Sunteți pe pagina 1din 35

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2010; 84:1222–1256


Published online 25 May 2010 in Wiley Online Library (wileyonlinelibrary.com). DOI: 10.1002/nme.2941

A theoretical study on the smoothed FEM (S-FEM) models:


Properties, accuracy and convergence rates

G. R. Liu1,2 , H. Nguyen-Xuan3,4, ∗, † and T. Nguyen-Thoi3,4


1 Center for Advanced Computations in Engineering Science (ACES), Department of Mechanical Engineering,
National University of Singapore, 9 Engineering Drive 1, Singapore 117576, Singapore
2 Singapore-MIT Alliance (SMA), E4-04-10, 4 Engineering Drive 3, Singapore 117576, Singapore
3 Department of Mechanics, Faculty of Mathematics and Computer Science, University of Science,

National University of Vietnam—HCM, 227 Nguyen Van Cu, Dist. 5, Ho Chi Minh City, Vietnam
4 Faculty of Civil Engineering, Ton Duc Thang University, 98 Ngo Tat To St., War 19, Binh Thanh Dist.,

Ho Chi Minh City, Vietnam

SUMMARY
Incorporating the strain smoothing technique of meshfree methods into the standard finite element method
(FEM), Liu et al. have recently proposed a series of smoothed finite element methods (S-FEM) for
solid mechanics problems. In these S-FEM models, the compatible strain fields are smoothed based on
smoothing domains associated with entities of elements such as elements, nodes, edges or faces, and
the smoothed Galerkin weak form based on these smoothing domains is then applied to compute the
system stiffness matrix. We present in this paper a general and rigorous theoretical framework to show
properties, accuracy and convergence rates of the S-FEM models. First, an assumed strain field derived
from the Hellinger–Reissner variational principle is shown to be identical to the smoothed strain field
used in the S-FEM models. We then define a smoothing projection operator to modify the compatible
strain field and show a set of properties. We next establish a general error bound of the S-FEM models.
Some numerical examples are given to verify the theoretical properties established. Copyright 䉷 2010
John Wiley & Sons, Ltd.

Received 11 May 2009; Revised 8 March 2010; Accepted 10 April 2010

KEY WORDS: numerical methods; meshfree methods; displacement model; smoothed finite element
method (S-FEM); equilibrium model; node-based smoothed finite elements (NS-FEM);
edge-based smoothed finite elements (ES-FEM)

∗ Correspondence to: H. Nguyen-Xuan, Department of Mechanics, Faculty of Mathematics and Computer Science,
University of Science, National University of Vietnam—HCM, 227 Nguyen Van Cu, Dist. 5, Ho Chi Minh City,
Vietnam.
† E-mail: nxhung@hcmuns.edu.vn, URL: http://www.math.hcmuns.edu.vn/∼nxhung/

Contract/grant sponsor: A*Star, Singapore; contract/grant number: SERC 052 101 0048
Contract/grant sponsor: State Key Laboratory of Advanced Technology of Design and Manufacturing for Vehicle
Body, Hunan University, People’s Republic of China; contract/grant number: 40915001

Copyright 䉷 2010 John Wiley & Sons, Ltd.


A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1223

1. INTRODUCTION

Meshfree methods have been studied intensively in the last decade, and many techniques have been
developed [1]. The numerical operations used in meshfree methods are in generally beyond the
elements, and hence are distinct in many ways from the standard finite element method (FEM). First,
the function approximations or interpolations are essentially node-based without element-based
mapping. Second, the integration and the interpolation are basically performed separately. Third,
strain fields in meshfree methods are often constructed based on ‘information’ of the surrounding
cells. Last, the generation of the cells used in a meshfree method is performed automatically.
One of the main issues in many meshfree methods is the instable phenomenon due to the under-
integration of the weak form inherent in the nodal integration [2–5]. To stabilize this instability,
a strain smoothing technique was proposed by Chen et al. [6] in the context of Galerkin meshfree
method, and it was then further applied to the natural element methods by Yoo et al. [7], and a
series of linearly conforming point interpolation methods (PIMs) by Liu and coworkers [8–12].
In [13], Cescotto and Li have shown that such a strain smoothing field can be recovered in their
approach based on the Fraeijs de Veubeke (FdV) variational principle.
Incorporating the strain smoothing technique of meshfree methods into the finite element
methods, Liu et al. have formulated a series of smoothed FEM (S-FEM) models named as cell-based
S-FEM (CS-FEM) [14–23], node-based S-FEM (NS-FEM) [24–26], edge-based S-FEM (ES-FEM)
[27–32], face-based (FS-FEM) [33, 34] and alpha-FEM [35] that use linear interpolations. In these
S-FEM models, the finite element mesh is used similarly as in the standard FEM. However, these
S-FEM models evaluate the weak form based on smoothing domains created from the entities of
the element mesh such as cells/elements (Figure 1), or nodes (Figure 2), or edges (Figure 3) or
faces (Figure 4). These smoothing domains can be located inside the elements (CS-FEM) or cover
parts of adjacent elements (NS-FEM, ES-FEM and FS-FEM). They are linearly independent and

3 3
6
4 4

1 2 1 2
x 5
(a) (b)

7 3 3
4 4
9 6
y 8

1 5 2 1 2
x
(c) (d)
: field nodes : virtual nodes to form the smoothing domains

Figure 1. Division of quadrilateral element into the smoothing domains (SDs) in the
CS-FEM by connecting the mid-segment-points of opposite segments of smoothing
domains: (a) 1 SD; (b) 2 SDs; (c) 4 SDs; and (d) 8 SDs.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1224 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

Figure 2. n-sided polygonal elements and the smoothing domains (k) (shaded areas)
associated with node k in the NS-FEM.

Figure 3. Division of domain into triangular element and smoothing domains (k) connected to
edge k of triangular elements in the ES-FEM.

ensure stability and convergence of the S-FEM models. Applying the strain smoothing technique
[6] on smoothing domains will help to soften the over-stiffness of the standard FEM model, and
hence can improve significantly the accuracy of solutions in both displacement and stress. Owing
to using the different smoothing domains in the S-FEM models, the softening effect of strain
smoothing technique on the over-stiffness of the standard FEM model will be different. Therefore,
each of the S-FEM models has different properties, advantages and disadvantages.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1225

Figure 4. Two tetrahedral elements and the smoothing domains (shaded areas)
associated with face k in the FS-FEM.

The numerical examples have so far illustrated many interesting properties of the S-FEM models
compared with the counterpart standard FEM models. For example (1) the S-FEM models are
more accurate and have higher convergence rates in displacement or stress solutions or both; (2) no
mapping or coordinate transformation is involved in the S-FEM models because of using the
simple PIM [8–12]. Therefore, the numerical procedure in the S-FEM models is performed only
on the physical elements, the problem domain can be discretized in more flexible ways and even
polygonal elements with arbitrary number of sides can be used; (3) the NS-FEM also possesses
many interesting properties similar to those of an equilibrium model [36–40] such as the upper
bound in strain energy, freedom of the volumetric locking, super-accuracy and super-convergence
in the stress solution, and many other properties. Besides the interesting advantages mentioned
above, the S-FEM models also possess a disadvantage compared with the FEM counterpart. Because
the smoothing domains have more supporting nodes than the elements have, the bandwidths of
stiffness matrix in the S-FEM models (NS-FEM, ES-FEM and FS-FEM) are larger than those
of the FEM models and hence the computational cost of the S-FEM models are also higher for
models of the same sets of nodes. However, when the efficiency of computation (computation time
for the same accuracy) in terms of the error estimator versus computational cost is considered, the
S-FEM models are in general more efficient than the FEM models [24, 27, 33].
So far we have only proved that the S-FEM models were stable and convergent to the exact
solution [41]. However, the theoretical bases of properties of the S-FEM models have not yet
been set up and properly proven. This paper hence attempts to establish a rigorous theoretical
framework to prove some of the crucial properties, accuracy and convergence rate of these S-FEM
models. We first prove that an assumed strain field derived from the variational principle is identical
to the smoothed strain field used in the S-FEM models, implying that the S-FEM models are
variationally consistent. We then define a smoothing projection operator to modify compatible
strain fields and show a set of properties including the orthogonal property. Next, we establish a
general bound of solution error of the S-FEM models. In the special case, we reveal that when the
smoothing projection operator is an identity, the general error bound is reduced into the standard
form of priori error estimation of a compatible displacement model. Discussions given in this

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1226 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

paper are in a general fashion and hence applicable to all the S-FEM models. In the specific
theoretical discussions and numerical illustrations, we work on two representative S-FEM models
in two-dimensional (2D) problems: ES-FEM using 3-node triangular elements (ES-FEM-T3) and
NS-FEM using 4-node quadrilateral elements (NS-FEM-Q4). The theoretical properties of the
S-FEM models are confirmed by some numerical examples.
The paper is organized as follows: In the next section, finite element formulations are briefed
for later reference. Section 3 presents a brief on the S-FEM models. Section 4 shows a relation
between a smoothed strain field and the Hellinger–Reissner variational principle. In Section 5, we
present some properties of a smoothing projector. The convergence of S-FEM models is studied
in Section 6. Displacement and energy error norms are defined in Section 7. In Section 8, two
numerical examples are conducted and discussed to verify the theoretical properties established.
Section 9 concludes the main remarks.

2. BRIEF ON THE FEM

2.1. Basic equations for elasticity


Consider a 2D linear elastic solid defined in a domain with a Lipschitz continuous boundary .
A body force b acts within the domain. Boundary  is split into two parts, namely u where
displacements u are prescribed (Dirichlet conditions), and t where tractions t are prescribed
(Neumann conditions). Those two parts form the boundary seamlessly  = u ∪t , u ∩t = ∅.
The relations between the displacement field u, the strain field e and the stress field r are
(1) The compatibility relation between the strains and the displacements

∀i, j = 1, 2 : εi j = 12 (* j u i +*i u j ) or e = *u in 
(1)
u i = u i on u

where *i = */*xi are the first partial derivatives corresponding to xi ∈ {x, y} and * denotes
a matrix of differential operators:
⎡ ⎤T
* *
⎢ *x 0
⎢ *y ⎥

*=⎢ ⎥ (2)
⎣ * *⎦
0
*y *x

(2) The constitutive relations for solids can be given as

i j = Di jkl εkl in  or r = De in  (3)

where D is a matrix of elastic constants of stable materials that is bounded uniformly and
a symmetric positive-definite matrix (SPD). This implies that we work with stable solids of
positive definite in this paper.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1227

(3) The equilibrium equations


* j i j +bi = 0 in  (4)
i j n j = ti on t (5)

2.2. Briefing on a H1 space [42–44]


We first define the inner product of any pair of functions in a linear space over a domain  as

(v, w) = vw d (6)


A Lebesgue space L () with the inner product and the associated norm is given by
2


L2 () = (v, w)L2 () = vw d < ∞,


v
L2 () = [(v, v)]1/2 (7)


In a similar manner, a Hilbert space H1 () is defined as


H1 () = {(v, w)H1 = (∇v, ∇w)L2 () +(v, w)L2 () < ∞,
v
H1 = [(v, v)H1 ]1/2 } (8)

where (∇v, ∇w)L2 () =  (∇v)T ∇w d =  (*1 v*1 w +*2 v*2 w) d is the inner product for the
gradient of functions v and w.
Let v = {v1 , v2 } and w = {w1 , w2 } be two 2D vector functions. We often use the notation of
‘semi-norm’ defined as
|v|H1 = [(∇v, ∇v)]1/2 =
∇v
L2 (9)
Let the two spaces of kinematically admissible displacements be
⎧  1/2 ⎫
⎨ 2 ⎬
V = v ∈ (H1 ())2 , v = v on u |
v
H1 =
vi
2H1 (10)
⎩ i=1 ⎭

V0 = {v ∈ V, v = 0 on Cu } (11)

These spaces lead to a bounded energy in a stable solid



εi j (v)Di jkl εkl (v) d < ∞ (12)

From Equation (12), both V and V0 may be equipped with the ‘energy’ inner product and the
induced energy norm defined as

a(v, w) = (D*v, *w) = εi j (v)Di jkl εkl (w) d (13)

 1/2

*v
E = εi j (v)Di jkl εkl (v) d (14)


Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1228 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

2.3. Briefing on a G1 space


A G space of a set of discrete functions was established in the recent work by Liu [41] and
Liu and Zhang [45], especially in the meshfree methods [1]. This space was defined to include
some continuous/discontinuous functions allowing the use of more types of methods/techniques
to produce shape functions for numerical models. This is achieved by defining the energy norm
(or G1 norm) in a special manner based on the smoothing domains. In this work we focus on
the S-FEM models, and hence we consider only continuous functions created based on elements.
We need not invoke all G space theories. However, since H1 ⊂ G1 and using smoothing domains
in the S-FEM models, some inequalities derived in [41] can be readily used. Therefore, we provide
here a necessary briefing on the G1 space. More details on the properties of a G space and detailed
proofs can be found in [46].
The G1 space related to this work can be defined as follows. Consider a problem domain
 discretized into Ne elements with Nn nodes. The problem domain  is then partitioned into
 Ns s
a set of Ns smoothing domains sk associated with nodes or edges such that  = k=1 k and
is ∩sj = ∅, i = j. By this way, these smoothing domains are linearly independent. In addition,
the minimum number of smoothing domains (which is the number of non-constrained nodes) is
also satisfied because the minimum number of smoothing domains is always smaller or equal to
the number of smoothing domains associated with nodes or edges. Then we have the following
definition.
For a given set of minimum number of linearly independent smoothing domains sk , (k =
1, 2, . . . , Ns ), we have
 
(v, w)G1 = (∇v, ∇w)+(v, w), (v, w) < ∞,
G () =
1
(15)
(∇v, ∇v) > 0, ∀v = 0,
v
G1 = [(v, v)G1 ]1/2
 n
where v(x) = N I =1 N I (x)d I ≡ N(x)d, d ∈ R
Nn
and N I (x) is the nodal shape functions for node I ,
and these N I nodal shape functions form the basis of the G1 space.
The smoothed (or approximated) gradient is defined for each smoothing domain as

⎛ ⎞
 ⎜    ⎟
*w *w ⎜ 1 1 ⎟
∇w =
T ⎜
=⎜ s w(x)n x d s w(x)n y d⎟
*x *y ⎟
⎝ Ak k
s A k k
s

   
=ḡ1 , constant in sk =ḡ2 , constant in sk

1
= nT (x)w(x) d ≡ (ḡx (w) ḡ y (w)) (16)
Ask sk

where ḡi (w) denotes the smoothed derivatives of w with respect to xi via the smoothing domains
sk as discussed in Section 3; n = [n x n y ]T is the outward normal vector on the boundary sk .
We often use the notation of ‘semi-norm’ for functions in a G space that is defined as

|v|G1 = [(∇v, ∇v)]1/2 =


∇v
L2 (17)

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1229

Now a new ‘energy’ inner product and an induced energy norm based on the smoothing domain
sk are defined as


Ns
ā(v, w) = (D*v, *w) = ε̄i(k)
j (v)Di jmn ε̄mn (w)Ak
(k) s
(18)
k=1
 1/2

Ns

*v
E = ε̄i(k)
j (v)Di jmn ε̄mn (v)Ak
(k) s
(19)
k=1

where the subscript ‘E’ stands for energy norm of the smoothed models, and ε̄i(k)
j is the smoothed
strain over the smoothing domain sk , which will be defined precisely in Section 3.3.

2.4. Weak statement for the standard FEM models


The standard displacement approach finds a displacement field u ∈ V that satisfies the standard
Galerkin weak form
  
eT (u)De(v) d = bT v d+ t¯T v d ∀v ∈ V0 (20)
  t

or in the classical weak statement of the problem: find u ∈ V such that

(D*u, *v) = f (v) ∀v ∈ V0 (21)


 
a(u,v)

where a(., .), f (.) are the bilinear and linear forms, respectively, defined as
  
(D*u, *v) = eT (u)De(v) d, f (v) = bT v d+ t̄T v d (22)
    t
a(u,v)

It has been proven that Equation (21) has a unique solution for all stable solids. This is ensured
by the classical inequality known as the Poincaré–Friedrichs inequality for all functions in a more
general Sobolev space that contains the Hilbert spaces.

2.5. Finite element formulation


Now let Vh ⊂ V be an approximation space. The statement of the discrete problem becomes finding
a discrete solution uh ∈ Vh that satisfies

(D*uh , *vh ) = f (vh ) ∀vh ∈ V0h (23)


 
a(uh ,vh )

Equation (23) has a unique solution for all stable solids. Once the problem is properly posed, the
issue now is to create proper Sobolev spaces of finite dimension for approximated solutions. These
common models are FEM [42–44] and meshfree models [1]. On FEM settings, the nodal shape
functions are used to form the basis of a proper Sobolev space of finite dimension for displacement
field approximation.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1230 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

Let {N I (x)} be such nodal basis (functions) for the finite dimension space Vh . In a displacement
approach, for instance, the displacement is first assumed using
$ %
Nn N I (x) 0
u (x) =
h
d I ≡ N(x)d (24)
I =1 0 N I (x)

where d I = [u I v I ]T are the unknown displacement associated with node I , N I (x) is the shape
function associated with node I , and d ∈ R Nn is the nodal displacement vector for all the unknowns
of the problem.
Then, the discrete compatible strain field is obtained as

e(uh ) = *uh = B(x)d (25)

where B(x) = *N(x) is the discretized strain–displacement matrix.


By substituting Equations (24) and (25) into Equation (23), we obtain a linear algebraic system
of equations

Kd = f (26)

with the stiffness matrix given by



K= BT DB d (27)


which is an SPD and the load vector by


 
f = NT b d+ NT t d (28)
 t

3. BRIEF ON THE S-FEM MODELS

The S-FEM models in 2D domains (CS-FEM, NS-FEM and ES-FEM) can use general n-sided
polygonal elements [17, 24, 28] including 3-node triangular and 4-node quadrilateral elements
[14, 15, 18, 24, 27, 28]. In this paper, we choose the 3-node triangular and 4-node quadrilateral
elements to analyze in detail.

3.1. Smoothing domain creation of the S-FEM models


In the S-FEM models, a mesh of elements is required, which can be created in the same manner
as in the standard FEM. Consider now a mesh of Ne elements with Nn nodes and Neg edges that
has already been created.
The problem domain  is also then divided into a set of Ns smoothing domains sk such that
 Ns s
 = k=1 k and is ∩sj = ∅, i = j. In theory, such a division can be arbitrary when continuous
shape functions are used. In practice, however, it is usually performed based on the element entities,
such as cells resided inside elements, or nodes, or edges (or faces) of the elements. By this way,
the smoothing domains created are linearly independent and the solutions of the S-FEM models
can be ensured to be stable and convergent.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1231

Figure 5. A NS-FEM-Q4 model using quadrilateral elements. The shaded area is the smoothing domain
s1 associated with node 1 that consists of four sub-domain cells: The symbols (•), (◦) and () denote
the field nodes, the mid-edge point and the intersection point of two bi-medians of element, respectively.
Solid line: element edges, dash lines: segments of the boundary of the smoothing domains.

Specifically, the NS-FEM-Q4 uses a mesh of Ne 4-node quadrilateral elements, and Ns = Nn


smoothing domains are created based on nodes. The smoothing domain sk associated with the
node k is created by connecting sequentially the mid-edge points to the intersection point of two
bi-medians of all 4-node quadrilateral elements around the node k as shown in Figure 5.
Similarly, the ES-FEM-T3 uses a mesh of triangular elements (ES-FEM-T3) and Ns = Neg
smoothing domains are created based on edges. The smoothing domain sk associated with the
edge k is created by connecting two end-edge points to centroids of adjacent elements as shown
in Figure 3.

3.2. Shape functions of the S-FEM models


For 3-node triangular elements, the shape functions used in the S-FEM-T3 are similar to those
used in the standard FEM-T3. They are linear shape functions and are constructed directly based
on physical coordinates of the three nodes of triangular elements. The union of these linear shape
functions of all the elements will create a piecewise linear, continuous nodal shape functions for
the whole problem domain.
For 4-node quadrilateral elements, the shape functions used in the S-FEM-Q4 models are
different from those used in the standard FEM-Q4. In the FEM-Q4, the isoparametric bilinear

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1232 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

shape function is used and mappings are required. On the contrary, in the S-FEM-Q4 models, the
mappings are not required and the shape functions are constructed directly on the physical element
via a linear PIM detailed in [14, 17, 24]. In addition, because only linear interpolation is used in
the S-FEM-Q4 models, there is no need to construct the shape functions explicitly. Only some
discrete shape function values interpolated linearly along the boundary segments of the smoothing
domains are necessary as presented in Section 3.7.

3.3. Smoothed strain and smoothing projector


In the S-FEM models, we do not use the compatible strain fields given in (25) but strains ‘smoothed’
over smoothing domains, and naturally the integration for the stiffness matrix is no longer based on
elements, but on these smoothing domains. For instance, the NS-FEM-Q4 uses smoothing domains
associated with the nodes (Figure 5) whereas the ES-FEM-T3 uses smoothing domains associated
with the edges (Figure 3). The smoothed strain over the smoothing domains is in general defined
as follows:

ē =
(k)
e(x)(k) (x) d (29)
sk

where (k) (x) is the smoothing function and is chosen to be a Heaviside-type function given by

1/Ask , x ∈ sk
 (x) =
(k)
(30)
0, x∈/ sk

where Ask = s d is the area of the smoothing domain sk . Using the smoothing function (30),
k
the smoothed strain ē(k) becomes constant in the smoothing domains sk and has the form

1
ē = s
(k)
*uh d (31)
Ak sk

where uh ∈ Vh is assumed using Equation (24).


Next, to define a smoothing projector succinctly, it is necessary to define the following spaces.
First, the space S containing arbitrary strains and stresses is defined by

S = {e, r ∈ [L2 ()]3 } (32)

and the subspace Sc ⊂ S in which the stresses are related to compatible strains (1) through Hooke’s
law (3) is defined by

Sc = {e, r ∈ S|r = De, e = *u, u ∈ V} ⊂ S (33)

Also, we define the subspace S̄c ⊂ S in which the stresses are related to smoothed strains (31)
through Hooke’s law (3) as
  
1
S̄c = ē, r̄ ∈ S|r̄ = Dē, ē = s *u d, u ∈ V ⊂ S (34)
Ak sk

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1233

In the mathematics terminology, smoothing the compatible strain e(x) in the smoothing domain
sk to give the smoothed strain ē(k) (31) is similar to a projection Ph from e(x) ∈ Sc into ē(k) ∈ S̄c .
We hence have the definition of a projector Ph as

Ph : e(x) → ē(k) , e(x) ∈ Sc ,


ē(k) ∈ S̄c such as

1 (35)
ē(k) = *vh = Ph *vh = s *vh d ∀vh ∈ V0h
Ak sk

The projector (35) can be called as a smoothing or constant projector. It is hence easy to have
two following remarks.

Remark 1
Projection defined in (35) is from a higher-dimension space Sc onto a lower-dimension space S̄c .

Remark 2
The projector defined in (35) is an idempotent: Ph2 = Ph .

Now, by performing the integration by part for Equation (31), the smoothed strain ē(k) can be
calculated using

1
ē(k) = *uh = Ph *uh = s uh (x)n(k) (x) d (36)
Ak sk

where sk is the boundary of the domain sk as shown in Figure 5, and n(k) (x) is the outward
normal vector matrix on the boundary sk and has the form
⎡ (k) ⎤
nx 0
⎢ ⎥
n(k) (x) = ⎢
⎣ 0 ny ⎦
(k) ⎥
(37)
n (k)
y n (k)
x

3.4. Weakened weak statement for the S-FEM models


The discrete solution of the problem associated with an S-FEM model is to find ūh ∈ Vh such that
the following ‘smoothed’ Galerkin weak form is satisfied
(DP *ūh , P *vh ) = f (vh ) ∀vh ∈ V0h (38)
 h  h
ā(ūh ,vh )

where ā(., .) is a ‘smoothed’ bilinear form of smoothed derivatives of functions and has the
following summation form:
Ns 

ā(ūh , vh ) = (DPh *ūh , Ph *vh ) = ē(k)T (ūh )Dē(k) (vh ) d (39)
k=1 sk

where Ph is the smoothing projector defined in Equation (35).


It has been shown that ā(ūh , vh ) is also an ‘energy’ bilinear form but is evaluated using smoothed
strains. It satisfies all the necessary conditions of a bilinear form, and hence is a ‘legal’ bilinear

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1234 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

form satisfying all the inequalities and properties of a standard bilinear form [1]. This is important
for our later functional analysis in Section 6. Equation (38) is known as a weakened weak (W2 )
statement, because in evaluating the smoothed bilinear form, we use only the assumed displacements
themselves, and no derivatives of the assumed displacements are needed as in the usual weak
form [41]. The consistent requirement on the assumed displacement functions is further weakened
compared with the standard weak form (21).

Remark 3 (Existence and uniqueness)


The variational problem of linear elasticity given in Equation (38) for stable solids has a unique
solution if the smoothing domains used are linearly independent and are not smaller than a minimum
number of smoothing domains.

The proof of Remark 3 can be found in [41].


Note that from Remark 3, we have ā(vh , vh ) = (DPh *vh , Ph *vh ) to be coercive and because
a(vh , vh ) = (D*vh , *vh ) is also proven coercive [42], one obtains

∃G H > 0 :
*vh
2E = (DPh *vh , Ph *vh )G H (D*vh , *vh ) = G H
*vh
2E ∀vh ∈ V0h (40)

where G H is a general positive constant that is independent of vh ∈ V0h .

3.5. Smoothed strain–displacement matrix of the S-FEM models


Substituting Equation (24) into Equation (36), the smoothed strain for the smoothing domain
sk can be formulated in the following matrix form:

ē(k) = B̄ I (xk )d̄ I (41)
I ∈(k)
n

where d̄ I is the unknown displacement of S-FEM models at node I ; (k) n is the set of nodes
supporting to the smoothing domain sk . For example for the ES-FEM-T3 as shown in Figure 3, the
set of nodes supporting smoothing domain associated with the inner edge k is (k) n = {A, B, C, D},
and with the boundary edge m is (k) n = {E, G, H } and B̄ I (x k ) is termed as the smoothed strain–
displacement matrix on the smoothing domain sk and is calculated numerically by
⎡ ⎤
b̄ I x (xk ) 0
⎢ ⎥
B̄ I (xk ) = ⎢
⎣ 0 b̄ I y (xk )⎥ ⎦ (42)
b̄ I y (xk ) b̄ I x (xk )

where

1
b̄ I h (xk ) = s N I (x)n (k)
h (x) d (h = x, y) (43)
Ak sk

From the above equation, it is observed that we need only the values of N I on the boundary
of the smoothing domain sk , which can be obtained by a simple linear interpolation technique
given in [14, 17, 24]. In such cases, one Gaussian point is sufficient for line integration along each

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1235

segment of the boundary skb ∈ sk . Equation (43) can be further simplified to its algebraic form

n eg
(k) (k)
b̄ I h (xk ) = N I (xGP
b )n bh lb (h = x, y) (44)
b=1

where n eg is the total number of the segments skb ∈ sk ; xGP


b is the midpoint (Gaussian point) of
(k) (k)
the boundary segment of kb ; lb and n bh are the length and the outward unit normal of skb ,
s

respectively.

3.6. Smoothed stiffness matrix of the S-FEM models


The smoothed stiffness matrix K̄ of the system is then assembled by a similar process as in the FEM

Ns
K̄ = K̄(k) (45)
k=1

where K̄(k) is the stiffness matrix associated with smoothing domain sk and is calculated by
  T  
1
K̄ =
(k)
B̄ DB̄ d = s
T (k)
n N d D (k)
n N d (46)
sk Ak sk sk

The force vector obtained in the S-FEM models is the same as that in the FEM. Equation (46)
shows again that in the S-FEM models, field gradients are computed directly using only the shape
function values at the Gauss points along the segments of the boundary and no derivatives of the
shape functions are needed. Therefore, in the S-FEM models, we do not need to construct the
shape function N I explicitly, and hence no mapping is required. This is one of the key differences
between the S-FEM models and the standard FEM.

3.7. Evaluation of the shape function values in the S-FEM models


The most essential requirement in evaluating the shape function values in the S-FEM models is to
ensure the compatibility of the shape functions on all the interfaces of these smoothing domains.
The simple PIM [24] satisfies this requirement and is easy to use, and hence is used in the S-FEM
models. Using PIM, the shape function values at each Gauss point are evaluated by simple linear
interpolation (or averaging) using two endpoints of the segment containing the Gauss point. For
example for the NS-FEM-Q4 as shown in Figure 5, the shape function values at the Gauss point
#a on segment A–B are the simple average of those at two end-segment points: points #A and #B;
the shape function values at the Gauss point #b on segment A-1 are the simple average of those
at two end-segment points: points #A and #1. Therefore, in order to facilitate the evaluation of
the shape function values at the Gauss points on the smoothing domain boundaries in the S-FEM
models, we need first to evaluate the shape function values at the endpoints of segments such as
#A, #B, . . . ,#H, #I and #1.
Similarly, the evaluation of the shape function values at the endpoints of segments is also
conducted by PIM. For example, for the NS-FEM-Q4 as shown in Figure 5, the shape function
values at the mid-edge point #A on the edge 1–2 are the average of those at two field nodes on
the edge: nodes #1 and #2; or the shape function values at central point #B (the intersection of
two bi-medians A-K and C-J) are the average of those at four field nodes of the element: nodes
#1, #2, #3 and #4.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1236 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

It should be mentioned that the introduction of the central points and the mid-edge points aims
to facilitate the evaluation of the shape function at the Gauss points along the segments of the
smoothing domains. No extra degrees of freedom (DOFs) associated with these points are required.
In other words, these points carry no additional independent field variable. Therefore, the total
DOFs of an S-FEM model will be exactly the same as the standard FEM using the same set of
nodes.
Note that in the standard FEM, which in general uses the isoparametric elements and mapping,
the shape functions obtained using the above PIM in the S-FEM models are computed directly on
the physical elements. Hence, the problem domain can be discretized in more flexible ways with
any shape of elements bounded by straight edges, including severely distorted 4-node quadrilateral
elements [24] and general n-sided polygonal elements [17].

4. RELATION BETWEEN SMOOTHED STRAIN AND THE HELLINGER–REISSNER


VARIATIONAL PRINCIPLE

Using the Hellinger–Reissner variational principle in the S-FEM models in which the stress r and
the displacement u are considered as independent fields, one has
   
1 1
HR (r, u) = − i j Di−1
jkl kl +i j (*i u j +* j u i )−bi u i d− ti u i d (47)
 2 2 t

In the S-FEM models, the stresses (or strains) are constants in sk . Thus, Equation (47) can be
expressed in a summation over all the smoothing domains as follows:

   

Ns 1 1
HR (r, u) = − i j Di−1
jkl kl +i j (*i u j +* j u i )−bi u i d− ti u i d (48)
k=1 sk 2 2 skt

where skt is the portion of the element boundary over which the prescribed surface tractions t̄ are
applied. The second term in the right-hand side of Equation (48) becomes
  
1 1
i j (*i u j +* j u i ) d = i j (n i u j +n j u i ) d− u i * j i j d (49)
sk 2 sk 2 sk

Also because the stress is constant in sk , the last term in (49) vanishes. Substituting Equation (49)
into Equation (48), one has

    

Ns 1 1
HR (r, u) = − i j Di−1
jkl kl Ak +i j
s
(n i u j +n j u i ) d− bi u i d− ti u i d
k=1 2 2 sk sk stk

(50)

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1237

The variation with respect to stress field leads to



1
Di−1  A
jkl kl k
s
= (n i u j +n j u i ) d (51)
2 sk

This implies that there exists the strain field ε̄i j = Di−1
jkl kl such that

1
ε̄i j = s (n i u j +n j u i ) d (52)
2Ak sk

This shows that the smoothed strain in Equation (36) used in the S-FEM models can be derived
from the Hellinger–Reissner principle. Therefore, we obtain the variational principle of two fields
based on the smoothed strain ē and the displacement u as
   
Ns 1 (k) (k) s
HR (ē, u) = ε̄i j Di jkl ε̄kl Ak − bi u i d− ti u i d (53)
k=1 2 sk skt

whose form is identical to the often used mixed approach. The difference is that operations are
based on the smoothing domains instead of elements.
From this relation between the smoothed strain and the Hellinger–Reissner principle (note that
the Hellinger–Reissner variational principle is a particular case of the FdV variational principle
[13]), the S-FEM models in general can be considered as a quasi-equilibrium model with two
following properties: (1) equilibrium is satisfied inside the smoothing domains but is not satisfied
on the boundary of smoothing domains; (2) on the contrary, the compatibility is not satisfied inside
the smoothing domains but is satisfied on the boundary of smoothing domains. The S-FEM models
are not as ‘fully’ compatible as the standard FEM. Therefore, the S-FEM models are ‘softer’ than
the standard FEM, and can give better results with many interesting properties. For example, the
NS-FEM-Q4 model can give an upper bound to the exact solution in the energy norm for forced
driven elasticity problems [24] and gains also super-convergence in energy norm. The NS-FEM-Q4
is even free from volume-locking, which is another typical behavior of a fully equilibrium model
[36–40, 47].

5. PROPERTIES OF THE SMOOTHING PROJECTOR

We now study the properties of the smoothing projector defined in Equation (35) that is used in
the formulation of the S-FEM models. We start again with the more general Hellinger–Reissner
variational principle: the exact solution (r, u) makes the following functional:

HR (r, u) = − 12 (r, D−1 r)+(r, *u)− f (u) (54)

maximum for all stress fields r ∈ S and minimum for all displacement fields u ∈ V. In the weak
form, the exact solution (r, u) has to satisfy

(s, *u) = (s, D−1 r) ∀ ∈ S


(55)
(r, *v) = f (v) ∀v ∈ V0

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1238 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

Let Vh ⊂ V and Sh ⊂ S be the respective finite element spaces. The statement of the discrete
problem is to find a pair (rh , uh ) ∈ Sh ×Vh such that

(sh , *uh ) = (sh , D−1 rh ) ∀sh ∈ Sh (56a)


(rh , *vh ) = f (vh ) ∀vh ∈ V0h (56b)

Equation (56a) can be rewritten as

(sh , (D−1 rh −*uh )) = 0 ∀sh ∈ Sh (57)

We now observe that if D−1 rh is an orthogonal projection of *uh onto Sh , Equation (56a) is
satisfied automatically. Therefore, if we can find an orthogonal operator Ph such that

(rh , *vh ) = (rh , Ph *vh ) ∀vh ∈ V0h (58)

we shall always have

D−1 rh = Ph *uh or rh = DPh *uh (59)

which leads to

(rh , *vh ) = (DPh *uh , Ph *vh ) ∀vh ∈ V0h (60)

Using Equations (56b) and (60), one has

(DPh *uh , Ph *vh ) = f (vh ) ∀vh ∈ V0h (61)

which is a displacement-like formulation with only displacement as the unknown field (no
smoothing), and the formulation is variationally consistent. In the very special case of an identity
operator Ph , which means that the stress field is obtained using the compatibility equation, the
compatible displacement model is recovered.

Remark 4 (Orthogonal projector)


We claim now that the projector defined in (35) is an orthogonal projector.

Proof

 

Ns 
Ns
(r , *v ) = (DPh *u , *v ) =
h h h h
(DPh *u ) *v d = (DPh *u )
h T h h T
*vh d
  k=1 sk   k=1 sk
rh constant in sk


Ns 1 
Ns
= Ask (DPh *uh )T *vh d = (DPh *uh )T Ph *vh Ask
k=1 Ask sk k=1

= (DPh *uh , Ph *vh ) = (rh , Ph *vh ) (62)

which satisfies Equation (58). 

It is seen again that the W2 statement equation (38) is variationally consistent.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1239

Remark 5 (Distributive projector in an inner product)


We claim now that the projector defined in (35) is a distributive projector in an inner product,
meaning that
(Ph *wh , *vh ) = (Ph *wh , Ph *vh ) ∀wh , vh ∈ V0h (63a)

(DPh *wh , *vh ) = (DPh *wh , Ph *vh ) ∀wh , vh ∈ V0h (63b)

Proof
For any wh , vh ∈ V0h , we have
 

Ns 
Ns
(Ph *wh , *vh ) = (Ph *wh )T *vh d = (Ph *wh )T *vh d
k=1 sk k=1 sk


Ns 1 
Ns
= Ask (Ph *wh )T *vh d = (Ph *wh )T Ph *vh Ask
k=1 Ask sk k=1

= (Ph *wh , Ph *vh ) (64)


which is Equation (63a). By the same token, Equation (63b) can be also proven. 
Remark 6 (Commutative projector in an inner product)
The projector defined in (35) is a commutative projector in an inner product, meaning that
(Ph *wh , *vh ) = (*wh , Ph *vh ) ∀wh , vh ∈ V0h (65a)

(DPh *uh , *vh ) = (D*uh , Ph *vh ) (65b)

Proof
For any wh , vh ∈ V0h , we have
Ns  
Ns 
(Ph *w , *v ) =
h h
(Ph *w ) *v d = (Ph *w )
h T h h T
*vh d
k=1 sk k=1 sk

   T

Ns 1 
Ns 1
= Ask (Ph *wh )T s *vh d = Ask *wh d Ph *vh
k=1 Ak sk k=1 Ask sk
 T

Ns
= *w d
h
Ph *vh = (*wh , Ph *vh ) (66)
k=1 sk

which is Equation (65a). By the same token, Equation (65b) is also proven. 
Remark 7 (Softening effect)
The work of the applied forces of the solution of S-FEM models is always greater than that of the
solution of the standard fully compatible displacement approach:
& &
f (uh ) f (ūh ) (67)

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1240 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

Proof
For simplicity, we consider only homogenous essential boundary conditions, and hence Vh = V0h .
Let uh be the solution of the classical (fully compatible) displacement model. It verifies that

(D*uh , *vh ) = f (vh ) ∀vh ∈ Vh (68)

Setting vh = uh and using the definition of energy norm (14), one has


*uh
2E = f (uh ) (69)

Now, from Equation (38) and (63), and using the Cauchy–Schwarz inequality for an inner
product (which can be proven in the same way as in [41]), we arrive at the following inequality:
& &
f (uh ) = (DPh *ūh , *uh )
Ph *ūh
E
*uh
E = f (ūh ) f (uh ) (70)

which leads to Equation (67). 

The softening effect is an essential reason for the fact that an S-FEM model can perform better
in energy norm than the standard FEM model using the same mesh.

6. CONVERGENCE OF THE S-FEM MODELS

In this section, we establish a general error bound of the S-FEM models and explain further the
rate of the convergence of the S-FEM models.

6.1. Relation between the S-FEM solution and the exact solution
The weak statement for the exact solution for stable solids is given in Equation (21), and we have
the following proven properties (continuity and stability) of the bilinear form:

∃H > 0 : |(D*w, *v)|H


*w
E
*v
E
∀w, v ∈ V0 (71)
∃H > 0 : (D*v, *v)H
*v
2E

where H relates to the coercivity constant and H is the continuity constant for weak formula-
tion based on the Hilbert space theory. In fact, in our case, H = 1 due to the Cauchy–Schwarz
inequality.
Now, by manipulating (DPh *u, *v):

(DPh *u, *v) = (D*u, *v)+(DPh *u−D*u, *v) ∀v ∈ V0


= f (v)+(D(Ph *u−*u), *v) (by Equation (21)) (72)

and because vh ∈ V0h ⊂ V0 , we should have

(DPh *u, *vh ) = f (vh )+(D(Ph *u−*u), *vh ) ∀vh ∈ V0h (73)

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1241

Using Equation (38) and Equation (63b), we have

(DPh *ūh , *vh ) = f (vh ) ∀vh ∈ V0h (74)

Substituting Equation (74) into (73), we obtain

(D(Ph *ūh − Ph *u), *vh ) = (D(*u− Ph *u), *vh ) ∀vh ∈ V0h (75)

which gives the relationship between the S-FEM solution ūh and the exact solution u.

6.2. A general error bound on the S-FEM solutions


We now derive the general error bound and point out some remarks on convergence rate for the
S-FEM models. We first state

Theorem 1 (General solution bound for the S-FEM models)


When a minimum number of linearly independent smoothing domains are used in the S-FEM
models, the error in energy norm of the S-FEM solution ūh satisfies


*u− Ph *ūh
E  CP
*u− Ph *u
E + CD inf
*u−*vh
E (76)
    ∀vh ∈V0h
solution error error from the projection of exact solution  
error from the assumed displacements

where CP and CD are general constants.

Proof
We now need to show that the error comes from two terms: one from the projection of the exact
solution and one from the assumed displacements that deviate from the exact displacement field.
We start with the following functional analysis: ∀vh ∈ V0h , we have

(DPh (*ūh −*vh ), Ph (*ūh −*vh )) = (DPh (*ūh −*vh ), *ūh −*vh ) (distributive)
= (D(Ph *ūh − Ph *u), *ūh −*vh )
+(D(Ph *u− Ph *vh ), *ūh −*vh ) (bilinear) (77)

The first term in the third line in Equation (77) is now evaluated as

(D(Ph *ūh − Ph *u), *ūh −*vh ) = (D(*u− Ph *u), *ūh −*vh ) (by Equation (75))
 H
*u− Ph *u
E
*ūh −*vh
E (by Equation (71a)) (78)

Using Equation (40), we have


Ph (*ūh −*vh )
2E = (DPh (*ūh −*vh ), Ph (*ūh −*vh ))G H
*ūh −*vh
2E (79)

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1242 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

Equation (78) now becomes


(D(Ph *ūh − Ph *u), *ūh −*vh ) √ H
*u− Ph *u
E
Ph (*ūh −*vh )
E ∀vh ∈ V0h (80)
G H

The second term in the third line in Equation (77) is

(D(Ph *u− Ph *vh ), *ūh −*vh )


= (D(*u−*vh ), Ph (*ūh −*vh )) (Communicative)
H
*u−*vh
E
Ph (*ūh −*vh )
E ∀vh ∈ V0h (by Equation (71a)) (81)

From Equation (71b), one obtains

H
Ph (*ūh −*vh )
2E  (DPh (*ūh −*vh ), Ph (*ūh −*vh ))

 √ H
*u− Ph *u
E
Ph (*ūh −*vh )
E
G H

+H
*u−*vh
E
Ph (*ūh −*vh )
E , ∀vh ∈ V0h (82)

which leads to
H 

Ph (*ūh −*vh )
E  √
*u− Ph *u
E + H
*u−*vh
E , ∀vh ∈ V0h (83)
H  G H H

On the other hand, we have


*u− Ph *ūh
E 
*u− Ph *u
E +
Ph *u− Ph *ūh
E

*u− Ph *u
E +
Ph (*u−*vh )
E +
Ph (*vh −*ūh )
E

*u− Ph *u
E +
*u−*vh
E +
Ph (*vh −*ūh )
E
   
H H
 1+ √
*u− Ph *u
E + 1+
*u−*vh
E , ∀vh ∈ V0h (84)
H  G H H

Finally, we have
   
 H

*u− Ph *ū
E  1+ √H
h

*u− Ph *u
E + 1+ inf
*u−*vh
E
H  G H H ∀vh ∈V0h
   
CP CD
   
error from the projection of exact solution error from the assumed displacements

= CP
*u− Ph *u
E +CD inf
*u−*vh
E (85)
∀vh ∈V0h

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1243

which is Equation (76). It is clear now that the error comes from two terms: one from the projection
of the exact solution onto the smoothed domains, and one from the assumed displacement field
that deviates from the exact solution. 

It is evident that when Ph is an identity operator, the first term of the right-hand side in (85)
vanishes, and hence the standard priori error estimation of the compatible displacement model is
recovered.
Remark 8 (Applicability of Theorem 1: a general bound)
In establishing Theorem 1 we do not use any condition on how the smoothing domains are
constructed, as long as a minimum number of independent smoothing domains are used. This
implies that Theorem 1 is applicable to all S-FEM models in which the assumed displacement
fields are in a proper Sobolev space based on the weak formulation.
Remark 9 (h-dependence of solution error in energy norm)
In our past work on W 2 formulations [41], it was often found that the rate of convergence for the
solution error in energy norm is around 1.5. However, we could not give a proof on these findings
yet. In the analysis of interpolation error in G1 norms, we found an ideal rate of 1.5 [41]. Up to
now, the link between the solution error and the interpolation error is yet to be quantified. Even for
the 1.5 rate of interpolation error, it is with the conditions of (1) symmetric smoothing domain (first
moment of the smoothing function vanishes); (2) constant second derivative of the exact solution
in the two elements sharing any smoothing domain. The first symmetric condition is essential for
the integral representation to produce exactly the first gradient of an approximate function [1].
When this condition is satisfied, the smoothing operation will reproduce the first derivative exactly,
leading to second-order accuracy. The only error will be on the boundary where the symmetry
condition cannot be satisfied, because of the cut-off of the geometry on boundary. The second
condition of constant second derivative of the target function seems to be very strong, but it can
be rather easy to be quite closely satisfied, because the second derivative of the exact solution is
locally constant in the two halves of the elements sharing a smoothing domain. When the mesh is
refined, we can expect the second derivative to be approximately constant locally. Therefore, the
rate of 1.5 given in [41] can be expected in actual modes when the mesh is reasonably refined.
In practical applications, on the other hand, the first condition is very difficult to meet for 2D and
3D domains, simply because it is rare to have uniform division of elements for practical problems
of complicated geometry, and hence difficult to create symmetric smoothing domains. However, we
can in fact expect the smoothing domains to be approximately ‘symmetric’ locally. In such cases,
we can still expect a convergence rate close to 1.5. This has been confirmed in many numerical
examples presented in all S-FEM methods [12, 24–28], where numerical rates of about 1.45 were
often found, which will be shown also in the following numerical examples. We were excited
about the high numerical convergence rates observed, but could not give a good explanation yet.
We only knew that the rate should be in between 1.0 and 2.0 [24–28] and far higher than the
standard FEM models.

Remark 10 (Convergence rates in energy error norm for different models)


We now have the following understanding:
1. The rate of convergence for a pure equilibrium model is 2.0: a well-known fact (see e.g. [47]).
2. The rate of convergence for a standard weak form, such as the standard FEM model, is 1.0:
a well-known fact [42–44].

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1244 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

3. An ideal rate of convergence for the S-FEM models can be 1.5 (not theoretically proven yet),
which is between the pure displacement model and the pure equilibrium model.
4. NS-FEM models can be viewed as a quasi-equilibrium model, which has the feature properties
as discussed in Section 4.

The S-FEM models can be classified as a W2 model [41] that is more general and applicable
also for meshfree methods.

7. DISPLACEMENT AND ENERGY ERROR NORMS

In next section, some numerical examples of two representative S-FEM models, NS-FEM-Q4 and
ES-FEM-T3, are performed to verify the theoretical properties of the S-FEM models established.
The results of NS-FEM-Q4 and ES-FEM-T3 will be compared with those of the standard FEM using
pure compatible quadrilateral elements (FEM-Q4) and pure equilibrium quadrilateral elements
(FEM-EQ4) [37]. For quantitative study of the error and convergence rate of these numerical
methods, two types of error norms are used: displacement error norm and energy error norm.

7.1. Displacement error norm


The displacement error norm is based on Ne elements ie and is defined by
 1/2  N  1/2
e

u−u
L2 () =
h h T h
(u−u ) (u−u ) d = h T h
(u−u ) (u−u ) d (86)
 i=1 ie

where u is the exact or analytical solution for the displacement and uh is the numerical solution
for the displacement of any numerical methods.

7.2. Energy error norms


The energy error norm is more involved and hence has to be defined in detail for different cases.
For the FEM-Q4 and FEM-EQ4, the energy error norm is also evaluated based on Ne elements
ie and is defined by
  1/2

Ne

*u−*uh
E = (r−rh )T D−1 (r−rh ) d (87)
i=1 ie

where r is the exact or analytical solution for the stresses and rh is the numerical solution for the
stresses.
For the NS-FEM-Q4 and the ES-FEM-T3, the energy error norm is evaluated based on Ns
smoothing domains sk and is defined by [4, 41]
 1/2

Ns

*u− Ph *ūh
E = (r− r̄(k) )T D−1 (r− r̄(k) )Ask (88)
k=1

where r̄(k) = Dē(k) is derived from smoothed strain solutions ē(k) on smoothing domains sk .

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1245

In addition, to conduct ‘fairest’ possible comparisons among the NS-FEM-Q4, ES-FEM-T3 and
FEM-Q4, we need to define an energy error norm that can be applied for all numerical models.
Herein, we use a recovery energy error norm based on Ne elements ie . In this case, the three-
mentioned numerical methods are named correspondingly: NS-FEM-Q4-Re, ES-FEM-T3-Re and
FEM-Q4-Re. The recovery energy error norm is defined by
  1/2

Ne
T −1

*u−*û
E = (r− r̂) D (r− r̂) d (89)
i=1 ie

where r̂ is the recovery stress for each element in the mesh and is defined by
e

nn
r̂ = N I (x)r(x I ) (90)
I =1

where n en is the number of the nodes of the element (n en = 3 for triangular elements and n en = 4
for the quadrilateral elements); N j (x) is the shape functions of the elements. The shape function
is the same as that used in the standard FEM (linear for the triangular elements and bilinear
isoparametric for the quadrilateral elements); r(x I ) is the numerical stress at the nodes x j of the
element. For the NS-FEM-Q4-Re, r(x I ) is computed directly from the smoothed strain ē(x I ) at
node x I by r(x I ) = Dē(x I ), whereas for the ES-FEM-T3-Re, r(x I ) is computed by (area-weighted)
averaging of stresses of smoothing domains surrounding node x j [25, 27]. For the FEM-Q4-Re,
the stress r(x I ) at the node x j is computed by the following three-step procedure [48]:
(1) Evaluate the stresses at the Gauss points in the element.
(2) Extrapolate the stresses at the Gauss points to the nodes of the element.
(3) Average the stresses computed for the same field node from the adjacent elements.
Note that in order to evaluate the integrals in Equations (86), (87) and (89) accurately, the
mapping procedure using the Gauss integration is performed on each element and then a summation
on all elements is performed. In each element, a proper number of Gauss points depending on the
order of the integrand are used.

7.3. Characteristic length


To evaluate the convergence rates of the displacement and the energy error norms, it is necessary
to define the ‘characteristic length’ of the sides of the elements. In this paper, because the elements
used in a mesh are different in dimensions and of irregular shape, the average length h of the sides
of the elements is considered to be the characteristic length. For the quadrilateral elements, h is
evaluated by
'
A
h= (91)
Ne

where A is the area of the whole problem domain. For the triangular elements, h is evaluated by
'
2A
h= (92)
Ne

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1246 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

8. NUMERICAL RESULTS

8.1. Cantilever loaded at the end: a plane strain problem


A cantilever with length L and height D is studied as a benchmark here. It is subjected to a
parabolic traction at the free end as shown in Figure 6. The plane strain condition is assumed
and the related parameters are taken as Young’s modulus E = 3.0×107 kPa, D = 4 m, L = 8 m and
P = 250 N. The analytical solution is available and can be found in a textbook by Timoshenko and
Goodier [49].
(  )
Py D2
ux = (6L −3x)x +(2+ ¯ ) y 2 −
6 Ē I 4
( ) (93)
P D2 x
uy = − 3¯ y (L − x)+(4+5¯)
2
+(3L − x)x 2
6 Ē I 4
where the moment of inertia I for a beam with rectangular cross-section and unit thickness is
given by I = D 3 /12 and
 
E,  for plane stress
Ē = ¯ = (94)
E/(1−2 ), /(1−) for plane strain
The stresses corresponding to the displacements in Equation (93) are
 
P(L − x)y P D2
x x (x, y) = ,  yy (x, y) = 0, x y (x, y) = − −y 2
(95)
I 2I 4
The domain of the cantilever is discretized into two types of meshes using two different elements:
3-node triangular and 4-node quadrilateral elements as shown in Figure 7. Herein, the DOFs of the
FEM-Q4, NS-FEM-Q4 and ES-FEM-T3 elements are located at the corner nodes of quadrilateral
or triangular elements whereas those of the FEM-EQ4 element are at the middle points along the
element edges.
Under plane strain conditions and the Poisson ratio  = 0.3, the convergence curves of the strain
energy solutions of the NS-FEM-Q4 and the ES-FEM-T3 in comparison with FEM-Q4 and FEM-
EQ4 are plotted in Figure 8. It can be seen that the NS-FEM-Q4 gives upper-bound solutions in the
strain energy that is similar to the FEM-EQ4 (pure equilibrium element), whereas the FEM-Q4 and
the ES-FEM-T3 produce the lower-bound solutions. In addition, the solutions of the ES-FEM-T3
are more accurate than those of the FEM-Q4.

x
D

Figure 6. Cantilever loaded at the end.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1247

0 1 2 3 4 5 6 7 8
(a)
2

0 1 2 3 4 5 6 7 8
(b)

Figure 7. Domain discretization of the cantilever using: (a) 3-node triangular


elements and (b) 4-node quadrilateral elements.

0.0395
Exact

0.039

0.0385

0.038
Strain energy

0.0375

0.037

0.0365

0.036

0.0355

0.035
0 50 100 150 200 250 300 350 400 450 500
Number of elements

Figure 8. Convergence of the strain energy solution of the NS-FEM-Q4 and the ES-FEM-T3 in comparison
with FEM-Q4 and FEM-EQ4 for the cantilever problem.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1248 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

log10(displacement error norm)

0
log h
10

Figure 9. Displacement error norms for the cantilever problem using the
NS-FEM-Q4, ES-FEM-T3 and FEM-Q4.

Table I. Error in energy norm obtained using different methods for the cantilever
problem using the same set of meshes.
Mesh 16×8 Mesh 32×16 Mesh 64×32 Mesh 128×64
h(m) 0.5 0.25 0.125 0.0625
FEM-Q4 2.7071e−02 1.3584e−02 6.7983e−03 3.3999e−03
NS-FEM-Q4 9.3691e−03 3.3287e−03 1.1815e−03 4.1832e−04
ES-FEM-T3 1.4488e−02 5.1613e−02 1.8288e−03 6.4778e−04
FEM-Q4-Re 1.4213e−02 5.1310e−03 1.8282e−03 6.4838e−04
NS-FEM-Q4-Re 6.6741e−03 2.5287e−01 9.2881e−04 3.3483e−04
ES-FEM-T3-Re 1.5317e−02 5.3954e−03 1.9078e−03 6.7509e−04
FEM-EQ4 3.9969e−03 9.9338e−04 2.4794e−04 6.1960e−05

Figure 9 plots the convergence rate in the displacement error norm of three methods: ES-FEM-T3,
NS-FEM-Q4 and FEM-Q4. It is seen that the ES-FEM-T3 stands out clearly. When the fine mesh
(h = 0.125) is used, the error of the ES-FEM-T3 is about 13 of the FEM-Q4 whereas the error of
the FEM-Q4 is about 13 of the NS-FEM-Q4. In terms of convergence rate, the super-convergence is
also observed for the ES-FEM-T3 with a rate of 2.63 that is even much higher than the theoretical
value of 2.0 for linear displacement models based on the weak formulation. Note that we did not
mention the results of FEM-EQ4 in this comparison, because this requires a recovery process of
the nodal displacement solution from the average displacements of the equilibrium model that is
slightly involved.
The energy error norms of the methods obtained by the corresponding equations (87), (88)
and (89) are listed in Table I and are plotted in Figure 10. It is seen that the error of the

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1249

log (energy error norm)


10

log h
10

Figure 10. Energy error norm for the S-FEM models in comparison with those
of other methods for the cantilever problem.

FEM-EQ4 is smallest due to the use of a direct approximation of the stress field in the equilibrium
approach. Without using the recovery energy error norm (89), the errors of the NS-FEM-Q4
and the ES-FEM-T3 are smaller than those of FEM-Q4. By using the recovery energy error
norm (89), the error of the NS-FEM-Q4-Re is still smaller than that of the FEM-Q4-Re, whereas
the error of the ES-FEM-T3-Re is slightly equivalent to that of the FEM-Q4-Re. In terms of
convergence rate, it is seen that the convergence rate of all S-FEM models (from 1.45 to 1.5) is
higher than that of the FEM-Q4 (0.998) and lower than that of FEM-EQ4 (2.0). The results of the
FEM-Q4 and the FEM-EQ4 are quite consistent with the theoretical convergence rates of the fully
compatible model and the fully equilibrium model, respectively, as mentioned in Remark 10, and
the convergence rates of the S-FEM models are very close to the possible theoretical rate of 1.5
given in [41].
Figure 11 plots the energy error norm against the Poisson ratio changing from 0.4 to 0.4999999
by using 512 four-node quadrilateral elements (for FEM-Q4, FEM-EQ4 and NS-FEM-Q4) and
1024 three-node triangular elements (for ES-FEM-T3). The results show that both FEM-EQ4 and
NS-FEM-Q4 are free of volumetric locking, whereas the FEM-Q4 is subjected to the volumetric
locking resulting in a drastic accuracy loss in the numerical solutions. As observed, although
the ES-FEM-T3 can produce much better solutions than the FEM-Q4, it is also slightly sensi-
tive to locking. More details on remedy locking behavior of the ES-FEM-T3 can be found
in [27].
In general for this problem, if we only consider the models using the displacements at nodes
as the main variables (FEM-Q4, NS-FEM-Q4 and ES-FEM-T3), the NS-FEM-Q4 gives the best
results in the energy error norm, and the ES-FEM-T3 gives the best results in the displacement
error norm. However, if we consider both the accuracy (in both displacement and stress) and the
efficiency in mesh discretization, the ES-FEM-T3 is the best choice. In addition, the NS-FEM-Q4

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1250 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

10

Energy error norm


6

0.4 0.49 0.499 0.4999 0.49999 0.499999 0.4999999


Poisson’s ratio

Figure 11. Energy error norm for the ES-FEM-T3 and the NS-FEM-Q4 in comparison with those of the
FEM-Q4 and the FEM-EQ4 for the cantilever problem with different Poisson ratios.

y y

b
r
a x
o
p
p
o x

Figure 12. A hollow cylinder subjected to an inner pressure and its quarter model with symmetric conditions
imposed on the left and the bottom edges.

also possesses three interesting properties similar to those of an equilibrium FEM model: (1) the
upper-bound property in strain energy; (2) ultra-accuracy and super-convergence of stress solutions;
(3) natural immunization from the volumetric locking.

8.2. A hollow cylinder subjected to inner pressure: a plane stress problem


Figure 12 shows a hollow cylinder, with internal radius a = 1 m, external radius b = 5 m, subjected to
an internal pressure p = 3×104 kN/m2 . Because of the axi-symmetric characteristic of the problem,
only the upper right quadrant of the cylinder is modeled. Figure 13 gives the discretizations
of the domain using 3-node triangular and 4-node quadrilateral elements. Plane stress condition

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1251

5 5

4.5 4.5

4 4

3.5 3.5

3 3

2.5 2.5

2 2

1.5 1.5

1 1

0.5 0.5

0 0
0 1 2 3 4 5 0 1 2 3 4 5
(a) (b)

Figure 13. Domain discretization of the hollow cylinder using: (a) 3-node triangular elements
and (b) 4-node quadrilateral elements.

is considered and Young’s modulus E = 3×107 kN/m2 , the Poisson ratio  = 0.25. Symmetric
conditions are imposed on the left and bottom edges, and the outer boundary is traction free.
The exact solution for the stress components [49] is
   
a2 p b2 a2 p b2
r (r ) = 1− ,  (r ) = 1+ , r  = 0 (96)
b2 −a 2 r2 b2 −a 2 r2

whereas the radial and the tangential exact displacements are given by

a 2 pr b2
u r (r ) = 1−+ 2 (1+) , u = 0 (97)
E(b2 −a 2 ) r

where (r, ) are the polar coordinates, and  is measured counter-clockwise from the positive
x-axis.
The convergence curves of the strain energy solution of the NS-FEM-Q4 and the ES-FEM-T3 in
comparison with the FEM-Q4 and the FEM-EQ4 are plotted in Figure 14. Again, the NS-FEM-Q4
gives upper-bound solutions in the strain energy, and the solution of the ES-FEM-T3 is more
accurate than those of the FEM-Q4.
Figure 15 plots the convergence rate in the displacement error norm of three methods:
ES-FEM-T3, NS-FEM-Q4 and FEM-Q4. Again, the ES-FEM-T3 stands out clearly. When the
finest mesh (h = 0.06) is used, the error of the ES-FEM-T3 is about 12 of the FEM-Q4 whereas
the error of the FEM-Q4 is about 14 of the NS-FEM-Q4. In terms of convergence rate, the
super-convergence is again observed for the ES-FEM-T3 with a rate of 2.36 that is much higher
than the theoretical value of 2.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1252 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

32.4 Exact

32.2

32

Strain energy
31.8

31.6

31.4

31.2

31
10 15 20 25 30 35 40 45 50
Mesh density

Figure 14. Convergence of the strain energy solution of the NS-FEM-Q4 and the ES-FEM-T3 in comparison
with the FEM-Q4 and the FEM-EQ4 for the hollow cylinder problem.
log (displacement error norm)
10

log h
10

Figure 15. Displacement error norms for the hollow cylinder problem using the
NS-FEM-Q4, ES-FEM-T3 and FEM-Q4.

The energy error norms of the methods obtained by the corresponding equations (87), (88) and
(89) are listed in Table II and are plotted in Figure 16. The same results as in the previous example
are obtained. First, the error of the FEM-EQ4 is smallest due to the use of a direct approximation
of stress field in the equilibrium approach. Without using the recovery energy error norm (89), the

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1253

Table II. Error in energy norm obtained using different methods for the hollow
cylinder using the same set of meshes.
Mesh 12×12 Mesh 24×24 Mesh 48×48 Mesh 72×72
h(m) 0.362 0.181 0.090 0.060
FEM-Q4 9.1349e−01 4.6487e−01 2.3354e−01 1.5583e−01
NS-FEM-Q4 5.4925e−01 2.0293e−01 7.3729e−02 4.0532e−02
ES-FEM-T3 6.1346e−01 2.3480e−01 8.6321e−02 4.7591e−02
FEM-Q4-Re 7.2458e−01 3.0001e−01 1.1635e−01 6.5472e−02
NS-FEM-Q4-Re 3.5371e−01 1.4536e−01 5.6234e−02 3.1606e−02
ES-FEM-T3-Re 6.9870e−01 2.6871e−01 1.1039e−01 6.1935e−02
FEM-EQ4 2.2143e−01 5.8096e−02 1.4720e−02 6.5588e−03

0
log (energy error norm)
10

log10h

Figure 16. Energy error norm for the S-FEM models in comparison with those of other
methods for the hollow cylinder problem.

errors of the NS-FEM-Q4 and the ES-FEM-T3 are much smaller than those of the FEM-Q4. By
using the recovery energy error norm (89), the error of the NS-FEM-Q4-Re is still much smaller
than that of the FEM-Q4-Re, whereas the error of the ES-FEM-T3-Re is slightly smaller than that
of the FEM-Q4-Re. In terms of convergence rate, it is seen that the convergence rate of all S-FEM
models (from 1.348 to 1.455) is higher than that of the FEM-Q4 (0.987) and lower than that of
the FEM-EQ4 (1.965). Again, the convergence rates of the S-FEM models are quite close to the
theoretical rate of 1.5 given in [41].
It is again confirmed in this example that the S-FEM models can reveal all interesting properties
as shown in the previous problem.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1254 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

9. CONCLUSIONS

In this work, we have presented and proven a number of properties and theorems for the S-FEM
models. These properties and theorems have also been confirmed via numerical examples solved
using two representative S-FEM models in 2D problems: NS-FEM-Q4 and ES-FEM-T3. From
these studies, the following conclusions can be drawn:
• The S-FEM models are created using the linear PIM for assuming the displacement field
without mapping, and the smoothing projection for strain field modification. The smoothing
projection is proven to be orthogonal, distributive and commutative in bi-linear forms.
• As long as a minimum number of linearly independent smoothing domains are used, the
S-FEM models are always stable and convergent. The convergence of the S-FEM models is
bounded in general by two terms: one from the accuracy of the integral representation of a
function and another from the accuracy of the displacement field approximation.
• The S-FEM models are displacement-like models without using any additional DOFs and are
always variationally consistent due to the orthogonal property of the smoothing projection.
• The S-FEM models are softer than the fully compatible FEM model and hence the S-FEM
models are more accurate and have higher convergence rates in displacement solution or stress
solution or both.
Finally, we note that the present theoretical studies can be applicable to the Natural Neighbour
Method based on the FdV principle [13] in which it is assumed that the strains and the stresses
are constant over each Voronoi cell associated with each node.

ACKNOWLEDGEMENTS
The authors acknowledge Professor J. F. Debongnie from the University of Liège for his suggestion and
advice in the error analysis of the S-FEM models for this work. This work is partially supported by
A*Star, Singapore (SERC Grant No: 052 101 0048). It is also partially supported by the Open Research
Fund Program of the State Key Laboratory of Advanced Technology of Design and Manufacturing for
Vehicle Body, Hunan University, People’s Republic of China under the grant number 40915001.

REFERENCES
1. Liu GR. Mesh-Free Methods: Moving Beyond the Finite Element Method (2nd edn). CRC Press: Boca Raton,
2009.
2. Nagashima T. Node-by-node meshless approach and its applications to structural analyses. International Journal
for Numerical Methods in Engineering 1999; 46:341–385.
3. Dohrmann CR, Heinstein MW, Jung J, Key SW, Witkowski WR. Node-based uniform strain elements for three-
node triangular and four-node tetrahedral meshes. International Journal for Numerical Methods in Engineering
2000; 47:1549–1568.
4. Puso MA, Solberg J. A stabilized nodally integrated tetrahedral. International Journal for Numerical Methods in
Engineering 2006; 67:841–867.
5. Puso MA, Chen JS, Zywicz E, Elmer W. Meshfree and finite element nodal integration methods. International
Journal for Numerical Methods in Engineering 2008; 74:416–446.
6. Chen JS, Wu CT, Yoon S, You Y. A stabilized conforming nodal integration for Galerkin mesh-free methods.
International Journal for Numerical Methods in Engineering 2001; 50:435–466.
7. Yoo JW, Moran B, Chen JS. Stabilized conforming nodal integration in the natural-element method. International
Journal for Numerical Methods in Engineering 2004; 60:861–890.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
A THEORETICAL STUDY ON THE SMOOTHED FEM (S-FEM) MODELS 1255

8. Liu GR, Zhang GY, Dai KY, Wang YY, Zhong ZH, Li GY, Han X. A linearly conforming point interpolation
method (LC-PIM) for 2D solid mechanics problems. International Journal of Computational Methods 2006;
2:645–665.
9. Zhang GY, Liu GR, Wang YY, Huang HT, Zhong ZH, Li GY, Han X. A linearly conforming point interpolation
method (LC-PIM) for three-dimensional elasticity problems. International Journal for Numerical Methods in
Engineering 2007; 72:1524–1543.
10. Liu GR, Li Y, Dai KY, Luan MT, Xue W. A linearly conforming radial point interpolation method for solid
mechanics problems. International Journal of Computational Methods 2006; 3:401–428.
11. Liu GR, Zhang GY. Edge-based smoothed point interpolation method (ES-PIM). International Journal of
Computational Methods 2008; 5(4):621–646.
12. Liu GR. A generalized Gradient smoothing technique and the smoothed bilinear form for Galerkin formulation
of a wide class of computational methods. International Journal of Computational Methods 2008; 5(2):199–236.
13. Cescotto S, Li X. A natural neighbour method for linear elastic problems based on Fraeijs de Veubeke variational
principle. International Journal for Numerical Methods in Engineering 2007; 71:1081–1101.
14. Liu GR, Dai KY, Nguyen-Thoi T. A smoothed finite element for mechanics problems. Computational Mechanics
2007; 39:859–877.
15. Liu GR, Nguyen-Thoi T, Dai KY, Lam KY. Theoretical aspects of the smoothed finite element method (SFEM).
International Journal for Numerical Methods in Engineering 2007; 71:902–930.
16. Nguyen-Thoi T, Liu GR, Dai KY, Lam KY. Selective smoothed finite element method. Tsinghua Science and
Technology 2007; 12(5):497–508.
17. Dai KY, Liu GR, Nguyen-Thoi T. An n-sided polygonal smoothed finite element method (nSFEM) for solid
mechanics. Finite Element Analysis and Design 2007; 43:847–860.
18. Nguyen-Xuan H, Bordas S, Nguyen-Dang H. Smooth finite element methods: convergence, accuracy and properties.
International Journal for Numerical Methods in Engineering 2008; 74:175–208.
19. Nguyen-Xuan H, Bordas S, Nguyen-Dang H. Addressing volumetric locking and instabilities by selective
integration in smoothed finite elements. Communications in Numerical Methods in Engineering 2009; 25:19–34.
20. Liu GR, Nguyen-Thoi T, Nguyen-Xuan H, Dai KY, Lam KY. On the essence and the evaluation of the shape
functions for the smoothed finite element method (SFEM) (Letter to Editor). International Journal for Numerical
Methods in Engineering 2009; 77:1863–1869.
21. Nguyen-Xuan H, Rabczuk T, Bordas S, Debongnie JF. A smoothed finite element method for plate analysis.
Computer Methods in Applied Mechanics and Engineering 2008; 197:1184–1203.
22. Nguyen-Thanh N, Rabczuk T, Nguyen-Xuan H, Bordas S. A smoothed finite element method for shell analysis.
Computer Methods in Applied Mechanics and Engineering 2008; 198:165–177.
23. Nguyen-Xuan H, Nguyen-Thoi T. A stabilized smoothed finite element method for free vibration analysis of
Mindlin–Reissner plates. Communications in Numerical Methods in Engineering 2009; 25(8):882–906.
24. Liu GR, Nguyen-Thoi T, Nguyen-Xuan H, Lam KY. A node based smoothed finite element method (NS-FEM)
for upper bound solution to solid mechanics problems. Computers and Structures 2009; 87:14–26.
25. Nguyen-Thoi T, Liu GR, Nguyen-Xuan H. Additional properties of the node-based smoothed finite element method
(NS-FEM) for solid mechanics problems. International Journal of Computational Methods 2009; 6:633–666.
26. Nguyen-Thoi T, Liu GR, Nguyen-Xuan H, Nguyen-Tran C. Adaptive analysis using the node-based smoothed
finite element method (NS-FEM). Communications in Numerical Methods in Engineering 2009; DOI: 10.1002/
cnm.1291.
27. Liu GR, Nguyen-Thoi T, Lam KY. An edge-based smoothed finite element method (ES-FEM) for static, free
and forced vibration analyses of solids. Journal of Sound and Vibration 2009; 320:1100–1130.
28. Nguyen-Thoi T, Liu GR, Nguyen-Xuan H. An n-sided polygonal edge-based smoothed finite element method (nES-
FEM) for solid mechanics. Communications in Numerical Methods in Engineering 2009; DOI: 10.1002/cnm.1375.
29. Nguyen-Xuan H, Liu GR, Nguyen-Thoi T, Nguyen-Tran C. An edge-based smoothed finite element method
(ES-FEM) for static, free and forced vibration analyses of solids. Smart Materials and Structures 2009; 18:12,
065015.
30. Nguyen-Xuan H, Liu GR, Thai-Hoang C, Nguyen-Thoi T. An edge-based smoothed finite element method with
stabilized discrete shear gap technique for analysis of Reissner–Mindlin plates. Computer Methods in Applied
Mechanics and Engineering 2010; 199(9–12):471–489.
31. Nguyen-Thoi T, Liu GR, Vu-Do HC, Nguyen-Xuan H. An edge-based smoothed finite element method (ES-FEM)
for visco-elastoplastic analyses of 2D solids using triangular mesh. Computational Mechanics 2009; 45:23–44.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme
1256 G. R. LIU, H. NGUYEN-XUAN AND T. NGUYEN-THOI

32. Tran TN, Liu GR, Nguyen-Xuan H, Nguyen-Thoi T. An edge-based smoothed finite element method for
primal–dual shakedown analysis of structures. International Journal for Numerical Methods in Engineering 2010;
82(7):917–938.
33. Nguyen-Thoi T, Liu GR, Lam KY, Zhang GY. A face-based Smoothed Finite Element Method (FS-FEM) for
3D linear and nonlinear solid mechanics problems using 4-node tetrahedral elements. International Journal for
Numerical Methods in Engineering 2009; 78:324–353.
34. Nguyen-Thoi T, Liu GR, Vu-Do HC, Nguyen-Xuan H. A face-based smoothed finite element method (FS-FEM)
for visco-elastoplastic analyses of 3D solids using tetrahedral mesh. Computer Methods in Applied Mechanics
and Engineering 2009; 198:3479–3498.
35. Liu GR, Nguyen-Thoi T, Lam KY. A novel Alpha Finite Element Method (FEM) for exact solution to mechanics
problems using triangular and tetrahedral elements. Computer Methods in Applied Mechanics and Engineering
2008; 197:3883–3897.
36. Fraeijs De Veubeke B. Displacement sand equilibrium models in the finite element method. In Stress Analysis,
Chapter 9, Zienkiewicz OC, Holister G (eds). Wiley: New York, 1965; 145–197. Reprinted in International
Journal for Numerical Methods in Engineering 2001; 52:287–342.
37. Nguyen-Dang H. Finite element equilibrium analysis of creep using the mean value of the equivalent shear
modulus. Meccanica (AIMETA, Italy) 1980; 15:234–245.
38. Nguyen-Dang H. Sur la plasticité et le calcul des états limites par éléments finis. Ph.D. Thesis, Publication No.
98, Faculty of Applied Sciences, University of Liège, 1985.
39. Debongnie JF, Zhong HG, Beckers P. Dual analysis with general boundary conditions. Computer Methods in
Applied Mechanics and Engineering 1995; 122:183–192.
40. Debongnie JF, Nguyen-Xuan H, Nguyen HC. Dual analysis for finite element solutions of plate bending.
Proceedings of the Eighth International Conference on Computational Structures Technology. Civil-Comp Press:
Stirlingshire, Scotland, 2006.
41. Liu GR. A G space and weakened weak (W2) form for a unified formulation of compatible and incompatible
methods, part I—theory and part II—application to solid mechanics problems. International Journal for Numerical
Methods in Engineering 2009; DOI: 10.1002/nme.2720.
42. Bathe KJ. Finite Element Procedures. Prentice-Hall, MIT: Englewood Cliffs, NJ, MA, 1996.
43. Hughes TJR. The Finite Element Method. Prentice-Hall: Englewood Cliffs, NJ, 1987.
44. Brezzi F, Fortin M. Mixed and Hybrid Finite Element Methods. Springer: New York, 1991.
45. Liu GR, Zhang GR. A normed G space and weakened weak (W2) formulation of a cell-based Smoothed Point
Interpolation Method. International Journal of Computational Methods 2009; 6(1):147–179.
46. Liu GR. On a G space theory. International Journal of Computational Methods 2009; 6(2):257–289.
47. Johnson C, Mercier B. Some equilibrium finite element methods for two-dimensional elasticity problems.
Numerische Mathematik 1978; 30(1):103–116.
48. Felippa C. Stress recovery. Introductions to Finite Element Methods (ASEN 5007), Chapter 28, 2009. Available
from: http://www.colorado.edu/engineering/CAS/courses.d/IFEM.d/.
49. Timoshenko SP, Goodier JN. Theory of Elasticity (3rd edn). McGraw-Hill: New York, 1987.

Copyright 䉷 2010 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2010; 84:1222–1256
DOI: 10.1002/nme

S-ar putea să vă placă și