Acta Mech

DOI 10.1007/s00707-015-1300-z

D. Garijo · Ó. F. Valencia · F. J. Gómez-Escalonilla

Global interpolating MLS shape functions for structural

problems with discrete nodal essential boundary conditions

Received: 16 August 2014 / Revised: 13 December 2014

© Springer-Verlag Wien 2015

Abstract The moving least squares (MLS)-based element-free Galerkin method (EFGM) has been widely
studied, yielding valuable results and becoming a robust mesh-free technique for structural analysis. Due to the
non-interpolating nature of MLS shape functions, the EFGM procedure leads to stiffness and mass matrices
based on nodal parameters instead of true nodal displacements. The interpolating moving least squares (IMLS)
methods with singular weight functions or the weighted nodal least squares (WNLS) methods were proposed in
order to arrange the formulation in terms of actual displacements at the nodes. In this work, a later transformation
matrix to retrieve the classic stiffness and mass matrices is discussed and numerically tested in problems with
discrete prescribed displacements at boundaries, for which a straightforward finite element method (FEM)-type
imposition of boundary conditions is possible.
Mathematics Subject Classification 35J25 · 65N99 · 74S30

1 Introduction

In many mesh-free methods, the techniques employed to construct the shape functions are non-interpolating
approximation algorithms which do not possess the Kronecker delta property. This is the case for of one of the
most popular mesh-free methods, the element-free Galerkin method (EFGM) [1], based on the moving least
squares (MLS) approximation [2]. The system of equations derived from the discretization of the weak form is
solved in terms of nodal parameters, but not actual nodal displacements. As a consequence, in solid mechanics
the stiffness and mass matrices lack direct physical interpretation. Besides, enforcement of essential boundary
conditions or coupling with other schemes is not straightforward and requires notable additional computational
effort. In this sense, strategies such as the Penalty method [3], Nitsche’s method [4] or the Lagrange multipliers
[5] are efficient resources for the imposition of boundary conditions in displacements, but they require specific
programming and introduce new unknown variables in the analysis.
Lancaster and Salkauskas anticipated the use of singular weight functions to provide MLS with interpolating
capability [2]. This idea was explored and consolidated by Kaljevic and Saigal [6], whose research led to the
D. Garijo (B)
Safran Engineering Services (Safran Group), Calle Arnaldo de Vilanova, 5, 28906 Getafe, Madrid, Spain
E-mail: [email protected]
Tel.: +34 914 913 005

Ó. F. Valencia
FACC AG, Kammer 32, 4974 Ort im Innkreis, Austria
E-mail: [email protected]

F. J. Gómez-Escalonilla
Airbus Defence and Space, Avda John Lennon s/n, 28906 Getafe, Madrid, Spain
E-mail: [email protected]
 D. Garijo et al.

interpolating moving least squares (IMLS) methods. The IMLS methods were aimed at alleviating one of the
major limitations of MLS, the non-straightforward prescription of boundary conditions [7]. Improvements in
IMLS formulation were introduced by Ren and Cheng in order to increase its computational efficiency, yielding
the interpolating element-free Galerkin (IEFG) method, tested in elasticity [8] and potential [9] problems.
An enhancement to the MLS weighting procedure that leads to an interpolating scheme with non-singular
values of the weights and the coefficient matrices in the whole domain was developed by Most and Bucher
[10]. Further research drove to a new MLS technique in which least squares coefficients are calculated only at
the nodes, the so-called weighted nodal least squares (WNLS) method [11], that was posed and numerically
experimented with success in classic structural benchmarks.
Instead of dealing with modified singular or non-singular weight functions, Chen et al. [12,13] had proposed
a matrix inversion to achieve an approach with the Kronecker delta property in their studies of nonlinear
problems with reproducing kernel particle methods (RKPM). In this case, once stiffness and mass matrices
have been assembled, a later transformation can restore the formulation of non-interpolating methods in terms
of nodal displacements.
In general problems, however, verification of Kronecker delta property does not suffice to allow direct FEM-
type imposition of essential boundary conditions [14]. Since interpolating shape functions of interior nodes
vanish at the nodes laying on the boundary but generally not at the interstices between them, the enforcement
of distributed essential boundary conditions along the whole boundary still requires the contribution of the
internal nodes laying on a domain around the studied point. The general use of interpolating shape functions
with a straightforward FEM-type imposition of boundary conditions is then limited to problems with discrete
(nodal) prescribed displacements at boundaries. This lack of generality is a major drawback of the interpolating
MLS, but there is a relatively large number of applications which can be tackled with this strategy.
In this work, the later matrix transformation procedure suggested by Chen et al. is developed and discussed
for the EFGM. The presented numerical tests focus on the relative performance of the modified interpolating
MLS shape functions versus the standard MLS approach in structural applications. Benchmark problems of
elastostatics and modal analysis with discrete nodal essential boundary conditions are subjected to experimen-
tation in order to evaluate the deviations in error and computational effort between both approximations. In
the standard EFGM, the prescribed displacements are enforced using Lagrange multipliers; in the modified
EFGM, the size of the transformed stiffness and mass matrices is directly reduced by deleting the prescribed
degrees-of-freedom. Thus, the penalty in terms of machine time due to the required operations for the genera-
tion and inversion of the transformation matrix is measured versus the cost associated to the classic introduction
of Lagrange multipliers. Problems with distributed boundary conditions in displacements between nodes are
also analysed in order to find a correlation between the assumed error at the essential boundaries and the global
error of the interpolating approximation.

2 Standard element-free Galerkin method

The EFGM [1] constructs the numerical approximation uh to a function u in a given point x ∈ Ω from the
contributions of a discrete set of nodes laying on a surrounding domain Ωe of x. Those contributions are
determined by the MLS [2] shape functions φi , leading to:

N
uh (x) = φi (x) ai = Φ (x) as , (1)
i=1
Φ (x) = (φ1 (x) φ2 (x) . . . φ N (x)) , (2)
asT = (a1 a2 . . . a N ) , (3)
φi (x) being the shape function associated with evaluation point xi and ai the ith evaluation point parameter,
which h
 is not equal to the value of u in xi due to the lack of Kronecker delta property of MLS shape functions,
φi x j  = δi j .
In solid mechanics, Galerkin’s weak form constrained via Lagrange multipliers for general dynamic prob-
lems yields [5]:
  
ρδu üdΩ +
T
δ (Lu) C (Lu) dΩ −
T
δuT bdΩ −
Ω Ω
  Ω
− δu t̄dΓ −
T
δλ (u − ū) dΓ −
T
δuT λdΓ = 0, (4)
Γt Γu Γu
Global interpolating MLS shape functions

where L is the governing differential operator in the domain Ω, Γt is the contour where natural boundary
conditions t̄ are applied, Γu is the contour in which essential boundary conditions ū are prescribed, λ is the
set of Lagrange multipliers to constrain essential boundary conditions, ρ is the material density, and C is the
characteristic matrix of the material.
For elastostatic analysis, Eq. (4) is reduced to
    
δ (Lu)T C (Lu) dΩ − δuT bdΩ − δuT t̄dΓ − δλT (u − ū) dΓ − δuT λdΓ = 0. (5)
Ω Ω Γt Γu Γu

Inserting approximation (1) into Eq. (5) gives

N N  
N  
N 
 
δaiT
(Lφi ) C Lφ j dΩa j −
T
δai
T
φi bdΩ − δai
T
φi t̄dΓ −
i=1 j=1 Ω i=1 Ω i=1 Γt
(6)

nλ 
N  
N 
nλ 
− δλiT NiT (φk ak − ū) dΓ − δaiT φi Nk dΓ λk = 0.
i=1 k=1 Γu i=1 k=1 Γu

The array of Lagrange multipliers

 
λT = λ1 λ2 . . . λn λ (7)
can be expressed in compact form as follows:
⎡⎤⎧ ⎫

nλ Ni 0 0 ⎨ λxi ⎬

nλ
λ= Ni λi = ⎣ 0 Ni 0 ⎦ λ yi , (8)
i=1 i=1 0 0 Ni ⎩ λzi ⎭
where the Lagrange interpolants Nk of order n in a two-dimensional analysis are given by
 (s − sh )
Nkn (s) = , (9)
0≤h≤n
(sk − sh )
h=k

s being the arc length parameter along the boundary Γu .

Classic simplification n = 1 (see Dolbow and Belytschko [15]) leads to the interpolants:
s − s1 s − s0
N0 (s) = N1 (s) = (10)
s0 − s1 s1 − s0
for s = s0 and s = s1 , respectively.
In a similar way to interpolation (8), the virtual Lagrange multipliers can be written as follows:

nλ
δλ = Ni δλi . (11)
i=1

Once virtual displacements and virtual Lagrange multipliers are suppressed, the matrix form of the set of
Eq. (6) results in
    
K G as f
= (12)
GT 0 λ q
with:
  
Ki j = DiT CD j dΩ, fi = φi bdΩ + φi t̄dΓ,
Ω Ω Γt
 
Gik = − φi Nk dΓ, qk = − NkT dΓ ū,
Γu Γu (13)
⎡ ∂ ∂ ∂
⎤
∂x 0 0 0 ∂z ∂y
⎢ ⎥
Di = Lφi = ⎣ 0 ∂∂y 0 ∂z ∂
0 ∂∂x ⎦ φi .
∂ ∂ ∂
0 0 ∂z ∂ y ∂ x 0
 D. Garijo et al.

System (12)–(13) allows to retrieve the array of nodal parameters as . Nodal displacements ui are computed
immediately by means of application of Eq. (1) as:


N
uh (xi ) = ui = φ j (xi ) a j (14)
j=1

and, in general, ui  = ai . Therefore, matrices K, G and f are based on nodal parameters and cannot be directly
compared with the corresponding ones of interpolating-type schemes, including the FEM approach.
Extension of discretized weak form (6) to the free vibration analysis is performed by adding the inertial
terms:
N  N 
 
δaiT ρ (φi )T φ j dΩa j (15)
i=1 j=1 Ω

and then system (12) turns into

    
K − ω2 M G as 0
= (16)
GT 0 λ 0

with the mass matrix based on nodal parameters:


 
Mi j = ρ (φi )T φ j dΩ, (17)
Ω

ω in (16) being the natural frequencies of the structure. Solving for the eigenvalues of the system of equations,
the eigenvectors as can be obtained immediately. This way, as in elastostatic analysis, Eq. (14) is applied, and
displacements at the evaluation points can be finally determined.

3 Element-free Galerkin method based on nodal displacements

In this section, a modification of the calculation process of displacements at the evaluation points is introduced.
When Eq. (6) is reached, instead of extracting nodal parameters ai out of the integrals, authors follow the
transformation proposed by Chen et al. [12,13].
Nodal parameters ai can be transformed into actual displacements ui at such points using the approximation
given by (1), considering that shape functions are known by the MLS technique:
U = Sas (18)
with UT = (u1 u2 . . . u N ), asT = (a1 a2 . . . a N ) and where S is the transformation matrix
⎡ ⎤
φ1 (x1 ) φ2 (x1 ) · · · φ N (x1 )
⎢ φ1 (x2 ) φ2 (x2 ) · · · φ N (x2 ) ⎥
S=⎢ ⎣ .. .. .. .. ⎥,
⎦ (19)
. . . .
φ1 (x N ) φ2 (x N ) · · · φ N (x N )
that is, Si j = φ j (xi ). Therefore,


N 
N
uh (x) = φi (x) ai = Φ (x) S−1 U = Ψ (x) U = ψ j (x) u j , (20)
i=1 j=1
 
where ψ j (x) are the modified MLS shape functions, which possess interpolating nature, ψi x j = δi j .
h (x) = Ψ (x) U = Φ (x) S−1 U. Rearranging Eq. (5) in
Consequently, the derivatives of uh (x) become u,x ,x ,x
compact notation,
 T  T  T   
δUT S−1 KS−1 U + S−1 Gλ − S−1 f + δλT GT S−1 U − q = 0, (21)
Global interpolating MLS shape functions

and forcing verification of previous equation for all the virtual displacement and Lagrange multipliers fields,
    
K̂ Ĝ U
= f̂ , (22)
Ĝ
T
0 λ q

where the matrices K̂, Ĝ and f̂ are now based on true nodal displacements:
 T
K̂ = S−1 KS−1 , (23)
 T
Ĝ = S−1 G, (24)
 T
f̂ = S−1 f. (25)
For free vibration problems, the modified MLS mass matrix is derived in a similar way as
 T
M̂ = S−1 MS−1 , (26)
leading to the system
    
K̂ − ω2 M̂ Ĝ U 0
= . (27)
Ĝ
T
0 λ 0

Furthermore, if the essential boundary conditions CΓ (u) = 0 are prescribed at k discrete nodes,
u j − ū j = 0 , j = 1, 2, . . . , k, (28)
then it is possible to constrain directly the problem by deleting the displacements involved in Eq. (28), leading to
reduced modified stiffness (K̂red ) and mass (M̂red ) matrices in a system assembled accordingly to the enforced
degrees-of-freedom. In a typical configuration with fixed nodes,

K̂red Ured = f̂red (29)

for the elastostatic analysis and
 
K̂red − ω2 M̂red Ured = 0 (30)

for the vibration analysis. This procedure is of general application to one-dimensional boundary value prob-
lems, in which the boundaries Γu become the end points of the interval of definition of the differential equation.
For problems with non-discrete essential boundary conditions, if the IMLS shape functions are employed in
combination with a straightforward procedure to enforce the boundary conditions, the local error in displace-
ments at boundaries Γu is given by the expression:
   N


 
error Γu = ū − ψi (x̄) ui  dΓ, x̄ ∈ Γu . (31)
Γu  
i=1

The one-dimensional modified shape functions ψi are shown together with the original MLS shape functions
 T  T
in Fig. 1. In the general case, the products S−1 KS−1 and S−1 MS−1 result in full matrices K̂ and M̂; then,
shape functions ψi lose their local support and become globally extended or globally supported interpolating
functions, as illustrated in Figs. 1 and 2.
The typical oscillation of interpolating functions is tuned in the modified shape functions ψi by means of
the parameter dmax [1] (Fig. 3), which determines the size of the support domain in classic MLS [15].
The transformation (23)–(26) is also applicable when dealing with non-equally spaced nodal patterns
provided no spatial location is shared by two or more evaluation points, xi  = x j ∀ i  = j. As a result of MLS
weighting, S is, in the general case, a non-singular band matrix and then transformation is feasible. The aspect
of interpolating MLS shape functions will, however, be much less predictable than that in the case of equally
spaced nodal distributions. In Fig. 4, two examples of interpolating MLS shape functions for random nodal
distributions are provided. Random distributions are tuned by the parameter Λ, which controls the maximum
width of the random deviation of nodes from original equally spaced locations.
 D. Garijo et al.

Fig. 1 MLS shape functions ψi and φi , cubic spline weight function, a 7 nodes, dmax = 2.0; b 15 nodes, dmax = 3.0

Fig. 2 Interpolating MLS shape functions ψi in a two-dimensional domain, a 7 × 7 nodes, dmax = 2.0; b 9 × 9 nodes, dmax = 3.0

Fig. 3 Interpolating MLS shape functions ψi , cubic spline weight function, a 11 nodes, dmax = 2.0; b 11 nodes, dmax = 4.0

The loss of the local support of the original MLS shape functions with the introduction of the matrix S may
also imply the loss of the original sparse structure and specific properties of the matrices in (12). This issue is
analysed in the numerical examples by means of the comparison of the L 2 norm-based condition numbers of
the K matrices of MLS and IMLS approximations:
     −1   
   
κMLS (K) = K−1 2 · K2 ; κIMLS K̂ = K̂  · K̂ . (32)
2 2

Additionally, the conditioning of the IMLS system after direct enforcing of essential boundary conditions is
included in the comparisons by plotting the condition number of the reduced stiffness matrix:
Global interpolating MLS shape functions

Fig. 4 MLS shape functions ψi and φi , cubic spline weight function, a 11 nodes, dmax = 2.5, Λ = 0.3; b 15 nodes, dmax = 3.0,
Λ = 0.4

   −1   
   
κ I M L Sred K̂red = K̂red  · K̂red  (33)
2 2

In the case of free vibration analysis, the conditioning of the problems solved is studied using the condition
numbers of the numerical eigenvector matrices computed.
Besides, numerical dissipation induced by matrix–matrix multiplication may also arise and affect the
accuracy of the IMLS approach for large problems. This question is also investigated by solving the numerical
examples presented with high nodal densities and performing convergence analysis.

4 Numerical experimentation

The interpolating MLS scheme obtained by means of transformation matrix S is tested in seven structural
applications. The relative performance of interpolating MLS and standard MLS approximations when they
are incorporated into a variational Galerkin procedure is analysed. In all cases (except the problem of a hole
in an infinite plate), distributions of equally spaced nodes are considered; the numerical integration of the
terms of the matrix systems is performed by means of a classic Gaussian quadrature; a first-order polynomial
intrinsic basis is selected and the weight function chosen is a cubic spline as referred in Belytschko et al. [16].
In the standard EFGM approximation, the essential boundary conditions are enforced by means of Lagrange
multipliers; in the EFGM with global interpolating MLS shape functions, the straightforward imposition of
boundary conditions leading to reduced stiffness and mass matrices is applied.

4.1 One-dimensional rod elastostatic analysis

This benchmark of meshless analysis has been discussed in-depth in the literature, see Dolbow and Belytschko
[15], Liu and Gu [17] or Valencia et al. [18]. The boundary where essential conditions are prescribed is a single
node; therefore, the displacement at this node can be directly enforced if interpolating MLS shape functions
are used. The size of the matrix S is in this case equal to N xN , N being the number of nodes to discretize the
rod. The numerical accuracy is measured by means of the global norm of error in strain energy εe as
 
1
εe = log10 εT E ε dΩ (34)
2 Ω

where ε = ε num − εexact , εnum being the array of numerical strains, εexact the array of exact strains, and E
the material’s Young’s modulus.
The rate of convergence in norm εe is presented in Fig. 5a. The relative computational cost is provided
in Fig. 5b with the parameter rt = tIMLS /tMLS , tIMLS being the elapsed time from the start of the assembly
of system (29) until the retrieval of the nodal parameters (nodal values U, in case of IMLS) and tMLS is the
 D. Garijo et al.

Fig. 5 One-dimensional rod elastostatic analysis, a rate of convergence in strain energy, interpolating MLS shape functions; b
ratio of computational cost rt

Fig. 6 One-dimensional rod elastostatic analysis up to 201 nodes, a condition numbers of MLS and IMLS approximations; b εe
versus CPU time

same time defined for the standard MLS approach. The evolution of the condition numbers κMLS , κIMLS and
κ I M L Sred with the increasing number of degrees-of-freedom is presented in Fig. 6a. The diagram in Fig. 6b
illustrates the MLS and IMLS performances εe versus CPU time for high nodal densities. Several values of
the parameter dmax are tested.

Figure 5a shows that the transformation using the matrix S modifies neither the accuracy nor the convergence
rate of the original EFGM [15]. In Fig. 5b, it can be noticed that, after initial widespread data for small problems,
for all the dmax tested, the ratios rt converge to the band rt = 1–1.1 as N increases, where they keep confined
for large problems. The penalty of computational cost due to the introduction of matrix S can be considered
negligible and is balanced by a simpler FEM-type implementation. The impact of the S matrix transformation
on the conditioning of this problem is also very small, as read from Fig. 6a. Even for high nodal densities, the
FEM-type procedure yields low and stable values of κ I M L Sred . If the G matrix terms in (12) are not considered
to compute the MLS condition numbers, it can be checked that κIMLS /κMLS = κ(K̂)/κ(K) ∼ 1 for all nodal
densities and dmax values. This is verified in all the numerical examples presented in this work, as shown
in the corresponding included diagrams. Table 1 compiles results for several nodal densities and values of
dmax .
Global interpolating MLS shape functions

Table 1 Results for one-dimensional rod elastostatic analysis

Nodes dmax Interpolating MLS Standard MLS εe (%) rt

Norm εe t(s) Norm εe t(s)
11 2.0 −2.60283923 0.0940 −2.60283923 0.0470 0 2.00
2.5 −2.97336082 0.0150 −2.97336082 0.0150 0 1.00
4.0 −3.59272036 0.0460 −3.59272036 0.0310 0 1.48
51 2.0 −3.65260640 0.1250 −3.65260640 0.1250 0 1.00
2.5 −4.01178249 0.1250 −4.01178249 0.1250 0 1.00
4.0 −4.69459815 0.1250 −4.69459815 0.1410 0 0.89
101 2.0 −4.10431442 0.2340 −4.10431442 0.2340 0 1.00
2.5 −4.42949726 0.2500 −4.42949726 0.2490 0 1.00
4.0 −5.14671598 0.2810 −5.14671598 0.2650 0 1.06

4.2 One-dimensional rod free vibration analysis

Another remarkable problem of meshless literature is the one-dimensional rod free vibration analysis, studied
with EFGM, for example, in Tiago and Leitão [19] or Valencia et al. [20]. The exact solution for the natural
frequencies ωn can be found in Timoshenko [21]:
!
π EA
ωn = (2n − 1) , n = 1, 2 . . . , (35)
2 m
where E is the material’s Young’s modulus, A is the cross-sectional area, and m is the mass of the rod. For a
numerical approach, the accuracy is typically computed by means of the norm of error E n :
" "
"ωnum − ωexact "
n n
E n = log10 . (36)
ωnexact

The straightforward procedure for the imposition of essential boundary conditions is also applicable to
this one-dimensional case. The rates of convergence in the norm of error E n for IMLS are provided in Fig. 7.
The results presented in Fig. 8a–c correspond to the deviation of norm E n between interpolating and standard
MLS approximations:

E nIMLS − E nMLS
Δω (n) = (37)
E nMLS
for the first three natural frequencies of the rod.

Fig. 7 One-dimensional rod free vibration analysis, rates of convergence in E n , interpolating MLS shape functions, a natural
frequency 1; b natural frequency 2
 D. Garijo et al.

Fig. 8 One-dimensional rod free vibration analysis, a error Δω in natural frequency 1; b error Δω in natural frequency 2; c error
Δω in natural frequency 3; d ratio of computational cost rt

Fig. 9 One-dimensional rod free vibration analysis up to 201 nodes, a condition numbers of MLS and IMLS approximations;
b E 1 versus CPU time
Global interpolating MLS shape functions

Fig. 10 Timoshenko beam elastostatic analysis, a rate of convergence in strain energy, interpolating MLS shape functions,
dmax = 2.0; b ratio of computational cost rt

The relative computational cost is shown in Fig. 8d, where the parameter rt = tIMLS /tMLS is plotted as a
function of the nodal density used in the discretization. Times tIMLS and tMLS are defined as the elapsed times
from the start of the assembly of the free vibration systems of equations until the eigenvalues of the problem
are calculated. Figure 9a gathers the evolution of the condition numbers of the eigenvector matrices up to
201 nodes. Figure 9b presents the corresponding E 1 versus CPU cost diagram, comparing MLS and IMLS
approximations. Again, several values of dmax are tested.
Figure 8a–c verifies that deviation Δω (n) is negligible for the three natural frequencies when the number
of nodes N is small. A slight increment in Δω (n) is observed as N becomes larger, although the order of
the magnitude of the deviation is still very small (10−5 –10−8 ). The condition number κIMLS obtained from
reduced stiffness and mass matrices remains below κMLS for all the nodal densities and dmax values, as can be
observed in Fig. 9a. This suggests that the increment in Δω (n) may be caused by numerical dissipation, note
that the size of transformation matrix is N × N . Deviation arises specially for dmax = 2.5 and dmax = 3.5.
As in the elastostatic case, Fig. 8d shows clearly the convergence with h-refinement of relative computational
cost to the range rt = 1–1.1 for all the dmax values.

4.3 Timoshenko beam elastostatic analysis

A key benchmark problem in computational mechanics and, in particular, in meshless literature is the next
application case. Timoshenko beam elastostatic analysis introduces interesting features such as a high-order
patch solution and the simultaneous enforcement of essential and natural boundary conditions. It is often used
to test the convergence rates of numerical implementations to be compared with FEM. A detailed study of a
Timoshenko beam with EFGM can be found in Belytschko et al. [1], Dolbow and Belytschko [15], Liu [5]
or Liu and Gu [17]. The analytical solution was obtained by Timoshenko and Goodier [22]. The numerical
accuracy is measured by means of the norm of error in strain energy εe as
 
1
εe = log10 εT C ε dΩ , (38)
2 Ω

where ε = ε num − ε exact , εnum being the tensor of numerical strains, ε exact the tensor of exact strains, and
C the characteristic matrix of the material.
The discrete nodal essential boundary conditions [15] of this two-dimensional case allow the direct enforce-
ment when using the interpolating shape functions resulting from introduction of matrix S. The results of
experimentation are presented next. Plane stress is considered. The ratio of computational cost rt is defined as
in the previous cases. Again, several values of the parameter dmax are tested.
Regarding the computational error, the conclusions are similar to the ones in the two first cases. Figure 10a
gathers the norm of error in strain energy for dmax = 2.0, showing that the interpolating MLS approximation
preserves the original rate of convergence [15]. Figure 10b reveals that large problems require higher values
 D. Garijo et al.

Fig. 11 Timoshenko beam elastostatic analysis, a condition numbers of MLS and IMLS approximations; b εe vs CPU time for
multiple nodal patters (up to 1,394 dof)

Table 2 Results for Timoshenko beam elastostatic analysis

Nodes dmax Interpolating MLS Standard MLS εe (%) rt

Norm εe t(s) Norm εe t(s)
8×2 2.0 −0.8092917991 0.1560 −0.8092917991 0.1400 0 1.11
2.5 −0.8362705660 0.1710 −0.8362705660 0.1870 0 0.91
4.0 −0.8472977326 0.2650 −0.8472977326 0.2490 0 1.06
16 × 4 2.0 −1.2918752866 0.8110 −1.2918752866 0.7950 0 1.02
2.5 −1.2158757469 1.1540 −1.2158757469 1.1700 0 0.99
4.0 −1.3077953523 2.6360 −1.3077953532 2.6210 −6.88E−08 1.01

of the parameter dmax to minimize the impact of transformation in machine time. In those cases, the parameter
rt remains below 1.1. Note that dmax ∼ 2–4 is recommended for static analysis [15].
The size of the matrix S is increased up to 2N x2N for two-dimensional problems. Figure 11a shows that
the condition number of the IMLS approximation with direct imposition of boundary conditions increases
as the size of the problem becomes larger. However, experimentation up to 3,500 degrees-of-freedom yields
κ I M L Sred < κMLS and κIMLS /κMLS ∼ 1. This result suggests that the conditioning of the problem is not
worsened by the introduction of matrix S in the formulation.
Table 2 compiles additional results for several nodal densities and values of dmax , showing that the original
performance of the MLS shape functions is not modified in terms of accuracy and convergence rates by the
transformation matrix S.

4.4 Euler–Bernoulli beams subjected to uniform distributed load

Another benchmark problem to test the applicability of the IMLS shape functions is a two-dimensional approach
to the Euler–Bernoulli beam subjected to uniform distributed load. Two support scenarios are studied: beam
with simply supported ends (SS) and beam with clamped ends (CC). In the former case, the hinged end
conditions are represented by enforcing displacements equal to zero at the two nodes of the neutral axis of the
beam located at the extremes x = 0 and x = L. Thus, the problem is constrained at discrete nodes and the
essential boundary conditions can be imposed directly when using IMLS shape functions. In the latter case, the
essential boundary conditions are applied along the boundaries Γu ; thus, the imposition of zero displacement
at nodes xi ∈ Γu does not suffice to constrain the problem correctly. The use of interpolating MLS shape
functions with direct enforcement of boundary conditions at nodes leads to a global error of the approximation
and to a local error at boundaries Γu , which can be measured by means of expression (31). One of the objectives
of this test is to find a correlation between both errors.
The exact solution for the deflection curves can be found in Timoshenko [23]:
q  3 
yexact (x) = − L x − 2L x 3 + x 4 (SS), (39)
24E I
Global interpolating MLS shape functions

Fig. 12 Deformation of beam with simply supported ends, interpolating MLS shape functions, a 21 × 3 nodes, dmax = 4.0;
b 43 × 5 nodes, dmax = 2.0

Fig. 13 Deformation of beam with clamped ends, interpolating MLS shape functions, a 21 × 3 nodes, dmax = 4.0; b 43 ×
5 nodes, dmax = 2.0

Fig. 14 Rates of convergence in displacement error for beam with simply supported ends, interpolating MLS shape functions,
a ν = 0.0; b ν = 0.5

q  2 2 
yexact (x) = − L x − 2L x 3 + x 4 (CC), (40)
24E I
q being the distributed load, L the beam span, E the material’s Young’s modulus, and I the inertia of the beam
cross section.
As a representative value of the global accuracy, it is taken the error of span mid-point deflection:
yIMLS (x = L/2) − yexact (x = L/2)
error = (41)
yexact (x = L/2)
Figures 12 and 13 show the deformation of the beams with two different nodal densities. The convergence
analyses are provided in Figs. 14 and 15, in which the error (41) is plotted versus the longitudinal distance
between nodes h, keeping five nodes in transversal direction. Results for two values of Poisson’s ratio ν are
presented. Table 3 compares numerical and exact values of the deflection of the mid-point of the span for
several beam slenderness ratios, nodal discretizations and parameters dmax for the SS case.
Figure 16 illustrates the deflection of the neutral axis of the SS beam, comparing IMLS and exact results for
two different dmax values. In Fig. 16a, it is noticed that the behaviour of the IMLS approximation with dmax =
1.0 resembles the locking effect of FEM. When dmax = 1.0, matrix S is the identity matrix and the interpolating
functions ψi become the bilinear shape functions of FEM. Using the same nodal density, in Fig. 16b the IMLS
approximation restores the expected accuracy by setting dmax = 2.0.
In Fig. 17, the deflection of the neutral axis of the CC beam is represented, comparing IMLS and exact
results for different nodal densities and dmax values.
 D. Garijo et al.

Fig. 15 Rates of convergence in displacement error for beam with clamped ends, interpolating MLS shape functions, a ν = 0.0;
b ν = 0.5

Table 3 Analysis results for beam with simply supported ends

Slenderness Nodes dmax = 2.0 dmax = 4.0 dmax = 6.0 Exact

L/D = 10 5×3 −2.622265E−4 −2.586300E−4 −2.422231E−4 −2.604167E−4
11 × 3 −2.638504E−4 −2.646242E−4 −2.646600E−4 −2.604167E−4
21 × 3 −2.640727E−4 −2.647550E−4 −2.646901E−4 −2.604167E−4
5×5 −2.637252E−4 −2.610165E−4 −2.428043E−4 −2.604167E−4
11 × 5 −2.651634E−4 −2.655986E−4 −2.655949E−4 −2.604167E−4
21 × 5 −2.654030E−4 −2.657304E−4 −2.657507E−4 −2.604167E−4
L/D = 50 5×3 −1.479676E−1 −1.136164E−1 −1.254811E−1 −1.627604E−1
11 × 3 −1.622673E−1 −1.626817E−1 −1.628322E−1 −1.627604E−1
21 × 3 −1.624613E−1 −1.628504E−1 −1.628549E−1 −1.627604E−1
5×5 −1.486637E−1 −1.192865E−1 −1.254767E−1 −1.627604E−1
11 × 5 −1.626041E−1 −1.626936E−1 −1.628470E−1 −1.627604E−1
21 × 5 −1.627882E−1 −1.628766E−1 −1.628786E−1 −1.627604E−1
L/D = 100 5×3 −1.869743E+0 −1.225595E+0 −1.967088E+0 −2.604167E+0
11 × 3 −2.590232E+0 −2.597046E+0 −2.600748E+0 −2.604167E+0
21 × 3 −2.597956E+0 −2.604143E+0 −2.604301E+0 −2.604167E+0
5×5 −1.885907E+0 −1.206698E+0 −1.957164E+0 −2.604167E+0
11 × 5 −2.595518E+0 −2.596061E+0 −2.600476E+0 −2.604167E+0
21 × 5 −2.602962E+0 −2.604343E+0 −2.604136E+0 −2.604167E+0

Fig. 16 Deflection of neutral axis of beam with simply supported ends, interpolating MLS shape functions, a 11 × 5 nodes,
dmax = 1.0; b 11 × 5 nodes, dmax = 2.0

The local error at boundaries Γu (31) is plotted versus the deflection error (41) in its non-relative form
in Fig. 18. In most of the curves, a close-to-linear correlation holds between both quantities and both are
Global interpolating MLS shape functions

Fig. 17 Deflection of neutral axis of beam with clamped ends, interpolating MLS shape functions, a 11 × 5 nodes, dmax = 2.0;
b 61 × 3 nodes, dmax = 6.0

Fig. 18 Error εΓu versus deflection error at CC beam mid-point, a 3 nodes in transversal direction, b 5 nodes in transversal
direction

Fig. 19 Beams with simply supported and clamped ends, elastostatic analysis, a condition numbers of IMLS approximation;
b IMLS deflection error vs degrees-of-freedom for multiple nodal patters (up to 1,414 dof)

progressively reduced with h-refinement. The order of magnitude of the local error is two or three times lower
than the deflection error at mid-point. In general terms, the interpolating approximation yields acceptable
accuracy and convergence rates for this problem with non-discrete essential boundary conditions.
 D. Garijo et al.

The evolution of the condition numbers for the SS and CC problems solved with IMLS approximation is
provided in Fig. 19a. Two runs with different number of nodes in transversal direction are plotted: three nodes
(black) and seven nodes (blue). Nodal patterns up to 1,414 degrees-of-freedom are tested. The associated decay
of the deflection error with the nodal density is presented in Fig. 19b. The results suggest that the interpolating
shape functions with direct imposition of boundary conditions do not lead to poorly conditioned problems,
even for high nodal densities. Besides, the penalty in the accuracy of the IMLS approximation when solving
the CC beam problem due to the incorrect imposition of distributed boundary conditions is small.

4.5 Cantilever beam free vibration analysis

In this section, interpolating MLS shape functions are tested in a two-dimensional approach to the cantilever
beam free vibration problem. As in the previous CC case, the use of IMLS shape functions with direct
enforcement of boundary conditions—system (30)—instead of the classic MLS implementation (16) implies
an incorrect procedure with an intrinsic numerical error. Only if this associated error is negligible, the simpler
IMLS programming can be accepted as a competitive advantage over the standard MLS.
Modal analysis of beams with different support conditions by means of EFGM has been widely treated in
the literature, for instance, see the works published by Liu [5], Tiago and Leitão [19] or Wang et al. [24]. The
exact natural frequencies are given by the expression [25]:
!
EI
ωn = (an L) 2
, n = 1, 2 . . . , (42)
m L4

Fig. 20 Cantilever beam free vibration analysis, standard and interpolating MLS shape functions, a rates of convergence in E 1 ;
b rates of convergence in E 2 ; c rates of convergence in E 3 ; d ratio of computational cost rt
Global interpolating MLS shape functions

Table 4 Results for cantilever beam free vibration analysis

Frequency Nodes IMLS MLS Exact

dmax = 2.0 dmax = 3.0 dmax = 2.0 dmax = 3.0
ω1 11 × 3 1.23535 1.16365 1.23282 1.15827 1.11186
31 × 3 1.13068 1.11798 1.12939 1.11602 1.11186
51 × 3 1.11974 1.11457 1.11873 1.11336 1.11186
11 × 5 1.23404 1.13921 1.23398 1.13870 1.11186
31 × 5 1.12923 1.11567 1.12861 1.11538 1.11186
51 × 5 1.11832 1.11348 1.11806 1.11340 1.11186
ω2 11 × 3 7.80481 7.26006 7.78826 7.22743 6.96792
31 × 3 7.08348 7.00136 7.07537 6.98925 6.96792
51 × 3 7.01369 6.98081 7.00776 6.97334 6.96792
11 × 5 7.79830 7.14927 7.79203 7.14558 6.96792
31 × 5 7.07437 6.98796 7.07181 6.98630 6.96792
51 × 5 7.00466 6.97406 7.00270 6.97311 6.96792
ω3 11 × 3 22.24841 20.32213 22.19901 20.23513 19.51037
31 × 3 19.82112 19.58535 19.79865 19.55218 19.51037
51 × 3 19.62204 19.52901 19.60557 19.50841 19.51037
11 × 5 22.23189 20.08626 22.21323 20.07509 19.51037
31 × 5 19.79494 19.54980 19.78760 19.54540 19.51037
51 × 5 19.59581 19.50964 19.59072 19.50682 19.51037

where E is material’s Young’s modulus, L is the beam length, I is the beam inertia, m is the beam mass, and
(an L) is the nth root of the characteristic equation

cos (an L) · cosh (an L) + 1 = 0. (43)

The numerical accuracy is again measured via the norm of error E n :

" "
"ωnum − ωexact "
n n
E n = log10 (44)
ωnexact

and the ratio of computational cost rt is used to compare the performances of MLS and IMLS.
The rates of convergence in E n obtained by MLS and IMLS for the first three natural frequencies are
provided in Fig. 20a–c, h being the longitudinal distance between nodes, while the ratios rt are presented in
Fig. 20d. The deviations in accuracy and convergence rates between MLS and IMLS are very small for this
problem. The ratio of computational cost is typically in the range 1.0–1.1; therefore, it can be assumed that
the additional 10 % of IMLS machine time can be counterbalanced by its simpler code programming. Table 4
compiles a complete set of results for the first three natural frequencies, comparing numerical and exact results
for ωi , for several nodal discretizations and parameters dmax .

4.6 Hole in infinite plate subjected to uniform axial stress

The next benchmark case is a circular hole in an infinite plate subjected to unidirectional far-field stress. The
problem is posed following Belytschko et al. [1] and Zhu and Atluri [3] methodology, that is, representing only
the upper right quadrant and imposing symmetry boundary conditions at the left and bottom edges. Thus, this
analysis deals with the discretization of the hole geometry, distributed essential boundary conditions and the
generation of IMLS approximation from non-equi-spaced nodal distribution (see Fig. 21a), leading to a matrix
S of variable band width. The exact solution for unitary far-field stress (Timoshenko and Goodier [22]):
# $
a2 3 3a 4
σx (x, y) = 1 − 2 cos2θ + cos4θ + 4 cos4θ, (45)
r 2 2r
# $
a2 1 3a 4
σ y (x, y) = − 2 cos2θ − cos4θ − 4 cos4θ, (46)
r 2 2r
# $
a2 1 3a 4
τx y (x, y) = − 2 sin2θ + sin4θ + 4 sin4θ, (47)
r 2 2r
 D. Garijo et al.

Fig. 21 Hole in an infinite plate problem, a 8 × 9 nodes discretization; b hoop stress at θ = π/2, 143 nodes

Fig. 22 Hole in an infinite plate analysis, rates of convergence in εe , a 13 circumferential nodes; b 15 circumferential
nodes

Fig. 23 Hole in an infinite plate analysis, a condition numbers of MLS and IMLS approximations; b εe versus degrees-of-freedom

where a is the hole radius and r and θ are polar coordinates, is employed to derive the applied stresses at the
top and right edges. The numerical accuracy is measured by means of the norm of error in strain energy (38).
The problem is solved using an MLS approximation with rectangular support domains and cubic spline
weight functions, instead of the circular domains and exponential weight functions selected by Belytschko et
al. [1] and Zhu and Atluri [3]. The size of the support domains is controlled by means of the parameter αs ,
Global interpolating MLS shape functions

Fig. 24 Edge-cracked panel analysis, a rate of convergence in SIF, IMLS Ω J 1 ; b rate of convergence in SIF, IMLS Ω J 2 ;
c condition numbers of MLS and IMLS approximations; d ratio of computational cost rt

which is a fraction of the maximum vertical and horizontal distance between nodes at the upper right corner
of the domain.
Values of E = 103 and ν = 0.3 are used. The hoop stress obtained at θ = π/2 with a discretization
of 143 nodes is shown in Fig. 21b, note that the IMLS approximation yields incorrect results at boundaries
Γu , although in terms of the global norm of error εe the results are acceptable compared with those of the
standard MLS. The convergence analysis is given in Fig. 22, h being a measurement of the nodal density
in radial direction. The comparison of MLS and IMLS condition numbers is provided in Fig. 23a, with the
corresponding decay of norm εe as the number of degrees-of-freedom grows shown in Fig. 23b.
The presented numerical experimentation provides satisfactory results when applying the global IMLS
procedure to a problem defined in a non-rectangular domain with a discretization of non-equi-spaced nodes.

4.7 Edge-cracked panel subjected to loading mode I

Finally, the IMLS approximation is tested in the evaluation of the stress intensity factor (SIF) corresponding to
an edge-cracked panel under loading mode I. This problem is another benchmark of meshless literature and,
in particular, of EFGM analysis (Belytschko et al. [1]; Rao and Rahman [26]).
The main interest of this test is the study of the behaviour of the IMLS approximation when the local
support domains of the original MLS shape functions are modified according to the visibility criterion [16] to
represent the material discontinuity generated by the crack. The calculation of the SIF is implemented via J
integral [27] computed through an equivalent domain integral method (EDI) procedure [28].
The reference solution for a crack of length a in a finite panel of height H under far-field stress σ f f in
linear elastic fracture mechanics (LEFM) regime is given by the following expression that can be found, for
 D. Garijo et al.

Fig. 25 Edge crack problem, components of stress tensor obtained with IMLS approximation of 49 × 31 nodes and dmax = 1.7,
a σx x ; b σ yy

Table 5 Results for edge crack problem

ΩJ Nodes dmax K IMLS K Iexact εMLS (%) K IIMLS K Iexact εIMLS (%)
1 9×6 1.5 2233.322 1712.856 30.39 2233.322 1712.856 30.39
25 × 16 1.5 1725.329 1712.856 0.73 1725.329 1712.856 0.73
49 × 31 1.5 1708.300 1712.856 0.27 1708.299 1712.856 0.27
9×6 2.0 2571.424 1712.856 50.12 2571.424 1712.856 50.12
25 × 16 2.0 1911.708 1712.856 11.61 1911.708 1712.856 11.61
49 × 31 2.0 1745.938 1712.856 1.93 1745.938 1712.856 1.93
2 9×6 1.5 2022.183 1712.856 18.06 2022.183 1712.856 18.06
25 × 16 1.5 1720.465 1712.856 0.44 1720.465 1712.856 0.44
49 × 31 1.5 1709.457 1712.856 0.20 1709.457 1712.856 0.20
9×6 2.0 2375.059 1712.856 38.66 2375.059 1712.856 38.66
25 × 16 2.0 1893.038 1712.856 10.52 1893.038 1712.856 10.52
49 × 31 2.0 1741.338 1712.856 1.66 1741.338 1712.856 1.66

instance, in Anderson [29]:

√ a
K Iexact = σ f f H f ,
% H
a πa    (48)
2tan 2H a πa 3
f = πa 0.752 + 2.02 + 0.37 1 − sin .
H cos 2H H 2H

A domain Ω J bounded by rectangular contours surrounding the crack tip is selected for the EDI implemen-
tation, with a q1 plateau function [29] based on the cubic spline function [16]. The panel is constrained only at
two nodes laying on the crack vertical symmetry axis in order to prevent rigid body displacement; therefore, the
essential boundary conditions are discrete and no deviations in terms of numerical error are expected between
MLS and IMLS approximations. Numerical results confirm the lack of penalty due to direct enforcement of
boundary conditions. Large distributions of equi-spaced nodes (homogeneous nodal density in the whole panel
is used instead of nodal concentration around crack tip) are experimented to check convergence and condition
numbers.
Numerical results obtained with the MLS and IMLS approximations show moderate dependence on
the size of" the EDI domain " Ω J . In Fig. 24a, b, diagrams with the rates of convergence in SIF error
εIMLS = " K IIMLS − K Iexact " /K Iexact for two different sizes of Ω J are provided. Low values of dmax yield
better performances for the nodal arrangements experimented. Figure 24c illustrates the evolution of the con-
dition numbers of MLS and IMLS approximations up to high nodal densities. Figure 24d assesses the penalty
in CPU time due to S matrix transformation, showing increasing ratios with the h-refinement, even though
high dmax values alleviate the relative increase in computational effort (for instance, rt ∼ 1.17 for dmax = 3.0
and 49 × 31 nodes).
Global interpolating MLS shape functions

The components σx x and σ yy of the stress field around the crack tip computed with the IMLS technique
using 49 × 31 nodes and dmax = 1.7 are plotted in Fig. 25. Additional results comparing the accuracy of the
interpolating and non-interpolating approximations are gathered in Table 5.

5 Conclusions

The main disadvantage of MLS shape functions is the need of strategies such as the Penalty method, Nitsche’s
method or the Lagrange multipliers to enforce the essential boundary conditions or to couple the approximation
with FEM-type schemes. These strategies are powerful resources and provide excellent results; however,
their implementation is usually penalized since specific programming codes are required, with the associated
investment of time in the pre-analysis stage.
In the last years, remarkable advances in the study of the MLS functions to correct their non-interpolating
nature have led to modified MLS approximations such as IMLS, IEFGM and WNLS. In this context, a numerical
investigation of a later transformation of the mass and stiffness matrices in structural analysis is presented in
this research.
This transformation, which avoids the introduction of singular weight functions, arranges the formulation
in terms of actual nodal displacements and generates interpolating MLS shape functions which are not locally
supported but globally extended to the whole domain of analysis. Its use is limited to problems with discrete
prescribed displacements at boundaries, for which a straightforward FEM-type imposition of boundary condi-
tions is possible, otherwise the resulting structural constraint is incorrect. When inserted into an element-free
Galerkin implementation, experiments in classic structural benchmarks with discrete and non-discrete bound-
ary conditions show that the numerical accuracy and convergence of the global IMLS approach are practically
the same as those of the standard MLS with boundary displacements imposed by means of Lagrange multipli-
ers. The relative computational cost of both approximations remains in the range of 1.0–1.3 and, at least for
the structural examples presented, the condition numbers of the full matrices of IMLS approximation imply
problems sufficiently well conditioned when compared with standard MLS.
Further research of the IMLS implementation in problems with arbitrary nodal distributions, complex
geometry and different choice of parameters of the MLS approximation (enrichment of intrinsic basis, weight
function, etc) is proposed to support the results presented in this work.

Acknowledgments The authors wish to thank the support and collaboration of Safran Group, FACC AG and Airbus Defence
and Space.

References

1. Belytschko, T., Lu, Y.Y., Gu, L.: Element-free Galerkin methods. Int. J. Numer. Methods Eng. 37, 229–256 (1994)
2. Lancaster, P., Salkauskas, K.: Surfaces generated by moving least squares methods. Math. Comput. 37, 141–158 (1981)
3. Zhu, T., Atluri, S.N.: A modified collocation method and a penalty formulation for enforcing the essential boundary conditions
in the element free Galerkin method. Comput. Mech. 21, 211–222 (1998)
4. Fernández-Méndez, S., Huerta, A.: Imposing essential boundary conditions in mesh-free methods. Comput. Methods Appl.
Mech. Eng. 37(12), 1257–1275 (2004)
5. Liu, G.R.: Mesh Free Methods. Moving Beyond the Finite Element Method. CRC Press, Boca Raton (2003)
6. Kaljevic, I., Saigal, S.: An improved element free Galerkin formulation. Int. J. Numer. Methods Eng. 40(16), 2953–
2974 (1997)
7. Chen, J.-S., Wang, H.-P.: New boundary condition treatments in meshfree computation of contact problems. Comput.
Methods Appl. Mech. Eng. 187, 441–468 (2000)
8. Ren, H., Cheng, Y.: The interpolating element-free Galerkin (IEFG) method for two-dimensional elasticity problems. Int. J.
Appl. Mech. 3(4), 735–758 (2011)
9. Ren, H., Cheng, Y.: The interpolating element-free Galerkin (IEFG) method for two-dimensional potential problems. Eng.
Anal. Bound. Elem. 36(5), 873–880 (2012)
10. Most, T., Bucher, C., Macke, M.: A natural neighbor based moving least squares approach with interpolating weighting
function. In: Gürlebeck, K., Könke, C. (eds.) 17th International Conference on the Application of Computer Science and
Mathematics in Architecture and Civil Engineering, pp. 12–14. Weimar, Germany (2006)
11. Most, T., Bucher, C.: New concepts for moving least squares: an interpolating non-singular weighting function and weighted
nodal least squares. Eng. Anal. Bound. Elem. 32, 461–470 (2008)
12. Chen, J.-S., Pan, C., Wu, C.-T., Liu, W.K.: Reproducing Kernel Particle Methods for large deformation analysis of non-linear
structures. Comput. Methods Appl. Mech. Eng. 139, 195–227 (1996)
13. Chen, J.-S., Pan, C., Wu, C.-T.: Large deformation analysis of rubber based on a reproducing kernel particle method. Comput.
Mech. 19, 211–227 (1997)
 D. Garijo et al.

14. Wagner, G.J., Liu, W.K.: Application of essential boundary conditions in mesh-free methods: a corrected collocation
method. Int. J. Numer. Meth. Eng. 47(8), 1367–1379 (2000)
15. Dolbow, J., Belytschko, T.: An introduction to programming the meshless element free Galerkin method. Arch. Comput.
Methods Eng. 5(3), 207–241 (1998)
16. Belytschko, T., Krongauz, Y., Organ, D., Fleming, M., Krysl, P.: Meshless methods: an overview and recent develop-
ments. Comput. Methods Appl. Mech. Eng. 139(1–4), 3–47 (1996)
17. Liu, G.R., Gu, Y.T.: An Introduction to Meshfree Methods and Their Programming. Springer, Berlin (2005)
18. Valencia, Ó.F., Gómez-Escalonilla, F.J., López, J.: Influence of selectable parameters in EFGM. 1d bar axially loaded
problem. Proceedings IMechE, Part C: J. Mech. Eng. Sci. 222(11), 1621–1633 (2008)
19. Tiago, C.M., Leitão, V.M.: Analysis of free vibration problems with the element-free Galerkin method. In: 11th International
Conference on Numerical Methods in Continuum Mechanics, Žilina, Slovak Republic (2003)
20. Valencia, Ó.F., Gómez-Escalonilla, F.J., Garijo, D., Díez, J.L.: Bernstein polynomials in EFGM. Proceedings IMechE, Part
C: J. Mech. Eng. Sci. 225(8), 1808–1815 (2011)
21. Timoshenko, S.: Vibration Problems in Engineering, 2nd edn. D. Van Nostrand Company, Inc., New York (1973)
22. Timoshenko, S., Goodier, J.N.: Theory of Elasticity. McGraw Hill, New York (1951)
23. Timoshenko, S.: Strength of Materials, Part I: Elementary Theory and Problems, 2nd edn. D. Van Nostrand Company,
Inc., New York (1940)
24. Wang, Y.H., Li, W.D., Tham, L.G., Lee, P.K.K., Yue, Z.Q.: Parametric study for an efficient meshless method in vibration
analysis. J. Sound Vib. 255(2), 261–279 (2002)
25. Paz, M.: Structural Dynamics, Theory and Computation, 3rd edn. Van Nostrand Reinhold Company, Inc., New York (1991)
26. Rao, B.N., Rahman, S.: A coupled meshless-finite element method for fracture analysis of cracks. Int. J. Press. Vessels
Pip. 78, 647–657 (2001)
27. Rice, J.R.: A path independent integral and the approximate analysis of strain concentration by notches and cracks. J. Appl.
Mech. ASME 35, 379–386 (1968)
28. de Lorenzi, H.G.: On the energy release rate and the J -integral for 3-D crack configurations. Int. J. Fract. 19(3), 183–
193 (1982)
29. Anderson, T.L.: Fracture Mechanics Fundamentals and Applications, 2nd edn. CRC Press LLC, Boca Raton (1995)