Built upon sand: Theoretical ideas inspired by granular flows

Leo P. Kadanoff*
The James Franck Institute, The University of Chicago, Chicago, Illinois 60637

Granulated materials, like sand and sugar and salt, are composed of many pieces that can move
independently. The study of collisions and flow in these materials requires new theoretical ideas
beyond those in the standard statistical mechanics or hydrodynamics or traditional solid mechanics.
Granular materials differ from standard molecular materials in that frictional forces among grains can
dissipate energy and drive the system toward frozen or glassy configurations. In experimental studies
of these materials, one sees complex flow patterns similar to those of ordinary liquids, but also
freezing, plasticity, and hysteresis. To explain these results, theorists have looked at models based
upon inelastic collisions among particles. With the aid of computer simulations of these models they
have tried to build a ‘‘statistical dynamics’’ of inelastic collisions. One effect seen, called inelastic
collapse, is a freezing of some of the degrees of freedom induced by an infinity of inelastic collisions.
More often some degrees of freedom are partially frozen, so that there can be a rather cold clump of
material in correlated motion. Conversely, thin layers of material may be mobile, while all the
material around them is frozen. In these and other ways, granular motion looks different from
movement in other kinds of materials. Simulations in simple geometries may also be used to ask
questions like When does the usual Boltzmann-Gibbs-Maxwell statistical mechanics arise?, What are
the nature of the probability distributions for forces between the grains?, and Might the system
possibly be described by uniform partial differential equations? One might even say that the study of
granular materials gives one a chance to reinvent statistical mechanics in a new context.
[S0034-6861(99)00701-1]

CONTENTS them all. In addition, the properties of a granular mate-

rial can depend upon its history. Tamped sand is differ-
I. Introduction: A Basic Description of Statistics ent from loose sand.
and Flow 435 But in many ways, a granular material is like an ordi-
II. Experiments on Shaken Sand 436 nary fluid. Both types of material are composed of many
III. Shaken Motion: Some Evidence Against Simple small particles, and each has a bulk behavior that hides
Hydrodynamics 437
the material’s graininess. It is thus natural to ask
A. Rapid variation 437
B. History dependence 438
whether the same equations, concepts, and theories that
IV. Statistical Dynamics 439 work for molecular material also apply to the granular
A. Bunching 439 form of matter.
B. Inelastic collapse 440 So let us go back to the fundamental statistical de-
V. Low-Dimensional Systems 441 scription of matter as developed by Boltzmann, Max-
VI. Conclusions 443 well, Gibbs, and many others. The macroscopic state of
Acknowledgments 443
a statistical system is described by a probability distribu-
References 443
tion, which contains a set of parameters like the tem-
perature, chemical potential, and velocity of the system.
The number of such parameters is equal to the number
of independent conserved quantities. For the standard
I. INTRODUCTION: A BASIC DESCRIPTION
OF STATISTICS AND FLOW
one-component fluid, the parameters can be taken to be
the mass density r, the average velocity of the system u,
Granular materials show a wonderfully diverse set of and the temperature T. We use the symbols r a to denote
behaviors.1 Make a sand castle, and the material appears the densities of the different conserved quantities, and
solid. Push on the castle and it can fall down in an the symbols m a to denote the macroscopic parameters.
avalanche-like pattern. Sometimes the avalanche moves In part, the conserved quantities and the parameters
the bulk of the material, sometimes it is confined to a are observable in the thermodynamics of the system.
thin layer on the surface. Shake up crushed ice in a mar- They are also visible in the dynamics. We concentrate
tini shaker, and it moves like a gas. Try to pour salt upon the slow dynamics that arise when the time varia-
through an orifice, and it has a characteristic tendency to tion within the material is very slow in comparison to
choke up and clog the orifice. Gas, liquid, solid, plastic the typical collision and relaxation time of the constitu-
flow, glassy behavior—a granular material can mimic ent particles. We also require that spatial variations be
slow in comparison to mean free paths and to the size of
any one of the constituent particles. One source of pos-
*Electronic address: [email protected] sible slow motion is the gradual flow of conserved quan-
1
For recent reviews see Behringer and Jenkins (1997). For a tities from one part of the system to another. These mo-
more specialized view, close to the outlook of the present tions may be described by saying that each part of the
work, see Jaeger, Nagel, and Behringer (1996a and 1966b). system is in a local thermodynamic equilibrium and then

Reviews of Modern Physics, Vol. 71, No. 1, January 1999 0034-6861/99/71(1)/435(10)/$17.00 ©1999 The American Physical Society 435
436 Leo P. Kadanoff: Granular flows

using the conservation laws as flow equations. Addi- ditions for the derivation of hydrodynamic equations,
tional slow motions may arise when there is a broken which are mostly concerned with the slowness of varia-
symmetry in the system. We shall not worry much about tions in space and time. Maybe hydrodynamics does de-
the broken symmetries here. A less well understood scribe motion in granular materials. Indeed experiments
source of slow relaxation is glassy behavior, with its par- in fluid and in granular materials look similar. Many
tial freezing of degrees of freedom and its concomitant people, going back to Faraday2 (1831), have seen local-
very slow relaxation to full equilibrium. We shall see a ized and delocalized excitations on the surface of a con-
considerable glassiness in granular systems. tainer of shaken sand. These patterns of motion look
To derive equations for the slow flow, one starts from quite similar to the Faraday crispation patterns devel-
the conservation laws and uses the local equilibrium to oped on the surface of a shaken fluid. Even the new
derive what are described as the hydrodynamic equa- localized sand excitations called oscillons (Umban-
tions for the system. For each conserved quantity there
hower, Melo, and Swinney, 1996) look very similar to
is a corresponding current ja (r,t) describing its flow.
localized excitations seen in water by Lathrop and
The hydrodynamic equations are the set of conservation
Putterman.3
laws:
However, there is a major difference between mol-
] ecules and grains. An ordinary fluid conserves energy
r ~ r,t ! 1¹+ja ~ r,t ! 50 (1)
]t a within the observed degrees of freedom. Thus, if one
puts heat into a fluid, this heat contributes to the kinetic
supplemented by thermodynamic relations and constitu- energy of each of the molecules as part of the process of
tive equations that define how the density and current raising the temperature of the fluid. This added energy is
depend upon the thermodynamic parameters m a . We never lost, except perhaps by a slow leakage through the
assume and assert that the densities are local functions walls of the vessel. Thus a macroscopic flow, once
of the parameters, e.g., that the energy density at r,t started, is likely to continue for a long time. Moreover,
depends only upon the mass density, temperature, and the relative motion of the basic particles, called heat,
velocity at the same point in space-time. We further take will never die away.
the currents to be only functions of the local parameters In contrast, in a granular material, some energy can be
and their gradients. For example, Galilean invariance lost to heat in each collision. Heat energy is stored in the
gives the mass current as r u, while a constitutive equa- (unobserved) molecules within the grains, and not as ki-
tion gives the heat current as 2 k ¹T, where k is the netic energy of the observed grains. From the point of
thermal conductivity. view of the grains, the system dissipates energy very rap-
All this is classical and gives us the usual hydrody- idly. If left alone, the system would get stuck in a solid
namic equations as partial differential equations in space or glassy configuration and relative motion would come
and time. There are no long-range effects, that is, no to a virtual or complete halt. This complete relaxation
integrals over space. There are no memory effects, that might happen in one region of the material and not in a
is, no integrals over previous history. The entire descrip- neighboring one. Thus sand may never show the relax-
tion is in the few partial differential equations and their ation to overall uniform equilibrium that is required in
boundary conditions. the usual derivations of hydrodynamic equations (Chap-
Along with this nonequilibrium theory, we inherit man and Cowling, 1990). Because of this failure, we can-
from the old masters an equilibrium theory in which the not be at all sure that hydrodynamic equations describe
probability of anything is given by the Gibbs form of the behavior of granular materials in any general way.
Maxwell-Boltzmann statistics, i.e., the probability is an But it would be nice if a set of stable hydrodynamics
exponential linear in such conserved quantities as the equations did apply. We would like to be able to de-
energy and momentum. The parameters m a are the ad- scribe the granular material by a small set of local vari-
justable constants in these probability functions, which ables. Perhaps the variables would be a local velocity, a
are then used to set the average values of the various density, and an effective temperature, as in an ordinary
conserved quantities. one-component fluid. We would further wish that we
could write a set of local equations connecting the local
II. EXPERIMENTS ON SHAKEN SAND values of these quantities. These local equations would,
in our dreams, be partial differential equations, maybe
Many authors have done studies of the effects of shak- very closely similar to those of ordinary hydrodynamics.
ing a granular material, perhaps composed of glass Then the equations would be supplemented by bound-
beads, or grains of sand, or rice, or coal. [See, for ex- ary conditions, and we would have a complete descrip-
ample, Larouche, Douady, and Fauve (1989); Knight, tion of the space and time development of the system
Jaeger, and Nagel (1993); Pak, Doorn, and Behringer (Haff, 1983; Jenkins and Savage, 1983; Jenkins, 1992;
(1993); Melo, Umbanhowar, and Swinney (1994, 1995);
Metcalf, Knight, and Jaeger (1997)]. In these experi-
ments, there are many many individual constituents, 2
Faraday (1831) himself credits Chladni for the discovery of
each one in motion and bumping into its neighbors. The sand patterns. More information on Chladni can be found in
objects are small and the relaxations are rapid. At first Waller (1961).
3
sight it would appear to be quite easy to satisfy the con- Private communications.

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

 Leo P. Kadanoff: Granular flows 437

Hayakawa, Yue, and Hong, 1995; Sela and Goldhirsch,

1996; Grossman, Zhou, and Ben-Naim, 1997). Are our
dreams connected to reality?
Before we look at the experimental answers to this
question we should contrast hydrodynamic behavior
with the possible alternative behaviors that might arise.
The hydrodynamic situation occurs when the system is
fully defined by a set of partial differential equations and
their boundary conditions. We also have to specify the
values of some small set of state variables at some initial
time. Except for the description of walls that define the
container, there is no explicit coordinate dependence in
the equations. All coefficients in the equations are com-
pletely independent of time. These hydrodynamic equa-
tions remain uniformly valid throughout space and time.
Unless the partial differential equations themselves de-
velop troubles in the form of singularities or infinities FIG. 1. The positions of grains of sand at the same phase in
the motion of two successive shakes. In plate (a), the particles
(Kadanoff, 1997), the equations will hold everywhere.
are colored in black or white depending on their position at
We call the situation in which the system is described by
that time. These same particles are depicted in plate (b). No-
uniform and uniformly valid partial differential equa-
tice the downward motion in a boundary layer very near the
tions hydrodynamic behavior, or more precisely uniform wall.
and local hydrodynamics.
In the next level of complexity, instead of partial dif-
ferential equations, the system is described by integral 1. What is the pattern of the grains during a shake?
equations. If there are time integrals, we say that the 2. What is the long-term pattern of motion? Specifically,
system has a history dependence; if there are space in- can one give a qualitative description of how the various
tegrals, we say that it is nonlocal. Either way, the equa- grains will move over the long term? Naturally, the long-
tions are more complex than in the hydrodynamic situ- term motion is an expression of the composite effects of
ation. Another kind of complexity arises when the the short-term motion.
system is quite sensitive to local fluctuations. In this situ-
A. Rapid variation
ation, small fluctuations can give large and unpredict-
able deviations from uniformity. These nonuniform We have information about this situation from both
cases can be found in turbulence, flame fronts, and other experiments (Jaeger and Nagel, 1992) and simulations
chaotic situations. When, as in the case of granular ma- (Grossman, 1997a, 1997b). We look in detail at the re-
terials, there is an inherent nonuniformity in the system sults of a simulation done by Grossman. This simulation
itself, any mechanism for magnifying fluctuations can shows many of the same features as the experiment it
give us all kinds of complications. On large scales, the was intended to model. We want to know whether the
equations describing the system may end up as complex motion during a shake and the long-term motion are
as a set of nonlinear partial differential equations con- hydrodynamic in character.
taining stochastic coefficients and/or forcing. [See, for Figure 1 shows sand flying at the maximum of two
example, Shinbrot (1997).] That kind of situation can be successive shakes. The balls are colored in black and
as complex as anything in quantum field theory. So ask white so that one can follow the net motion integrated
once more, What kind of problem is posed by a granular over one full shake. Note that the top is flying free and
material? that there is considerable motion in boundary layers at
the side walls. Integrated over the entire shake, there is
III. SHAKEN MOTION: SOME EVIDENCE AGAINST little net motion in the central bulk of the material. Re-
SIMPLE HYDRODYNAMICS call that the derivation of hydrodynamic equations re-
quires slow variation in space. However, the ‘‘convec-
In this section, I look at some of the evidence that tion experiment’’ of Fig. 1 involves motion in a thin
suggests that granular materials might not be describ- boundary layer, where—in fact—the motion is most
able by uniform and local hydrodynamic equations. I am rapid (Hayakawa, Yue, and Hong, 1995). In both the
going to follow over some of the ground covered by real experiment and the simulation, the layer is only a
recent review papers (Jaeger, Nagel, and Behringer, few grains thick. This thickness is not suitable for the
1996a, 1996b) but emphasize different aspects of the development of a uniform, hydrodynamic description. It
data. may well be that, in most granular flows, thin layers of
There is a very interesting series of experiments in distinctive behavior will dominate the flow of hydrody-
which sand is brought into motion by successive vertical namic materials and ruin the possibility of a uniform
shakes (Jaeger, Knight, Liu, and Nagel, 1994). An accel- hydrodynamic theory, [for further work on shaking and
eration of more than 1 g sends the sand near the top of fluidization, see Clément and Rajchenbach (1991); Gal-
the container flying. One then asks two questions: las, Hermann, and Sokolowski (1992); Esipov and

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

438 Leo P. Kadanoff: Granular flows

function of the number of shakes. The system starts with

a low density. For a very long period, including many
shakes, the density slowly rises. In between the shakes,
the material does not move. So time-independent be-
havior is observed at many different densities. The den-
sities are determined by the prior history of the sample.
We see then that the system has a history dependence in
which slow changes over long periods can have a
marked effect. This dependence is particularly impor-
tant because the material’s flow properties are very sen-
sitive to the density, with higher densities producing en-
hanced resistance to flow (Ristow, Strassburger, and
Rehberg, 1997).
The data in Fig. 2 are fit by curves described by Eq.
FIG. 2. Compaction of a granular material. See Nowak et al. (2) (Nowak, Knight, Ben-Naim, Jaeger, and Nagel, 1995;
(1995). A granular material is initially ‘‘fluffed up’’ by shaking Ben-Naim, Knight, Nowak, Jaeger, and Nagel, 1998).
it with a stream of nitrogen gas. Then it is shaken with a maxi- This fit shows that in the long run the density will rise to
mum acceleration G, measured in g’s. As the shaking goes on it
an ‘‘equilibrium’’ value which depends on the accelera-
compacts more and more. The data points show how the den-
tion G. The slow rise and the G dependence show that
sity r depends upon the number of shakes t. The curves are
constructed as the best fit to Eq. (2).
the density does not just depend upon the conditions at
one time. Instead, it has an important history depen-
dence. This kind of dependence is not found in any or-
Pöschel (1997)]. Such boundary layers can also appear dinary liquid.
on free surfaces [see de Gennes (1997) and Boutreus Ben-Naim developed a theory4 of automobile parking
and de Gennes (1997)]. that effectively explains the shake dependence of Fig. 2
The time dependence also shows rapid variation. Fig- (Nowak, Knight, Ben-Naim, Jaeger, and Nagel, 1995). In
ure 1 shows that the grains are thrown into the air where this theory a parking spot appears via the collection of
they move freely for a time comparable with their relax- many bits and pieces of empty space until, by a random
ation time at that density. Is such free-particle motion process, sufficient space is assembled to make a full
hydrodynamic? This free motion further casts doubt on parking spot. Thus, for high density, the time to park
the validity of hydrodynamic theory for this situation. varies exponentially with the available space. Con-
In fact, the possibility of hydrodynamic breakdown versely, the density varies in a logarithmic law involving
via short-wavelength behavior is well known in another time. When translated into the granular-compaction
context. If you shear a granular material, or indeed an process, this model suggests a law for compaction which
ordinary solid, sufficiently hard it can produce thin lay- gives the density as a function of the number of shakes t
ers of weakly bound material (Oda, Iwasbhita, and (or equivalently the time since the material was last
Kakiuchi, 1997). These so-called shear bands have been fluffed up) as
conjectured to play a role in earthquake fault zones.
r ~ ` ! 2 r ~ t ! 5A/log~ 11t/ t ! , (2)
where A, r(`), and t are all parameters that can depend
B. History dependence
upon G. As one can see from Fig. 2, the phenomenologi-
In addition, there are classical experiments on granu- cal theory does fit the observed facts. A little later
lar materials which show that the state of these materials Gavrilov (1997)5 developed a parallel theory based upon
is quite dependent on their past history. Such history highway congestion, which measures the number of clus-
dependence is possible for glasses or spin glasses or fro- ters of different sizes. His theory gave a result like Eq.
zen materials, but it is not consistent with equilibrium (2) and also described the fluctuations in density after
statistical mechanics. I describe one of these experi- the system reached steady state (Gavrilov, 1998).
ments here (Knight, Fandrich, Lau, Jaeger, and Nagel, This and analogous results leave the theory in a very
1995). Imagine that you fill a tall glass vessel with granu- uncomfortable position. On the one hand, we would like
lar material and measure the density of the material by a hydrodynamic theory that is local in space and time.
placing the glass tube inside a set of capacitor plates. On the other hand, we see that the system can come to a
The capacitance gives a direct measurement of the rela-
tive volume of interior filled by air and by the grains. 4
You start out by fluffing up the granular material by Ben-Naim’s theory was developed in Boston, where one can
have ample time to speculate on mechanisms for producing
swirling it around with compressed gas, and then you
suitable parking spots. A competitive theory was developed by
shake the tube in a vertical motion many, many times. Gavrilov (1997).
The entire experiment is described by two variables: G 5
Gavrilov, a graduate student at Chicago, developed his
the maximum acceleration relative to g and t the number (1997) theory of clustering by observing traffic patterns from a
of shakes. train window. He then applied this theory to granular materi-
Figure 2 shows typical pictures of the density as a als (unpublished.)

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

 Leo P. Kadanoff: Granular flows 439

stop with a wide range of different densities. Hence the

density, at least, is history dependent. It is hard to make
a system of partial differential equations that comes to a
stop in this way. Thus it is quite possible that no differ-
ential equation theory will apply to this system. In the
models, the relaxation of a shaken system comes from
the filling in of defects and vacancies in the material. To
describe processes like these, we need some statistical
theory of defect production and removal (Shinbrot et al.,
1997; Venkataramani, 1998). Hydrodynamic equations
for a uniform system do not provide for these kinds of
statistical processes. The theories of Ben-Naim and FIG. 3. Simulations of granular motion. Taken from Mc-
Namara and Young (1994). Here there are 1024 inelastic disks
Gavrilov thus suggest that we have to reach beyond hy-
with periodic boundary conditions. The cases are (a) r50.99
drodynamics to include stochastic processes as part of
and (b) r50.6. In both cases the system is run from random
the basic equations of the system. When such processes initial data until there are a large number of collisions per
act in space and time, they can produce very rich behav- particle. The particles marked in black have taken part in the
ior, so that the resulting theory may be extremely diffi- last 200 collisions. Note the clumping of particles and colli-
cult. sions.

IV. STATISTICAL DYNAMICS

ticles shown in black are those which have participated
A. Bunching in one of the last 200 collisions.
In case (a), the system is almost elastic. Only weak
Let us start again. Let us look at simulations of iso- correlations develop among the particles and the recent
lated systems containing a relatively large number of collisions. In case (b) there is more inelasticity and the
particles. For the moment, we take these particles to particles are more bunched up in space. Figure 4 shows a
have center-of-mass motion but no spinning motion. simulation with considerably more particles carried out
They all have equal masses. (In this paper, we shall take by Goldhirsch and Zanetti (1993). It has the same r
the masses of all particles to be equal to one.) The pe- value as in Fig. 3(b). Notice once more the very consid-
culiar properties of granular materials are represented erable bunching up of the particles. In fact, these au-
by having the particles lose a little energy in each colli- thors argue that the bunching is the result of a hydrody-
sion. Specifically take the overall collision to conserve namic instability in which regions of reduced
momentum and to conserve for each particle the com- temperature have reduced pressure, which then fills
ponent of the velocity perpendicular to the line of cen- them with more particles so as to equalize the pressure.
ters. In an elastic collision, described in the center-of- But then these regions have more frequent collisions,
mass system, the components of each particle’s velocity which reduces the temperature still further, and so forth
along the line of centers is reversed. In our case, we shall until some small region has a very high density. Since
take the velocity to be diminished by a factor r, called hydrodynamics fails when the system’s dynamics pro-
the coefficient of restitution, so that the new normal duces very rapid variations in space or time, one can
component is related to the old, v n , by
v n8 52r v n 21<r<1. (3)
The last condition is required to make the energy con-
tinually decrease. The usual elastic case is r51.
Figure 3 shows two simulations due to McNamara and
Young (1996), in which they look at inelastic collisions
among 1024 particles in a two-dimensional box with pe-
riodic boundary conditions. The simulation is of the
‘‘event-driven’’variety (Rapaport, 1980). The program
looks ahead to the next collisions that can occur, picks
the one that actually occurs first, readjusts the velocities
of the two colliding particles, and then looks ahead once
again with the appropriately adjusted set of possible col-
lisions. In the simulation, the particles start out with ran-
domly chosen positions and velocities picked from a
Gaussian ensemble. As time goes on the particles slow
down considerably, but this slowing does not compro-
mise the effectiveness of the computer program. Com- FIG. 4. Another picture of granular motion. From Goldhirsch
pare the simulations. Both pictures show the situation and Zanetti (1993). Here there are 40 000 particles with r
after a large number of collisions. In both cases, the par- 50.6. Notice the very evident clustering.

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

440 Leo P. Kadanoff: Granular flows

worry that the bunching shown in Figs. 3(b) and 4 might in a finite period of time. In this model, after a finite
invalidate any hydrodynamic-style theory of the granu- time the ball comes to rest on the surface of the table.
lar material. Thus inelastic collapse does indeed occur in a very
simple everyday situation.
Another case open to analysis involves three particles
B. Inelastic collapse
with equal mass which are free to move on a line (Con-
stantin, Grossman, and Mungan, 1995). Their velocities
There is another, independent, symptom of the indefi-
are, in order of increasing x coordinate, u, v , and w.
nite bunching of particles and collisions. As pointed out
Their entire motion can be described in terms of the
by Haff (1983), Bernu and Mazighi (1990), and Mc-
ratio of the relative velocities, h 5(u- v )/(w- v ). The
Namara and Young (1992), subsets of the particles in
subsequent math is very easy once one sees how to set
this hard-sphere granular system can undergo an infinite
up the problem. Assume that initially the two outer par-
number of collisions in a finite period of time. One can
ticles approach the inner one. The subsequent process is
in fact see this phenomenon in Fig. 3(b). Notice how the
one in which the outer particles collide successively with
dark particles, those which have participated in the last
the inner particle. If the particles with velocities u and v
200 collisions, form a weakly curved connected region.
collide, the subsequent value of h is
These particles will soon collide an infinite number of
times, so that all their motion in the lateral directions 11r 1
will go to zero. In the meantime, however, the motion in 1/h 8 5 2 .
2r rh
the transverse directions will continue, and after a time
the particles will move apart. However, before this hap- If the other two particles collide, the new value of h is
pens much of the energy of their relative motion will given by
have been lost. This phenomenon of infinitely repeated 11r h
collisions is called inelastic collapse. h 85 2 .
2r r
There has been some controversy about the impor-
tance of such collapse for real granular materials.6 We Given these relations, one can follow the entire subse-
shall discuss this further below. However, before passing quent motion and figure out the range of r and initial
judgment upon its significance, we should try to under- values of h that will produce one collision, or two, or
stand the collapse by studying it in the simplest possible ten, or an infinite number. The last is the case of col-
situations. lapse. The net result is very simple: collapse can occur
Imagine a ball in vertical motion. It is pulled down by for suitable values of h only when r stands in the range
gravity and is bounced up by partially inelastic collisions
0<r<724)50.073 . . . . (5)
with a table top. This is a high-school physics problem,
with a tiny modification for inelasticity. Say that the The entire collapse is a kind of approach to a ‘‘fixed
speed just after the mth bounce is s m . The speed on the point’’ in which the value of h which appears after the
next bounce is assumed to be given by an expression like nth u- v collision relaxes to an n-independent value for
that used in Eq. (3), namely, large n. Thus, in the long run, the process simply repeats
itself.
s m11 5rs m . (4) As this calculation is extended to higher dimensions,
Here the coefficient of restitution r is restricted to be one finds that there are additional variables which de-
between zero and one. The speeds approach zero as a scribe the state of the system at the point of collapse
geometric series, s m 5s 0 r m , while the heights of the nth (Zhou and Kadanoff, 1996). First there is the angle u
bounce obey h m 5s 2m /(2g);r 2m and so the time be- between the particles’ lines of centers at the moment of
tween collisions, being proportional to h m /s m goes to collapse. For each value of r, there is a range of u values
zero as r m . Thus we have an infinite number of collisions that will permit collapse. This range is shown in Fig. 5.
Notice that only small values of r permit collapse and
that the permissible range of u gets smaller and smaller
6
For example, a referee of this paper said: ‘‘The issue of in- as r is increased.
elastic collapse is in one sense a computational artifact, in that A second set of new variables are the transverse com-
nothing physical happens. Rather, computations grind to a halt ponents of the momentum. In one region of Fig. 5, la-
as the number of collisions grows,’’ I neither fully hold to nor beled stable collapse, nonzero values of these variables
fully deny this view. On the one hand, for dimensions greater will not upset the approach to a fixed point in which the
than one, there remains an unquenched component of velocity motion is repeated again and again over decreasing
perpendicular to the line of collisions. Thus, as stated, motion scales of distance and lateral velocity. Another region,
continues after collapse. Further, collapse does not occur in
labeled unstable collapse, requires these values to be set
the same way if the particles have a soft—but dissipative—
to zero. If these transverse components are not set to
interaction. On the other hand, the collapse time is an instant
of extremely singular velocity correlations. Mathematicians are zero, the particles can have many collisions but will
interested in questions about how singularities can arise from eventually veer away from one another, preventing in-
classical mechanics. In addition, the collapse signals the onset elastic collapse.
of strong correlations, which may be robust even if the collapse Note that Fig. 5 in part explains why the collapsing
is not. particles in Fig. 3(b) essentially lie on a straight line. For

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

 Leo P. Kadanoff: Granular flows 441

FIG. 5. Permitted values of r and u for inelastic collapse. See

Zhou and Kadanoff (1996).

three particles, we see that the collapse occurs over the

largest range of r, when u is small. Thus collapse is fa-
FIG. 6. Collapse with rotation. From Schörghofer and Zhou
vored when u is small. This requirement apparently ex-
(1996). A simulation of elastic collision among particles that
tends itself to the many-particle case and produces the
are free to rotate, in a two-dimensional box with periodic
roughly collinear observed behavior. boundary conditions. The inset in the upper right-hand corner
All this describes a situation in which the particles shows a similar simulation without rotation. Notice that the
cannot rotate. Schörghofer and Zhou (1996) extended collapsed particles (in black) form a zig-zag pattern in the ro-
this calculation to include rotational motion. They found tating case and a much straighter pattern in the absence of
that, when the particles rotate, high-u collapse configu- rotation. This result indicates large values of u in the rotating
rations become possible—and indeed quite probable— case and smaller ones in the absence of rotation.
for three particles. When they simulate collapse for a
many-particle configuration (see Fig. 6), they find that
the collapsing particles form a zig-zag pattern, reflecting systems. Start with a large number N of particles moving
a high-u value during the collapse. along a line (Du, Li, and Kadanoff, 1995). In this one-
These and a variety of other calculations show that dimensional example, we shall pick the inelasticity, «
collapse can be understood, and is indeed mostly under- 512r, to be a small number so that the system may be
stood. However, this understanding does not contribute expected to have a substantial similarity to a set of fully
much to our knowledge about either the nature of elastic particles. The left-hand wall of the system will be
granular motion or the overall effect of the collapse hot, defined so that any particle which hits it comes off
upon that motion. Several workers in the field have ar- with a perpendicular component of velocity chosen from
gued that the collapse phenomenon is so dependent the Maxwell-Boltzmann distribution7 v exp(2v2/2) (see
upon the hard-sphere system that it is essentially uncon- Fig. 7). The right-hand wall of the system is taken to be
nected with the sorts of slowdowns and collapse that will a reflecting wall. A particle that hits it is reflected with-
occur in real granular materials. If collapse is not very out changing its speed.
robust, it will probably offer little insight into the inter- Before we studied the simulation, we expected that
esting things that happen in real materials. On the other the system would relax to a situation like that in the
hand, collapse might be a robust symptom of the bunch- usual equilibrium statistical mechanics: a uniform distri-
ing and slowing down that certainly occurs in a granular bution in space with a Gaussian distribution in each
system. It might produce isolated regions in space-time component of each particle’s velocity. The expectation
which are so frozen that they form defects in the hydro- was that, as « got smaller, the fit to the usual statistical
dynamics and, in fact, produce regions in which addi- mechanics would get better and better. Instead some-
tional variables or information are necessary to com- thing entirely different happened. In the limiting situa-
plete the hydrodynamics. The last word has not been tion in which N is big and N« is of the order of unity, the
said about the importance or unimportance of this col-
lapse.
7
This probability distribution is not itself the Gaussian distri-
bution connected with the names of Maxwell and Boltzmann.
V. LOW-DIMENSIONAL SYSTEMS However, this distribution of wall-induced velocities produces
the appropriate distribution in phase space for the particles
To go beyond the question of collapse, let us look at moving away from the wall, prior to their first collision with
what is happening to very simple examples of dissipative other particles.

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

442 Leo P. Kadanoff: Granular flows

and reflecting walls on the top and the bottom. There

are two simple instructive cases, the first is a ‘‘pipe’’ that
is only a bit wider than a single particle, and the second
is a square box. The pipe is substantially different from
the one-dimensional case because the structure of the
equilibrium situation is changed entirely. As soon as
there is a transverse component of the velocity sharing
FIG. 7. A group of slightly inelastic particles in one dimension. energy in a scattering event, the system loses all the
After Du, Li, and Kadanoff (1995). The particles end up in a extra conservation laws that dominated the one-
situation in which one particle moves fast and the others sit dimensional situation.
almost still near the elastic wall. Simulations by Grossman, Zhou, and Ben-Naim
(1997) show that both the pipe and the square box have
system can do one of two things. It can undergo inelastic a much smoother behavior than that of the one-
collapse. We put those collapse behaviors aside and dimensional case. No longer is there a single particle
study the noncollapse behavior. In this situation, the sys- that moves much faster than all of the others. Instead
tem splits into two parts. One part is composed of a there is a gradual and continuous falloff of temperature
single fast-moving particle near the hot wall; the other is as one moves away from the hot wall. It seems that this
a group of slower particles huddled together near the situation can, in fact, be described by using quasi-
reflecting wall. hydrodynamic equations connecting the density r, aver-
This situation is maintained through the collisions. age velocity u5(u,0), and temperature T. As usual, the
When the fast-moving particle, with speed of order one, hydrodynamics is given in terms of conservation laws
hits the first member of the clump, it transfers all its (Haff, 1983; Savage, 1988), which form equations for the
momentum, save a fraction «, to the slower particle and time derivatives of densities of conserved quantities r,
is itself slowed down to a speed of order «. This process momentum density5 r u, and energy density r T
happens again and again, with the fast momentum being 1 r u2 /2. In a steady situation the equation for mass con-
reduced by a factor (12«) n 'exp(2n«), after n colli- servation can be satisfied if the average velocity u is
sions. The momentum reaches the far wall, is reflected, zero. The equation for momentum conservation will be
and returns by the same process. Finally, one particle true if the pressure is independent of position. The final
with momentum reduced by a factor exp(22N«) moves equation, for energy conservation, has the form in which
leftward out of the pack. It is still hot, but leaves behind net heat flow from conduction is balanced against en-
it particles with speeds of order «. Thus the equilibrium ergy dissipation:
state of the system is one in which there is a clumping of
¹ ~ k ¹T ! 52 ~ energy dissipation rate! . (6)
particles, but one particle is left out of the clump. This
behavior could not have been predicted from any hydro- Here k is the thermal conductivity. Grossman, Zhou,
dynamic style of analysis. and Ben-Naim (1997) use kinetic theory to obtain a de-
One might have expected that the one-dimensional termination of the unknown functions in the theory,
problem would cause trouble. One can see this trouble namely, the dependence of the pressure, the dissipation
directly from the Boltzmann equation analysis of Mc- rate, and the thermal conductivity upon r and u. The
Namara and Young (1992) or of Grossman and Roman simulations show that neither T nor r is position inde-
(1996) or from a consideration of the basics of statistical pendent, with the first one rising and the other falling
mechanics. Recall that in statistical mechanics the equi- near the hot wall. The equation of state will determine
librium state of any system is described as an exponen- the pressure as a function of T and r, and the constancy
tial of linear combinations of conserved quantities. In of the pressure will give one relationship between r and
this exponential, there is one undetermined parameter T. These quantities will be fully determined when the T
for each conserved quantity. For one dimension and « equation [Eq. (6)] is solved with boundary conditions
50, each collision simply interchanges the velocities of provided by the known temperature on the hot surface
the colliding particles. Therefore the number of particles and a no-current-flow condition on the reflecting sur-
with any given value of the momentum is conserved. face. A few parameters are needed to generate kinetic
McNamara and Young (1992) set up a Boltzmann equa- theory results for the pressure, the thermal conductivity,
tion to describe this situation and then said that the and the energy dissipation via collisions. The net result
equilibrium solution of this Boltzmann equation is com- is a temperature that falls off as 1/cosh(an«1/2) away
posed of two pieces: any spatial distribution of particles from the hot plate, where n is the number of molecules
at rest and, on top of this, any distribution of particles in counted from the hot plate. Here « is the inelasticity
velocity, all of these velocities being uniformly distrib- defined here as (12r 2 )/2 while a is a coefficient of order
uted through space. The first piece corresponds to our unity. Correspondingly, as n increases the density rises
clump; the second to our single particle in motion. Thus and then saturates. These results fit the simulational
the «→0 limit does reduce to statistical mechanics, but data quite well.
to a very nonstandard statistical situation. The probability distributions for the velocity are not
Next consider the same situation in two dimensions. of Gaussian form. They are skewed by the large energy
Once again there is a hot left wall, a reflecting right wall, current that flows through the system. However, they

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

 Leo P. Kadanoff: Granular flows 443

sis of the sort done in the problem with one hot wall. We
are far from sure that any hydrodynamic description will
work even in this very simple situation. Thus we are left
with a distrust of uniform hydrodynamics as a possible
description for a granular system.

VI. CONCLUSIONS

Throughout this paper, our main question has been

Can a granular material be described by hydrodynamic
equations, most specifically those equations which apply
to an ordinary fluid? It seems to me that the answer is
‘‘no!’’ Glassy behavior is familiar, but it is not fully de-
scribed by any simple set of partial differential equa-
FIG. 8. Behavior of inelastic spheres confined between two tions. It is also not fully understood. In fact, there are a
hot walls. From Zhou and Kadanoff (1998 unpublished). The rich variety of familiar but weakly understood nonequi-
top picture shows the geometry and a disordered initial state. librium problems. These range from plastic flow to crack
The bottom picture is a sketch of the configuration reached propagation to charge density waves and to ordinary
after a long time. The particles near the wall produce a hot gas, and spin glasses. Granular materials form a particularly
with the temperature declining away from the hot walls. In the rich example of this kind of system, with behaviors
central region, there is a clump of material with very low rela- which are, at this moment, not fully understood.
tive velocities. This clump moves as a unit and acts as a Brown-
ian particle, which has been formed in the midst of an inelastic
gas. ACKNOWLEDGMENTS

The research work on this paper was supported by the

obey a sort of scaling law so that the velocity divided by
University of Chicago Materials Research Science and
AT has the same probability distribution throughout the
Engineering Center, the Office of Naval Research, and
sample. Experiments by Kudrolli and Gollub (1997) and
the National Science Foundation. I should like to thank
Kudrolli, Wolpert, and Gollub (1997), conducted with
my Chicago colleagues, particularly Eli Ben Naim, Peter
balls on a slanted table, show a qualitatively similar be-
Constantin, Susan Coppersmith, Konstantin Gavrilov,
havior to that found in the simulation. Thus, in a very
Elizabeth Grossman, Heinrich Jaeger, Sidney Nagel,
simple situation, hydrodynamics seems to work quite
Norbert Schörghofer, Thomas Witten, and Tong Zhou,
well.8 However, there is still room for some doubt. The
for their help in developing my ideas about granular ma-
authors forced the agreement by using two adjustable
terials. Many colleagues outside the University of Chi-
parameters, one for each wall. This adjustment might be
cago have also been of considerable help to me. I men-
necessary precisely because the system is far from an
tion particularly Jerry Gollub, Michel Louge, Daniel
equilibrium steady state, so that the usual conditions for
Rothman, and Troy Shinbrot, but there are many others.
the derivation of hydrodynamic equations do not hold.
Clumping can ruin the hydrodynamic description. In
very recent work, Zhou (1998) have been studying a REFERENCES
pipe in which both right and left walls are hot. In this
system, the majority of the particles clump up in a mass, Behringer, Robert, and James Jenkins, 1997, Eds., Powders &
which is separated from these walls by a smaller number Grains 97 (A.A. Balkema, Rotterdam).
Ben-Naim, E., J. B. Knight, E. R. Nowak, H. M. Jaeger, and S.
of fast-moving particles (see Fig. 8). This clump seems to
R. Nagel, 1998, Physica D, in press.
undergo Brownian motion, fed by momentum fluctua-
Bernu, B., and R. Mazighi, 1990, J. Phys. A 23, 5745.
tions that are provided by the particles added at the hot Boutreuz, T., and P. G. de Gennes, 1997, in Powders & Grains
wall. The speed of the correlated motion can be very 97, edited by Robert Behringer and James Jenkins (A.A.
much greater than the thermal velocity describing the Balkema, Rotterdam), pp. 439–442.
relative motion of the particles at the center of the Chapman, S., and T. G. Cowling, 1990, The Mathematical
clump. The clump contains most of the particles, which Theory of Non-Uniform Gases (Cambridge University Press,
can move all together and quite rapidly. We can use Cambridge, UK).
kinetic theory to calculate the motion of the clump, but Clément, E., and J. Rajchenbach, 1991, Europhys. Lett. 16,
this calculation includes the clump’s Brownian motion. 133.
That motion is left out in a purely hydrodynamic analy- Constantin, P., E. L. Grossman, and M. Mungan, 1995, Physica
D 83, 409.
de Gennes, P. G., 1997, in Powders & Grains 97, edited by R.
8
For another, earlier, investigation which supports two- Behringer and J. Jenkins (A.A. Balkema, Rotterdam), pp.
dimensional hydrodynamics, see Louge, Jenkins, and Hopkins 3–12.
(1993). A parallel publication, Louge (1994), suggests that Du, Yunson, Hao Li, and Leo P. Kadanoff, 1995, Phys. Rev.
there are some difficulties with boundary conditions in the hy- Lett. 74, 1268.
drodynamic formulation. Esipov, S. E., and T. Pöschel, 1997, J. Stat. Phys. 86, 1385.

Rev. Mod. Phys., Vol. 71, No. 1, January 1999

444 Leo P. Kadanoff: Granular flows

Faraday, Michael, 1831, ‘‘On a Peculiar Class of Acoustical Larouche, C., S. Douady, and S. Fauve, 1989, J. Phys. (France)
Figures; And on Certain Forms Assumed by Groups of Par- 50, 699.
ticles Upon Vibrating Elastic Surfaces,’’ Philos. Trans. R. Louge, M., 1994, Phys. Fluids 6, 2253.
Soc. London 52, 299. Logue, M. Y., J. T. Jenkins, and M. A. Hopkins, 1993, Mech.
Gallas, J. A. C., H. J. Hermann, and S. Sokolowski, 1992, Mater. 126, 199.
Physica A 189, 437. McNamara, S., and W. R. Young, 1992, Phys. Fluids A 4, 496.
Gavrilov, K. 1997, ‘‘Self-Organization of Uninterrupted Traffic McNamara, S., and W. R. Young, 1994, Phys. Rev. E 50, R28–
Flow,’’ in Powders & Grains 97, edited by Robert Behringer R31.
and James Jenkins (A.A. Balkema, Rotterdam), pp. 523–526. McNamara, S., and W. R. Young, 1996, Phys. Rev. E 53, 5089.
Gavrilov, K., 1998, Phys. Rev. E 58, 2107. Melo, F., P. B. Umbanhowar, and H. L. Swinney, 1994, Phys.
Goldhirsch, I., and G. Zanetti, 1993, Phys. Rev. Lett. 70, 1619. Rev. Lett. 72, 172.
Grossman, E. L., 1997a, Phys. Rev. E 56, 3290. Melo, F., P. B. Umbanhowar, and H. L. Swinney, 1995, Phys.
Grossman, E. L., 1997b, Ph.D. thesis (University of Chicago). Rev. Lett. 75, 3838.
Grossman, E. L., and B. Roman, 1996, Phys. Fluids 8, 3218. Metcalf, T., J. B. Knight, and H. M. Jaeger, 1997, Physica A
Grossman, E. L., T. Zhou, and E. Ben-Naim, 1997, Phys. Rev. 236, 202.
E 55, 4200. Nowak, E. R., J. B. Knight, E. Ben-Naim, H. Jaeger, and S. R.
Haff, P. K., 1983, J. Fluid Mech. 134, 401. Nagel, 1995, Phys. Rev. E 51, 3957.
Hayakawa, Hiaso, Su Yue, and Daniel C. Hong, 1995, Phys. Oda, M., K. Iwasbhita, and T. Kakiuchi, 1997, in Powders &
Rev. Lett. 75, 2328. Grains 97, edited by Robert Behringer and James Jenkins
Jaeger, H. M., and S. R. Nagel, 1992, Science 255, 1523. (A.A. Balkema, Rotterdam), pp. 207–211.
Jaeger, H. M., S. R. Nagel, and R. P. Behringer, 1996a, Phys. Pak, H. K., E. van Doorn, and R. P. Behringer, 1993, Phys.
Today 49(4), 32. Rev. Lett. 71, 1832.
Jaeger, H. M., S. R. Nagel, and R. P. Behringer, 1996b, Rev. Rapaport, D. C., 1980, J. Comput. Phys. 34, 184.
Mod. Phys. 68, 1259. Ristow, G. H., G. Strassburger, and I. Rehberg, 1997, Phys.
Jaeger, H. M., J. B. Knight, C.-H. Liu, and S. R. Nagel, 1994, Rev. Lett. 79, 833.
Mater. Res. Bull. 19, 25. Savage, S. B., 1988, J. Fluid Mech. 194, 457.
Jenkins, J. T., and S. B. Savage, 1983, J. Fluid Mech. 134, 187. Schörghofer, N., and T. Zhou, 1996, Phys. Rev. E 54, 5511.
Jenkins, J. T., 1992, J. Appl. Mech. 59, 120. Sela, N., and I. Goldhirsch, 1996, Phys. Fluids 7, 507.
Kadanoff, L. P., 1997, Phys. Today 50(9), 11. Shinbrot, T., D. Khakhar, J. J. McCarthy, and J. M. Ottino,
Knight, J. B., C. G. Fandrich, Chun-Ning Lau, H. M. Jaeger, 1997, Phys. Rev. Lett. 79, 829.
and S. R. Nagel, 1995, Phys. Rev. E 51, 3957. Umbanhower, B., F. Melo, and H. L. Swinney, 1996, Nature
Knight, J. B., H. M. Jaeger, and S. R. Nagel, 1993, Phys. Rev. (London) 382, 793.
Lett. 70, 3728. Venkataramani, S. C., 1998, unpublished.
Kudrolli, A., and J. P. Gollub, 1997, in Powders & Grains 97, Waller, M. D., 1961, Chlandi Figures: A Study in Symmetry (G.
edited by Robert Behringer and James Jenkins (A.A. Bell, London).
Balkema, Rotterdam), pp. 535–538. Zhou, Tong, 1998, unpublished.
Kudrolli, A., M. Wolpert, and J. P. Gollub, 1997, unpublished. Zhou, Tong, and L. P. Kadanoff, 1996, Phys. Rev. E 54, 623.

Rev. Mod. Phys., Vol. 71, No. 1, January 1999