2nd Reading

August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

International Journal of Computational Methods

Vol. 13, No. 5 (2016) 1630001 (42 pages)

c World Scientific Publishing Company
DOI: 10.1142/S0219876216300014

An Overview on Meshfree Methods: For

Computational Solid Mechanics

G. R. Liu
Department of Aerospace Engineering and Engineering Mechanics
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

University of Cincinnati, Cincinnati, OH 45221-0070, USA

Consultant, College of Mathematics
Taiyuan University of Technology, P. R. China
Int. J. Comput. Methods Downloaded from [Link]

liugr@[Link]

Received 2 June 2016

Accepted 26 June 2016
Published 8 August 2016

This review paper presents a methodological study on possible and existing meshfree
methods for solving the partial differential equations (PDEs) governing solid mechanics
problems, based mainly on the research work in the past two decades at the authors
group. We start with a discussion on the general steps in a meshfree method based on
nodes, with the displacements as the primary variables. We then examine the major tech-
niques used in each of these steps: (1) techniques for displacement function approxima-
tions using nodes, (2) approximation of the gradient of the displacements or strains based
on nodes and a background T-cells that can be automatically generated and refined,
and (3) formulation techniques for producing algebraic equations. The function approx-
imation techniques include node-based interpolation methods, cell-based interpolation
methods, function smoothing techniques, and moving least squares approximation tech-
niques. The gradient approximation includes direct differentiation, gradient smoothing,
and special strain construction. Formulation techniques include strong-form, weakform,
local weakform, weak-strong-form, and weakened weakform (W2). In theory, a mesh-
free method can be developed using a combination of function approximation, gradient
approximation, and formulation techniques, which can lead to matrix of a large number
of possible methods. This review attempts to provide an overall methodological review,
rather than a usual review of comparing different methods. We hope to show readers the
differences between the forests, and just between the trees.

Keywords: Meshfree method; FEM; S-FEM; PIM; S-PIM; numerical method; interpola-
tion techniques; PDE; stability; convergence; efficiency.

1. Introduction
Problems of solid mechanics are typically governed by a set of simultaneous
second-order partial differential equations (PDEs), with a proper set of displace-
ment (Dirichlet, essential) and force (Neumann, natural) boundary conditions.

1630001-1
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

Throughout this work we assume that the original mechanics problem is well-posed,
meaning that it has physically one and only one solution. In the design of a struc-
ture made of solids for engineering applications, it is essential to solve the PDE
for solutions in displacements and then in strains and stresses, so as to ensure the
functionality and safety of the structural systems. Because of the complexity of the
PDE, it is generally not possible to solve it in analytic means, especially for higher
(2D and 3D) dimensional problems with complicated geometry. One naturally uses
the so-called numerical methods that seek for an approximated solution, and typi-
cally by means of discretization using nodes, elements, or cells or grids. This paper
discusses meshfree methods in a loose term that allows the use of nodes and may
also a background mesh or cells to help in the numerical treatments, but the mesh
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

or cells must be generatable automatically, as defined in Liu [1]. In actual fact,

with a set of nodes are available, one can always generate a set of T-mesh or T-
Int. J. Comput. Methods Downloaded from [Link]

cells (Triangular cells for 2D and Tetrahedral cells for 3D) for complex geometries,
using existing algorithms such as the well-known Delaunay and advancing front
algorithms. Therefore, in terms of practical applications, as long as a mesh can be
generated automatically without human interventions, there is no reason not to
make use of it. What is of practical importance is that the methods should be not
sensitive to cell/mesh distortion, because the automatically generated mesh may
not be always of good quality. Based on this argument, a more practical definition
of meshfree method should be a method that does not rely on a mesh that cannot
be generated automatically and the mesh is used in free and not a confined manner.
There are quite a large number of numerical methods developed in the past
based on element mesh [Hughes (1987); Belytschko (2000); Liu and Quek (2013)],
nodes or particles [5] and nodes with a background mesh or cells [Liu and Gu
(2005); Liu (2002)]. This paper aims to provide an overall methodological review on
a large number of groups of meshfree methods that have been studied intensively in
the past, especially those studied by the author’s groups in Singapore, China, and
the United States. We shall finally provide a matrix of possible groups meshfree
methods, including many of the major existing methods and possible promising
future methods. A discussion on the features and the suitability of different group
methods to different types of problems of are also provided. Our discussion will
be conducted in reference to the standard finite element method (FEM), which is
currently the dominating and widely used method in the field and well-known to
our readers.
In the development of any meshfree method, we need essentially the following
five major steps:

(1) Domain discretization using nodes, in which process we may also create a cor-
responding set of cells known as background mesh in an automated manner;
(2) Displacement function approximation based on nodes;
(3) Displacement function-gradient approximation based on nodes, and may be also
on cells;

1630001-2
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

(4) Establish a set of algebraic system of equations that is a “good” representation

of the original PDEs, so that the solution is stable and converges to the exact
solution of the PDEs, when the nodes are refined;
(5) Solve the set of algebraic system of equations for solutions in first displacements
and then strains and stresses.

This paper assumes that the discretization is already done using a proper pre-
processor and a set of nodes are available with at least a set of T-cells or T-mesh.
We assume also that once the set of algebraic system equations are obtained, it can
be routinely solved using the standard algebraic equation solvers widely available.
The development of efficient such solver is another important area of research for
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

large scale problems, but it is out of the scope of this paper. Therefore, this paper
focuses our discussions on the key numerical techniques needed in Steps 2, 3 and 4:
Int. J. Comput. Methods Downloaded from [Link]

(1) Techniques for displacement function approximation (or shape function con-
structions) based on nodes;
(2) Techniques for approximating the gradient (or derivatives or strains) of the
displacement function;
(3) Formulation techniques aiming to convert the PDEs to a set of algebraic system
of equations.

The function approximation techniques include node-based interpolation

methods, cell- or element-based interpolation methods, smoothing (integral rep-
resentation in general) techniques, and moving least squares (MLS) approximation
techniques. The gradient approximation includes direct differentiation, gradient
smoothing, and special strain construction techniques. Formulation techniques
include strong-form, weakform, local weakform, weak-strong-form, and weakened
weakform (W2). Therefore, in theory, a numerical method can be developed using
a combination of function and derivative approximation techniques and formulation
techniques, which can lead to at least 7×3×7 = 147 possible numerical methods, as
will be shown in detail in Table 3. A general discussion of these different numerical
methods is provided, and an overview of the major existing numerical methods on
their features and the applicability for different types of problems is conducted. We
briefly mention also some of the possible promising future methods to be developed
in the future.

2. Setting of the Problem

Consider a problem of solid mechanics defined in domain Ω bounded by Γ = Γu ∪Γt .
The standard governing equations are those given in the form of PDEs, with a set
of boundary conditions:

LTd σ + b = 0 Equilibrium equation in problem domain Ω, (1)

1630001-3
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

u = uΓ Displacements given on the Dirichlet boundary Γu , (2)

LTn σ = tΓ Traction given on the Neumann boundary Γt , (3)
where
b is the given body force vector containing force components,
tΓ is the prescribed traction on the natural (stress) boundaries,
uΓ denotes the prescribed displacement on the essential (displacement)
boundaries,
Ld is the matrix of differential operators defined by:
 
∂/∂x1 0 0
 
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

   0 ∂/∂x2 0 
∂/∂x1 0  
 0 ∂/∂x3 
   0 
Ld =  0 ∂/∂x2  for 2D, Ld =   for 3D.
Int. J. Comput. Methods Downloaded from [Link]

 0 ∂/∂x ∂/∂x 
 3 2 
∂/∂x2 ∂/∂x1 3×2  
∂/∂x3 0 ∂/∂x1 
∂/∂x2 ∂/∂x1 0 6×3
(4)
Ln is the matrix of the components of the unit outwards normal on the natural
boundary:
 
n1 0 0
  0 n 0
 2 
n1 0  
  0 0 n 
Ln =  0 n2  for 2D, Ln =   3  for 3D. (5)

 0 n3 n2 
n2 n1 3×2  
n3 0 n1 
n2 n1 0 6×3

σ is the vector of the stress components that relates to the strain vector ε through
a matrix of the material constants c that is symmetric positive definite (SPD) in
the form of
σ = cε. (6)
The strain vector ε relates to the displacement vector u via the compatibility
equation:
ε = Ld u. (7)
The strains obtained using the forgoing equation are called compatible strains (with
the given displacement function).

3. Techniques for Function Approximation

To solve the problem defined in Sec. 2 using any discretized numerical method,
the problem domain is discretized with a set of Nn nodes located in the interior

1630001-4
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

of the domain Ω and on the boundary Γ. As set of Nc T-cells can also be created
automatically with the given set of nodes, the unknown function u of any of the
displacement components is often approximated using the nodal displacements dn
and the so-called shape functions φn defined over the problem domain:

Nn
u= φn dn = Nd. (8)
n=1

Because the shape function is often assigned to the nodes, it is thus also called nodal
shape function. Once a set of linearly independent shape functions is found, the
actual unknowns are only the nodal displacements. Therefore, the shape functions
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

can serve as (nodal) bases for the function approximation or interpolation to evaluate
the displacement function at any desired location within the problem domain Ω.
Int. J. Comput. Methods Downloaded from [Link]

With Eq. (8), a distributed function is now approximated using the discretized Nn
nodal displacements, and thus the dimension of the displacement vector d is Nn . For
d-dimensional solids, the total dimension of the space will be d × Nn , because each
node carries d degrees of freedoms (DOFs). Therefore, the approximation of the
displacement function becomes the creation of the shape functions, and it is one of
the most important tasks in developing a meshfree method. Many techniques have
been developed for this purpose, which will be discussed below in detail.

3.1. Requirements on shape functions

The challenge is now how to efficiently create shape functions using the given nodes
in the problem domain. In such a creation process, the following requirements [1]
may be considered:

(1) Arbitrary nodal distribution: The nodal distribution can be arbitrary within
reason, and at least is more flexible than that used in the standard FEM.
(2) Stability: The algorithm must be stable with respect to irregularity (within rea-
son) of the node distribution. This allows the method insensitive to background
cell distortion.
(3) Consistency: The shape functions should possess a certain order of consis-
tency — the capability to locally reproduce exactly the polynomials of that
order. This is essential for the method producing solutions converging to the
exact one when nodal spacing is reduced.
(4) Delta function property: Ideally, the shape function should possess the Kro-
necker delta function property. This is important for easy imposition of essential
(displacement) boundary conditions.
(5) Compact support: The support (or influence or smoothing) domain for field
variable approximation/interpolation should be small to include only a small
number of nodes. This can lead to a sparse set of algebraic system questions
that can be solved very efficiently using the standard routines of equation
solvers.

1630001-5
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

(6) Efficiency: The algorithm should be computationally efficient. The compu-

tational complexity should be of the same order of complexity as that of
FEM. This is an ultimate measure on any method for practical engineering
applications.
(7) Compatibility: Preferably, the nodal shape functions should be continuous
throughout the problem domain. This is an essential requirement when a weak-
form is used to produce the algebraic system equations.
(8) Linearly independence: The set of the nodal shape functions of all the nodes in
the problem domain should be, in general, linearly independent and hence forms
a basis for a unique displacement interpolation. This is essential for producing
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

a stable and unique solution.

It is not mandatory for a method to satisfy all the above-mentioned requirements.
Int. J. Comput. Methods Downloaded from [Link]

The essential requirements should correspond to the choice of the gradient approx-
imation techniques and the formulation techniques to convert the PDEs to a set of
algebraic system equations. More discussions will be given in Sec. 7.

3.2. Types of methods for creating shape functions

A number of ways to construct shape functions have been proposed, which may be
grouped into three major categories:

3.2.1. Series representation methods

(a) Point interpolation methods (PIM):
(i) Polynomial PIM (or PIM or PPIM) that uses local surrounding nodes [1].
It uses a very small number of nodes in the locally supported domain of
a point of interest to perform interpolation. The approximation function
passes through the nodal value, and the polynomial bases can be repro-
duced, as proven in [1]. Because of this, it uses the least number of nodes
and hence it is most efficient. However, the moment matrix could be sin-
gular, the construction process may fail, and special techniques are needed
to overcome the singularity issue [46]. Also the approximation function
may not be continuous over the global problem domain. Thus, it works
best with weakened weak (W2) formulations [Liu and Zhang (2013); Liu
(2009, 2010)] that based on the G-space theory accommodating discontin-
uous functions. In addition, in the W2 formulation, no domain integration
is needed, and it works well for T-cells [98].
(ii) Radial PIM (or RPIM) that uses local surrounding nodes [Wang and
Liu (2000, 2002a, 2002b); Liu and Gu (2001c); Wang et al. (2002,
2005); Liu et al. (2003b, 2004b, 2005a, 2005b, 2006a, 2006b, 2006c,
2006d, 2007a; 2008a; 2011a, 2013); Dai et al. (2003, 2004); Wu and Liu
(2003); Hardy (1990); Kansa (1990); Golberg et al. (1999); Powell (1992);

1630001-6
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

Coleman (1996); Sharan et al. (1997); Fasshauer (1997); Wendland (1998);

Zhou and Liu (2006); Kee et al. (2007, 2008a, 2008b); Zhang et al. (2007);
Zhao et al. (2008, 2009a, 2009b); Cui et al. (2010a, 2010b)]. RPIM is similar
to PIM, but it uses the radial basis functions as basis for the interpolation,
which effectively resolved the singular moment matrix problem [1]. RBF
was used earlier by many [Hardy (1990); Kansa (1990); Golberg et al.
(1999); Powell (1992); Coleman (1996); Sharan et al. (1997); Fasshauer
(1997); Wendland (1998)] for solving PDEs via collocation, but they use all
the scattered nodes in the entire problem domain, resulting in full stiffness
matrices that are often bad-conditioned. RPIM developed by Liu’s group
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

uses only local nodes in various types of formulations: strong-form with

regularization techniques for stabilization [Liu et al. (2005b; 2006c); Zhou
and Liu (2006); Kee et al. (2007, 2008a, 2008b)]; weakform to improve the
Int. J. Comput. Methods Downloaded from [Link]

stability [Wang and Liu (2000, 2002); Liu and Gu (2001c); Wang et al.
(2002, 2005); Liu et al. (2003b, 2004b, 2005, 2006, 2007a); Dai et al. (2003,
2004); Wu and Liu (2003)], and in a weakened weakform to ensure the
stability [Liu et al. (2006d, 2008a, 2011a, 2013); Zhang et al. (2007); Zhao
et al. (2008, 2009a, 2009b); Cui et al. (2010a, 2010b)]. Because only a small
number of nodes are used, the moment matrix is well-conditioned and the
moment matrix is always invertible. Therefore, RPIM works practically for
randomly distributed nodes. Pure RPIM is not able to produce the pure
linear polynomial. Therefore, in practice, one often uses also lower order
polynomial bases in RPIM. Because the approximation function may not
be continuous over the global problem domain. Thus, it works best with
W2 [Liu and Zhang (2013); Liu (2009, 2010)] using T-cells [98].
Note that the radial basis functions are essentially a distance func-
tion that contains shape parameters. The performance of the RPIM is
affected by the choice of these parameters [90]. Intensive studies at GR
Liu’s group has led the following general understandings [1]: (1) strong-
form methods are in general very sensitive to these parameters, and some
kind of regularization techniques are needed to obtain stabilized solutions
[Liu et al. (2005b; 2006c); Zhou and Liu (2006); Kee et al. (2007, 2008a,
2008b)]; (2) weakform methods are less sensitive, and much better man-
ageable [90]; and the W2 methods is very insensitive to these parameters
[Liu et al. (2006d, 2008a, 2011a, 2013); Zhang et al. (2007); Zhao et al.
(2008, 2009a, 2009b); Cui et al. (2010a, 2010b)]. Because of this, RPIM
work best with W2 formulations.
(iii) Radial polynomial PIM (or RP-PIM) [Wang and Liu (2000, 2002a, 2002b);
Liu and Gu (2001c); Wang et al. (2002, 2005); Liu et al. (2003b, 2004b,
2005a, 2005b, 2006a, 2006b, 2006c, 2006d, 2007a; 2008a; 2011a, 2013);
Dai et al. (2003, 2004); Wu and Liu (2003); Hardy (1990); Kansa (1990);
Golberg et al. (1999); Powell (1992); Coleman (1996); Sharan et al. (1997);

1630001-7
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

Fasshauer (1997); Wendland (1998); Zhou and Liu (2006); Kee et al. (2007,
2008a, 2008b); Zhang et al. (2007); Zhao et al. (2008, 2009a, 2009b); Cui
et al. (2010a, 2010b)]. This is an effective combination of PIM and RPIM.
It is often simply termed as RPIM. It is recommended to use only the
linear polynomial basis, so that linear function can be reproduced, and at
the same time the robustness against irregular node distribution can be
maintained.
(b) MLS methods:
(i) MLS approximation that uses local nodes [Lancaster and Salkauskas (1981);
Nayroles et al. (1992); Belytschko et al. (1994, 1996); Krongauz and
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Belytschko (1996); Fleming et al. (1997)]. The MLS shape function does
not passes through the nodal value, but the polynomial bases can be repro-
Int. J. Comput. Methods Downloaded from [Link]

duced. A proof on this can be found in [1]. The singular moment matrix
problem is resolved by using excessive nodes compared with the number of
basis terms. Because bell-shaped weight functions are used, the MLS shape
functions can be made continuous, and hence it works best with weakform
formulations, which provides the stability. In addition, due to the use of
weakform, the natural boundary condition is “naturally” imposed in the
same weakform equation. However, proper means are required to impose
the essential boundary conditions, because the approximation function does
not pass through the nodal values. In addition, because of the weak formu-
lation, domain integration is required and thus a background mesh or cells
are needed. Special treatments are also needed for imposing the essential
boundary conditions. The efficiency is affected by the use of excessive nodes
leading to larger bandwidth.
(ii) Modified MLS approximation (Chapter 3 of Liu [1]). Similar to the MLS,
but enforcements are introduced to ensure the approximation function pass-
ing through the nodal value only for the nodes on the essential boundary.
Through such a modification to the shape functions, the imposition of the
essential boundary conditions can be done in the same way as in the FEM.
(c) FEMs: a standard method in the FEM community
(i) Element-based interpolation with mapping [Hughes (1987); Belytschko
(2000); Liu and Quek (2013)] is widely used in the standard FEM for-
mulation. Shape functions are created based on regular (square, cube, etc.)
elements in the natural coordinate system assigned to the element of regu-
lar shape. A mapping procedure links the regular element to the physical
irregular element via a Jacobian matrix. Such a mapping is a must to ensure
that (1) the shape functions can be created without dealing with the sin-
gular moment matrix, and (2) the approximation function is continuous
along the interfaces of the elements. Too heavy distortion of the elements
is prohibited, to avoid the ill-condition of the Jacobian matrix. In addition,
domain integration for each of the elements is required.

1630001-8
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

(ii) Lower order element-based interpolation without mapping [Liu and Quek
(2013); Liu et al. (2007b); Dai et al. (2007); Liu and Nguyen (2010); Liu
and Zhang (2013)]. This is widely used in the so-called smoothed finite
element method (S-FEM) [Liu et al. (2007b); Dai et al. (2007); Liu and
Nguyen (2010)], where no mapping and no domain integration is needed.
It is the simplest special case of the so-called smoothed point interpolation
methods (S-PIMs), when lower polynomial basis are used [98]. Because of
the smoothing operations, the S-FEM works well with T-cells, and is found
insensitive to mesh distortion. In addition, the elements can be of arbitrary
polygonal shape, and the solution can have the properties of typical W2
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

models, such as lower boundary properties, volumetric-free, ultra-accuracy,

etc. [97]. Note also that when the S-PIM is used, the interpolation is per-
formed for each of the Gauss points on the segments of the smoothing
Int. J. Comput. Methods Downloaded from [Link]

domain. Such an interpolation is entirely independent of elements, but only

the surrounding nodes: a typical meshfree approach [98].

3.2.2. Integral representation methods

(i) Smoothed particle hydrodynamics (SPH) method that is particle based [Liu
and Liu (2003); Liu and Gu (2005); Liu (2002); Lucy (1977); Gingold and
Monaghan (1977); Monaghan (1982); Liu et al. (2002, 2008b); Zhou et al.
(2007)].
The function value at a particle becomes a weighted-average of those at
the surrounding particles, using a weight function that is often of bell-shaped
[13]. Function derivative value at a particle becomes the sum of the function
values at the surrounding particles weighted by the derivatives of the weight
function. It works best with strong-form formulation (the simplest collocation
approach), which is very straightforward. The SPH method is often found quite
stable, despite the use of simple collocation approach. This is because of the
use of smoothing operations that “weakened” the function (and its derivatives)
approximations. This type of method was regarded as weakform-like [1]. Due to
the collocation approach, it is however difficult to impose boundary conditions,
especially the natural boundary conditions.
(ii) Gradient Smoothing Method (GSM) that is cell-based [Liu et al. (2008c); Liu
and Xu (2008); Xu et al. (2010, 2012); Li et al. (2007, 2012a, 2012b); Yao et al.
(2013, 2015); Yao and Liu (2014, 2025); Wang et al. (2013, 2014); Zhang et al.
(2008); Xu et al. (2010)].
The GSM is a systematic technique to approximate various orders of the
gradient of a function that is approximated using nodal function values, using
smoothing operations, which is in an integral form (rather than the usual
differential forms). It uses various (node-based, edge-based, face-based, and
cell-based) smoothing domains in carefully designed nested manners, so that
different orders of the gradient can be properly approximated. It works for

1630001-9
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

all types of background cells and insensitive to cell-distortions, including the

T-cells that can be automatically generated. It suits well to the simple strong-
form formulation such as collocation. The GSM is found very stable, despite the
use of simple collocation. This is because of the use of smoothing operations
for the function derivatives approximations. Due to the use of background
cells, it is easy to impose boundary conditions including the natural boundary
conditions. It uses the simple node-passing PIM shape function for function
approximation, can reproduce the linear function exactly, and is a conservative
formulation. It was initially developed for CFD problems, and was extended for
solid mechanics problems [27] and fluid–structural interaction (FSI) problems
[Wang et al. (2013, 2014)]. The GSM is a significant improvement from the
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

SPH approximations.
(iii) Reproducing kernel particle method (RKPM) [Liu et al. (1993, 1995, 1997);
Int. J. Comput. Methods Downloaded from [Link]

Chen et al. (1996)]

Similar to SPH, but the reproducibility of the basis function is enforced
by effectively introducing a correction function. The RKPM shape function
has similar behavior as the MLS, and an analysis of their relationship can be
found in Liu [1]. The stability of this approximation process is achieved by using
excessive nodes compared with the number of basis terms. It works best with
weakform formulations, which provides the stability. In addition, due to the
use of weakform, the natural boundary condition is naturally imposed. Special
techniques are needed to impose the essential boundary conditions, because the
approximation function does not pass through the nodal values. The efficiency
is affected by the use of excessive nodes leading to larger bandwidth. Because of
the weak formulation, domain integration is required and hence a background
mesh/cells are need to carry out the integration numerically.

3.2.3. Differential representation methods

(i) Finite difference method (FDM) using regular and irregular grids [Liszka and
Orkisz (1980); Jensen (1980); Nayroles et al. (1992); Liu et al. (2006a)].
This technique is based on the Tayler series expansion, which works well for
regular grids, as in the standard FDMs. When the grids become irregular, the
moment matrix can be singular, and the approximation process can fail. One
needs resorting to special algorithms to select nodes to avoid such a singularity
issue [101]. When a node selection scheme is used, one need also making sure
that the shape function constructed will be linearly independent. However,
the examination of the linear independence of the approximation is in general
difficult.
(ii) The regular-grid-based Taylor series expansion with the RPIM [64]. It uses the
Taylor series expansion based on regular-grids, which takes care of the linear
independence. It uses RPIM for interpolation, which takes care of the node
singularity issue. It is thus an excellent combination in this regard.

1630001-10
 Int. J. Comput. Methods Downloaded from [Link]
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Table 1. Properties of existing shape functions.

Shape Arbitrary nodal/ Delta Compact Global Linearly Preferred to pair

August 4, 2016 14:24 WSPC/0219-8762

functions particle distribution Stability Consistency function support Efficiency compatibility independency with formulation
√ √ √√√
PIM Yes Yes Yes Yes/No Yes W2 form
√√√ √√√ √√
RPIM (Kriging)a Yes Yes Yes No Yes W2 form
√√ √√ √
MLS Yes No Yes Yes Yes Weak form
√√√ √√√ √√√
SPH No No Yes NA Yes Strong-form
√√ √√ √
196-IJCM

RKPM Yes No Yes Yes Yes Weak form

1630001-11
√√ √√ √
LS (LS) Yes No Yes No Yes W2 form
√√ √√ √
MLM Yes No Yes No Yes W2 form
Irregular- FDM
√ √ √ √ √ √√√
or Hermite-type NA Yes/No Strong-form
1630001

Note: a Kriging approximation is essentially the same as the RPIM [57].

2nd Reading

An Overview on Meshfree Methods

 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

3.2.4. Other techniques

There are a number of other useful methods for constructing shape functions:

(i) NURBS used in the isogeometric analysis, as well documented in [104].

(ii) Finite point method (FPM, irregular grids) [105].
(iii) Kringing interpolation that is essentially the same as the RPIM [Gu (2003);
Dai et al. (2003)].
(iv) Minimum length method (MLM) that is a nodal value passing interpolation
technique [95].
(v) hp-clouds method and the partition of unity useful in creating the generalized
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

FEM or GFEM [Duarte and Oden (1996); Melenk and Babuska (1996)].

Table 1 lists some of the shape function techniques together with the properties.
Int. J. Comput. Methods Downloaded from [Link]

4. Techniques for Function Derivatives Approximation

4.1. Procedure for derivatives approximation
The traditional procedure of the displacement methods (with the FEM being the
most popular one) has essentially very simple two basic steps: (1) displacement field
construction (via shape function creation), and (2) weak formulation to derive a dis-
cretized set of algebraic equations that can be solved using standard routines. The
most popular and standard FEM follows this process [Hughes (1987); Belytschko
(2000); Liu and Quek (2013)]. To further advance, for example in the meshfree
methods [1], we may need to introduce one more additional step: strain field con-
struction or approximation of (displacement) function derivatives. This additional
step creates a room to significantly improve the solution accuracy, and to solve the
compatibility issues in the standard FEM so that incompatible methods can be
formulated based on the W2 formulation [1].
Although in the so-called hybrid or mixed FEM, one may also assume the strain
filed using three or two field variational principles [Pian and Wu (2006); Wu (1982)],
but the operations are confined within the elements. Based on the meshfree con-
cepts, the construction of strain field is much more flexible and the operations are
often beyond the elements. Such operations are based on the G-space theory and W2
formulations, and there is no need to introduce any of the derivatives as additional
DOFs [Liu et al. (2007b); Liu (2010)].
Once the displacement function is assumed, one can obtain its derivatives natu-
rally by simple differentiation. Because of this natural practice, other techniques for
derivative construction seem to be unnecessary, and hence are relatively less devel-
oped. The basic requirements for novel techniques for function derivative approxi-
mation are as follows.

(1) The procedure should at least not affect negatively the stability and convergence
of the numerical model;

1630001-12
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

(2) The procedure should be straightforward without too many additional numer-
ical operations;
(3) It should be cost-effective without introducing additional DOFs to the equations
system;
(4) It should have sufficient improvement with respect to the standard FEM, with-
out increasing the DOFs, or offer some attractive additional properties;
(5) It should work well with T-cells that can be automatically generated.
If the techniques for function derivative approximation do not require the
existence of the compatible strain (or derivatives of the displacement functions),
it is regarded as a weakened weak (W2) formulation. The existing W2 for-
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

mulations include the generalized smoothed Galerkin (GS-Galerkin) and the

so-called strain-constructed Galerkin (SC-Galerkin) formulations (see Chap-
ter 5 in [1]). We now discuss some of the techniques for function-derivative (or
Int. J. Comput. Methods Downloaded from [Link]

function-gradient or strain field) approximation.

4.2. The need for function derivative approximation

Once the displacement is properly assumed, the derivatives are already available
by simple standard differentiation. The strains can be obtained by the strain–
displacement relation given in Eq. (7), which is known as the compatible strains,
as in the standard FEM. Because the assumed displacement field is compatible
over the entire problem domain and the strain is obtained exactly by differentiation
(no approximation), the standard FEM model is said fully compatible. There are,
however, three major issues associated with this type of fully compatible formu-
lation: (1) The so-called “locking” behavior for many problems due to the well-
known “overly stiff” phenomenon; (2) less accuracy in the stress solutions due to
the strain jumping across the element interfaces; and (3) poor solution using trian-
gular/tetrahedral elements; (4) significant accuracy loss when the element mesh is
heavily distorted, and its demands for good quality mesh of quadrilateral elements,
due to the bad-conditioning problems of the Jacobian matrix. On the other hand,
the T-cells are often preferred as they can be much easily created automatically for
complicated geometries.
Efforts have been made in resolving the overly stiff phenomenon, such as the
hybrid or mixed FEM using three or two field variational principles [Pian and Wu
(2006); Wu (1982)] within the framework of elements. In meshfree settings, the com-
patible strain field may not be accessible on the discontinuous points/lines/surfaces
(even in the so-called distributive sense). One has to construct a new strain field
or approximate the derivatives of the assumed displacement functions. Even if the
compatible strain is accessible, such a strain can be erroneous because the differen-
tiation is known as a “hashing” operator [109]. We thus often want to construct a
better strain field for more accurate solutions with desired properties. The simplest
way is to smooth out the errors in strains induced by the approximated displace-
ments, this leads to the S-PIM [98] and S-FEM models [97]. A meshfree model that

1630001-13
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

uses constructed strains (instead of the compatible strains) is generally called a

strain-constructed model or SC-model [98].

4.3. Admissible conditions for constructed strain fields

The minimum condition for a constructed strain field over the problem domain Ω
is that the constructed strain ε̂(v) using the assumed displacements v is squarely
integrable over the entire problem domain: ε̂ ∈ L2 (Ω). This is to ensure that the cor-
responding strain potential energy in the entire solid is bounded for all the possibly
constructed strain fields [1]. The following three additional admissible conditions
are needed for establishing stable and convergent SC-models.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

4.3.1. Condition 1: Orthogonal condition

Int. J. Comput. Methods Downloaded from [Link]

For any assumed displacement function v ∈ H1h,0 (Ω), the constructed strain field
ε̂(v) and the compatible strain field ε̃(v) = Ld v should satisfy the following orthog-
onal condition
 
ε̂ (v)cε̂(v)dΩ =
T
ε̂T (v)c(Ld v)dΩ. (9)
Ω Ω

In this case, a Galerkin model using the strain-constructed strain filed can be
derived directly from the single-field Hellinger–Reisnner’s principle [110] and hence
is said variationally consistent. Not here that we need the assumed displacement
functions in an H1 space (continuous in the entire problem domain), so that the
compatible strain ε̃(v) is available over the entire problem domain.

4.3.2. Condition 2a: Norm equivalence condition

A proof of the stability of a W2 model is given in [1]. In such a proof, a known
reference strain field of a compatible FEM model is used, which leads to the so-called
norm equivalence condition that is briefed below.
For any displacement field v ∈ G1h,0 (Ω) for a model with a total of Nn nodes, a
vector U that collects the nodal values of the displacement function can always be
arranged in the form of
 T

 

U = u1(1) u2(1) u3(1) u1(2) u2(2) u3(2) · · · u1(Nn ) u2(Nn ) u3(Nn ) .

| {z } | {z } | {z }

node(1) node(2) node(Nn )

(10)
We now introduce a base (discrete and compatible) constant strain field ε̃b (U):
ε̃b (U) = Ld Φe U, (11)
  
Bb

1630001-14
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

where Bb is the global strain matrix has entries assembled from element strain
matrices of a standard FEM model using linear triangular elements:

[Bb ]6×12 = [Ld ]6×3 [Φe ]3×12 (12)

in which the matrix of FE shape functions has the form of

 
φb φb2e φb3e φb4e 0 
 1e 0 0 0 0 0 0 0 
 
Φe = 
 0 φb1e 0 0 φb2e 0 0 φb3e 0 0 φb4e 0 .
 b 
 0 0 φb1e 0 0 φb2e 0 0 φb3e 0 0 φ4e 
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

           
node 1 element e node 2 element e node 3 element e node 4 element e
(13)
Int. J. Comput. Methods Downloaded from [Link]

The global strain matrix Bb is usually a very sparse matrix with a dimension of
6 × 3Nn for 3D models. Because all the nodal shape functions are linearly indepen-
dent, Bb has 3Nn linear independent columns for 3D models. The (global) strain
vector ε̃b (U) can always be written in the form of
 T
ε̃T
b(1) ε̃T
b(2) ··· ε̃T
b(Nc )
ε̃b (U) =      , (14)
element 1 element 2 element Nc

where Nc is the total number of the triangular elements for the base model, and

ε̃b(i) = {ε̃11 ε̃22 ε̃33 2ε̃23 2ε̃13 2ε̃12 }T

(i) , (i = 1, 2, . . . , Nc ). (15)

When linear triangular elements are used, any strain components are constant
in an element.
The constructed strain field ε̂(U) can also be written in the similar form of
 T

 

 
ε̂(U) = ε̂T ε̂T ··· ε̂T , (16)


(1)

(2)
    
(Nq )

 
cell 1 cell 2 cell Nq

where Nq is the total number of the quadrature or integration cells for the SC-model,
and

ε̂(i) = {ε̂11 ε̂22 ε̂33 2ε̂23 2ε̂13 2ε̂12 }T

(i) , (i = 1, 2, . . . , Nq ). (17)

These constructed strain components ε̂ij are, in general, not constant in an

integration cell. The constructed strain field ε̂(U) needs to be equivalent in a norm
to the base constant strain field ε̃b (U), meaning that there exists a nonzero positive

1630001-15
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

constants cac and cca such that

ε̂(U)L2 ≤ cac ε̃b (U)L2 and
(18)
ε̃b (U)L2 ≤ cca ε̂(U)L2 , ∀U ∈ RdN
0
n
,

where RdN 0
n
is a subspace of RdNn where the displacements at the nodes on the
essential boundary are imposed. Constants cac and cca should be independent of
RdN
0
n
. Since our strain field is defined in a vector form, measuring in L2 norm is
handy, and workable for both ε̃b (U) and ε̂(U) (although they may be different in
length). Clearly, to satisfy conditions (18), the minimum requirement should be
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Nq ≥ Nc (19)

which means that we should “sample” more locations for the (non-negative) strain
Int. J. Comput. Methods Downloaded from [Link]

energy in the construction of an SC-model, as practiced in the S-FEM [Liu et al.

(2007b); Dai et al. (2007); Liu and Nguyen (2010)], and proved for general settings
based on the argument of ensuring positivity [100].
The norm equivalence ensures only the stability and hence the convergence, but
we do not know where it converges to. To ensure the solution converges to the exact
solution, we need a strain convergence condition.

4.3.3. Condition 2b: Strain convergence condition

The convergence condition for the constructed strain is defined as [136]:

lim ε̂(x, U) → ε̃b (x, U), ∀x ∈ Ω, ∀U ∈ RdN

0
n
. (20)
h→0
Nn →∞

When both conditions (18) and (20) are satisfied, the SC-Galerkin model is sta-
ble and converges to the exact solution of the original PDEs, and it is a typical W2
formulation. In this case, it may not be variationally consistent in the conventional
sense, because v is not, in general, in an H 1 space.

4.3.4. Condition 3: Zero-sum condition

For any v ∈ H1h,0 , we may also construct the strain in each of the quadrature
domains in the following form

ε̂(v) = ε̃b (v) + αε̂m (v), (21)

where ε̃b is the (compatible) strain field of a base model, α ∈ R is a finite adjustable
parameter for regularizing the gradient of the modification, and ε̂m is the modified
portion of the strain field. The zero-sum condition is defined as

ε̂m (v)dΩ = 0, ∀v ∈ H1h,0 . (22)
Ω

1630001-16
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

When both Eqs. (21) and (22) can be found, the Galerkin-like weakform is then
applicable (see, Chapter 5 of [1]), and the formulation is variationally consistent.
The zero-sum condition is similar to orthogonal condition that used in FEM for
the stabilization formulation for quadrilateral element using reduced integration
[Belytschko and Bachrach (1986); Belytschko and Bindeman (1993)].

Table 2. Some schemes for construction and modification of the strain fields.

No Schemes Brief description Examples

1. Strain gradient Using a factor to scale down or up the αFEM [127] VCαFEM [138]
scaling gradient of the compatible strain
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

field based on elements

2. Strain The compatible strain field based on S-FEM [113] NS-FEM [117]
smoothing elements is smoothed over ES-FEM [118] GSM [14]
Int. J. Comput. Methods Downloaded from [Link]

(orthogonal smoothing domains that can be SCNI [134]

projection) element-based, cell-based,
node-based, edge-based, and
face-based. The constructed strain
is assumed to be constant in the
smoothing domains.
3. Generalized The strain field is constructed by the S-PIM [98] S-RPIM [Liu
smoothing generalized smoothing technique et al. (2006d, 2008a);
over the smoothing domains that Zhang et al. (2007); Zhao
can be element-based, cell-based, et al. (2008, 2009a,
node-based, edge-based, and 2009b)] NS-PIM [Liu
face-based. The displacements et al. (2005c); Liu and
maybe discontinuous. The Zhang (2007); Liu (2008)]
constructed strain is assumed to be NS-RPIM [132] ES-PIM
constant in the smoothing domains. [123]; Liu (2008)] CS-PIM
[Liu et al. (2011a, 2013);
Cui et al. (2010a, 2010b)]
4. PIM At points in the problem domain, the SC-PIM [Liu et al. (2009f);
strains are assigned with smoothed Liu (2008)] PIM-CS [Liu
strains over smoothing domains. et al. (2009a, 2009b)]
The constructed strain field is then
constructed via a PIM using these
strains at these points.
5. Least square At points in the problem domain, the LS-PIM [51]
approxima- strains are assigned with smoothed
tion strains over smoothing domains.
(orthogonal The constructed strain field is then
projection) constructed via least squares
approximation using these strains
at these locations.
6. Finite difference Tailor series is used to approximate RPIM-FD [64]
(Tailor Series the derivatives at regular
Expansion) background grids, utilizing the
values of the primary variable at
the grid points, which values are in
turn approximated using RPIM (or
PIM) locally.

Note: This table may not be exclusive.

1630001-17
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

4.4. Strain construction techniques: A summary

Table 2 lists some of the possible and practical techniques for constructing strain
fields, with a brief description and sources for each one of them.

4.4.1. Scaling the strain gradient

This technique is to change the gradient of the compatible strain field obtained using
the assumed displacement field that is in an H1 space. This was first used in the so-
called αFEM presented in [127], based on a FEM setting of quadrilateral elements.
It was discovered that the compatible strain field in a quadrilateral elements can be
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

divided into a constant portion and a portion varying linearly with the coordinates.
The gradient of the varying portion can be modified by scaling without affecting
the ability to pass the standard patch tests [138]. Therefore, the convergence of the
Int. J. Comput. Methods Downloaded from [Link]

model can be ensured, as long as the stability is kept. A quadrilateral element was
then formulated to obtain “exact” or best possible solution for a given problem by
scaling the gradient of strains in the natural coordinates and Jacobian matrices with
a scaling factor α ∈ (−∞, +∞). It is shown that as long as α is not zero, the stability
of the model can be preserved, and will always converge to the exact solution of the
original strong-form PDEs. The method is not necessarily variationally consistent
but stable and convergent. The αFEM can produce “nearly exact” solutions in the
strain energy for all overestimation problems, and the “best” possible solution to
underestimation problems. The general procedure of strain gradient scaling works
is as follows.
For any assumed uh ∈ H1h,0 , the compatible strain field within an element can
be obtained as
ε̃(x) = Ld uh (x). (23)
The compatible strain field is then split into two portions:
ε̃(x) = ε̃c (xc ) + ε̃v (x), (24)
  
ε̃c

where ε̃c = ε̃(xc ) is the constant portion evaluated at the origin of the natural
coordinate xc defined for the element, and ε̃v (x) varies with the coordinate and
should satisfy the zero-sum condition equation (22). The strain field can then be
constructed in the following form:
ε̂(x) = ε̃c + αε̃v (x), (25)
where α ∈ (−∞, +∞) is an adjustable parameter. The stability is ensured by using
a small α = 0. Note ε̃c can be in general different form ε̃b in Eq. (21). It is only the
constant portion of the compatible strain field.
The strain gradient scaling technique described above has been tested for quadri-
lateral elements where the compatible strain field is linear (in the natural coordi-
nate) [135]. Extension of these techniques to other types of elements is of course

1630001-18
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

possible, but should be performed with caution, as the proof for general settings for
this technique has not yet been performed.

4.4.2. Strain construction by smoothing

(1) Strain field construction by strain smoothing for uh ∈ H1h
A strain field can be constructed by smoothing the compatible strain field
obtained from assumed displacement field that is in an H 1 space. Smoothing oper-
ation is an often used integral technique to smooth a function, and has been used in
the nonlocal continuum mechanics [143] to obtain a smoother stress field. The SPH
[Lucy (1977); Gingold and Monaghan (1977); Monaghan (1982); Liu et al. (2002,
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

2008b); Zhou et al. (2007)] uses it for the approximations of functions and their
derivatives. The strain smoothing was used as a means of resolving the material
Int. J. Comput. Methods Downloaded from [Link]

instabilities and in the special instability in the nodal integrated meshfree methods
[134]. It is used later in finite element settings: such as the node-based smoothed
FEM or NS-FEM [117] and the edge-based smoothed FEM or ES-FEM [118] with
a number of attractive properties. It was also used for the derivative approximation
to establish the GSM models of strong-formulation for solids [141], compressible
fluids [Xu et al. (2010); Liu and Xu (2008)], and incompressible fluids [Xu et al.
(2012); Yao and Liu (2014)].
Since this construction is by modifying the strain field, it can only be done when
the strain field is available, meaning that the assumed displacement there must be
at least differentiable in the local smoothing domain Ωsx associated with the point
of interest x.
(2) Strain field construction by generalized smoothing for uh ∈ G1h
When the assumed displacement function is discontinuous, we need to construct
a strain field through alternative means. Based on the existing works on strain
smoothing operations, Liu [133] has proposed a generalized smoothing technique for
functions that are discontinuous, leading to the G space theory. A strain field can
then be constructed using the assumed displacement field in a G space. A generalized
smoothed Galerkin (GS-Galerkin) weakform has then been established for models
based on elements [133] and PIM models [123, 133]. The strain field constructed
and used in the GS-Galerkin is piecewise constant, and the formulation is a typical
case of weakened weak W2 formulation and the stability and convergence has been
proven [1]. The GS-Galerkin offers a unified theoretical foundation for this class of
compatible and incompatible methods. The strain modification is a special case of
strain construction by generalized smoothing. Since a function in a H1h space is also
in the corresponding G1h space, the GS-Galerkin formulation works also for models
with assumed displacement functions in a H1h space. It is variational consistent, if
the solution is sought from an H1h space, and is stable and convergent if the solution
is sought from a G1h space.
The numerical methods developed based on the GS-Galerkin form can possess
three major important properties: (1) very accurate solutions can be obtained using

1630001-19
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

T-cells; (2) the stiffness of the discretized model can be reduced compared to com-
patible FEM model, and the exact model, which allows us to obtain upper bound
solutions with respect to both the FEM solution and the exact solution; (3) the
solution of a numerical method developed using the smoothed bilinear form are
less sensitive to the quality of the mesh, and T-cells can be well used for problems
of complicated domains; (4) the continuity of the trial and test functions can be
reduced as long as they are in a G space, which allows us to use much more types
of methods to create shape functions.

4.4.3. Strain construction by point interpolation

by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

In this case, the strain field ε̂ is constructed using the PIM interpolation by the
following steps.
Int. J. Comput. Methods Downloaded from [Link]

(1) Select a set of points for strain interpolation based on the element mesh (in
FEM settings) or the background cells (in meshfree settings). The density of the
points should usually be higher than the nodes to satisfy the norm equivalence
condition. We may have some of these points coincide with the nodes so that
the strain convergence condition equation (20) can be easier to satisfy. For
example, when a triangular mesh is used, we first choose all the field nodes
as strain interpolation points, and then add in the middle points of edges and
the centroids of the triangular cells/elements as the interpolation points. Such
selection of strain interpolation points are used in PIM-CS [136], SαFEM [139],
and SC-PIM [140].
(2) For any assumed admissible displacement uh , at a strain interpolation point, the
strain is the assigned as either the compatible strain obtained using Eq. (7) (if
uh ∈ H1h ), or a smoothed strain of the compatible strain field over a proper local
domain. For example, when a triangular mesh is used, we choose node-based
smoothing domains for points coinciding with nodes, and edge-based smoothing
domains for points at the mid edge points. For points at the centroids, we
may simply use the compatible strains, as in PIM-CS [136], SC-PIM [140], and
SαFEM [139].
(3) A new set of (preferably triangular) cells for the set of strain interpolation
points is then created for strain interpolation. This set of cells is used also for
integration of the weakform, and it is called quadrature or integration cells. For
example, when triangular cells are used, the new triangular quadrature cells are
hosted by the original triangles, and typically a triangular cell is divided into
1, 3, 4, or 6 integration cells, as in the SC-PIM [140].
(4) The strain field in each of the quadrature cells is then constructed using point
interpolation based on these strains at the set of points using shape functions
created using the strain interpolation points and any interpolation technique
presented in Sec. 3.2.1. For example, when a triangular mesh is used, we simply
use linear shape functions created for each of the integration cells. In such

1630001-20
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

a case, the integration of the weakform can be performed analytically using

the area coordinates that is the same as the linear shape functions [112] and
no numerical integration is needed. In this case, the strain field within each
integration cell will be obtained using linear interpolation:

3
ε̂(x) = Φi (x)ε̆i , (26)
i=1

where ε̂ is the vector of constructed strains over the triangular integration cell, and
ε̆ is the vector of strains at each vertex which are either the compatible strains ε̃


or the smoothed strains ε obtained using a local smoothing domain. In practice,
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

the linear interpolation is often used for easy integration. In this case, the strain
energy for a triangular integration cell can be obtained exactly so that no numerical
integration is needed [98].
Int. J. Comput. Methods Downloaded from [Link]

4.4.4. Strain construction by least square approximation

Least square approximation is a kind of orthogonal projection, found useful in
strain field construction, and was implemented in the least square point interpola-
tion methods (LS-PIM) [51]. The constructed strain field can easily satisfy the norm
equivalence condition, and as long as the strain convergence condition equation (20)
can be satisfied, we have a convergent SC-model. The procedure for constructing
strain field using least square approximation is quite similar to the method of inter-
polation. The four steps procedure given in Sec. 4.4.3 stand, except to replace the
interpolation by a least square approximation. When least square approximation
is used, we can have large integration cells, use more strain points for approxima-
tion, and have wider choice of polynomial basis terms for creating models of desired
properties. More details are given in [51].

5. Formulation Techniques
5.1. Strong-form (asymmetric)
In strong-formulations, we discretize the governing PDE (1) directly with the bound-
ary conditions. We first obtain the shape functions using local nodes in a small set
Sn , and the displacements can be approximated using (for 2D problems as example),
 h   
u  φH I 0 uI 
h
u = = H
= ΦH
I uI , (27)
v I∈Sn
0 φI vI I∈Sn
     
ΦH uI
I

where ΦH H
I is the matrix of shape functions φI , and uI is the vector of the (unknown)
nodal displacements. Using Eq. (7), the compatible strain becomes
  
ε = Ld uh = Ld ΦHI uI = Ld ΦH
I uI = BI uI . (28)
  
I∈Sn I∈Sn I∈Sn
BI

1630001-21
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

Substitute now Eq. (28) into Eq. (6) and then into (1) for each of the nodes in
the interior of Ω. Substitute next Eq. (6) into (3) for all the nodes on the Neumann
boundary, and finally impose Eq. (2) at all the nodes on Dirichlet boundary. This
would give us a total of 2Nn equations, each for a node, for a total of 2Nn unknown
nodal displacements (for 2D problem as an example), which can be written in a set
of linear algebraic system equations:
KU = F. (29)
Because all the equations (equilibrium and boundary conditions) are all enforced
at the nodes, it is called collocation method. This process is very straightforward,
and can be easily coded. However, it can be seen that there is no means in the
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

process to control the stability and the convergence of the solution. In fact, the
collocation methods are often found not stable and the solutions can vary a lot
Int. J. Comput. Methods Downloaded from [Link]

when the locations of the nodes change.

Note also that the K matrix in Eq. (29) is not symmetric in general, because
it is a special case of the Petrov–Galerkin formulation (with the Delta function as
the test function) [1]. Moreover, the positivity of K matrix (hence the stability of
the discretized equations) is also questionable, due to the use of the Delta func-
tion. In addition, the satisfaction of the governing equilibrium equation and those
of the boundary conditions are entirely separated at individual nodes. Therefore,
there could be conflicts for nodes near the boundaries: a node is enforced by the
equilibrium equations, but a nearby node is enforced by a boundary condition. This
totally disconnected situation could be at least one of the most critical causes of the
instability issues with the colocation methods. If the stability is not ensured, discus-
sion about the convergence is meaningless. Therefore, special (and often expensive)
stabilization techniques must be used to at least ensure the stability, as per the
intensive studies conducted by GR Liu’s group [Liu et al. (2005b; 2006c); Zhou and
Liu (2006); Kee et al. (2007, 2008a, 2008b)].

5.2. Weakform (symmetric)

In a weak formulation, we solve PDE (1) indirectly, by converting it to a weakform,
or simply use a proper energy principle. Detailed discussions on various weakforms
and energy principles can be found in [1]. Typical weakforms can be derived from
the general weighted residuals approach, which is a general mathematical approach.
It is done by pre-multiply weight (test) functions to the strong-form of the governing
(equilibrium) equations, in which the unknown functions replaced by trial functions
that are approximated using the shape functions [1]. The most widely used weakform
for solid mechanics problems is the Galerkin weakform, in which the test and trial
functions are from the same space. The standard Galerkin weakform used in the
FEM and some of the meshfree methods has the following form
  
δ(Ld u)T (cLd u)dΩ − δuT bdΩ − δuT tΓ dΓ = 0. (30)
Ω Ω Γt

1630001-22
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

Substitute now Eq. (27) into Eq. (30), execute the variation, and then impose
Eq. (2) at all the nodes on the Dirichlet boundary. A set of 2Nn equations will
be obtained for a total of 2Nn unknown nodal displacements (for 2D problem),
which can also be written in the form of Eq. (29). Note that in this formulation, the
Neumann boundary equation (3) is naturally satisfied in an integrated manner in
Eq. (30) (via the third term). For this reason, the Neumann boundary equation (3)
is also called naturally boundary condition. Note also that if the displacement func-
tions are in a proper H1 space, the stability and the convergence of the solution
can be theoretically proven [2], meaning that the K matrix will be positive definite.
Because the Galerkin form is symmetric (the transport of the first term in Eq. (30)
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

is itself), the K matrix obtained will be symmetric. Therefore, we shall have SPD
K matrix. With such a K matrix, Eq. (29) has one and only one solution. When
the nodes are refined the solution shall approach to the exact solution to the prob-
Int. J. Comput. Methods Downloaded from [Link]

lem defined by Eqs. (1)–(3), because the finite element space approaches to the H1
space that hosts the exact solution.
Note that if the approximation defined in Eq. (27) does not pass through the
nodal displacements, such as the MLS or RKPM approximations, proper constraints
need to be imposed. In this case, the Galerkin weakform needs to be modified to:
  
T
δ(Ld u) (cLd u)dΩ − δu bdΩ −
T
δuT tΓ dΓ
Ω Ω Γt
 
− δλT (u − uΓ )dΓ − δuT λdΓ = 0. (31)
Γu Γu

where the last two terms in Eq. (31) are produced by the method of Lagrange
multipliers for handling essential boundary conditions for cases when u − uΓ = 0.
The Lagrange multipliers λ here can be viewed physically as smart forces that force
u − uΓ = 0. If the trial function u can be chosen so that u − uΓ = 0, the last two
terms will vanish. In this case the discretized system equation has the following
extended form:
    
K G U F
= . (32)
GT
0 λ q
The matrices in Eq. (32) can be much larger than the stiffness matrix obtained
using Eq. (30), because of the presence of matrix G needed to enforce the essential
boundary conditions. Depending on the number of nodes on the essential boundaries
in relation to the total number of nodes in the problem domain, solution efficiency
can be reduced significantly. From Eq. (30), it can also be clearly seen that the
matrix of the enlarged system is still symmetric but no longer positive definite.
This further reduces computational efficiency in solving the system equations.

5.3. Local weak form (asymmetric)

Local weakforms can also be derived using the general form of the weighted resid-
uals formulation, by using the so-called Petrov–Galerkin approach where the test

1630001-23
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

functions from a different space as that of the trial functions. A typical local weak-
form has the following matrix form [1]:
 T   





VI (cLd u)dΩ − WI bdΩ − WI tdΓ − WI tΓ dΓ = 0, (33)
ΩQ ΩQ ΓQu ΓQt

where WI is a weight function defined for each nodes in the problem domain, and


VI is a matrix that collects the derivatives of the weight functions:


W I,x 0
 



VI =  0 W . (34)
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

 I,y 



W I,y W I,x
Int. J. Comput. Methods Downloaded from [Link]

It is in fact the “strain” matrix caused by the weight (test) functions.

Note that in the MLPG [Atluri and Zhu (1998); Atluri et al. (1999)] approach,
the integration domain in Eq. (33) is different from that in Eq. (30). It is over a
small local integration ΩQ corresponding to a node rather than the global problem
domain Ω. In addition, the third term in Eq. (33) is performed only on the Dirichlet
boundary and the fourth term is only for the Neumann boundary. Note that when
ΩQ is entire within the Ω, the third and fourth terms vanishes. The approximation
of the trail function can still be performed using Eq. (27), but a bell-shaped locally
supported test function will be used for each of the nodes. Therefore, the imple-
mentation of Eq. (33) is performed for each of the nodes. This would give us a total
of 2Nn equations, each for a node, for a total of 2Nn unknown nodal displacements
(for 2D problem), which can also be written in the form of Eq. (29).
Because the test and trail functions are now from different spaces, the stiffness
matrix K obtained will be no longer symmetric. In this regard, it is quite similar as in
the strong-formulation: equations are created node by node. However, because of the
use of the (local) integral form, the imposition of the (domain) equilibrium equation
and the boundary conditions are somewhat connected in the same equations, which
provides a way to resolve the conflicts, and hence some insurances of the stability
to the discretized system equation. However, the solution properties are not very
well understood, due to the use of Petrov–Galerkin approach.
Note that if the approximation defined in Eq. (27) does not pass through the
nodal displacements, such as the MLS or RKPM approximations, proper constraints
need to be imposed. The treatments are the same as those for the weak formulations.
Note also that when local weakforms are used, the trail functions do not have to be
in a H1 (Ω) space. If the derivatives of the smoothing function is continuous (which is
usually true), we require trail function to be only locally continuous (in distributive
sense) over ΩQ of any node. Therefore, there is no reason to create MLS shape
functions, and all the PIM and RPIM shape functions created locally for ΩQ of a
node shall work well. Intensive studies in this direction can be found in Gu and Liu
[2001], Liu and Gu [2001b, 2002], [51] and Liu et al. [2003b, 2005a].

1630001-24
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

It can be clearly observed from Eq. (27) that when the Delta function is used
as the test function for each node, the local weakform becomes the strong-form
discussed in Sec. 5.1.

5.4. Weak-strong-form (asymmetric)

Based on the arguments given in Secs. 5.1 and 5.3, a meshfree weak-strong (MWS)
is proposed [Gu and Liu (2005); Liu and Gu (2003a, 2003b); Liu et al. (2004)]. In
the MWS method, one uses strong-form for all the interior nodes with ΩQ that
is also within domain Ω, and local weakform for nodes with ΩQ that intersects
domain boundary Γ. The invention of the MWS method is originally based on the
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

realization that (1) the treatments to these near boundary nodes have conflicts in
satisfying the governing equations and boundary conditions; (2) when a strong-
Int. J. Comput. Methods Downloaded from [Link]

form collocation method with a set of linearly independent nodal shape functions
is applied to problem with only Dirichlet boundary conditions, the solutions are
often much more stable; (3) weak formulations can naturally handle the Neumann
boundary conditions; and (4) the formulation in both strong-form and local weak-
form are all based on individual nodes, and hence can be combined together without
difficulty. Therefore, a simple combination of both shall make a good sense. The
MWS has been applied to solve static [145] and time-dependent problems [144] and
incompressible flow problems [147].

5.5. Weakened weak (W2) form (symmetric)

In the effort of establishing a more general theoretical framework for meshfree meth-
ods, W2 forms have been proposed [Liu and Zhang (2013); Liu (2009, 2010)]. The
first W2 form was the generalized smoothed Galerkin or GS-Galerkin weakform
built using the gradient smoothing technique [134, 113] and the generalized gradi-
ent smoothing technique [133]. The foundation is on the normed G space theory [99].
The second W2 form was the strain constructed Galerkin or SC-Galerkin weakform
that is a more general W2 form using functions from basically un-normed G spaces.
The weakened-weak formulations are relatively new but have been used for a num-
ber of meshfree methods for important properties including upper bound property,
ultra-accuracy, and super-convergence [140].
The requirement on the assumed functions for field variables is further reduced
upon the already reduced weak formulations. For the solid mechanics problem
defined in Sec. 2 the assumed displacements can be in a G1h space and hence may
be discontinuous. When the solution is sought from a H1h space (a subspace of a G1h
space), the W2 model becomes softer and the accuracy and the convergence rate of
the solution is often found much higher than the weak formulation. It offers a sys-
tematical and efficient way for both compatible and incompatible meshfree methods.
Similar to the weak formulation, the discrete system equations of a W2 formu-
lation are symmetric for irregularly distributed nodes for problems of symmetric
operators. Therefore, one should always use a W2 formulation for such problems

1630001-25
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

to preserve the symmetry property and hence excellent stability, accuracy, and effi-
ciency. The typical weakened weakform can be written as
Ns  
T
Asi (δε̂i ) cε̂i − δuT bdΩ − δuT tΓ dΓ = 0, (35)
i=1 Ω Γt

where Ns is the number of the smoothing domains or the sub-domains over which
the strains are constructed. When the smoothed strains are used, a sufficient number
of smoothing domains are required [98], and the displacement functions can be in a
proper G1 space. In general, the constructed strains should satisfy these admissible
conditions given in Sec. 4.3 to ensure the stability and convergence of W2 model.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

A typical W2 model is the S-PIM, where the strains are computed by looping over
each of the smoothing domains.
Int. J. Comput. Methods Downloaded from [Link]

(a) (b)

(c) (d)

Fig. 1. A typical process to obtain smoothed strains in S-PIM [1, 98]. The problem domain is
represented by a set of arbitrarily distributed nodes (four of which are shown as filled circle).
The domain is divided into a set (eight in this case) of smoothing domains. For each of the
smoothing domain (Ωs2 for this case), numerical line-integration along the four segments of the
smoothing domain is performed to calculate the smoothed strain that may be used in Eq. (35).
The Gauss points are the stars, at which the displacements need to be approximated (using for
example PIM or RPIM and locally selected nodes). (a) First Gauss point on the left segment; (b)
2nd Gauss point on the left segment; (c) first Gauss point on the bottom segment; (d) 2nd Gauss
point on the bottom segment. The similar is done for the right and top segments of the smoothing
domain. The local node selection may be assisted by a background triangular mesh that can be
automatically generated.

1630001-26
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

6. Possible Meshfree Methods

Combining the techniques for function approximation given in Sec. 3, gradient
approximation techniques given in Sec. 4, and formulation techniques given in Sec. 5,
we can now in theory formulate a large number of numerical methods. Table 3 lists
(7 × 3 × 7 =) 147 possible meshfree methods. The existing methods are filled into
this table with references. Mark “NA” indicates the method is either not applicable
or it does not made a good sense to do so. The single question “?” marks indicate
possible methods that yet to be developed. The double question marks “??” indicate
possible methods that may potentially have good performance. Interested readers
are encouraged to investigate further.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

7. Applications: A Number of Issues

Int. J. Comput. Methods Downloaded from [Link]

We now discuss some key issues related to the application of a number of methods
to practical engineering problems. To facilitate the discussion, we categorize the
application problems now into two major groups:

• Group 1: Problems that are essentially governed by linear or nonlinear elliptic

PDEs (in terms of spatial derivatives), such as the solid mechanics problems that
may be linear or nonlinear;
• Group 2: Problems that are essentially governed by nonlinear hyperbolic PDEs
(in terms of spatial derivatives), such as the computational fluid dynamics (CFD)
problems.

7.1. Pre-processing
7.1.1. Pre-processing in FEM
Pre-processing is usually needed for solving a practical engineering problem of
large scale. In using the standard FEM, we have now quite a large number of
r ANSYS
,
choices of commercially available software packages, such as ABACUS
, r
r r
NASTRAN
, DYNA
, etc. A brief discussion on this can be found in Liu and
Quek [2013]. These FEM packages usually have their own preprocessors, but one
r etc.
can also use dedicated software package for preprocessing, such HyperMesh
,
Generally speaking, the time for analyst to create a FEM model can be a domi-
nate cost factor to carry out a complex FEM project. This is mainly because the
standard FEM demands a quality element mesh to ensure a quality solution. This
is the bottleneck to the ultimate goal of automation in computation for practical
engineering problems.

7.1.2. Automatic mesh generator

The ultimate solution to the expensive manpower cost in creating a numerical model
is, in the opinion of the author, the automation in mesh generation. Ideally, an

1630001-27
 Int. J. Comput. Methods Downloaded from [Link]
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

G. R. Liu
Table 3. A matrix of possible and the existing numerical methods.

Formulation techniques⇒
Function Strain Galerkin Petrov–Galerkin Weak-strong Weakened weakform Least Boundary integral
approximation approximation Strong-form weakform weakform -form (W2) Galerkin) squares equations

PIM Compatible ? PIMf LPIMm ? NA ? BPIMx

Smoothing GSMa NA ? ?? S-PIMq ? ?
Construction NA NA ? ? SC-PIMr ? ?
RPIM Compatible RPICMb RPIMg LRPIMn MWSp NA LS-RPIM v ??
Smoothing ?? NA ? ?? S-RPIMs ? ?
August 4, 2016 14:24 WSPC/0219-8762

Construction ? NA ? ? ?? ? ?
MLS Compatible ? EFGh MLPGo ?? NA ? BNMy
Smoothing ?? SCNIi ? ? NA ? ?
Construction ? NA ? ? NA ? ?
RKPM Compatible ? RKPMj ? ? NA ? ?

1630001-28
i
196-IJCM

Smoothing ?? SCNI[134] ? ? NA ? ?
Construction ? NA ? ? NA ? ?
SPH Compatible NA NA NA NA NA NA NA
Smoothing SPHc ? ? ?? ? ? ?
Construction NA NA NA NA NA NA NA
1630001

Irregular-FDM or Compatible RPIM-FDd NA NA NA NA ? NA

GFDMe
Hermite-type Smoothing GSMa NA NA NA NA ? NA
Construction ?? NA NA NA NA ? NA
Element-based Compatible NA FEMk NA NA NA LS-FEMw BEMz
Smoothing NA NA NA NA S-FEMt ? NA
Construction NA Hybrid FEMl NA NA αFEMu ? NA
2nd Reading
 Int. J. Comput. Methods Downloaded from [Link]
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Note: NA — not applicable or no advantages; ? — Not yet known to the author; ?? — Advised to try by the author; Underlined-symmetry formulation.
a Liu et al. [2008c], Liu and Xu [2008], Xu et al. [2010, 2012], Li et al. [2007, 2012a, 2012b], Yao et al. [2013, 2015], Yao and Liu [2014, 2025], Wang

et al. [2013, 2014], Zhang et al. [2008] and Xu et al. [2010].

b Liu et al. [2005b] and Zhou and Liu [2006].
c Liu and Liu [2003], Liu and Gu [2005], Liu [2002], Lucy [1977], Gingold and Monaghan [1977], Monaghan [1982], Liu et al. [2002, 2008b] and Zhou

et al. [2007].
d Liu et al. [2006a].
e Liszka and Orkisz [1980] and Jensen [1980].
f Liu and Gu [1999, 2001a, 2001b, 2002, 2003], Gu and Liu [2001] and Liu et al. [2002, 2003a, 2004a, 2009a].
g Wang and Liu [2000, 2002], Liu and Gu [2001c], Wang et al. [2002, 2005], Liu et al. [2003b, 2004b, 2005, 2006, 2007a], Dai et al. [2003, 2004] and
August 4, 2016 14:24 WSPC/0219-8762

Wu and Liu [2003].

h Lancaster and Salkauskas [1981], Nayroles et al. [1992], Belytschko et al. [1994, 1996], Krongauz and Belytschko [1996] and Fleming et al. [1997].
i Chen et al. [2001].
j Liu et al. [1993, 1995, 1997] and Chen et al. [1996].
k Hughes [1987], Belytschko [2000] and Liu and Quek [2013].
l Pian and Wu [2006] and Wu [1982].
m Gu and Liu [2001] and Liu and Gu [2001b, 2002].
196-IJCM

1630001-29
n [51] and Liu et al. [2003b, 2005a].
o Atluri and Zhu [1998] and Atluri et al. [1999].
p Gu and Liu [2005], Liu and Gu [2003a, 2003b] and Liu et al. [2004].
q Liu and Zhang [2008], Liu et al. [2005c] and Zhang et al. [2007].
r Liu et al. [2009a, 2009b, 2009f] and Xu et al. [2010].
1630001

s Liu et al. [2006d, 2008a, 2011a, 2013], Zhang et al. [2007], Zhao et al. [2008, 2009a, 2009b] and Cui et al. [2010a, 2010b].
t Liu et al. [2007b], Dai et al. [2007] and Liu and Nguyen [2010].
u Liu et al. [2009c, 2009g, 2011b].
v Kee et al. [2007, 2008a, 2008b] and Liu et al. [2006c].
w Bochev et al. [2009].
x Gu and Liu [2002, 2003] and Liu and Gu [2004].
y Mukherjee and Mukherjee [1997a, 1997b].
z Brebbia et al. [1984].
2nd Reading

An Overview on Meshfree Methods

 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

analyst needs only to create (or import it from a CAE package) the geometry of the
problem domain, set the boundary and loading conditions, and to prescribe some
solution accuracy requirements. All the rest of tasks should be done automatically
by the computer. For this to happen, we must use automatic mesh generator. In
addition, with such an automatic mesh generator, we can perform the so-called
adaptive analyses [Xu et al. (2010); Yao et al. (2015); Zhang et al. (2008); Xu
et al. (2010); Liu et al. (2006b, 2006c) Kee et al. (2007, 2008a); Tang et al. (2012)].
Currently, what we have available now is the automatic generators for T-cells or
T-mesh. Therefore, any numerical method should be made working well with the
T-cells. It is one of the most important considerations for any practical method for
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

automation in modeling and simulation.

7.2. Computational efficiency of a method

Int. J. Comput. Methods Downloaded from [Link]

In terms of efficiency for solving large practical engineering problems, the compu-
tational efficiency is usually dominated by the CPU time in solving Eq. (29). Nor-
mally, the operations leading to establish the K (or F), is very minimum, unless the
method is extremely complicated. In solving Eq. (29) efficiently, the property of K
is essential, because iterative solver is required for large systems. Therefore, a good
method must be able to produce a well-conditioned K. The condition number of K
should be, in general, as close to 1 as possible (1 is for the simplest identity matrix).
Intensive discussions and evaluations on how the condition number affects the CPU
time of solving Eq. (29) can be found in [98]. Our intensive study has concluded
that methods using smoothing operations generally produce better conditioned K
with small condition numbers, compared to those formulated based on weakforms.
We now suggest some of the high-performance methods that are already avail-
able, based on the author’s experiences and best knowledge. The suggestions are
made considering the overall high performance, in terms of workability with T-cells,
stability (hence reliability), accuracy, and efficiency. These methods are highlighted
in blue in Table 3 and the key features are given below.

7.3. Methods for Group 1 problems such as the solid mechanics

SC-PIM [136]: SC-PIM, W2 form, symmetric K (hence very efficient), works well
with T-cells, best using simple lower order interpolations, best for linear and non-
linear solid mechanics problems (Elliptical PDEs), supper convergent solutions.
S-PIM [123] [Zhang et al. (2007)]: S-PIM, W2 form, symmetric K (hence very
efficient), works well with T-cells, best using lower order interpolations, best for lin-
ear and nonlinear solid mechanics problems (Elliptical PDEs), NS-PIM for upper
bound and volumetric looking free solution, ES-PIM for accurate and efficient
solution.
S-RPIM [Liu et al. (2006d, 2008a, 2011a, 2013); Zhang et al. (2007); Zhao et al.
(2008, 2009a, 2009b); Cui et al. (2010a, 2010b)]: S-RPIM, W2 form, symmetric K

1630001-30
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

(hence very efficient), works well with T-cells, best using lower order interpolations,
best for linear and nonlinear solid mechanics problems. NS-RPIM for upper bound
and volumetric looking free solution, ES-PRIM for accurate and efficient solution.
Particularly robust against irregular nodes.

S-FEM [Liu et al. (2007b); Dai et al. (2007); Liu and Nguyen (2010)]: S-FEM, W2
form, symmetric K (hence very efficient), works well with T-elements, best using
lower order elements, best for linear and nonlinear solid mechanics problems, NS-
FEM for upper bound and volumetric looking free solution, ES-FEM for accurate
and efficient solution.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

αFEM [Liu et al. (2009c, 2009g, 2011b)]: Alpha FEM, W2 form, symmetric K
(hence very efficient), works well with T-elements, best using lower order elements,
best for linear and nonlinear solid mechanics problems (Elliptical PDEs), capable
Int. J. Comput. Methods Downloaded from [Link]

of producing both FEM and NS-FEM solutions, and ultra-accurate solutions.

FEM [Hughes (1987); Belytschko (2000); Liu and Quek (2013)]: FEM, weakform,
symmetric K (hence very efficient), works less well with T-elements, best for solid
mechanics problems (Elliptical PDEs), best for lower bound solutions. Need a qual-
ity quadrilateral (hexahedral) mesh.

7.4. Methods for Group 2 problems such as the CFD

GSM [Liu et al. (2008c); Liu and Xu (2008); Xu et al. (2010, 2012); Li et al.
(2007, 2012a, 2012b); Yao et al. (2013, 2015); Yao and Liu (2014, 2025); Wang
et al. (2013, 2014); Zhang et al. (2008); Xu et al. (2010)]: GSM, strong-form, a
systematic approach to perform smoothing operations to various orders of gradi-
ents/derivatives, and hence made collocation method usable and stable, asymmet-
ric K, works well with T-cells (unstructured grids), best for general CFD problems
(nonlinear hyperbolic PDEs), generally better performance than the widely used
finite volume method (FVM).

SPH [Liu and Liu (2003); Liu and Gu (2005); Liu (2002); Lucy (1977); Gingold and
Monaghan (1977); Monaghan (1982); Liu et al. (2002, 2008b); Zhou et al. (2007)]:
method of SPH, strong-form, asymmetric K, works well with arbitrarily distributed
particles, best for CFD problems (nonlinear hyperbolic PDEs) with heavy domain
fragmentation during the process.

LRPIM [Gu and Liu (2001); Liu and Gu (2001b, 2002)]: Local Radial Point Inter-
polation Method, local weakform, asymmetric K, works less arbitrarily distributed
nodes, best for general CFD problems (nonlinear hyperbolic PDEs). Robust against
node irregularity.

MWS [Gu and Liu (2005); Liu and Gu (2003a, 2003b); Liu et al. (2004)]:
MWS form method, weak-strong-form, asymmetric K, works well for arbitrarily

1630001-31
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

distributed nodes, best for CFD problems (nonlinear hyperbolic PDEs). Simple in
formulation.

7.5. Symmetry versus asymmetry

To solve Eq. (29) efficiently in computation, we hope K is symmetric (in addition
to being positive definite). This is because if K is not symmetric, the CPU time
will be at least doubled, and so the storage (assuming everything else is the same).
Therefore, for Group 1 problems, one shall always use Galerkin type of formulation
underlined in Table 3, unless there are other overwarming reasons. Methods for
group 1 problems may work also for group 2 problems, but it does not make a
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

very good sense either, because the spatial derivatives in CFD PDEs have first-
order terms (for convection effects). This means that K will not be symmetric,
Int. J. Comput. Methods Downloaded from [Link]

even if one uses the symmetric Galerkin formulation. Therefore, the computational
efficiency benefit of a SPD K cannot be gained. Moreover, the use of the Galerkin
form can lead to a large condition number K. This is because of the first term
in Eq. (30) containing (Matrix)T (Matrix) operations, that practically squared the
condition number (that is always larger than one). For this reason, if we can find
a stable strong-form formulation, such as GSM, the computational efficiency shall
usually be much better, due to the simpler formulation as well as possibly smaller
condition number.
On the other hand, methods for group 2 problems may work also for group 1
problems, but it does not make any sense to do so, because it destroy the symmetry
of the original elliptic PED and hence lose at least half the efficiency by doing so.
Therefore, unless there is a very special reason, we shall not use the methods for
group 2 problems for group 1 problems.

7.6. On the terms of weak and strong

As discussed in Sec. 5.2, in a weak and W2 formulation, a weakening operation
is applied directly to the governing equations, and it can generally provide much
better stability to the discretized algebraic system equations. In fact, we can prove
the stability, and many of the proofs can be found in [1]. On the other hand, it is well
known that the strong-form method is in general not stable, because the formulation
process described in Sec. 5.1 itself has no control at all the stability of the resultant
algebraic system equations. However, from the discussions in Secs. 7.3 and 7.4, we
may notice that when smoothed derivatives are used rather than the compatible
ones, the strong-form method becomes much more stable. This is because the system
is somehow weakened at the stage when the function derivatives are approximated.
Although the “weakening” process is different from the weak formulation, some sort
of similar effects are evident. For this reason, strong-form using smoothed derivatives
was termed in as weakform-like formulation [1].
Note also that in the W2 formulation, weakening operations are performed
at both formulation stage and derivative approximation stage. For this double

1630001-32
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

weakening, W2 formulation offers not only the important stability, but also a unique
and important softening effects, enabling being insensitivity to mesh distortions,
working well with T-cells, more accurate solutions, upper boundary solutions, and
volumetric locking-free [1, 97, 98].

7.7. On dynamic problems

The discussion in the paper is about the spatial discretization. For dynamic prob-
lems, we need also to handle the time derivatives. The techniques for time deriva-
tives in meshfree methods are essentially the same as those used in the standard
FEM, which are rather well developed in the FEM community. We can just borrow
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

these techniques, as long as the meshfree method is spatially stable. However, when
not-fully-compatible methods (strong-form, local weakform, weak-strong from, W2
Int. J. Comput. Methods Downloaded from [Link]

form, and boundary equation formulations) are used, there may be the so-called
temporal stability issues. Such temporal stability is rooted at the softness of a W2
model, and hence choice of types of smoothing domains is critical. Detailed studies
can be found in [1, 97].

7.8. On nonlinear problems

The discussion in the paper focuses on linear elastic problems. For nonlinear prob-
lems, we need also to perform iterative analysis, for which the techniques used in the
meshfree methods are essentially the same as those used in the standard FEM (or
FVM for CFD problems). These nonlinear formulation and techniques are already
well developed, and can be naturally extended to the meshfree methods discussed in
this paper, as long as the method is spatially stable for linear problems. For example,
when a W2 method is used, one needs to perform the smoothing operations to the
displacement gradient for nonlinear problems, rather than smoothing the strains in
linear problems. These techniques are also well studied in Chen et al. [2012], Zhang
et al. [2012, 2013] and Jiang et al. [2014, 2015].

7.9. On fracture mechanics problems

When dealing with fracture mechanism problems, these W2 methods have a clear
advantage over other types of methods. This is because the so-called enrichments
can be done with ease in W2 formulations. When a crack is present, there will be
stress singularity at the crack tip, and hence the displacement needs to be enriched
with special terms in order to capture the singularity. Because no differentiation
is performed in the W2 formulations, such enrichments can be done by simple
inclusion of the desired terms in the basic functions at the stage of shape function
construction, without worrying about the integrability in the weakform. In addition,
because the W2 methods work particularity well with T-cells, the automatic mesh
adaptation can be done with ease to capture the new crack-tip positions during
the crack propagation. Moreover, fatigue analysis can also be conducted effectively.

1630001-33
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

These techniques are also well studied in Liu et al. [2010a, 2010b, 2011a, 2011c,
2012], Jiang et al. [2011]; Nguyen-Xuan [2012, 2013]; Chen et al. [2010, 2011a,
2011b, 2011c, 2012].

8. Conclusions
This paper presents a methodological study on possible meshfree methods and on
the exiting methods for solving the PDEs. We have shown that hundreds of meth-
ods can be formulated by combinations of techniques for function approximation,
gradient approximation, and formulation procedures. Our discussion is focused on
solid mechanics problems, and extended also to CFD problems. In summary, the
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

author may suggest the following:

• For solid mechanics problem (nonlinear hyperbolic PDEs), good choices may be
Int. J. Comput. Methods Downloaded from [Link]

SC-PIM, S-PIM, S-RPIM, S-FEM, αFEM, and FEM.

• For CFD problems (Elliptical PDEs), good choices may be GSM, FVM, SPH,
LRPIM, and MWS.
The author has released (for free) the basic source codes for meshfree methods
including SPH, S-PIM, and S-FEM at [Link] Interested
readers may further develop the possible methods using these basic codes.

Acknowledgments
This theoretical basic research is partially sponsored by NSF under the Award
No. DMS-1214188 and the National Natural Science Foundation of China (Grant
No. 11472184). The author would like to thank the valuable support leading to this
paper.

References
Atluri, S. N. and Zhu, T. [1998] “A new meshless local Petrov–Galerkin (MLPG) approach
in computational mechanics,” Comput. Mech. 22, 117–127.
Atluri, S. N., Cho, J. Y. and Kim, H. G. [1999] “Analysis of thin beams, using the meshless
local Petrov–Galerkin method, with generalized moving least squares interpolations,”
Comput. Mech. 24, 334–347.
Belytschko, T. and Bachrach, W. E. [1986] “Efficient implementation of quadrilaterals
with high coarse-mesh accuracy,” Comput. Methods Appl. Mech. Eng. 54, 279–301.
Belytschko, T. and Bindeman, L. P. [1993] “Assumed strain stabilization of the 4-node
quadrilateral with 1-point quadrature for nonlinear problems,” Comput. Methods Appl.
Mech. Eng. 88, 311–340.
Belytschko, T., Krongauz, Y., Organ, D., Fleming, M. and Krysl, P. [1996] “Meshless
method: An overview and recent developments,” Comput. Methods Appl. Mech. Eng.
139, 3–47.
Belytschko, T., Liu, W. K., Moran, B. and Elkhodary, K. [2000] Nonlinear Finite Elements
for Continua and Structures (John Wiley & Sons, New York).
Belytschko, T., Lu, Y. Y. and Gu, L. [1994] “Element-free Galerkin methods,” Int. J.
Numer. Methods Eng. 37, 229–256.

1630001-34
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

Bochev, P. B. and Gunzburger, M. D. [2009] Least-Squares Finite Element Methods

(Springer, New York).
Brebbia, C. A., Telles, J. C. and Wrobel, L. C. [1984] Boundary Element Techniques
(Springer Verlag, Berlin).
Chen, J. S., Pan, C., Wu, C. T. and Liu, W. K. [1996] “Reproducing kernel particle methods
for large deformation analysis of nonlinear structures,” Comput. Methods Appl. Mech.
Eng. 139, 195–228.
Chen, J. S., Wu, C. T. and Yoon, S. Y. [2001] “A stabilized conforming nodal integration
for Galerkin mesh-free methods,” Int. J. Numer. Methods Eng. 50, 435–466.
Chen, L., Liu, G. R. and Zeng, K. Y. [2011c] “A combined extended and edge-based
smoothed finite element method (es-xfem) for fracture analysis of 2d elasticity,” Int.
J. Comput. Methods 8(4), 773–786.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Chen, L., Liu, G. R., Jiang, Y., Zeng, K. and Zhang, J. [2011] “A singular edge-based
smoothed finite element method (ES-FEM) for crack analyses in anisotropic media,”
Eng. Fract. Mech. 78(1), 85–109.
Int. J. Comput. Methods Downloaded from [Link]

Chen, L., Liu, G. R., Nourbakhsh-Nia, N. and Zeng, K. [2010a] “A singular edge-based
smoothed finite element method (ES-FEM) for bimaterial interface cracks,” Comput.
Mech. 45(2–3), 109–125.
Chen, L., Liu, G. R., Zeng, K. and Zhang, J. [2011b] “A novel variable power singular
element in G space with strain smoothing for bi-material fracture analyses,” Eng.
Anal. Boundary Elem. 35(12), 1303–1317.
Chen, L., Rabczuk, T., Bordas, S. P. A., Liu, G. R., Zeng, K. Y. and Kerfriden, P. [2012]
“Extended finite element method with edge-based strain smoothing (ESm-XFEM) for
linear elastic crack growth,” Comput. Methods Appl. Mech. Eng. 209, 250–265.
Chen, L., Zhang, Y. W., Liu, G. R., Nguyen-Xuan, H. and Zhang, Z. Q. [2012] “A stabi-
lized finite element method for certified solution with bounds in static and frequency
analyses of piezoelectric structures,” Comput. Methods Appl. Mech. Eng. 241, 65–81.
Coleman, C. J. [1996] “On the use of radial basis functions in the solution of elliptic
boundary value problems,” Comput. Mech. 17, 418–422.
Cotteral, J. A., Hughes, T. J. R. and Bazilevs, Y. [2009] Isogeometric Analysis (Wiley,
Chichester).
Cui, X. Y., Liu, G. R. and Li, G. Y. [2010a] “Analysis of mindlin-reissner plates using cell-
based smoothed radial point interpolation method,” Int. J. Appl. Mech. 2(3), 653–680,
doi:10.1142/S1758825110000706.
Cui, X. Y., Liu, G. R. and Li, G. Y. [2010b] “A cell-based smoothed radial point interpola-
tion method (CS-RPIM) for static and free vibration of solids,” Eng. Anal. Boundary
Elem. 34(2), 144–157.
Dai, K. Y. and Liu, G. R. [2007] “Free and forced vibration analysis using the smoothed
finite element method (SFEM),” J. Sound Vib. 301, 803–820.
Dai, K. Y., Liu, G. R. and Nguyen, T. T. [2007] “An n-sided polygonal smoothed finite
element method (nSFEM) for solid mechanics,” Finite Elem. Anal. Design 43, 847–
860.
Dai, K. Y., Liu, G. R., Lim, K. M. and Gu, Y. T. [2003] “Comparison between the radial
point interpolation and the kriging interpolation used in meshfree methods,” Comput.
Mech. 32(1–2), 60–70.
Dai, K. Y., Liu, G. R., Lim, K. M., Han, X. and Du, S. Y. [2004] “A meshfree radial
point interpolation method for analysis of functionally graded material (FGM) plates,”
Comput. Mech. 34(3), 213–223.
Duarte, C. A. and Oden, J. T. [1996] “An hp adaptive method using clouds,” Comput.
Methods Appl. Mech. Eng. 139, 237–262.

1630001-35
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

Eringen, A. C. [1972] “Nonlocal polar elastic continua,” Int. J. Eng. Sci. 10, 1.
Fasshauer, G. E. [1997] “Solving partial differential equations by collocation with radial
basis functions,” in Surface Fitting and Multiresolution Methods, eds. Mehaute, A. L.,
Rabut, C. and Schumaker, L. L., pp. 131–138 (Vanderbilt University Press, New York).
Fleming, M., Chu, Y. A., Moran, B. and Belytschko, T. [1997] “Enriched element-free
Galerkin methods for crack tip fields,” Int. J. Numer. Methods Eng. 40, 1483–1504.
Gingold, R. A. and Monaghan, J. J. [1977] “Smooth particle hydrodynamics: Theory and
applications to non-spherical stars,” Mon. Not. R. Astron. Soc. 181, 375–389.
Golberg, M. A., Chen, C. S. and Bowman, H. [1999] “Some recent results and proposals
for the use of radial basis functions in the BEM,” Eng. Anal. Boundary Elem. 23,
285–296.
Gu, L. [2003] “Moving Kriging interpolation and element-free Galerkin method,” Int. J.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Numer. Methods Eng. 56, 1–11.

Gu, Y. T. and Liu, G. R. [2001] “A local point interpolation method for static and dynamic
analysis of thin beams,” Comput. Methods Appl. Mech. Eng. 190(42), 5515–5528.
Int. J. Comput. Methods Downloaded from [Link]

Gu, Y. T. and Liu, G. R. [2002] “A boundary point interpolation method for stress analysis
of solids,” Comput. Mech. 28(1), 47–54.
Gu, Y. T. and Liu, G. R. [2003] “Hybrid boundary point interpolation methods and their
coupling with the element free galerkin method,” Eng. Anal. Boundary Elem. 27(9),
905–917.
Gu, Y. T. and Liu, G. R. [2005] “A meshfree weak-strong (MWS) form method for time
dependent problems,” Comput. Mech. 35(2), 134–145.
Hardy, R. L. [1990] “Theory and applications of the multiquadrics — biharmonic method
(20 years of discovery 1968–1988),” Comput. Math. Appl. 19, 163–208.
Hughes, T. J. R. [1987] The Finite Element Method: Linear Static and Dynamic Finite
Element Analysis (Prentice-Hall, Englewood Cliffs, New Jersey).
Jensen, P. S. [1980] “Finite difference techniques for variable grids,” Comput. Struct. 2,
17–29.
Jiang, C., Liu, G., Han, X., Zhang, Z. and Zeng, W. [2015] “A smoothed finite element
method for analysis of anisotropic large deformation of passive rabbit ventricles in
diastole,” Int. J. Numer. Methods Biomed. Eng. 31(1), 1–25.
Jiang, C., Zhang, Z., Han, X. and Liu, G. [2014] “Selective smoothed finite element meth-
ods for extremely large deformation of anisotropic incompressible bio-tissues,” Int. J.
Numer. Methods Eng. 99(8), 587–610.
Jiang, Y., Liu, G. R., Zhang, Y. W., Chen, L. and Tay, T. E. [2011] “A singular ES-
FEM for plastic fracture mechanics,” Comput. Methods Appl. Mech. Eng. 200(45–46),
2943–2955.
Kansa, E. J. [1990] “A scattered data approximation scheme with application to compu-
tational fluid-dynamics — I & II,” Comput. Math. Appl. 19, 127–161.
Kee, B. B. T., Liu, G. R. and Lu, C. [2007] “A regularized least-squares radial point
collocation method (RLS-RPCM) for adaptive analysis,” Comput. Mech. 40(5), 837–
853.
Kee, B. B. T., Liu, G. R. and Lu, C. [2008a] “A least-square radial point collocation
method for adaptive analysis in linear elasticity,” Eng. Anal. Boundary Elem. 32(6),
440–460.
Kee, B. B. T., Liu, G. R., Zhang, G. Y. and Lu, C. [2008b] “A residual based error
estimator using radial basis functions,” Finite Elem. Anal. Design 44(9–10), 631–645.
Krongauz, Y. and Belytschko, T. [1996] “Enforcement of essential boundary conditions in
meshless approximations using finite elements,” Comput. Methods Appl. Mech. Eng.
131(1–2), 133–145.

1630001-36
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

Lancaster, P. and Salkauskas, K. [1981] “Surfaces generated by moving least squares meth-
ods,” Math. Comput. 37, 141–158.
Li, E., Liu, G. R., Xu, G. X., Vincent, T. and He, Z. C. [2012a] “Numerical modeling and
simulation of pulsatile blood flow in rigid vessel using gradient smoothing method,”
Eng. Anal. Boundary Elem. 36(3), 322–334.
Li, E., Tan, V., Xu, G. X., Liu, G. R. and He, Z. C. [2011] “A novel linearly-weighted
gradient smoothing method (lwgsm) in the simulation of fluid dynamics problem,”
Comput. Fluids 50, 104–119.
Li, E., Tan, V., Xu, G. X., Liu, G. R. and He, Z. C. [2012b] “A novel alpha gradient
smoothing method (alpha GSM) for fluid problems,” Numer. Heat Transfer B Fundam.
61(3), 204–228.
Liszka, T. and Orkisz, J. [1980] “The finite difference method at arbitrary irregular grids
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

and its application in applied mechanics,” Comput. Struct. 11, 83–95.

Liu, G. R. [2002] Meshfree Methods: Moving Beyond the Finite Element Method (CRC
Press, Boca Raton).
Int. J. Comput. Methods Downloaded from [Link]

Liu, G. R. [2008] “A generalized Gradient smoothing technique and the smoothed bilin-
ear form for Galerkin formulation of a wide class of computational methods,” Int. J.
Comput. Methods 5(2), 199–236.
Liu, G. R. [2009] “On G space theory,” Int. J. Comput. Methods 6(2), 257–289.
Liu, G. R. [2010] “A G space theory and a weakened weak (W2) form for a unified formula-
tion of compatible and incompatible methods: Part I theory and Part II applications,”
Int. J. Numer. Methods Eng. 81(9), 1093–1156.
Liu, G. R. and Gu, Y. T. [1999] “A point interpolation method,” in Proc. 4th Asia-Pacific
Conf. Computational Mechanics, December, Singapore, pp. 1009–1014.
Liu, G. R. and Gu, Y. T. [2001a] “A point interpolation method for two-dimensional
solids,” Int. J. Numer. Methods Eng. 50, 937–951.
Liu, G. R. and Gu, Y. T. [2001b] “A local point interpolation method for stress analysis
of two-dimensional solids,” Struct. Eng. and Mech. 11(2), 221–236.
Liu, G. R. and Gu, Y. T. [2001c] “A local radial point interpolation method (LRPIM) for
free vibration analyses of 2-D solids,” J. Sound Vib. 246(1), 29–46.
Liu, G. R. and Gu, Y. T. [2002] “Comparisons of two meshfree local point interpolation
methods for structural analyses,” Comput. Mech. 29(2), 107–121.
Liu, G. R. and Gu, Y. T. [2003] “A matrix triangularization algorithm for the polynomial
point interpolation method,” Comput. Methods Appl. Mech. Eng. 192(19), 2269–2295.
Liu, G. R. and Gu, Y. T. [2003a] “A meshfree method: Meshfree weak-strong (MWS) form
method for 2-D solids,” Comput. Mech. 33(1), 2–14.
Liu, G. R. and Gu, Y. T. [2004] “Boundary meshfree methods based on the boundary
point interpolation methods,” Eng. Anal. Boundary Elem. 28(5), 475–487.
Liu, G. R. and Gu, Y. T. [2005] An Introduction to Meshfree Method Methods and their
Programming (Springer, The Netherlands).
Liu, G. R. and Han, X. [2003] Computational Inverse Techniques in Nondestructive Eval-
uation (CRC Press, Boca Raton).
Liu, G. R. and Liu, M. B. [2003] Smoothed Particle Hydrodynamics: A Meshfree Particle
Method (World Scientific, New Jersey).
Liu, G. R. and Nguyen, T. T. [2010] Smoothed Finite Element Methods (CRC, USA).
Liu, G. R. and Quek, S. S. [2003] The Finite Element Method: A Practical Course
(Butterworth Heinemann, Oxford).
Liu, G. R. and Quek, S. S. [2013] The Finite Element Method: A Practical Course, 2nd
edn. (Butterworth Heinemann, Oxford).

1630001-37
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

Liu, G. R. and Xu, G. X. [2008] “A gradient smoothing method (GSM) for fluid dynamics
problems,” Int. J. Numer. Methods Fluids 56(10), 1101–1133.
Liu, G. R. and Zhang, G. Y. [2008] “Upper bound solution to elasticity problems: A
unique property of the linearly conforming point interpolation method (LC-PIM),”
Int. J. Numer. Methods Eng. 74(7), 1128–1161.
Liu, G. R. and Zhang, G. Y. [2008] “Edge-based smoothed point interpolation methods,”
Int. J. Comput. Methods 5(4), 621–646.
Liu, G. R. and Zhang, G. Y. [2013] Smoothed Point Interpolation Methods — G Space
Theory and Weakened Weakforms (World Scientific, Singapore).
Liu, G. R., Chen, L., Nguyen-Thoi, T., Zeng, K. Y. and Zhang, G. Y. [2010a] “A novel
singular node-based smoothed finite element method (NS-FEM) for upper bound solu-
tions of fracture problems,” Int. J. Numer. Methods Eng. 83(11), 1466–1497.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Liu, G. R., Dai, K. Y. and Lim, K. M. [2004b] “Static and vibration control of composite
laminates integrated with piezoelectric sensors and actuators using the radial point
interpolation method,” Smart Mater. Struct. 13(6), 1438–1447.
Int. J. Comput. Methods Downloaded from [Link]

Liu, G. R., Dai, K. Y. and Nguyen, T. T. [2007b] “A smoothed finite element method for
mechanics problems,” Comput. Mech. 39, 859–877.
Liu, G. R., Dai, K. Y. and Varadan, V. K. [2003a] “Meshfree point interpolation meth-
ods for analysis of piezoelectric structures,” Smart Mater., Struct., Syst. 5062(Pts. 1
and 2), 305–317.
Liu, G. R., Dai, K. Y., Lim, K. M. and Gu, Y. T. [2002] “A point interpolation mesh free
method for static and frequency analysis of two-dimensional piezoelectric structures,”
Comput. Mech. 29(6), 510–519.
Liu, G. R., Dai, K. Y., Lim, K. M. and Gu, Y. T. [2003b] “A radial point interpola-
tion method for simulation of two-dimensional piezoelectric structures,” Smart Mater.
Struct. 12(2), 171–180.
Liu, G. R., Gu, Y. T. and Dai, K. Y. [2004a] “Assessment and applications of point
interpolation methods for computational mechanics,” Int. J. Numer. Methods Eng.
59(10), 1373–1397.
Liu, G. R., Huynh, D. B. P. and Gu, Y. [2006b] “An adaptive meshfree collocation method
for static and dynamic nonlinear problems,” eds. Liu, G. R., Tan, V. B. C. and Han, X.,
doi:10.1007/978-1-4020-3953-9 70.
Liu, G. R., Jiang, Y., Chen, L., Zhang, G. Y. and Zhang, Y. W. [2011a] “A singular cell-
based smoothed radial point interpolation method for fracture problems,” Comput.
Struct. 89(13–14), 1378–1396.
Liu, G. R., Kee, B. B. T. and Chun, L. [2006c] “A stabilized least-squares radial point
collocation method (LS-RPCM) for adaptive analysis,” Comput. Methods Appl. Mech.
Eng. 195(37–40), 4843–4861.
Liu, G. R., Li, Y., Dai, K. Y., Luan, M. T. and Xue, W. [2006d] “A linearly conform-
ing radial point interpolation method for solid mechanics problems,” Int. J. Comput.
Methods 3(4), 401–428.
Liu, G. R., Liu, M. B. and Lam, K. Y. [2002] “A general approach for constructing smooth-
ing functions for meshfree methods,” Ninth Int. Conf. Computing in Civil and Building
Engineering, 3–5 April, pp. 431–436.
Liu, G. R., Nguyen, T. T. and Lam, K. Y. [2008d] “A novel Alpha Finite Element Method
(αFEM) for exact solution to mechanics problems using triangular and tetrahedral
elements,” Comput. Methods Appl. Mech. Eng. 197, 3883–3897.
Liu, G. R., Nguyen, T. T., Dai, K. Y. and Lam, K. Y. [2007] “Theoretical aspects of
the smoothed finite element method (SFEM),” Int. J. Numer. Methods Eng. 71,
902–930.

1630001-38
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

Liu, G. R., Nguyen-Thoi, T. and Lam, K. Y. [2009c] “A novel FEM by scaling the gradient
of strains with factor alpha (alpha FEM),” Comput. Mech. 43(3), 369–391.
Liu, G. R., Nguyen-Thoi, T. and Lam, K. Y. [2009e] “An edge-based smoothed finite
element method (ES-FEM) for static, free and forced vibration analyses of solids,”
J. Sound Vib. 320(4–5), 1100–1130.
Liu, G. R., Nguyen-Thoi, T., Nguyen-Xuan, H. and Lam, K. Y. [2009d] “A node-based
smoothed finite element method (NS-FEM) for upper bound solutions to solid mechan-
ics problems,” Comput. Struct. 87(1–2), 14–26.
Liu, G. R., Nguyen-Xuan, H. and Nguyen, T. T. [2011b] “A variationally consistent
alpha FEM (VC alpha FEM) for solution bounds and nearly exact solution to solid
mechanics problems using quadrilateral elements,” Int. J. Numer. Methods Eng. 85(4),
461–497.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Liu, G. R., Nguyen-Xuan, H., Nguyen-Thoi, T. and Xu, X. [2009g] “A novel galerkin-
like weakform and a superconvergent alpha finite element method (S alpha FEM) for
mechanics problems using triangular meshes,” J. Comput. Phys. 228(11), 4055–4087.
Int. J. Comput. Methods Downloaded from [Link]

Liu, G. R., Nourbakhshnia, N. and Zhang, Y. W. [2011c] “A novel singular ES-FEM

method for simulating singular stress fields near the crack tips for linear fracture prob-
lems,” Eng. Fract. Mech. 78(6), 863–876.
Liu, G. R., Nourbakhshnia, N., Chen, L. and Zhang, Y. W. [2010b] “A novel general
formulation for singular stress field using the es-fem method for the analysis of mixed-
mode cracks,” Int. J. Comput. Methods 7(1), 191–214.
Liu, G. R., Wu, Y. L. and Ding, H. [2004] “Meshfree weak-strong (MWS) form method
and its application to incompressible flow problems,” Int. J. Numer. Methods Fluids
46(10), 1025–1047.
Liu, G. R., Xu, X., Zhang, G. Y. and Gu, Y. T. [2009a] “An extended galerkin weakform
and a point interpolation method with continuous strain field and superconvergence
using triangular mesh,” Comput. Mech. 43(5), 651–673.
Liu, G. R., Xu, X., Zhang, G. Y. and Nguyen-Thoi, T. [2009b] “A superconvergent point
interpolation method (SC-PIM) with piecewise linear strain field using triangular
mesh,” Int. J. Numer. Methods Eng. 77(10), 1439–1467.
Liu, G. R., Zhang, G. Y., Dai, K. Y., Wang, Y. Y., Zhong, Z. H., Li, G. Y. and Han,
X. [2005c] “A linearly conforming point interpolation method (LC-PIM) for 2D solid
mechanics problems,” Int. J. Comput. Methods 2(4), 645–665.
Liu, G. R., Zhang, G. Y., Gu, Y. T. and Wang, Y. Y. [2005a] “A meshfree radial point
interpolation method (RPIM) for three-dimensional solids,” Comput. Mech. 36(6),
421–430.
Liu, G. R., Zhang, G. Y., Liu, G. R. and Zhang, G. Y. [2009f] “A novel scheme of strain-
constructed point interpolation method for static and dynamic mechanics problems,”
Int. J. Appl. Mech. 1(1), 233–258.
Liu, G. R., Zhang, G. Y., Wang, Y. Y., Zhong, Z. H., Li, G. Y. and Han, X. [2007a] “A nodal
integration technique for meshfree radial point interpolation method (NI-RPIM),” Int.
J. Solids Struct. 44(11–12), 3840–3860.
Liu, G. R., Zhang, G. Y., Zong, Z. and Li, M. [2013] “Meshfree cell-based smoothed alpha
radial point interpolation method (CS-alpha RPIM) for solid mechanics problems,”
Int. J. Comput. Methods 10(4), 1350020.
Liu, G. R., Zhang, J. and Lam, K. Y. [2008c] “A gradient smoothing method (GSM) with
directional correction for solid mechanics problems,” Comput. Mech. 41, 457–472.
Liu, G. R., Zhang, J., Li, H., Lam, K. Y. and Kee, B. B. T. [2006a] “Radial point interpo-
lation based finite difference method for mechanics problems,” Int. J. Numer. Methods
Eng. 68(7), 728–754.

1630001-39
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

Liu, G. R., Zhao, X., Dal, K. Y., Zhong, Z. H., Li, G. Y. and Han, X. [2008a] “Static and
free vibration analysis of laminated composite plates using the conforming radial point
interpolation method,” Compos. Sci. Technol. 68(2), 354–366.
Liu, M. B., Liu, G. R. and Zong, Z. [2008b] “An overview on smoothed particle hydrody-
namics,” Int. J. Comput. Methods 5(1), 135–188.
Liu, P., Bui, T. Q., Zhang, C., Yu, T. T., Liu, G. R. and Golub, M. V. [2012] “The singular
edge-based smoothed finite element method for stationary dynamic crack problems in
2D elastic solids,” Comput. Methods Appl. Mech. Eng. 233, 68–80.
Liu, W. K., Jun, S. and Zhang, Y. [1995] “Reproducing kernel particle methods,” Int. J.
Numer. Methods Fluids 20, 1081–1106.
Liu, W. K., Li, S. F. and Belytschko, T. [1997] “Moving least-square reproducing kernel
methods: I Methodology and convergence,” Comput. Methods Appl. Mech. Eng. 143,
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

113–154.
Liu, X., Liu, G. R., Tai, K. and Lam, K. Y. [2005b] “Radial point interpolation collocation
method (RPICM) for the solution of nonlinear Poisson problems,” Comput. Mech.
Int. J. Comput. Methods Downloaded from [Link]

36(4), 298–306.
Lucy, L. [1977] “A numerical approach to testing the fission hypothesis,” Astron. J. 82,
1013–1024.
Melenk, J. M. and Babuska, I. [1996] The Partition of Unity Finite Element Method: Basic
Theory and Applications, TICAM Report 96-01, University of Texas, Austin.
Monaghan, J. J. [1982] “Why particle methods work,” Siam J. Sci. Sat. Comput. 3(4),
423–433.
Mukherjee, Y. X. and Mukherjee, S. [1997a] “Boundary node method for potential prob-
lems,” Int. J. Numer. Methods Eng. 40, 797–815.
Mukherjee, Y. X. and Mukherjee, S. [1997b] “On boundary conditions in the element-free
Galerkin method,” Comput. Mech. 19, 264–270.
Nayroles, B., Touzot, G. and Villon, P. [1992] “Generalizing the finite element method:
Diffuse approximation and diffuse elements,” Comput. Mech. 10, 307–318.
Nguyen-Thoi, T., Liu, G. R., Lam, K. Y. and Zhang, G. Y. [2009] “A face-based smoothed
finite element method (FS-FEM) for 3D linear and geometrically non-linear solid
mechanics problems using 4-node tetrahedral elements,” Int. J. Numer. Methods Eng.
78(3), 324–353.
Nguyen-Xuan, H., Liu, G. R., Bordas, S., Natarajan, S. and Rabczuk, T. [2013] “An
adaptive singular ES-FEM for mechanics problems with singular field of arbitrary
order,” Comput. Methods Appl. Mech. Eng. 253, 252–273.
Nguyen-Xuan, H., Liu, G. R., Nourbakhshnia, N. and Chen, L. [2012] “A novel singular
ES-FEM for crack growth simulation,” Eng. Fract. Mech. 84, 41–66.
Onate, E. et al. [1996] “A finite point method in computational mechanics applications to
convective transport and fluid flow,” Int. J. Numer. Methods Eng. 39, 3839–3866.
Pian, T. H. H. and Wu, C. C. [2006] “Hybrid and Incompatible finite element methods,”
(CRC Press, Boca Raton).
Powell, M. J. D. [1992] “The theory of radial basis function approximation in 1990,” in
Advances in Numerical Analysis, ed. Light, F. W. (Oxford University Press, Oxford),
pp. 303–322.
Sharan, M., Kansa, E. J. and Gupta, S. [1997] “Application of the multiquadric method
for numerical solution of elliptic partial differential equations,” Appl. Math. Comput.
84, 275–302.
Simo, J. C. and Hughes, T. J. R. [1998] Computational Inelasticity (Springer-Verlag, New
York).
Tang, Q., Zhang, G. Y., Liu, G. R., Zhong, Z. H. and He, Z. C. [2012] “An effi-
cient adaptive analysis procedure using the edge-based smoothed point interpolation
1630001-40
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

An Overview on Meshfree Methods

method (ES-PIM) for 2D and 3D problems,” Eng. Anal. Boundary Elem. 36(9), 1424–
1443.
Wang, J. G. and Liu, G. R. [2000] “Radial point interpolation method for elastoplastic
problems,” Proc. 1st Int. Conf. Structural Stability and Dynamics, Taipei, China, 7–9
December, pp. 703–708.
Wang, J. G. and Liu, G. R. [2002a] “A point interpolation meshless method based on
radial basis functions,” Int. J. Numer. Methods Eng. 54, 1623–1648.
Wang, J. G. and Liu, G. R. [2002b] “On the optimal shape parameters of radial basis
functions used for 2-D meshless methods,” Comput. Methods Appl. Mech. Eng. 191,
2611–2630.
Wang, J. G., Liu, G. R. and Lin, P. [2002] “Numerical analysis of biot’s consolidation
process by radial point interpolation method,” Int. J. Solids Struct. 39(6), 1557–
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

1573.
Wang, J. G., Yan, L. and Liu, G. R. [2005] “A local radial point interpolation method for
dissipation process of excess pore water pressure,” Int. J. Numer. Methods Heat Fluid
Int. J. Comput. Methods Downloaded from [Link]

Flow 15(5–6), 567–587.

Wang, S., Khoo, B. C., Liu, G. R. and Xu, G. X. [2013] “An arbitrary lagrangian–eulerian
gradient smoothing method (gsm/ale) for interaction of fluid and a moving rigid body,”
Comput. Fluids 71, 327–347.
Wang, S., Khoo, B. C., Liu, G. R., Xu, G. X. and Chen, L. [2014] “Coupling GSM/ALE
with ES-FEM-T3 for fluid-deformable structure interactions,” J. Comput. Phys. 276,
315–340.
Wendland, H. [1998] “Error estimates for interpolation by compactly supported radial
basis functions of minimal degree,” J. Approx. Theor. 93, 258–396.
Wu, H. C. [1982] Variational Principle in Elasticity and Applications (Scientific Press,
Beijing).
Wu, Y. L. and Liu, G. R. [2003] “A meshfree formulation of local radial point interpolation
method (LRPIM) for incompressible flow simulation,” Comput. Mech. 30(5–6), 355–
365.
Xu, G. X., Li, E., Tan, V. and Liu, G. R. [2012] “Simulation of steady and unsteady
incompressible flow using gradient smoothing method (gsm),” Comput. Struct. 90–
91, 131–144.
Xu, G. X., Liu, G. R. and Tani, A. [2010] “An adaptive gradient smoothing method (gsm)
for fluid dynamics problems,” Int. J. Numer. Methods Fluids 62, 499–529.
Xu, X., Liu, G. R. and Zhang, G. Y. [2009] “A point interpolation method with least
square strain field (PIM-LSS) for solution bounds and ultra-accurate solutions using
triangular mesh,” Comput. Methods Appl. Mech. Eng. 198(17–20), 1486–1499.
Xu, X., Liu, G. R., Gu, Y. T., Zhang, G. Y., Luo, J. W. and Peng, J. X. [2010] “A
point interpolation method with locally smoothed strain field (PIM-LS2) for mechanics
problems using triangular mesh,” Finite Elem. Anal. Design 46(10), 862–874.
Yao, J. and Liu, G. R. [2014] “A matrix-form gsm-cfd solver for incompressible fluids and
its application to hemodynamics,” Comput. Mech. 54(4), 999–1012.
Yao, J., Lin, T., Liu, G. R. and Chen, C. L. [2015] “An adaptive gsm-cfd solver and its
application to shockwave boundary layer interaction,” Int. J. Numer. Methods Heat
Fluid Flow 25(6), 1282–1310.
Yao, J., Liu, G. R. and Chen, C.-L. [2013] “A moving-mesh gradient smoothing method
for ompressible cfd problems,” Math. Models Methods Appl. Sci. 23(2), 273–305.
Zhang, G. Y., Liu, G. R., Nguyen, T. T., Song, C. X., Han, X., Zhong, Z. H. and Li, G. Y.
[2007] “The upper bound property for solid mechanics of the linearly conforming radial
point interpolation method (LC-RPIM),” Int. J. Comput. Methods 4(3), 521–541.

1630001-41
 2nd Reading
August 4, 2016 14:24 WSPC/0219-8762 196-IJCM 1630001

G. R. Liu

Zhang, G. Y., Liu, G. R., Wang, Y. Y., Huang, H. T., Zhong, Z. H., Li, G. Y. and
Han, X. [2007] “A linearly conforming point interpolation method (LC-PIM) for three-
dimensional elasticity problems,” Int. J. Numer. Methods Eng. 72, 1524–1543.
Zhang, J., Liu, G. R., Lam, K. Y., Li, H. and Xue, G. [2008] “A gradient smoothing
method (GSM) based on strong-form governing equation for adaptive analysis of solid
mechanics problems,” Finite Elem. Anal. Design 44(15), 889–909.
Zhang, Z., Liu, G. R. and Khoo, B. C. [2012] “Immersed smoothed finite element method
for two dimensional fluid-structure interaction problems,” Int. J. Numer. Methods Eng.
90(10), 1292–1320.
Zhang, Z., Liu, G. R. and Khoo, B. C. [2013] “A three dimensional immersed smoothed
finite element method (3D IS-FEM) for fluid-structure interaction problems,” Comput.
Mech. 51(2), 129–150.
by NANYANG TECHNOLOGICAL UNIVERSITY on 08/08/16. For personal use only.

Zhao, X., Liu, G. R., Dai, K. Y., Zhong, Z. H., Li, G. Y. and Han, X. [2008] “Geometric
nonlinear analysis of plates and cylindrical shells via a linearly conforming radial point
interpolation method,” Comput. Mech. 42(1), 133–144.
Int. J. Comput. Methods Downloaded from [Link]

Zhao, X., Liu, G. R., Dai, K. Y., Zhong, Z. H., Li, G. Y. and Han, X. [2009a] “Free-
vibration analysis of shells via a linearly conforming radial point interpolation method
(LC-RPIM),” Finite Elem. Anal. Design 45(12), 917–924.
Zhao, X., Liu, G. R., Dai, K. Y., Zhong, Z. H., Li, G. Y. and Han, X. [2009b] “A linearly
conforming radial point interpolation method (LC-RPIM) for shells,” Comput. Mech.
43(3), 403–413.
Zhou, B. H. and Liu, G. R. [2006] “3-D heat transfer analysis using a collocation method
together with RPIM shape functions and FIC boundary conditions,” eds. Liu, G. R.,
Tan, V. B. C. and Han, X., doi:10.1007/978-1-4020-3953-9 84.
Zhou, C. E., Liu, G. R. and Lou, K. Y. [2007] “Three-dimensional penetration simulation
using smoothed particle hydrodynamics, Int. J. Comput. Methods 4(4), 671–691.

1630001-42