Lecture notes on the smoothed particle hydrodynamics (SPH)

G. R. Liu
(In collaboration with Drs. M. B. Liu and Li Zirui)

Abstract ............................................................................................................................... 1
1 Introduction.................................................................................................................. 2
1.1 Traditional grid based numerical methods............................................................ 2
1.2 Meshfree methods................................................................................................. 3
1.3 Smoothed particle hydrodynamics........................................................................ 4
2 SPH approximation techniques.................................................................................... 6
2.1 Kernel approximation of a function...................................................................... 7
2.2 Kernel approximation of derivatives .................................................................... 8
2.3 Particle approximation ........................................................................................ 11
2.4 Techniques for deriving SPH formulations ........................................................ 12
2.5 SPH formulations for Navier-Stokes (N-S) equations........................................ 13
3 SPH smoothing function ............................................................................................ 15
3.1 Review on commonly used smoothing functions ............................................... 15
3.2 Generalizing constructing conditions ................................................................. 19
3.3 Constructing SPH smoothing functions.............................................................. 21
4 Consistency ................................................................................................................ 23
4.1 Consistency in kernel approximation (kernel consistency) ................................ 24
4.2 Consistency in particle approximation (particle consistency) ............................ 25
4.3 Review on approaches for restoring consistency................................................ 26
4.4 An SPH formulation for Discontinuity ............................................................... 29
4.5 A general approach to restore particle inconsistency ......................................... 35
4.6 Finite particle method ......................................................................................... 37
4.7 Consistency vs. stability...................................................................................... 44
5 Special topics ............................................................................................................. 44
5.1 Solid boundary treatment.................................................................................... 44
5.2 Representation of solid grains............................................................................. 49
5.3 Material interface treatment................................................................................ 51
5.4 Tensile instability................................................................................................ 55
6 Applications ............................................................................................................... 58
6.1 High strain hydrodynamics with material strength............................................. 58
6.2 Detonation, explosion and underwater explosion............................................... 64
6.3 Microfluidics and micro drop dynamics ............................................................. 70
6.4 Ocean and coastal hydrodynamics and offshore engineering............................. 75
6.5 Environmental and geophysical flows ................................................................ 77
7 Summary .................................................................................................................... 80
References......................................................................................................................... 81

1
 Lecture notes on the smoothed particle hydrodynamics (SPH)

G. R. Liu
Office: EA-5-05; Tel: 6516-6481; E-mail: [email protected]
In collaboration with
Dr. Li Zirui (E3A-02-06, Tel : 6516-4795 ; email : [email protected])
Centre for Advanced Computations in Engineering Science (ACES), Department of
Mechanical Engineering, National University of Singapore, 10 Kent Ridge Crescent,
Singapore 119260.
3
SMA Fellow, Singapore-MIT Alliance.
And Dr. M. B. Liu
Institute of Mechanics, Chinese Academy of Science, #15, Xisihuanbei Road, Beijing,
China, 100190.

Abstract
This lecture notes is largely from the invited review paper by the Archives of
Computational Methods in Engineering (ARCME). The contents of this reviewer paper
fits very well to the objectives of Part II of this SMA5233 course that is to provide a
general and overall understanding of the method of smoothed particle hydrodynamics
(SPH). Some of the detailed formulations and technicall treatments can be found in the
reference book [6].
SPH is a meshfree particle method based on Lagrangian formulation, and has been
widely applied to different areas in engineering and science. This paper presents an
overview on the SPH method and its recent developments, including 1) the need for
meshfree particle methods, and advantages of SPH, 2) approximation schemes of the
conventional SPH method and numerical techniques for deriving SPH formulations for
partial differential equations such as the Navier-Stokes (N-S) equations, 3) the role of the
smoothing kernel functions and a general approach to construct smoothing kernel
functions, 4) kernel and particle consistency for the SPH method, and approaches for
restoring particle consistency, 5) several important numerical aspects, and 6) some recent
applications of SPH. The paper ends with some concluding remarks.

Keywords: Numerical methods, Meshfree method, Smoothed particle hydrodynamics

(SPH), Particle method, Smoothing function, Particle consistency.

1
1 Introduction

1.1 Traditional grid based numerical methods

Computer simulation has increasingly become a more and more important tool for
solving practical and complicated problems in engineering and science. It plays a
valuable role in providing tests and examinations for theories, offering insights to
complex physics, and assisting in the interpretation and even the discovery of new
phenomena. Grid or mesh based numerical methods such as the finite difference methods
(FDM), finite volume methods (FVM) and the finite element methods (FEM) have been
widely applied to various areas of computational fluid dynamics (CFD) and
computational solid mechanics (CSM). These methods are very useful to solve
differential or partial differential equations (PDEs) that govern the concerned physical
phenomena. For centuries, the FDM has been used as a major tool for solving partial
differential equations defined in problem domains with simple geometries. For decades,
the FVM dominates in solving fluid flow problems and FEM plays an essential role for
solid mechanics problems with complex geometry [1-3]. One notable feature of the grid
based numerical models is to divide a continuum domain into discrete small subdomains,
via a process termed as discretization or meshing. The individual grid points (or nodes)
are connected together in a pre-defined manner by a topological map, which is termed as
a mesh (or grid). The meshing results in elements in FEM, cells in FVM, and grids in
FDM. A mesh or grid system consisting of nodes, and cells or elements must be defined
to provide the relationship between the nodes before the approximation process for the
differential or partial differential equations. Based on a properly pre-defined mesh, the
governing equations can be converted to a set of algebraic equations with nodal
unknowns for the field variables. So far the grid based numerical models have achieved
remarkably, and they are currently the dominant methods in numerical simulations for
solving practical problems in engineering and science [1-5].
Despite the great success, grid based numerical methods suffer from difficulties in
some aspects, which limit their applications in many types of complicated problems. The
major difficulties are resulted from the use of mesh, which should always ensure that the
numerical compatibility condition is the same as the physical compatibility condition for
a continuum. Hence, the use of grid/mesh can lead to various difficulties in dealing with
problems with free surface, deformable boundary, moving interface, and extremely large
deformation and crack propagation. Moreover, for problems with complicated geomatry,
the generation of a quality mesh has become a difficult, time-consuming and costly
process.
In grid based numerical methods, mesh generation for the problem domain is a
prerequisite for the numerical simulations. For Eulerian grid methods such as FDM
constructing a regular grid for irregular or complex geometry has never been an easy task,
and usually requires additional complicated mathematical transformation that can be even
more expensive than solving the problem itself. Determining the precise locations of the
inhomogeneities, free surfaces, deformable boundaries and moving interfaces within the
frame of the fixed Eulerian grid is also a formidable task. The Eulerian methods are also
not well suited to problems that need monitoring the material properties in fixed volumes,

2
e.g. particulate flows [1, 2, 6]. For the Lagrangian grid methods like FEM, mesh
generation is necessary for the solids and structures, and usually occupies a significant
portion of the computational effort. Treatment of extremely large deformation is an
important issue in a Lagrangian grid based method. It usually requires special techniques
like rezoning. Mesh rezoning, however, is tedious and time-consuming, and may
introduce additional inaccuracy into the solution [3, 7].
The difficulties and limitations of the grid based methods are especially evident when
simulating hydrodynamic phenomena such as explosion and high velocity impact (HVI).
In the whole process of an explosion, there exist special features such as large
deformations, large inhomogeneities, moving material interfaces, deformable boundaries,
and free surfaces [8]. These special features pose great challenges to numerical
simulations using the grid based methods. High velocity impact problems involve shock
waves propagating through the colliding or impacting bodies that can behave like fluids.
Analytically, the equations of motion and a high-pressure equation of state are the key
descriptors of material behavior. In HVI phenomena, there also exist large deformations,
moving material interfaces, deformable boundaries, and free surfaces, which are, again,
very difficult for grid based numerical methods to cope with [9].
The grid based numerical methods are also not suitable for situations where the main
concern of the object is a set of discrete physical particles rather than a continuum.
Typical examples include the interaction of stars in astrophysics, movement of millions
of atoms in an equilibrium or non-equilibrium state, dynamic behavior of protein
molecules, and etc. Simulation of such discrete systems using the continuum grid based
methods is often not be a good choice [10, 11].
1.2 Meshfree methods

Over the past years, meshfree methods have been a major research focus, towards the
next generation of more effective computational methods for more complicated problems
[4, 6]. The key idea of the meshfree methods is to provide accurate and stable numerical
solutions for integral equations or PDEs with all kinds of possible boundary conditions
using a set of arbitrarily distributed nodes or particles [4, 6]. The history, development,
theory and applications of the major existing meshfree methods have been addressed in
these monographs and review articles [4, 6, 7, 12-15]. The readers may refer to these
literatures for more details of the meshfree methods. To avoid too much detour from our
central topic, this article will not further discuss these meshfree methods and techniques,
except for mentioning briefly some of the latest advancements.
For solid mechanics problems, instead of weak formulations used in FEM, we now
have a much more powerful weakened weak (W2) formulation for general settings of
FEM and meshfree methods [16-18]. The W2 formulation can create various models
with special properties, such upper bound property [19-22], ultra-accurate and supper
convergent solutions [21, 23-28], and even nearly exact solutions [29, 30]. These W2
formulations have a theoretical foundation on the novel G space theory [16-18]. All most
all these W2 models work very well with triangular mesh, and applied for adaptive
analyses for complicated geometry.

3
 For fluid dynamics problems, the gradient smoothing method (GSM) has been
recently formulated using carefully designed gradient smoothing domains [31-34]. The
GSM works very well with unstructured triangular mesh, and there is can be used
effective for adaptive analysis [33]. The GSM is an excellent alternative to the FVM for
CFD problems.
One distinct meshfree method is the method of smoothed particle hydrodynamics or
SPH. The SPH is a very powerful method for types of CFD problems governed by the
Navier-Stokes equations.
1.3 Smoothed particle hydrodynamics

Smoothed particle hydrodynamics is a “truly” meshfree, particle method originally

used for continuum scale applications, and may be regarded as the oldest modern
meshfree particle method. It was first invented to solve astrophysical problems in three-
dimensional open space [35, 36], since the collective movement of those particles is
similar to the movement of a liquid or gas flow, and it can be modeled by the governing
equations of the classical Newtonian hydrodynamics.
In SPH, the state of a system is represented by a set of particles, which possess
material properties and interact with each other within the range controlled by a weight
function or smoothing function [6, 37, 38]. The discretization of the governing equations
is based on these discrete particles, and a variety of particle-based formulations have been
used to calculate the local density, velocity and acceleration of the fluid. The fluid
pressure is calculated from the density using an equation of state, the particle acceleration
is then calculated from the pressure gradient and the density. For viscous flows, the
effects of physical viscosity on the particle accelerations can also be included. As a
Lagrangian particle method, SPH conserves mass exactly. In SPH, there is no explicit
interface tracking for multiphase flows – the motion of the fluid is represented by the
motion of the particles, and fluid surfaces or fluid-fluid interfaces move with particles
representing their phase defined at the initial stage.
SPH has some special advantages over the traditional grid based numerical methods.
1. SPH is a particle method of Lagrangian nature, and the algorithm is Galilean
invariant. It can obtain the time history of the material particles. The
advection and transport of the system can thus be calculated.
2. By properly deploying particles at specific positions at the initial stage before
the analysis, the free surfaces, material interfaces, and moving boundaries can
all be traced naturally in the process of simulation regarless the complicity of
the movement of the particles, which have been very challenging to many
Eulerian methods. Therefore, SPH is an ideal choice for modeling free surface
and interfacial flow problems.
3. SPH is a particle method without using a grid/mesh. This distinct meshfree
feature of the SPH method allows a straightforward handling of very large
deformations, since the connectivity between particles are generated as part of
the computation and can change with time. Typical examples include the SPH

4
 applications in high energy phenomena such as explosion, underwater
explosion, high velocity impact, and penetrations.
4. In SPH method, a particle represents a finite volume in continuum scale. This
is quite similar to the classic molecular dynamics (MD) method [11, 39] that
uses a particle to represent an atom or a molecule in nano-scale, and the
dissipative particle dynamics (DPD) method [40, 41] that uses a particle to
represent a small cluster of molecules in meso-cale. Thus, it is natural to
generalize or extend SPH to smaller scales, or to couple SPH with molecular
dynamics and dissipative particle dynamics for multiple scale applications,
especially in biophysics, and biochemistry.
5. SPH is suitable for problems where the object under consideration is not a
continuum. This is especially true in bio- and nano- engineering at micro and
nano scale, and astrophysics at astronomic scale. For such problems, SPH can
be a natural choice for numerical simulations.
6. SPH is comparatively easier in numerical implementation, and it is more
natural to develop three-dimensional numerical models than grid based
methods.
The early SPH algorithms were derived from the probability theory, and statistical
mechanics are extensively used for numerical estimation. These algorithms did not
conserve linear and angular momentum. However, they can give reasonably good results
for many astrophysical phenomena. For the simulations of fluid and solid mechanics
problems, there are challenges to reproduce faithfully the partial differential equations
governing the corresponding fluid and solid dynamics. These challenges involve
accuracy and stability of the numerical schemes in implementing the SPH methods.
With the development of the SPH method, and the extensive applications to a wide
range of problems, more attractive features have been showcased while some inherent
drawbacks have also been identified. Different variants or modifications have been
proposed to improve the original SPH method. For example, Gingold and Monaghan
found the non-conservation of linear and angular momentum of the original SPH
algorithm, and then introduced an SPH algorithm that conserves both linear and angular
momentum [42]. Hu and Adams also invented an angular-momentum conservative SPH
algorithm for incompressible viscous flows [43].
Many researchers have conducted investigations on the SPH method on the numerical
aspects in accuracy, stability, convergence and efficiency. Swegle et al. identified the
tensile instability problem that can be important for materials with strength [44]. Morris
noted the particle inconsistency problem that can lead to poor accuracy in the SPH
solution [45]. Over the past years, different modifications or corrections have been tried
to restore the consistency and to improve the accuracy of the SPH method. Monaghan
proposed symmetrization formulations that were reported to have better effects [46-48].
Johnson and his co-workers gave an axis-symmetry normalization formulation so that, for
velocity fields that yield constant values of normal velocity strains, the normal velocity
strains can be exactly reproduced [49, 50]. Randles and Libersky derived a normalization
formulation for the density approximation and a normalization for the divergence of the
stress tensor [51]. Chen et al. proposed a corrective smoothed particle method (CSPM)

5
which improves the simulation accuracy both inside the problem domain and around the
boundary area [52, 53]. The CSPM has been improved by Liu et al. in resolving
problems with discontinuity such as shock waves in a discontinuous SPH (DSPH) [54].
Liu et al. also proposed a finite particle method (FPM), which uses a set of basis function
to approximate field variables at a set of arbitrarily distributed particles [55, 56]. FPM
can be regarded as an improved version of SPH and CSPM with better performance in
particle consistency. Batra et al. concurrently developed a similar idea to FPM, and it is
named modified SPH (MSPH) [57] with applications mainly in solid mechanics. Fang et
al. further improved this idea for simulating free surface flows [58], and they later
developed a regularized Lagrangian finite point method for the simulation of
incompressible viscous flows [59, 60]. A stress point method was invented to improve
the tensile instability and zero energy mode problems [61-64]. Other notable
modifications or corrections of the SPH method include the moving least square particle
hydrodynamics (MLSPH) [65, 66], the integration kernel correction [67], the reproducing
kernel particle method (RKPM) [68, 69], the correction for stable particle method [70,
71], and several other particle consistency restoring approaches [6, 55, 72]. Belytschko
and his co-workers have conducted a series of stability and convergence analyses on
meshfree particle methods, and some of the numerical techniques and analyses can also
be applicable to SPH [13, 71, 73].
This article is organized as follows. In Section 1, the background of meshfree particle
methods is first addressed with highlights on overcoming the limitations of the grid based
numerical models. The invention, features and developments of the SPH method are then
briefly introduced. In Section 2, the approximation schemes of the SPH method are
discussed, while some numerical techniques for developing SPH formulations are
presented. The SPH formulations for the N-S equation, which governs the general fluid
flow problems, are also given. Section 3 presents a review on the smoothing kernel
function. Conditions for constructing smoothing functions are developed with examples
of smoothing functions constructed. Section 4 introduces consistency concept of SPH
including kernel consistency and particle consistency, and also provides an in-depth
review on the existing approaches for restoring particle consistency. Several new
approaches are also presented, which include a discontinuous SPH for simulating
problems with discontinuity, a general approach to restore particle inconsistency, and a
finite particle method. In Section 5, some important numerical topics in SPH are
discussed. These special topics include 1) solid boundary treatment, 2) representation of
solid obstacles, 3) material interface treatment, and 4) tensile instability. Different
applications of the SPH method have been reviewed in Section 6. Some concluding
remarks are given in Section 7.
2 SPH approximation techniques

The conventional SPH method was originally developed for hydrodynamics problems
in which the governing equations are in strong form of partial differential equations of
field variables such as density, velocity, energy, and etc. There are basically two steps in
obtaining an SPH formulation. The first step is to represent a function and/or its
derivatives in continuous form as integral representation, and this step is usually termed
as kernel approximation. In this kernel approximation step, the approximation of a

6
function and its derivatives are based on the evaluation of the smoothing kernel function
and its derivatives. The second step is usually referred to as particle approximation. In
this step, the computational domain is first discretized by representing the domain with a
set of initial distribution of particles representing the initial settings of the problem. After
discretization, field variables on a particle are approximated by a summation of the values
over the nearest neighbor particles.
2.1 Kernel approximation of a function

The kernel approximation in the SPH method involves representation of a function

and its derivatives using a smoothing function. The smoothing function should satisfy
some basic requirements, and it has been called kernel, smoothing kernel, smoothing
kernel function, or sometimes even weight function in some SPH literature [37, 45, 48,
74]. A detailed discussion on smoothing function, basic requirements and constructing
conditions will be given in Section 3.
The kernel approximation of a function f ( x ) used in the SPH method starts from the
following identity
f ( x )   f ( x ) ( x  x )dx  , (1)


where f is a function of the position vector x , and  ( x  x) is the Dirac delta function
given by
1 x  x
 ( x  x )   . (2)
0 x  x
In equation (1),  is the volume of the integral that contains x . Equation (1) implies that
a function can be represented in an integral form. Since the Dirac delta function is used,
the integral representation in equation (2) is exact and rigorous, as long as f ( x ) is defined
and continuous in  .
The Delta function  ( x  x) with only a “point” support, and hence equation (1)
cannot be used for establishing discrete numerical models. If replacing the Delta function
 ( x  x) by a smoothing function W ( x  x , h) with a finite spatial dimension h, the
kernel approximation of f ( x ) ,  f ( x )  , becomes

 f ( x )    f ( x )W ( x  x, h)dx , (3)



where h is the smoothing length defining the influence or support area of the smoothing
function W. Note that as long as W is not the Dirac delta function, the integral
representation shown in equation (3) can only be an approximation, except for special
cases. Therefore equation (3) can be written as
 f ( x )   f ( x )W ( x  x , h)dx  . (4)


7
 A smoothing function W is usually chosen to be an even function for reasons given
later in Section 3. It should also satisfy a number of conditions. The first one is the
normalization condition that states

 W ( x  x , h)dx   1 .

(5)

This condition is also termed as unity condition since the integration of the smoothing
function produces the unity.
The second condition is the Delta function property that is observed when the
smoothing length approaches zero
lim W ( x  x , h)   ( x  x ) . (6)
h 0

The third condition is the compact condition

W ( x  x , h)  0 when x  x   h . (7)
where  is a constant related to the smoothing function for a particle at x, and  h defines
the effective (non-zero) area of the smoothing function. This effective area is usually
called as the support domain of the smoothing function for a point at x (or the support
domain of that point). Using this compact condition, integration over the entire problem
domain is localized as integration over the support domain of the smoothing function.
Therefore, the integration domain  can be the same as the support domain.
In the SPH literatures, the kernel approximation is often said to have h2 accuracy or
second order accuracy [45, 46, 48, 74-76]. The observation can be obtained easily using
Taylor series expansion on equation (4). Note from equation (7) that the support domain
of the smoothing function is x   x  h , the errors in the SPH integral representation can
be roughly estimated by using the Taylor series expansion of f ( x ) around x in equation
(4). If f ( x ) is differentiable, we can get


 f ( x )  [ f ( x )  f ( x )( x   x )  r (( x   x ) 2 )]W ( x  x , h)dx 

, (8)
 
 f ( x ) W ( x  x , h)dx   f ( x ) ( x   x )W ( x  x , h) dx   r ( h 2 )
 

where r stands for the residual. Note that W is an even function with respect to x, and
( x   x )W ( x  x , h) should be an odd function. Hence we should have

 ( x  x)W ( x  x, h)dx  0 .


(9)

Using equations (5) and (9), equation (8) becomes

 f ( x )  f ( x )  r (h 2 ) . (10)
It is clear that SPH kernel approximation of an arbitrary field function is of second order
accuracy.
2.2 Kernel approximation of derivatives

The approximation for the spatial derivative   f ( x ) is obtained simply by

substituting f ( x ) with   f ( x ) in equation (4), which gives

8
    f ( x )   [  f ( x )]W ( x  x , h)dx  , (11)


where the divergence in the integral is operated with respect to the primed coordinate.
Considering
[  f ( x)]W ( x  x, h)    [ f ( x)W ( x  x, h)]  f ( x ) [W ( x  x, h)] , (12)
the following equation is obtained
  f ( x )     [ f ( x )W ( x  x , h)]d x  -  f ( x ) [W ( x  x , h)]d x  . (13)
 

The first integral on the right hand side (RHS) of equation (13) can be converted using
the divergence theorem into an integral over the surface S of the domain of the
integration,  , as

   f ( x )   f ( x)W ( x  x , h)  ndS   f ( x) W ( x  x , h)dx  , (14)
S 

where n is the unit vector normal to the surface S . Since the smoothing function W is
usually defined to have compact support (see equation (7)), the value of W on the surface
of the integral in equation (14) is zero in SPH. Therefore, the surface integral on the right
hand side of equation (14) is also zero. Hence, the kernel approximation of the
derivatives can be written from equation (14) as
   f ( x )    f ( x ) W ( x  x , h)dx  , (15)


It is clear that the differential operation on a function is transformed into a differential

operation on the smoothing function. In other words, the SPH kernel approximation of
the derivative of a field function allows the spatial gradient to be determined from the
values of the function and the derivatives of the smoothing function W, rather than from
the derivatives of the function itself.
Kernel approximation of higher order derivatives can be obtained in a similar way by
substituting f ( x ) with the corresponding derivatives in equation (4), using integration by
parts, divergence theorem and some trivial transformations. Another approach is to
repeatedly use equation (15) to obtain the kernel approximation of the higher order
derivatives, since any higher order derivative can always be regarded as the first order
derivative of its next lower order derivative.
Following similar analyses based on Taylor series expansion, it is easy to show that
the kernel approximation of the derivative is also of second order accuracy. Since the
SPH kernel approximations for a field function and its derivatives are of second order
accuracy, that is why the SPH method has usually been referred as a method of second
order accuracy. However, equation (10) is not always true because equations (5) and (9)
are sometimes not satisfied. For example, in a 1D problem space, if the support domain
is within the problem domain under consideration, the integration of the smoothing
function is unity (equation (5)), and the integration of the first moment of the smoothing
function (see equation (9)) is zero. Also the surface integral in equation (14) is zero.

9
Hence the SPH kernel approximations are of second order accuracy, and this is shown in
Figure 1.
However, there are scenarios in which the support domain intersects with the problem
domain boundary, as shown in Figure 2. Therefore, the smoothing function W is
truncated by the boundary, and the integration of the smoothing function is no longer
unity. The integration of the first moment term of the smoothing function and the surface
integral in equation (14) are also no longer zero. At such scenarios, the SPH kernel
approximations are not of second order accuracy.

Figure 1 Schematic illustration of the scenarios in which the support domain is located
within the problem domain. For such scenarios, the SPH kernel approximations are of
second order accuracy.

Figure 2 Schematic illustration of the scenarios in which the support domain intersects
with the problem domain. For such scenarios, the SPH kernel approximations are not
exactly of second order accuracy.

10
 2.3 Particle approximation

The second step of SPH method is the particle approximation, which involves
representing the problem domain using a set of particles, and then estimating field
variables on this set of particles. Considering a problem domain  filled with a set of
particles (usually arbitrarily distributed, see Figure 3 for illustration in a two-dimensional
domain). These particles can either be centered particles initially generated using
existing mesh generation tools or concentrated particles initially generated using some
kind of space discretization model such as the particle-fill model in AUTODYN [77].
The state of the system is represented by these particles, each associated with field
properties. These particles can be used not only for integration, interpolation or
differencing, but also for representing the material. The volume of a subsection is
lumped on the corresponding particle. Therefore one particle i is associated with a fixed
lumped volume Vi without fixed shape. If the particle mass and density are concerned,
the lumped volume can also be replaced by the corresponding mass to density ratio
mi i . These particles can be fixed in an Eulerian frame or move in a Lagrangian frame.

Figure 3 SPH particle approximations in a two-dimensional problem domain  with a

surface S. W is the smoothing function that is used to approximate the field variables
at particle i using averaged summations over particles j within the support domain
with a cut-off distance of  hi .

After representing the computational domain with a finite number of particles, the
continuous form of kernel approximation expressed in equations (4) can be written in
discretized form of a summation of the neighboring particles as follows
N mj
 f ( x )   f ( x j )W ( x  x j , h) , (16)
j 1 j
where N is the total number of particles within the influence area of the particle at x. It is
the total number of particles that are within the support domain which has a cut-off

11
distance, characterized by the smoothing length, h , multiplied by a scalar constant  .
This procedure of summation over the neighboring particles is referred to as particle
approximation, which states that the value of a function at a particle can be approximated
by using the average of the values of the function at all the particles in the support
domain weighted by the smoothing function. Following the same procedure, the particle
approximation of a derivative can be obtained as
N mj
   f ( x )   f ( x j ) W ( x  x j , h) , (17)
j 1 j
where the gradient W in the above equation is evaluated at particle j. Equation (17)
states that the value of the gradient of a function at a particle located at x can be
approximated by using the summation of those values of the function at all the particles
in the support domain weighted by the gradient of the smoothing function. The particle
approximation in equations (16) and (17) converts the continuous form of kernel
approximation of a field function and its derivatives to the discrete summations over a set
of particles. The use of particle summations to approximate the integral is, in fact, a key
approximation that makes the SPH method simple without using a background mesh for
numerical integration, and it is also the key factor influencing the solution accuracy of the
SPH method.
One important aspect is that the particle approximation in the SPH method introduces
the mass and density of the particle into the equations. This can be conveniently applied
to hydrodynamic problems in which the density is a key field variable. This is probably
one of the major reasons for the SPH method being particularly popular for dynamic fluid
flow problems. If the SPH particle approximation is applied to solid mechanics problems,
special treatments are required. One of the ways is to use the SPH approximation to
create shape functions, and to establish the discrete system equations [4].
The particle approximation is, however, related to some numerical problems inherent
in the SPH method, such as the particle inconsistency and the tensile instability, as will
be addressed in the following sections. One basic reason is that the discrete summation is
only taken over the particles themselves (collocation). In general, in meshfree methods,
to achieve stability and accuracy, the number of sampling points for integration should be
more than the field nodes (particles). This is especially true for meshfree methods based
on weak forms for solid mechanics problems [4]. Otherwise, it may (not always) lead to
some kind of instability problems.

2.4 Techniques for deriving SPH formulations

By using the above-described procedure of kernel approximation and particle
approximation, SPH formulations for partial differential equations can always be derived.
There are in fact a number of ways to derive SPH formulation of PDEs. Benz used one
approach to derive the SPH equations for PDEs that is to multiply each term in the PDEs
with the smoothing function, and integrate over the volume with the use of integration by
parts and Taylor expansions [78]. Monaghan employed a straightforward approach of
directly using equations (16) and (17) [48]. In that approach, the following two identities
are employed to place the density inside the gradient operator,

12
 1
  f ( x)  [  (  f ( x ))  f ( x )  ] , (18)

f ( x) f ( x)
  f ( x )   [  ( )  ] . (19)
 2

The above two identities may be substituted into the integral in equation (11). The same
procedure of the particle approximation to obtain equation (17) is applied to each
gradient term on the right hand side of equations (18) and (19). Note that each expression
at the outside of every gradient term is evaluated at the particle itself, the results from
equations (18) and (19) for the divergence of f ( x ) at particle i are obtained as

1  
N
   f ( xi )  
 m j  f ( x j )  f ( xi )  iWij  ,
i  j 1   
(20)


and
 N  f (x j ) f (x )  

   f ( xi )  i  m j
 j 1
 ( 2 )  ( 2i )  iWij  .
  j i  
(21)
 
One of the good features for the above two equations is that the field function f ( x )
appears pairwisely and involves asymmetric and symmetric SPH formulations. These
asymmetric and symmetric formulations can help to improve the numerical accuracy in
SPH simulations [6, 48, 55].
Besides the above-mentioned two identities, some other rules of operation can be
convenient in deriving the SPH formulations for complex system equations [6]. For
example, for two arbitrary functions of field variables f1 and f2, the following rules exist.
f1  f 2  f1  f 2 . (22)
f1 f 2  f1 f2 . (23)
Hence, an SPH approximation of the sum of functions equals to the sum of the SPH
approximations of the individual function, and an SPH approximation of a product of
functions equals to the product of the SPH approximations of the individual functions.
If f1 is a constant denoted by c, we should have
cf 2  c f 2 . (24)
It is clear that the SPH approximation operator is a linear operator. It is also easy to show
that the SPH approximation operator is commutative, i.e.,
f1  f 2  f 2  f1 , (25)
and
f1 f 2  f 2 f1 . (26)
For convenience, the SPH approximation operator “  ” is omitted in later sections.
2.5 SPH formulations for Navier-Stokes (N-S) equations

Using the afore-mentioned kernel and particle approximation techniques with

necessary numerical tricks, it is possible to derive SPH formulations for partial

13
differential equations governing the physics of fluid flows. For example, for Navier-
Stokes equations controlling the general fluid dynamic problems, we have
 D v 
   
 Dt x
 Dv

1  
  F , (27)
 Dt  x 

 De   v
 
 Dt  x 
where the Greek superscripts  and  are used to denote the coordinate directions, the
summation in the equations is taken over repeated indices, and the total time derivatives
are taken in the moving Lagrangian frame. The scalar density  , and internal energy e ,
the velocity component v , and the total stress tensor   are the dependent variables.
F is the external forces such as gravity. The spatial coordinates x and time t are the
independent variables. The total stress tensor   is made up of two parts, one part of
isotropic pressure p and the other part of viscous stress  , i.e.,     p     . For
Newtonian fluids, the viscous shear stress should be proportional to the shear strain rate
denoted by  through the dynamic viscosity  , i.e.,      , where
v  v 2
     (  v )  .
x x  3
Substituting the SPH approximations for a function and its derivative (as shown in
equations (16) and (17)) to the N-S equations, the SPH equations of motion for the N-S
equations can be written as
 D i N Wij
   m j vij
 Dt j 1 xi
  
 Dvi N
 i  j Wij

Dt
   m j (
 2

 2
)   Fi
x
. (28)
 j 1 i j i
 De 1 N p Wij i  
 i   m j ( p2i  2j )vij  i i
 Dt 2 j 1 i  j xi 2 i

where vij  vi  v j . Equation (28) is a set of commonly used SPH equations for the N-S
equations. It should be noted that by using different numerical tricks, it is possible to get
other different forms of SPH equations for the same partial differential equations. The
obtained SPH formulations may have special features and advantages suitable for
different applictions [6]. One typical example is the approximation of density. If the
field function is the density, equation (16) can be re-written as
N
i   m jWij . (29)
j 1

14
This is another approach to obtain density directly from the SPH summation of the mass
of all particles in the support domain of a given particle, rather than from the continuum
equation. Compared to the SPH formulations on density change in (28), this summation
density approach conserves mass exactly, but suffers from serious boundaries deficiency
due to the particle inconsistency. A frequently used way to remediate the boundaries
deficiency is the following normalization form by the summation of the smoothing
function itself [51, 52]
N

m W
j 1
j ij

i  . (30)
N  mj 
 
j 1   j
 Wij


3 SPH smoothing function

3.1 Review on commonly used smoothing functions

One of the central issues for meshfree methods is how to effectively perform function
approximation based on a set of nodes scattered in an arbitrary manner without using a
predefined mesh or grid that provides the connectivity of the nodes. In the SPH method,
the smoothing function is used for kernel and particle approximations. It is of utmost
importance in the SPH method as it determines the pattern to interpolate, and defines the
cut-off distance of the influencing area of a particle.
Many researchers have investigated the smoothing kernel, hoping to improve the
performance of the SPH method, and/or to generalize the requirements for constructing
the smoothing kernel function. Fulk numerically investigated a number of smoothing
kernel functions in one-dimensional space, and the obtained results are basically valid for
regularly distributed particles [38, 74]. Swegle et al. revealed the tensile instability,
which is closely related to the smoothing kernel function [44]. Morris studied the
performances of several different smoothing functions, and found that by properly
selecting the smoothing function, the accuracy and stability property of the SPH
simulation can be improved [45, 79]. Omang provided investigations on alternative
kernel functions for SPH in cylindrical symmetry [80]. Jin and Ding investigated the
criterions for smoothed particle hydrodynamics kernels in stable field [81]. Capuzzo-
Dolcetta gave a criterion for the choice of the interpolation kernel in SPH [82]. Cabezon
and his co-workers proposed a one-parameter family of interpolating kernels for SPH
studies [83].
Different smoothing functions have been used in the SPH method as shown in
published literatures. Various requirements or properties for the smoothing functions
have been discussed. Major properties or requirements are now summarized and
described in the following discussion.
1. The smoothing function must be normalized over its support domain (Unity)

15
  W ( x  x, h)dx  1 .

(31)

This normalization property ensures that the integral of the smoothing function
over the support domain to be unity. It can be shown in the next section that it
also ensures the zero-th order consistency (C0) of the integral representation of a
continuum function.
2. The smoothing function should be compactly supported (Compact support), i.e.,
W ( x  x )  0 , for x  x    h . (32)
The dimension of the compact support is defined by the smoothing length h and a
scaling factor  , where h is the smoothing length, and  determines the spread of
the specified smoothing function. x  x    h defines the support domain of the
particle at point x. This compact supportness property transforms an SPH
approximation from a global operation to a local operation. This will later lead to
a set of sparse discretized system matrices, and therefore is very important as far
as the computational efforts are concerned.
3. W ( x  x )  0 for any point at x  within the support domain of the particle at point
x (Positivity). This property states that the smoothing function should be non-
negative in the support domain. It is not mathematically necessary as a
convergent condition, but it is important to ensure a physically meaningful (or
stable) representation of some physical phenomena. A few smoothing functions
used in some literatures are negative in parts of the support domain. However in
hydrodynamic simulations, negative value of the smoothing function can have
serious consequences that may result in some unphysical parameters such as
negative density and energy.
4. The smoothing function value for a particle should be monotonically decreasing
with the increase of the distance away from the particle (Decay). This property is
based on the physical consideration in that a nearer particle should have a bigger
influence on the particle under consideration. In other words, with the increase of
the distance of two interacting particles, the interaction force decreases.
5. The smoothing function should satisfy the Dirac delta function condition as the
smoothing length approaches to zero (Delta function property)
lim W ( x  x , h)   ( x  x ) . (33)
h0

This property makes sure that as the smoothing length tends to be zero, the
approximation value approaches the function value, i.e. f ( x )  f ( x ) .
6. The smoothing function should be an even function (Symmetric property). This
means that particles from same distance but different positions should have equal
effect on a given particle. This is not a very rigid condition, and it is sometimes
violated in some meshfree particle methods that provide higher consistency.
7. The smoothing function should be sufficiently smooth (Smoothness). This
property aims to obtain better approximation accuracy. For the approximations of
a function and its derivatives, a smoothing function needs to be sufficiently
continuous to obtain good results. A smoothing function with smoother value of
the function and derivatives would usually yield better results and better
performance in numerical stability. This is because the smoothing function will

16
 not be sensitive to particle disorder, and the errors in approximating the integral
interpolants are small, provided that the particle disorder is not extreme [6, 48, 74].
Any function having the above properties may be employed as SPH smoothing
functions, and many kinds of smoothing functions have been used. Lucy in the original
SPH paper [35] used a bell-shaped function
(1  3R )(1  R )3 R 1
W ( x  x , h)  W ( R, h)   d  , (34)
0 R 1

where  d is 5 4h , 5  h 2 and 105 16 h3 in one-, two- and three-dimensional space,

respectively, so that the condition of unity can be satisfied for all the three dimensions.
r x  x
R is the relative distance between two points (particles) at points x and x  , R   ,
h h
where r is the distance between the two points.
Gingold and Monaghan in their original paper [36] selected the following Gaussian
kernel to simulate the non-spherical stars
2
W ( R, h )   d e  R , (35)
3
where  d is 1  1 2 h , 1  h2 and 1  3 2 h , respectively, in one-, two- and three-dimensional
space, for the unity requirement. The Gaussian kernel is sufficiently smooth even for
high orders of derivatives, and is regarded as a “golden” selection since it is very stable
and accurate especially for disordered particles. It is, however, not really compact, as it
never goes to zero theoretically, unless R approaches to infinity. Because it approaches
zero numerically very fast, it is practically compact. Note that it is computationally more
expensive since it can take a longer distance for the kernel to approach zero. This can
result in a large support domain with more particles for particle approximations.
The most frequently used smoothing function may be the cubic B-spline function,
which was originally used by Monaghan and Lattanzio [84]
 2  R 2  1 R3 0  R 1
3 2
 1
W ( R, h)   d   6 (2  R )3 1 R  2. (36)

0 R2

In one-, two- and three-dimensional space,  d  1 h , 15 7 h2 and 3 2 h3 , respectively.
The cubic spline function has been the most widely used smoothing function in the
emerged SPH literatures since it closely resembles a Gaussian function while having a
narrower compact support. However, the second derivative of the cubic spline is a
piecewise linear function, and accordingly, the stability properties can be inferior to those
of smoother kernels.
Morris have introduced higher order (quartic and quintic) splines that are more
closely approximating the Gaussian and more stable [45, 79]. The quartic spline is

17
 ( R  2.5) 4  5( R  1.5) 4  10( R  0.5)4 0  R  0.5

(2.5  R ) 4  5(1.5  R ) 4 0.5  R  1.5
W ( R, h)   d   , (37)
(2.5  R ) 4 1.5  R  2.5

0 R  2.5

where  d is 1 24h in one-dimensional space. The quintic spline is

(3  R)5  6(2  R )5  15(1  R)5 0  R 1

(3  R)5  6(2  R )5 1 R  2
W ( R, h)   d   , (38)
(3  R)5 2 R3

0 R3

where  d is 120 h , 7 478 h 2 and 3 359 h3 in one-, two- and three-dimensional space,
respectively.
Johnson et al. used the following quadratic smoothing function to simulate high
velocity impact problems [85]
3 2 3 3
W ( R, h)   d ( R  R ) 0 R 2, (39)
16 4 4
where in one-, two- and three-dimensional space,  d  1 h , 2  h 2 and 5 4 h3 ,
respectively. Unlike other smoothing functions, the derivative of this quadratic
smoothing function always increases as the particles move closer, and always decreases
as they move apart. This was regarded by the authors as an important improvement over
the cubic spline function, and it was reported to relieve the problem of compressive
instability.
Some higher order smoothing functions that are devised from lower order forms have
been constructed, such as the super-Gaussian kernel [84]
3 2
W ( R, h )   d (  R 2 )e  R 0  R  2, (40)
2
where  d is 1  in one-dimensional space. One disadvantage of the high order
smoothing function is that the kernel is negative in some region of its support domain.
This may lead to unphysical results for hydrodynamic problems [74].
The smoothing function has been studied mathematically in detail by Liu and his co-
workers. They proposed a systematical way to construct a smoothing function that may
meet different needs [37]. A new quartic smoothing function has been constructed to
demonstrate the effectiveness of the approach for constructing a smoothing function as
follows.
 2 9 2 19 3 5 4
(  R  R  R ) 0R2
W ( R, h)   d  3 8 24 32 . (41)
0 R2


where  d is 1 h , 15 7 h2 and 315 208 h3 in one-, two- and three-dimensional space,

respectively. Note that the centre peak value of this quartic smoothing function is
defined as 2 3 . The quartic smoothing function behaves very much like the widely used

18
cubic B-spline function given in Equation (36), but has only one piece, and hence is
much more convenient and efficient to use. More discussions on this quartic smoothing
function will give in the next section.
3.2 Generalizing constructing conditions

Major requirements of an SPH smoothing function have been addressed in Section 3.1.
Some of these requirements can be derived by conducting Taylor series analysis. This
analysis is carried out at the stage of the SPH kernel approximation for a function and its
derivatives. It shows that, to exactly approximate a function and its derivatives, certain
conditions need to be satisfied. These conditions can then be used to construct the
smoothing functions.
Considering the SPH kernel approximation for a field function f ( x ) as show in
equation (4), if f ( x ) is sufficiently smooth, applying Taylor series expansion of f ( x ) in
the vicinity of x yields
1
f ( x )  f ( x )  f ( x )( x   x )  f ( x )( x   x ) 2  
2
, (42)
(1) k h k f ( k ) ( x ) x  x  k x  x
n
 
k 0
k!
(
h
)  rn (
h
)

where rn is the remainder of the Taylor series expansion. Substituting equation (42) into
equation (4) leads to

x  x
n
f ( x)  A
k 0
k f ( k ) ( x )  rn (
h
), (43)

where

(1) k h k x  x k
Ak 
k!

( h
) W ( x  x , h)dx  . (44)

Comparing the LHS with the RHS of equation (43), in order for f ( x ) to be
approximated to n-th order, the coefficients Ak must equal to the counterparts for f ( k ) ( x )
at the LHS of equation (43). Therefore, after trivil transformation, the following
conditions for the smoothing function W can be obtained as follows

19
 

M 0  W ( x  x , h)dx   1

 
M 1  ( x  x )W ( x  x , h)dx   0 

 


M 2  ( x  x ) W ( x  x , h)dx   0  ,
2
(45)
 

 



M n  ( x  x ) n W ( x  x , h)dx   0 



where M k is the the k-th moments of the smoothing function. Note that the first equation
in (45) is, in fact, the unity condition expressed in equation (31), and the second equation
in (45) stands for the symmetric property. Satisfaction of these two conditions ensures
the first order consistency for the SPH kernel approximation for a function.
Also performing Taylor series analysis for the SPH kernel approximation of the
derivatives of a field function f ( x ) , using the concept of integration by parts, and
divergence theorem with some trivial transformation, the following equations
W ( x  x , h) |S  0 (46)
and


M 0  W ( x  x , h)dx   0

 
M 1  ( x  x )W ( x  x , h)dx   1 
 


2
M 2  ( x  x ) W ( x  x , h)dx   0  (47)
 

 


    
M n  ( x  x ) W ( x  x , h ) dx  0 
n

 


can be obtained. Equation (46) actually specifies that the smoothing function vanishes on
the surface of the support domain. This is compatible to the compactness condition of the
smoothing function. Equation (47) defines the conditions with which the derivatives of
the smoothing function should be satisfied. Note that Equations (45) and (47) are
actually compatible considering integration by parts, divergence theorem and the
boundary value vanishing effects (equation (46)) of the smoothing function.
Performing Taylor series analysis on the SPH kernel approximation for the second
derivatives, similar equations can be obtained. Except for the requirements on the second
derivatives of the momentums, the first derivative of the smoothing function also needs to
vanish on the surface, which is
W ( x  x , h) |S  0 . (48)

20
 Equations (45)-(48) can be used to construct smoothing functions. It can be seen that
the conditions of smoothing functions can be classified into two groups. The first group
shows the ability of a smoothing function to reproduce polynomials. Satisfying the first
group, the function can be approximated to n-th order accuracy. The second group
defines the surface values of a smoothing function as well as its first derivatives, and is
the requirements of the property of compact support for the smoothing function and its
first derivative. Satisfying these conditions, the first two derivatives of the function can
be exactly approximated to the n-th order.
3.3 Constructing SPH smoothing functions

By using above-mentioned conditions, it is possible to have a systematic way to

construct the SPH smoothing functions. If the smoothing function is assumed to be a
polynomial dependent only on the relative distance of the concerned points, it can be
assumed to have the following form in the support domain with an influence width of  h .
W ( x  x , h)  W ( R )  a0  a1 R  a2 R 2  ...  an R n . (49)
It is clear that a smoothing function in the above-mentioned form is a distance
function since it depends on the relative distance. It is easy to show that for the second
derivative of the smoothing function to exist, a1 should vanish. Substituting this
polynomial form smoothing function into the conditions (equations (45)-(48)), the
parameters a0, a2, …, an can be calculated from the resultant linear equations, and then
the smoothing function can be determined.
There are several issues that need further consideration. Firstly, a smoothing function
derived from this set of conditions (see equations (45)) will not necessarily be positive in
the entire support domain, especially when high order reproducibility is required. Such a
negative smoothing function may result in unphysical solutions, for example, negative
density (mass) and negative energy. For this reason, smoothing functions used in SPH
literatures are generally non-negative for CFD problems. On the other hand, for the even
moments ( k  2, 4, 6... ) to be zero, a smoothing function has to be negative in some parts of
the region. This implies that one cannot have both non-negativity and high-order
reproducibility at the same time.
Secondly, in constructing a smoothing function, the center peak value is a factor that
needs to be considered. The center peak value of a smoothing function is very important
since it determines how much the particle itself will contribute to the approximation.
Revisiting equation (45), if a positive smoothing function is used, the highest order of
accuracy for the function approximation is second order. Therefore, the second
momentum ( M 2   ( x  x )2 W ( x  x , h)dx  ) can be used as a rough indicator to measure the

accuracy of the kernel approximation. The smaller the second moment M 2 is, the more
accurate the kernel approximation is. The center peak value of a smoothing function is
closely related to M 2 . A positive smoothing function with a large center peak value will
have a smaller second moment M 2 . This implies that a smoothing function is closer to
the Delta function, and therefore is more accurate in terms of kernel approximations.

21
 Thirdly, in some circumstances, a piecewise smoothing function is preferable since
the shape of the piecewise smoothing function is easier to be controlled by changing the
number of the pieces and the locations of the connection points. For example, consider
the general form of a smoothing function with two pieces,
W1 ( R) 0  R  R1

W ( R )  W2 ( R) R1  R  R2 . (50)
0 R2  R

The function itself and the first two derivatives at the connection points should be
continuous, i.e., W1 ( R1 )  W2 ( R1 ) , W1( R1 )  W2 ( R1 ) and W1( R1 )  W2( R1 ) . Considering the
requirements at these points as well as the compact support property, one possible form
of the smoothing function is
b1 ( R1  R) n  b2 ( R2  R ) n 0  R  R1

W ( R )   d b2 ( R2  R ) n R1  R  R2 . (51)
0 R2  R

It is also feasible to construct smoothing function with more pieces using similar
expressions.
To show the effectiveness of this approach to constructing general SPH smoothing
functions Liu et al. [37] derived a new quartic smoothing function using the following
conditions
 the unity condition,
 compact support of the smoothing function,
 compact support of the first derivative of the smoothing function,
 centre peak value.
2 9 19 3 5 4
The constructed smoothing function is given as W ( R, h)   d (  R 2  R  R ) , for
3 8 24 32
0R2 ,where  d is 1 h , 15 7 h2 and 315 208 h3 in one-, two- and three-dimensional
space, respectively. Note that the centre peak value of this quartic smoothing function is
defined as 2 3 .
As defined, this quartic function satisfies the normalization condition, while the
function itself and its first derivative have compact support. It is very close to the most
commonly used cubic spline (equation (36)) with a same center peak value of 2 3 , and
monotonically decreases with the increase of the distance as show in Figure 4. However,
this quartic function produces a smaller second momentum than the cubic spline function,
and therefore can produce better accuracy for kernel approximation. Also this quartic
smoothing function has a smoother second derivative than the cubic spline smoothing
function, thus the stability properties should be superior to those of the cubic spline
function, as reported by many researchers that a smoother second derivative can lead to
less instability in SPH simulation [79, 86].

22
 1.5

0.5

0 W'

Functions   d
W

-0.5

-1

Cubic spline kernel

-1.5 W'' The new quartic kernel

-2

-2.5
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2
R

Figure 4 The quartic smoothing function constructed by Liu et al. by using the
smoothing function constructing conditions [37]. The shapes of the quartic function and
its first derivative are very close to the shapes of the cubic spline function and its first
derivative. However, this one piece smoothing function is expected to produce better
accuracy as it has smaller second momentum. It is also expected to be more stable since
it has a continuous second derivative.

4 Consistency

In early studies, the SPH method has usually been reported to have second order
accuracy. This is reasonable, as the kernel approximation of a field function has a
remainder of second order (see equation (8)). However, the kernel approximation of a
function and/or its derivatives may not necessarily be of second order accuracy as the
smoothing function may not be an even function since it is truncated when approaching
the boundary region. Moreover, high order accuracy of the kernel approximation does
not necessarily mean high order accuracy of the SPH simulations, as it is the particle
approximation rather than the kernel approximation that eventually determines the
accuracy of the SPH simulations.
Following the Lax-Richtmyer equivalence theorem, we know that if a numerical
model is stable, the convergence of the solution to a well-posed problem will then be
determined by the consistency of the function approximation. Therefore the consistency
of SPH approximation is crucial. In finite element method, to ensure the accuracy and
convergence of an FEM approximation, the FEM shape function must satisfy a certain
degree of consistency. The degree of consistency can be characterized by the order of the
polynomial that can be exactly reproduced by the approximation using the shape function.
In general, if an approximation can reproduce a polynomial up to k-th order exactly, the
approximation is said to have k-th order consistency or Ck consistency. The concept of
consistency from traditional finite element methods can also be used for meshfree particle
methods such as smoothed particle hydrodynamics [3, 6, 13, 71, 73, 87]. Therefore it is

23
similarly feasible to investigate the consistency of SPH kernel and particle
approximations in reproducing a polynomial by using the smoothing function.
4.1 Consistency in kernel approximation (kernel consistency)

For a constant (0th order polynomial) function f ( x )  c (where c is a constant) to be

exactly reproduced by the SPH kernel approximation, following equation (1), we require
f ( x )   cW ( x  x , h)dx   c , (52)
or

 W ( x  x, h)dx  1 . (53)

Equation (53) is exactly the normalization condition described previously.
Further, for a linear function f ( x )  c0  c1 x (where c0 and c1 are constants) to be
exactly reproduced, we must have
f ( x )   (c0  c1 x)W ( x  x , h)dx   c0  c1 x . (54)
Using equation (52), equation (54) can be simplified as

 xW ( x  x, h)dx = x . (55)

Multiplying x to both side of equation (53), we have the following identity

 xW ( x  x, h)dx = x . (56)

Subtracting equation (55) from the above identity yields

 ( x  x)W ( x  x, h)dx  0 . (57)

Equation (57) is the symmetric condition described in Section 3.1.
More generally, by performing Taylor series analyses on the kernel approximation of
a function f ( x ) (   f ( x)W ( x  x , h)dx  ) in a one-dimensional space, we have already
obtained a set of requirements of the smoothing function in Section 3.2, as expressed in
equation (45). Equations (53) (normalization condition) and (57) (symmetric condition)
are actually components in equation (45), which describes the 0th and 1st moments.
For the integrations expressed in equations (45), the integration domain is assumed to
be a full continuous support domain that is not truncated by the boundaries. Equation (45)
states the requirements on the moments of a smoothing function to reproduce certain
order of polynomials.
Equation (45) can thus be used as an approximation accuracy indicator. If a
smoothing function satisfies equations (45), a function can be approximated to n-th order
accuracy. Furthermore the 0th moment in equations (45) states the normalization
condition, and the 1st moment states the symmetry property of the smoothing function.

24
 Similar to the consistency concept in the traditional FEM, if an SPH approximation
can reproduce a polynomial of up to nth order exactly, the SPH approximation is said to
have nth order or Cn consistency. If the consistency of an SPH kernel approximation in
continuous form is termed as kernel consistency, the kernel consistency of an SPH kernel
approximation is of nth order when the smoothing function satisfies equation (45).
Therefore the expressions in equations (45) are also the kernel consistency conditions of
the smoothing function for an SPH kernel approximation.
Note that if the SPH kernel approximations are carried out for regions truncated by
boundaries, constant and linear functions can not be reproduced exactly since equations
(53) and (57) are not satisfied for these regions. Therefore we can conclude that, since a
conventional smoothing function satisfies the normalization and symmetric conditions,
the conventional SPH method has up to C1 consistency for the interior regions. However,
for the boundary regions, it even does not have C0 kernel consistency.
4.2 Consistency in particle approximation (particle consistency)

Satisfying the consistency conditions at the kernel approximation stage does not
necessarily mean that the discretized SPH model will have such a consistency. This is
because such a consistency can be distorted by the particle approximation process in
disrete SPH model. Therefore, the consistency analysis should be conducted for the
discrete SPH model in the particle approximation process, and this consistency can be
termed as particle consistency.
The discrete counterparts of the constant and linear consistency conditions as
expressed in equations (53) and (57) are
N

W ( x  x , h)v
j 1
j j  1, (58)

and
N

 ( x  x )W ( x  x , h)v
j 1
j j j 0. (59)

These discretized consistency conditions are not satisfied in general. One obvious
and simple example is the particle approximations at the boundary particles (Figure 5a).
Even for uniform particle distribution, due to the unbalanced particles contributing to the
discretized summation, the LHS of equation (58) is smaller than 1 and the LHS of
equation (59) will not vanish, due to the truncation of the smoothing function by the
boundary. For cases with irregularly distributed particles (Figure 5b), it is also easy to
verify that even for the interior particles whose support domains are not truncated, the
constant and linear consistency conditions in discretized forms may not be exactly
satisfied. Therefore the original SPH method does not even have C0 consistency in the
particle approximation. It is clear that the inconsistency caused by the particle
approximation is closely related to the corresponding kernel approximation and particles
involved in the approximation. Such an inconsistency problem results in directly the
solution inaccuracy in the original SPH method.

25
 Figure 5 SPH particle approximations in one-dimensional cases. (a) Particle
approximation for a particle whose support domain is truncated by the boundary. (b)
Particle approximation for a particle with irregular particle distribution in its support
domain.

Besides the particle approximation features associated with boundary particles or

irregular distributed particles, the choice of the smoothing length is also important in the
particle approximation process. In a one-dimensional domain with the cubic spline
smoothing function, it is easy to verify that for uniformly distributed interior particles, the
original SPH method has C0 particle consistency if the smoothing length is taken exactly
as the particle spacing ( h  x ) since equation (58) is satisfied. However, varying the
smoothing length can results in a dissatisfaction of equation (58), leading to poor
accuracy in the original SPH method. This is a reason why we often need to examine the
influence of the smoothing length on the SPH approximation results.
In summary, the original SPH models, in general, do not have even C0 consistency.
Such an inconsistency originates from the discrepancy between the SPH kernel and
particle approximations. Boundary particles, irregular distributed particles, and variable
smoothing length can usually produce inconsistency in the particle approximation process.
In the next section, we discuss ways to restore the consistency in SPH models.
4.3 Review on approaches for restoring consistency

It has been shown that the original SPH method even does not have 0th particle
consistency. Different approaches have been proposed to improve the particle
inconsistency and hence the SPH approximation accuracy. Some of them involve
reconstruction of a new smoothing function so as to satisfy the discretized consistency
conditions. However, these approaches are usually not preferred for hydrodynamic
simulations because the reconstructed smoothing function can be partially negative, non-
symmetric, and not monotonically decreasing. Approaches which improve the particle
consistency without changing the conventional smoothing function are usually more
preferable in simulating hydrodynamics.
One early approach [48, 51] is based on the anti-symmetric assumption of the
derivative of a smoothing function

26
 N

W  v
j 1
i, j  0, (60)


where Wi ,  Wi ( x ) x , in which  is the dimension index repeated from 1 to d (d is
the number of dimensions). Therefore when approximating the derivative of a function f,
the particle approximation can be rewritten as

fi ,    f j  fi Wi , v j ,
N
(61)
j 1

or

fi ,    f j  fi Wi , v j .
N
(62)
j 1

It should also be noted that equation (60) is not necessarily valid, even if its
corresponding continuous counterpart  Wi , dx  0 is valid (for interior regions). This is
also a manifestation of the particle inconsistency. Therefore equations (61) and (62)
actually use the particle inconsistency in approximating the derivative of the smoothing
function to offset or balance the particle inconsistency in approximating the derivatives of
a field function, with a hope to improve the accuracy of the approximations.
Randles and Libersky [51] derived a normalization formulation for the density
approximation
N

  W v
j 1
j ij j

i  N
, (63)
Wij v j
j 1

and a normalization for the divergence of the stress tensor 

N

 (
j 1
j   i )  iWij v j
(   )i  N
, (64)
(x
j 1
j  xi )  iWij v j

where  is the tensor product. Again, equations (63) and (64) also use the inconsistency
in approximating the smoothing function and its derivatives to offset the inconsistency in
approximating a field function and its derivatives, also with an aim to improve the
accuracy of the approximations.
Based on Taylor series expansion on the SPH approximation of a function, Chen et al
[53] suggested a corrective smoothed particle method (CSPM). In one-dimensional space,
the process of CSPM can be briefed as follows.
Performing Taylor series expansion at a nearby point xi , a sufficiently smooth
function f ( x) can be expressed as

27
 ( x  xi ) 2
f ( x )  fi  ( x  xi ) fi , x  f i , xx  ... . (65)
2!
Multiplying both sides of the equation (65) by the smoothing function W and integrating
over the entire computational domain yield

 f ( x)W ( x)dx
i

fi , xx . (66)
 fi  Wi ( x)dx  fi , x  ( x  xi )Wi ( x)dx   ( x  xi ) Wi ( x)dx  ...
2

2
If the terms involving derivatives in this equation are neglected, a corrective kernel
approximation for function f(x) at particle i is obtained as

fi 
 f ( x)W ( x)dx . i
(67)
 W ( x)dx i

For a conventional smoothing function (non-negative and symmetric), the second term at
the RHS of equation (66) is zero for interior region and not zero for boundary region.
Therefore the corrective kernel approximation expressed in equation (67) is also of 2nd
order accuracy for interior region and 1st order accuracy for boundary region. Comparing
equation (67) with equation (16), it is found that for the interior regions, the kernel
approximations in the original SPH and CSPM are actually the same due to the
satisfaction of the normalization condition (in continuous form). For the boundary
regions, since the integral of the smoothing function is truncated by the boundary, the
normalization condition cannot be satisfied. By retaining the non-unity denominator,
CSPM restores the C0 kernel consistency.
The corresponding particle approximation for function f(x) at particle i can be
obtained using summation over nearest particles for each term in equation (66) and again
neglecting the terms related to derivatives
N

f
j 1
j Wij v j
fi  N
. (68)
W v
j 1
ij j

It is noted that the particle approximation of the second term at the RHS of equation (66)
is not necessarily zero even for the interior particles due to the irregularity of the particles.
Therefore strictly speaking, the particle approximation expressed in equation (68) is of 1st
order accuracy for both the interior and boundary particles. Only if the particles are
uniformly distributed can the particle approximation of the second term at the RHS of
equation (66) be zero. In this case, the particle approximation expressed in equation (68)
is of 2nd order accuracy for the uniformly distributed interior particles.
If replacing Wi ( x) in equation (66) with Wi , x and neglecting the second and higher
derivatives, a corrective kernel approximation for the first derivative is generated as

28
 fi , x 
  f ( x)  f ( x )W ( x)dx .
i i,x
(69)
 ( x  x )W ( x)dx i i,x

The particle approximations corresponding to equations (69) is

N

( fj 1
j  f i )Wi , x v j
fi , x  N
. (70)
(x
j 1
j  xi )Wi , x v j

Similarly, The CSPM kernel approximations for the derivatives are also of second
order accuracy (or 1st order consistency) for interior regions, but 1st order accuracy (or 0th
order consistency) for boundary regions. Except for cases with uniformly distributed
interior particles, the CSPM particle approximations for the derivatives are of 1st order
accuracy (or 0th order consistency) for both the interior and boundary particles.
4.4 An SPH formulation for Discontinuity

It is clear that CSPM can have better accuracy than the conventional SPH method as
it improves the boundary deficiency problem. It was reported that the CSPM can also
reduce the so-called tensile instability inherent in the traditional SPH method [88, 89].
One notable point in CSPM is that it is based on Taylor series analysis on the kernel
approximation of a field function and its derivatives. To perform Taylor series analysis,
the function under consideration should be sufficiently smooth. Therefore, CSPM is not
applicable to problems with discontinuities such as hydrodynamic problems that generate
shock waves.
Liu and his co-workers proposed a further improvement on SPH in resolving
discontinuity problems, and they call this as a discontinuous SPH, or DSPH [54]. The
idea of DSPH one-dimensional space is briefed here as follows.
Examine the kernel approximation for any function f ( x ) in the support domain of xi .
The support domain is bounded by a and b with a dimension of 2 h , shown in Figure 6.
Assuming that function f ( x ) has an integrable discontinuity at d in the support domain
and that it is located in the right half of the support domain, i.e., xi  d  b , the integration
of the multiplication of f ( x ) and the smoothing function W over the entire support
domain can be divided into two parts
b d b
 a
f ( x )Wi ( x )dx   a
f ( x )Wi ( x )dx   d
f ( x )Wi ( x )dx . (71)

29
 Figure 6 Kernel approximations for a one-dimensional function with a discontinuity at
point d. The support domain of xi is bounded by a and b with a dimension of 2  h .

Expanding f ( x ) in the first integral on the right hand side around point xi , and around
another arbitrary point xk in the second integral, where d  xk  b gives
b
 a
f ( x )Wi ( x )dx
d b
 f ( xi )  a
Wi ( x )dx  f ( xk )  d
Wi ( x )dx
. (72)
d b
 f ( xi )  a
( x  xi )Wi ( x )dx  f ( xk )  d
( x  xk )Wi ( x )dx

 r (h 2 )
Rearranging by combining some similar terms with some transformations yields
b
 a
f ( x )Wi ( x )dx
b b
 f ( xi )  a
Wi ( x )dx   f ( xk )  f ( xi )   d
Wi ( x )dx
b
. (73)
 f ( xi )  a
( x  xi )Wi ( x )dx
b
  d
( x  xk ) f ( xk )  ( x  xi ) f ( xi )Wi ( x )dx  r (h2 )
Since the kernel W is assumed to be even, normalized, and has a compact support, we
have x - xk   h . In the above equation, we assume that f ( x ) must exist and be bounded
in [a, d )  (d , b] . In other words, the derivative of f(x) exists, and when x approaches d
from both sides, it is bounded within a finite limit. Hence, the last two terms (excluding
the residual term) on the RHS of the above equation can be bounded by terms of order of
h respectively. Therefore, the above equation can be rewritten as:
b b b
 a
f ( x )Wi ( x )dx  f ( xi )  a
Wi ( x )dx   f ( xk )  f ( xi )   d
Wi ( x )dx  r (h) . (74)

This equation can be re-written as

b  b 
 f ( x )Wi ( x )dx 
  f ( xk )  f ( xi )  d Wi ( x )dx 
  r ( h) ,
a
f (x ) 
i b
 b (75)
 a
Wi ( x )dx 
 a Wi ( x )dx 


30
which is the kernel approximation of the field function with a discontinuity.
Similarly, the kernel approximation of the derivative is obtained as
b b

f ( x ) 
 a
 f ( x )  f ( xi ) Wix ( x )dx
{
 f ( xk )  f ( xi )   d
Wix ( x )dx
i b b
 a
( x  xi )Wix ( x )dx  a
( x  xi )Wix ( x )dx
b
. (76)


 d
[( x  xk ) f ( xk )  ( x  xi ) f ( xi )]Wix ( x )dx
}  r (h)
b
 a
( x  xi )Wix ( x )dx

The above two equations are the kernel approximations of a field function (and its
derivative) with a discontinuity in the support domain. The kernel approximations at the
RHS of equations (75) and (76) consist of two parts. The first parts are the same as those
at the RHS of equations (67) and equations (69), which describe the kernel
approximations of a field function and its derivatives. It is the second parts in the big
brackets that describe the behavior of the discontinuity. Neglecting the second parts of
equations (75) and (76) should result in numerical errors since the resultant kernel
approximations in the presence of a discontinuity are inconsistent. If the second terms are
retained, the resultant kernel approximations are consistent up to the first order.
The particle approximations for the field function and its derivatives at particle i are a
little bit different from the particle approximations in the conventional SPH and CSPM
due to the presence of the discontinuity. When the domain is discretized by particles,
since two particles cannot be located at the same position, the discontinuity should
always be located between two particles. In deriving equations (75) and (76), since the
point xk is arbitrarily selected, it can be taken as the particle that is nearest to and on the
right hand side of the discontinuity (Figure 7).

Figure 7 Particle approximations for a function with a discontinuity at point d. In the

process from the kernel approximation to particle approximation, an arbitrary point xk is
associated with a particle k that is the nearest particle on the right hand side of the
discontinuity. The total number of particles in the support domain of [a, b] is N.

The particle approximation of the discontinuous function is

31
 N  mj   N   
   
mj
 f jWij   f ( xk )  f ( xi )   Wij 
     
j 1 j  j k  j  
fi 
N 
 . (77)
mj  
N  mj  
 
 
j 1  j 
 Wij

   W
   ij
j 1  j 



And similarly, the particle approximation of the derivative of the discontinuous function
is
N  mj   N 
mj 
 
 j 
j 1   

( f ( x j )  f ( xi ))iWij   f ( xk )  f ( xi )  

 W
  i ij
j k  j 
fi  
N 
mj  
N 
mj 
 
 
j 1  j 
( x j  xi )iWij 

 
  j
j 1  j 
( x  xi )iWij
. (78)
N 
mj  
 
   j
j k  j 
 ( x  xk ) f ( xk )  ( x j  xi ) f ( xi )  iWij 

 
N 
mj  
  
 j  j
j 1  
( x  xi ) W
i ij 

Also, the particle approximations consist of two parts, the first primary parts similar to
those in the CSPM approximation and the second additional parts in the big brackets
developed for the discontinuity treatment. It should be noted that the summations of the
numerators of the additional parts are only carried out for particles in the right portion of
the support domain [d, b] shaded in Figure 7. The coordinates of these particles are
x j ( j  k , k  1, , N ) that satisfy ( xk  x j )( xk  xi )  0 . Particles at the same side as i do not
contribute to the summation in the additional parts. Otherwise, the additional part would
become zero, and the modification term would disappear.
Equations (75) and (76), (77) and (78) are the essential DSPH formulations for
approximating a discontinuous function and its first derivative. It is possible to extend
the DSPH to higher order derivatives, and multi-dimensional space.
The different performances of the SPH, CSPM and DSPH in resolving discontinuity
have been demonstrated by a number of numerical tests. One of them is the sod tube
problem, which has been studied by many researchers for validation purposes [46, 90].
The shock-tube is a long straight tube filled with gas, which is separated by a membrane
into two parts of different pressures and densities. The gas in each part is initially in an
equilibrium state of constant pressure, density and temperature. When the membrane is
taken away suddenly, a shock wave, a rarefaction wave and a contact discontinuity will
be produced. The shock wave moves into the region with lower density gas; the
rarefaction wave travels into the region with higher density gas; while the contact
discontinuity forms near the center and travels into the low-density region behind the
shock. Exact solution is available for comparison for this one-dimensional problem.
In this example, the cubic spline function is used as the smoothing function to
simulate this shock tube problem. The initial conditions are same as those introduced by
Sod [91]

32
 x  0   1 v  0 e  2.5 p  1 x  0.001875
x  0   0.25 v  0 e  1.795 p  0.1795 x  0.0075
where  , p e , and v are the density, pressure, internal energy, and velocity of the gas,
respectively. x is the particle spacing. A total of 400 particles are deployed in the one-
dimensional problem domain. All particles have the same mass of mi =0.001875. 320
particles are evenly distributed in the high-density region [  0.6, 0.0], and 80 particles are
evenly distributed in the low-density region [0, 0.6]. The initial particle distribution is to
obtain the required discontinuous density profile along the tube. The equation of state for
the ideal gas p  (  1)  e is employed in the simulation with   1.4 .
In the simulation, the time step is 0.005 s and the simulation is carried out for 40 steps.
In resolving the shock, the Monaghan type artificial viscosity is used, which also serves
to prevent unphysical penetration [48]. Figure 8, Figure 9, Figure 10 and Figure 11
show, respectively, the density, pressure, velocity and internal energy profiles.
It can be seen that the obtained results from the DSPH agree well with the exact
solution in the region [  0.4, 0.4]. The shock is observed at around x = 0.3 within several
smoothing lengths. The rarefaction wave is located between x =  0.3 and x = 0. The
contact discontinuity is between x = 0.1 and x = 0.2. Just as reported by Hernquist and
Katz [90] and Monaghan [46], the traditional SPH with artificial viscosity yields
comparable results in capturing shock physics. The DSPH gives slightly better results
than the traditional SPH method though only 40 steps passed. If more dynamic steps are
involved, the better performance of the DSPH is more evident [54].
CSPM failed to captures the major shock physics. In this example, CSPM performs
poorer even than the traditional SPH method. The poor performance of the CSPM in
simulating shock waves should arise from the denominator that in fact acts as a
normalization factor. Revisiting equation (67), it is found that the summation of the
smoothing function around the discontinuity region is far from the unity. When it is used
to normalize the summation of the density, the local features of shock wave can be
smoothed out, and hence the shock physics can be concealed. In contrast, the traditional
SPH does not suffer from this smoothing effect, and the DSPH benefits from the addition
part in resolving the shock physics. Their performance, therefore, should be better than
that of the CSPM. It is clear that the different performances of CSPM and DSPH
originate from the additional corrective parts in DSPH that approximate the discontinuity.

33
 1

0.8

Density (Kg/m )
3
0.6

0.4

Exact solution
0.2 DSPH
Traditional SPH
CSPM
0
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
x (m)

Figure 8 Density profiles for the shock tube problem obtained using different versions of
SPH formulation. DSPH gives the best result. CSPM failed to capture the major shock
physics. (t = 0.2 s).

Exact solution
1.2
DSPH
Traditional SPH
1 CSPM
Pressure (N/m )
2

0.8

0.6

0.4

0.2

0
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
x (m)

Figure 9 Pressure profiles for the shock tube problem obtained using different versions
of SPH formulation.

34
 1
Exact solution
DSPH
Traditional SPH
0.8 CSPM

Velocity (m/s)
0.6

0.4

0.2

0
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
x (m)

Figure 10 Velocity profiles for the shock tube problem obtained using different versions
of SPH formulation.

2.7
Exact solution
2.6 DSPH
Traditional SPH
2.5 CSPM

2.4
Internal energy (J/Kg)

2.3

2.2

2.1

1.9

1.8
-0.4 -0.3 -0.2 -0.1 0 0.1 0.2 0.3 0.4
x (m)

Figure 11 Energy profiles for the shock tube problem obtained using different versions
of SPH formulation.

4.5 A general approach to restore particle inconsistency

Liu et al gave a general approach to restore particle inconsistency through

reconstructing the smoothing function [6]. In general, a smoothing function can be
written in the following form

35
 x  xj x  xj
W ( x  x j , h)  b0 ( x ,h)  b1 ( x ,h)( )  b2 ( x ,h)( )2  ......
h h
. (79)
k x  xj
  bI ( x ,h)( ) I

I 0 h
By substituting the above smoothing function into equation (45), and after some trivial
transformation, the discretized form of equation (45) can be written as
k N x  xj 
 bI ( x ,h) ( h
) I x j  1 
I 0 j 1 
k N x  x j I 1 

I 0
bI ( x ,h) (
j 1 h
) x j  0 
. (80)
 

k N x  x j I k 

I 0
bI ( x ,h) (
j 1 h
) x j  0 

The k  1 coefficients bI ( x ,h) can then be determined by solving the following matrix
equation
 m0 ( x ,h) m1 ( x ,h)  mk ( x ,h)  b0 ( x ,h)  1 
 m ( x ,h) m ( x ,h)  m ( x ,h)   b ( x ,h)  0 
 1 2 1 k  1   
     , (81)
          
mk ( x ,h) mk 1 ( x ,h)  mk  k ( x ,h)  bk ( x ,h)  0 
  
M b I

or
Mb  I , (82)
where
N x  xj
mk ( x, h)   ( ) k x j (83)
j 1 h ,
M is a moment matrix, b is a vector of coefficients, I is a vector of given constants.
After determining the coefficients bI ( x ,h) , the smoothing function expressed in
equation (79) can be calculated. The procedure ensures particle consistency to kth order.
Therefore, the particle consistency restoring process actually gives an approach to
construct some kind of smoothing function for the SPH methods.
Comparing with the traditional smoothing function, which is only dependent on the
particle distance and applicable for all the particles, the consistency restored smoothing
function is particle-wise. It therefore depends on both the distance and positions of the
interacting particles. The cost-effectiveness for this approach in constructing particle-
wise smoothing functions needs to be considered since it will require additional CPU

36
time to solve the particle-wise equation (79) for all the particles. Moreover, since all
particles are moving, the particle location is changing as well. Hence, the particle-wise
smoothing functions need to be computed for every time step. Another problem is that,
to solve equation (81), the moment matrix M is required to be non-singular. Therefore,
the particle distribution must satisfy certain conditions to avoid singular momentum
matrix. This implies that when we enforce on consistency, we will face the stability
problem shown as the bad-conditioned moment matrix in the SPH settings. In contrast,
in the original SPH method, particles can be arbitrary distributed, though the obtained
results may be less accurate.
As far as the approximation is concerned, restoring particle consistency is an
improvement on the accuracy of the particle approximation, provided that the moment
matrix M is not singular. However, it is noted that restoring the consistency in discrete
form leads to some problems in simulating hydrodynamic problems. Firstly, the resultant
smoothing function is negative in some parts of the region. Negative value of smoothing
function can leads to unphysical representation of some field variables, such as negative
density, negative energy that can lead to a breakdown of the entire computation.
Secondly, the resultant smoothing function may not be monotonically decreasing with the
increase of the particle (node) distance. Moreover, the constructed smoothing function
may not be symmetric and using this non-symmetric smoothing function violates the
equal mutual interaction in physics.
4.6 Finite particle method

Considering the disadvantages of the above-mentioned particle inconsistency

restoring approach in constructing a point-wise smoothing function, Liu et al. devised
another particle consistency restoring approach, which retains the conventional non-
negative smoothing function instead of reconstructing a new smoothing function [55, 56].
This approach has been termed as Finite Particle Method (FPM), in which a set of basis
functions can be used in the numerical approximation.
Performing Taylor series expansion at a nearby point xi  {xi , yi , zi } and retaining the
second order derivatives, a sufficiently smooth function f ( x ) at point x  {x, y, z} can
be expressed as follows

  ( x   xi )( x   xi )
f ( x )  fi  ( x  xi ) f i ,  f i ,  r (( x  xi )3 ) , (84)
2!

where ,  are the dimension indices repeated from 1 to 3 (or from x to z). r (( x  xi )3 )
is the remainder of the expansion. fi , fi , and fi , are defined as

fi  f ( xi ) , (85)

37
 fi ,  f ( xi )  (f  x )i , (86)

fi ,  f ( xi )  ( 2 f  x x  )i . (87)

Multiplying both sides of equation (84) with a function 1 ( x  xi ) and integrating

over the problem space  can yield the following equation

 f ( x) ( x  x )dx = f   ( x  x )dx  f   ( x


1 i i 1 i i,  xi )1 ( x  xi )dx
  
. (88)
f i ,
 ( x  xi )( x  xi )1 ( x  xi )dx  r (( x  xi ) )
   
 3

2 

It is seen that the integration is carried out over the entire problem space, and can be
quite time-consuming. One usual assumption is that a field variable at point xi is only
strongly influenced by the field variables at nearby points and that the influence of the
field variables at points far away from point xi is very weak and hence can be neglected.
Therefore, the global integration can be converted into a local integration by defining a
support domain for point xi in which the field variables at point xi can be determined.
The shape of the support domain can be conveniently taken as a circle (in 2D) or a sphere
(in 3D) with a radius of  h , in which  is a constant scalar factor, and h is a length
characterizing the support domain. The function 1 ( x  xi ) is also limited to the local
support domain, and can be rewritten as 1 ( x  xi , h) .
Since the points distributed in the problem space are actually particles, each
occupying individual lumped volume, equation (88) can be numerically approximated by
summation over the particles surrounding point xi as follows
N N


j 1
f ( x j )1 ( x j  xi , h) V j = f i  1 ( x j  xi , h)V j
j 1
N
 fi ,  ( xj  xi )1 ( x j  xi , h)V j , (89)
j 1

f i , N

2
 ( x  x )( x   x  ) ( x
j 1
j i j i 1 j  xi , h)V j

where N is the number of particles within the support domain of particle i. The
remainder term r (( x  xi )3 ) in equation (88) is omitted in equation (89) for the sake of
conciseness. The summation over the particles is illustrated in Figure 3.
Equation (89) can be further simplified as the following equation at point xi

38
 B1i = A1ki Fki , (90)

where

Fki  [ f i fi , fi , ]T , (91)

N
B1i =  f ( x j )1 ( x j  xi , h)V j , (92)
j 1

N
A1ki  [ 1 ( x j  xi , h)V j
j 1
N

 ( x  x ) ( x
j 1
j i 1 j  xi , h)V j . (93)

1 N 
 ( x j  xi )( x j  xi )1 ( x j  xi , h)V j ]
2 j 1

Corresponding to 1, 2, and 3 dimensional cases, there are one function value, 1, 2 and
3 first derivatives, and 1, 3 and 6 second derivatives that will be approximated. It is clear
that k in equations (91), (93) and (90) is 3, 6, and 10 respectively corresponding to 1, 2,
and 3 dimensional cases. To calculate the function value, the first and the second
derivatives at xi , 2, 5, and 9 other equations similar to equation (90) are required.
Therefore in 1, 2, and 3 dimensional cases, totally 3, 6, and 10 functions (  M ( x  xi , h) ,
M  3, 6, or 10 ) are required in order to approximate the function value, the first
and second derivative. These functions form a set of basis functions used for
approximating the function value, its first and second derivatives. A conventional SPH
smoothing function and its first and second derivatives can form a set of basis function in
the FPM. For example, in 2D space, a smoothing function W , its two first order
derivatives, W and W , and its three second order derivatives, W , W and W forms
a set of 6 basis functions.
In summary, multiplying a set of basis functions on both sides of equation (84),
integrating over the problem domain, summing over the nearest particles within the local
support domain of particle i , a set of matrix equation can be produced to approximate the
function value as well as the first and second derivatives at particle i . The matrix
equations at particle i at can be written as
BMi = AMki Fki or B = AF (94)

39
where

N
BMi =  f ( x j ) M ( x j  xi , h)V j (95)
j 1

N
AMki  [  M ( x j  xi , h)V j
j 1
N

 ( x  x )
j 1
j i M ( x j  xi , h)V j (96)

1 N 

2 j 1
( x j  xi )( x j  xi ) M ( x j  xi , h)V j ]

Equation (94) is the basis of the finite particle method and can be used to approximate
a function value and its derivatives for a field variable. It is seen that only if the
coefficient matrix A is not singular, can these M equations determine a unique set of
solutions at particle i for the M unknowns in vector F. Solving the above pointwise
matrix equations, the function value as well as the first and second derivatives at every
particle can be simultaneously approximated. Note that the conditioning of A matrix
refelects the stability of the FPM model.
Since the governing equations in CFD only involve the first and second derivatives,
only the derivatives up to the second order are retained in equation (84). For problems in
other areas such as computational solid mechanics, high order derivatives may be
involved. If third or higher order derivatives are to be approximated, in expanding f ( x )
at xi , the interested derivatives need to be retained in equation (84). To obtain the
increased number of unknowns, more functions like  M ( x  xi , h) are necessary to
complete the matrix equation (94). Therefore, except for the increased number of
unknowns, increased number of basis functions, and therefore more computational efforts,
the solution procedure for the interested unknowns is the same.
Comparing conventional SPH and FPM, it is clear that both FPM and SPH are
meshfree particle methods in which particles with lumped volumes are used to represent
the state of a system. The particles form a frame for interpolation, differencing or
integration in a certain approximation. Both FPM and SPH can be used as Lagrangian
methods if allowing the particles to move in the problem space. However, the difference
between FPM and SPH is obvious.
1. FPM uses a set of basis function to approximate the function value and its
derivatives, whereas SPH employs a smoothing function and its derivatives to
approximate a function value and the corresponding derivatives. The smoothing
functions in SPH should have some special properties as described in Section 3.
However, the basis functions in FPM are more general. Any set of functions

40
 which do not lead to a singular coefficient matrix A can be used as basis functions.
Therefore the smoothing function and its specific derivatives actually can be one
possible choice as a suitable set of basis functions.

2. SPH can be regarded as a special case of FPM, whereas FPM is a generalized

version of SPH with modifications. In equation (84), if all the terms related to
derivatives are neglected, multiplying both sides of equation (84) with the
smoothing function W, and integrating over the problem space can lead to the
approximation in SPH. Summation over the nearest particles within the support
domain of a particle further produces the particle approximation of the field
variable at that particle (equation (89)).

3. FPM should have better accuracy than SPH. Since no derivative term is retained
in equation (84), the SPH method actually is of first order accuracy. If a
symmetric smoothing function is used, the terms related to the first order
derivatives are actually zero for the interior particles in the problem domain.
Therefore SPH is of second order accuracy in interior parts. In contrast, since up
to second order derivatives are retained in the expansion process, the accuracy of
FPM is of third order. Moreover, if higher derivatives are retained, better
accuracy can be achieved. FPM should have a better accuracy than SPH both for
the interior particles and boundary particles.

4. The accuracy of FPM is not sensitive by the selection of smoothing length, and
extremely irregular particle distribution. This has been demonstrated by Liu and
his co-workers in testing one- and two-dimensional cases.

5. As the solution is to be obtained from solving the matrix equation (94), a good
matrix inversion algorithm is necessary to prevent the co-efficient matrix to be
negative.

6. The matrix equation (94) is to be solved at every particle, and every time step. It
can be more computationally expensive than the conventional SPH method.
One simpler example of FPM is to only consider the first derivative in equation (84).
Using the smoothing function and its first derivatives as the basis functions, the following
equations can be obtained

 f ( x)W ( x)dx  f  W ( x)dx  f   ( x


i i i i,  xi )Wi ( x )dx , (97)
and

 f ( x)W  dx  f  W  dx  f   ( x

i, i i, i,  xi )Wi ,  ( x )dx . (98)

Again  is the dimension index repeated from 1 to d. The corresponding discrete forms
for equations (97) and (98) are
N N N

 f W v
j 1
j ij j  f i   Wij v j  fi ,   ( xj  xi )Wij v j ,
j 1 j 1
(99)

41
and
N N N

j 1
f jWij ,  v j  f i   Wij ,  v j  f i ,   ( xj  xi )Wij ,  v j .
j 1 j 1
(100)

There are d  1 equations for d  1 unknowns ( fi and f i , ). Equations (99) and (100)
are therefore complete for solving with respect to fi and fi , , and the solutions for fi and
fi , are
1
 N N
  N 
  Wij v j    f j Wij v j 
 
( x j  x i )Wij v j 
 fi   j 1 j 1
  j 1 .
f  N N  N 
(101)
 i , 
  Wij ,  v j  ( xj  xi )Wij ,  v j    f j Wij ,  v j 
 j 1 j 1   j 1 
In equations (99) and (100), the terms related to the function and the first order
derivatives are all retained, only the terms related to second or high order derivatives are
neglected. Therefore the resultant particle approximations for a function and its
derivatives (equation (101)) are able to exactly reproduce a constant and a linear function
(C0 and C1 consistency). Hence the algorithm shown in equation (101) actually restores
the particle consistency that conventional SPH method does not have. Again, this
particle consistency restoring approach is independent of the particle distribution (either
regular or irregular), and the choices of the smoothing kernel and smoothing length.
Another advantage is that this particle consistency restoring approach does not change the
conventional smoothing function and should be preferable in simulating hydrodynamics.
In summary, the consistency of the conventional SPH method, CSPM, and FPM (if
using equation (101), rather than equation (94), which is associated with higher order
particle consistency than equation (101)) are described in the following two tables.

Table 1 Kernel consistency of SPH, CSPM, and FPM

Interior domain Boundary area

SPH 1st order less than 0th order

CSPM 1st order 0th order

FPM 1st order 1st order

42
 Table 2 Particle consistency of SPH, CSPM, and FPM

Interior domain Boundary area

Regular distribution Irregular distribution

SPH 1st order less than 0th order less than 0th order

CSPM 1st order 0th order 0th order

FPM 1st order 1st order 1st order

Figure 12 Approximation results for a linear function f ( x, y )  x  y using (a) SPH

and (b) FPM in a domain of [0, 1, 0, 1].

Figure 12 shows the approximation results for a linear function f ( x, y )  x  y using

discrete models of SPH and FPM. The particles are uniformly distributed. It is seen that
the SPH results are associated with oscillation especially near the boundary, whereas the
particle consistency restoring approach can obtain much better results, which are in very
good agreement with the analytical solution. It is worth noting that if the particles are
randomly or irregularly distributed, FPM can also obtain accurate results, whereas, SPH
results can be far away from the analytical solutions. It is also worth noting that in this
simple numerical test, the approximation is carried out by obtaining numerical values at
current step from the existing numerical values at the particles at the previous step. There
is no physical dynamics involved in these two numerical performance studies. Only the
accuracy of the numerical approximations of the given function is studied using different
approximation methods. It can be expected that numerical errors existing in the present
approximated results can propagate to the next approximation steps and can even be
magnified. However, when these methods are used to simulate a well-posed physical
problem, the physics is governed by the conservation equations. If the disrete model is

43
stable, the error in the current step is controlled by the consistency. Therefore, the errors
should be more or less within some level, and will not be magnified from one step to the
next step.

4.7 Consistency vs. stability

We have seen in the two previous sections that a typical dilemma exists for many
numerical methods: consistency or stability. For a given “setting” of a numerical model
such as a particle method, we may choose one over another, but probably a “balanced”
one can be difficult to choose. The original SPH has clearly chosen the stability (and also
flexibility) over the consistency, which gives the SPH a distinct feature of working well
for many complicated problems with good efficiency, but less accuracy. It always tries to
deliver some reasonably good results for the price paid. This seems to be a very practical
choice for many practical engineering problems, as should be regarded as an advantage of
the SPH method. Attempts to improve the accuracy of SPH via restoring the consistency
can be helpful, provided that the stability and efficiency is not too much compromised.
The consistency restoring approaches, such as the FPM, put more emphases on the
consistency (hence hopefully accuracy), but the stability (the conditioning of the M or A
matrix) can be in question for some types of problems. Hence proper measures are
needed to establish the stability ensuring the accuracy of the solution.
The question is that can we have both the consistency and the stability at the same
time? The answer is yes, provided we are willing to change the “setting” and pay the
price. The recently proposed GSM method [31-34] is a typical example that guarantees
both excellent stability for arbitrary grids and 2nd order accuracy. It uses also the gradient
smoothed technique, but in a very carefully designed fashion. However, the GSM is not
a particle method any more; the simplicity and efficiency features of the particle methods
are lost. The GSM requires precise evaluation of the integrals for carefully chosen types
of smoothing domains, and it works more like FVM. Therefore, the final question
depends on what we want and at what cost: choosing a numerical method should be
closely related to the nature of the problem, the requirement for the the solutions and the
resources we have.

5 Special topics

5.1 Solid boundary treatment

Particle methods such as MD, DPD, and SPH are very different from the grid based
numerical methods in treating solid boundary, and care must taken in this regards. For
finite element methods, proper boundary conditions can be well defined and proper
imposed without affecting the stability of the discrete model. In the finite difference
method, because the domain is relatively regular, proper ways to handling the boundary
conditions have already been developed. In either FEM or FDM, the implementations of
boundary conditions (either Neumann boundary conditions or Dirichlet boundary
conditions, or mixed boundary conditions) are generally straightforward [2, 3, 5, 92]. In
contrast, in the MD, DPD and SPH, the implementation of the boundary conditions is not

44
as straightforward as in the grid based numerical models. This has been regarded
historically as one weak point of the particle methods [12, 15, 76, 78, 93-98]. For
molecular dynamics and dissipative particle dynamics, it is common to fill the solid
obstacle areas with frozen particles. These fixed particles can prevent the mobile particles
from penetrating the solid walls, and interact with the mobile particles with proper
interaction models. Different complex solid matrix and solid-fluid interface physics can
thus be modeled by appropriately deploying fixed particles plus a suitable interaction
model with mobile particles. Preventing fluid particles from penetrating solid walls by
exerting a repulsive force simulates the boundary condition with zero normal velocity.
Different models of reflection or mirror of fluid particles or interaction models of solid
particles with fluid particles can also be applied to exert slip or non-slip boundary
condition [11, 39, 96, 99-102]. In principal, the boundary treatment techniques in MD
and DPD can also be used for SPH. However, SPH is a continuum scale particle method,
and the field variables on boundaries such as pressure need to be directly calculated.
Therefore solid boundary treatment in SPH is more difficult and numerical techniques for
implementing solid wall boundary conditions are also more diversified. Since the
invention of the SPH method, a lot of works have been published addressing the
treatments of solid wall boundary conditions [93, 96, 103-110].
As discussed in Section 4, the conventional SPH model suffers from particle
inconsistency that is closely related to the selection of smoothing function, smoothing
length, and most importantly the distribution of particles. An extreme case is the particle
distribution near the boundary area. Only particles inside the boundary contribute to the
summation of the particle interaction, and no contribution comes from the outside since
there are no particles beyond the boundary. This one-sided contribution does not give
correct solutions, because on the solid surface, although the velocity is zero, other field
variables such as the density are not necessarily zero.

Interior particles
Boundary particles

Figure 13 SPH kernel and particle approximations for interior and boundary particles. The
boundary particles suffer from particle insufficiency, and no contribution comes from outside
since there are no particles beyond the boundary.

Some improvements have been proposed to treat the boundary condition. Monaghan
used a line of ghost or virtual particles located right on the solid boundary to produce a

45
highly repulsive force to the particles near the boundary, and thus to prevent these
particles from unphysically penetrating through the boundary [111]. Campbell treated
the boundary conditions by including the residue boundary terms in the integration by
parts when estimating the original kernel integral involving gradients [112], as discussed
in Section 3 in deriving the constructing conditions for the smoothing kernel function.
Libersky and his co-workers introduced ghost or image ( or imaginary) particles to reflect
a symmetrical surface boundary condition with opposite velocity on the reflecting image
particles [113]. Later, Randles and Libersky proposed a more general treatment of the
boundary condition by assigning the same boundary value of a field variable to all the
ghost particles, and then interpolating smoothly the specified boundary ghost particle
value and the calculated values of the interior particles [51].
In general, it is feasible to use virtual particles to implement the solid boundary
conditions. These virtual particles can be allocated on and outside the boundary. These
virtual particles can be classified into two categories, virtual particles that are located
right on the solid boundary (Type I) and virtual particles that are outside the boundary
(Type II) (see Figure 14). Type I virtual particles are similar to what Monaghan used
[111], and they are used to exert a repulsive boundary force on approaching fluid (real)
particles to prevent the interior particles from penetrating the boundary. The penalty
force is calculated using a similar approach for calculating the molecular force of
Lennard-Jones form [114]. If a Type I virtual particle is the neighboring particle of a real
particle that is approaching the boundary, a force is applied pairwisely along the
centerline of these two particles as follows

Figure 14 Schematic illustration of flow region, solid boundary, real fluid particles and
virtual particles.

  r r  xij r0
 D  ( 0 ) n1  ( 0 ) n2  2 ( ) 1
  rij rij  rij rij
PBij    . (102)
 r
0 ( 0 ) 1
 rij

46
where the parameters n1 and n2 are empirical parameters, and are usually taken as 12 and
4 respectively. D is a problem dependent parameter, and should be chosen to be in the
same scale as the square of the largest velocity. The cutoff distance r0 is important in the
simulation. If it is too large, some particles may feel the repulsive force from the virtual
particles in the initial distribution, thus leads to initial disturbance and even blowup of
particle positions. If it is too small, the real particles may have already penetrated the
boundary before feeling the influence of the repulsive force. In most practices, r0 is
usually selected approximately close to the initial particle spacing. To more effectively
prevent fluid particles from penetrating the solid wall, Type I virtual particles are usually
more densely distributed at the solid boundary than the fluid particles (Figure 14).
Type II virtual particles are usually obtained through reflecting the real fluid particles
along the solid boundary. Scalar properties of the virtual particles can be taken as the
same as the reflected real particles, while velocity of the virtual particles are usually
taken as opposite of the reflected counterparts. As such the non-slip boundary condition
is implemented. Due to the compactness of the smoothing kernel function, only the real
particles that are within  h to the solid boundary are necessary to be reflected to form
Type II virtual particles. Hence there are only several layers of Type II virtual particles
outside the solid boundary. For example, if the cubic spline smoothing function is used,
there are two layers of Type II virtual particles (if the real particles are regularly
distributed). Therefore, for a fluid particle i close to the solid boundary, all the
neighboring particles NN (i) that are within its influencing area of  hi can be categorized
into three subsets ((see Figure 14))
a) I (i) : all the interior particles that are the neighbors of i (real particles);
b) B(i) : all the boundary particles that are the neighbors of i (virtual particles of
type I);
c) E (i) : all the exterior particles that are the neighbors of i (virtual particles of type
II).
These two types of virtual particles are specially marked for contribution in the later
summation on the real particles. Type I virtual particles exert repulsive force on an
approaching particle, but they usually do not contribute in the particle approximation.
Type II virtual particles take part in the kernel and the particle approximations of the real
particles. It is clear that by using type II virtual particles, the one-sided particle
distribution problem in the conventional SPH method can be somewhat remedied. In
different implementations, the position and physical variables for type II virtual particles
can be fixed (determined once a set of virtual particles are obtained), or can adapt in the
simulation process with the evolution of the corresponding counterpart real particles. In
general, adapting type II virtual particles can result in better results. However, the
adapting process may be time-consuming, and may be complicated especially when the
solid wall is not able to be defined in a simple curve.

47
 v
vv
f df dv
vf

Figure 15 A general approach to construct type II virtual particles and assign velocities.
Only one virtual particle is plotted in the obstacle region for illustration.

In many circumstances, the computational domain can be complicated, and can

contain solid obstacles of arbitrary shapes. Generating type II virtual particles may not
be an easy task. For such cases, a feasible way is to deploy particles into the
computational domain either by placing all the particles in some kind of lattice, or
injecting particles into the computational domain and then running SPH simulation for
the entire system to reach equilibrium, just as in MD and DPD simulation. The particles
outside of the boundary (or within the solid obstacles) can be used as type II virtual
particles. This approach has some advantages since the system starts from a quiet or
equilibrium state of the same number density without unwanted noise. For implementing
the non-slip boundary condition, these virtual particles can move with the velocity as the
obstacles to which they are attached. For better accuracy, the velocity of the virtual
particles can be obtained from extrapolating the velocity of a neighboring fluid particle,
as illustrated in Figure 15. Assuming the normal distance of a virtual particle to the solid
boundary is dv , and the normal distance of a fluid particle to the solid boundary is d f ,
with non-slip boundary condition (velocity on the boundary is zero), the velocity of the
virtual particle under consideration should be vv    dv d f  v f . Hence the velocity
 d 
difference between the fluid and virtual particles is v fv  1  v  v f , which can be used in
 df 
calculating viscous forces as addressed in Section 5. Note that the positions of the virtual
particles do not need to evolve. Also if the obstacle is a moving object, the velocity of
the fluid particle should be taken as the relative velocity to the obstacle. To prevent
numerical singularities caused by a fluid particle closely approaching the solid boundary
( d f  0 ), a safety parameter  can be used in calculating the velocity difference

48
  dv 
between the fluid and virtual particles as v fv   v f , where   min  max ,1   .  max is

 df 
an empirical parameter, it is reported that taking 1.5 can result in good results [110].
5.2 Representation of solid grains

(a) (b) (c)

Figure 16 Schematic illustration of the treatment of solid obstacles. (a) The cells in the
entire computational domain are first labeled, “0” for fluid (void) cells and “1” for solid
(obstacle) cells, (b) The SPH particles in the obstacle cells are frozen. (c) Only the frozen
particles that are close to the fluid cells (within 1 cut-off distance) are retained as boundary
particles.

To model the geometries of a complicated solid matrix, the entire computational

domain can be discretized using a ‘shadow’ grid. The grid cells are labeled “0” for
regions occupied by pore spaces and “1” for solid-filled regions (Figure 16a). This
simple identification of fluid and solid cells can be used to represent any arbitrary
complex geometry. The unit vectors normal to the solid-fluid interfaces define the local
orientation of the fluid-solid interface and can be obtained by simply calculating the
surface gradient from the indicator numbers (“0” for liquid regions and “1” for solid
regions). As discussed in Section 6.1 in constructing type II virtual particles, we can
place particles in the entire computational domain regularly in some kind of lattice
according to a designated number density. It is also applicable to inject particles into the
computational domain, and run SPH simulation until the system arrives at an equilibrium.
At the beginning of each simulation, the particles are initialized and positioned randomly
within the entire computational domain until a pre-defined particle number density is
reached, and the system is then run to equilibrium. The particles within the solid cells
(marked as “1”) are then ‘frozen’ (their positions are fixed) to represent the solid grains
(Figure 16b). The solid grains in obstacles can occupy a considerable fraction of the
entire computational domain, and hence the number of frozen particles representing the
solids can be very large. Most of the frozen particles inside the solid grains are more than

49
1 cut-off distance away from the adjacent fluid cells. These particles do not contribute to
the solid-fluid interactions and consequently they have no influence on the movement of
the flow particles within the fluid cells. Therefore, only the frozen particles that are
within 1 cut-off distance from the solid-fluid interface are retained as boundary particles
(Figure 16c), and the rest of the particles further inside the solid grains are removed from
the model domain.
Figure 17 and Figure 18 show two example of solid obstacle representation. In
Figure 17, the flow region is a porous media system, in which the circular obstacles
occupy a great amount of the computational domain. Figure 17a shows the original
porous media system where the circular obstacles are filled with (solid) particles. Figure
17b shows the porous media system for later simulation where only the (solid) particles
within 1 cut-off distance away from the adjacent fluid cells are retained. In Figure 18,
the flow region is a fractured network system and the obstacles also occupy most of the
computational domain (Figure 18a). After treatment, only the fracture particles within 1
cut-off distance away from the adjacent fluid cells are retained. For both cases, it is
convenient and flexible to treat the complex geometries by identifying the flow and
obstacle regions separately with the indicator numbers. Also since only the solid
particles that are within 1 cut-off distance away from the adjacent fluid cells are retained
for later simulation, the number of particles can be greatly reduced, and therefore the
computational efficiency can be significantly improved.

(a) (b)

Figure 17 An example of solid obstacle representation in a porous media flow region with
circular obstacles. (a) Original porous media system. (b) Solid particle distribution for later
simulation.

The above approach of deploying frozen particles to represent solid grains and
interact with flow particles is applicable to particle methods such as MD, DPD, and SPH,
and it is especially effective for complex boundaries [98, 115-117]. For problems with
comparatively simpler geometries, some other approaches are commonly used. These
approaches may also involve particles fixed on the solid boundaries with or without

50
equilibrium process, and may involve non-fixed boundary particles updated with the
neighboring flow particles. Although fixed boundary particles are more frequently used
in SPH literature, non-fixed boundary particles updated with the neighboring flow
particles may produce better results [6, 109, 110, 118].

(a) (b)

Figure 18 An example of solid obstacle representation in a fracture network in which most

of computational domain is full of obstacles. (a) Original fractured network system. (b) Solid
particle distribution for later simulation.

5.3 Material interface treatment

For modeling systems with multiple materials, an effective algorithm is required for
efficiently describing the material interface dynamics. It is somewhat different from the
grid based numerical methods in correctly detecting contact and effectively treating it in
particle methods like SPH. Traditionally SPH uses particles to represent the state of a
system under consideration. One point is that the SPH particles are interpolation points
on which mass of each particle has been lumped. An SPH particle has a finite mass and
density, and hence can have a finite volume. However, in contrast to FEM, FDM and
FVM, in which a cell or an element has a definite shape, an SPH particle has never
assumed as a definite shape. On the other hand, the shape of a particle can also be
associated with the influencing area (or support domain, that is kh ) of the particle.
Generally an SPH particle is assumed to be circular (or spherical) when using a
smoothing kernel function with a scalar smoothing length (Figure 19a). In some cases,
an SPH particle can be elliptical (or ellipsoidal) when a tensor smoothing length is used
[119-123] (Figure 19b). In few simulations, it even can be taken as a square or a cube
[124] (Figure 19c). In the SPH simulation, the mass of a particle is usually assumed to
be constant, and updating the density means change of the volume. If the particle takes a

51
definite shape such as a circular, then the size of the particle also changes. Therefore
different shapes and sizes of the SPH particles increase the difficulty in implementing the
contact algorithm. During the course of SPH development, different contact algorithms
have emerged [125-141]. For simplicity, only SPH models with circular or spherical
shaped particles are discussed here. The SPH particles are assumed to be hard-spheres
rather than soft balls, as models with soft balls also increase the difficulty in
implementing the contact algorithm.

khy khx

(a) (b) (c)

Figure 19 Possible shapes of an SPH particle. (a) circular shape corresponding to a scalar
smoothing length, (b) elliptical shape corresponding to a tensor smoothing length at
different directions, (3) square shape whereas the smoothing length can also be a tensor.

Figure 20 Illustration of material interface with SPH particle model. (a) Initial SPH particle
distribution, and (b) SPH particle evolution. Interaction between particles from different
materials can introduce shear and tensile stress and can prohibit sliding and separation of
different materials.

52
 In the SPH particle models, if particles from different materials are not bonded
together, particles from one material can influence, and also be influenced by the
particles from the other material, according to conventional SPH algorithm. Figure 20
shows an example of the material interface at the initial stage, and at a later stage in the
evolution. Initially the SPH particles are distributed regularly, and there is no
intersection and deformation of particles. With the SPH evolution, particles from
different materials influence each other, and the conventional SPH algorithm can
introduce errors since particles from material A influence the particles from material B in
calculating the strain and strain rate, and vice versa. This interaction between particles
from different materials can introduce shear and tensile stress, which prevent sliding and
separation of different materials. Special algorithms are required if an SPH model is used
to for simulation of problems with sliding and separation.
Major steps to be considered in a contact algorithm in the SPH simulation include 1)
identifying the boundary, 2) detecting the contact and 3) applying repulsive contact
forces. Boundary (or interface) has usually been considered to be a surface one half of
the local smoothing length away from the boundary SPH particles [51, 126]. A
convenient way to detect particle-particle contact is to check the proximity of two
approaching particles from different materials, and compare with the summation of the
smoothing lengths of the particles. As shown in Figure 21, the contact and penetration
between two particles can be identified by
khi  kh j
pe   1, (103)
rij

where rij is the distance between particle i and particle j . After detection of the contact
and penetration, a restoring contact force needs to be applied along the centerline of the
two particles (position vector of particle i and particle j ).

Figure 21 An illustration of particle and particle contact.

There are different choices for the restoring contact force, which is usually a function
of penetration. One possible choice is the penalty force in Lennard-Jones form. Liu et al.
had used this penalty force for simulating underwater explosion when dealing with
different material interface [142]. In their work, the Lennard-Jones form penalty force is

53
applied pairwisely on the two approaching particles along the centerline of the two
particles
 xij
 p ( pe  pe ) 2 pe  1
n1 n2

f ij   rij , (104)
0 pe  1

where the parameters p , n1, n2 are usually taken as 105, 6 and 4 respectively. In fact,
these parameters can be adjusted to suit the needs of different problems, and the involved
driven force in underwater explosions can vary in a very big range [143]. The application
of the penalty force in combination with the summation between the interface particles
has been shown to well prevent the unphysical penetration in the simulation of
underwater explosion, though numerical oscillations near the vicinity of the interface can
still exist.
It is reasonable that when the penetration of two contacting particle increases (or
particle distance decreases), the magnitude of the restoring force should also increase.
Monaghan [144] used the following form of repulsive force, which is controlled by the
kernel and the global distribution of the particles
Wij
fij  , (105)
W ( p )
where p denotes the average particle spacing in the neighborhood of particle i
(proportional to smoothing length h).
Vignjevic et al. [145] used the following repulsive force for treating possible contact
and penetration between solid particles
N m j Wij
fi  K f  , (106)
j 1 i  j xi
where K f is a user defined empirical parameter. In their work, the force was applied to
boundary particles that are within the influencing area kh of each other. This repulsive
force introduces the derivatives of the smoothing kernel, which in general reaches the
maximum at a certain location between the origin and the boundary of the influencing
area, and vanishes at the origin and the boundary point (see Figure 4). Considering that
decreasing the inter-particle distance means bigger repulsive force, using the kernel itself
rather than its derivative physically would be more reasonable. For example, one
possible choice can be
N mj xij
fi  K f  Wij . (107)
j 1 i  j rij2

54
 5.4 Tensile instability

When using the SPH method for hydrodynamics with material strength, one
numerical problem called tensile instability [44, 61, 62, 73, 144, 146-151] may arise. The
tensile instability is the situation that when particles are under tensile stress state, the
motion of the particle becomes unstable. It could result in particle clumping or even
complete blowup in the computation. Swegle et al. demonstrated the tensile instability
effects in a two-dimensional space as illustrated in Figure 22 [44]. It is a simple test in
which the particles representing the object initially are distributed regularly and under
uniform initial stress. The boundary particles are fixed (two layers) and the interior
particles are free to move due to some kind of external perturbation. If applying a very
small velocity perturbation on a center particle, theoretically, it will take a long time for a
particle under consideration to move over single particle spacing. However, in an SPH
simulation with a common cubic spline smoothing kernel function, if the stress is tensile,
after a number of time steps (while the number of steps is not sufficiently to move the
particles even a single particle spacing), the interior particles can clumped together
unphysically, and form a void. It shows that the particle system exhibits a numerical
instability from tensile stress.

(a) (b)

Figure 22 Schmatic illustration of the tensile instability in a two-dimensional space. (a)

Initial particle distribution; (b) Particle distribution in a certain stage under the effect of
tensile instability.

According to Swegle, the tensile instability depends neither on the artificial viscosity,
no on the time integration scheme. It is closely related to the selection of smoothing
kernel function. In a one-dimensional von Neumann stability analysis, Swegle et al. gave
a criterion for being stable or instable in terms of the stress state and the second
derivative of the smoothing function, i.e., a sufficient condition for the unstable growth is

W    0 , (108)

55
where W is the second derivative of the smoothing function.
In the SPH method, the cubic spline smoothing function (illustrated in Figure 4) is
mostly commonly used. The initial smoothing length is usually set to be equal to the
particle spacing. Under such circumstances, the first nearest neighbor particles are
located at r h  1 ; and the next nearest neighbor particles are at r h  2 . As can be seen
from Figure 23, the second derivatives of the cubic spline function from r h  1 to r h  2
are always positive. Therefore it is expected that, according to equation (108), the SPH
method with the cubic spline function would be stable in a compressed state but could be
unstable in a tensile state in this region.

2.5

2
Tension, stable Tension, unstable
1.5 Compression, unstable Compression, stable

1 W ’’ < 0 W ’’ > 0

0.5 W
F u n ctio n s

-0.5

-1 W’

-1.5
W ’’
-2

-2.5
0 0.5 1 1.5 2
r/h

Figure 23 Schematic illustration of tensile instability with the cubic spline function and
its first and second derivatives.

Several remedies have been proposed to improve or avoid such tensile instability.
Morris suggested using special smoothing functions since the tensile instability is closely
related to the second order derivative of the smoothing function [45]. Though successful
in some cases, they do not always yield satisfactory results generally. Chen and his co-
workers proposed the corrective smoothed particle method (CSPM), which was reported
to improve the tensile instability [88]. Recently Monaghan and his colleagues proposed
an artificial force to stabilize the computation [144, 152].
The basic reason of tensile instability is that the SPH method is essentially a
collocation method, in which the particle approximations are conducted ONLY over the
particles that represent the entire system. This leads to insufficient “sampling” points for

56
establishing equations, and can result in numerical instability problem [4]. The situation
is very much similar to the so-called “node integration” in the implementation of the
element free Galerkin method (EFG) [153]. In the EFG method, the instability is restored
by adding stabilization terms in the Galerkin weak form. Based on this analysis, we need
somehow bring in more information from other points, in addition to these particles.
One of such a method is to make use of the information at additional points in the
support domain, rather than use only these collocation particles. Dyka et al. first
introduced additional stress points other than the normal particles in a one-dimensional
algorithm aimed at removing the tensile instability in SPH [61, 62]. The stress points
were also shown to be stable in tension and contributed considerably to the accuracy in
wave propagation problems. Later, this approach has been further extended to multi-
dimensional space by staggering the SPH particles using stress points so that there are
essentially an equal number in each set of points [63, 107, 150]. Basically in this
approach, two sets of particles are used. One set of SPH particles carry velocity, and are
referred to as “velocity particles”. The other companion set of particles carry all required
field variables except for the velocity, and are referred to as “stress particles”. Figure 24
shows two possible setups of velocity and stress particles.

Velocity particles

Stress particles

Figure 24 Two possible setups of velocity and stress particles.

Randles and Libersky pointed out that, the tensile instability for problems involving
material strength generally is latent. The growth rate of damages in solid continuum
models is often much faster than the grow rate of the tensile instability [63].
Except for problems with material strength which can experience tensile instability,
fluid mechanics problems sometimes can also meet tensile instability. Melean et al. had
showed the tensile instability in a formation of viscous drop [154], and the instability can
be removed by using the artificial stress proposed by Monaghan [144, 152].

57
6 Applications

The original applications of the SPH method is in astrophysical phenomena, such as

the simulations of binary stars and stellar collisions [48, 155, 156], supernova [157, 158],
collapse as well as the formation of galaxies [159, 160], coalescence of black holes with
neutron stars [161, 162], single and multiple detonation of white dwarfs [163], and even
the evolution of the universe [164]. It also has been extended to a vast range of problems
in both fluid and solid mechanics because of the strong ability to incorporate complex
physics into the SPH formulations [6]. The applications of SPH to many other
engineering applications include
 multi-phase flows [118, 165-173],
 coastal hydrodynamics including water wave impact, dam break, sloshing and
overtopping [111, 174-199],
 environmental and geophysical flows including flood and river dynamics,
landslide, flow in fractures and porous media, seepage, soil mechanics and
mudflow [116, 200-220],
 heat and/or mass conduction [52, 221-226],
 ice and cohesive grains [227-233],
 microfluidics and/or micro drop dynamics [122, 151, 168, 234-245],
 high explosive detonation and explosion [54, 124, 142, 246-255],
 underwater explosions and water mitigation [124, 142, 246-248, 254],
 elastic and/or plastic flow [51, 113, 256-258],
 fracture of brittle solids [259],
 metal forming and high pressure die casting [67, 260-269],
 magneto-hydrodynamics and magnetic field simulation [270-281],
 problems with fluid-solid interactions [175, 195, 209, 282-287], and
 many other problems like blood flow [288-291], traffic flow [292].
In this review, we will discuss applications of SPH to the following areas:
 high strain hydrodynamics with material strength,
 high explosive detonation and explosions, and underwater explosion,
 microfluidics and micro drop dynamics,
 coast hydrodynamics and offshore engineering, and
 environmental and geophysical flows.
6.1 High strain hydrodynamics with material strength

High strain hydrodynamics is generally characterized by the presence of shock waves,

intense localized materials response and impulsive loadings. Numerical simulation of
high strain hydrodynamics with material strength such as high velocity impact (HVI) and
penetrations is one of the formidable but attractive tasks in computational solid
mechanics. Most of the wave propagation hydro-codes use traditional grid based
methods such as finite difference methods and finite element methods to simulate high
strain hydrodynamics. Some of them are associated with advanced features which
attempt to combine the best advantages of FDM and FEM. Examples include arbitrary
Lagrange-Eulerian (ALE) coupling and coupling Eulerian-Lagrangian (CEL). Though

58
many successful achievements have been made using these methods, some numerical
difficulties still exist. These numerical difficulties generally arise from large
deformations, large inhomogeneities, and moving interfaces, free or movable boundaries
[255, 293-296].
SPH method has been intensively used for simulating such phenomena of high strain
hydrodynamics with material strength, due to its special features of meshfree, Lagrangian
and particle nature. Many researches have investigated impact (either high velocity or
not) and penetration using SPH method [49, 50, 85, 128, 131, 133, 138, 139, 297-334].
The applications are mainly in defense industry, and typical numerical examples range
from Taylor bar impacting on a solid wall and spherical projectile/bumper
collision/penetration [113, 329], projectile impact and penetration and an aircraft fuselage
nose impacting a rigid target [299], to bird strike impacting on an aircraft Wing [334],
and even to a high-quality high-fidelity visualization of the September 11 attack on the
world trade center [327].
Libersky and his co-workers have carried out the pioneering work of applying the
SPH method to high strain hydrodynamic problems including hyper velocity impact
(HVI), fracture and fragmentation [51, 113, 257]. The group from the Applied Physics
Division of Los Alamos National Laboratory (LANL) has modeled high velocity impacts
ranging from the very small size (femtogram-scale projectiles that have been accelerated
by a Van de Graaf machine to create craters in various types of targets) to the very large
size (Shoemaker-Levy comet impact with the planet Jupiter). Johnson et al. [50, 313]
proposed a normalized smoothing function (NSF) for axisymmetric problems based on
the condition of uniform strain rate, and have made outstanding contributions in the
application of SPH to impact problems. Attaway, Pampliton and Swegle et al have
worked in coupling the SPH processor with a transient-dynamics FEM code, PRONTO,
in which high-strain areas that typically tangle or break conventional finite element
meshes are resolved using the SPH method [125, 335].
Zhou and Liu have revisited the Taylor-Bar-Impact with the focus on the variation of
results corresponding to the different model parameters which represent varied SPH
implementation in a series of three-dimensional computational simulations, and have
provided informative data on appropriate SPH implementation options [258]. They also
have investigated normal and oblique hypervelocity impacts of a sphere on the thin plate,
producing large deformation of structures. Figure 25 shows the snapshots of three-
dimensional simulation of normal hypervelocity impact of a sphere on the thin plate at t =
0, 10, 15, and 20 s. Figure 26 shows the top view (a) and three dimensional view (b) of
oblique hypervelocity impact of a sphere on the thin plate at t = 20 s with a striking
angle of 30o. They observed that for oblique hypervelocity impacts of a sphere on the
thin plate, the shape of the crater is no longer in circle, and the shape of the debris cloud
changes with different striking angles while the ratio of the debris cloud (length over
width) remains approximately 1.3.

59
 (a) t = 0 s (b) t = 10 s

(c) t = 15 s (d) t = 20 s

Figure 25 Snapshots of three-dimensional simulation of normal hypervelocity impact of a

sphere on the thin plate at t = 0, 10, 15, and 20 s (from [258]).

60
 (a) (b)

Figure 26 Top view (a) and three dimensional view (b) of oblique hypervelocity impact of a
sphere on the thin plate at t = 20 s with a striking angle of 30o (from [258]).

The standard SPH method uses an isotropic smoothing kernel which is characterized
by a scalar smoothing length. One of the problems associated with the standard SPH is
that the isotropic kernel of SPH can be seriously mismatched to the anisotropic volume
change that generally occur in many problems. To closely match the anisotropic volume
changes, an anisotropic smoothing kernel which can be characterized by a matrix (in two-
dimensional space) or tensor (in three-dimensional space) smoothing length can be
efficacious. This leads to the development of the adaptive smoothed particle
hydrodynamics (adaptive SPH or ASPH) in which the smoothing length can be adapted
with the volume changes or other dimension-dependent features. The idea of using
anisotropic kernel with SPH dates back to Bicknell and Gingold [336]. Fulbright et al.
also presented three-dimensional SPH designed to model systems dominated by
deformation along a preferential axis using spheroidal kernels [337]. Later on Shapiro et
al. systematically introduced anisotropic kernels, tensor smoothing and shock-tracking to
SPH to create ASPH [120]. Owen et al presented an alternative formulation of the
ASPH algorithm for evolving anisotropic smoothing kernels [119]. Except for problems
with anisotropic deformations, the concept of elliptical kernel has also been applied to
channel flows with very large length width ratio for saving computational efforts [122].
The numerical results presented in the references further demonstrated that ASPH has
significantly better performance than the standard SPH in terms of resolving ability for a
wide range of problems.
Liu and his co-workers have developed an adaptive smoothed particle hydrodynamics
method (ASPH) for high strain Lagrangian hydrodynamics with material strength [123,
338]. In ASPH, the isotropic kernel in the standard SPH is replaced with an anisotropic
kernel whose axes evolve automatically to follow the mean particle spacing as it varies in
time, space, and direction around each particle. Except for the features inherited from the
standard SPH, ASPH can capture dimension-dependent features such as anisotropic

61
deformations with a more generalized elliptical or ellipsoidal influence domain. A series
of comparative studies show that ASPH has better accuracy than the standard SPH when
being used for high strain hydrodynamic problems with inherent anisotropic deformations.
A typical example is the simulation of an aluminum (Al) cylinder on an Al thin plate,
as shown in Figure 27. In the simulation, the cylinder is of 1.0 cm diameter. The plate is
0.4 cm thick. The plate length is of 10 cm. The particles are all initialized as squares of
0.02 cm in side dimensions. The particles in the cylinder, which is an infinite cylinder in
plane symmetry, are arranged in circumferential rings as this gives a most realistic
representation of the geometry. The particles in the plate are arranged in a rectangular
Cartesian array. There are 500 particles along the length and 20 particles along the
thickness of the plate. There are 1956 particles in the cylinder and 10000 particles in the
plate, for a total of 11956 particles. The cylinder is initially in contact with the center of
the plate. The problems are run with the plate free of constraints. The impact speed of
the cylinder is 6180 m/s. The problems are run to 20 µs after the impact. The
environmental and initial temperature of the cylinder and plate are set to 0 oC. Figure 28
shows the particle distributions obtained by using SPH at 0, 5, 15, and 20 μs
respectively. Figure 29 shows the particle distributions obtained using ASPH at the
same stages. In general, both SPH and ASPH can get agreeable results with the
experimental observations [339]. The symmetry of the problem is well preserved.
Figure 30 shows the close-up view of the particle distributions near the penetrated edge
of the plate obtained using SPH (a) and ASPH (b) at 20 μs. In comparison with the SPH
simulation, the orientation and anisotropy of the deformation of the particles can be
clearly seen in the ASPH simulation.

Figure 27 Initial particle distribution in the vicinity of the contact area for both the
SPH and ASPH simulation of a cylinder impacting on a plate (from [123]).

62
 (a) 0 μs (b) 5 μs

(c) 15 μs (d) 20 μs

Figure 28 Particle distributions obtained using the SPH method for simulating an Al
cylinder penetrating an Al plate at 0, 5, 15, and 20 μs respectively (from [123]).

(a) 0 μs (b) 5 μs

(c) 15 μs (d) 20 μs

Figure 29 Particle distributions obtained using ASPH for simulating an Al cylinder

penetrating an Al plate at 0, 5, 15, and 20 μs respectively (from [123]).

63
 (a) (b)

Figure 30 Close-up view of the particle distributions near the penetrated edge of the
plate obtained using SPH (a) and ASPH (b) at 20 μs (from [123]).

6.2 Detonation, explosion and underwater explosion

The explosion of a high explosive (HE) charge can rapidly convert the original
explosive charge into gaseous products with extremely high pressure through a chemical
reaction process. The high pressure can lead to damages to nearby personnel and
structures. A typical HE explosion consists of the detonation process through the HE at a
constant detonation velocity and the later expansion process of the gaseous products to
the surrounding medium (Figure 31). The detonation process is accompanied by the
propagation of the reactive wave that advances through the explosive with a constant
velocity related to the particular type of explosive concerned. In a steady state detonation
process, the reaction rate is essentially infinite and the chemical equilibrium is attained.
After the completion of the detonation, the detonation-produced explosive gas expands
outwards. This gas expansion generally involves moving material interfaces if the
explosive is surrounded by outside medium or free surfaces if the explosion occurs in the
vacuum.
When the detonation and explosion happen underwater, the explosion physics can be
more complicated. The underwater explosion (UNDEX) [340, 341] produced by the
detonation of a submerged high explosive poses a serious threat to the integrity of nearby
structures. Issues related to the underwater explosion include 1) the detonation process of
the high explosive (HE) charge, 2) the expansion process of the detonation-produced
explosive gas into the surrounding water, and 3) the interaction of underwater shocks
with the nearby structures. In the detonation process, the high explosive is converted into
gaseous products at very high temperature and pressure through a violent chemical
reaction, occurring with extreme rapidity and releasing a great deal of heat. The
propagation of the detonation wave through an explosive is so rapid that the gaseous
products directly behind the wave front are not in pressure equilibrium with the gas
further behind the wave front. As the detonation wave reaches the interface between the
explosive and the surrounding water, a high pressure shock wave of step exponential type

64
is transmitted to propagate through the water, followed by a series of bubble pulsation
associated with the repeating expansion and contraction of the bubble of the explosive
gas. In the entire process of the underwater explosions, some special features such as
large deformations, large inhomogeneities, moving material interfaces, deformable
boundaries, and free surfaces usually exist.

Figure 31 Detonation of a 1D high explosive. The reaction end plane is an interface of

the pressurized high explosive charge and the explosive gas produced in the detonation
process.

Theoretical solutions to the detonation and explosion of high explosive, and

underwater explosion are only limited to some simple cases. Experimental studies need
to resort to dangerous and expensive firing trials, and sometimes certain physical
phenomena related to the explosions cannot be scaled in a practical experimental setup.
Recently, more and more analyses of detonation, explosion and underwater explosions
are based on numerical simulations with the advancement of the computer hardware and
computational techniques [8, 342, 343]. However, numerical simulations of the high
energy phenomena are generally very difficult for the conventional grid based numerical
methods. First, during the detonation process in the explosion, a very thin reaction zone
divides the domain into two inhomogeneous parts and produces large deformations.
Second, in the expansion process, there are free surfaces and moving interfaces involved.
It is even more difficult to simulate underwater explosion, as underwater explosion
phenomena are subject to a number of physical laws and properties, including the
physical conditions at the interface of the explosive gas and the surrounding water. The
surrounding water has such properties as large density that is approximately 1000 times
of the air density, low compressibility, and large sound speed. Owing to the dynamic
properties of the water (especially in the regions surrounding the explosive gas), the
pressures are generally very high and the wave velocities are dependent on the magnitude
of the pressure and the displacement of the water as it progresses. These complications
for waves of finite amplitude are expressed in much more difficult mathematical

65
statements than those which suffice to explain the propagation of small amplitude waves
whose velocities are practically independent on the magnitude of the pressure.
Traditional Lagrangian techniques such as the finite element methods are capable of
capturing the history of the detonation and explosion events associated with each material.
It is, however, difficult to apply practically, since the severely distorted mesh may result
in very inefficient small time step, and may even lead to the breakdown of the
computation. Traditional Eulerian techniques, such as the finite difference methods or
finite volume methods, can well resolve the problem due to the large deformations in the
global motions, but it is very difficult to analyze the details of the flow because of the
lack of history and the smearing of information as the mass moves through the fixed-in-
space Eulerian mesh [255]. Considering the difficulties of grid based numerical models,
the meshfree particle methods can be a good alternate for simulating detonation,
explosion and underwater explosions [124, 135, 142, 246, 247, 251, 254, 255, 344-349].
Liu and his co-workers have conducted a series of original work in extending the SPH
method for simulating explosion phenomena including high explosive detonation,
explosion [247, 350], shaped charge [246], and underwater explosions [142, 254]. Since
the detonation and expansion speed are extremely high, the gaseous products can be
assumed to be inviscid and the explosion process is adiabatic. The Euler equations can
be used to model the explosion process together with a suitable equation of state. The
SPH equations of motion derived in Section 2.5 can thus be used to model high explosive
explosion phenomena, with the viscosity terms in the momentum and energy equations
ignored. For the explosive gas, the standard Jones-Wilkins-Lee (JWL) [351] equation of
state has been employed. These works not only show the feasibility of applying the SPH
method to explosion phenomena, but also identify some important and unresolved issues.
A Gruneisen form of equation of state for water has been used, which is a polynomial
form either in compressed or expanded state [352].
To show the effectiveness of the SPH method in dealing with detonation and
explosion, a benchmark one-dimensional TNT slab denotation and explosion is provided
here, in which a 0.1 m long TNT slab detonates at one end of the TNT slab.
If the solid wall boundary condition is used to prevent material transport from
everywhere, a symmetric setup can be employed to deploy the particles, and thus makes
the detonation of the 0.1 m long slab from one end to the other end equivalent to the
detonation of a 0.2 m long slab from the middle point to both ends. Before detonation,
particles are evenly distributed along the slab. The initial smoothing length is one and a
half times the particle separation. After ignition, a plane detonation wave is produced.
According to the detonation velocity, it takes around 14.4 s to complete the detonation
to the end of the slab. Figure 32 show the pressure profile along the slab at 2 s interval
from 2 to 14 s by using 4000 particles. The dashed line in Figure 32 represents the
experimentally determined C-J detonation pressure, which is, according to the Chapman
and Jouguet’s hypothesis, the pressure at the tangential point of the Hugoniot curve and
the Rayleigh line, and represents the pressure at the equilibrium plane at the trailing edge
of the very thin chemical reaction zone. For this one-dimensional TNT slab detonation
problem, the experimental C-J pressure is 2.1  1010 N/m2. It can be seen from Figure 32

66
that, with the process of the detonation, the detonation pressure converges to the C-J
pressure.

10
x 10
2.5
Experimental PCJ

2
2 s 4 s 6 s 8 s 10 s 12 s 14 s

Pressure (N/m2)
1.5

0.5

0
0 0.02 0.04 0.06 0.08 0.1
Distance along the TNT slab (m)

Figure 32 Pressure profiles along the TNT slab during the detonation process (from [247]).

Figure 33 Velocity transients at 1 and 2 s for the TNT slab detonation-dispersion process
(from [247]).

If the free boundary applies to the end of the detonation, the explosive gas behind the
C-J plane disperses outwards with the forward propagating detonation wave. A
theoretical solution for this one-dimensional TNT detonation-dispersion problem exists.
Figure 33 shows the comparisons of velocity profiles between theoretical values and the
presented SPH results at 1 and 2 s. The presented SPH results are in close agreement
with the theoretical values.
Figure 34 shows the pressure evolution in an underwater explosion process in a
confined square chamber [142]. The initial outward propagating shock wave, reflection

67
wave from the solid wall, explosive gas expansion and later compression can all be seen
from the figure. Right after the detonation, a shock wave is generated in the water and
propagates outwards. At the same time, the rarefaction wave is also produced within the
explosive gas and advances inward. With the advancement of the shock wave through the
water, the gas bubble expands within the surrounding water. The shock wave reaches the
solid wall around the instant of 200 s, and then reflects from the solid wall. This
reflection wave propagates inward and tends to compress the expanding gas. At around
400 s, the gas bubble reaches to the maximum and then reduces. With the continued
process of shock reflection and explosive gas contraction, the gas bubble then will reach
a minimum size, and then expands afterwards. This repeated process of expansion and
contraction maintains for many circles, and will finally reach equilibrium at larger time
step.

68
Figure 34 Pressure distributions in a confined underwater explosion (from [142]).

69
 6.3 Microfluidics and micro drop dynamics

By integrating mechanical elements, sensors, actuators, and electronic components

using micro-fabrication technology, micro-electro-mechanical systems (MEMS) are fast
in response, capable of achieving high spatial resolution, and cost-effective due to the
batch micromachining techniques. Characterization of fluid flows in microfluidic
devices has increasingly becoming a very important topic since the fluidic behavior in
MEMS is very different from what observed in daily life. Flows in microfluidic devices
usually involve small or ignorable inertial force, but dominant viscous, electro-kinetic
and surface effects especially when the surface-to-volume ratio increases [353].
Analytical or semi-analytical solutions for microfluidics are generally limited to a very
few simple cases, whereas experimental studies are usually expensive. Numerical
simulation of flows in microfluidic devices, as an effective alternate, has been attracting
more and more researchers. However, simulation of microfluidic devices is not easy due
to the involved complex features including movable boundaries (free surfaces and
moving interfaces), large surface-to-volume ratio, and phenomena due to micro scale
physics. Numerical studies with reliable models are needed to develop a better
understanding of the temporal and spatial dynamics of multiphase flows in microfluidic
devices.
On the other hand, drop formation and break-up in micro/nano scales are
fundamentally important to diverse practical engineering applications such as ink-jet
printing, DNA and protein micro-/nano-arraying, and fabrication of particles and capsules
for controlled release of medicines. Numerical studies provide an effective tool to
improve better understanding of the inherent physical dynamics of drop formation and
breakup. Computational models for drop formation and breakup in micro/nano scales
must be able to handle movable boundaries such as free surfaces and moving interfaces,
large density ratios, and large viscosity ratios. These requirements together with micro
scale phenomena and possible complex boundaries (fluid-fluid-solid contact line
dynamics and fluid-fluid interface dynamics) in microfluidic devices present severe
challenges to conventional Eulerian-grid based numerical methods such as FDM and
FVM which require special algorithms to treat and track the interfaces. Algorithms based
on Lagrangian-grid based methods such as FEM have been shown to agree quantitatively
with experimental measurements, but they are only capable of modeling the dynamics of
formation of a single drop or the dynamics until the occurrence of the first singularity.
A number of meso-scale methods have been developed for simulating micro- and
nano-fluidics such as the direct simulation Monte Carlo (DSMC) for rarefied gas flows
[354-358], and dissipative particle dynamics for complex fluid flows [40, 41, 117, 359].
Hoover and his co-workers have first noted the similarity between the particle methods of
molecular dynamics and smoothed particle hydrodynamics, and described the inherent
links between them [360-362]. Espanol and his colleagues, when studying the dissipative
particle dynamics method for meso-scale applications, proposed a fluid particle dynamics
model, which is actually a synthesis of dissipative particle dynamics and smoothed
particle dynamics [95, 363]. Later on, they invented a smoothed dissipative particle
dynamics for micro- or meso-scale applications, by introducing a fluctuation term into
the conventional smoothed particle hydrodynamics [364].

70
 There have been a lot of literatures addressing the applications of SPH method to
simulating microfluidics and micro drop dynamics. Nugent and Posch described an
approach to modeling liquid drops and surface tension for a van der Waals fluid [235].
The cohesive pressure in the equation of state for the van der Waals fluid actually acts an
attractive force between SPH particles. Melean and his co-workers investigated the
formation of micro drops using SPH method [154]. It was reported that the tensile
instability also exists when using SPH for simulating viscous liquid drops and using an
artificial stress proposed by Monaghan et al. [144] can help greatly to remove the tensile
instability. Later on The group extended their work to the simulation of coalescence of
colliding van der Waals liquid drops [239] and oscillation of viscous drops [238]. Liu et
al. applied the SPH method to multiphase fluid flow in micro channels with applications
to flip-chip underfill encapsulation process with both isotropic and anisotropic smoothing
kernels [122, 365]. Li et al. proposed an SPH model for simulating droplet collision and
coalescence [366]. Zhang et al. developed an SPH model for free surface and
solidification problems, which is also applicable to microfluidics and micro drop
dynamics such as droplet spreading, splashing and solidification, substrate melting and
deformation [242-244]. A revised surface tension model was developed for macro-scale
particle methods including SPH by Zhou et al. [245]. Fang et al. also developed an
improved SPH model for the simulations of droplet spreading and solidification [237].
Tartakovsky, Meakin and their co-workers have conducted a series of excellent work in
applying the SPH method to the modeling of surface tension and contact angle [234],
miscible flow in three-dimensional fractures and the two-dimensional Rayleigh Taylor
instability [367], unsaturated flow in complex fractures [116], reactive transport and
precipitation [368], mixing-induced precipitation [369] and non-aqueous phase liquid
flow and dissolution [370].
There are basically two approaches in modeling the multiphase fluid flow in micro
fluidic devices, and multiphase drop dynamics. The first approach is to use the
continuum surface force (CSF) model proposed by Brackbill et al. [371], and introduced
the surface tension force into the momentum equation as presented in Section 5. This
approach is straightforward, and real physical parameters are used in the simulation. It is
important to note that the surface tension parameters in this approach are user-input
parameters, and surface curvature needs to be calculated to get the surface tension force.
Some researchers have used this approach in SPH method to model multiphase fluid flow
[122, 372].
With this CSF model, the surface tension is added to the momentum equation as an
external source force. The surface tension force F can be formulated as follows
F  2 k ( x)V V , (109)

where  is the surface tension coefficient, k is the surface curvature. V is the volume of a
fluid cell. Let n be the surface normal which can be computed from the gradient of the
fluid cell as n  V , the surface curvature k is defined as

71
 1  n  
k    n     n     n   , (110)
n  n 
 

where n is the unit surface normal defined as

n
n  (111)
n

For fluid cells near a solid wall, a wall adhesion model is applied to adjust the surface
normal through using the contact angle  that a fluid is assumed to make with the wall.
The dynamic contact angle can be different for different contacting points. However, in
micro channel flows with a small buck velocity, the dynamic contact angle can be
assumed to be constant along the flow direction as the starting contact angle. The unit

normal n is adjusted as follows for fluid cells close to the wall

n  n w cos   n t sin  (112)

where n w and n t are the unit vectors of the surface normal ( n  V ) normal and
tangential to the wall, respectively. Thereby the unit surface normal n for fluid cells
near the wall is adjusted using the contact angle. Whereas the unit surface normal n for
fluid cells one cell away from the wall is normally calculated. After determination of the
unit surface normal, the local surface curvature and therefore the surface tension can be
calculated.

Figure 35 Snapshots of the moving flow leading edge in an SPH simulation of a

multiphase fluid flow in a micro channel using the CSF model (from [122]).

72
 Figure 35 gives an example of CSF model in simulating micro channel fluid flow. It
is an underfill encapsulation process. The encapsulant is used to fill the space between
the solder joints under the chip and the encapsulation process is either driven by a
capillary action or by a pressurized injection [122].
Another approach is to introduce an inter-particle interaction force (IIF) implicitly in
the SPH equations, and can be termed as IIF model. This attractive force between every
pair of SPH particles contribute to the surface tension. Considering the SPH equations
presented in Section 2.5, an equation of state can be used to close the equation system.
An equation of state describes the relationship of the pressure p, density  and the
internal energy per unit mass, e, respectively. For example, the van der Waals (vdW)
equation of state can be used to model the behavior of the fluid under consideration. The
van der Waals equation of state was derived from statistical mechanics as the mean-field
limit for the free energy density of a system of hard particles with a superimposed long-
range, attractive pair potential. It is realistic to display a gas-to-liquid phase transition
similar to that of a real fluid. The van der Waals equation of state can written as

 kT
p  a 2 , (113)
1  b

and

e  kT   a . (114)

In the above two equations, k  k B T , where k B is the Boltzmann’s constant. T is the

2
system temperature. a  a m , and b  b m , where a and b are the parameter
describing a van der Waals fluid. a controls the strength of the attractive force, and b
relates to the size of the particle.
The second part in equation (113) describes the cohesion between particles. Using
SPH particle approximation for pressure (equation (28)), and considering this cohesive
pressure part separately, we can get

Dvi N W
 2a  m j ij . (115)
Dt j 1 xi

Similarly substituting the cohesive pressure into the energy (equation (28)), it is possible
to get its analogous contribution to energy

Dei N Wij
 a  m j vij . (116)
Dt j 1 xi

It is clear that the attractive, long-range, inter-atomic vdW force can be transformed
into similarly attractive forces between SPH particles. Equation (115) describes a
volume force on continuum scale to account for the formation of co-existing liquid-gas

73
phases. Comparing with the above-mentioned approach in using CSF model in
calculating surface force, this approach does not need to locate the surface, and then to
calculate the local surface curvature. Also, the surface tension is not user-input
parameters, while it is implicitly obtained from inter-particle interactions. If the particles
are distributed regularly (please refer to Section 4.2 on particle consistency), the force
obtained from equation (115) vanishes for interior particles in both liquid and gas phases.
While for boundary particles, e.g., particles near the gas-liquid interface, since the inter-
particle interaction force between different phases are generally different, a surface force
is produced, which is basically perpendicular to the surface, pointing towards the dense
phase. Nugent and Posch identified that to obtain acceptable results, a large influencing
area (e.g., two times the smoothing length for approximating other field variables and the
first part in equation (113)) is necessary to conduct the particle approximations in
equations (115) and (116).
Figure 36 shows an example of IIF model for the formation of a liquid drop. The
initial square-shaped liquid (Figure 36a) gradually becomes rounded at the corner due to
the inter-particle interaction force, and finally forms a circular liquid drop.

(a) (b)

Figure 36 SPH simulation of the formation of a liquid drop using the IIF model. (a)
initial stage, (b) final stage of liquid drop.

Except for the inter-particle interaction force implicitly calculated from the cohesive
pressure in the van der Waals equation of state, Tartakovsky et al. added another inter-
particle interaction force, which can be written as
Fij  sij cos(1.5 / kh)rij , (117)

where sij is an interaction coefficient. Figure 37 shows the shape of the inter-particle
interaction force expressed in the above equation ( cos(1.5 / kh) only, where k  2 ). It
is clear that this inter-particle force is repulsive at short range, and attractive at long-
distance. Similar to inter-particle force calculated from the cohesive pressure in the van
der Waals equation of state, this inter-particle also describes a volume force on

74
continuum scale and can account for the formation of co-existing liquid-gas phases. The
force obtained from equation (117) also vanishes for interior particles in both liquid and
gas phases, with small fluctuation around the overall direction of the macroscopic density
gradient. For boundary particles, e.g., particles near the gas-liquid interface, a surface
force is produced, pointing towards the denser phase. The IIF approach is comparatively
simply and straightforward since it does not need to calculate the surface curvature,
which is not an easy task for particle methods like SPH. One problem is that as the IIF
model implicitly calculates the surface tension force with parameters from atomistic level,
it needs parameter calibration, which relates the physical parameters from atomistic level
to continuum level.

0.5

-0.5

-1

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

x

Figure 37 The shape ( cos(1.5 / kh) ) of the inter-particle interaction force expressed in
equation (117) with k  2 .

6.4 Ocean and coastal hydrodynamics and offshore engineering

Flow phenomena in ocean and coastal hydrodynamics and offshore engineering are
significantly important as they can greatly influence the nearly personnel and structures.
The flow phenomena include
 wave dynamics (waver generation, wave breaking, and wave interaction with
other structures),
 dam breaking,
 water filling and water discharge (to and from a water tank or reservoir),
 shallow water flows,
 entry of water, sloshing phenomena with fluid-solid interaction, and
 different other problems.
These phenomena involve special features, which make it difficult for numerical
simulation. For example, water waves can propagate shoreward where they undergo

75
changes induced by the near-shore topography and increase in height. Upon reaching the
shoreline, they can break into pieces, and travel inland for large distances with potential
damage of property and loss of life. Experimental setups for fluid flow in coast
hydrodynamics and offshore engineering are expensive and only limited to laboratory
applications. Numerical simulation has thus become a great tool to predicting fluid flow
in ocean and coast hydrodynamics and offshore engineering.
However, numerical simulation of fluid flow in these related areas is a formidable
task as it involves not only complex geometries and free surfaces, but also fluid-solid
interaction as well as other complex physics in a comparably very large scale. In many
circumstances, violent fluid–structure interactions lead to air entrapment and multi-phase
flows, where the dynamics of the entrapped air at the impact may play a dominant role
during the process and contribute to the high pressure maxima and pressure oscillations.
Though conventional grid based methods like FDM, FVM and FEM have achieved
greatly in simulating fluid flow in coast hydrodynamics and offshore engineering, there is
still a long way to go for practical engineering applications.
Smoothed particle hydrodynamics, due to its meshfree, Lagrangian and particle nature,
has been attracting more and more researchers in coast hydrodynamics and offshore
engineering. From the very early simulation of a simple dam break problem [111], there
have been a lot of literatures addressing the applications of SPH method in related areas.
Shao and his colleagues simulated near-shore solitary wave mechanics by an
incompressible SPH method [373]. They later extended the work to simulation of wave
breaking and overtopping with turbulence modeling [374], plunging waves using a 2-D
sub-particle scale (SPS) turbulence model [375], solitary wave interaction with a curtain-
type breakwater [376], and water entry of a free-falling object [189].
Frank and Reich addressed the conservation properties of SPH applied to shallow
water equation [377]. Ata and Soulaimani proposed a stabilized SPH method for inviscid
shallow water flows [378]. Rodriguez-Paz and Bonet also developed a corrected smooth
particle hydrodynamics formulation of the shallow-water equations [379].
Investigation of water wave including water wave generation, water-structure
interaction, and water wave breaking has been a very attractive application of the SPH
method. Gomez-Gesteira and Dalrymple had investigated wave impact on a tall structure
using a three-dimensional SPH method [179]. The research group also investigated water
waves and waves breaking using the SPH method [174, 188]. Idelsohn and his co-
workers applied the particle finite element method to solve incompressible flows with
free-surfaces and breaking waves [380]. Gotoh et al. developed an SPH-LES model for
numerical investigation of wave interaction with partially immersed breakwater [177].
They also simulated the coupled motion of progressive wave and floating curtain wall by
using the developed SPH-LES model [176]. Gotoh and Sakai also addressed some key
issues in the particle method for computation of wave breaking [182]. Recently, the
research team developed a corrected incompressible SPH method for accurate water-
surface tracking in breaking waves [184]. Crespo et al. presented an example of 3D SPH
simulation of large waves mitigation with a dike [175]. Qiu studied the water waves
generated by landslide [187]. Yim investigated the water wave generation by a vertical

76
plunger using RANS and SPH models [192]. Cleary and Prakash discussed the
feasibility of using SPH method for modeling tsunami effects [204, 265].
There have some references in addressing the entry of water, e.g., air cushion effects
in a wedge water entry [381], waves produced by a falling mass into a reservoir [382],
water entry of a free-falling object [189], and wedge water entries [383].
Simulation of sloshing problems using SPH is a promising research direction.
Iglesias et al. simulated the anti-roll tanks and sloshing type problems [195]. Rhee and
Engineer studied liquid tank sloshing with Reynolds-averaged Navier-Stokes [198].
Souto-Iglesias et al. assessed the liquid moment amplitude in sloshing type problems with
smooth particle hydrodynamics [199]. Anghileri investigated the fluid-structure
interaction of water filled tanks during the impact with the ground [284]. Delorme et al.
simulated the sloshing loads in LNG tankers with SPH [384].
Figure 38 gives an example of the SPH simulation of the dam-break flow and impact
again a vertical wall at (a) single phase (water only) with free surface, and (b) two phases
(water and air) with moving interface. The flow pattern of flow along the surge front, the
impact of fluid against the right hand side vertical wall and the entrapment of air bubble
in water are in good agreement with the results from other sources [385].

(a) (b)

Figure 38 SPH simulation of the dam-break flow and impact against a vertical wall at (a)
single phase (water only) with free surface, and (b) two phases (water and air) with
moving interface (from [386]).

6.5 Environmental and geophysical flows

Environmental and geophysical flows are fundamentally important for human life and
economic development. In large scale, there are many typical environmental flow
problems including
 flood due to dam break, or overflow of river and reservoir,
 landslide, mudslide, mud-rock flow triggered by heavy rain infiltration or even
earth quake,
 sand/dust or dust storm,
 polluted liquid (e.g. oil ) and air transport, and
 many others.

77
Unwanted environmental flows can be hazardous to nearby people and structures. In
general, these large scale environmental flows are difficult to be accurately simulated and
precisely predicted. On one hand, these large scale environmental flows involve complex
geometries and arbitrarily moving interfaces. It is a formidable task to identify the origin
of the flows. Therefore, it is hard to determine the initial and boundary condition for
numerical simulations. On the other hand, large scale environmental flows are associated
with complicated multiple fluid phases and multiple physics with solid-liquid-air phase at
different scales. Theses also add difficulties to the numerical simulations for grid based
numerical models.

It is very appealing to use SPH to model such large-scale environmental flow

problems due to its meshfree, Lagrangian and particle nature. Actually in section 7.4,
some large-scale environmental flow problems such as dam breaking and tsunami effects
have been discussed. They are also other references addressing the SPH applications in
large-scale environmental flow problems. Ghazali and Kamsin presented a real time
simulation and modeling of flood hazard [205]. Shen et al. conducted SPH simulation of
river ice dynamics [232]. Kipfer and Westermann also investigated the realistic and
interactive flows in river hydrodynamics [206]. Xu and Shen studied the fluid-structure
interaction of hydrodynamic damper during the rush into the water channel [387].

Landslide and mudslide as well as mud-rock flows can be simulated using the SPH
method. Cleary and his co-workers discussed the feasibility of using SPH method for
modeling dam break, tsunami, landslide and volcano flows [204, 265]. The SPH method
is coupled with discrete element method (DEM) for modeling solid-fluid interaction.
Ataie-Ashtiani and Shobeyri simulated landslide impulsive waves by using
incompressible smoothed particle hydrodynamics [388]. Qiu presented a two-
dimensional SPH simulations of landslide-generated water waves [187]. McDougall and
Hungr developed a numerical model for the analysis of rapid landslide motion across
three-dimensional terrain [215]. Pastor et al. proposed a depth-integrated, coupled SPH
model for flow-like landslides and related phenomena [219].

There are more references with applications to other large-scale environmental flow
problems. Bui et al. developed an SPH model for large deformation and failure flows of
geo-material using elastic-plastic soil constitutive model [208]. They also presented a
numerical simulation of soil-water interaction using the SPH method [209]. Laigle and
his colleagues developed an SPH-based numerical investigation of mudflow and other
complex fluid flow interactions with structures [213].

Small scale environmental and geophysical flows are also very important, but are
usually difficult to simulate because of the associated multiple fluid phases and multiple
physics, as well as the existence of complex geometries and arbitrarily moving interfaces.
For example, fluid motion in the vadose zone is very important for groundwater recharge,
fluid motion and contaminant transport. Flow through fractures and fractured porous
media can lead to exceptionally rapid movement of liquids and associated contaminants
[98, 389, 390]. The physics of fluid flows in unsaturated fractures and porous media is
still poorly understood due to the complexity of multiple phase flow dynamics.
Experimental studies of fluid flow in fractures and fractured porous media are limited,

78
and in computer simulations it is usually difficult to take into account the fracture surface
properties and microscopic roughness. Predictive numerical models can be divided into
two general classes: volume-averaged continuum models (such as those based on
Richard’s equation) and discrete mechanistic models. Knowledge of the physical
properties of the fluids and the geometry of the fracture apertures is required in both
classes. Volume-averaged continuum models are more suitable for large-scale systems,
and they usually involve the representation of fractures as porous media with porosity
and permeability parameters adjusted to mimic flow within fractures. However, volume-
averaged continuum models are unable to describe the details of flow dynamics in
fractures, they do not reproduce the spatio-temporal complexity of multiphase fluid flow
in fractures, and they often fail to predict the rapid fluid motion and contaminant
transport observed in the fractured vadose zone. Small-scale studies with discrete
mechanistic models are needed to develop a better understanding of the temporal and
spatial dynamics of fracture flows. However, the complexity of fracture flow dynamics
makes it difficult to develop successful numerical models for fluid flows in fracture
networks. A broadly applicable model must be able to simulate a variety of phenomena
including film flow with free surfaces, stable rivulets, snapping rivulets, fluid
fragmentation and coalescence (including coalescence/fragmentation cascades), droplet
migration and the formation of isolated single-phase islands trapped due to aperture
variability.
Realistic mechanistic models for multiphase fluid flows in fracture and fractured
porous media must be able to handle moving interfaces, large density ratios (e.g.,
1000:1 for water and air), and large viscosity ratios (e.g., 100:1 for water and air).
These requirements combined with the complex geometries of natural fractures present
severe challenges to mechanistic models. Grid based numerical methods such as finite
difference methods, finite volume methods and Eulerian finite element methods require
special algorithms to treat and track the interface between different phases. However,
continuum grid based numerical models usually do not take account of the detailed void
and obstacle geometries, fluid-fluid interface dynamics within pores and complex fluid-
fluid-solid contact line dynamics. They rely on constitutive equations that describe the
coarse-grained behavior and can, at least in principle, be derived from the results of pore
scale simulations or experiments. Therefore, small-scale simulations with mechanistic
models are needed to develop a better understanding of the temporal and spatial
dynamics of multiphase flow through porous media. Pore-scale flows have been studied
extensively using grid based methods including finite difference method [391], finite
volume method [392, 393], and finite element method [394]. However, due to the
difficulties associated with geometrically complex boundaries, fluid-fluid-solid contact
line dynamics, and fluid-fluid interface dynamics, it is difficult to apply conventional grid
based multiphase simulation methods based on computational fluid dynamics (CFD)
coupled with interface tracking algorithms [371, 395-397] to pore-scale multiphase flow
modeling.
The SPH manhood has been recently modified for much smaller scale, typically low
Reynolds number, applications The performance of the SPH method was demonstrated
for two-dimensional single-phase flows through idealized, pore-scale porous media
composed of spatially periodic square and hexagonal arrays of cylinders [201, 398], and

79
two-phase (miscible and immiscible) flows through pore-scale fractures and fracture
junctions [116, 367]. One obvious advantage of SPH over conventional grid based
methods is that SPH does not require explicit and complicated interface tracking
algorithms, and thus there is no need to explicitly track the material interfaces, and
processes such as fluid fragmentation and coalescence can be handled without difficulty.
SPH also does not require contact angle models since contact angles can be naturally
calculated from the shape of the moving particle distributions, and can vary spatially and
temporally, depending on the dynamic balance of viscous, capillary and gravitational
forces. Figure 39 illustrates the snapshots of SPH simulation of multiphase fluid motion
in a fracture at 4 representative stages. Water was injected into the top entrance of the
fracture using a syringe pump and drained out through the bottom of the fracture.

Figure 39 Snapshots of SPH simulation of multiphase fluid motion in a fracture at 4

representative stages.

7 Summary

Smoothed particle hydrodynamics (SPH) is a meshfree particle method, which not

only uses particles as the computational frame for interpolation or differencing, but also
uses the particles to carry the material properties. In this survey, the smoothed particle
hydrodynamics method, and its recent development in numerical algorithms and
applications have been reviewed. In methodology and numerical approximation, the
basic concepts of kernel and particle approximations as well as different techniques for
developing SPH formulations have been addressed. Different smoothing kernel functions
have been reviewed with constructing conditions provided. The emphasis is on the
consistency problem, which traditionally limits the accuracy of the SPH method. A
number of consistency restoring approaches have been surveyed. An SPH formulation
for discontinuity, a general approach for restoring particle consistency, and a finite
particle method are described. Several important numerical topics have been investigated,

80
including 1) solid boundary treatment, 2) representation of solid obstacles, 3) material
interface treatment and 4) tensile instability. In applications, different applications of the
SPH method have been reviewed, with emphasis on 1) high strain hydrodynamics with
material strength, 2) high explosive detonation and explosions, and underwater explosion,
3) microfluidics and micro drop dynamics, 4) coast hydrodynamics and offshore
engineering, and 5) environmental and geophysical flows.
The last decades have witnessed the great success of SPH developments in
methodology and applications, and there are still many tasks and challenges remaining.
To achieve a reliable solution, the computational accuracy, consistency, efficiency,
stability and convergence need to be incorporated into good SPH algorithms.
In general, SPH has great potential in many problems in engineering and science. It
has salient advantages over traditional grid based numerical models in treating large
deformation, tracking free surfaces, moving interfaces, and deformable boundaries,
resolving moving discontinuities such as cracks and shock waves. The attraction of the
SPH method has been showcased in diversified applications, as reviewed in Section 7.
One obvious advantage is that the SPH method provides a feasible physical model for
non-continuum, as it has some same features as the classic molecular dynamics method
and the dissipative particle dynamics method. Therefore it would be very attractive to
apply the SPH method to simulating problems where the main concern of the object is a
set of discrete physical particles rather than a continuum, e.g., the interaction of stars in
astrophysics, movement of millions of atoms in an equilibrium or non-equilibrium state,
dynamic behavior of protein molecules, and environmental flows with solid and fluid
particles such as landslide and mudslide. There is no need for discretization to begin with
for such situations. Also since SPH shares many similarities with other particle methods
such as MD and DPD, it is natural to couple these three methods, from molecular
dynamics (at nano- and micro- scales), to dissipative particle dynamics (at meso-scales),
and then to smoothed particle hydrodynamics (at macro-scales) for simulations with
multiple scale physics. It is attractive to develop reliable models of length scale coupling
for problems with multiple physics and multiple scales, such as ink-jet printing, DNA and
protein micro-/nano-arraying, and fabrication of particles and capsules for controlled
release of medicines.

References

1.Chung TJ (2002) Computational fluid dynamics. Cambridge University Press.

2.Anderson JD (2002) Computational fluid dynamics: the basics with applications. McGraw Hill, New
York.
3.Zienkiewicz OC, Taylor RL (2000) The finite element method. Butterworth-Heinemann.
4.Liu GR (2002) Meshfree methods: moving beyond the finite element method. CRC Press.
5.Hirsch C (1988) Numerical computation of internal & external flows: fundamentals of numerical
discretization. John Wiley & Sons, Inc. New York, NY, USA.
6.Liu GR, Liu MB (2003) Smoothed particle hydrodynamics: a meshfree particle method. World Scientific.
7.Liu GR, Gu YT (2005) An Introduction to meshfree methods and their Programming. Springer Dordrecht,
2005. 479.
8.Zhang SZ (1976) Detonation and its applications. Press of National Defense Industry, Beijing.
9.Zukas JA (1990) High velocity impact dynamics. Wiley&Sons, New York.

81
10.Hockney RW, Eastwood JW (1988) Computer simulation using particles. Institute of Physics Publishing.
11.Allen MP, Tildesley DJ (1987) Computer Simulation of Liquids. Oxford University Press.
12.Li S, Liu WK (2002) Meshfree and particle methods and their applications. Appl Mech Rev 55(1):1-34.
13.Belytschko T, Krongauz Y, Organ D, Fleming M, Krysl P (1996) Meshless methods: An overview and
recent developments. Comput Method Appl M 139(1-4):3-47.
14.Idelsohn SR, Onate E (2006) To mesh or not to mesh? That is the question. Comput Method Appl M
195(37-40):4681-4696.
15.Nguyen VP, Rabczuk T, Bordas S, Duflot M (2008) Meshless methods: A review and computer
implementation aspects. Math Comput Sim 79(3):763-813.
16.Liu GR (2008) A generalized Gradient smoothing technique and the smoothed bilinear form for
Galerkin formulation of a wide class of computational methods. Int J Comput Meth 5(2):199-236.
17.Liu GR (2009) A G space theory and weakened weak (W2) form for a unified formulation of
compatible and incompatible methods, Part I: Theory and Part II: Applications to solid mechanics
problems. Int J Numer Meth Eng (In press).
18.Liu GR (2009) On the G space theory. Int J Comput Meth (In press).
19.Liu GR, Nguyen-Thoi T, H. N-X, Lam KY (2008) A node-based smoothed finite element method (NS-
FEM) for upper bound solutions to solid mechanics problems. Comput Struct (In press).
20.Liu GR, Zhang GY (2008) Upper bound solution to elasticity problems: A unique property of the
linearly conforming point interpolation method (LC-PIM). Int J Numer Meth Eng 74(1128-1161).
21.Liu GR, Nguyen-Thoi T, Lam KY (2009) An edge-based smoothed finite element method (ES-FEM)
for static, free and forced vibration analyses in solids. J Sound Vib 320(1100-1130).
22.Zhang GY, Liu GR, Nguyen TT, Song CX, Han X, Zhong ZH, Li GY (2007) The upper bound
property for solid mechanics of the linearly conforming radial point interpolation method (LC-
RPIM). Int J Comput Meth 4(3):521-541.
23.Liu GR, Zhang GY (2009) A normed G space and weakened weak (W2) formulation of a cell-based
smoothed point interpolation method. Int J Comput Meth 6(1):147-179.
24.Liu GR, Zhang GY (2008) Edge-based Smoothed Point Interpolation Methods. Int J Comput Meth
5(4):621-646.
25.Liu GY, Zhang GY (2009) A novel scheme of strain-constructed point interpolation method for static
and dynamic mechanics problems. Int J Appl Mech 1(1):233-258.
26.Liu GR, Xu X, Zhang GY, Nguyen-Thoi T (2009) A superconvergent point interpolation method (SC-
PIM) with piecewise linear strain field using triangular mesh. Int J Numer Meth Eng 77:1439-
1467.
27.Liu GR, Xu X, Zhang GY, Gu YT (2009) An extended Galerkin weak form and a point interpolation
method with continuous strain field and superconvergence using triangular mesh. Comput Mech
43:651-673.
28.Nguyen-Thoi T, Liu GR, Lam KY, Zhang GY (2009) A face-based smoothed finite element method
(FS-FEM) for 3D linear and nonlinear solid mechanics problems using 4-node tetrahedral
elements. Int J Numer Meth Eng 78:324-353.
29.Liu GR, Nguyen-Thoi T, Lam KY (2008) A novel Alpha finite element method (FEM) for exact
solution to mechanics problems using triangular and tetrahedral elements. Comput Method Appl
M 197:3883-3897.
30.Liu GR, Nguyen-Xuan H, Nguyen TT, Xu X (2009) A novel weak form and a superconvergent alpha
finite element method for mechanics problems using triangular meshes. J Comput Phys
228(11):3911-4302.
31.Liu GR, Zhang J, Lam KY (2008) A gradient smoothing method (GSM) with directional correction for
solid mechanics problems. Comput Mech 41:457-472.
32.Liu GR, Xu XG (2008) A gradient smoothing method (GSM) for fluid dynamics problems. Int J Numer
Meth Fl 58:1101-1133.
33.Xu XG, Liu GR (2008) An adaptive gradient smoothing method (GSM) for fluid dynamics problems.
Int J Numer Meth Fl (In press).
34.Xu XG, Liu GR, Lee KH (2009) Application of gradient smoothing method (GSM) for steady and
unsteady incompressible flow problems using irregular triangles. Int J Numer Meth Fl (Submitted).

82
35.Lucy LB (1977) A numerical approach to the testing of the fission hypothesis. Astron J 82(12):1013-
1024.
36.Gingold RA, Monaghan JJ (1977) Smoothed particle hydrodynamics-Theory and application to non-
spherical stars. Mon Not R Astron Soc 181:375-389.
37.Liu MB, Liu GR, Lam KY (2003) Constructing smoothing functions in smoothed particle
hydrodynamics with applications. J Comput Appl Math 155(2):263-284.
38.Fulk DA, Quinn DW (1996) An analysis of 1-D smoothed particle hydrodynamics kernels. J Comput
Phys 126(1):165-180.
39.Frenkel D, Smit B (2002) Understanding molecular simulation: from algorithms to applications.
Academic Press.
40.Hoogerbrugge PJ, Koelman J (1992) Simulating microscopic hydrodynamic phenomena with dissipative
particle dynamics. Europhys Lett 19:155.
41.Groot RD (1997) Dissipative particle dynamics: Bridging the gap between atomistic and mesoscopic
simulation. J Chem Phys 107(11):4423.
42.Gingold RA, Monaghan JJ (1982) Kernel estimates as a basis for general particle method in
hydrodynamics. J Comput Phys 46:429-453.
43.Hu XY, Adams NA (2006) Angular-momentum conservative smoothed particle dynamics for
incompressible viscous flows. Phys Fluids 18:101702. 101702.
44.Swegle JW, Hicks DL, Attaway SW (1995) Smoothed particle hydrodynamics stability analysis. J
Comput Phys 116(1):123-134.
45.Morris JP (1996) Analysis of Smoothed Particle Hydrodynamics with Applications. Monash University.
46.Monaghan JJ (1982) Why particle methods work. Siam J Sci Stat Comp 3(4):422-433.
47.Monaghan JJ (1985) Particle methods for hydrodynamics. Comput Phys Rep 3:71-124.
48.Monaghan JJ (1992) Smooth particle hydrodynamics. Annu Rev Astron Astr 30:543-574.
49.Johnson GR (1996) Artificial viscosity effects for SPH impact computations. Int J Impact Eng
18(5):477-488.
50.Johnson GR, Beissel SR (1996) Normalized smoothing functions for SPH impact computations. Int J
Numer Meth Eng 39(16):2725-2741.
51.Randles PW, Libersky LD (1996) Smoothed particle hydrodynamics: Some recent improvements and
applications. Comput Method Appl M 139(1):375-408.
52.Chen JK, Beraun JE, Carney TC (1999) A corrective smoothed particle method for boundary value
problems in heat conduction. Int J Numer Meth Eng 46:231-252.
53.Chen JK, Beraun JE (2000) A generalized smoothed particle hydrodynamics method for nonlinear
dynamic problems. Comput Method Appl M 190:225-239.
54.Liu MB, Liu GR, Lam KY (2003) A one-dimensional meshfree particle formulation for simulating
shock waves. Shock Waves 13(3):201-211.
55.Liu MB, Liu GR (2006) Restoring particle consistency in smoothed particle hydrodynamics. Appl
Numer Math 56(1):19-36.
56.Liu MB, Xie WP, Liu GR (2005) Modeling incompressible flows using a finite particle method. Appl
Math Model 29(12):1252-1270.
57.Batra RC, Zhang GM (2004) Analysis of adiabatic shear bands in elasto-thermo-viscoplastic materials
by modified smoothed-particle hydrodynamics (MSPH) method. J Comput Phys 201(1):172-190.
58.Fang JN, Owens RG, Tacher L, Parriaux A (2006) A numerical study of the SPH method for
simulating transient viscoelastic free surface flows. J Non-Newton Fluid 139(1-2):68-84.
59.Fang JN, Parriaux A (2008) A regularized Lagrangian finite point method for the simulation of
incompressible viscous flows. J Comput Phys 227(20):8894-8908.
60.Fang JN, Parriaux A, Rentschler M, Ancey C (2009) Improved SPH methods for simulating free
surface flows of viscous fluids. Appl Numer Math 59(2):251-271.
61.Dyka CT, Ingel RP (1995) An approach for tension instability in smoothed particle hydrodynamics
(SPH). Comput Struct 57(4):573-580.
62.Dyka CT, Randles PW, Ingel RP (1997) Stress points for tension instability in SPH. Int J Numer Meth
Eng 40(13):2325-2341.
63.Randles PW, Libersky LD (2000) Normalized SPH with stress points. Int J Numer Meth Eng
48(10):1445-1462.

83
64.Vignjevic R, Campbell J, Libersky L (2000) A treatment of zero-energy modes in the smoothed particle
hydrodynamics method. Comput Method Appl M 184(1):67-85.
65.Dilts GA (1999) Moving-Least-Squares-particle hydrodynamics. I: Consistency and stability. Int J
Numer Meth Eng 44:1115-1155.
66.Dilts GA (2000) Moving least square particle hydrodynamics ii: conservation and boundaries. Int J
Numer Meth Eng 48:1503-1524.
67.Bonet J, Kulasegaram S (2000) Correction and stabilization of smooth particle hydrodynamics methods
with applications in metal forming simulations. Int J Numer Meth Eng 47(6):1189-1214.
68.Chen JS, Pan C, Wu CT, Liu WK (1996) Reproducing kernel particle methods for large deformation
analysis of nonlinear structures. Comput. Methods Appl. Mech. Engrg 139:195-227.
69.Liu WK, Chen Y, Jun S, Chen JS, Belytschko T, Pan C, Uras RA, Chang CT (1996) Overview and
applications of the reproducing kernel particle methods. Arxiv Comput Methods Engrg 3(1):3-80.
70.Rabczuk T, Belvtschko T, Xiao SP (2004) Stable particle methods based on Lagrangian kernels.
Comput Method Appl M 193(12):1035-1063.
71.Belytschko T, Krongauz Y, Dolbow J, Gerlach C (1998) On the completeness of the meshfree particle
methods. Int J Numer Meth Eng 43(5):785-819.
72.Zhang GM, Batra RC (2004) Modified smoothed particle hydrodynamics method and its application to
transient problems. Comput Mech 34(2):137-146.
73.Belytschko T, Guo Y, Liu WK, Xiao SP (2000) A unified stability analysis of meshless particle
methods. Int J Numer Meth Eng 48:1359-1400.
74.Fulk DA (1994) A numerical analysis of smoothed particle hydrodynamics. Air Force Institute of
Technology.
75.Monaghan JJ (2005) Smoothed particle hydrodynamics. Rep Prog Phys 68(8):1703-1759.
76.Hernquist L (1993) Some cautionary remarks about smoothed particle hydrodynamics. Astrophys J
404(2):717-722.
77.Century Dynamics Incorporated (1997) AUTODYN release notes Version 3.1. AUTODYN™
Interactive Non-Linear Dynamic Analysis Software.
78.Benz W (1990). Smooth particle hydrodynamics: A review. The numerical modelling of nonlinear
stellar pulsations, problems and prospects, Kluwer Academic, Boston.
79.Morris JP (1996) A study of the stability properties of smooth particle hydrodynamics. Publ Astron Soc
Aust 13(1):97-102.
80.Omang M, Borve S, Trulsen J (2005) Alternative kernel functions for Smoothed Particle
Hydrodynamics in cylindrical symmetry. Shock Waves 14(4):293-298.
81.Jin HB, Ding X (2005) On criterions for smoothed particle hydrodynamics kernels in stable field. J
Comput Phys 202(2):699-709.
82.Capuzzo-Dolcetta R, Di Lisio R (2000) A criterion for the choice of the interpolation kernel in
smoothed particle hydrodynamics. Appl Numer Math 34(4):363-371.
83.Cabezon RM, Garcia-Senz D, Relano A (2008) A one-parameter family of interpolating kernels for
smoothed particle hydrodynamics studies. J Comput Phys 227(19):8523-8540.
84.Monaghan JJ, Lattanzio JC (1985) A refined particle method for astrophysical problems. Astron
Astrophys 149(1):135-143.
85.Johnson GR, Stryk RA, Beissel SR (1996) SPH for high velocity impact computations. Comput Method
Appl M 139(1-4):347-373.
86.Swegle JW, Attaway SW, Heinstein MW, Mello FJ, Hicks DL (1994) An analysis of smoothed
particle hydrodynamics. Sandia National Labs. Albuquerque, NM, US.
87.Liu GR, Quek SS (2003) The finite element method: a practical course. Butterworth Heinemann.
88.Chen JK, Beraun JE, Jih CJ (1999) An improvement for tensile instability in smoothed particle
hydrodynamics. Comput Mech 23(4):279-287.
89.Chen JK, Beraun JE, Jih CJ (1999) Completeness of corrective smoothed particle method for linear
elastodynamics. Comput Mech 24(4):273-285.
90.Hernquist L, Katz N (1989) TREESPH- A unification of SPH with the hierarchical tree method.
Astrophys J Suppl S 70(2):419-446.
91.Sod GA (1978) A survey of several finite difference methods for systems of nonlinear hyperbolic
conservation laws. J Comput Phys 27:1-31.

84
92.Fletcher CAJ (1991) Computational techniques for fluid dynamics: 1; fundamental and general
techniques. Springer Verlag.
93.Agertz O, Moore B, Stadel J, Potter D, Miniati F, Read J, Mayer L, Gawryszczak A, Kravtosov A,
Nordlund A, Pearce F, Quilis V, Rudd D, Springel V, Stone J, Tasker E, Teyssier R, Wadsley
J, Walder R (2007) Fundamental differences between SPH and grid methods. Mon Not R Astron
Soc 380:963-978.
94.Schussler M, Schmitt D (1981) Comments on smoothed particle hydrodynamics. Astron Astrophys
97(2):373-379.
95.Espanol P (1998) Fluid particle model. Phys Rev E 57(3):2930-2948.
96.Wang L, Ge W, Li J (2006) A new wall boundary condition in particle methods. Comput Phys
Commun 174(5):386-390.
97.Revenga M, Zuniga I, Espanol P (1998) Boundary models in DPD. Int J Mod Phys C 9(8):1319-1328.
98.Liu MB, Meakin P, Huang H (2007) Dissipative particle dynamics simulation of pore-scale flow. Water
Resour Res 43. W04411, DOI:10.1029/2006WR004856.
99.Revenga M, Zuniga I, Espanol P (1998) Boundary models in DPD. Int J Mod Phys C 9(8):1319-1328.
100.Revenga M, Zuniga I, Espanol P (1999) Boundary conditions in dissipative particle dynamics.
Comput Phys Commun 121(122):309-311.
101.Willemsen SM, Hoefsloot HCJ, Iedema PD (2000) No-slip boundary condition in dissipative particle
dynamics. Int J Mod Phys C 11(5):881-890.
102.Duong-Hong D, Phan-Thien N, Fan X (2004) An implementation of no-slip boundary conditions in
DPD. Comput Mech 35(1):24-29.
103.Crespo AJC, Gomez-Gesteira M, Dalrymple RA (2007) Boundary conditions generated by dynamic
particles in SPH methods. CMC-Comput Mater Con 5(3):173-184.
104.Gong K, Liu H (2007). A new boundary treatment for smoothed particle hydrodynamics. Asian and
Pacific Coasts 2007, Nanjing, China.
105.Hieber SE, Koumoutsakos P (2008) An immersed boundary method for smoothed particle
hydrodynamics of self-propelled swimmers. J Comput Phys 227(19):8636-8654.
106.Klessen R (1997) GRAPESPH with fully periodic boundary conditions: fragmentation of molecular
clouds. Mon Not R Astron Soc 292(1):11-18.
107.Randles PW, Libersky LD (2005) Boundary conditions for a dual particle method. Comput Struct
83(17-18):1476-1486.
108.Yildiz M, Rook RA, Suleman A (2009) SPH with the multiple boundary tangent method. Int J Numer
Meth Eng 77(10):1416-1438.
109.Takeda H, Miyama SM, Sekiya M (1994) Numerical simulation of viscous flow by smoothed particle
hydrodynamics. Prog Theor Phys 92(5):939-960.
110.Morris JP, Fox PJ, Zhu Y (1997) Modeling low Reynolds number incompressible flows using SPH. J
Comput Phys 136(1):214-226.
111.Monaghan JJ (1994) Simulating free surface flows with SPH. J Comput Phys 110(2):399-406.
112.Campbell PM (1989). Some new algorithms for boundary value problems in smooth particle
hydrodynamics. Technical Report, Mission Research Corp, Albuquerque, NM.
113.Libersky LD, Petschek AG, Carney TC, Hipp JR, Allahdadi FA (1993) High strain Lagrangian
hydrodynamics: a three-dimensional SPH code for dynamic material response. J Comput Phys
109(1):67-75.
114.Rapaport DC (2004) The art of molecular dynamics simulation. Cambridge University Press.
115.Liu MB, Meakin P, Huang H (2007) Dissipative particle dynamics simulation of fluid motion through
an unsaturated fracture and fracture junction. J Comput Phys 222(1):110-130.
116.Tartakovsky AM, Meakin P (2005) Simulation of unsaturated flow in complex fractures using
smoothed particle hydrodynamics. Vadose Zone J 4(3):848-855.
117.Liu MB, Meakin P, Huang H (2007) Dissipative particle dynamics simulation of multiphase fluid flow
in microchannels and microchannel networks. Phys Fluids 19(3):033302.
118.Colagrossi A, Landrini M (2003) Numerical simulation of interfacial flows by smoothed particle
hydrodynamics. J Comput Phys 191(2):448-475.
119.Owen JM, Villumsen JV, Shapiro PR, Martel H (1998) Adaptive smoothed particle hydrodynamics:
methodology. II. Astrophys J Suppl S 116(2):155-209.

85
120.Shapiro PR, Martel H, Villumsen JV, Owen JM (1996) Adaptive smoothed particle hydrodynamics,
with application to cosmology: methodology. Astrophys J Suppl S 103(2):269-330.
121.Fulbright MS, Benz W, Davies MB (1995) A method of smoothed particle hydrodynamics using
spheroidal kernels. Astrophys J 440(1):254-262.
122.Liu MB, Liu GR (2005) Meshfree particle simulation of micro channel flows with surface tension.
Comput Mech 35(5):332-341.
123.Liu MB, Liu GR, Lam KY (2006) Adaptive smoothed particle hydrodynamics for high strain
hydrodynamics with material strength. Shock Waves 15(1):21-29.
124.Swegle JW, Attaway SW (1995) On the feasibility of using Smoothed Particle Hydrodynamics for
underwater explosion calculations. Comput Mech 17(3):151-168.
125.Attaway SW, Hendrickson BA, Plimpton SJ, Gardner DR, Vaughan CT, Brown KH, Heinstein MW
(1998) A parallel contact detection algorithm for transient solid dynamics simulations using
PRONTO3D. Comput Mech 22(2):143-159.
126.Campbell J, Vignjevic R, Libersky L (2000) A contact algorithm for smoothed particle hydrodynamics.
Comput Method Appl M 184(1):49-65.
127.Drumm C, Tiwari S, Kuhnert J, Bart HJ (2008) Finite pointset method for simulation of the liquid-
liquid flow field in an extractor. Comput Chem Eng 32(12):2946-2957.
128.Huang H, Dyka CT, Saigal S (2004) Hybrid particle methods in frictionless impact-contact problems.
Int J Numer Meth Eng 61:2250-2272.
129.Li Y, Liu GR, Luan MT, Dai KY, Zhong ZH, Li GY, Han X (2007) Contact analysis for solids
based on linearly conforming radial point interpolation method. Comput Mech 39(4):537-554.
130.Limido J, Espinosa C, Salauen M, Lacome JL (2007) SPH method applied to high speed cutting
modelling. Int J Mech Sci 49(7):898-908.
131.Mehra V, Chaturvedi S (2006) High velocity impact of metal sphere on thin metallic plates: A
comparative smooth particle hydrodynamics study. J Comput Phys 212(1):318-337.
132.Parshikov AN, Medin SA (2002) Smoothed particle hydrodynamics using interparticle contact
algorithms. J Comput Phys 180(1):358-382.
133.Parshikov AN, Medin SA, Loukashenko, II, Milekhin VA (2000) Improvements in SPH method by
means of interparticle contact algorithm and analysis of perforation tests at moderate projectile
velocities. Int J Impact Eng 24(8):779-796.
134.Plimpton S, Attaway S, Hendrickson B, Swegle J, Vaughan C, Gardner D (1998) Parallel transient
dynamics simulations: Algorithms for contact detection and smoothed particle hydrodynamics. J
Parallel Distr Com 50(1-2):104-122.
135.Prochazka PP, Kravtsov AN, Peskova S (2008) Blast impact on structures of underground parking.
Underground Spaces: Design, Engineering and Environmental Aspects 102:11-19.
136.Sauer RA, Li SF (2008) An atomistically enriched continuum model for nanoscale contact mechanics
and Its application to contact scaling. J Nanosci Nanotechno 8(7):3757-3773.
137.Schieback C, Burzle F, Franzrahe K, Neder J, Henseler P, Mutter D, Schwierz N, Nielaba P (2009)
Computer simulations of complex many-body systems. in: High Performance Computing in
Science and Engineering '08. Stuttgart, Germany.
138.Seo S, Min O, Lee J (2008) Application of an improved contact algorithm for penetration analysis in
SPH. Int J Impact Eng 35(6):578-588.
139.Vignjevic R, Campbell J (1999) A penalty approach for contact in smoothed particle hydrodynamics.
Int J Impact Eng 23(1):945-956.
140.Vignjevic R, De Vuyst T, Campbell JC, Source C (2006) A frictionless contact algorithm for
meshless methods. CMES-Comp Model Eng 13(1):35.
141.Belytschko T, Neal MO (1991) Contact-impact by the pinball algorithm with penalty and Lagrangian
methods. Int J Numer Meth Eng 31(3):547-572.
142.Liu MB, Liu GR, Lam KY, Zong Z (2003) Smoothed particle hydrodynamics for numerical
simulation of underwater explosion. Comput Mech 30(2):106-118.
143.Liu MB, Liu GR, Lam KY, Zong Z (2001) A new technique to treat material interfaces for smoothed
particle hydrodynamics. in: Computational Mechanics – New Frontiers for New Millennium.
Elsevier Science.
144.Monaghan JJ (2000) SPH without a tensile instability. J Comput Phys 159(2):290-311.

86
145.Vignjevic R, De Vuyst T, Campbell J (2002). The use of a homogeneous repulsive force for contact
treatment in sph. WCCM V, Vienna, Austria.
146.Balsara DS (1995) von Neumann stability analysis of smoothed particle hydrodynamics suggestions
for optimal algorithms. J Comput Phys 121(2):357-372.
147.Belytschko T, Xiao S (2002) Stability analysis of particle methods with corrected derivatives. Comput
Math Appl 43(3-5):329-350.
148.Bonet J, Kulasegaram S (2001) Remarks on tension instability of Eulerian and Lagrangian corrected
smooth particle hydrodynamics (CSPH) methods. Int J Numer Meth Eng 52:1203-1220.
149.Lanson N, Vila JP (2007) Renormalized meshfree schemes I: Consistency, stability, and hybrid
methods for conservation laws. Siam J Numer Anal 46(4):1912-1934.
150.Randles PW, Petschek AG, Libersky LD, Dyka CT (2003) Stability of DPD and SPH. in: Meshfree
Methods for Partial Differential Equations. Springer.
151.Sigalotti LDG, Lopez H (2008) Adaptive kernel estimation and SPH tensile instability. Comput Math
Appl 55(1):23-50.
152.Gray JP, Monaghan JJ, Swift RP (2001) SPH elastic dynamics. Comput Method Appl M
190(49):6641-6662.
153.Beissel S, Belytschko T (1996) Nodal integration of the element-free Galerkin method. Comput
Method Appl M 139(1-4):49-74.
154.Melean Y, Sigalotti LDG, Hasmy A (2004) On the SPH tensile instability in forming viscous liquid
drops. Comput Phys Commun 157(3):191-200.
155.Benz W (1988) Applications of smooth particle hydrodynamics (SPH) to astrophysical problems.
Comput Phys Commun 48(1):97-105.
156.Frederic AR, James CL (1999) Smoothed particle hydrodynamics calculations of stellar interactions. J
Comput Appl Math 109:213-230.
157.Hultman J, Pharayn A (1999) Hierarchical, dissipative formation of elliptical galaxies: is thermal
instability the key mechanism ? hydrodynamic simulations including supernova feedback multi-
phase gas and metal enrichment in cdm: structure and dynamics of elliptical galaxies. Astron
Astrophys 347:769-798.
158.Thacker RJ, Couchman HMP (2001) Star formation, Supernova feedback, and the angular momentum
problem in numerical cold dark matter cosmogony: halfway there. Astrophys J 555(1):L17-L20.
159.Monaghan JJ, Lattanzio JC (1991) A simulation of the collapse and fragmentation of cooling
molecular clouds. Astrophys J 375(1):177-189.
160.Berczik P (2000) Modeling the star formation in galaxies using the chemo-dynamical SPH code.
Astrophys Space Sci 271(2):103-126.
161.Lee WH, Kluzniak W (1999) Newtonian hydrodynamics of the coalescence of black holes with
neutron stars II. Tidally locked binaries with a soft equation of state. Mon Not R Astron Soc
308(3):780-794.
162.Lee WH (2000) Newtonian hydrodynamics of the coalescence of black holes with neutron stars- III.
Irrotational binaries with a stiff equation of state. Mon Not R Astron Soc 318(2):606-624.
163.Senz DG, Bravo E, Woosley SE (1999) Single and multiple detonations in white dwarfs. Astron
Astrophys 349 177-188.
164.Monaghan JJ (1990) Modelling the universe. Astronomical Society of Australia Proceedings 8(3):233-
237.
165.Hong JM, Lee HY, Yoon JC, Kim CH (2008) Bubbles alive. ACM T Graphic 27(3).
166.Monaghan JJ, Kocharyan A (1995) SPH simulation of multi-phase flow. Comput Phys Commun
87:225-235.
167.Hu XY, Adams NA (2006) A multi-phase SPH method for macroscopic and mesoscopic flows. J
Comput Phys 213(2):844-861.
168.Garg R, Narayanan C, Lakehal D, Subramaniam S (2007) Accurate numerical estimation of
interphase momentum transfer in Lagrangian-Eulerian simulations of dispersed two-phase flows.
Int J Multiphas Flow 33(12):1337-1364.
169.Liu J, Koshizuka S, Oka Y (2005) A hybrid particle-mesh method for viscous, incompressible,
multiphase flows. J Comput Phys 202(1):65-93.

87
170.Password F (2003) Cosmological smoothed particle hydrodynamics simulations: a hybrid multiphase
model for star formation. Mon Not R Astron Soc 339(2):289-311.
171.Ritchie BW, Thomas PA (2001) Multiphase smoothed-particle hydrodynamics. Mon Not R Astron Soc
323(3):743-756.
172.Hu XY, Adams NA (2009) A constant-density approach for incompressible multi-phase SPH. J
Comput Phys 228(6):2082-2091.
173.Hu XY, Adams NA (2007) An incompressible multi-phase SPH method. J Comput Phys 227(1):264-
278.
174.Dalrymple RA, Rogers BD (2006) Numerical modeling of water waves with the SPH method. Coast
Eng 53(2-3):141-147.
175.Crespo AJC, Gomez-Gesteira M, Dalrymple RA (2007) 3D SPH simulation of large waves mitigation
with a dike. J Hydraul Res 45(5):631-642.
176.Shao SD, Gotoh H (2004) Simulating coupled motion of progressive wave and floating curtain wall by
SPH-LES model. Coast Eng J 46(2):171-202.
177.Gotoh H, Shao SD, Memita T (2004) SPH-LES model for numerical investigation of wave interaction
with partially immersed breakwater. Coast Eng 46(1):39-63.
178.Shao SD (2006) Incompressible SPH simulation of wave breaking and overtopping with turbulence
modelling. Int J Numer Meth Fl 50(5):597-621.
179.Gomez-Gesteira M, Dalrymple RA (2004) Using a three-dimensional smoothed particle
hydrodynamics method for wave impact on a tall structure. J Waterw Port C-Asce 130(2):63-69.
180.Crespo AJC, Gomez-Gesteira M, Carracedo P, Dalrymple RA (2008) Hybridation of generation
propagation models and SPH model to study severe sea states in Galician Coast. J Marine Syst
72(1-4):135-144.
181.Crespo AJC, Gomez-Gesteira M, Dalrymple RA (2008) Modeling dam break behavior over a wet bed
by a SPH technique. J Waterw Port C-Asce 134(6):313-320.
182.Gotoh H, Sakai T (2006) Key issues in the particle method for computation of wave breaking. Coast
Eng 53(2-3):171-179.
183.Issa R, Violeau D (2008) Modelling a plunging breaking solitary wave with eddy-viscosity turbulent
SPH models. CMC-Comput Mater Con 8(3):151-164.
184.Khayyer A, Gotoh H, Shao SD (2008) Corrected incompressible SPH method for accurate water-
surface tracking in breaking waves. Coast Eng 55(3):236-250.
185.Monaghan JJ, Kos A, Issa N (2003) Fluid motion generated by impact. J Waterw Port C-Asce
129(6):250-259.
186.Panizzo A (2005). SPH modelling of underwater landslide generated waves. Coastal Engineering 2004
(Proceedings of the 29th International Conference), Lisbon, Portugal.
187.Qiu LC (2008) Two-dimensional SPH simulations of landslide-generated water waves. J Hydraul Eng-
ASCE 134(5):668-671.
188.Rogers BD, Dalrymple RA (2005). SPH modeling of breaking waves. Coastal Engineering 2004
(Proceedings of the 29th International Conference), Lisbon, Portugal.
189.Shao SD (2009) Incompressible SPH simulation of water entry of a free-falling object. Int J Numer
Meth Fl 59(1):91-115.
190.Shao SD, Ji CM, Graham DI, Reeve DE, James PW, Chadwick AJ (2006) Simulation of wave
overtopping by an incompressible SPH model. Coast Eng 53(9):723-735.
191.Violeau D, Buvat C, Abed-Meraim K, de Nanteuil E (2007) Numerical modelling of boom and oil
spill with SPH. Coast Eng 54(12):895-913.
192.Yim SC, Yuk D, Panizzo A, Di Risio M, Liu PLF (2008) Numerical Simulations of wave generation
by a vertical plunger using RANS and SPH models. J Waterw Port C-Asce 134(3):143-159.
193.Zou S, Dalrymple RA (2005). Sediment suspension modeling by smoothed particle hydrodynamics.
Coastal Engineering 2004 (Proceedings of the 29th International Conference), Lisbon, Portugal.
194.Bulgarelli UP (2005) The application of numerical methods for the solution of some problems in free-
surface hydrodynamics. J Ship Res 49(4):288-301.
195.Iglesias AS, Rojas LP, Rodriguez RZ (2004) Simulation of anti-roll tanks and sloshing type problems
with smoothed particle hydrodynamics. Ocean Engrg 31(8-9):1169-1192.
196.Kim Y (2007) Experimental and numerical analyses of sloshing flows. J Eng Math 58(1-4):191-210.

88
197.Lohner R, Yang C, Onate E (2006) On the simulation of flows with violent free surface motion.
Comput Method Appl M 195(41-43):5597-5620.
198.Rhee SH, Engineer L (2005) Unstructured grid based Reynolds-averaged Navier-Stokes method for
liquid tank sloshing. J Fluid Eng 127:572.
199.Souto-Iglesias A, Delorme L, Perez-Rojas L, Abril-Perez S (2006) Liquid moment amplitude
assessment in sloshing type problems with smooth particle hydrodynamics. Ocean Engrg 33(11-
12):1462-1484.
200.Bursik M, Martinez-Hackert B, Delgado H, Gonzalez-Huesca A (2003) A smoothed-particle
hydrodynamic automaton of landform degradation by overland flow. Geomorphology 53(1-2):25-
44.
201.Zhu Y, Fox PJ, Morris JP (1999) A pore-scale numerical model for flow through porous media. Int J
Numer Anal Met 23(9):881-904.
202.Zhu Y, Fox PJ (2002) Simulation of pore-scale dispersion in periodic porous media using smoothed
particle hydrodynamics. J Comput Phys 182(2):622-645.
203.Tartakovsky AM, Meakin P (2006) Pore scale modeling of immiscible and miscible fluid flows using
smoothed particle hydrodynamics. Adv Water Resour 29(10):1464-1478.
204.Cleary PW, Prakash M (2004) Discrete-element modelling and smoothed particle hydrodynamics:
potential in the environmental sciences. Philos T Roy Soc A 362(1822):2003-2030.
205.Ghazali JN, Kamsin A (2008). A real time simulation and modeling of flood hazard. 12th WSEAS
International Conference on SYSTEMS, Heraklion, Greece.
206.Kipfer P, Westermann R (2006). Realistic and interactive simulation of rivers. Graphics Interface 2006,
Quebec, Canada
207.Bui HH, Fukagawa R, Sako K (2006). Smoothed particle hydrodynamics for soil mechanics Taylor &
Francis.
208.Bui HH, Fukagawa R, Sako K, Ohno S (2008) Lagrangian meshfree particles method (SPH) for large
deformation and failure flows of geomaterial using elastic-plastic soil constitutive model. Int J
Numer Anal Met 32(12):1537-1570.
209.Bui HH, Sako K, Fukagawa R (2007) Numerical simulation of soil-water interaction using smoothed
particle hydrodynamics (SPH) method. J Terramechanics 44(5):339-346.
210.Gallati M, Braschi G, Falappi S (2005) SPH simulations of the waves produced by a falling mass into
a reservoir. Nuovo Cimento Della Societa Italiana Di Fisica C-Geophysics and Space Physics
28(2):129-140.
211.Herrera PA, Massabo M, Beckie RD (2009) A meshless method to simulate solute transport in
heterogeneous porous media. Adv Water Resour 32(3):413-429.
212.Hui HH, Fukagawa R, Sako K (2006) Smoothed particle hydrodynamics for soil mechanics.
Terramechanics 26:49-53.
213.Laigle D, Lachamp P, Naaim M (2007) SPH-based numerical investigation of mudflow and other
complex fluid flow interactions with structures. Comput Geosci 11(4):297-306.
214.Maeda K, Sakai H (2007) Seepage failure analysis with evolution of air bubbles by SPH. in: New
Frontiers in Chinese and Japanese Geotechniques ( Proceedings of the 3rd Sino-Japan
Geotechnical Symposium). Chongqing, China.
215.McDougall S, Hungr O (2004) A model for the analysis of rapid landslide motion across three-
dimensional terrain. Can Geotech J 41(6):1084-1097.
216.McDougall S, Hungr O (2005) Dynamic modelling of entrainment in rapid landslides. Can Geotech J
42(5):1437-1448.
217.Moresi L, Muhlhous H, Dufour F (2001). An overview of numerical methods for Earth simulations.
218.Morris JP, Zhu Y, Fox PJ (1999) Parallel simulations of pore-scale flow through porous media.
Comput Geotech 25(4):227-246.
219.Pastor M, Haddad B, Sorbino G, Cuomo S, Drempetic V (2009) A depth-integrated, coupled SPH
model for flow-like landslides and related phenomena. Int J Numer Anal Met 33(2).
220.Sakai H, Maeda K (2006) Seepage failure of granular ground accounting for soil-water-gas interaction.
in: Geomechanics and Geotechnics of Particulate Media (Proceedings of the International
Symposium on Geomechanics and Geotechnics of Particulate Media). Ube, Yamaguchi, Japan.

89
221.Cleary PW (1998) Modelling confined multi-material heat and mass flows using SPH. Appl Math
Model 22(12):981-993.
222.Jeong JH, Jhon MS, Halow JS, Van Osdol J (2003) Smoothed particle hydrodynamics: Applications
to heat conduction. Comput Phys Commun 153(1):71-84.
223.Jiang F, Sousa ACM (2006) SPH numerical modeling for ballistic-diffusive heat conduction. Numer.
Heat Tr. B-Fund 50(6):499-515.
224.Jiang FM, Sousa ACM (2006) SPH numerical modeling for ballistic-diffusive heat conduction. Numer
Heat Tr B-Fund 50(6):499-515.
225.Rook R, Yildiz M, Dost S (2007) Modeling transient heat transfer using SPH and implicit time
integration. Numer Heat Tr B-Fund 51(1):1-23.
226.Sousa ACM, Jiang FM (2007) SPH as an inverse numerical tool for the prediction of diffusive
properties in porous media. Diffusion in Solids and Liquids: Heat Transfer - Microstructure &
Properties 553:171-189.
227.Gutfraind R, Savage SB (1998) Flow of fractured ice through wedge-shaped channels: smoothed
particle hydrodynamics and discrete-element simulations. Mech Mater 29(1):1-17.
228.Oger L, Savage SB (1999) Smoothed particle hydrodynamics for cohesive grains. Comput Method
Appl M 180(1):169-183.
229.Ji SY, Li H, Shen HT, Wang RX, Yue QJ (2007) A hybrid Lagrangian-Eulerian numerical model for
sea-ice dynamics. Acta Oceanol Sin 26:12-24.
230.Ji SY, Shen HT, Wang ZL, Shen HH, Yue QJ (2005) A viscoelastic-plastic constitutive model with
Mohr-Coulomb yielding criterion for sea ice dynamics. Acta Oceanol Sin 24(4):54-65.
231.Schafer C, Speith R, Kley W (2007) Collisions between equal-sized ice grain agglomerates. Astron
Astrophys 470(2):733-739.
232.Shen HT, Su JS, Liu LW (2000) SPH simulation of river ice dynamics. J Comput Phys 165(2):752-
770.
233.Wang RX, Ji SY, Shen HT, Yue QJ (2005) Modified PIC method for sea ice dynamics. China Ocean
Eng 19(3):457-468.
234.Tartakovsky A, Meakin P (2005) Modeling of surface tension and contact angles with smoothed
particle hydrodynamics. Phys Rev E 72(2):26301.
235.Nugent S, Posch HA (2000) Liquid drops and surface tension with smoothed particle applied
mechanics. Phys Rev E 62(4):4968-4975.
236.Apfel RE, Tian Y, Jankovsky J, Shi T, Chen X, Holt RG, Trinh E, Croonquist A, Thornton KC,
Sacco JA (1997) Free oscillations and surfactant studies of superdeformed drops in microgravity.
Phys Rev Lett 78(10):1912-1915.
237.Fang HS, Bao K, Wei JA, Zhang H, Wu EH, Zheng LL (2009) Simulations of droplet spreading and
solidification using an improved SPH model. Numer Heat Tr A-Appl 55(2):124-143.
238.Lopez H, Sigalotti L, Di G (2006) Oscillation of viscous drops with smoothed particle hydrodynamics.
Phys Rev E 73(5):51201.
239.Melean Y, Sigalotti LD (2005) Coalescence of colliding van der Waals liquid drops. Int J Heat Mass
Tran 48:4041-4061.
240.Melean Y, Sigalotti LDG (2005) Coalescence of colliding van der Waals liquid drops. INT J HEAT
MASS TRAN 48(19-20):4041-4061.
241.Wang W, Huang Y, Grujicic M, Chrisey DB (2008) Study of impact-induced mechanical effects in
cell direct writing using smooth particle hydrodynamic method. J Manuf Sci E-T ASME 130(2).
242.Zhang MY, Zhang H, Zheng LL (2007) Application of smoothed particle hydrodynamics method to
free surface and solidification problems. Numer Heat Tr A-Appl 52(4):299-314.
243.Zhang MY, Zhang H, Zheng LL (2008) Simulation of droplet spreading, splashing and solidification
using smoothed particle hydrodynamics method. Int J Heat Mass Tran 51(13-14):3410-3419.
244.Zhang MY, Zhang H, Zheng LL (2009) Numerical Investigation of Substrate Melting and
Deformation During Thermal Spray Coating by SPH Method. Plasma Chem Plasma P 29(1):55-68.
245.Zhou GZ, Ge W, Li JH (2008) A revised surface tension model for macro-scale particle methods.
Powder Technol 183(1):21-26.
246.Liu MB, Liu GR, Lam KY, Zong Z (2003) Meshfree particle simulation of the detonation process for
high explosives in shaped charge unlined cavity configurations. Shock Waves 12(6):509-520.

90
247.Liu MB, Liu GR, Zong Z, Lam KY (2003) Computer simulation of high explosive explosion using
smoothed particle hydrodynamics methodology. Comput Fluids 32(3):305-322.
248.Liu MB, Liu GR, Lam KY (2002) Investigations into water mitigation using a meshless particle
method. Shock Waves 12(3):181-195.
249.Liu MB, Liu GR (2004) Smoothed particle hydrodynamics: some recent developments in theory and
applications. J. Beijing Polytech. Univ. 30:61-71.
250.Liu MB, Liu GR, Lam KY, Zong Z (2003) Computer simulation of shaped charge detonation using
meshless particle method. Fragblast 7(3):181-202.
251.Alia A, Souli M (2006) High explosive simulation using multi-material formulations. Appl Therm Eng
26(10):1032-1042.
252.Bromm V, Yoshida N, Hernquist L (2003) The first supernova explosions in the universe. Astrophys J
596(2):L135-L138.
253.Busegnies Y, Francois J, Paulus G (2007) Unidimensional SPH simulations of reactive shock tubes in
an astrophysical perspective. Shock Waves 16(4-5):359-389.
254.Liu MB, Liu GR, Lam KY (2003) Comparative study of the real and artificial detonation models in
underwater explosions. Electronic Modeling 25(2):113-124.
255.Mair HU (1999) Review: Hydrocodes for structural response to underwater explosions. Shock Vib
6(2):81-96.
256.Swegle JW (1992) Report at Sandia National Laboratories.
257.Libersky LD, Petschek AG (1991). Smooth particle hydrodynamics with strength of materials.
Advances in the Free-Lagrange Method Including Contributions on Adaptive Gridding and the
Smooth Particle Hydrodynamics Method: Proceedings of the Next Free-Lagrange Conference,
Jackson Lake Lodge, Moran, WY, USA.
258.Zhou CE, Liu GR, Lou KY (2007) Three-dimensional penetration simulation using smoothed particle
hydrodynamics. Int J Comput Meth 4(4):671-691.
259.Benz W, Asphaug E (1995) Simulations of brittle solids using smooth particle hydrodynamics. Comput
Phys Commun 87(1):253-265.
260.Cleary P, Ha J, Alguine V, Nguyen T (2002) Flow modelling in casting processes. Appl Math Model
26(2):171-190.
261.Cleary PW, Ha J, Ahuja V (2000) High pressure die casting simulation using smoothed particle
hydrodynamics. Int J Cast Metal Res 12(6):335-355.
262.Chen JS, Pan C, Roque C, Wang HP (1998) A Lagrangian reproducing kernel particle method for
metal forming analysis. Comput Mech 22(3):289-307.
263.Cleary PW, Ha J (2000) Three dimensional modelling of high pressure die casting. Int J Cast Metal
Res 12(6):357-365.
264.Cleary PW, Prakash M, Ha J (2006) Novel applications of smoothed particle hydrodynamics (SPH) in
metal forming. J Mater Process Tech 177(1-3):41-48.
265.Cleary PW, Prakash M, Ha J, Stokes N, Scott C (2007) Smooth particle hydrodynamics: status and
future potential. Prog Comput Fluid Dy 7(2-4):70-90.
266.Ha J, Cleary PW (2000) Comparison of SPH simulations of high pressure die casting with the
experiments and VOF simulations of Schmid and Klein. Int J Cast Metal Res 12(6):409-418.
267.Ha J, Cleary PW (2005) Simulation of high pressure die filling of a moderately complex industrial
object using smoothed particle hydrodynamics. Int J Cast Metal Res 18(2):81-92.
268.Hu W, Yao LG, Hua ZZ (2007) Parallel point interpolation method for three-dimensional metal
forming simulations. Eng Anal Bound Elem 31(4):326-342.
269.Prakash M, Cleary PW, Grandfield J, Rohan P, Nguyen V (2007) Optimisation of ingot casting
wheel design using SPH simulations. Prog Comput Fluid Dy 7(2-4):101-110.
270.Ala G, Francomano E, Tortofici A, Toscano E, Viola F (2007) Corrective meshless particle
formulations for time domain Maxwell's equations. J Comput Appl Math 210(1-2):34-46.
271.Ala G, Francomano E, Tortorici A, Toscano E, Viola F (2006) Smoothed particle electromagnetics: a
mesh-free solver for transients. J Comput Appl Math 191(2):194-205.
272.Ala G, Francomano E, Tortorici A, Toscano E, Viola F, di Ingegneria Elettrica D, e delle
Telecomunicazioni E (2007) A mesh-free particle method for transient full-wave simulation. IEEE
T Magn 43(4):1333-1336.

91
273.Borve S, Omang M, Trulsen J (2001) Regularized smoothed particle hydrodynamics: A new approach
to simulating magnetohydrodynamic shocks. Astrophys J 561(1):82-93.
274.Dolag K, Bartelmann M, Lesch H (1999) SPH simulations of magnetic fields in galaxy clusters.
Astron Astrophys 348(2):351-363.
275.Francomano E, Tortorici A, Toscano E, Ala G, Viola F (2009) On the use of a meshless solver for
PDEs governing electromagnetic transients. Appl Math Comput 209(1):42-51.
276.Jiang FM, Oliveira MSA, Sousa ACM (2006) SPH simulation of transition to turbulence for planar
shear flow subjected to a streamwise magnetic field. J Comput Phys 217(2):485-501.
277.Meglicki Z (1994) Verification and accuracy of smoothed particle magnetohydrodynamics. Comput
Phys Commun 81(1-2):91-104.
278.Pimenta LCA, Mendes ML, Mesquita RC, Pereira GAS (2007) Fluids in electrostatic fields: An
analogy for multirobot control. IEEE T Magn 43(4):1765-1768.
279.Price DJ, Monaghan JJ (2004) Smoothed particle magnetohydrodynamics - I. Algorithm and tests in
one dimension. Mon Not R Astron Soc 348(1):123-138.
280.Price DJ, Monaghan JJ (2004) Smoothed particle magnetohydrodynamics - II. Variational principles
and variable smoothing-length terms. Mon Not R Astron Soc 348(1):139-152.
281.Price DJ, Monaghan JJ (2004) Smoothed Particle Magnetohydrodynamics: Some Shocking Results.
Astrophys Space Sci 292(1):279-283.
282.Antoci C, Gallati M, Sibilla S (2007) Numerical simulation of fluid-structure interaction by SPH.
Comput Struct 85(11-14):879-890.
283.Prakash M, Cleary PW, Ha J, Noui-Mehidi MN, Blackburn H, Brooks G (2007) Simulation of
suspension of solids in a liquid in a mixing tank using SPH and comparison with physical
modelling experiments. Prog Comput Fluid Dy 7(2):91-100.
284.Anghileri M, Castelletti LML, Tirelli M (2005) Fluid structure interaction of water filled tanks during
the impact with the ground. Int J Impact Eng 31(3):235-254.
285.Chikazawa Y, Koshizuka S, Oka Y (2001) A particle method for elastic and visco-plastic structures
and fluid-structure interactions. Comput Mech 27(2):97-106.
286.Guilcher PM, Ducrozet G, Doring M, Alessandrini B, Ferrant P (2006). Numerical simulation of
wave-body interactions using a modified SPH solver. Proceedings of the Sixteenth International
Offshore and Polar Engineering Conference, San Francisco, California.
287.Hosseini SM, Amanifard N (2007) Presenting a modified SPH algorithm for numerical studies of fluid-
structure interaction problems. IJE Trans B: Applications 20:167-178.
288.Muller M (2004) Interactive blood simulation for virtual surgery based on smoothed particle
hydrodynamics. Technol Health Care 12(1):25-31.
289.Hieber SE (2004) Remeshed smoothed particle hydrodynamics simulation of the mechanical behavior
of human organs. Technol Health Care 12(4):305-314.
290.Tanaka N, Takano T (2005) Microscopic-scale simulation of blood flow using SPH method. Int J
Comput Meth 2(4):555-568.
291.Tsubota K, Wada S, Yamaguchi T (2006) Simulation Study on Effects of Hematocrit on Blood Flow
Properties Using Particle Method. J Biomech Sci Eng 1(1):159-170.
292.Rosswog S, Wagner P (2002) Towards a macroscopic modeling of the complexity in traffic flow. Phys
Rev E 65(3):36106.
293.Benson DJ (1992) Computational methods in Lagrangian and Eulerian hydrocodes. Comput Method
Appl M 99(2-3):235-394.
294.Walters WP, Zukas JA (1989) Fundamentals of shaped charges. Wiley.
295.Anderson CE (1987) An overview of the theory of hydrocodes. Int J Impact Eng 5(1-4):33-59.
296.Anghileri M, Castelletti LML, Invernizzi F, Mascheroni M (2005) A survey of numerical models for
hail impact analysis using explicit finite element codes. Int J Impact Eng 31(8):929-944.
297.Attaway SW, Heinstein MW, Swegle JW (1993). Coupling of smooth particle hydrodynamics with the
finite element method. IMPACT-4: structural mechanics in reactor technology conference, Berlin,
Germany.
298.Batra RC, Zhang GM (2008) Modified Smoothed Particle Hydrodynamics (MSPH) basis functions for
meshless methods, and their application to axisymmetric Taylor impact test. J Comput Phys
227(3):1962-1981.

92
299.Brown K, Attaway S, Plimpton S, Hendrickson B (2000) Parallel strategies for crash and impact
simulations. Comput Method Appl M 184:375-390.
300.De Vuyst T, Vignjevic R, Campbell JC (2005) Coupling between meshless and finite element methods.
Int J Impact Eng 31(8):1054-1064.
301.Fahrenthold EP, Koo JC (1997) Energy based particle hydrodynamics for hypervelocity impact
simulation. Int J Impact Eng 20(1-5):253-264.
302.Fawaz Z, Zheng W, Behdinan K (2004) Numerical simulation of normal and oblique ballistic impact
on ceramic composite armours. Compos Struct 63(3-4):387-395.
303.Fountzoulas CG, Gazonas GA, Cheeseman BA (2007) Computational modeling of tungsten carbide
sphere impact and penetration into high-strength-low-alloy (HSLA)-100 steel targets. J Mech
Mater Struct 2(10):1965-1979.
304.Groenenboom PHL (1997) Numerical simulation of 2D and 3D hypervelocity impact using the SPH
option in PAM-SHOCK(TM). Int J Impact Eng 20(1-5):309-323.
305.Hayhurst CJ, Clegg RA (1997) Cylindrically symmetric SPH simulations of hypervelocity impacts on
thin plates. Int J Impact Eng 20(1-5):337-348.
306.Hayhurst CJ, Clegg RA, Livingstone IA, Francis NJ (1996). The application of SPH techniques in
Autodyn-2D to ballistic impact problems. 16th International Symposium on Ballistics, San
Francisco, USA.
307.Hiermaier S, Schafer F (1999) Hypervelocity impact fragment clouds in high pressure gas numerical
and experimental investigations. Int J Impact Eng 23(1):391-400.
308.Ioan A, Brizzolara S, Viviani M, Couty N, Donner R, Hermundstad O, Kukkanen T, Malenica S,
Temarel P (2007) Comparison of experimental and numerical impact loads on ship-like sections.
in: Advancements in Marine Structures. Taylor & Francis.
309.Ji XY, Li YC, Huang XC, Chen G (2007). Coupled FE-SPH simulation of truncated-conical
projectiles penetrating steel sheets. the 7th International Conference on Shock & Impact Loads on
Structures, Beijing, China.
310.Johnson AF, Holzapfel M (2003) Modelling soft body impact on composite structures. Compos Struct
61(1-2):103-113.
311.Johnson GR (1994) Linking of Lagrangian particle methods to standard finite element methods for
high velocity impact computations. Nucl Eng Des 150(2-3):265-274.
312.Johnson GR, Beissel SR, Stryk RA (2000) A generalized particle algorithm for high velocity impact
computations. Comput Mech 25(2):245-256.
313.Johnson GR, Peterson EH, Stryrk RA (1993) Incorporation of an sph option into the epic code for a
wide range of high velocity impact computations. Int J Impact Eng 14:385-394.
314.Johnson GR, Stryk RA, Beissel SR, Holmquist TJ (2002) An algorithm to automatically convert
distorted finite elements into meshless particles during dynamic deformation. Int J Impact Eng
27(10):997-1013.
315.Kikuchi M, Miyamoto B (2006) Numerical simulation of impact crush/buckling of circular tube using
SPH method. Key Eng Mat 306-308:697-702.
316.Kitsionas S, Whitworth AP (2007) High-resolution simulations of clump-clump collisions using SPH
with particle splitting. Mon Not R Astron Soc 378(2):507-524.
317.Lam KY, Shen YG, Gong SW (2001) A study of axial impact of composite rods using SPH approach.
Shock Vib 8(5):303-312.
318.Lee M, Yoo YH (2001) Analysis of ceramic/metal armour systems. Int J Impact Eng 25(9):819-829.
319.Libersky LD, Randles PW, Carney TC, Dickinson DL (1997) Recent improvements in SPH modeling
of hypervelocity impact. Int J Impact Eng 20(6):525-532.
320.Lukyanov AA, Reveles JR, Vignjevic R, Campbell J (2005). Simulation of hypervelocity debris
impact and spacecraft shielding performance. Proceedings of the 4th European Conference on
Space Debris.
321.Ma S, Zhang X, Qiu XM (2009) Comparison study of MPM and SPH in modeling hypervelocity
impact problems. Int J Impact Eng 36(2):272-282.
322.Michel Y, Chevalier JM, Durin C, Espinosa C, Malaise F, Barrau JJ (2006) Hypervelocity impacts
on thin brittle targets: Experimental data and SPH simulations. Int J Impact Eng 33(1-12):441-451.

93
323.Nizampatnam LS, Horn WJ (2008). Investigation of multi-material bird models for predicting impact
loads. Proceedings of the Asme Turbo Expo, Berlin, Germany.
324.Park YK, Fahrenthold EP (2005) A kernel free particle-finite element method for hypervelocity impact
simulation. Int J Numer Meth Eng 63(5):737-759.
325.Rabczuk T, Eibl J (2006) Modelling dynamic failure of concrete with meshfree methods. Int J Impact
Eng 32(11):1878-1897.
326.Randles PW, Carney TC, Libersky LD, Renick JD, Petschek AG (1994) Calculation of oblique
impact and fracture of tungsten cubes using smoothed particle hydrodynamics. Int J Impact Eng
17(4):661-672.
327.Rosen P, Popescu V, Hoffmann C, Irfanoglu A (2008) A high-quality high-fidelity visualization of
the September 11 attack on the World Trade Center. IEEE T Vis Comput Gr 14(4):937-947.
328.Shintate K, Sekine H (2004) Numerical simulation of hypervelocity impacts of a projectile on
laminated composite plate targets by means of improved SPH method. Compos Part A-Appl S
35(6):683-692.
329.Stellingwerf RF, Wingate CA (1993). Impact modeling with Smooth Particle Hydrodynamics.
Conference: Smooth particle hydrodynamics in astrophysics OAT workshop, Trieste, Italy.
330.Taylor EA (2001) Simulation of hollow shaped charge jet impacts onto aluminium whipple bumpers at
11 km/s. Int J Impact Eng 26(1-10):773-784.
331.Vignjevic R, Lepage S, De Vuyst T (2005) Simulation of high velocity impacts on thin metallic
targets II (Discrete elements). WIT Trans Eng Sci 49:575-582.
332.Wang P, Shao JL, Qin CS (2009) Effect of loading-wave-front width on micro-jet from aluminum
surface. Acta Phys Sin 58(2):1064-1070.
333.Zhang QM, Long RR, Liu ZF, Huang FL (2008) SPH simulations of hypervelocity impact of al
spheres on multi-plate structures. Int J Mod Phys B 22(9-11):1604-1611.
334.McCarthy MA, Xiao JR, Petrinic N, Kamoulakos A, Melito V (2004) Modelling of bird strike on an
aircraft wing leading edge eade from fibre metal laminates, Part 1: material modelling. Appl
Compos Mater 11(5):295-315.
335.Brown K, Attaway S, Plimpton S, Hendrickson B (2000) Parallel strategies for crash and impact
simulations. Comput Method Appl M 184(2):375-390.
336.Bicknell GV, Gingold RA (1983) On tidal detonation of stars by massive black holes. Astrophys J
273:749-760.
337.Fulbright MS, Benz W, Davies MB (1995) A method of smoothed particle hydrodynamics using
spheroidal kernels. Astrophys J 440:254.
338.Chin GL (2001) Smoothed particle hydrodynamics and adaptive smoothed particle hydrodynamics
with strength of materials. National University of Singapore.
339.Hiermaier S, Konke D, Stilp AJ, Thoma K (1997) Computational simulation of the hypervelocity
impact of Al-spheres on thin plates of different materials. Int J Impact Eng 20(1-5):363-374.
340.Bangash MYH (1993) Impact and explosion. Blackwell Scientific Publications, Oxford.
341.Cole RH (1948) Underwater explosions. Princeton University Press.
342.Fickett W, W. WW (1966) Flow calculations for pulsating one-dimensional detonations. Phys Fluids
9:903-916.
343.Mader CL (1998) Numerical modeling of explosives and propellants. CRC Press, New York.
344.Libersky LD, Randles PW (2006) Shocks and discontinuities in particle methods. Shock Compression
of Condensed Matter - 2005, Pts 1 and 2 845:1089-1092.
345.Ma S, Zhang X, Lian YP, Zhou X (2009) Simulation of high explosive explosion using adaptive
material point method. CMES-Comp Model Eng 39(2):101-123.
346.Miyoshi H (2008). Numerical simulation of shaped charges using the SPH solver: Jet formation and
target penetration. Explosion, Shock Wave and Hypervelocity Phenomena in Materials II.
347.Tanaka K (2008) Numerical studies of explosive welding by SPH. Explosion, Shock Wave and
Hypervelocity Phenomena in Materials Ii 566:61-64.
348.Wataru K, Akiko M (2006) Explosion simulation by Smoothed Particle Hydrodynamics.
Computational Methods, Pts 1 and 2:1397-1403.
349.Xu JX, Liu XL (2008) Analysis of structural response under blast loads using the coupled SPH-FEM
approach. J Zhejiang Univ-Sci A 9(9):1184-1192.

94
350.Liu MB, Liu GR, Lam KY (2001). Simulation of the explosive detonation process by using SPH
methodology. Computational Fluids and Solid Mechanics, Elsevier Science.
351.Dobratz BM (1981) LLNL explosive handbook. Lawrence Livermore National Laboratory, Livermore,
CA.
352.Chiesum JE, Shin YS (1997) Explosion gas bubbles near simple boundaries. Shock Vib 4(11):11-25.
353.Karniadakis GE, Beskok A, Aluru A (2005) Microflows and nanoflows: fundamentals and simulation.
Springer.
354.Bird GA (1994) Molecular gas dynamics and the direct simulation of gas flows. Oxford University
Press.
355.Beskok A, Karniadakis GE (2001) Microflows: fundamentals and simulation. Springer, Berlin.
356.Pan L, Ng TY, Xu D, Liu GR, Lam KY (2002) Determination of temperature jump coefficient using
the direct simulation Monte Carlo method. J Micromech Microeng 12(1):41-52.
357.Pan LS, Liu GR, Khoo BC, Song B (2000) A modified direct simulation Monte Carlo method for
low-speed microflows. J. Micromech. Microeng 10:21-7.
358.Pan LS, Liu GR, Lam KY (1999) Determination of slip coefficient for rarefied gas flows using direct
simulation Monte Carlo. J Micromech Microeng 9(1):89-96.
359.Duong-Hong D, Wang JS, Liu GR, Chen YZ, Han J, Hadjiconstantinou NG (2007) Dissipative
particle dynamics simulations of electroosmotic flow in nano-fluidic devices. Microfluid
Nanofluid 3:1-7.
360.Hoover WG, Hoover CG, Kum O, Castillo VM, Posch HA, Hess S (1996) Smooth particle applied
mechanics. Comput Methods Sci Engrg 2:6572.
361.Hoover WG, Pierce TG, Hoover CG, Shugart JO, Stein CM, Edwards AL (1994) Molecular
dynamics, smoothed-particle applied mechanics, and irreversibility. Comput. Math. Appl 28(10-
12):155-174.
362.Hoover WG, Hoover CG (2003) Links between microscopic and macroscopic fluid mechanics. Mol
Phys 101(11):1559-1573.
363.Espanol P (1997) Fluid particle dynamics: A synthesis of dissipative particle dynamics and smoothed
particle dynamics. Europhys Lett 39(6):605-610.
364.Espanol P, Revenga M (2003) Smoothed dissipative particle dynamics. Phys Rev E 67(2):026705.
365.Liu MB, Liu GR, Lam KY (2003) Computer simulation of flip-chip underfill encapsulation process
using meshfree particle method. Int J Comput Engrg Sci 4(2):405-408.
366.Li Q, Cai TM, He GQ, Hu CB (2006) Droplet collision and coalescence model. Appl Math Mech-
Engl 27(1):67-73.
367.Tartakovsky AM, Meakin P (2005) A smoothed particle hydrodynamics model for miscible flow in
three-dimensional fractures and the two-dimensional Rayleigh Taylor instability. J Comput Phys
207(2):610-624.
368.Tartakovsky AM, Meakin P, Scheibe TD, Eichler West RM (2007) Simulations of reactive transport
and precipitation with smoothed particle hydrodynamics. J Comput Phys 222(2):654-672.
369.Tartakovsky AM, Redden G, Lichtner PC, Scheibe TD, Meakin P (2008) Mixing-induced
precipitation: Experimental study and multiscale numerical analysis. Water Resour Res 44(6).
370.Tartakovsky AM, Meakin P, Ward AL (2009) Smoothed particle hydrodynamics model of non-
Aqueous phase liquid flow and dissolution. Transport Porous Med 76(1):11-34.
371.Brackbill JU, Kothe DB, Zemach C (1992) A continuum method for modeling surface tension. J
Comput Phys 100(2):335-354.
372.Morris JP (2000) Simulating surface tension with smoothed particle hydrodynamics. Int J Numer Meth
Fl 33(3):333-353.
373.Lo YME, Shao S (2002) Simulation of near-shore solitary wave mechanics by an incompressible SPH
method. Appl Ocean Res 24(5):275-286.
374.Shao S (2006) Incompressible SPH simulation of wave breaking and overtopping with turbulence
modelling. Int J Numer Meth Fl 50(5).
375.Shao S, Ji C (2006) SPH computation of plunging waves using a 2-D sub-particle scale (SPS)
turbulence model. Int J Numer Meth Fl 51(8).
376.Shao SD (2005) SPH simulation of solitary wave interaction with a curtain-type breakwater. J Hydraul
Res 43(4):366-375.

95
377.Frank J, Reich S (2003) Conservation properties of smoothed particle hydrodynamics applied to the
shallow water equation. Bit 43(1):41-55.
378.Ata R, Soulaimani A (2005) A stabilized SPH method for inviscid shallow water flows. Int J Numer
Meth Fl 47(2):139-159.
379.Rodriguez-Paz M, Bonet J (2005) A corrected smooth particle hydrodynamics formulation of the
shallow-water equations. Comput Struct 83(17-18):1396-1410.
380.Idelsohn SR, Onate E, Pin FD (2004) The particle finite element method: a powerful tool to solve
incompressible flows with free-surfaces and breaking waves. Int J Numer Meth Eng 61(7):964-
989.
381.Oger G, Alessandrini B, Ferrant P (2005). Capture of air cushion effects in a wedge water entry SPH
simulation. Proceedings of the Fifteenth International Offshore and Polar Engineering Conference.
382.Callati M, Braschi G, Falappi S (2005) SPH simulations of the waves produced by a falling mass into
a reservoir. Nuovo Cimento-C 28(2):129-140.
383.Oger G, Doring M, Alessandrini B, Ferrant P (2006) Two-dimensional SPH simulations of wedge
water entries. J Comput Phys 213(2):803-822.
384.Delorme L, Iglesias AS, Perez SA (2005). Sloshing loads simulation in LNG tankers with SPH.
International Conference on Computational Methods in Marine Engineering, Barcelona.
385.Greco M, Landrini M, Faltinsen OM (2004) Impact flows and loads on ship-deck structures. J. Fluids
Struct. 19(3):251-275.
386.Liu MB, Liu GR, Zong Z (2008) An overview on smoothed particle hydrodynamics. Int J Comput
Meth 5(1):135-188.
387.Xu QX, Shen RY (2008) Fluid-structure interaction of hydrodynamic damper during the rush into the
water channel. J Hydrodyn 20(5):583-590.
388.Ataie-Ashtiani B, Shobeyri G (2008) Numerical simulation of landslide impulsive waves by
incompressible smoothed particle hydrodynamics. Int J Numer Meth Fl 56(2):209-232.
389.Scanlon BR, Tyler SW, Wierenga PJ (1997) Hydrologic issues in arid, unsaturated systems and
implications for contaminant transport. Rev Geophys 35(4):461-490.
390.Nativ R, Adar E, Dahan O, Geyh M (1995) Water recharge and solute transport through the vadose
zone of fractured chalk under desert conditions. Water Resour Res 31(2):253-261.
391.Schwartz LM, Martys N, Bentz DP, Garboczi EJ, Torquato S (1993) Cross-property relations and
permeability estimation in model porous media. Phys Rev E 48(6):4584-4591.
392.Huang H, Meakin P, Liu MB (2005) Computer simulation of two-phase immiscible fluid motion in
unsaturated complex fractures using a volume of fluid method. Water Resour Res. W12413,
DOI:10.1029/2005WR004204.
393.Huang H, Meakin P, Liu MB, McCreery GE (2005) Modeling of multiphase fluid motion in fracture
intersections and fracture networks. Geophys Res Lett 32. L19402, DOI:10.1029/2005GL023899.
394.Snyder LJ, Stewart WE (1966) Velocity and pressure profiles for Newtonian creeping flow in regular
packed beds of spheres. AIChE J 12(1):167-173.
395.Sussman M, Smereka P, Osher S (1994) A level set approach for computing solutions to
incompressible two-phase flow. J Comput Phys 114(1):146-159.
396.Hirt CW, Nichols BD (1981) Volume of fluid (VOF) method for the dynamics of free boundaries. J
Comput Phys 39(1):201-225.
397.Unverdi SO, Tryggvason G (1992) A front-tracking method for viscous, incompressible, multi-fluid
flows. J Comput Phys 100(1):25-37.
398.Zhu Y, Fox PJ (2001) Smoothed particle hydrodynamics model for diffusion through porous media.
Transport Porous Med 43(3):441-471.

96