Development of Stresses in Cohesionless Poured Sand

Development of stresses in
cohesionless poured sand

By M. E. Cates1 , J. P. W i t t m e r1 , J.-P. B o u c h a u d2 a n d P. Claudin2
1
Department of Physics and Astronomy, University of Edinburgh,
James Clerk Maxwell Building, King’s Buildings, Mayfield Road,
Edinburgh EH9 3JZ, UK
2
Service de Physique de l’Etat Condensé, CEA, Ormes des Merisiers,
91191 Gif-sur-Yvette Cedex, France
The pressure distribution beneath a conical sandpile, created by pouring sand from
a point source onto a rough rigid support, shows a pronounced minimum below the
apex (‘the dip’). Recent work by the authors has attempted to explain this phe-
nomenon by invoking local rules for stress propagation that depend on the local
geometry, and hence on the construction history, of the medium. We discuss the fun-
damental difference between such approaches, which lead to hyperbolic differential
equations, and elastoplastic models, for which the equations are elliptic within any
elastic zones present. In the hyperbolic case, the stress distribution at the base of a
wedge or cone (of given construction history), on a rough rigid support, is uniquely
determined by the body forces and the boundary condition at the free (upper) sur-
face. In simple elastoplastic treatments, one must in addition specify, at the base
of the pile, a displacement field (or some equivalent data). This displacement field
appears to be either ill-defined, or defined relative to a reference state whose physical
existence is in doubt. Insofar as their predictions depend on physical factors unknown
and outside experimental control, such elastoplastic models predict that the obser-
vations should be intrinsically irreproducible. This view is not easily reconciled with
the existing experimental data on conical sandpiles, which we briefly review. Our
hyperbolic models are based instead on a physical picture of the material, in which:
(1) the load is supported by a skeletal network of force chains (‘stress paths’) whose
geometry depends on construction history; (2) this network is ‘fragile’ or marginally
stable, in a sense that we define. Although perhaps oversimplified, these assumptions
may lie closer to the true physics of poured cohesionless grains than do those of con-
ventional elastoplasticity. We point out that our hyperbolic models can nonetheless
be reconciled with elastoplastic ideas by taking the limit of an extremely anisotropic
yield condition.
Keywords: granular materials; sandpiles; stress propagation; fragile matter
1. Introduction
Recently, a new strategy for the modelling of stress propagation in static cohesionless
granular media was developed (Bouchaud et al . 1995; Wittmer et al . 1996, 1997a, b).
The medium is viewed as an assembly of rigid particles held up by friction. The
static indeterminacy of frictional forces within the assembly is circumvented by the
assumption of certain local constitutive relations (CRs) among components of the
Phil. Trans. R. Soc. Lond. A (1998) 356, 2535–2561 c 1998 The Royal Society

Printed in Great Britain 2535 TEX Paper
2536 M. E. Cates, J. P. Wittmer, J.-P. Bouchaud and P. Claudin
stress tensor†. These are assumed to encode the network of contacts in the granular
packing geometry; they therefore depend explicitly on the way in which the medium
was made, i.e. its construction history. The task is then to postulate, and/or jus-
tify, physically suitable CRs among stresses, of which only one (the primary CR) is
required for systems with two-dimensional symmetry, such as a wedge of sand; for a
three-dimensional symmetric assembly (the conical sandpile) a secondary CR is also
needed.
Among the primary constitutive relations of Wittmer et al . (1996, 1997a) are a
certain class (called the ‘oriented stress linearity’ (OSL) models) which have simpli-
fying features. Indeed, in two-dimensional geometries these combine with the stress-
continuity equation to give a wave equation for stress propagation, in which the
horizontal and vertical directions play the role of spatial and temporal coordinates,
respectively (Bouchaud et al . 1995). A distinguishing feature of the OSL models is
that the characteristic rays for stress propagation (analogous to light or sound rays
in ordinary wave propagation) are then fixed by the construction history: they do not
change direction under subsequent reversible loading. (Irreversible loadings, which
can in these models be infinitesimal, are discussed in § 6.) As discussed by Bouchaud
et al . (1998), the characteristics of the differential equation can be viewed as repre-
senting, in the continuum, the mean behaviour of ‘force chains’ or ‘stress paths’ in
the material (Dantu 1967; Liu et al . 1995; Jäger et al . 1996; Thornton & Sun 1994).
Of the OSL models, a particularly appealing member, with special symmetry prop-
erties, is called the fixed principal axes (FPA) model. This has the additional property
that the characteristics everywhere coincide in orientation with the principal axes
of the stress tensor. The FPA model therefore supposes that these principal axes
have an orientation fixed at the time of burial‡. This is arguably the simplest possi-
ble choice for a history-dependent CR among stresses. For the case of a sandpile in
which grains are deposited by surface avalanches, which we presume to apply for a
conical pile constructed from a point source (though see § 5 c below), the orientation
of the major axis at burial is constant, and known from the fact that the free surface
in such a pile must be a yield surface. The resulting constitutive equation among
stresses, for the sandpile geometry, then has a singularity at the centre of a cone or
wedge; this is physically admissible since the centreline separates material which has
avalanched to the left from material which has avalanched to the right. This singu-
larity leads to an ‘arching’ effect, as previously invoked to explain the stress-dip by
Edwards & Oakeshott (1989) and others (Trollope 1968; Trollope & Burman 1980).
The OSL models were developed to explain experimental data on the stress distri-
bution beneath a conical sandpile, built by surface avalanches of sand, poured from a
point source onto a rough rigid support (Smid & Novosad 1981; Jokati & Moriyama
1979; Brockbank et al . 1997). Such data show, unambiguously, the presence of a
† In solid mechanics the term ‘constitutive relation’ normally refers to a material-dependent equation
relating stress and strain. In fluid mechanics one has instead equations relating stress and (in the general
case of a viscoelastic fluid) strain-rate history. Our models of granular media entail equations relating
stress components to one another, in a manner that we take to depend on the construction history of the
material. Clearly, such equations describe constitutive properties of the medium: they relate its state of
stress to other discernible features of its physical state. We see no alternative to the term ‘constitutive
relations’ for such equations. The same equations could, of course, be obeyed by some solutions of models
whose constitutive definition was quite different; in that context they would not be CRs.
‡ Gudehus (1974) used a related idea, that the principal axes should be locally specified as inputs
to the stress-continuity equations. This he employs as a calculation method for generating a variety of
stress distributions based on ‘gutem statischen Gefühl’.
Phil. Trans. R. Soc. Lond. A (1998)

Stresses in poured sand 2537
Figure 1. Comparison of FPA model using a uniaxial secondary closure (Wittmer et al . 1996,
1997a) with scaled experimental data of Smid & Novosad (1981) and (*) that of Brockbank
et al . (1997), which was averaged over three piles. Upper and lower curves denote normal and
shear stresses. The data are used to calculate the total weight of the pile, which is then used as
a scale factor for stresses. The horizontal coordinate is scaled by the pile radius.
minimum (‘the pressure dip’) in the vertical normal stress below the apex of the
pile. With a plausible choice of secondary CR (of which several were tried, with only
minor differences resulting), the FPA case, in particular, was found to give a fairly
good quantitative account of the data of Smid & Novosad (1981), and of Brockbank
et al . (1997); see figure 1. This is remarkable, in view of the radical simplicity of the
assumptions made.
We accept, of course, that such models may be valid over only a limited regime in
some larger parameter space. For example, since strain variables are not introduced,
these models cannot, of themselves, examine the crossover to conventional elastic
or elastoplastic behaviour that must presumably arise when the applied stresses are
significant on the scale of the elastic modulus of the grains themselves. Some further
remarks on this crossover, in the context of anisotropic elastoplasticity, are made in
§ 3 d.
In this paper we discuss the physical content of our general modelling approach
(of which the FPA model is one example), based on local stress-propagation rules
that depend on construction history, as encoded in constitutive relations among
stresses. In particular we contrast the approach with more conventional ideas, espe-
cially the ideas of elastoplasticity. For simplicity, our mathematical discussion is
mainly limited to two dimensions (although our models were developed to describe
three-dimensional piles) and to the simplest, isotropic forms of elastoplastic theory.

The discussion aims to sharpen some conceptual issues. These concern not the details
of particular models, but the general question of what sort of description we should
aspire to: what sort of information do we need as modelling input, and what can be
predicted as output? An equally important (and closely related) question is: what are
the control variables in an experiment that must be specified to ensure reproducible
behaviour, and what are the observables that can then be measured to depend on
these? For experiments on sandpiles (briefly reviewed in § 5), we believe these to
be open physics questions, and to challenge some widely held assumptions of the
applicability of traditional elastoplastic modelling strategies to cohesionless poured
grains.
The proposal that granular assemblies under gravity cannot properly be described
by the ideas of conventional elastoplasticity has been opprobiously dismissed in some
quarters: we stand accused of ignoring all that is ‘long and widely known’ among
geotechnical engineers (Savage 1997). However, we are not the first to put forward
such a subversive proposal. Indeed workers such as Trollope (1968) and Harr (1977)
long ago developed ideas of force-transfer rules among discrete particles, not unre-
lated to our own approaches, which yield continuum equations quite unlike those of
elastoplasticity. More recently, dynamical hypoplastic continuum models have been
developed (Kolymbas 1991; Kolymbas & Wu 1993) which, as explained by Gudehus
(1997) describe an ‘anelastic behaviour without [the] elastic range, flow conditions
and flow rules of elastoplasticity’. Our own models, though not explicitly dynamic,
are similarly anelastic, as we discuss in § 6. They should perhaps be classified as
hypoplastic models, although their relation to extremely anisotropic elastoplastic
models is examined in § 3 d.
2. Continuum models of cohesionless granular matter

Initially we review (in their simplest forms) some well-known modelling approaches
based on rigid-plastic and elastoplastic ideas. This is followed by a brief summary of
the mathematical content of the FPA model and its relatives.
(a) Stress continuity and the Coulomb inequality

The equations of stress continuity express the fact that, in static equilibrium, the
forces acting on a small element of material must balance. For a conical pile of sand
we have, in d = 3 dimensions,
∂r σrr + ∂z σzr = β(σχχ − σrr )/r, (2.1)
∂r σrz + ∂z σzz = g − βσrz /r, (2.2)
∂χ σij = 0, (2.3)
where β = 1. Here z, r and χ are cylindrical polar coordinates, with z the downward
vertical. We take r = 0 as a symmetry axis, so that σrχ = σzχ = 0; g is the force
of gravity per unit volume; σij is the usual stress tensor which is symmetric in i, j.
The equations for d = 2 are obtained by setting β = 0 in (2.1), (2.2) and suppressing
(2.3). These describe a wedge of constant cross-section and infinite extent in the
third dimension. They also describe a layer of grains in a thin, upright Hele–Shaw
cell, but only if the wall friction is negligible.

The Coulomb inequality states that, at any point in a cohesionless granular me-
dium, the shear force acting across any plane must be smaller in magnitude than
tan φ times the compressive normal force. Here φ is the angle of friction, a material
parameter which, in simple models, is equal to the angle of repose. We accept this
here, while noting that (i) φ in principle depends on the texture (or fabric) of the
medium and hence on its construction history; (ii) for a highly anisotropic packing,
the existence of an orientation-independent φ is questionable (see § 3 d); and (iii)
the identification of φ with the repose angle ignores some complications such as the
Bagnold hysteresis effect.
(b) Rigid-plastic models

The model that Wittmer et al . (1996, 1997a) refer to as ‘incipient failure every-
where’ (IFE), is more commonly called the rigid-plastic model. It postulates that
the Coulomb condition is everywhere obeyed with equality (Nedderman 1992). That
is, through every point in the material there passes some plane across which the
shear force is exactly tan φ times the normal force. By assuming this, the IFE model
allows closure (modulo a sign ambiguity discussed below) of the equations for the
stress without invocation of an elastic strain field. The IFE model has therefore, as
its ‘constitutive relation’ (Wittmer et al . 1997a),
1 2
p
σrr = σzz [sin φ + 1 ± 2 sin φ 1 − (cot φσzr /σzz )2 ], (2.4)
cos2 φ
whereas the Coulomb inequality requires only that σrr lies between the two values
(±) on the right.
The postulate that a Coulombic slip plane passes through each and every material
point is not usually viewed as being accurate in itself; the rigid-plastic model is more
often proposed as a way of generating certain ‘limit-state’ solutions to an underlying
elastoplastic model. In the simplest geometries, these solutions correspond to taking
the − or + sign in (2.4). It is a simple exercise to show that for a sandpile at its repose
angle, only one solution of the resulting equations exists in which the sign choice is
everywhere the same. This requires the negative root (conventionally referred to as
an ‘active’ solution) and it shows a hump, not a dip, in the vertical normal stress
beneath the apex. Savage (1997), however, draws attention to a ‘passive’ solution,
having a pronounced dip beneath the apex. This solution actually contains a pair of
matching planes between an inner region where the positive root of (2.4) is taken,
and an outer region where the negative is chosen.
In principle there is more than one such ‘passive’ solution. For example, one can
seek an IFE solution in which all stress components are continuous across the match-
ing plane. This requires a discontinuity in the gradients of the stresses at the centre-
line (see figure 2). The latter does not contradict equation (2.4), although it might be
thought undesirable on other grounds (for example if the IFE equation is thought to
bound the behaviour of a simple elastoplastic body, for which the resulting displace-
ment fields might not be admissible). An alternative, which avoids this, is to instead
have a discontinuity of the normal stress parallel to the matching plane itself. This
gives a second passive solution (Savage 1997, 1998). These solutions do not exhaust
the repertoire of IFE solutions for the sandpile: there is no physical principle that
limits the number of matching surfaces. By adding extra ones, a very wide variety
of results can be achieved.

Figure 2. Vertical normal stress found for the BCC model described in § 2 d, for a pile at angle
of repose φ = 30◦ , compared to the active and two passive IFE solutions (obtained by shoot-
ing from the midplane) discussed in the text. (For numerical reasons the continuous IFE uses
P = 12 (σzz + σxx ) and the polar angle θ as functions of the direction of the principal axis Ψ .)
Note that active and passive IFE solutions do not bound the stress, either in the BCC model or
in the simple elastoplastic model of Cantelaube & Goddard (1997), which, for a certain param-
eter choice, yields identical results. The two-dimensional FPA solution is also included (dotted
line).
The emphasis placed on the rigid-plastic approach, at least in some parts of the
literature, seems to rest on a misplaced belief that the limit-state solutions can be
‘generally regarded as bounds between which other states can exist, i.e. when the
material is behaving in an elastic or elastoplastic manner’ (Savage 1997). A simple
counter-example is shown in figure 2. This shows the active and two passive solutions
(as defined above) for a two-dimensional pile (wedge), along with an elementary
elastoplastic solution as presented recently by Cantelaube & Goddard (1997) and
earlier by Samsioe (1955). The latter is piecewise linear with no singularity in the
displacement field on the central axis; it happens to coincide mathematically with the
solution of a simple hyperbolic model (Bouchaud et al . 1995) for the same geometry
(there is no stress dip in this particular model). Clearly, the vertical normal stress
does not lie everywhere between that of the active and passive IFE solutions, which
are therefore not bounds.
(c) Elastoplastic models

In two dimensions at least, it has been argued that the pressure dip can be
explained within a simple conventional elastoplastic modelling approach. This is

certainly possible if the base is allowed to sag slightly (Savage 1997). Here, however,
we are concerned with piles built on a rough rigid support. Even in this case, it has
been argued that results similar to those of the FPA model can be obtained (Can-
telaube & Goddard 1997). The simplest elastoplastic models assume a material in
which a perfectly elastic behaviour at small strains is conjoined onto perfect plastic-
ity (the Coulomb condition with equality) at larger ones. In such an approach to the
sandpile, an inner elastic region is matched, effectively by hand, onto an outer plastic
one. In the inner elastic region the stresses obey the Navier equations, which follow
from those of Hookean elasticity by elimination of the strain variables. The corre-
sponding strain field is not discussed, but tacitly treated as infinitesimal, since the
high-modulus limit is taken. However, for FPA-like solutions, which show a cusp in
the vertical stress on the centreline, the displacement shows singular features which
are not easily reconciled with a purely elastoplastic interpretation. The fact that
the plastic zone is introduced ad hoc also has drawbacks; for example, it is hard to
explain the continued presence of such a zone if the angle of the pile is reduced to
slightly below the friction angle φ (to allow for the Bagnold hysteresis effect, say). In
OSL approaches, an outer zone is not assumed but predicted, and remains present
in this case, although the material in this zone is no longer at incipient failure.
The existence of FPA-like solutions to simple elastoplastic models in three dimen-
sions, on a non-sagging support, remains very much in doubt. But in any case, our
objections to the elastoplastic approach to modelling sandpiles lie at a more fun-
damental level. Specifically, it appears that, to make unambiguous predictions for
the stresses in a sandpile, these models require boundary information that has no
obvious physical meaning or interpretation. We return to this physics problem in
§ 3 b.
(d ) Local-rule models of stress propagation

In the FPA model (Wittmer et al . 1996, 1997a) one hypothesizes that, in each
material element, the orientation of the stress ellipsoid became ‘locked’ into the
material texture at the time when it last came to rest, and does not change in
subsequent loadings (unless forced to: see § 6). This is a bold simplifying assumption,
and it may indeed be far too simple, but it exemplifies the idea of having a local
rule for stress propagation that depends explicitly on construction history. For the
sandpile geometry, where the material comes to rest on a surface slip plane, this
constitutive hypothesis leads to the following relation among stresses:
σrr = σzz − 2 tan φσzr , (2.5)
where φ is the angle of repose. Equation (2.5) is algebraically specific to the case of
a pile created from a point source by a series of avalanches along the free surface.
A consequence of equation (2.5), for a pile at repose, is that the major principal
axis everywhere bisects the angle between the vertical and the free surface. It should
be noted that in Cartesian coordinates, the FPA model reads:
σxx = σzz − 2 sgn(x) tan φσxz , (2.6)
where x is horizontal. From equation (2.6), the FPA constitutive relation is seen to
be discontinuous on the central axis of the pile: the local texture of the packing has
a singularity on the central axis which is reflected in the stress-propagation rules of
the model. The paradoxical requirement, on the centreline, that the principal axes

are fixed simultaneously in two different directions has a simple resolution: the stress
tensor turns out to be isotropic there.
The FPA model is one of the larger class of OSL models in which the primary
constitutive relation (in the sandpile geometry) is, in Cartesian coordinates,
σxx = ησzz + µ sgn(x)σxz , (2.7)
with η, µ constants. As explained by Wittmer et al . (1997a), these models (in two
dimensions) yield hyperbolic equations that have fixed characteristic rays for stress
propagation. (Unless µ = 0, these are asymmetrically disposed about the vertical
axis, and invert discontinuously at the centreline x = 0.) The constitutive property
that OSL models describe, is that these characteristic rays, or force chains, have
orientations that are ‘locked in’ on burial of an element. The boundary condition, that
the free surface of a pile at its angle of repose, φ, is a slip plane, yields one equation
linking η and µ to φ; thus the OSL scheme represents a one-parameter family of
models. Note that, as soon as η is not exactly unity (the FPA case) the orientation
of the principal axes rotates smoothly as one passes through the centreline of the
pile. The assumption of fixed principal axes, though appealing, is thus rather delicate,
and arguably much less important than the idea of fixed characteristics, since these
represent the average geometry of force chains in the medium. The experimental data
(figure 1) support models in the OSL family with η close, but perhaps not exactly
equal, to unity.
Note that, unless the OSL parameter is chosen so that µ = 0, a constitutive
singularity on the central axis remains. The case µ = 0 corresponds to one studied
earlier by Bouchaud et al . (1995); this ‘BCC’ model is the only member of the OSL
family to have characteristics symmetric about the vertical. (Their angles ±θ to the
vertical obey tan2 θ = c20 = η.) This latter model could be called a ‘local Janssen
model’ in that it assumes local proportionality of horizontal and vertical compressive
stresses—an assumption which, when applied globally to average stresses in a silo,
was first made by Janssen (1895).
The local-rule models just discussed do not account for the presence of ‘noise’ or
randomness in the local texture. Such effects have been studied by Claudin et al .
(1998), and, if the noise level is not too large, lead at large length-scales to effective
wavelike equations with additional gradient terms giving a diffusive spreading of
the characteristic rays. The limit where the diffusive term dominates corresponds to
a parabolic differential equation (Harr 1977), similar to those arising in scalar force
models (Liu et al . 1995) which have, in effect, a single downward characteristic (so the
main interest lies in the diffusive spreading). It is possible that under extreme noise
levels, this picture changes again, although this conclusion is based on assumptions
about the noise itself that may not be valid (Claudin et al . 1998). The discussions
that follow therefore apply to local-rule models with moderate, but perhaps not
extreme, noise.
Note, finally, that the fact that two continuum models, based on different constitu-
tive hypotheses, can give identical results for the stresses in some specified geometry,
obviously does not mean that the models are equivalent. (Equivalence requires, at
least, that the Green function of the two models is also the same.) Thus, for example,
models such as FPA are not equivalent to Trollope’s model of ‘clastic discontinua’
(Trollope 1968; Trollope & Burman 1980). In Appendix B we outline the relationship
between our work and the marginal packing models studied by R. Ball & S. Edwards

(personal communication), Huntley (1993), Hong (1993), Bagster (1978) and Liff-
mann et al . (1992), as well as the work of Trollope (1968).
3. Strain and displacement variables

In the FPA model and its relatives, strain variables are not considered. A partial jus-
tification for this was given by Wittmer et al . (1997a), namely that the experimental
data obey a form of radial stress-field (RSF) scaling: the stress patterns observed at
the base are the same shape regardless of the overall height of the pile. Formally, one
has for the stresses at the base
σij = ghsij (r/ch), (3.1)
where h is the pile height, c = cot α and sij is a reduced stress: α is the angle between
the free surface and the horizontal so that for a pile at repose, α = φ. This form of
RSF scaling, which involves only the forces at the base (Evesque 1997), might be
called the ‘weak form’ and is distinct from the ‘strong form’ in which equation (3.1)
is obeyed also with z (an arbitrary height from the apex) replacing h (the overall
height of the pile).
This scaling implies that there is no characteristic length-scale. Since elastic defor-
mation introduces such a length-scale (the length-scale over which an elastic pile
would sag under gravity) the observation of RSF scaling to experimental accuracy
suggests that elastic effects need not be considered explicitly. We accept, however,
(correcting Wittmer et al . (1997a)) that this does not of itself rule out elastic or
elastoplastic behaviour which, at least in the limit of large modulus, can also yield
equations for the stress from which the bulk strain fields cancel. Note that it is
tempting, but entirely wrong, to assume that a similar cancellation occurs at the
boundaries of the material; we return to this below (§ 4).
The cancellation of bulk strain fields in elastoplastic models is convenient, since
there appears to be no clear definition of strain or displacement for piles constructed
by pouring sand grainwise from a point source. To define a physical displacement or
strain field, one requires a reference state. In (say) a triaxial strain test (see, for exam-
ple, Wood 1990), an initial state is made by some reproducible procedure, and strains
measured from there. The elastic part is identifiable, in principle, by removing the
applied stresses (maintaining an isotropic pressure), and seeing how much the sample
recovers its shape. In contrast, a pile constructed by pouring grains onto its apex
cannot convincingly be described in terms of the plastic and/or elastic deformation
from some initial reference state of the same continuous medium: the corresponding
experiments are unrealizable. Even were the load (which consists purely of gravity)
to be removed, the resulting unstrained body would comprise grains floating freely
in space with no definite positions. It is unsatisfactory to define a strain or displace-
ment field with respect to such a body. The problem occurs whenever the solidity
of the body itself only arises because of the load applied. A similar situation occurs,
for example, in colloidal suspensions that flow freely at small shear stresses but (by
jamming) can support larger ones indefinitely (Cates et al . 1998).
Although one cannot uniquely define the strain in a granular assembly under
gravity, it may of course be possible to define incremental strains in terms of the
displacement of grains when a small load is added. However, the range of stress
increments involved might in practice be negligible (Kolymbas 1991; Gudehus 1997).

4. The role of boundary conditions

(a) Boundary conditions in hyperbolic (and parabolic) models
Models that assume local constitutive equations among stresses (including all OSL
models, and also the IFE or rigid-plastic model) provide hyperbolic differential equa-
tions for the stress field. Accordingly, if one specifies a zero-force boundary condition
at the free (upper) surface of a wedge, then any perturbation arising from a small
extra body force (a ‘source term’ in the equations) propagates along two charac-
teristics passing through this point. (In the OSL models, these characteristics are,
moreover, straight lines.) Therefore, the force at the base can be found simply by
summing contributions from all the body forces as propagated along two characteris-
tic rays onto the support; the sandpile problem is, within the modelling approach by
Bouchaud et al . (1995) and Wittmer et al . (1996, 1997a), mathematically well-posed.
Note that in principle, one could have propagation also along the ‘backward’ char-
acteristics (see figure 3a). This is forbidden since these cut the free surface; any such
propagation can only arise in the presence of a non-zero surface force, in violation of
the boundary conditions. Therefore, the fact that the propagation occurs only along
downward characteristics is not related to the fact that gravity acts downward; it
arises because we know already the forces acting at the free surface (they are zero).
Suppose we had, instead, an inverse problem: a pile on a bed with some unspeci-
fied overload at the top surface, for which the forces acting at the base had been
measured. In this case, the information from the known forces could be propagated
along the upward characteristics to find the unknown overload. More generally, in
OSL models of the sandpile, each characteristic ray will cut the surface of a (convex)
patch of material at two points. Within these models, the sum of the forces tangen-
tial to the ray at the two ends must be balanced by the tangential component of the
body force integrated along the ray (see figure 3b). We discuss this physics (that of
force chains) in § 6 a.
In three dimensions, the mathematical structure of these models is somewhat
altered (Bouchaud et al . 1995), but the conclusions are basically unaffected. Note,
however, that for different geometries, such as sand in a bin, the problem is not well-
posed even with hyperbolic equations, unless something is known about the interac-
tion between the medium and the sidewalls. Ideally, one would like an approach in
which sidewalls and base were treated on an equal basis; this is the subject of ongo-
ing research. Note also that the essential character of the boundary-value problem
is not altered when appropriate forms of randomness are introduced. For although
the response to a point force is now spread about the two characteristics, even in
the parabolic limit (where the underlying straight rays are effectively coincident and
only spreading remains) the sandpile boundary-value problem remains well-posed.
(b) The physics of elastic indeterminacy

The well-posedness of the sandpile does not extend to models involving the elliptic
equations for an elastic body. For such a material, the stresses throughout the body
can be solved only if, at all points on the boundary, either the force distribution or a
displacement field is specified (Landau & Lifshitz 1986). Accordingly, once the zero-
stress boundary condition is applied at the free surface, in principle nothing can be
calculated, unless either the forces or the displacements at the base are already known

Figure 3. (a) The response to a localized force is found by resolving it along characteristics
through the point of application, propagating along those which do not cut a surface on which
the relevant force component is specified. For a pile under gravity, propagation is only along
the downward rays. (b) Admissible boundary conditions cannot specify separately the force
component at both ends of the same characteristic. If these forces are unbalanced (after allowing
for body forces), static equilibrium is lost.
(and the former amounts to specifying in advance the solution of the problem). From
an elastoplastic perspective, it is clearly absurd to try to calculate the forces on the
support, which are the experimental data, without some further information about
what is happening at the bottom boundary. We have called this the problem of ‘elastic
indeterminacy’ (Bouchaud et al . 1998) although perhaps ‘elastic ill-posedness’ would
be a better term. The problem does not arise from any uncertainty about what to
do mathematically: one should specify a displacement field at the base. Difficulties
nonetheless arise if, as we argued above, no physical significance can be attributed
to this displacement field for cohesionless poured sand.
To give a specific example, consider the static equilibrium of an elastic cone or
wedge of finite modulus resting on a completely rough rigid surface (which one could
visualize as a set of pins; figure 4). Starting from any initial configuration, another
can be generated by pulling and pushing parts of the body horizontally across the
base (i.e. changing the displacements there); if this is rough, the new state will still
be pinned and will achieve a new static equilibrium. This will generate a stress
distribution, across the supporting surface and within the pile, that differs from the
original one. If a large enough modulus is now taken (at fixed forces), this procedure
allows one to generate arbitrary differences in the stress distribution while generating
neither appreciable distortions in the shape of the cone, nor any forces at its free
surface. Analogous remarks apply to any simple elastoplastic theory of sandpiles,
in which an elastic zone, in contact with part of the base, is attached at matching
surfaces to a plastic zone.
In contrast, experimental reports (reviewed in § 5) indicate that for sandpiles on a
rough rigid support, the forces on the base can be measured reproducibly. They also
suggest that these forces, although subject to statistical fluctuations on the scale

Figure 4. Starting from an elastic cone or wedge on a rough support, any initial stress distribution
can be converted to another by displacements with respect to the rough ‘pinning’ surface (a) →
(b). Taking the limit of a high modulus (b) → (c) at fixed surface forces, an arbitrary stress
field remains, while recovering the initial shape of the cone and satisfying the free surface
boundary conditions. This shows the physical character of ‘elastic indeterminacy’ for an elastic
or elastoplastic body on a rough support.
of several grains, do not vary too much from one pile to another, at least among
piles constructed in the same way (e.g. by avalanches from a point source), from
the same material. This argues strongly against the idea that such forces in fact
depend on a basal displacement field, which is determined either by the whim of
the experimentalist, or by some as-yet unspecified physical mechanism acting at the
base of the pile. Note that basal sag is not a candidate for the missing mechanism,
since it does not resolve the elastic indeterminacy in these models; the latter arises
primarily from the roughness, rather than the rigidity, of the support. Note also,
however, that elastic indeterminacy can be alleviated in practice if the elastoplastic
model is sufficiently anisotropic; we return to this point in § 3 d.
Evesque (personal communication), unlike many authors, does confront the issue
of elastic indeterminacy and seemingly concludes that the experimental results are
and must be indeterminate; he argues that the external forces acting on the base of a
pile can indeed be varied at will by the experimentalist, without causing irreversible
rearrangements of the grains (see also Evesque & Boufellouh 1997). To what extent
this viewpoint is based on experiment, and to what extent on an implicit presumption
in favour of elastoplastic theory, is to us unclear.
(c) Displacements to the rescue?

Let us boldly suppose, then, that the experimental data are meaningful and repro-
ducible, at least as far as the global, ‘coarse-grained’ features of the observations
are concerned. (Noise effects at the level of individual grains may, in contrast, be
exquisitely sensitive to temperature and other poorly controlled parameters; Claudin

& Bouchaud (1997).) Adherents to traditional elastoplastic models then have three
choices. The first is to consider the possibility that, after all, the problem of cohe-
sionless poured sand may be better described by quite different governing equations
from those of simple elastoplasticity. This possibility, which represents our own view,
has certainly been suggested before. For example, hypoplastic models in which there
is negligible elastic range (Gudehus 1997; Kolymbas 1991; Kolymbas & Wu 1993) do
not suffer from elastic indeterminacy.
The second choice is to postulate various additional constraints, so as to eliminate
some of the infinite variety of solutions that elastoplastic models allow (unless basal
displacements are specified). For example, it is tempting to impose (in its strong
form) RSF scaling: for a wedge, as shown by Samsioe (1955) and Cantelaube &
Goddard (1997) this reduces the admissible solutions to a piecewise-linear form. Such
a postulate may seem quite harmless: after all, we have emphasized already that the
observations do themselves show (weak) RSF scaling. However, according to these
models, the central part of the pile can correctly be viewed as an elastic continuum;
hence, from any solution for the stresses, it should be physically possible to generate
another by an infinitesimal pushing and pulling of the elastic material along the rough
base. Accordingly, one has no reason to expect even the weak RSF scaling observed
experimentally. Setting this aside, one could impose weak RSF scaling by assuming
a basal displacement field of the same overall shape for piles of all sizes. However,
as pointed out by Evesque (1997), even this imposition does not require the strong
form of RSF scaling assumed by Cantelaube & Goddard (1997). In summary, simple
elastoplastic models of sandpiles require that the experimental results for the force
at the base depend on how the material was previously manipulated. Any attempt
to predict the forces without specifying these manipulations is misguided.
A third reaction, therefore, is to start modelling explicitly, the physical processes
going on at the base of the pile. As mentioned previously, one is required to specify
a displacement field at the base of the elastic zone; more accurately, it is the product
of the displacement field and the elastic modulus that matters. This need not vanish
in the large modulus limit (§ 3 b); one possible choice, nonetheless, is to set the
displacement field to zero at a finite modulus (which might then be taken to infinity).
The simplest interpretation of this choice is by appeal to a model in which the
‘sandpile’ is constructed as follows (figure 5a): an elastoplastic wedge, floating freely
in space, is brought to rest in contact with a rough surface, in a state of zero strain.
Once in contact, gravity is switched on with no further adjustments in the contact
region allowed. This might be referred to as the ‘spaceship model’ (or perhaps the
‘floating model’) of a sandpile. This illustrates two facts: (1) in considering explicitly
the displacement field at the bottom surface, elastoplastic modellers are obliged
to make definite assumptions about the previous history of the material; (2) these
assumptions do not usually have much in common with the actual construction
history of a sandpile made by pouring. A possible alternative to the spaceship model,
in which unstressed laminae of elastoplastic material are successively added to an
existing pile (figure 5b) is discussed in Appendix A.
(d ) Determinacy and anisotropy

We shall now show that hyperbolic behaviour can be recovered from an elasto-
plastic description by taking a strongly anisotropic limit (Cates et al . 1998). For
simplicity we restrict attention to the FPA model.

Figure 5. (a) The ‘spaceship’ model of a sandpile. An unstrained isotropic elastoplastic cone
or wedge is brought into contact with a rough surface and gravity then switched on. (b) ‘Lam-
inated elastoplastic model’ of a sandpile. Layers are added in a state of zero stress (thereby,
it is argued, zero displacement) to a pre-existing, gravitationally loaded pile. Such a pile (if
gravity is removed) will spring into a new shape, characterized by a non-zero internal stress field
(Appendix B).
The FPA model describes, by definition, a material in which the shear stress must
vanish across a pair of orthogonal planes fixed in the medium; those normal to the
(fixed) principal axes of the stress tensor. According to the Coulomb inequality, which
the model also adopts, the shear stress must also be less than tan φ times the normal
stress, across planes orientated in all other directions. Clearly, this combination of
requirements can be viewed as a limiting case of an elastoplastic model with an
anisotropic yield condition:
|σtn | 6 σnn tan Φ(θ), (4.1)
where θ is the angle between the plane normal, n, and the vertical (say). An aniso-
tropic yield condition should arise, in principle, in any material having a non-trivial
fabric, arising from its construction history. The limiting choice corresponding to the
FPA model for a sandpile is Φ(θ) = 0 for θ = 14 (π − 2φ) (this corresponds to planes
where n lies parallel to the major principal axis), and Φ(θ) = φ otherwise. (There is
no separate need to specify the second, orthogonal plane across which shear stresses
vanish, since this is assured by the symmetry of the stress tensor.) By a similar
argument, all other OSL models can also be cast in terms of an anisotropic yield
condition, of the form |σtn − σnn tan Ψ (θ)| 6 σnn tan Φ(θ), where Φ(θ) vanishes, and
Φ(θ) is finite for two values of θ. (This fixes a non-zero ratio of shear and normal
stresses across certain special planes.)
At this purely phenomenological level, there is no difficulty in connecting hyper-
bolic models smoothly onto highly anisotropic elastoplastic descriptions. Specifically,

consider a medium having an orientation-dependent friction angle Φ(θ) that does not
actually vanish, but is instead very small (less than or equal to , say) in a narrow
range of angles (say of order ) around θ = 14 (π − 2φ), and approaches φ elsewhere.
(One interesting way to achieve the required yield anisotropy is to have a strong
anisotropy in the elastic response, and then to impose a uniform yield condition the
strains, rather than stresses.)
Such a material will have, in principle, mixed elliptic/hyperbolic equations of the
usual elastoplastic type. The resulting elastic and plastic regions must nonetheless
arrange themselves so as to obey the FPA model to within terms that vanish as → 0.
If is small but finite, then for this elastoplastic model the results will depend on
the basal boundary condition, but only through these higher-order corrections to the
leading (FPA) result. We show in § 6 that the case of small but finite is exactly
what one would expect if a small amount of particle deformability were introduced
to a fragile skeleton of rigid particles obeying the FPA constitutive relation.
Although somewhat contrived (from an elastoplastic standpoint), the above choice
of anisotropic yield condition establishes an important point of principle, and may
point toward some important new physics. Although elastoplastic models do suffer
from elastic indeterminacy (they require a basal displacement field to be specified),
the extent of the influence of the boundary condition on the solution depends on
the model chosen. Strong enough (fabric-dependent) anisotropy, in an elastoplastic
description, might so constrain the solution that, although it suffers elastic indeter-
minacy in principle, it does so only harmlessly in practice. Under such conditions it
is primarily the fabric, and only minimally the boundary conditions, which actually
determine the stresses in the body. For models such as that given above there is a
well-defined limit where the indeterminacy is entirely lifted, hyperbolic equations are
recovered, and it is quite proper to talk of local stress-propagation ‘rules’ which are
determined, independently of boundary conditions, by the fabric (hence construction
history) of the material.
Our modelling framework, based precisely on these assumptions, will be valid for
sandpiles if, as we contend, their physics lie close to this limit of ‘fabric dominance’
(see § 6). This contention is consistent with, though it does not require, belief in the
existence of an underlying elastoplastic continuum description.
5. Experimental results
Before discussing in more detail the physical interpretation of our models, we give
a brief account of the experimental data. In doing this, it is important to draw
a distinction between (axially symmetric) cones, and (translationally symmetric)
wedges of sand. The latter is a quasi-two-dimensional geometry. The main question
is, to what extent the pressure-dip can be trusted as a reproducible experimental
phenomenon for a sandpile constructed by pouring onto a rough rigid support. In
particular, Savage (1997) has drawn attention to the possible role of small deflections
in the base (‘basal sag’) in causing the dip to arise.
(a) Cones
The earliest data we know of, on conical sandpiles, are that of Hummel & Finnan
(1920), who observed a pronounced stress dip. However, their pressure cells were

apparently subject to extreme hysteresis, and these results cannot be relied upon.
Otherwise, the only data before Smid & Novosad (1981) for cones are those of Jokati
& Moriyama (1979). Although a stress dip is repeatedly observed by these authors,
their results (on rather small piles) do not show consistent RSF scaling. The well-
known data of Smid & Novosad (1981) show a clear stress minimum at the centre of
the pile. Even this data-set is not completely satisfactory: the observation of the dip
is based on the data from a single (but calibrated) pressure cell beneath the apex.
However, the data for different pile heights show clear (weak) RSF scaling, and are
quantitatively fit by the FPA model with either of the secondary closures shown in
figure 1. Savage (1997) points out that ‘it is not possible from the information given
to estimate the deflections [at the base] that might result from the weight of the
pile’. Smid & Novosad (1981), however, describe their platform as ‘rigid’†.
Recently, Brockbank et al . (1997) have performed a number of careful measure-
ments on relatively small piles of sand (as well as flour, glass beads, etc.). The pressure
transducers comprise an assembly of steel ball-bearings lying atop a thin blanket of
transparent rubber on a rigid glass plate; material is poured from a point source
onto this assembly. The deflection of the ball-bearings is estimated as 10 µm. By
calibrating and optically monitoring their imprints on the rubber film, the vertical
stresses can be measured. Perhaps the most interesting feature of this method is that,
although the basal deflection is certainly not zero, it is of a character quite unlike
basal sag. Indeed, the supporting ball-bearings are deflected downward (indenting
the rubber film) in a manner that depends on the local compressive stress, as opposed
to the cumulative (i.e. non-local) effect of sagging. The latter is bound to be maximal
under the apex of the pile, whereas the indentation is maximal under the zone of
maximum vertical compressive stress, wherever that may be. If the stress pattern is
controlled by slight deformations of the base, there would be no reason to expect a
similar stress pattern to arise for an indentable base, as for a sagging one.
But in fact, a very similar stress pattern is seen (figure 1). The data shown here
involve averaging over several piles, since the set-up measures stresses over quite small
areas of the base (the ball bearings are 2.5 mm in diameter) and these stresses fluc-
tuate locally, as is well known (Liu et al . 1995; Claudin & Bouchaud 1997). Although
still subject to relatively large statistical scatter, the data show an unambiguous dip
of very similar magnitude to that reported by Smid & Novosad (1981); moreover,
the dip is spread over several transducers, rather than a single transducer.
It is, of course, important to distinguish conceptually the noisiness of these data
(arising from fluctuations at the granular level) from any intrinsic irreproducibility
of the results. If the results are reproducible, then for large enough piles one might
expect the averaging over several piles to be obviated by binning the data over
many transducers. This is, in effect, what Smid & Novosad (1981) do (since their
transducers are much larger). More careful experimental investigations of this point
would, nonetheless, be welcome.
We conclude from this recent study, which substantially confirms the earlier work
of Smid & Novosad (1981), that the attribution of the stress dip to basal sag is not
justified for the case of conical piles of sand. Brockbank et al . (1997) also saw a stress
† Savage (1997) also criticizes the reduction method used to analyse these data by Wittmer et al.
(1997a), as shown in figure 1; when normalizing stresses by the mean density of the pile, he apparently
prefers to use a separate measurement of the bulk density (in a different geometry), rather than the
density deduced by integrating the vertical normal stresses to give the weight of the pile.

dip for small, but not large, glass beads. This difference suggests that to observe the
dip requires a large enough pile compared to the grain size; perhaps to allow an
anisotropic mesoscale texture to become properly established. No dip was seen for
lead shot (deformable) or flour (cohesive).
(b) Wedges
The experiments on wedges appear very different. The papers of Hummel & Finnan
(1920), and Lee & Herington (1971) include data-sets for which the construction
history is described as being effectively from a line source. These results, as well
as others cited by Savage (1997), offer support for his conclusions (made earlier by
Trollope & Burman 1980) that the construction history of the wedge does not much
matter, and that there is only a very small or negligible dip for wedges supported by
a fully rigid base. These studies also suggest that a dip appears almost immediately,
if the base under the wedge is allowed to sag. These results, if confirmed by careful
repetition of the experiments, would certainly cast doubt on FPA-type models as
applied to wedges.
(c) Specifying the construction history

Such historic experiments, measuring the stress distribution for wedges made sup-
posedly from a line source, need careful repetition. This is because, even from a point
source (conical pile) or line source (wedge), at least two different types of construction
history are possible. The first is when, as assumed in FPA-type models, the grains
avalanche in a thin layer down the free surface. The second, which, like the first, has
clearly been observed in three-dimensional work on silo filling (Munch-Andersen &
Nielsen 1990; Munch-Andersen & Askegaard 1993) is called ‘plastic-cone’ behaviour.
It entails the impacting grains forcing their way downwards at the apex into the body
of the pile, which then spreads sideways. A parcel of grains arriving at the apex ends
up finally as a thin horizontal layer. (A transition between this and surface avalanche
flow may be controlled by varying the height from which grains are dropped, among
other factors.) A third possibility is that of ‘deep yield’ (see Evesque & Boufellouh
1997): a build-up of material near the apex followed by a deep avalanche in which a
thick slab of material slumps outwards (Evesque 1991).
These different construction histories, even among piles created from a point or
line source, would lead one to expect quite different stress patterns. For example,
the plastic cone construction should lead to a texture with local symmetry about
the vertical, as assumed by Bouchaud et al . (1995). This model, which we also
expect to describe a conical pile built by sieving sand uniformly onto a supporting
disc (Wittmer et al . 1997b) does not give a pressure dip. Although in point-source
experiments on cones the surface avalanche mechanism is usually seen (Evesque
1991; Evesque et al . 1993) we do not know whether the same applies for wedges; the
classical literature is ambiguous (Hummel & Finnan 1920; Lee & Herington 1971).
For these reasons, such experiments must be repeated, with proper monitoring of
the construction history, before conclusions can be drawn.
There are, in fact, good reasons why the surface avalanche scenario, on which
models such as FPA depend, may be very hard to observe in the wedge geometry.
Recall that for the wedge geometry at repose, all OSL models predict an outer
sector of the wedge, of substantial thickness, in which the Coulomb inequality is

saturated. Clearly, if avalanches take place on top of a thick slab of material already
at incipient failure, it may be impossible to avoid rearrangements deeper within
the pile, leading either to ‘deep-yield’ or ‘plastic-wedge’ behaviour. To this extent,
the application of FPA-type models to a wedge geometry is not necessarily self-
consistent. The same does not apply in the conical geometry, where the solution
of these models predicts only an infinitesimal plastic layer at the surface of the
cone (Wittmer et al . 1996). Accordingly, it would be very interesting to compare,
experimentally, wedges and cones of the same material to see whether the character
of the avalanches is fundamentally different, as FPA-like models might lead one
to expect. Further experiments involving comparison of histories are suggested by
Wittmer et al . (1997a, b).
Although there are, so far, few data showing a clear dependence of measured
stresses on construction history in free-standing cones or wedges, the effect is well
established in experiments on silos. Specifically, for flat-bottomed silos filled by sur-
face avalanches from a point source, the vertical normal force at the centre of the
base is less than at the edge (Munch-Andersen & Nielsen 1990). This effect, which is
readily explained within an FPA-type modelling approach (Wittmer et al . 1997a),
is not reported in silos filled by sieving, nor when a plastic-cone behaviour is seen at
the apex (Munch-Andersen & Askegaard 1993).
6. Sandpiles as fragile matter

As we have emphasized, the continuum mechanics represented by our hyperbolic
models is not that of conventional elastoplasticity. In what follows, we develop an
outline interpretation of this continuum mechanics as that appropriate to a mate-
rial in which stresses propagate primarily along force chains. Simulations of frictional
spheres offer some support for the force-chain picture, at least as a reasonable approx-
imation: most of the deviatoric stress is found to arise from strong, normal forces
between particles participating in force chains; tangential forces (friction) and the
weaker contacts transverse to the chains contribute mainly to the isotropic pressure
(Thornton & Sun 1994; Thornton 1997, and this issue). In addition to this, the con-
tent of our models is to assume that the skeleton of force chains is fragile, in a specific
sense defined below.
(a) Force chains

Informally speaking, the hyperbolic problem posed by OSL models is determined
once half of the boundary forces are specified. More precisely (figure 3b), one is
required to specify the surface force tangential to each characteristic ray, at one end
and one end only. The corresponding force acting at the other end is obliged to
balance the sum of the specified force, and any body forces acting tangentially along
the ray. If it does not do so, then within our modelling approach, the material ceases
to be in static equilibrium. This is no different from the corresponding statement for a
fluid or liquid crystal; if boundary conditions are applied that violate the conditions
for static equilibrium, some sort of motion results. Unlike a fluid, however, for a
granular medium we expect such motion to be in the form of a finite rearrangement
rather than a steady flow. Such a rearrangement will change the microtexture of the
material, and thereby alter the constitutive relation among stresses. We expect it to

Figure 6. (a) A force chain (‘stress path’) of hard particles can support only tangential compres-
sive loads in static equilibrium. This is to avoid torques on particles in the chain (gravitational
torques acting directly on the particles within a chain are ignored). (b) A simple realization of
the FPA model as a rectilinear arrangement of force chains under tangential loading.
do so in such a way that the new network of force chains (new constitutive relation)
is able to support the newly imposed forces.
Although simplified, we believe that this picture correctly captures some of the
essential physics of force chains. Such chains are load-bearing structures within the
contact network and, in the simplest approximation of straight chains of uniform
orientation, these must have the property described above: any difference in the
forces on two ends of a path must be balanced by a body force. Note that if one
makes a linear chain of more than two rigid particles with point contacts, then to
avoid torques, this can indeed support only tangential forces, regardless of the local
friction coefficient between the grains themselves; see figure 6a. Force chains should,
we believe, be identified (on average) with the characteristic rays of our hyperbolic
equations. The mean orientation of the force chains is then reflected in a constitutive
equation such as FPA or OSL.
Our modelling approach thus assumes that the mean orientation of force chains,
in each element of the material, is fixed at burial. (This does not necessarily require
that the individual chains are themselves fixed.) We think it reasonable to assume
that the force chains will not change their average orientations so long as they are
able to support subsequent applied loads. But if a load is applied that they cannot
support (one in which the tangential force difference and body force along a path do
not match), irreversible rearrangement is inevitable (P. Evesque, personal communi-
cation). This causes some part of the pile to adopt a new microtexture and thereby
a new constitutive relation. In other words, incompatible loadings of this kind must
be seen as part of the construction history of the pile.
(b) Fragile matter

There is a close connection between these ideas and recent work on the ‘marginal
mechanics’ of periodic arrays of identical grains. (This receives further consideration
in Appendix B.) The marginal situation is where the (mean) coordination number

of the grains is the minimum required for mechanical integrity; in two dimensions
this is three for frictional and four for frictionless spheres. (Larger coordination num-
bers are needed for aspherical grains.) Indeed, each OSL model rigorously describes
the continuum mechanics of a certain ordered array of this kind (see Appendix B).
Marginal packings are exceptional in an obvious sense: most packings of grains one
can think of do not have this property, and the forces acting on each grain cannot be
found without further information. However, we can interpret this correspondence
between continuum and discrete equations, not at the level of the packing of indi-
vidual grains (for which the marginal coordination state would be hard to explain)
but at the level of a granular skeleton made of force chains. The OSL models (in two
dimensions) can then be viewed as postulating a simplified, marginally stable geome-
try of the skeleton, in which a regular lattice of force chains (bearing tangential forces
only) meet at four-fold coordinated junctions. (For the FPA model, though not in
general, this lattice is rectangular; see figure 6b.) Such a skeleton leads to hyperbolic
equations (or perhaps parabolic ones if enough disorder is added); its mechanics are
determinate in the absence of a displacement field specified at the base.
In the present context, fragility arises from the requirement of tangential force
balance along force chains. If this is violated at the boundary (within the models as
so far defined, even infinitesimally), then internal rearrangement must occur, causing
new force chains to form, so as to support the load. It seems reasonable to assume
that when rearrangements are forced upon the system, it responds in an ‘overdamped
manner’, that is, the motion ceases as soon as the load is once again supported. If
so, one expects the new state to again be marginally stable. This suggests a scenario
in which the skeleton evolves dynamically from one fragile state to another. By such
a mechanism, marginally stable packings, although exceptional in the obvious sense
that most packings one can think of are not marginal, may nonetheless be generic
in unconsolidated dry granular matter. Thornton (1997) reports that, in simulations
of frictional spheres, force chains do rearrange strongly under slight reorientations of
the applied load.
Consider, finally, a regular lattice of force chains, for simplicity rectangular (the
FPA case), which is fragile if the chains can support only tangential loads. This is
the case so long as such paths consist of linear chains of rigid particles meeting at
frictional point contacts: as mentioned above, the forces on all particles within each
chain must then be collinear, to avoid torques. This imposes the (FPA) requirement
that there are no shear forces across a pair of orthogonal planes normal to the force
chains themselves (see § 3 d). Suppose now a small degree of particle deformability
is allowed (Cates et al . 1998). This relaxes slightly the collinearity requirement, but
only because the point contacts are now flattened. The ratio, , of the maximum
transverse load to the normal one, will therefore vanish as some power of the mean
deformation. This yield criterion applies only across two special planes; failure across
others is governed by some smooth yield requirement (such as the ordinary Coulomb
condition: the ratio of the principal stresses lies between given limits). The granular
skeleton just described, which was fragile in the limit of rigid grains, is now governed
by a strongly anisotropic elastoplastic yield criterion of precisely the kind described in
§ 3 d. The skeleton can support loads that do violate the tangential balance condition,
but only through terms that vanish as → 0. To escape the hyperbolic regime of
‘fabric dominance’, must be significant, which in turn requires significant particle
deformation under the influence of the mean stresses applied.

This indicates how a non-fragile packing of frictional deformable rough particles,

displaying broadly conventional elastoplastic features when the deformability is sig-
nificant, can approach a fragile limit when the limit of a large modulus is taken at
fixed loading. (It does not, of course, imply that all packings become fragile in this
limit.) Conversely, it shows how a packing that is basically fragile (in its response
to gravity), could nonetheless support very small incremental deformations, such as
sound waves, by an elastic mechanism. The question of whether sandpiles are better
described as fragile, or as ordinarily elastoplastic, remains open experimentally. To
some extent it may depend on the question being asked. However, we have argued,
on various grounds, that in calculating the stresses in a pile under gravity a fragile
description may lie closer to the true physics.
7. Conclusion
From the perspective of geotechnical engineering, the problem of calculating stresses
in the humble sandpile may appear to be of only marginal importance. The physicist’s
view is different: the sandpile is important, because it is one of the simplest problems
imaginable in granular mechanics. It therefore provides a test-bed for existing models
and, if these show shortcomings, may suggest ideas for improved physical theories of
granular media.
There are, in physics, certain types of problem for which the fundamental princi-
ples or equations are clear, and the difficulty lies in working out their consequences.
An example is the use of the Navier–Stokes equation in studies of (say) turbulence.
The form of the Navier–Stokes equation can be deduced by considering only the
symmetries and conservation laws of an isotropic fluid. Accordingly, its status is not,
as sometimes assumed, that of an approximation based on constitutive hypotheses
that happen to be very accurate for certain materials. Rather, it describes a lim-
iting behaviour, which all members of a large class of materials (viscoelastic fluids
included) approach with indefinite accuracy in the limit of long length- and time-
scales. (We are aware of no theory of elastoplasticity having remotely similar sta-
tus.) There are other types of problem in which the fundamentals are not clear. For
such problems, the governing equations must first be established, before they can be
solved. We remain convinced that the static modelling of poured assemblies of cohe-
sionless grains under gravity is of this second type. This view is not particularly new,
either among physicists (Edwards & Oakeshott 1989), or among engineers (Gudehus
1985, 1997).
From this perspective, we can see no reason why the starting points of simple
rigid-plastic or elastoplastic continuum mechanics should offer significant insights
into the sandpile problem. Simple elastoplastic approaches, in particular, give only
one unambiguous physical prediction: that a sandpile supported by a rough base
should have no definite behaviour. Experimentalists, who believe themselves to be
measuring a definite result, are likely to be baffled by such predictions. For if, as these
models require, the forces acting at the base of a pile can be varied at will without
causing its static equilibrium to be lost (by making small elastic displacements at
the base), then all the published ‘measurements’ of such forces must be dismissed as
artefact. An alternative view is that these represent rather haphazard investigations
of some unspecified physical mechanism that does somehow determine a displacement
field at the base of the pile. (As mentioned previously, basal sag is certainly not an

adequate candidate.) The challenge of whether, for cohesionless poured sand, such a
displacement field can sensibly be defined, remains open.
Given the present state of the data, a conventional elastoplastic interpretation
of the experimental results for sandpiles may remain tenable; more experiments are
urgently required. In the meantime, a desire to keep using tried-and-tested modelling
strategies until these are demonstrably proven ineffective is quite understandable. We
find it harder to accept the suggestion (Savage 1997) that anyone who questions the
complete generality of traditional elastoplastic thinking is somehow uneducated.
Our own position is not that elastoplasticity itself is dead, but that macroscopic
stress propagation in sandpiles is determined much more by the internal fabric of the
material (therefore the construction history), and much less by boundary conditions,
than simple elastoplastic models suggest. Reasons for this, based on the idea of a
fragile skeleton of force chains, have been discussed above. By considering a particular
form of yield condition, we have shown how a fragile model can be matched smoothly
onto a relatively conventional, but strongly anisotropic, elastoplastic theory. Thus it
is possible in principle to have a model which, although strictly governed by the
mixed hyperbolic/elliptic equations of elastoplasticity, leads to solutions that obey
purely hyperbolic equations, to within (elastically indeterminate) corrections that
are small in a certain limit. In such a system, the results will depend less and less
on boundary conditions, and more and more on fabric, as that limit is approached.
Moreover, for certain well-defined fragile packings of frictional grains, the limit is the
rigid-particle one, in which the elastic modulus of the grains is taken to infinity at
fixed loading.
In summary, we have discussed a new class of models for stress propagation in gran-
ular matter. These models assume local propagation rules for stresses which depend
on the construction history of the material and which lead to hyperbolic differen-
tial equations for the stresses. As such, their physical basis is substantially different
from that of conventional elastoplastic theory (although they may have much more
in common with ‘hypoplastic’ models). Our approach describes a regime of ‘fragile’
behaviour, in which stresses are supported by a granular skeleton of force chains
that must undergo finite internal rearrangement under certain types of infinitesi-
mal load. Obviously, such models of granular matter might be incomplete in various
ways. Specifically, we have discussed a possible crossover to elastic behaviour at
very small incremental loads, and to conventional elastoplasticity at very high mean
stresses (when significant particle deformations arise). However, we believe that our
approach, by capturing at least some of the physics of force chains, may offer impor-
tant insights that lie beyond the scope of conventional elastoplastic or rigid-plastic
modelling strategies. The equivalence between our fragile models and limiting forms
of extremely anisotropic elastoplasticity has been pointed out.
We are grateful to S. Edwards, P. Evesque, J. Goddard, G. Gudehus, J. Jenkins, D. Kolymbas,
D. Levine, S. Nagel, S. Savage, C. Thornton and T. Witten for discussions. This research was
funded in part by EPSRC (UK) Grants GR/K56223 and GR/K76733.
Appendix A. Laminated elastoplastic cone

As an alternative to the ‘spaceship model’, one might envisage (figure 5b) the creation
of a pile by incremental addition of thin layers of elastoplastic material to its upper
surface (in imitation of an avalanche). It might then be argued that this thin layer,

being under negligible stress, must be characterized by a zero-displacement field (Sav-

age 1998). On a rough support, one would then expect the displacement at the base
to remain zero as further additions to the pile are made, giving a zero-displacement
boundary condition at the base of what has, by now, presumably become a simple
elastoplastic body.
This reasoning is flawed: the same argument entails that, at any stage of the pile’s
construction, the last layer added is in a state of zero displacement, not just where it
meets the base, but along its entire length. If so, then not only the base but also the
free surface of the pile is subject to a zero-displacement boundary condition. For a
simple elastoplastic cone or wedge, this is incompatible with the zero-stress boundary
condition already acting at the free surface. (Such a body, in effect suspended under
gravity from a fixed upper surface, will exert forces across that surface, as well as
across the supporting base.)
The paradox is resolved by noticing that this ‘laminated elastoplastic’ model in fact
involves the addition of thin, stress-free elastoplastic layers to an already deformed
body. The result will not be a simple elastoplastic continuum, but a body in which
internal stresses and displacements are present even when all body forces are removed
(like a reinforced concrete pillar, or a tennis racket made of laminated wood); see
figure 5b. Such a body can, if carefully designed with a specific loading in mind,
simultaneously satisfy a zero-stress and zero-displacement (more properly, constant-
displacement) boundary condition at any particular surface. These rather intriguing
properties may well be worth investigating further, but they are still a long way
from a realistic description of the construction history of a sandpile. In any case,
it is misleading to suggest (Savage 1998) that such considerations can justify the
adoption of a zero-displacement basal boundary condition within an ordinary (i.e.
not pre-strained), isotropic elastoplastic continuum model.
Appendix B. Microscopic force-transmission models

First, note that a very large class of discrete models lead directly to OSL models in
the continuum limit. A simple example is defined in figure 7a. As shown by Bouchaud
et al . (1995), this model gives a wave equation with two characteristic rays symmet-
rically arranged about the vertical axis. If the symmetry in the stress propagation
rules is broken, an asymmetric OSL model arises instead (figure 7b).
Second, when the continuum limit of such force-transfer models is taken, one has
(in two dimensions) only two characteristic rays even if the force-transfer rules involve
more than two neighbours in the layer below. An example (Claudin et al . 1998) is
shown in figure 7c. Broadly speaking, one recovers an OSL model, in the contin-
uum limit, whenever the forces passed from a grain to its downward (or sideways)
neighbours obey a deterministic linear decomposition of the ‘incident force’ (fx , fz ),
defined as the vector sum of the forces acting from grains in the layer above, plus
the body force on the given grain.
Trollope’s model, whose force-transfer rules are as shown in figure 7d–f , is not
a member of this class. (Indeed it has three characteristic rays in the continuum
limit, rather than two.) This is because the vector sum of the incident forces on a
grain is not taken before applying a rule to determine the outgoing forces from that
grain; the latter depend separately on each of the incident forces. As a description
of hard frictional grains, we consider this unphysical. For, if the grain in figure 7d

Figure 7. (a) Force-transfer rules for a simple discrete model (Hong 1993; Bouchaud et
al. 1995). The forces obey (f2 − f1 ) sin θ = fx and (f1 + f2 ) cos θ = fz + w, with w
that of gravity. A first-order difference equation for fx is found by writing fx (x, z) =
[f2 (x − ∆x, z − ∆z) − f1 (x + ∆x, z − ∆z)] sin θ, with ∆x = d cos θ and ∆z = d sin θ, and
eliminating f1,2 in favour of fx,z (d is the grain diameter). A similar procedure is then followed
for fz . In the continuum limit, the resulting first-order differential equations give the BCC
model (with c0 = tan θ) with two characteristics (right). (b) The same, with asymmetric prop-
agation rules, leading to an asymmetric OSL model. (c) A simple model with three downward
neighbours. The force-assignment rule for the middle ray is f2 = α(f1 + f3 ), where α is some
constant. As shown by Claudin et al . (1998), the result is still an OSL model (in fact BCC with
c0 = tan θ0 < tan θ). (d) In Trollope’s model, the outgoing granular forces (p0 , q 0 , r0 ) depend
separately on the incoming ones (p, q, r) rather than on their vector sum: p0 − p = w/[c(1 + k)],
q 0 − q = wk/[c(1 + k)] and r0 = r + (1 − k)wt/(1 + k). Here, w is the weight of a grain and c
and t denote cos θ, tan θ. (e) As a result, for 0 < k < 1 a symmetrical extra loading from two
neighbours above whose resultant 2f c is directly downwards, is not equivalent to an increase
(f ) in grain weight w = 2f c.
is subjected to two equal small extra forces f , from its two neighbours in the layer
above (whose vector sum is vertical), the net effect on the outgoing forces should be
equivalent to a small increase in its weight, w = 2f cos θ. Within Trollope’s model,
this is not the case. Since its propagation rules are linear, any attempt to rectify this
feature (by taking the vector sum of the forces before propagating these on to the
next layer) will give an OSL model instead.

References
Bagster, D. F. 1978 An idealised model of granular material behaviour in ore heaps and hoppers.
J. Powder Bulk Solids Technol. 2, 42–46.
Bouchaud, J.-P., Cates, M. E. & Claudin, P. 1995 Stress-distribution in granular media and
nonlinear-wave equation. J. Physique I 5, 639–656.
Bouchaud, J. P., Claudin, P., Cates, M. E. & Wittmer, J. P. 1998 Models of stress propagation in
granular media. In Proc. Physics of Dry Granular Media, Cargese, France, 15–26 September
1997 (ed. H. J. Herrmann, J.-P. Hovi & S. Luding), pp. 97–122. NATO Advanced Study
Institute. Kluwer.
Brockbank, R., Huntley, J. M. & Ball, R. C. 1997 Contact force distribution beneath a three-
dimensional granular pile. J. Physique II 7, 1521–1532.
Cantelaube, F. & Goddard, J. D. 1997 Elastoplastic arching in 2D heaps. In Proc. 3rd Int. Conf.
on Powders and Grains, Durham, NC, USA, 18–23 May (ed. R. P. Behringer & J. T. Jenkins),
pp. 231–234. Rotterdam: Balkema.
Cates, M. E., Wittmer, J. P., Bouchaud, J.-P. & Claudin, P. 1998 Jamming, force chains and
fragile matter. Phys. Rev. Lett. 81, 1841–1844.
Claudin, P. & Bouchaud, J.-P. 1997 Static avalanches and giant stress fluctuations in silos. Phys.
Rev. Lett. 78, 231–234.
Claudin, P., Bouchaud, J.-P., Cates, M. E. & Wittmer, J. P. 1998 Stress fluctuations in granular
media. Phys. Rev. E 57, 4441–4457.
Dantu, P. 1967 Contribution à l’étude méchanique et géométrique des milieux pulvérulents.
Ann. Ponts Chaussées 4, 144–148.
Edwards, S. F. & Oakeshott, R. B. S. 1989 The transmission of stress in an aggregate. Physica
D 38, 88–92.
Evesque, P. 1991 Analysis of the statistics of sandpile avalanches using soil-mechanics results
and concepts. Phys. Rev. A 43, 2720–2740.
Evesque, P. 1997 About the scaling hypothesis of the stress field in a conic sandpile. Comment
on ‘Stress propagation and arching in static sandpiles’, by J. P. Wittmer et al . J. Physique
I 7, 1305–1307.
Evesque, P. & Boufellouh, S. 1997 Stress distribution in an inclined pile: soil mechanics cal-
culation using finite element technique. In Proc. 3rd Int. Conf. on Powders and Grains,
Durham, NC, USA, 18–23 May (ed. R. P. Behringer & J. T. Jenkins), pp. 295–298. Rotter-
dam: Balkema.
Evesque, P., Fargeix, D., Habib, P., Luong, M. P. & Porion, P. 1993 Pile density is a control
parameter of sand avalanches. Phys. Rev. E 47, 2326–2332.
Gudehus, G. 1974 Konstruktion statisch möglicher Spannungsfelder in Erdkörpern. Strasse/
Bruecke/Tunnel 6, 157–161.
Gudehus, G. 1985 In Mechanics of geomaterials: rocks, concrete, soils (ed. Z. P. Bazant), pp. 47–
63. New York: Wiley.
Gudehus, G. 1997 Attractors, percolation thresholds and phase limits of granular soils. In Proc.
3rd Int. Conf. on Powders and Grains, Durham, NC, USA, 18–23 May (ed. R. P. Behringer
& J. T. Jenkins), pp. 169–183. Rotterdam: Balkema.
Harr, M. E. 1977 Mechanics of particulate media. New York: McGraw-Hill.
Hong, D. C. 1993 Stress-distribution of a hexagonally packed granular pile. Phys. Rev. E 47,
760–762.
Hummel, F. H. & Finnan, E. J. 1920 The distribution of pressure on surfaces supporting a mass
of granular material. Proc. Inst. Civil Engng 212, 369–392.
Huntley, J. M. 1993 Vacancy effects on the force distribution in a two-dimensional granular pile.
Phys. Rev. E 48, 4099–4111.

Jäger, H. M., Nagel, S. R. & Behringer, R. P. 1996 The physics of granular materials. Phys.
Today 49, 32–38.
Janssen H. A. 1895 Versuche über Getreidedruck in Silozellen. Z. Vert. Dt. Ing. 39, 1045.
Jokati, T. & Moriyama, R. 1979 On the bottom pressure distribution of the bulk materials piled
with angle of repose. J. Soc. Powder Tech. Jpn 60, 184–189.
Kolymbas, D. 1991 An outline of hypoplasticity. Arch. Appl. Mech. 61, 143–151.
Kolymbas, D. & Wu, W. 1993 Introduction to hypoplasticity. In Modern approaches to plasticity
(ed. D. Kolymbas), pp. 213–223. Rotterdam: Elsevier.
Landau, L. D. & Lifshitz E. M. 1986 Theory of elasticity, 3rd edn. Oxford: Pergamon.
Lee, I. F. & Herington, J. R. 1971 Stresses beneath granular embankments. In Proc. 1st Aust.–
N.Z. Conf. Geomech., vol. 1, pp. 291–297.
Liffmann, K., Chan, D. Y. C. & Hughes, B. D. 1992 Force distribution in two-dimensional
sandpiles. Powder Technol. 72, 255–267.
Liu, C. H., Nagel, S. R., Schecter, D. A., Coppersmith, S. N., Majumdar, S. & Witten, T. A.
1995 Force fluctuations in bead packs. Science 269, 513–515.
Munch-Andersen, J. & Nielsen, J. 1990 Pressures in slender grain silos. In Proc. CHISA’ 90,
Praha, 26–31 August 1990.
Munch-Andersen, J. & Askegaard, V. 1993 Silo model tests with sand and grain. In Proc.
RELPOWFLO II, Oslo, 23–25 August 1993.
Nedderman, R. M. 1992 Statics and kinematics of granular materials. Cambridge University
Press.
Samsioe, A. F. 1955 Stresses in downstream part of an earth or a rock fill dam. Géotechnique
5, 200–223.
Savage, S. B. 1997 Problems in the statics and dynamics of granular materials. In Proc. 3rd
Int. Conf. on Powders and Grains, Durham, NC, USA, 18–23 May (ed. R. P. Behringer &
J. T. Jenkins), pp. 185–194. Rotterdam: Balkema.
Savage, S. B. 1998 Modelling and granular material boundary value problems. In Proc. Physics
of Dry Granular Media, Cargese, France, 15–26 September 1997 (ed. H. J. Herrmann, J.-P.
Hovi & S. Luding), pp. 25–96. NATO Advanced Study Institute. Kluwer.
Smid, J. & Novosad, J. 1981 Pressure distribution under heaped bulk solids. In Proc. of 1981
Powtech Conference, Ind. Chem. Eng. Symp. 63, D3V 1–12.
Thornton, C. 1997 Force transmission in granular media. In KONA Powder and Particle 15,
81–90.
Thornton C. & Sun, G. 1994 Numerical simulation of general 3-D quasi-static shear deforma-
tion in granular media. In Numerical methods in geotechnical engineering (ed. I. M. Smith),
pp. 143–148. Rotterdam: Balkema.
Trollope, D. H. 1968 The mechanics of discontinua or clastic mechanics in rock problems. In Rock
mechanics in engineering practice (ed. K. G. Stagg & O. C. Zienkiewicz), ch. 9, pp. 275–320.
New York: Wiley.
Trollope, D. H. & Burman, B. C. 1980 Physical and numerical experiments with granular wedges.
Géotechnique 30, 137–157.
Wittmer, J. P., Claudin, P., Cates, M. E. & Bouchaud, J.-P. 1996 An explanation for the central
stress minimum in sand piles. Nature 382, 336–338.
Wittmer, J. P., Claudin, P. & Cates, M. E. 1997a Stress propagation and arching in static
sandpiles. J. Physique I 7, 39–80.
Wittmer, J. P., Claudin, P., Cates, M. E. & Bouchaud, J.-P. 1997b A new approach for stress
propagation in sandpiles and silos. In Friction, arching, contact dynamics (ed. D. E. Wolf &
P. Grassberger), pp. 153–167. Singapore: World Scientific.
Wood, D. M. 1990 Soil behaviour and critical state soil mechanics. Cambridge University Press.

Development of Stresses in Cohesionless Poured Sand

Uploaded by

Document Informationclick to expand document information

Copyright:

Available Formats

Development of Stresses in Cohesionless Poured Sand

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Development of Stresses in Cohesionless Poured Sand

Uploaded by

Copyright:

Available Formats

Development of stresses in

cohesionless poured sand

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

2. Continuum models of cohesionless granular matter

(a) Stress continuity and the Coulomb inequality

Phil. Trans. R. Soc. Lond. A (1998)

(b) Rigid-plastic models

Phil. Trans. R. Soc. Lond. A (1998)

(c) Elastoplastic models

Phil. Trans. R. Soc. Lond. A (1998)

(d ) Local-rule models of stress propagation

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

3. Strain and displacement variables

Phil. Trans. R. Soc. Lond. A (1998)

4. The role of boundary conditions

(b) The physics of elastic indeterminacy

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

(c) Displacements to the rescue?

Phil. Trans. R. Soc. Lond. A (1998)

(d ) Determinacy and anisotropy

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

(c) Specifying the construction history

Phil. Trans. R. Soc. Lond. A (1998)

6. Sandpiles as fragile matter

(a) Force chains

Phil. Trans. R. Soc. Lond. A (1998)

(b) Fragile matter

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

This indicates how a non-fragile packing of frictional deformable rough particles,

Phil. Trans. R. Soc. Lond. A (1998)

Appendix A. Laminated elastoplastic cone

Phil. Trans. R. Soc. Lond. A (1998)

being under negligible stress, must be characterized by a zero-displacement field (Sav-

Appendix B. Microscopic force-transmission models

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

Phil. Trans. R. Soc. Lond. A (1998)

You might also like