Editorial Board:
E.Guyon
Ecole Nonnale
Lui Lam
San Jose State University
Dominique Langevin
Laboratoire de Physique ENS
H.E. Stanley
Boston University
Advisory Board:
J. Charvolin
Institut LaueLangevin
W. Helfrich
Patrick A. Lee
Massachusetts Institute
of Technology
John D. Litster
Massachusetts Institute
of Technology
David R. Nelson
Harvard University
F. Hoffman  La Roche
Martin Schadt
&Co.
Jacques Duran
Springer
Jacques Duran
Directeur de Recherche
CNRS
University of Paris VI
France
Translated by
Axel Reisinger
Lockheed Martin
Lexington, MA
USA
Editorial Board:
E. Guyon
Ecole Normale
45 RueD'Ulm
F75005 Paris
France
Lui Lam
Department of Physics
San Jose State University
One Washington Square
San Jose, CA 95192
USA
Dominique Langevin
Laboratoire de Physique ENS
24 Rue Lhomond
F75231 Paris, Cedex 05
France
H.E. Stanley
Center for Polymer
Studies
Physics Department
Boston University
Boston. MA 02215
USA
987 6 5 4 3 2 I
ISBN 9781461267904
Foreword
vi
Foreword
a funnel, a few hoses, and a glass container. As it turns out, a thousand technical
details can upset an experiment and lead to nonsensical results. The physics of
granulars may not require costly hardware, but it demands enormous care. Nor is
it exempt of real dangers: Industrial reactors can generate violent explosions, even
when they are ostensibly empty.
I am thrilled to see French physics actively engaged in the field of grains and
powders, where almost everything is yet to be discovered. Within the last decade,
a number of highly creative teams have popped up in places like Paris, Lyon, and
Rennes. Jacques Duran leads one of those teams at the Pierre and Marie Curie
University in Paris. This book draws on his firstrate expertise. I personally had
the opportunity to read various drafts as it was taking shape. I learned a great deal
from them in the process. There was a time when we had to steer anyone interested
in sand to the classic text by Bagnold. From now on, Duran will be high on the
required reading list. This book will fill a very real need. I sincerely wish it great
success.
PierreGilles de Gennes
Paris, France
Preface
viii
Preface
up concepts and theories about a medium whose behavior is most often surprising
and counterintuitive. My goal has been to gather, in as selfcontained a manner
as possible and in a unified language, the background necessary to understand the
latest developments, as well as to present the rudiments of granular physics.
Given the current status of this young discipline, any progress is likely to come
from experiments. Henri Poincare once said: "It has been quite some time since
anyone has thought of getting ahead of experimentation or to construct the world
on the basis of a few hasty hypotheses." The remark is particularly relevant to the
physics of granular media. This book is rooted in that very principle. In this spirit, it
contains many descriptions of experimental devices and accounts of experimental
results obtained under conditions that are as controlled as possible. These provide
the backdrop for concepts, models, and ideas that are pushed to levels that some
might find too speculative. The adopted strategy, whose limitations must be borne
in mind, goes with the very nature of this discipline, which is still very much in its
infancy. In this context, this book discusses concepts and results as they are known
to us at the present moment. As such, it is only a snapshot, as our understanding
of this topic is bound to evolve, perhaps even undergo major revisions.
Lest it become a dry compilation of facts, writing a book of this kind entails
somewhat arbitrary choices that remain the sole responsibility of the author. On
more than one occasion, I felt compelled to leave out this particular result or that,
even though I may come to regret it in the future. I did so simply because it did
not readily fit in with the flow of the presentation or because it would have taken
me too far afield. I beg for those whose work is mentioned too cursorily or not at
all to forgive me. This by no means implies a negative judgment on my part. My
choice was guided exclusively by pedagogical concerns and my own concept of
the logical flow of the material. My decisions were often agonizingly difficult.
My thanks go to my colleagues and friends in my own laboratory (Eric Clement,
Jean Rajchenbach, and Touria Mazozi). I am also grateful to the members of the
Physics of Heterogeneous and Complex Matter Research Group affiliated with
the National Center of Scientific Research (CNRS) and with the European HCM
(Human Capital and Mobility) Network. I am enormously indebted to them not
only for much of the technical material in the book but also for infusing in it
their irrepressible enthusiasm which I hope will brighten the experience of the
reader. I acknowledge the careful reading of the french manuscript and comments
by Ko Okumura and VJ.P. Faroux. Axel Reiseinger successfully carried out the
translation from the French edition. He kept within the spirit of the text while
suggesting numerous improvements and clarifications.
Finally, I owe a large debt of gratitude to PierreGilles de Gennes and Etienne
Guyon, both of whom played an instrumental role in the early stages of the research
discussed here. I was fortunate to have their unrelenting support and encouragement
to write and publish the material in this book. It borrows a great deal from their
own work.
Jacques Duran
Paris, France
Contents
Foreword
Preface
1.
2.
vii
Introduction
1.1 Some Orders of Magnitude Defining the Problem
1.2 Economic Implications and Industrial Problems
1.2.1 Industrial Processing of Granulars .
Construction Materials . . . . . . . . . .
Processing Industries . . . . . . . . . . .
An Example: Casting by Sacrificial Polystyrene
The Agriculture Industry
1.2.2 Flow Problems
.
1.2.3 Problems of Segregation . .
1.3 Granular Materials and Geophysics
1.4 A Brief Historical Review . . . . .
1.5 Prerequisites and Selected Bibliography .
10
10
12
14
16
18
19
19
20
22
23
1
3
4
4
6
7
Contents
2.2
2.3
2.4
3.
24
25
27
27
28
28
30
31
32
32
32
34
34
36
36
39
41
42
44
46
49
49
50
53
3.1
54
54
54
56
59
60
62
64
64
64
65
66
69
70
70
74
Contents
xi
74
75
75
76
76
77
79
80
82
84
86
87
87
89
89
90
91
91
91
93
96
96
98
101
104
104
105
105
107
108
110
111
111
114
116
117
117
xii
4.
Contents
4.2
5.
5.2
Introduction....................
5.1.1 Oyama's Cylindrical Drum
.
5.1.2 Potential Energy of a Heterogeneous Pile
Superposition of StacksTwo Compact Stacks
Where Are the Defects Concentrated? .
Segregation by Vibration
.
5.2.1 Simulation of Segregation by Size.
TwoDimensional Model
.
ThreeDimensional Model . . . . .
5.2.2 Experiments on Segregation by Vibration.
Experiments on Continuous and Intermittent Ascent .
Convection or Arch Effect?
.
119
119
123
123
126
126
127
128
128
130
131
132
133
134
136
138
145
145
146
146
147
149
150
151
154
154
155
156
157
158
160
160
162
164
164
166
167
Contents
5.3
5.4
6.
X111
168
168
171
171
176
178
178
179
181
181
182
Numerical Simulations
6.1 Introduction . . . . . . . . . . . . . . . . . . . .
6.1.1 The Challenges of Numerical Simulation
6.1.2 The Different Simulation Methods. . . .
Hard Spheres and Soft Spheres . . . . .
Duration of Collisions and Chronology Problems
6.1.3 The Transition from a Discrete
to a Continuous Description.
6.2 Simulations of Collisions . . . . . . . .
6.2.1 Introduction...........
6.2.2 OneDimensional LRV Procedure
6.3 Molecular Dynamics (MD) Simulations.
6.3.1 Elastic and Friction Forces . . .
Linear and Nonlinear Equations
Mechanical Analogies . . . . .
6.3.2 MD Collision Model . . . . . .
Linear Model of a Binary Collision
Nonlinear Model of a Binary Collision
The Detachment Effect . . . . . .
6.4 Simulation of the Dynamics of Contacts. . . .
6.5 Monte Carlo (MC) Simulations
Monte Carlo Technique for Stacking and Relaxation
6.6 Sequential Model of a Pile . . . . . . . . . . . . . .
184
184
185
185
186
186
187
189
189
189
190
191
191
193
194
194
196
197
199
202
203
206
Bibliography
209
Index
213
1
Introduction
1. Introduction
The physical laws governing the behavior of granular media actually apply to
objects whose dimensions cover several orders of magnitude. From grains, a few
hundred microns each, to ice floes drifting across the polar seas (over distances of
1000 km),2 not forgetting Saturn's rings (made of icy particles about 1 cm wide
distributed in a band roughly 1 km thick), the science of granulars covers at least
twelve decades of sizes.
The fact that aggregates seem to obey universal laws applicable over such a wide
range of dimensions and characteristics is a strong incentive to pursue fundamental studies in that area. For instance, phenomena of segregation and intermittent
blockages are pervasive in numerous industrial processes involving granular materials. Accordingly, the remainder of this chapter will be devoted to a brief and
selective review of a few techniques and processes that in one way or another are
subject to phenomena such as convection, segregation, blockages by arching, all
of which are routinely encountered in industry. These effects will be examined in
greater depth in subsequent chapters.
of the movements of icebergs, notably in the vicinity of port facilities, has been the object of
several research contracts with the Canadian Navy. It turns out to be closely related to the physics of
granular materials.
3 A recent report by
1. Introduction
and pharmaceutical industries, specialized chemistry, and the food industry, which
demand increasingly sophisticated processing technologies.
1.2.1.
Construction Materials
The construction industry (housing, hydraulic concrete needs, public works projects, and so on) consumes aggregates at the rate of seven tons per capita per year,
which generated revenues of some fifteen billion francs (three billion dollars) in
1994. 4 Next to water, aggregates are the most heavily used material. Natural
aggregates are of alluvial, volcanic, or sedimentary origin. After being extracted
from quarries or underwater deposits, they go through a series of steps that require
the handling of huge quantities of material in granular form. Among these operations are shortdistance conveying (often by forced flow through pipes), scalping
(i.e., the separation of nonvaluable fine residues, which removes at least part of the
problem for subsequent processing), crushing and grinding, followed by sifting,
washing, hydrocycloning, and desliming (particularly to improve the properties of
sand and gravel). The sequence generally ends with storage and ultimately transportation to the point of use. Most of these operations suffer from a troublesome
list of disruptive phenomena such as flow obstructions, formation of plugs, and
segregation, all of which we will have an opportunity to analyze in detail later on.
Industrial requirements are so diverse, and the properties of the corresponding
materials so complex, that it takes no fewer than 15 parameters to fully specify
an aggregate used, for instance, in the preparation of hydraulic concretes. They
describe a menu of characteristics like thickness, cleanliness, degree of crushing,
cohesion, coefficient of flatness, and a host of others. s This proliferation of empirical coefficients attests to the inherent complexity of the industrial world, while
physicists try to reduce them to no more than two or three fundamental micromechanical parameters, perhaps because that is the most they can handle. It would
seem that physics has a long way to go to catch up with the real world. Yetand
this is a very encouraging signthe troublesome phenomena that plague industrial
processes out in the field (blockages, arches, segregation) show up readily even in
our idealized theoretical models. 6 In this regard, it is instructive to at least briefly
discuss some problems associated with the concrete industry.
4The numbers quoted here come from data published by LafargeCoppee, a worldwide leader in the
manufacture of all types of construction materials.
5Interestingly, the official list of parameters characterizing industrial aggregates does not include even
one of the micromechanical parameters considered of fundamental importance by physicists. Industrial engineers do not care about things like the coefficient of elastic restitution of the coefficient
of friction. These omissions are ample evidence that practical problems in the real world often demand a more complex, albeit more empirical, set of parameters than seem necessary in the idealized
atmosphere of a laboratory.
6This touches on a fundamental aspect of the scientific method, which can at times seem too reductive.
Physicists have to choose objects of study that are simplified enough to be amenable to analysis, yet not
so simplified as to lose any relevance to the real process they are intended to simulate, which is invariably far more complex. We will confront several such choices during the course of this book. Making
the COlTect decision is a difficult task which, for the most part, rests on no more than sound intuition.
FIGURE 1. Around 200 B.C., Appolonius of Perga invented the arrangement shown here
to tile a surface as fully as possible with circles of various sizes. This object enjoys the
property of selfsimilarity.
Paradoxically, how aggregates contribute to the strength of concrete was rediscovered only very recently. It has only been a few years since the resilience of
bituminous or highperformance concretes was proven to be due in large part to the
characteristics of the aggregates used in their manufacture. Everyone knows that
concrete is essentially a mixture of aggregates and cement, and that it undergoes a
slow chemical transformation (on a scale of tens of years). Since it is a composite
material, the key to success is to minimize the influence of the more fragile binder
material, and maximize that of the much stronger granular material in its midst.
The practical solution is embodied in hypercompact arrangements, which constitute a natural answer to the problem posed many centuries ago by Appolonius of
Perga (ca. 262, ca. 190 B.C.) [7]. It is illustrated in Figure 1.7
This idea is probably what inspired engineers to add granulars of various sizes,
spanning several orders of magnitude, to the mixtures they prepare to make highstrength concretes. The finest ingredients in these mixtures are of submicron size;
they are called silica fume. Their cost is extremely high, but they have made it
possible to produce concretes which such high strength that the construction of
skyscrapers a full kilometer high has become technologically feasible. 8 Indeed, the
city ofTokyo is planning to erect just such structures. Nevertheless, itis not difficult
to anticipate the enormous problems encountered when trying to mix such a granular medium compactly, given the unavoidable phenomenon of size segregation.
It is only fitting to acknowledge that, although they were not familiar with the
dauntingly complex physics involved, engineers and specialists in the fields of
7Very much to the point, there exists an Arab saying that can be loosely translated as: "You think your
basket is full when you have stuffed it with oranges. In truth, it is full of void, for you can still add
nuts, and even chickpeas after thal."
8The quality of a highstrength concrete is often expressed as the height of a column with a onesquaremeter base that withstands collapse. This maximum height has increased by more than one
order of magnitude in the last few years.
1. Introduction
concretes and aggregates have proven remarkably adept at building on their rich
experience and mastering the difficult art of devising fluid binders in which each
grain can find its ideal spot. No one can deny that systematic experimentation has
yielded the right answers, even though manufacturing costs are so high that applications are for the time being restricted to the construction of highly specialized
buildings. It is reasonable to expect that the possible contributions of physics to a
better understanding of the properties of construction materials may substantially
lower the costs of producing, conveying, and mixing them. These processes are
intimately related to problems of flow and segregation, which are to be taken up
in Sections 1.2.2 and 1.2.3.
We will refrain from treating here the principles of soil mechanics, which fall
largely in the domain of geophysics and geology, or the kinds of problems civil
engineers have long been accustomed to dealing with. There exist in this area
numerous models painstakingly put together, and with a respectable track record.
For instance, most soilmechanics engineers tend to view granular materials as
continuous objects that obey the usual laws of classical mechanics (including
friction, stressstrain relations, plastic deformation, and the like), such as they
apply to homogeneous solids. As we will see, there are cases when this approach
raises serious questions. From a practical point of view, however, it does offer
the advantage of permitting simple calculations to predict the behavior of various
structures made of granular materials, such as roadways, with generally adequate
safety margins.
There is, however, one related phenomenon that deserves to be singled out,
because it shows up routinely in everyday life. We are referring to segregation
by shearing, an effect that is characteristic of most granular materials. Figure 2
describes the phenomenon in question, which civil engineers are quite familiar with. They have to contend with it as they plan and manage construction
projects.
Processing Industries
Generally speaking, modern manufacturing methods of hightechnology materials
rely heavily on the processing of granular materials. We are not about to review
each of the many applications, transformations, and various processes involving
granulars at some stage of the technology. One book would not be enough to
cover them all. We will simply give a few examples selected for the purpose of
highlighting the manyfaceted problems faced by industry, and their connections
with the fundamental approach advocated in this book.
We start by making a clear distinction between two separate branches of the
chemical industry. On the one hand are industries that typically deal with large
quantities of materials in granular or powder form. Their problems are generally
similar to the ones we have already mentioned, such as transport through pipes,
storage, or granular segregation. On the other hand are the chemical or pharmaceutical industries where materials are prepared in relatively small quantities, but
with highvalueadded materials. In this case, the demands in terms of purity and
reproducibility are obviously far more stringent. Accordingly, their manufacturing
FIGURE 2. Grading the embankment on the side of a road with a bulldozer causes segregation by shearing. The concentration of large rocks is much greater on the suti'ace of the
embankment than in the interior.
The technique discussed here has been used recently by the French automobile
maker Peugeot to produce engine blocks which would have been extremely difficult
to manufacture any other way. A number of electromechanical firms and steel
foundries have studied and developed similar processes. In principle, the method,
illustrated in Figure 3, is quite simple and of potentially low cost. Unfortunately,
its practical implementation runs into difficult problems because of a fundamental
property of granular materials, which we will examine in detail in Chapter 3, known
as convection by vibration. The sequence of steps required in this technique is as
follows:
A polystyrene model of the object to be duplicated in metal form is prepared.
The model is placed inside a container that is filled with fine sand.
The whole system is shaken vigorously by means of a vibrating support placed
under the container.
At the conclusion of this step, the sand occupies every bit of space between
the model and the container's wall. Highly compacted, it is now ready to act
as the mold.
1. Introduction
Polystyrene model
Vibration
10
1. Introduction
As is well known, many Western developed nations are major producers of agricultural and food products that often rank first on their list of exports. Despite
years of steady growth, the industry has been unable to bring much improvement
to its basic processing technology, particularly of granular materials such as wheat,
com, powdered milk, cocoa, and the like. A number of research centers have been
created around the world in the hope of changing this state of affairs. Engineers
have become increasingly aware of some of the outstanding problems concerning
the processing, transport, and storage of all manner of granulars.
As is the case in the chemical industry, the processing of foodstuffs falls in
one of two categories, depending on the degree of valueadded materials. The
prototype in the lowvalueadded category is the animal feed industry, which uses
a relatively primitive technology to handle enormous quantities of granular materials derived from agriculture. A typical plant treats hundreds of thousands of
tons of various grains per year. A small portion of its budgetusually around
9%is earmarked for processing, while 83% covers the raw materials (mostly of
agricultural origin), and 8% is devoted to transport. It is not uncommon to notice
in such facilities hoppers pockmarked with numerous dents left by hammer blows
intended to free flow blockages caused by arch effects. Materials such as flour are
prone to creating solid plugs. It is a fact of life that these blockages have to be
routinely cleared with pick axes, if not jackhammers.
A quick survey of the problems affecting daily operations in such plants reveals
that they are almost exclusively of two typesblockages of flow by the arch effect,
and granular separation.
1.2.2.
Flow Problems
Granular flows have properties that give rise to numerous problems in industrial
settings. Much can be learned by observing the flow of grains in an ordinary
hourglass filled with sand [8]. Contrary to what happens in sealedoff hourglasses
of the type once used to keep track of time, the flow of sand in coneshaped funnels
is generally not smooth. We know from experience that the coupling between the
granular flow and the ambient fluid can cause intermittent events that impart to
the exiting flux of material a discontinuous character. How granular materials
flow down industrialsize hoppers turns out to be a rather tricky problem that has
been the subject of a great many studies. Yet the physics of this seemingly simple
phenomenon remains poorly understood.
Virtually all industrial or laboratory facilities designed to handle granular materials such as grains (food industry) or gravels (construction industry and public
works) are plagued by blocked flows which can be extremely disruptive. Blockages of this type occur with annoying regularity, for instance, while trucks or barges
are being loaded with gravel through giant chutes designed to handle tens of tons
of material, as schematically illustrated in Figure 4. The frequency of these events
depends on the diameter of the opening, as well as on the size and mechanical properties of the aggregates. They can invariably be traced to the formation of arches
(b)
(e)
11
f\
FIGURE 4. Diagrams (a) and (b) show flow blockages due to the formation of arches near
the orifice of an hourglass or a hopper. Diagram (c) depicts an arch structure with maximum
stability (inverted chain).
in the vicinity of the discharge orifice. Depending on the circumstances, the flow
can either become sporadic or be completely blocked by remarkably stable arches.
For the same reason, it is extremely difficult to reproducibly mix two granular
materials by merging two separate flows into a common discharge channel. That
has proven to be a major problem in the food industry and in chemical plants
producing polymers, where raw materials are typically dry granulars. Blockages
can form spontaneously as a result of some kind of seizing mechanism. 9 The
phenomenon is even known to have caused dangerous explosions by allowing
fermentation gases to accumulate in grain silos plugged up by such arches.
Given the many uncertainties surrounding the static properties of a granular
system, it would seem that describing the mechanism of arch formation would be
quite a daunting problem. Nevertheless, it is possible to state a number of general principles which will later prove useful. It is important to realize that arches,
which are a direct consequence of gravitational forces, can form because of purely
geometrical and steric reasons. This can be easily understood by considering a
stack of identical frictionless spheres resting on either side against inclined walls,
as depicted in Figure 4(b). In this example, each sphere is kept in equilibrium at
two contact points whose connecting line lies below the center of gravity. While
such arrangements can obviously form spontaneously in twodimensional systems,
they would seem rather more improbable in reallife threedimensional situations,
because they require the cooperation of a much larger number of spheres, each stabilizing its neighbors. Nevertheless, they are frequently encountered in industrial
or laboratory conditions. Figure 5 illustrates some examples actually observed in
simple experiments.
Not surprisingly, the phenomenon of dry friction greatly increases the likelihood
of forming arches anchored onto the side walls. In this case, it is not even necessary
9The phenomenon of seizing is frequently observed in the industry, although it is not clearly understood.
It manifests itself in a gradual hardening of arches forming in hoppers. It is not uncommon for workers
in the agriculture industry to have to use pick axes to remove solid plugs in silos that have remained
idle for some time. These plugs can be the result of high humidity (see Section 2.1) or of poorly
known physical and chemical processes causing grains to stick together during storage.
12
1. Introduction
FIGURE 5. Stable arches form readily when a 1 em diameter test tube filled with fine sand
(100 !Lm) is turned upside down. The photo on the left provides a side view, while the one
on the right is taken from below to reveal the arch pinned against the wall.
for the side walls to be slanted. Indeed, we will run into several examples of arches
pinned against vertical walls (e.g., Section 3.2.4), particularly in silos. Effects of
this type can have dangerous consequences in industrial facilities. Engineers have
come up with a number of more or less effective solutions, such as those illustrated
in Figure 6.
1.2.3.
Problems of Segregation
The industry processes enormous quantities of granular materials year in and year
out. Virtually every stage of its operations is at the mercy of segregationa most
irksome phenomenon that tends to separate the components of a mixture supposed
(a)
(b)
(e)
FIGURE 6. Three methods used in industry to prevent or remedy blockages by arch effect.
They include (a) an Archimedes screw, and (b) a conveyor belt with a corrugated surface.
In (c), a plant worker pounds an obstructed hopper with a sledge hammer to get the flow
started again. The latter is the method of choice in industries producing lowvalueadded
granular materials.
13
Mixture
FIGURE 7. Two techniques used in industry to try to avoid problems of granular segregation. The principle is the same in both cases. The idea is to force the materials through a
section of hardware that promotes mixing.
14
1. Introduction
(a)
1
B
(b)
<~~
B
(e)
pieces are placed on top of each other (Figure S(c, resulting in an AlB I AI B
structure. The process is repeated until the thickness of each individual layer is
roughly equal to that of the largest of the two particles A or B. At the conclusion
of N operations, the number of layers is obviously 2N . Not only is the process
extremely laborintensive, but the final product is only a superposition of different
layers, rather than a true threedimensional mixture.
The process just described has been adapted to massproduction with socalled
Kenics mixers, illustrated in Figure 9, which accomplish, automatically and continuously, the discrete sequence of steps of the pastrymaker technique. 10
The pastes or fluids to be mixed are injected at points A and B shown in the figure.
The ultimate speed of the process is limited by the onset of turbulence phenomena
which may appear within the reactor and can potentially upset the quality and
smoothness of the mixing. Such phenomena depend primarily on the dimensions
of the apparatus and the properties of the products to be mixed.
It is clear that implementing this technology on the factory floor requires a
large capital investment and has to contend with difficult processing techniques
(the preparation of pastes can itself be a challenge). As such, it is largely incompatible with the processing of industrial quantities of lowvalueadded materials,
whose cost has to be kept as low as possible. A better understanding of the mechanism of segregation in granular materials is of great importance not only from a
fundamental physics point of view but for economic reasons as well.
11 The
1.3
15
B
FIGURE 9. The first two elements of a Kenics mixer (after [115]).
16
1. Introduction
Another discipline that, in many respects, has much in common with the physics
of dry granular materials is seismology. How forces of friction are stored and suddenly released is just as much of interest to geophysicists as it is to fundamental
theorists trying to understand the basic properties of granulars. The collective
behavior of particles in quasipermanent contact is the central focus of the geophysical sciences. Granted, geophysicists typically deal with interactions that are
far more complex than those to which we will deliberately restrict our attention.
They have to worry about phenomena such as cohesion effects, contact fatigue,
erosion, strain hardening, and many others that we shall barely touch upon. Yet,
the phenomenon of fracturing, which is at the basis of geophysics, shows up just
as faithfully, albeit in an elementary form, in the most grossly simplified versions
of the physics of granulars, as we will see in Chapter 3.
Avalanches and free flows are treated in Chapter 4, in which we draw liberally
from the work of geophysicists. We do so for at least two reasons. First, an
accurate description of what is called the stickslip phenomenon requires a clear
understanding of the interdependence between friction and velocity, which also
happens to be at the heart of the physics of granular materials. Second, the simulation of socalled selforganized critical (SOC) systems has been called upon to
explain both granular avalanches and the scaling laws of the distribution of earthquakes (typified by the GutenbergRichter law). The validity of this approach
remains a topic of lively debate in both cases.
17
experimental observations on the equilibrium of earthen embankments, the stability of stone structures, and other edifices. It puts the physics of granular materials
on a foundation that is difficult to contest even today. For instance, it ultimately
led to the celebrated Coulomb laws of dry friction between solids, which in time
would be extended to granular materials. In that sense, Coulomb's seminal paper
can be viewed as heralding an entirely new discipline.
In 1780, Ernst Chladni (17561827) noticed a number of interesting differences
(which remain rather puzzling even today) between the behavior of light granulars
(such as horse hair) and that of coarser and heavier ones (such as sand). He observed
what came to be known as Chladni's complementary figures, which we will discuss
in Chapter 2. His experiments were duplicated and confirmed a short while later by
Christiaan Oerstedt (17771851), who used lycopodium powderan extremely
fine material that was instrumental in many other discoveries. 14 Felix Savart
(17911841) was interested, among other things, in music. He would make use
in 1827 of Chladni's geometric figures to study the frequencies and wavelengths
of sound waves.
Michael Faraday (17911867) was another illustrious figure who, against the
backdrop of his primary work on hydrodynamic instabilities, also had a strong interest in how vibrations induce the formation of sand piles [12]. This phenomenon,
closely related to Chladni's experiments, remained a mystery until recently. The
problem was to figure out the role of air as sand collects in regular patterns. Chapter 3 will provide an opportunity to discuss the problem of granular instabilities,
reminiscent of those observed in liquids by Faraday, as well as the formation of
sand piles.
William Rankine (18201872) examined the theoretical implications of friction
in granular materials as early as 1857 [13]. Starting from Coulomb's ideas, he
established a number of principles which remain fully valid to this day. He defined
what is now called Rankine's passive and active states. These concepts are clearly
and thoroughly discussed in the book by Brown and Richards, and we will not
dwell on them here [5].
The problem of the equilibrium distribution of forces in a granular medium
stored in a silo has been studied by several researchers, who have published the
results of their work. 15 In 1884, I. Roberts noted that "in a structure whose lateral
walls are parallel, the pressure exerted by the stored grain onto the base ceases
to increase when the structure is filled to a height more than twice the diameter
of the inscribed circle" [14]. A few years later, H. Janssen, a German engineer
from the city of Bremen, proposed a model based on a coefficient describing the
redirection offorces toward the wall [15]. He failed to refer to the work of Roberts,
presumably because he was unaware of it. We will present the model in Chapter 3,
140ur great ancestors in science deserve much admiration for their inquisitiveness. Their curiosity was
rarely restricted to a single discipline.
15This problem remains quite current. Given the various uncertainties concerning the forces at play in
a granular mediumwhich we will study in some detailthe correct way to model the distribution
of forces on the base of a silo remains in dispute.
18
1. Introduction
2
Interactions in Granular Media
20
that such idealized objects might be fairly accurately approximated in the laboratory by placing solid particles in a vacuum chamber. In practice though, things are
rarely that simple because of a major intrinsic problem. As it turns out, the energy
dissipated through mutual collisions or friction is inevitably accompanied by the
appearance of surface electrical charges on both the particles themselves and the
walls of the container. Such surface charges are particularly difficult to eliminate
in vacuum; the resulting mediumrange electrostatic interactions can greatly complicate the analysis of experimental results. Adding a surrounding fluid makes it
much easier to drain away those pesky surface charges. For that reason, such a
fluid is not only desirable, but indeed essential. Fortunately, provided a few precautions are observed, a fluid matrix does not necessarily preclude treating many
laboratory, or even industrial, situations involving granulars as though the materials were truly dry. The purpose of this section is to define the range of validity
of the methodologies which we will develop throughout this book. For strictly
pedagogical reasons, we start out by discussing the interactions of a single particle
with its fluid environment.
Laminar Drag
Laminar
Turbulent
2.1
21
where the integral in the denominator extends over the interstitial space I around
the particle. It is not essential to work out this integral exactly. In order to get an
order of magnitude for R[, all we need to do is estimate this integral by considering
that all the lengths entering into it are of the order of R. In that case, we find that
the denominator is of the order of lOJTr]R 2 v, while the numerator is equal to
1PbJT R 3 v 2 . Within these approximations, we find that
R[
Pb
Rv.
r]
(21)
We might point out in passing that the problem could have been solved directly
by using Stokes's formula for the drag of a sphere. We also note that (21) bears a
striking similarity to the standard expression for Reynolds's number, which gives
the ratio of convective flux to diffusive flux in a viscous fluid in motion. The reader
will recall that Reynolds's number is given by
Re
UL
v
=
P
UL,
r]
where U is the velocity of the moving fluid, and L is a length characteristic of that
fluid. It is clear that U and L playa role similar to v and R, respectively, in (21).
The only difference between the two definitions is that in the case of a particle in
motion in a fluid, it is the volumetric density of the particle that matters, rather
than that of the moving fluid itself. With this restriction in mind, the number R[
introduced above can be thought of simply as the Reynolds number for the problem
at hand.
It is useful to clarify the practical implications of (21) with the help of a numerical calculation that will put some orders of magnitude in perspective. If we
consider a particle of silicate sand, of density 2.2 g/cm 3 and diameter 1 mm, moving at a speed of 1 cm/s in dry air, R[ turns out to be roughly equal to 104 . This is a
reassuring result which gives us confidence that it is perfectly legitimate to regard
particles with this particular mass and speed as a dry granular. If, on the other hand,
we were to deal with a powder made of particles whose typical size is only 10 !Lm
(e.g., lycopodium powder), rather than the 1 mm used in the previous example,
the number R[ would decrease proportionately. It could easily get smaller than 10,
which would sound a warning that the viscosity of the surrounding fluid can no
longer be safely ignored. This simple calculation illustrates that, even in favorable
experimental circumstances (such as a noncritical system of particles in motion),
it is advisable to exercise caution. It is imperative to be alert to the possibility that
our approximations for a dry granular medium may break down when interactions
with the surrounding fluid are no longer negligible. 1
1Some
recent experiments, whose significance will become more transparent after reading Chapter 3,
have clarified some of the concepts introduced here [31].
22
Before leaving this subject, the reader will appreciate that the situation is likely
to be problematic in a fluid such as water, which has a coefficient of viscosity a
hundred times larger than air. Even with particles as small as I mm in diameter,
viscous interactions cannot be neglected in this case.
Turbulent Drag
Hydrodynamics taught us that, as soon as the relative speed of a particle in motion
in a fluid matrix becomes sufficiently high, instabilities appear in the boundary
layer between the fluid and the object in motion (Figure lOb). There is at that
moment a transition from a laminar flow to one that is turbulent. We also know
that the transition between the two regimes occurs for relative speeds of the order
of a few cmfs in dry air. As it happens, most granular flows involve velocities that
are precisely in that range. We can thus anticipate that, depending on the details of
the circumstances, energy dissipation will be either laminar (directly linked to the
viscosity of the fluid) or turbulent (related to the difference in dynamic pressure
ahead of and behind the moving object).
Here again, we define a number R t (where the subscript t now stands for turbulent) as the ratio of the kinetic energy to the work done by the opposing force
of turbulence F t over the applicable characteristic length, which, as before, is
typically of the order of the radius R of the particle.
This time around, though, the force Ft is proportional to the frontal crosssectional area S of the particle, to the square of the relative speed v, and to a
coefficient kt whose value is typically 0.24 for a sphere. Specifically, we write
Pov
Ft =kt 
S,
where Po is the density of the fluid. We then find that the number R t is approximately equal to
Rt
1 Pb
;::::::.
kt Po
This equation tells us that the number R t does not depend on the velocity (assuming
that the regime is indeed turbulent). It also shows that the concept of dry granular,
i.e., of a system of particles whose interactions with the surrounding fluid are
negligible, is valid for sufficiently heavy particles such as grains of silicate sand
moving through a gas such as air. 2 In this case the number R t is greater than
2The cmcial importance of the density of the medium in the physics of granulars was vividly illustrated
by a simple observation due to Chladni in the eighteenth century. He had noticed that small pieces of
horse hair torn loose from a bow mbbing against the strings of a violin tended to collect on the wooden
body of the instmment according to specific patterns. He proceeded to carry out a series of simple
experiments by depositing on a violin either fine sand or lycopodium powder. The sand piled up at
the nodes of vibration, while the lycopodium powder gathered at the vibration peaks, giving rise to
complementary figures (known as Chladni's figures). The explanation is that the very light lycopodium
powder responds to the vibrations of the surrounding air, while the sand is sensitive primarily to the
vibrations of the instmment itself. This, incidentally, constitutes a magnificent example of a physicist's
sense of observation, as reported by Michael Faraday in 1830.
2.1
23
FIGURE 11. Growth of a dendritic pattern when air is injected into a confined granular
medium. The photograph on the left shows the beginning of the process. The one on the
right shows a welldeveloped pattern roughly 10 cm in diameter (after [26]).
24
There are plenty of indications, both in the laboratory and in industrial situations,
that other perturbations can complicate things and tax our ability to model the
behavior of "dry granulars" as we have defined them earlier. Ambient humidity,
for instance, can cause serious disruptions by creating clumps of particles that are
more or less mobile. We know from common experience that wet sand can be fairly
cohesive, whereas dry sand crumbles apart readily. It is also intuitively obvious
that the smaller the particles, the greater the perturbationin the context of the
present discussionsince capillary forces are then apt to be of the same order of
magnitude as the gravitational forces that would otherwise largely determine the
behavior of these objects. Once again, some systematic simplifications enable us
to estimate how small the radius of spherical particles has to be in order for two
of them to remain stuck together by a thin layer of water, against the action of
gravity. The problem is illustrated in Figure 12.
Calculating the capillary force that keeps two wet spheres in contact is far from
trivial. Several methods have been proposed to avoid the difficulties associated
with solving the LaplaceYoung equation. The simplest one is known as the "pulley
method," because the contour of the meniscus is approximated by a pulley in the
plane tangent to the two particles. The "pulley" has an inner radius rj, while
the outer radius (of the ridge) is rz [33]. This approach gives an expression for
the capillary force in the form
Fc
= nYzvrz
(1 + ~~),
where YZv is the surface tension at the airliquid interface. We assume that, when
2.1
25
the weight of the lower spherical particle is just equal to the capillary force, the
radius of the liquid is some fraction a of the radius of the particles, or r2 ~ aR,
and the ratio r2f rl is of the order of 5. This series of approximations leads to a
result, correct within an order of magnitude, given by
R~ j4agPbYtv .
When applied to glass beads (Pb = 2.2g/cm3 ) and water as the liquid (Ytv
73 x 10 3 N/m), the above expression suggests that the beads will remain stuck
for diameters up to 1 mm when a = 1. With a more modest amount of water
(a = 0.01), the diameter of the particles should not exceed 100 {tm. This exercise
highlights the need to control humidity levels when doing experiments with granular materials assumed to be dry. We may also conclude that industrial processes
dealing with powders (particles smaller than 100 {tm) in open air are very likely
not to conform to the type of physics presented in this book.
It is less straightforward to incorporate electrostatic interactions in a simple
model, primarily because of the difficulty in quantifying even approximately the
electric charges that develop on the surface of granulars in relative motion. It is
somewhat easier to estimate them empirically on the basis of simple experiments,
such as the following. Drop a few steel balls, 1.5 mm in diameter, into a small
plastic tube (Figure 13). Shake the tube and its contents vigorously and place it
back vertically on a table. You will then notice that a certain number of balls remain
stuck to the walls of the tube. The position of each of these balls in precarious
equilibrium obviously depends on the position of its neighbors, since they all
interact electrostatically.
Knowing the weight of each steel sphere, and assuming that it is exactly offset
by electrostatic repulsion, it is easy to calculate that each one carries an electric
charge of the order of 3 x 109 C (Coulomb), which corresponds approximately
to 300 {t C/kg, or a surface charge density of 4 x 10 8 CJcm2 . These numbers
seem to agree with those usually observed in industry. Experience shows that
organic materials are much less sensitive than minerals (by about a factor of 100),
and that the amount of accumulated charge depends on the physical and chemical
nature of the surfaces. Furthermore, as mentioned earlier, humidity minimizes this
type of perturbation, although the actual mechanism is not entirely clear. Finally,
the existence of these surface charges constitutes a danger in industrial facilities,
which tend to store large quantities of dry grains (such as corn). The reason is that
they can trigger explosions of gases generated in silos by the decomposition of
organic substances. A standard way to mitigate the risk in laboratories is to coat
sensitive surfaces with antistatic sprays. This solution is unfortunately completely
impractical on an industrial scale.
Classification of Granular Materials and Definitions
It is useful to familiarize oneself with the terminology traditionally used in the field
granular materials. The following definitions are taken from Brown and Richards,
26
FIGURE 13. Small steel spheres suspended by electrostatic interactions on the walls of a
plastic tube.
In light of these definitions and of our previous remarks, it should be clear that
the physics of dry granular materials is limited in practice to granular solids and
aggregates. Excluded are particulate systems such as powders, suspensions, or
fluidized beds, in which interactions with the surrounding fluid are predominant.
2.2
27
2.2.1.
The "macroscopic" laws governing the static and dynamic properties of two solids
in contact are based strictly on experiments. They are remarkably simple and
have stood the test of time with amazing durability. They have survived virtually
untouched for almost five centuries, even though they describe phenomena that
have yet to be fully understood on a fundamental level. Leonardo da Vinci gets
credit for being the first to realize in the sixteenth century that, in order to get a solid
to start sliding against another, it is necessary to apply to it a tangential force that is
proportional not to the surface area in contact but, rather, to the force pressing the
two objects together. This property was rediscovered in the seventeenth century by
Guillaume d' Amontons, and verified independently by La Hire at the insistence of
the members of the French Academy of Sciences who were at first incredulous. In
1750, Leonhard Euler (17071783) introduced the concepts of static and dynamic
friction. All these properties are summarized in a set of three fundamental laws.
Interestingly, it was not until the second half of the twentieth century that they
received a credible explanation in terms of interactions at the microscopic level
[34].
28
~
...
FIGURE 14. Leonardo da Vinci's traction experiments. The force of traction T necessary
to set the disks PI, P2, and P3 is the same in both configurations.
A Microscopic Explanation
Micrographs of the surface of ordinary solids such as metals reveal a ragged topography of the type schematically indicated in Figure 15. It is crucial to understand
that Coulomb's second law, which states that the force oftraction required to set a
solid object in motion is independent of the surface area in contact with the base,
is justification enough for studying the physics of friction on the scale of a single
protrusion (since the result does not depend on the number of such features).
These ragged features, whose typical size is of the order of 1 t.tm, are so small
that, even under weak stress conditions, the deformations they experience substantially exceed the elastic regime in which Hooke's law remains valid (see some
2.2
29
E::i..
1:
Ol
'Qj
10
Distance (flm)
FIGURE 15. Schematic magnified view of the surface of a polished metal sample.
orders of magnitude in Section 2.2.2). 3 This implies that the protrusions in contact
deform at constant pressure p until they exactly offset the normal load force N
(normal in the sense of being perpendicular to the average plane of contact, as
shown in Figure 16).
If A is the surface area in contact, then the relation N = pA holds when equilibrium has been reached. The effect of a sufficiently large tangential force T is
to separate the surfaces in contact. There is a constant s which characterizes the
ability of the material to resist this shearing action; it is given T = sA. In this
picture, the static friction coefficient {Ls defined previously can depend only on the
physical constants of the materials in contact. Its value is
{Ls
T
L
= 
= p
As it happens, the constants sand p vary roughly in the same proportion for many
different surfaces. 4 In particular, metals are characterized by
s
0.6::::  :::: 1.2.
p
This model accounts for the tightly clustered range of values of {Ls. It also explains the experimental observation that the contact regions between two solids
deform slowly under the weight of the upper object, which is fully consistent with
the behavior expected in the plastic regime. We shall present in Section 6.4 a
more accurate description of these laws of friction in the context of numerical
simulations.
3 This
becomes even more obvious if we appreciate that the real contact surface area between two solids
can be hundred to thousand times smaller than the "apparent" area. When distributed among a few
tiny spots, the stress due to weight can easily be much larger than the elastic limit.
4This statement is not true of materials whose mechanical properties are strongly anisotropic. For
instance, materials with a lamellar structure such as graphite exhibit a strong resistance to deformations
in a direction perpendicular to the stacked layers, but a weak one to shearing in the plane of these
layers. That is precisely the property that is being exploited when mechanical parts are lubricated
with graphite.
30
where Rand R f are frames of reference attached to solids (S) and (Sf), respectively.
As written, the above equation describes the motion of point M in the frame of
reference of solid (Sf). For the purpose of simplification, we assume the solid (Sf)
to be fixed. At any given instant, a generalized rotation (pivoting or rolling) is
given as a sum of two orthogonal components
W
= W n + WI.
The vector WI corresponds to a rotation in the plane of the figure, while Wit describes
spinning about a vertical axis. The dynamic equation written above includes three
5The notion of a rigid solid is purely theoretical. It clearly conflicts with the picture we have painted
of solid friction, which is based on contacts at a number of points undergoing plastic defonnation.
This is only one of several difficult issues raised in making the transition from classical mechanics to
the mechanics of real solids, particularly in granular fonn.
31
(8')
Vp
and
Vr
= MI X Wt
Frustrated rotation
Rolling without gliding
32
We will deal on several occasions (see in particular Sections 3.1.1 and 4.2.2) with a
type of motion frequently observed in situations involving dry friction. The motion
in question is known as stickslip. It typically results from interactions between
particles that are subject to the CoulombEuler laws of friction and also exhibits
an elastic behavior. Given that the granular materials we are specifically interested
in often conform to both requirements simultaneously, it should come as no surprise
that the stickslip mechanism is commonly observed in granular matter. Although
we will have an opportunity to give a more general treatment of the phenomenon
later (see Section 4.2.2), we nevertheless introduce here a simplistic but instructive
description [35]. It is based on the model illustrated in Figure 19.
An object of mass m is placed on a conveyor belt moving at a speed v. The object
is also tied to a stationary post by a string of stiffness k. The friction between the
object and the belt is characterized by coefficients fts (static) and ftd (dynamic),
33
such that j.1,d ::'S j.1,s, in accordance with the CoulombEuler laws. For the sake of
simplicity, we assume that j.1,d is negligibly small.
We start at time t = 0 with the spring in a state of rest, at which point its length
is x = xo. Because of dry friction, the object starts moving at a constant speed v.
The horizontal component of the reaction associated with the friction opposes the
traction of the spring, which we write as
T = k(x  xo) = kvt.
The uniform horizontal motion continues as long as Coulomb's condition is verified, that is, as long as
T
IN
I=
The condition is met for a duration t ::'S t[ such that t[ = mgj.1,slkv. As soon as
t > t[, the force of friction gives way (suddenly becomes equal to 0), and the
motion of the mass m obeys the usual differential equation
d 2x
mdt 2
k(x  xo).
+ A sin[woCt 
tl)
+ a],
vJt?
+ 1/w6.
The motion is sinusoidal until time t2 when the velocity of the object relative to
the belt is zero. At that instant, static friction kicks in again. The time t2 is given
by one of the roots (the one that is different from t2 = tl) of the equation
COS[WoCt2  tl)
+ a]
= cos(a).
34
:E
'"
&
Time
FIGURE 20. Dynamics of a periodic stickslip motion. Straightline segments (stick phase)
are connected to portions of sinusoids (slip phase).
This leads to
t2
= t] + 2[(n
x  Xo
= A sin[woCt2 
t])
+ a] = A sin(a).
Once again, the object is dragged by the belt at constant speed v until time t3. The
distance covered during the interval t3  t2 is 2A sin(a). The complete periodic
motion xCt) is schematically depicted in Figure 20.
It is composed of straightline segments joined by sinusoidal sections. The
period to is given by
2
2A sin(a)
to=(na)+
.
Wo
v
The amplitude is the same as that of the sinusoid. At low speed, the diagram
becomes a sawtooth made of a series of straight lines. In that case, t] = mg Itslkv.
We might also point out that the value of the static friction coefficient Its can be
determined by measuring the maximum amplitude of the motion. As a further
exercise, the reader is encouraged to look at the shape of the function F (v) which
describes the friction force as a function of the speed of the belt, and to show
that it has regions of negative slope (see Sections 2.3 and 4.2.2). This simplistic
model can also explain the forced vibration of a violin string; because the bow is
normally made of hair coated with colophane, its interaction with the string is a
classic example of dry friction. In a more prosaic vein, stickslip is responsible
for the squeaking of poorly lubricated doors, the bucking of machine tools, and
other similar phenomena.
2.2.2.
35
FIGURE 21. Two elastic spheres exchange momenta during a headon collision.
collinear. This is, of course, a highly improbable event with real granular solids,
which most often behave as inelastic objects, subject to friction, and experiencing
collisions at various angles. Nevertheless, the ideal case is useful to look at because it involves two fundamental tenets of classical mechanicsconservation of
momentum and of kinetic energy.
Referring to the notation indicated in Figure 21, the speed after impact is given by
Uj
m2
Vj
+m2
mj mj
2m2
v2.
mj +m2
In actuality, collisions between real granulates always entail some loss of momentum and kinetic energy. Some fraction of the incident momentum is given up to the
colliding particles by exciting sound waves propagating in their interior. In the process, a portion of the elastic energy stored in both particles is dissipated in the form
of acoustic waves or phonons that relax by heating up the mass of both particles. A
loss of kinetic energy can also result from a permanent deformation incurred during
a collision. In any event, we observe experimentally that a ball hitting a vertical
wall of infinite mass with a velocity v rebounds with a smaller velocity cp v (with
cp :s 1). To a first approximation, and in the case of frontal collision between two
identical spheres, it is convenient to describe the velocity shortfall in the frame of
reference of the center of gravity of the system by a matrix equation [37]
[::J ~c,t] ~
rT :n[~:l
(22)
36
with some caution. On the other hand, it is often sufficient to describe experimental
observations, as we ourselves will realize when we discuss numerical simulations
in Chapter 6. As long as we are prepared to adopt this approach, it is worthwhile to
emphasize once again that E is truly a measure of the momentum loss of colliding
particles. If P and pi designate the momentum of the total system immediately
before and immediately after impact, we may write in the frame of reference of
the center of mass
p
= m12(vl
 V2),
pi = ml2(ul  U2),
where m 12 = m 1m2/ (m 1 + m2) is the reduced mass of the system of two colliding
objects. It is straightforward to verify that (22) and (23) are indeed consistent
with the definition of E in the form
E
pi
Ul  U2
VI  V2
=  = 
where the minus signs in front of the ratios account for the fact that the velocities
reverse direction after the collision. The change in kinetic energy llEkin after
impact is easily shown to be given by
llEkin
As mentioned earlier, a frontal elastic collision, that is to say, the impact of rotationless particles along a path joining their centers, is an extremely improbable
event that, to a first approximation, does not involve friction. The reality is far
more complex in granular media. It includes nonfrontal collisions and frictional
effects giving rise to gliding and rolling motions that can impart angular momenta
to neighboring particles. We proceed to give a somewhat simplified treatment of
the problem [38], [39]. A more accurate description turns out to be extremely
complicated.
As an introduction to a subsequent section in which we will outline a technique
useful for numerical simulations of rotations, we develop next a classically mechanistic description of the gliding and rotation motions of two solids in contact [40].
A Ball Thrown Against a Wall
We consider the simple problem of the collision between a spherical ball and a
vertical wall. Let V x and v y be the components of the velocity of the center of
mass of the ball, as shown in Figure 22, at the instant to of impact. Let Wo be
its angular momentum just prior to impact. We assume that the rotation vector is
perpendicular to the plane of the figure. Under these conditions, it is intuitively
obvious that the entire problem must retain the symmetry of the plane. Our goal
is to determine the components U x and u y of the velocity immediately after the
collision, as well as the final instantaneous rotation WI. We designate by X and Y
37
x
x
Y~_"'l<I>
If, on the other hand, we introduce a nonzero friction between the ball and the wall
in the form of a single coefficient fL, we must distinguish two separate cases.
1. The Gliding Velocity Remains Positive for the Entire Duration of the Collision
In this case, the system of equations describing the exchange of momenta (linear
and angular) reads
m(u x
vx )
m(u y
vy)
~ma2(wl  wa)
=
=
X,
fLX,
= afLX,
where m is the mass of the ball (assumed solidly filled), and a is its radius. It
is immediately apparent that this system of four equations with four unknowns
(X, u X , u y , and wI> can be solved if all relevant quantities before the collision are
given and if the coefficients of friction fL and elastic restitution 8 are known. We
leave this as an exercise, and focus our attention on the phenomenology of the
collision, i.e., on the gliding and rotation motions of the ball during the collision.
We note first that the normal momentum X = mvx(l + 8) is indeed positive,
as it should, since V x is negative. The final gliding velocity must also be positive,
38
2: fh (1
+ E)
V y  awo
< '
v x
2. The Gliding Velocity Drops to Zero at any Time t1 During the Collision
In that case, the general relations read
m(u x
v x ) = X,
m(u y
vy)
= Y,
u y  aWl
= O.
This time around, we have a system of five equations in five unknowns, which can
easily be solved. We still have to verify that the linear momentum makes an angle
less than tan 1 (fh) and, in addition, that
The last two conditions turn out to be mutually exclusive. It follows that the
behavior will be of one type or the other depending on the value of the friction
coefficient fh. 7
We may now ask ourselves whether these mutually exclusively situations are the
only two possible outcomes. Is it at all possible for the gliding velocity to change
sign during the interval [to, td of the collision? If the answer is yes, the condition
IY I < fhX could still be verified without us being able to make any prediction
whatsoever about the final value of (u y  aWl). The following reasoning gives the
answer.
Consider an instant t E [to, td, and let I] and l; designate the components of the
reaction force during impact. The equations of motion at that particular instant are
dvx
mdt
dv y
md"t
2
sma
2 dw
=1],
=l;,
dt =
l;a.
7We note here that this mechanistic model assumes that the forces of friction come into play instantaneously during the collision. From the point of view of a physicist, this condition is unlikely to
be realized in all situations. Indeed, phenomena of friction (Section 2.2.1) and collisions between
solids (Section 2.2.2) involve plastic deformation, a process that is notorious for being not at all
instantaneous.
(a)
39
(b)
FIGURE 23. Two spheres (a) before and (b) after a nonfrontal collision. The parameters
are defined in the text.
= :is.
7
As it turns out, the quantity t;, which is the tangential component of the reaction
force during contact, has a sign opposite to that of (v y  aw). It follows that it can
only decrease in magnitude. Should it reach zero at any time during the collision,
it will remain zero up to the end. Stated in plain language, should the gliding
speed happen to drop to zero at any time during a collision, the ball would then
go on gliding for the remainder of the collision. This simple argument shows that
the two conditions considered previously are indeed mutually exclusive and that
they are the only two possibilities. One or the other must prevail depending on the
values of the friction coefficient JL and of the components of the incident velocity.
The same distinction between two regimes will come up again when we study
nonfrontal collisions between two spherical particles.
40
The relative velocity of the two particles at the point of contact is given by
where Vi and Wi are the translation and rotation velocities, respectively, of particle
i before the collision. Note that the magnitude of the relative speed IVe I increases
when the individual velocities point in opposite directions and when the rotation
vectors of these particles point in the same direction. The velocity Ve has a normal
component V~,) = (v e . n)n, while its tangential component is v~r) = Ve  V~').
The vector v~r) defines a unit tangential vector such that t = v~)/ Iv~t) I. The angle
of impact y is defined as the angle between the normal n and the relative velocity
Ve , and is such that y E [n/2, n]. Figure 23(a) shows the situation of the two
colliding particles in the particular case when WI = W2 = O.
We now consider what happens to the kinetic momentum during the collision,
using the same convention as before, namely that Oi designates the speedfollowing
the collision. We have
(24)
The normal component of ~p has no effect on the angular rotation speed, but the
tangential component does. By creating a torque on the arm defined by the vector
(dJf2)n, the change in momentum ~p causes a change in angular momentum
such that
nx ~p=
21
dew w),
(25)
where 1 is the moment of inertia with respect to the center of the particle and Wi
is the unknown angular rotation speed after the collision. It should be appreciated
that (25) indicates that the change in angular momentum is the same for both
particles. Figure 25(b) depicts the situation after impact.
If ~p is known, it is possible to calculate all relevant velocities following the
collision with the help of (24) and (25). The results are
~p
OJ = V j +  ,
mj
~p
02
Wi
1
Wj 
dj
n x
W2 
 n
V2   ,
m2
2It
d?
x
2h
~p
'
~P.
ev(lI)
'
(26)
41
which relates the translation velocities before and after the collision. Equation (26) applied to the normal component of the sum ~P/ml + ~P/m2 allows
us to calculate the normal component of the change in linear momentum, giving
the result
~p(n)
mdl
+ s)v~n),
(27)
where, as before, m 12 is the reduced mass of the system composed of m 1 and m2.
Coulomb's law establishes a relation between the normal and tangential components of ~P, namely 1~p(t) 1 = ~ I ~p(n) I. Since collisions are always dissipative,
the vector ~p(t) must be aligned along to We can thus write
~p(t)
(28)
where we have v~n) = Vc cos(y), since cos(y) is always negative. We also have
the relation t = v~t) /[v c sin(y)]. Combining (27) and (28) yields an expression
for the change in momentum ~p
(29)
The second term in (210) has the same form as the first, except for a factor of ~
which comes about through the bias of the moment of inertia of a solidly filled
sphere. 9 In light of our previous remarks, the coefficient of tangential restitution fJ
9This factor is obviously different for other shapes of colliding particles. For instance, it is equal to ~
for disks, and 2 for thin rings.

42
must be equal to the smallest of two values fJo or fJl corresponding to one of two
different regimes:
Large values of the angle y (y :::: Yo) correspond to a gliding contact, in which
the contacts have been broken. In this case, the appropriate choice for fJ is
fJ = fJo, withfJo E [1, +1] .
Small values of y (y ::::: Yo) correspond to a collision for which the interaction
can be described in terms of dry friction and elastic restitution. In that case,
the correct choice is fJ = fJl, with fJl = 1  ~ fL(l + E') cot(y).
The angle Yo marks the dividing line between the two regimes, which occurs when
fJo = fJl. The value of Yo is given by the equation
tan(yo)
1 + E'
1 + fJo
fL.
2
Reference [41] gives additional details on some practical aspects of the two modes
of collision by considering the ratio of the normal and tangential collision velocities.
This allornothing simplification makes it possible to pursue numerical calculations of situations involving nonfrontal collisions and rotations. We will discuss
such models using rigid spheres to simulate particles undergoing multiple collisions (see Sections 3.2.4 and 6.1.2).
Before doing so, however, it is useful to first explore in more depth a number
of problems associated with the mechanics of collisions. To that end, we proceed
to calculate some orders of magnitude of a few phenomena occurring during the
frontal collision of two spheres that can penetrate each other, as is almost always
the case in the real world.
Penetration During Frontal Collision: Hertz's Problem
Consider two identical spheres of radius R and mass M, approaching each other
at a relative velocity v, as shown in Figure 24. Hertz has shown that the elastic
energy E e stored by two spheres deformed over a depth h is given by
5/ 2
E e  lkh
2
'
43
= 4~ _E_v'R.
15 1 
(211b)
(J2
The quantities E and (J entering in (211b) are Young's modulus and Poisson's
coefficient, respectively. Upon impact, the initial kinetic energy is converted partly
into a reduced kinetic energy and partly into stored elastic energy. Thus, we write
Mv
dh)2
= kh 5/ 2 + M ( dt
(212)
The velocity drops to zero when the two spheres have penetrated each other by a
distance h o. At that moment, dh/dt = 0, and we have
ho 
M
k
2/5
)
V4/ 5
(213)
rho
r=2 Jo
/v 2
dh
4fir(~) (M
(k/M)h 5 / 2 = 5r(!o)
k2
2)I/5
v
(214)
M 2)1/5
= 2.94 ( k 2 v
6 x 1011 dynes/cm 2 ,
(J 8:::i
0.3,
10 10 ( cgs units).
44
very much larger than the number calculated above. The beads need only
have a velocity two or three times larger to reach the plastic regime. Most
importantly, the deformation is of the order of the height of the surface roughness. In accordance with what was said in Section 2.2.1, the deformation of
the surface protrusions during a collision is necessarily plastic. The implication isand experiments confirm itthat the degree of polishing of the
interacting objects plays a crucial role in the process of restitution during
impact [34J. We will see shortly an approach to describe the penetration of
inhomogeneous spheres with a surface layer softer than the core.
Part of the energy may be dissipated by exciting sound waves inside the solid
beads. As an example, we may calculate the time Tph it takes a phonon to
complete a round trip across a sphere, knowing the propagation velocity Vph'
The result is Tph = 2R/v ph = 0.15 cm/6 x 105 cm/s = 2.5 x 107 s. In other
words, there is plenty of time for several round trips during a typical collision.
This effect constitutes a potentially important perturbation that in some cases
may invalidate the model embodied in (212).
(b)
e
FIGURE 25. Schematic ofthe zone affected during the penetration of two (a) homogeneous
and (b) inhomogeneous spheres.
2.2
45
homogenous spheres. In order to understand the reason for the change, it might
be worthwhile to first outline an argumentoriginally advanced by de Gennesexplaining the usual ~ exponent.
Consistent with the notation of Figure 25, h is the depth of penetration during a
collision, and R is the radius of each sphere. From elementary geometry, we can
easily derive an expression for the radius a of the contact circle. Assuming that
h is very much smaller than R, which is justified if we insist on remaining within
the elastic regime, the result is a 2 ~ Rh.
It is not unreasonable to assume that the deformation of colliding spheres affects
a depth roughly equal to the radius a of the contact circle. Under these conditions,
and in the presence of a stress P, the classical definition of Young's modulus gives
a relation of the type
~E.
This justifies the ~ power law. If we now consider a sphere covered with a soft cmst
of thickness e much smaller than the radius R of the sphere, but sufficiently large
compared to the penetration depth h, we can assume that the deformation will be localized within the outer layer whose Young's modulus is E e E. In that case, the
depth affected will be of the order of e, rather than a as in the previous case. In other
words, a stress P acting on the sphere now produces a penetration depth given by
The relation between penetration depth and force has changed from h ex p2/3,
valid in the case of Hertz's homogeneous spheres, to h ex pI /2. The deformation
increases more slowly with applied force. This makes sense if we note that, in
the presence of a soft crust, the deformation is more localized to regions near the
contact point and does not spread as deeply into the core.
Such a dependence can explain, at least in part, certain experimental observations
that deviate notably from what would be expected on the basis of Hertz's model
[44]. This is not the only possible explanation. Others have been proposed, invoking
notions of disorder and enrichment in the vicinity of contact points under the
influence of stress [45]. We shall discuss a particular example in Section 3.1.2.
46
FIGURE 26. A sphere rolling down a rough surface looses kinetic energy by friction and
successive collisions. Its trajectory is a mix of ballistic arcs and sections of curves following
the rough surface.
2.3.
.
F
r = sm(8)
.
mg
(215)
A uniform motion results whenever r = 0, i.e., when the friction force exactly
offsets the effect of gravity. It then becomes evident that the problem boils down
to working out a plausible expression for the force of friction F, as we will see in
several examples later in this book (for instance, in Section 4.2.2). The idealized
2.3
47
case described above helps us to understand that the force of friction corresponds
to a loss of energy as the particle proceeds down the slope. This allows us to write
F = aEjax, where x corresponds to the distance covered. The energy of the
particle of mass m has two components:
(1) its kinetic energy E b which is dissipated over a characteristic distance A as
and
Ep
= mgh cos(e),
(!
mgf"...h
1
2D
F p = ~ ~ 'img v 2
The equation just derived suggests that the corresponding dissipative force
becomes less and less effective as the speed of the particle increases, which
certainly conforms to our intuition.
At low velocity, the path of the particle follows the profile of the surface of
the granular support. The resulting loss of potential energy should no longer
depend on speed. It corresponds only to the dry friction experienced by the
particle as it moves down the surface. In this case we have
F p = mg.
1
1 + {3(v 2 jgD)'
~mg:,
48
+ Fc =
mg [
t )
l+b v 2 jgD
+ c (~)],
gD
(216)
LLI~ 3
c
t5
it
49
a=b=1
c= 0.25
""0
co
0 ++..~
o
2
3
Normalized Velocity
FIGURE 27. Effective friction force F as a function of the particle's velocity. The curve
does not go through the origin and exhibits a minimum.
2.4.
2.4.1.
Having mastered the laws of friction between solids reviewed earlier in this chapter,
we are now in a position to explore how a real granular material reacts to the kind
of shearing force involved in Leonardo da Vinci's experiments. The experimental
device illustrated in Figure 28 was originally devised and used by Dawes [5].
It consists of two open boxes with individual compartments shaped like a comb. 10
The two boxes are filled with a granular medium (such as a powder or river sand)
and placed upside down on top of each other. A lateral shearing force F is applied
to the upper box, which is also subjected to a vertical load P.
Without going into the details ofthe experiments, we state the following experimental observations:
The shearing force required to set the upper half in lateral motion is strictly
proportional to the load P and is independent of the surface area under shear
F=~sP.
This result is entirely consistent with the first law of solid friction reviewed
in Section 2.2.1. 11
IOSoilmechanics scientists routinely use boxes of this type, roughly a meter long, to measure granular
friction coefficients out in the field.
II It appears that Dawes did not try to verify that the force F was independent of the surface area of
the boxes he was using [51. It also appears that Coulomb implicitly accepted this independence by
writing that granular materials obey the laws of friction as stated by d' AmontonsLa Hire [II]. This
was indeed confirmed in recent experiments.
50
FIGURE 28. Typical experimental setup used to measure the coefficient of the static friction
fhs of a granular material.
The coefficient fts is of the order of 0.7, which is quite comparable to values
typically encountered in the case of friction between solids. The angle e,
given by the formula e = tanlefts), is approximately equal to 35. It is
useful to keep this value in mind, as it will resurface in Chapter 4.
We note in passing that these results hold equally well for cohesive and noncohesive granular materials (cohesive materials are made of particles kept together
by cohesive forces). The only difference is that in the cohesive case the above
expression must be modified by the addition of a constant term C; the expression
then takes on the form F = fts P + C.
These results have actually been long accepted by people working in the field of
soil mechanics, who routinely make use of the concept of stress. 12 Nevertheless,
these properties constitute a rather stunning discovery. Noone could have predicted
them, given the complexity of granular piles and the many unknowns concerning
their equilibrium (see Section 3.3.1). In particular, it seems quite unlikely that a
model based on plastic deformations, such as we invoked in our discussion of the
laws of friction between solids, could be extrapolated to the present situation. The
best we can do at this point is to accept that a medium as complex and heterogeneous as a granular can lead to exceedingly simple macroscopic laws. As an aside,
the same can be said of two sheets of cardboard in contact [48].
2.4.2.
Bagnold's Number
51
FIGURE 29. The flow of a granular material medium normally involves only a few layers
near the free surface. The mean velocity of the particles decreases rapidly with depth.
Consider a cylindrical container filled with a granular suspension, i.e., a collection of granules in a viscous liquid. A second smaller cylinder, concentric with
the first, and dotted with large protuberances, is rotated inside the first. We are
interested in the resulting flow pattern of the particles. The following discussion
also applies to an "avalanche" flow, obtained for instance by tilting a table covered
with a thick layer of granules, as illustrated in Figure 29.
Experiments show that any flow takes place in sheets. The movement of the
granular material occurs in stratified layers parallel to each other, each layer being
characterized by its own velocity, different from that of the adjacent layers. The
medium can be viewed as a succession of layers slowed down by friction with
layers on either side. We can introduce a parameter called the rate of shearing of a
sheet of granular material in motion on top of another sheet moving at a different
velocity. This parameter is a measure of the average velocity gradient and is defined by y = (avjaz), where z is the depth. It is evident that the rate of shearing
must figure prominently in constructing our model, particularly to account for the
loss of kinetic energy by friction. Let D be the diameter of individual particles
of mass m. The quantity Dy measures the relative velocity of this particle with
respect to the underlying sheet. The excess loss of kinetic energy, which occurs
over a characteristic length Ae typically of the order of a few diameters D, is due
to the friction force Fe. This we may write
Fe _(aEe)~mD2
  ~   y2
.
ax
2A e
We also note that the quantity 17Y, where 17 is the viscosity of the interstitial
fluid, represents the friction force F v due to viscous interaction with the fluid,
divided by the crosssectional area of the solid in motion, which is of the order
of D 2
Bagnold introduced a number B defined as the ratio of the forces due to friction
52
and collisions between solids to the forces exerted by the viscous fluid. 13 In other
words
Fc
my
B=~
Fv
2A e 17
13 Actually,
Bagnold started out by introducing the notion of dilation 8 of a granular mixture, defined as
8 = D/A e . Bagnold's number is then equal to 8 1/ 2 Ps D 2 Y/1J. This quantity merges with the number
B defined in the text in the limit of concentrated solutions, for which (Dy)I/2 is indeed of the order
of D.
3
Fluidization, Decompaction,
and Fragmentation
Common sense tells us that, in the absence of any external perturbation, a pile
of granular material is at rest. It can be inclined, gently shaken, and submitted to
various mild stimuli without anything drastic happening, in other words without
triggering a flow or collapse, and without affecting the relative positions of the
granules. We also know from experience that a sufficiently high inclination can set
off an avalanche, possibly even a continuous or intermittent flow; that vibrations
cause a granular material to behave somewhat like a liquid; and that sand deforms
readily when crushed. All these properties are markedly different from those of
the normal solids, liquids, and gases we are accustomed to. The main difference
appears to be that granular materials must be "sufficiently" perturbed in order to
change. This observation strongly suggests that we will be dealing with issues
such as thresholds, nonlinearities, and hysteresis effects.
The aim of this chapter is to clarify some principles and highlight some of the
difficulties encountered when trying to understand why and how a granular pile
changes configuration. We are particularly interested in phenomena of decompaction,jiuidization, and fragmentation. As the chapter progresses, we will come
to realize that these phenomena can overlap, sometimes even act in concert, and
that they constitute different facets of a more complex reality. For instance, we
will show that progressive decompaction and convection, which show up in longterm observations, can actually result from a subtle superposition of discontinuous
processes and fragmentations that typically occur over short periods of time. What
this means is that, as is sometimes the case in other branches of physicsmost
notably in hydrodynamicsthe evidence for a particular phenomenon depends on
the duration of the observation.
However, before tackling the problem of how granular stacks deform, we will
take the time to analyze briefly the static properties of a sand pile. Paradoxically,
J. Duran, Sands, Powders, and Grains
SpringerVerlag New York, Inc. 2000
54
...
. . .
.. ...
..
..
.'
....
... . ...
._. .. . _,,_,_
_M"~__
_~
Contact Points
FIGURE 30. The classic "cannon ball" stack. Even though the structure appears to have a
high degree of order, many parameters are actually undetermined.
the static properties pose even more delicate challenges than their dynamic counterparts, which we will examine later.
The very nature of contacts between solids and, more importantly, the mechanism
of solid friction (see Section 2.2.1) govern the static and dynamic properties of
granular media in a complex way. When dealing with systems of many particles,
it becomes quickly apparent that the problem is almost inextricable because it
has a large number of solutions that depend very much on the past history of the
material. The solution Nature chooses is almost never predictable, as it involves
some degree of disorder of the forces of contact between particles. It is essential
to clearly understand this point before proceeding any further. The goal of the
next few paragraphs is to convince the reader of the intrinsic complexity of the
problem at hand, as well as to highlight some key ideas that will prove of great
use subsequently.
The Stacking of Cannon Balls
interesting to note that the static properties of this structure continue to be the object of a great many
theoretical articles, whose results, incidentally, do not always agree with the (difficnlt) measurements
that can be made.
3.1
55
FIGURE 31. Collision of three particles. Classical dynamics requires a precise knowledge
of the order in which the collisions occur, as well as of the nature of the surfaces.
56
related to the plastic deformations of the microcontacts, the nature and magnitude
of this reaction force are perfectly known. The tangential reaction, on the other
hand, is quite another matter. The corresponding force, called "resistance to friction" because it opposes lateral motion, is inherently indeterminate, which shatters
any hope of describing precisely and exactly the static properties of objects in friction. While it is clear that the force R opposed to gliding cannot exceed a critical
value F/Ls, where F is the force that keeps the objects together, it is no less clear
that, as long as the system is at rest, R has an unknown value that can be anywhere
between 0 and F/Ls. We can think of the problem by imagining that the microscopic protrusions get deformed and oppose subsequent displacements, although
we do not know precisely the details of these deformations. To complicate matters
further, the deformations in question may well take place in the plastic regime,
which implies slow changes. In other words, strictly speaking, we ought to treat
the force of resistance to gliding as a timedependent problem. It is also apparent
that a complete description of all the forces involved between two solids depend on
the prior history of their contact, specifically on how that contact was established.
The situation is illustrated in Figure 32. Part (b) of this figure shows that the
force acting on a brick placed on a support featuring two perpendicular slanted
surfaces depends on the way the brick was placed. It might have first been set
free on the left surface and subsequently pushed against the perpendicular wall, or
the other way around. The same ambiguity exists with regard to the equilibrium
of spheres resting on each other, as shown in Figure 32(c). In all these cases, the
solution to the problem depends on the prior history of the system.
It would then seem that any attempt to describe the forces of contact in a granular
edifice is doomed from the outset. It turns out that this is not always the case, as
calculations are often possible as long as enough information is available about the
way equilibrium was reached. This is in fact quite a general property of granular
materials. We will illustrate the point with a particularly simple example which
we will revisit from a different angle in Chapter 4.
(b)
?
(c)
FIGURE 32. The forces of friction between solids can make it impossible to determine
unambiguously the magnitude of all the forces at equilibrium.
3.1
57
subjected to are perfectly known. They include its weight P, the normal reaction
S of the surface, the reaction R = kx of the spring (where x is the deviation
of its length from equilibrium and k is its stiffness, or spring constant), and the
force of friction with the surface, i.e., the tangential reaction F. The angle e can
vary between 0 and n /2. We know the coefficient of static friction fhs, and for
simplicity we assume that the coefficient of dynamic friction fhd is zero. Our goal
is to determine the compression x of the spring as a function of the angle e.
Finding the exact solution to this problem requires a bit of caution. We start
with two extreme cases. First, when e = n /2, the solution is trivial since the force
pressing the brick against the surface vanishes. Therefore, the friction force F is
itself zero, in which case we have x (n /2) = P / k. The case of a horizontal surface (e = 0) is already more complicated. The solution is actually undetermined
because the spring may have been left initially in a compressed state defined by x,
with x allowed to have any value as long as the reaction R is less than the tangential
reaction due to friction. In other words, x < Pfhs/ k. Here is a case where the
position XQ of the brick cannot be determined without some additional knowledge
about the way the configuration of the system was arrived at.
Suppose now that we start with the spring at rest (xQ = 0) and we gradually
incline the surface from 0 to n /2. We denote by e+ angles that are reached by
increasing the inclination, and e in the opposite case. The tangential component
ofthe load due to the weight increases as P sin(e+). Since the spring is initially at
rest, this load is the only force that is offset by an equal and opposite friction force,
until the moment the brick starts moving when tan(e+) = fhs. The brick then
drops down without friction (since we assumed fhd = 0) along the surface until
the spring stops in a new position Xe+,
which will happen when kxe2+ = P sin(e2+).
2
The brick being once again at rest, the friction force is reactivated, and the surface
must be inclined to a new larger angle ei to trigger the next motion. The angle
ei is such that kx e: = P sin(ei) = P sin(ei)  fhs cos(ei). We can thus see
that the descent of the brick (analogous to the stickslip mechanism discussed in
Section 2.2.1) is marked by successive pauses determined by
~xet =
(ei~l) 
fhs cos
(ei~l)'
(31)
since the friction force now opposes upward, rather than downward, displacements
of the brick. In other words, the angles at which the brick stops are different on the
return trip, even though the starting and ending points are the same. The behavior
is illustrated in Figure 33.
Some simple experiments confirm the overall validity of the model just described, although a particular behavior is often observed in the vicinity of e = n /2.
58
~
c
o
~
OJ
C
o
iIi
10
20
30
40
50
60
Tilt angle
80
70
80
90
FIGURE 33. Mechanical hysteresis effect. The successive elongations of the spring shown
in Figure 32(a) is plotted as a function of the tilt angle. The calculation assumes a static
friction coefficient fts of 0.3.
The mechanism of friction must, by definition, involve a zone of microscopic asperities operating in the plastic regime (see Section 2.2.1), even when the brick
is almost vertical. This requires that the force of contact be sufficient to break
through the very thin layer of oily residue that, barring special precautions, invariably coats the surface of most ordinary materials. This condition is rarely met.
As a result, the operating point often follows the smooth sinCe) curveat least
for a whileas the angle is decreased from Jr /2. Stated differently, the force of
friction appears to obey a certain threshold, corresponding to a minimum value of
the pressing force.
Other than this, the hysteresis effect discussed in the context of the simplified
model considered above can be generalized in the form of a recursion relation
between adjacent rest stages of the system
where Nand T are unit vectors, respectively, parallel and perpendicular to the
displacement of the object and E = I depending on the direction of the displacement. We know that fts = tanCes ), where es is known as Coulomb's angle. It is
the angle of inclination required for the first slip of the brick when the tilt of the
support is increased from horizontal.
It is easy to appreciate the subtle and ambiguous character of the forces involved
in a system at equilibrium by considering the following situation. Imagine that we
tilt the support by a given angle ei greater than tan 1 Clls) and start with the spring
in a quiescent state. What will the final position of the brick be when we place it
on its support? In light of the previous discussion, it should be obvious that the
result will depend on precisely how we conduct the experiment. In particular, do
we first place the brick on its support and then let the spring do what it will, or do
we first find the equilibrium position of the spring and then let the brick down?
3.1
59
60
FIGURE 34. Stress pattern observed in a twodimensional granular material under compression (after Dantu [51]).
Let p be the linear density of the chain, and let F A and F B be the two forces acting
on the extremities of an elemental segment dl, as shown in Figure 35. The figure
helps us understand that the arch will be stable if no torque or shearing force acts
3.1
61
i~
F.
J1~dl
p gdl
on any of the contact points between contiguous spheres. Some shearing force is
actually tolerable, as long as we remain in the dry friction regime, which requires
particular conditions that are left as an exercise. For simplicity, we simply seek
here the conditions that make these forces, as well as the torque, null. We do this
by writing that the resultant of all the forces involved must be tangent to the curve
at points A and B. Assuming that the segment dl is rigid, we may write
FA +FB + pgdl = O.
If is the angle the segment dl makes with the horizontal, the following relations
hold in a Cartesian coordinate system: dxjdl = cos(e), dyjdl = sinCe), and
dx 2 + d y 2 = d1 2 . The equilibrium equation projected on the horizontal axis yields
dx
dl
F  =Fh
(33)
'
(FdYdl )
(F dYdl )
_ pg dl
= 0,
d(ddxY) +pgFh  0 .
dl
(34)
62
We thus arrive at a differential equation describing the curve we are looking for.
It reads
pg dt
Fh dx
=  ;; [COSh ( ;~ x )
I}
(35)
As we might have guessed, this last result is similar to the equation describing a
flexible rope hanging loose between its two suspended ends. The requirement of
"flexibility" means in the present context that the particles the arch is made of may
roll on each other but are not allowed to slip laterally. In other words, (35) correctly
describes a granular arch ifand only ifthe constituent particles are subject to
rolling without gliding, which implies some restrictions specified in Section 2.2.2.
Figure 36 shows a magnificent example of architecture that owes its stability to the
principles we have just discussed. Essentially the same approach can be followed to
derive the equation corresponding to an arch supporting a load whose distribution
is known. We will treat the simple case of an evenly distributed weight.
Arch Supporting an Evenly Distributed Load
We now assume that the arch has a negligible linear density, but that it supports
an evenly distributed linear mass mo. In this case, we have mag dx = pg dt, and
(34) must be modified to
d y
dx 2
mag _ 0
F ,
h
whose solution is obviously an inverted parabola. The highest point of the parabola
has a horizontal tangent, from which we get
1 mag
y = zx
Fh
3.1
63
FIGURE 36. Photograph ofthe aqueduct ofMaintenon, built during the reign of Louis XlV.
It is a magnificent example of arches that have remained stable through several centuries.
64
3.1.2.
StressStrain Relations
more detailed discussion of this problem can be found in an article by S. Roux [24].
65
co
'
.2
OJ
v
(External Force)
when
V> Vt,
1=0,
when
V:S
VI,
3.1.3.
66
Rubber pouch
containing sand
+ colored liquid
3.1
67
FIGURE 39. Parallelogram used to model the unit cell of a twodimensional uniform
granular medium.
The lines connecting the centers of the four disks form a parallelogram that
changes shape when forces are applied in the manner shown in Figure 39. We
focus our attention on the closed geometrical figure composed of the four disks
themselves (assumed to be rigid) and the void between them. More particularly,
we are interested in how the surface area SI of this figure varies during the deformation. Let h v and hi be the lengths of the two diagonals of the parallelogram.
We can show that the surface area of the four disk sectors situated within the parallelogram is constant and equal to rr R 2 . It follows that SI has the following simple
expression
(36)
The diagonals hi and h v are restricted to values such that ht + h~ = l6R 2 . The
variation /::"SI of the surface area of interest can then be written in terms of one of
the two diagonals only, say hi
where hi can take on values only between 2R (when the left and right disks
touch each other) and 4R cos(rr /6) (when the top and bottom disks come in contact).
Figure 40 is a graph of /::"SI (normalized to 4R 2) plotted against the length hi
of the horizontal diagonal (normalized to 2R). The graph differs markedly from
what would be expected on the basis of the mechanical properties of homogeneous
and isotropic solids. As the horizontal deformation begins in response to a vertical
stress, the curve /::"St(h l ) starts out by increasing, until it reaches a maximum.
In this regime, the volume of the material goes up. This behavior agrees with
Reynolds's principle, and runs counter to what is observed with the usual homogeneous solids. We should point out, however, that the material does not necessarily
comply with the requirement that it be "strongly compacted," except to the extent
68
OJ
a:
""
1.0
~
ell
()
""
"
Reynolds's " "I
ell
CD
""
0.9
{g
regime
:::J
CfJ
I" "
I
I
0.8
1.0
1.2
Solid
regime
1.4
""
""
1.6
1.8
FIGURE 40. Total surface area of the object defined in the text (normalized to 4R 2 ) plotted
against the horizontal distance between spheres (normalized to 2R). The inclined straight
line describes the behavior of a traditional homogeneous twodimensional solid, whose
surface area always decreases when compressed.
that adjacent particles are already in contact. To the right of the maximum, on
the other hand, the volume does decrease, in accordance with the classical laws of
elasticity. The conclusion is that Reynolds's principle can, indeed must, depend
in a subtle way on the mode of stacking of the particles. We begin to appreciate
how ambiguous the statement "strongly compacted" really is. Finally, we have
ignored what might happen toward the left side of Figure 40. We can at least conceive that a "loose" granular medium should see its volume initially shrink as it is
being compressed. Figure 41 illustrates that some arrangements obey Reynolds's
principle, while others do not.
We may attempt to calculate effective deformation parameters by resorting to
the standard methodology used in the mechanics of solids. Since we have no idea
what the elastic stiffness of the object considered might be, except that it must be
(a)
16x16
Planar square lattice
(b)
16x16
Triangular lattice
FIGURE 41. Two possible lattice configurations in two dimensions. Stack (a) has minimum
compactness; its volume decreases upon application of any external stress. By contrast,
stack (b) exhibits maximum compactness; it does conform to Reynolds's dilatancy principle.
3.1
69
anisotropic, it is out of the question to try to define effective elastic constants. What
we can do, however, is to define the equivalent of Poisson's coefficient (]". In the
case of a homogeneous and isotropic material subjected to a vertical load, Poisson's
coefficient is defined as the ratio of the horizontal dilatation strain Uz = dhz/ h z to
the vertical contraction strain Uv = dh v / h v
Uz
Uv
dhz h v
dh v hi
(]"= =     .
When the two disks on the horizontal axis are in contact, we deal with a compact
triangular stack in two dimensions (which is equivalent to a compact hexagonal
stack in three dimensions). In this case, the effective Poisson coefficient is equal
to 3. This value is abnormally large when compared to the usual solids for which
thermodynamic stability considerations impose the restriction (]" :s 0.5. 4 The coefficient (]" decreases as the two disks lined up vertically get closer to each other.
When the parallelogram becomes a square, which conesponds to the maximum of
the curve in Figure 40, we have (]" = 1. This can be demonstrated by differentiating
(36), which gives
dS! = !(h zdh v
+ h v dh z).
!h zdhvCl  (]").
It is clear from this last equation that the variation in total surface area d S! changes
sign when (]" = 1. In other words, Reynolds's dilatancy principle ceases to be valid
as soon as the effective Poisson coefficient becomes smaller than 1.
Deformation of a Row of Parallelograms Placed Between Two Walls
Containers playa crucial role in the physics of granular materials, as we will observe many times in the remainder of this chapter. We might point out, incidentally,
that, although we did not spell it out, the model discussed above implicitly assumes
the presence of vertical walls to prevent particles from squirting out sideways. It
also requires dry friction to ensure the stability of the overall structure. 5 The walls
4This does not imply that the principle of thermodynamic stability does not apply here. The large value
of the coefficient (Y simply results from the local anisotropy of the material considered.
5 As
an exercise, the reader can work out the stability conditions of the structure by introducing a
coefficient of static friction between adjacent disks. It turns out that the structure will collapse when
the force applied on the topmost disk exceeds a critical value determined by Coulomb's angle of
friction.
70
must, of course, be deformable. They effectively play the role of Young's modulus by resisting lateral movements. In the interests of simplicity, and without
undermining the argument, we assume that the vertical walls undergo a uniform
horizontal displacement Uz = (5z E, where (5z is the stress exerted by the disks on
the walls and E is Young's modulus of the material the walls are made of.
Stability having been ensured and dry friction neglected (the disks are assumed
perfectly rigid and smooth), we are ready to focus on the problem we want to solve:
What is the stress exerted on the lateral walls of the container when a vertical stress
(5 v is applied to the system?6 For the time being, we consider a system made of only
three layers of a compact triangular stack of the type illustrated in Figure 41 (b). The
reason we choose to do so is that such a stack is made of a series of parallelograms
similar to the ones we have dealt with in our previous example. We will see later
how to solve the problem for a large number oflayers (see Section 3.1.4).
By itself, an individual parallelogram exerts no resistance to the stress applied to
the disks lined up vertically. 7 It merely redirects the displacements (or strain in the
language of continuous mechanics) toward the lateral walls. Since the system is
nondissipative, the principle of conservation of work applies and the stress exerted
on the lateral walls is immediately seen to be such that
(37)
=
.jCi(5v
K(5v.
(38)
3.1.4.
speaking, we ought to speak in terms of forces rather than stresses. But we have deliberately
opted to resort somewhat loosely to the traditional terminology in the mechanics of continuous media.
8We will often use the word "redirection" in this book. It means that part of the vertical stress applied
to a granular material creates a horizontal component that presses against the vertical walls. The word
"reorientation" is also sometimes used.
3.1
71
had a marked proclivity to redirect vertically applied stresses toward the sides, as
described above in the context of simple models. 9 The model was elaborated on a
few years later by Lord Rayleigh [16]. It rests on two principles that are important
to understand clearly:
From a mathematical vantage point, the medium is treated as though it were
continuous. This approximation is highly arguable in a granular medium.
Nevertheless, it makes it possible to write differential equations whose solutions describe the behavior of these materials with rather surprising
accuracy.
A vertical force Pv (or stress) applied to the material automatically generates
a proportional horizontal force Ph (or stress) such that Ph = K Pv.
This is entirely consistent with our previous analysis of a string of parallelograms.
Janssen and Rayleigh postulated that the behavior we found earlier could be extended to more complex structures than those we have considered so far. To
treat a granular medium as if it were continuous may strike some as too crude an
approximation. A few comments are in order.
Since the material is made of discrete constituents with a finite size, the
relevant space variables by definition cannot approach zero, and writing differential equations obviously raises serious questions. However, this type of
approximation has been used successfully in other branches of physics and
has led to solutions that for the most part seem to agree well with experiments.
A case in point is the transition from the mechanics of atomic structures to
that of continuous media. It should be kept in mind, however, that the approach works only as long as we deal with large numbers of particles. It
breaks down when we try to explore properties on a local level, such as when
the number of particles involved is small. We will see an example of this
limitation when we study the dynamic behavior of granulars, where a model
based on continuous media fails to account for the phenomenon of fragmentation in granular media that are vibrated or in forced flow. The message
here is that the approach applies only to structures made of a large number of
partic1es. lO
Problems related to the lack of cohesion of a dry granular medium are even
more fundamental. By definition, a granular material is heterogeneous and
contains empty spaces. Given the constraints we have specified early on in
this book, namely that the particles not interact with their gas environment, the
intergranular voids inevitably impose some restrictions on how deformations
9This property is not the exclusivity of granular materials. Soft media, such as lUbber, exhibit it
too. Unlike lUbber, however, the granulars we will deal with have no cohesion. We will see some
consequences of this fact a little later in this chapter.
10 A
practical method useful in simulation work will be presented in Section 6.1.3. It permits a
smooth transition from discrete variables (position, individual velocity, mass) to continuous variables
consistent with thermodynamic quantities (density, mean speed, temperature, etc.).
72
and stresses can propagate within the medium. A differential quantity such as
a variation in pressure becomes meaningless the moment the particles are not
in contact. Likewise, the concept of individual rotations presents difficulties.
Classical mechanics, as it applies to continuous and homogeneous solids, can
handle shear forces routinely but is powerless to deal with local rotations. We
may anticipate that a description based on the theory of continuous media
requires that the network of forces be transmitted throughout the medium
without interruption. A variable such as pressure can be defined in a granular
material only inasmuch as the contact chain is unbroken and rotations of
individual particles do not occur. We will make use on these remarks in what
follows, notably in our analysis of the fragmentation of granulars in forced
flow (see Section 3.2.4).
Problems related to the finite size and discontinuities of these materials can be
particularly bothersome when it comes to their dynamic properties, as we will
appreciate later on. Given the wide range of relaxation times in such systems,
a continuous theory may be useful for sufficiently longterm observations,
giving the system plenty of time to relax, but completely inadequate for
shortterm observations. Indeed, it is not all that unusual in the world of
physics for phenomena and their description to be quite different depending
on the duration of the observations. We will see examples of this in this
chapter.
Referring to Figure 42, we consider a sheet of thickness dh situated at a height h
in a cylinder of surface area A and perimeter P. Such a sheet is in equilibrium
under the combined effect of several forces.
Since pressure increases with depth (we adopt the convention that h = 0 at
the top of the cylinder, and h > 0 toward the bottom), the slice of interest
experiences a force directed toward the top and equal to A dpv'
The weight of a slice of thickness dh constitutes a force directed toward
the bottom and equal to pg dAh, where p is the volumetric density of the
material, assumed constant throughout the slice.
Surface Area A
Perimeter P
FIGURE 42. Cylindrical configuration defining the parameters used in Janssen's model.
3.1
73
The forces of friction with the walls, resulting from an infinitesimal movement of the slice toward the bottom, are directed upward. This is not an
arbitrary choice, as it amounts to assuming that the material slowly settles
under the action of gravity, a trend opposed by friction. 11 As we will see later,
there are some experimental situations that would justify making the opposite
choice for the direction of the forces of friction. In any case, the relevant force
exerts itself all around the wall, over an area P dh. Its value is thus equal
to fLs Ph P dh. Taking into account (38), which relates the horizontal and
vertical components of the stress, the forces of friction become KfLs Pv P dh.
We are now in a position to write the equilibrium condition for the particular slice
considered. It reads
Adpv
+ KfLsPpv dh
= pgA dh.
It can be rewritten as
Pv
+ pvoexp (KfLs~h).
(311)
74
... Hydrostatic
regime
Psat
1:
0>
'iii
Saturated
regime
When h becomes larger than about A/ P Kf.L" the vertical pressure Pv saturates,
as it approaches asymptotically a limit given by Pv + pg(A/ P Kf.Ls).
Denoting by X the argument appearing in the decreasing exponential in (311),
we have X = (Ph/ A)Kf.Ls, where Ph represents the outer vertical surface area of
the container, while A is its crosssectional area. Because it will reappear later on,
we define the ratio of these two areas as a parameter S, which we call the aspect
ratio (S = Ph/A). In terms of this new parameter S, the argument, which we
denote X, appearing in the exponential is simply X = S Kf.Ls. For reasons that will
become clear later on, we call this argument X the decompaction parameter. It
is a dimensionless number that completely characterizes the distribution of forces
in a cylindrical pile. To show that this is true, consider the mass m of granular
material contained within a depth h counted from the top; it is given by m = pAh.
The vertical force F v exerted on a layer at that particular depth h is equal to
Fv
_ mg
x
Pv A  (1  e ).
(312)
Through the combined effects of dovetailing and friction with the walls, the apparent weight of the cylindrical column is reduced by a factor that depends solely
on the dimensionless decompaction parameter X.
Specific Applications
2D Cell
75
3D Cell
TwoDimensional Container
A single layer of a granular is confined between two large flat frontal plates. The
granular medium has no friction with those plates, although it does with the small
end walls. We will make repeated use of such a twodimensional stack in the
remainder of this book. Let e be the thickness of the material, L the length of the
cell, and h its height (see Figure 44). The aspect ratio S of such a stack is given
by
Ph
2he
2h
S A  Le  L'
(313)
The numerator includes the active part of the perimeter only, ignoring that which
does not contribute to friction. The decompaction parameter reads X = SK!J,s =
2K!J,shjL, and (311) becomes
Pv(h)
= p gL  [ 1 2K!J,s
exp (  2K!J,s
  h )] .
L
76
(a)
(b)
N
spheres
FIGURE 45. Spheres excited by a vibrating plate. Diagram (a) is the classic problem of a
"bouncing ball." Diagram (b) corresponds to several spheres stacked vertically.
3.2.1.
Analyzing the behavior of a column of spheres is of interest for two reasons [37],
[55], [56]. First, the exercise reveals a serious difficulty in modeling collisions
when the particles involved coalesce into a cluster. Second, it shows the existence
of two distinct regimesone in which the particles remain stuck together and
move collectively as a solid bloc, as it were, and the other in which they encounter
each other only sporadically during brief collisions and spend most of their time
spread apart.
Before deriving equations and working out their solutions, it useful to keep in
mind some orders of magnitude.
12The importance of making a distinction between these two terms will become apparent as we go
further in our discussion. Fluidization of a dry granular gives it dynamic properties reminiscent of
those of a nonviscous liquid or gas. Decompaction is a phenomenon that gives it the ability to execute
internal movements of reorganization, for instance by convection.
3.2
77
(314)
where ZiQ and Via are measured immediately following the previous collision. The
collisions are described by linear equations of the form
(315)
where Ui and Vi are the velocities of bead i immediately before and immediately
after a collision measured in the frame of reference attached to the center of mass
of the two colliding objects. By convention, Ua and Va refer to the vibrating plate.
The beads are numbered sequentially from bottom to top.
How to implement such a calculation constitutes an important physics problem.
It involves making some decisions on the sequence of eventsin this case ballistic flights and collisions between the N particles and the vibrating plate. Without
going into the details of the computer simulation, which will be covered in Chapter 6, we will simply mention the simplest and most natural algorithm. It belongs in
the class of socalled "eventdriven" algorithms, which analyze events sequentially
78
10 ..:::,.....y,...,,
0.00
",
.........
..........
0.02
0.04
Time (s)
FIGURE 46. Dynamics of a collection often spheres, calculated with the help of (314) and
(315), using 8 p = 0.6 and 8 = 1. The upper diagram shows the positions of the spheres
plotted against time. The lower diagram gives the time intervals between successive collisions (after [59]).
in the order in which they occur. 13 For instance, starting at some time t, we can
construct a matrix T which contains a sequence of times ti at which future events
are predicted to take place. In particular, the smallest element tim of that matrix is
the time of the very next event. When it has occurred and its consequences have
been calculated, a new matrix T is constructed, and the cycle is started all over
again. There is absolutely nothing wrong with this procedure as long as we are
not faced with two or more simultaneous events (in the sense of physics or of the
computer's ability to distinguish them), in which case some inherently arbitrary
choices have to be made. Erroneous decisions can, of course, seriously taint the
subsequent chain of events. More pragmatically, the reader may recall a calculation we did in Section 2.2.2, which convinced us that collisions take some finite
amount of time (of the order of a microsecond in the example treated), and that
this time further depends on the velocity of the colliding objects. Including this
effect leads to additional uncertainties.
Summarizing, the calculation technique based on Newton's equation for ballistic flights (314) and eventdriven collision matrices (315) can run into major
problems when the sequence of events becomes difficult to keep track of. The point
is illustrated in Figure 46, which tells the story in the case of ten beads. The figure
shows the trajectories Zi (t) over the duration of one excitation period, as well as the
time intervals separating sequential events. We notice immediatelyand that is a
very general resultthat the time interval between consecutive collisions becomes
exceedingly small (less than 107 s in the present case) when the ten beads get so
tightly clustered as to be practically in contact. From a mathematical point of view,
we may even claim that contact will be achieved when the interval between events
13 "Eventdriven"
algorithms are to be distinguished from "timedriven" algorithms and other sequential techniques. They are also sometimes referred to under the heading of "collision method."
79
goes to zero, although that may not be too realistic because even the most simplistic
models indicate that two colliding objects remain in contact for durations as long
as several tens of time intervals in this limiting regime. Hence a serious problem.
It is clear that the difficulty has to do with the existence of two separate characteristic times of the systemthe duration fl Tc of a collision and the interval flt m ,ml
between two successive events. The validity of the model is assured only to the
degree that fl Tc fltm ,ml. Some might further object that in this troublesome
regime the distance between particles becomes comparable to the deformations or
microasperities of the surfaces, which in itself may invalidate (315). This complication, which is a constant source of vexation in numerical simulations, has come
to be known in the literature as inelastic collapse or inelastic catastrophe [57][59].14
Numerical solutions of Newton's equations, obtained with the precautions discussed above, reveal several specific and generic behaviors that seem extendible
to two and threedimensional configurations. We proceed next to describe these
behaviors and to emphasize their differences, although the dividing line between
them is not always clearcut.
14Needless to say, it is a catastrophe only from the point of view of developing a model. As we might
guess, the term was coined by people doing computer simulations.
15This is a common technique designed to get around problems of accumulation in numerical simulations. We will encounter it again in Chapter 6.
16The LRV criterion will be discussed in detail in Section 6.2.2.
80
FIGURE 47. Behavior of five spheres initially grouped in two blocks of 3 and 2, respectively. The horizontal axis represents time. The spheres rearrange themselves after the
collision. The simulation was done by applying the LRV criterion (see text) (after [59]).
When the coefficient of elastic restitution is high, as in the case of steel balls for
which s = 0.9, the key parameter is the normalized acceleration r. As shown in
Figure 48, a colunm of ten balls submitted to a strong acceleration is virtually in
a fluid state. By contrast, when the acceleration is lowered, the column remains
clustered together in the form of a single block that moves collectively in phase
with the vibrating plate. The system is then said to be "locked" to the excitation.
This behavior turns out to be quite stable over a wide range of excitation parameters
(acceleration r and frequency f). It can be understood, at least intuitively, with
the help of the following simple argument. Consider the ballistic trajectory of
a group of spheres moving as one. Since the spheres are tightly clustered, most
of the energy they receive from the vibration plate is dissipated in multiple and
repeated collisions. Under such conditions, the coefficient of elastic restitution of
the column is quite small. I? The problem then becomes identical to that of a single
inelastic ball, as depicted in Figure 45(a). The advantage of the much simpler case
is that it has been solved completely. One of its wellknown properties is to evolve
toward chaos (Feigenbaum's scheme). Without dwelling on the details, Figure 49
enables us to understand the mechanisms involved.
The figure illustrates the trajectory of an inelastic object (simulating the column we were dealing with) placed on a vibrating plate whose vertical motion
is sinusoidal with amplitude A and frequency f. The object takes off when the
acceleration imparted by the plate is such that Au/ > g. It then leaves the plate
and follows a parabolic ballistic trajectory, falling back toward the plate which in
the meantime keeps on oscillating. It is crucial to realize that, since the impact is
nearly inelastic, the object does not bounce back. Rather, it "sticks" to the oscillating plate until it is once again launched up into the air. We note that inelasticity
will also tend to cause a colunm of several beads to clump together, ensuring the
stability of the system.
17 A
collection of clustered particles presents a small global coefficient of elastic restitution even if each
individual member is nearly perfectly elastic. That is why a packed bag of marbles does not bounce
back when falling on a hard floor, even though a single marble does so spectacularly.
81
(a)
(b)
FIGURE 48. Calculated trajectories of ten spheres with 8 p = 1 and 8 = 0.9. The system
is vibrated at 10 Hertz with two different accelerations ['. In (a), r = 8.0, and the system
is "fluidized." In (b), [' = 1.7, and the system is "condensed."
The following simple calculation is informative. Let t* be the time when the
normalized acceleration r becomes equal to 1. It is given by
At that moment, the object is launched up with an initial velocity v* equal to that
of the vibrating plate. All other parameters of the movement are straightforward
to calculate. We might note that, even if the coefficient of elastic restitution is not
strictly zero, the locking on the second harmonic (see Figure 49(b)) is particularly
stable. That is because the motions of the object and of the vibrating plate then
take place in the same direction, minimizing the velocity differences and reducing
(a)
(b)
FIGURE 49. Locking of the trajectory of a column of particles on the vibration of the plate.
Locking can occur on a number of frequencies, including (a) the fundamental, and (b) the
second harmonic.
82
the likelihood of rebound. 18 The same situation can exist, of course, at other
frequencies besides the second harmonic, including the fundamental.
Returning to the case of a column of beads, the considerations developed above
provide a plausible picture as to why a largeamplitude excitation can lead to
fluidization, as depicted in Figure 48(a). Even if we start with a situation in which
all the beads are clustered and form effectively a single block, enough energy can be
transmitted along the entire string to cause even the topmost bead to separate from
the block and initiate its own individual ballistic trajectory. 19 What can happen to
the topmost bead can a fortiori happen to any other bead in the stack as well.
Summarizing, a column of beads or other elastic granular material vibrated
vertically exhibits two distinct regimes depending on the acceleration imparted by
the bottom plate. For small accelerations, the system is in a "condensed" state,
with the beads practically in contact with each other and moving in unison. For
large accelerations, the system is in a "fluidized" state where the beads move
about individually much like particles in a gas or a fluid. In this latter state, the
acceleration Aw 2 is no longer the pertinent variable. Instead, that role is taken
over by the kinetic energy, proportional to A 2w 2 , or, equivalently, by the square
of the mean vibration velocity. The transition between the two regimes involves
portions of the system that are fluidized and others that are still condensed. All
of this has been satisfactorily simulated numerically and verified experimentally
[56], [60].
The existence of these two regimes and the transition from one to the other also
depend on the value of the coefficient of elastic restitution as well as on the number
of particles involved. This point is examined further in the next development.
Fluidization and Condensation as Functions ofHeight
83
hardened steel. We then find that the critical number of balls is 8 for aluminum
and 39 for steel. With very few exceptions, the granular materials typically handled by the food or pharmaceutical industries have much a lower coefficient E;.
Therefore, in the vast majority of practical situations, the variable X is such that
X 3, and the materials are most often in a condensed regime.
Case X :s 3: As noted above, there is a possible transition from a condensed
state prevailing at weak accelerations to a fluidized state when the acceleration
increases.
Case X ~ 3: Here we have a sufficiently tall column of material in which the
energy dissipation by collisions is rather substantial. The column acts as a
single block and the duration of contacts between particles can become quite
longat any rate much longer than the vibration period. It is important to
realize that, strictly speaking, we are not dealing with a condensed state in
quite the sense defined previously, since the number of collisions between
particles approaches zero. What we do have is a truly compacted column
whose dynamic properties approximate closely that of a single inelastic ball.
Indeed, that was confirmed by simulations and experiments. Figure 50 shows
bifurcation diagrams for a set of ten aluminum balls with increasing accelerations. Note the agreement between experimental results and calculations
done with a coefficient of elastic restitution of 0.6.
The bifurcation diagram in Figure 50 shows a classic tendency toward chaos. In
particular, for certain accelerations, the system "hesitates" between two possible
states. Experimentally, we generally find that the system oscillates back and forth
between two such states. That is a telltale signature of the onset of chaos.
ob
:<
.J
t>
t>
.....
.....
o
o
Eo<
<
g~
t>
t>
t>
10
FIGURE 50. Bifurcation diagram of the time of occurrence of collisions. The product f x
Teall is plotted against the normalized acceleration ['. In the present case, N = 10, f = 30
Hertz, and sp = s = 0.6. Circles are experimental points, while triangles show the results
of numerical simulations. The diagram clearly shows locking on the fundamental, as well
as on the second and third harmonics (after [56]).
84
3.2.2.
85
Pz
15
z+
I
I
I_~~
FIGURE 51. Statistics of the trajectories of elastic and frictionless beads at the surface
of a twodimensional pile. The photograph reveals that the lower part of the pile remains
compacted. The histograms show the distribution of velocities along the two principal axes
(after [61]).
86
Faraday instabilities in liquids and are currently the focus of active research. We
will briefly discuss them later in this chapter (see Section 3.2.5).
The experimental apparatus described above allows us to study the nature of
the fluidized top few layers of pellets which, as shown in Figure 51, seem to
follow complex ballistic trajectories above the compacted reservoir. This can be
done by visually recording the pellets with a CCD camera over a duration of
many thousands of excitation periods. A particularly useful technique is to use a
stroboscopic lighting system flashing pulses of light at regularly spaced instants
synchronized with multiples of the excitation period. The technique lends itself
to direct measurements of the instantaneous velocities of the fluidized particles.
Subsequent data manipulation generates the statistics of the velocities at different
times during one period. The results can then be plotted in the form of histograms.
Figure 51 includes such histograms, showing the probability distribution functions
of the horizontal and vertical components of the velocities centered about their
mean values (v x ) and (v z ). The two histograms have approximately the same
width, indicating that the medium behaves isotropically, in conformance with the
symmetry of the problem. As it turns out, this is one of the characteristics of a
fluid or gas in thermal equilibrium.
Some Comments on Scaling
87
3.2.3.
Once again we resort to the same notation we used in our discussion of Janssen's
model, particularly when we derived (39), in which A represents the surface area
of the cylinder's base and P its perimeter. However, for reasons of convenience,
we adopt a different convention for the heights h. From here on, they will be
referenced to the base (h = 0) and counted positively toward the top. The geometry
is illustrated in Figure 52.
88
ho
A
Decompacted
dm
p
Compacted
Ih
tr
Ih'
(316)
89
Equation (316) enables us to define a threshold height h t below which the material is not allowed to separate from the wall and must therefore remain compacted,
while it will undergo decompaction above hi. The equation defining hi can be
recast in the form
pg_A_[l_ exp
PKfLs
(KfLs~(hl A
h o))],
where h o is the total height of the stack. Combining the last two equations yields
an expression giving hi
(317)
If we introduce once again a parameter So = P hoi A as we did in Section 3.1.4,
we may define what we call the rate of decompaction a as a = h t I h o. It is given
in terms of the normalized acceleration r by
hi
In(2  r)
a==l+
ho
SoKfLs
=1+
In(2  r)
X
(318)
The rate of decompaction a turns out to depend only on the dimensionless parameter X (aside from the normalized acceleration r). This is the reason why we
chose earlier to give X the name decompaction parameter.
90
1.0
0.8
0
c:
o
0f$
<1\
0
E
o
21 The
91
This technique makes it possible to obtain snapshots such as the one reproduced
in Section 3.2.2 [61]. It requires a stroboscopic flash that can be synchronized
with some characteristic time of the system, for instance, the excitation period of a
vibrating plate. The duration b. T of the flash is adjusted in order for the particles
to leave a visible trace whose length is proportional to their speed at the time of
the illumination. b.T must be sufficiently short to avoid random collision events
which would interfere with the interpretation of the record. Images are stored
separately and processed with statistical packages designed to compute the mean
values of the relevant quantities. It is not uncommon to be dealing with a few
thousand images, which requires automated data processing. We might note here
that this technique applies to shortduration phenomena (short compared to the
excitation period). We will see several applications in what follows.
Measurements of the Relative Motion ofParticles
92
IImageKh
1
tHE
1
This table tells us that the resulting display would show not only the points that have
not moved between the two shots, but also the traces of those that did move. That
is how the images of convection reproduced in a later paragraph were produced.
With an "exclusive or" (or) operation, the image would show only the traces of
the particles that had moved, which may also be of interest.
The sequence of operations just described must obviously be repeated over
a sufficiently long time for something interesting to show up. In other words,
the technique is suitable to examine longerterm phenomena (compared to the
excitation period). This technique, known as "Computer Posed Photograph," or
CPP for short, was used to produce the images shown in Figure 54.
We now describe the experiment that was used to generate Figure 54. Roughly
50 x 50 aluminum beads are stacked in the form of a regular triangular lattice inside
a twodimensional cell characterized by a shape factor So = 2h o/ L. The outer
surface of the beads is roughened by means of an appropriate surface treatment.
This is accomplished by shaking the beads for some time in air. The work hardening
resulting from the many collisions increases the coefficient offriction between balls
from an initial value of 0.2 to about 0.6. We recall that the coefficient of elastic
restitution s of aluminum balls is also about 0.6. We showed in Section 3.2.2 that
the reduced variable describing the dissipation of the collisional energy in such a
configuration is given by X z = Nz(ls). Under the conditions of this experiment,
we are therefore definitely in a condensed regime since a column six balls high is
FIGURE 54. Computer posed photographs of a twodimensional vibrated stacle The photo
on the right is a magnification of the region marked at the left by a circle. Convection rolls
generated by shearing at the wall are clearly evident (after [67]).
3.2
93
94
(a)
(b)
r
>1
Vibrations
FIGURE 55. (a) Photograph of a "Chinese hat" in the process of developing. (b) A simple
experiment demonstrating the crucial role played by lateral walls in pile formation.
95
This simple twodimensional experiment demonstrates conclusively that convection processes in the vicinity of the lateral walls are directly responsible for pile
formation.
Similar experiments using threedimensional geometries confirm the key role
played by the walls of the container. In one set of experiments, illustrated in
Figure 56, two types of granules were usedsome transparent and some painted
black [20]. As long as the stack is not too thick, it is possible to visually track
the path of the black granules while the cylindrical container is given a series of
vertical impulses. Figure 56 shows that the particles situated near a roughened
wall creep down along that wall. No convection is observed if the wall is smooth,
in agreement with our previous discussion.
All these observations support the contention that walls do indeed playa determining role in the mechanism of pile formation. The Janssen model also implies
that the parameter governing the onset of decompactionand, by extension, pile
formation via convectionis the acceleration r. This has been abundantly confirmed experimentally both in two and in three dimensions.
96
(0)
FIGURE 56. Threedimensional convection experiments. In configuration (a), the left wall
is polished smooth. Only the right wall, which has been roughened to intensify friction,
induces convection movements. In configuration (b), convection takes place in a direction
opposite to that which is observed in a cylindrical container (after [20]).
.........
..
Q)
Bead diameter
...
0
"'0
::l
c..
'"
O'l
Co
1
'"
0.2mm
0.4 mm
1.0 mm
C\a
97
oOe
\
0
8
,0
2
1.0
1.2
1.4
1.6
1.8
2.0
2.2
log(m/2n)
FIGURE 57. Loglog plot of the vibration's amplitude threshold for threedimensional pile
formation against frequency, for several bead sizes. The diagram shows that the acceleration
is indeed the determining factor, since the straight line has a slope of 2 (after [66]).
98
en
"0
ro
25
Q)
.0
.....0
20
= 1.39
en
15
ro
Q)
0
.....0
c:
0
:;:::;
10
5
'iii
0
a..
0
10
10
10
10
10
Time (seconds)
FIGURE 58. Separation between each of the two mounds and the nearest wall plotted
against time for two accelerations. The experiment covers about 30 hours. Distances are
expressed in numbers of particles (after [67]).
The variable x is a measure of the distance from the summit of each of the two
mounds to the wall. Given the geometry of the system, it is also proportional to the
depth of the Chinese hat. In the above equation, the coefficient a, which describes
the efficiency of the pile formation process, turns out to depend on the acceleration
r through the quantity r  1, and the coupling via friction between beads and
walls. 26 The coefficient C which figures in the exponential defines a characteristic
length in the experiment; it can be assimilated with the typical dimension of a
convection roll, involving the coupling between beads. In the absence of any solid
theoretical basis, it would be hazardous to venture any further interpretation of
what are after all purely phenomenological parameters. We simply conclude by
listing the experimental values of the parameters a and C, in units of number of
beads, for two values of the acceleration r.
r
a
C
1.15
3.9
1.0
1.39
16.8
2.7
~::~I
1.0
0.0
0.2
004
0.6
0.6
1.0
t$ 0.8
'.
(e)
"
O
hh'
:
.
.
.
Decompacted
0 .6
I 10
Compacted
0.0 I_"~o1.0
1.2
104
1.2
a=
004
I.L 0.2
99
So
hl/h o
=0.67
__1
1.6
So
0.6
1.0
(b)
EOA
.c
<1
0.2
0.0
~..
a
~
'Kf=011
h (cm)
t$
0.8
L..
0.6
t5
004
CIl
'"
'.
Decompacted
So= 1.26
Compacted
I.L 0.2
0.0
(d)
I
1.0
1.2
Ie
104
10
1.6
FIGURE 59. Experimental results confirming the gradual decompaction model (after [67]).
See text.
100
(a)
(b)
(c)
Pile Formation
No Boundaries
Dual Lattice
K>O
!l>0
K>O
!l::0
1<::0
!l >0
FIGURE 60. Synopsis of results consistent with the gradual decompaction model. Three
configurations correspond to different combinations of the parameters K and /L (after [97]).
we will see later that this behavior holds only in cells of relatively small
lateral dimensions. Section 3.2.5 will describe what happens in larger cells. As
predicted by the model, the acceleration needed for collective liftoff of a stack
is indeed a monotonically increasing function of its height. The data points fall
on a theoretical curve with a single adjustable parameter Kf = 0.11. These
results were obtained with a different cell than the one used in the previous
experiment.
Figure 59(b) refers to a completely different experiment, although, this time,
it is done in the same cell as above. Stimulated by the vibrations, the pellets
can move relative to the cell. Provided that the particles leave a record of their
movements on the front and back windows of the cell, we can then calculate
the amplitudes of the displacements /')"h(h) as a function of height h. Here
again, there is excellent agreement between theory and experiment.
Figure 60 brings together a series of contrastenhanced images of granular stacks
viewed through the transparent windows of various twodimensional cells. It is
a fitting conclusion to this paragraph devoted to experimental verifications of the
model in that it summarizes all the results obtained thus far.
Inasmuch as the model identifies the quantity X = 50 K/h as the key parameter,
we have, for one thing, been able to verify the effect of the aspect ratio 50 when K
and /h are both nonzero (Figure 60(a)). The twodimensional cylindrical container
enabled us to explore the case K > 0 and /h = 0 (Figure 60(b)).
Finally, the case K = 0 and /L > 0, depicted in Figure 60(c), requires a trick
based on the discussion in Section 3.1.3. In that section, we analyzed how the
surface occupied by a horizontal row of parallelograms changes as a function of
a vertically applied stress. We pointed out that the dilatancy properties of such
an object vanish when the disks lined up vertically come into contact. If we now
consider a twodimensional stack of such parallelograms in the same configuration,
it is clear that the vertical lattice, named dual lattice, must have a coefficient K
such that K = O. This configuration is illustrated in Figure 61. As it turns out,
such stacks exhibit sufficient stability to withstand moderate accelerations r in
Triangular Lattice
K>O
101
Dual Lattice
K=O
FIGURE 61. Stacking in a triangular lattice and its dual counterpart, rotated by 90. The
configuration on the left is the only one that redirects stresses toward the walls.
the range of 1 to 1.5, which are typical of the experiments conducted previously
with conventional triangular lattices. Unlike the conventional case, though, and
in full agreement with the model, there is here neither gradual decompaction, nor
convection, nor pile formation. In other words, the meshed character of such
structures is indeed one of their fundamental properties explaining not only their
dilatancy properties but their resistance to decompaction under vibrations.
Short Term Decompaction: Fragmentation
Although the model described in the previous paragraphs is based on the mechanics
of continuous media, it accounts fairly satisfactorily for the gradual decompaction
properties of granular stacks subjected to vibrations. Before proceeding any further, it is essential to appreciate the deeper implications of this theory. The model
rests fundamentally on Janssen's hypotheses, supplemented by the dynamical law
expressed in (316), which was arrived at empirically. Our purpose is not to reexamine here the derivation of Janssen's equation (see Section 3.2.3), but we do need
to take time out to reflect on the nature of the approximations underlying (316),
which is rewritten here for convenience:
fg dm  g dm ::: dFtric!.
We will focus on two key points: (1) the assumption that the system can be decomposed in uniform horizontal sheets; and (2) the simplified description of dry
friction .
To begin with, we recall that the model assumes that the material is made
of horizontal sheets of mass dm, rubbing against the walls where a frictional force d Ffrie! is developed. This force can be calculated with the help
of Janssen's hypothesis. Furthermore, (316) implies that the acceleration
fg imparted by the vibrating plate is homogeneous throughout any horizontal slice of material. Assuming that acceleration, massand, therefore,
frictionare all uniform in a given sheet is manifestly in flagrant contradiction with our earlier description of the network of contact points within a
granular material (see Section 1.2.2). The reader will recall that we invoked
102
arches anchored onto the lateral walls to describe the equilibrium forces in
such materials. The notion of arch is an essential and unavoidable component
of the static and dynamic properties of granulars. From this vantage point
alone, we can expect substantial deviations between theory and reality.
The extremely simplified way the friction forces with the walls are treated
may have even more serious consequences. Without going into details that
will be discussed in greater depth in Section 6.4, we merely point out here that
(316) expresses nothing more than a simple rupture of the forces of contact at
the boundaries when a sheet experiences a sufficient force. This completely
neglects the indeterminacy of forces which was discussed in Section 3.1.1,
and assumes that these forces are fully mobilized and directed straight toward
the top, in accordance with Janssen's hypotheses at equilibrium. It is highly
likely that the random nature of the contact chains, for one thing, and the
more or less elastic coupling between particles involved in arch formation,
for another, will in fact lead to a rather large range of contact forces at the
walls. Invoking elementary uniform horizontal sheets to model a medium that
is inherently inhomogeneous and partly undetermined may seem fraught with
danger. Furthermore, (316) also ignores the coefficient of dynamic friction
and, perhaps more importantly, the velocity dependence of the friction force.
All these considerations may induce the skeptics to think that the agreement between theory and experiment is purely fortuitous. That conclusion would be
wrong, though, at least as long as the model is used with proper caution and within
its own limitations. It is appropriate to point out, for instance, that the simulation of gradual decompaction merely establishes a phase diagram that predicts
the average height at which a stack undergoes decompaction, without specifying anything further about what might be going on in the decompacted phase.
Except for the calculation of particle trajectories, which requires an additional
assumption to be discussed shortly, the model does not go beyond working out
the height of the compacted phase. The procedure usually takes into account the
irregular character of Coulomb's friction law, as we will discuss in more detail in
Section 6.4.
The experiments we have described thus far essentially deal with longduration
observations. In actuality, the decompaction processes observed experimentally
proceed by way of successive relaxation states in which a system remains long
enough (typically several seconds) to be recorded by conventional techniques, such
as CPP photography. This approach misses the details of any fast event (lasting
no more than a fraction of the excitation period) involving shortduration and reversible distortions. While that much is clear from an experimental point of view,
what consequences does it entail in terms of the simplified model we have developed? In what fundamental way is itin spite or because of its imperfectionsmore suitable to account for longduration phenomena? There is no clear answer
to this question. All we can do at this point is to simply acknowledge that the model
does seem to describe fairly accurately systems in their reorganized states, after
they have relaxed back to equilibrium. It does not address the largely indeterminate
103
Fracture
Lines
FIGURE 62. Shortterm observations, using a strobe light synchronized with the vibration
period, reveal that decompaction of a vibrated twodimensional pile actually results from
a series of fragmentations propagating from top to bottom during ballistic flights. Such
fragmentations occur during each cycle of the excitation (after [69]).
solutions, subject to random fluctuations, that would result from a more complete
analysis of the problem.
Given the fundamentally dual natureshortduration and longdurationof
the relevant phenomena, it is naturally tempting to observe the time evolution of a
granular stack by imposing a slight shift in the frequency of the stroboscopic light
flashes relative to that of the excitation. Figure 62 shows the results of just such
an experiment. 27
The image contains a lot of information about the shortterm behavior of a
granular structure, which deserves a number of comments. 28 First of alland
that cannot be conveyed by a frozen snapshotthere are fractures, such as the
one indicated in the figure, that have dynamic properties of their own. A video
would show that these fractures are initiated near the top and propagate toward the
bottom. This mechanism of successive fragmentation occurs during each cycle
of the excitation and persists during the ballistic phase of the stack, which can be
quite short compared to the excitation period (typically in the ratio of 10 ms to
0.2 s). These fractures have a characteristic Vshaped form (in the present case
pointing toward the bottom) and are more or less dislocated as they progress within
27 This
discussion does not respect the actual chronological sequence of events that led to this experiment. In reality, it was while trying to measure directly with a video camera the movements of
particles in a granular stack that researchers came to realize that gradual decompaction resulted from
a series of more or less irreversible fragmentations.
28Note that CPP images and traces left on frontal windows can only provide information averaged over
large numbers of periods.
104
the structure. We shall see in the next section how we can extract more detailed
information about this phenomenon, which appears to be a very general mode of
decompaction of a guided stack in ballistic flight.
3.2.4.
TwoDimensional Experiment
Particles are stacked in a regular triangular arrangement in a vertical cell of the
type described previously. The bottom of the cell is blocked by a springloaded
metal plate that can be dropped rapidly (with an initial acceleration of 3g). A
CCD camera records the fall of the stack, which lasts approximately onetenth of
a second. Figure 63 shows a timelapse sequence obtained during the experiment.
As expected in light of the results presented in the preceding paragraph, the stack
breaks up into separate Vshaped blocks whose apex now points upward, always
in a direction opposite to gravity. 3D Also as noted before, successive fractures
propagate toward the top of the pile during the fall. All these observations are fully
consistent with those described in connection with a vibrated stack, provided that
the sign of the acceleration r imparted to the structure in ballistic flight be reversed.
While it was directed toward the top in the case of a vibrated cell (upward ballistic
flight prior to recompaction), it now points down. Given our previous discussion
of the formation arches (see Section 3.1.1), the Vshaped structures observed in the
present experiment can be understood as contact chains with enhanced strength
supporting the granular mass above them.
29Drake reported experiments involving inclined falls in which he was studying structural modifications by direct observation, although he did not draw the conclusions derived here concerning the
fragmentation process [68].
observations apply to threedimensional structures. If a small tube about I cm in diameter is
half filled with compacted sand and the bottom is suddenly opened, the content escapes by fragmenting
from the bottom up into a series of blocks separated by arches analogous to those shown in Figure 5.
30 Similar
105
Theoretical Modeling
An interesting question is whether our previous theoretical model, which is a dynamic extension ofJanssen's model, can account for the process offragmentation. 31
We will later show how a properly adapted numerical simulation can indeed reproduce the appearance and propagation of fractures within granular piles [69].
Fall Without Fragmentation
The first question that comes to mind has to do with the fall of a simple block
assumed, for the time being, to be free of fractures. This problem was actually
solved in Section 3.2.3, and we use the same notation here, including the convention
that heights be counted positively from the bottom up. The heightdependence of
the vertical stress was given by
PvCh)
pg_A_[l_ exp
PKfLs
(KfLs~(h
A
)].
ho
106
where, for convenience, the characteristic coefficients are lumped into a single
parameter C; = Lj(2Kfls).
A straightforward solution can be obtained, considering the stack as a whole
whose apparent weight has been given by equation 3.12. In this case and supposing
that the pressure and friction forces remain constant during the downfall process
of a pile whose height is h o, we get the reduced acceleration of the stack as
where X = .
!;
An alternative approach would first consist in supposing that the pressure equilibrium has been established before opening the springloaded metal plate. Second,
and because of the particular discontinuous nature of the material, we may consider a thin slab of granulate and suppose that the previously mobilized friction
forces are still acting at the wall thereby reducing the downwards acceleration of
this particular piece of material. We know that the friction force developed at the
walls at this particular thin slab of granulate can be written as
dFfrict
 = Pv = g [ 1 dm
pC;
hO)]
exp (h
 C;
= gf(h) = g 
dFfrict
dm
,
where
[0, hoJ.
Strictly speaking, and in the context of the hypotheses discussed earlier, this equation is valid only at the precise instant when the downward fall is initiated. In the
interest of simplification, and for a semiquantitative analysis, we shall assume that
it holds for the entire duration of the fall. We will see later on that experiments
justify this approximation.
We begin by noting that r(h) is an increasing function of the height h. This
implies that a stack subject to its own weight and to the applicable forces of friction
will tend to remain compacted as it falls. In other words, the stack experiences no
spontaneous tendency to fracture, unless external causes come into play.
To identify external causes likely to initiate fractures that can propagate and
lead to fragmentation, we need to modify the surface properties of the lateral
walls. When we do so, we observe the following:
400
.s..9!
"6.
Fractures
One particle
ho=20 mm
ho
300
107
ho=44 mm
Q)
:S
'0 200
Cl.
ho==' 93mm
h o= 130mm
Q)
:S
'0 100
1:
ho==' 195mm
0>
"iii
J:
0
0
50
100
150
200
250
Time (ms)
FIGURE 64. Altitude at the top of the pile as a function of fall time. The filled symbols are
experimental data. The solid curves correspond to parabolas calculated with KfLs = 0.12.
The arrows indicate the appearance of at least one fracture in the pile (after [69]).
Surfaces prepared with a specular polish, that is to say, with a roughness of the
order of a fraction of a micron, have a coefficient of friction fLs that is virtually
the same as for regular surfaces after they have gone through normal wearout. With polished lateral surfaces, stacks can fall without fragmentation,
although with a reduced acceleration compared to that due to gravity alone, in
accordance with (320). These conclusions emerge from experiments whose
results are shown in Figure 64. They were obtained with identical oxidized
aluminum pellets placed in the same cell, but stacked to different heights.
The data indicate that within experimental errors, and as long as no fracture
appears, the stack falls with a normalized acceleration that is governed by the
initial height and is adequately described by our equation for r(h). The solid
curves in Figure 64 were calculated using a single adjustable parameter KfLs
of 0.12, which is somewhat smaller than the value determined previously. The
discrepancy is probably due to the fact that what matters here is the dynamic,
rather than static, coefficient of friction.
Lateral surfaces with a roughness of a few microns or more trigger the apparently random appearance of a series of ascending fractures, as noted previously.
Where Do Fractures Initially Appear?
It is extremely difficult to accurately model interactions between a stack and the
surface of the lateral walls. Accordingly, we will restrict ourselves here to the
issue of the stability of a fracture initiated at one of the lateral walls by a random
fluctuation of the topography, or of the micromechanical parameters (which can
disrupt the regularity of a lattice), or even by a loss of contact with the wall or
any other instability apt to promote fracturing. Let us then assume that a fracture
108
appears at height h f measured, as usual, from the base of the pile. The fracture
causes an interruption in the contact chain and a reorganization (assumed to be
instantaneous) of the equilibrium forces in what is now two daughter pilespile
A on top and pile B at the bottom. Each daughter pile is from now on subjected
to its own acceleration r A and r B .
Since r is a monotonic function of height, it is immediately apparent that a
fracture will remain stable or can even get amplified as long as r A ~ r B, which
implies the condition h f .:::; h al2. Stated in words, the fracture must be initiated
in the lower half of the stack if it is not to close up on itself during the fall.
We next examine how an ab initio simulation can reproduce many of the results
we have just discussed.
Numerical Simulation of a Stack in Guided Fall
We have learned in Section 2.2.2 how to write down useful equations for simulating numerically the dynamics of stacks. Still more techniques will be presented
in Chapter 6. Until then, we further our goal to understand the physics of fragmentation of granular stacks in guided falls by examining a particular numerical
technique which provides additional insight into a number of experimental observations discussed above.
The method we are about to use is rooted in the concept of rigid spheres (in a
sense to be specified in Chapter 6). It is an "eventdriven" approach quite consistent
with the spirit of that which we introduced to treat the problem of a onedimensional
column in Section 3.2.1. As we stressed then, the technique is fundamentally dynamic in nature and is not really suitable to account for static situations such as the
initial state of a stack. Its implementation presupposes the use of an artificial "thermal agitation" enabling the particles to sense, as it were, their immediate surroundings, even in a static condition. 32 In this picture, equilibria seem themselves described in terms of entities like impulses, linear and angular momenta, all of which
are related to translations and rotations. A real equilibrium or quasiequilibrium
situation contains inherent uncertainties that may obscure the connection between
experimental reality and numerical simulations. As far as experiments are concerned, random irregularities built in the surface of the lateral walls often set off
fractures. By the same token, a thermal agitation artificially inserted in a numerical
model constitutes from the outset a random phenomenon capable of initiating fractures that are independent ofthe quality ofthe walls. That being understood, numerical simulations turn out to reproduce satisfactorily many aspects of experimental
results, as illustrated in Figure 65. Without going into the details of the simulation,
which are discussed in detail in reference [71], we highlight a few salient results:
32The introduction of an artificial thennal agitation to simulate, through the bias of a dynamic numerical
technique, the static properties of a stack hides a profound reality. As we shall see in Section 6.4, the
dynamic properties of a stack are less subject to indeterminacies (in the sense of Section 3.1.1) than
are its equilibrium characteristics. Forces are mobilized and well defined at all times during multiple
collisions, whereas unknowns persist at rest. Furthennore, thennal agitation implies a fictitious
source of fluctuations which has no basis in reality. This touches on a fundamental problem in
matters of simulation and understanding of the behavior of granular materials.
t =0 s
0.02 s
0.04 s
0.06 s
109
0.08 s ,
FIGURE 65. Simulation of the fracturing of a stack in guided fall under conditions similar
to the experiments. Note that thermal agitation, required by the simulation, causes a kind of
artificial boiling effect at the top of the pile. The time t is recorded in seconds. The aspect
ratio and micromechanical coefficients match the experimental values (after [69]).
Numerical simulations produce satisfactory descriptions only if the angular momenta of particles are included. Otherwise, the results look more
like a gradual dilution of the stack, from which inverted Vshaped structures
are conspicuously missing, even though they are invariably present in actual experiments. The implication is that rotations playa key role in this
phenomenon. This point will be examined further as we go along.
Fractures appear and grow near the bottom of the stack, in agreement not
only with observations but also with the simplified model discussed above.
As predicted, those fractures that occasionally appear in the upper portion of
the stack are quickly squelched by collapsing back on themselves during the
fall.
Fractures typically get started on one side of the structure, violating the symmetry of the system. They typically emanate from a boundary, much like
what is observed in cantilevered fractures of solids. 33 This trend can also be
observed in numerous snapshots of real experiments.
Having recognized that simulations must take into account the rotation of particles
if they are to produce results in agreement with reality, we still have to understand
why that is. The best strategy to that end is to rely on experiments. That is precisely
the course of action we will follow. We will begin by considering the distribution
33 A
cantilevered fracture occurs, for instance, when we attempt to cut a solid object with a knife. A
fracture often opens up at the point of contact of the knife and propagates according to a dynamics
familiar in geophysics.
110
~
c
0.20
Ijl
:::J
:e
[1!
0.15
0.10
:::J
<Jl
<Jl
[1!
Q..
0.05
0.00
0.05
0.00
z(m)
0.05
0.10
FIGURE 66. Numerical simulation of the pressure against the walls as a function of height.
The integration time is 10 ms, and each point is obtained by averaging over a height
equivalent to six rows (after [69]).
of stresses in a stack in guided fall, and go on from there to examine in some detail
the modes of selforganization of the rotation of a collection of particles in the
immediate vicinity of a fracture.
Distribution of Pressure in a Stack: Arch Effects
Numerical simulations allow us to keep track of the velocities of the particles and
the rate of collisions between themselves as well as with the walls at any point
in space and time. As we have emphasized earlier, we are dealing here with a
dynamic model. As such, we are not in a position to calculate a static pressure.
Rather, we consider that the pressure exerted by the particles results from repeated
collisions and transfer of momenta to the lateral walls such as described by (210),
in which the normal component is conserved. This approach is entirely consistent
with the traditional kinetic theory of gases. The time integration of the momentum
I~"t P dt is carried out over a reasonable duration of the order of onehundredth of
a second. The results of such calculations, obtained for a series of different times,
are shown in Figure 66.
The appearance offractures within the stack (at 0.04 and 0.06 s) corresponds to
a significant increaseby one order of magnitudeof the local pressure on the
walls. This is very much in keeping with our intuitive notion of an arch whose
primary function, as we have seen in Section 3.1.1, is precisely to transfer the
pressure exerted by the upper portions of the stack laterally to the walls. Thus,
through the bias of a numerical simulation, we arrive at some objective and quantifiable information about the mechanism of fragmentation. It appears to result
from the successive formation and disappearance of triangularly shaped arches
pointed upward, as we may have anticipated.
It is likewise possible to study in detail what causes fractures to form in the stack.
The model suggests that a fracture opens up with a single particle whose translation
and rotation velocities (see Section 2.2.2) suddenly have to adjust to the relative
III
(b)
I""~
(a)
/\
(c)
Impossible Configuration
FIGURE 67. Diagram (a) shows a simulation of rotations in a stack in guided fall. The
rotations organize themselves around fractures indicated by arrows. Diagram (b) shows an
organization compatible with Vshaped contact chains of the type observed experimentally.
The arrangement depicted in diagram (c) is incompatible with the arch model and is in fact
never observed in simulations.
speed of the wall. Stated differently, unlike its cousins in the interior of the stack, a
particle interacting with a wall goes into a regime of rolling without gliding, in the
sense defined in Section 2.2.2, which triggers the onset of a fracture. Furthermore,
the model also favors the propagation of a fracture toward the interior along easy
dislocation lines of the lattice. 34 This propagation results in a dramatic increase
in the number of collisions between particles along and around the contact chain.
It would then appear that the mechanism responsible for initiating a fracture is
a kind of organization of the rotations of particles in relation to the wall. Given
the concomitant increase in the number of collisions in the immediate vicinity
of the event, it is logical to inquire whether such a local organization can in fact
propagate along a contact chain. That is our next topic.
SelfOrganization of Rotations
Figure 67 depicts a map of the "spins" of particles in a stack right after the moment two fractures, indicated by arrows, have opened up at the wall. 35 White circles
indicate a counterclockwise rotation with an axis perpendicular to the plane of the
figure, while black circles correspond to a rotation in the opposite direction.
34We use here the standard terminology in crystallography. Given the triangular symmetry of a compact
twodimensional stack, the dislocation lines run horizontally or at a 60 angle with the horizontal.
35 In the present context, the word "spin" designates the rotation of a particle with an angular momentum
112
Figure 67(a), obtained by numerical simulation, shows clearly that the spins
tend to organize themselves in an alternating pattern in the immediate vicinity of
and just above the fractures. Figure 67(b) and (c) may help us understand that the
angle of these patterns is compatible with the notion of arch developed earlier in our
discussion. A contact chain inclined at a 60 angle relative to the horizontal implies
that the rotations of its strongly bound particles alternate. It is intuitively obvious
that the particles forming such a contact chain, being in intimate contact with each
other, will experience frustrated rotation relative to the chains immediately above
and below (see Section 2.2.1). It follows that the weakest points of the structure
run between superposed contact chainsparallel to the direction of the archand
that fractures can develop only along those lines. Here again, we see that the
selforganization of the rotations of particles conforms to the notion of arches as
contact chains which we developed earlier. 36
3.2.5.
Many of the phenomena we have described thus far, such as convection or fragmentation, can be traced to interactions between granular media and the walls of
the containers. Before moving on to the topic of free inclined flows in Chapter 4,
we end this chapter with a variation on the previous case, where the stack is now
much wider than it is high. It is useful to recall some considerations we invoked
in connection with Janssen's static model (Section 3.1.4) and its dynamic extension (Section 3.2.3). We pointed out that, within the constraints of the model, the
behavior of a stack of height h and perimeter P is governed by the decompaction
parameter X = S K f.t through the exponential factor exp ( X). Recall that the aspectratio S = Phi A is a measure of the lateral area normalized to the crosssection
area A. In a twodimensional configuration, the aspect ratio becomes S = 2hl L,
where L is the lateral length. The attenuation of the stresses in the stack may be
written as
exp(x) = exp
(~) =
LIKf.t
exp (_ 2h)
A
'
(a)
113
(b)
A
~ Vibrations
Vibrations
Ph
== 
4h
= 
nd
Rj
0.01.
114
(a)
(b)
(c)
FIGURE 69. Three typical organization patterns on the surface of a 1.2mmthick threedimensional granular pile vibrated at a frequency of 67 Hertz. Diagram (a) shows striations
at f /2 (f' = 4.0). Diagram (b) is an example of competition between squares and striations
at f /4 (f' = 6.0). Diagram (c) shows hexagons at f/4 (f' = 7.4) (after [74]).
ll5
3
Disorder
Hexagons
.....
..c
'5
Ql
i=
Flat
o
2
Acceleration
FIGURE 70. Bifurcation diagram of inelastic particles and organization patterns of the
surface of an extended threedimensional granular medium (after [74]).
The formation of these geometrical patterns results from the combined effects
of two phenomena, at least one of which is specific to granular matter:
First is a phenomenon of successive bifurcations associated with the vibrational perturbation of a granular material behaving as a completely inelastic
object, analogous to what we encountered in Section 3.2.1. 39
Second is a mechanism of parametric excitation of surface waves of the type
observed in a liquid (the ones Faraday observed). This phenomenon involves
a complex filtering effectvia anharmonic couplingof the harmonics and
subharmonics of the excitation frequency.
Additional insight is provided by Figure 70, which reproduces in a slightly different
format the bifurcation diagram shown in Figure 40, combined with the results of
experimental observations relative to the formation of selforganization patterns
such as striations, squares, hexagons, and various combinations thereof.
This bifurcation diagram plots the product f . tf' t f being the time of free flight
of the granular layer, against the normalized acceleration r. It reveals the existence
of critical points separating regions corresponding to different geometrical figures.
As the acceleration r is increased, we observe successively striations at f /2,
hexagons at f, striations and squares at f/4, hexagons also at f/4, and eventually
almost complete chaos when r reaches the value 8. For frequencies between 10
39Considering how thin the layers are in this experiment, we might legitimately be concerned that the
condition N" (1 8) :c: 3, required to ensure that the stack will not be fully fluidized, may be violated.
The reader may convince himself that eight balls with a coefficient of restitution of 0.6 are sufficient
for the condition to be verified. As it turns out, compliance with the above condition is not crucial
for this particular experiment.
116
o
.000
.002
1/f
.004
2
.006
(sec 2 )
FIGURE 71. Wavelength of organization patterns plotted against the inverse of the square of
the excitation frequency. Both straight lines have a slope of 1, indicating a linear dependence
for spheres of diameter 0.4 mm (square data points) or 0.2 mm (triangular data points). Only
the ordinate at the origin depends on the size of the spheres. In this case, r = 3.5 (after
[74]).
and 100 Hertz, these experiments confirm that it is indeed the parameter r that
governs the appearance of selforganized surface patterns, as well as their contrast,
validating the model of an inelastic ball.
Wavelength Dependence on Vibration Frequency
where Amin depends only on the diameter d of the particles and is approximately
equal to lId. A similar equation appears in the book Fluid Mechanics by Landau
and Lifshitz, in a paragraph dealing with gravitational surface waves. 40 In this
light, the parameter geff can be construed as an effective acceleration which must
be a fraction of gravity's acceleration. That is indeed the case here, since we find
2
geff ~ 3.1 m/s .
40See page 36 of L. D. Landau and E. M. Lifshitz, Fluid Mechanics (Oxford: Pergamon Press, 1987).
117
(a)
Hertz. The horizontal scale is different in the two cases. In reality, the wavelengths are
such that Al = 2A2.
G eff
f2'
Summary
The selforganization of surface patterns in extended two or threedimensional
granular media result from the superposition of two wellknown phenomena. One
118
is related to the bifurcation diagram associated with an inelastic ball, and the
other is Faraday's instabilities in liquids. The latter connection is intriguing and
suggests the possibility of making a parallel between the viscosity of a liquid and
the degree of looseness of a granular sheet. Experiments indicate that a loose
granular medium indeed has a lower apparent viscosity (in the sense defined for
liquids) than a compacted one (see, for instance, Section 2.4.2). Intriguing as it
may be, this type of analogy remains largely in the realm of speculation.
4
Granular Media in a State of Flow
4.1.
We all know that with dry sand it is impossible to make a pile with sides that are
vertical or even steeply inclined relative to the horizontal. As soon as the slope
exceeds a certain value, the pile collapses (i.e., relaxes back) until the slope returns
to an angle e which, curiously enough, always seems to be near 35. This angle
is referred to as the "angle of repose." It might actually be more accurate to speak
120
FIGURE 73. The figure on the left shows the embankment angle of a twodimensional pile.
The one on the right depicts one of many ways to create a coneshaped sand pile. Grains
are dropped one at a time on a roughened horizontal base.
of angles in the plural, for it turns out that there is more than one, as we shall see
shortly. The phenomenon is illustrated in Figure 73.
The first individual to make quantitative observations on the angle of repose was
Charles de Coulomb, who back in the eighteenth century was a military engineer
responsible for building fortifications [11]. Being well acquainted with friction
between solids, as mentioned in Section 2.2.1, he proposed a simple explanation
that is still considered the authoritative word on the topic. His model is rooted in
the idea that two contiguous sheets of a dry granular material cannot slide relative
to each other unless their inclination e is at least equal to tanj(ILs)' By analogy
with what we know about friction between solids, ILs is to be interpreted as a
coefficient characteristic of the friction forces involved.
This analogy suggests that Coulomb's friction law applies to granular media as
well, which is indeed supported by experimental observations, as pointed out in
Section 2.4.1. Yet, for all its simplicity, this line of thinking raises a number of
questions on closer examination. For starters, how do we define the "weight" of a
granular sheet? We have already emphasized repeatedly that the distribution of
forces exerted on the surface of such a sheet is anything but uniform and a far
cry from the simple description applicable to a massive solid placed on a support.
Based on the arguments we invoked earlier concerning the microscopic mechanisms of solid friction, we can easily appreciate that the problem is even thornier in
the case of a granular sheet. The many indeterminacies and hysteresis phenomena
alluded to earlier all contribute to complicating the task of defining unambiguously
what we mean by angle of repose. Indeed, both experiments and theory show that
there are not one but several angles of repose depending, among other things, on
how a pile was prepared.
A wealth of technical details on this topic can be found in the book by Brown
and Richards [5]. We will simply summarize them here by stating that the relative
indeterminacy of the angle of repose stems from two main factors:
4.1
Concave
121
Convex
FIGURE 74. A convex pile generally has a lower embankment angle than a concave file,
such as exists near the opening of an hourglass .
The first factor is geometrical, as it involves the shape or, more precisely, the
curvature of a pile. The point is illustrated in Figure 74, which may help us
realize intuitively that particles in the vicinity of the free surface tend to be
more densely surrounded in a concave pile (in the form of a crater) than in a
convex one (shaped like a mound).
Any difference is expected to vanish when the radius of curvature of the
surface becomes much larger than the mean diameter of the granules. These
predictions are essentially confirmed by experimental results such as those
reproduced in the following table:
Material
Geometry
"Mound"
angle
"Crater"
angle
Dynamic
angle
Tapioca
Sand
Coal
spherical
angular
angular
30
37
37.5
37.5
39
41
32
36.5
34
122
Angle of movement
before relaxation
Angle of repose
after relaxation
FIGURE 75. Definitions of the angle of movement em and the angle of repose er . The
difference.5 = em  er is called the relaxation angle. Its value is typically about 2.
4.1
123
FIGURE 76. In a pile made of a small number of particles, the relaxation angle 8 goes to
zero.
concerns the transition from a regime of intermittent flow to one that is continuous.
We go on to examine these two problems.
difference between angle of repose and angle of movement may be artificially reduced, if not completely eliminated. With reference to Figure 76, D is the diameter
of individual spheres and L is the length of the path of an avalanche. If we start at
the angle of repose er , adding a single sphere may be enough to exceed the angle
of movement em. Under such conditions, the two angles become indistinguishable
since both correspond to the onset of an avalanche.
Let us designate by N the number of spheres in a threedimensional pile assumed
to be in the shape of a cone with an angle close to the critical value. N is related to
the relevant linear dimensions by (L / D) :=:::; (3N) 1/3. Within these approximations,
the relaxation angle (, may not exceed the value (3N) 1/3. This means that there is
a minimum number N min of particles below which the relaxation effects described
above will cease to be relevant for purely geometrical reasons. This number is
given by
Nmin
~ (~
r3~3'
We have pointed out earlier that (, is typically equal to 2 for real piles involving
a large number of particles. N min is then approximately equal to 8000 granules.
These considerations show that in matters related to the embankment angle, it is
important to keep track of the number of particles constituting the pile.
Q,
For slow rotations (typically less than 0.1 revolution per minute, or rpm),
there is an intermittent flow in which the upper sheet of material oscillates
back and forth between the two angles em and er defined earlier. Varying
quantities of granulars are driven toward the bottom of the pile during each
124
(b)
FIGURE 77. Profile of the free surface of a pile in a slowly rotating drum. (a) The surface
is essentially straight when the rotation speed is just beyond the threshold for continuous
flow. (b) At higher speeds, we witness the appearance of a surge wave, as the influence of
the boundaries at the top and bottom of the flowing layer become noticeable.
Intermittent
flux
"
Continuous
flux
Q
Q+
     +  1    +  1~> Q
FIGURE 78. The hysteresis effect between intermittent and continuous flow. The arrows
indicate whether the rotation speed of the drum is being increased or decreased.
4.1
125
same in the intermittent and in the continuous regimes, for which the fall times are
designated f) and f2, respectively. The transition from intermittent to continuous
flow occurs when the fall time of a granule becomes synchronized with the interval
T separating two successive avalanches. As we will see shortly, T is actually a
random variable. Nevertheless, its fluctuations are small enough for the argument
to hold. In particular, the regimes switch over when Q+ = 8/f) and Q_ = 8/f2,
with f) < f2.
While the angle of movement 8m is clearly a fundamental property of sand and
evidently reflects a critical phenomenon, things are not so clear when it comes
to the angle of repose 8r after the system has relaxed. Even for largesized piles,
the angle of repose is generally affected by the fact that the flowing sheet has to
come to a stop at the bottom of the drum. Such finitesize effects are less of a
nuisance than in the case of smallsized piles, but they are real nonetheless. The
influence of the wall must be taken into account if we are to develop an accurate
description of a phenomenon that turns out to be far more complex than it appears
at first.
Since we are dealing with a critical phenomenon, it is logical to want to examine
how the flux of material varies as a function of the inclination angle 8. ) In particular,
the question is whether this particular phenomenon might fit in with other critical
transitions by obeying a law of the form
(41)
where L is the length of the drum, and R its radius. Note that J is expressed as a
volume per unit time.
The results shown in Figure 79 correspond to a 19cmIong drum halffilled
with particles of diameter 0.3 mm. Practical considerations impose that the rotation speed Q be restricted to a range between 0.5 and 12 rpm. The lower limit
corresponds to the onset of continuous flow, while at the upper end a substantial
fraction of the flowing sheet is sent flying off due to the centrifugal force. The
experiment consists simply in measuring the inclination angle of the flowing sheet
as a function of rotation speed of the drum. Since we are looking for a power law,
the results are plotted in a loglog format. The graph shows that in the range of
variables investigated there is indeed a power law of the type J ex (8  8e n , with
m = 0.5 0.1. It is interesting to look for the physical significance of the exponent m. Some useful insight can be obtained by reviewing the case, well known in
hydrodynamics, of an ordinary flowing liquid (Brownian flow).
lThe angle e, called dynamic angle, defines one of the characteristics of avalanches. It obviously
depends on the speed of rotation. Under certain conditions, it also depends on the size of the granules.
We will see an application of this poorly understood property in Section 5.4.
126
20
0
0
Q
10
f*
f
3 4 5
10
.n (rpm)
FIGURE 79. Loglog plot of the angular deviation from the critical value versus transport
flux during avalanches in the continuous regime (after [75]).
pgh 3
=   sinCe),
31]
4.2
Avalanche Models
127
In any event, according to this model, the friction force resulting from a velocity
gradient V v is proportional to (V v)2. We can thus write the equilibrium between
the friction force and the force of motion as
av)2
ex ( az
+ pgz[sin(B) 
JLcos(B)] = 0,
(43)
= '3
This explains the J ex (B Be) 1(2 dependence suggested by de Gennes and observed
experimentally. Based on this elementary calculation, Bagnold's law in (VV)2 and
a simple law of dry friction are all that is needed to account for the exponent
characterizing the behavior of a granular flow near the critical point.
128
ill
FIGURE 80. Principle of the onedimensional cellular automaton model (CAM) (after
[77]).
model that establishes a relation between the slope of a pile and the flux of particles in motion. Finally, we will examine a model based on coupled variables that
accounts satisfactorily for a number of characteristics of avalanches.
4.2.1.
The model was originally proposed by Bak, Tang, and Wiesenfeld (BTW) specifically for the purpose of studying systems in a selforganized critical state [77].
Nonetheless, the cellular automaton model (or CAM for short) is directly related to avalanche processes, although some caution is warranted in making this
generalization. 3
The Principle
Squares are stacked on top of each other to form contiguous columns according
to a set of extremely simple rules:
(l) The height difference between two adjacent columns cannot be greater than
two units. This effectively simulates the angle of repose, which cannot
exceed a critical value without collapse.
(2) When a column rearranges itself because of excessive height relative to its
neighbors, it involves the movement of a set of two unit cells. This is akin
to the domino effect in an avalanche.
Any starting configuration is allowed to relax back to its equilibrium in accordance with the rules stipulated above. The end result is a stable system, such as is
illustrated in Figure 80, which becomes the starting configuration for subsequent
experiments. Individual squares are then added at random to the edifice, one at
a time. Each such event triggers a relaxation process of its own, subject to the
3In this chapter, the term "selforganized criticality" (SOC) applies to a system evolving spontaneously
toward a critical state with no memory of the initial conditions. In other words, it is a system for which
the critical state is an aUraetor as far as its dynamic properties are concerned. The interested reader
is referred to the abundant literature on this topic. There has been much debate about whether this
model is at all pertinent to avalanches. Suffice to say that caution is strongly advised, if for no other
reason than because of the existence of two different angles em and er , rather than a single one.
129
same rules. Since the bottom surface is assumed to be of finite size, squares cannot accumulate there but, instead, drop off the base. The objective of this thought
experiment is to count the number of squares that drop off after each new square
is added to the pile. The release of one square can trigger mini "avalanches" of 0,
2, 4, and so on, squares escaping off the side. As it turns out, small avalanches are
far more frequent than larger ones.
This elementary CAM process, described here in one dimension, can be generalized to the case of 2, 3, or more, dimensions by developing suitable computer
algorithms. As artificial as such generalizations may seem in terms of describing real avalanches, they are of considerable interest because theories can often
predict the value of the exponent in the power laws applicable to any dimension,
even higher than 3. Exercising such algorithms often requires the use of parallel
computing so as to cut down on the computation time and minimize roundoff
errors which are the bane of conventional sequential machines.
As an example, we can go through the algorithm describing the sequence of steps
in the onedimensional case discussed above. The difference in height between
adjacent columns is defined as Zn = hen)  hen + 1). When a single square is
added to the nth column, it entails the following changes:
Zn
J>
Zn
+1
and
Whenever the height difference becomes greater than a critical value Zc, the system
relaxes back in such a way that
and
Znl J> Znl 
The pile is bounded on the left (corresponding to the index 0), and open on the right
(index N), where squares drop off the pile. Therefore, events at the boundaries are
written as
Zo
0,
ZN
J>
ZN 1,
ZNl
J>
ZNl
+1
for
Zn > Zc.
gives a total number Z;: of stable configurations. They are not all equally stable,
however. A simple technique to arrive at the least stable of these configurations is to
artificially construct a system in which all columns are unstable (meaning that Zn >
Zc for all values of n), and let it relax spontaneously to some equilibrium which
becomes the fresh starting point from which new squares are then added. Such a
minimumstability configuration is a critical state in the sense that any subsequent
130
disturbance can propagate through the entire pile. This situation is, therefore,
analogous to the wellknown problem of percolation in one dimension. The issue
of stability is considerably more complicated in structures with dimensions greater
than 1. It is decidedly beyond the scope of this book, but we refer the interested
reader to the many articles on this topic published in the wake of the original BTW
paper. Extrapolating the algorithm itself, on the other hand, to more than one
dimension is straightforward. We will do it here for the twodimensional case. For
a square pile described by Z (x, y), the equivalent of the chain of events considered
previously becomes
Z(x  1, y)
Z(x  1, y)  1,
Z(x, y  1)
Z(x, y  1)  1,
Z(x, y) ~ Z(x, y)
+ 2.
Z(x, y)  4,
1) ~ Z(x, y
1, y)
Z(x
1) + 1,
1, y) + 1
for
where the exponent T turns out to be approximately equal to 1.0 in two dimensions,
and 1.37 in three dimensions.
It is also of interest to examine the probability distribution function of the
durations (or lifetimes) of these avalanches. In other words, given a system in a
critical state on which an additional square is dropped randomly, how long will the
ensuing slide last? Intuitively, we expect some degree of correlation between the
size of a slide and its lifetime. An event triggering a large number of cells to fall
down the side ought to take longer than another one triggering just a few. Although
it takes a little more work to prove it than in the case of the size, "experiments"
131
show that lifetimes do in fact obey similar power laws, expressed in the form
D(T) ex T"',
(44)
where ex is found to be 0.43 and 0.92 in two and three dimensions, respectively.
goodquality hifi amplifier, for instance, transmits all frequencies between 20 Hz and 20 KHz
without distortions. We might wonder what the frequency spectrnm of the output noise of such an
amplifier might be. Under ideal conditions (with all other noise sources eliminated), the noise turns
out to have a II! spectrum. This remarkable property is due to the shot noise associated with the
input impedance of the amplifier.
132
iiiii
S(!JI/( fJj 2
p'EJwer
Spectrum
11 fO
~
rS229 ~
I~~~~~~~'''~rn
Time I
log f
FIGURE 81. Three types of commonly encountered noise, and their power spectral densities (f0, r 1, and r 2) (after [78]).
of attention on the part of researchers. This type of noise is one of the characteristics
of systems that enjoy selfsimilarity, also known as fractal objects. The reason is
that the noise power in a small frequency window df is given by S(f) df; when
S(f) = Ilf, this noise power becomes df/f. In other words, the noise power
remains invariant provided the bandwidth df is scaled to match the mean frequency
f [79]. That is one of the fundamental properties of socalled selfsimilar systems,
which exhibit the same properties on all scales.
As we noted in connection with (45), the cellular automaton model (CAM)
leads to a frequency probability distribution function very close to a 1If law. This
immediately suggests the possibility that the behavior of such automatons may be
assimilated with that of a selfsimilar system. As we are about to see, this turns
out to be a complicated issue whose answer is not black and white.
The Statistics ofAvalanches
We have stated in Section 4.1 that real granular piles are characterized by no fewer
than two critical angles em and er as long as the pile is made of a large number
of particles. If the number of particles is small, the two angles merge into one.
We may therefore expect different behaviors depending on the number of particles
involved. This is precisely what is observed in actual experiments.
Several research groups, inspired by the results just described, have tried to verify that the distributions of real avalanches conform to the power laws suggested
by the cellular automation model. To that end, they have devised a number of
experiments based on different principles, illustrated in Figure 82. In Figure 82(a),
particles are dropped one at a time on a coneshaped pile supported in the pan
of a balance interfacing with a computer. The weight of the particles falling off
the pile can be measured with reasonable accuracy and gives an estimate of the
(b)
(c)
I r I r
Scale
Condenser
Condenser
133
(d)
f7
Microphone
FIGURE 82. Four techniques used to study the statistics of avalanches (after [20]).
size of successive avalanches. In Figure 82(b), a condenser is used to count almost
onebyone the particles as they roll off a pile that is being resupplied continually.
A condenser is also used in Figure 82(c) to count the number of particles escaping
from a partially closed halfcylinder as it rotates slowly around its axis. Finally,
in Figure 82(d), a drum almost halffilled with a granular material spins slowly
around its axis, causing periodic avalanches of varying sizes. A tiny microphone
placed near the apparatus records the noise of these avalanches. All these methods have yielded results that clearly depended on the size of the population of
granules.
134
~;~J
0 1 2
t (ks)
2..r,.,
{j)
C;
..Q
2
4
3
(b)
2
1
lag(r)
FIGURE 83. Statistics of avalanches observed in a rotating drum. Diagram (a) shows the
number of particles transported as a function of time as the drum rotates at a constant speed
of 1.3 degree/min. Diagram (b) shows the power spectral density. The dotted line indicates
a 1/f dependence. The granular material is made of particles 0.5 mm in diameter (after
[80]).
surface makes with the horizontal, thereby dropping it below the critical
value, with drastic repercussions for the physics of the process. This is a
classic example of a finitesize effect. It acts somewhat like a highpass
filter, as it were, favoring larger events.
(2) As noted above, the splitting of critical angles into two valuesangle of
movement and angle of reposeimplies that a system that has just experienced an avalanche is no longer in a critical state. It is necessary to incline
it by an additional angle em  er in order to reach a new critical state. It is
then not obvious that real avalanches involving many particles truly constitute critical systems, even if they can still be considered selforganized.
In light of some of our previous observations, however, it seem likely that
small avalanches may not suffer from these restrictions. To some extent, the
experiments we are about to describe support this view.
10
Avalanche size (number of particles)
135
100
FIGURE 84. Results of experiments on limitedsize piles made of different particles. The
fractional rate of avalanche occurrence (i.e., the number of avalanches of size s normalized
to the total number) is plotted vertically against the number of particles s involved in the
corresponding event. In (a), the triangles correspond to steel balls, and the circles to glass
beads. In (b) the squares refer to polystyrene beads and the circles to glass beads (after
[78]).
pile supported in the pan of a balance. The balance keeps track of the weight of
material falling off the pan during each avalanche event. We may expect that the
shape of the pile should playa fairly important role, since a convex pile is inherently
twodimensional, in accordance with the model developed in Section 4.2.1.
In a configuration of this type, the sizes of successive avalanches do exhibit a
considerable dispersion [78], [81]. Smallsize avalanches are plentiful, while large
ones are relatively rare. Figure 84 also reveals that, regardless of the nature of
the materials (steel, glass, or polystyrene), the statistics of avalanche sizes indeed
conforms to a power law, as predicted by the cellular automaton model. It may be
objected, however, that the agreement has been demonstrated over not much more
than one decade, which some may deem too small a dynamic range to truly test
the validity of the law.
Although several experiments conducted under different conditions have all
confirmed that the behavior of a granular pile differs according to whether it is
made of a large or a small number of particles, the reasons for the crossover
from one regime to another are still not fully elucidated. It is possible that a
more sophisticated cellular automaton model may be able to account for such
finitesize effects. Yet, the existence of the two angles em and en as well as the
relaxation oscillations between these two extreme states, constitute a fundamental
objection difficult to overcome when modeling a dry granular material. A number
136
(a)
(b)
'2
o
.............................
24
21
18
_,
ff
Vibrations
(c)
3
2
.. ,
loglOCf(sec 1)J
g;
15
12
~..
....
"........
..
.... .....
.9u =26.1
.. 8..=22.2
.9..=19.5
.9..=16.4
FIGURE 85. Results of experiments done with a rotating drum subjected to vertical vibrations. The angle ess, which corresponds to the steadystate angle obtained for a rotation
speed of 1.3 degree/min, is a measure of the excitation intensity. The dotted line in (b)
indicates the 1/f behavior predicted theoretically. In (c), Q = 0, and the structure relaxes
according to a log(t) behavior, whereas the CAM model predicts a relaxation proportional
to t (after [80]).
4.2
Avalanche Models
137
in the static case. The question we seek to answer is whether vertical vibrations
can put a pile in a supercritical state characterized by a single critical angle, such
that the power law predicted by the cellular automaton model is observed. Based
on the curves shown in Figure 85(b), the answer is decidedly no. Even when the
avalanche angle, which started out at 39, is lowered by more than 2enough
to cancel the difference 8the power spectrum of the avalanches continues to
exhibit a peak, which is inconsistent with the SOC model. As the critical angle is
lowered by even greater amounts, the power spectrum begins to look more like a
power law, but with an exponent that deviates substantially from the 1If behavior
expected theoretically. The experiments suggest a dependence closer to II fo. 8 .
Nevertheless, this is not enough to conclude that the poor agreement between the
CAM model and the real systems is solely due to the existence of the two angles
em and er 5
Other measurements done with the same apparatus provide a wealth of information. The pertinent results are shown in Figure 85(c). Here the drum no longer
rotates, but it is stopped in a position such that the free surface of the granular
material is at an angle e just below the value er When the vertical vibration is
then turned on, the angle of repose turns out to decrease (or relax) with time.
Furthermore, the rate of relaxation depends on the amplitude of the current in the
loudspeaker. The results shown in Figure 85(c) constitute fairly compelling evidence that the embankment angle e evolves as log(t), while the CAMSOC model
suggests that it should vary proportionally to t. This rather remarkable result has
been interpreted in terms of a simple model involving a relaxation process induced
by an agitation (caused by the loudspeaker) that may be considered the equivalent
of a "thermal" agitation.
In the context of this model, the intensity of the vibration plays the role of an
"effective" temperature Teff [21], [80]. There is a perhaps rather bold analogy to be
made with the phenomenon of electrical conductivity, where a flux j of electrons
results from an applied electrical field E through the bias of the conductivity a,
the latter being of a magnitude governed by the degree to which electrons are
released from randomly distributed traps. Here, the electric field Ethe driving
forceis assimilated with the angle e, and the current density j with the rate of
change deldt. By analogy with the electrical case, a granule can be thought of
as trapped, much like electrons in a conductor, by neighboring granules. Rather
than reviewing the many assumptions involved in this problem, we will simply
calculate the average height of an effective barrier V as a function of the angle e.
Expanding V to first order about the starting point of the experiment, we may write
V ~ V o + VI (e  er ). Knowing that a flow is triggered when e = em, we impose
the condition V (em) = 0, which leads to 8 == e  er = Vol VI. Pursuing our
electrical analogy, we write that the rate of change del dt, like the current density,
5 As
we will show in Section 4.2.2, the dragging effect of the rotating disk, even as slow as it is in
this case, can introduce an artificial periodicity in the sequence of avalanches. This point is still
being actively debated. Nevertheless, very recent experiments have for the most part confirmed the
conclusions presented in this paragraph.
138
de
 =
dt
where both A == A o exp( Vol kTeff) and 1'3 = VI! kTeff are independent of e.
The solution of the differential equation can be expressed by means of the
exponential integral function E 1 (tJe). Approximations valid when the argument
M I lead to
I
e ~ er  loglQ(tJAert
1'3
+ 1),
which is consistent with the loglQCt) behavior observed in Figure 85(c) for times
greater than to = 111'3 Aer . The agreement remains reasonably good even for shorter
times. This derivation thus accounts for the loglQCt) behavior of the embankment
angle of a vibrated granular pile.
Interpreting avalanches in terms of thermal detrapping is fundamentally different
from the critical angle model proposed by BTW. The experiments described above
clearly show that the simplistic view of the angle of repose er as a critical angle
does not stand the test of a detailed analysis of the relaxation mechanism.
4.2.2.
Chronologically, the model we are about to present came into existence after the
ones described above (see Fauve et al. in [27]), against a backdrop of mounting
evidence that the SOC model had serious shortcomings. It starts from an altogether
different point of view. The underlying idea is remarkable simple. It is based on
a system of phenomenological coupled equations describing the behavior of two
pertinent observable quantities, namely, the angle e of a sheet in motion and the
flux D of particles sliding down a slope during an avalanche. 6 The basic principle
is embodied in a particular relation between friction force and velocity and in
the coupling of friction with a spring, as outlined in Section 3.1.1. As such, it
takes advantage of an explicit parallel, already suggested in Chapter 3, between
the alternate stickslip motion related to solid friction and intermittent series of
avalanches.
Figure 86(b) illustrates the approach. It shows a pad rubbing against a surface
with static and dynamic coefficients of friction Ihs and Ihd (x), respectively. As
usual, Ihd is smaller than Ihs and can depend explicitly on the velocity, as explained
in Section 2.3. The pad is also connected to a spring of stiffness K and moving
at a velocity V. This arrangement is entirely similar to that in Section 2.2.1, and
was subsequently used to analyze the phenomenon of hysterisis in Section 3.1.1.
If i; designates the deformation of the spring and Q) the speed of rotation of the
6The paTticles in flow are sometimes referred to as a "moving species." The idea of describing the
problem by means of these two coupled variables has recently been exploited from a different angle.
It has led to several interesting developments which we will discuss later on.
(a)
139
(b)
FIGURE 86. Illustration of the correspondence between (a) avalanche processes and (b) the
stickslip mechanism. Both phenomena can be described by the same set of differential
equations (after [27]).
cylinder depicted in Figure 86(a), it turns out that the two systems are governed
by the same set of equations, with the following correspondence
w ++ V.
In particular, when the pad is in motion, the elongation of the spring is given by
the equation
if
dx
=0
dt
and
140
til
OJ
(jJ/y
:0
co
<P f
<Pd
Variables
<Ps
e or 1;
de
(46)
=0
dt
'
dD
=0
dt
'
(47)
if
D =0
and
(48)
When a flow occurs, the preceding analogy allows us to use the equations established in Sections 2.2.1 and 3.1.1
de
=wyD,
dt
dD
dt = P[sin(e)
 Itd(D) cos(e)],
if
D=O
and
where P = gh, h is the thickness of the moving sheet, and y depends on the
geometry of the flow; to a first approximation, y can be considered a constant. The
first of these equations expresses the conservation of the mass of material driven
141
to the bottom and replenished at the top by the rotation of the cylinder. The second
equation is obtained by analogy with the results found in Section 3.1.1.
The notation can be simplified by introducing an angle cD d defined by
or
We should bear in mind that fJd depends on the flux D, which is itself given by
D = _w__d_e/_d_t .
y
Eliminating D between the first two of our set of equations yields a secondorder
differential equation in e
where p is defined as p = p / cos (cD d). This last equation describes oscillations
of the angle e with a period (l/y p )1/2. The steadystate solution, corresponding
to a constant flow of the rolling species, is given by
and
As expected, the stability of this solution and, more generally, the overall behavior
of the system, is critically dependent on how the coefficient of dynamic friction
fJd depends on the flux D.
Expanding cDd(D) to first order around the steadystate value DO, we get
cDd
cD + (acD
 d)
aD
= cD  (acD
1 de
 d) ,
aD y dt
(D  D )
0
which can be substituted back into the secondorder differential equation to give
(49)
142
In plain language, at very low rotation speeds, the relative velocity imparted by the
cylinder becomes negligible, and the angular windows ,6.(} and ,6. <P both converge
toward the difference em  er , which is what we called the relaxation angle in
Section 4.1 and is defined unambiguously only when the rotation speed is slow.
Next, we determine the effect of the rotation of the disk on the duration r of an
avalanche. That quantity can be derived from (49) and is given by
r =
2 { rr
ffi
 tan
1[ffi
~(<Ps
 <Pd)
]}
143
o
~
a
1:5
CD
.~
I
cD d
Angle
cDs
FIGURE 88. Numerical calculations of the trajectories D(e) with a very low rotation
speed (w = 10 3 rad/s). The coefficient of dynamic friction decreases with speed, which
introduces an asymmetry in the curves, while the angular difference between trajectories
gradually increases (after [27]).
increasing function of the flux D. For instance, in the limit of a small and constant
variation a =  ~ (B<Pd/B D)o, (49) can be solved in closed from, giving the result
More general cases can only be solved numerically. Figure 88 shows an example
of such a calculation. Experiments done with noninvasive observation techniques
reveal remarkably good agreement with the present model, in spite of its great
simplicity.7 Further details on the comparison between theory and experiment can
be found in the original article on this topic [27]. By way of example, Figures 89 and
90 reproduce the results of two experiments conducted in the same drum rotating
at two different speeds. The similarities with Figures 87 and 88 are striking.
As a general comment, we might point out that avalanches are rather chaotic
for very slow rotations and tend to become more regularly periodic as the rotation
speeds up. This is suggested by the model and certainly appears confirmed experimentally. An important consequence of this observation is that it is generally
difficult to carry out reliable and controlled measurements in a rotating cylinder.
The same comment actually holds true for any other configuration. Gradually increasing the inclination angle so as to drive the system out of equilibrium must
always be done as slowly as possible.
As pointed out previously, this particular stickslip model for avalanches leads
to a secondorder differential equation, in which the dissipative term (in de / dt)
involves a nonlinear dependence of the friction coefficient f.id on the flux D. The
7The term "noninvasive" technique means that the measurements do not disturb the object being
measured. These techniques are most often optical in nature. They include the imaging methods
described in Chapter 3.
144
6
5
4
3
'CD
o
1
0.4
1.2
0.8
1.6
2.0
(arbitrary units)
FIGURE 89. Optical measurements of the slope of piles and its time derivative for a series
of avalanches. The relation between the amplitude of avalanches and the angle at which
they begin is evident. The trajectories exhibit the asymmetry predicted by calculations. The
rotation speed used here is a rather slow 0.023 rpm (after [27]).
2
CD
2
0.5
1.0
1.5
2.0
8 (Arbitrary units)
FIGURE 90. Same as Figure 89, but with a faster rotation speed of 0.52 rpm. The trajectories
become more symmetrical and nearly periodic (after [27]).
145
dependence in question can take on varied fonns, which determine to a great extent
the precise behavior of the flow.
Put another way, the key to any analysis of stickslip mechanisms is to learn
something about the F (v) dependence, F being the force of friction and v the
velocity. The functional shape of F (v) has in fact been the focus of a great many
investigations, notably among geophysicists and geologists, who are interested
primarily in earthquakes. A comprehensive review of these largely heuristic models
would be beyond the scope of this book. It nonetheless serves a useful purpose to
review a few general principles, which we now proceed to do.
The BurridgeKnopoff pad model is in many ways the paradigm among a slew
of approaches aimed at deriving functional laws for F (v). The model does suffer
from some weaknesses, which have inspired numerous variations attempting to
remedy them. We will present only a brief description.
The principle behind the BK model is illustrated in Figure 91 [82]. A series of
pads, each of mass m, rests on a horizontal base. The pads are connected to a rope
by meanS of identical springs of stiffness k. The pads are also interconnected by
a different set of springs with stiffness K. As it moves horizontally at a constant
velocity v, the rope drags along with it all the pads and causes them to slide against
the base (it is important that the pads not bounce off the base during their motion so
as to maintain friction at all times). We sense intuitively that such a system can be
the seat of complex oscillations. Our objective is to find the applicable fonn of the
force of friction, which we know must depend on both k and K. The differential
equation governing the motion of the j th pad, flanked On either side by pads j  1
and j + 1, reads
2
d_X,
_
J
dt2
K(X+ 1  2X,
J
+ X J'I) 
,)
k(X  vt)  F (dX
_J
J
dt '
vt  kF(v).
Constant velocity
146
Velocity
FIGURE 92. Dependence of the dynamical friction force on velocity in the BK model.
A detailed analysis of these equations shows that the F (v) has a functional dependence such as is shown in Figure 92. Except very near the origin, the curve F(v)
features a negative slope which, as described earlier as well as in Section 2.3, is
likely to give rise to amplification and possibly chaotic oscillations. This particular function, and others of similar shape, translate a very common phenomenon
known in geophysics as the velocity weakening friction law. In plain language, it
simply means that the friction forces decrease as the velocity of the relative motion
increases. 8
Other Friction Laws F(v)
deadly manifestation of this trend is the phenomenon of aquaplaning of vehicle tries on wet roads.
Cal
2:Q
.g1l5
LLo
(.J
A sphere on an
inclined sheet.
"Half W" curve
Vshaped curve
c al
O'u
t;:E
'C;
LL
Velocity
Contact fatigue:
"InvertedV" curve
(Carlson and Langer, 1989)
"u ~~
"i::
LL
~
Velocity
IC719~)
Velocity
c al
0'(3
147
vv
Nshaped curve
(Barenblatt et al. 1981)
cal
0'(3
UtE
"C ~
LL
Velocity
FIGURE 93. Various proposed functional dependencies of the friction force on velocity.
A spatial coupling that accounts for the propagation of the rolling species after
the movement has been initiated. It also explains the duration of avalanches.
Obviously, the model has limitations. For one thing, it does not enable us to
calculate the profile of the granular flux. There exist other, more sophisticated,
models that can do that. They too start from a system of coupled differential
equations, but they involve a different set of variables (in this case, the flux of
rolling species and the height of the pile) and are based on a rather different
principle. We will take the time to briefly discuss the broad outlines of these
models, which are still in the process of being developed. We deliberately limit
ourselves to a phenomenological description.
4.2.3.
The basic idea behind such models is to work with two carefully chosen variables,
inspired by those used in deriving the fundamental equations of hydrodynamics
[84][86]. We consider the simplest possible case, which is of dimension I + 1.
A single spatial variable x suffices to describe the dimension along which the
flow of particles proceeds. The height h(x, t) of the pile depends on both position
x and time t. At a point of abscissa x, the slope of the pile is measured by the
derivative with respect to x. For simplicity of notation, we denote this derivative
axh, the minus sign being there to remind us that the free surface slopes down
in the direction of increasing x. A critical situation will be reached when the slope
axh exceeds a certain threshold Sc that can obviously be identified with the angle
of repose er discussed earlier.
We introduce a second variable R (x, t) corresponding to the density of rolling
species. It characterizes the flux of matter sliding down the slope. It too depends
on both position and time. Some authors have proposed that the density R(x, t)
148
 =
3t
3x (vR)
(410)
where v represents the velocity of the sheet toward the bottom of the pile, and D is
a diffusion constant that can drive particles in both directions. The first two terms
on the righthand side of (410) are quite familiar; they correspond to the usual
convection and diffusion mechanisms. The last term feR, h), on the other hand,
is the crux of the matter, because it must be able to account for the properties of
avalanches. As such, what form to chose for it bears some discussion. The function
f (R, h) must act as a mathematical operator capable of stopping a particle in motion or, conversely, setting one that is standing still in motion. Taking our cue from
the phenomenology discussed earlier in this chapter, we may begin to have a feeling
for what this function might look like by imposing a certain number of conditions:
A granule at rest can be set in motion only if it is dislodged by another granule
already in motion.
This dislodging process is effective only if the local slope axh exceeds Sc
(in the context of our previous discussion, this means 8r < 8 < 8m ). In order
to further simplify the notation, we subtract Sc throughout from the slope.
Under these circumstances, the gradient of hex, t) will be a small quantity
at all points along the xaxis. Note that, with our sign convention, axh >
indicates a slope that is less than the critical value.
If 3x h > 0, the granules are no longer in unstable equilibrium, since we are in
a subcritical situation. The dislodged particles then roll down independently
of each other without triggering a largescale avalanche. Steady state demands
that they come to rest and "stick" to the surface at the same rate R as they are
being supplied.
If 3x h = 0, that is to say, if the slope matches precisely the critical angle,
firstorder derivatives vanish and it becomes necessary to push the analysis to
secondorder by including h. We may anticipate that the system, controlled
through the bias of the operator f, will evolve so as to erode any roughness of
the free surface. Put another way, the local curvature will tend to be reduced,
here again in proportion to R.
a;
Given these constraints, all based on what we already know about avalanches, we
are now in a position to look for a plausible form for f. The simplest possible way
for f to comply with our four requirements is to be of the form
(411)
where y and K are two positive constants. We might point out that this expression
depends on hand R to first order, which is a distinct advantage from the standpoint
of analytical calculations, as we shall see shortly. We also note that the operator f
is now proportional to R. That is a major difference from the equations considered
earlier. Previously, the flux of rolling species experienced amplification not as a
result of a domino effect, as is the case here, but solely because of the negative
149
curvature of the function a<t>lax in the vicinity of its equilibrium value [see (49)].
In the previous stickslip model, the driving force determining the flux was an angular deviation. Here, the flux is determined both by the deviation from the critical
slope and by the quantity of material already in motion.
In order to account for material at rest, all we need to do is write h =  r in the
previous equation. Thus, the height of the pile obeys the equation
(412)
so that the total quantity of matter h + R is conserved locally. We further note that
(411) does reproduce the metastability of a pile on the verge of avalanching which,
as we know, is a defining characteristic of granular piles. We can indeed convince
ourselves that, in the absence of any flow, the equation does not spontaneously
generate any avalanche. Thus the surface appears as if frozen in a static equilibrium.
If, on the other hand, we start in such a situation and introduce a perturbation on
the surface of the pile, at least a few grains will start rolling down the slope. The
slide will take place in a finite amount of time and lead to a stable state. That too
can be recognized as one of the fundamental characteristics of avalanches.
It is not essential here to examine further the details of this formalism. Suffice
it to say that it can be made use of either analytically (in the simplest cases) or
numerically (in the tougher ones). The interested reader is encouraged to consult the
basic literature about the theory [84][87]. We will restrict ourselves to discussing
two applications of this model.
time t
time t+dt
FIGURE 94. The model predicts that a perturbation created anywhere on the slope works
its way back up. In the process, it gets attenuated and widened because of diffusion effects.
150
efficiency of this process as measured by ah / at, and hence its propagation velocity,
depends linearly on y and R o, as is evident in (412). Things are far less intuitively
obvious, on the other hand, in the case of a mound. We would be more inclined
to think that the lower part of the mound would collapse. Instead, because it is
based on a set of coupled equations, the model suggests that the material located
above the mound can somehow "sense" the perturbation below and undergo local
minislides that rearrange material in such a way as to generate a new mound that
appears to propagate toward the top. Perhaps this unexpected property is a generic
signature of this type of model.
Simulation ofAvalanches
The ultimate test of the model is, of course, whether it can correctly simulate an
avalanche. To resolve that question, we consider a pile in a metastable state such
that axh(x, 0) = SoC <0). That can be easily accomplished by increasingat least
mentallythe inclination of the pile, initially assumed to be at the angle of repose,
until it approaches the angle of movement. Barring any additional perturbation, the
new state is metastable, and nothing happens (as long as the angle of movement
is not exceeded). In this situation, the upper part of the pile is at an angle greater
than the angle of repose.
What happens now if we add a few grains near the bottom of the pile? Those who
have correctly grasped the essence of the correlations associated with the coupled
equations will readily guess that the result is going to depend quite critically on
the value chosen for So.
To clarify this point, we need to consider in some detail how the process unfolds. Assume that at time t = 0 a localized perturbation is created at point x in
the form of a few moving grains, which may be written as R (x', 0) = flo (x'  x).
Two opposite effects combine to produce the final outcome. First, as we have just
seen, the small perturbation propagates up the slope toward the top of the pile.
That is directly implied by (410) in which we neglect the coupling term r. The
ascending perturbation decreases with time as R(x, t) ex exp( v 2 t /4D), as is
characteristic of all diffusion processes. On the other handand that is the second
competing phenomenonthe grains rolling down at abscissa x will dislodge other
grains initially at rest. Equation (410) shows that, when v = D = 0 (velocity and
diffusion are both zero), the density of the rolling species grows exponentially as
R(x, t) ex exp(y Sot). Stated in words, the initial perturbation sets off two competing phenomena that both grow exponentially. The situation is critical in the sense
that either one can win out. We find that if So > Sd ~ v 2 / y D, the number of grains
that are dislodged grows faster than the number of grains that make it back up the
slope. A runaway eventrunaway because governed by an exponentialis triggered in the form of an avalanche. 9 In the other case (corresponding to So < Sd),
the diffusive term dominates; it attenuates the effect of the perturbation and the
system remains at rest.
9We have chosen to preserve the notation used by the authors of this model. This should help those
inspired to read the original articles. The correspondence with the angles em and e,., used earlier in
this chapter, as well as their difference 8 = em  e,., is obvious. Here Sc' (== e,.') is set equal to 0, and
Sd(== em) becomes equivalent to what we called the angle of movement.
151
Starting from a simple consideration of an angle of repose Sc', the present model
leads quite naturally to the definition of a critical angle Sd which corresponds to our
familiar angle of movement em. It is the angle required to trigger an avalanche. 10
We might add that numerical solutions of the applicable equations confirm these
interpretations.
The model can, of course, be applied to a variety of situations, from filling
a silo to far more complex scenarios. We must keep in mind, however, that the
formalism accords a great deal of importance to what we have referred to as a
diffusive process, which in actuality boils down to a mechanism of attenuation and
upward propagation competing with the rate at which grains are knocked loose.
De Gennes has pointed out that such a diffusive term was unsatisfactory on at least
two grounds [86]:
The first has to do with the range of this diffusion. For any variation of the
density of a rolling species R of size L, the diffusive term corresponds to a
perturbation of the order of D / lv, which is itself of the order of d/ L (where d
is the mean diameter of the grains) relative to the convection term. In practice,
this means that the diffusive term must be rather small for smallsize slides.
The second objection is that the physical interpretation of the diffusive term
is a sort of Brownian motion that enables some of the particles to make their
way back up the slope. The problem is that such an interpretation is plausible
on a very small spatial scale only, of the order of D / v ex d, in which case a
continuous description looses its meaning.
This opens to door to the possibility of modifying the model, while still preserving
the spirit of the initial equations.
De Gennes's Modified Model: Rotating Drum Experiment
The previous model suggests that rolling species respond to a small disturbance by
convection, amplification, and diffusion, as long as the slope is inclined by more
than the angle of repose. The magnitude R of the flow at a given location x is given
at some subsequent time t by an equation of the form
Rex
exp[y(e  e
r) 
:~}.
The basic idea behind the modified approach is to keep the general form of (410)
and (412) but discard the diffusive term. The objective is to replace the original
function by one that serves the same purpose but is more realistic. A useful clue
is that the system is nearly always subjected to an external source of noise, which
depends on what might be called ambient noise rather than temperature. On that
basis, we can develop a model of "thermal" detrapping that remains consistent with
the spirit of the original approach. At the onset of an avalanche at angle e > e,,
the flux of rolling species was given in the previous model by an equation of
lOWe also note, incidentally, that the present model leads to a process analogous to a firstorder
phase transition, consistent with the conclusions arrived at in Section 4.2.1. This is, of course, in
contradiction with the predictions of the SOC model.
152
the form
R(x, t) = Ri(x
+ vt, 0) exp[y(e 
ee)t].
We retain the same functional form here, except that R i (x, t) now represents a
source of ambient noise due to the mechanical agitation of the system, and ee is
an angle corresponding to equilibrium. This can be justified by picturing particles
trapped in a potential well U that depends on the angle e through a relation of the
type U = mg dj(e), where j(e) is adecreasingfunctionofe which must obviously
vanish when e = Jr 12. The mechanical noise is described in terms of an effective
temperature Teff. The source of rolling species is then of the formd exp( U 1kTeff),
which varies as d exp[a(e  em)], where em is defined as the angle at which the
detrapping potential equals the effective "thermal" energy associated with the
ambient noise. Under these conditions, we have U ~ kTeff and = mgdl kTeff,
and it can be verified that nucleation is suppressed as long as e < em. It only
begins when e = em' We should point out that in this model the angle em depends
critically on noise.
In short, the essence of this model is to give a preeminent role to the ambient
noise. This noise is responsible for detrapping the rolling species and is instrumental in determining the angle of movement.
De Gennes has shown that this approach accounts for both types of movement
identified in Section 4.Ia continuous flow (class A) at high velocities, and an
intermittent one (class B) at low velocities. In other words, there are always some
grains in motion in a classA movement and, in particular, there exists a steadystate solution for which the profile h(x, t) is constant, which is perfectly consistent
with experimental observations.
Intermittent or continuous avalanches in a rotating drum can be described by
the following system of equations
ah
ah
=yR+wx,
at
ax
aR
aR
ah
=v+yR,
at
ax
ax
where the flow is assumed to take place from left to right. Note the presence of the
wx term, describing the effect of the rotation. How to solve these equations is left
as an exercise. In case of difficulties, the reader is referred to the original article
[86]. We conclude this chapter by seeking the steadystate solution pertaining to a
classA regime in a rotating drum of diameter 2L. The geometry is illustrated in
Figure 95.
The relevant boundary conditions are R = 0 at x = L. By further imposing
the restriction that the timederivatives vanish throughout, we find steadystate
solutions of the form
W
Rss(x) = (L  x )
2v
(a)
153
(b)
L+_:
L
+L
FIGURE 95. Geometries used in conjunction with de Gennes's amended model based on
a system of coupled differential equations (see text).
and
These solutions predict a slope with a parabolic correction, illustrated in Figure 95,
although the deviation from a straight line remains small.
It is useful to remember that solutions of type A (continuous flow) and type B
(intermittent slides) crop up systematically in granular flow problems. They reflect
the metastable nature of piles whose slope evolves between the angles em and e,..
As a further example, we may want to determine the steadystate solutions when
a silo of diameter 2L is being filled from the top by a flux Q of particles falling
through a hopper. We assume that the flow is continuous and free of arch effects.
The orifice of the hopper is at x = 0, and the lateral walls of the silo are at x = L.
Using the same conditions as in the previous example, we find that the steadystate
solutions are
Q ( 1Rss(x) = 2v
LIX I)
and
ess 
e = (ah)
ax

ss
y L
Ixl
Here again, the deviation of the steadystate angle from the angle of repose is
small, being only of the order of d / L.
This modified model allows a variety of situations to be analyzed, although,
at the present time, it does have limitations related to the fact that it treats a
granular as a continuous medium. One thing it cannot do, for instance, is account
for phenomena of fracturing and fragmentation (see Section 3.2.4) which, as we
now know, are likely to occur in piles subjected to any sudden change, such as
when the door of a filled silo is opened abruptly.
5
Mixing and Segregation
5.1.
Introduction
liquids are made of particles subject to Brownian motion, thermal agitation alone produces
mixing. As pointed out in Chapter I, the Brownian motion of granular particles is entirely negligible.
Another source of energy is then required to achieve mixing. Vibration is a logical candidate, even
though the final result is often the exact opposite of what is being sought!
5.1
Introduction
155
:.:
Wavelength A
FIGURE 96. Oyama's horizontal drum rotates around its axis. An initially homogenous
granular mixture segregates in vertical bands with a spatial periodicity A.
experiments by shaking a mixture of different grains (wheat, corn, rice, salt, are
all good candidates) in a test tube. 3 Likewise, farmers know very well that tilling
fields causes large rocks, in seemingly endless supply, to work their way up to
the surface. Peasants in India take advantage of the segregation properties of dry
granulars by shaking their harvest of chickpeas in baskets in order to separate
them from other materials. When Brazil nuts, mixed with other smaller nuts, are
transported in pickup trucks over the rough back roads of South America, they
invariably end up on top of the load. 4 In short, the phenomenon of segregation in
dry granulars is universally recognized, even though it may not be well understood.
The first recorded observation of segregation in a threedimensional medium
was described by Oyama in 1939.
5.1.1.
156
knows why. Spectacular as it may be, this experiment remains poorly understood. It
continues to be the object of numerous studies, some of which rely on sophisticated
techniques such as nuclear magnetic resonance (see Section 3.2.3) [90].
Be that as it may, the experiment underscores the amazing efficacy of segregation
by rotation or, to be more precise, by shearing. As mentioned earlier, segregation is
a very general property of mixtures in which particles are in motion relative to one
other. A careful observation of the mechanisms potentially involved has led some
authors to distinguish at least two different modes of granular segregation. They are:
5.1.2.
5.1
Introduction
157
M and m, and radii Rand r (with M > m and R > r), stacked on top of each
other. If the large sphere is on top, the potential energy E p of the stack can be
written as
E p = mgr
+ Mg(R + 2r),
where the horizontal base on which the edifice rests is taken as reference. Both
spheres are assumed to be made of the same material and, therefore, have the same
volumetric density. This additional piece of information leads to an expression for
the potential energy in terms of sizes only
E p ocr 4
+ R 4 + 2rR 3 .
Interestingly, this expression is not symmetrical in rand R, which means that the
energy depends on whether the large or the small sphere is on top. Not surprisingly,
the configuration with the small sphere on top is energetically favorable. We may
find a degree of solace in this result. Yet, our intuition clearly fails us when it comes
to the phenomenon of segregation by size which, as already indicated, invariably
drives the largesthence, heaviestparticles toward the top.
Superposition of StacksTwo Compact Stacks
We next consider the stack shown in Figure 97. The structure, depicted here in
two dimensions, is made of two kinds of marble made of the same material. The
marbles differ only by their diameter and are arranged in two zones stacked in a
compact triangular lattice (which gives it maximum compactness, in accordance
with Section 3.1.3). They are placed in a cylindrical container of crosssectional
area S. We designate by V and v, respectively, the volumes of each of the zones
occupied by marbles of radii Rand r. Assuming that the largest marbles occupy
FIGURE 97. Two types of particles A and B of different sizes are stacked in superposed
layers. In the absence of any structural defect, AlB and BIA configurations are equivalent
from an energy point of view.
158
16x16
Planar square lattice
(b)
16x16
Triangular lattice
FIGURE 98. Two stacking modes for twodimensional structures. The triangular configuration (b), being as compact as possible, has a lower potential energy than configuration
(a). It is therefore energetically favorable and more stable.
the upper part of the container, the potential energy of the entire pile is given
by
This time, which type of marble is on top has no bearing from an energy point of
view, and there is no reason to anticipate any segregation, at least as long as interface effects are absent. This result is obvious, inasmuch as two layers of identical
density are always in equilibrium.
Without the benefit of a more detailed analysis, we may begin to suspect that
structural defects in a lattice made of particles of different sizes may have something to do with the tendency of large particles to work their way back up. To
explore this possibility, we return to two configurations already discussed in Section 3.1.3. The relevant structures are reproduced in Figure 98.
An elementary analysis of the two stacks shows that the triangular lattice isbased on simple energy argumentsmore probable and stable than its square
counterpart. As we have seen, the compact triangular lattice is the only one to
exhibit characteristics consistent with Reynolds's dilatancy principle. The reader
will recall that a compact stack subjected to any distortion can only respond by
expanding, which increases its potential energy. If we realize that distorting a
compact triangular lattice necessarily entails the creation of defects in the stack,
we may legitimately inquire whether such defects may tend to concentrate in the
lower or the upper portion of the pile. In particular, what can energy arguments
tell us about this question?
5.1
Introduction
159
Faults
Compacted
Zone
FIGURE 99. Defects are created when a large disk is inserted into a twodimensional stack.
The photograph was obtained by backilluminating a real stack. The lower part of the stack
remains compacted, while the triangular symmetry of the upper part is greatly disturbed by
the introduction of the large disk (after [93]).
energy is denoted E pu if the defects are in the upper region, and E pi otherwise. If
the defects involve an increase in volume dv, the calculation indicates that
E pl
E pu ex vdv > 0,
which implies that defects near the top are indeed energetically preferred. Another
way to express the same result is that, in a cylindrical container, the potential
energy is minimum when the less dense material is on top.
An equivalent twodimensional experimental configuration, similar to those we
have discussed in Chapter 3, is easy to implement with suitable provisions for
agitation. s All we need to do is to introduce into the container a single larger disk
of the same volumetric density as the rest of the pile, and we can readily witness
the phenomenon just described. As demonstrated in Figure 99, the intruder causes
local distortions in the lattice by creating defects which tend to migrate to the upper
part of the structure. 6
When viewed in this context, the process of segregation by size emerges as one
of the consequences of the dilatancy principle. The introduction of an intruder
necessarily causes a local distortion of the lattice manifesting itself in a local
expansion. The expanded and, therefore, less dense portion of the pile tends to
5The meaning of the phrase "suitable agitation" deserves to be thought out carefully. It may be said
that shaking (or vibrating) a complex granular edifice enables us to explore perhaps not all but at least
many of the possible configurations of a pile. Several simulation approaches (notably the Monte Carlo
technique, to be discussed in Chapter 6) capitalize on this observation by minimizing the energy after
each perturbational event through various relaxation processes.
6We will often use the term "intruder" to refer to a particle whose size or other properties differ from
those of the "normal" sea of particles which, for the sake of simplicity, we will consider uniform.
160
move toward the top, dragging the intruder along with it. This would suggest that
the shape of the intruder may playa crucial role in the process of segregation,
depending on how readily it may fit in the surrounding lattice.
There are too many potential objections to accept such a crude explanation at
face value. In an attempt to get a better grasp of the phenomenon, we will have to
refine our understanding of the successive relaxation states of a pile undergoing
segregation. Until then, the one idea that should stand out and be remembered in
what follows is that size segregation implies defects created by the intruder in its
environment.
5.2.1.
Before proposing a model that can account for the behavior of a vibrated twodimensional pile composed of a single intruder in an otherwise homogeneous
environment, it is useful to pause and consider Figure 100, as it highlights a
number of remarkable characteristics of such structures.
For starters, we note that both computergenerated and reallife stacks feature
stacking faults in the form of distortions and dislocations in the upper part of the
structure. This is fully consistent with our previous elementary energy considerations. In addition, we see that the perturbation caused by an intruder whose
size is different from that of the particles composing the matrix develops in an
essentially triangular pattern. We might think that this pattern is simply a consequence of the symmetry characteristic of the type of twodimensional lattice
considered here. However, as we shall soon see, this property is not restricted to
twodimensional lattices. Computers can serve as an inexhaustible source of synthetic threedimensional piles, whose geometry can be easily studied [91]. It turns
out that, in this case too, defects are generated in the upper part of the pile, and their
geometry is quite similar to what is observed in two dimensions. That provides confidence for extending the arguments we are about to develop to threedimensional
piles.
Lastlyand this is crucial for the model discussed nextwe can see that a
largesize intruder, such as in Figures lOO(a) and (c), does not necessarily have to
(a)
(b)
(e)
(d)
161
FIGURE 100. Different configurations of a twodimensional inhomogeneous stack. Diagrams (a) and (b) are computergenerated structures, while (c) and (d) are photographs of
actual reallife stacks. Note that a large intruder can be in a stable position without having
to be in contact with particles immediately below (arch effect) (after [93]).
rest on lattice lines associated with the matrix. Instead, they can be propped up
above such lines by lateral particles, somewhat like the arch of a cathedral rests on
stones that transmit its weight to side columns. Pursuing this metaphor, we refer
to this phenomenon as the arch effect or vault effect.
When trying to model the dynamic properties of such a system, we need to
inventory all possible stable positions of the intruder. Stability occurs under one
of two conditions:
The intruder rests on a lattice line defined by the ordered arrangement of
spheres constituting the environment.
The intruder is kept above a lattice line because it sticks at two points marked
by arrows in Figure 10 1. The line formed by joining these two points traverses
the intruder below its center of gravity. Should that process fail, the intruder
drops back down to the next lowest lattice line.
Modeling this situation involves solving a topological problem, which goes something as follows: The intruder is raised a very small step at a time, as depicted
in Figure 102, and the system is left to reorganize itself by relaxing around the
intruder. The new arrangement is examined to determine if it is stable or unstable. If it is unstable, the intruder is raised some more by a tiny amount, and the
stability is examined again. The process makes it possible to find eventually all
stable configurations.
162
FIGURE 101. Simulation of the equilibrium of an intruder by the arch effect. On the left
is a photograph of an actual stack of metal spheres containing a foreign disk (after [93]).
TwoDimensional Model
The relevant geometry is illustrated in Figure 103(a). Let cD = R/ r be the ratio
(> 1) of the radii of the particles. As indicated earlier, the effective part of the pile
is confined within walls Bl (T) and B 2 (T), which in the present case intersect at an
angle of 60. Our purpose is to determine all stable positions of the intruder as it is
being raised in a stepbystep fashion up the pile. The first thing we notice is that,
because of the geometry of the structure, we do not have to explore a height greater
than one period of the structure, which is given by 8 = 2r~, or 3.46 times the
radius of the dominant particles. Let h be the altitude of the intruder's center, as
shown in Figure 103(a). We keep track of each particle by its row (index i) and
column (index j). Simple geometry considerations show that the stable positions
FIGURE l02. Method for seeking the equilibrium positions in a stack by exploring different
possible configurations.
(a)
163
(b)
FIGURE 103. Diagrams of two and threedimensional stacks used to develop a model of
segregation by the arch effect.
of the intruder, i.e., when it rests on a lattice line, are obtained when
where the superscript S indicates that the position is stable. The index k runs
from [(_l)i+l + 1]/2 to Int((i + 1) /2), the operatorInt(m) designating the nearest
truncated integer of the argument m.
Next we look for stable positions of the intruder via vault effects, that is to say,
when it rests on two particles located below its center of gravity. We start from
a situation where the intruder is in contact with the two lateral walls B 1(T) and
B2(T), which occurs when h vI = 2r. When the intruder is raised gradually, it will
find a new stable position when two small particles can just squeeze below its
center of gravity. This happens when h v2 = (R + 8),J3 ;:;" r,J3(cD + 2), where 8
is the space between the intruder and the lateral walls.? The fraction S of stable
vault configurations over a period e is given by
S=l
h v2
h v1
2
,J3
=~cD;:;"O.077cD.
2y3
164
 
~ 20
'"
OJ
.0
'0 15
d
E.
_.
~o
c.
OJ
.!9
5
0
t
:0
(f)


I
~ 10 1o
lih
<I>
13
11
9
7
5
3
1.2
'*'c
= 3 + .)2
3.)2
i':::i
2 78
.,
5.2.2.
The first quantitative experiments aimed at studying the phenomenon of segregation by size were begun only in the late 1970s. For lack of a better technique,
the experimental approach consisted simply in measuring the time taken by an
intruder placed at the bottom of a container filled with a granular material to make
its way back up to the surface when subjected to vibrations [94]. This method was
much too primitive to reveal the subtleties of granular segregation which we are
about to discuss.
165
(b)
(a)
2D
1
FIGURE 105. Diagram of experimental methods used to study the segregation by shaking
of a large particle in either (a) twodimensions or (b) threedimensions. In (a), a large
marked intruder is immersed in a population that includes a few tracers (black particles),
whose progression is tracked by image processing techniques. In (b), the movement of
tracer particles in their environment is monitored by direct visual observation.
Modem techniques take advantage of the possibilities offered by image processing (see Section 3.2.3), as well as nuclear magnetic resonance, which makes it
possible to monitor what goes on inside a threedimensional opaque system. More
direct methods rely on suitably prepared samples containing tracer particles. The
technique, schematically illustrated in Figure 105, has been used to study both
two and threedimensional configurations [93], [95].
A typical implementation of the approach is shown in the photograph of Figure 106. The apparatus uses two small cameras. The first, placed next to a vibrating
Tracking video camera
Marked intruder
Vibration source
Camera monitoring
the vibration
amplitude
166
(a)
(b)
_
"
.......... ..
..
....
.. .. .. ..
. ..
.. .. .. .. .. .. .. ..
10
It.
...
..
4
:'I%i\\~H~~1ii~~!~~
..
plate, measures accurately the amplitude of the vibration. The second is mounted
on a translation stage and connected to the image processing electronics. With
thresholding and the image addition techniques described in Section 3.2.3, it is
possible to track one or more particles during the course of an experiment. 8 Alternatively, by moving the camera horizontally at constant speed, the vertical position
h(t) of the intruder can be displayed directly on the monitor's screen as a function
of time. We will see several examples of this technique below.
Experiments on Continuous and Intermittent Ascent
An apparatus similar to the one just described has been used to study the different
modes of ascent of cylindrical disks of various sizes immersed in a uniform granular
medium. Experiments show that a small intruder "sees" the discontinuities of
the granular environment, which is to be expected. A large disk, on the other
hand, rises smoothly without any pauses, which is not at all intuitive. Disks of
intermediate sizes exhibit continuous rises interspersed by plateaus, in agreement
with the predictions of the ascent diagram.
The model does seem to predict fairly well whether an intruder will ascend continuously (Figure 107(a)) or intermittently (Figure 107(b)). On the other hand, it
tells us nothing about what causes the ascent in the first place. A careful observation
of the vibrated container offers some hints as to what drives the intruder toward
the top.
With the aid of a suitable stroboscopic lighting system, it is easy to observe
fractures with various lifetimes appearing in the immediate vicinity of the intruder,
as shown in Figure 108.
S"Thresholding" consists in defining a particular level of brightness. Anything brighter than that level
is considered white (or I), and anything darker becomes black (or 0).
167
Longlived
fractu res
FIGURE 108. Photograph showing the relatively random appearance of fractures around
the intruder (after [93]).
Referring to the ascent diagram of Figure 104, we can make the following observations:
A small intruder (one that is characterized by <P < 12.9) requires a large relative displacement 8 between itself and its environment in order to overcome
each step in the ascent diagram. Accordingly, we may anticipate a sizable vibration threshold to initiate this kind of motion. Experiments indeed confirm
this .
A large intruder (one with <P> 12.9) proceeds much more readily while remaining continuously in an arch configuration. We can predictand it is
experimentally verifiedthat the amplitude threshold is much lower than in
the previous case. In other words, a larger intruder moves up far more easily
than a small one.
As we will see, this is further confirmed by diagrams of upward speed as a function
of the ratio <P = R / r, at least on condition that we avoid the convection regime.
In this respect, experiments that make it possible to observe simultaneously the
upward drift of the intruder and the relative motions of the surrounding particles
are highly revealing.
Finally, we might point out that the phenomena of continuous and discontinuous ascent just discussed have never been observed up to now in three dimensions,
even though simulations predict them. That does not necessarily mean that they
do not exist.
168
lo}
(b)
(e)
E
~
169
Q)
l.l
5
::J
til
.8 10
Q)
l.l
rn
1il
15
200
100
200
Number of impulses
FIGURE 110. Successive positions of different intruders plotted against the number of
impulsions applied to the container. The ratio <t> is equal to 9.5 (crosses), 6 (circles), and
1 (squares), respectively. Note that large intruders remain trapped at the surface after completing their climb, while small ones (square data points) are dragged back down to the
bottom by convection (after [95]).
twodimensional pile vibrated vertically with varying accelerations. Figure III (a)
reveals a process of convection identical to the one just discussed in three dimensions. By contrast, Figure 111 (b) is indicative of an arch effect squeezing the intruder. As emphasized above, this phenomenon depends on the intruder's size. The
velocity of the upward movement is presumably a function of the diameter ratio <P.
This was entirely confirmed by a set of experiments recording the altitude h(t) of
intruders of various sizes as a function of time for a given acceleration and container
configuration. The results of these experiments are reproduced in Figure 112.
convection
(a)
(b)
FIGURE 111. Diagram (a) shows a typical segregation mechanism by convection observed
with a relatively strong excitation (f' = 1.6). Diagram (b) was obtained with a weaker
acceleration (f' = 1.2). It reveals an arch effect mechanism that displaces markers laterally
below the intruder. This latter mode of segregation acts differently on intruders of different
sizes (after [93]).
170
I /<I>=12~9
.",.
/<1>10.5
,"
<1>=9
i) I
<1>=16.3
10
.
()
'"
.....
/ ..
Q)
"'C
::l
;;:::;
t1r.... .lntermittent
<1>=5.3
. I
Events
<1>=3
2
0
0
10
20
30
40
50
60
Time (minutes)
FIGURE 112. Positions h(t) of intruders of various sizes immersed in a bath of particles
1.5 mm in diameter. The larger the size of the intruder, the faster its ascent (after [93]).
Here the acceleration was held constant at r = 1.25. The various curves h(t)
have been displaced along the horizontal axis to make it easier to tell them apart.
The results clearly demonstrate a process of size segregation consistent with the
arch effect model discussed earlier. Smallsize intruders (<l> < 12.9) experience a
discontinuous ascending motion marked by a series of steps and plateaus. The
greater the size of the intruder, the less discontinuous its upward movement, in
agreement with the model. All of this is consistent with the ascent diagram shown
in Figure 104. Furthermore, with small enough intruders (characterized by <l> < 3),
no ascending movement occurs at all, at least not for this particular acceleration
and over the duration of the experiment (about 1 hour). The results, summarized
in Figure 113, prove that there is indeed a threshold diameter below which any
14
12
'E
......
E 10
.s
~
'u
0
~
C
OJ
Ul
4
2
0
10
15
20
5.3
Segregation by Shearing
171
ascending motion is inhibited. The existence of the threshold diameter makes sense
on the basis of the ascent diagram in Figure 104. We noted at the time that, for
weak accelerations, the fluctuations 8h of the intruder's positions can be smaller
than what is required to bridge the quantum jumps between successive steps. This
is consistent with the nonlinear behavior observed in Figure 113.
Finally, the size dependence of the ascending velocity of intruders may find useful industrial applications, since it provides a means to separate particles immersed
in a granular medium. We can envision the possibility of "filtering out" components
of different sizes by a proper choice of the acceleration imparted to the cell.
5.3.
Segregation by Shearing
5.3.1.
As pointed out in Chapter 4, a rotating drum is a convenient tool to study the flow
properties of sheets of granular materials, and we will once again resort to this
familiar device depicted in Figure 114 [98], [99].10 This cylindrical drum with a
horizontal axis is similar to the one that served us so well in Chapter 4 for studying
avalanches. To take advantage of the image processing techniques described in
Section 3.2.3, we want to select objects with a high visual contrast. For instance,
we may use a matrix of white spheres and a single, black tracer particle, whose
movements we wish to track. We learned in Chapter 4 that when the drum is rotated
slowly about its axis, a series of more or less periodic avalanches of different sizes
is set off. The avalanches are confined to a narrow layer near the free surface of
the material, which constitutes what might be called a liquid phase. The rest of the
structure is in a compacted state that can be thought of as a solid phase. The solid
lOPigure 114 is a drawing, not a photograph taken during an actual experiment. The distinction is
significant, inasmuch as it is not obvious that the configuration depicted here can truly occur in a
real rotating drum. This particular structure was arranged so as to ensure the local equilibrium of
every disk during the stacking process. There is a fundamental difference between that and the global
stability ofthe entire edifice. An avalanche boils down to disrupting the local stability, which in turn
affects the global stability of a granular pile.
172
Rotation
Witness
particle
Roughened
inner wall
portion of the pile remains effectively bound to the cylinder. As the tracer particle
is being dragged along by an avalanche to the bottom of the pile, it is reinserted
and buried into the solid phase. The rotation of the drum then causes it to rise back
up toward the free surface.
It is important to have a clear picture of the sequence of events. The evolution
of the intruder is illustrated in Figure 115, which shows the outcome of a real
experiment. The experiment reveals that the process of segregation toward the
center or edges of the drum occurs in the flowing region at the surface of the
pile. Based on what we know about avalanches, particularly about their statistical
FIGURE 115. Velocity diagram of a tracer particle dragged in the liquid phase and reinserted in the solid phase. In the present case, R < r. The intruder is seen to converge toward
the center (after [90]).
5.3
Segregation by Shearing
173
picked out and all other particles are discarded. This produces a map with
a single black pixel.
(2) The brightness of the picture is then divided into 256, and the result is stored
in a buffer (a block of memory in electronics parlance).
(3) A new image is recorded at a later time t2. Its brightness is also divided into
256. This second image is added to the first one in the buffer. At this point,
the buffer will contain either one pixel with a brightness of 2/256 if the
intruder has not moved, or two separate pixels with a brightness of 1/256
if it has moved, in a sea of pixels with zero brightness corresponding to all
the other sites that have yet to be visited.
(4) The cycle is then repeated from the top, for a total of up to 256 times.
The lower part of Figure 116 shows the results of this type of exercise. The data is
analyzed by slicing time in discrete frames, denoted by an index i, during which
one or more avalanches may occur. If the intruder is inserted in the flowing sheet at
the radial coordinate ri at time ti, what will its coordinate ri+ 1 be in the very next
frame attime ti+ I? The corresponding correlation diagram ri+l = f (rd, referred to
as the map offirst iteration, is shown in the upper part of Figure 116. The diagram
is drawn for a region encompassing a 40 wide sector S, bounded on the short side
by a radius R 1 corresponding to the exclusion zone of the particles in flow, and on
the long side by the inner radius R 2 of the rotating cylinder.
IIThis is an important observation. It means that segregation by shearing does not result merely from
reinsertion of an intruder during successive avalanches. A more plausible picture is for the intruder
to undergo segregation by size during an avalanche and be transported across a distance that is
determined by its relative size.
174
(a)
ri+1J
ri+l~
"~;.'..
:/;:I~':.:;~J
if.'
R/r
(e)
(b)
0.,,..,,0
ri
rj
=2/3
R/r
=1
R/r
=4/3
FIGURE 116. The bottom part of this figure is the superposition of 12,000 snapshots taken
every 5 seconds. It shows the spatial distribution of the sites visited by particles whose
diameter is (a) 1 mm, (b) 1.5 mm, and (c) 2 mm, placed in a bath of particles of diameter
1.5 mm. The upper part of the figure is discussed in the text. The cylinder had a diameter
of 160 mm, and its rotation speed was typically 2 degrees/s (after [99]).
The results show that, depending on the size of the intruder relative to the
majority particles, it tends to converge toward the center, to explore all the available
space, or to take refuge near the periphery. A uniformly gray area indicates near
perfect mixing.
We note first that the correlation diagrams ri+ 1 = f (ri) are all symmetrical with
respect to the principal diagonal. This observation is not insignificant. It means that
a true steady state is reached as early as the first iteration. If that were not the case,
we would see over the course of measurements a flight of the intruder either toward
large radii or toward small ones. That would translate into an accumulation of data
points either above or below the principal diagonal in the correlation diagrams. It
also means that the steady state can be described in terms of a relation of the type
IT (rj+rI rj)
IT(r;jri+1)
where P (r) is the probability of finding the intruder in region S, and IT (rj +1/ rj)
is the conditional probability of finding the particle at rj+1 when it is at rj during
the preceding frame. These results can be normalized by writing
per)
R2
Rr
IT (r /r')P(r') dr
and
R2
Rl
per) dr = 1.
5.3
Segregation by Shearing
175
5.0  ,                 ,
0'08~
a 0.04
4.0
~......
.....
0.00
0.04
0.08
30
.
I::"
CD
.........
Q 2.0
''''''
1.0
0.0
\~
'";1.
, .. ..
0.P
::.::::&< "
ollo
,."
'"
t,r''::..,Ir='.r1r=!
20.0
40.0
r (mm)
60.0
80.0
FIGURE 117. Probability ofpresence P (r) inregion S (see text). The triangles, squares, and
parallelograms correspond to particles ofdiameter 1, 1.5, and 2 mm in diameter, respectively.
They are immersed in a bath of particles whose diameter is 1.5 mm. The inset shows the
dependence of the parameter a (the inverse of the localization length) on the diameter ratio
CD (after [99]).
The experimental results have been averaged over a distance /:).r equal to three
particle diameters.
Figure 116 shows convincingly that the region of space covered by the intruder
depends on whether its diameter is larger or smaller than that of the majority
particles. The same information is conveyed by a graph displaying the probability
P (r), as shown in Figure 117. To a first approximation, the probability P (r) can
be described by a function of the type per) ex exp(ar), where a represents the
inverse of a length characteristic of the segregation process. Evidently, a changes
sign when the ratio <1> goes through unity. A positive sign indicates that the intruder
flees toward the periphery, a negative sign that it tends to converge toward the center.
Without pursuing the analysis of experimental results any further, we may note
that our avalanche model discussed in Chapter 4 is not completely accurate. In
particular, we had claimed that an avalanche, being of random size, should reinsert
a particle anywhere along a flowing sheet. Actually, as revealed by the maps of
first iteration or by tracking an individual marked particle of congruent size, both
the center and the periphery of the cylinder are attractors for the dynamics of
the system. If a particle is introduced near the periphery of the cylinder, it will
tend to stay there. Likewise, a particle introduced near the center will tend not to
wander off vary far. This implies a degree of correlation between trapping events
taking place within avalanches, somewhat consistent with the model developed in
Section 4.2.2. We might conclude that the dynamics of size segregation is governed
by two attractors, one at the center, and the other on the periphery of the cylinder.
According to this hypothesis, segregation in a rotating cylinder could be construed
176
as a mechanism favoring one attractor over the other. Such situations often involve
the phenomenon of bistability. 12
5.3.2.
We now consider a mixture of two distinct populations of particles. The first question that comes to mind is whether the observations just made in the case of a single
intruder in a sea of identical particles can be extrapolated to predict the behavior
of our mixture of two materials. In other words, can the process of separation in a
binary mixture be treated as a succession of independent steps affecting individual
particles, such as we were dealing with in the preceding section? The answer is
not at all obvious, as we are about to find out. We will show that the kinetics and
geometry of the segregation process depends on the shape of an incipient cluster
growing with a fractal structure.
We now define the problem more rigorously. Let N A be the number of particles
of diameter d A, which is to be mixed with, or segregated from, a number N B
of particles of diameter dB. As usual, we define a ratio <P = dA/d B. The entire
system is placed in a twodimensional rotating cylinder of the type used earlier,
which amounts to a simplified twodimensional version of Oyama's drum. The
experiment starts by mixing particles A as completely and randomly as possible
with particles B. These may be, for instance, black disks mixed in with slightly
larger white disks, as depicted in Figure 118.
The experiment reveals that after just a few turns of the cylinder, the smaller
particles have gathered in the central part of the drum. The connected mass created this way is called the reference mass. 13 Its surface area, reached in principle
12 A
bistable system has two equilibrium states. It can switch from one to the other under the effect of
some external perturbation.
13 A
mass is said to be "connected" when its particles actually touch each other.
5.3
Segregation by Shearing
177
after an infinite timein reality, this time is quite short, as we will seeis denoted Soo.
We proceed to highlight a few general ideas relevant to this situation. A more
detailed analysis can be found in [90]. In order to characterize the state of the
mixture, it is useful to introduce a parameter describing its degree of order. Let
S Ct) be the surface area of type A disks absorbed in the reference mass at time t. It
is clear that S Ct) has to be smaller than Soo. The degree of segregation is quantified
by a parameter aCt) defined as
aCt)
Set)
==.
Soo
At this point, it is natural to introduce an ordering parameter poet) that can vary
between 0 (for a completely random and homogenous mixture) and 1 (for a fully
developed reference mass). This parameter is defined in terms of aCt) by
P (t) _ aCt)  a(O)
a
1  a(O) .
1.0
C....L
0.8
"*E~
0.6
....Q)
"E
0.4
a..
O;~o
rn
I1l
.Q
0.2
0.0
0 00
Time constant
'r
Time
200
400
600
Time t (seconds)
FIGURE 119. Order parameter PaCt) plotted against time for a mixture of disks with
diameters of 6 and 10 mm. Thirty percent of the surface is occupied by the smaller disks.
The inset shows the dependence of [1  PaCt)] on time Cafter ref. [90]).
178
The ordering parameter grows exponentially. The inset in Figure 119 shows
that it evolves as Po(t) ex 1  exp( t Ir), which describes a kinetic process
of first order. This remarkably simple behavior remains totally unexplained.
When studying mixtures of different concentrations of particles A and B, the
conclusion is that the time constant T is practically independent of the composition.
As of yet, this too is unexplained.
1.2<D
with
<D E [0.2,0.8].
5.3
Segregation by Shearing
179
This is entirely consistent with the results reported in Section 4.1, namely, that
the flow switches from discontinuous to continuous at around 0.3 rpm. Second,
the speed of 1.3 rpm is well below the 12 rpm needed for the onset of inertial
effects.
Both the previous experiments and the present one were done with 600 small
disks 6 mm in diameter, and 720 large disks 10 mm in diameter, giving area ratios
of 25% and 75%, respectively. Two surprising results emerged. First, the time
constant remained unchanged at 25 s as the rotation velocity varied from 1.3 to
8 rpm. Second, at velocities higher than 8 rpm, segregation disappeared altogether
and the mixture remained substantially homogeneous, even over periods of several
hours.
These results are completely unexpected and, for the time being, without explanation. We would normally anticipate that the process of segregation should
become more efficient as the small disks pass more frequently through the portion
of granular material flowing down the slope. In fact, nothing of the sort happens.
Quite on the contrary, segregation proceeds at the same pace even though the
number of crossings through the liquid phase varies by more than a factor of 6.
Furthermore, since segregation was shown to behave as a kinetic process of first
order, we would expect it to depend monotonically on rotation speed, rather than
to suddenly drop to zero at 8 rpm.
Perhaps a plausible interpretation can be proposed if we go back to the principles
we invoked to explain the role of arch effects in the phenomenon of segregation
by vibration. We argued that a granular system must have enough time to relax
between excitations in order to adapt to the intruder's geometry. Only then can the
intruder migrate efficiently. In other words, segregation is sensitive to the size and
geometry of the objects involved and requires a finite amount of time to manifest
itself. When viewed in that light, the results described above may not be so puzzling
after all. As the rotation velocity increases, the granular surface flow becomes too
fast and chaotic, leaving too little time to adjust to the geometry of the particles.
We may even push the argument a step fmther and envision that the flowing
sheet, which is liquidlike at low speed, gradually turns into a gas as the number
of collisions between particles increases with speed. This hypothetical "phase
change" between liquid and gas may occur rather abruptly and could very well
explain why segregation suddenly disappears above a certain rotation velocity (in
the present case, at 8 rpm). Between 1.3 and 8 rpmthe regime where segregation
proceeds efficientlythe experimental results show that T increases linearly with
rotation velocity (provided that T be expressed in number of revolutions).
180
(a)
Boundary line
(b)
FIGURE 120. Diagram (a) shows particle clustering after the drum, filled as indicated in the
text, has rotated for 300 s. Diagram (b) shows a portion of the boundary of the accumulated
cluster.
There exists a prolific literature on the subject of materials synthesis by a variety of growth mechanisms, including thinlayer deposition, DiffusionLimitedAggregation (or DLA, for short), [lOO]directed percolation, and several others.
Quite often, these various growth mechanisms lead to selfsimilar geometrical
forms, that is to say, structures with fractal properties.
Defining the segregated cluster as that delineated by the greatest possible number
of connected points, we designate by M (r) the length of the jagged line that forms
its boundary. The objective is to calculate this length as a function of the radius r
of a circle centered on a particular point of the line, as shown in Figure 120. Let
M (rs ) be the length obtained when the radius r of the circle matches the radius r s
of the small disks composing the segregated mass. We can then generate diagrams
of the quantity log[M (r) / M (rs )] plotted against log(r/ r s )' The result is shown
in Figure 121. The graph reveals that the boundary of the segregated cluster has
a fractal structure with a dimension d = 1.62 0.2, which allows us to write the
functional dependence M (r) ex rd.
As it happens, numerical simulations have also been done, although in a somewhat different context [l 00]. Based on a model of directed and uncorrelated growth
of a twodimensional structure, they gave an exponent of 1.76, not too different
from the result reported above. It can be shown, incidentally, that the exponent
should decrease when finitesize effects are present, which is most certainly the
case for our segregated cluster. To be sure, these considerations are only semiquantitative. Although they have great pedagogical value, the results described
5.4
181
a
0.0
0.4
0.8
1.2
1.6
log(r)
FIGURE 121. Normalized length M plotted against the radius r of the measurement circle.
here have yet to be confirmed by more careful experimental work and further
theoretical analysis. 14
5.4.1.
Experimental Observations
The experiments described here are done by filling onethird of the volume of a
glass cylinder 70 cm long and 10 cm in diameter, whose internal wall is lined
with roughened spheres. The mixture introduced into the cylinder is composed
of 50% of colored spheres 1 mm in diameter, and 50% of spheres of a different
color and 3 mm in diameter. The cylinder is then rotated around its horizontal
14De Gennes has recently developed a model for segregation via avalanches [43]. The model is based
on a set of coupled variables discussed in Section 4.2.3. Although it deals with a situation that is
different from a rotating drum, it leads to power laws with fractional exponents which may well have
a bearing on the experiments described here.
182
Wavelength A
axis at speeds ranging from 15 to 65 rpm. After 10 or 20 min, the mixture has
separated in bands characterized by a wavelength A, as depicted in Figure 122.
The most salient features of the experimental results can be summarized thus:
Between 15 and 65 rpm, there is little or no dependence of the wavelength on the
angular velocity. When the rotation speed drops below 15 rpm, segregation ceases
altogether, and the smallest bands tend to be the first ones to vanish.
Several authors have reported that the steady state, reached after long periods,
consists of three bands.
5.4.2.
Savage's Model
the dependence of the kinetic angle on particle size has apparently never been observed
in twodimensional geometries, perhaps because the total number of particles is then too limited.
Expanding on an idea advanced in Section 4.1 in connection with avalanches of varions sizes, we
may surmise that the aspect ratio manifests itself only when the number of particles is large enough,
regardless of dimensionality.
183
This we write
with
Focusing our attention on typeB particles, their flux will result from two competing effects. On the one hand, there is a flux <PBx (,6,,8) due to the difference
in kinetic angles, which tends to drive particles along the direction of the xaxis
(see Figure 122). On the other, there is an opposing flux <P BD due to diffusion and
described by Fick's law involving a diffusion coefficient D. The total flux <P Bx of
Btype particles along the xaxis is then given by
<PBx
aCB
aCB
= ,6,,8  D.
ax
ax
To summarize, the horizontal flux is created by differences in the kinetic angle, but
it is opposed by a diffusive component that tends to equalize the concentrations.
The formation of bands would result from competition between the two effects.
6
Numerical Simulations
6.1. Introduction
Numerical simulations aimed at modeling various aspects of the physics of granular materials, which we have touched upon throughout the earlier chapters, have
a twofold objective. l On the one hand, there are pressing incentives to solve a
number of practical problems related to the treatment of granular matter in industry. Whether the issue is pesky segregation, blockages of flows by arch effects,
or disruptive internal convection phenomena (see Chapter 1), the requirements
of the industrial sector are many and, needless to say, almost always immediate.
The urgency of industry's needs and the increasingly rapid developments of creative numerical simulation techniques have prompted many researchers to devote
a great deal of effort to devising algorithms suitable for describing the behavior of
granular materials. 2
On the other hand, numerical simulations are of considerable interest from a
more fundamental point view as well. They offer the possibility to explore the
effect of many parameters which are simply not accessible to experimentation. 3
In that sense, numerical simulation has truly become an integral part of basic
1 An
excellent introduction to the topic of numerical simulations of granular materials can be found in
[59].
2 At
the present time, the number of researchers engaged in numerical simulations in this particular
domain of physics substantially exceeds those pursuing experimental work.
3That is the case, notably, for the coefficients of elastic restitution E and the coefficients of friction
fL, which can be valied at will on a computer, whereas nature offers the experimenter a very limited
range of choices.
6.1
Introduction
185
6.1.1.
The comments we have just made could apply just as well to any other area of
the physical sciences. Yet, the case of granular materials is rather unique. We have
already hinted in Chapter 2 at the difficulties physicists face when trying to model
collisions and friction between solids. Modern numerical techniques can deal with
socalled nbody problems, even when n is quite large. That is no longer an obstacle.
The main challenge is to incorporate in the formulation of these problems the basic
micromechanical properties describing the relevant interactions, and to do so as
accurately as possible. Many decisions have to be made while setting up a model.
What is the duration tc of a collision between two particles? What is the penetration
distance (Section 2.2.2)? How should the proper time step be chosen in relation to
both the time interval between two successive events and the collision duration tc
(Sections 2.2.2 and 3.2.1)? What is the best way to model Coulomb's laws of dry
friction (see Section 6.4)? Perhaps the toughest problem of all is how to account
correctly for the effect of microcontacts and their erosion over time, as mentioned
in Section 2.2.1. Is there away to include phenomena of wear and tear and strain
hardening in the model? All these questions have to be answered before devising
a realistic simulation.
6.1.2.
186
6. Numerical Simulations
6.1
Sequential
Method
II
Loss of
Event
I I I I
Introduction
187
I I
?
Events
11111
Time
Jammed
Point
EventDriven
Method
1111
Events
11111
Time
FIGURE 123. Schematic illustration of the type of difficulties encountered in numerical simulations using either sequential algorithms (top) or eventdriven algorithms
(bottom).
6.1.3.
We have emphasized in Chapter 3 the difficulties faced by continuum descriptions when applied to a real granular medium, which is inherently discontinuous.
Predictably, differential equations become increasingly inadequate as the number
of particles involved diminishes. This type of problem was particularly evident in
our analysis of granular avalanches in Chapter 4.
188
6. Numerical Simulations
There are clearly several possible answers to this question, some more realistic
than others. The technique we describe next has the merit of being fairly intuitive.
It is rooted in the notion of "cloud," which effectively spreads out the mass of each
particle over a region larger than its actual volume [57]. In this picture, the clouds
associated with two particles can overlap, which ensures a continuous passage
from one to the other. The cloud function her) must satisfy several conditions.
They are
00
h(r)2:rrr dr
(61)
1,
her) + 0
when
r +
00,
(62)
(63)
her) :::: O.
(r
where 0' is larger than the diameter d of a particle (for instance, 0' = 6d) and determines the extent of the clould. This enables us to define a density p, a macroscopic
4We have already defined in Section 4.2.1 a granular temperature as the driving force of a thermal
agitation by vibration which causes the detrapping necessary to set off an avalanche. It is not in the
least proven that this definition is identical to the one we are about to put forth in this section.
6.2
Simulations of Collisions
189
p(x) =m I>(lxi
 xl),
;=1
N
p(X)V(X) = m I>;h(lx; 
xl),
;=1
p(x)T(x) =m
L
N
;=1
U2
'h(lxi 
V 2 (X)
xl)  p(X),
2
where N is the number of particles used in the simulation, and m is the mass of
each. These macroscopic quantities are continuous in space and time. It is possible to calculate their gradients, and as such they can be viewed as the usual
thermodynamic variables.
Introduction
6.2.2.
The socalled LRV procedure (for Largest Relative Velocity) is useful when granules in a multiparticulate system come in contact and form what we have referred
to as blocks [59]. The power of the technique lies in its ability to avoid infinite computationalloops that arise when the particles, assumed to be made of hard spheres,
remain clustered. The algorithm relies on a logical test to make predictions about
the outcome. As such, it avoids situations that would result in impractically lengthy
calculations. 5 Figure 124 illustrates a specific case.
5 Researchers doing numerical simulations frequently resort to such predictive techniques.
which consist
in bypassing accumulation points leading to endless loops by deciding ahead of time the state a system
will find itself in. This type of trick, as it were, speeds up the computation time. It only works, of
course, to the extent that the predictions are indeed correct.
190
6. Numerical Simulations
FIGURE 124. Evolution of a group of five spheres, initially grouped in two blocks, for
E = 0.8 (after [59]).
Consider five spheres originally arranged in two blocks of two and three members, respectively, about to collide. What is the trajectory of the five spheres after
the collision? Do the spheres coalesce into a single block? If not, do they separate
and perhaps rearrange themselves in a different pattern? Before solving this problem, it is useful to recall the definition given earlier of a "block" in the context of a
numerical simulation. Two colliding particles form a block if their relative velocity
is smaller than a predetermined value V c chosen according to the characteristics of
the computer used. From a practical standpoint, the velocity differences between
all pairs of adjacent particles forming a block have to be computed at every instant
identified by the ED algorithm. Let 1'..vi = Vil  Vi designate those differences
for all adjacent pairs of a block. Pairs for which 1'.. Vi < 0 do not collide, while those
for which 1'.. Vi > 0 are likely to. The LRV procedure works as follows:
(1) We pick the adjacent pair with the largest value of 1'.. Vi , or 1'..Vj = max(1'..vi),
at the moment of impact, and we let the particles (j, j  1) collide.
(2) Collision matrices of the type described in Section 3.2.1 are used to calculate
an updated set of differences 1'.. Vi.
(3) The previous two steps are repeated until all the differences 1'.. Vi become
either smaller than V c (in which case the corresponding particles form a
block) or negative (in which case they fly apart).
It has been shown that this type of predictive approach does indeed lead to the
same result as the conventional ED method.
6.3
191
of soft spheres and on a sequential calculation. The primary difference with eventdriven methods is that in the present case the duration tc of a collision is not zero.
The principle behind MD methods is to solve in regular incremental steps the
equations governing the changes in linear and angular momenta of the colliding
particles. The objective is to solve the following vector equations
!1p = !1(mv) = mv  mvo =
10t' Fcrndt
(64a)
and
!1(Iw)
fw  fwo
10t' (r x
F)dt,
(64b)
where f is the moment of inertia of the solid around its axis of rotation, r x F is
the torque exerted by the force F, and Fern is the component of the force acting on
the center of mass. Once again, we emphasize that this strategy is quite different
from the one followed in ED models, which starts from the equations governing
momentum exchanges, in the manner described by (22) and (23). In the present
case, solving (64) requires a knowledge of the forces F and Fern, of how they vary
in time, and of the duration tc of the collision. As a prerequisite to any molecular
dynamics simulation, it is essential to model as exactly as possible the forces
of elastic restitution and friction involved during collisions between particles. We
have already stressed on numerous occasions how fundamentally difficult this task
can be (see, in particular, Section 3.1.1), due to the inherently indeterminate nature
of the equilibrium forces in a granular stack, as they depend on its prior history, or
to our limited understanding of contact interactions between solids. This explains
the many forms of equations proposed by various researchers working on this
problem. The situation is not unlike that discussed in Section 4.2.2, in which we
reviewed a variety of functional dependences of the friction forces on velocity.
The goal of the next few paragraphs is to discuss the different types of behavior
we might encounter in modeling the contact forces F and Fern, which feed directly
into (64).
6.3.1.
192
6. Numerical Simulations
contact forces are involved, at least under the assumption that angular momenta
can be neglected. In simplified vector notation, these forces are:
A force of elastic restitution, related to the elastic energy stored during the
penetration of the two particles. This force is given by
(65)
where llij is the unit vector along the line connecting the centers of the particles i and j. This is simply the usual relation for the deformation of a spring
with stiffness K. It is obviously linear and, as such, incompatible with Hertz's
penetration model (Section 2.2.2), which predicts a power law with an exponent of ~ to describe how the force depends on the penetration distance. To
allow for this nonlinearity, (65) is modified to a slightly more general form
(66)
where f3 = ~ in the Hertz model, and f3 =  ~ in the case of a soft crust (see
Section 2.2.2).
A friction force which opposes the rupture of contacts. It plays a dissipative
role similar to that of the EulerCoulomb dynamic friction. For generality,
two components are distinguished. The normal component is
(67)
where mij is the reduced mass of the system of two colliding particles i
and j, V;j is the difference between their velocities, and D n is a dissipation
coefficient characterizing the separation of contact along the direction of llij.
Likewise, the tangential component of the friction force is
(68)
where tij is a vector tangent to the contact, that is to say, perpendicular to llij,
along the slip direction, and D t describes the corresponding dissipation.
Here again the linear approximation contained in (67) and (68) is sometimes too
limiting. The equations are often generalized in the form
6.3
193
Oashpot
FIGURE 125. Mechanical model simulating contact interactions by a coupled springdashpot arrangement. The dashpot acts as a shock absorber.
Mechanical Analogies
The preceding equations were introduced purely phenomenologically. They may,
however, be interpreted in terms of more concrete models that give a physical
meaning to the parameters figuring in the equations. The simplest analogy is depicted in Figure 125. It features a spring (simulating elastic restitution) coupled to
a linear dampener. 6
Such a simple system obviously cannot account for the subtleties of contact
interactions, such as the plastic deformations that typically occur when two colliding spheres penetrate each other, as pointed out in Section 2.2.2. More elaborate
variants have been proposed to simulate these more complicated effects [38]. An
example is depicted in Figure 126. With enough creative imagination, other arrangements can undoubtedly be contrived, but we should keep in mind that such
mechanical analogies have limitations and remain crude pictures of reality. Figure
126(b) describes the behavior of the system shown in Figure 126(a). The spring of
stiffness K 1 is compressed, simulating the two particles colliding and penetrating
each other. On Figure 126(b), the operating point moves up along the straight line
of slope K 1 until its abscissa is equal to et. At that point, the ratchet mechanism
jumps down one notch, which causes the stiffness to suddenly increase to K2.
If the system is allowed to relax in its new configuration, the operating point
moves back down along a different line of slope K 2 , until it reaches the point of
abscissa eto. Since the force has now returned to zero, the system is clearly left in a
different state relative to what it was, which is consistent with the phenomenon of
6 Such
a dampener is sometimes referred to as LSD, for linear spring dashpot. Those fond of acronyms
will shortly be treated to an example of PLS, for partially latching spring.
194
6. Numerical Simulations
(b)
(a)
~o
u.
"'iii
Ratchet
Mechanism
E
o
Normal Elongation
FIGURE 126. Mechanical model of the phenomenon of plasticity. It uses a set of coupled
springs, one of which activates a ratchet mechanism.
plastic deformation. As more and more notches get engaged, the operating point
describes increasingly skinny triangular sectors, as shown in Figure 126(b). The
plastic limit corresponds to the first notch on the ratchet. As long as that condition
is not exceeded, the regime remains linear with a stiffness K j. Beyond that point,
two offset springs act in parallel, with a net stiffness K j  K 2 . As intriguing as
this device may be, it still cannot reproduce Hertz's nonlinear penetration regime.
6.3.2.
MD Collision Model
This section deals with the equations that govern the collision between two particles. We begin with a linear elastic model, and move on next to discuss the
nonlinear elastic regime?
mi
mj
where fei) = fe~) + f//), since only the normal force comes into play in a headon collision. For simplicity, the vector notation has been dropped. The previous
equation applies only when x = ~ (d; + d j )  rij > O. Under these conditions, we
have
d 2x
dx
2
 2 +~+wox=O,
dt
7 So
dt
(69)
as to limit ourselves to these regimes, we will refrain from discussing the simulations done by
Taguchi, who added to the equations a viscous dissipative term [105], [106].
6.3
195
~o e Mt
x(t) =
(610)
sin(iVt) ,
where Vo is the relative velocity just before the collision, and iV is the frequency of
the dampened oscillation, with iV =
ft2. The rate at which the distance x
varies is given by
JW5 
dx
dt
Vo Mt [
. (_)
_
(_)]
iVe
ftsmwt +wcoswt .
(611)
n:
n:
iv
J(K/m)  (D/m)2
= =
Contact ends when x(tc) becomes negative. Note that in the present model, tc is
independent of the relative velocity of the particles. We may define the equivalent
of the coefficient of restitution E: introduced in Section 2.2.2 by writing
E
== 
[dx / dt]t=t,
[dx/dt]t=o
which leads to
This last relation clearly demonstrates the link between the loss of momentum
during collision and the dissipative term D n (or ft). The coefficient of restitution
also turns out to be independent of the relative velocity.
We are now in a position to calculate the maximum penetration depth X max along
the same line we followed for Hertz's model (Section 2.2.2). Maximum penetration
is obtained when the penetration velocity dx/dt vanishes at time t = tmax ' From
(610) and (611), the result is
X max
Wo
If the system is only slightly dissipative (for instance, when E: ::: 0.9), then wo ft,
and tmax approaches the value 2tc , as in the case of Hertz's model. Under these
circumstances, X max reduces to
X max
Vo
=.
Wo
In other words, the penetration depth is then proportional to the relative velocity of
the colliding particles. This result differs significantly from Hertz's model, which
196
6. Numerical Simulations
predicts a much weaker dependence (as v I/5 ). We thus arrive at the conclusion
that a linear elastic model deviates substantially from the physical behavior of
real collisions. 8 It seems necessary to devise a more realistic model incorporating
the nonlinear nature of the contact interactions. That is the objective of the next
paragraph.
Nonlinear Model of a Binary Collision
Using the same notation as previously, we write a generalized form of the differential equation (69)
(X)Y
dx
(X)fJ
d x
m+1]d
+Ed
x=O,
dt 2
d
dt
f3
(612)
(3)
It should be fairly evident by now that modeling collisions between particles is not
easy. The physics of contact interactions is inherently complex and remains poorly
elucidated. Furthermore, good numerical algorithms are tricky to develop, because
they have to scrupulously take into account all the time constants involved (such
as the duration of collisions, the relative velocities, the time of free flights, and
others). Carelessness is likely to lead to unphysical results. To illustrate the point,
we proceed to discuss a completely artificial effect that comes up in models based
8We might come to the elToneous conclusion that the present simple model, based on coupled spring
and dampener, is useless. In fact, it can be shown that, as long as the contact duration tc is chosen
judiciously, in other words, realistically from the standpoint of the physics of the materials involved,
MD simulations yield results that turn out to be fairly satisfactory.
6.3
197
on soft spheres [60]. It has come to be known as the "detachment effect," because
it causes an unphysical separation of particles undergoing multiple collisions. It
can best be understood by pursuing the simple model used in Section 3.2.1 to
describe the behavior of a onedimensional stack of spheres subjected to a vertical
sinusoidal excitation.
where Eo and E f are the initial and final kinetic energies (before and after the
collision). It is important to choose a suitable variable to analyze this problem.
Numerical simulations suggest that one such variable is the ratio cr = so/(votc ),
where So is the initial distance separating the colliding particles. Figure 127 shows
how ceff depends on cr. The trend indicated in the figure seems to be "universal" in
0.5
EDLRV
0.0
6
4
2
log(a)
FIGURE 127. Effective restitution coefficient 8 plotted against the ratio SO/vote (see text).
The horizontal line at 8eff "'" 0.34 corresponds to the result of the EDLRV procedure
described in Section 6.2.2. The present results were obtained for a column of ten spheres
and with fixed walls. The parameters used for the calculation were: d = 1 mm, 8 = 0.9
(the true coefficient of restitution), tc = 0.0022 s, and va = 0.03 m/s (after [59]).
198
6. Numerical Simulations
the sense that the results of MD simulations obtained for an extremely wide range
of values of tc (over three decades) and Va (varying by a factor of 400) all line up
along the same curve.
The graph reveals a sudden change in the behavior of a column of spheres when
CJ = 1, that is to say, when the separation between particles becomes comparable to
the distance traveled during the duration of a collision. As the separation becomes
smaller than this critical value, the effective coefficient of restitution approaches
unity. Such a result is in flagrant contradiction not only with experiments, but
also with theoretical predictions that 8 eff should be a decreasing function of the
number of particles. Here we find, instead, that it becomes equal to or larger than
the coefficient of a single sphere, for which 8 = 0.9. Put another way, the column
appears far too "elastic," which from a practical point of view leads to an artificial
separation of the colliding particles. No such problem exists when the initial separation is sufficiently large (CJ > 1), in which case the molecular dynamics model
agrees quite well with the results of the ED method. The latter technique, coupled
with an LRV procedure, correctly predicts that 8 eff does not depend on CJ. The artificial decompaction just discussed is at the origin of the designation "detachment
effect," whose meaning is further illustrated in Figure 128.
Here the effect is clearly demonstrated when the particles are initially in contact.
If we were to repeat the same calculations with particles initially separated by about
0.01 mm, the two techniques would produce virtually identical results.
There is, incidentally, another related phenomenon, known as the brake failure
effect, when particles collide tangentially [108]. It comes about for very much
the same reason. Here again, particles are slowed down considerably less in MD
simulations than in other, more realistic, mechanicsbased models.
We conclude this brief review of moleculardynamics models with a more
general remark, which in fact applies to all other simulation techniques as well.
0.2
ED
E
oS
0.1
t5Q)
.~
I
0.0
0.25
0.5
Time (ms)
0.75
0.25
0.5
0.75
Time (ms)
FIGURE 128. Trajectories of the centers of ten spherical particles. The MD model was
carried out using exponents f3 = ~ and y = 0 (Hertz's model). Other parameters were:
6
E: = 0.86, and t c = 6 x 10 s for a binary collision, va = 0.2 mis, and Sa = O. The ED
model used the same values (after [59]).
6.4
199
The results generally converge as long as the colliding particles spend most of
their time sufficiently far apart, in which case the dynamic behavior of the entire
system can be accurately modeled by a series of binary collisions. As soon as
more than two particles come in contact at the same time, several questions come
up. Are the collisions binary or ternary, or worse? Are we dealing with blocks?
The answers are never simple, even from a straight physics point of view. All
simulation techniques pay a price for this fundamental indeterminacy, although
the symptoms may differ in each case. These difficulties manifest themselves in
the form of inelastic collapse in ED models, or the detachment effect in their MD
counterparts. As we pointed out in Section 3.1.1, short of knowing the details
of interactions on a microscopic scale, we find ourselves rather helpless when it
comes to predicting the dynamical behavior of a simple stack of as few as three
particles when they are almost in contact.
200
6. Numerical Simulations
/!N
FIGURE 129. The diagrams on the left correspond to Coulomb's law of dry friction. T and
N are tangential and normal forces. The diagrams on the right show Signorini's conditions.
D is the distance between contact points. Both upper diagrams are discontinuous. The discontinuities have been partly mitigated in the lower figures. The gentler form of Coulomb's
law implies a viscous interaction in the vicinity of the contact, while that corresponding to
Signorini's condition assumes an elastic reaction when the solids get close to each other
(after [Ill]).
A similar analysis can be done on the basis of Signorini's conditions, which deal
with the normal, rather than tangential, forces. They apply to hard objects, considered impenetrable in the sense defined earlier:
When V n = Yn = 0, the normal force opposing penetration can have any
value N :::: 0.
Contact is broken the moment V n = and Yn > 0, in which case the normal
force N must vanish.
From this point of view, Signorini's conditions exhibit very much the same type
of discontinuity as Coulomb's law.
20 I
Static
T
Dynamic
!I,N
!IN
!Id N
0
!IdN
!I,N
and
where <t>t and <t>n are the normal and tangential components of the reaction force
due to friction. These components depend on the mode of contact between the
two particles, but not on the external forces, since we deliberately treat the two
separately. If we work in the frame of reference attached to the contact point
between the two particles, the fundamental equations are represented in the above
diagrams by straight lines with a positive slope. These straight lines would intersect
the discontinuous curves at a single point, as shown in Figure 130, which implies
a unique solution. The problem is somewhat more complicated in the case of
dry friction. Whether the solution is unique or not depends on the experimental
conditions and the way Coulomb's friction is modeled. Figure 130 reveals the
following:
If dry friction is modeled with a single coefficient fh = fhs = fhd, the solution
is always unique for a dynamic interaction.
For a static interaction, the straight line describing the fundamental equation
becomes vertical, and the solution becomes undetermined (with an infinity of
solutions).
9 An
excellent analysis of these indeterminacies and how to handle them mathematically can be found
in [111].
lOHere we neglect any possible rotation of the particles. It could easily be added to the equations, but
it would not materially change the argument.
202
6. Numerical Simulations
If dry friction is modeled with two different friction coefficients /is and fLd,
with fLs > /id, the straight line can intersect Coulomb's graph at two distinct
points, and the solution is obviously not unique. Which solution the system
chooses depends on its prior history, which opens the door to the kinds of
hysteresis effects discussed in Sections 2.3 and 3.1.1. That is a commonly
recognized characteristic studied in structural analysis [112].
These considerations may well elicit growing skepticism that it will ever be possible
to accurately model any granular system that is subject to such intrinsic indeterminacies. As we now know, these indeterminacies all come from the discontinuous
character of the static resistance force. One way to get around this problem is to
resort to wellbehaved functional dependences of the type depicted in the lower
part of Figure 129. Another way is to consider the static situation simply as a
limiting case of the dynamic problem (when v + 0). Such arguments may indeed
be viewed as a posteriori justifications of the MD and ED simulations techniques
which, being inherently dynamical approaches, avoid these problems entirely. It
is also essential to bear in mind that we have considered only hard objects (in the
sense of the hard spheres in ED simulations). The creation and rupture of microcontacts is unlikely to be as discontinuous as implied by the standard constitutive
laws. It is in fact quite plausible that smoother functions might describe real phenomena more realistically. What we can say with some confidence is that various
models based on the arguments presented above generally lead to results in good
agreement with experiments [Ill]. This includes ED and MD simulations, as well
as others to be discussed in the latter part of this chapter. There is no compelling
reason to promote anyone technique over another. In all likelihood, a particular
approach, based on specific simplifying assumptions, can be perfectly adequate in
certain circumstances, and completely break down in others. The best strategy is
to be flexible and keep an open mind.
We proceed next to discuss two more simulation techniques, based on procedures for synthesizing piles. These methods may appear somewhat primitive
when compared to the ones reviewed thus far. Yet, they too turn out to produce
very satisfactory results, at least when the geometry of a pile is an important factor.
6.5
203
Let T be one period of the stacking:relaxation cycle. It is worthwhile to pause a moment to understand the implications and limitations of this strategy, in the light of
what we have learned in previous chapters about collision models and the behavior
of granular piles. First of all, this procedure obviously overlooks the dynamical
properties of collisions. Barring additional refinements, it ignores all the problems
associated with solid frictional dissipation, whether static or dynamic. Accordingly, we should not expect this approach to properly describe the behavior of a
collection of particles in frequent collisions. To be more specific, we designate by
TI the time interval between the two closest sequential events defining the dynamics of the pile. In the language of ED modeling of a vibrated onedimensional stack
(Section 3.2.1), TI would be the time between two successive collisions. As we
have seen, this time can become infinitesimally small, giving rise to what is known
as "inelastic collapse." Under these same circumstances, tracking the evolution of
the system with an MD method would require sampling with a period T < TI,
which could easily entail prohibitive computational times. Short of that, the subtle
details of the mechanics of systems undergoing multiple collisions would be at risk
of being missed. This would be equivalent to neglecting events on a short spatial
scale A (of the order of the distance separating particles), which could lead to erroneous results. The Monte Carlo method specifically deals with successive states
of a granular medium after it has relaxed. As such, it is particularly well suited to
describing the physics of granular objects over fairly long time intervals, such as
when they are excited periodically and sufficiently slowly to leave enough time for
the pile to relax between successive excitations. I I With these precautions in mind,
the stacking techniques discussed here can be extremely valuable, notably for the
purpose of analyzing the phenomenon of segregation by size [113]. The next paragraph outlines the practical steps required to implement a Monte Carlo simulation.
11 It
204
6. Numerical Simulations
by following fairly closely the traditional way of using Monte Carlo calculations.
We will subsequently discuss the specifics of applying the technique to granular
materials.
Although we could, without unduly complicating the problem, treat the case of
a threedimensional pile of dissimilar granules, we consider, instead, a collection
of identical disks of diameter d. These disks, assumed to be impenetrable, are
initially arranged randomly in a hypothetical vertical twodimensional container
without walls. 12 In practice, this is approximated by using a ringshaped container.
The initial configuration of such a system of N disks is described by a generalized
vector encompassing the coordinates of all centers
E g (1) =
mgZj,
(613)
j=l
where m is the mass of an individual disk and Z j is the altitude of its center.
The Monte Carlo method is based on analyzing the probability P of different
configurations r, each of which has an energy E g (1). Thermodynamics tells us that
P[E g (1)]
(1)]
= 1 exp [E
g
 ,
Q
kT
where Q is the partition function of the system, and T is its absolute temperature.
Note that this last expression characterizes all the configurations that are equivalent
from an energy point of view, in equilibrium at temperature T. They only differ
by the actual positions of the individual disks.
The technique consists in examining the probabilities of all possible configurations arrived at by moving every disk in the population within a small region of
area 82 . We write this process as a set of equations
(6l4a)
and
(6l4b)
where ~x and ~z are independent random variables equally distributed in the interval [1, + 1], and 8 > O. So as to ensure that the disks do not penetrate each
other during the successive trials, we require that the interaction between adjacent
particles be governed by a potential energy U (s) of a pair such that
U(s) = 0
if s?:. d
(6l5a)
12This is a crucial restriction. We have seen in Chapters 3 and 5 that walls induce convection effects
in granular media. By getting rid of them, convection is conveniently eliminated. All that is then left
are geometrical phenomena, such as "arch effects" of the type described in Section 5.2.1.
6.5
205
and
U(s)
= 00
if s < d.
(615b)
The trials conducted according to (613) through (615) must be evaluated for
plausibility against the following criteria:
If the quantity
t"E
= E(I")
 E(I') :::: 0,
(616)
the new configuration has a lower energy than the one we started with. It is
therefore retained for subsequent calculations.
If t"E > 0, the solution I' is not necessarily rejected, as it may well be
accessible via simple thermal agitation. It is therefore assigned a probability
given by
(t"E)
P t"E  P[E(I")] e x
(
)  P[E(i')]
P kT'
(617)
In tum, this probability is compared to a random number ~ uniformly distributed between 0 and 1. If P (t"E) ::: ~, the solution is retained. If not, it is
discarded, and another one is tried.
The procedure calls for jiggling every single particle of the system, until each
member of the population has had its tum, which completes one iteration. The
new configuration is then used as a fresh starting point for the next iteration, in
which all particles are moved about all over again and allowed to relax, and so
forth.
Some comments on the temperature of the system are in order. The method
we have just described is essentially what is used traditionally for Monte Carlo
simulations of Brownian systems. When it is applied to macroscopic objects like
granular materials, the significance of equations involving the thermal energy kT
raises some legitimate questions.
As pointed out in Chapter 1, the Brownian motion of typical systems of interest here is entirely negligible, the ratio mg t"z/ kT being of the order of 10 12 at
ordinary temperatures. If so, (617), and the criterion associated with it, gives a
probability that is always practically zero. In other words, the only really relevant
equation for a granular system is (616), which means that the potential energy
can only decrease at each step of the iteration. This is equivalent to assuming that
the temperature of the system is at absolute zero. The clear implication is that the
system traps particles in potential wells from which they cannot escape without
collisions on a microscopic level. In accordance with our earlier discussions, it
is clear that this simulation strategy amounts to neglecting the shortrange interactions normally associated with multiple collisionswhich are equivalent to a
local temperature of the granular. Rather, it deals fundamentally with systems in
their relaxed states.
206
6. Numerical Simulations
Despite these restrictions, this type ofsimulation technique has proven extremely
useful to model a number of situations, such as a pile of dissimilar particles. Byway
of summary, we emphasize again that, by its very nature, this particular technique
is not a good choice to describe nonrelaxed configurations, where particles spend
only a fraction of their time in actual contact. One important example is that of
fluidized beds, which are more suitably treated by ED or MD simulations.
5
Stop
6.6
207
the centers of the two underlying particles. We note, incidentally, that this
assumes the absence of any rebound when the particle raches that favorable
spot, in accordance with earlier remarks on these stacking methods.
Once a particle stops, it becomes permanently embedded in the pile.
Agitation can be simulated, for instance, by perturbing the entire system upward
and leaving it to relax on its own. This can be accomplished in the following
manner:
(1) A pile is first generated, by randomly depositing particles one at a time
and allowing them to relax after each addition, using the algorithm just
described.
(2) The stacked particles are numbered in ascending order starting from the
bottom.
(3) The entire pile is raised (fictitiously), and each particle is left to fall down
individually, again using the above algorithm. The process starts with the
lowestnumbered particles and gradually works its way up. To some extent,
it preserves a memory of the pile's prior configuration. 13
(4) Steps (2) and (3) are repeated many times, thereby simulating a vertical
vibration.
This type of simulation is relatively frugal in terms of computation time. It is the
technique of choice to treat cases involving large numbers of particles in three
dimensions. However, the limitations discussed in the context of the Monte Carlo
method apply here as well. Both techniques are good choices to treat a series of
relaxed states, to the specific exclusion of rapid interactions and multiple collisions
that may occur in real systems. The method of steepest descent is particularly well
suited to dealing with problems in which geometry is of paramount importance.
It has produced results in relatively good agreement with experiments. Perhaps its
greatest claim to fame is to have predicted the existence of critical diameters in the
Brazil nut problem, similar to the ones we found analytically in Section 5.2.1. 14
13We encourage the reader to refresh his or her memory by going back to the part of Chapter 3 that
describes the various modes of decompaction of a pile nnder vertical excitation, particularly in one
and two dimensions. This will provide further opportunities to reflect on the degree of realism of the
present algorithm.
14The algorithm described here was originally developed by Jullien et al. [91]. Interestingly, early
versions did not include noise, that is to say, random fluctuations of the particles's positions during
the stacking of the pile, making it entirely deterministic. One consequence was that segregation was
precluded for <P < <Pc, whereas the analytical model in Section 5.2.1 predicts merely a change in
behavior as the critical value <Pc is crossed. A noise source was subsequently added to the model,
and a more realistic behavior did indeed result from this improvement.
Bibliography
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
210
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[49]
[50]
[51]
[52]
[53]
[54]
Bibliography
J. Duran, "La Physique du tas de sable," (The physics of a sand pile), Revue du
Palais de la Decouverte 21 (1994),23224.
R. M. Nedderman, Statics and Kinematics ofGranular Materials (Cambridge, UK:
Cambridge University Press, 1992).
D. Bideau and J. Dodds, eds., Physics of Granular Media (Commack, NY: Nova
Science, 1991).
H. J. Heimann and S. Roux, eds., Statistical Modelsfor the Fracture ofDisordered
Media (New York: Elsevier, 1990).
K. K. Bardhan, B. K. Chakrabarti, and A. Hansen, eds., NonLinearity and Breakdown in Soft Condensed Matter (Berlin: SpringerVerlag, 1994).
E. Guazzelli and L. Oger, eds., Mobile Particulate Systems (Dordrecht: Kluwer
Academic, 1995).
S. B. Savage and D. I Jeffrey, J. Fluid Mech. 110 (1981), 255.
M. W. Richman, Acta Mech. 75 (1988), 227.
M. W. Richman, J. Rheo/. 33 (1989), 1293.
H. K. Pak, E. Van Doorn, and R. P. Behringer, Phys. Rev. Lett. 74 (1995), 4643.
D. D. Joseph, European J. Mech. B (Fluids) 9 (1990), 565.
K. Hotta, K. Takeda, and K, Iinoya, Powder Technol. 10 (1974), 231.
F. P. Bowden and D. Tabor, The Friction and Lubrication of Solids (Oxford:
Clarendon Press, 1950).
E. Rabinowicz, Friction and Wear of Materials (New York: Wiley, 1965).
W. Goldsmith, Impact: The Theory of Physical Behaviour of Colliding Solids
(London: Arnold, 1960).
B. Bernu and R. Mazhigi, J. Phys. A 23 (1990), 5745.
O. R. Walton and R. L. Braun, J. Rheol. 30 (1986), 949.
S. Luding, Phys. Rev. E 52 (1995), 4442.
L. D. Landau and E. M. Lifshitz, Mechanics (Oxford: Pergamon Press, 1976).
S. F. Foerster, M. Y. Louge, H. Chang, and K. Allia, Phys. Fluids 6 (1994), 1108.
J. Duffy and R. D. Mindlin, J. Appl. Mech. (ASME) 24 (1957), 585.
P.G. de Gennes, private communication.
S. N. Domenico, Geophysics 42 (1977), 1339.
J. D. Goddard, Proc. Roy. Soc. (London) 430 (1990), 105.
F. X. Riguidel, R. Jullien, G. Ristow, A. Hansen, and D. Bideau, J. Physique I4
(1994),4973.
G. G. Batrouni, S. Dippel, and L. Samson, Phys. Rev. E 53 (1996), 6496.
F. Heslot, T. Baumberger, B. Perrin, B. Caroli, and C. Caroli, Phys. Rev. E (1994),
4973.
J. Grindlay, Amer. J. Phys. 61 (1993), 469.
E. Guyon, in Physics of Granular Media, edited by D. Bideau and J. Dodds
(New York: Nova Science, 1991).
P. Dantu, in Proc. of the Fourth International Con! on Soil Mechanics and Foundation Eng., vol. 1 (London: Butterworths Scientific, 1957), p. 144.
S. F. Edwards and C. C. Mounfield, Phys. A 210 (1994), 290.
D. Stauffer, H. J. Hermann, and S. Roux, J. Physique 48 (1987), 437.
T. Travers, M. Amrni, D. Bideau, A. Gervois, J.C. Messager, and IP. Troadec,
Europhys. Lett. 4 (1987),329.
Bibliography
[55]
[56]
[57]
[58]
[59]
[60]
[61]
[62]
[63]
[64]
[65]
[66]
[67]
[68]
[69]
[70]
[71]
[72]
[73]
[74]
[75]
[76]
[77]
[78]
[79]
[80]
[81]
[82]
[83]
[84]
[85]
[86]
[87]
[88]
[89]
211
212
[90]
[91]
[92]
[93]
[94]
[95]
[96]
[97]
[98]
[99]
[100]
[101]
[102]
[103]
[104]
[105]
[106]
[107]
[108]
[109]
[110]
[111]
[112]
[113]
[114]
[115]
[116]
Bibliography
Index
Convection, 7, 87, 93
CPP, 92, 102, 179
Critical, 122, 125
Critical ratio, 163
Aggregate, 3, 5
Angle of movement, 48, 121
Angle of repose, 48, 93, 119
Arching, 10,60, 104, 110, 161,
204
Aspect ratio, 74
Attractor, 176
Avalanche, 16,48,53, 127
Electrostatics, 24
Event driven, 77, 108
CAM,128
Cannon ball, 54, 64, 101
Chaos, 80, 85
Cohesion, 5, 16
Collision, 34, 55, 77,193
Compaction, 7
Condensation, 82
Fluctuations, 1
Fluidization, 53, 76, 82
Fluidized bed, 15
Fractal, 23, 132, 176, 179
Fracture, 16, 107, 166
Fragmentation, 53, 72, 101,
104
213
214
Index
Penetration, 42
Percolation, 55, 65, 130, 178
Plug, 3,10
Redirection, 17, 70
Restitution, 36, 195
Reynolds, 18,21,65
Rotating drum, 152, 155, 171
Rotation, 30, 109
Segregation, 2, 12, 154
Selforganization, 86,111,114,
128
Sifting, 178
Signorini conditions, 191, 200
Silo, 3, 70, 153
SOC, 16, 128,203
Soft crust, 44, 192
Soft spheres, 186, 191
Steepest descent, 207
Stickslip, 32, 138