Computers and Structures 87 (2009) 1451–1460

Contents lists available at ScienceDirect

Computers and Structures

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

A triangular six-node shell element

Do-Nyun Kim, Klaus-Jürgen Bathe *
Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA

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

Article history: We present a triangular six-node shell element that represents an important improvement over a
Received 15 February 2009 recently published element [1]. The shell element is formulated, like the original element, using the MITC
Accepted 1 May 2009 procedure. The element has the attributes to be spatially isotropic, to pass the membrane and bending
Available online 3 June 2009
patch tests, to contain no spurious zero energy mode, and is formulated without an artificial constant.
In particular, the improved element does not show the instability sometimes observed with the earlier
Keywords: published element. We give the convergence behavior of the element in discriminating membrane-
Shells
and bending-dominated benchmark problems. These tests show the effectiveness of the element.
Finite element
Triangular element
Ó 2009 Elsevier Ltd. All rights reserved.
Spatial isotropy
MITC method
Six-node element

1. Introduction uum by introducing the Reissner–Mindlin kinematical hypothesis

and the plane stress assumption for the mid-surface and the mate-
A large amount of research has been expended over the past rial layers parallel to that surface. Ideally, the shell element should
four decades on the development of shell finite elements, and yet then converge reliably and optimally to the exact solution of the
more effective triangular shell elements are still much needed, mathematical model and for any well-posed shell problem. How-
see Refs. [1–3] and the references therein. In particular, the search ever, the usual displacement interpolation leads to locking and a
for a general and uniformly effective six-node triangular shell ele- scheme needs to be used to alleviate this detrimental behavior.
ment continues, and indeed the development of such an element Successful quadrilateral general shell elements have been devel-
represents one of the remaining key challenges in finite element oped using the mixed-interpolated-tensorial-component ap-
analysis. While such an element is, in the first instance, sought proach, that is, the MITC procedure [7–11]. The advantage of this
for linear analysis, of course, the formulation should, as well, be di- approach is that the elements are general, that is, they can be used
rectly extendable to general nonlinear analysis. for general shell geometries in linear and nonlinear analyses, and
Numerous shell analyses are conducted routinely but very fine the elements have only the degrees of freedom of displacement-
discretizations and quadrilateral elements are typically used [4]. based elements with negligible additional computational cost.
An effective general curved six-node shell element would be very The MITC4 element is now widely used [4] and can also be em-
useful in that: (i) it can be employed to discretize virtually any ployed in a hierarchical manner to model additional 3D effects
shell geometry, (ii) it can be used to model shells overlaid on [12]. While tight mathematical convergence proofs of the MITC
three-dimensional solids that are represented in free-form mesh- shell elements are not available, and indeed for general geometries
ing by 10 or 11-node tetrahedral solid elements, and (iii) it would may be out of reach, the elements have been thoroughly tested on
give accurate solutions when using relatively coarse meshes. appropriate ‘discriminating and revealing’ test problems [2,11–16].
Originally, to a large extent, shell elements were developed by However, these studies largely focused on the use of quadrilateral
simply superimposing plate bending and in-plane membrane elements, equally successful general triangular shell elements are
behavior, and flat facet-shell elements were proposed. As now well more difficult to develop.
known, such elements are not truly representing shell behavior On the other hand, the family of MITC plate bending elements
and indeed may not even converge depending on which shell prob- contains quadrilateral and triangular elements that are very effec-
lem is solved [2]. The most promising formulation approach for a tive, and for plate bending solutions practically optimal [17–19].
general shell element is based on the use of the ‘‘basic shell model” Thorough mathematical convergence analyses and results of
[2,5,6]. This mathematical model is obtained from the 3D contin- numerical studies have been published, see e.g. Refs. [20–23].
However, except for the MITC4 element, the elements contain
* Corresponding author. Tel.: +1 617 253 6645. internal nodes with rotational degrees of freedom only, which ren-
E-mail address: [email protected] (K.J. Bathe). ders them not effective for extension to shell analyses and general

0045-7949/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.compstruc.2009.05.002
1452 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460

s
s
1 ert
1
err s2
s2 est
ess
eqt
eqq s3
common
s1
s1
r
r 0 r1 r3 r2 1
0 r1 r2 1

Fig. 1. Interpolations and tying points used for the MITC6 shell element; r1 ¼ s1 ¼ 12  2p1 ffiffi3 ; r 2 ¼ s2 ¼ 12 þ 2p1 ffiffi3 and r3 ¼ s3 ¼ 13

‘discriminating and revealing’ shell test problems. In particular,

Free these shell test problems include the analysis of a hyperboloid shell
with, at both ends, either clamped or totally free conditions. We con-
sider these two problems to be excellent benchmark problems to
test an element formulation for its capacity to predict membrane-
dominated and bending-dominated shell behaviors.
y
x However, additional testing of the element by Chapelle et al.
[24,25] showed a surprising element peculiarity. Namely, when
used to model certain shell geometries and boundary conditions,
the solution becomes unstable, although the single element does
not contain a spurious zero energy mode. An unphysical oscillatory
response is predicted, somewhat like observed in some solutions
θ
Clamped z with the 4/1 element of the displacement–pressure formulation
for incompressible materials [19]. Chapelle et al. stabilized the for-
mulation by replacing a part of the mixed-interpolated shear strain
energy by the unreduced displacement-based shear strain energy.
As is typical in such techniques of stabilization, a factor is intro-
duced to allocate the amount of stabilization [2,19,24]. Depending
on the shell problem solved, if the factor is too large the element
behavior deteriorates significantly and if the factor is too small,
the instability shows up. While the magnitude of the stabilizing
factor is based on some analysis, ideally, we would have a stable
and effective formulation without such factor. This is particularly
desirable when the element formulation is to be used in general
nonlinear analysis. Hence we continued our search for a more reli-
able and accurate triangular shell element.
In the search for more effective elements, the fundamental dif-
Fig. 2. Analysis of a hyperboloid shell problem. The mid-surface is given by
ference between the MITC formulation approach and the ‘en-
x2 þ z2 ¼ 1 þ y2 ; ð1  y  1Þ. The shell is fixed at its bottom and free at its top;
E = 2.0  1011, m = 1/3, t/L = 1/10000 (where t denotes the thickness of the shell, see
hanced assumed strain’, or EAS, formulation approach is
Section 4, and L = 1); the loading is the pressure loading p(h) = p0cos(2h), important [26], although, of course, they are theoretically related.
p0 = 1.0  106. The problem is solved using the original MITC6 shell element of Both techniques start with the displacement formulation and aim
Ref. [1]. to improve its predictive capability. Then, in the MITC formulation,
the strain assumptions inherently used in the displacement formu-
lation are improved by not including certain terms of the displace-
nonlinear analysis. Still, the fact that excellent MITC triangular
ment-based strain space. In this way, many MITC elements can, in
plate bending elements exist encourages the search for an effective
principle, be developed even for the same displacement assump-
MITC triangular shell element.
tions, and the key is to identify the optimal formulation. Hence,
A triangular six-node shell element based on the MITC approach
when searching for an effective six-node MITC triangular shell ele-
was recently presented by Lee and Bathe [1].1 This element has the
ment, many possibilities arise, some of which were studied in Refs.
desirable properties of not containing a spurious zero energy mode
[1,24,25].
or artificial factor, being spatially isotropic, having the same degrees
On the other hand, in the EAS formulations, new strain fields are
of freedom at every node, passing the plate bending and membrane
added to those already inherently used in the displacement formu-
patch tests, showing good convergence behavior in plate bending
lation, like first proposed by Wilson et al. with incompatible dis-
analyses, and reasonable convergence behavior in the analysis of
placement modes, see Refs. [27,19] and the references therein.
The EAS approach is generally implemented using static condensa-
1
When we refer to the MITC6 shell element of Ref. [1], we mean the MITC6a shell tion for the additional strain terms on the element level. This re-
element formulated and tested in that reference. sults into some additional cost, and complexity in nonlinear
 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460 1453

(0.0,1.0,1.2)

(0.5,0.5,0.8)

(0.7,0.3,0.5) (0.7,1.2,0.5)
(1.0,0.0,0.7)

(1.0,1.0,0.0)

Fig. 3. Isotropic element test of the six-node triangular shell element, taken from Ref. [1].

y Table 1
Basic test results of MITC6 shell elements.

Element Isotropic Zero energy Membrane Bending

(0,10) (10,10) element test mode test patch test patch test
The original MITC6 Pass Pass Pass Pass
The improved MITC6 Pass Pass Pass Pass

(4,7)
(8,7)
but also on the discriminating problems proposed in Refs. [2,13]
and used, for example, in Refs. [1,2,6,11,15,24,25]. The actual per-
formance of a shell element formulation will only be revealed
when solving these or equivalent problems and measuring the
(8,3) solution errors in appropriate norms.
(2,2) The objective in this paper is to present a further development
of the MITC6 shell element of Ref. [1]. The improved MITC6 shell
element represents a simple but effective extension of the original
x development. The element is not based on a stabilization scheme
(0,0) (10,0) and does not contain any factor to be set. The same membrane
strain and transverse shear strain interpolations as in Ref. [1] are
used, but the interpolated covariant strain components are re-
Fig. 4. Mesh used for patch tests. ferred to an element constant contravariant basis. Of course, the
geometry and the displacement-based strains used in the tying
process are calculated using the varying quantities, as defined
analysis, not present in the MITC formulations. While there exists through the discretization of the ‘basic shell mathematical model’.
potential in developing elements based on the EAS method, a diffi- For plate problems, the improved element reduces to the original
culty encountered is that stable formulations in linear analysis may element and hence the results obtained using the original and
become unstable in nonlinear analysis [28,29]. the improved elements are identical. Indeed, this is one reason
In an additional approach to obtain more effective elements, the why we use this specific interpolation of strain components.
‘discontinuous Galerkin (DG) method’ can be pursued and shell In the next sections we first briefly review the original MITC6
elements can be formulated within this framework [30,31]. In this shell element, then we present the formulation of the improved
approach, stability parameters are used and significant additional element, and finally we give the numerical results obtained in
computational cost is present, even when static condensation can the solution of the test problems. These benchmark tests include
be employed. The performance of such shell elements in nonlinear the discriminating test problems referred to above. While we con-
solutions need also still be studied. sider in this paper only linear analysis, the element formulation
An important point is that any newly formulated element can directly be extended to general nonlinear analysis, which is
should not only be tested on rather simple shell analysis problems, an inherent property of the MITC formulations [2,8,19].
1454 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460

0.0

t/L=1/100
t/L=1/1000
-0.6
t/L=1/10000

-1.2

log (relative error)

-1.8

Fig. 5. Shell problem of Fig. 2 solved with the improved MITC6 shell element. -2.4

2. The formulation of the MITC6 shell element

-3.0
As for displacement-based shell elements, the geometry of the
six-node shell element is interpolated using
X
6
t X6 -3.6
~
xðr; s; tÞ ¼ hi ðr; sÞ~
xi þ ai hi ðr; sÞ~
V in ð1Þ
i¼1
2 i¼1
-4.0
-1.8 -1.2 -0.6
where hi is the 2D interpolation function of the standard isopara-
log(h)
metric procedure corresponding to node i, ~xi is the position vector
at node i in the global Cartesian coordinate system, and ai and ~ V in Fig. 7. Convergence curves for the clamped plate problem. The bold line shows the
denote the shell thickness and the director vector at node i, optimal convergence rate.
respectively.
The displacements of the element are given by nij
X k 
~eij ðr; s; tÞ ¼ hij ðr; sÞeij ðrk ;sk ;tÞ ð5Þ
X
6
t X6
k¼1
ij ij
~
uðr; s; tÞ ¼ hi ðr; sÞ~
ui þ ai hi ðr; sÞð~
V i2 ai þ ~
V i1 bi Þ ð2Þ
i¼1
2 i¼1 where nij is the number of tying points for the covariant strain com-
k
where for node i, ~ ui is the nodal displacement vector in the global ponent ~eij and the hij are the assumed interpolation functions
Cartesian coordinate system, ~ V i1 and ~
V i2 are unit vectors orthogonal satisfying
to ~
V in and to each other, and ai and bi are the rotations of the direc- k
hij ðr lij ; slij Þ ¼ dkl ; l ¼ 1; . . . ; nij ð6Þ
tor vector ~ V in about ~
V i1 and ~
V i2 , respectively.
The covariant strain components are calculated using with dkl the Kronecker delta. This tying procedure is carried out on
the elemental level for each individual element. We next express
1
gi  ~
eij ¼ ð~ u;j þ ~
gj  ~
u;i Þ ð3Þ the displacement-based covariant strain components in terms of
2
the nodal displacements and rotations
where
eij ¼ Bij U ð7Þ
@~
x @~
u
~
gi ¼ and ~
u;i ¼ with r 1 ¼ r; r 2 ¼ s; r3 ¼ t ð4Þ where B is the strain–displacement matrix and U is the nodal dis-
@r i @ri
placement/rotation vector. Thus we obtain
The basic step in the MITC formulation is to select a set of tying "n #
Xij

points k = 1, . . ., nij on the shell mid-surface with coordinates ~eij ¼ k
hij ðr; sÞBij ðrk ;sk ;tÞ e ij U
U¼B ð8Þ
ðr kij ; skij Þ, and define the assumed covariant strain components ~ eij as k¼1
ij ij

Fig. 6. Clamped plate subjected to uniform pressure; L = 1.0, E = 1.7472  107, m = 0.3 and q = 1.0.
 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460 1455

z (a) (b)

C
β

D θ
B
Fig. 10. Meshes used for 1/8th of the hyperboloid shell (8  8 element mesh) with
symmetry boundary conditions applied. The geometry, material properties and
x A 2L y loading are as in Fig. 2. (a) The graded mesh is used when both ends are fixed and
R

(b) the uniform mesh is used when both ends are free. The boundary layer of width
pffiffi
6 t is meshed in the graded mesh [1].
α

problems. The interpolations presented in Ref. [1] are leading to

Fig. 8. Cylindrical shell problem; pressure loading p(h) = p0 cos(2h); both ends are quite an effective element and are,
either clamped or free, see Refs. [1,2]; L = R = 1.0, E = 2.0  105, m = 1/3 and p0 = 1.0
~err ¼ a1i þ b1i r þ c1i s
~ess ¼ a2i þ b2i r þ c2i s ð10Þ
The strain–displacement matrix in Eq. (8) gives the covariant ~eqq ¼ a3i þ b3i r þ c3i ð1  r  sÞ
strain components as a function of the element coordinates r, s,
and t. The constitutive tensor is defined with respect to the local for the in-plane strains, as denoted by the subscript i on the coeffi-
Cartesian coordinate system in which the plane stress assumption cients, where ~eqq ¼ 12 f~err þ ~ess g  ~ers , and
holds. Hence the assumed covariant strains in Eq. (8) are trans- ~ert ¼ a1t þ b1t r þ c1t s þ sðdt r þ et sÞ
formed into that coordinate system at each integration point to ob- ð11Þ
~est ¼ a2t þ b2t r þ c2t s  rðdt r þ et sÞ
tain the stiffness matrix. The local Cartesian coordinate axes are
Er , ~
given by (~ Es , ~
Et ) where [19] for the transverse shear strains, as denoted by the subscript t on the
~
gs ~
gt coefficients. We refer to Ref. [1] for details on how to obtain the
~
Er ¼ ~
Et ; ~
Es ¼ ~
Et  ~
Er ; ~
Et ¼ ð9Þ coefficients in Eqs. (10) and (11). The interpolations with the tying
k~
gs k k~
gt k
points used are shown in Fig. 1.
The key ingredients in the element formulation are the specific However, as mentioned above already, and reported first in Ref.
interpolations used for the membrane and transverse shear strains. [24] the resulting element shows an instability in the analysis of
Many different possibilities are available but the difficulty is to ob- certain shell problems, depending on the curvature and the bound-
tain an effective element that does not contain a spurious zero en- ary conditions of the shell structure. Fig. 2 shows this instability in
ergy mode, is spatially isotropic, passes the patch tests, and the analysis of a hyperboloid shell, clamped at the bottom and free
performs well in bending-dominated and in membrane-dominated at the top. These instabilities, even when seen only in the solution

(a) 0.0 (b) 0.0

t/L=1/100 t/L=1/100
t/L=1/1000 t/L=1/1000
-0.6 -0.6 t/L=1/10000
t/L=1/10000

-1.2 -1.2
log (relative error)
log (relative error)

-1.8 -1.8

-2.4 -2.4

-3.0 -3.0

-3.6 -3.6

-4.0 -4.0
-1.8 -1.2 -0.6 -1.8 -1.2 -0.6
log(h) log(h)

Fig. 9. Convergence curves for the cylindrical shell problem (a) when both ends are clamped and (b) when both ends are free. The bold lines show the optimal convergence
rate.
1456 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460

of certain problems, are clearly undesirable and a remedy needs to Considering shell analyses, we note that when the element size
be introduced. Chapelle et al. [24,25] discussed in depth the diffi- becomes small, the base vectors are almost constant within the
culty to obtain an improved triangular shell element that shows element, and hence the improved shell element must be expected
all the desirable properties and no instability, and presented a sta- to perform like the original element. We shall see in the tests given
bilization of the MITC6 element. However, as is typical in stabilized below that the improved element does not display the instability of
formulations [2,19] a stabilization factor is introduced. In the next the original element and performs quite well. The reason is the
section we improve the original element formulation in a different coupling of strain components used in Eqs. (12) and (13). In some
way, without the use of a factor, while preserving the other desir- respects, a more natural approach is to use membrane and shear
able properties of the original element. strain interpolations as in Eq. (5), but with different interpolation
functions and tying points than employed in the original MITC6
3. The improved MITC6 shell element shell element. Many different schemes can be explored but – with
the criteria to be satisfied – it appears difficult to reach in this way
The basic approach in this formulation is as presented above. a significantly improved shell element, see Refs. [1,25].
However, instead of using Eq. (5), we use the interpolation
nij 4. Solution of test problems
X 
~^e ðr; s; tÞ ¼ k
hij ðr; sÞ^eij ðrk ;sk ;tÞ ð12Þ
ij
k¼1
ij ij In this section we report on the performance of the improved
MITC6 shell element. As mentioned already, the element is isotro-
where
pic, hence the test of Fig. 3 is passed. The element contains only the
g^i Þð~
gk  ~
^eij ¼ ekl ð~ g^j Þ with ~
gl  ~ g^i ðr; s; tÞ ¼ ~
g i ð1=3; 1=3; tÞ ð13Þ rigid body modes, no spurious zero energy mode, and passes the
membrane and bending patch tests, see Fig. 4 and Table 1.
Here we imply summation over the indices k and l, and the Of particular interest is the solution of the problem considered
interpolation functions are those introduced in Ref. [1], see Fig. 1. in Fig. 2, to see whether spurious displacements are obtained. Fig. 5
Therefore, the same form of interpolation as given in Eqs. (5) and shows the result using the improved MITC6 shell element and we
(6), is used in the improved element, but the interpolated strains see that no spurious displacements occur.
are given in the basis ð~ g^s ; ~
g^r ; ~ g^t Þ and the ^ eij are employed instead In the further tests, we evaluate the s-norm introduced in Ref.
of the eij to evaluate the coefficients of the interpolation functions. [15] to measure the rate of convergence, since this norm can be ap-
Except for using the base vectors ð~ g^s ; ~
g^r ; ~ g^t Þ, constant in r and s, in plied in bending-dominated and membrane-dominated shell prob-
the interpolations instead of the base vectors ð~ gr ; ~
gs; ~
g t Þ, there is lems. The relative error is defined as [1,2,15]
no difference in the element formulations. Using Eq. (12) the strain
k~
uref  ~uh k2s
terms in the Cartesian basis aligned with the normal shell direction relative error ¼ ð14Þ
are calculated for use of the plane stress constitutive law. k~uref k2s
Note that, when the element is flat and straight-sided, the base where ~ uref denotes the reference solution. We consider below the
vectors ð~ gr ; ~
gs ; ~
g t Þ are constant throughout the element, and the problems solved in Ref. [1]. For each problem, we use as ~ uref the
interpolations given in Eqs. (12) and (13) reduce to those of the ori- solution obtained with a fine enough reference mesh.
ginal MITC6 element. Hence identical results are obtained when In these tests we consider only structures of constant thickness t
plate problems are solved. and t/L denotes the thickness over length ratio, as e.g. in Refs. [1,2].

(a) 0.0
(b) 0.0
t/L=1/100
t/L=1/1000
-0.6 -0.6
t/L=1/10000

-1.2 -1.2
log (relative error)

log (relative error)

-1.8 -1.8

-2.4 -2.4

-3.0 -3.0

t/L=1/100
-3.6 -3.6
t/L=1/1000
t/L=1/10000
-4.0 -4.0
-1.8 -1.2 -0.6 -1.8 -1.2 -0.6
log(h) log(h)

Fig. 11. Convergence curves for the hyperboloid shell problem (a) when both ends are clamped and (b) when both ends are free. The bold lines show the optimal convergence
rate.
 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460 1457

(a) 2.4

t/L=1/100
1.8
t/L=1/1000
1.2 t/L=1/10000

0.6

0.0

log (relative error)

-0.6

-1.2

-1.8

-2.4

-3.0

-3.6

-4.2

-4.8
-1.8 -1.2
log(h)

(b) 2.4 (c) 2.4

1.8 1.8

1.2 1.2

0.6 0.6

0.0 0.0
log (relative error)

log (relative error)

-0.6 -0.6

-1.2 -1.2

-1.8 -1.8

-2.4 -2.4

-3.0 -3.0

-3.6 -3.6 t/L=1/100

t/L=1/100
t/L=1/1000 t/L=1/1000
-4.2 -4.2
t/L=1/10000 t/L=1/10000

-4.8 -4.8
-1.8 -1.2 -1.8 -1.2
log(h) log(h)

Fig. 12. Convergence curves in the Am norm for the fully clamped hyperboloid shell problem solved using (a) the displacement-based six-node element, (b) the original MITC6
element and (c) the improved MITC6 element. Graded meshes are used as shown in Fig. 10(a).

4.1. Analysis of clamped plate problem dominated problem (clamped boundary conditions) and a bend-
ing-dominated problem (free ends) are obtained [1,2]. We solve
The plate problem considered is shown in Fig. 6 and the conver- both problems and the results are given in Fig. 9. The same good
gence results are given in Fig. 7. These results should be identical to convergence behavior as reported in Ref. [1] is seen.
those reported for the MITC6 shell element in Ref. [1], and indeed
are for individual nodal displacements. However, slight differences 4.3. Analysis of hyperboloid shell problems
in the relative errors are observed because the s-norms were calcu-
lated using different implementations. The MITC6 shell element performs very well in the analysis of
the plate and cylindrical shell problems. However, these shells
4.2. Analysis of cylindrical shell problems have rather simple surfaces, the plate is flat and the cylinder has
one principal curvature equal to zero.
The geometry and the loading of the problems are defined in Two much more discriminating problems are obtained when
Fig. 8. Depending on the boundary conditions used, a membrane- considering the hyperboloid shell shown in Fig. 2. A membrane-
1458 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460

(a) 0.0

t/L=1/100
-0.6 t/L=1/1000
t/L=1/10000

-1.2

-1.8

log (relative error)

-2.4

-3.0

-3.6

-4.2

-4.8

-5.4
-1.8 -1.2
log(h)

(b) 0.0
(c) 0.0

t/L=1/100 t/L=1/100
-0.6 t/L=1/1000 -0.6
t/L=1/1000
t/L=1/10000 t/L=1/10000
-1.2 -1.2

-1.8 -1.8
log (relative error)

log (relative error)

-2.4 -2.4

-3.0 -3.0

-3.6 -3.6

-4.2 -4.2

-4.8 -4.8

-5.4 -5.4
-1.8 -1.2 -1.8 -1.2
log(h) log(h)

Fig. 13. Convergence curves in the Am norm without shear terms for the fully clamped hyperboloid shell problem solved using (a) the displacement-based six-node element,
(b) the original MITC6 element and (c) the improved MITC6 element. Graded meshes are used as shown in Fig. 10(a).

dominated problem is obtained by considering clamped-clamped Finally we calculate the convergence curves of the element in
conditions and a bending-dominated problem is obtained when the solution of the clamped hyperboloid in the Am norm, that is,
considering both edges to be free. It is important to mesh appropri- we evaluate
ately the boundary layer in the case of the clamped case [2,32], and Am ð~
uref  ~
uh ; ~uref  ~
uh Þ
we use the meshing of Ref.p[1], relative error ¼ ð15Þ
ffiffi where half the mesh is used in the Am ð~uref ; ~
uref Þ
boundary layer of width 6 t , see Fig. 10. The very thin boundary
layer present in the free case is not specially meshed. Fig. 11 shows where Am(  ,  ) is the exact bilinear form containing the membrane
the results obtained which are quite close to those reported for the and shear strain contributions. As well known, displacement-based
original MITC6 shell element [1]. elements show excellent convergence in this norm when mem-
While the convergence behavior is quite good, of course, the brane dominated problems are solved, and indeed display optimal
element does not show optimal behavior, which would correspond behavior when properly graded meshes are used [2]. To calculate
to the optimal rate of convergence and no shift in the convergence Am and ~
uref we use the displacement-based six-node triangular shell
curves when the ratio t/L decreases. element with a mesh of 128  128 elements.
 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460 1459

(a)

(b)

(c)

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Fig. 14. Rotation magnitudes ( a2 þ b2 ) of the fully clamped hyperboloid shell
problem solved using (a) the displacement-based six-node element, (b) the original
MITC6 element and (c) the improved MITC6 element. The 16  16 graded mesh is
used with t/L=1/10000. Fig. 15. Shell problem of Fig. 2 solved with the stabilized MITC6 shell element. (a)
C = 0.1, (b) C = 0.2 and (c) C = 0.4

Since the solution ~uh obtained with the MITC6 shell element
will be different from the displacement-based solution, this mea-
Table 2
sure for convergence is very discriminating. Any small difference Normalized maximum displacements of the clamped plate problem in Fig. 6.
in the calculated shell section displacements and rotations is
Improved MITC6 t/L = 1/10000 t/L = 1/10000 t/L = 1/10000
magnified in the norm by the fact that the displacement-based
element formulation locks when solving bending dominated 44 1.020307 1.014698 1.014628
88 1.012138 1.009567 1.009512
problems.
16  16 1.006941 1.005911 1.005839
Figs. 12 and 13 show the results obtained using Eq. (15). We
Stabilized MITC6 (C = 0.1) t/L = 1/10000 t/L = 1/10000 t/L = 1/10000
also show the behaviors of the displacement-based six-node trian-
gular shell element and the original MITC6 shell element, and the 44 1.019691 0.967412 0.306133
results when excluding the transverse shear strain effects. The 88 1.012110 1.008506 0.964094
16  16 1.006935 1.005893 1.004342
figures show that reasonable convergence is measured with the
improved MITC6 shell element, and that the errors are substan- Stabilized MITC6 (C = 0.4) t/L=1/10000 t/L=1/10000 t/L=1/10000

tially less when the shear strain effects are excluded. The shear 44 1.010283 0.698071 0.026907
strain error is largely due to errors in the nodal rotations which 88 1.011874 0.996715 0.673236
16  16 1.006914 1.005626 0.994437
cause spurious shear stresses. Fig. 14 displays the rotations for
one case of number of elements used, and we see that the
improved MITC6 shell element result, compared with the original
element result, is closer to the displacement-based solution. cern using a stabilization approach is that a factor has to be set.
Hence we focus on the use of different values of the stabilization
4.4. A brief study using a stabilized shell element formulation factor.
We obtain a stabilized shell element of the original MITC6 shell
Here we want to briefly show how a formulation like the one gi- element by replacing a part of the mixed-interpolated shear strain
ven in Ref. [24] based on stabilization performs in the solution of by the unreduced displacement-based shear strain, see Refs.
the problem of Fig. 2. As pointed out already above, the major con- [2,19,24].
1460 D.-N. Kim, K.J. Bathe / Computers and Structures 87 (2009) 1451–1460

 
h hT MITC6 shell element, and Dr. D. Chapelle of INRIA, Rocquencourt,
~ert ¼ 1  C T ~eMITC6 þ C eDI
L rt L rt France, for his comments on this work.
  ð16Þ
h hT
~est ¼ 1  C T ~eMITC6 þ C eDI
L st L st References
where ~eMITC6
rt and ~eMITC6
st are the shear strains calculated from the
[1] Lee PS, Bathe KJ. Development of MITC isotropic triangular shell finite
MITC6 strain interpolation in Eq. (11) and eDI DI
rt and est are the strains elements. Comput Struct 2004;82:945–62.
obtained from the displacement interpolation in Eq. (3). In Eq. (16) [2] Chapelle D, Bathe KJ. The finite element analysis of shells-
C is the stabilization factor to be set, hT is a measure of the element fundamentals. Springer; 2003.
[3] Lee PS, Noh HC, Bathe KJ. Insight into 3-node triangular shell finite elements:
size and L is the characteristic length. For the problems we consider, the effects of element isotropy and mesh patterns. Comput Struct
L = 1 and we use hT to be the radius of the circumscribed circle 2007;85:404–18.
around the corner points of the triangular element. This stabiliza- [4] Bathe KJ. The finite element method. In: Encyclopedia of computer science and
engineering. J. Wiley and Sons; 2009. p. 1253–64.
tion operates on the transverse shear strains whereas the procedure [5] Chapelle D, Bathe KJ. The mathematical shell model underlying general shell
of Ref. [24] operates on the shear strain energy. We expect that a elements. Int J Numer Meth Eng 2000;48:289–313.
similar stabilization is achieved with the two techniques. [6] Lee PS, Bathe KJ. Insight into finite element shell discretizations by use of the
basic shell mathematical model. Comput Struct 2005;83:69–90.
Fig. 15 shows the deformations of the shell considered in Fig. 2 [7] Dvorkin E, Bathe KJ. A continuum mechanics based four-node shell element for
when three different values of C are used. As seen, the deforma- general nonlinear analysis. Eng Comput 1984;1:77–88.
tions of the shell are quite sensitive to the value of C, but once [8] Bathe KJ, Dvorkin EN. A formulation of general shell elements – the use of
mixed interpolation of tensorial components. Int J Numer Meth Eng
the stabilization factor is large enough, the instability of the origi-
1986;22:697–722.
nal MITC6 element is no longer present. Hence it appears that sim- [9] Bucalem ML, Bathe KJ. Higher-order MITC general shell elements. Int J Numer
ply a large enough value of C needs to be selected. Meth Eng 1993;36:3729–54.
However, clearly, if the stabilization factor is too large, the error [10] Bucalem ML, Bathe KJ. Finite element analysis of shell structures. Arch Comput
Meth Eng 1997;4:3–61.
in the response prediction (displacements and stresses) is signifi- [11] Bathe KJ, Lee PS, Hiller JF. Towards improving the MITC9 shell element.
cant, see Ref. [24]. We demonstrate this deterioration of the re- Comput Struct 2003;81:477–89.
sponse prediction in Table 2 for the analysis of the clamped plate [12] Kim DN, Bathe KJ. A 4-node 3D-shell element to model shell surface tractions
and incompressible behavior. Comput Struct 2008;86:2027–41.
problem of Fig. 6. Here the stabilization is not needed but when [13] Chapelle D, Bathe KJ. Fundamental considerations for the finite element
used with values as in Fig. 15, the response prediction is much analysis of shell structures. Comput Struct 1998;66:711–2.
deteriorating. [14] Bathe KJ, Iosilevich A, Chapelle D. An evaluation of the MITC shell elements.
Comput Struct 2000;75:1–30.
Hence a major difficulty when using this stabilization approach [15] Hiller JF, Bathe KJ. Measuring convergence of mixed finite element
is to choose the optimal stabilization factor automatically for each discretizations: an application to shell structures. Comput Struct 2003;81:
element for any shell analysis, including nonlinear analysis. This is 639–54.
[16] Bathe KJ, Iosilevich A, Chapelle D. An inf-sup test for shell finite elements.
hardly possible but assuming that it is achieved, we may find Comput Struct 2000;75:439–56.
thereafter that the accuracy of the solution is not acceptable. [17] Bathe KJ, Dvorkin E. A four-node plate bending element based on Mindlin/
Reissner plate theory and a mixed interpolation. Int J Numer Meth Eng
1985;21:367–83.
5. Conclusions [18] Brezzi F, Bathe KJ, Fortin M. Mixed-interpolated elements for Reissner/Mindlin
plates. Int J Numer Meth Eng 1989;28:1787–801.
The objective in this paper was to present a triangular shell ele- [19] Bathe KJ. Finite element procedures. Prentice-Hall; 1996.
[20] Bathe KJ, Brezzi F, Cho SW. The MITC7 and MITC9 plate bending elements.
ment which represents a significant improvement over an earlier Comput Struct 1989;32:797–814.
published element [1]. Like the earlier presented element, the im- [21] Iosilevich A, Bathe KJ, Brezzi F. On evaluating the inf-sup condition for plate
proved six-node element is based on the MITC formulation ap- bending elements. Int J Numer Meth Eng 1997;40:3639–63.
[22] Lyly M, Niiranen J, Stenberg R. A refined error analysis of MITC plate elements.
proach and has all the attractive attributes of MITC shell Math Models Meth Appl Sci 2006;16:967–77.
elements, with respect to ease of use and computational effective- [23] Lyly M, Niiranen J, Stenberg R. Superconvergence and postprocessing of MITC
ness. Actually, the changes in the formulation of the earlier ele- plate elements. Comput Meth Appl Mech Eng 2007;196:3110–26.
[24] Beirão da Veiga L, Chapelle D, Paris Suarez I. Towards improving the MITC6
ment to reach the improvements are quite simple.
triangular shell element. Comput Struct 2007;85:1589–610.
The formulation of the improved MITC6 shell element given here [25] Chapelle D, Paris Suarez I. Detailed reliability assessment for triangular MITC
specifically addresses the peculiar unstable behavior reported in Ref. elements for thin shells. Comput Strut 2008;86:2192–202.
[26] Simo J, Rifai MS. A class of mixed assumed strain methods and the
[24] observed with the earlier published six-node element [1] in the
method of incompatible modes. Int J Numer Meth Eng 1990;29:
solution of certain shell problems. Specific shell geometries and 1595–638.
boundary conditions allow the instability to occur. The improved [27] Wilson EL, Ibrahimbegovic A. Use of incompatible displacement modes for the
MITC6 shell element does not show this behavior and in the other calculation of element stiffness and stresses. Finite Elements Anal Des
1990;7:229–41.
test problems performs practically as well as the earlier published [28] Wriggers P, Reese S. A note on enhanced strain methods for large
element. In plate analyses the same results as earlier are obtained. deformations. Comput Meth Appl Mech Eng 1996;135:201–9.
While the shell element does not show uniformly optimal behav- [29] Pantuso D, Bathe KJ. On the stability of mixed finite elements in large
strain analysis of incompressible solids. Finite Elements Anal Des 1997;28:
ior in all analyses, a property that is extremely difficult to reach [2], 83–104.
the element shows good convergence behavior. A mathematical [30] Güzey S, Stolarski HK, Cockburn B, Tamma KK. Design and development of a
analysis of the discretization scheme would be very valuable and discontinuous Galerkin method for shells. Comput Meth Appl Mech Eng
2006;195:3528–48.
could yield insight into how the element might be further improved. [31] Güzey S, Cockburn B, Stolarski HK. The embedded discontinuous Galerkin
method: application to linear shell problems. Int J Numer Meth Eng
Acknowledgements 2007;70:757–90.
[32] Pitkäranta J, Sanchez Palencia E. On the asymptotic behavior of sensitive shells
with small thickness. Comptes Rendus de l’Academie des Sciences Serie II
The authors would like to thank Prof. P. S. Lee of KAIST, Korea, Fascicule B-Mechanique Physique Chimie 1997;325:127–34.
for sharing his experiences in the implementation of the original