COMMUNICATIONS IN NUMERICAL METHODS IN ENGINEERING

Commun. Numer. Meth. Engng 2005; 21:757–767

Published online 11 August 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/cnm.791

A rotation-free quadrilateral thin shell

subdivision
nite element

Seth Green and George M. Turkiyyah∗; †

Department of Civil and Environmental Engineering; University of Washington;
Seattle; WA 98195-2700; U.S.A.

SUMMARY
We present a rotation-free, subdivision-based, isoparametric quadrilateral
nite element for the analysis
of doubly-curved thin shells. The element is a generalization of spline-based C 1
nite elements to
curved domains, and general mesh connectivity. It uses semi-local basis functions that span the two-
neighbourhood of a node. The traditional diculty with this class of elements comes from the fact that
these basis functions depend on the valence of the nodes of the element. For elements with 4-valence
nodes the basis functions are bicubic polynomials, but there are no simple closed-form expressions for
them when the nodes have general connectivity patterns. In the paper, we show how the basis functions
for the general case may be eciently evaluated as linear combinations of translation and scaling
transformations of the bicubic polynomials. The performance of the resulting element is illustrated on
a hemispherical shell model problem. Copyright ? 2005 John Wiley & Sons, Ltd.

KEY WORDS: shells;
nite elements; subdivision surfaces; rotation-free

1. INTRODUCTION

We develop the formulation of a quadrilateral
nite element with translation-only degrees of

freedom that has the necessary continuity requirements for thin-shell simulations. This element
belongs to a class of
nite elements that have been recently called subdivision
nite elements
[1] because of the shape functions used in their formulations. These shape functions were
originally popularized in the geometric modelling literature, where they were used to describe
tangent continuous C 1 curved surfaces termed ‘subdivision surfaces’. (The use of the subdi-
vision name is due to a procedural subdividing process commonly used in the construction of
these surfaces.) The subdivision shell element we develop here is isoparametric, using these

∗ Correspondenceto: George M. Turkiyyah, Department of Civil and Environmental Engineering, University of

Washington, Seattle, WA 98195-2700, U.S.A.
† E-mail:
[email protected]

Contract=grant sponsor: National Science Foundation; contract=grant number: ACI-0086093

Received 3 June 2004
Revised 13 May 2005
Copyright ? 2005 John Wiley & Sons, Ltd. Accepted 18 May 2005
758 S. GREEN AND G. M. TURKIYYAH

shape functions to describe both the geometry of the model and the solution of the thin shell
equilibrium equations.
Subdivision elements use ‘semi-local’ nodal shape functions whose support spans the two-
neighbourhood of a node. This is in contrast to the more standard
nite element shape
functions which have support over the one-neighbourhood of a node only, and thus re-
quire high-order polynomials with rotations and curvature degrees of freedom to produce
conforming elements. A triangular subdivision element has been previously described by Cirak
et al. [1, 2] which has inspired further work investigating the accuracy and eciency of these
elements [3, 4].
There are several advantages to this formulation. The absence of rotational degrees of
freedom makes it easier to perform simulations that couple thin shells to other displacement-
only continuous media. From a computational perspective, one of the advantages of subdivi-
sion elements is that the total number of degrees of freedom is only three times the num-
ber of nodes. As the model gets larger, this can result in signi
cant computational savings
for the numerical solution of the equilibrium equations, relative to formulations that include
rotational degrees of freedom. For large-scale problems where iterative solutions strategies
are necessary, and used in conjunction with sparse data structures or element-by-element
computations, subdivision elements are particularly attractive. The slight increase in element
evaluation is oset by the solution cost that generally dominates the total simulation ex-
pense. In addition, the subdivision element proposed has a natural hierarchy that can be used
for eective preconditioning. For brevity, the hierarchical aspect of the element will not be
described here.
Because the shape functions span a two-neighbourhood of a node, the size of the stiness
matrix of the element is determined by the connectivity of all faces adjacent to the element.
For the regular connectivity case, where every node is connected to exactly four neighbours,
the at rectangular version of this element is essentially the spline plate element described by
Leung et al. [5]. However, meshes which require this 4-neighbour connectivity everywhere are
too restrictive to model general geometries and local re
nement. The generalization of these
elements to arbitrary topologies and curved geometry has been dicult. The main contribution
of this paper is to show the derivation and implementation details of the general case of
n-neighbour connectivity for curved elements.

2. SUBDIVISION SURFACES

Subdivision surfaces are piecewise smooth surface descriptions which are de
ned as the limit
of an in
nite sequence of smooth re
nements applied to an initial control mesh. An example
of a subdivision re
nement hierarchy is shown in Figure 1. With each step of re
nement the
mesh converges toward a limit surface which has guaranteed smoothness properties that are
important for thin shell
nite element analysis. Many techniques, including those described in
this document, exist to work with the limit surface directly in constant time.
A de
ning characteristic of subdivision approaches is that each step of smooth re
nement
involves the addition and repositioning of vertices by averaging a neighbourhood of connected
vertices on the parent mesh. The speci
c way in which subdivision re
nement takes place is
determined by a chosen re
nement scheme. In this paper we use the Catmull–Clark scheme
[6], for quadrilateral control meshes. The vertices p at level j + 1 in a subdivision hierarchy

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
 QUADRILATERAL SHELL ELEMENT 759

Figure 1. Subdivision re
nement hierarchy.

are related to those at level j by the global subdivision matrix A

 [7] as
 jj+1 p j
p j+1 = A (1)

Naturally A is rectangular as level j + 1 contains more vertices than level j. Local neighbour-
hoods of
xed size are de
ned by the local subdivision matrix A.
Traditionally subdivision surfaces have been de
ned by the limit of a sequence of re
ned
meshes. In this paper we wish to deemphasize the subdivision nature of the representation,
and instead view geometry as the union of parameterized surface patches corresponding to
faces of a discretized mesh E


s= se (u; v); (u; v) ∈ [0; 1]×[0; 1] (2)
e∈E

Many popular subdivision schemes including the Catmull–Clark scheme generalize spline
surfaces to arbitrary topologies, and reduce to spline descriptions in areas where connectivities
are regular. Speci
cally, Catmull–Clark surfaces are tensor products of cubic splines over a
quadrilateral face whose neighbourhood consists of 4-valence vertices.

3. A QUADRILATERAL THIN SHELL FINITE ELEMENT

In this section we describe an isoparametric quadrilateral thin shell
nite element that uses
Catmull–Clark subdivision basis functions. The element described uses the Kirchho–Love
theory where the thin shell is represented by a ‘middle surface’ such that all lines initially
normal to the middle surface remain straight and normal during deformation. The deformations
may be expressed as
U = Dm d; Z = D
d (3)
where U and Z are the strains due to changes in surface area and curvature, respectively,
expressed in terms of the displacement
eld d through the respective dierential operators Dm
and D
. Mathematical background of shell theories are presented in References [8, 9].

3.1. Catmull–Clark basis functions

The Catmull–Clark basis functions generalize the one-dimensional spline basis functions to
surfaces of arbitrary connectivity. Reif and Schroder have shown that the Catmull–Clark basis

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
760 S. GREEN AND G. M. TURKIYYAH

13 14 15 16

9 10 11 12

5 6 7 8

1 2 3 4

Figure 2. Catmull–Clark
nite element and neighbourhood.

functions have square integrable principal curvature tensors, uniquely enabling their use as
shape functions in
nite element simulation of thin shells within a strict Kirchho–Love frame-
work [10]. It is worth noting that the Catmull–Clark basis functions dier from standard
nite
element basis functions in three signi
cant ways: (1) the basis functions are non-interpolating;
(2) the support of the basis functions extends over a two-neighbourhood of their associated
control node; and (3) the de
nition of the basis functions are dependent on the local connec-
tivity of the elements they aect. Although the basis functions do not directly interpolate the
control nodes, they do form a partition of unity. The values at control nodes de
ne a smooth

eld, and represent generalized displacements.
An example of a Catmull–Clark element surrounded by its control nodes is shown in
Figure 2. The central element is shaded in white, while its control neighbourhood is dark-
ened. The control neighbourhood consists of all nodes which are in the one-neighbourhood
of the element. The fact that the de
nition of the basis functions are dependent on the lo-
cal connectivity of the mesh is non-standard in
nite element implementations. Instead of a
de
ning a single element, the Catmull–Clark paradigm de
nes a family of elements for every
possible local mesh connectivity. A discussion of the representation and evaluation of these
basis functions is detailed in Sections 3.2 and 3.3.
Determining the neighbourhood of an element is an important step in evaluating the
element’s local stiness matrix. Elements may be organized into three classes by their
local connectivities. A typical example from each class is shown in Figure 3, with the
valence (number of connected edges) of each node of the central face indicated. The three
classes include regular, extraordinary and super extraordinary elements, which are discussed in
Sections 3.2, 3.3 and 3.4, respectively.

3.2. Regular elements

Figure 3(a) shows a quadrilateral element with regular connectivity. Each vertex of the central
face has a valence of four. The basis functions over this element, shown in Equation (4), are
de
ned simply by a tensor product of cubic spline polynomials. The numbering of the basis
functions corresponds to the local control mesh shown in Figure 2
Ni (u; v) = Bj (u) · Bk (v); i ∈ 1 : : : 16; j = div(i − 1; 4) + 1; k = mod(i − 1; 4) + 1 (4)

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
 QUADRILATERAL SHELL ELEMENT 761

4 4 4 4 4 5

4 4 3 4 3 4

(a) (b) (c)

Figure 3. Types of quadrilateral elements: (a) regular element; (b) extraordinary element; and
(c) super extraordinary element.

The cubic spline basis functions are simple polynomials

B1 = 16 (1 − )3 B2 = 16 (33 − 62 + 4) B3 = 16 (−33 + 32 + 3 + 1) B4 = 16 3 (5)
which sum to unity over their domain

4
Bi () = 1;  ∈ [0; 1] (6)
i=1

The tensor product of the cubic spline basis functions of Equation (4) also forms a partition
of unity. The parametric displacement d (u; v) over the element is given as the sum of the
parametric basis functions Ni weighted by nodal displacements di

16
d (u; v) = di Ni (u; v); (u; v) ∈  = [0; 1]×[0; 1] (7)
i=1

Regular elements are C 2 on their interiors and across patch boundaries. Using these shape
functions, the stiness matrix of a shell element may then be evaluated as
    @N @N 
T
C T
C  
K= (Dm N) m (Dm N) + (D
N)
(D
N)  @u × @v  du dv (8)


where Cm and C
are the membrane and curvature constitutive relations, respectively, and
the last term is the determinant of the Jacobian of the isoparametric mapping. Numerical
quadrature can be used to evaluate K.

3.3. Extraordinary elements

Elements which do not have regular neighbourhoods are known as irregular elements. Histor-
ically irregular elements are divided into two categories: extraordinary elements which con-
tain exactly one non-regular (extraordinary) valence node, and super extraordinary elements
which contain two or more irregular nodes. An example extraordinary element is illustrated in
Figure 3(b). An extraordinary element, which contains a single extraordinary node of valence

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
762 S. GREEN AND G. M. TURKIYYAH

V will have a neighbourhood of size K = 2V + 8 control nodes. In the same manner of Equa-
tion (7), the displacement at any parametric location d(u; v) over an extraordinary element
may be found as

K
d (u; v) = d i Ñi (u; v); (u; v) ∈ ([0; 1]×[0; 1])\(0; 0) (9)
i=1

A dierent set of K basis functions Ñi (u; v) exists for every possible valence V ∈ [3; ∞). The
de
nition of the Ñi (u; v) are more complicated than the Ni (u; v) of the regular case, de
ned
in Equation (4). Until recently, the de
nition of the Ñi (u; v) were not known in closed
form. Previous implementations required brute force subdivision for evaluation, relying on the
fact that each step of subdivision produces three regular and one extraordinary child regions
for each extraordinary element of the original mesh. After enough steps of subdivision, any
parametric location can be made to lie within an ordinary child element and integration could
be performed numerically over ordinary child regions and propagated back to the parent
element. However, as the desired evaluation point approaches an extraordinary vertex the
amount of work required for evaluation increases exponentially.
A closed form expression for the Ñi (u; v) in terms of the regular Ni (u; v) was
rst described
by Stam [11]. Stam’s paper provides three key results:
• Repeated subdivision is equivalent to exponentiation of the linear subdivision
operator A.
• The eigenstructure of A may be found in closed form, allowing for evaluation in
constant time.
• The basis K functions for the extraordinary element may be written as sums of translates
and dilates of the 16 basis functions of the regular element.
For each valence V , the Ñi (u; v) are sums of translated and dilated versions of the regular
basis polynomials Ni (u; v). Let N (u; v) and Ñ (u; v) be column vectors containing the regular
and extraordinary shape functions, respectively
N (u; v)T = [N1 (u; v) ; : : : ; N16 (u; v)] Ñ (u; v)T = [ Ñ1 (u; v) ; : : : ; ÑK (u; v)] (10)
The shape functions for the extraordinary element may be expressed in terms of the regular
shape functions as
Ñ (u; v) |kn = Ak; n N(tk; n (u; v)) (11)
where Ak; n is a function of the eigenstructure of A, and tk; n is an ane transformation of
variables. Each of these quantities are dependent on subscripts k and n which are functions
of the parametric location (u; v). For n ∈ [3; ∞) and k ∈ (1; 2; 3), the kn are de
ned implicitly
as functions of (u; v) as shown in Table I, along with the de
nition of the corresponding tk; n .
The de
nition of Ak; n is more complicated, and may be found in Reference [11]. The domains
kn exhaust the parametric space [0; 1]×[0; 1] with an in
nite set of tiles limiting to the point
(0; 0). This tiling is illustrated in Figure 4(a).
The extraordinary basis functions for an element with an extraordinary vertex of valence
6 are shown in Figure 5. These basis functions retain much of the character of their regu-
lar (bicubic) counterparts, and are C 2 everywhere except at the point (0; 0) where they are
formally C 1 only, yet still possess fully H 2 square integrable curvature tensors. It must be

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
 QUADRILATERAL SHELL ELEMENT 763

Table I. Extraordinary element parameterization.

Region Domain Transformation



k =1 n1 = 21n ; 2n−11
× 0; 21n t1; n (u; v) = (2n u − 1; 2n v)



k =2 n2 = 21n ; 2n−1
1
× 21n ; 2n−1
1
t2; n (u; v) = (2n u − 1; 2n v − 1)



k =3 n3 = 0; 21n × 21n ; 2n−1 1
t3; n (u; v) = (2n u; 2n v − 1)

υ υ

Ω13 Ω12
Ω4 Ω3

Ω23 Ω22
Ω1 Ω2
Ω11
Ω33 Ω32
Ω21
Ω31

(a) (b)

Figure 4. Parameterization: (a) extraordinary parameterization; and (b) super

extraordinary parameterization.

emphasized that though Equation (11) is more complicated to evaluate than standard
nite
element shape functions, for the purposes of numerical quadrature it serves as a black box
allowing the shape functions and all of their derivatives to be evaluated in constant time. Due
to the transformation of variables t (u; v) an extra Jacobian factor must be included whenever
derivatives are evaluated; as this transformation is ane, higher derivatives of t (u; v) are
not needed.

3.4. Super extraordinary elements

Figure 3(c) shows an element containing two non-regular valence vertices. One approach for
dealing with such elements is to subdivide the mesh by splitting every element into four
quadrilaterals. This isolates the non-regular vertices. All elements in a once-subdivided mesh
may be evaluated using the technique described in the previous sections. However, subdividing
all meshes for this purpose needlessly introduces elements and degrees of freedom which
greatly increase the computational cost of solution. The ability to evaluate super extraordinary
faces is quite important, and worth the added cost of implementation, especially for creating
a multilevel hierarchy.
Super extraordinary elements contain Q control nodes where
Q = 2G − 16; G = V1 + V2 + V3 + V4 (12)
and G is the sum of the individual nodal valences. Displacement over a super extraordinary
element may be written as the sum of the parametric super extraordinary basis functions

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
764 S. GREEN AND G. M. TURKIYYAH

0.5 0.5 0.5 0.5

0 0 0 0
1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
v 0 0 u v 0 0 u v 0 0 u v 0 0 u

0.5 0.5 0.5 0.5

0 0 0 0
1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
v 0 0 u v 0 0 u v 0 0 u v 0 0 u

0.5 0.5 0.5 0.5

0 0 0 0
1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
v 0 0 u v 0 0 u v 0 0 u v 0 0 u

0.5 0.5 0.5 0.5

0 0 0 0
1 1 1 1 1 1 1 1
0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5
v 0 0 u v 0 0 u v 0 0 u v 0 0 u

0.5 0.5

0 0
1 1 1 1
0.5 0.5 0.5 0.5
v 0 0 u v 0 0 u

Figure 5. Extraordinary basis functions for an element with one valence 6 node.

weighted by generalized displacements as


Q
d (u; v) = d̃i Ñi (u; v); (u; v) ∈ ]0; 1[×]0; 1[ (13)
i=1

The basis functions over these elements may be evaluated using an extension to Stam’s
approach. A super extraordinary element may be viewed as the union of four child regions,
each of which is evaluatable using techniques already discussed in this section. To evaluate a
super extraordinary element, it is sucient to evaluate the basis functions on the appropriate
child region and propagate the result to the parent without introducing new elements or degrees
of freedom to the model. The parametric domains of the four child elements are illustrated
in Figure 4(b).
Let d̃ and Ñ (u; v) be the vector of displacements for a super extraordinary element
T
d̃ = [d̃1 ; : : : ;d̃Q ]; Ñ (u; v)T = [N1 (u; v) ; : : : ; NQ (u; v)] (14)

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
 QUADRILATERAL SHELL ELEMENT 765

4 3

1 2

(a) (b) (c)

Figure 6. Super extraordinary element child regions: (a) super extraordinary element; (b) child regions;
and (c) control topology for child region 1.

Table II. Quadrilateral child region parameterization.

Region Domain Transformation Jacobian

1
1
k =1 1 = 0; × 0; t1 (u; v) = (2u; 2v) [2; 0; 0; 2]

1 2 
21 
k =2 2 = ; 1 × 0; t2 (u; v) = (2v; 2 − 2u) [0; 2; −2; 0]

21 
1 2 
k =3 3 = ;1 × ;1 t3 (u; v) = (2 − 2u; 2 − 2v) [−2; 0; 0; −2]

2 1 
21 
k =4 4 = 0; 2 × 2 ; 1 t4 (u; v) = (2 − 2v; 2u) [0; −2; 2; 0]

The displacements of the control vertices of the child regions are given as

d k = Pk Ad̃ (15)

where d k are column vectors containing the K nodal displacements of each child region
k ∈ 1 : : : 4 illustrated in Figure 6, A is the W ×Q local element subdivision matrix where W
is the size of the full once subdivided element neighbourhood, and Pk is a K ×W picking
matrix (a rectangular permutation matrix) which chooses and orders the control vertices for
each of the four child regions. The shape functions for a super extraordinary element may
therefore be written as

Ñ (u; v) |k = AT PkT Ñ(tk (u; v)) (16)

where k is a function of the parametric location and tk is a linear transformation of variables

as listed in Table II. For the purposes of evaluating derivatives the Jacobian of tk is also
given; this term is constant and, no higher derivatives are required.
The process of evaluating the Ñ is illustrated in Figure 6. A super extraordinary element and
a parametric evaluation point are shown in Figure 6(a). The eect of applying the subdivision

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
766 S. GREEN AND G. M. TURKIYYAH

operator A to generate the control neighbourhoods for all possible subregions is shown in
Figure 6(b). Noting that for this example the point (u; v) lies in child region 1, the eect of
using P1 to select the control vertices for this child region is shown in Figure 6(c). Pseudo-
code for evaluation of the shape functions is given in Algorithm 1.
Algorithm 1. Computation of super extraordinary basis functions.

function: Ñ(element, u, v)
k ⇐ Find Region(u; v) {See Table II}
(Pk ; regionk ) ⇐ Pick Child Region(element; k)
(u ; v ) ⇐ tk (u; v)
Ñ ⇐ AT ˙ PkT ˙ Ñ (regionk ; u ; v )
Q×1 Q×W W ×K K×1

As with the extraordinary elements, the de
nition of the super extraordinary basis func-
tions is more involved than those used in typical
nite element approaches. However, once
implemented, the basis functions again serve as a black box for numerical quadrature and
integration.

3.5. Convergence
We illustrate the convergence of this element by solving a well known benchmark problem
from an obstacle course of thin shell model problems [12]. The model problem of a hemi-
spherical cap stressed by four opposing radial point loads is shown in Figure 7(a). Using
symmetry boundary conditions we modelled one quarter of the hemisphere using our quadri-
lateral shell elements with a 2×2 Gauss quadrature scheme. Figures 7(b) and (c) show the
shell in the undeformed and deformed con
gurations respectively. The convergence of this
simulation is illustrated in Figure 8 as a plot of relative error in maximal displacement ver-
sus the number of elements used for both the full and quarter hemisphere models. This plot
illustrates that this element converges with almost second-order accuracy in displacements. A
full description of the application of boundary conditions and further convergence tests are
detailed in Reference [4].

P
P

P
(a) (b) (c)

Figure 7. Hemispherical shell model: (a) schematic; (b) undeformed; and (c) deformed.

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767
 QUADRILATERAL SHELL ELEMENT 767

Full and Quarter Hemisphere

1
Full
Quarter

0.1

Error
0.01

0.001

0.0001
10 100 1000 10000
Elements

Figure 8. Hemispherical shell convergence.

REFERENCES
1. Cirak F, Ortiz M, Schroder P. Subdivision surfaces: a new paradigm for thin-shell
nite-element analysis.
International Journal for Numerical Methods in Engineering 2000; 47(12):2039 –2072.
2. Cirak F, Ortiz M. Fully C 1 -conforming subdivision elements for
nite deformation thin-shell analysis.
International Journal for Numerical Methods in Engineering 2001; 51(7):813 – 833.
3. Green S, Turkiyyah G, Storti D. Subdivision-based multilevel methods for the large scale simulation of thin
shells. Seventh ACM Proceedings on Solid Modeling and Applications. ACM, 2002.
4. Green S, Turkiyyah G. Second order accurate constraints for subdivision
nite elements. International Journal
for Numerical Methods in Engineering 2004; 61(3):380 – 405.
5. Leung AYT, Au FTK. Spline
nite elements for beam and plate. Computers and Structures 1990; 37(5):
717–729.
6. Catmull E, Clark J. Recursively generated B-spline surfaces on arbitrary topological meshes. Computer Aided
Design 1978; 10(6):350 –355.
7. Schroder P, Zorin D (eds). SIGGRAPH: Subdivision Course Notes, CDROM Supplement, 2000.
8. Simo JC, Fox DD. On a stress resultant geometrically exact shell model. Part I: formulation and optimal
parameterization. Computer Methods in Applied Mechanics and Engineering 1989; 72:267–304.
9. Bathe K, Chapelle D. The Finite Element Analysis of Shells: Fundamentals. Springer: Berlin, 2003.
10. Reif U, Schroder P. Curvature integrability of subdivision surfaces. Advances in Computational Mathematics
2000; 1–18.
11. Stam J. Exact evaluation of Catmull–Clark subdivision surfaces at arbitrary parameter values. Computer
Graphics. ACM: New York, 1998; 395 – 404.
12. Belytschko T, Stolarski H, Liu WK, Carpenter N, Ong JSJ. Stress projection for membrane and shear locking
in shell
nite-elements. Computer Methods in Applied Mechanics and Engineering 1985; 51:221–258.

Copyright ? 2005 John Wiley & Sons, Ltd. Commun. Numer. Meth. Engng 2005; 21:757–767