Paulo

C Philippi

Lectures on the
Lattice-Boltzmann
method
Lectures on the Lattice Boltzmann method
Paulo C Philippi
Mechanical Engineering Graduate Program.
Pontical Catholic University of Paraná.
Curitiba, PR, Brazil

March 27, 2022

2
 3

Preface
This set of lectures is intended to be a rst course in the Lattice-Boltzmann
Method (LBM) for undergraduate and graduate students, who, in addition to
learn the method, want to understand the theoretical fundamentals behind it.
LBM is a kinetic numerical method and this means that the behavior of a
physical system, usually a uid, is numerically predicted by solving a kinetic
model of this system.
In the LBM framework, uids are considered as a system of particles and
we have, mainly, three hierarchic scales in studying such systems: a) the mi-
croscopic or molecular scale; b) the mesoscopic or kinetic scale and c) the
macroscopic scale.
The microscopic or molecular scale is the subject of the Molecular Dynam-
ics, in which the motion of each individual molecule is followed by solving its
equation of motion, usually, the Newton's second law of mechanics and con-
sidering all the interaction forces with the other molecules and the boundaries.
Molecular dynamics simulations are usually performed in very large clusters of
computers and the main physical parameters of the whole system is obtained
through statistical ensemble averages.
The mesoscopic or kinetic framework is a less detailed level of description,
where the interest is not to follow each individual particle along its trajectory,
but to nd the expected number of particles that will be found in a given state,
or the distribution f of particles over the several possible states of the system.
The state of a particle is usually given by its molecular translational velocity, ξ,
but it can be also given by the angular velocity, ω , when the particle shape and
rotational degrees of freedom are considered and by the vibrational quantum
states when the molecular modes of internal vibration are considered. In this
level of description, the distribution f is found by solving a kinetic equation.
Historically, the rst and more well-known kinetic equation is the Boltzmann
equation, derived by Ludwig Boltzmann, in the late 19th century.
Finally, the macroscopic level of description is the less detailed one, where
the interest is to follow the time and space evolution of a given set of moments
of the distribution function f, the density, ρ, the macroscopic velocity u and
the temperature T, usually given in terms of a set of balance equations, the
Fourier-Navier-Stokes equations.
The purpose of LBM is not to nd the true distribution f of particles, but to
consistently retrieve the macroscopic behavior of a physical system. Therefore,
numerically solving a kinetic equation, the framework of LBM, must, rstly,
be thought as a method for solving a physical problem, with, in certain cases,
some advantages with respect to classical CFD due to its intrinsic Lagrangian
nature and hyperbolic basis.
In the reverse sense, the LBM framework enables to reveal, or to put in
evidence, the inuence of a number of molecular processes on the macroscopic
behavior of a physical system. In fact, the kinetic level of description, may be
thought as a two-way bridge linking the molecular and the macroscopic levels
of description. Therefore, although derived from the molecular domain and,
4

in consequence, featuring some properties of uids that are inaccessible to the

macroscopic level of description, the kinetic model may be written in such a
manner as to recover the correct macroscopic equations.
Numerical methods based on kinetic models have been the scenery of an
impressive growth in the last two decades both in their theoretical foundations
as well as in their range of applications in physics and engineering, including
microscopic, nanoscale and multiscale physics for emerging technologies, inter-
facial phenomena and micro and nanouidics. Without being restricted to very
small scales as in molecular dynamics (MD), when compared to classical com-
putational uid dynamics (CFD), kinetic methods have as their main features:
i) the hability of solving macroscopic problems that require information from
the underlying molecular scale; ii) an hyperbolic basis, the eld information
travelling along the characteristics lines, avoiding to solve a large system of lin-
ear equations at each time step and iii) an intrinsic Lagrangian nature avoiding
to capture the interface in multiphase ow problems.
There are two main classes of kinetic methods. The rst class the lattice-
Boltzmann method (LBM) bases its approach on solving a discrete form of a
kinetic equation with a given numerical scheme. Such methods include the nite
dierence discretization and the collision-propagation method. A second class
of kinetic methods is based on the construction of numerical models to directly
simulate the behavior of particle systems. The most well-known method of
such type is the Direct Simulation Monte Carlo (DSMC) method and include
the smoothed-particle hydrodynamics (SPH), the moving particle semi-implicit
(MPS) and the dissipative particle dynamics (DPD) method.
Present Lectures are restricted to the rst class of kinetic methods. In
particular, we focus on the establishment of the formal link between the con-
tinuous Boltzmann equation and its dierent discrete forms. This link was the
starting point for the high-order discretization of continuous kinetic models
and can be considered as a breakthrough point toward high-order mesoscopic
modelling of complex physical systems. These discrete forms have been applied
to non-isothermal and fully compressible ows and to the numerical simulation
of non-ideal uids, whose description requires to consider the attraction forces
between the molecules and, so, to nd improved kinetic equations, beyond the
Boltzmann equation, restricted to rareed gases. At this point, it is, perhaps,
important to clarify that the name `Boltzmann ' in `Boltzmann equation ' is
here used for a kinetic model based on an hyperbolic evolution equation for the
expected number f of particles, with a streaming and an interaction term.

Paulo C Philippi

Curitiba, December 2021

Contents

1 Introduction 15
1.1 Early history . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.1.1 Molecular dynamics (MD) . . . . . . . . . . . . . . . . . 15
1.1.2 MD simulation of phase transition . . . . . . . . . . . . 17
1.1.3 Lattice Gas Automata (LGA) . . . . . . . . . . . . . . . 17
1.2 Towards the lattice-Boltzmann method . . . . . . . . . . . . . . 21
1.3 Connection with the Boltzmann equation . . . . . . . . . . . . 23
1.3.1 High-order lattice Boltzmann . . . . . . . . . . . . . . . 24

2 The Boltzmann equation 27

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 The Boltzmann equation . . . . . . . . . . . . . . . . . . . . . . 31
2.3 The equilibrim solution . . . . . . . . . . . . . . . . . . . . . . 35
2.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3 A rst insight into LBM 43

3.1 From the Boltzmann equation to LB schemes . . . . . . . . . . 43
3.2 Simulation in LBM . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.3 Some features of the D2Q9 LBE . . . . . . . . . . . . . . . . . 53
3.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.5 Boundary Conditions: rst part . . . . . . . . . . . . . . . . . . 55
3.6 Bounce back schemes . . . . . . . . . . . . . . . . . . . . . . . . 56
3.7 Inamuro et al. (1995) counter-slip condition . . . . . . . . . . . 58
3.8 Zou and He (1997) boundary condition for at surfaces . . . . 59
3.9 Zou and He (1997) conditions for inlet/outlet . . . . . . . . . . 59
3.10 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.11 Sample case: lid driven cavity ow . . . . . . . . . . . . . . . . 60
3.11.1 Numerical simulation work . . . . . . . . . . . . . . . . 61

4 Forces in LBM 63
4.1 Lattice Boltzmann equation . . . . . . . . . . . . . . . . . . . . 63
4.1.1 Discretization of the stream term . . . . . . . . . . . . . 65
4.2 A smart procedure for overcoming the impliciteness of the nu-
merical scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.3 Another alternative: changing variables . . . . . . . . . . . . . 68

5
6 CONTENTS

4.3.1 Step by step procedure . . . . . . . . . . . . . . . . . . . 69

4.4 Simulation work: plane Poiseuille ow . . . . . . . . . . . . . . 69

5 The discretization problem 71

5.1 Hermite polynomials in one dimension . . . . . . . . . . . . . . 71
5.1.1 General expression . . . . . . . . . . . . . . . . . . . . . 72
5.1.2 Internal products . . . . . . . . . . . . . . . . . . . . . . 72
5.1.3 Maxwell Boltzmann expansion . . . . . . . . . . . . . . 73
5.2 Hermite polynomial tensors . . . . . . . . . . . . . . . . . . . . 75
5.2.1 General expression . . . . . . . . . . . . . . . . . . . . . 75
5.2.2 Internal products . . . . . . . . . . . . . . . . . . . . . . 75
5.2.3 Maxwell Boltzmann expansion . . . . . . . . . . . . . . 76
5.3 Velocity discretization . . . . . . . . . . . . . . . . . . . . . . . 77
5.3.1 Example in one-dimension . . . . . . . . . . . . . . . . . 81
5.3.2 Example in two-dimensions . . . . . . . . . . . . . . . . 84
5.3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.4 Cartesian products . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.4.1 The D2Q9 LBE . . . . . . . . . . . . . . . . . . . . . . . 89
5.4.2 The D3Q27 LBE . . . . . . . . . . . . . . . . . . . . . . 91
5.4.3 High-order Cartesian products . . . . . . . . . . . . . . 92
5.5 The moments space . . . . . . . . . . . . . . . . . . . . . . . . 95

6 Chapman-Enskog analysis 99
6.1 The continuous case . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.1 Macroscopic equations . . . . . . . . . . . . . . . . . . . 102
6.1.2 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
6.2 Scale analysis of a LB equation . . . . . . . . . . . . . . . . . . 107
6.3 Zero-th order transport equations . . . . . . . . . . . . . . . . . 109
6.3.1 Mass conservation . . . . . . . . . . . . . . . . . . . . . 109
6.3.2 Momentum balance . . . . . . . . . . . . . . . . . . . . 110
6.3.3 Energy conservation . . . . . . . . . . . . . . . . . . . . 111
6.3.4 Internal energy balance equation . . . . . . . . . . . . . 112
6.4 First-order transport equations . . . . . . . . . . . . . . . . . . 113
6.4.1 Mass conservation equation . . . . . . . . . . . . . . . . 114
6.4.2 Momentum balance equation . . . . . . . . . . . . . . . 114
6.4.3 Energy conservation . . . . . . . . . . . . . . . . . . . . 116
6.4.4 Internal energy balance equation . . . . . . . . . . . . . 118
6.5 Macroscopic equations . . . . . . . . . . . . . . . . . . . . . . . 119
6.5.1 Mass conservation equation . . . . . . . . . . . . . . . . 119
6.5.2 Momentum balance equation . . . . . . . . . . . . . . . 119
6.5.3 Internal energy balance equation . . . . . . . . . . . . . 119
6.6 Summary of the necessary conditions for retrieving the macro-
scopic transport equations . . . . . . . . . . . . . . . . . . . . . 120
6.7 The viscous stress tensor . . . . . . . . . . . . . . . . . . . . . . 121
6.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.9 LBE with a forcing term . . . . . . . . . . . . . . . . . . . . . . 123
CONTENTS 7

6.9.1 First-order discretization of the stream term . . . . . . . 123

6.9.2 Second-order discretization for the source term . . . . . 127

7 Beyond BGK: moments-based methods 129

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7.2 Kinetic Projections . . . . . . . . . . . . . . . . . . . . . . . . . 133
7.3 Increasing the order of approximation . . . . . . . . . . . . . . 137
7.3.1 Adding Hermite polynomials to subspace HN . . . . . . 138
7.3.2 Increasing the order of the lattice-Boltzmann equation . 142
7.4 Moments-based methods . . . . . . . . . . . . . . . . . . . . . . 143
7.4.1 A sample case: Double periodic shear layer ow . . . . . 145
7.5 Toward high-order moments-based LBE: recurrence relations . 148
7.6 Moments-based LBE with recurrence relations . . . . . . . . . 153
7.6.1 Isothermal hydrodynamics . . . . . . . . . . . . . . . . . 153
7.6.2 Non-Isothermal hydrodynamics . . . . . . . . . . . . . . 156

8 Boundary conditions: part II 159

8.1 Mass conservation . . . . . . . . . . . . . . . . . . . . . . . . . 160
8.2 Boundary nodes on plane surfaces . . . . . . . . . . . . . . . . 162
8.2.1 Latt et al. model . . . . . . . . . . . . . . . . . . . . . . 162
8.2.2 Malaspinas et al. model . . . . . . . . . . . . . . . . . . 164
8.2.3 Mohammed and Reis model . . . . . . . . . . . . . . . . 164
8.2.4 Hegele et al. model . . . . . . . . . . . . . . . . . . . . . 164
8.3 Corner nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
8.4 Comparative analysis . . . . . . . . . . . . . . . . . . . . . . . . 169
8.4.1 Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . 169
8.4.2 Mass leakage . . . . . . . . . . . . . . . . . . . . . . . . 172
8 CONTENTS
List of Figures

1.1 Lennard-Jones 12-6 potential. . . . . . . . . . . . . . . . . . . . 16

1.2 Molecular dynamics simulation of a cooling process . . . . . . . 17
1.3 A regular triangular lattice . . . . . . . . . . . . . . . . . . . . 18
1.4 Collision step in LGA. . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 1985 article in the Washington Post. . . . . . . . . . . . . . . . 19
0
1.6 Transition matrix α (s, s ) for the HPP model. . . . . . . . . . . 20
1.7 Sets of velocity vectors for high-order LBM . . . . . . . . . . . 25

2.1 Ludwig Boltzmann (1844 1906) . . . . . . . . . . . . . . . . . 27

2.2 Front cover of 'Hydrodynamica' by Daniel Bernoulli, published
in 1788. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Considering all the molecules inside a parallelepiped of sides ∆x,
∆y and ∆z, only f (x, ξ, t)∆x∆y∆z∆ξx ∆ξy ∆ξz have velocities
between ξ and ξ + ∆ξ . . . . . . . . . . . . . . . . . . . . . . . . 31
2.4 Trajectory deviation of the bullet particle produced by the radial
frainage due to the repulsion forces. Vector
→
−α is a unitary vector
0
along γ −γ . Parameter b is the impact factor and angle ε is the
azimuth angle of the collision plane. . . . . . . . . . . . . . . . 33
2.5 Elementary volume in cylindrical coordinates. . . . . . . . . . . 34

3.1 P.L. Bhatnagar (1912-1976) . . . . . . . . . . . . . . . . . . . . 44

3.2 D2Q9 [79] and RD2Q9 [74] sets of lattice vectors. . . . . . . . . . . 45
3.3 Ilustration of the collision-propagation scheme in LBM. The amount
of particles in each direction is represented by the thickness of the
vector pointing in that direction. The collision step tries to equi-
librate the dierent populations and some particles along the most
peopled populations, in the northeast directions, are thrown to the
other directions. . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4 In the standard bounce-back scheme the boundary nodes are wet
nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.5 Half-way bounce-back boundary condition: the wall surface is
supposed to be located in the half-way between the boundary
and the uid sites. . . . . . . . . . . . . . . . . . . . . . . . . . 57

9
10 LIST OF FIGURES

3.6 Half way bounce back: population f1 is supposed to propagate

to the solid site when if nds the solid surface at time t + δ/2
and returns back to the uid site. So f3 at time t+δ will be the
same f1 at time δ. . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.7 A wet boundary node is always to be considered as a uid node. 58
3.8 Zou and He [94] boundary conditions for the inlet. . . . . . . . 60
3.9 Streamlines for a square lid driven cavity ow when Re=1000 [ ? ]. 61
3.10 Numerical results for Re=100, 400, 1000 and 5000 showing the:
a) the horizontal component of the velocity along the vertical
mid-line ; b) the vertical component of the velocity along the
horizontal mid-line of the cavity [ ? ]. Results are compared
with Ghia et al. [30], Botella and Peyret [8] and Erturk et al. [25]. 62

4.1 Rayleigh-Bénard instability . . . . . . . . . . . . . . . . . . . . 64

4.2 Rayleigh-Taylor Instability. . . . . . . . . . . . . . . . . . . . . 64
4.3 Velocity prole for a steady-state, fully developed incompressible
channel ow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

5.1 The Hilbert space. . . . . . . . . . . . . . . . . . . . . . . . . . 71

5.2 Two-dimensional lattices. . . . . . . . . . . . . . . . . . . . . . 85
5.3 Three dimensional counterparts of the D2Q9 set of lattice-vectors 86
5.4 Lattices D2Q17 and D2Q21. For the D2Q21 lattice there will
be seven unknowns as , W0 , W1 , W2 , W3 , W4 .W5 for six norm re-
strictions, after eliminating identical equations. Letting as be a
free variable, the system gives a solution with real positive roots
when as is inside the interval 0.659836 and 1.16208. The val-
ues as =0.659836 and as =1.16208 are roots of the polynomials
W0 (as ) and W3 (as ) , respectively. Therefore, when the value
as =1.16208 is chosen, W3 =0, corresponding to the lattice vec-
tors in red, giving a modication of the D2Q17 lattice, which
was named D2V17. . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.5 The D2Q9 LBE as related to the Hilbert subspace H2 . . . . . . 88
5.6 The D2V17 LBE as related to the Hilbert subspace H3 . . . . . 88
5.7 Hermitian representations for the D2Q9 LBE. (a) Second order
Hermitian basis; (b) Full Hermitian basis as obtained from the
Cartesian product of one-dimensional representations. . . . . . 91
5.8 D3Q27 full Hermitian basis. All the Hermite polynomials in
n o
(1) (2)
red results from the Cartesian product H (0) , Hx , Hxx ×
n o n o
(1) (2) (1) (2)
H (0) , Hy , Hyy × H (0) , Hz , Hzz . . . . . . . . . . . . . 92

5.9 The fourth-order D2Q25 LBE. . . . . . . . . . . . . . . . . . . 94

5.10 Hermitian basis of the D2Q25-ZOT LBE. Hermite polynomials
in red result from the Cartesian product of two D1Q5-ZOT LBE.
Hermite polynomials in blue are Hermite polynomials modied
by factor γ given in Eq. (5.22) . . . . . . . . . . . . . . . . . . 95
LIST OF FIGURES 11

7.1 The Hilbert space solution of the Boltzmann equation. . . . . . 133

7.2 Subspace Hq=10 of H generated by a nite set of 10 Hermite

polynomial tensors and related to the D2V17 LBE. . . . . . . . 136

7.3 Hermitian basis of the Hilbert subspaces Hq=6 in (a) and Hq=9
in (b), related to the D2Q9 LBE. . . . . . . . . . . . . . . . . . 138

7.4 The eect on stability of adding third order Hermite polynomials

to the second order equilibrium distribution in a D2Q9 LBGK
[85]. Comparison with the MRT model of Ref. [56] . . . . . . . 141

7.5 Results of a von-Neumann stabilty analysis showing the eect

of increasing the order of the LBE for enlarging the stability
∗
ranges. Parameter ν is the dimensionless kinematic viscosity
∗
and umax is the maximum local velocity beyond which the nu-
merical scheme becomes unstable. . . . . . . . . . . . . . . . . . 142

7.6 an Gogh clouds showing the Kelvin-Helmholtz instability. . . . 145

7.7 Kelvin-Helmholtz instability in shear ow. A small velocity per-

turbation, perpendicular to the shear ow direction, initiates a
Kelvin-Helmholtz instability and causes roll up of the shear layers.146

7.8 Initial velocity proles used for the simulation of the double pe-
riodic shear ow. . . . . . . . . . . . . . . . . . . . . . . . . . . 147

7.9 Decay of average kinetic energy for an initial Reynolds num-

ber, Re = 32000, L = 128 lattice unities, u∗0 =1/32 and ν ∗ =
4/30000. The average values are normalized with respect to
2
ρu∗o /2. The LBGK scheme becomes unstable early on, around
∗
time t = 0.3, corresponding to 1230 time steps. The moments-
based LB scheme remain stable during the whole simulation, [66] 147

7.10 Vorticity eld for t∗ = 1 showing the simulation results for the
LBGK model (left) and the moments-based method (right). It
is clearly seen that the LBGK model is plagued by spurious
secondary vortex even for Re = 32000 whereas the moments-
based scheme produces accurate and stable results up to Re =
80000 [66] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

7.11 Third-order Hermitian basis of the subspace Hq=10 and its re-
lated D2V17 set of lattice vectors. . . . . . . . . . . . . . . . . 155

7.12 Fourth-order Hermitian representation (a) and its correspondng

D2V37 set of lattice vectors (b). . . . . . . . . . . . . . . . . . 157

8.1 A boundary node. . . . . . . . . . . . . . . . . . . . . . . . . . 160

8.2 A concave corner node . . . . . . . . . . . . . . . . . . . . . . . 167

8.3 Lid-driven cavity ow for Re=1000 [ ? ]. . . . . . . . . . . . . . 169

8.4 Numerical results for Re=400, 1000 and 5000 showing the rel-
ative velocity u∗x /u∗0 along the mid-line y ∗ = L/2. Results are
compared with Ghia et al. [30], Botella and Peyret [8] and Er-
turk, Corke and Gökçöl [25]. Continuous lines are from Bazarin
et al. [ ? ]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
12 LIST OF FIGURES

8.5 Numerical results for Re=400, 1000 and 5000 showing the rel-
ative velocity u∗y /u∗0 along the mid-line x∗ = L/2. Results are
compared with Ghia et al. [30], Botella and Peyret [8] and Er-
turk, Corke and Gökçöl [25]. Continuous lines are from Bazarin
et al. [ ? ]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
8.6 Stream function and vorticity isolines for (a) Re = 100, (b) Re
= 400, (c) Re = 1000, (d) Re = 3200 and (e) Re = 5000. The
top row illustrate the stream function and the bottom row the
vorticity, Bazarin et al. [ ? ]. . . . . . . . . . . . . . . . . . . . . 171
8.7 Average density in terms of the number of time steps. We com-
pare the results of Latt et al. (2008), Malaspinas et al. (2011)
and Mohammed and Reis (2017) with the solution methods pre-
sented by Bazarin et al. (2021). Both Malaspinas et al. and Latt
et al. models fail in preserving the total mass of the system.
Method IV based on the concept of density is neither a mass
preserving method, but mass leaking is much smaller, Bazarin
et al. [ ? ]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
8.8 Average density measured with a more accurate weigth balance
in terms of the number of time steps. method IV based on den-
sity, presents the same linear behavior as observed in Malaspinas
et al. (2011) and Latt et al. (2008) models, although with a
much less mass variation, Bazarin et al. [ ? ]. . . . . . . . . . . 173
8.9 Local mass leakage for Mohammed and Reis (2017) model after
50000 time steps: (a) north and south boundaries; (b) east and
west boundaries; . . . . . . . . . . . . . . . . . . . . . . . . . . 174
List of Tables

5.1 Hermitian representation of the D2Q9 LBE considered as the

cartezian product of the corresponding one-dimensional Hermite
representations, . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

13
14 LIST OF TABLES
Chapter 1

Introduction

1.1 Early history

Although the name `Boltzmann ' in `lattice-Boltzmann ' suggests that this
method was established having the Boltzmann equation as its foundation stone,
to nd the true origin of the lattice-Boltzmann method it is necessary to go
down toward much smaller scales governed by the exclusion principle and by
the Fermi-Dirac equilibrium distribution.

1.1.1 Molecular dynamics (MD)

If we consider a uid as a system of material points numbered i = 1, 2, ..., n
and that the motion of each one of these particles satisfy the Newton's laws

d2 xi X
m = Fij , i = 1, ..., n (1.1)
dt2
j6=i

the position xi and velocity ξi = dxi /dt of each particle may be followed
along the time by solving the above system of equations. Two conditions are,
nevertheless, required to be satised for this solution:

1. the forces Fij between particles i and j must be known;

2. the initial position and velocity must be known for each particle.

Intermolecular forces
Molecules attract themselves with an electrostatic force that is dependent
on their polar moment. Asymmetric molecules such as H2 0 and CO2 have
a permanent dipole, the geometric center of the electronic clouds does not
coinciding with the geometric center of the positive charges.
The dipole length and the dipolar moment of a polar molecule are subjected
to uctuations produced by electrostatic elds originated from other molecules.

15
16 CHAPTER 1. INTRODUCTION

In the same way, a non-polar molecule will acquire a dipolar moment, µind , un-
der the induction of an electrical eld, E. We dene the molecular polarizibility,
α, by

µind = αE,
giving a measure of the easiness a molecule becomes polarized, or changes its
dipolar moment, when subjected to an electrical eld.
Symmetrical molecules do not have any permanent electrical dipole and the
attractive forces among these molecules was rst described by London in 1930
as due to very rapid uctuations of the geometric center of their molecular
electronic clouds.

Figure 1.1: Lennard-Jones 12-6 potential.

All these attractive forces between neutral molecules can be writen in terms
of a two-particles r−6 potential energy given by

B
Φ12 = − ,
r6
where B is a molecular property dependent on the electrostatic properties of
the interacting molecules 1 and 2 and r = |r2 − r1 |.
When molecules are close enough they experience a repulsion force due to
the overlapping of their electronic orbitals. The combined attraction-repulsion
eect can be described by the 12-6 Lennard-Jones potential [61] shown in Figure
1.1.

A B
Φ12 = − 6
r12 r
In molecular dynamics simulations it is usual to replace parameters A and
B by

A = 4σ 12 , B1−2 = 4σ 6 ,

where σ is the molecular diameter or the distance where Φ12 = 0 and  is the
potential strength.
1.1. EARLY HISTORY 17

Figure 1.2: Molecular dynamics simulation of a cooling process

Initial conditions
Molecular dynamics is a deterministic science, meaning that the initial
state

xi , ξi , i = 1, ..., n
is considered to be known at time t = 0.
Simulation of uid ow is performed by taking the ensemble average of
several realizations, starting from dierent initial conditions.

1.1.2 MD simulation of phase transition

Figure 1.2 shows a molecular dynamics simulation of a vapor condensation
process based on an N-body simulation of the Newton second law of motion.
Each one of the N particles is subjected to a trajectory, in the physical space,
given by the solution of the Eqs. (1.1), where Fij is the force between each
i−particle and all the remaining particles, evaluated by supposing a Lennard-
Jones interaction potential among the particles.
Particles are spherical with a diameter that is given by the inversion point
of the Lennard-Jones potential, where the attractive forces become repulsive.
Each time a particle collide with the container surface, it is reected back
following a specular reection and with only a previously established fraction
of the kinetic energy it had before the collision, trying to reproduce a cooling
process at the walls. Attractive forces between the wall container and the
particles where not considered for avoiding condensation at the wall surface.
The initial state was randomly set.
We can see that this molecular dynamics simulation gives a good picture of
what is expected to happen in a condensation process at the molecular scale,
although the analysis was performed on a mechanical deterministic system of
particles, subjected to Newton's second law of motion, without any help of
thermodynamic concepts such as thermodynamic energy and entropy.

1.1.3 Lattice Gas Automata (LGA)

 A new paradigm. A virtual simplied micro-world is constructed as an
18 CHAPTER 1. INTRODUCTION

Figure 1.3: A regular triangular lattice

Figure 1.4: Collision step in LGA.

automaton universe based not on a realistic description of interacting particles

(as in molecular dynamics simulations) but merely on the conservation laws of
microscopic physics and the laws of symmetry of macroscopic physics of uids.
(Frisch, Haslacher, Pomeau, 1986 [26])

Molecular dynamics simulations are computer expensive and limited to

nanoscopic systems. Design of a molecular dynamics simulation should ac-
count for the available computational power, which is limited even considering
large computational clusters.
On the other hand, the macroscopic behavior of uids appear to be not
sensitive to the details of molecular interactions. So, let us consider that our
uid is a system of particles, as in molecular dynamics, but we restrict these
particles to be in the sites of a regular triangular lattice as in Figure 1.3
Site x can be occupied by 6 particles with unitary velocities e1 , ..., e6 . Each
one of these particles is represented by a Boolean variable ni = {0, 1} , i =
1, ..., 6. When ni = 1 direction i is occupied by a particle, otherwise ni = 0.
After each time step, these particles are propagated for the neighbors sites
x + ei h shown in the gure.
Now, consider that at time step t, site x is occupied by 3 particles after
they were propagated from the neighbors sites, as shown in Figure 1.4.a.
At time t there are 3 particles in site x and the local momentum at this
1.1. EARLY HISTORY 19

Figure 1.5: 1985 article in the Washington Post.

site is zero. So, if we want to simulate a collision process, we must preserve the
mass

X
mni = ρ, (1.2)
i
and the momentum

X
mni ei = ρu. (1.3)
i
Thus, we must choose, between the possible post-collisional congurations,
the ones that satisfy Eqs. (1.2-1.3). In this case, there are only two post-
collisional congurations that satisfy these restrictions. Figure 1.4.b show one
of such congurations.
Boolean models or `Lattice Gas Automatas (LGA)'  as they were named
represent a drastic simplication of the molecular dynamics where:

1. the innite space of molecular velocities is replaced by a discrete and

nite set;

2. the collision process is only required to preserve the mass and momentum
of the particles.

Nevertheless, they are able to produce a sketch of uid ow and represented a
new paradigm in CFD methods for hydrodynamics (Figure 1.5).
They were invented by Hardy, Pomeau and de Pazzis in 1973 [37], [36] who
adopted a square lattice with 4 neighbors for each site, the HPP lattice.
The HPP Boolean model was criticized by Frish, Hasslacher and Pomeau
[26]. In the words of these authors:  When density and momentum are varied
slowly in space and time, "macrodynamical" equations dier from the nonlinear
Navier-Stokes equations in three respects. The discrepancies may be classied
as (1) lack of Galilean invariance, (2) lack of isotropy, and (3) a crossover
dimension problem .
20 CHAPTER 1. INTRODUCTION

Figure 1.6: Transition matrix α (s, s0 ) for the HPP model.

In fact, the HPP does not have enough symmetry properties to insure the
isotropy of fourth rank lattice tensors and, so, fails to relate the viscous stress
tensor to the velocity gradients.

In this same paper, Ref. [26], Frish, Hasslacher and Pomeau proposed a
new LGAthe FHP Boolean model based on the triangular lattice shown
in Figure 1.3, avoiding the HPP drawbacks.

The algorithm
In the collision step the Boolean variables ni (x, t) related to site x are
modied in accordance with

n0i (x, t) = ni (x, t) + ωi (n1 , ..., nb ) , (1.4)

where ωi is the collision operator and takes the value 1, 0 or -1 in accordance

with the pre-collisional state (n1 , ..., nb ) of site x.
b
For each site of the lattice with b directions there are 2 possible states
represented by the vectors s∈S, s = (s1 , ..., sb ). Thus, the collision operator
can be dened through a transition operator:

α : S × S → {0, 1} .

The transition operator α can be represented by a Boolean transition ma-

0
trix, α (s, s ), with 2b X 2b elements and which assumes the value 1 when, and
only when, the collision process changes the state s to the state s'.

Figure 1.6 shows the transition matrix for the HPP model. There are 16
possible states. The only transition that is allowed in the HPP model is the
one that changes the state (0, 1, 0, 1) to state (1, 0, 1, 0) and vice-versa.

The collision operator can be written in terms of the transition matrix in

accordance with
1.2. TOWARDS THE LATTICE-BOLTZMANN METHOD 21

ωi (n1 , ..., nb )
b
s
XX Y 1−sj
= α(s, s0) (s0i − si ) nj j (1 − nj ) . (1.5)
s s0 j=1

Therefore, in the HPP model if the pre-collisional state is n = (0, 1, 0, 1),

the only state s for which the product

b
s
Y 1−sj
nj j (1 − nj ) 6= 0,
j=1

is the state s = (0, 1, 0, 1) and since α(s, s0) is diferent from zero only when
s0 = (1, 0, 1, 0), we will get

ωi (n1 , ..., n4 ) = α(s, s0) (s0i − si )

or

ω1 = 1, ω2 = −1, ω3 = 1, ω4 = −1.
and these values changes the state s = (0, 1, 0, 1) into the state s0 = (1, 0, 1, 0).
In the propagation step the Boolean variables are propagated to the neighbor
sites.
In FHP model there are 64 possible states and more than one state s0 for
some states s with two or three particles. In this case, we use a random variable
to decide which post-collisional Boolean state is to be chosen.
Finally, in some later versions of the FHP model a rest particle was added
and the Boolean state was written as

(n0 , n1 , ..., nb )
increasing the number of possible states to 128 and with benecial eects for
the simulations.

1.2 Towards the lattice-Boltzmann method

Historically, the lattice-Boltzmann equation (LBE) was introduced by Mc-
Namara and Zanetti [70] in 1988, as an alternative technique to the lattice-gas
automata for the study of hydrodynamic properties, replacing the Boolean vari-
ables in the discrete collision-propagation equations by their ensemble averages.
This approach eliminated the statistical noise that plagued the lattice-gas sim-
ulations.
We can think about the description of a lattice gas model in three levels.
At a more detailed level, the ni (x, t) are specied for all points in the network
and for each time step. This provides a complete description of the system. In
general this description is much more detailed than is required. A less detailed
22 CHAPTER 1. INTRODUCTION

description is obtained with a set of kinetic equations for the expected values
of Ni = hni (x, t)i, obtained over a set of realizations. However, even this
description is redundant for many purposes. Thus, on a third level we look for
a closed system of equations involving the moments of Ni leading, e.g., under
certain hypotheses, to the Navier-Stokes equations.
So, consider several realizations of a Boolean model and the ensemble aver-
age of Eq. (1.4)

Ni0 (x, t) = Ni (x, t) + Ωi (N1 , ..., Nb ) , (1.6)

where

Ωi (N1 , ..., Nb ) = hωi (n1 , ..., nb )i .

It should be noted that ωi (n1 , ..., nb ) is given by Eq. (1.5) as a product
of Boolean variables associated with the particles present in x at time t, prior
to the collision process. These variables are uncorrelated, assuming the molec-
ular chaos hypothesis, i.e., considering that the value of the Boolean variable
assigned by the process to one direction in x at time t does not depend on the
value assigned to the other direction in the same point x and at the same time
t. Thus:

Ωi (N1 , ..., Nb )
b
s
XX Y 1−sj
= A(s, s0) (s0i − si ) Nj j (1 − Nj ) . (1.7)
s s0 j=1

In the above equation

A(s, s0) = hα(s, s0)i ,

is a real number between 0 and 1 and must satisfy

X
A(s, s0) = 1.
s0

It must rst be noted that

b
s
Y 1−sj
Nj j (1 − Nj ) ,
j=1

is now diferent from 0 or 1 and can be seen as the probability P (s) of Boolean
state s = (s1 , ..., sb ) considering all the dierent realizations. In the same way

b
Y s0 1−s0j
P (s0 ) = Nj j (1 − Nj ) ,
j=1

is the probability of state s0 = (s01 , ..., s0b ).

1.3. CONNECTION WITH THE BOLTZMANN EQUATION 23

In equilibrium, when state s can change to state s0 , i.e., when they have the
same mass and momentum, their probability must be identical

P (s) = P (s0 ) .
So

b b
Y s 1−sj
Y s0 1−s0j
Nj j (1 − Nj ) = Nj j (1 − Nj ) .
j=1 j=1

Thus, by taking the logarithm of the above equation

b b b
Y s 1−sj
X Nj X
ln Nj j (1 − Nj ) = sj ln + ln (1 − Nj ) ,
j=1 j=1
1 − Nj j=1
b b b
Y s0 1−s0j
X Nj X
ln Nj j (1 − Nj ) = s0j ln + ln (1 − Nj ) ,
j=1 j=1
1 − Nj j=1

and so

b b
X Nj X Nj
sj ln = s0j ln ,
j=1
1 − Nj j=1
1 − Nj

We know that s and s0 must satisfy the mass and momentum preservation
conditions
b
X b
X b
X b
X
sj = s0j and sj ej = s0j ej ,
j=1 j=1 j=1 j=1

Nevertheless, sj and s0j can have dierent values (0 or 1) for the same j. So
this identity is only satised if

Nj
ln ,
1 − Nj
is a collisional invariant.

1.3 Connection with the Boltzmann equation

In 1989, Higuera and Jimenez [45] proposed a linearization of the collision
term derived from the Boolean models, recognizing that this full form was un-
necessarily complex when the main purpose was to retrieve the hydrodynamic
equations. Starting from this linearized form, Higuera et al. [46] proved that
one can dene a Boltzmann equation with enhanced collisions, independently
of any underlying boolean microscopic dynamics. The collisions are dened by
a matrix, whose structure is essentially dictated by symmetry arguments.
24 CHAPTER 1. INTRODUCTION

Following this line of reasoning, Chen et al. [16] suggested replacing the
collision term by a single relaxation-time term, followed by Qian et al. [79] and
Chen et al. [13], who introduced a model based on the celebrated kinetic-theory
idea of Bhatnagar, Gross, and Krook (BGK) [4], but adding rest particles and
retrieving the correct incompressible Navier-Stokes equations, with third-order
non-physical terms in the local Mach number.
The single relaxation time BGK collision term describes the relaxation of the
distribution function to an equilibrium distribution. This discrete equilibrium
distribution was settled by writing it as a second-order polynomial expansion
in the particle velocity, with parameters adjusted to retrieve the mass density,
the local velocity, and the momentum ux equilibrium moments, which are
necessary conditions for satisfying the Navier-Stokes equations.

1.3.1 High-order lattice Boltzmann

Until some years ago, LBM was mostly restricted to isothermal, incom-
pressible ows. LB schemes for non-isothermal ows include higher-order terms
in the equilibrium distribution function, requiring to increase the lattice di-
mensionality, i.e., the number of vectors in the nite and discrete velocity set
ξi , i = 0, 1, ..., nb − 1. The connection between the LBE and the continuous
Boltzmann equation was rst established by He and Luo [40] and by [1], in 1997,
who directly derived the LBE from the Boltzmann equation for some widely
known lattice-Boltzmann equations: D2Q9, D2Q6, D2Q7, D3Q27. This was
performed by the discretization of the velocity space, using the Gauss-Hermite
and Gauss-Radau quadrature. Excluding the above mentioned lattices, the dis-
crete velocity sets obtained by this kind of quadrature do not generate regular
space lling lattices.
Shan et al. [82] and Philippi and co-workers [75], [84] reopened the prospect
of using the lattice Boltzmann method to simulate non-isothermal and/or high
Knudsen number ows through the direct resolution of the continuous Boltz-
mann equation, by establishing a systematic link between the kinetic theory
and the lattice Boltzmann method and enabling to determine the necessary
conditions for the discretization of the velocity space. The lattices obtained
by the method proposed by these authors (Figure 1.7), a prescribed abscissas
quadrature, proved to be stable in ows over a wide range of parameters by the
use of the high-order lattice-Boltzmann schemes [85], leading to velocity sets
which, when used in a discrete velocity kinetic scheme, ensure accurate recovery
of the high-order hydrodynamic moments and assuring increasingly higher or-
der of isotropy of the lattice tensors.Lattices found with the prescribed abcissas
method [85].
With the advent of faster computers, mesoscopic particle models, together
with numerical simulations, provide scientists with a very powerful tool to in-
vestigate new challenges in complex ow problems, both at the fundamental
and the applied levels. The recent years have witnessed many promising ad-
vances with the development of new theoretical frameworks and applications
of the mesoscopic particle models.
1.3. CONNECTION WITH THE BOLTZMANN EQUATION 25

Figure 1.7: Sets of velocity vectors for high-order LBM

Nowadays, LBM subjects range from the fundamentals of the kinetic the-
ory and quantum transport to applied subjects, including some ones such as
such as: magnetohydrodynamics (MHD); quantum lattice Boltzmann mod-
els; unstructured lattice Boltzmann equation (ULBE); non-Newtonian ows;
molecular dynamics. They also include lattice Boltzmann approaches for the
analysis of precursor lms; droplet spreading on solid surfaces; biopolymer
translocation; nano and microuidics; bioengineering and biophysics; granular
ows; combustion; hybrid methods and high performance computing.
26 CHAPTER 1. INTRODUCTION
Chapter 2

The Boltzmann equation

2.1 Introduction
The concept of `heat' or `internal energy' of a thermodynamic system as
the result of the kinetic energy of its individual molecules was long advocated
by philosophers and scientists such as Francis Bacon (1561-1626), John Locke
(1632-1704) and Thomas Hooke (1635-1703).  Heat is motion  in the words
of Francis Bacon (around 1600).  Heat is nothing else but the motion of the
particles that form a body  (Hooke, around 1650).

Figure 2.1: Ludwig Boltzmann (1844 1906)

Nevertheless, the caloric theory and mechanics evolved as independent

sciences since the invention of the thermoscope by Galileu in 1600, the built-up

27
28 CHAPTER 2. THE BOLTZMANN EQUATION

Figure 2.2: Front cover of 'Hydrodynamica' by Daniel Bernoulli, published in

1788.

of the mercury thermometer and the establishment of measurements methods

for the specic and latent heat in terms of the heat unity, the `calorie ' (the
0
amount of heat that raises the temperature of water by 1 C ).
In his `Hydrodynamica' (Figure 2.2), Bernoulli (1700-1782) proposed that
gases consist of great numbers of molecules moving in all directions, that their
impact on a surface causes the gas pressure and that heat transfer occurs from
regions with high energetic molecules to those with low energetic molecules.
In the end of the 18th century, Lavoisier [60], the father of chemistry and
the one of the most known and outstanding scientists in Europe, postulated the
principle of the conservation of the caloric as a  uid without mass that ows
from the hotter to colder bodies . Presently, this principle would be written as

δQ = dU
where U is the internal energy of a thermodynamic system, i.e., the energy of
its molecules.
On the mechanics side, people were worried with the movement of the plan-
ets and moons, i.e., with systems for which the mechanical energy is preserved.
Nowadays, the conservation of mechanical energy would be written as

δW = dEc + dEp

The rst half of the 19th century saw the birth of thermodynamics as a uni-
fying theory and the law of energy conservation was postulated by Helmholtz,
in 1847 [43],
δW + δQ = dU + dEc + dEp
2.1. INTRODUCTION 29

The Joule's principle of the mechanical equivalent of the caloric [49] can be
derived by integrating this equation, when a thermodynamic system performs
a cycle, returning to its initial state


(δW + δQ) = 0

James Joule equivalence principle and the XIX Century mechanical theory
of heat were highly inuenced by the atomistic philosophy.
Nevertheless, the main question that imposed severe diculties for the ac-
ceptance of the heat and work as equivalent forms of energy was only solved
by Clausius, a German scientist [18], [20].
Indeed, in the words of William Thompson  Nature has an arrow, and the
natural tendency is to transform vis-viva (or energy) into heat and not the
inverse .
So how to advocate the Joule's equivalence principle?
Clausius postulated this as the second law of thermodynamics, showing that
 if this was possible, heat would ow from a colder to a hotter body .
These two principles are stated in his 1867 book [20]: " Die Energie der
Welt ist constant. Die Entropie der Welt strebt einem Maximum zu."
But how to conciliate the dissipative second law of thermodynamics with
the conservative laws of mechanics? Indeed, in an isolated universe of particles
satisfying the Newton's laws, the mechanical energy is preserved and this sys-
tem is reversible: its previous states can always be attained by reversing the
time in the equations of motion.
So, if `heat ' has a `mechanical foundation ' how to explain irreversibility
based on the mechanical laws of motion?
1
Joule was tutored as a young man by the famous scientist John Dalton and
2
was strongly inuenced by chemist William Henry . His equivalence principle
was based on the atomistic philosophy and he also  explained the pressure of
gases by the impact of their molecules, and has calculated the velocity which
they must have in order to produce the pressure observed in particular gases.
Maxwell [67].
In accordance with Maxwell [67], Clausius paper  On the kind of motion
3
we call heat  is  a complete exposition of the molecular theory adopted in this
paper. After reading his investigation of the distance described by each molecule

1 The most important of all Dalton's investigations are those concerned with the beginning
atomic theory in chemistry. The main points of Dalton's atomic theory were: i) elements are
made of extremely small particles called atoms; ii) atoms of a given element are identical in
size, mass, and other properties; atoms of dierent elements dier in size, mass, and other
properties; iii) atoms cannot be subdivided, created, or destroyed; iv) atoms of dierent
elements combine in simple whole-number ratios to form chemical compounds; v) in chemical
reactions, atoms are combined, separated, or rearranged.
2 Henry W (1803) Experiments on the Quantity of Gases Absorbed by Water, at Dierent
Temperatures, and under Dierent Pressures. Philos Trans R Soc London 93:29274. doi:
10.1098/rstl.1803.0004 [44].
3 Originally published under the title  Ueber die Art der Bewegung, welche wir Winne
nennen , Annalen der Physik, Vol. 100, pp. 353-80 (1857); English translation in Philosoph-
ical Magazine, Vol. 14, pp. 108-27 (1857).
30 CHAPTER 2. THE BOLTZMANN EQUATION

between successive collisions, I published some propositions on the motions and

collisions of perfectly elastic spheres, and deduced several properties of gases,
especially the law of equivalent volumes, and the nature of gaseous friction. I
also gave a theory of diusion of gases, which I now know to be erroneous, and
there were several errors in my theory of the conduction of heat in gases which
M. Clausius has pointed out in an elaborate memoir on that subject .
Maxwell made a fully statistical model of gases, in which the velocities
were  distributed according to the same formula as the errors are distributed in
4
observations (the normal distribution). In 1873 Maxwell wrote :  The mod-
ern atomists have adopted a method which is, I believe, new in mathematical
physics, though it has long been in use in Statistics. The data of the statis-
tical method as applied to molecular science are the sums of large numbers of
molecular quantities. In studying the relation between quantities of this kind,
we meet with a new kind of regularity, the regularity of averages. (...)  If a
great many equal spherical particles were in motion in a perfectly elastic vessel,
collisions would take place among the particles, and their velocities would be
altered at every collision, so that after a certain time the vis viva will be di-
vided among the particles according to some regular law, the average number of
particles whose velocity lies between certain limits being ascertainable, though
the velocity of each particle changes at every collision .
But, the irreversibility as a principle for mechanical systems was only ex-
5,6
plained by Ludwig Boltzmann in the second half of the 19th century, based
on the concept of probability in mechanics.
When the exact position and velocity of each particle of a mechanical system
cannot be ascertain without errors, the best it is possible to do is to try to
prescribe the probability that this particle will have a certain velocity when it
is in a position x at time t.
7
As written by Max Planck in the Preface of his book :  The full content
of the second law can only be understood if we look for its foundation in the
known laws of the theory of probability as they were laid down by Clausius and
Maxwell, and further extended by Boltzmann .
In 1877, Boltzmann was in Graz, as professor of experimental physics. He
submitted two memoirs to the Academy of Sciences in Vienna. In the rst of
these, presented on January, 1877, "On Some Problems of the Mechanical The-
8

ory of Heat ", Boltzmann made an incisive analysis of the formula δQ/T ≤ 0
and argued that its validity was not based on the inherent laws of nature but
rather on the choice of the initial conditions. According to him, one had to

4 Maxwell, James Clerk (1873). 'Molecules', Nature : 437-441 [68].

5 L. Boltzmann. Vorlesungen uber Gas Theory, (1896) . J.A. Barth, Leipzig [ ? ]. This
book was translated in English by Stephen Brush and edited by University of California
Press. A Dover edition is presently available .
6 L. Boltzmann. Vorlesungen uber Gas Theory (1898) J.A. Barth. Leipzig [7]. This book
was translated in English by Stephen Brush and edited by University of California Press. A
Dover edition is presently available .
7 Planck, M. (1903). Treatise on Thermodynamics. Published in German in 1897. Trans-
lated to English by Alexander Ogg. London: Longmans, Green & Co [77].
8 Ludwig Boltzmann (1878). On some problems of the mechanical theory of heat Philo-
sophical Magazine. Series 5. Vol. 6, pp. 236-237.
2.2. THE BOLTZMANN EQUATION 31

Figure 2.3: Considering all the molecules inside a parallelepiped of sides ∆x,
∆y and ∆z, only f (x, ξ, t)∆x∆y∆z∆ξx ∆ξy ∆ξz have velocities between ξ and
ξ + ∆ξ .

assume on the basis of the kinetic theory of gases that the perfectly elastic
balls representing the molecules always tended to change their actual posi-
tions. Consequently, any conguration, however improbable, could conceivably
occur as time went on: "The calculus of probabilities teaches us precisely this:
any non-uniform distribution, unlikely as it may be, is not strictly speaking
impossible."
It now remained to give to this conclusion a rigorous derivation and quan-
titative applicability, which Boltzmann did in his memoir submitted to the
academy on October, 1877, "On the Relation between the Second Law of the
Mechanical Theory of Heat and the Probability Calculus with respect to the
Propositions about Heat-Equivalence ". These memoirs were published in Ger-
man in 1896 and 1898 [7]. He concluded that the entropy of a state is propor-
tional to the probability of the conguration of its component particles. In this
way, entropy became a purely statistical attribute of the overall system. As the
entropy increases, the conguration of the system becomes more probable.

2.2 The Boltzmann equation

In the following the Boltzmann equation is derived.
Let f (x, ξ, t)∆x∆ξ be the number of particles with velocities between ξ and
ξ + ∆ξ that are found at time t in the elementary volume ∆x between x and
x + ∆x, Figure 2.3. Consider now the development of f (x + ∆x, ξ + ∆ξ, t + ∆t)
in a Taylor series around (x, ξ, t),

f (x + ∆x, ξ + ∆ξ, t + ∆t) = f (x, ξ, t) + (∂t f ) ∆t

+ (∇x f ) · ∆x + (∇ξ f ) · ∆ξ + ..., (2.1)

or, dividing by ∆t with ∆t → 0,

32 CHAPTER 2. THE BOLTZMANN EQUATION

(∂t f ) + ξ · (∇x f ) + g · (∇ξ f )

f (x + ∆x, ξ + ∆ξ, t + ∆t) − f (x, ξ, t)
= lim , (2.2)
∆t→0 ∆t
∆ξ
where g=∆t is an acceleration.
Therefore, the above expression is null, if at time t + ∆t the particles that
were in x, with velocity ξ are, presently, in x + ∆x because they were trans-
ported with ξ , considering that these particules will also have the velocity
ξ + ∆ξ due to the acceleration g. So,

∂t f + ξ · (∇x f ) + g · (∇ξ f ) = 0. (2.3)

+
Nevertheless, there are (∂t f ) ∆x∆ξ∆t particles in the volume ∆x that
acquired the velocity ξ in the course of the time interval ∆t because they have
−
collided with other particles. On the other hand, (∂t f ) ∆x∆ξ∆t particles in
∆x, loss, in this time interval, the velocity ξ due also to the collisions they
suered with other particles.
So

+ −
∂t f + ξ · (∇x f ) + g · (∇ξ f ) = (∂t f ) − (∂t f ) . (2.4)
| {z } | {z } | {z }
stream term external forces colision term

In Boltzmann's model, the particles behave like billard balls, each collision
producing a sharp change in their velocity. A molecule may be thought as a
material point with a force eld around it. This force eld has a strong repulsion
kernel. When two molecules have a frontal collision, they will experience this
strong repulsion when their electronic orbits begin to intercept with themselves.
Around this repulsion kernel, molecules have a soft attraction eld produced
by electrostatic forces.
Considering σ to be a length related to a parameter possible to be identied
with the molecular diameter, the two particles will experience a repulsion force
when their centers are at a distance equal to σ. At this distance the attraction
and repulsion elds canceal themselves but the kinetic energy of the bullet with
respect to the target particle is not null and the center of the bullet particle will
penetrate into the σ -sphere, where it will be frained along the radial coordinate.
This radial frainage will produce a deviation in the bullet trajectory de-
picted in Figure (2.4). Using the label 1 for the bullet particle, the relative
velocity of the bullet with respect to the target before the collision is γ = ξ1 − ξ
when the target and bullet have the velocities ξ and ξ1 , respectively.
The collision term Ω takes account of these losses and gains in the particle
populations due to radical changes of the velocity ξ happening in a very small
time interval.
Considering that our reference is on a single target particle with velocity ξ,
this particle will change its velocity to ξ0 if it is hit by a bullet particle with
velocity ξ1 during the time interval ∆t.
2.2. THE BOLTZMANN EQUATION 33

Figure 2.4: Trajectory deviation of the bullet particle produced by the radial
frainage due to the repulsion forces. Vector
→
−
α is a unitary vector along γ −γ
0
.
Parameter b is the impact factor and angle ε is the azimuth angle of the collision
plane.

Consider a single bullet particle with velocity ξ1 and

γ = |γ| = |ξ1 − ξ| . (2.5)

z -ordinate of our reference frame be oriented along the direction

Let the
given by the vectorγ . The bullet particle will hit the target in the time interval
∆t when: a) the impact factor b is smaller than σ ; b) it is found in the
elementary volume bdεdbγ∆t in this time interval (Figure 2.5), where ε is the
azimuthal angle.
Inside the volume bdεdbγ∆t there are

f (x, ξ1 , t) dξ1 × bdεdbγ∆t , (2.6)

| {z } | {z }
bullet part. per unit vol.. volume in cyl. coord..

bullet particles.
Therefore, since there are f (x, ξ1 , t) ∆ξ∆x target particles inside ∆x, dur-
ing the time interval ∆t, the number of target particles that loss the velocity
ξ, during ∆t will be,

−
(∂t f ) ∆tdξ∆x
  2π  σ 
= f (x, ξ, t) f (x, ξ1 , t) bdεdbγdξ1 ∆t∆ξ∆x. (2.7)
ξ1 ε=0 b=0
34 CHAPTER 2. THE BOLTZMANN EQUATION

Figure 2.5: Elementary volume in cylindrical coordinates.

considering all the possible velocities ξ1 in the velocity space and all the impact
parameters b in each azimuthal plane ε,
So

  2π  σ 
−
(∂t f ) = f (x, ξ, t) f (x, ξ1 , t) bdεdbγdξ1 (2.8)
ξ1 ε=0 b=0
In the same way,

  2π  σ 
+
(∂t f ) = f (x, ξ 0 , t) f (x, ξ10 , t) b0 dεdb0 γ 0 dξ1 (2.9)
ξ1 ε=0 b=0

where ξ 0 , ξ10 must be such that ξ 0 , ξ10 → ξ, ξ1 after the collision. When the
potential energy is originated by a central eld, i.e., when this potential energy
is dependent only on the relative distance between two molecules Φ = Φ (krk),
0
it can be shown that γ 0 = γ and b = b. It is also possible ξ 0 , ξ10
to calculate
from known values of ξ, ξ1 using the mechanics of Newton.
In its nal form, the Boltzmann equation is written as,

∂t f + ξ · (∇x f ) + g · (∇ξ f )
  2π  σ 
f (x, ξ 0 , t) f (x, ξ10 , t)
 
= bdεdbγdξ1 (2.10)
ξ1 ε=0 b=0 −f (x, ξ, t) f (x, ξ1 , t)
The acceleration g is given by

g = g(e) + g(`d) (2.11)

(e) (`d)
where g is related to the external body forces and g is the attraction
force per unit mass due to the long-range intermolecular interaction. In the
Boltzmann original model, developed for rareed gases, there is no long-range
attraction between the molecules, g(`d) = 0, resulting in a model which equa-
tion of state is the one for ideal gases, P = nkT . In fact, in the classical
approach, developed for rareed gases, the long-distance attraction is included
into the collision term, the molecules being free of any interaction potential in
the time interval between any two collisions, when their centers are distanced
by |x1 − x| > σ .
2.3. THE EQUILIBRIM SOLUTION 35

2.3 The equilibrim solution

The Boltzmann equation

 
∂t f + ξ · (∇x f ) + g(e) + g(`d) · (∇ξ f ) = Ω, (2.12)

is an integro-diferential equation and the term,

  2π  σ
f (x, ξ 0 , t) f (x, ξ10 , t)
  
Ω= bdεdbgdξ1 , (2.13)
ξ1 ε=0 b=0 −f (x, ξ, t) f (x, ξ1 , t)

gives the net balance between the molecules that acquire the velocity ξ and the
ones that loss this velocity. In equilibrum conditions this balance must be null,
otherwise the distribution f would vary in the course of the time and in the
space. So,

 0

f x, ξ , t f (x, ξ10 , t) = f (x, ξ, t) f (x, ξ1 , t) , (2.14)

or

ln f (x, ξ 0 , t) + ln f (x, ξ10 , t) = ln f (x, ξ, t) + ln f (x, ξ1 , t) , (2.15)

meaning that

ln f = ln f (eq) = collisional invariant. (2.16)

The collisional invariants are

m = mass, (2.17)

mξ = momentum, (2.18)
1 2
mξ = kinetic energy. (2.19)
2
Therefore, any linear combination of these invariants is also a collisional
invariant,

ln f (eq) = A + B · ξ + Cξ 2
2
= a + b (ξ − c) , (2.20)

or

2 2
f (eq) = ea eb(ξ−c) = deb(ξ−c) . (2.21)

Parameters d, b e c are determined from

36 CHAPTER 2. THE BOLTZMANN EQUATION


n = f (eq) dξ = number density of molecules, (2.22)

1
u = f (eq) ξdξ = mean molecular velocity, (2.23)
n

1 1 2
ec,f = f (eq) m (ξ − u) dξ = mean peculiar kinetic energy.(2.24)
n 2
In fact, for a uid in equilibrium the quantities

n= f dξ, (2.25)


1
u = hξi = f ξdξ, (2.26)
n
must be constants and given by replacing the distribution f by the distribution
the system has at equilibrium, i.e., f (eq) .
The third integral deserves a more lengthy discussion. The mean kinetic
energy per molecule is given by

1
 1 1
ec = m ξ 2 = f mξ 2 dξ. (2.27)
2 n 2
Writting,
2
ξ 2 = (ξ − u) + 2 (ξ − u) · u + u2 , (2.28)

it can be easily seen that this kinetic energy has two components

  
J 1 1 2 1
ec = f m (ξ − u) dξ + mu2 . (2.29)
molecule n 2 2
| {z }
| {z }
=ec,f advection energy

The thermodynamic internal energy, e, per molecule, is the sum of the the
peculiar kinetic energy ec,f , i.e., the energy due to the molecular random motion
of the molecules and the potential energy due to the intermolecular interaction
among these molecules,

Φ
e = ec,f + , (2.30)
n
where Φ is the intermolecular potential energy per unit volume.
For thermodynamic systems in equilibrium, the thermodynamic internal
energy, satises,

   
∂P
de = cv dT + T − P dv . (2.31)
| {z } ∂T v
energy due to molecular motion | {z }
intermolecular potential energy Φ
2.3. THE EQUILIBRIM SOLUTION 37

For ideal gases, with R=universal gas constant,

kT
P = nkT = , (2.32)
v
where,

R
k= , (2.33)
N
9
N being the Avogadro number and k= Boltzmann constant. So, the second
term in Eq. (2.31) is null and,

de = cv dT, (2.34)

meaning that solely the molecular motion contributes to the thermodynamic

energy. Nevertheless, for real uids, the thermodynamic energy must take the
intermolecular potential energy into account. For a uid with, e.g., a van der
Waals equation of state

kT a
P = − , (2.35)
v − b v2
the parameter ”a” is related to the intermolecular forces and we get

a
de = cv dT + dv. (2.36)
v2
In both cases, the molal heat cv is a molecular parameter related to the
10
motion degrees of freedom of the molecules ,

9 The Avogadro constant is named after the early nineteenth century Italian scientist
Amedeo Avogadro, who, in 1811, rst proposed that the volume of a gas (at a given pressure
and temperature) is proportional to the number of atoms or molecules regardless of the
nature of the gas. The French physicist Jean Perrin in 1909 proposed naming the constant
in honor of Avogadro. Perrin won the 1926 Nobel Prize in Physics, in a large part for his
work in determining the Avogadro constant by several dierent methods.
The value of the Avogadro constant was rst indicated by Johann Josef Loschmidt who, in
1865, estimated the average diameter of the molecules in air by a method that is equivalent
to calculating the number of particles in a given volume of gas. This latter value, the number
density of particles in an ideal gas, is now called the Loschmidt constant in his honour, and
is approximately proportional to the Avogadro constant. The connection with Loschmidt is
the root of the symbol L sometimes used for the Avogadro constant, and German language
literature may refer to both constants by the same name, distinguished only by the units of
measurement.
Accurate determinations of Avogadro's number require the measurement of a single quan-
tity on both the atomic and macroscopic scales using the same unit of measurement. This
became possible for the rst time when American physicist Robert Millikan measured the
charge on an electron in 1910. The charge of a mole of electrons is the constant called the
Faraday and had been known since 1834 when Michael Faraday published his works on elec-
trolysis. By dividing the charge on a mole of electrons by the charge on a single electron the
value of Avogadro's number is obtained. Since 1910, newer calculations have more accurately
determined the values for Faraday's constant and the elementary charge. (from Wikipedia )

10 With monatomic gases, thermal energy comprises only translational motions. Trans-
lational motions are ordinary, whole-body movements in 3D space whereby particles move
38 CHAPTER 2. THE BOLTZMANN EQUATION

cv = cv,transl + cv,rot + cv,vib . (2.37)

For molecules with solely translational degrees of freedom (our case )

D
cv = cv,transl = k, (2.38)
2
where D= space dimension.
Therefore, by integrating Eq. (2.36), we get for a van der Waals equation
of state,

 
D 1
e (T, v) = kT + −a (in J/molecule), (2.39)
2 v
plus a constant C (T0 , v0 ) that depends only on the initial state.
This means that, in a molecular basis,


1 1 2 D
ec,f = f (eq) m (ξ − u) dξ = kT (in J/molecule). (2.40)
n 2 2
about and exchange energy in collisionslike rubber balls in a vigorously shaken container
(see animation here). These simple movements in the three X, Y, and Zaxis dimensions of
space means individual atoms have three translational degrees of freedom. A degree of free-
dom is any form of energy in which heat transferred into an object can be stored. This can
be in translational kinetic energy, rotational kinetic energy, or other forms such as potential
energy in vibrational modes. Only three translational degrees of freedom (corresponding to
the three independent directions in space) are available for any individual atom, whether it
is free, as a monatomic molecule, or bound into a polyatomic molecule.
As to rotation about an atom's axis (again, whether the atom is bound or free), its energy
of rotation is proportional to the moment of inertia for the atom, which is extremely small
compared to moments of inertia of collections of atoms. This is because almost all of the
mass of a single atom is concentrated in its nucleus, which has a radius too small to give a
signicant moment of inertia. In contrast, the spacing of quantum energy levels for a rotating
object is inversely proportional to its moment of inertia, and so this spacing becomes very
large for objects with very small moments of inertia. For these reasons, the contribution
from rotation of atoms on their axes is essentially zero in monatomic gases, because the
energy-spacing of the associated quantum levels is too large for signicant thermal energy
to be stored in rotation of systems such small moments of inertia. For similar reasons, axial
rotation around bonds joining atoms in diatomic gases (or along the linear axis in a linear
molecule of any length) can also be neglected as a possible "degree of freedom" as well, since
such rotation is similar to rotation of monatomic atoms, and so occurs about an axis with a
moment of inertia too small to be able to store signicant heat energy.
In polyatomic molecules, other rotational modes may become active, due to the much
higher moments of inertia about certain axes which do not coincide with the linear axis of
a linear molecule. These modes take the place of some translational degrees of freedom for
individual atoms, since the atoms are moving in 3-D space, as the molecule rotates. The
narrowing of quantum mechanically-determined energy spacing between rotational states
results from situations where atoms are rotating around an axis that does not connect them,
and thus form an assembly that has a large moment of inertia. This small dierence between
energy states allows the kinetic energy of this type of rotational motion to store heat energy
at ambient temperatures. Furthermore (although usually at higher temperatures than are
able to store heat in rotational motion) internal vibrational degrees of freedom also may
become active (these are also a type of translation, as seen from the view of each atom). In
summary, molecules are complex objects with a population of atoms that may move about
within the molecule in a number of dierent ways (see animation at right), and each of these
ways of moving is capable of storing energy if the temperature is sucient. (from Wikipedia)
2.4. EXERCISES 39

Also, for a van der Waals equation of state,

D
e (T, n) = kT + (−an) (in J/molecule). (2.41)
2
Eq. (2.40) is, in fact, the denition of temperature for systems composed
of molecules with only translational degrees of freedom. For a van der Waal
equation of state, the intermolecular potential energy per unit volume becomes
Φ = −an2 and is always negative. When the number density of molecules
increases, meaning a larger number of molecules in the same volume, this energy
increases in absolute value, meaning that the molecules are strongly linked,
When the number density decreases, Φ → 0. For an ideal gas, the molecules
are free from the intermolecular forces and Φ = 0.

With these restrictions Eqs (2.22-2.24) give for the equilibrium distribution,
the Maxwell-Boltzmann distribution

 m D/2 − (ξ−u)2
2kT
f (eq) = n e m . (2.42)
2πkT

2.4 Exercises
1- Consider a two-dimensional system of particles and the transformation
ξx , ξy → ξf x = ξx − ux , ξf y = ξy − uy . We know that dξx dξy = Jdξf x dξf y
where J is the Jacobian of the transformation

 ∂ξ ∂ξx 

x
∂ (ξx , ξy ) 
 ∂ξf x ∂ξf y 
J= =  ∂ξy ∂ξy 
∂ (ξf x , ξf y )  ∂ξ ∂ξ

fx fy

Show that J = 1. ( Hint: remember that the molecular velocity and the
position are independent variables )

2- Maxwell [67] made a fully statistical model of gases, in which, in equi-

librium, the velocities were  distributed according to the same formula as the
errors are distributed in observations (the normal distribution)

2
ξf
f eq = Ae− B ,
where A and B can be determined by considering that the number density of
molecules is given by

  ∞  ∞  ∞ 2 +ξ2 +ξ2
ξf,x f,y f,z
n= eq
f dξ = A e− B dξx dξy dξz ,
−∞ −∞ −∞

and that the kinetic energy related to the `molecular agitation' is given by
40 CHAPTER 2. THE BOLTZMANN EQUATION


1
ec,f = f eq mξf2 dξ
2
 ∞  ∞  ∞ 2 +ξ2 +ξ2
ξf,x f,y f,z 1
Ae− 2 2 2


= B m ξf,x + ξf,y + ξf,z dξx dξy dξz .
−∞ −∞ −∞ 2

FindA and B in terms of the number density n and the mean kinetic energy
of uctationsec,f . What are the independent variables of n and ec,f ?
Hint. Dene new integration variables such as ξ√f,x
B
ξ
= x, √f,y
B
ξ
= y, √f,z
B
=z
and remember that

 ∞  ∞
2 √ 2 1√
e−x dx = π, e−x x2 dx = π.
−∞ −∞ 2

2- Boltzmann derived the equilibrium distribution in the form

2 2
f (eq) = ea e−b(ξ−c) = de−b(ξ−c) .
Find d, b and c using


n = f (eq) dξ = number density of molecules,

1
u = f (eq) ξdξ = mean molecular velocity,
n

1 1 2
ec,f = f (eq) m (ξ − u) dξ = mean peculiar kinetic energy.
n 2

3- Consider a material volume V(t), i.e., a volume that is composed by the

same molecules. The total number of molecules in this volume is


N= n (x, t) dV,
V (t)

and the time derivative of N is null


dN d
= ndV = 0,
dt dt V (t)

Use the Reynolds transport theorem

  
d ∂n
ndV = dV + nu · en dA
dt V (t) V (t) ∂t A(t)

where A (t) is the area of the surface around V(t) and en is the unitary vector
pointing outside A (t), to show that
2.4. EXERCISES 41

∂ρ
+ ∇ · (ρu) = 0 (2.43)
∂t
where ρ = nm and m is the mass of each molecule.

4- Equation 2.43 can be also written as

∂ρ
+ u · ∇ρ = −ρ∇ · u
∂t
Consider now a Taylor expansion of ρ (x + ∆x, t + ∆t) around ρ (x, t)

∂ρ
ρ (x + ∆x, t + ∆t) = ρ (x, t) + ∆t + ∇ρ · ∆x + ...
∂t
In the limit ∆t → 0 we get

ρ (x + ∆x, t + ∆t) − ρ (x, t) ∂ρ

= + u · ∇ρ
∆t ∂t
Show that when ∇ · u = 0, the density does not vary along a uid trajectory
in the physical space (we say that the ow is incompressible).

5- A system of particles is described by giving the expected value f (x, ξ, t) =

f (x, y, z, ξx , ξy , ξz , t) of the amount of molecules inside an elementary volume
dx =dxdydz that have velocities between ξx and ξx + dξx , ξy and ξy + dξy ,
ξz and ξz + dξz . Based on Exercise 4 and using words, try to give a physical
meaning for

f (x + ∆x, ξ + ∆ξ, t + ∆t) − f (x, ξ, t)

lim ,
∆t→0 ∆t
when this system is in equilibrium.

6- The pressure tensor P can be written as the momentum m (ξα − uα ) that

is transported along a direction β with the velocity uctuation (ξβ − uβ )

Pαβ = f m (ξα − uα ) (ξβ − uβ ) dξ. (2.44)

a) Show that in equilibrium, this tensor is isotropic

Pαβ = P δαβ
and reduces to the thermodynamic pressure P = nkT (the Clapeyron equation ).
b) Under the light of Eq. (2.44) try to describe the meaning of an isotropic
tensor, using your own words.
c) Consider a 2D case and the integral

 ∞  ∞
f eq mξf x ξy dξx dξy
ξy =0 ξfx =−∞
42 CHAPTER 2. THE BOLTZMANN EQUATION

giving the momentum mξf x that is transported into the positive direction ξy
and the integral

 0  ∞
f eq mξf x ξy dξx dξy
ξy =−∞ ξfx =−∞

giving the momentum mξf x that is transported into the negative direction ξy
and answer the question:
In equilibrium, the momentum mξf x is transferred along direction y ?...Justify
your answer.
Chapter 3

A rst insight into LBM

In this lecture, we will have a rst insight into the LB discretization prob-
lem enabling the Boltzmann equation to be solved by numerical methods in a
1
computer .

3.1 From the Boltzmann equation to LB schemes

Consider the Boltzmann equation in the form

∂t f + ξ · (∇x f ) = Ω, (3.1)

where the term Ω is given by

 
Ω = Ω(sd) − g(e) + g(`d) · (∇ξ f ) , (3.2)

with the short distance repulsion term written as,

  2π  σ
Ω(sd) = (f 0 f10 − f f1 ) bdεdbgdξ1 .
ξ1 ε=0 b=0

It is possible to make f = f eq in the evaluation of ∇ξ f on the right hand

side of Eq. (3.2) without any eects on the macroscopic equations,

(ξ − u)
∇ξ f = ∇ξ f eq = − kT
f eq .
m
So we get a kinetic model for Ω
  (ξ − u)
Ω = Ω(sd) + g(e) + g(`d) · kT f eq .
m
By `kinetic model' we mean that the Boltzmann equation is replaced by a
model that is able to retrieve the main, or some of the main properties of the

1 Readers well acquainted with the rudiments of the LB method can avoid this lecture.

43
44 CHAPTER 3. A FIRST INSIGHT INTO LBM

Figure 3.1: P.L. Bhatnagar (1912-1976)

Boltzmann equation. In LBM, the more widely used model for Ω(sd) is the
2
BGK model [4]

f eq − f
Ω(sd) = , (3.3)
τ
meaning that the Boltzmann collision term is replaced by a relaxation term
with a single relaxation time τ. Therefore, this model preserves the main eect
of collisions leading the distribution of particles towards equilibrium, when a
uid is in a non-equilibrium state.
We will talk more about kinetic models in the following lectures. In this
lecture we focus our attention on LBM discretization and consider Ω as a known
function Ω (f, f eq ).
Discretization means to replace the entire continuous physical space, repre-
sented by the continuous variable x by some points xi and the entire velocity

2 "In 1951, Bhatnagar went to Harvard University, Cambridge as a Fullbright scholar

for two years. This handsome tall scholar from India was often mistaken in the University
corridors for a student. Once he took his place at the lecture rostrum, the students realized
that he was indeed a senior faculty. He lectured on mathematical theory of gases based on the
mathematically formidable book by S. Chapman and T.G. Cowling. At Harvard, he produced
two very important publications: (1) a book Stellar Interiors jointly with D.H. Menzel and
H.K. Sen, published in the International Astrophysical Series and (2) a research paper in
Physical Review in 1954 which contained the famous BGK (Bhatnagar, Gross, Krook) model.
The Bolzmann equation, governing the evolution of a state of molecules in gases, contains
an extremely complicated integral term . Faced with the reality that the Boltzmann equation
was too dicult to handle due to this collision integral term, Bhatnagar, Gross and Krook
used their deep understanding of relaxation process of a swarm of molecules towards an
equilibrium state to replace this term by a much simpler term free, which has since been
used as alternative to the Boltzmann equation in solving problems in rareed gas dynamics,
plasma physics and the kinetic theory itself. " Phoolan Prasad, P.L. Bhatnagar and the
BGK Model, 19th P.L. Bhatnagar Memorial Award Lecture - 2006, delivered at 72nd Annual
Conference of the Indian Mathematical Society at Jabalpur, 27 - 30 December, 2007.
3.1. FROM THE BOLTZMANN EQUATION TO LB SCHEMES 45

Figure 3.2: D2Q9 [79] and RD2Q9 [74] sets of lattice vectors.

space represented by the continuous variable ξ by some velocity vectors ξi ,

i = 0, ..., nb − 1, for which the distribution f will be calculated in dierent
times separated by a time interval δ. Figure 3.2 represents two versions of
a 9-velocity lattice, suitable for solving two-dimensional incompressible ow
problems.
Therefore, considering a single velocity ξi from this set of nb velocities, the
Boltzman equation can be written as

∂t fi + ξi · (∇x fi ) = Ωi ,
where fi = f (x, ξi , t) indicates the value of f for ξ = ξi , i.e., the packet of
particles with velocity ξi that are found in the point x at time t.
We further consider that the uid is free from external (no gravity ) and
intermolecular forces (ideal gas assumption ), or that

fieq − fi
Ωi =
τ
At the time t + δ, the population fi of particles along direction i can be
found using a Taylor series

∂fi 1 ∂ 2 fi
fi (x, t + δ) = fi (x, t) + δ + δ 2 2 + ...,
∂t 2 ∂t
and will be known from the value this packet had at time t, when all the time
∂fi ∂ 2 fi
derivatives
∂t , ∂t2 , ...are known at time t.
This is not the case. Nevertheless, when δ → 0, it is possible to neglect all
the derivatives of order 2 and higher.

∂fi
+ O δ2


fi (x, t + δ) = fi (x, t) + δ
∂t
giving

∂fi fi (x, t + δ) − fi (x, t)

= + O (δ)
∂t δ
In the same way, for calculating the value fi in a point x + ∆x, at the time
t + ∆t
46 CHAPTER 3. A FIRST INSIGHT INTO LBM

Figure 3.3: I lustration of the collision-propagation scheme in LBM. The amount of

particles in each direction is represented by the thickness of the vector pointing in
that direction. The collision step tries to equilibrate the dierent populations and
some particles along the most peopled populations, in the northeast directions, are
thrown to the other directions.

∂fi ∂fi 
2

fi (x + ∆x, t + δ) = fi (x, t) + δ + ∆x · + O (∆x) , δ 2 .
∂t ∂x
Now consider that the physical and velocity space are coupled through

∆x = ξi δ,
meaning that after each time step δ , the population fi of particles is transferred
from site x to site x + ∆x This is the main feature of almost all LB methods.
In this case,

 
∂fi ∂fi
+ O δ2 ,


fi (x + ξi δ, t + δ) = fi (x, t) + δ + ξi ·
∂t ∂x
or

 
∂fi ∂fi fi (x + ξi δ, t + δ) − fi (x, t)
+ ξi · = + O (δ) .
∂t ∂x δ
The discrete Boltzmann equation will, then, be written as

fieq − fi
+ O δ2 ,


fi (x + ξi δ, t + δ) = fi (x, t) + ∗
(3.4)
τ
where τ∗ is a dimensionless relaxation time
∗
τ = τ /δ .
The term

|ξi | δ
,
|∆x|
3.1. FROM THE BOLTZMANN EQUATION TO LB SCHEMES 47

3
is the Courant-Friedrich-Lewy number , CF L, and in the present numerical
scheme CF L = 1. This condition implies that at the time t + δ the particles
with velocity ξi will be found at the site x + ∆x , at the time t + δ . This scheme
is at the origin of the collision propagation LB method, which is an explicit
numerical method because the interacton term Ωi is evaluated at time t.
In LB simulation, Eq. (3.4) is usually replaced by

eq
fi − fi
+ O δ2 ,


f i (x + ei h, t + δ) = f i (x, t) + ∗
(3.5)
τ
where

3
fi ξ
fi = ,
n0
is a dimensionless population, n0 being a number density of reference and

r
kT
ξ= ,
m
4
is related to a mean molecular speed .
Now, we need to relate the molecular velocity ξi with the lattice vectors ei
and the Maxwell-Boltzmann equilibrium distribution f eq (ξi ) with the discrete
eq
equilibrium populations fi .
A fundamental quantity in LBM is the lattice speed c dened as

h
c= ,
δ
h being the orthogonal distance between any two contiguous sites, meaning that
the particles will be transported along direction i with a velocity cei . Since
these particles have a molecular velocity ξi , we can conclude that the lattice
speed is such that

ξi = cei .
3 In mathematics, the CourantFriedrichsLewy condition (CFL condition) is a necessary
condition for convergence while solving certain partial dierential equations (usually hyper-
bolic PDEs) numerically (it is not in general a sucient condition). It arises when explicit
time-marching schemes are used for the numerical solution. As a consequence, the time
step must be less than a certain time in many explicit time-marching computer simulations,
otherwise the simulation will produce wildly incorrect results. The condition is named after
Richard Courant, Kurt Friedrichs, and Hans Lewy who described it in their 1928 paper [21].

4 The mean molecular speed is given by the square root of

D E f eq ξf2 dξf
ξf2 =
eq ,
f dξf
and can be calculated as

√
r
kT
D .
m
48 CHAPTER 3. A FIRST INSIGHT INTO LBM

Another important parameter in LBM is the scaling factor as , which relates

the dimensionless molecular velocity

ξi
ξ0,i = ,
ξ
with the lattice vectors ei ,

ξ0,i = as ei .
The scaling factor as can be shown to be a lattice parameter and will have
dierent values when we change the discrete set of lattice vectors. It is more
usual to write it as

1
cs = ,
as
since cs can be shown to be the dimensionless speed of sound in isothermal ow
in LBM simulations.
From the above equations,

ξ
ξi = cei = ξξ0,i = ei
cs
so

r
kT
ξ= = ccs .
m
Now, we drive our attention to the relation between the equilibrium popu-
eq
lations fi and the Maxwell-Boltzmann equilibrium distribution f eq (ξ),
 m D/2 − (ξ−u)2
eq 2kT
f (ξ) = n e m .
2πkT
Since the MB equilibrium distribution refers to an innite velocity space,
eq
we cannot require that the discrete populations fi reproduce this distribution.
eq
What it is possible to require [69] is that the rst moments of fi are the same
rst moments of f eq (ξ).
eq
So, the discrete populations fi are required to retrieve the number density
of molecules

 X eq
n= f eq (ξ) dξ = n0 fi ,
i
or


n 1 X eq
ρ∗ = = f eq (ξ) dξ = fi . (3.6)
n0 n0 i

In the same way, we require that

3.1. FROM THE BOLTZMANN EQUATION TO LB SCHEMES 49


1 X eq
ρ∗ u∗ = f eq (ξ) ξdξ = f i ei . (3.7)
n0 c i

where u∗ = u/c
We also require that


1 X eq
ρ∗ u∗α u∗β ∗
+ P δαβ = f eq (ξ) mξα ξβ dξ = f i ei,α ei,β , (3.8)
mn0 c2 i

where

P n kT 1
P∗ = 2
= = ρ∗ c2s
mn0 c m c2
n0 |{z}
|{z}
=ρ∗ =c2 c2s

which is the dimensionless equation of state for ideal gases the Clapeyron
equation written in dimensionless variables.

The above conditions are all we need to satisfy when dealing with second
order models like the D2Q9 or D3Q27.

It is convenient to write Maxwell-Boltzmann (MB) distribution [67], [7] in

the following form

1 (ξ0 −u0 )2
f eq = n D
e− 2 ,
D/2
(2π) ξ

where u0 = u/ξ is a dimensionless local velocity.

In its simplest form, the interaction term Ω is modeled by a BGK single

relaxation term [4]

f − f eq
Ω=− , (3.9)
τ
where τ is a relaxation time.
eq
Now, the dimensionless form f of f eq may be developed in terms of the
Hermitian basis of H as

D
eq f eq ξ 1 (ξ0 −u0 )2
f = = ρ∗ D/2
e− 2
no (2π)
ξ2
− 20 (θ)
e X aeq,nx ,ny ,nz
= ρ∗ D/2
× Hn(θ)
x ,ny ,nz
(ξ0 ) ,
(2π) nx !ny !nz !
θ

where ρ∗ = n/n0 = ρ/ρ0 and nα is the number of times the index α = x, y or

z appears repeated, nx + ny + nz = θ
50 CHAPTER 3. A FIRST INSIGHT INTO LBM

In 2D, the rst Hermite polynomial tensors are,

H (0) = 1,
(1)
H1 = Hx(1) = ξ0,x ,
(1)
H2 = Hy(1) = ξ0,y (3.10)
(2) (2)
H1,1 = Hxx = ξ0,x ξ0,x − 1,
(2) (2)
H1,2 = Hxy = ξ0,x1 ξ0,y , (3.11)
(2) (2)
H2,2 = Hyy = ξ0,y ξ0,y − 1 (3.12)

Due to the orthogality of Hermite polynomials, the dimensionless equilib-

(θ)
rium moments aeq(nx ,ny ,nz ) are given by

(θ)
aeq(nx ,ny ,nz ) (u0 )

1 (ξ0 −u0 )2
= D/2
e− 2 Hn(θ)
x ,ny ,nz
(ξ0 ) dξ0 .
(2π)
In components notation, the rst equilibrium moments are

a(0)
eq = 1,

a(1)
eq,α = u0,α ,
(2)
aeq,αβ = u0,α u0,β ,

Consider a 2D example. When the Hermite polynomial expansion is trun-

cated after second order terms,

2
ξ0 " #
(0) (1) (1) (1) (1)
eq e− 2 aeq H (0) + aeq,x Hx + aeq,y Hy
f = D ρ∗ 1 (2) (2) 1 (2) (2) 1 (2) (2)
(2π) 2 + 2! aeq,xx Hxx + 2! aeq,yy Hyy + + 1!1! aeq,xy Hxy
2
ξ0
e−
 

1 + 1u0,x ξo,x + u0,y
ξo,y
2
∗
= ρ 1 2 2
(2π) 2
D
+ 2! u0,x ξo,x − 1 + 2! 2
u20,y ξo,y − 1 + u0,x u0,y (ξo,x ξo,y )

In its discrete form

2
ξ0,i
e−
 

1 + 1u0,x ξ0,i,x + u0,y
ξo,i,y
eq 2
∗
fi = D ρ 1 2 2 2 2
(2π) 2 + u
2! 0,x ξ 0,i,x − 1 + 2! 0,y ξ0,i,y − 1 + u0,x u0,y (ξ0,i,x ξ0,i,y )
u
u∗ e u∗ ei,y
 
1 + xc2i,x + yc2
= W i ρ∗  ∗2
s

1 u∗y2 2
s

u∗x u∗y .
1 ux 2 2 2
+ 2 c4 ei,x − cs + 2 c4 ei,y − cs + c4 ei.x ei.y
s s s

This discrete equilibrium distribution can be written in a compact form as

3.2. SIMULATION IN LBM 51

u∗ e
" #
eq ∗ 1 + αc2iα +
fi = Wi ρ 1 ∗ ∗
s
2

, (3.13)
2c4 uα uβ eiα eiβ − cs δαβ
s

where repeated index means summation, in accordance with Einstein notation.

The weights Wi = W (|ei |) follows from the quadrature conditions Eqs.(3.6
- 3.8) when the polynomial given by Eq. (3.13) is required to retrieve the same
integrals of the MB distribution. For the D2Q9 LBE we have four unknows:
W0 related to the rest population, W1 related to the populations along the main
axes,W2 assigned to the populations along the diagonal directions and, nally,
the scaling factor cs , which until, presently must be considered as another
unknown. Eqs.(3.6 - 3.8) represent, in fact, 6 restrictions: a rst equation
for the density, two equations for the velocities u∗x and u∗y and 3 equations
for the second order moments. Nevertheless only four of these equations are
independent and their solution gives [75],

1 16 4 1
cs = √ , W0 = , W1 = , W2 = .
3 36 36 36
Since the particles behave as the molecules of an ideal gas as the long-
range intermolecular forces were suppressed the incompressibility condition
must be assured by keeping the Mach number as small as possible. The Mach
number is dened as the ratio between a uid characteristic speed U and the
speed of sound Us ,
U
Ma = ,
Us
which can also be written as

U∗
Ma = ,
cs
where, as before, U ∗ = U/c and cs = Us /c are dimensionless speeds. The LBM
simulation of incompressible ows with second order models requires M a ≤ 0.1.

3.2 Simulation in LBM

LB simulation has three steps.
a) Calculation of the equilibrium distribution
eq
. In LBM simulation,
we must calculate the equilibrium populations fi at each time step and in
every site in the physical domain. In isothermal problems, this is accomplished
by rst calculating the dimensionless density and velocity


X 1
ρ∗ = fi = f dξ
i
n0

X 1
ρ∗ u∗ = f i ei = f ξdξ
i
n0 c
52 CHAPTER 3. A FIRST INSIGHT INTO LBM

replacing these variables into Eq. (3.13).

Therefore, when the external and intermolecular forces are neglected and
the repulsion term Ω(sd) is modelled using a BGK relaxation term the LB
method proceeds in accordance with two steps.
b) Collision step: Populations fi are pictured in Figure 3.3 with lines with
thicknesses related to the amount of particles of that population. Population
f1 came from site D, f2 from site G and so on, where these populations of
particles were found at the previous time step. So, with the exception of the
zero-velocity population, f 0, the information we have in site x at time step t is
the information that came from the neighbor sites x + ei h at time step t − δ.
Then, populations fi are recalculated in each site x using the local information
in this site at time t,
eq
out f i (x, t) − f i (x, t)
fi (x, t + δ) = f i (x, t) + , (3.14)
τ∗
and the value we have for relaxation time τ∗ = 1 τ ∗. When all the outgoing
∗ 1 ∗ 1
populations are the equilibrium ones and the lower limit for τ
2 . For τ ≤ 2 is
the numerical scheme is unstable. In fact, it is possible to show that, in this
∗
rst order scheme, the dimensionless kinematic viscosity ν is related to the
relaxation time by

 
∗ ∗ 1
ν = c2s τ − .
2
On the other hand, since the Reynolds number is dened as

U ∗ (L/h) U ∗ (L/h)
Re = =
,
ν∗ c2s τ ∗ − 21
where L is the linear dimension of the physical domain. Since U∗ must be
kept as small as possible, if we want to simulate larger Re number ows, the
relaxation time τ∗ 1
2 as possible for reducing the
must be kept as close to
computational cost represented by an excessively linear number of sites N =
L/h.
Another limit in LBM simulation is represented by the Knudsen number.
In kinetic theory it is dened as

`
Kn =
L
where ` is the mean free path,

` = ξτ ∗ δ = cs τ ∗ h.
So, in LBM the Kn number is, usualy, dened as

cs ν ∗
Kn = .
N
3.3. SOME FEATURES OF THE D2Q9 LBE 53

considering that the limit

1
τ∗ =
2 is due to the use of a rst order discretization
of the stream term, Eq. (3.4) and that second and higher order discretization
∗ ∗ ∗ ∗ 1 5
2


results in ν = τ (and not ν = cs τ −
2 .

In continuum problems it is required that the Knudsen number be as small

as possible, the accorded limit being Kn = 0.001. This condition can be satis-
ed by increasing the linear number of sites N or by reducing the dimensionless
kinematic viscosity ν∗. The second option is to be chosen when it is wanted to
6
reduce the computational costs .

c) Propagation step: after the collision step, the new values of the distri-
out
butions f i are propagated along the i-directions

out
f i (x + ei h, t + δ) = f i (x, t + δ) , (3.15)

and we return to step a).

3.3 Some features of the D2Q9 LBE

The equilibrium distribution for the D2Q9 LBE is given by a second order
expansion in the orthogonal Hermite polynomials [84] H0 =1, H1,α = ξ0i,α =
eiα 1


cs , H2,αβ = ξ0i,α ξ0i,β − δαβ = c2s eiα eiβ − c2s δαβ giving

u∗ e
" #
eq 1 + αc2iα +
fi = W i ρ∗ 1 ∗ ∗
s
2

,
2c4 uα uβ eiα eiβ − cs δαβ
s

These populations will be written as

5 This question will be dicussed further in these lectures.

6 In fact, working with very low values of ν ∗ is a very actual challenge in LBM, considering
the stability issues of LB numerical schemes when τ∗ approaches its limiting value.
54 CHAPTER 3. A FIRST INSIGHT INTO LBM

 
4 ∗ 3 ∗2 3 ∗2
f¯0eq = ρ 1 − uy − ux
9 2 2
 
1 ∗ 3 ∗2
f¯1eq = ∗ ∗2
ρ 3ux + 3ux − uy + 1
9 2
 
1 ∗ 3 ∗2
f¯2eq = ∗
ρ 3uy − ux + 3uy + 1 ∗2
9 2
 
1 ∗ 3 ∗2
f¯3eq = ∗2 ∗
ρ 3ux − 3ux − uy + 1
9 2
 
1 ∗ 3 ∗2
f¯4eq = ∗2
ρ 3uy − ux − 3uy + 1 ∗
9 2
1 ∗
f¯5eq ρ 3u∗x + 3u∗y + 9u∗x u∗y + 3u∗2 ∗2


= x + 3uy + 1
36
1 ∗
f¯6eq ρ 3u∗y − 3u∗x − 9u∗x u∗y + 3u∗2 ∗2


= x + 3uy + 1
36
1 ∗
f¯7eq ρ 9u∗x u∗y − 3u∗y − 3u∗x + 3u∗2 ∗2


= x + 3uy + 1
36
1 ∗
f¯8eq ρ 3u∗x − 3u∗y − 9u∗x u∗y + 3u∗2 ∗2


= x + 3uy + 1
36

It is easy to verify that the main conditions for the discretization are satised
up to second order velocity polynomials

8 
X 1
f¯ieq = f¯eq dξ = ρ∗
i=0
n0
8 
X 1
f¯ieq ei,x = f¯eq ξx dξ = ρ∗ u∗x
i=0
n0 c
8 
X 1
f¯ieq ei,y = f¯eq ξy dξ = ρ∗ u∗y
i=0
n0 c
8 
X 1
f¯ieq ei,x ei,y = f¯eq ξx ξy dξ = ρ∗ u∗x u∗y
i=0
n0 c2
8 
X 1 1
f¯ieq ei,x ei,x = f¯eq ξx ξx dξ = ρ∗ + ρ∗ u∗2
x
i=0
n0 c2 3
8 
X 1 1
f¯ieq ei,y ei,y = f¯eq ξy ξy dξ = ρ∗ + ρ∗ u∗2
y
i=0
n0 c2 3

as it would to be expected.
3.4. EXERCISES 55

3.4 Exercises
1- In Figure 3.3 the dimensionless lattice vectors are such that
√ |ei | = 1
for i = {1, 2, 3, 4} and |ei | = 2 i = {5, 6, 7, 8}. In the same way, the
for
quadrature weights can be found as W0 = 16/36, Wi = 4/36 for i = 1, 2, 3, 4
and Wi = 1/36 for i = 5, 6, 7, 8. Consider that you initialize a LB simulation
attributing a macroscopic velocity u∗ = 0 and a dimensionless density ρ∗ = 1
for all the 1000×1000 sites with a D2Q9 symmetry. Calculate the equilibrium
eq
populations fi in this case.

√
2- The sound speed for this lattice is cs = 1/ 3. Suppose that you are
performing a LB simulation and that, at a given time step, we nd the following
dimensionless populations at a given site x

fi = 0.1 for i = 0,
fi = 0.08 for i = 1, 2, 5,
fi = 0.06 for i = 3, 4, 6, 7, 8,

a) calculate the dimensionless densityρ∗ and the dimensionless velocity

∗ ∗
components ux , uy ;
b) calculate the equilibrium distribution as it is given by Eq. (3.13)
c) calculate the Mach number at this site for this time step,
d) calculate the Reynolds and the Knudsen number supposing that you are
using a 1000 × τ ∗ = 0.502.
1000 physical domain and a relaxation time
out
e) calculate the new populations f i after the collision step using Eq.
(3.14). Are there any negative populations? What do you think is the source
of this problem?

3- Repeat itens d) and e) of Exercise 2 , supposing now that τ ∗ = 1. What

is the physical meaning of this relaxation time?...What are the consequences of
such a choice?...

3.5 Boundary Conditions: rst part

Dealing with boundary conditions (BC) was ever considered a puzzling
question in the LatticeBoltzmann (LB) method. The most popular BC mod-
els are based on Ad-Hoc rules and, although these BC models were shown to
be suitable for low-order LB equations, their extension to highorder LB was
shown to be a very dicult problem and, at author knowledge, never solved
with satisfaction. In fact, the main question to be solved is how to deal with a
problem when the number of unknowns (the particle populations coming from
the outside part of the numerical domain ) is greater than the number of equa-
tions at our disposal at each boundary site. Recently, BC models based on the
regularization of the LB equation, or moments-based models, were proposed.
56 CHAPTER 3. A FIRST INSIGHT INTO LBM

These moments replace the discrete populations as unknowns, independently of

the number of discrete velocities that are needed for solving a given problem.
The full set of moments-based BC leads, nevertheless, to an overdetermined
system of equations and what distinguish one model from another is the way
this system is solved.
In this chapter we only present the most widely used populations-based
models. They were the former models introduced into LB theory and are still
being used in applications. In a later chapter we will be presented to the later
generation of moments-based models for boundary conditions.
For getting familiarity with the subject, we restrict ourselves to isothermal
problems in the incompressible limit (M a << 1) and to the two-dimensional
D2Q9 LBE.

3.6 Bounce back schemes

There are dierent models based on the idea of reecting back the incoming
populations from the uid nodes on the boundary nodes for assuring no mass
leaking and zero slip. In the standard bounce-back scheme (BB), Figure 3.4 the
boundary nodes are located on the wall surface and the reected populations
at time t+δ are considered to be the same incoming populations from the uid
at time t.

f −i (xb , t + δ) = f i (xb , t) .
Therefore, in the standard bounce-back scheme (MBB) the collision process is
not carried out on the boundary nodes.
A second bounce-back scheme is a modication of the rst when the colli-
sion is also imposed. In this scheme, the pre-collisional populations are recon-
structed in accordance with the following rule

f −i (xb , t) = f i (xb , t) .
So, in this modied bounce-back scheme all pre-collisional populations be-
comes determined on the boundary nodes and it is then possible to perform
the collision step on these boundary nodes.
The half-way bounce back (HWBB) is the most popular boundary condi-
tion in LBM. As for the BB scheme, for each boundary site xb the outgoing
populations f−i at time t+δ are made equal to the incoming populations f i,
at time t, but now the wall surface is supposed to be located in the half-way
between the boundary and the uid sites (Figures 3.5, 3.6).
This boundary condition assures, for plane walls, no mass leaking and zero
slip. Roughness and surface iregularity are sources of errors in numerical sim-
ulations.
Bounce-back schemes can be extended to complex geometries. Their main
advantage is their simplicity and easy implementation and, being local models,
are frequently used for ows in complex geometries such as ows in porous me-
dia. Actually, some studies have shown that the half-way bounce-back scheme
3.6. BOUNCE BACK SCHEMES 57

Figure 3.4: In the standard bounce-back scheme the boundary nodes are wet
nodes.

Figure 3.5: Half-way bounce-back boundary condition: the wall surface is sup-
posed to be located in the half-way between the boundary and the uid sites.

Figure 3.6: Half way bounce back: population f1 is supposed to propagate to

the solid site when if nds the solid surface at time t + δ/2 and returns back
to the uid site. So f3 at time t+δ will be the same f 1 at time δ .
58 CHAPTER 3. A FIRST INSIGHT INTO LBM

Figure 3.7: A wet boundary node is always to be considered as a uid node.

is of second-order accuracy, while the standard one is only of rst order . Sim-
ilar to the half-way bounce-back scheme, the modied bounce-back scheme is
also of second-order accuracy [38]. It is clear that in the bounce-back schemes,
the momentum of a particle is just reversed before and after it hits the wall,
which means that the macroscopic velocity at the wall is zero. Therefore, these
bounce-back schemes work for stationary walls. As the wall moves with a given
velocity, these schemes should be modied to include the contribution of the
wall motion.

3.7 Inamuro et al. (1995) counter-slip condition

Inamuro et al. [48], consider that the particles given by f 1, f 5, f 8 came
from cticious solid sites external to the uid domain (Figure 3.7). In this way,
he considers these distributions as equilibrium distributions given by

eq
f i = f i (ρ0 , u∗ , v ∗ + v 0 ), i = 1, 5, 8 (3.16)

where ρ0 and v0 are unknowns and are chosen so as to adjust

X
f i = ρW , (3.17)

X
f i eix = ρW u∗ , (3.18)

X
f i eiy = ρW v ∗ . (3.19)

This set of equations is used to nd ρW , ρ0 and v0 . Velocity v0 can be

considered as a counter-slip velocity that cancels the slip of the tangential
component at the boundary site. For north and south walls, Inamuro uses a u0
counter-slip. The normal component has no-slip when we use a diuse-reexion
condition such as Eq. (3.16)
3.8. ZOU AND HE (1997) BOUNDARY CONDITION FOR FLAT SURFACES59

3.8 Zou and He (1997) boundary condition for

at surfaces
In the Zou and He [94] boundary condition, the wall is supposed to be
moving with velocity u in the x-direction and v
y -direction. Site is located
in
at the wall (Figure 3.7). After propagation, the values of f 0, f 2, f 3, f 4, f 6, f 7
are known, but the pressure (ρw ) and the distributions f 1 , f 5 , f 8 are unknown.
These distributions must satisfy Eqs. (3.17-3.19).
The fourth equation is considered as the following bounce-back condition
for the non-equilibrium populations in the normal direction

neq neq
f1 = f3 . (3.20)

This last condition was formulated considering the simmetry of the pres-
sure tensor. Thus, the Zou and He approach replaces only unknown particle
populations. The strong point is numerical accuracy, especially in 2D, and the
weak point is the decient numerical stability at high Reynolds numbers. In
the Zou/He method, a bounce-back principle is applied to o-equilibrium parts
of the particle populations. This boundary condition is easy to implement in
2D and 3D.

3.9 Zou and He (1997) conditions for inlet/outlet

The unknowns are the pre-collisional populations f 1, f 5, f 8 and u∗ . Pres-
∗
sure P = ρ∗ c2s is known. The equations are
X
f i = ρ∗

X
f i eix = ρ∗ u∗

X
f i eiy = 0

and the above bounce-back condition for the non-equilibrium populations

neq neq
f1 = f3

3.10 Exercise
1. Find f 1, f 5, f 8 for Inamuro and Zou and He BC for at solid surfaces
and for Zou and He BC for the outlet of a uid domain.
60 CHAPTER 3. A FIRST INSIGHT INTO LBM

Figure 3.8: Zou and He [94] boundary conditions for the inlet.

3.11 Sample case: lid driven cavity ow

The incompressible ow in a square cavity whose top wall moves with a
uniform velocity has been widely used as a model problem for testing and
evaluating numerical techniques, in spite of the singularities that we have at
the two top corners [30, 47, 8, 92, 83, 10]. These ows are of great scientic
interest because they display almost all uid mechanical phenomena in the
simplest of geometrical settings. In this kind of ow problem, a main primary
vortex is formed near the center of the cavity and secondary and even tertiary
vortices appear in the corners and are intensied when the Reynolds number
increases (Figure 3.9). In LBM framework, numerical solutions were based on
the BGK model [47], multi-relaxation times [89], [62], entropic approach [3]
and regularization [72].
3.11. SAMPLE CASE: LID DRIVEN CAVITY FLOW 61

Figure 3.9: Streamlines for a square lid driven cavity ow when Re=1000 [ ? ].

Figure 3.10.a presents the simulations results using a moments-based method

for the core ow and boundary conditions showing the horizontal velocity com-
ponent along the mid-line x∗ = L∗ /2. Figure 3.10.b shows the vertical velocity
∗ ∗
component along the the mid-line y = L /2. Results are compared with pre-
vious work: the works of Ghia et al. [30] and Erturk et al. [25] solved this
problem using discretization schemes based on the vorticity-stream formula-
tion of the Navier-Stokes equations. The work of Botella and Peyret [8] uses a
highly-accurate solution based on Chebyshev collocation method for the spatial
discretization in an assymptotic expansion of the Navier-Stokes equations.

3.11.1 Numerical simulation work

Using the BGK lattice-Boltzmann scheme for the core ow, Eqs. 3.14 and
3.15 and the Zou & He set of boundary conditions, nd the steady-state velocity
eld for a square cavity whose lid is moving with a constant dimensionless
velocity u∗0 = 0.05. You are free to choose the dimensionless linear dimension
∗
of the cavity L (the number of boundary nodes along directions x or y), but
begin to work trying to simulate the ow for Re=100 and, if you are succesful,
Re=400.
62 CHAPTER 3. A FIRST INSIGHT INTO LBM

[25]

Figure 3.10: Numerical results for Re=100, 400, 1000 and 5000 showing the:
a) the horizontal component of the velocity along the vertical mid-line ; b) the
vertical component of the velocity along the horizontal mid-line of the cavity
?
[ ]. Results are compared with Ghia et al. [30], Botella and Peyret [8] and
Erturk et al. [25].
Chapter 4

Forces in LBM
Forces play a central role in many hydrodynamic problems. This is, for
instance, the case of the gravitational acceleration g, which acts as a source of
momentum leading to a number of hydrodynamic eects. If two uids with dif-
ferent densities are mixed or if the temperature in a uid is nonhomogeneous,
density gradients in the gravitational eld lead to buoyancy eects and phe-
nomena like the Rayleigh-Bénard instability or the Rayleigh-Taylor instability.
In the Rayleigh-Bénard instability, which is essential in studies of heat transfer,
convection patterns develop when warmed uid rises from a hot surface and
falls after cooling (Figure 4.1). The Rayleigh-Taylor instability can occur when
a layer of denser uid is over a lower-density uid layer (4.2). Gravity waves at
a free water surface are another example. Apart from gravity, there are several
other physical problems where forces are important. In incompressible ows,
the driving mechanism of the pressure gradient eld may be, in some cases,
described by a body force. There exist cases where the problem physics spec-
ify pressure gradients, but where it is convenient to replace these with forces.
One reason for this is that the LB method may lose accuracy when solving
pressure elds due to compressibility errors. While this change is possible in
arbitrarily complex ow geometries, the task of nding an equivalent driving
force eld is only trivial in periodic ow congurations. This is often explored
in LB simulations of porous media ows. Forces are also commonly used to
model multiphase or multi-component ows. Furthermore, some algorithms
for uid-structure interactions, e.g. the immersed boundary method, rely on
forces mimicking boundary conditions.

4.1 Lattice Boltzmann equation

When the force term is considered, the Boltzmann equation in its dimen-
sionless, discrete form can be written as

eq
∂t∗ f i + ei .∇f i = Ω i + g∗ . (ei − u∗ )f i , (4.1)

63
64 CHAPTER 4. FORCES IN LBM

Figure 4.1: Rayleigh-Bénard instability

Figure 4.2: Rayleigh-Taylor Instability.

4.1. LATTICE BOLTZMANN EQUATION 65

D
where the overbar indicates a dimensionless distribution
p f i = fi ξ /n0 , ξ is
the mean molecular speed, ξ = kT /m and is related to the lattice speed
c through ξ = ccs , g∗ is a dimensionless acceleration, g∗ = gδ 2 /h, Ω i is
eq
the collision term and f i is the LBE discrete approximation to the Maxwell-
Boltzmann equilibrium distribution.
In lattice-Boltzmann framework it is a common practice to simplify the
2
second term on the r..h.s. of Eq. (4.1) by neglecting all the u∗ and higher
order terms. It results

 
1 ∗ ∗ eq 1 ∗ ∗ 1 2

∗
g . (ei − u )f i u 2 Wi ρ g . ei + 2 ei ei − cs δ .u . (4.2)
cs2 cs cs

Therefore, Eq. 4.1 can be written in index notation, as,

∂t∗ f i + eiα ∂α∗ f i = Ω i + S̄ig , (4.3)

where

 
1 ∗ ∗ 1 ∗ ∗ ∗ ∗

1 2


S̄ig = W i ρ g e
α iα + g u + g u e e
iα iβ − c δ
s αβ . (4.4)
c2s 2 α β β α
c2s
 
∗ ∗ 1 ∗ ∗ ∗ ∗
above equation we have replaced gα uβ by
In the
2 gα uβ + gβ uα because

this replacement increases the symmetry and showed to have benecial eects
on the algorithm.

4.1.1 Discretization of the stream term

Most LB numerical schemes deal with a rst order discretization for the
streaming term,

f¯i (x + ei h, t + δ) = f¯i (x, t) + Ω̄i + O δ 2 .

(4.5)

The Navier-Stokes equations require the exact retrieval of the viscous stress
tensor, which is a second order moment of the non-equilibrium distribution
fineq . So if the NS equations are to be correctly retrieved in a LBE simulation,
without errors up to the rst order in Knudsen it is required to: a) use a second
order approximation to the streaming term; b) use a third order aproximation
to the equilibrium distribution and a corresponding third order set of lattice
vectors ([75], [82] ). When Ω̄i is the BGK relaxation term, Ω̄i = −fineq /τ ∗ ,
the LBM classical schemes based on Eq. (4.5) are only successful in simu-
lating hydrodynamics phenomena, because the second order errors given by
(∂tt fi + eiα eiβ ∂αβ fi + 2eiα ∂tα fi ) contribute as a numerical viscosity which is
absorbed in the momentum balance equation. When the LB equation includes
a force term as in Eq. (4.3) a second or higher order approximation has to be
used for avoiding the presence of spurious terms in the Navier-Stokes equations
[41].
66 CHAPTER 4. FORCES IN LBM

Second order approximations for the stream term of Eq. (4.3) can be found
by considering a second-order Taylor series expansion of the populations

f¯i (x + ei h, t + δ) = f¯i (x, t) + ∂t∗ f i + eiα ∂α∗ f i

1

+ ∂t∗ t∗ f i + eiα eiβ ∂α∗ β ∗ f i + 2eiα ∂t∗ α∗ f i .
2

The last term on the r..h.s. can be rearranged, giving

∂t∗ t∗ f i + eiα eiβ ∂α∗ β ∗ f i + 2eiα ∂t∗ α∗ f i = ∂t∗ ∂t∗ f i + eiα ∂α∗ f i


+ eiα ∂α∗ ∂t∗ f i + eiβ ∂β ∗ f i .

Now, from Eq. (4.3),

∂t∗ f i + eiα ∂α∗ f i = Ω i + S̄ig , (4.6)

and, so,

f¯i (x + ei h, t + δ) = f¯i (x, t) + Ω i + S̄ig

1


+ ∂t∗ Ω i + S̄ig + eiα ∂α∗ Ω i + S̄ig . (4.7)
2

Consider, now, a rst-order Taylor series expansion of the r.h.s. of Eq. (4.3)

Ω i (x + ei h, t + δ) = Ω i (x, t) + ∂t∗ Ω i + eiα ∂α∗ Ω i , (4.8)

S̄ig (x + ei h, t + δ) = S̄ig (x, t) + ∂t∗ S̄ig + eiα ∂α∗ S̄ig , (4.9)

and so, Eq. (4.7) becomes

1
f¯i (x + ei h, t + δ) = f¯i (x, t) + Ω i + S̄ig +


Ω i (x + ei h, t + δ) − Ω i (x, t)
2
1

+ S̄ig (x + ei h, t + δ) − S̄ig (x, t) ,
2
leading to

1
f¯i (x + ei h, t + δ) = f¯i (x, t) +


Ω i (x + ei h, t + δ) + Ω i (x, t)
2
1

+ S̄ig (x + ei h, t + δ) + S̄ig (x, t) (4.10)
2
4.2. A SMART PROCEDURE FOR OVERCOMING THE IMPLICITENESS OF THE N

4.2 A smart procedure for overcoming the im-

pliciteness of the numerical scheme
Eq. (4.10) represents a trapezoidal numerical scheme, which is an implicit
scheme since Ω i (x + ei h, t + δ) is dependent on f¯i (x + ei h, t + δ). By restrict-
ing ourselves to a BGK collision term, this implicitness can, nevertheless, be
overcome with the following smart procedure [76];
1. Make the following variables shif,

1
f¯i (x, t) = f¯i (x − ei h, t − δ) +


Ω i (x, t) + Ω i (x − ei h, t − δ)
2
1

+ S̄ig (x, t) + S̄ig (x − ei h, t − δ) . (4.11)
2
2. Since all populations are known in the neighbor sites x − ei h at the
previous time step t − δ, it is possible to nd the density ρ(x, t) by summing
Eq. (4.11) on i,

X
ρ∗ (x, t) = f¯i (x − ei h, t − δ)
i
" ! !#
1 X X
+ Ω i (x − ei h, t − δ) + S̄ig (x − ei h, t − δ) .
2 i i

3. After calculating the density ρ∗ (x, t) it is possible to update the velocity

from Eq. (4.11)

X
ρ∗ u∗γ (x, t) = f¯i (x − ei h, t − δ)
i
!
1 X
+ Ω i (x − ei h, t − δ)eiγ
2 i
!
1 X
+ ρ∗ gγ∗ + S̄ig (x − ei h, t − δ)eiγ ,
2 i

because

X 1 X
S̄ig eiγ = ρ∗ gα∗ Wi eiα eiγ = ρ∗ gγ∗ .
i
c2s i

4. We can then calculate the equilibrium distribution fieq (ρ∗ , u∗ ) and the
source term S̄ig (x, t) from Eq. (4.4) .
5. Population f¯i (x, t) can be nally calculated by using a slight modication
of Eq. (4.11)
68 CHAPTER 4. FORCES IN LBM

 ¯1 
1 + ∗ f (x, t) = f¯i (x − ei h, t − δ)
2τ i
 
1 1 ¯eq
+ f + Ω i (x − ei h, t − δ)
2 τ∗ i
1

+ S̄ig (x, t) + S̄ig (x − ei h, t − δ) .
2

4.3 Another alternative: changing variables

The implicitness of Eq. (4.10) can be, also, overcome with the following
change of variables,

1
f˜i = f¯i −


Ω i + S̄ig , (4.12)
2
which for a BGK collision term, can be also written as

 
1 1 ¯eq
f˜i = f¯i − ¯


f − fi + S̄ig ,
2 τ∗ i
enabling to retrieve the original distribution f¯i from known values of f˜i , f¯ieq (ρ∗ , u∗ ),
∗
S̄ig (ρ )
1 ¯eq
f˜i + 1


2τ ∗ fi + 2 S̄ig
f¯i = . (4.13)
1 + 2τ1∗
With the transformation Eq. (4.12) , Eq. (4.10) becomes

f¯eq − f˜i
 
1
f˜i (x + ei h, t + δ) = f˜i (x, t) + i + 1− S̄ig (x, t), (4.14)
τ̃ 2τ̃

where τ̃ = τ ∗ + 1/2. Eq. (4.14) represents an explicit numerical scheme for the
modied distribution f˜i .
From Eq. (4.12)

X
f˜i = ρ∗ , (4.15)
i
X 1X
f˜i eiγ = ρ∗ u∗γ − S̄ig eiγ ,
i
2 i

and, by replacing S̄ig from Eq. (4.4)

1X 1
S̄ig eiγ = ρ∗ gγ∗ ,
2 i 2
because,
4.4. SIMULATION WORK: PLANE POISEUILLE FLOW 69

X
Wi eiα eiγ = c2s δαγ .
i

Therefore,

P ˜
fi eiγ 1
u∗γ = i ∗ + gγ∗ . (4.16)
ρ 2

4.3.1 Step by step procedure

In this alternative a suggested step by step algorithm proceeds as follows.
After the distributions f˜i are propagated,
∗
1. The density ρ and velocity u∗γ can be updated from Eqs. (4.15) and
(4.16), respectively;
2. Calculate the equilibrium distributions f¯ieq (ρ∗ , u∗ ) and the source term
∗
S̄ig (ρ );
3. Find f˜i (x + ei h, t + δ) from Eq. (4.14).

4.4 Simulation work: plane Poiseuille ow

Figure 4.3: Velocity prole for a steady-state, fully developed incompressible

channel ow.

Consider, as in Figure 4.3, the incompressible ow in a channel, whose

width is 2H ∗ in lattice-units and that we are interested in nding the velocity
prole, when the following conditions are satised: a) steady state and b) fully
developed ow. This problem has a very simple analytic solution for the x
velocity component,

2 !
y∗

3
u∗x = hu∗x i 1 − , (4.17)
2 H∗
70 CHAPTER 4. FORCES IN LBM

where hu∗x i is the dimensionless average velocity along the x-direction.

Nevertheless, the computational cost can be high if we try to simulate this
problem by using a pressure dierence P0∗ − PL∗ between the entry and the
channel exit, because the ow requires a very large channel length to be fully
developed. Therefore, instead of a pressure gradient, it is a common LB practice
in this kind of problem to use a body force oriented along the x-direction and
periodic conditions for the channel entry and exit.
Using a D2Q9 LBE and Zou & He boundary conditions, try to retrieve
the velocity prole given by Eq. (4.17) for some Reynolds neumber. Consider
hu∗
x iH
∗
ν ∗ = c2s τ ∗ − 1


Re = ν∗ , where
2 is the dimensionless kinematic viscosity.
Chapter 5

The discretization problem

The distribution f (x, ξ, t) solution of the Boltzmann equation may be consid-
ered, for each position and time as a map from the continuous velocity space
onto the real numbers [75]. In fact, it belongs to the Hilbert space H of square
integrable functions f : ξ D → R, and may be written in terms of an orthogonal
basis of H, which will be considered as the innite set of Hermite polynomial
(1) (2)
tensors H (0) , Hα , Hαβ , ... (Figure 5.1). So, for each ξ in the velocity space,
(1) (2)
H (0) (ξ) = 1, Hα (ξ) = ξα , Hαβ (ξ) = ξα ξα − δαβ and so on, [33].
To get familiarity with Hermite polynomial tensors, let us start with Her-
mite polynomials in a single dimension.

5.1 Hermite polynomials in one dimension

1
Hermite polynomials were dened by Laplace in 1810 though in scarcely
2
recognizable form, and studied in detail by Chebyshev in 1859 . Chebyshev's

1 P.S. Laplace, Théorie analytique des probabilités (1812) livre 2, 321-323; Oeuvres VII
2 P.L.Chebyshev: Sur le développement des fonctions à une seule variable Bull. Acad.
Sci. St. Petersb. I 1859 193-200; Oeuvres I 501-508.

Figure 5.1: The Hilbert space.

71
72 CHAPTER 5. THE DISCRETIZATION PROBLEM

work was overlooked and they were named later after Charles Hermite who
3
wrote on the polynomials in 1864 describing them as new . In his later 1865
papers, Hermite was the rst to dene the multidimensional polynomials.

5.1.1 General expression

The general expression of Hermite polynomials is

dθ
 
ξ2 2
(θ) θ − ξ2
H (ξ) = (−1) e 2 e , (5.1)
dξ θ
where θ is the polynomial degree and ξ is a one-dimensional independent vari-
able. The rst Hermite polynomials are

H (0) (ξ) = 1,
(1)
H (ξ) = ξ,
H (2) (ξ) = ξ 2 − 1,
H (3) (ξ) = ξ ξ2 − 3 ,

H (4) (ξ) = ξ 4 − 6ξ 2 + 3.

Exercise
Derive the above expressions for the Hermite polynomials using Eq. (5.1)

5.1.2 Internal products

The internal product in the space of functions f : (−∞, ∞) → R that are
4
square integrable is dened as

 ∞
1 ξ2
f ∗g = √ e− 2 f gdξ,
2π −∞
ξ2
where e− 2 is a weight function.
Hermite polynomials are orthogonal, in the sense that

 ∞
1 ξ2
√ e− 2 H (θ) (ξ) H (η) (ξ) dξ = 0,
2π −∞

when θ 6= η .
When θ = η,
 ∞ 2
1 ξ2

√ e− 2 H (θ) dξ = θ!
2π −∞
3 C. Hermite: Sur un nouveau développement en série de fonctions C. R Acad. Sci. Paris
58 1864 93-100; Oeuvres II 293-303
4 Square
∞
integrable functions f : (−∞, ∞) → R are functions such that
−∞ |f (x)|2 dx is
a nite number
5.1. HERMITE POLYNOMIALS IN ONE DIMENSION 73

Consider now the following expansion in Hermite polynomials,

X a(θ)
f (ξ) = H (θ) (ξ) .
θ!
θ

ξ2 H (η) (ξ)
Mutiplying both sides by e− 2 √
2π
and integrating,

 ∞ X a(θ)  ∞
1 ξ2 1 ξ2
√ e− 2 H (η) (ξ) f (ξ) dξ = √ e− 2 H (η) (ξ) H (θ) (ξ) dξ ,
2π −∞ θ! 2π −∞
θ | {z }
=θ!δηθ

resulting in a closed integral expression for the coecients a(η)

 ∞
1 ξ2
a(η)
=√ e− 2 f (ξ) H (η) (ξ) dξ.
2π −∞

Exercise
Show that
 ∞ 2  ∞
1 − ξ2
2  1 ξ2
√ e H (3)
dξ = 3! , √ e− 2 H (3) H (1) dξ = 0.
2π −∞ 2π −∞

5.1.3 Maxwell Boltzmann expansion

Consider the MB distribution in one-dimension

2
f eq 1 (ξ−u)
− 2kT
=
1/2 e m .
n 2π kT
m

The exponent can be written as

2 2
(ξ − u) (ξ0 − u0 ) T0
2kT
= ,
m
2 T

where T0 is a reference temperature and

q ξ0 = ξ/ξ , u0 = u/ξ , are dimensionles
kTo
variables with ξ =
m , related to the mean molecular speed at a reference
temperature T0 . So, in its dimensionless form, the one-dimensional MB equi-
librium distribution can be written as

(ξ0 −u0 )2
eq f eq ξ 1 1 −
2 T
f = = ρ∗ 1/2 1/2
e T0
,
n0 (2π)

T

T0

where ρ∗ = n/n0 .
Dening
74 CHAPTER 5. THE DISCRETIZATION PROBLEM

T
Θ= − 1,
T0
as a relative temperature deviation with respect to T0 , we get

eq 1 1 (ξ0 −u0 )2
f = ρ∗ 1/2
√ e− 2(Θ+1) .
(2π) Θ+1
or

 
(ξ0 −u0 )2

eq 1 ξ2 √ 1
Θ+1
e− 2(Θ+1)
− 20
f = ρ∗ e .
 
1/2 ξ2

(2π) e − 20

5
The function
(ξ0 −u0 )2
√ 1
Θ+1
e− 2(Θ+1)

2
ξ0
,
e− 2

is now expanded in Hermite polynomials

1 (ξ0 −u0 )2 2 X
ξ0 1 (θ) (θ)
√ e− 2(Θ+1) = e− 2 a H (ξ0 ) .
Θ+1 θ!
θ
1
Multiplying both members by √
2π
H (η) (ξ0 ) and integrating, we get the
Hermite coecients of the expansion

 ∞
1 1 (ξ0 −u0 )2
a(η) = √ √ e− 2(Θ+1) H (η) (ξ0 ) dξ0 .
2π −∞ Θ+1
ξ0 −u0
The variable ξ0 is now changed to x = √
Θ+1
resulting in the following

expression for a(η)

 ∞ √
1 x2

a(η)
=√ e− 2 H (η) Θ + 1x + u0 dx. (5.2)
2π −∞
It is easy to verify that the rst Hermite coecients of this expansion are

a(0) = 1,
a(1) = u0 ,
a(2) = u20 + Θ,
a(3) = u0 u20 + Θ

and so on.
So, in one-dimension the MB distribution becomes a series of Hermite poly-
nomials with the following form

2 ∞
eq 1 ξ0 X 1 (θ)
f = ρ∗ 1/2
e− 2 a (u0 , Θ) H (θ) (ξ0 ) .
(2π) θ!
θ=0
5 When Θ = 0, this function is the generating function of the Hermite polynomials.
5.2. HERMITE POLYNOMIAL TENSORS 75

Exercise
Starting from Eq. (5.2) show that the rst Hermitian moments in one-
dimensional spaces are

a(0) = 1, a(1) = u0 , a(2) = u20 + Θ, a(3) = u0 u20 + Θ

5.2 Hermite polynomial tensors

5.2.1 General expression
In multidimensional velocity spaces a general recurrence relation for the
Hermite polynomial tensors can be written as

H (0) (ξ) = 1.

n
(n−1)
X
Hα(n+1)
0 α1 ...αn
= ξα0 Hα(n)
1 ...αn
− δα0 αk Hα1 ...αn /αk (5.3)
k=1

The rst Hermite polynomial tensors are

H (0) = 1,
Hα(1) = ξα ,
(2)
Hαβ = ξα ξβ − δαβ ,
(3)
Hαβγ = ξα ξβ ξγ (5.4)

− (δβγ ξα + δαγ ξβ + δαβ ξγ ) ,

(4)
Hαβγδ = ξα ξβ ξγ ξδ
 
δγδ ξα ξβ + δβδ ξα ξγ + δβγ ξα ξδ
−
+δαδ ξβ ξγ + δαγ ξβ ξδ + δαβ ξγ ξδ
+ (δαβ δγδ + δαγ δβδ + δαδ δβγ ) ,

Exercise
Derive the expressions for Hermite polynomial tensors in Eq. (5.4) using
the recurrence equation, Eq. (5.3).

5.2.2 Internal products

Internal products in the space of the functions f : ξD → R can be dened
as


1 ξ2
f ∗g = D/2
e− 2 f gdξ.
(2π)
76 CHAPTER 5. THE DISCRETIZATION PROBLEM

Hermite polynomial tensors are also orthogonal with respect to the above
internal product and their norm is given by


1 ξ2 2

D/2
e− 2 H (θ) dξ = n1 !...nD !,
(2π)
where n1 is the number of times the index x appears repeated in the polynomial
tensor,n2 is the number of times the index y appears repeated in the polynomial
tensor and D is the dimension of the velocity space (D = 1, 2, 3).
Let
X a(θ)
f (ξ) = H (θ) (ξ) .
n1 !...nD !
θ

be the expansion of a function f (ξ) in Hermite polynomial tensors.

It is easy to verify that


1 ξ2
a(η) = D/2
e− 2 f (ξ) H (η) (ξ) dξ.
(2π)

Exercise
Considering a two-dimensional space, show that


1 ξ2 2

D/2
e− 2 (2)
Hxx dξ = 2!
(2π)
and that


1 ξ2 2

D/2
e− 2 (2)
Hxy dξ = 1!
(2π)

5.2.3 Maxwell Boltzmann expansion

Consider the dimensionless form of the MB equilibrium distribution

eq 1 1 (ξ0 −u0 )2
f = ρ∗ D/2 D e− 2(Θ+1) . (5.5)
(2π) (Θ + 1) 2

The function
(ξ0 −u0 )2
√ 1
Θ+1
e− 2(Θ+1)

2
ξ0
,
e− 2

is now expanded in Hermite polynomial tensors

1 (ξ0 −u0 )2 2
ξ0 1
e−
X
2(Θ+1) = e− 2 a(θ) H (θ) (ξ0 ) . (5.6)
(Θ + 1)
D
2 n1 !...nD ! rθ
θ
1
Multiplying both members by
(2π)D/2
H (η) (ξ0 ) and integrating
5.3. VELOCITY DISCRETIZATION 77


1 1 (ξ0 −u0 )2

D/2 D e− 2(Θ+1) H (η) (ξ0 ) dξ0 = a(η) .

(2π) (Θ + 1) 2

ξ0 −u0
Changing the variable ξ0 by r= √
Θ+1
we get

 √
1 r2

a(η) = D/2
e− 2 H (η) Θ + 1r + u0 dr. (5.7)
(2π)
In components notation, the rst equilibrium moments are

a(0)
eq = 1,
a(1)
eq,α = u0,α ,
(2)
aeq,αβ = (u0,α u0,β + Θδαβ ) ,
(3) u0,α u0,β u0,γ
aeq,αβγ = ,
+Θ (δβγ u0,α + δαγ u0,β + δαβ u0,γ )
(4)
aeq,αβγδ = u0,α u0,β u0,γ u0,δ
 
δαβ u0,γ u0,δ + δαγ u0,β u0,δ
+Θ  +δαδ u0,β u0,γ + δβγ u0,α u0,δ 
+δβδ u0,α u0,γ + δγδ u0,α u0,β
+Θ2 (δαβ δγδ + δαγ δβδ + δαδ δβγ ) .

Exercise
Starting from Eq. (5.7), show that

a(0) = 1, a(1) (2) 2 (2)

x = u0,x , axx = u0,x + Θ, axy = u0,x u0,y .

5.3 Velocity discretization

Velocity discretization is a critical step in deriving the lattice Boltzmann
equation from the continuous Boltzmann equation since it is intended, in this
step, to replace the entire continuous velocity space by a nite set of discrete
velocities ξi , i = 0, ..., nb − 1.
Let f (x) be a real function [a, b].
of a real variable dened on an interval
As in all discretization problem, we want to replace f (x) by the values f (xi )
of this function on certain points xi ∈ [a, b] , i = 0, ..., n in such a way that,
although we are conscious that the set f (xi ) is not able to fully represent the
original function f (x), we expect that some properties of f (x) can be preserved
by this discrete set.
One of such properties is the zeroth moment of f (x) or the area below the
curve f (x), calculated as
78 CHAPTER 5. THE DISCRETIZATION PROBLEM

 b
f (x) dx,
a

and we can formulate the discretization problem as a quadrature problem or

to nd suitable poles xi and weights Wi in such a way that this moment is
preserved by the discrete representation,

X  b
Wi f (xi ) = f (x) dx.
i a

In our case, when working with dimensionless variables, we want to nd

a suitable set ξ0,i , i = 0, ..., nb − 1 of dimensionless discrete velocities and
weights ωi in such a way that the moments of interest of the MB equilibrium
eq
distribution f (ξ0 ) are preserved, or

eq
f (ξ0 ) ϕ (ξ0 ) dξ0

for ϕ (ξ0 ) = 1, ξ0,α , ξ0,α ξ0,β , ξ0,α ξ0,β ξ0,γ , .... In fact, as it was demonstrated by
McNamara [69], the preservation of the equilibrium moments in the discrete
space is the main condition for the exact retrieval of the macroscopic hydrody-
namic equations and this is so, because when the Knudsen number Kn → 0, all
the non-equilibrium moments can be expressed in terms of spatial derivatives
of equilibrium moments.
Nevertheless, we are, here, asked to decide the order N of ϕ (ξ0 ) up to
which the quadrature condition must be satised and N will give the order of
approximation of the LBE to the continuous Boltzmann equation.
So, considering the velocity discretization as a quadrature problem, the
quadrature equation becomes

b −1
nX 
eq eq
ωi f (ξ0,i ) ϕ (ξi,0 ) = f (ξ0 ) ϕ (ξ0 ) dξ0 .
i=0
2
ξ0,i
D/2
Further it is convenient, although not mandatory, to replace ωi by Wi (2π) e 2 .

b −1
nX 2

ξ0 eq eq
D/2
Wi (2π) e 2f (ξ0,i )ϕ (ξi,0 ) = f (ξ0 ) ϕ (ξ0 ) dξ0 . (5.8)
| {z }
i=0
ωi
| {z }
=fieq
eq
and, so, our discrete populations fi becomes

2
ξ0
eq D/2 eq
f i = Wi (2π) e 2 f (ξ0,i )
Therefore, as before, solving this quadrature problem is what is wanted,
leading to a set of discrete velocities ξ0,i and weights Wi in such a manner that
all the equilibrium moments of interest are preserved.
5.3. VELOCITY DISCRETIZATION 79

The dimensionless MB distribution is, then, developped in a series of Her-

mite polynomial tensors H (θ) as in Eqs. (5.5, 5.6)

2
eq 1 ξ0 X 1
f (ξ0 ) = ρ∗ D e− 2 a(θ) (u0 , Θ) H (θ) (ξ0 ) , (5.9)
(2π) 2 n1 !...nD !
θ

Replacing Eq. (5.9) into Eq. (5.8), the following equation is obtained

nXb −1
X 1 (θ)
a (u0 , Θ) Wi H (θ) (ξi,0 ) ϕ (ξi,0 )
n1 !...nD ! i=0
θ
 2
X 1 1 ξ0
= a(θ) (u0 , Θ) D e− 2 H (θ) (ξ0 ) ϕ (ξ0 ) dξ0 . (5.10)
n1 !...nD ! (2π) 2
θ

Nevertheless, the velocity monomials ϕ = 1, ξ0,α , ξ0,α ξ0,β ... can be, also,
6
expanded in a nite sum of Hermite polynomial tensors

N
X
ϕ (ξ0 ) = b(η) H (η) (ξ0 ) . (5.11)
η=0

So, by replacing Eq. (5.11) into Eq. (5.10)

N nXb −1
XX 1
b(η) a(θ) (u0 , Θ) Wi H (θ) (ξi,0 ) H (η) (ξi,0 )
η=0
n1 !...nD ! i=0
θ
N  2
XX 1 1 ξ0
= b(η) a(θ) (u0 , Θ) D e− 2 H (θ) (ξ0 ) H (η) (ξ0 ) dξ0 ,
η=0
n1 !...nD ! (2π) 2
θ

or

N
XX 1
b(η) a(θ) (u0 , Θ)
η=0
n1 !...nD !
θ
 Pnb −1 
i=0 Wi H (θ) (ξi,0 ) H (η) (ξi,0 )
×
2
ξ0 
− 1
D e− 2 H (θ) (ξ0 ) H (η) (ξ0 ) dξ0
(2π) 2
= 0
6 In eect

1 = H0 ,
ξ0,α = H1,α ,
ξ0,α ξ0,β = H2,αβ + δαβ H0 ,

and so on.
80 CHAPTER 5. THE DISCRETIZATION PROBLEM

The above equation is only satised when

 2
1 ξ0

D/2
e− 2 H (θ) (ξ0 ) H (η) (ξ0 ) dξ0
(2π)
b −1
nX
= Wi H (θ) (ξ0,i ) H (η) (ξ0,i ) , (5.12)
i=0

for all θ, η smaller or equal than N . In fact, in the discrete space, the nite
expansion of ϕ (ξ0 ) up to order N restricts the expansion of f eq to the same
order, meaning that the poles ξi and weights Wi are only required to satisfy a
nite set of equations.
Solving a discrete LBE gives an approximation to the full kinetic equation,
whose accuracy can be as high as it can be achieved by increasing the order of
approximation of the Hermite expansion and, consequently, the number nb of
discrete velocities used in the representation of the continuous velocity space.
Nevertheless, since the main interest in numerically solving a LBE is to
describe the behaviour of a physical system whose macroscopic equations are
known, or supposed to be known, the expansion of the MB distribution in a
nite set of orthogonal polynomial tensors of order N of approximation has the
important feature of preserving the moments hϕp i of order p, smaller or equal
than N,
 
eq
fN ϕp dξ = f eq ϕp dξ for p ≤ N.

By dening the inner product in the discrete space ξ (N ) that maps the
velocity space onto the real numbers as

b −1
nX
f ∗g = Wi f (ξ0,i ) g (ξ0,i ) , (5.13)
i=0
it can be seen that Eq. (5.12) requires that the norm and orthogonality of
Hermite polynomial tensors are preserved in the discrete space ξ (N ) for all θ, η
smaller or equal than N. This means that the metrics of the continuous velocity
space must be preserved for all velocity functions whose order is smaller than
N.
In Philippi et al. [75], it is shown that when the discrete velocity space is
invariant under π/2 rotations and reections about the x, y and z axis, the
norm preservation of the Hermite polynomial tensors in ξ (N )

 2 i2 b −1
nX i2
1 ξ0
h h
D/2
e− 2 H (θ) (ξ0 ) dξ0 = Wi H (θ) (ξ0,i ) , (5.14)
(2π) i=0

is a sucient condition for the orthogonality between these functions.This is

important, since this property reduces the discretization problem to nd the
weights Wi and poles ξ0,i satisfying solely these norm restrictions.
5.3. VELOCITY DISCRETIZATION 81

If a set ei , i = 0, ...., nb − 1, of lattice vectors is chosen, the discretization

problem reduces to nd the weights Wi and a scaling factor as , such that
ξ0,i = as ei , by solving the norm preservation condition, Eq. (5.12), with θ = η .
Considering that the poles ei are previously known this method was called
by Philippi et al. [75] as quadrature with prescribed abcissas . When the set ei ,
7
i = 0, ...., nb − 1, is regular, the lattice is space-lling and enables the use of
collision-propagation schemes in the numerical solution of the LBE. When this
set is not space lling, another numerical method, such as nite dierence, must
be used. Surmas et al., [87] and Shan, [81], show several discrete space-lling
and non space-lling velocity sets in one, two and three dimensions, obtained
with this method.

Finaly, from Eqs. (5.9) and (5.8), it is seen that the projections of the MB
distribution along the discrete poles ξ0,i are,

N
eq X 1
f i = ρ∗ Wi a(θ) (u0 , Θ) H (θ) (ξ0,i )
n1 !...nD !
θ=0

5.3.1 Example in one-dimension

Consider a second-order equilibrium distribution, with N =2

H (0) = 1,
(1)
H = ξi,0 = as ei ,
(2) 2
H = ξi,0 − 1 = a2s e2i − 1.

The norm preservation condition of these 3 Hermite polynomials provides 3

equations for as and Wi . So, a natural candidate for satisfying these restrictions
is the set e0 = 0, e1 = 1 and e−1 = −1
The constants as , W0 = W (ke0 k) , W1 = W (ke1 k) are determined by solv-
ing

b−1 i2  2 i2
X h 1 ξ0
h
Wi H (θ) (ξ0,i ) = D/2
e− 2 H (θ) (ξ0 ) dξ0 ,
i=0 (2π)

for θ = 0, 1, 2, giving

7 Space-lling LBE are sometimes called as `regular ' or `on-lattice '.

82 CHAPTER 5. THE DISCRETIZATION PROBLEM

 ∞ 2
1 ξ0
W0 + W1 + W1 = 1/2
e− 2 dξ0 , for θ = 0,
(2π) −∞
2 2 2
W0 (as e0 ) + W1 (as e1 ) + W1 (as e−1 )
 ∞
1 ξ2
− 20 2
= 1/2
e ξ0 dξ0 , for θ = 1,
(2π) −∞

2
2
2
W0 a2s e20 − 1 + W1 a2s e21 − 1 + W1 a2s e2−1 − 1
 ∞
1 ξ2
2
− 20
= 1/2
e ξ02 − 1 dξ0 , for θ = 2.
(2π) −∞

Since

 ∞
1 x2

1/2
e− 2 dx = 1,
(2π) −∞
 ∞
1 − x2
2

1/2
e x2 dx = 1,
(2π) −∞
 ∞
1 x2
2
1/2
e− 2 x2 − 1 dx = 2,
(2π) −∞

we get the following linear system of equations

W0 + W1 + W1 = 1,
W1 a2s + W1 a2s = 1,
2
2
W0 + W1 a2s − 1 + W1 a2s − 1


= 2,
2 1
√
whose solution is: W0 = , W1 = , as = 3.
3 6
The discrete equilibrium distributions are

 
eq ∗ (0) (0) (1) (1) 1 (2) (2)
f0 = W0 ρ a H (as e0 ) + a H (as e0 ) + a H (as e0 )
2
 
∗ 1 2


= W0 ρ 1 + u0 as × e0 + (u0 u0 + Θ) as e0 e0 − 1
2
 
1
W0 ρ∗ 1 − u2 + Θ ,


=
2 0
 
eq ∗ (0) (0) (1) (1) 1 (2) (2)
f1 = W1 ρ a H (as e1 ) + a Hα (as e1 ) + a H (as e1 )
2
 
∗ 1 2


= W1 ρ 1 + u0 as × e1 + (u0 u0 + Θ) as e1 e1 − 1
2
 
∗ 1 2
2

= W1 ρ 1 + u0 as + u + Θ as − 1 ,
2 0
5.3. VELOCITY DISCRETIZATION 83

 
1
eq
f−1 = W1 ρ∗ a(0) H (0) (as e−1 ) + a(1) Hα(1) (as e−1 ) + a(2) H (2) (as e−1 )
2
 
1
= W1 ρ∗ 1 + u0 as × e−1 + (u0 u0 + Θ) a2s e−1 e−1 − 1


2
 
1 2
= W1 ρ∗ 1 − u0 as + u0 + Θ a2s − 1 .



2
Writing u0 = as u∗ , the above expressions become

 
1  ∗2
f0eq = W0 ρ∗ 1 − 3u + Θ ,
2
h  2 i
f1eq = W1 ρ 1 + 3u∗ + 3u∗ + Θ ,
∗

h  2 i
eq
f−1 = W1 ρ∗ 1 − 3u∗ + 3u∗ + Θ .

We see that

X
fieq = ρ∗ ,
i
X
fieq ei = ρ∗ u∗ ,
i
1 X eq 2 1 ∗ 1 2
f e = ρ (1 + Θ) + ρ∗ u∗ .
2 i i i 6 2

Now let us express the pressure P = nkT in dimensionless quantities. Pres-

sure P can be written as

  
n kT0 T
P = nkT = mn0 .
n0 m To
But

kT0 2 1
= ξ = c2 c2s with cs = .
m as
So

P
P∗ = = ρ∗ c2s (Θ + 1) .
mn0 c2
For the kinetic energy of uctuations
 
D D kT0 n T
nec,f = nkT = n0 m
2 2 m n0 T0
| {z }
=c2 c2s
84 CHAPTER 5. THE DISCRETIZATION PROBLEM

and, therefore, since D=1 in this case,

n ec,f 1
ρ∗ e∗c,f = = ρ∗ c2s (Θ + 1) (5.15)
n0 mc2 2
√
Parameter cs Θ + 1 can be shown to be related to the dimensionless speed
of sound in the lattice. In fact, the speed of sound us is given by


∂T 
 ∂s ∂s
∂P ∂ρ ∂T P ∂ρ
u2s = = − ∂s
P = − ∂s
∂T
P
∂ρ s ∂P ρ ∂T ρ ∂P ρ

For an ideal gas, the entropy per molecule s is given by


dT dρ dT m ∂ρ
ds = cv −k = cp + 2 dP
T ρ T ρ ∂T P
ρ
Since P = m kT ,

∂T
cp ∂ρ cp
 
kT0 T D+2 2 2
u2s =−
P = = c c (Θ + 1)
cv ∂T
∂P ρ
cv m T0 D } s
| {z
|{z} | {z }
= D+2 =c2 c2s =3
D

So, for one-dimensional lattices, the dimensionless speed of sound is given

by

2
 u 2
s
u∗s = = 3c2s (Θ + 1)
c

5.3.2 Example in two-dimensions

The D2Q9 lattice is shown in Figure 5.2. The Hermite polynomials up to
the second order in two dimensions are,

H (0) = 1,
Hx(1) = ξ0x = as ei,x ,
Hy(1) = ξ0y = as ei,y ,
(2) 2
Hxx = ξ0x − 1 = a2s e2i,x − 1,
(2) 2
Hyy = ξ0y − 1 = a2s e2i,y − 1,
(2)
Hxy = ξ0x ξ0y = a2s ei,x ei,y .

The constants as , W0 , W1 , W2 are determined by solving

b −1
nX i2  2 i2
h 1 ξ0
h
Wi H (n) (ξ0,i ) = D/2
e− 2 H (n) (ξ0 ) dξ0 , (5.16)
i=0 (2π)
5.3. VELOCITY DISCRETIZATION 85

Figure 5.2: Two-dimensional lattices.

giving

W0 + 4W1 + 4W2 = 1,
4
X 8
X
W0 a2s e20,x + W1 a2s e2i,x + W2 a2s e2i,x = 1,
i=1 i=5
4
X 8
X
W0 a2s e20,y + W1 a2s e2i,y + W2 a2s e2i,y = 1,
i=1 i=5

4 8

2 X
2 X
2
W0 a2s e20,x − 1 + W1 a2s e2i,x − 1 + W2 a2s e2i,x − 1 = 2,
i=1 i=5

4 8

2 X
2 X
2
W0 a2s e20,y − 1 + W1 a2s e2i,y − 1 + W2 a2s e2i,y − 1 = 2,
i=1 i=5

4 8

2 X
2 X
2
W0 a2s e0,x e0,y + W1 a2s ei,x ei,y + 4W2 a2s ei,x ei,y = 1.
i=1 i=5

Only four of the six above equations are linearly independent, giving the
16 4 1
√
values of
36 , W1 = 36 , W2 = 36 and the scaling factor as =
W0 = 3.
The equilibrium distribution written in terms of discrete variables reads
86 CHAPTER 5. THE DISCRETIZATION PROBLEM

Figure 5.3: Three dimensional counterparts of the D2Q9 set of lattice-vectors

2
eq X 1
f i = Wi ρ∗ a(θ) (u0 , Θ) H(θ) (as ei ) .
n1 !...nD !
θ=0

For the D2Q9 model

a(0) + a(1) 1
a(2)


a2s ei,x ei,x − 1

|{z} α as ei,α + 2 xx
 =1 |{z} |{z} 
u0,α =as u∗ =a2s u∗ 2
eq α x +Θ
= Wi ρ∗ 
 
fi  + 1 a(2) 2

(2) 2

 (5.17)
 2 yy as ei,y ei,y − 1 + axy as ei,x ei,y  
|{z} |{z}
=a2s u∗ 2
y +Θ
=a2s u∗ ∗
x uy

Figure 5.3 shows two three-dimensional counterparts of the D2Q9 set of

lattice vectors. Both representations are second-order representations with
√ √ √
kei k = 0, 1, 2 for the D3Q19 and kei k = 0, 1, 2, 3 for the D3Q27 LBE.
The D2Q13 and the next lattices are also able to run second-order models.
In these cases, the number of unknowns is greater than the number of disposable
equations, and several solutions will be available, satisfying the quadrature
problem.
Nevertheless, this lattice is not able to run full third-order models. Indeed,
when N = 3, it is impossible to nd real positive values for as , W0 , W1 , W2 , W3
satisfying all the norm restrictions. This result is the same for the D2Q17
lattice (Figure 5.4).
Considering the D2Q21 lattice, Figure 5.4, as a next candidate for third-
as , W0 , W1 , W2 , W3 , W4 .W5
order models, there will be, in this case, seven unknowns
for six norm restrictions,after eliminating identical equations. Letting as be
a free variable, the system gives a solution with real positive roots when as
is inside the interval 0.659836 and 1.16208. The values as =0.659836 and
as =1.16208 are roots of the polynomials W0 (as ) and W3 (as ) , respectively.
5.3. VELOCITY DISCRETIZATION 87

Figure 5.4: Lattices D2Q17 and D2Q21. For the D2Q21 lattice there will
be seven unknowns as , W0 , W1 , W2 , W3 , W4 .W5 for six norm restrictions, after
eliminating identical equations. Letting as be a free variable, the system gives
a solution with real positive roots when as is inside the interval 0.659836 and
1.16208. The values as =0.659836 and as =1.16208 are roots of the polynomials
W0 (as ) and W3 (as ) , respectively. Therefore, when the value as =1.16208 is
chosen, W3 =0, corresponding to the lattice vectors in red, giving a modication
of the D2Q17 lattice, which was named D2V17.

Therefore, when the value as =1.16208 is chosen, W3 =0 and the lattice loses
an energy level, giving a modication of the D2Q17 lattice, which was named
D2V17 [75]. The weights, with six signicant digits, are W0 =0.402005, W1 =0.116155,
W2 =0.0330064, W3 =0, W4 =0.0000790786, and W5 =0.000258415.

5.3.3 Summary
Consider the subspace HN of H generated by the rst Hermite polynomial
tensors

(2) (N )
H (0) , Hα(1) , Hαβ , ...Hαβ...θN ,
There is a relationship between the order N of the Hilbert space HN and the
discrete and nite velocity set ei , i = 0, ..., nb − 1. Figure 5.5 shows this rela-
tionship for the two-dimensional D2Q9 lattice, a second order LBE and Figure
5.6 this same relationship for the D2V17 LBE. In fact, velocity discretization
means to replace the MB equilibrium distribution by its projection onto the
space HN and the key question in the velocity discretization problem is to nd
the set ei , i = 0, ..., nb −1 in accordance with the order N of HN . This question
has not a single answer and the prescribed abcissas method above presented
leads to dierent velocity sets, with the same order N, when they are required
to be space lling or not. Further details can be found in [75], [82], [84], [87]
and [81].
88 CHAPTER 5. THE DISCRETIZATION PROBLEM

Figure 5.5: The D2Q9 LBE as related to the Hilbert subspace H2 .

Figure 5.6: The D2V17 LBE as related to the Hilbert subspace H3 .

5.4. CARTESIAN PRODUCTS 89

1 as eix a2s e2ix − 1

1 1 as eix a2s e2ix − 1

2
as eiy as eiy as eix eiy
as eiy a
2s e2ix − 1

a2s e2iy − 1 a2s e2iy − 1 as eix a2s e2iy − 1 as eiy − 1 a2s e2ix − 1
2 2

Table 5.1: Hermitian representation of the D2Q9 LBE considered as the

cartezian product of the corresponding one-dimensional Hermite representa-
tions,

5.4 Cartesian products

5.4.1 The D2Q9 LBE
Consider the one-dimensional D1Q3 LBE, whose lattice vectors are

ei = {0, 1, −1} .
The D2Q9 set of lattice vectors is the Cartesian product of two D1Q3 LBE
along the directions x and y

eix = {0, 1, −1} and eiy = {0, 1, −1} ,

in such a way that

 
(0, 0) , (1, 0) , (0, 1) , (−1, 0) , (0, −1) ,
ei = .
(1, 1) , (−1, 1) , (−1, −1) , (1, −1) ,
√
Now, for the D1Q3 LBE we found as = 3 and W0 = W (ke0 k) = 64 , W1 =
1
√
W (ke1 k) = 6 . For the D2Q9 LBE the scaling factor is, also, as = 3 and the
weights are the products of the corresponding one-dimensional weights

4 4 16
W0 = W (ke0 k) × W (ke0 k) = × = ,
6 6 36
4 1 4
W1 = W (ke0 k) × W (ke1 k) = × = ,
6 6 36
1 1 1
W2 = W (ke1 k) × W (ke1 k) = × = .
6 6 36
The Hermitian representation of this LBE can be also retrieved as the Carte-
sian product of the corresponding one-dimensional representations (Table 5.1).
(2)
In fact, consider, e.g., the Hermite polynomial Hαβ . From the recurrence
equation, Eq. (5.3)

(2) (1) (1)

Hαβ = Hα(1) Hβ − δαβ H0 = Hα(1) Hβ ,
because, in Cartesian products between functions of dierent coordinates eix ,
eiy or eiz α 6= β . In the same way,
90 CHAPTER 5. THE DISCRETIZATION PROBLEM

(3) (2) (1) (2)

Hαβγ = Hα(1) Hβγ − δαβ Hγ(1) − δαγ Hβ = Hα(1) Hβγ , (5.18)

because if α = x, β and γ are, necessarily equal to y in a Cartesian product.

For the fourth order products

(4) (3) (2) (2) (2)

Hαβγδ = Hα(1) Hβγδ − δαβ Hγδ − δαγ Hβδ − δαδ Hβγ .
So, if α=β and γ = δ, but dierent from α (or β) meaning a Cartesian
(2) (2)
product between two second order Hermite polynomials, say Hxx and Hyy
(4) (3) (2)
Hαβγδ = Hα(1) Hβγδ − Hγδ .
From Eq. (5.18)

(3) (1) (2)

Hβγδ = Hβ Hγδ .
So

(4) (1) (2) (2)

Hαβγδ = Hα(1) Hβ Hγδ − Hγδ ,
and since α=β ,

(2) (1)
Hαβ = Hα(1) Hβ − 1.
So, when α=β and γ = δ, but dierent from α (or β)
(4) (2) (2)
Hαβγδ = Hαβ Hγδ .
This result can be generalized for higher order Cartesian products between
Hermite polynomials.
Therefore

H (0) (eix , eiy ) = H (0) (eix ) × H (0) (eiy ) = 1,

Hx(1) (eix , eiy ) = H (1) (eix ) × H (0) (eiy ) = as eix ,
Hy(1) (eix , eiy ) = H (0) (eix ) × Hy(1) (eiy ) = as eiy ,
(2) (2)
Hxx (eix , eiy ) = Hxx (eix ) × H (0) (eiy ) = a2s e2ix − 1,
(2)
Hyy (eix , eiy ) = H (0) (eix ) × Hyy
(2)
(eiy ) = a2s e2iy − 1,
(2)
Hxy (eix , eiy ) = Hx(1) (eix ) × Hy(1) (eiy ) = a2s eix eiy .
Nevertheless, we have, in this Hermitian basis, three other elements that
are possible to be obtained from the cartesian products of the corresponding
one-dimensional representations

(3) (2)
(eix ) × Hy(1) (eiy ) = a2s e2ix − 1 as eiy ,


Hxxy (eix , eiy ) =
Hxx
(3)
(eix , eiy ) = Hx(1) (eix ) × Hyy
(2)
(eiy ) = a2s e2iy − 1 as eix ,


Hxyy
(4) (2) (2)
(eiy ) = a2s e2ix − 1 a2s e2iy − 1 .



Hxxyy (eix , eiy ) = Hxx (eix ) × Hyy
5.4. CARTESIAN PRODUCTS 91

Figure 5.7: Hermitian representations for the D2Q9 LBE. (a) Second order Her-
mitian basis; (b) Full Hermitian basis as obtained from the Cartesian product
of one-dimensional representations.

This set of 9 Hermitian polynomials is represented in Figure 5.7.b. It is

easy to show that this Hermitian set is orthogonal. Nevertheless, it cannot be
considered as a fourth-order Hermitian basis because polynomials Hxxx , Hyyy ,
Hxxxx , Hyyyy , Hxxxy , Hxyyy are lacking in this representation and, so, the
third and fourth order moments continue to be retrieved with errors.
The inclusion of the two third order and the fourth order Hermite poly-
nomials in the D2Q9 representation showed, however, to be benecial to the
stability of the D2Q9 LB algorithm in numerical simulations [85]. This matter
will be discussed in its full length in a next chapter dealing with the stability
of LB equations.
In isothermal problems Θ = 0 and the equilibrium distribution given by Eq.
(5.17) reduces to

 
eq ∗ ∗ 9 ∗ 2 3 ∗2
f i,q=6 = Wi ρ 1 + 3u · ei + (u · ei ) − u , (5.19)
2 2
where q represents the number of Hermite polynomials in the basis.
When the D2Q9 LBE is considered as the cartesian product of two D1Q3
one-dimensional LBE, the equilibrium distribution is given by

 2 2 
1 + 3u∗ · ei + 29 (u∗ · ei ) − 23 u∗
∗2 ∗
+ 27 1
2


2 ux uy eix − 3
eiy
 
eq
f i,q=9 = Wi ρ∗  , (5.20)
 
∗2 ∗
 + 27 2 1
2 uy ux eiy −
3 eix

81 ∗2 ∗2
+ 4 ux uy eix − 3 e2iy − 31
1
2

5.4.2 The D3Q27 LBE

The three-dimensional D3Q27 LBE shown in Figure 5.3 can be also derived
from the cartesian product of

eix = {0, 1, −1} , eiy = {0, 1, −1} and eiz = {0, 1, −1}

in such a way that

92 CHAPTER 5. THE DISCRETIZATION PROBLEM

Figure 5.8: D3Q27 full Hermitian basis. All the Hermite polynomials in red
n o n o
(1) (2) (1) (2)
results from the Cartesian product H (0) , Hx , Hxx × H (0) , Hy , Hyy ×
n o
(1) (2)
H (0) , Hz , Hzz

 

 (0, 0, 0) , 

(1, 0, 0) , (0, 1, 0) , (0, 0, 1) , (−1, 0, 0) , (0, −1, 0) , (0, 0, −1)

 


 

(1, 1, 0) , (1, 0, 1) , (0, 1, 1) (−1, 1, 0) , (−1, 0, 1) , (0, −1, 1)

 

 
ei = (−1, −1, 0) , (−1, 0, −1) , (0, −1, −1) , (1, −1, 0) , (1, 0, −1) (0, 1, −1) .
(1, 1, 1) ,

 


 

(1, 1, −1) , (1, −1, 1) , (−1, 1, 1) , (1, −1, −1) , (−1, 1, −1) , (−1, −1, 1) ,

 


 

(−1, −1, −1)
 

√
For the D3Q27 LBE the scaling factor is, also, as = 3 and the weights are
the products of the corresponding one-dimensional weights

4 4 4 64
W0 = W (ke0 k) × W (ke0 k) × W (ke0 k) = × × = ,
6 6 6 216
4 4 1 16
W1 = W (ke0 k) × W (ke0 k) × W (ke1 k) = × × = ,
6 6 6 216
4 1 1 4
W2 = W (ke0 k) × W (ke1 k) × W (ke1 k) = × × = ,
6 6 6 216
1 1 1 1
W3 = W (ke1 k) × W (ke1 k) × W (ke1 k) = × × = .
6 6 6 216
The D3Q27 full Hermitian representation is shown in Figure 5.8.

5.4.3 High-order Cartesian products

Both the D2Q9 and the D3Q27 LBE are second order Hermitian represen-
tations and can only be used in the simulation of isothermal and incompressible
ows. Full compressible and non-isothermal ows require to increase the or-
der of the Hermitian representation and, consequently, the number of discrete
velocities.
5.4. CARTESIAN PRODUCTS 93

Considering these high-order LBE as Cartesian products of one-dimensional

LBE, let us try, as our starting point a fth velocity one, say

ei = {−2, −1, 0, 1, 2} .
There are three weights W0 = W (0), W1 = W (1) and W2 = W (2) and a
scaling factor as to be found and the norm preservation condition Eq. (5.12)

 2 2 b −1
nX 2
1 ξ0
 
D/2
e− 2 H (θ) (ξ0 ) dξ0 = Wi H (θ) (as ei ) , (5.21)
(2π) i=0

leads, for θ = 0, 1, 2, 3 to the following system of equations

W0 + 2W1 + 2W2 = 1,
2
2W1 a2s + 2W2 (2as ) = 1,
W0 + 2W1 a4s − 2a2s + 1 + 2W2 a4s 24 − 2a2s 22 + 1



= 2,
  2   2
2 2
2W1 (as ) − 3 as + 2W2 (as 2) − 3 2as = 6.

This system has no real solution.

Therefore, let us now try the D1Q5ZOT or the Zero, One and Three one-
dimensional set of lattice vectors [17],

ei = {−3, −1, 0, 1, 3} .

This set leads, for θ = 0, 1, 2, 3 to

W0 + 2W1 + 2W2 = 0!
2
2W1 a2s + 2W2 (3as ) = 1!
W0 + 2W1 a4s − 2a2s + 1 + 2W2 a4s 34 − 2a2s 32 + 1



= 2!
2W1 9a2s − 6a4s + a6s + 2W2 9 × 32 a2s − 6 × 34 a4s + 36 a6s



= 3!

This system has two solutions with positive weights and scaling factor.
One of such solutions is

√

4 4 + 10
W0 = ,
45√

3 8 − 10
W1 = ,
80√
16 − 5 10
W2 = ,
s 720 √

5 + 10
as = .
3
94 CHAPTER 5. THE DISCRETIZATION PROBLEM

Figure 5.9: The fourth-order D2Q25 LBE.

The corresponding Hermitian basis is not straightforward anymore because

the number of lattice vectors is 5 and, so, is dierent from the number of
Hermitian polynomials in the representation. We, nevertheless, try to increase
the order of the Hermitian representation, including Hxxxx into this basis after
which we check for the orthogonality condition: it is observed that these base
vectors are indeed orthogonal with respect to the weighted inner product, but
the normalization factor for Hxxxx does not conform with the norm preservation
condition, i.e.,

2 √
X 
(4) 16 + 5 10
Wi Hxxxx (as eix ) = 4!,
i
6

and not 4!
Therefore, we rewrite the fourth order Hermite polynomial as

(4)
(4) Hxxxx
Kxxxx = ,
γ

where

s √
16 + 5 10
γ= . (5.22)
6

With this implementation, the D1Q5-ZOT becomes a fourth-order LBE

and, so, able to describe full compressible and non-isothermal problems. In
consequence, the Cartesian products of the one-dimensional D1Q5-ZOT LBE,
or the two-dimensional D2Q25-ZOT and the three-dimensional D3Q125-ZOT
are, also, fourth-order LBE.

Figure 5.9 presents the D2Q25-ZOT LBE and Figure 5.10 its corresponding
Hermitian basis.
5.5. THE MOMENTS SPACE 95

Figure 5.10: Hermitian basis of the D2Q25-ZOT LBE. Hermite polynomials

in red result from the Cartesian product of two D1Q5-ZOT LBE. Hermite
polynomials in blue are Hermite polynomials modied by factor γ given in Eq.
(5.22)

5.5 The moments space

Consider the D2Q9 full Hermitian representation shown in Figure 5.7.b
and that we represent the LB populations fi in terms of this Hermitian basis,

(θ)
X
fi = Ki a(θ) , (5.23)
(θ)
Hi ∈Hq

(θ)
(θ) Wi Hi (θ)
where Ki = nx !ny ! , with Hi = H (θ) (as ei ),

(0) (8) 1
K0 = W (ke0 k) H (0) (as e0 ) , ..., K0 = (4)
W (ke0 k) Hxxyy (as e0 )
2!2!
(0) (8) 1
K1 = W (ke1 k) H (0) (as e1 ) , ..., K1 = (4)
W (ke1 k) Hxxyy (as e1 )
2!2!
(0) (8) 1
K2 = W (ke2 k) H (0) (as e2 ) , ..., K2 = (4)
W (ke2 k) Hxxyy (as e2 )
2!2!
.............................................................
(0) (8) 1
K8 = W (ke8 k) H (0) (as e8 ) , ..., K8 = (4)
W (ke8 k) Hxxyy (as e8 )
2!2!

This gives a 9×9 matrix for the D2Q9 LBE, which can be written as
96 CHAPTER 5. THE DISCRETIZATION PROBLEM

4
0 0 − 29 − 29 0 0 0√ 1 

9 √ 9
3
1
9 9 0
√
1
9
1
− 18 0 0√ − 183 1 
− 18 
3
1
9 0√ 9
1
− 18 1
9 0 − 183 0
√
1 
− 18 
1 3 1 1 3 1 
9 − 9 0√ 9 − 18 0 0
√ 18 − 18 
1
−√ 93 3
1 1

1 
K= 
9 √
0 − 18 9 0 18
√ √
0 − 18 
1 3 3 1 1 1 3 3 1 
 36 36
√ 36
√ 36 36 12 36
√ 36
√ 36 
1 3 3 1 1 1 3 3 1 
 36 −√36 36
√ 36 36 − 12 36
√
−√36 36 

1

36 −√363 −√36
3 1
36
1
36
1
12 −√36
3
−√363 1 
36
3
1
36 36 − 363 1
36
1
36
1
− 12 − 363 3
36
1
36

Since

X Wi H (θ) H (η)
i i
= δθη
i
nx !ny !
due to the orthogonality condition for the Hermite polynomials, each moment
a(η) can be written as

(η)
X
a(η) = Hi f i (5.24)
i

for each one of the q Hermitian polynomials in the Hermitian basis of Hq .

In the particular case of the D2Q9 LBE, H is, also, a 9×9 matrix of real
numbers and is written as

 
H (0) (e0 ) ... H (0) (e8 )
(1) (1)
H =  Hx (e0 ) ... Hx (e8 )  ,
 
(4) (4)
Hxxyy (e0 ) ... Hxxyy (e8 ) 9×9
or

 H = K−1 = 
1 √1 1 1
√ 1 √1 1
√ 1
√ √1
0
 3 √0 − 3 √0 √3 −√ 3 −√ 3 √3 

0
 0 3 0 − 3 3 3 − 3 − 3 
−1 2 −1 2 −1 2 2 2 2 
 
−1 −1 2 −1 2 2 2 2 2 
 
0
 0 √0 0 √ 0 √3 −3
√ 3√ −3
√ 

0
 0
√ − 3 √0 3 2√3 2 √3 −2√3 −2√ 3

0 − 3 0 3 0 2 3 −2 3 −2 3 2 3
1 −2 −2 −2 −2 4 4 4 4
Matrix H can be considered as a linear transformation from the space gen-
 T
erated by the 9 populations fb = f (ei ) , i = 0, ..., 8 to the space generated
by the 9 moments

h iT
a = a(0) , a(1)
b (1) (2) (2) (2) (3) (3) (4)
x , ay , axx , axy , ayy , axxy , , axyy , axxyy
5.5. THE MOMENTS SPACE 97

and K as its inverse H−1 , retrieving the populations f i from a set of 9 moments,

a = Hfb
b fb = Kb
a
Therefore, since H is independent of the position and time, it can be cal-
culated with its inverse K and stored in the computer memory, allowing to
compute on each point x, for each time step, the moments a(θ) from the in-
coming populations fi and vice-versa. The three rst moments are equilibrium
eq
moments and dene the equilibrium populations f i . The non-equilibrium
neq eq
populations are given by f i = f i − f i and the non-equilibrium moments by
neq
aneq = Hfb .
b
The collision step can thus be written as

 
out eq 1 neq
fb = fb + 1 − ∗ fb ,
τ
where

fb eq = aeq
Kb (5.25)

fb neq = Kb
aneq
. (5.26)

In this procedure, the collision step is thus performed with the information
we have for the equilibrium and non equilibrium moments at (x, t). It is a regu-
larized kinetic scheme because the non equilibrium populations are recalculated
by Eq. (5.26) previously to the collision step, ltering the high-order moments
that are outside the Hq=9 Hermitian representation. The moments method 
together with the idea of replacing the BGK collision step by the relaxation of
the non-equilibrium moments with dierent relaxation times  was introduced
into the LBM framework by D'Humières [23]. Nevertheless, contrary to nd-
ing a set of orthogonal polynomials by the Gram-Schmidt orthogonalization
method as in D'Humières methodwhich can be a cumbersome process, espe-
cially for high order LBEthese polynomials are here considered as discrete
forms of the Hermite polynomials. The main condition is that these polyno-
mials satisfy Eq. (5.14), preserving, in the discrete space, the same metrics of
the continuous space.
This procedure can be extended to all LBE that are Cartesian products of
one-dimensional LBE.
98 CHAPTER 5. THE DISCRETIZATION PROBLEM
Chapter 6

Chapman-Enskog analysis
In this chapter we investigate the necessary conditions for a LBE for retrieving
the macroscopic transport equations. First we remember some main points of
the Chapman-Enskog analysis [11], [12], restricting ourselves to a system of
material points without external forces.

6.1 The continuous case

For a system of material points without external forces, the Boltzman
equation can be written as

∂t f + ξ · ∇f = Ω. (6.1)

Let τ be a scale of time for the molecular collisions and Γ a macroscopic

time. If ` is the mean free path and ξ the mean molecular speed, τ = `/ξ . Let
+ +
us dene a dimensionless time t = t/Γ and a dimensionless length x = x/L,
+
where L is a macroscopic length and consider ξ = ξ/ξ . Since the characteristic
+
time for collisions is τ , we write Ω = Ω/τ.
With these denitions Eq. (6.1) can be written as

τ /Γ Ω+
∂t+ f + ξ + · ∇+ f = . (6.2)
`/L `/L
When collisions are dominant as it happens in continuous systems, we ex-
pect that

`  L,
τ  Γ.

We dene the Knudsen number as

`
= .
L

99
100 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

The dimensionless group multiplying the rst term on the l.h.s. of Eq. (6.2)
can be written as

τ /Γ χ
= ,
`/L 
where χ = τ / Γ. The term χ/  is expected to be of O(1), although it can be an
important dimensionless number in problems with high oscillating macroscopic
1
variables. This term is, in fact, usually identied as the Strouhal number ,

L
= Sr.
Γξ
Now, when the molecular collisions prevail over advection, we can consider
that the distribution function f is close to the equilibrium distribution and
write it as an asymptotic expansion

f = f (0) + f (1) + 2 f (2) + .... (6.3)

Consider, for instance, the momentum balance equation derived by multi-

plying Eq. (6.1) by mξ and integrating the result in the velocity space,

 
∂t (ρu) + ∇ · ρuu + P = 0. (6.4)

In this equation the pressure tensor P is given by


Pαβ = mf ξα ξβ dξ,

and so, the asymptotic expansion Eq. (6.3) will induce the decomposition of
the pressure tensor in accordance with the order of the Knudsen number

  
(0) (1) 2
Pαβ = mf ξα ξβ dξ +  mf ξα ξβ dξ +  mf (2) ξα ξβ dξ + ...,
| {z } | {z } | {z }
(0) (1) (2)
Pαβ Pαβ Pαβ

meaning that the time derivative of the momentum in Eq. (6.4) is due to
2
contributions with dierent orders in the Knudsen number. Enskog proposed
to decompose this time derivative in accordance with

∂t = ∂0 + ∂1 + 2 ∂2 + ... (6.5)

In such a way that the momentum balance equation was decomposed as

1 Strouhal, V. (1878) "Ueber eine besondere Art der Tonerregung" (On an unusual sort of
sound excitation), Annalen der Physik und Chemie, 3rd series, 5 (10) : 216251.
2 Enskog, D. Kinetische theorie der vorgänge in mässig verdünnten gasen. I. Allgemmeiner
teil. Inaugural Dissertation, Published by Alqvist & Wiksells, 1917, Uppsala, 1917
6.1. THE CONTINUOUS CASE 101

 
(0)
∂0 (ρu) + ∇ · ρuu + P = 0, (6.6)

 
(1)
∂1 (ρu) + ∇ · P = 0, (6.7)

 
(2)
2 2
 ∂2 (ρu) + ∇ ·  P = 0, (6.8)

and so on.
When Eqs. (6.3) and (6.5) are replaced into Eq. (6.2), we get the following
equation

χ  
(∂0+ + ∂1+ + ∂2+ ...) f (0) + f (1) + 2 f (2) + ...
  
+ξ + · ∇+ f (0) + f (1) + 2 f (2) + ...
(0)+
Ω (1)+ (2)+
= +Ω + Ω + ....

By rearranging the terms in accordance with their order

χ χ   χ  
∂0 f (0) + ξ + · ∇+ f (0) + ∂0 f (1) + ∂1 f (0) + ξ + · ∇+ f (1) +
| {z } |  {z }
order 1 order 
χ χ χ +
∂0 2 f (2) + ∂1 f (1) + 2 ∂2 f (0) + 2 ξ · ∇+ f (2) + ...
|  {z }
order 2
(0)+
Ω (1)+ (2)+
= +Ω + Ω + ....

The above equation is then split into the following equations,

(0)+
Ω = 0,

χ (1)+
∂0 f (0) + ξ + · ∇+ f (0) = Ω ,

χ  (1)  χ   (2)+
∂0 f + ∂1 f (0) + ξ + · ∇+ f (1) = Ω ,
 
and so on.
We can then return to the original variables, obtaining

Ω(0) = 0, (6.9)
(0) (0) (1)
∂0 f + ξ · ∇f = Ω , (6.10)
   
(0) (1)
∂1 f + ∂0 f + ξ · ∇ f (1) =  Ω 2 (2)
, (6.11)
102 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

and so on.
Therefore, from Eq. (6.9), the order-zero solution f (0) is the Maxwell-
Boltzmann equilibrium distribution

f (0) = f eq .

6.1.1 Macroscopic equations

Order zero in Knudsen

Multiplying Eq. (6.10) by the mass m and integrating in the velocity space
 
∂0 mf (0) dξ + ∂β mξ f (0) dξ = 0,
β
or

∂0 ρ + ∂β (ρuβ ) = 0, (6.12)

giving the mass conservation equation at order zero.

The momentum balance equation at this order is obtained by multiplying
Eq. (6.10) by the momentum mξ and integrating in the velocity space,
 
∂0 mξα f (0) dξ + ∂β mξα ξ f (0) dξ = 0,
β
or

(0)
∂0 (ρuα ) + ∂β Pαβ = 0. (6.13)

Therefore, at this scale, the pressure tensor reduces to


(0)
Pαβ = mf (0) ξα ξβ dξ =ρuα uβ + P δαβ ,

where P = nkT is the ideal gas pressure. The hydrodynamic equations are the
Euler equations and there is no viscous transfer of momentum.
Consider now that we multiply Eq. (6.10) by the molecular kinetic energy
1 2
2 mξ
 
1 2 (0) 1 2
∂0 mξ f dξ + ∂β mξ ξ f (0) dξ = 0. (6.14)
2 2 β
1 2
We know that the molecular kinetic energy
2 mξ may be decomposed,
following

1 2 1 2 1
mξ = m (ξα − uα ) + mξα uα − mu2α .
2 2 2
In addition,

ξα ξβ = (ξα − uα ) (ξβ − uβ ) + ξα uβ + ξβ uα − uα uβ .
6.1. THE CONTINUOUS CASE 103

So, Eq. (6.14) becomes

    
1 1 2
∂0 ρec,f + ρu2 + ∂β ρec,f + ρu uβ + P uβ = 0, (6.15)
2 2

where


1 2
ρec,f = m (ξα − uα ) f (0) dξ,
2
is the peculiar kinetic energy per unit volume of the system of particles and,
at the zeroth-order, Eq. (6.15) gives the balance of the total kinetic energy,
1 2
considered as composed by ρec,f 2 ρu .
and the advective energy
At order zero in Knudsen, the balance equation for the internal energy can
be found by subtracting the balance equation for the momentum at this order
from Eq. (6.15).
So, when we perform the internal product of Eq. (6.13) by the macroscopic
velocity,

uα ∂0 (ρuα ) + uα ∂β (ρuα uβ + P δαβ ) = 0, (6.16)

the rst term on the l.h.s. of the above equation gives

 
1 2 1
uα ∂0 (ρuα ) = ∂t ρu + u2 ∂0 ρ,
2 2
whereas

 
1 2 1
uα ∂β (ρuα uβ ) = ∂β ρu uβ + u2 ∂β (ρuβ ) ,
2 2
in such a manner that

   
1 2 1 2
uα ∂0 (ρuα ) + uα ∂β (ρuα uβ ) = ∂0 ρu + ∂β ρu uβ .
2 2
On the other hand

δαβ uα ∂β (P ) = ∂β (P uβ ) − P ∂β (uβ ) .
So, the balance equation for the advection energy becomes,

   
1 2 1 2
∂0 ρu + ∂β ρu uβ + P uβ = P ∇ · u. (6.17)
2 2
Eq. (6.17) is then subtracted from Eq. (6.15) giving

∂0 (ρec,f ) + ∇ · [ρec,f u] = −P ∇ · u, (6.18)

which is the balance equation for the internal energy at the Kn zeroth-order.
104 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

Order 1 in Knudsen
At the much faster time scale t1 = t0 , the solution f (1) is retrieved by solving
Eq. (6.10). Considering for simplicity a BGK collision model [4], then

f (1)
Ω(1) = − ,
τ
and, in this case the rst order approximation to the distribution function f (1)
and the viscous stress tensor ταβ are found by writting the l.h.s. of Eq. (6.10)
in terms of spatial derivatives of the hydrodynamic elds.
So, when Eq. (6.11) is multiplied by the mass m and integrated in the
velocity space, it gives

∂1 ρ = 0. (6.19)

Consider now, multiplying Eq. (6.11) by the momentum mξα and integrat-
ing the result in the velocity space

   
∂1 mξα f (0) dξ + ∂β mξα ξβ f (1) dξ = 0,

or

∂1 (ρuα ) + ∂β ταβ = 0, (6.20)

where ταβ is the viscous stress tensor

  
ταβ = mξα ξβ f (1) dξ.

1 2
Finally, we multiply Eq. (6.11) by the kinetic energy
2 mξα and integrate
the result in the velocity space

 
1 2 (0) 1 2
∂1 mξ f dξ + ∂β mξ ξβ f (1) dξ = 0. (6.21)
2 2
| {z }
=ρec,f + 21 ρu2

For the second term on the l.h.s.

  
1 2 1 2 1 2 (1)
mξ ξβ f (1) dξ = mξ (ξβ − uβ ) f (1) dξ+uβ mξ f dξ .
2 2 2
| {z }
=0

But

1 2 1
mξ = m (ξα − uα ) (ξα − uα ) + m (ξα − uα ) uα − muα uα
2 2
and so,
6.1. THE CONTINUOUS CASE 105

 
1 2 1 2
mξ ξβ f (1) dξ = mξ (ξβ − uβ ) f (1) dξ
2 2
  
1
= m (ξα − uα ) (ξα − uα ) (ξβ − uβ ) f (1) dξ
2
| {z }
(1)
=qβ

+uα m (ξα − uα ) (ξβ − uβ ) f (1) dξ
| {z }
(1)
=ταβ

− (uα uα ) m (ξβ − uβ ) f (1) dξ ,
| {z }
=0

where the vector giving the ux of heat q is given by the diusive transport of
the peculiar kinetic energy

1 2
q= m (ξ − u) (ξ − u) f neq dξ.
2
Finally

 
1 
(1) (1)

∂1 ρec,f + ρu2 + ∂β qβ + uα ταβ = 0, (6.22)
2
or

 
1
∂1 ρec,f + ρu2 + ∂β (qβ + uα ταβ ) = 0 (6.23)
2

Macroscopic equations up to the rst order in the Knudsen number

By summing Eqs. (6.12) and (6.19), the mass conservation equation is
retrieved

∂t ρ + ∂β (ρuβ ) = 0. (6.24)

The momentum balance equation is obtained by summing Eqs. (6.13) and

(6.20)

∂t (ρuα ) + ∂β (P δαβ + ρuα uβ + ταβ ) = 0. (6.25)

Finally, we sum Eqs. (6.15) and (6.23) for retrieving the equation for the
energy conservation,

 
1 2  1 2


∂t
2
ρu + ρec,f + ∂β 2 ρu + ρec,f uβ + P uβ + ταβ uα + qβ = 0.
(6.26)
106 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

Balance equation for the internal energy

1 2
2 ρu my be obtained by
The balance equation for the advection energy
making the inner product of Eq. (6.25) by the macroscopic velocity u. The
procedure is the same as the one used for obtaining the the Knudsen zeroth-
order balance equation for the advection energy, Eq. (6.17) and gives

   
1 2 1 2
∂t ρu + ∂β ρu uβ + uβ P + uα ταβ = P ∇ · u + ταβ ∂β uα . (6.27)
2 2
The balance equation for the internal energy can be then obtained by sub-
tracting Eq. (6.27) from Eq. (6.26),

∂t (ρec,f ) + ∂β (ρec,f uβ + qβ ) = −P ∇ · u−ταβ ∂β uα . (6.28)

The last term on the r.h.s. is the viscous dissipation term φ = −ταβ ∂β uα ,
which can be shown to be always non-negative and represents the irreversible
transformation of advective energy into internal energy.

6.1.2 Exercise
When the collision operator Ω is modelled following a BGK relaxation term
f − f eq
Ω=−
τ
this relaxation term is decomposed in accordance with

f (0) − f eq
Ω(0) = −
τ

f (1)
Ω(1) = −
τ

f (2)
Ω(2) = −
τ
and so on.
a) Using Eq. (6.10), show that the viscous stress tensor components, in
(1)
order 1 in Knudsen, ταβ = ταβ , when calculated as the moment

Ω(1) ξα ξβ dξ,

is given by

ταβ = −µ (∂α uβ + ∂β uα ) + κ∇ · uδαβ ,

2
where D is, as usually, the Euclidean dimension of the space and µ= D τ ρec,f
2 ∗
is the dimensionlessmshear viscosity and κ = D µ is the dimensionless bulk
viscosity.
6.2. SCALE ANALYSIS OF A LB EQUATION 107

b) Using Eq. (6.10), nd the heat ux vector , in order 1 in Knudsen,
(1)
qα = qα , when calculated as the moment

1 2
Ω(1) m (ξβ − uβ ) (ξα − uα ) dξ.
2

6.2 Scale analysis of a LB equation

Let us restrict ourselves to the more usual LBM rst-order collision-propagation
schemes, when the discrete form of the Boltzmann equation, Eq. (6.1), is given
by

f i (x + ei h, t + δ) − f i (x, t) = Ωi + O h2 .


(6.29)

3
where, as before, f i = fni ξ̄0 is the population of particles fi = f (ξi ), in dimen-
ξ¯ = kT0 /m is the mean molecular speed, and T0 and n0 are,
p
sionless form,
respectively, a reference temperature and a reference number density.
We develop f i (x + ei h, t + δ) in a Taylor series around f i (x, t),obtaining

1 1
δ∂t f i + heiα ∂α f i + δ 2 ∂tt f i + h2 eiα eiβ ∂αβ f i + hδeiα ∂tα f i + ... = Ωi . (6.30)
|2 2 {z }
second order errors

So, in the rst-order LB, the second-order terms in the above equation must
be considered as errors with respect to the continuous Boltzmann equation.
Let L be a macroscopic length and Γ a macroscopic time and dene

h
= ,
L
and

δ
χ= .
Γ
Although we are using the same symbols  and χ as before, their meanings
are not the same. Symbol  is now being used for the spatial resolution and the
symbol χ for the ratio between the time step δ and the macroscopic scale of
time. In LBM, the Knudsen number continues to be given by the ratio between
the mean free path and a macroscopic characteristic length,

` ξτ h
Kn = = ,
L h L
and since ξ = ccs = hδ cs , it can also be written as

ν∗
Kn = cs τ ∗  = .
cs
108 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

Remark that, in addition to the geometric resolution  and the scaling factor
cs a lattice constant the Knudsen number depends on the dimensionless
relaxation time τ ∗ = τ /δ . This relaxation time is a measure of the importance
of the collision term with respect to the stream term. Small values of τ∗ means
a fast relaxation to an equilibrium state and corresponds to a low kinematic
viscosity. So, in low-viscous uids the interactions between the particles are
so numerous and eective that momentum and energy between neighboring
particles is quickly balanced; the system remains close to the equilibrium state.
As a result, the net ux of momentum is small in the moving frame of reference
or, in other words, transport of momentum is sustained mainly by advection
rather than by diusion.
Both  and χ are required to be small parameters, the rst one being related
to the geometric resolution of the physical system and the second one to the
ratio between the time step δ and the time Γ that gives a scale for the time
variation of the local elds. We consider χ ≈  and replace t by t+ = t/Γ and
x by x+ = x/L in Eq. (6.30),

2
1χ
 
2  ∂t t f i
+ +
χ Ωi
∂t+ f i + eiα ∂α+ f i +  + 12 eiα eiβ ∂α+ β + f i  + ... = .
 
| {z } +χeiα ∂t+ α+ f i |{z}
O(1) | {z } O(−1 )
O()

Therefore, when  → 0, the collision term on the right-hand side of the

above equation becomes dominant with respect to the stream terms on the
left-hand side. Consider, now, the following expansion of fi when  → 0,
(0) (1) (2)
fi = fi + f i + 2 f i + ... .

Using the Enskog decomposition of the time derivative

∂t+ = ∂0+ + ∂1+ + 2 ∂2+ + ...,

we obtain

χ (0) (0)
∂0+ f i + eiα ∂α+ f i +
| {z }
O(1)
(1) (1) (0) 2 (0)
!
1χ
χ∂0+ f i + eiα ∂α+ f i + χ∂1+ f i +
2  ∂0 0 f i
+ +
(0) (0) + ...
+ 12 eiα eiβ ∂α+ β + f i + χeiα ∂0+ α+ f i
| {z }
O()
(0)
Ωi (1) (2)
= + Ωi + Ωi + ... . (6.31)

| {z } |{z} | {z }
O(1) O()
O(1/)

So, in O(1/):
6.3. ZERO-TH ORDER TRANSPORT EQUATIONS 109

(0)
Ωi = 0.
(0)
and this equation gives the zero-th order approximation fi to the solution as
related to the equilibrium solution

(0) eq
fi = fi ,
usually written as a polynomial approximation to the Maxwell-Boltzmann equi-
librium distribution.
In O(1) :
χ (0) (0) (1)
∂ + f + eiα ∂α+ f i = Ωi , (6.32)
 0 i
or, returning to the LB variables,

(0) (0) (1)

∂0∗ f i + eiα ∂α∗ f i = Ωi , (6.33)

where t∗ = t/δ and x∗ = x/h are dimensionless LB variables.

Finally, in O()

(1) (1) (0) 1χ

2 (0)
χ∂0+ f i + eiα ∂α+ f i + χ∂1+ f i +
2  ∂0 0 f i
+ +
(0) (0) (2)
+ 12 eiα eiβ ∂α+ β + f i + χeiα ∂0+ α+ f i = Ωi
or, returning to the LB variables,

(1) (1) (0) (0)

∂0∗ f i + eiα ∂α∗ f i + ∂1∗ f i + 12 ∂0∗ 0∗ f i
(0) (0) (2) (6.34)
+ 12 eiα eiβ ∂α∗ β ∗ f i + eiα ∂0∗ α∗ f i = 2 Ω i .
Eqs. (6.32) and (6.34) are the starting points to nd the conditions which
must be fullled by a rst-order LB equation to retrieve the hydrodynamic
macroscopic equations.

6.3 Zero-th order transport equations

6.3.1 Mass conservation
When we sum Eq. (6.33) on i,
X (0) X (0) X (1)
∂0∗ fi + ∂α∗ f i eiα = δ Ωi ,
i i i
th
this gives the zero order aproximation of the mass conservation equation in
LB dimensionless variables

∂0∗ ρ∗ + ∇∗ · (ρ∗ u∗ ) = 0, (6.35)

when the following conditions are satised

110 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS


X (0) 1
fi = ρ∗ = f eq dξ, (6.36)
i
n0

X (0) ∗ ∗ 1
f i ei =ρ u = f eq ξdξ, (6.37)
i
cn0

and

X 
Ωi = Ωdξ =0. (6.38)
i
(0) eq
Therefore, since fi = fi , we see that, for retrieving the zero-th approx-
imation to the mass conservation equation, the Hermite polynomial approxi-
eq
mation for fi must satisfy the conditions given by Eqs. (6.36-6.37) and, in
addition, the modelled collision term must be written in such way as to sat-
isfy Eq. (6.38), meaning that mass must be preserved by the colision model
considering the discrete set ei , i = 0, ..., b − 1 of lattice vectors.

6.3.2 Momentum balance

Now, multiply Eq. (6.33) by eiβ and sum on i,
X (0) X (0)  X (1)
∂0∗ f i eiβ + ∂α∗ f i eiα eiβ = Ω eiβ ,
i i
δ i i

giving the Euler hydrodynamic equations

∂0∗ (ρ∗ u∗ ) + ∇∗ · (ρ∗ u∗ u∗ + P ∗ δ) = 0, (6.39)

when Eq. (6.37) is satised and when


X (0) 1
f i eiα eiβ = ρ∗ u∗α u∗β ∗
+ P δαβ = f eq ξα ξβ dξ, (6.40)
i
n 0 c2

and

X 
Ωi ei = Ωξdξ = 0. (6.41)
i

In Eq. (6.40) the dimensionless pressure is given by

P ∗ = ρ∗ c2s (Θ + 1)
and corresponds to the Clayperon pressure for an ideal gas, P = nkT . Eq.
(6.41) means that the collision model and the set ei , i = 0, ..., b − 1 of lattice
vectors are required to satisfy the momentum preservation in collisions.
(0) eq
Eq. (6.40) is only satised when fi = fi is, at least, a second order
Hermitian representation.
6.3. ZERO-TH ORDER TRANSPORT EQUATIONS 111

Exercise
Show that the D2Q9 LBE with an equilibrium distribution given by

a4s
eq
h    i
f i,q=6 = Wi ρ∗ + a2s ρ∗ u∗α eiα + 2 ρ∗ u∗α u∗β + Θ eiα eiβ − 1
a2s δαβ ,
(6.42)
satises the condition given by Eq. (6.40).

6.3.3 Energy conservation

1
Multiply Eq. (6.33) by
2 eiβ eiβ and sum on i,

! !
1 X (0) 1 X (0) X (0) 1
∂0∗ f eiβ eiβ + ∂α∗ f eiα eiβ eiβ = Ωi eiβ eiβ
2 i i 2 i i i
2

which gives the zero − th order aproximation of the energy conservation equa-
tion

    
1 ∗ ∗2 1 ∗ ∗2
∂0∗ ρ u + ρ∗ e∗c,f + ∇∗ · ρ u + ρ∗ e∗c,f u∗ + P ∗ u∗ = 0, (6.43)
2 2
3
where the internal energy per unit volume is given in its dimensionless form
by

D ∗ 2
ρ∗ e∗c,f = ρ cs (Θ + 1) ,
2
For retrieving the energy conservation equation at order zero, Eq. (6.43),
the following conditions must be satised


1 X (0) 1 2 1 1
f eiβ eiβ = ρ∗ u∗ + ρ∗ e∗c,f = f eq ξβ ξβ dξ, (6.44)
2 i i 2 n0 c2 2

and

  
1X (0) 1 ∗ ∗2 1 1
mf i eiα eiβ eiβ = ρ u + ρ∗ e∗c,f u∗α +P ∗ u∗α = f eq ξα ξβ ξβ dξ,
2 i 2 n0 c3 2
(6.45)
which means a null net heat ow vector at this order,

3 The internal energy is the kinetic energy resulting from the velocity uctuations of the
molecules,

1
ρe = m (ξ − u)2 dξ.
2
112 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

(0)
q = 0.
In addition, the following condition must also be satised,

X 
Ωi eiβ eiβ = Ωξ 2 dξ = 0, (6.46)
i

meaning that the collision model Ωi and the set ei , i = 0, ..., b − 1 of lattice
vectors are required to satisfy the preservation of kinetic energy in collisions.

Exercise
It is important to remark that Eq. (6.45) requires more than a second
eq
order Hermitian representation of fi . Verify that the D2Q9 LBE with the
equilibrium distribution given by Eq. (6.42) do not satisfy the condition given
by Eq. (6.45) .

6.3.4 Internal energy balance equation

Multiply Eq. (6.39) by u∗β

u∗β ∂0∗ ρ∗ u∗β + u∗β ∂α∗ ρ∗ u∗α u∗β + u∗β ∂β ∗ P ∗ = 0,

But

 
1 ∗ ∗2 1
u∗β ∂0∗ ρ∗ u∗β = ∂0∗ + u∗2 ∂0∗ ρ∗ ,


ρ u
2 2
and

 
1 ∗ ∗2 ∗ 1
u∗β ∂α∗ ρ∗ u∗α u∗β = ∇∗ · + u∗2 ∇∗ · (ρ∗ u∗ ) .


ρ u u
2 2
Finally,

u∗ · ∇∗ P ∗ = ∇∗ · (P ∗ u∗ ) − P ∗ ∇∗ · u∗ ,
giving the balance equation for the advection energy at the zero-th order,

   
1 ∗ ∗2 1 ∗ ∗2 ∗
∂0∗ ρ u + ∇ · ρ u u + P u = |P ∗ ∇{z
∗ ∗ ∗ ∗
· u∗}. (6.47)
2 2
source term

Thus, the pressure P∗ that is advected with the ow contributes, to the
1 ∗ ∗2
balance of the advective energy
2 ρ u , at this order of approximation. The
∗ ∗ ∗
source of the advection energy is σ (Eadv ) = P ∇ · u . Since the potential
energy, related to external forces is not being considered, the total energy Et
has only two contributions, the advection Eadv and the internal energy E

Et = Eadv + E.
6.4. FIRST-ORDER TRANSPORT EQUATIONS 113

Consequently,

σ (Et ) = σ (Eadv ) + σ (E) = 0,

and so,

σ (E) = −σ (Eadv ) = −P ∗ ∇∗ · u∗ ,
giving for the internal energy, the following balance equation at order zero

∂0∗ (ρ∗ e∗ ) + ∇∗ · (ρ∗ e∗ u∗ ) = −P ∗ ∇∗ · u∗ . (6.48)

Therefore, at the time scale t0 , the internal energy is only transported

by advection along the current lines, with a source term σ (E) related to the
divergence of the local velocity u, in compressible ows. The heat ux, or the
diusive transport of internal energy due to the random motion of molecules
is null at this scale t0 of time.
In the following we will see what are the macroscopic equations at the faster
time scale t1 .

6.4 First-order transport equations

We rewrite Eq. (6.34),

(1) (1) (0)

∂0∗ f + eiα ∂α∗ f + ∂1∗ f i
1h (0) (0) (0)
i
+ ∂0∗ 0∗ f i + eiα eiβ ∂α∗ β ∗ f i + 2eiα ∂0∗ α∗ f i
2
(2)
= 2 Ω i .

The term between brackets in the l.h.s. of the above equation can be,
further rearranged

(0) (0) (0)

∂0∗ 0∗ f i + eiα eiβ ∂α∗ β ∗ f i + 2eiα ∂0∗ α∗ f i
 (0) (0)

∂0∗ ∂0∗ f i + eiα ∂α∗ f i
=  (0) (0)
 ,
+eiβ ∂β ∗ ∂0∗ f i + eiα ∂α∗ f i

and, in accordance with Eq. (6.33)

(0) (0) (1)

∂0∗ f i + eiα ∂α∗ f i = Ωi .

So, we obtain
114 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

(1)
 (1)
 (0)
∂0∗ f i + ∂α∗ eiα f i + ∂1∗ f i
1h (1) (1)
i
+ ∂0∗ Ωi + eiβ ∂β ∗ Ωi
2
(2)
= 2 Ω i . (6.49)

6.4.1 Mass conservation equation

Multiply Eq. (6.49) by 1 and sum on i. We obtain the mass conservation
equation at the time scale t1 = t0

∂1 ρ∗ = 0, (6.50)

when conditions Eqs. (6.38, 6.41) are satised and from

X (1) X (1)
fi = eiα f i = 0,
i i

resulting from the solvability conditions

 
eq
f ϕ (ξ) dξ = f ϕ (ξ) dξ,

for ϕ (ξ) = m, mξ, 12 mξ 2 corresponding to the equilibrium moments mass, mo-

mentum and kinetic energy. Therefore, the solvability conditions are satised
when Eqs. (6.36), (6.37), and (6.44) are satised.

6.4.2 Momentum balance equation

We need, in the next, a collision model for the collision term. Consider a
BGK model. In this case,

(1)
(1) f
Ωi = − i∗ . (6.51)
τ
Eq. (6.49) becomes

(0)
∂1∗ f i
(1)
!
f (1)
1
∂0∗ i∗ + eiα ∂α∗ f i


+ 1− 2τ ∗ τ
(2)
=  2 Ωi . (6.52)

Multiply Eq. (6.52) by eiβ and sum on i, using the previous relations, and
the condition
6.4. FIRST-ORDER TRANSPORT EQUATIONS 115

X (2)
Ωi eiβ = 0. (6.53)
i

We obtain

  
1
ρ∗ u∗β


∂1∗ + ∂α∗ 1− ∗ ταβ = 0, (6.54)
2τ
4
where

(1)
X (1)

ταβ = ταβ = f i eiα eiβ ,
i

From Eq. (6.33) and (6.51)

(1)
(0) (0) f i
∂0∗ f i + eiα ∂α∗ f i =− ,
τ∗
and so, making the inner product of the above equation by eiα eiβ ,
X  (0) (0)

ταβ = −τ ∗ eiα eiβ ∂0∗ f i + eiγ ∂γ ∗ f i . (6.55)
i

The following conditions must be satised by the LBE for correctly retriev-
ing the viscous stress tensor


X (0) 1
eiα eiβ f i = f eq ξα ξβ dξ, (6.56)
i
n 0 c2

X (0) 1
eiα eiβ eiγ f i = f eq ξα ξβ ξγ dξ. (6.57)
i
n 0 c3

The rst condition Eq. (6.56) is, in fact, the condition given by Eq. (6.44)
and results in

X (0) 2 ∗ ∗
eiα eiβ f i = ρ e δαβ + ρ∗ u∗α u∗β .
i
D

The second condition, Eq. (6.57) requires a third order Hermite polynomial
(0)
expansion for fi . When this condition is satised it gives

4 Rigorously, the viscous tension tensor is given by

(1) (2)
ταβ = ταβ + 2 ταβ + ...
Considering

(1)
ταβ = ταβ ,
corresponds to a Knudsen rst-order Chapman-Enskog approximation.
116 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

X (0)
eiα eiβ eiγ f i
i
2 ∗ ∗
ρ e δαβ u∗γ + δαγ u∗β + δβγ u∗α + ρ∗ u∗α u∗β u∗γ .


= (6.58)
D
After some cumbersome calculations the viscous stress tensor is given by

ταβ = −µ∗ ∂α∗ u∗β + ∂β ∗ u∗α + κ∗ ρ∗ e∗ ∇∗ · u∗ δαβ .

(6.59)

where

   
2
∗ ∗ 1 ∗ ∗ ∗ ∗ 1
µ = τ − ρ e =ρ τ − c2s (Θ + 1) .
D 2 2
is the shear viscosity and

2 ∗
κ∗ = µ ,
D
is the bulk viscosity for a mechanical system of material points, without internal
degrees of freedom.
Replacing ταβ in Eq. (6.54) by its expression given by Eq. (6.59), the
momentum balance equation at order 1 is obtained as

∂1∗ (ρ∗ u∗α ) = ∂β ∗ µ∗ ∂α∗ u∗β + ∂β ∗ u∗α




−∂α∗ [κ∗ ∇∗ · u∗ ] , (6.60)

So, the second order error terms resulting from a rst order approximation
of the stream term in Eq. (6.29) are absorbed into the momentum balance
equation as a numerical viscosity. The only restriction is that the dimensionless
relaxation time τ∗ must be greater than
1
2.

6.4.3 Energy conservation

1
Multiply Eq. (6.52) by
2 eiβ eiβ and sum on i,

   X !
1 1 (1) 1
∂1∗ ρ∗ e∗c,f + ρ∗ u∗2 + ∂α∗ 1− ∗ f i eiα eiβ eiβ = 0.
2 2τ i
2

Using

eiα eiβ eiβ = (eiα − u∗α ) eiβ eiβ + u∗α eiβ eiβ
= (eiα − u∗α ) eiβ − u∗β eiβ − u∗β + 2 eiβ − u∗β u∗β





+u∗α eiβ eiβ ,

6.4. FIRST-ORDER TRANSPORT EQUATIONS 117

we obtain

∗
=qα
z }| {
X (1) 1 1 X (1) ∗ 1 ∗ 2


f i eiα eiβ eiβ = f i (eiα − uα ) eiβ − uβ
i
2 2 i 2
X (1)
+u∗β f i (eiα − u∗α ) eiβ − u∗β .


i
| {z }
∗
=ταβ

So, at this order, the total energy conservation equation becomes

    
1 1
ρ∗ e∗c,f + ρ∗ u∗2 qα∗ + u∗β ταβ
∗


∂1∗ + ∂α 1− = 0.
2 2τ ∗
The heat ow vector is given by

1 X (1)
2
qα∗ = f i (eiα − u∗α ) eiβ − u∗β .
2 i
From Eq. (6.33)

(1)
(0) (0) f
∂0∗ f i + eiγ ∂γ ∗ f i = − i∗ ,
τ
Therefore

1 X
2 (0)
qα∗ = −τ ∗ ∂0∗ (eiα − u∗α ) eiβ − u∗β f i
2 i
1 X
2 (0)
−τ ∗ ∂γ ∗ (eiα − u∗α ) eiβ − u∗β eiγ f i ,
2 i

meaning that the following conditions must be satised for correctly retrieving
the heat ow vector


X
2 (0) 1 2
(eiα − u∗α ) eiβ − u∗β fi = f eq (ξα − uα ) (ξβ − uβ ) dξ = 0,
i
n0 c3
(6.61)

X
2 (0)
(eiα − u∗α ) eiγ eiβ − u∗β fi
i

1 2
= f eq (ξα − uα ) (ξγ − uγ ) (ξβ − uβ ) dξ. (6.62)
n0 c4
5
Condition given by Eq. (6.62) means that

5 Note that
118 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

X
2 (0) 2
(eiα − u∗α ) eiγ − u∗γ eiβ − u∗β = ρ∗ c4s (Θ + 1) δαγ


fi
i

The heat ow vector will, then, be given by

1
qα∗ = −τ ∗ ∂α∗ ρ∗ c2s (Θ + 1) c2s (Θ + 1)


2
∗1 2
= −τ cs ∂α∗ (P ∗ (Θ + 1))
2
τ∗ τ∗
= − (Θ + 1) c2s ∂α∗ P ∗ − ρ∗ c4s (Θ + 1) ∂α∗ Θ.
2 2

Therefore, the heat ow vector is dependent on the temperature gradient

as given by Fourier's law, with a dimensionless thermal conductivity

ρ∗ c4s
 
1
λ∗ = τ∗ − (Θ + 1) ,
2 2

but is also dependent on the pressure gradient.

6.4.4 Internal energy balance equation

Multiply Eq. (6.54) by u∗β

      
1 ∗ ∗2 1 δ
u∗β ταβ ταβ ∂α∗ u∗β ,


∂1∗ ρ u + ∂α∗ 1− ∗ = 1−
2 2τ 2τ

which is the kinetic energy balance equation. Therefore, the internal energy
balance equation will be in the rst order in Knudsen,

    
1 1
∂1 ρ∗ e∗c,f + ∂α∗ qα = − 1 − ∗ ταβ ∂α∗ u∗β .


1− (6.63)
2τ ∗ 2τ


1
2
f eq (ξα − uα ) (ξγ − uγ ) ξβ − uβ dξ
n0 c4
 2  2 
1 kT0 T 1 C2
= ρ∗ D
e− 2 Cα Cγ C 2 dC
c4 m T0 (2π) 2

1 C2
= ρ∗ c4s (Θ + 1)2 D
e− 2 Cα Cγ C 2 dC.
(2π) 2
| {z }
=δαγ
6.5. MACROSCOPIC EQUATIONS 119

6.5 Macroscopic equations

6.5.1 Mass conservation equation
The mass conservation equation is obtained by summing Eqs. (6.35) and
(6.50)

∂0∗ ρ∗ + ∂α∗ (ρ∗ u∗α ) = 0,

∂1∗ ρ∗ = 0,
giving

∂t∗ ρ∗ + ∂α∗ (ρ∗ u∗α ) = 0. (6.64)

6.5.2 Momentum balance equation

The momentum balance equation is obtained by summing Eqs. (6.39) and
(6.54)

∂0∗ ρ∗ u∗β + ∂α∗ ρ∗ u∗α u∗β + P ∗ δαβ = 0,

  
1
∂1∗ ρ∗ u∗β + ∂α∗


1− ταβ = 0,
2τ ∗
giving

  
1
ρ∗ u∗β ρ∗ u∗α u∗β ∗



∂t∗ +∂ α∗ + P δαβ = −∂ α∗ 1− ∗ ταβ .
2τ

6.5.3 Internal energy balance equation

The internal energy balance equation is obtained by summing Eqs (6.48)
and (6.63)

∂0∗ ρ∗ e∗c,f + ∂β ∗ ρ∗ e∗c,f u∗β = −P ∗ ∂β ∗ u∗β ,

    
1 1
∂1∗ ρ∗ e∗c,f + ∂α∗ 1 − ∗ qα = − 1 − ∗ ταβ ∂α∗ u∗β ,


2τ 2τ

giving

    
1 1
∂t ρ∗ e∗c,f ρ∗ e∗c,f u∗β ∗
∂β u∗β −∂α∗ qα − 1 − ∗ ταβ ∂α∗ u∗β .



+∂β = −P 1− ∗
2τ 2τ
120 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

6.6 Summary of the necessary conditions for re-

trieving the macroscopic transport equations
The collision model must satisfy

X
Ωi = 0,
i
X
Ωi eiβ = 0,
i
X
Ωi e2i = 0,
i

related, respectively to mass, momentum and kinetic energy preservation in

collisions.
The equilibrium distribution function, i.e., the solution of Ω i = 0, must
satisfy

X (0)
fi = ρ∗ , (6.65)
i
X (0)
f i eiα = ρ∗ u∗α , (6.66)
i
X (0)
f i eiα eiβ = ρ∗ u∗α u∗β + P ∗ δαβ , (6.67)
i

X (0) 1
eiα eiβ eiγ f i = f eq ξα ξβ ξγ dξ. (6.68)
i
n0 c3

X
2 (0)
(eiα − u∗α ) eiγ eiβ − u∗β fi
i

1 2
= f eq (ξα − uα ) (ξγ − uγ ) (ξβ − uβ ) dξ (6.69)
n 0 c4
When conditions given by Eqs. (6.65-6.68) are satised, the momentum
balance equation becomes

∂t∗ ρ∗ u∗β ∂0∗ ρ∗ u∗β + ∂1∗ ρ∗ u∗β

=
 
∗ ∗ ∗ ∗

1 ∗
= −∂α∗ ρ uα uβ + P δαβ − 1 − ∗ ∂α∗ ταβ ,
2τ
or

∂t ρ∗ u∗β + ∂α ρ∗ u∗α u∗β = −∂β P ∗ + ∂α µ∗ ∂α u∗β + ∂β u∗α − ∂β (κ∗ ∇ · u∗ ) .




(6.70)
6.7. THE VISCOUS STRESS TENSOR 121

Eq. (6.68) means that the equilibrium distribution must be represented by

a third order Hermite polynomial expansion. In two and three dimensions this
representation corresponds to a D2V17 and D3V39 LBE, respectively. When
a second order LBE such as the D2Q9 or D3Q27 is used, the Navier-Stokes
 3 
equations are only retrieved with errors O u∗ , u∗ θ .
Energy transfer in non-isothermal and full compressible problems can only
be correctly described when the condition given by Eq. (6.69) is satised.
This condition requires that the equilibrium distribution be represented by a
fourth-order Hermite polynomial expansion.

6.7 The viscous stress tensor

We rewrite Eq. (6.55) which gives the rst order approximation to the
viscous stress tensor
 (0) (0)

(1)
X
ταβ = −τ ∗ eiα eiβ ∂0∗ f i + eiγ ∂γ ∗ f i .
i
We need, rst, to calculate

!
X (0) 2
∂0∗ (ρ∗ e∗ ) δαβ + ∂0∗ ρ∗ u∗α u∗β .


∂0∗ eiα eiβ f i = (6.71)
i
D
From Eq. (6.48)

2
∂0∗ (ρ∗ e∗ ) = −∂γ ∗ ρ∗ e∗ u∗γ − ρ∗ e∗ ∂γ ∗ u∗γ .


D
On the other hand

∂0∗ ρ∗ u∗α u∗β = u∗β ∂0∗ (ρ∗ u∗α ) + u∗α ∂0∗ ρ∗ u∗β − u∗α u∗β ∂0∗ ρ∗ .

From the mass conservation equation

∂0∗ ρ∗ = −∂γ ∗ ρ∗ u∗γ

and from the momentum balance equation Eq. (6.39)

 
2 ∗ ∗
∂0∗ (ρ∗ u∗α ) ∗ ∗ ∗
= −∂γ ∗ ρ uα uγ + ρ e δαγ .
D
So, it results for Eq. (6.71)

!
X (0)
∂0∗ eiα eiβ f i
i
2 2
−u∗β ∂γ ∗ρ∗ u∗α u∗γ − u∗β ∂α∗ (ρ∗ e∗ ) − u∗α ∂γ ∗ ρ∗ u∗β u∗γ − u∗α ∂β ∗ (ρ∗ e∗ )



=
D  D
∗ ∗ ∗ ∗

2 ∗ ∗ ∗

2 ∗ ∗ ∗
+uα uβ ∂γ ∗ ρ uγ − ∂γ ∗ ρ e uγ + ρ e ∂γ ∗ uγ δαβ .
D D
122 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

Now, from Eq. (6.58)

!
X (0)
∂γ ∗ eiα eiβ eiγ f i
i
2 2 2
∂γ ∗ ρ∗ e∗ u∗γ δαβ + u∗β ∂α∗ (ρ∗ e∗ ) + ρ∗ e∗ ∂α∗ u∗β


=
D D D
2 2
+ u∗α ∂β ∗ (ρ∗ e∗ ) + ρ∗ e∗ ∂β ∗ u∗α + ∂γ ∗ ρ∗ u∗α u∗β u∗γ


D D
Adding the two last equations

! !
X (0) X (0)
∂0∗ eiα eiβ f i + ∂γ ∗ eiα eiβ eiγ f i
i i
−u∗β ∂γ ∗ ρ∗ u∗α u∗γ − u∗α ∂γ ∗ ρ∗ u∗β u∗γ + u∗α u∗β ∂γ ∗ ρ∗ u∗γ




=
 2
∗ ∗ ∗ ∗ 2 2
ρ∗ e∗ ∂γ ∗ u∗γ δαβ + ρ∗ e∗ ∂α∗ u∗β + ∂β ∗ u∗α




+∂γ ∗ ρ uα uβ uγ −
D D

But

∂γ ∗ ρ∗ u∗α u∗β u∗γ

u∗α ∂γ ∗ ρ∗ u∗β u∗γ + u∗β ∂γ ∗ ρ∗ u∗γ u∗α − u∗β u∗α ∂γ ∗ ρ∗ u∗γ

=

So, the viscous stress tensor, Eq. (6.55) is, nally, given by

(1)
ταβ = −µ∗ ∂α∗ u∗β + ∂β ∗ u∗α + κ∗ ∇∗ · u∗ δαβ .


(6.72)

where

2 ∗ ∗
µ∗ = τ ∗ ρ e
D
and

2 ∗
κ∗ = µ
D

6.8 Exercises
1. Show that when the D2Q9 LBE is used with an equilibrium distribution
given by

ρ∗ + ρ∗a2su∗α eiα +
" #
eq ∗
f i,q=6 = Wi ρ a4s , (6.73)
 
2 ρ∗ u∗α u∗β + Θ eiα eiβ − 1
a2s δαβ

then Eq. (6.68) is not satised.

6.9. LBE WITH A FORCING TERM 123

2. At order 1, when the collision operator is represented by a BGK relax-

ation term

  f (1)
∂0 f (0) + ∂α ξα f (0) = − . (6.74)
τ
the Knudsen rst order approximation to the viscous stress tensor is given by

 
(1)
τβγ = f (1) (ξβ − uβ ) (ξγ − uγ ) dξ = f (1) ξβ ξγ dξ, (6.75)

Show that, in two-dimensional systems, when the equilibrium distribution

is given in terms of its full Hermitian representation

H (0) , Hx(1) , Hy(1) , Hxx

(2) (2)
, Hxy (2)
, Hyy (3)
, Hxxy (3)
, Hxyy (4)
.Hxxyy ,
(1) (1) (1)
the component τxy is correctly retrieved, but not τxx and τyy . What do you
think would be the consequences of this nding?

6.9 LBE with a forcing term

6.9.1 First-order discretization of the stream term
For further reference we will consider a Chapman-Enskog analysis of the
following rst order LBE, written in its dimensionless form given by

f i (x + ei h, t + δ) − f i (x, t) = Ωi + S̄i uO h2 ,


(6.76)

where S̄i is a momentum source term given by

eq
S̄i = g∗ . (ei − u∗ ) f i ,
gis a dimensionless acceleration, g∗ = gδ 2 /h and Ωi is a collision term. usually
a BGK relaxation term [4].
By expanding the equilibrium distribution in Hermite polynomials and ne-
2
glecting all the terms of order u∗ and higher, the source term S̄i can be also
written as

 
1 ∗ ∗ 1 ∗ ∗ ∗ ∗

2


S̄i ≈ W i ρ gα e iα + g u + g u e iα e iβ − c δ αβ (6.77)
c2s 2c2s α β β α s

Similarly to the LBE without a forcing term, the Taylor expansion of

f i (x + ei h, t + δ)
in Eq. (6.76) leads to
124 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

1 1
δ∂t f i + heiα ∂α f i + δ 2 ∂tt f i + h2 eiα eiβ ∂αβ f i + hδeiα ∂tα f i + ... = Ωi + S̄i .
2
| 2 {z }
second order errors

(6.78)

As before, let L be a macroscopic length and Γ a macroscopic time and

dene

h
= ,
L

and

δ
χ= .
Γ

Both  and χ are required to be small parameters, the rst one being related
to the geometric resolution of the physical system and the second one to the
ratio between the time step δ and the time Γ that gives a scale for the time
variation of the local elds. We consider χ ≈  and replace t by t+ = t/Γ and
x by x+ = x/L in Eq. (6.78),

2
1χ
 
2  ∂t t f i
+ +
χ Ωi S̄i
∂t+ f i + eiα ∂α+ f i +  + 12 eiα eiβ ∂α+ β + f i  + ... = + ,
| {z } +χeiα ∂t+ α+ f i

|{z} 
|{z}
O(1) | {z } O(−1 ) O(1)
O()

where we have considered that the source term S̄i is O(). Therefore, when
 → 0, the collision term on the right-hand side of the above equation becomes
dominant with respect to the stream terms on the left-hand side and the source
term.

Consider, now, the following expansion of fi when  → 0,

(0) (1) (2)

fi = fi + f i + 2 f i + ... .

Using the Enskog decomposition of the time derivative

∂t+ = ∂0+ + ∂1+ + 2 ∂2+ + ...,

we obtain
6.9. LBE WITH A FORCING TERM 125

χ (0) (0)
∂0+ f i + eiα ∂α+ f i +
| {z }
O(1)
(1) (1) (0) 2 (0)
!
1χ
χ∂0+ f i + eiα ∂α+ f i + χ∂1+ f i +
2  ∂0 0 f i
+ +
(0) (0) + ...
+ 12 eiα eiβ ∂α+ β + f i + χeiα ∂0+ α+ f i
| {z }
O()
(0)
Ωi S̄i (1) (2)
= + + Ωi + Ωi + ... . (6.79)
 } |{z}
| {z  |{z} | {z }
O(1) O()
O(1/) O(1)

So, in O(1/):
(0)
Ωi = 0.
(0)
and this equation gives the zero-th order approximation fi to the solution as
related to the equilibrium solution

(0) eq
fi = fi ,
usually written as a polynomial approximation to the Maxwell-Boltzmann equi-
librium distribution.
In O(1), returning to the LB variables,

(0) (0) (1)

∂0∗ f i + eiα ∂α∗ f i = Ωi + S̄i . (6.80)

Finally, in O()
(1) (1) (0) (0)
∂0∗ f i + eiα ∂α∗ f i + ∂1∗ f i + 12 ∂0∗ 0∗ f i
(0) (0) (2) (6.81)
+ 12 eiα eiβ ∂α∗ β ∗ f i + eiα ∂0∗ α∗ f i = 2 Ω i .
Eq. (6.81) can be rearranged using

1 (0) 1 (0) (0)

∂0∗ 0∗ f i + eiα eiβ ∂α∗ β ∗ f i + eiα ∂0∗ α∗ f i
2 2
1  (1) 
= (∂0∗ + eiα ∂α∗ ) Ωi + S̄i .
2

For a BGK model

(1)
(1) f
Ωi = − i∗ ,
τ
and so, Eq. (6.81) becomes
126 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

 
1 (1) (0) 1 S̄i (2)
1− ∗ (∂0∗ + eiα ∂α∗ ) f i +∂1∗ f i + (∂0∗ + eiα ∂α∗ ) = Ωi . (6.82)
2τ 2 

When we sum Eq. (6.80) on i, we get the zero − th order aproximation of

the mass conservation equation in LB dimensionless variables

∂0∗ ρ∗ + ∂α∗ (ρ∗ u∗α ) = 0, (6.83)

Nevertheless, the sum on i of Eq. (6.82) gives

1
∂1∗ ρ∗ = − ∂α∗ (ρ∗ gα∗ ) . (6.84)
2
Summing Eqs (6.83) and (6.84),

1
∂t∗ ρ∗ + ∂α∗ (ρ∗ u∗α ) = − ∂α∗ (ρ∗ gα∗ ) , (6.85)
2
which is the mass conservation equation with a spurious, nonphysical, term on
6
its right hand side .
The momentum balance equation can be also obtained from Eqs. (6.80)
and (6.82), giving

  
1
∂t∗ ρ∗ u∗β + ∂α∗ ρ∗ u∗α u∗β + P ∗ δαβ = ρ∗ gβ∗ − ∂α∗ ∗



1− ταβ + Eβ ,
2τ ∗
(6.86)
where Eβ is, also, a spurious nonphysical term given by

1
∂0∗ ρ∗ gβ∗ + ∂α∗ ρ∗ gα∗ u∗β + gβ∗ u∗α .



Eβ = −
2
Therefore, we see that although the second order errors were absorbed by
the viscous term, the use of a rst order approximation for the stream term is
at the origin of the spurious term into the momentum balance equation, Eq.
(6.86), and is responsible for breaking the mass conservation, Eq. (6.85). In
the following we consider a second order scheme for the stream term.

Exercise
Try to derive the momentum balance equation, Eq. (6.86) from Eqs.
(6.80) and (6.82).

6 In physical unities Eq.(6.85) would be written as

δ
∂t ρ + ∂α (ρuα ) = − ∂α (ρgα )
2
.
6.9. LBE WITH A FORCING TERM 127

6.9.2 Second-order discretization for the source term

We adopt a mixed scheme based on the second-order discretization of the
stream term when the only term on the r.h.s. of the Boltzmann equation is the
source term, For a BGK collision model this procedure leads to

f¯ieq − f¯i 1

f i (x + ei h, t + δ) − f i (x, t) = + S̄i (x, t) + S̄i (x + ei h, t + δ) .
τ∗ 2
(6.87)
In this case, we must evaluate S̄i (x + ei h, t + δ) in terms of S̄i (x, t). We
get

S̄i (x + ei h, t + δ) = S̄i (x, t) + (∂t∗ + eiα ∂α∗ ) S̄i + O h2 .

So, Eq. (6.87) decomes

f¯ieq − f¯i 1
f i (x + ei h, t + δ) − f i (x, t) = + S̄i + (∂t∗ + eiα ∂α∗ ) S¯i ,
τ∗ 2
and the equation analogous to Eq. (6.79) is given by.

χ (0) (0)
∂0+ f i + eiα ∂α+ f i +
| {z }
O(1)
(1) (1) (0) 2 (0)
!
1χ
χ∂0+ f i + eiα ∂α+ f i + χ∂1+ f i +2  ∂0 0 f i
+ +
(0) (0) + ...
+ 12 eiα eiβ ∂α+ β + f i + χeiα ∂0+ α+ f i
| {z }
O()
¯(0) (1) (2)
1 f¯ieq − fi S̄i fi fi 1
= ∗
+ − ∗
+ − ∗
+ (∂0∗ + eiα ∂α∗ ) S̄i . (6.88)
| τ
{z } | {z τ } | τ 2 {z }
O(1/) O(1) O()

Grouping the terms with the same order we get

in O(1/):
(0) eq
fi = fi ,
usually written as a polynomial approximation to the Maxwell-Boltzmann equi-
librium distribution.
In O(1):
(1)
(0) (0) fi
∂0∗ f i + eiα ∂α∗ f i = − + S̄i . (6.89)
τ∗
Finally, in O()
128 CHAPTER 6. CHAPMAN-ENSKOG ANALYSIS

 (1) 
(1) (1) (0) 1 fi
∂ 0∗ fi + eiα ∂ α∗ fi + ∂ 1∗ fi + 2  (∂0
∗ + eiα ∂ ) − τ ∗ + S̄i .
α∗
(2)
fi
= −2 τ∗ + 12  (∂0∗ + eiα ∂α∗ ) S̄i .
(6.90)
The above equation reduces to

 (1) 
(1) (1) (0) f
∂0∗ f i + eiα ∂α∗ f i + ∂1∗ f i + 12  (∂0∗ + eiα ∂α∗ ) − τi∗ .
(6.91)
(2)
fi
= −2 τ∗ ,

Eq. (6.91) has the same form of Eq. (6.34): the second order expansion of
the stream term canceled the spurious non-physical terms related to the mo-
mentum source term S̄i and the macroscopic equations are correctely retrieved.
Chapter 7

Beyond BGK:
moments-based methods
Instability issues are frequent in LB simulation with the BGK relaxation
term and increasing the stability ranges is a very actual subject in LBM re-
search. In this section we review the theory that was developed to increase our
understanding on these issues and the methods to enhance the stability limits.

7.1 Introduction
Picturing a uid as a mechanical system of particles has a long and suc-
cessful history, written by outstanding philosophers and scientists, and leading
to the solution of important challenges to the human knowledge. Indeed, the
principle of energy conservation was only postulated by Helmholtz [43], when
matter was recognized as composed by molecules. The works of Joule [49],
Clausius [19], Maxwell [67], Boltzmann [5], [6] and Gibbs [31], were funda-
mental for the borning of the `Mechanical Theory of Heat ' and for clarifying the
meaning of the second principle of thermodynamics and entropy for a system
of particles ruled by the deterministic Newtonian mechanics.
`Dissipation ' means a one-way transformation of macroscopic energy into
heat  the mechanical energy of the moleculesand this happens because
molecules collide between themselves exchanging momentum and kinetic en-
ergy.
When a uid ows, these molecular properties are transferred from one
point to another by two kind of mechanisms: diusion and advection. Diu-
sion is related to the random motion of molecules and is responsible for entropy
growth. In the reverse sense, advection means organization: molecular proper-
ties are transferred with the bulk ow, along the streamlines.
Diusion is dominant at the small Reynolds regime, but when the Reynolds
number increases, diusion becomes less eective and properties are mainly
transferred by advection.

129
130 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

A fast relaxation to an equilibrium state corresponds to a low kinematic

viscosity. So, in low-viscous uids the interactions between the particles are
so numerous and eective that momentum and energy between neighboring
particles is quickly balanced; the system remains close to the equilibrium state.
As a result, the net ux of momentum is small in the moving frame of reference
or, in other words, transport of momentum is sustained mainly by advection
rather than by diusion, which conforms to a high Reynolds number ow.
This aspect is used in this chapter as the guiding principle when designing
a particular lattice-Boltzmann scheme for numerical simulation of uid ows
with improved stability properties.
A mechanical system of particles can be described by a distribution function
f (x, ξ, t) governed by the Boltzmann equation. The Boltzmann integral colli-
sion term is dicult to handle and kinetic models were proposed for describing
the interaction between molecules, based on the relaxation of this distribu-
tion to the equilibrium due to collisions [4], [34], and on the intermolecular
attraction forces, responsible for cohesion [88], [80], [39].
For each point x and time t, the distribution f (x, ξ, t) can be considered as
a map from the velocity space onto the real numbers and, so, as an element of
the Hilbert space H: ξ D → R, it can be developped into Hermite or any other
innite set of orthogonal polynomials that generate the space H.
eq
This distribution can be decomposed into an equilibrium f and a non-
eqilibrium part f neq .
The coecients of the f neq expansion are the non-equilibrium macroscopic
moments and each one of these macroscopic moments has an advective and
a diusive component. The diusive parts are dened as the peculiar or
central moments of the non-equilibrium distribution and hence they are re-
lated to the ux of momentum, energy and other higher-order moments in
the moving frame of reference. The advective non-equilibrium moments are
polynomials of local ow velocities with convective non-equilibrium moments
of lower order as coecients. They do not have any contribution to the ow
when the local velocity is null, but are dominant in high Reynolds number (Re)
ows.
In LBM discrete framework, the purpose is not to reproduce the continuous
kinetic model itself. In most times, what is wanted is to describe thermo-
hydrodynamics when the Knudsen number Kn → 0. As a consequence, the
Fourier-Navier-Stokes equations are dierential equations involving low-order
moments. Nevertheless, in the kinetic approach to thermohydrodynamics, the
distribution f embodies all the macroscopic moments, including the ones whose
order are higher than the order of the moments one wants to retrieve [23]. These
`ghost ' moments can contribute to stability issues in LB simulations.
In fact, the most commonly used lattice Boltzmann collision model is the
single relaxation-time BGK model, [4]. This model is able to asymptotically
represent near incompressible ows. Nevertheless, it suers from instability
drawbacks, especially at high Reynolds numbers and among the several conjec-
tures for the appearance of these issues, one of them is based on the inuence
of ghost-moments on the simulation.
7.1. INTRODUCTION 131

Until presently, the great majority of LB models are second order models
whose moment space are tailored to a particular discrete velocity set (D2Q9,
D3Q27, D3Q19) by, e.g., resorting to the Gram-Schmidt orthogonalization pro-
cedure.
Improving stability of these LB schemes has been addressed by several
authors and some solutions have been proposed. For recent reviews, see e. g.
Brownlee et al. [9], Golbert et al. [32] and references therein.
The rst solutions were based on the use of multiple relaxation times (MRT)
tuned with the help of a linear stability analysis, D'Humieres [23], D'Humières
et al. [24], Dellar [22], Lallemand and Luo [57], and Xu and Sagaut [90]. For a
recent paper dealing with MRT coupled with the Kupershtokh [53], [54] exact
dierence method (EDM) in the study of non-isothermal droplet evaporation,
see Albernaz et al. [2] and references therein.
The entropic LBGK method appears as a second solution and was conceived
based on the maximization of the entropy, by locally tuning the BGK single
relaxation time, Karlin et al. [51]. Recently, a new extension of LB schemes
was proposed, namely the entropic stabilizer, Karlin et al. [50]. Unlike the
entropic LBGK model, this Karlin et al. extension do not locally alter the
viscosity, but rather relies on modifying the relaxation time for the higher-order
moments (i.e. the moments beyond the stress tensor) which do not contribute
to the viscosity. In this respect, this extension is akin to the already mentioned
relaxation parameter tuning for MRT schemes.
Finally, the cascaded model and its subsequent improvements, factorized
and cumulant LB schemes [28], [78], [29], base its approach on peculiar mo-
ments and particular choices for the relaxation times for assuring Galilean in-
variance. Nevertheless, these models are, also, restricted to second order D2Q9
and D3Q27 LBE.
Considering the Boltzmann equation in fact, a continuous kinetic model
able to fulll the main conditions that are required for the description of a given
physical system a lattice Boltzmann equation (LBE) can be considered as
a discrete form of a kinetic projection of this kinetic model onto a nite Hq
subspace of H.
The connection between the LBE and such a projection was rst established
by He and Luo [40] and Abe [1] in 1997, who directly derived the LBE from
the Boltzmann equation for some widely known lattice-Boltzmann equations:
D2Q9, D2Q6, D2Q7 and D3Q27. This was performed by the discretization
of the velocity space, using the Gauss-Hermite and Gauss-Radau quadrature.
Excluding the above mentioned lattices, the discrete velocity sets obtained with
this kind of quadrature do not generate regular space lling lattices.
Shan et al. [82] and Philippi and co-workers [75], [84] established a sys-
tematic link between Hermitian representations and the sets of LBE discrete
velocities. This approach led to recognize a LB equation as an approximation
of the Boltzmann or any kinetic equation based on a nite and discrete
set of nb molecular velocities ξi , i = 0, ..., nb − 1.
This set is directly related to the projection of the kinetic equation onto a
subspace generated by the rst q Hermite polynomials and the order N of this
132 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

set gives the order of approximation of the LBE to its continuous counterpart.
As considered by Higuera and Jimenez [45] in 1989, the main purpose
of LBM continues to be, nowadays, to solve hydrodynamic problems, but,
presently, the conception of accurate and stable LB equations enabling to
achieve high Reynolds ows and to describe uid systems under strong tem-
perature non-equilibrium and/or complex interactions are the main challenges
for this kinetic method. Instability issues are frequent in LB simulation and
increasing the stability ranges is a very actual subject in LBM research.
As mentioned above, LBGK equations are low-order approximations to the
continuous BGK equation and, so, their solutions suer from errors related to
uncontrolled high-order moments. Consequently, instability issues are to be
expected.
Therefore, considering the nature of these issues, three main alternatives
appear at hand for enlarging the stability ranges without the help of free pa-
rameters, ad-hoc assumptions and/or an optimization step. The rst one is to
use lattices with increased dimensionality, i.e., based on high-order LB equa-
tions with increased velocity sets. This alternative was investigated by Siebert
et al. [85]. The second alternative is to add high order Hermite polynomial
tensors to the equilibrium distribution, trying to reduce the eect of their re-
lated moments on stability. This alternative was, also, utilized by Ref. [85] for
improving the stability of the D2Q9 LBE.
The third alternative is to work with moments-based LB equations, i.e., to
rewrite the LB equation in such a manner as to lter the undesirable ghost
moments from the numerical scheme. This alternative was investigated by
Ladd [55], Chen et al. [15], Zhang et al. [91] and by Latt and Chopard [58] for
the D2Q9 model and was very recently compared with the entropic stabilizer
method of Ref. [50] by Mattila et al. [66]. Although it is also a moments based
method, it diers from MRT because instead of being attributed with a given
relaxation time, all the ghost moments are ltered,
In this chapter, we present a method based on Mattila et al. [65] for
dealing with general ow problems which require high-order LBE, based on a
systematic improvement of the standard regularization idea.
In this method, for each Hermitian representation, both the equilibrium,
f eq , and non-equilibrium, f neq , parts are expanded on the nite Hermitian
basis of the subspace Hq of H.
Nevertheless, the high-order non-equilibrium moments do not t into the
Hermitian representation and, consequently, are sources of errors.This is the
case of the viscous stress tensor ταβ , which
requires a full third order Hermitian
representation and is the source of O u3 errors in the macroscopic equations
when a second order LBE such as the D2Q9 or D3Q27 is used for the ow
simulation. This is also the case of the heat ow vector qα , which requires a
full fourth order Hermitian representation.
In Mattila et al. proposal, [65], the expansion of the non-equilibrium distri-
bution is carried out on a pair with the equilibrium distribution, but the diu-
sive parts are retained only up to the last physically relevant non-equilibrium
moment. Beyond that, the diusive parts are ltered out which leads to spe-
7.2. KINETIC PROJECTIONS 133

cic recurrence relations expressing the high-order non-equilibrium moments

in terms of the local ow velocity and lower rank non-equilibrium moments.
Technically the above procedure corresponds to regularization in the pecu-
liar (or central) moment space. The procedure is conveniently and eciently
implemented using the recurrence relations here derived. Note that the usual
regularization procedure operates with the convective moments and, in particu-
lar, the selected high-rank non-equilibrium moments are ltered out completely.
The motivation is to acknowledge the essential role of advective transport
in high Reynolds number ows as discussed at the beginning. Similar reasoning
motivated the Cascaded lattice-Boltzmann scheme also operating with central
moments [27]. Nevertheless, as above commented, Cascaded LB schemes as well
as its subsequent improvements, Factorized and Cumulant LB schemes [28],
[78], [29], have so far been presented only for second-order, standard discrete
velocity sets (e.g., D2Q9 and D3Q27). The procedure here presented is general
and directly applicable to high-order velocity sets.
Furthermore, the proposed scheme relies purely on Hermite polynomial ex-
pansions of the distribution function, i.e. the appropriate moments are imme-
diately known. Thus, there is no need to tailor a moment space representation
for the populations fi related to a particular discrete velocity set by, e.g., re-
sorting to the Gram-Schmidt orthogonalization procedure. These features are
estimulating factors for an increasingly number of papers on the use of LB
schemes based on Hermite polynomial tensors (see e.g. [91], [14], [63], [73]).

7.2 Kinetic Projections

Figure 7.1: The Hilbert space solution of the Boltzmann equation.

The distribution f (x, ξ, t), solution of a kinetic equation,

∂t f + ξ·∇x f = Ω, (7.1)
134 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

may be considered, for each (x, t) as a map from the continuous velocity space
ξD onto the space R of real numbers (Figure 7.1). In fact, it belongs to the
Hilbert space H of square integrable functions f : ξ D −→ R and may be
written in terms of an orthogonal basis of H that will be considered as the
innite set of Hermite polynomial tensors,

X a(θ) (x, t)
f (x, ξ, t) = H (θ) (ξ0 ) , (7.2)
nx !ny !nz !
θ

where θ indicates the order of tensor H (θ) , nα is the number of times the index
α = x, y or z appears repeated, nx + ny + nz = θ, ξ0 = ξ/ξ is a dimension-
1
less velocity, ξ = (kT0 /m) 2 is a mean microscopic speed, k is the Boltzmann
constant, T0 is a reference temperature and m is the mass of each particle.
Using the more common components notation for the Hermite polynomials
(1) (N )
H (0) , Hα0 , ..., Hα0 α1 ...αN , where αi = x, y or z, the rst Hermite polynomial
tensors are,

H (0) = 1,
Hα(1) = ξ0,α ,
(2)
Hαβ = ξ0,α ξ0,β − δαβ ,
(3)
Hαβγ = ξ0,α ξ0,β ξ0,γ (7.3)

− (δβγ ξ0,α + δαγ ξ0,β + δαβ ξ0,γ ) ,

(4)
Hαβγδ = ξ0,α ξ0,β ξ0,γ ξ0,δ
 
δγδ ξ0,α ξ0,β + δβδ ξ0,α ξ0,γ + δβγ ξ0,α ξ0,δ
−
+δαδ ξ0,β ξ0,γ + δαγ ξ0,β ξ0,δ + δαβ ξ0,γ ξ0,δ
+ (δαβ δγδ + δαγ δβδ + δαδ δβγ ) ,
and the following recurrence relation can be derived for the Hermite polynomial
tensors

n
(n−1)
X
Hα(n+1)
0 α1 ...αn
= ξα0 Hα(n)
1 ...αn
− δα0 αk Hα1 ...αn /αk , (7.4)
k=1
where the notation α1 ...αn /αk means that for a given k, only the indices α1 ,
..., αk−1 , αk+1 , ..., αn are taken into account.
The Boltzmann equation, Eq. (7.1) has an equilibrium solution, the Maxwell-
Boltzmann (MB) distribution [67], [5], [7],

(ξ0 −u0 )2

eq 1 e− 2(Θ+1)
f =n D/2
, D
D/2
(Θ + 1)
(2π) ξ
where n is the number density of particles, D is the Euclidean dimension of the
velocity space, D = 1, 2 or 3 and u0 = u/ξ is a dimensionless local velocity,
and Θ = T /T0 − 1 is the relative deviation from the reference temperature T0 .
7.2. KINETIC PROJECTIONS 135

In its simplest form, the interaction term Ω in Eq. (7.1) is modeled with a
BGK single relaxation term [4]

f − f eq
Ω=− , (7.5)
τ
where τ is a relaxation time.
Now, recall that f eq belongs to the Hilbert space H and its dimensionless
eq
form f may be developed in terms of the Hermitian basis of H as

2
ξ0
D
eq f eq ξ ∗ e− 2
f = =ρ D/2
no (2π)
X a(θ)
eq (u0 , Θ)
× H (θ) (ξ0 ) ,
nx !ny !nz !
θ

where ρ∗ = n/n0 = ρ/ρ0 , n0 , ρ0 are, respectively, a reference number and mass

(θ)
density of particles, and the dimensionless equilibrium moments aeq are given
by

a(θ)
eq (u0 , Θ)
 (ξ0 −u0 )2
1 e− 2(Θ+1)
= D/2 D/2
Hn(θ)
x ,ny ,nz
(ξ0 ) dξ0 .
(2π) (Θ + 1)

In components notation, the rst equilibrium moments are

a(0)
eq = 1,

a(1)
eq,α = u0,α ,
(2)
aeq,αβ = (u0,α u0,β + Θδαβ ) ,
(3) u0,α u0,β u0,γ
aeq,αβγ = , (7.6)
+Θ (δβγ u0,α + δαγ u0,β + δαβ u0,γ )
(4)
aeq,αβγδ = u0,α u0,β u0,γ u0,δ
 
δαβ u0,γ u0,δ + δαγ u0,β u0,δ
+ Θ  +δαδ u0,β u0,γ + δβγ u0,α u0,δ 
+δβδ u0,α u0,γ + δγδ u0,α u0,β
+ Θ2 (δαβ δγδ + δαγ δβδ + δαδ δβγ ) .

Consider now the subspace Hq of H generated by a nite set of q Hermite

polynomial tensors (Figure 7.2). The projection of the Boltzmann equation,
Eq. (7.1) onto this subspace will have a solution that can be written in its
dimensionless form as,
136 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

Figure 7.2: Subspace Hq=10 of H generated by a nite set of 10 Hermite poly-

nomial tensors and related to the D2V17 LBE.

2
ξ0
D
fξ e− 2
f= = D/2
n0 (2π)
X a(θ) (x, t)
× H (θ) (ξ0 ) . (7.7)
(θ)
n x !n y !n z !
H ∈Hq

Function f is unambiguously determined by their Hermitian moments a(θ) (x, t)


a(θ) = f H (θ) (ξ0 ) dξ0 . (7.8)

The subspace Hq is directly related to the set of b discrete velocities ξ0,i ; i =

0, ..., b − 1 used to represent the velocity space This formal relationship was
found by Shan et al. [82] and Philippi and co-workers [75], [84] and is based
on the preservation of the inner product

   
H (θ) ∗ H (η) = H (θ) ∗ H (η) , (7.9)
d c

between any two Hermite polynomials H (θ) that form the basis of Hq , where

 
H (θ) ∗ H (η)
c
 2
1 ξ0
= D e− 2 H (θ) (ξ0 ) ∗ H (η) (ξ0 ) dξ0 , (7.10)
(2π) 2
7.3. INCREASING THE ORDER OF APPROXIMATION 137

 
H (θ) ∗ H (η)
d
b −1
nX
= Wi H (θ) (as ei ) ∗ H (η) (as ei ) . (7.11)
i=0

In the above equations, Wi are quadrature weights and as is the scaling

factor that relates the lattice vectors ei to the dimensionless velocities, ξ0,i =
as ei , [75]. Since the Hermite polynomials are orthogonal in the continuous
velocity space, the above condition means that, when considered as mappings
from the discrete velocity space onto the real numbers, their inner product
must be zero or the square of their norm, when identical,

b −1
nX h i2
Wi H (θ) (as ei )
i=0
 2 i2
1 ξ0
h
= D/2
e− 2 H (θ) (ξ0 ) dξ0 . (7.12)
(2π)

The above relationship results in a set of equations for H (θ) ∈ Hq , that

can be solved for the scaling factor as and weights Wi when the poles ei are
prescribed (see Philippi et al. [75] for details about the prescribed abcissas
method)
A very popular example of such a subspace is the second order Hq=6
Hilbert subspace, related to the D2Q9 LBE and generated by the rst 6 two-
(0) (1) (1) (2) (2) (2)
dimensional Hermite polynomial tensors H0,0 , H1,0 , H0,1 , H2,0 , H0,2 and H1,1
. This Hermitian basis leads to the well known second order equilibium distri-
bution

2 ∗
" #
eq ∗
1+
 as uα eiα +
f i,q=6 = Wi ρ a4s , (7.13)

2 u∗α u∗β eiα eiβ − 1
a2s δαβ

where u∗ = as u0 and the equilibrium distribution parameters were found from

√ 4 1 1
the solution of Eq. (7.12) as as = 3, W0 = 9, W1 = 9 and W2 = 36 .

7.3 Increasing the order of approximation

LB equations are low-order approximations to the continuous Boltzmann
equation and their solutions suer from errors related to uncontrolled high-
order moments. Consequently, instability issues are to be expected. Therefore,
considering the nature of these issues, three main alternatives appear at hand
for enlarging the stability ranges without the help of free parameters and/or
an optimization step.
138 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

Figure 7.3: Hermitian basis of the Hilbert subspaces Hq=6 in (a) and Hq=9 in
(b), related to the D2Q9 LBE.

7.3.1 Adding Hermite polynomials to subspace HN

The rst alternative is to add high order Hermite polynomial tensors to the
equilibrium distribution, trying to reduce the eect of their related moments
on instability. This alternative was utilized by Siebert et al. [85] for improving
the stability of the D2Q9 LBE.
In the second order Hilbert subspace H2 , related to the D2Q9 space lling
lattice, the Hermite polynomial tensors that generate this subspace are,

(0) (1) (1) (2) (2) (2)

H0,0 , H1,0 , H0,1 , H2,0 , H0,2 , H1,1 (7.14)

(θ)
where, as usual, the indexes nx and ny in Hnx ,ny indicates, respectively, the
number of times the index x and y appear repeatedly in the polynomial tensor,
with nx + ny = θ .
Recall [75], [84], that the norm preservation condition of these polynomial
tensors in the discrete space

b −1
nX  2
Wi Hn(θ)
x ,ny
(as ei )
i=0
 2 2
1 ξ0

= D/2
e− 2 Hn(θ)
x ,ny
(ξ0 ) dξ0 = nx !ny !, (7.15)
(2π)

results in a system of 4 independent equations that denes the weights

√ W0 =
16 4 1
,W1 = ,W2 = and the scaling factor as = 3 for the D2Q9 LBE. In
36 36 36
isothermal conditions, the equilibrium distribution for this LBE is then, given,
in lattice dimensionless variables, by

2 eq(θ)
eq
X anx ,ny (θ)
fi,q=6 = Wi H (as ei ) (7.16)
nx !ny ! nx ,ny
θ=0

where
7.3. INCREASING THE ORDER OF APPROXIMATION 139


1 (ξ0 −u0 )2 nx ny
aeq(θ)
nx ,ny = D/2
e− 2 Hn(θ)
x ,ny
(ξ0 ) dξ0 = ρ∗ u∗x u∗y aθs .
(2π)
(3) (3) (4)
Nevertheless, two third order H2,1 and H1,2 and one fourth order H2,2
polynomial tensors have their norm, Eq. (7.15) unaltered when the D2Q9
weigths Wi and scaling factor as are used. This allows the inclusion of the
related moments in the equilibrium distribution, which takes on the following
modied form

aeq(θ)
 
P2 nx ,ny (θ)
 θ=0 nx !ny ! Hnx ,ny (as ei ) 
eq eq(3) eq(3)
fi,q=9 = Wi ρ∗  + a2,1 H (3) (as ei ) + a1,2 H (3) (as ei ) . (7.17)
 
 2!1! 2,1 1!2! 1,2 
eq(4)
a2,2 (4)
+ 2!2! H2,2 (as ei )

This full set of q=9 Hermite polynomial tensors is show in Figure 7.3.b
and can be understood [17], as the cartesian product of the sets
n o
(0) (1) (2)
H0 (as eix ) , H1 (as eix ) , H2 (as eix )

and
n o
(0) (1) (2)
H0 (as eiy ) , H1 (as eiy ) , H2 (as eiy ) .

Consider, now, the more usual Hermitian second order basis of the D2Q9
LBE, composed by the rst 6 two-dimensional Hermite polynomial tensors
(1) (1) (2) (2) (2)
H (0) , Hx , Hy , Hxx , Hxy and Hyy (Figure 7.3.a).

Table 2. Hilbert subspace Hq=9 and corresponding

equilibrium moments
(θ) eq(θ)
Order θ Hnx ,ny Expression anx ,ny
(0)
0 H0,0 1 ρ∗
(1)
1 H1,0 as eix ρ∗ as u∗x
(1)
1 H0,1 as eiy ρ∗ as u∗y
(2) 2
2 H2,0 a2s e2ix − 1 ρ∗ a2s u∗x
(2) 2
2 H0,2 a2s e2iy − 1 ρ∗ a2s u∗y
(2)
2 H1,1 a2s eix eiy ρ∗ a2s u∗x u∗y
(3) 2
ρ∗ a3s u∗x u∗y


3 H2,1 a2s e2ix − 1 as eiy
(3) 2
ρ∗ a3s u∗y u∗x


3 H1,2 a2s e2iy − 1 as eix
(4) 2 2
ρ∗ a4s u∗x u∗y



4 H2,2 a2s e2ix − 1 a2s e2iy − 1

The inner product of ξβ ξγ by the second term on the l.h.s. of Eq. (6.33)
(3) (3) (3) (3)
requires the full set Hxxx , Hxxy , Hxyy and Hyyy , to belong to Hq . Since the
140 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

D2Q9 is a nine velocity set, the best it is possible to do is to add the polynomials
(3) (3) (4)
Hxxy , Hxyy and Hxxyy to the Hq subspace, because the discretization condition
given by Eq. (7.12),

b −1
nX h i2
Wi H (θ) (as ei ) = nx !ny !,
i=0
√
is satised by all these added polynomials when 3, W0 = 49 , W1 = 19
as =
1
and W2 =
36 (see Refs. [75], [85] for further details)
From Figure 7.3.b it is clearly seen that τxy is the only non-equilibrium
second-order moment that can be retrieved from the Hermitian basis of the
(3) (3)
subspace Hq=9 because Hxxy and Hxyy belong to the Hermitian basis of Hq=9 .
(3) (0)
The moments τxx and τyy are also dependent on the inner products Hxxx ∗ f
(3) (0) (3)
and Hyyy ∗ f , respectively and these inner products are zero because Hxxx
(3)
and Hyyy do not belong to the Hermitian basis of Hq=9 . So, τxx and τyy are
3


retrieved with errors O M a .
Therefore, the D2Q9 LBGK is a second order approximation to the con-
tinuous BGK equation and will retrieve the momentum balance equation with


third order errors O M a3 in both the two versions of the Hermitian basis,
Hq=6 and Hq=9 , restricting this LBE to be used for isothermal, nearly incom-
pressible ows. This order of approximation is given in Figure 7.3.a and b, by
the last row with a full set of Hermite polynomial tensors.
Nevertheless, in the rst version, all second order non-equilibrium moments
are ghost, in the sense that they cannot be represented in terms of the Her-
mitian basis of the subspace Hq=6 , while, in the second version, the viscous
stress tensor τxy can be represented in terms of the spatial derivatives of the
(3) (3)
equilibrium moments aeq,xxy and aeq,xyy .
It is, here, important to note that the above reasoning leads to the con-


clusion that contrary to what is usually believed, the O M a3 errors that
characterize the D2Q9 LBE and, also, the D3Q27 are not due to the dis-
cretization process itself, but to a poor second order Hermitian representation.
Indeed, this conclusion reects the dependence of the discrete velocity set on the
order of approximation of the Hermitian representation (and not the inverse ).
Better representations can be built without changing the set of discrete
velocities and the LBE parameters as and Wi . So, more accurate and stable
eq
LBE can be achieved by only replacing the equilibrium distribution fi , Eq.
(7.13), by its expansion in terms of the full set of 9 Hermitian polynomial
tensors shown in Figure 7.3.b.
The question that remains to be answered is whether the inclusion of these
third-order Hermite polynomials has any eect on the LBE stability.
The D2Q9 stability was analyzed by comparing the D2Q9 LBGK with the
Hq=6 Hermitian representation, Eq. (7.16), and a Hermitian representation
(3) (3)
where the third-order polynomials Hxxy and Hxyy were included into the
equilibrium distribution, Siebert et al. [85]. These models were also compared
with the Lallemand and Luo [56], multiple relaxation time MRT model, since
7.3. INCREASING THE ORDER OF APPROXIMATION 141

Figure 7.4: The eect on stability of adding third order Hermite polynomials to
the second order equilibrium distribution in a D2Q9 LBGK [85]. Comparison
with the MRT model of Ref. [56]

this model was also built with the aim of improving the LBE stability.

The results presented in Siebert et al. were based on a von Neumann

stability analysis for a two-dimensional isothermal problem described by the
LBGK model. In this kind of analysis, the aim is to obtain the response of
a system, described by a set of equations, which is slightly removed from a
given equilibrium state by a small perturbation. When this perturbation is
not absorbed by the system itself, such an equilibrium state is considered to
be unstable. This is performed using spatial wave perturbations, the eects
of which can be superposed, when this physical system is described by linear
equations.

The present analysis can be found in Figure 7.4 and is focused on values
1
of the relaxation time very close to its singular limit
2 . The abscissa was
∗
chosen as 1/τ in order to compare with previous results from Ref. [56]. It can
be observed that both the second and the third-order LBGK models present a
homogeneous decrease in the local speed stability limit when the relaxation time
approaches 1/2, whereas this limit remains insensitive to the τ∗ variation in the
∗
MRT model, up to τ =0.50251. Although the results presented by Lallemand
and Luo are not related to quasi-incompressible models, as the present LBGK
ones are, these authors attributed the better performance of the MRT model
with respect to the second-order LBGK to the use of high frequency relaxation
terms in modeling the collision term.
142 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

Figure 7.5: Results of a von-Neumann stabilty analysis showing the eect of

increasing the order of the LBE for enlarging the stability ranges. Parameter
ν∗ is the dimensionless kinematic viscosity and u∗max is the maximum local
velocity beyond which the numerical scheme becomes unstable.

Figure 7.4 shows, nevertheless, that the third-order LBGK model has a
considerably better performance when compared with the second-order one
and with the MRT model in what concerns its stability limits. So, the addition
of third-order velocity polynomials largely improve the stability range and this
improvement is due to the equilibrium distribution representation itself and
not to the use of extra relaxation terms in the collision model. This is an
important conclusion, since it avoids the use of MRT dispersion relations for
the adjustable parametersrelated to the short wavelength nonhydrodynamic
momentsto increase numerical stability.

Further details can be found in ref. [85].

7.3.2 Increasing the order of the lattice-Boltzmann equa-

tion
In fact, the improvement of the stability limits by increasing the order of
the polynomial approximation to the full MB equilibrium distribution in LBGK
models has shown to be a general result. This can be seen in Figure 7.5, where
the second order LBGK D2Q9 model is compared with the full third order
D2V17 and with the full fourth-order D2V37 models, derived by Philippi et al.
[75] .

This alternative was discussed by Siebert et al. [85].

In the following, we discuss a third alternative in relationship with the

concept of kinetic projection.
7.4. MOMENTS-BASED METHODS 143

7.4 Moments-based methods

After propagation, all the populations fi are known on every lattice node
x and time t. Therefore all the Hermitian rst moments, or the Hermitian
moments that are needed for the simulation, can be computed by making the
inner product of

X a(θ) (x, t)
f i = Wi H (θ) (as ei ) , (7.18)
nx !ny !nz !
θ

by H (η) (as ei ), giving,

X
f i H (η) (as ei ) = a(η) .
i

because

X H (θ) (as ei ) H (η) (as ei )

Wi = δθη .
i
nx !ny !nz !

For isothermal second-order models, the Hermitian moments that are needed
are

X
a(0) = f i = ρ∗ ,
i
X
a(1)
α = f i as eiα = as ρ∗ u∗α ,
i
(2)
X
f i a2s eiα eiβ − δαβ = a2s ρ∗ u∗α u∗β + ταβ
∗



aαβ = .
i

Therefore, in a moments-based method the propagated populations are for-

gotten and the pre-collisional populations are reconstructed based on the in-
formation of their rst order moments. For a second order isothermal model
we would thus rewrite the pre-collisional populations as

 
1 (2) 2
fbi = Wi a(0) + a(1)


α a e
s iα + a a e e
iα iβ − δ αβ ,
2 αβ s
or

 
1 1
fbi = Wi ρ∗ + 2 ρ∗ u∗α eiα + 4 ρ∗ u∗α u∗β + ταβ
∗
eiα eiβ − c2s δαβ .



(7.19)
cs 2cs

Although fbi does not alter the mass and momentum of site x, satisfying
144 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

X Xb
fi = f i = ρ∗ ,
i i
X Xb
f i eiα = f i eiα = ρ∗ u∗α ,
i i
X Xb
f i eiα eiβ = f i eiα eiβ = P ∗ δαβ + ρ∗ u∗α u∗β + ταβ
∗
.
i i

these reconstructed populations dier from the populations fi that were origi-
nally propagated to site x. Some authors attribute this dierence to the noise
produced by numerical errors due to non-controlled higher-order moments. In

this case, the reconstruction of fbi based on its rst momentsthe ones that t
into the LBE Hermitian representation may be seen as a regularization proce-
dure, performed for each site and at each time step [58]. So, when we rewrite
 
the populations as fbi ρ∗ , u∗α , ταβ
∗
as in Eq. (7.19), we are implicitly requiring

(θ)
that the inner products f ∗ H = 0 for every θ > 2. In other words, we are
b
i

requiring that the projection of populations fbi on a subspace of order higher

than 2 is null.
In another point of view, we can argue that the only local information
that we can use for site x is the one related to the moments ρ∗ , u∗α , ταβ
∗
, the

populations fi originally propagated to this site loosing their roles in favor

of these moments. Therefore, as in d'Humières moments method ([23]), the
collision step proceeds by modifying the non-equilibrium moments and the post-
collisional populations are then calculated in terms of these modied moments.
Both point of views lead to the same results.
The collision step proceeds as

 
out 1 eq 1 b
fi (x, t + δ) = f (x, t) + 1 − f i (x, t) . (7.20)
τ∗ i τ∗
eq
where fi is given by

 
eq 1 1
f i = Wi ρ∗ + 2 ρ∗ u∗α eiα + 4 ρ∗ u∗α u∗β eiα eiα − c2s δαβ


cs 2cs
The main idea behind this method is to lter the high-order non-equilibrium
moments that cannot be represented in terms of the Hermitian basis of the
Hilbert subspace Hq and so, are sources of errors aecting both the accu-
racy and the stability of the method. This idea was investigated by Latt and
Chopard [58] for the second-order D2Q9 LBE, who observed for this method,
an stability range 7.7 times the stability range of the BGK numerical scheme,
in the simulation of lid driven cavity ows.
It is, perhaps, important to emphasize that the distribution given by Eq.
(7.19) is the projection of fi on the subspace Hq=6 and that all the calculations
are based on this projection at each time step. So all the spurious information
7.4. MOMENTS-BASED METHODS 145

Figure 7.6: an Gogh clouds showing the Kelvin-Helmholtz instability.

that is carried out by the propagated populations is ltered at the pre-collisional

step.

Mattila and co-workers [66] investigated the properties of this method, by

setting up a perturbed double periodic shear layer ow as a benchmark case.
In this kind of problem, a small velocity perturbation, perpendicular to the
shear ow direction, initiates a Kelvin-Helmholtz instability and causes roll-up
of the antiparallel shear layers that is dissipated along the time. The method
was compared with the BGK stencil and with the entropic stabilizer method
of Karlin et al. [50]. Comparison details can be found in Ref. [66].

7.4.1 A sample case: Double periodic shear layer ow

Clouds showing a Kelvin-Helmholtz instability are formed when two layers
of air  one fast, the other slow  collided at just the right speed to create rises
and dips that caused the clouds to curl in on themselves and crash, just like
waves on a beach (Figure 7.6). An example of a numerical simulation showing
the benecial eects of the regularization for improving stability is found in
Mattila et al. [66], where a perturbed double periodic shear layer ow was set
up as a benchmark case. There, a small velocity perturbation, perpendicular to
the shear ow direction, initiates a Kelvin-Helmholtz instability and causes roll
up of the antiparallel shear layers (Figure 7.7). Periodic boundary conditions
were used.

The initial conditions are shown in Figure 7.8 and are given by
146 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

Figure 7.7: Kelvin-Helmholtz instability in shear ow. A small velocity pertur-

bation, perpendicular to the shear ow direction, initiates a Kelvin-Helmholtz
instability and causes roll up of the shear layers.

y
 ∗  1

 L
 uo tanh λ L − 4 for |y| ≤ 4
u∗x = ,
y
u∗o tanh λ 3 L


− for |y| >

4 L 4

and

  
x 1
u∗y = εu∗o sin 2π + .
L 4
In the above equationsu∗o is the velocity of the shear layer. The prole
is controlled by parameter λ, which was considered in present sample case as
λ = 80. With this value of λ, the velocity prole, although continuous, is nearly
constant in the central layer and has a step gradient in the contact lines with
the lateral layers, producing a shear ow. An initial sinusoidal velocity pertur-
bation u∗y is imposed in the vertical direction and controlled by parameter ε.
Parameter ε = 0.05 in present simulations. Boundary conditions are periodic.
The simulation starts from these imposed velocity proles and follows the ve-
locity eld till the end of the ow, when all the kinetic energy is dissipated by
viscous forces.
The Mach number, M = u∗ /cs was kept very small (M = 5.4127 × 10−2
) and the simulations were performed with the second-order D2Q9 LBE, in
despite of its errors. At each time step the average kinetic energy is measured
2
and normalized with respect to ρu∗o /2. Fig. 7.9 shows the time decay of the
average kinetic energy comparing the simulation results between the LBGK
model, given by Eq. (7.5) with the moments-based LBE, Eq. (7.19) , Mattila
et al. [66]. The initial Reynolds number is Re = 32000, the numerical domain
has a length L = 128 lattice unities, and velocity u∗0 =1/32. The dimensionless
∗
kinematic viscosity is very low ν = 4/30000. The plotted average values of
∗2
the kinetic energy are normalized with respect to ρuo /2. It is clearly seen that
∗
the LBGK scheme becomes unstable very early, around time t = 0.3, corre-
sponding to the rst 1230 time steps, whreas the moments-based LB scheme
remain stable during the whole simulation.
7.4. MOMENTS-BASED METHODS 147

Figure 7.8: Initial velocity proles used for the simulation of the double periodic
shear ow.

Figure 7.9: Decay of average kinetic energy for an initial Reynolds number,
Re = 32000, L = 128 lattice unities, u∗0 =1/32 and ν ∗ = 4/30000. The average
∗2
values are normalized with respect to ρuo /2. The LBGK scheme becomes
unstable early on, around time t∗ = 0.3, corresponding to 1230 time steps.
The moments-based LB scheme remain stable during the whole simulation,
[66]
148 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

Figure 7.10: Vorticity eld for t∗ = 1 showing the simulation results for the
LBGK model (left) and the moments-based method (right). It is clearly seen
that the LBGK model is plagued by spurious secondary vortex even for Re =
32000 whereas the moments-based scheme produces accurate and stable results
up to Re = 80000 [66]

Being a global measure, the average kinetic energy, does not capture well
the dierences between local ow features. Fig. 7.10 shows the simulated
t∗ = 1, comparing the LBGK and the moments-based scheme
vorticity eld, at
at this time step. Now Re = 80000 for a simulation domain with L = 256
∗
and a dimensionless velocity u0 =1/16. The dimensionless kinematic viscosity
∗ −4
is ν = 2 × 10 . The LBKG case was simulated with Re = 32000 since
beyond this limit the simulation was unstable. Simulation with LBGK shows
small spurious secondary vortexes. The moments-based LB scheme remained
stable during the whole simulation and was shown to be accurate without the
appearance of unphysical secondary vortex.

7.5 Toward high-order moments-based LBE: re-

currence relations
For second order models like the D2Q9 and D3Q27, both the BGK and
the moments-based methods given by Eq. (7.19) suer from errors O (M a3 )
and are restricted to isothermal and low Mach number ows. In the follow-
ing sections high-order moments-based schemes are proposed for avoiding this
restriction and suitable for non-isothermal problems.
Non-equilibrium moments have an advective component, which is null when
the local velocity is null and a diusive component related to the peculiar
microscopic velocity. For concreteness, consider, for instance, a third order
non-equilibrium moment
7.5. TOWARD HIGH-ORDER MOMENTS-BASED LBE: RECURRENCE RELATION

∗
X neq
παβγ = fi eiα eiβ eiγ . (7.21)
i

The velocity function eiα eiβ eiγ can be written as

eiα eiβ eiγ

   
(eiα − u∗α ) eiβ − u∗β eiγ − u∗γ

 + eiγ − u∗γ u∗α u∗β + (eiα − u∗α ) u∗β u∗γ




   
+ eiβ − u∗β u∗α u∗γ
 
 
   
= + eiβ − u∗β eiγ − u∗γ u∗α
, (7.22)


 
+ (eiα − u∗α )eiγ − u∗γ u∗β


 
 
+ (eiα − u∗α ) eiβ − u∗β u∗γ
 
 
+u∗α u∗β u∗γ

and the above equation is a general relationship, independent of the lattice.

Replacing Eq. (7.22) into Eq. (7.21) it results

∗ ∗
παβγ =π
eαβγ + πadv,αβγ ,
∗
where π
eαβγ is the diusive component

∗
X neq
(eiα − u∗α ) eiβ − u∗β eiγ − u∗γ ,



π
eαβγ = fi
i

and the advective component is given in terms of second-order non-equlibrium

moments by

∗
πadv,αβγ = u∗α τβγ
∗
+ u∗β ταγ
∗
+ u∗γ ταβ
∗
.
A Chapman-Enskog analysis shows that, in isothermal problems, the vis-
cous stress tensor ταβ is the highest non-equilibrium moment that is needed for
closing the system of hydrodynamic equations at the Navier-Stokes level. In
isothermal problems, the appropriate truncation point for the Hermite polyno-
mial expansion is at N = 3, when the M a3 errors are, also, to be avoided. That
is, for moments up to third-order, both equilibrium and non-equilibrium part,
should be included into the expansion. However, as becomes evident from this
analysis, the diusive component of the third-order non-equilibrium moment
is dened in terms of fourth-order equilibrium moments which do not belong
∗
to the chosen Hermite basis. Hence this third-order diusive component π
eαβγ
must be considered as a spurious  or ghost  moment, and thus should be
ltered from the representation in isothermal hydrodynamics, resulting in the
following recurrence relation

∗
παβγ = u∗α τβγ
∗
+ u∗β ταγ
∗
+ u∗γ ταβ
∗
,
150 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

which expresses a third order non-equilibrium moment in terms of second-order

ones.
Recurrence relations will be now generalized for high-order Hermitian mo-
ments. The microscopic velocity ξ0 can be referred to the local velocity u0 ,
considering the peculiar velocity

c0 = ξ0 − u0

as the independent variable, instead of ξ0 .

The equilibrium distribution can be, thus, written in dimensionless form as

c2
0
eq ρ∗ e− 2(Θ+1)
f = D D , (7.23)
(Θ + 1) 2 π2

and its expansion in Hermite polynomials H

e = H (c0 ) is given by

c2
− 0 (θ)
∗e aeq
eq 2 X
f =ρ D
e e (θ) ,
H
π 2 nx !ny !nz !
θ

where the expansion coecients are given by

 c2
0
1 e− 2(Θ+1) e (θ)
a(θ)
eeq = D D H (c0 ) dc0 .
(2π) 2 (Θ + 1) 2
Since the Hermite polynomials are written in terms of the peculiar velocity
c0 and considering Eq. (7.23):

 c2
0
1 e− 2(Θ+1)
a(0)
eeq = D D 1dc0
(2π) 2 (Θ + 1) 2

1 c2
= D e− 2 dc = 1,
(2π) 2

p 
(Θ + 1) c2
a(1)
eeq,α = D e− 2 cα dc = 0,
(2π) 2

(2) 1 c2
aeq,αβ
e = D e− 2 (cα cβ Θ + cα cβ − δαβ ) dc =Θδαβ
(2π) 2

an so on.
Consider now the non-equilibrium distribution and its expansion in Hermite
polynomials e (θ) ,
H
c2
0 (θ)
neq e− 2 X aneq
e e (θ) ,
f = D H (7.24)
π 2 nx !ny !nz !
θ
7.5. TOWARD HIGH-ORDER MOMENTS-BASED LBE: RECURRENCE RELATION

(θ)
where parameters aneq
e are given by


neq
a(θ)
eneq = f e (θ) dc0 .
H

Hermite polynomials e (θ) (c0 ) can be written

H as a sum of n+1 terms of
(θ)
decreasingly order H (ξ0 ) in accordance with,

(n−1)
X
e α(n)...α (c0 ) = Hα(n)...α (ξ0 ) −
H u0,αθ Hα1 ...αn /αθ (ξ0 )
1 n 1 n
θ
(n−2)
X
+ u0,αθ u0,αη Hα1 ...αn /αθ αη (ξ0 )
θ,η
n
+ ... + (−1) u0,α1 ...u0,αn H (0) (ξ0 ) , (7.25)

where the notation α1 ...αn /αθ ...αη means that all the indices are considered
except the indices αθ ... αη . For each rth term in the sum, r = 1, ..., n + 1, the
n
above equation considers all the possible Cr−1 combinations of the n indices
{α1 , ..., αn } taken as sets of r − 1 elements. For instance, if n = 4, the third
4
term on the right hand side of Eq. (7.25) will be a sum of C2 = 6 terms,

(2) (2) (2)

u0,α u0,β Hγδ + u0,α u0,γ Hβδ + u0,α u0,δ Hβγ
(2) (2) (2)
+ u0,γ u0,δ Hαβ + u0,β u0,δ Hαγ + u0,β u0,γ Hαδ .

Eq. (7.25) gives rise to a general relationship for the non-equilibrium mo-
neq
ments. By multiplying Eq. (7.25) by f and integrating in the velocity space

(n−1)
X
a(n)
e (n)
neq,α1 ...αn = aneq,α1 ...αn − u0,αθ aneq,α1 ...αn /αθ
θ
(n−2)
X
+ u0,αθ u0,αη aneq,α1 ...αn /αθ αη
θ,η
n
+ ... + (−1) u0,α1 ...u0,αn a(0)
neq . (7.26)

So, for the rst non-equilibrium moments written in terms of LB variables

152 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

a(0) a(0)
neq = eneq = 0,

a(1) a(1)
neq,α = eneq,α = 0,
(2) (2) 1 ∗
aneq,αβ = e
aneq,αβ = τ ,
c2s αβ
 ∗ ∗
uα τβγ + u∗β ταγ
∗

(3) (3) 1
aneq,αβγ = e
aneq,αβγ + 3 , (7.27)
cs +u∗γ ταβ
∗

(4) (4)
aneq,αβγδ = e
aneq,αβγδ
!
(3) (3)
1 u∗α aneq,βγδ + u∗β aneq,αγδ
+ (3) (3)
cs +u∗γ aneq,αβδ + u∗δ aneq,αβγ
u∗α u∗β τγδ
∗ ∗
+ u∗α u∗γ τβδ
 
1
− 4  +u∗α u∗δ τβγ∗
+ u∗γ u∗δ ταβ
∗ .
cs ∗ ∗ ∗ ∗ ∗ ∗
+uβ uδ παγ + uβ uγ παδ

The continuum hypothesis means that the Knudsen number  given by

the ratio between the mean free path and a characteristic length L should
1
be very small, close to zero . As stated above, in isothermal problems, the
viscous stress tensor is the hightest non-equilibrium moment that is needed for
closing the system of hydrodynamic equations. All the diusive components
of non-equilibrium moments with order higher than 2 are to be considered as
ghost and ltered from the representation. So, isothermal hydrodynamics can
be described based on Hilbert subspaces generated by a third order Hermitian
basis.
The above considerations also mean that when the third-order non-equilibrium
(3)
diusive moments aneq,αβγ
e and higher are ltered from the representation, Eq.
(7.24), the following recurrence relations are obtained from Eq. (7.27)

(3) 1
u∗α τβγ
∗
+ u∗β ταγ
∗
+ u∗γ ταβ
∗


aneq,αβγ = 3
, (7.28)
cs
 ∗ ∗ ∗
u u τ + u∗α u∗γ τβδ ∗ 
(4) 1  α∗ β∗ γδ∗
aneq,αβγδ = 4 +uα uδ τβγ + u∗β u∗γ ταδ
∗ 
. (7.29)
cs ∗ ∗ ∗ ∗ ∗ ∗
+uβ uδ ταγ + uγ uδ ταβ

Eqs. (7.28) and (7.29) are recurrence relations for the third and fourth
order non-equilibrium moments in terms of the second order viscous stress
tensor. Malaspinas [63] found similar relations, but using a dierent approach
and not considering the role of the diusive component of these moments.

1 In the lattice Boltzmann framework, the Knudsen number is related to the ratio of
the lattice spacing h and a characteristic length L. It is a nite number and cannot be
considered as close to zero. So, errors due to high-order Knudsen terms are to be expected in
the hydrodynamic equations. The inuence of nite Knudsen numbers on the error formation
in isothermal hydrodynamics is, nevertheless, outside the scope of this paper.
7.6. MOMENTS-BASED LBE WITH RECURRENCE RELATIONS 153

These ndings open the route to improved Hermitian representations and to

increase the stability ranges of LB algorithms based on high order schemes for
the relaxation of the non-equilibrium moments, without increasing the number
of lattice vectors in the representation.

7.6 Moments-based LBE with recurrence rela-

tions
In this section some moments-based LBE suitable for isothermal hydrody-
namic and non-isothermal problems will be presented and discussed. The main
idea is to write these LBE as in Eq. (7.19)

eq
f i (x, t)
(θ)
X aeq
= Wi H (θ) (as ei ) , (7.30)
nx !ny !nz !
H (θ) ∈Hq

and

f i (x, t)
X a(θ)
= Wi H (θ) (as ei ) . (7.31)
nx !ny !nz !
H (θ) ∈Hq

Therefore, each derived lattice Boltzmann equation will be dependent on the

subspace Hq of Hermite polynomials that was chosen for the representation and
on the kind of problem it is wanted to be solved: isothermal or non-isothermal.
We restrict ourselves to 2D LBE. Three-dimensional LBE can be built using
the same reasoning.

7.6.1 Isothermal hydrodynamics

In this case Θ=0 and the highest order non-equilibrium moment to be
∗
retrieved is the viscous stress tensor ταβ .

The D2Q9 LBE

Figure 7.3.b shows the Hermitian basis of the subspace Hq=9 . This Her-
mitian basis with third and fourth order Hermite polynomial tensors will be
used for the D2Q9 representation. The third and fourth-order non-equilibrium
moments can be written in terms of second-order ones in accordance with Eqs.
(7.28-7.29)
154 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

∗ (2) ∗ (2)
a(3)
neq,xxy = uy aneq,xx + 2ux aneq,xy ,
∗ (2) ∗ (2)
a(3)
neq,xyy = 2uy aneq,xy + ux aneq,yy ,
 2 
∗ (2) ∗ ∗ (2) ∗2 (2)
a(4)
neq,yyy = ux aneq,yy + 2ux uy aneq,xy + uy aneq,yy .

So, a moments-based discrete form of the BGK equation can be written as

in Eq. (7.19) where

1 ∗ 1 ∗ ∗
 2


1+ c2s uα eiα + 2c4s uα uβ eiα eiα − cs δαβ
27 ∗2 ∗ 1
2


+ 2 ux uy eix − 3 eiy
 
eq
f i = W i ρ∗  , (7.32)
 
∗2 ∗
+ 27 1
2


2 uy ux eiy −
3 eix
 
∗2 ∗2
+ 81 e2ix − 13 e2iy − 31


4 ux uy

and

fˆ i = f i
eq

 9 ∗ 1


2 ταβ eiα eiβ −
3 δ2αβ
∗ ∗ ∗ ∗
+ 92 1



 uy τxx + 2ux τxy
eix − 3
eiy


 + 92 u∗x τyy
∗
+ 2u∗y τxy
∗
e2iy − 1
3 eix

+ Wi  (7.33)
 
∗2 ∗
 
ux τyy +



2u∗x u∗y τxy∗ 
 27 
e2ix − 13 1

2
 +4 eiy − 3 

2
+u∗y τxx ∗

Eqs. (7.32) and (7.33) are to be compared with the standard moments-
based D2Q9 LBE. This new form of the D2Q9 moments-based LBE, also re-


trieves the isothermal Navier-Stokes equations with errors O M a3 and, as
commented above, is also restricted to very low Mach number ows, in the in-
compressible limit. The addition of third and fourth order ltered terms in the
non-equilibrium distribution has the purpose of enhancing its stability limits,
for high Reynolds number ows, without the interference of third and higher
order diusive ghost moments.

The D2V17 LBE

Figure 7.11 shows the Hermitian basis of the subspace Hq=10 and its cor-
responding D2V17 set of discrete velocities [75]. The norm preservation con-
ditions, Eq. (7.12) are satised by the scaling factor as and weigths Wi given
in Table 1 of [84]. This Hermitian basis is a full third order approximation
of the continuous BGK equation and, so, all the second order non-equilibrium
moments τxx , τyy and τxy are retrieved without errors up to the rst order in
the Knudsen number. In the same way the momentum balance equation.
The third-order non-equilibrium moments can be written in terms of second-
order ones in accordance with Eq. (7.28)
7.6. MOMENTS-BASED LBE WITH RECURRENCE RELATIONS 155

Figure 7.11: Third-order Hermitian basis of the subspace Hq=10 and its related
D2V17 set of lattice vectors.

∗ (2)
a(3)
neq,xxx = 3ux aneq,xx ,
∗ (2) ∗ (2)
a(3)
neq,xxy = uy aneq,xx + 2ux aneq,xy , (7.34)

a(3)
neq,xyy = 2u∗y a(2)
neq,xy + u∗x a(2)
neq,yy ,

a(3)
neq,yyy = 3u∗y a(2)
neq,yy .

Considering that all the second-order non-equilibrium moments can be re-

trieved up to order 1 in the Knudsen number, a moments-based discrete form
of the BGK equation can be written as

eq
fi
1 + c12 u∗α eiα
 
s
+ 1 u∗ u∗ e e − c2s δαβ


= Wi ρ∗   2c4s α β iα iβ
 
u∗ ∗ ∗
eiα eiβ eiγ ,
 
α uβ uγ

+ 6c6s −c2s (δβγ eiα + δαγ eiβ + δαβ eiγ )

fˆ i = f i
eq

 ∗ 
ταβ

e e − c2s δαβ
 2c4s iα iβ  
+ Wi 
 + 6c16 u∗α τβγ∗
+ u∗β ταγ
∗
+ u∗γ ταβ
∗ .

 s 
× eiα eiβ eiγ − c2s (δβγ eiα + δαγ eiβ + δαβ eiγ )

It is not possible to improve the Hermitian representation of the D2V17

LBE. In fact, contrary to the D2Q9 LBE, which Hermitian basis can be im-
th
proved by the addition of third and fourth order Hermite polynomials, 4 and
higher order Hermite polynomials are not orthogonal to the Hermitian basis
presented in Figure 7.11.
156 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS

7.6.2 Non-Isothermal hydrodynamics

(3)
In non-isothermal hydrodynamic problems, the diusive moment aneq,αβγ
e
has a physical meaning and is related to the dimensionless heat ow vector. In
fact, the molecular difusive ux of the peculiar kinetic energy is given by

1 2
qα = f neq m (ξβ − uβ ) (ξα − uα ) dξ ,
2
or

3 neq 1 2
qα = ρ0 ξ f (ξ0,β − u0,β ) (ξ0,α − u0,α ) dξ0 .
2
On the other hand, from Eq. (7.3) the peculiar Hermite polynomial e (3)
Hαββ
is written as

e (3) = (ξ0,α − u0,α ) (ξ0,β − u0,β )2

H αββ
 
(ξ0,α − u0,α )
− .
+2δαβ (ξ0,β − u0,β )

and since ρ and ρu are equilibrium moments


31 neq (3)
qα = ρ0 ξ f H e
αββ (c0 ) dc0
2
3 1 (3)
= ρ0 ξ e a .
2 neq,αββ
2
Remembering that ξ = c2 c2s , in its dimensionless form the heat ux qα∗

(3)
= qα / ρ0 c3 can be aneq,αββ in its discrete form by,
related to e

(3) 2qα∗
aneq,αββ =
c3s
e
1 X neq
2
= 3 f eiβ − u∗β (eiα − u∗α ) . (7.35)
cs i i

(3)
Therefore, from Eq. (7.27) the sum aneq,αββ is related to the heat ux qα∗
by

(3) qα∗ 1
+ 3 u∗α τββ
∗
+ 2u∗β ταβ
∗


aneq,αββ = 2 3
.
cs cs
Third-order diusive moments require fourth-order Hermite representations.
th
Figure 7.12 shows a 4 -order Hermitian representation and its corresponding
D2V37 LBE.
(4)
The difusive moments aneq,αβγ
e are to be ltered from the representation
(4)
and each moment aneq,αβγδ is calculated as in Eq. (7.27), in terms of the
7.6. MOMENTS-BASED LBE WITH RECURRENCE RELATIONS 157

Figure 7.12: Fourth-order Hermitian representation (a) and its correspondng

D2V37 set of lattice vectors (b).

(3) (3)
aneq,αβγ . The peculiar moment aneq,αβγ
e is calculated at each site and time
step in accordance with

(3) 1 X neq
f i (eiα − u∗α ) eiβ − u∗β eiγ − u∗γ ,



aneq,αβγ =
e 3
cs i
related to the heat ow vector. In two dimensions and in accordance with Eq.
 
(3) (3)
(7.35) the sum c3s eaneq,αxx + e
aneq,αyy is twice the heat ux component along

the direction α, or 2qα∗ .

158 CHAPTER 7. BEYOND BGK: MOMENTS-BASED METHODS
Chapter 8

Boundary conditions: part II

Fluid ows are ubiquitous in nature and technology. Assuring a correct
description of ows is, then, of utmost importance. Solving analytically the
Navier-Stokes equations has been a challenge for almost two centuries, since
only few analytical solutions, generally for very simple ows and geometries,
are known. The alternative to the formal methodology is to resort to numer-
ical methods, usually discretizing, in several dierent ways, the Navier-Stokes
equations.

In this regard, the lattice Boltzmann equation (LBE) has been successfully
used in the last decades as a computational uid dynamics tool, solving a
discrete analogue of the Boltzmann equation for the mass, momentum and
energy balance equations ([52], [86], [35]) . The LBE has proved to be valuable
in solving simple and complex ows.

One of the key issues in a numerical method is the application of boundary

conditions. While solving the core ow itself is the main goal of every numerical
method, the way boundary conditions - either Dirichlet, Neumann or Robin
 are specied is usually considered a less important problem. Nevertheless,
it has been shown [42] that a careful approach for the uid-solid interaction
might lead to a signicant improvement of the description of the ow and the
numerical stability of the method itself.

A lot of eort has been devoted to the boundary conditions for the lattice
Boltzmann framework, but the most recent thorough comparison of dierent
models for boundary conditions was published a decade ago [58], [62]. In this
work, we focus the extension of moments-based populations to boundary con-
ditions (BC's). Schemes based on moments-based method generally leads to
improved general stability of LB equations, but, as we will see in this chapter,
there is no single solution to solve the boundary condition problem.

In this chapter, only Dirichlet boundary conditions will be considered and

we restrict ourselves to second order 2D LBE suitable for the simulation of
isothermal, incompressible ows.

159
160 CHAPTER 8. BOUNDARY CONDITIONS: PART II

Figure 8.1: A boundary node.

8.1 Mass conservation

Consider a uid node located in a uid node (xb , yb ) that is adjacent to a
solid surface in Figure 8.1 and suppose the populations shown in solid lines are
the populations that were propagated to from neighbor uid sites. Populations
shown in dot lines, are unknown.
Adherence conditions require that at the boundary node (xb , yb ),
X
f¯i eix = ρ∗ u∗x ,
i

X
f¯i eiy = ρ∗ u∗y .
i

where u∗x and u∗y are the dimensionless velocity components for x = xb , that
are supposed to be known. A third equation can be written as

X
f¯i = ρ∗ , (8.1)
i

giving the mass density ρ∗ at the boundary node, which is here to be reinter-
preted as the pressure since P ∗ = ρ∗ c2s in the LB incompressible limit. For the
node depicted in Figure 8.1, we have thus 3 equations for the unknowns ρ∗ , f¯4
, f¯7 , f¯8 . This is the main diculty for establishing local boundary conditions
in LB framework since the number of unknowns is greater than the number
of equations at our disposal. This is especially important when we deal with
high-order LB equations, with a great number of discrete lattice vectors [75],
[84], [82].
In addition, although the collision process does not aect the mass density
at the boundary node and we can always assure that

X X
f¯i = f¯iout , (8.2)
i i
8.1. MASS CONSERVATION 161

there is no guarantee that that there is no mass-leakage, i.e., that the amount
of particles that reach the boundary node from the uid phase is the same
amount that leaves this node back to the uid phase, after collision.
To be clear, let us consider the set

I = {0, 1, 2, 3, 5, 6} ,
in Figure 8.1, as the set of incoming directions from the uid phase to the solid
surface and that

O = {0, 1, 3, 4, 7, 8} ,
is the set of outgoing directions from the solid to the uid sites. For avoiding
loss or gain of particles to or from the external environment, instead of Eq.
(8.2) the mass preservation condition requires that the total amount of particles
incoming from the uid phase before collision be the same as the total amount
of particles leaving the solid after collision. In other words,

X X
f¯i = f¯iout . (8.3)
i∈I i∈O

Indeed, Eq. (8.2) means that

X X X X
f¯i + f¯i = f¯iout + f¯iout , (8.4)
i∈I i∈I
/ i∈O i∈O
/

and, under the light of Eq. (8.3), we have

X X
f¯i = f¯iout . (8.5)
i∈I
/ i∈O
/
For the boundary node shown in Figure 8.1, the above equation reduces to

f¯4 + f¯7 + f¯8 = f¯2out + f¯5out + f¯6out , (8.6)

meaning that the amount of particles entering the uid before the collision step
must compensate the amount of particles leaving the uid after the collision
step. This is necessary for avoiding gain or loss of particles by the uid phase.
It should be remarked that the bounce-back boundary condition shown in
Section ........... for solid surfaces at rest,

f −i (xb , t + δ) = f i (xb , t) ,
also avoids the mass leaking, by making

f¯4 (xb , t + δ) = f¯2 (xb , t) , f¯7 (xb , t + δ) = f¯5 (xb , t) , f¯8 (xb , t + δ) = f¯6 (xb , t) .
Nevertheless, the above condition must not be confused with the mass
preservation condition given by Eq. (8.6). While Eq. (8.6) gives post-collisional
populations that are dependent on the relaxation time, BB condition gives post-
collisional populations f¯i (xb , t + δ) independent on it.
162 CHAPTER 8. BOUNDARY CONDITIONS: PART II

8.2 Boundary nodes on plane surfaces

By considering a second-order Hermitian representation based on the Her-
(1) (1) (2) (2) (2)
mite polynomials H (0) , Hx , Hy , Hxx , Hxy , Hyy , the following set of
equations are to be satised for every boundary site

X X
f¯i = f¯iout , (8.7)
i∈I i∈O
X
f¯i eix = ρ∗ u∗x , (8.8)
i
X
f¯i eiy = ρ∗ u∗y , (8.9)
i
X 2
f¯i eix eix = ρ∗ u∗x + ρ∗ c2s + τxx
∗
, (8.10)
i
X
f¯i eix eiy = ρ∗ u∗x u∗y + τxy
∗
, (8.11)
i
X 2
f¯i eiy eiy = ρ∗ u∗y + ρ∗ c2s + τyy
∗
, (8.12)
i

where, in most models, the mass conservation condition is replaced by Eq.

(8.1). Since we are considering that the boundary node is moving with a known
velocity whose components are u∗x and u∗y , we have thus an overdetermined
system of 6 equations for 4 unknowns in a moments-based method. They are
ρ∗ , τxx
∗ ∗ ∗
, τxy , τyy , because each unknown population can be written in terms of
these unknowns in accordance with

fˆ i = f i = Wi
h  
i
ρ∗ + 1 ∗ ∗
c2s ρ uα eiα + 1
2c4s ρ∗ u∗α u∗β + ταβ
∗
eiα eiβ − c2s δαβ .

Therefore, the system of equations (8.7)-(8.12) does not have a single so-
lution and dierent models were proposed for solving this boundary condition
problem

8.2.1 Latt et al. model

Latt et al. (2008) base their approach on the concept of density, Eq.(8.1),
and on the bounce-back of the non-equilibrium parts of all unknown popula-
tions, similarly as in Zou and He [93] model.
Therefore, considering the boundary node shown in Figure 8.1

fˆ
¯neq τ ∗ , τ ∗ , τ ∗ = f¯2 − f¯eq (ρ∗ ) ,


4 xx xy yy 2

fˆ
¯neq τ ∗ , τ ∗ , τ ∗ = f¯5 − f¯eq (ρ∗ ) ,


7 xx xy yy 5
8.2. BOUNDARY NODES ON PLANE SURFACES 163

fˆ
¯neq τ ∗ , τ ∗ , τ ∗ = f¯6 − f¯eq (ρ∗ ) ,


8 xx xy yy 6
∗ ∗ ∗
and the following expressions are obtained for τxx , τxy , τyy ,when cs is replaced
√
by 1/ 3

∗
= 4 f¯5 + f¯6 − 2f¯2 − ρ∗ u∗x ,


τxx

 
1
∗
= 2 f¯5 − f¯6 − ∗
+ uy ρ∗ u∗x ,


τxy
3
1
∗
= 2 f¯2 + f¯5 + f¯6 − 1 + u∗y ρ∗ u∗y + ,



τyy
3
∗
where the density ρ can be evaluated as

1  ¯
ρ∗ = f0 + f¯1 + f¯3 + 2 f¯2 + f¯5 + f¯6 .



(8.13)
u∗y +1
as in Zou & He [93] model.
The unknown populations can be then calculated as

¯4 = f¯2 − 2 ρ∗ u∗ ,
fˆ y
3

¯7 = f¯5 − 1 ρ∗ u∗ + u∗ ,
fˆ


x y
6

¯8 = f¯6 + 1 ρ∗ u∗ − u∗ ,
fˆ


x y
6
The above equations give the following adherence conditions

X
f¯i eiy = ρ∗ u∗y , (8.14)
i
X 1
f¯i eix = ρ∗ u∗x , (8.15)
i
3

which can only be satised for the x-component when the known populations f¯1
and f¯3 are replaced by their moments-based forms fˆ¯1 and fˆ¯3 , whose dierence
gives

fˆ ¯3 = 2 ρ∗ u∗ .
¯1 − fˆ
x
3
Thus the main particularity of Latt et al. model is the replacement of the
known populations by reconstructed ones, based on the values of the viscous
stress tensor components. Nevertheless, Latt et al. model does not satisfy
Eq.(8.3) and, so, it is unable to prevent the leaking of particles across the
boundaries.
164 CHAPTER 8. BOUNDARY CONDITIONS: PART II

Exercise
Demonstrate Eqs. (8.14) and (8.15).

8.2.2 Malaspinas et al. model

Malaspinas et al. [64] base their approach on the solution of a system of
equations composed by

eq
f i = f i (ρ∗ ) + fˆ
¯neq τ ∗ , τ ∗ , τ ∗ , i ∈ I,


i xx xy yy

X
ρ∗I = f i = ρ∗I ρ∗ , τxx
∗ ∗ ∗


, τxy , τyy
i∈I

Therefore, for the boundary node depicted in Figure 8.1, the above equa-
tions give a system of 7 equations and 4 unknowns, because fi is know for each
i ∈ I. A minimization procedure is used for solving this overdetermined system
of equations.

8.2.3 Mohammed and Reis model

Mohammed and Reis [71] use a MRT approach for solving the LBE and, so,
(3)
based on a full Hermitian basis with 9 Hermite polynomials, polynomials Hxyy ,
(3) (4)
Hxxy , Hxxyy being added to the representation (see Figure .............................).
Solution is based on the system of equations for the rst moments, Eqs.
(8.7)-(8.12). For boundary nodes adjacent to plane surfaces such as the one
shown in Figure 8.1, this system was reduced by Mohammed and Reis (2017),
choosing the following set of equations

X
f¯i eix = ρ∗ u∗x , (8.16)
i
X
f¯i eiy = ρ∗ u∗y , (8.17)
i
X 2
f¯i eix eix = ρ∗ u∗x + ρ∗ c2s , (8.18)
i
∗
and requiring that τxx = 0, because u∗x is constant along direction x.

8.2.4 Hegele et al. model

Independently of the number of lattice vectors in the representation, when
the velocity of the boundary node is known, the only unknowns are the density
and the three components of the viscous stress tensor in this kind of problem.
In fact, each population f¯i may be written in terms of its moments ρ∗ , τxx
∗ ∗
, τxy ,
∗
τyy , in accordance with
8.2. BOUNDARY NODES ON PLANE SURFACES 165

 
ˆ
¯ ∗ 1 ∗ ∗ 1 ∗ ∗ ∗ ∗

2


fi = Wi ρ + 2 ρ uα eiα + 4 ρ uα uβ + ταβ eiα eiβ − cs δαβ .
cs 2cs
Since mass must be conserved and node (xb , yb ) is adjacent to a solid sur-
face, the sum of the moments-based populations of particles outgoing from the
boundary node to the uid sites, after the collision step, must be identical to
the sum of the populations incoming from uid sites before the collision step,
or

fˆ
X X
f¯i = ¯out .
i (8.19)
i∈I i∈O

So, the physics underlying Eq. (8.19) and Eq. (8.3) is the same. Never-
theless, in Eq. (8.19) all the outgoing post-collisional populations are kinetic
projections dependent on the dimensionless quantities ρ∗ , τxx
∗ ∗ ∗
, τxy , τyy in ac-
cordance with

¯out (x, t + δ) = ωf eq (ρ∗ ) + (1 − ω) fˆ¯i ρ∗ , τ ∗ , τ ∗ , τ ∗ , ,

fˆ


i i xx xy yy

where ω = 1/τ ∗ is a collision frequency.

Recall that the l.h.s. of Eq. (8.19) is a known positive real number and
that, after this replacement, the r.h.s of this equation has a linear dependence
on ρ∗ , τxx
∗ ∗ ∗
, τxy , τyy .
The remaining equations are, for the rst-order moments,

fˆ
X X
f¯i eiα + ¯i eiα = ρ∗ u∗ ,
α (8.20)
i∈I i∈I
/

and, for the second-order moments,

X
Xˆ
f¯i eiα eiβ − c2s δαβ + f¯i eiα eiβ − c2s δαβ = ρ∗ u∗α u∗β + ταβ
∗


. (8.21)
i∈I i∈I
/

As before, there are dierent methods for solving the overdetermined system
of Eqs. (8.19)-(8.21) for the boundary node on plane walls such as the one
depicted in Figure 8.1, leading to dierent solutions. We present the results of
two of these methods presented in Bazarin et al. [ ? ]

Solution based on mass conservation

In contrast with previous works Bazarin et al. [ ? ] base their approach on
the conservation of mass and on the regularization of second-order moments
searching for the solution of Eqs. (8.19), (8.21). For the node depicted in Figure
8.1 the mass conservation equation, Eq, (8.19) gives a relationship between the
dimensionless mass density and the normal component of the viscous stress

 
∗ 5 1 ∗ ∗

1 ∗
ρ − u 1 + uy + (ω − 1) τyy = ρI . (8.22)
6 2 y 2
166 CHAPTER 8. BOUNDARY CONDITIONS: PART II

∗
On the other hand, Eq. (8.21), for τ22 , results in the following relationship

2 ∗ 1  2 
τyy + ρ∗ 6u∗y + 3u∗y − 1 = (myy )I , (8.23)
3 9
ρI = iI fi and (ρmαβ )I = i f¯iI eiα eiβ − c2s δαβ
P ¯ P

where are known quan-
tities, since they are only dependent on the incoming populations that reachs
the boundary node, coming from the uid phase.
Eqs. (8.22) and (8.23) can be solved for the density ρ∗ and the normal
∗
component τyy of the viscous stress, giving

4ρI + 3 (myy )I (1 − ω)
ρ∗ = 3 ,
ω + 9 − 3uy (1 + ω) − 6ωu2y

∗
2ρI + 15 (myy )I − 3 2ρI + 3 (myy )I uy − 3 4ρI + 3 (myy )I u2y
τyy = .
ω + 9 − 3uy (1 + ω) − 6ωu2y

The remaing components of the viscous stress can be then be found as

∗ 6 4ρI − 3 (myy )I (ω − 1)
τxx = (mxx )I − 3u2x ,
5 ω − 3uy (ω + 1) − 6ωu2y + 9


∗ ux 4ρI − 3 (myy )I (ω − 1) (3uy + 1)
τxy = 2 (mxy )I − .
ω − 3uy (ω + 1) − 6ωu2y + 9
As expected all the moments become dependent on the relaxation frequency
ω.

Solution based on the density

For comparison purposes, a solution method can be also implemented for
analyzing the eect of the mass conservation boundary condition on the so-
lution. In this case, we replace the mass conservation condition, Eq.(8.19),
by

¯fˆ ρ∗ , τ ∗ , τ ∗ , τ ∗ .
X X
ρ∗ = f¯i +


i xx xy yy (8.24)
i∈I i∈I
/

For the node depicted in Figure 8.1 the above equation, Eq, (8.19) also
gives a relationship between the dimensionless mass density and the normal
component of the viscous stress

1 ∗ ∗ 2
 1
ρ 3uy − 3u∗y + 5 − τyy ∗
= ρI ,
6 2
and Eq. (8.21) for α = β = 2 closes the system of equations for ρ∗ and
∗
τyy
1 ∗  ∗2  2
ρ 6uy + 3u∗y − 1 + τyy ∗
= (myy )I .
9 3
∗
Solving this system gives the following relationship for ρ ,
8.3. CORNER NODES 167

Figure 8.2: A concave corner node

1
ρ∗ =



4ρI + 3 (myy )I .
3 1 + u∗y
The remaing second order moments can be written in terms of ρ∗ ,

∗ 3 1  2 
τyy = (myy )I − ρ∗ 6u∗y + 3u∗y − 1 ,
2 6

∗ 1
= 2 (mxy )I − ρ∗ u∗x 3u∗y + 1 ,


τxy
3

∗ 6 2
τxx = (mxx )I − ρ∗ u∗x .
5

8.3 Corner nodes

Figure 8.2 shows a concave corner node. Pre-collisional populations f¯1 , f¯4 ,
f¯8 , f¯5 and f¯7 are unknowns and should be replaced by their kinetic projections.
In this case, the set related to the incoming populations is I = {0, 2, 3, 6}and
O = {0, 1, 4, 8} designate the outgoing directions from the boundary node to
the uid phase. Remark that the pre-collisional populations f¯5 and f¯7 do not
belong to anyone of these population-sets and, consequently, do not contribute
to the mass conservation equation, Eq. (8.19). Furthermore, they give

1
f¯5 − f¯7 = ρ∗ u∗x + u∗y ,


6
being identical only for the case when the corner boundary node is at rest,
u∗x = u∗y =0.
In Latt et al. [59] approach, Eq. (8.13) is useless for evaluating the density
at corner nodes. Therefore Latt et al. propose to extrapolate the density of
the next neighbors with a second order precision.
Extending the approach to concave non-slip corners, Mohammed and Reis
[71] propose a system of 6 equations to determine the 5 unknown populations
168 CHAPTER 8. BOUNDARY CONDITIONS: PART II

and density. The system is given by Eqs (8.7)-(8.12) considering the tensions
∗ ∗ ∗
τxx , τxy and τyy null.
For nding the unknown populations in Hegele et al model [42], the proce-
dure is exactly the same as presented in Section 8.2.4 for the plane surfaces.
We solve the conservation of mass Eq.(8.19) requiring

X bout X
fi = f¯i ,
i∈O i∈I

with I = {0, 2, 3, 6} and O = {0, 1, 4, 8} and Eq. (8.21) is used for the
∗ ∗ ∗
moments τxx , τxy and τyy .
The solution for ux = u∗y = 0
∗
is obtained as

12 
ρ∗


= 3ρI + (1 − ω) 3 (myy )I + 3 (mxx )I + 7 (mxy )I
9ω + 16
∗ 2  
τxx = 12ρI − 12 (myy )I (ω − 1) + 15 (mxx )I (ω + 4) + 10 (mxy )I (ω − 6)
3 (9ω + 16)
∗ 2  
τxy = 42ρI − 15 (myy )I (ω − 6) − 15 (mxx )I (ω − 6) − 50 (mxy )I (2ω + 9)
9 (9ω + 16)
∗ 2  
τyy = 12ρI + 15 (myy )I (ω + 4) − 12 (mxx )I (ω − 1) + 10 (mxy )I (ω − 6)
3 (9ω + 16)

Alternatively, when the density concept replaces the mass conservation con-
dition, we solve Eq. (8.24), instead of Eq. (8.19), and Eqs.(8.21). The solution
for u∗x = u∗y = 0 is, now, independent on the collision frequency and is given by

 
∗ 9 7
ρ = ρI + (myy )I + (mxx )I − (mxy )I
4 3
∗ 1  
τxx = ρI + (myy )I + 5 (mxx )I − 5 (mxy )I
2 
∗ 1 7
τxy = ρI + 5 (myy )I + 5 (mxx )I − 25 (mxy )I
4 3
∗ 1 
τyy = ρI + 5 (myy )I + (mxx )I − 5 (mxy )I
2
This procedure is the same for convex corner sites and, therefore, can be
extended to staircase surfaces.

Exercises
1. For the boundary node shown in Fig. 8.2, nd the populations f¯1 , f¯4 ,
f¯8 , f¯5 and f¯7 and the density ρ ∗
by solving Eqs. (8.7)-(8.12), following the
method of Mohammed and Reis [71].
2. Using Hegele et al. [42] method, presented in Secion 8.2.4, nd the
expressions for the density ρ∗ and second order moments
∗
τxx ∗
, τxy and
∗
τyy
∗
when the corner boundary node is moving with a velocity ux 6= 0.
8.4. COMPARATIVE ANALYSIS 169

Figure 8.3: Lid-driven cavity ow for Re=1000 [ ? ].

8.4 Comparative analysis

The incompressible ow in a square cavity whose top wall moves with a
uniform velocity has been widely used as a model problem for testing and
evaluating numerical techniques (Figure 8.3), in spite of the singularities that
we have at the two top corners ([30], [47], [8]; [92], [83], [10] ). This ow
is of great scientic interest because it displays almost all uid mechanical
phenomena in a simple geometrical setting. In this kind of ow problem, a
main primary vortex is formed near the center of the cavity and secondary
and even tertiary vortices appear in the corners and are intensied when the
Reynolds number increases.
The purpose of this section is to show some numerical simulations that were
performed to test the new proposed set of boundary conditions presented in
Section 3.2 with respect to accuracy, grid convergence, mass conservation and
stability. Simulations were performed for Re= 100, 400, 1000, 3200 and 5000.
Simulation results were organized as follows:

8.4.1 Accuracy
The accuracy of the presented methods was tested by Bazarin et al.[ ? ],
comparing the results obtained with moments-based methods with the classical
results of Ghia et al. [30], and the more recent benchmarks of Botella and
Peyret [8] and Erturk, Corke and Gökçöl [25]. Ghia et al. and Erturk et
170 CHAPTER 8. BOUNDARY CONDITIONS: PART II

Figure 8.4: Numerical results for Re=400, 1000 and 5000 showing the relative
velocity u∗x /u∗0 along the mid-line y ∗ = L/2. Results are compared with Ghia et
al. [30], Botella and Peyret [8] and Erturk, Corke and Gökçöl [25]. Continuous
lines are from Bazarin et al. [ ? ].

al. solved this problem using discretization schemes based on the vorticity-
stream formulation of the Navier-Stokes equations. The work of Botella and
Peyret uses a highly-accurate solution based on Chebyshev collocation method
for the spatial discretization in an assymptotic expansion of the Navier-Stokes
equations.
Comparing the velocity proles with other works, Figures 8.4 and 8.5 shows
for Re = 100, 400, 1000 and 5000 the numerical results for the relative velocity
u∗x /u∗0 along the mid-line y ∗ = L/2 and the relative velocity u∗y /u∗0 along the
∗
mid-line y = L/2 of the cavity, respectively. Results were based on simulations
performed using a 512 x 512 numerical domain. Adherence conditions are
veried for all methods.
Figure 8.6 shows plots of the stream function and the vorticity isolines.
These stream function isolines give a very clear picture of the complex ow
structure that develops in the cavity ow and the eect of the Reynolds number
on the patterns of the steady recirculating eddies in the cavity. The increas-
ingly complexity of the ow structure with the Reynolds number can be also
appreciated in the second row of Figure 8.6, which shows the vorticity isolines
for the lid driven cavity when Re varies from 100 to 5000.
A primary vortex and a pair of counterrotating eddies of much smaller
strength develop in the lower corners of the cavity are observed for all values
of Re presented. At Re = 1000 and 5000 a second pair of tertiary vortices are
seen in the lower corners of the cavity and exclusively for Re = 5000 a left top
vortex is observed. Additionally, as shown in Table 3, for low Re (Re=100), the
center of the primary vortex is located in the upper-right quadrant of the cavity
and moves down towards the geometric center of the cavity as the Reynolds
number increases.
The velocity proles shown in Figures 8.4 and 8.5, and stream function and
vorticity elds in Figure 8.6 were obtained using the method based on the mass
8.4. COMPARATIVE ANALYSIS 171

Figure 8.5: Numerical results for Re=400, 1000 and 5000 showing the relative
velocity u∗y /u∗0 along the mid-line x∗ = L/2. Results are compared with Ghia et
al. [30], Botella and Peyret [8] and Erturk, Corke and Gökçöl [25]. Continuous
lines are from Bazarin et al. [ ? ].

Figure 8.6: Stream function and vorticity isolines for (a) Re = 100, (b) Re =
400, (c) Re = 1000, (d) Re = 3200 and (e) Re = 5000. The top row illustrate
the stream function and the bottom row the vorticity, Bazarin et al. [ ? ].
172 CHAPTER 8. BOUNDARY CONDITIONS: PART II

conservation condition and on second-order moments. Nonetheless, there are

no meaningful dierences when these velocity proles are obtained using other
methods. These velocity proles were measured along the midlines x∗ = L/2
∗
and y = L/2, but there is, also, a good agreement for the main ow patterns
shown in Figure 8.6 concerning the location of the vortex centers and the values
of the stream function and vorticity. In addition, the appearance and position
of secondary and tertiary vortices are predicted with the same accuracy in
all methods I-IV. In other words, all these methods lead to almost the same
macroscopic ow patterns and these ow patterns are in good agreement with
previous work.
Therefore, .....................

8.4.2 Mass leakage

The lid driven cavity is a closed system and, so, the total mass of the uid
inside it is to be conserved with time. Nevertheless, as asserted in Section 8.1,
it is not possible to prevent mass leakage when the boundary conditions are
based on the density concept, Eq. (8.1), instead of the much more restrictive
condition represented by Eq. (8.3).
Therefore the average density

¯
P ∗
x,y ρ
ρ= ,
L∗2
was measured for each time step for all the models that were analyzed in this
work. Results are shown in Figures 8.7 and 8.8.
As expected, both methods (I and II), based on Eq. (8.3), are clearly mass
conservative but Method IV also appears as mass conservative. For clarifying
this issue Bazarin et al. (2021) changed the average density scale in Figure 8.8,
measuring the total mass variation with a more precise weight balance. While
in both methods based on Eq. (8.3) the total mass is constant with time,
method IV based on density, presents the same linear behavior as observed in
Malaspinas et al. (2011) and Latt et al. (2008) models, although with a much
less mass variation. Curiously, Mohammed & Reis (2017) model preserves the
total mass although being a density based model.
Therefore, for investigating this issue more carefully, local mass leaking was
also veried by computing

¯fˆ out (xb , t) ,

X X
4ρ∗ (xb , t) = f¯i (xb , t) − i
i∈I i∈O

at each time step and at each boundary node, using the information on the
pre-collisional populations f¯i that were propagated to the boundary sites xb
from the uid phase and the information on the populations ¯fˆ out that leave
i
these sites after collision and that were calculated based on the information we
have for the rst moments ρ∗ , τxx
∗ ∗
, τxy and
∗
τyy .
Figure 8.9 shows the local mass leakage for Mohammed and Reis (2017)
model after 50000 time steps. It is clear that Mohammed and Reis model is
8.4. COMPARATIVE ANALYSIS 173

Figure 8.7: Average density in terms of the number of time steps. We compare
the results of Latt et al. (2008), Malaspinas et al. (2011) and Mohammed
and Reis (2017) with the solution methods presented by Bazarin et al. (2021).
Both Malaspinas et al. and Latt et al. models fail in preserving the total mass
of the system. Method IV based on the concept of density is neither a mass
preserving method, but mass leaking is much smaller, Bazarin et al. [ ? ].

Figure 8.8: Average density measured with a more accurate weigth balance in
terms of the number of time steps. method IV based on density, presents the
same linear behavior as observed in Malaspinas et al. (2011) and Latt et al.
(2008) models, although with a much less mass variation, Bazarin et al. [ ? ].
174 CHAPTER 8. BOUNDARY CONDITIONS: PART II

Figure 8.9: Local mass leakage for Mohammed and Reis (2017) model after
50000 time steps: (a) north and south boundaries; (b) east and west boundaries;
8.4. COMPARATIVE ANALYSIS 175

not mass conservative from a local standpoint. In fact, although the model is
globally mass conservative there is a local mass leaking on the upper corners.
Nevertheless, the addition of mass on the left upper corner is compensated by
a loss of mass on the right upper corner.
176 CHAPTER 8. BOUNDARY CONDITIONS: PART II
Bibliography
[1] T. Abe, Derivation of the Lattice Boltzmann Method by Means of the
Discrete Ordinate Method for the Boltzmann Equation, Journal of Com-
putational Physics, 131 (1997), pp. 241246.

[2] D. Albernaz, M. Do-Quang, and G. Amberg, Multirelaxation-time

lattice Boltzmann model for droplet heating and evaporation under forced
convection, Physical Review E, 91 (2015), p. 043012.

[3] P. Asinari and I. V. Karlin, Quasiequilibrium lattice Boltzmann models

with tunable bulk viscosity for enhancing stability, Physical Review E, 81
(2010), p. 016702.

[4] P. L. Bhatnagar, E. P. Gross, and M. Krook, A Model for Collision

Processes in Gases. I. Small Amplitude Processes in Charged and Neutral
One-Component Systems, Physical Review, 94 (1954), pp. 511525.

[5] L. Boltzmann, Weitere Studien uber das Warmegleichgewicht unter Gas-

molekun, Sitzungberichte der Akademie der Wissenschaften zu Wien,
mathematisch-naturwissenschaftliche Klasse, 66 (1872), pp. 275370.

[6] , On some problems of the mechanical theory of heat, Philosophical

Magazine Series 5, 6 (1878), pp. 236237.

[7] , Vorlesungen uber Gas Theory, Part II, J.A. Barth, Leipzig, 1896.

[8] O. Botella and R. Peyret, Benchmark spectral results on the lid-

driven cavity ow, Computers and Fluids, 27 (1998), pp. 421433.

[9] R. A. Brownlee, J. Levesley, D. Packwood, and A. N. Gor-

ban, Add-ons for Lattice Boltzmann Methods: Regularization, Filtering
and Limiters, Progress in Computational Physics, (2013), pp. 3152.

[10] C.-H. Bruneau and M. Saad, The 2D lid-driven cavity problem revis-
ited, Computers and Fluids, 35 (2006), pp. 326348.

[11] C. Cercignani, Mathematical Methods in Kinetic Theory, London:

Macmillan, 1969.

177
178 BIBLIOGRAPHY

[12] S. Chapman and T. Cowling, The Mathematical Theory of Non-

uniform Gases, Cambridge University Press, 1970.

[13] H. Chen, S. Chen, and W. H. Matthaeus, Recovery of the Navier-

Stokes equations using a lattice-gas Boltzmann method, Physical Review
A, 45 (1992), pp. 53395342.

[14] H. Chen and X. Shan, Fundamental conditions for N-th-order accurate

lattice Boltzmann models, Physica D: Nonlinear Phenomena, 237 (2008),
pp. 20032008.

[15] H. Chen, R. Zhang, I. Staroselsky, and M. Jhon, Recovery of full

rotational invariance in lattice Boltzmann formulations for high Knudsen
number ows, Physica A: Statistical Mechanics and its Applications, 362
(2006), pp. 125131.

[16] S. Chen, H. Chen, D. Martinez, and W. Matthaeus, Lattice Boltz-

mann model for simulation of magnetohydrodynamics, Physical Review
Letters, 67 (1991), pp. 37763779.

[17] S. S. Chikatamarla and I. V. Karlin, Lattices for the lattice Boltz-

mann method, Physical Review E, 79 (2009), p. 046701.

[18] R. Clausius, Über die Wärmeleitung gasförmiger Körper, Annalen der

Physik, 125 (1850), pp. 353-400.

[19] , Ueber die bewegende Kraft der Wärme und die Gesetze, welche sich
daraus für die Wärmelehre selbst ableiten lassen, Annalen der Physik und
Chemie, 155 (1850), pp. 500524.

[20] , The mechanical theory of heat, with its applications to the steam-
engine and to the physical properties of bodies, John van Voorst, London,
1867.

[21] R. Courant, K. Friedrichs, and H. Lewy, Uber die partiellen Dif-

ferenzengleichungen der mathematischen Physik, Mathematische Annalen,
100 (1928), pp. 3274.

[22] P. J. Dellar, Incompressible limits of lattice Boltzmann equations using

multiple relaxation times, Journal of Computational Physics, 190 (2003),
pp. 351370.

[23] D. D'Humieres, Generalized Lattice-Boltzmann Equations, in Rareed

Gas Dynamics: Theory and Simulations, vol. 159, American Institute of
Aeronautics and Astronautics, Washington DC, jan 1994, pp. 450458.

[24] D. D'Humières, I. Ginzburg, M. Krafczyk, P. Lallemand, and

L.-S. Luo, Multiple-relaxation-time lattice Boltzmann models in three di-
mensions., Philosophical transactions. Series A, Mathematical, physical,
and engineering sciences, 360 (2002), pp. 43751.
BIBLIOGRAPHY 179

[25] E. Erturk, T. C. Corke, and C. Gökçöl, Numerical solutions of

2-D steady incompressible driven cavity ow at high Reynolds numbers,
International Journal for Numerical Methods in Fluids, 48 (2005), pp. 747
774.

[26] U. Frisch, B. Hasslacher, and Y. Pomeau, Lattice-Gas Automata for

the Navier-Stokes Equation, Physical Review Letters, 56 (1986), pp. 1505
1508.

[27] M. Geier, A. Greiner, and J. G. Korvink, Cascaded digital lattice

Boltzmann automata for high Reynolds number ow, Physical Review E -
Statistical, Nonlinear, and Soft Matter Physics, 73 (2006), pp. 110.

[28] M. Geier, A. Greiner, and J. G. Korvink, A factorized central mo-

ment lattice Boltzmann method, The European Physical Journal Special
Topics, 171 (2009), pp. 5561.

[29] M. Geier, M. Schönherr, A. Pasquali, and M. Krafczyk, The cu-

mulant lattice Boltzmann equation in three dimensions: Theory and valida-
tion, Computers and Mathematics with Applications, 70 (2015), pp. 507
547.

[30] U. Ghia, K. Ghia, and C. Shin, High-Re solutions for incompressible

ow using the Navier-Stokes equations and a multigrid method, Journal of
Computational Physics, 48 (1982), pp. 387411.

[31] J. Gibbs, On the equilibrium of heterogeneous substances. Part II, Trans-

actions of the Connecticut Academy of Arts and Sciences, 3 (1874),
pp. 343524.

[32] D. Golbert, P. Blanco, A. Clausse, and R. Feijóo, On the search

of more stable second-order lattice-Boltzmann schemes in conned ows,
Journal of Computational Physics, 294 (2015), pp. 605618.

[33] H. Grad, Note on N-dimensional hermite polynomials, Communications

on Pure and Applied Mathematics, 2 (1949), pp. 325330.

[34] E. P. Gross and E. A. Jackson, Kinetic Models and the Linearized

Boltzmann Equation, Physics of Fluids, 2 (1959), p. 432.

[35] Z. Guo, S. Chang, and C. Shu, Lattice Boltzmann Method and its
Applications in Engineering, vol. 5, World Scientic Publishers, Singapore,
2013.

[36] J. Hardy, O. de Pazzis, and Y. Pomeau, Molecular dynamics of a

classical lattice gas: Transport properties and time correlation functions,
Physical Review A, 13 (1976), pp. 19491961.

[37] J. Hardy, Y. Pomeau, and O. de Pazzis, Time Evolution of a Two-

Dimensional Classical Lattice System, Physical Review Letters, 31 (1973),
pp. 276279.
180 BIBLIOGRAPHY

[38] X. He and G. D. Doolen, Lattice Boltzmann method on curvilinear

coordinates system: Flow around a circular cylinder, J. Comput. Phys.,
134 (1997), pp. 306315.

[39] , Thermodynamic foundations of kinetic theory and lattice Boltzmann

models for multi-phase ows, Journal of Statistical Physics, 107 (2002),
pp. 309328.

[40] X. He and L.-S. Luo, Theory of the lattice Boltzmann method: From the
Boltzmann equation to the lattice Boltzmann equation, Physical Review E,
56 (1997), pp. 68116817.

[41] X. He, X. Shan, and G. Doolen, Discrete Boltzmann equation model

for nonideal gases, Physical Review E, 57 (1998), pp. R13R16.

[42] L. A. Hegele, A. Scagliarini, M. Sbragaglia, K. K. Mattila,

P. C. Philippi, D. F. Puleri, J. Gounley, and A. Randles, High-
Reynolds-number turbulent cavity ow using the lattice Boltzmann method,
Physical Review E, 98 (2018), p. 043302.

[43] H. Helmholtz, On the conservation of force, Transactions of the Physical

Society of Berlin, (1847), pp. 114162.

[44] W. Henry, Experiments on the Quantity of Gases Absorbed by Water,

at Dierent Temperatures, and under Dierent Pressures, Philosophical
Transactions of the Royal Society of London, 93 (1803), pp. 29274.

[45] F. J. Higuera and J. Jiménez, Boltzmann Approach to Lattice Gas

Simulations, Europhysics Letters (EPL), 9 (1989), pp. 663668.

[46] F. J. Higuera, S. Succi, and R. Benzi, Lattice Gas Dynamics with

Enhanced Collisions, Europhysics Letters (EPL), 9 (1989), pp. 345349.

[47] S. Hou, Q. Zou, S. Chen, G. Doolen, and A. Cogley, Simulation of

Cavity Flow by the Lattice Boltzmann Method, Journal of Computational
Physics, 118 (1995), pp. 329347.

[48] T. Inamuro, M. Yoshino, and F. Ogino, A non-slip boundary condi-

tion for lattice Boltzmann simulations, Physics of Fluids, 7 (1995), p. 2928.

[49] J. P. Joule, On the Mechanical Equivalent of Heat, Philosophical Trans-

actions of the Royal Society of London, 140 (1850), pp. 6182.

[50] I. V. Karlin, F. Bösch, and S. S. Chikatamarla, Gibbs' principle for

the lattice-kinetic theory of uid dynamics, Physical Review E, 90 (2014),
p. 031302.

[51] I. V. Karlin, A. Ferrante, and H. C. Öttinger, Perfect entropy

functions of the Lattice Boltzmann method, Europhysics Letters (EPL), 47
(1999), pp. 182188.
BIBLIOGRAPHY 181

[52] T. Krüger, H. Kusumaatmaja, A. Kuzmin, O. Shardt, G. Silva,

and E. M. Viggen, The Lattice Boltzmann Method, Graduate Texts in
Physics, Springer International Publishing, Cham, 2017.

[53] A. Kupershtokh and D. Medvedev, Lattice Boltzmann equation

method in electrohydrodynamic problems, Journal of Electrostatics, 64
(2006), pp. 581585.

[54] A. Kupershtokh, D. Medvedev, and D. Karpov, On equations of

state in a lattice Boltzmann method, Computers and Mathematics with
Applications, 58 (2009), pp. 965974.

[55] A. J. C. Ladd, Numerical simulations of particulate suspensions via a

discretized Boltzmann equation. Part 1. Theoretical foundation, Journal of
Fluid Mechanics, 271 (1994), p. 285.

[56] P. Lallemand and L.-S. Luo, Theory of the lattice Boltzmann Method:
Dispersion, Dissipation, Isotropy, Galilean Invariance, and Stability,
Physical Review E, 61 (2000), pp. 65466562.

[57] , Theory of the lattice Boltzmann method: Acoustic and thermal

properties in two and three dimensions, Physical Review E, 68 (2003),
p. 036706.

[58] J. Latt and B. Chopard, Lattice Boltzmann method with regularized

pre-collision distribution functions, Mathematics and Computers in Sim-
ulation, 72 (2006), pp. 165168.

[59] J. Latt, B. Chopard, O. Malaspinas, M. Deville, and A. Mich-

ler, Straight velocity boundaries in the lattice Boltzmann method, Physical
Review E, 77 (2008), p. 056703.

[60] A. Lavoisier, Mémoire sur la chaleur lu à l'Académie royale des sciences,

le 28 juin 1783, Mémoires de l' Académie Royale des Sciences, (1783).

[61] J. E. Lennard-Jones, On the Determination of Molecular Fields. II.

From the Equation of State of a Gas, Proceedings of the Royal Society
A: Mathematical, Physical and Engineering Sciences, 106 (1924), pp. 463
477.

[62] L.-S. Luo, W. Liao, X. Chen, Y. Peng, and W. Zhang, Numerics

of the lattice Boltzmann method: Eects of collision models on the lattice
Boltzmann simulations, Physical Review E, 83 (2011), p. 056710.

[63] O. Malaspinas, Increasing stability and accuracy of the lattice Boltzmann

scheme: recursivity and regularization, arXiv:1505.06900v1 [physics.u-
dyn] 26 May 2015, (2015), pp. 131.

[64] O. Malaspinas, B. Chopard, and J. Latt, General regularized bound-

ary condition for multi-speed lattice Boltzmann models, Computers and
Fluids, 49 (2011), pp. 2935.
182 BIBLIOGRAPHY

[65] K. K. Mattila, P. C. Philippi, and L. A. Hegele, High-order reg-

ularization in lattice-Boltzmann equations, Physics of Fluids, 29 (2017),
p. 046103.

[66] K. K. K. Mattila, L. L. A. L. Hegele, and P. P. C. Philippi,

Investigation of an entropic stabilizer for the lattice-Boltzmann method,
Physical Review E, 91 (2015), p. 063010.

[67] J. C. Maxwell, On the dynamical theory of gases, Philosophical trans-

actions of the Royal Society, 157 (1867), pp. 4988.

[68] , Molecules, Nature, (1873), pp. 437441.

[69] G. McNamara and B. Alder, Analysis of the lattice Boltzmann treat-

ment of hydrodynamics, Physica A: Statistical Mechanics and its Applica-
tions, 194 (1993), pp. 218228.

[70] G. R. McNamara and G. Zanetti, Use of the boltzmann equation

to simulate lattice-gas automata, Physical Review Letters, 61 (1988),
pp. 23322335.

[71] S. Mohammed and T. Reis, Using the Lid-Driven Cavity Flow to Vali-
date Moment-Based Boundary Conditions for the Lattice Boltzmann Equa-
tion, Archive of Mechanical Engineering, 64 (2017), pp. 5774.

[72] a. Montessori, G. Falcucci, P. Prestininzi, M. La Rocca, and

S. Succi, Regularized lattice Bhatnagar-Gross-Krook model for two- and
three-dimensional cavity ow simulations, Physical Review E, 89 (2014),
p. 053317.

[73] A. Montessori, P. Prestininzi, M. La Rocca, and S. Succi, Lattice

Boltzmann approach for complex nonequilibrium ows, Physical Review E
- Statistical, Nonlinear, and Soft Matter Physics, 92 (2015), pp. 16.

[74] M. Namburi, P. Kumar, S. Krithivasan, and S. Ansumali, Crystal-

lographic Lattice Boltzmann method for Hydrodynamics, in 24th Discrete
Simulation of Fluid Dynamics (DSFD 2015), Edinburgh, 2015, p. 96.

[75] P. C. Philippi, L. A. Hegele Júnior, L. Dos Santos, and R. Sur-

mas, From the continuous to the lattice Boltzmann equation: The dis-
cretization problem and thermal models, Physical Review E, 73 (2006),
p. 056702.

[76] P. C. Philippi, K. K. Mattila, D. N. Siebert, L. O. E. dos Santos,

L. A. Hegele Júnior, and R. Surmas, Lattice-Boltzmann equations for
describing segregation in non-ideal mixtures, Journal of Fluid Mechanics,
713 (2012), pp. 564587.

[77] M. Planck, Treatise on Thermodynamics, Dover Publications, New York,

1917.
BIBLIOGRAPHY 183

[78] K. N. Premnath and S. Banerjee, On the Three-Dimensional Central

Moment Lattice Boltzmann Method, Journal of Statistical Physics, 143
(2011), pp. 747794.

[79] Y. H. Qian, D. D'Humières, and P. Lallemand, Lattice BGK Mod-

els for Navier-Stokes Equation, Europhysics Letters (EPL), 17 (1992),
pp. 479484.

[80] J. S. Rowlinson, Cohesion: a Scientic History of Intermolecular Forces,

Cambridge University Press, 2004.

[81] X. Shan, General solution of lattices for Cartesian lattice Bhatanagar-

Gross-Krook models, Physical Review E, 81 (2010), p. 036702.

[82] X. Shan, X.-F. Yuan, and H. Chen, Kinetic theory representation of

hydrodynamics: a way beyond the Navier-Stokes equation, Journal of Fluid
Mechanics, 550 (2006), pp. 413441.

[83] P. N. Shankar and M. D. Deshpande, Fluid Mechanics in the Driven

Cavity, Annual Review of Fluid Mechanics, 32 (2000), pp. 93136.

[84] D. Siebert, L. A. Hegele Júnior, R. Surmas, L. dos Santos, and

P. Philippi, Thermal lattice Boltzmann in two dimensions, International
Journal of Modern Physics C, 18 (2007), pp. 546555.

[85] D. N. Siebert, L. A. Hegele, P. C. Philippi, L. A. Hegele Júnior,

and P. C. Philippi, Lattice Boltzmann equation linear stability analysis:
Thermal and athermal models, Physical Review E, 77 (2008), p. 026707.

[86] S. Succi, The Lattice Boltzmann Equation for Fluid Dynamics and Be-
yond (Numerical Mathematics and Scientic Computation) 1st Edition,
Clarendon Press, 2001.

[87] R. Surmas, C. E. Pico Ortiz, and P. C. Philippi, Simulating ther-

mohydrodynamics by nite dierence solutions of the Boltzmann equation,
The European Physical Journal Special Topics, 171 (2009), pp. 8190.

[88] J. D. van der Waals, Over de Continuiteit van den Gas en Vloeistoftoe-
stand, PhD thesis, Leiden University, 1873.

[89] J.-S. S. Wu and Y.-L. L. Shao, Simulation of lid-driven cavity ows

by parallel lattice Boltzmann method using multi-relaxation-time scheme,
International Journal for Numerical Methods in Fluids, 46 (2004), pp. 921
937.

[90] H. Xu and P. Sagaut, Optimal low-dispersion low-dissipation LBM

schemes for computational aeroacoustics, Journal of Computational
Physics, 230 (2011), pp. 53535382.

[91] R. Zhang, X. Shan, and H. Chen, Ecient kinetic method for uid
simulation beyond the Navier-Stokes equation, Physical Review E, 74
(2006), p. 046703.
184 BIBLIOGRAPHY

[92] W. Zhang, C. Zhang, and G. Xi, An explicit Chebyshev pseudospectral

multigrid method for incompressible Navier-Stokes equations, Computers
and Fluids, 39 (2010), pp. 178188.

[93] Q. Zou and X. He, On pressure and velocity boundary conditions for the
lattice Boltzmann BGK model, Physics of Fluids, 9 (1997), pp. 15911598.

[94] , On pressure and velocity ow boundary conditions and bounceback

for the lattice Boltzmann BGK model, Physics of Fluids, 1591 (1997), p. 18.
Bibliography
[1] T. Abe, Derivation of the Lattice Boltzmann Method by Means of the
Discrete Ordinate Method for the Boltzmann Equation, Journal of Com-
putational Physics, 131 (1997), pp. 241246.

[2] D. Albernaz, M. Do-Quang, and G. Amberg, Multirelaxation-time

lattice Boltzmann model for droplet heating and evaporation under forced
convection, Physical Review E, 91 (2015), p. 043012.

[3] P. Asinari and I. V. Karlin, Quasiequilibrium lattice Boltzmann models

with tunable bulk viscosity for enhancing stability, Physical Review E, 81
(2010), p. 016702.

[4] P. L. Bhatnagar, E. P. Gross, and M. Krook, A Model for Collision

Processes in Gases. I. Small Amplitude Processes in Charged and Neutral
One-Component Systems, Physical Review, 94 (1954), pp. 511525.

[5] L. Boltzmann, Weitere Studien uber das Warmegleichgewicht unter Gas-

molekun, Sitzungberichte der Akademie der Wissenschaften zu Wien,
mathematisch-naturwissenschaftliche Klasse, 66 (1872), pp. 275370.

[6] , On some problems of the mechanical theory of heat, Philosophical

Magazine Series 5, 6 (1878), pp. 236237.

[7] , Vorlesungen uber Gas Theory, Part II, J.A. Barth, Leipzig, 1896.

[8] O. Botella and R. Peyret, Benchmark spectral results on the lid-

driven cavity ow, Computers and Fluids, 27 (1998), pp. 421433.

[9] R. A. Brownlee, J. Levesley, D. Packwood, and A. N. Gor-

ban, Add-ons for Lattice Boltzmann Methods: Regularization, Filtering
and Limiters, Progress in Computational Physics, (2013), pp. 3152.

[10] C.-H. Bruneau and M. Saad, The 2D lid-driven cavity problem revis-
ited, Computers and Fluids, 35 (2006), pp. 326348.

[11] C. Cercignani, Mathematical Methods in Kinetic Theory, London:

Macmillan, 1969.

185
186 BIBLIOGRAPHY

[12] S. Chapman and T. Cowling, The Mathematical Theory of Non-

uniform Gases, Cambridge University Press, 1970.

[13] H. Chen, S. Chen, and W. H. Matthaeus, Recovery of the Navier-

Stokes equations using a lattice-gas Boltzmann method, Physical Review
A, 45 (1992), pp. 53395342.

[14] H. Chen and X. Shan, Fundamental conditions for N-th-order accurate

lattice Boltzmann models, Physica D: Nonlinear Phenomena, 237 (2008),
pp. 20032008.

[15] H. Chen, R. Zhang, I. Staroselsky, and M. Jhon, Recovery of full

rotational invariance in lattice Boltzmann formulations for high Knudsen
number ows, Physica A: Statistical Mechanics and its Applications, 362
(2006), pp. 125131.

[16] S. Chen, H. Chen, D. Martinez, and W. Matthaeus, Lattice Boltz-

mann model for simulation of magnetohydrodynamics, Physical Review
Letters, 67 (1991), pp. 37763779.

[17] S. S. Chikatamarla and I. V. Karlin, Lattices for the lattice Boltz-

mann method, Physical Review E, 79 (2009), p. 046701.

[18] R. Clausius, Über die Wärmeleitung gasförmiger Körper, Annalen der

Physik, 125 (1850), pp. 353-400.

[19] , Ueber die bewegende Kraft der Wärme und die Gesetze, welche sich
daraus für die Wärmelehre selbst ableiten lassen, Annalen der Physik und
Chemie, 155 (1850), pp. 500524.

[20] , The mechanical theory of heat, with its applications to the steam-
engine and to the physical properties of bodies, John van Voorst, London,
1867.

[21] R. Courant, K. Friedrichs, and H. Lewy, Uber die partiellen Dif-

ferenzengleichungen der mathematischen Physik, Mathematische Annalen,
100 (1928), pp. 3274.

[22] P. J. Dellar, Incompressible limits of lattice Boltzmann equations using

multiple relaxation times, Journal of Computational Physics, 190 (2003),
pp. 351370.

[23] D. D'Humieres, Generalized Lattice-Boltzmann Equations, in Rareed

Gas Dynamics: Theory and Simulations, vol. 159, American Institute of
Aeronautics and Astronautics, Washington DC, jan 1994, pp. 450458.

[24] D. D'Humières, I. Ginzburg, M. Krafczyk, P. Lallemand, and

L.-S. Luo, Multiple-relaxation-time lattice Boltzmann models in three di-
mensions., Philosophical transactions. Series A, Mathematical, physical,
and engineering sciences, 360 (2002), pp. 43751.
BIBLIOGRAPHY 187

[25] E. Erturk, T. C. Corke, and C. Gökçöl, Numerical solutions of

2-D steady incompressible driven cavity ow at high Reynolds numbers,
International Journal for Numerical Methods in Fluids, 48 (2005), pp. 747
774.

[26] U. Frisch, B. Hasslacher, and Y. Pomeau, Lattice-Gas Automata for

the Navier-Stokes Equation, Physical Review Letters, 56 (1986), pp. 1505
1508.

[27] M. Geier, A. Greiner, and J. G. Korvink, Cascaded digital lattice

Boltzmann automata for high Reynolds number ow, Physical Review E -
Statistical, Nonlinear, and Soft Matter Physics, 73 (2006), pp. 110.

[28] M. Geier, A. Greiner, and J. G. Korvink, A factorized central mo-

ment lattice Boltzmann method, The European Physical Journal Special
Topics, 171 (2009), pp. 5561.

[29] M. Geier, M. Schönherr, A. Pasquali, and M. Krafczyk, The cu-

mulant lattice Boltzmann equation in three dimensions: Theory and valida-
tion, Computers and Mathematics with Applications, 70 (2015), pp. 507
547.

[30] U. Ghia, K. Ghia, and C. Shin, High-Re solutions for incompressible

ow using the Navier-Stokes equations and a multigrid method, Journal of
Computational Physics, 48 (1982), pp. 387411.

[31] J. Gibbs, On the equilibrium of heterogeneous substances. Part II, Trans-

actions of the Connecticut Academy of Arts and Sciences, 3 (1874),
pp. 343524.

[32] D. Golbert, P. Blanco, A. Clausse, and R. Feijóo, On the search

of more stable second-order lattice-Boltzmann schemes in conned ows,
Journal of Computational Physics, 294 (2015), pp. 605618.

[33] H. Grad, Note on N-dimensional hermite polynomials, Communications

on Pure and Applied Mathematics, 2 (1949), pp. 325330.

[34] E. P. Gross and E. A. Jackson, Kinetic Models and the Linearized

Boltzmann Equation, Physics of Fluids, 2 (1959), p. 432.

[35] Z. Guo, S. Chang, and C. Shu, Lattice Boltzmann Method and its
Applications in Engineering, vol. 5, World Scientic Publishers, Singapore,
2013.

[36] J. Hardy, O. de Pazzis, and Y. Pomeau, Molecular dynamics of a

classical lattice gas: Transport properties and time correlation functions,
Physical Review A, 13 (1976), pp. 19491961.

[37] J. Hardy, Y. Pomeau, and O. de Pazzis, Time Evolution of a Two-

Dimensional Classical Lattice System, Physical Review Letters, 31 (1973),
pp. 276279.
188 BIBLIOGRAPHY

[38] X. He and G. D. Doolen, Lattice Boltzmann method on curvilinear

coordinates system: Flow around a circular cylinder, J. Comput. Phys.,
134 (1997), pp. 306315.

[39] , Thermodynamic foundations of kinetic theory and lattice Boltzmann

models for multi-phase ows, Journal of Statistical Physics, 107 (2002),
pp. 309328.

[40] X. He and L.-S. Luo, Theory of the lattice Boltzmann method: From the
Boltzmann equation to the lattice Boltzmann equation, Physical Review E,
56 (1997), pp. 68116817.

[41] X. He, X. Shan, and G. Doolen, Discrete Boltzmann equation model

for nonideal gases, Physical Review E, 57 (1998), pp. R13R16.

[42] L. A. Hegele, A. Scagliarini, M. Sbragaglia, K. K. Mattila,

P. C. Philippi, D. F. Puleri, J. Gounley, and A. Randles, High-
Reynolds-number turbulent cavity ow using the lattice Boltzmann method,
Physical Review E, 98 (2018), p. 043302.

[43] H. Helmholtz, On the conservation of force, Transactions of the Physical

Society of Berlin, (1847), pp. 114162.

[44] W. Henry, Experiments on the Quantity of Gases Absorbed by Water,

at Dierent Temperatures, and under Dierent Pressures, Philosophical
Transactions of the Royal Society of London, 93 (1803), pp. 29274.

[45] F. J. Higuera and J. Jiménez, Boltzmann Approach to Lattice Gas

Simulations, Europhysics Letters (EPL), 9 (1989), pp. 663668.

[46] F. J. Higuera, S. Succi, and R. Benzi, Lattice Gas Dynamics with

Enhanced Collisions, Europhysics Letters (EPL), 9 (1989), pp. 345349.

[47] S. Hou, Q. Zou, S. Chen, G. Doolen, and A. Cogley, Simulation of

Cavity Flow by the Lattice Boltzmann Method, Journal of Computational
Physics, 118 (1995), pp. 329347.

[48] T. Inamuro, M. Yoshino, and F. Ogino, A non-slip boundary condi-

tion for lattice Boltzmann simulations, Physics of Fluids, 7 (1995), p. 2928.

[49] J. P. Joule, On the Mechanical Equivalent of Heat, Philosophical Trans-

actions of the Royal Society of London, 140 (1850), pp. 6182.

[50] I. V. Karlin, F. Bösch, and S. S. Chikatamarla, Gibbs' principle for

the lattice-kinetic theory of uid dynamics, Physical Review E, 90 (2014),
p. 031302.

[51] I. V. Karlin, A. Ferrante, and H. C. Öttinger, Perfect entropy

functions of the Lattice Boltzmann method, Europhysics Letters (EPL), 47
(1999), pp. 182188.
BIBLIOGRAPHY 189

[52] T. Krüger, H. Kusumaatmaja, A. Kuzmin, O. Shardt, G. Silva,

and E. M. Viggen, The Lattice Boltzmann Method, Graduate Texts in
Physics, Springer International Publishing, Cham, 2017.

[53] A. Kupershtokh and D. Medvedev, Lattice Boltzmann equation

method in electrohydrodynamic problems, Journal of Electrostatics, 64
(2006), pp. 581585.

[54] A. Kupershtokh, D. Medvedev, and D. Karpov, On equations of

state in a lattice Boltzmann method, Computers and Mathematics with
Applications, 58 (2009), pp. 965974.

[55] A. J. C. Ladd, Numerical simulations of particulate suspensions via a

discretized Boltzmann equation. Part 1. Theoretical foundation, Journal of
Fluid Mechanics, 271 (1994), p. 285.

[56] P. Lallemand and L.-S. Luo, Theory of the lattice Boltzmann Method:
Dispersion, Dissipation, Isotropy, Galilean Invariance, and Stability,
Physical Review E, 61 (2000), pp. 65466562.

[57] , Theory of the lattice Boltzmann method: Acoustic and thermal

properties in two and three dimensions, Physical Review E, 68 (2003),
p. 036706.

[58] J. Latt and B. Chopard, Lattice Boltzmann method with regularized

pre-collision distribution functions, Mathematics and Computers in Sim-
ulation, 72 (2006), pp. 165168.

[59] J. Latt, B. Chopard, O. Malaspinas, M. Deville, and A. Mich-

ler, Straight velocity boundaries in the lattice Boltzmann method, Physical
Review E, 77 (2008), p. 056703.

[60] A. Lavoisier, Mémoire sur la chaleur lu à l'Académie royale des sciences,

le 28 juin 1783, Mémoires de l' Académie Royale des Sciences, (1783).

[61] J. E. Lennard-Jones, On the Determination of Molecular Fields. II.

From the Equation of State of a Gas, Proceedings of the Royal Society
A: Mathematical, Physical and Engineering Sciences, 106 (1924), pp. 463
477.

[62] L.-S. Luo, W. Liao, X. Chen, Y. Peng, and W. Zhang, Numerics

of the lattice Boltzmann method: Eects of collision models on the lattice
Boltzmann simulations, Physical Review E, 83 (2011), p. 056710.

[63] O. Malaspinas, Increasing stability and accuracy of the lattice Boltzmann

scheme: recursivity and regularization, arXiv:1505.06900v1 [physics.u-
dyn] 26 May 2015, (2015), pp. 131.

[64] O. Malaspinas, B. Chopard, and J. Latt, General regularized bound-

ary condition for multi-speed lattice Boltzmann models, Computers and
Fluids, 49 (2011), pp. 2935.
190 BIBLIOGRAPHY

[65] K. K. Mattila, P. C. Philippi, and L. A. Hegele, High-order reg-

ularization in lattice-Boltzmann equations, Physics of Fluids, 29 (2017),
p. 046103.

[66] K. K. K. Mattila, L. L. A. L. Hegele, and P. P. C. Philippi,

Investigation of an entropic stabilizer for the lattice-Boltzmann method,
Physical Review E, 91 (2015), p. 063010.

[67] J. C. Maxwell, On the dynamical theory of gases, Philosophical trans-

actions of the Royal Society, 157 (1867), pp. 4988.

[68] , Molecules, Nature, (1873), pp. 437441.

[69] G. McNamara and B. Alder, Analysis of the lattice Boltzmann treat-

ment of hydrodynamics, Physica A: Statistical Mechanics and its Applica-
tions, 194 (1993), pp. 218228.

[70] G. R. McNamara and G. Zanetti, Use of the boltzmann equation

to simulate lattice-gas automata, Physical Review Letters, 61 (1988),
pp. 23322335.

[71] S. Mohammed and T. Reis, Using the Lid-Driven Cavity Flow to Vali-
date Moment-Based Boundary Conditions for the Lattice Boltzmann Equa-
tion, Archive of Mechanical Engineering, 64 (2017), pp. 5774.

[72] a. Montessori, G. Falcucci, P. Prestininzi, M. La Rocca, and

S. Succi, Regularized lattice Bhatnagar-Gross-Krook model for two- and
three-dimensional cavity ow simulations, Physical Review E, 89 (2014),
p. 053317.

[73] A. Montessori, P. Prestininzi, M. La Rocca, and S. Succi, Lattice

Boltzmann approach for complex nonequilibrium ows, Physical Review E
- Statistical, Nonlinear, and Soft Matter Physics, 92 (2015), pp. 16.

[74] M. Namburi, P. Kumar, S. Krithivasan, and S. Ansumali, Crystal-

lographic Lattice Boltzmann method for Hydrodynamics, in 24th Discrete
Simulation of Fluid Dynamics (DSFD 2015), Edinburgh, 2015, p. 96.

[75] P. C. Philippi, L. A. Hegele Júnior, L. Dos Santos, and R. Sur-

mas, From the continuous to the lattice Boltzmann equation: The dis-
cretization problem and thermal models, Physical Review E, 73 (2006),
p. 056702.

[76] P. C. Philippi, K. K. Mattila, D. N. Siebert, L. O. E. dos Santos,

L. A. Hegele Júnior, and R. Surmas, Lattice-Boltzmann equations for
describing segregation in non-ideal mixtures, Journal of Fluid Mechanics,
713 (2012), pp. 564587.

[77] M. Planck, Treatise on Thermodynamics, Dover Publications, New York,

1917.
BIBLIOGRAPHY 191

[78] K. N. Premnath and S. Banerjee, On the Three-Dimensional Central

Moment Lattice Boltzmann Method, Journal of Statistical Physics, 143
(2011), pp. 747794.

[79] Y. H. Qian, D. D'Humières, and P. Lallemand, Lattice BGK Mod-

els for Navier-Stokes Equation, Europhysics Letters (EPL), 17 (1992),
pp. 479484.

[80] J. S. Rowlinson, Cohesion: a Scientic History of Intermolecular Forces,

Cambridge University Press, 2004.

[81] X. Shan, General solution of lattices for Cartesian lattice Bhatanagar-

Gross-Krook models, Physical Review E, 81 (2010), p. 036702.

[82] X. Shan, X.-F. Yuan, and H. Chen, Kinetic theory representation of

hydrodynamics: a way beyond the Navier-Stokes equation, Journal of Fluid
Mechanics, 550 (2006), pp. 413441.

[83] P. N. Shankar and M. D. Deshpande, Fluid Mechanics in the Driven

Cavity, Annual Review of Fluid Mechanics, 32 (2000), pp. 93136.

[84] D. Siebert, L. A. Hegele Júnior, R. Surmas, L. dos Santos, and

P. Philippi, Thermal lattice Boltzmann in two dimensions, International
Journal of Modern Physics C, 18 (2007), pp. 546555.

[85] D. N. Siebert, L. A. Hegele, P. C. Philippi, L. A. Hegele Júnior,

and P. C. Philippi, Lattice Boltzmann equation linear stability analysis:
Thermal and athermal models, Physical Review E, 77 (2008), p. 026707.

[86] S. Succi, The Lattice Boltzmann Equation for Fluid Dynamics and Be-
yond (Numerical Mathematics and Scientic Computation) 1st Edition,
Clarendon Press, 2001.

[87] R. Surmas, C. E. Pico Ortiz, and P. C. Philippi, Simulating ther-

mohydrodynamics by nite dierence solutions of the Boltzmann equation,
The European Physical Journal Special Topics, 171 (2009), pp. 8190.

[88] J. D. van der Waals, Over de Continuiteit van den Gas en Vloeistoftoe-
stand, PhD thesis, Leiden University, 1873.

[89] J.-S. S. Wu and Y.-L. L. Shao, Simulation of lid-driven cavity ows

by parallel lattice Boltzmann method using multi-relaxation-time scheme,
International Journal for Numerical Methods in Fluids, 46 (2004), pp. 921
937.

[90] H. Xu and P. Sagaut, Optimal low-dispersion low-dissipation LBM

schemes for computational aeroacoustics, Journal of Computational
Physics, 230 (2011), pp. 53535382.

[91] R. Zhang, X. Shan, and H. Chen, Ecient kinetic method for uid
simulation beyond the Navier-Stokes equation, Physical Review E, 74
(2006), p. 046703.
192 BIBLIOGRAPHY

[92] W. Zhang, C. Zhang, and G. Xi, An explicit Chebyshev pseudospectral

multigrid method for incompressible Navier-Stokes equations, Computers
and Fluids, 39 (2010), pp. 178188.

[93] Q. Zou and X. He, On pressure and velocity boundary conditions for the
lattice Boltzmann BGK model, Physics of Fluids, 9 (1997), pp. 15911598.

[94] , On pressure and velocity ow boundary conditions and bounceback

for the lattice Boltzmann BGK model, Physics of Fluids, 1591 (1997), p. 18.