perform the N-R algorithm to find coefficients of reference distributions three times,
while the new kinetic model JJ only needs to do the N-R algorithm two times.

Let us consider the use of the linearized reference distribution f,, discussed in
Section 4.2.3. From the data and graphs obtained from the two steps mentioned in
Section 4.4.1, the following conclusions can be drawn.

For the BGK model, the average relative errors for almost all macroscopic
parameters and test situations (which are cases 4.1-4.12) are smaller than 0.001.

For the v(c)-BGK model, the average relative errors for p , u,, T , p and a,,
are smaller than 0.01 for almost all test situations, while the average relative
errors for a,,and a, are smaller than 0.02 for almost all test situations.

For the ES-BGK model, and the new kinetic model 11, there are two types of
linearized f,, one is obtained based on an isotropic Gaussian distribution,
another one is obtained based on an anisotropic Gaussian distribution. The
average relative errors for p , u,, T , p and a,,are smaller than 0.005 (0.0005)
for almost all test situations for the first (second) type of linearized f,, of these
two kinetic models, while the average relative errors for a,,and a,,are smaller
than 0.02 (0.005) for almost all test situations for the first (second) type of
linearized f, of these two kinetic models. Therefore, the second type of
 97
linearized f, is better than the first type of linearized f, for these two kinetic
models.

For the first type linearized f,, of the new kinetic model I with hard sphere

molecules, diverged results (NaNQ appear in the output of the FORTRAN

computation program) was obtained for all test situations; the reason for this is
still being considered. The first type of linearized f,, of the new kinetic model I
with Maxwell molecules works well, such as, average relative errors for p , u, ,

T , p and a,,are smaller than 0.005 for almost all test situations, and average
relative errors for o,,and a,, are smaller than 0.02 for almost all test situations.

For the second type linearized f, of the new kinetic model I, no matter with hard

sphere molecules or Maxwell molecules, it works well, such as, average relative
errors of p , u, , T , p and o,,for almost all test situations are smaller than 0.005
for hard sphere molecules (0.0005 for Maxwell molecules), and average relative
errors of o,,and o, for almost all test situations are smaller than 0.02 for hard
sphere molecules (0.003 for Maxwell molecules).

When the computational time for each time step is considered, using the
linearized f,can save 20-40% of computational time compared to using the
original f,, . This point is applicable for all kinetic models.

Let us consider the effect of step of space grid on computational results, based on the
comparison of results of cases 4.17-4.20 with results of cases 4.7,4.8,4.3 and 4.4. From
the data and graphs obtained from the two steps mentioned in Section 4.4.1, it is realized
that the average relative errors for p , u, , T and p are smaller than 0.001 for almost all
test situations and all kinetic models, while the average relative errors for o,,, a, and
a,,are smaller than 0.02 for almost all test situations and all kinetic models. Therefore,
our choices of step of space grid in Table 4.3 are acceptable.
 Let us consider the effect of bounds of velocity grid on computational results, based
on the comparison of results of cases 4.21-4.22 with results of cases 4.19-4.20. From the
data and graphs obtained from the two steps mentioned in Section 4.4.1, it is realized that
the average relative errors for p , u, , T , p and a,,are smaller than 0.002 for almost all
test situations, while the average relative errors for a,,and a,, are smaller than 0.015 for
almost all test situations. Therefore, the effect of bounds of velocity grid on results can be
neglected as long as the choice of bounds satisfied the standard given in Section 3.2.6.

Let us consider the effect of step of velocity grid on computational results, based on
the comparison of results of cases 4.13-4.16 with results of cases 4.1,4.2,4.10 and 4.12,
and also the comparison of results of cases 4.23-4.24 with results of cases 4.13 and
4.16.From the data and graphs of the above first series comparison obtained from the two
steps mentioned in Section 4.4.1, it is realized that the average relative errors for p , u, ,
T , p and o,,are smaller than 0.02 for almost all test situations and all kinetic models,
while the average relative errors for o,,and o,,sometimes are unacceptable large, such
as 0.1, sometimes even higher for the v(c)-BGK model. While the average relative errors
for all macroscopic parameters are smaller than 0.01 in the comparison of cases 4.23 and
4.24 with cases 4.13 and 4.16. Doubling the number of discrete velocities will increase
the computational time of each iteration eight times, and a larger number of discrete
velocities also imply smaller time steps by Eq. (4.4), which further means more iterations
are needed to reach converged results. Therefore, tests with smaller step of velocity grid
than the step of velocity grid shown in Table 4.3 is not suggested to be done, before some
ways are applied to decrease the computational time, such as implicit numerical scheme,
effective criterion to check whether the results are converged or not, parallel computing,
etc.

As some example figures, Figure 4.2 shows the use of two types linearized fd and
the effect of step of velocity grid in the new kinetic model I1 for situation Sa (Kn=0.025,
hard sphere molecules, 300.0 m/s plate velocity) (cases 4.1 and 4.13); Figure 4.3 shows
 99
the effect of bounds of velocity gird and the effect of step of space gird in situation Sc
( K n d . 1, hard sphere molecules, 300.0 m/s plate velocity) (cases 4.3,4.19 and 4.21) from
the ES-BGK model with original frd .
 0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L
(b)
Figure 4.2: Results to show the use of linearized f w and the effect of step of velocity grid (new kinetic
model 11, situation Sa: Kn=0.025, hard sphere molecules, 300.0 m/s plate velocity) (Triangle for original
f, in case 4.1, Triangle filled for first type linearized fmf in case 4.1, Diamond for second type linearized
f , in case 4.1, Diamond filled for original fw in case 4.13, Star for original f , in case 4.23)
 DimensionlessDornain w i d t h y / L
(c)

0.2 0.4 0.6 0.8 1

D i m e n s i o n l e s s Domain w i d t h y / L

(4
Figure 4.2: Results to show the use of linearized fwj and the effect of step of velocity grid (new kinetic
model 11, situation Sa: Kw0.025,hard sphere molecules, 300.0 mis plate velocity) (Triangle for original
frej in case 4.1, Triangle filled for first type linearized f , in case 4.1, Diamond for second type linearized

f, in case 4.1, Diamond filled for original fnl. in case 4.13, Star for original fw in case 4.23)
 0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

Figure 4.2: Results to show the use of linearized fie, and the effect of step of velocity grid (new kinetic
model 11,situation Sa: Kn=0.025, hard sphere molecules, 300.0 m/s plate velocity) (Triangle for original
f, in case 4.1, Triangle filled for first type linearized fm, in case 4.1, Diamond for second type linearized
f , in case 4.1, Diamond filled for original f , in case 4.13, Star for original f, in case 4.23)
 0.2 0.4 0.6 0.8
Dimensionless Domain width ylL
(h)
Figure 4.2:Results to show the use of linearized f, and the effect of step of velocity grid (new kinetic
model 11, situation Sa: Kn=0.025,hard sphere molecules, 300.0m/s plate velocity) (Triangle for original
fie, in case 4.1,Triangle filled for first type linearized f, in case 4.1,Diamond for second type linearized
f , in case 4.1, Diamond filled for original ,
f in case 4.13,Star for original f, in case 4.23)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y / L
(9

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(i)

Figure 4.2: Results to show the use of linearized f , and the effect of step of velocity grid (new kinetic
model 11, situation Sa: Kn=0.025,hard sphere molecules, 300.0 mls plate velocity) (Triangle for original
f , in case 4.1, Triangle filed for first type linearized fw in case 4.1, Diamond for second type linearized
f , in case 4.1, Diamond filled for original f r f l in case 4.13, Star for original fnf in case 4.23)
 0.2 0.4 0.6 0.8
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L
(b)
Figure 4.3: Results to show the effect of bounds of velocity grid and the effect of step of space grid (ES-
BGK model, situation Sc: Knd.1, hard sphere molecules, 300.0 m/s plate velocity) (Triangle for case 4.3,
Triangle filled for case 4.19, Diamond for case 4.21)
 Dimensionless Domain width y 1 L
(c)

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y l L

Figure 4.3: Results to show the effect of bounds of velocity grid and the effect of step of space grid (ES-
BGK model, situation Sc: Kn=O.l, hard sphere molecules, 300.0 m/s plate velocity) (Triangle for case 4.3,
Triangle filled for case 4.19, Diamond for case 4.21)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

(0
Figure 4.3: Results to show the effect of bounds of velocity grid and the effect of step of space grid (ES-
BGK model, situation Sc: K n 4 . 1 , hard sphere molecules, 300.0 mls plate velocity) (Triangle for case 4.3,
Triangle filled for case 4.19, Diamond for case 4.21)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

DimensionlessDomain width ylL
01)
Figure 4.3: Results to show the effect of bounds of velocity grid and the effect of step of space grid (ES-
BGK model, situation Sc: Kn=O.1, hard sphere molecules, 300.0 m/s plate velocity) (Triangle for case 4.3,
Triangle filled for case 4.19, Diamond for case 4.21)
 0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L
(9

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
ti)
Figure 4.3: Results to show the effect of bounds of velocity grid and the effect of step of space grid (ES-
BGK model, situation Sc: Kn=O.l,hard sphere molecules, 300.0 mls plate velocity) (Triangle for case 4.3,
Triangle filled for case 4.19, Diamond for case 4.21)
 110
4.4.4 Comparison with DSMC
There are too many data and graphs obtained from the numerical tests mentioned in
Section 4.4.1, so only some of them are shown in this thesis". Tables D.2 to D.4 in
Appendix D show average relative errors of macroscopic parameters from kinetic models
we used compared to DSMC in cases 4.1-4.12. Figures 4.4 to 4.7 show the results of
kinetic models and DSMC for cases 4.14,4.6,4.9 and 4.12.

First, let us consider the v(c)-BGK model. There are two main points against this
model in Couette flow. One is that for larger Knudsen numbers (e.g. Kn> OS), the shape
of the density profile is too flat compared to the shape of the density profile from the
DSMC (examples refer to graph (a) in Figures 4.5-4.7). This phenomenon was also
pointed out in [34], and it is shown that this might be due to different possibilities for
choosing the Knudsen number, but most likely it is a general failure of the v(c)-BGK
model 1341. Moreover, the result of q, is the worst among all kinetic models. For small
plate velocities (which corresponds to small Mach numbers), we even get the heat flux q,
with opposite signs, compared to values from the DSMC (examples refer to graph (h) in
Figures 4.4 and 4.7).

Then, let us consider the two new kinetic models. It is found that results from the new
kinetic model I are very similar to results from the new kinetic model I1 for all test cases.
Results from the new kinetic models locate in between results from the v(c)-BGK
model and results from the ES-BGK model for hard sphere molecules, while the new
kinetic models are identical to the ES-BGK model for Maxwellian molecules, for the
collision frequency is a constant and independent of the microscopic velocity for
Maxwellian molecules. Combining with the analysis about the v(c)-BGK model in the
previous paragraph, the ES-BGK model is a little better for tested lDCF than the new
kinetic models from results and computational time.

17
Graphs and data from these two steps are too many to be shown in this thesis. If interested in graphs and
data which are not shown, contact the author (or his supervisor) for them.
 At last, let us consider the BGK model and the ES-BGK model. For small Knudsen
number situations (Kn 5 O.l), where the Prandtl number is meaningful, the ES-BGK
model gives better results than the BGK model, and examples refer to Figure 4.4. This is
meaningful, for Pr = 1 in the BGK model, while Pr = 213 in the ES-BGK model. While
for large Knudsen number situations, it is realized that none of the kinetic models (BGK
and ES-BGK) gives better results than another one for all tests. We suggest to using the
ES-BGK model, instead of the BGK model for further consideration, especially for small
Knudsen number situations.

Let us consider the effect of Knudsen number and Mach number on the accuracy of
kinetic models, which means the difference between results from kinetic models and
results from the DSMC. One parameter to show this difference is the average relative
error, which values are shown in Table D.2-D.4 in Appendix D. When the Knudsen
number changes with fixed plate velocity, it is found that there are no big changes in the
accuracy of kinetic models. Accuracy of kinetic models at small Knudsen numbers seems
to be lower than at large Knudsen numbers, since there are strong oscillations in the
results of the DSMC. These are due to the strong stochastic noise of the DSMC method at
smaller Knudsen numbers, which can not be considered as a good benchmark in these
cases [47]. When the plate velocity is increased with fixed Knudsen number, it can be
seen that the accuracy decreases for any parameter and any kinetic model. Therefore, the
Mach number has a stronger effect on the accuracy of kinetic models than the Knudsen
number.

When the degree of moments becomes higher, the accuracy of moment from kinetic
models will be lower. As an example, differences of all,o,,, o12
and q, between
kinetic models and the DSMC are bigger than differences of p , u,, T , and p for the
same test situation.
 0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

I l , , , , , , , . ,

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L
(b>
Figure 4.4: Comparison of kinetic models with DSMC (case 4.14: Kn=0.025,Maxwell molecules, 300.0
rnls plate velocity) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4, 0.6 0.8
Dimensionless Domain width y / L
(c)

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L
(dl
Figure 4.4: Comparison of kinetic models with DSMC (case 4.14: Kn=0.025,Maxwell molecules, 300.0
mls plate velocity) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L
(e)

I . . . , . . . , i
0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

(0
Figure 4.4:Comparison of kinetic models with DSMC (case 4.14: Kn=0.025,Maxwell molecules, 300.0
m/s plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model II)
 -0.0165
0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L
O
Figure 4.4: Comparison of kinetic models with DSMC (case 4.14: Kn=0.025,Maxwell molecules, 300.0
mls plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model II)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y / L
(0

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
ti)
Figure 4.4: Comparison of kinetic models with DSMC (case 4.14: Kn=0.025,Maxwell molecules, 300.0
mls plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for V(C)-BGK Diamond filled for new kinetic model I, Stat for new kinetic model 11)
 0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L
(b>
Figure 4.5: Comparison of kinetic models with DSMC (case 4.6: Kn=0.5, Maxwell molecules, 300.0 m/s
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 I I I

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y l L
6)

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L
(dl
Figure 4.5: Comparison of kinetic models with DSMC (case 4.6:Kn=0.5, Maxwell molecules, 300.0 mls
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for V(C)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L
(e)

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L

Figure 4.5: Comparison of kinetic models with DSMC (case 4.6: Kn=0.5, Maxwell molecules, 300.0 m/s
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y / L

(g)

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(h)
Figure 4.5: Comparison of kinetic models with DSMC (case 4.6: Kn=0.5, Maxwell molecules, 300.0 m/s
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 I I , , I , I
0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L
(0

0.2 0.4 0.6 0.8

Dimensionless Domain w i d t h y / L

Figure 4.5: Comparison of kinetic models with DSMC (case 4.6: Kw0.5, Maxwell molecules, 300.0 mls
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 . . . . . . . . . . - -. . . . . . . . . . .
0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(b>
Figure 4.6: Comparison of kinetic models with DSMC (case 4.9: Kn=0.5, hard sphere molecules, 600.0
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y / L
(c)

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L
(a
Figure 4.6: Comparison of kinetic models with DSMC (case 4.9: Kn=0.5, hard sphere molecules, 600.0
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y / L
(el

Figure 4.6: Comparison of kinetic models with DSMC (case 4.9: Kn=0.5, hard sphere molecules, 600.0
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

(g)

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L
01)
Figure 4.6: Comparison of kinetic models with DSMC (case 4.9: Kn=0.5, hard sphere molecules, 600.0
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK. Diamond filled for new kinetic model I, Star for new kinetic model II)
 Dimensionless Domain width y / L
(9

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(i)

Figure 4.6: Comparison of kinetic models with DSMC (case 4.9: Kn=0.5, hard sphere molecules, 600.0
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 0.2 0.4 0.6 0.8
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L
(b)
Figure 4.7: Comparison of kinetic models with DSMC (case 4.12: Kn=0.5, Maxwell molecules, 1000.0 m/s
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model II)
 0.2 0.4 0.6 0.8
Dimensionless Domain width y / L
(dl
Figure 4.7: Comparison of kinetic models with DSMC (case 4.12: Kn=0.5, Maxwell molecules, 1000.0mls
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 Q
A'
u
P'

P'
d

0.2 0.4 0.6 0.8 1

Dimensionless Domain w i d t h y / L
(el

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(0
Figure 4.7: Comparison of kinetic models with DSMC (case 4.12: Kn=0.5, Maxwell molecules, 1000.0 mls
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model II)
 -Z
- +
- \
I
!'
I
i.
- \
1 \
Lu+~+~tw+t+-ce+~t+t-c+t~~t++t+-~te~~~
f
0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(h)
Figure 4.7: Comparison of kinetic models with DSMC (case 4.12: Kn=0.5, Maxwell molecules, 1000.0 mls
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for V ( C ) -BGK, Diamond filled for new kinetic model I, Star for new kinetic model II)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L
(0

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L
(i)

Figure 4.7: Comparison of kinetic models with DSMC (case 4.12: Kn=0.5, Maxwell molecules, 1000.0 mis
plate velocity) ) (Line without symbols for DSMC, Triangle for BGK, Triangle filled for ES-BGK,
Diamond for v(c)-BGK, Diamond filled for new kinetic model I, Star for new kinetic model 11)
 132
4.5 Conclusion
For the one-dimensional Couette flow problem, the following conclusion can be
drawn from the above discussion.

When the computational time of each iteration is considered, the new kinetic model I
needs the longest time, followed by the new kinetic model 11, the v(c)-BGK model, the
ES-BGK model and the BGK model. The time ratio of the new kinetic model I to the ES-
BGK model to the BGK model is about 3:2:1.

When comparison with DSMC is considered, the v(c)-BGK is the worst, followed by
the two new kinetic models. The ES-BGK model is better than the BGK model for small
Knudsen number situations (Kn 10.1), while it is hard to say which one of the ES-BGK
and the BGK is better for large Knudsen number situations.

Using linearized f@ instead of original f in program works well and gives almost
identical results with shorter computational time (20-40% saved) for almost all test
situations. The exception is the first type of linearized fM for the new kinetic model I
with hard sphere molecules, which becomes unstable.

The second type of linearized f is much better than the first type of linearized f,
for the ES-BGK model, and the new kinetic models I and 11.

As long as the size of finite volume cell is small enough (such as smaller than one half
of the mean free path), making it smaller does not change the results too much, but makes
the computational time for each iteration larger and requires more iterations (for smaller
time step) to reach converged results. Therefore, one suggestion is to not use a very large
number of cells.
 133
Equation (3.37) determines the bounds of velocity grid with sufficient accuracy.
Considering the computational time of each iteration and the value of time step, very
large values of bounds of velocity grid are not suggested.

Based on Eq. (3.37) to determine of step of velocity grid is acceptable enough for
consideration of density, velocity, temperature and pressure, while not enough when
trace-free pressure and heat flux are considered. Some ways to decrease computational
time, such as implicit numerical scheme, a proper criterion to check the convergence,
parallel computing, and so on, are suggested to be applied before computation with high
number of discrete velocities is done.
 134
Chapter 5 Comparison of NSF, Burnett, Grad 13 and R13s equations to
kinetic models

5.1 Introduction
From Chapters 3 and 4, one can see that when a kinetic model is used to compute a
flow problem, the number of dimension of the variables (distribution function) in the
program are at least five, three for velocity space, one for time step and one for position
at one direction. There will be three indices for position nodes in a three-dimensional
problem. Moreover, due to discretization, the actual number of is huge, e.g.
12x 11x 11x 2 = 2904 variables for each position node in the case 3.1 in Table 3.1. This
requires the memory of computer to be very large for the computation of kinetic models,
and the computational time will be very long.

Compared with computation of kinetic models, in the program of computation of

macroscopic continuum theory, the dimension of one variable will be at most four (max
three for position nodes, and one for time step), and the number of variables for each
position node, which is the number of moments considered, is not large, e.g. five
variables are considered in the NSF and Burnett equations, and 13 variables are
considered in the Grad 13 equations, and their regularization (R13).

Therefore, a macroscopic continuum theory is preferred to be applied to solve rarefied

gas flow problems, instead of microscopic theory (such as DSMC, kinetic models), in
order to save computational time and simplify the computational program. In the scope of
macroscopic continuum theory of rarefied gas dynamics, there are the NSF equations, the
Burnett equations from the C-E expansion method [ll], while also there are the Gradl3,
the R13s (original and slightly linearized) from the moment method or related methods
[15,26]. It must be asked which set of macroscopic equations (NSF, Burnett, Gradl3, or
R13s) can give a good description of the flow, e.g. in relation to range of Knudsen
number. To contribute to the answer on this question is the goal of this part of the work.
In this chapter, the above sets of macroscopic equations from kinetic models will be
compared with the kinetic models themselves. The kinetic models considered are the
 135
BGK model and the ES-BGK model. Since the only difference between results for the
1
BGK model and results for the ES-BGK model is that the Prandtl number Pr = -, and
1-b
b = 0 for the BGK model, and [- 0.5,1.0] as the range of b for the ES-BGK model,
therefore in the following description, we will focus on the ES-BGK model, and deal
with the BGK model is a specific case for the ES-BGK model.

5.2 NSF and Burnett equations for BGK and ES-BGK models

5.2.1 NSF and Burnett equations in three dimen~ions'~

The aim of the C-E method [3, 5, 91 is to obtain expressions for and qi from
kinetic theory. The distribution function f in the form of an expansion in powers of the
small parameter Kn (Eq. (1.6)) is looked for in the C-E method. In this method, a factor
YKnis inserted in front of the collision term, that is to say, the ES-BGK model to be
considered is

where the reference distribution f, is defined in Eq. (2.3). Substituting the expression of
fin Eq. (1.6), rewrite as follows

into Eq. (5.1), and equating the coefficients of the same power of Kn yields a set of
equations for f '"',By solving this set of equations, an asymptotic solution of the
distribution functionf is obtained (note that Kn=l is set in the ultimate expression).

According to the above ordering hypothesis, the trace-free pressure tensor aijand heat

flux q, can also be written as a series

'*This section and Section 5.2.2 are the main part of the paper "Burnett equations for the ellipsoidal
statistical BGK model", by Zheng and Struchtrup [ll]
The conserved quantities p , ui, T, however, are not expanded. That is in the C-E
expansion one assumes that each order of the expansion gives p , ui,T, which is
tantamount to

b, pu, , 3 p ] = j{l, c, , c 2 } f d c =j{l, c, , ~ ' } f ( ~ ',d c

and imples the so-called compatibility conditions

j{l, { 0 0, ) f o r a l l n > l .
c , , ~ ~ } f ( ~ ) d c =0,

Note that, because of ps = p6, + 5,it follows that ':P = 0:"' for all n 2 1.

Since the reference distribution function f , is expressed in terms of the pressure

tensor in the ES-BGK model, f , must also be expressed in the form of powers of Kn in
the expansion

This expression for fig) will be derived later in this section.

Similar to the distribution function, time derivatives of density p, macroscopic flow

velocity ui and temperature T are likewise expressed as a series of operators,

where U,= p, ui,T , (r=1.. .5). Put Eqs. (5.2,5.5) into the conservation law Eq. (1.5), and
equating the coefficients of the same power of Kn yields the following expressions.
where m 2 1.

The next assumption of the C-E method is that the distribution function f depends on
position xi and time t only through the hydrodynamic variables U, (= p,ui,T,
(r-1 ...5))and their space derivatives. That is to say, f'"' = ~'"'(u,,vu,, ...,vnuGi),
~ r k 2 1.
but f '"Iis not a function of V n + k with

At last, based on the above line of arguments, the sequence of equations for f '")can
be written as

where
 ak k m u r ) = Vm(%) , since the position xi and time t are independent to each
at

other.

In order to obtain the expression for f '*'and the Burnett equations, we need to know
f&'up to the second order of Kn .From the definition of the matrix in Eq. (2.10) one
obtains

It follows for the expansion of the Gaussian

f, = fi:) + f ; ' ~ n + f g ' ~ n ' + C ( K ~ ) ),

where

Here we have anticipated that 010' = 0, as will be shown below. fM is the local
Maxwellian distribution.
 139
First, let us consider the order of C(yKn)
in Eq. (5.7), which yields the constitutive

relations for the Euler equations,

f (O) = fg)= fM ,

pi' = p s i ,op= O , q,(O) = 0 .

Next, we consider the order ah0)in Eq. (5.7), which yields

where

After some manipulation, one obtains

The constitutive relations in Eq. (5.14) are the laws of Navier-Stokes and Fourier where

p=-- is the viscosity, and K =-- Rp is the thermal conductivity. It follows that
1-bv, 2 VB
the Prandtl number is related to the coefficient b of the ES-BGK model by the relation
 140
By means of the Navier-Stokes and Fourier laws Eq. (5.14), the distribution can be
rewritten in a form that is familar from the Chapman-Enskog expansion of the Boltzmann
equation of Maxwell molecules, viz.

It is straightforward to show that this expression indeed satisfies the compatibility

relations (5.3). The relations between viscosities and collision frequency in Eq. (5.14) and
Eq. (2.32) can be used to obtain an explicit relation between the latter and the variables
p , T which reads

RT,"
where v,, = is a constant.
(1 - b)po

At last, let us consider the order C(Kn) in Eq. (5.7), which yields

where

ap)
-=-.- a p au, + a p a(vu,)
a3 au, ax, a(vu,)' ax,
After a tedious manipulation, the expression of f ' 2 ) is obtainec

where
 +--- 2 au<, au, + z(3.5-O) c,c, a~,, au,
---- c ~ , --- , a ~ , x i c j --
auk au, auk
I -b a ~ , ,ax, 3(1- b ) RT axI , a ~ , (I - ~ ) R Tax, ax, (1 - ~ ) R Ta~,, ax,

cic2a T a u j +-
---- 2-b , ca~,,
~ ~ c , c-------
aT ~ s ( ~ - o ) R ~ dT
T
R T ~ax, axi 2(i-b) R'T~ axtax,, 2 T ax, ax,

+--CiCjC4 aT aT
~ R ax, ~ ax,
T ' ~

It is straightfonvard, but cumbersome, to show that this expression indeed satisfies the
compatibility conditions (5.3).

Pressure tensor and heat flux at the second order, the Burnett order, are computed as

where
The above expressions for 0:' and q!2' are irreducible forms in terms of the gradients of
density, velocity, and temperature.

For the special choice b = 0 , the ES-BGK model reduces to the standard BGK model,
and the corresponding Burnett expressions for 0:' and q!Z' read

where

The Burnett equations for the BGK model are available in the literature, see [72-741.
In [72, 731, only hard sphere molecules ( w = 11 2) were considered, and our results agree
with those in [72,73], when we set 0 = 112. The Burnett BGK equations reported in [74]
are questionable: For example, since is a trace-free tensor, one finds that

a,/a8= -2 should hold in Eq. (I 1) of [74], but the authors report a,/a8= 3.5 instead.
Similar problems arise with other coefficients. Thus, it seems that the equations in [74]
are not correct.
 143
In order to compare the Burnett equations for the ES-BGK model with the Burnett
equations for the Boltzmann equation, e.g. see [12, 13, 241, we rewrite Eqs. (5.19-5.20)
as

where the so-called Burnett coefficients Cir, and 8, are given by

In particular we are interested in the case where the Prandtl number has its proper value
Pr = 2/3 which is the case for b = -l/2. Then, the coefficients ma and 8, have the
values

For the special case of Maxwell molecules, where 8 = 1 , the above expressions Eqs.
(5.23, 5.25) are identical to the expressions given in [13]. For any molecule model with
power interaction potentials, the above expressions are identical to the expressions given
 144
in [12]19, which are the Bumett equations of the Boltzmann equation for a first order
solution of the resulting integral equation in terms of Sonine polynomials.

Table 5.1 shows the values of the Burnett coefficients for different values of n (or w )
compared to values obtained by Reinecke and Kremer [24] from high accuracy
computations for the full Boltzmann equation, i.e. a fifth-order expansion in Sonine
polynomials. Note that the expansion in Sonine polynomials is necessary to solve the
integral equations that appear in the Chapman-Enskog expansion of the full Boltzmann
equation, so that in this case the Chapman-Enskog expansion of a given order can only be
computed in an approximate manner. It must be noted here, that the fifth order expansion
yields very accurate results, so that the results in [24] can be considered as almost exact.

For the ES-BGK model this additional difficulty does not arise due to the simpler
collision term, and the Chapman-Enskog expansion of a given order can be computed
exactly, without further approximations. However, we emphasize that the ES-BGK
collision term itself is an approximation of the Boltzmann collision term and does not
preserve the full physics of the Boltzmann equation.

Indeed, from the values in Table 5.1 it becomes apparent that the Bumett equations
from the ES-BGK model agree with the Burnett equations from the Boltzmann equation
only for Maxwell molecules ( w = l ) . For other molecule models, the coefficients are
close to those of the full Boltzmann equation, but different nevertheless. Thus we can
state that the ES-BGK model agrees with the Boltzmann equation for Maxwell molecules
up to second order in the Knudsen number, while for more realistic molecule models (e.g.
o = 415 ), some differences occur.

l9 Take into account that, by Eq. (2.27), = a,.

P dT
 145
Table 5.1: Burnett coefficients tiT, to tiT6 and 8, to e5 for different power index n and

n @
4 92 83 84 8s
7.518 4518 3 3 3914
5 1
(798) (4518) (3) (3) (3914)
10.13 5.625 3 3 9.55
2313 415
(10.159655) (5.655703) (3.014430) (2.760746) (9.018759)
11.25 5.625 3 3 9.25
00 112
(11.652480) (5.826240) (3.095605) (2.415493) (8.100700)
1
NOTES: The values in brackets are those from the full Boltzmann equation [24] where values for n=23/3
are interpolated.

As a conclusion of this section, the Burnett equations for the ES-BGK model with
power interaction potentials have been computed by the Chapman-Enskog method to the
second order. When the Prandtl number is adjusted to its proper value Pr = 213
(b = -1/2), the ES-BGK Burnett equations are found to be identical to the Burnett
equations for the Boltzmann equation only in the case of Maxwell molecules, while the
Burnett coefficients exhibit some differences for other interaction types (e.g. n = 2313 or
n = 00 - hard spheres). This indicates that computations of processes with the ES-BGK
equation will show some discrepancies to computations that are based on the Boltzmann
equation. For processes at not too large Knudsen numbers these differences might be
small. Since our analysis is based on an expansion in terms of the Knudsen number only
 146
to second order, we can not make any prediction as of how close the ES-BGK model will
be to the Boltzmann equation, when the Knudsen number is large.

5.2.2 Linear stability analysis of the Burnett equations for the ES-BGK model

The main point of critizism on the Burnett equations is the question of linear stability.
It is well known [75] that the Burnett equations for the Boltzmann equation with Maxwell
molecules are linearly unstable in transient processes. The ES-BGK model gives us some
freedom to adjust the Prandtl number, and we reconsider the stability of the Burnett
equations here, and will show that for some Prandtl numbers, the ES-BGK-Burnett
equations are stable. Thus, in this section, we consider the linear stability of the Burnett
equations for the ES-BGK model where we restrict ourselves to one-dimensional
processes. We follow the procedure for stability analysis outlined in [75].

We are interested only in processes with small deviations from an equilibrium state
po,To, and % , = 0.Then, we can introduce dimensionless variables P , f , rii, &u and
4i by

The new variables measure the deviation from a global equilibrium state, an&in the
linear c a s e a r e much smaller than unity.

Moreover, we use the mean free path I to introduce non-dimensional space and time
variables ii and t^ as

Linearizing Eqs. (5.19, 5.20, 1.5) in the new variables yields the dimensionless
linearized three dimensional Burnett equations as
Note that the value of the viscosity exponent w plays no role in the linear equations. The
hats will be omitted in the sequel in order to simplify the notation.

For the case of one-dimensional processes where all variables depend only on x, = x ,
and where ui = {u,o,o), the above equations reduce further to

We now assume plane wave solutions of the form

p=fi.exp(h+ikx), ~ = F . e x ~ ( h + i k x u) ,= i i . e x p ( h + i h )
where k is the positive real wave number, while A is a complex number, whose real part
can be interpreted as the negative of the damping of the wave, and whose imaginary part
is the frequency of the wave. Stability of the system requires negative damping, so that
ReA 5 0 must hold for all possible values of k.

Substitution of these solutions into Eqs. (5.27) yields a homogeneous algebraic system
for the amplitudes fi , ii ,and T which reads
A-fi+ik.ii=O,

(1 - b)(l-Sb) 3 5(1-b)k2
. u + -a+-
(ik +
3 ik3) - 2( 2
 148
The condition for existence of a nontrivial solution of above system is that the
determinant of the system equals to zero, that is

Note that the ES-BGK model requires that - 112 5 b 5 1. We will divide the whole range
of b into three parts to discuss the roots of the above cubic equation for A, and the linear
stability of one-dimensional processes.

P a r t I ( - l / 2 5 b < O and l / 5 < b c l )

We follow the line of arguments of [66].For values of b in this range, it is easy to see
that SZ(A = 0, k ) > 0 and SZ(A+ m, k ) + +.o while for very large values of k

holds. Therefore, Eq. (5.29) has at least two positive real roots A . It follows that the
Burnett equations for the ES-BGK model are linearly unstable for values of b in the
intervals-1/25b<O and 1 / 5 c b c 1 .

Part I1 (b = 1)
In this particular case, Eq. (5.29) simplifies to

The three roots for this cubic equation are

Therefore, Re&,,,, 5 0 , and the Burnett equations for the ES-BGK model are linearly
stable for b = 1. Note that by this case corresponds to a gas with infinite viscosity and
Pr = co
 A cubic equation has either three real roots, or one real root and two conjugate
complex roots [ X I . We shall analyze the roots of Eq. (5.29) for 0 I b I 1/5 in the
following.

Three real roots

Assume that the three roots of Eq. (5.29) are all real. Since the function o ( A , k ) (the
left hand side of Eq. (5.29)) increases when A increases ( A 2 0 assumed), and
Sl(R = 0,k ) > 0,it follows that Eq. (5.29) cannot have positive roots, and that all three
real roots (so they exist) will be negative.

One real root and two conjugate complex roots

In this case, we assume that the three roots of Eq. (5.29) are one real value A, and two
conjugate complex values A,, = AR + iAi ,where AR and Ai themselves are real numbers.
Since for A 2 0 the function GI(/Z,k) increases when A increases, and since
C@ = 0, k ) > 0, A, must be negative. Furthermore, from Eq. (5.29), we obtain by
inserting A = il,+ i&

Eliminating Ai between these two equations yields

where the k-dependent coefficients are given by

 b(9 -5b)(l- b)(l-56) 123- 1636 + 1056' - 25b3 15 - 56 k 4
k6 +-
+

J3 =
81 108 36
Since the coefficients J, , c2,and J3 are positive for 0 I b 1115, there are no positive
roots of Eq. (5.3 1) and it follows that Re il I 0.

Thus, Re /Z I 0 holds for all values of b in the interval [O, 1/51 and we conclude that
the Burnett equations for the ES-BGK model are linearly stable for 0 Ib 1115, that is for
Prandtl numbers in 1 I Pr I 514. For all other values of b (or Pr) except Pr = -, however,
the Burnett equations for the ES-BGK model are linearly unstable - this includes the
physically relevant case of Pr = 2/3.

The stability of the BGK-Burnett equationkwhere P-1---can be found in the

literature [77]. Our anlysis shows that the stability can be obtained for a wider range of
values for the Prandtl number.

As a conclusion of this section, our analysis of the linear stability of the ES-BGK

For 1 I Pr I 514 and for Pr = -

Burnett equations revealed that the stability depends on the value of the Prandtl number.
the ES-BGK Burnett equations are linearly stable, while
they are linearly unstable 213 < Pr < 1 and 514 <Pr < -. It follows that higher order
Chapman-Enskog expansions do not necessarily lead to unstable equations.

5.2.3 NSF and Burnett equations for one-dimensional Couette flow at steady state
In the lDCF at steady state we consider (one plate is stationary, while another plate is
moving along X1 direction, and the direction perpendicular to the plates is the X2
direction), unknown variables in trace-free pressure and heat flux are o,,, 02,, o,,, q, ,
 151
+ 0 , ), o,, = o,,, 03,= o13= 0,
and q, , while o,, = -(o1, 03,= = 0 , and u, = 0,

From Eq. (5.14), after some manipulation, one obtains the following expressions for
the NSF equations

From Eq. (5.19), after some manipulation, one obtains the following expressions for
the Burnett equations,

where
 From the above Eqs. (5.32-5.34), one can see that the expressions for o,, and 9,
from the NSF equations and the Burnett equations are the same, while the expressions for
a,,, o, and q, from the NSF equations and the Burnett equations are different. Non-
zero values of a,,, a,, and q, refer to rarefaction effects which are not described by the
NSF equations, but by the Burnett equations, and the moment equations which will be
discussed now.

5.3 Grad13 and R13s equations for BGK and ES-BGK models

5.3.1 Grad13 and R13s in three dimensions

There are 13 unknown variables, and 13 moment equations in the Grad 13 and the
R13s for a three-dimensional problem. The unknown parameters are p , u, , u2, u, , T ,
o,,, oZ, o,,
, o,,, 023,
q, , q2 and q3, while other parameters can be derived from
these parameters or equal to zero, such as, p = pRT, =- ( + 0 2 )02,= OI2,

The Grad13 and the original R13 for the kinetic model (BGK or ES-BGK) are
obtained from the same idea and procedure as for the Boltzmann equation with the
Maxwellian molecular model, which are given in [26].

After multiplying the kinetic model equation Eq. (2.8) by polynomials of the peculiar
1
velocity, viz. 1, C, , C2, C,,Cj, and -C2ci, and subsequent integration over velocity
2
space, the basic 13 moment equations for the kinetic model are obtained as
where the definition of moments refer to Eq. (1.3), and Eqs. (2.12, E.16) have been used
to compute the right hand sides of the last two expressions in Eq. (5.35). Note that this set
of 13 equations does not form a closed set of equations for the 13 variables, since it
contains the higher degree moments p,,, , p,,,,,and p, .

There are some rules about the computation of trace free part of a tensor [78].

1
A($, = + 4 + Ajki+ Ajik+ AkV+ Akji).
Therefore, for three-degree and fourth-degree moments, we have
2 1
Pijk = p<ijk>+j(qi6jk+ q j s i k +qk%). Prrik = Prr<ik> +- ~rrss'ik ' (5.37)
3

In order to close the system for the 13 variables, a closure assumption is required that
allows to relate the higher degree moments to the 13 variables, which requires to express
the distribution function as a function of the 13 variables. As mentioned in Chapter 1,
there are two ways to find the closure assumption, and here Grad's moment method is
considered. In Grad's moment method, the distribution is related to the moments as an
expansion in Hermite polynomials about the equilibrium distribution [5,6,26],
where N is the number of moments considered in the system, YA ( A = 1,...,N ) means a
set of N polynomials of the peculiar velocity, which corresponding macroscopic
parameters are the N moments p, considered, that is to say,

PA= j%fiNc (5.39)

and the coefficients A, are determined from the N moments considered. In particular, for
1
the Grad's 13 moment equations, N=13, Y, =( 1, C,, c2,C,,C,, - c 2 c i , and pAare
2
the 13 moments p , ui , T , a, and qi, and one has

From this expression for the distribution and Eqs. (1.3,5.37), we have
GI3
= 0.0, G13
p<m>u=7RToU, /YE:: = 15-P (5.41)
P

Grad's 13 moment equations form a hyperbolic system, which implies finite wave
speeds and discontinuities in shocks that make the system difficult to handle [26]. In
order to overcome these disadvantages of the Grad's moment system, Struchtrup and
Torrilhon developed a method, named as regularization, in which some parabolic terms
are added, which changes the character of the equations so that no discontinuities will
occur [26, 271. The additional terms are obtained from the Grad's moment equations for
higher degree moments (than moments considered). In their work, the Boltzmann
equation with Maxwell molecules is considered, while here the ES-BGK model (which
includes the BGK model as a specific case) with Maxwell molecules will be considered
with the same idea and procedure. Since we are interested in 13 moment equations, we
will describe the idea and procedure based on this situation.

In order to perform the regularization, we define the derivations of p,ij,>, and

p, from their values obtained with the Grad 13 closure as

For the Grad 13 moment equations, we thus have

vk
= 4;13 = AG13-
-0.

The regularization of Grad's 13 moment equations is based on the Grad's moment

equations for the three-degree and fourth-degree moments p,*, , prr+ and p, ,which
have, for the ES-BGK model, the form

In the derivation of the above equation (5.43, the following expression for f, (Eqs.
(E. 16, E.24)) and Eq. (2.12) have been used,
Ic,c,c,f,dc = 0,

After some manipulation, one obtains the equations for m f l , Rij and A from Eqs.
(5.35,5.42,5.44) as
 Then at last, computing the left hand side terms in the above equation (5.46) with 13
moment distribution, one obtains the mvk, Rv and A of the original R13 as

1
The idea of last step is a Chapman-Enskog like procedure [26], which is adding -
E

(which will be set equal to one at the end of the calculations) on the right hand side of Eq.
(5.46) (not for Eq. (5.35)), and expanding moments (only for m, , R, and A ,not for the
13 variables) as
Then, it is straightforward to see that the zeroth-order approximation (balancing terms
1
with the factor -) just gives Grad's 13 moment case (Eq. (5.43)), and the first-order
E

correction results (balancing terms with the factor EO) gives the Regularized 13 moment
case (Eq. (5.47)).

The slightly linearized R13 equations is obtained from a new method [15], which is
based on a different way to account orders in the Kn number, and provides a direct link
between Grad type moment equations and the Kn number. The expression of muk, Rii
and A of the slightly linearized R13 equations are found to be the original R13 equations
(5.47) omitting the highest order terms, which are underlined in Eq. (5.47).

5.3.2 Grad13 and R13s in one-dimensional Couette flow at steady state

In the lDCF we consider, there are nine unknown variables in the moment equations,
which are p , u,, u, , T , a,,, a;,, a,,, q, ,and q, ,and other parameters can be derived
from these parameters or are zero value, p = pRT, u3 = 0 , a,, = -(o,,+ 02,),
d d
a,, = 0,,, 03,= o13
= 0, = 0 , q3 = 0 , -= -= 0 . Here only the
c3,= oZ3
ax, ax,
equations for lDCF are given, and their derivations from the general equations (refer to
corresponding equations in the above Section 5.3.1) has been omitted.

The nine basic moment equations, which are the same for the Grad 13, the R13s, and
kinetic models (BGK and ES-BGK), are
 Note that the above equations do not form a closed set of equations for the nine
variables, since they contain the higher degree moments p,,,,,, p,,,,,, p,, , p m<,2,,
P,,~, and p, . The difference between the Grad13 equations, and their regularization
comes in through the expressions for the higher degree moments, which can be
decmposed according to Eq. (5.42),
P C 1 12> = mll2 9 P<122>= m122 9 P<222>= m222 3 (5.49)
 For the Grad 13 moment equations, from Eq. (5.43), one gets
GI3 - GI3 G13 - GI3 = R2?3 = ~ --0.
G l 3
m112 -m122 = m 2 2 2 -4 2

For the original R13 moment equations, from Eq. (5.47)

Eq. (5.51) omitting the highest order terms (terms with underlined in equations) are
the expressions for the slightly linearized R13 moment equations [IS],
 From Eq. (5.48), one obtains the following equations for the steady state of IDCF,

If one uses the computational results from kinetic models for all moments, that is to
say, without considering the closure relations (5.49-5.5 I), the above Eq. (5.52) should be
satisfied. Indeed, the first four expressions in Eq. (5.52), that u,, o,, , p,, , and
4, + u p l 2 are constant in the whole domain at steady state, have been applied in Chapter
4 already, are used also here to check whether the converged computational results are at
 161
steady state. The verification of the last five expressions in Eq. (5.52) is difficult, since
the computation is related to the calculation of derivatives. If a proper expression for the
derivative is chosen, the last five expressions in Eq. (5.52) are found to be satisfied from
the results of kinetic models as will be shown in Section 5.5.

Expressions to compute trace-free pressure tensor oijand heat flux q,, which are
values at the right hand side of Eq (5.52), are rewritten from Eq. (5.52) for lDCF at
steady state as

The test of the different macroscopic equations will be done as follows: The closure
relations from the Grad13 equations (5.49-5.50), or from the R13s equations (5.49,5.51),
are used to relate the third and fourth degree moments to the lower degree moments.
Results from the kinetic models for first and second degree moments (such as density,
heat flux, etc.) are then used in the right hand side of Eq. (5.53). This gives values for
a and qM ,i.e. the left hand side of Eq. (5.53). The values computed this way will not
:
necessarily be equal to those computed from the kinetic models. From the difference, one
 162
can infer which set moment equations fits the kinetic model best. A similar method is
used for the NSF equations and the Burnett equations that are presented in Section 5.2.

5.4 The Knudsen layer

At small Kn numbers, the behavior of a gas is conveniently expressed in the sum of two terms, which
are the bulk part (or fluid-dynamic part, whose length scale of variation is of the order of geometry length,
describes the overall behavior of the gas) and the boundary layer part, or Knudsen layer part, which is
appreciable only in a thin layer with thickness of the order of the mean free path adjacent to the boundary,
and represents a correction to the bulk part. Each term can be expanded in a power series of the Knudsen
number [I, 5,8,9,79].

The NSF equations are the equations at the first order of Kn, the Burnett equations are obtained at the
second order of Kn, the Grad13 equations are equations between second and third order of Kn, the slightly
linearized R13 equations are equations at the third order of Kn, and the original R13 are equations between
third and fourth order of Kn [IS]. Therefore, it will be expected that at small Knudsen numbers the R13s
give the best results, followed by the Gradl3, the Bumett equations, and that the NSF equations will give
the worst results. This expectation will be seen to be true from the results in Section 5.7.2. In [lo],
Struchtrup gives detailed discussion about the Knudsen layer for the Bumett and super-Bumett equations
and the original R13 equations, and shows that the boundary layer of the heat flux q, (parallel to the
plates) from the original R13 equations has a different sign than that from the Bumett and super-Bumett
equations (this issue is also shown in Section 5.7.2, such as graph (c) in Figure 5.2). Also in Ref. [lo], the

shape of the q, boundary layer of the R13 equations is found to be A sinh

[ )
- 0a5 , A denotes
Kn
the amplitude, x is the position considered, and L is the domain width.

Since macroscopic continuum equations describe the bulk part better than the
Knudsen layer part, it makes sense that the accuracy of macroscopic continuum equations
in the middle part of the flow domain is higher than near the boundary. When the
Knudsen number increases, the thickness of the boundary layer, which is about several
mean free paths, also increases, and the bulk part becomes thinner and thinner. and the
effect is lower accuracy of macroscopic continuum equations. Finally, the boundary layer
extends over the full domain, and no bulk part in the domain.

5.5 Description about the test method

 163
Since the boundary conditions for macroscopic continuum equations (except the NSF
equations) are still developing and non-mature [15,26,49], a specific way will be used in
the tests, in order to avoid the use of boundary conditions.

The test of the different macroscopic equations is done as follows: The closure
relations from the Grad13 equations (5.49-5.50), or from the R13s equations (5.49, 5.51),
are used to relate the third and fourth degree moments to the lower degree moments.
Results from the kinetic models for 13 moments (i.e. p , ui , T , o, and qi ) are then used

in the right hand side of Eq. (5.53). This gives values for 0; and q" , i.e. the left hand
side of Eq. (5.53). The values computed this way will not necessarily be equal to those
computed from the kinetic models. From the difference, one can infer which set moment
equations fits the kinetic model best. A similar method is used for the NSF equations and
the Burnett equations in the Section 5.2, Eqs. (5.32-5.34). Some third-degree and fourth-
degree moments will be treated similarly, in case of Grad13 and the R13s equations.

5.6 Test examples

The numerical test problem is one-dimensional Couette flow (IDCF) at steady state
and kinetic models used are the BGK and the ES-BGK models with Maxwellian
molecules, as described in Chapter 4. Table 5.2 shows the situations for the numerical
tests used in this chapter, and Table 5.3 shows the parameters used in the numerical tests
of kinetic models, while common parameters in the tests are the same as in Section 4.3.
Some analyses, same as done in Chapter 4, has been done to show that the final
computational results are converged and correspond to steady state. This time, the direct
output macroscopic parameters are p , u,, u,, T , p , p,,, p,, , q,, q,, and the
o,,
third-degree and fourth-degree moments p,,,, p,,,, p,,, pm,,, pm2,and prrs. Other
interesting parameters can be calculated from these.

Table 5.2: Situations of numerical tests of kinetic models for lDCF in chapter 5
Situation Knudsen Plate velocity Domain Mach Reynolds
number (mls) width (m) number number
Sb 0.025 300.0 0.3533 0.975 50.345
Sd 0.1 9M.O 008833 0.974 12.587
 ible 5.3: Parameters in the numerical tests of kinetic models for lDCF in chapter 5
Cas Situa Numb Number of Ratio of cell Number of Bound of velocities (mls)
e tion er of iteration width over velocities Xl direction X2 (or X3)
- cells mean free path
5.1 200 60000 5.0 11*10*10 - 1 100.0, -1 100.0,
- 1400.0 1 100.0
5.2 -1 100.0, -1 100.0,
-
5.3
1400.0
-1 100.0,
1 100.0
-1 100.0.
1300.0 1100.0
5.4 - 1 100.0, -1 100.0,
1300.0 1100.0
5.5 -1 100.0, -1300.0,
1700.0 1300.0
5.6 -1200.0, -1600.0,
2200.0 1600.0
5.7 -1 100.0, -1200.0,
-
5.8
1700.0
- 1 100.0,
1200.0
-1400.0,
- 2100.0 1400.0

5.7 Results and discussion

5.7.1 Some important notes on dealing with data

Four choices to compute the derivative have been used to compute trace-free pressure
tensor avand heat flux q, based on Eq. (5.53) and results of kinetic models (BGK or ES-
BGK) for all moments at the right hand side of Eq. (5.53): three point formula (central
difference) in [a]five
, point formula in [a]and
, the simplest and best expression
mentioned in [80]

Let us consider the function f = f (x) ,x is discretized as xi, where i = 1,2,..., and the
df = f ', and the second order
step is Ax, the first order derivative is denoted as -
dx
d 2 f = f ". Then the three-point formula in [64] is
derivative is denoted as -
dx2
The five-point formula in [64] gives

The simplest expression in [80]gives

The best choice in [80]gives

where the definition of the minmod function refer to Section 3.2.1.

After some numerical tests, it is found that using the central difference formula (5.54)
to compute the first order derivatives gives the computed values of o; and q: closest to

the results of ovand q, from the same kinetic model directly.

Another thing that needs be mentioned is that if we do the above computation directly,
the computed ooand qi have some small oscillations and a jump near the boundaries,
see Figure 5.1. In order to reduce and eliminate that oscillation and jump phenomena, we
smoothed the original results from kinetic models by half position points, so that values
at new positions are the average values of two adjacent old position points, and then
apply these smoothed data, instead of original results, in the computation. It is found that
: and qf based on the smoothed data
the oscillation phenomenon of the computed a
decreases a lot, while at the same time, the profile from smoothed data is the same as the
profile from those original results from kinetic models.

Therefore, two points, central differences for derivatives and smoothed original data,
are applied in the following computations. Figure 5.1 shows an example (case 5.2:
 166
Kn=O.1, 300.0 m/s plate velocity; BGK model) of the original and computed :
a and qF
without and with the smoothness procedure. Note that since the profile of a,,is similar
to the profile of -a,,, and no new information can be obtained from graphs of a, ,
graphs of a, are omitted in the figures in this chapter. Tables D.5-D.6 in Appendix D
show average relative errors of the computed ovcompared to (smoothed) original data,
which are smaller than 0.01 for all test cases. For relative error is meaningless when a
quantity approaches zero, and values of q, are near to zero in the middle domain, thus it
is hard to choose a domain to compute relative error for qi . Therefore for the heat flux
q, ,the average relative errors are not computed

Normally, performances shown by the ES-BGK model and its corresponding sets of
macroscopic equations are also shown by the BGK model and its corresponding sets of
macroscopic equations. Therefore, in the following discussion, we will mention kinetic
model generally, and not specifically mention the BGK model or the ES-BGK model.
 0.2 0.4 0.6 0.8
Dimensionless Domain width y / L

. A -

0.2 0.4 0.6 0.8 1

DimensionlessDomain width y / L

Figure 5.1: Comparison of original and computed trace-free pressure and heat flux (BGK model, case 5.2:
Kn9.1, 300.0 rnls plate velocity) (Triangle for original results, Triangle filled for computed results from
Eq. (5.53), Diamond for original results smoothed by half number position points, Diamond filled for
computed results from Eq. (5.53) with smoothed original data)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y/L

0.2 0.4 0.6 0.8

Dimensionless Domain width ylL

(dl
Figure 5.1: Comparison of original and computed trace-free pressure and heat flux (BGK model, case 5.2:
KnlO.1, 300.0 m/s plate velocity) (Triangle for original results, Triangle filled for computed results from
Eq. (5.53), Diamond for original results smoothed by half number position points, Diamond filled for
computed results from Eq.(5.53) with smoothed original data)
 169
5.7.2 Results and discussion of a, and qi
Figures 5.2 to 5.5 show the (smoothed) original o,j and qi, and the computed a, and
qi based on the NSF equations (5.32), the Burnett equations (5.33-5.34), the Grad13
equations (5.49, 5.50, 5.53), the original and slightly linearized R13 equations (5.49,
5.51, 5.53), and the kinetic model itself, Eq. (5.53), with all (smoothed) original data at
the right hand side. The test cases considered are case 5.1 (Kn=0.025, 300.0 m/s plate
velocity) with the BGK model, case 5.7 (Kn=O.1,600.0 m/s plate velocity) with the BGK
model, case 5.6 (Kn=0.5, 1000.0 m/s plate velocity) with the ES-BGK model, and case
5.4 (Kn=l.O, 300.0 m/s plate velocity) with the ES-BGK model. Tables D.5-D.6 in
Appendix D show the average relative errors of a, between the above sets of
macroscopic continuum equations and the kinetic model.

Generally, the computed values of a, and qi from all sets of macroscopic equations
fit the original data well at small plate velocities (i.e. small Mach numbers) and small
Knudsen numbers, and not so well at large plate velocities orland large Knudsen
numbers. In case 5.1 (Kn=0.025, 300.0 m/s plate velocity, Figure 5.2), the computed
values of o,,, o,, q, and q, give a very good agreement with the original data in the
middle part for all models but the NSF equations. Recall that the NSF equations do not
account for any rarefaction effects of a,,, a,, and q, ,and just gives their values as zero,
while their values are not zero even in the middle domain. The R13s equations give the
same shape as the kinetic model near the boundary, while the Burnett equations differ in
shape (if the curve from the R13s is upward, then it from the Burnett is downward]. The
computed values of o,, are a good fit with the original data in the middle part, but not so
good near the boundary. Therefore, at small Knudsen numbers, the R13 equations give
the best results, followed by the Gradl3, the Burnett equations, and the NSF equations,
which give the worst result. When the plate velocity orland the Knudsen number
increase, the fitness becomes worse. As an example, in case 5.6 (Kn=0.5, 1000.0 m/s
plate velocity, Figure 5.4), the computed a, and qi from the R13s even all have the
 170
opposite sign as the original data. The above results are not surprising based on the
description about the Knudsen layer in Section 5.4.

Let us consider the original and slightly linearized R13s. At small Knudsen numbers
and plate velocities, results from both models are similar, while at large Knudsen number
or large plate velocities, results from the slightly linearized R13 fit the original data from
kinetic model better than results from the original R13. The original R13 equations only
have some terms of the fourth order of Kn, but not all terms [IS], so it might not be
surprising that their results at some situations would be worse than results from the
slightly linearized R13. Related to the complexity of equations, between these two
R13models, the slightly linearized R13 is suggested, instead of the original R13, for their
simpler form.

Next, let us compare the Grad13 and the R13s equations. At small Kn number and
small plate velocities, results from the R13s fit the original data better than results from
the Gradl3, while at large Kn number and large plate velocities, it seems the conclusion
is opposite, and it is hard to say which one is better for intermediate situations. Normally,
the original values of o, and qi locate in between the values from the Grad13 and them
from the R13s for large Kn numbers or large plate velocities, while for other situations,
the original values of ov and qi locate beyond the values from the Grad13 and them
from the R13s.

Then, let us consider the NSF equations and the Burnett equations. The expressions of
o,, and q, are the same from these two sets of equations. The NSF equations do not
account for any rarefaction effect on oll, 02,
and q, , while the Burnett equations do. At
small Kn numbers, results from the Burnett equations fit the data from the kinetic model
better than results from the NSF equations, while it is hard to say which one is better for
other situations. Normally the original values of oll,a,, and q, locate in between the
values from the Burnett equations and them from the NSF equations (except q, for case
5.1-5.2)
 Finally, let us consider the Grad13 and the Burnett equations. It is found that results
from the Grad13 fit the original data from kinetic model better than results from the
Burnett equations for almost all test cases.
 Dimensionless Domain width y / L

(a)

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

Figure 5.2: Results of trace-free pressure and heat flux from the BGK model and its corresponding sets of
macroscopic equations (case 5.1: Kn=0.025, 300.0 plate velocity) (Triangle for original values from the
BGK, Triangle filled for computed values from the BGK based on Eq. (5.53), Diamond for the NSF,
Diamond filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the slightly
linearized R13)
 0.2 0.4 0.6 0.8
Dimensionless Domain width y / L

6)

Figure 5.2: Results of trace-free pressure and heat flux from the BGK model and its corresponding sets of
macroscopic equations (case 5.1: Kn=0.025, 300.0 plate velocity) (Triangle for original values from the
BGK, Triangle filled for computed values from the BGK based on Eq. (5.53), Diamond for the NSF,
Diamond filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the slightly
linearized R13)
 Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L

Figure 5.3: Results of trace-free pressure and heat flux from the BGK model and its corresponding sets of
macroscopic equations (case 5.7: Kn=O.l,600.0 plate velocity) (Triangle for original values from the BGK,
Triangle filled for computed values from the BGK based on Eq. (5.53), Diamond for the NSF, Diamond
filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the slightly linearized
R13)
 0.2 0.4 0.6 0.8
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L
(d)
Figure 5.3: Results of trace-free pressure and heat flux from the BGK model and its corresponding sets of
macroscopic equations (case 5.7: Kn=O.1,600.0 plate velocity) (Triangle for original values from the BGK,
Triangle filled for computed values from the BGK based on Eq. (5.53), Diamond for the NSF, Diamond
filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the slightly linearized
R13)
 0.2 0.4 0.6 0.8
D i m e n s i o n l e s s D o m a i n w i d t h y/L

(a)

- ~ ~ + - * - t ~ + ~ * - * - * ~ + - + ~ - t - + + + + - - t -
/ , , , I , , , , , I

0.2 0.4 0.6 0.8 1

D i m e n s i o n l e s s D o m a i n w i d t h y/L

Figure 5.4: Results of trace-free pressure and heat flux from the ES-BGK model and its corresponding sets
of macroscopic equations (case 5.6: Kn=0.5, 1000.0 plate velocity) (Triangle for original values from the
ES-BGK, Triangle filled for computed values from the ES-BGK based on Eq. (5.53), Diamond for the
NSF, Diamond filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the
slightly linearized R13)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

DimensionlessDomain width y / L

Figure 5.4: Results of trace-free pressure and heat flux from the ES-BGK model and its corresponding sets
of macroscopic equations (case 5.6: Kn=0.5, 1000.0 plate velocity) (Triangle for original values from the
ES-BGK, Triangle filled for computed values from the ES-BGK based on Eq. (5.53), Diamond for the
NSF, Diamond filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the
slightly linearized R13)
 0.2 0.4 0.6 0.8
DimensionlessDomain width y / L

Dimensionless Domain width y / L

Figure 5.5: Results of trace-free pressure and heat flux from the ES-BGK model and its corresponding sets
of macroscopic equations (case 5.4: Kn=l.O, 300.0 mls plate velocity) (Triangle for original values from
the ES-BGK, Triangle filled for computed values from the ES-BGK based on Eq. (5.53), Diamond for the
NSF, Diamond filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the
slightly linearized R13)
 0.2 0.4 0.6 0.8
Dimensionless Domain width y / L

-40 t , I I
1 1 1 , 1 1 1 1 1

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

Figure 5.5: Results of trace-free pressure and heat flux from the ES-BGK model and its corresponding sets
of macroscopic equations (case 5.4: Kn=l.O, 300.0 m/s plate velocity) (Triangle for original values from
the ES-BGK, Triangle filled for computed values from the ES-BGK based on Eq. (5.53), Diamond for the
NSF, Diamond filled for the Burnett, Star for the Gradl3, Star filled for the original R13, Box for the
slightly linearized R13)
 180
5.7.3 Result. and discussion about third-degree and fourth-degree moments
The only difference among the Gradl3, the R13s (original and slightly linearized) and
the kinetic model to compute oii and qi is how the third-degree and fourth-degree

moments P , ~ ~ ,p,,,,
, and p, used in Eq. (5.53) are computed. In this section, their
values from the Gradl3, the R13s and the kinetic model will be compared, and average
relative errors among them are shown in Tables D.7-D.9 of Appendix D. As two
examples, Figure 5.6 shows the comparison in case 5.1 (Kn=0.025, 300.0 m/s plate
velocity) for the BGK model, while Figure 5.7 shows the comparison in case 5.6
(Kn=0.5, 1000.0rn/s plate velocity) for the ES-BGK model.

After analysis, it is found that the computed values of the moments R,, , p,,,,, from
the Grad13 fit the data from the kinetic model better than or at least similar to results
from the R13s for all test cases, while for R,, and p,,, , the conclusion is
opposite, that is the R13s equations give better results. At small Knudsen numbers or
small plate velocities, the computed values of p, from the Grad13 and the R13s fit the
original data very well, while not good for large Kn and large plate velocities. The
computed A from the Grad13 and the R 13s do not fit the original data.
 DimensionlessDomain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L

(b)
Figure 5.6: Results of third-degree and fourth-degree moments from the BGK model and its corresponding
sets of moment equations (case 5.1: Kn=0.025, 300.0 m/s plate velocity) (Triangle for the BGK, Diamond
for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L

Figure 5.6: Results of third-degree and fourth-degree moments from the BGK model and its corresponding
sets of moment equations (case 5.1: Kn=0.025, 300.0 m/s plate velocity) (Triangle for the BGK, Diamond
for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 Dimensionless Domain width y / L

(e)

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L

(0
Figure 5.6: Results of third-degree and fourth-degree moments from the BGK model and its corresponding
sets of moment equations (case 5.1: Kn=0.025, 300.0 mls plate velocity) (Triangle for the BGK, Diamond
for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 000Q0000000000W000OO0000000000000000000000000000900000000000000000QO~000000000Q04Q0000000000000

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

0.2 0.4 0.6 0.8

Dimensionless Domain width y / L

Figure 5.6: Results of third-degree and fourth-degree moments from the BGK model and its corresponding
sets of moment equations (case 5.1: Kn=0.025, 300.0 mls plate velocity) (Triangle for the BGK, Diamond
for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

6)
Figure 5.6: Results of third-degree and fourth-degree moments from the BGK model and its corresponding
sets of moment equations (case 5.1: Kn=0.025, 300.0 m/s plate velocity) (Triangle for the BGK, Diamond
for the Gradl3, Star for the original R13, Box for the slightly linearized R13)

400
-
"E
1
E
200
N
rl

6
c,
C
a, 0
E
a,
a,
&

$m
&
E.4

-400

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

(a)
Figure 5.7: Results of third-degree and fourth-degree moments from the ES-BGK mode1 and its
corresponding sets of moment equations (case 5.6: Knd.5, 1000.0 m/s plate velocity) (Triangle for the ES-
BGK, Diamond for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 Dimensionless Domain width y / L

0)

0.2 0.4 0.6 0.8

DimensionlessDomain width y / L

Figure 5.7: Results of third-degree and fourth-degree moments from the ES-BGK model and its
corresponding sets of moment equations (case 5.6: Kn=0.5, 1000.0 m/s plate velocity) (Triangle for the ES-
BGK, Diamond for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

Figure 5.7: Results of third-degree and fourth-degree moments from the ES-BGK model and its
corresponding sets of moment equations (case 5.6: Kw0.5, 1000.0 1x11splate velocity) (Triangle for the ES-
BGK, Diamond for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 0.2 0.4 0.6 0.8 1
Dimensionless Domain width y / L

O O O Q Q Q 4 Q O O O Q O Q Q Q O O Q O O Q O O O O O o O o O D O o o v v v o

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y l L

(g)
Figure 5.7: Results of third-degree and fourth-degree moments from the ES-BGK model and its
corresponding sets of moment equations (case 5.6: Kn=0.5, 1000.0 mls plate velocity) (Triangle for the ES-
BGK, Diamond for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 0.2 0.4 0.6 0.8 1
DimensionlessDomain width y / L

0.2 0.4 0.6 0.8 1

Dimensionless Domain width y / L

(0
Figure 5.7: Results of third-degree and fourth-degree moments from the ES-BGK model and its
corresponding sets of moment equations (case 5.6: Kn=0.5, 1000.0 m/s plate velocity) (Triangle for the ES-
BGK, Diamond for the Gradl3, Star for the original R13, Box for the slightly linearized R13)
 190
5.8 Conclusion
In this chapter, the Grad13 equations, the original R13 equations, the slightly
linearized R13 equations, the NSF equations and the Burnett equations for the BGK
model and the ES-BGK model are compared with the kinetic models themself for steady
state 1DCF.

At small Kn numbers and small plate velocities (Situations Sb and Sd, Kn 10.1, plate
velocity 1300.0 rnts), all (except the NSF equations for a,,, a, and q,) work very well,
especially at the middle part of the flow domain. As the Knudsen number or plate
velocity increases, the accuracy of results from these sets of macroscopic equations
becomes worse. Among these sets of equations, the slightly linearized R13 equations is
better than the original R13 equations; the Grad13 equations is better than the Burnett
equations and the NSF equations; while between the Grad13 equations and the R13s
equations, between the Burnett equations and the NSF equations, it is hard to say which
one is better (except at small Kn situations (Kn lO.l), in which the R13s equations are
better than the Grad13 equations, the Burnett equations is better than the NSF equations).

Therefore the slightly linearized R13 equations is suggested to be used as macroscopic

continuum equations in the rarefied gas flow at Kn 5 0.1, while for rarefied gas flow at
Kn > 0.1, none tested macroscopic continuum equations can be considered to be perfect.
While the higher order models, in particular the R13s equations can describe rarefaction
effects qualitatively, they certainly fail to give an accurate quantitative description. This
can be attributed to their failure in describing boundary layer effects. For smaller
Knudsen numbers, all higher order models give a good description of the bulk behavior,
e.g. see Figure 5.2. When the Knudsen number grows, and boundary effects dominate, all
models fail.

This matches with the observations in the computation of shock structures [26]:
Burnett, Grad 13, and R13 equations yield shock structures in agreement with DSMC
computations for Mach numbers below Ma=3.0, but not for higher Mach numbers, where
rarefaction effects and deviations from equilibrium states become more important.
 In conclusion we can state that the failure to describe Knudsen boundary layers, and
strong shocks, puts strong limitations on the applicability of all macroscopic models
considered.
Chapter 6 Conclusions and outlook

6.1 Conclusions
The Boltzmann equation is the basic equation to describe rarefied gas flows. Some
kinetic models with simple collision term expressions have been proposed to reduce the
mathematical complexity of the Boltzmann equation. The Boltzmann equation and
kinetic models describe rarefied gases at the microscopic levewith great accuracy, but a
cumbersome, and expensive, to solve.

Macroscopic models can be derived form the Boltzmann equation, or kinetic models,
which reduce the number of variables drastically, but these can be applied only to flows
with not too large Knudsen numbers. Important macroscopic models considered in this
thesis are the Navier-Stokes-Fourier equations and the Burnett equations, which both
follow from the Chapman-Enskog expansion, and extended sets of moment equations, in
particular Grad's 13 moment equations, and their regularization (R13).

In the first part of this work, two new kinetic models (new kinetic model I and new
kinetic model 11) were proposed. Both models use a meaningful expression for the
collision Erequency, while retaining the important properties for a kinetic model at the
same time. Both models can be reduced to the ES-BGK model and the V(C)-BGK model
for suitable choices of parameters, and thus can be considered as more general kinetic
models. The H-theorem of these two new kinetic models could be proven only for small
Kn number situations.

In the second part of this work, several kinetic models (BGK, ES-BGK, v(C)-BGK,
two new kinetic models) were tested numerically for the situations of one-dimensional
shock waves at steady state and plane steady state Couette flow. Computational results
from the kinetic models were compared to results obtained with the Direct Simulation
Monte Carlo Method (DSMC), which serves as a benchmark in this work. The effects of
space grid spacing, bounds and step size of the velocity grid on the computational results
were discussed, and it was shown that shock wave problem at not small Mach numbers,
 193
and Couette flow problem at small Knudsen numbers and large plate velocities, require
very fine space and velocity grids to eliminate the numerical error due to discretization.
This implies large computational times, which for the problems considered in the thesis
are of the same order as those encountered in DSMC simulations. The benefit of discrete
velocity models for kinetic schemes over DSMC simulations is that a deterministic
solution is computed, and not a stochastic solution. Accordingly, the results do not suffer
from stochastic noise, so that it will be no trouble to compute the derivative of
macroscopic parameters and further transport coefficients (such as viscosity and thermal
conductivity). Moreover, discrete velocity schemes allow to considering time dependent
(means non steady state) problems, which are difficult to access for the DSMC method.
The advantage of using kinetic models instead of the Boltzmann equation lies in the
much easier numerical treatment of the collision term.

Results from the two new kinetic models are very similar, but the new kinetic model I1
is better for a simpler program structure, which further affects the computational time.
For hard sphere molecules, results from the two new kinetic models locate in between
results from the ES-BGK model and the v(C)-BGK model, while for the Maxwell
molecules, the two new kinetic models are identical to the ES-BGK model.

For the shock wave problem, the new kinetic model 11 gives the best agreement to the
DSMC simulation (See Table D.l, the average relative errors of density, velocity and
temperature of the new kinetic model I[ comparing to the DSMC are smaller than 0.05
even for Ma=6.0); for Couette flow, however, the ES-BGK model gives best results (See
Tables D.2-D.4, the average relative errors of density, velocity and temperature of the
ES-BGK model comparing to the DSMC are smaller than 0.05 for all test situations
(Kn 5 1.0)). The v(C) -BGK model gives results of insufficient accuracy in both cases,
and requires the longest computational time (for it need more iterations to reach certain
total time step than other kinetic models), and therefore should not be used to model
flows
 194
Also in the second part of this work, a modification of Mieussens's original numerical
scheme was developed. The basic idea is to use linearized expressions for the reference
distribution function f, (for the ES-BGK model and the two new kinetic models, there
are two types linearized f, ), instead of its exact expression, in the numerical program.
Results from this modified scheme are almost identical to results from the original
scheme for most test cases, while about 20-40 percent of computational time can be
saved. For the ES-BGK model and the two new kinetic models, the second type
linearized fief, which is based on an anisotropic Gaussian, is better than the first type
linearized f,, which is based on an isotropic Gaussian. Some pity things about this
modified numerical scheme are: the first type linearized f, of the new kinetic model I
with hard sphere molecules is unstable for all test cases (NaNQ are output from the
Fortran program), the second type linearized f , of the new kinetic model I with hard
sphere molecules is unstable for high Mach (e.g. Ma=6.0) situations (NaNQ are output
from the Fortran program), and the linearized f, of the v ( C )-BGK model can not give
results with accuracy enough for middle and high Mach situations.

In the third part of this work, several macroscopic continuum equations, the Navier-
Stokes-Fourier equations, the Burnett equations, Grad 13 moment equations, and the
regularized 13 moment equations (in original form and in slightly linearized form), were
derived from the BGK and the ES-BGK models, and were then compared with the
corresponding kinetic model for steady state Couette flow. At small Knudsen numbers
and small plate velocities, all (except the NSF equations for a,,, a, and q,) work very
well, especially at the middle part of the flow domain. As the Knudsen number or the
plate velocity increases, results from these sets of macroscopic equations become
increasingly different to results from kinetic models. Among these sets of equations, the
slightly linearized R13 is better than the original R13, the Grad13 is better than the
Burnett and the NSF, while it is hard to say which one of the Grad13 and the R13s, or the
Burnett and the NSF, is better. For Knudsen numbers Kn 50.1,the R13 equations are
better than the Grad13 equations, and the Burnett equations are better than the NSF
equations. Therefore the slightly linearized R13 equations are suggested to be used as
 195
macroscopic continuum equations in rarefied gas flows at Kn 10.1, while for rarefied gas
flows at larger Knudsen numbers none of the tested macroscopic continuum equations
can be recommended without reservation.

6.2 Outlook
The results presented in this thesis show that numerical solutions of kinetic models
can be used to model rarefied gas flows with sufficient accuracy (new kinetic models for
shock wave, and ES-BGK model for Couette flow). In this thesis the standard test cases
of shock waves and Couette flow were considered, and the natural next step is to simulate
rarefied gas flows in more realistic applications, which will be two-dimensional or three-
dimensional problems. However, the computational times involved are substantial, and in
order to reduce these as much as possible, proper criteria for convergence should be
developed which should allow to maximize time and space steps, in order to minimize
computational time. Moreover, in order to reduce computational time, one also can
consider to using parallel computing orland implicit numerical schemes.

The examination of the macroscopic continuum models showed that these can
describe rarefied gas flows outside the Knudsen layer in good agreement to solutions of
microscopic models (i.e. kinetic models, andlor DSMC), but fail in the description of the
Knudsen layer. Since the numerical effort for macroscopic models is substantially smaller
than the effort to solve microscopic equations, the former should be considered wherever
possible, that is outside of Knudsen layers and strong shocks. The Knudsen layer itself
(and strong shocks) should be simulated with a microscopic model. This will lead to
hybrid schemes, where microscopic and macroscopic models are solved in their
respective domains. Hybrid models which connect the Boltzmann or kinetic model
solutions to the Navier-Stokes-Fourier equations have been shown to be an effective way
to save computational time and memory, e.g. [81-841. The Navier-Stokes-Fourier
equations cannot describe rarefaction effects, which are, however, described in the
Burnett, the Grad 13, and the R13 equations. Since the Burnett equations suffer from
instabilities, and the Grad 13 equations describe unphysical shocks, one should consider
the (slightly linearized) R13 equations in flow areas where mild rarefaction plays a role,
 196
and the Navier-Stokes-Fourier equations in regions were rarefaction effects are
unimportant, together with a suitable kinetic model (e.g. the ES-BGK model) in areas of
large rarefaction.
 197
References
C. Cercignani, Rarefied gas dynamics: from basic concepts to actual calculations,
Cambridge University Press, 2000.

M. S. Ghidaoui, J. Q. Deng, W. G. Gray, and K. Xu, A Boltzmann based model

for open channelflows, International Journal for Numerical Methods in Fluids 35,
449-494 (2001).

C. Cercignani, Theory and application of the Boltzmann equation, Scottish

Academic Press Ltd., 1975.

C. Cercignani, The Boltzmunn equation and its applications, Springer-Verlag

New York Inc., 1988.

S. Harris, An introduction to the theory of the Boltzmann equation, Holt, Rinehart

and Winston, Inc., 1971

I. Muller, Thermodynamics, Pitman Advanced Publishing Program, 1985.

G. S. Springer, Heat transfer in rarefied gases, Advances in Heat Transfer 7, 163-

218 (1971).

G. E. Karniadakis and A. Beskok, Micro flows: fundamentals and simulation,

Springer-Verlag, New York, 2002.

M. N. Kogan, Rarefied gas dynamics, Plenum Press, 1969.

H. Struchtrup, Failures of the Burnett and super-Burnett equations in steady state

processes, accepted for publication in Continuum Mechanics and
Thermodynamics (2004).

Y. Zheng and H. Struchtrup, Burnett equations for the ellipsoidal statistical BGK
model, Continuum Mechanics and Thermodynamics 16 (1-2), 97-108 (2004).

S. Chapman and T. G. Cowling, The mathematical theory of non-uniform gases,

Third Edition, Cambridge Univerisity Press, 1970.

J. H. Ferziger and H. G. Kaper, Mathematical theory of transport processes in

gases, North-Holland Publishing Company, 1972.

H. Struchtrup, Some remarks on the equations of Burnett and Grad, Transport in

Transition Regimes, IMA Volume Series 135, Springer New York (2003).

H. Struchtrup, Stable transport equations for rarefied gases at high orders in the
Knudsen number, Physics of Fluids 16 (1I), 3921-3934 (2004).
H. Grad, Principles of the kinetic theory of gases, Handbuch der Physik XII, 205-
294 (1958)

C. D. Levermore, Entropy-based moment closures for kinetic equations,

Transport Theory and Statistical Physics 26,591-606 (1997)

C. D. Levermore, Moment closure hierarchies for kinetic theories, Journal of

Statistical Physics 83, 1021-1065 (1996).

C. D. Levermore, W. J. Morokoff, The Gaussian moment closure for gas

dynamics, SIAM Journal of Applied Mathematics 59 (I), 72-96 (1998).

W. Weiss, Continuous shock structure in extended thermodynamics, Physical

Review E 52,5760 (1995).

I. Muller and T. Ruggeri, Rational extended thermodynamics, Springer, New

York, 1998.

P. Charrier, B. Dubroca and J. L. Feugeas, Levermore's moment closure of

discrete Boltzmann equationsfor non-equilibrium kineticflows, 21st International
Symposium on Rarefied Gas Dynamics I, Marseille France, 95-102 (1998).

M. Junk, Maximum entropy moment problems and extended Euler equations,

IMA workshop "Simulation of Transport in Transition Regimes", 2000,
ww w .mathematik.uni-kl.de/-iunk/~ublications/list.
html

S. Reinecke and G. M. Kremer, Bumett's equationsfrom a (13+9N) -field theory,

Continuum Mechanics and Thermodynamics 8, 121- 130 (1996).

S. Reinecke and G. M. Krerner, Method of moments of Grad, Physical Review A

42 (2), 8 15-820 (1990).

H. Struchtrup and M. Torrilhon, Regularization of Grad's 13 moment equations:

Derivation and linear analysis, Physics of Fluids 15,2668-2680 (2003).

M. Torrilhon and H. Struchtrup, Regularized 13-Moment equations: shock

structure calculations and comparison to Bumett models, Journal of Fluid
Mechanics 513, 171-198 (2004).

P.L. Bhatnagar, E.P. Gross and M. Krook, A model for collision processes in
gases. I: small amplitude processes in charged and neutral one-component
systems, Physical Review 94,s 11-525 (1954).

J. Lowell H. Holway, New statistical models for kinetic theory: methods of

construction, Physics of Fluids 9,1658-1673 (1966).
J. Lowell H. Holway, Kinetic theory of shock structure using an ellipsoidal
distributionfunction, Rarefied Gas Dynamics I, 193-2 15 (1965).

F. Sharipov and V. Seleznev, Data on internal rarefled gas flows, Journal of

Physical and Chemical Reference Data 27,657-706 (1998).

P. Andries and B. Perthame, The ES-BGK model equation with correct Prandtl
number, AIP Conference Proceedings 585 (I), 30-36 (2001).

P. Andries, P. Le Tallec, J. Perlat and B. Perthame, The Gaussian-BGK model of

Boltzmann equation with small Prandtl numbers, European Journal of Mechanics:
B Fluids 19(6), 8 13-830 (2000).

L. Mieussens and H. Struchtrup, Numerical comparison of BGK-models with

proper Prandtl number, Physics of Fluids 16(8), 2797-28 13 (2004).

H. Struchtrup, The BGK-model with velocity-dependent collision frequency,

Continuum Mechanics and Thermodynamics 9,23-3 1 (1997).

E. M. Shakhov, Generalization of the Krook kinetic relaxation equation, Fluid

Dynamics (English Translation of Izvestiya Akademii Nauk SSSR, Mekhanika
Zhidkosti i Gaza) 3,9596 (1968).

E. M. Shakhov, Kinetic model equations and numerical results, the 14th

International Symposium on Rarefied Gas Dynamics, Tsukuba Japan, 1984.

G. Liu, A method for constructing a model form for the Boltzmann equation,
Physics of Fluids A 2,277-280 (1990).

V. Garzo, Transport equations from the Liu model, Physics of Fluids A 3, 1980-
1982 (1991).

V. Garzo and M. L. d. Haro, Kinetic model for heat and momentum transport,
Physics of Fluids 6,3787-3794 (1994).

L. Mieussens, Discrete-velocity models and numerical schemes for the

Boltzmann-BGK equation in plane and axisymmetric geometries, Journal of
Computational Physics 162(2), 429-466 (2000).

L. Mieussens, Convergence of a discrete-velocity model for the Boltzmann-BGK

equation, Computers and Mathematics with Applications 41,83-96 (2001).

L. Mieussens, Discrete velocity model and implicit scheme for the BGK equation
of rarefied gas dynamics, Mathematical Models and Methods in Applied Sciences
10,1121-1149 (2000).
P. Charrier, B. Dubroca and L. Mieussens, Discrete velocity BGK model and
implicit scheme for computing rarefied gas flows, 21st International Symposium
on Rarefied Gas Dynamics I, Marseille France, 95-102 (1998).

H. Babovsky and R. Illner, A convergence proof for Nanbu's simulation method

for thefull Boltzmunn equation, SIAM Journal of Numerical Analysis 26, 45-65
(1989).

E. S. Oran, C. K. Oh and B. Z. Cybyk, Direct simulation Monte Carlo: recent

advances and applications, Annual Review of Fluid Mechanics 30, 403-441
(1998).

G. A. Bird, Molecular gas dynamics and the direct simulation of gas flows,
Oxford Science Publications, Oxford, 1994.

Toby Thatcher's master thesis (in progress), Department of mechanical

engineering, University of Victoria.

P. Le Tallec, A hierarchy of hyperbolic models linking Boltzmann to Navier

Stokes equationsforpolyatomic gases, ZAMM 80 (1 1-12), 779-789 (2000).

J. 0. Hirschfelder, C. F. Curtiss, and R. B. Bird, Molecular theory of gases and

liquids: John Wiley & Sons, Inc., 1954.

R. L. Liboff, The theory of kinetic equations, Wiley and Sons, New York, 1969.

F. Bouchut and B. Pertharne, A BGK model for smull Prandtl number in the
Navier-Stokes approximation, Journal of Statistical Physics 71, 191-207 (1993).

L. B. Barichello, A. C. R. Bartz,M. Camargo and C. E. Siewert, The temperature-

jump problem for a variable collision frequency model, Physics of Fluids 14 (I),
382-39 1 (2002).

S. K. Loyalka, Velocity profile in the Knudsen layer for the Kramer's problem,
Physics of Fluids 18 (12), 1666-1669 (1975).

S. K. Loyalka and K. A. Hickey, Velocity slip and defect: Hard sphere gas,
Physics of Fluids A 1(3), 612-614 (1989).
 201
K. Xu, A gas-kinetic BGK scheme for the Navier-Stokes equations and its
connection with artijicial dissipation and Godunov method, Journal of
Computational Physics 171,289-335(2001).

W. G. Vincenti and C. H. Kruger, Jr., Introduction to physics gas dynamics, John

Wiley & Sons, Inc., 1965.

H. C. Yee and J. L. Shinn, Semi-implicit and fully implicit shock-capturing

methodfor hyperbolic conservation laws with stiflsource temzs, AIAA Paper, 87-
1116 (1987).

H. C. Yee, Construction of explicit and implicit symmetric TVD schemes and their
applications, Journal of Computational Physics 68, 151- 179 (1987).

W. H. Press, B. P. Flannery, S. A. Teukolsky and W. T. Vetterling, Numerical

Recipes, Cambridge University Press, 1986.

R. L. Burden and J. D. Fakes, Numerical Analysis, Sixth Edition, BrooksICole

Publishing Company, 1997.

J. E. Dennis and R. B. Schnabel, Numerical methods for unconstrained

optimization and nonlinear equations, Prentic Hall Inc., 1983.

L. Mieussens (Toulouse France), private communicaton, 2003.

A. Schuetze, Direct simulation by Monte Carlo modeling one dimensional shock

wave using dsmclas$ A user's manual, Report, Department of mechanical
engineering, University of Victoria, 2003.

G. C. Pham-Van-Diep, D. A. Erwin and E. P. Muntz, Testing continuum

descriptions of low-Mach-number shock structures, Journal of Fluid Mechanics
232,403-413 (1991).

H. Struchtrup, Kinetic schemes and boundary conditions for moment equations,

ZAMP 51,346-365 (2000).
A. Schuetze, Direct simulation by Monte Carlo modeling Couette flow using
dsmc1as.f: A user's manual, Report, Department of mechanical engineering,
University of Victoria, 2003.

[72] V. Garzo, On the derivation of the Burnett hydrodynamic equations from the
Hilbert expansion, Physica A 149,551-560 (1988).
C. Y . Cha and B. J. McCoy, Third-order constitutive equations for a monatomic
gas, Journal o f Chemical Physics 34,4369-4372 (1971).

R. K. Agarwal, K. Y . Yun, and R. Balakrishnan, Beyond Navier-Stokes: Bumett

equations for flows in the continuum-transition regime, Physics o f Fluids 13,
3061-3085 (2001).

A. V . Bobylev, The Chapman-Enskog and Grad methods for solving the

Boltzmann equation, Soviet Physics. Dokl. 27,29-3 1 (1982).

M. Kurtz, Handbook of Applied Mathematics for Engineers and Scientists,

McGraw-Hill, Inc., 199 1.

R. Balakrishnan, Entropy consistent formulation and numerical simulation of the

BGK-Burnett equations for hypersonicflows in the continuum-transition regime,
Dissertation, Wichita State University 1999.

H. Struchtrup, Derivation of 13 moment equationsfor rarefied gas flow to second

order accuracy for arbitrary interaction potentials, Multiscale Modeling &
Simulation, (2004)be in press.

Y . Sone, Flows induced by temperature jields in a rarefied gas and their ghost
efect on the behavior of a gas in the continuum limit, Annual Review of Fluid
Mechanics 32,779-8 1 1 (2000).

S. F. Liotta, V . Romano, and G. Russo, Central schemes for balance laws of

relaxation type, SIAM Journal of Numerical Analysis 38, 1337-1356 (2000).

P. L. Tallec and J. P. Perlat, Coupling kinetic models with Navier-Stokes

equations, in Computational fluid dynamics review, M. Hafez and K. Oshima,
Eds. World Scientific, 833-855 (1998).

J. Schneider, Direct coupling offluid and kinetic equations, Transport Theory and
Statistical Physics 25,681-698 (1996).

S. P. Popov and F. G. Cheremisin,Example of simultaneous numerical solution of

the Boltzmann and Navier-Stokes equations, Computational Mathematics and
Mathematical Physics 41,457-468 (2001)
[84] J. F. Bourgat, P. L. Tallec, B. Perthame, and Y. Qiu, Coupling Boltzmann and
Euler equations without overlapping, Contemporary Mathematics 157, 377-398
(1994).

[85] J. M. Gere and W. Weaver, Jr., Matrix algebra for engineers, 2nd Edition,
Wadsworth Inc., 1983.
 -
Appendix A: Flow diagrams of N-R algorithm and computational
programs

Q Start this part

Initial guess of coefficients

Reference distribution functi

I

Functions for Newton Algorithm

Linear System for &+I-&

RexFac=RexFac/2.0

Yes

Figure A.l: Flow diagram of N-R algorithm used for reference distribution function
 +
Input Part
I
I~ i information about discretized scheme
ebasic I
Boundaw condition Part 1. Itel=l I
I
Initial Guess Distribution function

I Dim Less Coll Freq I

I

i
I

Coll Freq

I Output 3

- -

Figure A.3: Flow diagram of program for one-dimensionalCouette flow

 207
Appendix B Expression of functions and Jacobian in the N-R algorithm
for all kinetic models
In this appendix, a subscript k is used to show the iteration number in the N-R
algorithm, and the meaning and expression of dimensionless distributions and their
coefficients, and so on, refer to corresponding part in Chapters 3 and 4. Figure A.l shoe
the flow diagram of the N-R algorithm used for reference distribution function.

B.l BGK model

For the IDSW, functions we compute in the N-R algorithm are

where the dimensionless distribution

The coefficients are obtained as follows

where the Jacobian Jtkused in the computation is

Since
aF&K.i.j& -
- F&K.i,jL aF;~~.i.j.k = -($j y- BGK ,i, j.k '
at> aTk
J

aPl:,
-=C a aFiGK,i,j*
- j=,
F;GK,i,jk
,then
aa,> j=l a F 'jVjk aa&

For the lDCF, functions we compute in the N-R algorithm are

j=1

Jacobian used in the computation is

B.2 ES-BGK model
For the IDSW, functions we compute in the N-Ralgorithm are

where 883: = %( By'i

1 P"
+( 1 - ~ ) P L , ~BB4: --;[
) , = 2 1~ i P nB GRK . ~(~I - - i r ) P k , i ) -
+

Jacobian used in the computation is

 For the lDCF, functions we compute in the N-R algorithm are
* J I J
=
j=l
~&i.j,k 9 F2;k =
j=1
'l;i,F,&.i, j.k - 9 '3; = C
j=l
$j2 F,&,i.j* -0 3

where

J
Note that P7tk = xqcjl77cj2F A , 8 . .
- BB7: 1.D - 5 . Jacobian
for situation A~s(BB~:)<
j==l

used in the computation is

JEk = 6:k.I 2 J k J:,k,5 J:,k,6 ~:k.7 I* (Bag)
 --
-

BE;
Note that for situation A ~ S ( B B)~<: 1.D- 5 ,

B.3 V ( C ) -BGK model

For the lDSW, functions we compute in the N-R algorithm are
 J f.". J fiyj
=zQ;lj'"lAc, j 2 ; = z ~ i ' f j 1 9 ~ j Tf3;
- J A?j
A ~= ,~ ( ? ? : j Y ' i ' f j - ~ ~ * Note
where
j=I Pi" j=l Pi j=i Pi"

that F2tk = x$j,

J

j=1
9;j~;i,j,k - f 2; for situation Abs(j2:)< 1 .D - 5 (which may happen

when u:,~= 0.0. Jacobian used in the computation is

(B.10)

Note that for situation Abs(j2;)< 1

For the IDCF, functions we compute in the N-Ralgorithm are

 214

(B. 11)

74;= c($,y ~ pi": ~

J

j=l
f.".
~ A C Note
. that P2; =
J
C ~ ~ ; , Q :-,721
j=1
F ; ~for, ~situation
,~
~ b s G 2 ; ) e1 .D - 5 (which may happen when u,", = 0.0)) and
J
93:, = Cq;j2~;j~;i,j,k
- 7 3 : for situation ~ b s ( j 3 : ) el.D - 5 (which may happen
j l

when ujf,= 0.0 ). Jacobian used in the computation is

( B .12)
Note that for situation ~bs(72:)<1 .D - 5,

and for situation Abs(73; )< 1 l - 5 ,

B.4 New kinetic model I

For the lDSW, functions we compute in the N-R algorithm are

J
that F2:, = CT,I;,,V;~F;;,
, - 72;
,. , for situation Abs(f2;) < l.D - 5 (which may
j=1

happen when u:, = 0.0).Jacobian used in the computation is

 L

Note that for situation Abs(f2;) < 1 .D- 5 ,

For the lDCF, functions we compute in the N-R algorithm are

(B. 15)

J f .". J
where 9:72; = C$jlPl:j
T1;=CPI>-Ac9 +Ac9 73; = Cqn.P-Ac,fi?j

j=l Pi" j=1 Pi j=1 l J pin

~ b s ( 7 3 ); < 1 - 5 . Jacobian used in the computation is

 217

(B. 16)

I
j ( j ~ l , i , j , k j j l ( j , . j , j j j j k kp:j~~j(q:jpF~l.i.j.k
j=l j=l j=1 - j=l
J;k.4 = 74; 74; 74;

Note that for situation ~bs(f"2:) c 1.D - 5 ,

and for situation ~ b s ( J 3 ;) < 1.D -5 ,

B.5 New kinetic model I1

For the lDSW, functions we compute in the N-R algorithm are
 J
Note that 82;, = ~ q ~ j l O ~ j ~ ~-l lf21 , k situation ~bs(f2:) < l.D - 5 (which may
, i , jfor
j=1

happen when u:,~= 0.0). Jacobian used in the computation is

Note that for situation ~bs(f"2;)< 1 .D- 5 ,

For the lDCF, functions we compute in the N-R algorithm are

where
":
~ l ; = ~ v ~ j - ~ ~ ,
j=l pin
 220
J
Note that 82; = xq:j,f l : j ~ : f f , i , ,, - f 2; for situation Abs(y2:) < 1 .D - 5 ,
j=1

J
P3:, = ~ ? ? [ j Z 3 ~ j F ~ I I , i . j ,-k S3: for situation Abs(f3:) < 1 .D - 5 , and
j=l

J
, F:',,~,,,,, - f71 for situation ~ b s ( y 7 "c 1.D - 5 . Jacobian used in
P7tk = ~ ( ~ i ' #j2f ~ )PC,
j=l

the computation is
Note that for situation Abs(f"2;) < 1 .D- 5 ,

and for situation Abs(f"3;) < l.D - 5 ,

and for situation Abs(f"71) < 1 .D- 5 ,

 223
Appendix C: Linearized dimensionless reference distribution F&,

C.l BGK model

For the lDSW, linearized dimensionless reference distribution is

F,&K,i,j=ex*(- ( q ; j r l a l : +a2;q:jI +a3:(qtj Y), (c. 1 )

and three conditions to determine the three coefficients are

where

which is a symmetric matrix.

For the lDCF, linearized dimensionless reference distribution is

F;GK,i,j= exp(- r
(6lalI + a21 rl,:, + a3;ql. + a4; ( q n )'
a.1~ 1.1 r
and four conditions to determine the four coefficients are
where

J; =

(C.6)
which is a symmetric matrix.

C.2 ES-BGK model

From section 3.2.4 and section 4.2.4, there are two types linearized reference
distribution f&,iej for the ES-BGK model. Here, only the discussion of first type
linearized f A i , j is given, while the discussion of second type linearized f&,i.j is omitted,
which is similar to the discussion of first type.

For the lDSW, linearized dimensionless reference distribution is

and four conditions to determine the four coefficients are

For the ZDCF,linearized dimensionless reference distribution is
(C. 12)
 ) ,which is a symmetric matrix.

C.3 V(C)-BGKmodel
For the lDSW, linearized dimensionless reference distribution is

F;~,, = exp(- (q:, p). (al; +a21q:j1+ a3: (qcj ), (c. 13)

and three conditions to determine the three coefficients are

J "j
C P ; ~ F ; , ~ =all J;ll + a21 J;,, + a31 J L , = CP:j~ A c ,
j=1 j=l Pi
J J h"j
xP:j$'.
j=1
YJ.I
I.JI F n . . = all J;, + a21 J;, + a31 J;, =~ ~ ; j ? # j l
j=1
f AC,
pi
(C. 14)

-
J
x9:jq:j,e~p(-(q:jP) i+:j(V;jPeXP(-h:jP)
j=1 j=l

i ~ ; ~ ( ~ t ~ ~ p e x p ( - ( q$cp ~t j P
( q): j ~ q : j l e x ~ ( - ( q t j P )
j=1 j=1
J I
j l j l ( j ~ ( ( j ~pcj(qtjP(~:jPeXP(-(q~jp)
)
j=1 j=l -

(C.15)
 229
which is a symmetric matrix.

For the lDCF, linearized dimensionless reference distribution is

,
F;, = exp(- (q:j Plull + a21q:, + a4; (n:j ) ),
+ a3; gj2
2
(C. 16)

and four conditions to determine the four coefficients are

J fi;
x P ; l jF;,, = all J:,
j=1
+ a21 J;, +a31 J1!'', + a41 J;, = XP; PAC,
,=I Pi"

j=1 j=1

where J; = [J;, Jr2 J;, J;, 1, which is a symmetric matrix.

9 (C.18)
 230
C.4 New kinetic model I
From section 3.2.4 and section 4.2.4, there are two types linearized reference
distribution f:,i,j for the new kinetic model I. Here, only the discussion of first type

linearized f:,,,,is given, while the discussion of second type linearized f:, ,. is omitted,
which is similar to the discussion of first type.

For the lDSW, linearized dimensionless reference distribution is

F:,+i,j= exp(-($,~).(al; +a2;qtjl +a3;Atj), (C. 19)

, ~)
where At:, = a 3 ~(qfjt
2
+ a4'& ((q: j2 r + r ),
(qfj3 and three conditions to determine the
three coefficients are

iqj
($, P
j=1
F;;,,~,,= a11 J;, + a2; J;' +a3: J;,
J
= cP:,
j=1
2 fi)
(v;~) ,:j)Ac,
Pi
where

(C.21)
which is not a symmetric matrix.

For the lDCF, linearized dimensionless reference distribution is

FNLiJ= exp(- (vcj )2). (al; +a2;%, + a31qfj2+ a41A;), (C.22)

 232
linearized f;// . .jis given, while the discussion of second type linearized f;lI,i,j is omitted,
which is similar to the discussion of first type.

For the lDSW, linearized dimensionless reference distribution is

and four conditions to determine the four coefficients are

where

J; = [J;, J;, J:, J:, ] ,which is not a symmetric matrix.

 For the 1DCF7linearized dimensionless reference distribution is

and seven conditions to determine the seven coefficients are

J
F;ll,i,j = al,"JiIz1+a21 J:22+a31 J:23 +a47ilfz4+ a5,"J;,
$:j~:jI + a6"ilf2, + a7,"J,!27 = BB2,"
j=l

where
JtI J;, If3 J;, Jf5 J:, J;], which is a symmetric matrix.
Appendix D: Average relative errors in comparisons

Table D. 1: Average relative errors of kinetic models corn lared to DSMC in lDSW
Parameters Density Velocity Momentum Temperature Pressure p

u1 pul

3.6 New I .18E-01 .19E-01 .46E-03

3.6 New I1 .13E-01 .14E-01 S3E-03
 q q - T011
Table D.2: Average relative errors of kinetic models compared to DSMC in 1DCF. Part I
I Parameters Density Velocity Tempera Pressure Pressure Pressure Pressure

I Case and Kinetic

model
P 012

4.3 BGK
4.3 ES-BGK
4.3 "(c)-BGK
4.3 New I
4.3 New I1
4.4 BGK
4.4 ES-BGK
4.4 v(c)-BGK
4.4 New I
4.4 New I1
 238
~bleD.3: Aver lge relati .e errors of kinetic models c ~mparedto DSMC in lDCF Part I1
Parameters Density Velocity Tempera Pressure Pressure Pressure Pressure
1P u1
ture T P
Case and
Kinetic model
4.5 BGK .17E-2
4.5 ES-BGK .78E-3
4.5 v(c)-BGK .29E-2
4.5 New I .15E-2
4.5 New I1 .15E-2
4.6 BGK .99E-3
4.6 ES-BGK .13E-2
4.6 "(c)-BGK .36E-2
4.6 New I .13E-2
4.6 New I1 .13E-2
4.7 BGK .73E-3
4.7 ES-BGK .11E-2
4.7 v(c)-BGK .ME-2
4.7 New I .16E-2
4.7 New I1 .16E-2
4.8 BGK .1OE-2
4.8 ES-BGK .16E-2
4.8 4 ~ ) - B G K .28E-2
4.8 New I .16E-2
4.8 New I1 .16E-2
 239
Table D.4: Average relative errors of kinetic models compared to DSMC in 1DCF. Part
m
Parameters

Case and Kinetic

model
4.9 BGK
4.9 ES-BGK
4.9 v(c)-BGK
Density
P

I
Velocity Tempera Pressure
P
Pressure 1 Pressure I Pressure I

4.9 New I
4.9 New I1
4.10 BGK
4.10 ES-BGK

4.12 New I
Table D.5: Average relative errors of macroscopic continuum equations compared to
kinetic models. Part I
Grad 1 3 R13I R13II Kinetic R13IyRl
equations 31

.24E+O
.20E+O
.33E+O
.34E-1
.26E+O
.39E+O
.39E-1
Notes: 1. R13I means the original R13, R13II means th slightly lit
2. Values in column "Kinetic equations" mean the average relative errors between the computed values from the kinetic
model based on Eq. (5.53) and the original values from the kinetic model.
3. Values in column "R13II/R13I" mean the average relative emrs between the computed values from the R13II and
the computed values from the R13I.
Table D.6: average relative errors of macroscopic continuum equations compared to
kinetic models. Part 11

z::; I1 k
Case and et of equations NSF

Es-
o1 Pressure
BGK
Pressure 0Z2
Pressure 012
ES- ~ressureell
BGK
Pressure 0Z2
Pressure 012
5.3 Es- ~ressureall
BGK
Pressure OZ2
Pressure 012
5.4 ES- Pressure ol1
BGK
II Pressure ba
Pressure o,,

Pressure 0,-

BGK

5.7 ES-
BGK

BGK

Pressure oi2 I I
Notes: 1. R13I means the origi [I means the slightly linearized R13
2. Values in column "Kinetic equations" mean the average relative errors between the computed values from the kinetic
model based on Eq. (5.53) and the original values from the kinetic model.
3. Values in column "R13IYR13I" mean the average relative errors between the computed values from the R13II and
the computed values from the R13I.
 242
Table D.7: Average relative errors of thirddegree and fourth-degree moments of moment equations compared to
kinetic models. Part I
Set of equations BGK ES-BGK

zzz?- Grad13

.42E-1
R131

.30E-1
R13II

.29E-1
R13IYR13I Grad13 R13I

.60E-4 .81E-1 .61E-1

R13II

.61E-1
R13IYR13I

.82E-4
Pl12
.10E+1 .69E+O .69E+O S3E-2 .10E+1 .75E+O .75E+O .48B2
P<ll2,

P122
.58E+O .13E+O .13E+O .14E-2 .59E+O .18E+O .18E+O .18E-2
.10E+1 .20E+O .20E+O .31E-2 .1OE+1 .28E+O .28E+O .40E-2
P<122,
.20E-1 .14E-1 .14E-1 .16E-4 .38E-1 .28E-1 .28E-1 .22E-4
 243
Table D.8:Average relative errors of thirddegree and fourthdegree moments of moment equations compared to
kinetic models. Part I1
et of equations BGK
\
ES-BGK
R13II R13IVR13I W l 3 R131
Parameters Grad13 R131
5.4 P,u .67E+O .22E+O
Table D.9: Average relative errors of thirddegree and fourthdegree moments of moment equations compared to
kinetic models. Part 111
Case I ES-BGK
R13II 1 R13IVR13I 1 Grad13 1 R131 I R13II 1 R13IIlR13I
5.7

5.8

I I

PA,,, 1 .lOE+l 1 .31E+1

- - - -- - - .- - - - .-- - - - .- - - .- -

Notes: Values in column "R13IYR131" mean th I average relative errors between the computed values from the R13II
(the slightly linearized R13) and the computed values from the R13I (the original R13).
 245
Appendix E: Moments computed from the reference distribution f, in
the ES-BGK model

E.l Some basic knowledge

The reference distribution f, in the ES-BGK model is Eq. (2.3), which is rewritten as

I:[
where the vector of peculiar velocity C = C , ,the superscript Ton a vector or a matrix

means the Transpose of the vector or the matrix, matrix

where I is the unit matrix, a is the trace-free pressure tensor, b a parameter that serves
to adjust the Prandtl number, and E = 1-I, the superscript -1 on a matrix means the
inverse matrix, I I on a matrix means the determinate of the matrix.

In order to compute moments from f, at any conditions, we need to define some

matrixes, which are 11 (the modal matrix of matrix of E (normalized eigenvectors),
which is an orthogonal matrix), M (the inverse matrix of 11),E (the spectral matrix of E

(eigenvalues along the principal diagonal), which is a diagonal matrix), vector Y = MC .

The following relations among these matrixes and vectors can be obtained (basic
knowledge about matrix refers to [76, 851)
For the integration, we have

where a > 0 , k = 0,1,2,...,t = &y , and l?( ) is the l? function [86]. Some examples of

Eq. (E.lO): l e - ' ~ ' d ~ = [y2e-ay2dy=-

E.2 Moments computed from f,

Let us consider the zeroth-degree moment, which is

From Eqs. (E.5, E.6, E.9), one get

Since Y, , Y2 and Y3 are independent, therefore

Then utilizing Eq. (E. lo), we get

At last, we get the zeroth-degree moment as

 247
If&= P 9 (E. 12)
which is the first expression in Eq. (2.12).

Let us consider the first-degree moment, which is

(E. 13)

From Eqs. (ES, E.6, E.9), one can get

Since I'IilYl+ I'Ii2Y2+ n,Y, is an odd function of Y, , Y2 and Y,, while

1
~x~(-$E,,Y: - -2E , c is an even function of Yl , Y2 and Y,, therefore,
2
we have
IfEs C , ~ =C0,
which is the second expression in Eq. (2.12).

Let us consider the @-degree moments (where N is an odd number), from the same
idea as the first-degree moment, one get
Ici,...CiNf,dC =0 , (E. 15)
such as
p i c 2f, =o,

Ic,c,c,f,dC=0, (E. 16)

which is the first expression in Eq. (5.45)

Let us consider the second-degree moments, which is

From Eqs. (E.5, E.6, E.9), one get

Since Y, , Y2and Y, are independent, therefore

Then utilizing Eq. (E. lo), one get

jl ni2n j2
= P[- +-
Ell E22

At last, combining Eq. (E.2, E.8), we get

Jc2f,&c= 3 p ,
which is the last expression in Eq. (2.12).

let us consider fourth-degree moments, which is

From Eq. (E.61, one can get

 249
where altogether there are 34 = 81 terms in the expansion, while only terms, which are
even function of Y, , Y, and Y3, will have non-zero value when integration in Eq. (E.20)
is done. Combining Eqs. (E.5, E.9, E.20, E.2 I), one get

After some straightforward but tedious manipulation, we get

n,nll ni2n12
+ P(-
Ell
+- E22
At last, utilizing Eq. (E.8), we have

Let us check the correction of the above equation through consider one specific situation
b = 0.We know at this situation, Aj = RTaV, f, = f, from Eqs. (E.l, E.2), then from

Let us consider some specific fourth-degree moments, which are used in Chapter 5,
from Eq. (E.22),
 Ic,c,c<~c~,f , d ~= p (E.23)

f U d c = P('~~'~ + ' s r 5 + asraSr

) = p(9R2T2+ 2asr& )
In the analysis, if we only consider the linear terms, and omit higher order terms, such as
o p , ,then we have

= R 2 T 2 G , + 2 R T b- a ,
P
b b
&i/Zj,, =.- R ' T ' ~ , , +2RT-a<,, = 2 R T - 5 .
P P
At last, we get

~ C , C , C ~f,dC ) =P
C , = p ( 9 ~ ' ~+' 2 / 2 , ~ , ~ 15-,
P

[c,c~c,c,~,~c=~ =7bRTo,,

which are the second and third expressions in Eq. (5.45).