Journal of Computational Physics: Xiaosong Sun, Mikio Sakai, Yoshinori Yamada

Journal of Computational Physics 248 (2013) 147176
Contents lists available at SciVerse ScienceDirect
Journal of Computational Physics

journal homepage: www.elsevier.com/locate/jcp
Three-dimensional simulation of a solidliquid ow

by the DEMSPH method
Xiaosong Sun a, Mikio Sakai b,, Yoshinori Yamada a
a
b
Department of Systems Innovation, School of Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan
Department of Nuclear Engineering and Management, School of Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-8656, Japan
a r t i c l e
i n f o
Article history:
Received 19 October 2012
Received in revised form 11 February 2013
Accepted 7 April 2013
Available online 25 April 2013
Keywords:
Solidliquid ow
Two-phase ow
Fluidparticle interaction
Discrete element method
Smoothed particle hydrodynamics
Wall boundary conditions
a b s t r a c t
In this paper, we describe a new LagrangianLagrangian algorithm, which is referred to be
the DEMSPH method, for solidliquid ows involving free surfaces. The DEM solid phase
and the SPH liquid phase are coupled using the local averaging technique described by
Lagrangian approaches, 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 DEMSPH interaction via a variational approach. By
introducing a correction to the SPH approximation 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 efciently construct and evaluate this boundary model. To examine
the validity of the present method, we perform three-dimensional simulations of a
dynamic ow in a solidliquid dam break and a quasi-steady ow in a rotating cylindrical
tank; and we conduct validation experiments to justify the simulation results. In the dambreak 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 ow, e.g. the shape, dimension and velocity prole 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 DEMSPH method is proved to be adequate in modeling solidliquid ows through this study.
2013 Elsevier Inc. All rights reserved.
1. Introduction
Solidliquid ows are encountered in many applications such as sedimentation, slurry transport and rheology. The solid
liquid ows are widely studied in science and engineering. Accurate numerical 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, modeling of the moving boundary, a free surface uid ow and solidliquid coupling become important, where a Lagrangian method can be promising. In comparison with some traditional Eulerian formulations, Lagrangian
schemes may be effective for modeling the moving boundaries and free surface ows. 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 uid ows are concerned, some truly mesh-free methods, i.e. smoothed particle
hydrodynamics (SPH) [1] and the moving particle semi-implicit method (or moving particle simulation) (MPS) [2,3], have
Corresponding author. Tel.: +81 0358416977.
E-mail addresses: song@dem.t.u-tokyo.ac.jp (X. Sun), mikio_sakai@n.t.u-tokyo.ac.jp (M. Sakai).
0021-9991/$ - see front matter 2013 Elsevier Inc. All rights reserved.
http://dx.doi.org/10.1016/j.jcp.2013.04.019
148
X. Sun et al. / Journal of Computational Physics 248 (2013) 147176
been thoroughly studied thus far. These Lagrangian 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 ow relatively easily. These methods can also simulate moving objects in a uid
ow with ease. Therefore, these Lagrangian approaches have been applied across a broad range of engineering disciplines
to compute various environmental or industrial uid ows, for example, in marine [4,5], coastal [69], civil engineering
[10], casting [11], nuclear engineering [12] and membrane phenomenon [13]. As for simulating discontinuous systems consisting of granular matters and powders, the discrete element method (DEM) [14] is widely recognized for its simplicity and
capability. The DEM is also a Lagrangian method simulating the behavior of individual solid particles based on Newtons 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 applied to various powder systems, e.g. uidized beds [1517], ball mills [18,19], and pneumatic conveying [20].
When it comes to the application of these Lagrangian methods to soliduid ows, 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 DNSbased methods, the hydrodynamics forces acting on a solid are evaluated by solving the NavierStokes equation directly. The
disadvantage is that extremely ne resolution (i.e. the number of the SPH or MPS particles) is required to reproduce the uid
motion precisely [5,6,22]. As a result, the DNS-based method is difcult 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 ow 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. gassolid ows such as uidized beds. Although empirical equations are needed to calculate the drag force,
the number of computation points becomes signicantly smaller than that of the DNS-based method. A LagrangianLagrangian approach was developed by coupling the DEM with the MPS method, where the effects of hydrodynamics forces and
physical properties on macroscopic 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 system, 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 ows. For multi-uid ows, a variety of SPH works have been proposed to
capture the discontinuity and model uids motion [24,25]. Inception of applying SPH method to multiphase ows with interpenetrating particles could be dated back to some early literatures: a two-uid model [21,26] was resolved in SPH [27] and
applied to a blend of gas and dust [28], while recently an alternative attempt was to simulate the gassolid 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 approach seemed to have some problems: not showing the variational consistency
in the underlying SPH system, traditional SPH formulations were xed 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 condition was treated in a spring-dashpot algorithm such as
the DEM, which was not a uid dynamical statement at all; also it was suspected that some fundamental principles, both the
momentum balance according to Newtons third law of motion and the uid 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 boundary condition in arbitrary shapes remained to be an
arguable aspect asking for further research in the SPH method. Consequently, adequacy of the modeling of solidliquid ows
has not been proved by using the DEM and SPH method.
To handle those problems, in the present study, we carry out a new DEMSPH algorithm, where DEM and SPH phases are
coupled using the local averaging technique. We present a novel variational formulation of the pressure-based interphase
interaction to ensure conservative properties. Sufce to say other fundamental aspects, both the reciprocal principle of Newtons third law of motion 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 solidliquid dam-break ow is performed, through which the evolutions of the wave front are evaluated and compared with those gained from a validation test. Another numerical study is the three-dimensional
simulation of a quasi-steady solidliquid ow 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. Consequently,
the DEMSPH method is shown to be adequate through this study.
2. Physical model and governing equations

As a place of departure, we will rst review the elementary concept of local mean values and the local averaging technique [21] for the mechanics of solidliquid ows in this section. The work by Anderson and Jackson [21] established a
mechanical description, namely the two-uid model, for the dynamics of particleuid uidizations. With the introduction
of local mean variables, it was able to transform the equations of motions for solid and liquid phases to represent inter-phase
momentum exchange and balance.
149
In their approach, the mathematical concept of the local mean voidage, or the local volume fraction of liquid phase, is the
core in dening the spatial distribution of phase density. Let e be the local mean voidage, qf be the actual density of the uid
phase. The locally averaged density of the liquid phase could be expressed to be
qf eqf :
According to the local averaging technique, in a solidliquid ow, the continuity and momentum equations for uids are
written as
Dqf
qf r v f 0;
Dt
qf
Dv f
f
erp f r s qf g;
Dt
where v f , p, f f, s and g stand for the uid velocity, pressure, drag force per unit volume on uid 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
ms
Dv s
ms
F fs
rp F Cs ms g F M
s ;
Dt
qs
where ms , v s , F fs and F Cs are the particle mass, particle velocity, drag force on solid phase and net force of solid contact forces,
respectively. F M
s stands for a sum of other miscellaneous effects (e.g. lubrication and virtual mass forces, see [23]). We note
that the original two-uid 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 DEMSPH coupling method in Section 4, where we will
see that, the local mean voidage e is extensively interwoven with the dynamics of solidliquid ows: it denes the constitutive behaviors of inter-phase momentum transfers, and will be used in deriving a variationally consistent form of momentum equations. However, before these arguments could be achieved, one will have to nd a well-dened form for the local
averages within the context of Lagrangian systems, which is to be resolved in the following discussions.
3. Local averaging technique in terminology of a particle method
In this section, we rst go through some fundamental parts in the SPH method. Next, we extend the local averaging technique to a discrete representation within the framework of SPH regime.
3.1. The SPH fundamentals
In the SPH method, the compute domain occupied by continuums is divided into assemblies of particles. These particles
carry physical quantities such as mass, velocity and pressure with themselves and move according to the Lagrangian material velocity. To better distinguish between SPH uid particles and real solid particles, in this paper, we use subscripts i and
j for uid particles, and preserve k and l for solid particles.
The mathematical basis of SPH frames lies in the integral interpolant theory. Consider a function 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:
hAix
ArWx r; hdr;
where W(r, h) is the smoothing kernel and h is called the smoothing length which denes the inuence 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. 1,
hAix
X
Ai V i W i x;
i
hrAix
X
Ai V i rW i x;
6
7
where Vi is the tributary volume associated to particle i and it is generally evaluated as the quotient of the particle mass
divided by the density. We note that Ai = A(xi) and the abbreviations with the kernel are short for
W i x Wx xi ;
rW i x
@Wx xi
:
@x
8
9
150
Fig. 1. The SPH kernel approximation.
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
Wr Wjrj and rWr
r dWjrj
:
jrj
dr
10
The smoothing kernel is a vital ingredient critically determining the performance of SPH simulations: 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 [3235]. 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:
8
> 1 32 q2 34 q3 ; 0 6 q < 1;
1 <1
Wr; h
2 q2 ;
1 6 q < 2;
ph3 >
:4
0;
q P 2;
11
where q = r/h. Apparently, the support domain of the cubic spline kernel is a sphere of which the radius is 2h. This kernel is
plotted on XY plane with a unit smoothing length as shown in Fig. 2.
Fig. 2. Three-dimensional cubic spline kernel on xy plane.
151
Fig. 3. Kernel integral near the boundary.
3.2. A corrected kernel approximation

The kernel approximation with Eq. (6) is adequate when the support domain of the kernel is completely enclosed inside a
given body V. However, it is not the case if the support domain intersects the boundary. Take Fig. 3 as an example. In this
case, the kernels support domain is truncated by the boundary and Eq. (6) will yield a result decreased unphysically.
This phenomenon has been gured out by many works and a variety of variants of the kernel approximation have been
proposed to solve the problem (e.g. see [3739]). A lightweight x to this problem could be done with a least-square tting
process (the derivation is shown in Appendix A) or equivalently following the interpretation by Kulasegaram et al. [40],
through which the kernel approximation could be corrected by
Aj V j W j x
;
Cx
12
Wx rdr;
13
hAix
Cx
Z
X
T
with X suppW V. From the corrected kernel integral of Eq. (12), we can see that the numerator is the conventional SPH
summation of Eq. (6), and the denominator C (which is referred to as the correction factor in [40]) shows up as an integral of
T
the kernel function on X. Apparently, when the region X is lled by the uid, that is X @V , the value of C is 1 and Eq.
T
(12) regresses to the conventional SPH kernel approximation; if near the boundary, say X @V , then dividing C < 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 at boundary (see Fig. 4), it deserved to regularize the correction factor as a function taking a
dimensionless number as its parameter,
Cjx Cv;
14
with v = y/h and y denoting the distance between point x and the nearest rigid boundary. Moreover, provided that the kernel
function is specied, the correction factor itself could be evaluated analytically.1 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
correction factor C and its rst-order derivative C0 as
8
3v6 9v5 20v3 42v 30
>
>
>
;
0 < v 6 1;

>
<
60
Cv v6 9v5 30v4 40v3 48v 28
>
; 1 < v 6 2;
>
>
60
>
:
1;
v > 2;
15
8
>
18v5 45v4 60v2 42
>
>
;
0 < v 6 1;
>
<
60
C0 v 6v5 45v4 120v3 120v2 48
>
; 1 < v 6 2;
>
>
60
>
:
0;
v > 2;
16
respectively. Plots of these functions are shown in Fig. 5.

3.3. Local mean variables in SPH
In the preceding section all the mathematical tools required to convert the local averaging to Lagrangian form have been
set. Using the corrected kernel approximation in Eq. (12), we are able to estimate the averaged density directly as
1
In the work by Kulasegaram et al. [40], the correction factor and its derivative are approximated with numerical quadrature and curve tting.
152
Fig. 4. Variation of the correction factor near the boundary.
Fig. 5. Correction factor and its derivative.
qi
mj W j xi
Ci
17
where Ci = C(xi) and mj 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 differential equation (2) might disconcert some readers,
but it is actually a discretized form of the Eulerian 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
nd 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,
ei 1
k V k W k xi
Ci
18
with Vk the volume of a single solid particle. For the random close packing of spheres gives a lling rate of 63.4% [42], a minimum value for the local volume fraction is set to be 0.366. In fact, this statement seldom gets infringed throughout the computation. Subsequently we dene the reference uid density by
qi
qi
:
ei
19
The averaged density q estimated by SPH scheme shows the way how neighboring particles are utilized as sample points to
interpolate values. In the case of single-phase ows, the SPH density is a synonym to that of the actual values. However, in
153
solidliquid ows, the number of neighboring uid particles shall fall as solid particles intrude into the inuence domain. As
a result, both the local SPH density q and the voidage e decrease on the existence of solid particles in the vicinity, while the
corresponding reference density q is expected to keep unchanged roughly. Accordingly as a macroscopic phenomenon, it can
be observed that solids occupy the place and uids are displaced.
4. Numerical methodology
In the present study, the DEMSPH method is developed for a three-dimensional simulation of a solidliquid ow involving free surfaces. Details of the methodology for the simulation are shown in the following sections. Again, we use subscripts
i and j for uid particles, and k and l for solid particles; sometimes we use characters s and f to indicate the solid and the uid
if necessary.
4.1. Solid phase: the DEM model
In the solidliquid ows, various forces are acting on a solid particle, e.g. drag, lubrication and lift forces. In many solid
liquid ows, drag force might be dominant. Therefore, drag, buoyancy, contact and gravitational forces have often been the
only forces considered in some past studies [18,43]. In this study, the forces exerted on a single particle are taken into consideration as follows:
mk
X L
dv k X C
F kl mk g
F kl F Dk F Bk ;
dt
l
l
20
of which mk is the particle mass, v k is the velocity of solid particle k; on the right-hand side, F Ckl , F Lkl , F Dk and F Bk 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
Ik
dxk X C
T kl ;
dt
l
21
where Ik, xk and T Ckl are the moment of inertia, angular velocity, and torque of contact forces.
4.1.1. The contact force
The contact force acting on a solid particle is estimated by the DEM. In the DEM, the particleparticle and particlewall
contact forces are determined from the particle overlap proles using a spring-dashpot model as shown in Fig. 6. Decomposed orthogonally according to the contact surface, the net contact force is the sum of the normal component and the tangential component,
Ct
F Ckl F Cn
kl F kl ;
22
where the superscript n and t denote normal and tangential components, respectively. The normal component of the contact
force is given by
n
n
F Cn
kl kdkl gv kl ;
dnkl ,
23
where k,
g and v are the stiffness, normal component of the displacement, the damping coefcient and the normal component of relative velocity, respectively. The tangential component of the contact force is given by
n
kl
t
t
F^ Ct
kl kdkl gv kl ;
F Ct
kl
24
8
>
< kdtkl gv tkl

Cn
jF^ Ct
kl j < lF kl
v tkl
>
: ljF Cn
kl j t
Cn
jF^ Ct
kl j P ljF kl j
jv j
kl
25
Fig. 6. Voigt model in DEM.
154
t
t
where F^ Ct
kl is a trial force to determine if the contact surface slides, and l, dkl , v kl are the friction coefcient, tangential displacement accumulation and velocity component in the tangential direction, respectively. g, or the damping coefcient, satises the following equation
s
k
2mk ml
;
g 2 lne
2
2
m
k ml
ln e p
26
where e indicates the coefcient of restitution.

4.1.2. The lubrication force
Lubrication force [44] acts as a cushion between two wet surfaces and thus tends to reduce the friction effect between
them. The lubrication force between two spheres reads
2
F Lkl
3plf dkl
v kl xkl
8jxkl j dkl
x2kl
xkl ;
27
with v kl v k v l , xkl = xk xl, and dkl dk dl =2, given that dk is the diameter of a solid particle. A cut-off distance of 2dkl
is adopted for this lubricant effect.
4.1.3. The hydrodynamic force
The hydrodynamic force could be divided into two parts, the drag force and the buoyancy. The drag force refers to the
resistant effect which acts on an object in the opposite direction of its velocity relative to the ow eld, and the buoyancy
comes from the uid pressure acting on the surface of a submerged object. Note that both the drag force and the lift force
arise from the interaction between solid particles and surrounding uids, through which momentum and energy are transported between different phases. That is why they are mentioned as interactive forces. Apparently, they must be dened
and evaluated properly to verify the reciprocal principle of interphase momentum balance. Additionally, in the local averaging technique, the form of the drag force keeps undetermined until some experimental relation could be made. In this section, we give a description of these forces in the viewpoint of solid phase and address the constitutive aspect of the drag
force.
The drag force is calculated at each particle. In the case of solidliquid ows, it depends on not only the relative uid ow
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,
e V W x
ek Pi i i i k ;
28
i V i W i xk
where W is the same function as the one used in SPH kernel approximations. The drag force is generally given by the following form,
F Dk
bk
v f v k V k ;
1 ek k
29
where bk is the interphase momentum transfer coefcient, and v fk is the average ow velocity around particle k. A proper
drag model for the description of b is vital in soliduid interaction problems. In the current study, a combination of the
equations of Ergun [45] and Wen and 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, b is given by
bk
8
2
< 150 1eek l2f 1:751 ek qd f jv fk v k j
d
ek 6 0:8
: 0:75C
ek > 0:8
ek 1ek
dk
qf jv fk v k jek2:65
30
where lf, qf and Cd are the viscosity, reference density of uid and drag coefcient for an isolated particle, respectively. The
average ow velocity is estimated using a Shepard lter as
v fk
P
v j V j W j xk
Pj
:
j V j W j xk
31
The drag coefcient Cd is given by
Cd
24
Rek
1 0:15Re0:687
Rek 6 1000
k
Rek > 1000:
0:44
32
And the particle Reynolds number Rek is given by
Rek
jv fk v k jek qf dk
lf
33
155
On the other hand, the buoyancy, generated by the gradient of uid pressure at the particle position, is modeled by
F Bk V k hrpixk V k
X mi p
i
Ci qi
rW i xk :
34
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 solidliquid system, which will be further explained in Section 4.2.1.
4.2. Liquid phase: Lagrangian-based SPH equations
For liquid phase, one will have to obtain the SPH momentum equations supervised by the local-averaging corrections. In
the present SPH practice, we do this starting with the Lagrangian mechanics of the coupled solidliquid particle system. This
could be achieved following the variational approach based on Lagrangian as done in several previous works [4749]. This
kind of formulation is prevalent in SPH literatures. As it is gured out by Dr. Price in an interesting speech [50], the non-dissipative part of SPH hydrodynamics could be naturally derived in a variational way by putting the continuity equation as a
constraint on the particles Lagrangian, while other non-conservative effects are to be introduced independently. Equivalently
herein we will adopt a procedure similar to that used in [40] to derive the equation of motion for the uid particles composing the liquid phase, which will lead us to a system of momentum equations including explicit treatment of boundary
force terms.
Consider the discrete system consisting of SPH particles, for which we assume that the pressure 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 Lagrangian of uids [51], we have the discrete Lagrangian of SPH particles as

X 1
mj
v 2j xj g Uj :
2
j
35
Also, as gured out by Bonet et al. [52] and Monaghan and Raee [53], with other non-conservative generic forces denoted
by G, the EulerLagrange equation with respect to an individual particle i is given by

d @L
@L

Gi :
dt @ v i
@xi
36
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 Pi Ki ;
37
where Pi is the viscous force and Ki is reaction force of the drag reected from solid phase.
Expanding Eq. (36) gives
mi
dv i
T i mi g Pi Ki ;
dt
X @U j
T i mj
:
@xi
j
38
39
Herein we call Ti the internal force of the Lagrangian system, which incidentally corresponds to the pressure gradient term in
well-known NavierStokes equations.
4.2.1. The internal force
The rst law of thermodynamics tells us that without thermal effects, the internal energy turns out to be a function of the
uid density only,
U Uq and
dU
p
:
dq q2
40
Note that q is not the averaged density, but the actual density. This is because no matter the way uids are mixed with solids,
the discontinuity in energy between different materials must keep sharp and should not be smeared out. Of course, nally
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 uid volume fraction to this internal energy, say
U Uq; e:
41
156
With Eq. (19) the denition of the actual density, we are able to obtain the partial derivative of Uj with respect to xi,
2
@U j ej pj
2
@xi
qj
!
1 @ qj qj @ ej

:
ej @xi e2j @xi
42
The derivatives of the averaged density and local voidage with respect to particle positions are available through direct differentiation of the SPH summations in Eqs. (17) and (18),
@ qj
1
1X
qi @ Ci
mi rW j xi dij
m rW i xI dij
;
@xi Cj
Ci I I
Ci @xi
43
(
)
@ ej
1X
1 ei @ Ci
dij
V rW k xi
;
@xi
Ci k k
Ci @xi
44
and
where dij is the Kronecker delta. As for the derivative of the correction factor, it is evaluated following [40],

@ Ci 1 dC
C0
nBi i nBi ;

@xi h dv vyi =h
h
45
where yi and nBi are the distance and the unit vector from the nearest point on the rigid boundary to the particle position,
respectively. This is illustrated in Fig. 4.
Finally, inserting Eqs. (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),
T i T fi T si T w
i ;
where
T fi
X
mi mj
ej pj ei pi
rW j xi ;
Cj q2j Ci q2i
T si mi
Tw
i mi
46
pi X
V k rW k xi ;
Ci qi
47
48
pi C0i B
n :
Ci q i h i
49
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 actions2:
The rst term T f is due to the interaction with other uid particles. Its form looks very alike as a locally averaged version
of the pressure gradient term recognized in ordinary SPH formulations.
The second term T s contains the positions of vicinal solid particles, so it could be ascribed to the action of uid 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 expected to be an interactive force as well. Consequently, we use T s to model
the buoyant force in the underlying solidliquid system. In fact, the antisymmetric form of Eq. (48) is used to dene the
buoyancy on solid particles as we have done in Eq. (34).
The last term T w arises from the contacting between uid particles and wall boundaries and corresponds to the boundary
force model in [40]. Note that this force shows up only if the inuence domain of the particle intersects the boundary; for
those far from the boundary, or C0 = 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 corresponding information, which e tells about the solid distribution and C about the rigid boundary, has been taken into consideration evaluating the density. Meanwhile, the evaluation of the density eld 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 gure 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, xed, or image particles, see Section 4.4 for a review) are needed.
2
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.
157
Unfortunately, this formulation is not a painless walk-through. To fully implement and utilize this formulation, rstly it is
vital to efciently tell the distance between a particle and the surrounding 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 nite
set of facets, segments or points. The original work by Kulasegaram et al. [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 extent. 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.
4.2.2. Fluid pressure and the equation of state
For the system described by NavierStokes equations, some relations involving the pressure could be added to give the
closure. In the original SPH method, the uid is assumed to be weakly compressible, by which the pressure is related to the
density of the uid, and thus the pressure and the velocity could be decoupled. Herein we respect this statement in our
DEMSPH implementation.
The equation of state is the one proposed by MacDonald [56]. For a weakly compressible uid, the uid pressure pi can be
calculated as

8
2
qi c
< qi0 ci
1 ; qi > ei qi0 ;
pi
c
ei qi0
:
0;
qi 6 ei qi0 ;
50
where ci and qi0 are the speed of sound in the uid at the reference density, and the density of the uid when not compressed, respectively. For simple ideal uids, c is often set to be 7.0. This pressure calculation has been long employed by
the SPH method as the equation of uid state. It represented the change in uid pressure resulting from the variation from
standard uid 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 uctuation would impose a repelling force on the particle pair. This is also the
case when solid particles are submerging in the uid region: local voidage e decreases and pressure increases, causing the
uid particle pushed out of the way.
Numerically, the sound speed c in Eq. (50) could be justied as a calculation parameter rather than some true physical
property. Its value was normally determined by taking values 10 times higher than the maximum uid velocity magnitude
to constrain the density uctuation below 1% according to [57].
4.2.3. Dissipative forces
In this work, the dissipative force is regarded as the laminar viscosity in Newtonian uids. The viscous force on a uid
particle has been proposed by Lo and Shao [58] as
Piv isc hr mrv rv T ii
X
4mxij rW i xj
mj
qi qj jxij j2
j
v ij ;
51
where m is the kinematic viscosity of the uid.

4.2.4. Reaction force of drag force
In Eq. (38) the term Ki is considered the reaction force of the drag on solid particles according to Newtons third law of
motion. It is dened as a partition of the drag force in proportion to the weight of each particle,
Ki
X
Kik ;
52
V i W i xk D
Kik P
Fk :
j V j W j xk
53
4.2.5. Other special treatments in SPH simulation

Incidentally, it has been well known that some pathologies, such as high-frequency oscillations in pressure and the socalled tensile instability, might slip into the computation and grow up to spoil the solution. Hence, some special treatments
are required to alleviate them.
In the SPH, it is common to introduce another dissipative term, or the so-called articial viscosity, into the momentum
equation to damp out unphysical spurious oscillations and thus improve the numerical stability. The articial viscosity on an
SPH particle is often represented in the following way,
X
Part
mi mj part
i
ij rW j xi :
j
54
158
A widely applied model for
part
ij
part is proposed by Monaghan [32] as
ij
8
< a ci cj /ij ;
qi qj
:
0;
/ij h
v ij xij < 0;
v ij xij P 0;
55
v ij xij
;
x2ij g2
56
where g2 = 0.01h2 is given to avoid the singularity when two particles coincide. a is a free parameter controlling the extent
on dissipative effects and it is to be chosen according to specic problems. Generally, the articial 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 a = 0.075 which is chose for consideration to numerical stability only.
Meanwhile, the articial pressure term proposed by Monaghan [59] is employed to suppress tensile instability developing. It is written as
X
Ri mi mj Rij fij4 rW j xi ;
57
where
fij
W j xi
:
Wr 0:667; h 1:0
58
The factor Rij is determined as
8
!
>
pj
pi
>
>
>
;
if pi > 0 and pj > 0;
0:01
>
<
q2i q2j
Rij
!
>
pi pj
>
>

>
>
: 0:2 q2 q2 ; else:
i
j
59
Both the articial viscosity and the articial pressure could be integrated with the internal force (Eq. (47)) as
X
b f T f Part Ri mi mj
T
i
i
i
j
ej pj ei pi
4
part
rW j xi :
ij Rij fij
Cj q2j Ci q2i
60
4.3. Algorithm and time-stepping

The DEMSPH method follows a fully explicit scheme, where all the terms in the momentum equations of both the solid
phase Eq. (20) and the liquid phase Eq. (38) are calculated according to the current time step. Sequentially, the process of a
single compute loop proceeds as:
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
The correction factor Eq. (13) evaluated using the present particle distribution;
The SPH density and local voidage evaluated with Eqs. (17) and (18);
The uid pressure calculated through the equation of state (50);
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;
The DEM contact forces (22) evaluated using the inter-particle displacements, and their torques (21) evaluated to
modify angular velocities;
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;
Drag forces (29) evaluated on solid particles and reected to uid particles;
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 positions are modied with an explicit Euler integrator. With any physical
quantity q and its superscript denoting time steps, the scheme works as
nk ;
q_ n1
q_ nk Dtq
k
61
qn1
qnk Dtq_ n1
k
k :
62
Other time stepping algorithms (e.g. the predictorcorrector scheme [60], the modied Euler scheme [48] and the modied
Verlet scheme [61]) are also applicable targeting higher order convergence.
159
4.4. Treatment on wall boundaries

4.4.1. Wall boundaries in SPH
When it comes to the wall boundary condition for the SPH method, various models have been proposed in SPH literatures.
Conventionally, almost all these models are featured with using of special wall particles, and generally they could be categorized into three main methodologies, with
(A) Boundary particles that exert repelling forces usually shaped as some kinds of molecular forces on the uid 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 symmetric or anti-symmetric
velocities against real uid 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 [8,6466], 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 neighborhood of uid 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.
As long as the particle boundary is deployed, inherently it comes with a major drawback that the shape data dening the
containers 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 conguration 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 DEMSPH coupling method, this drawback could even be more obvious: as DEM depending on
mesh-based walls, there will be duplicated yet different representations of the same boundary. That is not admirable in the
sense of engineering. For instance, on modeling of a cylindrical tank, different views based on particles and meshes are
shown in Fig. 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 implementation 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.
4.4.2. The variational formulation of boundary forces
An alternative boundary model based on a variational approach was proposed and applied to two-dimensional system by
Kulasegaram et al. [40]. In the present study, we reinvented and extended that approach. With the inclusion of local mean
voidage and a corrected average density, we gave formulation to the internal force of this two-phase coupled system; meanwhile a boundary force equation (49), which coincided with that in [40].
In the context of three-dimensional solidliquid ows, this boundary force model outperforms the traditional models
based on particles. It eliminates the redundant wall particle and thus unies the boundary representations in both SPH
and DEM,3 which is a virtue appreciated in two-way couplings. Moreover, it is more computationally efcient since fewer particles are involved. Additionally, we note that any triangle-based CAD data could be used to dene the boundary layout seamlessly for practical simulations.
4.4.3. Distance functions as a shortcut
In this section, we show how the boundary proles, i.e. the boundary distance and boundary normal, are calculated. They
are vital to evaluate the correction factor C. In the current work, fast evaluation of these items is achieved by using distance
functions.
The distance function (abbreviated to DF hereinafter) is a class of scalar functions indicating Euclidean distances from a
given point to the closest boundary point. Its inception of being applied in uid dynamics should be owned to the level set
method (LSM) [68], where it is used to dene the uid surface. Three-dimensional DFs have been long employed in calculations carrying with complex shapes to reduce computational burden [69]. However, only a few attempts have been done
with the application of DFs in SPH. In an SPH computer graphics work [70], DF was used to implement non-penetrable walls
by dening repulsive forces inversely proportional to the boundary distance. In [71], DF was constructed in uid regions to
present the SPH uid surface, for fast free-surface detection and post processing.
Herein, we explain the denition and some features of DFs following [55]. Mathematically, the DF has its root in the implicit function theorem. For a closed region X and its boundary oX, a DF is dened at "x 2 X as
3
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.
160
Fig. 7. Different representations of wall boundary. They are sectional views colored according to Z coordinates. The cylinder in (a) consists of 15,845
particles aligned radially and the one in (b) is made up of 1440 triangles. (For interpretation of the references to color in this gure legend, the reader is
referred to the web version of this article.)
ux minjx xB j for all xB 2 @ X:
63
Or equivalently, we rst x the position for the nearest point on the boundary,
xC arg minjx xB j;
64
xB 2@ X
then
ux jx xC j:
65
A natural inference with DF is that evaluating ru(x) directly gives the unit vector pointing from xC to x, which is the inward
boundary normal vector,
nB x rux:
66
Note that the denition by Eq. (63) implies that
ux 0 for all x 2 @ X;
67
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.
161
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 rst 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 Fig. 8. In fact, in the DEMSPH method,
the canonical denition of DF by Eq. (63) could be relaxed. The reason is that, for those particles sufciently far from the
border, neither DEM rigid contact nor SPH boundary force really appears. So it is not mandatory to inform them of the precise
boundary distance; instead a fake value larger than max({dk}, 2h) is returned. This simplies 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 DEMSPH or the original LSM. The reader is referred to [55]
for an excellent explanation of DFs and their role in LSM.
Using DF for boundary denition has some advantages in that:
DFs could be constructed procedurally from any mesh-based CAD data;
DFs hold for rigid motions of walls;
DFs are computationally efcient since once built they are cached and reused.
Again, we want to notify that using distance function is a unied solution for particle-based physical simulations. Both the
DEM and the SPH can benet from the same routines for boundary detections.
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 dened by extruding a curved border. The metric of this layout is
featured with a curvature radius of 13.33h. As for the DF itself, Fig. 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 C directly substituting the distance yielded by the DF in Eq. (15). At the
same time, the actual value of C is calculated using a numerical quadrature scheme, where the integral Eq. (13) is evaluated
upon very ne grids. This is illustrated in Fig. 10 from which very close results could be conrmed. The relative error between those two are shown in Fig. 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) instead of exactly working with
Eq. (13), will cause a ctitious location of the boundary, or a visible horizon. Take the concave surface in Fig. 12 for
instance. While DF accurately telling about the nearest boundary point, a tangentially sliced region, (xA0 B0 ) in this gure,
is recognized as the integral domain, although the exact domain is (xAB). The same could be said for convex boundaries as
Fig. 8. Implementation of distance function.
162
Fig. 9. DF near a curved wall.
Fig. 10. Correction factors near a curved wall.
well. In effect, curves along the boundary are to be approximated by piecewise segments. 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 sufciently ne, the inuence domain would be much smaller than the metric dimension of the boundary shape. Thus the discrepancy will be minor, as shown in Fig. 11, and the evaluation of boundary terms
could be handled gracefully without much loss of accuracy.
4.4.5. An articial force for wall boundaries
Specically, we can attach an articial force to particles near the wall boundary. This articial force imitates the articial
viscosity term used in SPH [72] and it is designed to prevent the fast penetration through the rigid boundary. This force takes
the form like
163
Fig. 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.
Fig. 12. Boundary round-off.
(
F art
nBi ;
mi v Bi n1
Dt
0;
v Bi < 0;
v Bi P 0;
v Bi v i v boundary nBi :
68
69
The damping coefcient satises 1 6 n 6 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
articial viscosity is not an indispensable part of the computation although 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 rst test case shows a violent free surface ow to justify the application of the present rigid boundary model. The second example simulates the dynamic problem of a solidliquid two-phase dam break ow and the third example computes a quasi-steady solidliquid ow
in a cylindrical tank, in order to examine the adequacy of the DEMSPH algorithm.
5.1. Single-phase ow test case: a dam break
In order to verify the present boundary model, we carried out a three-dimensional numerical test of dam break. We
choose the dam break for both its simplicity and energetic interaction between uids and boundaries. The dam break prob-
164
lem is often employed to validate the model of free surface ows: it has been used by the MPS method [2,73], SPH method
for interfacial ows [24], as well as other mesh-based numerical methods such as the volume of uid method (VOF) [74].
In the dam break ow, a rectangular water column stands to a vertical wall, as shown in the inset to Fig. 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 ows out along the oor. When the ow reaches the right, it impacts on the wall and
generates a plunging wave causing the surface to fold on itself.
In this study, with a unit length a, the SPH particles were set up with a regular lattice conguration, 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 composed of 12 triangle meshes, from which the DF eld was constructed to give
the boundary representation. Fig. 13 shows some typical snapshots of the simulation by the present SPH implementation,
where it could be seen that the DF-based boundary conned the uid motion properly. The uid 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 Fig. 14 where both the
SPH and VOF solutions were illustrated. It can be seen that these two agreed very well.
5.2. A solidliquid dam-break ow on a dry plane
This test case presents an analysis about the dynamic wave propagation on a dry bottom during the early-time evolution
of a solidliquid dam break. For a three-dimensional simulation of a mixture of water and glass beads, the DEMSPH model
Fig. 13. Typical snapshots of dam break.
165
Fig. 14. Position of water front during collapse of water dam.
is employed to predict the behavior of wave fronts after the dam is removed. A coherent validation experiment has been
performed and compared with the DEMSPH solutions.
As mentioned in the previous section, the dam-break is a good problem to study as a fast transient phenomenon in free
surface hydrodynamics. Although in [75] experiments of collapsing water dams have been reported, it seems that solid
liquid dam-break ows under similar conditions have never been studied thoroughly, which motivates our design of the
present test case. Very recently a DNS based uid-particle modeling has been proposed by coupling the VOF and the DEM
Fig. 15. Setting up of the solidliquid dam-break test.
166
Fig. 16. The solidliquid dam break: typical snapshots of the experiment (top) and the simulation (bottom).
167
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 involve experiments; and whats 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 problem setting quite different from ours.
5.2.1. Calculation conditions
Fig. 15 shows the initial setting up of the solidliquid dam-break test. Holding the notations used in the previous singlephase dam-break computation, the reference unit size a is 0.05 m. The full domain of the tank is set to be 4a 3a 2a, which
overlays the directions of Cartesian coordinates. A water gate stands at x = a.
At the initial conditions, a column of solidliquid mixture with a dimension of a 2a 2a is set to the left border of the
tank. This set-up conguration is achieved by putting solid particles regularly aligned at the bottom of the reservoir, and letting the liquid fall freely to ll the vacancy. The water gate is modeled by an array of ctitious 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 solidliquid 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
coefcient and friction coefcient 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 quasiincompressible uid to 100 m/s. A xed time increment of 1.0 105 s is used throughout the simulation. A DF eld similar
to the one in the single-phase dam break is built to represent the wall boundary of the tank.
5.2.2. Validation test
The experimental apparatus has been designed to give similar circumstances comparable to those in the numerical simulation. A rectangular tank, which is 20 cm long, 15 cm high and 10 cm wide, is used. At a quarter of the way along the tank,
rails and a watertight door are emplaced. The water gate itself is quite light and it is connected to weights through steel
wires and pulley sets. In the cabin separated by the gate, particle bed of 200 g glass beads is embedded and then compressed
lightly to atten the upper surface. After that water is poured into the space until it reaches a water level of 10 cm.
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 tting. This value is adopted for the simulation to better recover the situation in the experiment.
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 ow,
i.e. before the ow reaches the opposite end of the tank, which is similar to several past literatures [74,76]. Fig. 16 shows
Table 1
The solidliquid dam break: front positions of the particle bed obtained from the experiment.
t (s)
z a (m)
Stddev.
t ()
z ()
Stddev.
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.197
0
0
0
0
0.0104
0.0133
0.0209
0.0260
0.0315
0.0393
0.0457
0.0537
0.0610
0.0698
0.0798
0.0892
0.0996
0.1104
0.1224
0.1335
0.1412
0
0
0
0
0.0015
0.0006
0.0006
0.0007
0.0010
0.0004
0.0004
0.0008
0.0014
0.0020
0.0012
0.0018
0.0022
0.0020
0.0025
0.0022
0.0022
0
0.198
0.396
0.594
0.792
0.990
1.188
1.386
1.584
1.782
1.980
2.178
2.376
2.574
2.772
2.970
3.168
3.366
3.564
3.762
3.900
1
1
1
1
1.208
1.266
1.418
1.521
1.631
1.786
1.915
2.074
2.221
2.396
2.596
2.785
2.991
3.208
3.448
3.670
3.824
0
0
0
0
0.030
0.012
0.012
0.013
0.021
0.008
0.008
0.016
0.027
0.040
0.025
0.036
0.043
0.039
0.050
0.044
0.044
With a = 0.05 m, t t
p
2g=a and z = z/a.
168
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 s. For both the experiment and the simulation, it appears that the liquid wave moves slightly faster than the
solids, and nally the particle beds are spread out on the bottom of the tank. The macroscopic behaviors of the dam-break
ow seem to match very well.
Next for a quantied 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 Fig. 14), we dene a dimensionless number z by the surge front position z normalized by the initial dam size a:
z
z ;
a
70
Table 2
The solidliquid dam break: front positions of the water wave obtained from the experiment.
t (s)
z a (m)
Stddev.
t ()
z ()
Stddev.
0
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
0.1
0.11
0.12
0.13
0.14
0.15
0.16
0.17
0.18
0.19
0.197
0
0
0.0088
0.0121
0.0166
0.0201
0.0235
0.0291
0.0355
0.0420
0.0479
0.0569
0.0643
0.0734
0.0836
0.0925
0.1020
0.1144
0.1264
0.1396
0.1478
0
0
0.0006
0.0007
0.0004
0.0007
0.0007
0.0007
0.0011
0.0010
0.0016
0.0015
0.0011
0.0018
0.0020
0.0016
0.0028
0.0026
0.0027
0.0028
0.0028
0
0.198
0.396
0.594
0.792
0.990
1.188
1.386
1.584
1.782
1.980
2.178
2.376
2.574
2.772
2.970
3.168
3.366
3.564
3.762
3.900
1
1
1.175
1.243
1.332
1.401
1.469
1.582
1.709
1.841
1.958
2.138
2.286
2.469
2.671
2.850
3.040
3.288
3.528
3.791
3.955
0
0
0.012
0.015
0.008
0.013
0.015
0.014
0.021
0.020
0.031
0.029
0.023
0.036
0.040
0.033
0.056
0.053
0.054
0.055
0.055
With a = 0.05 m, t t
p
2g=a and z = z/a.
Fig. 17. The solidliquid dam break: temporal variation of the wave fronts.
169
and a characteristic time
t t
r
2g
;
a
71
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 image, intending to reduce human errors. Those measurements of wave
front positions, in raw and normalized values for both phases, are provided in Tables 1 and 2 for reference.
Fig. 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 difcult 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 ow.
5.3. Solidliquid ows in a rotating cylindrical tank
A three-dimensional simulation is performed in this study applying the DEMSPH method to a quasi-steady solidliquid
ow 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.
5.3.1. Calculation conditions
Fig. 18 shows a schematic diagram of the compute domain, which was a cylindrical tank with inner diameter and depth
both of 100 mm. The cylindrical tank was rotated around its central axis at 104 rpm in the simulation.
The properties of the solid and uid 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 stiffness, restitution coefcient and friction coefcient
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/
Fig. 18. Schematic diagram of the cylindrical tank.
Table 3
Physical properties and computational parameters.
Item
Value
Solid phase
DEM particle number
Particle diameter (m)
Density (kg/m3)
Stiffness (N/m)
Restitution coefcient
Friction coefcient
7755
2.7e3
2.5e3
1.0e3
0.9
0.3
Liquid phase
SPH particle number
Average particle distance (m)
Smoothing length (m)
Density (kg/m3)
Laminar viscosity (Pa s)
Virtual sound speed (m/s)
11,007
3.0e3
3.75e3
1.0e3
1.0e3
75.0
DF conguration
Number of triangle meshes
Interpolation resolution (m)
1440
1.35e3
170
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 ctitious sound speed in the equation of state (Eq. (50)) took a value of 75 m/s, which would be
sufcient for preserving the incompressibility of uid phase. The calculation was conducted with a time step of 1.0 105 s.
Totally 7755 of DEM particles and 11,007 of SPH particles were used in the simulation. The total mass of DEM particles
counted 1.99 102 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 being when their mixture came to a state of rest. The DF dening the cylindrical boundary was constructed beginning with 1440 triangle meshes.
5.3.2. Validation test
The experimental device is composed of a platform and a cylindrical jar made of glass. The cylinders inner diameter and
depth were both 100 mm. The platform was equipped with two rollers by which the cylindrical tank was driven to rotate at
104 rpm. The rotation speed was measured using a tachometer. Fig. 19 shows the initial conguration as an example.
Glass beads of 1.99 102 g were placed in the vessel. These glass beads had an average size of 2.7 mm in diameter, and
the particle density was 2500 kg/m3. The uid 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 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.
5.3.3. Results and discussions
In this study, we focused on the behavior at a quasi-steady state. The results are described below. A typical experimental
photograph is shown in Fig. 20(a). The bed slope has become bilinear in shape, as indicated by the lines in the gure. From a
Fig. 19. Initial congurations of the test case.
171
Fig. 20. Macroscopic behaviors of the solid bed.
series of photos (10 pictures) picked randomly, we measured the bed widths and heights, of which the mean results read
73.41 mm (standard deviation 1.25 mm) and 62.18 mm (standard deviation 1.71 mm), respectively.
A representative snapshot of the simulation result is shown in Fig. 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 sufciently 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 proles from the experiment and the simulation were compared for further validation. To be compared with
those PIV results, the compute domain was divided into grids and DEM velocities inside those grids were ltered using a
smoothing kernel. As shown in Fig. 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 simulated and experimental results implies that the
DEMSPH method can simulate the solidliquid ow accurately.
Table 4
Macroscopic aspects of the solid bed.
Experiment
Simulation
Error
Bed shape
Bed width (standard deviation)
Bed height (standard deviation)
Bi-linear
Bi-linear
N/A
73.41 (1.25) mm
71.45 (0.79) mm
2.67%
62.18 (1.71) mm
59.80 (0.61) mm
3.83%
172
Fig. 21. Velocity proles in the quasi-steady state.
6. Conclusions
In this study, the DEMSPH method is developed to perform three-dimensional simulations of solidliquid ows 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 applications has been implemented.
The DEMSPH method is validated using the three-dimensional examples of a solidliquid dam-break ow and a solid
liquid ow 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 compared with the validating experiment. It is seen that the temporal evolutions
of both the solid bed and the liquid ow have been well captured using the present model. For the quasi-steady solidliquid
ow in the rotating tank, we compare some macroscopic aspects such as solid bed widths and heights and the velocity proles between the calculational and the experimental results. Alike bi-linear bed shapes and circular ow 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 DEMSPH method is an effective tool to simulate solidliquid ow involving free surfaces.
Acknowledgement
This study was nancially supported by a Grant (22760579) from the Ministry of Education, Culture, Sports, Science and
Technology (MEXT), Japan.
Appendix A. An alternative interpretation of kernel approximation in SPH
It is known that the SPH method has many parts in common with Galerkin methods based on weak forms [77,78], and
even SPH itself could be formulated in a Galerkin style as shown by [79]. Thus, it is plausible to describe the kernel approx-
173
imation as an optimization process in terms of a weighted residuals scheme. Hence, we choose the best approximation by
nding a value to minimize the following error functional,
JAx
Ar Ax 2 Wr xdr;
72
hAix arg min JAx :
73
Ax 2R
This could be simply achieved by
@JAx
@Ax
2Ax ArWr xdr 0:
74
Moving the variable out of the integral, one obtains
R
hAix
XRArWr
X
xdr
;
Wr xdr
75
where the right-hand side numerator is incidentally the conventional SPH approximation by Eq. (5) while the denominator is
the correction factor C.
Readers are referred to [40] for their original idea of density smoothing. In addition, for those familiar with state-of-theart in SPH simulations, specically we note that this approach has a similar form but is actually different from the so-called
Shepard kernel [37,80], in which the normalizing factor is obtained by summing over neighbor particles,
CShepard x
V j W j x:
76
The Shepard kernel is a well-dened 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.
Appendix B. A derivation of the internal forces
Before we can start this, one must understand that the evaluation depends on whether the smoothing length is variable.
Following [40], we consider the constant smoothing length only. For the case of variable smoothing length, one may nd
some help in [81,82].
Lets consider a full Lagrangian system containing all DEM solid particles and SPH uid particles. As claimed in the beginning of Section 4.2, the dissipationless dynamics could be sufciently described by the Lagrangian mechanics of the overall
particle system. On the other hand, dissipative parts, that are the physical/articial viscosity and drags in our case, are able to
be attached to this framework as external forces. It is well known that the so-called dissipation functions could be introduced
to help dene 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.
With all the same notations, subscriptions i and j are used for SPH and k and l for DEM; additionally, subscription g denotes the union of all particles. Readily the discrete Lagrangian is written as
X 2

X v 2
v
mi i xi g U i
mk k xk g Ek ;
2
2
i
k
77
where Ek for solid particles is dened 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 congurations of the solid and liquid
phases simultaneously. Specically, 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 EulerLagrange equation (36),
which is consistent with the Hamiltonian framework [83]. Subsequently, expanding EulerLagrange equations,

d @L
@L

0
dt @ v g
@xg
78
provides momentum equations for SPH phase as
mi
X @U j X @El
X @U j
dv i
mi g
mj

ml
mi g
mj
dt
@xi
@xi
@xi
j
j
l
79
and for DEM phase
mk
X @U j X @El
dv k
mk g
mj

ml
:
dt
@xk
@xk
j
l
80
174
Incidentally,
@El
l ml @xk
gives the contact force in DEM phase.
Using Eqs. (40) and (19), it is able to evaluate the derivative of Uj with respect to xg as a total differential,

e2j pj
@U j dU j @ qj
p @ qj @ qj @ qj @ ej
2
@xg dqj @xg qj @ qj @xg @ ej @xg

qj
!
1 @ qj qj @ ej
:

ej @xg e2j @xg
81
When the SPH phase is concerned, the symbol xg stands for the position of an SPH particle. Therefore, the spatial derivative of
averaged density (Eq. (17)) is calculated as
@ qj
1 X @W j xI qj @ Cj
1X
qj dCj @xj
mI
m d dji rW Ij
@xi
@xi Cj I
Cj @xi Cj I I Ii
Cj dxj @xi
Cj
mi rW ij dij
1X
qj dCj
mI rW Ii dij
C
j dxj
I
Ci
82
with rWab = rWb(xa). In a same way, differentiating Eq. (18) yields the spatial derivative of the local voidage as
@ ej
1 X @W k xj

V
@xi
Cj k k @xi
k V k W k xj
2
j
(
)
@ Cj
1X
1 ej dCj
:
dij
V rW jk
@xi
Cj k k
Cj dxj
83
Insert Eqs. (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).
For the DEM phase, Eq. (81) could be simplied as
2
@U j ej pj
qj @ ej
2 0 2
@xk
qj
ej @xk
!
pj @ ej
:
qj @xk
84
Similarly, the spatial derivative of ej is given by
@ ej
1 X @W l xj
1X
1

V

V 0 dlk rW jl V k rW kj ;
@xk
Cj l l @xk
Cj l l
Cj
85
where the invariant rWjk = rWkj 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
X @U j
X mj pj
mj

V rW kj ;
@xk
Cj q j k
j
j
86
which is actually the buoyant force shown in Eq. (34).

References
[1] J.J. Monaghan, An introduction to SPH, Computer Physics Communications 48 (1988) 8996.
[2] S. Koshizuka, Y. Oka, Moving-particle semi-implicit method for fragmentation of incompressible uid, Nuclear Science and Engineering 123 (1996)
421434.
[3] Y. Yamada, M. Sakai, S. Mizutani, S. Koshizuka, M. Oochi, K. Murozono, Numerical simulation of three-dimensional free-surface ows with explicit
moving particle simulation method, Transactions of the Atomic Energy Society of Japan 10 (2011) 185193.
[4] K. Shibata, S. Koshizuka, Numerical analysis of shipping water impact on a deck using a particle method, Journal of Waterway, Port, Coastal, and Ocean
Engineering 34 (2007) 585593.
[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) 751769.
[6] H. Gotoh, T. Sakai, Key issues in the particle method for computation of wave breaking, Coastal Engineering 53 (2006) 171179.
[7] J.J. Monaghan, A. Kos, Solitary waves on a Cretan Beach, Journal of Waterway, Port, Coastal, and Ocean Engineering 125 (1999) 145155.
[8] R.A. Dalrymple, B.D. Rogers, Numerical modeling of water waves with the SPH method, Coastal Engineering 53 (2006) 141147.
[9] D. Violeau, C. Buvat, K. Abed-Meraim, E. de Nanteuil, Numerical modelling of boom and oil spill with SPH, Coastal Engineering 54 (2007) 895913.
[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) 741753.
[11] P. Cleary, J. Ha, V. Alguine, T. Nguyen, Flow modelling in casting processes, Applied Mathematical Modelling 26 (2002) 171190.
[12] X. Sun, M. Sakai, K. Shibata, Y. Tochigi, H. Fujiwara, Numerical modeling on the discharged uid ow from a glass melter by a Lagrangian approach,
Nuclear Engineering and Design 248 (2012) 1421.
[13] J. Ye, J. Yang, J. Zheng, X. Ding, I. Wong, W. Li, et al, A multi-scale ow analysis in hydrogen separation membranes using a coupled DSMCSPH method,
International Journal of Hydrogen Energy 37 (2012) 894902.
[14] P.A. Cundall, O.D.L. Strack, A discrete numerical model for granular assemblies, Gotechnique 29 (1979) 4765.
[15] M. Sakai, Y. Yamada, Y. Shigeto, K. Shibata, V.M. Kawasaki, S. Koshizuka, Large-scale discrete element modeling in a uidized bed, International Journal
for Numerical Methods in Fluids 64 (2010) 13191335.
[16] M. Sakai, H. Takahashi, C.C. Pain, J.-P. Latham, J. Xiang, Study on a large-scale discrete element model for ne particles in a uidized bed, Advanced
Powder Technology 23 (2012) 673681.
[17] S. Yuu, T. Umekage, Y. Johno, Numerical simulation of air and particle motions in bubbling uidized bed of small particles, Powder Technology 110
(2000) 158168.
[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) 230239.
175
[19] M. Sakai, K. Shibata, S. Koshizuka, Effect of nuclear fuel particle movement on nuclear criticality in a rotating cylindrical vessel, Journal of Nuclear
Science and Technology 42 (2005) 267274.
[20] M. Sakai, S. Koshizuka, Large-scale discrete element modeling in pneumatic conveying, Chemical Engineering Science 64 (2009) 533539.
[21] T.B. Anderson, R. Jackson, Fluid mechanical description of uidized beds. Equations of motion, Industrial and Engineering Chemistry Fundamentals 6
(1967) 527539.
[22] A.V. Potapov, M.L. Hunt, C.S. Campbell, Liquidsolid ows using smoothed particle hydrodynamics and the discrete element method, Powder
Technology 116 (2001) 204213.
[23] M. Sakai, Y. Shigeto, X. Sun, T. Aoki, T. Saito, J. Xiong, et al, LagrangianLagrangian modeling for a solidliquid ow in a cylindrical tank, Chemical
Engineering Journal 200202 (2012) 663672.
[24] A. Colagrossi, M. Landrini, Numerical simulation of interfacial ows by smoothed particle hydrodynamics, Journal of Computational Physics 191 (2003)
448475.
[25] X.Y. Hu, N.A. Adams, An incompressible multi-phase SPH method, Journal of Computational Physics 227 (2007) 264278.
[26] D. Gidaspow, Multiphase Flow and Fluidization: Continuum and Kinetic Theory Descriptions, Academic Press, 1994.
[27] J.J. Monaghan, A. Kocharyan, SPH simulation of multi-phase ow, Computer Physics Communications 87 (1995) 225235.
[28] J.J. Monaghan, Implicit SPH drag and dusty gas dynamics, Journal of Computational Physics 138 (1997) 801820.
[29] Q. Xiong, L. Deng, W. Wang, W. Ge, SPH method for two-uid modeling of particleuid uidization, Chemical Engineering Science 66 (2011) 1859
1865.
[30] D. Gao, J.A. Herbst, Alternative ways of coupling particle behaviour with uid dynamics in mineral processing, International Journal of Computational
Fluid Dynamics 23 (2009) 109118.
[31] D.J. Price, Modelling discontinuities and KelvinHelmholtz instabilities in SPH, Journal of Computational Physics 227 (2008) 1004010057.
[32] J.J. Monaghan, Smoothed particle hydrodynamics, Annual Review of Astronomy and Astrophysics 30 (1992) 543574.
[33] J.W. Swegle, D.L. Hicks, S.W. Attaway, Smoothed particle hydrodynamics stability analysis, Journal of Computational Physics 116 (1995) 123134.
[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) 106118.
[35] D. Violeau, Dissipative forces for Lagrangian models in computational uid dynamics and application to smoothed-particle hydrodynamics, Physical
Review E 80 (2009).
[36] J.J. Monaghan, J.C. Lattanzio, A rened particle method for astrophysical problems, Annual Review of Astronomy and Astrophysics 149 (1985) 135143.
[37] P.W. Randles, L.D. Libersky, Smoothed particle hydrodynamics: some recent improvements and applications, Computer Methods in Applied Mechanics
and Engineering 139 (1996) 375408.
[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) 289307.
[39] J. Bonet, S. Kulasegaram, Correction and stabilization of smooth particle hydrodynamics methods with applications in metal forming simulations,
International Journal for Numerical Methods in Engineering 47 (2000) 11891214.
[40] S. Kulasegaram, J. Bonet, R.W. Lewis, M. Prot, A variational formulation based contact algorithm for rigid boundaries in two-dimensional SPH
applications, Computational Mechanics 33 (2004) 316325.
[41] P. Morrison, Hamiltonian description of the ideal uid, Reviews of Modern Physics 70 (1998) 467521.
[42] C. Song, P. Wang, H.A. Makse, A phase diagram for jammed matter, Nature 453 (2008) 629632.
[43] G. Li, Y. Wang, R. He, X. Cao, C. Lin, T. Meng, Numerical simulation of predicting and reducing solid particle erosion of solidliquid two-phase ow in a
choke, Petroleum Science 6 (2009) 9197.
[44] C.T. Crowe, M. Sommerfeld, T. Yutaka, Multiphase Flows with Droplets and Particles, Taylor & Francis, 1997.
[45] S. Ergun, Fluid ow through packed columns, Chemical Engineering Progress 48 (1952) 8994.
[46] C. Wen, Y. Yu, Mechanics of uidization, Chemical Engineering Progress Symposium Series 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) 97115.
[48] J.J. Monaghan, Smoothed particle hydrodynamics, Reports on Progress in Physics 68 (2005) 17031759.
[49] Y. Suzuki, S. Koshizuka, Y. Oka, Hamiltonian moving-particle semi-implicit (HMPS) method for incompressible uid ows, Computer Methods in
Applied Mechanics and Engineering 196 (2007) 28762894.
[50] D.J. Price, Smoothed Particle Hydrodynamics: Things I wish my mother taught me, (2011) 10.
[51] C. Eckart, Variation principles of hydrodynamics, Physics of Fluids 3 (1960) 421.
[52] J. Bonet, S. Kulasegaram, M.X. Rodriguez-Paz, M. Prot, Variational formulation for the smooth particle hydrodynamics (SPH) simulation of uid and
solid problems, Computer Methods in Applied Mechanics and Engineering 193 (2004) 12451256.
[53] J.J. Monaghan, A. Raee, A simple SPH algorithm for multi-uid ow with high density ratios, International Journal for Numerical Methods in Fluids 71
(2012) 537561.
[54] G. Oger, M. Doring, B. Alessandrini, P. Ferrant, An improved SPH method: towards higher order convergence, Journal of Computational Physics 225
(2007) 14721492.
[55] S. Osher, R. Fedkiw, Level Set Methods and Dynamic Implicit Surfaces, Springer, 2002 (Google eBook).
[56] J. MacDonald, Some simple isothermal equations of state, Reviews of Modern Physics 38 (1966) 669679.
[57] J.J. Monaghan, Simulating free surface ows with SPH, Journal of Computational Physics 110 (1994) 399406.
[58] E.Y.M. Lo, S. Shao, Simulation of near-shore solitary wave mechanics by an incompressible SPH method, Applied Ocean Research 24 (2002) 275286.
[59] J.J. Monaghan, SPH without a tensile instability, Journal of Computational Physics 159 (2000) 290311.
[60] M. Gomez-Gesteira, B.D. Rogers, A.J.C. Crespo, R.A. Dalrymple, M. Narayanaswamy, J.M. Dominguez, SPHysics development of a free-surface uid
solver Part 1: theory and formulations, Computers & Geosciences 48 (2012) 289299.
[61] D. Molteni, A. Colagrossi, A simple procedure to improve the pressure evaluation in hydrodynamic context using the SPH, Computer Physics
Communications 180 (2009) 861872.
[62] G.-R. Liu, M.B. Liu, Smoothed Particle Hydrodynamics: A Meshfree Particle Method, World Scientic, 2003.
[63] D. Violeau, R. Issa, Numerical modelling of complex turbulent free-surface ows with the SPH method: an overview, International Journal for
Numerical Methods in Fluids 53 (2007) 277304.
[64] M. Gmez-Gesteira, D. Cerqueiro, C. Crespo, R.A. Dalrymple, Green water overtopping analyzed with a SPH model, Journal of Waterway, Port, Coastal,
and Ocean Engineering 32 (2005) 223238.
[65] E.-S. Lee, C. Moulinec, R. Xu, D. Violeau, D. Laurence, P. Stansby, Comparisons of weakly compressible and truly incompressible algorithms for the SPH
mesh free particle method, Journal of Computational Physics 227 (2008) 84178436.
[66] J.P. Hughes, D.I. Graham, Comparison of incompressible and weakly-compressible SPH models for free-surface water ows, Journal of Hydraulic
Research 48 (2010) 105117.
[67] P.W. Cleary, J. Ha, M. Prakash, T. Nguyen, 3D SPH ow predictions and validation for high pressure die casting of automotive components, Applied
Mathematical Modelling 30 (2006) 14061427.
[68] S. Osher, J.A. Sethian, Fronts propagating with curvature-dependent speed: algorithms based on HamiltonJacobi formulations, Journal of
Computational Physics 79 (1988) 1249.
[69] M.W. Jones, J.A. Baerentzen, M. Sramek, 3D distance elds: a survey of techniques and applications, IEEE Transactions on Visualization and Computer
Graphics 12 (2006) 581599.
176
[70] T. Harada, S. Koshizuka, Y. Kawaguchi, Smoothed particle hydrodynamics in complex shapes, in: Spring Conference on Computer Graphics, 2007, pp.
235241.
[71] S. Marrone, A. Colagrossi, D. Le Touz, G. Graziani, Fast free-surface detection and level-set function denition in SPH solvers, Journal of Computational
Physics 229 (2010) 36523663.
[72] J.J. Monaghan, On the problem of penetration in particle methods, Journal of Computational Physics 82 (1989) 115.
[73] A. Khayyer, H. Gotoh, Modied moving particle semi-implicit methods for the prediction of 2D wave impact pressure, Coastal Engineering 56 (2009)
419440.
[74] C. Hirt, B. Nichols, Volume of uid (VOF) method for the dynamics of free boundaries, Journal of Computational Physics 39 (1981) 201225.
[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) 312324.
[76] S.-Y. Lin, Y.-C. Chen, A pressure correction-volume of uid method for simulations of uidparticle interaction and impact problems, International
Journal of Multiphase Flow 49 (2013) 3148.
[77] G.A. Dilts, Moving-least-squares-particle hydrodynamics I: consistency and stability, International Journal for Numerical Methods in Engineering. 44
(1999) 11151155.
[78] G.A. Dilts, Moving least-squares particle hydrodynamics II: conservation and boundaries, International Journal for Numerical Methods in Engineering.
48 (2000) 15031524.
[79] L. Cueto-Felgueroso, I. Colominas, G. Mosqueira, F. Navarrina, M. Casteleiro, On the Galerkin formulation of the smoothed particle hydrodynamics
method, International Journal for Numerical Methods in Engineering. 60 (2004) 14751512.
[80] N. Grenier, M. Antuono, A. Colagrossi, D. Le Touz, B. Alessandrini, An Hamiltonian interface SPH formulation for multi-uid and free surface ows,
Journal of Computational Physics 228 (2009) 83808393.
[81] J.J. Monaghan, SPH compressible turbulence, Monthly Notices of the Royal Astronomical Society 335 (2002) 843852.
[82] J. Bonet, M.X. Rodrguez-Paz, Hamiltonian formulation of the variable-h SPH equations, Journal of Computational Physics 209 (2005) 541558.
[83] D.J. Price, Smoothed particle hydrodynamics and magnetohydrodynamics, Journal of Computational Physics 231 (2012) 759794.

Journal of Computational Physics: Xiaosong Sun, Mikio Sakai, Yoshinori Yamada

Uploaded by

Copyright:

Available Formats

You might also like

Journal of Computational Physics: Xiaosong Sun, Mikio Sakai, Yoshinori Yamada

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Journal of Computational Physics: Xiaosong Sun, Mikio Sakai, Yoshinori Yamada

Uploaded by

Copyright:

Available Formats

Journal of Computational Physics 248 (2013) 147176

Contents lists available at SciVerse ScienceDirect

Journal of Computational Physics

Three-dimensional simulation of a solidliquid ow

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

2. Physical model and governing equations

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 1. The SPH kernel approximation.

Wr Wjrj and rWr

Fig. 2. Three-dimensional cubic spline kernel on xy plane.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 3. Kernel integral near the boundary.

3.2. A corrected kernel approximation

respectively. Plots of these functions are shown in Fig. 5.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 4. Variation of the correction factor near the boundary.

Fig. 5. Correction factor and its derivative.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 6. Voigt model in DEM.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

where e indicates the coefcient of restitution.

The drag coefcient Cd is given by

And the particle Reynolds number Rek is given by

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Piv isc hr mrv rv T ii

where m is the kinematic viscosity of the uid.

4.2.5. Other special treatments in SPH simulation

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

A widely applied model for

part is proposed by Monaghan [32] as

The factor Rij is determined as

4.3. Algorithm and time-stepping

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

4.4. Treatment on wall boundaries

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

ux minjx xB j for all xB 2 @ X:

Note that the denition by Eq. (63) implies that

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 8. Implementation of distance function.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 9. DF near a curved wall.

Fig. 10. Correction factors near a curved wall.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 12. Boundary round-off.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 13. Typical snapshots of dam break.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

Fig. 14. Position of water front during collapse of water dam.

Fig. 15. Setting up of the solidliquid dam-break test.

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

X. Sun et al. / Journal of Computational Physics 248 (2013) 147176

and a characteristic time

hAix arg min JAx :