Granular Matter (2012) 14:381–392

DOI 10.1007/s10035-012-0342-3

ORIGINAL PAPER

A three-dimensional discrete-grain model for the simulation

of dam-break rectangular collapses: comparison between
numerical results and experiments
L. Girolami · V. Hergault · G. Vinay · A. Wachs

Received: 14 August 2011 / Published online: 17 March 2012

© Springer-Verlag 2012

Abstract In this paper, we used a 3-D discrete-element List of symbols

model, Grains3D, which allows the simulation of unsteady δij Overlap distance between a particle
granular flows of monodisperse soft spherical particles in a i and j during the collision
common situation (i.e., down a rectangular channel). A series t Computational time step
of numerical dam-break experiments was performed to pre- n Coefficient of restitution
dict the behavior of granular columns that propagate down γn Damping coefficient in the normal direction
a rough horizontal surface from different initial conditions γt Dissipative friction coefficient in the
(varying the initial aspect ratio). Numerical results were com- tangential direction
pared to those obtained experimentally by Lajeunesse et al. μc Interparticle Coulomb friction coefficient
(Phys Fluids 17:103302, 2005) from a similar configuration. μw Particle-wall Coulomb friction coefficient
Runout distance, temporal flow evolution, deposit morphol-
ωij Relative angular velocity between a particle i and j
ogy and internal flow structures of similar laboratory experi-
ρi Density of a particle i
ments were quantitatively reproduced as well as prediction of
σ Poisson coefficient
empirical and theoretical scaling laws. This paper highlights
τc Characteristic free fall time of the granular column
that such fully 3-D simulations of soft-spheres can remark-
θc Internal friction angle
ably capture dam-break collapses performed in a rectangular
channel. Moreover, Grains3D can provide a complete physi- a Initial aspect ratio of the column in the reservoir
cal description of such complex unsteady systems which will dp Particle diameter
be the topic of future on-going studies. E Young modulus
Fi Sum of the forces applied to a particle i
Keywords Discrete-element modeling · Fij Colliding force between a particle i and j
Dam-break flows · Quantitative reproduction Fij,dn Normal component of the dissipative force
Fij,el Hookean elastic restoring force
Fij,t Shear component of the dissipative force
h Flow thickness at the time t
hf Deposit thickness taken at x = Li
H Reservoir height
Hi Initial height of the column
kms Coefficient of rolling resistance
L. Girolami (B) · V. Hergault · G. Vinay
kn Stiffness coefficient
Fluid Mechanics Department, IFP Energies Nouvelles,
1-4 Avenue de Bois Préau, 92852 Rueil-Malmaison, France lf Length of the final deposit taken at x = 0
e-mail: [email protected] Li Initial length of the column
Mi Mass of a particle i
A. Wachs
Mrf Artificial rolling friction
Fluid Mechanics Department, IFP Energies Nouvelles, Rond-point
de l’Echangeur de Solaize, BP 3, 69360 Solaize, France mij Reduced mass of two particles i and j in contact

123
382 L. Girolami et al.

N Total number of particles constituting the system propagation and/or for granular columns characterized by a
nC Unit normal vector at the contact point high initial aspect ratio [35].
R Particle radius Therefore, it follows that the most accurate modeling for
Rij Reduced particle radius complex geophysical and industrial flows is provided by a
tC Unit tangential vector at the contact point set of kinematic equations which describe the trajectory of
t Time from gate opening each individual particle. The role of particle interactions on
Urn Relative velocity between a particle i and j the flow dynamics constitute typical insights that could be
in the normal direction gained into such particle-based modeling methods [7]. Dis-
Urt Relative velocity between a particle i and j crete-element methods (DEM), first introduced by Cundall
in the tangential direction [11] and Cundall and Strack [12], are particularly well suited
w Channel width to model flows of particulate solids [6–9,18,29,45,50] which
x Flow length at the time t is the purpose of this paper. The principal interest of such a
xf Flow length measured from x = Li method is to represent granular media as an assembly of con-
tacting particles, termed as discrete elements (DE). However,
calculations are associated with a heavy computational load
1 Introduction and resource requirements which come from the small time
step required to properly handle particle contacts and thus
Granular flows are common phenomena in natural and man- the very large number of time increments. Therefore, appli-
made processes. Understanding their dynamics represents a cations of DEM are generally limited to 2-D systems or 3-D
crucial objective for hazard assessment and industrial optimi- systems involving a weak number of particles. Paralleliza-
zation, namely in the fields of agriculture, chemical engineer- tion is then an obvious option for the solution of natural and
ing, fundamental physics, and pharmaceutical industry. An industrial-scale problems which require increasing compu-
important feature of a granular material is its mobility or the tational capabilities.
degree to which it is capable of moving freely from its source Discrete-grain simulations have shown a promising poten-
[13,15,44,49]. This is exploited in industrial processes as a tial to reproduce the general behavior of unsteady collapses of
mechanism by which a solid material can be transported from a granular column and the key scaling laws observed exper-
one place to another. It is also important in many environmen- imentally [27,33]. Zenit [50] and Staron and Hinch [45]
tal flows, such as the geophysical flows produced by some first modeled the 2-D axisymmetric collapse of undeform-
natural catastrophes where large, dangerous, and destructive able rigid circular disks using the contact dynamics method
currents of material can travel up to several kilometers [4]. (CDM). Their reported results qualitatively reproduced the
Although field observations from deposits [4] and approx- experiments [2,27,33], but not quantitatively. Then, Cleary
imate measurements of frontal velocity [30,32] exist, their and Franck [10] used a 3-D DEM to examine the effect of
hazardous behaviour developed on a broad range of spatial the initial parameters on the dynamics of granular collapses.
and temporal scales requires more detailed studies. Compu- The model parameters were adjusted by repeating a single
tational modeling can play an important role both in helping experiment of an axisymmetric column. Results were very
to describe precisely the physical processes developed dur- satisfactory but the model fails to provide quantitative results
ing transport and in predicting environmental risks induced in this geometrical configuration. More recently, Lacaze et
by flow and deposition. al. [26] reproduced 2-D planar collapses of soft spheres using
Previous numerical studies dealing with the simulation of DEM simulations. Numerical results matched very well the
granular flows [19,24,34,35,38,40,42] mostly focused on similar experiments. However, the planar configuration leads
the development of depth-averaged models. This approach to behaviors that disrupt the collapse dynamics. An effect
is essential to capture the first order feature of granular flows of crystallization, caused by large-scale lattice structures,
in simple situations such as flows down inclined surfaces in occurs when the channel width (w) is equal to the particle
which the flow achieves a steady regime [20,40] as well as diameter (dp ). Otherwise, an effect of jamming is observed
for large scale simulations over natural terrains with existing when w is taken greater (about 2 dp ) since the gap between
computational power [34,38]. In these models, simplifica- the sidewall is large enough for two particles to wedge each
tions come from assuming no vertical variations, resulting other due to their softness. The authors described that both
in the depth averaged shallow-water equations. These mod- crystallization and jamming effects have an influence on the
els are commonly applied with a Coulomb basal friction law final state of the static pile and lead then to nongeneric results.
[40] which is a leading order parameter in the characteris- Therefore, they were constrained to fix w equal to 1.2 dp to
tics of the final stopping phase and deposit. However, this minimize these effects.
approach remains limited and fails to capture unsteady flows In this paper, we show that such 3-D discrete-grain mod-
for which the volume fraction fluctuates significantly during els of soft spheres are particularly well suited to the accurate

123
Discrete-grain model for dam-break flows 383

reproduction of dam-break rectangular collapses. A series of where the stiffness coefficient (kn ) is expressed as a function
flows generated from granular columns and characterized by of the Young modulus (E) and the Poisson coefficient (σ ),
different initial aspect ratios have been performed. A mes- such as: kn = 3(1−σ
ER
2 ) for two identical spheres of radius R.
sage passing interface (MPI) library was implemented in the δij denotes the overlap distance between the particles i and
code to speed up the parallel computations. This allows to j during the collision and nC : the unit normal vector at the
introduce a reasonable number of particles in the reservoir contact point.
as well as to obtain a reasonable computation time. Details The normal component of the dissipative force (Fij,dn ) is
on the parallel strategy and associated profitability will be expressed by:
the heart of a following paper and is beyond the scope of the
Fij,dn = −2γn mij Urn (4)
present one.
Here, we first present the governing equations and the where γn represents the damping coefficient in the normal
numerical scheme of Grains3D. We also shortly outline Mi Mj
direction, mij = Mi +M j
: the reduced mass of the two par-
the parallelization technique using the MPI library and its ticles i and j in contact and Urn : their relative velocity in
implementation in the code. Then, we present simulations the normal direction. γn can be expressed as  a function of the
of 3-D dam-break rectangular collapses. Finally, concluding
coefficient of restitution (n ) such as γn = − mknij π 2 +(ln
lnn
e n )2
.
remarks and perspectives are presented in the last section.
Note that the normal force is set to 0 when the model yields
a negative value to prevent unrealitic attraction between the
two colliding particles [41].
2 The three-dimensional DEM model
The shear component (Fij,t ) of the dissipative force is
defined as follows:
2.1 Contact force model
Fij,t = − min{μc |Fel |, |Fdt |}tC (5)
DEM [3,5,48] constitute the most accurate models of gran-
ular flows since they provide individual trajectories of par- with:
ticles, including their interactions with each other and with Fdt = −2γt mij Urt (6)
their environment. In Grains3D, presented in this paper, the
distinct elements, termed hereafter as grains, interact with where μc is the interparticle Coulomb friction coefficient
each other through a force-displacement law and the New- expressed as a function of the internal friction angle (θc ),
ton’s second law of motion. The interaction forces between such as: μc = tan(θc ), and tC is the unit tangential vector
grains are calculated using the soft-sphere approach [12,46]. at the contact point. γt is the dissipative tangential friction
This latter allows the grains-overlap during the shock which coefficient (set as γt = γn ) and Urt : the relative velocity of
generates a repulsive force pushing grains back apart by a the two particles i and j in contact in the tangential direc-
distance proportional to the length of interpenetration and to tion.
the relative velocity of the two grains at the contact point. In order to prevent the free rolling of particles, we used a
The total contact force includes a dissipative frictional con- rolling resistance model (termed “type A” in the classifica-
tribution from the tangential motion after collision. The sum tion of Ai et al. [1]) and in which the artificial rolling friction
of the forces (Fi ) applied to a particle i are those induced by (Mrf ) is based on a torque model as commonly used in DEM
gravity and by collisions with another particle j, Fij , such [25,31,36,51,52], such as:
as: ωij
Mrf = −kms |Fij n
|Rij (7)
−1 |ωij |

N
Fi = Mi g + Fij (1) where kms represents the coefficient of rolling resistance,
j =0,j =i n = F n , R = Ri Rj is the reduced radius and ω =
Fij ij C ij Ri +Rj ij
ωi − ωj is the relative angular velocity.
 Mi represents the mass of the particle i, such as Mi =
where
ρi 43 π R 3 with R: its radius and ρi : its density; and N : the
total number of particles constituting the system. 2.2 DEM algorithm and time integration scheme
The total contact force (Fij ) resulting from a collision
between a particle i and j can be decomposed as follows: In order to calculate the contact forces applied on each par-
ticle, three main tasks are performed at each time step:
Fij = Fij,el + Fij,dn + Fij,t (2)

The Hookean elastic restoring force (Fij,el ) reads: (1) The construction of a particle near-neighbour interaction
list which allows the reduction of the computational
Fij,el = kn δij nC (3) complexity of this problem. Furthermore, a detection

123
384 L. Girolami et al.

of particle contacts is performed using a linked-cell Table 1 Summary of the input parameters used for the different numer-
method in order to obtain a simple cellular decomposi- ical simulations performed in this study
tion [16]. Calculus N dp ( mm) W (cm) Li (cm) a Figs.
(2) The calculation of the total force applied on each parti-
1 15,000 3 4.5 5.3 3.2 4, 5, 8
cle in the near-neighbour list.
2 3,800 3 4.5 2.6 3.2 4, 8
(3) The resolution of the kinematic equations predicting the
particle position from the velocity whilst solving New- 3 3,300 5 4.5 5.3 3.2 4, 8
ton’s equations gives the velocities from the net forces. 4 8,500 3 4.5 10 0.5 2, 3, 6, 7, 8
5 9,500 3 4.5 7.5 1.0 7, 8
6 8,500 3 4.5 5.0 2.0 2, 6, 8
To ensure the stability and accuracy of the calculations, 7 5,900 3 4.5 3.5 3.0 7, 8
the computational time step (t) was chosen much smaller 8 8,500 3 4.5 2.5 7.5 2, 6, 7, 8
than the contact duration, thus typically equals to 10−7 s for
9 8,300 3 4.5 1.75 15.0 7, 8
the dam-break simulations reported in this paper. Therefore,
10 8,100 3 4.5 1.6 18.0 7, 8
a second-order leap-frog scheme was used, allowing a single
evaluation of the contact forces per time step. Thus, around
10 to 50 time steps were needed to appropriately integrate a
contact.
3 Dam-break collapses

2.3 MPI implementation 3.1 Set-up, materials, and procedures

Parallelization of Grains3D is based on the domain decom- In this paper, we consider only dam-break collapses of gran-
position strategy [39,43]. This method consists in parti- ular columns. The interest of such configurations, frequently
tioning the computational domain into equal sub-domains, used in hydraulic applications, is to define precisely the input
each being assigned to a processor. The particle features parameters involving the geometry of the granular column in
(index, position, linear and angular velocities) are distrib- the reservoir.
uted among the appropriate processors, such as all required The lock-exchange flume is made from a rectangular res-
data are local and in control of that specific sub-domain. ervoir and horizontal channel. The flume has a 30-cm-long
In this configuration, each processor computes only the and 30-cm-high rectangular reservoir with a variable length
interaction forces and position updates of particles it con- (Li ) in which particles were inserted before being released
tains at each time step. However, particles can move freely down the horizontal surface. For each run (reported in
into the computational domain during simulations, from Table 1), a first simulation was performed to fill the res-
one processor to another. Therefore, processors need to ervoir by releasing each individual particle from a random
exchange informations with each other through their bound- cross sectional position located at its top. In this way, each
aries about each particle located nearby the sub-domain new particle is inserted in the system as soon as a geometric
boundary. Inter-processor communication must be imple- slot is available, then falling down the reservoir base by grav-
mented in the DEM code using message passing interface ity. The height and width of the channel were 30 and 4.5 cm
(MPI) library [37]. respectively (Fig. 1). The experimental procedure consists
Although flexible but complicated dynamic decompo- of partially filling the reservoir by randomly inserting the
sition algorithms [17] allow to keep load balancing of N particles of the system to form a loosely-packed heap of
individual processors through moving boundaries during length Li , height Hi , and width w (Fig. 1). Repeated runs
simulation, we used a fixed domain decomposition with sta- of the same experiment provided reproducible results asso-
tic sub-domains boundaries [21]. The aim was to devise ciated to a weak uncertainty (of around 5 %). These small
a simple and efficient 3-D parallel DEM code where the variations may be probably explained by the fast filling pro-
load balancing should minimize the latency time during cedure used in this study and that can lead to different initial
the inter-processors data exchange. The inherent synchro- configurations of particles in reservoir and associated voi-
nization of the message passing algorithm and non-block- dages. This process can cause small variations in the flow
ing communication routines significantly improve the par- properties and dynamics as highlighted by Kadau and Herr-
allel efficiency of the code, ensuring the good performance mann [23]. In simulations, the gate is moved along the flume
of computations on the distributed memory PC clusters. In in order to vary Li and then the initial aspect ratio of the col-
addition, such parallelization strategy allows performing 3-D umn (a). The gate opening velocity (0.1 m s−1 ) was chosen
discrete-grain simulations involving a reasonable number of to avoid the particle mobilization (rotation, translation) or
particles. any other disturbances in the reservoir before release. This

123
Discrete-grain model for dam-break flows 385

the horizontal plane until it came to rest. The final deposit

had a maximum height hf and total length lf (which corre-
sponds to the sum of the reservoir length Li and runout dis-
tance xf ). These laboratory investigations [27] showed that
the flow dynamics and the deposit morphologies are observed
to depend primarily on the initial aspect ratio of the granular
column (a = Hi /Li ) in the reservoir. Moreover, Lajeunesse et
al. [27] inferred empirical scaling laws for the runout distance
dependent on the initial aspect ratio of the column and on
the lock-width (w). For dam-break experiments in a narrow
channel, they observed three flow regimes: (1) for small a, an
avalanche of the column flanks produces a truncated deposit.
The flow is dominated by friction, such as the vertical accel-
eration is negligible and pressure gradients are small in com-
parison with friction forces. (2) for large a, a free fall column
leads to a conical deposit. The flow is dominated by iner-
tia, such as the vertical acceleration becomes important and
pressure gradients are significant in comparison with friction
forces. (3) for intermediate values of a, a transitional regime
between small-a flows and large-a flows, in which frictional
Fig. 1 Lock-exchange flume designed to simulate dam-break rectan- forces balanced the inertial ones, might be identified.
gular collapses described in this paper The deposits morphology leads to the following scaling
laws:
value is also similar to that obtained experimentally from an
equivalent set-up in which the gate aperture is ensured by a • For the non-dimensionalized runout distance (xf /Li ):
counterweight. The granular material considered in our sim-
ulations was monodisperse spherical particles of density ρp xf /Li ∝ a (a ≤ 3)
(8)
= 2,500 kg m−3 and of diameter dp = 3 or 5 mm. Details on xf /Li ∝ a2/3 (a ≥ 3)
each calculation are summarized in Table 1.
In order to assess the experimental scatter arose in simula- • For the non-dimensionalized deposit thickness (hf /Li ):
tions, we ran 10 times the computations 1, 2 and 3 (reported
in Table 1) and obtained a reasonable typical uncertainty hf /Li ∝ a (a ≤ 0.7)
(9)
of approximately 5 %. Results were then averaged and pre- hf /Li ∝ a1/3 (a ≥ 0.7)
sented in Fig. 4. A single run was performed for the other
computations (from 4 to 10, reported in Table 1) and pre- Note that deposit heights and runout distances both exhibit
sented directly in the following sections, taking into account regimes depending on a but characterized by different power-
the representative uncertainty for all simulations. law exponents. This crossover may be interpreted as a con-
sequence of the truncated cone deposits observed for small
values of a and in which only a small quantity of particles are
3.2 Summary of experimental results
involved, such as: Hf = Hi while L increases linearly with
Hi . Moreover, Balmforth and Kerswell [2] observed more
A series of experiments initiated by Lajeunesse et al. [27]
recently that the power-law dependence for runout distances
has generated a lot of interest and several computational
varies with the channel width. They found different scaling
studies [2,24,26,28,35]. This experimental situation, that
laws for narrow channels (xf /Li ∝ a0.65 ) and for wide chan-
models unsteady flows with variable aspect ratio in a simple
nels (xf /Li ∝ a0.9 ). The constants of proportionality of the
geometry, allows to explore the limits of the depth-averaged
power-laws were found to depend on the internal friction
approach. This paper aims to compare the results obtained
angle of the granular material.
from 3-D DE-simulations with those obtained from a similar
experimental configuration.
In this work [27], the authors have reported findings from 3.3 Choice of the model parameters
experimental dam-break granular collapses. The initial gran-
ular pile of height Hi and length Li (see Fig. 1) was then The input parameters required for each simulation are listed
released by quickly opening the sliding gate of the reser- hereafter: the Young modulus (E), the Poisson coefficient
voir closed. After release, the granular material spread on (σ ), the interparticle Coulomb friction coefficient (μc ), the

123
386 L. Girolami et al.

Table 2 Summary of the

physical parameters used both in Refs. n kn μc N dp (mm) a
the literature and in this study
This paper 0.75 4.105 0.5 3,300–15,000 3/5 0.5–18
Zenit et al. [50] 0.75 7.105 0.3/0.57 100–10,000 3.5 0.33–12.38
Staron and Hinch [45] 0.5 1 1,000–8,000 0.21–17
Cleary and Franck [10] 0.4 0.3 165,000 1.9–2.1 1.91
Lacaze et al. [26] 0.5 0.15/0.35 160–6,000 2.5/5 1.2–23.7

particle-wall Coulomb friction coefficient (μw ), the coeffi- observations of Cleary and Franck [10] from which they
cient of restitution (n ), the coefficient of rolling resistance concluded that the dynamics of the column and its final
(kms ), the particle diameter (dp ), the particle density (ρp ) shape is insensitive to this parameter. In this study, we
and the channel dimensions (Li , w, H). For input parame- chose to fix n = 0.75. Finally, from simulations we
ters related to the material properties (dp , ρp , σ ) and to the observed that the coefficient of rolling resistance (kms )
flume configuration (Li , w, H), we fixed the exact experi- characterizing the interparticle contacts has no effect on
mental values adopted in Lajeunesse et al. [27]. For the other the flow and deposits and was set to 0 in all simulations.
ones, we explored their individual effect on the flow char- We chose to fix kms = 0.133 for the particle-wall con-
acteristics. During this step, we observed that (1) E, n and tacts which is in good agreement with the values used
kms has no significant effect (i.e., negligible or less than the in Ai et al. [1] and that can also somehow reproduce
typical numerical uncertainty determined from simulations) a subtratum roughness of around 200 µm, as used in
on the flow dynamics and deposit morphologies while (2) μc Lajeunesse et al. [27].
and μw have a more important effect (i.e., slightly greater (2) The calibration of μc and μw was based on the range
than the typical uncertainty) on the flow. Therefore, E, n of values commonly used in previous analog model-
and kms were assigned to a specific value commonly used in ing efforts [10,26,45,50] and taken as references in this
the literature, whilst μc and μw were adjusted carefully in paper. Once this narrow range of values was determined,
order to both fit experimental results of Lajeunesse et al. [27] we ran simulations in order to get the best combination
and to remain consistent with the range of theoretical values of input parameters that collapsed as much as possible
reported in previous numerical studies. Details of these two with the analog laboratory experiments [27]. Result-
different steps are reported hereafter. ing values of μc and μw remained then constant for
all simulations presented in this paper. The interparticle
(1) The typical value of the Young modulus (E) for glass Coulomb friction coefficient (μc ) represents one of the
beads is commonly taken around O(109 ) Pa. This value most important parameter that controls the flow proper-
implies a short-lived contact duration that is therefore ties. Since this parameter is not well constrained exper-
very expensive in terms of computation time. From imentally, we repeated simulations by taking different
repeated runs, we observed however that smaller val- values of μc ranged from 0.2 to 0.6. We observed that
ues (taken around O(106 ) Pa) produces similar results the final run-out decreased slightly with increasing fric-
in terms of flow and deposit morphology. Moreover, tion, that could be probably explained by the increase
the associated value of kn (set as 4.105 N.m−1 ), which of the rate of dissipation of kinetic energy. However,
is then lower than in experiments, does not modify the this variation was found inferior to 8 % in each case,
flow characteristics and appears to be well suited to our indicating a variation of the internal angle of friction
simulations. This observation is confirmed by Kadau et (θc ) less than 2◦ . From calibration’s runs, we fixed μc
al. [22] who systematically investigated this issue and = 0.5 which is associated to a reasonable friction angle
observed that the stiffness coefficient has no significant of tan−1 (0.5) ∼ 26.5◦ . The same value was assigned
influence on the flow. We adopted therefore these values to μw in all simulations. Values of specific parameters
in our simulations as used in Lacaze et al. [26]. Thus, used in this paper as well as in the referenced previous
the coefficient of restitution (n ), related to the damping works are summarized in Table 2.
coefficient, can potentially influence the rate of dissipa-
tion of energy during the granular collapse, especially 4 Dynamics of the granular column collapse
by controlling the inelasticity of collisions during the
shocks. However, Staron and Hinch [45] have shown 4.1 Flow regimes
that its influence is negligible on the characteristics of
the final static pile, except for the extreme case in which When released, the granular column flowed down the flume
n = 1. This result was also in good agreement with the at speeds of up to 0.75 m s−1 until motion ceased. A set of

123
Discrete-grain model for dam-break flows 387

Fig. 2 Three sequences of images corresponding to different situations high c (a = 7.5). The time interval between each image corresponds to
of granular collapses: a the initial aspect ratio is small (a = 0.5); b the τc , except for the last image taken at the end of the flow when the heap
initial aspect ratio is moderate (a = 2); and the initial aspect ratio is is at rest

numerical experiments allowed us to investigate the different idized granular columns (see Girolami et al. [14] for the
flow regimes obtained for a wide range of initial aspect ratios description) in which the material mobilization is promoted
(0.5 ≤ a ≤ 18). In agreement with laboratory experiments, by the fluid. The temporal evolution of the flow profiles is
two different dynamics were observed. captured from simulations and are reported in Fig. 2, whilst
Flows generated from small initial aspect ratio (a = 0.5) reproducing precisely the similar laboratory experiments (see
simply consists of the fall of the edges of the initial column. Fig. 3).
Grains located inside the column have no motion and play no This paper focuses principally on the quantitative repro-
role in the spreading. Therefore, the central region remains duction of analog laboratory experiments [27]. Following this
unchanged and the shape of the final deposit looks like a target, three simulations were performed to observe the tem-
truncated cone (Fig. 2a). Flows generated from intermedi- poral evolution of the flowing material down the rectangular
ate initial aspect ratios (a = 2), i.e., for an initial column channel. In these simulations, the granular pile was charac-
approximately as tall as it is wide, give rise to a more volu- terized by a constant initial aspect ratio (a = 3.2); while the
mic avalanche of the edges only leaving a central triangular particle diameter, the reservoir length, and the initial gran-
region unchanged (Fig. 2b). The shape of this inner cone is ular mass vary from one simulation to another (see values
likely to be related to the frictional properties of the mate- of the initial conditions in Table 1). Runout distances were
rial [27]. Finally, flows generated from high initial aspect determined from the delimitation of the flow front which was
ratios (a = 7.5) involve the whole material in a vertical clearly defined in the numerical experiments, and such as the
motion, in response to gravity, causing most of the grains small amount of individual particles that overpassed the main
to take part in the dynamics. Particles moving down collide current were not taken into account in that measurement. Dur-
with the static inner pile and are pushed out with the main ing the collapse, non-dimensionalized flow thickness (h/Li )
flow. In this regime, the flow spreads out and thins until it and length (x/Li ), measured at different times, are reported in
comes to rest, leaving a smooth, elongated deposit (Fig. 2c) Fig. 4. In these numerical experiments, the characteristic time
which looks like more to those obtained from initially flu- corresponds to the free-fall time of the granular column (τc ),

123
388 L. Girolami et al.

By this way, we observed that, independently of the initial

aspect ratio and of the flow geometry, the collapse is initi-
ated by a Coulomb-type failure forming an angle of around
52◦ ± 3◦ with the horizontal plan: above which the mate-
rial slides down and below which it remains static (Fig. 5a,
b). During the flow, this interface separating the static and
flowing region, remains approximately constant (and steeply
inclined: up to 52◦ ; Fig. 5a, b) while it decreases with time
(becoming minimum at the end of the flow) and with increas-
ing initial aspect ratio. Then, for small values of a(a = 0.5),
the edge of the truncated deposit forms an angle α1 of around
22◦ with the horizontal plan (Fig. 6a). For intermadiate val-
ues of a (a = 2), the angle formed between the triangular
deposit and the horizontal (α2 ) reached around 18◦ (Fig. 6b).
For high values of a (a = 7.5), the elongated deposit forms an
inflection point that separates a steep slope of angle α3 = 16◦
from a larger, almost flat region of angle α4 = 10◦ with the
horizontal (Fig. 6c). Comparison between the internal flow
structures developed in the numerical simulations and deter-
mined from laboratory experiments (Fig. 5a, b) indicates the
precise reproduction of the dam-break collapses from our
model. A more detailed study describing the internal flow
structures developed in 3-D no-walled dam-break flows will
be established in an on-going paper in order to assess the
Fig. 3 Images sequences pointing the temporal evolution of (a) a effect of lateral spreading on the flow dynamics (i.e., the
numerical and (b) an experimental granular collapse. The time inter- evolution of internal flow structures).
val between each image corresponds to τc , except for the last image Moreover, images sequences analysis highlights that the
taken at the end of the flow when the heap is at rest. The numerical particle motion during the granular collapse down the chan-
simulation is associated with an initial column length: Li = 10 cm, an
initial aspect ratio a = 0.5 and a time interval t = 70 ms; while the nel does not mobilize the entire initial volume of the gran-
experiment of Lajeunesse et al. [27] is characterized by: Li = 10.2 cm, ular mass in the reservoir. For small a(a ≤ 1), only the
a = 0.6, t = 80 ms superficial zone of the column collapses while the principal
 underlying region remains static. Thickness of the mobile
such as τc = Hgi . According to the results of Lajeunesse region decreases with time, whilst increasing that of the sta-
et al. [27], the three different profiles reported at different tic region. A transition towards a different flow regime is
times proportional to τc collapse. Results indicate that for a observed when a is increased (1 ≤ a ≤ 3).
given granular material and substrate, the flow dynamics and For high a(a ≥ 3), the upper part of the granular mass falls,
deposit morphology depend only on the initial aspect ratio of conserving its shape, while the foot of the pile propagates
the column in the reservoir (a) and not on the volume released along the channel [27]. In this case, a significant amount of
and particle diameter. Moreover, results of Lajeunesse et al. particles is involved in the flow, increasing then the flow run-
[27] reported in Fig. 4 indicate that simulations are in good out and duration, as well as forming more elongated depos-
agreement with laboratory experiments. This highlights that its. Figure 6 indicates particles (reported in blue) which were
such 3-D DE-model involving a relevant set of input parame- mobilized in the flow. It is clear that when the initial aspect
ters can provide a precise prediction of both runout distances ratio is low (a = 0.5), only a small part of the initial volume
and runout durations, as well as the reproduction of the col- (15–20 %) is mobilized (Fig. 6a). Conversely, when the ini-
umn morphology at each time during propagation, within the tial aspect ratio is high (a = 7.5), a substantial portion of the
computation uncertainties assessed for simulations. initial volume (95 %) is mobilized (Fig. 6c). In the intermedi-
ate case (a = 2), around half of the initial volume (40–60 %)
4.2 Internal flow structure is mobilized during the collapse (Fig. 6b). This confirms the
limitations of depth-averaged models which assume that the
From the set of numerical experiments involving different entire volume of material is involved in flows characterized
initial aspect ratios, internal flow structures developed dur- by a small initial aspect ratio (a ≤ 1). Such findings are
ing propagation were first determined using a software based henceforth observed experimentally, reproduced numerically
on the motion analysis from images sequences (Fig. 5a). in this study, and explained in part theoretically in [27].

123
Discrete-grain model for dam-break flows 389

3.5 3.5

3 3

2.5 2.5

2 2

1.5 (a) 1.5 (b)

1 1

0.5 0.5

0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6

3.5 3.5

3 3

2.5 2.5

2 2

1.5 (c) 1.5 (d)

1 1

0.5 0.5

0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6

3.5 3.5

3 3

2.5 2.5

2 2

1.5 (e) 1.5 (f)

1 1

0.5 0.5

0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6

Fig. 4 Sequences of scaled profiles h(x, t)/ Li of three different granu- 650 g, a column length Li = 5.3 cm, and a particle diameter dp = 3 mm.
lar heaps characterized by the same initial aspect ratio a = 3.2. The three Numerical profiles (represented by dashed lines with triangles, squares,
different profiles have been obtained using different initial volumes or and crosses) correspond to the computations 1, 2 and 3 respectively (as
particle diameters and were measured at different times proportional refered in Table 1). The curves are barely distinguishable from experi-
to τc . The plain line and circles corresponds to the experimental val- ments
ues of Lajeunesse et al. [27] obtained from an initial granular mass of

5 Particle transport and deposit morphologies phases of transport, independent of the initial aspect ratio and
the flow regime. First, a transient acceleration phase during
5.1 Kinematics and deposit morphology which the column collapse forms a triangular pile and that
lasts during approximately 0.8τc . Then, a principal phase
Similarly to the analog laboratory experiments [27], dam- during which the flow front propagates at a nearly constant
break flows reproduced numerically exposed three distinct spreading velocity for around 2τc . Most of the total distance

123
390 L. Girolami et al.

(a) (b) 5.2 Scaling laws

From this set of numerical experiments, runout distances (xf )

defined as the difference between the total length of the final
deposit (lf ) and that of the initial column (Li ) as well as
maximum deposit thicknesses (hf ) were measured system-
atically (Fig. 8a, b). Experiments of Lajeunesse et al. [27]
are also reported in Fig. 8. From simulations, we inferred
scaling laws for the deposit morphology similarly to those
defined empirically [27]. Non-dimensionalized runout dis-
tances turned out to be proportional to the initial aspect ratio,
such as: xf /Li ∝ a for small values of it (a ≤ 3) and are well
represented by a simple power-law dependence: xf /Li ∝
a2/3 at large a (a ≥ 3). In the same manner, non-dimensional-
ized thicknesses are proportional to a, such as: hf /Li ∝ a for
small values of it (a ≤ 0.7) while it follows a simple power-
Fig. 5 Snapshot of the internal flow structures developed during the law dependence: hf /Li ∝ a1/3 at larger a (i.e., a ≥ 0.7) in
propagation of (a) a numerical and (b) an experimental granular col-
lapse. Measurements were obtained from a column of an initial gran-
agreement with results described in [27] and summarized in
ular mass of 650 g, an initial aspect ratio a = 3.2, a column length Sect. 3.2. The constants of proportionality (which both equal
Li = 5.3 cm, and a particle diameter dp = 3 mm for an averaged time to 0.99 for runout distances, while varying respectively from
of 0.3 τc . The mean slope of the interface between the static pile and 1.66 to 2.4 for thicknesses) depend on the internal friction
flowing layer is shown by the black line
angle of the granular material. In this paper, simulations took
into account only dam-break granular collapses generated in
a narrow channel. Further on-going works will deal with the
effect of the lateral spreading on the variation of such expo-
nent.

6 Summary and perspectives

We have performed dam-break rectangular collapses of gran-

ular columns using a 3-D parallel DE-model involving
soft spherical particles. Satisfactory results obtained from
our simulations evidence that quantitative reproduction and
precise prediction of dam-break rectangular collapses was
feasible in this way. Indeed, the flow dynamics, deposit
morphologies, and internal flow structures developed during
propagation have been captured properly in simulations and
demonstrate that such 3-D DE-model can provide a complete
Fig. 6 Deposits resulting from dam-break avalanches of different ini-
tial aspect ratios a a = 0.5; b a = 2; c a = 7.5 showing the parti- physical description of dam-break systems. Moreover, the
cles involved in the granular collapse. Gray particles represent particles scaling laws, indicating three different flow regimes depend-
which remain static during the total duration of the granular collapse. ing on a, are also in good agreement with the experimental
Conversely, blue particles correspond to the particles which were mobi-
predictions [27]. Therefore, it appears that these results of
lized during the flow. (Color figure online)
good quality are better than those obtained from 2-D con-
tact dynamics models of monodisperse rigid circular disks
[45,50] or of weakly polydisperse, almost rounded disks [29]
(85–90 %) traveled by the foot of the heap is covered dur- as well as from 3-D DE-models of monodisperse, soft spheres
ing this time interval. Finally, the flow front decelerates and [10]. The principal explanation of this discrepency lies proba-
comes to rest in a time interval of the order of 0.6τc . The total bly in two main factors: (1) the use of 2-D models [29,45,50]
duration of the flow lasts about 3.2τc . These observations involving rigid disks which impede the proper reproduction
highlight that the dynamics of the column collapse is gov- of the material properties (i.e., soft spheres as used in the
erned by the initial aspect ratio and by the conversion of experiments); and (2) the axisymmetric configuration [10]
potential energy into kinetic energy. that involves a different evolution of the flow thickness during

123
Discrete-grain model for dam-break flows 391

(a) (b)

Fig. 7 a Temporal evolution of the frontal position of different granular avalanches obtained from simulations (see Table 1 for the details of
calculations); b Profiles of the final deposits obtained
2 4

10
9
(a) 3 (b)
7 2
5
4
3 slope 1 slope 2/3 1
9
2 slope 1 slope 1/3
7
1 5
9
7 4
5 3
4
3 2
2

0.1 0.1
0.1 2 3 4 5 7 91 2 3 4 5 7 9 10 2 3 0.1 2 3 4 5 7 91 2 3 4 5 7 9 10 2 3

Initial aspect ratio (a) Initial aspect ratio (a)

Fig. 8 a Scaled runout distance (xf /Li ) and b scaled deposit thickness configuration using glass beads of different diameters (dp = 1.15 or
(hf /Li ) as functions of the initial aspect ratio (a). Squares correspond 3 mm). Circles and triangles correspond to the numerical collapse for
to the experiments of Lajeunesse et al. [27] performed in the dam-break small a and large a respectively

the lateral spreading. Lajeunesse et al. [27] mentionned that 2. Balmforth, N.J., Kerswell, R.R.: Granular collapse in two dimen-
this effect does not result from an experimental artifact due sions. J. Fluid Mech. 538, 399 (2005)
3. Barker, G.C.: Computer simulations of granular materials in gran-
to the sidewall friction but should result from a geometrical ular matter: an inter-disciplinary approach. In: Mehta, A. p. 317.
effect. Moreover, we should highlight also that the extreme Springer (1993)
case corresponding to a dam-break planar configuration (i.e., 4. Branney, M.J., Kokelaar, P.A.: Pyroclastic density currents and
taking w = dp ) as used in Lacaze et al. [26] provides some the sedimentation of ignimbrites. Geol. Soc. Lond. Mem. 27, 152
(2002)
limitations attributed to the cristallization effect that leads to
5. Campbell, C.S.: Rapid granular flows. Ann. Rev. Fluid Mech. 22,
nongeneric results. Consequently, a more detailed study that 57–92 (1990)
aims to examine the geometrical effects presented by the 6. Campbell, C.S., Brennen, C.E.: Computer simulation of granular
two different configurations (rectangular channel and axi- shear flows. J. Fluid Mech. 151, 167 (1985)
7. Campbell, C.S., Cleary, P.W., Hopkins, M.: Large scale land-
symmetric column), as well as involving non-spherical par-
slide simulations: global deformation, velocities and basal fric-
ticles [47], becomes a priority in order to better define the tion. J. Geophys. Res. 100, 8267–8283 (1995)
scope of such 3-D DE-models. 8. Cleary, P.W.: Discrete element modelling of industrial granular
flow applications. TASK Q. 2, 385–416 (1998)
Acknowledgments The authors thank G. Ferrer for his technical 9. Cleary, P.W., Prakash, M.: Discrete-element modelling and
assistance during numerical modeling as well as E. Lajeunesse for hav- smoothed particle hydrodynamics: potential in the environmental
ing shared his experimental data. The manuscript also benefited from sciences. Phil. Trans. R. Soc. Lond. 362, 2003–2030 (2004)
valuable comments and suggestions from two anonymous reviewers. 10. Cleary, P.W., Franck, M.: Three-dimensional discrete element sim-
ulation of axi-symmetric collapses of granular columns. http://
www.mathematik.uni-kl.de/~frank/GranularCollapsePC.pdf. Pre-
print (2006)
References 11. Cundall, P.A.: A computer model for rock-mass behaviour using
interactive graphics for the input and output of geometric data.
1. Ai, J., Chen, J.F., Rotter, J.M., Ooi, J.Y.: Assessment of rolling U. S. Nat. Tech. Inf. Serv. Rep.AD/A-001, 602 (1974)
resistance models in discrete element simulations. Powder Tech- 12. Cundall, P., Strack, O.: Discrete numerical model for granular
nol. 206(3), 269–282 (2011) assemblies. Geotechnics 29, 47–65 (1979)

123
392 L. Girolami et al.

13. Druitt, T.H., Avard, G., Bruni, G., Lettieri, P., Yates, P.: Gas reten- 33. Lube, G., Huppert, H.E., Sparks, R.S.J., Hallworth, M.: Axisym-
tion in fine-grained pyroclastic flow materials at high temperatures. metric collapse of granular columns. J. Fluid Mech. 508, 175 (2004)
Bull. Volcanol. doi:10.1007/s00445-007-0116-7 (2007) 34. Mangeney, A., Heinrich, P., Roche, R., Boudon, G., Che-
14. Girolami, L., Druitt, T.H., Roche, O., Khrabrykh, Z.: Propagation minée, J.L.: Modeling of debris avalanche and generated water
and hindered settling of laboratory ash flows. J. Geophys. Res. 113. waves: Application to real and potential events in Montser-
doi:10.1029/2007JB005074 (2008) rat. Phys. Chem. Earth 25, 741 (2000)
15. Girolami, L., Roche, O., Druitt, T.H., Corpetti, T.: Particle 35. Mangeney-Castelnau, A., Bouchut, F., Vilotte, J.P., Lajeunesse, E.,
velocity fields and depositional processes in laboratory ash Aubertin, A.: On the use of Saint-Venant equations for simulat-
flows, with implications for the sedimentation of dense pyro- ing the spreading of a granular mass. J. Geophys. Res. 110 doi:10.
clastic flows. Bull. Volc. 72, 747–759 (2010). doi:10.1007/ 1029/2004JB003161 (2005)
s00445-010-0356-9 36. Mc Clung, D., Schaerer, P.: The Avalanche Handbook. pp. 342 The
16. Grest, G.S., Duenweg, B., Kremer, K.: Link cell fortran code Mountaineers (Seattle) edition, USA (1993)
for molecular dynamics simulations for a large number of parti- 37. Pacheco, P.: Parallel Programming with MPI. pp. 500 Morgan
cles. Comp. Phys. Com. 55(3), 69–285 (1989) Kaufmann, San Francisco, CA (1997)
17. Hendrickson, B., Devine, K.: Dynamic load balancing in com- 38. Pitman, E.B., Nichita, C.C., Patra, A., Bauer, A., Sheridan, M., Bur-
putational mechanics. Comp. Meth. App. Mech. Eng. 184(2–4), sik, M.: Computing granular avalanches and landslides. Phys. Flu-
485–500 (2000) ids 15, 3638 (2003)
18. Hopkins, M., Hibler, W., Flato, G.: On the simulation of the sea 39. Plimpton, S.: Fast parallel algorithms for short range molecular
ice ridging process. J. Geophys. Res. 96, 4809–4820 (1991) dynamics. J. Comput. Phys. 117, 1 (1995)
19. Hutter, K., Siegel, M., Savage, S.B., Nohguchi, Y.: Two-dimen- 40. Pouliquen, O., Forterre, Y.: Friction law for dense granular flows:
sional spreading of a granular avalanche down an inclined application to the motion of a mass down a rough inclined
plane. Acta Math. Acad. Sci. Hung. 100, 37 (1993) plane. J. Fluid Mech. 453, 131 (2002)
20. Hutter, K., Koch, T., Pluss, C., Savage, S.B.: The dynamic of ava- 41. Radjai, F., Dubois, F.: Discrete Element Modeling of Granular
lanches of granular materials from initiation to runout. Part II: Materials. In: Hermes Sciences, Lavoisier (2010)
experiments. Acta Mech. 109, 127 (1995) 42. Savage, S.B., Hutter, K.: The motion of a finite mass of granular
21. Jabbarzadech, A., Atkinson, J.D., Tanner, R.I.: A parallel algo- material down a rough incline. J. Fluid Mech. 199, 177 (1989)
rithm for molecular dynamics simulation of branched mole- 43. Smith, B., Bjorstad, P., Gropp, W.: Domain Decomposition: par-
cules. Comp. Phys. Com. 150, 65–84 (2003) allel multilevel methods for elliptic partial differential equa-
22. Kadau, D., Schwesig, D., Theuerkauf, J., Wolf, D.E.: Influence of tions. pp. 224 Cambridge University Press, Cambridge (1996)
particle elasticity in shear testers. Gran. Matters 8(1), 35–40 (2006). 44. Sparks, R.S.J.: Particle size variations in ignimbrites and impli-
doi:10.1007/s10035-005-0217-y cations for the transport of pyroclastic flows. Sedimentology 23,
23. Kadau, D., Herrmann, H.J.: Density profiles of loose and col- 147–188 (1976)
lapsed cohesive granular structures generated by ballistic depo- 45. Staron, L., Hinch, E.J.: Study of the collapse of granular columns
sition. Phys. Rev. E 83, 031301 (2011) using 2D discrete-grains simulation. J. Fluid Mech. 545, 1 (2005)
24. Kerswell, R.R.: Dam break with coulomb friction: a model for 46. Tsuji, Y., Takana, T.: Discrete particule simulation of flow pat-
granular slumping?. Phys. Fluids 17, 057101 (2005) terns in two-dimensional gas-fluidized beds. Int. J. Mod. Phys. 7,
25. Khan, K.M., Bushell, G.: Comment on rolling friction in the 1889–1898 (1993)
dynamic simulation of sandpile formation. Phys. A. 352, 522– 47. Wachs, A., Girolami, L., Vinay, G., Ferrer, G.: Grains3D, a flex-
524 (2005) ible DEM approach for particles of arbitrary convex shape—Part
26. Lacaze, L., Phillips, J., Kerswell, R.R.: Planar collapse of a gran- I: numerical model and validations. Powder Technol. submitted
ular column: experiments and discrete-element simulations. Phys. (2011)
Fluids 20, 063302, 1–12 (2008) 48. Walton, O.R.: Numerical simulation of inelastic frictional parti-
27. Lajeunesse, E., Monnier, J.B., Homsy, G.M.: Spreading of a gran- cle-particle interaction. In: Particulate two-phase flow, pp 1002.
ular mass on a horizontal plane. Phys. Fluids 17, 103302 (2005) M. C. Roco, ch. 25, Butterworth-Heinemann Edition, Oxford
28. Larrieu, E., Staron, L., Hinch, E.J.: Raining into shallow (1994)
water as a description of the collapse of a column of 49. Wilson, C.J.N.: The role of fluidization in the emplacement of pyro-
grains. J. Fluid Mech. 554, 259–270 (2006) clastic flows: an experimental approach. J. Volcanol. Geotherm.
29. Lee, J., Herrmann, H.J.: Angle of repose and angle of marginal Res. 8, 231–249 (1980)
stability: molecular dynamics of granular particles. J. Phys. A 26, 50. Zenit, R.: Computer simulations of the collapse of a granular col-
373–383 (1993) umn. Phys. Fluids 17, 031703 (2005)
30. Levine, A.H., Kieffer, S.W.: Hydraulics of the august 7, 1980, 51. Zhang, J., Hu, Z., Ge, W., Zhang, Y., Li, T., Li, J.: Application of the
pyroclastic flow at Mount St. Helens, Washington. Geology 19, discrete approach to the simulation of size segregation in granular
1121–1124 (1991) chute flow. Ind. Eng. Chem. Res. 43, 5521–5528 (2004)
31. Li, S., Yao, Q., Chen, B., Zhang, X., Ding, Y.L.: Molecular dynam- 52. Zhou, Y.C., Wright, B.D., Yang, R.Y., Xu, B.H., Yu, A.B.: Roll-
ics simulation and continuum modelling of granular surface flow ing friction in the dynamic simulation of sandpile forma-
in rotating drums. Chin. Sci. Bull. 52(5), 692–700 (2007) tion. Phys. A. 269, 536–553 (1999)
32. Loughlin, S.C., Calder, E.S., Clarke, A.B., Cole, P.D., Luckett, R.,
Mangan, M., Pyle, D., Sparks, R.S.J., Voight, B., Watts, R.B.: Pyro-
clastic flows and surges generated by the 25 June 1997 dome col-
lapse in: Eruption of Soufriere Hills Volcano, Montserrat from 1995
to 1999. Geol. Soc. Lon. Mem. 21, 191–209 (2002)

123