International Journal of Non-Linear Mechanics 81 (2016) 129–138

Contents lists available at ScienceDirect

International Journal of Non-Linear Mechanics

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

An eigenanalysis-based bifurcation indicator proposed

in the framework of a reduced-order modeling technique
for non-linear structural analysis
Ke Liang a,b,n, Martin Ruess c, Mostafa Abdalla b
a
Qian Xuesen Laboratory of Space Technology, China Academy of Space Technology, Beijing 100094, China
b
Aerospace Structures and Computational Mechanics, Delft University of Technology, Kluyverweg 1, 2629 HS Delft, The Netherlands
c
School of Engineering, University of Glasgow, Rankine Building, Oakfield Avenue, Glasgow G12 8LT, United Kingdom

art ic l e i nf o a b s t r a c t

Article history: The Koiter–Newton method is a reduced order modeling technique which allows us to trace efficiently
Received 3 July 2015 the entire equilibrium path of a non-linear structural analysis. In the framework of buckling the method
Received in revised form is capable to handle snap-back and snap-through phenomena but may fail to predict reliably bifurcation
10 November 2015
branches along the equilibrium path. In this contribution we extend the original Koiter–Newton
Accepted 19 January 2016
approach with a reliable and accurate bifurcation indicator which is based on an eigenanalysis of the
Available online 27 January 2016
reduced order tangent stiffness matrix. The proposed indicator has a negligible numerical effort since all
Keywords: computations refer to the reduced order model which is typically of very small dimension. The extension
Koiter–Newton method allows the identification of bifurcation points and a tracing of corresponding bifurcation branches in each
Bifurcation indicator
sector of the equilibrium path. The performance of the method in terms of reliability, accuracy and
Eigenanalysis
computational effort is demonstrated with several examples.
Tangent stiffness
Reduced order model & 2016 Elsevier Ltd. All rights reserved.

1. Introduction general path-following methods has created a number of available

In aerospace engineering, many structures are prone to be
limited in their load carrying capability by buckling, while staying
in the linear elastic range of the material [1–4]. It is very important
to trace the complete equilibrium paths, i.e. load-displacement
curves, in order to know and to comprehend the characteristics of
structures subjected to loads [5–8]. For structural buckling two
main phenomena are generally recognized, namely limit-point
buckling and bifurcation-point buckling. The instability properties
of some structures may even include both buckling phenomena.
An accurate assessment of buckling is highly significant in the
structural design, since the load carrying capability of a structure
is largely overestimated if a bifurcation point that lies before the
limit point is missed during the non-linear path-following analy-
sis. The consideration of possible bifurcations during the path
tracing of the non-linear response of a structure includes two key
features, the detection of bifurcation points and the branch-
switching from one equilibrium path to another equilibrium
path of the structure. The implementation of the two features into
n
Corresponding author at: Qian Xuesen Laboratory of Space Technology, China
Academy of Space Technology, Beijing 100094, China. Tel./fax: þ 86 010 68 113099.
E-mail address: liangke@qxslab.cn (K. Liang).
numerical techniques for bifurcation analysis.
Non-linear path-following techniques, e.g. the arc-length
method and its variants [9–11], work well for applications where
pure limit-point-type buckling with a unique equilibrium path is
present, but have difficulties to converge to bifurcation points and
generally prevent an automatic switch from the pre-instability
path to one of the bifurcation paths. In general, a small physical
imperfection is introduced to allow a switch of the primary equi-
librium path to a neighboring bifurcated equilibrium path [1,2,12].
The approach bears the risk to influence the path prior to the
bifurcation point with changed characteristics of the considered
structure. Another, more accurate possibility to allow for a branch
switch at the bifurcation point is the consideration of a suitable
fictitious perturbation added to the equilibrium path near bifur-
cation points [13].
A modification of arc-length-type methods for reliable bifur-
cation analysis has been in the focus for many years and is
repeatedly addressed by many researchers. One popular procedure
[9,10] to detect the bifurcation point is to check the sign of the
determinant of the tangent stiffness at the current calculation step
which changes if either a limit point or a bifurcation point is
passed. However, the tracing analysis after the bifurcation point
cannot continue unless some further, usually computationally
expensive, analyses are taken into consideration. Similarly, the

http://dx.doi.org/10.1016/j.ijnonlinmec.2016.01.013
0020-7462/& 2016 Elsevier Ltd. All rights reserved.
130 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138
sign of stiffness parameters and eigenvalues of the tangent stiff-
ness matrix of the current analysis step are often used to analyze
the direction of the equilibrium path [14]. Feng et al. [15–17] took
the history of the current equilibrium path into consideration to
produce a reliable path direction prediction even in the presence
of bifurcations. Eriksson [18] used the scalar product of two path-
related measures, the critical eigenvector and reference load vec-
tor, to have a reliable bifurcation indicator which well distin-
guishes between bifurcation and limit points. To achieve a branch-
switching, Wagner and Wriggers [19] added the critical eigen-
vector as a perturbation to the primary path to shift one of the
possible bifurcation branches. Recently, Zhou et al. [13] proposed
an arc-length method combined with an eigenanalysis-based
branch-switching method to discover additional bifurcation
points and corresponding secondary paths of cylindrical panels.
The critical eigenvector obtained from the eigenanalysis is com-
monly used in the aforementioned computational procedures to
allow a tracing of the bifurcation branch after the bifurcation point
which requires the critical eigenvector to be updated in each
iteration step. However, it is not practical to do the eigenanalysis of
the stiffness matrix in each iteration step in the view of the tre-
mendous computational cost, especially for large-scale finite ele-
ment models. Noguchi [20,21] has proposed an innovative
eigenanalysis-free idea to extract the buckling mode for bifurca-
tion instability from the LDLT-decomposed stiffness matrix. Wardle
[22,23] proposed an asymmetric meshing technique(AMT) to deal
with the case that the bifurcation buckling in asymmetric mode
exists before the first limit point on the equilibrium path.
The reduction method based on the Rayleigh–Ritz or pertur-
bation techniques [24,25], which can reduce the number of
degrees of freedom in non-linear mechanics, is a valuable alter-
native for structural buckling analysis. Some of the reduction
methods [26–30] work in combination with path-following tech-
niques to trace the entire non-linear equilibrium path based on a
predictor-corrector approach. The prediction is achieved using an
asymptotic expansion [24,25,30] to reduce the number of
unknowns in the original equilibrium equations. The solution of
the reduced equations is regarded as a predictor for the real
equilibrium path. These methods often use the path derivative
[28,30] as the path parameter in the expansion of the equilibrium
equations, which causes difficulties near bifurcation points. Bou-
tyour, Vannucci and Jamal [31–33] have proposed a bifurcation
indicator to detect bifurcation points in the framework of the
Asymptotic Numerical Method (ANM) [24,34,30]. After introdu-
cing a fictitious perturbation force to the structure at a point of the
solution branch, the bifurcation indicator is expressed as a scalar
function to measure the intensity of the system response to per-
turbation forces. The value of this indicator is evaluated through
the path-following procedure and becomes zero exactly at the
bifurcation point. Recently, the bifurcation indicator that was
proposed by Boutyour et al. [33] has been successfully imple-
mented in various mechanical problems of thin films on compliant
substrates to detect multiple bifurcations [35,36]. In Boutyour's
work [33,37], the detection of a bifurcation point can also rely on
the analysis of the roots of the denominator of the Padé approx-
imants. In addition, Lopez [38] selected the expansion parameter
based on the residual vector, and Mottaqi [39] discussed some
other numerical techniques to detect the bifurcation point along
the equilibrium path, based on Potier-Ferry's work [34].
The Koiter–Newton method [40–42] is a recently proposed
reduced order modeling methodology to trace the entire equili-
brium path in a stepwise manner. At each step, the approach
combines a prediction using a non-linear reduced order model
(ROM) based on Koiter's initial post-buckling expansion [43–46]
with a Newton arc-length correction, thus allowing the algorithm
to use fairly large step sizes. The possibility of some perturbation
loads that may excite the secondary branches is taken into account
to deal with bifurcation-point-type buckling. We have found that
for an exactly perfect structure, i.e. a flat plat under axial com-
pression, the proposed method can directly trace one of the
bifurcation branches after the bifurcation point [42], with no need
to introduce any physical imperfection. The effectiveness of the
proposed method for the limit-point-type buckling has been
proven in [40], and a force imperfection model has been intro-
duced in [41] to show the ability of imperfection analyses. How-
ever, recently we noticed that if the structure bifurcates before the
limit point the proposed method may pass by the bifurcation
branch leading to an overestimated stability design. Hence, a
bifurcation indicator should be introduced in the Koiter–Newton
method to detect the bifurcation point and to allow tracing pos-
sible branches accurately.
The contribution of this paper distinguishes significantly from
previous publications in the strategy followed to find the bifur-
cation branches. Previously, these were found by the introduction
of perturbation loads, a strategy which may fail in the case that the
bifurcation branches lie before the limit point. In this paper we
overcome this shortcoming by using a bifurcation indicator based
on the eigenanalysis of the tangent stiffness. Before reaching the
first limit point the energy preserving properties of the system are
indicated by a positive definite tangent stiffness matrix, a property
which is lost with the appearance of instabilities in the structure
signaled by negative eigenvalues. Here, the tangent is obtained
from the reduced order model with, in general, less than 10
degrees of freedom. This fact allows the use of this bifurcation
indicator in each iteration step during tracing the equilibrium
path. In this respect, this method differs not only from the pre-
vious ones presented in [40–42] but also distinguishes from the
classical arc-length method, which always fails to deal with the
bifurcation-point-type buckling if no physical imperfection is
introduced. Another point which characterises this paper is the
use of the critical mode as the displacement perturbation at the
bifurcation point to let the primary path change into the expected
bifurcation branch. The proposed bifurcation indicator designed in
the framework of the Koiter–Newton method is also different from
the one proposed by Boutyour et al. [33] which is well adapted to
the Asymptotic Numerical Method. Boutyour's bifurcation indi-
cator introduces a fictitious perturbation force in the problem and
requires the computation of a second series at each step. The
indicator proposed in this paper is simpler with respect to
implementation aspects and the computation only refers to a
lower-order reduced order model.
The paper is organized as follows. The developed bifurcation
indicator in the framework of the Koiter–Newton method for
detecting bifurcation branches is presented in Section 2. In this
section we also briefly revisit the fundamental aspects of the ori-
ginal Koiter–Newton method to allow a clear comprehension of
the strategy and the mechanism of the proposed bifurcation
indicator. Numerical examples which demonstrate the success of
the method are provided in Section 3. We summarize the paper
and draw conclusions in Section 4.
2. A bifurcation indicator proposed in the framework of the
Koiter–Newton method
The Koiter–Newton method was proposed, similar to classical
path following techniques, as a step by step procedure to trace the
complete equilibrium path of a deforming structure. In each
expansion step of the Koiter–Newton method three basic steps are
involved: (1) construction of the reduced order model, (2) iterative
solution of the reduced order model, (3) correction of the pre-
dicted load-displacement step produced by the reduced order
 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138 131

model. In the following we introduce the three major analysis with respect to ξ:
steps and extend the analysis framework by a bifurcation indicator.
u ¼ uα ξα þ uαβ ξα ξβ þuαβγ ξα ξβ ξγ ; ð4Þ

2.1. Construction of the reduced order model where the subscripts α; β; γ ¼ 1; 2; …; m þ 1, and the summation
convention is applied. The first order displacements uα define the
For an elastic body, the non-linear equations of equilibrium are tangent plane to the equilibrium surface at the approximation
used to describe the non-linear response of the structure. An point. The second order displacements uαβ and third order dis-
approximation of the equilibrium equations in a discrete setting, placements uαβγ describe the interactions among first and second
e.g. by the use of a finite element model, reduces the governing order displacement fields, respectively.
equations to a set of non-linear algebraic equations: f int ðqÞ ¼ λf ext , The equilibrium surface may be parametrised with an infinite
where f int and f ext are the internal force vector and external force number of choices for ξ. To fix the paramterization, we choose the
vector, respectively, λ is the load parameter of the analysis, and q is vector ξ such that it is work conjugate to the load amplitudes ϕ:
the vector of primal values, the unknown displacements in
ðFϕÞt δu  ϕ δξ
t
ð5Þ
our model.
In the proposed Koiter–Newton method, a reduced order This leads to the following constraint equations:
model is established to approximate the equilibrium equations in 8
> f uβ ¼ δαβ
t
>
the neighborhood of a known equilibrium state ðq0 ; λ0 Þ, which < α
>
t
usually starts from the undeformed reference state of a structure. f α uβγ ¼ 0 ð6Þ
>
>
The vector q0 describes the configuration at this known equili- >
: ft u ¼0
α βγδ
brium state which we will refer to as the nominal configuration.
Let ðq; λÞ be the unknown state near this nominal state. We assume where δαβ is the Kronecker delta.
that the displacement vector q to be: Consistent with the displacement expansion, we assume that
the following expansion for the load amplitudes ϕ:
q ¼ u þq0 ð1Þ

where u describes the displacement field of the current config-

ϕ ¼ LðξÞ þQðξ; ξÞ þ Cðξ; ξ; ξÞ ð7Þ
uration with respect to the nominal configuration. where L, Q and C are still to be determined linear, quadratic and
In order to simplify the algebra, the equilibrium equations can cubic forms of the approximated load amplitudes and are repre-
be expanded to the third order with respect to u on the nominal sented by a two-dimensional tensor L, a three-dimensional tensor
configuration using a Taylor series expansion, as given by: Q and a four-dimensional tensor C of order ðm þ 1Þ, respectively.
Using the displacement expansion (4) and the load expansion
LðuÞ þQðu; uÞ þ Cðu; u; uÞ þ Oð J u J 4 Þ ¼ Δλf ext ð2Þ (7) in the equilibrium Eq. (3), equating the coefficients of the
where L, Q and C denote the linear, quadratic and cubic forms of various powers of ξ with zero, and taking the orthogonality con-
the approximated elastic forces, respectively. These forms can also ditions (6) into account result in the following two sets of gov-
be considered as a two-dimensional tensor L, a three-dimensional erning systems of equations:
tensor Q and a four-dimensional tensor C of order N, where N is   " # " #
Kt F uα 0
the total number of degrees of freedom in the finite element ¼ ð8Þ
 Ft 0 Lα  Eα
model. For example, Qðu; vÞ ¼ Q pij ui vj , where the Einstein sum-
mation convention is applied. We further use Δλ ¼ λ  λ0 and the   " #  
Kt F uαβ  Qðuα ; uβ Þ
fact that f int ðq0 Þ ¼ λ0 f ext .
t ¼ ð9Þ
The proposed method considers the higher-order extended F 0 Q αβ 0
equilibrium equations for buckling sensitive structures and
structures with imperfections. In the presence of buckling several where Kt ¼ L is the tangent stiffness matrix, Eα is the α-th unit
secondary branches may exist that intersect the primary equili- vector, and where L α and Q αβ represent column vectors of order
brium path at bifurcation points. For such cases, some perturba- ðm þ 1Þ of the multi-dimensional tensors L and Q, respectively.
tion loads which excite the secondary branches during the analysis The two sets of linear systems of Eqs. (8) and(9), have identical
should be taken into account. Recently, we have found that these coefficient matrices, hence factorization is needed only once. The
perturbation loads work well only for the perfect structure thus a dimension of the coefficient matrix is ðN þ m þ1Þ, where N is the
new bifurcation indicator is proposed and introduced in this paper. number of degrees of freedom of the full finite element model and
The third order form of the equilibrium equations may be exten- ðm þ 1Þ is the number of degrees of freedom of the reduced order
ded to consider multiple loadings of the form: model. Since ðm þ 1Þ is generally very small compared to N, these
two sets of systems can be regarded as augmented linear finite
LðuÞ þQðu; uÞ þ Cðu; u; uÞ þ Oðkuk4 Þ ¼ Fϕ ð3Þ element(FE) systems [47].
where F is a load matrix column-wise formed by ðm þ 1Þ load vectors In addition to (8) and(9), the load and displacement expansions
f α ðα ¼ 1; 2; …; m þ 1Þ, denoted as sub-loads. The first sub-load vector result in the relation:
f 1 ¼ f ext corresponds to be the external load. The remaining m load 2h i
C αβγδ ¼ Cðuα ; uβ ; uγ ; uδ Þ  utαβ Lðuδγ Þ þ utβγ Lðuδα Þ þ utγα Lðuδβ Þ
vectors are perturbation loads derived from m closely spaced linear 3
buckling modes of the structure. The selection of perturbation loads ð10Þ
will be addressed later. The vector ϕ represents load amplitudes, the
where C αβγδ denotes the components of tensor C. The relation (10)
form of which will be presented in (12).
does not need any factorization.
The solution of u, Eq. (3), in general, lies on an ðm þ 1Þ dimen-
The solutions of (8)–(10) give the unknown displacements uα ,
sional surface. The numerical construction of such a surface is uαβ and the unknown components of L, Q, C. At this point, the
computationally prohibitive. To circumvent this, an approximate expansion (7) of load amplitudes is achieved. We rewrite (7) in the
solution can be obtained using a series expansion. The equilibrium following form, to get the reduced order model of order ðm þ 1Þ:
surface is parametrised in terms of generalised displacements ξ,
and the equilibrium displacement is expanded to the third order LðξÞ þ Qðξ; ξÞ þCðξ; ξ; ξÞ ¼ ϕ ð11Þ
132 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138
where the left hand side of (11) has already been obtained after
solving (8)–(10). The right hand side, which represents the load
amplitudes ϕ, can be expressed by the load parameter Δλ of the
equilibrium Eq. (2) as:
ϕ ¼ ΔλE1
where E1 is the 1st unit vector. The dimension of the load
amplitudes ϕ is ðm þ 1Þ. The first value of Δλ corresponds the
increment of the external load f 1 . The other components are
parameters of the perturbation loads f α ; α ¼ 2; …; m þ 1 and all are
set equal to zero to simulate the response of the structure to the
actual loads.
In the case where no buckling occurs in the analysis, only the
external load f 1 and Δλ need to be considered resulting in only a
single degree of freedom in the reduced order model. Otherwise,
the perturbation loads f α ; α ¼ 2; …; m þ 1 are used to excite the
buckling branches, where m is the number of closely spaced
buckling modes of the analyzed structure. The construction of the
perturbation loads is related to the buckling modes of structures,
for the interested reader this is discussed in detail in [40]. In cases
where many closely spaced buckling modes exist, e.g. buckling
modes associated with equal or nearly equal buckling loads, an
interaction of these modes is likely with influence on the buckling
characteristics of the structure, thus the selection of relevant
perturbation loads should also consider this case. As a simple rule,
buckling modes with buckling loads within 120% of the first
buckling load are considered to be closely spaced buckling modes
and are used to construct the perturbation loads.
During the construction of the reduced order model, a set of
augmented linear FEM Eq. (8)–(9), needs to be solved. Since all
these systems share the same coefficient matrix, only one stiffness
matrix factorization is needed. Hence, the computational cost for
constructing the reduced order model is almost equivalent to the
cost of solving one linear FE system, assuming that the computa-
tional cost is dominated by a factorization and not by a forward
and back substitution.
2.2. Iterative solution of the reduced order model, implementing the
proposed bifurcation indicator
The reduced order model (11) obtained in Section 2.1 is a non-
linear system of equations, which can be solved using a classical
Newton-type or arc-length method. Using the solution of the
reduced order model the relation between the load parameter Δλ
and the displacement parameter ξ can be found from (11). In Fig. 1
we show exemplarily the nonlinear response curve of the Koiter–
Newton method using Eq. (1), the relation λ ¼ λ0 þ Δλ and a
Fig. 1. Path-following strategy of the Koiter–Newton method in the case where
bifurcation branches lie before the limit point. (For interpretation of the references
to color in this figure caption, the reader is referred to the web version of this
paper.)
displacement expansion (4) up to the second order, starting from
the undeformed state of the structure. The response curve
obtained from the reduced order model, cf. Fig. 1, red curve, is
used as a non-linear predictor in the analysis. A comparison with
the linear prediction of the Newton–Raphson method immediately
ð12Þ
reveals the pre-eminence of the Koiter–Newton approach in terms
of a significantly increased step size. Furthermore, the solution
effort of the reduced-order model is negligible due to few degrees
of freedom, usually less than ten, used.
As mentioned in the introduction part (Section 1), for some
nominal perfect structures the bifurcation branches may lie before
the limit point, as shown in Fig. 1. General non-linear path-fol-
lowing techniques, such as the Newton based arc-length methods,
often ignore these bifurcation branches and directly lead to a
limit-point-type buckling. In principle, the above case can be
avoided by checking the sign of the eigenvalues of the tangent
stiffness matrix at each iteration step. In the case where bifurca-
tion points do not exist prior to the limit point the equilibrium
path segment up to the limit point is stable and the corresponding
tangent stiffness matrix is positive definite, i.e. all eigenvalues are
positive. As soon as a negative eigenvalue appears in the tangent
stiffness along that segment positive definiteness is lost indicating
the existence of a bifurcation branch. However, the eigenanalysis is
computationally prohibitive due to the typically large dimension
of the tangent stiffness matrix generated from the full finite ele-
ment model of the structure. For these methods, the physical
imperfection, e.g. a force imperfection and/or geometric imper-
fection, must be introduced from the beginning of the non-linear
analysis to ensure a shift of the primary path to one of the acti-
vated bifurcation branches at the bifurcation point.
Fictitious perturbation loads have been implemented into the
Koiter–Newton method to excite the possible bifurcation branches,
and we noticed that this works well for the perfect structure, but
fails and directly leads to a limit-point-type buckling in the case of
bifurcation points located before the limit point. To overcome this
limitation we propose an eigenanalysis for the reduced order
model of small dimension which can be performed stably by direct
methods at reasonable computational cost [48]. Once a negative
eigenvalue appears in the system matrix of the reduced order
model of an incremental step, the solution process of the reduced
order model is halted. Returning to the initial state of the incre-
mental step and using the critical eigenvector corresponding to
the negative eigenvalue as a perturbation the primary path is
shifted to one of the possible bifurcation branches, cf. Fig. 1, blue
curves.
The flow chart of the proposed analysis strategy for the bifur-
cation indicator is given in Fig. 2 with the following explanations:
 The reduced order model is solved in an incremental iterative
procedure. In each iteration of the load increment step, the
current tangent stiffness matrix of the reduced order model is
formed, followed by a complete eigenvalue analysis. If all
eigenvalues are positive, the solution of the reduced order
model continues.
 A negative eigenvalue in the spectrum of the tangent matrix
indicates a bifurcation of the load-displacement curve and
requires a repetition of the current step starting from its initial
state. The associated mode is used as a perturbation of the
displacement field. In the current incremental step ði þ 1Þ the
initial displacement vector is perturbed as follows:
u0i þ 1 ¼ uni þ ϵv ð13Þ
where uni
is the total displacement vector obtained from the i-th
incremental step after n iterations; u0i þ 1 is the initial displace-
ment vector used in the increment step ði þ 1Þ; v is the
displacement perturbation mode which is the eigenvector
 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138
Fig. 2. Flow chart of the algorithmic strategy of the bifurcation indicator used in
the solution process of the reduced order model.
corresponding to the negative eigenvalue; and ϵ is a scaling
factor for the numerical perturbation, usually set to be 1e  3.
In step ði þ 1Þ the iteration starts with u0i þ 1 . After several
iterations, a new balance point in an instability path corre-
sponding to the mode v is obtained, thus the analysis shifts to
the expected path and the following steps can be continued
as usual.
 It is worth to mention that the proposed computer strategy for
bifurcation analysis also works in the case of one or multiple
bifurcations located at one single bifurcated branch. During
tracing the bifurcation branch, we can repeat the computing
procedures in Fig. 2 to detect and trace any possible bifurcated
branches located in the current branch, however we need to
notice the following issues: the dimension of the reduced
order model determines the total number of computable
eigenvalues, which also gives the maximum number of bifur-
cation branches that the proposed indicator is able to find,
hence a sufficiently high number of degrees of freedom must
be considered in the construction of the reduced order model
to ensure that none of the expected bifurcation branches is
missed. This issue is revisited in the following section and the
algorithmic approach proposed is demonstrated with the
numerical example 3.5.
2.3. Correction of the predictor
A classical drawback of reduced order models is that they have
a limited range of validity, however, by accounting for possible
branching into quickly softening modes, the proposed reduced
order model has a wider range of validity. Yet, it still has some
limited range of validity that cannot be determined a priori.
During the simulation of the reduced order model, the unba-
lanced force residual at the new predicted location on the equili-
brium path is calculated using (14), based on the full finite element
model. This residual force f r is used to decide whether the applied
reduced order model is sufficiently accurate. The simulation of the
133
Fig. 3. Flow chart of the Koiter–Newton method, implementing the bifurcation
indicator.
reduced order model is stopped for f r being below a given
threshold.
f r ¼ λf ext  f int ð14Þ
where f int is the internal force of the current iteration step, f ext is
the total external force and λ is the current load parameter. The
internal force calculation is very accurate as it is calculated using
the full finite element model.
In a subsequent corrector phase, the residual is driven to zero
in a manner similar to that used in traditional Newton-type arc-
length methods, as indicated in Fig. 1 by a small red arrow. Based
on the new found equilibrium state of the structure an updated
reduced order model is constructed which is used for further load
steps. During the correction of the predicted load-displacement
step produced by the reduced order model, a few Newton itera-
tions are used to increase the numerical stability of the procedure.
The computational cost for each iteration corresponds to the
solution of a linear FE system.
2.4. Summary of this section
The flow chart of the Koiter–Newton method extended by an
eigenanalysis-based bifurcation indicator is depicted in Fig. 3. The
flow chart shows that the improved version can be easily imple-
mented into the existing analysis scheme with little effort.
The total number N of linear FE systems which need to be
solved in the Koiter–Newton method is assessed with:
X
n
p
N¼ ð1 þ N c Þ ð15Þ
p¼1
where n is the total number of steps taken by the method, 1 indi-
cates the linear FE system which needs to be solved for the con-
struction of the reduced order model, and N pc is the number of
solved linear FE systems during the correction phase of step p. The
proposed Koiter–Newton method significantly improves the effi-
ciency of non-linear finite element analysis due to the much larger
step size provided by the non-linear reduced order model, espe-
cially for buckling sensitive structures.
The number N refers solely to the number of solutions of the
full FE model during the iterative solution procedure as is often
used to quickly assess the overall computational effort. The Koiter–
Newton method uses a high-order predictor associated with a
low-order corrector, hence the correction procedure may take
some computing time. To reduce this effort, a strict tolerance is
used to evaluate the accuracy of the predictor which results in a
reduced number of iterations for the correction phase.
134 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138
3. Numerical examples
We implemented the Koiter–Newton method in a finite ele-
ment environment to obtain a wider applicability of the approach.
In this section, we provide several examples which demonstrate
the performance of the proposed extended Koiter–Newton
approach to detect both, bifurcation points and branch-switching
after the bifurcation point. We compare our results to classical
non-linear path-following analysis results obtained from literature
and from ABAQUS [49] computations which use the full number of
equations of the finite element model. The discrete models in our
computations correspond to the models of the reference solutions,
with the same number of nodes and elements. The step size of the
proposed approach is indicated in the load-displacement diagrams
of the reference solution with markers.
Three of the selected examples are severe test cases for buck-
ling where the bifurcation branches are located prior to the limit
point of the equilibrium path. Both, ABAQUS and the original
Koiter–Newton method, failed to detect the bifurcation branches
and directly led to a limit-point-type buckling. With the fourth
example we have chosen a classical benchmark problem to study
secondary bifurcations. Finally, we tested our method for a
cylindrical roof structure which undergoes a snap-back and
branching behavior to reveal the overall performance of the
extensions for engineering relevant problems.
3.1. Straight beam example
The sketch of a single beam model is given in Fig. 4. For the
analysis, an isotropic material model was used with a Young's
modulus of 2000 MPa, a beam cross-section area A ¼ 391 mm2 and
a moment of inertia I ¼ 2000 mm4 . The structure was meshed with
16 beam elements. The non-linear response of the example is
characterized by two bifurcation branches which appear before
the limit point. The Newton arc-length method used in ABAQUS
directly led to the limit point and thus overestimated the load
carrying capability of the structure. Furthermore, it was found that
the original Koiter–Newton method also missed the bifurcation
branches [40]. In Chang's work [50], the first bifurcation branch of
this example was found using a linear bifurcation analysis to
predict the non-linear stability limits of structures. Here, the
proposed bifurcation indicator of the extended Koiter–Newton
method was used to reanalyze this example with the aim of
finding both of the bifurcation branches. The first buckling load of
this beam is separated from the others, hence only the first
buckling mode was chosen to construct a reduced order model of
dimension two.
We demonstrate the effect of the bifurcation indicator exten-
sion by comparison with the solution of the original Koiter–
Newton method and the ABAQUS solution for this problem. In a
first step we recall the solution of the original approach without
extension as depicted in Fig. 5. The solid line represents the
ABAQUS solution and the dotted curve shows the response curve
35 mm
400 mm
Fig. 4. Single beam model.
obtained from a single Koiter–Newton incremental load step. As
only one expansion step is applied at the undeformed state, and
the range of validity of the reduced order model is limited, the
response curve of the Koiter–Newton analysis is considered
accurate up to the triangle marker in the post-buckling regime, see
Fig. 5. Three consecutive corrector steps ensured equilibrium for
which the two curves fully coincide. To obtain the response curve
segment before the triangle marker, the Koiter–Newton method
required only a single step with 4 iterations of the full finite ele-
ment model, while the ABAQUS solution adopted 16 load steps
resulting in 43 iterations of the governing linear FE system. It can
be seen from Fig. 5 that both, the original Koiter–Newton method
and ABAQUS, led to a limit-point-type buckling load 100.25 N,
however the Koiter–Newton method is computationally much
more efficient.
Notation: In the following we indicate the converged Koiter–
Newton solution with markers on top of the reference solution to
identify the step size of the Koiter–Newton analysis.
In a second analysis we checked if the extended Koiter–Newton
method was capable to detect any bifurcation branch before the
limit point. During the solution of the reduced order model the
eigenvalues of the tangent stiffness revealed a first negative
eigenvalue at a total load of 20.47 N indicating a bifurcation branch
on the primary equilibrium path, see Fig. 6. Furthermore, it can be
seen from Fig. 6 that for a small physical perturbation, e.g. a load
150
100
ABAQUS
original K−N method (one step)
50
Load [N]
0
−50
−100
−150
0 10 20 30 40 50 60 70
Displacement [mm]
Fig. 5. Response curve of the single beam model, using nonlinear path-following
analysis.
Fig. 6. Response curve of the single beam model, using the proposed bifurcation
indicator.
 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138 135

perturbation, ABAQUS was also capable to detect the first bifur-
cation branch. The detected bifurcation point characterized the
failure behavior of the perfect structure and reduced the initially
predicted load carrying capability of 100.25 N by almost 400%. A
continuation of the bifurcation analysis with the Koiter–Newton
method revealed another bifurcation branch before the limit point,
a result which remained undetected by the ABAQUS solution.
3.2. Curved beam example
The detailed description of the curved beam loaded at its apex
by a concentrated load is illustrated in Fig. 7. The model has been
studied before by Li [51] and Meek [52] using improved arc-length
methods to trace the equilibrium path and to the detect the
intrinsic bifurcation branches. We conclude from their analyses
that only a single bifurcation point exists before the limit point. In
our computation with the extended Koiter–Newton method we
used 40 beam elements. The applied reduced order model of
dimension two was derived using the mode of the first buckling
load which was well separated from the succeeding modes.
The response curves of the vertical displacement at the load
position are shown in Fig. 8. The solid curve was achieved with
ABAQUS using 21 load steps with 52 iterations, while the original
Koiter–Newton method only used two load steps represented by
the markers, resulting in 8 iterations, to obtain the whole response
curve. Both, the original Koiter–Newton method and ABAQUS led
to a limit-point-type buckling at a load of 67.6 N, however, both
missed the bifurcation branch before the limit point. The reduced-
order eigenvalue analysis of the extended Koiter–Newton method
revealed the bifurcation branch in the second load step at a load of
60.3 N which corresponded exactly the result reported in [51].
3.3. Shallow arch example
In Fig. 9, the model of a shallow arch is depicted. The example
has been analyzed in [53] using Koiter's reduction method and in
[40] using the original Koiter–Newton method. The arch consists of
two flat plates, with a uniformly distributed pressure load along the
ridge line. The model was meshed with 160 shell elements using
isotropic material properties with E ¼ 1:0  106 MPa and ν ¼ 0:3.
Both reference solutions, [53] and [40] missed to detect the bifur-
cation branch before the limit point which was revealed with the
extended Koiter–Newton method during the initial load step. Again,
the reduced order model was constructed from a single, well
separated mode corresponding to the first buckling load.
The response curves of the vertical displacement at the load
position are shown in Fig. 10. The solid curve was achieved with
ABAQUS using 32 load steps with 73 iterations, while the original
Koiter–Newton method needed only three load steps to trace the
complete response curve, with a total of 12 iterations. Both, the
Koiter–Newton method and ABAQUS led to a limit-point-type
buckling at a load of 65.6 N. The proposed extended Koiter–
Newton method revealed a bifurcation branch in the first step,
prior to the limit point buckling, at a load of 50.99 N.
3.4. Flat plate under compression
A flat plate uniformly compressed in the longitudinal direction is a
classical benchmark that has been tested by many authors
[31,40,41,54]. The sketch of the plate is shown in Fig. 11. The plate was
meshed with 140 quadrilateral shell elements. The secondary bifur-
cation study which was done in [31] was reanalyzed using the pro-
posed bifurcation indicator in the framework of the Koiter–Newton
method. The plate exhibits two almost coincident buckling modes
with buckling loads of 343 N and 368 N respectively. To consider the
interactions of the first two closely spaced buckling modes, we used
three degrees of freedom in the reduced order model.

2.887 mm
R=3810 mm 0.346 mm
Y

Z
L=2540 mm Y

E = 68.98MPa X
A = 645 mm 2
L = 30 mm
I = 416200 mm 4
Fig. 7. Curved beam model. Fig. 9. Shallow arch model.

70 70

60 ABAQUS
60
original K−N (3 steps)
K-N: bifurcation branch
50 50
Load [N]

Load [N]

40 40

30 30

20 20
ABAQUS
10 original K−N (2 steps) 10
K-N: bifurcation branch

0 0
0 20 40 60 80 100 120 140 160 180 200 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Diaplacement [mm] Displacement [mm]

Fig. 8. Response curves of the curved beam. Fig. 10. Response curves of the shallow arch.
136 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138

It was demonstrated in [40] that a single Koiter–Newton step
with one iteration of the full FE model is enough to obtain the
complete response curve with sufficient accuracy up to the buck-
ling point. Hence, in the first Koiter–Newton step, the first and
second bifurcation branches, respectively, can be obtained as
shown in Fig. 12(a), using the proposed bifurcation indicator
whereas it is impossible for the general path-following techniques,
e.g. the arc-length method, to obtain the second bifurcation
branch. The intersection of these two branches is just the sec-
ondary bifurcation point, and the corresponding load value is
472.13 N and deviates by 8% from the value reported in [31]. The
error can be mainly contributed to the limited range of validity of
the reduced order in the post-buckling state, since the reduced
order model of the first expansion step was constructed at the
unloaded state.
In order to reduce the error we applied correction steps near
the bifurcation point of the initial post-buckling path before
applying the second expansion step. The effect of the additional
correction steps is depicted in Fig. 12(b). The load value at the
intersection point of the two bifurcation branches reduced to
439.51 N and thus stayed below a relative error of 1% with regard
to the reference value of 436.201 N, reported in [31]. Using a
second Koiter–Newton step improved the result significantly at a
total computational cost of five iterations of the full FE model.
3.5. Cylindrical roof
The cylindrical roof model depicted in Fig. 13 is a popular
benchmark to demonstrate the analysis of large deformation
a=140 mm
buckling of shells and post-buckling phenomena. The benchmark
problem has been studied by several researchers before [55–58],
all identifying a limit-point-type buckling and symmetric post-
buckling behavior using general path-following techniques. In
none of the aforementioned studies the physical buckling phe-
nomena was predicted correctly detecting a bifurcation point prior
to the snap-through response. Boutyour [33] and Zhou et al. [13]
revealed this phenomena using branch-switching methods which
were able to detect the bifurcation branch successfully.
The model was discretized with 200 quadrilateral shell ele-
ments resulting in a model size of 2646 degrees of freedom. The
complete fundamental equilibrium path was analyzed with ABA-
QUS in 63 load step and 123 iterations while the original Koiter–
Newton method needed 7 load steps and 26 iterations in total,
both missing the bifurcation branch and directly leading to the
limit point. In a refined analysis using the extended Koiter–New-
ton method to reveal the bifurcation point and corresponding
bifurcation branches the result reported in [13] was reproduced
exactly. The extended Koiter–Newton method used six steps and
22 iterations to trace the bifurcation branch 1 (A-B-C) compared to
116 iterations of the full FE system needed with the branch-
switching method of [13]. It should be noted that the method
proposed in [13] required a tremendous numerical effort due to
the repeated eigenvalue analysis of the full FE system in each load
step. At the bifurcation point a buckling load of 530.69 N was
found which corresponds to a critical load of 90% of the limit load
(593.13 N).
free
F
hinged
A1
A2

free hinged
b=100 mm
L = 508mm

R = 2540mm θ = 0.1 rad E = 3102.75 MPa

E = 70000MPa
v = 0.3
v = 0.3 tickness, t = 6.35mm
thickness = 0.5 mm

Fig. 11. Flat plate model. Fig. 13. Cylindrical roof model.

600
600

500 500
secondary bifurcation point: 472.13 N

400 secondary bifurcation point: 439.508 N

400
Load [N]

Load [N]

300 300

200 200

100 100
second bifurcation branch (one K−N step) second bifurcation branch (two K−N steps)
0 0
0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18 0 0.02 0.04 0.06 0.08 0.1 0.12 0.14 0.16 0.18
Out−of−plane displacement [mm] Out−of−plane displacement [mm]
One Koiter-Newton step Two Koiter-Newton steps
Fig. 12. Secondary bifurcation study of the plat plate.
 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138 137

800 800

600 600
B

400 400

Load [N]
200
Load [N]

200
D
0 0 A
ABAQUS
bifurcation path 1
in references
−200 −200 bifurcation path 2
C
E in references
bifurcation path 1
−400 −400 using K-N
ABAQUS
bifurcation path 2
original K−N using K-N
−600 −600
0 5 10 15 20 25 30 35 0 5 10 15 20 25 30 35
Displacement [mm] Displacement [mm]
Fundamental path Bifurcation branches

Fig. 14. Response curves of the cylindrical roof. (For interpretation of the references to color in this figure caption, the reader is referred to the web version of this paper.)

Buckling mode for limit point Buckling mode for bifurcation branch 1

Fig. 15. Limit-point buckling mode and bifurcation buckling mode.

The critical instability modes obtained from the two different
computations, the original Koiter–Newton method and the
extended version, are depicted in Fig. 15. The mode representing
limit-point buckling clearly shows symmetry at both center
planes, A1 and A2, cf. Fig. 15, whereas the bifurcation buckling
mode is asymmetric to the center plane A2, which corresponds to
the results presented in [13].
The proposed analysis strategy of the bifurcation indicator also
works in the case of multiple bifurcations and is considered in the
current shell roof example. To this end, a few more degrees of
freedom (4 DOFs) were considered for the construction of the
reduced order model and the bifurcation analysis was repeated
with a path tracing of the first bifurcation branch. The bifurcation
path 2 (A-B-D-E) is plotted as a green curve in Fig. 14(b).
4. Conclusions
The focus of the work presented in this paper was on the
reliable detection of bifurcation points and the tracing of corre-
sponding bifurcation branches that lie before limit point buckling
in the framework of a reduced order model. Both of the reference
methods considered in this paper, ABAQUS and the original
Koiter–Newton method, failed to predict correctly the bifurcation
response and instead led immediately to the limit-point-type
buckling which generally overestimates significantly the load
carrying capability of the structure. With an eigenanalysis-based
extension to the reduced order model we equipped the Koiter–
Newton method with the necessary capability to overcome this
shortcoming. The proposed bifurcation indicator of the extended
Koiter–Newton method predicts reliably all bifurcation branches
with convincing results with regard to accuracy and computa-
tional effort. During the solution process a repeated eigenvalue
analysis of the tangent stiffness matrix of the reduced order
model is performed at each iteration step. The very small
dimension of the tangent stiffness matrix, typically involving
only a few degrees of freedom, ensures the solution effort of the
eigenvalue problem to be insignificant, even for large discrete
analysis models. On basis of the computed eigenvalue spectrum
the existence of a bifurcation branch which intersects the pri-
mary equilibrium path is easily detected by a negative eigenva-
lue. The corresponding eigenmode is used as perturbation of the
displacement field to allow the primary path to branch out to a
secondary equilibrium path.
We demonstrated the potential of the proposed method and
the high quality of the analysis results with a set of benchmark
problems, some of which are considered to be severe test cases.
Both, reliability and accuracy of the approach are remarkable. The
method worked for all cases where established standard methods
failed. Furthermore, we demonstrated the practical relevance and
applicability of the method with a shell test problem from engi-
neering practice which revealed reliable results even in the pre-
sence of snap-back phenomena.
138 K. Liang et al. / International Journal of Non-Linear Mechanics 81 (2016) 129–138
Acknowledgment
The research leading to these results received funding from the
European Community's Seventh Framework Programme ([FP7/
2007-2013]) under Grant agreement no282522. This research work
was also sponsored by Laboratory Independent Innovation project
of Qian Xuesen Laboratory of Space Technology. Thank you Mir-
anda Aldham-Breary for reviewing the English text carefully.
References
[1] R. Degenhardt, A. Kling, A. Bethge, J. Orf, L. Kärger, R. Zimmermann, K. Rohwer,
A. Calvi, Investigations on imperfection sensitivity and deduction of improved
knock-down factors for unstiffened CFRP cylindrical shells, Compos. Struct. 92
(8) (2010) 1939–1946.
[2] S.G. Castroa, R. Zimmermannb, M.A. Arbeloa, R. Degenhardt, Exploring the
constancy of the global buckling load after a critical geometric imperfection
level in thin-walled cylindrical shells for less conservative knock-down factors,
Thin-Walled Struct. 72 (10) (2013) 76–87.
[3] K.M. Liew, Z.X. Lei, J.L. Yu, L.W. Zhang, Postbuckling of carbon nanotube-
reinforced functionally graded cylindrical panels under axial compression
using a meshless approach, Comput. Methods Appl. Mech. Eng. 268 (1) (2014)
1–17.
[4] M.W. Hilburger, J.H. Starnesjr, Buckling behavior of compression-loaded
composite cylindrical shells with reinforced cutouts, Int. J. Non-Linear Mech.
40 (7) (2005) 1005–1021.
[5] L. Kwasniewskia, A. Ziolkowski, Simplified analytical model and numerical
simulations of finite-disturbance buckling of columns, Int. J. Non-Linear Mech.
51 (9) (2013) 121–131.
[6] T. Kubiak, Static and Dynamic Buckling of Thin-Walled Plate Structures,
Springer Cham, Heidelberg, New York, Dordrecht, London, 2013.
[7] H. Kalathur, T.M. Hoang, R.S. Lakes, W.J. Drugan, Buckling mode jump at very
close load values in unattached flat-end columns: theory and experiment, J.
Appl. Mech. 81 (4) (2013) 041010-1–041010-8.
[8] N.P. Semenyuk, A.I. Morenko, M.J.A. Smith, On the stability and postbuckling
behavior of shells with corrugated cross sections under external pressure, J.
Appl. Mech. 81 (1) (2013) 011002-1–011002-8.
[9] M.A. Crisfield, An arc-length method including line search and acceleration,
Int. J. Numer. Methods Eng. 19 (9) (1983) 1269–1289.
[10] M.A. Crisfield, A fast incremental/iterative solution procedure that handles
snap-through, Comput. Struct. 13 (81) (1981) 55–62.
[11] B.W.R. Forde, S.F. Stiemer, Improved arc-length orthogonality methods for
nonlinear finite element analysis, Comput. Struct. 27 (5) (1987) 625–630.
[12] C.J. Stull, J.M. Nichols, C.J. Earls, Stochastic inverse identification of geometric
imperfections in shell structures, Comput. Methods Appl. Mech. Eng. 200 (25)
(2011) 2256–2267.
[13] Y. Zhou, I. Stanciulescu, T. Eason, M. Spottswood, Nonlinear elastic buckling
and postbuckling analysis of cylindrical panels, Finite Elements Anal. Des. 96
(2015) 41–50.
[14] M. Papadrakskis, Solving large-scale nonlinear problems in solid and struc-
tural mechanics, in: Solving Large-Scale Problems in Mechanics, John Wiley &
Sons, Chichester, England, 1993.
[15] Y.T. Feng, D. Peric, D.R.J. Owen, Determination of travel directions in path-
following methods, Math. Comput. Model. 21 (7) (1995) 43–59.
[16] Y.T. Feng, D. Peric, D.R.J. Owen, A new criterion for determination of initial
loading parameter in arc-length methods, Comput. Struct. 58 (3) (1996)
479–485.
[17] Y.T. Feng, D.R.J. Owen, D. Peric, On the sign of the determinant of the structural
stiffness matrix for determination of loading increment in arc-length methods,
Commun. Numer. Methods Eng. 13 (1) (1997) 47–49.
[18] A. Eriksson, On some path-related measures for non-linear structural F.E.
problems, Int. J. Numer. Methods Eng. 26 (8) (1988) 1791–1803.
[19] W. Wagner, P. Wriggers, A solution method for the calculation of postcritical
branches, Eng. Comput. 5 (2) (1988) 103–109.
[20] F. Fujii, H. Noguchi, The buckling mode extracted from the LDLT -decomposed
large-order stiffness matrix, Commun. Numer. Methods Eng. 18 (7) (2002)
459–467.
[21] H. Noguchi, F. Fujii, Eigenvector-free indicator, pinpointing and branch-
switching for bifurcation, Commun. Numer. Methods Eng. 19 (2003) 445–457.
[22] B.L. Wardle, P.A. Lagace, M.A. Tudela, Buckling response of transversely loaded
composites shells, Part2: numerical analysis, AIAA J. 42 (7) (2004) 1465–1473.
[23] B.L. Wardle, Solution to the incorrect benchmark shell-buckling problem, AIAA
J. 46 (2) (2008) 381–387.
[24] N. Damil, M. Potier-Ferry, A new method to compute perturbed bifurcations:
application to the buckling of imperfect elastic structures, Int. J. Eng. Sci. 28 (9)
(1990) 943–957.
[25] L. Azrar, B. Cochelin, N. Damil, M. Potier-Ferry, An asymptotic-numerical
method to compute the postbuckling behaviour of elastic plates and shells,
Int. J. Numer. Methods Eng. 36 (8) (1993) 1251–1277.
[26] D.A. Nagy, M. Konig, Geometrically nonlinear finite element behaviour using
buckling mode superposition, Comput. Methods Appl. Mech. Eng. 19 (3)
(1979) 447–484.
[27] B.O. Almroth, P. Stern, F.A. Brogan, Automatic choice of global shape functions
in structural analysis, AIAA J. 16 (5) (1978) 525–528.
[28] A.K. Noor, J.M. Peters, Reduced basis technique for nonlinear analysis of
structures, AIAA J. 18 (4) (1980) 455–462.
[29] A.S.L. Chan, K.M. Hsiao, Nonlinear analysis using a reduced number of vari-
ables, Comput. Methods Appl. Mech. Eng. 52 (1) (1985) 899–913.
[30] B. Cochelin, A path-following technique via an asymptotic-numerical method,
Comput. Struct. 53 (5) (1994) 1181–1192.
[31] P. Vannucci, B. Cochelin, N. Damil, M. Potier-Ferry, An asymptotic-numerical
method to compute bifurcating branches, Int. J. Numer. Methods Eng. 41 (8)
(1998) 1365–1389.
[32] M. Jamal, H. Elasmar, B. Braikat, Bifurcation indicators, Acta Mech. 139 (1)
(2000) 129–142.
[33] E.H. Boutyour, H. Zahrouni, M. Potier-Ferry, M. Boudi, Bifurcation points and
bifurcated branches by an asymptotic numerical method and Padé approx-
imants, Int. J. Numer. Methods Eng. 60 (12) (2004) 1987–2012.
[34] B. Cochelin, N. Damil, M. Potier-Ferry, Asymptotic-numerical methods and
Padé approximants for nonlinear elastic structures, Int. J. Numer. Methods
Eng. 37 (7) (1994) 1187–1213.
[35] F. Xu, M. Potier-Ferry, S. Belouettar, C. Yu, 3D finite element modeling for
instabilities in thin films on soft substrates, Int. J. Solids Struct. 51 (21–22)
(2014) 3619–3632.
[36] F. Xu, M. Potier-Ferry, S. Belouettar, H. Hu, Multiple bifurcations in wrinkling
analysis of thin films on compliant substrates, Int. J. Non-Linear Mech. 76
(2015) 203–222.
[37] E.H. Boutyour, H. Zahrouni, M. Potier-Ferry, M. Boudi, Asymptotic-numerical
method for buckling analysis of shell structures with large rotations, J. Com-
put. Appl. Math. 168 (1–2) (2004) 77–85.
[38] S. Lopez, An effective parametrization for asymptotic extrapolations, Comput.
Methods Appl. Mech. Eng. 189 (1) (2000) 297–311.
[39] H. Mottaqui, B. Braikat, N. Damil, Discussion about parameterization in the
asymptotic numerical method: application to nonlinear elastic shells, Comput.
Methods Appl. Mech. Eng. 199 (2010) 1701–1709.
[40] K. Liang, M. Abdalla, Z. Gürdal, A Koiter–Newton approach for nonlinear
structural analysis, Int. J. Numer. Methods Eng. 96 (12) (2013) 763–786.
[41] K. Liang, M. Ruess, M. Abdalla, The Koiter–Newton approach using von Kármán
kinematics for buckling analyses of imperfection sensitive structures, Comput.
Methods Appl. Mech. Eng. 279 (1) (2014) 440–468.
[42] K. Liang, A Koiter–Newton arc-length method for buckling sensitive structures
(Ph.D. thesis), Delft University of Technology, 2013.
[43] G. Garcea, R. Casciaro, G. Attanasio, F. Giordano, Perturbation approach to
elastic post-buckling analysis, Comput. Struct. 66 (5) (1998) 585–595.
[44] G. Garcea, Mixed formulation in Koiter analysis of thin-walled beams, Comput.
Struct. 190 (26–27) (2001) 3369–3399.
[45] T. Rahman, E.L. Jansen, Finite element based multi-mode initial post-buckling
analysis of composite cylindrical shells, in: 50th AIAA/ASME/ASCE/AHS/ASC
Structures, Structural Dynamics and Materials Conference, 2009.
[46] B. Geier, A method to compute Koiter's b-factor of anisotropic panels, Int. J.
Non-Linear Mech. 37 (4–5) (2002) 699–707.
[47] R. Seydel, Numerical computation of branch points in nonlinear equations,
Numer. Math. 33 (3) (1979) 339–352.
[48] M. Ruess, P.J. Pahl, An extended QR-solver for large profiled matrices, Int. J.
Numer. Methods Eng. 79 (11) (2009) 1419–1442.
[49] Abaqus 6.11, ABAQUS Theory Manual, Section 2.3.2, 2011.
[50] S.C. Chang, J.J. Chen, Effectiveness of linear bifurcation analysis for predicting
the nonlinear stability limits of structures, Int. J. Numer. Methods Eng. 23 (5)
(1986) 831–846.
[51] Y.Q. Li, Z.Y. Shen, Improvements on the arc-length-type method, Acta Mech.
Sin. 20 (5) (2004) 541–550.
[52] J.L. Meek, S. Loganathan, Large displacement analysis of space-frame struc-
tures, Comput. Methods Appl. Mech. Eng. 72 (1) (1989) 57–75.
[53] G. Zagari, Koiter's asymptotic numerical methods for shell structures using a
corotational formulation (Ph.D. thesis), Universita Della Calabria, 2009.
[54] P. Tiso, Finite element based reduction methods for static and dynamic ana-
lysis of thin-walled structures (Ph.D. thesis), Delft University of Technology,
2006.
[55] C.H. Yeom, S.W. Lee, An assumed strain finite element model for large
deflection composite shells, Int. J. Numer. Methods Eng. 28 (8) (1989)
1749–1768.
[56] C. Sansour, H. Bufler, Exact finite rotation shell theory, its mixed variational
formulation and its finite element implementation, Int. J. Numer. Methods
Eng. 34 (1) (1992) 73–115.
[57] K.Y. Sze, X.H. Liu, S.H. Lo, Popular benchmark problems for geometric non-
linear analysis of shells, Finite Elements Anal. Des. 40 (11) (2004) 1551–1569.
[58] L.G. Andrade, A.M. Awruch, I.B. Morsch, Geometrically nonlinear analysis of
laminate composite plates and shells using the eight-node hexahedral ele-
ment with one-point integration, Compos. Struct. 79 (4) (2007) 571–580.