Multiscale and Multiphysics Flow Simulations of Using The Boltzmann Equation

Jun Li
Multiscale and
Multiphysics
Flow Simulations
of Using the
Boltzmann Equation
Applications to Porous Media and MEMS
Multiscale and Multiphysics Flow Simulations
of Using the Boltzmann Equation
Jun Li
Multiscale and Multiphysics

Flow Simulations of Using
the Boltzmann Equation
Applications to Porous Media and MEMS
123
Jun Li
Center for Integrative Petroleum Research
College of Petroleum Engineering and Geosciences
King Fahd University of Petroleum and Minerals
Dhahran, Saudi Arabia
ISBN 978-3-030-26465-9 ISBN 978-3-030-26466-6 (eBook)

https://doi.org/10.1007/978-3-030-26466-6
© Springer Nature Switzerland AG 2020
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part
of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations,
recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission
or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar
methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt from
the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, expressed or implied, with respect to the material contained
herein or for any errors or omissions that may have been made. The publisher remains neutral with regard
to jurisdictional claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Dedicated to open-minded you.
Preface
This book presents a concise introduction to the kinetic theory, based on which
different simulation methods that were independently developed for solving prob-
lems of different fields can be naturally related to each other. Then, the advantages
and disadvantages of different methods will be discussed with reference to each
other. It mainly covers four advanced simulation methods based on the Boltzmann
equation (i.e., direct simulation Monte Carlo method, direct simulation BGK
method, discrete velocity method, and lattice Boltzmann method) and their appli-
cations with detailed results.
I tried to present an introduction to different theories of fluid dynamics starting
from the traditional macroscopic description (i.e., Navier-Stokes-Fourier equations),
which ends with discussion on its limitation in solving gas flows of high Knudsen
number (Kn), including low-speed gas flows in vacuum systems, micro/nano-
electro-mechanical systems and unconventional rock, as well as supersonic gas
flows. This limitation is usually due to the invalidity of the ordinary constitutive
equations and boundary conditions at high Kn. Subsequently, the microscopic sta-
tistical description based on the Boltzmann equation at the molecular scale as well as
the relevant simulation methods with applications is introduced.
In discussing the existing simulation methods, it was necessary to omit some
topics, which are detailed elsewhere as indicated in the references, to make the book
concise. This omission does not compromise the aim of the book, which is to
provide a basic understanding of the potentials of different simulation methods
based on the Boltzmann equation in solving different problems at multiscale. The
choice of reference coverage certainly reflects a personal preference and I apologize
for any omission of other important and relevant work. This book can serve readers
who are interested in the mathematical derivations of different theories in Chaps. 1
and 2, or the relevant simulation methods and applications in the subsequent
chapters.
Dhahran, Saudi Arabia Jun Li

May 2019
vii
Contents
1 Fluid Mechanics Based on Continuum Assumption . . . . . . . . . . . . . 1

1.1 Continuum Assumption . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Macroscopic Descriptions of Flow Phenomenon . . . . . . . . . . . . . 3
1.3 Derivations of Macroscopic Governing Equations . . . . . . . . . . . . 5
1.3.1 Force and Heat Interactions Between Fluid Parcels . . . . 5
1.3.2 Continuity Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.3 Momentum Equation . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.4 Energy Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.3.5 Entropy Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3.6 Convection-Diffusion Equation . . . . . . . . . . . . . . . . . . . 13
1.4 Transformation to Arbitrary Non-inertial Reference Frame . . . . . 14
1.5 Component Form in Arbitrary Orthogonal Curvilinear
Coordinate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 18
1.6 Limitations of Continuum Assumption . . . . . . . . . . . . . . . . .... 21
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .... 23
2 Boltzmann Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.1 Distribution Function in the Phase Space . . . . . . . . . . . . . . . . . . 25
2.2 Intermolecular Collision Dynamics . . . . . . . . . . . . . . . . . . . . . . 26
2.3 Molecular Models for Transport Coefficients . . . . . . . . . . . . . . . 33
2.4 Derivation of the Boltzmann Equation . . . . . . . . . . . . . . . . . . . . 38
2.5 Calculations of Macroscopic Properties . . . . . . . . . . . . . . . . . . . 41
2.6 Moment Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.7 H-Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.8 Properties at Equilibrium State . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.9 Simplification to Model Equation . . . . . . . . . . . . . . . . . . . . . . . 53
2.10 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.10.1 Specular Reflection Model . . . . . . . . . . . . . . . . . . . . . . 58
ix
x Contents
2.10.2 Diffuse Reflection Model . . . . . . . . . . . . . . . . . . . . . . . 59

2.10.3 CLL Reflection Model . . . . . . . . . . . . . . . . . . . . . . . . . 61
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3 Simulation Methods for Rarefied Gas Flows . . . . . . . . . . . . . . . . . . . 67
3.1 Direct Simulation Monte Carlo Method . . . . . . . . . . . . . . . . . . . 67
3.2 Direct Simulation BGK Method . . . . . . . . . . . . . . . . . . . . . . . . 72
3.2.1 Initialization Process . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
3.2.2 Molecular Motion and Intermolecular Collision . . . . . . . 74
3.2.3 External Body Force . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.4 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 80
3.2.5 Summary of the DSBGK Algorithm . . . . . . . . . . . . . . . 83
3.2.6 Extension to Gas Mixtures . . . . . . . . . . . . . . . . . . . . . . 84
3.3 Discrete Velocity Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.4 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.4.1 Couette Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.4.2 Lid-Driven Cavity Flows . . . . . . . . . . . . . . . . . . . . . . . 92
3.4.3 Thermal Transpiration Through Micro-channel . . . . . . . . 100
3.4.4 Apparent Permeability of Shale Gas . . . . . . . . . . . . . . . 104
3.4.5 Scaling Law of Gas Permeability in the Slip Regime . . . 113
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4 Multiscale LBM Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.1 Basic Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
4.2 Chapman–Enskog Expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.3 Shan–Chen Model for Multiphase Multicomponent Systems . . . . 126
4.3.1 Two-Phase Flows of Single Component . . . . . . . . . . . . 128
4.3.2 Flows of Two Immiscible Components . . . . . . . . . . . . . 131
4.4 Upscaled Lattice Boltzmann Method . . . . . . . . . . . . . . . . . . . . . 132
4.5 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
4.5.1 Large Eddy Simulation of Turbulence . . . . . . . . . . . . . . 134
4.5.2 Intrinsic Permeability . . . . . . . . . . . . . . . . . . . . . . . . . . 138
4.5.3 Buoyant Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
4.5.4 Two-Phase Flow Inside Real Digital Rock . . . . . . . . . . . 148
4.5.5 Upscaled Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . 151
4.6 Introduction to Fortran Code . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
5 Summary and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
Acronyms
BGK Bhatnagar-Gross-Krook
Ca Capillary number, the ratio of viscous force to interfacial tension
CFD Computational fluid dynamics
CFL Courant-Friedrichs-Lewy
D2Q9 Two-dimensional nine-velocity-grids model
DSBGK Direct simulation BGK
DSMC Direct simulation Monte Carlo
DUGKS Discrete unified gas-kinetic scheme
DVM Discrete velocity method
EOS Equation of state
FEM Finite element method
GDVM Godunov-type discrete velocity method
GHS Generalized hard sphere
GSS Generalized soft sphere
HPC High-performance-computing
HS Hard sphere
IFT Interfacial tension coefficient
Kn Knudsen number, the ratio of molecular mean free path to flow
characteristic length
LBM Lattice Boltzmann method
LES Large eddy simulation
LVDSMC Low-variance deviational simulation Monte Carlo
Ma Mach number, the ratio of flow speed to sound speed
MD Molecular dynamics
MEMS Micro-electro-mechanical systems
NEMS Nano-electro-mechanical systems
N–S–C–H Navier–Stokes and Cahn–Hilliard
N–S–F Navier–Stokes–Fourier
N–S–K Navier–Stokes–Korteweg
NTC No-time-counter
xi
xii Acronyms
PDE Partial differential equation

Pr Prandtl number, the ratio of momentum diffusivity rate to thermal
diffusivity rate
Re Reynolds number, the ratio of inertial force to viscous force
Rf Random fraction uniformly distributed between 0 and 1
RHS Right hand side
STP Standard temperature and pressure, i.e., 273.15 K and 1 atm
UGKS Unified gas-kinetic scheme
VHS Variable hard sphere
VSS Variable soft sphere
Chapter 1
Fluid Mechanics Based on Continuum
Assumption
The study of flow problem is very important for many applications and the approach
varies across disciplines. The widely used theory is based on the continuum assump-
tion by using macroscopic quantities (e.g., flow velocity, density, pressure and
temperature) that can be conveniently measured and are close to our ordinary
concepts about the flow problem. Its derivation could be classified into the Eulerian
and Lagrangian descriptions, and the later can be presented in a neat mathematical
form with clear implementations of the basic physical laws (e.g., mass, momentum
and energy conservations) to a closed fluid parcel composed of fixed material par-
ticles, as introduced in this chapter. The obtained system of governing equations
will be closed by introducing the constitutive equations, which are valid for ordinary
flow problems but become inaccurate as the characteristic scale of flow problem or
the density of gas media decreases, where the kinetic theory based on a statistical
description at the molecular scale can be used as discussed at the end of this chapter.
Additionally, the governing equations will be changed in non-inertial reference frame
and the component form in orthogonal curvilinear coordinate will be different from
that in the Cartesian coordinate, which are also discussed.
1.1 Continuum Assumption
Fluid mechanics is a branch of physics and can be divided into fluid statics, i.e.,
the study of fluids at rest, and fluid dynamics, i.e., the study of the effects of force
and heat on fluid motion. The traditional macroscopic theory of fluid mechanics is a
branch of continuum mechanics, which models matter without using the information
at microscopic scale (e.g., molecular position and velocity) [1]. The discussion of
this chapter will focus on the macroscopic description of fluid dynamics problems of
single-component system, where chemical reaction and ionization are neglected for
simplicity. The solution to a fluid dynamics problem typically involves calculating
© Springer Nature Switzerland AG 2020 1
J. Li, Multiscale and Multiphysics Flow Simulations of Using
the Boltzmann Equation, https://doi.org/10.1007/978-3-030-26466-6_1
2 1 Fluid Mechanics Based on Continuum Assumption
various distributions of the fluid properties, including flow velocity, pressure, density
and temperature, as functions of space and time.
All matters are made of atoms, have voids and thus are discrete at microscopic
scale. Consequently, the spatial distributions of physical properties are discontin-
uous. Nevertheless, many flow problems can be modeled as continuous using the
spatial and temporal distributions of macroscopic properties in partial differential
equations (PDEs), which is based on the continuum assumption. Under the contin-
uum assumption, macroscopic properties (e.g., density, pressure, temperature, and
bulk velocity or flow velocity) are taken to be well-defined at infinitesimal volume
elements, which are smaller than the characteristic length of the system but much
larger than the molecular average distance. Fluid macroscopic properties can vary
continuously from one volume element to another and are measured as the average
values of the relevant molecular properties within the volume element concerned [1].
Therefore, when continuum mechanics refers to a mathematical point in a continuous
body, it means an infinitesimal element of the body enclosing that point, instead of a
physical point in the interatomic space. This continuum assumption holds valid down
to a critical scale, above which the body concerned can be continually sub-divided
into smaller infinitesimal volume elements with measured macroscopic properties
of interest being almost unchanged. Further dividing the volume element to make it
much smaller than that critical scale will bring noticeable stochastic noise to the mea-
sured macroscopic properties due to insufficient samples (i.e., insufficient molecules)
within the volume element concerned. Temperature can be measured using the kinetic
energy of molecular random velocity and much more molecule samples are required
to obtain smooth measurement for temperature than for density that is subject only to
the randomness of molecular position when counting the molecular number within
the concerned volume element. Once the required molecular number for smooth
measurement and molecular average distance are known, the critical scale of each
macroscopic property can be determined, e.g., the critical scale will be 400 nm if the
molecular average distance is 4 nm and a million molecules are required to smooth
the particular measurement. The critical scale can be reduced if the time-averaging or
ensemble-averaging process is suitable to increase the samples. On the other hand, if
the size of volume element is very large and comparable to the characteristic length
of the flow problem, the measured macroscopic properties are usually not constant
due to their spatial variations at the macroscopic scale when the volume element is
continually sub-divided.
As we can see from the above discussion, the validity of macroscopic description
of flow problems based on the continuum assumption requires the existence of a
definition scale L d , around which the measurements of macroscopic properties are
definite and independent of the size of volume element, namely without the influence
due to spatial variation at the macroscopic scale or randomness at microscopic scale.
In two-dimensional problems, we can always choose an arbitrary scale much smaller
than the characteristic length as L d and assume the dimension L 3 in the direction
perpendicular to the two-dimensional plane is large enough such that the samples
within the volume element L 2d × L 3 are sufficient to smooth the measurements of
macroscopic properties. In some micro-scale three-dimensional problems, the tran-
1.1 Continuum Assumption 3
sient samples within a volume element L 3d might be insufficient since L d needs to

be much smaller than the characteristic length. Then, macroscopic properties can
be smoothly defined by averaging molecular properties over molecules not only in
each volume element L 3d but also for a certain period of time (i.e., time-averaging
process), or by averaging over molecules in each volume element of a large number
of similar systems (i.e., ensemble-averaging process). The time-averaging approach
is valid for steady state problems and the ensemble-averaging approach can be used
whenever an experiment is indefinitely reproducible. When both of them are valid,
the two averaging approaches can be expected to be identical and the molecular
motion is then deemed to be ergodic.
1.2 Macroscopic Descriptions of Flow Phenomenon
When analyzing the motion of fluid, it is necessary to describe the evolution of con-
figuration (e.g., spatial distribution of material particles) with time. The Lagrangian
description uses time t together with the material coordinate x0 , which usually is
the initial spatial coordinate of material particle, to describe the motion and focuses
on the evolution of each individual material particle with fixed x0 . By contrast, the
Eulerian description takes the motion as dynamic field distributions, which change
with the spatial coordinate x and time t, and gives attention to what is occurring at a
fixed spatial point x as time goes on.
In the Eulerian description, the object of study is an open control volume that usu-
ally has fixed spatial position and shape, through which the fluxes of mass, momentum
and energy are studied according to the balance laws to derive the governing equa-
tions of density, flow velocity and temperature, etc. In the Lagrangian description,
the object of study is fluid parcel, which is a closed and moving system composed
of fixed material particles, and thus the conservation laws can be directly applied to
obtain the governing equations. As a closed system, the mass of fluid parcel remains
constant but its volume may change in compressible flows. Its shape usually changes
due to the distortion by motion. The fluid parcel needs to be small enough so that
the macroscopic properties measured within each fluid parcel can accurately reflect
the spatial variation at the macroscopic scale. It also needs to be large enough so
that its macroscopic properties are not subject to stochastic noise associated with the
random molecular movements at the microscopic scale. The Lagrangian description
will be used in the following derivations.
According to the classical Newton’s laws, the motion of fluid parcel depends on
its external forces: surface force and body forces. Surface force or contact force,
expressed as force per unit area, can act either via the bounding surface of environ-
ment (i.e., external surface force) or an imaginary surface that separates the fluid
parcel from the surrounding ones (i.e., internal surface force). Body forces originate
from sources outside of the fluid parcel and act via force fields (e.g., gravitational
field, electromagnetic field) instead of direct contact. It is usually specified in terms
of force per unit mass or per unit volume. The macroscopic properties carried by
each fluid parcel will change with motion due to force and heat interactions between
adjacent fluid parcels and via external fields.
In Lagrangian description, the transient position x of the concerned material par-
ticle with given x0 can be expressed using a mapping function, i.e., x = f (x0 , t).
Since x0 is the original position, we have f (x0 , 0) = x0 . The mapping function needs
to have various properties so that it is physically admissible (also see [6]):
• one-to-one bijection, so that the inverse mapping function x0 = f −1 (x, t) exists
and thus can trace backwards where the material particle currently located at x
was located at the initial state,
• at lest twice continuously differentiable with respect to t, so that differential equa-
tions describing the motion can be formulated,
• globally invertible at all times, so that the body cannot intersect itself, e.g., the
material particles initially located within an imaginary/material surface cannot
move out of the material surface that has its shape changed constantly during the
motion but will keep closed for material particles at any subsequent instant,
• orientation-preserving, because a body cannot be continuously deformed into its
mirror image.
The Lagrangian description of arbitrary property Q can be transformed into a
Eulerian description using the inverse mapping function, and vice versa:
Q = f 1 (x0 , t) = f 1 ( f −1 (x, t), t) = f 2 (x, t), (1.1)
which will be simply denoted by Q(x0 , t) = Q(x, t) in the following discussion and
Q as a function will take different form when using different independent variables.
The time derivative of arbitrary property for a given material particle with fixed
x0 is called the material derivative or substantive derivative, D/Dt, which can be
thought as being measured by an observer traveling with the concerned material
particle. Note that the transient position x is also a property of material particle and
its material derivative is the instantaneous flow velocity u of the material particle
concerned, namely:
D
x(x0 , t) = u(x0 , t). (1.2)
Dt
Since x is the transient position of material particle x0 at t, Eq. (1.2) indicates that
the new position at t + Δt will be x + Δtu. Then, Eq. (1.1) takes the following form
at t + Δt:
Q(x0 , t + Δt) = Q(x + Δtu, t + Δt), (1.3)
which, according to the chain rule, implies the following transformation of derivatives
between the Lagrangian and Eulerian descriptions:
D ∂ Q(x, t) ∂ Q(x, t)
Q(x0 , t) = +u· , (1.4)
Dt ∂t ∂x
1.2 Macroscopic Descriptions of Flow Phenomenon 5
where the first term on the right-hand side is the local derivative and the second term
is the convective derivative due to material particle changing spatial position.
1.3 Derivations of Macroscopic Governing Equations
Following the rigorous and concise derivations given in [10], the macroscopic gov-
erning equations of flow problems will be derived in Sect. 1.3 using the Lagrangian
description. The obtained governing equations can be transformed into Eulerian
description using Eq. (1.4) if needed.
1.3.1 Force and Heat Interactions Between Fluid Parcels
As mentioned before, the movement of fluid parcel depends on its surface force and
body forces. The surface force can be defined according to the momentum exchange
during the motion through an imaginary surface that separates the concerned fluid
parcel from its surrounding ones. Similar to the fundamental interaction forces (e.g.,
gravitation and electromagnetic forces), the surface force happens between two bod-
ies that need to be closed systems. Thus, it makes sense to discuss the surface force
between fluid parcels. By contrast, it is not appropriate to define the force between
two adjacent open control volumes in the Eulerian description.
To calculate the net surface force of fluid parcel, we divide its closed surface
into many oriented differential surfaces d A. Each d A has an outward unit vector
normal to d A, n+ , and an inward unit normal vector, n− = −n+ . Correspondingly,
the surface force exerted by the surrounding fluid on the concerned fluid parcel via
d A and its counterforce are denoted by dFn+ and dFn− , respectively. Note that the
actual direction of dFn+ has nothing to do with the direction of n+ and will be outward
when the fluid is stretched (if physically possible to fluid), or inward when being
compressed. In non-static state, dFn+ is usually not normal to d A.
We use ei∈[1,2,3] as the unit vectors of Cartesian coordinate system of the reference
frame, where the flow problem is observed or modeled. Surface forces per unit area
or stresses are denoted by Pn+ = dFn+ /d A and Pn− = dFn− /d A=−Pn+ . Usually,
Pn+ can be expressed as a function of velocity gradient using constitutive equation.
But, it will be complicated to calculate the net effect of Pn+ on the whole closed
surface of fluid parcel concerned if the function form depends on n+ . Actually, there
is an inherent correlation between arbitrary Pn+ and the three special ones, i.e., Pe1 ,
Pe2 and Pe3 , which can be specifically expressed using the velocity gradient in a
constitutive equation.
As shown in Fig. 1.1, we study the correlation between stresses of different n+ at
the same spatial point d using a tetrahedron that has one vertex located at the point
d. The tetrahedron has three differential surfaces, d A1 , d A2 and d A3 , located within
the coordinate planes having outward unit normal vectors as −e1 , −e2 and −e3 . The
Fig. 1.1 Schematics of

stress balance; question
marks ‘?’ imply that the
directions of these surface
forces are not necessarily as
shown here
outward
‚ unit normal vector of the fourth differential surface, d An , is denoted by n+ .
Since n+ d A = 0 for arbitrary closed surface, we have:
d An n+ + d A1 (−e1 ) + d A2 (−e2 ) + d A3 (−e3 ) = 0, (1.5)
which implies:
dAj
n +, j = n+ · e j = , ( j = 1, 2, 3). (1.6)
d An
The Newton’s second law applied to the tetrahedron body is [9]:
ρΔV a = ρΔV g + P−e1 d A1 + P−e2 d A2 + P−e3 d A3 + Pn+ d An , (1.7)
where ρ is the mass density, ΔV is the volume of tetrahedron, a is the accel-

eration and g is the net body force (e.g., gravity) per unit mass. Substituting
limd An →0 (ΔV /d An ) = 0, P−e j = −Pe j ( j = 1, 2, 3) and Eq. (1.6) into Eq. (1.7),
we obtain the following inherent relationship of stress balance:
Pn+ = Pe1 n +,1 + Pe2 n +,2 + Pe3 n +,3 . (1.8)
Note that the values of four stresses in Eq. (1.7) might not be taken exactly at the
point d (instead, should be at a point determined by the mean value theorem for
integral) because, for example, we used P−e1 d A1 as the net surface force exerted
via d A1 , within which P−e1 still has possible spatial variation. But, Eq. (1.8) shows
the correlation between surface forces associated with different orientations, whose
values are taken at the same point d because the tetrahedron will shrink to the point
d as d An → 0 with fixed n+ .
1.3 Derivations of Macroscopic Governing Equations 7
According to Eq. (1.8), the stress Pn+ of arbitrary n+ can be generally expressed
using the three special stresses Pe j ( j = 1, 2, 3) measured on the coordinate planes.
Since Pe j ( j = 1, 2, 3) denoted with one subscript are vectors, they can be assembled
into a matrix P denoted without subscript:
⎛ ⎞
P11 e1 e1 P12 e1 e2 P13 e1 e3
P = P ji e j ei = ⎝ P21 e2 e1 P22 e2 e2 P23 e2 e3 ⎠ , (1.9)
P31 e3 e1 P32 e3 e2 P33 e3 e3
where P ji = Pe j · ei are components of P and the convention of Einstein summation

over each dummy index (i.e., a pair of each repeated index) is applied to shorten
the notation, unless stated otherwise. Note that the actual symbol used for a dummy
index is immaterial but the symbol of each dummy index cannot be repeated more
than twice to avoid confusion. Now, Eq. (1.8) can be rewritten into:
Pn+ = n +, j Pe j = n +, j P ji ei = n+ · P. (1.10)
We define the transformation coefficients αi j , which depends on time t in general,

between two arbitrary reference frames, R and R , associated with two Cartesian
coordinate systems, ei∈[1,2,3] and ej∈[1,2,3] :
αi j = ei · ej , (1.11)
which implies:
ei = (ei · ej )ej = αi j ej ,

(1.12)
ej = (ej · ei )ei = αi j ei .
The stress matrix P is an objective quantity and invariant when being observed in
different R. Additionally, to be a tensor, its components measured in different R
(i.e., measurements of using different coordinate systems) must satisfy a certain
correlation. In general, the components of an arbitrary nth-order tensor Q (n ≥ 1)
measured in different reference frames satisfy the following inherent correlation:
Q i1 i2 ···in = Q j1 j2 ··· jn αi1 j1 αi2 j2 · · · αin jn (1.13)
due to Eq. (1.12) and the following objective property:
Q i1 i2 ···in ei1 ei2 · · · ein = Q = Q j1 j2 ··· jn ej1 ej2 · · · ejn . (1.14)
According to the convention used in Eq. (1.9), the measurements of P in the refer-
ence frame R is P j1 j2 = Pej · ej2 , which can be expressed using the measurements
1
of P in R as follows [7]:
P j1 j2 = Pej · ej2

1
= [Pe1 (ej1 · e1 ) + Pe2 (ej1 · e2 ) + Pe3 (ej1 · e3 )] · ej2

= (Pe1 α1 j1 + Pe2 α2 j1 + Pe3 α3 j1 ) · αi2 j2 ei2 (1.15)
= (P1i2 α1 j1 + P2i2 α2 j1 + P3i2 α3 j1 )αi2 j2
= Pi1 i2 αi1 j1 αi2 j2 ,
where Eq. (1.8) has been substituted with n+ being replaced by ej1 . Note that
Eq. (1.12) implies αi j αk j = δik and αi j αik = δ jk , where the Kronecker delta δik equals
1 for i = k or 0 for i = k. Now, Eq. (1.15) can be rewritten as follows:
P j1 j2 αi1 j1 αi2 j2 = Pi3 i4 αi3 j1 αi4 j2 αi1 j1 αi2 j2

= Pi3 i4 δi3 i1 δi4 i2 (1.16)
= Pi1 i2 ,
which proves that the stress P is a second-order tensor according to Eq. (1.13).1 This
conclusion holds at both static and dynamic states and also for the stress of solid
body, where the essential condition of Eq. (1.8) is satisfied as well. Additionally, the
stress tensor can be proved symmetric using a derivation of angular momentum.
The stress as a symmetric second-order tensor can be modeled using other sym-
metric second-order tensors. Newtonian fluid is defined to be a fluid whose shear
stress is linearly proportional to the velocity gradient. This definition means regard-
less of the magnitude of external shear force acting on a fluid, it continues to flow
since the internal shear stress needs to be nonzero according to force balance. In con-
trast, solid can withstand shear forces at static state. The usual constitutive equation
of stress for Newtonian fluids is as follows:

2
Pi j = − p + μv − μ Skk δi j + 2μSi j , (1.17)
3
where p is the thermodynamic pressure, μ is the dynamic viscosity, μv is the second

viscosity or bulk viscosity, and Si j = 21 (∂u j /∂ xi + ∂u i /∂ x j ) is the components of
strain rate tensor S. The mechanical pressure −Pkk /3 = p − μv Skk is not equal to
p in the problems of compressible flows when μv is nonzero. The bulk viscosity
becomes important only when flow compressibility Skk = ∇ · u is significant (e.g.,
adsorption and attenuation of sound). It is mainly related to the vibrational energy of
molecules and zero for monatomic molecules [4]. Additionally, the construction of
constitutive equation should comply with physical restrictions, e.g., the second law
of thermodynamics, which will be discussed in the following governing equation of
entropy.
For a Newtonian fluid, the viscosity depends mostly on temperature and slightly
on pressure, but not on the forces acting upon it. By contrast, stirring a non-Newtonian
1 Similar proof of existence of stress tensor in Sect. 3.3.3 of [6] ends at Eq. (1.10) here.
fluid (e.g., honey) usually causes the viscosity to decrease and so the fluid appears
shear-thinning. Some fluids are shear-thickening, which is usually not observed in
pure materials but can occur in suspensions. Most fluids with long molecular chains
can react in a non-Newtonian manner. Additionally, fluids can be roughly divided
into ideal and non-ideal fluids in some applications. An ideal fluid is inviscid and
thus has no resistance to shear forces, which does not exist in reality but is an
acceptable model for the fluid far away from the solid boundaries since the viscous
effect is usually concentrated near the solid surfaces (i.e., within the boundary layers).
Correspondingly, the reduced momentum equation without the viscous term is called
the Euler equation [1].
In addition to the stress interaction between fluid parcels, there is also heat flux q,
which is specified in terms of thermal energy exchange per unit area per unit time.
According to the Fourier’s law, the usual constitutive equation of q is as follows:
∂T
qi = −ζ , (1.18)
∂ xi
where ζ is the thermal conductivity coefficient, T is the temperature, and the negative
sign means that heat flux moves from higher temperature regions to lower temperature
regions.
1.3.2 Continuity Equation
The mass density ρ and volume ΔV = Δx1 Δx2 Δx3 of a fluid parcel might change
with motion due to non-uniform distribution of the flow velocity u. But, as a closed
system, the total mass of fluid parcel is constant:
D Dρ DΔV
(ρΔV ) = ΔV +ρ = 0, (1.19)
Dt Dt Dt
where ρ, by a slight abuse of notation, should be the average density within ΔV .
The material derivative of ΔV can be calculated as follows:
1 DΔV 1 DΔx1 1 DΔx2 1 DΔx3
= + +
ΔV Dt Δx1 Dt Δx2 Dt Δx3 Dt
1 Dx1 1 Dx2 1 Dx3
= Δ + Δ + Δ (1.20)
Δx1 Dt Δx2 Dt Δx3 Dt
Δu k
= .
Δxk
Substituting Eq. (1.20) into Eq. (1.19), we obtain the following continuity equation
at Δx1 , Δx2 , Δx3 → 0:
Dρ
+ ρ∇ · u = 0, (1.21)
Dt
where ρ is the local density of a particular spatial point, to which the fluid parcel
shrinks as Δx1 , Δx2 , Δx3 → 0. Equation (1.21) holds inside the whole flow domain
because the fluid parcel can be selected arbitrarily. It shows that the density of fluid
parcel doesn’t change as long as ∇ · u = 0 that is the criterion for incompressible
flows. In this sense, gas at the standard temperature and pressure (STP) condition is
very easy to compress but its flow could be deemed incompressible if ∇ · u ≈ 0 due
to very low Mach number (Ma), which is defined as the ratio of the flow speed to the
sound speed. The Eulerian description of continuity equation can be obtained using
Eqs. (1.4) and (1.21):
∂ρ
+ ∇ · (ρu) = 0. (1.22)
∂t
1.3.3 Momentum Equation
The Newton’s second law applied to a fluid parcel has the following form:
˚ ˚ ‹
D
uρdV = gρdV + n+ · Pd A, (1.23)
Dt
where Pn+ = n+ · P of Eq. (1.10) has been substituted, the volume integral and
surface integral are over the volume and closed surface of the fluid parcel concerned,
respectively.
˝ Since the essence of integral is summation, the order of D/Dt and
can be changed and consequently we have D(uρdV )/Dt = uD(ρdV )/Dt +
ρdV Du/Dt = ρdV Du/Dt according to Eq. (1.19). The surface integral of Eq. (1.23)
can be transformed into a volume integral using the Gauss’s theorem (i.e., divergence
theorem) and thus Eq. (1.23) can be rewritten into:
˚
Du
ρ − gρ − ∇ · P dV = 0. (1.24)
Dt
Since Eq. (1.24) holds for arbitrary integral domain occupied by the corresponding
fluid parcel, the integrand must be zero inside the whole flow domain according to
the localization theorem:
Du
ρ = ρg + ∇ · P. (1.25)
Dt
Substituting the constitutive equation of Eq. (1.17) into Eq. (1.25), we obtain the
following Navier–Stokes (N–S) momentum equation:

Du 2
ρ = ρg − ∇ p + μ − μv ∇ · u + ∇ · (2μS). (1.26)
Dt 3
The Eulerian description of momentum equation can be obtained using Eqs. (1.4)
and (1.26):

∂u ∂u 2
ρ +u· = ρg − ∇ p + μ − μv ∇ · u + ∇ · (2μS), (1.27)
∂t ∂x 3
where the left-hand side can be further transformed to ∂(ρu)/∂t + ∇ · (ρuu) accord-
ing to the continuity equation of Eq. (1.22).
1.3.4 Energy Equation
The contributions to the energy change of fluid parcel include the work done by exter-
nal body force, the work done by surface force, and the heat flux from the surrounding
fluid to the fluid parcel concerned. The internal energy (associated with molecular
random translational motion, rotation, vibration and intermolecular potential energy,
but not chemical energy since chemical reaction and ionization are not considered as
stated at the beginning) and kinetic energy (associated with the macroscopic velocity
u) of fluid parcel per unit mass are denoted by e and u 2 /2 = u · u/2, respectively.
Note that similar to the cases of other conservative body forces, the gravitational
potential energy of fluid parcel should be neglected in the total energy when the
work done by gravity is considered in calculating the energy increment. For prob-
lems without volumetric heat sources/sinks, the conservation law of energy applied
to a fluid parcel has the following form:
˚ ˚ ‹
D u2
+ e ρdV = u · gρdV + (n+ · P) · ud A+
Dt 2
‹ (1.28)
−n+ · qd A,
where −n+ · q = n− · q is the heat flux toward the fluid parcel and (n+ · P) · u =
n+ · (P · u). Similar to Eq. (1.23), Eq. (1.28) can be rewritten into:
˚
D u2
ρ + e − u · gρ − ∇ · (P · u) + ∇ · q dV = 0, (1.29)
Dt 2
which holds for arbitrary integral domain occupied by the corresponding fluid parcel
and thus its integrand must be zero inside the whole flow domain:
D u2
ρ ( + e) = u · gρ + ∇ · (P · u) − ∇ · q. (1.30)
Dt 2
Note that the scalar product of u and Eq. (1.25) is:
D u2
ρ = u · gρ + (∇ · P) · u, (1.31)
Dt 2
which can be subtracted from Eq. (1.30) to obtain a simpler form as follows:
D
ρ e = P : ∇u − ∇ · q, (1.32)
Dt
where ∇ · (P · u) = (∇ · P) · u + P : ∇u has been substituted, which means that the

work done by surface force can be divided into the changes of kinetic energy and
internal energy of the fluid parcel concerned. For incompressible flows with Skk = ∇ ·
u = 0, the constitutive equation of Eq. (1.17) degenerates to: Pi j = − pδi j + 2μSi j .
Then, the transformation between the internal energy and the work done by surface
force, i.e., P : ∇u, is equal to the dissipated energy φ per unit volume per unit time,
i.e., P : ∇u = P : S = 2μS : S = 2μSi j Si j = φ ≥ 0, and thus this transformation
is completely irreversible. By contrast, for compressible flows, part of P : ∇u is
associated with the volume change of fluid parcel in a reversible manner without
entropy increase.
The Navier–Stokes–Fourier (N–S–F) equation system consists of three general
conservation equations, i.e., Eqs. (1.21), (1.25) and (1.32), and can be closed by
employing two constitutive equations, e.g., Eqs. (1.17) and (1.18), and two equations
of state, e.g., p = p(ρ, T ) and e = e(ρ, T ). This equation system with appropriate
boundary conditions is a well posed problem but so complicated that analytical solu-
tion is usually not available. The nonlinear convective term u · ∇u contained in the
momentum equation makes most problems difficult to solve and is the main con-
tributor to the turbulence of flows. For most problems, especially those involving
turbulence or complicated boundary geometry, solution of the N–S–F equation sys-
tem can currently only be approximated with the help of computers. This branch of
science is referred to as computational fluid dynamics (CFD). Despite a wide range
of computational applications, the N–S equation alone has not yet been proven in a
purely mathematical sense that its solutions in three-dimensional problems always
exist, or that they are smooth without any mathematical singularity if they do exist.
The Clay Mathematics Institute has called this Navier–Stokes existence and smooth-
ness problem one of the seven most important open problems in mathematics [5].
1.3.5 Entropy Equation
Entropy s of fluid parcel per unit mass is also a state variable (or state function)
of thermodynamics and, as a measurement of disorder, is related to the number of
microscopic configurations that a thermodynamic system can have at an equilibrium
state specified by several macroscopic state variables. For a closed system, entropy
will increase with temperature if the volume is fixed because the set of molecular
velocities has more admissible configurations at higher temperature. At a constant
temperature, the number of admissible configurations of molecular position increases
with the volume and thus entropy increases as well. This interpretation of entropy
will become clearer by using a probability distribution function defined inside a six-
dimensional phase space (a superposition of the unbounded molecular velocity space
and the physical space), which will be introduced in the next chapter.
For usual fluid systems of single component, there are only two independent state
variables and thus the material derivative of s can be calculated by applying the chain
rule to e(s, ρ):

De ∂e Ds ∂e Dρ
ρ =ρ |ρ + |s
Dt ∂s Dt ∂ρ Dt
(1.33)
Ds
= ρT − p∇ · u,
Dt
where two equations of thermodynamic functions, i.e., (∂e/∂s)|ρ = T and

(∂e/∂ρ)|s = p/ρ 2 , have been substituted, which presumes that the local dynamic
flow process is thermodynamically
√ quasi-static.√Additionally, the sound speed is
defined as csound = (∂ p/∂ρ)|s , which is equal to 7 p/(5ρ) for ideal gas of diatomic
molecules with e = 5 p/(2ρ).
Substituting Eq. (1.33) into Eq. (1.32), we obtain:
Ds
ρT = p∇ · u + P : ∇u − ∇ · q, (1.34)
Dt
where the right-hand side degenerates to φ = 2μSi j Si j ≥ 0 when ∇ · u = 0, q = 0
and P is modeled by Eq. (1.17), which means that the constitutive equation of
Eq. (1.17) complies with the second law of thermodynamics in the cases of incom-
pressible flows. For compressible flows with q = 0, the right-hand side of Eq. (1.34)
is equal to (μv − 2μ/3)(∇ · u)2 + φ when using Eq. (1.17) and then attentions are
needed to make it always positive.
1.3.6 Convection-Diffusion Equation
In some problems, we need to study the concentration evolution of a species (solute),

which is dissolved in the fluid (solvent), due to the processes of convection and
diffusion. The mole concentration of the species concerned is denoted by C (mol
m−3 ) and the diffusive flux of species, J (mol m−2 s−1 ), is modeled as follows
according to the Fick’s first law:
J = −D∇C, (1.35)
where D (m2 s−1 ) is the diffusion coefficient. We assume that C is sufficiently low
and thus its influences on ρ and u are negligible. For problems without volumetric
sources/sinks of the species concerned, the conservation law of substance amount
applied to the domain occupied by a fluid parcel has the following form:
˚ ‹
D
CdV = −n+ · Jd A. (1.36)
Dt
Substituting D(dV )/Dt = (∇ · u)dV of Eq. (1.20) into Eq. (1.36), we obtain:
DC
+ C∇ · u = −∇ · J. (1.37)
Dt
Note that the material particles of fluid will not move across the closed imaginary
surface of fluid parcel during the motion when selecting fluid parcel as the object of
study but the species concerned here is allowed to freely penetrate the surface.
1.4 Transformation to Arbitrary Non-inertial Reference

Frame
The N–S–F equation system is independent of any particular inertial reference frame
but its form will change in non-inertial reference frame. Most of physical quantities
contained in this equation system are objective, including P, q, J, ρ, p, T , e, s and
C. In contrast, the spatial position x, flow velocity u and acceleration a of material
particle depend on the reference frame R, where the flow problem is observed.
Additionally, the constitutive equations of Eqs. (1.17), (1.18) and (1.35) contain other
quantities S, ∇ · u, ∇T , and ∇C whose objectivities determine the objectivities of
these constitutive equations.
In the inertial reference frame R having the origin of coordinates located at the
−
→
point o, we use x for the vector od, where the point d is an arbitrary material particle
(or spatial point) concerned, and have the component form x = xi ei , where ei∈[1,2,3]
are the unit vectors of Cartesian coordinate system of the reference frame R. The
origin of coordinates of the arbitrary non-inertial reference frame R is located at o
−→ −
→
and vector o d is denoted by X. Additionally, let y denote the vector oo and thus:
y = x − X. (1.38)
The velocities of the same material particle observed in two different reference
frames R and R are two different quantities, denoted by u and U, respectively.
They are not equal when the non-inertial R has translational or rotational motion
relative to the inertial R and thus the velocity is variant. But, the divergences of
the two different velocities are always equal, namely ∂Uk /∂ X k ≡ ∂u k /∂ xk , which
implies that the velocity divergence is objective. Additionally, the strain rate tensor
1.4 Transformation to Arbitrary Non-inertial Reference Frame 15
and the gradient of temperature or any other objective scalar quantities are objective
as well, but the vorticity and gradient of velocity are variant. Note that when talking
about whether a quantity is invariant/objective, we compare two counterparts defined
equally but in different reference frames, e.g., compare U with u in discussing the flow
velocity, or compare ∂Uk /∂ X k with ∂u k /∂ xk in discussing the velocity divergence.
For a given vector, which is objective and thus independent of the reference frame,
it has different component forms when being measured in different reference frames:
x = xi ei = x j ej for example, where ej∈[1,2,3] are the unit vectors of Cartesian coor-
dinate system of the reference frame R . Since the Cartesian coordinate system is
used, ei and ej do not change with coordinates when observing the movement of
material point in R and R , respectively. Then, the flow velocity u of material particle
d observed in R can be simply calculated using the corresponding component form
of x, namely u = (Dxi /Dt)ei . Similarly, the flow velocity U of the same material par-
ticle d observed in R can be calculated as U = (DX j /Dt)ej . Note that (Dx j /Dt)ej
is the velocity of d relative to o observed in R , which is different from U that is the
velocity of d relative to o observed in R and also different from u that is the velocity
of d relative to o observed in R. Similarly, (DX i /Dt)ei = u and (DX i /Dt)ei = U.
Note that the temporal variation of a given vector might be different when being
observed in different reference frames. Thus, in Sect. 1.4, we apply the material
derivative only to scalar quantities (i.e., components of tensors) to avoid confusions
incurred when using it directly for tensors. For example, if Dx/Dt and DX/Dt are
used to denote u and U, respectively, we will literally obtain u = U if x ≡ X, which
is wrong because R can still have rotational motion relative to R making u = U
although their origins of coordinates remain overlapped, i.e., x ≡ X.
In general, we have the following correlation between u and U:
Dxi D(yi + X i )
u= ei = ei
Dt Dt
(1.39)
Dyi D(αi j X j ) Dyi Dαi j
= ei + ei = ei + U + X ei ,
Dt Dt Dt Dt j
where (Dyi /Dt)ei and (Dαi j /Dt)X j ei correspond to the translational and rotational
motions of the reference frame R relative to R, respectively. We introduce ω = ωi ei
to make the physical meaning of (Dαi j /Dt)X j ei clear and define ωi as follows:
1 Dαl j
ωi =
lik αk j , (1.40)
2 Dt
where the Levi-Civita symbol is defined as:

i jk = 0 if (i − j)(i − k)( j − k) =
0,
i jk = 1 if (i, j, k) ∈ {(1, 2, 3), (2, 3, 1), (3, 1, 2)}, and
i jk = −1 if (i, j, k) ∈
{(1, 3, 2), (3, 2, 1), (2, 1, 3)}. It can be proved that
i jk ω j = αkm Dαim /Dt using
i jk
l jn = δil δkn − δin δkl and thus:
Dαi j Dαi j
X ei = αk j X k ei =
i jk ω j X k ei = ω × X, (1.41)
Dt j Dt
where ω × X is the cross product of ω and X. Equation (1.41) shows that ω defined
by Eq. (1.40) is the rotational/angular velocity of R relative to R.
Note that yi , y j , αi j , ei and ej are independent of x and X in the Cartesian
coordinate system, which can simplify the calculation of each quantity and thus is
used in R and R for the convenience of discussion. Additionally, Eq. (1.12) implies
x j = αi j xi and xi = αi j x j . For the velocity divergence, we have:

∂ Dyi Dαi j
∇ · u = ek · ei + U + X j ei
∂ xk Dt Dt
∂U j ∂ X j Dαi j
= 0 + ek · e j + ek · ei
∂ xk ∂ xk Dt
∂ X i ∂U j ∂(x j − y j ) Dαi j
= αk j + e k · ei
∂ xk ∂ X i ∂ xk Dt
∂(xi − yi ) ∂U j Dαi j
= αk j + ek αk j · ei − 0
∂ xk ∂ Xi Dt (1.42)
∂U j Dαi j
= αk j αki − 0 + αi j
∂ X i Dt

∂U j 1 D(αi j αi j )
= δ ji +
∂ Xi 2 Dt

∂U 1 Dδ
= ei · ej +
j jj
∂ X i 2 Dt
= ∇ · U,
where ∇ ≡ ∂/∂X = (∂/∂ X i )ei . Since ∇ · u ≡ ∇ · U, the velocity divergence is

objective.
For the gradient of T or arbitrary objective scalar quantity, we have:
∂T ∂ X j ∂ T
∇T = ei = ei
∂ xi ∂ xi ∂ X j
(1.43)
∂(x j − y j ) ∂ T ∂T ∂T
= ei = αi j ei = e = ∇ T.
∂ xi ∂Xj ∂Xj ∂ X j j
Since ∇T ≡ ∇ T , the gradient of temperature is objective as well. Additionally, we

can prove ∇u = ∇ U + αk j (Dαi j /Dt)ek ei , which shows that the gradient of velocity
is variant. But, it can be further proved that (∇u + ∇uT ) ≡ (∇ U + ∇ UT ), where
∇uT is the transpose of ∇u. Thus, the strain rate tensor is also objective. Similarly,
we can prove ∇ × u = ∇ × U + 2ω, which means that the vorticity of velocity is
variant. The constitutive equations of Eqs. (1.17), (1.18) and (1.35) contain only
objective quantities and thus have invariant forms in both inertial and non-inertial
1.4 Transformation to Arbitrary Non-inertial Reference Frame 17
reference frames although their forms are usually validated in inertial reference frame
by experiments. Detailed derivations and discussion are given in [7].
The N–S momentum equation of Eq. (1.26) was derived in Sect. 1.3.3 using the
Newton’s second law in an inertial reference frame R. According to the above objec-
tivity analyses of different quantities, the objective quantities of Eq. (1.26) defined in
R can be replaced by the corresponding quantities defined in R and thus Eq. (1.26)
can be rewritten into:
Du i 2
ρ ei = ρg − ∇ [ p + ( μ − μv )∇ · U] + ∇ · [μ(∇ U + ∇ UT )], (1.44)
Dt 3
where we substituted ∇ · [μ(∇ U + ∇ UT )] = ∇ · [μ(∇ U + ∇ UT )], which can be

similarly proved as in Eq. (1.43).
Note that (Du i /Dt)ei is the acceleration of material particle concerned, which is
measurable in R but not in R . To obtain the momentum equation for R , we need
to make sure that all quantities contained in the equation are either measurable in R
or the properties of R . These properties of R are measured in R and independent
of X since they are associated with the whole reference frame R . Similarly, the
acceleration of material particle observed in R can be calculated as (DU j /Dt)ej .
The connection between this two accelerations of the same material particle observed
in R and R , respectively, is as follows:
Du i D2 yi DUi D(
i jk ω j X k )
ei = ei + ei + ei
Dt Dt 2 Dt Dt
D2 yi D(αi j U j ) Dω j DX k
= ei + ei +
i jk X k ei +
i jk ω j ei
Dt 2 Dt Dt Dt
D2 yi DU j Dαi j Dω j DX k
= ei + ej + αk j Uk ei + e j × X +
i jk ω j ei
Dt 2 Dt Dt Dt Dt
D2 yi DU j Dω j DX k
= ei + e +
i jk ω j Uk ei + e j × X +
i jk ω j ei
Dt 2 Dt j Dt Dt

D2 yi DU j Dω j DX k
= ei + e +ω×U+ e j × X +
i jk ω j ei , (1.45)
Dt 2 Dt j Dt Dt
where (D2 yi /Dt 2 )ei is the translational acceleration of R relative to R and
DX k D(αkl X l )

i jk ω j ei =
i jk ω j ei
Dt Dt
Dαkl
=
i jk ω j αml X m + αkl Ul ei (1.46)
Dt
=
i jk ω j (
klm ωl X m + Uk )ei
= ω × (ω × X) + ω × U.
Substituting Eqs. (1.45) and (1.46) into Eq. (1.44), we obtain the momentum equation,
which is valid in an arbitrary non-inertial reference frame R . Similar transformations
can be applied to other governing equations.
1.5 Component Form in Arbitrary Orthogonal Curvilinear

Coordinate
In the above Sect. 1.4, we discussed the replacement of a quantity defined in an iner-
tial reference frame R by its counterpart defined in an arbitrary non-inertial reference
frame R . This two counterparts will have the same form if the quantity concerned
is objective (e.g., divergence of velocity and strain rate tensor) but additional terms
are needed if not (e.g., acceleration in Eq. (1.45) and vorticity). By contrast, in this
section, the discussion will focus on the transformation between different compo-
nent forms of each given quantity contained in the governing equations of interest
such that the calculations of those quantities can be simplified in axisymmetric or
spherically symmetric flows by using non-Cartesian coordinate system. Note that
this transformation between component forms is also applicable to the governing
equations of R , e.g., Eqs. (1.44)–(1.46), although the quantities defined in R will
be used as examples in the following discussion that is based on the derivations of
[3].
For a given reference frame where the flow problem is observed, it is still possible
to use different coordinate systems for measurement to obtain different component
forms of each quantity of interest. As shown in Fig. 1.2, in the arbitrary orthogonal
curvilinear coordinate system S̃ of using (l1 , l2 , l3 ) as coordinates and (ẽ1 , ẽ2 , ẽ3 )

arbitrary orthogonal
curvilinear coordinate
system S̃ of using (l1 , l2 , l3 )
as coordinates and
(ẽ1 , ẽ2 , ẽ3 ) as coordinate
vectors; ẽi = ∇li /|∇li | for
given li (x1 , x2 , x3 ) or
ẽi = (∂x/∂li )/|∂x/∂li | for
given xi (l1 , l2 , l3 ), where the
repeated i does not imply
summation
1.5 Component Form in Arbitrary Orthogonal Curvilinear Coordinate 19
as coordinate vectors, each spatial point has three coordinates as in the Cartesian
coordinate system S of using (x1 , x2 , x3 ) as coordinates and (e1 , e2 , e3 ) as coordinate
vectors. There is a one-to-one bijection between (x1 , x2 , x3 ) and (l1 , l2 , l3 ) and this
correlation is given once S̃ is selected according to S. ẽi∈[1,2,3] are orthogonal at
arbitrary point but change with (l1 , l2 , l3 ), which is different from the case of S.
−→ −−→ −−→ −→
The increments of l1 associated with bb1 , b2 d3 , b3 d2 and d1 d are the same but
the corresponding spatial distances are usually different. Correspondingly, the Lame
coefficient Hi=1 (similar for i = 2, 3) at arbitrary point b is defined as follows:
H1 = lim (bb1 /Δl1 )

Δl1 →0

Δx12 + Δx22 + Δx32

= lim (1.47)
Δl1 →0
Δl2 =Δl3 =0
Δl1

∂ x1 2 ∂ x2 2 ∂ x3 2
= ( ) +( ) +( ) .
∂l1 ∂l1 ∂l1
Thus, we have (H1 , H2 , H3 ) = (1, r, 1) in the cylindrical coordinate system with

(l1 , l2 , l3 ) = (r, θ, z), and (H1 , H2 , H3 ) = (1, r sin ϕ, r ) in the spherical coordinate
system with (l1 , l2 , l3 ) = (r, θ, ϕ), where θ is the azimuthal angle and ϕ is the polar
angle.
The coordinate vectors ẽi∈[1,2,3] change with (l1 , l2 , l3 ) as shown in Fig. 1.3, where
−−→
b is the reference point. Then, d4 d1 = Δl3 ∂ ẽ2 /∂l3 , and d4 d1 = Δl2 ∂(H3 Δl3 )/∂l2
since bb3 = H3 Δl3 . By neglecting the higher-order infinitesimals, the corresponding
length ratios are equal for the two similar triangles:

variations of ẽ1 , ẽ2 , and ẽ3
with (l1 , l2 , l3 )
|Δl3 ∂ ẽ2 /∂l3 | Δl2 ∂(H3 Δl3 )/∂l2 1 ∂ H3

= = Δl3 , (1.48)
|ẽ2 | H2 Δl2 H2 ∂l2
which indicates that the magnitude of ∂ ẽ2 /∂l3 is equal to (1/H2 )∂ H3 /∂l2 as |ẽ2 | = 1.
Additionally, the direction of ∂ ẽ2 /∂l3 is the same as ẽ3 since ẽi∈[1,2,3] are always
orthogonal. By conducting similar analyses for other cases, we have the following
general formula:
∂ ẽi 1 ∂ Hj
= ẽ j , (i = j and i, j ∈ [1, 2, 3]), (1.49)
∂l j Hi ∂li
where the repeated indices do not imply summations. Substituting ẽi = ẽ j × ẽk with
(i, j, k) = (1, 2, 3), (2, 3, 1) or (3, 1, 2) into Eq. (1.49) and using the chain rule, we
obtain:
∂ ẽi 1 ∂ Hi 1 ∂ Hi
=− ẽ j − ẽk , (1.50)
∂li H j ∂l j Hk ∂lk
where the repeated indices do not imply summations.

Now, the calculation of operator ∇ in S̃ can be derived below. For arbitrary scalar
quantity Q, the gradient ∇ Q is a vector and determined once its three components
along the orthogonal coordinate vectors ẽi∈[1,2,3] are known. According to the def-
inition of directional derivative, ∇ Q · ẽi is the change rate of Q along the oriented
−
→
coordinate curve li and thus can be calculated as follows:
∇ Q · ẽi = lim (ΔQ/bbi )

Δli →0
Δl j =i =0
Δli ∂ Q/∂li
= lim (1.51)
Δli →0 Hi Δli
1 ∂Q
= ,
Hi ∂li
where the repeated i does not imply summation. Equation (1.51) indicates that
∇ Q = (1/H1 )(∂ Q/∂l1 )ẽ1 + (1/H2 )(∂ Q/∂l2 )ẽ2 + (1/H3 )(∂ Q/∂l3 )ẽ3 and the gra-
dient operator can be implemented as:
1 ∂ 1 ∂ 1 ∂
∇ = ẽ1 + ẽ2 + ẽ3 . (1.52)
H1 ∂l1 H2 ∂l2 H3 ∂l3
Equations (1.49), (1.50) and (1.52) can be used to calculate different quantities in
the governing equations using component forms. As an example, we can obtain the
following formulas for the divergence of velocity u = ũ i ẽi :
1.5 Component Form in Arbitrary Orthogonal Curvilinear Coordinate 21
1 ∂u 1 ∂u 1 ∂u
∇ · u = ẽ1 · + ẽ2 · + ẽ3 ·
H1 ∂l1 H2 ∂l2 H3 ∂l3
(1.53)
1 ∂(ũ 1 H2 H3 ) ∂(ũ 2 H3 H1 ) ∂(ũ 3 H1 H2 )
= + + ,
H1 H2 H3 ∂l1 ∂l2 ∂l3
which has the following form in the cylindrical coordinate system (l1 , l2 , l3 ) =
(r, θ, z):

1 ∂(ũ 1r ) ∂ ũ 2 ∂(ũ 3r ) 1 ∂ ũ 1 ∂ ũ 2 ∂ ũ 3
∇ ·u= + + = ũ 1 + r + +r ,(1.54)
r ∂r ∂θ ∂z r ∂r ∂θ ∂z
and has the following form in the spherical coordinate system (l1 , l2 , l3 ) = (r, θ, ϕ):

1 ∂ ũ 1 ∂ ũ 2 ∂ ũ 3
∇ ·u = 2 2ũ 1 r sin ϕ + r 2 sin ϕ +r + ũ 3 r cos ϕ + r sin ϕ . (1.55)
r sin ϕ ∂r ∂θ ∂ϕ
1.6 Limitations of Continuum Assumption
There are some problems, for example the Brownian motion of very small granule
in fluid, where the stochastic interaction associated with the intermittent molecular
movements is the dominant effect to model and thus the deterministic continuum
description is invalid. In this kind of problems, the object of study is the statistical
properties of stochastic quantity that is associated with the molecular-level activity,
instead of chaotic quantity, e.g., sound noise due to pressure oscillation at high
Reynolds number (Re), that still can be described by the time-dependent governing
equations of continuum description although it is challenging to obtain the numerical
solution.
Additionally, the continuum description of flow problems usually requires consti-
tutive equations, e.g., Eqs. (1.17), (1.18) and (1.35), to close the system of governing
equations. These ordinary constitutive equations become inaccurate in applications
like supersonic gas flows, gas flows in vacuum system, or gas flows of micro/nano
scale (e.g., shale gas flows at the pore scale or gas flows inside micro/nano-electro-
mechanical systems, i.e., MEMS/NEMS). This non-equilibrium phenomenon can
be easily understood in the Couette flow whose right boundary moves vertically in
the positive x2 direction at speed u 2,wall and left boundary moves vertically in the
the negative x2 direction at speed u 2,wall . When the molecular mean free path λ,
i.e., the average distance traveled by a moving molecule between its two successive
intermolecular collisions as shown in Fig. 1.4, is much larger than the distance of the
two boundaries, the intermolecular collisions become immaterial and the solution
depends mostly on the molecular reflections at the boundaries, namely boundary
condition. In this case, the molecules inside arbitrary volume element can be catego-
Fig. 1.4 Simplified illustration of calculating the mean free path λ, which is the average length of
all possible red free distances associated with random molecular velocity
Could be Burnett Eq.

+ high-order-slip B.C.
N-S Eq. + non-slip B.C. N-S Eq. + slip B.C. Simplified Boltzmann Eq.
Continuum regime Slip regime Transitional regime Free molecular
Fig. 1.5 Regime division of gas flows according to the Knudsen number K n
rized into two groups according to the component of molecular velocity c in the x1
direction: c1 > 0 for molecules moving towards the right boundary after reflecting on
the left boundary, and c1 < 0 for molecules moving towards the left boundary after
reflecting on the right boundary. The two groups have the same molecular number
on average according to mass balance at steady state. Assuming the complete diffuse
reflection at the boundaries, the average (c2 )c1 >0 of c2 equals −u 2,wall for the first
group while the second group has (c2 )c1 <0 = u 2,wall . Consequently, the flow velocity
u 2 (i.e., average of c2 over all molecules inside the volume element concerned) is
zero at arbitrary location and thus the strain rate tensor S is zero. On the other hand,
molecules carrying (c2 )c1 >0 = −u 2,wall will get (c2 )c1 <0 = u 2,wall after reflecting on
the right boundary and thus the shear stresses at the boundaries are nonzero, which
indicates that the shear stress inside the fluid is also nonzero according to force bal-
ance. Obviously, the linear constitutive relationship of Eq. (1.17) breaks down in this
case of free molecular regime with zero S and nonzero P. It is inaccurate although not
completely wrong in moderately non-equilibrium problems, where the intermolec-
ular collision is as important as the molecular reflection at the boundary and both
S and P have nonzero off-diagonal components. This non-equilibrium phenomenon
will be studied in Sect. 3.4.1.
These flow problems with non-equilibrium effect can be simulated using the
statistical mechanics theory based on the nonlinear integro-differential Boltzmann
1.6 Limitations of Continuum Assumption 23
equation. The Knudsen number (K n), which is defined as the ratio of λ to the char-
acteristic length L, can be used to determine whether or not the ordinary constitutive
equations and boundary conditions hold. As shown in Fig. 1.5, the gas flows can
be divided into different regimes according to K n = λ/L: (1) continuum regime
with K n < 0.01, where the N–S–F equation system with usual constitutive equa-
tions and boundary conditions is valid; (2) slip regime with 0.01 < K n < 0.1, where
complicated velocity-slip and temperature-jump boundary conditions are needed; (3)
transitional regime with 0.1 < K n < 10, where the Boltzmann equation of a statis-
tical description at the molecular level is necessary; (4) free molecular flow regime
with 10 < K n, where the Boltzmann equation can be significantly simplified and
analytical solutions are sometimes available. Note that the molecular size of methane
(CH4 ) is about 0.37 nm, the molecular average distance and λ at STP are about 3.46
and 50 nm, respectively, which implies that K n < 0.01 holds in usual problems
with L > 5 microns but is not satisfied in the pore-scale shale gas flows, where the
dominant pore size could be as small as 10 nm, although λ ∝ 1/ p could be much
smaller than 50 nm since the pore pressure p far away from the production wells is
much higher than the standard ambient pressure. In addition to the modifications of
boundary conditions, changing the constitutive equations as in the Burnett equation
can further extend the validity of continuum description up to K n < 1. By contrast,
Boltzmann-like equations can be applied in the whole range of K n [1, 2, 8].
References
1. Batchelor GK (1967) An introduction to fluid dynamics. Cambridge University Press, Cam-

bridge
2. Chapman S, Cowling TG (1970) The mathematical theory of non-uniform gases, 3rd edn.
Cambridge University Press, Cambridge
3. Cao SL (2005) Flow theory on fluid machinery. Lecture in Tsinghua, Beijing
4. Cramer MS (2012) Numerical estimates for the bulk viscosity of ideal gases. Phys Fluids
24:066102
5. Fefferman CL (2017) Existence and smoothness of the Navier-Stokes equation. Available via
Clay Mathematics Institute. http://www.claymath.org/sites/default/files/navierstokes.pdf
6. Gonzalez O, Stuart AM (2008) A first course in continuum mechanics. Cambridge University
Press, Cambridge
7. Li J (2014) Analysis on the invariant properties of constitutive equations of hydrodynamics in
the transformation between different reference systems. https://arxiv.org/abs/1401.0453
8. Shen C (2005) Rarefied gas dynamics: fundamentals, simulations and micro flows. Springer,
Berlin
9. Zhang ZS, Cui GX (2006) Fluid mechanics, 2nd edn. Tsinghua University Press, Beijing
10. Zhuang LX, Yin XY, Ma HY (2009) Fluid mechanics, 2nd edn. Press of University of Science
and Technology of China, Hefei
Chapter 2
Boltzmann Equation
Abstract When the mean free path of gas molecules becomes comparable to the
characteristic flow length (i.e., Knudsen number is not negligible), the traditional
Navier–Stokes–Fourier (N–S–F) equations fail and the kinetic theory could be used
to study the gas flows. This will occur due to either large mean free path at low
pressure or small characteristic length at micro/nano-scales. The distribution func-
tion and Boltzmann equation are introduced in this chapter, as the foundation of the
kinetic theory. Different intermolecular collision models are discussed to show the
selection of kinetic molecular models according to the macroscopic transport coeffi-
cients, which is based on the result of Chapman–Enskog expansion. The derivation
of N–S–F-like equation by computing moments of the Boltzmann equation is given
to show their correlation at the continuum regime. The entropy and second law of
thermodynamics can be better understood by using the H-theorem of the Boltzmann
equation. The definitions of mean free path and intermolecular collision frequency
can be obtained as properties of the Maxwell equilibrium distribution. Model equa-
tion is also introduced as a good approximation to the Boltzmann equation at low
speed. At the end, different boundary conditions are discussed together with their
implementation algorithms in Monte Carlo molecular simulations.
2.1 Distribution Function in the Phase Space
The Boltzmann equation was proposed to describe the statistical behavior of gas flow
problems and is the foundation of kinetic theory of gases. In contrast to deterministic
theory, e.g., the molecular dynamics (MD) method, where the velocity and position of
each molecule/particle are accurately studied according to the Newton’s laws, it con-
siders a probability distribution function of molecules in the six-dimensional phase
space (a superposition of the unbounded molecular velocity space c and the physical
space x). This statistical approach makes it computationally affordable to model sys-
tems with huge number of molecules as in usual problems (note: the number of ideal
gas molecules in a cubic centimeter at STP is about 2.687 × 1019 , i.e., the Loschmidt
number [21]) and can be justified by the fact that the quantities of interest in flow
problems usually are average properties of molecular variables, including density,

26 2 Boltzmann Equation
flow velocity, temperature, stress and heat flux. It models polyatomic molecule as
a particle without intramolecular structure but uses additional variables other than
c for molecular rotation and vibration. Also different from the Liouville equation
[17], the Boltzmann equation takes all molecules of the same chemical component
identical and then can employ a one-particle distribution function f (c, x, t) for the
molecular number density in the phase space. Accordingly, f dcdx is the molecular
number inside the differential element dcdx around (c, x) at the moment t. To be
more accurate, f dcdx is the probabilistic number of molecules defined using the
expected value of ensemble/time average. The molecular number inside dx at t is
R3 f dcdx and thus the molecular number density in the physical space is:

n(x, t) = f dc. (2.1)
R3
Mass density can be computed as ρ = nm, where m is the molecular mass. For a
mixture system, each chemical species has an independent one-particle probability
distribution function in question. For simplicity, we focus on the single-component
system without chemical reaction, ionization, and sources/sinks of the component
concerned. The derivations of [5, 21] are adopted in this chapter, but small modifi-
cations and more discussion are added here.
2.2 Intermolecular Collision Dynamics
Before discussing the change of f due to intermolecular collisions, we first introduce

the binary collision model and assume that the collisions involving more than two
molecules have negligible influence, which is reasonable when the density is not
very high because the probability of having ternary collision is about (D/δ)3 [21],
where the molecular diameter D 1 is usually much smaller than molecular average
distance δ = n −1/3 (e.g., DCH4 ≈ 0.37 nm and δCH4 ≈ 3.46 nm at STP). Note that this
probability is the conditional probability of having a ternary collision between the
pair of temporally encountered molecules and a third molecule, given that the binary
collision between the pair concerned has happened. This conditional probability is
estimated by the product of the lifetime (overlapping time) of the collision pair and
the collision frequency of each molecule, where the temporally encountered collision
pair is deemed a special molecule to estimate the probability of its collision with a
third molecule. Thus, (D/δ)3 1 means that the ternary collision is not important
compared to the binary collision.
To simplify the mathematical analysis, the intermolecular collisions are treated as
elastic such that the momentum and energy associated with the translational motion
before and after the collision are the same, respectively. We denote the pre-collision
1 By a slight abuse of notation, D denotes the diffusion coefficient introduced in Sect. 1.3.6 and the
molecular diameter for the intermolecular collision process introduced here.

2.2 Intermolecular Collision Dynamics 27
velocities by cp1 and cp2 of two colliding molecules and use cp1 and cp2 for the
post-collision velocities. The momentum and energy conservation are as follows:
m p1 cp1 + m p2 cp2 = m p1 cp1 + m p2 cp2 = (m p1 + m p2 )cm ,

(2.2)
m p1 cp21 + m p2 cp22 = m p1 (cp 1 )2 + m p2 (cp 2 )2 ,
where cm defined here is the velocity of their mass center, the light-face symbol
cp 1 is the magnitude of cp1 (the same for other vectors). In the reference frame of
mass center moving at cm , the observed movements of two molecules are antiparallel
before and after the intermolecular collision according to the following relationships:
m p2 mr
cp1 − cm = cr = cr ,
m p1 + m p2 m p1
−m p1 −m r
cp2 − cm = cr = cr ,
m p1 + m p2 m p2
m p2 mr (2.3)
cp1 − cm = c = c,
m p1 + m p2 r m p1 r
−m p1 −m r
cp2 − cm = c = c,
m p1 + m p2 r m p2 r
where m r = m p1 m p2 /(m p1 + m p2 ) is the reduced mass, cr = cp1 − cp2 and cr = cp1 −
cp2 . We assume that the molecules are the centers of intermolecular force and then
their trajectories are located inside a plane (i.e., collision plane) as shown in Fig. 2.1,
where b2 is the projected distance of the undisturbed trajectories of pre-collision
velocities and the two deflection angles χp2 and χp1 are equal since the trajectories
are antiparallel before and after the intermolecular collision.
From Eq. (2.3) we obtain:
m p1 cp21 + m p2 cp22 = (m p1 + m p2 )cm

2
+ m r cr2 ,
(2.4)
m p1 (cp 1 )2 + m p2 (cp 2 )2 = (m p1 + m p2 )cm
2
+ m r (cr )2 .
The energy conservation of Eq. (2.2) indicates cr = cr in Eq. (2.4). The conserva-
tion of angular momentum requires cr b = cr b , which indicates that the projected
distance b of the post-collision velocities is equal to b. The symmetric properties of
Eqs. (2.2)–(2.4) with respect to the pre-collision and post-collision velocities imply
that a collision, which is called the inverse collision of the original direct collision,
between two molecules having cp1 and cp2 at the beginning will end up with cp1
and cp2 if the pre-collision positions are adjusted appropriately to make the collision
plane and b unchanged, etc.
2 By a slight abuse of notation, b denotes an arbitrary reference point introduced in Sect. 1.5, the
projected distance of undisturbed trajectories for the intermolecular collision process introduced
here, and the slope parameter of the Klinkenberg correlation model introduced in Sect. 3.4.4.
parcle 1
parcle 2
Fig. 2.1 Trajectories of a binary collision observed in the reference frame moving together with
the mass center of the two molecular particles, b and χ are defined in the moving reference frame
Equation (2.3) shows that cp1 and cp2 can be computed using cr , which can be
determined via rotating cr by a deflection angle χp1 inside the collision plane. The
collision plane can be uniquely prescribed by the pre-collision positions xp1 and xp2
of the collision pair and the undisturbed trajectory of molecule p1 passing through
xp1 in the direction of cr (except for the head-on collision with b = 0 and thus
cr ≡ −cr ). As shown in Fig. 2.2, the collision plane can be parameterized using the
angle ε between it and a preset reference plane that always contains the undisturbed
trajectory of molecule p1 and an axis (e.g., the x1 axis used here) of the Cartesian
coordinate system Sm moving at cm . Sm is placed appropriately in such a way that
the undisturbed trajectory of molecule p1 will pass through the origin O of Sm to
form the reference plane as shown in Figs. 2.1 and 2.2. In the following calculations
of velocity components of cr , the unit coordinate vectors of Sm are always oriented
the same as in the static global S and thus the same symbols ei∈[1,2,3] will be used.3
We first introduce the concept of spherical trigonometry by Fig. 2.3, where three
points A, B and C 4 are located on the spherical surface centered at the point O.
3 This is different from the convention used in [5], where the moving coordinate system S
m changes
for each particular binary collision to have its em,1 always orientated in the direction of cr and thus
the definition of ε is different from the one used here.
4 By a slight abuse of notation, C denotes the mole concentration introduced in Sect. 1.3.6, an
arbitrary reference point on spherical surface introduced here, and the magnitude of molecular
peculiar velocity introduced in Sect. 2.5, and Cs denotes a coefficient of the Smagorinsky turbulence
model introduced in Sect. 4.5.1.
Fig. 2.2 Schematic of

calculating cr using cr , ε and
χp1 according to the
spherical trigonometry;
G OG is the collision plane
and G O A is the reference
plane; G, G , A, B and C are
points located on the surface
of a sphere centered at O;
only the part of spherical
surface stretched by G G A is
drawn; Sm is a Cartesian
coordinate system of the
moving reference frame
Similar to other arcs, AB is a part of great circle, which is the intersection of the
sphere surface with a diametral plane passing through A, B and O. The four points
determine a tetrahedron with O as the apex and the spherical triangle ABC as the
base. The part of sphere surface stretched by the triangle ABC is uniquely specified
as the part of the sphere surface cut by elongating the tetrahedron when moving the
base away from the apex. We use the symbols ε and ∠ for angle between two planes
and angle between two lines, respectively, and they are measured by rotating these
planes or lines within the tetrahedron or on its surfaces as shown in Fig. 2.3 (left),
which implies that these angles always rang from 0 to π .
Considering the spherical triangle G G A of Fig. 2.2, the cosine rule of spherical
trigonometry reads [21]:
cos ∠G O A = cos ∠G O A cos ∠G OG + sin ∠G O A sin ∠G OG cos ε, (2.5)

where cos ∠G O A = (cr · e1 )/cr = cr,1 /cr = cr,1 /cr , cos ∠G O A = (−cr · e1 )/cr =
−cr,1 /cr and thus sin ∠G O A = (cr,2 + cr,3 ) /cr since ∠G O A ∈ [0, π ] by conven-
2 2 1/2
tion and thus sin ∠G O A ≥ 0, cos ∠G OG = − cos χp1 and sin ∠G OG = sin χp1
since ∠G OG + χp1 = π . Thus, Eq. (2.5) can be rewritten into:

cr,1 = cr,1 cos χp1 + (cr,2
2
+ cr,3 ) sin χp1 cos ε.
2 1/2
(2.6)
Similarly, the cosine rule of spherical trigonometry for the spherical triangle
G G B has the following result:
proper wrong
Fig. 2.3 Schematic of spherical trigonometry, where the symbols ε and ∠ are used for angle
between two planes and angle between two lines, respectively
Fig. 2.4 One possible case

in determining the
relationship between the
angles of surface pairs,
εG G B = ε BG A − ε and
εG GC = εC G A − ε
regardless of χp1 ∈ [0, π ]

2
+ cr,3 ) sin χp1 cos εG G B .
2 1/2
(2.7)
Now, we need to make an assumption for each particular case. As a special case
of Fig. 2.4, we assume that the minor circular sector G OG (note: always consider
the minor sector since χp1 ∈ [0, π ] by convention) has intersections with the line
segments AB and AC at B and C , respectively, but consequently has no intersection
with the line segment BC (note: the three sides of a triangle cannot be cut at a
time). Thus, we have εG G B = ε BG A − ε and εG GC = εC G A − ε since ∀ε ∈ [0, π ]
by convention, which indicates that the rotating direction around cr will be different
for different ε. Then, cos εG G B in Eq. (2.7) can be computed via ε BG A using the
cosine rule of the spherical triangle BG A:
cos ∠B O A = cos ∠G O A cos ∠G O B + sin ∠G O A sin ∠G O B cos ε BG A , (2.8)

i.e.,
0 = cr,1 cr,2 + (cr,2

2
+ cr,3 ) (cr,1 + cr,3
2 1/2 2
) cos ε BG A ,
2 1/2
(2.9)
and therefore we have:

−cr,1 cr,2
cos ε BG A = , (2.10)
(cr,2
2
+ cr,3 ) (cr,1 + cr,3
2 1/2 2
)
2 1/2
and
|cr,3 |cr
sin ε BG A = , (2.11)
(cr,2
2
+ cr,3 ) (cr,1
2 1/2 2
+ cr,3 )
2 1/2
where the absolute value of cr,3 is used due to ε BG A ∈ [0, π ] by convention.5 Substi-
tuting Eqs. (2.10), (2.11) and cos εG G B = cos(ε BG A − ε) into Eq. (2.7), we obtain:
sin χp1
cr,2 = cr,2 cos χp1 + (|cr,3 |cr sin ε − cr,1 cr,2 cos ε). (2.12)
(cr,2
2
+ cr,3 )
2 1/2
By similar analysis using the spherical triangle G GC we obtain:
sin χp1
cr,3 = cr,3 cos χp1 + (|cr,2 |cr sin ε − cr,1 cr,3 cos ε). (2.13)
(cr,2
2
+ cr,3 )
2 1/2
Obviously, the above derivation requires more analyses via enumeration of Fig. 2.4
to reach general formulas, which exist as shown below. Additionally, it started with
an assumption of ε ∈ [0, π ] for the intermolecular collision in order to comply with
the convention of spherical trigonometry, but this assumption is not satisfied by all
possible intermolecular collisions having the same cr but random xp1 and xp2 , which
implies ε ∈ [0, 2π ].
A different derivation in a general form is given below. As shown in Figs. 2.2 and
2.4, the angle ε can be uniquely specified as follows: point the thumb of right hand to
the direction of cr and curl the fingers to rotate around the undisturbed trajectory of
molecule p1 from the half -collision-plane G OG containing the trajectory of p1 to
the half -reference-plane G O A containing the positive part of the x1 axis, and then
the angle of rotation ranging from 0 to 2π is defined as ε. Now, we first discuss the
case of rotating an arbitrary vector vb around another arbitrary but unit vector va by
an angle θ to the vector vc . To avoid confusion, all vectors are placed appropriately to
pass through the origin of Sm as in Fig. 2.2. An auxiliary Cartesian coordinate system

Sm with ei∈[1,2,3] as its unit coordinate vectors is introduced only to help understand
5 Thus,|cr,3 | and |cr,2 | of Eqs. (2.11)–(2.13) are used to replace cr,3 and cr,2 of [21], respectively,
although we have |cr,3 | = cr,3 for the particular case of Fig. 2.4.
the derivation of the calculation formula of vc = f (vb , va , θ ), which is independent

of Sm and thus generally valid.
Note that |va | = 1 and thus va ⊥[vb − (vb · va )va ]. Let e3 = va , and e1 points at
vb − (vb · va )va , and consequently e2 = e3 × e1 will point at va × [vb − (vb · va )va ]
and |vb − (vb · va )va | = |va × [vb − (vb · va )va ]|, which will be denoted by the
length vb,⊥a . Note that the part (vb · va )va of vb is unchanged during the rotation
around va but vb − (vb · va )va will be changed to vb,⊥a cos θ e1 + vb,⊥a sin θ e2 accord-
ing to the expression in a cylindrical coordinate system corresponding to Sm , which
leads to the following general formula:
vc = (vb · va )va + vb,⊥a cos θ e1 + vb,⊥a sin θ e2

(2.14)
= (vb · va )va + cos θ [vb − (vb · va )va ] + sin θ va × [vb − (vb · va )va ].
Equation (2.14) degenerates to the following simple form when vb ⊥va :
vc = cos θ vb + sin θ va × vb . (2.15)
In the case of intermolecular collision, cr × e1 = −cr,2 e3 + cr,3 e2 is a normal

vector of the half -reference-plane G O A along its rotating direction around cr .
Let vb = (−cr,2 e3 + cr,3 e2 )/(cr,2
2
+ cr,3 ) , va = cr /cr and θ = −ε, and then the
2 1/2
obtained vc is the unit normal vector of the half -collision-plane G OG along its
rotating direction (note: the determination of normal vector of half-plane along its
rotating direction around cr always complies with the above right-hand rule even
if the actual rotating direction on different purpose indicates the opposite normal
direction when θ < 0). Note that vb ⊥va is satisfied and Eq. (2.15) leads to:
−cr,2 e3 + cr,3 e2 cr,1 cr,2 e2 + cr,1 cr,3 e3 − cr,2

2
e1 − cr,3
2
e1
vc = cos ε − sin ε . (2.16)
(cr,2 + cr,3 )
2 2 1/2
cr (cr,2 + cr,3 )
2 2 1/2
Note that |vc | = |vb | = 1 and thus the obtained vc of Eq. (2.16) can be used as the
rotating axis (i.e., va ) of Eq. (2.15). Then, rotating cr around vc of Eq. (2.16) by χp1
and applying Eq. (2.15) again since cr ⊥vc , we obtain cr = cos χp1 cr + sin χp1 vc × cr
and the following general calculation formulas for its components:

2
+ cr,3 ) sin χp1 cos ε,
2 1/2
sin χ p1
cr,2 = cr,2 cos χp1 − 2 (c c sin ε + cr,1 cr,2 cos ε),
2 1/2 r,3 r
(cr,2 + cr,3 ) (2.17)
sin χp1
cr,3 = cr,3 cos χp1 + 2 (c c sin ε − cr,1 cr,3 cos ε).
2 1/2 r,2 r
(cr,2 + cr,3 )
We use εsole as ε of Eq. (2.17) to distinguish from ε of Eqs. (2.6), (2.12) and (2.13),
which then are consistent with the general formulas of Eq. (2.17) because cr,2 < 0,
cr,3 > 0 and εsole = 2π − ε as implied in the assumptions of Fig. 2.4.
The deflection angle χp1 generally depends on the intermolecular potential E p , b

and cr as detailed in [5, 21] but the derivation by using the integration of orbit equation
in the polar coordinate system is omitted here since the hard sphere (HS) model and
the variable hard sphere (VHS) model with variable diameter D will be used in the
following discussion and thus χp1 will depend only on b and D in a simple way,
b = cos(χp1 /2)(Dp1 + Dp2 )/2 as shown in the next Sect. 2.3. Additionally, although
the other impact parameter ε depends on the pre-collision positions of the collision
pair in addition to their cr , it can be treated as being uniformly distributed in [0, 2π ]
for arbitrary cr given the randomness of the pre-collision positions when discussing
the statistical behavior of many intermolecular collisions.
2.3 Molecular Models for Transport Coefficients
In the following, we adopt c and cp1 to replace cp1 and cp2 (similar for other molecular
variables), respectively, in using the previous formulas for simplicity when discussing
intermolecular collisions. As shown in Eq. (2.17) of Sect. 2.2, cr can be determined
by cr , ε and χ (use χ for χp1 and χp2 ), where χ depends on cr , b and the intermolec-
ular potential E p (note: hard sphere model is a model with special E p ). For arbitrary
molecular model with a given E p , we can use only two parameters b and ε to study
the statistical property of cr for each cr . As shown in Fig. 2.5, the position of the
surrounding molecule p1 relative to the test molecule with velocity c can be charac-
terized by using the differential collision cross-section bdεdb. Correspondingly, cr
will be deflected to cr located inside the differential solid angle dΩ = sin χ dεdχ ,
where dχ is associated with db. We define σ by σ dΩ = bdεdb:
b db
σ = | |, (2.18)
sin χ dχ
which is a function of χ and contains cr and E p as parameters in general.

The total collision cross-section σT is defined as:
2π
σT = bdεdb = 2π bdb, (2.19)
0
and can be calculated as follows (note: σ is independent of ε as shown in Eq. (2.18)):

4π π 2π π
σT = σ dΩ = σ sin χ dεdχ = 2π σ sin χ dχ . (2.20)
0 0 0 0
Equation (2.20) indicates that σT depends on cr and E p in general. But, the calcula-
tion of σT using Eq. (2.19) or Eq. (2.20) cannot be completed as a simple geometrical
integration, which leads to infinity for molecular models with usual intermolecu-
lar potential E p except the hard sphere model. The total collision cross-section σT

calculating collision dΩ = sin d d r
cross-section, where bdεdb
is the differential collision d
area corresponding to the
position distribution of the d
surrounding molecule p1
relative to the test molecule
concerned, and dχ is r
associated with db
d d p 1 located inside d d
determines the intermolecular collision frequency ν per molecule (see the follow-
ing Eq. (2.72)) and thus finite truncation of χ (and b accordingly since χ = χ (b)) is
needed to make ν finite, e.g., the intermolecular collisions with χ < χmin (or b > bmax
as χmin = χ (bmax )) are neglected. The molecular mean free path λ depends on ν and
can be clearly defined when the gas molecules are modeled as hard spheres, where
the subjective truncation can be avoided in calculating σT for ν and λ.
In addition to σT , there is also the viscosity cross-section σμ for single component
systems:
4π 2π
σμ = sin2 χ σ dΩ = sin2 χ bdεdb, (2.21)
0 0
which appears in the expression of dynamic viscosity μ obtained via the Chapman–
Enskog expansion of the Boltzmann equation [6, 21, 23]:
(5/8)(π mkB T )1/2

μ= ∞ , (2.22)
[m/(4kB T )]4 0 cr σμ exp[−mcr /(4kB T )]dcr
7 2
where kB ≈ 1.38064852 × 10−23 J/K is the Boltzmann constant. Note that although
the kinetic theory is based on a description at the molecular scale, the molecular inter-
action is modeled in a statistical way that is different from the physical modeling of
MD simulation based on the Newton’s laws of motion. Thus, the molecular param-
eters (e.g., molecular diameter) calibrated in the MD simulation is not valid in the
kinetic simulation. However, their approaches of determining molecular parameters
are similar, namely to match the fluid transport properties (e.g., viscosity, thermal
conductivity and diffusion coefficients).
The simplest molecular model is the hard sphere model. As shown in Fig. 2.6,
the two molecules collide as two billiards when b is less than the mean diameter
2.3 Molecular Models for Transport Coefficients 35
Fig. 2.6 Collision of hard

sphere (HS) molecules,
where cos(χ/2) = b/Dmean
Dmean of the two molecules and we have b = Dmean cos(χ /2). Substituting it into
Eq. (2.18), we obtain:
σ = Dmean
2
/4, (2.23)
which becomes independent of the orientation parameter χ . Additionally, we have

σT = π Dmean
2
according to Eq. (2.19) or, Eqs. (2.20) and (2.23).
In discussing the statistical behavior of many intermolecular collisions, the posi-
tion of surrounding molecule p1 relative to the test molecule concerned can be taken
as randomly distributed, which implies that the probability Pxp1 ∈bdεdb of the surround-
ing molecule p1 coming through the differential collision area bdεdb is proportional
to the area bdεdb itself. Consequently, the probability Pcr ∈dΩ of scattering the post-
collision velocity cr into the corresponding differential solid angle dΩ = sin χ dεdχ
is also proportional to bdεdb since Pcr ∈dΩ = Pxp1 ∈bdεdb . As bdεdb ≡ σ dΩ ∝ dΩ
according to Eq. (2.23), we obtain Pcr ∈dΩ ∝ dΩ, which indicates that the post-
collision velocity cr of hard sphere molecule has an isotropic distribution over all
possible orientations (i.e., uniformly points to a spherical surface).
In the single component systems, Dmean will be denoted simply by D. The actual
molecular size cannot be precisely and uniquely defined. Instead, the diameter D of
hard sphere model should be determined by the physical transport coefficient, i.e., the
dynamic viscosity μ or thermal conductivity ζ . Note that the Boltzmann equation has
correct Prandtl (Pr) number for monatomic gas molecules (Pr = cp μ/ζ = 2/3, see
Eqs. (6.52), (6.53) of [21], or Eqs. (12.1.6) and (13.1.2) of [6]), which is defined as the
ratio of momentum diffusivity rate to thermal diffusivity rate and nondimensionalized
by the isobaric specific heat capacity cp . Thus, the coefficient ζ will be automatically
matched if D is determined from μ. Substituting Eq. (2.23) into Eq. (2.21), we obtain
σμ = 2σT /3 = 2π D 2 /3, which is then substituted into Eq. (2.22) to get the following
correlation between D and μ [21]:
5(mkB T /π )1/2
μ= , (2.24)
16D 2
∞
where 0 y 7 exp(−β 2 y 2 )dy = 3/β 8 is used. Equation (2.24) indicates that the
dependence of μ on T is to the power of 1/2 if D is fixed while the powers for
real gases are usually about 0.75.
The above shortcoming can be successfully solved by using the variable hard
sphere (VHS) model [3], where the only difference from the original hard sphere
model is that σT is a function of cr as implied by Eq. (2.20). It decreases with
the increase of cr , which is consistent with the result of more realistic molecular
models having intermolecular forces of inverse power law if the same truncation
χmin is applied for all cr . The practical application shows that the simplification of
scattering cr in the HS and VHS models has little influence to the flow field of single-
component problem while the dependence of σT on cr is essential. The VHS model
can be generally defined as follows:
2 −2ξ
σT D cr
= = , (2.25)
σT,ref Dref cr,ref
where π D 2 = σT = σT,ref = π Dref 2

at cr = cr,ref by definition. Note that σ = D 2 /4
and σμ = 2σT /3 still hold although σ and σμ depend on cr via D. Substituting
σμ = 2σT,ref (cr /cr,ref )−2ξ /3 according to Eq. (2.25) into Eq. (2.22), we obtain:
15(π mkB )1/2 (4kB /m)ξ T (1/2+ξ )

μ= 2ξ
, (2.26)
8Γ (4 − ξ )σT,ref cr,ref
∞
where the gamma function is defined as Γ (z) = 0 y (z−1) exp(−y)dy. We have
Γ (z + 1) = zΓ (z) according to the integration by parts and Γ ( j + 1) = j! for inte-
ger j. In order to match μ = μref (T /Tref )ω as the desired dependence of μ on T ,
Eq. (2.26) requires that ξ = ω − 1/2 and D should vary with cr as follows6 :

15(m/π )1/2 (kB Tref )ω
DVHS = , (2.27)
8Γ (9/2 − ω)μref (mcr2 /4)(ω−1/2)
2
which degenerates to Eq. (2.24), i.e., DVHS (ω = 1/2) = 5(mkB Tref /π )1/2 /(16μref ).
In multi-component problems, the diffusion coefficient is as important as the
viscosity coefficient but the HS and VHS models cannot simultaneously match
4π
them (or the viscosity cross-section σμ and diffusion cross-section σ D = 0 (1 −
cos χ )σ dΩ). This defect can be solved by using the variable soft sphere (VSS)
model [14, 15], in which σT depends on cr as in the VHS model but the scatter-
ing rule of cr becomes b = D cosα (χ /2), where α depends on molecule species
and is usually larger than 1 to make χVSS < χHS = χVHS . The parameter α can be
determined for different molecules by matching σμ /σ D between the VSS model
and the inverse power law model. It can also be determined by fitting the exper-
imental data with a formula of viscosity that is similar to Eq. (2.26) but contains
6 mc2 /4 can be replaced by the relative translational energy m r cr2 /2 [5, 21].
r
2.3 Molecular Models for Transport Coefficients 37
an additional factor (α + 1)(α + 2)/(6α) in the right-hand side as σμ,VSS /σμ,VHS =

6α/[(α + 1)(α + 2)] [15], or by fitting the experimental data with a formula of self-
diffusion coefficient [5]. Similar to Eq. (2.27), we obtain:

5(α + 1)(α + 2)(m/π )1/2 (kB Tref )ω
DVSS = . (2.28)
16αΓ (9/2 − ω)μref (mcr2 /4)(ω−1/2)
In real applications [3, 5], Eq. (2.28) is replaced by:

kB Tref (ω−1/2) 1
DVSS = Dref,VSS ( ) , (2.29)
mcr /4
2 Γ (5/2 − ω)
where Dref,VSS is computed as follows according to Eq. (2.28):

5(α + 1)(α + 2)(mkB Tref /π )1/2
Dref,VSS = . (2.30)
4α(5 − 2ω)(7 − 2ω)μref
The general Eqs. (2.29), (2.30) take α = 1 for the VHS molecules and α = 1, ω =
1/2 for the HS molecules. In particular, Dref,VSS defined by Eq. (2.30) can degenerate
2
to Eq. (2.24), i.e., Dref,VSS (α = 1, ω = 1/2) = 5(mkB Tref /π )1/2 /(16μref ) = DHS
2
.
Additionally, the VHS and VSS models describe the interaction by a pure repul-
sive force while the intermolecular attractive force is noticeable at large distance in
reality, which makes it inaccurate to use only one term of power law to represent the
dependence of viscosity on the temperature in problems with a wide range of tem-
perature variation (e.g., spacecraft re-entry). Correspondingly, the generalized hard
sphere (GHS) model [13] is introduced to reproduce the effect of attractive-repulsive
potential by adding terms in σT = σT (cr ) but its scattering rule follows that of the
VHS model. It bears the same relationship to the Lennard-Jones molecular model
while the VHS and VSS models bear to the inverse power law model. Additionally,
the GHS model can be extended to the generalized soft sphere (GSS) model [9] by
using the scattering rule of the VSS model and the GSS model is very useful par-
ticularly in the cases of low temperature and polar molecules. More discussion on
the GHS, GSS models, inverse power law model, and the Maxwell model (a special
inverse power law model) are available in [5, 21].
At the end of this section, we discuss the implementations of the HS, VHS and VSS
models in Monte Carlo molecular simulations. The probability Pcr ∈dΩ of scattering
the post-collision velocity cr into the differential solid angle dΩ is proportional to
dΩ = sin χ dεdχ = −dεd cos χ in the HS and VHS models, which indicates that ε
and cos χ have uniform distributions. Correspondingly, the selections of ε ∈ [0, 2π ]
and cos χ ∈ [−1, 1] as χ ∈ [0, π ] are: ε = 2π R f 1 and cos χ = 2R f 2 − 1, where
R f 1 and R f 2 are two independent random fractions uniformly distributed between
0 and 1. Consequently, we obtain χ = arccos(2R f 2 − 1). Since the distribution
of cr of the HS and VHS molecular models is always isotropic regardless of the
pre-collision velocity cr , the computation of cr using cr , ε and χ in Eq. (2.17) can be
avoided. Instead, the isotropic cr can be selected directly in the static global spherical
coordinate system S̃, where ε and χ can be deemed the azimuthal and polar angles,
respectively, although their original definitions are associated with the orientations of
the collision plane and reference plane. Then, the three components of cr are simply

computed as follows: cr,1 = cr sin χ cos ε, cr,2 = cr sin χ sin ε and cr,3 = cr cos χ .
In the VSS model with b = D cosα (χ /2), the probability Pcr ∈dΩ is not propor-
tional to dΩ since σ depends on the orientation parameter χ and thus the above

selecting algorithm of cr,i∈[1,2,3] becomes invalid. On the other hand, it is easy to

select cr using the fact that the probability Pxp1 ∈bdεdb of the surrounding molecule
p1 coming through the differential collision area bdεdb is proportional to the area
bdεdb itself, which is guaranteed by assuming that the position of surrounding
molecule p1 relative to the test molecule concerned is randomly distributed. Note
that Pcr ∈dΩ = Pxp1 ∈bdεdb ∝ bdεdb = (1/2)dεdb2 and thus ε is uniformly distributed
inside [0, 2π ] and (b/D)2 is uniformly distributed inside [0, 1]. We rewrite the scat-
tering rule b = D cosα (χ /2) into cos χ = 2[(b/D)2 ](1/α) − 1 and then χ can be
(1/α)
randomly sampled as χ = arccos(2R f 2 − 1). Additionally, ε is randomly sam-
pled by ε = 2π R f 1 as in the HS and VHS models. Due to the anisotropic distribution
of cr of the VSS model, the selected ε and χ should be substituted together with cr
into Eq. (2.17) to compute cr .
2.4 Derivation of the Boltzmann Equation
The probability distribution f is a function of c, x and t as in the Eulerian description

but its evolution can be easily understood in the Lagrangian description. Suppose
there is no intermolecular collision and the external body force per unit mass is
g, the molecules located inside the differential element dcdx around (c, x) at the
moment t will move into dcdx around (c + dtg, x + dtc) at the moment t + dt,
which indicates [6]7 :
f (c + dtg, x + dtc, t + dt)dcdx = f (c, x, t)dcdx. (2.31)
Dividing Eq. (2.31) by dcdxdt, we obtain the following equation for problems without
intermolecular collisions:

∂f ∂f ∂f ∂f ∂f
= −g · −c· = + . (2.32)
∂t ∂c ∂x ∂t force ∂t move
7 This is similar to the conclusion of the Liouville’s theorem that the phase-space distribution function
is constant along the molecular trajectory, which is because of the fact that the ‘velocity field’ in
the phase space is non-divergent according to the Hamilton’s relations, and thus the differential
element has unchanged volume along the molecular trajectory as well.
2.4 Derivation of the Boltzmann Equation 39
Now, we neglect the molecular movements and external force to study (∂ f /∂t)coll
due to intermolecular collisions alone, and consider the change of f dcdx due to
intermolecular collisions between molecules of all possible c inside dx during dt.
Additionally, the two molecules of each collision pair are assumed to be uncor-
related prior to the intermolecular collision, i.e., molecular chaos assumption that
f (2) (c, cp1 , x, xp1 , t) = f (c, x, t) f (cp1 , xp1 , t), where f (2) is a two-particle distribu-
tion function. This assumption is removed in the Enskog equation developed for dense
gas problems, where the actual probabilistic number density f (2) (c, cp1 , x, xp1 , t)
will be very small due to the intermolecular repulsive force when xp1 is close to x
while the product f (c, x, t) f (cp1 , xp1 , t) is independent of |xp1 − x| by definition
and thus not accurate to approximate f (2) .
The number change of molecules of class c (note: class c means having c located
inside dc around c) in the element dx due to the intermolecular collisions during dt is
denoted by (∂ f /∂t)coll dcdxdt. We consider the collisions (c, cp1 → c , cp1 ) between
molecules of class c and molecules of class cp1 , in which cr = c − cp1 is scattered
into dΩ = sin χ dεdχ and correspondingly the post-collision velocities c and cp1
are determined by ε, χ , c and cp1 as discussed in Sect. 2.2. The molecules of class
c moves at the speed of cr relative to the molecules of class cp1 and sweeps out a
volume cr bdεdbdt = cr σ dΩdt, where the molecular number density of class cp1
is f p1 dcp1 = f (cp1 , xp1 , t)dcp1 ≈ f (cp1 , x, t)dcp1 . According to the assumption of
molecular chaos, the number of intermolecular collisions (c, cp1 → c , cp1 ) inside dx
during dt can be expressed in terms of the product of two single particle distribution
functions:
( f dcdx)( f p1 dcp1 cr σ dΩdt). (2.33)
Meanwhile, the number of inverse collisions (c , cp1 → c, cp1 ) inside dx during dt is:
( f dc dx)[ f p1 dcp1 cr (σ dΩ) dt], (2.34)
where f = f (c , x, t) and cr = cr . The symmetry between the direct and inverse
collisions guarantees that the transformation Jacobian (determinant of a matrix) of
changing variables for multiple integral equals unity, namely:
(σ dΩ) dc dcp1 = (σ dΩ)dcdcp1 . (2.35)
Note that the direct collision (c, cp1 → c , cp1 ) causes the depletion of molecules
of class c while the inverse collision (c , cp1 → c, cp1 ) leads to the replenishment of
molecules of class c. Subtracting Eq. (2.33) from Eq. (2.34) and applying Eq. (2.35),
we obtain the gain of molecular number of class c inside dx during dt due to the
above two types of intermolecular collisions:
( f f p1 − f f p1 )cr σ dΩdcdcp1 dxdt. (2.36)

The total gain of molecular number of class c can be computed by taking the inte-
gration over the whole volume swept out by the molecules of class c, equivalently
by taking the integration over the whole solid angle as bdεdb = σ dΩ. Additionally,
integration over the entire space of cp1 is needed to consider all possible intermolec-
ular collisions between the class c and the rest. Thus, the total number increase of
molecules of class c inside dx caused by the intermolecular collisions during dt is:

∂f 4π

dcdxdt = (f f p1 − f f p1 )cr σ dΩdcp1 dcdxdt. (2.37)
∂t coll R3 0
Adding (∂ f /∂t)coll of Eq. (2.37) to the right hand side of Eq. (2.32), we obtain
the Boltzmann equation for the single-component gas system:

∂f ∂f ∂f ∂f
= + +
∂t ∂t force ∂t move ∂t coll
(2.38)
∂f ∂f 4π

= −g · −c· + (f f p1 − f f p1 )cr σ dΩdcp1 ,
∂c ∂x R3 0
where f is the only unknown, c, x and t are the independent variables, cr = |c − cp1 |,
dΩ = sin χ dεdχ , σ = σ (χ , cr , E p ) in general, c and cp1 contained in f and f p1 ,
respectively, are functions of c, cp1 , χ and ε according to Eqs. (2.3) and (2.17). It
is almost impossible to obtain analytical solutions of the Boltzmann equation for
flows with complicated geometries or large perturbations. It is also very challenging
to obtain numerical solutions by using ordinary discretization schemes in the phase
space, where huge number of velocity points is needed to discretize the unbounded
molecular velocity space c at each spatial grid of the physical space x.
It is noteworthy that the derivation of Boltzmann equation has adopted two
assumptions: (1) binary collisions are dominant; (2) molecular chaos. These pre-
requisites usually require that the applications of Boltzmann equation are limited to
rarefied gas. By contrast, the model equations (e.g., BGK equation of Sect. 2.9) can
be used for dense gas since the intermolecular collisions are modeled as a relaxation
process without using these two assumptions. Additionally, the Enskog equation only
removes the molecular chaos assumption but has been successfully applied to model
dense gas.
In addition to the two assumptions, there is a simplification by modeling molecules
as particles without rotational and vibrational degrees of freedom. For diatomic and
polyatomic molecules, the dimensions of phase space are increased by the number
of internal degrees of freedom. To consider the exchanges between the translational
and intramolecular energies, more transport coefficients are to be introduced and the
collision term of the Boltzmann equation can be correspondingly modified [24].
The Boltzmann equation is the basic equation of rarefied gas dynamics and the
complicated collision term of its right-hand side can be taken as zero in problems
of the free molecular flow regime, where the intermolecular collision frequency is
negligible compared to the molecular collision frequency with boundary and the
2.4 Derivation of the Boltzmann Equation 41
solution depends mostly on the boundary condition of molecular reflection. The

N–S–F-like equation can be obtained from the Boltzmann equation as shown in the
following Eq. (2.59) and the constitutive equations can be derived by the Chapman–
Enskog expansion of the Boltzmann equation.8 The obtained N–S–F-like equation
can be extended to the Burnett equation by taking higher-order terms of the Chapman–
Enskog expansion to modify the constitutive equations. Additionally, the boundary
conditions of velocity slip and temperature jump used in the N–S–F-like equations
can be derived by the Chapman–Enskog expansion as well [21].
2.5 Calculations of Macroscopic Properties
After getting the solution f of the Boltzmann equation, the macroscopic flow prop-
erties of interest can be obtained by computing the moments of f . The moments are
usually defined by integrals with respect to c and the integrands are the products
of f and different quantities Q carried by each molecule. As shown in Eq. (2.1),
the number density is computed as the moment with Q ≡ 1. Note that f dcdx is the
molecular number in dcdx at t or simply refereed to as the molecular number of class
c inside dx at t. Thus, mc f dcdx is the momentum carried by the class c inside dx
at t and the mean molecular velocity c can be determined using the total momentum
and total mass:

( R3 mc f dc)dx 3 c f dc
c(x, t) = = R . (2.39)
( R3 m f dc)dx n
For single-component flow problems, c defined by Eq. (2.39) is the flow velocity
u used in the N–S-like equations. Generally speaking, the mean value of arbitrary
molecular quantity Q(c) at (x, t) is defined as follows:

R3 Q f dc
Q(x, t) = . (2.40)
n
The molecular velocity C observed in the reference frame moving at c is called the
peculiar velocity (i.e., thermal or random velocity) and thus C = c − c. Obviously,
we have C = 0 according to Eq. (2.39).
In calculating the force or heat flux via a surface element d A, the transports of
mass, momentum and energy associated with the molecular movements across d A
are concerned. The outward unit vector normal to d A is denoted by n+ . As shown
in Fig. 2.7, along the molecular trajectory of class c, there is a column (e.g., cylinder
here) that is situated oppositely to the c side of d A and has d A as the base and cdt
8 It
was independently proposed by Chapman and Enskog in about the same year to obtain approx-
imate solutions of the Boltzmann equation. Their results are identical from the application point of
view but their derivations differ widely in spirit and detail [8]. Enskog’s method was adopted by
Chapman and Cowling in their book [6].
Fig. 2.7 The volume

occupied by the molecules of
class c at the beginning of dt
that move across d A in dt
as the side length. Molecules of class c located inside this column at the beginning
of dt will move across d A after dt. The column volume is c · n+ dtd A and thus the
molecular number of class c moving across d A in dt is:
f dc(c · n+ )dtd A. (2.41)
Accordingly, the amount of arbitrary molecular quantity Q carried by the molecules

of class c crossing d A in dt is Q f dc(c · n+ )dtd A. The total transport of Q carried
by all classes of molecules crossing d A in dt can be computed by integration over
the entire c space and correspondingly the total flux (Q) is9 :

(Q) = Qc · n+ f dc = n Qc · n+ = n Qc · n+ , (2.42)
R3
where the definition of mean value by Eq. (2.40) is applied and n Qc is called the
flux vector of Q. Let Q = m and we obtain the mass flux vector nmc = nmc = ρu.
Sometimes, we need the flux of Q for the transport process crossing the surface only
in one direction, e.g., c·n+ >0 Qc · n+ f dc = n(Qc · n+ )c·n+ >0 as in Eq. (2.77).
When discussing the stress tensor P and heat flux vector q, the transports of
momentum and energy associated with the molecular movements across an imagi-
nary d A moving at the local flow velocity c = u should be considered such that the
defined P and q are the interactions between moving fluid parcels of the Lagrangian
description rather than the exchanges between fixed control volumes of the Eule-
rian description. Correspondingly, we shall discuss the flux vector n QC of arbitrary
molecular quantity Q(C). Let Q = mC, we obtain the momentum flux tensor caused
by the random thermal motion, i.e., the expression of stress tensor:

P=− mCC f dC = −nmCC = −ρCC. (2.43)
R3
9 The light-face symbol n without subscript is the number density rather than the magnitude of n+
that is unit.
2.5 Calculations of Macroscopic Properties 43
The physical meaning of P defined by Eq. (2.43) becomes clear after understanding
each of its components. For example, we consider the momentum transport of the
e1 direction across the surface d A having e2 as the outward normal direction. This
momentum component of each molecule is mC1 for the class C. The molecules
of this class crossing d A in dt are initially located in the volume C2 dtd A and the
corresponding molecular number is f dC(C2 dtd A). The momentum component of
the e1 direction carried by them is mC1 f dC(C2 dtd A). Note that the total momentum
flux across d A along e2 corresponds to the force exerted by the fluid parcel situated
on the −e2 side of d A to the fluid parcel situated on the e2 side of d A but, P21
corresponds to the counterforce according to the convention of Eq. (1.9) that the
first and second subscripts of P ji = Pe j · ei are associated with the outward normal
direction of d A and the direction of exchanged momentum component, respectively.
According to the Newton’s third law, we have:

P21 = − C2 mC1 f dC = −ρC2 C1 . (2.44)
R3
Obviously, the stress tensor defined by Eq. (2.43) is symmetric as required.

The pressure acting across the oriented surface element d A is the flux of momen-
tum component of the inward normal direction n− across d A along n− :

(C · n− )(mC · n− ) f dC = (n− · ρCC) · n− = −(n− · P) · n− , (2.45)
R3
which is always positive since the integrand of the left-hand-side integral is non-
negative. This indicates that the gas media having the momentum exchange via
molecular translational motions as the dominant effect always exerts on the surface a
compressing force but never stretching force. By contrast, the intermolecular repul-
sive and attractive forces have more contributions to the momentum exchange than
the molecular movement for solid and liquid media and thus stretching (e.g., cohesive
or adhesive) force is admissible when the intermolecular attractive force is dominant.
We can calculate the pressures at arbitrary point exerted across the three special coor-
dinate planes having e1 , e2 and e3 as their inward normal directions, respectively.
This three pressures are usually different due to non-equilibrium distribution, where
f is asymmetric with respect to C1 , C2 and C3 . But, the mean value of pressures
across any three orthogonal planes is independent of the coordinate system ei∈[1,2,3]
and thus is used to define the mean hydrostatic pressure [6]:
ρ 2 ρ
p= C1 + C22 + C32 = C 2 . (2.46)
3 3
Note that if the non-equilibrium effect is not significant, the mean hydrostatic pressure
p defined by Eq. (2.46) is close to the thermodynamic pressure, which is used in the
constitutive equation of stress tensor and determined in a different way, and thus the
same symbol p is used for both to simplify the notation. In the case of remarkable
non-equilibrium effect, the usual constitutive equation is invalid and thus p will be
used as the mean hydrostatic pressure, unless stated otherwise.
Similarly, we can calculate the heat flux vector as q = n QC, where Q = mC 2 /2
is the molecular kinetic energy associated with random translational motion (note:
Q contains more terms if molecular rotation and vibration are considered as well).
Additionally, the mean value of molecular translational energy per unit mass is:

R3 (C /2) f dC
2
etr = = C 2 /2. (2.47)
n
Correspondingly, the translational kinetic temperature Ttr is defined:
2metr
Ttr = . (2.48)
3kB
Substituting Eqs. (2.48)–(2.47) into Eq. (2.46), we obtain the following equation of
state:
p = nkB Ttr , (2.49)
which can be rewritten into p = ρ Rspecific Ttr as Rspecific = kB /m = kB NA /(m NA ) =

R/Mmole , where Rspecific is the specific gas constant, NA ≈ 6.022140857 × 1023
mol−1 is the Avogadro constant, R = kB NA ≈ 8.3144598 J/(mol·K) is the universal
gas constant, and Mmol is the molecular mass per mole. Note that the ideal gas equa-
tion of state, i.e., Eq. (2.49), is valid also at non-equilibrium state as long as p and
Ttr are defined by Eqs. (2.46) and (2.48), respectively.
Sometimes it is convenient to define Ttr using the average molecular velocity:
2m 2 2m c2 − c2
Ttr = (C1 + C22 + C32 )/2 = . (2.50)
3kB 3kB 2
Similarly, we can separately define the translational kinetic temperatures Ttr,i∈[1,2,3]

using three velocity components, respectively:
2m 2 m
Ttr,i = Ci /2 = Ci2 . (2.51)
kB kB
The differences between Ttr,i∈[1,2,3] reflect the degree of local non-equilibrium of the
translational motion among its three degrees of freedom.
Diatomic and polyatomic molecules also have rotational and vibrational degrees
of freedom. For the intramolecular energy mode with the number l of degrees of
freedom, the temperature Tintra of this energy mode can be similarly defined using
the average energy of this mode per unit mass eintra :
2meintra
Tintra = . (2.52)
lkB
2.5 Calculations of Macroscopic Properties 45
The energy equipartition principle implies Ttr = Ttr,∀i∈[1,2,3] = Tintra at equilibrium

state. For non-equilibrium state, the overall kinetic temperature Toa can be defined:
3Ttr + lTintra
Toa = . (2.53)
3+l
Note that the ideal gas equation of state does not apply to this overall temperature
at non-equilibrium state. In the following, Ttr will be denoted by T for simplicity,
unless stated otherwise.
2.6 Moment Equation
The definitions of different moments have been given in the last Sect. 2.5 and the
moment equation will be discussed in this section. The collision integral of arbitrary
quantity Q = Q(c) carried by a single molecule moving at c is the increase of Q
per unit physical volume per unit time due to intermolecular collisions and defined
using the collision term of the Boltzmann equation as follows:

∂f
Δ[Q] = Q( )coll dc
R3 ∂t
4π (2.54)
= Q( f f p1 − f f p1 )cr σ dΩdcp1 dc,
R3 R3 0
which can be rewritten into the following alternative forms thanks to the symmetry
between the direct and inverse collisions as used in Eq. (2.35) and the symmetry
between cp1 and c:
4π
1
Δ[Q] = (Q + Q p1 − Q − Q p1 )( f f p1 − f f p1 )cr σ dΩdcp1 dc
4 R3 R3 0
4π (2.55)
1
= (Q + Q p1 − Q − Q p1 ) f f p1 cr σ dΩdcp1 dc.
2 R3 R3 0
Equation (2.55) shows that Δ[Q] can be understood in a way different from that used
in its definition by Eq. (2.54). In Eq. (2.54), we first obtain (∂ f /∂t)coll due to all rele-
vant intermolecular collisions and then calculate the summation of Q(∂ f /∂t)coll over
all classes of c. In Eq. (2.55), we first obtain the increment (Q + Q p1 − Q − Q p1 )
due to each particular intermolecular collision (c, cp1 → c , cp1 ) and then calculate
the summation over all possible collisions, where the factor 1/2 is used to counteract
the double counting of collision times when integrating with respect to both c and
cp1 in the whole molecular velocity space R3 because the collision (cp1 , c → cp1 , c )
is already considered when counting the collision times of (c, cp1 → c , cp1 ).
In Eq. (2.33), we have seen that ( f dcdx)( f p1 dcp1 cr σ dΩdt) is the number of
intermolecular collisions (c, cp1 → c , cp1 ) inside dx during dt. Thus, the number of
all intermolecular collisions per unit physical volume per unit time is:
4π
1
Ncoll = f f p1 cr σ dΩdcp1 dc
2 R3 R3 0
(2.56)
1
= f f p1 cr σT dcp1 dc,
2 R3 R3
which is consistent with the expression in Eq. (2.55).

We call Q that satisfies Δ[Q] = 0 the summational invariant. Equation (2.55)
shows that Δ[Q] = 0 when Q is either of the collision invariants, e.g., Q =
m, mci∈[1,2,3] and mc2 /2 are collision invariants since Q + Q p1 − Q − Q p1 =
0, ∀(c, cp1 , Ω) according to the mass, momentum and energy conservation laws.
Furthermore, it can be proved that the collision invariants and their linear combina-
tions are the only summational invariants [12].
In Sect. 2.5, we defined the average value of arbitrary molecular quantity Q by
computing the corresponding moment of f using Eq. (2.40). Similarly, multiplying
the Boltzmann equation, i.e., Eq. (2.38), by Q = Q(c) and integrating each term
over the entire velocity space, we obtain the moment equation (i.e., a moment of the
Boltzmann equation):

∂f ∂f ∂f
Q +g· +c· dc
R3 ∂t ∂c ∂x
4π (2.57)
= Q( f f p1 − f f p1 )cr σ dΩdcp1 dc,
R3 R3 0

which can be rewritten into (note: R3 g · ∂(Q f )/∂cdc = 0 as Q f → 0 at c → ∞.):
∂
(n Q) + ∇ · (ncQ) − ng · ∂ Q/∂c = Δ[Q]. (2.58)
∂t
Historically, Eq. (2.58) was derived by Maxwell in 1866 before the derivation of
the Boltzmann equation in 1872 but its Δ[Q] is written in the form of Eq. (2.55),
instead of Eq. (2.54) by using the intermolecular collision term of the Boltzmann
equation [21]. Thus, Eq. (2.58) is also called the Maxwell transport equation. Let Q =
m, mc and mc2 /2, respectively, we obtain the moment equation of gas dynamics:
∂ρ
+ ∇ · (ρu) = 0,
∂t
∂
(ρu) + ∇ · (ρuu) − ∇ · P − ρg = 0, (2.59)
∂t
∂
(ρetr ) + ∇ · (ρuetr ) − P : ∇u + ∇ · q = 0,
∂t
2.6 Moment Equation 47
where u = c, P = −ρCC according to Eq. (2.43), etr = C 2 /2 according to Eq. (2.47),

and q = ρCC 2 /2 according to the definition before Eq. (2.47). Equation (2.59) con-
tains ρ, u, P, etr and q as unknowns, among which there are 13 independent scalars
because P is symmetric and P j j = −2ρetr . In the derivation of the above energy equa-
tion, the mass conservation equation and momentum equation are used to obtain the
following equality:
∂ u2 u2 D u2
(ρ ) + ∇ · (ρu ) = 0 + ρ
∂t 2 2 Dt 2
Du
= ρu ·
Dt (2.60)
∂u
= ρu · ( + u · ∇u)
∂t
= (∇ · P + ρg) · u.
Note that the equation system of Eq. (2.59) is consistent with that of Eqs. (1.22),
(1.25) and (1.32), which are derived by using the continuum assumption to present the
conservation laws of mass, momentum and energy. Another difference between the
two derivations is that P and q in Eqs. (1.25) and (1.32) are defined using their physical
meanings and thus the equation system will be closed by adding the constitutive
equations based on physical measurements plus assumptions, e.g., Eq. (1.17) for
P and Eq. (1.18) for q. By contrast, P and q in Eq. (2.59) are defined using the
moments of f and their constitutive equations can be derived by the Chapman–
Enskog expansion of the Boltzmann equation, where the expressions of transport
coefficients can be obtained as well, e.g., Eq. (2.22) for μ.
Additionally, Grad proposed a different method to solve Eq. (2.59) by taking the
13 independent scalars (moments) as equal unknowns and obtained the Grad thir-
teen moment equations [10]. The unknown distribution function f is expanded into
Hermite polynomials and explicitly expressed using ρ, u, T , P and q in addition to C,
which is based on assumption. Then, 8 additional governing equations different from
the 5 ones of Eq. (2.59) can be obtained by substituting different Q into Eq. (2.58)
[21]. Consequently, we obtain a closed system with 13 independent equations for 13
unknowns.
2.7 H-Theorem
For single-component gas system of monatomic molecules in the absence of exter-

nal force, the homogenous steady state solution of the Boltzmann equation is the
equilibrium distribution:
3/2
m −m(c − u)2
f eq (c, x, t) = n exp , (2.61)
2π kB T 2kB T
which was first obtained by Maxwell in 1860 before the establishment of the Boltz-
mann equation and thus is called the Maxwell distribution (see a derivation introduced
in [21]). In dynamic problems, n, u and T are defined by moments of f (c, x, t) and
thus depend on x and t for a general definition of local f eq .
Actually, Eq. (2.61) can be mathematically derived by using the Boltzmann’s H-
theorem. For a spatially homogeneous single-component gas system in the absence
of external force, the Boltzmann equation becomes:

∂f 4π
= ( f f p1 − f f p1 )cr σ dΩdcp1 . (2.62)
∂t R3 0
The H function10 introduced by Boltzmann is defined as follows:

H= f ln f dc = nln f . (2.63)
R3
Multiplying Eq. (2.62) by Q = ln f and integrating each term with respect to c, we

obtain:
∂H
= Δ[ln f ]
∂t
4π (2.64)
1 f fp
= ln 1 ( f f p1 − f f p1 )cr σ dΩdcp1 dc,
4 R3 R3 0 f f p1

where ∂ H/∂t
= R3 (∂ f /∂t) ln f dc + ∂n/∂t = R3 (∂ f /∂t) ln f dc is substituted
since n = R3 f dc is constant for systems uniform in the x space according to the
mass conservation law. Obviously, ln[ f f p1 /( f f p1 )] and ( f f p1 − f f p1 ) always have
opposite signs or equal zero together and thus H cannot increase with time, i.e., the
Boltzmann’s H-theorem:
∂H
≤ 0. (2.65)
∂t
It can be further proved that H monotonically decreases with time and converges
to a finite bound that corresponds to the steady state with ∂ H/∂t = 0 [6, 8, 21].
Since the integrand of Eq. (2.64) is always less than or equal to zero, ∂ H/∂t = 0
happens when and only when the following equality holds:
ln f + ln f p1 − ln f − ln f p1 = 0, or f f p1 = f f p1 , ∀(c, cp1 , Ω), (2.66)
10 By a slight abuse of notation,

Hi denotes the Lame coefficient of arbitrary orthogonal curvilinear
coordinate introduced in Sect. 1.5 and H denotes a special function of the Boltzmann’s H-theorem
introduced here.
2.7 H-Theorem 49
which indicates that ∂ f /∂t = 0 holds at ∂ H/∂t = 0 according to Eq. (2.62) with the
consideration of whole intermolecular collisions, and vice versa according to the defi-
nition of H by Eq. (2.63). Furthermore, ∂ f /∂t = 0 holds at ∂ H/∂t = 0 also because
of the detailed balancing between each pair of inverse and direct intermolecular
collisions, i.e., f f p1 cr (σ dΩ) dcp1 dc = f f p1 cr σ dΩdcp1 dc = f f p1 cr σ dΩdcp1 dc.
Additionally, Eq. (2.66) indicates that Q = ln f at steady state is a collision
invariant and thus can be represented by a linear combination of the only three
elementary collision invariants (i.e., m, mc and mc2 /2). Further analysis on this linear
combination can prove that the steady state f of Eq. (2.66) must be an equilibrium
distribution, i.e., taking the form of Eq. (2.61) [6, 21]. The molecular beam technique
can accurately measure the molecular velocity distribution at equilibrium state and
confirm the validity of the Maxwell distribution. Additionally, Eq. (2.66) obtained by
using the Boltzmann’s H-theorem shows that the Maxwell equilibrium distribution
of Eq. (2.61) is the sufficient and necessary condition for homogenous steady state.
Without the Boltzmann’s H-theorem, we can only prove that Eq. (2.61) satisfies
∂ f /∂t = 0 according to Eq. (2.62), namely the Maxwell distribution is a sufficient
condition of homogenous steady state. But, it is impossible to prove that Eq. (2.61)
is also the necessary condition of ∂ f /∂t = 0 according to Eq. (2.62) alone, where we
can have ( f f p1 − f f p1 ) > 0 somewhere and ( f f p1 − f f p1 ) < 0 somewhere else
to make ∂ f /∂t = 0. It is noteworthy that the equilibrium distribution of discrete
description of a N -particle system is a more general Boltzmann distribution (see
Eq. (1.68) of [21]).
As mentioned before, the entropy S 11 of the whole system (instead of per unit mass
as in Sect. 1.3.5) is a measurement of disorder of microscopic states and Boltzmann
obtained:
S = kB ln W, (2.67)
where W is the number of microscopic states admissible at a uniform steady state (i.e.,
equilibrium state) specified by several macroscopic state variables. For systems with
a continuous probability distribution function f (c, x, t), its entropy can be defined
as follows [6]:

S = −kB f ln f dcdx, (2.68)
R3 R3
which can be generally applied to non-uniform or non-steady states (i.e., non-

equilibrium states), at which there is no thermodynamic definition of entropy, e.g.,
dSthdy = Mcv dT /T + MkB dV /(V m) for ideal gases, where cv is the isochoric spe-
cific heat capacity, M and V are the total mass and volume of the system, respec-
tively. But, this kinetic definition is consistent with the thermodynamic definition of
entropy at equilibrium state, where S of Eq. (2.68) can be expressed using n and T
11 By a slight abuse of notation, Si j denotes the component of strain rate tensor introduced in
Sect. 1.3.1, S denotes the entropy of whole system introduced here, a characteristic channel size
introduced in Sect. 3.4.5 and the saturation introduced in Sect. 4.5.4.
via f = f eq and then Sthdy − S = const is independent of the gas state (see Sect. 4.2
of [6]). Additionally, the Boltzmann’s H-theorem indicates that the entropy S defined
by Eq. (2.68) will always increase with time due to intermolecular collisions (see
Sect. 4.13 of [6]). This irreversibility of natural processes of gas systems should
be understood in a statistical way since the Newton’s laws used for describing the
intermolecular collision dynamics is mathematically reversible with time [6, 8].
2.8 Properties at Equilibrium State
In this section, many useful scalar properties of gas at equilibrium state will be derived
using the spherical coordinate system (C, eq
θ, ϕ) since f eq of Eq. (2.61) is spherically
symmetric in the C space. Note that R3 ( f /n)dC = 1 and dC = C 2 sin ϕdCdθ dϕ.
Thus, the probability density function f P (C) of the absolute value of C can be
obtained as follows:
π 2π eq
f
f P (C) = C 2 sin ϕdθ dϕ
0 0 n
(2.69)
4 3 2
= √ β C exp(−β C ), 2 2
π
√
where β = m/(2kB T ).
It is easy to obtain the most probable thermal speed Cmp = 1/β, at which
d f P /dC = 0 and f P takes the maximum. By contrast, the most probable values
of C and C∀i∈[1,2,3] are zero according to Eq. (2.61). The mean thermal speed is:
∞

8kB T
C= C f P dC = . (2.70)
0 πm

The root-mean-square thermal speed Crms = C 2 is:

∞
3kB T
Crms = C 2 f P dC = . (2.71)
0 m
Thus, we have Cmp < C < Crms at equilibrium state as a consequence of the high-
speed tail of the distribution function.
When the molecules are modeled as spheres not subject to intermolecular force
fields, their motion between successive intermolecular collisions is free from any
mutual influences. The intermolecular collision frequency ν and mean free path λ
will be defined at equilibrium state by using the Maxwell distribution as well. The
number Ncoll of all intermolecular collisions per unit physical volume per unit time
is given in Eq. (2.56). Accordingly, the collision frequency ν of each molecule is:
2.8 Properties at Equilibrium State 51
ν = 2Ncoll /n

1 (2.72)
= f f p1 cr σT dcp1 dc,
n R3 R3
where f = f eq and the factor 2 is introduced because each molecule experienced

one instead of half collision if only one collision happened between two molecules,
for instance.
We can define the mean value of arbitrary Q = Q(cr ) as follows12 :

1
Q(cr ) = Q f f p1 dcp1 dc. (2.73)
n2 R3 R3
√
Then, we obtain (cr ) j = (2/ π )Γ [( j + 3)/2](2kB T /m r ) j/2 . In particular, the mean
value of the relative speed at equilibrium state is:

kB T √
cr = 4 = 2 × C. (2.74)
πm
For the particular HS molecules with fixed σT , we have ν = cr σT n = cr σT n and

thus:

kB T √ 5p
ν = 4σT n = 2σT nC = , (2.75)
πm 4μ
where μ = μ(σT ) of Eq. (2.24) is substituted. For the VSS molecules in general, σT
depends on cr and we have ν = 5(α + 1)(α + 2) p/[α(7 − 2ω)(5 − 2ω)μ] [5, 21].
The average time interval between two successive intermolecular collisions of
each molecule is 1/ν and thus the mean free path λ of HS molecules is:
λ = C/ν

4μ 8kB T
=
5p πm (2.76)

16μ m
= ,
5ρ 2π kB T
which will be used in the simulations of Chap. 3 to determine K n, unless stated

otherwise. For methane molecules at STP as an example, we have λ ≈ 50 nm that is
much larger than its molecular size of about 0.37 nm. Thus, the correlation between a
12 It is different from the mean value of collision quantity over all collisions, which is used in
discussing the chemical reaction rate can be generally defined via the collision frequency ν of
and
Eq. (2.72), i.e., Q(cr )coll = (nν)−1 R3 R3 Q f f p1 cr σT dcp1 dc [5, 21].
collision pair at the beginning of a free path (i.e., prior to the intermolecular collision)
is negligible and the molecular chaos assumption behind the Boltzmann equation is
reasonable.
In many cases (e.g., boundary condition) different from the applications of
Eq. (2.42), we need to calculate the molecular number flux across a surface element
d A only in the direction of n− . Let f = f eq by assuming that the non-equilibrium
effect around d A is negligible in calculating the number flux and then we obtain:

Nin = (c − ud A ) · n− f eq dc
(c−ud A )·n− >0
(2.77)
kB T √
=n [exp(−û in ) + π û in (1 + erf(û in ))],
2
2π m
√
where û in = (u − ud A ) · n− / 2kB T /m, ud A is the possible moving velocity of d A
and could be different from the√ flowvelocity u of neighboring gas, and the error func-
z
tion is defined as erf(z) = (2/ π) 0 exp(−y 2 )dy. In the case of û in = 0, we have:

kB T
Nin = n . (2.78)
2π m
The phenomenon of molecules at equilibrium state streaming through a small

hole in a thin wall (or through porous thin plate) into a vacuum chamber is called
molecular effusion or transpiration. The effusion speed of molecules can be com-
puted by Eq. (2.78) if the hole is small compared with λ in diameter (to avoid the
breach of equilibrium state due to noticeable flow speed) and in length (to avoid the
influence of molecular reflections at the hole surface to the distribution of effused
molecules). This result can be applied to the problem, where on each side of the
thin plate with small hole there is a tank containing the same species of gas but at
different temperature TL or TR . The effusion speed of each side to the other side is
Eq. (2.78). The two effusion speeds could be different at the beginning but become
equal at steady state by regulating the densities of the two sides, at which we have:

kB TL kB TR
nL = nR , (2.79)
2π m 2π m
√
which implies that pL / pR = TL /TR = 1 and is independent of the hole shape.
Additionally, Eq. (2.78) can also be applied to the differential effusion problem,
where a mixture of different gas species at equilibrium state effuses through a small
hole in a thin wall into a vacuum chamber. The number densities n a and n b of
two species during the effusion process are deemed unchanged and the backward
effusions are negligible in calculating the number densities n a and n b collected in
the vacuum chamber after a period of time. Note that Ta = Tb at equilibrium state.
According to Eq. (2.78), we have:
2.8 Properties at Equilibrium State 53

n a na mb
= , (2.80)
n b nb ma
which indicates that this differential effusion phenomenon can be used to separate
gases of different molecular weights and the purification effect can be improved by
multiple repetitions. The gas separation by the differential effusion at equilibrium
state is the simplest application based on the phenomena of rarefied gas dynamics.
More efficient dynamic separation can be achieved by the thermal transpiration
phenomenon driven by the temperature variation on the wall surface [16].
2.9 Simplification to Model Equation
Since the main challenge in solving the Boltzmann equation stems from the com-
plicated intermolecular collision term that has less influence to the solution at high
K n compared to the molecular collisions with the wall surface, attempts have been
made to model the collision term in a simple way. The widely used model equation
is the Bhatnagar–Gross–Krook (BGK) equation [2], which is supposed to be applied
to problems of small amplitude/perturbation as indicated in the title of the original
paper. As shown in the illustrative Fig. 2.8, the BGK approximation is based on
the fact that the effect of intermolecular collisions is to force the non-equilibrium
distribution inside the c space at arbitrary spatial point x back to a local equilibrium
distribution. It assumes that the approaching rate is proportional to the intermolecular
collision frequency ν of Eq. (2.72) and thus constant for all c. The BGK equation is
as follows:

relaxation process due to
intermolecular collisions
∂f ∂f ∂f f eq − f
+g· +c· = , (2.81)
∂t ∂c ∂x τ
where the relaxation time τ is independent of c, the local Maxwell distribution f eq
is defined by Eq. (2.61) and depends on x and t through local n, u and T , which
are defined by moments of the unknown f . Since f eq has the same n, u and T
as f , the BGK collision model satisfies the conservations of mass, momentum and
energy.
Additionally, it also satisfies the Boltzmann’s H-theorem, i.e., Δ[ln f ] =
R 3 (ln f ( f eq − f )/τ )dc ≤ 0.
It is obvious that the BGK equation gives correct solution in problems of collision-
less flows, where the collision term has negligible influence and thus the BGK equa-
tion is equivalent to the Boltzmann equation. For Maxwell molecular model with
σT ∝ cr−1 , the collision frequency ν of Eq. (2.72) is:

ν= f p1 cr σT dcp1 . (2.82)
R3
4π
Comparing − f /τ of the BGK equation with − R3 0 f f p1 cr σ dΩdcp1 = −ν f of
the Boltzmann equation, we have some kind of justification of the BGK model for
Maxwell molecules, e.g., by setting 1/τ = ν. For more realistic molecular models,
1/τ is different from the intermolecular collision frequency ν.
The N–S–F equations can be derived by using the Chapman–Enskog expansion
of the BGK equation as well and the obtained transport coefficients are related to the
relaxation time τ as follows [23]:
μ = τ nkB T,
5kB (2.83)
ζ = τ nkB T ,
2m
which indicates that the BGK equation has PrBGK = cp μ/ζ ≡ 2mcp /(5kB ). The
isobaric specific heat capacity cp and thermal conductivity coefficient ζ are related
to the excited degrees of freedom. For ideal gases of p = nkB T , we have cp = cv +
kB /m, where the isochoric specific heat capacity cv equals 3kB /(2m) for monatomic
gas molecules with three translational degrees of freedom in total. Thus, PrBGK is
equal to 1 for monatomic gas molecules while the correct value is 2/3 (see Eqs. (6.52),
(6.53) of [21], or Eqs. (12.1.6) and (13.1.2) of [6]). The actual Prandtl numbers for
usual gas molecules are around 0.7–1.0.
According to the phenomenological models, the energy exchange between the
internal and translational energy can be superimposed upon an otherwise monatomic
molecular model [5]. For polyatomic gas molecules at usual temperature condition,
the energy transformation associated with the vibrational mode is negligible but the
rotational mode is fully excited due to very low characteristic temperature of rotation.
Thus, the mean value of molecular rotational energy per unit mass erot is close to but
smaller than lkB T /(2m) at non-equilibrium state, where T is the translational kinetic
temperature as reference and l is the number of rotational degrees of freedom. The
2.9 Simplification to Model Equation 55
Prandtl number for polyatomic gas molecules can be roughly understood as follows:
we first neglect the rotation in modeling polyatomic gas molecules by using correct
m and μ and obtain an approximate solution of f (c, x, t), and then superimpose
additional molecular variable erot to each molecule,13 which doesn’t change μ since
the stress tensor P is unchanged, but increases cp by a factor (5 + l)/5 and ζ =
|n(mC 2 /2 + erot )C|/|∇T | by a factor smaller than (3 + l)/3. Thus, Pr ∝ cp /ζ of
polyatomic gas molecules is almost unchanged, i.e., around 2/3. The above analysis
assumed that the solution of f (c, x, t) will not be noticeably changed by adding
molecular variable erot , which is reasonable because the relaxation process of Trot
toward T is fast and completed within the order of five to ten mean collision time
[5] if neglecting (∂ f /∂t)force and (∂ f /∂t)move . Additionally, the relaxation process of
the translational modes (i.e., Ttr,i∈[1,2,3] toward T ) is slightly faster than that of the
rotational modes but the relaxation process of vibration mode at usual temperatures
is at least two or three orders of magnitude slower than the former two [5].
The incorrect Prandtl number of the BGK equation indicates that the actual values
of μ and ζ can not be matched simultaneously in the BGK equation. Consequently,
the BGK equation is inaccurate in solving problems, where the momentum and
energy exchanges are equally important even if the solution of f is close to its
local f eq (e.g., small perturbation) because the obtained local n, u and T for f eq
could be inaccurate. It is noteworthy that the ellipsoidal statistical model [1, 11] can
match realistic Pr < 1, where an anisotropic Gaussian distribution is used to replace
the Maxwellian (isotropic Gaussian) distribution f eq of the original BGK equation.
Additionally, accurate Pr can also be achieved in the BGK-like model equation if the
collision frequency/relaxation time is taken as a function of the molecular velocity
[4, 22]. A comparison of the two variants of BGK equation is given in [20].
In some studies, we are interested only in solving f (c, x, t). According to the
above superimposing rule, the translational movement associated with c (i.e., the
solution of f (c, x, t)) can be simulated by using molecular model without rota-
tional mode for both monatomic and polyatomic gas molecules. For problems with
heat exchange as the dominant mechanism in the molecular collisions and move-
ments but having μ given, the effective ζeff of polyatomic molecules after neglect-
ing the intramolecular energy mode can be determined by using Pr = 2/3 and
cp = 5kB /(2m), i.e., ζeff = cp μ/Pr = 15kB μ/(4m). The computed ζeff is equal to
the real ζ for monatomic molecules. Then, we match ζeff according to Eq. (2.83)
by using τ = 3μ/(2nkB T ) for both monatomic and polyatomic molecules. If ζ is
given instead of μ, we need to first determine μ by using the real values of Pr
and cp , i.e., μ = Pr ζ /cp , and then match ζeff (μ) via τ = 3μ/(2nkB T ) as discussed
above because the simulated molecules of current DSBGK simulation have no rota-
tional mode to reflect the real ζ . For problems with momentum exchange as the
13 By a slight abuse of notation, erot is usually used as the mean value but here also as the value
carried by each molecule for notation clarity, eα,i denotes the component of lattice velocity eα
introduced in Sect. 4.1.
dominant mechanism, we match μ by using τ = μ/(nkB T ) for both monatomic and

polyatomic molecules, where μ is either given or computed by using a given ζ via
μ = Pr ζ /cp as mentioned before.
2.10 Boundary Conditions
For open boundary, the associated physical process is that molecules are removed
from the flow domain when moving across the open boundary, where new molecules
moving toward the flow domain are simultaneously generated. The open boundary is
assumed to be static (i.e., ud A = 0) without loss of generality and the inward normal
direction (i.e., toward the inside of flow domain) of its surface element d A is denoted
by n− . In the boundary condition for the molecules of c · n− > 0, the distribution
function f is assumed to be the local Maxwell distribution f eq , where the number
density n and temperature T are usually prescribed while the flow velocity u could
be prescribed as well or dynamically computed by extrapolation from the inside of
flow domain (or using the value at the adjacent cell/grid without extrapolation for
simplicity). In Monte Carlo molecular simulations, the number flux of real molecules
can be calculated by Eq. (2.77) to determine the number of new simulated molecules
that will be generated at the open boundary. The representative velocity of each
new molecule can be randomly selected according to an algorithm derived from
f eq [5]. Note that the new simulated molecules include all molecules entering the
computational domain through the open boundary during the current time step Δt,
a part of which will be spent for the subsequent movement after crossing the open
boundary. Thus, a random time period (i.e., R f Δt) is assigned to each new molecule
to update its spatial position for the current time step.
This section focuses more on the boundary condition at solid surface, which
describes how molecule coming at a certain velocity toward the surface reflects on
it. The interaction of the gas molecules with the body surface is the origin of the
drag, lift forces and heat exchange between the surface and gas flow. The application
performance of the Boltzmann equation depends a lot on the boundary condition of
gas-surface interaction particularly at high K n. Unfortunately, both the theoretical
and experimental studies on this problem are far from attaining perfection due to
its physical complexity. The difficulties are mainly due to our lack of knowledge
about the structure and composition of surface layer at the molecular scale and
hence the effective interaction potential of the gas molecules with the surface [8]. In
aerospace and other engineering practice, the complete diffuse reflection model or
the Maxwellian type model (i.e., a combination of specular and diffuse reflections)
has been used for a long period of time. The specular reflection model assumes that
the incoming molecules reflect on the body surface as elastic spheres, i.e., the normal
velocity component reverses its direction while the tangential velocity components
remain unchanged. Thus, the total shear stress subjected by the surface and the total
energy exchange between the surface and gas flow are zero. The complete diffuse
reflection assumes that all molecules regardless of their incoming velocities will be
2.10 Boundary Conditions 57
scattered to form a half-velocity-space Maxwell distribution at the surface, which

depends on the local wall temperature Twall , velocity uwall and an effective n eff that
can be determined by the number flux of all incoming molecules according to the
mass conservation of the reflection process.
We preset a local Cartesian coordinate system Slocal , which moves at uwall in
general and uses the subscripts 2 and 3 for the tangential directions and 1 for the out-
ward normal direction n+ (i.e., toward the inside of flow domain). In the discussion
of boundary condition, the subscripts 1, 2 and 3 always represent the corresponding
components in Slocal , unless stated otherwise. The incoming and reflected velocities
are observed in Slocal and denoted by ci and cr ,14 respectively. In numerical simula-
tions, c and uwall will be stored in the component form of the unique global Cartesian
coordinate system S and we need the transformation of ci = c − uwall from S to Slocal
to obtain the components ci,1 , ci,2 , ci,3 used in the following boundary conditions.
Finally, the generated cr,1 , cr,2 , cr,3 will be transformed from Slocal to S to obtain the
component form of c = cr + uwall in S.
Usually, the reflection process on the surface can be directly characterized by
the so called scatter kernel R(ci , cr )15 that might depend on x and t via the surface
properties as additional parameters. R(ci , cr )dcr is the probability for the molecules
coming at ci to leave with the reflected velocity being located inside dcr around
cr . We assume that there is no adsorption or leakage of incoming molecules on the
surface and then the scatter kernel needs to satisfy the normalization condition:

R(ci , cr )dcr = 1, ∀ci . (2.84)
cr ·n+ >0
It also needs to satisfy the non-negativity condition, i.e., R(ci , cr ) ≥ 0 for ∀ci , cr .
Note that the scatter kernel R is different from the distribution function f and
used to select representative cr for each particular ci . In particular, R is the normal-
ized velocity distribution of the collected molecules that reflect on the same surface
element during a period of time, while f is the velocity distribution of the collected
molecules that locate inside the same spatial volume element at the moment con-
cerned or during a period of time. The time period used to collect these samples
could be extended to get sufficient samples for steady state problems. Additionally,
f determines the molecular number flux toward the surface but R has nothing to do
with the number flux. Equation (2.84) shows that R describes the probability distri-
bution of cr inside the half velocity space of cr · n+ > 0 as a function that generally
contains ci and the local surface properties as parameters. On the other hand, f at
the surface and R are correlated through the number flux of incoming molecules.
14 By a slight abuse of notation, cr denotes the relative velocity for the intermolecular collision
process introduced in Sect. 2.2 and the reflected velocity for the molecular reflection process on the
solid surface introduced here.
15 By a slight abuse of notation, R denotes gas constants introduced in Sect. 2.5 and the scatter kernel
for the molecular reflection process on the solid surface introduced here.
We introduce f B (c) observed in Slocal at the reflection point xwall at the time
t to replace f in the following discussion for simplicity, i.e., f B (c) = f (c +
uwall , xwall , t). After getting f B (c), the boundary condition for f is obtained accord-
ingly. The distribution f B (ci ) of incoming molecules with ci · n+ < 0 is known. Note
that − f B (ci )(ci · n+ )dci is the number of incoming molecules of class ci striking on
unit area of the surface in unit time and correspondingly the probabilistic number
− f B (ci )(ci · n+ )dci R(ci , cr )dcr of them will reflect with velocity being located inside
dcr around cr . Thus, the distribution f B (cr ) of reflected molecules with cr · n+ > 0
can be determined as follows:

f B (cr )(cr · n+ )dcr = − R(ci , cr ) f B (ci )(ci · n+ )dci dcr . (2.85)
ci ·n+ <0
Taking integration of Eq. (2.85) with respect to cr over its half velocity space and
using the normalization condition of R, we obtain:

f B (cr )(cr · n+ )dcr = − R(ci , cr ) f B (ci )(ci · n+ )dci dcr
cr ·n+ >0 cr ·n+ >0 ci ·n+ <0
(2.86)
=− f B (ci )(ci · n+ )dci ,
ci ·n+ <0
which represents the mass conservation of the molecular reflection process.
2.10.1 Specular Reflection Model
According to the definition of specular reflection, its scatter kernel is:
Rspecular (ci , cr ) = δ(ci − cr + 2n+ (n+ · cr )), (2.87)
where the Dirac delta function δ 16 is defined as follows:

δ(c − a)Q(c)dc = Q(a), ∀Q, a. (2.88)
R3
Substituting Eq. (2.87) into Eq. (2.85) and using Eq. (2.88), we obtain:
f B,specular (cr ) = f B (ci ) = f B (cr − 2n+ (n+ · cr )). (2.89)
16 Bya slight abuse of notation, δi j denotes the Kronecker delta function introduced in Sect. 1.3.1,
δ = n −1/3 molecular average distance introduced in Sect. 2.2, δ(c) the Dirac delta function intro-
duced here, and δ0 a mean rarefaction parameter introduced in Sect. 3.4.3.
The implementation of the specular reflection model in Monte Carlo molecular sim-
ulations is straightforward, i.e., the normal velocity component changes sign at the
surface and then the molecule moves at the new velocity in the remainder of the
current time step into a new position.
2.10.2 Diffuse Reflection Model
By contrast, the complete diffuse reflection is defined by directly prescribing that

f B,diffuse (cr ) is a Maxwell distribution and thus has the following form in Slocal :
3/2
m −mc2r
f B,diffuse (cr ) = n eff exp . (2.90)
2π kB Twall 2kB Twall
Similar to Eq. (2.77), the number flux of reflected molecules is (note: û in = 0):

Nout = f B,diffuse (cr )(cr · n+ )dcr
cr ·n+ >0
(2.91)
kB Twall
= n eff .
2π m

The number flux of incoming molecules is given by Nin = − ci ·n+ <0 f B (ci )(ci ·
n+ )dci , which is known but usually has no analytical expression. According to the
mass conservation of Eq. (2.86), n eff can be determined to complete f B,diffuse (cr ):

2π m
n eff = Nin . (2.92)
kB Twall
The scatter kernel Rdiffuse can be determined from f B,diffuse (cr ) as we assumed
that Rdiffuse is independent of ci in its definition, namely Rdiffuse = Rdiffuse (cr ). Using
Eqs. (2.85), (2.86), (2.90) and (2.91), we obtain:
f B,diffuse (cr )(cr · n+ )

Rdiffuse (cr ) =
− ci ·n+ <0 f B (ci )(ci · n+ )dci
f B,diffuse (cr )(cr · n+ )
= (2.93)
cr ·n+ >0 f B,diffuse (cr )(cr · n+ )dcr
cr · n+ m −mc2r
= ( )2 exp( ).
2π kB Twall 2kB Twall
In Monte Carlo molecular simulations, the method of inversion of the accumu-

lative density function can be used to select cr according to the probability density
function Rdiffuse (cr ) since its accumulative density function Fdiffuse (cr ), i.e., an inte-
gral of Rdiffuse with respect to its argument from minus infinity to cr , can be inverted
relative to the argument cr . If F(cr ) cannot be inverted, the acceptance-rejection
method is needed. For R(cr ) with singularities, the generalized acceptance-rejection
method or the combined accumulative distribution and acceptance-rejection method
cr,1 function R(cr,1 ) to
can be used [21]. Taking a one-dimensional probability density
illustrate the method of inversion of F, we have F(cr,1 ) = −∞ R(y)dy, which is the
probability that the random variable takes on a value less than or equal to cr,1 . Then,
a representative cr,1 can be sampled as cr,1 = F−1 (R f ), where F−1 is the inverse
function of F and R f is a random fraction.
To obtain the accumulative density function of Rdiffuse (cr ), we rewrite Eq. (2.93)
into a product of three independent probability density functions:
Rdiffuse (cr ) =Rdiffuse (cr,1 ) × Rdiffuse (cr,2 ) × Rdiffuse (cr,3 )

β β
=2β 2 cr,1 exp(−β 2 cr,1
2
) × √ exp(−β 2 cr,2 2
) × √ exp(−β 2 cr,3
2
),
π π
(2.94)
√
where β = m/(2kB Twall ). In order to get analytical expression of F for inversion,
we transform cr,2 and cr,3 to cr,τ 17 and θ of the polar coordinate system, i.e., cr,2 =
cr,τ cos θ and cr,3 = cr,τ sin θ , and then obtain:
β2
Rdiffuse (cr,2 )Rdiffuse (cr,3 )dcr,2 dcr,3 = exp(−β 2 cr,τ
2
)cr,τ dcr,τ dθ
π (2.95)
θ
= exp(−β 2 cr,τ
2
)d(β 2 cr,τ
2
)d( ).
2π
Taking β 2 cr,τ
2
∈ [0, ∞) as a random variable, its probability density function is
Rdiffuse (β cr,τ ) = exp(−β 2 cr,τ
2 2 2
) and its accumulative density function is F(β 2 cr,τ
2
)=
1 − exp(−β cr,τ ). Note that the random fraction R f is equivalent to 1 − R f . Thus,
2 2
√
cr,τ can be sampled as cr,τ = − ln(R f 1 )/β. Additionally, θ is uniformly distributed
between 0 and 2π since Rdiffuse (θ/(2π )) ≡ 1 and thus can be sampled as θ = 2π R f 2 .
Then, cr,2 and cr,3 are obtained from the sampled cr,τ and θ .
Equation (2.95) indicates that the probability density function of cr,τ is:
Rdiffuse (cr,τ ) = 2β 2 cr,τ exp(−β 2 cr,τ

2
), (2.96)
which is the same √ as Rdiffuse (cr,1 ) of Eq. (2.94). Thus, cr,1 can be sampled the same
as cr,τ , i.e., cr,1 = − ln(R f 3 )/β.
Additionally, Maxwell proposed a hybrid model composed of the specular reflec-
tion and diffuse reflection, where the α 18 portion of the incoming molecules reflects
17 By a slight abuse of notation, variable τ denotes the relaxation time of the BGK equation intro-
duced in Sect. 2.9 and subscript τ denotes the tangential direction introduced here.
18 By a slight abuse of notation, α denotes the coefficient of coordinate transformation introduced in
ij
Sect. 1.3.1; α denotes a soft parameter of the VSS molecular model for the intermolecular collision
diffusely and the other (1 − α) portion reflects specularly, which is called the
Maxwellian type boundary condition.
2.10.3 CLL Reflection Model
Experiments with engineering surfaces in contact with gases at usual temperatures

indicate that the reflection process approximates diffuse reflection with complete
thermal accommodation. This behavior may be a consequence of such surfaces being
microscopically rough with the incident molecules suffering multiple scattering, or of
the molecules being momentarily trapped or adsorbed on the surface [5]. For surfaces
processed by usual industrial means (i.e., not polished surface) and at usual tempera-
ture (i.e., with not too high incident energy of the incoming molecule), the complete
diffuse reflection model can serve as a fairly good approximation of the molecular
interaction with the body surface [21]. But, for clean and precisely processed surfaces
at high temperature and in high vacuum environment (no adsorption at the surface),
the experimental investigation shows that the complete diffuse reflection model is
not accurate. The molecular beam experiments also show that the distribution of
reflected velocity cr has the leaf blade form peaked in the specular direction of the
incoming velocity ci . Thus, more realistic molecular reflection models are needed
and they should satisfy the reciprocity principle or principle of the detailed balance
between the direct and inverse reflections (see a derivation introduced in [21]):
−mc2r −mc2i
(cr · n+ )R(−cr , −ci ) exp( ) = −(ci · n+ )R(ci , cr ) exp( ), (2.97)
2kB Twall 2kB Twall
where the direct and inverse reflections have (ci → cr ) and (−cr → −ci ), respec-
tively. Reciprocity is a subtle property of the circumstance, where the microscopic
dynamics is time reversible and the wall is assumed to be in a local equilibrium state,
which is not significantly disturbed by the impinging molecules [8].
Cercignani and Lampis proposed a more appropriate phenomenological model
RCL (ci , cr ) [7] that is positive and satisfies the normalization condition Eq. (2.84)
and the reciprocity principle Eq. (2.97). The numbers of molecules scattered into
given directions calculated by using this model agree very well with the experimental
results of the molecular beam measurement, where molecular beam with a certain
velocity distribution or collimated incident velocity is directed to strike onto the
body surface and the distribution of reflected velocity is determined by counting the
number of molecules scattered in certain direction [21]. Lord developed this model
and implemented it in Monte Carlo molecular simulations [18]. Thus, it is called the
CLL model. Its scatter kernel is the product of three independent parts associated
with the three velocity components, respectively:
process introduced in Sect. 2.3, and the accommodation coefficient for the molecular reflection
process on the solid surface introduced here, and subscript α denotes the velocity index of the
lattice Boltzmann method introduced in Sect. 4.1.
√
1 −(ĉr,2 − 1 − ατ ĉi,2 )2
RCL (ci , cr ) = √ exp ×
π ατ ατ
√
1 −(ĉr,3 − 1 − ατ ĉi,3 )2
√ exp ×
π ατ ατ
(2.98)
ĉr,1 −(ĉr,1
2
+ (1 − αn )ĉi,12
)
exp
π αn αn
2π √
2 1 − αn ĉr,1 |ĉi,1 |
exp cos θ dθ,
0 αn
√
where |ĉi,1 | is the absolute value of ci,1 / 2kB Twall /m as ci,1 = ci · n+ < 0,
RCL (ci,1 , cr,1 ) is the same as RCL (ci,τ , cr,τ ) that can be derived from RCL (ci,2 , cr,2 )
and RCL (ci,3 , cr,3 ) as in the diffuse reflection model, ατ is the accommodation coef-
ficient of kinetic energy for the tangential velocity components and satisfies:
∞ 2
2
ci,2 − cr,2 2 2
ci,2 − −∞ cr,2 RCL (ci,2 , cr,2 )dĉr,2
ατ = 2
= 2
, ∀ci,2 , (2.99)
ci,2 ci,2
√
because we have cr,2 = 1 − ατ ci,2 according to Eq. (2.98). Obviously, cr,2 ≡ 0 if
ατ is equal to 1 (i.e., completely accommodated to the situation of the surface) as
in the complete diffuse reflection model and cr,2 ≡ ci,2 if ατ = 0 as in the specular
reflection model.
Note that the energy accommodation coefficient used in other reflection mod-
els could be differently defined, e.g., α = [(ci2 ) − (cr2 )]/[(ci2 ) − diffuse (cr2 )],
where the energy fluxes are defined similar to Eq. (2.42) but in half velocity space, i.e.,
(Q i ) = − ci ·n+ <0 Q i ci · n+ f B (ci )dci and (Q r ) = cr ·n+ >0 Q r cr · n+ f B (cr )dcr
while diffuse (Q r ) is computed with f B,diffuse (cr ) of Eq. (2.90) by assuming that all
molecules are scattered according to the complete diffuse reflection model. For exam-
ple, the Maxwell’s diffuse-specular reflection model has R = (1 − α)Rspecular +
α Rdiffuse and thus f B (cr ) = (1 − α) f B,specular (cr ) + α f B,diffuse (cr ), which
means that
α should be defined above as cr ·n+ >0 Q r cr · n+ f B,specular (cr )dcr ≡ − ci ·n+ <0 Q i ci ·
n+ f B (ci )dci , ∀ f B (ci ). This different definition is valid for arbitrary Q = Q(c) as
long as Q i = Q r holds for the specular reflection model. Consequently, the tangen-
tial momentum accommodation coefficient is equal to the energy accommodation
coefficient for the Maxwell’s diffuse-specular reflection model, while they are usu-
ally different as observed in the MD simulations for surfaces of incomplete diffuse
reflection.
As we can see, f B,CL (cr ) (i.e., the boundary condition for f ) is determined from
Eqs. (2.85) and (2.98), which is very complicated for any given f B (ci ) (e.g., the
simplest case with f B (ci ) as a Maxwell distribution). On the other hand, its applica-
tion in Monte Carlo molecular simulations is simple [18]. As shown in Fig. 2.9,
we first introduce the coordinate transformation between (c2 , c3 ) and (v, w) to
make vi ≡ (ci,2 2
+ ci,3 ) and wi ≡ 0 for arbitrary incoming velocity components,
2 1/2

selecting representative cr,2
and cr,3 for given ci,2 and ci,3
in the CLL reflection model
and then cr,2 = vr cos θ − wr sin θ and cr,3 = vr sin θ + wr cos θ for the reflected
velocity components, where θ is the azimuthal angle of incoming velocity com-
ponent (ci,2 , ci,3 ) in the c2 c3 plane of Slocal , vr and wr will be selected by using the
given vi and wi . Furthermore, we introduce the coordinate transformation between
(vr , wr ) and (rτ , ϕτ ), i.e., vr = [(1 − ατ )(ci,2
2
+ ci,3
2
)]1/2 + (2kB Twall /m)1/2 rτ cos ϕτ ,
wr = (2kB Twall /m) rτ sin ϕτ . Similar to Eq. (2.98) for cr,2 and cr,3 , the CL scatter
1/2
kernel for vr and wr has the following property:
RCL (vi , vr )RCL (wi , wr )dv̂r dŵr

√
1 −(v̂r − 1 − ατ v̂i )2 −ŵr2
= exp exp dv̂r dŵr
π ατ ατ ατ
2
1 −rτ (2.100)
= exp rτ drτ dϕτ
π ατ ατ
2 2
−rτ r ϕτ
= exp d τ d ,
ατ ατ 2π
which indicates rτ = (−ατ ln R f 1 )1/2 , ϕτ = 2π R f 2 , where R f 1 and R f 2 are indepen-

dent random fractions. Then, the selected rτ and ϕτ are used to compute the represen-
tative
cr,2 and cr,3 via vr and wr . The representative normal component is selected as:
cr,1 = (1 − αn )ci,1
2
+ (2kB Twall /m)rn2 + 2(1 − αn )1/2 |ci,1 |(2kB Twall /m)1/2 rn cos ϕn ,
where rn = (−αn ln R f 3 )1/2 , ϕn = 2π R f 4 , and R f 3 , R f 4 are two additional random
fractions. This selection of cr,1 is the same as (cr,2
2
+ cr,3 ) = (vr2 + wr2 )1/2 .
2 1/2
We get cr,2 = ci,2 (1 − ατ ) + (2kB Twall /m) rτ cos(ϕτ + θ ) and cr,3 = ci,3
1/2 1/2
(1 − ατ )1/2 + (2kB Twall /m)1/2 rτ sin(ϕτ + θ ) after reorganizing the above formulas
of cr,2 and cr,3 . Note that ϕτ is selected uniformly from a periodic interval [0, 2π ]
and thus ϕτ + θ can be replaced simply by ϕτ for arbitrary θ , which implies that the
calculation of θ can be avoided to slightly improve the efficiency. Therefore, in the
CLL reflection model, the equivalent but simpler algorithm to compute the tangen-
tial components in Slocal is that cr,2 = ci,2 (1 − ατ )1/2 + (2kB Twall /m)1/2 rτ cos ϕτ and
cr,3 = ci,3 (1 − ατ )1/2 + (2kB Twall /m)1/2 rτ sin ϕτ [19]. This simpler algorithm also
degenerates to the complete diffuse reflection model when ατ = αn = 1.
References
1. Andries P, Le Tallec P, Perlat JP, Perthame B (2000) The Gaussian-BGK model of Boltzmann
equation with small Prandtl number. Eur J Mech B 19:813–830
2. Bhatnagar PL, Gross EP, Krook M (1954) A model for collision processes in gases. I. Small
amplitude processes in charged and neutral one-component systems. Phys Rev 94(3):511–525
3. Bird GA (1981) Monte Carlo simulation in an engineering context. In: Proceedings of the 12th
international symposium on rarefied gas dynamics, pp 239–255
4. Bouchut F, Perthame B (1993) A BGK model for small Prandtl number in the Navier-Stokes
approximation. J Stat Phys 71:191–207
5. Bird GA (1994) Molecular gas dynamics and the direct simulation of gas flows. Clarendon
Press, Oxford
7. Cercignani C, Lampis M (1971) Kinetic models for gas-surface interactions. Transp Theory
Stat Phys 1(2):101–114
8. Cercignani C (2006) Slow rarefied flows: theory and application to micro-electro-mechanical
systems. Birkhauser, Basel
9. Fan J (2002) A generalized soft sphere model for Monte Carlo simulation. Phys Fluids 14:4339–
4405
10. Grad H (1949) On the kinetic theory of rarefied gases. Commun Pure Appl Math 2:331–407
11. Holway LH (1966) New statistical models for kinetic theory: methods of construction. Phys
Fluids 9:1658–1673
12. Harris S (1971) An introduction to the theory of the Boltzmann equation. Holt, Rinehart, and
Winston, New York
13. Hassan HA, Hash DB (1993) A generalized hard sphere model for Monte Carlo simulations.
Phys Fluids A 5:738–744
14. Koura K, Matsumoto H (1991) Variable soft sphere model for inverse power law or Lennard-
Jones potential. Phys Fluids A 3:2459–2465
15. Koura K, Matsumoto H (1992) Variable soft sphere model for air species. Phys Fluids A
4:1083–1085
16. Kosyanchuk V, Kovalev V, Yakunchikov A (2017) Multiscale modeling of a gas separation
device based on effect of thermal transpiration in the membrane. Sep Purif Technol 180:58–68
17. Leaf B (1972) Derivation of the Boltzmann equation from the Liouville equation. Physica
59:206–227
18. Lord RG (1991) Some extensions to the Cercignani-Lampis gas-surface scattering kernel. Phys
Fluids 3(4):706–710
19. Li J (2012) Efficiency and stability of the DSBGK method. In: 28th international symposium
on rarefied gas dynamics. AIP conference proceedings, vol 1501, pp 849–856
20. Mieussens L, Struchtrup H (2004) Numerical comparison of Bhatnagar-Gross-Krook models
with proper Prandtl number. Phys Fluids 16:2797–2813
Berlin
References 65
22. Struchtrup H (1997) The BGK-model with velocity-dependent collision frequency. Continuum
Mech Thermodyn 9:23–31
23. Vincenti WG, Kruger CH Jr (1965) Introduction to physical gas dynamics. Wiley, New York
24. Wang Chang CS, Uhlenbeck GE (1964) Transport phenomena in polyatomic molecules. Uni-
versity of Michigan, Ann Arbor, CM 681
Chapter 3
Simulation Methods for Rarefied
Gas Flows
Abstract Low-speed gas flows of large Knudsen (Kn) number are characteristic
of micro/nano-electro-mechanical systems (MEMS/NEMS), vacuum system and
tight/shale gas reservoirs, and can be described by the kinetic theory. The traditional
direct simulation Monte Carlo (DSMC) method is the standard technique for mod-
eling gas flows at high Kn but computationally expensive and practically unfordable
at low speed due to stochastic noise. The discrete velocity method (DVM) is another
traditional method and deterministically solves the Boltzmann kinetic equation or
its simplified model equations. It can produce noise-free results that are accurate
when fine grid is used in the discretization of high-dimensional phase space, which
generally requires high computational cost in terms of memory usage and CPU time.
The direct simulation BGK (DSBGK) method was proposed recently to solve the
BGK model equation and has been comprehensively validated against the DSMC
method and experimental data in several benchmark problems over a wide range
of Kn. Although it is also a particle-based approach like the DSMC method, its
stochastic noise is very low and independent of the flow speed (or Mach number in
general), which is in sharp contrast to the DSMC method, where the stochastic noise
is inversely proportional to the square of Mach number. The algorithms of this three
simulation methods are introduced in this chapter and several benchmark problems
are studied to show their numerical performances in terms of accuracy, efficiency,
memory usage, as well as robustness associated with algorithm simplicity.
3.1 Direct Simulation Monte Carlo Method
Different from numerical simulation methods based on the Boltzmann-like equa-

tion, many physical simulation methods are based directly on the physics of the flow
instead of a mathematical model and they are usually implemented via the motions
and collisions of a large number of simulated molecules. The direct simulation Monte
Carlo method (DSMC) [3, 5] is a well-recognized physical simulation method for
rarefied gas flows of non-equilibrium state. The extension of DSMC method to prob-
lems of polyatomic molecules is straightforward by using appropriate relaxation
speed for each excited intramolecular energy mode modeled by additional molecu-

68 3 Simulation Methods for Rarefied Gas Flows
lar variable other than c, i.e., superimpose intramolecular modes upon an otherwise
monatomic molecular model with only translational modes associated with c. In
contrast, numerical simulation methods based on the Boltzmann-like equation for
polyatomic system, if available, will be very complicated. Additionally, the molecu-
lar dynamics (MD) method [1] is another well-known physical simulation method for
dense gas, liquid and solid phases. The major differences between the MD and DSMC
methods include: (1) the former uses a deterministic updating algorithm while the
later employs a random sampling scheme in handling the collisions; (2) the number
of simulated molecules of a given problem is not a free parameter in the MD simu-
lations while it can be arbitrarily prescribed and thus significantly smaller than the
number of real molecules in the DSMC simulations to save the computational cost;
and (3) the DSMC method uncouples the molecular motion from the intermolecular
collision over each time step, which should be much smaller than the mean collision
time, while the MD method models the real coupled process.
Note that both the Boltzmann equation and DSMC method have very close phys-
ical reasoning and are based on the same assumptions. Their major difference is
that the DSMC method does not depend on the existence of inverse collisions and
thus can be applied to complex phenomena such as ternary chemical reactions that
are inaccessible to the Boltzmann-like formulation [5]. Given the advantages of
the DSMC method over the numerical solvers of the Boltzmann-like equation, it is
unduly restrictive to limit the DSMC method to a solution of the Boltzmann equation.
However, it is still desirable if the DSMC algorithm for simple gas flow problem can
be derived from the Boltzmann equation because such derivation might help develop
new simulation methods for other Boltzmann-like equations, e.g., the Enskog equa-
tion as discussed in [9].
A derivation of the DSMC algorithm based on the Boltzmann equation is given
below. The DSMC algorithm in simple cases can be understood by using the impor-
tance sampling scheme to solve the Boltzmann equation [6, 22]. We consider the gas
flows of single-component monatomic molecules in the absence of external body
force. If the molecule is modeled by the hard sphere (HS) model with σ = D2 /4 and
σT = π D2 , the Boltzmann equation of Eq. (2.38) can be rewritten as follows:

∂f ∂f ∂f
+c· =
∂t ∂x ∂t coll
4π
(3.1)
1 D2
= (δp 2 + δp 1 − δp2 − δp1 )fp1 fp2 cr dΩdcp1 dcp2 ,
2 R3 R3 0 4
where f (c, x, t) is the unknown probability distribution function, c is the molec-

ular velocity, x is the spatial coordinate and t is the time, fp1 = f (cp1 , x, t) and
fp2 = f (cp2 , x, t), the Dirac delta functions δp1 = δ(cp1 − c), δp2 = δ(cp2 − c), δp 1 =
δ(cp1 − c), δp 2 = δ(cp2 − c) are defined by Eq. (2.88), cr = |cp1 − cp2 |, the post-
collision velocities cp1 , cp2 are determined by the pre-collision velocities cp1 , cp2
and the solid angle Ω according to Eqs. (2.3) and (2.17), dΩ = sin χ dεdχ , where
χ ∈ [0, π ] is the deflection angle of intermolecular collisions and ε ∈ [0, 2π ] is the
3.1 Direct Simulation Monte Carlo Method 69
angle between the collision plane and reference plane. Different boundary conditions
of the Boltzmann equation and their applications in Monte Carlo molecular simula-
tions have been discussed in Sect. 2.10. After getting the solution of f (c, x, t), the
number density n(x, t), flow velocity u(x, t) and temperature T (x, t) are computed
as follows:

n= f dc,
R
3
3 cf dc
u= R , (3.2)
n
3 (m/2)(c − u) f dc
2
T = R .
(3kB /2)n
In the DSMC simulation [3, 5], each simulated molecule l 1 carries two molecular
variables: position xl and velocity cl . In order to reduce the memory usage, each
simulated molecule represents number N of real molecules and thus the number of
simulated molecules could be much smaller than that of the real molecules contained
in the flow domain. Note that N is a constant and usually very large to make each
cell containing about 20 simulated molecules. The molecular position and velocity
are selected at the initial state according to the prescribed equilibrium distribution
and updated during the simulation process appropriately such that the simulated
molecules (i.e., [cl , xl ]all ) are distributed according to f in the phase space (c, x) at
arbitrary physical moment t.
The flow domain is divided into many cells and Vk denotes the volume of cell k.
Note that f dcdx is the number of real molecules in the velocity-space element dc
and the physical-space element dx. Thus, f dc of Eq. (3.2) can be replaced by N /Vk
to compute the macroscopic quantities by summation inside each cell k:

∈k N
nk = ,
Vk

∈k (N cl )
uk = , (3.3)
∈k N

[Nm(cl − uk )2 /2]
Tk = ∈k ,
(3kB /2) ∈k N

where ∈k is the summation over all simulatedmolecules l located inside the cell
k at the moment t of sampling. For example, ∈k N is the product of N and the
number of simulated molecules and so equal to the number of real molecules inside
the cell k.
1 By a slight abuse of notation, li denotes the coordinates of arbitrary orthogonal curvilinear coor-
dinate system introduced in Sect. 1.5; l denotes the number of degrees of freedom of the concerned
intramolecular energy mode introduced in Sect. 2.5, and subscript l denotes the index of simulated
molecule introduced here and the dominant density ρl introduced in Sect. 4.5.4.
During each time step Δt, we split ∂f /∂t into (∂f /∂t)move = −c · (∂f /∂x) due to
free molecular motions and (∂f /∂t)coll due to intermolecular collisions. As [cl , xl ]all
is a representative sample of f , (∂f /∂t)move can be reflected by updating xl alone via
uniform motions of simulated molecules.
For the intermolecular collision term (∂f /∂t)coll , we need to calculate the incre-
ment (Δf )coll after each Δt at all spatial points x and all velocity points c inside the
whole phase space. In order to make (Δf )coll tractable, we assume that the spatial
coordinates xl of those simulated molecules inside the same cell k are identical (i.e.,
notated by the same xcenter,k ). Then, we only need to compute (Δf )coll at xcenter,k of
each cell but still at all velocity points c. At spatial points different from xcenter,k inside
the cell k, we assumed f = 0, ∀c and thus (Δf )coll = 0. The distribution function f
at xcenter,k
is f k = ∈k δ(cl − c)N /Vk , which is consistent with Eqs. (3.2), (3.3), i.e.,
nk = R3 fk dc = ∈k N /Vk . At the end of each Δt and for each cell k, we compute
(Δfk )coll according to the Boltzmann equation2 :
∂fk
(Δfk )coll = Δt( )coll
∂t

Δt 4π
D2
= (δp 2 + δp 1 − δp2 − δp1 )fk,p1 fk,p2 cr dΩdcp1 dcp2
2 R3 R3 0 4
4π
D2 fk,p fk,p
= Npair G dΩ 1 dcp1 2 dcp2 , (3.4)
R3 R3 0 4σT nk nk

where fk,p1 = ∈k δ(cl − cp1 )N /Vk is the distribution function of cp1 at xcenter,k and
Δtn2k Vk (cr σT )max

Npair = ,
2N (3.5)
N cr σT
G= (δ + δp 1 − δp2 − δp1 ) .
Vk p2 (cr σT )max
The value of cr has upper bound here because fk is nonzero only at many but finite
velocity points cl . Although the value of (cr σT )max can be any constant in G, it should
be updated appropriately by the existing values of cr σT in all cells during each Δt
such that the ratio cr σT /(cr σT )max is always (note: practically will be almost always)
smaller than 1, which is required by the following acceptance-rejection scheme. On
the other hand, if (cr σT )max is much larger than required, the number Npair of tentative
collision pairs will be very large making the simulation process time-consuming as
shown in the following analysis.
4π
Note that 0 [D2 /(4σT )]dΩ = 1, R3 (fk,p1 /nk )dcp1 = 1, R3 (fk,p2 /nk )dcp2 = 1
and thus, (Δfk )coll is equal to Npair G, where G is the expected value of G in the
intermolecular collision process. The importance sampling
scheme is used to esti-
mate G, i.e., G ≈ ( ∈Nsample G j )/Nsample , where ∈Nsample G j is the sum of number
2 By a slight abuse of notation, G denotes the starting point of a molecular trajectory of a binary
collision introduced in Sect. 2.2, an auxiliary variable introduced here for notation clarity, and a
coefficient in the molecular interaction model of LBM introduced in Sect. 4.3.
3.1 Direct Simulation Monte Carlo Method 71
Nsample of representative G(Ω, cp1 , cp2 ) with Ω, cp1 , cp2 being selected according to
their probability density functions D2 /(4σT ), fk,p1 /nk , and fk,p2 /nk , respectively. Fur-
thermore, we let Nsample = Npair and obtain (Δfk )coll ≈ ∈Npair G j .
For each G j , we select simulated molecule pj1 randomly and uniformly from those
simulated molecules located inside the cell k and thus cp1 = cpj1 is selected accord-
ing to fk,p1 /nk as required because of fk,p1 = ∈k δ(cl − cp1 )N /Vk , which implies
that all simulated molecules should be selected with equal chance. The nonuniform
distribution of fk in the velocity space c is reflected by the number variation of sim-
ulated molecules inside different element dc of same volume. Similarly, we select
simulated molecule pj2 (j2 = j1 as molecular indices) randomly and uniformly inside
the cell k and use cpj2 as the jth representative value of cp2 , which also implies that
cp2 is selected according to fk,p2 /nk . Since D2 /(4σT ) is a constant, Ω is randomly
and uniformly selected from the whole surface of unit sphere, which is equivalent to
randomly selecting cpj and cpj according to the hard-sphere collision model because
1 2
Ω is used only to calculate the post-collision velocities. For other molecular mod-
els, we need to explicitly select Ω(ε, χ ) according to its probability distribution as
discussed in Sect. 2.3 and compute the post-collision velocities correspondingly.
Now, we have cpj1 , cpj2 , cpj , cpj and cr,j = |cpj1 − cpj2 | for each sampled G j .
1 2
Assuming that (cr σT )max has been updated for the current time step, G j is equal
to (N /Vk )(δp j + δp j − δpj2 − δpj1 )(cr,j σT )/(cr σT )max . Now, the acceptance-rejection
2 1
scheme is used to handle the fraction (cr,j σT )/(cr σT )max :
⎧
⎨ N (δ + δ − δ − δ ), if cr,j σT > Rf ,
pj2 pj1
G j = Vk pj2 pj1
(cr σT )max (3.6)
⎩
0, otherwise.

Note that fk = ∈k δ(cl − c)N /Vk and (Δfk )coll ≈ ∈Npair G j and so fk becomes

∈k δ(cl − c)N /Vk + ∈Npair G j after intermolecular collisions. This implies that

if (cr,j σT )/(cr σT )max > Rf , (δpj2 + δpj1 )N /Vk contained3 in ∈k δ(cl − c)N /Vk is
neutralized by (−δpj2 − δpj1 )N /Vk contained in G j and meanwhile (δp j + δp j )N /Vk
2 1
contained in G j is added to ∈k δ(cl − c)N /Vk , namely replacing (δpj2 + δpj1 )N /Vk
by (δp j + δp j )N /Vk in ∈k δ(cl − c)N /Vk . Till now, this replacement may con-
2 1
tribute nothing if we are discussing (Δfk )coll at velocity points c different from
the four special velocity points cpj1 , cpj2 , cpj and cpj because both the original
1 2
(δpj2 + δpj1 )N /Vk and the new (δp j + δp j )N /Vk are equal to zero at those nonspe-
2 1
cial points c. So, we consider (Δfk )coll at all velocity points c together and stipulate
that the same set of samples G j is used to compute (Δfk )coll at all points c of the
whole velocity space. Then, if (cr,j σT )/(cr σT )max > Rf , the contribution of G j to
(Δfk )coll in the whole velocity space is nonzero only at the four velocity points and
equivalent to changing the selected velocities cpj1 , cpj2 of fk to cpj , cpj , respectively,
1 2
which means that a pairwise intermolecular collision happens. In summary, we select
3 Since the simulated molecules pj1 and pj2 are selected from the cell k, the molecular index l of

∈k δ(cl − c)N /Vk equals j1 and j2 for these two simulated molecules, respectively.
number Npair of tentative collision pairs for each cell k at the end of each Δt and use
(cr,j σT )/(cr σT )max of each pair j1 , j2 as the acceptance probability to decide whether
a pairwise collision happens. This is the no-time-counter (NTC) DSMC algorithm
proposed based directly on the physical reasoning, e.g., Eqs. (3.5), (3.6) derived here
are the same as Eqs. (11.3), (11.4) obtained by physical analysis in [5].
For dense gases, the importance sampling scheme was used in [9] to solve the
Enskog equation, which is an extension of the Boltzmann equation by considering the
intermolecular repulsive force at short distance but still neglecting the intermolecular
attractive force at long distance. The attractive force is vital in simulating two-phase
flows [15]. For rarefied gas flows at low speed, the intermolecular collision integral
of the Boltzmann equation can be simplified and then evaluated by the importance
sampling scheme to improve the computational efficiency in the low-variance devi-
ational simulation Monte Carlo (LVDSMC) method [16]. The original LVDSMC
method conserves the mass on average in the intermolecular collision process and a
simple scheme was proposed in [23] to conserve the mass strictly.
3.2 Direct Simulation BGK Method
In the beginning of the development of the DSMC method, the mathematical tradition
in fluid mechanics was such strong that there were difficulties in gaining acceptance
for the emerging physical simulation method without any recourse to the conven-
tional mathematical models. A numerical solution of the approximate mathematical
model of the Boltzmann equation would be valued more highly than the results of
a physical simulation method that did not make such approximations.4 Even though
the relative advantage of the DSMC method increases with the complexity of the
flow and, for most cases of engineering interest, there was no practical alternative to
the DSMC method, it continued to suffer criticism for having a physical rather than
a mathematical foundation [5].
The direct simulation BGK (DSBGK) method is also a particle-based approach
and was recently proposed to improve the simulation efficiency of rarefied gas flows
at low speed [24]. It has been validated against the traditional DSMC simulations
and experimental study in several benchmark problems over a wide range of Kn [25,
28]. The development of the DSBGK method has been even worse than the DSMC
case due to the established tradition and the available options of other numerical
methods. The major dissent lies in the validity of the BGK relaxation model that is
the foundation of the DSBGK method. Some previous numerical solutions of the
BGK equation obtained by coarse solvers or at inappropriate conditions (e.g., high
speed) have left negative impressions. Additionally, the incorrect PrBGK number (see
the discussion in Sect. 2.9) also brings criticism although it is immaterial as shown
4 The computational approximations associated with the DSMC method are the ratio of the number
of simulated molecules to the number of real molecules, the finite time step over which the molecular
motion and collision are uncoupled, and the finite cell and sub-cell sizes in the physical space for
selecting neighboring collision pairs [5].
3.2 Direct Simulation BGK Method 73
by several independent studies in many low-speed problems, where the viscosity

and thermal conductivity coefficients are not equally important and thus only the
dominant one needs to be matched. Actually, the frequency of molecular collisions
with the wall surface is much higher than that of the intermolecular collisions at high
Kn, where the solution of gas flow depends more on the boundary condition and thus
the BGK simplification of intermolecular collision model has small influence.
We first consider gas flows of single component. In the absence of external body
force, the BGK equation of Eq. (2.81) can be rewritten into a Lagrangian form5 :
Df ∂f ∂f f eq − f
= +c· = , (3.7)
Dt ∂t ∂x τ
where the relaxation time τ is selected appropriately to satisfy either the viscosity
coefficient μ or the thermal conduction coefficient ζ according to Eq. (2.83), and the
equilibrium distribution function f eq is defined by Eq. (2.61).
As detailed in [25], the DSBGK simulation process is divided into a series of
time steps Δt and the flow domain is divided into many regular or irregular cells.
The average molecular displacement during a time step Δt and the cell size Δx
should be smaller than the molecular mean free path λ, which is the same as in the
DSMC method. A large number of simulated molecules are employed to represent
the distribution function f in the phase space and the evolution of f with time is
reflected by changing the molecular variables. The main idea of this method is to
track down the evolution of f along enormous molecular trajectories at constant
velocities, which are randomly selected according to the probability distribution
when simulated molecules are generated or reflected at the boundaries. In addition
to the molecular position xl and velocity cl used in the DSMC method, each simulated
molecule l of the DSBGK method carries two additional molecular variables: number
Nl of real molecules represented by the simulated molecule l, and Fl that is equal
to the representative value f (cl , xl , t) of f at the moment t and point (cl , xl ) in the
phase space. [cl , xl , Nl ]all is a representative sample of f and correspondingly [Fl ]all
is the representative value of f . The compatibility condition, namely [cl , xl , Nl ]all
and [Fl ]all are correlated via the same f , is required during the simulation process
and thus needs to be satisfied by the updating algorithms of free molecular motions,
intermolecular collisions and molecular reflections on the wall surface.
For the evolution of f due to free molecular motions, xl of [cl , xl , Nl ]all is changed
alone but [Fl ]all is unchanged along molecular trajectory. For the intermolecular col-
lisions, [Fl ]all is changed by integrating the BGK equation along molecular trajectory
and [cl , xl , Nl ]all is updated by changing Nl according to Fl .6 For the evolution of f
due to molecular reflection at xl on the wall surface, cl is changed to cnewl , Nl remains
unchanged to conserve the mass and then, Fl is updated to Flnew = f (cnew l , xl , t),
5 Different from the material derivative D/Dt = ∂/∂t + u · (∂/∂x) of material particles introduced
in Sect. 1.2, we have D/Dt = ∂/∂t + c · (∂/∂x) introduced here for molecules.
6 In the DSMC simulations, N ≡ N is a constant, and x is changed to represent the evolution of
l l
f due to free molecular motions, and cl is randomly changed for each selected collision pair to
represent the evolution of f due to intermolecular collisions.
which also satisfies the compatibility condition as both [cl , xl , Nl ]all and [Fl ]all have
been updated to represent the change of f .
To reduce the stochastic noises in the macroscopic variables of each cell k, the
transitional cell variables ntr,k , utr,k , and Ttr,k defined by Eq. (3.10) are adopted in
the BGK equation in place of the original nk , uk , and Tk defined by Eq. (3.3), respec-
tively, to update the molecular variables. In turn, these transitional cell variables will
be updated by using cl , xl and the increment (instead of transient value as in Eq. (3.3))
of Nl based on the mass, momentum and energy conservation laws of intermolecular
collision process. The increment of Nl is computed by an extrapolation of acceptance-
rejection scheme [24] using the smooth variation of Fl along molecular trajectory
due to the intermolecular collisions, which avoids the time-consuming process of
frequently generating random fractions (e.g., require several Rf for each tentative
collision pair in the DSMC method) and satisfies the compatibility condition by
directly correlating Nl with Fl . Consequently, this DSBGK collision algorithm sig-
nificantly reduces the stochastic noises of cell variables associated with the frequent
and random events of simulated molecules moving into and out of each cell.
3.2.1 Initialization Process
At the initial state, the cell variables ntr,k , utr,k and Ttr,k are usually uniform. The
initial molecular position xl and velocity cl are selected randomly as in the DSMC
simulation and then Fl is equal to f eq (cl , xl , 0), which contains ntr,k , utr,k and Ttr,k
as the macroscopic parameters that depend on x and t in general. The initial value
Nl,0 of Nl is usually the same for all simulated molecules and properly selected such
that the total number of simulated molecules takes an appropriate value for efficient
simulation. The smaller the value of Nl,0 is, the larger the total number of simulated
molecules at the initial state will be.
3.2.2 Molecular Motion and Intermolecular Collision
Each simulated molecule moves uniformly and in a straight line before encountering
the boundary. As we can see from Fig. 3.1, during each Δt, the molecular trajectory
may be divided into several segments by the cell interfaces or remain as a single
segment if not crossing any cell interfaces. As each segment is located inside a
particular cell k, Fl can be conveniently updated for each segment in sequence along
the trajectory according to the Lagrangian form of the BGK equation, i.e., Eq. (3.7).
Note that f eq of the BGK model is constant for each trajectory segment, associated
with a particular simulated molecule l moving at given cl inside a particular cell k
with given ntr,k , utr,k and Ttr,k . Thus, we can complete the integration of Eq. (3.7)
eq
along each segment, i.e., DFl /Dt = (fk − Fl )/τ , with respect to t and update Fl as
follows:
Fig. 3.1 Schematic of the DSBGK simulation, the trajectory of a simulated molecule during one
time step is divided into three segments by the cell interfaces (left) and the segments of all simulated
molecules located inside the same cell are used for summation to update the cell variables (right).
Different trajectory lengths over the same time step are due to different speeds (right)
eq eq
Flnew = fk + (Fl − fk ) exp(−Δk tl /τ ), (3.8)
where Flnew is the new value after the intermolecular collision, Δk tl is the time interval
used by the simulated molecule l during the current Δt to go through the segment
located inside the cell k. Since the molecular trajectory is divided first by the time
step Δt and then by the cell interfaces, Δk tl ≤ Δt. If the trajectory during the current
Δt is divided by the cell interfaces, Δk tl < Δt and Eq. (3.8) is used repeatedly to
update Fl for the successive segments in sequence.
This idea of tracking down the evolution of the velocity distribution function f
along each molecular trajectory is inspired by the lattice Boltzmann method (LBM)
[7, 14, 31] but there are noticeable differences between the LBM and DSBGK
method, e.g., their correlations of Eqs. (4.19) and (2.83) between relaxation time and
viscosity are different, and the discretization points in the spatial and velocity spaces
are fixed in the LBM but changing with the molecular movements and collisions in
the DSBGK method.
After updating Fl for each segment, Nl is updated correspondingly:
Nlnew = Nl Flnew /Fl , (3.9)
which is based on an extrapolation of acceptance-rejection scheme that if [cl , xl , Nl ]all

is a representative sample of f1 , [cl , xl , Nl (f2 /f1 )l ]all is a representative sample of
f2 , where (f2 /f1 )l is the ratio of f2 and f1 at the point (cl , xl ) [24]. By contrast,
assuming f2 < f1 , ∀(c, x), the original acceptance-rejection scheme takes (f2 /f1 )l as
the acceptance probability to decide for each simulated molecule l whether it should
be retained or removed, and then [cl , xl , Nl ]retained is a representative sample of f2 .
For the physical process along each trajectory segment, Fl is updated to keep
[Fl ]all representative, and meanwhile xl and Nl are updated to keep [cl , xl , Nl ]all
representative. In order to make Eq. (3.9) easy for understanding, we uncouple the
intermolecular collision from the molecular motion for each segment only in this
explanation (not in the algorithm implementation): in the free molecular motion, xl
is updated with t to the end of the segment concerned but cl , Fl , Nl keep unchanged as
f (cl , xl + Δk tl cl , t + Δk tl )=f (cl , xl , t); in the intermolecular collision, Fl is updated
to Flnew but cl , xl , t keep unchanged, and thus Nl is correspondingly updated to Nlnew
by Eq. (3.9). The precondition of using this extrapolation of acceptance-rejection
scheme is that the compatibility condition between [cl , xl , Nl ]all and [Fl ]all is satis-
fied at the beginning. Then, the updating algorithms of xl , Fl , Nl with t for the free
molecular motion and intermolecular collision processes will maintain the compat-
ibility condition thanks to using the extrapolation of acceptance-rejection scheme.
eq
The cell variables ntr,k , utr,k , Ttr,k of fk are used in Eq. (3.8) to update Fl and in
turn, updated at the end of each Δt by using the increment of Nl via Fl . During the
current Δt and for each cell k (see Fig. 3.1 right), some simulated molecules move
inside/into/out of the cell k and their increments Δk Nl = Nlnew − Nl over segments
inside the cell k are already known. The physical meaning of Δk Nl is the number
increment of real molecules of class cl due to the intermolecular collisions inside
the cell k during the current time step. We compute the summation ∈k Δk Nl over
those segments inside the cell k during the current Δt. A simulated molecule may
contribute more than one term to the summation if it reflects on the wall surface back
into the cell k. Obviously, ∈k Δk Nl means the number increment of real molecules
of all existing classes due to the intermolecular collisions inside the same cell k during
the same Δt. Consequently, ∈k Δk Nl is expected to be zero according to the mass
conservation law of the intermolecular collision process inside the cell k. Usually, this
summation
is not exactly equal to zero due to numerical error. Thus, we decrease ntr,k
if ∈k Δk Nl is positive such that ∈k Δk Nl will decrease towards zero at the next Δt
because each term Δk Nl decreases with ntr,k according to Eqs. (3.8), (3.9), and vice
versa. It works
as an auto-regulation
scheme that makes ∈k Δk Nl approaching zero.
Similarly, ∈k (Δk Nl mcl ) and ∈k (Δk Nl mc2l /2) are the momentum increment and
kinetic energy increment, respectively, of real molecules of all existing classes due
to the intermolecular collisions inside the same cell k and during the same Δt. They
are also supposed to be zero according to the momentum and energy conservation
laws and thus can be used to update utr,k and Ttr,k by auto-regulation schemes. The
proposed auto-regulation schemes are:

ntr,k Vk − ∈k Δk Nl
nnew
tr,k = ,
Vk

ntr,k Vk utr,k − ∈k (Δk Nl cl )
unew
tr,k = , (3.10)
nnew
tr,k Vk

ntr,k Vk (3kB Ttr,k /2 + mu2tr,k /2) − 2 new new 2
∈k (Δk Nl mcl /2) − ntr,k Vk m(utr,k ) /2
new =
Ttr,k ,
new
ntr,k Vk (3kB /2)
tr,k , utr,k , Ttr,k are the new values for the next
where nnew Δt and Vk
new new
is the volume
of the cell k. The updating schemes of Eq. (3.10) make ∈k Δk Nl , ∈k (Δk Nl cl ),

∈k (Δk Nl mcl /2)converging to zero and then ntr,k , utr,k , Ttr,k will fluctuate around
2
their steady state solutions due to stochastic effect.

We use Nl , Fl to denote the previous values at the origin of the segment located
inside the cell k during the current Δt and use Nlnew , Flnew for the new values at
the end of that segment after intermolecular collision as in Eqs. (3.8), (3.9). Note
that all possible representative trajectories are selected according to the probability
of reflected velocity cr . Thus, it can be expected that the statistical feature of all
existing classes represents that of all possible classes and the summation over all
existing classes on average is equivalent to the integration over all possible classes
(e.g., Eq. (3.3) is equivalent to Eq. (3.2) on average). Similar to the transformation
applied in Eq. (3.3), we replace Nl by Vk Fl dcl , where dcl is the velocity-space element
around cl , since the compatibility condition is satisfied. Note that Δk tl is the time
interval used by the simulated molecule l inside the cell k during the current Δt and so
Δk tl = Δt for those simulated molecules moving within the same cell. To simplify
the analysis, we assume that Δt is very small to make most simulated molecules
moving within the same cell during each Δt, i.e., Δk tl Δt, ∀l. The summation
form of mass conservation for each cell k can be rewritten into an integral form:

Δk Nl = Nlnew − Nl
∈k ∈k ∈k

= (Vk Flnew dcl ) − (Vk Fl dcl )

∈k ∈k

DFl
≈ Vk Δk tl dcl (3.11)
Dt
∈k

f eq − Fl
= Vk k
Δk tl dcl
τ
∈k
eq
fk − f
Vk Δt dc,
R3 τ
where the first approximate equality is because Flnew − Fl is actually determined by

Eq. (3.8) that is different from the form used in the derivation of Eq. (3.11) although
the two forms have very close results, the second approximate equality is because of
Δk tl Δt, ∀l, and the representative
value Fl of f at cl is replaced by f at the end.
Thus, after convergence with ∈k Δk Nl = 0, we obtain
eq
fk − f
dc = 0. (3.12)
R3 τ
But, it is not necessary to require Δt being very small to satisfy

eq
Δk tl Δt, ∀l.
Even if it is not satisfied, ∈k Δk Nl = 0 still implies R3 (1/τ )(fk − f )dc = 0 since
both of them represent the mass conservation of intermolecular collision of the same
eq
evolution equation Df /Dt = (fk − f )/τ inside the cell k.
Similarly, the integral forms ofmomentum and energyconservations can be

∈k (Δk Nl cl ) = 0 and ∈k (Δk Nl mcl /2) = 0.
2
obtained after convergence with
Thus, the following equation is satisfied for each cell k after convergence:
eq
fk − f
Qi dc = 0, (3.13)
R3 τ
whereQ1 = 1, (Q2 , Q3 , Q4 ) = c, Q5 = mc2/2. As the original BGK equation sat-

eq
isfies R3 (1/τ )(f eq − f )Qi dc = 0, we have R3 (fk − f eq )Qi dc = 0, which implies
that ntr,k = n, utr,k = u, Ttr,k = T at each cell k after convergence according to the
eq
definitions of fk and f eq .
So, the transitional cell variables ntr,k , utr,k , Ttr,k obtained by the DSBGK method
are equal to the discrete solutions nk , uk , Tk of the BGK equation after convergence
although they are defined by Eqs. (3.10) and (3.3), respectively. Then, the transitional
eq
fk used in the DSBGK method is equal to the original f eq of the BGK equation inside
each cell k. Consequently, [Fl ]all and [cl , xl , Nl ]all are the representative value and
sample, respectively, of the solution f of the BGK equation, which implies that any
higher-order moment, including stress tensor and heat flux, calculated by the DSBGK
method agrees with that of solving the BGK equation by other accurate numerical
methods, e.g., the discrete velocity method (DVM) [4] when using fine-grids in the
molecular velocity space.
Note that the updating scheme of Eq. (3.10) for the intermolecular collision pro-
cess conserves the overall mass carried by both the cell and molecular variables
inside each cell k, i.e., ntr,k Vk + ∈k Nl , as (nnew V
tr,k k − n V
tr,k k ) + ∈k Δ k Nl = 0.
Additionally, ntr,k and Nl are unchanged during the molecular reflection process
on
the wall surface. Thus, the overall mass inside the computational domain
∈Domain n tr,k Vk + ∈Domain Nl is constant during the simulation process. By con-
trast, the overall momentum and kinetic energy carried by the cell and molecular
variables are changed during the molecular reflection process on the wall surface
although they are conserved during the intermolecular collision process when using
Eq. (3.10). It should be pointed out that the conservations of the overall mass, momen-
tum and kinetic energy by Eq. (3.10) are physically unnecessary because the physical
conservation laws apply only to the quantities carried by the molecular variables as
used in Eq. (3.11). Thus, Eq. (3.10) can be modified by adding arbitrary but positive
factors before ∈k Δk Nl , ∈k (Δk Nl cl ), ∈k (Δk Nl mc2l /2) to regulate the conver-
gence speed in open problems. But, the overall mass should be conserved in closed
problems such that
Converge

Converge
(3.13)
Converge
1

Converge
Nl ntr,k Vk = Nl + ntr,k Vk
∈Domain
= ∈Domain 2 ∈Domain ∈Domain
Initial (3.14)
(3.10) 1

Initial
Initial
Nl + ntr,k Vk = Nl ,
= 2 ∈Domain ∈Domain ∈Domain
which satisfies an important definite condition for closed problems that the total
Converge
number ∈Domain Nl of real molecules represented by the simulated molecules after

convergence is equal to the actual total number of real molecules, i.e., Initial
∈Domain Nl .
Note that we use the term total for the summation ∈Domain over the computational
domain to distinguish from the previous term overall for the summation over both
the cell and molecular variables.
Now, we explain why the cell variables ntr,k , utr,k , Ttr,k are updated by the auto-
regulation schemes of Eq. (3.10) rather than directly computed by Eq. (3.3) alike.
As we can see, Δk Nl computed by Eq. (3.9) is smooth since Fl variessmoothly
along
the trajectory according2to Eq. (3.8). Consequently, the summations ∈k Δk Nl ,
∈k (Δk Nl cl ), ∈k (Δk Nl mc
l /2) used
in Eq. (3.10)
contain small stochastic noise.
By contrast, the summations ∈k Nl , ∈k (Nl cl ), ∈k (Nl mc2l /2) of transient molec-
ular variables still have large stochastic noise due to frequent and random events of
simulated molecules moving into and out of the cell k. Therefore, nk , uk and Tk
computed by using them in Eq. (3.3) alike will be very noisy since the number of
simulated molecules inside each cell on average is small. This is the cause of notice-
able stochastic noise in the DSMC and MD simulations. Instead of using transient
values of molecular variables, their increments along molecular trajectories are used
in the DSBGK method to updates/regulates the cell variables. Although the molec-
ular variables entering into each cell are still random and noisy, their variations by
integrating the BGK equation along molecular trajectories are smooth. Consequently,
the noise in the cell variables is significantly reduced in the DSBGK method, com-
pared to the DSMC method and other particle simulation methods that define the
cell variables by using the transient molecular variables.
Regarding to the discretization accuracy of the unbounded molecular veloc-
ity space, the justification is the same for both the DSBGK and DSMC methods.
Although the number of simulated molecules per cell used by the DSBGK method
is small, the molecular velocities inside each cell are dynamically updated via the
frequent and random events of simulated molecules moving into and out of each cell
(from the perspective of Eulerian description), and the molecular velocities along all
representative trajectories are also dynamically updated via the frequent and random
molecular reflections at the boundary (from the perspective of Lagrangian descrip-
tion). Thus, the dynamic discretization of using few simulated molecules per cell in
the DSBGK method can sample from the whole velocity space and therefore allow
as fine discretization of the unbounded molecular velocity space as desired with the
increase of simulation time, which is the same as in the DSMC method.
The DSBGK algorithm described here is valid for arbitrary cell division, e.g.,
using parallelepipeds or tetrahedrons. In the DSMC simulations of problems with
complicated surface configuration, it is favorable to use regular parallelepipeds to
divide the computational domain, which makes it efficient to determine which cell
the simulated molecules are situated in at the end of each Δt for the selection of
neighboring collision pairs. The situated parallelepiped cell can be efficiently deter-
mined by the final molecular position xl of the current Δt. By contrast, the situated
tetrahedron cell needs to be searched along the molecular trajectory, which usually
involves several steps when the trajectory during the current Δt crosses more than
one cell. Although the use of parallelepipeds makes it time-consuming to determine

the molecular reflection position on the complicated surface configuration, the num-
ber of simulated molecules running into the surface during each Δt is usually much
smaller than the total number. Compared with using tetrahedrons to divide the com-
putational domain, which makes the determination of surface reflection positions of
few simulated molecules efficient but the determination of the situated cells of all
simulated molecules time-consuming, the gain of using parallelepipeds usually out-
weighs its loss. But, in the DSBGK simulations, the computational efficiency depends
less on the cell type because the molecular trajectories always need to be divided
into segments by the cell interfaces for the summations over segments located within
the same cell, which makes the algorithm of using regular parallelepipeds similar to
that of using tetrahedrons. Thus, the selection of cell type in the DSBGK simulations
will depend mostly on preference. As a particle-based approach, the DSBGK method
has many numerical advantages, including simplicity, robustness, convenience for
complicated surface configuration and for parallel computation, because the basic
algorithmic structure of the DSMC method is adopted.
3.2.3 External Body Force
In the presence of external body force, the BGK equation is as follows:
∂f ∂f ∂f f eq − f
+c· +g· = , (3.15)
∂t ∂x ∂c τ
where g is the external body force per unit mass and could depend on x and t in
general. We split ∂f /∂t into (∂f /∂t)move = −c · (∂f /∂x), (∂f /∂t)coll = (f eq − f )/τ
and (∂f /∂t)force = −g · (∂f /∂c). To simplify the algorithm, we uncouple the effect
due to (∂f /∂t)force from the other two effects for each Δt. At the end of each Δt
of the previous DSBGK algorithm without external body force, the effects due to
(∂f /∂t)move and (∂f /∂t)coll are already incorporated into the simulation and thus we
add (∂f /∂t)force by updating cl of each simulated molecule to cl + Δtg and keep-
ing xl , Fl , Nl unchanged because of f (cl + Δtg, xl , t + Δt)=f (cl , xl , t) if neglecting
the molecular motions and intermolecular collisions to consider (∂f /∂t)force alone.
Correspondingly, utr,k of each cell is changed to utr,k + Δtg but ntr,k , Ttr,k keep
unchanged. When sampling the flow velocity at each cell, we use its average value
before and after implementing the external body force, i.e., sampling utr,k + 0.5Δtg.
3.2.4 Boundary Conditions
Different types of boundary condition have been discussed in Sect. 2.10 for deter-
ministic solvers of the Boltzmann-like equation as well as Monte Carlo molecular
simulation methods (e.g., the DSMC method). In this section, we focus on the dif-
ferences of boundary condition between the DSMC and DSBGK simulations.
For open boundary, simulated molecules are removed from the computational
domain when moving across open boundary during each Δt. Correspondingly, new
simulated molecules are generated at the end of each Δt at the open boundary with xl
and cl being selected randomly as in the DSMC simulations. Then, Fl is determined
eq
from xl , cl through fk by using the macroscopic quantities specified at the open
boundary or the transient values of adjacent cell (without extrapolation for simplic-
ity). The initial values of Nl of new simulated molecules at different open boundaries
can be different in the DSBGK simulations. For example, in the channel flow prob-
lem driven by the density difference ninlet − noutlet between the two ends, we can
use different initial values for Nlinlet and Nloutlet such that Nlinlet /ninlet = Nloutlet /noutlet ,
which makes the number of simulated molecules per cell almost the same for dif-
ferent cells of the same volume even if the number density of real molecules is not
uniform. Since the stochastic noise at each cell depends on the average number of
simulated molecules inside that cell, such selection of the initial Nl for new sim-
ulated molecules at different open boundaries achieves the trade-off of stochastic
noise among cells and thus reduces the sample size required to smooth the results in
the whole computational domain.
For molecular reflection at xl on the wall surface, cl is changed to cnew l = cr +
uwall , where the reflected velocity cr is selected in the local Cartesian coordinate
system Slocal moving at the wall velocity uwall according to the reflection model as in
the DSMC simulations, which is detailed in Sect. 2.10. The subscript l for the index
of simulated molecule is omitted for notation clarity of the incoming and reflected
velocities. Nl remains unchanged to conserve the mass. Then, Fl is updated to the
representative value of f at the point (cnew l , xl , t), i.e., Fl
new
= f (cnew
l , xl , t) = fB (cr ),
where the determination of fB (cr ) for different boundary conditions has been given
in Sect. 2.10.
In the specular reflection model, we have fB,specular (cr ) = fB (ci ), which implies
Flnew = Fl since Fl = f (cl , xl , t) and fB (ci ) = f (cl , xl , t) before reflection and
Flnew = fB,specular (cr ) after reflection.
In the complete diffuse reflection model (with n+ as the outward normal direction
of surface toward the inside of flow domain), the distribution fB (ci )|ci ·n+ <0 of the
incoming molecules is known from the molecular information in the adjacent cell,
and the distribution fB,diffuse (cr )|cr ·n+ >0 of reflected molecules is a Maxwell distri-
bution as given in Eq. (2.90), where Twall is the wall temperature at the reflection
position and neff needs to be determined according to the mass conservation law of
the reflection process. Theoretically, fB (ci ) depends on the incoming molecules. In
real applications, a simple boundary condition is proposed to further reduce stochas-
tic noise if the perturbation (Ma number in general) is small. That is, we use cell
variables rather than molecular variables to determine fB (ci ) as a local equilibrium
distribution:
3/2
m −m(ci − (utr,k − uwall ))2
fB,simple (ci ) = ntr,k exp , (3.16)
2π kB Ttr,k 2kB Ttr,k
where ntr,k , utr,k , Ttr,k are the quantities of cell k adjacent to the reflection point xl .
Then, the number flux of incoming real molecules is, i.e., Eq. (2.77):

Nin,simple = − fB,simple (ci )(ci · n+ )dci
ci ·n+ <0
(3.17)
kB Ttr,k √
= ntr,k exp(−ûin ) + π ûin (1 + erf(ûin )) ,
2
2π m

where ûin = −(utr,k − uwall ) · n+ / 2kB Ttr,k /m. Similarly,
√ Nout of reflected real
molecules is given by Eq. (2.91), i.e., Nout = neff kB Twall /(2π m). Let Nout =
Nin,simple according to the mass conservation and we obtain an estimate for neff :

Ttr,k √
neff,simple = ntr,k [exp(−ûin
2
) + π ûin (1 + erf(ûin ))]. (3.18)
Twall
Then, Fl can be updated to Flnew = fB,diffuse (cr ), where neff = neff,simple . As expected,
the estimation of incoming number flux by Nin,simple based on a local equilibrium
distribution becomes inaccurate at high Kn. But, the incurred errors in the solutions
of n, u and T , for example, are small even at high Kn = 8 of the lid-driven cavity
flow and occur mostly in the region with relatively small perturbations [18], which
will be detailed in Sect. 3.4.2.
For closed flow problems, density drift has been observed after very long sim-
ulation time (much longer than the period used by the convergence process) when
using this simple implementation of the boundary condition [25]. Usually, after a cer-
tain number of time steps, the transient distributions of u and T are converged except
small stochastic fluctuations but the transient distribution of n starts to have very slow
drift from the converged solution. The magnitude of the density drift can be reduced
by using a large number of simulated molecules per cell. But, the DSBGK accuracy
is almost unchanged when using only 10 simulated molecules per cell and more
time-averaging samples. Correspondingly, the sampling process of density could be
very short (e.g., using 100 time steps for 100 samples) to avoid deviation due to
the slow density drift, because the transient density distribution has low stochastic
noise. Note that this unphysical density drift is partially due to unideal sequence of
random fractions generated by a deterministic algorithm and thus can also be alle-
viated when slightly changing the sequence of random fractions Rf (e.g., skipping
the first few Rf ). Fortunately, it automatically disappears when simulating open flow
problems (e.g., channel flows [25]) because fixed number densities are applied at
the open boundaries. By contrast, flow velocity and temperature are not subject to
unphysical drift even in closed problems thanks to the specified constraints at the
boundary. Additionally, the density drift in closed problems becomes unnoticeable
if the perturbation is very small, e.g., u1,wall = 10−6 m/s in the lid-driven cavity
flows [26].
If the perturbation is not small, the density drift of closed problems can be elimi-
nated by using a statistically accurate boundary condition [26], in which the incoming
number flux is directly calculated by using the information of incoming simulated

molecules, although the correspondingly computed incoming flux is noisy. As in
the DSMC method, it is convenient for the DSBGK method to calculate the net
flux (Q) of any molecular quantity Q(c) in unit time and across unit area of the
boundary surface:
(Q) = Nl [Q(ci ) − Q(cr )]l , (3.19)

ΔtΔA ∈Δt,ΔA
where the summation is over the simulated molecules reflected on the sub-area ΔA
during the current time step Δt, Q(ci ) and Q(cr ) are the incoming and reflected
quantities, respectively. If Q = mc or mc2 /2, then (Q) represents the stress or heat
flux, respectively. Similarly, the incoming number flux is computed as:
Nin = Nl . (3.20)
ΔtΔA ∈Δt,ΔA
Let Nout = Nin again, we obtain a statistically accurate formula for neff :

2π m 1
neff = Nl , (3.21)
kB Twall ΔtΔA ∈Δt,ΔA

where ∈Δt,ΔA Nl usually contains large stochastic noise. Then, Fl is updated to
Flnew = fB,diffuse (cr )
during the simulation process and neff is updated after each Δt.
As we can see, ∈Δt,ΔA Nl of Eq. (3.21) usually contains large stochastic noise
and makes neff much more noisy than neff,simple . Thus, the previous simple diffuse
boundary condition of using neff,simple is still an advisable choice in closed problems
with small perturbation as well as in all open problems, unless the density drift occurs
and fails the goal of the simulations making the use of the statistically accurate neff
necessary. Additionally, the boundary condition of using noisy neff in problems of
small perturbation might lead to wrong results due to error accumulation, e.g., the
error accumulated till the previous Δt will make the error of using noisy neff at the
current Δt worse. This accumulation of stochastic error is similar to that of using few
(e.g., 5 instead of the recommended minimum number of 20) simulated molecules
per cell in the DSMC method and cannot be alleviated by time/ensemble-averaging
process.
3.2.5 Summary of the DSBGK Algorithm
The workflow of an ordinary DSBGK simulation can be summarized as follows:

1. In the initialization process, generate the domain cells and simulated molecules
and assign them initial values for ntr,k , utr,k , Ttr,k and xl , cl , Fl , Nl , respectively.
2. Each simulated molecule l moves on a uniform trajectory until encountering

boundaries. During each Δt, the trajectory of each simulated molecule l may be
divided into several segments by the cell interfaces. Then, xl , Fl , Nl are deter-
ministically updated for each segment in sequence along the trajectory. When
encountering wall boundaries, cl is updated to cr + uwall according to the reflec-
tion model, and Fl is correspondingly updated to fB (cr ). In open problems, sim-
ulated molecules are removed from the computational domain when they move
across open boundaries during each Δt. At the end of each Δt, new simulated
molecules are generated at the open boundaries and their molecular variables are
determined from the local Maxwell distribution, similar to the above initialization
process. The variables ntr,k , utr,k , Ttr,k of each cell k are updated at the end of each
Δt according to the conservation laws.
3. After convergence, ntr,k , utr,k , Ttr,k provide the discrete solutions of the BGK
equation at steady state.
3.2.6 Extension to Gas Mixtures
The DSBGK method can be extended [27] to simulate gas mixtures without
intramolecular energy mode and chemical reaction based on a consistent BGK-type
model [2]. This extension involves very few modifications to the original DSBGK
algorithm and other similar extensions are also possible by using different BGK-type
equations.
As in the original BGK equation, the macroscopic quantities of each component
σ ∈ [1, Ncomp ]7 are defined by using the distribution function fσ (c, x, t):

nσ = fσ dc,
R3

1
uσ = cfσ dc, (3.22)
nσ R3

mσ
Tσ = (c − uσ )2 fσ dc.
3kB nσ R3
Total number density n, overall flow velocity u and temperature T of the mixture
can be defined by using nσ , uσ , Tσ and molecular mass mσ of all components. In the
absence of external body force, the evolution of fσ is as follows [2]:
Dfσ ∂fσ ∂fσ

= +c· = υσ (fσeq − fσ ), (3.23)
Dt ∂t ∂x
7 By a slight abuse of notation, σ denotes the collision cross-section per unit solid angle introduced
in Sect. 2.3, and the component index introduced here and in Sect. 4.3.
Ncomp
in which the total collision frequency is υσ = σ =1 (υσ σ nσ )
8
and
3/2 eq
mσ −mσ (c − uσ )2
fσeq (nσ , ueq
σ , Tσ ) = nσ
eq
eq exp eq , (3.24)
2π kB Tσ 2kB Tσ
eq eq eq eq
where fσ essentially is a function of (c, x, t) although notation fσ (nσ , uσ , Tσ ) is
eq
adopted for the convenience of following discussion, and the auxiliary variables uσ ,
eq eq
σ = (3kB Tσ )/2 are defined as follows :9
Ncomp
mσ υσ ueq
σ = mσ υσ uσ + [2mr χσ σ nσ (uσ − uσ )] (3.25)
σ =1
and
mσ υσ eq
υσ σeq =υσ σ − (uσ − uσ )2
2
Ncomp

4mr χσ σ nσ mσ (uσ − uσ )2
(3.26)
+ σ − σ +
,
σ =1
mσ + mσ 2
where mr = mσ mσ /(mσ + mσ ) is the reduced mass and χσ σ 10 is the interaction

coefficient between components σ and σ . The coefficients υσ σ and χσ σ are defined
by using the molecular interaction potential [2].
During each Δt, the mass increment ΔMk,σ of component σ in the cell k of volume
Vk due to intermolecular collisions with all components is:

ΔMk,σ = ΔtVk mσ υσ (fσeq − fσ )dc
R3
(3.27)
= ΔtVk mσ υσ (nσ − nσ )
≡ 0,
which is consistent with the mass conservation.

During each Δt, the momentum increment ΔPk,σ of component σ in the cell k
due to intermolecular collisions with all components is:
8 As in [2], we use the coefficient υ in Sect. 3.2.6 instead of 1/τ used elsewhere.
9 By a slight abuse of notation, ijk denotes the Levi-Civita symbol introduced in Sect. 1.4 and
denotes average thermal energy per molecule introduced here.
10 By a slight abuse of notation, χ denotes the deflection angle of intermolecular collision introduced
in Sect. 2.2 and χσ σ denotes the interaction coefficient between components σ and σ introduced
here.

ΔPk,σ = ΔtVk (mσ c)υσ (fσeq − fσ )dc
R3
= ΔtVk υσ nσ mσ (ueq
σ − uσ ) (3.28)

Ncomp
= ΔtVk [2nσ nσ mr χσ σ (uσ − uσ )],
σ =1
where Eq. (3.25) is substituted. ΔPk,σ could be nonzero due to momentum exchange
among different components via intermolecular collisions but the overall momentum
Ncomp
conservation is satisfied as required, i.e., σ =1 ΔPk,σ ≡ 0.
During each Δt, the energy increment ΔEk,σ of component σ in the cell k due to
intermolecular collisions with all components is:

mσ c2
ΔEk,σ =ΔtVk υσ (fσeq − fσ )dc
R 3 2
mσ eq 2 mσ 2
=ΔtVk υσ nσ σeq + (uσ ) − σ − u
2 2 σ
(3.29)

2nσ nσ mr χσ σ
Ncomp
=ΔtVk ×
σ =1
mσ + mσ
[2σ − 2σ + (uσ − uσ ) · (mσ uσ + mσ uσ )],
where Eqs. (3.25) and (3.26) are substituted. ΔEk,σ could be nonzero due to energy
exchange between components via intermolecular collisions but the overall energy
Ncomp
conservation is satisfied as required, i.e., σ =1 ΔEk,σ ≡ 0.
In the DSBGK simulations of gas mixtures, each molecule l with a fixed com-
ponent index σl ∈ [1, Ncomp ] has four variables: xl , cl , Nl and Fl = fσ =σl (cl , xl , t) as
in the original algorithm. The magnitude of initial Nl of component σl = σ could
be proportional to the initial number density nσ,0 such that the average numbers of
simulated molecules per cell are almost equal for different components. Compared to
the DSMC method with a constant Nl = N for all simulated molecules, the DSBGK
method can significantly reduce the total number of simulated molecules when the
contrast of mole fractions of different components is high.
Each cell k has three original variables ntr,k,σ , utr,k,σ , Ttr,k,σ and two additional aux-
eq eq
iliary variables utr,k,σ , Ttr,k,σ for each component σ . At the initial state with given spa-
eq
tial distributions of nσ,0 , uσ,0 and Tσ,0 , we have ntr,k,σ = nσ,0 , utr,k,σ = utr,k,σ = uσ,0
eq
and Ttr,k,σ = Ttr,k,σ = Tσ,0 . The initial values of molecular variables of each compo-
eq
nent σ are selected according to the initial distribution fσ,0 = fσ (nσ,0 , uσ,0 , Tσ,0 ).
eq eq eq
During each Δt, Eq. (3.23) with fσ (ntr,k,σ , utr,k,σ , Ttr,k,σ ) is used to update the
molecular variables as in the original DSBGK algorithm. The discrepancy of the
numerical mass increment from the theoretical mass increment ΔMk,σ ≡ 0 of com-
ponent σ in the cell k due to intermolecular collisions with all components is:

ΔMk,σ
err
= mσ Δk Nl − ΔMk,σ = mσ Δk Nl ,
∈k ∈k
(3.30)
σl =σ σl =σ
where Δk Nl is the number increment of real molecules of class cl of component

σl due to intermolecular collisions with all components inside the cell k during the
current time step. The discrepancy of the numerical momentum increment from
the theoretical momentum increment ΔPk,σ of component σ in the cell k due to

ΔPk,σ
err
= mσ (Δk Nl cl ) − ΔPk,σ ,
∈k
(3.31)
σl =σ
where ΔPk,σ is computed by Eq. (3.28) using ntr,k,σ , ntr,k,σ , utr,k,σ , utr,k,σ in place
of nσ , nσ , uσ , uσ , respectively. The discrepancy of the numerical energy increment
from the theoretical energy increment ΔEk,σ of component σ in the cell k due to
mσ
ΔEk,σ
err
= (Δk Nl c2l ) − ΔEk,σ ,
2 ∈k (3.32)
σl =σ
where ΔEk,σ is computed by Eq. (3.29) using ntr,k,σ , ntr,k,σ , utr,k,σ , utr,k,σ ,
3kB Ttr,k,σ /2, 3kB Ttr,k,σ /2 in place of nσ , nσ , uσ , uσ , σ , σ , respectively.
The above numerical discrepancies are used to update the cell variables ntr,k,σ ,
utr,k,σ and Ttr,k,σ at the end of each Δt based on auto-regulation scheme:
ntr,k,σ Vk − ΔMk,σ
err
/mσ
tr,k,σ =
nnew ,
Vk
ntr,k,σ Vk utr,k,σ − ΔPk,σ
err
/mσ
tr,k,σ =
unew , (3.33)
nnew
tr,k,σ Vk
ntr,k,σ Vk (3kB Ttr,k,σ /2 + mσ u2tr,k,σ /2) − ΔEk,σ
err
− nnew
tr,k,σ Vk mσ (utr,k,σ ) /2
new 2
new
Ttr,k,σ = .
tr,k,σ Vk (3kB /2)
nnew
eq eq
Then, the cell auxiliary variables utr,k,σ and Ttr,k,σ can be updated by Eqs. (3.25) and
(3.26), where the updated discrete variables nnewtr,k,σ , utr,k,σ , Ttr,k,σ are used to replace
new new
nσ , uσ , Tσ , respectively.
3.3 Discrete Velocity Method
The discrete velocity method (DVM) deterministically solves the Boltzmann equa-
tion or its simplified models [4, 21, 33, 39]. It replaces the kinetic equation by a
set of equation of each discrete velocity and computes the collision integral by a
double summation over the discrete velocity grids. Although the DVM offers accu-
rate noise-free solutions, it generally requires a large number of grids to discretize
the distribution function in the high-dimensional phase space, which may lead to a
high demand in computational memory and time. For a given error tolerance, the
velocity grid number can be reduced by using a layout in the polar coordinate for
two-dimensional (2D) cases or in the spherical coordinate for three-dimensional (3D)
cases compared to the layout in the Cartesian coordinate [18]. This non-Cartesian
discretization of molecular velocity space is generally valid for arbitrary spatial
geometry since the projection from the non-Cartesian coordinate to the Cartesian
coordinate will be employed to determine the velocity components for the governing
equation of the Cartesian coordinate.
Taking the 2D lid-driven cavity simulation as an example, at each spatial grid, the
DVM needs at least 4 × 24, 4 × 40, 4 × 48 velocity points distributed in the polar
coordinate for Kn = 0.1, 1, 8, respectively, if we accept a maximum local relative
error (i.e., deviation from the accurate DVM results obtained by using a very fine
velocity grid of 8 × 80 points) of 10% in the u1 , u2 , n, T profiles along the horizontal
centerline parallel to the driving direction [18]. Note that this maximum discrepancy
occurs mostly in the region with relatively small perturbations, where small numeri-
cal error leads to relatively large discrepancy. In the 3D lid-driven cavity simulations,
the computational cost of DVM will significantly increase due to the presence of the
third molecular velocity component, e.g., the number of velocity points will increase
by 12 times due to the additional discretization of inclination angle of the spherical
coordinate compared to the 2D cases at arbitrary Kn [18]. By contrast, particle-
based methods (such as the DSBGK and DSMC methods) only need a number (e.g.,
10 ∼ 20) of simulated molecules per cell to discretize the molecular velocity space
at arbitrary Kn for both 2D and 3D problems. Although the number of simulated
molecules per cell used by particle-based methods is small, the molecular velocity
points inside each cell (i.e., at each spatial grid of discretization) are dynamically
updated via the frequent and random events of simulated molecules moving into
and out of each cell (from the perspective of Eulerian description), and the molec-
ular velocities along all representative trajectories are also dynamically updated via
the frequent and random molecular reflections at the boundary or intermolecular
collisions (from the perspective of Lagrangian description). Thus, the dynamic dis-
cretization of using few simulated molecules per cell in particle-based methods can
sample from the whole velocity space and therefore allow as fine discretization of
the unbounded molecular velocity space as desired with the increase of simulation
time.
If only the steady state solution is of interest, DVM can accelerate its convergence
rate by using implicit time-marching schemes or other iterative schemes [39]. Due
to the time-evolutionary nature, particle-based methods usually have no such accel-
eration opportunities without losing accuracy and the error of transport coefficients
in the DSMC method has been found to be proportional to Δt 2 [10, 13]. On the other
hand, DVM usually uses the same distance of spatial grid in different directions for
algorithm simplicity although it can be extended to non-uniform grids [17]. Particle-
based methods are unchanged when using different cell sizes in different directions
3.3 Discrete Velocity Method 89
to reduce the cell number for high aspect ratio as long as the maximum cell size is
smaller than the mean free path [25, 28] as shown in Sect. 3.4.3.
Taking 2D single-component problems as example [18], two reduced velocity
distribution functions are introduced to cast the 3D molecular velocity space into a
2D expression [39]11 :

g= f (c, x, t)dc3 ,
R
(3.34)
h= (c3 )2 f (c, x, t)dc3 .
R
For convenience, we denote c = (c1 , c2 ), C = (C1 , C2 ) and x = (x1 , x2 ) in this

By using g and h, macroscopic variablescan be computed as n = R2 gdc,
section.
nu = R2 cgdc, and n(u2 + 3kB T /m)/2 = (1/2) R2 (c2 g + h)dc. Taking the BGK
equation without external body force as example, the governing equations for the
two reduced distribution functions can be deduced in the form of generic function
φ = (h, g)12 :
∂φ ∂φ φ eq − φ
+c· = , (3.35)
∂t ∂x τ
where the reduced equilibrium distribution functions φ eq = (heq , g eq ) are:

nm −mC2
g (c, x, t) =
eq
f (c, c3 , x, t)dc3 =
eq
exp ,
2π kB T 2kB T
R (3.36)
kB T g eq
heq (c, x, t) = (c3 )2 f eq (c, c3 , x, t)dc3 = .
R m
Then, the continuous molecular velocity space c is projected into a set of fixed
Nc discrete velocities c(i) (i = 1, 2, .., Nc ). As a result, for the BGK model, the
governing equation (3.35) is replaced by a system of Nc independent equations.
Here, we discretize this system in time by a fully time-implicit Godunov-type scheme
[33, 39]:

1 1
+ c · ∇ + (ts) Δφ (ts) = RHS(ts) ,
(i)
Δt (ts) τ
(3.37)
1
RHS(ts) = (ts) φ eq,(ts) − φ (ts) − c(i) · ∇φ (ts) ,
τ
11 By a slight abuse of notation, g denotes the external body force per unit mass introduced in
Sect. 1.3.1, g denotes a reduced velocity distribution function of DVM introduced here, and subscript
g denotes the solubility ρg introduced in Sect. 4.5.4.
12 By a slight abuse of notation, φ denotes the dissipated energy introduced in Sect. 1.3.4 and the
generic function of DVM introduced here, and the porosity introduced in Sect. 3.4.5.
where Δφ (ts) = φ (ts+1) − φ (ts) needs to be determined at each time step ts. RHS(ts) is
the explicit part, and the spatial derivative is approximated by a third-order upwind
scheme. For instance, the derivative with respect to the x1 -direction at the jth point
is evaluated by:
(ts)
∂φ (ts) (2φj+1 + 3φj(ts) − 6φj−1
(ts) (ts)
+ φj−2 )/(6Δx1 ), c1(i) > 0 ,
= (3.38)
∂x1 j (ts)
(−2φj−1 − 3φj(ts) + 6φj+1
(ts) (ts)
− φj+2 )/(6Δx1 ), c1(i) < 0 .
The left-hand side of Eq. (3.37) is the implicit part, and the spatial derivative is
approximated by a first-order upwind scheme. By marching in the appropriate direc-
tion, e.g., increasing x1 in the case of c1(i) > 0, the unknown Δφ (ts) can be obtained
directly without solving a system of equations.
Note that Δt in Eq. (3.37) is a pseudo time step that is defined by the Courant–
Friedrichs–Lewy (CFL) condition, i.e., Δt = ηΔx1min /c1max , where η is the CFL num-
ber, Δx1min is the minimum spatial grid size, and c1max is the maximum discrete speed.
While η can be smaller than 1 to capture the transient behaviour, it can also be set as
large as 104 to only obtain the steady-state solution [18].
3.4 Simulation Results
3.4.1 Couette Flows
As discussed in Sect. 1.6, the ordinary constitutive equations of the continuum

description of flow problems become inaccurate at high Kn due to the non-equilibrium
phenomenon, which can be easily understood in the Couette flow problem.
The schematic is given in Fig. 3.2a, where the left and right plates move at
the same speed of 10 m/s but in opposite directions parallel to the x2 axis. With-
out loss of generality, the gas media is argon with the dynamic viscosity μ =
2.117 × 10−5 × (T /273)0.81 Pa · s and the molecular mass m = 6.63 × 10−26 kg.
The wall temperature is maintained at the initial temperature, i.e., Twall = T0 = 273
K. The initial gas number density is fixed at n0 = 2.6847 × 1025 m−3 and corre-
spondingly the initial mean free path is λ0 ≈ 63 nm. The distance L1 between the
two plates is adjusted for different Kn = λ0 /L1 . The cell number of discretizing L1
is 200, 20 and 20 for Kn = 0.01, 0.1 and 1, respectively, and each cell contains
about 2550 simulated molecules on average. The boundary condition of complete
Maxwell diffuse reflection is used and implemented in the DSBGK simulations by
using neff,simple of Eq. (3.18). The BGK relaxation time τ is selected here to match
the dynamic viscosity μ(T ), i.e., τ = μ/(nkB T ) where μ could also change with
local pressure if needed, as detailed in Sect. 2.9.
Figure 3.2b shows that the transient velocity distribution of the DSBGK simu-
lations is smooth, which is a big potential advantage over the DSMC method and
DVM that require ensemble-averaging process and iteration process, respectively,
3.4 Simulation Results 91
(a) (b)
(c) (d)
Fig. 3.2 Comparison between the DSBGK and DSMC methods in Couette flows
for each transient solution. Although the evolution of DSBGK velocity with time is
reasonable, it is subject to hysteresis effect due to using the auto-regulation schemes
of Eq. (3.10), which correct the current discrepancy at the next time step. This hys-
teresis effect can be mitigated by using small time step. Figure 3.2c shows very good
agreement of the steady state velocity distribution between the DSBGK and DSMC
methods at different Kn. Additionally, it also shows that the slippage effect (i.e., the
difference between the plate velocity and the gas velocity near the plates) is almost
negligible at Kn = 0.01 and becomes noticeable as Kn increases. In Fig. 3.2d, the
calculation of shear stress P12 exerted by the gas on the plate at x1 = 0 in the DSBGK
simulations is computed by using noisy molecular variables in Eq. (3.19) and thus the
time-averaging process is used to obtain smooth P12 that is normalized at Kn = ∞.
The agreement of P12 between the DSBGK and DSMC methods is also very good at
different Kn. Additionally, the steady state distribution of u2 (x1 ) in Fig. 3.2c becomes
nonlinear at Kn = 1, which implies the inaccuracy of the ordinary constitutive equa-
tion that has a linear relation between the stress tensor and the strain rate tensor.
As expected, the flow velocity u2 will become zero inside the whole computational
domain at the limit of Kn → ∞ because the two average (c2 )c1 <0 and (c2 )c1 >0 (i.e.,
10 and –10 m/s at the plates according to the complete Maxwell diffuse reflection)
are unchanged when molecules move away from the plates, and thus neutralize each
other when computing the overall average molecular velocity (i.e., c2 = u2 ). On the
other hand, P12 will not disappear but reach to its maximum at Kn → ∞, which
clearly indicates the breakdown of the ordinary constitutive equation.
3.4.2 Lid-Driven Cavity Flows
Low-speed gas flows at different Kn in lid-driven cavity are simulated to assess the
accuracy, efficiency and memory usage of the DVM and DSBGK method in solving
the BGK model equation. This problem is characterized by shear-driven and flow
compression/expansion phenomena and challenging due to the velocity discontinuity
at the boundaries. The DSMC results are used as the validation reference at moderate
speed and at low speeds the reference solutions are obtained by using fine-grid DVM.
Argon gas with the dynamic viscosity μ = 2.117 × 10−5 × (T /273)0.81 Pa · s and
the molecular mass m = 6.63 × 10−26 kg is used again. As shown in Fig. 3.3, the
cavity sizes in all directions are equal to L = 1 micron and x̂i∈[1,2,3] = xi∈[1,2,3] /L.
The boundary at x̂2 = 1 moves in the positive x̂1 -direction at a constant speed u1,wall .
Note that the DSBGK simulations use dimensional quantities while the DVM sim-
ulations employ only dimensionless quantities scaled by a relevant reference value,
e.g., L, n0 , u1,wall , T0 . Perturbed macroscopic quantities obtained by the two meth-
ods will be used in comparison: n̂ = (n − n0 )/n0 , û1 = u1 /u1,wall , û2 = u2 /u1,wall ,
T̂ = (T − T0 )/T0 . The definition of λ0 used in the original work [18] is different
from Eq. (2.76)but gives almost the same Kn = λ0 /L. The Mach number is defined
as Ma = u1,wall m/(γ kB T0 ), where γ is the specific heat ratio. The wall temperature
is maintained at the initial temperature, i.e., Twall = T0 = 273 K. The BGK relax-
ation time τ is selected here to match the dynamic viscosity μ, i.e., τ = μ/(nkB T ),
as detailed in Sect. 2.9.
(a) (b)
Fig. 3.3 Schematic of 2D lid-driven cavity (a) and 3D lid-driven cavity (b) [18]
In the 2D simulations, we choose the lid speeds to be u1,wall = 0.001, 1, 10 and

50 m/s, which correspond to Ma = 3.2 × 10−6 , 3.2 × 10−3 , 3.2 × 10−2 and 0.16,
respectively. The initial uniform number density n0 is adjusted to obtain different
Kn = 0.1, 1 and 8 to cover the slip, transitional and free-molecular flow regimes. The
boundary condition of complete Maxwell diffuse reflection is used and implemented
in the DSBGK simulations by using neff,simple of Eq. (3.18) at Ma < 0.16 or using
neff of Eq. (3.21) at Ma = 0.16. The number of uniform spatial cell/grid is 602
for both the DVM and DSBGK simulations. An 8 × 80 velocity grid of the polar
coordinate is used in the DVM for the fine-grid simulations. Correspondingly, the
DSBGK simulations use 2000 simulated molecules per cell and 500 time-averaging
samples at arbitrary
√ Kn and Ma. The time step in the DSBGK simulations is fixed at
Δt = 2.0Δx1 m/(2kB T0 ) for different Kn to make it easy to understand the relation
between the time step number and the corresponding CPU time.
Figure 3.4 shows the comparison between the DVM, DSBGK and DSMC profiles
of the perturbed û1 and û2 along the vertical centreline of x̂1 = 0.5 and the horizontal
centreline of x̂2 = 0.5, respectively, for various Ma and Kn. The DSMC results are
obtained by using the variable hard-sphere (VHS) molecular model [19]. As we can
see, the slip velocity increases considerably with Kn at the moving top boundary
(i.e., deviation from û1 = 1 due to slip), while its increase is negligible at the static
bottom boundary (i.e., deviation from û1 = 0 due to slip). At moderate Ma = 0.16,
the DVM and DSBGK method agree very well with the DSMC method for various
Kn. For Ma < 0.16, the DVM still agrees very well with the DSMC method but the
DSBGK method has small deviation due to using the simple boundary condition of
Eq. (3.18). Compared to the DVM in solving the same BGK equation, the maximum
local discrepancy of the DSBGK method is about 7.5% for û1 at Ma = 3.2 × 10−6 ,
Kn = 1, about 2.5% for û2 at Ma = 3.2 × 10−6 , Kn = 0.1; and about 4.7% for n̂,
2% for T̂ along the horizontal centerline at Ma = 3.2 × 10−6 if fixing Kn = 1. Note
that this maximum discrepancy occurs mostly in the region with relatively small
perturbations, where small numerical error leads to relatively large discrepancy.
So far, we have focused on the accuracy of the DVM and DSBGK method at their
best, i.e., by using high-resolution velocity grid in the DVM and a large number of
simulated molecules per cell for the implementation of accurate boundary condition
in the DSBGK method. For practical applications, it is important to strike a balance
between computational accuracy and efficiency. In the test of coarse-grid simulations
for Ma = 3.2 × 10−3 , the DVM velocity point number decreases to 4 × 24, 4 × 40
and 4 × 48 at Kn = 0.1, 1 and 8, respectively, and the DSBGK method uses only
10 simulated molecules per cell with the simple boundary condition and 5000 time-
averaging samples at arbitrary Kn. The maximum local relative errors of the coarse-
grid DVM and DSBGK results from the fine-grid DVM results are within 10%
in the û1 , û2 , n̂, T̂ profiles along the horizontal centreline. Figures 3.5, 3.6 and 3.7
show the contours of macroscopic quantities obtained by the coarse-grid DVM and
DSBGK simulations, alongside the reference contours obtained by the fine-grid DVM
simulations. The fluctuation of DSBGK results (particularly in the T̂ profiles) is due
to stochastic noise. The fluctuation of coarse-grid DVM results around the reference
solution is due to the ray effects, which exist in flow problems with discontinuous
(a) at Kn= 0.1 (b) at Kn= 0.1
(c) at Kn= 1 (d) at Kn= 1
(e) at Kn= 8 (f) at Kn= 8
Fig. 3.4 Profiles of the perturbed û1 (left column) and û2 (right column) along the vertical centreline
of x̂1 = 0.5 and horizontal centreline of x̂2 = 0.5, respectively, of the 2D lid-driven cavity [18]
Fig. 3.5 Contours of the perturbed û1 , û2 , n̂ and T̂ in the 2D lid-driven cavity flow obtained by
the DVM using an 8 × 80 velocity grid (black solid lines), the DVM using a 4 × 24 velocity grid
(blue dashed lines), and the DSBGK method using 10 simulated molecules per cell with the simple
boundary condition and 5000 samples (red dash-dot lines); Ma = 3.2 × 10−3 , Kn = 0.1 [18]
boundary and are caused by incompatibility of the velocity grid, spatial grid and
order of accuracy of the numerical scheme.
The computational cost can be further reduced by using coarse spatial grid/cell,
as long as the grid/cell size is smaller than the mean free path (as in the DSMC sim-
ulations) and the geometry is graphically preserved in discretization. We therefore
maintain high resolution in the velocity space and focus on the influence of spa-
tial grid/cell number on the accuracy for the case with Ma = 3.2 × 10−2 , Kn = 1.
Figure 3.8 shows the comparison of the DVM and DSBGK results obtained using
102 spatial grid/cell with the reference DVM results obtained using 602 spatial grid.
It can be seen that the n̂, T̂ contours given by both methods on the coarse spatial
grid/cell are in satisfactory agreement with the reference DVM results. However,
the û1 , û2 contours of the coarse solutions of both methods have noticeable devia-
tions from the reference DVM results. This numerical error is expected to occur also
with other simulation methods when using coarse spatial grids to save computational
boundary condition and 5000 samples (red dash-dot lines); Ma = 3.2 × 10−3 , Kn = 1 [18]
cost. Balancing computational accuracy with efficiency becomes a key issue when
simulating large-scale problems, e.g. gas flows in porous media.
In the 3D simulations of cavity flow, the initial uniform number density n0 is
adjusted to obtain different Kn = 0.1, 1 and 8. But, Ma is fixed at 3.2 × 10−3 by
using u1,wall = 1 m/s and thus neff,simple of Eq. (3.18) is used in the DSBGK boundary.
The number of uniform spatial cell/grid becomes 603 for both the DVM and DSBGK
simulations. We use the fine-grid DVM results as reference data at low speed, which
are obtained by using 4 × 80 × 40 velocity grid of the spherical coordinate. In the
coarse-grid simulations, the coarse DVM grids used in the 2D case are now extended
by using additional 12 points to discretize the inclination angle, i.e., 4 × 24 × 12,
4 × 40 × 12 and 4 × 48 × 12 for Kn = 0.1, 1 and 8, respectively, while the DSBGK
method still uses about 10 simulated molecules per cell and 5000 time-averaging
samples at arbitrary Kn as in the 2D case. The deviations of both methods from the
fine-grid DVM reference solution on the plane x̂3 = 0.5 are similar to the 2D case.
boundary condition and 5000 samples (red dash-dot lines); Ma = 3.2 × 10−3 , Kn = 8 [18]
As an example, Fig. 3.9 shows the û1 , û2 , n̂, T̂ contours on the planes x̂2 = 0.5 and
x̂3 = 0.5 for the case of Kn = 8. Additionally, by comparing the solutions of û1 , û2 ,
n̂, T̂ on the plane x̂3 = 0.5 of the 3D case with those of the 2D case (not illustrated
here), the side wall (x̂3 = 0, 1) effects on the middle plane are seen to be negligible
at Kn = 0.1 but increase with Kn, and change the T̂ profiles most significantly.
Overall, the DSBGK and DVM results are in satisfactory agreement for all exam-
ined 2D and 3D cases. When the accurate boundary condition and fine velocity grid
are used by them, respectively, they are also in good agreement with the DSMC
results, which implies that the BGK model equation solved by both the DVM and
DSBGK method is accurate at low speed although the intermolecular collision term
is significantly simplified compared to the original integral term of the Boltzmann
equation. As a statistical method, the stochastic noise of the DSBGK method is much
smaller than that of the traditional DSMC method, and is independent of Ma (i.e.,
Fig. 3.8 Contours of the perturbed û1 , û2 , n̂ and T̂ in the 2D lid-driven cavity flow obtained using
a 602 spatial grid (DVM: black solid lines) and 102 spatial grid/cells (DVM: blue dashed lines,
DSBGK: red dash-dot lines); Ma = 3.2 × 10−2 , Kn = 1 [18]
the time step number for averaging process is independent of Ma), which is in sharp
contrast to the DSMC method, where the stochastic noise is proportional to Ma−2 .
It is also important to compare the computational costs of the DVM and DSBGK
method in achieving the required coarse-grid accuracy of Figs. 3.5, 3.6 and 3.7, and
3.9. The DSBGK simulations need more CPU time than the DVM simulations for
the 2D case, i.e., 2–15 times more for the convergence process, and about 50–80
times more for the whole process when including the DSBGK time-averaging pro-
cess. However, for the 3D case, the third velocity component cannot be avoided in
the DVM and hence the CPU time ratio of DSBGK to DVM is now only 0.16–0.51
for the convergence process, and 1.6–5.8 for the whole process. For large-scale 3D
simulations, the efficiency advantage of DSBGK method is expected to be enhanced
since the CPU time used for the fixed time-averaging process (i.e. 5000 Δt, as we
used here) will become negligible compared to that for the prolonged convergence
process (e.g. increasing from 200 ∼ 2400 Δt used here to millions of Δt). Thus, the
comparison of CPU time used for the convergence process alone is also an important
(a) (DVM) (b) (DSBGK)
(c) (DVM) (d) (DSBGK)
(e) (DVM) (f) (DSBGK)
(g) (DVM) (h) (DSBGK)
Fig. 3.9 The 3D lid-driven case with Ma = 3.2 × 10−3 , Kn = 8: contours of the perturbed û1 , û2 ,
n̂ and T̂ on the planes x̂2 = 0.5 and x̂3 = 0.5, obtained by the DVM using a 4 × 48 × 12 velocity
grid (left column), and the DSBGK method using 10 simulated molecules per cell with the simple
boundary condition and 5000 samples (right column) [18]
indicator in large-scale problems, as well as in computing spatial-average quanti-

ties (e.g., mass flow rate and permeability) that are smooth at each moment and
don’t need time-averaging process at all. Additionally, the efficiency advantage of
DSBGK method will be further enhanced in problems, where different cell/grid sizes
in different directions are advisable to optimize the spatial discretization, which is
straightforward for the DSBGK method [25, 28] as shown in Sect. 3.4.3 but increases
the complexity of the DVM [17]. Note that efficiency improvements of DVM algo-
rithm have already been adopted in this comparison, e.g., large CFL number η = 104
for steady state flows and the sophisticated layouts of Gaussian quadrature veloc-
ity grid in non-Cartesian coordinates with the help of projection to the Cartesian
coordinate [18].
Additionally, the DSBGK simulations require much less memory than the DVM
simulations, because about 10 simulated molecules per cell in the DSBGK simu-
lations are sufficient for the required accuracy at arbitrary Kn in both 2D and 3D
cases, while the DVM simulations require at least 4 × 24 velocity points of the polar
coordinate and 4 × 24 × 12 velocity points of the spherical coordinate for the 2D
and 3D cases, respectively, even at low Kn = 0.1. Consequently, all tested DSBGK
serial simulations can run on an ordinary laptop while high-performance computing
(HPC) facility with large memory has been used to run the 3D DVM serial simu-
lations (about 0.25 GB used by the DSBGK method versus about 100 GB used by
the DVM). High demand of memory imposes the great limitation of DVM in real
applications.
It is worth pointing out that besides the above GDVM based on the Godunov-
type scheme of Eq. (3.37), there are other versions of DVM, including the recently
developed discrete unified gas-kinetic scheme (DUGKS) [12] that is developed from
the unified gas-kinetic scheme (UGKS) [37]. In a recent comprehensive comparison
[36], it shows that towards the continuum flow regime of small Kn, not only is the
DUGKS faster than the GDVM when using the same spatial mesh, but also requires
less spatial grids than the GDVM to achieve the same numerical accuracy; from the
slip to free molecular flow regimes of large Kn, however, the DUGKS is slower than
the GDVM, due to the complicated flux evaluation and the restrictive time step that
is smaller than the maximum effective time step of the GDVM.
3.4.3 Thermal Transpiration Through Micro-channel
It is well known that gas flows at high Kn through a tube with almost constant
pressure but variable temperature along the wall surface will experience apprecia-
ble bulk speed. The micro/nano-electro-mechanical systems (MEMS/NEMS) have
decreased to sub-microns in recent decades, where Kn could be large and the thermal
transpiration effect becomes important. Additionally, on-chip Knudsen pump can be
designed at micro-scale by using thermal transpiration effect without moving parts
and the compression ratio can be up to 50 by using 48 stages [11]. This gas flow
mechanism is also important in studying the heat transfer of electronic chips, and

thermal transpiration flows
through a micro-channel, the
wall temperature linearly
varies with the channel
length, pinlet = poutlet = p0
of interest
can be used to reduce the gas leakage back into the vacuum chamber through the
clearance of rotating vacuum pump.
Gaseous thermal transpiration flows through a real rectangular micro-channel are
simulated by the DSBGK method to compute the mass flow rates at different pressure
conditions for different gas species [28] and the simulation results are validated
against the experimental data [38]. As shown in Fig. 3.10, two tanks are added at the
channel ends to reflect the real end effects. The wall temperature Twall linearly varies
along the channel length but keeps unchanged in the two tanks, where Twall ≡ TL
or Twall ≡ TH , respectively. The gas pressures at the inlet and outlet are fixed at the
same initial pressure p0 specified for each simulation and thus the flow is driven by
the temperature variation instead of pressure difference. The gas temperatures at the
inlet and outlet are equal to the local wall temperatures, i.e., TL and TH , respectively,
and so the gas number densities at the two open boundaries are different. To reduce
the simulation cost, the cross-section area of tanks could be smaller than their real
area as long as they are much larger than that of the micro-channel. This is also
because the driving mechanism here is the temperature variation on wall surface of
the micro-channel, which is independent of the tank sizes used. By contrast, if the
flow is driven by pressure difference, the tank sizes need to be much larger than that
needed here to make the results independent of the tank sizes.
To investigate the influence of tank sizes to the solution inside the micro-channel,
we first study a small micro-channel with L1,channel × L2,channel × L3,channel = 7.3 ×
1 × 0.5 mm3 as its length, width and height. Different total sizes L1,all × L2,all × L3,all
of the whole computational domain are used to show the influence of tank sizes. Addi-
tionally, we use TL = 300 K, TH = 320 K, T0 = (TL + TH )/2 and p0 = 50 Pa. The
first gas is argon with the dynamic viscosity μ = 2.117 × 10−5 × (T /273)0.81 Pa · s
and the molecular mass m = 6.63 × 10−26 kg. At the initial state, we have μ0 =
2.346 × 10−5 Pa · s, λ0 = 0.1522 mm and √ thus Kn = λ0 /L3,channel = 0.3044. The
simulation time step is set as Δt = 0.8λ0 / 2kB T0 /m = 0.3389 × 10−6 s and the cell
sizes are Δx1 = Δx2 = Δx3 = 0.05 mm, which is smaller than λ0 as required. The
boundary condition of complete Maxwell diffuse reflection is used and implemented
in the DSBGK simulations by using neff,simple of Eq. (3.18) due to low speeds. About
20 simulated molecules per cell are used. The BGK relaxation time τ is selected
here to match the thermal conductivity coefficient ζ (μ), i.e., τ = 3μ/(2nkB T ), as
detailed in Sect. 2.9.
As shown in Figs. 3.11, 3.12, the influence of tank sizes to the solution inside
the micro-channel is investigated by comparing the results of two simulations using
L1,all × L2,all × L3,all = 10 × 5 × 5 mm3 and 20 × 10 × 5 mm3 , respectively. Over-
all, the results of T , n, u1 inside the micro-channel of the two simulations are almost
identical although noticeable discrepancy occurs outside the micro-channel. There is
noticeable discrepancy of p even inside the micro-channel because p has very small
relative variation and thus is sensitive to the numerical error associated with the finite
tank sizes. Although the comparison of u2 is impractical inside the micro-channel,
where u2 is almost zero and thus dominated by stochastic noise, the agreement of
u2 between the two simulations is good around the two ends of the micro-channel,
where the variation of u2 is appreciable for clear comparison. The overall good agree-
ment indicates that the influence of tank sizes to the solution inside the micro-channel
becomes negligible when L1,all × L2,all × L3,all ≥ 10 × 5 × 5 mm3 for this particular
case of L1,channel × L2,channel × L3,channel = 7.3 × 1 × 0.5 mm3 .
Figure 3.11 shows that the gas temperature T increases with x1 similar to Twall and n
varies in the opposite way because the pressures p = nkB T at the two open boundaries
are fixed at the same p0 . Nevertheless, p is not constant inside the micro-channel
although its relative variation is much smaller than these of T and n. Additionally,
pentry at the entry of micro-channel is lower than pexit at its exit, which provides the
pressure differences (driving mechanism, i.e., p0 − pentry and pexit − p0 ) inside the
two tanks to sustain the flow because the pressures at inlet and outlet of the whole
computational domain are equal to p0 . Although different tank sizes lead to different
pentry and pexit , their influences to the mass flow rate through the micro-channel are
small as observed in Figs. 3.11, 3.12 because they are still very close to p0 (thus Kn
is almost unchanged) and the driving mechanism of Twall (x1 ) is independent of the
tank sizes. Note that the simulation of micro-channel flow without two tanks will
lead to pentry > pexit , which will enhance the flow speed induced by Twall (x1 ) and thus
overestimate the mass flow rate [28].
Now, gas flows inside a real micro-channel with L1,channel ×L2,channel ×L3,channel =
73 × 6 × 0.22 mm3 [38] are simulated by using L1,all × L2,all × L3,all = 80 × 10 ×
1 mm3 . The temperature difference across the micro-channel is TH − TL = (347.1 −
289.2) K. A mean rarefaction parameter δ0 is used to characterize the mass flow rate
Ṁ [38]:
p0 L3,channel 0.9025
δ0 = √ ≈ . (3.39)
μ0 2kB T0 /m Kn0
Different cell sizes in different directions, i.e., Δx1 = Δx2 Δx3 , are used at
low pressure conditions to optimize the spatial discretization. The three cell sizes
are always smaller than λ0 at different pressure conditions as required, e.g., a total
of 3200 × 400 × 50 cells are used for argon gas flow at δ0 (p0 = 294 Pa) = 7.41
with λ0 ≈ 0.0268 mm (note: this simulation takes about one day for the mass flow
rate Ṁ to converge after 2000 time steps when using 40 CPU cores). In addition
to argon, we use μ = 1.865 × 10−5 × (T /273)0.66 Pa · s for helium molecules (m =
6.65 × 10−27 kg) and μ = 2.975 × 10−5 × (T /273)0.66 Pa · s for neon molecules
Fig. 3.11 Distributions of T /T0 , n/n0 , p/p0 inside and around the channel on the middle plane
of x3 = L3,all /2 computed by using L1,all × L2,all × L3,all = 20 × 10 × 5 mm3 (the black solid line
and colored background, where the vertical white bars are due to the visualization effect of different
subdomain data generated by a parallel simulation) and 10 × 5 × 5 mm3 (the white dash-dot line)
1
2
2
2
Fig. 3.12 Distributions of u1 and u2 inside and around the channel on the middle plane of x3 =
L3,all /2 computed by using L1,all × L2,all × L3,all = 20 × 10 × 5 mm3 (the black solid line and
colored background, where the vertical white bars are due to the visualization effect of different
subdomain data generated by a parallel simulation) and 10 × 5 × 5 mm3 (the white dash-dot line)
(m = 33.5 × 10−27 kg) [32]. Figure 3.13 and Table 3.1 show that the DSBGK results
of Ṁ agree very well with the experimental data over a wide rage of δ0 for different
gas species.
3.4.4 Apparent Permeability of Shale Gas
As we know, the prediction of low permeability of shale gas is very challenging. The
traditional experimental technique to measure the mass/volume flow rate at steady
Fig. 3.13 Mass flow rates

Ṁ (×10−10 kg/s) computed
by the DSBGK method and
the experimental data [38],
TH = 347.1 K and
TL = 289.2 K, cited from
[28]
Table 3.1 Mass flow rates Ṁ (×10−10 kg/s) computed by the DSBGK method and the experimental
data [38], TH = 347.1 K and TL = 289.2 K, cited from [28]
He, p0 ∈ [67.3, 799] Pa Ne, p0 ∈ [66.9, 532] Pa Ar, p0 ∈ [67.4, 294] Pa
δ0 ṀExp. ṀDSBGK δ0 ṀExp. ṀDSBGK δ0 ṀExp. ṀDSBGK
0.624 0.137 0.180 0.873 0.307 0.349 1.70 0.355 0.371
0.865 0.193 0.219 1.22 0.399 0.423 2.01 0.409 0.401
1.11 0.233 0.252 1.41 0.435 0.459 2.36 0.430 0.432
1.48 0.278 0.296 1.58 0.469 0.488 2.70 0.487 0.464
1.98 0.341 0.342 2.10 0.571 0.560 3.03 0.496 0.485
2.47 0.402 0.383 2.78 0.629 0.640 3.38 0.531 0.511
2.98 0.429 0.414 3.48 0.718 0.711 4.01 0.554 0.542
3.47 0.465 0.441 4.18 0.795 0.753 4.73 0.636 0.575
4.33 0.526 0.479 4.87 0.819 0.806 5.37 0.655 0.604
5.57 0.532 0.521 5.24 0.845 0.819 6.04 0.611 0.625
6.80 0.607 0.559 6.09 0.886 0.861 6.72 0.670 0.645
7.41 0.587 0.574 6.94 0.896 0.897 7.41 0.702 0.664
state is not applicable because it requires a considerable time due to low flow speed,
which is usually dominated by noise in the measurements. The current pulse-decay
approach is questionable, where the transient pressure variations at the two ends
of rock sample are measured and then applied in a mathematical model (PDE) to
inversely estimate the apparent permeability, which as a parameter of the PDE is
regulated until the pressure response at the outlet computed by the PDE matches
the measured one under the same inlet pressure pulse. Obviously, this experimental
approach is not pure measurement and its validity depends on the reliability of the
PDE and its empirical correction for the Kn effect.
We first present the derivation of permeability variation with the gas pressure
in a 2D channel flow problem of the slip regime. To make the results comparable
with usual experimental data, pressure difference instead of an external body force
is used to drive the flows as detailed in [32]. The channel length and width are
L1 and L2 , respectively, and the origin of Cartesian coordinate system is placed at
the channel center. The pressures at the two ends are p(x1 = −0.5L1 ) = pinlet and
p(x1 = 0.5L1 ) = poutlet , where the pressure difference pinlet − poutlet is used to drive
the flows in the x1 direction. In the case of low speed, the variations of temperature
and dynamic viscosity are negligible. Additionally, we also can neglect the compress-
ibility in the momentum equation although the density variation associated with the
pressure difference is nonzero. Usually, we have u = (u1 , 0) and u1 = u1 (x1 , x2 ), but
p, dp/dx1 and Kn = λ/L2 can be deemed functions of x1 alone.
The molecular
√ mean free path varies with the mass density according to Eq. (2.76),
i.e., λ = m/(2π kB T )16μ/(5ρ) that indicates λρ = const. The incompressible
Navier-Stokes momentum equation for this unidirectional problem is as follows:
∂ 2 u1 1 dp
= . (3.40)
(∂x2 )2 μ dx1
We assume that the α portion of the incoming molecules reflects diffusely and
the other (1 − α) portion reflects specularly, i.e., α is the accommodation coefficient
of the Maxwellian type boundary condition. Then, the slip boundary condition at
x2 ≡ 0.5L2 is (similar for the lower boundary at x2 ≡ −0.5L2 ):
−(2 − α) ∂u1
u1 = λ
α ∂x2
(3.41)
∂u1
= −γ ,
∂x2
where γ = λ(2 − α)/α 13 is introduced for notation clarity in the following deriva-
tion. Note that noticeable slip velocity could occur even at α = 1 although the incom-
plete diffuse reflection of α < 1 will make the slip velocity larger. The slip velocity
becomes negligible in ordinary problems, where the characteristic length of velocity
gradient is much larger than λ. The solution of u1 (x1 , x2 ) is:
−1 2 dp
u1 = (L2 − 4x22 + 4γ L2 ) . (3.42)
8μ dx1
13 By a slight abuse of notation, γ denotes the specific heat ratio introduced in Sect. 3.4.2 and an
auxiliary variable introduced here for notation clarity.
The average flow velocity u1 (x1 ) over the cross-section is:

0.5L2
1
u1 = u1 dx2
0.5L2 0

−1 dp 1 3
= L + 2γ L2
2
(3.43)
4L2 μ dx1 3 2

−L22 dp 2−α
= 1+6 Kn .
12μ dx1 α
The mass flow rate Ṁ through the channel of unit height in the x3 direction is:
0.5L2
Ṁ = 2ρ u1 dx2
0
= ρL2 u1 (3.44)

−ρ dp 1 3
= L2 + 2γ L22 ,
4μ dx1 3
which indicates that the pressure gradient varies with x1 even if the pressure dif-
ference at the two ends is very small, because Ṁ is independent of x1 while ρ and
γ ∝ 1/ρ depend on x1 . This is due to nonzero density variation although the incom-
pressibility is assumed to simplify the momentum equation. Additionally, we have
0.5L1
−0.5L1 Ṁ dx1 = Ṁ L1 and thus:

L32 1 2
Ṁ = (p − poutlet
2
)+
12L1 μTRspecific 2 inlet
(3.45)
6(2 − α)
λinlet pinlet (pinlet − poutlet ) ,
L2 α
where TRspecific = TkB /m = p/ρ. We define p0 = 0.5(pinlet + poutlet ) and β =

L2 (pinlet − poutlet )/(L1 p0 ), then Eq. (3.45) can be rewritten into:

βρ0 L22 2−α
Ṁ = ρ0 TRspecific 1 + 6 Kn0 , (3.46)
12μ α
where ρ0 = ρ(p0 ) and Kn0 = Kn(ρ0 ).14 The normalized mass flow rate obtained by
the N–S equation with slip boundary condition is as follows:

Ṁ 2 2−α
= √ 1+6 Kn0 , (3.47)
βρ0 L2 2TRspecific 15 π Kn0 α
14 Inthe derivation from Eqs. (3.40) to (3.49), the subscript 0 of Kn0 corresponds to the above
definition of p0 instead of the initial state, and Kn is used as a variable inside the computational
domain to replace x1 when needed.
which decreases as Kn0 increases. In reality, it reaches minima around Kn0 = 1,

which is the Knudsen minimum phenomenon as discussed in [32] and indicates the
invalidity of using the N–S equation with slip boundary at high Kn0 .
If we neglect the variation of pressure gradient with x1 and thus dp/dx1 ≡ (poutlet −
pinlet )/L1 = −βp0 /L2 , Eq. (3.42) of u1 (x1 , x2 ) can be rewritten into:

u1 1 4x22 2−α
= √ 1− 2 +4 Kn . (3.48)
β 2TRspecific 5 π Kn0 L2 α
In the pore-scale study of digital rock sample, the permeability scalar κ depends on
the gas pressure (i.e., p0 here) and is defined as κ = μL1 U1 /(pinlet − poutlet ), where
the volumetric velocity U1 15 is equal to u1 at the x1 with ρ = ρ0 for the channel
flows with 100% porosity, i.e., Ṁ = ρL2 u1 = ρ0 L2 U1 according to Eq. (3.44). Then,
substituting Eq. (3.46) for Ṁ , we obtain:

L22 2−α
κ= 1+6 Kn0 . (3.49)
12 α
A general Klinkenberg correlation formula is obtained in [20], i.e., κ = κ∞ (1 +

b/p0 ), where κ∞ is the intrinsic/absolute permeability of the rock sample concerned
and b is a parameter to be determined.
The gas permeability variation with the pore pressure (or Kn in general) can be
accurately studied by using the pore-scale flow simulations based on the kinetic
theory. Although the DSMC method is accurate in simulating gas flows at arbitrary
Kn, it is very time-consuming at low speed due to stochastic noise. Large pressure
ratio between the inlet and outlet is usually applied to increase the signal-to-noise ratio
and hence reduce the sampling cost required for smooth results [40]. Consequently,
the flow velocity will be large in the near-continuum regime (i.e., low Kn and high
pressure) and the obtained permeability is very likely not independent of the pressure
gradient due to the nonlinear Forchheimer effect. Additionally, the object of study
is the permeability variation with pore pressure, which requires each permeability
data to be computed precisely at/around a targeted pore pressure by keeping the
pressure variation small inside the computational domain. The DVM has also been
applied to solve a linearized BGK equation in studying the gas permeability of
2D geometries [35] and small pressure drop can be used since it is a deterministic
method without stochastic noise. But, to preserve the pore-scale accuracy of flow
field, its computational cost will be significantly increased in 3D simulations due to
the additional discretization of not only the physical space but also the molecular
velocity space in the third direction [18], which are usually avoided for efficient 2D
demonstrations.
The study of gas permeability variation with the pore pressure (i.e., Klinkenberg
slippage phenomenon) by using accurate simulation method and small pressure drop
15 Bya slight abuse of notation, U denotes the flow velocity of non-inertial reference frame intro-
duced in Sect. 1.4 and the volumetric velocity introduced here.
(e.g., 1%) is first conducted by the DSBGK method [29] and 3D simulations of real
digital shale rock have been performed to show the applicability [30].16
To show the permeability variation with the pore pressure from the slip to free
molecular regimes, we first present the DSBGK simulations in a 2D artificial geom-
etry [29], where the characteristic length can be clearly defined for accurate estimate
of Kn. As shown in Fig. 3.14, the pore size Lpore = 200 nm is the characteristic length
and the total sizes of the computational domain in both directions are L = 5Lpore . The
left and right boundaries are open with pressures fixed at p0 and 0.99p0 , respectively.
The top and bottom boundaries are periodic. For each simulation at a given p0 , the
volumetric velocity component U1 along the driving direction x1 at steady state is
used to compute the permeability κ:
μ
κ= U1
(p0 − 0.99p0 )/L
(3.50)
μ k∈void nk Vk u1,k
= ,
(p0 − 0.99p0 )/L n0 Vall
where nk , Vk , u1,k are the number density, volume and flow velocity component of
the void cell k, respectively, Vall is the total volume of both void and solid cells inside
the computational domain. The gas media is methane (m = 2.663 × 10−26 kg), and
the wall temperature Twall (the gas temperature at open boundaries as well) and
gas viscosity are fixed at T0 = 300 K and μ = 1.024 × 10−5 Pa · s, respectively, for
simplicity. The viscosity change with temperature can be conveniently implemented
as shown in the previous Sects. 3.4.1–3.4.3 but is negligible for isothermal flows at
low speed. In the current study, the flow is always driven in the x1 direction and thus
only the permeability along the x1 direction is computed to obtain κ(p0 ). Then, the
Knudsen number Kn = λ0 /Lpore will be used to analyze a more general variation of
κ(Kn). The boundary condition of complete Maxwell diffuse reflection is used and
implemented in the DSBGK simulations by using neff,simple of Eq. (3.18) due to low
speeds. About 100 simulated molecules per cell are used. The BGK relaxation time
τ is selected here to match the dynamic viscosity μ, i.e., τ = μ/(nkB T ), as detailed
in Sect. 2.9.
The permeability variation with the reciprocal pressure is shown in Fig. 3.15. As
we can see, the gas permeability (i.e., apparent permeability) κ approaches a constant
(i.e., the intrinsic permeability κ∞ ) as the pressure p0 increases making Kn < 0.01,
which is consistent with the observation of conventional reservoir with low Kn,
where the permeability is a rock property and independent of the flow conditions
and the nature of fluid. Additionally, the predictions of κ∞ by the DSBGK method
and the ordinary LBM are in good agreement, i.e., 1.99 × 10−15 mm2 and 2.05 ×
10−15 mm2 , respectively. In the case of Kn < 0.01, the traditional CFD methods
based on the N–S equation and non-slip boundary condition are valid and will predict
a constant permeability as expected. At another limit with very low pressure making
Kn > 10, the permeability becomes a linear function of the reciprocal pressure as
16 More details are available at https://sites.google.com/view/nanogassim/welcome.

Fig. 3.14 Schematic of methane permeability study with Lpore = 200 nm, and a representative
result of steady state streamlines at p0 = 0.4 MPa, Kn = 0.0644 [29]
illustrated by the red-dash line in Fig. 3.15, which can be explained by the kinetic
theory [29].
The simple Klinkenberg correlation model is obtained to predict the variation of
permeability with the gas pressure [20]:
κ = κ∞ (1 + b/p0 ), (3.51)
which satisfies the general features at two limits of the above discussion. We suggest
to take the Klinkenberg model as a data fitting formula [29], instead of a physi-
cal model that is derived from the N–S equation and slip boundary condition in
a simplified geometry (usually straight channel as shown by Eq. (3.49)) and thus
has physical definitions for its parameters (with empirical corrections for compli-
Fig. 3.15 Permeability

variation with the reciprocal
pressure; solid line: DSBGK
results by pore-scale
simulations, dash-dot line:
estimation by the
Klinkenberg correlation
model calibrated at the two
ends [29]
cated geometries). Then, the two model parameters κ∞ and b will be determined
by calibration using the permeability data accurately computed by the pore-scale
simulations at a low Kn and a high Kn, respectively. This calibration for each par-
ticular digital rock significantly improves the accuracy of the Klinkenberg model
for real complicated geometries, compared to its performance as a physical model
with empirical corrections for general application, as clearly shown at the end of
this section. Figure 3.15 shows that the permeability prediction by the calibrated
Klinkenberg correlation model agrees well with the accurate permeability data in the
whole flow regime (Note: the relative discrepancy is within 10%). The validity of
calibrated Klinkenberg correlation model has also been verified by the experimental
data of Klinkenberg as plotted in Fig. 3.16.
We also present the application of DSBGK method to a real 3D digital rock of shale
sample (porosity φ = 0.1647, 1003 voxels with voxel size of 2.82 nm) [30]. As shown
in Fig. 3.17, the intrinsic permeability κ∞ can be efficiently computed by the ordinary
LBM as used here (i.e., κ∞, LBM = 8.994 nD) or the traditional CFD methods. Then,
the slope b can be determined by κ∞ and a single gas permeability at high Kn that
can be efficiently obtained by a DSBGK simulation without cell refinement, which
will become required at low Kn to keep the cell size smaller than λ. Again, the
agreement between the prediction of the calibrated Klinkenberg correlation model
and the accurate DSBGK results over a wide range of Kn is very good as shown in
Fig. 3.17. For industry applications, the permeability of rock sample is measured to
establish geologic correlation and for quantitative calculation of production rate. In
these calculation, several approximations have to be made and thus it is not necessary
to know the permeability with a high degree of accuracy [20]. Thus, the calibrated
Klinkenberg correlation model is appropriate for industry applications. This hybrid
Fig. 3.16 Comparison

between the calibrated
Klinkenberg correlation
model and experimental data
of two samples, p is the
average pore pressure used in
experiments, cited from [29]
DSBGK-LBM simulation scheme [30] makes the best of each simulation method,
namely using the LBM at Kn = 0 to ensure its accuracy, and the DSBGK method
at large Kn to avoid cell refinement. Consequently, it can handle 10003 (a billion)
voxels by using about 1000 HPC processors for interesting applications.
As mentioned before, different formulas have been proposed to empirically deter-
mine the slippage factor b without resorting to the intensive pore-scale simulations
if the Klinkenberg correlation model is used as a physical model [41]. We use the
following formula proposed for nitrogen gas [8] and obtain the slippage factor for
the above real 3D geometry:
bN2 = 0.0094(κ∞ /φ)−0.5

−0.5
8.994 × 10−21
= 0.0094 (3.52)
0.1647
= 40.225 × 106 Pa.
Note that the Klinkenberg correlation model can be normalized for different gases,
i.e., κ/κ∞ = [1 + √ bLpore Kn/(p0 λ0 )], where Lpore is the characteristic pore size and
bLpore /(p0 λ0 ) ∝ b m/μ should be a constant for different gases flowing inside
the √same pore-scale geometry. Thus, we have bCH4 = bN2 mN2 /mCH4 (μCH4 /μN2 ) ≈
bN2 28/16(1.024/1.664) = 32.746 × 106 Pa, which is about 3 times larger than
b = 10.11 × 106 Pa obtained by calibration using accurate permeability data.
Fig. 3.17 Permeability of real digital rock in the x1 direction; solid line: DSBGK results by pore-
scale simulations, dash-dot line: estimation by the calibrated Klinkenberg correlation model [30]
3.4.5 Scaling Law of Gas Permeability in the Slip Regime
Although for each particular rock sample the gas permeability can be computed
by the pore-scale simulation and the calibrated Klinkenberg correlation model can
significantly reduce the simulation cost, it is still desirable if there is a general scal-
ing law to predict the gas permeabilities of different rock samples. We limit our
following discussion to the slip regime, where the Klinkenberg correlation model,
κ = κ∞ (1 + b/p0 ), can be derived and thus its parameters have physical meanings
with possible correlation. For arbitrary but similar geometries, the relation between
the dimensionless quantities (κ − κ∞ )/κ∞ and Kn should be the same according
to the dimensional analysis (e.g., Eq. (3.49) with κ∞ = L22 /12 for straight chan-
nel), which means that b/p0 is the same for these similar geometries at the same
√
Kn due to b/p0 = (κ − κ∞ )/κ∞ . Thus, we have b ∝ 1/Lpore ∝ 1/ κ∞ because of
b/p0 ∝ bLpore Kn and κ∞ ∝ Lpore .
2
For arbitrary geometries

√ without similarity, different independent studies indicate
a scaling law of b ∝ 1/√ κ∞ /φ [34]. This is consistent with the convention that
Lpore is often scaled to κ∞ /φ. In our numerical investigation, methane gas flows
are simulated by the DSBGK method in several bended nano-channels of Fig. 3.18.
Different dimensions L, W, S 17 are used to cover a wide range of pore body-to-throat
ratios, i.e., Case 1: L = 100 nm, W = 10 nm and S = 20 nm, Case 2: L = 500 nm,
W = 10 nm and S = 90 nm, Case 3: L = 500 nm, W = 10 nm and S = 10 nm, Case
4: L = 500 nm, W = 10 nm and S = 50 nm, Case 5: L = 500 nm, W = 20 nm and
S = 20 nm. The permeability data in the slip regime of Kn < 0.1 is used to extract
17 By a slight abuse of notation,

W denotes the number of microscopic states introduced in Sect. 2.7
and a characteristic channel size introduced here.
Fig. 3.18 Schematic of methane flows in 2D bended channels [34]
Fig. 3.19 Scaling law verified by the pore-scale DSBGK simulations, and experiments of
nanofluidics and core samples [34]
b and κ∞ for each case. In the DSBGK simulations, the wall temperature Twall (the
gas temperature at open boundaries as well) and gas viscosity are fixed at T0 = 300
K and μ = 1.024 × 10−5 Pa · s, respectively, for simplicity. The boundary condition
of complete Maxwell diffuse reflection is used and implemented in the DSBGK
simulations by using neff,simple of Eq. (3.18) due to low speeds. About 60 simulated
molecules per cell are used. The BGK relaxation time τ is selected here to match
the dynamic viscosity μ, i.e., τ = μ/(nkB T ), as detailed in Sect. 2.9. Experimentally,
steady-state methane gas flows are measured by using both reactive-ion etched nano-
channels with a controlled channel size on a silicon wafer (about 500 nm in depth)
√ and
core samples (carbonate and shale rocks). The correlations between b and 1/ κ∞ /φ
Fig. 3.20 Comparison of Case 2 between the pore-scale DSBGK simulation at Kn = 0.02 (p0 =
25.77 MPa, 6000 × 3000 cells in total) and the ordinary LBM simulation at Kn = 0 [34]
obtained from these independent studies vary across three orders of magnitude, yet
they all appear to collapse on a single scaling law as shown in Fig. 3.19.
Additionally, in order to cover the slip flow regime, the DSBGK method has been
applied at very high pressure, e.g., p0 = 25.77 MPa for Kn = 0.02 of Case 2 with
W = 10 nm. For low-speed gas flows driven by small pressure difference (small
flow compressibility as well), the incompressible N–S equation is valid and thus
the ordinary LBM can be used as reference for validating the DSBGK results. Note
that at Kn → 0 the BGK equation converges to the compressible N–S equation with
an equation of state p = nkB T that is inaccurate at high pressure. Nevertheless, at
high pressure but small pressure difference, the equation of state is immaterial, and
the solution of the compressible N–S equation is close to that of the incompressible
N–S equation because the difference (extra terms) between the two equation sys-
tems has negligible contribution when the specified flow compressibility is small.
Thus, only the pressures at the inlet and outlet need to be matched in the DSBGK
simulation at high pressure, regardless of the real density. Then, the pressure and
velocity distributions of the DSBGK simulation are accurate. Figure 3.20 shows that
the velocity distributions inside the two representative subdomains D and E obtained
by the DSBGK simulation of Case 2 at p0 = 25.77 MPa, Kn = 0.02 agree very well
with those by the LBM at Kn = 0 [34]. Near channel exits and entrances, the velocity
from DSBGK is slightly but consistently larger than that from LBM, showing small
effect of slippage in the DSBGK simulation at finite Kn = 0.02.
As expected, the BGK equation will be inaccurate at high pressure with noticeable
flow compressibility due to the inaccurate equation of state, which falls outside the
scope of the gas permeability study here. On the other hand, even with negligible flow
compressibility, the standard Boltzmann equation is deemed invalid at high pressure
due to the violation of its two assumptions, i.e., molecular chaos (see Sect. 2.4)
and the dominance of binary intermolecular collisions over ternary collisions (see
Sects. 2.2 and 2.4), which have been removed in the BGK equation. Additionally,
modification of kinetic equation (e.g., the BBGKY hierarchy equation) by including
the intermolecular potential energy can recover the usual equation of state for dense
fluid as discussed in [15].
References
1. Alder BJ, Wainwright TE (1957) Studies in molecular dynamics. J Chem Phys 27:1208–1209
2. Andries P, Aoki K, Perthame B (2002) A consistent BGK-type model for gas mixtures. J Stat
Phys 106:993–1018
3. Bird GA (1963) Approach to translational equilibrium in a rigid sphere gas. Phys Fluids 6:1518–
1519
4. Broadwell JE (1964) Study of rarefied shear flow by the discrete velocity method. J Fluid Mech
19(3):401–414
5. Bird GA (1994) Molecular gas dynamics and the direct simulation of gas flows. Clarendon
Press, Oxford
6. Baker LL, Hadjiconstantinou NG (2005) Variance reduction for Monte Carlo solution of the
Boltzmann equation. Phys Fluids 17:051703
7. Chen HD, Chen SY, Matthaeus WH (1992) Recovery of the Navier-Stokes equations using a
lattice-gas Boltzmann method. Phys Rev A 45:5339–5342
8. Civan F (2010) Effective correlation of apparent gas permeability in tight porous media. Transp
Porous Med 82:375–384
9. Frezzotti A (1997) A particle scheme for the numerical solution of the Enskog equation. Phys
Fluids 9:1329
10. Garcia AL, Wagner W (2000) Time step truncation error in direct simulation Monte Carlo.
Phys Fluids 12(10):2621–2633
11. Gupta NK, An S, Gianchandani YB (2012) A Si-micromachined 48-stage Knudsen pump for
on-chip vacuum. J Micromech Microeng 22:105026
12. Guo ZL, Xu K, Wang RJ (2013) Discrete unified gas kinetic scheme for all Knudsen number
flows: low-speed isothermal case. Phys Rev E 88(3):033305
13. Hadjiconstantinou NG (2000) Analysis of discretization in the direct simulation Monte Carlo.
Phys Fluids 12(10):2634–2638
14. He XY, Luo LS (1997) Lattice Boltzmann model for the incompressible Navier-Stokes equa-
tion. J Stat Phys 88:927–944
15. He XY, Doolen GD (2002) Thermodynamic foundations of kinetic theory and lattice Boltzmann
models for multiphase flows. J Stat Phys 107:309–328
References 117
16. Homolle TM, Hadjiconstantinou NG (2007) Low-variance deviational simulation Monte Carlo.
Phys Fluids 19:041701
17. Ho MT, Graur I (2014) Numerical study of unsteady rarefied gas flow through an orifice.
Vacuum 109:253–265
18. Ho MT, Li J, Wu L, Reese JM, Zhang YH (2019) A comparative study of the DSBGK and
DVM methods for low-speed rarefied gas flows. Comput Fluids 181:143–159
19. John B, Gu XJ, Emerson DR (2010) Investigation of heat and mass transfer in a lid-driven
cavity under nonequilibrium flow conditions. Numer Heat Transf Part B-Fund 58(5):287–303
20. Klinkenberg LJ (1941) The permeability of porous media to liquids and gases. In: Drilling and
productions practices. American Petroleum Institute, pp 200–213
21. Li ZH, Zhang HX (2009) Gas-kinetic numerical studies of three-dimensional complex flows
on spacecraft re-entry. J Comput Phys 228(4):1116–1138
22. Li J (2009) IP simulation of gas flows in the air bearing problems of head slider. PhD thesis,
Institute of Mechanics, CAS: 39–42
23. Li J, Shen C, Fan J (2010) Improvements to the low-variance deviational simulation Monte
Carlo method. Acta Aerodynamica Sinica 28(2):238–243
24. Li J (2010) Direct simulation method based on BGK equation. In: 27th international symposium
on rarefied gas dynamics, AIP conference proceedings, vol 1333, pp 283–288. (Presented first
at ESPCI, Paris, 2009)
25. Li J (2012) Comparison between the DSMC and DSBGK methods. https://arxiv.org/abs/1207.
1040
26. Li J (2014) Improved diffuse boundary condition for the DSBGK method to eliminate the
unphysical density drift. https://arxiv.org/abs/1403.3923
27. Li J (2017) DSBGK method to incorporate the CLL reflection model and to simulate gas
mixtures. https://arxiv.org/abs/1710.07795
28. Li J, Cai CP (2017) Numerical study on thermal transpiration flows through a rectangular
channel. https://arxiv.org/abs/1708.08105
29. Li J, Sultan AS (2017) Klinkenberg slippage effect in the permeability computations of shale
gas by the pore-scale simulations. J Nat Gas Sci Eng 48:197–202
30. Li J (2019) Efficient prediction of gas permeability by hybrid DSBGK-LBM simulations. Fuel
250:154–159
31. Qian YH, d’Humieres D, Lallemand P (1992) Lattice BGK models for Navier-Stokes equation.
Europhys Lett 17:479–484
Berlin
33. Titarev VA (2007) Conservative numerical methods for model kinetic equations. Comput Fluids
36(9):1446–1459
34. Tian Y, Yu X, Li J, Neeves KB, Yin XL, Wu YS (2019) Scaling law for slip flow of gases in
nanoporous media from nanofluidics, rocks, and pore-scale simulations. Fuel 236:1065–1077
35. Wu L, Ho MT, Germanou L, Gu XJ, Liu C, Xu K, Zhang YH (2017) On the apparent perme-
ability of porous media in rarefied gas flows. J Fluid Mech 822:398–417
36. Wang P, Ho MT, Wu L, Guo ZL, Zhang YH (2018) A comparative study of discrete velocity
methods for low-speed rarefied gas flows. Comput Fluids 161:33–46
37. Xu K (2014) Direct modeling for computational fluid dynamics: construction and application
of unified gas-kinetic schemes. World Scientific, Singapore
38. Yamaguchi H, Perrier P, Ho MT, Meolans JG, Niimi T, Graur I (2016) Mass flow rate measure-
ment of thermal creep flow from transitional to slip flow regime. J Fluid Mech 795:690–707
39. Yang JY, Huang JC (1995) Rarefied flow computations using nonlinear model Boltzmann
equations. J Comput Phys 120(2):323–339
40. Yang G, Weigand B (2018) Investigation of the Klinkenberg effect in a micro/nanoporous
medium by direct simulation Monte Carlo method. Phys Rev Fluids 3(4):1–17
41. Ziarani AS, Aguilera R (2012) Knudsen’s permeability correction for tight porous media.
Transp Porous Med 91:239–260
Chapter 4
Multiscale LBM Simulations
Abstract For flow problems of the continuum regime, the lattice Boltzmann method
(LBM) is a good alternative to the traditional CFD solvers based on the N–S-like
equations. It is efficient in modeling dynamic problems and very powerful for pore-
scale applications, where the simulation of interface dynamics on the real irregular
pore surface is challenging, if not impossible, to most of the traditional CFD solvers.
We start in this chapter with the basic LBM algorithm to show its correlation with
the N–S equation through the Chapman–Enskog expansion. Then, the widely used
Shan–Chen model will be introduced to simulate multiphase multicomponent flow
systems, having its applications detailed in the subsequent sections. We also present
the extension of LBM to the Darcy-scale simulations, where the LBM works as a
unified framework for simulations at different scales, i.e., both pore and Darcy scales,
and the detailed results are given at the end of this chapter. In the ordinary application
of LBM for computing the absolute permeability, we clarify the prevailed confusion
interpreted as viscosity-dependent permeability and reveal the underlying rarefaction
mechanism that has been commonly oversighted. Additionally, we also discuss the
application of large eddy simulation of turbulence in the LBM framework and the
same idea can be extended to model non-Newtonian fluids.
4.1 Basic Algorithm
To make it clear for applications, the variables of the lattice Boltzmann method
(LBM) [3, 10, 22] are used in a dimensional form of SI base units and the normal-
ization process can be applied to the results if needed. The computational domain
is discretized by uniform spatial grids with a constant distance Δx in the x1 , x2 , x3
directions and computational quantities are defined at those discrete grids. At each
spatial grid, we specify several lattice velocities eα indexed by α ∈ [0, Q − 1] for
Electronic supplementary material The online version of this chapter

(https://doi.org/10.1007/978-3-030-26466-6_4) contains supplementary material, which is
available to authorized users.

120 4 Multiscale LBM Simulations
Q 1 directions in total. The lattice velocity eα is either static for α = 0 or transports

particles from the current grid at x to its neighboring grids at x + Δteα after each
time step Δt. This implies that the construction of lattice velocities must complies
with the uniform layout of spatial grids, which imposes a great limitation for its
extension to the simulations of high K n as discussed in Chap. 3. The magnitude
of eα depends on c = Δx/Δt. For example, in the two-dimensional D2Q9 model
[22], e0 = (0, 0) and ω0 = 4/9, eα = (cos θα√, sin θα )c and θα = (α − 1)π/2 and
ωα = 1/9 for α ∈ [1, 4], eα = (cos θα , sin θα ) 2c and θα = (α − 5)π/2 + π/4 and
ωα = 1/36 for α ∈ [5, 8], where ωα 2 is the weighting factor.
The only unknown is the density distribution function f α (x, t), which is used to
compute the flow velocity u and mass density ρ (and then pressure p) at (x, t). Here,
we discuss the algorithm only for the mass and momentum transports in the absence
of external body force. The equilibrium distribution function of LBM is as follows:

3 9 3
f αeq = ρωα 1 + 2 eα · u + 4 (eα · u)2 − 2 u · u , (4.1)
c 2c 2c
where the definitions of ρ and u using f α will be introduced after Eq. (4.6).
The D2Q9 model as well as other possible
lattice models satisfy the following
Q−1
important properties (note: e0 = 0 and α = α=0 ):

E (n) = ωα eα,i1 eα,i2 · · · eα,in , (4.2)
α
where i 1 , · · · , i n ∈ [1, 3] are indices for the x1 , x2 , x3 directions and

E (0) = ωα = 1,
α
c2
E (2) = ωα eα,i eα, j = δi j ,
α
3 (4.3)
c4
(4)
E = ωα eα,i eα, j eα,k eα,l = (δi j δkl + δik δ jl + δil δ jk ),
α
9
E (2n+1) = 0, n = 1, 2, 3, · · · ,
where simpler indices i, j, k, l ∈ [1, 3] are used for the x1 , x2 , x3 directions and δi j
is the Kronecker delta function.
According to the above properties, we can compute the following zero-to-three-
eq
order moments of f α :
1 By a slight abuse of notation, Q denotes the total number of lattice velocities introduced here and
auxiliary variables used elsewhere for notation clarity.

2 By a slight abuse of notation, ω denotes the power of the dependence of μ on T introduced in
Sect. 2.3, ωα denotes the weighting factor associated with eα introduced here, and ω denotes the
acentric factor introduced in Sect. 4.5.3.
4.1 Basic Algorithm 121

f αeq = ρ,
α

eα,i f αeq = ρu i ,
α
c2 (4.4)
eα,i eα, j f αeq = ρδi j + ρu i u j ,
α
3
c2
eα,i eα, j eα,k f αeq = ρ(δi j u k + δik u j + δ jk u i ).
α
3
As shown later, the properties in Eq. (4.4) are very important in the derivation of
Navier–Stokes-like equation from LBM. The general rule in constructing new lattice
eq
models eα , ωα and f α is to satisfy Eq. (4.4) and then a Navier–Stokes-like equation
can always be recovered from LBM.
The explicit updating algorithm of the only unknown fα for each Δt is a relaxation-
propagation process and the single-relaxation-time scheme is3 :
eq
f α (x, t) − f α (x, t)
f α (x + Δteα , t + Δt) = f α (x, t) + , (4.5)
τ
where τ is the normalized relaxation time in the LBM and its selection will be
discussed at the end of Sect. 4.2. The LBM updates its unknown at the concerned
spatial grid by using the info of its immediate neighbors and thus exchanges only the
info of outmost layers of each subdomain handled by one of the processes of parallel
computation, which requires low cost for the communication between processes and
thus attains high efficiency of parallel computation.
The relaxation process of Eq. (4.5) should conserve mass and momentum as
follows:

f αeq = fα ,
α α
(4.6)
eα,i f αeq = eα,i f α ,
α α

which give the definitions ρ = α f α and u i = (1/ρ) α eα,i f α according to
Eq. (4.4). Now, the governing equation (4.5) is closed and has several parameters
Δx, Δt, τ .
In the boundary condition, we usually assume that the whole computational
domain is divided into many uniform voxels, which are occupied by either solid or
fluid, and the variables are defined at the voxel centers (i.e., uniform spatial grids). The
3 The lattice Boltzmann method actually solves a BGK-type equation rather than the Boltzmann
equation. Additionally, the relaxation time τ of the original BGK equation (2.81) has physical unit
(second) but τ in LBM is dimensionless. The correlation between τ and viscosity in LBM, i.e.,
Eq. (4.19), is completely different from that in the BGK equation, i.e., Eq. (2.83).
bounce-back boundary will be used for the non-slip boundary condition, i.e., allowing
f α (xfluid , t), which belongs to the fluid grid at xfluid and moves toward the solid grid at
xsolid in the eα direction, to bounce back at the same time step to update fα (xfluid , t) =
f α (xsolid , t) = f α (xsolid , t) = f α (xfluid , t), where α is the index of eα opposite to
eα . This implementation of non-slip boundary condition is not precise since the
actual solid surface is located at the middle between solid and fluid grids. Never-
theless, the simple bounce-back scheme makes the simulation efficient and robust.
Boundary with specified pressure or flow velocity constraint (e.g., open boundary
or moving solid boundary) will be modeled by the non-equilibrium extrapolation
eq eq
scheme proposed in [6], i.e., f α (xBC , t) = f α (xfluid , t) − f α (xfluid , t) + f α (xBC , t)
and then f α (xfluid , t) = f α (xBC , t), where the macroscopic quantities ρ( p) and u of
eq
f α (xBC , t) are either all prescribed as constraints, or extrapolated from the fluid
domain if one of them is not prescribed. As we can see, the non-equilibrium part
eq
of the distribution function, f α (xfluid , t) − f α (xfluid , t), is extrapolated from xfluid to
xBC to determine f α (xBC , t). This scheme is simple and robust for arbitrary geometry.
The LBM and its boundary schemes have been discussed in [27, 28].
The velocity space discretization in LBM is rather simple and thus crude although
its uniform discretization of physical space can be as fine as needed. LBM sacrifices
the numerical accuracy in solving the BGK-like equation to achieve the algorithm
simplicity. For example, in two-dimensional flows, the ordinary LBM adopts the
D2Q9 model, where only 9 points are used to discretize the unbounded velocity
space. By contrast, the similar deterministic DVM method requires at least 4 × 24,
4 × 40 and 4 × 48 velocity points distributed in the polar coordinate for the two-
dimensional lid-driven cavity flow at K n = 0.1, 1 and 8, respectively, as detailed
in Sects. 3.3 and 3.4.2. When the layout in the Cartesian coordinate is used as in
the LBM, the minimum number of velocity points of the DVM will increase, even
with the help of non-uniform discretization (e.g., half-range Gauss-Hermite velocity
grid). Thus, it is expected that the ordinary LBM is inaccurate at K n > 0.1 for
general problems (see the comparison between LBM D2Q9 and DVM D2Q1600 in
Fig. 4.6), except unidirectional flows (e.g., channel flow). As shown in the following
derivation of Sect. 4.2, LBM converges to the Navier–Stokes-like equation and thus
it works well at the limit of K n → 0.
4.2 Chapman–Enskog Expansion
The Chapman–Enskog expansion [2] can be used in the LBM to derive a Navier–
Stokes-like equation and a formula is consequently obtained to correlate the LBM
model parameters to the fluid viscosity, through which the viscosity is implicitly
implemented in LBM simulations. Although the validity of Chapman–Enskog expan-
sion that has a formal definition of time derivative without tangible mathematical
sense is not recognized by many mathematicians, the obtained correlation formula
usually works as long as the model parameters are carefully selected to make the
Mach number and Knudsen number small. Note that LBM works well at high Reynold
4.2 Chapman–Enskog Expansion 123
number as shown in the following Sect. 4.5.1. We present the following Chapman–
Enskog expansion for the LBM, which is based on the version of [10] but has mod-
ifications developed in [17] that lead to a general formula for computing the strain
rate tensor [16].
According to the Taylor expansion, we can rewrite Eq. (4.5) into4 :
∞
eq
Δt n f α (x, t) − f α (x, t)
Dtn f α (x, t) = , (4.7)
n=1
n! τ
where Dt = (∂t + eα · ∇). The Chapman–Enskog expansion in LBM is5 :

∞
∞

f α = f α(0) + f α(n) = f αeq + f α(n) ,
n=1 n=1
∞
(4.8)

∂t = ∂tn ,
n=0

where the expansion ∂t = ∞ n=0 ∂tn of the time derivative is just a formal definition
but not executable for any given analytical formula of f α (x, t) and thus this expansion
has no tangible mathematical sense.
Then, the terms in Eq. (4.7) can be sorted according to the order of magnitude
and Eq. (4.7) can be replaced by a series of equations arranged into a consecutive
order of magnitude:
−1 (1)
Δt (∂t0 + eα · ∇) f αeq = f ,
τ α
Δt 2 −1 (2) (4.9)
Δt (∂t0 + eα · ∇) f α(1) + Δt∂t1 f αeq + (∂t0 + eα · ∇)2 f αeq = f ,
2 τ α
···
In order to make each f α(n) tractable in Eq. (4.9) and meanwhile the conservation
rules of Eq. (4.6) still satisfied, the following harsh assumptions are used to replace
Eq. (4.6) (note: f α(0) = f α as assumed in Eq. (4.8)):
eq

f α(n) = 0, ∀n = 0,
α
(4.10)
eα,i f α(n) = 0, ∀n = 0.
α
4 By a slight abuse of notation, n denotes the molecular number density introduced in Sect. 2.1,
superscript n denotes the term index of model property introduced in Eq. (4.2) and the term indices
of Taylor and Chapman–Enskog expansions introduced here.
5 By a slight abuse of notation, f (2) denotes a two-particle distribution function introduced in
(2)
Sect. 2.4 and f α denotes the second-order term in the expansion of f α introduced here.
Note that the following commutative

properties hold for the binary opera-
eq eq eq eq
tions:
e.g., (∂
α t0 f α ) = ∂ t0 ( α f α ) and (e
α α · ∇ f α ) = α ∇ · (e α f α ) =∇·
eq
(e
α α f α ). By rewriting the second equation with the help of first one of Eq. (4.9)
and using Eqs. (4.4) and (4.10), the zero-order moments of the two equations of
Eq. (4.9) (note: see the definition of moment in Eq. (4.4)) are:
∂ρ ∂(ρu j )
+ = 0,
∂t0 ∂x j
(4.11)
∂ρ
= 0.
∂t1
Similarly, we can get the first-order moments of the two equations of Eq. (4.9):
2
∂(ρu i ) ∂ c
+ ρδi j + ρu i u j = 0,
∂t0 ∂x j 3
(4.12)
∂(ρu i ) 1 ∂
+ 1− eα,i eα, j f α(1) = 0.
∂t1 2τ ∂ x j α
Using Eqs. (4.4), (4.9), (4.11) and (4.12), we have:

eα,i eα, j f α(1) = −τ Δt eα,i eα, j (∂t0 + eα · ∇) f αeq
α α

∂ c 2
∂
= −τ Δt ( ρδi j + ρu i u j + eα,i eα, j eα,k f αeq ]
∂t0 3 ∂ xk α (4.13)

−c2 ∂ ∂(ρu i u j ) ∂
= −τ Δt δi j (ρu k ) + + eα,i eα, j eα,k f αeq ,
3 ∂ xk ∂t0 ∂ xk α
where
∂(ρu i u j ) ∂(ρu j ) ∂(ρu i ) ∂ρ
= ui +uj − ui u j
∂t0 ∂t0 ∂t0 ∂t0

∂ c2 ∂ c2 ∂(ρu k )
= −u i ρδ jk + ρu j u k − u j ρδik + ρu i u k + u i u j
∂ xk 3 ∂ xk 3 ∂ xk
(4.14)
c2 ∂ρ c2 ∂ρ ∂(ρu j u k ) ∂(ρu i u k ) ∂(ρu k )
= −u i −uj − ui −uj + ui u j
3 ∂x j 3 ∂ xi ∂ xk ∂ xk ∂ xk
c2 ∂ρ c2 ∂ρ ∂(ρu i u j u k )
= −u i −uj −
3 ∂x j 3 ∂ xi ∂ xk
4.2 Chapman–Enskog Expansion 125
and

∂ ∂ c2
eα,i eα, j eα,k f αeq = ρ(δi j u k + δik u j + δ jk u i )
∂ xk α ∂ xk 3
(4.15)
c2 ∂(ρu k ) c2 ∂ρ c2 ∂u j c2 ∂ρ c2 ∂u i
= δi j + uj + ρ + ui + ρ .
3 ∂ xk 3 ∂ xi 3 ∂ xi 3 ∂x j 3 ∂x j
Substituting Eqs. (4.15) and (4.14) into Eq. (4.13), we get6 :

c2 ∂u j ∂u i ∂
eα,i eα, j f α(1) = −τ Δt ρ + − (ρu i u j u k ) , (4.16)
α
3 ∂ xi ∂x j ∂ xk
which is first obtained in [16] to estimate the strain rate tensor for the application of
large eddy simulation (LES) of turbulence in LBM.
Assembling equations in Eq. (4.11) by using ∂t = ∞ n=0 ∂tn ≈ ∂t0 + ∂t1 , we get:
∂ρ ∂(ρu j )
+ = 0. (4.17)
∂t ∂x j
Assembling equations in Eq. (4.12) and using Eq. (4.16), we get:

∂(ρu i ) ∂(ρu i u j ) ∂ c2 ρ
+ =− +
∂t ∂x j ∂ xi 3

∂ c2 ∂u j ∂u i ∂
(τ − 0.5)Δt ρ + − (ρu i u j u k ) .
∂x j 3 ∂ xi ∂x j ∂ xk
(4.18)
The solutions of ρ and u in LBM simulations satisfy Eqs. (4.17) and (4.18),
which are different from the standard incompressible N–S equation. But, if we select
the model parameters √(i.e., Δx, Δt and τ ) carefully such that the Mach number is
small (i.e., |u| c/ 3 in LBM), the relative variation of ρ and the magnitude of
∂(ρu i u j u k )/∂ xk relative to that of strain rate tensor are negligible in Eq. (4.18).
Additionally, to avoid the undesired rarefaction effect (e.g., slippage effect) in the
conventional pore-scale flows, the selection √ of model parameters should also make
the Knudsen number small, i.e., K n LBM = π/6(τ − 0.5)/Nthroat [34], where Nthroat
is the number of spatial grids used to discretize the throat size in the pore-scale
simulations. Then, c2 ρ/3 and u of LBM correspond to p and u, respectively, of the
standard incompressible N–S equation, where the fluid kinematic viscosity ν 7 can
be implemented in LBM via Eq. (4.19) as indicated by Eq. (4.18):
6 Asa general formula, Eq. (4.16) contains arbitrary Δt and c.

7 By a slight abuse of notation, ν denotes the intermolecular collision frequency introduced in
Sect. 2.8 and the kinematic viscosity introduced here.
ν = (τ − 0.5)Δtc2 /3. (4.19)
Since ρ of ordinary LBM is just an auxiliary variable to compute p, its initial value
ρ0 could be different from the real ρreal of incompressible fluid. Then, the computed
pρreal /ρ0 corresponds to the real pressure distribution of the physical problem.
4.3 Shan–Chen Model for Multiphase Multicomponent

Systems
The Shan–Chen model [24, 25] is a numerically simple and physically justifiable
model, where the interaction between every two components/phases is modeled using
attractive or repulsive force depending on the particular application. According to
the Young equation, the static contact angle θeq at equilibrium state is a function
of three interfacial tension (IFT) coefficients. Thus, for each fluid-solid system, we
can calibrate the interaction force model/magnitude to match θeq (e.g., measured on
a flat solid surface) and the fluid-fluid IFT coefficient as well, and then apply the
force model (instead of parameters directly for θeq and the fluid-fluid IFT) to flow
simulations around arbitrary geometry surface of interest. The dynamic contact angle
θdy and movement of contact line will be automatically captured, as the consequence
of flow dynamics, by applying this calibrated interaction forces between different
components/phases. The validity of this methodology can be verified by comparing
the steady state components/phase distribution or the evolution of interface, if it
is not easy to accurately measure θdy . By contrast, geometric constraint of θdy is
usually imposed at the interfaces between different components/phases according to
correlation models (e.g., relating the deviation θdy − θeq to the contact-line speed) in
many traditional CFD simulations based on the N–S-like equation. These correlation
models are usually not unique (e.g., θdy also depends on the flow field in the vicinity of
the moving contact line) and thus empirical [29]. Additionally, the implementation
of geometric constraint requires high spatial resolution (e.g., grid refinement) for
irregular solid surface to determine the local tangential directions.
In the Shan–Chen model [24, 25], the number of molecules of the component σ
having the velocity eα at x and time t is denoted by f ασ (x, t), where σ ∈ [1, Ncomp ]
and Ncomp is the total number of chemical components. At the initial state, we pre-
scribe the spatial distributions of the mass density ρ σ and flow velocity uσ for each
component and then the initial value fασ (x, 0) is determined by the prescribed equilib-
σ,eq σ,eq
rium distribution as f ασ (x, 0) = f α (x, 0) = f α (ρ σ , uσ ). The explicit updating
σ
algorithm of the only unknown f α (x, t) is:
σ,eq
fα (x, t) − f ασ (x, t)
f ασ (x + Δteα , t + Δt) = f ασ (x, t) + , (4.20)
τσ
4.3 Shan–Chen Model for Multiphase Multicomponent Systems 127
σ,eq
where f α is defined using an auxiliary variable uσ,eq (x, t):

3 9 3 σ,eq σ,eq
f ασ,eq σ
= ρ ωα 1 + 2 eα · u σ,eq σ,eq 2
+ 4 (eα · u ) − 2 u ·u . (4.21)
c 2c 2c
The normalized relaxation time τ σ is determined

from the viscosity of component
σ via Eq. (4.19). As in the ordinary LBM, ρ σ = α f ασ , but the definition of uσ,eq
in the Shan–Chen model is:
ρ σ u + τ σ Fσ
uσ,eq = , (4.22)
ρσ
where Fσ (x, t) is related to the total volume force acting on the component σ , and
u is defined as follows to conserve momentum:
σ
σ
σ (1/τ ) α (eα f α )
u = σ
σ
, (4.23)
σ (1/τ ) α fα
Ncomp
where σ = σ =1 . Generally speaking [14], Fσ contains three parts: fluid-fluid
interaction F , fluid-solid interaction F2,σ and
1,σ
external
3,σ
σ body force F . For the
fluid-mixture
element having total mass of σ α f α , the momentum σincrease
σ,eq σ σ σ σ σ σ
σ α e α ( f α − f α )/τ = σ [(ρ u − e
α αα f )/τ + F ] = σ F after
each Δt through the relaxation process is equal to σ F3,σ if we consider the external

body force alone. This momentum increase physically should be Δtg σ α f ασ ,
where g is the external body force per unit mass. Thus, we have F3,σ = Δtρ σ g with
arbitrary Δt for a general expression. Note that even if F1,σ = F2,σ = F3,σ = 0,
there is still momentum exchange between different components via u in the Shan–
σ,eq
Chen model, i.e., α eα ( f α − f ασ )/τ σ = 0 for each component σ . Additionally,
in the
absenceσ,eq of external bodyand surface forces, i.e., F2,σ = F3,σ = 0, we have
σ α eα ( f α − f ασ )/τ σ = σ F1,σ ≡ 0 as required by the momentum conser-
vation for the fluid mixture as a whole.
The flow velocity u of the fluid mixture in the Shan–Chen model is equal to
the mean velocity before and after implementing the force term and computed as
follows:

(eα f σ ) + 0.5 σ Fσ
u = σ α α σ . (4.24)
σ ρ
The fluid-solid interaction F2,σ will be discussed in the following sections for
each particular application. Here, we discuss the fluid-fluid interaction F1,σ that
determines the equation of state in LBM simulations. In general, the resultant force
exerted on the fluid element of component σ at x by its nearest neighboring fluid
elements at x + Δteα of all components (including σ ) is8 :
Q−1
8 The summation over α in computing force interactions should be α=1 , which is replaced by
Q−1
α = α=0 for notation clarity thanks to e0 = 0.

F1,σ (x, t) = −G σ σ ψ σ σ (ρ σ (x, t)) ωα ψ σ σ (ρ σ (x + Δteα , t))eα ,(4.25)
σ α

where the effective mass ψ σ σ are positive functions of ρ σ and usually increase with

ρ σ , G σ σ are parameters controlling the magnitudes of different interactions and

G σ σ = G σ σ is required for σ = σ according to the Newton’s third law. Generally

speaking, ψ σ σ and ψ σ σ can take different forms when σ = σ . For example, we can
have ψ σ σ = ψconst
σ σ
exp(−ρconst /ρ σ ), where ψconst
σ σ
and ρconst are two parameters for

the interaction between molecules of the same component σ , and ψ σ σ = ρ σ for the
interaction between different components.
In the presence of fluid-fluid interaction F1,σ , the equation of state is different from

the previous form, i.e., p LBM = c2 ρ/3. We assume that ψ σ σ (ρ σ (x + Δteα , t)) =

ψ σ σ (x + Δteα , t) varies smoothly in the physical space and so has:

ψ σ σ (x + Δteα , t) ≈ ψ σ σ (x, t) + Δteα · ∇ψ σ σ (x, t). (4.26)
Substituting Eq. (4.26) into Eq. (4.25) and using Eq. (4.3), we obtain:
Δtc2
F1,σ (x, t) = −G σ σ ψ σ σ (x, t) ∇ψ σ σ (x, t)
σ σ σ
3
(4.27)
−Δtc2 1 σ σ σ σ σ σ
=∇ G ψ (x, t)ψ (x, t) .
3 2 σ σ

As indicated by the discussion after Eq. (4.23), σ F1,σ /Δt is the resultant force
exerted on the fluid-mixture per unit volume and thus contributes to −∇ p. Corre-
spondingly, the equation of state in the Shan–Chen model is [26]:

c2 1 σ σ σ σ σ σ
p LBM = ρσ + G ψ ψ . (4.28)
3 σ
2 σ σ
4.3.1 Two-Phase Flows of Single Component
For the simulations of liquid-gas coexistence (e.g., water-vapor system of the same
chemical component with possible phase change), the above general evolution algo-
rithm of Eq. (4.20) is unchanged. The notation σ can be omitted since Ncomp = 1
here. We first discuss the evolution equation that depends on the fluid-fluid interac-
tion F1 . The fluid-solid interaction F2 determines the interface dynamics on the solid
surface and will be discussed later. The attractive force exerted at x by the fluid at
neighboring grids x + Δteα is computed as:

F1 (x, t) = −Gψ(ρ(x, t)) ωα ψ(ρ(x + Δteα , t))eα , (4.29)
α
where G is a negative constant for the attractive force that is important for the sim-
ulations of two-phase coexistence of the same component. At fluid grids away from
the solid grids (i.e., grids located inside the solid phase), F = F1 + F3 is determined
and then f α (x, t) can be updated according to Eq. (4.20). Following the derivation
of [11] but here considering the general
case with arbitrary
Δx and Δt, the com-
puted macroscopic quantities ρ = α f α and u = ( α eα f√ α + 0.5F)/ρ satisfy the
following equations in the incompressible limit of |u| c/ 3:
∂ρ
+ ∇ · (ρu) = 0,
∂t
∂ρu
+ ∇ · (ρuu) = ρg + ∇ · PLBM + ∇ · [ρν(∇u + ∇uT )],
∂t
PLBM = − p LBM I + ς LBM , (4.30)

c2 1
p LBM
= ρ + Gψ , 2
3 2
1
ς LBM = κ LBM [(ψΔψ + |∇ψ|2 )I − ∇ψ∇ψ],
2
where I is the identity tensor and κ LBM = −Δt 2 c4 G/18 is positive as G is neg-
ative. The interfacial tension coefficient of LBM simulation depends on κ LBM . In
recovering PLBM of the above equation, we retained higher order terms in the Taylor
expansion of ψ(ρ(x + Δteα , t)) = ψ(x + Δteα , t):
∂
ψ(x + Δteα , t) ≈ψ + Δteα, j ψ+
∂x j
(4.31)
Δt 2 ∂ 2ψ Δt 3 ∂ 3ψ
eα, j eα,k + eα, j eα,k eα,l ,
2 ∂ x j ∂ xk 6 ∂ x j ∂ x k ∂ xl
where ψ is used to denote ψ(x, t) for notation clarity. Substituting Eq. (4.31) into
Eq. (4.29) and using Eq. (4.3), we obtain:
−c2 ∂ψ c4 Δt 3 ∂ 3ψ
Fi1 (x, t) = Δt Gψ − Gψ
3 ∂ xi 9 2 ∂ xi ∂ x j ∂ x j
2
∂ −c Δt
= Gψψδi j − (4.32)
∂x j 3 2

∂ c4 Δt 3 ∂ 2ψ 1 ∂ψ ∂ψ ∂ψ ∂ψ
G ψ δi j + δi j − .
∂x j 9 2 ∂ xk ∂ xk 2 ∂ xk ∂ xk ∂ x j ∂ xi
Again, F1 /Δt is the resultant force per unit volume and thus contributes to ∇ · PLBM ,
as reflected in Eq. (4.30).
The standard Navier–Stokes–Korteweg (N–S–K) equation as described in [9] is:
∂ρ
+ ∇ · (ρu) = 0,
∂t
∂ρu
+ ∇ · (ρuu) = ρg + ∇ · P + ∇ · [ρν(∇u + ∇uT )],
∂t
P = − pI + ς , (4.33)
ρ Rspecific T
p= − a2 ρ 2 ,
1 − a1 ρ
1
ς = κ[(ρΔρ + |∇ρ|2 )I − ∇ρ∇ρ],
2
where the van der Waals equation of state is used as an example and the parameter
κ controls the interfacial property IFT. Compared to the Navier–Stokes–Korteweg
equation, the drawback of the Shan–Chen model is that in order to match the tensor
ς we should let ψ = ρ and κ LBM = κ by appropriately selecting Δx, Δt and G.
Consequently, the equation of state in LBM simulations is different from the usual
ones. On the other hand, we can inversely determine ψ(ρ) according to the desired
preal (ρ) using the inherent relationship in LBM: p LBM = c2 (ρ + 0.5Gψ 2 )/3. After
getting ρ, we let p LBM = preal (ρ) and then ψ can be inversely determined from
p LBM and ρ using the above inherent relationship [11, 33]. We can match arbitrary
preal (ρ) including the popular Peng–Robinson equation of state, which makes the
LBM simulations stable at high density ratio up to 1000 between liquid and gas
phases [33]. But, the disadvantage of computing ψ using the prescribed preal (ρ) is
that ς cannot be exactly matched since ψ then is not linearly proportional to ρ [11].
It is noteworthy that although the LBM with an accurate equation of state can be
used to model flow problems with phase change/coexistence, its application is still
limited to low speed to make sure the compressibility of each phase during the flow
process is negligible because the constitutive equation between the stress tensor and
strain rate tensor of LBM in Eq. (4.18) is still different from the usual Eq. (1.17) of
compressible flow.
When a fluid grid xfluid is adjacent to the solid grids, the calculation of F(xfluid , t)
is performed as follows. First, the number of terms in F1 (xfluid , t) of Eq. (4.29) for the
fluid-fluid attractive interaction is reduced because some neighboring grids are solid
now and thus neglected in computing F1 (xfluid , t). Then, we compute the attractive
force F2 (xfluid , t) exerted by the solid phase to fluid at xfluid as described in [1]:

F2 (xfluid , t) = −Gψ(ρ(xfluid , t)) ωα ψ(ρsolid )eα
α∈αsolid
(4.34)
= −Gψ(ρ(xfluid , t))ψ(ρsolid ) ωα eα ,
α∈αsolid
where α takes those values corresponding to the neighboring solid grids and ρsolid is
a constant. If we set ρsolid close to the liquid density, the liquid phase is the wetting
phase and the gas phase is the non-wetting phase, and vise versa. Similarly in the
following simulations of two immiscible components (e.g., oil-water system), the
σ
component σ having a larger ρsolid is the wetting phase. Once F = F1 + F2 + F3 at
(xfluid , t) is determined, f α (x, t) can be updated by Eq. (4.20) for all fluid grids.
4.3.2 Flows of Two Immiscible Components
For modeling flows of two immiscible components/fluids (e.g., oil-water system),

the above general evolution algorithm of Eq. (4.20) is unchanged and Ncomp = 2
here. For simplicity, we assume that the densities of the two fluids are the same, and
neglect the attractive force within the same component. The external body force is

also neglected in the following discussion. Let ψ σ σ = ρ σ for simplicity and then
the repulsive force exerted on the component σ at x by the other component σ at
neighboring grids x + Δteα is:

F1,σ (x, t) = −G σ σ ρ σ (x, t) ωα ρ σ (x + Δteα , t)eα , σ = σ, (4.35)
α

where G σ σ = G σ σ is a positive constant for the repulsive force between two immis-
cible fluids. Similar to Eq. (4.34) as described in [1], the repulsive force exerted by
the solid phase as component σ to the component σ at xfluid is:

F2,σ (xfluid , t) = −G σ σ ρ σ (xfluid , t) σ
ωα ρsolid eα
α∈αsolid
(4.36)
σσ σ σ
= −G ρ (xfluid , t)ρsolid ωα eα , σ = σ.
α∈αsolid
We can change the parameters of ρsolid 1

and ρsolid
2
to adjust the static contact angle,
as mentioned before. After getting F and F , f ασ (x, t) is updated
1,σ 2,σ
using Eq. (4.20).
Meanwhile, the density of each component is computed as ρ σ = α f ασ and the flow
velocity u of the fluid mixture is computed by Eq. (4.24). Although a repulsive force
is used to model immiscibility, the density ρ σ of each component σ is close but not
equal to zero in the area dominated by the other component σ , which implies that
the Shan–Chen model used here is weakly miscible. The solubility and interfacial

tension coefficient depend on G σ σ .
As discussed in [30], the traditional Navier–Stokes and Cahn–Hilliard (N–S–C–
H) equations can be approximately recovered from the Shan–Chen LBM algorithm if

we assume that in their respectively dominated areas ρ σ and ρ σ are equal to the same
value ρ0 , which implies that the total density ρ of fluid mixture is almost constant, i.e.,

ρ = ρ σ + ρ σ ≈ ρ0 , inside the whole computational domain due to weak solubility.

We also assume that the kinematic viscosities ν σ and ν σ are equal to the same value ν,

which implies τ σ = τ σ = τ . We compute an order parameter as = ρ σ − ρ σ and

the pressure p LBM = c2 [ρ + G σ σ (ρ 2 − 2 )/4]/3 = c2 (ρ σ + ρ σ + G σ σ ρ σ ρ σ )/3.
Following the derivation of [30] but here considering arbitrary Δx and Δt, the com-
puted , u and p LBM satisfy
√ the following N–S–C–H-like equations in the incom-
pressible limit of |u| c/ 3:
∇ · u = 0,
∂
+ ∇ · (u) = ∇ · [M LBM ∇μLBM ],
∂t
∂u
ρ0 ( + u · ∇u) = ∇ · PLBM + ∇ · [ρ0 ν(∇u + ∇uT )],
∂t
ρ 2 − 2 (4.37)
M LBM = 0 (τ − 0.5),
ρ0

Δtc2 ρ0 + Δtc2 τ G σ σ Δt 3 c4 τ G σ σ
μLBM = ln − − Δ,
6 ρ0 − 6(τ − 0.5) 36(τ − 0.5)
1
PLBM = − p LBM I + κ LBM [(Δ + |∇|2 )I − ∇∇],
2

where κ LBM = Δt 2 c4 G σ σ /36 is positive as G σ σ is positive, M LBM ()9 is the mobil-
ity and μLBM ()10 is the chemical potential achieved in LBM simulations. The
momentum equation contains a third-order spatial derivative of , which is a more
reasonable description of the interface dynamics between two immiscible fluids,
compared to other mathematical models using only second-order spatial derivatives.
4.4 Upscaled Lattice Boltzmann Method
Flows at the Darcy scale can be simulated using modified LBM algorithms, where
LBM works as a unified framework for simulations at different scales, i.e., both pore
and Darcy scales. The resistance effect of the porous media to the fluid is usually
modeled by an external body force, which increases with the decrease of permeabil-
ity [8, 15]. Permeability is used as an input in the Darcy-scale simulations of flow
problems. We introduce the LBM algorithm proposed in [8], which converges to
the generalized Navier–Stokes equation described in [20] and uses the force model
proposed in [7]. We modify this algorithm by removing the convective term in the
momentum equation to obtain the Stokes, Darcy and Brinkman equations, and using
the Shan–Chen force model [24, 25] that is simpler than the original Guo et al. force
model [7]. The comparisons of the two force models and their numerical perfor-
mances in the Darcy scale simulation are given in [18].
9 By a slight abuse of notation, M LBM denotes the mobility in LBM and M denotes the total mass
of a system introduced in Sect. 2.7.
10 By a slight abuse of notation, μLBM denotes the chemical potential in LBM and μ denotes the
dynamic viscosity elsewhere.

4.4 Upscaled Lattice Boltzmann Method 133
Here, we only discuss single-phase single-component flows and the unknown

is f α (x, t) without notation σ . As a unified simulation framework, the evolution
equation of f α (x, t) is still Eq. (4.20) but with Ncomp = 1 here (i.e., Eq. (4.5)). We
eq
simplify Eq. (4.21) of f α by removing the nonlinear terms of ueq to recover the
Stokes, Darcy and Brinkman equations:
3
f αeq = ρωα (1 + eα · ueq ), (4.38)
c2

where ρ = α f α and ueq is defined according to the Shan–Chen model:

α eα f α + τ Δtρf m
u eq
= . (4.39)
ρ
The total effective body force f m per unit mass is:
φνDarcy φ Fφ
fm = − u − √ |u|u + φg, (4.40)
κ κ
where φ is the porosity, Fφ the coefficient of the nonlinear Forchheimer force, νDarcy
the fluid kinematic viscosity in the Darcy-scale simulation, κ and u are the local
permeability scalar and volumetric flow velocity,11 respectively. In general, we have
two kinematic viscosities, and νDarcy contained in the body force could be different
from ν = c2 Δt (τ − 0.5)/3, which is implicitly achieved by appropriately selecting
Δt, Δx, τ and has influence to the flow via the stress tensor. The flow velocity u is
defined as:

eα f α + 0.5Δtρf m
u= α . (4.41)
ρ
The explicit formula to compute u obtained by solving Eqs. (4.40), (4.41) is [8]:
v
u= ,
a0 + a02 + a1 |v|

v = (1/ρ) eα f α + 0.5Δtφρg ,
α (4.42)

φΔtνDarcy
a0 = 0.5 1 + ,
2κ
φΔt Fφ
a1 = √ .
2 κ
11 We denote the volumetric flow velocity of the Darcy-scale simulation by u, instead of U that is
used in Sect. 3.4.4, where u is used as the pore-scale flow velocity.
Then, ueq defined by Eq. (4.39) can be computed simply by:

u eq
= 2τ u + (1 − 2τ ) eα f α /ρ, (4.43)
α
which is √
obtained by solving Eqs. (4.39) and (4.41). In the incompressible limit of
|u| c/ 3, the computed p LBM = c2 ρ/3 and u satisfy the following equations:
∇ · u = 0,
−1 φνDarcy φ Fφ ∂u (4.44)
∇ p LBM + νΔu − u − √ |u|u + φg = ,
ρ0 κ κ ∂t
where the initial uniform density ρ0 could be different from the real density ρreal as
discussed at the end of Sect. 4.2. The parameters ν, νDarcy , φ, κ and Fφ can be set
independently and then the LBM solution at steady state approaches the solution of
the Stokes, Darcy, or Brinkman equation, respectively.
4.5 Simulation Results
4.5.1 Large Eddy Simulation of Turbulence
The time-evolutionary nature usually makes LBM more time-consuming than the
traditional CFD methods (e.g., finite difference/volume/element methods) in study-
ing steady state problems. In time-dependent dynamic problems (e.g., simulations of
turbulence and multiphase flow), however, LBM will become more efficient thanks to
using explicit updating algorithm for each physical time step. Large eddy simulation
(LES) based on the Smagorinsky model [23] can be conveniently incorporated into
the LBM [32] because the strain rate tensor S used to determine the eddy kinematic
viscosity can be calculated by the second-order moment of the non-equilibrium part
of density distribution function, which is obtained by the Chapman–Enskog expan-
sion, i.e., Eq. (4.16). In the Smagorinsky model, the eddy kinematic viscosity is
calculated as follows12 :

νeddy = (Cs Δx)2 2Si j Si j , (4.45)
where Cs is a constant. The correlation between dimensionless relaxation time and

kinematic viscosity in LBM will be changed to ν + νeddy = (τ + τeddy − 0.5)Δtc2 /3.
Similarly, τ in Eq. (4.16) should be replaced by τ + τeddy . Consequently, we obtain:
12 Similar idea can be used to model some kinds of non-Newtonian fluids.


τeddy = 0.5[ τ 2 + 18(Cs Δx)2 (ρc4 Δt 2 )−1 2Q i j Q i j − τ ], (4.46)

where Q i j = α eα,i eα, j f α(1) ≈ α eα,i eα, j ( f α − f α ). After getting τeddy by the
eq
above explicit algorithm, τ + τeddy will be used in the relaxation process to update
f α , which makes the current overall kinematic viscosity implemented in the LBM-
LES algorithm equal to the sum of the fluid kinematic viscosity ν and the current
eddy kinematic viscosity νeddy .
Additionally, by physical analysis of the momentum transport process, the strain
rate tensor can also be determined using the density distribution function f rel,α after
the relaxation process and its subsequent f α after the propagation process [16]. We
first rewrite the evolution equation (4.5) into:
f α (x + Δteα , t + Δt) = f rel,α (x, t),

eq (4.47)
f α (x, t) − f α (x, t)
f rel,α (x, t) = f α (x, t) + .
τ
Note that f rel,α (x, t) and f α (x, t) are the losses to and gains from the neighboring
grids of the concerned grid x, respectively. To compute the stress tensor, we need
both gains and losses. Accordingly, the transient distribution function during the
evolution process could be sorted as follows:
· · · → f rel,α (x, t − 2Δt) → f α (x, t − Δt) → f rel,α (x, t − Δt) → f α (x, t) → · · · ,
where each pair has the two values before and after a propagation process, i.e., the
first equation of Eq. (4.47). By contrast, the other sortation
f α (x, t − Δt) → f rel,α (x, t − Δt)
has the two values before and after a relaxation process, i.e., the second equation of
Eq. (4.47). Note that the stress tensor is essentially associated with the momentum
exchange through the transport process. Thus, we construct an effective distribution
function by using the mean value of the distribution function before and after the
propagation/transport process:
f αeff (x, t) = [ f rel,α (x, t − Δt) + f α (x, t)]/2, (4.48)
which is already known

when computing τ + τeddy to updatef rel,α (x, t). Then, we can
calculate ρ eff = α f αeff , p eff = c2 ρ eff /3, ueff = (1/ρ eff ) α eα f αeff , and the stress
tensor P as follows:

Pi j = − f αeff (eα,i − u ieff )(eα, j − u eff
j ), (4.49)
α
1,wall
0 1
Fig. 4.1 Schematic of 3D lid-driven cavity
which is similar to the original definition of using the continuous distribution function
of the Boltzmann equation, i.e., Eq. (2.43). Then, Si j can be inversely determined
from Pi j according to the inherent constitutive equation of LBM:
1
Si j = (Pi j + p eff δi j )
2νρ eff
(4.50)
3
= (Pi j + p eff δi j ),
2(τ − 0.5)c2 Δtρ eff
where ν and τ should be replaced by ν + νeddy and τ + τeddy in the LBM-LES

simulation, respectively, i.e., ν + νeddy = (τ + τeddy − 0.5)Δtc2 /3. Using Eq. (4.45)
of the Smagorinsky model again, we obtain the following alternative algorithm:

τeddy = 0.5[ (τ − 0.5)2 + 18(Cs Δx)2 (ρ eff c4 Δt 2 )−1 2Q ieffj Q ieffj − (τ − 0.5)],
(4.51)
where Q ieffj = Pi j + p eff δi j . After getting τeddy by the above explicit algorithm, τ +
τeddy will be used in the relaxation process to update f rel,α (x, t). This algorithm is
based on physical analysis rather than mathematical derivation of the previous one,
and its validity will be verified in the following simulations.
The three-dimensional lid-driven cavity flow of Fig. 4.1 at high Reynold (Re)
number is simulated by the two LBM-LES algorithms [16]. The size of cubic cavity is
L = 1 m, the wall boundary at x2 = L moves in the positive x1 direction at a speed of
1 m/s, and the fluid kinematic viscosity ν is 0.001 m2 /s making Re = 1000, at which
reliable literature results are available for validation. The computational domain
is divided into 513 uniform spatial grids (i.e., Δx = 0.02 m) and the time step is
Fig. 4.2 Steady state comparisons of Si j and τeddy on the symmetric plane of x3 = L/2 computed
by the traditional LBM-LES algorithm (solid line) and the alternative LBM-LES algorithm (dashed
line, which is overlapped by the solid line), Re = 1000, Cs = 0.17 [16]
Δt = 0.001155 s. We first compare Si j and τeddy computed by the two algorithms, i.e.,
Eqs. (4.16) and (4.46) versus Eqs. (4.50) and (4.51), respectively. Figure 4.2 shows
that the results of the two algorithms are almost identical and overlapping without
noticeable discrepancy at steady state. At the beginning, the signal propagated from
the moving lid is still small and thus the solution is sensitive to the numerical error
in the bottom area, where the two algorithms have small discrepancy that is omitted
here (e.g., at the 200th Δt [16]).
The alternative LBM-LES algorithm of Eq. (4.51) will be used below to present the
comparison of u with the traditional CFD simulations based on the incompressible N–
S equation. In the traditional CFD simulations of finite element method (FEM), both
uniform and non-uniform spatial grids are used and the non-uniform grid gives more
accurate results although the uniform grid leads to faster convergence speed [31].
Fig. 4.3 Steady state velocity profiles along the central axes on the symmetric plane of x3 = L/2
computed by the traditional FEM (solid line), and the alternative LBM-LES algorithm using Cs =
0.17 (dashed line) and Cs = 0.1 (dotted line), Re = 1000 [16]
Additionally, it uses 51 × 51 × 25 grids for simulating half computational domain

with the help of symmetric boundary condition. The velocity profiles of u 1 and u 2
along the central axes on the symmetric plane of x3 = L/2 are used to validate the
LBM-LES algorithm against the traditional FEM. As shown in Fig. 4.3, the steady
state velocity profiles obtained by the LBM-LES simulation of using Cs = 0.17 agree
well with those by the traditional FEM although the velocity magnitude of the former
is slightly smaller than that of the later around the local minimum and maximum.
In addition, this discrepancy increases when Cs decreases to 0.1, which differs from
the observation of [32], where Cs = 0.1 is found to yield better energy spectra than
Cs = 0.17 in the simulations of decaying homogeneous isotropic turbulence. Thus,
the influence of Cs on the numerical accuracy may depend on the specific problem
and also on which kind of results are considered.
It is worth pointing out that LBM is stable at high Re as shown here but becomes
eq
unstable at high Ma. According √to the construction of Eq. (4.1), f α of LBM will be
negative for some α if |u|/(c/ 3) is large (i.e., high local Ma), which is different
from the case of original definition of continuous f eq by Eq. (2.61) that is always
eq
positive. Negative f α will lead to negative f α through the relaxation process of
Eq. (4.5). Actually, Eq. (4.5) indicates that negative f α could occur before having
eq
negative f α when τ < 1. Once some f α become negative, this trend of generating
unphysical f α will be promoted by the relaxation process and eventually results into
negative infinite density ρ = α f α .
4.5.2 Intrinsic Permeability
The LBM has been widely used to compute the intrinsic/absolute permeability κ∞ of
porous media [5, 13, 19, 21], where it is implicitly assumed that the LBM at low Mach
number Ma is equivalent to the traditional CFD solvers based on the incompressible

N–S equation. However, high-order moments,13 which are completely neglected in
the N–S equation, are still present through the density distribution function in the
LBM simulations. To ensure that the LBM is accurately working at the N–S hydro-
dynamic level, the contributions of high-order moments have to be negligible. This
requires that the Knudsen number of LBM is small so that the undesired rarefaction
effect can be ignored, which has been commonly oversighted √ in simulating flows
inside porous media. In the LBM, we have K n LBM = π/6(τ − 0.5)/Nthroat [34],
where Nthroat is the number of spatial grids used to discretize the throat size in the
pore-scale simulations. Since it is difficult to determine Nthroat for complicated geom-
etry, the total voxel/grid number in the length direction will be used to replace Nthroat
in the following and the computed K n LBM is global Knudsen number.
In the LBM simulations, the model parameters can be correlated with the kine-
matic viscosity, i.e., ν = (τ − 0.5)Δtc2 /3. Therefore, we have flexibility in selecting
Δt, Δx = cΔt and τ for arbitrary ν. In the following examples, we will demonstrate
that for a given ν, the combination of Δt, Δx, τ always leads to finite K n LBM , which
may incur unintentional but noticeable rarefaction effect. This rarefaction effect was
interpreted as viscosity-dependent permeability [21], which could occur when ν as
well as K n LBM are changing with τ at constant Δt and Δx, for instance. But, the
interpretation of permeability change by variable ν is misleading because we can
easily find two simulations using the same τ and Δx (thus the same K n LBM ) but
different Δt (thus different ν) that will have the same permeability.
Actually, the variation of computed permeability κ is due to changing K n LBM and
thus can be mitigated by appropriately selecting τ and Δx 14 even with the single-
relaxation-time LBM. To validate this opinion, we first simulate the flow inside an
artificial two-dimensional porous media as shown in Fig. 4.4 and use small pressure
difference to drive the flow such that the Darcy law holds, namely the flow rate is
linearly proportional to the pressure gradient. Only the permeability along the driving
direction (i.e., a permeability scalar) will be discussed in Sect. 4.5.2. The geometry is
discretized by using N1 × N2 = 400 × 200 spatial grids in total (Δx1 = Δx2 = Δx).
The grid number is then increased to 800 × 400 and 1200 × 600, respectively, to
show the influence to permeability of Δx via K n LBM . Figure 4.5 shows that the
permeabilities obtained by using different Δx are very different even for the same τ
and geometry. This large difference is not due to spatial resolution Δx itself as in the
traditional CFD simulation based on the N–S equation, but due to different K n LBM .
Only when τ − 0.5 is close to zero (i.e., K n LBM → 0), they all approach the same
intrinsic permeability as expected although their Δx are still very different.
To further understand the influences of Δx and τ , we replot the permeability
curves against K n LBM and then find that the three curves of LBM collapse into a
single line especially at small K n LBM as shown in Fig. 4.6. It clearly shows that
the choice of τ, Δx should ensure K n LBM → 0 in computing the intrinsic perme-
13 To recover the incompressible N–S equation from the LBM, we only need the zero- and first-order
moments of Eq. (4.9).

14 Changing Δt alone at fixed τ and Δx doesn’t affect the computed permeability.
0.5
0 1 1
Fig. 4.4 Schematic of 2D pore-scale simulation, where the symmetric boundary condition is used
at the top and bottom boundaries while the pressure difference is applied between the left and right
boundaries [19]
ability, e.g., either by using small τ − 0.5 or small Δx (large grid/voxel number).
This underlying requirement of K n LBM → 0 should also be satisfied in the multi-
relaxation-time LBM simulation that though has a wider choice of model parameters.
This is especially important in simulating flows inside tight porous media, where the
voxel number used to discretize pore throat is usually small to avoid high computa-
tional cost and thus the feasible approach is restricted to using small τ − 0.5.
In Fig. 4.6, the DVM simulation with very fine molecular velocity grid of 1600
points (i.e., DVM D2Q1600) is also used to get a reference solution for the accu-
rate permeability variation with Knudsen number. As expected, the DVM D2Q1600
agrees well with the LBM D2Q9 in computing the intrinsic permeability while sig-
nificant deviation of the LBM data (green curve) from the DVM data (black curve)
can be seen when the global Knudsen number is larger than 10−2 (the corresponding
local Knudsen number defined by using the dominant pore throat size will be about
10−1 ). The ordinary LBM will become inaccurate even at smaller Knudsen number if
the pore-scale solutions (e.g., velocity distribution) instead of average property (e.g.,
permeability) are considered. Noted that in Fig. 4.6 both the LBM and the DVM used
the fully diffuse boundary condition on the wall surface, instead of the bounce-back
boundary condition, to facilitate their comparison.
We now use the LBM D3Q19 to simulate the force-driven flow in a three-
dimensional porous media of Fig. 4.7, as another example to study the intrinsic
permeability. Here, Δt, Δx and grid number of 1003 are fixed. The computed per-
meability will change with τ and is plotted against K n in Fig. 4.8, which is similar
to Fig. 4.6 of the previous two-dimensional simulations. Three different force mag-
nitudes |g|a > |g|b > |g|c are used to drive the flow and the obtained permeabilities
are the same except for small K n, where larger |g| leads to smaller permeability. This
Fig. 4.5 Computed permeability κ by using different τ and Δx(N1 ) in the LBM simulations [19]
Fig. 4.6 Computed permeability κ by both LBM and DVM versus K n [19]
Fig. 4.7 Schematic of 3D

pore-scale simulation, where
external body force is
applied to drive the flow in
the x1 direction and the
periodic boundary condition
is used at all boundaries [19]
is due to the increase of inertial effect as the kinematic viscosity decreases with K n
via τ at fixed Δt and Δx. The permeability computed with smaller |g| will also drop
if we keep on decreasing K n → 0 to make Re large enough. Therefore, in addition
to K n, the inertial force (or Re in general) has to be small as well in computing the
intrinsic permeability. The easiest way to avoid inertial effect is to solve the Stokes
equation without the convection term by removing the nonlinear terms of flow veloc-
eq
ity in f α as in Eq. (4.38), which makes the computed permeability independent of
the external body force (or the pressure gradient in the previous case).
4.5.3 Buoyant Flow
The buoyant flow of two immiscible fluids with different mass densities is simu-
lated here by using the Shan–Chen two-phase single-component model introduced
in Sect. 4.3.1. This problem is supposed to be modeled by the Shan–Chen model for
two immiscible components introduced in Sect. 4.3.2, which is usually valid only
for problems, where the density ratio is immaterial. In the simulation of Shan–Chen
two-phase single-component model, which will be simply referred to as the Shan–
Chen model in Sect. 4.5.3, we can take the two phases as two immiscible fluids as
long as the phase change is negligible during the flow process. This can be achieved
by making the pressure itself much larger than the pressure variation associated with
the gravity (i.e., ρg2 L 2 ). The following reduced Peng–Robinson equation of state
(EOS) will be used:
| |
| | <| |
∞
| | <| |
Fig. 4.8 Computed permeability κ by LBM using different external body forces versus K n [19]
3.2533ρr Tr 4.839ρr2 y(Tr )

pr = − ,
1 − 0.25307ρr 1 + 0.50614ρr − 0.064044ρr2 (4.52)

y(Tr ) = [1 + (0.37464 + 1.54226ω − 0.26992ω ) × (1 − 2
Tr )] ,
2
where Tr = T /Tc and Tc is the critical temperature, for instance, and ω is the
acentric factor. When the equation of state is specified to inversely determine
ψ(ρ) as discussed after Eq. (4.33), the negative coefficient G in the interaction
force model has no influence to the IFT anymore due to its dual usages in com-
puting ψ(ρ) and then F(ψ). Nevertheless, Δt and c contained in κ LBM are still
adjustable to match the IFT. Additionally, when the above reduced EOS is applied,
we have PLBM = − pI + ς LBM as the stress tensor at static state (if neglecting
1
2ρνS of the spurious velocity), ς LBM = κ LBM [(ψΔψ + |∇ψ|2 )I − ∇ψ∇ψ] and
2
ψ 2 = ( p − c2 ρ/3)/(0.5Gc2 /3). Since the IFT can be calculated by an integral of
the stress tensor across the interface (IFT ∝ |PLBM |) [25], we propose the following
steps for selecting the model parameters to match ρa , ρb , ν = νa = νb and IFT:
1. determine the spatial grid number (fix Δx), τ will be automatically determined
from the given ν and the parameter c that is selected below.
2. select appropriate pc and c to make the pressure much larger than its variation
associated with the gravity, and then adjust pc to make p(ρ) smaller but close to
c2 ρ/3 particularly in the area of light phase (see Fig. 4.9) such that the interface
will be thin as desired (see Fig. 4.11a).
Fig. 4.9 Equation of state used in the LBM simulation
3. select Tr to match the density ratio ρa /ρb = ρa,r /ρb,r .

4. adjust ρc and pc together, by keeping pc /ρc constant to maintain the distribution
pattern of ψ(x), to match the specific values of ρa and ρb , in which we have
IFT ∝ ρc . The initial density distribution should also be adjusted accordingly
when testing with an initial two-phase distribution (e.g., a bubble or droplet).
5. adjust c and pc together, by keeping pc /c2 constant to maintain the distribu-
tion pattern of ψ(x) that is actually unchanged in this adjustment, to match the
IFT, in which we roughly have IFT ∝ c2 as κ LBM ∝ c2 at fixed Δx. The IFT
will not precisely change as predicted, which is probably because the corre-
sponding adjustment of small contribution of spurious velocity is not concerted
(i.e., |2ρνS| ∝ |u| ∝ c = c2 at fixed ρ and ν, while in the previous step we have
|2ρνS| ∝ ρ ∝ ρc as desired).
As shown in Fig. 4.10, the light phase will rise in the column of heavy phase
due to the gravity, which has been simulated by the traditional CFD methods, i.e.,
the Eulerian level-set finite element method and the arbitrary Lagrangian–Eulerian
moving grid method [12]. The dimensions of the column is 1 × 2 and the initial
radius of circular bubble is r0 = 0.25 centered at (x1 , x2 ) = (0.5, 0.5). The non-slip
boundary condition is used at the top and bottom boundaries, whereas the free slip
condition is imposed at the left and right boundaries. Among the two tested cases,
we select the first one with ρa = 1000, ρb = 100, νa = νb = 0.01, g = (0, −0.98),
IFT = 24.5.
Fig. 4.10 Schematic of the buoyant flow, g = (0, −0.98), νa = νb = 0.01 and IFT = 24.5
(a) (b)
Fig. 4.11 Initial distributions of p(x2 ) and ρ(x2 ) along the vertical centreline of x1 = 0.5, Δp0 ≈
IFT/r0 for two-dimensional cases (a) and the representative transient locations of interface (b)
In the LBM simulation, the top and bottom boundaries are modeled as solid wall
while the left and right boundaries are modeled as periodic boundaries. We use
200 × 400 spatial grids in total for the 2D computational domain, Δx = 1/200, c =
Δx/Δt = 12, g = (0, −0.98), ω = 0.344, Tr = 0.900142166, ρc = 450.732475,
pc = 6602.80792, τ (νa ) = τ (νb ), and ρsolid = 900 that has little influence to the
dynamic process before the bubble touching the top boundary. The densities of the
two phases are ρa,0 = 1000, ρb,0 = 100 at the initial state and will be changed slightly
during the simulation process. The initial density distribution is specified as:

ρa,0 + ρb,0 ρa,0 − ρb,0 d − r0 d − r0
ρ(x, 0) = + tanh 20 × + ,
2 2 r0 d (4.53)

d = (x1 − 0.5)2 + (x2 − 0.5)2 .
The top and bottom boundaries are set as periodic for the arbitrary interface specified
by Eq. (4.53) to equilibrate during the initial 40,000 Δt (could be much shorter to save
the computational cost) and then changed to the solid boundary condition at t = 0,
which is at the end of the equilibration process. The external body force is activated at
t = 0. The transient interface locations at the subsequent 2400th Δt, 4800th Δt and
7200th Δt (corresponding to the physical moments of t = 1, 2 and 3 s, respectively)
are given in Fig. 4.11b to show the dynamic evolution, where the transient location at
t = 3 s of the LBM simulation agrees well with that of the traditional CFD methods
reported in [12]. Additionally, the transient streamlines around the interface and the
profiles of u 2 (x2 ) along the vertical centreline of x1 = 0.5 are given in Figs. 4.12,
4.13 and 4.14, where u 2 = 0 at the top and bottom boundaries as expected.
Fig. 4.12 Transient streamline around the interface (left) and velocity profile of u 2 (x2 ) along the
vertical centreline of x1 = 0.5 (right) at t = 1 s
4.5.4 Two-Phase Flow Inside Real Digital Rock
To be consistent with the prevailing convention, we refer to the flow system com-
posed of immiscible chemical components (e.g., oil-water system) as multiphase
flow although they are all liquid phase. We study the displacement/recovery pro-
cess by using the Shan–Chen model for two immiscible components introduced in
Sect. 4.3.2. The injecting speed used is very low (low Capillary (Ca) number in
general) to reflect the in-situ reservoir condition although this will make slow the
convergence process that depends on the interface propagation speed of the two
phases, instead of the propagation speed of pressure as in the simulation of single
phase flow. Thanks to low Ca, the wettability will be the dominant mechanism for
the pore-scale evolution of phase distribution and can be conveniently matched in
the LBM simulations as shown in Table 4.1. The density and viscosity ratios are set
to be about one for simplicity since they are immaterial now. Viscosity ratio different
from one can be achieved by using different τ for different phases if needed.
In the simulation of two immiscible phases 1 and 2, we will have subdomains
dominated by the phase 1 (e.g., droplets of phase 1), where the density ρl1 of phase
1 is much larger than the density ρg2 (solubility due to weak miscibility of the LBM
model) of phase 2, similar for the subdomains dominated by the phase 2. In the
test cases for parameter selection, a 2D computational domain is discretized by
1000 × 500 spatial grids with Δx = 10−6 , the top and bottom boundaries are solid
while the left and right boundaries are periodic. We initially place a droplet of the
phase 1 on the bottom and extract IFT and the contact angle θeq 1
occupied by the
phase 1 at the equilibrium state (e.g., the static state in the absence of external force).
We set ρl1 = 500 and ρg2 = 2 inside the droplet of phase 1 and ρg1 = 2 and ρl2 = 500
outside the droplet of phase 1. The interaction parameter is G = 0.003, effective
wall density ρsolid
1
= 5 and ρsolid
2
= 100 (thus the phase 2 is the wetting phase), the
lattice speed c = 1000, relaxation time τ 1 = 0.9 and τ 2 = 1.0. At the equilibrium
state, the droplet length L 1 at the bottom boundary, the droplet height L 2 , and the
pressures inside and outside the droplet can be accurately determined, although the
direct geometric determination of contact angle might be inaccurate. Then, the droplet
radius r and the contact angle θeq1
can be conveniently determined as follows [14]:
4L 22 + L 21
r= (4.54)
8L 2
and
⎧
⎪
⎪ L 1 /2
⎨arctan , if r > L 2 for 0 < θeq
1
< π/2,
θeq
1
= r − L 2 (4.55)
⎪
⎪ L 1 /2
⎩π + arctan , otherwise.
r − L2
Table 4.1 Calibration of LBM parameters according to the fluid properties IFT and θeq
1 , the LBM
parameters ρl = 500, ρg = 2 and G = 0.003 are fixed here; the symbol ... means the same as in
the first case
Case 1
ρsolid 2
ρsolid τ1 τ2 Δx c IFT 1
θeq
1st 2D droplet of phase 1 5 100 0.9 1.0 10−6 1000 22.87 π − 66.92
2nd Without wall B.C. ... ... ... ... ... ... 24.35 None
3rd 2D droplet of phase 2 ... ... ... ... ... ... 23.62 π − 68.51
4th 2D droplet of phase 1 100 5 ... ... ... ... 22.26 64.91
5th 2D droplet of phase 1 ... ... ... ... 2 × 10−6 ... ∝ Δx ...
6th 2D droplet of phase 1 ... ... ... ... ... 200 ∝ c2 ...
7th 2D droplet of phase 1 ... ... ... 0.9 ... ... 24.72 π − 66.46
8th 2D droplet of phase 2 ... ... ... 0.9 ... ... 24.18 π − 67.87
9th 3D droplet of phase 1 ... ... ... ... ... ... 28.74 π − 67.14
10th 2D droplet of phase 1 ... 130 ... ... ... ... 22.96 π − 60.28
Then, IFT can be determined according to the Young–Laplace equation, i.e., Δp2D =
IFT/r or Δp3D = 2IFT/r . When the same LBM parameters are used, the obtained
IFT and θeq 1
are almost the same for both 2D and 3D test cases as expected. Different
LBM parameters are used to show their influences to IFT and θeq 1
in Table 4.1, where
the influences of Δx and c to IFT can be understood by the Young–Laplace equation.
Then, the pore-scale displacement of the phase 2 by injecting the phase 1 is
modeled by the LBM simulation and Fortran MPI parallel computation is employed
to handle a real digital berea sample that has 5003 voxels with Δx = 2.5 microns as
shown in Fig. 4.15. Other LBM parameters except Δx are the same as in the first case
with θeq 1
= π − 66.92 of Table 4.1 and thus we have IFT = 22.87 × 2.5 = 57.17. As
discussed at the end of Sect. 4.5.1, the simulation will become unstable at large Ma,
which could happen with abrupt increase of flow speed sometimes due to interface
coalescence or pinch-off, or sudden breakthrough of the injected non-wetting phase
at the pore throat. To reduce Ma, we use very low injecting speed, which is also
needed to satisfy the in-situ reservoir condition of low Ca. The average flow speed
of the injected phase 1 is about 0.5754 at the 400,000th Δt (the saturation S1 of
phase 1 is about 0.03581) and decreases to about 0.2891 at the 6,000,000th Δt (S1
is about 0.8063 and the residual S2 = 1 − S1 is about 0.1937). Thus, we have Ca =
μ2 u 3 1 /IFT < 0.2083 × 0.5754/57.17 = 2.096 × 10−3 . Note that the displacement
pattern inside a capillary tube is almost independent of Ca at θeq
1
= π − 68 ± 5 when
−3
Ca < 6 × 10 [35]. Thus, the results will be almost unchanged when using lower
injecting speed, which increases the computational cost and can be avoided here.
The simulation takes about 6 millions of Δt due to low injecting speed and the sam-
pling of relative permeability and the corresponding saturation starts at the 400,000th
Δt with a fixed interval of 100,000 Δt between each two consecutive samplings. As
shown in Fig. 4.16, the permeability curves of two-phase simulation are below the
Fig. 4.15 A berea digital rock for the LBM two-phase flow simulations
single-phase absolute permeability as expected for usual cases. When decreasing the
wettability of phase 2 from about 67 to 80 degree, its permeability is enhanced but
the permeability change of phase 1 (non-wetting here) is negligible. Meanwhile, the
breakthrough saturation of phase 1 is increased, i.e., need injecting more phase 1 to
break through at the outlet, after which the phase 1 is connected from the inlet to
outlet and thus its permeability becomes nonzero (see marks 1 and 2). Additionally,
when the wettability of phase 2 becomes more neutral from about 67 to 80 degree,
the condition with equal permeability for both phases shifts towards the middle as
expected (see marks 3 and 4). Also, the condition with zero permeability of the phase
2 is shifted leftward (delayed), which means that more phase 2 can be easily pro-
duced at nonzero permeability (see marks 5 and 6). Although the phase 2 can still
be produced after having zero permeability (due to disconnected distribution) by
moving together with the injected phase 1, this production of almost confined phase
2 is very slow, which can be seen from the decreasing distance along S2 between
every two consecutive samplings.
Fig. 4.16 The influence of wettability to the relative permeability and recovery, blue curves: θeq
1 =
π − 66.92, green curves: θeq = π − 80.07

1
4.5.5 Upscaled Simulations
For many practical cases, the number of fine discretization points in the whole com-
putational domain is very large due to heterogeneity, making the memory usage
and computational time unaffordable. We use an upscaled simulation scheme [18]
to reduce the number of points by using a coarse grid with an effective permeabil-
ity distribution κ ∗ (x), which is obtained by solving many parallel local problems
on the fine grid but inside much smaller subdomain for each local problem. The
upscaled effective permeability is a tensor even though the original permeability
input is assumed to be a scalar, e.g., κ(x). With this approach we are able to capture
the average information of fine grid on the coarse grid in solving the flow equation at
any intermediate scale and transfer the information between different scales. In terms
of computational cost, the total number of unknowns will not be reduced when using
the coarse grid to replace the fine grid in general cases since all unknowns defined
on the original fine grid for a single simulation will need to be updated in the parallel
local simulations. However, the number of time steps (iteration times) needed to
converge will be significantly reduced in the decomposed simulations compared to
the fine-grid simulation, which eventually improves the overall efficiency by orders
of magnitude. A general overview of multiscale simulation methods by using coarse
grid is detailed in [4].
Fig. 4.17 Schematic of the fine and coarse grids in the upscaled LBM simulation [18]
In the upscaled simulation, the computational domain is divided into many sub-
domains and each subdomain is represented by a coarse point (see Fig. 4.17). This
substantially reduces the degrees of freedom in the coarse-grid simulation. To com-
pute the effective κ ∗ for each subdomain, we impose different external forces gconst
to drive flows in different directions in the local LBM simulations, which use a fine
grid located inside the corresponding subdomain and the distribution of κ(x) on the
fine grid. Then, the similar local LBM simulations usually need to be run with a
constant tensor κ ∗ as shown in Eq. (4.57). We seek κ ∗ such that the average veloci-
ties from local fine-grid simulations with the heterogeneous κ(x) and homogeneous
κ ∗ , respectively, are equal (see Eqs. (4.59), (4.60)). However, the onerous iterative
process by adjusting the unknown κ ∗ to match the fluxes computed using κ(x) is
avoided in our simulations since κ ∗ can be computed explicitly by Eq. (4.56).
We discuss two-dimensional problems as example. In the local LBM simulations
using κ(x), we drive flow in the x1 direction by g(1) const = (gconst , 0) and compute
the average velocity u(1)κ(x) , where · is defined as a volume average over the subdo-
(2) (2)
main concerned. We also compute uκ(x) by using gconst = (0, gconst ) in another local
simulation. Then, κ ∗ is computed as follows:
∗ (1)
∗ κ , ∗
κ12 νDarcy uκ(x) · (1, 0), u(2)
κ(x) · (1, 0)
κ = 11
∗ ∗ = . (4.56)
κ21 , κ22 gconst u(1)
κ(x) · (0, 1), u(2)
κ(x) · (0, 1)
Now, we validate that the computed κ ∗ satisfies the conservation principle of

average fluxes. Assuming that we run local LBM simulations using the constant
tensor κ ∗ computed by Eqs. (4.56), (4.40) is modified to be:
f m = −φνDarcy κ ∗ −1 · u + φg, (4.57)

where κ ∗ −1 is the inverse matrix of κ ∗ and the nonlinear Forchheimer term is

neglected to simplify the following analysis. The evolution of f α (x, t) is described
by Eqs. (4.5), (4.38), (4.39), (4.41) and (4.57). As κ ∗ and g are constant and
the periodic boundary conditions are used in the local simulations, the relation
f α (x + Δteα , t + Δt) = f α (x, t) holds at steady state. For arbitrary Δx, Δt, τ , φ,
νDarcy , κ ∗ and g = gconst , the steady state solution of f α is independent of x and equal
to:

3eα κ ∗ · gconst
f α = ρ0 ωα 1 + 2 · , (4.58)
c νDarcy

which implies that the uniform density is ρ = α f α ≡ ρ0 . This solution is inferred
from the understanding of force balance between the external body force g and the
resistance νDarcy κ ∗ −1 · u, i.e., f m ≡ 0. We validate the solution of Eq. (4.58) by the
following verification: substituting Eq. (4.58) into Eq. (4.41) and considering Eq.
(4.57), we get the uniform velocity u ≡ (κ ∗ · gconst )/νDarcy , and f m ≡ 0 by using u
in Eq. (4.57). Then, substituting f α and f m into Eq. (4.39), we get ueq = u, which
implies that the evolution equation (4.5) is satisfied at steady state since we have
eq
f α = f α according to Eqs. (4.38) and (4.58), and f α (x + Δteα , t + Δt) = f α (x, t).
According to the uniform solution of u ≡ (κ ∗ · gconst )/νDarcy and Eq. (4.56), the
average velocity u(1) ∗ (1)
κ ∗ of the local simulation using constant κ and gconst = (gconst , 0)
satisfies:
κ ∗ · (gconst , 0)
u(1)
κ∗ = = u(1)
κ(x) , (4.59)
νDarcy
which implies that the average flux is conserved when using the same external force
g(1)
const but different permeability distributions, namely using the heterogeneous κ(x)
and homogeneous κ ∗ , respectively. When driving flow by g(2) const = (0, gconst ), the
average flux is also conserved:
κ ∗ · (0, gconst )
u(2)
κ∗ = = u(2)
κ(x) . (4.60)
νDarcy
After getting the distribution of κ ∗ (x) of each coarse point on the coarse grid, we
implement two LBM simulations on the coarse and fine grids, respectively, inside
the whole computational domain. The parameters κ ∗ (x), Δxcoarse and Δtcoarse in the
coarse-grid simulation are different from κ(x), Δxfine and Δtfine , respectively. The
boundary conditions and the parameters ρ0 , ν, φ, νDarcy and g(x) in the coarse-grid
simulation are the same as in the fine-grid simulation. In order to clearly verify the
validity of the coarse-grid simulation of the whole computational domain, we use
periodic boundary conditions to eliminate potential numerical error, which occurs
when using fixed pressures, for example, at the two ends along the x1 direction
because fixed quantities are numerically imposed at the initial and last points along
the x1 direction and their spatial positions are changed when using different Δx.
In the following, we first show the validation of the computed effective permeabil-
ity in several benchmark problems, where analytical solutions are known. Then, we
implement upscaled LBM simulation using the computed effective permeabilities
on the coarse grid and present its validation against the fine-grid simulation.
4.5.5.1 Validation of Effective Permeability by Analytical Solutions
In the LBM simulations, the number of grid points is 100 × 100 and Δx = 0.01 m,
Δt = 0.0001 s, νDarcy = 2 × 10−6 m2 s−1 , ρ0 = 1000 kg m−3 , φ = 0.8. The periodic
boundary conditions are used. The average pressure over the computational domain
is subtracted from the computed pressure p = c2 ρ/3 in all figures of the pressure
distributions of Sect. 4.5.5. We run simulations in the computational domain with
prescribed distributions of κ(x) and verify the computed effective permeabilities κ ∗
against the analytical solutions.
We first study the problem with a layered distribution of κ(x), where the compu-
tational domain is uniformly divided into 10 layers parallel to the x2 axis. The odd
number layers have κa = 10−12 m2 and κb in the even number layers is constant with
values shown in Table 4.2 for different cases. Flow is driven in the x1 direction by a
uniform gconst = (2, 0) m s−2 . We set τ = 0.53 and so ν = 0.01 m2 s−1 . The results
∗
in Table 4.2 show that the computed κ11 by LBM agrees exactly with the analytical
solution although the analytical formula is derived from the Darcy equation. This
is because the steady state velocity is uniform and so the LBM simulations based
on the Brinkman equation with nonzero ν actually yield the solutions of the Darcy
equation at steady state.
When driving flow in the x2 direction by setting gconst = (0, 2) m s−2 , the velocity
distribution along the x1 direction is nonuniform. We set τ = 0.5 such that ν = 0 to
recover the Darcy equation. Note that the LBM simulation of using ν = 0 via τ = 0.5
is stable here because the external body force can be balanced by the resistance of
∗
permeability term for a small Ma. As we can see in Table 4.3, the computed κ22 by
LBM simulations agrees exactly with the analytical solution.
∗ , κ = 10−12 m2 and ν = 0.01 m2 s−1 [18]

Table 4.2 Validation of computed κ11 a
1 1 1 ∗ by LBM
κb /κa [ ( + )]−1 κ11
2 κa κb
2 1.33333 × 10−12 1.33333 × 10−12
10 1.81818 × 10−12 1.81818 × 10−12
50 1.96078 × 10−12 1.96078 × 10−12
100 1.98019 × 10 −12 1.98019 × 10−12
1000 1.99800 × 10 −12 1.99800 × 10−12
10000 1.99980 × 10 −12 1.99979 × 10−12
100000 1.99998 × 10 −12 1.99998 × 10−12
∗ , κ = 10−12 m2 and ν = 0 for the Darcy flow [18]

κb /κa (κa + κb )/2 ∗ by LBM
κ22
2 1.500000 × 10−12 1.499999 × 10−12
10 5.500000 × 10−12 5.499999 × 10−12
50 25.50000 × 10−12 25.49999 × 10−12
100 50.50000 × 10−12 50.49999 × 10−12
1000 500.5000 × 10−12 500.4999 × 10−12
10000 5000.500 × 10−12 5000.499 × 10−12
100000 50000.50 × 10−12 50000.49 × 10−12
∗ , κ = 10−12 m2 and ν = 0 for the Darcy flow [18]

√ ∗ by LBM ∗ by LBM (10002 points)
κb /κa κa κb κ11 κ11
2 1.41421 × 10−12 1.41418 × 10−12
10 3.16227 × 10−12 3.14081 × 10−12
50 7.07106 × 10−12 6.45938 × 10−12 7.01357 × 10−12
100 10.0000 × 10−12 8.25393 × 10−12 9.70489 × 10−12
1000 31.6227 × 10−12 12.2496 × 10−12 19.8897 × 10−12
10000 100.000 × 10−12 13.0133 × 10−12 23.2777 × 10−12
Now, we study the checkerboard problem, where the computational domain is

uniformly divided into 10 × 10 squares with each square containing 10 × 10 points.
The black squares of the checkerboard have κa = 10−12 m2 and κb in the white squares
takes different values for different cases as shown in Table 4.4. Flow is driven by
a uniform gconst = (2, 0) m s−2 and we set ν = 0 to get the solution of the Darcy
equation. The representative distributions of p, u 1 and u 2 are given in Fig. 4.18. The
∗
results in Table 4.4 show that the computed κ11 by LBM simulations agrees well with
the analytical solution when κb /κa is not very large but deviates significantly in the
case of high contrast. This deviation is due to the low spatial resolution of the grid
used in the LBM simulations at high contrast of permeability. We refine the grid by
increasing the total point number from 100 × 100 to 1000 × 1000 to show accuracy
improvement. Δx and Δt are changed to 10−3 m and 10−5 s, respectively. The results
∗
given in Table 4.4 show that the computed κ11 becomes very close to the analytical
solution when the permeability ratio is up to 100 but still significantly deviates from
the correct value if the permeability ratio is very high, where more spatial points are
required for higher resolution.
4.5.5.2 Upscaled Simulations of Darcy Flows
We choose a two-dimensional 1 m×1 m computational domain with periodic bound-

ary conditions and φ = 0.8, ρ0 = 1000 kg m−3 , νDarcy = 2 × 10−6 m2 s−1 . We set
Fig. 4.18 Distributions of p, u 1 and u 2 ; ν = 0 for the Darcy flow, gconst = (2, 0) m s−2 , κa = 10−12
m2 and κb /κa = 2 [18]
ν = 0 via τ = 0.5 in the simulations of both fine and coarse grids and also in the
calculation of κ ∗ . In order to have obvious variations in the results of the coarse-grid
simulation, a nonuniform external force g = (sin π x1 , sin π x2 ) m s−2 is used and the
distribution of permeability κ(x) in Fig. 4.19 is set according to Eq. (4.61) such that
the distribution of κ ∗ (x) is nonuniform as well:
⎧
⎪
⎪ κconst , 0.45 ≤ x1 , x2 ≤ 0.55
⎪
⎪
⎪
⎪ κconst , 0.2 ≤ x1 , x2 ≤ 0.3
⎪
⎪
⎨κ
const , 0.7 ≤ x1 , x2 ≤ 0.8
κ=
⎪
⎪ κ const , 0.2 ≤ x1 ≤ 0.3, 0.7 ≤ x2 ≤ 0.8
⎪
⎪
⎪
⎪ κconst , 0.7 ≤ x1 ≤ 0.8, 0.2 ≤ x2 ≤ 0.3
⎪
⎪
⎩
10[1 + sin(80x1 π ) cos(80x2 π )]κconst , elsewhere,
(4.61)
where κconst = 10−13 m2 . Additionally, we have Δxfine = 0.0025 m and Δtfine =

0.000025 s in the fine-grid simulation. The number of fine points is 400×400 inside
the whole computational domain, which is divided uniformly into 40×40 subdo-
mains. The average results over each set of 10 × 10 fine points located inside the
same subdomain are computed and used to verify the results of the coarse-grid sim-
ulation.
∗ ∗ ∗ ∗
We have κ11 = κ22 = κconst and κ21 = κ12 = 0 inside the subdomains with
κ ≡ κconst . For subdomains with κ = 10[1 + sin(80x1 π ) cos(80x2 π )]κconst , we use
∗ ∗
gconst = (2, 0) m s−2 to drive flow and get κ11 = 8.485κconst and κ21 = 0. The sym-
∗ ∗ ∗ ∗
metric property of κ(x) inside the subdomain implies that κ22 = κ11 and κ12 = κ21 .
∗ ∗
We define a scalar κ as the average value over all diagonal components of κ and
for all subdomains we have κ ∗ = κ ∗ I, where I is the identity tensor. Now, the gen-
eral formula Eq. (4.57) in the coarse-grid simulation can be replaced by Eq. (4.40),
where we change κ of the fine-grid simulation to κ ∗ for the coarse-grid simulation.
In the case of κ ∗ ≡ κ ∗ I, the algorithm in the coarse-grid simulation is the same as
in the fine-grid simulation but they use different scalar permeability distributions,
namely κ ∗ and κ, respectively. In the coarse-grid simulation, Δxcoarse = 0.025 m and
Fig. 4.19 Distribution of the permeability κ(x) on the fine grid [18]
Δtcoarse = 0.00025 s. The number of coarse points is 40 × 40 and the value of κ ∗

assigned to each coarse point with index (I1 , I2 ) is:
⎧
⎪κconst ,
⎪
⎪
19 ≤ I1 , I2 ≤ 22
⎪
⎪
⎪
⎪ κconst , 9 ≤ I1 , I2 ≤ 12
⎪
⎨κ
const , 29 ≤ I1 , I2 ≤ 32
κ∗ = (4.62)
⎪
⎪ κconst , 9 ≤ I1 ≤ 12, 29 ≤ I2 ≤ 32
⎪
⎪
⎪
⎪ κconst , 29 ≤ I1 ≤ 32, 9 ≤ I2 ≤ 12
⎪
⎪
⎩
8.485κconst , otherwise.
The distributions of the fine and coarse grids inside a representative area are given
in Fig. 4.19. Figures 4.20, 4.21 show that the agreement is very good between the
two simulations using the fine and coarse grids, respectively.
4.5.5.3 Upscaled Simulations of Brinkman Flows
The physical problem studied here is similar to that described in Sect. 4.5.5.2. The
differences are that we increase κconst from 10−13 to 10−7 m2 and set ν = 10−5 m2
s−1 such that the contribution by the viscosity term νΔu is noticeable as shown
in Fig. 4.22, which shows the comparison between the average results of two fine-
grid simulations using ν = 10−5 and 0 m2 s−1 , respectively. In this problem, flow
in some regions is close to the Stokes flow while in other regions, flow is close to
2 2 2
1 1 1
1 1 1
2 2 2
1 1 1
2 2 2
2 2 2
1 1 1
Fig. 4.20 Comparisons of p, u 1 and u 2 between the fine-grid results (left), fine-grid average results
(middle) and coarse-grid results using κ ∗ (right); ν = 0 for the Darcy flow, g = (sin π x1 , sin π x2 )
m s−2 , κconst = 10−13 m2 [18]
Fig. 4.21 Detailed comparisons of p, u 1 and u 2 between the fine-grid average results (red) and
coarse-grid results using κ ∗ (black); ν = 0, g = (sin π x1 , sin π x2 ) m s−2 , κconst = 10−13 m2 [18]
the Darcy flow. Since ν is nonzero here, we let ν = 10−5 m2 s−1 when comput-
∗ ∗ ∗ ∗
ing κ ∗ and get κ11 (κ, ν) = κ22 (κ, ν) = 7.367κconst and κ21 (κ, ν) = κ12 (κ, ν) = 0 for
subdomains with κ = 10[1 + sin(80x1 π ) cos(80x2 π )]κconst . For subdomains with
∗ ∗ ∗ ∗
κ = κconst , κ11 (κ, ν) = κ22 (κ, ν) = κconst and κ21 (κ, ν) = κ12 (κ, ν) = 0. Thus, we
∗
have κ (κ, ν) equal to 7.367κconst I or κconst I.
Fig. 4.22 Comparisons of p, u 1 and u 2 between two fine-grid simulations using ν = 10−5 (black)
and 0 (red) m2 s−1 , respectively; g = (sin π x1 , sin π x2 ) m s−2 , κconst = 10−7 m2 [18]
Fig. 4.23 Comparisons of p, u 1 and u 2 between the fine-grid average results (left), coarse-grid
results using κ ∗ (κ, ν) (middle) and κ ∗,err (κ) (right); ν = 10−5 m2 s−1 , g = (sin π x1 , sin π x2 )
m s−2 , κconst = 10−7 m2 [18]
As discussed in Sect. 4.5.5.2, we can use a scalar distribution of κ ∗ (κ, ν), which is
equal to 7.367κconst or κconst , in the coarse-grid simulation. We use Δxfine = 0.0025
m, Δtfine = 0.000025 s and τfine = 0.50012 in the fine-grid simulation, and use
Δxcoarse = 0.025 m, Δtcoarse = 0.00025 s and τcoarse = 0.500012 in the coarse-grid
simulation. Figures 4.23 and 4.24 show that the agreement of the coarse-grid simula-
tion using κ ∗ (κ, ν) with the fine-grid simulation is very good. In addition, we set ν =
Fig. 4.24 Detailed comparisons of p, u 1 and u 2 between the fine-grid average results (red) and
coarse-grid results using κ ∗ (κ, ν) (black); ν = 10−5 m2 s−1 , g = (sin π x1 , sin π x2 ) m s−2 , κconst =
10−7 m2 [18]
∗,err ∗,err ∗,err

0 when computing κ ∗,err and get κ11 (κ) = κ22 (κ) = 8.485κconst and κ21 (κ) =
∗,err
κ12 (κ) = 0 for subdomains with κ = 10[1 + sin(80x1 π ) cos(80x2 π )]κconst . For
∗,err ∗,err ∗,err ∗,err
subdomains with κ = κconst , κ11 (κ) = κ22 (κ) = κconst and κ21 (κ) = κ12 (κ) =
0, where we still use the superscript “err” since the local simulation procedure with
ν = 0 is wrong although the obtained κ ∗,err (κ) is the same as the above κ ∗ (κ, ν).
Now, we have κ ∗,err (κ) equal to 8.485κconst I or κconst I. We use a scalar distribution
of κ ∗,err (κ), which is equal to 8.485κconst or κconst , in another coarse-grid simulation.
Note that the difference between κ ∗,err (κ) and κ ∗ (κ, ν) is distinct in subdomains with
nonuniform κ(x). The results of the coarse-grid simulation using κ ∗,err (κ) are also
given in Fig. 4.23, which shows that the deviation of the coarse-grid simulation using
κ ∗,err (κ) from the fine-grid simulation is noticeable. Thus, the previous computation
of κ ∗ (κ, ν) as the effective permeability using nonzero ν is accurate for upscaling
the Brinkman equation.
4.6 Introduction to Fortran Code
Since the buoyant flow of Sect. 4.5.3 is most sophisticated among the cases demon-
strated above due to its demand of capturing the detailed flow pattern instead of
average quantities (e.g., permeability) of two-phase flows, a general Fortran code
for multiphase multicomponent simulation is provided here (see supplementary
material) but the parameters are set for this special buoyant flow problem. Assume
that all files of the source code are in the folder LBM
$ / . . . /LBM
and the current path of the application terminal on Mac OS or Linux sys-
tem is the directory of LBM, and the compiler gfortran or f77 is installed via
terminal. There are many online instructions of installing these compilers by
using few command lines on terminal when the computer is connected with the
internet. Type the following command line to compile all files inside LBM and then
4.6 Introduction to Fortran Code 161
get an executable file entitled xname that can be arbitrary according to our pref-
erence. The specification -O2 is applied if we want to optimize the computational
efficiency of running xname. The specification -o xname can be removed if we
want to use the default name a.out instead of xname.
$ gfortran −O2 ∗. f −o xname
Then we can run the simulation by using the following command line, which will
generate dynamic messages on the interface of terminal to show the simulation
status and key parameters that are specified in the programming. Additionally, data
files will be generated in the same folder LBM if there are more information to be
independently stored and checked at will. These data files are written in the format of
the software Tecplot, through which the visualization is straightforward although
it is also simple to use Matlab for the visualization.
$ . /xname
The only parameters that are at the disposal of the user for applications are those
in the files of variables and settings.f, which set the discretization grids
and the physical parameters for each particular problem. Modifications can be made
to other subroutines for different extensions.
References
1. Benzi R, Biferale L, Sbragaglia M, Succi S, Toschi F (2006) Mesoscopic modeling of a two-

phase flow in the presence of boundaries: the contact angle. Phys Rev E 74:021509
3. Chen HD, Chen SY, Matthaeus WH (1992) Recovery of the Navier-Stokes equations using a
lattice-gas Boltzmann method. Phys Rev A 45:5339–5342
4. Efendiev Y, Hou YT (2009) Multiscale finite element methods. Theory and applications. Sur-
veys and tutorials in the applied mathematical sciences. Springer, Berlin
5. Ferreol B, Rothman DH (1995) Lattice-Boltzmann simulations of flow through fontainebleau
sandstone. Transp Porous Media 20:3–20
6. Guo ZL, Zheng CG (2002) An extrapolation method for boundary conditions in lattice Boltz-
mann method. Phys Fluids 14:2007–2010
7. Guo ZL, Zheng CG, Shi BC (2002) Discrete lattice effects on the forcing term in the lattice
Boltzmann method. Phys Rev E 65:046308
8. Guo ZL, Zhao TS (2002) Lattice Boltzmann model for incompressible flows through porous
media. Phys Rev E 66:036304
9. Gomez H, Hughes JRT, Nogueira X, Calo MV (2010) Isogeometric analysis of the isothermal
Navier-Stokes-Korteweg equations. Comput Methods Appl Mech Eng 199:1828–1840
10. He XY, Luo LS (1997) Lattice Boltzmann model for the incompressible Navier-Stokes equa-
tion. J Stat Phys 88:927–944
11. He XY, Doolen GD (2002) Thermodynamic foundations of kinetic theory and lattice Boltzmann
models for multiphase flows. J Stat Phys 107:309–328
12. Hysing S, Turek S, Kuzmin D, Parolini N, Burman E, Ganesan S, Tobiska L (2009) Quantitative
benchmark computations of two-dimensional bubble dynamics. Int J Numer Methods Fluids
60:1259–1288
13. Hosa A, Curtis A, Wood R (2016) Calibrating lattice Boltzmann flow simulations and estimating
uncertainty in the permeability of complex porous media. Adv Water Res 94:60–74
14. Kang QJ, Zhang DX, Chen SY (2002) Displacement of a two-dimensional immiscible droplet
in a channel. Phys Fluids 14:3203–3214
15. Kang QJ, Zhang DX, Chen SY (2002) Unified lattice Boltzmann method for flow in multiscale
porous media. Phys Rev E 66:056307
16. Li J, Wang ZW (2010) An alternative scheme to calculate the strain rate tensor for the LES
application in the LBM. Math Probl Eng, 724578
17. Li J (2015) Appendix: Chapman-Enskog expansion in the lattice Boltzmann method. https://
arxiv.org/abs/1512.02599
18. Li J, Brown D (2017) Upscaled lattice Boltzmann method for simulations of flows in hetero-
geneous porous media. Geofluids 1740693
19. Li J, Ho MT, Wu L, Zhang YH (2018) On the unintentional rarefaction effect in LBM modeling
of intrinsic permeability. Adv Geo-Energy Res 2(4):404–409
20. Nithiarasu P, Seetharamu KN, Sundararajan T (1997) Natural convective heat transfer in a fluid
saturated variable porosity medium. Int J Heat Mass Transf 40:3955–3967
21. Pan CX, Luo LS, Miller CT (2006) An evaluation of lattice Boltzmann schemes for porous
medium flow simulation. Comput Fluids 35:898–909
22. Qian YH, d’Humieres D, Lallemand P (1992) Lattice BGK models for Navier-Stokes equation.
Europhys Lett 17:479–484
23. Smagorinsky J (1963) General circulation experiments with primitive equation. Mon Weather
Rev 91:99–164
24. Shan XW, Chen HD (1993) Lattice Boltzmann model for simulating flows with multiple phases
and components. Phys Rev E 47:1815–1820
25. Shan XW, Chen HD (1994) Simulation of nonideal gases and liquid-gas phase transitions by
the lattice Boltzmann equation. Phys Rev E 49:2941–2948
26. Shan XW, Doolen GD (1996) Diffusion in a multicomponent lattice Boltzmann equation model.
Phys Rev E 54:3614–3620
27. Succi S (2001) The lattice Boltzmann equation for fluid dynamics and beyond. Clarendon
Press, Oxford
28. Sukop MC, Thorne DT Jr (2006) Lattice Boltzmann modeling: an introduction for geoscientists
and engineers. Springer, Berlin
29. Sprittles JE (2010) Dynamic wetting/dewetting processes in complex liquid-solid systems.
PhD thesis, University of Birmingham
30. Scarbolo L, Molin D, Perlekar P, Sbragaglia M, Soldati A, Toschi F (2013) Unified framework
for a side-by-side comparison of different multicomponent algorithms: lattice Boltzmann versus
phase field model. J Comput Phys 234:263–279
31. Tang LQ, Cheng TW, Tsang TTH (1995) Transient solutions for three-dimensional lid-driven
cavity flows by a least-squares finite element method. Int J Numer Methods Fluids 21(5):413–
432
32. Yu H, Girimaji SS, Luo LS (2005) DNS and LES of decaying isotropic turbulence with and
without frame rotation using lattice Boltzmann method. J Comput Phys 209(2):599–616
33. Yuan P, Schaefer L (2006) Equations of state in a lattice Boltzmann model. Phys Fluids
18:042101
34. Zhang YH, Qin RS, Emerson DR (2005) Lattice Boltzmann simulation of rarefied gas flows
in microchannels. Phys Rev E 71:047702
35. Zhao BZ, Pahlavan AA, Cueto-Felgueroso L, Juanes R (2018) Forced wetting transition and
bubble pinch-off in a capillary tube. Phys Rev Lett 120:084501
Chapter 5
Summary and Outlook
It is our hope that this book has provided a concise but self-contained introduction to
the theories of describing various flow problems at different scales, and the difference
as well as correlation between different theories. We focused on the kinetic theory of
using the Boltzmann equation, which is valid for modeling flows of different regimes
from continuum to free molecular flow. Then, various fluid behaviors can be modeled
by different simulation methods developed based on the same framework of Boltz-
mann equation. As summarized in Fig. 5.1, four simulation methods are introduced
and they are the DSMC method, DSBGK method, DVM and LBM. The algorithm
differences among those simulation methods are noticeable and their performance
comparisons are detailed in several benchmark problems of the preceding chapters,
which provides some guidelines for those interested in using appropriate methods
for their particular applications.
Despite the fast developments of simulation methods, there are still many chal-
lenging problems. The simulations of real nano-gas flows inside MEMS/NEMS are
affordable to the DSBGK method, which can handle billions of cells with the help
of HPC, but the rock characterization of heterogenous shale sample needs to handle
billions of billions of cells/voxels (voxel size is about few nanometers) even for a
small elementary volume of few millimeters in size. Upscaled computation of effec-
tive permeability is probably the only feasible solution to bridge this scale gap of
few orders of magnitude. But, more research needs to be done to classify the subdo-
mains of few microns in size according to their mineral and chemical composition
such that subdomains of the same geological category need only one representative
pore-scale study to provide the permeability for all subdomains of this category as
the input of the subsequent upscaled computation of effective permeability for the
larger domain. Additionally, although the variation of methane gas permeability with
the pore pressure is dominated by the slippage mechanism of high K n, the varia-
tion curve will change due to the dense gas effect, the confinement effect when the
dominant pore size is close to the molecular size of methane, and the adsorption
effect when the dominant pore size is comparable to the thickness of adsorption
164 5 Summary and Outlook
Fig. 5.1 Comparison of different simulation methods based on the Boltzmann-like equations for
the gas flow problems, and possible extension of the LBM to other flow problems
layer (about the size of a methane molecule). Some of these mechanisms could also
occur in the nano-gas flows inside MEMS/NEMS. All these mechanisms and their
influences can be studied by the molecular dynamics (MD) simulations to quantify
their relative contributions compared to the dominant one of the slippage effect.
In the continuum regime, the two-phase flows with the evolution of dynamic con-
tact angle and movement of contact line on irregular solid surface can be efficiently
simulated by the LBM. But most LBM algorithms are still inaccurate in problems,
where the high density or viscosity ratio has noticeable influence. So far, we rec-
ommend to use the LBM for the pore-scale low-speed problems, where wettability
is the dominant mechanism and can be modeled accurately by the LBM. On the
other hand, the traditional CFD methods based on the N–S-like equation can handle
large density and viscosity ratios but their applications to the pore-scale flows are
usually unreliable due to the lack of generally valid boundary condition for inter-
face dynamics on the solid surface, and numerically unaffordable due to the grid
refinement required for irregular solid surface. In our opinion, developing a DVM
solver based on an extended kinetic theory at the molecular scale will be a rewarding
endeavor for simulating general two-phase flows. Its velocity grid could be coarse
for the continuum regime but the layout will be more sophisticated than that of the
LBM to improve the numerical performance.

Multiscale and Multiphysics Flow Simulations of Using The Boltzmann Equation

Uploaded by

Copyright:

Available Formats

Multiscale and Multiphysics Flow Simulations of Using The Boltzmann Equation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Multiscale and Multiphysics Flow Simulations of Using The Boltzmann Equation

Uploaded by

Copyright:

Available Formats

Jun Li

Multiscale and Multiphysics

ISBN 978-3-030-26465-9 ISBN 978-3-030-26466-6 (eBook)

Dhahran, Saudi Arabia Jun Li

1 Fluid Mechanics Based on Continuum Assumption . . . . . . . . . . . . . 1

2.10.2 Diffuse Reﬂection Model . . . . . . . . . . . . . . . . . . . . . . . 59

PDE Partial differential equation

1.1 Continuum Assumption

sient samples within a volume element L 3d might be insufficient since L d needs to

1.2 Macroscopic Descriptions of Flow Phenomenon

Q = f 1 (x0 , t) = f 1 ( f −1 (x, t), t) = f 2 (x, t), (1.1)

Q(x0 , t + Δt) = Q(x + Δtu, t + Δt), (1.3)

1.3 Derivations of Macroscopic Governing Equations

1.3.1 Force and Heat Interactions Between Fluid Parcels

Fig. 1.1 Schematics of

d An n+ + d A1 (−e1 ) + d A2 (−e2 ) + d A3 (−e3 ) = 0, (1.5)

The Newton’s second law applied to the tetrahedron body is [9]:

ρΔV a = ρΔV g + P−e1 d A1 + P−e2 d A2 + P−e3 d A3 + Pn+ d An , (1.7)

where ρ is the mass density, ΔV is the volume of tetrahedron, a is the accel-

Pn+ = Pe1 n +,1 + Pe2 n +,2 + Pe3 n +,3 . (1.8)

where P ji = Pe j · ei are components of P and the convention of Einstein summation

We define the transformation coefficients αi j , which depends on time t in general,

ei = (ei · ej )ej = αi j ej ,

Q i1 i2 ···in = Q j1 j2 ··· jn αi1 j1 αi2 j2 · · · αin jn (1.13)

due to Eq. (1.12) and the following objective property:

P j1 j2 = Pej · ej2

= [Pe1 (ej1 · e1 ) + Pe2 (ej1 · e2 ) + Pe3 (ej1 · e3 )] · ej2

P j1 j2 αi1 j1 αi2 j2 = Pi3 i4 αi3 j1 αi4 j2 αi1 j1 αi2 j2

where p is the thermodynamic pressure, μ is the dynamic viscosity, μv is the second

1.3.2 Continuity Equation

1.3.3 Momentum Equation

1.3.4 Energy Equation

Note that the scalar product of u and Eq. (1.25) is:

where ∇ · (P · u) = (∇ · P) · u + P : ∇u has been substituted, which means that the

1.3.5 Entropy Equation

where two equations of thermodynamic functions, i.e., (∂e/∂s)|ρ = T and

1.3.6 Convection-Diffusion Equation

In some problems, we need to study the concentration evolution of a species (solute),

1.4 Transformation to Arbitrary Non-inertial Reference

where the Levi-Civita symbol is defined as:

where ∇ ≡ ∂/∂X = (∂/∂ X i )ei . Since ∇ · u ≡ ∇ · U, the velocity divergence is

Since ∇T ≡ ∇ T , the gradient of temperature is objective as well. Additionally, we

where we substituted ∇ · [μ(∇ U + ∇ UT )] = ∇ · [μ(∇ U + ∇ UT )], which can be

where (D2 yi /Dt 2 )ei is the translational acceleration of R relative to R and

1.5 Component Form in Arbitrary Orthogonal Curvilinear

Fig. 1.2 Schematics of

H1 = lim (bb1 /Δl1 )

Δx12 + Δx22 + Δx32

Thus, we have (H1 , H2 , H3 ) = (1, r, 1) in the cylindrical coordinate system with

Fig. 1.3 Schematics of

|Δl3 ∂ ẽ2 /∂l3 | Δl2 ∂(H3 Δl3 )/∂l2 1 ∂ H3

where the repeated indices do not imply summations.

∇ Q · ẽi = lim (ΔQ/bbi )

1.6 Limitations of Continuum Assumption

Could be Burnett Eq.

1. Batchelor GK (1967) An introduction to fluid dynamics. Cambridge University Press, Cam-

2.1 Distribution Function in the Phase Space

© Springer Nature Switzerland AG 2020 25

2.2 Intermolecular Collision Dynamics