See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/222426965

Stable particle methods based on Lagrangian kernels

Article in Computer Methods in Applied Mechanics and Engineering · March 2004

DOI: 10.1016/j.cma.2003.12.005

CITATIONS READS
446 362

3 authors, including:

Timon Rabczuk Shaoping Xiao

Bauhaus-Universität Weimar University of Iowa
856 PUBLICATIONS 48,971 CITATIONS 82 PUBLICATIONS 4,555 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Shaoping Xiao on 16 January 2019.

The user has requested enhancement of the downloaded file.

 Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063
www.elsevier.com/locate/cma

Stable particle methods based on Lagrangian kernels

T. Rabczuk, T. Belytschko *, S.P. Xiao
Department of Mechanical Engineering, Northwestern University, 2145 Sheridan Road, Evanston, IL 60208-311, USA
Received 26 March 2003; received in revised form 20 September 2003; accepted 2 December 2003

Abstract

A large deformation particle method based on the Krongauz–Belytschko corrected-gradient meshfree method with
Lagrangian kernels is developed. In this form, the gradient is corrected by a linear transformation so that linear
completeness is satisfied. For the test functions, Shepard functions are used; this guarantees that the patch test is met.
Lagrangian kernels are introduced to eliminate spurious distortions of the domain of material stability. A mass allo-
cation scheme is developed that captures correct reflection of waves without any explicit application of traction
boundary conditions. In addition, the Lagrangian kernel versions of various forms of smooth particle methods (SPH),
including the standard forms and the Randles–Libersky modification are presented and studied. Results are obtained
for a variety of problems that compare this method to standard forms of SPH, the Randles–Libersky correction and
large deformation versions of the element-free Galerkin method.

2004 Elsevier B.V. All rights reserved.

Keywords: Meshfree methods; Particle methods; EFG; SPH

1. Introduction

Meshfree particle methods provide many advantages for modelling severe deformations and failure of
solids as compared to Lagrangian finite element and finite difference methods. Although most meshfree
methods are basically Lagrangian in character, the absence of a mesh enables them to deal with larger local
distortion than finite element methods. Furthermore, phenomena such as fracture and other material
instabilities are more easily modelled than with finite elements, particularly when they are not aligned with
the nodes.
Particle methods can be classified into those based on kernel approximations, as smooth particle
hydrodynamics (SPH) methods, and those based on field approximations, as the element-free Galerkin
method (EFG). One of the first meshfree methods was the SPH-method introduced by Lucy [32] and
Gingold and Monaghan [20]. It was first utilized for gas dynamic problems. The method was extended to
solid continua by Libersky et al. [28], who applied SPH to dynamic fracture and fragmentation of solids,

*
Corresponding author. Tel.: +1-847-491-4029; fax: +1-847-491-4011.
E-mail addresses: [email protected], [email protected] (T. Belytschko).

0045-7825/$ - see front matter
2004 Elsevier B.V. All rights reserved.
doi:10.1016/j.cma.2003.12.005
1036 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

where mesh-based methods have had difficulties. Fracture was modelled by letting the stresses go to zero
after a critical strain.
However, SPH suffers from some inherent difficulties: a lack of consistency that can lead to poor
accuracy and a tensile instability that results from the combination of an Eulerian kernel with a Lagrangian
description of motion. The latter is more generally a manifestation of the distortion of the domain of
material stability, see Belytschko et al. [6,7,47].
In the last 10 years, many improvements have been made in particle methods. Liu et al. [29,30] proposed
a correction function that restores linear (first-order) completeness of the kernel function. First-order
completeness appears to be a necessary condition for convergence. Furthermore, zeroth-order completeness
guarantees global conservation of linear momentum while first-order completeness guarantees global
balance of angular momentum as shown in Krongauz and Belytschko [27] and Belytschko et al. [2]. Be-
lytschko et al. [4] showed that these correction methods yield approximations that are identical to the
moving least square (MLS) approximation used in EFG [5]. The particle methods based on MLS (or
corrected functions), unlike SPH, do not satisfy local conservation of momentum or angular momentum.
Dilts [15–17] developed SPH methods based on MLS approximations and extensively studied the issue of
local conservation properties.
Johnson and Beissel [24] proposed a modification of the extensional strains of the SPH shape functions,
so that they would give the correct values of the extensional strains for linear fields. Subsequently, Randles
and Libersky [41] proposed a more general transformation of the gradient that gives the correct values of all
strains for linear fields, which they called normalization. Krongauz and Belytschko [27] developed a similar
transformation but showed that a standard Bubnov–Galerkin discretization then failed the patch test. Since
the patch test is apparently necessary for convergence, this may be quite detrimental to the method. They
showed that a Petrov–Galerkin method employing the Shepard functions as test functions and a kernel with
a corrected gradient as trial functions meets the patch test. The failure of corrected gradient methods to
meet the patch test was subsequently noted by Bonet et al. [13] who proposed solving a subsidiary set of
equations to enforce them.
As shown by Swegle et al. [45], particle methods suffer from tensile instabilities. To stabilize SPH, Dyka
et al. [18,19] introduced stress points into SPH. This approach was later extended to higher dimensions by
Randles et al. [42]. It was shown in [6] that the stress point technique stabilizes SPH by suppressing spurious
singular modes observed in SPH, but it does not eliminate the distortion of the domain of material
instability. Chen et al. [14] have developed an integration technique based on contour integrals that avoids
volumetric locking and appears to eliminate spurious singular modes.
In most of the SPH literature an Eulerian kernel is taken as the weighting function. Recently, Belytschko
et al. [6,7] investigated Eulerian and Lagrangian kernels. They pointed out that SPH with stress points and
an Eulerian kernel cannot eliminate the tensile instability, because Eulerian kernels distort the domain of
material instability; they showed that stress points only remove the instability that arises due to rank
deficiency, i.e. spurious singular modes. They also showed that the tensile instability can be avoided by
a Lagrangian kernel. However this limits the magnitude of the distortions that can be treated.
Other meshfree methods are the EFG method which uses the MLS approximation [5]. The main con-
tribution of the EFG was the development of methods for treating arbitrary discontinuities, see Belytschko
et al. [5,8,9]. It was also used for modeling nonlinear material response under dynamic loading, see [9].
Unlike SPH, the EFG method is discretized by a Galerkin method. The integrals are usually evaluated over
background cells based on an octree structure, see [5]. For dynamic applications, cell integration becomes
computationally expensive. Beissel and Belytschko [1] developed a nodal integration for EFG, which
corresponds to a particle method based on MLS.
In this paper, a meshfree particle method for large deformation, nonlinear problems that employs a
Lagrangian kernel with correction of the derivatives, as in Krongauz and Belytschko [27], is described.
Stress points are used to eliminate the rank deficiency. It is also shown that through an appropriate con-
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1037

struction of the mass matrix, the method accurately treats traction-free boundaries and boundaries with
applied tractions, without any special boundary treatment. The method is, however, limited in the amount
of deformation it can handle; it is not suitable for problems of fluid flow. Nevertheless, it appears to be able
to treat many problems of interest in the simulation of solids, where the deformations of interest are less
severe than in fluids. Furthermore, it is stable and robust for this class of problems.
This article is arranged as follows. First, the governing equations are given. After some remarks about
Eulerian and Lagrangian kernels, several meshfree approximations are summarized: the original SPH
method, the Randles–Libersky [41] correction of the derivatives, the Krongauz–Belytschko correction [27]
and the MLS approximation. The discrete governing equations are obtained by collocation as in SPH and
by Petrov–Galerkin and Bubnov–Galerkin procedures, respectively. In Section 5, nodal integration, nodal
integration with stress points and cell integration used in EFG are reviewed. Then, applications to concrete
and metals under quasistatic and dynamic loading are given and compared with experimental data.

2. Governing equations

The governing equations for large deformation problems will first be presented in the total Lagrangian
formulation, where the conservation equations and constitutive equations are expressed in terms of the
material coordinates X. They will then be presented in the updated Lagrangian formulation, where the
equations are described in the current configuration, see Belytschko et al. [10]. We will give both forms of
many of the equations.

2.1. Total Lagrangian formulation

We consider a body X in Rn (n ¼ 2 or 3) with boundary C; their images in the initial state are X0 and C0 ,
respectively. The initial state will also serve as the reference state. The motion is described by
x ¼ /ðX; tÞ; ð1Þ

where x are the spatial (Eulerian) coordinates and X the material (Lagrangian) coordinates. The dis-
placement is then given by
uðX; tÞ ¼ x
X ¼ /ðX; tÞ
X: ð2Þ
Neglecting thermomechanical and frictional forces, the conservation equations in the total Lagrangian
formulation are given by [10]
1 1 _
.J ¼ .0 J0 ; u¼
€ r0  P þ b; e_ ¼ F : PT ; ð3Þ
.0 .0
where J and J0 are the Jacobian determinant and initial Jacobian determinant, respectively, u is the dis-
placement vector, . is the current density, .0 is the initial density, P are the nominal stresses (note P is the
transpose of the first Piola Kirchhoff stress tensor as discussed in Belytschko et al. [10]), b are the body
forces per mass unit, e is the internal energy, r0 is the gradient or divergence operator expressed in material
derivatives and F denotes the deformation gradient. Note that the mass conservation equation is written in
algebraic form since it is integrable for a Lagrangian description.
The boundary conditions are:
uðX; tÞ ¼

uðX; tÞ on Cu0 ; ð4Þ

n0  PðX; tÞ ¼
tðX; tÞ on Ct0 ; ð5Þ

1038 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

u and
t are the prescribed displacement and traction, respectively, n0 is the outward normal to the
where

domain and Cu0 [ Ct0 ¼ C0 , Cu0 \ Ct0 ¼ 0.

2.2. Updated Lagrangian formulation

In the updated Lagrangian formulation the equations are expressed in the current domain of the body. It
should be understood that the governing equations in the total and updated Lagrangian formulation are
identical, they are only different descriptions (or transformations) of the same equations.
Neglecting thermomechanical and frictional forces, the conservation equations in the strong form are
1 r
._ ¼
.r  v; v_ ¼ r  r þ b; e_ ¼ : r  v; ð6Þ
. .
where v is the velocity vector, r is the Cauchy stress tensor, e is the internal specific energy and the r is the
gradient or divergence operator expressed in spatial coordinates. The energy conservation equation is only
necessary as a PDE if heat transfer is relevant. The continuity equation can also be in the form (3a) for an
updated Lagrangian description [10]. The boundary conditions are:
vðX; tÞ ¼
vðX; tÞ on Cv ; ð7Þ

n  rðX; tÞ ¼
tðX; tÞ on Ct ; ð8Þ

where
v and
t are the prescribed velocity and traction, respectively, n is the outward normal to the domain
and Cv [ Ct ¼ C, Cv \ Ct ¼ 0.
In the updated Lagrangian formulation the constitutive equations are commonly formulated in rate
form for large deformations:
rr ¼ rðD;
_ qÞ; ð9Þ
where q are internal state variables which depend on the material behavior, D is the rate of deformation or
velocity strain and rr is a frame invariant rate. The velocity strain is the symmetric part of the velocity
gradient L:
D ¼ 12ðr  x_ þ x_  rÞ: ð10Þ
The antisymmetric part of L
W ¼ 12ðr  x_
x_  rÞ ð11Þ
is the spin-tensor and is needed to calculate certain frame-indifferent rates.

3. Meshfree approximations

Meshfree approximations for a scalar function u in terms of the material (Lagrangian) coordinates can
be written as
X
uðX; tÞ ¼ UJ ðXÞuJ ðtÞ; ð12Þ
J 2S

where UJ ðXÞ are the shape functions and uJ is the value at the particle parameter (it is nearly equal to the
displacement) at the position XJ and S is the set of particles J for which UI ðXJ Þ 6¼ 0. Note, that the above
form is identical to the form of an FEM approximation. However, in contrast to FEM, UI ðXÞ is only an
approximant and not an interpolant, since uðXI Þ 6¼ uI . Therefore special techniques are needed to treat
displacement boundary conditions, see e.g. [2,17].
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1039

The approximation can also be formulated in terms of the spatial (Eulerian) coordinates. In that case,
it is conventional to approximate the velocities, instead of the displacements, so
X
_ tÞ ¼
uðx; UJ ðxÞu_ J ðtÞ: ð13Þ
J 2S

Eqs. (12) and (13) correspond to Lagrangian and Eulerian descriptions of motion, respectively. Note,
that (13) does not correspond to the standard description of motion for a solid as given in (1). Furthermore,
if all dependent variables such as the stress, strain and state variables are expressed in terms of the spatial
coordinates x and time t, then transport terms must be included in the update equations. The shape
functions UJ ðXÞ or UJ ðxÞ are obtained from the kernel functions, which are denoted by WJ ðXÞ or WJ ðxÞ for
Lagrangian and Eulerian kernels, respectively. Usually radial kernel functions are chosen. For Lagrangian
radial kernels WJ ðXÞ ¼ W ðr0 Þ where r0 ¼ kX
XJ k while for Eulerian kernels WJ ðx; tÞ ¼ W ðrÞ where
r ¼ kx
xJ ðtÞk. As noted later, the time dependence of Eulerian kernels is always neglected, see Monaghan
[33,34].
The kernel functions are chosen to have compact support. A widely used kernel which has been em-
ployed in our calculations is the cubic spline [14]:
8C
>
> ð1
1:5z2 þ 0:75z3 Þ 0 6 z < 1;
>
< hD
W ðrÞ ¼ C 3 ð14Þ
>
> ð2
zÞ 1 6 z 6 2;
> 4hD
:
0 z > 2;

where D is the dimension of the space, h is the interpolation radius or radius of the support, z ¼ r=h, and
C is a constant depending on the dimension
8
< 2=3 for D ¼ 1;
C ¼ 10=ð7pÞ for D ¼ 2; ð15Þ
:
1=p for D ¼ 3:
The support of the kernel function is also called the domain of influence since it determines which
neighboring particles effect a given particle and corresponds to S. The relationship of the kernel function to
the shape function depends on the approximation technique and is explained in the Sections 3.2–3.5.

3.1. Eulerian and Lagrangian kernels

In most SPH procedures [15–17,41,42] the shape function is directly expressed in terms of an Eulerian
kernel:
WJ ðxÞ ¼ W ðx
xJ ðtÞ; hðx; tÞÞ: ð16Þ
Note that the Eulerian kernel is expressed in terms of spatial coordinates. The radius h of the support can
depend on the spatial coordinates.
Belytschko et al. [6] have shown that particle discretizations of solids with an Eulerian kernel with or
without stress points lead to a distortion of the stable domain of the material in stress space; the tensile
instability analyzed in Swegle et al. [46] is one manifestation of this distortion.
The Lagrangian kernel is expressed in terms of material coordinates, so:
WJ ðXÞ ¼ W ðX
XJ ; h0 Þ: ð17Þ
For Lagrangian kernels, the neighbors of influence do not change during the course of the simulation,
but the domain of influence in the current configuration changes with time. For radial kernel functions, the
1040 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

domain of influence in the initial configuration is a circle in two dimensions and a sphere in three
dimensions.
The Lagrangian kernel eliminates the tensile instability, but the instability due to a rank deficiency of the
discrete equations remains unless stress points are added. Therefore, the representation of the kernel in
terms of material coordinates provides a more consistent procedure when simulating material fracture or
other material instabilities because instabilities will not occur due to numerical artifacts.
The distortion of material instability is illustrated for the case of a particular material in Fig. 1 [47],
which shows the domains of material stability for the Lagrangian kernel, an Eulerian kernel and that of the
governing partial differential equation, i.e. the momentum equation (3) for a particular hyperelastic
material. The stable domains are shown in the space of the two principal stretches, k1 and k2 . The hy-
perelastic material model is given by [25]:
1 h  1 2
i
r ¼ pffiffi ðc1 þ c2 I1 ÞB
c2 B2
c1 I33 þ 2c2 I33
k ln I3 I ; ð18Þ
I3
where B ¼ FFT . The material constants are c1 ¼ 1:265e5 N/m2 , c2 ¼ 1:012e4 N/m2 and k ¼ 1:012e7 N/m2 .
I1 ¼ trðBÞ and I3 ¼ detðBÞ. The material density is q ¼ 125:4 kg/m3 . In an ideal discretization, the stable
domains of the discretization and the PDE should coincide, which is almost the case for the Lagrangian
kernel (in fact, the two domains are almost indistinguishable in the figure, but differ numerically). However,
for the Eulerian kernel we can see a substantial domain (indicated by grayscale in Fig. 1) where the material
should be stable but the discrete model is not. In other words, the shaded area are states of deformation for
which the Eulerian kernel is unstable whereas the PDE indicates a stable response.
In the current configuration the domain of influence of a Lagrangian kernel can be extremely distorted.
This is a disadvantage in simulating fluid flow problems and other problems with very large distortions. In
the following sections the approximation techniques are described in terms of a Lagrangian kernel. The
equations for an Eulerian kernel can be derived similarly.

3.2. The SPH approximation

In the SPH method with a Lagrangian kernel, the shape functions are given by a product of the particle
volume and the weighting function:

Lagrangian kernel
3 Eulerian kernel
3 PDE
PDE

2
λ2

λ2

1
1

0
1 2 3 0 1 2 3
(a) λ1 (b) λ1

Fig. 1. (a) Stable domain for MLS particle method with stress point integration and Lagrangian kernel compared to the stable domain
for the PDE; (b) Stable domains of MLS particle methods for stress point integration with Eulerian and Lagrangian kernel for hy-
perelastic material; dashed and solid lines bound the stable domains for Lagrangian and Eulerian kernels, respectively.
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1041

UJ ðXÞ ¼ W ðX
XJ ; h0 ÞVJ0 ; ð19Þ

where VJ0is the volume associated with the particle J in the initial configuration. For an Eulerian kernel the
shape functions are
UJ ðxÞ ¼ W ðx
xJ ; hÞVJ ; ð20Þ
where VJ is the current volume associated with particle J .
In the application to PDEs, an approximation of a function gradient is needed. In SPH, an approxi-
mation of the gradient of a function in terms of the Lagrangian kernel is given by
X
ruh ¼
uJ r0 UJ with r0 UJ ¼ r0 W ðX
XJ ; h0 ÞVJ0 ; ð21Þ
J 2S

where the r0 indicates the gradient in terms of the material coordinates. The minus sign on the right-hand
side of Eq. (21) results from the integration by parts, that is standard in SPH [33,34].
One drawback of the standard SPH-method is its inability to accurately approximate a function when
particles are unevenly spaced. This is evidenced by its inability to reproduce linear fields (called the
reproducing conditions), which is generally necessary for convergence. The momentum equations with
stress rates depending on the velocity gradient are second-order PDEs, so linear completeness, i.e. the
ability to reproduce linear functions, is necessary for convergence. These are among the properties checked
by the well known patch test. For boundary particles, the constant reproducing conditions are violated even
for uniformly spaced particles.
The reproducing conditions (or completeness) of an approximation correspond to the order of the
polynomial which can be represented exactly. The following are the conditions for zeroth- and first-order
completeness in R3 :
X X
UJ ðXÞ ¼ 1 r0 UJ ðXÞ ¼ 0; ð22Þ
J 2S J 2S

X X X
UJ ðXÞXJ ¼ X UJ ðXÞYJ ¼ Y UJ ðXÞZJ ¼ Z: ð23Þ
J 2S J 2S J 2S

Approximations that satisfy (22) have zeroth-order completeness. Since the approximation (21) does not
fulfill zeroth-order completeness on the boundary even for a regular particle configuration, Monaghan [34]
introduced a so called symmetrization. In this procedure he assumed that
!
X
0
r0 W ðXI
XJ ; h0 ÞVJ uI  0; ð24Þ
J 2S

although this is only true for a uniform distribution of particles away from any boundary. Note that (24) is
equivalent to assuming zeroth-order completeness (note the quantity inside the parenthesis in (24)). Adding
Eq. (24) to Eq. (21) gives
X
ruh ðXI Þ ¼
ðuJ
uI Þr0 W ðXI
XJ ; h0 ÞVJ0 : ð25Þ
J 2S

A remarkable feature of the symmetrization procedure is that it yields zeroth-order completeness for the
derivatives of a function for an irregular particle arrangement.

3.3. Krongauz–Belytschko correction

Johnson and Beissel [24] noted that errors in the extensional strains due to lack of linear completeness
could be corrected by simple scaling, thus improving accuracy. A correction that enables the derivatives of
1042 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

constant or linear fields to be reproduced exactly was developed by Randles and Libersky [41] and
Krongauz and Belytschko [27]. We extend those ideas here to large deformations. The corrected derivatives
are approximated by
X
r0 uh ðX; tÞ ¼ GI ðXÞuI ðtÞ; ð26Þ
I2S

where GI is a linear combination of the derivatives of the Shepard functions. GI is obtained by a linear
transformation

GI ðXÞ ¼ aðXÞ  r0 WIS ðXÞ; ð27Þ

where aðXÞ are arbitrary parameters and WIS ðXÞ are the Shepard functions given by
WI ðXÞ
WIS ðXÞ ¼ P : ð28Þ
I2S WI ðXÞ

Note that aðXÞ is invariant with respect to time for a Lagrangian kernel, so they need only be computed
once and stored. The following reproducing condition for the derivative of a linear function must be ful-
filled:
X
GI ðXÞ  XI ¼ I; ð29Þ
I2S

where I is the identity matrix. Eq. (29) can be combined with (27) to yield
X
aðXÞ  r0 WIS ðXÞ  XI ¼ I: ð30Þ
I2S

If we let A be the outer product of the derivatives of the Shepard function and X
2 S S S
3
WI;X XI WI;Y XI WI;Z XI
6 S 7
A¼6 S S 7
4 WI;X YI WI;Y YI WI;Z YI 5
S S S
WI;X ZI WI;Y ZI WI;Z ZI
then (30) can be rewritten as
AaT ¼ I; ð31Þ
where
2 3
aXX aXY aXZ
a ¼ 4 aYX aYY aYZ 5:
aZX aZY aZZ
Thus, the parameters a can easily be determined from the set of three linear algebraic equation (31). Finally,
we obtain the approximation for the derivatives of a function from (27) and (29) as
X
r0 uh ðX; tÞ ¼ aðXÞ  r0 WIS ðXÞuI ðtÞ: ð32Þ
I2S

Since only the derivatives of the approximation are modified they are usually not integrable. Note that
for Lagrangian kernels, the correction parameters a only need to be computed once.
Krongauz and Belytschko [27] showed that correction of the derivatives for both the test functions and
the trial functions lead to a violation of the patch test. However, they showed that the patch test can be
satisfied if Shepard functions are the test functions in a Petrov–Galerkin method. In this case, the test
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1043

functions do not meet linear completeness, so that global angular momentum is not conserved. However, in
linear problems they observed excellent convergence. This procedure will herein be extended to large
deformation problems.

3.4. Randles–Libersky correction

Randles et al. [41] developed a similar correction which they called normalization (NSPH). We develop
it here for a Lagrangian kernel, which is a straightforward extension of their work. To fulfill the first-
order completeness, they modified the SPH approximation for the gradient of the function u with a
matrix B:
!
X
h 0
ru ðXI ; tÞ ¼
ðuJ ðtÞ
uI ðtÞÞr0 W ðXI
XJ ; h0 ÞVJ  BðXI Þ ð33Þ
J 2S

with
!
1
X
BðXÞ ¼
ðXJ
XÞ  r0 W ðX
XJ ; h0 ÞVJ0 : ð34Þ
J 2S

A B-matrix that correctly gives the derivatives of linear fields can also be considered using the Shepard
function W S instead of the symmetrization in (34). The expression for B becomes:
!
1
X
S
BðXÞ ¼
XJ  r0 W ðX
XJ ; h0 Þ ; ð35Þ
J 2S

which is similar to the Krongauz–Belytschko correction [27]. There are subtle differences between (34) and
(35). The approximation for the gradient of the function u can then be formulated in the unsymmetrized
form because the Shepard functions are zeroth-order complete by construction:
!
X
h S 0
ru ðX; tÞ ¼
uJ ðtÞr0 W ðX
XJ ; h0 ÞVJ  BðXÞ: ð36Þ
J 2S

3.5. The MLS-approximation

The MLS-approximation fulfills the reproducing conditions by construction, so no corrections are

needed. The MLS approximation was introduced in the EFG-method by Nayroles et al. [36] and Bely-
tschko et al. [5,31], and was first applied in a SPH setting by Dilts [16]. To satisfy the linear reproducing
conditions linear base functions p are chosen to be

pðXÞ ¼ ð1 X Y ZÞ 8X 2 R3 : ð37Þ

The MLS approximation is

X
uh ðX; tÞ ¼ pJ ðXÞaðX; tÞ ð38Þ
J 2S

with a chosen to minimize

X
J¼ ðpJ ðXÞT aðX; tÞ
uJ ðtÞÞ2 W ðX
XJ ; hÞ: ð39Þ
J 2S
1044 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

Minimizing Eq. (39) with respect to a leads to the approximation

X
uh ðX; tÞ ¼ uJ ðtÞUJ ðXÞ ð40Þ
J 2S

with
T
1
UJ ¼ pðXÞ  AðXÞ  pJ ðXÞW ðX
XJ ; hÞ; ð41Þ
X
AðXÞ ¼ pJ ðXÞpTJ ðXÞW ðX
XJ ; hÞ: ð42Þ
J 2S

For zeroth-order completeness (pðXÞ ¼ ½1) the shape function U is

WJI
UJ ðXÞ ¼ P ; ð43Þ
J 2S WJI

which is known as the Shepard function where WJI ¼ W ðXJ
XI ; hÞ. A fast evaluation procedure for the
gradient of the shape function UJ can be found in Belytschko et al. [3].

4. Discrete equations

4.1. Galerkin methods

4.1.1. The total Lagrangian formulation

In EFG, the discrete equations are obtained from a weak form of the momentum equation, which is
given in the total Lagrangian description by
Z Z Z Z
T
.0 du  €
u dX0 þ ðr0 duÞ : P dX0
.0 du  b dX0
du 
t dC ¼ 0 ð44Þ
X0 X0 X0 Ct0

8du 2 V0 where V0 is the space of test functions and u 2 V where V is the space of trial functions. The spaces
V0 and V are as follows:
V ¼ ðuðX; tÞjuðX; tÞ 2 H 1 ðXÞ; uðX; tÞ ¼

u on Cu Þ; ð45Þ

V0 ¼ V \ ðduðXÞjduðXÞ ¼ 0 on Cu Þ: ð46Þ
The test and the trial functions are approximated by
X
duh ðXÞ ¼ UJ ðXÞduJ ; ð47Þ
J 2S

X
uh ðX; tÞ ¼ WJ ðXÞuJ ðtÞ: ð48Þ
J 2S

In a Petrov–Galerkin procedure, different approximating functions duh and uh are chosen, so UJ differ
from WJ . We will use MLS shape functions (38) for the trial functions W and Shepard functions for the test
functions U. Furthermore, an assumed strain approach [44] is employed so that the gradient is approxi-
mated by (32). However, the formulation differs from [44] in that the gradient of the test function (not the
assumed gradient) is used in computing the internal forces; Black and Belytschko [12] have given a con-
vergence proof for this class of methods for the Laplace equation. Substituting (47) and (48) into (44) the
discrete form of the momentum equation is obtained
X
uJ ¼ f ext
mIJ € int
I
fI ; ð49Þ
J
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1045

where f ext
I and f int
I are the external and the internal forces, given by
Z Z
f ext
I ¼ . U
0 I b dX 0 þ UI
t dC0 ; ð50Þ
X0 Ct0

Z
f int
I ¼ r0 UI  P dX0 ð51Þ
X0

and
XZ
mIJ ¼ .0 WI ðXÞUJ ðXÞ dX0 : ð52Þ
I2S X0

The above mass matrix is the consistent mass matrix. It can be diagonalized by a row sum method (see
e.g. [10]). However, in particle methods the diagonal mass is usually obtained by an ad hoc procedure. It is
shown in Belytschko et al. [2] that if the test and the trial functions satisfy the linear reproducing conditions,
then global linear and angular momentum are conserved. Bonet et al. [13] have also demonstrated this
property.
As mentioned before, the standard derivation of SPH of the acceleration omits the time dependence of
xJ ðtÞ in (16). If we let
X
vðx; tÞ ¼ W ðx
xI ðtÞÞvI ðtÞ; ð53Þ
I2S

then the accelerations are

X
a¼ W ðx
xI ðtÞÞ_vI þ rW ðx
xI ðtÞÞx_ I  vI : ð54Þ
I2S

The second term is neglected in all versions of SPH that we know of, so it is also dropped in all cal-
culations reported here.

4.1.2. The updated Lagrangian formulation

In the updated Lagrangian method, the weak form of the momentum equation is
Z Z Z

ðrdvÞT : r dX þ dv  ð.b
._vÞ dX þ dv 
s dC ¼ 0: ð55Þ
X X Ct

Using the same procedure as for the total Lagrangian formulation, the discrete linear momentum
equation for a Petrov–Galerkin method is
Z Z Z XZ dvI

rUJ  r dX þ UJ
s dC þ .UJ b dX ¼ .UJ ðxÞWI ðxÞ dX : ð56Þ
X Ct X I X dt

4.2. Collocation methods

In SPH, the discrete equations are obtained from the strong form by collocation, i.e. the governing
equations are enforced at each particle. Neglecting body forces and boundary conditions, the momentum
equation with the Lagrangian kernel is
X
u ¼ VI0
mI € VJ0 r0 WIJ  PJ : ð57Þ
J 2S
1046 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

In Eulerian coordinates the momentum equation is

X
mI v_ ¼ VI VJ rWIJ  rJ : ð58Þ
J 2S

Usually the symmetrization given in Section 3.2 is applied to SPH methods. When the supports of all
particles are of the same size, the above discrete SPH equations are equivalent to the discrete equations
obtained from a nodal integration of the Galerkin form, except at the boundaries.

5. Integration schemes

5.1. Nodal integration

An efficient way to discretize the momentum equation in a Galerkin method is by nodal integration. In
most computations of transient phenomena spatial integration by background cells is quite expensive. To
obtain the discrete governing equations from the weak form, Beissel and Belytschko [1] proposed a nodal
integration:
Z X
f ðXÞ dX0  f ðXJ ÞVJ0 ; ð59Þ
X0 J 2S

where VJ0 is the initial particle volume. Applying this technique to the computation of the internal forces
(51) gives:
X
f int
I ¼ VJ0 r0 UI ðXJ Þ  PJ : ð60Þ
J 2S

The momentum equation is then

du_ I X
mI ¼ ð
r0 UI ðXJ Þ  PðXJ ÞVJ0 þ UI ðXJ Þ
sðXJ ÞC0J þ UI ðXJ ÞbðXJ Þ.ðXJ ÞVJ0 Þ: ð61Þ
dt J 2S

In Eulerian coordinates the momentum equation is

dvI X
mI ¼ ð
rUI ðxJ Þ  rðxJ ÞVJ þ UI ðxJ Þ
sðxJ ÞCJ þ UI ðxJ ÞbðxJ Þ.ðxJ ÞVJ Þ: ð62Þ
dt J 2S

5.2. Stress point integration

Stress point integration was proposed by Dyka et al. [18,19] to stabilize the SPH method. Randles et al.
[42] extended stress point integration to higher dimensions to stabilize the normalized form of SPH. Stress
point integration eliminates instabilities due to rank deficiency but not those due to the tensile instability,
see Belytschko et al. [6].
In stress point integration methods, stress points are interspersed between the particles and the contri-
butions of the stresses are added to the integration of (59) as described later. There are different methods for
arranging the stress points between the original particles, see Belytschko et al. [6]. In the applications in this
article, stress points are added as shown in Fig. 2. The additional stress points are only needed for the
computation of the internal forces. All kinematic values at the stress points such as displacements and
velocities are obtained from the original particles:
X X X
uSP
I ¼ UJ ðXSP P
I ÞuJ ; vSP
I ¼ UJ ðXSP P
I ÞvJ ; r0 vSP
I ¼ GJ ðXSP P
I ÞvJ ; ð63Þ
J 2S J 2S J 2S
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1047

Fig. 2. Arrangement of the stress points.

master particles
slave particles

Fig. 3. Scheme of stress point integration.

where the superscript SP indicate stress points and the P the original particles. UJ ðXSP
I Þ is the shape function
value of the supporting master particle J at XSP I . Note that the displacements are approximated by the trial
functions, whereas the gradient is obtained by the corrected derivatives.
The internal forces are calculated by
X X
f int
I ¼ VJ0P r0 UI ðXPJ Þ  PðXPJ Þ þ VJ0SP r0 UI ðXSP SP
J Þ  PðXJ Þ; ð64Þ
J 2S J 2S SP

where S is the set of supporting stress points to the particle XPI . The volumes VJ0P and VJ0SP are computed
SP

from a Voronoi diagram (see Fig. 3) so that their sums give the total initial volume:
X
NP X
NS
V0 ¼ VJ0P þ VJ0SP ; ð65Þ
J J

where NP is the number of particles and NS is the number of stress points.

5.3. Cell integration

In EFG the integrals are usually evaluated over background cells based on an quadtree structure [5]. In
each integration cell Gauss quadrature is performed. The number of Gauss points depends on the number
of particles in a cell. Here, as in [5], at least nQ  nQ Gauss points are used with
1048 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

pffiffiffiffi
nQ ¼ m þ 2; ð66Þ
where m denotes the number of particles in one cell.
In 2D the integral of a function is given by
Z Z þ1 Z þ1 X
m
f ðXÞ dX0 ¼ f ðn; gÞdet Jn ðn; gÞ dn dg ¼ wJ f ðnJ Þdet Jn ðnJ Þ; ð67Þ
X0
1
1 J ¼1

where n ¼ ðn; gÞ are scaled local coordinates, m is the total number of quadrature points, wJ ¼ wðnJ ÞwðgJ Þ
are the quadrature weights which is the product of the weight at the corresponding Gauss point in n- and
g-direction, and Jn is the Jacobian given by
oX
Jn ¼ : ð68Þ
on
The internal forces in 2D are then
X
m
f int ¼ wJ det Jn ðnJ Þr0 UðXðnJ Þ
XP ÞPðnJ Þ; ð69Þ
J ¼1

where the superscript P indicates the particle position.

6. Computation of the mass matrix

In dynamic problems with explicit time integration, a diagonal mass matrix is preferable for compu-
tational efficiency. A diagonal mass corresponds to assigning a mass to every particle. The masses are
calculated in terms of nodal volumes obtained from a Voronoi diagram:
mJ ¼ VJ0 .0J : ð70Þ
Crucial is how to consider and discretize the ÔrealÕ geometric volume of a body. It can be discretized with
particles arranged as in Fig. 4(a) where the particles are inside the volume. No particles lie on the
boundaries of the body. The body can also be discretized as illustrated in Fig. 4(b) where particles are
placed exactly on the boundaries. For Fig. 4(a), all masses are equal; this leads to poor reflection conditions
on boundaries. For the discretization in Fig. 4(b), the boundary particles have only half of the mass of the
interior particles for a Voronoi method. The masses of the particles at the corners are only a quarter of the
masses of the interior particles. We will show that this assignment of nodal masses leads to more accurate
wave reflections.
The boundary integrals for natural boundary conditions (external forces) (50) also differ for the two
models shown in Fig. 4. When the particles are located directly on the boundary, the value of the shape
function differs from when the particles are at a certain distance from the boundaries (see Fig. 4).

7. Boundary conditions

In standard SPH, no conditions are imposed on stress free boundaries. This is sometimes considered
justifiable because of compensating errors between the desired boundary conditions and errors in the
boundary derivatives. It can also be justified by the integration by parts in the SPH derivation underlying
(21) and the expression for the nodal forces, which is similar to the procedure in deriving a Galerkin weak
form from the strong form. However, Randles and Libersky reported that SPH with corrected derivatives
and no special treatment of boundary conditions led to errors. Traction boundary conditions in SPH are
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1049

(a) (b)

Fig. 4. Two arrangements of particles relative to a boundary and the corresponding shape functions.

investigated by Dilts [15,17] and Randles et al. [42]. We will show that the procedure developed here
correctly treats traction-free boundaries.

8. A summary of particle methods

The standard SPH-procedure, a Randles–Libersky correction,a Krongauz–Belytschko correction and an

MLS-approach have been reviewed. Many different discretizations can be constructed starting from these
approximations by using a Petrov–Galerkin or Bubnov–Galerkin method or by point collocation (SPH). In
the following we consider:

(1) The standard SPH-procedure with collocation (SPH).

(2) The Randles–Libersky correction with collocation (RLM ni).
(3) The Krongauz–Belytschko correction and Shepard functions in a Petrov–Galerkin method with nodal
integration and stress points (KBM si).
(4) The EFG method with nodal integration (EFG ni).
(5) The EFG method with cell integration (EFG ci).

The equations for the internal forces and the deformation gradient are summarized in Table 1 for a
Lagrangian kernel and in Table 2 for an Eulerian kernel. For the constitutive model in terms of the
Eulerian kernel the velocity gradient L is given instead of the deformation gradient. As mentioned in
Section 3.2, the equations can also be formulated in symmetrized versions. This is done for the SPH and the
Randles–Libersky correction (RLM) shown in Tables 1 and 2, where PI and rI , respectively, are the stresses
at the central particle.

9. Numerical results

9.1. Cantilever beam

In this section the cantilever beam illustrated in Fig. 5 is considered. Linear elastic material behavior is
assumed. The beam is loaded with a parabolic traction at the end of the beam as shown in Fig. 5. The
1050 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

Table 1
Discrete internal forces and deformation gradient for a Lagrangian kernel for various methods
Internal forces Deformation gradient
P P
SPH f int
I ¼ J 2S ðPJ þ PI Þ  r0 WJ VJ
0
FI ¼
J 2S ðuJ
uI Þ  r0 WJ VJ0
P oWJ 0 P oWJ 0
fiIint ¼ J 2S ðPij ðXJ Þ þ Pij ðXI ÞÞ
V FijI ¼
J 2S ðuiJ
uiI Þ V
oXj J oXj J
 P   P 
RLM (ni) f int
I ¼
J 2S ðPJ þ PI Þ  r0 WJ VJ
0
:B FI ¼
J 2S ðuJ
uI Þ  r0 WJ VJ0  B
   
P oWJ 0 P oWJ 0
fiIint ¼
J 2S ðPij ðXJ Þ þ Pij ðXI ÞÞ VJ Bkj FijI ¼
J 2S ðuiJ
uiI Þ VJ Bkj
oXk oXk
int P P P
KBM (si) f I ¼ J 2S r0 WI ðXJ Þ  PðXJ ÞVJ þ J 2S SP r0 WIS ðXSP
S P P 0P SP
J Þ  PðXJ ÞVJ
0SP P
F ¼ J 2S uJ  GðXJ Þ P
P
FSP ¼ J 2S SP uJ  GðXSP J Þ
P oWIS ðXPJ Þ X oW S ðXSP Þ P
fiIint ¼ J 2S Pji ðXPJ ÞVJ0P þ I J
Pji ðXSP
J ÞVJ
0SP P
FijP ¼ J 2S uiJ Gj ðXJ Þ
oXj J 2S SP
oXj P
FijSP ¼ J 2S SP uiJ Gj ðXSP
J Þ
P P
EFG (ni) f int
I ¼ J 2S PJ  r0 UI ðXJ ÞVJ0 FI ¼
J 2S ðuJ
uI Þ  r0 UI ðXJ Þ
P oUI ðXJ Þ 0 P oUI ðXJ Þ
fiIint ¼ J 2S Pij ðXJ Þ VJ FijI ¼
J 2S ðuiJ
uiI Þ
oXj oXj
Pm PNP
EFG (ci) f int
I ¼ J ¼1 wJ det JnJ PJ  r0 UðXJ
XPI Þ FI ¼ J ¼1 uJ  r0 UðXGP
I
XJ Þ

Pm oUðXJ
XPI Þ PNP oUðXGP

I
XJ Þ
fiIint ¼ J ¼1 wJ det JnJ Pij ðXJ Þ FijI ¼ J ¼1 uiJ
oXj oXj

WJ ¼ W ðXI
XJ Þ; UI ðXJ Þ ¼ UðXI
XJ Þ; PJ ¼ PðXJ Þ, etc.

SPH––symmetrized SPH with collocation
RLM (ni)––Randles–Libersky correction with nodal integration
KBM (si)––Krongauz–Belytschko correction with stress point integration
EFG (ni)––EFG with nodal integration, EFG (ci)––EFG with cell integration

analytic solution for this problem can be found in [21]. The five methods summarized in Section 8 are
considered. The normalized L2 error in the displacement as well as the error in the energy is checked. The
normalized L2 error in the displacement is computed by
kuh
uanalytic kL2
kerrkL2 ¼ ð71Þ
kuanalytic kL2
with
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Z ffi
kukL2 ¼ u  u dX0 : ð72Þ
X0

The error in the energy is obtained by

kuh
uanalytic kenergy
kerrkenergy ¼ ð73Þ
kuanalytic kenergy
with
Z
kukenergy ¼ P : FT dX0 : ð74Þ
X0

The integrals are evaluated numerically by the procedure that is used to evaluate the nodal forces, i.e. by
Gauss quadrature in EFG. The results are presented in Fig. 6. The best results are obtained with EFG with
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1051

Table 2
Discrete internal forces and deformation gradient for an Eulerian kernel for various methods
Internal forces Velocity gradient
P P
SPH f int
I ¼ J 2S ðrJ þ rI Þ  rWJ VJ LI ¼
J 2S ðvJ
vI Þ  rWJ VJ
P oWJ P oWJ
fiIint ¼ J 2S ðrij ðxJ Þ þ rij ðxI ÞÞ VJ LijI ¼
J 2S ðviJ
viI Þ VJ
oxj oxj
P  P 
RLM (ni) f int
I ¼ ð
J 2S ðrJ þ rI Þ  rWJ VJ Þ : B LI ¼
J 2S ðvJ
vI Þ  rWJ VJ  B
   
P oWJ P oWJ
fiIint ¼
J 2S ðrij ðxJ Þ þ rij ðxI ÞÞ VJ Bkj LijI ¼
J 2S ðviJ
viI Þ VJ Bkj
oxk oxk
P P P
KBM (si) f int
I ¼ J 2S rWIS ðxPJ Þ  rðxPJ ÞVJP þ J 2S SP rWIS ðxSP SP
J Þ  rðxJ ÞVJ
SP
LP ¼ r  vP ¼ J 2S vJ  GðxPJ Þ
P
LSP ¼ r  vSP ¼ J 2S SP vJ  GðxSP
J Þ
P oWIS ðxPJ Þ X oW S ðxSP Þ P
fiIint ¼ J 2S rji ðxPJ ÞVJP þ I J
rji ðxSP
J ÞVJ
SP
LPij ¼ J 2S viJ Gj ðxPJ Þ
oxj oxj
J 2S SP P
LSP
ij ¼ J 2S SP viJ Gj ðxSP
J Þ
P P
EFG (ni) f int
I ¼ J 2S rJ  rUI ðxJ ÞVJ LI ¼
J 2S ðvJ
vI Þ  rUI ðxJ Þ
P oUI ðxJ Þ P oUI ðxJ Þ
fiIint ¼ J 2S rij ðxJ Þ VJ LijI ¼
J 2S ðviJ
viI Þ
oxj oxj
Pm PNP
EFG (ci) f int
I ¼ J ¼1 wJ det JnJ rJ  rUðxJ
xPI Þ LI ¼ J ¼1 vJ  rUðxGP
I
xJ Þ
Pm oUI ðxJ
xPI Þ PNP oUI ðxGP
I
xJ Þ
fiIint ¼ J ¼1 wJ det JnJ rij ðxJ Þ LijI ¼ J ¼1 viJ
oxj oxj

WJ ¼ W ðxI
xJ Þ, UI ðxJ Þ ¼ UðxI
xJ Þ, rJ ¼ rðxJ Þ, etc.

SPH––symmetrized SPH with collocation
RLM (ni)––Randles–Libersky correction with nodal integration
KBM (si)––Krongauz–Belytschko correction with stress point integration
EFG (ni)––EFG with nodal integration, EFG (ci)––EFG with cell integration

D q

Fig. 5. Model for the cantilever beam.

cell integration. The Randles–Libersky and Krongauz–Belytschko corrections combined with a nodal
integration yield poor results. We have not considered the Randles–Libersky method with stress points
which would probably perform like the Krongauz–Belytschko correction with stress points.
SPH does not converge at all; this is a striking result that was also observed in Belytschko et al. [2].
Convergence for SPH can be achieved if the domain of influence is kept constant during refinement. Fig.
7(a) shows the error in the displacement norm, Fig. 7(b) shows the error in the energy norm for SPH and
EFG with cell integration if the domain of influence is kept constant when refining the discretization. Note
1052 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

0 0
10 10
L2 norm of error in the displacement

0.37

Error in the energy norm

-1
10

-1 0.28
1.79 10
-2
10
1.84
1.32
1.34 1.41
-3
10
-2
10
SPH
1.95 SPH RL M (ni)
-4 RL M (ni)
10 KBM (si)
KBM (si) EFG (ni)
EFG (ni) EFG (ci)
EFG (ci)
-5 -3
10 10
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
h h

Fig. 6. L2 norm of error in displacement and error in energy for different particle methods, see Section 8 for nomenclature.

-1
10
L2 norm of error in the displacement

Error in the energy norm

-1
10
-2
10 EFG (ci)
SPH
EFG (ci)
SPH
-3
10
-2
10

-4
10

0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Particle distance Particle distance
(a) (b)

Fig. 7. (a) L2 norm of error in displacement and (b) error in energy for EFG with cell integration and SPH and a constant domain
of influence.

that the EFG results are also more accurate for large R (R is the distance of the particles); compare Figs. 6
and 7. It should be mentioned that for practical reasons the domain of influence cannot be kept constant
during refinement since too many particles will contribute to the sum which increases the computational
cost enormously.
EFG with nodal integration and the Krongauz–Belytschko method with stress point integration provide
pretty good results. The rate of convergence is only a little bit less than for EFG with a cell integration, but
the relative error is larger.

9.2. Linear elastic rod with initial displacement

To investigate these methods for wave propagation and reflection, another simple example is studied: a
linear elastic rod. The major aim here is to examine the effect of the mass lumping described in Section 6.
The rod is discretized in 2D (plane strain). It is 60 mm long and its height is 5 mm. Again, the five methods
summarized in Section 8 are investigated. A rectangular impulse with a velocity of 50 m/s is applied to the
left hand end for 5.8 ms. The same number of particles (number of particles ¼ 19,721) is used in the
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1053

numerical simulation. This correspond to 481 particles in x-direction. The particle separation is 0.125 mm;
YoungÕs modulus E ¼ 210; 000 and PoissonÕs ratio m ¼ 0:0. The L2 error in the velocities is given before and
after the wave reflection:

kvh
vanalytic k
kerrkL2 ¼ : ð75Þ
kvanalytic k

Fig. 8. Velocity field in the rod for a rectangular impulse for different particle methods and integration schemes before the wave
reflection.

Fig. 9. Velocity field in the rod for a rectangular impulse for different particle methods and integration schemes after the second wave
reflection.
1054 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

In Fig. 8 the velocity is shown for the rectangular impulse before the wave reflection. Fig. 9 shows the
velocity after the wave reflects from the left-hand side. The best resolution is obtained by EFG with full
quadrature, but EFG with nodal quadrature is also able to reproduce the analytic velocity field quite well.
For SPH and the Randles–Libersky correction (we did not include the explicit enforcement of traction
boundary conditions that Randles and Libersky advocate), some oscillations before the maximum pressure
peak can be observed, which results in a significantly higher error. It should be mentioned that different
artificial viscosities with different damping factors were tried to keep the wave dispersion as small as
possible.

Fig. 10. Velocity profiles in the rod for an initial velocity condition (Gauss distribution of the velocities) for EFG with cell integration
at different times.

0.9
analytic solution
EFG (ci)
0.8 SPH

0.7

0.6
v[m/s]

0.5

0.4

0.3

0.2

0.1

0
0 10 20 30 40 50 60
X [mm]

Fig. 11. Velocity profile in the middle of the rod for SPH and EFG with cell integration compared to the analytical solution at
t ¼ 0:011563 ms.
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1055

2 pffiffiffiffiffiffiffiffi
Next, an initial velocity of v ¼ eð
aðx
30Þ Þ , a ¼ 0:025 was prescribed. The sound speed is c ¼ E=. with
E ¼ 210; 000 MPa and . ¼ 0:0078 g/mm3 , so the wave returns to its original position at t ¼ 0:011563 ms.
Fig. 10 shows the velocity field at different times for EFG with cell integration. It can be seen how the wave
propagates to the left and right hand ends, is reflected and after 0.011563 ms reaches its original position
again. Fig. 11 illustrates the velocity field at t ¼ 0:011563 ms for a section in the middle of the rod for SPH
and EFG with cell integration compared to the analytical solution. As can be seen, the EFG simulation and
the analytic solution match almost exactly. The wave dispersion is greater for SPH. Also, the maximum
velocity is not reached for the SPH computation. In Table 3, the error in the velocity is given for the
different methods at two times. At t ¼ 0:011563 ms, the wave first reaches its original position, at
t ¼ 0:023126 ms it reaches its original position for the second time. As expected, the best results are ob-
tained with EFG and cell integration; for SPH the error is approximately 5% (see Table 3).
Table 3 gives also the computation times for the different particle methods. Approximately 2500 time
steps were needed for the computation. The computations were performed on an IBM RS6000SP-SMP
with 192 Gflop/s performance. The longest computation is the EFG with cell integration.

9.3. Notched concrete beam in bending

This example is an unreinforced notched concrete beam under a static load as shown in Fig. 12, see
Peterson [38]. The beam is 2 m long and has a rectangular cross section 20 and 10 cm in height and depth,
respectively. In the middle, the beam has a 4 cm wide notch over half the height of the cross section. The
initial elastic modulus is 30,000 MPa and the concrete tensile strength is 3.3 MPa. The constitutive model
for concrete is a combined damage plasticity model as proposed in Rabczuk and Eibl [40]. It was originally
developed by Schmidt–Hurtienne [43] and modified in some points. The material parameters are listed in
Table 4. A Lagrangian kernel was used. The beam was discretized with 396,000 particles. The load was

Table 3
L2 error in the velocities for an initial condition (Gauss distribution of the velocities) for different particle methods before and after the
wave reflection
Error at Error at Number of Number of Number of CPU time
0.0116 ms 0.0231 ms particles str. points cells (min)
SPH (ni) 0.05026 0.04999 20,000 – – 6.2
RLM (ni) 0.01301 0.01297 20,000 – – 7.8
KBM (si) 0.00813 0.00842 10,500 10,200 – 8.4
EFG (ni) 0.008322 0.008232 20,000 – – 7.6
EFG (ci) 0.006817 0.006834 4100 – 15,700 20.1

0.2 m
0.1 m

0.04 m 0.1 m
2.0 m

Fig. 12. Notched unreinforced concrete beam of Peterson [38] under static concentrated loading.
1056 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

Table 4
Material parameters of the constitutive model for concrete
E0 ¼ 30; 000 MPa e0 ¼ 2:5  10
5 ep2 ¼ 0; 4
m ¼ 0:22 ed ¼ 2:2  10
4 . ¼ 2:4  10
3
c1 ¼ 0:0123424 gd ¼ 1:5 ndam ¼ 70
c2 ¼ 0:025166 rt ¼ 1:2 av ¼ 0:7
c3 ¼ 0:782058 rc ¼ 100 bv ¼ 3:5
c4 ¼ 0:346384 cp ¼ 0:9 ev ¼ 0:02
cc ¼ 0:08 ep1 ¼ 1:1  10
3 ev;th ¼ 0:008

applied as a linear function in time in the middle of the beam, so that the strain rate does not exceed
1 · 10
6 /s.
Fig. 13 compares the experimental load–midpoint displacement curves to the simulations with different
numbers of particles. In addition, the influence of the different methods is shown for the fine mesh. It can be
seen that the EFG method provides the best results, but the Krongauz–Belytschko method also reproduces
the experimental data quite well. The SPH results, for a reason we cannot explain, decay much faster than
the experimental results. Nevertheless, the influence of the numerics does not seem to be very high. Overall,
it can be said that the experimental results can be reproduced reasonably well, without spurious discreti-
zation dependence. The results for an Eulerian and Lagrangian kernel gave almost identical results. We did
not observe any tensile instabilities in SPH which is expected because of the low tensile strength of concrete.
For this example we will discuss the computational efficiency of the particle methods. We used 20
processors and a parallelization technique described in [39]. Approximately 12,000 time steps were needed
for the computation. Except for the full Gauss quadrature, the computation time for the different particle
methods do not differ too much when a Lagrangian kernel is used (see Table 5). Since for a Lagrangian
kernel the shape functions are computed only once and a neighbor search is performed only at the
beginning of the computation, the Lagrangian kernels perform better than the Eulerian kernel. For an
Eulerian kernel the neighbor search is performed every thirtieth time step. The most important factor is the

0.9
experiments
0.8 EFG-ni (fine)
EFG -ni (coarse)
KBM-si (fine)
RLM-ni (fine)
0.7 SPH (fine)

0.6
Load [KN]

0.5

0.4

0.3

0.2

0.1

0
0 0.2 0.4 0.6 0.8 1
Middle displacement [mm]

Fig. 13. Load–middle displacement curves of the experiment (the two curves give the range of results) of Peterson [38] compared with
the numerical results.
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1057

Table 5
Computation times for a Lagrangian kernel for 12,000 time steps
Method Number of particles Number of stress points Number of Gauss points Total CPU time (CPU h)
SPH (ni) 396,000 – – 34.8
RLM (ni) 396,000 – – 39.5
KBM (si) 175,000 173,000 – 39.3
EFG (ni) 396,000 – – 38.2
EFG (ci) 43,000 – 340,000 68.7

Table 6
Computation times for an Eulerian kernel for 12,000 time steps
Method Number of particles Number of stress points Number of Gauss points Total CPU time (CPU h)
SPH (ni) 396,000 – – 36.8
RLM (ni) 396,000 – – 45.0
KBM (si) 175,000 173,000 – 48.3
EFG (ni) 396,000 – – 84.0

number of computation of the shape functions, as can be seen from Tables 5 and 6. The CPU time for EFG
with nodal integration is twice that for SPH. Also the CPU-time for the Randles–Libersky and Krongauz–
Belytschko method is somewhat greater but not as severely.

9.4. Shear band propagation in metals

We next consider the simulation of shear band development in metals. The KBM with stress point
integration was used with Eulerian and Lagrangian kernel. Shear band formation is related to material
instabilities, so this is an area where Lagrangian kernels are expected to perform better. A square plate
under compressive load is considered. A viscoplastic constitutive model as described by Needleman [37] was
used. The model description and parameters can be found in Belytschko et al. [11] and in Fig. 14. The plate
is discretized with 250,000 particles. Fig. 16 shows the effective plastic strain (in [%]) in the specimen at

v(t)

v[m/s]

2 mm 30.0

2 mm
time [microsec]

0.1 0.2 0.3 0.4

Fig. 14. Model description for the quadratic plate under compression.
1058 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

1600 1600

stress in load direction [MPa]

stress in load direction [MPa]

1400 1400

1200
1200

1000
1000
800
800
600
600
400
400
200

200
0

0.1 0.2 0.1 0.2

strain in load direction strain in load direction
(a) (b)

1600 1600

stress in load direction [MPa]

stress in load direction [MPa]

1400 1400

1200 1200

1000 1000

800 800

600 600

400 400

200 200

0.1 0.2 1E-04 0.2501 0.5001 0.7501 1.0001

strain in load direction strain in load direction
(c) (d)

Fig. 15. Stress–strain curve of the viscoplastic model with (a) and (b) for an Eulerian and (c) and (d) for a Lagrangian kernel.

Fig. 16. Effective plastic strain of the plate under compressive loading to different time steps.
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1059

different time steps. They are in good agreement with the results obtained by Needleman [37] and Belyt-
schko et al. [11]. Fig. 15 illustrates the stress strain behavior for a particle in the middle of the shear band.
With an Eulerian kernel, the calculations evidences an instability in the softening domain; continuing the
calculations leads to further growth of the oscillations. This instability wasnÕt observed for a Lagrangian
kernel with stress points, see Fig. 15(c) and (d), although some slightly increasing oscillations are apparent.

9.5. Shear bands in rings

Here we will examine the performance of Lagrangian and Eulerian kernels for a problem exhibiting
several shear bands. The Krongauz–Belytschko method with stress point integration is used. A J2 flow
constitutive model is applied and the stress rate relation is

srJ ¼ CelsJ : De ¼ Cels J : ðD
Dp Þ; ð76Þ

where s is the Kirchhoff stress computed by s ¼ F  P. The yield condition is

f ðs; qÞ ¼ r


rY ð

Þ ¼ 0;

of 3 dev of d
¼ s ¼
rY ð

Þ ¼
H ð

Þ;
os 2
r o

d

where rY ð

Þ is the yield stress in uniaxial tension,
is the effective plastic strain, r

is the von Mises effective
3 dev dev 1=2 dev
stress ½2 s : s  , s is the deviatoric stress and H ð

Þ is the plastic modulus. For strain softening, H
is negative.
The steel ring is loaded along the inner ring. A pressure of 1.0e10 N/m2 is first applied and is then
decreased exponentially in time. The material constants are q ¼ 7800 kg/m3 , E ¼ 2:0e11 N/m2 , H ¼
E=10,
rY ð0Þ ¼ 1:0e10 N/m2 . One thousand and fifty particles and 900 stress points are used.
Fig. 17 shows the distribution of the effective strain in the deformed steel ring. The shear bands occur
because of the unstable material behavior. According to [26], the field of shear bands around a circular hole
loaded by a uniform pressure is known since at every point of the field the principal stress coincides with the
radial and peripheral directions. The shear bands will be logarithmic spirals emerging from the inner
surface at an angle  p4.
Regardless of whether we use a Lagrangian or an Eulerian kernel, the shear bands have the above
mentioned morphology. When a shear band has grown through the ring, the ring breaks into fragments
[22,23,35].
The shear bands usually are initiated where stress or strain concentrations occur due to geometric or
material inhomogeneities. In a numerical simulation with uniformly distributed stresses, a slight imperfection

Fig. 17. Deformed steel ring by stress points (t ¼ 0:5 ms): (a) Lagrangian kernal and (b) Eulerian kernal.
1060 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

from either the material or the numerics may trigger the localization. So shear bands will occur at different
positions in different patterns for different models. With an Eulerian kernel, the positions of the shear bands
differ from those in Lagrangian kernel simulation but it may simply be a reflection of these effects. The re-
sponse observed for the Eulerian kernel is quite acceptable. This is not always the case, as shown in the next
example.

9.6. Spurious material instability

This example shows the onset of a spurious instability for an Eulerian kernel in a 2D problem. Again, the
KBM with stress point integration is applied with Eulerian and Lagrangian kernels. The material model is
given in Eq. (18) and the related text. This material was studied in Xiao et al. [47] where it was shown
analytically that stress point integration with Lagrangian kernels can reproduce the material instability but
Eulerian kernels distort it severely.
A pressure of 6.2e7 N/m2 is applied for 0.1 ms on the inner surface of the rubber ring. The discrete model
with stress points consists of 540 particles and 480 stress points. Fig. 18 shows the deformed rubber ring at
the same time for the Lagrangian kernel and an Eulerian kernel computations. Fig. 18(b) exhibits a
clustering of particles in the Eulerian kernel which is typical of the tensile instability. Fig. 19 shows the
distribution of rhh in the deformed configuration. As can be seen, the stress is concentrated at discrete
locations around the circumference of the ring for the Eulerian kernel but it is still rotationally symmetric
for the Lagrangian kernel.

9.7. Fluid flow in a funnel

To illustrate a problem where the Eulerian kernel is more suitable than the Lagrangian kernel consider a
fluid flow in a funnel. The KBM with stress point integration is used. The material behavior of the fluid is
described via an equation of state:

Fig. 18. Deformed rubber ring by stress points (at t ¼ 0:12 ms): (a) Lagrangian kernel and (b) Eulerian kernel.

Fig. 19. rhh distribution at t ¼ 0:12 ms: (a) Lagrangian kernel and (b) Eulerian kernel.
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1061

Fig. 20. Fluid flow in a funnel using an Eulerian kernel, (a) at the beginning of the computation, (b) at the end of the computation;
the model sizes are rescaled to fit.

 c 
.
p¼B
1 : ð77Þ
.0
The initial density .0 for water is 0.001 g/mm3 , the exponent c ¼ 7 and B ¼ 0:4704 MPa. Fig. 20 shows the
flow out of the funnel for an Eulerian kernel. With a Lagrangian kernel the simulation becomes unstable.
The velocities explode suddenly, so that the flow cannot even be drawn after the instability.

10. Conclusions

A meshfree particle method based on Krongrauz and Belytschko [27] has been developed and tested. In
this method, the trial functions are corrected derivatives of the kernel functions and the test functions are
Shepard functions. Quadrature of the Galerkin weak form was performed over the particles and a set of
slave particles, often called stress points. It was shown that this method is substantially more robust for
solid mechanics problems than SPH and that it is able to deal with unstable phenomena such as shear
bands and fracture. The rate of convergence of the approximation is close to two in the displacement norm;
two is the optimal rate for an approximation with linear completeness.
We have also reviewed the original symmetrized SPH method, the Randles–Libersky normalization, and
MLS-based methods such as EFG and compared the results of EFG–KBM with these other methods. In
making these comparisons, we have considered both problems that have closed-form solutions, some where
only experimental results are available, and some where only qualitative evaluations can be made.
A noteworthy result, which was also reported in Belytschko et al. [2] for a nonsymmetrized version of
SPH, is that for a beam problem, symmetrized SPH solutions do not converge when the domain of
influence (smoothing length) is decreased at the same rate as the particle spacing. The method is convergent
when the domain of influence is kept constant in size. However, constant domains of influence are not
suitable for practical computations since the sparsity of the discrete equations is severely diminished.
We also showed that EFG–KBM is capable of treating natural boundary conditions, such as traction
free boundaries without any explicit enforcement of these conditions. This was demonstrated in a problem
of plane wave reflection; Randles and Libersky [41] have argued that a similar correction that they called
a normalization requires explicit treatment of the traction boundary conditions.
It is interesting to note that the best results were obtained for the EFG method with cell integration.
However, this method is significantly more expensive than SPH or stress point methods, and our results
1062 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063

indicate that its accuracy per computation time is inferior to EFG–KBM. The SPH and Randles–Libersky
methods are more dispersive than the other methods, which results in a higher error in the L2 velocity norm,
but the peak as well as the wave velocity are matched quite well.
This new particle method, EFG–KBM, was applied to failure problems involving concrete and metals and
its performance was compared to some of the other particle methods reviewed here, such as SPH. Concerning
the kernel, it can be said that the Lagrangian kernel seems to be more effective for solids than Eulerian kernels.
For a metal plate under compressive loading with viscoplastic material behavior, it was shown that the
Lagrangian kernel yielded stable solutions in the tensile regime while for the Eulerian kernel large oscillation
could be observed. Similarly, for a hyperelastic ring, the Eulerian kernel exhibited a distinct instability. These
instabilities confirm the analysis by Xiao et al. [47] that showed that Eulerian kernels distort the domain (in
stress space) of material stability. However, several cases were reported, such as the formation of shear bands
in the metallic ring, where Eulerian kernels performed quite well even in tension.
The influence of the Lagrangian and Eulerian kernel with respect to the computational effort cannot be
answered definitively. The advantage of the Lagrangian kernel is that a time consuming neighbor search and
the calculation of the shape functions needs to be done only once at the beginning of the calculation.
However, whereas Eulerian kernels naturally handle contact, Lagrangian kernels require contact algorithms.
However, the more pertinent issue is the relative capability of Eulerian and Lagrangian kernels to handle large
deformations. Lagrangian kernels cannot handle very large deformations, such as found in fluids, at least not
without some kind of reinitialization of the domains of influence. However, for the simulations of the failure
of solids, Lagrangian kernels appear to be more robust and they are easily able to handle the magnitude of
deformations encountered in this class of problems. Therefore, in fluid structure interaction problems, it
is probably preferable to model the solid with Lagrangian kernels and the fluid with Eulerian kernels.

Acknowledgements

The support of Office of the Naval Research and the Army Research Office is gratefully acknowledged.

References

[1] S. Beissel, T. Belytschko, Nodal integration on the element-free Galerkin method, Comput. Methods Appl. Mech. Engrg. 139
(1996) 49–74.
[2] T. Belytschko, Y. Krongauz, J. Dolbow, C. Gerlach, On the completeness of meshfree particle methods, Int. J. Numer. Methods
Engrg. 43 (1998) 785–819.
[3] T. Belytschko, M. Fleming, D. Organ, Y. Krongauz, W.K. Liu, Smoothing and accelerated computations in the element free
Galerkin method, J. Comput. Appl. Math. 74 (1996) 111–126.
[4] T. Belytschko, Y. Krongauz, D. Organ, P. Krysl, Meshless methods: An overview and recent developments, Comput. Methods
Appl. Mech. Engrg. 139 (1996) S.3–S.47.
[5] T. Belytschko, Y.Y. Lu, L. Gu, Element-free Galerkin methods, Int. J. Numer. Methods Engrg. 37 (1994) 229–256.
[6] T. Belytschko, Y. Guo, W.K. Liu, S.P. Xiao, A unified stability analysis of meshless particle methods, Int. J. Numer. Methods
Engrg. 48 (2000) 1359–1400.
[7] T. Belytschko, S.P. Xiao, Stability analysis of particle methods with corrected derivatives, Comput. Math. Appl. 43 (2000)
329–350.
[8] T. Belytschko, Crack propagation by element free Galerkin methods, Engrg. Fract. Mech. 51/2 (1995) 295–315.
[9] T. Belytschko, Y.Y. Lu, Element-free Galerkin methods for static and dynamic fracture, Int. J. Solids Strucut. 32 (1995) 2547–
2570.
[10] T. Belytschko, W.K. Liu, B. Moran, Nonlinear Finite Elements for Continua and Structures, John Wiley and Sons, Ltd,
New York, 2000.
[11] T. Belytschko, H.Y. Chiang, E. Plaskacz, High resolution two-dimensional shear band computations imperfections and mesh
dependency, Comput. Methods Appl. Mech. Engrg. 119 (1994) 1–15.
 T. Rabczuk et al. / Comput. Methods Appl. Mech. Engrg. 193 (2004) 1035–1063 1063

[12] T. Black, T. Belytschko, Convergence of corrected derivative methods for second-order linear partial differential equations, Int. J.
Numer. Methods Engrg. 44 (2) (1999) 177–203.
[13] J. Bonet, T.-S.L. Lok, Variational and momentum preservation aspects of Smooth Particle Hydrodynamics formulations,
Comput. Methods Appl. Mech. Engrg. 180 (1999) 97–155.
[14] J.S. Chen, C.T. Wu, S. Yoon, Y. You, A stabilized conforming nodal integration for Galerkin meshfree-methods, Int. J. Numer.
Methods Engrg. 50 (2001) 435–466.
[15] G.A. Dilts, Moving least squares particle hydrodynamics II: Conservation and boundaries, Int. J. Numer. Methods Engrg. (2000)
1503–1524.
[16] G.A. Dilts, Moving-least-squares-particle hydrodynamics I: Consistency and stability, Int. J. Numer. Methods Engrg. 44 (1999)
S.1115–S.1155.
[17] G.A. Dilts, Some recent developments for moving-least-squares particle methods, First M.I.T. Conference on Computational
Fluid and Soil Mechanics, June 12–14, 2001, Preprint, Massachusetts Institute of Technology, Cambridge, MA 02139, USA.
[18] C.T. Dyka, R.P. Ingel, An approach for tensile instability in smoothed particle hydrodynamics, Comput. Struct. 57 (1995) 573–580.
[19] C.T. Dyka, R.P. Ingel, Addressing Tension Instability in SPH Method, Naval Research Laboratory, Washington, 1994.
[20] R.A. Gingold, J.J. Monaghan, Smoothed particle hydrodynamics: theory and applications to non-spherical stars, Mon. Not.
R. Astr. Soc. 181 (1977) 375–389.
[21] K. Girkmann, Flaechentragwerke, 6, Springer Verlag, Wiley, New York, 1974.
[22] D.E. Grady, D.A. Benson, Fragmentation of metal rings by electromagnetic loading, Exp. Mech. 12 (187) (1983) 393–400.
[23] S. Hao, W.K. Liu, C.T. Chang, Computer implementation of damage models by finite element and meshfree methods, Comput.
Methods Appl. Mech. Engrg. 187 (2000) 401–440.
[24] G.R. Johnson, S.R. Beissel, Normalized smoothing functions for SPH impact computations, Int. J. Numer. Methods Engrg. 39
(1996) S.2725–S.2741.
[25] L.M. Kachanov, Time of the rupture process under creep conditions, Izv. Akad. Nauk SSR Otd. Tech. 8 (1958) 26–31.
[26] L.M. Kachanov, Foundations of the Theory of Plasticity, North Holland, London, 1971.
[27] Y. Krongauz, T. Belytschko, Consistent pseudo derivatives in meshless methods, Comput. Methods Appl. Mech. Engrg. 146
(1997) 371–386.
[28] L.D. Libersky, A.G. Petschek, Smooth particle hydrodynamics with strength of materials, Advances in the Free Lagrange
Method, Lecture Notes in Physics, 395, 1990.
[29] W.K. Liu, S. Jun, J. Adee, T. Belytschko, Reproducing kernel particle method for structural dynamics, Int. J. Numer. Methods
Engrg. 38 (1995) 1665–1679.
[30] W.K. Liu, S. Jun, Y.F. Zhang, Reproducing kernel particle methods, Int. J. Numer. Methods Fluid 20 (1995) 1081–1106.
[31] Y.Y. Lu, T. Belytschko, L. Gu, A new implementation of the element free Galerkin method, Comput. Methods Appl. Mech.
Engrg. 113 (1994) 397–414.
[32] L.B. Lucy, A numerical approach to the testing of fission hypothesis, Astronom. J. 82 (1977) 1013–1024.
[33] J.J. Monaghan, Kernel estimates as a basis for general particle methods in hydrodynamics, J. Comput. Phys. 46 (1982) S.429–
S.453.
[34] J.J. Monoghan, An introduction to sph, Comput. Phys., Commun. 48 (1988) 89–96.
[35] N.F. Mott, Fragmentation of shell cases, Proc. Roy. Soc. London 300 (1947) 300–308.
[36] B. Nayroles, G. Touzot, P. Villon, Generalizing the finite element method: diffuse approximation and diffuse elements, Comput.
Mech. 10 (1992) 307–318.
[37] A. Needleman, Dynamic shear band development in plane strain, J. Appl. Mech. 56/1 (March) (1989).
[38] P.E. Petersson, Crack growth development of fracture zone in plain concrete and similar materials, Report No. TVBM-1006,
Division of Building Materials, 1981, University of Lund, Sweden 1981.
[39] T. Rabczuk, Numerische Untersuchungen zum Fragmentierungsverhalten von Beton mit Hilfe der SPH-Methode, Dissertation,
Inst. Massivbau, Universitaet Karlsruhe, 2002.
[40] T. Rabczuk, J. Eibl, Simulation of high velocity concrete fragmentation using SPH/MLSPH, Int. J. Numer. Methods Engrg. 56
(2003) 1421–1444.
[41] P.W. Randles, L.D. Libersky, Recent improvements in SPH modeling of hypervelocity impact, Int. J. Impact Engrg. 20 (1997)
525–532.
[42] P.W. Randles, L.D. Libersky, Normalized SPH with stress points, Int. J. Numer. Methods Engrg. 48 (2000) 1445–1462.
[43] Schmidt-Hurtienne: Ein dreiaxiales Schaedigungsmodell zur Beschreibung des Dehnrateneffektes von Beton, Dissertation, Institut
fuer Massivbau und Baustofftechnologie, Universitaet Karlsruhe, 2000.
[44] J.C. Simo, T.J.R. Hughes, On the variational foundations of assumed strain methods, J. Appl. Mech. 53 (1986) 1685–1695.
[45] J.W. Swegle, S.W. Attaway, M.W. Heinstein, F.J. Mello, Hicks D.L., An Analysis of Smoothed Particle Hydrodynamics, Sandia
Report SAND93-2513,1994,SNL,Albuquerque, NM 87185.
[46] J.W. Swegle, D.A. Hicks, Smooth particle hydrodynamics stability analysis, J. Comput. Phys. 116 (1995) 123–134.
[47] S.P. Xiao, T. Belytschko, Material stability analysis of particle methods, Adv. Comput. Math., accepted for publication.

View publication stats