www.Ebook777.

com
 Collection Alea-Saclay:
Monographs and Texts in Statistical Physics

General editor : Claude Godn!che

C. Godreche ( ed.) : Solids Far from Equilibrium

P. Peretto : An Introduction to the Modeling of Neural Networks
C. Godreche and P. Manneville (eds) : Hydrodynamics and Nonlinear Instabilities
A Pimpinelli and J. Villain : The Physics of Crystal Growth
D. H. Rothman and S. Zaleski : Lattice-gas Cellular Automata
B. Chopard and M. Droz : Cellular Automata Modeling of Physical Systems

www.Ebook777.com
 Cellular Automata
Modeling of Physical
Systems

Bastien Chopard and Michel Droz

University of Geneva

. . .�! CAMBRIDGE
;:; UNIVERSITY PRESS

www.Ebook777.com
PUBLISHED BY THE PRESS SYNDICATE OF THE UNIVERSITY OF CAMBRIDGE

The Pitt Building, Trumpington Street, Cambridge CB2 1 RP, United Kingdom

CAMBRIDGE UNIVERSITY PRESS

The Edinburgh Building, Cambridge CB2 2RU, UK http:/ /www.cup.cam.ac.uk

40 West 20th Street, New York, NY 1001 1 -42 1 1 , USA http:/ /www.cup.org
10 Stamford Road, Oakleigh, Melbourne 3 1 66, Australia

© Bastien Chopard and Michel Droz 1998

This book is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without
the written permission of Cambridge University Press.

First published 1998

Printed in the United Kingdom at the University Press, Cambridge

Typeset in 1 1 pt Times [EPC]

A catalogue record for this book is available from the British Library

Library of Congress Cataloguing in Publication data

Chopard, Bastien, 1959-

Cellular automata modeling of physical systems / Bastien Chopard and Michel Droz.
p. em. - (Collection Alea-Saclay)
Includes bibliographical references and index.
ISBN 0 521 46 168 5 (he: alk. paper)
1. Statistical physics. 2. Cellular automata. 3. Diffusion.
4. Lattice gas. 5. Phase transformations (Statistical physics)
I. Droz, M. (Michel), 1944- . II. Title. III. Series.
QC174.8.C49 1998
530'.1 3'01'1-dc21 97-28284 CIP

ISBN 0 521 46168 5 hardback

www.Ebook777.com
 To Viviane
Clea, Tristan and Daphne

and Dominique
Anne-Virginie and Philippe
 This book provides a self-contained introduction to cellular automata and lattice
Boltzmann techniques.
Beginning with a chapter introducing the basic concepts of this developing
field, a second chapter describes methods used in cellular automata modeling.
Following chapters discuss the statistical mechanics of lattice gases, diffusion
phenomena, reaction-diffusion processes and nonequilibrium phase transitions. A
final chapter looks at other models and applications, such as wave propagation
· and multiparticle fluids. With a pedagogic approach, the volume focuses on the
use of cellular automata in the framework of equilibrium and nonequilibrium
statistical physics. It also emphasizes application-oriented problems such as
fluid dynamics and pattern formation. The book contains many examples and
problems. A glossary and a detailed list of references are also included.
This will be a valuable book for graduate students and researchers working in
statistical physics, solid state physics, chemical physics and computer science.
BASTIEN CHOPARD obtained a Masters degree and a PhD in theoretical physics
from the University of Geneva. After three years as a research associate at MIT
(Cambridge, Mass.) and in the Forschungzentrum in Jtilich (Germany), he is now
assistant professor in the Department of Computer Sciences (Parallel Computing
Group) of the University of Geneva. His research interests concern the modeling
and simulation of complex systems and parallel computing.
MICHEL D Roz obtained a Masters degree and a PhD in theoretical physics
from the University of Geneva. He spent three years as research associate at
Cornell University (Ithaca, NY) and Temple University (Phihi.delphia, Penn.). He
is now at the Department of Theoretical Physics of the University of Geneva
where his main fields of research are equilibrium and non-equilibrium statistical
mechanics, complex systems and cellular-automata modeling of physical systems.
Everything should be made as simple as possible but not simpler
A. Einstein
 Contents

Preface Xl

1 Introduction 1
1.1 Brief history 1
1 . 1 . 1 Self-reproducing systems 1
1 . 1 .2 Simple dynamical systems 3
1 . 1 .3 A synthetic universe 4
1 . 1 .4 Modeling physical systems 5
1 . 1 . 5 Beyond the cellular automata dynamics : lattice Boltzmann
methods and multiparticle models 7
1.2 A simple cellular automaton : the parity rule 8
1.3 Definitions 12
1 . 3 . 1 Cellular automata 12
1 .3.2 Neighborhood 14
1 .3.3 Boundary conditions 15
1 . 3.4 Some remarks 16
1 .4 Problems 18

2 Cellular automata modeling 21

2. 1 Why cellular automata are useful in physics 21
2. 1 . 1 Cellular automata as simple dynamical systems 21
2. 1 .2 Cellular automata as spatially extended systems 24
2. 1 . 3 Several levels of reality 26
2. 1 .4 A fictitious microscopic world 27
2.2 Modeling of simple systems : a sampler of rules 28
2.2. 1 The rule 1 84 as a model for surface growth 29
2.2.2 Probabilistic cellular automata rules 29
2.2.3 The Q2R rule 33
2.2.4 The annealing rule 37
2.2.5 The HPP rule 38
2.2.6 The sand pile rule 42

Vll
V111 Contents

2.2.7 The ant rule 46

2.2.8 The road traffic rule 51
2.2.9 The solid body motion rule 56
2.3 Problems 61

3 Statistical mechanics of lattice gas 66

3.1 The one-dimensional diffusion automaton 66
3. 1 . 1 A random walk automaton 67
3 . 1 .2 The macroscopic limit 68
3 . 1 . 3 The Chapman-Enskog expansion 71
3. 1 .4 Spurious invariants 74
3.2 The FHP model 75
3.2. 1 The collision rule 75
3.2.2 The microdynamics 78
3.2.3 From microdynamics to macrodynamics 80
3.2.4 The collision matrix and semi-detailed balance 105
3.2.5 The FHP-III model 107
3.2.6 Examples of fluid flows 1 10
3.2.7 Three-dimensional lattice gas models 1 12
3.3 Thermal lattice gas automata 1 12
3.3. 1 Multispeed models 1 12
3.3.2 Thermo-hydrodynamical equations 115
3.3.3 Thermal FHP lattice gases 116
3.4 The staggered invariants 1 17
3.5 Lattice Boltzmann models 1 22
3.5. 1 Introduction 1 22
3.5.2 A simple two-dimensional lattice Boltzmann fluid 125
3.5.3 Lattice Boltzmann flows 1 34
3.6 Problems 135

4 Diffusion phenomena 138

4. 1 Introduction 138
4.2 The diffusion model 139
4.2. 1 Microdynamics of the diffusion process 140
4.2.2 The mean square displacement and the Green-Kubo formula 147
4.2.3 The three-dimensional case 1 49
4.3 Finite systems 1 50
4.3 . 1 The stationary source-sink problem 151
4.3.2 Telegraphist equation 153
4.3.3 The discrete Boltzmann equation in 2D 1 57
4.3.4 Semi-infinite strip 1 59
4.4 Applications of the diffusion rule 1 63
4.4. 1 Study of the diffusion front 1 63
4.4.2 Diffusion-limited aggregation 1 66
4.4.3 Diffusion-limited surface adsorption
4.5
171
Problems 175
 Contents IX

5 Reaction-diffusion processes 178

5. 1 Introduction 178
5.2 A model for excitable media 179
5.3 Lattice gas microdynamics 181
5.3. 1 From microdynamics to rate equations 1 84
5.4 Anomalous kinetics 1 87
5.4. 1 The homogeneous A+ B --+ f/J process 188
5.4.2 Cellular automata o r lattice Boltzmann modeling 190
5.4.3 Simulation results 191
5.5 Reaction front in the A+ B --+ 0 process 193
5.5. 1 The scaling solution 195
5.6 Liesegang patterns 198
5.6. 1 What are Liesegang patterns 198
5.6.2 The lattice gas automata model 201
5.6.3 Cellular automata bands and rings 202
5.6.4 The lattice Boltzmann model 206
5.6.5 Lattice Boltzmann rings and spirals 208
5.7 Multiparticle models 209
5.7. 1 Multiparticle diffusion model 210
5.7.2 Numerical implementation 212
5.7.3 The reaction algorithm 214
5.7.4 Rate equation approximation 216
5.7.5 Turing patterns 217
5.8 From cellular automata to field theory 219
5.9 Problems 228

6 Nonequilibrium phase transitions 232

6. 1 Introduction 232
6.2 Simple interacting particle systems 234
6.2. 1 The A model 234
6.2.2 The contact process model (CPM) 241
6.3 Simple models of catalytic surfaces 242
6.3 . 1 The Ziff model 242
6.3.2 More complicated models 248
6.4 Critical behavior 249
6.4. 1 Localization of the critical point 250
6.4.2 Critical exponents and universality classes 252
6.5 Problems 254

7 Other models and applications 256

7. 1 Wave propagation 256
7. 1 . 1 One-dimensional waves 256
7. 1 .2 Two-dimensional waves 258
7. 1 . 3 The lattice BGK formulation o f the wave model 267
7. 1 .4 An application to wave propagation in urban environments 278
7.2 Wetting, spreading and two-phase fluids 280
X Contents

7.2. 1 Multiphase flows 280

7.2.2 The problem of wetting 282
7.2.3 An FHP model with surface tension 284
7.2.4 Mapping of the hexagonal lattice on a square lattice 287
7.2.5 Simulations of wetting phenomena 289
7.2.6 Another force rule 292
7.2.7 An Ising cellular automata fluid 293
7.3 Multiparticle fluids 298
7.3 . 1 The multiparticle collision rule 299
7.3.2 Multiparticle fluid simulations 302
7.4 Modeling snow transport by wind 305
7.4. 1 The wind model 305
7.4.2 The snow model 308
7.4.3 Simulations of snow transport 312

References 313

Glossary 327

Index 337
 Preface

The cellular automata approach and the related modeling techniques are
powerful methods to describe, understand and simulate the behavior of
complex systems. The aim of this book is to provide a pedagogical
and self-contained introduction to this field and also to introduce recent
developments. Our main goal .is to present the fundamental theoretical
concepts necessary for a researcher to address advanced applications in
physics and other scientific areas.
In particular, this book discusses the use of cellular automata in the
framework of equilibrium and nonequilibrium statistical physics and in
application-oriented problems. The basic ideas and concepts are illustrated
on simple examples so as to highlight the method. A selected bibliog­
raphy is provided in order to guide the reader through this expanding
field.
Several relevant domains of application have been mentioned only
through references to the bibliography, or are treated superficially. This
is not because we feel these topics are less important but, rather, because
a somewhat subjective selection was necessary according to the scope
of the book. Nevertheless, we think that the topics we have covered
are significant enough to give a fair idea of how the cellular automata
technique may be applied to other systems.
This book is written for researchers and students working in statistical
physics, solid state physics, chemical physics and computer science, and
anyone interested in modeling complex systems. A glossary is included
to give a definition of several technical terms that are frequently used
throughout the text. At the end of the first six chapters, a selection
of problems is given. These problems will help the reader to become
familiar with the concepts introduced in the corresponding chapter, or will
introduce him to new topics that have not been covered in the text. Some
problems are rather easy, although they usually require some programming

Xl
xii Preface

effort, but other problems are more involved and will demand significant
time to complete.
Most of the cellular automata simulations and results presented in this
book have been produced on the 8k Connection Machine CM-200 of
the University of Geneva. Others have been computed on an IBM SP2
parallel computer, also installed at the University of Geneva. Although
a parallel supercomputer is quite useful when considering large scale
simulations, common workstations and even modern personal computers
are well adapted to perform cellular automata computations, except for
on-line display which is alway very desirable. Dedicated hardware is also
available but, usually, less flexible than a general purpose machine.
Despite our effort, several errors and misprints are still likely to be
present. Please report them to us* (as well as any comment or suggestion).
We would like to thank all the people who have made this book
possible and, in particular Claude Godreche who gave us the opportunity
to write it. Special thanks go to Pascal Luthi and Alexandre Masselot
who made several original and important simulations which are presented
in this book. Other people have played a direct or indirect role in
the preparation of the manuscript. Among them, we thank Rodolphe
Chatagny, Stephen Cornell, Laurent Frachebourg, Alan McKane, Zoltan
Racz and Pierre-Antoine Rey.
Finally we acknowledge the Swiss National Science Foundation who
funded some of the research reported here and the Computer Science
Department and the Theoretical Physics Department of the University of
Geneva for making available the necessary environment and infrastructure
for this enterprise.

• Bastien. Chopard\Dcui. unige. ch or Michel. Droz\Dphysics. unige. ch

 1
Introduction

1.1 Brief history

Cellular automata (often termed CA) are an idealization of a physical

system in which space and time are discrete, and the physical quantities
take only a finite set of values.
Although cellular automata have been reinvented several times (of­
ten under different names), the concept of a cellular automaton dates
back from the late 1 940s. During the following fifty years of exis­
tence, cellular automata have been developed and used in many different
fields. A vast body of literature is related to these topics. Many con­
ference proceedings [1-8] ), special journal issues [9, 10] and articles are
available.
In this section, our purpose is not to present a detailed history of the
developments of the cellular automat 1 approach but, rather, to emphasize
some of the important steps.

1.1.1 Self-reproducing systems

The reasons that have led to the elaboration of cellular automata are
very ambitious and still very present. The pioneer is certainly John von
Neumann who, at the end of the 1940s, was involved in the design of
the first digital computers. Although von Neumann's name is definitely
associated with the architecture of today's sequential computers, his con­
cept of cellular automata constitutes also the first applicable model of
massively parallel computation.
Von Neumann was thinking of imitating the behavior of a human brain
in order to build a machine able to solve very complex problems. However,
his motivation was more ambitious than just a performance increase of
the computers of that time. He thought that a machine with such a

1
2 1 Introduction

complexity as the brain should also contain self-control and self-repair

mechanisms. His idea was to get rid of the difference which exists between
processors and the data, by considering them on the same footing. This
led him to envisage a machine capable of building itself, out of some
available material.
Rapidly, he considered the problem from a more formal viewpoint and
tried to define the properties a system should have to be self-replicating. He
was mostly interested to find a logical abstraction of the self-reproduction
mechanism, without reference to the biological processes involved.
Following the suggestions of S. Ulam [ 1 1 ] , von Neumann addressed this
question in the framework of a fully discrete universe made up of cells.
Each cell is characterized by an internal state, which typically consists of
a finite number of information bits. Von Neumann suggested that this
system of cells evolves, in discrete time steps, like simple automata which
only know of a simple recipe to compute their new internal state. The rule
determining the evolution of this system is the same for all cells and is a
function of the states of the neighbor cells. Similarly to what happens in
any biological system, the activity of the cells takes place simultaneously.
However, the same clock drives the evolution of each cell and the updating
of the internal state of each cell occurs synchronously. These fully discrete
dynamical systems (cellular space) invented by von Neumann are now
referred to as cellular automata.
The first self-replicating cellular automaton proposed by von Neumann
was composed of a two-dimensional square lattice and the self-reproducing
structure was made up of several thousand elementary cells. Each of these
cells had up to 29 possible states [12] . The evolution rule required the
state of each cell plus its four nearest neighbors, located north, south, west
and east. Due to its complexity, the von Neumann rule has only been
partially implemented on a computer [13] .
However, von Neumann had succeeded in finding a discrete structure of
cells bearing in themselves the recipe to generate new identical individuals.
Although this result is hardly even a very primitive form of life, it is
quite interesting because it is usually expected that a machine can only
build an object of lesser complexity than itself. With self-replicating
cellular automata, one obtains a "machine" able to create new machines
of identical complexity and capabilities.
The von Neumann rule has the so-called property of universal com­
putation. This means that there exists an initial configuration of the
cellular automaton which leads to the solution of any computer al­
gorithm. This sounds a surprising statement : how will such a discrete
dynamics help us to solve any problem ? It turns out that this property
is of theoretical rather than practical interest. Indeed, the property of
universal computing means that any computer circuit (logical gates) can
 1 . 1 Brief history 3

be simulated by the rule of the automaton. All this shows that quite
complex and unexpected behavior can emerge from a cellular automaton
rule.
After the work of von Neumann, others have followed the same
line of research and the problem is still of interest [14] . In particular,
E.F. Codd [15] in 1 968 and much later C.G. Langton [16] and Byl [17]
proposed much simpler cellular automata rules capable of self-replicating
and using only eight states. This simplification was made possible by giv­
ing up the property of computational universality, while still conserving
the idea of having a spatially distributed sequence of instructions (a kind
of cellular DNA) which is executed to create a new structure and then
entirely copied in this new structure.
More generally, artificial life is currently a domain which is intensively
studied. Its purpose is to better understand real life and the behavior of
living species through computer models. Cellular automata have been an
early attempt in this direction and can certainly be further exploited to
progress in this field [18, 19] .

1 . 1 .2 Simple dynamical systems

In a related framework, it is interesting to remember that it is precisely a
simple ecological model that has brought the concept of cellular automata
to the attention of wide audience. In 1 970, the mathematician John
Conway proposed his now famous game of life [20] . His motivation was
to find a simple rule leading to complex behaviors. He imagined a two­
dimensional square lattice, like a checkerboard, in which each cell can be
either alive (state one) or dead (state zero). The updating rule of the game
of life is as follows : a dead cell surrounded by exactly three living cells
comes back to life. On the other hand, a living cell surrounded by less than
two or more than three neighbors dies of isolation or overcrowdness. Here,
the surrounding cells correspond to the neighborhood composed of the
four nearest cells (north, south, east and west) plus the four second nearest
neighbors, along the diagonals. Figure 1 . 1 shows three configurations of
the game of life automaton, separated by 10 iterations.
It turned out that the game of life automaton has an unexpectedly
rich behavior. Complex structures emerge out of a primitive "soup" and
evolve so as to develop some skills. For instance, objects called gliders
may form (see problems, section 1 .4). Gliders correspond to a particular
arrangement of adjacent cells that has the property to move across space,
along straight trajectories. Many more such structures have been identified
in the vast body of literature devoted to the game of life [21,22] . As for
the von Neumann rule, the game of life is a cellular automata capable of
computational universality.
4 1 Introduction

. - •• "(t
· ·
0(10 ·-·- o0o · -·- . - •

0
I I •

0 l""'b-
• • I
..... '"I:.)
...

.) ..........
"'=
. . -·-

..... �

t+lO t+20

Fig. 1 . 1 . The game of life automaton. Black dots represents living cells whereas
dead cells are white. The figure shows the evolution of some random initial
configurations.

In addition to these theoretical aspects, cellular automata were used in

the 1 950s for image processing [23] . It was recognized early on that much
tedious picture analysis could be carried out automatically, according to a
cellular automata computing model : the pixels of an image can be treated
simultaneously, using simple local operations. Special-purpose machines
based on cellular automata logic have been developed for noise reduction,
counting and size estimation in images obtained from observations with
a microscope.
At the beginning of the 1 980s, S. Wolfram studied in detail a family of
simple one-dimensional cellular automata rules (the now famous Wolfram
rules [24,25] ). He had noticed that a cellular automaton is a discrete dy­
namical system and, as such, exhibits many of the behaviors encountered
in a continuous system, yet in a much simpler framework. A concept such
as complexity could be investigated on mathematical models allowing an
exact numerical computer calculation, because of their Boolean nature (no
numerical errors nor truncation as in more traditional models). Wolfram's
results have contributed to prove that cellular automata are important
objects to consider for statistical mechanics studies and, at the present
time, Wolfram's rule are still the topic of much research.

1.1.3 A synthetic universe

The property of many cellular automata rules being a universal computer
made several authors think that the physical world itself could be a
very large cellular automaton. Tommaso Toffoli [26] compares cellular
automata to a synthetic model of the universe in which the physical laws
are expressed in terms of simple local rules on a discrete space-time
structure.
 1.1 Brief history 5

T. Toffoli, N. H. Margolus and E. Fredkin recognized the importance

of cellular automata as a modeling environment for physical systems.
They were very interested in the analogy that exists between the theory of
information as it is used to describe numerical processing in a computer
and the laws of physics. Cellular automata provide an excellent framework
to develop these ideas. In particular, they showed how to build a fully time­
reversible logic from which any numerical operation can be implemented
without any loss of information. The so-called billiard ball [26] is a
cellular automata rule which is an example of such a reversible model of
computation.
The possibility of displaying, on a computer screen, the time evolution
of large cellular automata systems, at the rate of several updates per
second of the complete lattice offers a way of performing experiments live
on an artificial universe, whose evolution rules are set up by the observer.
By building their first general purpose cellular automata machines CAM-6
in the mid- 1980s, Toffoli and Margolus provided a very powerful cellular
automata environment with the capability of a supercomputer of that time,
at a very affordable price and with unique display facilities. This machine
has stimulated many developments of cellular automata techniques and
has contributed to the spreading of the main ideas to a wide audience of
scientists.
Toffoli and Margolus's book [26] : Cellular Automata Machines : a New
Environment for Modeling, is a wonderful source of inspiration in the field
of cellular automata and provide a complete description of the CAM-6
hardware. More recently, Toffoli, Margolus and coworkers have designed
CAM-8, a much more powerful hardware environment : a parallel, uni­
form, scalable architecture for cellular automata experimentation [27] .
This hardware platform offers high performance, a flexible approach, dis­
play facilities and is naturally appropriate to work on three-dimensional
systems. It has been successfully used for many different applications.

1 . 1 .4 Modeling physical systems

It was also in the 1 980s that an important step in the theory of cellular
automata was accomplished. It was recognized that the so-called HPP [28]
lattice gas models developed in the 1 970s by Hardy, Pomeau and de Pazzis
was in fact a cellular automata. This model consists of a simple and fully
discrete dynamics of particles moving and colliding on a two-dimensional
square lattice, in a such a way as to conserve momentum and particle
number.
The HPP dynamics was initially planned as a theoretical model to
study fundamental statistical properties of a gas of interacting particles.
The actual implementation of this model as a cellular automata rule and
6 1 Introduction

the visualization of the fast moving particle shed a different light on the
possibilities of such models : isn't it possible to simulate the behavior of a
real system of particles (like a fluid or a gas) as a cellular automata rule ?
After all, it is well known that the flows of a fluid, a gas or even a granular
medium are very similar at a macroscopic scale, in spite of their different
microscopic nature. A fully discrete and simplified molecular dynamics
could work too, provided the system is considered at an appropriate
observation scale.
Of course, the idea of using discrete systems as a model of real phenom­
ena has already been considered for several problems. The Ising model
of classical spin is a famous example which will be discussed in more
detail in the next chapter. From the fluid side, already at the end of the
nineteenth century, Maxwell [29] , had proposed a discrete velocity system
of interacting particles as a model of a gas. In fact, such lattice gas,
discrete velocity models have been developed independently from cellular
automata theory [30,3 1] .
However, cellular automata provide a new conceptual framework, as
well as an effective numerical tool, which retains important aspects of the
microscopic laws of physics, such as simultaneity of the motion, locality
of the interactions and time reversibility.
Cellular automata rules are viewed as an alternative form of the
microscopic reality which bears the expected macroscopic behavior. From
a numerical point of view it was expected, at the end of the 1 980s, that
a wind tunnel could be replaced by a fully discrete computer model.
The first cellular automata model to give credit to this possibility is the
famous FHP model proposed in 1986 by U. Frisch, B. Hasslacher and
Y. Pomeau [32] , and almost simultaneously by S. Wolfram [33] . These
authors showed that their model, despite its fully discrete dynamics, fol­
lows, in some appropriate limits, the behavior prescribed by the Navier­
Stokes equation of hydrodynamics.
Note that models like FHP or HPP are often termed lattice gas auto­
mata (LOA) to distinguish them from the less specific cellular automata
terminology. Clearly, from a mathematical point of view, a lattice gas
automata is a cellular automata, but the way one thinks, for instance,
of the game of life is quite different from the underlying philosophy
of the FHP model. This difference will become clear to the reader
as he or she becomes more familiar with the next chapter of this book.
Nevertheless, in this book, we will often use cellular automata to designate
a LGA.
Since the FHP rule was discovered, lattice gas automata or cellular
automata fluids as these kind of particle models are now often referred to,
have been developed intensively and several insufficiencies of the initial
model corrected. The Ecole Normale Superieure in Paris has been very
 1 . 1 Brief history 7

active and P. Lallemand and D. d'Humieres, in particular, have played a

pioneering role in this field [34-3 7] .
However, contrary to first expectations, lattice gas models of fluids
have not been able to surpass the traditional numerical methods of
hydrodynamics and compute high Reynolds flows. Their relatively high
viscosity, which is only determined by the cellular automata rule (and
therefore not adjustable), is a limiting factor to the practical study of
many of these flows. The finite spatial resolution of the cellular automata
lattice (physical phenomena must occur at a much larger scale than the
lattice spacing) is another limitation on the study and modeling of fully
developed turbulence, unless the system has such a large scale that the ad­
vantage of a cellular automata approach vanishes even on today's fastest
computers [38] .
However, lattice gas automata have been much more successful in
modeling complex situations for which traditional computing techniques
are not applicable. Flows in porous media [39-41], immiscible [42-
46] flows and instabilities, spreading of a liquid droplet and wetting
phenomena [47] , microemulsion [48] erosion and transport problems [49]
are some examples pertaining to fluid dynamics.
Other physical situations, like pattern formation, reaction-diffusion pro­
cesses [50] , nucleation-aggregation growth phenomena, are very well de­
scribed by cellular automata dynamics and will be investigated in detail
in this book.

1 . 1.5 Beyond the cellular automata dynamics : lattice Boltzmann methods

and multiparticle models
Very often, the advantage of the cellular automata (or lattice gas) approach
is most apparent when complex boundary conditions are present. Due
to the microscopic interpretation of the dynamics, these conditions can
be taken into account in a much more natural way than in a continuous
description (like a differential equation) in which our basic intuition of
the phenomena may be lost.
On the other hand, cellular automata models have several weakness
related to their fully discrete nature : statistical noise requiring systematic
averaging processes, and little flexibility to adjust parameters of a rule
in order to describe a wider range of physical situations. At the end
of the 1980s, McNamara and Zanetti [5 1 ] , and Higueras, Jimenez and
Succi [52] showed the advantage of extending the Boolean dynamics of the
automaton to directly work on real numbers representing the probability
for a cell to have a given state.
This approach, called the lattice Boltzmann method (LBM), is numer­
ically much more efficient than the Boolean dynamics and provides a
8 1 Introduction

new computational model much more suited to the simulation of high

Reynolds flows and many other relevant applications (for instance glacier
flow [53] ).
Lattice Boltzmann models retain the microscopic level of interpreta­
tion of the cellular automata approach but neglect many-body correlation
functions. However, this method now constitutes a very promising ap­
proach to modeling physical systems and is discussed on several occasions
throughout this book.
In between the strict cellular automata approach and the more flexible
lattice Boltzmann method, there is room for an intermediate description :
the multiparticle models which are still under development at the present
time. These models preserve the concept of a quantized state but an
infinite set of values is accepted. Consequently, numerical stability is
guaranteed (as opposed to the LBM), and many-body correlations taken
into account. The large number of possible states offers more flexibility
when modeling a physical system and yields less statistical noise. But a
multiparticle dynamics is more difficult to devise and numerically slower
than its lattice Boltzmann counterpart. Examples of this approach will be
presented in this book.
From our point of view, the cellular automata approach is not a
rigid framework. It is rather a philosophy of modeling which should
be considered with some pragmatism. The important issue in cellular
automata modeling is to capture the essential features of given phenomena
and translate them to a suitable form to obtain an effective numerical
model. To this end, it is acceptable (and even beneficial) to relax some
of the constraints of the original definition of a cellular automata. The
introduction of the lattice Boltzmann method is an illustration of this fact.
The point is to conserve the spirit of the approach and its relevant features
rather than its limitations. This remark is particularly in order because
present parallel computers offer an ideal and quite flexible platform to
implement cellular automata models without the restrictions imposed by
dedicated hardware.

1.2 A simple cellular automaton: the parity rule

In this section, we discuss a simple cellular automata rule, in order to

introduce and illustrate the concept. This should slowly familiarize the
reader with a more precise notion of cellular automata. Section 1 .3 will
present a more formal definition.
Although it is very basic, the rule we study here exhibits a surpris­
ingly rich behavior. It was proposed initially by Edward Fredkin in the
1 970s [54] and is defined on a two-dimensional square lattice.
 1 .2 A simple cellular automaton : the parity rule 9

( a) (b) (c )

Fig. 1 .2. The E!3 rule on a 256 x 256 periodic lattice : (a) initial configuration ; (b)
and (c) configurations after tb = 93 and tc = 1 10 iterations, respectively.

Each site of the lattice is a cell which is labeled by its position r = (i, j )
where i and j are the row and column indices. A function tpt (r ) is
associated to the lattice to describe the state of each cell at iteration t.
This quantity can be either 0 or 1 .
The cellular automata rule specifies how the states tpt+ l are to be
computed from the states at iteration t. We start from an initial condition
at time t = 0 with a given configuration of the values tpo (r ) on the lattice.
The state at time t = 1 will be obtained as follows

( 1 ) Each site r computes the sum of the values tpo (r ' ) on the four
nearest neighbor sites r ' at north, west, south and east. The system
is supposed to be periodic in both i and j directions (as on a torus)
so that this calculation is well defined for all sites.
(2) If this sum is even, the new state tp1 (r) is 0 (white), else, it is 1 (black).
The same rule (step 1 and 2) is repeated to find the states at time
t = 2, 3, 4, ... .
From a mathematical point of view, this cellular automata parity rule
can be expressed by the following relation
lpt+ l (i, j ) = tpt (i + 1, j ) Etl tpt (i - 1, j) Etl tpt (i, j + 1 ) Etl tpt (i, j - 1 ) ( 1. 1 )
where the symbol Etl stands for the exclusive O R logical operation. I t is
also the sum modulo 2 : 1 Etl 1 = 0 EB 0 = 0 and 1 Etl 0 = 0 EB 1 = 1 .
When this rule is iterated, very nice geometrical patterns are observed,
as shown in figure 1 . 2 . This property of generating complex patterns
starting from a simple rule is actually generic of many cellular automata
rules. Here, complexity results from some spatial organization which
builds up as the rule is iterated. The various contributions of successive
iterations combine together in a specific way. The spatial patterns that
are observed reflect how the terms are combined algebraically.
10 1 Introduction

Cellular

Automata

( a) (b) (c)

Fig. 1.3. The EB rule replicates any initial pattern when the number of iterations
is a power of two. Image (a) shows the initial pattern at time ta = 0. Images (b)
and (c) show successive iterations at times tb = 16 and tc = tb + 32.

On closer inspection, we also observe that the initial pattern is replicated

at some specific iteration. Figure 1.3 illustrates that point with a more
enlightening initial condition. The times at which this happens are a power
of two. Another surprising fact occurs when the system size L is a power
of two : after L/2 iteration the state of each cell vanishes for all possible
initial configurations.
These behaviors, as well as the way the pattern builds up can be
explained by working out the definition of the rule. Applying the rule 1 . 1
twice yields V't+ l as a function of V't- 1
V't+t(i, j ) V't-t (i + 2, j ) EB V't-t(i, j ) EB V't-t (i + 1, j + 1 )
=

EBV't-t( i + 1, j - 1 ) EB V't-t (i, j ) EB V't-t (i - 2, j )

EBV't-t( i - 1, j + 1 ) EB V't-t( i - 1, j - 1 ) EB V't-t( i + 1, j + 1 )
EBV't-t( i - 1, j + 1 ) EB V't-t(i, j + 2 ) EB V't-t( i, j )
EBV't-t (i + 1, j - 1 ) EB V't-t( i - 1, j - 1 ) EB V't-l (i, j )
EBV't-t( i, j - 2) ( 1 .2)
Since a EB a = 0 and a EB 0 = a, for all values of a, one obtains
V't+2 (i, j ) =
V't (i + 2, j ) EB V't (i - 2, j ) EB lf't (i, j + 2) EB lf't (i, j - 2) ( 1 .3 )
Thus, after two iterations, the action of the rule is to translate the initial
configuration by two lattice sites in the four directions and XOR them.
If now we compute similarly 1f't+ 3 (i, j ) as a function of V't (i, j ), we no
longer get such a simple expression. A relation similar to equation 1 . 2
is obtained, but without any cancellation. We see that 1f't+ 3 (i, j) is a
superposition of 1 6 translations. As a result, we obtain the rich geometrical
structure observed in figure 1 . 2 .
However, it is easy to prove that the behavior of the EB rule is simple
when the number of iterations performed is a power of two. To show this
 1 .2 A simple cellular automaton : the parity rule 11

property, suppose the following relation is true for a given value of T.

1JJr(i, j) = tpt-r(i+ T, j) EFYtpt-r(i - T, j) EFY1/Jr-r(i, j+ T) EFYtpt-r(i, j- T) ( 1.4)
We already know it is true for T = 1 and T = 2. Then, if we apply this
same relation to tpt-T on the right-hand side, we obtain
tpt(i, j) = lpt-2r(i + 2 T, j) EfJ lpt-2r(i, j) EFY lpt-2r(i + T, j + T)
EFYtpt-2r(i + T, j - T) EfJ tpt-2r(i, j ) EfJ tpt-2r(i - 2 T, j)
EFYtpt-2r(i - T, j + T) EfJ tpt-2r(i - T, j - T)
EfYtpt-2r(i + T, j + T) EfJ tpt-2r(i - T , j + T) EfJ tpt-2r(i, j + 2 T)
EfYtpt-2r(i, j) EfJ tpt-2r(i + T, j - T) EfJ tpt-2r(i - T, j - T)
EFYtpt-2r(i, j ) EFY lpt-2r(i, j - 2 T)
= lpt-2r(i + 2 T, j ) EFY lpt-2r(i - 2 T, j ) EFY lpt-2r(i, j + 2 T)
EFYtpt-2r(i, j - 2 T) ( 1.5)
This result shows that property 1 .4 is then also true for 2 T . Therefore,
it is true for any lag T which is a power of two. At these particular
values of time, the EFJ rule is equivalent to the superposition (in the sense
of the addition modulo two) of the initial pattern translated in the four
lattice directions by an amount T . When the spatial extension of the
initial pattern is small enough, we see it replicated four times. Otherwise
destructive interferences show up, which give rise to partial replication.
For a square lattice of size L = 2k , relation 1 .4 implies that, after L/2
iterations, the result of the EFJ rule is to superpose the initial condition
with itself and, therefore, to yield a zero configuration.
For a number of iterations that is not a power of two, some results
can also be obtained. In general, the configuration after T steps is the
superposition modulo 2 of 4k different translations of the initial pattern,
where k is the number of digits equal to 1 in the binary representation of
T.
This property can be proved as follows. First, we notice that the
rule is additive, that is any initial pattern can be decomposed into the
superposition of one-pixel images. Each of these simple configurations
can be evolved independently and the results superposed to obtained the
final complete pattern.
We will prove our statement for an initial configuration such that
tp 0) = 1, only. It is convenient to decompose T as a sum of powers of
o(O,
two. There exists a value of n such that we can write T as
n- 1
T = 2 n + L a t 2t = 2 n + T ' ( 1 .6 )
t=O
where, by construction, T ' ::;: 2 n - 1 (equality holds when all the at = 1 ) .
12 1 Introduction
Now, performing T iterations of the rule is equivalent to first do T' steps
and then the last 2 n iterations. Clearly, by the definition of the rule, tpT'(i, j)
will contain the terms tpo(i - T', j), tpo(i + T', j), tpo(i, j - T') and tpo(i, j +
T') and other terms involving only translations by a smaller amount.
Therefore, the configuration at time T' has a given spatial extension.
When performing the next 2 n iterations, we know from 1.4 that they will
result in the superposition of four translations by 2 n of the configuration
at time T'. None of these translations will give rise to cancellation
because the spatial extension of the configuration at time T' is smaller
than the translation length. Indeed, after the left translation by 2 n , the
right boundary will move to position T' - 2 n with respect to the original
pattern. Similarly, the left boundary will move to - T' + 2 n , due to the
right translation. There is no overlap between the patterns generated in
this way because, since T' < 2 n , one has
T' - 2 n � - T ' + 2n ( 1. 7)
Therefore, for each non-zero at in expression (1.6), four translations
are produced and the final result is composed of 4k non-over lapping
translations, with k = 2: at. When the initial image is not a single pixel,
destructive interference is observed.
As a result, we get the rich geometrical structure observed in figure 1.2.
As the number of iterations increases more and more terms are generated.
Therefore, the algorithmic complexity of the expression becomes larger,
thus reflecting the complexity of the cellular automaton configuration.
More precisely, k is bounded by the logarithm of T. In order to evaluate
the asymptotic complexity of the expression, we write T 2k . The number
"'

of translations generated is 4k = T 2 and the complexity goes as the square

of the number of iterations.
The above discussion has unraveled the mechanisms leading to the com­
plex structures produced by the E9 rule as being due to the superposition
of the initial pattern translated many times by a different amount. From
this analysis, we can conclude that this rule is not a self-replicating cellular
automaton in the sense imagined by von Neumann.
Finally, it is interesting to note that the E9 rule is a generalization in two
dimensions of the famous rule 90 of Wolfram which will be discusseded
in section 2. 1.1.

1.3 Definitions

1 .3.1 Cellular automata

In this section we shall present a more formal definition of a cellular
automaton. In general, a cellular automaton requires
 1 . 3 Definitions 13

(i) a regular lattice o f cells covering a portion of a d-dimensional space ;

(ii) a set <l> (r, t) = { <1>1 (r, t ), <1>2(r, t), ... , <I>m (r, t ) } of Boolean variables
attached to each site r of the lattice and giving the local state
of each cell at the time t = 0, 1, 2, ... ;

(iii) a rule R = {Rt, R2, ... , Rm} which specifies the time evolution of the
states <l> (r, t ) in the following way
<I>j(r, t + 1 ) = Rj(<I>(r , t), <I>(r + J 1 . t), <I>(r + J2, t ), ... , <I>(r + bq, t )) ( 1 .8)
where r + Jk designate the cells belonging to a given neighborhood
of cell r.
The example discussed in the previous section is a particular case in
which the state of each cell consists of a single bit <1> 1 (r, t) = lpt (r) of
information and the rule is the addition modulo 2.
In the above definition, the rule R is identical for all sites and is applied
simultaneously to each of them, leading to a synchronous dynamics. It is
important to notice that the rule is homogeneous, that is it cannot depend
explicitly on the cell position r. However, spatial (or even temporal)
inhomogeneities can be introduced by having some <I>j(r ) systematically 1
in some given locations of the lattice to mark particular cells for which
a different rule applies. Boundary cells are a typical example of spatial
inhomogeneity. Similarly, it is easy to alternate between two rules by
having a bit which is 1 at even time steps and 0 at odd time steps.
In our definition, the new state at time t + 1 is only a function of the
previous state at time t. It is sometimes necessary to have a longer memory
and introduce a dependence on the states at time t - 1 , t - 2, ... ,t-k. Such
a situation is already included in the definition if one keeps a copy of the
previous states in the current state. Extra bits <l>t can be defined for this
purpose. As an example, the one-dimensional second-order rule
<l>t( r, t + 1 ) = R(<l>t(r - 1, t), <l>t( r + 1, t)) ® <l>t( r, t-1 ) ( 1.9)
where EEl is the sum modulo 2, can be expressed as a first-order rule by
introducing a new state <l>2(r, t ) as follows
<l>t( r, t + 1 ) = R(<l>t(r-1, t ), <l>t( r + 1, t)) EEl <l>2(r, t) ( 1 . 10)
<1>2( r, t + 1 ) = <l>t( r, t )
Note that in the first case, the initial condition requires one to specify
<l>t( r, t = 0 ) and <l>t( r, t = 1 ) . In the second case, there are still two initial
conditions to specify, but now they are <l>t(r, t = 0 ) and <l>2(r, t = 0 ) .
An expression such as rule 1 .9 provides a general way to build a
time-reversible dynamics. Indeed, since the operation EEl is commutative,
14 1 Introduction

associative and a EB a = 0, one can reorganize the terms as

<l>r( r, t - 1) = R(<l>r(r- 1, t), <l>r( r + 1, t)) EB <l>r( r, t + 1) ( 1. 1 1)
which shows that the same rule also allows backward evolution : if one
gives <1>1 ( r, t + 1) and <1> 1 ( r, t) as initial conditions, the system will evolve
back to its own past.

1 .3.2 Neighborhood
A cellular automata rule is local, by definition. The updating of a given
cell requires one to know only the state of the cells in its vicinity. The
spatial region in which a cell needs to search is called the neighborhood.
In principle, there is no restriction on the size of the neighborhood, except
that it is the same for all cells. However, in practice, it is often made up
of adjacent cells only. If the neighborhood is too large, the complexity
of the rule may be unacceptable (complexity usually grows exponentially
fast with the number of cells in the neighborhood).
For two-dimensional cellular automata, two neighborhoods are often
considered : the von Neumann neighborhood, which consists of a central
cell (the one which is to be updated) and its four geographical neigh­
bors north, west, south and east. The Moore neighborhood contains, in
addition, second nearest neighbors north-east, north-west, south-east and
south-west, that is a total of nine cells. Figure 1.4 illustrates these two
standard neighborhoods.
Another useful neighborhood is the so-called Margolus neighborhood
which allows a partitioning of space and a reduction of rule complexity.
The space is divided into adjacent blocks of two-by-two cells. The rule
is sensitive to the location within this so-called Margolus block, namely
upper-left, upper-right, lower-left and lower-right. The way the lattice is
partitioned changes as the rule is iterated. It alternates between an odd
and an even partition, as shown in figure 1.5. As a result, information can
propagate outside the boundaries of the blocks, as evolution takes place.
The key idea of the Margolus neighborhood is that when updating
occurs, the cells within a block evolve only according to the state of that
block and do not immediately depend on what is in the adjacent blocks.
To understand this property, consider the difference with the situation
of the Moore neighborhood shown in figure 1.4. After an iteration, the
cell located on the west of the shaded cell will evolve relatively to its
own Moore neighborhood. Therefore, the cells within a given Moore
neighborhood evolve according to the cells located in a wider region. The
purpose of space partitioning is to prevent these "long distance" effects
from occurring. The sand rule described in section 2.2.6 give an example
of the use of the Margolus neighborhood.
 1 .3 Definitions 15

(a) (b)
Fig. 1 .4. (a) Von Neumann and (b) Moore neighborhoods. The shaded region
indicates the central cell which is updated according to the state of the cells
located within the domain marked with the bold line.

ul ur
11 lr

Fig. 1 .5. The Margolus neighborhood. The lattice is partitioned in 2 x 2 blocks.

Two partitionings are possible - odd and even partitions, as shown by the blocks
delimited by the solid and broken lines, respectively. The neighborhood alternates
between these two situations, at even and odd time steps. Within a block, the cells
are distinguishable and are labeled ul, ur, ll and lr for upper-left, upper-right,
lower-left and lower-right. The cell labeled lr in the figure will be become ul, at
the next iteration, with the alternative partitioning.

1 .3.3 Boundary conditions

In practice, when simulating a given cellular automata rule, one cannot
deal with an infinite lattice. The system must be finite and have bound­
aries. Clearly, a site belonging to the lattice boundary does not have
the same neighborhood as other internal sites. In order to define the
behavior of these sites, a different evolution rule can be considered, which
sees the appropriate neighborhood. This means that the information of
being, or not, at a boundary is coded at the site and, depending on
this information, a different rule is selected. Following this approach, it
16 1 Introduction

periodic fixed

adiabatic reflection

Fig. 1 .6. Various types of boundary conditions obtained by extending the

neighborhood. The shaded block represents a virtual cell which is added at the
extremity of the lattice (left extremity, here) to complete the neighborhood.

is also possible to define several types of boundaries, all with different

behavior.
Instead of having a different rule at the limits of the system, another
possiblity is to extend the neighborhood for the sites at the boundary .
For instance, a very common solution is to assume periodic (or cyclic)
boundary conditions, that is one supposes that the lattice is embedded in
a torus-like topology. In the case of a two-dimensional lattice, this means
that the left and right sides are connected, and so are the upper and lower
sides.
Other possible types of boundary conditions are illustrated in figure 1 .6,
for a one-dimensional lattice. We assume that the lattice is augmented
by a set of virtual cells beyond its limits. A fixed boundary is defined
so that the neighborhood is completed with cells having a pre-assigned
value. An adiabatic boundary condition (or zero-gradient) is obtained by
duplicating the value of the site to the extra virtual cells. A reflecting
boundary amounts to copying the value of the other neighbor in the
virtual cell.
The nature of the system which is modeled will dictate the type of
boundary conditions that should be used in each case.

1 . 3.4 Some remarks

According to its above definition, a cellular automaton is deterministic.
The rule R is some well-defined function and a given initial configuration
will always evolve the same way. However, as we shall see later on, it
may be very convenient for some applications to have a certain degree of
randomness in the rule. For instance, it may be desirable that a rule selects
one outcome among several possible states, with a probability p. Cellular
automata whose updating rule is driven by external probabilities are called
 1 .3 Definitions 17

probabilistic cellular automata. On the other hand, those which strictly

comply with the definition given above, are referred to as deterministic
cellular automata.
In practice, the difference between probabilistic and deterministic cel­
lular automata is not so important. Randomness enters into the rule
through an extra bit which, at each time step, is 1 with probability p
and 0 with probability 1 p , independently at each lattice cell. The
-

question is then how to generate a random bit. We shall see in sec­

tion 2. 1 . 1 that some very simple deterministic cellular automata rules have
an unpredictable behavior, that is there is no way to know what state
a cell will assume at a future stage, unless the evolution is actually per­
formed. Such a rule can be used to produce a pseudorandom bit which
is 1 with probability 1 /2. This mechanism can mimic a probabilistic
cellular automaton. As a matter of fact, it is interesting to note that
rule 30 of Wolfram (see section 2. 1 . 1 ) has been used to produce very
good quality pseudorandom numbers in the Connection Machine parallel
computer [55,24] .
Probabilistic cellular automata are a very useful generalization because
they offer a way to adjust some parameters of a rule in a continuous range
of values, despite the discrete nature of the cellular automata world. This
is very convenient when modeling physical systems in which, for instance,
particles are annihilated or created at some given rate.
As the complexity of a physical system increases, the number of bits
necessary to represents the phenomena at each site of the cellular automata
lattice may become quite large. Dedicated cellular automata hardware
usually imposes a restriction on the maximum number of bits available
at each site. Even on a general purpose computer, memory restrictions
may apply if the rule is expressed as a lookup table in which the updating
is computed in advance for all possible neighborhood configurations : the
size of a lookup table increases as 2n, where n is the total number of bits
in the entire neighborhood.
However, there is a case where a large number of bits per cell can
be easily dealt with using a general purpose computer : when these bits
are interpreted as the binary representation of an integer, or even, of a
floating point number. Then, the usual arithmetic operations +, , * and
-

/ can be used to build the rule.

Although this is not the original way to consider a cellular automata,
the definition we gave allows many interpretations. Strictly speaking, the
multiplication, on a computer, of two 32-bit floating points numbers can
be seen as a cellular automata rule. This generalization of the original
nature of a cellular automata may be quite useful in several applications
we have mentioned and which will be discussed in this book : multiparticle
automata and lattice Boltzmann models.
18 1 Introduction

Another interesting extension of a cellular automata is to allow an asyn­

chronous updating scheme. Clearly the laws of physics are parallel in the
sense that the events do not occur sequentially on a system with several
constituents. However, it may be too much to require full synchronism
of the processes. It turns out that many physical phenomena are well
described by a master e q uation which gives the probability that a process
occurs in the time interval [t, t + dt [. The master equation is sequen­
tial rather than synchronous and it is observed that the same dynamics
may converge to different states depending on whether it is implemented
sequentially or synchronously. Thus, an asynchronous updating scheme
makes sense from the physical point of view as it probably corresponds
to reality. Asynchronism can be achieved, for instance by using a ran­
dom cellular automata : at each site the rule is applied only with a given
probability and, if not, the state remains unchanged.
At this stage, the reader may think that any numerical scheme with
a discrete space and time can be regardeded as a cellular automaton.
This point of view would not be very helpful, however. The richness and
interest of the cellular automaton comes from the microscopic contents
of its rule : there is in general a clear physical or intuitive interpretation
of the dynamics directly at the level of the cell. On the other hand, a
numerical scheme like the discretized solution of a differential equation
is the result of a mapping from a high level mathematical abstraction to
a simpler and more tractable form. The philosophy of cellular automata
modeling works in the opposite direction.

1.4 Problems

1 . 1 . Seve ral ce l l u l a r auto m ata s i m u l ato rs ( l i ke ceflsim) are ava i l ab l e .

Search t h e I nte rnet t o f i n d s o m e o f the m * . Packages l i ke M atl ab o r
M ath e m atica a re envi ro n m e nts where it is poss i b l e to d eve l o p you
own ru l es and visual ize them [56] . Try to program the ru l es that a re
p roposed h e reafte r with M atlab o r M athematica.
1 .2 . Cons i d e r the game of l ife. The ce l l s can be either 0 (dead)
o r 1 (al ive) and the ru l e is: (i) a dead ce l l co mes to l ife if i t is
sou rrou nded by exactly th ree l i vi n g ce l ls among the e i g ht cel l s of
its Moore n e i g h borhood and ( i i ) a l i vi n g ce l l su rvi ves o n l y if it i s
s u rrou nded b y two or th ree n e i g h bors. Show that an isol ated 2 x 2
b l ock of l i vi n g ce l l s i s a sta b l e confi g u ration when the r u l e is ite rated .

• See also http:/ /liinwww.ira.uka.de/ -worsch/ca/prog-envs.html

 1 .4 Problems 19

Show that an iso l ated horizontal seg ment o f th ree l iv i n g cel l s i s a

peri od-two osc i l l at i n g conf i g u rati o n .
. 1 .3 . I n the game of l ife, show that the " o bj ect " d ef i ned by = 1,
ci,j
c i- l ,j-1 = 1, Ci-2,j- 1 = 1, ci- l ,j-2 = 1 and ci,j-2 = 1 , where cijdenotes
the state of ce l l (i, j ) , is a g l ider; that is it m oves at co nstant speed by
conti n u a l ly cyc l i ng t h ro u g h d i fferent i nte rnal co nfi g u rati o n s .
1 . 4. Us i n g a p rogram m i n g lang uage you k n o w (C, Fortra n , Pasca l )
exp ress t h e g a m e o f l ife ru l e assu m i ng that the state 0 o r 1 of each cel l
is sto red i n an array of s i ze n x n, where n i s the l attice s ize. Co n s i d e r
pe r i od ic boundary co nd itions.
1 . 5 . Genera l ize the parity ru l e by i nc l ud i ng the central ce l l in the XOR
ope rati o n . Does it change q u a l itatively the behav i o r? I ncl ude othe r
bound ary cond itions (ad i abatic, ref l ect i n g ) . R u n t h i s auto m aton on
yo u r com p uter and see the effect of these bound ary co nd itions.
1 . 6. Show that the ru l e
C ( t + 1 ) = ( S ( t).and.E ( t )).xor. W ( t).xor.N(t).xor.C(t)
is a random ru l e w h i c h gene rates a dyn a m i c patte rn of 0 and 1 ,
occu rri n g at each s i te with probab i l i ty 1 / 2 . C , N, E , W and S denote the
central cel l of the von N e u m a n n n e i g h borhood and its fo u r n e i g h bors
( n o rth , east, west and south ) .
1 . 7 . Co n s i d e r t h e fo l l owi ng r u l e , based o n t h e M a rg o l us n e i g h borhood :
the n ew conf i g u ration of a b l ock is obta i ned by swapp i n g d i agon a l l y
oppos i te ce l l s , except w h e n t h e co nfi g u ration i s ( ul, ll, lr, ur ) = ( 1 , 0, 1, 0)
or ( ul , ll, lr, ur ) = (0, 1, 0, 1 ) . In these cases , the ru l e a sym m etry with
respect to the h o r i zontal I i ne cross i n g the b l ock, namely ( 1 , 0, 1, 0) �
(0, 1, 0, 1 ) . What i s the behavior of t h i s r u l e (the M argol us b l ocks
a l ternates between the even and odd partition)?
1 .8 . S i m u l ate the behav i o r of the fo l l owi ng ru l e (see [26] , t ime-tunne l
ru l e) :
C ( t) + N ( t) + S ( t) + E (t) + W( t )
{
Sum (t)
C(t - 1 ) if Sum ( t) E {0, 5}
C (t + 1 )
1 - C ( t - 1 ) if Sum ( t) E { 1, 2 , 3 , 4}
where C, N, W, E and S have the same mean i ngs as i n p rob l e m
1 . 6 . Start w i t h an i n itial conf i g u ration where C ( t = 0 ) and C (t = 1) are
i d e ntical and fo rm a d isk in the m i dd l e of the l attice.
1 . 9 . Program the parity rule accord i n g to the fo l l ow i n g th ree meth ods.
There are seve ral tec h n i q u es to i m p l e m e nt a cel l u l a r auto m ata ru l e
o n a com p ute r. ( i ) D irect: the most o bv i ou s way i s t o d ef i n e o n e
20 1 Introduction

var i a b l e fo r each auto m ato n state and each l attice s ite . Then , the
ru l e is co m puted o n the fly, accord i n g to some arith metic o r p rog ram
i nstructions, ove r and over fo r each iteration of the ru n . ( i i ) Lookup
table: anoth e r tech n iq u e is to p re-co m p ute the o utputs of the ru l e fo r a l l
poss i b l e confi g u rati o ns of the n e i g h borhood a n d sto re them i n a lookup
tab l e . T h i s sol ution is fast s i nce the co m p utation red uces to cal c u l at i n g
an i nd ex i n t h e l o o k u p tab l e fo r each l attice s i te . But it uses q u ite a
l ot of m e m o ry if the n u m be r of b i ts n ecessary to specify the state of
each ce l l is l a rge (the tab l e g rows expo nenti a l l y fast) . ( i i i ) M u ltispin
coding: f i n a l l y a th i rd method is m u ltisp i n cod i ng . To save m e m o ry and
speed , o n e packs seve ral l attice s i tes in o n e co m p ute r wo rd (actua l ly,
32 s i tes in a 32-bit word ) . Then, b itwise l og ical operations ( l i ke AND ,
OR , . . . ) , can be used to i m p l ement the ru l e and co m p ute at once the
n ew state of 32 s i tes. Thus, even if one repeats the co m p utation fo r
each i terat i o n , there i s a co n s i d e ra b l e i m p rovement of perform ance
compared with m ethod ( i ) . In genera l , the b i ts a re packed in the words
in the fo l l owi n g way (fi nd out why this is advantageous) : the fi rst s ite is
stored in the fi rst b i t of the fi rst word ; the seco nd s i te goes to the first
b i t of the second wo rd and so o n . With n wo rds , s i tes n + 1 , n + 2 , . . . ,n + n
are stored i n the secon d b i t of a l l the words. T h u s , b i t b E {0, 1 , ... , 3 1 }
o f word k E { 1, . . , n} conta i ns o n e b i t of the state o f s ite l = b * n + k. Fo r
a ru l e req u i ri n g s b i ts per state, ooe n eeds s sets of n words. I n a 20
l attice, each l i ne is m u ltisp i n coded .
1 . 1 0. Snowfl a kes [56] : a s i m p l e CA ru l e can be d evised to model the
g rowth of an obj ect whose structu re s hows some s i m i larities with a
snowfl ake. C rysta l l ization can be vi ewed as a so l i d ificati o n p rocess i n
a sol i d-l i q u i d system . Altho u g h crysta l l i zati o n g rows from a n existi n g
seed , o n e o f the key featu re o f t h i s process is the growth inhibition
phenomenon whereby a crystal l i zed s i te p reve nts nearby s i tes from
so l i d ify i n g . T h i s effect can be i ncorpo rated in the s i m p l e fo l l owi n g
r u l e . Each s i te i s i n o n e o f two states: so l id o r l i q u i d ; a no n-so l i d
s i te that has exactly one so l i d n e i g hbor crystal l i zes . The oth e r s i tes
rem a i n u nchanged . S i m u l ate t h i s ru l e on a h exagonal l attice, sta rti n g
w i t h one i n itial sol i d seed i n the m id d l e . Mapp i n g a hexagonal g r i d
o n a squ are l attice c a n be performed b y co ns i d e r i n g two tru ncated
M oore n e i g h borhoods (see also section 7.2.4) : i n add ition to n o rth ,
south , east and west, odd l i n e s i tes of the l attice o n l y see no rth-west
and south-west n e i g h bors, whereas even l i ne s i tes see no rth-east and
south-east.
 2
Cellular automata modeling

2.1 Why cellular automata are useful in physics

The purpose of this section is to show different reasons why cellular

automata may be useful in physics. In a first paragraph, we shall consider
cellular automata as simple dynamical systems. We shall see that although
defined by very simple rules, cellular automata can exhibit, at a larger
scale, complex dynamical behaviors. This will lead us to consider different
levels of reality to describe the properties of physical systems. Cellular
automata provide a fictitious microscopic world reproducing the correct
physics at a coarse-grained scale. Finally, in a third section, a sampler of
rules modeling simple physical systems is given.

2. 1 . 1 Cellular automata as simple dynamical systems

In physics, the time evolution of physical quantities is often governed by
nonlinear partial differential equations. Due to the nonlinearities, solution
of these dynamical systems can be very complex. In particular, the solution
of these equation can be strongly sensitive to the initial conditions, leading
to what is called a chaotic behavior. Similar complications can occur in
discrete dynamical systems. Models based on cellular automata provide
an alternative approach to study the behavior of dynamical systems. By
virtue of their simplicity, they are potentially amenable to easier analysis
than continuous dynamical systems. The numerical studies are free of
rounding approximations and thus lead to exact results.
Crudely speaking, two classes of problem can be posed. First, given
a cellular automaton rule, predicts its properties. Second, find a cellular
automaton rule that will have some prescribed properties. These two
closely related problems are usually difficult to solve as we shall see on
simple examples.

21
22 2 Cellular automata modeling

The simplest cellular automata rules are one-dimensional ones for which
each site has only two possible states and the rule involves only the nearest­
neighbors sites. They are easily programmable on a personal computer
and offer a nice "toy model" to start the study of cellular automata.
A systematic study of these rules was undertaken by S. Wolfram in
1983 [24,25] . Each cell (labeled i) has at a given time, two possible
states S i = 0 or 1. The state si at time t + 1 depends only on the triplet
(si-l , si, si+ d at time t :
(2. 1 )
Thus t o each triplet o f sites one associates a value lik = 0 o r 1 according
to the following list :

'-v-' '-v-' '-v-' '-v-' '-v-' '-v-' '-v-' '-v-'

111 1 10 101 100 01 1 010 001 000 (2.2)
IX 3

Each possible cellular automata rule fll is characterized by the values

llo, ... , lL 7 . There are clearly 256 possible choices. Each rule can be identified
by an index � computed as follows
7-
�= I: 2j llj (2.3)
i= O
which corresponds to the binary representation lL?lL 6 lL 5 lL4 lL 3 lL2 li l lLO
Giving a rule and an initial state, one can study the time evolution
of the system. Some results can be deduced analytically using algebraic
techniques, but most of the conclusions follow from numerical iterations
of the rules. One can start from a simple initial state (i.e. only one cell
in the state 1 ) or with a typical random initial state. According to their
behavior, the different rules have been grouped in four different classes.
(1 ) Class 1. These cellular automata evolve after a finite number of time
steps from almost all initial states to a unique homogeneous state
(all the sites have the same value). The set of exceptional initial
configurations which behave differently is of measure zero when the
number of cells N goes to infinity. A example is given by the rule
40 (see figure 2. 1 (a)). From the point of view of dynamical systems,
these automata evolve towards a simple limit poin t in the phase
space.
(2) Class 2. A pattern consisting of separated periodic regions is produced
from almost all the initial states. The simple structures generated
are either stable or periodic with small periods. An example is given
by the rule 56 (see figure 2. 1 (b)) Here again, some particular initial
states (set of measure zero) can lead to unbounded growth. The
 2. 1 Why cellular automata are useful in physics 23

( a) (b ) (c) ( d)

Fig. 2. 1 . Example of the four Wolfram rules with a random initial configuration.
Horizontal lines correspond to consecutive iterations. The initial state is the
uppermost line. (a) Rule 40 belonging to class 1 reaches very quickly a fixed
point (stable configuration). (b) Rule 56 of class 2 reaches a pattern composed
of stripes which move from left to right. (c) Rule 18 is in class 3 and exhibits a
self-similar pattern. (d) Rule 1 10 is an example of a class 4 cellular automaton.
Its behavior is not predictable and as a consequence, we observe a rupture in the
pattern, on the left part.

evolution of these automata is analogous to the evolution of some

continuous dynamical systems to limit cycles.
(3) Class 3. These cellular automata evolve from almost all initial states
to chaotic, aperiodic patterns. An example is given by the rule 18 (see
figure 2. 1 (c)). Small changes in the initial conditions almost always
lead to increasingly large changes in the later stages. The evolution
of these automata is analogous to the evolution of some continuous
dynamical systems to strange a ttractors.
(4) Class 4. For these cellular automata, persistent complex structures are
formed for a large class of initial states. An example is given by the
rule 1 10 (see figure 2. 1 (d)). The behavior of such cellular automata
can generally be determined only by explicit simulation of their time
evolution.
The above discussion shows that the "toy rules" considered by Wolfram,
although very simple in construction, are capable of very complex behav­
ior. The validity of this classification is not restricted to the simple rules
described above but is somehow generic for more complicated rules. For
example, one can consider rules for which each cell can have k different
states and involve the r neighbors of the sites. In this case, the number of
possible rules is k (k <zr- l l ) . Several cases have been studied in the literature
and the different rules can be classified in one of the four above classes.
Many of the class 4 cellular automata (starting with k = 2, r = 2) have the
property of computational universality. This means that suitable initial
configurations can specify arbitrary algorithmic procedures and thus the
24 2 Cellular automata modeling

system can be used as a general purpose computer capable of evaluating

any computable function.
However, this "phenomenological" classification suffers drawbacks, the
most serious of which is its non-decidability. Culik and Yu [57] have shown
that for the Wolfram rules defined above, it is undecidable whether all
the finite configurations of a given cellular automaton eventually become
quiescent and consequently, undecidable to which class a given automaton
really belongs.
As we shall see later, some of these simple rules are closely related to
real physical systems.

2. 1 . 2 Cellular automata as spatially extended systems

Spatially extended systems are dynamical systems in which one or several
spatial dimensions are added. Typically, they are defined on a lattice
with each site characterized by a local state. This state evolves in time
according to a given function which usually depends on the values at the
site itself and in a local neighborh � od. A synchronous updating of all
sites is assumed.
According to this definition, a cellular automaton is clearly a spatially
extended system. But other systems, termed coupled map lattices, also
belong to this class. Coupled map lattices can be seen as cellular automata
with an infinite number of possible states at each site. The only difference
between the two types of systems is the continuous character of the local
variables. The definition of the rule is replaced by the choice of the local
dynamics of the continuous degrees of freedom sitting at each site. Lattice
Holzmann systems, which will be discussed in detail throughout this book,
are a particular case of a coupled map lattice.
Extended systems with local interactions and synchronous updating are
of great importance to understand the nature of complexity exhibited by
several natural processes because they contain the minimal ingredients
of many real systems : a dynamics which depends both on space and
time. Fully developed turbulence is an example for which basic models
are needed to provide some insights on the fundamental mechanisms that
are involved. For this reason, coupled map lattices are often use as a toy
model for turbulent systems.
An important problem which arises in the study of spatially extended
system is the existence of non-trivial collective behavior. A non-trivial col­
lective behavior is an unexpected behavior which is observed on macro­
scopic quantities such as, for instance, the total density (or the sum over
all the lattice sites of the local state). A typical example is a periodic or
quasi-periodic behavior of this global density, in complete disagreement
with -equilibrium statistical mechanics or mean-field calculations.
 2. 1 Why cellular automata are useful in physics 25

o ���--��-���
0 0.6
c(t)

Fig. 2.2. Non-trivial collective behavior for the simple 3D cellular automata rule
given by relation (2.4). The plot shows the behavior of the global density c(t) as
a function of its previous value.

Usually, non-trivial collective behaviors are likely to appear in high­

dimensional lattices. However, there is a simple cellular automata example
in three dimensions [ 58 ] : one considers a 3D Cartesian lattice with periodic
boundary conditions. Each site i can be either in state Si = 0 or in state
s i = 1 . The evolution rule is the following

s 1·(t + 1 ) = {0
1 if 2:-ne ig_hbors Sj( t)
otherw1se
E {0, 5}
(2.4)

Here the sum over the neighbors extends to the six nearest neighbors in
the cubic lattice, plus the central site i itself. Thus, the sum ranges between
0 and 7. The rule is that the new value si ( t + 1 ) is 1 if and only if this sum
equals 0 or 5.
As time goes on, one can record the density c ( t) defined as
c ( t) = L si ( t )
lattice
This quantity exhibits a non-trivial behavior in the sense that it does
not relax to a constant value at large time. Instead, it follows a low
dimensional dynamics which can be represented by the map shown in
figure 2.2 where c ( t + 1) is plotted as a function of c ( t ) .
Such non-trivial collective behaviors in spatially extended systems has
been observed on several totalistic cellular automata rules. A totalistic rule
means that the new state of the cell i is a function only of the sum of the
states Sj for cells j belonging to a given neighborhood of i. Rule (2.4) is an
example of a totalistic rule. Totalistic rules can be seen as mean-field-like
evolution since in the each site "sees" an average environment.
Non-trivial collective behaviors are not singular phenomena, requiring
the fine tuning of some conditions, but appear as general features, robust
26 2 Cellular automata modeling

to small modifications of local rules and initial conditions. The dynamics

is usually accompanied by noise which is related to the disorder of the
initial configuration. This noise disapears in the asymptotic regime.
Several types of collective behaviors have been observed and charac­
terized for both cellular automata and coupled map models. A uniform
treatment of both cases has been given by Chate and Manneville [59]
and their discussion leads to the classification of cellular automata rules
according to the various types of behaviors that are observed.
An important question concerns the minimum requirement for produc­
ing such non-trivial collective behaviors and several conclusions have been
drawn by Chate and Manneville. First, the deterministic character of the
evolution is not essential. Collective behaviors appear both for determin­
istic and probabilistic evolutions. Second, the synchronicity is essential.
To exhibit non-trivial behavior, the system requires a clock providing
an "instantaneous" global communication channel. Thus, the analogy
with more traditional systems studied in statistical mechanics (sequential
Monte-Carlo updating for example) should be made with great care.

2. 1 .3 Several levels of reality

The previous paragraph illustrated the point that cellular automata have
a strong similarity with complex systems, despite the simplicity of their
elementary dynamics. It turns out that the reason for such a similar
behavior is quite general and will be used all the time throughout this
book : often, the macroscopic behavior of a system composed of many
interacting constituents depends very little on the microscopic details of
the interactions. For instance, it is well known that the Navier-Stokes
equation governing the motion of a fluid is first of all an expression of
momentum conservation during the elementary interactions between the
molecules. The details of these interactions shows up in the coefficients of
the equation (for instance, the viscosity), but not in its algebraic structure.
We shall come back to this point many times and make it more explicit
,in various examples.
This is actually one of the most important points behind the concept of
cellular automata modeling. Several levels of reality exist when studying
a physical system. Roughly speaking, there are two main levels, corre­
sponding to a microscopic and macroscopic scale of observation. The
same system may look quite different when described at these various
levels. At a microscopic level, interactions may be governed by compli­
cated potentials and, sometimes, require the use of quantum mechanics
to be described properly. At a macroscopic level, the properties of the
system are dominated by the aggregate effect of all the microscopic inter­
actions. The resulting behavior, when seen at a large scale of observation,
 2. 1 Why cellular automata are useful in physics 27

is mostly related to the generic features of the microscopic interactions.

The complexity of the macroscopic world is apparently disconnected from
that of the microscopic world, although the former is driven by the latter.
The microscopic details are lost when the system is looked at through a
macroscopic filter and new collective properties emerge.
A well-known example of these two levels of reality is given by the study
of the magnetic properties of ferromagnetic materials. At a microscopic
level, the phenomena is driven by the distribution of electrons in atoms,
which, in principle, is described by the laws of quantum mechanics.
On the other hand, at a macroscopic level, one observes a collective be­
havior : below some critical temperature, the magnetic dipoles associated
to each atom align to form macroscopic magnetic domains. This property,
which makes sense only at a macroscopic scale of observation, is not re­
lated to the detailed form of the ferromagnetic interaction but rather comes
from the very general fact that the local energy of interaction is lowered
when two magnetic dipoles are oriented along the same direction. As a
matter of fact, the macroscopic behavior described here can be reproduced
by a very simple dynamical model, namely the famous Ising model. In this
model, all the microscopic complexity is reduced to simple interactions
between spins located at each site of a regular Cartesian lattice. Spins are
either 1 or -1, and each of them accounts for the resulting magnetic dipole
of the atom. It flips from up to down (or the converse) according the
local energy gain this movement results in. The Ising model offers a very
poor description of real microscopic interactions but retains the significant
aspects that show up when the system is considered at a larger scale.
Ising-like systems are very close in spirit to a cellular automata model.
They are based on a fictitious microscopic universe, which aims at cap­
turing the essential features of real microscopic interactions, but no more.
Kinetic Ising models are defined on a discrete space, with a discrete time,
as is the case of cellular automata. However, here the spin updating is done
sequentially, whereas, in a cellular automata, it is done synchronously. This
difference is of importance because a system of spins may not converge
to the same final configuration when the updating is done sequentially or
in parallel. We shall come back to this point in section 2.2.3.

2. 1 .4 A fictitious microscopic world

The cellular automata approach works when the simplification of the
microscopic laws pertaining to a given phenomena is not relevant at a
macroscopic scale of observation. We learn from statistical mechanics
that this fact is often true for systems whose complexity comes from
a collective behavior rather than from some distinctive aspects of the
microscopic interactions.
28 2 Cellular automata modeling

From the observation that the macroscopic behavior of many systems

(which is precisely the level of reality we are interested in) has little to do
with the true microscopic nature, it is a clear advantage to invent a much
simpler microscopic reality, which is more appropriate to our numerical
means of investigation. The following Einstein quotation :
Everything should be made as simple as possible but not simpler
reflects very well the spirit of cellular automata modeling. The important
point is to capture the essential and relevant ingredients of a real phe­
nomena and consider them as the fundamental physical laws of a new
imaginary system. In that sense, cellular automata are a caricature of the
real world rather than its portrait.
Modeling a system at a microscopic level of description has significant
advantages. The interpretation of cellular automata dynamics in terms
of simple microscopic rules offers a very intuitive and powerful approach
to modeling phenomena that are very difficult to include in more tradi­
tional approaches (such as differential equations). For instance, boundary
conditions are often naturally implemented in a cellular automata model
because it has a natural interpretation at this level of description (e.g.
particles bouncing back from an obstacle). The phenomena of wetting of
a solid substrate by a spreading liquid illustrates the difficulty in defining
appropriate boundary conditions (see section 7.2 ) .
The design of correct cellular automata models implies that the essential
aspects of a complex phenomena have been recognized and reduced to
a simple tractable form. This process of reducing a complex behavior
to the sum of simple mechanisms is an essential step in any scientific
investigation. Therefore, cellular automata models provide a very powerful
tool in fundamental research, real problems and pedagogical applications.
The numerical simplicity of many cellular automata rules makes them
well suited to massive computer simulations. The Boolean nature of the
operations that are to be performed leads to an exact evolution where no
truncation is necessary and no numerical instabilities are present. As a
result, the computer simulation is a totally faithful implementation of the
mathematical model expressed by the cellular automata rule. This is very
interesting when modeling N-body problems, because the many-particle
correlation functions are not corrupted by the numerical scheme.

2.2 Modeling of simple systems: a sampler of rules

The purpose of this section is to illustrate the general concept we have

presented so far. We will introduce a few simple cellular automata rules
aimed at describing the behavior of several physical process. From these
 2.2 Modeling of simple systems : a sampler of rules 29

examples, we shall develop our intuition and also identify some problems
related to the cellular automata approach. In particular, we shall see the
importance of developing techniques to assess the range of validity of a
model and its limitations.

2.2. 1 The rule 1 84 as a model for surface growth

Wolfram's rule 1 84, can be used to model the formation of spatio-temporal
structures for deterministic surface growth. An important surface growth
mechanism is through ballistic deposition. Krug and Spohn [60] studied
the following simple lattice model. Consider a one-dimensional surface
configuration parallel to the x axis of a square lattice. Particles move
on straight lines along the y axis and become part of the deposit when
they reach the empty nearest neighbor of an occupied site. Thus an
incoming particle may stick either to the side edge (with a rate rs)
or to the top edge of an existing column (with a rate r t ). The limit
p = r t frs ---+ oo corresponds to the so-called random deposition model.
One then considers the opposite limit p ---+ 0. With simultaneous up­
dating, the interface dynamics can be written is the following simple
form :
(2 .5 )
for all sites i, where hi (t) denotes the height of the surface at site i and
time t.
One can then consider a slightly modified version of the model, namely
deposition onto a one-dimensional surface parallel to the ( 1, 1 ) direction
of the square lattice. The particles are allowed to stick only at the minima
of the surface as shown on figure 2.3. Hence, the surface configurations
generated have a local gradient ± 1 . The particles have a simple dynamics.
With rate 1, they jump to the right except when the final site is occupied,
in which case they stay (hard-core repulsion). Thus the configuration at
time t + 1 is obtained from the configuration at time -t by a local rule
which is simply Wolfram's rule 1 84. For long times, the dynamics leads
to a trivial limit cycle. However, the approach to the final state has a rich
space-time structure when interpreted as a surface.

2.2.2 Probabilistic cellular automata rules

The Wolfram rules can be considered as the deterministic limit of prob­
abilistic cellular automaton rules which can exhibit even more complex
behavior as a function of the free parameters entering into the rule. A
simple, physically important example of a probabilistic rule is provided
by the problem of directed percolation. Consider the square lattice of
30 2 Cellular automata modeling

t+l

t
X

Fig. 2.3. Deterministic growth from a rough surface parallel to the (1, 1) direction.
At each time step, all local surface minima are filled with new particles. The
corresponding lattice gas evolves according to the automata rule 1 84.

figure 2.4. Any lattice site may be present with a probability p or absent
with probability ( 1 - p). Moreover, any bond may be present with prob­
ability q and absent with probability ( 1 - q). Assume that a single site is
"wet" at time t = 0. Present bonds conduct "water" only in the downward
(increasing time) direction, and when both sites connected by the bond
are present.
For q = 1, one speaks of site-directed percola tion, and for p = 1,
of bond-directed percola tion. Given the two probabilities p and q, one
can ask several questions : what is the probability of finding wet sites at
(time) level t and what is the typical size of the wet cluster ? In each
case, there are critical values (pc, qc), called the percolation thresholds,
above which the wet cluster spans the whole system. Directed perco­
lation is a well-known problem of "geometrical phase transition" for
which scaling theory and the renormalization group approach have been
developed.
This problem can easily be formulated in term of probabilistic cellular
automaton. Consider a linear chain or ring of sites ; with each site one
associates a variable 1JJ i ( t) = 0 or 1 . At odd (even) times, odd (even)
indexed sites change their state according to a probabilistic rule, and even
(odd) indexed sites stay in the same state. The full space-time history
can be presented on a two-dimensional lattice. The rules are defined
by the conditional probabilities P ( 1pi(t + 1 ) 1 1JJ i- l ( t ), 1JJ i+ l (t)) . For directed
 2.2 Modeling of simple systems : a sampler of rules 31

0
1
2
3
4
5
6

Fig. 2.4. Two-dimensional square lattice for directed percolation. The dots
represent the sites which can be present or absent. The heavy lines between the
sites represent the bonds. The vertical axis corresponds to time.

percolation, we have :
P ( 1 1 00 ) = 0, y = P ( 1 1 1 1 ) = pq , z
P ( l i01 ) = P ( 1 1 10) = p q(2 - q)
=
(2 .6 )
The situation in the parameter space (p , q) is represented in figure 2 .5.
At the transition point, the percolation cluster is a fractal object. More­
over, near the transition point, many physical quantities have a power law
behavior in terms of p - Pc or q - qc, with universal critical exponents.
Here again, one sees that a very simple cellular automaton rule leads to a
complex structure.
Another interesting and simple example of a probabilistic cellular au­
tomata rule is related to forest fire models [6 1,62] . This probabilistic
cellular automata model is defined on a d-dimensional hypercubic lattice.
Initially, each site is occupied by either a tree, a burning tree or is empty.
The state of the system is parallel updated according to the following rule :
1 . A burning tree becomes an empty site.
2. A green tree becomes a burning tree if at least one of its nearest
neighbors is burning.

3. At an empty site, a tree grows with probability p.

4. A tree without a burning nearest neighbor becomes a burning tree
during one time step with probability f (lightning).
32 2 Cellular automata modeling

1 .0 -�--:c�-;;;-
>..-.,
...
; ..
,----
-- ----�---

,.
,.
,.
,.
.·
z
,.
,.
,.
,.
,.
;
;
.·

;
;
.·

;
.·

0.5 ;
;
.·

;
.·

;
,
.·

,
.·

.·

,
I
•
•
•
.

0 0.5 y 1 .0
Fig. 2.5. Parameter space y = pq ; z = pq(2 - q ) defining the cellular automata
rules for generalized directed percolation. The dashed and dotted lines correspond,
respectively, to the bond and site percolation. The system percolates to the right
of the (solid) transition line.

Figure 2.6 illustrates the behavior of this rule, in a two-dimensional

situation. Provided that the time scales of tree growth and burning down
of forest clusters are well separated (i.e. in the limit f jp � 0), this models
has self-organized critical states [62] . This means that in the steady state,
several physical quantities characterizing the system have a power law
behavior. For example, the cluster size distribution JV ( s) and radius of a
forest cluster � ( s) vary with the number of trees s in the forest cluster as :

(2.7)

and

(2.8)

Scaling relations can be established between the critical exponents r and

J.l, and the scaling functions ((/ and :7 can be computed.

Note that there is a more general correspondence between d-dimensional

probabilistic cellular automata and d + 1 equilibrium statistical physics
problems. To each d-dimensional probabilistic cellular automata one can
associate a classical ( d + ! )-dimensional statistical model [63] .
 2.2 Modeling of simple systems : a sampler of rules 33

Fig. 2.6. The forest fire cellular automata rule : grey sites correspond to a
grown tree, black pixels represent burned sites and the white color indicates a
burning tree. The snapshot given here represents the situation after a few hundred
iterations. The parameters of the rule are p = 0.3 and f = 6 X w-5•

2.2.3 The Q2R rule

The Q2R rule was proposed by G. Vichniac [64] in the 1980s as a model
for Ising spin dynamics. The problem is to find a cellular automata rule
modeling the behavior described in section 2. 1 .3.
We consider a two-dimensional square lattice such that each site holds
a spin Si which is either up ( si = 1) or down ( si = 0). During their discrete
time evolution, the spins will flip (or not flip) according to a cellular
automata rule. Following our modeling philosophy, this rule will be built
up from very simple and general principles. In the present case, it will
be local energy conservation. This means we are considering a system
which cannot exchange energy with its surroundings. The model will be
a microcanonical cellular automata simulation of Ising spin dynamics,
without a temperature but with a critical energy.
In an Ising model, the energy of a spin s i comes from couplings with
the spins Sj located in its von Neumann neighborhood (north, west, south
and east). The contribution of the pair ( si , sj ) to this energy is some value
-J when two spins are aligned, and J when they are not aligned.
A spin s i can flip at time t to become 1 - s i at time t + 1 if and only
if this move does not cause any energy change. Accordingly, spin Si will
34 2 Cellular automata modeling

flip if the number of its neighbors with spin up is the same as the number
of its neighbors with spin down. However, one has to remember that the
motion of all spins is simultaneous in a cellular automata. The decision
to flip is based on the assumption that the neighbors are not changing. If
they flip too (because they obey the same rule), then energy may not be
conserved. This is simply illustrated by a one-dimensional chain of spins.
Suppose one has the following spin configuration at time t

'-v-" '-v-" '-v-" '-v-" '-v-"

time t : ... 0 1 0 0 1 0 ... ( 2.9)
+J +J -J +J +J
where the coupling energy is indicated for each pair of spin. Now, if all
spins flip according to our prescription, we obtain the situation

'-v-" '-v-" '-v-"

time t + 1 : ... 1 1 1 1 ... (2. 10)
-J -J -J
As we can see by counting the coupling energy for the middle spins, the
energy balance is incorrect. This is due to the simultaneity of the update.
A way to cure this problem is to split the updating into two phases : first,
one flips the spins located at an odd position along the chain, according
to the configuration of the even spins. In the second phase, the even
sublattice is updated according to the odd one. In a two-dimensional
lattice, the same procedure has to be carried out : the original lattice is
decomposed into two sublattices, as given by the white and black squares
of a chess board.
This strategy prevents the conflict of motion between adjacent spins.
However, at first sight, this method contradicts our definition of a cellular
automaton, which stipulates that all sites be updated simultaneously. To
conciliate the requirements of the problem with this definition, one can
simply extend the basic rule to include some spatial sensitivity : initially,
lattice sites hold an extra bit b of information whose value is 0 for the odd
sublattice and 1 for the even sublattice. For a two-dimensional system the
spatial "texture" given by b at time t = 0 would be
... 1 0 1 0 1 0 ...
0 1 0 1 0 1 ...
1 0 1 0 1 0 .. .

0 0 0 ( 2. 1 1 )
1 1 1 ...
... 1 0 1 0 1 0 ...
.. 0. 1 0 1 0 1 ...
The flipping rule described earlier is then regulated by the value of b. It
takes place only for those sites for which b = 1 . Of course, the value of b
is also updated at each iteration according to b ( t + 1 ) = 1 b ( t ), so that at -

the next iteration, the other sublattice is considered. Thus, at time t = 1,

 2.2 Modeling of simple systems : a sampler of rules 35

the space parity b of the system changes to

... 0 1 0 1 0 1 .. .
... 1 0 1 0 1 0 .. .
... 0 1 0 1 0 1
(2. 12)
... 1 0 1 0 1 0 .. .
... 0 1 0 1 0 1 .. .
... 1 0 1 0 1 0 .. .
In two dimensions, the Q2R rule can be the expressed by the following
relations

si1' ( t + 1 ) =
{ S1ij(t)- Sij(t) if bij = 1 and S i-l,j + S i + l,j + Si,j- 1 + S i,j+ l
otherwise
= 2

(2. 1 3 )
and
(2. 14)
where the indices (i, j) label the Cartesian coordinates and sij( t = 0 ) is
either one or zero.
The question is now how well does this cellular automata rule perform
to describe an Ising model. Figure 2.7 shows a computer simulation of
the Q2R rule, starting from an initial configuration with approximately
1 1 % of spins S ij = 1 (figure 2.7 (a)). After a transient phase (figures
(b) and (c)), the system reaches a stationary state where domains with
"up" magnetization (white regions) are surrounded by domains of "down"
magnetization (black regions).
In this dynamics, energy is exactly conserved because that is the way
the rule is built. However, the number of spins down and up may vary.
Actually, in the present experiment, the fraction of spins up increases from
1 1 % in the initial state to about 40% in the stationary state. Since there
is an excess of spins down in this system, there is a resulting macroscopic
magnetization.
It is interesting to study this model with various initial fractions P s of
spins up. When starting with a random initial condition, similar to that of
figure 2.7 (a), it is observed that, for many values of p5, the system evolves
to a state where there is, on average, the same amount of spin down
and up, that is no macroscopic magnetization. However, if the initial
configuration presents a sufficiently large excess of one kind of spin, then
a macroscopic magnetization builds up as time goes on. This means there
is a phase transition between a situation of zero magnetization and a
situation of positive or negative magnetization.
It turns out that this transition occurs when the total energy E of the
system is low enough (low energy means that most of the spins are aligned
36 2 Cellular automata modeling

(a) (b)

(c) (d)

Fig. 2.7. Evolution of a system of spins with the Q2R rule. Black represents the
spins down Sij = 0 and white the spins up Sij = 1. The four images (a), (b), (c)
and {d) show the system at four different times ta = 0 < tb < tc < < td .

and that there is an excess of one species over the other), or more precisely
when E is smaller than a critical energy Ec.
Therefore, the Q2R rule captures an important aspect of a real magnetic
system, namely a non-zero magnetization at low energy (which can be
related to a low temperature situation) and a transition to a non-magnetic
phase at high energy. However, Q2R also exhibits an unexpected behavior
that is difficult to detect from a simple observation. There is a loss of
ergodicity : a given initial configuration of energy Eo evolves without
visiting completely the region of the phase space characterized by E = E0 .
This is illustrated by the following simple example, where a ring of four
spins is considered. For the sake of convenience, we shall represent it as
 2.2 Modeling of simple systems : a sampler of rules 37

a linear chain with periodic boundary conditions (the first and last spins
are nearest neighbors).
t : 1001
t+1 : 1 100
t+2 : 0 1 10
t+3 : 001 1
t+4 : 100 1 (2. 15)
After four iterations, the system cycles back to its original state. The
configuration of this example has Eo = 0. As we observe, it never evolves
to 01 1 1, which is also a configuration of zero energy.
This non-ergodicity means that not only energy is conserved during
evolution of the automaton, but also other quantities which partition the
energy surface in independent regions. As we shall see, the study of the
dynamical invariants is very important when studying cellular automata
models.

2.2.4 The annealing rule

A natural class of cellular automata rules consists of the so-called majority
rules. Updating selects the new state of each cell so as to conform to the
value currently hold by the majority of the neighbors. Typically, in these
majority rules, the state is either 0 or 1 .
A very interesting behavior i s observed with the twisted majority rule
proposed by G. Vichniac [64] : in two dimensions, each cell considers
its Moore neighborhood (i.e itself plus its eight nearest neighbors) and
computes the sum of the cells having a value 1 in the Moore neighborhood.
This sum can be any value between 0 and 9. The new state SiJ (t+ 1) of each
cell is then determined from this local sum, according to the following
table
sumiJ(t) 0 1 23456789
Sij ( t + 1 ) 0000 1 0 1 1 1 1 (2. 16)
As opposed to the plain majority rule, here, the two middle entries of the
table have been swapped. Therefore, when there is a slight maj ority of 1 s
around a cell, i t turns t o 0 . Conversely, if there i s a slight majority o f Os,
the cell becomes 1 .
Surprisingly enough this rule describes the interface motion between two
phases, as illustrated in figure 2.8. Vichniac has observed that the normal
velocity of the interface is proportional to its local curvature, as required
by the Allen-Cahn [65] equation. Of course, due to its local nature, the
rule cannot detect the curvature of the interface directly. However, as the
38 2 Cellular automata modeling

( a) (b) (c)

Fig. 2.8. Evolution of the annealing rule. The inherent "surface tension" present
in the rule tends to separate the black phases s = 1 from the white phase s = 0.
The snapshots (a), (b) and (c) correspond to t = 0, t = 72 and t = 270 iterations,
respectively. The extra gray levels indicate how "capes" have been eroded and
"bays" filled : dark gray shows the black regions that have been eroded during
the last few iterations and light gray marks the white regions that have been
filled.

rule is iterated, local information is propagated to the nearest neighbors

and the radius of curvature emerges as a collective effect.
This rule is particularly interesting when the initial configuration is a
random mixture of the two phases, with equal concentration. Otherwise,
some pathological behavior may occur. For instance, an initial black
square surrounded by the white phase will not evolve : right angles are
not eroded and are stable structures.

2.2.5 The HPP rule

The HPP rule is the first of an important class of cellular automata
models. The basic working objects of such models are point particles that
move on a lattice, according to appropriate rules. The state of each cell
represents the presence or the absence of such particles.
In order to make this kind of fully discrete "molecular dynamics" com­
patible with cellular automata dynamics, one should prevent an arbitrary
number of particles being present simultaneously on a given site. This
is due to the definition of a cellular automaton, in which the states are
described by a finite number of bits. The rule should be defined so as to
guarantee an exclusion principle.
The purpose of the HPP rule is to model a gas of colliding particles. The
essential features that are borrowed from the real microscopic interactions
are the conservation laws, namely local conservation of momentum and
local conservation of particle number.
 2.2 Modeling of simple systems : a sampler of rules 39

I
T-
- ·

,_ - I
I

Fig. 2.9. Example of a configuration of HPP particles.

The HPP lattice gas automata is traditionally defined on a two­

dimensional square lattice. Particles can move along the main directions
of the lattice, as shown in figure 2.9. The model limits to 1 the number
of particles entering a given site with a given direction of motion. This
is the exclusion principle we were mentioning. Consequently, four bits
of information in each site are enough to describe the system during its
evolution. For instance, if at iteration t site r has the following state
s (r , t ) = ( 101 1 ), it means that three particles are entering the site along
direction 1,3 and 4, respectively.
The cellular automata rule describing the evolution of s (r , t ) is usually
split into two steps : collision and motion. The collision phase specifies how
particles entering the same site will interact and change their trajectories.
During the motion phase, or propagation, the particles are actually moved
to the nearest neighbor site they were traveling to. Figure 2. 10 illustrates
the HPP rules. This decomposition in two phases is another way to
partition space, as in the Margolus neighborhood. According to our
Boolean representation of the particles at each site, the collision part for
a two-particle head-on collision is expressed as
( 1 010) � (0101) (0 101) � ( 10 10) (2. 17)
all the other configuration being unchanged. During the propagation
phase, the first bit of the state variable is shifted to the east neighbor cell,
the second bit to the north and so on.
Since the collision phase amounts to rearranging the particles in a
different direction, it ensures that the exclusion principle will be satisfied,
provided that it was at time t = 0.
40 2 Cellular automata modeling

( a)

(b)
�

(c)
�

time t time t+l

Fig. 2.10. The HPP rule : (a) a single particle has a ballistic motion until
it experiences a collision. (b) and (c) The two non-trivial collisions of the
HPP model : two particles experiencing a head-on collision are deflected in the
perpendicular direction. In the other situations, the motion is ballistic, that is the
particles are transparent to each other when they cross the same site.

The aim of this rule is to reproduce some aspects of the real interaction
of particles, namely that momentum and particle number are conserved
during a collision. From figure 2. 10, it is easy checked that these properties
are obeyed : for the two situations (b) and (c) where the particles do
actually experience a collision, a pair of zero momentum particles along a
given direction is transformed into another pair of zero momentum along
the perpendicular axis.
The HPP rule captures another important ingredient of the microscopic
nature of a real interaction : invariance under time reversal. Figures 2. 10
(b) and (c) show that, if at some given time, the direction of motion of all
particles are reversed, the system will just trace back its own history. Since
the dynamics of a deterministic cellular automaton is exact, this allows
 2.2 Modeling of simple systems : a sampler of rules 41

(a)

(b)

Fig. 2. 1 1 . Time evolution o f a HPP gas. (a) From the initial state t o equilibrium.
(b) Illustration of time reversal invariance : in the rightmost image of (a), the
velocity of each particle is reversed and the particles naturally return to their
initial position.

us to demonstrate that physical systems return to their original situation

when all particles reverse their velocity.
Figure 2. 1 1 illustrates the time evolution of a HPP gas initially confined
in the left compartment of a container. There is an aperture in the wall of
the compartment and the gas particles will flow so as to fill the entire space
available to them. In order to include a solid boundary in the system,
the HPP rule is modified as follows : when a site is a wall (indicated by
an extra bit), the particles no longer experience the HPP collision but
bounce back to where they come from. Therefore, particles cannot escape
a region delimited by such a reflecting boundary.
If the system of figure 2. 1 1 is evolved, it reaches an equilibrium after
a long enough time and no trace of its initial state is any longer visible.
However, no information has been lost during the process (no numerical
dissipation) and the system has the memory from where it comes from.
Reversing all the velocities and iterating the HPP rule makes all particle
go back to the compartment in which they were initially located.
This behavior is only possible because the dynamics is perfectly exact
and no numerical errors are present in the numerical scheme. If one
introduces externally some errors (for instance, one can add an extra
particle in the system) before the direction of motion of each particle
is reversed, then the reversibility is lost. The result of such a change is
shown in figure 2.12 which is a snapshot of the HPP gas at a time all
particles would have returned to the left compartment if no mistake had
been introduced.
42 2 Cellular automata modeling

Fig. 2. 12. If a small error is introduced (e.g. one particle is added) in the situation
of the previous figure before the velocities are reversed, the system cannot return
to its initial state. The particles left in the right part of the system are those which
have been affected by the presence of the extra particle.

Fig. 2. 1 3 . Propagation of a compression wave in the HPP model. Initially, a high

concentration of particle is created inside a circle, at the middle of the system.
The two images represent snapshots at two time steps t1 < t2 . We can observe
how anisotropy builds up during time evolution.

The HPP rule is important because it contains the basic ingredients of

many models we are going to discuss in this book. We shall see in chapter 3
that the capability of this rule to model a real gas of particle is poor, due
to a lack of isotropy. This problem is illustrated in figure 2. 13, where a
compression wave propagates in the gas. Starting from an isotropic initial
condition of high particle density in the middle, we observe that the wave
does not propagate identically in all directions (the lattice axis show up
in the figure). A remedy to this problem is to use a different lattice (see
chapter 3).

2.2.6 The sand pile rule

The physics of granular materials has attracted much interest, recently [66-
69] . It is possible to devise a simple cellular automata rule to model basic
piling and toppling of particles like sand grains. The idea is that grains
 2.2 Modeling of simple systems : a sampler of rules 43

Fig. 2. 14. A stable pile of sand grains.

Fig. 2. 1 5. A conflict of motion results from the use of the Moore neighborhood.
The two gray particles are unstable and both want to move to the same new
location.

can stack on top of each other if this arrangement is stable. Of course, real
sand grains do not stand on a regular lattice and the stability criteria are
expected to depend on the shape of each grain. Despite this microscopic
complexity, the result is that sand piles that are too high topple.
In the framework of cellular automata modeling, the toppling mecha­
nisms can be captured by the following rule : a grain is stable if there is
a grain underneath and two other grains preventing it falling to the left
or right, as depicted in figure 2. 14. In the Moore neighborhood, it means
that a central grain will be at rest if the south-west, south and south-east
neighbors are occupied. Otherwise the gain topples downwards to the
nearest empty cell.
For instance, suppose one has the situation shown in figure 2. 1 5 . Since
the updating of all cells is simultaneous, the two shaded grains will move
to the same empty cell : there is a conflict. To prevent conflict, each
particle should know about the presence of the other one before they
move, and take a common decision as to which one is going to fall. This
requires a larger neighborhood than the Moore neighborhood and a more
complicated rule.
The Margolus neighborhood gives a simple way to deal with the syn­
chronous motion of all particles. Piling and toppling is determined by the
situation within 2 x 2 adjacent blocks. Of course (see section 1.3.2), for all
other steps, the space partitioning is shifted one cell down and one cell
left, so that, after two iterations, each cell sees a larger neighborhood.
Figure 2. 16 gives a possible implementation of a sand rule. It shows
how the configurations evolve due to toppling.
44 2 Cellular automata modeling

EB EE l8B EB � BBJ [8EH9S

� � � � � � � �

ffi EE � � � � [Sl�H�B
Fig. 2. 1 6. The sand rule expressed for the Margolus neighborhood. Only the
configuration that are modified are shown.

ffi
p � 1-p
ffi �
Fig. 2. 1 7. The sand rule allows for blocking (friction) with probability p when
two grains occupy the upper part of a Margolus block.

The configuration in which the upper part of a block is occupied by

two particles while the lower part is empty, is not listed in this figure,
although it certainly yields some toppling. Treating this configuration with
a probabilistic rule turns out to produce a more realistic behavior. When
this configuration occurs, the evolution rule is given in figure 2. 1 7. The
motivation for introducing a probabilistic toppling is that some friction
may be present between grains and some "arches" may appear to delay
collapse. Of course, the toppling of other configurations could also be
controlled by a random choice.
When no blocking occurs, the sand rule can be expressed by the follow­
ing algebraic relations. We call Sui . Sur , su and Sir the state before updating,
where the subscripts stand for "upper-left," "upper-right," "lower-left" and
"lower-right."
Sul (t + 1) SuiSII(Sir + ( 1 - Sir )sur) (2. 1 8 )
Sur( t + 1) SurSir(SII + ( 1 - su ) suz ) (2. 19)
su ( t + 1) Sll + ( 1 - su ) [Sul + ( 1 - Sui )SurSir] (2.20)
s lr( t + 1) - Sir + ( 1 - Slr) [Sur + ( 1 - Sur)SuiSII] (2.21 )
 2.2 Modeling of simple systems : a sampler of rules 45

t=20 t=lOO t= 4200 t= 4800

Fig. 2. 1 8. Evolution of the sand rule for an hour-glass simulation.

It is easy to check that the number of particles is conserved during the

evolution, namely that
Suz ( t + 1 ) + Sur( t + 1 ) + su ( t + 1) + Szr( t + 1) = Suz + Sur + Sl! + Szr (2.22)
Equation 2. 1 8 means that an upper-left particle will fall down unless both
the lower-left and the lower-right cells are occupied, or both an upper­
right and a lower-left particle are present. The other equations follow
from the same kind of arguments.
In order to make this rule more interesting, we must introduce the
ground, on which the grain will stop falling. This is necessary to observe
piling and then toppling. The ground will be represented as an extra bit g
in the state of each cell which, of course, does not evolve during updating.
A sand grain located on a ground g = 1 cell will be at rest for ever. This
is taken into account in the rule by modifying equations 2. 1 8 and 2. 19 as
Suz (t + 1 ) = guzSuz + ( 1 - guz ) Suz s u ( s zr + ( 1 - Szr ) Sur ) (2.23)
Sur (t + 1) = gurSur + ( 1 - gur ) SurSzr (Sll + ( 1 - su)suz ) (2.24)
which means that a particle can leave the upper part of the Margolus
block only if it is not a ground cell.
Equations 2.20 and 2.21 are modified in the following way
s u ( t + 1 ) = Sl! + ( 1 - su) [suz ( 1 - g uz ) + ( 1 - Suz )Sur ( 1 - gur )Szr] (2.25)
Szr (t + 1 ) = Szr + ( 1 - Szr HSur ( 1 - gur ) + ( 1 - Sur)Suz ( 1 - guz )su] (2.2 6 )
so that an upper particle is made invisible from the lower cells when it
stands on the ground.
The behavior of this rule is illustrated in figure 2. 1 8, which represents
a simulation of an hour-glass. We can observe that there is an angle of
repose of 45 degrees (slope for which a pile is stable and no avalanche
occurs). This angle is an obvious consequence of the lattice structure.
46 2 Cellular automata modeling

Note that more sophisticated models can be considered for describing

the flow of granular media. We shall discuss later in this book a model
of snow or sand transport by wind in which a different rule is adopted.
The cellular automata model developed by A. Karolyi and J. Kertesz [70]
contains many very interesting additional ingredients, such as dissipation
and static friction.
2.2. 7 The ant rule
The ant rule is a cellular automaton invented by Chris Langton [7 1] and
Greg Turk which models the behavior of a hypothetical animal (ant)
having a very simple algorithm of motion. The ant moves on a square
lattice whose sites are either white or gray. When the ant enters a white
cell, it turns 90 degrees to the left and paints the cell gray. Similarly, if it
enters a gray cell, it paints it white and turns 90 degrees to the right.
It turns out that the motion of this ant exhibits a very complex behavior.
Suppose the ant starts in completely white space. After a series of about
500 steps where it essentially keeps returning to its initial position, it
enters a chaotic phase during which its motion is unpredictable. Then,
after about 10 000 steps of this very irregular motion, the ant suddenly
performs a very regular motion which brings it far away from where it
started.
Figure 2. 1 9 illustrates these last two phases of motion. The path the ant
creates to escape the chaotic initial region has been called a highway [72] .
Although this highway is oriented at 45 degrees with respect to the lattice
direction, it is traveled by the ant in a way which makes one think of a
sewing machine : the pattern is a sequence of 104 steps which repeated
indefinitely.
The Langton ant is another example of a cellular automaton whose rule
is very simple and yet generates a complex behavior which seems beyond
our understanding. Somehow, this fact is typical of the cellular automata
approach : although we do know everything about the fundamental laws
governing a system (because we set up the rules ourselves !), we are often
unable to explain its macroscopic behavior. That is the opposite of
the usual scientific procedure : in principle, the physicist has access only
(through experiments) to the global properties of a system. From them,
he tries to find a universal law. The cellular automata example shows that
a fundamental law is quite important from a philosophical point of view,
but not sufficient. Complete knowledge of a physical process requires both
a microscopic and a macroscopic level of understanding.
To come back to Langton's ant, a general result about this rule is that,
during its motion, the ant visits an unbounded region of space, whatever
the initial space texture is (configuration of gray and white cells). This
observation is proved in a theorem due to Bunimovitch and Troubetzkoy.
 2.2 Modeling of simple systems : a sampler of rules 47

t=6900 t=1043 1 t=l 2000

Fig. 2.19. Langton's ant rule. The motion of a single ant starts with a chaotic
phase of about 10 000 time steps, followed by the formation of a highway. The
figure shows the state of each lattice cell (gray or white) and the ant position
(marked by the black dot). In the initial condition all cells are white and the ant
is located in the middle of the image.

The proof is as follows : suppose the region the ant visits is bounded.
Then, it contains a finite number of cells. Since the number of iterations
is infinite, there is a domain of cells that are visited infinitely often.
Moreover, due to the rule of motion, a cell is either entered horizontally
(we call it an H cell) or vertically (we call it a V cell). Since the ant turns
by 90 degrees after each step, an H cell is surrounded by four V cells, and
conversely. As a consequence, the H and V cells tile the lattice in a fixed
checkerboard pattern. Now, we consider the upper rightmost cell of the
domain, that is a cell whose right and upper neighbor are not visited. This
cell exists if the trajectory is bounded. If this cell is an H cell (and is so for
ever), it has to be entered horizontally from the left and exited vertically
downward and, consequently must be gray. However, after the ant has
left, the cell is white and there is a contradiction. The same contradiction
appears if the cell is a V cell. Therefore, the ant trajectory is not bounded.

Multi-ant motion. Strictly speaking, the motion rule of Langton's ant

is not sufficiently complete to specify a cellular automaton model in the
sense we defined it. Each cell may, in principle, be occupied by a different
ant because any initial configuration of two bits per cell (color and ant)
should be acceptable. However, if many ants are simultaneously present,
they may want to enter the same site at the same time, from different sides.
Langton's rule can be generalized to account for multi-ant situations.
48 2 Cellular automata modeling

Following the same idea as in the HPP rule, we will introduce n i ('r, t)
as a Boolean variable representing the presence (n i = 1 ) or the absence
(n i = 0) of an ant entering site r at time t along lattice direction Ci, where
c 1 , c2 , c3 and c4 stand for direction right, up, left and down, respectively.
Up to four ants may enter the same site at the same time, from the four
sides. If the color Jl (r , t) of the site is gray (/1 = 0), all entering ants turn
90 degrees to the right. On the other hand, if the site is white (/1 = 1 ),
they all turn 90 degrees to the left.
This motion can be represented by the following expression
(2.27)
where the index i is wrapped around the values 1 to 4 (i.e i + 1 = 1 when
i = 4, and i - 1 = 4 when i = 1 ). The color of each cell is modified
after one or more ants have gone through. Here, we choose to make this
modification depend on the number of ants present
Jl(r, t + 1 ) = Jl(r, t ) EB n r ( r, t ) EB n2 ( r, t) EB n 3( r, t) EB n4 ( r, t) (2.28)
where EB is the sum modulo 2.
When the initial conditions are such that the cell contains one single ant,
rules 2.27 and 2.28 are equivalent to the original Langton's ant motion.
When several ants travel simultaneously on the lattice, both cooperative
and destructive behavior are observed. First, the erratic motion of several
ants favors the formation of a local arrangement of colors allowing the
creation of a highway. One has to wait much less time before the first
highway. Second, once a highway is created, other ants may use it to
travel very fast (they do not have to follow the complicated pattern of the
highway builder). In this way, the term "highway" is very appropriate.
Third, a destructive effect occurs as the second ant reaches the highway
builder. It breaks the pattern and several situations may be observed.
For instance, both ants may enter a new chaotic motion, or the highway
may be traveled in the other direction (note that the rule is time reversal
invariant) and is thus destroyed. Figure 2.20 illustrates the multi-ant
behavior.
The problem of an unbounded trajectory occurs again with this gener­
alized motion. The assumption of Bunimovitch-Troubetzkoy's proof no
longer holds in this case because a cell may be both an H or a V cell.
Indeed, two different ants may enter the same cell, one vertically and the
other horizontally.
Actually, the theorem of unbounded motion is wrong in several cases
when two ants are present. Periodic motions may occur when the initial
positions are well chosen. Again, it comes as a surprise that, for many
initial conditions, the two ants follows periodic trajectories.
For instance, when the relative location of the second ant with respect to
 2.2 Modeling of simple systems : a sampler of rules 49

t=2600 t= 4900 t=8564

Fig. 2.20. Motion of several Langton's ants. Gray and white indicate the colors
of the cell at the current time. Ant locations are marked by the black dots.
Initially, all cells are white and a few ants are randomly distributed in the central
region, with random directions of motion. The first highway appears much earlier
than when the ant is alone. In addition the highway can be used by other ants
to travel much faster. However, the "highway builder" is usually prevented from
continuing its construction as soon as it is reached by following ants. For instance,
the highway heading north-west after 4900 steps is destroyed. A new highway
emerges later, as we see from the snapshot at time t = 8564.

the first one is (�x, �y ) = (2, 3), the two ants returns to their initial position
after 478 iterations of the rule (provided they started in a uniformly black
substrate, with the same direction of motion). A very complicated periodic
behavior is observed when (�x. �y) = ( 1 , 24) : the two ants start a chaotic­
like motion for several thousands of steps. Then, one ant builds a highway
and escapes from the central region. After a while, the second ant finds
the entrance of the highway and rapidly catches the first one. After the
two ants meet, they start undoing their previous path and return to their
original position. This complete cycle takes about 30 000 iterations.
More generally, it is found empirically that, when �x + �y is odd and
the ants enter the site with the same initial direction, the two-ant motion
is likely to be periodic. However, this is not a rule and the configuration
(�x, �y) = ( 1 , 0) yields an unbounded motion, a diamond pattern of
increasing diameter which is traveled in the same direction by the two
ants.
It turns out that the periodic behavior of a two-ant configuration is
not so surprising. The cellular automata rule 2.27 and 2.28 is reversible
50 2 Cellular automata modeling

in time, provided that there is never more than one ant at the same site.
Time reversal symmetry means that if the direction of motion of all ants
are reversed, they will move backwards through their own sequence of
steps, with the opposite direction of motion. Therefore, if at some point
of their motion the two ants cross each other (on a lattice link, not on
the same site), the first ant will go through the past of the second one,
and vice versa. They will return to the initial situation (the two ants being
exchanged) and build a new pattern, symmetrical to the first one, due to
the inversion of the directions of motion. The complete process then cycles
for ever. Periodic trajectories are therefore related to the probability that
the two ants will, at a some time, cross each other in a suitable way. The
conditions for this to happen are fulfilled when the ants sit on a different
sublattice (black or white sites on the checkerboard) and exit two adjacent
sites against each other. This explain why a periodic motion is likely to
occur when �x + �y is odd.

Time reversal invariance. Time reversal is an important concept in the

framework of cellular automata modeling because it is a very fundamental
symmetry of the physical world. Checking whether or not a rule is
time reversible can be done by considering the microdynamics of the
automaton. In our case, suppose the state of the automaton, when the
ants enter a site ' is the five-component vector s in = (11rin ' n i1n ' n2in ' ni3n ' n4in ) " As
a result of the color of the site, the directions of motion change and the
ants exit the site in other directions and modify the color of the site. Using
the updating rule, the output state sou t can be computed as the right-hand
side of equations 2.27 and 2.28. We shall write this evolution with an
operator U such that

(2.29)
If the dynamics is time reversal invariant, one can change the direction of
motion of all ants and the system traces its own past, with the opposite
directions of motion. Let us introduce the operator R as the operator
which reverses the directions of motion but does not change the color
( Rs ) o = so
(Rs)i = S i-2 (2.30)
Time reversal invariance implies that U Rsou t = Rsin since if one reverses
the directions of motion and updates the rule, one expects to obtain the
initial state with reverse velocities. Since so ut = Usin , the condition for
time reversal invariance reads
URU = R or RU = u- 1 R (2.3 1 )
 2.2 Modeling of simple systems : a sampler of rules 51

I t is easy to check that one-ant configurations satisfy the time reversibility

condition. For instance

On the other hand, some two-ant configurations do not satisfy the condi­
tion, because of the color rule 2.28

Clearly, time reversal symmetry can be restored by modifying rule 2.28 as

follows

f.l(t + 1 ) = { f.l(t) f.l(t)

1 -
if no ant is present
if there is at least one ant
(2.34)

What about real ants ? So far we have not considered the behavior of
real ants and how they may be connected to this cellular automata. As
mentioned before, artificial life is a fascinating domain to which cellular
automata models may contribute. The initial purpose of Langton's ant
model was apparently to study large-scale regularities in complex systems.
However, this rule has some analogies with what happens in the real ant
world.
It is well known that real ants leave behind them a substance (the
pheromone) which is used by other ants as a stimulus. The idea of modi­
fying the space texture (white or gray) captures an aspect this phenomena.
We have seen that the path created by an ant involve a spatial structure
that can be used by a companion to travel much more efficiently. It is
very tempting to relate this behavior with a famous experiment in which
real ants are able to find the shortest path to their food, as the result of a
collective effect.

2.2.8 The road traffic rule

Road traffic flow is an every day life problem which can be investigated (at
least partially) with cellular automata models [69,73,74] . Cars are treated
as point particles moving along a line of sites.
52 2 Cellular automata modeling

The microdynamic law of motion is taken from Wolfram's rule 1 84. A

car can move only if its destination cell is free. Even if this destination
cell will be freed by an occupying car at the time the motion occurs, the
first car cannot see it because its horizon of sight is finite and the vehicle
stays at rest until the cell in front of it is empty.
This cellular automata rule of motion can be expressed by the following
relation
(2.35)
where n i (t) is the car occupation number (n i = 0 means a free site, n i = 1
means a vehicle is present at site i). The quantity nln (t) represents the state
in the cell from which a car could come and nfut (t) indicates the state of
the destination cell. Rule 2.35 means that the next state of cell i is 1 if
a car is currently present and the next cell is occupied, or if no car is
currently present and a car is arriving.
The number of vehicles N mv t (t) that have moved between iteration t
and t + 1 are those which find n ou t empty
L
Nmv t (t) = L n i (t)(1 - nfu t (t)) (2.36)
i=l
where L is the system size (number of cells). The number of moving cars
is also given by the number of empty cells which will receive a new car
L
Nmv t (t) = L n ln (t)(1 - ni (t)) (2.37)
i=l
Therefore, the equation 2.35 of motion can be written as a balance
equation
L L
L ni (t + 1) - ni(t) = L [nin (t)(1 - ni(t)) - ni(t)(1 - nfut (t))] (2.38)
i=l i=l
For a simple one-dimensional ring-like road, the expressions for nin and
nout are very simple
(2.39)
and the microdynamics reduces exactly to Wolfram's rule 1 84.
In order to make the dynamics more interesting, some ingredients can
be added to the basic rule. For instance, a car can be blocked at some
given site [75] and released with some probability or only at some given
time steps, in order to mimic a traffic light or a slowing down due to road
construction or any other local events.
Another interesting problem is to generalize the rule to a two-dimensional
road network. The question is to define the motion rule of the cars at a
 2.2 Modeling of simple systems : a sampler of rules 53

Fig. 2.21 . Example of a traffic configuration near a junction. The four central
cells represent a rotary which is traveled counterclockwise. The gray levels indicate
the different traffic lanes : white is a northbound lane, light gray an eastbound
lane, gray a westbound lane and, finally, dark gray is a southbound lane. The
dots labeled a,b,c,d,e,f,g and h are cars which will move to the destination cell
indicated by the arrows. Cars without an arrow are forbidden to move.

road junction. This can be realized in a simple way if one assumes that a
rotary is located at each crossing. Road junctions are formed by central
circles around which the traffic moves always in the same direction. A
vehicle in a rotary has priority over any entering car.
The implementation we propose for this rule is illustrated in figure 2.21,
where a four-corner junction is shown. The four middle cells constitute the
rotary. A vehicle on the rotary (b or d) can either rotate counterclockwise
or exit. A local flag f is used to decide the motion of a car in a rotary.
If f = 0, the vehicle (d) exits in the direction allowed by the color of its
lane (see figure caption). If f = 1, the vehicle moves counterclockwise,
like b. The value of the local turn flag f can be updated according to the
modeling needs : it can be constant for some amount of time to impose a
particular motion at a given junction, completely random, random with
some bias to favor a direction of motion, or may change deterministically
according to any user-specified rule. For instance, figure 2.22 shows a
turn flag configuration which forces cars to move horizontally after the
junction.
As before, a vehicle moves only when its destination cell n ou t is empty.
Far from a rotary, the state of the destination cell is determined as the
occupation of the down-motion cell. This is also the case for a vehicle
turning in the rotary. On the other hand, a car wanting to enter the
rotary has to check two cells because it has no priority. For instance, car c
54 2 Cellular automata modeling

Fig. 2.22. Example of a "traffic-light" configuration. The values indicated on the

rotary are the turn flags f. Here, horizontal motion has the priority and vertical
traffic has to turn right or left according to a regulation "turn right on red traffic
light is permitted".

cannot enter the rotary because b is going to move to the white cell. Car e
cannot move either because it sees b (and cannot know whether or not b
will actually move). Car a, on the other hand can enter because it sees
that d is leaving the rotary and that the gray cell ahead is free.
Similarly, the incoming vehicle n in to a given cell is computed differently
inside and outside of the rotary. The light gray cell occupied by car b has
two possible inputs : in priority, it is the vehicle from the gray cell at west ;
if this cell is empty, the input will be the incoming lane, namely the car
labeled e.
Figure 2.23 shows two traffic configurations in a Manhattan-like city.
The difference between these two situations is the behavior at a rotary.
In the first case (a), on each rotary cell, a vehicle has a probability 1 /2
to exit. In (b), the turn flag f has an initial random distribution on the
rotary. This distribution is fixed for the first 20 iterations and then flips
to f = 1 f for the next 20 steps and so on. In this way, a junction acts
-

as a traffic light which, for some amount of time, allows only a given flow
pattern. We observed that the global traffic pattern is different in the two
cases : it is much more homogeneous in the random case.
A simple question that can be asked about this traffic model concerns
the so-called jamming transition. When the car density is low, all vehicles
 2.2 Modeling of simple systems : a sampler of rules 55

( a) (b)

Fig. 2.23. Traffic configuration after 600 iterations, for a car density of 30%.
Situation (a) corresponds to an equally likely behavior at each rotary junction,
whereas image (b) mimics the presence of traffic lights. In the second case, car
queues are more likely to form and the global mobility is less than in the first
case.

can move because there is a high probability that the next site will be
empty. However, if the density of cars is made large enough, there will
be an increasing number of vehicles that are blocked by a car occupying
the next cell. Actually, it is expected that above a critical car density P c ,
the average velocity of the vehicles in the road network depends on the
total number of cars and decreases as the car density is increased. The
traffic congestion can be expressed by measuring how the average velocity
(v) and the global traffic flow p(v) depend on the average car density.
Precisely, p and (v) are defined as

p = N/L

where L is the total length of the road network and N = I: i n i the total
number of cars.
We observe in figure 2.24 that above a critical car density (which is
quite small when the number of crossings is large), the traffic congestion
increases. Below this critical value, the cars move freely. We also observe
that the strategy adopted at each crossing affects the global traffic flow.
Figure 2.24 (a) corresponds to a case of "free rotary," where the direction
of motion is chosen randomly at each rotary cell. Three curves are
represented, corresponding to the length of the road segments between
56 2 Cellular automata modeling
( a) ( b)
0. 35 c--,-----,--,

o ������
car density car density
0

Fig. 2.24. (a) Plot of the average car velocity versus car density in the road
network for various road spacings. (b) Plot of the traffic flow p (v) for the minimal
road spacing allowed in the model (four cells) and various strategies of motion.

two successive intersections. For the case of a road network spacing equal
to 256, we observe a clear second-order transition from a free moving
phase to a situation of slowing down (partial j amming).
In Figure 2.24 (b) the road spacing is fixed at 4 and three different
strategies of motion are considered at the rotary : (i)"traffic light" where
the behavior at each rotary alternates between two possible paths over a
period of 20 iterations ; (ii) "free rotary," as in figure (a) and (iii) "flip-flop"
which refers to a situation where each car flips the turn flag f as it moves
across a rotary cell. In other words, a vehicle imposes the behavior of the
next car to come. The initial distribution of f is random. The flip-flop
strategy is the most efficient (it permits a higher traffic flow than the
others) since it balances the traffic load equally throughout the network.

2.2.9 The solid body motion rule

In the previous section, we have presented several cellular automata
models where point particles are moving on a lattice (HPP rule, sand
rule or traffic rule). Here, we shall present a rule allowing the motion of
more complex objects, such as a "solid body" which is composed of many
constituent particles and has a coherent structure over large distances.
The problem of solid body motion is quite difficult to address in
the framework of cellular automata modeling and, so far, only one­
dimensional objects or "strings" can be defined [76] . Strings consist of a
chain of pointlike particles which can travel in a three-dimensional space
and maintain their size and integrity during their motion. Deformations
 2.2 Modeling of simple systems : a sampler of rules 57

Fig. 2.25. A string configuration with five particles.

are allowed (these objects are not rigid), but the particles composing them
should not spread out in the entire space.
Here, we will restrict our attention to the motion of these strings in a
one-dimensional space and define collision rules between them.
A string can be thought of as a chain of masses linked by springs. More
precisely, it is composed of two kinds of particles, say the white ones
and the black ones, which alternate along a line. Two successive particles
along the chain are either nearest neighbors or separated by one empty
cell. Figure 2.25 shows such a chain composed of five particles.
The rule of motion is the following. The time evolution has two phases.
First, the black particles are held fixed and the white ones move according
to the rule below. Then, the white particles are held fixed and the black
ones move with the same prescription.
Let us consider the case where the white particles move. The rule of
motion is the following : the new configuration of the string is obtained
by interchanging the spacings that separate the white particles from their
two adjacent black neighbors at rest.
In other words, the motion of a white particle consists of a reflection
with respect to the center of mass of its two black neighbor particles. If
q(t) is the position of a white particle on the lattice, and q+ and q_ are
the positions of the right and left adjacent black particles (right and left
are defined according to the direction of the x-axis), then the new position
q(t + 1) of the white particle is given by
q(t + 1) = q+ + q_ - q(t) (2.40)
Of course, this rule is only valid for a particle inside the string which has
exactly two neighbors. At the extremity, an end particle moves so as to
get closer to the string center if its neighbor is two cells apart. On the
other hand if its neighbor is located on the next lattice cell, it moves away
from the middle of the string. This motion corresponds to a reflection
with respect to q+ + a, where a = 3/2 represents the length of the "spring"
linking the particles of the string together. To prevent the particles from
moving off the lattice, a has to be integer or half integer.
Thus, if the left end of the string is a white particle located at q(t), its
new position will be
q(t + 1) = 2(q+ - a) - q(t) (2.41)
58 2 Cellular automata modeling

� t=O

� t=l

� t=2

� t=3

f--{)e+(}+++O-- t=4

f--()-+-e()+++( t= 5

f----+--Oe+OeO- t=6

� t=7

� t=8

Fig. 2.26. A complete evolution cycle of a string. From t = 0 to t = 1 the black

particles move and the white ones are stationary. Then, the white particles move
while the black ones are held fixed. After t = 8 iterations, the configuration is the
same as at t = 0 and the string has traveled two lattice sites to the right ; we thus
conclude that its speed is 2/8.

and, similarly, if it is the right end, its new position will be

q(t + 1) = 2(q_ + a) - q(t) (2.42)
where q + and q _ refer, respectively, to the positions of the particles next
to the left end point and next to the right end point.
Figure 2.26 shows an evolution over eight iterations of the string defined
in figure 2.25. After these eight iterations, the string configuration is
identical to its initial state and new cycle can start. We observe that
the string has moved two lattice cells to the right during the basic cycle.
Therefore, its velocity is v = 2/ 8.
From a general point of view, mass, momentum and energy can be
associated with strings in simple ways. These quantities turn out to be
adjustable for a given string, and are conserved during the evolution. Mass
is defined as

(2.43)
where the particles of a string have been labeled from 1 to N. Note that
here, the indices refer to the particles and not to the lattice cells.
The momentum of the string is given by
N
P (t) = L mi vi (t) (2.44)
i=l
 2.2 Modeling of simple systems : a sampler of rules 59

where v;(t ) = q;(t + 1 ) - q;(t) is the distance particle i it will travel during
the next time step. Finally, the total string energy is
N- 1
1
E= -
2L
i= 1
[(x;+l - x; - a)2 ] (2.45)

Mass is obviously conserved by the evolution rule because no particle

disappears. Energy conservation is also obvious because X i+1 - x; is 1 or
2 and, with a = 3/ 2, one has (2 - 3/2) 2 = ( 1 - 3/2) 2 = 1 /4. Thus, energy
is proportional to the number of particles of the string. This is no longer
the case if the string is allowed to have a transverse motion, too (that
is, particles can also move along the y-axis in a two-dimensional system).
More generally, energy conservation is due to the fact that the cellular
automata rule moves the particles between symmetric positions.
In order to prove that momentum is conserved, it is convenient to write
the equations of motion as
2(q2 (t) - q 1 (t) - a)
q 1 (t - 1 ) - 2q 1 (t) + q1 (t + 1 )
q;- 1 (t) - 2q;(t) + q i+1 (t)
q;(t - 1 ) - 2q;(t) + q;(t + 1 )
2(qN-l (t) - qN (t) + a)
qN (t - 1 ) - 2qN (t) + qN (t + 1 ) (2.46)
which follows from the fact that at every other time step, q;(t - 1) = q;(t)
for the kind of particles which are held at rest. The above formulation is
manifestly time reversible and has the form of a discrete wave equation.
From 2.44 and 2.46, one has
N
P (t) = L m;(q;(t + 1 ) - q;(t))
i= 1
N- 1
1
2 (q1 (t + 1 ) - q 1 (t)) + L (q;(t + 1 ) - q;(t)) + 21 (qN (t + 1 ) - qN (t))
i =2
1
2 [2q 2 (t) - q1 (t) - 2a - q1 (t - 1 )] +
N- 1
2: [q;- 1 (t) - q;(t) + q;+ l (t) - q;(t - 1 n +
i=2
1
2 [2qN - 1 (t) - qN (t) - 2a - qN- 1 (t - 1 )]
1 N- 1 1
2 [q1 (t) - q 1 (t - 1 )] + L (q;(t) - q;(t - 1 )) + 2 (q N (t) - q N (t - 1))
i=2
= P (t - 1 ) (2.47)
which proves that momentum is indeed conserved during string motion.
60 2 Cellular automata modeling

0 time 512

Fig. 2.27. Elastic collisions between five strings o f different sizes. The system
is periodic with length 256 lattice sites (vertical axis). This figure shows the
trajectory of each string during 5 1 2 successive iterations. The total momentum is
zero.

The possible values of the momentum of a string range between M /2 -

and M /2, where M is the total mass. The smallest possible variation of
momentum for a given string is 11P = ± 1 .
I t i s possible t o define collisions between different strings s o that mass,
momentum and energy is conserved during the interaction. To this end,
we consider only strings which have the same kind of extremities (say,
white particles). Two situations are considered to deal with the conflicting
motion : (i) the left and right ends of two different strings are next to each
other and want to move past each other ; (ii) the left and right ends are
one site apart and want to move to the same site.
A way to solve these conflicts is to prevent the particles from moving
when they want to cross each other or to superpose. Momentum is
conserved in this way because, before collision, the two extremities have
opposite velocity (i.e their total momentum vanishes). Due to the collision,
the two particles are held motionless and, therefore, both have zero
momentum. With this mechanism, two colliding strings bounce back by
exchanging a quantum of their momentum. Note that such a collision may
last several time steps before each string has reversed its initial direction
of motion. These types of collisions may be referred to as elastic collisions.
They are illustrated in figure 2.27 in a situation of five interacting strings
moving in a periodic lattice 256 sites long. As we can observe, the motion
of a string is quite complex, due to the collision. In figure 2.28, we show
the position of the middle string over time. No regularity can be identify
during this interval of time which lasts more than 105 iterations.
Another type of collision can be defined. In case (ii) mentioned above,
it is also possible to superpose the two white extremities of mass 1 /2 and
 2.3 Problems 61
256 .---.---,---.

O L-__L___L___L_ __L___L___L___L___L___ L_�

0 1 3 1 100
time

Fig. 2.28. Erratic motion of the middle string of the previous figure, over a large
period of time.

time 512

Fig. 2.29. Trajectories o f five strings experiencing ballistic aggregation. The first
collision does not lead to aggregation because the appropriate conditions are not
fulfilled.

make one single white particle of mass 1 surrounded by black particles.

As a result, the two strings have stuck to each other. With the same
argument as before, it is easy to show that momentum is conserved during
this procedure. With this type of collision, the strings perform a ballistic
aggregation. Figure 2.29 illustrates this process in a system of five strings.

2.3 Problems

2 . 1 . Some cel l u l a r auto m ata ru l es exh i b it the i m portant s i m p l ify i n g

featu re o f add itive s u pe rpos i t i o n . A r u l e i s s a i d t o be add itive if the
62 2 Cellular automata modeling

evo l ution satisfi es the s u perposition pri n c i p l e so = to E9 uo +--+ s n = t n E9 U n ,

where s0 , to and u0 are th ree d i fferent i n itial co n d i t i o n and Ei3 i s , fo r
i nstance, the s u m mod u l o 2. F i n d , among Wolfra m ' s 256 r u l es , those
w h i c h h ave the above property.
2.2. A tota l istic ru l e is a ru l e w h i c h depends o n l y on the s u m of the
n e i g h bo rs and the s i te itself. Wh i c h of Wo lfra m ' s ru l es are tota l i stic? Is
an add itive ru l e tota l i stic?
2.3. A n o n u n iform ce l l u l ar auto m ata i s a CA whose ru l e is not the same
o n a l l l attice s i tes. A non u n iform cel l u l a r auto m ata may be used to
perfo rm tasks that a u n iform CA (of the same range) can n ot ach i eve.
Cons i d e r Wo lfra m ' s ru l e 3 1 o n a o n e-d i mensional period i c l attice and
" contam i n ate " the syste m by h avi ng one o r more s i tes obeyi n g ru l e 8 1
i nstead o f 3 1 . Show that th i s C A pe rfo rms t h e so-cal l ed syn c h ro n i zati o n
task[77] : that i s , from any i n itial conf i g u ration , t h e syste m conve rges
towards an osc i l l at i n g state where , at t i m e t, a l l s i tes are 1 and at ti m e
t + 1 t h e y a re a l l 0. Fi nd oth e r pai rs o f r u l es that l ead t o t h e same
behavior.
2.4. Wolfra m ' s ru l es can be genera l i zed to i nc l ude k states per ce l l
and a n e i g h borhood of s i ze 2r + 1 , where r i s cal l ed the range of the
rule. Stand a rd Wolfra m ' s ru l e are obta i ned fo r r = 1 and k = 2. What
i s ru l e 20 when r = 2 and k = 2 ? S i m u l ate i ts behavi o r on a co m p ute r.
2.5. Show that i te rat i n g twice an r = 1 , k = 2 Wo lfra m ' s ru l e is eq u i va­
l e nt to an r = 2, k = 2 r u l e .
2 . 6 . Study the stoch astic fo rest fi re cel l u l a r auto m ata i ntrod uced by
D rossel and Schwabl w h i c h is def i n ed as fo l l ows : one starts with a
squ are l attice of s i ze L x L with period ic bo u ndary cond itions. Each
s ite i s in one of the th ree fo l l owi n g states : 0 an e m pty s ite , T a s i te
occ u p i ed by a tree and B a s i te occ u p i ed by a b u rn i ng tree. The
u pdati n g ru l es use the von Neumann n e i g h borhood and read , in te rms
of the p roba b i l ities p, f and g: R 1 : An em pty s i te spro uts a tree with
p robab i l i ty p. R2: A tree catches fi re with p robab i l ity (1 - g), if at
l east o n e n ea rest n e i g hbor tree i s b u rn i n g . R3: A tree catch es fi re
with probab i l ity f(1 - g ) if no nearest n e i g h bor tree is b u rn i n g . R4: A
b u r n i n g tree b u rns down and beco mes an em pty s ite.
Ide ntify the d iffe rent reg i mes of behavior w h i c h are poss i b l e . Ve rify
that, when g = O, p � 1 and f � p, a state of self-organ i zed critical ity
is reached i n wh ich cl uste rs of a l l s i zes are b u rn i n g .
2 . 7 . Co n s i d e r a m u lti-lane traffic mod e l : cars move from l eft to r i g ht
a l o n g a r i n g , accord i n g to Wo lfram ru l e 1 84. I n add iti o n , when a car
i s b l ocked by another one in front, it can j u m p to the nearest l a n e ,
 2.3 Problems 63

p rovi ded t h i s m ove is poss i b l e . What k i n d of n e i g h borhood do you

n eed to i m p l e m e n t t h i s r u l e? S i m u l ate t h i s auto m ato n and meas u re
the f l ow d i ag ra m , assu m i ng some l a n es m ay be s l owed down by a
random obstacl e . Com pare the flow with the s ituation where cars
can not change l a n e .
2.8. Cons i d e r t h e a n t r u l e o f section 2 . 2 . 7 . Def i n e another strategy
to m od ify the " co l o r " of the ce l l afte r seve ral ants have crossed it.
S i m u l ate the model and observe the m od ificati o n with respect to the
origi nal rule.
2.9. Ass u m e that each a n t of t h e model o f section 2 . 2 . 7 h as an i nte rnal
" g e n o m e " which spec i f i es how the co l o r of the ce l l is to be mod ified
after the ant h as gone th roug h . For i nsta nce, i n a genome such as
0 1 1 00 1 a 0 means that the co l o r is not mod ified , whe reas 1 means i t
c h a n g e s . Afte r each t i m e step, t h e n ext d i g it i s read , i n a cyc l i c way.
P ropose a ru l e to deal with seve ral ants at a s i te and s i m u l ate the
model with ants h av i n g various genomes in s i ze and contents.
2 . 1 0 . General ize the ant ru l e by add i ng m o re than two states . Fo r
i nstance, each ce l l can be i n any state 0 , 1 , 2 o r 3 . Afte r the ant
leaves the cel l , the state is i ncreme nted by 1 mod 4. The ant moti on is
determ i ne by the state as fo l l ows : O=l eft tu r n , 1 = r i g ht turn 2=stra i g ht
and 3 reverse d i rect i o n .
2 . 1 1 . H o w l o n g does it takes f o r t h e h o u r g l ass o f section 2 . 2 . 6 to
e m pty, as a fu ncti o n of the h e i g ht of a g ra i n i n the u pper part? Is t i m e
accu rate ly measu red b y t h e n u m be r o f g r a i n s that h ave fa l l e n down?
2 . 1 2. P ropose a ru l e to i m p lement the sand mod e l (or a s i m i l a r " sand "
dynam i cs) without us i n g the Margol us n e i g h borhood . For i nstance,
co n s i d e r a n approach based on the HPP particle moti o n . Fo u r pop­
u l ations of g ra i n are a l l owed at each s ite: g rai ns at rest and th ose
mov i n g ve rtica l l y in d i rection south-west, south or south-east. A g ra i n
m oves p refe renti a l ly to t h e south (g ravity) u n l ess t h e desti n ati on s i te
is a l ready occ u p i e d . When th is is the case, it tries to fa l l l eft or r i g ht, if
poss i b l e and no conf l i ct can occur. Assu m i n g an exc l us i o n p r i n c i p l e fo r
the movi n g d i rection (0 or 1 parti c l es ) , how many rest particles s h o u l d
b e a l l owed on each s i te i n order t o b e certai n o f co nserv i n g t h e tota l
n u m be r of g ra i n s i n any s i tuation?
2 . 1 3. S i m u l ate a pop u l ation of robots mov i n g o n a 1 0 lattice. The ro bots
can go l eft, r i g ht or re mai n stationary. Thei r m otion is determ i ned by a
lookup tab l e , w h i c h m ay be d i fferent fo r each robot. T h i s tab l e i nd icates
whether the robot s h o u l d stop, reverse its d i rection or conti n u e stra i g ht
ahead , accord i n g to the presence o r absence of othe r robots i n the
64 2 Cellular automata modeling
n e i g h borhood . If two (or more) robots arrive on the same s i te they
col l i d e and d i e (they are removed from the l attice) . What is the s i ze of
the n e i g h borhood req u i red to i m plement th is ru l e? What is the s ize of
the lookup tab l es? S u p pose the i nte l l i gence of the robot popu l at i o n is
measu red as the n u m ber of robots than can coexi sts in th e l o n g ti m e
reg i m e ( i . e . when a l l col l i s i o n s h ave take n p l ace) . T h e n , starti n g from
a ran d o m i n itial popu l ation ( random tab l es) , i ncrease the fitness of the
robots by refres h i n g the popu l ati o n : create offs p r i n g obta i ned from a
" m ati n g " of two parents or, more precisely, by m i x i n g the tab l es of two
s u rvivi n g robots .
2. 1 4. The Prisoner 's Dilemma is a famous model of a s i m p l e soc i a l
behav i o r between two perso ns. The pro b l e m is we l l i l l utrated by t h e
case o f two prisoners who can not co m m u n icate before the i r tri a l . I f
they both coope rate ( i . e . do not accuse eac h oth e r o f t h e c r i m e) , they
wi l l be convicted to a l i g h t p u n i s h m e nt. If they both defect ( i . e . accuse
each oth e r of the cri me) they both get a severe p u n i s h ment. If o n e
o f the prisoner defects wh i l e the othe r coope rates , the l atte r goes
to j a i l and fo rmer i s freed . Thus, there are fo u r outco mes: CC, C D ,
DC o r D D where C and D stand fo r Coope rate or Defect and the two
l ette rs des i g n ate the behav i o r of each prisoner. There is a d ifferent
payoff associ ated with each of these outcomes. S i nce the prisoners
can not g u ess i n advance what the oth e r wi l l do, it is mean i ngfu l to
ass u m e that the payoff can be ra n ked as : DC, CC, D D , C D . That i s , the
best c h o i ce i s to defect, provided that the oth er coope rates . A d o u b l e
cooperation is good b u t cooperat i n g may b e q u ite dangerous i n case
the oth e r d efects . A spatial CA ve rs i o n of th is prisoner " ga m e " can
be co n s i d e red. O n a two-d i m e n s i o n a l l attice each s ite rep resents a
p l aye r with o n e of the two strateg i es (C o r D) . D u r i n g each ite ration
of the auto m ato n , each p l ayer i nte racts with its fo u r n e i g h bo rs (vo n
N e u m a n n e i g h bo rhood) and the tota l payoff of the p l ay is co m puted .
T h i s payoff i s cu m u l ated ove r the iterations and g i ves a score to each
p l aye r . Afte r each iterat i o n , the pl ayers adopt the strategy of p l aye r
with the h i g h e r score (sti l l i n the von N e u m a n n e i g h bo rhood ) . Study the
behav i o r of this system ( n u m be r of C and D p l ayers) fo r var i o us payoff
va l ues. Co n s i d e r fi rst a s ituation with DC=9, CC=8, DD=5, CD=3 and
c u m u l ate the score mod u l o 1 6.
2 . 1 5. Co n s i d e r the str i n g ru l e of section 2.2.9 but now on a 2D l attice
and with a transverse motion i nstead of the l o n g itud i nal one. Ass u m e
that success i ve particles occ u py adj acent x-coord i nates whereas the i r
y-coo rd i n ates d iffe r b y Y i - Y i- l E { 0, ±1 } . Odd particles move at odd
time steps and eve n part i c l es move at eve n time steps. The moti o n
i s a l o n g t h e vertical axi s s o a s to reach t h e sym metric pos i t i o n with
 2.3 Problems 65
respect to the center of m ass of the two stand i ng n e i g h bors. Show that
particles wi l l j u m p 0 , 1 or 2 s i tes u p or dow n . What n e i g h borhood do
you n eed to i m p l e m ent this ru l e? S i m u l ate the motion of a period ic
str i n g . What are the poss i b l e speeds you can obta i n?
2. 1 6. F i n d a stri n g ru l e w h i c h co m b i nes both the l o n g itud i n al and the
transverse motion .
2 . 1 7 . Def i n e a m ass and momentu m col l i s i o n mech a n i s m between
H PP-I i ke part i c l es and the str i n g model yo u obta i ned in the previous
prob l e m .
 3
Statistical mechanics of lattice gas

3.1 The one-dimensional diffusion automaton

In the previous chapter, we discussed several cellular automata rules which

definitely had something to do with the description of physical processes.
The question is of course how close these models are to the reality they
are supposed to simulate ?
In the real world, space and time are not discrete and, in classical
physics, the state variables are continuous. Thus, it is crucial to show
how a cellular automaton rule is connected to the laws of physics or to
the usual quantities describing the phenomena which is modeled. This is
particularly important if the cellular automaton is intended to be used as
a numerical scheme to solve practical problems.
Lattice gas automata have a large number of potential applications
in hydrodynamics and reaction-diffusion processes. The purpose of this
chapter is to present the techniques that are used to establish the con­
nection between the macroscopic physics and the microscopic discrete
dynamics of lattice gas automata. The problem one has to address is
the statistical description of a system of many interacting particles. The
methods we shall discuss here are very close, in spirit, to those applied in
kinetic theory : the N-body dynamics is described in terms of macroscopic
quantities such as the particle density or the velocity field. The derivation
of a Boltzmann equation is a main step in this process.
In order to illustrate and understand the main steps leading to a
macroscopic description of a cellular automaton model, we first consider
a linear rule for which a simple and rigorous calculation can be carried
out. From this result we justify the more sophisticated techniques that are
necessary when dealing with a nonlinear dynamics. Hydrodynamic models
such as the FHP automaton or the reaction-diffusion rules, require such
techniques.

66
 3.1 The one-dimensional diffusion automaton 67

3. 1 . 1 A random walk automaton

A random walk describes the motion of a particle which changes its
travel direction in a probabilistic way. Many physical situations exhibit a
diffusive behavior which is due to random motions at a microscopic level.
Here, we shall discuss a probabilistic cellular automaton rule modeling the
simultaneous random walk of many particles in one dimension. We will
consider the two- and three-dimensional cases in much detail in chapter 4.
We consider a one-dimensional system in which particles may travel
left or right. Usually, a random walk is simulated by selecting one
particle in the system and moving it at random to the left or to the right,
provided that the destination cell is empty. When all particle can move
simultaneously, as is the case in a cellular automaton model, this method
meets some difficulties : two particles may decide, at the same time to jump
to the same unoccupied site. Elaborate communication strategies among
the second nearest neighbors are required to solve a double occupancy
situation. This solution also means that the mobility of the particles are
subject to their density, which is not always desirable.
There is, however, a very simple way to produce a simultaneous random
walk compatible with an exclusion principle (i.e no arbitrarily large occu­
pancy). Our evolution rule is inspired by the HPP dynamics described in
section 2.2.5. At each site x of the lattice, we define two Boolean variables
n 1 (r, t) and n2 (r, t). These quantities are occupation numbers indicating
whether or not a particle is entering site r at time t in direction e 1 (right)
or ez (left), respectively.
Random motion is obtained by shuffling the two directions of motion
independently at each lattice site and at each time step. In other words,
what is traveling in direction e 1 will be exchanged with what is traveling in
direction e2 with a probability 1 - p. In order to write the microdynamics
of such a system, that is the rules governing the behavior of this lattice gas
automaton, we need a random bit Jl(r) at each lattice site. We suppose that
there is an external mechanism capable of producing, at each iteration, an
entire set of statistically independent Boolean variables Jl(r) which are 1
with probability p.
Our goal is to determine the macroscopic behavior of our cellular
automaton, that is its behavior in the limit of infinitely short time step r
and small lattice spacing 2 . To this end, it is convenient to introduce these
space and time units in our rule.
The microdynamics has two phases : shuffling and motion to the left or
right nearest neighbor. A particle entering site r +2 at time t with a velocity
pointing to the right must have been at site r at t. With a probability
p, this particle was the one which entered with a velocity pointing to the
right and with probability 1 - p the one which had a velocity pointing to
68 3 Statistical mechanics of lattice gas
the left. Therefore, the random walk rule reads
n 1 (t + r, r + 2) = p,(t, r)n 1 (t, r) + (1 - p,(t, r))nz (t, r)
nz (t + r, r - 2) = (1 - p,(t, r))n l (t, r) + p, (t, r)nz (t, r) (3. 1 )
As we can see p, (r) selects at random either n 1 (r, t) or n z (r, t). Note that
when both n 1 (r, t) nz (r, t) are one, shuffling has no effect, since the particles
are indistinguishable.

3. 1 . 2 The macroscopic limit

The first steps necessary to determine the macroscopic behavior of the
microdynamics 3. 1 is to average it. At a macroscopic scale, the variables
ni no longer appear as Boolean quantities but rather as quantities varying
continuously between 0 and 1. Indeed, a macroscopic point of coordinate
r designates a microscopic volume comprising many particles. Formally,
this averaging Ni(r, t) = (n i(r, t)) is treated as an ensemble average in the
sense of statistical mechanics. In practice, it is an average over space, or
time, or both, through a coarse-grain procedure.
Since p, (r, t) is statistically independent of n i (r, t) (the shuffling mecha­
nism is data blind), the average of ( p,n i) factorizes and we have
(3.2)
3.1 can be averaged and yield
Therefore, relations
N1 (t + r, r + 2) = pN1 (t, r ) + qNz (t, r)
Nz (t + r, r - 2) = qNl (t, r) + pNz (t, r) (3.3)
where we have defined q = 1 - p. These equations are linear and can be
solved exactly on a periodic system : any function defined on a periodic
lattice of size L = M2 is a superposition of plane waves exp(ikr) with
k = 0, (2n/L), ... , (2 n /L)m, ... , (2n/L)(M - 1). We will come back to this
point in more detail in chapter 4 but, for the moment, we will look for
solutions to 3.3 of the following form
N(t, r) = A(t) exp(ikr) (3.4)
N and A are the column matrices
(
where

N(t' r) = N N1 (t, r)
z (t, r)
) (3.5)
A 1 (t) and A z( t) are two functions to be determined. With the relation 3.4,
equation 3.3 reads
(
A(t + r) = pqexp(
exp(zk2) p exp(zkA)
)
-: ik 2) q exp( -: ikA ) A (t) (3.6)
 3.1 The one-dimensional diffusion automaton 69

We then write
A(t) = exp(ico t ) ( ) AO (3.7)

(
and one is left with the condition
p exp(-ikA) - exp(icor)
q exp( ikA)
q exp( -ikA)
p exp( ikA) - exp( icor)
) A(O) = O (3.8)

A solution A(O)
=I= 0 exists provided that the determinant of this matrix
is zero, which gives a relation between co and k (the dispersion relation).
This relation is
exp(2icor) - 2p cos(U) exp(icor) + 2p - 1 = 0 (3.9)
and can be solved for exp(icor)

exp(ico ± r) = p cos(U) ± ( 1 - p) 1-
2
� sin2 (kA)
(p 1 ) 2
(3. 10)

A second step when taking the macroscopic limit is to make the lattice
spacing A and the time step r tend to zero. The quantity giving the time
evolution is the factor exp ico ± t = exp(ico ± r) tl r. Therefore, one has to
compute

- t / r) = lim
lim (exp(ico+r)
r--+0)--+0 r--+0)--+0
[
p cos(kA) ± ( 1 - p) 1 -
2
� sin2 (U
(p 1 ) 2
f'
(3. 1 1 )
This limit is easy to compute for co_, since, when A ---+ 0, the term in the
square bracket goes to 2p - 1 . As 0 < p < 1, (2p - 1 ) t/ r goes to zero when
r ---+ 0.
For co+, this limit is of the type "1 00 ". This indetermination can be
removed by taking the logarithm of the expression and differentiating
both the numerator and the denominator with respect to r.
1. ln(exp(ico+ r)) 1" a, ln(exp(ico+ r))
lm t t r--+0)--+0
lm
r--+0)--+0 r a, r
t r--+0)--+0
lim a, ln( exp( ico+ r)) (3. 12)

To carry out this calculation we need to specify how A and r are related
when they go to zero. The quantity ln( exp( ico+ r)) as given by equation 3. 10
is an expression containing A. Therefore, it is more convenient to write

t r--+0)--+0
lim a, ln(exp(ico+r)) = t r--+0)--+0
lim exp(-ico+r)aA. exp(ico+r)
dA
dr
(3. 13)

By differentiating equation 3. 10, we obtain that the indetermination left

70 3 Statistical mechanics of lattice gas

[ ]
in the calculation is

t lim (\. ln(exp(iw+ r)) = t lim -pk sin kA -

o-+O,.l.-+0 .l.-+0 ( - p)
tp2
d't"
sin kA ( ) (3. 14)
dA.

because the other terms gives factors of 1 in the limit A. �_. . o

There are two natural choices to compute (dA.jdr) in the macroscopic
limit :
(a) A.jr = v = const =I= 0 which reflects a situation where the propagation
speed v of the particles remains finite. Therefore dA. = vdr and the
limit 3 . 1 4 gives 0. Consequently, we have
[exp(iw+ r)] tl' � 1 (3. 1 5)
and the system appears frozen since A(t) = A (O).

(b) A more interesting case is obtained when 22 jr � const =I= 0. In this

limit, the time step goes faster to 0 than the lattice spacing. We have
2AdA. = const dr, that is
dA. _!_ 22
=( ) (3. 1 6)
dr 22 r
The singularity in limit 3 . 1 4 is (sin kA./2) � k and we obtain
A_ 2
[exp(iw+r)] tl ' � exp(-tk 2
p
�) (3. 1 7)
2( 1 - p)
Therefore, in this limit, there is a quantity p(r, t) which evolves as
p(r, t) = L exp(ikr) exp(-tDk2 ) (3. 1 8 )
k
where D (the diffusion coefficient) is defined as
p A_2
-
2(1 - p) 't"
D= ( 3 . 1 9)

It turns out that p is precisely a solution of the one-dimensional diffusion

equation
(3.20)
Therefore, the simultaneous random walk rule we have defined here
simulates the behavior of a diffusive process in the continuous limit
(A.2 jr) =const. This is an expected result, of course, since diffusion
corresponds to a microscopic random motion. However, this very simple
example illustrates that a simplified microscopic dynamics may very well
capture the essential features of a macroscopic reality.
 3.1 The one-dimensional diffusion automaton 71
3.1.3 The Chapman-Enskog expansion
The derivation given in the previous section is possible because equa-
. tion 3.3 is linear. This is usually not the case and no explicit solution can
be expected to the lattice Boltzmann dynamics. Instead, we shall try to
obtain the differential equations describing, in the continuous limit, the
behavior of the macroscopic quantities such as the density of particle and
so on.
The Chapman-Enskog expansion technique is commonly used in
statistical mechanics to derive the macroscopic laws governing the
relevant physical quantities when the Boltzmann equation is known. The
Boltzmann equation is an equation for the quantity f(r, v, t) where
f(r, v, t)dr dv gives, at time t, the average number of particles in the
region dr around r and having a velocity between v and v + dv.
The quantity Ni defined above plays the role of the Boltzmann density
function f, except that the velocities are labeled with a discrete index
i instead of a continuous variable v. Equation 3.3 is then the lattice
Boltzmann equation of random walk microdynamics.
The local density of particle is the integral of f over the velocity variable.
In our case the density p reads
2
p(r, t) = L Ni(r, t) (3.21)
i=l
The idea o f the Chapman-Enskog expansion i s the following : t o a first
approximation, one expects that N1 = N2 = p / 2. Then, there are small
corrections that are found with perturbation analysis. More precisely, we
shall wFite
(3.22)
where e is a small parameter and the Nf!' l s are functions of r and t to be
determined.
The first step of the Chapman-Enskog expansion is to write equation 3.3
as
N1 (t + -r, r + A) - N1 (r, t) - (p - l)(N1 (t, r) - N2 (t, r))
N2 (t + -r, r - A) - N2 (r, t) (p - l)(N2 (t, r) - N1 (t, r)) (3.23)
By summing these two equations, we obtain
N1 ( t + r, r + A) + N2 ( t + r, r - A) - p(r, t) = 0 (3.24)
The fact that the right-hand side vanishes is not by accident. It reflects the
fact that the number of particles is locally conserved when the directions
are shuffled : the number of particles entering site r at time t are exiting at
72 3 Statistical mechanics of lattice gas
time t + r. For this reason, equation 3.24 is called the continuity equation.
This equation will play a crucial role in what follows.
The next step is to consider a Taylor expansion of the left-hand side of

� af + Aciar � er a; + dci atarl Ni (t, r)

equation 3.23. One has

[
Ni ( t +r, r +ACi) - Ni ( r, t) = r at + +

(3.25)
where we have defined c 1 = - c2 = 1 and neglected third-order terms in
the expansion.
From the calculation we performed in section 3 . 1 .2, we know that there
is only one way to take the continuous limit A � 0 and r � 0, namely
that r "'A2 . As a consequence, A and r are not of the same order of
magnitude. For the sake of convenience, we shall write
(3.26)
We can now compare, order by order, the two sides of equation 3.23.
Using relations 3.22 and 3.25, we have, for the order O(c0 )
N �O) = N�O) (3.27)
because the left-hand side is of order c, at least. The equation has no
unique solution. The key to the Chapman-Enskog expansion is to impose
the condition that the density is completely expressed in terms of the
p
Nl(O) s
2
P = L Nfo) (3.28)
i= l
Consequently, we have

(3.29)

and the condition that

2
2: NV) = 0 if t ;:::: 1 (3.30)
i= l

�l
The next order O(c) of the Boltzmann equation is iven by taking the
term Acia r N}0) of the Taylor expansion and the term Ni ) in the right-hand
side of equation 3.23. Since N}0l = /2, we obtain
p
A ( - 11 ) arp = (p - 1 ) ( -11 -1
1
)( � )
N( l )
( 3.3 1 )
2 N l)
 3.1 The one-dimensional diffusion automaton 73
This equation has a solution because the vector ( 1, - 1) is an eigenvector
of the right-hand side matrix
( - 11 - 11 ) ( -11 ) - 2 ( - 11 ) (3.32)
Since this matrix is not invertible, it is not expected that an equation
such as equation 3.3 1 always has a solution. However, the magic of the
Chapman-Enskog expansion is that, precisely, the left-hand side of the
equation is in the image of the matrix. We will return to this problem in
more detail in section 3.2. For the present time, it important to notice that
the reason why the matrix is not invertible is particle conservation.
The solution of equation 3.31 is then straightforward
A
Ni(1 l = C iOrP (3.33)
4(p _
1)
The equation governing the evolution of is given by the continuity
p
equation 3.24. The order O(c 1 ) reads
2
L ACi Or N}0l = 0 (3.34)
i= 1

[
which is obviously satisfied by our solution. The next order is

£;2 TOtNi(0) + AC iOr Ni( 1 ) + 2

Using the solution for N}0l and N}lJ, we obtain the equation
l
A1 2 ci2 ar2 Ni(0) - 0 _

(3.35)

a p+
,
A2 1
+
(1
2
a; p = o
) (3.36)
� 2(p - 1)
and, finally
at P = v a ; P (3.37)
which is the expected diffusion equation with the diffusion constant
A p2
D=� 2( 1 _
p)
(3.38)
The derivation here is particularly simple because the microdynamics of
the random walk cellular automaton is simple. For more complex rules,
such as the FHP rule we shall discuss in section 3.2, the principle of
the expansion will be identical but involves more complex steps that
were trivially verified here. For instance, the relation between A and T
is not known. In general, phenomena at various scales may be present.
In addition to diffusive behavior, convective phenomena may appear at
a scale where A "' T. In order to differentiate between these various
74 3 Statistical mechanics of lattice gas
orders of magnitude, a multiscale has to be used in conjunction with the
Chapman-Enskog method.

3.1 .4 Spurious invariants

Conserved quantities or invariants generally play a very important role
in physics and, in particular, in cellular automata modeling. Local con­
servation laws at the level of the automaton rule are still visible at the
macroscopic scale of observation. The diffusion equation obtained in the
previous section is a continuity equation
Dt P + divJ = 0 (3.39)
where J is the particle current whose expression is obtained by a Taylor
expansion of equation 3.24 and retaining only terms of order E

J � (N1( 1 ) - N ( 1 ) ) + A2 or (N1(O) + N (O) )

r 2 2r 2
-Dor p (3.40)
The expression for J depends on the microscopic nature of the particle
motion, but the continuity equation itself is only a consequence of the
conservation of particle numbers.
Are there any other conservation laws in our microdynamics that we
have not extracted with the Chapman-Enskog expansion and which could
play a role in a numerical simulation ? The answer is yes : there is a spurious
invariant in the random walk cellular automaton rule presented here. To
find it, let us return to the dispersion relation 3.10 :

exp(2iwr) - 2p cos (kA ) exp (iwr) + 2p - 1 = 0 (3.41)

A quantity will b e conserved i f II exp(iwr) ll = 1. The condition exp(iwr) =
1 implies
cos (kA) = 1 (3.42)
which is satisfied for kA = 2n n . Since k only takes the values
k = 0, (2n/ L), ... , (2 n / L)m, ... , (2n/ L)(M - 1)
the solution o f equation 3.42 i s k = 0.
If one expresses the density p as
p(r, t) = L Pk e i(w t+kr ) with Pk = L p ( r, O)e- ikr (3.43)
k rElattice
one sees that the physical quantity associated with the wave number k is
the total number of particle Po = L r p (r) = N.
 3.2 The FHP model 75

Equation 3. 10 has another solution for exp(iwr) = - 1 which is

cos(kA) = - 1 (3.44)
Within the possible range of values of k, this implies that kA = n.
Therefore, in our dynamics, there is a quantity which changes its sign at
each iteration. The quantity associated with kA = n is
Pn !). = L p(r, O ) e-in (r j).) (3.45)
rElattice
Thus,
Pn !). = Neven - Nodd (3.46)
where Neven and Nodd denote the total number of particles on even and
odd lattice sites, respectively. Our result is easily explained by the following
fact : all particles entering an even site at time t will move to an odd site
at time t + 1, and conversely. Therefore, there is no interaction, at any
time, between particles that are not located on the same odd or even
sublattice. Particle conservation holds independently for each sublattice
and our system actually contains two non-interacting subsystems. This
feature is known as the checkerboard invariant.
This spurious invariant may have undesirable effects in reaction-diffusion
processes, such as the annihilation reaction A + A � 0 in which two par­
ticles entering the same site are removed with a probability P reaction · If the
odd and even sublattices are not well distinguished, the amount of A par­
ticles will eventually reach a constant concentration because no reaction
can take place between particles on different sublattices.

3.2 The FHP model

The FHP rule is a model of a two-dimensional fluid which was introduced

by Frisch, Hasslacher and Pomeau [32] , in 1 986. We show here how the
fully discrete microscopic dynamics maps onto the macroscopic behavior
of hydrodynamics. Due to the nonlinearity of the rule, the calculation will
be much more complex than in the case of the one-dimensional diffusion
rule discussed in the previous section.

3.2. 1 The collision rule

The FHP model describes the motion of particles traveling in a discrete
space and colliding with each other, very much in the same spirit as the
HPP lattice gas discussed in section 2.2.5. The main difference is that, for
isotropy reasons that will become clear below, the lattice is hexagonal (i.e.
each site has six neighbors, as shown in figure 3. 1 ).
76 3 Statistical mechanics of lattice gas

Fig. 3. 1 . Two-body collision in the FHP model. On the right part of the figure,
the two possible outcomes of the collision are shown in dark and light gray,
respectively. They both occur with probability 1 /2.

The FHP model is an abstraction, at a microscopic scale, of a fluid.

It is expected to contain all the salient features of a real fluid. It is well
known that the continuity and Navier-Stokes equations of hydrodynamics
express the local conservation of mass and momentum in a fluid. The
detailed nature of the microscopic interactions does not affect the form
of these equations but only the values of the coefficients (such as the
viscosity) appearing in them. Therefore, the basic ingredients one has
to include in the microdynamics of the FHP model are the conservation
of particles and momentum after each updating step. In addition, some
symmetries are required so that, in the macroscopic limit where time and
space can be considered as continuous variables, the system is isotropic.
As in the case of the HPP model, the microdynamics of FHP is given
in terms of Boolean variables describing the occupation numbers at each
site of the lattice and at each time step (i.e. the presence or the absence
of a fluid particle). The FHP particles move in discrete time steps, with a
velocity of constant modulus, pointing along one of the six directions of
the lattice. The dynamics is such that no more than one particle enters the
same site at the same time with the same velocity. This restriction is the
exclusion principle ; it ensures that six Boolean variables at each lattice
site are always enough to represent the microdynamics.
In the absence of collisions, the particles would move in straight lines,
along the direction specified by their velocity vector. The velocity modulus
is such that, in a time step, each particle travels one lattice spacing and
reaches a nearest-neighbor site.
Interactions take place among particles entering the same site at the
same time and result in a new local distribution of particle velocities.
In order to conserve the number of particles and the momentum during
each interaction, only a few configurations lead to a non-trivial collision
 3.2 The FHP model 77

,-'- _y
-7\X,-/. -><
I
�

Fig. 3.2. Three-body collision in the FHP model.

(i.e a collision in which the directions of motion have changed). For

instance, when exactly two particles enter the same site with opposite
velocities, both of them are deflected by 60 degrees so that the output of
the collision is still a zero momentum configuration with two particles (see
figure 3. 1 ). As shown in figure 3 . 1 , the deflection can occur to the right or
to the left, indifferently. For symmetry reasons, the two possibilities are
chosen randomly, with equal probability.
Another type of collision is considered : when exactly three particles
collide with an angle of 120 degrees between each other, they bounce
back to where they come from (so that the momentum after the collision
is zero, as it was before the collision). Figure 3.2 illustrates this rule.
Several variants of the FHP model exist in the literature [3] , including
some with rest particles (models FHP-II and FHP-III).
For the simplest case we are considering here, all interactions come from
the two collision processes described above. For all other configurations
(i.e those which are not obtained by rotations of the situations given in
figures 3 . 1 and 3.2) no collision occurs and the particles go through as if
they were transparent to each other.
Both two- and three-body collisions are necessary to avoid extra conser­
vation laws. The two-particle collision removes a pair of particles with a
zero total momentum and moves it to another lattice direction. Therefore,
it conserves momentum along each line of the lattice. On the other hand,
three-body interactions deflect particles by 1 80 degrees and cause the net
momentum of each lattice line to change. However, three-body collisions
conserve the number of particles within each lattice line.
In the FHP rule, particles interact only at the lattice sites. It is tempting
to think of this dynamics as snapshots, at integer clock cycles, of a
continuous time evolution on a lattice. Between two consecutive time
78 3 Statistical mechanics of lattice gas
steps, the particles move along the lattice links. Strictly speaking, we may
ask what happens to them when they cross each other on a lattice link. It
is not specified in the model and somewhat irrelevant from the technical
point of view. If we want a precise picture of what occurs along the lattice
links, we may either imagine that the particles are transparent to each
other or, since particles are indiscernible, that they collide half way and
bounce back.
Note that a fully deterministic FHP model can also be considered, to
avoid generating random numbers to select among the possible outcomes
of the two-body collision. The usual strategy used for this purpose is
to alternate between the two outcomes. One method consists of using
time parity (i.e at even time steps, a n i n i+3 collision results in a n i+l n i+4
configuration and, at odd time steps, to a n i+5 n i+2 configuration). Another
technique is to add a seventh bit to the automaton state, which acts as a
control bit. When this bit is 1, the n i+l n i+4 configuration is chosen and
the bit is flipped to 0. Conversely, when the seventh bit is 0, the n i+5 n i+2
output is selected and the control bit is flipped to 1 .
A deterministic FHP dynamics is interesting also because it guarantees
time reversibility invariance, which is a fundamental symmetry of micro­
scopic physics. However, in the rest of this section, we shall consider the
probabilistic FHP rule initially described because no macroscopic differ­
ence is expected to show up when one dynamics is used in place of the
other.

3.2.2 The microdynamics

The full microdynamics of the FHP model can be expressed by evolution
equations for the occupation numbers : we introduce n i ('r, t) as the number
of particles (which can be either 0 or 1) entering site r at time t with a
velocity pointing along direction c i , where i = 1, 2, ... , 6 labels the six lattice
directions. The unit vectors c i are shown in figure 3.3.
We also define the time step as T and the lattice spacing as A. Thus, the
six possible velocities v i of the particles are related to their directions of
motion by

(3.47)

Without interactions between particles, the evolution equations for the

n i would be given by
(3.48)
which expresses that a particle entering site r with velocity along c i will
continue in a straight line so that, at the next time step, it will enter site
 3.2 The FHP model 79

Fig. 3.3. The direction of motion ci.

r + Xci with the same direction of motion. However, due to collisions, a

particle can be removed from its original direction or another one can be
deflected into direction c i .
For instance, if only n i and n i+3 are 1 at site r, a collision occurs and
the particle traveling with velocity v i will then move with either velocity
vi - 1 or Vi+ l (note that the operations on index i are taken to be modulo
6). The quantity
(3.49)
indicates, when D i = 1 that such a collision will take place. Therefore,
(3.50)
is the number of particles left in direction c i due to a two-particle collision
along this direction.
Now, when n i = 0, a new particle can appear in direction c i , as the
result of a collision between n i+l and n i+4 or a collision between n i- l and
ni+2 · It is convenient to introduce a random Boolean variable q (r , t) which
decides whether the particles are deflected to the right (q = 1 ) or to the
left (q = 0) when a two-body collision takes place. Therefore, the number
of particle created in direction c i is
qD i- 1 + ( 1 - q)Di+ l (3.5 1 )
Particles can also b e deflected into (or removed from) direction c i because
of a three-body collision. The quantity which expresses the occurrence of
a three-body collision with particles n i , ni+2 and n i+4 is
(3.52)
As before, the result of a three-body collision is to modify the number of
particles in direction c i as
(3.53)
80 3 Statistical mechanics of lattice gas

Fig. 3.4. Development of a sound wave in a FHP gas, due to particle overcon­
centration in the middle of the system.

Thus, according to our collision rules, the microdynamics of the FHP

model reads
n i(r + 2£\, t + r)ni(r, t)
=

-Di + qD i- 1 + ( 1 - q)Di+ l
- Ti + Ti+3 (3.54)
where the right-hand side is computed at position r and time t. These
equations are easy to code in a computer and yield a fast and exact
implementation of the model. As an example, figure 3.4 illustrates a
sound wave in the FHP gas at rest. Although the initial concentration is
analogous to the situation of figure 2. 1 3, we observe here a much more
isotropic behavior.

3.2.3 From microdynamics to macrodynamics

The macroscopic variables. The physical quantities of interest are not
so much the Boolean variables n i but macroscopic quantities or average
values, such as, for instance, the average density of particles and the
average velocity field at each point of the system. These quantities are
defined from the ensemble average Ni(r, t) (ni(r, t)) of the microscopic
=

occupation variables. Ni(r, t) is also the probability of having a particle

entering site r, at time t, with velocity v i (2/r)£\.
=

Following the usual definition of statistical mechanics, the local density

of particles is the sum of the average number of particles traveling along
each direction C i
6
p(r, t) =
L Ni (r, t) (3.55)
i= l
 3.2 The FHP model 81

Similarly, the particle current, which is the density p times the velocity
field u, is expressed as
6
p ( r , t ) u ( r , t) = 'L viNi ( r , t)
(3. 56)
i=l
Another quantity which plays an important role in the following derivation
is the momentum tensor II, which is defined as
6
IIap = L ViaVif3 Ni ( r , t) (3.57)
i=l
where the greek indices rx and f3 label the two spatial components of
the vectors. The quantity II represents the flux of the a-component of
momentum transported along the /3-axis. This term will contain the
pressure contribution and the effects of viscosity.

The multiscale and Chapman-Enskog expansion. It is important to show

that our discrete world is, at some appropriate scale of observation,
governed by admissible equations : the physical conservation laws and the
symmetry of the space are to be present and the discreteness of the lattice
should not show up. The connection between the microscopic Boolean
dynamics and the macroscopic, continuous world has to be established in
order to assess the validity of the model.
The starting point to obtain the macroscopic behavior of the FHP
automaton is to derive an equation for the Ni s. Averaging the micrody­
namics 3.54 yields
(3.58)
where
1 1
( Qi ) = - ( Ti ) + ( Ti+ 3 ) - (D i ) + 2 ( Di- 1 ) + 2 (D i+ l ) (3.59)
The factor 1 /2 comes from the fact that (q) = 1/2 because, as shown in
figure 3. 1, both collisions are chosen with equal probability.
Equation 3.58 is still discrete in space and time. The Ni s vary between 0
and 1 and, at a scale L > > A, T > > r, one can expect that they are smooth
functions of the space and time coordinates. Therefore, equation 3.58 can
be Taylor expanded up to second order to give
'(;2 _.12
r8 r Ni + ..1(1:\ V )Ni + 2 a ; Ni + l (i:\ · V ) 2 Ni + ..1r(£\ V)a t Ni = ( Qi ) (3.60)
· ·

The macroscopic limit of the FHP dynamics will require the solution of
this equation. However, under the present form, there is little hope of
solving it. Several approximations will be needed.
82 3 Statistical mechanics of lattice gas

The first one involves the Chapman-Enskog expansion presented in

section 3. 1 .3. We shall write
(3.6 1 )
where e i s a small parameter. A s i n section 3 . 1 .3, there will b e here again
some freedom in the choice of this expansion. In order to determine the
Nfl uniquely and consistently, it is necessary to impose the condition that
the macroscopic quantities p and pu are entirely given by the zero order
of the expansion
6 6
p= :L NfO) pu = :L vjNf0l (3.62)
i= l i= l

and, therefore
6 6
l:: Nfl = 0 � ....VI.
� N(t)
i - - 0' for t � 1 (3.63)
i=l i=l
These �relations will appear as a natural choice when we come to the
solution of equation 3.60.
In the case of FHP microdynamics, as opposed to the random walk,
we do not know how A and r are related in the continuous limit. On
physical grounds, we expect to observe phenomena on several time scales.
For instance, the ratio A/r is the velocity of the particles and should not
vanish in the continuous limit. In addition, dissipative phenomena (such
as viscosity) will be present too, and with a coefficient of the order A2 jr.
The difficulty is to capture, in equation 3.60, terms of the same magni­
tude. We could formally introduce, as in section 3 . 1 .3, quantities A 1 , r1 and
r z such that A = d t and r c r1 + e2rz. It is, however, more appropriate
=

to consider two macroscopic time scales T1 and Tz satisfying

'
- = O(e) (3.64)
Tt
and one macroscopic length scale
A
= O(e) (3.65)
Lt
We then introduce two time variables t1 and t2 such that
t1 tz
t=-+- ( 3.66)
e e2
As a consequence, the time derivative Ot becomes.
Ot = eOt1 + e2 0t2 ( 3.67)
 3.2 The FHP model 83

Similarly, a macroscopic space variable rt is introduced. It is related to r

through r = rt /c, and the space derivative reads

(3.68)

where lhx is defined as the derivative with respect to the a-component of

r1.
Except for the right-hand side, (ni), which is not yet expressed in terms
of the Ni,we are now in a position to identify the various orders in c
appearing in the left-hand side of equation 3.60.

The balance equation. We are interested in finding the equation of motion

of the particle density p and the particle current pu. Without solving
equation 3.60 and without developing the term (Qi), the general structure
of these equations can be derived.
Indeed, ni,
which is called the collision term, has some generic properties
6
I: vini = o (3.69)
i= l
The first equation holds because clearly

(3.70)
i=6 l i=6 l 6
L Di - L
i= l Di-1 L
i= l Di+l
= (3.7 1 )
i= l
The second equation in 3.69 is true because, due to their definition
(3.72)
and
(3.73)
Thus, after summation, all the terms cancel. Equation 3.69 is not true by
accident. It follows from the local conservation of particle numbers and
momentum during a collision.
Of course, equation 3.69 is also satisfied for the average (ni ).
Thus,
when equation 3.60 is summed over i, or multiplied by ci and then summed,
the right-hand side vanishes. This property provides the general form of
the equation of motion for p and pu.
Using equations 3.61, 3.67 and 3.68, we obtain for the order c of
84 3 Statistical mechanics of lattice gas

equation 3.60
6

L ['t" Ot1N}0) + AC;p O!pN}0>] = 0 (3.74)

i= l

and
6
L ['t" Ot1Cjer: N}0) + ACjer: C;p Otp N} )]
0
= 0 (3.75)
i= l

where we have adopted the convention that repeated Greek indices imply
a summation over the space coordinates

C;p O !p = C;x Olx O

+ C jy !y (3.7 6 )
We can perform the summation over i, using equ ations 3.62. We obtain

0 (€) : Ot1 P + div1 pii = 0 (3.77)

and

0(€) : (3.78 )
where n�� i s the zero-order Chapman-Enskog approximation o f the mo­
mentum tensor defined in 3.57. The reader will recognize in equation 3.77
the usual conti nuity e quation, reflecting particle conservation. Equa­
tion 3. 7 8 expresses momentum conservation and corresponds, as we shall
see, to the Euler equation of hydrodynamics, in which dissipative effects
are absent.
The same calculation can be repeated for the order 0 ( €'2). Remembering
relations 3.63, we find

(3.79)
and
1 1: 1: (o)
a t2 PUer: + a 1 p ller:( p) + l :'1 :'1 s<o) :'1 :'1
2
O t 1 PUa + l u 1 p u 1 y er:p y + 't"u t1 u tp ller: p
_
- 0 (3.80)
where S is a third-order tensor defined as
6
Sa.py = L V er: Vip V; N
j y i (3.8 1 )
i= l

These last two e qu ations can b e simplified using relations 3.77 and 3.78 .
Let u s first consider the case o f equation 3.79. One has

(3.82)

(3.83)
 3.2 · The FHP model 85

Thus, equation 3.79 reduces t o

(3.84)
which means that at the time scale T2, the density variations are negligible. ·

Similarly, since

(3.85)

equation 3.80 becomes

[
Ot2PUrx + 01 p llrx(1)p + 2r Ut1 llrxp
(O) ( �
(O)
+ U ly Srxpy )
� ] = 0 ( 3.86)

This last equation contains the dissipative contributions to the Euler

equation 3.78. The first contribution is ll�� which is the dissipative part

of the momentum tensor. The second part, namely (

� o11 n�W + O!ys!j1)
comes from the second-order terms of the Taylor expansion of the discrete
Boltzmann equation. These terms account for the discreteness of the lattice
and have no counterpart in standard hydrodynamics. As we shall see,
they lead to the so-called lattice viscosity.
The order e and e2 can be grouped together to give the general equa­
tions governing our system. Summing equations 3.86 and 3.78 with the
appropriate power of e as factor gives

�
OiPUrx + p llrxp +
O [ � ( EOt, n�W + a y s!jy � )l = 0 (3.87)

where we have used Ot = e011 + e20t2 and a�a = eO l rx· Similarly, equa­
tions 3.77 and 3.84 yield

o , p + divpu = o (3.88)
which is the standard continuity equation. Equation 3.87 corresponds to
the Navier-Stokes equation. With the present form, it is not very useful
because the tensors n and S are not given in terms of the quantities
p and u. To go further, we will have to solve equation 3.60 and find
expressions for N}0l and N}ll as functions of p and u. However, for the
time being, it is important to notice that the derivation of the continuity
equation 3.88 and the Navier-Stokes equation 3.87 are based on very
general considerations, namely mass and momentum conservation which
ensure that I: Q i = I: i\Qi = 0. The specific collision rules of the FHP
model do not affect the structure of these balance equations. However,
the details of the collision rule will play a role for the explicit expression
of ll and S.
86 3 Statistical mechanics of lattice gas

The Boltzmann equation. In order to solve equation 3.60 we first need to

express the right-hand side (Q i ) in a suitable form. Unfortunately, there
is no exact way to relate (Q i ) with the Ni s because, the collison term is
nonlinear and, in general, the average of a product is not equal to the
product of the average. However, if the quantities n i can be considered as
independent random variables, factorization is possible.
It is certainly an approximation to assume that (n i nj) = (n i) (nj) = Ni Nj
when i f. j because there are correlations among the particles (note that
when i = j, one has nf = n i , due to the Boolean nature of n i and
(nf) = (ni) ). These correlations may play a crucial role in the dynamics, as
we shall see in chapter 5, for a simple reaction-diffusion system. However,
in many situations, it turns out that, due to the numerous collisions the
particles experience, some form of molecular chaos takes place in the
system and it can be assumed that different occupation numbers n i and nj
are statistically independent. This is the so-called Boltzmann hypothesis
leading to a closed equation for the (n i )s.
It is quite difficult to prove that the Boltzmann hypothesis holds or
not in a given system. It is, however, a technical step which is necessary
to perform in order to progress with our derivation of the macroscopic
behavior of the lattice gas automaton. In section 5.8, we shall present
another method which does not require factorization. Unfortunately, this
method is limited to rather simple cellular automata dynamics, due to
technical difficulties.
The above discussion points out an important aspect of cellular au­
tomata modeling : a numerical simulation of, say, the FHP model does
not imply any simplification such as the Boltzmann hypothesis. The dy­
namics contains all degrees of freedom of the system and the computer
provides an exact N-body simulation (although the interactions are much
simpler than in real systems). Therefore, spontaneous fluctuations and
many-body correlation naturally build up as the dynamics goes on. Cel­
lular automata thus offer a very interesting way to study correlations
in nonequilibrium systems and to take into account the effect of the
fluctuations. We shall come back to these question in chapter 5.
For the moment, let us assume that the Boltzmann hypothesis is true
for the FHP model. Therefore,
(3.89)
and equation 3.60 reads
"[2 A_2
n
Ta t Ni + A-(ci · V ) Ni + 2 a; Ni + 2 (ci · V)2 Ni + A-T(ci · v)a t Ni = i ( N ) (3.90)
Using the explicit expression for Q i in terms of Ti and D i (see equations 3.49
 3.2 The FHP model 87

and 3.52), we obtain

Qi(N) =
- NiNi+2 Ni+4(1 - Ni+1 )(1 - Ni+3)(1 - Ni+5)
+Ni+1Ni+3Ni+ s (1 - Ni)(1 - Ni+2 )(1 - Ni+4)
-Ni Ni+3(1 - Ni+ l )(1 - Ni+2 )(1 - Ni+4)(1 - Ni+5)
1
+ 2 Ni+1Ni+4(1 - Ni)(1 - Ni+2 )(1 - Ni+3)(1 - Ni+ 5 )
1
+ 2 Ni+2 Ni+ s (1 - Ni)(1 - Ni+ l )(1 - Ni+3)(1 - Ni+4) (3.9 1 )
Equation 3.90 i s called the Boltzmann equation.

(a
The Chapman-Enskog expansion revisited. Equation 3.90 can be solved
using the Chapman-Enskog method. Using equations 3.61, 3.67 and 3.68
for the expressions of Nj, O t and I o rrx ) in terms of e, one obtains an
equation for each order O(et ). The right-hand side of equation 3.91 can

n n ( e t ( an��-(o) )
be expanded as

i(N) =
i(N 0) ) + Nj l ) + O(e2 ) (3.92)
1= 1 1
The left-hand side of 3.9 1 is of order e 1 , at least, whereas the right-hand
side has a contribution of order 0 ( e0). The two lowest-order equations
are simply (see equations 3.74 and 3.75).
O ( e0 ) : Qi(N(O) ) 0 =
(3.93)

( ) .
and

::) N �O) + Ulrx VIrx N �O)

Utl 1
::3 •
1
= !��
8Qi(N (O) ) � l )
· N1 ( 3 94)
• j= l oN1
0
The first equation determines the N� ) s. Once they are known, they can
be substituted into the second equation in order to obtain a solution
for the N? ) . Unfortunately, this procedure is not as simple as it first
looks, because the matrix (o.J).joN) (whose elements are 8Qi/8Nj) is not
invertible. Indeed, due to the conservation laws, one has

(3.95)

and similarly

( 3.96 )
88 3 Statistical mechanics of lattice gas

which shows that the columns of the matrix (an;aN)

are linear combi­
nations of each other and that the determinant is zero. The above two
equations can also be written as

( ;� ) T Eo = ( ;� ) T E 1 = ( ;� ) T E2 = 0 (3.97)

where the superscript T indicates a matrix transpose. The quantities

Eo, E1 and E2
are called the collisional invariants and are vectors of IR6
defined as
Eo = ( 1 , 1, 1, 1 , 1, 1 )
E1 (c u , c2 1 , C31 , c41 , c s1 , c6 1 )
(3.98)
where CiiX
denotes the spatial component rx of lattice direction vector Ci.
The reason the Ek
are called collisional invariants is because they describe
the conserved quantities of the dynamics, namely
N · Eo = p vN · E1 vN · E2 = pu2
= pu1 (3.99)
where · denotes the scalar product in IR6 and v = J.. f 'r:.

atl N[0) a l iXViiXN[01

In order for equation 3.94 to have a solution, it is necessary that
+ be in the image space of (an; aN).
It is well known
from linear algebra that the image of a matrix is orthogonal (in the sense

[ r
of the scalar product) to the kernel of its transpose

Im (;�) � Ker
( ;� r (3. 100)

Therefore, the solubility condition of equation 3.94 requires that + at1 N[0)
a l iXViiXN[0) Eo, E1 E2 .
N} l )
be orthogonal to and But that is precisely what
equations 3.77 and 3.78 express. Therefore, a solution is guaranteed
to exists provided that at1 N}0)
is suitably given in terms of the balance
equation. As we shall see, the appropriate choice will be to write
a N}-0) atJP + -a�-
atl Ni(0) -_ -� N[0) atJPUIX
up upu(X (3. 101)

and express at1 P and at1pu(X as given by equations 3.77 and 3.78.
Finally, note that when a solution to equation 3.94 exists, it is not
unique. As we said in section 3.2.3 we also impose the condition that
6
L iN ) 0, for t ;:::: 1
i=l i=l V V =
(3. 102)

This amounts to asking that the solution N ( l ) is also orthogonal to the

collisional invariants and belongs to Im(an;aN).
 3.2 The FHP model 89

The next subsections will be devoted to the explicit solution of equa­

tions 3.93 and 3.94.
The local equilibrium solution. The solutions N�O) which make the collision
term Q vanish are known as the local equilibrium solutions. Physically,
they correspond to a situation where the rate of each type of collision
r
equilibrates. Since the collision time is much smaller than the observation
times T1 and T2 , it is reasonable to expect, to a first approximation, that
an equilibrium is reached locally.
However, to find the N�0) s such that Qi (N(0l ) 0, one needs an extra
=

assumption. The quantity Qi is the sum of several contributions, as shown

by equation 3.59. The quantity Ti (N) gives the probability of a three­
body collision with one particle traveling along direction c i . In a local
equilibrium situation, the probability of the other three-body collision
Ti+ 3 has no reason to be different from Ti . Similarly, the probabilities
of all two-body collisions should be identical. Therefore, we can cancel
Qi (N(0l ) as follows
Ti(N(O) ) Ti+3(N(0) )
= (3. 103)

D i( N(O) ) D i+ l(N (O) ) D i l( N(0) )

= = - (3.104)
Since these relations should be satisfied for all i, equation 3 . 1 04 is equiv­
alent to simply D ( N(0l ) D i+ l(N(0l ). The first condition (Ti Ti+3)
i = =

reads
Ni(O) (O)
Ni+2 Ni+(O)4 _ Ni+(O)l Ni+(O)3 (O)
Ni+S
-
( 1 - N}0 l ) ( 1 - N�2) ( 1 - N�4 ) ( 1 - N�1 ) ( 1 - N�3 ) ( 1 - N�5 )
(3. 105)
We take the logarithm of this equation and introduce the notation
Mi
=
N �O)
1
(3. 106)
( 1 - N}0) )
so that our condition becomes
log(Mi ) + log(Mi+2 ) + log(Mi+4) - log(Mi+ l) - log(Mi+ 3) - log(Mi+S ) 0 =

(3. 107)
Note that this relation is the same for all i and we shall write it as
log M1 - log M2 + log M3 - log M4 + log Ms - log M6 0 (3. 108) =

The second set of conditions (equation 3 . 1 04) reads

log Mi + log Mi+3 log Mi+ l + log Mi+4
= (3.109)
which gives two independent relations
log M1 - log M2 + log M4 - log Ms 0 = (3. 1 10)
90 3 Statistical mechanics of lattice gas

for i = 1 and
log M2 - log M3 + log Ms - log M6 = 0 (3.111)
for i = 2. The other values of i yield relations that are a linear combination
of 3.110 and 3.111. Instead of equations 3.110 and 3.111, it is convenient
to replace them by their sum and difference
log M1 - log M3 + log M4 - log M6 = 0 (3.112)
and
log M1 - 2 log M2 + log M3 + log M4 - 2 log Ms + log M6 = 0 (3.113)
In order to combine equations 3.108, 3.112 and 3.113, it is convenient to
use the collisional invariants defined in equation 3.98
Eo (1, 1, 1, 1, 1, 1)
E1 (e ll , c2 1 , c31, C41, cs 1 , c6 1 )
E2 (3.114)
as well as the quantities
E3 (1, - 1, 1, - 1, 1, -1)
E4 (1, 0, - 1, 1, 0, - 1)
Es (1, -2, 1, 1, -2, 1) (3.115)
where Cirx denotes the spatial component of lattice direction6 vector Ci.
a
The six vectors Eo , ... , E5 form an orthogonal basis of IR because the
cis have the following properties
(3.116)
and
6
L Cirx Cif3 = 3 brxp (3.117)
i= l
The first two relations are obvious from figure 3.3. The last one can be
demonstrated as follows. First, we remark that it is equivalent to the
relation
(3.118)
where 1 is the 2 2 unity matrix and cici is the matrix whose components
x
are cirx Cif3 · Then, it is easy to check that if one multiplies the matrix I: i cic i
by any of the ck, one obtains 3ck because Ck · ck+ l = 1/2, ck · ck +2 - 1/2 =
and Ck + Ck +2 + Ck +4 = 0. With the vectors En, the local equilibrium
conditions 3.108, 3.112 and 3.113 can be written as
log M · E3 = 0 log M · E4 = 0 log M · E5 = 0 (3.119)
 3.2 The FHP model 91
where log M is a notation for (log M1 , log M2 , log M3 , log M4 , log M5 , log M6)
and · also represents the scalar product in IR6 . Since the Ens form a basis
of the space, log M is any linear combination of the collisional invariants
Eo, E 1 and E2
(3.120)
or, by using the definition of Eo , E 1 , E2 and the notation b = (b 1 , b2 )
log Mi a + b · ci = (3.121)
We can obtain the solution for N(O) using 3.106 and 3.121
N(O) = 1 (3.122)
1 + exp( -a - b · ci )
�

This expression has the form of a Fermi-Dirac distribution. This is a

consequence of the exclusion principle we have imposed in the cellular
automaton rule (no more than one particle per site and direction).
The Euler equation. The quantities a and b in 3.122 are functions of
the density p and the velocity field u and are determined according to
equations 3.62. In order to carry out this calculation, we consider a Taylor
expansion of N}0l , up to second order in the velocity field u.
Although straightforward, this calculation is rather tedious. We write
N�o) as
l

=
1=
Ni(0) f(Xi ) - 1 + exp(-X (3.123)
i)
where the six-dimensional vector X (X1 , X2 , ... , X6) is defined as
=

=
X a(p, u)Eo + b 1 (p, u )E1 + b2 (p, u )E2 (3.124)
The expansion of N}0l around u 0 yields =

N}0l = f(Xi( u = O)) + urx -a 1

f(Xi( u = O)) + -2 u upa
a
� f(Xi( u = O))
a Urx a Urx Up
(3.125)
We have
(3.126)
and
(3.127)
When the fluid has no macroscopic motion, that is when u 0, all the =
Nj0l should be identical, for symmetry reasons. Therefore, Xi( u 0) does =
92 3 Statistical mechanics of lattice gas

not depend on i and we write

Xi(p, u = 0) = x(p) (3.128)
Thus, equation 3.125 becomes
Ni(o) -_ f(x) +ucx ( - �
2
[
axi ) f, (x) + -1 ucxup a 2xi f, (x) + axi axi f (x)
a Ucx u=O a Ucx a Up -
a Ucx -
aUp
,

u=O
]
(3.129)
The unknowns in this relation are determined according to 3.99, which
states that
p = L Eoi · N�0) pu1 = v L E l i · N�o) pu2 v L E2i · N�0) (3.130)
=

First we observe that

L. EOi axi
aucx
aa
= 6 aucx
I

because
Eo · Eo = 6 E1 · E1 = E2 · E2 = 3 (3.132)
and Ek · E1 = 0 if k =/= l. Similarly,
L Eoi axi axi aa aa + 3 ab 1 ab 1 + 3 ab ab2
6 aucx 2
(3.133)
i
=
aucx aup aup aucx aup aucx aup

(3.134)
and
(3.135)
because
{
L E01· EOl· Ekl· -- L. Ekl· - 06 ifotherwise
k=O (3.136)
.
I I

if k, l =I= 0 (3.137)
and
if k, l, m =/= 0 (3.138)
These results follow from the geometric properties of the lattice directions
ci and, in particular, from 3.117.
 3.2 The FHP model 93
The condition p = Eo · N(0) now reads
p = 6f(x ) + 6urxarJ1 (x) + 3urxup arxpf'(x)
+ 21 urx up [6arx ap + 3 blrxblp + 3 b2rxb2p ] f" ( x ) (3.139)

( )
where we have introduced the notation
arx = (� ) a rxp iPa (3.140)
O Urx u=O =
O Urx O Up u=O
and similarly for b1 and b2 . Since 3.139 must be satisfied for all values of
p and u, one concludes that
6f(x) = p arx = 0 (3.141)
and
(3.142)
Now, since £ 1 · N(O) = p(udv), we also have
1
U l = 3urxblrxf1 (x ) + UrxUp [blrxfJ f1 ( x ) + (3arxblp + 3 blrx ap)f11 (x)] (3.143)
P -; 2
Consequently,
3buf' ( x ) f!_v=
b1 2 = 0 (3.144)
For symmetry reasons, the final condition E2 N(O) = p(u2 jv) gives
3 b22/1 ( x) = f!_v b2 1 = 0 b2rxp = 0
·

(3.145)
Finally, equation 3.1 42 reduces to
auf'(x ) + � brd"(x) = 0 a22f'( x ) + � b�2j" (x ) = 0 (3.146)

=
and the only non-zero quantities of our expansion are
f(x ) 6p bu b22 = 3vfp' (x)
= (3.147)
and
p 2 f" ( x )
au = a22 = - 18v (3.148)
2 (f' ( x )) 3
We can now substitute these results into the expression 3.125 for N}0)
p + p Ci . u - p2 u2 f"(x )
N'�O) 6 3v 36 v 2 (f1 (x)) 2
-+ -+
=
2p f"(x) up
+ 18 (jl (x)) 2 cirxCitJ -;Urx --;; (3.149)
94 3 Statistical mechanics of lattice gas

Since f is a Fermi-Dirac distribution, one also has

!' = / (1 - f) !" = / (1 - !)(1 - 2!) (3.150)
and, with f(x) = (p/6),
p2 f"(x) - 2p(3 - p) (3.151)
18 (f' (x))2 3(6 - p)
We now have the final expression for the local equilibrium solution of our
problem
p Cj U - -1 pG(p) 2u2 + pG(p) Cia Cip -Ua -
Ni(0) = -p6 + - Up (3.152)
3v 2 v v v
_. _.

with the function G defined as

G(p) = �3 (6(3 -- p)p) (3.153)
The Euler equation is the first approximation describing the fluid mo­
tion. It neglects corrections to local equilibrium, in particular those which
produce viscosity. The Euler equation is obtained from equation 3.78 with
the expression 3.57 for the momentum tensor. The calculation of II�� con­
sists of multiplying equation 3.152 by VjaVip and summing over i. We recall
that L i Via = L i ViaVipViy = 0, because Ci = - ci+3 and L i ViaVip = 3v2 b ap ·
We also need to evaluate fourth-order quantities in the V j. An important
relation is that
L CiaCip Ciy C i{) = 43 ( bapbyo + b aybp0 + b aobpy) (3.154)
6

i=l
The above relation is satisfied for any particular orientation of the six
vectors Cj. The fourth-order tensor E?=l Cia Cip Ciy Cio is isotropic and this
is the reason why a hexagonal lattice is considered for a hydrodynamical
model, rather than a square lattice. Isotropic fourth-order tensors made
of the lattice directions play a crucial role in the lattice gas models
of hydrodynamics because of the convective nonlinear term u · Vu in
the Navier-Stokes equation. The isotropy requirement constraints the
possible lattices which can be used. This is a problem in three dimensions
where such an isotropic lattice does not exist. One has to consider a
four-dimensional lattice and then project in three-dimensional space. As
a result, one obtains two different speeds of particles and some directions
with two particles per links.
The result of equation 3.154 can be obtained by general symmetry
consideration [33] but a direct derivation is also possible. For this purpose,
we write
Ck = (ck l , Ck2 ) = (cos(k(J + cf>), sin(kO + cf>)) (3.155)
 3.2 The FHP model 95

where
e 2n6 and ljJ [0, 2n]
= E (3. 1 56)
The quantities to be computed in relation 3. 1 54 are actually
L Ck io L Ck2• L c� l c�2 , L Ck l Ck2 , and L Ckl Ck2
k k k k k
(3.1 57)
Using �k=A rk (r n - 1 )/(r - 1) for f. 1 and exp(iO) cos e + i sin e,
= r =

one has
6 5
L (ckl + ick2 )4 L exp(4ik8) exp(4il/J)
k=l k=O
exp (4z.A.'+' ) exp(24i8) -1 =0
exp(4i8) - 1 (3. 1 58)

6 5

L ( Ckl + ick2 )\ckl - ick2 ) L exp(2ik8) exp(2il/J)

k=l k=O
exp(2 z. ljJ ) exp(12i8) -1
_
exp(2iO) 1 = 0 (3.1 59)

and
6 5
L ( Ckl + iq2 ) 2 (ck1 - ick2 ) 2 L 1 6
= = (3. 1 60)
k=l k=O
On the other hand, a directed calculation of the left-hand side of
relations 3. 1 5 8-3. 1 60 yields
6 6

L (ckl + iq2 ) 4 = L [(ck l + Ck2 - 6c� 1 c�2 ) + 2i(ck l Ck2 - Ck1Ck2 )] (3. 1 6 1 )
k=l k=l
6 6

L(ckl + iq2 ) 3 (ck1 - iq2 ) L [(ck l - Ck2 ) + 2i(ck 1 Ck2 + Ck1Ck2 )] (3. 1 62)
=

k=l k=l
and
6 6

L (ckl + iq2 ) 2 (ck1 - ick2 )2 L (ckl + Ck2 + 2c� l c�2 )

= (3. 1 63)
k=l k=l
Comparing equations 3 . 1 6 1-3. 1 63 with 3. 1 58-3 . 1 60 yields
L: c� 1 c�2 � L: ck l L ck2 � L ck l Ck2 L CklCk2 0
= = = = =

k k k k k
(3. 1 64)
96 3 Statistical mechanics of lattice gas

which, in a more compact form, is equivalent to relation 3 . 1 54, namely

6
L C irxCifJ C iy C i(j 4 ( <5rxp <5y[J + <5rxy <5p [! + <5rxb<5py)
3
i=l
= (3.1 65)

We now return to the calculation of fl (0l . Using equation 3. 1 54 and

equation 3. 1 52, we have
nrx(ofJ)

(3. 1 66 )
where
3
g ( p ) = -G(p p-3
2
) = -
p-6 (3. 1 67)

ci,
- )u2) <5rxp
As we can see, this result is independent of the lattice directions because
of the isotropy of the tensor 3 . 1 54.
(
The quantity � p � g (p is called the pressure term and pg (p )urx up
the convective part of the momentum tensor. Our microscopic dynamics

(v2
p 2 2 (p)u
p -
p g 2)
gives an explicit expression for the pressure

= (3. 1 68)

The term �p corresponds to a perfect gas contribution, at fixed tempera­

ture. In the FHP model, the temperature is not defined. In addition, the
balance equation for the kinetic energy is identical to the mass conserva­
tion equation, since all particles have the same velocities. Temperature has
been introduced in multispeed lattice gas models, through the equipartition
theorem [37,78,79] .
We are now in a position to write down the Euler approximation (i.e.
at scale T1 and Ll ) describing the behavior of our discrete fluid.
It is a standard approximation of hydrodynamics [80] that, at low Mach
number (i.e for u much smaller than the speed of sound), the density p
can be considered as a constant p = p o , except in the pressure term.
Within this limit and using the expression for fl (0l , equation 3.78 be­
comes
(3. 1 69)
 3.2 The FHP model 97

Compared with the standard Euler equation, the above relation differs
because of the g(p) factor in front of the convective term ( u · V) u. The
fact that g(p) 1- 1 is of course a failure of the model to properly describe
· a fluid flow. This factor comes from the non-Galilean invariance of the
cellular automata dynamics and implies that a t urx + g(p )( u · V) u is no longer
the total derivative ( du/ dt ) .
However, at low Mach, g(p) can be assumed constant and, by renor­
malizing the time t properly, one can transform a t , into g(p )a; and get rid
of the g(p) factor in the left-hand side of equation 3. 1 69.

The speed of sound. Within the Euler approximation (that is neglecting

the damping due to viscous effects), we can compute the speed of sound
of the FHP model. For this purpose, we consider a fluid at rest (u = 0)
with a given density P eq · Small fluctuations fo:.p and fo:.u will obey a wave
equation with a propagation speed which will be the speed of sound. We
have, according to 3.77 and 3.78
(3. 1 70)
and
(3. 1 7 1 )

where, t o first order i n u, IT�� = (v 2 /2)p� rxf3 (see equation 3 . 1 66). Since Peq
is constant, we have, to first order in fo:.p and fo:.u
(3. 172)
and
v2
Peqa t, ua + 2 a la !!P = 0 (3. 173)

Combining these two equation yields

v2
a t,2 !!P - 2 v2l !!P = o (3. 174)

This relation has the form of a wave equation with a speed

v2
cs2 =
2 (3. 175)

Thus, we can write

(O) - Cs2 p u rxp
rrrx/3_ x
(3. 176)
98 3 Statistical mechanics of lattice gas

The lattice viscosity. We shall now proceed to the explicit calculation of

[ ( )] -
the terms involved in the Navier-Stokes equation derived in 3.87, namely

atPUrx + ara IIrx(OJfJ + eiirx( lfJJ + 2r ea t1 IIrx(OJfJ + ara Srx(OfJJy -0 (3. 1 77)
p y
The new terms with respect to the Euler approximation of the pre­
vious section are of two kinds : those involving the zeroth order of the
Chapman-Enskog expansion and II �� which implies computing N? J from
equation 3.94. For symmetry reasons, these new contributions are expected
to be odd functions of the velocity flow u. Therefore, we consider only the
lowest-order terms in the expansion of N}0l and II �� .
We first investigate the two contributions
r a
2 arp
ea t! II rx(OJfJ and (3.1 78)

From equation 3 . 1 66 we approximate

(3. 1 79)

Since at1P = -divtpil (by equation 3.77), one has ea t1 P = -divpu and
r a (O J _ rv2 a
2 arp e a tj II rx fJ - 4 arp e a tjP() rxfJ

(3.1 80)

To compute the term involving

(3. 1 8 1 )

we first notice that the only contribution to N}0l given by equation 3 . 1 52

will be
(3.1 82)

because the other terms contain an odd number of v;. Thus, we have,
using 3. 1 54,
v2p 3
3 4 (6rxp6y o + 6rxy 6p0 + 6rx06py )u0
v2
4 p( 6rxpUy + 6rxy UfJ + 6py Urx)
 3.2 The FHP model 99

Therefore
T u� 2
� - Srxf3y
TV 2 TV 2 u� .
-
2- = - V2 P Urx + - -;- dtvpu
(0) �

(3. 1 83)
urp ry 8 4 urrx
Substituting the results 3. 1 80 and 3.183 into the Navier-Stokes equa­
tion 3. 1 77 yields

(3. 1 84)

We observe that the contribution of the two terms we just computed has
the form of a ViSCOUS effect Vtattice V 2 pu, where
2
Vtattice = - TV (3. 1 85)
S
where Vtattice is a negative viscosity. As mentioned previously, the origin
of this contribution is due to the discreteness of the lattice (S��Y and
O t 1 IT�� comes from the Taylor expansion). For this reason, this term is
referred to as a lattice contribution to the viscosity. The fact that it is
negative is of no consequence because, the contribution - 8�13 EIT�d which
we still have to calculate will be positive and larger than the present
one.

The collisional viscosity. The usual contribution to viscosity is due to the

collision between the fluid particles. This contribution is captured by the
term 8�13 eiT�d in equation 3 . 1 84. In order to compute it, we first have to
solve equation 3.94 for N} l ) . To lowest order in the velocity flow u, we
have

oNi(0) ' � (0)

uN
- -�-
�
i � f3 IT (O) + � N (0)
dlV tP U - -�--
up upUrx U l rxf3 UtrxVirx i
(3. 1 86)
where we have expressed the time derivative of N}0l in terms of the
derivatives with respect to p and pu rx , as mentioned in equation 3.101
u� Ni(O) u� Ni(O)
O t 1 Ni(0) -
-- - O t 1 P + -upU
�up �-rx O t 1 P Urx (3. 1 87)

and used equations 3.77 and 3.78 to express O t1 P and O t 1 P Urx . These
substitutions will ensure that the right-hand side of equation 3. 1 86 will be
in the image of (of!/ oN ).
100 3 Statistical mechanics of lattice gas
As we did for the lattice viscosity, we shall consider only the first order
in the velocity flow u. The omitted terms are expected to be of the order
O(u 3 ). From the expressions 3.152
and 3.166,
we have for the lowest order
in u

and (3.188)
Thus
oN �O) 1 (3.189)
_I_

ap 6 and

and we can rewrite 3.186 as

(3.190)
From this result, it is now clear that equation 3.186
will have a solution :
the six-dimensional vectors Qa.p of components Qa.p i given by (c ia.C ifJ - �c5a.p)
are orthogonal to the collisional invariant Eo, E1 and E2 defined in 3.98
( this is in particular due to the fact that L i C ia.C ifJ = 3c5a.p) . Since Eo, E1
and E2 are in the kernel of ( oO.joNf, then Qa.p is in the image space of
( oO.joN) ( see equation 3.100).
We can now consider the left-hand side of equation 3.186.
We compute
0
it for u = and introduce the notation

(3.191)
3.91 we obtain
From the explicit expression o f n i given b y equation
(anl ) -- -i (1 - s)3 - s3 (1 - s)2 - s (1 - s)4
oN1 u=O
_ !2 s2 (1 - s) 3 - !2 i (1 - s) 3
- -s (1 - s) 2 [s (1 - s) + s2 + (1 - sf + s (1 - s))
-s (1 - s)2
= (3.192)
 3.2 The FHP model 101
Similarly, we compute the other terms and we obtain
( an )
-
1
=
- s (1 - s)
2

aN u=O 2
-2 s + 1 -3s + 1 4s - 2 -3s + 1 s+1
s+1 -2 s + 1 -3s + 1 4s - 2 -3s + 1
X
-3s + 1 s + 1 -2 s + 1 -3s + 1 4s - 2
4s - 2 -3s + 1 s+1 -2 s + 1 -3s + 1
-3s + 1 4s - 2 -3s + 1 s+ 1 -2 s+1
s + 1 -3s + 1 4s - 2 -3s + 1 s+1 -2
(3.193)
We have already mentioned that the vectors Q rxp are in Im ( We g�).
observe here that they also are eigenvectors. To see that, let us consider
the product of the first row of (an; aN) with Q rxfJ· Using

Q rx{Ji = ( CirxCifJ - 2.1 brxp) (3.194)

one has

+s[l:)c irxCip ) - 4c 3rxC3fJ + 2c4rxC4fJ - 4c srxCsp]

i
= 3 brxp - 6c trxClfJ
+s[3brxp - 4(c lrxClfJ + C3rxC3fJ + c srx c sp) + 6c trxClp]
1
= -6(1 - s)(c trxClp - 2. brxp ) (3.195)
where we have again used 3. 1 17. Since the rows of (anjaN) are circular
permutations of each other, the same calculation can be repeated and we
find that
( an ) Qrxp = -3s( 1 - s) 3 Q rxp (3. 196)
aN
This yields immediately the solution for N( l ) as a multiple of Q rxfJ· Since
Q rxp is orthogonal to the collisional invariants, N( l ) will clearly satisfy the
extra conditions 3.63. Thus we have
_ - 2r_ 3 (cirxCip - 1 brxp)a l p P Urx
Ni( 1 ) - (3.1 97)
3p(1 s) 2.
We may now compute the correction EfJ( l ) to the momentum tensor. Since
102 3 Statistical mechanics of lattice gas

E81 {3 8p , We Obtain
=

EIIrxf3( l )
r: [ ]
3p(� : s) 3 � divpu<5rxp - i ( <5rxp <5ys + <5rxy <5ps + <5rxs<5py )8y pus
v 2 divpu<5rxp - ( 8rx pup + 8p purx )]
2p( ; _ s) 3 [ (3.198)
We can now rewrite explicitly, to first order in E and second order in
the velocity flow u , the Navier-Stokes equation obtained in 3.184. Using
expression 3.166 for II�t ' we have

[
8tPUrx + 8p (pg (p)urx up) - Vp - 8p 2p (�v� s) 3 (<5rxpdivpu - ( 8rxUf3 + 8p urx ))
= ]
(3. 1 99)
where the pressure p is given by relation 3.168.
Again, in the limit of low Mach number, the density can be assumed
to be a constant, except in the pressure term [80]. From the continuity
equation 3.88 , we then get divpu = 0 and JI( l ) contribution reduces to
1 TV2 arx8p pup + p 813u
2 rx]
-P 2p (1 - s) 3 [
TV2 V2 urx
2p(1 - s) 3
-vcon V 2 urx ( 3.200)
Finally, within this approximation equation 3. 1 99 can be cast into

8t u + g (p )(u · V )u _ !p Vp + vV2 u
= (3.201)
The quantity v is the kinematic viscosity of our discrete fluid, whose
expression is composed of the lattice and collisional viscosities

V =
(
TV 2 2p (1 � � ) 3 - �) ( 3.202 )
We observe that the viscosity of the FHP model depends strongly on the
density and may become arbitrarily large for the limiting values p = 0
and p = 6. Its minimal value is obtained for p = 3/2
Up to the factor g ( p ) = (3 - p ) / (6 - p ), equation 3.201 is the standard
Navier-Stokes equation. This term g ( p) f- 1 is an indication of the non­
Galilean invariance of the model. However, as we explained previously,
 3.2 The FHP model 103

since we assumed that p � const, the factor g(p) can be absorbed in a

renormalization of the time and our lattice dynamics is described by the
usual hydrodynamic equation.
Several models have been proposed to restore the Galilean invari­
ance [42,44, 8 1 ] , by adding more particles on the lattice with different
velocities. We shall return to this problem in the framework of the lat­
tice Boltzmann models which are a generalization of the fully discrete
dynamics of the FHP model.
In this section we have derived the macroscopic behavior of the FHP
model using the Chapman-Enskog and a multiscale expansion. Whereas
the form of equation 3.201 depends little on the type of collision the
particles experience, the expression 3.202 for the viscosity is very sensitive
to the collision processes. In a lattice gas dynamics, the viscosity is intrinsic
to the model and is not an adjustable parameter. The collision rules can
be modified in order to change the viscosity. Other techniques than the
full calculation we have presented here can be used to obtain the viscosity.
M. Henon [82,83] has developed a method in which a perturbed shear
flow can be considered to compute the viscosity.

Mean free path. It is interesting to note that there is a proportionality

relation between the collisional viscosity (see 3.202)

Vcoll =
( �)3 )
'W 2 2p(1 �
and the mean free path.
The mean free path is the average distance a particle travels between
two collisions. If we denote by the probability of having a collision
Pk
after exactly k steps, the mean free path is defined as
00

Amfp A k=2.::1 kpk

=

-
The probability of having a collision after k steps is the probability of
having no collision during the first k 1 iterations and then a collision.

- Pcon)k-1Pcoll
Thus, we write
Pk = (1
where is the probability for a particle to undergo a collision.
Pcoll
Amfp APcoll k=�1 k( - Peon)k- 1 PcolA l
Thus, the mean free path becomes

= � 1 = -
104 3 Statistical mechanics of lattice gas
This quantity can be computed from the microdynamics
'r,
ni (t + r + Ai:\) ni(t, r) i( )
= +Q n

A particle traveling with velocity

if Q = -r, + A!:\) =
i - 1 ni(t + r 0).
(i.e.
v i ni ( ,'f
(i.e.
t ) =1
From equation
undergoes a collision
) 3. 9 1,
we obtain
Qi(ni 1 )
= = -(1 -
niH )( 1 - ni1 +sHni+)]2ni+4( 1 -ni+3)
+ni+3( 1 -ni+2)( -niH
The probability Peon of having a collision is then the probability that
takes the value -1. Qi (ni 1 )
Since = can be either or 0 -1, Qi
Peon = Prob(Qi =
- 1 ) -(- 1 ·
= -1) + 0
Prob(Qi = 0)
· Prob(Q i = = - ( Qi )
Consequently,

Amfp =
1
In the Boltzmann approximation and at equilibrium (u 0, Ni s
= = =
(p/6)), we obtain

-(Qi) (1 - s)2 [s2( 1 - s) + s( 1 - s)2] � ( 1 - � r

= =

Therefore,

Amfp - e ( 1 A- e ) 3
6 6

v Amf
and
Veon =
12 p
This result is important because it shows that in order to have a small
viscosity, one has to decrease the mean free path. This is achieved by
increasing the collision probability. In conclusion, the more collisions we
have, the smaller the viscosity is.

Numerical viscometer. The viscosity can also be measured directly on

a simulation of the lattice gas automaton. The method developed by
Kadanoff et al. [ 8 4] is to build a numerical viscometer by considering two
Poiseuille flows [80],
opposite to each other. The system is periodic in both
x- and y-directions. Opposite uniform forces are applied to the upper and
lower half of the system by injecting momentum with a given probability
at each lattice site. Practically, this is achieved by reorganizing the local
directions of motion of the particle so as to modify the momentum in the
x-direction. Momentum is forced in the positive direction in the upper
part of the system and forced in the negative direction in the lower part.
 3.2 The FHP model 105

From the parabolic velocity profile (Poiseuille flow), one can deduce the
viscosity.
It turns out [84] that, as the system size gets larger, this viscosity
· increases logarithmically. Such a logarithmic divergence is expected for a
two-dimensional fluid, from general considerations of statistical mechanics.
This observation seems in contradiction with the result 3.202. The point
is that the transport coefficients, such as the viscosity given by 3.202,
are renormalized by hydrodynamic contributions (when going beyond the
Boltzmann approximation) whose effects are divergent with the system
size in two dimensions.
This problem is related to the existence of long time tails in the velocity
autocorrelation function in a fluid. It was first observed numerically

t-d/2 ,
by Alder and Wrainwright in 1968 that the velocity autocorrelation in
hydrodynamics decays as where d
is the dimensionality of the
t
system and the time. This long time tail is also well observed in lattice
gas models [85,86] .
The connection with the divergence of the viscosity is the following. It
is well known from statistical mechanics [87] that transport coefficients
(such as the viscosity) can be computed as an integral over time of
suitable autocorrelation functions. This is the Green-Kubo formalism. In
two dimensions, the velocity autocorrelation goes for long times as t- 1
and its integral diverges, as an indication that some pathology is present
in a two-dimensional fluid.

3.2.4 The collision matrix and semi-detailed balance

In section 3.2.2, we described the collision between FHP particles in terms
of two- and three-body collisions (see equation 3.54).
A more general formulation is often considered, in which no explicit
collision is given and which is generic of any lattice gas model. The idea
of this formulation is to introduce a collision matrix A(n, n')
which gives
the probab ility n
that an input state is transformed into an output state
n' n'
in a collision process (note that here designate the state just before
the propagation step takes place). For instance, in the plain FHP model,
figure 3 . 1 shows that the probability of the transition
( n1, nz , n3 , n4 , ns , n6 ) = (0, 1, 0, 0, 1, 0) � ( 1 , 0, 0, 1, 0, 0) = (n� , n; , n� , n� , n� , n� )
is
A (n, n ) = 21
I

Clearly the whole collision process can be defined by giving the the full
collision matrix A.
In the example of the FHP model, there are 64 input
states and 64 possible output states and thus, is a 64 x 64 matrix. A
106 3 Statistical mechanics of lattice gas
In general, A is a sparse matrix : A(n, n') is zero for most pairs (n, n').
This is the case in particular when the states n and n' do not have the
same momtentum or do not contain the same number of particles. Since
A(n, n') is a probability one has obviously
L A(n, n') 1 = (3.203)
n'
for any n.
In some models, the collision matrix is symmetrix : = A(n, n') A(n', n)
and one says that detailed balance
holds. This is the case with the HPP
and plain FHP models. However, in general, this is not true. We may for
instance modify the two-body collision rule in the FHP model so that,
with probability 1, one has
n (0, 1, 0, 0, 1, 0) � ( 1 , 0, 0, 1, 0, 0) n'
= =

n' ( 1 , 0, 0, 1, 0, 0) � (0, 0, 1, 0, 0, 1 ) n"

= =

n" (0, 0, 1, 0, 0, 1 ) � (0, 1, 0, 0, 1, 0) n

= =

Then 1 A(n, n') =I= A(n', n) 0 and detailed balance is violated.

= =

A weaker property is the so-called semi-detailed b alance which only

requires the symmetric of 3.203, namely that
L A(n, m) = 1
n

for any m. Obviously, detailed balance implies semi-detailed balance.

Semi-detailed balance is obeyed by most lattice gas models. In particu­
lar, by the above modified two-body FHP collision rule. However, if the
interaction was such that, for instance,
n' n
� n" � n n�n
then semi-detailed balance would not be satisfied. Note that this situation
also implies that time reversal is not obeyed.
When all input states appear with equal probability, semi-detailed bal­
ance guarantees that all output states are also equally likely. This property
is important because it is necessary to show the existence of an H-theorem
in lattice gases and prove rigorously [88] that the dynamics of a system
at equilibrium relaxes to the completely factorized, Fermi-Dirac universal
distribution 3 . 1 22.

The lattice Boltzman equation. Using the collision matrix A, the asso­
ciated lattice Boltzmann equation is easily written. In the Boltzmann
approximation (factorization assumption), if is the probability that
Ni
 FHP model
3.2 The 107

ni 1, the probability P(n) of an input state n can be expressed as

=

P(n) II N? ( 1 - Nj ) ( l j )
= (3.204) -n

j
Hence
L P(n)A(n, n')
n

is the probability of output state n' and the average value N; (n�) is =

N; L: n; P(n)A(n, n')
=

'
n,n

Since Ni can also be written as L niP(n) and L A(n, n')n n

' = 1, we have
N; - Ni L (n; - ni)P(n)A(n, n')
=

'
n,n

Using expression 3.204 for P(n) and remembering that N; Ni(r + A.ci, t + =

r), we finally obtain the following lattice Boltzmann equation

Ni(r+A.ci, t +r)- Ni(r, t) L (n;- ni)A(n, n') II N? ( 1 - Nj ) ( l j ) (3.205)
=
-n

n,n
' j

3.2.5 The FHP-III model

Several important problems in hydrodynamics occur at high Reynolds
),
numbers R = (tujv where t is the characteristic scale of the flow, its u
characteristic velocity and v the kinematic viscosity [80] . The Reynolds
number appears when the Navier-Stokes equation is scaled according to
the characteristic sizes of the physical flow. It indicates that the flows of
different fluids are similar when they have a similar Reynolds number.
This similarity of flows explains why, in principle, lattice gas automata
can be used to simulate any real fluid.
In the case of the FHP model, the factor g(p) has been absorbed in
a rescaling of time, which results in a rescaling of the viscosity v (p) --+
(v (p)jg(p)). The Reynolds number can then be rewritten as
R = tMR. (p) (3.206)
where M = (u/cs) is the the Mach number (cs is the speed of sound
derived in section 3.2.3) and R. (p) is defined as
Cs g(p)
R. (p) = (3.207)
v(p)
and is the reduced Reynolds number which contains the intrinsic proper­
ties of the lattice gas dynamics [88] . The maximum value of R. that can
108 3 Statistical mechanics of lattice gas
before after before after
""' I / ,,J / "--- I --""' ""' I I
� __.,.
/I' / w ""- / I ""- / .t / ""- / .w ""-
"-._ I / I
� __.,. "-.._ � / I
� __.,. � t/
/ t ""- / I ""- / .t , /.""-I / I ""'
""" �
"-._ I / "-._t "--- t --""' I
� __.,.
--""' I ' / .I /""- / t' /
/ � ""- / I ""'
""' I I
� __.,. ""' I / � t I
� __.,.
/ t/ ""- / � ""- / t' / .� / ""- /i""-
""" �
""' I "-._t / "--- I --""' "--- t --""' I
� __.,.
/ .I /, / .� ""- /.""-I / t' /
/.""-� /.I """
""" �
"-._ I / � t "--- t --""' I
� __.,.
_...,. . ,
I /I/ � / I ""- / t' / /.""-�
"-.._ � / "-._t / I
� __.,. "-.._ � / � t/
/ t' /.""- / I ""- _...,. . , /�
I I �

Fig. 3.5. The collision rule of the FHP-111 model. Before and after collision
configurations are shown. When two possible outputs are given, each of them is
selected with equal probability.

be obtained with the FHP model is

R�ax = 0.387 for p = 1.12 (3.208)
Considering a Mach number of M = 0.3, which is a standard choice for
assuming incompressibility, and a typical length scale of t = 1000 lattice
size, one obtains a Reynolds number of about R = 1 30.
A way to reach larger Reynolds numbers without increasing the lattice
size is to have the smallest possible viscosity [88] . For this purpose, it is
necessary to work with another set of collision rules.
The FHP-III model [3] is a variation of the FHP dynamics which is
significantly less viscous. Its particularity resides in the presence of rest
particles (i.e. up to seven particles per lattice site). In addition, the number
of configurations leading to a collision is increased with respect to the
original FHP automaton.
 3.2 The FHP model 109

Figure 3.5 illustrates the collision rules. We observe that while mass and
momentum are exactly conserved in a collision, kinetic energy is not. A
collision between a rest particle and a moving one results in two moving
· particles and there is production of kinetic energy. However, the inverse
collision is expected to occur with the same probability and, therefore,
energy is conserved in a statistical way.
The viscosity of the FHP-III model is found to be 3 []
v � <v2 ( 4p( l _ � ; '(I _ �) - �) (3.209)

which is smaller than the viscosity of the FHP model. Therefore, increasing
the number of collisions results in decreasing viscosity. For this model,
the maximum value of R. [88]
is
R!?ax = 2.22 for p = 2 (3.2 10)
Although the FHP-III model has a better physical behavior, it is much
more complex than the FHP dynamics. From the point of view of
numerical efficiency, it is highly desirable to implement the collision rule
as a lookup table in which all the 2 7 possible collisions are computed in
advance. Otherwise the microdynamics requires an enormous number of
terms. As an example of the complexity of the collision term, we give

population of rest particle

No,
below the algebraic expression for the FHP-III collision rules for the

No(t + r, r ) - No = - No [((N1N2 + N2N1 )N3 + N3N1 N2 )N4 Ns N6

+N1 N2 N3 ((N4Ns + NsN4 )N6 + N6N4 Ns )]
+ [(N1N3N2 N4 + N2N4N1 N3 )Ns N6
+(N3NsN4 N6 + N4N6N3 Ns )N1 N2
+(N1NsN2 N6 + N2N6N1 Ns )N3 N4 ](No -No)
1 - -- --
+ l [N6N2 (N1N4N3 Ns + N3NsN1 N4 )
+N4N6 (N2NsN1 N3 + N1N3N2 Ns )
+N2N4 (N3N6N1 Ns + N1NsN3 N6 )](No -No)
1 - -- --
2
+ [N1Ns (N2N4N3 N6 + N3N6N2 N4 )
+NsN3 (N2N6N1 N4 + N1N4N2 N6 )
+N3N1 (N4N6N2 Ns + N2NsN4 N6 )](No -No)
+ [N4(Ns(N2N6N1 N3 + N1N3N2 N6 )
+N2N3N6N1 Ns ) + N1(N2(N3NsN4 N6
+N4N6N3 Ns ) + N3NsN6N2 N4 )](No -No)
1 10 3 Statis tical mechanics of lattice gas
+ [N6(Ns(N4(N3(N2N1 + N1N2 ) + N1N2N3 )
+N1N2N3N4 ) + N1N2N3N4N5 )
(3.211)
where Ni, i = 1, ... , 6 label the moving particles. The right-hand side of this
equation is computed at time t and position r. The factors 1/2 account
for the two possible outcomes arising for some collisions. Quantity q is a
stochastic Boolean variable at each lattice site. The evolution law of the
moving particles Ni, i =f- 0 is even more complex.
3.2.6 Examp les offluid flows
In this section, we show, as an illustration, two examples of fluid flows
past an obstacle. Figure 3.6 show the streaklines lines resulting from the
presence of a plate in the way of a fluid flowing from right to left. The
flow pattern is produced by letting particles in suspension follow the fluid
streaklines (note that streaklines are different from streamlines in the sense
that the latter are defined as the path suspensions would follow if one
could "freeze" the velocity field at a given time).
The rule we have used to produce the motion of suspensions is the
following : each of the particles cumulates the local velocity flow, averaged
over a 32 x 32 neighborhood of sites and 10 consecutive time steps. When
it has gathered enough momentum along the x - or y-axis, it moves by one
lattice site in the corresponding direction by consuming an amount dx or
dy of its momentum.
Particles in suspension are injected along a vertical line on the right
part of the system, with regular spacing. One observes the formation of
eddies, resulting from a von Karman flow instability. Note that the fluid
particles are not shown in this figure, but only the suspensions.
The cellular automaton fluid we have considered here does not exactly
obey the FHP rule. It is a variant which enhance three- and four­
body collisions : all possible collisions with three or four fluid particles
entering a site are taken into account, as is the case for the FHP-III
model. However, here, there is no rest particle. In order to maximize the
frequency of collisions, the simulation has been performed at an average
density of p = 3 (i.e. three particles per site, on average).
The initial fluid speed is produced (and maintained) in a standard
way, by accelerating the particles in the first quarter of the system. This
is typically obtained by reversing the direction of motion of a particle
traveling rightward with a given probability.
In figure 3. 7 , we show the flow produced past a vertical plate. The
parameters of the simulation are identical to those of figure 3.6, except
that the system size is 5 1 2 x 128 and, consequently, the Reynolds number
 3.2 The FHP model 111

Fig. 3.6. Streaklines shown by particles in suspension in a FHP-like fluid. The

system size is 1024 x 1024 and periodic in the vertical direction.

Fig. 3.7. Streaklines shown by particles in suspension in a FHP-like fluid. The

system size is 5 1 2 x 128 and periodic in the vertical direction.

is smaller. Oscillations are seen along the streaklines, indicating the onset
of an instability.
Cellular automata fluids require quite large systems to exhibit non­
trivial flow patterns. This is due to their relatively high viscosity which
limits the Reynolds number that can be attained. On the other hand,
we can see that the symmetry can be spontaneously broken due to the
intrinsic fluctuations in the fluid automaton.
1 12 3 Statistical mechanics of lattice gas
3.2. 7 Three-dimensional lattice gas models
Hydrodynamics is usually a three-dimensional problem and it is clearly im­
portant to generalize the two-dimensional FHP fluid to three dimensions.
Although there is no difficulty in defining collision rules between particles
moving on a 3-d lattice, there are severe isotropy problems. It turns out
that no regular three-dimensional lattice exists with enough symmetry to
guarantee that fourth-order tensors such as (3. 1 54) are isotropic.
The solution to this problem has been proposed by d'Humieres, Lalle­
mand and Frisch [36,88] . The idea is to consider a four-dimensional
lattice, the so-called the FCHC lattice (face-centered-hyper-cubic) which
is isotropic and then to project the motion to three-dimensional space.
In the 4-d FCHC model, there are 24 directions of motion, correspond­
ing to the lines to connect a site to its nearest neighbors. These links are
of length .ji. The projection in three dimension results in six double links
of length 1 and 12 single links of length .ji. After projection, we obtain
a multispeed model with up to two particles along some directions.
The 24 lattice directions ci are the following [89] . First the double links
are
c1 = ( 1, 0, 0) c2 = (- 1, 0, 0) c 3 = (- 1, 0, 0)
c4 = ( 1, 0, 0) cs = ( 0, 1, 0) c6 = ( 0, - 1 , 0)
c 7 = ( 0, - 1, 0) cs = ( 0, 1, 0) c9 = ( 0, 0, 1 )
c1 0 = ( 0, 0, - 1 ) cu = ( 0, 0, - 1 ) c1 2 = ( 0, 0, 1 )
and, second, the single links
cu = ( 1, 1, 0) c1 4 = (- 1 , - 1 , 0) c 1 s = ( 1 , - 1 , 0)
c16 = (- 1 , 1, 0) c1 7 = ( 1, 0, 1 ) c l s = (- 1 , 0, - 1 )
c1 9 = ( 1, 0, - 1 ) c2o = (- 1 , 0 , 1 ) c 2 1 = ( 0, 1 , 1 )
c22 = ( 0, -1, -1) c 23 = ( 0 , 1 , - 1 ) c24 = ( 0, - 1 , 1 )
Figure 3 . 8 illustrates this projection in three-dimensional space. We refer
the reader to [88] for a detailed discussion of the collision rules.

3.3 Thermal lattice gas automata

3.3. 1 Multispeed models

In the HPP, FHP and related models discussed in the previous sections
we did not pay much attention to energy conservation during the collision
process. The reason is that these models are devised to describe athermal
systems, without a notion of temperature.
Here we would like to present some LGA models which may account
for thermal properties in a fluid. The key ingredient is to build a model
having a velocity distibution instead of one single speed. By velocity
 3.3 Thermal lattice gas automata 1 13

Fig. 3.8. The possible particle velocities of the FCHC lattice gas model of
a fluid. Bold links represent the direction in which two particles may travel
simultaneously.

distibution we mean here that the modulus of the particle speed may vary
in addition to its direction.
With several velocity populations, kinetic energy can be defined in a
non-trivial way and a local temperature associated to the fluid through
the equipartition theorem.
Since we work on a lattice and would like to keep the dynamics to
its maximum simplicity, we will not consider a large range of different
velocities for the particles. Typically we shall have "hot" (or fast) particles
that travel to second nearest neighbor sites in one time step, normal,
speed-one particles and, finally "cold" particles at rest. These models are
usually termed multisp eed LGA.
The main difficulty with a multispeed model is to define appropriate
collision rules. We would like to conserve mass, momentum and enery
independently and we expect that collisions result in a change of the
velocity distribution. The constraint that particles are moving on a lattice
reduces very much the number of possible collisions and the velocity set
should be chosen carefully.
The first multispeed model of this type was proposed by Lallemand
and d'Humieres [37] in 1986. It is defined on a square lattice and com­
prises particles of the same mass m = 1 but speed 0, 1 and J2. The
1 14 3 Statistical mechanics of lattice gas

'

time t time t+l

Fig. 3.9. Mass, momentum and energy conserving collision in the three-speed
square lattice automaton. The reverse collision is also possible.

fast particles move along the diagonals of the lattice while the speed­
one particles are identical to HPP particles. Collisions conserving mass,
momentum and energy can be defined between particles having the same
speed using the regular HPP rule. A non-trivially conserving energy
collision can be defined between particles of different speed. It is
illustrated in figure 3.9 and involves one fast, speed- J2 particle colliding
with one rest particles to produce two speed-one particles. Of course,
the reverse collision, in which two speed-one particles collide at right
angle and give one rest particle and one fast particle, is also taken into
account.
Energy is defined as pure kinetic energy

E = m-2v2
Rest particles have zero kinetic energy, speed-one particles have energy
1 /2 and fast particles have energy 2. With this definition, it is easy to
check that the collision described in figure 3.9 conserves mass, momentum
and energy. It is also interesting to note that this collision rule modifies
the velocity distribution : a pair of particles with speed 0 and J2 are
transformed into two particles of speed one. Thus rest particles couple
the population of speed-one and speed- J2 particles.
Of course, since this model is defined on a squre lattice we expect that
isotropy problems will show up with fourth-order tensors. Thus this model
is not appropriate to describe a fluid flow but, on the other hand, may
well simulate a fluid at rest, submitted to a temperature gradient. An
interesting application is the study of nonequilibrium fluctuations in such
a system [90] .
In section 3.3.3, we shall present other multispeed LGA that are defined
on a hexagonal lattice and can simulate isotropic flows.
 3.3 Thermal lattice gas automata 115

3.3.2 Thermo-hydrodynamical equations

When energy is taken into account, a fluid is decribed by the so-called ther­
. mohydrodynamic equations [87,9 1] . They are the Navier-Stokes equation
complemented with a new equation for the local energy dens ity or, using
thermodynamics transformations, an equation for the local temperature
at each point of the fluid.
In this section we shall briefly sketch how these equations can be
obtained from a multispeed LGA dynamics, without giving a complete,
nor a general derivation.
The microdynamics of a multispeed model can be written using the
usual formalism discussed in this chapter. We denote by n; E {0, 1 } the
presence or absence of a particle traveling with velocity v;
n; (r + TV;, t + r) - n; (r, t) O; ( n (r, t))
=

The right-hand side term 0; is the collision term which now obeys an
extra conservation law for energy

Of course the usual mass and momentum conservation are still required
L m;O; L m;vi ni = = 0
i i
Clearly, in a one-speed model, energy and mass conservation are identical.
Q
The precise expression for depends on the model, the type of collision
and the lattice topology. But the three above properties must be satisfied
in order to describe a thermal cellular automata fluid.
The standard procedure can be repeated to obtained the macrosopic
equations related to the LGA microdynamics : ensemble average, Boltz­
mann factorization and multiscale Chapman-Enskog expansion. The fluid
density p and velocity field u are defined as previously
p u :L m;v;N;
=

i
where N; = ( n i) is the averaged occupation number.
A new physical quantity is introduced : the local energy per particle e

and the corresponding internal energy e can be defined as

=
2
pe :L. m (v· -2 u) N; pe - p -
1
u2 =

l
2
1 16 3 Statis tical mechanics of lattice gas
A temperature T can be introduced in this model, assuming a local
version of the equipartition theorem [87] is valid. Thus we define, for a
two-dimensional system
e = kT
where k plays the role of the Boltzmann constant. Taking

turns out to be a natural choice for k in a lattice gas system.

Another important quantity is the heat current JQ

Q=
J-+ '"""' ( Vj
-+
-+ )2
-u -+ -+

( Vj - U
)
i
N
.
L.....,;
I
2
It is a rather tedious task to obtain the explicit thermohydrodynamic
equations governing p, u, e, starting from a given model and we will not
develop further this derivation. In general, we expect a loss of Galilean
invariance but, in some models [8 1], tricks may be used to remove this
problem.
As an illustration of the properties of a thermal model, it is also
interesting to consider the simpler case of a fluid at rest, in a stationary
state and submitted to a temperature gradient. In this case we typically
expect the following result [78]
JQ = - KVT
where K = K(p, T ) is the thermal conductivity which is built-in in the
collision rule.

3.3.3 Thermal FHP lattice gases

Multispeed models on a hexagaonal lattice which generalize the FHP
dynamics to include thermal properties have also been proposed. However,
due to six-fold symmetry, they are slightly more complicated than the
thermal HPP dynamics.
A three-speed FHP model is illustrated in figure 3.10. Particles may have
speed 0, 1 or 2 , in lattice units. We see that the collision which couples
the three populations involves a speed-two and a rest particle to produce
two speed-one particles (or conversely). But to conserve mass, momentum
and energy the different velocity populations must have different masses.
Rest particles have mass 3/2 and zero kinetic energy, speed-one particles
have mass 1 and energy 1/2 and, finally, speed-two particles have mass
1 /2 and energy 1.
 3.4 The staggered invariants 1 17

Fig. 3 . 1 0. Mass, momentum and energy conserving collision rule in a thermal

FHP model. The reverse collision is also possible. Note that particles with a
different speed have a different mass.

In order to define a multispeed dynamics on a hexagonal lattice with

all equal masses it is necessary to include a fourth velocity population.
The model proposed by Grosfils et al. contains up to 19 particles per site
(one rest particle plus three times six moving particles [79] . In addition to
speed 0, 1, and 2 , a population of speed J3 is included. These particles
travel along the lines bisecting the lattice directions and, in one time step,
they jump from one site to one of the six next-nearest neighbors.
In this model, a particle has kinetic energy 0, 1/2 , 3/2 or 2. The collision
rules are defined by selecting at random the output velocity configuration
among all configurations having the same mass, momentum and energy
as the input configuration. This approach requires one to classify all
situations in a table, according to their particle number, momentum and
energy. Figure 3.1 1 shows some of these classes and some configurations
belonging to it.
This four-speed FHP model has been shown to reproduce correctly
spontaneous fluctuations by observing the behavior of the dynamic struc­
ture factor which is expected to exhibit three peaks for a real fluid [79] .

3.4 The staggered invariants

As we have seen previously, quantities that are conserved by the micrody­

namics play a crucial role in the derivation of the macroscopic equations
governing the cellular automaton.
In the FHP model, the basic ingredients we have put in the dynamics
are the local conservation of mass and momentum, which are prescribed
by the laws of physics.
In principle, it is not desirable to have any other conserved quantities
in the dynamics. It turns out that extra conservation laws are frequent
1 18 3 Statistical mechanics of lattice gas
J..L= 2, 1t=(2,0), e=2 J..L= 3, 1t=(3,0), £=3

I •
� < f-.
J..L= 3, n=(O,-...J 3), e=9/2 J..L= 3, n=(O, ...J 3/2), e=3

�\ l
I L. •

Fig. 3. 1 1 . Example of particle configurations with same mass Jl, momentum n

and energy E in the four-speed FHP model. Particles with speed 0, 1 , .J3 and 2
are represented by a dot, a thin arrow, an open arrow and a large, black arrow,
respectively.

in lattice gas models. We have already observed that the HPP model
described in section 2.2.5 conserved momentum along each lattice line and
each lattice column. We have mentioned that both two- and three-body
collisions are necessary in the FHP rule to avoid spurious conservations.
These kinds of undesirable conservations are related to the choice of
the collision rule and generally show up when performing the Chapman­
Enskog expansion. However, in many cellular automata models, there are
other types of spurious conserved quantities, which are an artifact of the
discrete nature of the underlying lattice and have no physical counterpart.
They are the so-called staggered in v ariants, first discovered by MeN amara
and Zanetti [92]. They have received this name because they usually
depend on the parity of space and time.
For instance, we have already noticed in section 3.1.4 the existence of
a spurious invariant directly related to the checkerboard structure of the
square lattice
Neven(t) - Nodd( t) (-1) tf' [Neven(O ) - Nodd(O)]
= (3.212)
The checkerboard invariant generalizes to almost all lattice gas automata.
In the FHP model, this unphysical conserved quantity is called the stag­
gered momentum.
Consider for instance figure 3.12, and more particularly the horizontal
line of sites labeled + The particles sitting on this line at time t will
" ".

either stay on this line if they travel horizontally or move to one of the
horizontal lines labeled "-", above or below. Now, suppose we compute
 3.4 The staggered invariants 1 19

+ +

+ +

+ +

Fig. 3 . 1 2. Illustration of the staggered momentum in the FHP dynamics : The

vertical component of momentum due to particles sitting on a horizontal line
labeled "+" is transferred, at the next iteration, to the two lines labeled "-" above
and below.

the total momentum along the y-axis (vertical direction) carried by each
particles. The particles moving horizontally have, of course, no vertical
component of momentum and do not contribute. Therefore, all the v ertical
momentum of the line will be distributed, at time t 'L, on the two
"+" +
horizontal "-" lines.
If we continue to label the horizontal lines by alternating and "-"
we obtain similarly that all the vertical momentum of a "-" line at t is
"+"
transferred to the surrounding "+"
lines at the next step. We can build a

2:: 2:: 2::

globally conserved quantity by summing over all or "-" lines "+"
Py(t + 'L, r) = L Py(t, r ) Py(t + 'L, r) = Py(t, r ) (3.2 1 3)

where L+ and L_ denote the two sublattices obtained with our line
labeling and Py represent the y-component of momentum. If we add the
two equations, we obtain no more than the expected physical conservation
of momentum. But, by subtracting them, we obtain the typical form of a
staggered invariant, namely

L Py(t + 'L , r ) - L Py(t + 'L, r) =

-
(L Py(t, r ) - L Py(t, r ) )
(2:: )
rEL+ rEL- rEL+ rEL-

= (- l ) ( t/'r) Py(O, r ) - L Py(O, r)

rEL+ rEL_

Here, we see that the staggered momentum results from the detailed way
in which the momentum is conserved on the lattice.
Of course, the same construction we consider for the vertical component
of the momentum can be repeated for the lattice lines orientated 60 and
1 20 degrees with respect to the horizontal direction. As a result, we obtain
1 20 3 Statistical mechanics of lattice gas
a total of three staggered invariants, which are usually expressed as
G0 = L g0 ( t, r ) = L( -1) (t/Tl+ B· r L lJ · vi i r, t)
n( (3.214)
rEL rEL
where L is the set of lattice points. The quantity Li lJ vi i r t)
· n( , is the
momentum along lJ , which is defined as
lJ = ( 1 /Jc)(- sin(n( n /3)), cos(n( n /3))) n = 0, 1, 2 (3.2 1 5 )
For each value o f n, lJ i s orthogonal t o one o f the three families of
line orientations. The scalar product lJ · r specifies one sublattice division
because it take the same integer value (even or odd) on a line perpendicular
to lJ .
The local staggered invariants, such as g 0 often show a diffusive behavior
but sometimes may exhibit propagating modes, as shown by Ernst and
Brito.
Staggered invariants may play an important role in the physical behav­
ior of a lattice gas model because they couple to the nonlinear terms in
the Navier-Stokes equations [84,92-94] . For instance the u · Vu term will
be accompanied by an analogous term containing the average staggered
momentum density g 0 . However, no direct evidence of pernicious effects
have been detected in fluid flow simulations, at least qualitatively.
The influence of a staggered invariant on the physical behavior of a
cellular automata fluid can be minimized by properly choosing the initial
condition of the fluid particles [84] . It is indeed always possible to prepare
an initial state in which the total staggered momentum G0 vanishes exactly.
In addition, it is possible to chose the initial positions of the particles so
that the staggered momentum coincides with the regular momentum. For
this purpose, they are only placed on the lattice of spacing 2Jc shown in
figure 3 . 1 3. The other sites are left empty. These positions correspond to
choosing the intersection of the L + lines for the three orientations. Along
the bold lines of figure 3 . 1 3, one has
(3.2 1 6)
and, at time t = 0, G0 reads
G0 = lJ i lJ · Ptot
· LL
rEL i v irn ( , 0) = (3.21 7)

where Ptot denotes the total momentum of the system. Since both G0
and i\ot
are invariants, the fact that they are equal at the initial time
guarantees that they are always equal.
Usually, there are many staggered invariants in a lattice gas automa­
ton [95] and there is apparently no general way to find all of them. One
 3.4 The s taggered invariants 121

Fig. 3 . 1 3 . The bold lines represents the sublattice for which the total staggered
momentum reduces to the usual momentum.

often distinguishes between dynamic and geometric staggered invariants.

The former are staggered in both space and time (like the staggered mo­
mentum or the checkerboard invariant), whereas the latter are staggered
only in space.
As an example of a geometrical staggered invariant, consider the case
of a two-speed HPP model [78] . This model consists of two types of
interacting HPP particles : the slow particles obey the same dynamics as
discussed in section 2.2.5 ; the fast particles are also subject to the usual
HPP collisions but travel along the diagonals of the square lattice (which
forms a square lattice of spacing .jiA.). The slow and fast particles have
a different kinetic energy. Interactions between the two species which
conserves mass, momentum and energy can be devised. This system may
represent either a binary mixture or a gas with a temperature. In the latter
case, temperature is defined through the equipartition theorem. Figure 3.14
shows the result of a typical collision between a fast and a slow particle.
A simple geometric staggered invariant exists for the fast particles. Since
they move along the diagonals of the square lattice, there are two families
of fast particles : those moving on the "white" sites of the checkerboard
and those moving on the "black" sites. Therefore, their number (N����
and N��m is conserved independently of each other. The sum + N�'dJ N����
expresses the expected mass conservation of the fast particles and the
difference
Nofastdd _ Nfast
even
= ""' - r·B
� ( 1)
""""'
�
i=l,4
nfast (r t)
l ' (3.218)
rEL

is a geometric invariant, where e =

(1/ A.)(1, 1) characterizes the checker­
board division of the lattice.
1 22 3 Statistical mechanics of lattice gas

'

time t time t+l

Fig. 3. 1 4. Collision between a fast (long arrow) and a slow (short arrow) particle
in the two-speed HPP model. Mass, momentum and energy are conserved during
this interaction process.

Of course, the fast particles also yield the usual checkerboard dynamical
invariant,
) _ 1 ) t/1:H · il nfast t
L L (r, ) (3.219)
rEL i=l,4

which is now characterized by lJ = (1/ A.)(1, 0) or lJ = (1/ A.)(O, 1 ) and means

that a fast particle entering a vertical (respectively, horizontal) line at time
t will clearly enter, at the next step, one of the adjacent lines, at left or
right (up and down, respectively).

3.5 Lattice Boltzmann models

3.5.1 Introduction
Cellular automata fluids, such as those discussed so far in this chapter,
represent idealized N-body systems. Their time evolution can be performed
exactly on a computer, without many of the approximations usually done
when computing numerically the motion of a fluid. In particular, there is
no need, in a lattice gas simulation to assume some factorization of the
many-body correlation functions into a product of one-particle density
functions.
Of course, the cellular automata model may be inadequate to represent
a real situation but it includes naturally the intrinsic fluctuations present
in any system composed of many particles. This features is out of reach
of most tractable numerical techniques. In many physical situations,
spontaneous fluctuations and many-particle correlations can be safely
ignored. This is, however, not the case if one is interested in long time
tails present in the velocity autocorrelation (see section 3.2.3). We will
also consider in the next chapter other systems for which correlations or
 3.5 Lattice Boltzmann models 123
local fluctuations cannot be neglected and where the N-body nature of
the cellular automata approach is necessary.
On the other hand, a cellular automata simulation is very noisy (because
it deals with Boolean quantities). In order to obtain the macroscopic
behavior of system (like streaklines in the flow past an obstacle), one
has to average the state of each cell over a rather large patch of cells
around it (for instance a 32 x 32 square) and over several consecutive time
steps. This slows down very much the simulation speed and requires large
systems. Therefore, the benefits of the cellular automata approach over
more traditional numerical techniques becomes blurred [3 8 ] .
In addition, due to its Boolean nature, cellular automata models offer
little flexibility to finely adjust external parameters. Tuning is done through
probabilities, which is not always the most efficient way.
When correlations can be neglected and the Boltzmann molecular chaos
hypothesis is valid, it may be much more effective to directly . simulate on
the computer the lattice Boltzmann equation
(3.220)
with Qi given, for instance, by (3.91) or (3.211). It is more advantageous
to average the microdynamics before simulating it rather than after. The
quantities of interestNi are no longer Boolean variables but probabilities
of presence which are continuous variables ranging in the interval [0, 1].
This approach reflects the fact that, at a macroscopic level of observation,
the discrete nature of the fluid particles vanishes.
A direct simulation of the lattice Boltzmann dynamics was first con­
sidered by McNamara and Zanetti [51]. It considerably decreases the
statistical noise that plagues cellular automata models and considerably
reduces the computational requirements. As mentioned previously, the
main drawback is that this approach neglects fluctuations and many-body
correlation functions.
Strictly speaking, one can argue that a dynamics based on a set of
continuous variables is no longer a CA model because a real number
requires an infinite amount of information to be stored in a computer.
However, in practice, real numbers are approximated using a finite number
of bits (typically 32 or 64 in modern computers) and, therefore, the lattice
Boltzmann approach is also a cellular automata rule such that the state
of each cell is a large (but finite) number of bits. Of course, the use of a
lookup table is not possible and a direct calculation of the rule should be
repeated for each site and each iteration.
The lattice Boltzmann method or lattice Boltzmann equation (often
referred to as LBM and LBE) have been widely used for simulating various
fluid flows [96] and is believed to be a very serious candidate to overcome
traditional numerical techniques of computational fluid dynamics. Their
1 24 3 Statistical mechanics of lattice gas

microscopic levels of description provide a natural interpretation of the

numerical scheme and permit intuitive generalizations to complex flow
problems (two-phase flow [97,45,46] , magnetohydrodynamics [98] , flow in
porous media [40,41] or thermohydrodynamics [99] ).
In the following chapters, we shall consider some applications of the
lattice Boltzmann method to fluid problems (snow transport and depo­
sition by wind) and also to other domains such as pattern formation in
reaction-diffusion systems and wave propagation in heterogeneous media.
In lattice Boltzmann fluids, the most natural way to define the collision
term Qi is to average the microdymanics of a given underlying cellular
automata fluid and factorize it into a product of average, as we did
previously to get the Boltzmann approximation. We then obtain an
expression like those derived in 3.91 or 3.2 1 1 . It turns out that the
collision term given by 3.2 1 1 (and its counterpart for moving particles)
requires more than one thousand floating point operations at each lattice
site and time step. Even on a massively parallel computer, in which every
cell is computed simultaneously, this is usually not acceptable.
The first solution to this problem is to consider the same approximation
as we used with the Chapman-Enskog expansion when deriving the
macroscopic behavior of the FHP fluid. The idea is to linearize the
collision term around its local equilibrium solution. This approach has
been proposed by Higuera and coworkers [52] and considerably reduces
the complexity of the operations involved.
Following the same idea, a further simplification can be considered [100] :
the collision term need not be related to an existing cellular automata mi­
crodynamics, as long as particle and momentum are conserved. In its
simplest form, the lattice Boltzmann dynamics can be written as a relax­
ation equation [101]

flr Jo£\, t
+ + • ) - fi fr , t) = z (!}0l (r, t) - fi fr, t)) (3.22 1 )

or, equivalently as

(3.222)

where the local equilibrium solution f}0l is now a given function of the
density p = 1: fi
and velocity flow p u = 1: f/;i·
The function f}0l can
be chosen so as to produce a given behavior, first to guarantee mass and
momentum conservation but also to yield the appropriate local equilibrium
probability distribution. In particular, the lack of Galilean invariance that
plagues cellular automata fluid can be cured, as well the spurious velocity
contribution appearing in the expression 3 . 1 68 of the pressure term. In a
more general context, f}0l could include other physical features, such as a
 3.5 Lattice Boltzmann models 125
local temperature. Equation 3.221 is often referred to as the Lattice BGK
(or LBGK) method [102] (the abbreviation BGK stands for Bhatnager,
Gross and Krook [103] who first considered a collision term with a single
· relaxation time, in 1954).
Equation 3.221 has been studied by several authors [104], due to its
ability to deal with high Reynolds number flows. However, one difficulty
of this approach is the numerical instabilities which may develop in the
simulation, under some circumstances.
Finally, let us mention, in a similar context, the multiparticle lattice
gas which is an intermediate approach (see section 7.3) between a pure
automata and a real-valued description and allows an arbitrary number
of particles in each direction [105]. This technique is not yet very well
explored and could possibly alleviate some of the instability problems and
restore the many-body correlations.
3.5.2 A simple two-dimensional lattice Boltzmann fluid
In this section, we present a detailed derivation of a lattice Boltzmann
dynamics of a two-dimensional fluid. A natural starting point would
be to consider again the hexagonal lattice used for the FHP model.
However, in most computers (and especially on parallel computers), the
data structure representing a hexagonal lattice is more difficult to work
with than a regular square lattice which is naturally implemented as an
array of data. Although simple coordinate transforms can be devised to
map the hexagonal lattice onto a square lattice (see section 7.2), boundary
conditions should be treated in a separate way.
The choice of a hexagonal lattice stems from isotropic requirements. Ac­
tually, the FCHC model discussed in section 3.2.7 provides an alternative
since it can be projected onto a two-dimensional square lattice.
Figure 3.8 suggests that the resulting lattice is composed of eight ve­
locities i\ with and collinear with the horizontal and vertical
v1v,2V, 3v,4,vVs 6 V7vs
directions, while and correspond to the diagonals of the lattice.
These eight velocities are shown in figure 3.15.
Clearly, one has, for odd i,
(3.223)
r
where, as usual, 2 and are the lattice spacing and the time step, respec­
tively.
Since the square lattice results from the projection of the 24-particle
FCHC model, all velocities do not have the same weight. Actually, four
particles give the same projection along the main lattice directions. This
fact can be taken into account by assigning a mass mi 4 to the horizontal
=
126 3 Statistical mechanics of lattice gas

4
1

Fig. 3 . 1 5. The eight velocities in the lattice Boltzmann model of a two­

dimensional fluid (on the left) and the mass associated to each of these directions
(on the right).

and vertical directions and a mass mi 1 for the diagonal motions, as

=

illustrated in figure 3.15.

The macroscopic quantities, such as the local density p or the velocity
flow u will be defined as
8 8
p = l: mdi p u l: md/; i = (3.224)
i= l i= l
Isotropy. From the derivation we carried out for the FHP model, we
learned that isotropy is satisfied because the right-hand sides of equations
3.1 17 and 3.154 do not depend on a particular choice of the coordinate
system. In our case, since all velocity vectors do not have the same
modulus and all particles the same mass, the isotropy condition will take
the form
8
L mi vir�. VifJ fl(jrt./3
= (3.22 S )
i= l
and
8
L mi v ir�. VifJViy Vio v( <5r�.p <5yo + <5r�.y <5p o + <5r�.o <5 py )
= (3.226)
i= l
where f1 and v can be obtained with the help of some algebra. First,
since and are orthogonal, any vector x can be written as x
Vt vt )vtv3 v3)v3] =

(v -2 ) [(x
· + (x · where v (..1./r) is the modulus of i , for i odd.
= v
Consequently one must have
(3.227)
and similarly for and This is also true for v i and v i+2 when i is even,
vs v7.
except that the coefficient v 2 is replaced by 2v 2 , because of equation 3.223.
 3.5Lattice Boltzmann models 127
Therefore, with mi = 4 and mi = 1 for i odd and even, respectively, we
obtain
8
L miVirxVip = 12v 2 brxp ( 3.228 )
i=l
The fourth-order tensor I: i miVirxVifJViyVi.5 can be computed explicitly by the
method described by equations 3.155 through 3.164 . Since the summation
can be split into the odd and even values of i, we first compute
4
Trxpy .5 = L CjrxCipCiyCi.5 ( 3.229 )
i=l
where, for this calculation, c 1 , c2 , c3 and c4 are orthogonal unit vectors
defined as
ck = (cos(ke + lj>), sin(k8 + 4>)) ( 3.230)
E
with e = n /2 and 4> [0, 2n ]. Thus, we have
4
L (ckl + iq2 )4 = 4 exp(4i l/>) (3.231)
k=l
4
L (ckl + ick2 ) 3 (ckl - iq2 ) = 0 (3.232)
k=l
and
4
L ( Ckl + ick2 ) 2 (ckl - iq2 ) 2 = 4 ( 3.233 )
k=l
Therefore, we have
4
L [(ck l + ci2 - 6c� 1 c�2 )] = 4 cos 44> (3.234)
k=l
4
L [4(ck 1 Ck2 - ckl ck2 )] = 4 sin 44> ( 3.235 )
k=l
4 4
L [(ck l - Ck2 )] L [(ck 1 Ck2 + ck1 ck2 )] = 0
= (3.236)
k=l k=l
and
4
L [ck l Ck2 + 2c� 1 c�2] = 4
+ (3.237)
k=l
128 3 Statistical mechanics of lattice gas

Solving this system yields

� ck12 ck22 1 (1 - cos 4¢)
� = (3.238)
k=1 2
4 4
L Ck t Ck2 - L ckl ck2 � sin 4¢
= = (3.239)
k=1 k=1
and, finally
4 4
L ct1 L ck2 23 21 cos 4¢
= = + (3.240)
k=1 k=1
This result shows that, as expected, the fourth-order tensor Trxpyb is not
isotropic on a two-dimensional lattice. However, here, we also want to
consider the contribution of the other vectors, v2 , v4 , V6 and v8 , which are
n
rotated by I 4 with respect to the vectors, "h
= n
With ¢' ¢ + I 4 in the previous expression, we obtain cos 4¢' =

- cos 4¢ and sin 4¢' - sin 4¢. Therefore

=

4 L Vft Vf2 L Vft Vf2 +

k odd k even

4v 4 2 (1 - cos 4¢) 4v 4 21 (1 cos 4¢)

1 + +
4v 4 (3.241)
The two contributions come with a factor of 4v 4 because, in the first term
mk 4 and, in the second llvk 1 4 ( .y�'2)4 v4 4v4 . In a similar way, we
= = =

obtain
8 8
2: mk Vk t Vk2 2: mk Vk1 Vk2 0
= = (3.242)
k=1 k=l
and
8 8
L mk vk1 L mk vk2 12v 4
= = (3.243)
k=1 k=1
All these results can be combined to give
8
L miVirx VifJViy ViJ 4v4 ( <5rxpt5yb t5rxyt5pb t5rxbt5py )
= + + (3.244)
i=l
 3.5 Lattice Boltzmann models 129
The local equilibrium. The lattice Boltzmann dynamics is given by equa­
tion 3.221
+ + r) - fi('f, t) = ni = � (!}0) (1\ t) - fi('r, t))
flr TVi, t (3.245)
We will find a solution to this equation using, as before, a Chapman­
Enskog expansion
fi = f}0l + cf} 1l + ...
The main difference is that, here, f}0l is given, or more precisely, chosen,
in an appropriate way.
A natural choice for specifying f}0l as a function of the physical quan­
tities p and pu is to adopt a similar expression to that obtained for the
FHP model, namely equation 3.152. Accordingly, we define
b � ·� u2 h VirxVif3UrxUf3
fi(O) = ap 2 v PVi
+ u p e 2
v P
+ + 4
v (3.246)
where a, b, e and h are coefficients which will now be determined. There
are several constraints to fulfill. First, mass and momentum conservation
1mpose
8 8
L: mi ni = O and L: mi vi ni = O (3.247)
i=l i= l
This implies that
8 8
L: md}0l = p and � � �
6 miVi�i(0) = pu (3.248)
i =l i= l
because p and pu are defined through relation 3.224. This choice is also
compatible with the Chapman-Enskog expansion which requires that the
local equilibrium solution completely defines the physical variables (see
equation 3.62). A second condition we will use to determine a, b, e and
h is the Galilean invariance. This is a freedom which is possible in the
framework of the lattice Boltzmann method.
Using the fact that I: mi = 20, I: vi = 0 and I: miVirxVif3 = 12v 2 brxp, we
obtain
u2
20ap (12h + 20e)p 2
+ v
i= l
12bpurx (3.249)
130 3 Statistical mechanics of lattice gas

Therefore, mass and momentum conservation requires that

a = 1/20, 12h + 20e = 0 and b = 121 (3.250)
We can similarly compute the momentum tensor at zeroth order. We
have, by definition

and, after substitution of J}0) , we find

(3.251)
In order to satisfy Galilean invariance, the coefficient in front of the
quadratic term purx up must be 1. Accordingly, we must have
h=!
8
Since 12h + 20e = 0, we obtain
e = --
3
40
and the expression for II�� reads
II�� = pbrxp + PUrxUp (3.252)

( ) (1 2 )
where the pressure is defined as
h u2 3 u2
p = 12v 2 a + ( e + 4) p = 5 v2 - 3v p (3.253)
v2 2
Using the fact that II�� can be written as (see 3.176)
II�� = c; p brxp + O(u2)
we conclude that
s
12av2 = �v2
c2 =
5
We notice that a contribution proportional to u2 is present in the
expression of p. This is an artifact of the model which has no physical
meaning. This contribution can be removed when rest particles are
introduced in the system [101].
 3.5 Lattice Boltzmann models 131
The Navier-Stokes equation. In section 3.2.3, we obtained the following
result (see equation 3.87)
[ r
OtPUa + 0[3 IIaf3 + 2 (EOt1 ITa(O{3) + Oy Sa(O{J)y -_ 0 )] (3.254)
= !l
where t + � and r r tf takes into account the different time scales
= €
E E
of the problem (see section 3.2.3).
The derivation of 3.254 relies only on the fact that I: mi ni 0 and =

I: mi !J.i vi 0 (in section 3.2.3, mi 1, for all i). Therefore, this equation is
= =

still valid here. As we previously mentioned, the generality of this result is

due to the conservation laws which define the form of the equation. The
precise expression of the collision term, on the other hand, determines the
transport coefficient.
We obtained II (0) in the previous section. We still need to compute
S�� y and II ( . The quantity f ( l l is defined by a similar equation to that
l 1 l
obtained in relation 3.94, namely
! � ( (} Qi(f( O
1 )) ) f � l ) Ut::1 1 fI�O) + Ul::1 rxVIrx. fI�O)
J
=
r�
J=l u::1 1.

)
where O l a denotes the derivative with respect to the oc-component of r 1 .
Since Qi ! (!�0) (r, t) - fi(r, t) , the above equation simply reads
=

- 1 (1)
r � fi Ot Ji(0) + O l aViafi(0)
=

Following the derivation of equation 3.186, we have

_ _!__ l
'[� f ( )1 Ut::1 1 fi(O) + Ul::1 aViafi(O)
::1
� i(o) ::1 Ila(O{J) + Ul::1 aViafi(O) (3.255)
u::l fi(o) dlV" l PU - -of-UlfJ
- --

up ::1
u p Ua
with div 1 L a O l a ·
=

As earlier, we shall compute f? ) to the first order in u. We have

b
fi(0) ap + 2v PViaUa
= + 0 ( u2)
and

Thus,
al� o)
-- = a
op
132 3 Statistical mechanics of lattice gas

and

( 3.256)

( - )
Thus
2
!? ) = -T� � viy Vio a� <5y o a ly PUo ( 3.25 7)
t
Using EO y ay , the order O(E) contribution to ll reads
=

8
Enrxf3( l ) E L md[1 l virx Vif3
i=l
[
T� adivpu L miVirxVif3 - � L miVirxVif3Viy Vio ay p usJ
Tv2 � [(12a - 4b) <5 apdivpu - 4b( ap purx + oa pup )] ( 3.258)
where we have used expressions 3.228 and 3.244. We then obtain
ap Ell�� -Tv 2 � [(8b - 12a) Oa divpu + 4b a�pua ]
= ( 3.259)
From this expression, we obtain two viscosity coefficients (shear and bulk
viscosity), as is usual in compressible fluids.
Our final step is the calculation of the lattice viscosity. The first term in
3.254 giving a contribution to the lattice viscosity is 8p ( u/2) ar1 ll��- With
ll�� = 12av2 p<5 ap + O ( u2 ), we have
TE o) .
2 a t1 n rx( /3 = 6 TV 2 aEa t1 pu rxp - 6 TV 2 au ap dIVp u
s: s:
�
=

where we have used ar1 p + div 1 pu = 0 (see equation 3.77) and the definition
of the length scale Ediv 1 div. =

Therefore, we have
TE . U
(o) = - 6TV 2 aa rx dlVp
2 af3 a t1 n rxf3 (3.260)
�

Similarly, we must compute the contribution due to s!jy in equation 3.254

8
s!iy L miVirxVif3Viy /[0l
i=l
8
=
b
2_PUo L miVirxVif3 Viy ViJ
v i=l
 3.5 Lattice Boltzmann models 133
v2
3 PU15 ( �rxf3 �yi5 � rxy�fJ/5 �rxi5�py )
+ +
v2
3 p( uy�rx{J + Up �rxy + Urx �py ) (3.261)
Consequently, we obtain
2 0p 0y ( srx(O{J)y ) - ( Orx d vp u + 2 v2 PUrx )
r rv2
_
3
.

l
1�

The dissipative lattice contributions to the lattice Boltzmann dynamics is

obtained by combining the two terms we have obtained
� op (eo t1 II�� oy S!iY )
+ = rv2 [� 2 + (� - )
v (p urx ) ]
6a 8rxdivpu (3.262)
Finally, after substituting 3.262, 3.259 and 3.252 into the balance equa­
tion for momentum 3.254, we obtain
OtPUrx op [P �rxfJ p urxup]
+ + [(8 b - 12a)8rx divpu 4bV purx]
rv2 � + 2
-rv2 [� 2 + (� - ) ]
v (p urx ) 6a 8rxdivpu
where p is the scalar pressure given by 3.253. After rearranging the terms,
we have
DtPUrx p up opurx urxdivpu
+ + -OrxP
= + rv2 ( � - � ) V2purx
+rv2 [� ( � - 12a) - ( � - 6a)J 8rxdivpu
(3.263)
In the case of an incompressible fluid (at low Mach number, for instance)
one has divpu 0 and we recover the usual Navier-Stokes equation
=

8 t u (u · V)u !p Vp VJb V2 u
+ =
_
+ (3.264)
where VJb is the kinematic viscosity of our lattice Boltzmann fluid
VJb �-=
T� 2 ( � ) (3.265)
From this result, we observe that the lattice Boltzmann viscosity is in­
dependent of the particle density p. In addition, � is a free parameter
which can be tuned to adjust the viscosity within some range. We can
see that when � is small, relaxation to f (O) is fast and viscosity small.
<
However, � cannot be made arbitrarily small since � 1/2 would imply a
negative viscosity. Practically, more restrictions are expected, because the
dissipation length scale should be much larger than the lattice spacing.
134 3
. . . . . . . . . .•
Statistical mechanics of lattice gas
. .
-·
.
.. ... .. .. . - - · ·
.. . . ...
-
· · · - - · - . . .. . ..

C
- · ·· · ·· · · ·· · · ·· ····
- -
n

Fig. 3 . 1 6. Laminar flow past a disk obtained from the 2-D lattice Boltzmann
model. System size is 5 1 2 x 1 28, � = 10 and u00 = 0. 1 .

·. ·. l·
..-- · .
. .

Fig. 3 . 1 7. Non-stationary flow past a plate obtained from the 2-D lattice
Boltzmann model. System size is 5 1 2 x 1 28, � = 1 . and u00 = 0.025. From lef� to ·

right and top to bottom, the figure shows the different stages of evolution.

A more detailed discussion of the lattice Boltzmann approach to fluid

dynamics is given by Benzi, Succi and Vergassola [96].
3.5.3 Lattice Boltzmann flows
Finally, this section presents the results of simulations with the LBGK
model discussed in the previous section. Laminar flows, such as the one
presented in figure 3.16, are easy to obtain. On the other hand, a Karman
vortex street requires some care to set up the boundary conditions. First,
=
the initial condition fi('f, t 0) should be random, in order to break the
symmetry of the numerical scheme. If not, the pattern shown in figure 3.17
cannot develop (everything being symmetric with respect to horizontal
axis). Second, boundary conditions at the upper and lower limits of the
system are periodic. Other solutions can be considered, though, and a
condition of zero vertical velocity flow gave good oscillation patterns, too.
On the other hand, the flow is more sensitive to the boundary conditions
imposed at the left and right extremities.
 3.6 Problems 135
A simple, yet effective, solution is the following: periodic boundary
conditions are first imposed along the horizontal direction. The fluid
exiting on the right is re-injected on the left. Then, the driven force is
applied on the left to ensure the fluid motion and simulate an unperturbed
u 00
flow at velocity coming from the left. This is obtained by re-distributing
equally all the particles at the system input and systematically removing
some of them from direction c5 to put them in direction c 1 .

3.6 Problems

3. 1 . Cons i d e r an H P P gas at eq u i l i b r i u m with M part i c l es on V l atti ce

s i tes. Ass u m e that a l l poss i b l e conf i g u rations of M part i c l es o n V
s i tes have the same probabi l i ty p( eq u i l i br i u m ) . S how that =
(1/p )
V
(4 !)/(4 V - M) !M ! Ded uce that the ave rage occ u pat i o n n u m be r (ni(r))
r
fo r any s i te and d i rection i i s p/4. S h ow that (ni(r)nj (r ')) is ze ro
fo r an i nfi n i te system , u n l ess i = j and r r'
= . T h i s shows that the
Boltz m a n n facto r i zati o n hypothes is is obeyed , at l east at eq u i l i b ru m .
Com pute (ni(r)nj (r ')) fo r i j a n d =
= r r '.
3.2. Cons i d e r a th ree-speed H P P m od e l at eq u i l i b r i u m o n a two­
d i m e n s i o n a l sq u are l attice. There are th ree pop u l ations of part i c l es :
s peed 0 , speed 1 and speed Ji part i c l es . S u ppose each o f these
popu l ations h as a tota l of Nrest • Nsl ow and Nfast partic l es , respectively.
From the p r i nci p l e of eq u i partition of e n e rgy we m ay ass u m e that each
part i c l e contr i b utes of amou nt (dkB
T ) 2 to the total k i n etic e n e rgy E ,
/
d
where i s the spat i a l d i mens i o n , T
the te m pe ratu re a n d kB
the Boltz­
m a n n constant. S h ow how th is te m pe ratu re scal e depends on Nres t .
kB
Nsl ow and Nfast · What is a natu ra l choice fo r i n terms of the k i n etics
parameter of the model?
3.3. Cons i d e r anothe r te mperatu re sca l e fo r the p revious p rob l e m ,
based o n t h e m i croca n o n i cal defi n it i o n : (kB T )- 1
= W joE)v where
(8 ln
W i s the n u m be r of poss i b l e conf i g u rations with Nres t • Nsl o w a n d Nf ast
part i c l es , E the k i netic energy and V the l attice s i ze wh ich i s kept
constant. Com pare th is te mperatu re scal e with that of the p revi o us
p ro b l e m . W h i c h one seems the most approp r i ate?
3.4. Write down the m i crodynam ics fo r the H P P gas.
3.5. Derive the d ifferential eq uati on gove rn i ng the FHP dynam ics with
some s i m p l ified ass u m ptions. From the l attice Boltz m a n n eq u ation
Ni(r + 2ci, t + c) - Ni(r, t) Qi(N) =
136 3 Statistical mechanics of lattice gas
where n is the F H P col l is i o n te r m , take the Taylor expansion of the
A r.
Ni N{0l + rN{ 1 l +
l eft-hand side o n ly to first order in and Look fo r a sol ution of the
fo rm = r
us i n g the time step as the smal l expa n s i o n
...

paramete r. Fo l low the Chapman-E nskog fo rm a l ism , neg l ecti ng t h e

m u ltisca l e as pect o f the dynam i cs . Show that the conseq ue nce of
d roppi n g h i g he r orders i n andr A, as wel l as the m u ltiscal e part of
the deve l o p m ent, o n ly mod i f i es the viscos i ty by a constant (the l attice
viscos ity) .
3.6. Apply the same approx i m ations as i n the p revious p ro b l e m to the
HPP m icrodyn a m i cs . Show that a n i sotropy s hows u p in two p l aces : i n
front of the n o n l i near ve locity term (uV)u and i n the viscos i ty te r m .
3.7. Write the co l l i s i o n te rm o f the H P P ru l e i n the fo rm o f a x
16 16
co l l i s i o n m atrix. Write the m ost genera l fo rm o f t h i s matrix w h i c h is
com pati b l e with m ass and m o m e ntu m conse rvat i o n . Check whether
sem i-deta i l ed bal ance is obeyed .
3.8. Add a g ravity f i e l d o n a H P P g as and show that the equilibrium
density p(y) of particles at a h e i g ht y fo l l ows the barometric l aw i n the
l i m i t of l ow d e n s i ty. H i nt: g ravity can be added by writi n g the l attice
Boltz m a n n dynam i cs as
Ni (r + Aci, t + r) - Ni (r, t) = (1 - g )Qi(N) + gFi (N)
where n is the H P P co l l i s i o n te rm , g the p robab i l i ty that g ravity ope rat¢s
at s i te r and Fi is d efi ned as

w h i c h means that g ravity acts vertica l l y and m ay transform a partic l e

N
movi n g u p ( 2 ) i nto a parti c l e goi n g down (N4 ).
Ass u m e that = Ni
Ni (Y ) = p(y)/4 and show that t h i s i s a sol ution of the l attice Boltz m a n n
eq u ation p rov i d ed that
p(y) = (1 - g)Y I Jc p(O)
(Note : i n eq u i l i b ri u m the t i m e depende nce can be o m i tted and i n the
l i m i t of l ow density, o n e can neg l ect the p 2 te rms.) How s h o u l d the
A
p robab i l i ty g vary with the l attice spac i n g when tak i n g the conti n uous
A�
l i m it 0?
3.9. P ropose a n LBGK model fo r a o n e-d i mensional flow, with o r with­
out rest part i c l es . Derive the associ ated hyd rodynam ic equation . Ex­
p ress the l ocal eq u i l i b r i u m fu ncti o n d i rectly i n te rms of the density of
particle fi a n d d iscuss the n u m erical sta b i l ity of the eq u at i o n . Can the
pop u l ation densities becom e negative? What happens if the g rad i e nt
are l a rge?
 3.6 Problems 137
3 . 1 0. F o r the p rev i ou s p ro b l e m o f a 1 0 f l ow, c a n y o u choose the
l ocal eq u i l i b ri u m d istr i bution so as to m od e l the B u rger equati o n :
+
O t U UOxU = v o�u ?
3. 1 1 . I n the L B G K framework, d iscuss how a non-Newto n i an f l u i d cou l d
b e mod e l e d .
 4
Diffusion phenomena

4.1 Introduction

Diffusive phenomena play an important role in many areas of physics,

chemistry and biology and still constitute an active field of research.
There are many applications involving diffusion for which a particle based
model, such as a lattice gas dynamics, could provide a useful approach
and efficient numerical simulations.
For instance, processes such as aggregation, formation of a diffusion
front, trapping of particles performing a random walk in some speci�c
region of space [106,107], or the adsorption of diffusing particles on
a substrate are important problems that are difficult to solve with the
standard diffusion equation. A microscopic model, based on a cellular
automata dynamics, is therefore of clear interest.
Diffusion is also a fundamental ingredient in reaction-diffusion pheno­
mena that will be discussed in detail in the next chapter. Reaction processes
such as A + B C, as well as growth mechanisms are naturally imple­
--+

mented in the framework of a point particle description, often with the

help of threshold rules. Consequently, microscopic fluctuations are often
relevant at a macroscopic level of observation because they make symme­
try breaking possible and are responsible for triggering complex patterns.
Cellular automata particles can be equipped with diffusive and reactive
properties, in order to mimic real experiments and model several complex
reaction-diffusion-growth processes in the same spirit as a cellular auto­
mata fluid simulates fluid flow: these systems are expected to retain the
relevant aspects of the microscopic world they are modeling. However,
diffusion is much simpler than the motion of a fluid and this will be
reflected in the cellular automata rule.
At the beginning of chapter 3, we proposed a one-dimensional cellular
automata model of diffusion. Here, we shall generalize it to two- and
138
 4.2 The diffusion model 139
three-dimensional systems and consider the presence of various types of
complex processes. The purpose of this chapter is thus to investigate the
cellular automata approach to modeling diffusion and growth phenomena.
We shall discuss their reliability (the discrete artifacts of the models must
not show up) and their applicability to various problems. We shall first
spend some time on the pure diffusion rule and, in the next chapter, add
reaction mechanisms.

4.2 The diffusion model

Phenomenologically, diffusion follows from the mass conservation law, i.e.

the continuity equation
O t P + divJ = 0 (4.1)
where J , the particle current i s given i n terms o f the particle density p as
J = -Dgradp (4.2)
The quantity D is the diffusion constant. The last relation expresses that
the particle current is proportional to the spatial inhomogeneities of the
density.
When D is independent of p and independent of the spatial coordinates,
we have a linear and homogeneous diffusion. By substitution of ( 4.2) into
(4.1), one obtains the famous Pick's law
O DV2 p
tP = (4.3)
At a microscopic level, on the other hand, a diffusive phenomenon
corresponds to the random walk of many particles. Particle number is the
only quantity conserved by this dynamics.
This random motion is typically due to the properties of the environment
the particles are moving in. When one is not interested in an explicit
description of this environment, it can be considered as a source of
thermal noise and its effective action on the particles can be assumed to
be stochastic.
It is easy to imagine a simple mathematical model of a random walk
on a lattice. The first idea is to assume that each particle will pick at
random one neighbor lattice site and jump to it. This is the approach
considered in many numerical models in which only one particle moves
at a time. However, in the framework of a cellular automata dynamics
this may cause some problems. All the particles move synchronously and
several of them may choose to hop to the same lattice site. For instance,
r
on a two-dimensional square lattice, suppose that site is empty. Then, it
may happen that all neighboring sites (north, west, south and west) are
140 4 Diffusion phenomena
r
occupied by a particle and that all of them select precisely site as their
new destination.
In this example, if one allows four particles to sit at the same lattice
site then, at the next random motion, they will all be able to jump to a
neighboring site. There is some probability that two of them move to the
same neighbor. But this neighbor could very well receive other particles
from elsewhere and, finally, receive more than four particles.
Therefore, simultaneous motion may lead to an arbitrarily large number
of particles in some lattice sites. This is in conflict with the exclusion
principle which requires that the state of each site be coded with a small
and finite number of bits.
There is a simple solution to this problem, as already sketched in
section 3.1. We shall consider the same approach as used in the HPP
or FHP dynamics : at any site, there may be as many particles as there
are lattice links. Each particle has a velocity, defined by its direction of
motion. At each time step, the particles move to a nearest neighbor site,
according to its velocity. By reaching the site, the particles experience a
"collision" during which they change randomly their direction of motion.
In other terms, each particle selects, at random, a new velocity among
the possible values permitted by the lattice. But, as several particles may
enter the same site (up to four, on a two-dimensional square lattice),
the random change of directions should be such that there are never
two or more particles exiting a site with the same travel direction. This
would otherwise violate again the exclusion principle. In order to avoid
such a situation, it is sufficient to shuffle the directions of motion or,
more precisely, to perform a random permutation of the velocity vectors,
independently at each lattice site and each time step.
This evolution rule requires random numbers and corresponds to a
probabilistic cellular automaton.

4.2. 1 Microdynamics of the diffusion process

In this section, we shall write down the microdynamics related to the
diffusion rule we have just presented. Following the same formalism as
developed in chapter 3, we shall derive its macroscopic behavior and show
that Fick's law of diffusion is obtained in the limit of an infinite lattice
and a time step going to zero. Corrections due to the finite size of the
lattice spacing and the time step will be discussed in section 4.3.
Our model consists of particles moving along the main directions of a
hypercubic lattice (a square lattice in two dimensions or a cubic lattice in
three dimensions). As opposed to cellular automata fluids, we do not have
to consider here more complicated lattices. The reason is that diffusion
processes do not require a fourth-order tensor for their description.
 4.2 The diffusion model 141
The random motion is obtained by permuting the direction of the
incoming particles. If d is the space dimension, there are 2d lattice
directions. These 2d directions of motion can be shuffled in 2d ! ways,
which is the number of permutations of 2d objects. However, it is not
necessary to consider all permutations. A subset of them is enough to
produce the desired random motion. For the sake of simplicity, we restrict
ourselves to cyclic permutations. Thus, at each time step, the directions of
the lattice are "rotated" by an angle rxi chosen at random, with probability
Pi , independently for each site of the lattice. With this mechanism, the
direction a particle will take and exit a given site depends on the direction
it had when entering the site. The modification of its velocity determines
its next location on the lattice. The particle behaves as if it has some
inertia, as in a model proposed by G.I. Taylor [108] in 1921.
By labeling the lattice directions with the unit vectors c i we can introduce
the occupation numbers n i ("r) defined as the number of particles entering
the site r, at time t with a velocity pointing in direction c i .
With this notation, the CA rule governing the dynamics of our model
reads
-2d l
+
ni( r ACi , t + •) = t'=O
L f.lt (r, t)ni+t'(r, t) (4.4)
where i is wrapped onto { 1, 2, ... , 2d}. The f.lt E { 0, 1 } are Boolean variables
which select only one of the 2d terms in the right-hand side. Therefore
they must obey the condition
2d- 1
L f1t = 1 (4.5)
t'=O
Practically, this condition can be enforced in a simulation by dividing
the interval [0,1 ] into 2d bins of length Pt , each assigned to one of the
f.lt· Then, at each lattice site and each time step, a real random number
between 0 and 1 is computed (with a random number generator). The bin
it falls in will determine which f.lt is the one that will be non-zero.
In order to be more specific, we shall now consider the two-dimensional
case. This will correspond to many of the situations studied later in this
chapter.
On a two-dimensional square lattice, there are four possible random
rotations, as shown in figure 4.1 and equation 4. 4 reads
ni( r 2 ci , t • ) = f.loni f.l l ni+l f.12 n i+2 f.1 3 n i+3
+ + + + + (4.6 )
where the values of i are wrapped into {1,2,3,4} and the vectors c i are
defined in figure 4.2. Only one of the f.lt is equal to 1, at a given site
and time step. When flO = 1, the particles are not deflected and, locally,
142 4 Diffusion phenomena

Po p p

Fig. 4. 1 . How the entering particles are deflected at a typical site, as a result of
the diffusion rule. The four possible outcomes occur with respective probabilities
Po, Pt . pz and P3 · The figure shows four particles, but the mechanism is data-blind
and any one of the arrows can be removed when fewer entering particles are
present.

there is a ballistic motion. On the other hand, when J.l2 = 1 the particles
bounce back to where they came from (rotation by 180° degrees). The
other two cases (/-l l = 1 or J.l 3 = 1) give clockwise and counterclockwise
90° rotations. We also observed that if J.l l = J.l3 = 0 everywhere on the
lattice, equation 4.6 reduces to two one-dimensional random walks along
the horizontal and vertical directions.
The J.lt are 1 with probability pc, otherwise they are zero. Consequently
their average value is
=
(J.lc ) Pt (4.7)
For symmetry reasons, it is natural to impose the condition that
(4.8)
because the particle is equally likely to be deflected clockwise as counter­
clockwise. Thus, the probabilities Pt satisfy
Po + 2p + P2 1 = (4.9)
It is interesting to notice that a model similar to this one has been
 4.2 The diffusion model 143

Fig. 4.2. The four lattice directions.

considered by Hauge [109] , for the case of continuous space and time.
We shall now derive the macroscopic limit of equation 4.6 following the
same procedure as developed in section 3.1.
The operation o f random rotation i s spatially uniform and data-blind,
i.e. it is produced by an external mechanism which is the same for all sites
and which is statistically independent of the actual occupation numbers
ni of particles at a given site. Consequently,
( J.1t ni+t l = Pt (ni+tl = Pt Ni+t (4. 10)
and equation 4.6 simply becomes
Ni('r + A.£\, t + -r) - Ni('r, t) = (po - 1)Ni + Pt Ni+l + P2 Ni+2 + P3 Ni+3 (4. 1 1)
This equation is the lattice Boltzmann equation associated with the dif­
fusion rule. As opposed to the hydrodynamical situation discussed in
the previous chapter, it is exact and has not required any factorization
assumption. This is due to its linearity, of course.
We now write the left-hand side as a Taylor expansion

[-rat + � a; + A.(ci . Or) + � ]

Ni Cr + A.i:\, t + -r) - Ni(' r, t) =
(ci . Or) 2 + TAO t (Ci . Or) Ni
(4. 12)
and we write, as usual,
Ni = N}0 + cN[ i + ...l l (4. 13)
With the condition (see equation 3.26) that
and
the order 0(€0) of equation 4. 1 1 reads
O(c0 ) :
QN (O) = 0
144 4 Diffusion phenomena
where
·-1 Pl P2 P3 )
c
!1 = P3 Po - 1 Pl P2 (4.14)

( l
P2 P3 Po - 1 Pl
Pl P2 P3 Po 1
-
and
N1(O)
(0)
N(O) - NN2(0) (4.15)
3O
N4( )
Since L Pt = 1, N (O ) oc Eo is clearly a solution, where Eo = (1, 1, 1, 1)
i s the collisional invariant which expresses particle conservation. Except
for pathological values of the Pt this is the only solution. Using the
Chapman-Enskog requirement
4
p = I: NfOl (4.16)
i= l
we obtain that
N (O) = !!__4
l (4.17)
The next order O(c) in equation 4.1 1 is
O(c 1 ) : -i (ci · 8r )N{0l = I: nij NY l (4.18)
j
In order to solve this equation, it is interesting to notice that
3
2: PtCi+t = (1 - 2p)c i ( 4.19)
t=O
Indeed, we have ci = -ci+ 2 and
=
po c i + pci+ l + P2 Ci+2 + pci+3 (po - P2 )ci
Therefore, the vectors E1 and E2 defined as

(4.20 )

are eigenvectors of the matrix n, with eigenvalue Po - P2 - 1 = -2(p + P2 ) .

 4.2 The diffusion model 145
In other terms,
4
L Qij Cja -2(p + P2 )Cia
= (4.21)
j= l
With this observation it is easy to solve equation 4.18. We write
N? ) c/> ciaOa P
=

where we have used the summation convention over spatial coordinate

index 11. = 1, 2 and 4> is a constant to be determined. By substitution into
equation 4.18, we obtain

� (ci · or )P z::.:: nij Nj l l

j
= - 2(p + P2 ) c/> (ci · or)P
and

4> =
-
A (4.22)
8(p + P2 )
This yields the solution
A
8( p + P2 ) CiaOa P (4.23)
Note that when p + P2 = 0, this derivation is not possible. But, then, we
are in the situation where p0 = 1 which corresponds to a ballistic motion
without any direction change.
The equation for p is then obtained in the usual way, by summing over
i equation 4.1 1. Of course, I: Qi = 0 because particle number is conserved.
Using the Taylor expansion 4.12, one has, at order O(e)
4
O(E) : L [A(ci · or)N}0l ]
0 (4.24) =

i= l
which is obviously satisfied since N}0 l p/4 and I: ci 0. More interest­
= =

t, [TotN}0l + A(ci or)N?l + � (ci or )2 N}0ll

ing is the second order

· · = 0 (4.25)
With 4.23, the equation reduces to
(4.26)
1 46 4 Diffusion phenomena
Since 2::: c iaCif3 = 2l5ap , as obtained in section 3.5, we have
O t P + div [ -D gradp] 0 = (4.27)
where D is the diffusion constant
D
-1
=�
2 ( 4(p +
1 1 ) =
-1 2
�
( 4[1 -p +(p Po+ Po)] ) (4.28)
P2 ) -4
and J = -Dgradp is the particle current. It is interesting to observe that
J does not coincide with the expression 2::: N/J i which is the standard
definition of the particle current. Indeed, we have

ti=l N/Ji = - (D + 4'[,12 )

and thus we have

gradp (4.29)

This discrepancy by a factor of (-1 2 /4-r) can be interpreted as a renormaliza­

tion of the diffusion constant due to the discreteness of the lattice (because
it comes from the second-order terms in the Taylor expansion 4. 12). It is
analogous to the lattice viscosity obtained in section 3.2.3.
Relation 4.28 shows that the diffusion coefficient is adjustable in this
model. It can be made very small when P2 is close to 1 and very large
when Po is close to 1 . Practically speaking, however, not all choices of Pt
are as good as others in a finite system. We will return to this question in
section 4.3.
In the case p = 0, the particles follow one-dimensional trajectories. It is
important to notice that setting p = 0 into 4.28 does not give the correct
diffusion coefficient derived in equation 3.38, namely

D ld = ,1-r2 2(1 Po- Po)

Instead, there is a factor of 4 in the denominator. The reason is that,
when p = 0, there is another conserved quantity in addition to the particle
number. Due to the one-dimensional motion, the number Nx of particles
along the x-axis and the number Ny of particles along the y-axis are both
constant. As a result Nx - Ny is a conserved quantity. It is associated
to the extra collisional invariant E4 = ( 1 , - 1, 1 , - 1 ) which is the fourth
eigenvector of n, with eigenvalue -4p. Therefore, if p = 0, the local
 4.2 The diffusion model 147

equilibrium solution Ni =p / 4 is not the most general solution (which is

then N1 = N3 = P x/2 and N2 = N4 =
P y/2).

4.2.2 The mean square displacement and the Green-Kubo formula

In the previous section, we obtained the value of the diffusion coefficient D
by considering the continuous limit of the dynamics of a system of particles
performing a synchronous random walk, with an exclusion principle. The
same coefficient D (as found in 4. 2 8 ) can also be obtained by considering
only one particle and calculating its mean square displacement
0"2 ( t) = (x2 (t) + i(t))
where
t - l
(x(t), y(t)) = r L: v(n)
n=O
and v(n ) is the velocity the particle has at time step n (which is one of the
vi = (A./r)£\). Thus,
t - l t- l
four possible velocities

0" 2 (t) = \r2 n=O

I: L: v(n) . v(m))
m=O
There is a well-known relation between the mean square displacement and
the diffusion constant. When the probability of having a particle at a given
site obeys a diffusion equation, it is easy to show the following identity,
provided the probability and its gradient vanish on the boundaries
6 2 (t)
D = ___!_2d_ tlim
---+ctJ t (4. 30)

where d is the dimension of the space. This is Einstein's relation [1 10] .

0"2
The value of can b e calculated in our model. According to the cellular
automata rule, each particle performs a random walk, with probability Po
of going straight, p2 of bouncing back and p of being deflected by 90°
degrees right or left.
For the purpose of the calculation, let us assume that the particles
diffuse in the complex plane. Let us write
A. n ·e
v(n) = -el 1:

where 8 n is the angle formed by the velocity vector at iteration n. One has
148 4 Diffusion phenomena
where dj E {0, ±n/2, n } is the random variable giving the rotation at time
j . With these definitions and the notation v( m) = (A./r)e - i8m , we can write
t- 1
(J2 = r2 L (v(n)v(m))
n, m = O

n<m n=m n>m

t - 1
A2 t + A_2 L ( (e i(Bn -Bm ) ) + ( e -i(Bn -Bm ) ) )
n>m
Now, we have
n
( ei( Bn -Bm ) ) (ei l::;�m+l f:.j )
= = IT ( eif:.j)
j=m+1
because the dj are independent random variables. Since dj is 0, n/2, n or
- n /2 with probability po, p, P2 and p, respectively, we have
= =
(e if:. . ) Po + z.p - P2 - z.p Po - P2 ( e -if:. )
1 = 1

Therefore,

and we obtain
t- 1
(J 2 ( t) A2 t + 2A.2 L (Po - P2t-m
n>m
2 2 t- 1 k
A. t + 2A. L L (Po - P2 ) 1
k= 1 1 = 1
[k
A.2 t + 2A.2 (po - P2 ) � ((Po --Pl)) - 11 ] (4.31)
k= 1 Po P2 -
After summation over k, it is found that the mean square displacement is

(4.32)
provided that the lattice is infinite or the time short enough so that the
particle has not reached the limit of the system. This relation leads
 4.2 The diffusion model 149

to a value of

in perfect agreement with equation 4.28.

In statistical mechanics, it is common in the framework of the Green­
Kubo formalism to express transport coefficients in terms of the integral of
autocorrelation functions [87, 109] . We are now going to explicitly derived
the Green-Kubo relation for diffusion. We shall see that here, due to the
lattice discreteness, a correction to the usual formulation is present [1 1 1] .
Using equation 4.30 and the expression for CJ 2 in terms o f v(n) and v(m),
we can write D as (note that (v(n)i3(m)) = (v(n - m)i3(0)) and (v(O)i3(0)) =
(1) = 1)
T 1 N- 1 - N 1
D = lim - L (v(O)i3(0))
2 d N-> oo N k=O + _dT Nlim N1 n>m
L (v(n - m)i3(0))
----> 00
- (4.33)

Since (v(n - m)i3(0)) oc (po - P 2 ) n - m , the summation over n and m can be

easily performed and, after the limit N ---4 oo is taken, we get
1 N- 1 . 1 N- 1
k
lim
N-> L (v(n - m)i3(0))
oo N n>m hm - L L (v(l)v(O))
N-> oo N k=1 1=1
00

L (v(j)i3(0))
j=1
00

L (v(j)i3(0)) - (v(O)i3(0)) (4.34)

j=O
Therefore, the diffusion constant is given as a sum over the velocity
autocorrelation
00
T
D = d L (v(l)v(O)) - 2dT (v(O)v(O)) (4.35)
1=0
Equation 4.35 is equivalent to a Green-Kubo relation. In our case,
however, there is a correction -(r/2d) (v(O)i3(0)) due to the discreteness of
the lattice. This term has the same origin as the difference between the
particle flux J and pu we mentioned in section 4.2. 1. Mathematically, it
comes from the fact that the contributions = l in the double sum over
k
k and l is not negligible in a discrete system.

4.2.3 The three-dimensional case

In a three-dimensional space, our cellular automata model of diffusion
can be easily extended if the "rotations" that take place at each node of
150 4 Diffusion phenomena

the lattice are properly chosen. On a cubic lattice, we have six possible
directions of motion which can be mixed in several different ways, in
order to produce a random walk. Among the set of all the possibilities,
one of them gives equations very close to the two-dimensional case. For
this reason, we shall not repeat the details of the calculation. We shall
simply describe how to produce the random walk and write down the
corresponding macroscopic equation.
Let us first label the three main directions of the lattice c1, c2 and c 3 ,
with the property that c 3 = c1 x c2 . The three other directions are defined
according to the relation
i being wrapped onto { 1, 2, 3, 4, 5, 6 }
Now, we consider six operations �. k
0, .. , 5, that occur with probability
=

Pk · The result of the operation � is to deflect the particle entering a

site with velocity v i = ( Ji/r )ci into direction c i+k · The lattice Boltzmann
equation then reads
5
Ni (r + ACj, t + r ) = L Pk Ni+k (r, t)
k=O
Symmetry requires that we choose
=
P l P2 P4 Ps P
= = =

Since the probabilities sum up to 1, we have

Po + 4p + P3 = 1
Following the same calculations as we performed in section 4.2. 1, we
obtain that the particle density p = L: Ni obeys a Pick's equation with a
diffusion coefficient
A 2 (1 + Po - P3 ) Ji2 Po + 2p
Dm = =
r 6(1 - Po + P3 ) r 6( 1 - (Po + 2p))
4.3 Finite systems

The results obtained in the previous section are exact in the limit of
an infinite lattice with Ji ---+ 0, r ---+ 0 and Ji 2 jr ---+ const. In practice,
however, one never performs numerical simulation in this limit and some
corrections to Pick's equation are expected. In this section, our purpose
is to solve the lattice Boltzmann equation for finite systems, with various
boundary conditions. As a result, we will be able to determine the length
and time scale for which a fair modeling of diffusion is obtained.
 4.3 Finite systems 151
Since the cellular automata dynamics is linear, the lattice Boltzmann
equation is exact and also linear. For this reason, exact solutions can be
obtained and discrepancies with the continuous limit discussed rigorously.

4.3.1 The stationary source-sink problem

There is a simple situation in which our lattice Boltzmann diffusion equa­
tion can be solved exactly : this is the stationary source-sink problem
which consists of two plates separated by a distance L. One plate contin­
uously injects particles in the system while the other absorbs all particles
reaching it.
If one waits long enough, a density profile establishes in the system
and the average occupation number Ni no longer depends on time. For
the steady state, the lattice Boltzmann equation 4.1 1 associated with the
diffusion rule reads
N1(x + A, y) = po N1(x, y) + pN2 (x, y) + P2 N3 (x, y) + pN4 (x, y)
N2 (x, y + A ) = po N2 (x, y) + pN3 (x, y) + P2 N4 (x, y) + pN1(x, y)
N3 (x - A, y) = po N3 (x, y) + pN4 (x, y) + P2 N1(x, y) + pN2 (x, y)
N4 (x, y - A ) = po N4 (x, y) + pN3 (x, y) + P2 N4 (x, y) + pN1(x, y) (4.36 )
This system of equations has to be supplemented by boundary conditions.
We shall assume that the source and the sink are the vertical planes
x = -1 and x = L + 1, respectively. Consequently, we have
N1(0, y) = 1
For symmetry reasons, we shall look for a solution which does not depend
on the y variable and such that N2 = N4 .
In the case of a continuous diffusion equation, it is straightforward to
check that the solution of V2 p = 0 is a linear expression in x . Following
this indication we try the solution
N1(x, y) = ax + 1 N3 = a(x - L)
E
for x 0, A , 2A, . . . , L. After substitution, we are left with the conditions
a(p2 L + A) - 2pb Po - 1
apL + 2pb p
a(po L - A - L ) - 2pb P2 (4.3 7 )
Among these equations, only two are independent since adding the first
to the third gives the second. We then obtain
1
L + A (4D. + 1)
152 4 Diffusion phenomena

(
and

where
b=
1 - aL
2 =! 1+
2
1
1 + t (4D. + 1 )
D . i s the dimensionless diffusion coefficient
)
D. = .!_2 D = P + Po = P + Po
2 4(p + P2 ) 4 ( 1 - P - Po )

-
Therefore, we obtain
1
NI(X' y ) - 1 - 1 + t(4D. X
+ 1) L
1 (L - x)
N3 (x , y ) _- 1 + t (4D.
+ 1) L
and finally, since b = (1 - aL ) /2 ,

N2 = N4 = 21 (N1 + N3 )
We observe that, in the limit A. � �
0, a -(1 / L) and b = 1. Thus, the
density
p = N1 + N2 + N3 + N4 = 4ax + 2 ( 1 - aL)
tends to 4 ( L - x) /L , as expected for the solution of ( d2 pjdx2 ) = 0 with
the condition p (x = 0) = 4 and p (x = L) = 0.
In contrast, when A. is finite, some corrections (of the order £ ) are
present. However, the density profile is always a straight line. But the
values at x = 0 and x = L differ from 4 and 0, respectively. At x = L/2 ,
the discrete and continuous solution intersect, for the value p = 2.
Finally, it is interesting to compute, in this simple case, the particle
current j = v[N1(x + A.) - N3 (x) ] , that is the net number of particles
that have crossed an imaginary vertical plan located at x + ( A./2 ) (note :
N1(x + A.) is the probability of having a particle move from x to x + A.
'L
during the time while N3 (x) is the probability of having a particle move
from x + A. to x during the same time interval ; this not equivalent to
N3 (x + A. ) - N1(x)) .
We have
A. 22
J = - (1 + aL + a.A.) = 4-D.
1
= 4Da
'L L + A.(4D. - 1) -
'L

Since

�
 4.3 Finite systems 153
we find that the particle current is
j = -Dgradp
This relation, which is the usual one in a diffusion process is true even if
the lattice spacing does not tend to zero.

4.3.2 Telegraphist equation

When both A and r are finite, the propagation velocity v = Ajr is also finite
(note that, in contrast, v is infinite when A, r � 0 and r oc A -2 ). Therefore,
information propagates at a finite speed in the cellular automata and the
diffusion equation 8 1p = DV2 p cannot be true. Indeed, with this equation
any variation of p is instantaneously transmitted arbitrarily far away from
where it occurred. There is, however, an equation which combines diffusive
behavior with a finite propagation speed of information. This is the well­
known telegraphist equation. We will see that our cellular automata rule
naturally leads to this equation when a finite lattice is considered.
Using a matrix notation for the Ni , the lattice Boltzmann diffusion
equation can be cast in the following form
TN(t + r) = (Q + l)N(t)
where T is the spatial translation operator. In one dimension, this reads
T = Tx (
0 T_x
0 )
where T±x is a translation along the x-axis of ±A
T±xNi (x, t) = Ni (x ± A, t)
The collision operator Q is the usual matrix made of the Pi· In one
dimension, this reads

Q= ( po1 --Pol
1 - po
Po - 1
)
In two dimensions, T is diagonal and Tii is composed of either a transla­
tion T±x or a translation T±y along the y-axis. The collision matrix Q is
given in equation 4. 14.
The Q matrix can be diagonalized easily. For the sake of simplicity, we
shall consider the one-dimensional case. We introduce the variable M
N = PM
where P is the transformation matrix
1
p = J2 1 1
(
1 -1 ) (4.38)
1 54 4 Diffusion phenomena

( )
Its inverse reads
p -1 = _1 1 1 (4.39)
J2 -1 1
and one has for the two components of the column vector M
M1 = J21 (N1 + N2 ) = Jl
1p
and
1 ( - N1 + N2 ) = - -pu
M2 = J2 1
-
Jlv
One immediately checks that
p - 1 (Q + 1)P = D =
(� o
2po - 1
)
The microdynamics of the diffusion rule reads
TN(t + ) = (Q + 1)N(t)r

With the transformation N = PM , this is equivalent to

p - 1 T P M ( t + ) = DM(t) r

From this equation we can write two relations

p -1 T P M( t) = DM(t - r ) (4.40)
and
M ( t + ) = p - 1 r- 1 p DM(t)
r (4.41)
where r- 1 is the inverse translation
y-1 = (
Now, it is easy to check that
p -1 yp = !2 ( -TxTx ++TT_xx -TxTx ++TT_xx )
- -
= ! ( Tx + T_x Tx - T_x )
whereas
p - 1 y-1 p
2 Tx - T_x Tx + T_x
and we see that the off-diagonal terms come with opposite signs. More
generally, in higher dimensions, a similar result holds : some terms change
sign and others do not. This is related to the properties of the Mi to change
sign under the transformation x � - x (i.e N1 � N2 and N2 � NI) . This
fact can be used to easily eliminate some of the Mi from the equations. In
 4.3 Finite systems 155
one dimension, there is just M2 to eliminate and we obtain an equation
for M1 only.
The first equation of 4.40 and 4.41 gives

M1 (t + r )
1 2po - 1
2 ( Tx + T_x)Ml (t) - 2 ( Tx + T_x)M2 (t)
1 1
M1 (t + r) 2 ( Tx + T_x) Ml (t) + 2 ( Tx + T_x)M2 (t) (4.42 )
and by simple substitution of ( Tx + T-x)M2 (t) from the second line into
the first, we obtain
M1 (t + r) = Po
2 ( Tx + T_x) Ml (t) - ( 2po - 1)Ml (t - r) (4.43 )
In order to put this equation in a more convenient form, we introduce the
discrete differential operators
1
8A t Ml (t) = (M1 (t + r) - M1 (t - r))
2r

and
1
VA 2 M 1 (t, x) = ,P (M1 (t, x + A) - 2M(x, t) + M1 (t, x - A))
These finite difference operators tend to their usual counterpart in the
limit A � 0 and r � 0. They also correspond to the standard way of
discretizing space and time derivatives.
Equation 4.43 can be written as
� (M1 (t + r ) - M1 (t - r)) + b(M1 (t + r) - 2Ml
+ M1 (t - r)) = Po (Tx + T_x - 2 )Ml
with
a = 2(1 - Po ) b = Po
Therefore, we obtain

8A t M1 + rPo aA t2 M1 = A2 Po_ VA 2 M1
2(1 - Po ) � 2( 1
Po )
We have p = J2M 1 , D = (A2 /r ) [po / ( 2 - 2p0 ) ] , the diffusion coefficient
we defined previously and c2 = (A2 /r 2 ), the speed at which information
travels in the system. With these definitions, we obtain that the density of
particles obeys a discrete telegraphist equation
A D A2 = A 2
8 t p + 2a
c t
p DV p (4.44)
156 4 Diffusion phenomena

Equation 4.44 is exact and involves no approximation. It is true for all x

except at the system limits where boundary conditions prevail.
The telegraphist equation describes a situation in which a signal prop­
agates at finite speed and is subject to diffusion. It is the presence of the
second time "derivative" which guarantees a finite propagation speed of
information in the system. When c tends to infinity, it can be neglected
and one obtains a finite difference Fick's equation.
In the case of periodic boundary conditions, another way to obtain the
telegraphist equation is to consider again the dispersion relation derived
in section 3.1 (equation 3.9). It can also be written as (with the current
notation, Po stands for p)
cos kA = exp(ian) + (2po2p-o 1) exp( - ian)
exp(iwr) + exp(-iwr) + (1 Po ) exp(iwr) + exp( - iwr)
-
2 2po
consequently we have
Po
z. sm. wr + -1-- cos wr = --
Pv
cos
A k ,
- po 1 - po
Expanding this relation up to second order in wr and kA and using the
fact that in Fourier space one has the correspondence
a t iw V2 +-+ -k2
+-+

one gets back to the telegraphist equation.

The agreement between our dispersion relation and the' telegraphist
equation can be seen graphically. It is straightforward to show that
f ( x, t) = exp(iwr - kx) is solution of
D 2 2
c at p = DV p
8tP + 2
provided that
( GD2k2)
iw = _ !!__ 1 + y l - �
2D

. [ (
or equivalently
e1wr = exp - -c2· 1 +
2D.
where D. = (r/A.2 )D and c. = (r/A.)c are the dimensionless diffusion
coefficient and speed of sound, respectively. The above expression is the
dispersion relation for the telegraphist equation. It can be checked against
 4.3 Finite systems 1 57

D,= 1 . 5 D,= 1 .5

�
� o r-
* I
/
/

� 1
0 kf...
___ __________,._,

1t 0 ---- ---
- 1 '::-
kf...
-
:-;;----
-------"
1t

Fig. 4.3. Comparison of the real and imaginary parts of the dispersion relations
of (i) the cellular automata model (solid line) ; (ii) the telegraphist equation
(dashed line) and (iii) Fick's law of diffusion (dotted line). Notice that some
curves may have two branches because of the ± sign in front of the square root.

r
equation 3.10. This is done in figure 4.3, where eiwr is given as a function
of kA. We also show the dispersion relation of the pure diffusion equation
eiw = e-D. (U.jl
We can observe that our cellular automata diffusion model is much closer
to the telegraphist equation than Fick's law. Of course, at large time, only
small k contributions will survive (i.e those for which eiwr is close to 1 )
and in this limit all three evolutions correspond. Note however that k ,....., n
also produces long-lived terms. As we previously mentioned, this is related
to the checkerboard invariant that plagues cellular automata dynamics.
4.3.3 The discrete Boltzmann equation in 2D
The same approach as described in the previous section can be applied
to a two-dimensional system. Of course, the algebra becomes more
complicated. Here, we would like to mention some results, without giving
a complete derivation. More details can be found in [1 1 2] .
In the 2D case, the discrete Boltzmann equation reads
r =
T N(r, t + ) (Q + 1 ) N(r, t)
where N is the column vector composed of the Ni, Q + 1 the four by four

( P2�P P2:Po P�Po P:po2 )

matrix
0+1 =
158 4 Diffusion phenomena

and T the spatial translation matrix defined in the previous section

Tij f C r ) = f(r + A ci)<5ij

( ) ( )
The matrix n + 1 can be diagonalized with the linear transformation
1 -1 1
1 � �
p = 2 1 -1 --11 -11 �
and p - 1 = 2 1 -1 -11 - 1
1 � _1 �
1 -1 1 1 -1 1 -1
so that
0 0

JJ
Po - P2 0
0 Po - P2
0 0
In terms of the new quantities
M = p -1 N (4.45 )
and, following the procedure of section 4.3.2, one obtains
D. 2 A 2 = A2 A
ro" t M1 + 2r o t M 1 D.VM 1 + A2
Po - P (-oA x2 + oA 2 ) M4 (4.46)
c. 4(P + P2 ) y
A2 Po 2+ P ( _ a 2 + ay2 ) M1
X

+A2 PO 2- p (V2 ) M4 (4.47 )

where a t and a ; are the discrete time derivative and a� , a; and V2 the
second-order discrete space derivative.
The dimensionless diffusion constant D. and speed of sound c. are
d�fined in agreement with section 4.2.1 and read
D. = 4(1 -Po(p++P po )) and c2. = 21
and the coefficients a4 and b4 are
a4 = 1 - (po - P2 )(1 - 4p) and b4 = p + Po - 2p(po - P2 )
Similarly, it is found that M2 and M3 are related to M1 and M4 . In the
case Po = p, the equation for M1 decouples from the others and we simply
obtain that the density p = 2M1 (see 4.45 ) obeys the discrete telegraphist
equation
A + 2D oA t2 p = DVA 2 p
DtP c
From the equ ations for M1 and M2 , one can derive the dispersion
relation of our discrete diffusive process. In the case of a periodic system,
 4.3 Finite systems 159
(a)
,------,-""(c"-
) --=----:-=-� 1.5 ,--------c--'("-
e) --� 0.5
p=0.25 D,=O.i p=0.5 D.=0.25 p=0.05 D,=i .O

k,"A !l=(J.87

1!=0.93

J.L=0.99

(b)
,.------:--!
: (--'-
d) --=-----:::
:-:c :--1 1.5 ,-------:--!(1)-'---::-
-c -:
-:; -,
:-- 0.5
p=0. 1 8 D.=O.i p=0.3 D,=0.25 p=0.2 D,=l.O

Fig. 4.4. Contour lines of the diffusive part of the dispersion relation (J..L =
exp(iwr)) for various values of D. and p. The solid lines correspond to the CA
model, the dashed ones to the telegraphist equation and the dotted ones to Fick's
law of diffusion. For f.l = 0.99 the three dynamics are almost indistinguishable.
Note that the scales are not the same in all figures.

equations 4.46 and 4.47 can be Fourier transformed and the solution takes
the form
M1,4 - A 1,4 e i(wT+k·r) _

provided that w is a function of k. The dependence w = w(k) is shown

in figure 4.4. We observe that the choice of the Pi affects the dynamics
at short time and for small wave length. However, in the macroscopic
limit, the dispersion relation of the automaton is in agreement with the
behavior of a continuous diffusion process.
4.3.4 Semi-infinite strip
In this section, we consider the solution of the discrete Boltzmann equa­
tion in the case of a two-dimensional stationary problem. For a time­
independent diffusion process, Fick's law reduces to the Laplace equation
(4.48 )
with appropriate boundary conditions.
 160 4 Diffus ion p henomena

The geometry we want to study has the shape of a well whose vertical
walls are of infinite height and absorb particles. On the other hand, the
bottom line of the system acts as a source of particles. More precisely, we
shall impose the following conditions
=
p(x = O, y) = p(x L, y) = 0 for y � 0 (4.49)
p(x, y = 0) = cjJ = const for 0 x L < < (4.50)
and
ylim
-><XJ
p=0 (4.51)
The effect of a finite lattice spacing and finite time step are quite important
if L is not large enough. We shall also see that the choice of the Pi is
relevant even though the diffusion coefficient is not present in the Laplace
equation 4.48.
The exact solution of 4.48 with the above boundary conditions is
. ( )
known to be [113]
p=- 4 c/J " 1 nn y nn x 2 c/J _1 sin( y )
L.- - exp ( - - ) sm ( - ) = - tan (4.52)
n odd n
n
L L n sinh( i )
For a stationary state, the lattice Boltzmann dynamics 4.46 and 4.47
read
(4.53)
and
=
(-o" x2 + o" y2 )Ml + aV" 2 M4 f3 M4 (4.54)

--
where a and f3 are defined as
a = Po -+ pp f3 = 16pp(p++p P2 )
Po o
In order to solve 4.53 and 4.54 for the boundary conditions we defined
at the beginning of this section, we first assume that c 1 and c2 are,
respectively, parallel to the x- and y-axis. Next, we assume a solution of
the following form
M 1 ,4 = L A l ,4 (n) exp(z"y) sin(w"x) (4.55)
n

where
Q) n =­ nn (4.56)
L
z"
and has to be determined from 4.53 and 4.54.
 4.3 Finite systems 161
This calculation turns out to be straightforward because the discrete
derivatives have simple eigenfunctions. Indeed, the functions exp(z"y) and
sin(cvn x) are eigenfunctions of the discrete operators a� and a�
a� sin(wn x) = �� sin(cvn x)
A b
ay2 exp(z"y) = 2z exp(z"y)
A
From the simple properties of the sin and exp functions, we find
bw = 2(cos( L n n A) (4.57)
- 1)
and
bz = exp(z"A) + exp( -z"A) - 2 (4.58)
After substitutions of 4.55 into 4.53 and 4.54, we obtain
(bw + bz )Al + cx (-bw + bz )A4 0 =

(-bw + bz )A l + [cx ( bw + bz ) - {J ] A4 0 =

In order to find a non-zero solution, we obtain a relation between w" and

zn , namely
(4.59)
Thus, the ratio
ex

{3 16 P(o -+PP2 )
p p
(4.60)
is a free parameter of the model. Its value plays a role when the quantity
A/ L does not tend to zero. Clearly, rx/ f3 can be made arbitrarily large by
choosing p small enough. On the other hand, its smallest value is
!X 1
f3 8
=
obtained for p 1/2 and Po P2 0.= =
=
It is then straightforward to calculate the density p 2Ml from 4.55
=
because, for y 0, our expansion is identical to 4.52. Therefore
p =
4</> L ! [exp(z"A)] (Y /"-l sin( nnx ) (4.61)
n n
n odd L
Strictly speaking, the density is only defined on the lattice sites but
equation 4.61 provides an interpolation of p everywhere. Its similarity
162 4 Diffusion phenomena

�=0.0
IX

�=0.1
IX

�=0.5
1 (a) L=20'A
1 n 20
Fig. 4.5. Comparison between the solution of the Laplace equation and the
solution of the CA dynamics for various values of rx/ p and for two system
sizes (a) L = 202 and (b) L = 1 002. The dashed line corresponds to the exact
solution exp -nn (in a log scale). This line turns out to coincide very well with
rx/P = - 1 /24 � -0.04 1 7.

with 4.52 makes it very appropriate for a comparison with the continuous
solution of Laplace equation. Since
y

exp(zn y) ( [exp(z).)] (L/.l.) )

= r

the convergence of the discrete solution can be checked by comparing

exp(-nn) with [exp(zn.A.)] (L/.l.)
This is shown in figure 4.5 , where the logarithm of these two quantities
rx
is shown for various values of A./ L and / P. One sees that the significant
deviations that occur when the system size is too small become negligible
as the lattice spacing A. decreases. To draw this graph, we have solved
equation 4.58 for exp(zn .A.),
exp(zn .A.) 1 + i ± � Jb� + 4hz
=

and computed hz as a function of bw, as given by 4.59. In figure 4.5 , one

observes that the dashed line, which represents the solution in the limit
A. � 0, is approximated with more or less accuracy, depending on the
rx . rx =
value of jp The curve jp -1 /24 turns out to be quite good, even
=
for a small system L 202. It can be shown [1 12] that this value yields a
 4.4 App lications of the diffusion rule 163
correct solution to the Laplace equation up to order (2/L)6. Therefore, in
a finite system, the choice of the Pi may play a role in the quality of the
solution.
4.4 Applications of the diffusion rule

In section 4.2, we presented a cellular automata model of synchronous

random walks and discussed how well the discrete dynamics reproduces
the behavior of a diffusive process. The examples we have studied are
rather trivial from the view point of physics and there is no reason to
use a cellular automata to solve Laplace equation in a one-dimensional
source-sink problem, or in the semi-infinite strip.
However, there are several physical processes involving diffusing par­
ticles that are much more complex and difficult to describe in terms of
a solution of the diffusion equation. On the other hand, the cellular
automata approach is quite natural and well suited to perform numerical
simulation.
This section describes three applications of the diffusion rule, namely
the properties of a diffusion front, the so-called DLA aggregation process
and, finally, the problem of random deposition on a surface.
4.4. 1 Study of the diffusion front
The diffusion model we have introduced in this chapter can be used to
create and study rough interfaces between two media in contact. Most
solid interfaces observed in nature are irregular or rough on all length
scales and the fractal geometrical properties of such interfaces are of great
importance to materials science. The way to model an interface is to have
particles diffuse in a two-dimensional system, from a source on the left,
to a sink on the right. Before the density gradient reaches a steady state,
two regions can be identified: that made up of particles and connected to
the source and the other one, made up of holes (empty lattice sites), and
connected to the sink.
As practical examples of physical systems in which the interface plays
an important role, we can mention solid-solid chemical reactions, al­
loys formed out of the melt and many processes in solid state technol­
ogy [1 14,1 15]. Catalysis, corrosion, crystal growth and the physics of
superionic conduction [116] are other phenomena where the interfacial
properties are directly relevant. Finally, fractal interfaces occur naturally
in invasion percolation of a porous medium [117].
In addition to modeling fractal interfaces, the nonstationary diffusion
process is related to the problem of site percolation [1 18], and diffusion of
164 4 Diffusion p henomena

particles in a concentration gradient provide an ideal means to compute

percolation properties, notably the percolation threshold Pc and the fractal
dimension characterizing large clusters. The main advantage of this
approach is that one does not have to know or guess the value of Pc in
advance : it is determined self-consistently.
In the present example, we consider a system with a perfect line source
at y = 0 which replaces immediately particles diffusing into the system
=
and a line sink at y Ly which removes all arriving particles. In the
perpendicular direction, the system has a size Lx with periodic boundary
conditions. At time t = 0 the system is completely empty. Although the
diffusion process does conserve the number of particles inside the system,
the total number still changes, because of the source and the sink.
The diffusion front is defined through particle connectivity, as in site
percolation: any two particles sitting on nearest neighbor sites belong to
the same cluster. For percolation, a threshold Pc separates a region where
only finite clusters are present from the region where an infinite cluster
occupies a fraction of the system. In our model, p varies from p 0 to =
=
p 1 across the sample so that the finite cluster and the infinite cluster
>
regimes are present simultaneously. Close to the source p Pc and one
<
cluster spans the system. Close to the sink p Pc and only finite clusters
are present. In a weak gradient, large finite clusters have the structure of
percolation clusters and are located near Pc ·
More precisely, the diffusion front, also called external perimeter or
hull, is obtained according to the following procedure
(1) Determine the infinite cluster ("the land") as the set of all particles
which are connected to the source by nearest-neighbor particle­
particle bonds.
(2 ) The complementary infinite cluster ("the ocean") consists of the non­
land sites which are connected by nearest-neighbor bonds to the sink.
By this process, the islands (particles not connected to the source)
are removed.
(3) The diffusion front (or the "shore") is built as the set of land sites
which are nearest neighbor to at least one ocean site.
Different fronts can be defined by extending the connectivity from
nearest neighbors to second nearest neighbors [118]. Different geometrical
properties are then expected.
Note that the front construction, as given by the three above steps, is
the result of a cellular automata rule. The land and the ocean are obtained
by iterating a "coloring" rule : first, the source is painted, say, black. Then,
each cell which is occupied by a diffusing particle and which contains a
 4.4 App lications of the diffusion rule 165

.
. �

. . .

(a) ( b)

( c) (d )

Fig. 4.6. The diffusion front is the boundary between the source connected
cluster and the sink-connected cluster. The figures shows the four steps necessary
to extract the front : (a) the location of diffusive particles ; (b) the "land" cluster ;
(c) the "ocean" in gray ; and (d) the fractal front in black.

black site in its von Neumann neighborhood turns black, too. A similar
procedure is applied from the sink to determine the ocean.
Figure 4.6 illustrates these various steps, from the configuration of
diffusing particles to the construction of the front.
As we can see from figure 4.6, the diffusion front is a complex object.
Depending on the time at which it is obtained (i.e. the number of iterations
of the diffusion rule before rules (1) to (3) are applied), its width varies.
The longer the diffusion takes place, the wider the front is.
(J
Practically, the width is defined as follows. By symmetry, the front
properties are similar along the x-axis. Let us call P1(y, t) the average
166 4 Diffusion p henomena

density (over x ) of hull points at location y and time t. The average front
position YJ is defined by:
00
YJ ( t) = j0 yPPJj (y,(Y,tt)) ddyy
fow
and the width of the front a is given by
(J
2 - fow (y -00 YJ) 2 Pf (y, t) dy
'---"---
-- --:o=----'-----'c--.,---
-' -
j0 Pj (y, t ) dy
The relation between the width a and the time t determines the roughness
of the front. Typically, one expects that
oc
a ( t ) t'1 (4.62)
a
where is the roughness exponent.
In addition, the relation between the number
Nf ( t) = lo w Pf (y, t) dy
of particles in the front and its width a defines the fractal dimension DH
of the hull as
Nf = AaDH- l
where A is some amplitude. Here, we have subtracted -1 to DH because
we consider the dependence of Nf on a only. Clearly, Nf also depends
linearly on the front length Lx.
Roughness and fractal dimension have been measured according to the
above definitions and the results are given in figure 4.7. From the slope
of the log-log plot of Nf againts a and a against t, we obtain
a � 0.27
respectively. These values are in good agreement with those predicted by
independent theoretical arguments. The simulation which produced these
results has been obtained with the values p = 1/2 and Po = P2 = 0 in the
diffusion rule. Other values of the Pi yield the same fractal dimension and
roughness.
4.4.2 Diffusion-limited aggregation
Growth phenomena are frequent in many fields of science and technology.
A generic aspect of several growth processes is a far-from-equilibrium
formation situation and a fractal structure. Dendritic solidification in
an undercooled medium, electrodeposition of ions on an electrode and
viscous fingering are well-known examples of fractal growth [1 19].
 4.4 App lications of the diffusion rule 167
Fractal dimension Roughness

the curves give N1 "' u 0.41 and u "' t0·27 .

Fig. 4.7. Fractal dimension of the diffusion front and roughness. The slopes of

Aggregation constitutes an important growth mechanism : like particles

stick to each other as they meet and, as a result, form a complicated
pattern with a branching structure.
In many cases, growth is governed by a spatial quantity such as an
electric field, a local temperature, or a particle density field. Quite often,
the behavior of these quantities can be approximated by solving the
Laplace equation with moving boundary conditions.
A prototype model of aggregation is the so-called DLA model (diffusion­
limited aggregation), introduced by Witten and Sander [120] in the early
1980s to simulate the following process : a negative electrode is placed
in an electrolyte containing positive metallic ions. The ions diffuse in
the electrolyte and, when they hit the negative electrode, stick to it. In
addition, the ions can also stick to the already deposited aggregate and, as
time goes on, a complex rigid tree-like structure emerges. Screening plays
an important role in the growth process. As the aggregate gets larger,
a new ion is more likely trapped around the perimeter than deep inside.
Since its introduction, the DLA model has been investigated in great
detail. However, diffusion-limited aggregation is a far-from-equilibrium
process which is not described theoretically by first principles only. Spatial
fluctuations that are typical of the DLA growth are difficult to take into
account and a numerical approach is necessary to complete the analysis.
The DLA process can be readily modeled by our diffusion cellular
automata, provided that an appropriate rule is added to take into account
the particle-particle aggregation. The first step is to introduce rest particles
to represent the particles of the aggregate. Therefore, in a two-dimensional
system, a lattice site can be occupied by up to four diffusing particles,
or by one "solid" particle. Figure 4.8 shows a DLA cluster grown by
the cellular automata dynamics. The diffusion constant of the particles is
D. = 1/4, obtained with Po = p = P2 = 1/4.
168 4 Diffusion phenomena

Fig. 4.8. Two-dimensional cellular automata DLA cluster (black), obtained with
Ps = 1, an aggregation threshold of one particle and a density of diffusing particle
of 0.06 per lattice direction. The gray dots represent the diffusing particles not
yet aggregated.

At the beginning of the simulation, one or more rest particles are

introduced in the system to act as aggregation seeds. The rest of the
system is filled with particles in average concentration p.
When a diffusing particle becomes nearest neighbor to a rest particle,
it stops and sticks to it by transforming into a rest particle. Since several
particle can enter the same site, we may choose to aggregate all of them at
once (i.e. a rest particle is actually composed of several moving particles),
or to accept the aggregation only when a single particle is present.
In addition to this question, the sticking condition is important. If any
diffusing particle always sticks to the DLA cluster, the growth is very fast
and can be influenced by the underlying lattice anisotropy. It is therefore
more appropriate to stick with some probability Ps· Since up to four
particles may be simultaneously candidate for the aggregation, we can
also use this fact to modify the sticking condition. A simple way is to
require that the local density of particles be larger than some threshold
(say three particles) to yield aggregation. Similarly, the so-called technique
of noise reduction can also be implemented in the cellular automata rule :
the idea is to have a counter associated with each lattice site on which
aggregation is possible. The counter is initialized to some given threshold
 4.4 App lications of the diffusion rule 169

1 1 .4

2 L......J..---l.-..l..-..._J...J.-..l-...l...-J 2 '---'----'--1---'---"'

0 4.2
ln(box-size) 0 ln(box-size) 4.2

Fig. 4.9. Box-counting estimate of the fractal dimension of (a) a two-dimensional

and (b) a three-dimensional cellular automata DLA, using the box-counting
method. The slope of the lines give : (a) df = 1 .78 and (b) dt = 2.3.

(for instance four) and each time a diffusing particle reaches the site, the
counter is decremented by one. When it reaches 0, aggregation takes place.
Our approach has also some other differences compared with the origi­
nal Witten and Sanders model. All particles move simultaneously and can
stick to different parts of the cluster and we do not launch them, one after
the other, from a region far away from the DLA cluster. For this reason,
we may expect some quantitative variation of the fractal dimension for
the clusters generated with our method. Figure 4.9 shows the result of a
box-counting algorithm in order to estimate the fractal dimension dt of
the DLA cluster obtained from our cellular automata rule. The principle
of this technique is to tile the cluster with a regular array of square boxes
of size t x t. One then counts the number Nb of these boxes which contain
a particle of the cluster. For a two-dimensional object, Nb is proportional
to the number of boxes, namely t-2 • For a fractal object, one expects that
Nb "' t-df
and, from figure 4.9 (a), one obtains
dt 1.78 (2-D) �

Similarly, a three-dimensional cellular automata DLA cluster can be

generated by using the diffusion rule presented in section 4.2.3. In fig­
ure 4.10, such a cluster is shown, using a ray tracing rendering technique.
Although it looks less dendritic than its two-dimensional counterpart, its
fractal dimension is less than 3. The same box-counting method as before
170 4 Diffusion phenomena

Fig. 4. 10. Three-dimensional cellular automata DLA cluster, on a 128 3 lattice,

comprising around 80 000 particles (which is still not enough to obtain an accurate
estimate of the fractal dimension).

(now with cubic boxes of size t x t x t) yields

dt 2.3 �
(3-D)
as shown in figure 4.9 (b).
These fractal dimensions (2-D and 3-D) correspond to measurements
on a single cluster (that of figures 4.8 and 4.10, respectively). Due to
this limited statistic, we can conclude to a reasonable agreement with the
more classically established values [121,1 19] d1 = 1.70 (2-D) and dt = 2.53
(3-D).*
The cellular automata approach is also well suited to study dynam­
ical properties such as the DLA growth rate. The standard numerical
experiment is to distribute uniformly the initial diffusing particles on · the
lattice with a single aggregation seed in the middle. As time t goes on,
more and more particles solidify and the cluster mass M(t) increases. Our
simulations indicate (see figure 4.1 1) that this process has an intermediate
regime governed by a power law
M(t) ="-' trx.

• It is interesting to note that there is a mean-field prediction dt

= (d2 + 1)/(d+ 1) which corresponds
rather well to the numerical measurements, where d is the space dimensionality.
 4.4 Applications of the diffusion rule 171

2.3 L..-.J...8-.l....-.l..-...L-..J.._...J-...J...J.-__j__J
3 9.8
ln(time)

Fig. 4. 1 1 . Formation rate o f cellular automata DLA clusters i n two and three
dimensions. The lattice has periodic boundary conditions.

where
oc � 2
in both two and three dimensions. Although these results are not sufficient
to conclude definitely that the 2-D and 3-D exponents are the same, an
explanation would be that in 3-D there is more surface to stick to than
in 2-D, but also more space to explore before diffusing particles can
aggregate. These two effects may just compensate.
A phenomena very similar to DLA is the so-called diffusion-limited
deposition [119] which corresponds to a deposition-aggregation process
on a surface. Our cellular automata model can be applied to model such
a process by having a nucleation surface instead of a single seed.
In the next section we consider a different deposition situation, in which
particles do not stick to each other but need an empty slot on the surface
to become adsorbed.
4.4.3 Diffu sion-limited surface adsorption
As a last example of the application of our cellular automata diffusion
rule, we show here how it can be used to model irreversible particle
adsorption on a surface. This problem is common in many areas of
science. For instance, adsorption of proteins and other particles at the
solid-liquid interface is a process of fundamental importance in nature.
The typical ingredients of irreversible adsorption are (i) the transport of
particles from the bulk to the surface and (ii) their subsequent adsorption
172 4 Diffusion phenomena

onto the surface. In this process, the interaction between an already

adsorbed particle and a newly arriving one plays a crucial role.
For instance, two particles cannot overlap on the adsorption substrate.
When one is already present, the second one can only be adsorbed if its
center lands sufficiently far away from the first one. More generally, there
is an excluded surface around each deposited particle, in which no more
adsorption is possible. As the deposition process goes on, the surface
coverage increases. But, due to this excluded surface, some "holes" on
the surface will remain empty for ever because they are too small to
accommodate a new particle. As a result, the coverage fraction O(t) will
<
tend to a maximum jamming limit 800 1 as the time t tends to oo.
The excluded surface depends on the type of inter-particle interaction
and their shape. Clearly, in the case of hard core interactions, one expects
a different behavior when disks or needles are adsorbed on a surface.
The kinetics of this process (i.e. the time dependence of 8) and the
limit value 800 are, of course, quantities of particular importance when
describing irreversible adsorption.
Another relevant aspect of surface adsorption is the transport process
in the bulk. Particles which are brought in contact with the substrate
have their own dynamics. At some point of their motion, they hit the
adsorption surface and if they pick the right spot, are adsorbed. As a
result, the surface coverage increases by a quantity s corresponding to the
area of the particle. Otherwise, if the particle hits an already excluded
part of the surface, it is rejected and continues its motion according to
the bulk dynamics.
The most studied process which mimics these interactions, is the random
sequential adsorption (RSA) model (see [122] for a review). The RSA
model is characterized by sequential (i.e one after the other) deposition
attempts of particles of unit area s at a constant rate c per unit time
and area. Adsorption occurs only under the condition that there is no
resulting overlap. After adsorption or rejection, the next attempt occurs at
a location which is completely uncorrelated with respect to the previous
one.
Even though RSA goes deeper into the description of adsorption pro­
cesses than previous models such as the Langmuir isotherm [123], it
cannot be entirely correct since coupling between deposition and trans­
port is neglected.
A very natural transport mechanism in the bulk is diffusion. Several
theoretical [124,125] and numerical[126,127] works have focused on this
effect on the kinetics of the deposition and the structure of the jammed
configuration and have established departures from the RSA predictions.
In this section we consider the problem of diffusion-driven surface
adsorption, in the framework of a cellular automata dynamics. That is,
 4.4 Applications of the diffusion rule 173

Fig. 4. 12. Excluded surface on the lattice, due to the adsorption of a particle.
We can imagine that the deposited particle is a diamond shaped object (dark
gray) which overlaps the surrounding light gray cells, making them unavailable
for further adsorption. In the diffusion plan, this picture is not the same, of
course, and the particles are point-like.

we are going to assume that adsorption takes place on a two-dimensional

lattice, which is of course different from the original case where a particle
can deposit anywhere in a continuous space. However, if the substrate has
some crystalline structure, this may correspond to a legitimate assumption.
In addition, the same generic features as in the continuous case are present,
but in a conceptually simpler form which may help developing an intuition.
The cellular automata we want to consider is the following: particles
diffuse on the top a two-dimensional substrate, according to rule 4.4. These
particles can be adsorbed on an empty substrate site. Adsorbed particles
prevent further deposition on the entire von Neumann neighborhood
surrounding them. Thus, the excluded surface amounts to five lattice sites,
for each deposited particle. This is illustrated in figure 4.12.
More precisely, at each time step, adsorption obeys the following rules.
• When a site in the diffusion plan contains exactly one particle, this
particle is candidate for adsorption.
• Adsorption takes place if the corresponding site on the substrate is
empty and does not correspond to an excluded surface.
• Two candidate particles which are nearest neighbors are rejected.
This rule is necessary to avoid two adjacent empty sites adsorbing
simultaneously two particles.
• Particles that can be adsorbed are placed on the substrate and
removed from the diffusion plan. The other ones diffuse according
to the diffusion rule.
If the initial particle density in the diffusion plan is large enough, there are
only a few sites containing exactly one particle. Only these sites will yield
174 4 Diffusion phenomena

(a) (b) (c)

Fig. 4. 1 3 . Adsorption surface at different time steps. Black cells represent

adsorbed particles and the surrounding excluded surface is shown in gray. White
regions show the available space. Figure (a) and (b) are the upper left part of
the system, four times magnified. They show the adsorption surface after one
and five iterations, respectively. Figure (c) shows the entire surface, a 256 x 256
lattice, at the final stage. Light gray tortuous "paths" are the excluded surface
which separate close-packed regions (dark gray) having a checkerboard structure

adsorption at the first iteration. Thus, it is not necessary to introduce an

extra adsorption probability to avoid all particles becoming marked for
adsorption at the first iteration.
Figure 4.13 shows the evolution of the adsorption surface.
Due to the shape of the excluded surface, the highest coverage frac­
tion that can be theoretically achieved is Boo = 1/2. This close-packed
configuration would be attained if the particles would be smart enough
to organize themselves as the black squares of a checkerboard. But, dif­
fusion has certainly not the global knowledge required to achieve this
result. Random deposition is not better with this respect. One expects, on
the other hand, to observe local checkerboard structures with defects, as
shown in figure 4.13.
The global structure which emerges from the local adsorption process
has a well defined jamming coverage, in the limit of an infinite system.
From our cellular automata model, we have estimated Boo = 0.366. An
interesting observation is that this jamming limit depends on the transport
mechanism of the particle. For instance, for random sequential adsorption
with nearest-neighbor exclusion on a square lattice, theoretical analysis
yields the estimate Boo = 0.364 [122]. We found this result compatible with
a parallel random adsorption, that is a cellular automata in which particle
positions in the diffusion plan are decorrelated from one time step to the
other. Note that these values correspond to a two-dimensional substrate
and are not the same in one or three dimensions.
 4.5 Problems 175
(a) (b)

0.6 1.....-l--l--.J...._.J.._.l..<lilmiiO
1 1 66

Fig. 4. 14. (a) Kinetics of the cellular automata adsorption model in units of
iterations. (b) Available space as a function of the surface coverage.

Our analysis also shows that the jamming limit depends slightly on the
diffusion constant D and on the specific value of the Pi chosen for the
diffusion rule. This may be an indication that, for this non-Markovian
process, it is not sufficient to give D to uniquely specify the macroscopic
behavior.
Finally, we show in figure 4. 14 the kinetics of the diffusion-limited
adsorption process and the relation between the available space (the white
cells in figure 4. 13) and the coverage (the black cells). From the first plot,
we can conclude that the surface coverage obeys an exponential law
eCfJ - 8( t) exp( - tI Tc )
"'

where Tc is some characteristic time. The second plot describes how the
available space 4> decreases with the coverage fraction e. Clearly, for
e � �
0, we have ¢(8) - 5 8 since each adsorption reduces the number of
deposition sites by five cells. On the other hand, near the jamming limit
� �
8 800, only isolated empty sites remain and ¢(8) -8. In between, 4> is
a non-trivial function of e .

4.5 Problems

4. 1 . S i m u l ate a CA syste m of particles w h i c h bou nce i nto f i xed obtac l es

that are i n itial l y d istr i buted rando m ly across the l atti ce with p robab i l ity
q (Eh renfest mod e l ) . The obstac l es act as m i rrors and defl ect the
ente r i n g part i c l es by n 2 + / -/
,n 2 or n In the a bsence of o bstac l es
.
176 4 Diffusion phenomena
the part i c l es fo l l ow a stra i g h t l i n e m oti o n . How does th i s ru l e co m pare
with the d iffus i o n ru l e? Measu re the re l ation betwee n the mean square
d is p l acement and ti m e (co n s i d e r a s i m u l ation where al l particles start
movi n g off a ve rtical l i n e a n d meas u re the pos itions x they h ave
reached afte r t i te rati ons) . Does it obey the reg u l ar d iffusion l aw
(x2 (t)) "' t o r does it correspond to an anomalous random wal k?
4.2. P ropose a CA r u l e of d iffu s i o n where o n l y o n e particle per s i te is
perm itted ( i nstead of the usual fo u r partic l es) . Each particle is n o rth­
fac i n g , east-fac i ng , south-fac i n g o r west-fac i ng , random ly at eac h t i m e
step and l attice s ite. It c a n move t o t h e s i te it is po i nt i n g t o o n l y if no
double occu pation w i l l occ u r . If such a conf l i ct can happe n , the parti c l e
re m a i ns i n pos i t i o n . What is t h e s i ze o f the n e i g h borhood req u i red for
this ru l e? Is the d iffus i o n co nstant i ndependent of the dens ity? In the
l ow density l i m it, what i s the d iffu s i o n constant? What is, approx i m ately,
the critical dens ity b e l ow wh ich th i s l i m it i s val i d . Measu re the d i ffus i o n
constant us i ng a s i m u l ation acco rd i n g t o t h e m ethod o f t h e p revious
p rob l e m .
4 . 3 . Show h o w t h e M a rgo l us n e i g h borhood c a n be used t o exp ress
the d iffus i o n ru l e with o n l y one particle per s i te , with a m uch s m a l l e r
n e i g h borhood t h a n i n t h e p revious p rob l e m .
4.4. U s i n g t h e s a m e approx i m ations a s i n prob l e m 3 . 5 , derive the
d iffu s i o n eq u ati on start i n g from the d iffus i o n m ic rodyna m i cs p rese nted
at the beg i n n i ng of t h i s ch apte r. Com pare the d iffus i o n consta nt you
obta i n with the correct o n e .
4.5. I nstead o f a d iffu s i o n ru l e u s i n g random rotations, co n s i d e r a ru l e
based o n random sym metries. Derive t h e d i ffus i o n coeff i c i ent.
4.6. P ropose a n o n l i near d i ffus i o n ru l e such that the rotati o n probab i l ­
ities depend o n the n u m be r of particles ente r i n g a s ite.
4.7. C h e motaxi s d esc ri bes the capac ity of some o rgan isms to adj ust
thei r movement i n re l ati o n to the p resence of chem ical age nts . Con­
s i d e r a pop u l ation of d iffus i n g ants. Each of them adds one u n it of
p h e romone to the l attice s i te it visits . When a s i te is not visited at a
g i ven t i m e step, the l evel of p h e romone is red uced by o n e u n it, u nti l it
reaches zero. When an ant is alone on a s ite, it j u m ps to the nearest
(von N e u m a n n ) n e i g h bo r wh ich conta i n s the h i g h e r l evel of pheromone
(except the s i te it j ust l eft at the p revious i te ration) . If m o re than one
ant are p resent ( u p to fou r are poss i b l e) , they obey the usual d iffu s i o n
r u l e . S i m u l ate such a system and study its beh avi or. F i rst, con s i d e r a
homogeneous s i tu ati o n ; then ass u m e there a re two sou rces of ants i n
the system . See i f the a nts can f i n d a path from one sou rce to the other.
 4.5 Problems 177
4.8. Devise a CA model of d riven d iffus i o n to s i m u l ate fo r i nstance
a s ituation where an el ectric f i e l d b i as the d iffu s i o n towards a g iven
d i rect i o n . As in the previous pro b l e m , s u p pose that the f i e l d acts with
p robab i l i ty E only o n those s i tes which conta i n no more than one par­
t i c l e . When seve ral particles e nte r the same s ite , the reg u l ar d iffusion
ru l e is app l i e d .
4.9. P ropose a LBG K equation fo r t h e d iffus i o n r u l e with rest particles.
What i s the exp ress i o n to be chosen fo r the local eq u i l i br i u m d istri bu­
tion f (0) ? How is the d iffusion constant re l ated to the rel axation t i m e?
4. 1 0 . I n the d iffu s i o n model of the previous prob l e m , how can we make
the d iffus i o n co nstant depend on the spatial coord i n ates?
4. 1 1 . S i m u l ate a DLA p rocess with a probab i l ity of re-d isso l uti o n of the
aggregated part i c l es . How is the fracta l d i m e n s i o n affected?
4. 1 2. One cons i d e rs the d iffusion of particles o n a o n e-d i me n s i on al l at­
tice. O n e ass u m es that there are traps (act i n g as s i n ks of particles) d is­
tri b uted on the l attice with a concentration c. Two s i tu ati ons h ave to be
d i sti n g u ished: the q u e nched case where the traps are at f i xed pos itions
d u ri n g the d iffus i o n p rocess and the an nealed case where the traps
move o n a ti me scal e m uch faste r than the d iffus i ng spec i es . Assu m i n g
that there is a sou rce of parti cles at the ori g i n x = and a s i n k at x = oo ,
0
f i n d analytica l l y the stationary d iffusion p rof i l e P (x) f o r both cases.
4. 1 3 . One co n s i d e rs the s i tuation descri bed in the p revious exercise,
but now o n a two-d i m ensional l attice. Devise an al gorith m ( us i n g one of
the standard d iffu s i o n r u l e d iscussed i n t h i s ch apter) w h i c h s i m u l ates
the an nealed case. I nvestigate the properties of P (x) as a fu nction of
the concentrat i o n of traps c .
4. 1 4. Find a CA model fo r the random depos i t i o n of d i m e rs on a 2D
l attice. A d i m e r i s m ade of two adj ace nt ce l l s and can be o r i ented
horizonta l l y or vertical ly. It can be adsorbed on the s u bstrate o n l y
if two nearest-n e i g h bor s i tes a r e em pty ( h i nt: s e e section 6.3 fo r the
concept of m atc h i n g nearest n e i g h bor) . What is the ave rage j am m i ng
cove rage of the l attice? What is the asym ptotic t i m e evo l ution (power
l aw) of t h i s p rocess?
 5
Reaction-diffusion processes

5.1 Introduction

Systems in which different chemical species diffuse (for instance in a

solvent or a gel) and then react together are quite common in science.
Lattice gases provide a very natural framework to model such phenomena.
In the previous chapter, we discussed how diffusion can be implemented as
a synchronous random walk of many particles, using a velocity shuffling
technique.
The model can be extended so that several different "chemical" species
coexist simultaneously on the same lattice. It just requires more bits
of information to store the extra automaton state. Then, it is easy to
supplement the diffusion rule with the annihilation or creation of particles
of a different kind, depending on the species present at each lattice site
and a given reaction rule.
In agreement with the cellular automata approach, chemical reactions
are treated in an abstract way, as a particle transformation phenomena
rather than a real chemical interaction. Only the processes relevant at the
macroscopic level are taken into account.
Systems in which reactive particles are brought into contact by a dif­
fusion process and transform, often give rise to very complex behaviors.
Pattern formation [128-130], is a typical example of such a behavior in
reaction-diffusion processes.
In addition to a clear academic interest, reaction-diffusion phenomena
are also quite important in technical sciences and still constitute numerical
challenges. As an example, we may mention the famous problem of
carbonation in concrete [131,132].
Reaction processes are most of the time nonlinear and characterized
by a threshold dynamics. Therefore, they may be very difficult to analyze
theoretically and even numerically, with standard techniques, due to the
178
 5.2 A model for excitable media 179
important role that fluctuations may play.
For instance, even simple systems behave in an unexpected way: some
simple reaction-diffusion processes exhibit an anomalous kinetics : they
depart from the behavior predicted by the classical approach based
on differential equations for the densities. The reason is that they are
fluctuations-driven and that correlations cannot be neglected. In other
words, one has to deal with a full N-body problem and the Boltzmann
factorization assumption is not valid. For this kind of problem, a lattice
gas automata approach turns out to be a very successful approach.
As already mentioned, lattice gas automata are a particular class of
cellular automata. They are characterized by a two-phase dynamics : first,
a completely local interaction on each lattice point, and then particle
transport, or propagation, to nearest-neighbor sites. This is a way to
partition the space, thus preventing the problem of having a particle
simultaneously involved in several different interactions. Moreover, this
approach yields a quite intuitive interpretation of the system as a kind
of a fully discrete molecular dynamics and allows direct connection with
statistical mechanics.
However, some reactive phenomena can also be described by simple
cellular automata models which are not LGA. In the next chapter, we
shall see an example of a surface reaction on a catalytic substrate. Here,
we shall start our discussion with a model for an excitable media, in which
chemical waves are observed. Then we shall concentrate on the lattice gas
approach which is well suited to represent reaction-diffusion processes.
The microdynamics will be given as well as its link to macroscopic rate
equation. Various applications will be presented in this framework as well
as in the framework of a lattice Boltzmann model and a multiparticle
approach. The reader can find other examples and applications in [50].
A natural extension of what will be discussed in this chapter is to give
the particles a hydrodynamical dynamics instead of a diffusive one. Thus,
models for studying combustion phenomena can be devised.
5.2 A model for excitable . media

In this section we present a simple cellular automata model for chemical

waves in an excitable media. Strictly speaking, this is not an LGA since
there is no explicit propagation phase. Rather, the rule is defined on
the Moore neighborhood and the transport of information takes place
because the neighborhood has some spatial extension.
An excitable medium is basically characterized by three states [50] : the
resting state, the excited state and the refractory state. The resting state
is a stable state of the system. But a resting state can respond to a local
180 5 Reaction-diffusion processes

t=5 t=llO

t=l l5 t=120

Fig. 5 . 1 . Excitable medium : evolution of a stable initial configuration with

10% of excited states ¢ = 1 , for n = 10 and k = 3. The color black indicates
resting states. After a transient phase, the system sets up in a state where pairs
of counter-rotating spiral waves propagate. When the two extremities come into
contact, a new, similar pattern is produced.

perturbation and become excited. Then, the excited state evolves to a

refractory state where it no longer influences its neighbors and, finally,
returns to the resting state.
A generic behavior of excitable media is to produce chemical waves
of various geometries, traveling in the system [133,134]. Ring and spi­
ral waves are a typical pattern of excitations. Many chemical systems
 5.3 Lattice gas microdynamics 181
exhibits an excitable behavior. The Selkov model [135] and the Belousov­
Zhabotinsky reaction are examples. Chemical waves play an important
role in many biological processes (nervous systems, muscles) since they
can mediate the transport of information from one place to another.
The Greenberg-Hasting model is an example of a cellular automata
model of an excitable media. This rule, and its generalization, have been
extensively studied [136,137]. It has some similarity with the "tube-worms"
rule proposed by Toffoli and Margolus [26] (see problem 5 .1).
The implementation we propose here for the Greenberg-Hasting model
is the following : the state <f>(r, t) of site r at time t takes its value in
the set {0, 1, 2, ... , n - 1 } . The state </> = 0 is the resting state. The states
</> = 1, ... , n/2 (n is assumed to be even) correspond to excited states. The
rest, </> = n/2 + 1, ... , n - 1 are the refractory states. The cellular automata
evolution rule is the following:
1. If <f>(r, t) is excited or refractory, then <f>(r, t + 1) = <f>(r, t) + 1 mod n.
2. If <f>(r, t) = 0 (resting state) it remains so, unless there are at least
k excited sites in the Moore neighborhood of site r. In this case
<f>(r, t) = 1.
The n states play the role of a clock : an excited state evolves through the
sequence of all possible states until it returns to 0, which corresponds to
a stable situation.
The behavior of this rule is quite sensitive to the value of n and the
excitation threshold k. Figures 5.1 and 5.2 show the evolution of this au­
tomaton for two different sets of parameters n and k. Both simulation are
started with a uniform configuration of resting state, perturbed by some
excited site randomly distributed over the system. If the concentration
of perturbation is low enough, excitation dies out rapidly and the system
returns to the rest state. Increasing the number of perturbed states leads
to the formation of traveling waves and self-sustained oscillations may
appear in the form of ring or spiral waves.
5.3 Lattice gas microdynamics

After the cellular automata model of an excitable media, we now return

to the main purpose of this chapter, namely the discussion of a lattice gas
automaton for describing a reaction-diffusion process.
We would like to model the following irreversible reaction
K
A+B � c (5.1)
where A, B and C are different chemical species, all diffusing in the same
solvent, and K is the reaction constant. To account for this reaction,
1 82 5 Reaction-diffusion processes

t=5 t=20 t=250

Fig. 5.2. Excitable medium : evolution of a configuration with 5% of excited

states ¢ = 1 , and 9 5 % of resting states (black), for n = 8 and k = 3.

one can consider the following mechanism : at the "microscopic" level of

the discrete lattice dynamics, all the three species are first governed by a
diffusion rule, as described in the previous chapter. When an A and a B
particle enter the same site at the same time, they disappear and form a
C particle.
Of course, there are several ways to select the events that will produce a
C when more than one A or one B are simultaneously present at a given
site. Also, when Cs already exist at this site, the exclusion principle may
prevent the formation of new ones. A simple choice is to have A and B
react only when they perform a head-on collision and when no Cs are
present in the perpendicular directions. Other rules can be considered if
we want to enhance the reaction (make it more likely) or to deal with
more complex situations (2A + B C, for instance).
--+

Finally, a parameter k can be introduced to tune the reaction rate K by

controlling the probability of a reaction taking place.
In order to write down the microdynamic equation of this process,
E
we shall denote by ai("r, t), bi("r, t) and ci(r, t) {0, 1 } the presence or the
absence of a particle of type A, B or C, entering site r at time t pointing
in lattice direction i.
We will assume that the reaction process first takes place. Then, the
left-over particles, or the newly created ones, are randomly deflected
according to the diffusion rule. Thus, using equation 4.4, we can write
the reaction-diffusion microdynamics as (d is the dimensionality of the
Cartesian lattice)
2d - 1
ai ( r + A ej, t + r ) = L Jlt (r, t) [ai+tO\ t) + Rf+t ( a, b, c)] (5.2)
t=O
and similarly for the two other species B and C.
 5.3 Lattice gas microdynamics 183

Fig. 5.3. Automata implementation of the A + B � C reaction process.

As before, the J.lt(r, t) are independent random Boolean variables pro­

ducing the direction shuffling. The lattice spacing A. and time steps are T
introduced as usual and the lattice directions ei are defined in figure 4.2
(note that we have changed the notation in order to avoid confusion with
the occupation number c i of species C).
The quantity Rj (a, b, c) is the reaction term : it describes the creation or
the annihilation of an A particle in the direction j, due to the presence of
the other species. In the case of an annihilation process, the reaction term
takes the value Rj = - 1 so that a1 - Rj = 0. On the other hand, when
a creation process takes place, a1 = 0 and Rj = 1. When no interaction
occurs, Rj = 0.
For instance, in the case of the reaction 5. 1 (illustrated in figure 5.3),
the reaction terms could be written as
- Kai bi+2 [v (1 - Ci+l) + (1 - V )(1 - Ci_l )]
Ri+2
K( 1 - ci) [vai- t b i+l + ( 1 - v ) a i+l b i- tl (5.3)
Ri and Rf are annihilation operators, whereas Rf corresponds to particle
creation. One can easily check that, for each A (or B) particle which
disappears, a C particle is created. That is,
2d 2d 2d
L Ri = L Rf = - L Rf
i=l i=l i=l
The quantities v (r, t) and K(r, t) in equations 5.3 are independent random
bits, introduced in order to select among the various possible events :
v (r, t) is 1 with probability 1/2 and decides in which direction the reaction
product C is created. When v = 1, the new C particle forms a +90° angle
with respect to the old A particle. This angle is - 90° when v = 0.
The occurrence of the reaction is subject to the value of the Boolean
variable K. With probability k, K = 1. Changing the value of k is a
way to adjust the reaction constant K. We shall see that k and K are
proportional.
184 5 Reaction-diffusion processes
The presence of the terms involving ci in the right-hand side of equa­
tions 5.3 may seem unphysical. Actually, these terms are introduced here
in order to satisfy the exclusion principle : a new C cannot be created in
direction i if c i is already equal to 1. With this formulation, the reaction
slows down as the number of C particles increases. At some point one
may reach saturation if no more room is available to create new particles.
In practice, however, this should not be too . much of a problem if
one works at low concentration. Also, quite often, the C species also
undergoes a transformation: the reaction can be reversible or C particles
can precipitate if the concentration reaches some threshold. Or, sometimes,
one is only interested in the production rate LJ Rj '2:,1 Rj of the species
=
C and one can forget about them once they are created. In this case, one
=
simply puts ci 0 in the first two equations of 5.3.
Clearly, the exclusion principle may introduce some renormalization of
the reaction rate. If for some reason, this is undesirable, multiparticle
models offer an alternative to the LGA approach. This will be discussed
in section 5.7.
Due to the simple microscopic interpretation, equation 5.3 is easily
generalized to other reaction processes. A common situation is when one
species is kept at a fixed concentration. This means that the system is
fed a chemical by an external mechanism. In this case, the corresponding
occupation numbers (for instance the h i s ) can be replaced be random
Boolean variables which are 1 with a probability given by the selected
concentration of the species.
5.3.1 From microdynamics to rate equations
Several times, throughout this book, we have established a link between the
discrete cellular automata dynamics and the corresponding macroscopic
level of description. Here, again, we shall perform this calculation for the
case of the A + B � C process. Generalization to other reaction schemes
is straightforward. As in the previous chapters, we use the Boltzmann
molecular chaos assumption, in which correlations are neglected. Within
this approximation, we show that the microdynamics of reaction-diffusion
processes yields the usual rate equation, which, in our case, reads
( 5 . 4)
To derive the macroscopic behavior of our automata rule, we first
average equation 5.2
2d 2d
Ai(r + Aei, t + r) - Ai(r, t) L Qi}Aj(r, t) + LU5ij + Qij ) Rj(A, B, C)
= (5.5)
j=l j=l
 5.3 Lattice gas microdynamics 185
where A; = (a;) is the average value of the occupation numbers a;. The
matrix Q is the diffusion matrix defined in equation 4. 14 for a two­
dimensional system. Similar equations hold for B; and C;.
Using the Boltzmann hypothesis, the average value of the reaction term
is written as
( Rf (a, b, c, K, v ) ) � Rf (A, B , C, ( K ) , ( v ) ) (5.6)
that is the same function, but taken for the average values A, B and
C instead of the occupation numbers a, b and c, and for the average
values of the random Boolean fields. We shall see, in section 5.4 that
this factorization may be wrong for simple annihilation reaction-diffusion.
But, for the moment, we will assume that it is valid.
The second step is to replace the finite difference in the left-hand side
of 5.5 by its Taylor expansion

[ � af + Jc((\ · a,.) + � (c; · a-;. )2 + dat (c; · 8,)] A;

A;(r + Jce;, t + r) - A;(r, t) =
rat +
(5.7)
and similarly for the other species B and C. As usual, we will consider a
Chapman-Enskog-like expansion and look for a solution of the following
form
(5.8)
Since particle motion is governed by the diffusion process, we will use the
result derived in section 3.1.1, namely that, when taking the continuous
limit, the time and length scale are of the following order of magnitude
and (5.9)
In reactive systems, as opposed to hydrodynamics or pure diffusion, neither
momentum nor particle number are conserved in general. For instance, in
the annihilation process A + A � f/J, no conservation law holds.
On the other hand, the reaction term can be considered as a perturbation
to the diffusion process, which makes derivation of the macroscopic limit
rather simple. In equation 5.4, the reaction constant K is proportional to
the inverse of a time. This quantity defines at what speed the reaction
occurs. At the level of the automaton, this reaction rate is controlled by
the reaction probability k = ( K ) introduced in the previous section.
Now, when the continuous limit is taken, the automaton time step r
goes to zero. Thus, the number of reactions per second will increase
as r decreases, unless the reaction probability k also diminishes in the
right ratio. In other words, to obtain a finite reaction constant K in the
186 5 Reaction-diffusion processes

macroscopic limit, it is necessary to consider that k rx r .

Since is of the r
order c2 in our Chapman-Enskog expansion, the reaction term Rf is also
to be considered as an O(c2 ) contribution
(5. 10)
At the macroscopic level the physical quantities of interest are the particle
densities of each species. Following the usual method, we define the
density pA of the A species as
2d
PA = L A ) 0l
i=l
with the condition
2d
L A V') = 0 if t 2 1
i =l
Now we have to identify the different orders in c which appear in
equation 5.5, using the expressions 5.7, 5.8, 5.9 and 5.10. We obtain
( 5. 1 1)
j
.A. 1 (ei V)A )0l = L nijA)1l
· ( 5.12)
j
These equations are exactly similar to those derived in the case of pure
diffusion (see equations 4.1 1 and following). Thus we know that
A (l O) = PA

2d
and
_- .A. 1 1 eirx arx PA
A (i 1)
2d V
where V is the eigenvalue of the diffusion matrix n for the eigenvector
E rx = ( e 1 ,rx ; ez,rx ; ... ; e2d,x )
The equation for the density PA is now obtained by summing over i
equation 5.5, remembering that

Collecting all the terms up to O(c2 ), we see that the orders O(c0 ) and O(c)
 5.4 Anomalous kinetics 187
vanish and we are left with
c2 'r2 8 t � A �0) + c2 A t 2.:)ei
l l
· V )A � t ) + c 2 � �(ei V)2A�o)
t
l
·

j
Using the definition of 'r2 , A t , R�j and performing the summations yield
8 t PA = DV2PA + �1 '"' PA PB P ca
� Rj ( ld ' ld ' ld ) (5. 13)
1

where D is the same diffusion constant as would be obtained without the

chemical reactions (see equation 4.28).
It is interesting to note that the expression 5. 1 3 has been obtained
without knowing the explicit expression for the reaction terms R and
independently of the number of species. Actually, from this derivation, we
see that the reaction term enters in a very natural way in the macroscopic
limit: we just have to replace the occupation numbers by p j 2d, the random
Boolean fields by their average values and sum up this result for all lattice
directions.
For the case of the A + B C process in two dimensions, with the
�

reaction term given by 5.3, equation 5. 13 shows that the macroscopic

behavior is described by the rate equations
a tPA
4'r 4 (
DA V2 PA - !.__ 1 - P c PAPB )
a tPB DB V2 PB - 1- :"[ (
P
PAPB �)
8 tPC D c V2 p c + 'r 1 -:(
p
PAPB :) (5.14)
where, in principle a different diffusion constant can be chosen for each
species. We also observe that the reaction constant K is related to the
reaction probability k by
K = !_4'r
But, as explained previously, the exclusion principle introduces a correction
( 1 - P c/4) which remains small as long as C is kept in low concentration.

5.4 Anomalous kinetics

In the previous section, we derived the partial differential rate equation

governing the behavior of our cellular automata reagents with the assump­
tion that many-body correlations can be neglected. When the reaction
188 5 Reaction-diffusion processes

process is rare enough so that diffusion is efficient enough to mix the

particles thoroughly, this derivation makes sense. However, correlations
or intrinsic fluctuations often play a crucial role in reaction-diffusion
processes. Whereas the cellular automata approach naturally takes into
account this phenomenon, the rate equation may not always capture cor­
rectly the effect of these fluctuations, especially if the space dimensionality
is low.
In this section we present simple reaction-diffusion systems in which
fluctuations cannot be neglected or many-body correlations are non-zero.
This will be the annihilation processes A + B - 0 or A + A - 0 where 0
indicates that the reaction products of A and B do not play any role in
the system and can be forgotten.
5.4. 1 The homogeneous A + B - 0 process
The simple irreversible reaction-diffusion process A + B - 0 has received
a great deal of attention among chemists and physicists (see for instance
[138-143]).
Traditionally, this process is described in terms of rate equations for
the local concentrations a and b of the various species
o t a = DV2 a - R i3 t b = DV2 b - R (5.15)
where D is the diffusion constant and R the reaction term. Usually, R is
assumed to be proportional to the product of the a and b concentrations,
multiplied by K, a reaction constant :
R = Kab (5.16)
The approximation is often referred to as a mean-field approximation
since it assumes that the probability of a reaction taking place simply
factorizes into the product of the probabilities of the presence of an A
and a B particle.
Equations 5.15 with R = kab are easy to solve when the system is
homogeneous at time t = 0 (i.e a and b do not depend on space variables).
The Laplacian terms vanish and one is left with ordinary differential
equations. If the initial concentrations ao and bo are equal, a(t) will be
equal to b(t) at any time and one has to solve
da
= Ka2
dt -
whose solution is simply
ao
a(t) = (5.17)
1 + ao k t
 5.4 Anomalous kinetics 189
Note that if ao =F bo , say bo > ao , the solution is different and is found to
be
(bo ao) exp( -(bo - ao)kt)
a ( t ) = ao - ---, ---,-
-,----------------,- ( 5 .18 )
bo - ao exp( -(bo - ao)kt)
Equation 5.17 predicts a t- 1 behavior in the long time regime, whereas
equation 5.18 gives an exponential decay. Unfortunately, these results are
wrong if ao = bo . It is now widely recognized that when the number of A
and B particles are initially equal, the correct dynamics is (see [139], for
instance)
a(t) ,....., t-d/4 ( 5 .19 )
where d � 4 is the dimensionality of the system. Therefore, in dimen­
sion one, two and three, the predictions of the mean-field equations are
wrong and an anomalous kinetics is observed. This is due to fluctuations
and correlations which always exist in a many-particle system. In low
dimensions, diffusion is not a good enough mechanism to mix correctly
the particles.
This t-d/4 law can be understood as follows. Let us consider a time
interval r. Due to diffusion, a particle explores a typical distance l = Ji5T.
Suppose we divide our system in volumes of size l. If the size of the
system is L, there are (Ljl )d such volumes. Assuming that the average
initial density of particles is ao = bo , the average d number
d 12 N of A or B
=
particles in our smaller volumes is N = a l a (Dr) . However, this
quantity is subject to statistical fluctuations0 of the0 order JN. Therefore,
if the distribution of A and B particles is not correlated the number ofd/A
and B particles in our volumes typically differs from JN = .Jao(Dr ) 4 .
If, duringd the time interval r , all the A and B particles located in the same
volume l react together (which is reasonable since r is the time needed
to explore the volume), the total number of particles left after that time
will be the excess of particles of one of the two species. Thus
=
a( r ) JN ( �)d d
= L JiiO(Dr) -d/4 (5.20)
In addition to the time exponent, equations 5.20 and 5.17 differ by their
amplitudes : equation 5.20 depends on .JCiO and D, whereas 5.17 depends
on k but no longer on the density ao , in the long time regime.
This discussion suggests the following physical picture : when the species
A and B are randomly distributed in the solvent and the position of A
and B particles are uncorrelated, then local density fluctuations make
some region richer in A than in B particles. The reverse is true in some
other regions. If the reaction is fast enough compared to the diffusion, the
species which is more abundant will eliminate the other. A segregation
190 5 Reaction-diffusion processes

occurs, producing clusters of A surrounded by B clusters. The reaction is

then no longer a bulk reaction but a surface reaction since it is only at
the domain boundaries that 4annihilation takes place. This slows down the
dynamics
1
and yields the t-d/ decay law instead of the expected mean-field
c behavior.
We have insisted on the fact that the initial A and B particle distributions
should be noncorrelated in order for the t-d/4 to be true. When this does
not hold, another law is observed [142,143]. The system looks more like
the A + A 0 reaction, for which a t-d/2 decay law holds [139,144].
�

Thus, in one dimension, a non-mean-field behavior (i.e a deviation from

the predictions of equation 5.17) occurs.
At this point we may notice that there are two reasons to depart from a
mean-field behavior: (i) the initial condition or, more precisely, the nature
of the fluctuations in the initial distribution of the the A and B particles;
and (ii) the factorization of the joint probability that an A particle meets
a B particle into the product of the local density of A and B.
The fact that the initial condition plays such a crucial role makes the
A + B 0 system quite intricate. It is sometime difficult to differentiate
�

between the two effects. Actually, both initial condition and many-body
correlations may simultaneously affect the system. The slowest decay law
will be longer lived and mask the other behavior. A cd/4 will certainly
screen out a t-d/2 decay.
5.4.2 Cellular automata or lattice Boltzmann modeling
The A + B 0 reaction-diffusion process is therefore a good candidate for
�

cellular automata modeling where no mean-field assumptions are made

for the initial condition, or for the joint probability that an A and a B
annihilate. Using the reaction term given in relation 5.3, with ci always
zero provides a natural implementation of this process.
However, some care is needed to obtain reliable simulations. Earlier in
this book, we mentioned the checkerboard spurious invariant that plagues
many lattice gas automata. In our case, the checkerboard invariant means
that an A particle entering a "black" site of the lattice at time t will never
interact with a B particle entering a "white" site, and conversely. Thus
the total number of A particles left on the lattice as a function of time
is composed of two contributions, one from the black sublattice and the
other from the white sublattice. Since the two sublattices may not have
the same number of particles, the measurement of particle densities may
be erroneous.
The simplest solution is to consider an initial condition where all A and
B particles occupy only one sublattice, the other being left empty.
 5.4 Anomalous kinetics 191
The second technical difficulty is to build an initial condition with an
equal number of A and B particles. This is a crucial condition to observe a
power law instead of an exponential decay. The problem is to make both
A and B distributions statistically independent. One possibility to achieve
this result is to first take the B distribution equal to that of A. Then, the
B particles should be sent somewhere in the considered sublattice, so that
the original distribution is forgotten. The diffusion cellular automata rule
can be used for this purpose but for a number of iterations at least as
large as the square of the system size. On a general purpose computer,
faster ways can be devised : a random permutation of the lattice sites can
be performed to mix the position of the B particles [145].
Now, as already mentioned, we would like to identify more precisely
the role of the fluctuations in the initial configuration against the role
of the Boltzmann factorization assumption. For this purpose, a lattice
Boltzmann model (evolution rule 5.5) can be simulated and compared
with the cellular automata dynamics. The lattice Boltzmann approach
makes it possible (and natural) to prepare a random initial condition, in
which ai('r, t) and bi ( r, t) are either 1 or 0, in exactly the same way as in
the cellular automata case.
But the lattice Boltzmann method neglects many-body correlations,
and when both the cellular automata rule and the lattice Boltzmann
dynamics agree (and still depart from the mean-field predictions), we
may conclude that we are observing an effect of the fluctuations of the
initial configuration. Otherwise, we conclude that correlation cannot be
factorized as a product of the one-point densities.
5.4.3 Simulation results
The behavior of the CA model and the LBM simulations are illustrated
in figure 5.4. The number of A particle is exactly the same as the number
of B particles and there is no correlation between their respective initial
distributions. The power law 5.19 is well obeyed in both approaches.
On the other hand, when the initial distributions of A and B are
correlated, a different behavior is observed. Figure 5.5 shows the behavior
of the total number of A particles as a function of time, in a log-log
plot. The two curves correspond to the lattice Boltzmann and the cellular
automata dynamics, respectively. In both cases, the initial condition is
prepared identically: (i) the A particles are placed randomly on the lattice
and (ii) the B particles are obtained from the A particles by two diffusion
steps. In this way, a correlation between the positions of the A and B
particles is created.
Thus, the lattice Boltzmann dynamics predicts a long time regime which
is not the same as the one observed in the cellular automata model. The
192 5 Reaction-diffusion processes

(a) (b)
Fig. 5.4. Segregation in the A + B - 0 reaction-diffusion process, as obtained
from (a) the cellular automata model and (b) the lattice Boltzmann dynamics. The
situation corresponds to the evolution of a noncorrelated random, homogeneous
initial condition. In (a) the white dots represent lattice sites with a majority of A
particles and the gray ones a majority of B particles. The black dots correspond
to empty sites. In (b) the gray levels indicate the regions where A (or B) is
dominant : black when a(r, t)jb(r, t) > 1, dark gray when a(r, t)/b(r, t) > 6, light
gray when b(r, t)/ a(r, t) > 1, and white when b(r, t)/ a(r, t) > 6.

-1/2 time exponent found in the cellular automata is not so surprising

since the correlated A + B 0 reaction
---* presents many similarities with
the A + A 0 problem, for which t- 1 12 is the known decay law. It turns
---*

out that if one simulates the A + A 0 process along the same

�
1
lines, the
cellular automata dynamics indeed reproduces the expected t- behavior1 2
in one dimension, whereas the 1corresponding lattice Boltzmann simulation
gives the mean-field result t- , in spite of a randomly distributed initial
state.
Clearly, the factorization done in the lattice Boltzmann approach ne­
glects an important aspect of the problem, which is the non-mean-field
behavior of the dynamics itself, in one dimension. A cellular automata
model captures this behavior because it makes no assumption on the joint
probability of having an A particle meeting a B particle. We will return
:o this question in section 5.8, where it is explained how an exact solution
Jf the A + A 0 can be obtained analytically, in one dimension.
---*

At this point it is interesting to note that the A + B 0 reaction is even

---*

nore intricate than just a correlated or noncorrelated initial configuration

)f the As and Bs. Depending on the way the initial correlation is built,
iifferent time exponents may be observed in both approaches.
When the initial correlation is created by a translation of the one species
 5.5 Reaction front in the A + B � f/J process 193

"
[]

D
o D

slope�-0.9 3

3 L-�---L--�--��-��
0 12
In(time)

Fig. 5.5. Time behavior of the total number of A particles in the A + B ---+ 0
process, in one dimension. The two plots correspond to the lattice Boltzmann
(dots) and the cellular automata (squares) dynamics, both for a correlated initial
condition. The lattice Boltzmann takes into account the fluctuations of the initial
state but not the fact that the dynamics itself is not mean-field. The observed
time behavior is t'1, with rx � 1 /2 for the cellular automata and rx � 0.9 for the
lattice Boltzmann dynamics.

with respect to the other, the predictions found by Lindenberg et al. [142]
(namely a c(d+2l/4 long time regime and a cd/ 4 short time regime)
are verified in the lattice Boltzmann dynamics. The cellular automata
dynamics, on the other hand, does not agree well with these theoretical
predictions which do not take the fluctuations fully into account.
5.5 Reaction front in the A + B � f/J process

The problem we discussed in the previous section corresponds to a ho­

mogeneous initial situation where all the reagents are uniformly (up to
fluctuations) distributed in space. Another situation is obtained when the
chemical species A and B are initially separated (inhomogeneous systems).
For instance we may have the situation described in figure 5.6 : the A par­
ticles occupy the left part of the system while the B particles are on the
right. The A species try to diffuse into the B region, and conversely. A
reaction take place in which an A-B pair is annihilated (or transformed
into an inert substance).
The region where the reaction occurs is called the reaction front. In
figure 5.6, this corresponds to the white pixels. Reaction fronts are present
194 5 Reaction-diffusion processes

Fig. 5.6. Reaction front in the A + B � (/) reaction diffusion process (according
to cellular automata dynamics).

in many nonequilibrium physical or chemical processes and their study is

a first step to understanding various pattern formation phenomena.
On a macroscopic level, there are two important features to be described.
First, if the concentrations of A and B are not the same, the more abundant
species will penetrate the other (for simplicity, we assume here that both
diffusion coefficients are identical). As a results, the position of the front
will move and be a function of time. Second, as times goes on, the reaction
front becomes larger and larger because, due to the depletion of A and
B particles around the reaction zone, particles of one type have a greater
chance to enter deeper into the other region before reacting.
Therefore the question is : what are the front width W(t) and the front
location ro (t)?
Pioneering work on this subject has been done by Galfi and Racz [146].
Using the rate equations 5.4, they considered two species A and B diffusing
on a two-dimensional substrate and reacting to form an inert species C.
For an appropriate choice of the initial conditions (at time t 0, the =
system is uniformly filled with A for x :::;; 0, and uniformly filled with
>
B for x 0), the problem reduces to one dimension. Galfi and Racz
developed a scaling theory from which they found that the front position
obeys ro Jt (with a coefficient depending on the initial concentration
"'

difference Ao - Bo). Second, their scaling theory predicts that1 front width
W(t) follows, in the long time limit, the power law W ( t) t 16 ."'

Whereas the Jt law giving the front position holds in all generality,
the question of the width is more involved. It was shown [147], using
cellular automata models, that the upper critical dimension of the problem
is 2. That is, logarithmic corrections to the rate equation predictions
(i.e. the W t 1 16 behavior) are found for the space dimension d
"' =
=
2 and an important correction is measured in d 1. This reaction­
diffusion process has now been intensively studied due to its departure
from the mean-field description. The reason for this departure is of
course that the average reaction term (R) is not the product of the A
density times the B density. Instead, it is the joint probability of having
one A and one B particle simultaneously present at the same position.
 5.5 Reaction front in the A + B � 0 process 195
Unfortunately, this joint probability is not known in general. An equation
for it could be written down but it would involve other higher-order
correlation functions.
Thus the problem of finding the right behavior for W(t) in low­
dimensional systems is clearly a very difficult analytical task. Of course
the reader may argue that, since the upper critical dimension is d 2, =
everything will be fine in the 3d physical space and the mean-field result
is enough. However, the problem is very interesting from a theoretical
point of view and sheds some light on what may happen even in very
simple nonequilibrium systems. In addition, some experiment have been
realized in very thin tubes [148] but they are not thin enough to probe
the one-dimensional regime.
5.5.1 The scaling solution
The cellular automata approach is a very successful tool in the analysis of
reaction diffusion fronts. Moreover, a scaling theory can be considered in
order to interpret, in a theoretical framework, the simulation results.
The idea is to write macroscopic rate equations (with dependence only
on one spatial coordinate for symmetry reasons) as :
(5.2 1 )
but without specifying an explicit form for the reaction term R. Here, as
opposed to the homogeneous A + B 0 case, the spatial fluctuations of
�

the initial configuration does not play a crucial role since we assume that,
= = =
at time t 0, ai 1, bi 0 on the left part of the system and conversely
on the right part. Thus, we may argue that the departure from mean-field
behavior is due to the fact that
R =f. Rmr = ab
where Rmr denotes the standard mean-field approximation. In general, R
is the joint probability of having one particles of type A and one particles
of type B simultaneously present at the same point. Of course, since R
is not known, we cannot solve equation 5.21. However, in the spirit of
a scaling approach, we may say something about the asymptotic time
dependency of a(x, t) and b(x, t). Thus, we shall assume that
a(x, t) t-Ya(xt - rx )
b(x, t) t-Ya(xcrx )
R(x, t) cP R(xt-rx ) (5.22)
cx,
where fJ and y are the exponents we are looking for. From these
exponents, we can obtained the behavior of the width front W(t) defined
196 5 Reaction-diffusion processes

as
W 2 (t)2:�:-�Llx ;- (x) )2 R(x, t) (5.23)
Lx;:-L R(x, t)
=

Using the scaling form for R we have

W(t) "' {x
Similarly, we also obtain that the total number of particle annihilations is
related to ll and f3
1volume R "' cP+rx (5.24)
Now, the nice thing is that some general relations between these exponents
can be derived from two assumptions [146,149] based on the fact that A
and B react only in a localized region:
(i) The reaction zone increases more slowly than the characteristic length
of the diffusion process. This means that, in the long time limit, a and b
vary over a typical length governed by diffusion, i.e. grada "' gradb "' c 1 12
and thus
y + ll 1/2 = (5.25)
(ii) Due to the gradient concentration of A and B, a flux of particles
towards the reaction region is observed. Assuming that the reaction is
fed by these particle currents, the reaction-diffusion equations take on a
quasi-stationary form DV2 a KR and DV2 b KR. For the exponents,
= =

this amounts to the relation

y + 2 ll f3 =
( 5.26)
For consistency with assumptions (i) and (ii), we should also have
ll 1/2 < (5.27)
so that the width of the front grows more slowly than the depletion zone
produced by the diffusion. To ensure the quasi-stationarity of the solution,
we need furthermore that
DV2 a t -y-2rx
� 8 t a -t-y- l (ya + llXt-rxa')
= (5.28)
which is again consistent with the condition 5.27.
Relations 5.25, 5.26 and 5.29 are still quite general since no explicit
form of the reaction term R has been used. But, they are not enough to
find the value of ll, f3 and y. One of these exponent must be determined
numerically, using our cellular automata model.
Note, however, that we can already deduce that
f3 - ll,
1/2 = (5.29)
 5.5 Reaction front in the A + B ---+ 0 process 197
xJO-I
5_15 .-
---r--r-�---r---r-�--.-.---.""'�.---.-,--�,.--,---.--r-r--,
.. I=1049
0 1 =34408
0 1 =288669

3.43

L72

o m "'-""""-��-'---''---'-_L_-'---"'---'--'-_L___,---'-_,__,__:]1__"---cw�
-'-
-257 -L50 -0.42 0_65 L72 2.79
x/t o-292

Fig. 5. 7. Scaling profile of the reaction term, as obtained with a data collapse
from the cellular automata simulations.

Therefore, according to equation 5.24, the total annihilation rate decreases

with time as c 1 12 .
If we consider the mean-field approximation R ab, this introduces =
another relation among our exponents, namely
f3 = 2y (5.30)
and we obtain the following the values for the mean-field exponents
1 1
= Ymf = 3 (5.3 1)
(Xmf 2
From our one-dimensional cellular automata simulations, the values of
the non-mean-field exponents are found to be [1 50]
0.292 ± 0.005 (5.32)
f3 0.792 ± 0.006 (5.33)
y 0.216 ± 0.003 (5.34)
which are clearly different from their mean-field counterparts. On the other
hand, in two- and three-dimensional systems the mean-field exponents are
well obeyed (up to a logarithmic correction in 2D).
The shape of the reaction function R is not given by the scaling
analysis. However, it can be measured in the simulation. Figure 5. 7 shows
the observed profile. Superposition of the data obtained for different time
steps confirms our scaling hypotheses.
The analysis we presented about the A + B 0 reaction-diffusion pro­---+
cess, in both homogeneous and inhogeneous cases, can be generalized
198 5 Reaction-diffusion processes
along the same lines. Other situations, such as the rnA + nB 0 pro­
�

cess [143] can be considered or the problem of stationary front formatjon

in a system with a constant particle flux [151].
5.6 Liesegang patterns

In this section we shall study in detail a more complex system in which sev­
eral of the ingredients we have introduced so far are combined: reaction­
diffusion will be accompanied by solidification and growth phenomena.
This gives rise to nice and complex structures that can be naturally mod­
eled and analyzed in the framework of the cellular automata approach.
These structures are known as Liesegang patterns, from the german
chemist R.E. Liesegang who first discovered them at the end of the
nineteenth century [152] .
5.6. 1 What are Liesegang patterns
Liesegang patterns are produced by precipitation and aggregation in the
wake of a moving reaction front. Typically, they are observed in a test
tube containing a gel in which a chemical species B (for example AgN03)
reacts with another species A (for example HCl). At the beginning of
the experiment, B is uniformly distributed in the gel with concentration
b0. The other species A, with concentration a0 is allowed to diffuse into
the tube from its open extremity. Provided that the concentration a0 is
larger than bo , a reaction front propagates in the tube. As this A + B
reaction goes on, formation of consecutive bands of precipitate (AgCl in
our example) is observed in the tube, as shown in figure 5.8. Although
this figure is from a computer simulation, it is very close to the picture of
a real experiment.
The presence of bands is clearly related to the geometry of the system.
Other geometries lead to the formation of rings or spirals. The forma­
tion mechanism producing Liesegang patterns seems quite general. It is
assumed to be responsible for the rings present in agate rocks [153, 154]
to yield the formation of rings in biological systems such as a colony of
bacterium [1 52] or to play a role the the process of cement drying.
Although the Liesegang patterns have been known for almost one
hundred years, they are still subject to many investigations. The reason
is that they are not yet fully understood and can present unexpected
structures depending on the experimental situation (gravity, shape of the
container, etc.).
On the other hand, it has been observed that, for many different
substances, generic formation laws can be identified. For instance, after a
 5.6 Liesegang patterns 199

x=O direction of the moving front

Fig. 5.8. Example of the formation of Liesegang bands in a cellular automata

simulation. The white bands correspond to the precipitate which results from the
A + B reaction.

transient time, these bands appear at some positions xi and times ti and
have a width wi. It is first observed that the center position Xn of the nth
band is related to the time tn of its formation through the so-called time
law Xn Jtn.
"'"'

Second, the ratio Pn = Xn/Xn- 1 of the positions of two consecutive bands

approaches a constant value p for large enough n. This last property is
known as the Jablczynski law [153] or the spacing law. Finally, the width
law states that the width Wn of the the nth band is an increasing function
of n.
In section 5.5, we explained that, in the A + B � 0 reaction process,
the front moves according to a square root t law with a coefficient factor
which depends on the difference of the concentrations a0 and b0.
As Liesegang patterns are formed in the wake of a moving reaction front
produced by an A + B � C process, the time law appears to be a simple
consequence of the diffusive dynamics. On the other hand, spacing and
width laws cannot be derived with reaction-diffusion hypotheses alone.
Extra nucleation-aggregation mechanisms have to be introduced, which
makes any analytical derivation quite intricate [155-157].
From an abstract point of view, the most successful mechanism that can
be proposed to explain the formation of Liesegang patterns is certainly
the supersaturation assumption based on Ostwald's ideas [158]. This
mechanism can be understood using the formation scenario proposed by
Dee [159] : the two species A and B react to produce a new species C
which also diffuses in the gel (this C species represents a colloidal state
which is observed in several experiments). When the local concentration of
C reaches some threshold value, nucleation occurs : that is, spontaneously,
the C particles precipitate and become solid D particles at rest. This
process is described by the following equations
8ta Da V2 a - Rab
8t b Db V2 b - Rab
(5.35)
200 5 Reaction-diffusion processes

where, as usual, a, b, c, d stand for the concentration at time t and position

r of the A, B, C and D species, respectively. The term Rab expresses
the production of the C species due to the A + B reaction. Classically,
=
a mean-field approximation is used for this term and Rab Kab , where
K is the reaction constant. The quantity nc describes the depletion of
the C species resulting from nucleation and aggregation on existing D
clusters. An analytical expression for this quantity is rather complicated.
However, at the level of a cellular automata model, this depletion term
can be included quite naturally. We will return to this point later in the
discussion.
Now, within this framework, the supersaturation hypothesis can be
stated as follows : due to aggregation, the clusters of nucleated D particles
formed at the reaction front deplete their surroundings of the reaction
product C. As a result, the level of supersaturation drops dramatically and
the nucleation and solidification processes stop. To reach again suitable
conditions to form new D nuclei, the A - B reaction has to produce
sufficient new C particles. But, the reaction front moves and this happens
at some location further away. As a result, separate bands appear.
A similar scenario can be considered even if the intermediate C species
does not exist as an individual diffusing molecule. In this case, the A
and B species coexist in the gel until the solubility product ab reaches
a critical value above which nucleation occurs according to the reaction
A+B ---+AB(solid). As nucleation has started, depletion of A and B
occurs in the surrounding volume and the same mechanism as described
above takes place to produce distinct bands [128].
Not all Liesegang structures obey the spacing and width laws. The
situation of inverse banding [160] can be observed in some cases, as well
as some exotic behavior [161]. Although the fundamental mechanism
described so far is probably still relevant, other effects should be taken
into account, such as local properties of the gel, spatial variation of the
diffusion coefficients and so on.
An important aspect of the mechanism of Liesegang patterns formation
is the role of spontaneous fluctuations. In section 5.5 we mentioned that,
in a one-dimensional system, the reaction front has an anomalous, non­
mean-field behavior. Therefore, since the motion of the front is responsible
for the creation of the bands of precipitate, it is likely that, in a true 1D
system, Liesegang patterns no longer obey the spacing law. But, in two­
and three-dimensional systems, we may safely ignore this problem. On
the other hand, the precipitation process and the aggregation (such as
a DLA process) are clearly dependent on local density fluctuations. For
instance, even if the average particle concentration of C particles is less
that the supersaturation threshold, it may be higher locally and give rise
to spontaneous nucleation. Similarly, aggregation is a function of the
 5.6 Liesegang patterns 201
particle density in the vicinity of an existing solid cluster, which is also a
locally fluctuating quantity.
The cellular automata approach naturally accounts for these fluctua­
tion phenomena and, in addition, captures the mesoscopic nature of the
precipitate cluster, that can be fractal. As we shall see in the next section,
this method turns out to be quite successful in modeling the formation of
Liesegang patterns. After that, we show that a mixed lattice Boltzmann
and lattice gas model can also be developed with the advantage of yielding
much faster simulations.
5.6.2 The lattice gas automata model
Most of the ingredients we need for modeling the formation of Liesegang
pattern have already been introduced in this chapter when describing the
A + B C reaction-diffusion process. In the case of Dee's scenario, we also
�

need to provide a mechanism for spontaneous nucleation (or precipitation)

in order to model the transformation of a diffusing C particle into a solid
D particle. Finally, aggregation of C particles on an existing D cluster will
be modeled in very much the same spirit as the DLA growth described in
section 4.4.2. The key idea will be to introduce threshold values to control
both of these processes.
More precisely, nucleation and aggregation phenomena are implemented
according to general principles of supersaturation theory, but applied at
a microscopic level. First, the C particles, once created, diffuse until their
local density (computed as the number of particles in a small neighborhood
divided by its total number of sites and lattice directions), reaches a
threshold value ksp· Then they spontaneously precipitate and become D
particles at rest (nucleation). Here, we typically consider 3 x 3 Moore
neighborhoods centered around each lattice site.
Second, C particles located in the vicinity of one or more precipitate D
particles aggregate provided that their local density (computed as before)
<
is larger than an aggregation threshold kp ksp· Finally, a C particle
sitting on top of a D particle always becomes a D (a third threshold k
could also be used here).
The parameters kp and ksp are the two main control parameters of
the model. The introduction of these critical values refers to the qualita­
tive models of solidification theory, relating supersaturation and growth
behavior [162].
As we said in section 4.4.2, it is common, from a microscopic point of
view, to describe the aggregation process in terms of a noise reduction
algorithm: C particles aggregate on a D cluster only after several en­
counters. The algorithm we use here is slightly different but allows us to
produce a fast enough aggregation process, compare to the speed of the
202 5 Reaction-diffusion processes

reaction front. This is important because a difference in these time scales

is crucial for the formation of separate bands of precipitate.
From experiments, it is observed that Liesegang patterns are obtained
only for a narrow range of parameters and fine tuning is necessary to
produce them. In particular, it is important that the initial A concentration
be significantly larger than the initial B concentration. In a cellular
automata model with an exclusion principle, such a large difference implies
having very few B particles. As a consequence, the production rate of
C particles is quite low because very few reactions take place. For this
reason, we have considered a pseudo-three-dimensional system composed
of several two-dimensional layers. The reaction has been implemented so
that particles of different layers can interact.
As a first case, we consider the situation of Liesegang bands (test tube
with axial symmetry). The initial condition is built as follows : at time
t = 0, the left part of the system (x ::;: 0) is randomly occupied by A
>
particles, with a density ao and the right part (x 0) is filled with B
particles with a density b0, as described in figure 5.8.
5.6.3 Cellular automata bands and rings
Figure 5.8 shows a typical example of a cellular automata simulation with
C particles, giving rise to bands. From the positions Xn and the formation
time tn of each band, we can verify the spacing and the time laws. For
instance, the plot given in figure 5.9 shows very good agreement for the
relation Xn/Xn- l � p. It is found that the so-called Jablczynski coefficient
p is 1.08, a value well within the range of experimental findings. Indeed,
the values obtained in different experiments with axial symmetries show
that 1.05 ::;: p ::;: 1.20. The way the value of p depends on the parameters
of the model has not been investigated yet.
As we said earlier, Liesegang patterns are found only if the parameters of
the experiment are thoroughly adjusted. In our simulation, kv and ksp are
among the natural quantities that control supersaturation and aggregation.
In practice, however, one cannot directly modify these parameters. On
the other hand, it is experimentally possible to change some properties
of the gel (its pH for example) and thus influence the properties of the
aggregation processes or the level of supersaturation.
Outside of the region where Liesegang patterns are formed, our simul­
ations show that, when kv and ksv vary, other types of patterns are
obtained. These various patterns can be classified in a qualitative phase
diagram, as shown in figure 5. 10. An example of some of these "phases"
is illustrated in figure 5.1 1. Note that the limits between the different
"phases" do not correspond to any drastic modification of the patterns.
There is rather a smooth crossover between the different domains. The
 5.6 Liesegang patterns 203

300

Xn 200

1 00

0 L_--�--�---L--�
0 1 00 200 300
Xn-1

Fig. 5.9. Verification of the spacing law for the situation with C particles. The
ratio Xn / Xn-1 tends to p = 1 .08.

No precipitation
I ,
� BANDS
\- - - - - - - .
Amorphous '. Dendrites'
solidification 1
I
:
I

Fig. 5. 10. Qualitative phase diagram showing the different possible patterns that
can be obtained with our cellular automata model, as a function of the values of
ksp and kp .

associated names are borrowed from the phenomenological theory of

solidification [162].
The terminology of dendrite comes from the tree-like structures that
are sometimes found on the surfaces of limestone rocks or plates and
that can be confused with fossils. The plant-shaped deposit is made
of iron or manganese oxides that appear when at some point in the
geological past the limestone was penetrated by a supersaturated solution
of manganese or iron ions. It turns out that the formation of these mineral
dendrites can be simulated by the same scenario as Liesegang patterns,
204 5 Reaction-diffusion processes

(a)

(b)

(c)

Fig. 5. 1 1 . Examples of patterns that are described in the phase diagram : (a)
corresponds to homogeneous clustering ; this is also the case of pattern (b) but
closer to the region of band formation. Pattern (c) shows an example of what
we called a dendrite structure. Amorphous solidification would correspond to a
completely uniform picture.

=
but with an aggregation threshold kp 0. Figure 5.12 shows the results
of such a modeling. The fractal dimension of these clusters is found to
be around 1.77, a value which is very close to that measured in a real
sample [163].
The patterns we have presented so far show axial symmetry, reflecting
the properties of the experimental setup. But the same simulations can be
repeated with different initial conditions. A case of interest is the situation
of radial symmetry responsible for the formation of rings or spirals. The
reactant A is injected in the central region of a two-dimensional gel initially
filled with B particles. The result of the cellular automata simulation is
shown in figure 5.13. In (a) concentric rings are formed, starting from
the middle of the system and appearing as the reaction front radially
moves away. In (b) a spiral-shaped structure is created. Although the two
situations are similar as far as the simulation parameter are concerned,
the appearance of a spiral stems from a spontaneous spatial fluctuation
which breaks the radial symmetry.
 5.6 Liesegang patterns 205

7.0

"'
,.-_

"'
«l
s (b)
'-"

.s

3.2
0.65 In (size) 2.8

-0.4

�
u
§
0..
::s
u
(c)
u
0
'-"

.s
-4.2
(a) 0.65 In (size) 2.8

Fig. 5 . 1 2. (a) Examples of mineral dendrite obtained from a cellular automata

simulation with kp = 0 ; in this figure, the reaction front moves from upward.
The two graphs on the right show the numerical measurement of the fractal
dimension using : (b) a sand-box method and (c) a box-counting technique.

(a) (b)

Fig. 5 . 1 3 . Liesegang rings (a) and spiral (b), as obtained after 2000 iterations of
the cellular automata model with C particles (indicated in gray).
206 5 Reaction-diffusion processes

5.6.4 The lattice Boltzmann model

In the previous section, we had to use a pseudo-3D simulation in order to
reach a suitable set of parameters. Consequently, a pure cellular automata
approach can be very demanding as far as computation time is concerned.
In order to study systems with more bands or rings, it is desirable to
have a faster numerical scheme. In addition, for a systematic analysis,
it is also necessary to have more flexibility to play with the values of
parameters.
The natural solution is of course to consider a lattice Boltzmann ap­
proach where, instead of Boolean occupation numbers, this is the prob­
ability of the presence of a particle which is directly simulated on the
computer.
The lattice Boltzmann approach suppresses the inherent noise present in
a cellular automata model, due to the discrete nature of particles. Although
this is usually a desirable property, here, it removes an important ingredient
of the process : as we mentioned earlier, fluctuations trigger spontaneous
precipitation and also influence the aggregation processes.
Indeed, in the cellular automata model, nucleation occurs only at sites
whose neighborhood contains at least kspN particles, where N is the
product of the number of sites in the neighborhood times the number of
= =
directions of motion (typically, for a 3 x 3 neighborhood, N 3 2 x 4 36).
If, at a given site, the average particle density per direction is p, the
probability that nucleation occurs is given by:
Pnucl (P) =
"""
6 t
( )P
N t (1 - p) N -t (5.36)
t?.kspN
A similar relation can be derived for the aggregation probability. Thus, in
this approach, nucleation or aggregation may happen even if the average
density is below the critical thresholds, that is even if the system is not
macroscopically supersturated.
In order to restore the stochastic aspects given by equation 5.36 in
the lattice Boltzmann approximation, we may introduce a probabilistic
component for the nucleation and aggregation processes : nucleation and
aggregation will take place only with given probabilities, when the concen­
tration reaches some renormalized threshold ksp or kp. The determination
of these probabilities, as well as the new values of ksp or kp, can be done so
as to reproduce qualitatively (again, the microscopic details do not matter
so much) the behavior of 5.36.
The lattice Boltzmann method allows us to gain a factor of 100 in
the speed of the simulation, as compared with the equivalent, pseudo-3D
cellular automata model discussed in the previous sections.
 5.6 Liesegang patterns 207

Fig. 5. 14. Example of the formation of Liesegang bands in a lattice Boltz­

mann simulation. The model considered here corresponds to a scenario with
intermediate C particle.

1 .9 ,-�-----,--,----�-�-,--r-....,.--;,---,

1.8
1 .7
i 1 .6
.s 1.5 ks/a0=0.0081 •

1 .4 ks/a0=0.0088
ks/a0=0.0089
1 .3 ks/a0=0.0090 x

ks/a0=0.009 1
•

1 .2 ksja0=0.0092
1.1 '----�-_.._-'-----'--�---''---'-_j
2 2. 1 2.2 2.3 2.4 2.5 2.6 2.7 2.8 2.9 3
log(xn )

Fig. 5. 1 5. Dependence of the width Wn of the Liesegang bands as a function of

their position Xn, for various values of ksp. for the concentration ratio bo/ ao = 0.01.
We observe W n = x�·59, independently of the value of ksp ·

Figure 5.14 shows an example of a lattice Boltzmann simulation con­

taining up to 30 consecutive bands, in a system of sizes 1024 x 64.
From the lattice Boltzmann simulations we may study how the width
W n of a band depends on its position X n along the tube. Here, the width
has been measured as the total spatial extension of a band.
Whereas the time and spacing laws have been investigated in great
detail in many experiments, much less seems to be known about the width
law. In most experiments, it is observed that Wn increases with Xn . A
linear relation, derived from the width of only a few consecutive bands,
has been proposed in the literature [159,164]. Our simulations predict a
more general relation

from which we conclude (using the spacing law) that

Wn
-- --+
Wn -l
( )
Xn
--
X n -l
IX

-+ p
IX
(5.37)
where ()( is an exponent which depends on the initial concentration ao
and b0 . We find, for several simulations, that ()( is typically in the range
208 5 Reaction-diffusion processes

(a) (b)
Fig. 5. 1 6. Formation of (a) Liesegang rings and (b) spiral-shaped pattern, as
obtained after 2000 iterations of the lattice Boltzmann model with C particles.

0.5-0.6 (see figure 5.15). Note that, in this model, aggregation of several D
particles on top of each other is possible. If this superposition mechanism
is not permitted, it is found that the bands are wider and an exponent
rx � 1 is then obtained, in agreement with experimental observations [164].
5.6.5 Lattice Boltzmann rings and spirals
We can consider again the case of Liesegang rings and spirals, but now in
the framework of the lattice Boltzmann model.
For instance, figure 5.16(a) shows the situation where concentric rings
=
of precipitate are formed. The numerical parameters are : ao 1, bo / ao =
= = p =
0.013, Db/ Da 0.1, ksvl ao 0.0087, k / ao 0.0065. The nucleation
process takes place with a probability of 0.05 and aggregation with a
probability close to 1. This pattern turns out to be quite similar to real
Liesegang structures obtained in a similar experimental situation [152].
=
For the same set of parameters, but bo / ao 0.016, a different pattern is
observed in figure 5.16(b). There, a local defect produced by a fluctuation
develops and a spiral of precipitate appears instead of a set of rings. Such
a spiral pattern will never be obtained from a deterministic model without
a stochastic component.
From these data, we can check the validity of the spacing law for ring
formation. The relation
r /r - 1 � P
n n

holds, where r is the radius of the nth ring. In figure 5.17, the Jablczynski
n

coefficient p is plotted as a function of the concentration of B particles

 5. 7 Multiparticle models 209
1 .3 ---

.�
,...,

1 .2
(.)
s
<!.)
0
(.)

:Q
rfl

N l. l
�
(.)
�
......

1 .0 ---�--- - � -�-

0.009 0.01 1 0.0 1 3 0.0 1 5 0.0 1 7

' �- -

Initial concentration o f B

Fig. 5. 1 7. Jablczynski coefficients p as a function of the concentration b0 of B

particles (for ao = 1 ), for bands (lower curve) and rings (upper curve).

=
bo (for ao 1) both for axial (bands) and radial (rings) symmetries. We
notice that p decreases when bo increases in agreement with experimental
data. Moreover, for the same set of parameters, the value of p is found to
be larger in the case of rings than it is for bands.

5.7 Multiparticle models

Several times in this book, we have presented a cellular automata (or

lattice gas) model and then introduced a lattice Boltzmann version, in
order to improve the numerical efficiency, for the price of losing the
many-body correlations.
Clearly, lattice Boltzmann models are less noisy and provide more
flexibility than their Boolean counterpart. But they may exhibit some bad
numerical instabilities (that is the case of lattice BGK models of fluids)
and they sometimes fail to account for relevant physical phenomena. An
example is provided by the anomalous kinetics in simple reaction-diffusion
processes.
The objective we would like to consider here is to find a dynamics in
between lattice gas automata and lattice Boltzmann models. Multiparticle
models, as we call them, provide such a possible alternative. The main idea
is to conserve the point-like nature of particles, as in cellular automata,
but to allow an arbitrary number of them to be present at a lattice site.
In other words, we want to relax the exclusion principle that plagues the
210 5 Reaction-diffusion processes

cellular automata approach and which appears as a numerical artifact

rather than a desirable physical property.
Mathematically speaking, this means that the state of each lattice site
cannot be described with a finite number of information bits. However, in
practice, it is easy to allocate a 32- or 64-bit computer word to each lattice
site, to safely assume that "any" number of particles can be described at
that site.
If the number of particles per site is N, the intrinsic fluctuations due to
the discrete nature of the particles will typically be of the order JN. This
is small compared to N, if N is large enough. Therefore, we do not have
to perform much averaging to get a meaningful result.
In addition, with an arbitrary number of particle per site, we have
much more freedom to enforce a given boundary condition, or tune a
parameter of the simulation. Thus, for modeling a reaction process, it is
almost necessary to get rid of the exclusion principle. In equation 5.14
we showed how the exclusion principle slows down the reaction when the
concentration of C is not small. To describe more complicated situations,
such as rnA + nB � C, it is desirable to have more than four particles per
site.
Unfortunately, the numerical implementation of multiparticle models is
much more involved than lattice Boltzmann or cellular automata models
and the computation time is also much higher. On the other hand, we re­
store in a natural way the fluctuations that are absent in lattice Boltzmann
simulations and provide an intrinsically stable numerical scheme. Besides,
when compared to cellular automata, the extra computational time may
be well compensated by the fact that less averaging is required.
The purpose of this section is to introduce in more detail multiparticle
models. Here, the discussion will focus on the case of a reaction-diffusion
system, which is the topic of this chapter. In section 7.3, we shall also
present a multiparticle fluid model.
5. 7. 1 Multiparticle diffusion model
Our algorithm is defined on a d-dimensional Cartesian lattice of spacing
2. Each lattice site r is occupied, at time t, by an arbitrary number of
particles n (r , t). The discrete time diffusion process is defined as follows :
r,
during the time interval each particle can jump to one of its 2d nearest­
neighbor sites along lattice direction i with probability Pi , or stay at rest
=
with a probability Po 1 - L:: ?! 1 Pi ·
An advantage of dealing with multiparticle dynamics is that advection
mechanisms can be added to the diffusion process. When the probabilities
of jumping to a nearest-neighbor site are different in each direction, a drift
is introduced. This adds a density gradient term to the diffusion equation
 5.7 Multiparticle models 211
which then reads

where V is the advection velocity. Such an advection effect is difficult to

produce without an artifact when an exclusion principle holds.
For the sake of simplicity, we shall now consider a two-dimensional case.
The generalization is straightforward and follows the same reasoning.
The idea is to loop over every particle at each site, decide where it goes
and move it to its destination site. In terms of the particle numbers n(r, t),
our multiparticle rule can be expressed as
n(r ,t) n(r+Jce3 ,t)
n ( r, t + -r ) = L Pot (r, t) + L pu( r + Jc h t)
t=l t=l
n(r+Aet ,t) n(r+.ice4,t)
+ 2:::
t=l
P3t (r + Jc e 1 , t) + 2::: P2t( r + Jc e4 , t))
t=l
n(r+Jcht)
+ 2:::
i=l
P4t (r + Jc e2 , t)) (5.38)

The vectors e 1 = - h e2 = -e4 are the four unit vectors along the main
directions of the lattice. The stochastic Boolean variable P it (r, t) is 1 with
probability Pi and selects whether or not particle t chooses to move to site
r + Jc e i. Since each particle has only one choice, we must have
Pot + Ptt + P2t + P 3t + P4t = 1
The macroscopic occupation number N ( r, t) = (n(r, t) ) is obtained by
averaging the above evolution rule over an ensemble of equivalent systems.
Clearly, one has
n(r,t)
=
( t=l
2::: Pit (r, t) ) PiN ( r, t)

Thus, we obtain the following equation of motion:

N ( r, t + -r ) = po N ( r, t) + Pt N (r + Jc h t)
+p 3 N (r + Jci?t, t) + P2 N (r + Jc e4 , t) + p4 N (r + Jch t)
( 5.39)
Assuming N varies slowly on the lattice, we can perform a Taylor expan­
sion in both space and time to obtain the continuous limit. Using :Z::: Pi 1 =
212 5 Reaction-diffusion processes
and ei = - ei=2 , we obtain
2
rotNCr, t) + � 8t2 N(r, t) + CD(r3 ) = A [(p3 - pl )e1 + (p4 - P2 )e2] · VN(r, t)
+ � (p1 + p3)(el · vfN(r, t) + � (p2 + p4)(el · V)2 N ( r, t) + m( A 3 )
(5.40)
Since, e1 and e2 are orthonormal, we have
(e1 · V ) 2 + (e2 · v f = V2
In order to use this property it is necessary that Pl + P 3 = P2 + p4 ,
otherwise the lattice directions will be "visible". Thus we impose the
isotropy condition
1 - Po
Pl + P 3 = P 2 + P4 = -2-
and we obtain
8tN (r , t) + � 8t 2 N ( r, t ) + CD(r2 ) V·VN(r, t)
+DV2 N (r , t) + C9 ( A3 ) ( 5.41)
where V is the advection velocity
�
v = [(p 3 - pl ) e 1 + (P4 - P2) e21
!

and D the diffusion constant

A2
D=
4
/ 1 - Po)
We may now consider the limit A 0 and r
----+ 0 with A2 ! • constant,
----+ ----+

as usual in a diffusion process. However, here, some additional care is

needed. If P 3 -=/= Pl or P4 -=/= P2 , the advective term will diverge in this
limit. This means that P3 - Pl or P4 - P2 must decrease proportionally to A
when the limit is taken. Thus, with a halved lattice spacing, the difference
between P i and P i+2 must also be halved in order to produce the same
advection. With these assumptions, we obtain, in the macroscopic limit
8tN = V · VN + DV2 N

5. 7.2 Numerical implementation

The main problem when implementing our algorithm on a computer
(for instance, for the two-dimensional case we described in the previous
section) is to find an efficient way to select the particles at rest and
 5.7 Multiparticle models 213
distribute randomly the others among the four possible directions of
motion. More precisely, we have to compute quantities such as
n(r,t)
ni = L Pit (f, t)
t=l
In practice, we can loop over all t particles at every site and, for each of
them, choose a random number r, uniformly distributed in the interval
[0, 1] . Then, we consider a division of this interval in subintervals [rj, ri+1 [,
= = =
j 0, ... , 5, so that Pi ri+l - ri . We say that Pit 1 if and only
<
if ri+l ::::;; r ri . The quantities ni are thus distributed according to a
multinomial distribution.
This procedure is acceptable for small values of n but, otherwise, very
time consuming. However, when n is large (more precisely when npi(1 -
Pi) � 1, the statistical distributions of the ni is expected to approach
Gaussian distributions of mean npi and variance npi(1 -pi). This Gaussian
approximation allows us to be much more efficient because we no longer
have to generate a random number for each particle at each site.
= = =
For simplicity, take the case Po = p and Pl P2 P3 P4 (1 - p)/4. =
The n/s can be approximated as follows : we draw a random number
no from a Gaussian distribution of mean np and variance np(1 - p)
(for instance using the Box-Muller method [1 6 5] ) . This number is then
rounded to the nearest integer.
Thus, in one operation, this procedure splits the population into two
parts : no particles that will stay motionless and n - no that will move.
In a second step, the n - no moving particles are divided into two sub­
sets according to a Gaussian distribution of mean nm /2 and variance
nm (1/2)(1/2). Splitting up each of these subsets one more time yields the
number ni of particles that will move in each of the four lattice directions.
If advection is present, we can also proceed similarly. First, we divide
up the moving particle population into two parts : on the one hand,
those going to north and east, for instance, and on the other hand,
those going south and west. Second, each subpopulation is, in turn,
split into two subsets according to to the values of the Pi S. Of course,
as in traditional lattice gas automata, these splitting operations can be
performed simultaneously (in parallel) at each lattice site.
Empirical considerations, supported by theoretical arguments on bino­
=
mial distributions, show that ni 40 is a good threshold value in two
dimensions, above which the Gaussian procedure can be used. Below this
critical value, it is safer to have the algorithm loop over all particles. Note
that in a given simulation, important differences in the particle number
can be found from site to site and the two different algorithms may have
to be used at different places.
214 5 Reaction-diffusion processes
5.7.3 The reaction algorithm
We will now discuss how reaction processes can be implemented in the
framework of multiparticle models (see also [166]). Reaction-diffusion
phenomena can then be simulated by alternating the reaction process be­
tween the different species and then the diffusion of the resulting products,
according to the multiparticle diffusion algorithm just described.
=
A reaction process couples locally the different species A1, l 1, .. , q to
=
produce new species Bj, j 1, .. , m according to the reaction :
k
1X1 A 1 + 1X2A 2 + . . . + IXqAq /h B 1 + /h B2 + . . . + f3m Bm
�

The quantities IXz, /3j are the stoichiometric coefficients, and k is the reaction
constant.
In order to model this reaction scheme with a multiparticle dynamics,
one considers all the q-tuples that can be formed with IX l particles of A 1 ,
1X2 particles of A 2 , etc. These q-tuples are transformed into m-tuples of Bj

( � ) ( � ) (::q )
particles with probability k. At site r and time t, there are
n n 2
JV ( r, t) = 1 • • • ( r, t)
iX
ways to form these q-tuples, where nx(r, t) denotes the number of particles
of species X present at (r, t). If one of the nA; < IXi then obviously JV = 0.
This techniques offers a natural way to consider all possible reaction
scenarios. For instance, in the case of the annihilation reaction 2A 0, �

suppose we have three particles (labeled a 1 , a2 , a3) available at a given

lattice site. Then, there are three possible ways to form a reacting pair:
(a1, a2 ), (a 1 , a 3) and (a2 , a3). In principle, all these combinations have the
same chance of forming and reacting. However, if (a1, a2 ) react, then only
a 3 is left and there is no point in considering (a1 , a3) or (a2 , a3) as possible
candidates for reaction. Thus JV is the maximal number of possible
events, but it is likely that the available particles are exhausted before
reaching the end of this list of possible reactions.
The multiparticle reaction rule can therefore be summarized as follows :
• lXI
As long as there are enough particles left (i.e. at least of species A1,
for each l), but no more than JV times, choose a Boolean random K
which is 1 with probability k.
• =
If K 1, remove from each species A1 a number of particles IX[
(nAt � nA t - IX!) and add a number /3j of particles to each species Bj,
=
j 1, ... , m (nBj � nBj + /3j).
This algorithm can easily be extended to a reversible reaction.
 5. 7 Multiparticle models 215

k=0. 8

6 �������

0 ln(time) 10
Fig. 5. 1 8. Time decay of NA, the total number of A particles in the A + A ---+ f/J
reaction-diffusion process, with the multiparticle method. A non-mean-field power
law t-d/ 2 is observed.

When k is very small, we may assume that all the % q-tuples need to
be considered and the above reaction rule can be expressed as
%(r, t)
nA/r, t + T) = nAJr, t) - CX[ L Kh
h=l
%(r, t )
nBir, t + T) = nBj (r, t) + /3j L Kh (5.42)
h=l
where Kh is 1 with probability k.
This algorithm may become quite slow in terms of computer time if
the nx are large and k � 1. In this case, the Gaussian approximation
described in the previous section can be used to speed up the numerical
simulations: the number of accepted reactions can be computed from a
local Gaussian distribution of mean k..!V(r, t) and variance k(1 - k)JV(r, t).
In order to check that our multiparticle reaction rule captures the
true nature of fluctuation and correlation, we simulated the A + A 0 -+

reaction-diffusion process. We explained in section 5.4 that this reaction

exhibits a non-mean-field decay law in one-dimensional systems.
Figure 5.18 gives the behavior of a simulation performed on a line
of 64 536 sites, with an initial number of about 100 particles per site.
Diffusion and reaction processes are simulated with our multiparticle
algorithms with a probability 1 /2 that each particle moves left or right
and a reaction probability k 0.8. We observe that the total number of
=
216 5 Reaction-diffusion processes
A particles decreases with time as the power law NA (t) ,...., t- 1 12 , which is
=
the expected result in d 1 dimension.
5.7.4 Rate equation approximation
In a mean-field approximation, i.e. when the multipoint correlation func­
tions are factorized as a product of one-point functions and the reaction
probability k is much smaller than 1, our multiparticle dynamics gives the
expected rate equation given by the mass action law.
To show this, the first step is to average the discrete physical quantities
over an ensemble. The mean-field (or Boltzmann factorization approxi­
mation) yields

The first important result is that, if one assumes that each configuration
with a total of M particles over L sites appears with the same probability,

( )) ( ) ,....,
the quantities (n:,l ) can be computed
/ nA 1 ___!___ MA 1
__!__ (5.43)
\ rxz =
Lrx1 rxz rxz ! Ni1
where NA1 is the average particle number per site of species Az and the
number of lattice sites L is supposed to be large enough.
This result can be derived from the following combinatory arguments :
let us suppose that a number M of particles is homogeneously distributed
among L sites of the lattice. The average number N of particles on these
=
sites is then N MIL .
The probability for a particle to pick a given site r to stand on is 11 L
(and 1 - 1 I L to choose any other site). Thus the probability that a given
() ()
site contains n particles is
P (n) - _
M -1 n
1--
( )
1 M-n
-
M (L l) M - n
_
_

n L L n LM
The average we want to compute is

1 M M! (L - l) M -n
rx! �
(M - n) !(n - rx) ! LM
1 M' (M' + rx) ! (L - 1) M' -n'
rx! E(M' - n') !n' ! LM' +rx
 5. 7 Multiparticle models 217
( + a)
M' L
-rx M' M' ! (L - 1) M' -n'

(M'a+ a)
11. E (M' - n') !n' ! L M'
-
L rx (1 + 1) M' �� (5.44)
where M' = M - a and n' = n - a.Thus

where N M / L is the average number of particle per site. When the

=

system size is large (M � 1) we can approximate

which proves equation 5.43.

With this result, we can now obtain the mean-field approximation of
equation 5.42:

NB/ t + r ) - NBj(t) = k ' N�� N�� . . . n�:

where the reaction constant k' is
I
'
k
' '
k =
1/.t . 11.2 • • • • ll.q .

1:
In the limit � 0, we obtain the usual form of the rate equations for the
reaction process considered, namely
a t NA- ( t ) - k ' Nrx! Nrx2 Nrxq
'f:
I
- -- A 1 A2 · · · Aq

k Nrx2 N rxq
- -1:' Nrx1
a t NB,. ( t ) - A 1 A2 · · · Aq

5.7.5 Turing patterns

In this section, we use our multiparticle reaction-diffusion model to sim­
ulate the formation of the so-called Turing structures. Turing [167] was
the first to suggest that, under certain conditions, chemicals can react
and diffuse so as to produce steady-state heterogeneous spatial patterns
of chemical or concentrations [50]. Turing structures are believed to play
an important role in biological pattern formation processes, such as the
stripes observed on the zebra skin [130]. In contrast to most hydrody­
namical instabilities, the structure of Turing patterns is not related to any
218 5 Reaction-diffusion processes

imposed macroscopic length scales (like the size of the container). Turing
patterns exhibit regular structure with an intrinsic wavelength depending
on the diffusion constants and reaction rates. Typical examples of inho­
mogeneous stationary states observed in experiments have a hexagonal or
a striped structure [168].
For the sake of simplicity, we consider here only one of the simplest
models showing Turing patterns : the Schnackenberg reaction-diffusion
model [169] in two dimensions. It describes the following autocatalytic
reaction :
A�X
2X + Y � 3X (5.45)
where the densities of the species A and B are kept fixed (for instance
by external feeding of the system). This situation of having a fixed
concentration of some chemical is quite common in reaction-diffusion
processes. As a result, there is no need to include all the dynamics of
such reagents in cellular automata or multiparticle models. It is usually
enough to create randomly a local population of these particles at each
lattice sites.
Here we consider a two-dimensional multispecies multiparticle lattice
gas model with alternating reaction and diffusion steps, as explained in
sections 5.7.1 and 5.7.3, namely
k
• Reaction step : for the reaction nS1 +mS2 � qS3 all possible families
m
composed of n S1 and S2 particles have a probability kj to create
q particles s3 .
• Diffusion step : the particles jump to a nearest-neighbor site with
an equal probability for each direction. This gives a bare diffusion
coefficient of ..12 /4r. A simple way to tune it without having a
probability for a particle to be at rest is to vary the number t
of consecutive diffusion steps for a given species : this amounts to
=
introducing a different time step Trn rjt for this species and yields
D = tA-2 j4r.
The instability of the homogeneous state leading to Turing struc­
tures can be understood using the corresponding macroscopic rate equa­
tions [130] for the local average densities x and y
8t x k 1 a - k2 x + k 3 x2 y + Dx V2 x
8ty = k4 b - k3x2 y + Dy V2 y ( 5.46)
where a and b represent the densities of particles A and B, respectively.
A conventional analysis shows that for some values of the parameters a
 5.8 From cellular automata to field theory 219

• • •!.
(a) (b)

Fig. 5 . 1 9. Turing patterns obtained in the Schnackenberg reaction i n the long

time regime. (a) Multiparticle model and (b) mean-field rate equations.

homogeneous stationary state is unstable towards local density perturba­

tions. Inhomogeneous patterns can evolve by diffusion-driven instabilities
providing that the diffusion constants Dx and Dy are not the same.
The region of the parameter space (a, b, Dy/ Dx, ... ) for which homoge­
neous states of the system are unstable is called the deterministic Turing
space.
Figure 5.19 shows the configuration obtained in the long time regime
with our multiparticle model and the corresponding rate equations 5.46. In
both cases, a hexagonal geometry is selected. The right panel corresponds
to the solution of the rate equations, while the left panel corresponds
to the multiparticle simulation. As we can see, the two pictures are
quite similar. Thus, it is not clear that the multiparticle (which brings
fluctuations into play) adds anything compared with the predictions of
the mean-field rate equations (which use less computer time). However,
there are some indications [170] that the Turing space may be enlarged
when fluctuations are considered.
5.8 From cellular automata to field theory

As we have seen, reaction-diffusion problems have a rich structure which

can lead to complex cooperative behaviors in the long time regime. These
cooperative behaviors have some universal properties. This means that
some aspects of the dynamics are "forgotten" or irrelevant in the long
time regime, while more basic aspects related to symmetries and con­
servation laws govern the long time dynamics. Thus, different reaction­
diffusion systems can have a similar long time behavior. In a language
borrowed from the renormalization group [171] approach of critical phe­
nomena, one says that these systems belong to the same universality
class.
220 5 Reaction-diffusion processes

It is difficult from the inspection of the cellular automaton rules (or

multiparticle dynamics) to predict what will happen in the long time
regime. There are too many microscopic variables involved, and the
collective behavior shows up at a macroscopic scale. It is also difficult to
define a renormalization group transformation at the level of the cellular
automaton rule.
A better strategy is to find, starting from the microscopic rules, the
equations of motion of well chosen coarse-grained variables. These
"macroscopic" variables </Ji("r, t), i 1, 2, . take continuous values and
=

vary continuously in space and time. They describe the system at the
macroscopic scale, while keeping track of all the fluctuations related to
the disregarded microscopic degrees of freedom. Such a description is
called a continuous field theory. These field theories can then be studied
in their own right, using standard methods such as perturbation techniques
or renormalization group.
The first step to perform this program is to notice that the probabilistic
cellular automaton rules used for reaction-diffusion can be expressed as a
master equation (in discrete time) for generalized birth-death processes.
The way to associate a continuous field to a master equation for
a birth-death process (in continuous time) has been studied by many
people [172-174].
For the sake of simplicity, and to avoid as much as possible technical
difficulties, we shall start with a very simple one-cell problem (zero­
dimensional system). Moreover, we shall assume that only one type of
particle is present. The state of the cell is characterized by </J n (t), the
probability of having n particles at time t. We consider a multiparticle
model, with no exclusion principle, so that any value of n is allowed.
One considers a dynamical process described by the following master
equation
=
8 t </J n (t) L [w(n' � n )</J n' (t) - w(n � n' )</J n (t)] (5.47)
'
n =/= n

where w(n' � n) is the transition rate from the state with n' particles to
the state with n particles. One recognizes the degenerate form (only one
site) of a probabilistic cellular automaton rule. This stochastic process is
Markovian: there is no memory effect.
A natural way, in statistical physics, to deal with a problem in which
the particle number is not conserved is to introduce a so-called Fock space
structure. One works in a phase space which is the direct sum of all the
phase spaces corresponding to an arbitrary number of particles. Creation
and destruction of particles is described by creation and annihilation
operators acting on this Fock space. To each state { <P } one can associate
n

the state 1 <1>) considered as an element of a real vector space or a Fock

 5.8 From cellular automata to field theory 221
space §'
I <I>) = 2:= <f> n l n) (5.48)
n
where I n) is the state with exactly n particles. A useful quantity is the
generating function defined as
G(z, t) = 2:= </> n (t)z n (5.49)
n
where z is a complex number. The derivatives of G(z, t) generate the
factorial moments
ak
nk (</>) = (n(n - 1) ... (n - k + 1)) = a zk G(z, t)l z = l (5.50)
The definition of a scalar product on this vector space gives it a Hilbert
structure. Different scalar products are usually considered. We consider
here the so-called exclusive scalar product defined as
(nlm) = n ! b nm ( 5.51)
One then introduces the annihilation operator a and the creation operator
n defined as
aln) = nln - 1) (5.52)
and
n ln) = I n + 1) (5.53)
These (bosonic like) operators obey commutation relations
[a, n]_ = 1 (5.54)
and are Hermitian conjugate with respect to the exclusive scalar product.
It then follows that
<f> n (t) = 1
1 (nl <l>) = 11 (O i an I <I>( t)) ( 5.55)
n. n.
Finally, the generating function can be written down in this formalism as
G(z, t) = (z l<l>) (5.56)
where the lz) are the coherent states of Bargmann and Fock [175]
lz ) = exp(z * n) 1 0) (5.57)
One can then prove several useful properties
(z l () = ez (' ( 5.58)
J dzd2m� * e-zz" lz) (z I = I (5.59)
222 5 Reaction-diffusion processes

1 <1> ) = j d���· e-zz G(z, t) l z)

' (5.60)

a l z) = z * l z) (5.61)
In this approach the master equation takes the form
8 t i<I> ) = 2? I <D) (5.62)
the Liouvillian 2? is a polynomial in the creation and annihilation
operators and can always be written in normal form ( all the creation
operators are on the left of the annihilation operators)
(5.63)

The exact form of 2? i.e. the values that the constants CiJ take on, will
depend on the form of the transition rates w(n' � n) for the model under
consideration. Thus one can take either the w or the Liouvillian 2? to
define the model.
Before developing further this general formalism, let us consider a very
simple example, namely the simple decay A � 0 with a reaction rate w.
In this simple case, the different steps can be made explicitly. The master
equation is
D t cP n (t) -wnc/J n (t) + w(n + 1)¢ n+l (t)
= (5.64)
From equation 5.64 it follows that the generating function G(z, t) obeys
the equation
O t G(z, t) = w(l - z)oz G(z, t) (5.65)
whose solution is
G(z, t) = G((z - l) exp(-wt)) (5.66)
Thus the factorial moment of order one is
n 1 (¢) = (n) (t) =
a
= G(z, t) l z =l (n) (O) exp (-wt) (5.67)
oz
as expected from physical grounds. We can now construct the Liouvillian
2·'? . We have

O t l <l>)
= - L wnc/J n l n) + L w(n + 1)¢ n+l ln) -wnal<l>) + wa l <l>) (5.68)
=

n n
and the Liouvillan takes the simple form
2? = w(a - n a) (5.69)
Let us now return to the general case in zero dimension. To obtain a field
continuous field theory, one constructs from 5.62 and 5.63 a path-integral
 5.8 From cellular automata to field theory 223
representation for
P(z, tlzo, 0) = (zl exp t2� lzo ) (5.70)
by inserting 5.59 into 5.70 (N - 1) times, in the usual way [175]
N- 1 dz d k -
(
P(z, tlzo, 0) = IT j ; � e zkzZ ) (zN I exp c-2� 1 zN- 1 )
•

k= 1 m

x ( zN- 1 1 exp c-2� 1 ZN-z) ... (z1 1 exp c-2� 1 zo ) (5.71)

=
where ZN z and c- = t/ N. For small c- a typical term may easily be
evaluated
( zkl exp c-2� 1 zk - 1 ) ""' exp(c L Cij zt (zi:_ 1 )i + zkz;;_ 1 ) (5.72)
ij
Substituting 5.72 into 5.71 and taking the limit N � oo and c- � 0 with t
fixed gives for 5.71, after some algebra and up to boundary terms
P(z, tlzo, 0 ) = j d[z, z * ] exp -{lot dt [z z* - I.= CiJ Z i (z * )i]} (5.73)
ij

where the dot signifies differentiation with respect to time. One recovers
an object familiar in field theory. The long time properties are extracted
by studying the action
S(t, O) = lot dt' [zz·· - I.= Cij z i (z * )i ] (5.74)
ij

The basic formalism being established, we can now consider general

problems with many cells on a d-dimensional lattice. At each site r of
the lattice, one can have nr particles. One starts with a master equation
for <f>(n 1 , nz, ... , nr, ... ; t), the joint probability of finding nJ particles at site j
(j = 1, 2, ... ).
In the Fock space representation, the basis states of the system are of the
form I !!) = ln 1 , nz, ... nr, ... ). At each site r of the lattice, one associates a pair
of creation nr and annihilation ar operators satisfying the commutation
relations
=
[ar, ns] - �rs (5.75)
The macroscopic state is
(5.76)

and the generating function

G(z 1 , Zz, ... Zr, · · ; t) = rr rr </>!lz�' (5.77)
r
n
224 5 Reaction-diffusion processes
=
The Liouvillian se:F f£(rc 1 , a 1, ... , rcj, aj, ... ) contains terms which couple
site r and its nearest neighbors. Such non-local terms are needed to
describe diffusion of the particles on the lattice. Having the Liouvillian
se:F ' one can find its path-integral form.
As a simple illustration, let us consider the diffusion of particles in
one dimension. Particles at a given site r can jump with a rate w to a
nearest-neighbor site (labeled r ± e). The masters equation reads
=
8t</>(!1, t) L L w [(nr + 1 ) </> ( ... , nr+e - 1, nr + 1, .. , t)
e
- nr <f> ( ... , nr+e - 1, nr , . , t)] (5 .7 8)
The Liouvillan se:F , is
se:F L L w ( rcr+e ar - TCr ar)
= (5 .7 9)
r e
The action of the continuous field theory becomes
t
S(t, O) h dt'{ j da r [z( z* - DLlz * ] }
= (5.80)
=
where D w .A2 , A, being the lattice constant.
As a final example, we consider the case of the diffusion annihilation
process A + A � 0. Let D be the diffusion constant of the A particles and
k the reaction rate. The Liouvillian !£ff is easily found to be
seff" D L l: Urcr+ear - TCr ar ] - k ( rc; - 1 ) a; }
= (5.8 1)
e
where e denotes the nearest-neighbor sites of r. At the coarse-grained
level, the action is
(5.82)
The long time properties of this field theory have been analyzed using a
dynamical renormalization group approach [176,177].
Until now, we have assumed that only one particle species was present.
If two or more species participate or if the particles possess internal
degrees of freedom, the formalism can be easily extended by introducing
new fields [177].
There is another way to associate a field theory with the type of
master equation we consider. This second approach that we shall call
the Poisson approach has been introduced by Gardiner and Chaturvedi.
Although equivalent to the Fock space formalism [178], the Poisson
approach illustrates more explicitly the role played by the fluctuations.
Let us consider again first the Poisson method for the one-cell problem.
Specifically one assumes that a state of the system realized at time t
 5.8 From cellular automata to field theory 225
can be expanded as a superposition of multivariate uncorrelated Poisson
distributions. One then writes
( exp � ) n a a
</J n (t) j d
= a f(a, t) (5 .8 3)
n.
In this formalism, the generating function takes the form
G(z, t) j diX exp [(z - 1)1X]/(1X, t)
= (5 .84 )
If one is given the evolution equation for G in the form
a t G(z, t) =
!l'C§ G(z, t) (5.85)
then, by replacing G by its expression 5.84 and integrating by parts
(assuming that /(IX, t) and some of its derivatives vanish at the boundaries),
one obtains a Fokker-Planck-like equation for j(IX, t) [179,180]. This
equation is not a standard Fokker-Planck equation because /(IX, t) does
not have to remain positive and thus is not a probability density. Moreover,
IX is not necessarily real, but can be complex. This Poisson representation
leads to simple relations between the factorial moments of n and the
moments of IX. One has
nk (f) j diXIXkf(IX, t) ( 1Xk )
= =
( 5 .8 6)
The equivalence of the Fock and Poisson formalisms can be shown as
follows. One has first to determine the form of the Louvillian, !l'C§ ,
defined in 5.85, which describes the evolution of the generating function,
G(z, t), with time. Given the form 5.63 for !l':F this is straightforward.
Indeed,

(5 .87 )

where we have used 5.60. Now using 5.58 and 5.61 one has
a t G(z, t) L. . Cij Z i ( j_
=
)i (z J <D)
lj
az
!l'C§ G(z, t) ( 5 .88 )
where
(5 89 )
.

One sees the close relationship between !l':F and !l'C§ . To each creation
n
operator present in !l'g; one associates a factor z in !l'C§ and to each
226 5 Reaction-diffusion processes
annihilation operator a in seJF one associates the operator a I az in !l?cg .
It is now possible to proceed from 5.85 as described above to find the
evolution equation for f(a, t). One finds
atf(a, t) = se&f(a, t) (5.90)
where
!l? & = �
a . .
"""' Cij(1 - -)1a1 (5.91)
lj
.. aa
is the Liouvillian describing the evolution of the quasi-probability density
f(a, t). This Fokker-Planck-like equation is exact, no approximation or
truncation has been made in deriving it from the master equation 5.47. It
can be used as the basis for the study of processes described by 5.47, or as
the starting point for deriving a path-integral representation in terms of
the a variables. To carry out this latter program, one introduces operators
& and p [181] satisfying
[a,A pA ] _ = z [&, &]_ = 0 [p, p]_ = 0
•

(5.92)
with
p = -z­
A
a•

(5.93)
aa
in the a-representation. Then just as the master equation 5.62 is the Pock
space analogue of 5.85, one can write an operator analogue of 5.91
(5.94)
where
se m = l: Cj(1 + ip) i ( &)j (5.95)
ij
Introducing Ia) and lp), which are eigenkets of & and p with eigenvalues
a and p, respectively, one can derive a path-integral representation for
(5.96)
in exactly the same way as described in 5.70. One finds the analogous
expression to 5. 73 to be
j d[p, a] exp -{lot dt[ipa - I,: Cij(1 + ip) iaj] } (5.97)
lj

The "action" in the path-integral 5.73 and that in 5.97 are identical as
long as the identifications
a � z * , ip � (z - 1 ) (5.98)
 5.8 From cellular automata to field theory 227
are made. Although it was not obvious a priori that the two actions would
be identical, if 5.98 is expressed in the form
a a
a +-+ ' ( z - 1) (5.99)
az aa +-+
it becomes very much more plausible, given the conjugate nature of the
variables a and (z - 1) indicated in 5.84.
To show the equivalence of the two formalisms in the general d­
dimensional case, one can proceed in exactly the same way as for the
one-site problem. First of all, an analogous procedure to that carried out
in 5.87�5.89 shows that the Liouvillian describing the evolution of G(�, t)
is !f!q; = !f!(z 1 , 0�1 , , Zj, 0�j , ) and hence that governing the evolution of
••• •••

f(q,_, t) is !J!fl' = !1!( 1 - 0�1 , a 1 , ··· ' 1 - �j , aj , ... ). Secondly, the operator ana­
a

logue of !J!fl' is immediately seen to be !!! (!) = !1!(1 + ip 1 , & 1 , ... , 1 + ipJ , &1, . ).
. .
Finally, path-integral representations for the solution of the Fokker�
Planck equations can be constructed from !f!:F and !!!(!) and shown to be
the same, if the identification
(5. 100)
is made.
Instead of considering the generalized Fokker�Planck equation obtained
in the Poisson formalism, one can consider the associated generalized
Langevin equation. There is a one to one correspondence between the
two [179]. Let us consider again the simple case A + A 0 [182]. The �

( )
generalized Fokker�Planck equation is
N a2 a
a d ({ a(n) } n , t) = - k ; 2
aa2 (m)
- 2 aa(m) (a (m)f({a(n) } n , t))
N a
-d ; aa(m) ((a(m - 1) - 2a(m) + a(m + 1))f({a(n)} n , t)) (5. 101)

(D
The corresponding generalized Langevin equation is
2 )
da(r, t) = : 2 a(r, t) - 2ka(r, t) 2 dt + J-2k a(r, t) 2 dW(r, t) (5. 102)
r
This equation should be considered as a pair of coupled stochastic equa­
tions for the real variables aR(r, t) and a1 (r, t), respectively the real and
imaginary parts of a(r , t)

(D a 2 aR(r, t) - 2k(a2 t) - a2 (r, t)) dt

ar 2 R (r, 1
)
- Jikar (r, t)dW(r, t) (5.103)
228 5 Reaction-diffusion processes

drx1 (r, t) (: 2
)
D 2 rx1 (r, t) - 4krxR(r, t)rxJ (r, t) dt
r
+ JikrxR(r, t)d W(r, t) (5.104)
As one sees from the above equations, the noise terms arising from the
microscopic fluctuations are multiplicative with non-trivial correlation.
In particular, the noise-noise correlation is negative. This is an unusual
feature related to the nonequilibrium aspect of such reaction-diffusion
problems.
A Boltzmann-like approximation will overlook such complications and
leads to erroneous results. These problems show clearly the advantage
of working with cellular automata algorithms able to model important
fluctuations in reaction diffusion problems.

5.9 Problems

5. 1 . Toffo l i and M argo l us [26] p ropose a cel l u l a r auto m ato n (tu be­
worms ru l e) to model the Belousov-Zhaboti nsky react i o n . The state of
each s i te is either 0 ( refracto ry) or 1 (exc ited) and a l oca l t i m e r (wh ose
val ue is 3, 2, 1 or 0) contro ls the refracto ry period. Each iteration of
the r u l e can be exp ressed by the fo l l owi n g seq uence of operations: ( i )
where t h e t i m e r i s zero, t h e state i s exci ted ; ( i i ) t h e t i m e r i s decreased
by 1 u n l ess it is 0; ( i i i ) a s i te beco mes refracto ry whe n ever the t i m e r
is eq ual t o 2; ( i v) the t i m e r is reset t o 3 fo r t h e exc ited s i tes w h i c h
h ave two , o r m o re than fo u r, exc ited s i tes i n the i r Moo re n e i g h borhood .
S i m u l ate t h i s auto m ato n , start i n g from a random i n itial conf i g u ration
of the t i m e rs and the exci ted states. Observe the fo rmation of s p i ra l
pai rs o f exc i tations. C heck that th i s ru l e is ve ry sensitive t o smal l
mod ifications ( i n particu l a r to the o rder of operations ( i ) to ( i v) ) . What
are the d iffe rences betwee n th is ru l e and the G reenberg-Hast i n g m od e l
d i scussed i n section 5.2?
5.2. Mod ify the G reenberg-Hasti n g ru l e (or the Toffol i and M a rgol us
tu be-wo rms ru l e of the p revious exce rc ise) in order to m od e l a one- or
th ree-d i m e n s i o n a l exc i tab l e m ed i a .
5 . 3 . Contag i o n i n e p i d e m i c mod e l : a s i m i l a r dynam i cs t o that used for
d escri b i n g an exc i ta b l e c h e m i cal med i a can be con s i d e red to model
contag i o n in an e p i d e m i c d i sease [56] . Each s i te is occ u p i ed by an
i nd iv i d u a l who can be either suceptib/e, infectious or immune. A
healthy, s ucepti b l e i nd iv i d u a l i s i nfected o n l y if there is at l east one
 5.9 Problems 229
i nfectious person i n i ts von N e u m a n n n e i g h borhood . An i nfectious i nd i ­
v i d u a l rem a i ns so d u ri n g m steps a n d t h e n beco mes i m m u ne. I m m u n ity
stays fo r m steps: afte r that the i m m u n e i nd i v i d u a l s beco m e s ucepti b l e
agai n . Co n s i d e r a CA s i m u l ation o f th i s contag i o n p rocess , start i n g
w i t h a random conf i g u ration o f s ucepti b l e , i nfectious o r i m m u n e i n d i­
v i d u a l s . What is the d i ffe rence betwee n th is model and the G reenberg­
H asti n g ru l e? Modify the e p i d e m i c ru l e to m ake a p robab i l istic CA.
What is then the d ifference from the forest fi re mod e l s d iscussed i n
chapte r 2?

5.4. P ropose a CA model of a prey-p red ato r syste m where two spec i es
( p reys and p red ato rs) d iffuse on a 20 l attice and i nte ract accord i ng to
the fo l l ow i ng ru l es : n ew p reys come to l ife at a rate proportional to
the cu rrent n u m be r of p rey (e . g . one parent p rod uces one offspri ng) .
A p rey is eate n w i th probab i l ity k when it encou nte rs a predator.
P red ato rs die with a g i ve n constant p robab i l ity and co m e to l ife when
a parent predato r eats a p rey. Ass u m i n g that p reys and p redato rs are
homogeneously d i stri b uted , show that th i s syste m is ro u g h l y descri bed
by the mean-f i e l d equations p = ap - kpP and P = -bP + kpP where
p stands fo r the total n u m ber of preys and P the tota l n u m be r of
p redato rs . The q u antities a, b and k are paramete rs of the mode l .
T h i s syste m o f equations is known a s t h e Lotka-Vo lterra model and
yields an osc i l l atory behavior. S i m u l ate the auto m ato n mode l . Are
the speci es homogeneously d i stri buted in space o r do you observe
an osc i l l ato ry space-t i m e patte rn? Mod ify the r u l e as fo l l ows ( i ) eac h
p redato r has an i nte rnal " starv i n g c l ock" w h i c h is d ec reased by 1 at
each t i m e step. When a p rey is eate n , the c l ock is reset to i ts maxi m u m
val ue; when the c l ock ru ns down to zero , the p redator d i es . ( i i ) Add a
" p regnancy c l ock" w h i c h determ i n es the ti m e n ecessary for a p redato r
t o g i ve b i rth t o a baby. H o w do these m od if i cati o n affect t h e g l obal
behav i o r? I m ag i n e othe r mod ifications that are sound from a b i o l og ical
po i nt of v i ew. Can you eas i ly write down the correspond i ng d iffe renti a l
eq u ation?

5.5. G i ve a cel l u l a r automata ru l e fo r the A -+ B -+ C -+ A cyc l i c

reaction [ 1 82] . T h i s model is characte rized b y th ree poss i b l e states
on each cel l : A, B or C . There is a cyc l ic h i e rarchy of react i o n : A i s
transfo rmed i nto B if a t l east one o f the v o n N e u m a n n n e i g h bors i s a
B. S i m i larly, B changes to C if and o n l y if a C i s arou n d . And, f i n a l ly,
a C turns i nto a n A if an A is p resent in the n e i g h borhood . Study the
behav i o r of th is system in 1 0 or 20 l attices. G e n e ra l i ze th i s p rocess
with m o re than th ree spec i es . From the s i m u l at i o n , how many speci es
do you th i n k are n ecessary to obta i ned a " froze n " confi g u ration?
230 5 Reaction-diffusion processes
5.6. Write the m ic rodyn a m i cs of the two-d i m ens i o n a l A +A � 0 reaction
in w h i c h any two part i c l es at the same s i te an n i h i l ate (not o n l y those
with oppos ite ve l oc i t i es ) . What i s the reaction rate?
5.7. Write the m ic rodyn a m i cs of the reve rs i b l e reaction d i ffus i o n p ro­
cess A + B +--+ C w h e re the rate of the d i rect and i nverse reactions are
not necessa r i l y i d e ntica l .
5.8. Write t h e m ic rodyn a m i cs o f t h e two-d i m e n s i o n a l 2A + B � 0 reac­
tion .
5.9. P ro pose a reaction ru l e correspo nd i n g to the so-cal l ed Sel kov
mode l : A +--+ X, X + 2Y +--+ 3 Y and Y +--+ B where X and Y are
the " acti ve " spec i es and A and B are h e l d at fixed concentration by
external mechan i s ms .
5. 1 0. M eas u re the ro u g h n ess of the i nterface between the A and B
spec i es i n the A + B � 0 reaction d i ffus i o n p rocess as a function of
t i m e . What a re the effects of the d iffus i o n constant and the reaction
rate o n th i s ro u g h n ess?
5. 1 1 . Cons i d e r the reactio n-d iffu s i o n p rocess A + B � 0 in a two­
d i m e n s i o n a l box with the same n u m be r of A and B particles , but with
correl ated i n it i a l confi g u rati ons. The i n itial state i s prepared as fo l l ows .
The A part i c l es a re rand o m l y d i str i b uted i n the box (Poisson p rocess) .
I n a fi rst step , the same d istr i bution i s a l so used fo r the B particles .
But then the B part i c l es a re a l l owed t o d iffuse f o r a few iterations,
wh i l e the A rem a i n at rest. In th i s way, the A and B d i str i b utions a re
m ad e statistica l l y dependent. The fu l l reaction-d iffu s i o n p rocess is then
started . Co m p ute the time dependance of the pop u l ations. Why does the
long ti m e behav i o r not ag ree with the p red i cti on NA (t) = NB (t) "' c 1 1 2 ?
Co n s i d e r the same p ro b l e m but now with the i n itial B confi g u ration
obta i ned from a tra n s l ation of the A d istri buti o n . Does th is change the
ti m e behav i o r?
5. 1 2 . O n e cons i d e rs the reactio n-d iffus i o n p rocess A + B � 0 where
the two co m po n e nts a re i n iti a l ly separated i n space {the As are l ocated
i n the dom a i n x < 0 and the Bs i n x > 0) . Auss u m e there is now, at
x = 0, a sem i-permea b l e wal l a l l owi n g the A particles to cross the wal l
but not the B s . D i scuss the dynam i cs of the reaction fro nt.
5. 1 3 . Cons i d e r the o n e-spec i es reaction-an n i h i l ation process A +A � 0
i n a fo u r-d i me n s i o n a l s pace . I n iti a l ly, al l the particles s i t on a one­
d i m e n s i o n a l l i n e . Write a CA a l gorith m e w h i c h s i m u l ates th is dynam ics.
Show that in the long ti m e reg i m e the n u m be r of A part i c l es n ever
van i s h es . Why i s th is so?
 5.9 Problems 23 1
5. 1 4. Cons i d e r the revers i b l e reactio n-d iffu s i o n p rocess A + B � C
with rate k and C � A + B with rate g . The C spec i es is also d iffusi n g .
If t h e two reactants A a n d B are i n it i a l ly separated i n space (the As
are l ocated i n the d o m a i n x < 0 and the Bs i n x > 0) and no Cs are
p resent at t i m e t = 0, what w i l l be the dynam i cs of the C reaction front
as a fu nction of k and g?
5. 1 5. One of the best-known reactio n-d iffus i o n p rocesses lead i n g to
b i stab i l ity is the Sch l o g l mode l . The chem i cal reactions c h aracte r i z i n g
th is m o d e l are:
A� �� X, 2X + B � �! 3X
(i) Write a cel l u l a r auto m aton algorithm that s i m u l ates this process i n
a c l osed system . ( i i ) Co nsider the case of a n open syste m i n wh ich A
and B are created u n ifo r m l y with a constant rate . Genera l ize the CA
a l gorithm fo r t h i s case.
5. 1 6. Cons i d e r the d ecay of a po i nt-l i ke rad i oactive sou rce. Write the
correspond i n g m aste r equat i o n . Derive the Liouvi l l i a n in te rms of the
creation and a n n i h i l ation ope rato rs . Co m p ute the time dependance of
the n u m be r of part i c l es l eft.
5 . 1 7 . Cons i d e r the fo l l owi ng reaction-a n n i h i l atio n-coag u l ation p ro­
cesses . The A part i c l es d iffuse and react acco rd i n g to the p rocesses:
A + A � A with a reaction rate g and A + A � 0 with a react i o n rate k.
Write the correspo nd i ng action of the associ ated q u antu m f i e l d theory.
 6
Nonequilibrium phase transitions

6.1 Introduction

The concept of phase transition plays a very important role in many

domains of physics, chemistry, living matter and biology. The simplest
case is the one in which a phase transition arises between two equi­
librium states when varying a control parameter. Examples of control
parameters are temperature, pressure or and concentration. The most
familiar phase transitions are the paramagnetic-ferromagnetic transition,
the liquid-gas transition and the supraconducting-normal phase transition
in a metal. Equilibrium statistical mechanics gives a coherent description
of the physics of these states. The determination of the presence of a
phase transition and the study of its properties is not generally an easy
problem, however there exists a well-defined framework to approach it.
When, at the transition some quantities have a discontinuity, one speaks of
a first-order phase transition. If the physical quantities vary continuously
one speaks of second-order phase transitions [183]. This latter case is par­
ticularly tricky because one has to face a multiscale problem. Fluctuations
of all wavelengths (from a few angstrom to the size of the system) play an
equivalent important role at the transition. However, even in these com­
plicated situations, such tools as the renormalization group method have
been developed to solve these problems [171]. Although several analytical
methods exist in equilibrium statistical mechanics, one often has to have
recourse to numerical simulations. In particular, the Monte-Carlo method
is widely used. As we have seen in chapter 2, Monte-Carlo algorithms can
be viewed as probabilistic cellular automata providing that some care is
taken with the evolution rules (checkerboard invariants).
The case of nonequilibrium phase transitions that we are now going to
discuss is even more interesting for several reasons. In this case the phases
transitions take place between different steady states of an open system.
232
 6. 1 Introduction 233
Open system means that the system is in contact with reservoirs and that
fluxes (of particles, energy, etc.) go through the system. The equilibrium
situation is thus the particularly simple case in which the fluxes are zero.
Such open systems are often encountered and can have complex be­
haviors. Examples are provided by driven-diffusive systems [184], kinetic
models with competition among different dynamical processes [185], auto­
catalytic systems which exhibit spatia-temporal patterns [167], or by living
matter [130].
One remarkable feature of equilibrium phase transitions is their uni­
versal properties in the vicinity of the phase transition point. This fact
has been noticed by the end of the nineteenth century. When expressed in
reduced units, the coexistence curves of many different liquid-gas systems
collapse onto a single curve [183]. Later it was realized that the study of
the behavior of physical quantities near criticality was of great interest.
Generically, in the vicinity of the critical temperature Tc of a continuous
phase transition, the temperature dependence of a physical quantity Y
can be written as :
Y ( T ) Yr( T ) + Ys(r)
= (6.1)
where Yr is an analytic function of T showing no singularity at T Tc, =

while Ys is a nonanalytic function of the reduced temperature r ( T -=

Tc)/ Tc. This nonanalyticity usually takes the form of a power law:
(6.2)
and y is called the critical exponent for the variable Y . Several critical
exponents can be defined. In the equilibrium case, hypothesis concerning
the homogeneity of the thermodynamic potentials leads to relations among
these exponents, called scaling relations.
To illustrate the problem, let us consider again a liquid-gas system.
At a given critical density, pressure and temperature, the system changes
continuously (without latent heat) from the gaseous to the liquid phase.
Two quantities that can be experimentally determined are the isothermal
compressibility and the specific heat at fixed density. It is found that
both quantities diverge (with different critical exponents) as the critical
point is approached. The universal character of the phenomenon is
shown by the fact that the values of these exponents are the same for
different liquids. More strikingly, these exponents turn out to have the
same values for an order-disorder transition in a binary alloy. This
universal behavior is nowadays well understood for equilibrium phase
transitions. The so-called renormalization group theory [171,186] gives
a clear explanation and classifies the different systems into universality
classes. The different universality classes are characterized only by a small
number of parameters, namely, for systems with short-range interactions,
234 6 Nonequilibrium phase transitions
the dimensionality of the system d and the number of components n of
the order parameter describing the change of symmetry at the transition.
The situation is not so clear in the framework of nonequilibrium systems.
Nonequilibrium statistical mechanics is still in its infancy and there is no
general theory available to describe such systems. Most of the known
results are based on numerical simulations. However, here again the
concept of universality classes appears to be relevant although we do not
completely understand how the universality classes are characterized.
Accordingly, we shall discuss some simple examples to illustrate the
basis ideas and possible approaches to study such nonequilibrium systems.
First we shall sketch the theoretical approaches and their difficulties. Then
we shall discuss the modeling in terms of cellular automata and review
the problems and difficulties associated with this approach.
6.2 Simple interacting particle systems

A simple class of models exhibiting nonequilibrium phase transitions is

the class of so-called in teracting particle systems [187]. These systems
are Markov processes on a lattice [188]. Each site can be in one of two
states : vacant or occupied. The enumeration of the state of occupancy
of all the lattice sites defines a configuration or a microstate of the
system. Transitions between different configurations occur via elementary
processes, related to creation, annihilation or hopping of particles. An
example is provided by the problem of directed percolation discussed in
chapter 2. We shall consider here two new models, the so-called A m odel
(or AM) and con tact process m odel (or CPM).
6.2. 1 The A model
This model has been introduced by Dickman and Burschka [189] as a
simple model describing poisoning transitions similar to the ones observed
on catalytic surfaces. One considers a d-dimensional substratum covered
by a regular hypercubic lattice. Each site has two possible states : empty
or occupied by an A particle. The first step of the dynamical process is
adsorption. The probability of a vacant site becoming occupied during a
short time interval M is pM. The second step of the process is desorption.
x
The probability of an occupied site becoming vacant is rM, provided
x
that at least one of the nearest neighbors of is vacant. During the time
interval �t, one of the two processes occurs at each site, depending on its
state. For simplicity we shall restrict ourselves to the case r (1 - p) ; thus
=

p is the only control parameter of the problem. Qualitatively speaking,

one expects that if p is large enough, an initially empty substratum will be,
 6.2 Simple interacting particle systems 235
after some time, completely covered by A particles. This is the poisoned
phase or the adsorbing state. But, if p is small enough, the desorbing
mechanism will be efficient enough to prevent such poisoning. Thus one
may anticipate the existence of a threshold value Pc such that, in the
stationary state, the covering fraction of A on the substratum XA will
be 1 for p 2 Pc (poisoned phase) and smaller than 1 for p Pc· If XA <
varies continuously across Pc the transition will be of second order and its
behavior near the threshold will be characterized by the critical exponent
/3 :
(6 . 3)
This model looks very simple and one may think that an exact analytical
solution should be possible without too much effort. This is unfortunately
not true as we shall see.

Analytical approaches. Several analytical approaches have been developed

for this type of problem. They are based on more or less well-controlled
approximations or perturbative-like expansions.
The first one is a mean-field type approximation. The simplest version
is the one-site approximation in which one writes down the evolution
equation for the probability that one cell is in a given state at a given
time. Let us illustrate this method for a one-dimensional system.
In the mean-field spirit, all the sites play an equivalent role. Let PA (t) be
the probability that a site is occupied by a particle A at time t, and po(t)
the probability that a site is empty. Assuming a discrete time evolution,
we have
PA (t + 1) = p� (t) + PPA (t) ( 1 - p� (t)) + PPo (t) (6.4)
The first term on the right-hand side expresses the probability that a given
site and its two neighbors are occupied by A particles. The second term
is the probability that a particle A having at least one nearest neighbor
site empty does not desorb during one time step. The final term is the
probability that an empty site is filled during one time step. Similarly, one
has
po(t + 1) = (1 - p)pA (t)(1 - p� (t)) + (1 - p)po(t) (6.5)
The stationary state value p� is a solution of the third-order equation
p� 3 (1 - p) - p� + p 0 and p0 1 - p�. More elaborated approximations
= =

take into account the correlations between cells. For example, one can
write the evolution equation for the joint probabilities Pij that two nearest
neighbor sites are respectively in the states i, j. In one dimension, these
236 6 Nonequilibrium phase transitions
evolution equations are
PAA (t + 1) = [
PAA (t)D(t) P�A (t) + � P6A (t) + PPOA (t)PAA (t)]
+P2 PoA (t) + P2 Poo(t) (6.6)

PoA (t + 1 ) PAA (t)D(t) [

p(1 p)
;
P6A (t) + (1 - P )PoA (t)PAA (t) ]
+ 2(1 - p)poA (t) + 2p(1 - p)poo(t) (6.7)
(1 - p) 2 2
Poo (t + 1) PAA (t)D(t) (t) + (1 - p) 2PoA(t) + Poo (1 - p) 2 (6.8)
4 PoA
=

where D(t) (PAA (t) + iPBA (t)). By symmetry, POA PAO · Moreover, the
= =

one-site probabilities Pi are obtained from the joint probabilities PiJ by

summing over all the possible values of j.
The results for the one-dimensional case are summarized in the typical
phase diagram given in figure 6. 1. The stationary values of the coverage
of A particles is given as a function of the control parameter p, both
for the one-site mean-field approximation and the direct simulation of
the CA rules which will be introduced later. Near the critical point, all
the different mean-field like approximations lead to 1 - XA (P) "' (p� - p),
i.e. to a critical exponent f3 1. The critical value of the probability p�
=

depends on the approximation considered and for the one-site mean-field

approximation, Pc 0.676. In fact, the value f3 1 is not correct when
= =

the dimension of the system is lower than an upper critical dimension

de 4. These mean-field theories often provide a reasonable qualitative
=

description but cannot make quantitative predictions because they neglect

fluctuations which play a very important role in the vicinity of the phase
transition. Higher-order joint probabilities can be taken into account and
each approximation scheme gives an approximate value for the critical
probability Pc · However, the algebra soon becomes very tedious. Note
that this approach is not restricted to one dimension; similar equations of
motion can be derived in all dimensions.
A different analytical approach allowing one to take into account the
fluctuations in a systematic way has been proposed by Dickman [190] .
By analogy with what is done in equilibrium statistical mechanics, one
performs a perturbative series expansion. The expansion is made around
a model whose dynamics may be solved exactly and the development is
performed directly in terms of the kinetic parameters rather than in terms
of a power expansion of the interaction.
We shall not discuss this formalism in detail here but simply give the
main ideas of this approach using again the one-dimensional A model as
 6.2 Simp le interac ting partic le systems 237

1 .00

0.80

0.60

><"'
0.40

0.20

0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80

Fig. 6. 1 . Phase diagram for the one-dimensional A model. The full line
correspond to the one-site mean-field approximation and the dotted line to the
results of numerical simulations of the CA rules described later.

an example. As we have seen in chapter 5, it may be convenient to describe

such a Markov process in terms of an operator formalism. However, if
we do not allow multiple occupancy of the sites, the method developed in
section 5.8 has to be modified and a different type of creation annihilation
operators must be introduced.
Each site i can be in one of the two basic states I <I>o, i ) and I <I>u)
corresponding, respectively, to site i vacant or occupied. These vectors are
orthonormal:
j,l
(<I> · ·I<I>k, l) �],k = ·
(6.9)
A configuration of the system can be written as a product:
I'P ) l {h , h , ... , A, ... }) II I <I>ki )
= =(6. 10)

Creation and annihilation operators ai , ai are then defined as

ai I <I>o,i ) I <I>l,i ) ( 6. 1 1)
ai i <I>u) 0 (6. 12)
ai I <I> l,i) I <I>o,i ) (6. 1 3)
ail <I>u) 0 (6. 14)
Note that
(6.15)
and for i f. j,
ai a} - a} ai = 0 (6. 16)
where I is the identity operator.
238 6 Nonequilibrium phase transitions
A physical state of the system at time t is given by
=
l'l' )( t) L P ( {j1 , h , ... , A, ... }, t) l {j1 , h , ... , }k , . . . }) ( 6. 1 7)
{ik}

where the sum is over all the configurations and p is the probability dis­
tribution. Physical states 1'1' ) must satisfy the positivity and normalization
conditions :
( {h , h , .. . , }k , ... } 1'1' ) � 0, V{j1 , h , . .. , A, ... } (6. 1 8 )
and
( ·1 '1' ) 1= (6.19)
where
=
(I = IJ ( (<I>o,A I + (<I>1,A I ) L( {h , h , ... , )k, . .. } l (6.20)
{ik} {ik}

Observables are represented by operators which are diagonal in the

occupation representation and the expectation of an observable d in the
state 1'1' ) is given by
(d ) 'f' ( - ldl'l' ) = ( 6 . 2 1)
The evolution of the probability distribution can be written in the form
of a Schrodinger-like equation:
d l 'l') S l ) =
'l' dt
(6.22)
where the evolution operator is defined in terms of the creation and
annihilation operators. The conservation of normalization takes the simple
form, ( -I S 0.=

The idea of the perturbation theory is to divide S into two parts,

=
as S S 0 + V , such that the evolution under S 0 is solvable and that
= = =
( ·I S 0 (I V 0. Note that ('1'0 I ( -I ; thus the stationary left eigenvector
of S 0 is the normalization state. One would like to find a formal expression
for the stationary state of the interacting 1 problem, which is a solution of
=
S l'l' ) 0, in terms of 1'1'0 ) . Let [S 0] - be the inverse of [S 0] in1 the
non-zero subspace. Multiplying the equation S l'l') 0 by ( S 0 ) - and =
1 =
using ( S 0 ) - S 0 1 - 1 '1'0 ) ( '1'0 1 , one finds
(6.23)
This is the relation needed.
 6.2 Simple interacting particle systems 239
Let us illustrate the method on the model A in one dimension. The
evolution operator S can be written in the form
(6.24)
where
(6.25)
and
(6.26)
The problem whose dynamics is given by S0 is readily solvable. sp has
two eigenvalues 0, -(1 + p), corresponding to the right eigenvectors :
=
llfJo, i) (1 - v) l<l>o, i) + v l <l>u ) (6.27)
=
llfJ u ) l <l>o,i) - l <l> t,i) (6.28)
and the left eigenvectors :
=
(lfJo,i I ( <l>o,i I + ( <l>u I (6.29)
( lfJu I v ( <l>o,i I ( 1 - v) ( <t> u I
= (6.30)
=
where v p/(1 + p).
S 0 has a unique steady state 1 \f/0 ) Tij llfJo,i) . The other eigenstates of
=

S 0 are the states l i t , i2 , ... , im) for which all the sites i t , i2 , ... , im are in the
excited state llfJ u ), while all other sites are in their ground state l lfJo,i ) . The
corresponding eigenvalues are Am - m(1 + p) with m 1, 2, ... .
= =
Let Pm be the projection onto the subspace of states with exactly m
excited sites. Then:
(jJ (jJ

[S 0r t L Am - t Pm -(1 - v) L m- t Pm
= = (6.3 1)
m=t m=t
It is necessary to introduce new creation and annihilation operators bj
and bi defined as
bJ llfJo,i) llfJt ,i) (6.32)
hi llfJt,i) 0 (6.33)
bil lfJ u ) llfJo,i) (6.34)
bi l tpo, i) 0 (6.35)
They are related to the creation and annihilation operators a i , ai as follows :
a i b i (1 + (1 - v)bi ) - (1 - v)bj((1 - v) + bi)
= (6.36)
240 6 Nonequilibrium phase transitions
a; = b i (vb j - 1) - vbJ (bi - v) (6.37)
The steady-state occupation fraction can be written as
x= ( 'P0 1(vb;b j - b; )I'P) (6.38)
The interaction terms Vi can be expressed in terms of the bs and b t s.
[S 0] - 1 V is then a six-order polynomial in v, while p appears to all order
in this operator. One therefore adopts v as the expansion parameter. If
the lowest-order calculations are easy to perform, the higher orders soon
become very tedious. Moreover, it is not clear how the series converges, if
it does.
This shows that, even more than for the equilibrium case, numerical
simulations are very necessary to study such systems.
Cellular automata approaches. The cellular automata version of this
model is straightforward. One considers a d-dimensional lattice. Each cell
of the lattice j has two possibles states : 1 '¥1 ) = 1 0) or lA). The cellular

.
automata probabilistic rules are
=
If I 'I'J )( t) 1 0) then :
{
!
I 'I'J )( t + 1) = 10lA)) ', w1t
with probability (1 - p),
h probab·1·1 1ty p . ( 6. 3 9 )
If I 'I'J )( t) = lA) then :
lA), with probability p if the site j has one nearest
neighbor empty,
I'PJ )( t + 1) = lA), with probability 1, if all the neighbors are occupied,
10), with probability (1 - p) if the site j has at least one
nearest neighbor empty.
(6.40)
Simulation of the above cellular automata rule has been performed for
chains of lengths L between 30 and 65 ' 536 [191]. As can be seen on
figure 6.1, a second-order phase transition is obtained at Pc = 0.465. The
order parameter critical exponent f3 extracted by fitting the data with the
relation 6.3 over the range 0.001 � (Pc - p) � 0.02 is f3 = 0.280 ± 0.010.
This shows that a mean-field-like approximation gives a qualitatively
reasonable phase diagram, but overestimates the critical probability Pc
and gives a poor value for the exponent {3.
For the two-dimensional case, one-site and pair mean-field approxima­
tions predict a second-order phase transition for, respectively, Pc = 0.800
and 0. 785. Here again the value of the order parameter critical exponent is
f3 = 1. The results of numerical simulations of the CA rules lead to a phase
diagram topologically similar to the one obtained for the one-dimensional
case. The critical probability is Pc = 0.747. The order parameter critical
 6.2Simp le interac ting partic le systems 241
exponent f3 extracted by fitting the data with the relation 6.3 over the
=
range 0.001 ::::;; (Pc - p) ::::;; 0.02 is f3 0.52 ± 0.01.

6.2.2 The contac t process model (CPM)

The contact process model is in many respects similar to the A model.
However, the desorption mechanism is more subtle. In the A model, there
is an equiprobable desorption each time that at least one of the nearest
neighbors of a cell is empty. In the CPM, the probability of desorption
decreases when the number of nearest neighbors increases. The cellular
automata version of this model is defined by the following rules :
=
If 1 \fj ) ( t) 10) then :
=
{
1 \fj ) ( t + 1) l10A))', Wit with probability (1 - p),
. h prob ab·1· 1 1ty p. ( 6.41)
=
If 1 \fj ) ( t ) l A ) then:
=
1 \fj ) ( t + 1) { 10) ' with probability (1 - nj/z)q,
l A ) , w1t. h prob ab1" l"1ty (1 - (1 - nj I z ) q ) . ( 6.42)
where nj is the number of occupied nearest-neighbor cells of the cell j,
=
and z the coordination number of the lattice (z 2d for a d-dimensional
hypercubic lattice). Here again to restrict the number of independent
control parameters, we shall assume that q ( 1 - p) . =

The one-dimensional phase diagram is again topologically similar to

the one obtained for the A model (see figure 6. 1) . A second-order phase
=
transition occurs at Pc 0.28 1 and the order parameter critical exponent
=
is f3 0.260 ± 0.020. In two dimensions, the second-order phase transition
=
occurs at Pc 0.438 and the order parameter critical exponent f3 extracted
=
by fitting the data in the range 0.00 1 ::::;; (Pc - p) ::::;; 0.020 is f3 0.52 ± 0.04.
Several conclusions can be can drawn from the above results concerning
the possible differences between a fully parallel dynamics and a sequential
one. There are two categories of quantities. The first category contains the
quantities which are "universal" i.e. independent of the type of dynamics
used. This is the case of the critical exponents. The second category
contains the "nonuniversal", quantities, i.e. ones depending on the chosen
dynamics. The critical probabilities Pc are typical nonuniversal quantities
and their values for a given model may be quite different in the cellular
automata approach and the sequential updating. For example, for the A
model in one dimension, one has Pc = 0.465 with parallel updating while
sequential updating gives Pc 0.574 [190]. =
242 6 Nonequilibrium phase trans itions
6.3 Simple models of catalytic surfaces

In this section, we describe more complicated models related to the

important problem of heterogeneous catalysis.
6.3.1The Ziff model
The original Ziff model with sequential dynamics. One of the simplest
models describing the adsorption-dissociation-desorption process on a
catalytic surface has been introduced by Ziff et al. [192]. This model is
based upon some of the known steps of the reaction A - B2 on a catalyst
surface (for example CO - 02 ). The basic steps in heterogeneous catalysis
are the following:
• =
A gas mixture with concentrations XB2 of B2 and XA (1 - XB2 ) of
A sits above the surface and can be adsorbed. The surface can be
divided into elementary cells. Each cell can adsorb one atom only.
• The B species can be adsorbed only in the atomic form. A molecule
B2 approaching an empty cell will be dissociated into two B atoms
only if another cell adjacent to the first one is empty. Otherwise it is
rejected. The first two steps correspond to the reactions
A � A(ads) B2 � 2B(ads) (6.43)
• If two nearest neighbor cells are occupied by different species they
chemically react according to the reaction
A(ads) + B(ads) � AB (6.44)
and the product of the reaction is desorbed. In the example of the
CO - 0 2 reaction, the desorbed product is a C02 molecule.
This final desorption step is necessary for the product to be recovered and
for the catalyst to be regenerated. However, the gas above the surface
is assumed to be continually replenished by fresh material. Thus its
composition is constant during the whole evolution. The original Ziff
dynamics is sequential and defined as follows :
• A trial begins with the random collision of a gas molecule on a
square lattice, which represents the surface. The colliding molecule
is chosen to be A with probability XA and B2 with probability XBr
If the colliding molecule is A, then the following happens :
(1) A site on the lattice is randomly chosen.
(2) If the site is already occupied then the trail ends.
 6.3 Simp le models of catalytic surfaces 243
(3) Otherwise the A adsorbs and the four nearest neighbors are
checked in random order.
(4) If a B is present in any of the nearest-neighbor sites, a reaction
takes place between the central A particle and first B particle
found. A a result, both sites become vacant.
• If the colliding molecule is B2 , then the following happens:
(1) Two adjacent sites are randomly chosen on the surface.
(2) If either site is occupied, the trial ends.
(3) Otherwise, the B2 dissociates and adsorbs on the two sites as
two Bs.
(4) All the three nearest neighbors sites of both Bs are then checked
in random order for an A particle.
(5) The first A particle found in any of these two neighborhood
sites reacts with the corresponding B atom by vacating the two
sites.
It is found by numerical simulation [192] that a reactive steady state
occurs only in a window defined by
X1 < XA < X2
= =
where xl 0.389 ± 0.005 and x2 0.525 ± 0.001.
Outside this window, the steady state is a "poisoned" catalyst of pure
> >
A (XA X2 ) or pure B (XA < Xl). For XA X1 , the coverage fraction
varies continuously with XA and one can speak of a continuous (or second­
=
order) nonequilibrium phase transition. At XA X2 , the coverage fraction
varies discontinuously with XA and one can speak of a discontinuous (or
first-order) nonequilibrium phase transition.
The revis ited model with parallel dynamics . It is not clear from a physical
point of view if the dynamics of such a system is sequential or parallel.
It is then legitimate to study a similar model with fully parallel, cellular
automata dynamics.
In a simple approach [193], the cells of the automaton correspond to the
elementary cells of the catalyst. In order to model the different processes,
=
each cell j can be in four different states denoted lll'j) 10), lA), I B) or
I C).
The state 10 ) corresponds to an empty cell, lA) to a cell occupied by an
atom A, and I B) to a cell occupied by an atom B. The state I C) is artificial
and represents a precursor state describing the conditional occupation of
the cell by an atom B. Conditional means that during the next evolution
244 6 Nonequilibrium phase trans itions

Fig. 6.2. Typical microscopic configuration in the stationary state of the CA Ziff
model, where there is coexistence of the two species. This simulation corresponds
to the generalized model described in equation 6.53. The gray and black dots
represent, respectively, the A and B particles, while the empty sites are white. The
control parameter XA is larger in the right image than it is in the left.

1 0)
step of the automaton, I C) will become I B) or depending upon the
fact that a nearest neighbor cell is empty and ready to receive the second
B atom of the molecule B2 . This conditional state is necessary to describe
the dissociation of B2 molecules on the surface.
The time evolution of the CA is given by the following set of rules,
fixing the state of the cell j at time t + 1, llJlj)( t + 1), as a function of the
state of the cell j and its nearest neighbors (von Neumann neighborhood)
at time t:
= 10)
{
R1 : If l lJlj)( t) then
l A ) with probability XA
lJll j)( t + 1)= (6.4 5 )
I C) with probability (1 - XA)
=

{ 10)
R2 : If llJlj)( t) l A ) then
if at least one of the nearest neighbor cells
=
l lJlj)( t + 1) of j was in the state IE) at time t ( 6.46)
l A ) otherwise
=

{ 10)
R3 : If l lJlj)( t) IE) then
if at least one of the nearest neighbor cells
=
llJlj)( t + 1) of j was in the state l A ) at time t ( 6.47 )
I B) otherwise
 6.3 Simple models of catalytic surfaces 245

0.40

0.20

Fig. 6.3. Stationary state phase diagram in the simple CA model of the Ziff
problem. XA is the concentration of the A species injected on the surface. XA. and
XB are respectively the coverage fractions of A and B.

{
R4 : If l lfJj ) ( t ) IC) then
1 0) if none of the nearest neighbors was
=
l lfJj ) ( t + 1) in the state IC) at time t (6.48)
I B) otherwise
This last rule expresses the fact that the atoms of Bz can be adsorbed
only if they had been dissociated on two adjacent cells, that is if at least
two adjacent cells were empty at time t 1. Rules R1, R4 describe
-

the adsorption-dissociation mechanism while rules R2, R3 describe the

reaction-desorption process.
Figure 6.2 shows typical stationary configurations obtained with a cel­
=
lular automata version of the Ziff model. At time t 0, all the cells are
empty and a randomly prepared mixture of gases with fixed concentration
of one species (XA ) sits on top of the surface. The rules are iterated until
a stationary state is reached. The stationary state is a state for which the
mean coverage fractions XA. and X� of atoms of type A or B does not
change in time, although microscopically the configurations of the surface
changes.
As expected, it is found that the reactive steady state occurs only in
< < =
a window limited by X1 XA Xz, where X1 0.5761 ± 0.0004 and
=
X2 0.6515 ± 0.0004. The phase diagram is given in figure 6.3.
As anticipated, the limits of the window differ from those obtained
246 6 Nonequilibrium phase transitions

with sequential updating. However, there is also an important qualitative

difference. Both transitions are now of second order. This behavior is due
to the fact that the above CA rules do not reproduce exactly the physics
of the Ziff model. Indeed, a given cell occupied at time t by a B atom
can take part simultaneously in the formation of several AB pairs; the
situation is of course similar for the A atoms and the correct stoichiometry
is not obeyed here. In addition, the adsorption phase B2 2B is also not
-

correct microscopically: three adjacent B atoms can be adsorbed because

the same C can participate in the formation of two different pairs B - B.
This situation not allowed in the Ziff model where the B2 adsorption
phase is geometrically much more constrained than the adsorption of an
A molecule. The CA rules do not capture this asymmetric aspect of the
reaction.
These difficulties are intimately related to the fact that all the cells are
updated simultaneously in a CA and that a cell does not know completely
the current situation of its neighbors. In particular it does not know
whether or not these neighbors can also react with other particles in their
own neighborhood.
This drawback can be cured [194] by adding a vector field to every
site in the lattice, as shown in figure 6.4. A vector field is a collection
of arrows, one at each lattice site, that can point in any of the four
directions of the lattice. The directions of the arrows at each time step
are assigned randomly. Thus, a two-site process is carried out only on
those pairs of sites in which the arrows point toward each other (matching
nearest-neighbor pairs (MNN)). This concept of reacting matching pairs
is a general way to partition the parallel computation locally, very much
in the same spirit as in lattice gas automata where the interaction takes
place only among particles at the same site.
With reacting matching pairs, new CA rules can be defined as
=
R1 : If ltpj ) ( t ) 1 0) then
=
ltpj ) ( t + 1) { lA ) with probability XA

IC ) with probability (1 - XA)

(6.49)

=
R2 : If ltpj ) ( t ) lA ) then

=
ltpj ) ( t + 1) { IO) if the MNN
of j was in the state I B ) at time t
lA ) otherwise
(6.50)

=
R3 : If ltpj ) ( t ) I B) then

=
ltpj ) ( t + 1) { IO) if the MNN
of j was in the state lA ) at time t
I B) otherwise
(6.51)
 6.3 Simple models of catalytic surfaces 247

Fig. 6.4. Illustration of the modified rules R2 and R3. The arrows select which
neigbor is considered for a reaction. Dark and white particles represent the A
and B species, respectively. The shaded region correspond to cells that are not
relevant to the present discussion such as, for instance, cells occupied by the
intermediate C species.

Fig. 6.5. Illustration of the modified rule R4. Light gray particles represent
the intermediate C species and the arrows indicate how the matching pair is
done. The shaded region corresponds to cells that are not relevant to the present
discussion such as, for instance, cells occupied by A or B particles. White cells
are empty cells.

R4 : If l1pJ ) ( t ) = I C ) then
11pj ) ( t + 1) =
{ I1 0)B )
if MNN is in the state I C ) at time t
otherwise (6.52)
The above rules R2 and R3 are illustrated in figures 6.4 and 6.5.
The simulations made on this model [194] reproduce the correct orders
of the transitions (second and first order), but the limits of the reactive
window are respectively 0.0495 and 0.0546, illustrating once more the
nonuniversal character of these quantities.
248 6 Nonequilibrium phase transitions

6.3.2 More complicated models

The models described above are the simplest possible and they may not
be rich enough to describe the physical reality. Many new features can be
added and again described by CA rules.
Let us first consider the A model again. One natural generalization is to
assume that once adsorbed, the A particles can diffuse on the surface. What
will be the effect of this diffusion? This question has been investigated by
Bagnoli et al. [195]. The dynamics alternates the adsorption-desorption
rule discussed above with the diffusion rule described in section 4.2.
By changing the relative frequency of the diffusion and adsorption­
desorption phases, the diffusion constant D can be tuned and different
regimes can be investigated. As expected from physical grounds, the
value of the critical probability Pc increases monotonically with D. In the
limit D � oo, one recovers the mean-field value. The situation is more
complicated for the critical exponents and will be discussed in the next
section.
Let us now return to the two-components Ziff model. More complicated
dynamics can be considered. An empty site can remain empty with some
probability. One has then two control parameters to play with: XA and
XB2 that are the arrival probability of an A and a B2 molecule, repectively.

{
For example rule 6.49 can be modified as follows :
If l lJJj ) ( t) 1 0) then
=

l A ) with probability XA
llJJj ) ( t + 1)
= I C ) with probability XB2 (6.53)
1 0) with probability 1 - XA - XB2
In addition, since empty sites may now be present on the lattice during the
adsorption B2 2B a generalization of rule 6.52 can be considered: a cell
�

in the intermediate state C will give two adjacent B atoms if its matching
arrow points to an empty site which is not pointed to by another C state.
This new rule is illustrated in figure 6.6.
The phase diagram obtained for this generalized CA Ziff model is given
in figure 6.7, with the value XB2 0.1. Here again, the phase diagram is
=

topologically similar to the case XB2 1 - XA, but the locations of the
=

critical points are different.

The influence of several other mechanisms on the location of the active
window have been investigated. Besides diffusion, the role of adatom­
adatom interactions [196], desorption activation energy and precursor
adsorption have been considered [197-199]. By tuning the parameters
controlling the different effects, it is possible to obtain reasonable quanti­
tative agreement with experimental situations.
 6.4 Critical behavior 249

(a) (b)

Fig. 6.6. A less constrained adsorption rule B2 - 2B when some cells may be
empty. Here the B 2 molecule (or C state) is represented as two disks on top
of each other. Dissociation is possible if the upper disk can move to the site
indicated by the arrow without conflict with other moves. Note that we still
accept dissociation when a pair of matching Cs sit next to each other.

o B coverage

• A coverage

o �����--����
0.04 0.06

Fig. 6.7. Stationary state phase diagram corresponding to the CA Ziff model
given by relation 6.53 and the generalized dissociation rule.

6.4 Critical behavior

As we have seen at the beginning of this chapter, a problem of great

interest is the study of the critical behavior at a second-order phase
transition. We shall briefly review some methods which can be used to
extract the critical exponents for nonequilibrium systems from numerical
simulations.
250 6 Nonequilibrium phase transitions

6.4.1 Localization of the critical point

Good determination of the critical exponents requires precise knowledge
of the critical point. Due to finite size effects, this may be a difficult task.
Several approaches minimizing finite size effects have been devised. We
shall illustrate two of them in the context of the A model.
The first one was introduced by Grassberger for directed percolation
[200] . One starts with a lattice filled with A particles except for one single
site which is empty, and let the system evolve according to the CA rule.
One measures the density of empty sites n0 ( t ) , the average cluster size R2 ( t )
and the survival probability P ( t ) , that is the probability that the lattice is
not poisoned by A particles after t iterations. If the system is sufficiently
large that none of the clusters we simulate reaches the boundaries, the
data are free of finite size effects. At the critical probability Pc , all the
above quantities behave as a power law :
P ( t ) "' t -� (6.54)

no ( t ) "' t11 (6.55)

R2 ( t ) "' tz (6.56)
If one is not precisely at criticality, these quantities will have an upward
(downward) curvature in a log-log plot. Thus this method leads to a
precise determination of Pc·
Another approach is based on the concept of damage spreading [201] .
Here again, we discuss this method in the context of the A model.
Let us consider a system with N sites and two different configurations
A = { tp f } and B = { tpf } of this system. The Hamming distance between
these two configurations is defined as

DAB = 1 "' A B
N � llp i - lpi I
I
(6.57)

At time t = 0, one prepares two copies of the system such that

lpjA = lpjB I.f }. 0 _J_
T (6.58)
and
(6.59)
Moreover, one imposes that this last condition remains true for all time.
The initial Hamming distance is thus DAB ( t = 0) = -k· One then starts
the evolution in such a way that these two copies have strictly the same
dynamics, defined by the cellular automata rules for the A model discussed
previously. Locally, two types of damage can, in principle, occur.
At site i, tp f = 1 and tpf = 0 ; this may occur with a probability d f 0 .
 6.4 Critical behavior 251

At site i, tp f = 0 and tpf = 1 ; this may occur with a probability J[ 1 .

Let us now restrict ourselves to the stationary state and denote by
square brackets the time average in the stationary state. We have :
(6.60)
and
(6.6 1 )
thus
(6.62)
Let us now assume that we are not in the poisoned phase, i.e. p < P c ·
The dynamical process is not frozen in. Assuming ergodicity, we can
express the time averages in 6.62 by ensembles averages (denoted by ( .. ) )
on a nonconstrained nonequilibrium ensemble. One obtains :

[tp:4 ] ( tptpo tp)i)

= (6.63)
l ( o
[tp � ] ( (1 1--tptpo ) tp) i)
1
= (6.64)
( o
Indeed, the presence of the tpo insures that only configurations such that
=
tp� 1 and tpg = 0 are retained. Introducing the reduced correlation
function
(6.65)
and assuming translational invariance,
(6.66)
one has
r· - cOi
I - XA (1 - XA ) (6.67)

The probability of finding site i damaged is obviously D i = df 0 + d? 1 .

It is now important to notice that d? 1 = 0. Indeed, the probability of
a site i being occupied is always larger for the A configuration than for
the B one. Thus it is not possible that a given time tpf = 0 while tpf = 1 ,
hence D i = i . Introducing the total damage (which is just the Hamming
r

distance) as

(6.68)
252 6 Nonequilibrium phase transitions

one finds, in term of the nonequilibrium susceptibility defined as

x= L: coi ( 6.69 )

1 X
DAB = - --'-'--- (6.70)
( )
N XA 1 - XA
Thus the Hamming distance is exactly related to the nonequilibrium
one-point and two-point correlation functions, XA and Coi · In particular,
when p approaches Pc , the Hamming distance diverges allowing a precise
determination of Pc ·

6.4.2 Critical exponents and universality classes

The determination of the critical exponents presents its own difficulties.
Let us consider as an example the critical behavior of the A model with
diffusion. The behavior of the order parameter exponent f3 as a function
of the diffusion coefficient D has been studied by Bagnoli et al. [195] .
Simulation shows an apparent dependence of f3 upon D. Two limit cases
are clear. When D = 0, one finds the directed percolation value while,
when D � oo, one finds the mean-field value, f3 � 1. What is happening
in between ? The apparent D dependence does not mean that the true
critical exponent is continuously varying with D. Indeed, the presence of a
lattice of finite size prevents us going arbitrarily close to Pc · As D becomes
larger, the width of the critical region may shrink, and it is more and more
difficult to probe the real critical regime. In fact renormalization group
arguments support the following scenario. For all finite values of D, the
true critical behavior is one of directed percolation, while for D = oo one
is in the mean-field regime. The continuously varying exponent measured
reflects the crossover between the two limit cases. This situation can be
described in terms of an effective exponent (see [195] for more details).
This example shows that it may be difficult to extract an exponent from
the numerical data if one does not have a clear theoretical idea of what
is happening.
Similar studies can be made for the two-species models and in these
cases one also finds that the critical exponents belong to the directed
percolation universality class. This is not the place to discuss in detail
the attributes of the directed percolation universality class. A thourough
analysis of this question is given in the review article by Kinzel [202] .
This universality class appears to be very large and robust and one can
wonder if there exist some systems belonging to a different class.
Such models indeed exist and Grassberger et al. have shown [203] that
simple Wolfram's rules with very special added noise belong to a different
 6.4 Critical behavior 253

II
Fig. 6.8. Behavior of probabilistic rule 94 for p=O.O, 0.05, 0.25 and 1 .0, respec­
tively. A kink corresponds to two nearest neighbors black sites while an anti-kink
corresponds to two nearest-neighbor white sites.

universality class from directed percolation. Let us consider for example

the Wolfram's rule number 94
[( 1 1 1 ) ; ( 1 0 1 ) ; (010) ; ( 1 00) ; (00 1 ) ; (000)] � [0 ; 0 ; 1 ; 1 ; 1 ; 0] (6.7 1 )
and add the following noise :

[(01 1 ) ; ( 1 10)] � { [0 ; 0]
[1 ; 1]
with probability p
with probability ( 1 - p )
(6.72)

For very small p, the system orders itself spontaneously : there are two
symmetric adsorbing states. The corresponding patterns are, respectively,
vertical stripes with lp i = 0, lp i+ l = 1 . Starting from a random initial state,
after some short transients one finds small domains separated by doubly
occupied neighbors called kinks. For p = 0, the kinks are frozen and
between two kinks one observes periodic patterns with very short periods.
For small p the kinks are no longer frozen. They move by annihilating
random walks and the kink density decays in time as (pt ) 1 1 2 . If p grows
further, in addition to enhancing the random walk it leads to a splitting
of the kinks according to : kink� kink+ [kink+antikink] . For small p, the
annihilation procedure dominates but if p is larger than a critical value
Pc = 0. 1 3 ± 0.02, a single kink in the initial state is enough to create a
completely disordered state. Note that in the limit p = 1 , the rule is simply
Wolfram's rule 22 which has a chaotic behavior. These different situations
are illustrated on figure 6.8.
At p = Pc, the density of kinks decreases as
nkink ,...., crx, rx = 0.27 ± 0.08 (6.73)
and for p � Pc from above, the density of kinks in the stationary state
tends to zero as
nkink "' (p - pc)f1, f3 = 0.6 ± 0.2 (6.74)
These exponents are not those of directed percolation.
254 6 Nonequilibrium phase transitions

This cellular automata model is in fact closely related to the more gen­
eral problem of the branching annihila ting random walk (BAR W) [204] .
In the BARW, particles on a lattice are subject to the following sequen­
tial dynamics. First, one chooses one particle randomly. With probability
q this particle jumps to a nearest-neighbor site. If the new site is already
occupied by one particle, both particles annihilate. Secondly, with proba­
bility ( 1 q) the particle gives birth to n offspring distributed randomly
-

on the nearest-neighbor sites. In this problem there is a phase transition

in the stationary states for q = qc between the empty phase and a phase
with a non-zero particle density. Two cases have to be distinguished : if
n is odd the exponents associated to the phase transition are those of
directed percolation while if n is even, the exponents are those of the
cellular automata described above [205] .
The physical reason for this difference is simple. The even n case is
special as the particle number is conserved modulo 2. This means that
the absorbing state exists only if we start out with an even number of
particles.
From the numerous models which have been studied, the following rule
seems to be correct. Only systems possessing a unique absorbing state,
or systems with many absorbing states such that, with probability one,
the system falls into one of the many available absorbing states, belong
to the directed percolation universality class [206] . This explains why
the directed percolation universality class is large but also why there are
exceptions.

6.5 Problems

6. 1 . Cons i d e r a var i ation of the A model i ntrod uced i n section 6.2 . 1 I n

th i s n ew vers i o n , the parti c l es are adsorbed i n pai rs at vacant nearest­
n e i g h bor s i tes . Study the p h ase d i ag ra m of t h i s model i n mean-f i e l d-l i ke
theory and by n u me r i ca l CA s i m u l ations.
6.2. Cons i d e r the CA i ntrod uced in section 6.3. 1 fo r the Ziff mode l .
The ru l es a re mod i f i ed as fo l l ows : a s i te occ u p i ed b y a B ato m
can also r.eact with an A atom arrivi n g o n t h i s s ite. Study the p h ase
d i ag ram of th is mod e l i n mean-fi eld-l i ke theory and by ce l l u l ar automata
s i m u l ations.
6.3. Study the k i n etics of the Z i ff model ( i ntrod uced in section 6.3. 1 )
with a sequential dynamics . Com pare with the s i te and pa i r mean-fi e l d
approx i m ati o n .
 6.5 Problems 255

6.4. Cons i d e r a n exte n s i o n of the model A i ntrod u ced in section 6.2. 1 .

O nce the A atoms are adsorbed o n the s u rface , they can d i ffuse. Study
the ro l e of d iffu s i o n o n the l ocation of the critical th res h o l d Pc and on
the critical expon e nt {3.
6.5. D r i ve n-d iffusive syste ms provide s i m p l e exa m p l es of noneq u i l i b­
ri u m p rob l e m s . Co n s i d e r a one-d i mensional l attice of N s i tes with
period ic bou ndary co n d itions and such that nearest-n e i g h bo r particles
are s u bj ect to a n attractive i nte ract i o n . At each s ite, o n e can have
at most one part i c l e . I n eq u i l i b ri u m , the particles can hop to em pty
nea rest- n e i g h bo r s i tes with a rate co mpati b l e with the fact that the
particles are in thermal eq u i l i b r i u m with a h eat bath at tem pe ratu re T .
The syste m can be put out of eq u i l i b r i u m by add i n g an external f i e l d
favori n g the hopp i ng o f the particles i n the d i rection o f the f i e l d .
( i ) Write the m aste r equation gove rn i ng the p robab i l i ty P (O., t)
that t h e
conf i g u ration 0. i s rea l ized a t t i m e t ( w i t h and without the e l ectric f i e l d ) .
( i i ) I ntrod uce a p robab i l istic C A w h i c h s i m u l ates the above dynam ics.
( i i i ) Compute fo r d iffe rent s i tu ati ons the H am m i ng d istance hAB ( t) at
t
t i m e betwee n two conf i g u rations o.A
and o.B .
The H am m i n g d istance
is d efi ned as the s u m of the l ocal d i fferences:
1 � A B
hAB ( t) =
� IJi ( t ) EB IJi ( t )
N i=l
where IJ i represents the state of s i te i and EB stands fo r the exc l usive-o r
operation ( h i nt: see [207] ) .
6.6. Vote r mod e l : co n s i d e r an abitrary l attice a n d ass u m e that each
s ite i s occ u p i ed by a vote r who m ay h ave one of two opi n i o ns labeled
+ and - . Each s i te keeps its o p i n i o n d u ri n g some t i m e i nterval ,
expo nenti a l l y d i stri b uted , and then assu m es the o p i n i o n of a rando m ly
chosen n e i g h bo r i n g s ite . Let Pn ( t)
the fraction of vote rs who change
n
th e i r op i n i o n t i m es u p to t i m e t.
Co m pute and Pn ( t) (n)
the ave rage
n u m be r of o p i n i o n chan ges in the mean-f i e l d approx i m ati o n .
6 . 7 . Write a C A ru l e w h i c h s i m u l ates t h e vote r model descri bed i n the
p revious p rob l e m . I nvestigate the one-, two- and th ree-d i mensional
cases and meas u re ( n ( t)) ,
the ave rage n u m be r of op i n i o n changes u p
t o t i m e t.
 7
Other models and applications

This chapter presents a few more applications of the cellular automata

and lattice Boltzmann techniques. We introduce some new ideas and
models that have not been discussed in detail previously and which can
give useful hints on how to address different problems.
We shall first discuss a lattice BGK model for wave propagation in
a heterogeneous media and show how it can be applied to simulate a
fracture process or make predictions for radio wave propagation inside
a city. Second, we shall present how van der Waals and gravity forces
can be included in an FHP cellular automata fluid in order to simulate
the spreading of a liquid droplet on a wetting substrate. Then, we shall
define a multiparticle fluid with a collision operator inspired by the lattice
BGK method in order to avoid numerical instabilities and re-introduce
fluctuations in a natural way. Finally, we shall explain how particles in
suspension can be transported and eroded by a fluid flow and deposited
on the ground. Snowdrift and sand dunes formation is a typical domain
where this last model is applicable.

7.1 Wave propagation

7. 1 . 1 One-dimensional waves
In this book, we have already encountered one-dimensional wave propa­
gation. The chains of particles, or strings, discussed in section 2.2.9 move
because of their internal shrinking and stretching. From a more physical
point of view, this internal motion is mediated by longitudinal backward
and forward deformation waves. This is easy to see from equation 2.46
q l ( t - 1 ) - 2q l (t) + q l ( t + 1 ) 2(q2 ( t) - q 1 ( t) - a )
qi( t - 1 ) - 2qi( t) + qi( t + 1 ) q i- l ( t) - 2qi (t) + q i+l( t)
qN (t - 1 ) - 2qN ( t) + qN (t + 1 ) 2(qN-l ( t) - qN ( t) + a )
256
 7.1 Wave propagation 257

_ .,.__ .,.__ _

Fig. 7. 1 . The string motion i s due to forward and backward traveling waves f t
and h

where q i ( t ) is the location along the x-axis of particle i in the chain and
a the equilibrium separation between consecutive particles. Now if we
define Ll i as the deformation from the equilibrium distance
Ll i = q i+ l - q i - a
we obtain the following equations of motion
Ll t ( t + 1 ) - 2c5 t ( t ) + Ll t ( t - 1 )
L'l 2 ( t ) - 3 L'l t ( t ) ( 7. 1 )
Lli ( t + 1 ) - 2L'l i ( t ) + Lli ( t - 1 )
Lli -t ( t ) - 2 L'li ( t ) + Lli + l (t) (7.2)
Ll N-l ( t + 1 ) - 2Ll N-! ( t ) + LlN-! ( t - 1 )
Ll N_2 ( t ) - 3 LlN-l ( t ) ( 7.3 )
Disregarding momentarily equations 7. 1 and 7.3, it is interesting to note
that equation 7.2 admits a very simple set of solutions
Ll i ( t ) = f ( i - t) + g ( i + t ) ( 7.4 )
as can be seen by direct substitution. The functions f and g can be chosen
arbitrarily and correspond to a backward and forward traveling wave.
As we shall see, equations 7. 1 and 7.3 impose reflection and transmission
boundary conditions on these longitudinal deformation waves.
Now, let us return to the interpretation of the motion of our string
particles, as presented in section 2.2.9. Black and white particles alternate
along a chain and each type of particle moves at every other time step
as if they were linked by a spring. Suppose we define ft ( i, t ) as the
deformation on the left of a moving particle (say a black one) and h ( i, t )
as the deformation on its right
ft ( i, t ) qi - qi-1 - a
h ( i, t ) q i+ l - q i - a
for i odd or even depending on which kind of particle is going to move.
According to the rule of motion, our black particle moves so as to
exchange ft and h, as shown in figure 7. 1. Then, at the next iteration, ft
and h are no longer defined with respect to the black particle that just
moved, but with respect to the white particles around it.
This choice of defining ft and h as the left and right deformation,
respectively, is less general than the solution 7.4. This is a particular
258 7 Other models and applications

case where f1 is zero when h is not and vice versa. But this choice
corresponds to the actual motion of the string and leads to the same kind
of formulation as found in lattice gas automata
( h ((ii +- 11,' tt ++ 11 )) ) = ( 1 1 ) ( h ((i,i, tt)) )
h
0
0 h

With this formulation, it is obvious that the string deformation is mediated

by a forward and a backward wave h and /2 , respectively.
The boundary conditions 7. 1 and 7.3 can be easily expressed in terms
of the fi s. The particles at the right (or left) extremity of the string move
so as to take a symmetrical position with respect to qN_ 1 + a (or q2 - a ) .
Thus we have for the right end
and qN ( t + 1 ) = qN_ 1 ( t ) + a - f l ( N, t )
with, also qN_ 1 ( t + 1 ) = qN_ 1 ( t ) since the next-to-the-end particle does not
move when the last one does. Therefore :
h ( N - 1, t + 1 ) qN ( t + 1 ) - qN _ 1 ( t + 1 ) - a
qN_1 ( t ) + a - f l ( N, t ) - qN_ 1 ( t ) - a
-f l ( N, t )
A similar result holds for the left end of the string. The interpretation is
quite simple : at the extremity of the string, the forward wave is reflected
back with opposite sign and so is the backward wave at the left string
end.
Therefore, the evolution rule for the deformation of the string can be
written as
( f/2l ((ii +- 1,1, tt ++ 11 )) ) = ( RT RT ) ( hfl ((i,i, tt)) ) (7.5)

where T = 1, R = 0 inside the string and T = 0, R = - 1 at the free

ends. Another interesting case is when the right string extremity is fixed
and cannot move. Then, clearly f2 ( N - 1, t + 1 ) = f 1 ( N, t ) and R should
be set to R = + 1.

7. 1 .2 Two-dimensional waves
The problem of modeling two-dimensional waves is more difficult because
waves spread as they propagate in space. Thus, a representation of
the deformation in terms of a field of integer values is in general not
possible. However, the construction proposed in the previous section can
be generalized provided that the fi s are real-valued.
Let us assume we have a two-dimensional lattice of black and white
particles organized in a checkerboard fashion. Figure 7.2 shows a basic
 7. 1 Wave propagation 259

Fig. 7.2. Alternate motion of black and white particles in a two-dimensional

system. The quantities fi , h, h, and !4 measure the distance, along the x­
axis, separating the central particle from its four connected neighbors. The cross
indicates the location of the center of mass of the white particles. Other quantities
gis can be defined to describe the separation along the vertical y-axis.

element of this lattice, made of a central black particle and its four white
connecting neighbors.
As in the one-dimensional case, the black particles move at even time
steps and the white ones at odd time steps. In figure 7.2, the motion is
relative to the positions of the four connected white neighbors at rest.
The rule of motion for the black particles is to j ump to the symmetrical
position with respect to the center of mass (the cross in the figure) of the
four white surrounding particles.
Let us denote the location of the black particle by
ri,J = ( x i,J , Yi,J )
The surrounding white particles will be at positions r i-l,J , r i+l,J , ri,J-1 and
ri,J+l · We define their x-axis separation to the central black particle as
f l ( i, j, t )
X i-l,J ( t ) - X i,j ( t )
= /3 ( i, j, t ) X i+l,j ( t ) - X i,j ( t )
=

X i,j-l ( t ) - X i,j ( t
f2 ( i, j, t ) = ) j4 ( i, j, t ) Xi,J+l ( t ) - X i,j ( t ) (7.6)
=

Similarly, we can define g1 , g2 ,g 3 , and g4 as the distance, along the y-axis,

separating the white particles from the black ones.
If the origin of the coordinate system is the black particle (that is
ri 1,J ( t ) = 0), the x-coordinate Xe M of the center of mass of the four white
particles is
1
XcM = 4 (fl + f2 + h + f4 ]
The new position of the black particle is then
1
X i ,j ( t + 1 ) = 2XcM = 2: (f l + f2 + h + /4 ] (7.7)
260 7 Other models and applications

because the dynamics is to move to the symmetrical position with respect

to the center of mass.
The motion of the black particle results in a new distribution of the
·
fi s. Since at time t + 1 the white particles will move and the black
will be stationary, the fi ( t + 1 )s have now to be computed with respect
to each white particle location. For instance, f i ( i + 1, j, t + 1 ) will be
the separation to the black particle as seen by the particle located at
r I· +l ,]·
fi ( i + 1, j, t + 1 ) - 2XcM - X i +l,j
1
2 [f l + h + h + !4 1 - h
1
- 2 [f l + h - h + /4 ] (7.8)

Similarly, h ( t + 1 ), h ( t + 1 ) and j4 ( t + 1) can be computed for each white

( ) ( )
particle in figure 7.2. Thus we obtain the following evolution law for the
x-separation between black and white particles
fi ( i + 1, j, t + 1) f i ( i, j, t )
h ( i, j + 1, t + 1 ) w h ( i, j, t )
h ( i - 1, j, t + 1 ) = h( i, j, t ) (7.9)

( )
f4 ( i, j + 1, t + 1 ) f4 ( i, j, t)
where the propagation matrix W reads
1 1 -1 1
1 1 1 1 -1
Wjree = 2 (7. 10)
-1 1 1 1
1 -1 1 1
where the subscript "free" indicates propagation in a medium with neither
reflection nor absorption.
It is important to notice that the above dynamics requires to work with
real-valued quantities /;. In other words, by following our motion rule 7.7,
the white and black particles will eventually jump off lattice sites. Thus,
equation 7.9 is no longer a pure cellular automata rule as it was in the
one-dimensional case but, rather, a lattice Boltzmann dynamics. Figure 7.3
illustrates this dynamics.
Equation 7.9 with W given by 7.10 leads to the propagation of two­
dimensional waves. This will be proven below but, at this point, it should
be clear from the physical interpretation we gave : masses are vibrating
around their local equilibrium positions, and the deformation propagates
like a wave. Note that equation 7.9 concerns only the deformations along
the x-axis. This is enough to simulate two-dimensional wave propagation
since an x-deformation also propagates along the y-axis.
 7. 1 Wave propagation 26 1

fl - �
, - -�·� · -
----- -<---

2
_l

Fig. 7.3. Scattering of an incoming flux h at a node. The value of each outgoing
flux is indicated on the arrows.

However, to describe the overall motion of our system of white and

black particles (the two-dimensional counterpart of the string rule dis­
cussed in section 2.2.9), it would be necessary to include the y-motion in
terms of the other fields g i s. Then we could simulate the motion of a solid
body made of many particles whose own internal movement mediates the
large-scale motion of the whole system. Note that the interaction we have
considered between black and white particles does not correspond exactly
to that of a regular spring : we have here completely decoupled the x and
y motions, which is not the case with a spring where the force depends
on the Euclidean distance between the pairs of particles. However, our
approach captures the essential features of the system, with fewer compli­
cations. Figure 7.4 shows an example of the motion of such a solid body
bouncing off the wall of a container.
Using this interpretation of local internal deformation, a possible appli­
cation of this model is to simulate dynamic fracture formation and cracks
in a solid material. This problem is important because the way things
break is still not well understood [208] and numerical simulations may
help to develop a better intuition.
A typical experiment is to apply an initial stress to the system, for
instance on the left and right sides of the sample. This gives non-zero
initial values to the fi s and gi s at the boundaries. Then, a deformation
wave will propagate and, if the value of a given fi('r, t ) exceeds some pre­
assigned threshold at some site r, we may assume that the corresponding
bond breaks. The following dynamics of that site is then governed by
three neighbors instead of four. In the present model, when a bond is
broken, one adds a virtual particle at a distance 2ro from the next-to­
the-last particle [209] so that the same four-neighbor rule can be applied.
The quantity r0 is the equilibrium separation between adjacent "atoms".
262 7 Other models and applications

t= IO t=20 t=30

t=40 t=50

Fig. 7.4. Motion of a deformable object made of many particles interacting

according to the lattice Boltzmann wave rule. An initial momentum is given to
the object by choosing the values of the quantities f; and g;. The propagation
of this initial deformation produces an overall motion. The center of mass of
the object follows the straight line shown on the pictures until its boundary
reaches the limit of the simulation space. By forbidding the particles to cross the
boundary, the entire object naturally bounces on the wall.

t=l t= ll t=2 1 t=23 t=25

Fig. 7.5. Dislocation of a two-dimensional media obeying the lattice Boltzmann

wave model. The sample shown in this figure is subject to an external force
pulling the left and right side in opposite directions. The breaking threshold is
chosen so that it is more difficult to break a vertical bond than a horizontal one.

It differs for the horizontal and vertical coupling (it is zero in the latter
case, for the x-motion).
A bond breaks if one of the f; (or g; ) becomes too large. Each time a
bond breaks a virtual particle is added. Figure 7.5 illustrates the various
stages of this process. Of course, it is possible to distribute the breaking
thresholds randomly among the bonds, so as to produce a disordered
media.
Some modifications to the model can be considered. For instance,
instead of comparing the distance f; directly with the breaking threshold,
 7.1 Wave propagation 263

(a) (b) (c) (d)

Fig. 7.6. Different patterns of fracture obtained with our model. Dissipation is
an important parameter to obtain a smooth fracture, as in (d) rather than many
micro-fractures and branching as in (b).

0
'-'
-..
,.-.,
"0
0
0
0..
"'
�
()
«!
....
()
'-'

0
0 1 000
time

Fig. 7.7. Propagation speed of the factures shown in figures 7.6 (b) and (d)
(upper and lower curve, respectively).

it is more natural to consider the deformation fi - ro. It is also possible

to add some attenuation in the vibration (the way to do this is explained
in the next paragraph). These parameters play an important role in the
fracture behavior. Also the stress rate and the amount of disorder modify
quantitatively the shape of the broken region. Figure 7.6 illustrates several
patterns obtained with a breaking threshold based on fi - ro.
The fracture propagation speed, as measured from the simulation, is
shown in figure 7.7. It may be seen that propagation is slower than
sound speed c 0 = 1/ .Ji and that the fracture advances faster in the case
where branching occurs. These facts are in agreement with experimental
data.
264 7 Other models and applications

Note finally that another version of this kind of fracture model IS

discussed in [210] , though with a more complicated dynamics.

Reflection, absorption and source. Let us now return to the modeling

of wave propagation. Equation 7. 10 describes the dynamics inside a
homogeneous media. In general, a system contains boundaries or obstacles
at which waves are reflected. Following what we did for one-dimensional
waves at the end of section 7. 1 . 1 , reflection at a boundary site can easily

( )
be implemented by bouncing back the incoming flux with the same or the
opposite sign. Thus we modify the free propagation matrix 7. 10 as follows
0 0 ±1 0
+ 0 0 0 ±1
W;;f = ±1 0 0 0
(7. 1 1 )
0 ±1 0 0
where the sign ± 1 depends whether or not there is a change of phase
during the reflection step. A positive value corresponds to a free end while
a negative one corresponds to an absorbing boundary.
With this approach, we can simulate free wave propagation and total
reflection at obstacles. The lattice sites that represent obstacles are charac­
terized by a "scattering" matrix wrt ' while free sites are evolved according
to the scattering matrix Wf ree· In the next section, we will return to the
general problem of modeling a change of medium and different speeds
of propagation in each of them. For the time being let us explain how
attenuated or damped waves can be modeled.
In the interpretation of black and white moving particles, attenuated
waves may be seen as resulting from a particle movement with friction :
the displacement is amortized and the moving particles jump closer to
their initial position than they would otherwise do. Therefore, similarly to
what we did in equation 7.7 we shall now write

X i,j ( t + 1 ) = X i,j ( t ) + 0: (XcM - X i,j ( t )) (7. 1 2)

where o: is some real coefficient between 0 and 2 describing the amount
of friction. Particular values are : o: = 0 meaning that the particle has not
moved at all, o: = 1 meaning it has moved to the center of mass and o: = 2
corresponding to free motion. Thus, in an arbitrary coordinate system, we
have
fl ( t + 1 ) X i,j ( t + 1 ) - X i+l,j
o:XcM + ( 1 - o: ) x i,j - X i+ 1 ,j
0:
- ( x·l - 1 ,j· + x l·+ 1 ,j· + x·l,J·- 1 + x·l,j·+ 1 ) + ( 1 - o: ) x·l,j· - x·l + I ,j
·

4
 7.1 Wave propagation 265

( � -1 ) (x ·+1 · - x· · ) + � (x ·-1 · - x· · ) + � (x· ·-1 - x· · )

4 1 ,] 1,] 4 1 ,] 1,] 4 1,] 1,]
0(
+ 4 ( x;,j + 1 - x;,j )
- �h + �h + (� -1 ) h + � !4

( )
Repeating the same calculation for the other f;s yields the attenuation
propagation matrix

Watt(oc) =
!4 ; �;1 ��1 ( 7. 1 3 )
ex ex ex ex
4 -1 4 4 4
� �-1 � �
Clearly, when oc = 0, Watt = Wr--;f and when oc = 2, Watt = Wtree · In the
general case, Watt can be written as
0( 0(
Watt = Wf ree + ( 1- 2 ) Wr--;f
2
Damped waves are very useful to mimic an infinite system with a finite
lattice : by smoothly absorbing the waves at the boundary of the lattice, it
is possible to prevent them from coming back into the system as if they
had freely continued their propagation beyond the limits of the simulation.
For the sake of completeness, we also give here the expression for
Watt for a one-dimensional system, which can be obtained with a similar
derivation.

1D : Watt(oc) = � .:_ 1 2 �
� (
�- 1 )
( 7. 1 4 )

Another important ingredient to include in our lattice wave model is

the presence of wave sources such as an emitting antenna. For instance a
lattice site can excite a sine oscillation by forcing the time dependence of
the fi to follow some given function of time (a sine in this case). Similarly,
plane waves can be produced by having an entire line of sites oscillating
in phase.
As an illustration of our wave model, we show, in figure 7.8, the result
of a simulation of a plane wave being reflected by a parabolic mirror.
Observe that the wave naturally converges toward the focal point of the
mirror, as a result of the collective effect of the reflector sites organized
as a parabola.
Finally, let us mention that the propagation matrix 7. 10 has been
proposed and re-invented by several authors for different applications.
Probably the first domain of application consisted of actual circuits or­
ganized as a two-dimensional lattice. Tension and current were used to
266 7 Other models and applications

Fig. 7.8. Lattice Bolztmann wave model and focusing of a plane wave by a
parabolic mirror.

study wave propagation in a kind of non-digital computer. This tech­

nique is widely known to the community studying wave guides and has
received the name transmission line matrix (TLM) [21 1 ] . More recent
works related to this wave propagation model include [212-214] .

The dispersion relation. By using A and r as the lattice spacing and time
step respectively, equation 7.9 can be rewritten as
4
fi ('r + TC j, t + r) = L Wi,Jfj('r, t )
J =l
where W is given by equation 7.10. The dispersion relation of this lattice
Bolztmann wave model can be obtained by considering a solution of the
form
fi ( r , t ) = A i e i(w t+kr )
After substitution, it is found that the above expressiOn 1s a solution
provided that w and k satisfy

e iwr = �( K ± i �4 - K2 ) ( 7. 1 5 )
where K = cos(k l A)+cos(kzA) and k1 and kz are the two spatial components
of k. A Taylor expansion of 7. 1 5 gives, with k 2 = k y + k�
w 2 r2
2
A2k 2 Ak
iwr - -- + O(r 3 ) = - 4 + 0 (A4 ) ± i J2 + 0 (A 3 )
[ ]
( 7. 16 )

In view of taking the limit A � 0 and r � 0 of this equation, it is

interesting to consider the multiscale formalism introduced in section 3.2.3.
The question is always to know how A and r are related to each other
when they tend to zero. In principle, one assumes that both convective
 7. 1 Wave propagation 267

and diffusive phenomena are present. In other words r is expected to have

components of the order A and also A2 • Thus, we shall write

where E is the parameter which will tend to zero.

We can now match the various terms of equation 7. 1 6 according to
their order of magnitude. We obtain
1 A1
O(e) : w=-
+ - -k
J2 r 1
which is exactly the dispersion relation expected for wave propagation
with speed
1 1. A
J2 ,___.o r
co = - Im -

Similarly, for the second order in E, we obtain

. z - w2 Ayk 2
zwr - r21 = - --
2 4
Since we have just obtained that w 2 ry ( Ayk 2 /2 ) , we find that
A k 2 Ai k 2 0
iwr z = - i + =
4 4
Hence, the dissipative term vanishes at the order E 2 and we are left with
pure wave propagation. Of course, in finite systems, corrections of higher
order may show up.

7. 1 . 3 The lattice BGK formulation of the wave model

Here we shall show that the lattice Boltzmann wave model introduced in
the previous sections can be formulated as a lattice BGK equation, very
much in the same spirit as the fluid model discussed in section 3.5.2. This
fact is not so surprising since we have already noticed in section 3.2.3 that
sound waves can propagate in a fluid. The key ingredients to obtaining a
wave equation were to have mass and momentum conservation and also
to neglect nonlinear velocity terms and viscous effects.
We shall no'v consider these ingredients again and formulate a lattice
BGK model of colliding particles which privileges wave propagation over
the normal fluid behavior. In a lattice BGK model, we write the evolution
rule as
(7. 1 7)
268 7 Other models and applications

or, equivalently

�
fi(r + -rvj , t + -r) - fi(r, t ) = (!�0) (r, t ) - fi(r, t) ) ( 7. 1 8 )

where, here, v i designates the velocity vector along the ith lattice direction.
The quantities fi are interpreted as the average number of particles
traveling with velocity V i at site r and time t. At this point the relaxation
parameter � is a free quantity. We shall, later, see that only one specific
value is acceptable to get clean wave propagation.
We now introduce the conserved quantities
'¥ = L mdi j= L mdi v i ( 7.19 )
where mi are free parameters corresponding to the mass of the particles
traveling in lattice direction i.

The local equilibrium. In order to be in a situation similar to that

discussed in section 7.1 .2, we now consider a two-dimensional case and a
square lattice. However, for more generality (and to be able to adjust the
propagation speed), we shall assume that rest particles are possible and
accounted for by the field fo . Thus, the index i runs from 0 to 4 and, of
course, the speed v0 associated with the rest particles is null
vo = 0
The physics involved in equation 7. 1 8 is determined by the expression
for the local equilibrium distribution ff0 l. To describe wave propagation
we define

( 7.20 )

where ai and b i are constants to be determined and v 2 = ( A./ -r) 2 is the

modulus of the V i . for i =/= 0.
Compared with the hydrodynamical situation (see equation 3.246 ) , we
have here an exp ression for the local equilibrium which is linear in the
particle current J. This a priori choice can be understood from the fact
that sound waves are seen in hydrodynamics when the velocity field is
small and nonlinear terms neglected. We prove below that equation 7.20 is
indeed the correct choice to produce waves instead of a hydrodynamical
behavior.
This freedom in the choice of the local equilibrium distribution ff0l
is actually a very interesting property of the lattice Boltzmann scheme :
different physical behaviors can be captured by suitable choice of ff0l ,
while still keeping the same evolution equation.
 7. 1 Wave propagation 269
The value of the constants ai and bi can be calculated by imposing mass
and momentum conservation. As usual, we require that the right-hand
side of equation 7. 1 8 vanishes when multiplied by mi or mi v i and then

4
summed over i. From equation 7.19, this amounts to asking that

'¥ = 2: md�0)
i =O

4� ( 4 )
Using the expression 7.20 for f�0) , we have

1 ...
'¥ = 'l' � mI· a·I + -2 �
� m1· b·v·
1 1 ·J
1=0 2V 1=0

tz=O (t1=0 )
and

2� 2 mi b i ( v i ] ) v i
J = '¥ mi a i i\ + ·

But we know from chapter 3 that 2:� 1 V i = 0 and 2:� 1 VirxVip = 2v 2 �rxp,
since the v i s point along the main directions of a square lattice. Thus, if

(�4 )
mi b i = 1 we simply have (remember that v0 = 0 )
1
mi b i ( Vi · J ) v i = J
� -+ ...
-+

2v 2
Therefore mass and momentum conservation can be enforced if
mi = m ai = a bi = b for i = 1, 2 , 3, 4
and

aomo + 4ma = 1 bo = 0 b = _!_ ( 7.21 )

m
We can summarize these results as follows :
1 vi · J
- a'¥ + - -­
2 if i f 0
m 2v
ao'¥
aomo + 4ma = 1 (7 .22 )

The propagation matrix. We can now rewrite the evolution law 7. 1 8 using
the above expression for f� 0) . We obtain

fo (r, t + -r ) = z [ao'¥ + (� - 1 )fo ]

270 7 Other models and applications

and, for i =I= 0,

fi(r + t"i\, t + t") = � [alJI + ! v;�[ + (� - 1 )/i]

o
By replacing lJI and J with their expressions 7. 1 9 in terms of f� ) we obtain

/o(7, t+') � [<aomo + � - 1 )/o + aom t, t.]

�

fi(r + t"Vj, t + t") - � [ (am + � - � ) + amfi+ 1 h

( l ( l
+ (am - � ) fi+2 + amfi+3 + amofo] (7.23)

This can be written in a matrix form

fo(r , t + t") fo(r, t)
h(r + t"V 1 , t t") + (r, t ) h
h (r+ t"V2 , t + t") = w
t) h(r, (7.24)
h (r + t"V3, t + t") h (r, t )

(
f4(r + t"V4, t + t") f4(r, t )

)
where
aomo + � - 1
W = -;1 ; amo am + �
amo am + �
�1 am
�i
am
�1 �i
<:. amo am - � am am + � - � am
amo am am - � am am + � - �
(7.25)
By choosing the parameters in matrix W in an appropriate way, we find
the matrix Wfree given in equation 7. 10. Indeed, in the case that there is
no rest particle (i.e. 0,fo = ao =
0), we have

a = -1m 4
Now, if � is chosen as

� = -1 2
W reduces to
-mo1 1
0
1
0
1
0
1
0
2m 2 2 -2 2
!!!!!. 1 1 1 1
W= 2m 2 2 2 -2
mo 1
-2
1 1 1
2m 2 2 2
mo 1 1 1 1
2m 2 - 2 2 2
 7. 1 Wave propagation 27 1
which, except for the first row and column, is exactly the same as 7. 10.
Thus, with fo = 0 the lattice BGK formulation 7. 1 8 is equivalent to the
wave model developed in section 7. 1.2.
Note also that the quantities fi are not guaranteed to be positive. For
some choice of 1p and J , it is certainly possible to make ff0l negative as can
be seen from the definition 7.20. Actually, it is clear that fi will be negative
since they represent an oscillating wave. This is not very satisfactory if the
fis are interpreted as an average number of particles. However, since the
dynamics is linear, we can always think that fi describes the fluctuations
around some large positive constant value.
Adjusting the propagation speed. Equation 7.25 is a five by five matrix and
the presence of rest particles allows us to change the wave propagation
speed. In fact, each lattice site can have its own refraction index and
it is then possible to model propagation through different media or in a
disordered medium.
To obtain the resuting wave equation, we first compute the equation
governing the evolution of the conserved quantities 'I' and J . To this end,
we shall use the same formalism (Chapman-Enskog expansion) as used in
section 3.5.2. Since we have mass and momentum conservation we know
that we have (see equations 3.87 and 3.88 )
at 'I' + opJp o ( 7.26 )
[
otlrx + op IIrxp + � (eot1 II�t + oySrxpy ) - 0
] ( 7.27 )

where Ot = E Ot1 + e2 ot2 is expressed in terms of the two relevant time scales
and e is the expansion parameter.
Using the usual definitions (see chapter 3 ) , the momentum tensor is
II = rr (O) + eii ( l)
and its the zeroth order is given by
rr�o; - L mdf0l VjrxVip
= am'I' L VirxVip + ::2 L VjrxVipViy
2amv 2'Pb rxp
Hence
( 7.28 )
and
( 7.29 )
where we have used the continuity equation 3.77 for the last equality.
272 7 Other models and applications

Now, the O ( c ) correction to the momentum tensor is

rr�}/ = L md? )via Vip
As obtained in section 3.5.2, J} 1 l has the solution
(7.30)

where iha is the order c of the spatial derivative with respect to coordinate
a, i.e. Oa = E O!a· Thus,

Errap( l)

where Tapyo is, by definition, the fourth-order tensor

Tapyo L ViaVipViyVio
=

Remember that, since we are here working on a square lattice, without the
diagonals, Tapyo is not isotropic (it depends on the particular orientation
of the vectors v i) . This is of course not desirable and we shall see below
how this problem is solved. For the time being we have

opciTap(1 ) _- 2(amv 2 roadrvJ - 2rv(2 Tapyoopoylo

.
�

(7.3 1 )
Finally, we also have to compute
s�jy L miViaVipViy /}0)
= � I: ViaVipViy Vio) fo
2 2 (

The term involving 'I' vanishes since it contains the product of three v i s.
Therefore
T
apoyS (0) _- T 2 Tapyoopoylo
apy (7.32)
2 4v
We can now collect relations 7.28, 7.29, 7.3 1 and 7.32, and plug them
into equation 7.27. After grouping the terms, we have
Ot la + 2amv 2 oa'l' + (2( - 1 ) amrv 2 o a divJ - (2( - 1 ) 4Tv 2 Tapyoopoylo
�

0 =

(7.33)
As we can see, by choosing ( = 1 /2, we cancel the anisotropic term
containg Tapyo as well as the dissipative term (2( - 1 ) amr v 2 8a divl
In the fluid model of section 3.5.2, ( plays a similar role in the sense
that ( = 1 /2 corresponds to the limit of zero viscosity. But there, this
 7. 1 Wave propagation 273

limit is numerically unstable, due to nonlinear instabilities. In the wave

model, on the other hand, this limit is stable and is the only one which is
acceptable due to isotropy (note that for � > 1 /2, the fluid model is still
isotropic since the lattice is not the same as it is here).
With � = 1 /2 we can now rewrite the momentum conservation equa­
tion 7.33 as
(7.34)
The wave equation can now be obtained easily. We first differentiate the
continuity equation 7.26 with respect to time
a ; '¥ + 8 t8pJp = 0
and we take the spatial derivative 8rx of 7.34 and sum over r:x

OtOrxlrx + 2amv 2 V2 '¥ = 0

After substitution of the second equation into the first one, we obtain
(7.35)
Thus, as expected, '¥ obeys the wave equation with propagation speed c,
where
C = V�
As opposed to the situation described in section 7. 1 .2 (without "rest
particles"), this propagation speed c is now adjustable. Indeed, in condi­
tions 7.21 we still have the freedom to choose the quantity aomo.
Equation 7. 1 8 is linear in the f;. Thus, in principle there is no need
to consider a sophisticated method (the Chapman-Enskok expansion) to
derive the wave equation. However, in the LBGK framework, this method
is straightforward to apply and makes it unnecessary to compute the
eigenvalues of a 5 x 5 matrix.
In summary, with � = 1/2 the wave model with ajustable speed is given
by the propagation matrix
m m m m
2n2 - 4 - (4n2 - 4) - (4n 2 - 4) - (4n 2 - 4) - (4n2 - 4)
mo mo mo mo

1 1 - 2n2
mo

1 - 2n2
1 mo
W(n) =
2n2 m

1 - 2n2
mo

1 - 2n2 1 1
mo

m
(7.36)
274 7 O ther models and applications

Fig. 7.9. Lattice Bolztmann wave model and focusing of a plane wave by a
convex lens with a refraction index of 1 .7. The s ource is on the leftmost side of
the system. Energy is shown in this simulation. Light gray denotes high energy.

where n is the refraction index defined as the ratio of the free propagation
speed v / J2 to the propagation speed c = 2am v 2
v2
n2 = -
2c2
Figure 7.9 illustrates the propagation of a plane wave through a lens
shaped medium. The lattice points outside the lens are characterized by
a matrix W ( l ) and fo Ci\ t) = 0 while the internal sites have W( n ) with
n ?.: 1. Due to the change of speed between the two media and the shape
of the interface, we observe a focusing of the incident plane wave.

Time reversal invariance. In the previous section, we saw that the relax­
ation time � should be set to 1 /2 in order to obtain a good wave model.
There is another intere sting way to see this, using time reversal symmetry.
Time reversal is a fundamental symmetry of microscopic physics and is
clearly satisfied by the wave equation which has only second-order time
derivative. We shall see here that time reversal implies
� 1
s = -
2
For this purpose, let us write equation 7. 1 8 as

( 7. 3 7 )

where J; designates the outgoing flux and W the propagation matrix.

As explained in section 2.2.7, a dynamics is reversible in time if, after
reversing the direction of the flux, the system traces back its own past. In
our case, reversing the flux is achieved with a reversal matrix R such that
and ( Rf) o = fo (7.38)
 7. 1 Wave propagation 275

where .f stands for the five-element column vector whose ith component
is h Thus, in order for equation 7 . 3 7 to be invariant under time reversal,
one should have ( see section 2.2.7)

W R.f ' =
Rf
Using that ( see equation 7.20)
b· �
a z· 'P(.f) + _l2 iJ '. · J(.f)
f
�)
. t
=
. 2v
_ __
.
we have

( WR.f' )t =
(1 - z) (R.f ' )t + z[ a t 'I' (Rf ' ) + �2 vt · l (R.f ' ) ] (7.39)

Let us now compute 'P(R.f') and I (R.f ' )

4

'P(R.f ' ) =
L mJ (Rf ' )J =
mo (R.f ' ) o + L m (Rf ' )J
j=O j f'O
4

mofb + L mf.i+2 = L m.if}

J=!=O j=O
4

L mJfJ (because of mass conservation)

j=O
'P(f)

Similarly,
4
I (Rf ' ) L 11lj Uj (Rf ' ); =
L u_; m (Rf);
} =0 j fO
4

- L fnVJ+2i"J+2 =
- L m_;v_;f}
J=!=O j=O
4

L.. m}-ii..;f";
� (because of momentum conservation )
} =0

- l (f)

Substituting these expressions for 'P( Rf' ) and I ( Rf ' ) into 7.39 yields ( for
i j 0)

( W Rf ' )t =
( 1 - l ) f'i+2 +
l fa n�
l rC )
· r· b �
- 2v2 Vi .
J� ( f) l
. J
Z Z
276 7 Other models and applications

( ) fi+2 + [alJI(f) + 2vb2 vi+2 . J(f)]

1
1 1 1 �
�

( zZ ) [ ( zZ) fi+2 + zfi+2] + zfi+2

-

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

( z ) 2 fi+2 + z ( z ) fi+2
1 -
1 1
2 -
1 (0)

Since we require that

we have the conditions

whose solution is
( = 1 /2
Similarly, the condition ( WRf ' ) o = ( Rf) o = fo also imposes this value for
(.
I n conclusion, with ( = 1 /2, the LBGK formulation o f the wave model
becomes
(7.40)
This expression gives a natural way to extend the model to a three­
dimensional cubic lattice (see also [21 2] ) since our derivation did not use
any specific properties of the two-dimensional space.
In the previous section, we saw that ( = 1 /2 makes the viscosity
vanish, that is requiring time reversal invariance is equivalent to removing
dissipation (actually breaking of time reversal symmetry is the standard
definition of dissipation). This observation shows that the LBGK fluid
models of section 3.5.2 is not reversible in time. Thus, this symmetry is
lost when compared with the cellular automata fluids such as FHP. In the
cellular automata approach, the microdynamics is time reversal invariant.
This symmetry breaks only when the average behavior is considered to
obtain the Navier-Stokes equation.

Energy conservation. Our first interpretation of the wave model was to

say that the fi represent local elastic deformations in a solid medium
made of alternating black and white particles connected by springs. From
this physical picture, we may argue that 2::::: J1 corresponds to some po­
tential energy. This is the way we defined energy in the string model of
section 2.2.9. Now, since the black and white particles move between time
t and time t + r in such a way that the springs are always in maximal
 7. 1 Wave propagation 277

deformation state when one looks at them, we may also argue that kinetic
energy vanishes at the observation time steps.
In this section we want to obtain the condition that the evolution matrix
W must satisfy in order to conserve the energy
4
E (r, t ) = L il = (!If)
i=O
where (f l f) denotes the scalar product.

( OJ
It is easy to check that the matrix W given by 7.25 obeys
RWR = W
where R is the flux reversal operator introduced in 7.38
1 0 0 0
0 0 0 1 0
R= 0 0 0 0 1
0 1 0 0 0
0 0 1 0 0
The above property means that W and R commute (note that R 2 = 1 ) and
reflect a natural spatial symmetry of the dynamics : if an incoming flux fi
enters a site it produces outgoing flux fj = Wji/i . Now the incoming flux
is fi+2 = ( Rf) i we shall get the same outgoing flux, but pointing in the
reverse direction, namely ( Rf' )j .
Energy conservation requires that the energy E ' after scattering is equal
to the energy before scattering
E ' = (f ' lf' ) = ( Wf i Wf) = ( W j_ Wf l f)
where W j_ is the transpose of matrix W. Thus energy will be conserved if
w j_ w = 1
In the case that the dynamics is invariant under time reversal (and that
is really what we want here), we learned in the previous section that
WR W = R
Since we have also shown that R W R = W we obtain that
W2 = 1
Therefore the condition of energy conservation demands
W j_ = W
or, in other words that W be symmetric. We can return to equation 7.36
and use the extra freedom in the choice of � to make W symmetric. The
278 7 O ther m odels and applications

condition is

- = 2 v�
mo m . mo
- - - ( 4n2 - 4) or n2 - 1
m mo m
Thus, with energy conservation and time reversal invariance, the wave
propagation matrix in a medium of refraction index n � 1 is

- 4 2 #-1 2 �-=1 2 -)n2 - i 2 .}�2 - 1

2n2
2 "n 2
-l 1 1 1 - 2n2 1
1
W(n) = -
·
2 �"ii2
-1 1 1 1 - 2n 2
2n 2
2 Jn 2
��1 1 - 2n 2 1
2 )n 2
-1 1 1 - 2n 2 1 1
(7. 4 1 )
Note that here the limit n = 1 corresponds to taking mo = ao = 0 in
equation 7 .22.

7. 1 . 4 An application to wave propagation in urban environments

As an illutration of our wave propagation model we consider, in this

section, the problem of simulating how a radio wave propagates in an
urban environment. The reason for studying this question stems from the
rapid development of mobile, personal systems such as cellular phones.
Large cities constitute an interesting situation because radio waves are
absorbed, reflected and scattered in a complicated way by the buildings
and the amplitude pattern of a wave emitted by any antenna surrounded
by such obstacles may be very complicated and beyond analytical calcula­
tion. In addition to the physical problem of solving the wave equation in
a disordered environment, there is the problem of resource optimization :
some regions of a city will experience a high mobile communications traf­
fic and the base stations in charge of relaying these calls should be sized so
that they can service everyone. The solution is to place more base stations
in regions of dense traffic. Thus, a city is divided into cells (defined as the
region covered by a base station) whose size is chosen so that the potential
number of calls within this cell matches the limited number of channels
available at the corresponding base station. For continuity reasons during
transmission, adj acent cells must have overlapping b oundaries.
An optimized set of base stations should ensure complete coverage with
a minimum number of cells of appropriate area. Therefore the starting
point for the planning or maintenance of any mobile communication
network is based on a prediction of the radio wave attenuation around
the transmitter ( antenna). In this problem, it is important to note that
the antenna may well be located at an elevation which is less than the
building height.
 7. 1 Wave propagation 279

Fig. 7. 10. Wave propagation in a region of the city of Bern, as predicted by the
lattice Boltzmann model. The gray levels show, from white to black, the intensity
of the wave. The source antenna is located at the dot mark. The white straight
line indicates the path along which intensity measurements have been compared
with real measurements.

In our simulation we consider the map of a city and its discretization on

a square lattice. Two different choices can be considered for the lattice sites
corresponding to the boundary of buildings. First, they can be governed
by a propagation matrix r:t. Wr ef where r:t. < 1 is an empirical constant which
accounts for the energy absorbed by the bulidings. Another choice is to
take the matrix Waa (cx) on the building walls.
The other lattice sites (corresponding to streets) obey the Wf ree propa­
gation matrix. Finally, a source can be located anywhere along the streets
to simulate the presence of an antenna.
The simulation of such wave propagation in a city is given in figure 7. 10.
When the transmitting antenna is below the rooftop, the essential fea­
tures of a building layout are assumed to be captured by a two-dimensional
simulation (which is equivalent to a line source and infinitely high build­
ings) . A real antenna, however, is considered as a point source in a
three-dimensional space. Quantitative comparisons of simulations with
real measurements become possible only after an appropriate renormal­
ization is performed.
In addition to this 2D propagation, the simulation shown in figure 7 . 1 0
has another problem. It has not been performed a t the frequency used in
280 7 Other models and applications

personal mobile communication systems ( 900 and 1 800 MHz). The reason
is that the discreteness of the lattice is not fine enough to work at such a
high frequency. The renormalization proposed is the following [215] .
We suppose that the measured (real) amplitude A 3 v of the wave can be
approximated by the outcome azv of our simulation using the following
relation :
( 7.42 )
where A is the actual wavelength used in the measurement and A. the
simulated one. The quantity (> represents the distance around the buildings
separating the transmitter from the receiver.
The rationale for this expression is that it is true for wave propagation
in vacuum. It is then assumed that the effects of multiple reflections
and _ scattering on buildings is already captured in azv : these are the
geometrical contributions which are expected to be the same in two
and three dimensions. Finally, since the distance traveled by the wave
is not accounted for in the same way in 2D and 3D, the rest of the
renormalization is assumed to be included in the function f>Ol Here we
have used for f>(r) the so-called Manhattan distance which is the shortest
distance following the lattice edges representing the discrete city map.
Figure 7. 1 1 gives a comparison between the predictions of our model
and real measurements. The renormalization given by equation 7.42
has been used to compute the wave intensity along the path shown in
figure 7.10. We observe good qualitative agreement between the two curves
and also acceptable quantitative agreement knowing that our simulation
neglects all small-scale details of the real city and that each building is
assumed to produce the same absorption and reflection. Finally, it should
be noted that, in the simulations, some streets are too narrow to permit
wave penetration which may explain why some intensity peaks are not
correctly reproduced by the simulations.

7.2 Wetting, spreading and two-phase fluids

7.2. 1 Multiphase flows

Sytems in which different species of fluids coexist, move and interact
are frequently encountered in science. Such systems are often termed
multiphase flows and represent an important domain of application of
the lattice gas approach. Models for miscible, immiscible and reactive
two-phase flows have been proposed [ 35,216,97,217,46] . Initially cellular
automata models were developed, and more recently, lattice Boltzmann
models have been proposed.
 7.2 Wetting, spreading and two-phase fluids 28 1

�
- 1 05 r
� -115 iI
"' t\
tl..�
� - 1 25 �-�
.. I \
I \

. \

- 1 35 '
i \/

- 1 45 L--�--�----�j
0 1 00 200 300 400 0 1 00 200 300 400
distance (in meters) along the street

Fig. 7. 1 1 . Lattice Boltzmann predictions of wave intensity (dashed line) and in

situ measurements (thick lines) along a street in the city of Bern (courtesy of Swiss
Telecom PTT). The left panel corresponds to a simulation with a propagation
oc w;;1 on building sites, whereas the right panel shows the result obtained with
an attenuation propagation matrix Wau(oc) at each building wall.

Multiphase flows are necessary to model, for instance, a Rayleigh­

Taylor instability where two immiscible fluids lie on top of each other,
so that the heavy one is above the light one. Small perturbations cause
the heavy fluid to move down, giving rise to a spectacular mushroom­
like unstable interface. Chemical processes such as combustion or front
flame [2 1 8,2 1 9] can also be investigated within this framework.
Multispecies fluids can be modeled in much the same way as the
reaction-diffusion systems discussed in chapter 5, except that now, particles
obey hydrodynamics instead of a random walk. Thus, an FHP model or
an isoptropic lattice Boltzmann model with momentum conservation is
necessary to represent the basic dynamics of each species. For practical
reasons, one often associates a different "color" to each species involved,
in order to mark and distinguish them.
Interactions between the species are the key ingredient of multiphase
flows and, depending on the nature of the physical processes, different
interactions will be considered.
For instance, an immiscible two-phase flow can be simulated with a
local attraction between particles of the same type and a local repulsion
between particles of different type. For this purpose, a "color" gradient is
computed at each lattice site, using the color densities on the neighboring
grid points. The direction of this gradient actually determines the local
interface between the species. The output of particle collisions is then
chosen so that a particle of a given color preferentially moves in the
direction where its own species resides. Of course in such a collision, mass
(color) is conserved separately for each type of fluid, whereas momentum
282 7 Other models and applications

can be exchanged. A realistic surface tension naturally emerges from such

a rule, as a consequence of these microscopic attractions/repulsions.
The lattice gas approach has the important advantage that the interface
between the different fluids appears naturally as a consequence of the
way the fluids are modeled in terms of particles. The evolution of this
interface results from the collective behavior of all particles and does not
need to be incorporated explicitely in the model (as is the case with more
traditional approaches, which introduces many complications).
For more detail on how the lattice Boltzmann method can be applied
to multiphase flows, we refer the reader to the literature cited at the
beginning of this section and in particular to ref. [46] . For the remainder
of this discussion, we present two specific applications where body forces
or surface tension are included. The first one is a cellular automata
fluid with inter-molecular attracting forces to simulate the wetting and
spreading of a liquid droplet on a solid substrate. The second application
is an Ising fluid model. Our main purpose is to illustrate how lattice gas
techniques can be extended to include more complex physical processes.

7.2.2 The problem of wetting

The phenomena of wetting occurs, for instance, when a fluid is in contact
with a solid interface, such as the wall of a container. It is well known
that, at equilibrium, the free surface of the fluid joins the solid with some
given contact angle, determined by the respective interaction between the
molecules in the solid, those in the fluid and those in the surrounding air.
For instance (see figure 7. 12), a liquid droplet sitting on a flat solid
surface rests in equilibrium with a static contact angle (} given by Young's
relation
Ylv COS (} = Ysv - Ysl
where Y tv . Ysv and Ysl are the liquid/vapor, solid/vapor and solid/liquid
surface tensions, respectively.
When such a balance between the surface tensions is achieved, one has
a situation of partial wetting. However there are some situations where
Y t v . Ysv and Ysl are such that (} goes to zero. There is a transition between
partial wetting to a situation of complete wetting where the liquid droplet
spreads over the solid surface. Complete wetting can be observed in
simple every day life experiments : when a droplet of oil is deposited on
the surface of water, it spreads slowly to form a thin layer which may
eventually break if there is free space left on the water surface.
The spreading of volatile and non-volatile liquids is different. In the first
case there is equilibrium between the liquid and its vapor and molecules
of the liquid can be efficiently transported through the vapor phase and
 7.2 Wetting, spreading and two-phase fluids 283

vapor

Fig. 7. 1 2. A liquid droplet in equilibrium with its vapor.

rapidly recondense into a thin layer over the substrate. On the other hand
a non-volatile liquid covers the substrate much more slowly than a volatile
substance, because it has to flow.
It has long been observed that the spreading of non-volatile liquids
are often accompanied by the formation of a so-called precursor film
at the interface between the substrate and the liquid [220] . This film
spreads faster than the bulk of the droplet and is typically a monolayer
of molecular thickness [22 1] . Sophisticated experimental techniques [222]
have also shown that the fluid spreads as a series of superposed molecular
layers.
As a result of these features, the spreading of liquid is a very interesting
phenomena for which much is still to be understood from first principles.
The Navier-Stokes equation, commonly used to described fluid motion, is
not the appropriate approach in this case. The usual boundary condition
in the Navier-Stokes equation is a "no slip" condition at the liquid-solid
interface. A zero velocity field at the boundary is clearly incompatible with
droplet spreading [223] and the existence of a precursor film. Furthermore,
the Navier-Stokes equation assumes that each fluid element is composed
of many molecules. This hypothesis is not true in our case since the
layering which is observed occurs at a molecular scale.
Several theoretical models have been proposed to describe the spreading
of a droplet [224, 225] . They assume the existence of two-dimensional
layers of fluid, moving on top of each other with some viscous force.
These models give predictions of the speed of the precursor film, the
droplet profile and the partial/complete wetting transition.
Wetting and spreading phenomena are of great importance in industrial
processes used in the manufacture of lubricant and paints. From a nu­
merical point of view, spreading can be investigated in terms of molecular
dynamics simulations, since it results from inter-molecular attractive or
repelling forces. However, molecular dynamics requires massive compu­
tational resources and we are interested here in the long time regime for
which this approach may not be appropriate.
284 7 Other models and applications

Therefore, 1t 1s natural to turn to the cellular automata approach

which provides a simple description of the fluid particles at a microscopic
level. We start from the FHP model to which we add nearest-neighbor
interactions among the fluid particles, in order to produce surface tension.
We also consider external body forces to account for the attraction due
to a solid wall or gravity.
As we shall see, simulations of our model show the formation of a
precursor film, provided suitable boundary conditions are used at the
liquid-solid interface. The motion of the upper fluid layers (above the
precursor film) are found to obey power laws. The droplet profile is found
to be parabolic.

7.2.3 An FHP model with surface tension

The plain FHP was introduced in section 3.2. Here we shall use the same
terminology and denote by n i ( 1 \ t ) = 0 (or n i (' r , t ) = 1 ) the absence (or
presence) of a particle entering site r at time t with velocity Vj.
In order to simulate wetting phenomena, we need first to include a
surface tension in the fluid. Second, since spreading is due to Lennard­
Janes interactions between the fluid particles and the solid substrate, we
need to include external forces acting on the fluid. The resulting force
F ( r , t ) acting on each particles produces a new velocity distribution n;( r , t )
so that
L n;( r , t ) = L ni ( r , t) = n ( r , t ) (7.43)
i=l ,6 i=l ,6
and
P ' ( t, r ) - P ( t, r ) = n ( r , t ) F ( r, t ) (7.44)
The quantities P ( t, r ) and n ( r , t ) are the total momentum and the total
number of particles at site r and time t, respectively. P ' ( t, r ) is the total
momentum just after collision. The momentum is defined as
P ( r , t) = L ni( r, t) v i (7.45)
i=l,6
Equation 7.44 states that particle number is conserved such that the vari­
ation of momentum for each particle is equal to the total force (Newton's
law of mechanics). Equation 7.43 states that no particles are lost or
created during the interaction. After the interaction has taken place, the
particles move to a nearest-neighbor site, according to the direction of
their new velocity.
Clearly, in a cellular automata model, P can take only a finite number
of values and it is not expected that equation 7.44 will always have a
solution. We may even say that usually it has no solution and the new
 7.2 Wetting, spreading and two-phase fluids 285

particle distribution should be computed so as to minimize the error. As a

consequence, if F is the mutual attraction between the fluid particles, the
quantity L r P (r, t ) may not be exactly conserved during time evolution.
A better way to ensure momentum conservation while still having
the existence of mutual attraction would be to follow the approach of
Rothman [2 1 6] : two fluids are considered which could be interpreted
here as the liquid and the st1rrouding gas, respectively. Momentum
can be exchanged between the two fluids so that global conservation is
guaranteed. However, within each fluid, momentum may not be conserved.
Our approach may be seen as a simplified version of this approach,
in which the second fluid is disregarded. Empty lattice sites may be
considered to act as reservoirs of momentum.

Surface tension. Let us now discuss the way to compute the force F in
our model. Surface tension is produced by short-range mutual attraction.
Keeping only interactions with nearest-neighbor particles, we define
F('r, t) = f L Cj L nj ( r + Cj, t) (7.46)
i,=1,6 j=1,6
where, as usual, Cj denotes the unit vector joining two nearest-neighbor
sites (j is labeling the direction) and f is some constant describing the
magnitude of the elementary attraction.
Figure 7. 1 3 illustrates the result of such an interaction. The force F
given in equation 7.46 (with f = 1 ) causes the particles at the middle site
to change their velocities according to equation 7.44. This example shows
a situation where equation 7.44 can be fulfilled exactly. This is usually not
the case and one has to find the "best" approximation to equation 7.44
by inspection of all possible distributions of the nis. Of course, a lattice
Boltzmann version of this model would give much more flexibility in
the implementation of equation 7.44 because the n i would no longer be
restricted to the values 0 or 1 .
I n order to adjust the surface tension, a threshold quantity can be
introduced in the rule : the mutual interaction F between fluid particles
is retained only when it is larger than this threshold value. In particular,
this prevents a single particle from capturing another one and forming
with it a stable oscillating pair,
In the same spirit, we can also introduced in the rule a "temperature
effect," that is some probability that surface tension does not operate at a
site. When this happen, the particles are submitted only to the action of
the external force. The purpose of this probability is to cause evaporation,
i.e to let particles escape from the attraction of the liquid bulk.
286 7 Other models and applications

't,
-
',

',

--- ----L- ----

\
\

, ,

'

',

Fig. 7. 1 3 . Example o f surface tension forces. We show the result o f the

1ttraction of the gray particles on the two black ones. The resulting force is a
1orizontal vector pointing to the left which causes the black particles to change
:heir momentum along the horizontal direction.

External forces. External forces, like those produced by gravity or by

the resulting attraction due to the substrate, sum up to the internal forces.
fhis external force field is modeled as follows. At each lattice site, one
1ssociates the angle and the amplitude of the external force. For instance,
�ravity is constant in amplitude and direction at each site. Attractions
jue to a solid wall give a force normal to the surface, with a l jt 3 decay,
where t is the distance to the wall. This value is obtained as the sum of
1ll Lennard-lones contributions from every solid particles at the interface.
Usually, it is not appropriate to have the external force act at every
lattice site : a small fraction of them are sufficient but not always the
;;arne ones. For this reason, a probabilistic force is considered. The force
[s taken into account only when its amplitude is larger than a random
number which is generated at each time step and each lattice site.

Boundary conditions. Finally, we have to introduce the rule for the

particles reaching the solid substrate or the limit of the system. We can
;:;onsider three different possible behaviors when a particle encounter a
;;olid wall, as shown in figure 7. 14. We call them : (a) specular reflection ;
(b) bouncing back ; and (c) trapping wall.
The first two conditions are standard boundary conditions in lattice gas
models. We have introduced condition (c) in order to take into account
the strong attraction that may exist for particles right on the wall. The
idea is that, once a particle has hit such a wall, strong short-range forces
trap it. The particle is then restricted to move parallel to the wall. Of
course, due to the exclusion principle, such trapping is only possible when
 7.2 Wetting, spreading and two-phase fluids 287

Fig. 7. 14. The various boundary conditions that have been used when a particle
meets a solid wall. (a) Specular reflection, (b) bounce back condition and (c)
trapping wall condition. Condition (c) is necessary to produce a precursor film.

there is no particle already traveling in the considered direction. If this

is the case, the particle will bounce back from where it came. As we
shall see, the trapping boundary condition appears essential to produce
the precursor film.

7.2.4 Mapping of the hexagonal lattice on a square lattice

Simulating a hexagonal lattice on a computer requires some mapping,
since the natural data structures on a computer are Cartesian grid (arrays).
There are two simple ways to map the hexagonal lattice on a grid, as shown
in figures 7. 1 5 and 7. 1 6.
In the first case (figure 7. 1 5), the square lattice undergoes a deformation
which amounts to tilting the vertical lines by 60 degrees and shrinking the
vertical scale by a factor of -/3 /2. If each lattice site is connected to its
north, east, west, south, north-east and south-west neighbors, then in the
transformed system a hexagonal topology is obtained.
If ( i, j ) is the location of lattice site r in the square lattice computer
topology, the corresponding coordinates ( x, y) in the physical topology
(hexagaonal) are

y =
where .A is the length of the lattice spacing.
In the second mapping technique (see figure 7. 1 6), odd and even hor­
izontal lines are treated differently. Odd lines are shifted half a lattice
spacing with respect to even lines. In addition to north, east, west and
south neighbors, lattice sites on an even line are also connected to their
north-east and south-east neighbors. On the other hand, lattice sites re­
siding on an odd line are connected to their north-west and south-west
288 7 Other models and applications

Fig. 7. 1 5. Mapping of a hexagonal lattice onto a square lattice, by means of a

60 degree deformation.

Fig. 7. 1 6. Mapping of a hexagonal lattice onto a square lattice, by means of a

translation of half a lattice spacing for every second horizontal lines.

neighbors. Thus the mapping can be expressed as

Even j : X = Ai

Odd j :

The advantage of the first technique over the second one is that the
transformation is the same for all j and that the connectivity of each
lattice site is identical. On the other hand, boundaries are more easily dealt
with in the second method. Periodic boundary conditions in the vertical
direction are preserved in the mapping of figure 7. 1 6 while they are not
in figure 7. 1 5. Also, a vertical wall is inclined in the computer topology
of the first mapping, while, in the second case it can be implemented as a
straight vertical line. Thus, despite its more difficult implementation, the
second method is probably safer.
 7.2- Wetting, spreading and two-phase fluids 289

(a) (b) (c)

Fig. 7. 1 7. Free surface of a liquid in a capillary tube : (a) a wetting situation ; (b)
inert walls ; and (c) a non-wetting situation. Gravity is acting downwards.

7.2.5 Simulations of wetting phenomena

Capillary tube. First, we consider a "capillary tube" simulation in order
to see the formation of a meniscus. The result is shown in figure 7.17,
in which three kinds of solid-liquid interaction are considered. Situation
(a) corresponds to an attracting l jt3 force, where t is the distance to the
wall of the container. In (b), we have a situation with no liquid-solid
interactions. Situation (c) is equivalent to (a) except for the sign of the
force, which is repulsive.

Spreading. Now, we consider the situation of spreading. A drop is

deposited on a horizontal attracting substrate. Its surface is ideally fiat
and produces a 1 jt 3 force above it. Gravity is neglected. As the simulation
goes on, the droplet spreads over the substrate, as illustrated in figure 7. 18.
Note that particles can be exchanged between adjacent layers. This motion
takes place mostly downwards, due to the direction of the substrate
attraction and, as a result, the fluid layers must flow horizontally.
Boundary conditions are periodic along the horizontal axis (direction
of spreading). The upper limit of the system is a reflecting wall. On the
substrate, we have tried the three types of boundary conditions discussed
in the previous section.
290 7 Other models and applications

Fig. 7. 18. Spreading of a liquid droplet on an attractive substrate. Bouncing

back condition is used at the substrate. These snapshots are shown after 300,
1000 and 3500 iterations. Evaporation can occur, as shown by dot particles.

With both specular reflection and bouncing back conditions, the droplet
is seen to spread away. However, no precursor film are observed. The
fluid may be considered as being at the transition between partial and
complete wetting. We can measure the position x0 (t ) of the first fluid layer
as a function of time (the existence of layers of molecular size is obvious
here, due to space discreteness). We observe, in figure 7. 19(a), a t 1 12
behavior
xo(t ) "" t 1 12
in agreement with the results of [226,227] . Upper layers are found to
advance also with power law, but with different exponents.
The droplet profile can be measured. It turns out to fit relatively well
a parabola, at least for the lower part of the droplet, as indicated on
figure 7. 1 9(b ) This result agrees with the prediction of [227] for the
.

solid-on-solid model described in [225,226] .

Finally, if we consider our "trapping wall" boundary condition on the
substrate, a rapidly advancing precursor film is formed at the liquid-solid
interface. As in the model [226] , it moves at constant speed (xo(t ) "" t ) ,
much faster than the rest of the fluid. This film is fed by particles flowing
 7.2 Wetting, spreading and two-phase fluids 29 1

(a) (b)

3 .6 IL-L......J--l----L---L---l..---1...--l.--'--' l --'---'---..L.--'-----'---L---'
0 2.8 1 10 36020
ln(t)

Fig. 7. 19. (a) Measurement of the time behavior of the first fluid layer at the
liquid-solid interface (horizontal advance versus time in a log-log scale) ; the
slope gives x '"" t0·5. (b) Measurement of the droplet profile at a given time, with
an appropriate change of scale ; only the first 3 1 layers are taken into account
(one-third of the total height) ; this plot indicates a parabolic profile, at least for
the foot of the droplet.

Fig. 7.20. Snapshots of a spreading droplet, as given by our model with the
"trapping" boundary condition. A rapidly advancing precursor film is present at
the liquid-solid interface.
292 7 Other models and applications

Fig. 7.2 1 . Mutual interaction of two particles separated by a distance of r = 4

lattice sites. Here the repulsive force of range 4 is depicted. Black and gray colors
show the particles before and after the interaction, respectively.

down from the bulk of the droplet. Figure 7.20 illustrates the formation
of the precursor film.
Although the way we introduced the interaction between the particle
is not exact from the point of view of momentum conservation, we see
that this microscopic approach exhibits several key features of the wetting
and spreading phenomena : the formation of a meniscus in capillary tubes
and the existence of a precursor film during the spreading of a droplet.
The trapping boundary condition provides a very intuitive explanation of
the formation of the precursor film. More complex situations, such as
spreading over a rough surface, could be studied easily with the present
approach.

7.2.6 Another force rule

Other techniques have been proposed to represent particle interactions
within a cellular automata model [228] . They are appropriate to implement
both short and long ran¥ e force �. The rule is the following : suppose two
particles (of mometum P1 and Pz) are separated by a distance r, which
is the exact range of the force. Then, these two particles may exchange
some momentum !1P so that

In this way, momentum is exactly conserved. Practically, this force rule

requires matching the pairs of particles at distance r that can effectively
exchange this momentum !1P . If the interaction is repulsive and the two
particles move toward each others, we may say that the particle having
a po sitive component of momentum loses some of it at the expense
of the other particle having a negative component of momentum. If the
interactio n is attractive, it then applies to particles moving away from each
other . Figure 7.21 illustrates this force rule. Finally, note that a crystal-like
 7.2 Wetting, spreading and two-phase fluids 293

interaction can be produced by switching from a repulsive to an attractive

force depending on the distance separating the two interacting particles.

7.2. 7 An Ising cellular automata fluid

In this section we consider a two-phase cellular automata fluid. Each
particle of the fluid can be in two possible states, say s = 1 or s = - 1.
I f w e call this extra degree o f freedom a spin, this fluid can b e compared
with the Ising system we defined in section 2.2.3 but in which, now, the
spins can move according to the rules used in lattice gas models for
hydrodynamics.
In order to produce a surface tension effect at the interface between
the two phases, we shall also introduce an interaction between nearest­
neighbor spins using the same approach as found in classical dynamical
Ising models.
This interaction will be defined through real-valued fields and, thus, the
present model can be viewed as a mixture of a cellular automata rule and
a coupled map lattice dynamics.
It is interesting to remark that a temperature can be defined in a
natural way through the Ising interaction. In addition to being a binary
fluid model, this system has some of the ingredients of a ferrofluid [229] if
the spin is interpreted as the magnetization carried by the particles. Note
that a lattice Boltzmann model for magnetic fluids has also been proposed
in the literature [230] .

The collision rule. Before we define more precisely the spin interaction,
let us return to the particle motion. A collision rule which conserves
mass, momentum and spin can be defined in analogy with the FHP rule
described in section 3.2.
We denote by s i(' r , t ) E { -1, 0, 1 } the state of the automaton along lattice
direction i at site r and time t ( s i = 0 means the absence of a particle).
Clearly the presence of a particle is characterized by sr = 1, regardless
of its spin. Thus, the collision term can be obtained by using sf as an
occupation number.
When a collision takes place, the particles are redistributed among the
lattice directions but the same number of spin + 1 and - 1 particles should
be present in the output state as there were in the input state. A way
to guarantee this spin conservation is to assume that the particles are
distinguishable, at least as far as their spin is concerned.
In figure 7.22 we show the case of a collision between two spin 1
particles and one spin - 1 particle. We assume that all particles bounce
back to where they came from (this is not the only possibility to conserve
momentum and any permutation of the three particles could be considered
294 7 Other models and applications

Fig. 7.22. Three-particle collision in a FHP Ising fluid. The black arrow indicates
a particle with spin + 1 while the gray arrows show particles with spin - 1 .

too). Thus, for such a three-body collision, the state after collision is
given by
si('f + 2c i, t + -r ) =
Si - SiSf+2 sf+4 (1 - sf+ 1 )(1 - Sf+ 3 )(1 - sf+ s )
+si+JSf+l Sf+ s (1 - Sf )(1 - Sf+2 )(1 - Sf+4 )
The identity of the particle appearing or disappearing in direction i is
expressed by the variables s i and s i+3, whereas the presence or absence of
other particles in direction j is indicated by s].
Similarly, by including the two-body collisions, we can express the full
collision term as
si("f + ACi, t + -r ) =
si
-SiSf+2 Sf+4 (1 - Sf+ 1 )(1 - Sf+ 3 )(1 - Sf+ S )
+si+JSf+1 st+ 5 (1 - st )(1 - sf+2 )(1 - st+4 )
-si sf+ 3 (1 - sf+ 1 )(1 - sf+2 )(1 - sf+4 )(1 - sf+ s )
+pqsi+1Sf+4 (1 - st )(1 - sf+2 )(1 - Sf+ 3 )(1 - sf+ s )
+p(1 - q)si+4 st+ 1 (1 - st )(1 - st+2 )(1 - Sf+ 3 )(1 - sf+ s )
+( 1 - p)(1 - q)si+2 st+5 (1 - st )(1 - Sf+t )(1 - sf+ 3 )(1 - st+4 )
+( 1 - p)qsi+ s st+2 (1 - st )(1 - sf+ 1 )(1 - sf+ 3 )(1 - st+4 )
( 7.47 )
where p and q are random Boolean variables that are 1 with probability
1/2, independently at each site and time step. These quantities are
necessary because, with two-body collisions, several collision outputs are
possible and one has to select randomly one of them.

Spin interaction. An important part of this Ising fluid model is the

interaction between spins at the same sites and spins on adjacent lattice
 7.2 Wetting, spreading and two-phase fluids 295

sites. This interaction produces surface tension and can be adjusted

through a parameter which corresponds to the temperature of the system
(which is asumed to be uniform here).
The interaction we propose here does not conserve the number of spins
of each sign. It conserves only the number of particles and, for this reason,
does not represent two different fluids but two possible state of the same
fluid. Of course, the miscibility or immiscibility of the two phases can be
tuned through the temperature.
The updating rule for the spin dynamics is taken from the Monte-Carlo
method [23 1] , using the so-called Glauber transition rule. The main idea is
that a spin flips (changes sign) if it can lower the local energy of the system.
The energy of the pair of spins Si and Sj is computed as E = -J1 sisj if the
two spins are nearest neighbors on the hexagonal lattice and E = - JosiSj
if they are both on the same site (remember that up to six particles can
populate a given site).
However, a spin can flip even if this results in a local increase of energy.
But, then, the change is accepted only with a probability W(s � -s) which
depends on the temperature. In the Glauber dynamics, this probability is
given by [23 1]
1
W(si � -si) = 2 ( 1 - si tanh (Ei))

where Ei is the energy before the flip

1
Ei = (Jomi + J1 Mi) si
kB T
and m i = Lj=/= i Sj is the on-site "magnetization" seen by spin Si and
Mi = L<ji > Sj is the "magnetization" carried by all the particles j on
the neighboring sites of spin i. The quantity T is the temperature and
kB the Boltzmann constant that we can set to 1 when working with an
appropriate temperature scale. When more than one particle are present
at a site, only one of them, chosen at random, is checked for such a spin
flip.
The above transition rule is obtained from the detailed balance condi­
tion, namely
W(si � -si) exp (-E (-si)/(kB T ))
W(-si � si) exp (-E (si)/(kB T))
where E (±si) denotes the Ising energy as a function of Si and has the
properties to drive an ergodic system to thermodynamic equilibrium.
As opposed to the standard the Monte-Carlo approach, where the
lattice sites are visited sequentially and in a random way, here we update
synchronously all the sites belonging to a given sub-lattice. Indeed, for
coherence of the dynamics it is important not to update simultaneously
296 7 Other models and applications

2 0 1 2 0 1 2 0 1 2 0 1
0 1 2 0 1 2 0 1 2 0 1 2
2 0 1 2 0 1 2 0 1 2 0 1
0 1 2 0 1 2 0 1 2 0 1 2

Fig. 7.23. The three sub-lattices on the hexagonal lattice used for the syn­
chrounous spin update. The values 0, 1, 2 label the sites according to the
sub-lattice to which they belong.

any two spins that are neighbors on the lattice. This is for the same reason
as explained in section 2.2.3 when we discussed the Q2R rule.
In a hexagonal lattice, it is easy to see that the space can be partitioned
into three sub-lattices so that all the neighbors of one sub-lattice always
belong to the other two (see figure 7.23).
Therefore, the spin interaction rule described above cycles over these
three sub-lattices and alternates with the FHP particle motion discussed
in the previous section.
It is of course possible to vary the relative frequency of the two rules
(Glauber and FHP). For instance we can perform n successive FHP steps
followed by m successive steps of the Ising rule in order to give more or
less importance to the particle motion with respect to the spin flip. When
n = 0 we have a pure Ising model on a hexagonal lattice but with possibly
a different number of spins per site.
If the temperature is large enough and a periodic boundary imposed,
the system evolves to a configuration where, on average, there are the same
number of particles with spins up and down. Of course, the situation is
not frozen and the particles keep moving and spins continuously flip. As
in a regular Ising systems, there is a critical temperature below which we
can observe a collective effect like a global magnetization (an excess of
particles of a given spin) and the growth of domains containing one type
of spin. This situation is illustrated in figure 7.24 and corresponds to the
case n = m = 1, namely one spin update cycle followed by one step of
FHP motion. It is observed that the critical temperature depends on the
frequency updating n and m.
Another interesting situation corresponds to the simulation of a Raleigh­
Taylor instability. Two immiscible fluids are on top of each other and
the heavier is above the lighter. Due to gravity, the upper fluid wants to
penetrate through the lower one. Since the two fluids are immiscible, the
interface between them becomes unstable and, as time goes on, gives rise
to a mushroom-like pattern.
Gravity can be added to our model in the same way as explained in
 7.2 Wetting, spreading and two-phase fluids 297

Fig. 7.24. Three snapshots of the evolution of the Ising FHP model below the
critical temperature. Particles with spin + 1 are shown in black while gray points
show particles with - 1 . White cells indicate empty sites.

t=50 t=l50

t=300 t=350

Fig. 7.25. Rayleigh-Taylor instability of the interface between two immiscible

fluids. Particles with spin + 1 are shown in black and are "lighter" than gray
particles having spin - 1 . An approximate immisciblity is obtained by choosing a
low temperature in the model.

section 7.2.3. Two immiscible fluids can be modeled by having a low

temperature T in the Glauber dynamics and producing the necessary
surface tension. The upper fluid layer is initialized with only particles of
spin - 1, whereas the lower layer contains only spins + 1 (see figure 7.25).
Gravity is adjusted so that "light" particles go up and heavy particles go
down. After a few iterations, the flat interface destabilizes as shown in the
final panel of figure 7.25.
298 7 Other models and applications
7.3 Multiparticle fluids

In this section we present a method to go beyond the cellular automata

fluid models which have an exclusion principle limiting the number of
particles per lattice direction to one. Although it may be interesting to
increase this limit to two, three or four particles per direction, here we
would like to define a lattice hydrodynamics with an arbitrary number
of particles on each site. Hence, the particle density can vary from zero
to infinity (actually infinity stands for the largest number that can be
represented in the computer).
Our purpose is very similar to that in section 5.7. The advantage of
such an approach is that it reduces considerably the statistical noise that
plagues cellular automata hydrodynamics (such as the FHP model) and,
also, more effectively fine tunes the boundary conditions. More flexibility
is offered because the model variables may vary over a much larger range.
With respect to the lattice Boltzmann fluids discussed in section 3.5,
the present approach will remove the numerical instabilities because it
deals only with integers that can be exactly processed in a computer. In
addition, spontaneous fluctuations will be restored because particle are
indivisible. From what we learned with reaction-diffusion phenomena (see
section 5.4) fluctuations can be important in some reactive processes such
as combustion [2 1 8,21 9] .
The main drawback of a multiparticle fluid model, compared with the
lattice Boltzmann approach, is the performance of the algorithm. Defining
a collision between an arbitrary number of particles which conserve mass
and momentum is not an easy task : particles are indivisible and fractions
of them cannot be distributed among the lattice directions to satisfy the
conservation laws. Furthermore, it is not possible to pre-compute all
possible collisions (as we do in a cellular automaton) because there are
an infinite number of configurations. Thus, more sophisticated algorithms
should be devised which slow down computation of collision output.
There are several requirements we would like to impose on such a
multiparticle model. First, its algorithmic complexity should grow more
slowly than linearly with the number of particles. Otherwise it would be
as fast to run several instances of a standard cellular automata fluid to
obtain the same statistical accuracy on the physical quantities of interest.
Second, we would like to be able to exploit the fact that more types of
collisions are possible when many particles collide, and define a collision
rule that lowers the intrinsic model viscosity. As an example of the new
possiblities to define particle interactions, figure 7.26 shows two examples
of collisions involving several particles traveling along the same link.
Other aspects of multiparticle fluid models can be found in [105,232]
 7.3 Multiparticle fluids 299

1 3
.\2
� - - -

' --I-

\._2 - ==> -

Fig. 7.26. Two examples of multiparticle collisions which conserve mass and
momentum. The values indicated on the arrows give the number of particles
traveling on the corresponding link. It is interesting to note that these collisions
have a non-zero global momentum, as opposed to the usual FHP model.

(note that these models are also referred to as Integer Lattice Gases). Here
we present a multiparticle model for which the algorithmic complexity
grows as the the square root of the number of particles present at a given
site. For this reason, it is expected to eventually win over a basic cellular
automata model as far as the CPU time required to obtained a given
accuracy is concerned. In addition, the rule we propose is rather simple
and allows us to adjust the viscosity with an external parameter.

7.3.1 The multiparticle collision rule

As earlier (see chapter 3), we shall describe the multiparticle micrody­
namics in terms of quantities fi(' r, t ) describing the number of particles
entering site r at time t along lattice direction i and which now take their
value in { 0, 1, 2, ... , oo } . The evolution equation will be, as usual
fi (r + rvi, t + r ) = ffi (f (r, t ))
where v i denotes the elementary particle velocities, r the time between two
consecutive iterations and ff i the collision term. The physical quantities
such as the particle density p and velocity field u are defined as
p(r , t ) = "L fi (r , t) pu(r, t) = L fi (r, t)vi
i
where index i runs over the lattice directions.
300 7 Other models and applications
In order to devise a collision operation f/' i between an arbitrary number
of particles, we shall assume that the main effect of the interaction is to
restore a local equilibrium distribution
b
fi(O) = ap + 2PVi ..
.. u 2 h
· U + pel + P 4 VirxVipUrxUp
v v v
along each lattice direction. Note that fi is an integer whereas f�O) is
a real number. The above expression is the standard form of the local
equilibrium that builds up in a cellular automata fluid (see 3. 1 52 or 3.246 )
and the parameters a, b, e and h should be determined according to
the geometry of the lattice, with the condition that the Navier-Stokes
equation describes the dynamics of the system, and
p(r, t ) = L !�0) 0\ t ) pu(r, t ) = L f�0) (r, t)v i
i i
The relaxation to the local equilibrium is governed by a parameter � .
Thus, a s in the LBGK situation, we shall require that the number of
particles f� leaving (after collision) a given site along direction i be given
by
( 7.48 )
where �fi is a random quantity accounting for the fact that (after collision)
the actual particle distribution may depart from its ideal value.
In practice f� can be obtained as follows. Let N = Li fi the total
number of particles at the given site. Let us denote by Pi the probability
that a particle leaves the site along direction i. We define

where M is a normalization constant adjusted so that Li Pi = 1. Clearly,

if none of the Pi is zero, Li Pi = (I:;d;)/M (since I:: d�o) = I:: di ) and
M = N the total number of particles on the site. From a numerical
analysis, the risk that

z f�O) + ( 1 - z) fi < 0
decreases rapidly as the relaxation time � deviates from the zero viscosity
limit � = 1/2 (see section 3.5 ) . On the other hand, it increases when the
number of particles decreases.
To compute the collision output, we run through each of the N particles
and place them in direction i with probability P i · This will give us a
temporary particle distribution /i which then must be corrected to obtain
f� , in order to ensure exact momentum conservation.
 7.3 Multiparticle fluids 301
Practically, Ji is obtained by choosing a random number s E [0, 1 [ and
comparing it with the quantity
i
so = 0 Si = L Pi
}=1
If si- 1 ::;; s < s i , the particle is placed in direction i (by construction, the
probability that this happens is indeed Pi = si - Si-d· Thus we write
N
Ji = L (si- 1 ::;; sh < si ) ( 7.49 )
h=1
where (si- 1 ::;; sh < s i ) is to be taken as a Boolean value which is 1 when
the condition is true and zero otherwise.
Since sh is a random number uniformly distributed in the interval [0, 1 [,
Ji is also a random quantity. If none of the Pi is zero, its expectation value
lS

!!}0) + ( 1 - !)� fi
N
(Ji ) = L Pi = ( 7.50 )
h=1 �
Note that when N is large enough, equation 7.49 can be computed using
the same Gaussian approximation as used in section 5.7. In this way, the
algorithmic complexity of the operation is independent of the number of
particles N.
While the distribution Ji of output particles obviously conserves the
number of particle, equation 7.50 shows that it only conserves momentum
on the average. Some particles must be pushed to other directions.
Momentum tuning is performed iteratively, according to the following
steps
• At each site where momentum is not correctly gtven by L_1 f/v1,
choose at random one lattice direction i.
• If f1 > 0 move one particle randomly to an adjacent direction.
• Accept the change if it does not worsen the momentum.
• Iterate this procedure until a particle distribution f� is obtained
which satisfies momentum conservation L-1 fj v1 = L- 1 fi v1.
From the way the particles are distributed with the probabilities Pi, we
expect that roughly .j(N ) of them are misplaced. This gives an estimate
of the number of iterations necessary to re-adjust the particles direction.
According to the above discussion it is reasonable to argue that, on
average, the quantity f').ji defined in equation 7.48 vanishes. This fact is
302 7 Other models and applications

confirmed numerically. Consequently, we have

fl r + T V i , t + T ) � z!}0l (1\ t) + ( 1 - z) fi (r, t) ( 7.5 1 )

where we have used fi (r + T V i , t + r ) = f�(r, t ) , due to the definition of

particle motion.
Equation 7.5 1 is identical to the BGK microdynamics, except that now it
approximates a multiparticle dynamics, that is fi are integer quantities. In
the limit where the correlations beween the fi can be neglected (remember
that f}0lis a nonlinear function of all fj ) we may take the average of the
above equation to obtain

(fi (r + TV i , t + r )) = z!}0\ (p) , (pu) ) + ( 1 - z) (h ) ( 7.52 )

If the average distribution (fi (r, t )) is smooth in space and time, it can be
differentiated. Then, the same derivation as that given in section 5.7 can be
repeated to show that a correct hydrodynamical behavior emerges from
our multiparticle model : equation 7.52 is equivalent to the Navier�Stokes
equation with viscosity

where c; is the sound speed whose value is model dependent (different for
hexagonal, square or cubic lattices) [233 ] .
The present muliparticle scheme is intrinsically stable because it is
based on integer numbers and probabilities. No small fluctuation will be
amplified unphysically to make the arithmetic to blow up as is the case
with the LGBK model. In principle, any value of the relaxation parameter
� can be considered. Of course if � :::; 1 /2 it is not clear whether or not a
hydrodynamical behavior is reproduced.

7.3.2 Multiparticle fluid simulations

In this subsection we present some simple results of our multiparticle
lattice gas fluid. We have considered a two-dimensional hexagonal lattice,
including the possibility of having rest particles. Using the standard

{
LBGK expression for such a situation [233] we can write the appropriate
local equilibrium function as f}0l
f(O) = p [ 1 + 2 Vi . U + 8 rx Uf3 (Virx VifJ - 4 U rxf3 ) ]
4 u v2 s:
7
� �

i
/0) = 12[ 1 _ 2 Vurx Uf32 ]
o
!!__
( 7.53 )
2 v
 7.3 Multiparticle fluids 303

Fig. 7.27. Flow instability past a plate. The system size is 80 x 500 (although
only half of it is shown here). The parameters of the simulations are : u00 = 0. 1 5,
� = 0.55 and p = 3500 on average.

The first simulation is shown in figure 7.27. It pictures the flow instability
which develops in a fluid flow past an obstacle (von Karman street).
A second experiment (figure 7.28) shows the velocity time autocorrela­
tion in an equilibrium multiparticle fluid. We consider a periodic system of
size 400 x 400, with a random initial configuration of zero average velocity
and 500 particles per site. The quantity of interest is the velocity-velocity
time correlation
C ( t ) = (v( i\ t)v(r, o ) ) - (v(r, o ) ) 2
where v(r, t ) is defined as
v(r, t) = l: vdi(r, t )
i
This quantity is known to exhibit the so-called long time tails, that
is it obeys a power law [87,234,86] instead of the fast decay exponential
predicted by a linearized Boltzmann equation [235] . The plot in figure 7.28
shows the behavior of

which is one of the basic components of C ( t ) . Here, since the system

in at equilibrium and homogeneous in space, the averaging process is
performed by summing the contribution of all lattice points r. By choosing
(fi ) = 0.0837 (instead of the theoretical value (fi ) = 1 / 1 2 = 0.0835 given
by equation 7.53 with u = 0), we obtain that c ( t ) behaves as
c ( t ) ,...., c 1
as expected for a two-dimensional fluid [234,86] .
Finally, we consider a multiparticle simulation of a Poiseuille flow. Fluid
particles are injected on the left side of a channel of length L and width
W with a rightward velocity. On the upper and lower channel limits, the
usual no-slip condition is imposed, by bouncing back incoming particles
in order to produce a zero speed flow at the boundary. According to
304 7 Other models and applications

- 1 3 L__L __ L__L__L_-L�L_��--��

1 .5 ln (t) 6

Fig. 7.28. Long time tail in the multiparticle fluid model. Simulations are
performed on a 400 x 400 lattice, with an average density of 500 and a relaxation
time � = 1 . The slope of the straight line is - 1 .

x-velocity 0.25

Fig. 7.29. Velocity profile in a multiparticle Poiseuille flow. The plot shows
the horizontal average velocity (ux(Y )) as a function of y the vertical position
between the upper and lower boundaries. The solid line corresponds to the best
parabola fitting the data.
 7.4 Modeling snow transport by wind 305

hydodynamics (see [80] for instance), we expect to observe a parabolic

velocity profile. Figure 7.29 shows the measured velocity profile.

7.4 Modeling snow transport by wind

As mentioned several times in this book, the lattice gas approach makes
it quite natural to introduce complex boundary conditions. It is, for
instance, easy to deal with a dynamically changing boundary. This kind of
problem appears in particular in sedimentation processes where particles
in suspension in a fluid are transported by the flow and, under the effect
of gravity, deposit on the ground. As a result, the shape of the ground
continually changes according to its own dynamics, like for instance the
piling and toppling of a granular material. The fluid flow has of course
to self-adjust to this moving boundary. This change of flow will, in turn,
change the way deposition takes place.
Thus, we see that the problem can be quite intricate and difficult to
formulate in terms of differential equations. Moreover, in addition to
deposition, the whole process also includes erosion phenomena : the fluid
may pick up some of the particles located at the top of the deposit and
transport them downstream again according to various possible mecha­
nisms such as creeping, saltation or suspension.
Such processes take place, for instance, in sand dune formation, sedi­
mentation by water and snowdrift formation. Although complete formu­
lation of the sedimentation problem is difficult, the lattice gas approach
give a very natural framework to propose intuitive rules accounting for the
flow motion, particle transport, erosion and deposition. One of the major
difficulties is to deal (on the same lattice) with the very different orders of
magnitude that exist between the characteristic scales governing the fluid
dynamics and the suspension. But, as usual, the hope is that, provided
a scale separation exist, the correct phenomena can be captured even
though the real scale separation is much larger than the one in the model.
In this section we present such a lattice gas model to describe snow
transport by wind and the mechanisms of deposition and erosion.

7.4. 1 The wind model

The problem of snow transport by wind is important in exposed areas such
as mountain environments. Predicting the location of snow deposits after
a windy period is crucial to protect roads against snowdrifts or control the
formation of wind-slab (responsible for about 80% of accidentally caused
avalanches). Placing snow fences at appropriate locations can substantially
modify the shape of the deposit, by screening out the wind. This technique
306 7 Other models and applications
is widely used in practice to control and prevent overaccumulation of snow
at critical spots or, in contrast, to store snow on ski trails.
From the simulation point of view, such a problem has two main
components. First, there is the fluid (the wind here) and, second, the
particles (the snow flakes) that are transported. We shall consider here a
two-dimensional system, for the sake of simplicity.
From what we have learned so far, a lattice Boltzmann model like the

104
one discussed in section 3.5 (that is a LBGK model on a square lattice

106
with diagonals) is the natural choice to simulate wind flow. Indeed, our
aim is to simulate high Reynolds number flows such as R > for
small-scale modeling and up to for large-scale situations such as a
mountain area.

0 0
The main difference from the situation described in section 3.5 is that
now we shall include rest particles. As a consequence, the lattice directions
will be labelled from i = to i = 8, where i = refers to the rest particle
population (in particular the associated velocity vo is zero). This model is
often referred to as the D2Q9 model [233] (2 dimensions and 9 particle
populations).
The physical quantities of interest are
• the densityp = 2:::�=0 h
• and the velocity field u = � I:�=l fivi.
The lattice Boltzmann dynamics is expressed by the usual equation

flr + TVi, t + r)
fi(r, t) + z (f7q(u(r, t), p(r, t)) - f; (r, t))
=

where � is the relaxation time and f7q ( u , p) is the equilibrium distribu­

tion [233]

where
1
cs2 = 3
is the speed of sound and the quantity ti are
4
to = -
9
for the rest particles,
 7.4 Modeling snow transport by wind 307

for the horizontal and vertical particles and, finally

1
t2 = t4 = t6 = ts = -
36
for the diagonal particles.
The shear viscosity v is given in terms of the relaxation parameter �, as
in section 3.5, according to the relation
2� - 1
v= ---
6
Decreasing the relaxation time too much leads to instability and the
limit of high Reynolds number (i.e v ;::::; 0) cannot be attained. The solution
is to consider what is called a subgrid model. This is a frequent approach
in standard CFD resolution techniques. The idea is to assume that an
effective viscosity results from the unresolved scales, that is scales below
the lattice spacing A.
Our goal is not to give a theoretical discussion of subgrid models but,
rather, to adopt a pragmatic approach and show how to introduce eddy
viscosity in a lattice Boltzmann model, following S. Hou et al.'s work [104] .
Practically, in a subgrid model, one adapts the relaxation time locally,
according to the value of local strain tensor

Sap = 21 ( 8p ua + 8a up )
Instability appears where the magnitude I SF I = J2SapSap of this tensor is
large.
One common subgrid model is the Smagorinski model. It amounts to
correcting the relaxation time as
t = � + 3 C]magoA.2 I SF I
where Csmago > 0 is the so-called Smagorinsky constant (in practice, a
parameter of the model set, for instance, to 0 . 3 ) .
It turns out that the magnitude of the tensor Sap can be computed
locally, without taking extra derivatives, just by considering the nonequi­
librium momentum tensor
rr;}l = Li V ia V ip(fi - f�q )
Then, the quantity I SF I is obtained directly as
-� + �2 + 18A.2C}mago rr;}Jrr;}J
ISF I = -- -'----
-
6 A_ 2 C}mago
--=-----=------

Therefore the bare viscosity v is transformed to

v ' = v + Vt
308 7 Other models and applications

with Vt the Smagorinski eddy viscosity, given by

Vt = clmago A2 1 SF I

7.4.2 The snow model

On the same lattice described in the previous subsection, we now define
another kind of particle Si ( r , t ) (S stands for snow, sand or solid). These
particles interact with the wind particles and have a set of local evolution
rules to account for deposition, collisions, cohesion, erosion and gravity
effects.
All the processes pertaining to sedimentation and erosion are not fully
understood, at least for the case of snow. To model these phenomena,
we follow the cellular automata philosophy of keeping only the key
microscopic interactions, hoping that they will be sufficient to capture the
macroscopic behavior of the system.
On each lattice site, snow distribution is represented by Si ( r , t ) , i E
0, . . . , 8. Si are integer values equal to the number of particles moving in
the ith direction (So for the particles remaining motionless). Therefore,
snow is described with a multiparticle-like dynamics.
In real life, the influence of wind during the snowfall is important, but
transport after the fall is essential for re-shaping of the deposit. It is
assumed that the snow transport can be split into three main processes :

• creeping of particles along the ground, consisting of very small jumps

of the snow particles ;

• saltation, which occurs mainly in the first 30 centimeters above the

surface. This process is essential for snow transport ;

• suspension or turbulent diffusion / dispersion of particles. Particles are

caught in the saltation layer by strong eddies and can be transported
over long distances. This process is often seen as "smoke" above the
mountain crests.

Unlike other numerical models [236,237] , the aim of our lattice model is
not to treat these processes (mainly saltation and suspension) separately.
By setting reasonable microscopic rules we expect to naturally recover the
macroscopic reality as the emergence of a complex behavior.
These rules governing the snow particles and their interaction with
wind particles have several components : transport mechanisms, including
gravity, erosion and, finally deposition rules. They are now described in
more detail.
 7.4 Modeling snow transport by wind 309

Transport by wind. Neglecting the inertial effects, we may assume that a

r
snow particle follows the wind velocity field u( , t ) and experiences gravity.
Due to friction, the main effect of gravity is to impose vertical falling
speed -ug to the particle in suspension. Thus, to a first approximation, a
r
flake at site will move, after • units of time, to
r( t + •) =
r ( t ) + (u - ug ) •
r
Unfortunately, the quantity ( t + • ) defined in this way no longer has any
reason to correspond to a lattice site. Thus, it is no longer possible to use
a lattice gas approach to describe the motion of snow particles if we do
not use a trick.
One possibility, however, is to view the particle transport as a proba­
bilistic process. We write
r( t + •) =
r ( t ) + vi •
where V i is one of the nine velocity vectors associated with the D 2Q9
square lattice (see figure 3 . 1 5 and remember that vo = 0). For each
particle motion, one of the vi is chosen so that, on average
r
( ( t + • ) - r ( t ) } = (u - ug ) •
Practically, we can proceed as follows. We first assume that I ii - ug I < v
where v = (A./ • ) is the maximum speed pemitted on the lattice. Then we
introduce four probabilities of motion Px. P-x. py, P-y corresponding to
motion along the x- or y-axis, in the plus or minus direction. We define

Px = max 0,
( v 1 · (u - ug )
2 v
)
P-x = max 0,
vs (u - u )
2
( ·

v g
)
Py = max
(0, v3 · (u - ug )
v2
) P-y = max
(0, v1 · (u - ug )
v2
)
Since v 1 = -vs and v 3 = -v7 one cannot have Px and P-x simultaneously
non-zero (and similarly for Py and P-y ). On the other hand, we may have
any combination of x- and y-motion.
To select the direction of motion of a particle with this approach, two
random numbers qx and qy, uniformly distributed in the interval [0, 1],
are chosen. If qx < Px (or qx < P-x) the particle will move along the
positive x-direction (or the negative x-direction). Otherwise, if qx is larger
than both Px and P-x• no motion along the horizontal axis is permitted.
The same algorithm is applied to qy for the y-axis motion. When both
the x- and y-axis are selected, the particle moves with velocity v2 , V4, V6 or
vs, according to the sense of motion.
The average velocity (v } of the particle is then
(v } = Px Vt + P-x Vs + Py V3 + P-y V7
3 10 7 Other models and applications

Fig. 7.30. Example of particle transport in a fixed, circular velocity field. The
lattice contains 128 x 128 sites and a particle source is set in the middle of
the small circle. The particles follow their · path with a dispersion due to the
probabilistic algorithm.

V3 -v7,
Since Px and P- x are incompatible (and so are py, P-y) and v1 = -vs,
V = (vl · (uv - ug)) V...l (v3 · (uv - ug)) V3 = U -Ug
= we may as well write
( ... ) ... ...
2 2
-+

v1 v3
since and are orthogonal. This last relation shows that, although the

the desired velocity flowu -ug.

particles in suspension move on the lattice, their average position follows

As we may have many snow particles at a site, the above procedure

should be repeated for each of them. Of course, it is possible to speed up
this phase by using the Gaussian approximation described in section 5.7,

possible velocities Vj.

and know in a few steps how many particles will move with each of the

An example of such a particle transport over a lattice is shown in fig­

ure 7.30. We observe an intrinsic particle dispersion due to the probabilitic
nature of the algorithm. This dispersion, which is significant over a long
distance, can be given a physical content such as, for instance, the effects
of microscopic fluctuations present in the system.

Deposition. Deposition is a crucial ingredient of our model. Particles

in suspension eventually reach the bottom of the system, which follows
the ground profile. As they fall down, particles pile up and modify the
shape of the terrain. In our model, a lattice site can either be solid (like
"hard" soil or deposited snow) or free (where fluid and snow particles
can move). When a snow particle should move to a neighbor solid site, it
remains in position as a result of being solidified. Since we deal with a
 7.4 Modeling snow transport by wind 311

Fig. 7.3 1 . Four sequential stages in a snow deposit simulation. The wind is
blowing from left to right and the presence of a fence (with a ground clearance)
is shown by the vertical line segment. Light gray regions indicate solid snow (the
deposit) and white dots show the flying snow (particles in suspension). In order
to feed the system with snow, a source of snow is placed at the lower left corner
of the system.

Fig. 7.32. Three sequential stages of snow deposit in a trench. A source of snow
is placed on the left side of the system.

multiparticle snow model, a site will become a solid site only after it has
received more than a given threshold number of particles. This technique,
which takes into account some cohesion between snow particles, makes
the use of complex piling-toppling rules (such as described in section 2.2.6)
unnecessary.

Erosion. The erosion process is very complex and not yet fully understood.
Deposited particles can be picked up again by the wind, but the physical
conditions that are necessary to produce this and the interactions that are
involved are not easy to identify [238] . Some empirical laws exist that
express the erosion rate in terms of the local properties of the system, such
as the wind speed at the erosion surface. Several rules can be considered
in the lattice gas approach to reproduce qualitatively the experimental
observations. The basic erosion mechanism we use is that a solid snow
particle is ejected vertically with a random speed when the appropriate
erosion conditions are met.
312 7 Other models and applications

7.4.3 Simulations of snow transport

In this section we present two numerical simulations of our lattice model
for snow transport by wind. The first situation is to obtain the deposition
of snow around a fence. This is a case frequently studied in wind tunnels
and which has many practical applications. The purpose of placing a
fence is to screen out the wind and force the snow to accumulate on the
leeward site of the fence instead of flying further away. Usually, such
a fence has a bottom gap which allows a continual wind flow below it.
This prevents the snow from accumulating at this place and thus rapidly
burying the fence, which would make it completely ineffective.
Figure 7.3 1 shows several stages of the deposit formation. The snow
deposit grows in a non-regular way and the wind pattern evolves as the
accumulation becomes larger. It is also interesting to note the presence of
a smaller deposit in front of the fence, due to eddies forming windward
of the fence. These simulations are very close to wind tunnel experiments.
As a second example, we consider in figure 7.32 the filling of a trench
due to snow transport during a wind episode. The way the various layers
appears as time goes on is not easy to predict since the wind pattern
in the cavity changes as snow deposits in the trench. The results of the
simulation are in good qualitative agreement with real experiments.
 References

[1] R. Livi, S. Ruffo, S. Ciliberto, and M. Buiatti, editors. Chaos and

Complexity. World Scientific, 1 988.
[2] I.S.I and R. Monaco, editors. Discrete Kinetic Theory, Lattice Gas
Dynamics and Foundations of Hydrodynamics. World Scientific, 1 989.
[3] G. Doolen, editor. Lattice Gas Method for Partial Differential Equations.
Addison-Wesley, 1 990.
[4] P. Manneville, N. Boccara, G.Y. Vichniac, and R. Bideau, editors.
Cellular Automata and Modeling of Complex Physical Systems. Springer
Verlag, 1 989. Proceedings in Physics 46.
[5] A. Pires, D.P. Landau, and H. Herrmann, editors. Computational Physics
and Cellular Automata. World Scientific, 1 990.
[6] J.M. Perdang and A. Lejeune, editors. Cellular Automata : Prospect
in Astrophysical Applications. World Scientific, 1 993.
[7] Minnesota IMA cellular automata bibliography
http :/ jwww.ima.umn.edu/bibtexjca.bib.
[8] Santa-Fe cellular automata bibliography.
ftp :/ /alife.santafe.edujpub/topicsjcasjca-faq.bib.
[9] T. Toffoli D. Farmer and S. Wolfram, editors. Cellular Automata,
Proceedings of an Interdisciplinary Workshop, volume 10. Physica D,
North-Holland, 1 984.
[10] J.-P. Boon, editor. Advanced Research Workshop on Lattice Gas Automata
Theory, Implementations, and Simulation, J. Stat. Phys, 68 (3/4), 1 992.
[1 1] S. Ulam. Random processes and transformations. Proc. Int. Congr.
Math., 2 :264-275, 1952.
[1 2] A.W. Burks. Von Neumann's self-reproducing automata. In A.W. Burks,
editor, Essays on Cellular Automata, pp. 3-64. University of Illinois Press,
1970.
[1 3] U. Pesavento. An implementation of von Neumann's self-reproducing
machine. Artificial Life, 2 :337-354, 1 995.

313
3 14 References
[14] A. Reggia, S.L. Armentrout, H.-H. Chou, and Y. Peng. Simple systems
that exhibit self-directed replication. Science, 259 : 1 282, 1 993.
[1 5] E.F. Codd. Cellular Automata. Academic Press, 1 968.
[1 6] C.G. Langton. Self-reproduction in cellular automata. Physica D,
10 : 1 35�144, 1 984.
[1 7] J. Byl. Self-reproduction in small cellular automata. Physica D,
34 :259�299, 1 989.
[1 8] C.G. Langton, C. Taylor, J.D. Farmer, and S. Rasmussen, editors.
Artificial Life II. Addison-Wesley, 1 992.
[1 9] C. G. Langton. Editor's Introduction. Artificial Life, 1 :v�viii, 1 994.
[20] M. Gardner. The fantastic combinations of John Conway's new solitaire
game life. Scientific American, 220(4) : 1 20, 1 970.
[2 1] E.R. Berlekamp, J.O. Conway, and R.K. Guy. Winning Ways for your
Mathematical Plays, volume 2. Academic Press, 1 982. Chapter 25.
[22] M. Gardner. Wheels, Life and Other Mathematical Amusements.
Freeman, 1983.
[23] K. Preston and M. Duff. Modern Cellular Automata : Theory and
Applications. Plenum Press, 1984.
[24] S. Wolfram. Theory and Application of Cellular Automata. World
Scientific, 1986.
[25] S. Wolfram. Cellular Automata and Complexity. Addison-Wesley,
Reading MA, 1 994.
[26] T. Toffoli and N. Margolus. Cellular Automata Machines : a New
Environment for Modeling. The MIT Press, 1 987.
[27] Information Mechanics Group. CAM8 : a parallel, uniform, scalable
architecture for CA Experimentation. http :/ jwww.im.lcs.mit.edujcam8.
[28] J. Hardy, Y. Pomeau, and 0. de Pazzis. Molecular Dynamics of a
classical lattice gas : Transport properties and time correlation functions.
Phys. Rev. A, 13 : 1 949�60, 1976.
[29] J.C. Maxwell. Scientific Paper II. Cambridge University Press, 1 890.
[30] H. Cornille. Exact (2+ 1 )-dimensional solutions for two discrete velocity
Boltzmann models with four independent densities. J. Phys. A,
20 :L1063�67, 1 987.
[3 1 ] J.E. Broadwell. Shock structure i n a simple discrete velocity gas. Phys.
Fluids, 7 : 1 243, 1 964.
[32] U. Frisch, B. Hasslacher, and Y. Pomeau. Lattice-gas automata for the
Navier�Stokes equation. Phys. Rev. Lett., 56 : 1 505, 1 986.
[33] S. Wolfram. Cellular automaton fluid : basic theory. J. Stat. Phys.,
45 :47 1 , 1 986.
 References 315

[34] D. d'Humieres, Y. Pomeau, and P. Lallemand. Simulation of 2-D von

Karman streets using a lattice gas. C.R. Acad. Sci. Paris, II, 301 : 1 39 1 ,
1 9 8 5 . Reprinted i n Lattice Gas Methods for Partial Differential
Equations, ed. G. Doolen, p.423, Addison-Wesley, 1 990.
[35] P. Calvin, D. d'Humieres, P. Lallemand, and Y. Pomeau. Cellular
automata for hydrodynamics with free boundaries in two and three
dimensions. C.R. Acad. Sci. Paris, II, 303 : 1 1 69, 1 986. Reprinted in
Lattice Gas Methods for Partial Differential Equations, ed. G. Doolen,
p. 4 1 5, Addison-Wesley, 1990.
[36] D. d'Humieres, P. Lallemand, and U. Frisch. Lattice gas models for
3D hydrodynamics. Europhys. Lett., 2 :29 1 , 1 986.
[37] D. d'Humieres and P. Lallemand. Lattice gas models for 3D
hydrodynamics. Europhys. Lett., 2 :29 1 , 1 986.
[38] S.A. Orszag and V. Yakhot. Reynolds Number Scaling of Cellular
Automata Hydrodynamics. Phys. Rev. Lett, 56 : 1 69 1-1 693, 1 986.
[39] S. Chen, K. Diemer, G.D. Doolen, K. Eggert, C. Fu, S. Gutman,
and B.J. Travis. Lattice Gas Automata For Flow Through Porous Media.
Physica D, 47 :72-84, 1 99 1 .
[40] E. Aharonov and D. Rothman. Non-Newtonian flow (through porous
media) : a lattice Boltzmann method. Geophys. Res. Lett., 20 :679-682,
1 993.
[41] D.W. Grunau, T. Lookman, S.Y. Chen, and A.S. Lapedes. Domain
growth, wetting and scaling in porous media. Phys. Rev. Lett.,
71 :4198-4201 , 1 993.
[42] D.H. Rothman. Immiscible lattice gases : new results, new models.
In P. Manneville, N. Boccara, G.Y. Vichniac, and R. Bideau, editors,
Cellular Automata and Modeling of Complex Physical Systems,
pp. 206-23 1 . Springer Verlag, 1 989. Proceedings in Physics 46.
[43] M. Bonetti, A. Noullez, and J.-P. Boon. Lattice gas simulation of 2-d
viscous fingering. In P. Manneville, N. Boccara, G.Y. Vichniac, and
R. Bideau, editors, Cellular Automata and Modeling of Complex Physical
Systems, p. 239-24 1 . Springer Verlag, 1 989. Proceedings in Physics 46.
[44] A.K. Gunstensen, D.H. Rothman, S. Zaleski, and G. Zanetti. Lattice
Boltzmann Model of Immiscible Fluids. Phys. Rev. A, 43 :4320-4327,
1 99 1 .
[45] D. Grunau, Shiyi Chen, and K. Eggert. A Lattice Boltzmann Model
for Multi-phase Fluid Flows. Phys. Fluids A, 5 :2557-2562, 1 993.
[46] D. Rothman and S. Zaleski. Lattice-Gas Cellular Automata : Simple
Models of Complex Hydrodynamics. Collection Alea. Cambridge
University Press, 1 997.
316 References
[47] U. D'Ortona, M. Cieplak, R.B. Rybka, and J.R. Banavar. Two-color
nonlinear cellular automata : surface tension and wetting. Phys. Rev. E,
51 :37 1 8-28, 1 995.
[48] B. Boghosian, P. Coveney, and A. Emerton. A lattice-gas model of
microemulsions. Proceedings of the Royal Society of London,
452 : 1 22 1-1250, 1 996.
[49] J.T. Wells, D.R. Janecky, and B.J. Travis. A lattice gas automata model
for heterogeneous chemical reaction at mineral surfaces and in pores
network. Physica D, 47 : 1 1 5-123, 199 1 .
[50] J.-P. Boon, D. Dab, R. Kapral, and A. Lawniczak. Lattice Gas Automata
for Reactive Systems. Phys. Rep., 273 :55-148, 1 996.
[5 1] G.G. McNamara and G. Zanetti. Use of the Boltzmann Equation to
Simulate Lattice-Gas Automata. Phys. Rev. Lett., 61 :2332-2335, 1988.
[52] F. Higuera, J. Jimenez, and S. Succi. Boltzmann approach to lattice gas
simulations. Europhys. Lett, 9 :663, 1 989.
[53] D.B. Bahr and J.B. Rundle. Theory of lattice Boltzmann simulation of
glacier flow. J. Glaciology, 41(1 39) :634--40, 1 995.
[54] E. Banks. Information processing and transmission in cellular automata.
Technical report, MIT, 197 1 . MAC TR-8 1 .
[55] Thinking Machines Corp. CM Fortran Libraries Reference Manual,
January 1 994. Version 2. 1 , p.28.
[56] R. J. Gaylord and K. Nishidate. Modeling Nature with Cellular Automata
using Mathematica. Springer-Verlag, 1996.
[57] K. Culick II and F. Yu. Undecidability of CA classification scheme.
Complex System, 2 : 1 77-190, 1988.
[58] J. Hemmingsson. A totalistic three-dimensional cellular automaton with
quasiperiodic behavior. Physica A, 183 :255, 1 992.
[59] H. Chate and P. Manneville. Collective behaviors in spatially extended
systems with local interactions and synchronous updating. Prog. Theor.
Phys., 87 : 1-67, 1 992.
[60] J. Krug and H. Spohn. Universality classes for deterministic surface
growth. Phys. Rev. A, 38 :427 1-83, 1988.
[6 1] K. Chen P. Bak and C. Tang. A forest-fire model and some thoughts
on turbulence. Phys. Lett. A, 147 :297, 1 990.
[62] B. Drossel and F. Schwab!. Self-Organized Critical Forest-Fire Model.
Phys. Rev. Lett., 69 : 1 629, 1 992.
[63] P. Rujim. Cellular Automata and Statistical Mechanical Models.
J. Stat. Phys., 49 : 1 39, 1 987.
[64] G. Vichniac. Simulating physics with cellular automata. Physica D,
10 :96-1 1 5, 1984.
 References 317
[65] J.D. Gunton and M. Droz. Introduction to the Theory of Metastable and
Unstable States. Springer Verlag, 1983.
[66] J. Rajechenbach, E. Clement, J. Duran, and T. Mazozi. Experiments on
bidimensional Models of sand : study of the dynamics. In J. Vannimenius
A. McKane, M. Droz and D. Wolf, editors, Scale Invariance and
Non-equilibrium Dynamics, pp. 3 1 3-27. NATO ASI Series, Plenum Press,
1 995.
[67] E. Guyon and J.-P. Troadec. Du sac de bille au tas de sable. Edition
Odile Jacob, 1 994.
[68] H.M. Jaeger and S.P. Nagel. La physique de l'etat granulaire.
La Recherche, 23(249) : 1 380--87, December 1 992.
[69] D.E. Wolf, M. Schreckenberg, and A. Bachem, editors. Traffic and
Granular Flow. World Scientific, 1 996.
[70] A. Karolyi and J. Kertesz. Hydrodynamics cellular automata for
granular media. In R. Gruber and M. Tomassini, editors, Proceeding of
the 6th Joint EPS-APS International Conference on Physics Computing :
PC '94, pp. 675-68 1 , 1 994.
[7 1] I. Stewart. The ultimate in anty-particle. Scientific American, pp. 88-9 1 ,
July 1 994.
[72] J. Propp. Trajectory of generalized ants. Math. Intelligencer, 16( 1 ) :37-42,
1 994.
[73] 0. Biham, A.A. Middleton, and D. Levine. Self-organization and
dynamical transition in traffic-flow models. Phys. Rev. A, 46 :R6 1 24, 1992.
[74] K. Nagel and H.J. Herrmann. Deterministic models for traffic jams.
Physica A, 199 :254, 1 993.
[75] M. Kikuchi S. Yukawa and S. Tadaki. Dynamical phase transition in
one-dimensional traffic flow model with blockage. J. Phys. Soc. Jpn,
63( 1 0) :3609-3 6 1 8, 1 994.
[76] B. Chopard. A Cellular Automata Model of Large Scale Moving
Objects. J. Phys. A, 23 : 1 67 1-1 687, 1 990.
[77] M. Sipper. Co-evolving Non-Uniform Cellular Automata to Perform
Computations. Physica D, 92 : 1 93-208, 1 996.
[78] B. Chopard and M. Droz. Cellular automata model for heat conduction
in a fluid. Phys. Lett. A, 126 :476, 1988.
[79] P. Grosfils, J.-P. Boon, and P. Lallemand. Spontaneous fluctuation
correlation in thermal lattice gas automata. Phys. Rev. Lett., 68 : 1077,
1 992.
[80] D.J. Triton. Physical Fluid Dynamics. Clarendon Press, 1988.
[8 1] K. Molvig, P. Donis, R. Miller, and G. Vichniac. Multi-species
Lattice-Gas Automata for Realistic Fluid Dynamics. In P. Manneville,
N. Boccara, G.Y. Vichniac, and R. Bideau, editors, Cellular Automata and
318 References
Modeling of Complex Physical Systems, pp. 206-23 1 . Springer Verlag,
1 989. Proceedings in Physics 46.
[82] M. Hi�non. Viscosity of a lattice gas. Complex Systems, 1 :763-789, 1 987.
Reprinted in Lattice Gas Methods for Partial Differential Equations, ed.
G. Doolen, pp. 1 79, Addison-Wesley, 1 990.
[83] K. Diemer, K. Hunt, S. Chen, T. Shimomura, and G. Doolen. Density
and velocity dependence of Reynolds numbers for several lattice gas
models. In G. Doolen, editor, Lattice Gas Methods for Partial
Diffe rential Equations, pp. 1 37-1 77. Addison-Wesley, 1 990.
[84] L.P. Kadanoff, G.R. McNamara, and G. Zanetti. From automata to fluid
flow : comparison of simulation and theory. Phys. Rev. A, 40 :4527-4541 ,
1989.
[85] N. Margolus, T. Toffoli, and G. Vichniac. Cellular automata
supercomputers for fluid-dynamics modeling. Phys. Rev. Lett.,
56 : 1 694--1 696, 1 986.
[86] D. Frenkel. Long-time decay of velocity autocorrelation of
two-dimensional lattice gas cellular automata. In P. Manneville,
N. Boccara, G.Y. Vichniac, and R. Bideau, editors, Cellular Automata
and Modeling of Complex Physical Systems, pp. 144--1 54. Springer
Verlag, 1 989. Proceedings in Physics 46.
[87] R. Balescu. Equilibrium and Nonequilibrium Statistical Mechanics.
John Wiley & Sons, 1 975.
[88] U. Frisch, D. d'Humieres, B. Hasslacher, P. Lallemand, Y. Pomeau, and
J.-P. Rivet. Lattice Gas Hydrodynamics in two and Three Dimension.
Complex Systems, 1 :649-707, 1987. Reprinted in Lattice Gas Methods for
Partial Differential Equations, ed. G. Doolen, p. 77, Addison-Wesley,
1 990.
[89] M. A. van der Hoef. Introduction to lattice gas cellular automata as a
model for simple fluids. In J.M. Pergang and A. Lejeune, editors, Cellular
Automata : Prospects in Astrophysical Applications, pp. 87-1 18. World
Scientific, 1 993.
[90] B. Chopard and M. Droz. Cellular automata approach to nonequilibrium
correlation functions in a fluid. Helvetica Physica Acta, 61 :893-6, 1988.
[9 1] P. Resibois and M. Leener. Classical Kinetic Theory of Fluids. John
Wiley, 1 977.
[92] G. Zanetti. Hydrodynamics of lattice-gas automata. Phys. Rev. A,
40 : 1 539-1 548, 1 989.
[93] T. Naitoh, M.H. Ernst, and J.M. Dufty. Long-time tails in
two-dimensional cellular automata fluids. Phys. Rev. A, 42 :71 87, 1 990.
[94] R. Brito and M.H. Ernst. Propagating Staggered Wave in Cellular
Automata Fluids. J. Phys. A, 24 :333 1 , 1 99 1 .
 References 319
[95] D. d'Humieres, Y.H. Quian, and P. Lallemand. Invariants in Lattice gas
Models. In I.S.I and R. Monaco, editors, Discrete Kinetic Theory, Lattice
Gas Dynamics and Foundations of Hydrodynamics, pp. 102-1 1 3. World
Scientific, 1 989.
[96] R. Benzi, S. Succi, and M. Vergassola. The lattice Boltzmann equation :
theory and application. Phys. Rep., 222(3) : 145-1 97, 1 992.
[97] A. Gustensen, D.H. Rothman, S. Zaleski, and G. Zanetti. Lattice
Boltzmann model of immiscible fluids. Phys. Rev. A, 43 :4320--4 3 27, 199 1 .
[98] S. Succi, M. Vergassola, and R. Benzi. Lattice-Boltzmann scheme for
two-dimensional magnetohydrodynamics. Phys. Rev. A, 43 :4521, 1 99 1 .
[99] F.J. Alexander, S. Chen, and J.D. Sterling. Lattice Boltzmann
thermohydrodynamics. Phys. Rev. E, 47 :2249-2252, 1 993.
[100] F. Higuera, J. Jimenez, and S. Succi. Lattice gas dynamics with enhanced
collision. Europhys. Lett, 9 :345, 1989.
[101] Hudong Chen, Shiyi Chen, and W.H. Matthaeus. Recovery of
Navier-Stokes equations using a lattice-gas Boltzmann method. Phys.
Rev. A, 45 :R5339--42, 1 992.
[102] Y.H. Quian, D. d'Humieres, and P. Lallemand. Lattice BGK Models for
Navier-Stokes Equation. Europhys. Lett, 17(6) :470--8 4, 1 992.
[103] P. Bhatnager, E.P. Gross, and M.K. Krook. A model for collision process
in gases. Phys. Rev., 94 : 5 1 1 , 1954.
[104] S. Hou, J. Sterling, S. Chen, and G.D. Doolen. A lattice Boltzmann
subgrid model for high Reynolds number flows. Fields Institute
Communications, 6 : 1 5 1-1 66, 1996.
[105] R. Chatagny and B. Chopard. Multiparticle lattice gas models for
hydrodynamics. In J.-M. Alimi, A. Serna, and H. Scholl, editors, Science
on the Connection Machine System : Proceedings of the Second European
CM Users Meeting 1 993, pp. 239--47. Thinking Machines Corporation,
1 995.
[106] G.H. Weiss. Random walks and their applications. American Scientist,
71 :65, 1983.
[107] G.H. Weiss, editor. Contemporary Problems in Statistical Physics. SIAM,
1994.
[108] G.I Taylor. Diffusion by continuous movements. Proc. Lond. Math. Soc,
20 : 1 96--212, 1 92 1 .
[109] E.H. Hauge. What can we learn from Lorentz models ? In G. Kirczenow
and J. Marro, editors, Transport Phenomena, p. 337. Springer, 1974.
[1 10] N.G. van Kampen. Stochastic Processes in Physics and Chemistry.
North-Holland, 1 98 1 .
[1 1 1] X.P. Kong and E.G.D. Cohen. Anomalous diffusion in lattice-gas
wind-tree model. Phys. Rev. B, 40 :4838, 1 989.
320 References
[1 12] B. Chopard and M. Droz. Cellular automata model for diffusion
processes. J. Stat. Phys., 64 :859-892, 1 99 1 .
[1 1 3] J.D. Jackson. Classical Electrodynamics. John Wiley, 1 975.
[1 14] A.S. Nowick and J.J. Burton. Diffusion in Solids. Academic Press, 1 978.
[1 1 5] J. Philibert. Diffusion et Transport de Matiere dans les Solides. Les
Editions de Physique, Les Ulis, 1985.
[1 16] W. Dieterich, P. Fulde, and I. Peschel. Theoretical models for superionic
conductors. Adv. Phys., 29 :527, 1 980.
[1 1 7] J.P. Hulin, E. Clement, C. Baudet, J.-F. Gouyet, and M. Rosso.
Qualitative analysis of an invading-fluid invasion front under gravity.
Phys. Rev. Lett., 61 :333, 1988.
[1 1 8] B. Chopard, M. Droz, and M. Kolb. Cellular automata approach to
nonequilibrium diffusion and gradient percolation. J. Phys. A, 22 : 1 609,
1 989.
[1 19] T. Vicsek. Fractal Growth Phenomena. World Scientific, 1 989.
[1 20] T.A. Witten and L.M. Sander. Diffusion-limited aggregation. Phys. Rev.
B, 27 :5686, 1983.
[1 2 1] S. Tolman and P. Meakin. Off-lattice and hypercubic-lattice models for
diffusion-limited aggregation in dimension 2-8. Phys. Rev. A, 40 :428-37,
1 989.
[ 1 22] J.W. Evans. Random and cooperative sequential adsorption. Reviews of
Modern Physics, 65 : 1 28 1-1 329, 1 993.
[123] I. Langmuir. Adsorption and gases on glass, mica and platinium. J. Am.
Chern. Soc., 40 : 1 36 1 , 1 9 1 8 .
[1 24] P. Schaaf, A. Johner, and J. Talbot. Asymptotic behavior of particle
deposition. Phys. Rev. Lett., 66 : 1 603-1 605, 1 99 1 .
[125] G. Tarjus and P. Viot. Generalized Car-Parking Problem as a Model
Particle Deposition with Entropy-Activated Rate Process. Phys. Rev.
Lett., 68 :2354, 1 992.
[1 26] B. Senger, P. Schaaf, J.-C. Voegel, A. Johner, A. Schmitt, and J. Talbot.
Influence of bulk diffusion on the adsorption of hard spheres on a flat
surface. J. Chern. Phys., 97 : 3 8 1 3-3820, 1 992.
[127] B. Senger, J. Talbot, P. Schaaf, A. Schmitt, and J.-C. Voegel. Effect of the
Bulk Diffusion on the Jamming Limit Configurations for Irreversible
Adsorption. Europhys. Lett., 21 : 1 3 5-140, 1 993.
[1 28] B. Chopard, P. Luthi, and M. Droz. Microscopic approach to the
formation of Liesegang patterns. J. Stat. Phys., 76 :66 1-677, 1 994.
[1 29] J. E. Pearson. Complex Patterns in a Simple System. Science,
261 : 1 89-192, July 1 993.
[130] J.D. Muray. Mathematical Biology. Springer-Verlag, 1 990.
 References 321

[1 3 1] L.M. Brieger and E. Bonomi. A stochastic cellular automaton model of

nonlinear diffusion and diffusion with reaction. J. Comp. Phys.,
94 :467-486, 1 99 1 .
[1 32] E.J. Garboczi. Permeability, Diffusivity and Microstructural Parameters :
a Critical Review. Cement and Concrete Res., 20 :59 1-601 , 1 990.
[1 3 3] R. Kapral and K. Showalter, editors. Chemical Waves and Patterns.
Kluwer Academic, 1 995.
[1 34] J.P. Keener and J.J. Tyson. The dynamics of scroll waves in excitable
media. SIAM Rev., 34 : 1-39, 1 992.
[1 3 5] E.E. Selkov. Self-oscillation in glycolysis : a simple kinetic model.
Eur. J. Biochem., 4 :79, 1 968.
[1 3 6] R. Fisch, J. Gravner, and D. Griffeath. Threshold-range scaling of
excitable cellular automata. Statistics and Computing, 1 :23, 1 9 9 1 .
[1 37] J. Gravner and D. Griffeath. Threshold grouse dynamics. Trans. Amer.
Math. Soc., 340 :837, 1 993.
[1 3 8] A.A. Ovchinnikov and Y.B. Zeldovich. Role of density fluctuation in
bimolecular reaction kinetics. Chern. Phys., 28 :21 5, 1 978.
[1 39] D. Toussain and F. Wilczek. Particle-antiparticle annihilation in diffusive
motion. J. Chern. Phys., 78 :2642, 1983.
[140] K. Kang and S. Redner. Scaling approach for the kinetics of
recombination processes. Phys. Rev. Lett., 52 :955, 1 984.
[1 4 1] H. Schnorer, I.M. Sokolov, and A. Blumen. Fluctuations statistics in the
diffusion-limited A + B - 0 reaction. Phys. Rev. A, 42 :7075, 1990.
[142] K. Lindenberg, B.J. West, and R. Kopelman. Diffusion-limited A + B - 0
reaction : Correlated initial condition. Phys. Rev. A, 42 :890, 1 990.
[143] S. Cornell, M. Droz, and B. Chopard. Some properties of the
diffusion-limited reaction nA + mB - C with homogeneous and
inhomogeneous initial conditions. Physica A, 188 :322-336, 1 992.
[1 44] R. Kopelman. Rate processes on fractals : theory, simulations and
experiments. J. Stat. Phys., 42 : 1 85, 1 986.
[ 145] B. Chopard, L. Frachebourg, and M. Droz. Multiparticle lattice gas
automata for reaction-diffusion systems. Int. J. Mod. Phys. C, 5 :47-63,
1 994.
[146] L. Galfi and Z. Racz. Properties of the reaction front in an A + B - C
type reaction-diffusion process. Phys. Rev. A, 38 : 3 1 5 1 , 1988.
[147] B. Chopard and M. Droz. Microscopic study of the properties of the
reaction front in an A + B - C reaction-diffusion process. Europhys.
Lett., 15 :459-464, 1 99 1 .
[148] Y.E. Koo, L. Li, and R. Kopelman. Reaction front dynamics in
diffusion-controlled particle antiparticle annihilation : experiments and
simulations. Mol. Cryst. Liquid Cryst., pp. 1 8 7-192, 1 990.
322 References
[149] Z. Jiang and C. Ebner. Simulation study of reaction fronts. Phys. Rev. A,
42 :7483, 1 990.
[1 50] S. Cornell, M. Droz, and B. Chopard. Role of fluctuations for
inhomogeneous reaction-diffusion phenomena. Phys. Rev. A, 44 :4826-32,
1 99 1 .
[1 5 1] S. Cornell and M. Droz. Steady-state reaction-diffusion front scaling for
rnA + nB --+ [inert] . Phys. Rev. Lett., 70 :3824-27, 1 993.
[1 52] H. K. Henisch. Crystals in Gels and Liesegang Rings. Cambridge
University Press, 1988.
[1 5 3] K. Jablczynski. La formation rythmique des pecipites : Les anneaux de
Liesegang. Bull. Soc. Chim. France, 33 : 1 592, 1 923.
[1 54] R.E. Liesegang. U ber einige Eigenschaften von Gallerten. Naturwiss.
Wochenschr., 1 1 :353, 1 896.
[1 5 5] S. Prager. Periodic precipitation. J. Chern. Phys., 25 :279, 1 956.
[1 56] Y.B. Zeldovitch, G.I. Barrenblatt, and R.L. Salganik. The
quasi-periodical formation of precipitates occuring when two substances
diffuse into each other. Sov. Phys. Dokl., 6 :869, 1962.
[1 57] D.A. Smith. On Ostwald's supersaturation theory of rhythmic
precipitation (Liesegang rings). J. Phys. Chern., 81 :3 102, 1 984.
[1 5 8] W. Ostwald. Lehrbuch der allgemeinen Chemie. Engelman, 1 897.
[1 59] G.T. Dee. Patterns Produced by Precipitation at a moving Reaction
Front. Phys. Rev. Lett., 57 :275-78, 1 986.
[1 60] B. Mathur and S. Ghosh. Liesegang Rings-Part I : Revert system of
Liesegang Rings. Kolloid-Zeitschrift, 159 : 143, 1958.
[1 6 1] E. Karpati-Smidroczki, A. B ki, and M. Zrinyi. Pattern forming
precipitation in gels due to coupling of chemical reaction with diffusion.
Colloid. Polym. Sci., 273 :857-865, 1 995.
[1 62] H. K. Henisch. Periodic Precipitation. Pergamon Press, 199 1 .
[1 63] B. Chopard, H.J. Herrmann, and T. Vicsek. Structure and growth
mechanism of mineral dendrites. Nature, 353 :409-41 2, October 199 1 .
[1 64] K.M. Pillai, V.K. Vaidyan, and M.A. Ittyachan. On the theory of
Liesegang phenomena. Colloid Polym. Sci., 258 :83 1-38, 1 980.
[1 65] W.H. Press, B.P. Flannery, S.A. Teukolsky, and W.T. Vetterling. Numerical
Recipes : The Art of Scientific Computing. Cambridge University Press,
1 989.
[1 66] T. Karapiperis and B. Blankleider. Cellular automata model of
reaction-transport process. Physica D, 78 :30-64, 1994.
[1 67] A. Turing. The chemical basis of morphogenesis. Phil. Trans. Roy. Soc.
London, B237 :37, 1952.
 References 323

[1 68] V. Dufiet and J. Boissonnade. Conventional and unconventional Turing

patterns. J. Chern. Phys., 96 :664, 1 99 1 .
[1 69] J. Schnackenberg. Simple chemical reaction systems with limit cycle
behaviour. J. Theor. Biol., 81 :389, 1 979.
[1 70] B. Chopard, M. Droz, S. Cornell, and L. Frachebourg. Cellular automata
approach to reaction-diffusion systems : theory and applications. In J.M.
Perdang and A. Lejeune, editors, Cellular Automata : Prospects in
Astrophysical Applications, pp. 1 57-1 86. World Scientific, 1 993.
[ 1 7 1] Shang-Keng Ma. Modern Theory of Critical Phenomena. W.A. Benjamen,
1 976.
[1 72] M. Doi. Stochastic theory of diffusion-controlled reaction. J. Phys. A,
9 : 1479, 1 976.
[173] P. Grassberger and M. Scheunert. Fock-space methods for identical
classical objects. Fortschrift. Phys., 28 :547, 1 980.
[ 1 74] L. Peliti. Path integral approach to birth-death processes on a lattice. J.
de Physique, 46 : 1469, 1985.
[ 1 75] C. Itzykson and J.-B. Zuber. Quantum Field Theory. McGraw-Hill, 1985.
[1 76] L. Sasvari and M. Droz. A renormalization-group approach to simple
reaction-diffusion phenomena. Phys. Rev. E, 48 :R2343, 1 993.
[1 77] B. P. Lee. Renormalisation group calculation for the reaction kA -+ 0. J.
Phys. A, 27 :2633, 1 994.
[178] M. Droz and A. McKane. Equivalence between Poisson representation
and Fock space formalism for birth-death processes. J. Phys. A,
27 :L467, 1 994.
[1 79] C.W. Gardiner. Handbook of Stochastic Methods. Sprinder-Verlag, 1983.
[1 80] C.W. Gardiner and S. Chaturvedi. The Poisson representation. A new
technique for chemical master equation. J. Stat. Phys., 17 :469, 1 978.
[1 8 1] D. Elderfield. Exact macroscopic dynamics in non-equilibrium chemical
systems. J. Phys. A, 18 :2049, 1985.
[1 82] L. Frachebourg. Sur le role des fluctuations en mecanique statistique hors
de l'equilibre. PhD thesis, Universite de Geneve, 1994.
[1 8 3] H.E. Stanley. Introduction to Phase Transitions and Critical Phenomena.
Oxford University Press, 197 1 .
[1 84] B. Schmittmann and R.K.P. Zia. Statistical Mechanics of Driven Diffuse
Systems. Volume 1 5 of Phase Transitions and Critical Phenemena.
Academic Press, 1996.
[1 8 5] M. Droz, Z. Racz, and J. Schmidt. One-dimensional kinetic Ising model
with competing dynamics : Steady-state correlations and relaxation
times. Phys. Rev. A, 39 :2141, 1989.
[1 8 6] K. G. Wilson and J. Kogut. Renormalization group and critical
phenomena. Phys. Rep. C, 12 :75, 1974.
324 References
[1 8 7] T. M. Liggett. Interacting Particle Systems. Springer, 1985.
[1 88] W. Feller. An Introduction to Probability Theory and its Applications. J.
Wiley, 1 962.
[189] R. Dickman and M. Bruschka. Nonequilibrium critical poisoning in a
single species model. Phys. Lett. A, 127 : 1 32, 1 987.
[1 90] R. Dickman. Nonequilibrium lattice models : series analysis of steady
states. J. Stat. Phys., 55 :997, 1 989.
[1 9 1] Ph. Beney, M. Droz, and L. Frachebourg. On the critical behaviour of
cellular automata models of nonequilibrium phase transtions. J. Phys. A,
23 :3353, 1 990.
[1 92] R.M. Ziff, E. Gulari, and Y. Barshad. Kinetic Phase Transitions in an
Irreversible Surface-Reaction Model. Phys. Rev. Lett., 56 :2553, 1 986.
[193] B. Chopard and M. Droz. Cellular automata approach to non
equilibrium phase transitions in a surface reaction model : static and
dynamic properties. J. Phys. A, 21 :205, 1 987.
[194] R.M. Ziff, K. Fichthorn, and E. Gulari. Cellular automaton version of
the AB2 reaction model obeying proper stoichiometry. J. Phys. A,
24 :3727, 1 99 1 .
[195] F. Bagnoli, B. Chopard, M. Droz, and L. Frachebourg. Critical behavior
of a diffusive model with one adsorbing state. J. Phys. A, 25 : 1085, 1 992.
[196] H.-P. Kaukonen and R.M. Nieminen. Computer simulations studies of a
catalytic oxidation of carbon monoxide on platinum metals. J. Chern.
Phys., 91 :4380, 1 989.
[1 97] M. Ehsasi, F.S. Rys, and W. Hirschwald. Steady and nonsteady rates of
reaction in a heterogeneously catalyzed reaction : oxidation of CO on
platinium, experiments and simulations. J. Chern. Phys., 91 :4949, 1 989.
[1 98] B. Sente. Contribution a l'etude des reactions de surface. PhD thesis,
Universite de Mons-Hainaut, Belgium, 1 992.
[199] J. Mai and W. von Niessen. Cellular automata approach to a surface
reaction. Phys. Rev. A, 44 :R6 1 65, 1 99 1 .
[200] P. Grassberger. Directed percolation in 2 + 1 dimensions. J. Phys. A,
22 :3673, 1 989.
[201] M. Droz and L. Frachebourg. Damage spreading in cellular automata
models of nonequilibrium phase transitions. Phys. Lett. A, 148 :447, 1 990.
[202] W. Kinzel. Directed percolation. In G. Deutscher and J. Adler, editors,
Percolation Structures and Processes, volume 5. Annals of the Israel
Physical Society, 1983.
[203] P. Grassberger, F. Krause, and T. von der Twer. A new type of kinetic
critical phenomenon. J. Phys. A, 17 :L105, 1 984.
[204] J.L. Cardy and U.C. Tauber. Theory of branching and annihilating
random walks. Phys. Rev. Lett., 77 :4780, 1 996.
 References 325

[205] I. Jensen. Critical behavior of nonequilibrium models with infinitely

many absorbing states. Int. J. Mod. Phys. B, 8 :3299, 1 994.
[206] M.A. Munos, G. Grinstein, R. Dickman, and R. Livi. Infinite numbers
of adsorbing states : critical behavior. Physica D, 103 :485-490, 1 997.
[207] F. Bagnoli, M. Droz, and L. Frachebourg. Ordering in a one-dimensional
driven diffuse system with parallel dynamics. Physica A, 179 :269-276,
1 99 1 .
[208] M. Marder and J. Fineberg. How things break. Physics Today,
pp. 24--2 9, September 1 996.
[209] B. Chopard and P.O. Luthi. A lattice Holtzman model and its
application. In S. Bandini and G. Mauri, editors, ACRJ'96 Proceedings
of the Second Conference on Cellular Automata for Research and Industry,
pp. 1 3-24. Springer-Verlag, 1997.
[210] P. Ossadnick. Cellular Automaton for the Fracture of Elastic Media.
In K. Schilling T. Lippert and P. Ueberholz, editors, Science on the
Connection Machine, pp. 127-136. World Scientific, 1 993.
[2 1 1] W. J. R. Hoeffer. The Transmission-Line Matrix method. Theory and
applications. IEEE Trans. on Microwave Theory and Techniques,
MTT-33( 10) :882-893, October 1985.
[2 1 2] H. J. Hrgovcic. Discrete representation of the n-dimensional wave
equation. J. Phys. A, 25 : 1 329-1350, 199 1 .
[2 1 3] C. Vanneste, P. Sebbah, and D. Sornette. A wave automaton for
time-dependent wave propagation in random media. Europhys. Lett.,
17 : 7 1 5, 1 992.
[214] D. Sornette, 0. Legrand, F. Mortessagne, P. Sebbah, and C. Vanneste.
The wave automaton for the time-dependent Schroedinger, classical wave
and Klein-Gordon equations. Phys. Lett. A, 178 :292-300, May 1993.
[2 1 5] P.O. Luthi, B. Chopard, and J.-F. Wagen. Wave Propagation in Urban
Microcells : a massively parallel approach using the TLM method. In
J. Dongarra, K. Madsen, and J. Wasniewski, editors, Applied Parallel
Computing : Computations in Physics, Chemistry and Engineering Science,
pp. 429-435. Springer, 1996. Lecture Notes in Computer Science ; Vol.
104 1 .
[2 1 6] D.H. Rothman and J.M. Keller. Immiscible Cellular Automaton Fluids.
J. Stat. Phys, 52 :275-282, 1988.
[2 1 7] R. Holme and D.H. Rothman. Lattice Gas and Lattice Boltzmann
Models of Miscible Fluids. J. Stat. Phys., 68 :409-429, 1 992.
[2 1 8] P. Calvin, P. Lallemand, Y. Pomeau, and G. Searby. Simulation of free
boundaries in flow systems by lattice-gas models. J. Fluid Mechanics,
188 :437, 1988.
[2 1 9] V. Zehnlt� and G. Searby. Lattice gas experiments on a non-exothermic
diffusion flame in a vortex field. J. de Physique, 50 : 1 083-1097, 1989.
326 References
[220] W.B. Hardy. The Spreading of Fluid on Glass. Phil. Mag., 38 :49-:S5,
1 9 1 9.
[221] P. Ball. Spreading it about. Nature, 338 :624, April 1 989.
[222] F. Heselot, N. Fraysse, and A.M. Cazabat. Molecular layering in the
spreading of wetting liquid drops. Nature, 338 :640-642, April 1 989.
[223] P. Thompson and M. Robbins. To slip or not to slip. Physics World,
pp. 35-38, November 1 990.
[224] P.G. de Gennes and A.M. Cazabat. Etalement d'une goutte stratifiee
incompressible. C.R. Acad. Sci.Paris, II, 310 : 1 601-1 606, 1 990.
[225] D.B. Abraham, P. Collet, J. De Connick, and F. Dunlop. Langevin
dynamics of spreading and wetting. Phys. Rev. Lett., 65 : 195-198, 1 990.
[226] D.B. Abraham, J. Heinio, and K. Kaski. Computer simulation studies of
fluid spreading. J. Phys. A, 24 :L309, 1 9 9 1 .
[227] B. Chopard. Numerical simulations of a Langevin dynamics of wetting.
J. Phys. A, 24 :L345-350, 1991.
[228] C. Appert and S. Zaleski. Lattice gas with a liquid-gas transition. Phys.
Rev. Lett., 64 : 1-4, 1 990.
[229] R.E. Rosensweig. Magnetic Fluids. Scientific American, pp. 1 24-1 32,
October 1 982.
[230] V. Sofonea. Lattice Boltzmann Approach to Collective-Particle
Interactions in Magnetic Fluids. Europhys. Lett., 25 :385-390, 1 994.
[23 1] K. Binder and D.W. Heermann. Monte Carlo Simulation in Statistical
Physics. Springer-Verlag, 1 992.
[232] B.M. Boghosian, J. Yepez, F.J. Alexander, and N.H. Margolus. Integer
lattice gases. Phys. Rev. E, 55 :41 37-4147, 1 997.
[233] Y.H. Qian, S. Succi, and S.A. Orszag. Recent Advances in Lattice
Boltzmann Computing. In D. Stauffer, editor, Annual Reviews of
Computational Physics III, pp. 195-242. World Scientific, 1 996.
[234] M.E. Colvin, A.J.C. Ladd, and B.J. Alder. Maximally Discretized
Molecular Dynamics. Phys. Rev. Lett., 61 :38 1-384, 1 988.
[235] E.H. Hauge. Time Correlations Functions from the Boltzmann Equation.
Phys. Rev. Lett., 72 : 1 501-1 503, 1 972.
[236] T. Uematsu, T. Nakata, K. Takeuchi, Y. Arisawa, and Y. Kaenada.
Three-dimensional numerical simulation of snow drift. Cold Regions Sci.
Technol., 20( 1 ) :25-39, 1 9 9 1 .
[237] B. Bang, A. Nielsen, P.-A. Sundsb0, and T. Wiik. Computer simulation
of wind speed, wind pressure and snow accumulation around buildings
(SNOW-SIM). Energy and Buildings, 21 :235-243, 1 994.
[238] A. Clappier. Influence des particules en saltation sur la couche limite
turbulente. Technical report, Ecole polytechnique Federale de Lausanne
(Switzerland), LASEN, 1 99 1 .
 Glossary

A model A simple model discribing poisoning transitions similar to

1-
those observed on a catalytic surface. A-particles can be adsorbed on
vacant sites with some rate p. An adsorbed particle having at least one
nearest neighbor empty is desorbed with a probability p. As a
function of the value of p the stationary state is poisoned or not.
advection Transport of some quantity p due to an underlying flow
with speed u, or due to an external drift. An advection term is usually
written as u · V p. Convection is often used as a synonym of advection.
annealed disorder One possible type of disorder in statistical systems.
The annealed degrees of freedom responsible for the disorder are in
thermal equilibrium with the other degrees of freedom of the system. In
some CA models, an annealed disorder can be produced by an
independant random process at each time step.
BGK models Lattice Boltzmann models where the collision term Q
is expressed as a deviation from a local equilibrium distribution f ( 0l ,
namely Q = (f (O) - f)/�, where f is the unknown particle distribution
and � a relaxation time (which is a parameter of the model). BGK
stands for Bhatnager, Gross and Krook who first considered such a
collision term, but not specifically in the context of lattice systems.
Boltzmann equation A balance equation which expresses how the
average number of particles with a given velocity changes between
(t + dt, r + dr) and (t, r), due to inter-particle interactions and ballistic
motion ; t and r are the time and space coordinates, respectively.
CA Abbreviation for cellular automata or cellular automaton.
cellular automaton System composed of adjacent cells or sites
(usually organized as a regular lattice) which evolves in discrete time

327
328 Glossary
steps. Each cell is characterized by an internal state whose value belongs
to a finite set. The updating of these states is made in parallel according
to a local rule involving only a neighborhood of each cell.
chaos Way to describe unpredictable and apparently random
structures.
Chapman-Enskog method Expansion around the local equilibrium
distribution function f (O) , often used to solve a Boltzmann equation. One
expresses the solution as f = f (O) + ej U) + e 2 j ( 2 ) + ... where e is the
expansion parameter. By identifying the same order in e, the Boltzmann
equation yields f ( i) in terms of j U - 1 ) . However, the relation is not
invertible unless some assumptions (that are the essence of the
Chapman-Enskog scheme) are made.
collision term The right-hand side of a Boltzmann equation or of the
microdynamics associated to an LGA model. Describes the balance of
particles entering and leaving a lattice site due to collisions.
conservation law A property of a physical system in which some
quantity (such as mass, momentum or energy) is locally conserved during
the time evolution. These conservation laws should be included in the
microdynamics of a CA model because they are essential ingredients

0
governing the macroscopic behavior of any physical system.
continuity equation An equation of the form O t P + divpu =
expressing the mass (or particle number) conservation law. The quantity
p is the local density of particles and u the local velocity field.
coupled map lattice A dynamical system defined on a spatial lattice.
Each lattice site is characterized by an internal state taking a continuous
range of possible values. The dynamics of each site is defined by a
function (map) giving the evolution in terms of the local state and those
of the nearest neighbors. For descrete time, it can be viewed as a CA
with continuous states.
critical phenomena The phenomena which occur in the vicinity of a
continuous phase transition, and are characterized by very long
correlation length.
diffusion A physical process described by the equation O t P = DV2 p,
where p is the density of a diffusing substance. Microscopically, diffusion
can be viewed as a random motion of particles.
directed percolation Percolation phenomenon in which the
connectivity between the lattice sites (bonds or sites) has a preferred
direction.
 Glossary 329
DLA Abbreviation of Diffusion Limited Aggregation. Model of a
physical growth process in which diffusing particles stick on an existing
cluster when they hit it. Initially, the cluster is reduced to a single seed
particle and grows as more and more particles arrive. A DLA cluster is a
fractal object whose dimension is typically 1.72 if the experiment is
conducted in a two-dimensional space.
driven diffusion Diffusion with a drift, possibly caused by an external
field, in some preferred direction.
dynamical system A sytem of equations (differential equations or
discretized equations) modeling the dynamical behavior of a physical
system.
equilibrium states States characterizing a closed system or a system
in thermal equilibrium with a heat bath.
equipartition Postulate according to which all the microstates of an
isolated system formed by a large number of particles are realized with
the same probability.
ergodicity Property of a system or process for which the
time-averages of the observables converge, in a probabilistic sense, to
their ensemble averages.
Euler equation Equation of fluid mechanics describing the evolution
of the velocity field u of a perfect fluid (without viscosity). It reads
1
a tu� + (�u · V )�u = - - VP
p
where p is the density and P the pressure.
exclusion principle A restriction which is imposed on LGA or CA
models to limit the number of particles per site and/ or lattice directions.
This ensures that the dynamics can be described with a cellular automata
rule with a given maximum number of bits. The consequence of this
exclusion principle is that the equilibrium distribution of the particle
numbers follows a Fermi-Dirac-like distribution in LGA dynamics.
FCHC model Abbreviation for Face-Centered-Hyper-Cubic model.
This is a four-dimensional lattice gas model (generalizing the FHP rule)
whose projection onto the 3D space provides a model of hydrodynamic
flows.
FHP model Abbreviation for the Frisch, Hasslacher and Pomeau
lattice gas model which was the first serious candidate to simulate
two-dimensional hydrodynamics on a hexagonal lattice.
330 Glossary
FHP-111 An extension of the FHP lattice gas model, involving rest
particles and more complicated collision rules in order to achieve lower
viscosity and higher Reynolds number flows.
field theory The theory that describes the dynamics of fields defined
on space-time and obeying partial differential equations.
Fock space A Hilbert space useful to study processes in which the
number of particles is not conserved. The creation and annihilation of
particles are described in terms of creation or annihilation operators
acting on the vectors of this Hilbert space.
fractal Mathematical object usually having a geometrical
representation and whose spatial dimension is not an integer. The
relation between the size of the object and its "mass" does not obey that
of usual geometrical objects. A DLA cluster is an example of a fractal.
front The region where some physical process occurs. Usually the
front includes the locations in space that are first affected by the
phenomena. For instance, in a reaction process between two spatially
separated reactants, the front describes the region where the reaction
takes place.
Galilean invariance Fundamental physical properties characterizing a
system that is invariant under a constant velocity shift (i.e. the system

1
obeys the same equation whether or not it moves with respect to the
observer). Galilean invariance is violated in several LGA and this is
reflected by a factor g =I= in front of the convective term uVu in the
hydrodynamical description.
geometrical phase transition Qualitative change of behavior (similar
to a thermodynamical phase transition) in a system without thermal
effects. Percolation and directed percolation are examples of geometrical
phases transitions.
Glauber dynamics Dynamical model aiming at describing the
relaxation of an Ising spin system, from an initial state to the equilibrium
state, with no conservation law.
Hamming distance A way to define a distance between two
microstates by counting the numbers of sites on which the two
microstates differ.
HPP model Abbreviation for the Hardy, de Pazzis and Pomeau
model. The first two-dimensional LGA aimed at modeling the behavior
of particles colliding on a square lattice with mass and momentum
 Glossary 331

conservation. The HPP model has several physical drawbacks that have
been overcome with the FHP model.
invariant A quantity which is conserved during the evolution of a
dynamical system. Some invariants are imposed by the physical laws
(mass, momentum, energy) and others result from the model used to
describe physical situations (spurious, staggered invariants). Collisional
invariants are constant vectors in the space where the Chapman-Enskog
expansion is performed, associated to each quantity conserved by the
collision term.
Ising model Hamiltonian model describing the ferromagnetic
paramagnetic transition. Each local classical spin variables s i = ± 1
interacts with its neighbors.
isotropy the property of continuous systems to be invariant under
any rotations of the spatial coordinate system. Physical quantities defined

2D
on a lattice and obtained by an averaging procedure may or may not be
isotropic, in the continuous limit. It depends on the type of lattice and
the nature of the quantity. Second-order tensors are isotropic on a
square lattice but fourth-order tensors need a hexagonal lattice.
LB Abbreviation for lattice Boltzmann.
lattice Boltzmann model A physical model defined on a lattice where
the variables associated to each site represent an average number of
particles or the probability of the presence of a particle with a given
velocity. Lattice Boltzmann models can be derived from cellular
automata dynamics by an averaging and factorization procedure, or be
defined per se, independently of a specific realization.
lattice gas A system defined on a lattice where particles are present
and follow a given dynamics. Lattice gas automata (LGA) are a
particular class of such system where the dynamics is performed in
parallel over all the sites and can be decomposed in two stages : (i)
propagation : the particles jump to a nearest-neighbor site, according to
their direction of motion and (ii) collision : the particles entering the
same site at the same iteration interact so as to produce a new particle
distribution. HPP and FHP are well-known LGA.
lattice spacing The separation between two adjacent sites of a
regular lattice. Throughout this book, it is denoted by the symbol A.
LBGK model Abbreviation for Lattice BGK models. See BGK for a
definition.
332 Glossary

LGA Abbreviation for Lattice Gas Automaton. See lattice gas

model for a definition.
local equilibrium Situation in which a large system can be
decomposed into subsystems, very small on a macroscopic scale but large
on a microscopic scale such that each sub-system can be assumed to be
in thermal equilibrium. The local equilibrium distribution is the function
which makes the collision term of a Boltzmann equation vanish.
lookup table A table in which all possible outcomes of a cellular
automata rule are pre-computed. The use of a lookup table yields a fast
implementation of a cellular automata dynamics since however
complicated a rule is, the evolution of any configuration of a site and its
neighbors is directly obtained through a memory access. The size of a
lookup table grows exponentially with the number of bits involved in the
rule.
Margolus neighborhood A neighborhood made of two-by-two blocks
of cells, typically in a two-dimensional square lattice. Each cell is updated
according to the values of the other cells in the same block. A different
rule may possibly be assigned dependent on whether the cell is at the
upper left, upper right, lower left or lower right location. After each
iteration, the lattice partition defining the Margolus blocs is shifted one
cell right and one cell down so that at every other step, information can
be exchanged across the lattice. Can be generalized to higher dimensions.
master equation An equation describing the evolution of the
probability of a state at a given time as the balance between transitions
leading to this state and transitions removing the system from this state.
mean-field approximation An approximation, often used in statistical
mechanics, which approximates an interacting many-body problem by a
sum of effective one-body problems.
microdynamics The Boolean equation governing the time evolution
of a LGA model or a cellular automata system.
Monte-Carlo method Method for computing integrals in a large
dimensional space. In statistical physics, the method is used to compute
ensemble averages. The ensemble average is replaced by a time average.
The dynamics defining the time evolution is given by a master equation
such that only the most important states are visited during the time
evolution.
Moore neighborhood A neighborhood composed of the central cell
and all eight nearest and next-nearest neighbors in a two-dimensional
square lattice. Can be generalized to higher dimensions.
 Glossary 333

multiparticle models A discrete dynamics modeling a physical system

in which an arbitrary number of particles is allowed at each site. This is
an extension of an LGA where no exclusion principle is imposed.

multiphase flows A flow composed of several different fluid species

or different phases of the same fluid. The different phases may be
miscible or immiscible.

multiscale expansion A method for solving a differential equation in

which phenomena at different scales are present. For instance, the time
evolution can be governed by two scales (e.g. diffusive and convective
processes) so that the time derivative is split as Ot = EOt1 + e2 ot2 where t 1
and t2 are the two time scales. The solution is then obtained by
identifying the same orders in e.

multispeed models Designate a lattice gas model in which at least

1
two velocity amplitudes (fast and slow particles) are allowed. For
instance, on a 2D square lattice, a two-speed model may include particles
of speed which jump to the nearest neighbors in one time step, while
speed- .J2 particles jump to the next nearest neighbors (along the
diagonal). Interactions conserving mass, momentum and energy are
expected between the different particle populations. Multispeed models
are useful in having a distinct mass and kinetic energy conservation laws.

multispin coding A way to code several lattice sites in the same

computer word, using each bit to represent the state of a site. An
N -states automaton requires N computer words to be represented. Only
one dimension of the lattice can be multispin coded in a given
implementation, for efficiency reasons. For 32-bit computers, this
technique is expected to produce a speedup of 32 since 32 sites can be
computed at once. This technique also saves on computer storage but is
more involved from a programming point of view and less flexible when
tuning a rule.

Navier-8tokes equation The equation describing the velocity field u

in a fluid flow. For an incompressible fluid (otp = 0), it reads

8tu + (u V)u = _ !vp + vV2 u

·

p
where p is the density and P the pressure. The Navier-Stokes equation
expresses the local momentum conservation in the fluid and, as opposed
to the Euler equation, includes the dissipative effects with a viscosity
term vV2 u. Together with the continuity equation, this is the fundamental
equation of fluid dynamics.
334 Glossary

neighborhood The set of all cells necessary to compute a cellular

automaton rule. A neighborhood is usually composed of several adjacent
cells organized in a simple geometrical structure. Moore, von Neumann
and Margolus neighborhoods are typical examples.

occupation numbers Boolean quantities indicating the presence or

absence of a particle in a given physical state.

open system A system communicating with the environment by

exchange of energy or matter.

parallel Refers to an action which is performed simultaneously at

several places. A parallel updating rule corresponds to the updating of
all cells at the same time as if resulting from the computations of several
independent processors.

partitioning A technique consisting in dividing space in adjacent

domains (through a partition) so that the evolution of each block is
uniquely determined by the states of the elements within the block.

percolation A "geometrical phase transition" of connectivity in a

system. Imagine an infinite lattice whose bonds (sites) are randomly
occupied with probability p. When p is larger that a percolation
threshold Pc , there is an infinite path of connected bonds (or contiguous
sites) from one side of the system to the other. If p < Pc , clusters of
connected bonds (sites) remain finite.

phase transition Change of state obtained when varying a control

parameter such as the one occurring in the boiling or freezing of a liquid,
or in the change between ferromagnetic and paramagnetic states of a
magnetic solid.

quenched disorder One possible type of disorder in statistical systems.

The quenched degrees of freedom responsible for the disorder are not in
thermal equilibrium with the other degrees of freedom of the system.
They are frozen or evolve at a much longer time scale that the other
degrees of freedom.

random walk A series of uncorrelated steps of length unity

describing a random path with zero average displacement but
characteristic size proportional to the square root of the number of steps.

rate equation The evolution equation for the local particles densities
in a reaction-diffusion systems in which the reaction terms are treated in
mean-field approximation.
 Glossary 335
reaction-diffusion systems Systems made of one or several species of
particles which diffuse and react among themselves to produce some new
species.
renormalization group Method which systematically implements
some form of coarse-graining to extract the properties of the large-scale
phenomena, in physical systems where many scales are important.
Reynolds number In fluid dynamics, a dimensionless quantity defined
as the ratio (uwd)jv, where u00 is the generic speed of the flow (e.g speed
far away from obstacles), d the typical size characterizing the system
(characteristic size of an obstacle) and v the kinematic viscosity of the
fluid. For the same geometry, flows with identical Reynolds numbers are
similar in the sense that they differ by scaling factors.
scaling hypothesis A hypothesis concerning the analytical properties
of the thermodynamic potentials and the correlation functions in a
problem invariant under a change of scale.
scaling law Relations among the critical exponents describing the
power law behaviors of physical quantities in systems invariant under a
change of scale.
self-organized criticality Concept aimed at describing a class of
dynamical systems which naturally drive themselves to a state where
interesting physics occurs at all scales.
spatially extended systems Physical systems involving many spatial
degrees of freedom and which, usually, have a rich dynamics and show
up complex behaviors. Coupled map lattices and cellular automata
provides a way to model spatially extended systems.
spin Internal degree of freedom associated to particles in order to
describe their magnetic state. A widely used case is the one of classical
Ising spins. To each particle, one associates an "arrow" which is allowed
to take only two different orientations, up or down.
telegraphist equation The partial differential equation
atP + (D j c2 )a� pDV2 p, where p is the quantity which is considered, D a
=

diffusion constant and c a propagation speed. The telegraphist equation

describes a combination of diffusive phenomena and wave propagation.
It prevents information from traveling infinitely fast.
thermal models In the context of lattice models, it designates a
model (such as a multispeed LGA) able to include thermal properties
such as, for instance, temperature and heat conductivity.
336 Glossary

thermohydrodynamics A branch of physics describing fluid flows and

including local energy exchanges between the different fluid regions. An
equation describing energy balance is added to the continuity and the
Navier-Stokes equations. Using thermodynamical laws, this new
equation can be written in terms of a local temperature function.
time step Interval of time separating two consecutive iterations in
the evolution of a discrete time process, like a CA or a LB model.
Throughout this book the time step is denoted by the symbol r.
Turing patterns Inhomogeneous stationary state observed in some
open reaction-diffusion systems. The diffusion is a destabilizing factor for
a homogeneous stationary state.
universality The phenomenon whereby many microscopically
different systems exhibit a critical behavior with quantitatively identical
properties such as the critical exponents.
updating operation consisting of assigning a new value to a set of
variables, for instance those describing the states of a cellular automata
system. The updating can be done in parallel and synchronously as is the
case in a CA dynamics or sequentially, one variable after another, as is
usually the case for a Monte-Carlo dynamics. Parallel, asynchronous
updating is less common but can be envisaged too. Sequential and
parallel updating schemes may yield different results since the
interdependencies between variables are treated differently.
viscosity A property of a fluid indicating how much momentum
"diffuses" through the fluid in a inhomogeneous flow pattern.
Equivalently, it describes the stress occurring between two fluid layers
moving with different velocities. A high viscosity means that the resulting
drag force is important and low viscosity means that this force is weak.
Kinematic viscosity is usually denoted by v and dynamic viscosity is
denoted by Yf = v p where p is the fluid density.
von Neumann neighborhood On a two-dimensional square lattice, the
neighborhood including a central cell and its nearest neighbors north,
south, east and west.
Ziff model A simple model describing adsorption-dissociation-
desorption on a catalytic surface. This model is based upon some of the
known steps of the reaction A B2 on a catalyst surface (for example
-

C0-0 2 ).
 Index

A + A ---> 1/J process, 2 1 5 periodic, 1 6

A + B ---> 1/J process, 1 8 8 reflecting, 1 6
Box-Muller method, 2 1 3
A-model, 234, 248 box-counting method, 1 69, 205
additive rules, 6 1 Burger equation, 1 37
adsorption, 1 7 1 By!, 3
o f dimers, 1 7 7
advection, 210 CA, 1
velocity, 2 1 2 capillary tube, 289
aggregation, 7 , 61, 1 6 7 car traffic, 5 1
Allen-Cahn theory, 37 carbonation, 178
annealed, 177 catalytic surface, 242
annihilation reaction, 188 cellular automata, 1
anomalous deterministic, 1 7
diffusion, 176 classification, 22
kinetics, 187 definition, 12
ant probabilistic, 17, 29
Langton, 46, 5 1 chaos, 2 1
rule, 63 Chapman-Enskog method, 7 1 , 8 2 , 1 29, 185
artificial life, 3, 51 checkerboard invariant, 75, 190
asynchronous updating, 18 chemotaxis, 176
autocorrelation, 1 22, 149, 303 Codd, 3
velocity, 105 collective behavior, 26
averaging procedure, 68 collision, 60
collision matrix, 105, 1 3 6
ballistic collision term, 83
aggregation, 6 1 collisional invariants, 88, 90
deposition, 29 combustion, 1 79, 281, 298
barometric law, 1 3 6 computational universality, 23
Belousov-Zhabotinsky reaction, 1 8 1 , 228 congestion, 55
BGK lattice method, 125, 267 conservation laws, 74
billiard-ball rule, 5 contact angle, 282
binary fluid, 1 2 1 , 293 contact process model, 234
block, Margolus, 14 contagion, 228
Boltzmann equation, 71, 87 continuity equation, 72, 76, 84, 85, 1 39
Boltzmann hypothesis, 86 control parameter, 232
bounce back condition, 4 1 , 286 Conway, John, 3
boundary conditions, 1 5 cooperative behavior, 48
adiabatic, 1 6 correlation, 86, 1 22
fixed, 1 6 coupled map lattice, 24, 293

337
338 Index
crack, 261 exclusion principle, 38, 67, 76, 140, 209
creeping, 305, 308 exponent
critical critical, 3 1 , 233
behavior, 249 mean-field, 197
exponent, 3 1 , 233, 252
phenomena, 2 1 9 FCHC model, 1 1 2
point, 250 ferrofluid, 293
current of particle, 139 FHP model, 6, 75, 77, 78, 1 1 8, 276
FHP-III, 77, 107, 109
D2Q9 model, 306 Fick's law, 139
damped waves field theory, 220
one dimension, 265 fluctuations, 86, 1 1 1, 1 22, 200
two dimensions, 264 fluid,multiparticle method, 298
decay process, 222 Fock space, 220
dendrite, 203 force
deposition, 305 external, 286
detailed balance, 106 gravity, 1 36, 296
diffusion long range, 292
constant, 1 39, 146, 1 50, 2 1 2, 2 1 8 short range, 292
front, 163 forest fire model, 31, 62
microdynamics, 140 fractal, 3 1 , 1 63, 201
one dimension, 70, 146 fracture, 261
rule, 1 39, 141, 1 76 Fredkin rule, 8
three dimensions, 149 front
two dimensions, 1 57 reaction, 193
with traps, 1 77 width, 194
diffusion-limited deposition, 1 7 1
dimer, 1 7 7 Galilean invariance, 97, 102, 103, 1 1 6, 1 24, 1 29,
directed percolation, 29 1 30
dispersion relation game of life, 3, 1 8
diffusion, 1 5 6 Gaussian approximation, 2 1 3 , 2 1 5 , 3 0 1 , 3 1 0
random walk, 6 9 geometrical phase transition, 30
wave model, 266 Glauber dynamics, 295
dissipation, 276 glider, 3, 19
DLA, 166, 1 77, 200 granular, 305
driven diffusion, 1 77, 255 gravity, 1 36, 296, 309
dynamical system, 2 1 Green-Kubo
formalism, 105
eddies, 1 10 relation, 149
Ehrenfest model, 175 Greenberg-Hasting model, 181, 228
Einstein, 28
Einstein relation, 14 7 H-theorem, 106
elastic collision, 60 Hamming distance, 255
energy, 96, 1 1 5 heterogeneous catalysis, 242
conservation, 276 hexagonal lattice, 75, 94, 295, 296
equipartition, 1 3 5 mapping on a square lattice, 287
internal, 1 1 5 homogeneous, 1 8 8
epidemic model, 228 HPP
equilibrium states, 232 collision rules, 39
equipartition, 96, 1 1 3, 1 3 5 model, 5, 6, 38, 1 3 5
ergodicity, 3 6 two-speed, 121
erosion, 305
Euler equation, 94, 96 immiscible flows, 7, 280
excitable media, xi, 179 inertia, 141
excluded surface, 1 72 inhomogeneous systems, 193
 Index 339

instability, 1 10, 298, 307 local equilibrium distribution, 268

numerical, 125, 307 logarithmic divergence, 105
integer lattice gases, 299 long time tails, 303
interacting particle systems, 234 lookup table, 17, 20
interaction
attractive, 292 Mach number, 96
repulsive, 292 macroscopic behavior, 26
interface, 1 63, 230 magnetic fluid, 293
motion, 37 magnetization, 35, 293, 295
invariant, 1 1 8 majority rules, 37
checkerboard, 1 1 8, 1 57, 190 Margolus neighborhood, 39, 176
dynamic, 121 master equation, 18, 220, 255
geometric, 1 2 1 mean-field approximation, 188, 235
spurious, 1 1 8 mean-free-path, 103
staggered, 1 1 7 meniscus, 289
Ising model, 27, 33, 293 microdynamics, 80
isotropy, 42 diffusion, 140
hexagonal lattice, 7 5 FHP, 78
square lattice, 126 FHP-III, 109
wave model, 273 HPP, 1 3 5
multiparticle, 299
J ablczynski coefficient, 202, 208 random walk, 67
jamming reaction-diffusion, 1 8 2
limit, 172 microscopic behavior, 2 6
transition, 54 miscible flows, 280
molecular chaos, 86
momentum conservation, 59, 267, 269, 281
Karman vortex street, 1 34
Monte-Carlo method, 232, 295
kinetic energy, 96, 109, 1 1 3, 1 14, 1 16, 1 1 7, 121,
Moore neighborhood, 14
135
multiparticle
diffusion model, 210
Langton, 3
fluid, 298
Laplace equation, 159
models, 8, 1 7, 1 25, 1 84, 209
lattice Boltzmann
reaction model, 214
equation, 7 1 , 106, 123, 1 5 1
multiphase flows, 280
fluid, 125
multiscale expansion, 74, 266
method, 7
multispeed models, 96, 1 12, 1 1 3, 1 1 5, 1 1 6
model, 1 7, 1 22, 206, 209
multispin coding, 20
reaction diffusion, 1 9 1
lattice gas, 6 Navier-Stokes equation, 6, 26, 76, 85, 102, 1 3 1 ,
lattice gas automata, 6 1 33, 302
lattice spacing, 78 neighborhood, 14
LBGK model, 125, 306 Margolus, 14, 39, 43
Lennard-Janes interaction, 284 Moore, 14
level of reality, 26 von Neumann, 14, 165
LGA, 6 noise reduction, 168
Liese gang nondecidability, 24
bands, 198 nonlinear diffusion, 176
patterns, 198 nontrivial collective behavior, 24
rings, 198, 208 nonuniform CA, 62
spiral, 208 nucleation, 7, 199, 201
spirals, 198
limit occupation numbers, 78
cycle, 23 open system, 232
point, 22
local equilibrium, 89, 128 particle current, 146
340 Index
partitioning, 14, 39 Selkov model, 1 8 1 , 230
percolation, 29, 1 63, 252-254 semi-detailed balance, 106, 1 3 6
periodic motion, 48 sequential updating, 1 8, 254
phase transition, 232 simulator, cellular automata, 1 8
first-order, 232 Smagorinsky constant, 307
nonequilibrium, 232 snow transport, 305
second order, 232 snowdrift, 305
plane wave, 265 snowflake, 20
Poiseuille flow, 104, 303 solid body motion, 56, 261
Poisson method, 224 solubility condition, 88
porous media, 7 sound speed, 97, 1 30, 263, 302, 306
precursor film, 283, 292 source-sink problem, 1 5 1
prey-predator automaton, 229
spacing law, 199, 202
prisoner's dilemma, 64
spatially extended systems, 24
spatia-temporal structure, 29
Q2R rule, 33
speed of sound, 97, 1 30, 263, 302, 306
quenched, 177
spin, 27, 33, 293
spreading, 7, 282
Raleigh-Taylor instability, 296
spurious invariant, 74
random rule, 19
stoichiometric coefficients, 214
random walk, 67, 139
stoichiometry, 246
range, of a rule, 62
strain tensor, 307
rate equation, 1 84
strange attractors, 23
reacting matching pairs, 246
streaklines, 1 10
reaction
string model, 56, 261
constant, 1 8 1
subgrid model, 307
multiparticle rule, 2 1 4
term, 1 8 3 sublattice, 34, 296
reaction-diffusion, 7 , 1 78 supersaturation, 199, 201
refraction index, 271, 274 surface
renormalization group, 2 1 9, 232 adsorption, 1 7 1
repose, angle, 45 growth, 29
reversible reaction, 214 tension, 282, 293
Reynolds number, 7, 107, 108, 1 1 1 suspension, 1 10, 305
road traffic, 5 1 synchronization task, 62
robots, 63
roughness, 1 63, 1 66
tails, long time, 105, 122
RSA, random sequential adsorption, 172
telegraphist equation, 1 53, 1 55, 1 58
temperature, 96, 1 1 3, 1 1 6, 1 2 1 , 1 3 5 , 293
saltation, 305, 308
tensor
sand
fourth order, 94, 127, 272
dune, 305
pile, 42 isotropic, 94, 128
rule, 43 momentum, 81, 96, 271, 272, 307
sand-box method, 205 strain, 307
scaling thermal
relation, 32, 233 conductivity, 1 1 6
theory, 195 FHP models, 1 1 6
Schli:igl model, 231 LGA, 1 12
Schnackenberg model, 218 thermohydrodynamics, 1 1 5, 1 24
second-order rule, 13 three-dimensional
sedimentation, 305 diffusion, 149
segregation, 189 fluids, 1 12
self-organized criticality, 32 wave model, 276
self-reproduction, 2 time law, 199
 Index 341

time reversal, 50, 274, 276 von Neumann, 1

time step, 78 von Neumann neighborhood, 14
TLM method, 266 voter model, 255
toppling, sand, 42
totalistic CA, 25, 62 wave, 256
traffic equation, 267
congestion, 55 one-dimensional, 256
flow, 55 ring, 180
model, 62 source, 265
trap, 177 spiral, 1 80
trapping wall, 286 three-dimensional, 276
tube-worms rule, 1 8 1 two-dimensional, 258
turbulence, 7 wetting, 7, 282
Turing patterns, 2 1 7 complete, 282
partial, 282
universality, 219 width law, 199, 207
class, 219, 233 Wolfram
upper critical dimension, 236 generalized rule, 62
rules, 4, 12, 2 1 , 29, 52, 62, 253
viscometer, 104 Wolfram S., 4, 22
viscosity, 104, 1 3 3
kinematic, 102 Young relation, 282
lattice, 85, 99, 1 32, 146
von Karman street, 1 10, 303 Ziti model, 242, 248