Professional Documents
Culture Documents
PII: S0021-9991(13)00268-4
DOI: http://dx.doi.org/10.1016/j.jcp.2013.04.019
Reference: YJCPH 4589
Please cite this article as: X. Sun, M. Sakai, Y. Yamada, Three-dimensional simulation of a solid-liquid flow by the
DEM-SPH method, Journal of Computational Physics (2013), doi: http://dx.doi.org/10.1016/j.jcp.2013.04.019
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
Three-dimensional simulation of a solid-liquid flow by the
DEM-SPH method
a
Department of Systems Innovation, School of Engineering, The University of Tokyo,
7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan;
b
Department of Nuclear Engineering and Management, School of Engineering, The
University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan
Abstract
In this paper, we describe a new Lagrangian-Lagrangian algorithm, which is referred to be the
DEM-SPH method, for solid-liquid flows involving free surfaces. The DEM solid phase and the
SPH liquid phase are coupled using the local averaging technique described by Lagrangian ap-
proaches, where both the continuity equation and the interaction force, i.e. drag force, are connected
with the local mean voidage. Conservative forms of momentum transformation are derived for the
DEM-SPH interaction via a variational approach. By introducing a correction to the SPH approxi-
mation with explicit inclusion of boundary information, arbitrary boundaries can be modeled without
any extra wall particles, where the boundary is used commonly for both DEM and SPH phases. We
deploy level-set distance functions to efficiently construct and evaluate this boundary model. To
examine the validity of the present method, we perform three-dimensional simulations of a dynamic
flow in a solid-liquid dam break and a quasi-steady flow in a rotating cylindrical tank; and we con-
duct validation experiments to justify the simulation results. In the dam-break problem, positions of
wave fronts during the collapse are computed and compared with experimental measurements; for
the circulating tank, some macroscopic aspects of the steady flow, e.g. the shape, dimension and ve-
locity profile of the solid bed, are obtained for validation data. In both cases, the simulation results
are in good agreement with those of the experiment. Consequently, the DEM-SPH method is proved
to be adequate in modeling solid-liquid flows through this study.
*Corresponding author.
Email addresses: song@dem.t.u-tokyo.ac.jp (Xiaosong Sun),
mikio_sakai@n.t.u-tokyo.ac.jp (Mikio Sakai)
1
Keywords: Solid-liquid flow, two-phase flow, fluid-particle interaction, discrete element me-
thod, smoothed particle hydrodynamics, wall boundary conditions
1. Introduction
Solid-liquid flows are encountered in many applications such as sedimentation, slurry transport
and rheology. The solid-liquid flows are widely studied in science and engineering. Accurate numer-
ical simulations would be helpful both for understanding the complex phenomena and optimizing the
design and investigation of operational conditions in these processes. For these simulations, model-
ing of the moving boundary, a free surface fluid flow and solid-liquid coupling become important,
where a Lagrangian method can be promising. In comparison with some traditional Eulerian formu-
lations, Lagrangian schemes may be effective for modeling the moving boundaries and free surface
flows. This is because the Lagrangian approaches usually deal with discrete elements and do not
need complicated techniques like mesh generation or adaptive mesh.
As far as the Lagrangian approaches for fluid flows are concerned, some truly mesh-free me-
thods, i.e. smoothed particle hydrodynamics (SPH) [1] and the moving particle semi-implicit method
(or moving particle simulation) (MPS) [2,3], have been thoroughly studied thus far. These Lagran-
gian methods have some advantages, in that they use straightforward convection models which are
not affected by the so-called numerical diffusion, and can simulate large deformation, fragmentation
and folding of the free surface flow relatively easily. These methods can also simulate moving ob-
jects in a fluid flow with ease. Therefore, these Lagrangian approaches have been applied across a
broad range of engineering disciplines to compute various environmental or industrial fluid flows,
for example, in marine [4,5], coastal [6–9], civil engineering [10], casting [11], nuclear engineering
[12] and membrane phenomenon [13]. As for simulating discontinuous systems consisting of granu-
lar matters and powders, the discrete element method (DEM) [14] is widely recognized for its sim-
plicity and capability. The DEM is also a Lagrangian method simulating the behavior of individual
solid particles based on Newton's second law of motion. In the DEM, the contact force is evaluated
using the Voigt model consisting of springs, dashpots and a friction slider. The DEM has been ap-
plied to various powder systems, e.g. fluidized beds [15–17], ball mills [18,19], and pneumatic con-
veying [20].
When it comes to the application of these Lagrangian methods to solid-fluid flows, two types of
approaches have been employed so far. One is the direct numerical simulation (DNS) and the other
is the local averaging technique [21]. In DNS-based methods, the hydrodynamics forces acting on a
solid are evaluated by solving the Navier-Stokes equation directly. The disadvantage is that ex-
2
tremely fine resolution (i.e. the number of the SPH or MPS particles) is required to reproduce the
fluid motion precisely [5,6,22]. As a result, the DNS-based method is difficult to be applied to
large-scale practical systems, though the hydrodynamics force can be calculated accurately without
any experimental equations. On the other hand, the local volume average technique may be able to
be used to calculate a multiphase flow in a large-scale system. The governing equations transformed
by the local volume average technique have often been employed in the modeling of practical-scale
systems, e.g. gas-solid flows such as fluidized beds. Although empirical equations are needed to
calculate the drag force, the number of computation points becomes significantly smaller than that of
the DNS-based method. A Lagrangian-Lagrangian approach was developed by coupling the DEM
with the MPS method, where the effects of hydrodynamics forces and physical properties on ma-
croscopic behavior of solid particle beds were investigated in a two-dimensional system [23]. It is
illustrated that the other hydrodynamics forces except for the drag force were negligible in this sys-
tem, whilst relaxed contact forces could be applied by softer solid particles from the viewpoint of
engineering.
Similar to the MPS method, the SPH method is also featured as a Lagrangian approach that is
truly mesh-free. The SPH method has been applied to multi-component flows. For multi-fluid flows,
a variety of SPH works have been proposed to capture the discontinuity and model fluids motion
[24,25]. Inception of applying SPH method to multiphase flows with interpenetrating particles could
be dated back to some early literatures: a two-fluid model [21,26] was resolved in SPH [27] and ap-
plied to a blend of gas and dust [28], while recently an alternative attempt was to simulate the
gas-solid bubbling using SPH [29]. In the past study, the SPH method was coupled with the DEM,
where the volume average technique was introduced into the liquid phase [30]. However, this ap-
proach seemed to have some problems: not showing the variational consistency in the underlying
SPH system, traditional SPH formulations were fixed by local averaging directly in their work (in
fact, in [30] the SPH form used for the continuity equation was less compatible in the sense of weak
formulations, e.g. see [31] for a discussion of the SPH continuity equations); the wall boundary con-
dition was treated in a spring-dashpot algorithm such as the DEM, which was not a fluid dynamical
statement at all; also it was suspected that some fundamental principles, both the momentum balance
according to Newton’s third law of motion and the fluid displacement caused by immersed objects,
were not modeled precisely, since the net force of interaction terms on a single particle were not
represented properly in summations of pairwise relations. Furthermore, modeling of the wall boun-
dary condition in arbitrary shapes remained to be an arguable aspect asking for further research in
the SPH method. Consequently, adequacy of the modeling of solid-liquid flows has not been proved
by using the DEM and SPH method.
3
To handle those problems, in the present study, we carry out a new DEM-SPH algorithm, where
DEM and SPH phases are coupled using the local averaging technique. We present a novel varia-
tional formulation of the pressure-based interphase interaction to ensure conservative properties.
Suffice to say other fundamental aspects, both the reciprocal principle of Newton’s third law of mo-
tion and the displacement of immersed bodies, are modeled properly. Besides, an implementation of
wall boundaries using Level set functions is brought in for practical simulations. As a numerical
example, the three-dimensional simulation of a solid-liquid dam-break flow is performed, through
which the evolutions of the wave front are evaluated and compared with those gained from a valida-
tion test. Another numerical study is the three-dimensional simulation of a quasi-steady solid-liquid
flow in a cylindrical tank. Correspondingly a validation experiment is conducted as well. For both
test cases, the simulation results are shown in good agreements with validating experiments. Conse-
quently, the DEM-SPH method is shown to be adequate through this study.
In their approach, the mathematical concept of the local mean voidage, or the local volume
fraction of liquid phase, is the core in defining the spatial distribution of phase density. Let ε be
the local mean voidage, ρf be the actual density of the fluid phase. The locally averaged density of
the liquid phase could be expressed to be
f f . (1)
According to the local averaging technique, in a solid-liquid flow, the continuity and mo-
mentum equations for fluids are written as
D f
fvf 0, (2)
Dt
4
Dv f
f p f f
f g , (3)
Dt
where v f , p, f , and g stand for the fluid velocity, pressure, drag force per unit volume on
f
fluid phase, stress deviator tensor and gravity, respectively. Note that the drag force is a function of
the local mean voidage and empirical relations are used. On the other hand, the governing equation
describing the motion of solid particles is given as
Dv s m
ms Fs f s p FsC ms g FsM , (4)
Dt s
f C
where ms , v s , Fs and Fs are the particle mass, particle velocity, drag force on solid phase and
M
net force of solid contact forces, respectively. Fs stands for a sum of other miscellaneous effects
(e.g. lubrication and virtual mass forces, see [23]). We note that the original two-fluid model is a
continuum description and the dynamics of solid particles are transformed into continuum equations.
Extending the spirit of local mean approach, we will develop our DEM-SPH coupling method
in Section 4, where we will see that, the local mean voidage ε is extensively interwoven with the
dynamics of solid-liquid flows: it defines the constitutive behaviors of inter-phase momentum trans-
fers, and will be used in deriving a variationally consistent form of momentum equations. However,
before these arguments could be achieved, one will have to find a well-defined form for the local
averages within the context of Lagrangian systems, which is to be resolved in the following discus-
sions.
5
fluid particles and “real” solid particles, in this paper, we use subscripts i and j for fluid particles, and
preserve k and l for solid particles.
The mathematical basis of SPH frames lies in the integral interpolant theory. Consider a func-
tion A(x) of which the values are scattered among arbitrarily placed points. In the SPH method, the
conventional way to approximate or to interpolate the value of A(x) is to use the following spatial
convolution:
where W(r, h) is the smoothing kernel and h is called the smoothing length which defines the influ-
ence domain of the kernel. This integral form and its gradient, within discrete notations of the SPH
method, could be approximated by the following point-wise quadrature scheme, as indicated by Fig-
ure 1,
A ( x ) AiViWi ( x ) , (6)
i
where Vi is the tributary volume associated to particle i and it is generally evaluated as the quo-
tient of the particle mass divided by the density. We note that Ai A( xi ) and the abbreviations
with the kernel are short for
Wi ( x) W ( x xi ) , (8)
W ( x xi )
Wi ( x ) . (9)
x
Generally, kernel functions are nonnegative scalar functions with isotropic radial shapes
and a compact support radius of nh / 2(n 1,2,) . Provided these properties, it is known that
6
r dW ( r )
W (r ) W (| r |) and W (r ) . (10)
r dr
The smoothing kernel is a vital ingredient critically determining the performance of SPH simula-
tions: not only does the kernel contribute to the interpolation accuracy of SPH approximations, but
also it is related to some aspects of numerical stability. Essential properties and stability criteria of
kernels have been explored in past studies [32–35]. In this study we deployed the most commonly
used cubic spline kernel [36] (sometimes called the M4 spline kernel), which takes the following
form in three-dimension:
3 2 3 3
1 2 q 4 q 0 q 1
1 1
W ( r ; h) 3 ( 2 q ) 2 1 q 2 (11)
h 4
0 q2
where q=r/h. Apparently, the support domain of the cubic spline kernel is a sphere of which the ra-
dius is 2h. This kernel is plotted on XY plane with a unit smoothing length as shown in Figure 2.
This phenomenon has been figured out by many works and a variety of variants of the kernel
approximation have been proposed to solve the problem (e.g. see [37–39]). A lightweight fix to
this problem could be done with a least-square fitting process (the derivation is shown in Appendix
A) or equivalently following the interpretation by [40], through which the kernel approximation
could be corrected by
A ( x)
A V W ( x) ,
j j j
(12)
( x )
7
( x ) W ( x r )dr , (13)
with supp(W ) V . From the corrected kernel integral of Eq. (12), we can see that the nume-
rator is the conventional SPH summation of Eq. (6), and the denominator (which is referred to as
the correction factor in [40]) shows up as an integral of the kernel function on . Apparently, when
the region is filled by the fluid, that is V , the value of is 1 and Eq. (12) re-
gresses to the conventional SPH kernel approximation; if near the boundary, say V ,
then dividing 1 intends to account for the absence of particles beyond the rigid boundary. In
the present study, we used Eq. (12) for kernel approximation in place of the conventional Eq. (6).
Following [40], assuming a flat boundary (see Figure 4), it deserved to regularize the cor-
rection factor as a function taking a dimensionless number as its parameter,
x ( ) , (14)
with y / h and y denoting the distance between point x and the nearest rigid boundary.
Moreover, provided that the kernel function is specified, the correction factor itself could be eva-
luated analytically*. For the cubic spline kernel used in this study, integrating on the portion of the
spherical domain cut off by the plane with the help of the CAS software Maxima, we obtain the cor-
rection factor and its first-order derivative as
3 6 9 5 20 3 42 30
, 0 1,
6 60
9 5 30 4 40 3 48 28
( ) , 1 2, (15)
60
1, 2,
*In the work by [40], the correction factor and its derivative are approximated with
numerical quadrature and curve fitting.
8
18 5 45 4 60 2 42
, 0 1,
5 60
6 45 4 120 3 120 2 48
( ) , 1 2, (16)
60
0, 2,
i
m W (x ) .
j j i
(17)
i
where i ( xi ) and m j is the mass associated with the particle. We note that Eq. (17) gives an
SPH form of continuity equation as well. The summation differing much from the well-known dif-
ferential equation (2) might disconcert some readers, but it is actually a discretized form of the Eule-
rian density [41]. Thus it represents an integral formulation of the continuity equation, while its time
derivative leads to the continuity equation often used. With the summation in Eq. (17) one is able to
find out the density directly, not having to evolve the density according to Eq. (2).
In a similar way, we can express the local mean voidage as the complement of volume
fraction of nearby solid particles smoothed by the kernel function,
i 1
V W (x )
k k k i
(18)
i
with Vk the volume of a single solid particle. For the random close packing of spheres gives a fill-
ing rate of 63.4% [42], a minimum value for the local volume fraction is set to be 0.366. In fact, this
9
statement seldom gets infringed throughout the computation. Subsequently we define the reference
fluid density by
i
i . (19)
i
The averaged density estimated by SPH scheme shows the way how neighboring particles are
utilized as sample points to interpolate values. In the case of single-phase flows, the SPH density is a
synonym to that of the actual values. However, in solid-liquid flows, the number of neighboring flu-
id particles shall fall as solid particles intrude into the influence domain. As a result, both the local
SPH density and the voidage decrease on the existence of solid particles in the vicinity,
while the corresponding reference density is expected to keep unchanged roughly. Accordingly
as a macroscopic phenomenon, it can be observed that solids occupy the place and fluids are dis-
placed.
4. Numerical methodology
In the present study, the DEM-SPH method is developed for a three-dimensional simulation of
a solid-liquid flow involving free surfaces. Details of the methodology for the simulation are shown
in the following sections. Again, we use subscripts i and j for fluid particles, and k and l for solid
particles; sometimes we use characters s and f to indicate the solid and the fluid if necessary.
dvk
mk FklC mk g FklL FkD FkB , (20)
dt l l
of which mk is the particle mass, v k is the velocity of solid particle k ; on the right-hand side,
FklC , FklL , FkD and FkB are, respectively, the contact force, the lubrication force, the drag force and
the buoyancy. For rotational motion, the angular acceleration of a solid particle is given by
10
dωk
Ik TklC , (21)
dt l
where I k , ωk and Tkl are the moment of inertia, angular velocity, and torque of contact forces.
C
where the superscript n and t denote normal and tangential components, respectively. The normal
component of the contact force is given by
where k , δkl , and v kl are the stiffness, normal component of the displacement, the damping
n n
coefficient and the normal component of relative velocity, respectively. The tangential component of
the contact force is given by
11
is a trial force to determine if the contact surface slides, and , δkl , v kl are the friction
Ct t t
where F̂kl
coefficient, tangential displacement accumulation and velocity component in the tangential direction,
respectively. , or the damping coefficient, satisfies the following equation
k 2mk ml
2 ln(e) , (26)
ln (e) mk ml
2 2
12
The drag force is calculated at each particle. In the case of solid-liquid flows, it depends on
not only the relative fluid flow velocity, but also the local density of neighboring solid particles,
which could be obtained with the local mean voidage, which is estimated at each solid particle by
smoothing the nearby values of SPH particles,
k
VW (x ) .
i i i i k
(28)
VW (x )
i i i k
where W is the same function as the one used in SPH kernel approximations. The drag force is
generally given by the following form,
k
FkD (vkf vk )Vk , (29)
1 k
around particle k. A proper drag model for the description of is vital in solid-fluid interaction
problems. In the current study, a combination of the equations of Ergun [45] and Wen-Yu [46] has
been employed and a void fraction of 0.8 has been adopted as the threshold to toggle between these
two regimes. With these circumstances, is given by
(1 k ) 2 f
150 1.75(1 k ) f v kf v k k 0.8
k dk2
dk
k (30)
0.75Cd k (1 k ) f v kf v k k2.65 k 0.8
dk
where f , f and C d are the viscosity, reference density of fluid and drag coefficient for an
isolated particle, respectively. The average flow velocity is estimated using a Shepard filter as
vkf
v V W (x ) .
j j j j k
(31)
V W (x )j j j k
13
24
Re (1 0.15Rek ) Rek 1000
0.687
Cd k (32)
0.44 Rek 1000
vkf vk k f d k
Rek . (33)
f
On the other hand, the buoyancy, generated by the gradient of fluid pressure at the particle
position, is modeled by
mi pi
FkB Vk p ( xk ) Vk Wi ( xk ) . (34)
i i i
It must be underlined that the pressure gradient term appearing in Eq. (34) is not a direct imposition
of the SPH gradient model shown in Eq. (7). In fact, the unique form of the buoyancy is found from
the coupled Lagrangian dynamics of the solid-liquid system, which will be further explained in Sec-
tion 4.2.1.
14
Consider the discrete system consisting of SPH particles, for which we assume that the pres-
sure p is found from the internal energy density U (this is true in the sense of thermodynamics
and is consistent with the conventional weakly-compressible statement in SPH, see Section 4.2.2),
and any change in entropy is negligible. As a self-explanatory discretization of the continuum La-
grangian of fluids [51], we have the discrete Lagrangian of SPH particles as
1
L m j v 2j x j g U j . (35)
j 2
Also, as figured out by [52,53], with other non-conservative generic forces denoted by G, the Eu-
ler-Lagrange equation with respect to an individual particle i is given by
d L L
Gi . (36)
dt vi xi
In our scope of this paper, the non-conservative generic force is given by the sum of dissipative
viscous forces and reaction forces from solid particles,
Gi Π i Λi , (37)
where Π i is the viscous force and Λi is reaction force of the drag reflected from solid phase.
dvi
mi Ti mi g Π i Λi , (38)
dt
U j
Ti m j . (39)
j xi
Herein we call Ti the internal force of the Lagrangian system, which incidentally corresponds to
the pressure gradient term in well-known Navier-Stokes equations.
15
4.2.1. The internal force
The first law of thermodynamics tells us that without thermal effects, the internal energy
turns out to be a function of the fluid density only,
dU p
U U ( ) and 2. (40)
d
Note that ρ is not the averaged density, but the actual density. This is because no matter the way
fluids are mixed with solids, the discontinuity in energy between different materials must keep sharp
and should not be smeared out. Of course, finally we will have to convert this into its corresponding
SPH form somehow. And it will be convenient if we could connect the averaged density and local
fluid volume fraction to this internal energy, say
U U ( , ) . (41)
With Eq. (19) the definition of the actual density, we are able to obtain the partial deriva-
U j 2j p j 1 j j j
. (42)
xi j2 j xi 2j xi
The derivatives of the averaged density and local voidage with respect to particle positions are
available through direct differentiation of the SPH summations in equations (17) and (18),
j 1 1 i i
xi
j
mi W j ( xi ) ij
i
m W ( x )
I
I i I ij
i xi
, (43)
and
j 1 1 i i
xi
ij V W ( x ) , (44)
i i xi
k k k i
16
where ij is the Kronecker delta. As for the derivative of the correction factor, it is evaluated fol-
lowing [40],
i 1 d i B
niB ni , (45)
xi h d yi / h
h
B
where yi and ni are the distance and the unit vector from the nearest point on the rigid boun-
Finally, inserting equations (43) and (44) into the energy variation Eq. (42), one comes to
the internal force in the following form (a complete derivation for this are available in Appendix B),
Ti Ti f Ti s Ti w , (46)
where
j p j i pi
Ti f mi m j W j ( xi ) , (47)
2
2
j j j i i
pi
Ti s mi
i i
V W (x ) ,
k k k i (48)
pi i B
Ti w mi ni . (49)
i i h
This form of internal force is very illuminating in that, as shown by the right-hand side of Eq. (46),
the total internal force working on an SPH particle is found from three different actions*:
* Actually, other auxiliary terms may show up if traction boundaries are involved.
However, since no work is done on free surfaces where the pressure is zero, that term
vanishes naturally for free surfaces.
17
f
The first term T is due to the interaction with other fluid particles. Its form looks very alike
as a locally averaged version of the pressure gradient term recognized in ordinary SPH formu-
lations.
s
The second term T contains the positions of vicinal solid particles, so it could be ascribed to
the action of fluid pressure pressing the surrounding solid particles. Apparently, it acts as some
kind of buoyant effect acting on immersed bodies; also, since solids are involved, this term is
s
expected to be an interactive force as well. Consequently, we use T to model the buoyant
force in the underlying solid-liquid system. In fact, the antisymmetric form of Eq. (48) is used
to define the buoyancy on solid particles as we have done in Eq. (34).
w
The last term T arises from the contacting between fluid particles and wall boundaries and
corresponds to the boundary force model in [40]. Note that this force shows up only if the in-
fluence domain of the particle intersects the boundary; for those far from the boundary, or
0 , this term becomes zero.
In particular, we want to remark on the emergence of the buoyant force Ts and the boundary
force Tw. The reason why these terms appear as parts of the internal force is because the correspond-
ing information, which ε tells about the solid distribution and Γ about the rigid boundary, has
been taken into consideration evaluating the density. Meanwhile, the evaluation of the density field
is possessed by the variational description of the Lagrangian system. Thus in this model, the buoyant
force and the boundary force are discovered naturally for the underlying Lagrangian system, of
which it has been called the “variational consistency” or “coherence principle” addressing theoretical
completeness in some SPH literatures [47,54]. Furthermore, from the viewpoint of SPH practitioners,
we can figure out that the boundary model are incorporated within the governing equations explicitly
and thus no other special techniques utilized by most SPH implementations (e.g. repulsive, fixed, or
image particles, see Section 4.4 for a review) are needed.
Unfortunately, this formulation is not a painless walk-through. To fully implement and utilize
this formulation, firstly it is vital to efficiently tell the distance between a particle and the surround-
ing wall; if the wall lies within 2h distance, then the normal direction pointing from the wall to the
particle is required. In the context of computational geometry, it could be regarded as a variant of the
nearest-neighbor searching problem as the rigid boundary could be discretized using a finite set of
facets, segments or points. The original work by [40] talked about using approximations to calculate
these quantities with applications in two dimensions. However, in real-world applications, it is not as
straightforward as the 2D case since the 3D geometry complicates the solving process to a great ex-
18
tent. In the present study, the SPH was equipped with an excellent mathematical tool, the distance
function [55], to solve this problem for 3D simulations. Details will be shown in Section 4.4.
The equation of state is the one proposed by MacDonald [56]. For a weakly compressible
fluid, the fluid pressure pi can be calculated as
c 2
i 0 i i 1 i i i 0
i i 0
pi (50)
0 i i i0
where ci and i 0 are the speed of sound in the fluid at the reference density, and the density of
the fluid when not compressed, respectively. For simple ideal fluids, is often set to be 7.0. This
pressure calculation has been long employed by the SPH method as the equation of fluid state. It
represented the change in fluid pressure resulting from the variation from standard fluid density. In
this approach, when a pair of particles had come closer, the pressure between them would increase
rapidly due to Eq. (50) and this large pressure fluctuation would impose a repelling force on the par-
ticle pair. This is also the case when solid particles are submerging in the fluid region: local voidage
decreases and pressure increases, causing the fluid particle pushed out of the way.
Numerically, the sound speed c in Eq. (50) could be justified as a calculation parameter ra-
ther than some true physical property. Its value was normally determined by taking values 10 times
higher than the maximum fluid velocity magnitude to constrain the density fluctuation below 1%
according to [57].
19
ivisc ( (v v T ))
i
4x W ( x ) (51)
j
mj
ij i
v
2 ij
( i j ) xij
j
Λi Λik , (52)
k
ViWi ( x k )
Λik FkD . (53)
j V jW j ( xk )
In the SPH, it is common to introduce another dissipative term, or the so-called artificial
viscosity, into the momentum equation to damp out unphysical spurious oscillations and thus im-
prove the numerical stability. The artificial viscosity on an SPH particle is often represented in the
following way,
20
(ci c j )ij
vij xij 0
i j
ijart (55)
0 vij xij 0
vij xij
ij h , (56)
xij2 2
where 2 0.01h 2 is given to avoid the singularity when two particles coincide. is a free
parameter controlling the extent on dissipative effects and it is to be chosen according to specific
problems. Generally, the artificial viscosity produces a damping force between particles when they
approach and this force vanishes for receding particles. The computations in this study have been
assigned with α=0.075 which is chose for consideration to numerical stability only.
Meanwhile, the artificial pressure term proposed by [59] is employed to suppress tensile
instability developing. It is written as
where
W j ( xi )
f ij . (58)
W (r 0.667; h 1.0)
p p
0.01 i2 2j if pi 0 and p j 0,
Rij i j
(59)
pi pj
else.
0.
2 2
2
j
i
21
Both the artificial viscosity and the artificial pressure could be integrated with the internal
force (Eq. (47)) as
Tˆi f Ti f Π iart Ri
j p j i pi (60)
mi m j ijart Rij f ij4 W j ( xi )
j j i i
2 2
j
1) The correction factor Eq. (13) evaluated using the present particle distribution;
2) The SPH density and local voidage evaluated with equations (17) and (18);
4) Non-interactive forces in the SPH equations (38) and (46), i.e. the pressure gradient term (47),
boundary force (49), gravity and dissipative terms (51), evaluated literally;
5) The DEM contact forces (22) evaluated using the inter-particle displacements, and their torques
(21) evaluated to modify angular velocities;
6) Non-interactive forces in the DEM equation (20), i.e. the contact forces, gravity and lubrication
forces (27), evaluated in the way as they are;
7) Drag forces (29) evaluated on solid particles and reflected to fluid particles;
8) Buoyant forces (34) and (48) evaluated and accumulated for the solid phase and liquid phase
respectively.
It could be announced that now the accelerations of both the DEM and SPH particles have been
obtained and the solid-liquid particle system is ready to be updated. Any kind of time integrator
might be eligible with proper time increments. In this study, the velocities, angular velocities and
22
positions are modified with an explicit Euler integrator. With any physical quantity q and its su-
perscript denoting time steps, the scheme works as
q kn 1 q kn tqkn , (61)
Other time stepping algorithms (e.g. the predictor-corrector scheme [60], the modified Euler scheme
[48] and the modified Verlet scheme [61]) are also applicable targeting higher order convergence.
A) Boundary particles that exert repelling forces usually shaped as some kinds of molecular forces
on the fluid particles [53,62], which is a straightforward implementation for non-penetration
conditions;
B) Mirror particles analogous to the ghost cells in differential methods [24,54,63], where symme-
tric or anti-symmetric velocities against real fluid particles are assigned to imaginary particles
in the opposite side out of the wall to enforce slip or non-slip conditions;
C) Fluid particles whose positions and velocities are constrained to follow the boundary [8,64–66],
which is also the wall model used by the MPS method [2].
No matter which kind of model is used, these boundary particles are included in the neighbor-
hood of fluid particles as well as in both the continuity equation and the momentum equations to
evolve their states. In this way, boundary forces are handled naturally in SPH; although special
treatments are required to properly assign physical quantities such as density, pressure and velocity
to these wall particles.
23
As long as the particle boundary is deployed, inherently it comes with a major drawback that
the shape data defining the container’s geometry could not be used directly and some extra work is
required to generate the SPH boundaries. Especially for a curved boundary, one will have to control
the particle configuration carefully in order to get a smooth surface, which will not be a so trivial
work for three-dimensional cases (i.e. see [67] for the method applied to create particle models for
dies with complex layouts). In our DEM-SPH coupling method, this drawback could even be more
obvious: as DEM depending on mesh-based walls, there will be duplicated yet different representa-
tions of the same boundary. That is not admirable in the sense of engineering. For instance, on mod-
eling of a cylindrical tank, different views based on particles and meshes are shown in Figure 7.
Possibly another drawback is that the number of wall particles might grow to a considerable
amount, because sometimes wall particles must be multi-layered to prevent particle penetrations. In
that case, computational overheads might grow expensive because of these auxiliary wall particles.
Of course, whether the computational cost deserves our concern depends both on the concrete im-
plementation and the scale of the problem. But for practitioners of numerical simulations, most of
the time the importance of high performance could never be emphasized too much.
In the context of three-dimensional solid-liquid flows, this boundary force model outper-
forms the traditional models based on particles. It eliminates the redundant wall particle and thus
unifies the boundary representations in both SPH and DEM*, which is a virtue appreciated in
two-way couplings. Moreover, it is more computationally efficient since fewer particles are involved.
Additionally, we note that any triangle-based CAD data could be used to define the boundary layout
seamlessly for practical simulations.
* Note that what the DEM needs for evaluating boundary contacts are the same things:
the distance and normal vector from the wall, as shown in Section 4.1.1.
24
4.4.3. Distance functions as a shortcut
In this section, we show how the boundary profiles, i.e. the boundary distance and boundary
normal, are calculated. They are vital to evaluate the correction factor Γ. In the current work, fast
evaluation of these items is achieved by using distance functions.
Herein, we explain the definition and some features of DFs following [55]. Mathematically,
the DF has its root in the implicit function theorem. For a closed region Ω and its boundary ∂Ω, a
DF is defined at x as
Or equivalently, we first fix the position for the nearest point on the boundary,
then
( x ) x xC . (65)
A natural inference with DF is that evaluating (x) directly gives the unit vector pointing from
xC to x , which is the inward boundary normal vector,
n B ( x) ( x) . (66)
25
Note that the definition by Eq. (63) implies that
which gives an implicit representation of the boundary. In brief, one can identify the DF as a black
box whose input is a point and output is the steepest ascent of distance from the boundary.
Usually, DFs are implemented using a background mesh where the values of boundary
distance are calculated using prescribed boundary shape data and cached on each mesh node. When
the DF at some point is requested, we first transform the point into the local coordinates of the DF
system, next perform a simple lookup to locate the corresponding grid and then interpolate the value
using those stored on the nodes. This process is illustrated in Figure 8. In fact, in the DEM-SPH me-
thod, the canonical definition of DF by Eq. (63) could be relaxed. The reason is that, for those par-
ticles sufficiently far from the border, neither DEM rigid contact nor SPH boundary force really ap-
pears. So it is not mandatory to inform them of the precise boundary distance; instead a fake value
larger than max d k ,2h is returned. This simplifies the work when building the DF.
It is somewhat beyond the scope of this paper to devise the construction of DFs from
scratch and other tricks to avoid singularity in DFs, which would better be left for other literatures.
Anyway, as a universal tool for computational geometry, the interface of distance function makes no
difference in either DEM-SPH or the original LSM. The reader is referred to [55] for an excellent
explanation of DFs and their role in LSM.
DFs are computationally efficient since once built they are cached and reused.
Again, we want to notify that using distance function is a unified solution for particle-based physical
simulations. Both the DEM and the SPH can benefit from the same routines for boundary detections.
26
4.4.4. Boundary integral using DFs
As the preceding section described the DFs as the fundamental treatment of boundary terms,
this section will focus on the evaluation of distances and boundary forces adopting DFs. We show an
example in three-dimension, a curved boundary, to discuss the handling of boundary integrals.
The case is a circular sector cut out from the region defined by extruding a curved border.
The metric of this layout is featured with a curvature radius of 13.33h. As for the DF itself, Figure 9
shows its value normalized by the smoothing length and the boundary normal obtained by evaluating
its gradient. It can be seen that DF changes its value smoothly with respect to the distance from the
wall, while its gradient pointing inward readily estimates the normal direction.
We are now able to evaluate the correction factor directly substituting the distance
yielded by the DF in Eq. (15). At the same time, the actual value of is calculated using a numer-
ical quadrature scheme, where the integral Eq. (13) is evaluated upon very fine grids. This is illu-
strated in Figure 10 from which very close results could be confirmed. The relative error between
those two are shown in Figure 11. Comparing the immediate values and those accurate ones, we
know that merely some tiny errors reside near the boundary.
It must be noted that this discrepancy itself is totally systematic: easily using Eq. (15) in-
stead of exactly working with Eq. (13), will cause a fictitious location of the boundary, or a “visible
horizon”. Take the concave surface in Figure 12 for instance. While DF accurately telling about the
nearest boundary point, a tangentially sliced region, ( x A' B' ) in this figure, is recognized as
the integral domain, although the exact domain is ( x A B) . The same could be said for convex
boundaries as well. In effect, curves along the boundary are to be approximated by piecewise seg-
ments. It is to say that the original boundary shape might not be reproduced precisely. Nevertheless,
impenetrability still rigorously holds for this segmented view. Moreover, if the compute resolution is
sufficiently fine, the influence domain would be much smaller than the metric dimension of the
boundary shape. Thus the discrepancy will be minor, as shown in Figure 11, and the evaluation of
boundary terms could be handled gracefully without much loss of accuracy.
27
1 B
mi viB ni viB 0
Fi art t (68)
0 v 0
B
i
The damping coefficient satisfies 1 1 . A cut-off distance is set to 0.3h. If a particle gets too
close to the wall, this effect will damp out its velocity component approaching the wall like a bulk
viscosity.
With this wall viscosity, it is observed that particles tend to distribute less disordered near
the boundary. Note that this artificial viscosity is not an indispensable part of the computation al-
though it might be vital for high-velocity impingement problems. In our simulation, neither particle
leakage nor any instability has been observed even if this effect is excluded from the computation.
5. Numerical example
The following three-dimensional examples are performed to validate the proposed model. The
first test case shows a violent free surface flow to justify the application of the present rigid boun-
dary model. The second example simulates the dynamic problem of a solid-liquid two-phase dam
break flow and the third example computes a quasi-steady solid-liquid flow in a cylindrical tank, in
order to examine the adequacy of the DEM-SPH algorithm.
In the dam break flow, a rectangular water column stands to a vertical wall, as shown in the
inset to Figure 14. The water dam is 1 unit long in x-axis, 2 units high in y-axis, and 1.5 units wide
in z-axis. As the simulation begins, the water dam collapses due to downward gravity and flows out
along the floor. When the flow reaches the right, it impacts on the wall and generates a plunging
wave causing the surface to fold on itself.
28
In this study, with a unit length a, the SPH particles were set up with a regular lattice confi-
guration, where the particle distance was a/20 and the smoothing length was 1.25a/20. Physical
properties of water were used in the simulation. The rectangular tank was modeled by a box com-
posed of 12 triangle meshes, from which the DF field was constructed to give the boundary repre-
sentation. Figure 13 shows some typical snapshots of the simulation by the present SPH implemen-
tation, where it could be seen that the DF-based boundary confined the fluid motion properly. The
fluid behavior observed is similar to those reported in past studies. The leading position of the water
front toe was obtained against time as validation data. The result obtained from the present method
was compared with the VOF solution from [74], as plotted in Figure 14 where both the SPH and
VOF solutions were illustrated. It can be seen that these two agreed very well.
As mentioned in the previous section, the dam-break is a good problem to study as a fast tran-
sient phenomenon in free surface hydrodynamics. Although in [75] experiments of collapsing water
dams have been reported, it seems that solid-liquid dam-break flows under similar conditions have
never been studied thoroughly, which motivates our design of the present test case. Very recently a
DNS based fluid-particle modeling has been proposed by coupling the VOF and the DEM methods
via an immersed boundary (IB) approach [76]. Their authors have also simulated a collapse of a
submerged sphere stack in a rectangular tank. We note that, however, their exploration does not in-
volve experiments; and what’s more, in their simulations the DEM particles are quite large and
comparable in size when compared with the initial dam length, which makes their theme and prob-
lem setting quite different from ours.
29
At the initial conditions, a column of solid-liquid mixture with a dimension of a×2a×2a is
set to the left border of the tank. This set-up configuration is achieved by putting solid particles reg-
ularly aligned at the bottom of the reservoir, and letting the liquid fall freely to fill the vacancy. The
water gate is modeled by an array of fictitious particles that are restricted to prescribed positions.
After some damping steps, the simulation begins and the water gate is moved upwards along the
y-axis with a speed of Vgate=0.68 m/s (for this value, see the section below). Thus the solid-liquid
mixture is released and collapses down to the deck.
The physical properties and computational parameters used in this test are given as follows.
The stiffness, restitution coefficient and friction coefficient of the solid particles are chosen as 1000
N/m, 0.9 and 0.2, respectively. The mean diameter of solid particle is 2.7 mm and the density is 2500
kg/m3. The mass of solid particles is 200 g in total. For the liquid phase, physical properties of water
under room temperature are assumed. The average distance between liquid particles is set to be 3.0
mm, and the smoothing length is selected to have a support radius of 2h=7.5 mm. We set the sound
speed in the quasi-incompressible fluid to 100 m/s. A fixed time increment of 1.0×10-5 sec is used
throughout the simulation. A DF field similar to the one in the single-phase dam break is built to
represent the wall boundary of the tank.
When the experiment starts, the weights are unlocked. This causes the water gate to be
hitched up. We have performed independent tests of this instrument and take pictures at 1000 frames
per second as the gate rises. By measuring the distance on those photos, it is found that the gate
seems to be heaved up almost linearly and its speed is around 0.68 m/s if decided with a
least-squares fitting. This value is adopted for the simulation to better recover the situation in the
experiment.
30
5.2.3. Results and discussions
In this test we study the transitional behaviors of both the solid and liquid phases in the early
time of the dam-break flow, i.e. before the flow reaches the opposite end of the tank, which is similar
to several past literatures [74,76]. Figure 16 shows some typical snapshots of the experiment and
their counterparts during the simulation, for selected time instants of t=0.05, 0.1, 0.15 and 0.2 sec.
For both the experiment and the simulation, it appears that the liquid wave moves slightly faster than
the solids, and finally the particle beds are spread out on the bottom of the tank. The macroscopic
behaviors of the dam-break flow seem to match very well.
Next for a quantified comparison, we evaluate the propagation of the wave fronts obtained
from the simulation and experiment. By extending the notations in the previous single-phase dam
break (see the inset of Figure 14), we define a dimensionless number z* by the surge front position z
normalized by the initial dam size a:
z
z* , (70)
a
2g
t* t (71)
a
where g is absolute value of gravity. To gain the front position z from the experimental data, we have
to measure the distances in pixel on the photos and convert them to metric units afterwards. This
operation is independently repeated for three times using different baselines of reference in the im-
age, intending to reduce human errors. Those measurements of wave front positions, in raw and
normalized values for both phases, are provided in Table 1 and Table 2 for reference.
Figure 17 plots the normalized front positions against the characteristic time in the two-phase
dam break. We note that, however, just after the release of the water gate opens some tiny jets have
been observed which makes it difficult to judge the exact surge fronts. Nevertheless, it is seen that
the simulation yields results very close to those of the experimental data, for both the particle bed
and the liquid flow.
31
5.3. Solid-liquid flows in a rotating cylindrical tank
A three-dimensional simulation is performed in this study applying the DEM-SPH method to
a quasi-steady solid-liquid flow in a cylindrical tank. The results obtained from the simulations are
compared with those obtained from the experiments in the macroscopic behaviors of the solid phase,
namely, bed shape, dimensions and solid velocity distribution.
The properties of the solid and fluid phases and other parameters are shown in Table 3. The
same values of the physical properties were used in all the simulations. For the solid phase, the stiff-
ness, restitution coefficient and friction coefficient were set to be 1000 N/m, 0.8 and 0.3, respectively.
The diameter of solid particles was 2.7 mm and the density was 2500 kg/m3. For the liquid phase,
the average distance between particles was set to be 3.0 mm, and the smoothing length h was 1.25
times larger than that. The fictitious sound speed in the equation of state (Eq. (50)) took a value of 75
m/s, which would be sufficient for preserving the incompressibility of fluid phase. The calculation
5
was conducted with a time step of 1.0 10 sec.
Totally 7755 of DEM particles and 11007 of SPH particles were used in the simulation. The
total mass of DEM particles counted 1.99 10
2
g and the number of SPH particles was adjusted to
set the water level to be half the height of the cylindrical tank. The initial condition was taken as be-
ing when their mixture came to a state of rest. The DF defining the cylindrical boundary was con-
structed beginning with 1440 triangle meshes.
Glass beads of 1.99 10 g were placed in the vessel. These glass beads had an average size of
2
2.7 mm in diameter, and the particle density was 2500 kg/m3. The fluid was water. The experiment
was performed at room temperature. Photographs and videos of the rotating cylinder were recorded
by a high-speed camera PHANTOM v9.1. From these visual data, the macroscopic behaviors of the
32
solid particle bed, viz. the bed shape, width and height, were measured for validation. The velocity
distribution of the bed was measured by those pictures by using the Particle Image Velocimetry
(PIV) technique. The PIV was performed by using the DANTEC Dynamics Studio.
A representative snapshot of the simulation result is shown in Figure 20 (b). In that scene, the
shape of the bed seemed to be bilinear as well. Similarly, by picking a set of simulated snapshots
with time intervals sufficiently large, the values of solid bed width and height were averaged to be
71.45 mm (standard deviation 0.79 mm) and 59.80 mm (standard deviation 0.61 mm), respectively.
These resultant values were quantitatively comparable to the experiment, as shown in Table 4.
The velocity profiles from the experiment and the simulation were compared for further valida-
tion. To be compared with those PIV results, the compute domain was divided into grids and DEM
velocities inside those grids were filtered using a smoothing kernel. As shown in Figure 21, similar
patterns could be observed for both of them: the solid bed appeared to be circulating, the maximum
velocity was around 0.2 m/s and the central part of the bed was much slower than other parts.
Hence, it has been shown that the simulation results were reasonably close to the experimental
results in both the bed shape and the velocity distribution. The correspondence between the simu-
lated and experimental results implies that the DEM-SPH method can simulate the solid-liquid flow
accurately.
6. Conclusions
In this study, the DEM-SPH method is developed to perform three-dimensional simulations of
solid-liquid flows involving free surfaces. The solid and liquid phases are coupled using the local
averaging technique, while a pressure-based interaction term is derived with a variational approach.
By introducing distance functions, a boundary force model suitable for three-dimensional applica-
tions has been implemented.
33
The DEM-SPH method is validated using the three-dimensional examples of a solid-liquid
dam-break flow and a solid-liquid flow involving a free surface in a rotating cylindrical tank. In the
dam break test, the dynamic behaviors of the propagation of the wave fronts are studied and com-
pared with the validating experiment. It is seen that the temporal evolutions of both the solid bed and
the liquid flow have been well captured using the present model. For the quasi-steady solid-liquid
flow in the rotating tank, we compare some macroscopic aspects such as solid bed widths and
heights and the velocity profiles between the calculational and the experimental results. Alike
bi-linear bed shapes and circular flow patterns are observed, and the bed dimensions and velocities
are reasonably close between the simulation and the experiment. For both numerical examples, the
results obtained by the simulations are in good agreement with those of the experiments. Hence, we
have shown that the DEM-SPH method is an effective tool to simulate solid-liquid flow involving
free surfaces.
Acknowledgement
This study was financially supported by a Grant (22760579) from the Ministry of Educa-
tion, Culture, Sports, Science and Technology (MEXT), Japan.
J Ax
2 Ax A(r ) W (r x )dr 0 . (74)
Ax
34
A ( x)
A(r )W (r x )dr
, (75)
W (r x )dr
where the right-hand side numerator is incidentally the conventional SPH approximation by Eq. (5)
while the denominator is the correction factor .
Readers are referred to [40] for their original idea of density smoothing. In addition, for
those familiar with state-of-the-art in SPH simulations, specifically we note that this approach has a
similar form but is actually different from the so-called Shepard kernel [37,80], in which the norma-
lizing factor is obtained by summing over neighbor particles,
Shepard( x) V jW j ( x) . (76)
The Shepard kernel is a well-defined partition of unity. However, it does not contain any explicit
information about the geometry of rigid boundaries, thus it cannot help much with the boundary
force.
Let’s consider a full Lagrangian system containing all DEM solid particles and SPH fluid
particles. As claimed in the beginning of Section 4.2, the dissipationless dynamics could be suffi-
ciently described by the Lagrangian mechanics of the overall particle system. On the other hand,
dissipative parts, that are the physical/artificial viscosity and drags in our case, are able to be at-
tached to this framework as external forces. It is well known that the so-called dissipation functions
could be introduced to help define friction-like forces (in forms of relative velocities), e.g. see [35]
for a comprehensive discussion on this topic in the SPH frame. For clarity, all dissipative terms such
as drags, frictions and viscosities will be omitted in forthcoming discussions. And the conservative
forces will be referred to as the “internal” forces which include the normal contacts in solid phase,
the gradient of pressure in liquid phase, and the pressure-intermediated interaction between these
two phases.
35
With all the same notations, subscriptions i and j are used for SPH and k and l for DEM;
additionally, subscription η denotes the union of all particles. Readily the discrete Lagrangian is
written as
v2 v2
L mi i xi g U i mk k x k g Ek , (77)
i 2 k 2
where Ek for solid particles is defined as the energy density of elastic potential stored in the DEM
spring model, so we know that it is a function of the positions of the DEM particles, which is owned
to the DEM phase independently of the SPH phase. This is quite different from the SPH internal
density in that the latter depends on the configurations of the solid and liquid phases simultaneously.
Specifically, notice that both the thermal energy Ui and the elastic energy Ek are able to be written
directly as functions of particle coordinates. So one is free to straightforwardly apply the Eu-
ler-Lagrange equation (36), which is consistent with the Hamiltonian framework [83]. Subsequently,
expanding Euler-Lagrange equations,
d L L
0 (78)
dt v x
dv i U j E
mi mi g m j ml l
dt j xi l xi
(79)
U j
mi g m j
j xi
dv k U j E
mk mk g m j ml l . (80)
dt j x k l x k
Incidentally, m l
El
l x k gives the contact force in DEM phase.
36
Using equations (40) and (19), it is able to evaluate the derivative of U j with respect to
x as a total differential,
U j p j j j j
dU j j
x d j x 2j j x j x
(81)
2j p j 1 j j j
2
j j x 2j x
When the SPH phase is concerned, the symbol x stands for the position of an SPH particle.
j 1 W j ( x I ) j j
xi
j
m
I
I
xi
j xi
1 j d j x j
j
m (
I
I Ii ji )WIj
j dx j xi
(82)
1 1 j d j
j
mi Wij ij
i
m W
I
I Ii ij
j dx j
with Wab Wb ( xa ) . In a same way, differentiating Eq. (18) yields the spatial derivative of the
local voidage as
j 1 Wk ( x j ) V W (x ) j
V k k k j
Insert equations (82) and (83) into the differentiation of the thermal energy (81), and put the resultant
back into the generic SPH equation (79), then the internal force in the SPH phase could be obtained
and represented as Eq. (46).
37
For the DEM phase, Eq. (81) could be simplified as
U j 2j p j j p j
0 j j . (84)
x k j2 2j x k j x k
j 1 Wl ( x j ) 1
x k
j
Vl
l
x k
j
V (0
l
l lk )W jl
(85)
1
Vk Wkj
j
where the invariant W jk Wkj has been used. Subsequently, insert Eq. (85) into Eq. (84),
again insert the resultant form into the right-hand second term in Eq. (80). Then an interaction force
is obtained as
U j mj pj
mj Vk Wkj , (86)
j x k j j j
References
38
[3] Y. Yamada, M. Sakai, S. Mizutani, S. Koshizuka, M. Oochi, K. Murozono, Numerical
Simulation of Three-Dimensional Free-Surface Flows with Explicit Moving Particle
Simulation Method, Transactions of the Atomic Energy Society of Japan. 10 (2011)
185–193.
[5] S. Koshizuka, A. Nobe, Y. Oka, Numerical analysis of breaking waves using the
moving particle semi-implicit method, International Journal for Numerical Methods
in Fluids. 26 (1998) 751–769.
[6] H. Gotoh, T. Sakai, Key issues in the particle method for computation of wave
breaking, Coastal Engineering. 53 (2006) 171–179.
[7] J.J. Monaghan, A. Kos, Solitary Waves on a Cretan Beach, Journal of Waterway, Port,
Coastal, and Ocean Engineering. 125 (1999) 145–155.
[8] R.A. Dalrymple, B.D. Rogers, Numerical modeling of water waves with the SPH
method, Coastal Engineering. 53 (2006) 141–147.
[10] J.W. Fernandez, P.W. Cleary, M.D. Sinnott, R.D. Morrison, Using SPH one-way
coupled to DEM to model wet industrial banana screens, Minerals Engineering. 24
(2011) 741–753.
[11] P. Cleary, J. Ha, V. Alguine, T. Nguyen, Flow modelling in casting processes, Applied
Mathematical Modelling. 26 (2002) 171–190.
[13] J. Ye, J. Yang, J. Zheng, X. Ding, I. Wong, W. Li, et al., A multi-scale flow analysis in
hydrogen separation membranes using a coupled DSMC-SPH method, International
Journal of Hydrogen Energy. 37 (2012) 894–902.
39
[14] P.A. Cundall, O.D.L. Strack, A discrete numerical model for granular assemblies,
Géotechnique. 29 (1979) 47–65.
[16] M. Sakai, H. Takahashi, C.C. Pain, J.-P. Latham, J. Xiang, Study on a large-scale
discrete element model for fine particles in a fluidized bed, Advanced Powder
Technology. (n.d.).
[17] S. Yuu, T. Umekage, Y. Johno, Numerical simulation of air and particle motions in
bubbling fluidized bed of small particles, Powder Technology. 110 (2000) 158–168.
[18] H. Mori, H. Mio, J. Kano, F. Saito, Ball mill simulation in wet grinding using a
tumbling mill and its correlation to grinding rate, Powder Technology. 143-144 (2004)
230–239.
[22] A. V Potapov, M.L. Hunt, C.S. Campbell, Liquid-solid flows using smoothed particle
hydrodynamics and the discrete element method, Powder Technology. 116 (2001)
204–213.
40
[25] X.Y. Hu, N.A. Adams, An incompressible multi-phase SPH method, Journal of
Computational Physics. 227 (2007) 264–278.
[26] D. Gidaspow, Multiphase Flow and Fluidization: Continuum and Kinetic Theory
Descriptions, Academic Press, 1994.
[28] J.J. Monaghan, Implicit SPH Drag and Dusty Gas Dynamics, Journal of
Computational Physics. 138 (1997) 801–820.
[29] Q. Xiong, L. Deng, W. Wang, W. Ge, SPH method for two-fluid modeling of
particle–fluid fluidization, Chemical Engineering Science. 66 (2011) 1859–1865.
[30] D. Gao, J.A. Herbst, Alternative ways of coupling particle behaviour with fluid
dynamics in mineral processing, International Journal of Computational Fluid
Dynamics. 23 (2009) 109–118.
[32] J.J. Monaghan, Smoothed particle hydrodynamics, Annual Review of Astronomy and
Astrophysics. 30 (1992) 543–574.
[33] J.W. Swegle, D.L. Hicks, S.W. Attaway, Smoothed Particle Hydrodynamics Stability
Analysis, Journal of Computational Physics. 116 (1995) 123–134.
[34] M.B. Liu, G.R. Liu, K.Y. Lam, Z. Zong, Smoothed particle hydrodynamics for
numerical simulation of underwater explosion, Computational Mechanics. 30 (2003)
106–118.
[35] D. Violeau, Dissipative forces for Lagrangian models in computational fluid dynamics
and application to smoothed-particle hydrodynamics, Physical Review E. 80 (2009).
[36] J.J. Monaghan, J.C. Lattanzio, A refined particle method for astrophysical problems,
Astronomy and Astrophysics. 149 (1985) 135–143.
41
[37] P.W. Randles, L.D. Libersky, Smoothed Particle Hydrodynamics: Some recent
improvements and applications, Computer Methods in Applied Mechanics and
Engineering. 139 (1996) 375–408.
[38] J.-S. Chen, C. Pan, C.M.O.L. Roque, H.-P. Wang, A Lagrangian reproducing kernel
particle method for metal forming analysis, Computational Mechanics. 22 (1998)
289–307.
[41] P. Morrison, Hamiltonian description of the ideal fluid, Reviews of Modern Physics.
70 (1998) 467–521.
[42] C. Song, P. Wang, H.A. Makse, A phase diagram for jammed matter., Nature. 453
(2008) 629–32.
[43] G. Li, Y. Wang, R. He, X. Cao, C. Lin, T. Meng, Numerical simulation of predicting
and reducing solid particle erosion of solid-liquid two-phase flow in a choke,
Petroleum Science. 6 (2009) 91–97.
[44] C.T. Crowe, M. Sommerfeld, T. Yutaka, Multiphase Flows with Droplets and
Particles, Taylor & Francis, 1997.
[45] S. Ergun, Fluid flow through packed columns, Chem. Eng. Prog. 48 (1952) 89–94.
[46] C. Wen, Y. Yu, Mechanics of fluidization, Chem. Eng. Prog. Symp. Ser. 62 (1966) 100.
[47] J. Bonet, T.-S.L. Lok, Variational and momentum preservation aspects of Smooth
Particle Hydrodynamic formulations, Computer Methods in Applied Mechanics and
Engineering. 180 (1999) 97–115.
42
[49] Y. Suzuki, S. Koshizuka, Y. Oka, Hamiltonian moving-particle semi-implicit (HMPS)
method for incompressible fluid flows, Computer Methods in Applied Mechanics and
Engineering. 196 (2007) 2876–2894.
[50] D.J. Price, Smoothed Particle Hydrodynamics: Things I wish my mother taught me,
(2011) 10.
[53] J.J. Monaghan, A. Rafiee, A simple SPH algorithm for multi-fluid flow with high
density ratios, International Journal for Numerical Methods in Fluids. 71 (2012)
537–561.
[55] S. Osher, R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces (Google
eBook), Springer, 2002.
[57] J.J. Monaghan, Simulating Free Surface Flows with SPH, Journal of Computational
Physics. 110 (1994) 399–406.
[59] J.J. Monaghan, SPH without a Tensile Instability, Journal of Computational Physics.
159 (2000) 290–311.
43
[61] D. Molteni, A. Colagrossi, A simple procedure to improve the pressure evaluation in
hydrodynamic context using the SPH, Computer Physics Communications. 180
(2009) 861–872.
[62] G.-R. Liu, M.B. Liu, Smoothed Particle Hydrodynamics: A Meshfree Particle Method,
World Scientific, 2003.
[63] D. Violeau, R. Issa, Numerical modelling of complex turbulent free-surface flows with
the SPH method: an overview, International Journal for Numerical Methods in
Fluids. 53 (2007) 277–304.
[67] P.W. Cleary, J. Ha, M. Prakash, T. Nguyen, 3D SPH flow predictions and validation
for high pressure die casting of automotive components, Applied Mathematical
Modelling. 30 (2006) 1406–1427.
[69] M.W. Jones, J.A. Baerentzen, M. Sramek, 3D distance fields: a survey of techniques
and applications., IEEE Transactions on Visualization and Computer Graphics. 12
(2006) 581–99.
44
[71] S. Marrone, A. Colagrossi, D. Le Touzé, G. Graziani, Fast free-surface detection and
level-set function definition in SPH solvers, Journal of Computational Physics. 229
(2010) 3652–3663.
[73] A. Khayyer, H. Gotoh, Modified Moving Particle Semi-implicit methods for the
prediction of 2D wave impact pressure, Coastal Engineering. 56 (2009) 419–440.
[74] C.. Hirt, B.. Nichols, Volume of fluid (VOF) method for the dynamics of free
boundaries, Journal of Computational Physics. 39 (1981) 201–225.
[75] J.C. Martin, W.J. Moyce, Part IV. An Experimental Study of the Collapse of Liquid
Columns on a Rigid Horizontal Plane, Philosophical Transactions of the Royal
Society A: Mathematical, Physical and Engineering Sciences. 244 (1952) 312–324.
[76] S.-Y. Lin, Y.-C. Chen, A pressure correction-volume of fluid method for simulations of
fluid–particle interaction and impact problems, International Journal of Multiphase
Flow. 49 (2013) 31–48.
[78] G.A. Dilts, Moving least-squares particle hydrodynamics II: conservation and
boundaries, International Journal for Numerical Methods in Engineering. 48 (2000)
1503–1524.
[81] J.J. Monaghan, SPH compressible turbulence, Monthly Notices of the Royal
Astronomical Society. 335 (2002) 843–852.
45
[82] J. Bonet, M.X. Rodríguez-Paz, Hamiltonian formulation of the variable-h SPH
equations, Journal of Computational Physics. 209 (2005) 541–558.
46
Figures
47
Figure 1. The SPH kernel approximation.
48
Figure 2. Three-dimensional cubic spline kernel on X-Y plane.
49
Figure 3. Kernel integral near the boundary.
50
Figure 4. Variation of the correction factor near the boundary.
51
Figure 5. Correction factor and its derivative.
52
Figure 6. Voigt model in DEM.
53
(a) A particle wall
54
Figure 8. Implementation of distance function.
55
(a) Value of DF;
56
(a) Correction factor evaluated directly using DF;
57
Figure 11. Relative error in correction factor based on DF.
The uneven error distribution is because DFs are stored on structured grids: while projected on a
curved surface, numerical errors are introduced.
58
Figure 12. Boundary round-off.
59
Figure 13. Typical snapshots of dam break.
60
Figure 14. Position of water front during collapse of water dam.
61
Figure 15. Setting up of the solid-liquid dam-break test.
62
Figure 16. The solid-liquid dam break: typical snapshots of the experiment (top) and the
simulation (bottom).
63
Figure 17. The solid-liquid dam break: temporal variation of the wave fronts.
64
Figure 18. Schematic diagram of the cylindrical tank.
65
(a) Experiment;
(b) Simulation;
Figure 19. Initial configurations of the test case.
66
(a) Experiment;
(b) Simulation;
Figure 20. Macroscopic behaviors of the solid bed.
67
(a) PIV;
68
Tables
Table 1. The solid-liquid dam break: front positions of the particle bed obtained
from the experiment. ............................................................................................ 70
Table 2. The solid-liquid dam break: front positions of the water wave obtained from
the experiment. ..................................................................................................... 71
Table 3. Physical properties and computational parameters. .................................. 72
Table 4. Macroscopic aspects of the solid bed. ........................................................... 72
69
Table 1. The solid-liquid dam break: front positions of the particle bed obtained from the
experiment.
t (s) z-a (m) Stddev. t* (-) z* (-) Stddev.
0 0 0 0 1 0
0.01 0 0 0.198 1 0
0.02 0 0 0.396 1 0
0.03 0 0 0.594 1 0
0.04 0.0104 0.0015 0.792 1.208 0.030
0.05 0.0133 0.0006 0.990 1.266 0.012
0.06 0.0209 0.0006 1.188 1.418 0.012
0.07 0.0260 0.0007 1.386 1.521 0.013
0.08 0.0315 0.0010 1.584 1.631 0.021
0.09 0.0393 0.0004 1.782 1.786 0.008
0.1 0.0457 0.0004 1.980 1.915 0.008
0.11 0.0537 0.0008 2.178 2.074 0.016
0.12 0.0610 0.0014 2.376 2.221 0.027
0.13 0.0698 0.0020 2.574 2.396 0.040
0.14 0.0798 0.0012 2.772 2.596 0.025
0.15 0.0892 0.0018 2.970 2.785 0.036
0.16 0.0996 0.0022 3.168 2.991 0.043
0.17 0.1104 0.0020 3.366 3.208 0.039
0.18 0.1224 0.0025 3.564 3.448 0.050
0.19 0.1335 0.0022 3.762 3.670 0.044
0.197 0.1412 0.0022 3.900 3.824 0.044
70
Table 2. The solid-liquid dam break: front positions of the water wave obtained from the
experiment.
t (s) z-a (m) Stddev. t* (-) z* (-) Stddev.
0 0 0 0 1 0
0.01 0 0 0.198 1 0
0.02 0.0088 0.0006 0.396 1.175 0.012
0.03 0.0121 0.0007 0.594 1.243 0.015
0.04 0.0166 0.0004 0.792 1.332 0.008
0.05 0.0201 0.0007 0.990 1.401 0.013
0.06 0.0235 0.0007 1.188 1.469 0.015
0.07 0.0291 0.0007 1.386 1.582 0.014
0.08 0.0355 0.0011 1.584 1.709 0.021
0.09 0.0420 0.0010 1.782 1.841 0.020
0.1 0.0479 0.0016 1.980 1.958 0.031
0.11 0.0569 0.0015 2.178 2.138 0.029
0.12 0.0643 0.0011 2.376 2.286 0.023
0.13 0.0734 0.0018 2.574 2.469 0.036
0.14 0.0836 0.0020 2.772 2.671 0.040
0.15 0.0925 0.0016 2.970 2.850 0.033
0.16 0.1020 0.0028 3.168 3.040 0.056
0.17 0.1144 0.0026 3.366 3.288 0.053
0.18 0.1264 0.0027 3.564 3.528 0.054
0.19 0.1396 0.0028 3.762 3.791 0.055
0.197 0.1478 0.0028 3.900 3.955 0.055
71
Table 3. Physical properties and computational parameters.
Item Value
Solid phase
DEM particle number 7755
Particle diameter (m) 2.7e-3
Density (kg/m3) 2.5e3
Stiffness (N/m) 1.0e3
Restitution coefficient 0.9
Friction coefficient 0.3
Liquid phase
SPH particle number 11007
Average particle distance (m) 3.0e-3
Smoothing length (m) 3.75e-3
Density (kg/m3) 1.0e3
Laminar viscosity (Pa s) 1.0e-3
Virtual sound speed (m/s) 75.0
DF configuration
Number of triangle meshes 1440
Interpolation resolution (m) 1.35e-3
72
Highlights
We develop a new Lagrangian model for solid-liquid flows involving free surfaces.
The present model employs conservative forms and a compatible boundary model.
We analyze the dynamic wave propagation in a solid-liquid dam-break flow.
We simulate the quasi-steady solid-liquid flow in a rotating tank.
Simulation results are in good agreement with experimental data for validation.