Comput. Methods Appl. Mech. Engrg.

191 (2002) 3775–3810

www.elsevier.com/locate/cma

Element formulation and numerical techniques for

stability problems in shells
Anders Eriksson *, Costin Pacoste
Department of Mechanics, Royal Institute of Technology, KTH, SE-100 44 Stockholm, Sweden
Received 1 March 2001; received in revised form 12 February 2002; accepted 6 March 2002

Abstract

In the context of instability problems in shells or shell like structures, the objective of the present paper is twofold.
Primarily, the paper describes how quasi-static, conservative instability problems can be considered in a multi-para-
metric context, where generalized path-following procedures for augmented equilibrium problems are used as com-
putational tools. These allow systematic treatment of the higher-dimensional solution sets generated under the
variations of certain parameters deemed relevant for the given problem. The efficient implementation of the above
mentioned procedures requires however, as an essential ingredient, a non-linear finite element which is not only ac-
curate but also inexpensive. To this end, a systematic view on a corotational Total Lagrangian formulation is described.
The TRIC element of Argyris and coworkers is slightly modified, and introduced as core element formulation. Special
emphasis is given to the alternative methods for treatment of finite three-dimensional rotations, with reference to both
the element definition and solution algorithms. Numerical examples verify the element capabilities, and the possi-
bility to completely describe instability phenomena of large, discretized models. Ó 2002 Elsevier Science B.V. All
rights reserved.

1. Introduction

The development of an accurate and efficient analysis strategy for instability problems in shells or shell
like structures requires the solution of two interconnected problems. The first one is related to the critical
and post-critical behavior of the idealized structure while the second is related to the sensitivity of the
critical behavior to variations in the geometrical and mechanical parameters which define the structure.
The solution to the first problem involves, on one hand, the detection and isolation of critical states on
the fundamental path of the perfect structure, but also an accurate prediction of the post-critical behavior,
i.e. secondary bifurcations, mode jumping and so on. The analysis is based on idealized models with respect
to geometrical, mechanical and loading parameters and the equilibrium-paths of the system are normally

*
Corresponding author. Fax: +46-8-796-9850.
E-mail address: [email protected] (A. Eriksson).

0045-7825/02/$ - see front matter Ó 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 0 4 5 - 7 8 2 5 ( 0 2 ) 0 0 2 8 8 - 8
3776 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

traced using incremental-iterative algorithms. The underlying viewpoint is that the behavior of the system is
a function of a single variable, a load factor.
For the second problem, however, the above mentioned viewpoint is scarcely relevant, since it leads to
equilibrium paths which only correspond to certain sections of a much more complex equilibrium manifold
in a multi-parametric space. An efficient solution method in this context requires a different type of ap-
proach. Thus, the response should no longer be seen solely as a function of only the load variable for a
specified structure, but rather as a function of several variables; apart from the load, additional variables
describing geometrical or mechanical properties of the structure, are included. A specific structure is hereby
viewed as an instance of a larger family, correspondingly parameterized.
Along this line of work, the aim of the present paper is to describe how quasi-static, conservative in-
stability problems can be analyzed in a multi-parametric setting using augmented equilibrium systems. As
tools for the multi-parametric analyses, generalized incremental-iterative algorithms are used. The treated
augmented systems contain auxiliary conditions defining subsets of equilibrium states which share a
common property, e.g. criticality. The auxiliary conditions will allow the addition of extra control pa-
rameters interesting for the sensitivity investigations. Specific details of the basic solution methods are
discussed to some extent in [1,2].
Considering the complexity and the setting of the desired instability investigations, an accurate and
efficient shell element is a prerequisite. For these reasons, the paper also discusses some theoretical and
numerical aspects related to the formulation of a non-linear facet triangular element. The method of de-
scription is corotational, which here means that the motion of an element is decomposed into a rigid body
motion accompanied by a pure deformation through the use of a local coordinate system, which contin-
uously translates and rotates with the element. Following mainly ideas from [3], the definition of the ele-
ment resorts to a systematic change of variables from the local frame to the global one. This is done
through the use of a projector matrix which extracts the rigid body modes from the global displacements.
At the level of the element definition, one issue of special relevance is the parameterization of finite 3D
rotations. Two distinct alternatives, previously discussed in literature [4–6], are explored for the present
purpose. The first one uses a three-parameter representation, based on the rotation vector. Compared with
the formulation in [3], this parameterization requires an additional change of variables defined by the linear
operator which relates the spatial components of angular variations to variations of the rotation vector [4].
In order to allow rotations of any magnitude, a vector like parameterization within each increment is also
explored [5]; this is based on the spatial form of the incremental rotation vector. Consequently, additive
updates will still apply but this time only within an increment. The transformation matrices for the two
cases are the same, considerably simplifying the implementation and comparison of the two alternative
parameterizations. The use of additive rotation variables is essential for the multi-parametric analyses [2,7].
The main feature of the adopted corotational formulation is its independence of the core assumptions
used to derive the internal forces and tangent stiffness in local coordinates. Using this property, any type of
core formulation can be adopted. The option considered in this paper is based on the linear TRIC element
of Argyris et al. [8–10]. Defined using the natural mode concept, this element is accurate and efficient in
many problems. As the element, however, shows a less accurate behavior for certain classes of problems, a
version, here named MTRIC, with slightly modified membrane behavior has been developed and imple-
mented.
The organization of this paper is as follows: The issues related to the element formulation are discussed
in Section 2, while Section 3 is a brief description of the multi-parametric setting of instability problems and
numerical solution methods. A set of numerical examples is presented in Section 4. The examples dem-
onstrate the ability of the shell element to deal with problems involving arbitrarily large displacements and
rotations, but also the ability of the multi-parametric viewpoint to demonstrate complex instability phe-
nomena and its capacity in parameter sensitivity analyses. Conclusions and suggestions for future work are
presented in Section 5.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3777

2. Element formulation

The objective of this section is to discuss the formulation of a corotational facet triangular shell element.
The central idea in this context is to introduce a local coordinate system, which continuously rotates and
translates with the element. The remaining local displacement coordinates
qeI are then expressed as func-
tions of the total, according to

qeI ðqgJ Þ
qeI ¼
ð1Þ
qeI
for ðI; J ¼ 1; 2; 3Þ. The local displacements thereby represent the deformational quantities in the element


system, and are used to formulate the strain energy function U ¼ Uð
qeI Þ. The contribution to the internal
force vector and tangent stiffness matrix in the global degrees of freedom are then obtained through suc-
cessive differentiation, according to

 ( )
g oU oU o
qeI
f ¼ ¼ ;
oqgJ qeI oqgJ
o

 2  " 2 # ð2Þ
g oU o U o
qeJ oU o2
qeI
qeI o

k ¼ ¼ þ e g g ;
oqgK oqgL o
qeI o
qeJ oqgL oqgK o
qI oqL oqK

where the corotational framework can be seen as a chain-rule transformation. The above expressions
contain the differentials of the strain energy with respect to the local displacements:
( ) " #
e oU e o2 U
f ¼ ; k ¼ ð3Þ
qeI
o
qeI o

o
qeJ

which give the expressions for the internal forces and tangent stiffness in local coordinates. These ex-
pressions depend on the definition of the local strains, i.e. strains expressed for the core element. They also
contain the partial derivatives of local versus global displacements, giving the transformation matrices
necessary in order to re-express f e and ke in global coordinates:
" # " #
qeI
o
o2

qeI
Ae ¼ ; AeI ¼ : ð4Þ
oqgJ oqgL oqgK
These matrices, which actually define the corotational framework, depend on the non-linear functions in
Eq. (1), which in turn depend solely on the choice of the local coordinate system. Thus, various corotational
elements defined using different core assumptions but the same type of local coordinate system will share
the same transformation matrices; the corotational formulation as described is element independent.
The term corotational has been used in many contexts, and with many different interpretations [11]. The
term as used above agrees with several other approaches, but is perhaps most adequately described by the
terms ‘ghost reference’ [12], and ‘(continuously) adapted reference’ [13], formulations. In the above form, it
is obvious that the method is really a total Lagrangian formulation, but using a coordinate transformation
in the description.
With the form chosen, the local ‘core’ element formulation can use any conjugate pair of stress and
strain measures. Seeing this exactness in the core element formulation, it is further obvious that the only
source of error lies in the discontinuity of the element connections at the nodes. Although not creating any
problems for the linear core element proposed here, any non-linear element formulation must be ascer-
tained to state the element contributions to forces and stiffness at coinciding geometrical positions, elim-
inating the energy stored in the fictitious gaps between connected element nodes.
Following mainly [3], the change of variables in (1) and the transformation matrices in Eq. (4) are
constructed using a projector matrix, which subtracts rigid body modes from the total increment of motion.
3778 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

The main difference from [3] lies in the parameterization of finite 3D rotations. Two distinct alternatives
from literature were explored in the present work. The first one uses a three-parameter representation of
finite rotations based on the rotation vector [4]. The main advantage here is that the rotation variables are
additive, which avoids a special updating procedure. The second alternative was also based on a vector like
parameterization, but only within each increment [5]. Additive updates will still apply but this time only for
iterative corrections within an increment. In both cases, an additional change of variables is required
compared to the formulation in [3]. The issues pertaining to the proposed parameterization of finite ro-
tations will be addressed in the following subsection.

2.1. Parameterization of finite 3D rotations

One of the central issues in the development of a non-linear shell element is the parameterization of the
orthogonal matrix R, used in order to represent large 3D rotations. This matrix has nine components, but
can be represented in terms of only three independent parameters. One frequently used alternative [4,6,14–
16], is based on the so called rotation vector, defined by
e Þ;
R ¼ expð W W ¼ ew; ð5Þ
where the tilde denotes the skew-symmetric matrix obtained with the components of the vector. The
geometrical significance of the above definition is that any finite rotation can be represented by a unique
rotation with an angle w about an oriented axis defined by the unit vector e. The properties of these
representations, and other ways of expression, are extensively discussed in [6].
The rotation matrix R maps an orthonormal Cartesian frame E i into another, ti . An incremental ro-
tation, which carries the moving frame ti into a new position t0i , can be expressed through a spatial rotation
W:
t0i ¼ Rs ti ¼ Rs RE i ; fÞ
Rs ¼ expð W ð6Þ
which affects the already rotated axes. It can also be seen as a material rotation H, applied to the original
axes, before rotating by R:
t0i ¼ R0 E i ¼ RRm E i ; e Þ;
Rm ¼ expð H ð7Þ
where the spatial and material incremental rotation vectors are related by W ¼ RH. Variations to the
rotation matrix dR can then be computed as either of [6]:
d d d
dR ¼ ½Re e¼0 ¼ ðexpðex ~ R ¼ ðR expðe~hÞÞe¼0 ¼ Rd~h;
~ ÞRÞe¼0 ¼ dx ð8Þ
de de de
where dx denotes spatial and dh material angular variations, i.e. infinitesimal rotations superimposed onto
the existing R. The parameterization of admissible variations dR is thereby independent of the technique
used to parameterize R itself. Consistent parameterizations, dx or dh, must only be projected onto the
parameter space adopted for R. This issue, which is the key observation for the implementation and
evaluation of different rotation parameterizations, will be addressed in the remainder of this subsection. As
the basic definitions of these descriptions are given in literature, the discussion will focus on the imple-
mentation of these into the present context.

2.1.1. Rotation vector

When R is described by a rotation vector W, Eq. (5), and considering the material form of Eq. (8),
variations dW to the rotation vector are found from
e e Þ ¼ R expðe d~
Re ¼ expð W hÞ; We ¼ W þ e dW ð9Þ
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3779

which leads to

expðe d~ e Þ expð W
hÞ ¼ expð W e e Þ: ð10Þ
Further treatment [4–6], gives the key result:
dh ¼ T m ðWÞdW ð11Þ
with
2
sin w sin w 1 sinðw=2Þ e:
T m ðWÞ ¼ I3 þ 1  eeT  W ð12Þ
w w 2 ðw=2Þ

The spatial angular variations are correspondingly given by

dx ¼ T s ðWÞdW ð13Þ
with
2
sin w sin w 1 sinðw=2Þ e:
T s ðWÞ ¼ RT m I3 þ 1  eeT þ W ð14Þ
w w 2 ðw=2Þ

Evaluating the determinant of T s ¼ T Tm shows that the corresponding mappings cease to be bijections for
w ¼ 2kp, k ¼ 1; 2; . . . [6,16]. Avoiding these points, the inverse mappings are well defined, and can be
computed according to
ðw=2Þ ðw=2Þ e;
T 1
m ¼ I3 þ 1  eeT þ 12 W ð15Þ
tanðw=2Þ tanðw=2Þ

ðw=2Þ ðw=2Þ 1e
T 1
s ¼ I3 þ 1  eeT  W : ð16Þ
tanðw=2Þ tanðw=2Þ 2

2.1.2. Incremental rotation vector

For a stepwise defined rotation, a new and an old rotation matrix are related by [14]:
e Þ:
Rn ¼ Ro expð H ð17Þ
Considering the material form of Eq. (8), admissible variations dH, which lead to the same perturbed
matrix RnðeÞ , are found from
e e Þ ¼ Rn expðe d~
Ro expð H hÞ; He ¼ H þ e dH: ð18Þ
Using Eq. (17) gives:
expðe d~ e Þ expð H
hÞ ¼ expð H e e Þ: ð19Þ
Following a procedure entirely similar to the one used in the preceding subsection leads to
dh ¼ T m ðHÞ dH ð20Þ
and, similarly [5]:
dx ¼ T s ðWÞdW: ð21Þ
In the equations, the increments H and W are seen as additions to a known old rotation state W, which has
significance on the storage demands in a computational formulation.
3780 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

2.2. Coordinate systems, element kinematics

The development presented in this subsection is based on the use of a local coordinate system, which
continuously rotates and translates with the element, as shown in Fig. 1. The origin of the local system is
taken at point C, which is the centroid of the triangle. The orientation of the local frame in the initial
configuration X 0 is specified by the orthogonal matrix Ro .
The motion of the element from the initial to the final deformed configuration can be seen as split into
two stages. First, a rigid translation and rotation of the local element frame is considered. The rigid
translation is defined by the displacement of point C, denoted by ugc , expressed in the global reference frame
X g . The rigid rotation is such that the new orientation of the local frame X e is defined by an orthogonal
matrix Rr , which can be chosen with some freedom. This rigid motion is accompanied by local, deforma-
tional, displacements with respect to the element axes.
Using the notation defined in Fig. 1, the displacement of an element point P within the local element
frame is given by

uep ¼ RTr ðrgp þ ugp  rgc  ugc Þ  rop ;

ð22Þ
T
where
uep ¼ ½

uep ;
vep ; w

ep  , and the over-bar is used to denote a deformational kinematic quantity. This ex-
pression for the element nodes p ¼ I ðI ¼ i; j; kÞ gives the local nodal translations ð
uei ;
uej ;
uek Þ.
The global rotations at element point P can be written in terms of the rotation of the local axes, defined
by Rr , accompanied by a local rotation relative to these axes; this is defined by another orthogonal matrix
e
Rp . The latter should be viewed as a material rotation, as the deformations are seen in relation to the initial
geometry. Consequently, the orientation of the local triad X pi at point P can be obtained by means of the

Fig. 1. Corotational formulation: element kinematics and coordinate systems.

 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3781

e
product Rr Rp . On the other hand, this orientation can also be obtained through the product Rp Ro with Rp
denoting the rotation matrix at point P. Thus, utilizing the orthogonality of Rr :
e
Rp ¼ RTr Rp Ro ð23Þ

and the local rotations are obtained from

~ e ¼ log ðRT R R Þ:


# ð24Þ
p e r p o

Evaluating Eqs. (22) and (24) for the nodes p ¼ I ðI ¼ i; j; kÞ, the deformational rotations at the element
nodes are obtained. Using these nodal values together with appropriate interpolation schemes gives the
strain energy according to
uei ;

U ¼ Uð
uej ;

e; #
uek ; #
e; #

e Þ: ð25Þ
i j k

The local internal forces f e and tangent stiffness ke can then be computed through successive differentiation.
The resulting blocks are of the forms, for I and J ðI; J ¼ i; j; kÞ:
2 3 2 2 3
oU " # oU o2 U
6 o

ueI 7 6 o
e
e 7
ueJ o
ueI o#
7¼ n
eI 6 uI o
J 7
f eI ¼ 6
4 oU 5 ; ke
¼ 6 7: ð26Þ

eI
m IJ
4 o2 U o2 U 5

e
o#
e o

o# ue o#
e o#
e
I I J I J

Any formulation for a triangular element, linear or non-linear, can be placed at the core of the corotational
formulation and tested for efficiency and accuracy, remembering the connective description of the element
as discussed above. The development of a specific linear core element will be addressed in a separate
subsection.

2.3. Choice of local system

It is obvious from the previous section that there are some degrees of choice for the transformations ugc
and Rr . As the local element formulation is not affected by any rigid translation, the key point in the
evaluation of the local displacements according to Eqs. (22) and (24) is the definition of the orthogonal
matrix which specifies the orientation of the moving frame. For this purpose, let ei denote the orthonormal
basis vectors of the local frame in the current deformed configuration. The required orthogonal matrix is
then given by Rr ¼ ½e1 e2 e3 .
One possible choice is to take e1 as parallel to the side i  j of the element in its current configuration.
The three axis vectors are then
xij xij xik
e1 ¼ ; e3 ¼ ; e2 ¼ e3 e1 ; ð27Þ
kxij k kxij xik k
where
xij ¼ rgj þ ugj  rgi  ugi ; xik ¼ rgk þ ugk  rgi  ugi : ð28Þ
This choice leads to an element formulation which is dependent on the node ordering. In the present im-
plementation, the local frame is instead re-oriented by a procedure first described in [17]. The procedure
minimizes the Euclidean norm of the local nodal translation components:
X X
ueT

ueI ¼ min )
I
d
ueT
I
ueI ¼ 0: ð29Þ
I¼i;j;k I¼i;j;k
3782 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

The variation of the local nodal displacements with respect to variations in the orientation of the local
frame can be computed using Eq. (22) and the second of Eq. (8):
ueI ¼ d~
d
hr RTr ðrgI þ ugI  rgc  ugc Þ ¼ d~
hr reI ; ð30Þ
where reI is found from Fig. 1. Using the above expression in the second of Eq. (29) gives
" #
X X
e e T e o
uI ¼ dhr
uI T

d
~rI rI ¼ 0: ð31Þ
I¼i;j;k I¼i;j;k

Since dhr is arbitrary, the above equation leads to the condition

X
~reI roI ¼ 0: ð32Þ
I¼i;j;k

Thus, starting with an initial orientation of the local frame as described at the beginning of this subsec-
tion, a solution to Eq. (32) can be obtained through a Newton procedure. The system to be solved for the
iterative corrections is found by taking the variations of Eq. (32):
" #
X X X
e o e o e o
dð~rI rI Þ ¼ dðrI rI Þ ¼  ~rI ~rI dhr ¼ Kr dhr : ð33Þ
I¼i;j;k I¼i;j;k I¼i;j;k

Once the above equation is solved, a new orientation of the local frame is obtained as
Rr :¼ Rr expðd~
hr Þ ð34Þ
which is repeated until the correction is negligible.

2.4. Internal forces and tangent stiffness

The transformation matrices used to convert the internal forces and tangent stiffness from local to global
coordinates need the variations of Eqs. (22) and (23):
uep ¼ dRTr ðrgp þ ugp  rgc  ugc Þ þ RTr ðdugp  dugc Þ;
d
ð35Þ
e
dRp ¼ dRTr Rp Ro þ RTr dRp Ro : ð36Þ
The variations of the orthogonal matrix Rr in the above equations are computed according to the second of
Eq. (8). Thus
dRr ¼ Rr d~
hr : ð37Þ
With this, Eq. (35) becomes, under the assumption that the core element is invariant under pure translation
[3]:
uep ¼ d~
d
hr rep þ duep ¼ ~rep dhr þ duep ; ð38Þ

where rep ¼

uep þ rop , according to Eq. (22).
Similarly, the variation of the rotation matrix can be found as
e e
dRp ¼ ½d~
hep  d~
hr RTr Rp Ro ¼ ½d~
hep  d~
hr Rp ð39Þ

which can also be written

dRp ¼ d~
e e e
h
p Rp ; d

hep ¼ dhep  dhr : ð40Þ
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3783

Further, the transformation matrix Ae , Eq. (4), can be evaluated according to

Ae ¼ PE T ; ð41Þ
where E ¼ diagðRr ; Rr ; Rr ; Rr ; Rr ; Rr Þ, while the 6 6 blocks of the matrix P, coupling nodes I and J, are
obtained as
2 e 3
o

uI o
ueJ
6 oue ohe 7
6 J J 7
PIJ ¼ 6 e
e 7: ð42Þ


4 oh o h I 5
I
oueJ
oheJ
Using Eqs. (38) and (40):
2 3
e ohr ohr
6 I 3 dIJ þ ~rI oueJ ~reI
oheJ 7
PIJ ¼ 6
4
7 ¼ I 6 dIJ  AI G T ; ð43Þ
ohr ohr 5 J
 e I 3 dIJ  e
ouJ ohJ
where the blocks:
2 3
ohr
   1 T 
~reI 6 oueJ 7 o
AI ¼ ; GJ ¼ 6 7 ¼ ðKr ~rJ Þ ð44Þ
I3 4 ohr 5 03
oheJ
using Eq. (33)
Constructed as above, P is a projector matrix, i.e. P ¼ P2 , which extracts the deformational part from
the total increment of motion [3,17].
Using the previously introduced equations, the global internal element forces are computed as
f g ¼ ATe f e ¼ EPT f e : ð45Þ
Noting the derivation above, the specific content of the general Eq. (26) to be used, can be written:
2 3
oU
 e
6 o

ueI 7
f eI ¼ 6 7 ¼ nIe ð46Þ
4 oU 5 mI
o

heI
to be consistent with the definition of P in Eq. (42). Further, the tangent stiffness matrix in global coor-
dinates can be computed by taking the variations of Eq. (45), which, after some algebra, yields
kg ¼ E½PT ke P  GF T1 P  F 2 G T E T ; ð47Þ
where the blocks:
2 2 3
oU o2 U
6 o
e
ueJ o
heJ 7
ueI o

kIJ ¼ 6
e u I o
7 ð48Þ
4 2 5
o Ue o2 U
oheI o

uJ o

hI o

e
hJe

and
e eT
F 1 ¼ ½N 03 e eT
N 03 e eT
N 03  ;
T
i j k
ð49Þ
e eT
F 2 ¼ ½N f eT
M e eT
N f eT
M e eT
N f eT T :
M
i i j j k k
3784 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

In the above equations, N eI and M eI are the block components of the vector
T
PT f e ¼ ½N eT
i M eT
i N eT
j M eT
j N eT
k M eT
k  ð50Þ
which are the quantities compatible with the used element variations.

e ! d

2.5. Change of variables: d# he

Comparing the expressions for the local forces in Eqs. (26) and (46), it is noted that the differentiation
is performed with respect to different types of rotation variables. As d#
e ¼ T 1 ð#

e Þd
he , Eq. (46) can be
s
reformulated as the blocks
2 3
2 3 oU
I 3 03 T 6 e 7
f eI ¼ 4
e 5 6 o

uI 7 T
03
o# I 4 oU 5 ¼ hI f I ð51Þ


ohIe

e
o# I

with a transformation for each node as

 
I3 03
hI ¼
e ð52Þ
03 T 1
s ð #I Þ

and f I the blocks of a vector f

Through a further differentiation, the corresponding transformation equation for the tangent stiffness
matrix is obtained as
ke ¼ hT kh þ kh ; ð53Þ
where k is, similarly as in Eq. (51), obtained as the blocks
2 2 3
oU o2 U
6 o
e e
e 7
ueI o#
6 u o
u o
J 7
kIJ ¼ 6 I2 J 2 7 ð54Þ
4 oU oU 5

e o
ue o#
o#
e o#

e
I J I J

and the matrix h has

h ¼ diagðhi ; hj ; hk Þ ð55Þ
and
kh ¼ diagð03 ; khi ; 03 ; khj ; 03 ; khk Þ: ð56Þ
The three diagonal 3 3 blocks khI for (I ¼ i; j; k), are found from
o h i
½T T
v ¼ g½
e vT  2v#
#
eT þ ð#

eT þ #
eT vÞI 3  þ l#

e2 ½v#

eT   1~v T 1 ð#

e Þ; ð57Þ
o

he s 2 s

where
2 sin #
e  #
e ð1 þ cos #
e Þ #
e ð#
e þ sin #
e Þ  8 sin2 ð#
e =2Þ
g¼ ; l¼ ð58Þ
2#
e2 sin #
e 4#
e4 sin2 ð#
e =2Þ
with #
e ¼

heI , #
e ¼ k#

e k, and v ¼ m
eI .
I
At least for small increments, #
is a small quantity. Consequently, T 1 ð#
e
e Þ is close to identity. For these
s
reasons, the transformations described by Eqs. (51) and (53) are often omitted in the formulation of co-
rotational elements [18]. For accurate stability evaluations and fast convergence, the full matrix is, however,
desirable, and has been implemented in the presented work.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3785

2.6. Finite rotation parameters

The procedure described in the previous subsections for computing the internal forces and tangent
stiffness in global coordinates is summarized as

W
Eq: ð51Þ Eq: ð45Þ f
f ! fe ! fg !
fW

W
Eq: ð53Þ Eq: ð47Þ k
k ! ke ! kg !
kW
where the last step allows a choice of rotation parameters, as discussed below.
Computed as above, f g and kg are consistent with spatial angular variations dx, as defined in the first of
Eq. (8). Consequently, the updating of the rotation matrix for each node of the structure at each iteration
step is done according to Rn ¼ expðdx ~ ÞRo , where the subscripts n and o refer to the ‘new’ and ‘old’ con-
figurations, respectively.
The descriptions sofar being independent of rotation parameters, two different alternatives from liter-
ature have been explored. The first one is based on the rotation vector W, Eq. (5), while the second is based
on the spatial incremental rotation vector W (cf. Eqs. (6) and (21)) [5]. Additional changes of variables from
dx to dW, or alternatively from dx to dW, are therefore introduced in order to make the force expressions
conjugate to the displacement variables. These changes of variables will be defined in the sequel.

2.6.1. Change of variables: dx ! dW

Using Eq. (13), with U denoting the strain energy, and uI , WI the translation and rotation vectors at node
I, respectively, the internal forces at node I can be obtained through differentiation. Thus, using the chain
rule,
2 3 2 3
oU 2 3T oU
I 0  
6 ouI 7 3 3 6 ouI 7 T NI
fW ¼ 6 7¼4 ox 5 6 7 ¼ H ¼ H TI f gI ; ð59Þ
I 4 oU 5 03
I 4 oU 5 I
MI
oWI
oWI oxI
where
 
I 03
HI ¼ 3 : ð60Þ
03 T s ðWI Þ
Through further differentiation, the corresponding transformation equation for the tangent stiffness matrix
is obtained as
kW ¼ H T kg H þ K h ; ð61Þ
where H ¼ diagðH i ; H j ; H k Þ and
K h ¼ diagð03 ; K hi ; 03 ; K hj ; 03 ; K hk Þ: ð62Þ
The 3 3 blocks K hI are obtained from
sin w 1 T sin w 1 T
K hI ¼ cos w  ½ve  ðeT vÞeeT  þ 1  ½ev  2ðeT vÞeeT þ ðeT vÞI 3 
w w w w
!
2 2
sin w sinðw=2Þ 1 sinðw=2Þ
  ðe vÞeT þ ~v ð63Þ
w ðw=2Þ 2 ðw=2Þ

with W ¼ we ¼ WI , and v ¼ M I .
3786 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

The main advantage of the previously introduced change of variables is that the rotation variables
become additive and a special updating procedures is avoided.

2.6.2. Change of variables: dx ! dW

The main drawback of a parameterization based on the rotation vector is that the magnitude of the
rotation vector must be restricted to values less than 2p. In order to remove this drawback, a parame-
terization based on the spatial incremental rotation vector W and its iterative increment dW [5], has also
been implemented.
The transformation matrices required in this second case are constructed on the basis of Eq. (21) which
provides the connection between dx and dW. Following a procedure entirely similar to that used above, the
internal forces at node I are obtained from f WI ¼ H TI f gI with f gI as before and H I given by
 
I 03
HI ¼ 3 ; ð64Þ
03 T s ðW I Þ
where W I is the spatial incremental rotation vector at node I. The transformation equation for the tangent
stiffness matrix is also formally identical to Eq. (61), but with the 6 6 blocks H I evaluated according to
Eq. (64), and the 3 3 blocks K hI from Eq. (63) with v ¼ M I , but with W I replacing WI .
The transformation matrices required for this second type of parameterization are essentially con-
structed using the same linear operators as before. Information about the rotations at each converged step
must also be stored. The representation complicates the description of external nodal moments.

2.7. Core element formulation

The main feature of the adopted corotational framework is its independence of the core assumptions
used to derive the internal forces and tangent stiffness in local coordinates. Using this property, any type of
local formulation can be implemented, but the specific form considered in this work is based on the linear
TRIC element [8–10,19]. Two changes are, however, introduced in order to better model some test prob-
lems.

2.7.1. Natural and Cartesian elastic stiffness matrices

According to Fig. 2, the definition of the TRIC element uses two sets of local coordinates. The first one
X1e X2e X3e is the local Cartesian coordinate system, defined as discussed in Section 2.2. The second is the so-
called natural system abc, whose axes span the triangle edges.
Following the notation of [8,19,20], the state of displacement in the element is described as: q ¼ ½q0 qN ,
where q0 is the vector of rigid body motion while qN contains the amplitudes of the natural modes. The
natural straining modes are defined for each side: axial stretching, symmetric and antisymmetric bending
together with a transverse shear mode. In addition, three azimuth rotation modes wi , wj , wk are included.
The vector qN thus takes the form
qN ¼ ½ eta etb etc wi wj wk wSa wAa wSb wAb wSc wAc  ð65Þ
with the corresponding vector of natural internal forces given by
f N ¼ ½ fa ljk fb lki fc lij Mi Mj Mk MSa MAa MSb MAb MSc MAc ; ð66Þ
where fi , Msi , MAi ði ¼ i; j; kÞ are the axial load, symmetric and antisymmetric bending moments for side i,
while MI ðI ¼ i; j; kÞ denotes the azimuth moment associated with node I. The internal natural forces are
given by a 12 12 stiffness matrix, according to
f N ¼ kN qN ; kN ¼ diagðkM ; kB Þ; ð67Þ
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3787

Fig. 2. The three node flat shell element TRIC [8–10,19].

where kM and kB denote the membrane and bending stiffnesses, respectively. Since the element is flat, the
membrane and bending stiffnesses are uncoupled.
The natural force vector and the stiffness matrix are transformed from the natural to the local reference
system by a matrix [8,19,20]:
" #
oqNl
aN ¼ ; l ¼ 1; 12; k ¼ 1; 18; ð68Þ
qek
o

ueT
qe ¼ ½

where the vector

eT

# ueT
eT

# ueT
eT T contains the local displacements at the three nodes
#
i i j j k k
of the element. The internal force vector and tangent stiffness in local Cartesian coordinates are then given
by
f ¼ aTN f N ¼ kN

qe ; k ¼ aTN kN aN : ð69Þ
The definition of an element needs the natural stiffness kN and the transformation aN . The modifications to
the original TRIC element will thus affect both these aspects.

2.7.2. Membrane stiffness matrix

The membrane stiffness of the TRIC element is in [8,19] constructed from
kM ¼ diagðkn ; kz Þ; ð70Þ
T
where kn is the natural stiffness pertaining to the total natural strains et ¼ ½eta etb etc  , evaluated according
to:
kn ¼ Xtjct ; jct ¼ B 1 DB T ð71Þ
with X denoting the element area, t its thickness, D the constitutive matrix for two-dimensional (2D) plane
stress problems and B the transformation matrix connecting the natural and Cartesian strain components,
T
i.e. et ¼ B T e with e ¼ ½e11 e22 2e12  .
The matrix kz denotes the natural stiffness corresponding to the azimuth rotations. This matrix contains
three fictitious diagonal stiffness coefficients, taken typically as 106 maxðdiagðkB ÞÞ, at the three vertices.
3788 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

For the TRIC element [8,20], the terms in the transformation matrix aN , Eq. (68), pertaining to the
membrane stiffness are constructed from
xk  xj yk  yj
eta ¼ ð
uk 
uj Þ þ 2 ð
vk 
vj Þ;
l2jk ljk
xi  xk yi  yk
etb ¼ ð
ui 

uk Þ þ 2 ð
vi 
vk Þ; ð72Þ
l2ki lki
xj  xi yj  yi
etc ¼ ð
uj 

ui Þ þ 2 ð
vi 
vj Þ;
l2ij lij

and

wi ¼ #
e3i  #03 ; wj ¼ #
e3j  #03 ; wk ¼ #
e3k  #03 : ð73Þ

In the above equations, lIJ are the lengths of the three sides; xI , yI are the coordinates of the three nodes
measured along the local X1e and X2e axes, respectively;
uI ,
vI , #
e3I are the local in-plane translations and
rotations at the three vertices and #03 is the average rotation of the triangle, given by
1
#03 ¼ ½ðxj  xk Þ
ui þ ðxk  xi Þ
uj þ ðxi  xj Þ
uk þ ðyj  yk Þ
vi þ ðyk  yi Þ
vj þ ðyi  yj Þ
vk : ð74Þ
4X
Constructed as above, the membrane stiffness matrix of the TRIC element is equivalent to the CST element,
but equipped with three independent fictitious springs at the nodes in order to model the ‘drilling’ stiffness.
Although the TRIC element is for most cases accurate and efficient, it will for certain problems fail to
transfer twisting moments correctly. In the modification to the element, here denoted MTRIC, the mem-
brane stiffness was reformulated using the third order interpolation scheme introduced by Allman [21]. The
interpolation is defined in terms of the area coordinates ni , nj , nk over an arbitrary triangle with vertices I,
J ¼ i, j, k as
X X
ue ¼

ðAIJ lIJ cos gIJ nI nJ ðnJ  nI Þ þ BIJ lIJ cos gIJ nI nJ Þ þ
ueI nI ; ð75Þ
IJ ¼ij;jk;ki I¼i;j;k

X X

ve ¼ ðAIJ lIJ sin gIJ nI nJ ðnJ  nI Þ þ BIJ lIJ sin gIJ nI nJ Þ þ
veI nI ; ð76Þ
IJ ¼ij;jk;ki I¼i;j;k

where gIJ denotes the angle between the outward normal to the side IJ and the X1e local axis (Fig. 2).
According to [21], the coefficients AIJ and BIJ are connected to the nodal azimuth rotations. In the present
context, the coefficients are

AIJ ¼ 12ðwI þ wJ Þ; BIJ ¼ 12ðwJ  wI Þ: ð77Þ

It can be shown that the rotation description satisfies, for ðI ¼ i; j; kÞ

!

e
e
e 1 o
ve o
ue
#3I ¼ #3 ðx1 ; y1 Þ; #3 ðx; yÞ ¼  : ð78Þ
2 ox oy

The modified membrane displacement field gives a new connection between the total natural strains and the
local Cartesian displacements at the three nodes, Eq. (72). The three in-plane Cartesian strain components
are computed as
ue
o
o
ve o
ve o
ue
e11 ¼ ; e22 ¼ ; 2e12 ¼ þ : ð79Þ
ox oy ox oy
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3789

The resulting second order expressions are reformulated as

Z
1
e ¼ ea þ eh ; ea ¼ e dV ð80Þ
V V
with e ¼ ðe11 ; e22 ; 2e12 ÞT , and V ¼ Xt. The average strain components ea were symbolically evaluated using a
three-point Gauss rule. The total natural strains are then obtained as
xk  xj yk  yj X
eta ¼ uj Þ þ 2 ð
vk 
vj Þ  2 ð#
e3k  #
e3j Þ;
ð
uk 

l2jk ljk 3ljk
xi  xk yi  yk X
etb ¼ ð
uk Þ þ 2 ð
vi 
vk Þ  2 ð#
e3i  #
e3k Þ;
ui 
ð81Þ
l2ki lki 3lki
xj  xi yj  yi X
etc ¼ ð
ui Þ þ 2 ð
vj 
vi Þ  2 ð#
e3j  #
e3i Þ
uj 

l2ij lij 3lij

which gives the matrix aN in Eq. (68) as a replacement for the one in Eq. (72).
The modified displacement field also leads to a modification of the stiffness pertaining to the azimuth
rotation modes. In the present approach, this matrix is evaluated from the strain energy expression:
Z
1
Uh ¼ ðeh ÞT Dðeh Þ dV ; ð82Þ
2 V
corresponding to the higher order terms in Eq. (80). The integral in Eq. (82) was evaluated symbolically
using a four-point Gauss rule, giving a closed expression. As eh ¼ e  ea , Uh depends only on the azimuth
rotation modes.
Consequently, the expression for kz is obtained by differentiation with respect to the wI :
 2 
o Uh
kz ¼ ð83Þ
owI owJ
replacing the fictitious spring terms.
An interesting connection can here be established between the membrane part of the MTRIC element,
i.e. the Allman incompatible triangle, and the element with rotation degrees of freedom by Bergan and
Felippa [22]. The latter is constructed using the free formulation, which essentially amounts to the same
splitting of the strain components into a constant and a higher order component. In [22], the constant
component is computed according to
1 T
ec ¼ L v ð84Þ
X
where v is a vector which groups the Cartesian in-plane degrees of freedom and L ¼ LðaÞ is the so-called
‘lumping matrix’ which depends on a scaling factor a. Then ea as given by Eq. (80) can be obtained from
Eq. (84) evaluated for a ¼ 1.
If the interpolation scheme defined in Eqs. (75) and (76) is modified such that all higher order terms, i.e.
the terms involving AIJ and BIJ , are multiplied by a factor a, then Eqs. (84) and (80) will produce identical
results for any a. Consequently, using the terminology in [22], the two elements will share the same ‘basic
stiffness matrix’. As argued in [12,22], such a choice for the basic stiffness matrix will automatically ensure
the fulfilment of the so-called ‘individual element test’ [23]. If Eqs. (75) and (76) are affected by a coefficient
a in the way described above, then the expressions for the total natural strains, Eq. (81), and subsequently
the expression of the transformation matrix aN , Eq. (68), will be correspondingly parameterized by the same
coefficient a. Eq. (71) will however remain unchanged. Using a symbolic software tool such as Maple [24],
the implementation of such a parameterization in the present context is a straightforward matter.
3790 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

In addition, following the same references, the stiffness matrix pertaining to the higher order straining
modes, i.e. kz , can also be scaled by a coefficient b. In such a case, Eq. (70) would become
kM ¼ diagðkn ; bkz Þ: ð85Þ
The two coefficients a and b can then be tuned to optimise the response of the element to certain defor-
mation patterns [22]. This possibility has not been further exploited in the present work.

2.7.3. Bending stiffness

According to [8,19], the transversal displacements for the TRIC element can be seen as

ve ¼ N c wS þ N h wA ; ð86Þ
T T
where wS ¼ ½wSa wSb wSc  , wA ¼ ½wAa wAb wAc  while N c and N h are the constant strain and higher order
shape functions. Following a procedure similar to the one in the previous subsection, the expressions of the
total natural strains induced by the bending action are split into a constant and a higher order term ac-
cording to
ct ¼ cct þ cht : ð87Þ
In the above equation, the term cht is computed according to
cht ¼ Bth wA : ð88Þ
As pointed out in [19], the specific expression for B th used in the formulation of the TRIC element is not
energy orthogonal, in the sense that [12,23]
Z
Bth dV 6¼ 0: ð89Þ
V

In the present formulation, this property was enforced by replacing Bth as given in [8,19], by
Z
1
B th :¼ B th  Bth dV : ð90Þ
V V
This modification to the TRIC element was, however, verified to be of minor importance.
As a total, the computational demands for evaluating one element stiffness matrix is about 5% higher for
the MTRIC than for the TRIC element.

3. The multi-parametric setting

The purpose of this section is to discuss the multi-parametric setting of an instability problem, from the
level of the general concept, including its mathematical formulation based on augmented equilibrium
systems, and up to details concerning the numerical solution algorithms to be used in this context. This
description is a brief summary of previously published work [2,7,25].
The main ideas which lie behind the analysis strategy advocated in this paper can be intuitively illus-
trated by means of a simple structure. The response to a single load of a clamped toggle frame in Fig. 3a is
investigated for different heights H [26]. Fig. 3b shows common load–deflection curves for six different
heights, whereas Fig. 3c indicates a continuous higher-dimensional equilibrium surface in an (N þ 2)-
dimensional space ðq; k; H Þ, where q are the discrete displacements, and k the load factor. The response is
thereby no longer viewed only as a function of one variable (the load) for a specified height, but rather as a
function of two variables, i.e. the height of the frame is explicitly included in the analysis, which allows an
improved phenomenon description.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3791

Fig. 3. Equilibrium subset in multi-parametric structural equilibrium problem: (a) definition; (b) load–displacement curves; (c) curves
spanning an equilibrium surface; (d) projections of fold line.

As a conclusion from this viewpoint, the dash-dotted lines in Fig. 3c thus show that the two limit points
obtained for various heights, are situated on a continuous curve, a ‘fold line’. Projections of this fold line on
the ‘height-deflection’ and ‘height-load’ planes are shown in Fig. 3d. As a geometric interpretation, the
common load–deflection curves in Fig. 3b are the intersections of the full equilibrium manifold with co-
ordinate-constant planes, whereas the fold lines represent the intersections of the equilibrium manifold with
another manifold of critical states, i.e. states where the tangential stiffness matrix is singular [2].

3.1. Problem formulation

The basic formulation for the solution methods for a discrete model of a quasi-static elastic structural
problem, with N degrees of freedom is written in the generic residual form:
FðzÞ f ðq; KÞ  pðKÞ q
GðzÞ ¼ ¼ ¼ 0; z ; ð91Þ
gðzÞ gðq; KÞ K

where F is the set of N non-linear equilibrium equations, and g is a set of r auxiliary conditions, which
defines the required subset of equilibrium states. The solution to the problem gives combinations z of N
state variables q and p control variables K; the latter can include load factors and load directions, but can
also be related to the structural model [2,7]. The internal forces, here denoted f , are assumed dependent
on the current displacements q, but also on parameters introduced in the structural model, whereas the
3792 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

external forces p are dependent on, at least, K1 which is assumed to be the principal load factor. The so-
lution to the equilibrium equations alone is a set of p-dimensional manifolds.
The formulation introduces a demand that all variables are additive, and history-less, which complicates
the representation of the finite rotations in the present context. If a parameterization in terms of the in-
cremental rotation vector is adopted, the updating of the rotation variables at an iterative solution qjþ1
kþ1
must be performed according to
Rjþ1 jþ1 j j
kþ1 ¼ Rk expðDW k Þ ¼ Rk expðDW k þ dW k Þ ðj ¼ 0; 1; . . . ; ‘Þ; ð92Þ

Rkþ1 ¼ R‘þ1
kþ1 ; ð93Þ
where Rkþ1 is the stored rotation matrix corresponding to equilibrium point k. As already mentioned, an
additive updating is here preserved just at the level of the iterative corrections.
The auxiliary conditions allow conditions that are very broad in their scope (cf. [7,27]). A specially
interesting case are the criticality indicators, of which zero values indicate critical states. Here, scalar
measures are used to represent criticality [25,28,29].

3.1.1. Differential matrix

The differential of the mapping GðzÞ in Eq. (91) is important for many aspects of a solution algorithm. It
is formally given by the submatrices
 
F ;q F ;K
G ;z  ; ð94Þ
g ;q F ;K
where a subscript comma denotes differentiation with respect to the index variable. The accuracy of this
matrix is thus of paramount importance for the efficiency of the iteration process, but also for the inter-
pretation of solved parameterized equilibrium states [2,7,25,27].

3.1.2. Special solution sets

In [7], a term ‘special solution points’ was introduced as a generalization of the critical points occurring
on a load–displacement curve. In the multi-dimensional setting, the special solutions will become mani-
folds, rather than points. In general, these special solution manifolds will be of a dimension one lower than
the studied solution manifold, i.e. curves on a 2D solution surface. The present algorithms check the degree
of instability, noting the exchanges of stability between any equilibrium states. Also points where a control
variable vanishes, easily found from the solution points z, are often interesting [2,27].

3.2. Solution manifolds

The solution manifolds to the stated problems are represented by a number of discrete solution points,
fulfilling the equilibrium equations, but also the included augmenting equations. As an extension of the arc-
length constraining equations on load displacement paths [7,30], constraining functions NðzÞ are used as
position measures on the manifold.

3.2.1. Solution algorithms

Two different solution algorithms have been developed for structural mechanical problems in the general
form of Eq. (91). They are developed as general algorithms for a wide class of non-linear sets of equations.
When applied to numerical models based on finite element discretizations, a general finite element library
[31], is utilized in formulating internal force vectors and tangential stiffness matrices, needed in the for-
mulation of current and incremental, linearized, equilibrium expressions. This allows for easy creation of
main parts of the total expressions. The auxiliary equations are, however, often of types that are not easily
obtained from general library formulations, but must be found from a problem dependent description.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3793

The solution to the augmented system defined in Eq. (91) is a set of (p–r)-dimensional manifolds, with
curves as special cases when r ¼ p  1. A particular case is the common equilibrium problems, involving
only the load factor, p ¼ 1, and no auxiliary equations, r ¼ 0. The possibilities to solve for one-dimensional
(1D) generalized equilibrium paths are extensively discussed in [7,25].
Ref. [2] more tentatively discusses the possibilities to generalize and extend the 1D numerical procedure,
in order to investigate 2D equilibrium manifolds. The definition of the problem will thus involve p P 2
control parameters, and r ¼ p  2 P 0 auxiliary equations. The solution manifold to Eq. (91) now consists
of a set of surface segments in the (N þ p)-dimensional space. Since only a limited part of the total manifold
is of interest, the total problem is seen as variations described by the additional parameters, around a basic
case [2].

4. Numerical examples

The numerical examples presented in this section are grouped in two categories, presenting verification
for element formulation and instability treatment, respectively.

4.1. Element tests

The test problems included in this category had two main objectives. The first one was to motivate the
modifications to the TRIC element introduced in Section 2.7, checking the accuracy and efficiency of the
proposed element MTRIC. The section will also discuss the relative merits of the two alternative parame-
terizations of 3D rotations.

4.1.1. The Raasch problem

A curved strip depicted in Fig. 4 was analyzed, assuming the strip to be fully restrained
 at one end  and
loaded by an in-plane shear load at the other, represented by forces of the form ðP =N Þ 12; 1; . . . ; 1; 12 . This
problem is investigated by Knight [32], who shows that shell elements with shear stiffness predict a much
too flexible behavior and ultimately fail to converge in this case. According to Hoff et al. [33], this type of
behavior comes from the inability of the elements to transfer twisting moments correctly in curved shell
models. As the mesh is refined, an increasingly smooth description of the shell geometry across element

Fig. 4. The Raasach problem: (a) definition; (b) typical mesh (2 14).
3794 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

Table 1
The Raasch problem: normalized mid-node tip deflections in the load direction. Reference solution dref ¼ 4:9352 [32]
Element Mesh
2 14 4 28 8 56 16 112 20 136 40 272
6
TRIC kz ¼ 10 1.091 1.166 1.377 2.097 2.553 –
TRIC kz ¼ 103 1.089 1.159 1.351 1.987 2.377 –
MTRIC 0.996 0.995 0.998 1.008 1.011 1.0185
Rebel [34] 0.9733 0.9887 1.0049 1.0169 1.0188 –

boundaries is provided, which will result in a rank deficiency of the stiffness matrix, if no drilling stiffness is
provided.
The MTRIC element showed good results for this problem (cf. Table 1). Convergence took place es-
sentially at the same value as for the 9-node elements used in [34]. This result was about 1.8% more flexible
than the reference solution obtained by Knight [32], using a 20 136 2 mesh of assumed-stress hybrid
brick elements, but only 0.3% more flexible than the solution of Hoff [33]. The proposed modification of the
membrane stiffness was thus validated for this case, and can be compared to the solutions using the TRIC
element, where the fictitious drilling stiffness was not sufficient to prevent an over-flexible behavior, for any
value of kz .

4.1.2. Deployable ring

A ring with geometrical and material characteristics as shown in Fig. 5 [35], was analyzed as fully
clamped at one end and loaded at the other end by a twisting moment. At the loaded end, the center point
of the section was restricted to move along and rotate around the x axis. The loading was applied through a
square rigid plate attached to the ring [34]. The height and width of the plate were equal to the height of the
ring. Its thickness was taken as 10 times the thickness of the ring in order to provide a rigid transfer of the
load.
After a complete rotation cycle, i.e. HxðAÞ ¼ 2p, the ring was wrapped around itself and transformed into
a smaller ring with a radius of only one third of the original one. One more cycle, i.e. HxðAÞ ¼ 4p, brought
the ring back to the original configuration. The results obtained for this problem, using a 2 80 mesh of
MTRIC elements, are shown in Fig. 6. Both types of rotation parameterization were considered in the

Fig. 5. Deployable ring: definition.

 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3795

Fig. 6. Deployable ring: load–displacement curves. 2 80 mesh.

analysis. For a parameterization based on the total rotation vector, a quadratic convergence rate was
observed for values HxðAÞ < 2p. In the vicinity of 2p, the tangent stiffness matrix became ill conditioned, the
quadratic convergence was gradually lost; ultimately, the iterations failed to converge. In contrast, the
formulation based on the incremental rotation vector was unaffected and the computations gave quadratic
convergence up to HxðAÞ ¼ 4p. The two formulations produced identical results for 0 6 HxðAÞ < 2p. The
results were in excellent agreement with the results of Rebel [34], obtained using a 1 124 mesh of 9-node
shell elements.
For the same mesh, the results obtained using the TRIC element were less satisfactory (Fig. 6). Further
analyses, with a 4 496 mesh (Fig. 7) showed considerable sensitivity to the values of kz . For this mesh, the
deflections obtained using MTRIC elements were within 1% of those obtained using the 2 80 mesh.

Fig. 7. Deployable ring: load–displacement curves for TRIC element models, with various of kz . 4 496 meshes. MTRIC results for
comparison.
3796 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

Fig. 8. Pinched cantilever cylinder: definition.

4.1.3. Pinched cantilever cylinder

This example tested the element formulation for the cylindrical shell shown in Fig. 8. The cylinder, which
is clamped at one end and free at the other, is subjected to two opposite loads P. Using symmetry, only one
quarter of the cylinder was modelled with a regular 16 16 mesh. The material and geometrical properties
of the cylinder are also shown in Fig. 8.
The analyses were performed up to a displacement under the load of about 1.5R, ignoring the physically
occurring contact. Although this limit is not physically reasonable, the problem as such is an excellent test
for large deformation analysis. The obtained results are shown in Fig. 9.
For the MTRIC element, both types of rotation parameterization produced identical results; a quadratic
convergence rate was observed in both cases. For the TRIC element, the value of kz had very little impact

Fig. 9. Pinched cantilever cylinder: load–displacement curves.

 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3797

on the results. Fig. 9 was computed using kz ¼ 103 , but almost identical load–deflection curves were
obtained for 106 6 kz 6 103 . However, the quadratic convergence rate was lost for values kz 6 105 .
The calculated load–displacement curves were in excellent agreement with a solution provided by Okstad
and Mathisen [36], up to a displacement of about 1.3R. The latter solution is obtained using an adaptive
mesh procedure, which for displacements over 1.2R involved 976 triangular elements.

4.2. Stability investigations

The numerical test problems included in this subsection were aimed to prove, on one hand, the reliability
of the algorithms put forth in Section 3 in dealing with complex instability problems, and, on the other
hand, their versatility in performing parameter sensitivity analyses. Only the MTRIC element was utilized.
Since all test problems involve rotations with magnitudes less than 2p, the two alternative rotation de-
scriptions were identical.

4.2.1. Compressed plate

A rectangular plate was subjected to compressive stresses in the longitudinal direction and simply
supported along all four edges (Fig. 10). The in-plane displacements along the edges were assumed as free.
The finite element model used a 32 16 mesh for the entire plate.
The critical compressive stress for a uniformly compressed plate is
!
2
p2 E t i2 K22
pcr ¼ þ þ2 ; ð95Þ
12ð1  m2 Þ ly K22 i2
|fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
p0

where the lowest value of the critical load pcr;min ¼ 4p0 is obtained for K2 ¼ ðlx =ly Þ ¼ i (i ¼ 1; 2 . . .).
The dependence of the critical behavior of the plate on variations in its geometry and loading conditions
was studied. Apart from the main load factor K1 ¼ k ¼ p1 =p0 , three additional control parameters were
introduced (cf. Fig. 10): lx ¼ K2 ly , l1 ¼ K3 ly , and p2 ¼ K4 p1 . Only one of the additional parameters was
varied at a time.
The influence of the length to width ratio was examined first, i.e. variable K2 , but K3 ¼ 0, K4 ¼ 1. For a
basic problem with ly ¼ 400 and K2 ¼ 1, the first three bifurcation points, i ¼ 1, 2, 3, were isolated on the
fundamental path. Then, fold lines for variable length were evaluated, giving in Fig. 11 the bifurcation load
factors versus the ratio K2 .
The influence of K3 on the critical behavior of the plate was also examined. Five fixed loading config-
urations, i.e. K4 ¼ 1; 0.5; 0; 0.5; 1, were considered, for K2 ¼ 1:5. For K3 ¼ 1, the first critical load, with

Fig. 10. Compressed plate: definition.

3798 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

Fig. 11. Compressed plate: fold lines for first three bifurcation load levels, as functions of length to width ratio K2 .

load p1 ¼ pref was first isolated on the fundamental path and used as a starting point in the fold line
evaluation. Fig. 12 shows the obtained bifurcation load versus K3 .
The influence of K4 ¼ p2 =p1 , was also examined. The analysis was performed for four chosen values of
K3 , and fixed K2 ¼ 1:5. The obtained fold lines are plotted in Fig. 13. The reference value used in this case is
pref ¼ p0 , with p0 obtained from Eq. (95) with t ¼ t2 ¼ 1:2, i ¼ 2 and K2 ¼ 1:5. For K3 ¼ 0, the theoretical
results provided by Timoshenko and Gere [37], are also shown.
The stability investigation also studied mode jumping in connection with simply supported rectangular
plates. The term is here used to denote the sudden change in the wave number of the buckled configuration,
during the quasi-static loading process. Such a phenomenon was experimentally outlined by Stein [38,39],
for the case of rectangular compressed plates, simply supported along the longitudinal edges but clamped
along the transversal ones. The problem is further analyzed in [40], where it is shown that the conditions for

Fig. 12. Compressed plate: fold lines for first bifurcation load, as functions of variations in thickness change parameter K3 ¼ l1 =ly .
Various load distributions.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3799

Fig. 13. Compressed plate: fold lines for first bifurcation load, as function of load distribution factor K4 ¼ p2 =p1 . Various thickness
distributions K3 ¼ l1 =ly .

the occurrence of mode jumping are intrinsically related to the clamped boundary conditions and will not
be met for simply supported conditions.
For the simply supported case, however, it is argued in [41] that mode jumping is still possible under
some special conditions. These conditions are related to the existence of a special disturbing load which is
responsible for initiating and driving the mode jumping. This was studied for the simply supported plate of
Fig. 10. The length to width ratio was taken as K2 ¼ 1:40, with ly ¼ 100. The plate had uniform thickness
t1 ¼ t2 ¼ 1 and was subjected to uniform compression, i.e. K4 ¼ 1. The obtained results are plotted in Fig.
14. Buckling with one half-wave took place at kb;1 ¼ 4:472, immediately followed by buckling with two
half-waves, at kb;2 ¼ 4:535. The one half-wave post-buckling branch was rising and stable. The two half-
wave branch was initially unstable, but at ks ¼ 4:56 gave a secondary bifurcation; above this it was also

Fig. 14. Compressed plate: post-buckling behavior for K2 ¼ 1:40.

3800 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

Fig. 15. Compressed plate: branch connecting path. Disturbing load versus central deflection.

stable. This fact is of primary importance, since the occurrence of mode jumping requires the existence of
two stable modes [41].
At k ¼ 1:1kb;2 on the two half-wave branch, a disturbing load, acting at the middle of the plate in the
transversal direction, was applied. The disturbing load was described as Pd ¼ kd P0 where the reference load
P0 was taken as 103 times the total applied in-plane compressive force. The main load factor k ¼ 1:2kb;2
was kept constant, while the loading process varied the disturbing load, by incrementing the transversal
deflection.
On this path, a limit point L (Fig. 15), was first reached. The deflected shape of the plate at this point
contained a dominant two half-wave component (Fig. 16). Increasing the disturbing load beyond this value
would physically produce a jump from L to L0 , where the two half-wave component would be replaced by a
dominant one half-wave mode. Obviously from Fig. 15, reversing the direction of the perturbing load
would produce a jump in the opposite direction, i.e. from the one half-wave to the two half-wave con-
figuration.

Fig. 16. Compressed plate: deformed central line. Various points along branch connecting path.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3801

4.2.2. Cylindrical shell segment

A cylindrical shell segment, with material and geometrical properties as shown in Fig. 17a was analyzed.
The shell was subjected to a concentrated load P ¼ 1000K1 . Using symmetry conditions, only one half of
the shell was modelled with a 10 5 mesh.
The sensitivity of the critical behavior of the shell to variations in its geometrical characteristics was
studied. Two additional control parameters were used to describe variations in the thickness t and radius R
of the shell, according to t ¼ ð1 þ K2 Þt0 and R ¼ ð1 þ K3 ÞR0 , with t0 ¼ 6:35 and R0 ¼ 2540.
An analysis of a basic equilibrium problem defined by K2 ¼ 0 and K3 ¼ 0 was performed (Fig. 17b). The
results are compared with those obtained from a 3D brick model [1], with excellent agreement. The fun-
damental path exhibits a snap-through behavior with two limit points. Before the limit point, an unstable
symmetric bifurcation point was also detected.
A fold line for limit states was evaluated under variation of K2 , while K3 ¼ 0. The obtained results are
shown in Fig. 18a; the snap-through behavior only exists for t 6 22. A fold line was also evaluated for the

Fig. 17. Cylindrical shell segment: (a) definition; (b) load deflection curves for the basic equilibrium problem t ¼ t0 ¼ 6:35, R ¼
R0 ¼ 2540.

Fig. 18. Cylindrical shell segment: fold lines for critical load as functions of shell thickness t; (a) limit points; (b) bifurcation points.
3802 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

Fig. 19. Cylindrical shell segment: fold line for limit states evaluated as function of shell thickness. Projection on load–displacement
space.

bifurcation state. Fig. 18b shows the bifurcation load against the thickness of the shell. The fold line
presents a turning point with respect to the thickness, which indicates that the bifurcation state exists for
t 6 10:6.
Fig. 19 shows the obtained limit load as a function of the displacement at the limit state. The fold line
intersects each load–deflection curve at the two limit points. For the border case t ¼ 22, the two limit points
merged into an inflection point and the fold line was tangent to the load–deflection curve.
Two points where the bifurcation and limit point fold lines are tangent, were also isolated (Fig. 18b).
These points revealed the existence of hilltop branching states, for t  8:0 and t  8:6, respectively.
The analysis was repeated, for variations in the radius of the shell, i.e. K3 variable and K2 ¼ 0 (Fig. 20a).
Hilltop branching points occurred for R ¼ 1:2578R0 and R ¼ 1:3744R0 while the bifurcation state was
shown to exist for all values of the radius up to R ¼ 1:6484R0 .

Fig. 20. Cylindrical shell segment: (a) fold lines for all critical states as functions of shell radius; (b) hilltop branching states as function
of radius R and thickness t.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3803

Fig. 21. Cylindrical shell segment: solution surfaces for two-parametric model, including vertical force and apex moment; (a) apex
moment value as color shading; (b) extraction of surface parts where moment is vanishing.

As seen in Figs. 18b and 20a, hilltop branching points occurred for various combinations of radius and
thickness of the shell. Evaluation of the fold line connecting all the hilltop critical states required variations
in both K2 and K3 , simultaneously. The obtained results are plotted in Fig. 20b, as combinations of R and t
giving such behavior; the loads for these are shown by another projection of the curves.
This example was also used to illustrate the evaluation of a 2D equilibrium surface. For the basic geo-
metry, and in addition to the basic load, an apex moment was applied at point A, with magnitude de-
scribed by a second control variable, M ¼ K2 M0 . The moment was applied around the lengthwise symmetry
line, representing the first buckling mode. With only equilibrium equations, a solution surface was ob-
tained, essentially describing the system with 2 degrees of freedom. Fig. 21 shows the obtained, rather
strange, equilibrium surface. Fig. 21a shows the moment value as a color shading, whereas Fig. 21b shows
an extraction of all states on the surface where the moment is vanishing, 1 6 M 6 1. This information is
obtained from interpolation in the triangles, and although it did not require any kind of isolation of critical
states very clearly shows the bifurcation in the behavior. As a complement to this, Fig. 22 shows the ob-
tained degree of instability for all points on the surface, interpolated over the triangles. Due to the coarse
triangularization used, the transitions between zero, one and two negative eigenvalues of the tangential
stiffness matrix are not too clearly visible from this figure without further refinement of the mesh.

4.2.3. Compressed C profile

A simply supported beam of symmetric C section was studied, with material and geometrical proper-
ties shown in Fig. 23. A uniform compressive force P ¼ 125 000K1 was applied at the two ends. The
boundary conditions were defined as uðLx =2; Ly =2; 0Þ ¼ 0, vð0; Ly =2; 0Þ ¼ vðLx ; Ly =2; 0Þ ¼ 0 and wð0; y; 0Þ ¼
wðLx ; y; 0Þ ¼ 0. The finite element model was based on a 90 ð4 þ 4 þ 12Þ mesh, and 11 466 degrees of
freedom.
The sensitivity of the critical behavior of the beam to variations in its geometrical characteristics was
studied. Two additional control parameters were included in the model, giving t ¼ K2 t0 , L ¼ K3 L0 with
t0 ¼ 1, L0 ¼ 900 the standard case.
The investigation was initiated with a basic equilibrium problem defined by K2 ¼ 1, K3 ¼ 1. As also
noted in [1,42], a cluster of buckling modes was present on the fundamental path. Thus, 12 bifurcation
points were isolated for 1:43 6 k 6 1:64. The obtained results, shown in Table 2, are in excellent agreement
3804 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

Fig. 22. Cylindrical shell segment: degree of instability as a color shading on two-parametric equilibrium surface.

Fig. 23. Compressed C profile: definition of basic case.

with results in [1], using 400 20-node brick elements, p ¼ 6, and 65 032 degrees of freedom. The main critical
behaviors can be described as: local buckling with a certain number of half-waves (h-w), global flexural
buckling and torsional buckling (Fig. 24).
The influence on the buckling loads of the thickness t was analyzed, with variable K2 , K3 ¼ 1. Four of the
12 critical states were included in the analysis. The corresponding critical modes are shown in Fig. 24. The
results obtained from the fold line evaluations are shown in Fig. 25, with the critical load factors plotted
against the thickness t. For the two local modes, the critical loads varied essentially as third order functions
of t. For the bending mode, the dependence of the critical load on t was almost linear. A twofold critical
point was present for t ¼ 1:0277. This point involved the local mode with 11 h-w and the bending mode; the
10 h-w critical state was situated immediately above (0.3% difference in critical load). Over this threshold
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3805

Table 2
Compressed C profile: load factors and mode shapes for the first 12 bifurcation points for standard geometry
No. MTRIC 3D model [1] Mode shape
1 1.4311 1.4351 11 h-w
2 1.4352 1.4393 12 h-w
3 1.4464 1.4494 10 h-w
4 1.4656 1.4657 13 h-w
5 1.4734 1.4806 9 h-w
6 1.4872 1.4844 Flexural
7 1.5103 1.4986 14 h-w
8 1.5291 1.5408 15 h-w
9 1.5603 1.5455 8 h-w
10 1.6195 1.5919 16 h-w
11 1.6285 1.6227 Torsional
12 1.6383 1.6594 7 h-w

Fig. 24. Compressed C profile: a few critical modes; (a) local, 11 h-w; (b) local, 10 h-w; (c) flexural; (d) torsional.

value, which can be decided to high accuracy by the fold line calculations, the bending mode will first occur
on the fundamental path of the structure, giving substantially simpler behavior and computations.
A similar analysis was performed with variable K3 , K2 ¼ 1. Only the local modes were considered in the
analysis. The obtained results are shown in Fig. 26; while the value of the first critical load remained ap-
proximately constant, 1:41 < k < 1:42, the associated critical mode changed from 11 h-w for L ¼ 900 to 6
h-w for L ¼ 450. The transition from odd to even modes took place at twofold critical points, while the
transition between two odd or two even modes was very smooth. The corresponding fold lines came close to
3806 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

Fig. 25. Compressed C profile: fold lines for bifurcation loads as functions of length ratio.

Fig. 26. Compressed C profile: fold lines for bifurcation load as function of length. Numbers on curves indicate number of half-waves
in critical mode.

each other, switched shapes and then diverged again. The situation is very similar to that outlined in [1] for
a transversally loaded T-beam.
The behavior of the beam was also studied in the post-critical range. The analysis was performed only
for the basic problem, i.e. t ¼ 1, L ¼ 900. For this case, the post-buckling behavior was strongly influenced
by the interaction between the local and bending modes. Thus, buckling was initially induced in the 11 h-w
mode at the first critical point. Immediately after the critical point, the bending mode was also activated,
giving the asymmetric behavior described by the post-buckling path in Fig. 27a and the corresponding
deformed configuration in Fig. 27c. Regardless of the direction in which the local mode was initiated, the
global mode was finally always triggered downwards, which corresponded to an increased compression in
the flanges.
The post-buckling path associated with the second (symmetric) bifurcation point, appeared as a closed
curve (Fig. 27b) which intersected the first post-buckling path at point I as indicated in the 3D diagram of
Fig. 27d.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3807

Fig. 27. Compressed C profile: post-critical behavior; (a) first post-buckling path; (b) second post-buckling path; (c) deformed con-
figuration on first post-buckling path; (d) 3D view of both paths.

5. Conclusions

Two topics of relevance for the stability analysis of shells or shell like structures are addressed in the
present paper. The first one is related to the formulation of a non-linear facet shell element which is not only
accurate but also inexpensive, whereas the second is related to the numerical solution techniques appro-
priate for these analyses.
The element definition is addressed within the framework of a corotational approach, with this concept
discussed in some detail. The issues emphasized in this context are the parameterization of finite 3D ro-
tations and the core element formulation.
Concerning the parameterization of 3D rotations, two alternatives are discussed in the paper. The first
one uses the common rotation vector. The advantage here is that the rotation variables are additive and
3808 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

that there is no need for a special updating procedure. This feature is extremely important for the efficient
implementation of the numerical techniques advocated in the paper. The drawback associated with this
type of parameterization is that the magnitude of the rotation vector must be restricted to values less than
2p, in order to avoid singularity of the tangent stiffness matrix.
The second alternative for rotation parameterization is based on an spatial incremental rotation vector
[5]. Additive updates still apply but this time restricted within an increment, at the level of the iterative
correction. The benefit of this compromise, is that the ill-conditioning problems are avoided. The relative
merits of the two alternatives are illustrated by the numerical examples. Thus, if the problem at hand in-
volves rotations larger than 2p, the second option is the only available. However, for problems where this
limit is not reached, the first alternative is more efficient in the context of fold line evaluations, due to its
lower storage requirements. The two parameterizations are also based on transformation matrices, which
are very similar. An efficient implementation will thus offer the choice.
The core element formulation is essentially based on the so-called TRIC element of Argyris et al. [8–
10,19]. Two modifications are, however, implemented in the present approach, giving an element MTRIC.
The first is related to the membrane stiffness, which is redefined by adding aspects from Allman’s third
order interpolation scheme. As shown by the numerical examples, this modification removed a certain
vulnerability of the original TRIC element to the choice of the fictitious drilling stiffness coefficient kz . The
second modification is at the level of the bending stiffness. This modification, enforcing energy orthogo-
nality in the sense of [12,23], has, however, no significant impact on the behavior of the element, at least for
the numerical examples examined in this paper.
With respect to computational efficiency, the linear TRIC or MTRIC elements in a corotational setting
is an attractive possibility when compared to many other non-linear shell elements. Some comparisons have
also been made to shell instability analyses based on solid elements. Although not conclusive, due to dif-
ferent overall implementations and a limited number of comparisons, it is concluded that similar accuracies
in results can be obtained with substantially fewer degrees of freedom, when using a shell element for the
discretization.
The second major topic covered in the paper concerns the numerical viewpoint to be used in the analyses
of instability problems. With a view to parameter dependence investigations, the discussion is focussed
from, the onset on a multi-parametric description of a problem. Solution algorithms for both 1D and 2D
equilibrium manifolds are briefly discussed.
The instability examples presented show that improved understanding of occurring phenomena can be
reached by evaluating generalized equilibrium paths, primarily fold lines. Although more complex than a
conventional equilibrium analysis, their use is very efficient in the overall economy of the analysis, since the
parameter dependence of the response is obtained without complete re-evaluations of modified structural
models. As illustrated in the numerical examples, the control parameters can be systematically varied and
seemingly uncommon instability situations can be revealed to high accuracy. Moreover, the procedure
allows a complete freedom in the choice of the control parameters, with no restriction placed on their
magnitude. The 2D situation, i.e. the equilibrium surfaces, is treated as a generalization of the 1D case. A
numerical example, illustrating the type of information that can be obtained from such an analysis is given.
It is concluded that the multi-parametric setting and the different dimensions of solution manifolds are
important tools in numerical instability investigations, but need further development of procedures for
evaluation and interpretation.

References

[1] A. Eriksson, C. Pacoste, A. Zdunek, Numerical analysis of complex instability behaviour using incremental-iterative strategies,
Comput. Meth. Appl. Mech. Engrg. 179 (1999) 265–305.
 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810 3809

[2] A. Eriksson, C. Pacoste, Solution surfaces and generalised paths in non-linear structural mechanics, Int. J. Struct. Stab. Dyn. 1
(2001) 1–30.
[3] B. Nour-Omid, C.C. Rankin, Finite rotation analysis and consistent linearization using projectors, Comput. Meth. Appl. Mech.
Engrg. 93 (1991) 353–384.
[4] C. Pacoste, Co-rotational flat facet triangular elements for shell instability analyses, Comput. Meth. Appl. Mech. Engrg. 156
(1998) 75–110.
[5] A. Ibrahimbegovic, On the choice of finite rotation parameters, Comput. Meth. Appl. Mech. Engrg. 149 (1997) 49–71.
[6] M. Geradin, A. Cardona, Flexible Multibody Dynamics. A Finite Element Approach, Wiley, Chichester, 2001.
[7] A. Eriksson, Structural instability analyses based on generalised path-following, Comput. Meth. Appl. Mech. Engrg. 156 (1998)
45–74.
[8] J. Argyris, L. Tenek, L. Olofsson, TRIC: a simple but sophisticated triangular element based on 6 rigid-body and 12 straining
modes for fast computational simulations of arbitrary isotropic and laminated composite shells, Comput. Meth. Appl. Mech.
Engrg. 145 (1997) 11–85.
[9] J. Argyris, M. Papadrakakis, C. Apostolopoulou, S. Koutsourelakis, The TRIC shell element: theoretical and numerical
investigation, Comput. Meth. Appl. Mech. Engrg. 182 (2000) 217–245.
[10] J. Argyris, L. Tenek, M. Papadrakakis, C.C. Apostolopoulou, Postbuckling performance of the TRIC natural mode triangular
element for isotropic and laminated composite shells, Comput. Meth. Appl. Mech. Engrg. 166 (1998) 211–231.
[11] K. Mattiasson, On the corotational finite element formulation for large deformation problems, Dr. Thesis, Chalmers University of
Technology, G€ oteborg, 1983.
[12] P.G. Bergan, M.K. Nyg ard, Finite elements with increased freedom in choosing shape functions, Int. J. Numer. Meth. Engrg.
20 (1984) 643–664.
[13] A. Eriksson, On a thin shell element for non-linear analysis, based on the isoparametric concept, Comp. Struct. 42 (1992) 927–
939.
[14] A. Cardona, M. Geradin, A beam finite element non-linear theory with finite rotations, Int. J. Numer. Meth. Engrg. 26 (1988)
2403–2438.
[15] C. Pacoste, A. Eriksson, Beam elements in instability problems, Comput. Meth. Appl. Mech. Engrg. 144 (1997) 163–197.
[16] A. Ibrahimbegovic, F. Frey, I. Kozar, Computational aspects of vector like parameterization of three-dimensional finite rotations,
Int. J. Numer. Meth. Engrg. 38 (1995) 3653–3673.
[17] C.C. Rankin, On choice of best possible corotational element frame, in: S.N. Atluri, P.E. O’Donoghue (Eds.), Modeling and
Simulation Based Engineering, vol. 1, Tech Science Press, Forsyth, GA, 1998, pp. 772–777.
[18] M.A. Crisfield, A consistent co-rotational formulation for non-linear, three-dimensional, beam-elements, Comput. Meth. Appl.
Mech. Engrg. 81 (1990) 131–150.
[19] J. Argyris, M. Papadrakakis, C. Apostolopoulou, S. Koutsourelakis, TRIC: A cost effective and reliable shell element, in: W.
Wunderlich, E. Stein (Eds.), European Conference on Computational Mechanics, ECCM’99, Munich, 1999.
[20] J. Argyris, L. Tenek, A natural triangular layered element for bending analysis of isotropic, sandwich, laminated composite and
hybrid plates, Comput. Meth. Appl. Mech. Engrg. 109 (1993) 197–218.
[21] D.J. Allman, Evaluation of the constant strain triangle with drilling rotations, Int. J. Numer. Meth. Engrg. 26 (1988) 2645–2655.
[22] P.G. Bergan, C.A. Felippa, A triangular membrane element with rotational degrees of freedom, Comput. Meth. Appl. Mech.
Engrg. 50 (1985) 25–69.
[23] P.G. Bergan, Finite elements based on energy orthogonal functions, Int. J. Numer. Meth. Engrg. 15 (1980) 1541–1555, Addendum
17 (1981) 154–155.
[24] B.W. Char, K.O. Geddes, G.H. Gonnet, B.L. Leong, M.B. Monagan, S.M. Watt, Maple V Language Reference Manual,
Springer, Berlin, 1991.
[25] A. Eriksson, Fold lines for sensitivity analyses in structural instability, Comput. Meth. Appl Mech. Engrg. 114 (1994) 77–101.
[26] C. Pacoste, A. Eriksson, Element behaviour in post-critical plane frame analysis, Comput. Meth. Appl. Mech. Engrg. 125 (1995)
319–343.
[27] A. Eriksson, Equilibrium subsets for multi-parametric structural analysis, Comput. Meth. Appl. Mech. Engrg. 140 (1997) 305–
327.
[28] P.J. Rabier, G.W. Reddien, Characterization and computation of singular points with maximum rank deficiency, SIAM,
J. Numer. Anal. 23 (1986) 1040–1051.
[29] R.-X. Dai, W.C. Rheinboldt, On the computation of manifolds of foldpoints for parameter-dependent problems, SIAM,
J. Numer. Anal. 27 (1990) 437–446.
[30] M.A. Crisfield, New solution procedures for linear and non-linear finite element analysis, in: J. Whitman (Ed.), The Mathematics
of Finite Elements and Applications V, Academic Press, London, 1986, pp. 49–81.
[31] A. Olsson, Object-oriented finite element algorithms, Lic. Thesis, Dept. Struct. Engng., Royal Inst. Techn., Stockholm, 1997.
[32] N.F. Knight Jr., The Raasch challenge for shell elements, in: Proceedings of the 37th AIAA/ASME/ASCE/AHS/ASC Structures,
Structural Dynamics and Material Conference, CP. 962, Salt Lake City, UT, 1996, pp. 450–460.
3810 A. Eriksson, C. Pacoste / Comput. Methods Appl. Mech. Engrg. 191 (2002) 3775–3810

[33] C.C. Hoff, R.L. Harder, G. Campbell, R.H. MacNeal, C.T. Wilson, Analysis of shell structures using MSC/NASTRAN’s shell
elements with surface normals, in: The 1995 MSC World Users’ Conference, Universal City, CA, 1995, The MacNeal-Schwendler
Corporation.
[34] G. Rebel, Finite rotation shell theory including drill rotations and its finite element implementation, Ph.D. Thesis, Delft University
of Technology, 1998.
[35] Y. Goto, Y. Watanabe, T. Kasugai, M. Obata, Elastic buckling phenomenon applicable to deployable rings, Int. J. Solids Struct.
29 (1992).
[36] K.M. Okstad, K.M. Mathisen, Towards automatic adaptive geometrically non-linear shell analysis, Part I: Implementation of an
h-adaptive mesh refinement procedure, Int. J. Numer. Meth. Engrg. 37 (1994) 2657–2678.
[37] S.P. Timoshenko, J.M. Gere, Theory of Elastic Stability, second ed., McGraw-Hill, New York, 1961.
[38] M. Stein, The phenomenon of change of buckling patterns in elastic structures, Technical Report R-39, NASA, 1959.
[39] M. Stein, Loads, deformations of buckled rectangular plates, Technical Report R-40, NASA, 1959.
[40] D. Schaeffer, M. Golubitski, Boundary conditions and mode jumping in the buckling of a rectangular plate, Commun. Math.
Phys. 69 (1979) 209–236.
[41] H. Suchy, H. Troger, R. Weiss, A numerical study of mode jumping of rectangular plates, ZAMM 65 (1985) 71–78.
[42] A.D. Lanzo, G. Garcea, Koiter’s analysis of thin-walled structures by a finite element approach, Int. J. Numer. Meth. Engrg.
39 (1996) 3007–3031.