Computers and Structures 182 (2017) 368–391

Contents lists available at ScienceDirect

Computers and Structures

journal homepage: www.elsevier.com/locate/compstruc

A meshfree unit-cell method for effective planar analysis of cellular

beams
A.R. Zainal Abidin a,b, B.A. Izzuddin a,⇑, F. Lancaster a,c
a
Department of Civil and Environmental Engineering, Imperial College, London SW7 2AZ, United Kingdom
b
Faculty of Civil Engineering, Universiti Teknologi Malaysia, 81310 Skudai, Johor, Malaysia
c
Eckersley O’Callaghan, 137 Varick Street, #406, New York, NY 10013, United States

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a novel approach for accurate and efficient planar response analysis of cellular beams,
Received 12 April 2016 which provides the necessary input for local out-of-plane buckling analysis of web components. The pro-
Accepted 1 September 2016 posed approach utilises the super-element concept defined for unit-cells, achieving further efficiencies
Available online 9 January 2017
through an enhanced Element-Free Galerkin (EFG) approach for establishing the planar super-element
response. Several examples are presented, firstly at the level of unit-cells, where the computational ben-
Keywords: efits of the EFG method are highlighted, and finally at the overall level of cellular beams, where the supe-
Meshfree methods
rior performance of the unit-cell approach with virtually no compromise in accuracy is demonstrated.
Element-Free Galerkin method
Planar response
Ó 2016 Elsevier Ltd. All rights reserved.
Cellular beams
Unit-cell formulation

1. Introduction Although some guidance on the design and assessment of cellu-

lar beams is provided in codes of practice [10,11], the selection and
Cellular beams have become in recent years a popular form of design of cellular beams is typically undertaken in practice with
steel construction, largely due to enabling large spans while facil- the aid of specialist software (e.g. [12]). However, in considering
itating the integration of services within the floor depth [1–3]. As the local buckling of web elements, simplified models were
illustrated in Fig. 1, cellular beams are easily fabricated from stan- adopted [13–17], utilising for example a strut buckling analogy
dard I-beams through appropriate cutting, offsetting and joining with empirical calibration against nonlinear finite element models
operations, enhancing the bending stiffness and resistance via an [18–20]. While the resulting simplified models are computation-
increased distance between the flanges, while reducing weight ally efficient, these are not rational and sufficiently general for
via the resulting openings in the web. application to a wide range of cellular beam geometric configura-
Besides the above relative structural benefits of cellular beams, tions, as they ignore both the nonlinear planar stress distribution
their structural response is very different in certain respects from and the two-dimensional plate buckling behaviour of the web
the parent I-beam [4]. The existence of regular holes influences components. Of course, accurate predictions can be achieved in
the structural behaviour particularly in the web region where local this respect using nonlinear finite element analysis, though the
elements, such as the web-posts and top/bottom tees, are subject computational and modelling demands of such an approach are
to local actions throughout the beam span. These actions consist still prohibitive for application in design practice.
of not only contributions to the overall cross-sectional shear and The present work is motivated by the need for an efficient and
bending moment, but are also superimposed by secondary effects practical modelling approach for predicting the nonlinear planar
leading to a highly nonlinear stress distribution over the cellular stress distribution in cellular beams, which could in turn be used
beam depth and length [5,6]. One of the most important and dis- for local buckling assessment of the web components in a separate
tinctive structural characteristics of cellular beams relates to local model of the nonlinear out-of-plane response. Focusing on elastic
buckling of the web elements [6–9], which in turn depends on the local buckling as a performance consideration that may be used
nonlinear distribution of planar stresses in the perforated web. along with material strength limit for cellular beam design, this
paper is mainly concerned with the accurate and efficient determi-
nation of the nonlinear planar stress distribution in cellular beams
⇑ Corresponding author.
arising under planar loads within the linear elastic range. The
E-mail address: [email protected] (B.A. Izzuddin).

http://dx.doi.org/10.1016/j.compstruc.2016.09.002
0045-7949/Ó 2016 Elsevier Ltd. All rights reserved.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 369

Fig. 1. Typical profiling of cellular beams.

assessment of local buckling could then be undertaken considering nature of the holes over the beam length, thus introducing the
the out-of-plane geometric stiffness associated with the predicted notion of a unit-cell as a ‘super-element’ which interacts with adja-
planar stress distribution as well as the out-of-plane material stiff- cent unit-cells via a reduced number of DOF, as illustrated in Fig. 2.
ness, though this aspect is not covered in the present paper but is Towards this end, the EFG method is adopted and suitably
referred to previous work by the authors on beams with irregular enhanced for planar analysis at the unit-cell level, while the overall
openings [21]. response of the cellular beam is efficiently obtained via the assem-
For an accurate prediction of the nonlinear planar stress distri- bly of super-element contributions.
bution in linear elastic cellular beams, the most obvious approach The paper proceeds with presenting the numerical discretisa-
is to utilise finite element analysis, though this can still be compu- tion of a unit-cell using the EFG method, highlighting the enhance-
tationally demanding for typical cellular beams, even in the linear ments undertaken for application to cellular beams. The
elastic range. Meshfree methods have recently emerged as an formulation of the super-element associated with the unit-cell is
alternative approach [22–24] with potentially significant relative then provided, including the treatment of edge unit-cells and the
benefits, particularly where geometric features complicate the assembly of the overall cellular beam response. Finally, several
meshing with finite elements, as in the case of beams with open- examples are presented, ranging from the level of an individual
ings. Moreover, in such cases, the meshfree approach has the unit-cell to a whole cellular beam, which demonstrate the accuracy
potential to achieve significant computational savings with a con- and computational benefits of the proposed approach towards pre-
siderable reduction in the number of degrees of freedom (DOF) for dicting the planar response of cellular beams.
comparable accuracy to finite element analysis.
Amongst the various meshfree methods, the Element-Free 2. Numerical discretisation with EFG method
Galerkin (EFG) method [25] and the meshless local Petrov-
Galerkin (MLPG) method [26] have been considered for the present The Element-Free Galerkin (EFG) method is used to discretise
work. In contrast with the MLPG method which is claimed as ‘truly the problem domain at the unit cell level. The main feature of
meshless’, the EFG method requires ‘background cells’ for numeri- the EFG method is the utilisation of a smooth and continuous
cal integration of the governing equation over the problem domain. domain function via the application of a moving least squares
While this may be seen as a shortcoming for the EFG method, it is (MLS) approach. As this MLS approximation requires only a set of
actually a benefit in the present context, since the evaluation of nodes to generate shape functions, it is acknowledged as an
planar stresses at a fixed set of integration points, rather than a element-free method which offers significant savings on
variable set associated with the testing function as in the MLPG computational effort, since a major cost of element meshing is
method, is required for determining the out-of-plane geometric eliminated [25]. However, such benefit comes with some addi-
stiffness in subsequent local buckling assessment. Furthermore, tional costs, especially in relation to the imposition of boundary
the MLPG method suffers from the fact that rigid body testing conditions, since the MLS function does not represent the actual
modes are not accurately represented, leading to only an approxi- nodal values.
mate satisfaction of equilibrium between the applied loading and The EFG method is constructed on the basis of a Galerkin formu-
the reactions at the support boundaries. Accordingly, the EFG lation with the adoption of moving least-squares approximation to
method is selected with due consideration of its following benefits: produce the shape functions, as presented in detail by [25]. To view
(i) it is a meshless method that can be easily applied to complex the application of EFG method in 2D solid mechanic problems, a
geometric domains, (ii) it facilitates numerical integration via the domain of X with a boundary C, being in a state of equilibrium,
use of fixed integration points over the domain, and (iii) it ensures is considered:
external equilibrium at sub-domain level between loading and
boundary actions. LTd r þ b ¼ 0 in X ð1Þ
Besides the selection of the EFG approach, this work capitalises
on a major characteristic of cellular beams, which is the repetitive with boundary conditions specified as:
370 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

Unit cell

Super-element
nodes

Fig. 2. Unit-based discretisation of cellular beam.

u¼u
on Cu On the other hand, aj(x) is an arbitrary set of coefficients that needs
ð2Þ
Ln r ¼
t on Ct to be solved, and yet it should not be taken as a constant for accu-
racy reason [25]. Therefore, in solving for the unknown coefficients,
in which r is the stress vector corresponding to the displacement a functional of weighted residuals between the actual nodal param-
fields, Ld is a matrix of differential operators for equilibrium, Ln is eters (uI) and their approximated values, uh(xI), is used:
a matrix of the unit normal to the boundary C, and b is the body

and
t are the prescribed boundary dis- X
N
force of the domain X. u D¼ wðx  xI Þ½uh ðxI Þ  uI 
2
ð8Þ
placements and tractions respectively. The well-known Galerkin I¼1
weak form of the equilibrium Eq. (1) is given as:
Z Z Z where N is the number of local nodes, w(x  xI) is a built-in weight
deT  r dX  duT  b dX  duT  t C dC ¼ 0 ð8duÞ ð3Þ function that determines the influence of the residuals at different
X X Ct nodes (xI) around the focus point (x).
or in the form of the displacements function u: Considering Eqs. (6) and (8), the selection of a(x) that minimises
Z Z Z the residual functional can be obtained from the condition of sta-
ðrs duÞT  D  ðrs uÞ dX  duT  b dX  duT  t C dC ¼ 0 ð8duÞ tionary D with respect to a(x), which leads to the following solu-
X X Ct tion of a linear equation:
ð4Þ
AðxÞaðxÞ ¼ BðxÞu ) aðxÞ ¼ ðAðxÞ1 BðxÞÞu ð9Þ
/
in which u is often defined in terms of shape functions (N ) and the
where
nodal DOFs (u):
X
N

u ¼ N/u ð5Þ AðxÞ ¼ wðx  xI Þ pðxI Þ pT ðxI Þ

I¼1
ð10Þ
Note that discretisation with the finite element method (FEM) uses BðxÞ ¼ ½ b1 ; . . . ; bN ; bI ¼ wðx  xI Þ pðxI Þ
an element-based approach to construct the pre-defined shape
Finally, by considering Eq. (9), the initial Eq. (6) can be rewritten in
functions, for which a well-generated mesh is necessary to obtain
terms of the nodal parameters (u) and their corresponding shape
a good approximation. This is always a major issue in the FEM par-
functions (/I ) as follows:
ticularly when dealing with irregular domains. The EFG method, on
the other hand, benefits from the MLS numerical discretisation, in X
m X
N X
m

which only a free form of distributed local nodes is required and uh ðxÞ ¼ pj ðxÞaj ðxÞ ¼ pj ðxÞðA1 ðxÞBðxÞÞjI uI
j¼1 I¼1 j¼1
completely exempted from any element connectivity.
X
N

2.1. Shape functions by MLS approximation ¼ /I ðxÞuI ¼ Uu ð11Þ

I¼1

The MLS approximation is a general technique of fitting a curve where u is a vector of nodal displacements (I = 1, 2, . . . , N) and U is
or surface to scattered data, which is commonly used in statistical the vector of shape functions:
analysis. In the EFG method, such a technique is enhanced to gen- X
m
erate a higher-order surface fit function of displacements uh(xI) /I ðxÞ ¼ pj ðxÞðA1 ðxÞBðxÞÞjI ð12Þ
based on nodal values (uI). This is accomplished by the use of a j¼1
polynomial-based expansion defined as:
It is important to note that uh(xI) – uI as a consequence of the
X
m
MLS interpolation, unless the prescribed data uI is based on a poly-
uh ðxÞ ¼ pj ðxÞaj ðxÞ  pT ðxÞaðxÞ ð6Þ
nomial function [27].
j¼1

in which pj(x) is the polynomial basis function with m number of 2.2. Choice of weight function
terms, and aj(x) features as the unknown coefficients which per-
form curve-fitting. Both terms are a function of the space coordi- In the EFG method, the weight function w(x  xI) play two
nates x ¼ ½x y. Considering Pascal’s triangle of monomials, pj(x) major roles: (i) as nodal weighting parameters and (ii) as a smooth-
simply reaches the desired degree of continuity by the employment ing factor on the overall MLS function. The selection of the shape of
of sufficient monomial terms. In this work, the quadratic basis is the support domain is almost arbitrary, but a circular shape is the
employed, hence m ¼ 6, as given by: most convenient choice. The support radius (rI) is used to control
the limit of the ‘domain of influence’ and also to avoid a singularity
pT ðxÞ ¼ ½1; x; y; x2 ; xy; y2  ð7Þ
issue of matrix A(x) in Eq. (10), in the sense that the support
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 371

domain has to be large enough to include a sufficient number of two planar displacement fields, and B/ is the strain-displacement
local nodes, but at the same time it needs to be as small as possible matrix, as given by:
so as to preserve the locality of the influence area.
In the present approach, an exponential-based Gaussian weight ð17Þ
function [25] is chosen which has the form:
8
< exp½ðdI =cÞ2k exp½ðrI =cÞ2k  ; 0 6 d 6 r
I I
wðx  xI Þ ¼ 1exp½ðr I =cÞ2k  ð13Þ
: 0; d >r ð18Þ
I I

where k and c are constants. As can be easily illustrated numeri-

cally, the parameter c is used to obtain a smooth connection Considering Eq. (18), the first-order derivatives of the shape
between the support domain and the external area (Fig. 3). In the functions are obtained as:
present study, c is typically taken as rI/4 as suggested by Atluri /I ðxÞ;x ¼ pðxÞ;x AðxÞ1 BðxÞ þ pðxÞðAðxÞ1 Þ;x BðxÞ
and Zhu [26], which provides the required smoothness condition
as shown in Fig. 3. Parameter k is always taken as 1 for computa- þ pðxÞAðxÞ1 BðxÞ;x ð19Þ
tional simplicity.
The first derivatives of the weight function with respect to x where
(with k = 1) is thus obtained as: ðAðxÞ1 Þ;x ¼ AðxÞ1 ðAðxÞÞ;x AðxÞ1 ð20Þ
8
<  2dI ðdI;x Þexp½ðdI =cÞ2  ; 0 6 d 6 r
I I and
wðx  xI Þ;x ¼ c2 ð1exp½ðrI =cÞ2 Þ ð14Þ
:
0; dI > r I X
N
ðAðxÞÞ;x ¼ wðx  xI Þ;x pT ðxI ÞpðxI Þ ð21Þ
with subscript ‘‘,x” representing differentiation with respect to the I¼1
spatial coordinates x.
In the above equation, w(x  xI),x is as given by Eq. (13), and p(xI)
represents the monomial basis functions evaluated at xI. For com-
2.3. Galerkin method with MLS interpolation putational efficiency, the derivatives in Eq. (19) can be obtained
without the explicit evaluation of the inverse of A(x) by applying,
Considering the displacement shape functions resulting from for example, Gaussian elimination or LU decomposition methods
MLS interpolation in Eq. (11), the variational weak form in Eq. (4) [28]. It is important to note that the Galerkin equation in (15) does
becomes: not deal with the essential boundary conditions of Eq. (2), which are
Z Z addressed in the following section.
T T
ðB/ duÞ  D  ðB/ uÞ dX  ðN / duÞ  b dX
X X
Z 2.4. Essential boundary conditions
T
 ðN / duÞ  t C dC ¼ 0 ð8duÞ ð15Þ
Ct
As previously noted, the EFG method cannot satisfy the essen-
where D is the constitutive matrix for the plane stress case given tial boundary conditions via nodal constraints, since the MLS func-
by: tion does not pass through the nodal parameters, and thus the
2 3 enforcement of the essential condition at the boundary nodes leads
1 t 0 to violation of the condition on the boundary edge (e.g. u–uC on
Et 6 7
D¼ 4t 1 0 5 ð16Þ Cu ). In the original EFG method by Belytschko et al. [25], additional
ð1  t2 Þ
0 0 ð1  tÞ=2 terms of Lagrange multipliers are employed in addressing essential
boundary conditions. Such an approach comes at the cost of addi-
with E being Young’s modulus, t being Poisson’s ratio and t being tional unknowns that enlarge the size of the discrete system of
the plate thickness. N / is the matrix of shape functions for the equations. It was also noted to be computationally inefficient

0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-1

-2
w(x-xi) for c = ri/1
w(x-xi) for c = ri/4
-3 discontinuity
w'(x-xi) for c = ri/1
w'(x-xi) for c = ri/4
-4

Fig. 3. Exponential weight function and its first-order derivatives for different c parameters (rI = 1 and k = 1).
372 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

owing to the destruction of the positive definiteness of the stiffness can be performed by ignoring all Gauss points outside the domain
matrix [24]. during the summation. This basically considers the domain inte-
In the current method, essential boundary conditions are grand function to retain its values within the domain X but
imposed through a penalty method, which is also often used in become zero outside X. Due to the discontinuity introduced in
other numerical methods, including conventional FEM. The penalty the function within the integration cells across the boundary, inac-
term comprises of the following variational form: curacy in the integration could occur if limited quadrature points
Z are used. More advanced methods include the adoption of triangu-
 dðu  uC ÞT  a  ðu  uC Þ dC ¼ 0 ð8duÞ ð22Þ lation integration schemes to fit an irregular domain as presented
Cu
in [29,30], particularly in the context of moving discontinuities as
where a is a diagonal matrix of penalty factors (ax, ay) and (u  uC) in the modelling of evolving cracks. Such methods require a more
is the residual condition to be penalised so that u  uC along the complex algorithm, including the modification of quadrature
boundary with essential conditions. This method does not enforce weights, so as to deal with an arbitrary integration domain that
the given condition fully unless a very large penalty terms are used changes with crack growth.
for a, though the specific choice should be made with great care to In this work, a simple and accurate approach for a planar anal-
avoid numerical problems arising from ill-conditioning. Liu [24] ysis problem is proposed based on a multi-level rectangular grid
suggested that the value of a should be between 104 and 1013 mul- (MLRG), in which the cells that intersect the boundary are hierar-
tiplied by the maximum of the diagonal terms in the stiffness chically reduced in size (e.g. by dividing the original cell side into
matrix. two). Ultimately, the cells that lie outside the boundary are not
Considering the penalty term, the updated form of Eq. (15) can considered, as illustrated in Fig. 4. This is nothing more than pre-
be written as: venting Gauss points from falling outside the domain so that no
Z Z elimination of Gauss points has to take place. Despite having a con-
T T
ðB/ duÞ  D  ðB/ uÞ dX  ðN / duÞ  b dX centration of quadrature points near the curved boundary, it is
X X
Z Z shown later that this technique is quite efficient, as only a small
T T
 ðN / duÞ  t C dC  ðN / duÞ  a  ðN / uÞ dC ð23Þ number of quadrature points per cell is required whilst preserving
Ct Cu the simplicity of having background cells based on a rectangular
Z
þ
T
ðN / duÞ  a  uC dC ¼ 0 ð8duÞ grid.
Cu Numerical integration in this study is performed using a Gaus-
sian quadrature scheme via a set of rectangular background cells.
or in a more generalised form of discrete linear equations:
In each cell, Ng  Ng Gauss quadrature points (xq) are employed,
Z Z 
which makes the total number of Gauss points in the domain
B/T DB/ dX þ wT aw dC u
X Cu MX = Nc  N2g, where Nc is the number of cells. With regard to line
Z Z Z integration, such as over the boundary, this is performed by parti-
T
¼ N t C dC þ N /T b dX þ
/T
N / auC dC ð8duÞ ð24Þ tioning the line into NL segments, each of which containing 9
Ct X Cu
quadrature points, hence MC = 9  NL. Having determined the
This can be symbolised in terms of stiffness matrix (k), penalty stiff- location of Gauss points (xq), their weights (wq) and the corre-
ness matrix (ka), displacements vector (u) and an equivalent load- sponding Jacobian (J e ), the components in Eq. (26) can be obtained
ing vector (f) to yield the following form: using Gaussian quadrature as:
a
½k þ k u ¼ f ð25Þ
where the associated terms are defined as:
X
MX
T
kIJ ¼ wq B/I ðxq ÞDB/J ðxq ÞJ e
Z q¼1
kIJ ¼ B/T /
I DBJ dX X
MX
ZX a
kIJ ¼ wq wTI ðxq ÞawJ ðxq ÞJ e
a
kIJ ¼ wTI awJ dC ð26Þ q¼1
Cu
Z Z Z X
MC
T X
MC
T X
MC
T

fI ¼
T T
N / t C dC þ N / b dX þ
T
N / auC dC fI ¼ wq N /I ðxq Þt C Je þ wq N /I ðxq ÞbJ e þ wq N /I ðxq ÞauC J e
Ct X Cu q¼1 q¼1 q¼1

a ð28Þ
with the components of kIJ defined as:
   
/I ðxÞ 0 ax 0
wI ¼ ; a¼ ð27Þ where kIJ is the 2  2 matrix component of the full 2N  2N matrix
0 /I ðxÞ 0 ay a
k, and the same applies to kij . N /I ðxq Þ and B/I ðxq Þ are, respectively,
the shape function and strain-displacement matrices of node I eval-
2.5. Background cells for integration uated at Gauss points xq. The term wI ðxq Þ is also a matrix of shape
functions which is given by Eq. (27).
One of the issues relating to the application of the EFG method
is the need for background cells for numerical integration, as a
result of which the method is claimed to be not a ‘truly meshless’
method [26]. Nevertheless, the application of such a background 3. Unit-cell formulation
‘mesh’ for integration is much more flexible than the FEM due to
the fact that the cells are only utilised for numerical integration, The planar response for a cellular beam is assembled from con-
thus a simple grid of rectangular shapes would be satisfactory. tributions of super-elements, each representing the characteristic
An important issue arising with the EFG method relates to inte- response of a unit-cell. The formulation of the unit-cell super-
gration over irregular domains, since a simple rectangular grid element, its contribution to the global cellular beam response,
without modifications cannot perfectly fit a domain with curved and the recovery of the planar stresses after determination of the
boundary. Belytschko et al. [25] proposed that irregular domains super-element nodal parameters are presented hereafter.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 373

3
1

* the numbers 1, 2 and 3 indicate the cells level

Fig. 4. Numerical integration for irregular domain with two hierarchic subdivisions.

3.1. Unit-cell level the web as well as flange regions. This allows the planar response
to be modelled using a 2D model in which the web and flange
A unit-cell refers to a portion between two adjacent openings of regions have different out-of-plane thicknesses, where the thick-
a cellular beam or between an opening and the beam end, as pre- ness for the flanges is taken equal to their width. The issue con-
viously illustrated in Fig. 2. The unit-cell response is approximated cerning material discontinuity between the web and flange
by means of a super-element utilising three DOF (ui vi hi) at the tee components, as pointed by Liu [24], can lead to discontinuity in
centroids, corresponding to four ‘super-element nodes’ for a single the transverse normal stress ryy and shear stress sxy, though this
super-element. As for the outer edges of the end cells, the cross- can still be approximated well with continuous MLS functions sub-
section of the top and bottom tees extends to include web areas ject to adequate transverse discretisation at these locations.
up to mid-depth of the cell. As shown in Fig. 5, the associated com- The applied tractions of fxi, fyi and mi along the side edges
ponents of internal forces, horizontal force fxi, vertical force fyi and account for changes in the thickness, as demonstrated in Fig. 6.
moment mi, are used to establish the characteristic super-element In the analysis, anti-clockwise moments are considered positive,
response as discussed later, and minimum restraint of rigid body and the action of vertical shear in the tee section is assumed to
movement is applied at three locations along the mid-depth of be effective only along the web area, which is realistic for relatively
the web-post. The unit cell is considered for illustrative purposes thin flanges.
under the action of a uniformly distributed load (UDL) q acting
along the top edge, which allows a step-wise distributed load over 3.2. Representative actions
the full beam span, though more sophisticated load distributions
can also be easily considered in the proposed approach. In establishing the overall planar response of a cellular beam,
Where the cellular beam is subject to sagging bending moment typical cell configurations are subjected to EFG discretisation and
under applied loading, the top and bottom flanges are affected by considered under representative actions to establish their
compressive and tensile stresses, respectively. Provided the flanges characteristic super-element response, and standard assembly
are relatively narrow, shear lag effects may be ignored [31], and procedures are then employed to obtain the discrete response of
thus a plane stress idealisation may be reasonably assumed for the overall cellular beam. The characteristic unit cell response is

Fig. 5. Unit cell under representative actions.

374 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

Fig. 6. Unit cell tractions on boundary C4.

determined on the basis of a flexibility-based formulation, by where ux2 is the vector of horizontal displacements for all points
considering alternative load cases of unit tractions, each along C2. This is where meshfree method is of benefit since the
corresponding to one of the super-element DOF nodes 2–4, while recovery of rx2 and ux2 values along C2 (at the Gauss points) is very
equilibrium of the unit cell is satisfied by similar tractions at node straightforward due to the continuous approximated field by the
1 (Fig. 5). MLS shape function.
A brief formulation of the representative actions is presented With each load case associated with a set of nodal displace-
here. Consider a unit cell (Fig. 5) as a super-element possessing ments in accordance with the EFG approximation, this provides a
12 DOF defined by: flexibility matrix which can be inverted for the planar stiffness
matrix of the 4-noded super-element representing the individual
UA ¼ hu1 ; v1 ; h1 iT for node 1 unit cell response. Note that displacement compatibility along
ð29Þ
UB ¼ hu2 ; v2 ; h2 ; u3 ; v3 ; h3 ; u4 ; v4 ; h4 iT for node 2; 3 and 4 the edges between adjacent unit cell super-elements is satisfied
only in an approximate weak sense, as can be observed for exam-
The associated components of forces, PA and PB, are defined as: ple from Eq. (31) where the same u2 does not guarantee the same
PA ¼ hfx1 ; fy1 ; m1 iT for node 1 ux2 along the whole edge. Accordingly, a model for the whole beam
based on the assembly of super-element contributions is expected
PB ¼ hfx2 ; fy2 ; m2 ; fx3 ; fy3 ; m3 ; fx4 ; fy4 ; m4 iT for node 2; 3 and 4 to be associated with some compatibility defaults which can lead
ð30Þ to a more flexible response compared to a full 2D model of the
overall beam. However, as shown in subsequent application exam-
The effect of a unit force, for example horizontal force at node 2, is
ples, the proposed unit-cell approach combines computational effi-
studied by applying a unit traction at that particular point,
R ciency arising from the super-element simplification with high
fx2  C2 rx2 dC ¼ 1, while equilibrating the system with additional
levels of solution accuracy.
equivalent forces at node 1, fx1 = 1 and fy1 = m1 = 0, so as to avoid The full expression of the flexibility formulation can be written
reactions and hence stress concentrations at the web-post supports, as:
as illustrated in Fig. 7. Based on the principle of virtual work along 8 9 2 3 8 9 2 3
the boundary of node 2 (C2), the nodal displacement u2 associated >
< UA > = FA >
< uqA >
= I33
6 7 6 7
with the applied forces can be determined by:    ¼ 4    5 PB þ q    þ 4    5 UA
Z Z >
: >
; >
: >
; ð32Þ
fx2 ¼1 UB FB uqB TB
u2  fx2 ¼ T
r
x2 ux2 d C ) u2 ¼ r T
C
x2 ux2 d ð31Þ |fflfflffl{zfflfflffl} |fflfflfflffl{zfflfflfflffl}
C2 C2 129 123

Γ1 Γ2
fx1 fx2

fx1 ≡ − ∫ σ x1 dΓ = −1 fx 2 ≡ ∫ σ x 2 dΓ = 1 u 2 = ∫ σ x 2 T ux 2 dΓ
Γ1 Γ2 Γ2

Fig. 7. Effect of a unit horizontal force at node 2 and the equilibrating force at node 1.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 375

in which [FA, FB]T is the flexibility matrix obtained by separate anal- putational cost. Note that KE is a symmetric matrix, and hence the
yses of the unit cell under specific traction forces in PB. The effect of overall stiffness matrix, which is simply the assembly of the individ-
uniform-distributed loads, q, is accounted for in the equation by the ual unit stiffness matrices KE, is also a symmetric matrix. In addi-
component huqA ; uqB iT that is the flexibility vector associated with q. tion, the overall stiffness matrix could be banded with

A ¼ hu
The vector U
1 ;


1 ;v h1 iT corresponds to additional displacements appropriate numbering of the super-element nodes.
at node 1 which allow for rigid body movement, where I33 is a
3  3 identity matrix, and TB is a rigid body transformation matrix 3.3. Global analysis
given by:
The global stiffness matrix is obtained by assembling the stiff-
ness matrices of individual super-elements, and the linear system
of equilibrium equations is solved using standard techniques
allowing for essential boundary conditions at the super-element
nodes. The set of linear equations associated with the global sys-
ð33Þ
tem is given by:
in which S and Dct denote the horizontal and vertical distances 1
between the super-element nodes (Fig. 5). Equilibrium conditions KG UG ¼ PG þ PGP ) UG ¼ ðKG Þ ðPG þ PGP Þ ð37Þ
are also used to relate PA to PB and q as follows: G
where K is the global stiffness matrix of size 3Ne  3Ne, Ne being
PA ¼ TTB  PB þ qTq ð34Þ the number of super-element nodes, and:

2 T UG ¼ hU1 ; V1 ; H1 ; U2 ; V2 ; H2 ; . . . ; UNe ; VNe ; HNe iT

where Tq ¼ h0; S; S =2i . ð38Þ
The flexibility Eq. (32) can now be converted to a stiffness- PG ¼ hFx1 ; Fy1 ; M1 ; Fx2 ; Fy2 ; M2 ; . . . ; FxNe ; FyNe ; MNe iT
based formulation, by solving for PB and eliminating UA as follows:
2 3 8 9 8 9 2 3 PGp is a vector of equivalent global nodal loads corresponding to dis-
FA >
< UA >
= >
< uqA >
= I33 tributed loading q, which has the form:
6 7 6 7
4   5 PB ¼   q   4   5 UA ) D ET
>
: > ; >
: >
;
FB UB uqB TB PGp ¼ hPEp i1 ; hPEp i2 ; . . . ; hPEp iNe ð39Þ
|fflffl{zfflffl} |fflfflfflffl{zfflfflfflffl}
129 123
T
F where hPEp ii ¼ fPEp gi is the local components vector obtained in Eq.
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
2 3 8 9 8 9
  FA  > UA >  > uqA > (36):
. . < = . < =
6 7 08 9 8 91
TB .. I99 4  5 PB ¼ TB .. I99   q TB .. I99  ) >
> > > > n o < uqA >
= >< Tq > =
|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} FB |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} : UB ; |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} : uqB ; PEp
B
¼ q@KE      A
C
ð40Þ
912 |fflffl{zfflffl} 912 912 > > > >
i : ; : ;
129
8 9 8 9 uqB 091
 >< UA >=  >< uqA > =
1 .. 1 .. Having obtained the global displacements UG from Eq. (37), local
PB ¼ F TB . I99   qF TB . I99 
> > > >
|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} : UB ; |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} : uqB ;
reactions of the modular unit cells at each super-node can be read-
912 912 ily determined, which in turn enable the recovery of planar stresses,
ð35Þ as discussed hereafter.

Considering Eq. (34), this expression can be extended to include the 3.4. Recovery of planar stresses over unit-cell
resistance forces at all four super-element nodes:
PB The following expression summarises the determination of the
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
0 ffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl1
ffl{
8 9 2 3 8 9 8 9 planar stress field of the current approach:
< PA >
> = TB B  > < UA > =  > < uqA > =C
6 7B . . C uI
  ¼ 4   5 BF1 TB .. I99    qF1 TB .. I99  C 8 9 zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
0 8 91
: >
> @ : >
> > >
;A
PB
;
I99 |fflfflfflfflfflfflfflfflffl {zfflfflfflfflfflfflfflffl ffl } UB
; |fflfflfflfflfflfflfflfflffl {zfflfflfflfflfflfflfflfflffl } :
uqB < rx >
> = .
>
< PB > =
ð41Þ
|fflfflfflffl{zfflfflfflffl} 912 912
rj ¼ ry ¼ Dej ¼ DBj B
@½FB ..uqB     A
C
129
8 9 >
: >
; >
: > ;
< Tq >
> =
sxy j q
þ q 
>
: >
; in which D is the plane stress constitutive matrix, and Bj is the
091 strain displacement matrix where subscript j is the index for Gauss
PE UE PEp
zfflfflffl
KE points in unit cells. The terms in bracket produce the planar dis-
8ffl}|fflfflffl9
ffl{ zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
2 3 zfflfflffl
8 ffl}|fflfflffl9ffl{ zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
0 8 ffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
9 8 91 ffl{
> TB placements vector uI of nodes in the super-element domain, where
< PA > = 
.
>< UA > = < uqA >
> = > < Tq > =
6 7 B C .
 ¼ 4  5 ½F1 TB .. I99   q@KE     A ½FB ..uqB  is a transformation matrix relating the EFG nodal parame-
: >
> ; >
: >
; >
: >
; >
: >
;
PB I99 |fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl} UB uqB 091 ters to the discrete forces of the super-element nodes PB and dis-
|fflfflfflffl{zfflfflfflffl} 912
129 tributed load q. This equation encapsulates the two levels of
PE ¼ KE  UE  PEp discretisation employed for planar analysis, namely the EFG and
super-element levels.
ð36Þ
It is important to note that since a unit-based analysis of the full
This is the final expression for the unit-cell super-element response, beam is implemented, the obtained stresses are continuous only
which is characteristic of a linear discrete system. The above within the local cell domain, though the associated inaccuracy is
expression needs to be evaluated only once for identical unit cells negligible for typical beams, as illustrated later in the verification
which are repeated along a cellular beam, where the overall beam examples. Moreover, a continuous field of the planar stress could
response can then be assembled without an undue increase in com- be achieved by applying the MLS approximation at the overall
376 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

beam level, though the computational effort would then be rela- through several illustrative examples presented in the subsequent
tively excessive. Indeed, the proposal of unit-cell super-elements sections.
in this work is primarily aimed at overcoming this excessive com-
putational demand while retaining good levels of accuracy in the
4. Numerical examples and discussion
planar stress distribution.
Here, numerical examples are presented to evaluate the perfor-
mance of the developed EFG method considering the MLRG inte-
3.5. Overall numerical procedure gration approach for the planar analysis of the web panel in a
unit-cell (without the flanges). Two specific cases are chosen to
A numerical procedure for planar analysis of a whole cellular examine the ability of the EFG method in dealing with irregular
beam using the proposed EFG method is illustrated in Fig. 8. The domains under general loading conditions.
calculation involves two different levels of analysis, namely the The selection of support radius and domain integration param-
global beam and local unit-cell levels. For a regular cellular beam eters is discussed. As suggested earlier, parameters in the weight
subjected to a UDL on top of its flange, only three unit-cells, left function of Eq. (13) are kept as k = 1 and c = rI/4 in all calculations,
end, right end and internal cell, need to be analysed locally by unless stated otherwise. A quadratic basis function is used herein.
the EFG method. Each unit-cell is then considered as a super- Discretisation over the domain is made on the basis of a rectangu-
element, where the internal super-element is repeatedly applied lar grid (as close to square as possible), where the EFG nodes are
as a modular computational entity for the various internal cells located at all grid intersections within the domain, and additional
along the span. The significant advantages of this approach in nodes are introduced along curved edges. Another layer consisting
terms of accuracy and computational efficiency are highlighted of a different rectangular grid is also employed for numerical

Fig. 8. Flowchart of the proposed EFG method for cellular beams planar analysis.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 377

integration. These two sets of grid could be of the same formation, 4.1.1. Influence of support radius
but for generality they are allowed to be different. This is to enable The effect of the support radius rI is conducted here for the
a complete order integration rule to be employed with a coarse dis- irregular domain, where irregularity in the nodal distribution
cretisation of the displacement fields. occurs near the curved edges. Two cases considering different sets
For the two considered examples, the MLS approximation of nodes (a) 124 nodes based on a rectangular grid of 10  13, and
employs quadratic basis functions. Comparison is made between (b) 392 nodes based on a 20  25 grid, are used in conjunction with
the developed EFG method and detailed FEA models, where rela- 3  3 or 6  6 quadrature within a sufficient number of rectangular
tive performance is investigated in relation to the total DOFs. Since background cells, as elaborated in Section 4.1.2. Results are sum-
a theoretical solution is not available for these problems, the ‘exact’ marised in Fig. 10(a) and (b) for the normalised displacement
solution is estimated from a detailed FEA model with a very fine and strain energy in terms of j = rI/h, where h is the nodal spacing,
mesh. taken as the maximum of the two element edge lengths in the x
and y directions. It is observed that for j exceeding 6.0, the deflec-
4.1. Unit cell under transverse compressive load tion prediction deteriorates for the two levels of discretisation,
while for j less than 4.0 the strain energy tends to be underesti-
As illustrated in Fig. 9, a unit cell of a cellular beam with a depth mated. On the other hand, the use of different quadrature rules
Dp = 1.0 m, hole diameter Do = 0.6 m, spacing S = 0.8 m and web has a relatively a small effect. Accordingly, rI is henceforth fixed
thickness tw = 0.001 m is considered for the first case. The panel to 5.0  h, as this provides good accuracy for both regular and
is loaded with 1 kN/m UDL along the top and the bottom edges irregular domains.
in opposite directions to cause significant compression, where
the material parameters are E = 210  109 N/m2 and t = 0.25. As 4.1.2. Domain integration
mentioned earlier, the three nodal supports across the web-post The multi-level rectangular grid (MLRG) approach developed
are used to prevent rigid body movement. for integration of irregular domains is investigated here. Consider-
As can be seen, the domain discretisation of the panel using the ation is given to the most computationally efficient way to inte-
EFG method is performed using a rectangular grid of nodes, exclud- grate the domain function without compromising accuracy,
ing nodes that would be located outside the domain boundary. possibly by optimising the number of quadrature points. Since
Additional nodes are employed along the curved edges (Fig. 9(b)) the distribution of Gauss points is based on rectangular cells, the
in order to improve the accuracy of the MLS approximation first important concern is to ensure that the coverage of cells over
functions. the domain is maximised. In the MLRG approach, this is mainly
Fig. 9(b) shows the deflected shape of a unit cell under trans- controlled by the level of rectangular grid. The second concern
verse compression, comparing the EFG prediction with the FEA relates to the required number of quadrature points per integration
result. In the following discussion, the vertical displacement of cell. Note that the size of the cells reduces in the vicinity of the
point A (uA) is used as a reference, and where appropriate reference curved edges, as discussed earlier. Therefore, in order to avoid
to the strain energy (e) of the system is also considered. From a unnecessary computational effort in integrating small cells, a
detailed finite element analysis performed using ADAPTIC, the esti- simplification is made here whereby the quadrature order per cell
mated exact solution for this problem is: uA = 1.0343  105 m is reduced as the cell size decreases. To address these two
and e = 5.6945 Nm. concerns, a numerical parametric investigation is conducted,

–1 kN/m
y
1.0
point A

0.8

0.6
FEA ADAPTIC
FEA-ADAPTIC
Dp=1.0m
EFGmethod
EFG method (13x10)
(124 nodes)
Do=0.6m
0.4
y ,v

0.2

x ,u 0.0
-0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 x
1 kN/m

S= 0.8m
(a) (b)
Fig. 9. Unit cell under transverse compression: (a) applied loads and boundary conditions and (b) distribution of nodes and deflected shapes with EFG and FEA models.
378 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

100.1 100.1

100.0 100.0

Strain Energy , e (%)

99.9 99.9
Deflecon, uA (%)

99.8 99.8

99.7 99.7

3x3 quadrature (124 nodes) 3x3 quadrature (124 nodes)

99.6 3x3 quadrature (392 nodes) 99.6 3x3 quadrature (392 nodes)
6x6 quadrature (124 nodes) 6x6 quadrature (124 nodes)
6x6 quadrature (392 nodes) 6x6 quadrature (392 nodes)
99.5 99.5
2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0
Support radius / mesh size, κ Support radius / mesh size, κ
(a) (b)
Fig. 10. Effect of support radius on (a) deflection of point A and (b) system strain energy.

150 150

3x3 quadrature 2x2 quadrature

140 4x4 quadrature 140 4x4 quadrature
6x6 quadrature 6x6 quadrature
9x9 quadrature 9x9 quadrature
130 130
Strain energy, e (%)

120 120

110 110

100 100

90 90
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Level of rectangular grid Level of rectangular grid
(a) Main grid of 8 x 10 (b) Main grid of 16 x 20
Fig. 11. Prediction of strain energy for different quadrature schemes.

enabling the most appropriate combination of the number of cells, ent quadrature rules and rectangular grid levels. Except for the
quadrature order and the level of rectangular grid to be coarse 8  10 grid with 3  3 quadrature, it is clear from the figure
established. that the effect of increasing the Gauss points appears to be insignif-
Fig. 11 demonstrates the integration result of the total strain icant. In fact, the use of more integration cells appears to be more
energy for the current problem. The EFG discretisation is based important, as evident from comparing the results for the 8  10
on 20  25 grid with a total of 392 nodes. Two sets of background and 16  20 grids. It is also noted that a high level of MLRG,
grid are chosen for integration cells, each of which employs differ- between 6 and 8 levels, is required to establish accurate
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 379

102.5
Belytschko & Lu [25] - 8x10 grid
102.0 Belytschko & Lu [25] - 16x20 grid
MLRG - 16x20 grid

Strain energy, e (%)

101.5

101.0

100.5

100.0

99.5

99.0
3 4 5 6 7 8 9
Quadrature rules (Ng x Ng)

Fig. 12. Comparison between MLRG integration and approach of Belytschko et al. [25].

integration – higher than that, the return in accuracy for computa- 4.1.4. Convergence of energy
tional effort is significantly diminished. For consistency, a grid of The distribution of the normal stresses rxx and ryy over the unit
16  20 cells with 4  4 quadrature rules and 6 levels of MLRG is cell are shown in Figs. 14 and 15 respectively. Comparison of these
proposed as the optimal integration scheme for a unit cell panel. stresses against FEA predictions is also made in detail, considering
Comparing the proposed MLRG integration to the original single- the stress along the panel depth at x = S/2 as illustrated in Figs. 14
level approach of Belytschko et al. [25] in Fig. 12, it is evident that (c) and 15(c). It can be seen that the EFG model is capable of
the new MLRG approach offers a more consistent integration with achieving high-accuracy even for a coarse distribution of nodes
a greater level of accuracy. (e.g. 124 nodes based on 10  13 grid). More importantly, the
stresses are continuous within the domain unlike the FEA model
4.1.3. Convergence of displacements which would normally require post-processing (e.g. by taking aver-
The vertical displacement at point A of the unit cell under trans- aged nodal stresses between elements) for obtaining a smooth
verse compression has been shown earlier in Fig. 9(b), where good stress distribution. Finally, the prediction of strain energy by the
agreement of the nodal displacements is observed between the EFG and FEA models, considering different levels of discretisation,
EFG and FEA models using a much smaller number of EFG nodes is presented in Fig. 16. Clearly, the EFG model is superior to the FEA
(NEFG  NFEA). Detailed comparison considering the different levels model, where excellent prediction of the strain energy can be
of discretisation is provided in Fig. 13, where it is evident that the achieved with as few as 124 nodes, whereas the FEA model
EFG model converges much faster than the FEA model utilising the employing the 9-noded quadratic shell elements requires more
conforming 9-noded shell elements of ADAPTIC [32,33] with than 1300 nodes for comparable accuracy.
increasing levels of nodal discretisation. Note that although the
converged value of the EFG model appears to be slightly lower than 4.2. Unit cell under shear
the estimated exact solution based on the convergence rate of the
FEA solution with conforming shell elements, all EFG predictions The same unit cell as in the previous example is now considered
offer excellent accuracy with an error of less than 0.1%, even for with different boundary conditions and loading, where, as shown
the coarsest discretisation level of 124 nodes. in Fig. 17(a), the panel is clamped along the left side edges and is

-1.022

-1.024 EFG method

FEA ADAPTIC
Displacement, uA (x10-5 m)

-1.026 Esmated exact soluon

-1.028

-1.030

-1.032

-1.034

-1.036

-1.038
100 300 500 700 900 1100 1300
Number of nodes, N
Fig. 13. Convergence of displacements for unit cell under compression.
380 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

2500

2000

1500

1000

Stress, Pa
500

2.500e+03 -500
2.050e+03
1.600e+03
1.150e+03
7.000e+02
-1000
2.500e+02
-2.000e+02
-6.500e+02 -1500
-1.100e+03
-1.550e+03
-2.000e+03
-2000
(a) FEA ADAPTIC (2261 nodes) (b) EFG method (392 nodes)

3000

2500 FEA ADAPTIC

y EFG method (124 nodes)
2000 EFG method (392 nodes)

1500
Stress, Pa

1000

500
y
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-500

-1000

-1500
(c)
Fig. 14. Distribution of normal stress rxx (a) and (b) over unit cell and (c) over depth at x = S/2.

horizontally restrained along the right side edges where shear trac- with the previous example of the unit cell under transverse
tions are applied. The two alternative levels of discretisation as in compression.
the preceding example, namely 124 nodes based on a 10  13 grid
and 392 nodes based on a 25  16 grid, are also considered. The
estimated exact solution, obtained from a detailed FEA model 4.2.2. Convergence of displacements
using ADAPTIC, is determined for the vertical displacement as The deflected shape of the unit cell under the applied shear
uA = 1.7308  105 m and for the total strain energy as loading is depicted in Fig. 17. The convergence of displacements
e = 3.4904 N m. for this problem is presented in Fig. 19 for the EFG model com-
pared with the FEA model utilising the 9-noded conforming shell
4.2.1. Domain integration element, specifically considering the displacement of point A. For
The prediction of strain energy is summarised in Fig. 18 for the same number of nodes, the EFG model exhibits better accu-
cases with background cells of grid 8  10 and grid 16  20. The racy than the FEA model, and it converges quickly to the exact
latter provides a better prediction of the exact solution especially solution when the total nodes approach 600 compared to over
for a low level of MLRG subdivision. However, employing at least 1300 for the FEA model. Except for the coarsest discretisation
6 levels of rectangular grid for both sets is necessary to achieve of 124 nodes, all subsequent EFG models provide accuracy for
good accuracy with an error less than 0.5%, which is consistent uA within 0.1%.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 381

-500

-1000

-1500

-2000

Stress, Pa
-2500

-3000
0.000e+00
-5.000e+02 -3500
-1.000e+03
-1.500e+03
-2.000e+03
-2.500e+03
-4000
-3.000e+03
-3.500e+03
-4.000e+03 -4500
-4.500e+03
-5.000e+03
-5000
(a) FEA ADAPTIC (2261 nodes) (b) EFG method (392 nodes)

Fig. 15. Distribution of normal stress ryy (a) and (b) over unit cell and (c) over depth at x = S/2.

5.71

5.70
Strain Energy, e (Nm)

5.69

5.68

5.67
EFG method
FEA ADAPTIC
5.66
Esmated exact soluon

5.65
100 300 500 700 900 1100 1300
Number of nodes, N
Fig. 16. Convergence of strain energy for unit cell under compression.
382 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

y
–1 kN/m
1.0

point A

0.8

0.6
Dp=1.0m
Do=0.6m FEA ADAPTIC
FEA-ADAPTIC
0.4 EFG method(124
EFG method (13x10)
nodes)
y ,v

–1 kN/m
0.2

x ,u 0.0 x
0.0 0.2 0.4 0.6 0.8

S=0.8m -0.2

(a) (b)
Fig. 17. Unit cell under shear: (a) applied loads and boundary conditions and (b) distribution of nodes and deflected shapes with EFG and FEA models.

150 150

3x3 quadrature 2x2 quadrature

140 4x4 quadrature 140 4x4 quadrature
6x6 quadrature 6x6 quadrature
9x9 quadrature 9x9 quadrature
130 130
Strain energy, e (%)

120 120

110 110

100 100

90 90
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Level of rectangular grid Level of rectangular grid
(a) Main grid of 8 x 10 (b) Main grid of 16 x 20
Fig. 18. Strain energy resulting from different quadrature schemes.

4.2.3. Convergence of energy in Fig. 20(c), where the prediction of the EFG model with 124 nodes
The distribution of shear stresses sxy in the unit cell is depicted shows slight discrepancies near the two ends and at mid-depth,
in Fig. 20, where good agreement is observed between the EFG and though these discrepancies are considerably reduced with the finer
FEA results. For a detailed quantitative comparison, the distribu- model using 392 nodes. Finally, Fig. 21 depicts the prediction of
tion of sxy across the depth of the panel (at x = S/2) is presented strain energy using the EFG and FEA models, where it is again
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 383

-1.66

-1.67 EFG method

FEA ADAPTIC

Displacement, uA (x10-5 m)
-1.68 Esmated exact soluon

-1.69

-1.70

-1.71

-1.72

-1.73

-1.74

-1.75
100 300 500 700 900 1100 1300
Number of nodes, N
Fig. 19. Convergence of displacements for unit cell under shear.

500

-500

Stress, Pa
-1000
5.000e+02
2.500e+02
0.000e+00
-2.500e+02
-5.000e+02 -1500
-7.500e+02
-1.000e+03
-1.250e+03
-1.500e+03
-1.750e+03
-2.000e+03
-2000
(a) FEA ADAPTIC (2261 nodes) (b) EFG method (392 nodes)

500

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
y

-500
Stress, Pa

-1000

FEA ADAPTIC
-1500
EFG method (124 nodes)
EFG method (392 nodes)

-2000
(c)
Fig. 20. Distribution of shear stress sxy(a) and (b) over unit cell and (c) over depth at x = S/2.
384 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

3.54

3.50

Strain Energy, e (Nm)

3.46

3.42

EFG method
3.38 FEA ADAPTIC
Esmated exact soluon

3.34
100 300 500 700 900 1100 1300
Number of nodes, N
Fig. 21. Convergence of strain energy for unit cell under shear.

evident that the EFG models offer a better convergence rate com- shown in Fig. 22. The material properties are taken as Young’s
pared to the FEA models. modulus E = 210  109 N/m2 and Possion’s ratio t = 0.3.
Fig. 22 demonstrates the deflected shape of the simply-
5. Application to cellular beams supported beam under UDL, comparing the results between
detailed FEA and the proposed EFG approach. Point A in the figure
Several examples of cellular beam problems are presented here depicts the location with the maximum vertical displacement,
to demonstrate the performance of the proposed unit-based EFG which is being used as a reference point for detailed comparison,
method. In particular, the ability of this method to obtain accurate where the results are shown in Fig. 23. Although the EFG method
predictions of the planar response is highlighted in comparison is considered with the unit-cell approach, the total number of
with 2D FEA models using ADAPTIC in terms of the number of nodes N over the whole beam is used for comparing the relative
nodes. In all cases, the support radius rI is taken as 5  h, while accuracy of FEA and EFG predictions. The estimated exact solution
the weight function parameters are taken as k = 1 and c = rI/4. uA(exact) is 2.9279  105 m, and the FEA-ADAPTIC using the
The previously established quadrature rule of 4  4 per cell with 9-noded conforming shell elements appears to predict the value
6 levels of MLRG is employed for domain integration. Again, the reasonably well. On the other hand, the EFG method through the
‘exact’ solution for these problems is estimated from a detailed use of a unit-cell model provides a slightly more flexible result
FEA model utilising a very fine mesh. which is 1.5% lower than uA(exact). This relatively small disagree-
ment is attributed to the unit-based approach used for establishing
5.1. Simply supported cellular beam under UDL the overall cellular beam response. Since a flexibility-based super-
element is adopted, where the displacements at the super-element
A relatively short cellular beam of length L = 7.92 m and overall nodes are averaged approximations from boundary displacements,
depth Dp = 1.603 m, hole diameter Do = 0.8 m and spacing there would be small displacement discrepancies, so-called com-
S = 1.472 m is considered. The beam is simply-supported and pro- patibility defaults, between adjacent unit cells, which lead to the
duced from an original I-section of UB1016  305  222 with web EFG unit-based solution being slightly more flexible.
thickness tw = 16 mm, flange thickness tf = 21.1 mm and flange As illustrated in Fig. 24, a favourable comparison against the
width bf = 300 mm. Five holes are symmetrically introduced along detailed FEA model is achieved by the proposed EFG approach,
the length and 1.0 kN/m UDL is applied on top of the beam, as despite the use of a modular formulation for the local unit-cell

2.5
y
2.0

1.5

1.0

0.5

0.0 x
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0
-0.5
point A FEA ADAPTIC
-1.0
proposed EFG (828 nodes)
-1.5

Fig. 22. Simply supported cellular beam under uniform distributed loading.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 385

-2.86
Proposed EFG
-2.88 Converged value (EFG)

Deflecon, uA (x 10 -5 m)
FEA-ADAPTIC
-2.90 Esmated exact soluon

-2.92

-2.94

-2.96

-2.98

-3.00
0 2000 4000 6000 8000 10000
Number of nodes, N
Fig. 23. Convergence of the deflection at point A for the simply supported beam.

response. It can be seen that the EFG stress functions are readily the other end free thus corresponding to a cantilever condition.
continuous over the unit cells, whereas some post-processing An identical beam geometry to the previous example is used, and
would be required in FEA for smoothing the stresses value over a UDL of 1.0 kN/m is applied to top of the beam.
the domain. Note that additional stress concentration near the
beam supports in the EFG model is caused by the vertical restraints
5.2.1. Fully clamped cellular beam
that are applied at the location of super-element nodes, while
Since the developed model is based on 4-noded super-elements,
these restraints are distributed along the edges in the FEA model.
the enforcement of the fixed end supports can only be made based
The strain energy predicted by the proposed EFG approach is
on associated freedoms of super-element nodes, though this
depicted in Fig. 25, where the prediction is slightly higher than that
should not cause significant discrepancy. The deformed shape of
of detailed FEA owing to the more flexible response by the EFG
the fully clamped beam under UDL is depicted in Fig. 27, while
unit-based solution as discussed earlier. By considering only the
the result comparisons based on point A are provided in Fig. 28.
result of the EFG analysis, favourable predictions with less than
Again, the simplified EFG approach slightly overestimates the
0.1% error compared to the converged value are achieved when
the total number of nodes exceeds 2500, which is basically equiv- detailed FEA solution ðuAðexactÞ ¼ 1:3984  105 mÞ due to possible
alent to an average 300 nodes per unit cell. This performance is discrepancy between the connecting super-elements and also at
consistent with the previously obtained result for a unit-cell the fixed supports. The same applies to the estimation of the corre-
analysis. sponding strain energy shown in Fig. 29, where a slight overesti-
Finally, the convergence rate of the FEM and EFG solutions is mate standing at less than 1.5% is observed. Considering the EFG
investigated by considering the relative error (re) in the energy solutions, an error within 0.1% of the converged solution is
norm, which is related to the error in the predicted strain energy: achieved when the total number of nodes reaches approximately
3000 nodes, compared to more than 10,000 nodes with FEM.
jeconv  epred j Finally, in terms of the convergence rate, the proposed EFG unit-
re ¼ ð42Þ
jeconv j cell method exhibits again a better performance compared to
FEM, with an accelerated rate as the level of discretisation becomes
where epred is the predicted strain energy and econv is the converged/
finer, as depicted in Fig. 30.
estimated exact solution. A comparison between the FEM and EFG
convergence performance is depicted Fig. 26 in terms of a non-
dimensional characteristic mesh size he, which for a 2D problem 5.2.2. Cantilever cellular beam
is related to the number of nodes (N) by: The same beam subjected to UDL along the top edge is now con-
.pffiffiffiffi sidered under cantilever conditions. As depicted in Fig. 31, the pro-
he ¼ 1 N ð43Þ posed EFG approach provides favourable comparison against
detailed FEA in relation to the deformed shape. In terms of the pre-
It can be observed that the FEM solution achieves the optimal con- diction of the maximum deflection at point A, the EFG approxima-
4
vergence rate expected from a quadratic element with re ¼ Oðhe Þ. tion shows a promising result with 0.2% compared to the estimated
On the other hand, the proposed EFG unit-cell method attains a exact solution (uA(exact) = 2.3142  105 m) as presented in
similar convergence rate for a coarse level of discretisation, though Fig. 32. The corresponding strain energy prediction is also within
it notably achieves an improved and accelerated rate as the level of 0.35% of the estimated exact solution (Fig. 33). Furthermore, con-
discretisation becomes finer. sidering the convergence of both methods shown in Fig. 34, a bet-
ter convergence rate is achieved by the proposed method similar to
5.2. Cellular beams with different support conditions the two previous cases. Nevertheless, it is noted that, the perfor-
mance of the EFG method for the cantilever beam is marginally
Further verification for cellular beams with different support better than the two previous cases, which suggests that the com-
conditions is presented here. Two particular cases are chosen: (1) patibility defaults for the cantilever condition are reduced in com-
both beam ends are fully fixed and (2) only one end is fixed with parison with the simply supported and fully clamped conditions.
386 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

Normal stress, σxx x 10

4

1
Proposed EFG

0.5

Stress, Pa
0

-0.5

-1
FEA-ADAPTIC

1.000e+0
8.000e+0
6.000e+0
4.000e+0
2.000e+0
0.000e+0
-2.000e+0
-4.000e+0
-6.000e+0
-8.000e+0
-1.000e+0

Normal stress, σyy x 10

4

1
Proposed EFG

0.5

Stress, Pa
0

-0.5

-1
FEA-ADAPTIC

1.000e+0
8.000e+0
6.000e+0
4.000e+0
2.000e+0
0.000e+0
-2.000e+0
-4.000e+0
-6.000e+0
-8.000e+0
-1.000e+0

Shear stress, τxy

8000
6000
Proposed EFG

Stress, Pa
4000
2000

0
-2000
-4000
-6000

-8000
FEA-ADAPTIC

8.000e+0
6.400e+0
4.800e+0
3.200e+0
1.600e+0
0.000e+0
-1.600e+0
-3.200e+0
-4.800e+0
-6.400e+0
-8.000e+0

Fig. 24. Comparison of stress patterns between the proposed EFG and detailed FEA.

5.3. Computational benefits of EFG method for increasing number of suffers from the need to use more elements and nodes for accurate
holes prediction in proportion to the number of holes. To highlight these
benefits, simply supported cellular beams with an increasing num-
The proposed EFG approach benefits from the repetitiveness of ber of web openings, varying from 2 to 21 holes, are analysed, and
cells in typical cellular beam profiles, whereas conventional FEA the results are summarised in Fig. 35. With an excellent compar-
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 387

7.58

7.54

Strain energy, e (x 10 2 Nm)

7.50

7.46

7.42

7.38 Proposed EFG

Converged value (EFG)
7.34 FEA-ADAPTIC
Esmated exact soluon
7.30
0 2000 4000 6000 8000 10000
Number of nodes, N
Fig. 25. Convergence of strain energy predictions for simply-supported cellular beam.

-1.0
-1.5 4

-2.0 1

-2.5
-3.0
log(re)

-3.5
-4.0
-4.5
-5.0 Proposed EFG
FEA-ADAPTIC
-5.5
-6.0
-2.25 -2 -1.75 -1.5 -1.25 -1
log(he)

Fig. 26. Convergence rates of EFG and FEM predictions for simply-supported cellular beam.

2.5
y
2.0

1.5

1.0

0.5

0.0
x
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0
-0.5
FEA ADAPTIC
-1.0 point A
proposed EFG (828 nodes)
-1.5

Fig. 27. Fully clamped cellular beam under uniform distributed loading.

ison between the EFG and FEA results in relation to the predicted recovery. The superior performance of the proposed EFG unit-cell
strain energy, the proposed EFG method consumes a remarkably approach is mainly due to the focus of the computation on the
small computing time regardless of the number of holes compared three unit-cells (left, right and internal cells), where the internal
to the FEA models, where the reported computing times include all unit-cells are duplicates of the same internal super-element. It
element/unit-cell computations, solution of equations and stress should be noted though that the comparison of computing time
388 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

-1.37
Proposed EFG
-1.38 Converged value (EFG)

Deflecon, uA (x 10 -5 m)
FEA-ADAPTIC
Esmated exact soluon
-1.39

-1.40

-1.41

-1.42

-1.43
0 2000 4000 6000 8000 10000
Number of nodes, N
Fig. 28. Convergence of the deflection at point A for the fully clamped cellular beam.

3.46

3.44
Strain energy, e (x 10 2 Nm)

3.42

3.40

3.38

3.36

3.34
Proposed EFG
3.32 Converged value (EFG)
FEA-ADAPTIC
3.30 Esmated exact soluon
3.28
0 2000 4000 6000 8000 10000
Number of nodes, N
Fig. 29. Convergence of the strain energy for the fully clamped cellular beam.

0.0
- 0.5 4
- 1.0 1

- 1.5

- 2.0
log(re)

- 2.5

- 3.0

- 3.5

- 4.0 Proposed EFG

- 4.5 FEA-ADAPTIC

- 5.0
-2.25 -2 -1.75 -1.5 -1.25 -1
log(he)

Fig. 30. Convergence rate of EFG and FEM predictions for fully clamped cellular beam.

is only indicative, since the proposed EFG approach is implemented time, between the proposed EFG method and a full scale model
using MATLAB, while the FEA model employs ADAPTIC which is utilising only the EFG method (without the unit-cell formulation)
implemented using the FORTRAN programming language. An also exhibited a far better performance by the novel unit cell idea
attempt to compare the results, particularly the computational than the conventional discretisation approach.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 389

2.5
y FEA ADAPTIC
2.0
proposed EFG (828 nodes)

1.5

1.0

0.5

0.0
0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0
-0.5

-1.0 point A
x

-1.5

Fig. 31. Cantilever cellular beam subjected to uniformly distributed loading.

-2.29
Proposed EFG
Converged value (EFG)
Deflecon, uA (x 10 -4 m)

FEA-ADAPTIC
-2.30
Esmated exact soluon

-2.31

-2.32

-2.33
0 2000 4000 6000 8000 10000
Number of nodes, N
Fig. 32. Convergence of the deflection at point A for the cantilever cellular beam.

0.414

0.413
Strain energy, e (Nm)

0.412

0.411

0.410

0.409
Proposed EFG
0.408 Converged value (EFG)
FEA-ADAPTIC
0.407
Esmated exact soluon
0.406
0 2000 4000 6000 8000 10000
Number of nodes, N
Fig. 33. Convergence of the strain energy for the cantilever beam.
390 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391

-1.0

-1.5 4
-2.0 1

-2.5

-3.0
log(re) -3.5

-4.0

-4.5

-5.0 Proposed EFG

-5.5 FEA-ADAPTIC

-6.0
-2.25 -2 -1.75 -1.5 -1.25 -1
log(he)

Fig. 34. Convergence rate of EFG and FEM predictions for cantilever cellular beam.

60.0 300

Energy - Proposed EFG

50.0 Energy - FEA-ADAPTIC 250

Compung me (seconds)

Time - Prosposed EFG
Time - FEA-ADAPTIC
Strain energy (Nm)

40.0 200

30.0 150

20.0 100

10.0 50

0.0 0
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Number of holes
Fig. 35. Strain energy and computing time for cellular beams with varying number of holes.

6. Conclusion sent EFG model offers significantly better results with a smaller
number of DOFs compared to FE models. In this respect, the use
This paper presents an effective approach for modelling the pla- of nodes along curved boundaries, in addition to a simple arrange-
nar response of cellular beams, in which the concept of a unit-cell ment of nodes based on a regular rectangular grid, is found essen-
super-element is introduced, which benefits further from the tial to improve the accuracy of EFG predictions.
adoption and enhancement of the Element-Free Galerkin (EFG) Besides the specific enhancements of the EFG approach for
method for establishing the super-element response characteris- irregular domains, more significant computational benefits arise
tics. The main motivation of this work is the accurate and efficient from the proposed concept of a ‘unit-cell’ formulation, which takes
determination of the planar stress distribution under in-plane advantage of the repetitive profile in cellular beams with regular
loads, which is required for assessing local out-of-plane buckling web holes. In this approach, the individual cell response is cap-
in the web components of a cellular beam. tured with a super-element utilising only four nodes, with a
The EFG method offers significant advantages through its reduced number of DOFs and a corresponding set of representative
smooth and continuous shape functions, governed by a proper actions, the characteristic response of which needs only to be
choice of the polynomial basis and weight functions, thus avoiding determined once for a set of identical cells. Application examples
the need for post-processing of discontinuous stress predictions considering cellular beams under different boundary conditions
that are typical in finite element (FE) analysis. The essential bound- illustrate that deflection, stress pattern and strain energy predic-
ary conditions are imposed via a penalty method, in which a care- tions compare favourably against FE analysis, offering further an
ful selection of penalty factor is crucial to avoid numerical errors. improved representation of the continuous stress field without
An enhanced method of numerical integration is proposed in this the need for post-processing. Finally, the computational advantage
work to deal with irregular domains, particularly in the vicinity of the proposed method are highlighted considering cellular beams
of curved boundaries, where a multi-level rectangular grid (MLRG) with an increasing number of holes, where a drastically reduced
approach is shown to provide optimal efficiency and accuracy in computing time, proportional to the number of holes, is achieved
the solution. The numerical examples demonstrate that the pre- compared to FE analysis.
 A.R. Zainal Abidin et al. / Computers and Structures 182 (2017) 368–391 391

The computational and modelling benefits of the proposed [11] Institution BS. Eurocode 3: design of steel structures. Part 1-1: general rules
and rules for buildings. London (UK): British Standard Institution; 2005.
approach in predicting the planar response of cellular beams, paves
[12] Westok. The Westok Cellular BeamAvailable from: <http://www.
the way for effective local buckling analysis of web components in kloecknermetalsuk.com/kloecknerwestok/software-download/>2016
such systems, as demonstrated in previous work of the authors [accessed 01.04.16].
[21] in the context of beams with arbitrary and irregular web [13] Lawson RM. Design for openings in the webs of composite beams. CIRIA
Special Publication 51/SCI Publication 068; 1987.
openings. The extension of this local buckling approach to cellular [14] Ward JK. Design of composite and non-composite cellular beams: SCI. SCI
beams with regular openings will be the subject of future work Publication 100; 1990.
building on the current developments, thus offering a real prospect [15] Zaarour W, Redwood R. Web buckling in thin webbed castellated beams. J
Struct Eng-ASCE 1996;122:860–6.
for accurate local buckling assessment of cellular beams within a [16] Redwood R, Demirdjian S. Castellated beam web buckling in shear. J Struct
computationally efficient framework, as necessary for practical Eng-ASCE 1998;124:1202–7.
application in design practice. [17] Redwood RG. Behaviour of composite castellated beams; 2000.
[18] Lawson RM, Lim J, Hicks SJ, Simms WI. Design of composite asymmetric
cellular beams and beams with large web openings. J Constr Steel Res
Acknowledgement 2006;62:614–29.
[19] Wong VB, Burgess I, Plank R. Behaviour of composite cellular steel - concrete
beams at elevated temperatures. Int J Steel Struct 2009;9:29–37.
The funding provided by Universiti Teknologi Malaysia (UTM) [20] Panedpojaman P, Thepchatri T, Limkatanyu S. Novel design equations for shear
for the PhD work of the first author at Imperial College London is strength of local web-post buckling in cellular beams. Thin-Wall Struct
2014;76:92–104.
gratefully acknowledged. [21] Zainal Abidin AR, Izzuddin BA. Meshless local buckling analysis of steel beams
with irregular web openings. Eng Struct 2013;50:197–206.
[22] Belytschko T, Krongauz Y, Organ D, Fleming M, Krysl P. Meshless methods: an
References overview and recent developments. Comput Methods Appl Mech Eng
1996;139:3–47.
[1] British Steel Sections Plate and Commercial Steels. Design in steel 4: [23] Atluri SN, Zhu TL. New concepts in meshless methods. Int J Numer Meth Eng
castellated and cellular beams; 1995. 2000;47:537–56.
[2] ArcelorMittal. Long carbon Europe sections and merchant bars. ACB Cellular [24] Liu GR. Meshfree methods: moving beyond the finite element. 2nd ed. CRC
Beams; 2008. Press; 2010.
[3] New Steel Construction. In the grove. NSC annual review Januari 2016. The [25] Belytschko T, Lu YY, Gu L. Element-free Galerkin methods. Int J Numer Meth
British Constructional Steelwork Association Ltd; 2016. p. 16. Eng 1994;37:229–56.
[4] Sonck D, Van Impe R, Belis J. Experimental investigation of residual stresses in [26] Atluri SN, Zhu TL. The meshless local Petrov-Galerkin (MLPG) approach for
steel cellular and castellated members. Constr Build Mater 2014;54:512–9. solving problems in elasto-statics. Comput Mech 2000;25:169–79.
[5] Srimani SL, Das PK. Finite element analysis of castellated beams. Comput Struct [27] Nayroles B, Touzot G, Villon P. Generalizing the finite element method: diffuse
1978;9:169–74. approximation and diffuse elements. Comput Mech 1992;10:307–18.
[6] Hoffman RM, Dinehart DW, Gross SP, Yost JR. Analysis of stress distribution [28] Nguyen VP, Rabczuk T, Bordas S, Duflot M. Meshless methods: a review and
and failure behavior of cellular beams. In: Proceedings of the 2006 ANSYS computer implementation aspects. Math Comp Simul 2008;79:763–813.
conference. Pittsburgh (PA). [29] Rabczuk T, Belytschko T. Adaptivity for structured meshfree particle methods
[7] Aglan AA, Redwood RG. Web buckling in castellated beams. Proc Inst Civil Eng in 2D and 3D. Int J Numer Meth Eng 2005;63:1559–82.
Part 2-Res Theory 1974;57:307–20. [30] Rabczuk T, Areias PMA, Belytschko T. A meshfree thin shell method for non-
[8] Kerdal D, Nethercot DA. Failure modes for castellated beams. J Constr Steel Res linear dynamic fracture. Int J Numer Meth Eng 2007;72:524–48.
1984;4:295–315. [31] Trahair NS, Bradford MA, Nethercot DA. The behaviour and design of steel
[9] Okubo T, Nethercot DA. Web post strength in castellated steel beams. In: structures to BS5950. 3rd ed. New York: Spon Press; 2001.
Proceedings of the institution of civil engineers, part 2; 1985. [32] Izzuddin BA. Nonlinear dynamic analysis of framed structures. London
[10] Institution BS. Structural use of steelwork in building. Part 1: code of practice (UK): Imperial College London; 1991.
for design - rolled and welded sections. London (UK): British Standard [33] Izzuddin BA, Li ZX. A co-rotational formulation for large displacement analysis
Institution; 2000. of curved shells. Develop Mech Struct Mater 2004;2:1247–53.