Co-rotational beam elements

in instability problems

by

Jean-Marc Battini

Department of Mechanics

January 2002
Technical Reports from
Royal Institute of Technology
Department of Mechanics
SE-100 44 Stockholm, Sweden
Akademisk avhandling som med tillstånd av Kungliga Tekniska Högskolan i Stock-
holm framlägges till offentlig granskning för avläggande av teknologie doktorsexamen
fredagen den 18:e januari 2002 kl 13.00 i Kollegiesalen, Administrationsbyggnaden,
Kungliga Tekniska Högskolan, Valhallavägen 79, Stockholm.

Jean-Marc
c Battini 2002
 To Amra

iii
Abstract

The purpose of the work presented in this thesis is to implement co-rotational beam
elements and branch-switching procedures in order to analyse elastic and elasto-
plastic instability problems.
For the 2D beam elements, the co-rotational framework is taken from Crisfield [23].
The main objective is to compare three different local elasto-plastic elements.
The 3D co-rotational formulation is based on the work of Pacoste and Eriksson [73],
with new items concerning the parameterisation of the finite rotations, the definition
of the local frame, the inclusion of warping effects through the introduction of a
seventh nodal degree of freedom and the consideration of rigid links. Different
types of local formulations are considered, including or not warping effects. It is
shown that at least some degree of non-linearity must be introduced in the local
strain definition in order to obtain correct results for certain classes of problems.
Within the present approach any cross-section can be modelled, and particularly,
the centroid and shear center are not necessarily coincident.
Plasticity is introduced via a von Mises material with isotropic hardening. Numer-
ical integration over the cross-section is performed. At each integration point, the
constitutive equations are solved by including interaction between the normal and
shear stresses.
Concerning instabilities, a new numerical method for the direct computation of elas-
tic critical points is proposed. This is based on a minimal augmentation procedure as
developed by Eriksson [32–34]. In elasto-plasticity, a literature survey, mainly con-
cerned with theoretical aspects is first presented. The objective is to get a complete
comprehension of the phenomena and to give a basis for the two branch-switching
procedures presented in this thesis.
A large number of examples are used in order to assess the performances of the
elements and the path-following procedures.

Keywords: instability, co-rotational method, branch-switching, beam element,

warping, plastic buckling, post-bifurcation.

v
List of papers

The work presented in this thesis is based on 6 papers and a licentiate thesis accord-
ing to the list below. The correspondence between the chapters of this manuscript
and these previous publications is given in Section 1.2.

Journal papers

J.-M. Battini and C. Pacoste

Co-rotational beam elements with warping effects in instability problems
Accepted by Computer Methods in Applied Mechanics and Engineering

J.-M. Battini and C. Pacoste

Plastic instability of beam structures using co-rotational elements
Submitted to Computer Methods in Applied Mechanics and Engineering

J.-M. Battini, C. Pacoste and A. Eriksson

Minimal augmentation procedure for the direct computation of critical points
Submitted to Computer Methods in Applied Mechanics and Engineering

Conference papers

C. Pacoste, J.-M. Battini and A. Eriksson

Parameterisation of rotations in co-rotational elements
Proceedings Euromech 2000, Metz

J.-M. Battini and C. Pacoste

Eléments poutres co-rotationels avec gauchissement
Proceedings CSMA 5ème colloque national en calcul des structures, Giens 2001

C. Pacoste and J.-M. Battini

Calcul des chemins d´équilibres post-critiques
Proceedings CSMA 5ème colloque national en calcul des structures, Giens 2001

Licentiate thesis: J.-M. Battini

Plastic instability analysis of plane frames using a co-rotational approach
Department of Structural Engineering, KTH, Stockholm 1999

vii
Preface

The research reported in the present thesis was carried out first at the Department
of Structural Engineering and later at the Department of Mechanics, at the Royal
Institute of Technology in Stockholm.
The work was initiated by Associate Professor Costin Pacoste and conducted under
his supervision.
This project was financed by KTH and partially by the Swedish Research Council
for Engineering Sciences (TFR).
First of all, I want to express my gratitude to my supervisor Costin Pacoste for
his encouragement, valuable advice and for always having time for discussions. I
want particularly to thank him for continuing to help me during the writing of this
manuscript and the preparation of the defence, although he had resigned his position
at KTH in October 2001.
I also wish to thank Professor Anders Eriksson for reviewing the manuscript of this
thesis, providing valuable comments for improvement, and specially for accepting to
present me to the doctoral examination.
I am also grateful to Gunnar Tibert who helped me with the LaTex code and spent
several hours in proof reading this manuscript.
Finally, I would like to thank Amra for her love and support.

Stockholm, December 2001

Jean-Marc Battini

ix
Contents

Abstract v

List of papers vii

Preface ix

1 Introduction 1
1.1 Aims and scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 General structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Plastic instabilities – review 7

2.1 Simple models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.1 Shanley’s column . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.1.2 Hutchinson’s model . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Hill’s criterion of uniqueness . . . . . . . . . . . . . . . . . . . . . . . 16
2.2.1 Linear comparison solid . . . . . . . . . . . . . . . . . . . . . 17
2.3 Euler beam analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.1 Elastic case . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3.2 Reduced versus tangent modulus load . . . . . . . . . . . . . . 19
2.3.3 Theoretical plastic analysis . . . . . . . . . . . . . . . . . . . . 22
2.4 Discretised systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.4.1 Constitutive framework and discretised rate problem . . . . . 28
2.4.2 Tangent comparison solid . . . . . . . . . . . . . . . . . . . . 30
2.4.3 Bifurcation and stability . . . . . . . . . . . . . . . . . . . . . 30
2.4.4 Energy approach . . . . . . . . . . . . . . . . . . . . . . . . . 32

xi
3 2D beam element formulation 35
3.1 Co-rotational framework . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.1.1 Beam kinematics . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1.2 Virtual displacements . . . . . . . . . . . . . . . . . . . . . . . 38
3.1.3 Internal forces . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1.4 Tangent stiffness matrix . . . . . . . . . . . . . . . . . . . . . 39
3.2 Local linear Bernoulli element . . . . . . . . . . . . . . . . . . . . . . 40
3.2.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.2.2 Gauss integration . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.3 Local internal forces . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.4 Local tangent stiffness matrix . . . . . . . . . . . . . . . . . . 42
3.2.5 Constitutive equations . . . . . . . . . . . . . . . . . . . . . . 43
3.3 Local shallow arch Bernoulli element . . . . . . . . . . . . . . . . . . 43
3.3.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.2 Local internal forces . . . . . . . . . . . . . . . . . . . . . . . 43
3.3.3 Local tangent stiffness matrix . . . . . . . . . . . . . . . . . . 44
3.4 Local linear Timoshenko element . . . . . . . . . . . . . . . . . . . . 46
3.4.1 Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.4.2 Local internal forces . . . . . . . . . . . . . . . . . . . . . . . 46
3.4.3 Local tangent stiffness matrix . . . . . . . . . . . . . . . . . . 47
3.4.4 Constitutive equations . . . . . . . . . . . . . . . . . . . . . . 47

4 3D beam element formulation 55

4.1 Parameterisation of finite 3D rotations . . . . . . . . . . . . . . . . . 56
4.1.1 Rotational vector . . . . . . . . . . . . . . . . . . . . . . . . . 58
4.1.2 Incremental rotation vector . . . . . . . . . . . . . . . . . . . 59
4.2 Co-rotational framework . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.2.1 Coordinate systems, beam kinematics . . . . . . . . . . . . . . 60
4.2.2 Change of variables δ ϑ̄ −→ δ θ̄ . . . . . . . . . . . . . . . . . . 63
4.2.3 Change of variables δpa −→ δpgg . . . . . . . . . . . . . . . . 64

xii
 4.2.4 Eccentric nodes, rigid links . . . . . . . . . . . . . . . . . . . . 70
4.2.5 Finite rotation parameters . . . . . . . . . . . . . . . . . . . . 71
4.2.6 Formulation with warping . . . . . . . . . . . . . . . . . . . . 73
4.3 Local element formulation in elasticity . . . . . . . . . . . . . . . . . 74
4.3.1 Local beam kinematics, strain energy . . . . . . . . . . . . . . 75
4.3.2 Element types . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.3.3 Beams with thin-walled or open cross-sections . . . . . . . . . 79
4.3.4 Beams with solid or closed cross-sections . . . . . . . . . . . . 80
4.4 Local element formulation in elasto-plasticity . . . . . . . . . . . . . . 80
4.4.1 Strain definition . . . . . . . . . . . . . . . . . . . . . . . . . . 81
4.4.2 Finite element formulation . . . . . . . . . . . . . . . . . . . . 82
4.4.3 Beams with thin-walled or open cross-sections . . . . . . . . . 82
4.4.4 Beams with solid or closed cross-section . . . . . . . . . . . . 84
4.5 Applied loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5.1 Eccentric forces . . . . . . . . . . . . . . . . . . . . . . . . . . 85
4.5.2 External moments . . . . . . . . . . . . . . . . . . . . . . . . 85

5 Path following techniques 87

5.1 Non-critical equilibrium path . . . . . . . . . . . . . . . . . . . . . . 88
5.1.1 Procedure at the structural level . . . . . . . . . . . . . . . . . 88
5.1.2 Procedures at the element level . . . . . . . . . . . . . . . . . 90
5.2 Direct computation of elastic critical points . . . . . . . . . . . . . . 91
5.2.1 Classical approach . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2.2 Alternative extended system . . . . . . . . . . . . . . . . . . . 92
5.2.3 Modified algorithm . . . . . . . . . . . . . . . . . . . . . . . . 94
5.2.4 Initialisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.3 Branch-switching in elasto-plasticity . . . . . . . . . . . . . . . . . . . 97
5.3.1 Mode injection method . . . . . . . . . . . . . . . . . . . . . . 97
5.3.2 Minimisation procedure . . . . . . . . . . . . . . . . . . . . . 98

xiii
6 2D examples in elasto-plasticity 101
6.1 Example 1: cantilever beam . . . . . . . . . . . . . . . . . . . . . . . 101
6.2 Example 2: Lee’s frame . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.3 Example 3: Euler beam . . . . . . . . . . . . . . . . . . . . . . . . . 107
6.4 Example 4: asymmetrical frame . . . . . . . . . . . . . . . . . . . . . 112
6.5 Example 5: toggle frame . . . . . . . . . . . . . . . . . . . . . . . . . 114

7 3D examples in elasticity 117

7.1 Example 1, deployable ring . . . . . . . . . . . . . . . . . . . . . . . . 117
7.2 Example 2, helical beam . . . . . . . . . . . . . . . . . . . . . . . . . 119
7.3 Example 3, right angle frame . . . . . . . . . . . . . . . . . . . . . . 120
7.4 Example 4, non-linear torsion . . . . . . . . . . . . . . . . . . . . . . 122
7.5 Example 5, channel-section beam . . . . . . . . . . . . . . . . . . . . 123
7.6 Example 6, torsional buckling . . . . . . . . . . . . . . . . . . . . . . 124
7.7 Example 7, lateral torsional buckling . . . . . . . . . . . . . . . . . . 126
7.8 Example 8, lateral buckling of a cantilever . . . . . . . . . . . . . . . 127
7.9 Example 9, lateral buckling with warping . . . . . . . . . . . . . . . . 130
7.10 Example 10, L frame . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

8 Elastic isolation – examples 133

8.1 Example 1, 2D clamped arch . . . . . . . . . . . . . . . . . . . . . . . 133
8.2 Example 2, 2D simply supported arch . . . . . . . . . . . . . . . . . . 134
8.3 Example 3, 2D shallow arch . . . . . . . . . . . . . . . . . . . . . . . 136
8.4 Example 4, deep circular arch . . . . . . . . . . . . . . . . . . . . . . 137
8.5 Example 5, narrow cantilever . . . . . . . . . . . . . . . . . . . . . . 138
8.6 Example 6, L frame . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
8.7 Example 7, cable hockling . . . . . . . . . . . . . . . . . . . . . . . . 141
8.8 Example 8, spatial frame . . . . . . . . . . . . . . . . . . . . . . . . . 142

9 3D examples in elasto-plasticity 145

9.1 Example 1, channel-section beam . . . . . . . . . . . . . . . . . . . . 145

xiv
 9.2 Example 2, framed dome . . . . . . . . . . . . . . . . . . . . . . . . . 146
9.3 Example 3, lateral buckling of an I beam . . . . . . . . . . . . . . . . 150
9.4 Example 4, lateral buckling of a T cantilever . . . . . . . . . . . . . . 152
9.5 Example 5, buckling by torsion and flexure . . . . . . . . . . . . . . . 156

10 Conclusions and future research 161

10.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
10.2 Future research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

Bibliography 165

A 3D elastic local formulation 173

A.1 t3d element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
A.2 tw3d element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
A.3 b3d element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
A.4 bw3d element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

B Warping function 177

B.1 Basic equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
B.2 Saint-Venant torsion . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
B.3 Determination of the warping function . . . . . . . . . . . . . . . . . 178
B.4 Cross-section quantities . . . . . . . . . . . . . . . . . . . . . . . . . . 180

xv
Chapter 1

Introduction

The analysis of structural instabilities is an important part of the design process in

civil, mechanical and aeronautical engineering. Despite the great interest surround-
ing these problems, most of commercial finite elements codes can model such phe-
nomena only partially. As a matter of fact, in these programs bifurcation loads are
often calculated through a linearised buckling analysis, which may give inaccurate
results in certain cases. Concerning post-bifurcation paths, these are often studied
by introducing small initial imperfections. In addition to the difficulties related to
the choice of the form and the magnitude of these imperfections, such an approach
removes the bifurcation and does not allow a complete physical understanding.
However, an accurate evaluation of bifurcation points is necessary for two different
purposes. First, these points define critical conditions for the functionality of struc-
tures. Second, in order to get a complete description of the instability, secondary
paths must be computed using perfect structures. This requires on one hand a
procedure for detecting and isolating bifurcation points along fundamental paths,
and, on the other hand, a procedure for performing branch-switching to secondary
paths. One of the purposes of this thesis is to develop such procedures for elastic
and elasto-plastic cases.
Most of the work done about instability concerns elastic structures under quasi-static
loading. The reason is that the analysis of such problems requires the introduction
of geometrical non-linearities which, in itself, is a rather complicated task to work
with. However, the assumption that the structure behaves elastically up to the
bifurcation point may not always hold. If some part of the structure develops pre-
buckling plastic deformations, then the analysis of the instability must include both
geometrical and material non-linearities.
Stability analysis of inelastic structures is complicated by the fact that the princi-
ple of minimum potential energy, which is the basic tool for the stability analysis
of elastic structures, cannot be applied. This is due to the dissipation of energy
inherent in plastic deformations. Consequently, the theory developed by Koiter or
the more refined catastrophe theory, which proposes a classification of the critical
points as well as methods to investigate post-bifurcation paths and imperfection
sensitivity aspects, cannot be applied. In fact, the irreversibility of plastic deforma-

1
tions produces new phenomena. While in elasticity bifurcation occurs at isolated
critical points and is characterised by a loss of stability on the fundamental path, in
plasticity, bifurcation along the fundamental path may occur at a continuous range
of equilibrium points.
Although that the first comprehensive description of this phenomenon has been
given in 1947 by Shanley [88], the complete description of bifurcation and insta-
bility in time-independent plasticity was mostly developed in the two last decades.
The absence until recent years of such a theoretical basis and the complexity of the
problem explain why most of the books about stability of structures treat plastic
instabilities only superficially or even not at all. It explains also why so little nu-
merical research has been performed concerning the topic. Most of the work with
finite elements concern elastic structures and often under quasi-static evolutions.
Naturally, finite beam elements are very common: a huge amount of research has
been carried out within this topic and commercial codes propose often several el-
ements which include both geometrical and material non-linearities. However, two
problems remain. The first one concerns warping effects which are usually not in-
cluded in non-linear formulations; if they are considered, it is often by assuming
bi-symmetric cross-sections. The second problem is related to plastic instability
problems in the sense that most of elements are too crude to model correctly such
phenomena. As an example, it will be shown in this thesis that elements which ne-
glect hardening or use yield criteria expressed in function of stress resultants cannot
be used. In fact, concerning beam elements, attempts to correctly model plastic
bifurcation problems and compute post-bifurcation paths are not very common in
literature. Thus another purpose of this thesis is to develop efficient non-linear beam
elements which can include warping effects for arbitrary cross-sections and which
are accurate enough in order to model elastic and elasto-plastic instability problems.
In this context, the co-rotational approach has generated an increased amount of
interest in the last decade. However, most of the work done on co-rotational beam
elements concern trivial cross-sections, as rectangular ones. It appears then inter-
esting to investigate the co-rotational approach in order to introduce warping effects
and arbitrary cross-sections.

1.1 Aims and scope

An important research concerning numerical stability analysis of elastic structures

under quasi-static loading has been carried out in the last years by the Structural
Mechanics Group at KTH. Based on a co-rotational approach and under the as-
sumption of small deformations, efficient beam [72,73] and shell elements [59,68,71]
have been developed in order to model large displacements problems in general
and stability problems in particular. At the same time, advanced path-following
methods [30–34, 65] including branch-switching procedures and parameter sensitiv-
ity analyses (fold lines) have been implemented. An incursion into dynamics [37]
has also been performed.

2
The first aim of this thesis was to introduce material non-linearity in this previous
work and to investigate how the co-rotational beam elements and the path-following
procedures have to be modified in order to account for plastic deformations. For
this purpose, the co-rotational approach is well suited since it leads to an artificial
separation of the material and geometrical non-linearities. Consequently, only local
internal force vectors and tangent stiffness matrices need to be modified.
The work first focused on 2D beam elements. Three local elasto-plastic formulations
have been developed and tested. Based on the Bernoulli assumption, the first two
local elements use a linear and a shallow arch local strain definition, respectively.
The third element is based on the Timoshenko assumption with linear interpolations.
Concerning path following aspects, two methods of branch-switching in elasto-
plasticity have been implemented. In the first one, branch-switching is operated
by using as predictor the eigenvector associated to the negative eigenvalue at the
bifurcation point. In the second one, introduced by Petryk [76–81], an energy ap-
proach is used to select automatically the stable post-bifurcation path.
Before dealing with numerical aspects, a literature review on plastic instabilities
was carried out. The first objective was to get a complete picture of the physical
phenomena involved, and thus enable a correct numerical modelling. For that, the
models of Shanley and Hutchinson have been carefully analysed. The second objec-
tive was to get the theoretical background for the two branch-switching procedures.
This background has been provided by the works of Hill [45] and Petryk [76–81].
This work constitutes the Licentiate Thesis, Plastic instability analysis of plane
frames using a co-rotational approach [5], presented by the author in June 1999.
The second aim of this thesis was to further develop the 3D beam elements developed
by Pacoste and Eriksson [73] and also to incorporate material non-linearity. With
respect to this previous publication, the new items concerning the co-rotational
framework are a new definition of the local frame, the use of the spatial form of
the incremental rotational vector to parameterise finite rotations, the inclusion of
warping effects through the introduction of a seventh nodal degree of freedom and
the consideration of rigid links. As regard the local formulation, a systematic study
of partly non-linear expressions for beam deformations has been carried out and it
has been shown that some degree of non-linearity must be introduced in the local
strain definition in order to obtain correct results for certain problems involving
torsional effects.
With these improvements, the elements presented in this thesis can be used to model
any problems involving large displacements and rotations, under the assumption of
small strains. In addition, arbitrary beam cross-sections can be considered and
particularly, the centroid and shear center are not necessarily coincident, as it is
often assumed in non-linear beam elements.
The third aim of this thesis was to adapt the work of Eriksson [32–34] about fold line
algorithms and develop a new procedure for the direct computation of elastic critical
points. Compared to the classical approach of Wriggers et al. [104, 105], two main

3
modifications have been introduced. First, following Eriksson [32–34], the condition
of criticality is expressed by a scalar equation instead of a vectorial one. Next, the
present procedure does not use exclusively the extended system obtained from the
equilibrium equations and the criticality condition, but also introduces intermediate
iterations based purely on equilibrium equations under load or displacement control.
Finally, concerning the modelling of inelastic instabilities, the choice between an
incremental flow plastic theory and a total strain or deformation theory must be
discussed. As a matter of fact, experimental results have paradoxically but persis-
tently shown that the deformation theory is superior to the flow one in predicting
plastic buckling loads for certain problems, e.g. the inelastic axial-torsional buck-
ling of cruciform columns. However, contrary to the flow theory which relates the
increment of plastic strains to the stresses so that the plastic strains depend on
the loading history, the deformation theory relates the total plastic strains to the
stresses and the plastic strains are independent of the loading history. Consequently,
the deformation theory cannot describe phenomena associated with loading and un-
loading from the yield surface. In fact, this theory is restricted to the particular
type of stress history known as proportional loadings [11] in which the components
of the stress tensor increase in constant ratio to each other. This assumption is
not respected in many cases and particularly it makes the study of secondary paths
impossible. For this reason, and despite the previously mentioned paradox, an in-
cremental flow theory has been adopted in this thesis, and a von Mises material with
isotropic hardening has been taken. Two additional arguments against the defor-
mation theory is its lack of physical ground and the difficulty of finding numerical
examples in the literature.

1.2 General structure

To get an overview of the general structure of this thesis, the contents of the chapters
are presented in the following.
In Chapter 2, a review on plastic instabilities is presented. As mentioned before,
both physical and theoretical aspects are emphasised. Since a lot of work has been
done on it, a special section is devoted to the plastic buckling of the Euler beam.
In Chapter 3, internal force vectors and tangent stiffness matrices for three elasto-
plastic 2D beam elements are derived. An important part is devoted to the resolution
of the constitutive equations for the Timoshenko element.
In Chapter 4, a complete description of the co-rotational framework for 3D beam ele-
ments is presented. Several local formulations in elasticity, including or not warping
effects and based on Timoshenko or Bernoulli assumptions are discussed. Plasticity
is further included in the local formulation of the Timoshenko elements.
In Chapter 5, path following techniques are developed. First, the procedure used
to compute non-critical paths is presented, both at structural and element levels.
Then, the new algorithm for the direct computation of elastic critical points and the

4
two branch-switching methods in elasto-plasticity are explained in detail.
In Chapter 6, five 2D numerical examples in elasto-plasticity are studied in order to
assess the performances of the branch-switching procedures and the 2D elements.
In Chapter 7, ten 3D numerical examples are presented in order to assess the perfor-
mances of the elastic elements. Several examples are devoted to the parameterisation
of finite rotations.
In Chapter 8, eight numerical examples are studied in order to compare the conver-
gence properties of the new algorithm for the direct computation of elastic critical
points with the classical approach of Wriggers et al. [104, 105]
In Chapter 9, five 3D numerical examples in elasto-plasticity are used in order to
assess the performances of the elements and the branch-switching procedures.
In Chapter 10, conclusions and directions for future research are presented.

Correspondence with previous publications

Chapters 2 and 3 are reproduced from [5] with some minor modifications. In Chapter
4, the description of the co-rotational framework and the elastic local formulation
are taken from [6], while the plastic local formulation is taken from [9]. Minors
modifications have been introduced and particularly, the evaluations of matrices G
and Kg in Section 4.2.3 are extended. In Chapter 5, Section 5.2 is taken from [8],
while Section 5.3 is taken from [9]. The numerical examples in Chapter 6 are similar
to the ones published in [5]; However, some modifications have been introduced,
e.g. the implementation of the minimisation procedure. Chapters 7 and 8 correspond
to the sections headed “Numerical examples” in [6] and [8], respectively. Chapter
9 corresponds mainly to the section headed “Numerical examples” in [9]; the only
difference is that the two 2D examples presented in [9] are not reproduced.

5
Chapter 2

Plastic instabilities – review

Historically, plastic buckling of columns has been studied since the early works of
Considère [20] (1891) and von Karman [102] (1910). The issue under consideration
at that time was the determination of the maximal load that columns can support.
Despite considerable efforts in the following decades, this apparently simple problem
did not receive a comprehensive solution until the work of Shanley [88] (1947).
Based on experimental results, Shanley showed that the plastic bifurcation of a
perfect column occurs at the so-called tangent modulus load and is characterised
by the apparition of a zone of elastic unloading. Moreover, by using a very simple
model of the column, consisting of two rigid parts connected with two springs,
Shanley was able to give a correct qualitative description of the phenomenon of
plastic bifurcation and outlined the existence of a continuous range of bifurcation
points. The most interesting feature of Shanley’s model lies in its simplicity: with
very simple mathematics a complete solution to the problem can be obtained. An
improved model, where the two rigid parts are connected by a continuous range
of springs was introduced by Hutchinson [49] (1973). This model has been used
to study theoretically and numerically imperfection sensitivity and secondary post-
bifurcation paths in plastic buckling problems.
Based on the study of Shanley’s model, one very important difference between elas-
tic and plastic bifurcations becomes apparent: in elasticity, bifurcation occurs at
isolated critical points and is characterised by a loss of stability on the fundamen-
tal path; in plasticity, a continuous range of stable bifurcation points along the
fundamental path may occur.
Despite its significant phenomenological insights, the work of Shanley remains lim-
ited in scope. Its conclusions are essentially restricted to the plastic buckling of
compressed columns. The first theoretical general approach to bifurcation and sta-
bility in elasto-plastic solids was given by Hill [45] (1959). By taking into account
the change in geometry during the deformation process, Hill derived criteria for the
uniqueness and stability of a solution and introduced the notion of linear comparison
solid. However, several problems remained. On the theoretical side, one important
question is how to interpret the stability of the points along the fundamental path
beyond the first bifurcation point since these points have no apparent physical mean-

7
ing. On the numerical side the notion of linear comparison solid does not lend itself
to a direct implementation which would result in a reliable numerical procedure to
calculate the lowest bifurcation point.
Within the framework of an energetic approach, Petryk [76–81] proposed a solution
to these two problems by introducing the notions of tangent comparison solid and
stability of a deformation path. His results, obtained by assuming a discretisation
of the structure and the symmetry of the constitutive moduli give a complete theo-
retical description of the phenomenon and can be considered as a generalisation of
the conclusions obtained with Shanley’s model. Furthermore, they are the basis for
the numerical branch-switching procedures used in this thesis.
It should be noted here that an energetic approach to Hill’s criteria was also given
by Nguyen [62, 63] in 1987. This theory, which has also applications in fracture and
friction mechanics, will not be investigated in this thesis.
Following this brief introduction, the remainder of this chapter is organised in four
parts. The first part is a review of the work done on the simple models introduced
by Shanley and Hutchinson. The second part presents the criterion of uniqueness
and the notion of linear comparison solid introduced by Hill. As an application,
the theoretical analysis of the plastic buckling of the Euler beam is presented in
the third part. In this context, emphasis is given to the study of the secondary
post-bifurcation path. Finally, based on the work of Petryk, the last part presents a
complete theoretical approach to instability problems in time-independent plasticity
in the context of discrete structures.

2.1 Simple models

Two simple models are presented and analysed in this section. The first one origi-
nates from Shanley. Its study is interesting for several purposes. From a pedagogical
point of view, Shanley’s column is the simplest structure presenting plastic bifurca-
tion and can be studied with very easy mathematics. From a historical point of view,
the work done by Shanley in 1947 gives the first correct description of the problem.
The second model has been introduced by Hutchinson in 1973 and has been inves-
tigated later by several authors. Its study generates qualitative conclusions about
the effects of both geometrical and material non-linearities on post-bifurcation be-
haviour.

2.1.1 Shanley’s column

In his famous article, Shanley [88] studied a model (cf. Figure 2.1) consisting of a
rigid ⊥ frame loaded by a vertical downward force P . The frame is maintained in
equilibrium by two springs k1 and k2 . The stiffness of the springs is either E or Et ,
depending on whether plastic deformation occurs. The system has two degrees of
freedom, z and θ. A linearised study is performed.

8
 P

θ
H

F1
L F2 z
L
k1
k2

Figure 2.1: Shanley’s column.

Equilibrium equations
The equilibrium equations are

P = F1 + F2 (2.1a)
P H θ = (F2 − F1 ) L (2.1b)

where F1 and F2 are the forces in the springs given by

F1 = k1 (z − L θ) (2.2a)
F2 = k2 (z + L θ) (2.2b)

which finally gives

P = (k1 + k2 ) z + (k2 − k1 ) L θ (2.3a)


(k1 + k2 ) H z − L2 θ = (k2 − k1 ) L z (2.3b)

By differentiation the following rate equations are obtained

Ṗ = (k1 + k2 ) ż + (k2 − k1 ) L θ̇ (2.4a)


 
(k1 + k2 ) H z − L2 θ̇ + H θ ż = (k2 − k1 ) L ż (2.4b)

Fundamental path
The fundamental path is defined by

θ = 0 k1 = k2 = E P = 2 E z in the elastic range

(2.5)
θ = 0 k1 = k2 = Et P = 2 Et z in the plastic range

Bifurcation paths
The possibility of bifurcation from the fundamental path with θ̇ > 0 is now in-
vestigated. Since a linearised analysis is performed, only the initial tangent of the
post-bifurcation paths is possible to evaluate. Consequently, the angle θ is set to 0.

9
Two trivial cases can easily be found if k1 = k2 . Equation (2.4b) is then reduced to


H z − L2 θ̇ = 0 (2.6)
and bifurcation is possible if
L2
z= (2.7)
H
which by taking account of (2.3a) gives as bifurcation loads
2 L2
Pe = E if k1 = k2 = E
H (2.8)
2 L2
Pt = Et if k1 = k2 = Et
H
where Pe is the elastic buckling load and Pt is the tangent modulus load.
Other solutions are searched by assuming
k1 = E k2 = Et (2.9)
which implies
Ḟ1 < 0 Ḟ2 > 0 (2.10)
and therefore, from (2.2a) and (2.2b)
−L θ̇ < ż < L θ̇ (2.11)
Equation (2.4b) can be rewritten as
 
E + Et H
ż = 1 − 2 z L θ̇ (2.12)
E − Et L
Introducing equation (2.12) in (2.4a), and taking (2.3a) into account gives
H E + Et
Ṗ = (Pr − P ) θ̇ (2.13)
L E − Et
where
2 L2 2 E Et
Pr = Er Er = (2.14)
H E + Et
Pr is the reduced modulus load and Er is the reduced modulus.
By using (2.12) and (2.3a) the conditions (2.11) provide after some work
Pt < P < Pe (2.15)

Conclusions
From the equations (2.13) and (2.15), it can be concluded that there is a continuous
range of bifurcation loads between Pt and Pe . Moreover, the initial tangent of the
post-bifurcation path is positive at Pt , negative at Pe , and range monotonically
between Pt and Pe with a horizontal tangent at Pr . These results are represented in
Figure 2.2.

10
 P

Pe

Pr

Pt

Figure 2.2: Linearised analysis of Shanley’s column.

Three remarks concerning the study of this model should be mentioned [87]:

• The infinite number of bifurcations does not come from the non-linearity of the
stress–strain relation but from the irreversible character of the process: a strain
configuration can correspond to an infinite number of stress configurations,
corresponding to different histories of the loading.

• The stability aspect will be studied in details in the following sections. How-
ever, it can be inferred intuitively that the fundamental path will be stable in
a dynamical sense up to Pr (upward tangent) and unstable after (downward
tangent).

• The conditions (2.10) assume that both springs are plasticised (k1 = k2 = Et )
before the load reaches Pt . If PY is the yield limit of the springs and if Pt < PY ,
then the bifurcations between Pt and PY do not exist.

2.1.2 Hutchinson’s model

The model introduced by Hutchinson [48–50] and shown in Figure 2.3 differs from
the previous one in that the ⊥ frame is supported by a continuous distribution of
springs. Imperfections are represented by an initial rotation θ̄ from the vertical in the
unloaded state. Linearised analysis is performed once again and geometrical non-
linearities are artificially introduced through a horizontal non-linear spring which
develops a force F (θ) = k L2 θ2 (k > 0).
This model has been studied with two different elasto-plastic laws as shown in
Figure 2.4. The first law is the classical bilinear one while the second is of Ramberg-
Osgood type defined by the equation
 n
ε σ σ
= +α (2.16)
εY σY σY

11
 P
F (θ) = k L2 θ2

H θ + θ̄

L L
u
x

Figure 2.3: Hutchinson’s continuous model of Shanley’s column.

σ σ

σY Et
E

ε ε

Figure 2.4: Bilinear and Ramberg-Osgood (α = 0.2, n = 3) elasto-plastic laws.

Equilibrium equations
The linearised equilibrium equations are
 L
P = σ dx (2.17)
−L



 L
P H θ + θ̄ + F H = σx dx (2.18)
−L

with

ε = u + xθ (2.19)

and
σ̇ = Et ε̇ for plastic loading
(2.20)
σ̇ = E ε̇ for elastic unloading or within the elastic range

A review of some results found in literature is presented below.

12
Continuous range of bifurcation points
The purpose of the study performed by Cimetière, Elkoulani and Léger [18] was to
prove mathematically the results obtained by Shanley on the basis of the two springs
3 3
model, i.e. every point within the interval [Pt , Pe ] where Pt = 2L
3H t
E and Pe = 2L
3H
E
is a bifurcation point while no bifurcation points are present outside this interval.
In this study, a constant value of Et was adopted and geometrical non-linearities
were neglected (k = 0). The results, adapted from [18], are shown in Figure 2.5.

Pe

Pr

Pt

Figure 2.5: Bifurcation diagram for E/Et = 5.

Imperfection analysis
Hutchinson [48, 50] calculated numerically the post-buckling path in the case of a
Ramberg-Osgood model with α = 0.2 and n = 3. The results (adapted from [48])
with and without imperfections are shown in Figure 2.6. The imperfect paths are
obtained for θ̄ = 0.001. The imperfect structure gives equilibrium paths which
tend to the path branching at Pt when θ̄ tends to 0. This suggests that a perfect
model will bifurcate at Pt , and therefore, only the path branching at Pt needs to be
investigated. The other post-bifurcation paths do not have any physical meaning.
Su and Lu [94] studied the case with constant Et and k > 0 and found similar
results.
Asymptotic expressions
The curves shown in Figure 2.6 present a maximum point for relatively small values
of θ. Hutchinson calculated asymptotic expressions for the stable portion of the
post-bifurcation path and then deduced expressions for the maximum load. For the
structure without imperfection, he obtained expansions in the form

P = Pt + λ1 θ + λ2 θ3/2 + λ3 θ2 + ... (2.21)

The difference from similar expressions obtained using Koiter’s theory for elastic
structures (cf. e.g. [96]) is the presence of a term λ2 θ3/2 , which comes from the
apparition of a growing elastic unloading zone at bifurcation. However, it has to
be emphasised that such expressions are only valid for sufficiently small values of θ.

13
 P / Pt

1.1

1.05

1
k =0
0.95
k / E =1

0 0.1 0.2 0.3

θ

Figure 2.6: Post-buckling behaviour and imperfection sensitivity in the case of a

Ramberg-Osgood relation (H/L = 1).

Hutchinson compared these asymptotic expansions with numerical calculations and

found that good accuracy is preserved until the maximal load only in the case of
strong geometrical non-linearity (k/E = 1). Van der Heijden [99] proposed a differ-
ent method to calculate more accurate asymptotic expansions, but strong divergence
from numerical simulations were still observed for cases without strong geometrical
non-linearity.
Combination of geometrical and material non-linearities
An analysis without linearisation (and thus without a horizontal spring) was pro-
posed by Cimetière and Léger [19]. A better description of the model is needed and
the two cases shown in Figure 2.7 were studied. In case (a) the springs are tied at
their bottom and free to move along the x-axis at their top. In case (b) the contrary
applies.

(a) (b)

Figure 2.7: Models with non-linear geometry.

In case (a), equations (2.19) and (2.18) are replaced by

ε = u + x tan θ (2.22)

14
  L
σ
P H sin θ = x dx (2.23)
−L cos2 θ

while in case (b) they are replaced by

ε = u + x sin θ (2.24)

 L
P H sin θ = x σ cos θ dx (2.25)
−L

The models have been tested with a bilinear elastic-plastic law, without geometrical
imperfections. The numerical results for case (a) show a monotonically increasing
post-bifurcation curve. In case (b), the bifurcated branch can be either monotoni-
cally strictly increasing or it can present a maximum, depending on the ratio Et /E.
The conclusions presented in [19] are now summarised.
The cumulative effects of geometrical and material non-linearities are difficult to
study without numerical simulations, e.g. different ratios Et /E in the bilinear elastic-
plastic law or different exponents n in the Ramberg-Osgood relation can lead to
different qualitative solutions (cf. also [61]).
With respect to the effects of material non-linearities, it can be concluded by com-
paring Figures 2.5 and 2.6 that the decreasing stiffness of the material in the plastic
range (Ramberg-Osgood law) has destabilising effects.
The geometrical non-linearity can have both stabilising or destabilising effects. The
differences in the post-buckling paths of the cases investigated can be partly under-
stood by considering only geometrical non-linearity. Four models have been studied:

• model 1 : linearised equations without geometrical imperfections

• model 2 : linearised equations with geometrical imperfections

• model 3 : non-linearised equations, case (a)

• model 4 : non-linearised equations, case (b)

The elastic post-buckling paths for these models are shown in Figure 2.8. It can
be concluded that an unstable elastic post-critical curve accentuates the possibility
of occurrence of a maximum point and a stable one accentuates the possibility of a
monotonic path.
Moreover, in the vicinity of the bifurcation, material non-linearity prevails over the
geometric one and a stable path is always obtained. More detailed conclusions
cannot be derived.

15
 P/Pc

model 3

2 .0

1.0 model 1
model 4
model 2
θ
0 .2 0 .6

Figure 2.8: Elastic post-buckling of the studied models.

2.2 Hill’s criterion of uniqueness

This section presents the theoretical results, due to Hill [45], concerning uniqueness
of the solution of an elastic-plastic deformation.
A general solid body subjected to a quasi-static loading defined by the parameter λ is
considered. The general boundary-value problem in elastic-plastic deformation can
be defined as follow: at a generic stage in the loading process, the current shape of
the body and the internal stress distribution are supposed to have been determined
already, together with the existing state of hardening and mechanical properties in
general. The incremental changes in all these variables have now to be calculated
for a further infinitesimal variation λ̇ of the loading parameter. The question of the
uniqueness of the solution is then set.
The differentiation of the principle of virtual work can be written in the form
  
Ṅij δuj,i dv = ḃj δuj dv + Ṫj δuj ds (2.26)
v v s

where Tj are the surface tractions applied on the surface s and bj are the body forces.
Nij are the nominal stresses, i.e. the stresses acting in the current configuration on
an infinitesimal surface element in the reference configuration. δuj is a kinematically
admissible virtual displacement.
Two different incremental solutions, u̇j and u̇∗j , are considered. By taking as virtual
displacement

δuj = u̇j − u̇∗j = ∆uj (2.27)

the following equation is obtained

  
Ṅij (∆uj ) ,i dv = ḃj ∆uj dv + Ṫj ∆uj ds (2.28)
v v s

16
Equation (2.28) must hold also if u̇j and u̇∗j are interchanged, that is, if Ṅij , ḃj and
Ṫj are replaced by Ṅij∗ , ḃ∗j , and Ṫj∗ . Subtracting these two equalities and introducing
the notations

∆Ṅij = Ṅij − Ṅij∗ ∆ḃj = ḃj − ḃ∗j ∆Ṫj = Ṫj − Ṫj∗ (2.29)

gives
  
∆Ṅij (∆uj ) ,i dv = ∆ḃj ∆uj dv + ∆Ṫj ∆uj ds (2.30)
v v s

The loading is assumed conservative, which means that Tj and bj are only depending
on λ, which further implies Ṫj = Ṫj∗ and ḃj = ḃ∗j . Equation (2.30) is then reduced to

H = ∆Ṅij (∆uj ) ,i dv = 0 (2.31)
v

Hence, according to Hill, uniqueness of the solution is ensured if


H = ∆Ṅij (∆uj ) ,i dv > 0 (2.32)
v

for every kinematically admissible ∆uj .

2.2.1 Linear comparison solid

The criterion defined in equation (2.32) is difficult to apply in practical problems.

By introducing the notion of linear comparison solid, Hill proposed a second cri-
terion which is easier to handle. A fictitious solid having the same configuration
and stresses under the current loading as the real one, but a different incremental
constitutive law is considered. Namely, at each material point, the fictitious solid is
assumed as incrementally linear with the constitutive relationship
L
Ṅij = Cijkl ul,k (2.33)
L
By using the notations in (2.27) and (2.29), the modulus Cijkl is chosen such that
L
∆Ṅij (∆uj ) ,i ≥ Cijkl (∆uj ) ,i (∆ul ) ,k (2.34)

for every pair of incremental solutions u̇j and u̇∗j .

The following functional F is then defined

L
F = Cijkl (∆uj ) ,i (∆ul ) ,k dv (2.35)
v

From (2.31) and (2.34), it can be concluded that

H≥F (2.36)

17
and by using (2.32), the criterion of uniqueness can be expressed as

L
F = Cijkl (∆uj ) ,i (∆ul ) ,k dv > 0 (2.37)
v

for every kinematically admissible ∆uj .

The problem related to equation (2.37) lies in choosing the optimal linear comparison
solid so that the difference between F and H is as small as possible. An example of
such a choice is given in Section 2.3.3.

2.3 Euler beam analysis

The purpose of this section is to present a review of the different works done on the
plastic buckling of a pin-ended column based on the Bernoulli plane beam assump-
tions. After presenting the elastic case and the two classical theories of the tangent
and reduced modulus load, a theoretical analysis is performed, starting from Hill’s
criterion of uniqueness. A post-bifurcation analysis is also presented.
The column, see Figure 2.9, has a circular cross-section. It is assumed thick enough
so that elastic buckling is prohibited, but slender enough so that beam theory can
still be applied. The model adopted is based on a shallow arch strain formulation
which, under Bernoulli hypothesis, is defined by
1
ε = u + (v  )2 − y v  (2.38)
2
where a prime denotes differentiation with respect to x.
A bilinear elastic-plastic law is considered.

P R

uL

v( x ) L

x
x

y y

Figure 2.9: Simply supported column with circular cross-section.

18
2.3.1 Elastic case

The elastic buckling analysis is performed by using the principle of virtual work

σ δε dv = −P δuL (2.39)
v

If the beam is inextensible, equation (2.38) gives

1
u + (v  )2 = 0 (2.40)
2
and therefore
 

δuL = δu = − v  δv  dx (2.41)
L L

Equations (2.38), (2.40) and (2.41) are introduced in (2.39). After some work, the
second order terms give the classical Euler equation

E I v  + P v = 0 (2.42)

The solution is given by

π2 E I π 3 E R4
Pe = = (2.43)
L2 4 L2
The eigenmode associated with Pe is
π x
v = R sin (2.44)
L
and the related strains are
π2 π x
ε = 2 R y sin (2.45)
L L
Equation (2.42) can also be derived by using a linear strain theory and by writing
the equilibrium equations in the deformed configuration. This analysis, based on
a shallow arch strain definition, is usually called linearised buckling or Euler buck-
ling. It gives the right critical load, but the post-bifurcation path obtained is the
horizontal line P = Pe , while the exact one is an upward parabola [29, 72].

2.3.2 Reduced versus tangent modulus load

These two theories based on linearised buckling assumption were in concurrence

until Shanley showed that a perfect column begins to bend at the tangent modulus
load Pt . Shanley’s results are based on experiments and on the assumption that a
column behaves qualitatively in the same way as the discrete model he introduced.
The theoretical rigorous proof requires Hill’s criterion of uniqueness and will be
described in the next section.

19
Apart from the historical aspect, the interest in presenting these two theories lies in
the physical description of the buckling phenomenon. Moreover, although the theory
of the reduced modulus load Pr is not correct, this load corresponds to the limit
of stability along the fundamental path and has therefore a theoretical importance
(cf. Section 2.4.3)
In order to simplify the calculations, a column with a rectangular cross-section
(A = b h) is considered.
Reduced modulus load
It is assumed that the column remains straight while the axial load is increased
beyond the yield point, after which the column bends, or tries to bend, at a con-
stant compressive force. The problem is then the same as for Euler’s theory, i.e. to
determine deformed configurations which are in equilibrium. The calculations based

;;;;;
on Figure 2.10 are summarised below.

Pr Pr
∆σ1
∆σ2
stresses

;;;;; h

before buckling
loading

∆ε
unloading

during buckling
z

strains

Figure 2.10: Reduced modulus load assumptions.

∆σ1 = Et ∆ε in the loading area

(2.46)
∆σ2 = E ∆ε in the unloading area

Buckling occurs at a constant axial load, which implies


∆σ = 0 (2.47)
A

Equation (2.47) gives the position of the z axis as function of the moduli E and Et .
The bending moment is then calculated according to

M= z ∆σ (2.48)
A

20
which gives

M = Er I/ρ (2.49)

with

 −2
1 −1/2 −1/2
Er = E + Et (2.50)
2

Er is called the reduced modulus and depends on the shape of the cross-section.
By introducing the classical equations

M = −P y 1/ρ = y  (2.51)

the following differential equation is obtained

Er I y  + P y = 0 (2.52)

By similarity with Euler equation, the solution is given by

π2
Pcr = Pr = Er I (2.53)
L2
This result is valid under the assumption that the column is plasticised at Pr , i.e. if
Pr > A σY (σY is yield limit defined in Figure 2.4).
Tangent modulus load
Actually, the column is free to bend at any time. There is nothing to prevent it
from bending simultaneously with increasing axial load. The tangent modulus load
theory assumes that the column remains straight until the critical load Pt is reached
and that an infinitesimal lateral deflection occurs when applying an infinitesimal
increment load ∆P in such a way that the tensile strain caused by the deflection
is compensated by the axial shortening due to ∆P . Then there is no unloading
point in the cross-section (cf. Figure 2.11) and ∆σ = Et ∆ε still applies everywhere.
The analysis is therefore the same as in the elastic case by replacing E by Et . The
solution (Euler) is

π2
Pt + ∆P = Et I (2.54)
L2
If ∆P is assumed infinitesimal, the critical load is obtained as

π2
Pcr = Pt = Et I (2.55)
L2
This result assumes also that the column is plasticised at Pt , i.e. Pt > A σY .
Remark
This approach presents a paradox. There is no elastic unloading and Et applies
everywhere. Therefore, according to Euler’s theory of buckling, the load cannot

21
 ;;;;;
Pt Pt + ∆P
∆σ

;;;;;
stresses

∆ε
strains

before buckling during buckling

Figure 2.11: Tangent modulus load assumptions.

exceed Pt , which is inconsistent with the application of ∆P . This paradox was

explained first by Shanley [88]. By measuring the strain distribution in a column
test, he showed that bending in a perfect column begins at the tangent modulus
load Pt with one line of non-loading (in case of a column with rectangular cross-
section). After that, a further increase of the load is supported by the apparition of
a zone of elastic unloading. This zone expands when the load increases as shown in
Figure 2.12.

line of non loading at Pt

zone of elastic unloading

Figure 2.12: Zone of elastic unloading in a column with rectangular cross-section.

2.3.3 Theoretical plastic analysis

This analysis can be seen as an application of the concepts presented in Section

2.2. In addition, based on the work done by Hutchinson [49, 50], a study of the
post-bifurcation path is performed.

22
The beam is assumed uniformly compressed beyond the plastic limit (P > σY π R2 ).
The problem is to determine whether it is possible to obtain a bifurcation when
an infinitesimal load increment ∆P is applied. The incremental changes of the
fundamental solution are denoted by ∆σf and ∆εf while those of the bifurcated
solution are denoted by ∆σb and ∆εb with ε defined according to equation (2.38).
The difference between these two solutions is denoted by

∆ε = ∆εb − ∆εf ∆σ = ∆σb − ∆σf (2.56)

The function H, defined in (2.31), can be rewritten as


H = ∆σ ∆ε dv (2.57)
v

At P the whole beam is plasticised since the fundamental path is a pure axial
L
compression. The linear comparison operator Cijkl defined in Section 2.2 is taken as
L
Cijkl = Et

Hence, the function F (2.35) can be rewritten as


F = Et ∆ε2 dv (2.58)
v

For this particular problem, the relation F ≤ H (2.36) can easily be proved as follow

H − F = T dv (2.59)
v

with

T = (∆σ − Et ∆ε) ∆ε (2.60)

= [(∆σb − ∆σf ) − Et (∆εb − ∆εf )] (∆εb − ∆εf )

Depending on the sign of ∆εb and ∆εf , T can take the following values

• ∆εb ≤ 0 ∆εf ≤ 0 → ∆σb = Et ∆εb ∆σf = Et ∆εf

→ T =0

• ∆εb ≤ 0 ∆εf ≥ 0 → ∆σb = Et ∆εb ∆σf = E ∆εf

→ T = (Et − E) ∆εf (∆εb − ∆εf ) ≥ 0 (Et < E)

• ∆εb ≥ 0 ∆εf ≥ 0 → ∆σb = E ∆εb ∆σf = E ∆εf

2
→ T = (E − Et ) (∆εb − ∆εf ) ≥ 0

• ∆εb ≥ 0 ∆εf ≤ 0 → ∆σb = E ∆εb ∆σf = Et ∆εf

→ T = (E − Et ) (∆εb − ∆εf ) ∆εb ≥ 0

23
This shows that T ≥ 0 always applies and therefore H − F ≥ 0.
Bifurcation load
Let Pc be the lowest value for which the condition F = 0 is satisfied. The solution
of the variational principle δF = 0 is given in Section 2.3.1. The only modification
required is the replacement of E by Et . Hence, the critical load is

π 2 Et I π 3 Et R4
Pc = = = Pt (2.61)
L2 4 L2
and the associated eigenmode is defined by
(1)
π x (1) π2 π x
v = R sin ε= R y sin (2.62)
L L2 L
Equation (2.61) confirms that the lowest bifurcation occurs at the tangent modulus
load Pt as it was originally shown by Shanley.
Initial tangent
A bifurcated solution is searched as a linear combination of the fundamental path
and the eigenmode of the elastic comparison solid
(1)
∆εb − ∆εf = ξ ε (2.63)

with ξ denoting the amplitude of the eigenmode defined by (2.62) which is taken as
the independent variable in the post-buckling expansion. With (2.63) F = 0, but
not H, unless both ∆εb and ∆εf have the property that no elastic unloading occurs
at any point in the column. The load ratio λ and its incremental variation ∆λ are
introduced as
P
λ= (2.64)
Pc

∆λ = λ − λc = λ1 ξ (λc = 1) (2.65)

If λ1 has to represent the initial tangent to the bifurcated path, a differential form
of equations (2.63) and (2.65) can be obtained by the following transformations
∆εb ∆εf (1) ∆λ
− =ε = λ1 (2.66)
ξ ξ ξ
When ξ → 0, equations (2.66) give
◦ ◦ (1)  (1)
εb = εf + ε = λ1 εf + ε (2.67)

with

◦ ∂ ()  ∂ () 
() = () = (2.68)
∂ξ ∂λ λc

24
For the column
σf −λ Pc  −π 2 R2
εf = = → εf = (2.69)
Et π R2 Et 4 L2
Equations (2.67) and (2.69) give

 π x 
◦ π 2 R2 λ1 y
εb = − + sin (2.70)
L2 4 R L
By taking λ1 large enough it is obviously possible to obtain
◦
εb < 0 (2.71)

everywhere in the column so that no elastic unloading occurs, which further implies
H = 0. The problem is to determine which value(s) of λ1 can be solution to the
boundary-value problem. For this purpose, the slope for the elastic comparison solid
at bifurcation is denoted by λhe he he
1 . Euler analysis gives λ1 = 0. Then λ1 is such that

λhe y π x
− 1
+ sin >0 (2.72)
4 R L
in some part of the column and therefore1

λhe
1 < λ1 (2.73)

It can then be inferred that the initial slope of the elastic-plastic solid λ1 must be the
smallest value consistent with (2.71). The reason for this is that if λ1 were larger,
◦
then by continuity there would be some range of positive ξ where εb would be lower
than zero everywhere. Consequently, the behaviour of the elastic-plastic solid would
initially coincide with that of the comparison solid so that λhe 1 = λ1 . However, this
possibility is contradicted by (2.73) which implies that λ1 is the smallest value such
that

 π x 
◦ π 2 R2 λ1 y
∀x ∈ [0; L] and y ∈ [−R; R] εb = − + sin ≤0 (2.74)
L2 4 R L
The solution is then

λ1 = 4 (2.75)
◦
and one point of non-loading (εb = 0) is obtained at x = L/2 and y = R. This
argumentation proves that a zone of elastic unloading spreads from this point after
the bifurcation.
Post-bifurcation path
By performing a perturbation expansion of λ about the bifurcation point, Hutchin-
son calculated an expression for the post-bifurcation path under the form

λ = 1 + 4 ξ + λ2 ξ 1+β (2.76)
1
The inequality (2.73) is verified in most of the common elastic-plastic problems.

25
This approach is similar to that developed in Section 2.1.2 for the ⊥ model. Here,
λ2 and β are determined by an approximation of the lowest order non-vanishing
terms in the equation of virtual work (2.39), which gives
 1/3
Et 1
λ2 = −6.337 β= (2.77)
E − Et 3
The analysis which leads to the above expression is rather lengthy and will not be
described here. It can be noted that Hutchinson showed that the second term of
the perturbation expansion comes essentially from the apparition and expansion of
the zone of elastic unloading. The same result was also obtained using a different
method by Leger and Potier-Ferry [58]. Using the expression (2.76), an estimation
of the maximal load can be calculated as
E − Et E − Et
λmax = 1 + 0.106 ξmax = 0.106 (2.78)
Et Et
In the case of a rectangular cross-section, the same analysis leads to
 2/5
Et
λ = 1 + 3 ξ + λ2 ξ 7/5
λ2 = −5.003 (2.79)
E − Et

Remark 1: Contrary to the elastic case, the post-buckling path in the plastic range
depends on the geometry of the cross-section. It comes from the growth of the elastic
unloading zone which depends on the shape of the cross-section (cf. Figure 2.13).

section A-A
A A

Figure 2.13: Zone of elastic unloading for circular and rectangular cross-sections.

According to the author’s opinion, the two following remarks about expressions (2.76)
and (2.79) can be stated.
Remark 2: The strain assumption (2.38) is to crude to represent correctly the
elastic post-bifurcation. This study is therefore based on the assumption that in
the vicinity of the post-bifurcation path material non-linearity prevails over the
geometrical one. However it is not known how far this assumption is valid and
especially if it is still valid when the maximal load is reached. The results presented

26
in Section 2.1.2 suggest that geometrical non-linearity influences the post-bifurcation
path in the plastic range and therefore the value of the maximal load.
Remark 3: The expressions (2.76) and (2.79) are only valid for small values of ξ, but
it is not known how large ξ can be. They may diverge from the correct path before
the occurrence of the maximal load and therefore the estimations of the maximal
load may not be correct. In the case of a rectangular cross-section, Cheng [16], using
a different approach, calculated the following term in the expansion
 4/5
7/5 9/5 Et
λ = 1 + 3 ξ + λ2 ξ + λ3 ξ λ3 = 2.207 (2.80)
E − Et

The expressions (2.79) and (2.80) are plotted in Figure 2.14. Here again, it is not
known how far the expression (2.80) is valid. It is not certain that the addition of
another term in the expansion will increase the accuracy of the maximal load. The
unusual high values of λmax (≈ 2) and ξmax (≈ 1.8) suggest that (2.80) diverges from
the correct path before the occurrence of the maximal load.
λ
2

including 9/5 term

1.5

not including 9/5 term

0.5 ξ
0 0.5 1 1.5 2 2.5

Figure 2.14: Post-critical paths according to (2.79) and (2.80) with Et /E = 0.2

2.4 Discretised systems

In the light of the application presented in the last section two fundamental ques-
tions remain. The first one concerns the choice of the constitutive moduli in the
elastic comparison solid. Until now, no method has been proposed to find a com-
parison solid which does not overestimate the range of non-uniqueness and gives by
eigenvalue analysis the lowest bifurcation point. For the Euler beam it has only
been shown that taking Et as elastic comparison modulus everywhere in the beam,
works. This choice is rather intuitive since the fundamental solution is a uniform
plastic compression. However, in a finite element context this choice is not obvi-

27
ous, especially in cases where the fundamental solution is more complicated than a
uniformly compressed state.
The second question concerns the existence of an infinite number of secondary post-
critical paths. In the last section, only the secondary path branching at Pt has
been studied. The essential question in a numerical context is then whether it
is worthwhile to determinate the other secondary paths. An additional question
refers to the nature of the fundamental path beyond Pt . On one hand imperfection
analyses (cf. e.g. Section 2.1.2) suggest that only the first secondary path has a
physical meaning. However, on the other hand, the points along the fundamental
path are stable up to Pr .
The answers to these questions have been given by Petryk [76–81]. This work, which
is summarised in this section, gives a complete theoretical description of the phe-
nomenon of bifurcation and instability in time-independent plasticity. The notions
of uniqueness and stability of the solutions to the boundary-value problem in elastic-
plastic deformation are discussed and an energy interpretation is given. Moreover,
this work gives the theoretical basis for the two numerical branch switching proce-
dures presented in Section 5.3.
The theory developed by Petryk is based on the assumptions that the problem has
been spatially discretised and that the constitutive moduli are symmetric.

2.4.1 Constitutive framework and discretised rate problem

A time-independent elastic-plastic material is considered. The constitutive equation

between the rate of stresses and strains is expressed in the form

∂ Ṡ
Ṡ = C(Ḟ, H) · Ḟ C= (2.81)
∂ Ḟ
where S is the first Piola-Kirchoff stress tensor (i.e. the transpose of the nominal
stress tensor) and F is the deformation gradient. The symbol H represents the
influence of the deformation history.
It is assumed that relation (2.81) admits a potential U (Ḟ, H) so that

∂U ∂ 2U
Ṡ = C= (2.82)
∂ Ḟ ∂ Ḟ∂ Ḟ
which is equivalent to imposing that the instantaneous modulus C is symmetric.
The problem is supposed spatially discretised. The velocity fields v are then re-
stricted to having the classical form

v (ξ) = φα (ξ) vα α = 1, ..., N (2.83)

where ξ are the position vectors in the reference configuration, φα the shape func-
tions and vα the velocities of the nodal degrees of freedom. In the same way, the

28
displacements u and the velocity variations w are given by

u (ξ) = φα (ξ) uα α = 1, ..., N

(2.84)
w (ξ) = φα (ξ) wα α = 1, ..., N

The numeration of the shape functions is chosen such that the boundary conditions
give prescribed values vα = v̄α and wα = 0 for α = M + 1, ..., N .
By assuming a conservative loading, components of the prescribed load vector are
given, in the rate form, by
 
Ṗα = ḃ φα dv + Ṫ φα ds α = 1, ..., M (2.85)
v s

and the rates of the internal forces are given by


Q̇α (ṽ) = Ṡ (∇v) · ∇φα dv α = 1, ..., N (2.86)
v

where a tilde over a symbol denotes a spatial field defined over the body volume in
the reference configuration.
The equilibrium equations of the first order problem in velocities are

Q̇α (ṽ) = Ṗα α = 1, ..., M (2.87)

The components of the tangent stiffness matrix are defined by


∂ Q̇α (ṽ)
Kαβ (ṽ) = = ∇φα · C (∇v) · ∇φβ dv (2.88)
∂vβ v

which allows the reformulation of the equilibrium equations (2.87) as

Kαβ (ṽ) vβ = Ṗα α = 1, ..., M (2.89)

From (2.82) and (2.88), it follows that K is symmetric, i.e. Kαβ = Kβα .
Since the existence of a potential U is assumed, a functional J can be defined as

M M
1
J (ṽ) = U (∇v) dv − Ṗα vα = Kαβ (ṽ) vβ vα − Ṗα vα (2.90)
v α=1
2 α=1

The equations (2.87) or (2.89) can then be given the variational formulation

∂J (ṽ)
=0 α = 1, ..., M (2.91)
∂vα
From (2.90), the tangent stiffness matrix can also be expressed as

∂ 2 J (ṽ)
Kαβ (ṽ) = (2.92)
∂vα ∂vβ

29
2.4.2 Tangent comparison solid

The theory developed by Hill (cf. Section 2.2) does not require any solution to be
known in advance. However, in practice a fundamental solution is known and the op-
timal choice for the linear comparison solid can be determined from the fundamental
solution, as explained in the following.
e
The elastic operator is Cijkl . Where the yield condition is satisfied, the elasto-plastic
e p
operator is Cijkl in case of elastic unloading, and Cijkl in case of plastic loading. The
L
linear comparison operator Cijkl is defined by
L p
Cijkl = Cijkl (2.93)

at every point in the body where the yield condition is currently satisfied, indepen-
dent of the case (plastic loading or elastic unloading) concerned, and by
L e
Cijkl = Cijkl (2.94)

at every point where the stresses lie within the yield surface. With this choice, it
can be proved that the condition (2.34) is respected for elastic-plastic material with
smooth yield surface.
In other words, the actual tangent moduli of the fundamental solution is taken as
linear comparison solid. For this reason Petryk introduced the notion of tangent
comparison solid.
L
It can be noted that in the analysis of Euler beam in Section 2.3.3, Cijkl = Et has
been assumed, which actually corresponds to the tangent comparison solid.

2.4.3 Bifurcation and stability

Directional stability of an equilibrium state

The notion of stability is often referred to the definition given by Liapunov. Roughly
speaking, an equilibrium state is stable if after applying a small perturbation at
constant loading, the damped structure returns to the previous equilibrium state.
For elastic-plastic structures, this cannot be expected in general. Moreover, even
if the equilibrium state after perturbation is very close to the original one, there is
no guarantee that the continuation of the loading process will not give divergence
between the two paths. To avoid this problem, Petryk proposed a different definition
of stability of an equilibrium state: “an equilibrium configuration is said to be
stable if the distance from that configuration in any dynamic motion caused by a
disturbance can be made as small as we please if a measure of the disturbance itself
is sufficiently small.” He introduced a restriction to this definition by considering
that the perturbed motion from the equilibrium state is such that variations of the
direction of the velocity field is negligible. In that restrictive sense, he introduced
the notion of directional stability.

30
Under these assumptions, it can be proved that an equilibrium state is directionally
stable if
Q̇α (w̃) wα > 0 for every w̃
= 0 (2.95)

Uniqueness of the solution

It is supposed that a fundamental solution ṽf to the equation (2.87) is known. The
respective tangent stiffness matrix is denoted Kf . Petryk has shown that if the
condition (2.95) is respected, uniqueness of the solution is ensured if ṽf assigns to
J a strict and absolute minimum value, i.e. if
J (ṽf ) < J (ṽ) for every admissible ṽ (2.96)
By using equations (2.91) and (2.92) and the condition (2.96), it can be concluded
that uniqueness of the solution is ensured as long as Kf is positive definite. This
result can also be obtained from Hill’s criterion of uniqueness. The definition of
the tangent comparison solid allows to use results for elastic structures. Therefore
equation (2.37) ensures uniqueness of the solution as long as Kf is positive definite.
This criterion holds for elastic-plastic materials obeying the normality flow rule
relative to a smooth yield surface. This is the case for the von Mises material
which will be used in the numerical applications. Petryk proposed a less restrictive
condition
Ṡf · Ḟ − Ṡ · Ḟf ≥ 0 for every admissible Ḟ (2.97)

where Ṡ = C Ḟ. For the applications in this thesis, this condition does not need to
be further investigated.
Continuous range of bifurcation
The question investigated here is to determine what happens beyond the first bifur-
cation point when the tangent stiffness matrix Kf ceases to be positive definite. It
is possible that on the secondary post-bifurcation path the tangent stiffness matrix
becomes positive definite again. This is due to the local elastic unloading which
starts at the bifurcation (cf. e.g. the case of the Euler beam in Section 2.3). In that
case, uniqueness of the solution is guaranteed and the secondary path can be treated
exactly in the same way as the fundamental path before bifurcation.
The situation is different on the fundamental path where Kf becomes indefinite
after the first bifurcation point. If the condition (2.95) fails at the same point where
Kf becomes indefinite then this critical point is a limit point. If the condition (2.95)
is still valid beyond the first critical point, it can be shown that there is a bifurca-
tion at every point on the fundamental path along which Kf is indefinite and the
condition (2.95) holds. This defines a continuous range of bifurcation points along
the fundamental path. It has to be noted that every point within this interval is
stable in a dynamical sense. These results, illustrated in Figure 2.15, can be seen
as a generalisation of the results obtained for Shanley’s column, i.e., along the fun-
damental path every point within the interval [Pt , Pr ] is a bifurcation point and is
stable in the dynamical sense.

31
 λ

directional instability

continuous range of
Q̇α wα > 0
bifurcations

Kf uniqueness
positive definite

deviation from the

fundamental path

Figure 2.15: Schematic illustration of the results of Section 2.4.3.

2.4.4 Energy approach

Instability of a deformation path

The case of a continuous range of bifurcation points is considered. The question
investigated here is to determine among all these possible paths which ones are
realisable in a physical sense. The response can be found by introducing small im-
perfections: when the imperfections tend to zero, the equilibrium paths obtained
tend to the secondary path branching out at the lowest bifurcation point. This
proves that only this secondary path has a physical meaning. A probabilistic ap-
proach gives the same result: if at a bifurcation point both paths (fundamental and
secondary) have comparable chances to be followed, then, due to the infinite number
of bifurcation points, any segment of the fundamental path has zero probability to
be followed. Therefore it can be concluded that beyond the first bifurcation point,
the fundamental path, even if each equilibrium point along it is stable in a dynam-
ical sense, has no physical meaning and can be considered as unstable. The notion
of stability introduced here does not concern any more an equilibrium point but a
deformation path: a deformation path is said to be unstable if it cannot be followed
in a physical sense.
Energy interpretation
The purpose of this section is to give an energy interpretation of the notion of
instability of a deformation path.
The deformation work in the body during a deformation process is expressed as
 t
W = Qα vα dτ (2.98)
0

32
The potential energy of the loading is expressed as

M
Ω=− Pα uα (2.99)
α=1

The energy functional is then defined by

E =W +Ω (2.100)

An increment of the value of E can be interpreted as the amount of external energy

which has to be supplied to the system consisting of the body and the loading in
order to produce the deformation increment.
By time derivation, it can be shown that Ė is independent of ṽ, i.e.

Ė (ṽ) = constant (2.101)

and that if ṽ1 and ṽ2 are two admissible velocity fields at the equilibrium state
1 1
Ë (ṽ1 ) − Ë (ṽ2 ) = J (ṽ1 ) − J (ṽ2 ) (2.102)
2 2
where J is the functional defined in equation (2.90). Hence, the variational princi-
ple (2.91) can be equivalently written as

δ Ë (ṽ, w̃) = 0 for every admissible w̃

= 0 (2.103)

The following criterion of path instability is then adopted: along a stable deforma-
tion path, the deformation increment must minimise the value of the increment of
the energy functional E. This can be related to the physical hypothesis that the real
deformation path minimises the energy consumption. By using (2.101), the solution
ṽ1 is therefore stable in an energy sense if

Ë (ṽ1 ) < Ë (ṽ) for every admissible ṽ (2.104)

which, from (2.102) can be rewritten as

J (ṽ1 ) < J (ṽ) for every admissible ṽ (2.105)

By comparing (2.96) and (2.105) it can be concluded that the fundamental solution
ṽf is stable as long as the uniqueness of the solution is guaranteed. Beyond the first
bifurcation point the tangent matrix Kf becomes indefinite along the fundamental
path. Therefore, from (2.92), the fundamental solution ṽf does no longer minimise
Ë and the solution becomes unstable in the energy sense. It can be proved that
another solution which minimises Ë exists as long as the condition (2.95) holds.
As a conclusion, it has been shown that the definition of stability in an energy sense
(minimisation of the consumption of energy) proposed by Petryk is equivalent to the
notion of stability of a deformation path given in the previous section. Consequently,
the fundamental path ṽf becomes unstable exactly when the tangent stiffness matrix
Kf becomes indefinite.

33
Chapter 3

2D beam element formulation

In this chapter, internal forces and tangent stiffness matrices for three plane beam
elements are derived. All of them are based on the same co-rotational approach,
and differ by the strain definition used in the local co-rotational coordinate system.
Based on the Bernoulli assumption, the first two elements use a linear and a shallow
arch strain definition, respectively. The third element is based on the Timoshenko
assumption with linear interpolations for the displacements.
Co-rotational formulation
The co-rotational approach, viewed as an alternative way of deriving efficient non-
linear finite elements, has generated an increased amount of interest in the last
decade [22, 23, 25, 35, 47, 53, 64, 72, 73, 83, 84, 95]. The main idea in this context is
to decompose the motion of the element into rigid body and pure deformational
parts, through the use of a local coordinate system (xl , zl ), see Figure 3.1, which
continuously rotates and translates with the element. The motion of the element
from the original undeformed configuration to the actual deformed one can thus be
split in two steps. The first one is a rigid rotation and translation of the element. The
second step consists of a deformation in the local coordinate system. Assuming that
the length of the element is properly chosen, the deformational part of the motion is
always small relative to the local axes. Consequently, the local deformations can be
expressed with a low order of non-linearity. It must however be emphasised that this
low-order non-linearity is only apparent as the geometrical non-linearity is included
in the motion of the local coordinate system.
The main advantage of a co-rotational approach is that it leads to an artificial
separation of the material and geometrical non-linearities when a linear strain def-
inition in the local coordinate system is used (which is the case for the first and
third elements developed in this chapter): plastic deformations occur in the local
coordinate system where geometrical linearity is assumed; geometrical non-linearity
is only present during the rigid rotation and translation of the undeformed beam.
This leads to very simple expressions for the local internal force vector and tangent
stiffness matrix. Even when a low-order geometrical non-linearity is included in the
strain definition (which is the case of the second element), the expressions for the
local internal force vector and tangent stiffness matrix are still very simple.

35
Constitutive equations
The assumption of small strains but large displacements and rotations, is adopted.
Therefore, constitutive elasto-plastic equations based on an additive strain decom-
position can be used. In this thesis, a von Mises material with isotropic constant
hardening is considered. The reason for this choice has already been discussed in
Section 1.1. For the Timoshenko element, interaction between the normal and the
shear stress is considered and the constitutive equations are derived from the ones
under plane stress conditions by introducing the beam hypothesis.
Resultant forces against Gauss integration
The computation of the local internal forces and stiffness matrix involves several
integrations over the element. With respect to integration over a cross-section, two
approaches can be considered: the layered and non-layered ones [67].
The non-layered method is based on plastic equations at the section level. This
means that the yield criterion is defined as a function of the internal forces. Such
functions can be found in [1, 56, 89] for rectangular and bi-symmetric cross-sections.
However, these functions are only approximations of the plastic flow rules. More-
over, they do not model the propagation of plastification and they consider perfect
elasto-plastic material models without hardening. For the Bernoulli assumption,
Crisfield [24] proposes a function which takes into account the propagation of the
plastification, but without considering hardening. Another problem with this ap-
proach is the practical impossibility to derive functions for arbitrary cross-sections.
As shown in Chapter 2 for the case of the Euler beam, a growing zone of elastic un-
loading is developing after bifurcation. Furthermore, the bifurcation point and the
nature of the post-bifurcation path are dependent on the hardening. It seems there-
fore impossible to study plastic instability problems with a non-layered approach.
This method, though efficient from a computational point of view, is limited to quick
and approximate analyses.
In the layered approach, numerical Gauss integration is performed through the cross-
section which means that the plastic rate equations are solved at each Gauss point.
A relatively high number of Gauss points is needed in order to model correctly the
growing of the elastic unloading zone (case of the Euler beam) and obtain an accurate
post-bifurcation path. As alternative to Gauss integration, the Lobatto rule could
also be considered. This procedure has the advantage of taking integration points
on the surface. However, if a sufficient number of points is taken, both methods give
almost the same numerical results.

3.1 Co-rotational framework

The purpose of this section is to derive relations between the local and global ex-
pressions of the internal force vector and tangent stiffness matrix. This presentation
is mainly taken from Crisfield [23].

36
3.1.1 Beam kinematics

ū xl
zl

θ̄1 θ2
θ1 w2
θ̄2
α β

2
w1 u2
βo
1
u1
x

Figure 3.1: Co-rotational formulation: beam kinematics.

The notations used in this section are defined in Figure 3.1. The coordinates for the
nodes 1 and 2 in the global coordinate system (x, z) are (x1 , z1 ) and (x2 , z2 ). The
vector of global displacements is defined by
 T
pg = u1 w1 θ1 u2 w2 θ2 (3.1)
The vector of local displacements is defined by
 T
pl = ū θ̄1 θ̄2 (3.2)
The components of pl can be computed according to
ū = ln − lo (3.3a)
θ̄1 = θ1 − α (3.3b)
θ̄2 = θ2 − α (3.3c)
In the above equations lo and ln denote the initial and current lengths of the element
 1/2
lo = (x2 − x1 )2 + (z2 − z1 )2 (3.4)
  1/2
ln = (x2 + u2 − x1 − u1 )2 + (z2 + w2 − z1 − w1 )2 (3.5)
and α denotes the rigid rotation which can be computed as
sin α = co s − so c (3.6a)
cos α = co c + so s (3.6b)

37
with
1
co = cos βo = (x2 − x1 ) (3.7a)
lo
1
so = sin βo = (z2 − z1 ) (3.7b)
lo
1
c = cos β = (x2 + u2 − x1 − u1 ) (3.7c)
ln
1
s = sin β = (z2 + w2 − z1 − w1 ) (3.7d)
ln
Then, provided that |α| < π, α is given by

α = sin−1 (sin α) if sin α ≥ 0 and cos α ≥ 0

α = cos−1 (cos α) if sin α ≥ 0 and cos α < 0
(3.8)
α = sin−1 (sin α) if sin α < 0 and cos α ≥ 0
α = − cos−1 (cos α) if sin α < 0 and cos α < 0

3.1.2 Virtual displacements

The virtual local displacements are obtained through differentiation of equations (3.3)
 
δū = c (δu2 − δu1 ) + s (δw2 − δw1 ) = −c −s 0 c s 0 δpg (3.9a)
δ θ̄1 = δθ1 − δα = δθ1 − δβ (α = β − βo ) (3.9b)
δ θ̄2 = δθ2 − δα = δθ2 − δβ (3.9c)

Further, δβ can be calculated by differentiation of (3.7d)

1
δβ = [(δw2 − δw1 ) ln − (z2 + w2 − z1 − w1 ) δln ] (3.10)
c ln2
where δln = δū is given by (3.9a). Using (3.7d), the expression of δβ becomes
1  
δβ = (δw2 − δw1 ) − s c (δu2 − δu1 ) − s2 (δw2 − δw1 ) (3.11)
c ln
which, after simplifications gives
1  
δβ = s −c 0 −s c 0 δpg (3.12)
ln
Hence, the transformation matrix B, defined by

δpl = B δpg (3.13)

is given by
 
−c −s 0 c s 0
B =  −s/ln c/ln 1 s/ln −c/ln 0  (3.14)
−s/ln c/ln 0 s/ln −c/ln 1

38
3.1.3 Internal forces

The relation between the local internal force vector fl and the global one fg is ob-
tained by equating the virtual work in both the local and global systems
V = δpT T T T
g fg = δpl fl = δpg B fl (3.15)
Equation (3.15) must apply for any arbitrary δpg , and hence the global internal
force vector fg is given by

fg = BT fl (3.16)

in which the local internal force vector fl = [ N M1 M2 ]T depends on the element

definition and will be calculated in the following sections.

3.1.4 Tangent stiffness matrix

The global tangent stiffness matrix Kg defined by

δfg = Kg δpg (3.17)

is obtained by differentiation of (3.16)
δfg = BT δfl + N δb1 + M1 δb2 + M2 δb3 (3.18)
where b2 , for example, is the second column of BT .
The following notations are introduced
 T
r = −c −s 0 c s 0 (3.19)
 T
z = s −c 0 −s c 0 (3.20)
which by differentiation gives
δr = z δβ (3.21a)
δz = −r δβ (3.21b)
Equations (3.9a) and (3.12) can be rewritten as
δū = δln = rT δpg (3.22a)
zT
δβ = δpg (3.22b)
ln
With these notations
b1 = r (3.23a)
 T z
b2 = 0 0 1 0 0 0 − (3.23b)
ln
 T z
b3 = 0 0 0 0 0 1 − (3.23c)
ln

39
which by differentiation gives

z zT
δb1 = δr = δpg (3.24a)
ln
δz z δln 1

δb2 = δb3 = − + 2 = 2 r zT + z rT δpg (3.24b)
ln ln ln

The first term in equation (3.18) is computed by introducing the local tangent
stiffness matrix Kl , which depends on the element definition and will be calculated
in the following sections.

δfl = Kl δpl = Kl B δpg (3.25)

Finally, from (3.17), (3.18), (3.24a), (3.24b) and (3.25), the expression of the global
tangent stiffness matrix becomes

z zT 1

Kg = BT Kl B + N + 2 r zT + z rT (M1 + M2 ) (3.26)
ln ln

Equations (3.16) and (3.26) give the relations between the local values of the internal
forces fl and tangent stiffness matrix Kl and the global ones, fg and Kg . These
relations are independent on the local element definition and are therefore available
for all three elements developed in the following sections. This approach is different
from the one adopted by Crisfield [23] where different local displacements definitions
where used, leading to different relations between the local and global values.
The purpose of the following sections is to calculate the local internal force vector
fl and tangent stiffness matrix Kl for the three elements.

3.2 Local linear Bernoulli element

3.2.1 Definition

This element is based on the classical linear beam theory, using a linear interpolation
for the axial displacement u and a cubic one for the vertical displacement w.
x
u= ū (3.27a)
L
 x 2 x2  x
w =x 1− θ̄1 + − 1 θ̄2 (3.27b)
L L L
According to the Bernoulli assumption, the curvature k and strain ε are defined by
   
∂2w 4 x 2 x
k= = − + 6 2 θ̄1 + − + 6 2 θ̄2 (3.28a)
∂x2 L L L L

    
∂u ū 4 x 2 x
ε= −kz = +z − 6 2 θ̄1 + − 6 2 θ̄2 (3.28b)
∂x L L L L L

40
3.2.2 Gauss integration

Because of the non-linear relation between ε and σ, it is not possible to derive

analytical expressions for the internal forces and the local tangent stiffness matrix
and a Gauss quadrature has to be adopted. The number of Gauss points per section
has already been discussed: two cases, with seven and fifteen points are considered.
The number of sections, i.e. the number of Gauss points along the beam axis is taken
as two in order to obtain the exact solution for the elastic case (the elastic strain
energy defined by Φ = 12 v Eε2 dv is a polynomial of second order in x and therefore
is exactly evaluated using a two points Gauss integration). Taking more points along
the beam axis would of course increase the precision in the elasto-plastic case, but
at some computational cost.
The position (and weight) of the Gauss points are within 0 < x < L
       
L 1 L L 1 L
x1 = 1− √ x2 = 1+ √ (3.29)
2 3 2 2 3 2

3.2.3 Local internal forces

The local internal forces are calculated by using the theorem of virtual work

V = σ δε dv = N δū + M1 δ θ̄1 + M2 δ θ̄2 (3.30)
v

which by introducing (3.28b) gives


    
δū 4 x 2 x
V = σ +z − 6 2 δ θ̄1 + z − 6 2 δ θ̄2 dv (3.31)
v L L L L L


Equation (3.30) must apply for every δū, δ θ̄1 , δ θ̄2 and hence the local internal forces
are given by

σ
N= dv (3.32a)
v L
  
4 x
M1 = σ z − 6 2 dv (3.32b)
v L L
  
2 x
M2 = σ z − 6 2 dv (3.32c)
v L L
Using a two points Gauss integration gives

  
1
N= σ dA1 + σ dA2 (3.33a)
2 A1 A2
√  √ 
3+1 1− 3
M1 = σ z dA1 + σ z dA2 (3.33b)
2 A1 2 A2
√  √ 
3−1 3+1
M2 = σ z dA1 − σ z dA2 (3.33c)
2 A1 2 A2

41
3.2.4 Local tangent stiffness matrix

Differentiation of σ gives

    
δū 4 x 2 x
δσ = Ê δε = Ê +z − 6 2 δ θ̄1 + z − 6 2 δ θ̄2 (3.34)
L L L L L
where Ê = E in the elastic range and Ê = Et in the plastic range. The local tangent
stiffness matrix Kl is obtained by differentiation of equations (3.32)

1
δN = δσ dv (3.35a)
L v
  
4 x
δM1 = δσ z − 6 2 dv (3.35b)
v L L
  
2 x
δM2 = δσ z − 6 2 dv (3.35c)
v L L
which from (3.34) gives

∂N 1
Kl11 = = 2 Ê dv (3.36a)
∂ ū L v
  2
∂M1 4 x
Kl22 = = z 2
− 6 2 dv (3.36b)
∂ θ̄1 v L L
  2
∂M2 2 x
Kl33 = = z 2
− 6 2 dv (3.36c)
∂ θ̄2 v L L
  
∂N ∂M1 1 4 x
Kl12 = Kl21 = = = Ê z − 6 2 dv (3.36d)
∂ θ̄1 ∂ ū L v L L
  
∂N ∂M2 1 2 x
Kl13 = Kl31 = = = Ê z − 6 2 dv (3.36e)
∂ θ̄2 ∂ ū L v L L
   
∂M1 ∂M2 4 x 2 x
Kl23 = Kl32 = = = Ê z 2
−6 2 − 6 2 dv (3.36f)
∂ θ̄2 ∂ θ̄1 v L L L L
Finally, by using a two points Gauss integration

  
1
Kl11 = Ê dA1 + Ê dA2 (3.37a)
2 L A1 A2

√ 2 
√ 2 
3+1 3 − 1
Kl22 = Ê z 2 dA1 + Ê z 2 dA2 (3.37b)
2L A1 2 L A2

√ 2 
√ 2 
3−1 3+1
Kl33 = Ê z 2 dA1 + Ê z 2 dA2 (3.37c)
2L A1 2 L A2
√  √ 
3+1 1− 3
Kl12 = Ê z dA1 + Ê z dA2 (3.37d)
2L A1 2L A2
√  √ 
3−1 3+1
Kl13 = Ê z dA1 − Ê z dA2 (3.37e)
2L A1 2L A2
 
1 2 1
Kl23 = Ê z dA1 + Ê z 2 dA2 (3.37f)
L A1 L A2

42
3.2.5 Constitutive equations

A bilinear strain–stress relation, see Figure 3.2, with an isotropic hardening is

adopted. This model presents a constant modulus Et in the plastic range, and
therefore the influence of the decreasing of Et , cf. Section 2.1.2, cannot be studied.
σ

σY
Et
E

ε
εY
Figure 3.2: Bilinear elastic-plastic law.

3.3 Local shallow arch Bernoulli element

3.3.1 Definition

The second element uses a shallow arch definition for the local strain, according to
   2 
1 ∂u 1 ∂w
ε = εf − k z = + dx − k z (3.38)
L L ∂x 2 ∂x
In (3.38), an averaged measure for the axial strain is introduced in order to avoid
membrane locking. Using the same definition of the curvature and the same inter-
polations as in Section 3.2.1, gives

    
ū 1 2 1 1 2 4 x 2 x
ε= + θ̄ − θ̄1 θ̄2 + θ̄ + z − 6 2 θ̄1 + − 6 2 θ̄2 (3.39)
L 15 1 30 15 2 L L L L
For the same reason as in Section 3.2.2, a two points Gauss integration along the
beam axis is adopted.

3.3.2 Local internal forces

As before, the local internal forces are computed using the theorem of virtual work

V = σ δε dv = N δū + M1 δ θ̄1 + M2 δ θ̄2 (3.40)
v

43
with δε evaluated from (3.39)
δū 2 1 1 2
δε = + θ̄1 δ θ̄1 − θ̄1 δ θ̄2 − θ̄2 δ θ̄1 + θ̄2 δ θ̄2 (3.41)
L
15 30  30  15
4 x 2 x
+z − 6 2 δ θ̄1 + − 6 2 δ θ̄2
L L L L
Hence

σ
N= dv (3.42a)
v L
    
2 1 4 x
M1 = θ̄1 − θ̄2 σ dv + σ z − 6 2 dv (3.42b)
15 30 v v L L
    
2 1 2 x
M2 = θ̄2 − θ̄1 σ dv + σ z − 6 2 dv (3.42c)
15 30 v v L L
which, using a two points Gauss integration, gives

  
1
N= σ dA1 + σ dA2 (3.43a)
2 A1 A2
 
  
L 2 1
M1 = θ̄1 − θ̄2 σ dA1 + σ dA2
2 15 30 A1 A2
√  √  (3.43b)
3+1 1− 3
+ σ z dA1 + σ z dA2
2 A1 2 A2
 
  
L 2 1
M2 = θ̄2 − θ̄1 σ dA1 + σ dA2
2 15 30 A1 A2
√  √  (3.43c)
3−1 3+1
+ σ z dA1 − σ z dA2
2 A1 2 A2

3.3.3 Local tangent stiffness matrix

The local tangent stiffness matrix is calculated by differentiation of (3.42).


1
δN = δσ dv (3.44a)
L v
   
2 1 2 1
δM1 = δ θ̄1 − δ θ̄2 σ dv + θ̄1 − θ̄2 δσ dv
15 30 v 15 30 v
   (3.44b)
4 x
+ δσ z − 6 2 dv
v L L
   
2 1 2 1
δM2 = δ θ̄2 − δ θ̄1 σ dv + θ̄2 − θ̄1 δσ dv
15 30 v 15 30 v
   (3.44c)
2 x
+ δσ z − 6 2 dv
v L L

with δσ = Ê δε and δε given in (3.41). Using a two points Gauss rule, the local
tangent stiffness matrix is evaluated as

44

  
∂N 1
Kl11 = = Ê dA1 + Ê dA2 (3.45a)
∂ ū 2L A1 A2
 
  
∂N 1 2 1
Kl12 = = θ̄1 − θ̄2 Ê dA1 + Ê dA2
∂ θ̄1 2 15 30 A1 A2
√  √  (3.45b)
3+1 1− 3
+ Ê z dA1 + Ê z dA2
2L A1 2L A2
 
  
∂N 1 2 1
Kl13 = = θ̄2 − θ̄1 Ê dA1 + Ê dA2
∂ θ̄2 2 15 30 A1 A2
√  √  (3.45c)
3−1 3+1
+ Ê z dA1 − Ê z dA2
2L A1 2L A2
 2
  
∂M1 1 2 1
Kl22 = = L θ̄1 − θ̄2 Ê dA1 + Ê dA2
∂ θ̄1 2 15 30 A1 A2
 
  
2 1 √  √ 
+ θ̄1 − θ̄2 3+1 Ê z dA1 + 1 − 3 Ê z dA2
15 30 A1 A2

√ 2 
√ 2  (3.45d)
3+1 3 − 1
+ Ê z 2 dA1 + Ê z 2 dA2
2L A1 2 L A2

  
L
+ σ dA1 + σ dA2
15 A1 A2
 2
  
∂M2 1 2 1
Kl33 = = L θ̄2 − θ̄1 Ê dA1 + Ê dA2
∂ θ̄2 2 15 30 A1 A2
 
  √  
2 1 √
+ θ̄2 − θ̄1 3−1 Ê z dA1 − 3+1 Ê z dA2
15 30 A1 A2

√ 2 
√ 2  (3.45e)
3−1 3 + 1
+ Ê z 2 dA1 + Ê z 2 dA2
2L A1 2L A2

  
L
+ σ dA1 + σ dA2
15 A1 A2
  
  
∂M1 1 2 1 2 1
Kl23 = = L θ̄1 − θ̄2 θ̄2 − θ̄1 Ê dA1 + Ê dA2
∂ θ̄2 2 15 30 15 30 A1 A2
√   √   
3−1 2 1 3+1 2 1
+ θ̄1 − θ̄2 + θ̄2 − θ̄1 Ê z dA1
2 15 30 2 15 30 A1
√   √   
3+1 2 1 3−1 2 1
− θ̄1 − θ̄2 + θ̄2 − θ̄1 Ê z dA2
2 15 30 2 15 30 A2

  
  
1 L
+ 2
Ê z dA1 + Ê z dA2 −
2
σ dA1 + σ dA2
L A1 A2 60 A1 A2
(3.45f)
Kl21 = Kl12 Kl31 = Kl13 Kl32 = Kl23 (3.45g)

45
Constitutive equations
The same elastic-plastic model as for the previous element is used.

3.4 Local linear Timoshenko element

3.4.1 Definition

A classical two noded Timoshenko beam element is defined with linear interpolations
for u, w and θ in the local co-rotational coordinate system.
x
u = ū (3.46a)
L
w=0 (3.46b)
 x x
θ = 1− θ̄1 + θ̄2 (3.46c)
L L
The curvature k, shear deformation γ and strain ε are defined by
∂θ θ̄2 − θ̄1
k= = (3.47a)
∂x L
∂w  x x
γ= −θ =− 1− θ̄1 − θ̄2 (3.47b)
∂x L L
∂u ū θ̄2 − θ̄1
ε= −kz = − z (3.47c)
∂x L L

3.4.2 Local internal forces

The local internal forces are here also calculated by using the theorem of virtual
work, which, taking account of the shear deformation can be expressed as

V = (σ δε + τ δγ) dv = N δū + M1 δ θ̄1 + M2 δ θ̄2 (3.48)
v

Further, δε and δγ are calculated by differentiation of (3.47b–c) which gives

  
σ
 x x 
V = δū − z δ θ̄2 − δ θ̄1 − τ 1 − δ θ̄1 + δ θ̄2 dv (3.49)
v L L L
Using a one point Gauss integration in order to avoid shear locking, the local internal
forces are computed from (3.48) and (3.49) as
 
σ
N= dv = σ dA (3.50a)
L
v  A
 
σ τ L
M1 = z− dv = σ z dA − τ dA (3.50b)
v L 2 A 2 A
   
σ τ L
M2 = − z− dv = − σ z dA − τ dA (3.50c)
v L 2 A 2 A

46
3.4.3 Local tangent stiffness matrix

The consistent tangent operator defined by




σ̇ Cct1 Cct3 ε̇
= (3.51)
τ̇ Cct3 Cct2 γ̇

will be calculated in Section 3.4.4. Equation (3.51) can be rewritten as

δσ = Cct1 δε + Cct3 δγ (3.52a)

δτ = Cct3 δε + Cct2 δγ (3.52b)

which, by using (3.47b–c) gives

Cct1
 Cct3

δσ = δū + zδ θ̄1 − zδ θ̄2 − δ θ̄1 + δ θ̄2 (3.53a)
L 2
Cct3
 Cct2

δτ = δū + zδ θ̄1 − zδ θ̄2 − δ θ̄1 + δ θ̄2 (3.53b)
L 2
Differentiation of (3.50) gives

δN = δσ dA (3.54a)
A
 
L
δM1 = δσ z dA − δτ dA (3.54b)
A 2 A
 
L
δM2 = − δσ z dA − δτ dA (3.54c)
A 2 A

Finally, from (3.53) and (3.54), the local tangent stiffness matrix is

∂N 1
Kl11 = = Cct1 dA (3.55a)
∂ ū L A
  
∂M1 1 L
Kl22 = = Cct1 z dA −
2
Cct3 z dA + Cct2 dA (3.55b)
∂ θ̄1 L A 4 A
 A 
∂M2 1 L
Kl33 = = Cct1 z 2 dA + Cct3 z dA + Cct2 dA (3.55c)
∂ θ̄2 L A A 4 A
 
∂N 1 1
Kl12 = Kl21 = = Cct1 z dA − Cct3 dA (3.55d)
∂ θ̄1 L A 2 A
 
∂N 1 1
Kl13 = Kl31 = = − Cct1 z dA − Cct3 dA (3.55e)
∂ θ̄2 L A 2 A
 
∂M1 1 L
Kl23 = Kl32 = = − 2
Cct1 z dA + Cct2 dA (3.55f)
∂ θ̄2 L A 4 A

3.4.4 Constitutive equations

The plastic rate equations for the Timoshenko beam are more complicated, since two
strains (ε, γ) and two stresses (σ, τ ) are involved. Many different algorithms exist

47
to integrate these equations. The simplest one is the forward-Euler scheme, which
avoids iteration at the Gauss point level. However, even if sub-incrementation is
used, the forward-Euler scheme does not give stresses which lie on the yield surface.
Another method to get an explicit integration is to assume a constant strain direction
(i.e. proportional loading) at each increment [75, 85]. However, this assumption,
based on the deformation theory of plasticity [11], is too restrictive for stability
problems. Another argument against this method is the implicit violation of the
plastic flow rules.
Among the iterative procedures (see e.g. [66]), the most popular one is the backward-
Euler scheme. It takes a simple form in case of the von Mises yield criterion (which
is adopted here) and it allows the generation of a consistent tangent operator which
maintains the quadratic convergence of the Newton-Raphson method used at the
structure level. Consistent tangent matrices can also be derived from other implicit
algorithms such as the mid-point rule, but this requires the computation of the
contact point when the initial stresses lie inside the yield surface. For all these
reasons, the backward-Euler scheme has been adopted.
Plane beam equations
The relations for the plane beam are derived from the von Mises material with
isotropic hardening under plane stress conditions [23] by setting

σz = 0 (3.56)

The von Mises yield function is


2 1/2
f = σx2 + 3τxz − σo (εps ) = σe − σo (εps ) (3.57)

with

2

εps = ε̇ps dt ε̇ps = √ ε̇2px + ε̇2pz + ε̇px ε̇px + γ̇pxz
2
/4 = λ̇ (3.58)
3
The hardening parameter H is calculated from the uniaxial stress strain law
∂σo Et
H= = (3.59)
∂εps 1 − Et /E
which gives

σo = σY + H εps (3.60)

where σY is the yield stress. The Prandtl-Reuss flow rules associated to (3.57) are
   
ε̇px 2 σx
∂f λ̇ 
ε̇p =  ε̇pz  = λ̇ = λ̇ a = −σx  (3.61)
∂σ 2 σe
γ̇pxz 6 τxz
The stress changes are related to the strain changes via

σ̇ = C (ε̇ − ε̇p ) = C ε̇ − λ̇ a (3.62)

48
where
     
1 ν 0 σ̇x ε̇x
E  
C= ν 1 0 σ̇ =  σ̇z  ε̇ =  ε̇z  (3.63)
1 − ν2
0 0 (1 − ν) /2 τ̇xz γ̇xz

By assuming that the stresses must remain on the yield surface in case of plastic
loading (λ̇ > 0), it is obtained

∂f T ∂f ∂σo
f˙ = σ̇ + ε̇ps = aT σ̇ − H λ̇ = 0 (3.64)
∂σ ∂σo ∂εps

which by using (3.62) gives

aT C ε̇
λ̇ = (3.65)
aT C a + H
Since σ̇z = 0, the second equation of (3.62) gives
 
λ̇ σx 1
ε̇z = −ν ε̇x + ν− (3.66)
σe 2

which by introduction in the first and third equations of (3.62) gives

   
σx 3 τxz
σ̇x = E ε̇x − λ̇ τ̇xz = G γ̇xz − λ̇ (3.67)
σe σe

Substitution of equations (3.66) and (3.67) in (3.65) gives after some algebra

σe (E σx ε̇x + 3 G τxz γ̇xz )

λ̇ = (3.68)
(E σx2 + 9 G τxz
2 ) + H σ2
e

Finally, introducing (3.65) in (3.62) gives

 
a aT C
σ̇ = Ct ε̇ = C I − ε̇ (3.69)
aT C a + H

where Ct is the tangent operator.

The backward-Euler algorithm shown in Figure 3.3 consists in applying an elastic
forward step (AB) followed by a return mapping (BC) on the yield surface.

σ C = σ B − ∆λ C aC = σ A + C ∆ε − ∆λ C aC (3.70)

The vector aC is normal to the yield surface at C, which is not known and therefore
an iterative procedure must be used.
Since σzC = σzA = 0, the second equation of (3.70) gives

∆λ σxC
∆εz = −ν ∆εx + (2ν − 1) (3.71)
2 σeC

49
 aC
σ1 B

σ2

Figure 3.3: Backward-Euler scheme.

which, by substitution into the first and third equations, gives after some algebra
 
σxC
σxC − σxA = E ∆εx − ∆λ (3.72a)
σeC
 
3 G τxzC
τxzC − τxzA = G ∆γxz − ∆λ (3.72b)
σeC

By introducing the notations

σ = σx τ = τxz ε = εx γ = γxz (3.73)

and new definitions for C, σ, ε, a




E 0 σ ε
C= σ= ε= (3.74)
0 G τ γ

equations (3.61), (3.67), (3.68) and (3.72) can be combined and rewritten as


λ̇ σ
ε̇p = λ̇ a = (3.75)
σe 3τ

σ̇ = C (ε̇ − ε̇p ) = C ε̇ − λ̇ a (3.76)

aT C ε̇
λ̇ = (3.77)
aT C a + H

σ C = σ B − ∆λ C aC = σ A + C ∆ε − ∆λ C aC (3.78)

It is therefore proved by comparing (3.62), (3.65) and (3.70) with (3.76), (3.77) and
(3.78) that the equations for the plane beam can be written in the same form as

50
those under plane stress conditions. It can be noted that the elastic forward step in
the backward-Euler scheme does not give σzB = 0, but the equation (3.78) proves
that σzB does not need to be calculated. A similar procedure is adopted in [23] to
derive equations under plane stress condition from the three-dimensional case.
Backward-Euler scheme
The algorithm of the backward-Euler scheme is taken from [23], with the notations
introduced at the end of the previous section.
The first estimation of σ C is calculated with
σ C = σ B − ∆λ C aB = σ A + C ∆ε − ∆λ C aB (3.79)
where aB is calculated from (3.75) and ∆λ is determined from a first order Taylor
expansion of the yield function around point B
∂f T ∂f ∂σo
f = fB + ∆σ B + B ∆σ B − H ∆λ
∆εps = fB + aT (3.80)
∂σ ∂σo ∂εps
∆σ B is then calculated from (3.76) as
∆σ B = C (∆ε − ∆λ aB ) (3.81)
with ∆ε = 0 since the total strain has already been applied in the elastic step (AB).
Introducing (3.81) in (3.80) and setting f = 0 gives
fB
∆λ = (3.82)
aT
B C aB + H
Equation (3.79) gives stresses which do not satisfy the yield function since the normal
at B is not the same as the normal at the final position C.
The iterative process is performed by introducing the vector r defined by the differ-
ence between the current stresses and the backward-Euler ones
r = σ C − (σ B − ∆λ C aC ) (3.83)
A truncated Taylor expansion of (3.83) gives
∂a
r = ro + σ̇ + λ̇ C a + ∆λ C σ̇ (3.84)
∂σ
where


∂a ∂ 2f 3 τ 2 −στ
= = (3.85)
∂σ ∂σ 2 σe3 −στ σ 2

and σ̇ is the change in σ C and λ̇ is the change in ∆λ. The subscript C is from
now dropped in order to simplify the notations (all the notations without subscripts
refer to C). Setting r to zero in (3.84) gives
 −1  
∂a
σ̇ = − I + ∆λ C ro + λ̇ C a = −Q−1 ro + λ̇ C a (3.86)
∂σ

51
A truncated Taylor expansion of the yield function at C gives

∂f T ∂f ∂σo
f = fo + σ̇ + ε̇ps = fo + aT σ̇ − H λ̇ (3.87)
∂σ ∂σo ∂εps

further, setting f = 0 and using (3.86) gives

fo − aT Q−1 ro
λ̇ = (3.88)
aT Q−1 C a + H
The backward-Euler scheme is summarised by a flowchart in Figure 3.4.
Consistent tangent operator
The tangent operator Ct defined in equation (3.69) assumes infinitesimal strain
and stress changes. However, in order to eliminate numerical elastic unloading, a
path-independent strategy [28] is adopted. This means that at each iteration, the
strain changes are calculated with reference to the beginning of the increment (last
equilibrium state) and not with reference to the last iteration. This is the reason
why the use of Ct leads to low convergence rates for the Newton-Raphson method at
the structural level, unless the increments are infinitesimal. In order to maintain the
quadratic convergence inherent in the Newton-Raphson method, a tangent operator
Cct , consistent with the backward-Euler scheme is derived. The concept of consistent
tangent operator has been first introduced by Simo and Taylor [90,91] and then used
by several authors. The derivation presented here follows [23].
Differentiation of (3.78) gives (subscript C is dropped)
∂a
σ̇ = C ε̇ − λ̇ C a − ∆λ C σ̇ (3.89)
∂σ
where the last term in equation (3.89) is omitted in the derivation of the standard
tangent operator. From (3.89), it is obtained
 −1  
∂a
σ̇ = I + ∆λ C C ε̇ − λ̇ a = R ε̇ − λ̇ a (3.90)
∂σ

The stresses must remain on the yield surface, which, from (3.64), gives

f˙ = aT σ̇ − H λ̇ = aT + R ε̇ − λ̇ a − H λ̇ = 0 (3.91)

and hence
aT R ε̇
λ̇ = (3.92)
aT R a + H
Introducing (3.92) in (3.90) finally gives
 
a aT R
σ̇ = Cct ε̇ = R I − ε̇ (3.93)
aT R a + H

52
 input
σA εpsA ∆ε

elastic predictor
σ B = σ A + C ∆ε
σo = σY + H εpsA
(3.57) → fB

if fB > 0 first estimation

(3.75) → aB (3.82) → ∆λ
σ = σ B − ∆λ C aB
εps = εpsA + ∆λ

σo = σY + H εps
(3.57) → f

while f > 0 (tolerance)

(3.75) → a (3.83) → r
∂a
(3.85) → (3.86) → Q
∂σ
(3.88) → λ̇ (3.86) → σ

∆λ = ∆λ + λ̇
εps = εpsA + ∆λ
σo = σY + H εps

(3.57) → f

Figure 3.4: Backward-Euler algorithm.

Chapter 4

3D beam element formulation

As for the 2D case, the main idea of the co-rotational approach in 3D [22,23,25,35,47,
53,64,72,73,83,84,95] is to decompose the motion of the element into rigid body and
pure deformational parts, through the use of a reference system which continuously
rotates and translates with the element. The deformational response is captured at
the level of the local reference frame, whereas the geometric non-linearity induced by
the large rigid-body motion, is incorporated in the transformation matrices relating
local and global quantities. Assuming the pure deformation part to be small, a
linear theory can be used from the onset in the local system. As pointed out by
Haugen and Felippa in their review of the topic [43, 44], the main benefit of such an
assumption is the possibility to reuse existing high-performance linear elements.
In order to fully exploit this possibility, Rankin and co-workers [64,83,84] introduced
the so called “element independent co-rotational formulation”. The definition of the
element resorts to several changes of variables from the local frame to the global one.
This is done through the use of a projector matrix which relates the variations of
the local displacements to the variations of the global ones, by extracting the rigid
body modes from the latter. A very similar line of work was adopted by Pacoste
and Eriksson [73] and will also be followed in this chapter. With respect to the cited
reference, several improvements have been introduced.
The first one concerns the co-rotational framework. A different definition of the local
coordinate system is adopted. Besides, the number of local displacements is kept
to seven (one translation and six rotations) and therefore, the projector matrix is
never explicitly computed. Instead, only certain block components in this matrix are
evaluated and used directly in the expressions of the internal force vector and tangent
stiffness in global coordinates. The objective here was to obtain a formulation which
is more efficient from a computational point of view. Rigid links and eccentric forces
are also introduced.
The second one is related to the parameterisation of finite 3D rotations. Thus, in [73]
a 3-parameter representation based on the so-called “rotational vector”, is used.
The main advantage here is that the rotational variables become additive and the
necessity of a special updating procedure is avoided. However, this parameterisation
can not be retained globally [51,73] as the magnitude of the rotational vector must be

55
restricted to values less than 2π. In order to overcome this inconvenience, a different
possibility is explored in the present work. Essentially it amounts to using a vector
like parameterisation but this time within each increment. Using the terminology
in [51], this alternative is based on the spatial form of the incremental rotation
vector. Additive updates will still apply but this time only within an increment at
the level of the iterative corrections.
The third improvement is related to warping effects which are introduced by adding
a seventh degree of freedom at each node of the element. With respect to [64, 73],
this requires an adaptation of the co-rotational framework.
As already mentioned, the main feature of the adopted co-rotational formulation is
its independence on the assumptions used to derive the internal forces and tangent
stiffness in local coordinates. This means that for elements with the same number
of nodes and degrees of freedom the co-rotational framework is the same. Using
this property, different local formulations can be tested and compared. In many
instances, see e.g. [22, 43, 44, 64, 73], a linear strain definition is adopted in the local
system. However, as the numerical applications will demonstrate, such a choice
will lead to incorrect results for certain classes of problems, especially when the
torsional effects are important. For this reason a different approach is advocated in
this work. At the level of the local frame, the kinematic description proposed by
Gruttmann et al. [41] is adopted and used to construct a consistent second order
approximation to the Green-Lagrange strains. Various possibilities to simplify the
resulting expressions are further discussed and tested in numerical examples.
The organisation of this chapter is as follows. Section 4.1 discusses some aspects re-
lated to the parameterisation of finite 3D rotations, whereas the general co-rotational
framework for a two node beam element is introduced in Section 4.2. Section 4.3 is
devoted to the local element formulation in elasticity. Several local elements, based
on Timoshenko and Bernoulli assumptions and including or not warping effects are
presented. Based on the same strain assumptions as in elasticity, elasto-plastic local
Timoshenko elements are derived in Section 4.4. Finally, Section 4.5 addresses cer-
tain issues related to the applied external forces, e.g. rigid eccentricities, definition
of external moments.

4.1 Parameterisation of finite 3D rotations

One of the central issues in the development of a non-linear beam element is the
treatment of finite 3D rotations. In general, finite rotations are represented through
an orthogonal tensor R which is an element of the SO(3) rotation group. Its co-
ordinate representation is a 3 × 3 orthogonal matrix involving nine components.
However, owing to its orthonormality, it can be described in terms of only three
independent parameters. One alternative in this context, is based on the so-called
“rotational vector” [51, 68, 72, 73], defined by

Ψ = uψ (4.1)

56
The geometrical significance of the above definition is the following: any finite rota-
tion can be represented by a unique rotation with an angle ψ about an axis defined
by the unit vector u. The magnitude of ψ is given by

ψ = Ψ21 + Ψ22 + Ψ23 (4.2)

where Ψi , (i = 1, 2, 3) are the components of Ψ. In terms of Ψ, the orthogonal

matrix R admits the following representation

2
sin ψ 1 sin(ψ/2) 2
R=I+ Ψ̃ + Ψ̃ (4.3)
ψ 2 ψ/2
Here, as in the sequel, a tilde denotes the skew-symmetric matrix obtained with the
components of the corresponding vector, whereas I denotes a 3 × 3 identity matrix.
If the trigonometric functions in the above equation are expanded in Taylor series,
R can be written as
1 2
R = I + Ψ̃ + Ψ̃ + · · · = exp(Ψ̃) (4.4)
2

Following the notations introduced in Figure 4.1, Ro is a rotation operator which

maps the orthogonal cartesian frame ei into the orthonormal frame ti , which implies
ti = Ro ei , (i = 1, 2, 3). An incremental rotation which carries the moving frame ti

Rs t2
t1
t2 t2
t1 t1
Ro
e2
t3
t3
t3
e1
R
e3
R

Figure 4.1: Parameterisation of finite 3D rotations.

into a new position ti can be expressed through a spatial rotation Θ which is viewed
as a rotation applied to ti

ti = R ei = Rs Ro ei Rs = exp(Θ̃) (4.5)

or through a material rotation W, which is viewed as a rotation applied to the

material frame ei

ti = R ei = Ro Rm ei Rm = exp(W̃) (4.6)

57
In the terminology of Argyris [1], the sequence of rotation in equation (4.5) is denoted
as “rotation around follower axes”, i.e. axes rotated by the previous rotation in the
sequence, whereas the sequence described in equation (4.6) is denoted as “rotation
around fixed axes”, i.e. the rotation axes remain fixed but the rotation sequence itself
is inverted. The relationship connecting the spatial Θ and material W incremental
rotation vectors is thus given by

Θ = Ro W (4.7)

Let now consider an infinitesimal rotation which carries the frame ti into ti , such
that ti = R ei , (i = 1, 2, 3). Using the same dualism as above, admissible variations
δR are computed as
d d 
δR = [R ]=0 = exp( θ̃) R = δ θ̃ R (4.8)
d d =0
d
= (R exp( ω̃))=0 = R δ ω̃
d
where δ ω̃ denotes material and δ θ̃ spatial angular variations, i.e. infinitesimal ro-
tations superimposed onto the existing R. As defined in the above equation, the
parameterisation of admissible variations δR is largely independent of the technique
used to parameterise R itself. In order to obtain a consistent parameterisation, δθ
(or alternatively δω) must be projected onto the parameter space adopted for R.
This issue will be addressed in the remainder of this subsection, for the types of
parameterisation adopted in this work.

4.1.1 Rotational vector

Considering the spatial form of equation (4.8), admissible variations δΨ which lead
to the same perturbed matrix R are found from

R = exp(Ψ̃ ) = exp( δ θ̃) R (4.9)

where Ψ = Ψ +  δΨ. Taking into account that R = exp(Ψ̃), this leads to

exp( δ θ̃) = exp(Ψ̃ ) exp(−Ψ̃) (4.10)

Using the formula for compound rotations in terms of the rotational pseudo-vector
(i.e. Rodrigues parameters) and differentiating with respect to , gives

δθ = Ts (Ψ) δΨ (4.11)

with
   2
sin ψ sin ψ 1 sin(ψ/2)
Ts (Ψ) = I+ 1− T
uu + Ψ̃ (4.12)
ψ ψ 2 (ψ/2)
Similar equations involving material angular variations δω can be easily derived by
noting that δθ = R δω. For details concerning the above derivation the interested
reader is referred to [51, 72].

58
In connection with the linear operator Ts it should also be noted that
2(1 − cos ψ)
det(Ts ) = (4.13)
ψ2
which shows that the corresponding mapping ceases to be a bijection for ψ = 2kπ,
k = 1, 2, ... [51, 72]. Avoiding these points, the inverse mapping is well defined as

δΨ = T−1
s (Ψ) δθ (4.14)

with
 
(ψ/2) (ψ/2) 1
T−1
s = I+ 1− u uT − Ψ̃ (4.15)
tan(ψ/2) tan(ψ/2) 2

4.1.2 Incremental rotation vector

Considering equation (4.5), admissible variations δΘ which lead to the same per-
turbed matrix R are found from

R = exp(Θ̃ ) Ro = exp( δ θ̃) R (4.16)

where Θ = Θ +  δΘ. Taking equation (4.5) into account leads to

exp( δ θ̃) = exp(Θ̃ ) exp(−Θ̃) (4.17)

Following a procedure entirely similar to the one used in the preceding subsection
(for details see [51]) leads to

δθ = Ts (Θ) δΘ (4.18)

Note that a similar relation connecting δω and δW can also be obtained. This issue
however will not be further utilised in this thesis.

4.2 Co-rotational framework

The objective of this section is to define the co-rotational framework for a two node
beam element. The central idea in this context is to introduce a local coordinate
system which continuously rotates and translates with the element. With respect to
the moving frame, local (deformational) displacements pl are defined by extracting
the rigid body modes from the global displacements pgg . The local displacements
are expressed as functions of the global ones, i.e.

pl = pl (pgg ) (4.19)

and used in order to compute the internal force vector fl and tangent stiffness matrix
Kl in local coordinates. Note that fl and Kl depend only on the definition of the
local strains and not on the particular form of equation (4.19).

59
The expression of the internal force vector in global coordinates fg , can be obtained
by equating the internal virtual work in both the global and local systems
gT
V = δpT
l fl = δpg fg (4.20)

Using equation (4.19), the connection between the variations of local and global
displacements is obtained as

δpl = B δpgg (4.21)

which, upon substitution into (4.20), provides the result of interest

fg = BT fl (4.22)

The expression of the tangent stiffness matrix in global coordinates Kg , is obtained

by taking the variations of equation (4.22), which gives

∂B
Kg = BT Kl B + :f (4.23)
∂pgg l

The second term on the right-hand side of the above equation is the so-called “geo-
metric” stiffness, with the symbol “:” used to denote a contraction.
In equations (4.22) and (4.23), B and ∂B/∂pgg play the role of transformation ma-
trices required in order to re-express fl and Kl in global coordinates. These ma-
trices, which actually define the co-rotational framework, depend on the non-linear
functions in equation (4.19) and thus on the choice of the local coordinate system.
However, they are independent of the particular strain definition used in order to
derive fl and Kl . Consequently, various co-rotational elements defined using differ-
ent local strain assumptions but the same type of local coordinate system will share
the same transformation matrices, i.e. the co-rotational formulation is “element in-
dependent” [64]. Using this property, various local assumptions can be placed at
the core of the co-rotational formulation and tested for efficiency and accuracy. The
issues pertaining to the development of a specific form for the local internal forces
fl and tangent stiffness Kl will however be addressed in the following sections.
The remainder of this section will instead focus on the definition of the co-rotational
framework. The approach put forth follows mainly [64, 73]. Thus, the change of
variables (4.19) and subsequently the definition of the transformation matrices in
equations (4.22) and (4.23) are performed by extracting the rigid body modes from
the variations of the global displacements.

4.2.1 Coordinate systems, beam kinematics

The definition of the co-rotational two node beam element described in this section,
involves several reference systems. First (see Figure 4.2), a global reference system
defined by the triad of unit orthogonal vectors eα , α = 1, 2, 3. Second, a local system
which continuously rotates and translates with the element. The orthonormal basis

60
 R̄1 r2
Rg1
t11
t12 t21
eo2 t22 R̄2
1 r1
r3 t13 2
1
eo1 t23
eo3
2 Rg2

e2
Ro

e1 Rr

e3

Figure 4.2: Beam kinematics and coordinate systems.

vectors of the local system are denoted by rα , α = 1, 2, 3. In the initial (undeformed)

configuration, the local system is defined by the orthonormal triad eoα . In addition,
t1α and t2α , α = 1, 2, 3, denote two unit triads rigidly attached to nodes 1 and 2.
According to the main idea of the co-rotational formulation, the motion of the
element from the initial to the final deformed configuration is split into a rigid body
component and a deformational part. The rigid body component consists of a rigid
translation and rotation of the local element frame. The origin of the local system is
taken at node one and thus the rigid translation is defined by ug1 , i.e. the translation
at node 1. Here as in the sequel, the superscript g indicates quantities expressed in
the global reference system.
The rigid rotation is such that the new orientation of the local reference system is
defined by an orthogonal matrix Rr , given by

R r = [ r1 r2 r3 ] (4.24)

The first coordinate axis of the local system is defined by the line connecting nodes
1 and 2 of the element. Consequently, r1 is given by
xg2 + u2g − x1g − ug1
r1 = (4.25)
ln
with xgi , i = 1, 2 denoting the nodal coordinates in the initial undeformed configu-
ration and ln denoting the current length of the beam, i.e.

ln = xg2 + ug2 − xg1 − ug1  (4.26)

61
The remaining two axes are determined with the help of an auxiliary vector q. In
the initial configuration q is directed along the local eo2 direction, whereas in the
deformed configuration its orientation is obtained from
1
q= (q1 + q2 ) qi = Rgi Ro [ 0 1 0 ]T i = 1, 2 (4.27)
2
where Rg1 and Rg2 are the orthogonal matrices used to specify the orientation of the
nodal triads t1α and t2α respectively, and Ro specifies the orientation of the local
frame in the initial configuration, i.e. Ro = [ eo1 eo2 eo3 ]. The unit vectors r2 and r3
are then computed by the vector products
r1 × q
r3 = r2 = r 3 × r1 (4.28)
r1 × q
and the orthogonal matrix Rr in equation (4.24) is completely determined.
The rigid motion previously described, is accompanied by local deformational dis-
placements with respect to the local element axes. In this context, due to the
particular choice of the local system, the local translations at node 1 will be zero.
Moreover, at node 2, the only non zero component is the translation along r1 . This
can easily be evaluated according to

ū = ln − lo (4.29)

with lo denoting the length of the beam in the original undeformed configuration.
Here as in the sequel, an overbar denotes a deformational kinematic quantity.
The global rotations at node i can be expressed in terms of the rotation of the local
axes, defined by Rr , followed by a local rotation relative to these axes; this is defined
by the orthogonal matrix R̄i . Using the terminology of Section 4.1, the latter should
be viewed as a material rotation. Consequently, the orientation of the nodal triad
tiα can be obtained by means of the product Rr R̄i . On the other hand, (see Figure
4.2) this orientation can also be obtained through the product Rgi Ro , which gives
g
R̄i = RT
r Ri Ro i = 1, 2 (4.30)

The local rotations are then evaluated from


ϑ̄i = log R̄i i = 1, 2 (4.31)

The explicit form of equation (4.19) is thus completely specified through equations
(4.29), (4.30) and (4.31), with pl given by
 
T T T
pl = ū ϑ̄1 ϑ̄2 (4.32)

Correspondingly, the local internal force vector is given by

 
T T
fl = n mT 1 m2 (4.33)

where n denotes the axial force whereas m1 and m2 denote the moments at nodes
1 and 2, respectively.

62
The next step will then be to determine the transformation matrix B in equation
(4.21). This matrix relates the variations of the local displacements
 T
T T
δpl = δū δ ϑ̄1 δ ϑ̄2 (4.34)

to their global counterparts

 T
δpgg = δug1 T δθ 1g T δug2 T δθ 2g T (4.35)

with δθ ig , i = 1, 2 denoting spatial angular variations as defined in (4.8). The

expression of B, and subsequently the transformations in equations (4.22) and (4.23)
are derived using a sequence of two changes of variables, as described in the following
subsections.

4.2.2 Change of variables δ ϑ̄ −→ δ θ̄

The general procedure for evaluating the transformation matrix B involves the vari-
ations of equations (4.29) and (4.30). Referring to (4.30), admissible variations δ R̄i ,
i = 1, 2 are computed – cf. equation (4.8) – according to
˜ R̄
δ R̄i = δ θ̄ (4.36)
i i

The angular variations δ θ̄ i involved in the above equation, are of a different nature
than δ ϑ̄i as defined in (4.31). A change of variables from δ ϑ̄i to δ θ̄ i is thus required.
In order to construct such a change of variables, using equations (4.14), (4.31) and
(4.36), δ θ̄ i and δ ϑ̄i are connected by

δ ϑ̄ = T−1
s (ϑ̄) δ θ̄ (4.37)

which, by introducing the notation

 
T T T
δpa = δū δ θ̄ 1 δ θ̄ 2 (4.38)

gives
 
1 01,3 01,3
 
δpl = Ba δpa Ba =  −1
 03,1 Ts (ϑ̄1 ) 0 
 (4.39)
−1
03,1 0 Ts (ϑ̄2 )

Here, as in the sequel 0i,j denotes an i × j zero matrix. For a 3 × 3 zero matrix the
notation 0 is however used.
Further, a virtual work equation gives

fa = BT
a fl (4.40)

63
with fa denoting the internal force vector consistent with δpa . The corresponding
transformation for the local tangent stiffness matrices, i.e. Kl and Ka , is obtained
by taking the variations of (4.40)

δfa = BT T
a δfl + δBa fl (4.41)

where, by definition

δfl = Kl δpl δfa = Ka δpa (4.42)

Using equations (4.39), (4.41) and (4.42) gives the required transformation
 
0 01,3 01,3
 
Ka = BT a Kl Ba + Kh Kh =   0 3,1 K h1 0 
 (4.43)
03,1 0 Kh2

The expressions of Kh1 and Kh2 are computed from

∂ ∂ ∂ ϑ̄ ∂
[T−T
s v] = [T−T
s v] = [T−T v] T−1 (4.44)
∂ θ̄ ∂ ϑ̄ ∂ θ̄ ∂ ϑ̄ s s

with the vector v maintained constant during differentiation. Using (4.14) gives
after some algebra


∂ T T ˜ 2 T 1
[T v] = η [ϑ̄ v − 2 v ϑ̄ + (ϑ̄ · v) I] + µ ϑ̄ [v ϑ̄ ] − ṽ T−1
−T T
s (ϑ̄) (4.45)
∂ θ̄ s 2
with the coefficients η and µ given by
2 sin α − α (1 + cos α) α (α + sin α) − 8 sin2 (α/2)
η= µ= α = ϑ̄ (4.46)
2 α2 sin α 4 α4 sin2 (α/2)
Thus, Khi , i = 1, 2 are evaluated from equation (4.45) with ϑ̄ = ϑ̄i and v = mi ,
with mi as defined in (4.33).
Finally, it should be mentioned that, at least for small increments, ϑ̄ is a small
quantity. Consequently, T−1 s (ϑ̄) is close to identity. For this reason, the transfor-
mation described by equations (4.40) and (4.43) is often omitted in the formulation
of co-rotational elements (cf. e.g. [22]). The numerical tests performed in [73] showed
that this omission has scarcely any effect on the correctness of the results, although
it has an impact on the convergence properties of the element.

4.2.3 Change of variables δpa −→ δpgg

The second step of the variable change involves δpa and δpgg , as defined in equations
(4.38) and (4.35), respectively.
Referring first to the local axial translation ū, the variations of equation (4.29) give
 
δū = δln = r δpgg r = −rT T
1 01,3 r1 01,3 (4.47)

64
For the rotational terms, the variations of equation (4.30) are needed
g g
δ R̄i = δRT T
r Ri Ro + Rr δRi Ro (4.48)

where δ R̄i is defined in (4.36) whereas δRgi and δRr are computed using the spatial
form of equation (4.8), i.e.
g g
δRgi = δ θ̃ i Rgi δRr = δ θ̃ r Rr (4.49)

δRT T
r is calculated from the orthogonality condition Rr Rr = I which, by differenti-
ation and introduction of (4.49) gives

δRr RT T
r + Rr δRr = 0 (4.50a)
g
δ θ̃ r Rr RT
r + Rr δRT
r =0 (4.50b)

and then
g
r = −Rr δ θ̃ r
δRT T
(4.51)

Using (4.36), (4.49) and (4.51), equation (4.48) can be rewritten as

g
˜ R̄ = −RT δ θ̃ Rg R + RT δ θ̃ Rg R g
δ θ̄ i i r r i o r i i o
g g g g
= −Rr δ θ̃ r Rr Rr Ri Ro + Rr δ θ̃ i Rr RT
T T T
r Ri Ro (4.52)
e e
= (δ θ̃ i − δ θ̃ r ) R̄i

where use has been made of equation (4.30) and of the fact that Rr transforms a
vector and a tensor from global to local coordinates according to

xe = RT
r x
g
x̃e = RT g
r x̃ Rr (4.53)

Thus, equation (4.52) gives

δ θ̄ i = δθ ie − δθ re i = 1, 2 (4.54)

Further, let
 
Rr 0 0 0
 
 0 Rr 0 0 
δpge =E T
δpgg E=
 0
 (4.55)
 0 Rr 0 

0 0 0 Rr

Then, using the chain rule, δ θ̄ i is evaluated as

∂ θ̄ i ∂pge g ∂ θ̄ i T g
δ θ̄ i = e g δpg = E δpg i = 1, 2 (4.56)
∂pg ∂pg ∂pge

Substituting from equation (4.54) gives

     T 
δ θ̄ 1 0 I 0 0 G
= − ET δpgg = P ET δpgg (4.57)
δ θ̄ 2 0 0 0 I G T

65
where the matrix G is defined by
∂θ re
G= (4.58)
∂pge

Hence, from (4.47) and (4.57), the connection between δpa and δpgg is given by
 
r
δpa = Bg δpgg Bg = (4.59)
P ET

Expression of G
The expression of G is obtained from (4.49) which can be rewritten as
g
δ θ̃ r = δRr RT
r (4.60)

and after the transformation (4.53) as

e
δ θ̃ r = RT
r δRr (4.61)

From (4.24) and the above equation, it can easily be found that
   
δθre1 −rT2 δr3
 e   
δθ re = 
 δθr 2
 =  −rT
  3 δr 1

 (4.62)
δθre3 rT
2 δr 1

Introducing the notation ugi = [ ugi1 ugi2 ugi3 ]T (i = 1, 2), differentiation of (4.25) gives
 g 
δu21 − δug11
1   g 
δrg1 = I − r1 r T 1 
 δu 22 − δu g 
12  (4.63)
ln
δug23 − δug13

and after transformation (4.53) in the local coordinate system, it is obtained

 
δue21 − δue11
1  e 
δre1 =  δu − δu e 
(4.64)
ln  22 12 
δue23 − δue13

Hence, since the local expressions of r2 and r3 are [ 0 1 0 ]T and [ 0 0 1 ]T , equation

(4.62) gives
1
δθre2 = (δue13 − δue23 ) (4.65a)
ln
1
δθre3 = (δue22 − δue12 ) (4.65b)
ln

66
The evaluation of δθre1 is more complicated and can be performed as follows. Differ-
entiation of (4.27) gives
1
δq = (δRg1 + δRg2 ) Ro [ 0 1 0 ]T
2
1 g g
= (δ θ̃ 1 Rg1 + δ θ̃ 2 Rg2 ) Ro [ 0 1 0 ]T (4.66)
2
1 g g
= (δ θ̃ 1 q1 + δ θ̃ 2 q2 )
2
The local expressions of the vectors q, q1 and q2 are denoted by
     
q1 q11 q21
     
RT
r q =  q2  RT r q1 =  q12  RTr q2 =  q22  (4.67)
0 q13 q23
The last coordinate of RT
r q is zero since q is perpendicular to r3 .

The local expression of δq can be deduced from equation (4.66) as

   
q11 q21
e 1 e  1 e 
δq = δ θ̃ 1  q12  + δ θ̃ 2  q22  (4.68)
2 2
q13 q23
which after calculation gives
 
e e e e
− q12 δθ13 + q13 δθ12 − q22 δθ23 + q23 δθ22
1  
δqe =  e
+ q11 δθ13 e
− q13 δθ11 e
+ q21 δθ23 e
− q23 δθ21  (4.69)
2 
e e e e
− q11 δθ12 + q12 δθ11 − q21 δθ22 + q22 δθ21
The following notations are introduced
q1 q11 q12 q21 q22
η= η11 = η12 = η21 = η22 = (4.70)
q2 q2 q2 q2 q2
The differentiation of r3 is calculated from its definition (4.28). By noting that
r1 × q = q2 , the first line of equation (4.62) can be rewritten as
 
e rT e rT e 1
δθr1 = − δr1 × q −
2 2
r1 × δq − δ 2 (r1 × q)
rT (4.71)
q2 q2 q2
The last term in the above equation is zero. The two others terms can be evaluated
from equations (4.64), (4.67) and (4.69). The result, after some work, is
η η11 e η12 e η21 e η22 e
δθre1 = (δw1e − δw2e ) − δθ12 + δθ11 − δθ22 + δθ21 (4.72)
ln 2 2 2 2
Finally, the expression for the matrix G is
 η η12 η11 η η22 η21 
0 0 − 0 0 0 − − 0
 ln 2 2 ln 2 2 
 
 1 1 
G =
T
 0 0 0 0 0 0 0 − 0 0 0 
 (4.73)
 ln ln 
 1 1 
0 − 0 0 0 0 0 0 0 0 0
ln ln

67
Internal forces
Using (4.59), the internal force vector in global coordinates is computed as

fg = BT
g fa (4.74)

Note that, according to the sequence of variable changes previously defined, the
matrix B in equation (4.21) is explicitly given by the product Ba Bg .
Tangent stiffness matrix
Differentiation of (4.74) gives
g
δfg = BT T T
g Ka Bg δpg + δr fa1 + δ(E P ) m (4.75)

with
 T
m = fa2 fa3 fa4 fa5 fa6 fa7 (4.76)

where fai , i = 1, 7 denotes the i-th component of the vector fa .

From (4.47) and (4.63), it can easily be derived that
 
D3 0 −D3 0
 0 
 0 0 0  1
δrT = D δpgg D=  D3 = (I − r1 rT
1 ) (4.77)
 −D3 0 D3 0  ln
0 0 0 0

The last term in expression (4.75) is evaluated from

δ(E PT ) m = δE PT m + E δPT m (4.78)

By introducing
 
n1
 
 m1 
PT m =   (4.79)
 n2 
m2

and using (4.55) and (4.61), the first term in (4.78) can be expressed as
 e    e 
Rr δ θ̃ r 0 0 0 n1 δ θ̃ r n1
   
m1 
e e
 0 Rr δ θ̃ r 0 0    δ θ̃ r m1 
δE P m = 
T


  = E 

δ θ̃ r n2 
e e
 0 0 Rr δ θ̃ r 0  n2   
e e
0 0 0 Rr δ θ̃ r m2 δ θ̃ r m2
(4.80)

which, using the relation

ã b = −b̃ a (4.81)

68
gives
 
ñ1
 
 m̃1 
δE PT m = −E Q δθ re Q=  (4.82)
 ñ2 
m̃2
Then, by using (4.58), it is obtained
δE PT m = −E Q GT ET δpgg (4.83)
The calculation of the second term of (4.78) requires the value of δPT which can be
obtained by introducing the matrix A such as
 
0 0 0 0 0 0 0 0 0 0 0 0
 
AT =  0 0 0 0 0 0 0 0 −ln 0 0 0  (4.84)
0 0 0 0 0 0 0 ln 0 0 0 0
and by noting that
AT G = I (4.85)
Differentiation of the above equation gives
δAT G + AT δG = 0 (4.86)
and hence
δG = −A−T δAT G = −G δAT G (4.87)
Further, using the definition of P in (4.57) gives
 
I
δP = −C δGT C= (4.88)
I

which can be rewritten as

δPT = −δG CT = G δAT G CT (4.89)
Then, the second term of (4.78) becomes
E δPT m = E G δAT G CT m (4.90)
which can be simplified after symbolic matrix multiplications as
E δPT m = E G a δln (4.91)
with
 
0
 η 1 
 − 

a =  ln (fa 2 + f a 5 ) (fa3 + fa6 )  (4.92)
ln 
 1 
(fa4 + fa7 )
ln

69
Introducing (4.47) in (4.91) gives

E δPT m = E G a r δpgg (4.93)

Finally, from equations (4.75), (4.78), (4.83) and (4.93) the expression of the global
tangent stiffness matrix is

Kg = BT
g Ka Bg + Km Km = D fa1 − E Q GT ET + E G a r (4.94)

Note that the tangent stiffness matrix as given by equation (4.94), is not symmet-
ric. However, as shown in [64], it can be symmetrised without losing quadratic
convergence unless concentrated moments are applied on the structure.

4.2.4 Eccentric nodes, rigid links

A beam element is defined such that node 1 is rigidly linked to another node I of
the structure. Mathematically this can be expressed as

xgI = xg1 + vo (4.95)

where xg1 and xgI denote the position vectors of nodes 1 and I in the original un-
deformed configuration whereas vo denotes an eccentricity vector. Since the link is
assumed as rigid, the magnitude of the eccentricity vector will not change during
the deformation of the structure. The connection between the displacements at the
two nodes 1 and I can then be expressed as

ugI = ug1 + (Rg1 − I) vo Rg1 = RgI (4.96)

The modelling of a rigid link will thus require a change of variables from ug1 to ugI .
Taking the variations of equation (4.96), and using (4.81), gives
g
δugI = δug1 + δRg1 vo = δug1 + δ θ̃ 1 Rg1 vo = δug1 − ṽ δθ 1g (4.97)

with

v = Rg1 vo (4.98)

which, by introducing the notation

 T
g
δpec = δugI T δθ 1g T δug2 T δθ 2g T (4.99)

leads to
 
I ṽ 0 0
 
 0 I 0 0 
δpgg = g
Bec δpec Bec =   (4.100)
 0 0 I 0 
0 0 0 I

70
with δpgg as defined in equation (4.35). The internal force vector fec and tangent
g
stiffness matrix Kec consistent with δpec are then given by
fec = BT
ec fg Kec = BT
ec Kg Bec + Kn (4.101)
The matrix Kn is calculated by
     
03,1 03,1 03,1
 −δṽ n   ñ δ θ̃ g v   −ñ ṽ δθ 1g 
g    1   
Kn δpec = δ(BTec ) fg =  = = 
 03,1   03,1   03,1 
03,1 03,1 03,1
(4.102)
 
0 0 0 0
 0 −ñ ṽ 
 0 0  g
=  δpec
 0 0 0 0 
0 0 0 0

with n = [ fg1 fg2 fg3 ]T , fgi i = 1, 12 denotes the i-th component of the vector fg .

4.2.5 Finite rotation parameters

The global internal force vector fg and tangent stiffness matrix Kg computed ac-
cording to equations (4.74) and (4.94) are consistent with spatial angular variations
δθ g of the type defined in the first of equations (4.8). Consequently, for a certain
node of the structure, the updating of the rotation matrix at each iteration step
would have to be done according to
g
Rgn = exp(δ θ̃ ) Rgo (4.103)
where the subscripts n and o denote quantities corresponding to the “new” and
“old” configurations, respectively. Note that in this case the parameterisation of Rg
is largely independent on the technique used to parameterise its admissible varia-
tions. Using the terminology in [51], this alternative will be denoted as “intrinsic”
parameterisation of 3D rotations.
In addition to the “intrinsic” parameterisation, two alternatives are explored in the
present work. The first one is based on the rotation vector Ψ g , i.e. a 3-parameter
representation, cf. (4.3). In this case, the rotational variables are additive and the
necessity of a special updating procedure is avoided. The second alternative is based
on the spatial form of the incremental rotation vector Θ g , cf. (4.5) [51,68]. The rota-
tional variables are still additive, but this time only within an increment at the level
of the iterative corrections. Corresponding to these two choices, additional changes
of variables from δθ g to δΨ g or alternatively from δθ g to δΘ g , must be introduced.
These changes of variables are defined in the remainder of this subsection.
It should be noted here that in the context of multi-parametric analyses of instability
problems, additive variables, at least at the level of iterative corrections, are essential

71
for an efficient implementation of fold line evaluation algorithms [68, 71]. In the
context of dynamical problems and Newmark time-stepping schemes, an excellent
discussion on the relative merits of the different alternatives is presented in [51].

Change of variables: δθ g −→ δΨ g

For the purposes of this subsection, let prg denote the following vector of global
nodal displacements
 T
prg = ug1 T Ψ1g T ug2 T Ψ2g T (4.104)

where Ψig denotes the rotational vector at node i. The change to the new kinematic
variables in prg requires the connection between δpgg as defined in (4.35) and δprg .
This connection can be easily constructed using (4.11)
 
I 0 0 0
 
 0 Ts (Ψ1g ) 0 0 
δpgg = Br δprg Br = 
 0

 (4.105)
 0 I 0 
0 0 0 Ts (Ψ2g )

The global internal force vector fr and tangent stiffness matrix Kr , consistent with
prg , are then given by

fr = BT
r fg Kr = BT
r Kg Br + Kv (4.106)

where
 
0 0 0 0
 
 0 Kv1 0 0 
Kv =   (4.107)
 0 0 0 0 
0 0 0 Kv2

The expressions of Kv1 and Kv2 are obtained from

  2   2
∂ sin ψ sin(ψ/2) 1 sin(ψ/2)
[T v] = −
T
− (u × v) u +
T
ṽ
∂Ψ s ψ (ψ/2) 2 (ψ/2)
 
sin ψ 1  T 
+ cos ψ − v u − (uT v)u uT (4.108)
ψ ψ
 
sin ψ 1  T 
+ 1− u v − 2 (uT v)u uT + (uT v) I
ψ ψ

with the vector v maintained constant during the differentiation.

Thus, Kv1 and Kv2 are evaluated from equation (4.108) with Ψ = Ψ1g and Ψ = Ψ2g ,
respectively and v = [ fg4 fg5 fg6 ]T , v = [ fg10 fg11 fg12 ]T , respectively.

72
With this new parameterisation, the rotational variables are additive and the itera-
tive updates are performed according to

Ψng = Ψog + δΨ g Rgn = exp(Ψng ) (4.109)

The main drawback of a parameterisation based on the rotation vector lies in the
fact that it cannot be retained globally. As implied by equation (4.13), the linear
operator Ts becomes singular if the magnitude of the rotation vector reaches 2π.
This undesirable feature is then passed over to the tangent stiffness matrix computed
according to equation (4.106). In order to remove this drawback, a parameterisa-
tion based on the incremental rotation vector Θ g and its iterative increment δΘ g ,
cf. subsection 4.1.2, will be discussed in the next subsection.

Change of variables: δθ g −→ δΘ g

The transformation matrices required in this second case are constructed on the basis
of equation (4.18) which provides the connection between δθ g and δΘ g . Following
a procedure entirely similar to that used in the preceding subsection, the vector prg
is now defined as
 T
prg = ug1 T Θ1g T ug2 T Θ2g T (4.110)

where Θig denotes the spatial form of the incremental rotation vector at node i.
Note that in order to avoid singularities in the expression of T s , the magnitude
of Θ g should be restricted to values less than 2π. Since Θ g now represents an
incremental rotation between two consecutive equilibrium states, this restriction is
of no practical consequence.
The transformation equations for the internal force vector and the tangent stiffness
matrix are formally identical to equations (4.105)–(4.108), by replacing Ψig by Θig .
Using the incremental rotation vector, additive updates apply only within an in-
crement at the level of the iterative corrections. Thus, the updating of the nodal
rotations is performed according to

Θng = Θog + δΘ g Rgn = exp(Θng ) Rgn−1 (4.111)

where Rgn−1 is the rotational matrix obtained at the end of the previous increment.
The transformation matrices required for this second type of parameterisation are
essentially constructed using the same linear operators as before, which considerably
simplifies the implementation.

4.2.6 Formulation with warping

In order to include warping effects, additional (warping) degrees of freedom αi are

introduced at both nodes i = 1, 2 of the element. The local displacement vector will

73
thus become
 T
T T
pl∗ = ū ϑ̄1 ϑ̄2 α1 α2 (4.112)

However, since the warping is in itself a deformational quantity, these additional de-
grees of freedom remain constant during the sequence of transformations previously
defined. Hence, referring for instance to equation (4.59), the variations δpa and δpgg
are rewritten as
 T
∗ T T
δpa = δū δ θ̄ 1 δ θ̄ 2 δα1 δα2 (4.113a)
 T
δpgg ∗ = δug1 T δθ 1g T δug2 T δθ 2g T δα1 δα2 (4.113b)

and the corresponding connection is defined by


Bg 02
δpa∗ = Bg∗ δpgg ∗ Bg∗ = (4.114)
02 I2

with Bg as given in (4.59) and I2 denoting a 2×2 identity matrix. Including warping
effects, the global internal force vector and tangent stiffness matrix are given by


∗ ∗T ∗ ∗ ∗T ∗ ∗ ∗ ∗ Km 02
fg = Bg fa Kg = Bg Ka Bg + Km Km = (4.115)
02 02

with Km as defined in equation (4.94). The two other transformations are modified
in a similar way.

4.3 Local element formulation in elasticity

The purpose of this section is to define the internal force vector fl (fl∗ if warping
effects are considered) and tangent stiffness matrix Kl (Kl∗ ) in local coordinates.
Assuming elastic material behaviour, fl and Kl can be derived from the strain energy
Φ expressed as a function of the local displacements pl (or alternatively pl∗ ), through
successive differentiations
∂Φ ∂2Φ
fl = Kl = (4.116)
∂pl ∂p2l

For all element types presented in this section, the differentiations involved in equa-
tion (4.116) have been performed using the Maple symbolic software package. For
this reason, explicit expressions for fl and Kl will not be provided here. The remain-
der of this section will instead focus on the strain energy expression which is the
keystone in constructing the necessary Maple subroutines. To complete the element
definition, the Maple codes are given in Appendix A.
In order to provide an explicit expression for Φ(pl ), a certain definition of the local
strains must be adopted. As the main interest of the co-rotational formulation lies

74
in the fact that the geometrical non-linearity is embedded in the motion of the local
element frame, the local strains can be expressed in a simple manner. In fact, most
of the co-rotational elements found in the literature are based on local linear strain
assumptions. However, as the numerical tests presented in this thesis will show,
such elements can give inaccurate results for problems where the torsion effects are
important. For these reasons, in the present work a second order approximation of
the Green-Lagrange strains is constructed at the level of the local element frame.
The resulting expressions are subsequently simplified by neglecting certain second
order terms. The effects of these simplifications are then carefully examined in the
numerical examples.

4.3.1 Local beam kinematics, strain energy

As a general rule, all kinematic quantities introduced in this subsection are referred
to the local element frame ri , i = 1, 2, 3, as defined in equations (4.25), (4.27)
and (4.28). The origin of the local system is taken at node 1 with r1 directed along
the line of centroids. Note however that r2 and r3 are not necessarily directed along
the principal axes of the cross-section.
With respect to the local system, the Green-Lagrange strain components which
contribute to the strain energy of the beam are given by
1 2 1 1
ε11 = ū1,1 + ū1,1 + ū22,1 + ū23,1
2 2 2
2 ε12 = ū1,2 + ū2,1 + ū1,1 ū1,2 + ū2,1 ū2,2 + ū3,1 ū3,2 (4.117)
2 ε13 = ū1,3 + ū3,1 + ū1,1 ū1,3 + ū2,1 ū2,3 + ū3,1 ū3,3

where ū1 , ū2 , ū3 are the local (deformational) displacements of the current point P .
In the above equation, a comma followed by an index denotes differentiation with
respect to the corresponding variable.
In order to obtain a consistent second order approximation of the strain expressions
in equation (4.117), a second order approximation of the displacement field is needed.
Within the settings of classical beam theory, cf. e.g. [4], bending is defined around the
centroid G while torsion is referred to the shear center C. Consequently, whenever
C and G are not coincident, the derivation of a non-linear displacement field is not
an easy task since rotations are not defined around the same point. An attempt
can be found in [60], but the approach presented there lacks consistency. In fact,
with the classical beam theory, only a linearised displacement field, obtained by
superposition of linear stretching, bending and torsional effects can be derived.
In order to avoid such problems, the kinematic model proposed by Gruttmann et
al. [41] is adopted, see Figure 4.3. Thus, let xoP (x1 , x2 , x3 ) denote the position vector
of point P in the initial (i.e. rotated but still undeformed) configuration and let
xP (x1 , x2 , x3 ) denote the position vector of P in the current configuration. These

75
 x3

a3
a2 C
c3
r2 x2
a1
x1 c2 a2
a3 R̄ ū G
r3 r1
x1

Figure 4.3: Local beam configuration.

two vectors are given by

xoP (x1 , x2 , x3 ) = xoG (x1 ) + x2 r2 + x3 r3 (4.118)
xP (x1 , x2 , x3 ) = xG (x1 ) + x2 a2 (x1 ) + x3 a3 (x1 ) + α(x1 ) ω̄(x2 , x3 ) a1 (x1 )
with xoG and xG denoting the position vectors of G in the initial and current configu-
rations, respectively. The warping function ω̄(x2 , x3 ) is defined within Saint-Venant
torsion theory and refers to the centroid G, i.e.
ω̄ = ω − c2 x3 + c3 x2 (4.119)
where ω refers to the shear center C of coordinates c2 , c3 . Note that in connection
with ω, the following normality conditions hold
  
ω dA = 0 ω x2 dA = 0 ω x3 dA = 0 (4.120)
A A A
The orhonormal triad ai , i = 1, 2, 3 which specifies the orientation of the current
cross-section, is given by
ai = R̄ ri i = 1, 2, 3 (4.121)
where the orthogonal matrix R̄ defines a rotation relative to the local element axes,
see (4.30). Using (4.4), a second order approximation of R̄ can be constructed as
 
0 −ϑ3 ϑ2
R̄ = I + ϑ̄˜ + 1 ϑ̄
˜2 ˜ = ϑ
ϑ̄ 0 −ϑ1 

 3 (4.122)
2
−ϑ2 ϑ1 0
Introducing equation (4.122) into (4.118), a second order approximation of the dis-
placement vector ūP = [ ū1 ū2 ū3 ]T = xP − xoP can be evaluated as
1 1
ū1 = u1 − x2 ϑ3 + x3 ϑ2 + x2 ϑ1 ϑ2 + x3 ϑ1 ϑ3 + ω̄ α
2 2
1 1
ū2 = u2 − x3 ϑ1 − x2 (ϑ12 + ϑ32 ) + x3 ϑ2 ϑ3 + ω̄ α ϑ3 (4.123)
2 2
1 1
ū3 = u3 + x2 ϑ1 − x3 (ϑ12 + ϑ22 ) + x2 ϑ2 ϑ3 − ω̄ α ϑ2
2 2

76
where u1 , u2 , u3 are the displacements of the centroid G, i.e. the components of the
vector xG − xoG .
Using equations (4.119), (4.120) and (4.123), a second order approximation of the
Green-Lagrange strains defined in (4.117) can be evaluated. At this level, two ad-
ditional simplifications are introduced. First, the term 1/2 ū21,1 in the expression of
ε11 is neglected. Second, the non-linear strain components generated by warping are
omitted since warping effects are rationally taken into account in a linearised way
only [42]. Using these simplifications gives
1 2 2
ε11 = εc + x2 k2 + x3 k3 + r ϑ1,1 + ω α,1
2
2 ε12 = γ12 + ω̄,2 α − x3 k1 (4.124)
2 ε13 = γ13 + ω̄,3 α + x2 k1

with

r2 = x22 + x23
1
2 
εc = u1,1 + u2,1 + u23,1
2
1

k2 = −ϑ3,1 + ϑ1,1 ϑ2 + ϑ1 ϑ2,1 + u3,1 ϑ1,1 + c3 α,1
2
1

k3 = ϑ2,1 + ϑ1,1 ϑ3 + ϑ1 ϑ3,1 − u2,1 ϑ1,1 − c2 α,1 (4.125)
2
1

k1 = ϑ1,1 + ϑ2,1 ϑ3 − ϑ2 ϑ3,1
2
1
γ12 = u2,1 − ϑ3 + ϑ1 ϑ2 + u3,1 ϑ1 − u1,1 ϑ3
2
1
γ13 = u3,1 + ϑ2 + ϑ1 ϑ3 − u2,1 ϑ1 + u1,1 ϑ2
2
The strain energy can then be evaluated from
     
1 2 1  2 2
Φ= ΦA dx1 = E ε11 dA + G (2ε12 ) + (2ε13 ) dA dx1 (4.126)
lo lo 2 A 2 A

Integration through the cross-section gives


2 1
ε11 dA = A ε2c + I22 k22 + I33 k32 + Irr ϑ1,14
+ Iω α,21 +Io εc ϑ1,1
2
(4.127)
A 4
2 2 2
+ 2 I23 k2 k3 + I2r k2 ϑ1,1 + I3r k3 ϑ1,1 + Iωr α,1 ϑ1,1


 
(2ε12 )2 + (2ε13 )2 dA = A (γ12
2 2
+ γ13 ) + Io k12 + (Io − J) α2 (4.128)
A
− 2 (Io − J) α k1

where J denotes the Saint-Venant torsion modulus


J= [x2 (ω̄,3 +x2 ) − x3 (ω̄,2 −x3 )] dA (4.129)
A

77
and the other section quantities are defined by
  
2 2
I22 = x2 dA I33 = x3 dA I23 = x2 x3 dA
A A A
  
2 2
I2r = x2 r dA I3r = x3 r dA Irr = (x22 + x23 )2 dA (4.130)
A A A
Iω = ω 2 dA Iωr = ω r2 dA Io = (x22 + x23 ) dA
A A A

In deriving equations (4.128) and (4.129), the following relations have been used
(cf. the proof in [41])
 
ω̄,2 dA = ω̄,3 dA = 0 (4.131)
A A
 


ω̄,22 +ω̄,23 dA = − (ω̄,3 x2 − ω̄,2 x3 ) dA = Io − J (4.132)
A A

Bernoulli assumption
The bending shear strains γ12 and γ13 in equation (4.125) are neglected, which gives
1 1
ϑ2 = −u3,1 + u ϑ + u1,1 u3,1 ϑ3 = u2,1 + u ϑ − u1,1 u2,1 (4.133)
2 2,1 1 2 3,1 1
By introducing the above equations into (4.125), the bending curvatures and tor-
sional twist can be rewritten as
1

k1 = ϑ1,1 + u3,1 u2,11 − u2,1 u3,11
2
k2 = −u2,11 + u1,11 u2,1 + u1,1 u2,11 − u3,11 ϑ1 + c3 α,1 (4.134)
k3 = −u3,11 + u1,11 u3,1 + u1,1 u3,11 + u2,11 ϑ1 − c2 α,1

Since the bending shear stresses are neglected, it appears logical to neglect the
bi-shear Fω also. Thus
∂ΦA
Fω = =0 (4.135)
∂α
where ΦA is defined in equation (4.126). Introducing (4.127) and (4.128) into the
above expression, gives

α = k1 (4.136)

The integral (4.128) is then simplified under the form


 
(2ε12 )2 + (2ε13 )2 dA = J k12 (4.137)
A

In order to avoid membrane locking, ε11 in equation (4.124) is rewritten as

 
1 2 Io
ε11 = εav + x2 k2 + x3 k3 + r − 2
ϑ1,1 + ω α,1 (4.138)
2 A

78
where the “average” value εav is given by

 
1 1 2 2 Io 2
εav = u + u2,1 + u3,1 + ϑ1,1 dx1 (4.139)
lo lo 1,1 2 A
This gives
  
1 Io2
2
ε11 dA = 2
A εav+ I22 k22+ I33 k32
+ I − 4
ϑ1,1 + Iω α,21
A 4 rr A
2 2 2
+ 2 I23 k2 k3 + I2r k2 ϑ1,1 + I3r k3 ϑ1,1 + Iωr α,1 ϑ1,1 (4.140)

4.3.2 Element types

Both Timoshenko and Bernoulli elements were developed in the present work. The
Timoshenko elements are based on equations (4.125), (4.127) and (4.128), whereas
the Bernoulli elements are based on (4.134), (4.137), (4.139) and (4.140). Both
element types have been implemented by neglecting the nonlinear terms in the ex-
pressions of the curvatures k1 , k2 , k3 and bending shear strains γ12 , γ13 . Extensive
numerical testing has shown that this simplification has no effect whatsoever on
the accuracy of the results; i.e. including these terms generates differences of less
than 0.01% in the final results. On the other hand, certain tests have shown that
the Wagner term 12 r2 ϑ1,1
2
in the expression of ε11 can not be neglected or even be
replaced by its average value ( 12 IAo ϑ1,1
2
) over the cross-section.

4.3.3 Beams with thin-walled or open cross-sections

For beams with thin-walled or open cross-sections, warping effects are introduced
through an additional degree of freedom at each node. The vector of local displace-
ments is then given by
 T
T
pl∗ = ū ϑ̄1T ϑ̄2T α1 α2 ϑ̄i = [ ϑ1 ϑ2 ϑ3 ]i i = 1, 2 (4.141)

The Timoshenko element tw3d, is implemented using linear interpolations for all
variables, including the warping parameter α. The integrals in the strain energy
expression are computed using a one point Gauss rule in order to avoid shear locking.
Note that a linear interpolation scheme gives u2 ≡ u3 ≡ 0 since these displacements
have zero values at the nodes of the beam due to the particular choice of the local
system. Moreover, u2,1 ≡ u3,1 ≡ 0 and the only non-linear term in the expression of
ε11 is the so-called Wagner term.
For the Bernoulli element bw3d, Vlasov’s assumption is adopted, i.e.

α,1 = ϑ1,11 (4.142)

which, since the non-linear terms in the expression of the curvatures are neglected,
is the direct consequence of equation (4.136). Linear interpolation is used for the

79
axial displacement u, whereas Hermitian interpolations are used for u2 and u3 and
for the axial rotation ϑ1 . Note that in this case, the shallow arch terms u22,1 and u23,1
in the expression of ε11 , are non-zero.

4.3.4 Beams with solid or closed cross-sections

For such beams, the vector of local displacements is given by

 T
T T
pl = ū ϑ̄1 ϑ̄2 (4.143)

The cross-section is assumed to be bi-symmetric, with r2 and r3 directed along the

principal axes of the cross-section (see Figure 4.3). These assumptions give

I23 = I2r = I3r = Iωr = c2 = c3 = 0 (4.144)

The warping stresses are neglected and consequently the bi-shear Fω and the bi-
moment Mω are set to zero
∂ΦA ∂ΦA
Fω = =0 Mω = =0 (4.145)
∂α ∂α,1

Introducing equations (4.127) and (4.128) in the above expressions, gives

α = k1 Iω α,1 = 0 (4.146)

which are further substituted in the corresponding expressions of the strain energy.
It should be noted here that if the cross-section is not bi-symmetric, the second of
equation (4.146) is not valid; a more complicated equation is obtained.
The Timoshenko element t3d is implemented using linear interpolations for all vari-
ables and a one point Gauss rule. The element t3dl is obtained by keeping only
linear terms in the strain expressions. This element is the classical two-node linear
Timoshenko element.
The Bernoulli element b3d is implemented using linear interpolations for the axial
displacement u and the axial rotation ϑ1 and Hermitian interpolations for u2 and u3 .
The classical linear Bernoulli element b3dl is obtained by considering only linear
terms in the strain expressions.

4.4 Local element formulation in elasto-plasticity

The purpose of this section is to define the internal force vector fl (fl∗ if warping
effects are considered) and tangent stiffness matrix Kl (Kl∗ ) in local coordinates in
case of elasto-plasticity. Assuming a von Mises material with isotropic hardening,
Timoshenko elements are developed. Due to the material non-linearity, numerical
integration over the cross-section is required, see the introduction of Chapter 3. The

80
constitutive equations are solved at each integration point by including interaction
between the normal and shear stresses.
Assuming elasto-plastic material, fl and Kl can be derived from the virtual work
principle. If no volume or surface loads are acting on the element, this is written as

V = δεT · σ dv = δpT l fl (4.147)
v

4.4.1 Strain definition

The strain definition and displacement field are the same as in the elastic case, see
equations (4.117) and (4.118). However, in elasto-plasticity, the notion of shear
center does not have any meaning, and the warping function ω̄ is not replaced by
ω, see equation (4.119). As a matter of fact, due to the material non-linearity, the
normality conditions (4.120) do not introduce any simplifications, as it was the case
in elasticity. Equation (4.124) is then rewritten as
1 2 2
ε11 = εc + x2 k2 + x3 k3 + r ϑ1,1 + ω̄ α,1
2
2 ε12 = γ12 + ω̄,2 α − x3 k1 (4.148)
2 ε13 = γ13 + ω̄,3 α + x2 k1
with
r2 = x22 + x23
1
2 
εc = u1,1 + u2,1 + u23,1
2
1

k2 = −ϑ3,1 + ϑ1,1 ϑ2 + ϑ1 ϑ2,1 + u3,1 ϑ1,1
2
1

k3 = ϑ2,1 + ϑ1,1 ϑ3 + ϑ1 ϑ3,1 − u2,1 ϑ1,1 (4.149)
2
1

k1 = ϑ1,1 + ϑ2,1 ϑ3 − ϑ2 ϑ3,1
2
1
γ12 = u2,1 − ϑ3 + ϑ1 ϑ2 + u3,1 ϑ1 − u1,1 ϑ3
2
1
γ13 = u3,1 + ϑ2 + ϑ1 ϑ3 − u2,1 ϑ1 + u1,1 ϑ2
2
The warping function ω̄(x2 , x3 ) is defined within Saint-Venant torsion theory and
refers to the centroid G. For arbitrary cross-section, ω̄ is determined numerically in
a separate finite element analysis. The theoretical aspects of this computation are
given in Appendix B. In the numerical examples presented in this thesis, isopara-
metric quadratic plane elements with four Gauss points are used to discretise cross-
sections. It must be noted that within the Saint-Venant torsion theory, warping
effects are assumed to be small and are therefore taken into account in a linearised
way only [42]. Consequently, the warping function is determined by considering an
elastic material and linear deformations. Although frequently used, this assump-
tion is certainly debatable. However, even if the influence of plastic deformations

81
or finite displacements can be introduced by updating at each step the warping
function [2, 103], in practice, this operation requires too much computational time.

4.4.2 Finite element formulation

A finite element formulation can be easily derived from equations (4.148) and
(4.149). However, the numerical tests performed in the elastic range have shown
that these expressions can be simplified without affecting the results by neglecting
the non-linear terms in the expressions of the curvatures k1 , k2 , k3 and bending shear
strains γ12 , γ13 . With these simplifications, the strains can be rewritten as
1
2  1
ε11 = u1,1 + u2,1 + u23,1 − x2 ϑ3,1 + x3 ϑ2,1 + r2 ϑ21,1 + ω̄ α,1
2 2
2 ε12 = u2,1 − ϑ3 + ω̄,2 α − x3 ϑ1,1 (4.150)
2 ε13 = u3,1 + ϑ2 + ω̄,3 α + x2 ϑ1,1
which leads to a simple and computationally efficient local element.
The Bernoulli assumption α = ϑ1 , see section 4.3.1, can be introduced in equations
(4.150), leading to elements based on the same hypotheses as bw3d and b3d (see Sec-
tions 4.3.3 and 4.3.4). However, the formulation of such elements in elasto-plasticity
is computationally expensive and only Timoshenko elements will be considered in
the remainder of this section.

4.4.3 Beams with thin-walled or open cross-sections

Based on the same assumptions as the elastic element tw3d, see Section 4.3.3, a
Timoshenko element, called ptw3d is implemented. Thus, linear interpolations are
used for all variables, including the warping parameter α. Note that a linear inter-
polation scheme gives u2 ≡ u3 ≡ 0 (and consequently u2,1 ≡ u3,1 ≡ 0) since these
displacements have zero values at the nodes of the beam due to the particular choice
of the local system. Consequently, the only geometrical non-linear term in the strain
definition is the Wagner term 12 r2 ϑ1,1
2
.
Differentiation of equation (4.150) gives
 T
δε = A δ ε̂ ε̂ = u1,1 ϑ3 ϑ2 ϑ1,1 ϑ2,1 ϑ3,1 α α,1 (4.151)
with
 
1 0 0 r2 ϑ1,1 x3 −x2 0 ω̄
 
A=
 0 −1 0 −x3 0 0 ω̄,2 0 
 (4.152)
0 0 1 x2 0 0 ω̄,3 0
The virtual internal work in equation (4.147) can be rewritten as
 
V = δε · σ dv =
T
δ ε̂T · σ̂ dx1 (4.153)
v lo

82
where lo is the initial length of the element and σ̂ is the vector of stress resultants
defined by

σ̂ = AT σ dA (4.154)
A

To avoid shear locking, the second integral in equation (4.153) is evaluated by one
Gauss point integration. This gives

δ ε̂T · σ̂ dx1 = δpl∗ T GT σ̂ (4.155)
lo

where pl∗ is given in (4.141) and G is defined by

δ ε̂ = G δpl∗ (4.156)

From the choice of the interpolations, the matrix G is given as

 
1/lo 0 0 0 0 0 0 0 0
 0 
 0 0 0 1/2 0 0 1/2 0 
 
 0 0 1/2 0 0 1/2 0 0 0 
 
 0 −1/lo 0 
 0 0 1/lo 0 0 0 
G= 
 0 0 −1/lo 0 0 1/lo 0 0 0 
 
 0 0 0 −1/lo 0 0 1/lo 0 0 
 
 
 0 0 0 0 0 0 0 1/2 1/2 
0 0 0 0 0 0 0 −1/lo 1/lo
(4.157)

The vector of internal forces fl∗ in local coordinates is then given by

fl∗ = GT σ̂ (4.158)

where σ̂ is the vector of stress resultants for the section at the middle of the element.
The local tangent stiffness matrix defined by δfl∗ = Kl∗ δpl∗ can be calculated by
differentiation of equation (4.158), which gives

Kl∗ = GT D̂ G δ σ̂ = D̂ δ ε̂ (4.159)

The matrix D̂ can be obtained by taking the variations of (4.154)

 
 T   T 
δ σ̂ = A δσ + δ(A σ) dA =
T
A Cct A + L dA δ ε̂ (4.160)
A A

where Cct is the consistent tangent operator defined by δσ = Cct δε and L is ob-
tained from δ(AT σ) with σ maintained constant during the differentiation. Hence,
the only non zero term in the matrix L is

L4,4 = r2 σ11 (4.161)

83
Finally, σ̂ and D̂ are calculated trough a numerical integration over the cross-section,
by using the same mesh as for the determination of the warping function. Hence, the
cross-section is discretised using isoparametric quadratic elements and integration
over each of these elements is performed using 4 Gauss points. At each point, the
constitutive equations are solved iteratively using a backward Euler scheme [23],
which allows the generation of a consistent tangent operator Cct .
The constitutive equations for a 3D beam are derived from the von Mises equations
in three dimensions [23] by setting
σ22 = σ33 = σ23 = 0 (4.162)
By performing a similar work as in Section 3.4.4 for the 2D beam case, it can
be shown that the Backward-Euler algorithm and consistent tangent operator are
similar to those given in Figure 3.4 and equation (3.93) with
     
ε11 σ11 E 0 0
     
ε=  2 ε 12

 σ =  σ12 
  C =  0 G 0 
  (4.163)
2 ε13 σ13 0 0 G

2 
2 1/2
f = σ11 2
+ 3 σ12 + 3 σ13 − σo (εps ) = σe − σo (εps ) (4.164)
   
σ11 2
σ12 2
+ σ13 −σ11 σ12 −σ11 σ13
1  
 3 σ12  ∂a 3  
a= = 3 −σ11 σ12 2
σ11 2
+ 3 σ13 −3 σ12 σ13  (4.165)
σe   ∂σ σe  
3 σ13 −σ11 σ13 −3 σ12 σ13 2 2
σ11 + 3 σ12

4.4.4 Beams with solid or closed cross-section

For beams with solid or closed cross-sections, warping stresses can be neglected.
Assuming a bi-symmetric cross-section, it has been shown in Section 4.3.4 that the
local finite element formulation is obtained by introducing in equations (4.150) to
(4.157) the constraints
α = ϑ1,1 α,1 = 0 (4.166)
Hence, ε̂ and A, see equations (4.151) and (4.152), are rewritten as
 T
ε̂ = u1,1 ϑ3 ϑ2 ϑ1,1 ϑ2,1 ϑ3,1 (4.167)
and
 
1 0 0 r2 ϑ1,1 x3 −x2
 
A=
 0 −1 0 ω̄,2 −x3 0 0 
 (4.168)
0 0 1 ω̄,3 +x2 0 0
The internal force vector fl and tangent stiffness matrix Kl are then obtained by a
procedure similar to the one described in Section 4.4.3. This element is called pt3d.

84
4.5 Applied loads

The aim of the present section is to discuss two aspects related to the applied external
loads. The first one refers to the modelling of eccentrically applied forces. The sec-
ond refers to the applied external moments, in the context of the re-parameterisation
of finite rotations advocated in Section 4.2.5.

4.5.1 Eccentric forces

Assume that the vector of external forces fex I is applied at node I, and that node I
is rigidly connected to node 1 through the eccentricity vector vo , defined by
xgI = xg1 + vo ugI = ug1 + (Rg1 − I) vo Rg1 = RgI (4.169)
where xg1 and xgI denote the position vectors of nodes 1 and I in the original unde-
formed configuration.
The modelling of such a situation can be performed in two ways. The first one
consists of a change of variables from ug1 to ugI , as developed in Section 4.2.4. The
second one requires a transformation of fex I into fex 1 , which can be obtained by
considering that both forces must perform the same virtual external work
δpIg T fex I = δp1g T fex 1 (4.170)
with
 T  T
δpIg = δugI T δθ 1g T δp1g = δug1 T δθ 1g T (4.171)

Similar calculations as in Section 4.2.4 give

 
I −ṽ
fex 1 = BT
ef fex I Bef = v = R1g vo (4.172)
0 I

In addition, differentiation of equation (4.170) gives the stiffness correction

   
0 0 nI
Kef = fex I = (4.173)
0 ñI ṽ mI
which must be subtracted from the tangent stiffness matrix of the structure. It
should also be noted here that, equations (4.172) and (4.173) are consistent with
spatial angular variations. A different choice of finite rotation parameters will then
require a further transformation of these equations, cf. Section 4.2.5.

4.5.2 External moments

Consider that at node i, an external moment mθ performs the virtual work δθ ig T mθ ,

with δθ ig representing spatial angular variations. This means that if a parameterisa-
tions in terms of the total rotational vector Ψig is adopted, mθ should be accordingly

85
transformed into mΨ in order to represent the same physical problem. By consider-
ing that both moments must perform the same virtual external work, it is obtained

δθ ig T mθ = δΨig T mΨ (4.174)

which, by using equation (4.11) leads to

mΨ = TT
s (Ψ) mθ (4.175)

Differentiating the above equation gives the stiffness correction

∂  T 
KΨ = Ts (Ψ) mθ (4.176)
∂Ψ
which must be subtracted from the tangent stiffness matrix. The expression of KΨ is
given in equation (4.108). If a parameterisation in terms of the incremental rotation
vector is adopted, the required transformations are obtained by replacing Ψig with
Θig in equations (4.175) and (4.176). These aspects will be further discussed in
connection with one of the numerical examples where it is shown that if the above
transformations are not applied, the three parameterisations discussed in the paper
will produce totally different results. Note however that, if the moment is applied at
a point which is constrained to rotate around a fixed axis, Ts reduces to the identity
matrix and the above transformations are no longer necessary.

86
Chapter 5

Path following techniques

Concerning non-critical paths, the Newton-Raphson method, preceded by an Euler

forward predictor is chosen to solve the equilibrium equations at each step. In
most of commercial codes, the length of the step is defined by adding an arc-length
equation (see e.g. [23]) to the equilibrium equations. In this thesis, a displacement
or load control strategy has been adopted. The procedure for computing non-critical
paths is described at both structural and element levels in Section 5.1.
Since the purpose of the elements described in this thesis is the study of elastic and
elasto-plastic instability phenomena, special procedures are needed in order to first
detect and isolate bifurcation points along fundamental equilibrium paths and then
to perform branch-switching to secondary paths.
This approach is different to the one usually adopted in commercial finite element
codes. In these programs, the bifurcation loads are often evaluated by a linearised
buckling analysis based on the equation

(K1 + ω ∆K) φcr = 0 ∆K = K2 − K1 (5.1)

where K1 and K2 are the tangent stiffness matrices of two known equilibrium points
on the fundamental path of the perfect structure. The buckling load factor λcr is
then calculated by

λcr = λ1 + ω ∆λ ∆λ = λ2 − λ1 (5.2)

where λ1 , λ2 are the load factors of the two known equilibrium points.
This approach assumes a linear relation between the increment of the geometric
stiffness matrix (assumed to be ∆K) and the increment of the load factor ∆λ. This
assumption is satisfied when the pre-buckling displacements are sufficiently small so
that the shape of the stress distribution remains unchanged and only the stress levels
change with the factor λ. If this condition is not respected, a linearised buckling
analysis gives more or less inaccurate results.
As regard the post-bifurcation behaviour, this is normally investigated in commercial
codes by introducing small imperfections in the geometry or the loading of the
structure. However, apart from the difficulties related to the choice of the form and

87
the magnitude of these imperfections, such an approach does not allow a complete
understanding of the instability phenomenon.
Thus, the main purpose of the procedures developed in this chapter is to isolate
bifurcation points and to study post-critical behaviour of the perfect structure,
i.e. without introducing any imperfections.
In elasticity, branch-switching is performed classically, using the mode injection
method. This procedure consists of taking as predictor for the secondary path the
eigenvector associated to the zero eigenvalue of the tangent stiffness matrix at the
bifurcation point. The main innovation of this thesis resides in a new numerical
approach for the direct computation of critical points. This procedure, which can
also handle limit points, is explained in detail in Section 5.2.
In elasto-plasticity, numerical analyses of buckling problems are not so common in
the literature. This is probably due to the fact that the instability phenomenon in
itself is more complicated. Taking as starting point the theoretical bases presented
in Chapter 2, the purpose of Section 5.3 is to describe and discuss the advantages
and drawbacks of two different branch-switching procedures.

5.1 Non-critical equilibrium path

5.1.1 Procedure at the structural level

The discretised structure has n degrees of freedom. The values of the n components
of the displacement vector d and the load factor λ have to be calculated.
The n non-linear equilibrium equations can be written as

f (d, H) − fex (λ) = 0 (5.3)

where f are the internal forces. In case of elasto-plasticity, f depends on the history
of the deformations, which is represented by the parameter H. The external load
vector fex is assumed to be conservative and proportional (dead loading) and can be
written in the form

fex (λ) = λ p + p p (5.4)

where p and p p are two constant vectors. p represents the main load while p p stands
for a small constant perturbation load which can be applied in order to introduce
imperfections.
Under load control, a load increment ∆λ̄ = λ − λ̄ is applied at each step and the
auxiliary equation

g (λ) = λ − λ̄ = 0 (5.5)

88
is added to the system (5.3). Under displacement control, a displacement increment
∆d¯i = di − d¯i is applied at each step and the auxiliary equation becomes

g (d) = di − d¯i = 0 (5.6)

which implies that the value of the i th displacement di is constrained to d¯i .

In order to simplify the notations, the auxiliary equations (5.5) and (5.6) are ex-
pressed in the same form as

g (d, λ) = 0 (5.7)

The system of equations (5.3) and (5.7) is solved by Newton-Raphson iterations

preceded by an Euler forward step.
The differential matrix required in this procedure is given by
    
K −p ∆d f (d, H) − fex (λ)
=− (5.8)
g,d g,λ ∆λ g (d, λ)

where K is the tangent stiffness matrix.

The vector t tangent to the path at a non-critical equilibrium point is given by
 
K−1 p
t= (5.9)
1

The Euler forward predictor is defined under displacement control by

 
∆d ∆d¯i
= t (5.10)
∆λ ti

and under load control by

 
∆d
= ∆λ̄ t (5.11)
∆λ

Remarks

• In order to avoid numerical problems during the iteration process, the magni-
tude of the vector p is chosen such that ∆λ has the same order of magnitude
as the ∆di , i = 1..n.

• To calculate the whole equilibrium path, it is often necessary to switch between

different controls (e.g. in case of snap-throughs or snap-backs). An automatic
solution is to take as control parameter in each step the largest component of
the tangent vector t.

89
5.1.2 Procedures at the element level

The purpose of this section is to summarise the procedure for computing at the level
of the element the internal force vector fg and tangent stiffness matrix Kg in global
coordinates. These procedures have been described in detail in Chapters 3 and 4.

Elastic case

1. The local displacements pl are extracted from the global ones pg .

2. The local internal force vector fl and tangent stiffness matrix Kl are functions
of the local displacements pl and are computed using the Maple subroutines
given in Appendix A.

3. Finally, the global internal force vector fg and tangent stiffness matrix Kg are
obtained using the transformation matrices.

Elasto-plastic case

Steps 1 and 3 are identical in elasticity and elasto-plasticity. As a matter of fact,

these steps depend only on the definition of the co-rotational framework and are
independent on the deformations. The difference between the two material cases
lies in the computation of fl and Kl . In elasto-plasticity, due to the non-linear
relationship between the strains and the stresses, fl and Kl cannot be calculated
only from the current local displacements pl . A more complicated procedure, taking
into account the history of the deformation process, has to be implemented.
In the following the subscript o refers to old known values at the last equilibrium
state while the subscript n refers to new values at the current iteration.
Step 2.

• The current strains εn are calculated at each integration point from the local
displacements pl . In order to avoid spurious numerical unloading, an incre-
mental strategy is adopted. This means that the strains increment ∆ε at each
integration point is defined by reference to the last converged equilibrium state
and is calculated using the strains at the beginning of the increment εo by

∆ε = εn − εo (5.12)

• The current stresses σ n , equivalent plastic strain εpsn and consistent tangent
operator Cct (Ê for 2D Bernoulli elements) are calculated at each integration
point by solving the constitutive equations

(σ n , εpsn , Cct ) = f (σ o , εpso , ∆ε) (5.13)

• fl and Kl are then computed through numerical integration over the element.

90
Remark
This procedure requires to save at each integration point the stresses σ o , the strains
εo and the equivalent plastic strain εpso obtained at the last equilibrium state. How-
ever, in order to save memory space, it has been chosen to save the local displace-
ments pl at the beginning of the step and to recalculate at each iteration the strains
εo, which is a simple operation. This is particularly interesting for 3D beam elements
for which an important number of integration points are required.

5.2 Direct computation of elastic critical points

The simplest and most general method to isolate critical points along an equilib-
rium path is to perform successive bisections by monitoring the number of negative
eigenvalues of the tangent stiffness matrix at each equilibrium point. However, this
approach is not efficient since it requires the calculation of many intermediate equi-
librium points. A more efficient method can be developed on the basis of an extended
system obtained by augmenting the equilibrium equations with a criticality condi-
tion which forces the Newton-Raphson iterations to converge to the critical point.
Thus, once a critical point has been detected along an equilibrium path within a
load increment, its isolation is performed in only one step. Following the work of
Wriggers et al. [104, 105], a vectorial form of the criticality condition is the most
popular choice in this context.
The alternative presented in this section is based on two new ideas. First, following
Eriksson [32–34], the criticality condition is expressed by a scalar equation, thus
keeping the size of the extended system to a minimum. The main problem of the
direct computation of critical points based on extended systems is that convergence
is not assured in all cases, specially if the pre-buckling deformations are non-linear.
Then, in order to improve the convergence properties, a second modification is
introduced. This consists of combining iterations based on the extended system
with iterations based on equilibrium equations under load or displacement control.
The outline of this section is as follows. First, the classical method of Wriggers et
al. [104, 105] is briefly summarised. Next the alternative extended system and its
solution is presented. Finally, the whole procedure and its initialisation is described.

5.2.1 Classical approach

The purpose of this section is to briefly summarise the classical approach developed
by Wriggers et al. [104, 105].
Using the property that the tangent stiffness matrix K possesses a zero eigenvalue at
a critical point, the vectorial critical condition K φ = 0 is added to the equilibrium

91
equations. The system of non-linear equations becomes
 
f (d) − λ p
 
 K(d) φ  = 0 (5.14)
 
φ − 1
where f denotes the vector of internal forces, d the displacement vector, λ the load
factor, p the reference load vector and φ the critical eigenvector. Note, once again,
that in the above system and in all this chapter, the reference load vector p is
assumed constant (dead loading).
The solution to the extended system (5.14) is obtained through a full Newton-
Raphson procedure. The differential matrix required in this procedure is given by
    
K −p 0 ∆d f (d) − λ p
    
 (K φ),   ∆λ  = −  Kφ 
 d 0 K     (5.15)
0 0 φT / φ ∆φ φ − 1
The above systems contains 2 n + 1 equations and unknowns (n is the number of
active degrees of freedom). However, due to its particular form, it can be shown
(see e.g. [105]) that its solution can be performed by factorising only the tangent
stiffness matrix K. Consequently, compared to a classical iteration of the equilibrium
equations with n unknowns, only a small increase of computational time is needed.

5.2.2 Alternative extended system

The procedure for the direct computation of critical points based on an alternative
extended system is now presented. This approach, developed by Eriksson [32–34]
in the context of fold line evaluation algorithms, makes use of a scalar criticality
condition.

Extended system

The vectorial criticality condition K φ = 0 is replaced by the scalar equation

g(d) = φT K φ = 0 (5.16)
in which g represents the lowest (in absolute value) eigenvalue of the tangent stiffness
matrix K and φ its associated eigenvector. At a critical point, K is singular and
the condition g = 0 is obtained. The eigenvector φ is normalised according to
φT φ = 1 (5.17)
g and φ are determined in a separate procedure, based on the iterative sequence
 (i+1)   −1  
φ −K φ(i) 0
= (5.18)
g (i+1) φ(i) T 0 1

92
This system, which is obtained from equations (5.16) and (5.17), can be viewed as
one step of an inverse subspace iteration. Note that any random vector can be used
to initiate the process and convergence is usually obtained in very few iterations.
Using equation (5.16), the extended system can be written as
 
f (d) − λ p
=0 (5.19)
g(d)

The solution to the extended system (5.19) is obtained through a full Newton-
Raphson procedure. The differential matrix required in this procedure is given by
    
K −p ∆d f (d) − λ p
=− (5.20)
g,d 0 ∆λ g(d)

The vector g,d is evaluated numerically by using a finite difference formula (for
details see [32, 104])

φT
g,d ≈ [ K(d + γ φ) − K(d) ] (5.21)
γ
where γ is a numerical parameter taken as

γ = max di  × 10−6 (5.22)

1<i<n

with di denoting the i th component of d. Numerical tests have shown that the
procedure is not sensitive to the choice of the exponent coefficient in equation (5.22),
i.e. any values in the interval [10−8 10−3 ] can also be chosen.
In some works [52, 105], the directional derivative of the tangent stiffness matrix
in equation (5.21) is performed analytically. For the co-rotational elements, this
approach seems difficult to apply and would require additional developments.

Efficient numerical scheme

The numerical approach described in this section requires the solution of two sets of
equations: at each iteration of the extended system (5.20), the system (5.18) must
be solved in order to compute g. The solution procedures for these two systems are
given in Tables 5.1 and 5.2. As it can be seen, the only matrix which needs to be
factorised is K. Consequently, compared to a classical iteration involving only the
equilibrium equations, just a small increase in computational work is required. The
additional expense is due to one loop over the elements for the evaluation of g,d , one
back substitution step for the computation of ∆d2 and one or two back substitution
steps for the solution of (5.18). Hence, compared to the classical approach in [52,105],
the solution of the present system requires about the same computational time.

93
 Table 5.1: Solution of system (5.18).

Solve K ψ = φ(i)
1
Calculate new values g= φ(i+1) = ψ g
φ
(i) T
ψ

Table 5.2: Solution of system (5.20).

Solve K ∆d1 = p K ∆d2 = − (f (d) − λ p)

Calculate g,d within a loop over all elements

g + g,d ∆d2
Calculate new values ∆λ = − ∆d = ∆d1 ∆λ + ∆d2
g,d ∆d1

5.2.3 Modified algorithm

The main problem of the direct computation of critical points based on extended
systems is that convergence is not assured if the non-linearity between the starting
equilibrium point for the iterations and the critical one is too high. As a matter of
fact, during the iterative process, the matrix K does not correspond to an equilib-
rium point and it may not be reliable to use the properties of K in order to reach
the critical point.
In this context, the numerical tests performed by the author have shown that the
classical approach of Wriggers (Section 5.2.1) has better performances than the al-
ternative presented in Section 5.2.2. As a matter of fact, in the original formulation,
the critical condition is depending on both d and φ, which means that at each it-
eration both d and φ are iteratively updated. On the contrary, in the alternative
proposed in this thesis, the critical condition depends only on d and φ is recalcu-
lated at each iteration as the eigenvector associated to the lowest eigenvalue of K.
This constraint on φ, not present in the original approach, makes the procedure
more sensitive to the variations of K.
It has also been observed numerically that the first iteration of the extended sys-
tems (5.15) or (5.20), based on a K corresponding to an equilibrium point, often
gives very good approximations to the bifurcation load factors. On the contrary,
the following iterations, based on a K which does not correspond to an equilibrium
point, give less accurate approximations to the bifurcation load factors, specially in
case of high non-linearity.
Following these considerations, the algorithm presented in the previous section has
been modified in order to improve its convergence properties. For that, a distinction

94
between bifurcation and limit points is required.

Bifurcation points

The idea of this new approach, whose algorithm is given in Table 5.3, is to evaluate at
the end of each iteration gnew , the absolutely lowest eigenvalue of the tangent stiffness
matrix. If gnew  is lower than the previous value gold  or if the current point is
in equilibrium, then the next iteration will use the extended system (5.20). In the
opposite case, the information given by the critical condition may not be reliable.
It is then better to perform an iteration using the equilibrium equations under load
control in order to obtain a point nearer an equilibrium state, and consequently, a
more reliable K matrix for expressing the critical condition.
Another effect of this modification is that some iterations performed using the ex-
tended system (5.20) are replaced by pure equilibrium iterations which require less
computational work.

Limit points

The procedure described in Table 5.3 is not suitable for the isolation of limit points
since the equilibrium iterations based on a load control may fail in the neighbourhood
of a limit point. The solution, see Table 5.4, is then to perform equilibrium iterations
under displacement instead of load control. The degree of freedom corresponding to
the largest component of the tangent vector (computed by K−1 p) at the starting
equilibrium point for the procedure is taken as control.
This method can also be used for the isolation of bifurcation points. However, the
numerical tests performed by the author have shown that the convergence properties
of the procedure are better if a load control is adopted. It can also be noted that
the author has also tested different arc-length procedures as substitution of the load
and displacement controls, without obtaining better results for the tested examples.
It appears then judicious, but not absolutely necessary, to have two different isolation
procedures, one for bifurcation points and one for limit points. In practice, this is
scarcely a problem since such a distinction is also required at the level of the initial
values of the procedure. This issue will be addressed in the next subsection.

5.2.4 Initialisation

Along a solution branch, the detection of critical points is performed by monitoring

at each converged equilibrium state the number of negative eigenvalues of K. This
can e.g. be performed efficiently by a LDLT -factorisation. Once a critical point has
been detected within an increment, it is isolated by taking as starting point the left
or right bound of the increment. For that point, initial values go and φo are needed.

95
 Table 5.3: Modified algorithm – bifurcation points.

initial values : d λ K go φo

first iteration : solve (5.20) → d = d + ∆d λ = λ + ∆λ

assemble K and r = f (d) − λ p

solve (5.18) → gnew φnew

gold = go φold = φo

loop : while gnew  > tol or r > tol

if gnew  < gold  or r < tol

gold = gnew φold = φnew

solve (5.20) → d = d + ∆d λ = λ + ∆λ

else

solve K ∆d = −r → d = d + ∆d

end

assemble K and r = f (d) − λ p

solve (5.18) → gnew φnew

end

Table 5.4: Partial modified algorithm – limit points.

else  
 K ∆d − ∆λ p = −r  d = d + ∆d
solve →
 
∆di = 0 λ = λ + ∆λ
end

If a bifurcation point is expected, go is given by the LDLT -factorisation and is taken

as the lowest negative or positive eigenvalue, depending on the choice of the starting
point. Usually, only one critical point is situated in the neighbourhood of the initial
point and go is the lowest absolute eigenvalue. φo , which is the eigenvector associated

96
to go , can then be computed using the system (5.18). For that, any random vector
can be used to initiate the iterations and convergence is usually obtained in two or
three iterations, due to the favourable spectral properties of K close to the critical
point. However, if several critical points are situated in the neighbourhood of the
initial point, go may not be the lowest absolute eigenvalue, and φo must be calculated
using a more elaborate method.
If a limit point is expected, the lowest negative or positive eigenvalue and its associ-
ated eigenvector may not be a judicious choice for go and φo , especially if the initial
point is far from the critical one. A better choice is to take for φo the tangent vector
to the equilibrium path, which gives
φo = K−1 p go = φT
o K φo (5.23)

5.3 Branch-switching in elasto-plasticity

The physical and theoretical aspects of plastic instability phenomena have been
discussed in Chapter 2 (particularly in Section 2.4) and can be summarised as follow:

• In elasto-plastic situations, a continuous range of bifurcation points can be

obtained and the notion of stability of an equilibrium point normally used in
elasticity must be replaced by the notion of stability of a deformation path.
• Only the secondary path branching out at the lowest bifurcation point has a
physical meaning and needs to be computed.
• At the lowest bifurcation point, the tangent stiffness matrix K ceases to be
positive definite and becomes indefinite.

Using these properties, this section presents and discusses the advantages and draw-
backs of two different branch-switching procedures.

5.3.1 Mode injection method

This procedure is similar to the one used in elasticity. Using the property of the
tangent stiffness matrix, the lowest bifurcation point is isolated by successive one side
bisections and the branch-switching to the secondary path is performed by taken as
predictor the eigenvector v associated to the zero eigenvalue of K at the bifurcation
point. The expression “one side bisections” means that, due to the history aspect of
the deformations, the intermediate equilibrium states during the process are always
calculated by reference to the left (lower) limit.
This approach presents two drawbacks. First, contrary to the elastic case, the
eigenvector may not be a good predictor and it is not certain that the Newton-
Raphson iterations will converge on the secondary path. This is in particular the
case when the secondary path is tangent to the fundamental one [98].

97
Second, it can happen that the bifurcation is induced by a discontinuous drop of the
incremental stiffness of the material so that the tangent stiffness matrix along the
fundamental path becomes indefinite without being singular. One such example is
the Euler beam when A σY > Pt , see Section 2.3. The lowest bifurcation point is then
A σY and the tangent stiffness matrix along the fundamental path is discontinuous
at this point. As a matter of fact, the bifurcation is induced by the sudden change
of E to Et in the beam. In such cases, the bifurcation point can still be isolated
by successive bisections, but the notion of critical eigenvalue or eigenvector has no
meaning and the choice of a predictor for the branch-switching is not evident. An
alternative [26, 101] is to take the eigenvector associated to the negative eigenvalue
of K. However, it is not certain that this vector corresponds to the buckling mode
and the convergence on the secondary path is therefore not obvious.
In order to force the Newton-Raphson iterations to converge to the secondary path,
an alternative is to adopt a displacement control procedure and to choose a dis-
placement or a rotation which is zero along the fundamental path. Thus, if the
Newton-Raphson iterations converge, it will necessarily be on the secondary path.
However, if an arc-length procedure is adopted, this solution fails. Then, in order to
maximise the possibility of converging on the secondary path and not on the funda-
mental one, a predictor w perpendicular to the tangent vector to the fundamental
path u can be chosen [26, 101]. In this case, w is computed by
uT u
w =u+βv β=− (5.24)
uT v
where v is the eigenvector associated to the negative eigenvalue.
This previously described procedure has been implemented in the following way.
The bifurcation point is isolated by successive bisections until
λr − λl
< tol (5.25)
λr + λl
where λl and λr are the left (lower) and right (upper) approximations of the critical
load factor and tol is taken as 10−4 . Branch-switching is performed by taking as
predictor the eigenvector associated to the negative eigenvalue for the right point.
Switching to the secondary path is forced by choosing as displacement control a
degree of freedom which is zero on the fundamental path.

5.3.2 Minimisation procedure

The second approach, developed by Petryk [76–81], avoids the problems related to
the definition of the critical eigenvector and the choice of the predictor. This method,
for which the theoretical background has been given in Section 2.4, is based on the
fact that along a stable deformation path, the displacement vector d corresponds
to an absolute minimum of the functional J defined in equation (2.90). With the
notations introduced in Section 5.1.1, the potential J can be rewritten as
1
J = dT K d − dT fex (5.26)
2

98
Using this property, the branch-switching procedure is performed by finding an
incremental solution to the equilibrium equations (5.3) which minimises J. Thus, the
unstable fundamental path is rejected and the stable secondary path is automatically
selected. This method is therefore based on the implicit assumption that, at least
in a neighbourhood of the critical point, the post-buckling path is stable.
The bifurcation point is detected and isolated by classical one side bisections. In
fact, the isolation is not essential for this method. However, due to the historical
aspect of plastic deformations, it is important to initiate the secondary path from a
point which is close to the bifurcation point. The first point of the secondary path
is computed by minimising J. In order to ensure that the minimisation procedure
will not terminate on a saddle point corresponding to the fundamental path, the
predictor is constructed by adding a perturbation vector to the tangent vector at
the fundamental path u. Any random perturbation vector can be taken. However,
if an eigenvalue analysis has already been performed at the bifurcation point, it can
be judicious to take the eigenvector v associated to the negative eigenvalue of K.
In that case, the predictor w is computed by
1 u
w = u+αv α= (5.27)
100 v
The minimisation procedure which has been implemented [36] consists of solving at
each iteration k the system

(K(k) + β I) ∆d = −r(k) (5.28)

where r is the residual vector of the equilibrium equations and I the identity matrix.
The coefficient β is taken to 0 if K is positive definite and to 1.1 λmin otherwise (λmin
is the lowest eigenvalue of K). The displacements are then updated by

d(k+1) = d(k) + µ ∆d (5.29)

where µ is calculated by a line search procedure [57]. However, the tests performed
by the author have shown that the line search procedure was not essential and that
µ = 1 could be taken without affecting the convergence property of the procedure.
The following points along the secondary path are then computed by classical
Newton-Raphson iterations under displacement control.
As already pointed out, this minimisation procedure assumes that the bifurcation is
stable and is therefore not suited for elastic problems. For elasto-plastic instability
problems where buckling is induced by compressive forces, material non-linearity
prevails over the geometric one in the vicinity of the bifurcation point and a stable
post-buckling path is usually observed (cf. the study of the Hutchinson’s model in
Section 2.1.2). However, there is no guarantee that this is always the case. As
an example, unstable secondary path can be obtained if materials with very low
hardening are used. It can also happen that the stable region of the secondary path
is restricted to a small neighbourhood of the bifurcation point and thus is difficult
to catch numerically. All these cases will be illustrated by numerical applications in
Chapters 6 and 9.

99
Chapter 6

2D examples in elasto-plasticity

Three main objectives motivate the selection of the numerical examples included
in this chapter. The first one is to verify the ability of the element formulations
introduced in Chapter 3 in handling large displacement problems in the presence of
material non-linearity. Two classical problems, widely used as benchmark tests in
the literature, have been chosen for this purpose. These are a cantilever beam under
tip load and Lee’s frame.
The second objective of these examples is to test the two branch-switching proce-
dures presented in Section 5.3 and to compute elasto-plastic post-bifurcation paths
for structures exhibiting different types of instabilities in elasticity. Thus, the third,
fourth and fifth examples are characterised in elasticity by symmetric stable, asym-
metric and symmetric unstable bifurcations, respectively.
For these three examples, both the minimisation procedure and the mode injection
method worked very well and convergence to the first point on the secondary path
was obtained within 4–7 iterations. However, as expected, the minimisation proce-
dure failed for the case in the third example which presents an unstable bifurcation.
The third objective of these numerical applications is to examine comparatively the
three local element formulations, and especially test the differences between the two
Bernoulli elements. In addition, the influence of the number of Gauss points in
the cross-section is also analysed. All these comparisons are performed through the
study of the three first examples.

6.1 Example 1: cantilever beam

The purpose of this classical example, described in Figure 6.1, is to test and compare
the elements based on the Bernoulli assumption. The load–displacement diagrams
computed with ten linear elements and fifteen Gauss points along the beam depth
are presented in Figures 6.2 and 6.3. As shown in these figures, the results are in
excellent agreement with those obtained by Kondoh and Atluri [54].

101
 ;;
v

;;
u

0.5
P
L=5
0.1

E = 3 · 107 Et = E/30 σY = 3 · 104

Figure 6.1: Cantilever beam.

This example has also been tested with twenty linear elements and ten shallow
arch elements and the same load–displacement curves were obtained. This indicates
that convergence is already reached with ten linear elements. In the same way,
no significant difference was noted between results obtained with seven and fifteen
integration points.
It can also be noted that computation using ten Timoshenko elements (with ν =
0.3) gave the same load–displacement diagram as the one obtained with Bernoulli
elements. Consequently, the effect of shear deformations can be neglected, which
was not obvious since the height to length ratio is not small.

1400

1200
L+u v
1000
P

800
Load

600

400
present analysis
Kondoh and Atluri
200

0
0 1 2 3 4 5
Displacements

Figure 6.2: Cantilever beam: load–displacements diagram.

102
 160

140

120
v
100
P

80
Load

60
Kondoh and Atluri
40
present analysis
20

0
0 0.5 1 1.5 2 2.5
Displacements

Figure 6.3: Cantilever beam: partial load–displacements diagram.

6.2 Example 2: Lee’s frame

The second example is depicted in Figure 6.4. This structure exhibits a snap-through
behaviour and a limit point in the load–deflection diagram is obtained.
Several discretisations have been used, for comparative purposes in the analysis of
this problem. Thus, the results computed with twenty Timoshenko elements and
seven Gauss points in the cross-section are presented in Figure 6.5 for both elastic
and elasto-plastic cases. For the elasto-plastic analysis, deformed configurations at
typical load steps are shown in Figure 6.6. As noted from Figure 6.5, a very good
agreement with the results obtained by Park and Lee [74] and Cichon [17] is found.
The influence of the number of integration points along the beam depth has been
studied for the Timoshenko element. As shown in Figure 6.7, the difference between
results obtained with seven and fifteen Gauss points is very small.
Load–deflection curves have also been computed using the two types of Bernoulli
elements. Comparison with Timoshenko elements (Figure 6.8) shows that the effect
of shear deformations is almost negligible, which was expected since the two beams
are very slender. Moreover, it is shown in Figure 6.9 that results obtained with ten
shallow arch elements and twenty linear elements are almost identical.

103
 P

;
u
v

E = 720
120

2
ν = 0.3
3 Et = E/10
σY = 10.44

24 96

Figure 6.4: Lee’s frame.

1.5
u
v
1
P

0.5
Load

present analysis elastic

0 present analysis elasto−plastic
Park and Lee
Cichon
−0.5

−1
0 20 40 60 80 100
Displacements

Figure 6.5: Lee’s frame: diagram with 20 Timoshenko elements and 7 Gauss points.

104
 P

P = 1.49

P = 1.26

P = 0.49

P = 0.08

P = 1.24

Figure 6.6: Lee’s frame: elasto-plastic deformed profiles.

1.6

1.4
u
1.2
v
1
P

0.8
Load

0.6
7 Gauss points
0.4
15 Gauss points
0.2

−0.2
0 20 40 60 80 100
Displacements

Figure 6.7: Lee’s frame: diagram with 20 Timoshenko elements.

105
 1.6

1.4
u
1.2
v
1
P
0.8
Load

0.6
Timoshenko
0.4
Bernoulli−Linear
0.2

−0.2
0 20 40 60 80 100
Displacements

Figure 6.8: Lee’s frame: diagram with 20 elements and 7 Gauss points.

1.6

1.4
u
1.2
v
1
P

0.8
Load

0.6
10 linear elements
0.4
20 linear elements
0.2
10 shallow arch elements
0

−0.2
0 20 40 60 80 100
Displacements

Figure 6.9: Lee’s frame: diagram with Bernoulli elements and 15 Gauss points.

106
6.3 Example 3: Euler beam

;
d
P

0.1
Pp 0.1
L=1

E = 2 · 1011 ν = 0.3 Et = E/5 σY = 106

Figure 6.10: Euler beam.

The third example is an Euler beam with the geometrical and material parameters
presented in Figure 6.10. In order to study the imperfection sensitivity, a small
perturbing force P p , acting at the midpoint of the beam is introduced. The load
factor λ and the force P p are defined by
P π 2 Et I Pt
λ= Pt = Pp=
Pt L2 s
where s is varied to give different imperfection magnitudes.
Load–displacement curves computed using Timoshenko and Bernoulli shallow arch
elements are shown in Figures 6.11 and 6.12. As expected, the equilibrium paths
computed with imperfections tend to the path branching at the lowest bifurcation
point when the perturbation force P p tends to zero. This confirms the theoretical
conclusions of Petryk that only the secondary path branching at the lowest bifur-
cation point is stable in an energetic sense and has a physical meaning. Here also,
computations performed with seven and fifteen integration points along the beam
depth gave almost the same results, as shown in Figure 6.13.
A comparison between the two Bernoulli elements is presented in Table 6.1, where
the bifurcation and maximum loads have been computed using fifteen Gauss points.
It can be noted that results obtained with ten shallow arch elements, twenty linear
elements and forty linear elements are almost identical.
The two asymptotic expansions (2.79) and (2.80) presented in Section 2.3.3 are
compared in Figure 6.14 with the results obtained with ten shallow arch elements
and fifteen Gauss points. Very good agreement between expression (2.79) and finite
element calculations is observed up to λ = 1.3, where the two curves begin to diverge.
Expression (2.80) diverges from numerical calculation at λ = 1.1 and therefore the
validity of the term λ3 ξ 9/5 should be seriously questioned.

107
A number of reasons why asymptotic expressions should be considered with great
caution have already been discussed in Section 2.3.3. The examination of the growing
zone of elastic unloading, which appears at bifurcation, gives an additional argument.
The zones of elastic deformations at several increments are shown in Figure 6.16.
Calculations have been performed with forty shallow arch elements in order to get an
accurate representation of the zone of elastic unloading, but the load–displacement
curve is similar to the one depicted in Figure 6.14. It is shown in Figure 6.16
that a zone of plastic reversal stresses is developing before the occurrence of the
maximal load. This zone is not considered in the derivation of the asymptotic
expansions [49, 58]. The same phenomenon has been observed with different values
for the yield limit σY and the tangent modulus Et . Another consequence is that the
characteristic of the hardening cannot be neglected. The present method is based on
isotropic hardening. Similar computations with kinematic hardening would certainly
give different load–displacement diagrams.
Finally, the structure has been tested when Pt < A σY by taking σY = 5 · 108 . In
this case, the lowest bifurcation point is A σY and the tangent stiffness matrix along
the fundamental path becomes negative at this point without being singular. The
secondary paths computed with two different hardening parameters are presented
Figure 6.15. It can be observed that depending on the value of Et , the bifurcation
can be stable or unstable.

1.6

1.5

1.4
Load factor λ

1.3

1.2

1.1 without imperfection

1 s = 10000
s = 1000
0.9

0.8
0 0.05 0.1 0.15 0.2
Displacement d

Figure 6.11: Euler beam: diagram with 20 Timoshenko elements and 15 G.p.

108
 1.6

1.5

1.4

Load factor λ
1.3

1.2

1.1 without imperfection

s = 10000
1
s = 1000
0.9

0.8
0 0.05 0.1 0.15 0.2
Displacement d

Figure 6.12: Euler beam: diagram with 10 shallow arch elements and 15 G.p.

1.5

1.45

1.4

1.35
Load factor λ

1.3

1.25

1.2
15 Gauss points
1.15
7 Gauss points
1.1

1.05

1
0 0.05 0.1 0.15 0.2
Displacement d

Figure 6.13: Euler beam: diagram with 10 shallow arch elements.

109
 1.9

1.8
(2.80)
1.7

1.6
λ Load factor 1.5
F.E.M.

1.4
(2.79)
1.3

1.2

1.1

0.9
0 0.01 0.02 0.03 0.04 0.05
Displacement d

Figure 6.14: Euler beam: comparison between asymptotic expressions and numeri-
cal calculations.

5.5
Et = E/5
Applied load / 10 6

4.5

Et = E/10
4

3.5
0 0.01 0.02 0.03 0.04 0.05
Displacement d

Figure 6.15: Euler beam: diagram when Pt < A σY .

110
 Table 6.1: Euler beam: comparison of Bernoulli elements.

Number and type of elements λcr λmax

10 linear elements 1.017 1.501
20 linear elements 1.010 1.491
40 linear elements 1.009 1.488
10 shallow arch elements 1.008 1.489
20 shallow arch elements 1.008 1.488

λ = 1.02

λ = 1.06

λ = 1.16

λ = 1.40

λ = 1.47

λ = 1.49

Figure 6.16: Euler beam: zone of elastic unloading at different load factors. Circles
represent Gauss points where elastic deformation occurs.

111
6.4 Example 4: asymmetrical frame

; E = 2 · 1011
L=1

0.1
Et = E/2

0.1 σY = 106

L=1

Figure 6.17: Asymmetrical frame.

The frame described in Figure 6.17 exhibits an asymmetrical bifurcation for the case
of elastic behaviour, as shown in Figure 6.18. The load factor λ is defined by

4 π2 E I
P =λ in elasticity (6.1)
L2
and
4 π 2 Et I
P =λ in elasto-plasticity (6.2)
L2

The purpose of this example is to investigate the phenomenon of bifurcation in

elasto-plasticity. Sixteen shallow arch elements and fifteen Gauss points have been
used. The results are shown in Figure 6.19. Both post-bifurcation secondary paths
present a stable part before a limit point is reached. These results confirm the
conclusion derived from the analysis of Hutchinson’s model (Section 2.1.2) that in
the vicinity of the bifurcation, material non-linearity prevails over the geometric one
and, consequently, a stable path is usually obtained.

112
 0.65

0.6

λElastic load factor

0.55

0.5

0.45
−0.05 0 0.05 0.1 0.15
Displacement d

Figure 6.18: Asymmetrical frame: post-bifurcation paths in elasticity.

0.8

0.78

0.76
λElasto−plastic load factor

0.74

0.72

0.7

0.68

0.66

0.64

0.62

0.6
−0.05 0 0.05 0.1 0.15
Displacement d

Figure 6.19: Asymmetrical frame: post-bifurcation paths in elasto-plasticity.

113
 ;
6.5 Example 5: toggle frame

0.243
P 0.753
E = 107
0.5

Et = E/2
d
σY = 3 · 103
12.943 12.943

Figure 6.20: Toggle frame.

The toggle frame, described in Figure 6.20 has been analysed using a discretisation
with eight linear Bernoulli elements and seven Gauss points along the beam height.
The equilibrium paths for the elastic and elasto-plastic situations are shown in Fig-
ures 6.21 and 6.22. In the elastic case, the structure exhibits a symmetric unstable
bifurcation while in the elasto-plastic case, a stable bifurcation is observed. How-
ever, it must be noted that the stable part of the secondary path is very short and a
limit point is quickly reached. This observation confirms once again the conclusion
that in the vicinity of the bifurcation point, material non-linearity prevails over the
geometric one and a stable path is usually obtained.
Moreover, the results plotted in Figure 6.22 suggest that the structure presents a
tangent bifurcation in the elasto-plastic case. The existence of such a bifurcation has
been proved theoretically by Leger and Potier-Ferry [58] and an example has been
studied by Triantafyllidis [98]. According to these authors, a tangent bifurcation can
occur when the structure is only partially plasticised along the fundamental path
and is characterised by the apparition of a zone of elastic reloading. The analysis of
Figures 6.23 and 6.24 shows that along the fundamental path, a zone of plasticisation
is developing on the concave side of the two beams while the rest of the structure
deforms elastically. In addition, a zone of elastic unloading (ε̇ < 0) appears on the
convex side of the two beams. Along the secondary path, the zone of plasticisation
continues to grow in the left beam while it decreases and disappears in the right
beam. At the same time, the zone of elastic unloading disappears on the convex
side of the right beam. Consequently, the conditions for a tangent bifurcation stated
in [58] seem to be respected, but a rigorous proof of such a phenomenon cannot be
performed in a discretised and numerical context.

114
 35

30

25 elastic
elasto−plastic
Load P 20

15

10

−5

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8

Displacement d

Figure 6.21: Toggle frame: elastic and elasto-plastic diagrams.

28.5

28
Load P

27.5

27

26.5
0.08 0.085 0.09 0.095 0.1
Displacement d

Figure 6.22: Toggle frame: tangent bifurcation in elasto-plasticity.

115
 P = 27.2 d = 0.085
P = 26.9 d = 0.11
P = 18.2 d = 0.20

Figure 6.23: Toggle frame: deformed profiles before and after bifurcation.

fundamental path: inc. before bifurcation fundamental path: inc. after bifurcation

secondary path: inc. after bifurcation secondary path: 2nd inc. after bifurcation

Figure 6.24: Toggle frame: zone of plastic deformations at different increments. Cir-
cles represent Gauss points where plastic deformation occurs. Stars
represent Gauss points where elastic unloading occurs.

116
Chapter 7

3D examples in elasticity

The test problems included in this chapter have two objectives. The first one is to
test the ability of the elements in handling problems involving large 3D rotations.
As a general rule, the parameterisation in terms of the incremental rotation vector is
used throughout the whole set of test problems. However, for the first two examples,
the parameterisation based on the total rotation vector together with the more
traditional “intrinsic” parameterisation (see Section 4.2.5) were also included in
order to discuss the relative merits of the three possible choices. The second objective
is to test the ability of the elements to deal with problems involving coupling between
axial, bending and torsional-warping effects. For this purpose several examples
exhibiting pure torsional or flexural-torsional buckling were also included.
For all the instability problems included in this chapter, the procedure of direct
isolation of bifurcation points presented in Section 5.2 was used. Convergence to
the bifurcation point was obtained within 4–6 iterations. The performances of this
procedure will be investigated in more detail in Chapter 8.
For certain problems, the results obtained from the beam models have been com-
pared with results obtained using shell models. The shell models were constructed
using the flat facet triangular element developed by Pacoste [68, 71].
The characteristics of the cross-sections have been either taken from the literature
or determined with the finite element program presented in Appendix B.

7.1 Example 1, deployable ring

This first example, introduced by Goto [38], refers to the ring depicted in Figure 7.1.
The ring is 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 (point A) is restricted
to move along and rotate around the x axis.

117
 E = 200000
x G = 76923
R = 120

A
t = 0.6
h=6 M

Figure 7.1: Deployable ring.

The ring shows a very interesting behaviour under increasing rotation at A. Thus,
after a complete rotation cycle, i.e. ΘA = 2π the ring is 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. ΘA = 4π, will bring the ring back to the original configuration.
This example is therefore ideally suited in order to test the ability of the formulation
to handle large rotations.

10

8
shell
6
intrinsic
Applied moment / 100

4
rot. vec.
2

−2

−4

−6

−8
0 2 4 6 8 10 12
Rotation at A

Figure 7.2: Deployable ring: load–displacement curves.

118
The results shown in Figure 7.2 were obtained using 128 t3dl elements (with J =
0.432). The analysis included all three parameterisations of finite 3D rotations dis-
cussed in Section 4.2.5. Thus, for the total rotational vector, a quadratic convergence
rate was observed for values ΘA < 2π. In the vicinity of 2π, the tangent stiffness
matrix became ill-conditioned (see Section 4.1.1), the quadratic convergence was
gradually lost and ultimately, the iterations failed to converge. In contrast, the
parameterisation based on the incremental rotation vector as well as the usual “in-
trinsic” parameterisation, were unaffected and the computations were continued up
to ΘA = 4π. A quadratic convergence rate was maintained for the whole range of
the equilibrium path.
Note that since the point A is restrained to rotate around the x axis, all three types of
parameterisation produce identical results, regardless of whether the transformation
of the external moment is performed or not. Finally, note the excellent agreement
between the present solution and the results from the shell analyses presented in [71].

7.2 Example 2, helical beam

M
x

F
z

Figure 7.3: Helical beam.

This example was introduced by Ibrahimbegovic [51]. Its purpose is to illustrate the
effects of the transformations in equations (4.175) and (4.176) on the outcome of
the analysis, for problems involving concentrated moment loading. Thus, the beam
shown in Figure 7.3 is subjected to a concentrated moment M and an out-of-plane
force F , applied at its free end. Both M and F are increased proportionally up
to the values Mmax = 120π and Fmax = 30. The beam is modelled with 60 t3dl
elements. The geometric and sectional properties of the beam are

L = 10 EA = GA = 104 EI22 = EI33 = GJ = 102

The simultaneous application of a concentrated moment and a force, bends the beam
into a helical shape. At the end of the loading process, the beam is wrapped around
in six concentric circles. The results, obtained using the “intrinsic” parameterisation
of finite rotations are shown in Figure 7.4. Note that the out of plane z-displacement
oscillates around a zero value. Each passing through zero corresponds to a deformed
shape which is entirely situated in the plane xy.

119
 6

5
intrinsic param.
rot. vector with transf.
4
inc. rot. vector with transf.
Pseudo−time
rot. vector without transf.
3
inc. rot. vector without transf.
2

0
−2 −1 0 1 2 3 4 5 6
Free end displacement in z−direction

Figure 7.4: Helical beam: load–displacement curves.

Changing the type of parameterisation should not affect the outcome of the anal-
ysis since the physical problem does not change. Thus, absolutely identical results
should be obtained using either the total rotational or the incremental rotational
vector. However, as shown in Figure 7.4, this is the case only if the transforma-
tions described in Section 4.5.2 are applied on the external moment M . Without
such a transformation, a different parameterisation will imply a different physical
definition of M . As a consequence, the results obtained using different types of
parameterisation will be completely different.

7.3 Example 3, right angle frame

cross-section
E = 71240
0
G = 27191 24 30
0.6
y

z x
M M

Figure 7.5: Right angle frame.

120
This classical example (see Figure 7.5), has often been used in order to check the
ability of various beam formulations to properly account for torsional-bending cou-
pling [40, 44, 64, 73, 92, 93]. Due to the symmetry, only half of the frame is modelled.
At the support only translation along x and rotation around z are allowed. The
loading is given by a pair of concentrated moments applied at the support. The
non-trivial cross sectional characteristics are J = 2.133 and Iω = 40.43.
The problem is typical for situations where a local linear strain definition will lead to
an incorrect solution. Thus, the results obtained using 10 t3dl elements are shown
in Figure 7.6. As the moment increases, an unstable bifurcation point is reached.
These results are in excellent agreement with those found in the literature [40, 44,
64, 73, 92, 93].
A qualitatively different behaviour is predicted by the models using 10 tw3d or t3d
elements, as shown in Figure 7.7. A stable bifurcation point is now reached at a
slightly higher value of the critical moment. The secondary path loses its stability at
a limit point, for a value of the moment of about 1.20Mcr . In order to confirm these
results, the analysis was repeated using a model with 576 triangular shell elements
for one half of the frame. A good agreement with the beam model is observed up
to a value of the apex displacement of about one half of the frame height. Over this
value the quantitative differences are more pronounced. According to the author’s
opinion, these differences originate from the very large height to thickness ratio of
the cross-section. As a result, in the large displacement regime, the cross-section
can no longer be assumed as undeformable in its plane and a beam analysis is not
appropriate. Finally, it can be noted that similar results were obtained by Sauer [86].

800

600

400
Applied moment

200
t3dl
0
Nour−Omid
−200

−400

−600

−800
−200 −150 −100 −50 0 50 100 150 200
Apex displacement in z−direction

Figure 7.6: Right angle frame: post-buckling behaviour; t3dl element.

121
 1000

900

800

700
Applied moment
600

500

400
shell
300
tw3d
200
t3d
100

0
0 20 40 60 80 100 120 140 160 180
Apex displacement in z−direction

Figure 7.7: Right angle frame: post-buckling behaviour; tw3d and t3d elements.

7.4 Example 4, non-linear torsion

E = 2100000 G = 787500 cross-section

M
10

100 0.5

Figure 7.8: Non-linear torsion of a cantilever.

The non-linear torsion of the cantilever beam, depicted in Figure 7.8, was in-
vestigated. The non-trivial cross sectional characteristics are J = 0.40433 and
Iω = 0.85976. The results obtained using 20 tw3d elements are shown in Figure 7.9
and compared with the analytical solution [27] based on Vlasov’s beam theory. Very
good agreement between the two solutions is observed.
Note that for a beam subjected to pure torsion, the only non-zero nodal displace-
ments are the translations along and rotations around the longitudinal axis of the
beam. Consequently, the co-rotational framework in itself will not introduce any

122
non-linearity. The non-linear character of the solution shown in Figure 7.9, is en-
tirely generated by the so-called Wagner term included in the local element defini-
tion. The absence of this term, as in the case of the t3dl or b3dl elements, will result
in a purely linear solution.

9000

8000
tw3d
7000
analytical sol.
6000
Applied moment

5000

4000

3000

2000

1000

0
0 0.5 1 1.5 2
End rotation

Figure 7.9: Non-linear torsion of a cantilever: moment versus end-rotation.

7.5 Example 5, channel-section beam

10
x2
E = 21000 G = 8077
x2
x1 x3
x3
30

C O G
P
1
900
1.6

Figure 7.10: Channel-section beam.

123
The example presented in Figure 7.10 was introduced by Gruttmann et al. [40, 41].
It refers to a channel-section beam, clamped at one end and subjected to a tip force
at the free end. The parameters of the cross-section are

A = 58.8 J = 34.994 I22 = 8063.4 I33 = 564.25 I2r = 0

I3r = 10827 Irr = 1676005 Iω = 79917 OG = 2.449 OC = 3.567

The results, obtained using 20 bw3d elements, are shown in Figure 7.11. Very
similar results were obtained using 20 tw3d elements. These results are in excellent
agreement with those obtained in [41] using beam elements and with the results
provided by a shell element model. The shell model used a (4 + 12 + 4) × 180 mesh
with 7200 triangular elements.

20

18

16

14
Applied load

12

10

8
bw3d
6
Shell
4
Gruttmann et al.
2

0
0 50 100 150 200 250 300
x2 displacement of point O

Figure 7.11: Channel-section beam: load–displacement diagram.

7.6 Example 6, torsional buckling

The torsional buckling of the beam depicted in Figure 7.12 was analysed. The
parameters of the cross-section are

A = 12.8 J = 0.683 I22 = I33 = 136.62 Irr = 10493 Iω = 3.641

124
 x2
E = 21000 G = 8077
P x3

0.4

8
100
8

Figure 7.12: Torsional buckling.

All degrees of freedom including warping, are set to zero at the left end and only
the axial displacement is permitted at the other end. Since the cross-section is bi-
symmetrical and the axial load is applied at the centroid of the cross-section, the
bending and torsional buckling modes are uncoupled.
The analysis was initiated by isolating the first three critical points on the funda-
mental path of the beam. The critical modes are torsional modes with 1,2 and 3
half-waves, respectively. The corresponding critical loads, obtained using a mesh
with 10 bw3d elements, are listed in Table 7.1. These results are in excellent agree-
ment with the analytical solutions given by Timoshenko and Gere [97] based on
Bernoulli beam theory with Vlasov warping. Rigorously speaking, only the bw3d
element matches these assumptions. However, since for the present problem the
effects of the bending shear strains on the torsional buckling modes are negligible,
the tw3d element was also included in the comparison. Note that a finer mesh of
tw3d elements is needed in order to obtain the same level of accuracy as for bw3d,
especially for the higher buckling modes. This situation originates from the fact
that tw3d is defined using linear interpolation functions.

Table 7.1: Torsional buckling: critical loads.

No. Theory [97] 10 bw3d 10 tw3d 20 tw3d mode

1 272.57 272.57 273.55 272.80 torsional (1h-w)
2 287.35 287.38 291.99 288.40 torsional (2h-w)
3 314.99 315.17 334.00 318.89 torsional (3h-w)

Using a model with 20 bw3d elements, a post-buckling analysis was also performed.
The results, plotted in Figure 7.13, are in excellent agreement with those obtained
from a shell model with 1600 triangular elements.
Finally note that for the present problem, an analysis using a linear local strain
definition will completely miss the torsional modes. This comes from the fact that
the only non-zero displacements on the fundamental path are the axial ones. In
such a situation, as also shown in example 4, the co-rotational framework does not
induce any coupling between the axial and torsional effects.

125
 700

650

600
bw3d
550
Applied load shell
500

450

400

350

300

250

200
0 0.1 0.2 0.3 0.4 0.5 0.6
Axial rotation of the middle point

Figure 7.13: Torsional buckling: post-bifurcation behaviour.

7.7 Example 7, lateral torsional buckling

x2 E = 21000 G = 8077 0.2 x2 0.2

x3 x1 P 10
C G x3

150
10

Figure 7.14: Lateral torsional buckling.

This example, taken from [41], concerns the lateral torsional buckling of the beam
depicted in Figure 7.14. The characteristics of the cross-section are

A = 5.92 J = 0.0792 I22 = 110.8 I33 = 64.49 I2r = 0

I3r = 209.06 Irr = 6182.9 Iω = 1108.2 GC = 7.55

The axial load is applied at the centroid of the cross-section. The beam is assumed
as simply supported. The rotation around the axis of the beam is set to zero at

126
both supports. Warping deformations are however permitted. Since the cross-
section is not bi-symmetrical, both bending and torsional effects are included in
the buckling mode. The buckling loads, obtained using different discretisations are
listed in Table 7.2. Very good agreement with the theoretical solution given in [97]
is obtained.

Table 7.2: Lateral torsional buckling loads.

Theory [97] 10 bw3d 10 tw3d 20 tw3d

115.54 115.62 117.18 115.81

7.8 Example 8, lateral buckling of a cantilever

P t =1

h
E = 21000
L = 10h
G = 8077

Figure 7.15: Lateral buckling of a cantilever.

This example concerns the lateral buckling of the narrow rectangular cantilever
beam shown in Figure 7.15. At the free end, a transversal load is applied at the
centroid of the cross-section. The load is described as P = λPo with Po given by
1√ ht3 ht3
Po = 2 EIGJ I= J= (7.1)
L 12 3
The buckling loads were evaluated for various t/h ratios. The obtained buckling
load factors λcr are listed in Table 7.3. Both the b3d and b3dl elements were used in
the analysis (10 elements meshes). For comparison purposes the theoretical buckling
loads according to Timoshenko and Gere [97], are also listed in the table. Note that
for low values of the t/h ratio, significant pre-buckling deflections occur. Varying
this parameter will thus illustrate the influence of pre-buckling displacements on the
critical load.

Table 7.3: Lateral buckling of a cantilever: λcr for various heights h.

h Theory [97] 10 b3dl 10 b3d

50 4.013 4.039 4.039
10 4.085 4.099 4.099
5 4.324 4.313 4.314
3 5.030 5.072 5.077

127
One particularly relevant parameter for this problem refers to the specific point
of application of the force P . Thus, Table 7.4 lists the results obtained for three
different cases: P applied at the centroid, top and bottom of the end cross-section.
A very good agreement with the theoretical solutions given in [97] can be noted.

Table 7.4: Lateral buckling of cantilever: λcr for various load positions (h = 50).

position Theory [97] 10 b3dl 10 b3d

centroid 4.013 4.039 4.039
bottom 4.175 4.198 4.198
top 3.851 3.859 3.859

The post-buckling behaviour of the cantilever was also investigated with the load
applied at the centroid of the end cross-section. Two cases were considered. For
the first one, h = 10. All three Bernoulli elements, i.e. bw3d, b3d and b3dl, were
considered in the analysis. The Saint-Venant torsion modulus and the warping
constant were taken as J = 3.123 and Iω = 6.643. The results depicted in Figure 7.16
were obtained using 20 element meshes. The results obtained with the element bw3d
are in excellent agreement with those provided by a shell model using a mesh with
720 triangular elements. The results obtained with the elements b3d and b3dl are
a little bit different, which shows that warping has some minor effect.

12

11
shell
10
bw3d
Load factor λ

9
b3d
8
b3dl
7

4
0 0.1 0.2 0.3 0.4 0.5
Lateral tip displacement / L

Figure 7.16: Lateral buckling of a cantilever: post-buckling behaviour h/t = 10.

128
The numerical data for the second case were taken from the literature [55, 73].
Referring to the notations in Figure 7.15, the following values of the geometrical,
mechanical and cross-sectional parameters were used

L = 240 h = 30 t = 0.6 J = 2.133

Iω = 40.43 E = 71.24 · 106 G = 27.19 · 106

The results shown in Figure 7.17, were obtained using the same mesh and element
types as before. In addition, the results from a shell model with 576 triangular
elements, are also plotted in the figure. For the bw3d element, good agreement
with the shell results is obtained up to a lateral displacement of about 0.45L. The
situation is very similar to that in example 3. For very high h/t ratios, in the
large displacement regime, a beam analysis is not appropriate due to significant in-
plane deformations of the cross-section. This explanation is also supported by the
results obtained in the previous case h/t = 10 which showed an excellent agreement
between the shell and beam models. It is worth mentioning here, that the results
obtained using the bw3d and shell elements are significantly different from those of
Kouhia [55]. According to the author’s opinion, the differences are due to the fact
that in the cited reference, an averaged form of the Wagner term (see Section 4.3.2)
is used in the definition of the element.
Finally note that the b3dl element (which is also used in [73]) appears to be too
crude in order to model correctly the post-buckling behaviour of the cantilever beam,
especially when the ratio h/t is high.

10
shell
9
bw3d
b3d
8
λ

b3dl
Load factor

7
Kouhia

4
0 0.1 0.2 0.3 0.4 0.5 0.6
Lateral tip displacement / L

Figure 7.17: Lateral buckling of a cantilever: post-buckling behaviour h/t = 50.

129
7.9 Example 9, lateral buckling with warping

x2
x2 E = 21000
P
x3 x1 G = 8077 1 x3

10

2
L
4

Figure 7.18: Lateral buckling with warping.

This example concerns the lateral buckling of the simply supported I beam shown
in Figure 7.18. The characteristics of the cross-section are

A = 24 J = 24.67 I22 = 448 I33 = 22 Irr = 12352 Iω = 533.3

A transversal load P is applied at the centroid of the midsection. The load is

described as P = λPo with Po given by
1'
Po = EI33 GJ (7.2)
L2
A theoretical solution to this problem was given by Timoshenko and Gere [97].
This solution takes into account the effect of warping but neglects the effect of pre-
buckling bending deformations. Thus, in order to enable a comparison with the
numerical results, the value of I22 was multiplied by a factor of 100. The buckling
load factors λcr , obtained using a discretisation with 20 bw3d elements, are listed
in Table 7.5 for different values of the length L. An excellent agreement with the
theoretical solutions can be noted.

Table 7.5: Lateral buckling with warping: critical load factors λcr .
L2 GJ
EIω
L Theory [97] 20 bw3d
80 67.1 18.1 18.1
160 94.8 17.5 17.6
320 134.1 17.2 17.3

7.10 Example 10, L frame

The final example concerns the post-buckling behaviour of the L frame shown in
Figure 7.19. The non trivial characteristics of the cross-section are J = 2.133 and
Iω = 40.43. The frame is loaded by a horizontal force P applied at the centroid

130
of the end cross-section. Two different modelling alternatives were tested in the
analysis. The first one disregards the finite dimensions of the connection and uses
beam elements for the whole length of the frame members. For the second one,
the connection area is modelled using rigid links, as shown in Figure 7.19. In both
cases, a discretisation with 10 tw3d elements per member was used. The obtained
results are shown in Figure 7.20. The rigid link model gives a reasonably good fit
with shell results obtained using a mesh with 512 triangular elements. Note that
the use of rigid links amounts to assuming the connection area as infinitely rigid,
which is certainly not the case. This explains the slightly overstiff behaviour of the
rigid link beam model.

255
z
x
cross-section

0.6

255
E = 71240
y

30
G = 27191

P
A

Figure 7.19: L frame.

2.8

2.6
shell
2.4
rigid link
Applied load

2.2
model 1
2

1.8

1.6

1.4

1.2

1
0 10 20 30 40 50 60
z−displacement of point A

Figure 7.20: L frame: post-buckling behaviour.

131
Chapter 8

Elastic isolation – examples

The purpose of the eight problems included in this chapter is to compare the per-
formances of the procedure presented in Section 5.2 with the classical approach of
Wriggers et al. [104, 105]. For that, the number of iterations required for conver-
gence is given in tables. Note that for the present method, both the total number
of iterations and the number of the iterations which used the extended system (in
parenthesis) are given. In order to allow a better comparison, for each structure,
the isolation procedure was initiated from several equilibrium points.
The tolerance, tol, is given for each example. For the classical approach, this refers
to the norm of the residual of the extended system. For the present method, its
definition is given in Table 5.3.
The 2D numerical examples were solved by using a local linear Timoshenko element,
while the 3D examples were computed using the local Timoshenko element t3d
presented in Section 4.3.4.

8.1 Example 1, 2D clamped arch

6
d EA = GA = EI = 10

100

145

Figure 8.1: 2D clamped arch.

133
This example, introduced by Wriggers and Simo [104] is depicted in Figure 8.1. The
arch is discretised into 20 elements. The tolerance is set to tol = 10−6 . As shown
in Figure 8.2, the structure exhibits a limit point at a load Fcr = 972.9 and for a
displacement dcr = 72.13. The isolation procedure was initiated from five different
points. The results, presented in Table 8.1, show that if the starting point is far from
the critical one, the present method is more reliable than the classical approach.

1000

900

800

700

600
Load F

500

400

300

200

100

0
0 20 40 60 80 100 120
Displacement d

Figure 8.2: 2D clamped arch: equilibrium path.

Table 8.1: 2D clamped arch: isolation – dcr = 72.13

Starting point d = 20 d = 40 d = 60 d = 80 d = 100

Classical approach no 8 5 5 no
Present method 11 (6) 9 (4) 6 (4) 5 (4) 9 (4)

8.2 Example 2, 2D simply supported arch

This example, shown in Figure 8.3, is similar to the previous one, except that the
arch is simply supported at both ends. With these new boundary conditions, a
bifurcation point is obtained on the fundamental path for a force Fcr = 334.9, as
shown in Figure 8.4. The convergence properties of the two procedures, given in
Table 8.2 are very similar. In order to illustrate the switching between the extended
system and pure equilibrium iterations under load control, the details of the present
approach are given in Table 8.3 for the lowest starting point. Along the secondary
path, a limit point is obtained at Fcr = 423.1, dcr = 86.65. The results of its

134
isolation, presented in Table 8.4, show that the present method is better if the
starting point is far from the limit one.

6
d EA = GA = EI = 10

100

145

Figure 8.3: 2D simply supported arch.

600

500

400
Load F

300

200

100

0
0 20 40 60 80 100
Displacement d

Figure 8.4: 2D simply supported arch: equilibrium paths.

Table 8.2: 2D simply supported arch: isolation of the bifurcation pt – Fcr = 334.9

Starting point F = 200 F = 300 F = 400 F = 500

Classical approach 5 4 5 6
Present method 6 (4) 4 (3) 6 (3) 7 (4)

135
 Table 8.3: 2D simply supported arch: details of the present approach.

Iter. Residual norm Lowest eigenvalue Force Type of the next iteration
0 1.192 · 10−9 1.164 · 10−1 200.00 extended system
1 1.176 · 104 2.408 · 100 331.56 equilibrium equations only
2 3.209 · 102 5.631 · 10−3 331.56 extended system
3 7.804 · 101 3.224 · 10−2 334.89 equilibrium equations only
4 1.786 · 10−1 1.738 · 10−5 334.89 extended system
5 3.372 · 10−4 1.525 · 10−7 334.91 extended system
6 3.100 · 10−9 1.226 · 10−10 334.91

Table 8.4: 2D simply supported arch: isolation of the limit point – dcr = 86.65

Starting point d = 70 d = 80 d = 100 d = 110

Classical approach no 9 10 16
Present method 21 (5) 8 (4) 11 (5) 20 (5)

8.3 Example 3, 2D shallow arch

F E = 290000 G = 111538

d 0.95

R = 254 65.74
2.54

Figure 8.5: 2D shallow arch.

The structure presented Figure 8.5 is discretised into 40 elements. The tolerance is
set to tol = 10−6 . The load–displacement diagram in Figure 8.6 shows that a limit
point is reached at the values dcr = 2.917, Fcr = 45.82. The results concerning the
isolation are given in Table 8.5. For this structure, both methods worked very well.
This is probably due to the fact that the deformations are relatively small.

Table 8.5: 2D shallow arch: isolation – dcr = 2.917

Starting point d = 0.5 d = 1.0 d = 2.0 d = 4.0

Classical approach 8 7 6 6
Present method 8 (8) 8 (6) 6 (5) 8 (5)

136
 50

45

40

35

30

Load F 25

20

15

10

0
0 1 2 3 4 5
Displacement d

Figure 8.6: 2D shallow arch: equilibrium path.

8.4 Example 4, deep circular arch

d 6
E = 6 10
100 6
G = 3 10

145

Figure 8.7: Deep circular arch.

This example, shown in Figure 8.7 has been introduced by Cardona and Huespe [12].
The arch is fully clamped at one end and only the in plane rotation is permitted at
the other end. The tolerance is set to tol = 10−6 . The structure is discretised using
40 elements. The characteristics of the square cross-section are
A = 1.414 I22 = I33 = 0.1667 J = 0.4741 Irr = 0.1100
The structure (see Figure 8.8) presents a bifurcation point at a load value Fcr =
256.7, corresponding to an out of plane bifurcation mode. For this problem (see
Table 8.6), both isolation methods give almost similar results.

137
 450

400

350

300

Force F
250

200

150

100

50

0
0 5 10 15 20 25 30
Displacement d

Figure 8.8: Deep circular arch: equilibrium paths.

Table 8.6: Deep circular arch: isolation – Fcr = 256.7

Starting point F = 100 F = 200 F = 300 F = 400

Classical approach 6 6 6 12
Present method 8 (4) 6 (2) 6 (2) 9 (3)

8.5 Example 5, narrow cantilever

F t =1
E = 21000
h = 10

L = 100 G = 8077

Figure 8.9: Narrow cantilever.

This example concerns the lateral buckling of the narrow rectangular cantilever
beam shown in Figure 8.9. At the free end, a transversal load is applied at the
centroid of the cross-section. The load is described as F = λFo with Fo given by:
1√ ht3 ht3
Fo = EIGJ I= J= (8.1)
L2 12 3

138
The beam is discretised into 20 elements and the tolerance is set to tol = 10−8 . The
load–displacement diagram is given in Figure 8.10 and the convergence properties of
the isolation procedures are listed in Table 8.7. The lateral buckling is obtained for a
load factor λcr = 4.083, which is very close to the analytical result (4.085) obtained
by Timoshenko and Gere [97]. For this example, the pre-buckling deformations are
linear and very small, and both methods converge rapidly.

5
Load factor λ

0
0 0.5 1 1.5 2 2.5 3
Displacement d

Figure 8.10: Narrow cantilever: equilibrium paths.

Table 8.7: Narrow cantilever: isolation – λcr = 4.083

Starting point λ=2 λ=3 λ=5 λ=6

Classical approach no 5 5 6
Present method no 6 (5) 6 (4) 6 (6)

8.6 Example 6, L frame

This example, depicted in Figure 8.11, is also taken from Cardona and Huespe [12].
The frame is modelled by 16 elements. The tolerance is set to tol = 10−7 . As shown
in Figure 8.12, the structure develops very large pre-buckling deformations before
the bifurcation point (Fcr = 1.243) is reached. For this structure (see Table 8.8),
the present method gives better results than the classical approach.

139
 240

cross-section

1.2
E = 71240

240

3.5
G = 27191

F d

Figure 8.11: L frame.

2.5

1.5
Force F

0.5

0
0 20 40 60 80 100 120
Displacement d

Figure 8.12: L frame: equilibrium paths.

Table 8.8: L frame: isolation – Fcr = 1.243

Starting point F = 0.7 F = 1.0 F = 1.5 F = 2.0

Classical approach no 21 7 13
Present method no 6 (3) 6 (3) 8 (4)

140
8.7 Example 7, cable hockling

M E = 71240 A =1
G = 27190 J = 2.16
240 I = 0.0833

Figure 8.13: Cable hockling.

This classical example, see e.g. [52, 64], is described in Figure 8.13. It concerns an
initially straight cable built in at one end and submitted to a concentrated axial
moment at the other end. At the loaded end, only the axial displacement and axial
rotation are permitted. The finite element model consists of 20 elements. The toler-
ance is set to tol = 10−6 . The load–displacement diagram is plotted in Figure 8.14.
The fundamental path corresponds to a pure torsion while the secondary path, ini-
tiated at a value of the applied moment Mcr = 225.4, is characterised by bending
deformations. The results, given in Table 8.9, show that also for this structure, the
present method performs better than the classical approach.

350

300

250
Moment M

200

150

100

50

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
End rotation

Figure 8.14: Cable hockling: equilibrium paths.

Table 8.9: Cable hockling: isolation – Mcr = 225.4

Starting point M = 150 M = 200 M = 250 M = 300

Classical approach 11 26 14 26
Present method 4 (4) 3 (3) 3 (3) 4 (4)

141
8.8 Example 8, spatial frame

F cross-section

15
0.14

0.6

0.17
7
E = 10
6
G = 3.8462 10

Figure 8.15: Spatial frame.

The last example concerns the spatial dome structure depicted in Figure 8.15. The
structure is composed of three beams placed symmetrically with respect to the
vertical axis and is loaded by a vertical force applied at its top. Each beam is
discretised by 8 elements. The tolerance is set to tol = 10−7 .
As shown in Figure 8.16, several bifurcation points are present along the fundamental
path. These points are relatively close to each other, which generates difficulties for
the initialisation and convergence of the isolation procedures. This situation is
illustrated in Figure 8.17. The fundamental path was computed using displacement
control by imposing at each step a vertical displacement increment ∆d = 0.05 at the
top. Hence, at the equilibrium point corresponding to d = 0.1, the first bifurcation
point was detected by the appearance of a negative eigenvalue λ1 , see Table 8.10.
However, due to the proximity of the second bifurcation point, λ1 is not the lowest
one in absolute value. Consequently, in order to isolate the first bifurcation point,
φo , which is the eigenvector associated to λ1 , cannot be evaluated using the iterative
scheme (5.18). As a matter of fact, this system converges to λ2 and both isolation
procedures will then converge to the second bifurcation point. In this case, φo must
be evaluated using a different numerical method. This operation may take some
time, particulary for systems with large number of degrees of freedom.
A similar problem occurs at the equilibrium point corresponding to d = 0.15. The
second and third bifurcation points, are detected by the appearance of two additional
negative eigenvalues, see Table 8.10. Note that λ2 and λ3 are double eigenvalues. In
order to isolate the third bifurcation point, λ3 and its associated eigenvector must
be taken for go and φo . Since λ3 is the lowest absolute eigenvalue, the iterative
process (5.18) can be used to compute φo . However, in order to converge to the
second bifurcation point, λ2 and its associated eigenvector must be taken as initial
values. In this case, φo must be calculated using a different numerical method.
The convergence characteristics for the isolation of the first three bifurcation points
are reported in Table 8.11. For this example, the pre-buckling deformations are
small, and both methods work very well.

142
 9

5
Force F
4

−1
0 0.2 0.4 0.6 0.8 1
Vertical displacement at the top

Figure 8.16: Spatial frame: equilibrium paths.

6
Force F

0
0 0.05 0.1 0.15 0.2
Vertical displacement at the top

Figure 8.17: Spatial frame: near bifurcation points.

143
 Table 8.10: Spatial frame: lowest eigenvalues of K.

eq. pt. λ1 λ2 λ3
d = 0.10 −1.850 (1) +0.191 (2) +1.892 (2)
d = 0.15 −4.174 (1) −1.883 (2) −0.359 (2)

Table 8.11: Spatial frame: isolation.

Bifurcation point Fcr = 4.249 Fcr = 6.169 Fcr = 7.313

Starting point d = 0.10 d = 0.15 d = 0.15
Classical approach 4 4 3
Present method 4 (3) 4 (4) 3 (3)

144
Chapter 9

3D examples in elasto-plasticity

The test problems included in this chapter have two objectives. The first one is to
test the ability of the element to deal with elasto-plastic problems involving coupling
between axial, bending and torsional warping effects. For this purpose, examples
exhibiting lateral and flexural-torsional buckling were also included. The second
objective was to test and compare the two branch-switching procedures presented
in Section 5.3. For all instability problems included in the present set of examples,
the mode injection method worked very well and convergence to the first point on
the secondary path was obtained within 4–7 iterations. The minimisation procedure
also worked very well for all cases presenting a stable bifurcation. In these cases,
convergence to the secondary path was also obtained within few iterations. However,
as expected, the minimisation procedure failed for cases presenting an unstable
bifurcation.
In order to obtain converged results, the cross-sections have been discretised us-
ing large numbers of integration points. For many examples, fewer points would
certainly have given the same numerical results. No attempt of optimising the dis-
cretisation of the cross-sections has been performed. However, the tests performed
by the author have shown that in order to capture accurately the spreading of the
plasticity and get converged load–displacements curves, a large number of points is
often needed.

9.1 Example 1, channel-section beam

The example presented in Figure 9.1 was introduced by Gruttmann et al. [41]. It
refers to a channel-section beam, clamped at one end and subjected to a tip force at
the free end. An ideal elasto-plastic material with yield stress σY = 36 is considered.
The results obtained using 30 ptw3d elements and 336 integration points in the
cross-section are shown in Figure 9.2. These results are in very good agreement
with those presented in [41] based on shell elements.

145
 10
x2
E = 21000 G = 8077
x2
x1 x3
x3

30
C O G
P
1
900

1.6
P

Figure 9.1: Channel-section beam.

16

14
elastic
12

10
Applied load

plastic
8

6
Present
4
Gruttmann et al.
2

0
0 50 100 150 200
x2 displacement of point O

Figure 9.2: Channel-section beam: load–displacement diagram.

9.2 Example 2, framed dome

The instability of the framed dome shown in Figure 9.3 was analysed. Two different
load cases were considered. Case 1 consisted of a single vertical force applied at the
top of the dome while case 2 consisted of seven vertical loads of equal magnitude
placed at the top and at the end points of horizontal members. The dome was
modelled using N pt3d elements for each member, i.e. totally 18 × N elements.

146
 1.55
12.57

4.55
0.76
24.38
cross-section

1.22
E = 20690 G = 8830 Et = E/5 σY = 60

Figure 9.3: Framed dome.

For the elastic material, the results computed with N = 5 and N = 8 are presented
in Figures 9.4 and 9.5. Very good agreement with Kouhia [56], who used N = 5,
is observed. It can be noted that there are several bifurcation points along the
fundamental paths. In Figures 9.4 and 9.5, only the fundamental path and the
secondary path branching out at the lowest bifurcation point are represented. For
both loadings, the buckling mode is a rotation around a vertical axis, see Figure 9.6.

200

180

160

140
Applied load

120

100

80

60 present N = 8
40 present N = 5
Kouhia N = 5
20

0
0 2 4 6 8 10 12
Vertical displacement at the top

Figure 9.4: Framed dome: elastic material – case 1.

147
 70
present N = 5
60
present N = 8
50
Kouhia N = 5
Applied load
40

30

20

10

0
0 2 4 6 8 10 12
Vertical displacement at the top

Figure 9.5: Framed dome: elastic material – case 2.

Figure 9.6: Framed dome: elastic material – first buckling mode.

For the elasto-plastic material, the results computed with N = 8 and 10 × 16

integration points in the cross-section are shown in Figures 9.7 and 9.8. For case 1,
the deformations are first concentrated in the upper part of the dome and a limit
point is reached. After that, the load increases again and a bifurcation point is
observed. For case 2, a bifurcation point is quickly reached. In both cases, the
buckling mode is, as in the elastic case, a rotation around a vertical axis. For
case 2, see Figure 9.9, the bifurcation is stable. However, the stable part of the
secondary path is so short that very small increments must be taken in order to
catch it numerically. Consequently, the minimisation procedure is not well suited to
perform branch-switching in that case since it requires a very small load increment.

148
The mode injection method avoids this problem. In order to assure convergence on
the secondary path, this method was implemented by taken as control the rotation
around a vertical axis of the top of the structure. It can also be noted that Figure 9.9
suggests that the structure exhibits a tangent bifurcation.

70

60

50
Applied load

40

30
plastic
20
elastic
10

0
0 2 4 6 8 10
Vertical displacement at the top

Figure 9.7: Framed dome: elasto-plastic material – case 1.

30
plastic
25
elastic

20
Applied load

15

10

0
0 2 4 6 8 10
Vertical displacement at the top

Figure 9.8: Framed dome: elasto-plastic material – case 2.

149
 14.8

14.6

14.4

Applied load 14.2

14

13.8
fundamental
13.6
secondary
13.4

13.2
0.135 0.136 0.137 0.138 0.139 0.14 0.141 0.142 0.143
Vertical displacement at the top

Figure 9.9: Framed dome: elasto-plastic material – bifurcation for case 2.

9.3 Example 3, lateral buckling of an I beam

B
P
P
D

e
t

4.39 h

D = 0.2613 B = 0.1510 t = 0.0123 h = 0.0077

Figure 9.10: Lateral buckling of an I beam.

This example, introduced by Pi and Trahair [82], concerns the elasto-plastic lateral
buckling of the simply supported I beam depicted in Figure 9.10. The rotation
around the axis of the beam is set to zero at both supports, but warping deforma-
tions are permitted. The trilinear elastic-plastic stress-strain relationship shown in
Figure 9.11 is used.

150
 σ E = 2 · 1011
G = 8 · 1010
Es
σY σY = 25 · 107
Es = 6 · 109
E
ε εs = 0.01375
εs

Figure 9.11: Lateral buckling of an I beam: material parameters.

The results shown in Figure 9.12 were computed with 20 ptw3d elements and 168
integration points in the cross-section. Very good agreement with Reference [82]
is obtained. It can be noted that, due to the absence of hardening, the perfect
structure (e = 0) presents an unstable bifurcation.

130

120

110
Applied load x 1000

100

90
e=0
80
e = 0.001

70 Pi & Trahair

60
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4
Central twist rotation

Figure 9.12: Lateral buckling of an I beam: post-buckling behaviour.

151
9.4 Example 4, lateral buckling of a T cantilever

P
F
P O

6
60
6
1200

60

E = 70000 ν = 0.33 Et = E/10 σY = 500

Figure 9.13: Lateral buckling of a T cantilever.

The lateral buckling of the T cantilever depicted in Figure 9.13 were analysed.
Imperfections were introduced through a small lateral force F = P/1000. The beam
was modelled using 30 ptw3d elements and 288 integration points in the cross-
section. For this example, elastic shell and solid analyses were also performed.
The elastic shell model was constructed using the flat facet triangular co-rotational
element developed by Pacoste [68, 71]. A (10 + 10) × 150 mesh with 6000 triangular
elements has been used. The solid analysis was performed with the finite element
package SOLVIA. The mesh consisted of (8 + 8) × 60 = 960 isoparametric 20 noded
elements.
The bifurcation loads obtained with the different models are listed in Table 9.1.
For the solid analysis, the critical loads were calculated by applying very small
load increments and by checking the number of negative eigenvalues of the tangent
stiffness matrix at each converged state. The post-bifurcation paths for elastic and
elasto-plastic materials are presented in Figures 9.14 to 9.17.
With regard to the bifurcation loads, very good agreement between beam and solid
analyses is obtained. It can also be noted that the values obtained with elastic and
the elasto-plastic materials are almost similar. This unusual result can be explained
by the fact that at the bifurcation, the zone of plastic deformations, situated at
the lower part of the left end of the beam, is very small. Therefore, these plastic
deformations have almost no influence on the lateral buckling of the beam. Their
effect is, compared to the elastic case, to increase the bending pre-buckling deflec-
tions. However, it can be observed that in elasto-plasticity, a limit point is rapidly
reached along the secondary path. This shows that the elasto-plastic structure has,
as expected, a lower load carrying capacity than the elastic one.

152
The elasto-plastic post-bifurcation behaviours obtained with the beam and 3D mod-
els are similar up to a load of 3500. After this point, the curves diverge, which
indicates that the structure does no longer deform as a beam. This is confirmed by
Figure 9.18 which shows that although the structure is very slender, the deforma-
tions are concentrated in the first third part, near the clamped end. It can also be
noted that the beam analysis gives an accurate estimate of the maximum load (2%
difference from the solid model), which, from an engineering point of view, is of first
importance.
For the elastic case, the results obtained with the shell model are within a 4%
difference from those obtained using the beam or solid models. The reason for such
a difference is that using shell elements for modelling a T connection leads to an
overestimation of the cross-section. Consequently, the behaviour of the structure in
pure bending (fundamental path) is stiffer than in reality, see Figure 9.14.

Table 9.1: Lateral buckling of a T cantilever: bifurcation loads.

material beam 3D shell

elastic 3390 3375 3258
elasto-plastic 3400 3395

46

44

42
Load P / 100

40

38

36
beam F = 0
34 beam F = P / 1000

32
shell F = 0
solid F = P / 1000
30
100 120 140 160 180 200 220 240 260
Vertical displacement of point O

Figure 9.14: Lateral buckling of a T cantilever: elastic case.

153
 46
beam F = 0
44
beam F = P / 1000
42
shell F = 0
Load P / 100 40
solid F = P / 1000
38

36

34

32

30
0 10 20 30 40 50
Lateral displacement of point O

Figure 9.15: Lateral buckling of a T cantilever: elastic case.

40

39
beam F = 0
38
beam F = P / 1000
37
solid F = P / 1000
Load P / 100

36

35

34

33

32

31

30
100 150 200 250 300
Vertical displacement of point O

Figure 9.16: Lateral buckling of a T cantilever: elasto-plastic case.

154
 40
beam F = 0
39
beam F = P / 1000
38
solid F = P / 1000
37
Load P / 100

36

35

34

33

32

31

30
0 10 20 30 40 50 60
Lateral displacement of point O

Figure 9.17: Lateral buckling of a T cantilever: elasto-plastic case.

Figure 9.18: Lateral buckling of a T cantilever: elasto-plastic case – deformations

for P = 3500.

155
9.5 Example 5, buckling by torsion and flexure

6
y F
z P z
x O O

60
6
1800

60

E = 70000 ν = 0.33 Et = E/5 σY = 20

Figure 9.19: Buckling by torsion and flexure.

The buckling by torsion and flexure of the beam depicted in Figure 9.19 was anal-
ysed. All degrees of freedom including warping are set to zero at the left end and
only the axial displacement is permitted at the right end. In order to introduce
imperfections, a small lateral force F = P/1000 is applied in the midspan. The
beam model consists of 40 ptw3d elements and 288 integration points in the cross-
section. Elastic shell and solid analyses have also been performed, using the same
discretisations as in the previous example.
The bifurcation loads obtained with the different models are listed in Table 9.2. It
can be observed that the critical loads obtained using the different models differ by
less than 1%.
The load–displacement curves for elastic and elasto-plastic materials are presented
in Figures 9.20 to 9.24. In elasticity, good agreement between the three models
is obtained until a displacement in the z-direction of about 220. After that, due
to the apparition of local deformations, see Figure 9.22, the shell and solid load–
displacement curves present limit points. In the elasto-plastic case, the post-buckling
behaviour is dominated by the bending mode, and very good agreement between
beam and solid models is obtained.

Table 9.2: Buckling by torsion and flexure: bifurcation loads/10000.

material theory [97] beam 3D shell

elastic 8.452 8.444 8.405 8.482
elasto-plastic 1.834 1.845

156
 10

9.5

9
Load P / 10000

8.5

8
beam F = 0
beam F = P / 1000
7.5 shell F = 0
solid F = P / 1000
7
0 5 10 15 20 25 30 35 40
y−displacement of point O

Figure 9.20: Buckling by torsion and flexure: elastic case.

10

9.5

9
Load P / 10000

8.5

8
beam F = 0
beam F = P / 1000
7.5 shell F = 0
solid F = P / 1000
7
0 50 100 150 200 250 300
z−displacement of point O

Figure 9.21: Buckling by torsion and flexure: elastic case.

157
Figure 9.22: Buckling by torsion and flexure: elastic case – deformations after the
limit point (P = 91000) – solid analysis.

2.8

2.6

2.4

2.2
Load P / 10000

1.8
beam F = 0
1.6
beam F = P / 1000
1.4
solid F = P / 1000
1.2

1
0 1 2 3 4 5
y−displacement of point O

Figure 9.23: Buckling by torsion and flexure: elasto-plastic case.

158
 2.8

2.6

2.4

2.2
Load P / 10000

1.8
beam F = 0
1.6
beam F = P / 1000
1.4
solid F = P / 1000
1.2

1
0 10 20 30 40 50 60
z−displacement of point O

Figure 9.24: Buckling by torsion and flexure: elasto-plastic case.

159
Chapter 10

Conclusions and future research

In this last chapter, conclusions about the numerical computations performed in

Chapters 6 to 9 and suggested directions for future research are presented.

10.1 Conclusions

The purpose of this thesis was to implement efficient co-rotational beam elements
and branch-switching procedures in order to study elastic and elasto-plastic insta-
bility problems. These elements and procedures, presented in details in Chapters 3
to 5 have been tested through a large numbers of examples in Chapters 6 to 9. From
the analysis of these numerical applications, several conclusions can be stated.

2D beam element formulation

The co-rotational approach for 2D beam elements is a well known topic and the
objective of the present work was to compare three local elasto-plastic elements.
Both Timoshenko and Bernoulli elements gave close numerical results for frames
consisting of slender beams. This observation was expected since in such cases the
shear deformations can be neglected. It is therefore recommended to use Bernoulli
elements for these structures since the resolution of the constitutive equations at
each Gauss point does not require any implicit process.
Only small variations have been observed between load–displacement diagrams com-
puted with seven and fifteen integration points along the beam depth. From an
engineering point of view, these differences are negligible and a small number of
Gauss points should be used for the analysis of large structures.
Concerning Bernoulli elements, the shallow arch approach should be preferred to
the linear one since identical numerical results can be obtained with fewer elements.
This point, already observed in [73] for elastic structures, was not evident in elasto-
plasticity: by decreasing the number of elements, the number of sections in the
structure where Gauss integrations are performed is also reduced, and consequently,

161
zones of plastification or elastic unloading are modelled with lower accuracy. How-
ever, this does not seem to have any significant effect on load–displacement diagrams.

3D beam element formulation

Compared to Pacoste and Eriksson [73], the 3D co-rotational formulation developed

in Chapter 4 presents new items concerning the parameterisation of the rotations,
the definition of the local frame, the inclusion of warping effects through the intro-
duction of a seventh nodal degree of freedom and the consideration of rigid links. All
these improvements have been tested in a large number of examples in Chapters 7
and 9. The present results have been compared with results found in the literature
and also with analyses performed using shell and solid elements. It can then be
concluded that the elements developed in this thesis can be used to study any large
displacements and rotations problems involving non-linear coupling between axial,
bending and torsional-warping effects. In addition, any cross-section shape can be
considered. In fact, concerning the element formulation, the only restriction resides
in the small strain assumption.
One of the main properties of the co-rotational formulation adopted in this thesis, is
its independence of the local strains assumptions used to derive local internal force
vectors and tangent stiffness matrices. Using this property most authors use linear
strains at the level of the local element definition. However, as shown by several
of the numerical examples included in Chapter 7, this type of elements may give
incorrect results for problems where torsional effects are important.
For these reasons, the displacement field proposed by Gruttmann et al. [41] has
been adopted and used to construct a consistent second order approximation to
the Green-Lagrange strains, at the level of the local element frame. The resulting
expressions have been further simplified by neglecting the non-linear terms in the
expression of shear strains ε12 and ε13 and in the bending curvatures. In fact it
appears that the only important non-linear term is the Wagner term, i.e. 12 r2 ϑ21,1 ,
in the expression of ε11 . Some authors (see e.g. [55]) replace this term by its average
value over the cross-section. The tests performed by the author have shown that
this simplification is not acceptable.
Another characteristic of the co-rotational approach is that the material non-linearity
is only present at the level of the local deformations and does not affect the co-
rotational framework. Hence, due to the simplicity of the local strains expression,
the derivation of elasto-plastic local elements is rather easy. Using linear interpo-
lations, Timoshenko elements have been developed. Based on the work done in
Section 4.3.1, Bernoulli elements can also be derived without difficulties.

Branch-switching in elasto-plasticity

Two methods of branch-switching in elasto-plasticity have been presented in Section

5.3. In the first one, the bifurcation point is isolated by successive bisections and
the branch-switching is operated by using the eigenvector associated to the negative

162
eigenvalue. In the second one, introduced by Petryk [76–81], an energy approach is
used to select automatically the stable post-bifurcation path.
In most studied examples in Chapters 6 and 9, both methods work very well and
convergence to the first point of the secondary path is obtained after 4–7 iterations.
However, if the bifurcation is unstable (examples 6.3 and 9.3) or if the stable part
of the secondary path is very small (example 9.2), the minimisation procedure can
not be employed. On the other hand, if an arc-length procedure is used, the mode
injection method does not ensure convergence to the secondary path, in particular for
structures exhibiting a tangent bifurcation (examples 6.5 and 9.2). The conclusion of
the author is that the best alternative is to use the mode injection method together
with a displacement control procedure and to take as control a displacement or a
rotation which is zero along the fundamental path.

Direct computation of elastic critical points

In Section 5.2, a new procedure for the direct computation of elastic critical points
has been presented. Compared to the approach of Wriggers et al. [104, 105], two
new ideas have been introduced. First, the condition of criticality is expressed
by a scalar equation. Next, the present method combines iterations based on the
extended system obtained by the equilibrium equations and the criticality condition
with iterations based only on the equilibrium equations under load or displacement
control.
The performances of the present method and the classical approach [104, 105] have
been compared through eight numerical examples in Chapter 8. In most of the
cases, both approaches give similar results. However, if the non-linearity between
the initial point and the critical one is important, as in examples 8.1, 8.6 and 8.7, the
convergence properties of the present method are better. It can then be concluded
that the numerical procedure presented in this thesis is an interesting alternative
for the direct computation of critical points.

10.2 Future research

The work done in this thesis can be extended in three directions which are now
briefly presented.

Element formulation

The co-rotational formulation presented in this thesis is based on a spatial repre-

sentation of the rotations, see Section 4.1. Another possibility is to derive elements
using a material representation. As pointed out by Ibrahimbegovic [51], a material
approach is more suitable in the context of dynamics and time-stepping schemes.

163
For a certain class of problems, the small strains hypothesis may not be correct and
local elements based on large strains should be developed. Concerning co-rotational
2D beams and shells elements, such an approach was adopted by Kolahi [53].

Elasto-plastic models

The elasto-plastic model adopted in this thesis is based on an incremental flow

theory, a von-Mises yield criterion, a bilinear uniaxial stress-strain relation and an
isotropic hardening. However, a bilinear law is too crude to model correctly the
real behaviour of materials, in particular metals. A better approach, common in
finite elements codes, is to represent the plastic uniaxial behaviour by a succession
of segments. Such a modification would allow a more accurate description of the
post-buckling behaviour. As a matter of fact, the analysis of Hutchinson’s model
in Section 2.1.2 has shown that the decreasing stiffness in the plastic range has
destabilising effects and a significant influence on the post-bifurcation.
In order to model correctly plastic reversal deformations, kinematical and mixed
hardening should also be incorporated. The extension to other yield criteria such
as Tresca or Mohr-Coulomb can also be considered. However, it is not certain that
the bifurcation procedure based on the minimisation procedure holds in such cases.
Since these criteria are not characterised by a smooth yield surface, the condition
(2.97) must be investigated, which may be difficult.
Finally, the main interesting development would probably be to develop elements
based on a plastic deformation theory and to compare their performances with the
performances of the present elements in predicting elasto-plastic buckling loads.

New finite elements

The present work can also be extended to composite beam or shell elements. Con-
cerning shell elements, a co-rotational approach, similar to the one adopted for beam
elements, has been developed by Pacoste [68,71]. The main challenge concerns then
the local formulation. Many alternatives are available and the choice of an efficient
local strain definition is open. It would also be interesting to test the two elasto-
plastic branch-switching methods and the new procedure for the direct computation
of elastic critical points in a shell context.

164
Bibliography

[1] J.H. Argyris, B. Boni, U. Hindenlang, and M. Kleiber. Finite element analysis
of two- and three-dimensional elasto-plastic frames—the natural approach.
Comput. Methods Appl. Mech. Engrg., 35:221–248, 1982.

[2] S. Baba and T. Kajita. Plastic analysis of torsion of a prismatic beam. Int.
J. Numer. Meth. Engng., 18:927–944, 1982.

[3] J.L. Batoz and G. Dhatt. Modélisation Des Structures Par Eléments Finis,
Volume 1: Solides Elastiques. Hermès, Paris, 1990. In French.

[4] J.L. Batoz and G. Dhatt. Modélisation Des Structures Par Eléments Finis,
Volume 2: Poutres et Plaques. Hermès, Paris, 1990. In French.

[5] J.-M. Battini. Plastic Instability Analysis of Plane Frames Using a Co-
Rotational Approach. Licentiate Thesis, Department of Structural Engineer-
ing, KTH, Stockholm, 1999.

[6] J.-M. Battini and C. Pacoste. Co-rotational beam elements with warping
effects in instability problems. Accepted by Comput. Methods Appl. Mech.
Engrg.

[7] J.-M. Battini and C. Pacoste. Eléments poutres co-rotationels avec gauchisse-
ment. Proceedings CSMA 5ème colloque national en calcul des structures,
Giens 2001. In French.

[8] J.-M. Battini and C. Pacoste. Minimal augmentation procedure for the direct
computation of critical points. Submitted to Comput. Methods Appl. Mech.
Engrg.

[9] J.-M. Battini and C. Pacoste. Plastic instability of beam structures using
co-rotational elements. Submitted to Comput. Methods Appl. Mech. Engrg.

[10] Z.J. Bazant. Stability of Structures. Oxford University Press, 1991.

[11] B. Budianski. A reassessment of deformation theories of plasticity. J. Appl.

Mech., 26:259–264, 1959.

[12] A. Cardona and A. Huespe. Evaluation of simple bifurcation points and post-
critical path in large finite rotation problems. Comput. Methods Appl. Mech.
Engrg., 175:137–156, 1999.

165
[13] C.N. Chen. Study on the reliability of Shanley’s plastic buckling model by
using the finite element method. Proceedings of the 1994 Pressure Vessels and
Piping Conference, 277:127–134, 1994.
[14] N.C. Chen. A finite element study on Shanley’s plastic buckling theory of
short columns. Computers and Structures, 59(5):975–982, 1996.
[15] Y. Cheng, H. Fang, and W. Lu. Imperfection sensitivity analysis of a rect-
angular column compressed into the plastic range. Applied Mathematics and
Mechanics, 19(1):9–14, 1998.
[16] Y. Cheng, W. Lu, and H. Fang. Plastic post-buckling of a simply supported
column with a rectangular cross-section. Applied Mathematics and Mechanics,
16(8):713–722, 1995.
[17] C. Cichon. Large displacements in-plane analysis of elastic-plastic frames.
Computers and Structures, 19:737–745, 1984.
[18] A. Cimetière, A. Elkoulani, and A. Léger. Buckling and post-buckling of
an elastic-plastic simple model. Comptes rendus de l’académie des sciences,
Paris, 319:1263–1269, 1994.
[19] A. Cimetière and A. Léger. Some problems about elastic-plastic post-buckling.
Int. J. Solids Structures, 33(10):1519–1533, 1996.
[20] A. Considère. Résistance des pièces comprimées. In Paris Librairie Polytech-
nique, editor, Congrès International Des Procédés de Construction, 1891. In
French.
[21] R.D. Cook, D.S. Malkus, and M.E. Plesha. Concepts and Applications of
Finite Element Analysis. Wiley, Chichester, third edition, 1989.
[22] M.A. Crisfield. A consistent co-rotational formulation for non-linear, three-
dimensional, beam-elements. Comput. Methods Appl. Mech. Engrg., 81:131–
150, 1990.
[23] M.A. Crisfield. Non-Linear Finite Element Analysis of Solids and Structures,
Volume 1: Essentials. Wiley, Chichester, 1991.
[24] M.A. Crisfield. Non-Linear Finite Element Analysis of Solids and Structures,
Volume 2: Advanced Topics. Wiley, Chichester, 1997.
[25] M.A. Crisfield and D. Tan. A low-order large-strain quadrilateral shell element.
In European Congress on Computational Methods in Applied Sciences and
Engineering, Barcelona, 11-14 September 2000. ECCOMAS 2000.
[26] R. de Borst. Computation of post-bifurcation and post-failure behavior of
strain-softening solids. Computers and Structures, 25(2):211–224, 1987.
[27] V. de Ville de Goyet. L’analyse Statique Non Linéaire Par la Méthode Des
Éléments Finis Des Structures Spatiales Formées de Poutres a Section Non
Symétrique. PhD thesis, Université de Liège, 1988-1989.

166
[28] R.H. Dodds. Numerical techniques for plasticity computations in finite element
analysis. Computers and Structures, 26:767–779, 1987.

[29] M.S. El Naschie. Stress, Stability and Chaos in Structural Engineering: An

Energy Approach. McGraw-Hill, 1990.

[30] A. Eriksson. Fold lines sensitivity analyses in structural instability. Comput.

Methods Appl. Mech. Engrg., 114:77–101, 1994.

[31] A. Eriksson. Equilibrium subsets for multi-parametric structural analysis.

Comput. Methods Appl. Mech. Engrg., 140:305–327, 1997.

[32] A. Eriksson. Structural instability analyses based on generalised path-

following. Comput. Methods Appl. Mech. Engrg., 156:45–74, 1998.

[33] A. Eriksson and C. Pacoste. Solution surfaces and generalised paths in non-
linear structural mechanics. Int. J. Struct. Stability and Dynamic, 1(1), 2001.

[34] A. Eriksson, C. Pacoste, and A. Zdunek. Numerical analysis of complex in-

stability behaviour using incremental-iterative strategies. Comput. Methods
Appl. Mech. Engrg., 179:265–305, 1999.

[35] C.A. Felippa. A systematic approach to the element-independent corotational

dynamics of finite elements. In Int. Conf. Computational Methods for Shell
and Spatial Structures, Papadrakakis, M., Samartin, A. and Onate, E. (Eds.)
ISASR-NTUA, Athens, Greece 2000. IASS-IACM 2000.

[36] R. Fletcher. Practical Methods of Optimization. Wiley, second edition, 1987.

[37] K. Forsell. Instability Analyses of Structures under Dynamic Loads. PhD

thesis, Department of Structural Engineering, KTH, Stockholm, 2000.

[38] Y. Goto, T. Yoshimitsu, and M. Obata. Elliptic integral solutions of plane

elastica with axial and shear deformations. Int. J. Solids Structures, 26(4):375–
390, 1990.

[39] F. Gruttmann, R. Sauer, and W. Wagner. Shear stresses in prismatic beams

with arbitrary cross-sections. Int. J. Numer. Meth. Engng., 45:865–889, 1999.

[40] R. Gruttmann, R. Sauer, and W. Wagner. A geometrical nonlinear eccentric

3d-beam element with arbitrary cross-section. Comput. Methods Appl. Mech.
Engrg., 160:383–400, 1998.

[41] R. Gruttmann, R. Sauer, and W. Wagner. Theory and numerics of three-

dimensional beams with elastoplastic behaviour. Int. J. Num. Meth. Engng.,
48:1675–1702, 2000.

[42] W. Guggenberger. Nonlinear analysis of space frame structures–large rotation

formulations and their consistent approximations for moderate rotations. In
European Congress on Computational Methods in Applied Sciences and Engi-
neering, Barcelona, 11-14 September 2000. ECCOMAS 2000.

167
[43] B. Haugen and C.A. Felippa. A unified formulation of co-rotational finite
elements: I. theory. Technical Report CU-CAS-95-06, Center for Aerospace
Structures, Dept of Aerospace Engineering, U. Colorado, Boulder.

[44] B. Haugen and C.A. Felippa. A unified formulation of co-rotational fi-

nite elements: II. applications. Technical Report CU-CAS-95-06, Center for
Aerospace Structures, Dept of Aerospace Engineering, U. Colorado, Boulder.

[45] R. Hill. A general theory of uniqueness and stability in elastic-plastic solids.

J. Mech. Phys. Solids, 6:236–249, 1958.

[46] R. Hill and M. J. Sewell. A general theory of inelastic column failure. J. Mech.
Phys. Solids, 8:105–111, 1960.

[47] K.M. Hsiao and W.Y. Lin. A co-rotational finite element formulation for
buckling and postbuckling analyses of spatial beams. Comput. Methods Appl.
Mech. Engrg., 188:567–594, 2000.

[48] J.W. Hutchinson. Imperfection sensitivity in the plastic range. J. Mech. Phys.
Solids, 21:191–204, 1973.

[49] J.W. Hutchinson. Post-bifurcation behaviour in the plastic range. J. Mech.

Phys. Solids, 21:163–190, 1973.

[50] J.W. Hutchinson. Plastic buckling. Advances in Appl. Mech., 14:67–144, 1974.

[51] A. Ibrahimbegovic. On the choice of finite rotation parameters. Comput.

Methods Appl. Mech. Engrg., 149:49–71, 1997.

[52] A. Ibrahimbegovic and M. Al Mikad. Quadratically convergent direct calcu-

lation of critical points for 3d structures undergoing finite rotations. Comput.
Methods. Appl. Mech. Engrg., 189:107–120, 2000.

[53] A.S. Kolahi. Co-Rotational Methods for Small and Large Strain Beams, Sheets
and Shells. PhD thesis, Dept. of Aeronautics, Imperial College, London, 1998.

[54] K. Kondoh and N. Atluri. Large-deformation, elasto-plastic analysis of frames

under nonconservative loading, using explicitly derived tangent stiffness based
on assumed stresses. Comp. Mech., 2:1–25, 1987.

[55] R. Kouhia. On kinematical relations of spatial framed structures. Computers

and Structures, 40(5):1185–1191, 1991.

[56] R. Kouhia and M. Tuomala. Static and dynamic analysis of space frames using
simple Timoshenko type elements. Int. J. Num. Meth. Engng., 36:1189–1221,
1993.

[57] S.H. Lee. Rudimentary considerations for effective line search method in non-
linear finite element analysis. Computers and Structures, 32(6):1287–1301,
1989.

168
[58] A. Léger and M. Potier-Ferry. Sur le flambage plastique. J. Mécanique Théor.
Appl., 7:819–857, 1988. In French.

[59] T. Lidström. Computational Methods for Finite Element Instability Analyses.

PhD thesis, Department of Structural Engineering, KTH, Stockholm, 1996.

[60] K. Moon-Young, C. Sung-Pil, and K. Sung-Bo. Spatial stability analysis of

thin-walled space frames. Int. J. Num. Meth. Engng., 39:499–525, 1996.

[61] A. Needleman and V. Tvergaard. An analysis of the imperfection sensitivity

of square elastic-plastic plates under compression. Int. J. Solids Structures,
12:185–201, 1976.

[62] Q.S. Nguyen. Bifurcation and postbifurcation analysis in plasticity and brittle
fracture. J. Mech. Phys. Solids, 33:303–324, 1987.

[63] Q.S. Nguyen. Bifurcation and stability of time-independent standard dissipa-

tive systems. In Q.S. Nguyen, editor, CISM Courses and Lectures No. 327 :
Bifurcation and Stability of Dissipative Systems. Springer-Verlag, 1993.

[64] B. Nour-Omid and C.C. Rankin. Finite rotation analysis and consistent lin-
earization using projectors. Comput. Methods Appl. Mech. Engrg., 93:353–384,
1991.

[65] A. Olsson. Object-Oriented Finite Element Algorithms. Licenciate Thesis,

Department of Structural Engineering, KTH, Stockholm, 1997.

[66] M. Ortiz and E. P. Popov. Accuracy and stability of integration algorithms for
elastoplastic constitutive relations. Int. J. Num. Meth. Engng., 21:1561–1576,
1985.

[67] D.R.J. Owen and E. Hinton. Finite Elements in Plasticity: Theory and Prac-
tice. Pineridge Press Limited, Swansea UK, 1980.

[68] C. Pacoste. Co-rotational flat facet triangular elements for shell instability
analysis. Comput. Methods Appl. Mech. Engrg., 156:75–110, 1998.

[69] C. Pacoste and J.-M. Battini. Calcul des chemins d’équilibres post-critiques.
Proceedings CSMA 5ème colloque national en calcul des structures, Giens
2001. In French.

[70] C. Pacoste, J.-M. Battini, and A. Eriksson. Parameterisation of rotations in

co-rotational elements. Proceedings Euromech 2000, Metz.

[71] C. Pacoste and A. Eriksson. Element formulation and numerical techniques

for stability problems in shells. Accepted by Comput. Methods Appl. Mech.
Engrg.

[72] C. Pacoste and A. Eriksson. Element behaviour in post-critical plane frame

analysis. Comput. Methods Appl. Mech. Engrg., 125:319–343, 1995.

169
[73] C. Pacoste and A. Eriksson. Beam elements in instability problems. Comput.
Methods Appl. Mech. Engrg., 144:163–197, 1997.
[74] M.S. Park and B.C. Lee. Geometrically non-linear and elastoplastic three-
dimensional shear flexible beam element of von-Mises-type hardening material.
Int. J. Num. Meth. Engng., 39:383–408, 1996.
[75] U. Perego. Explicit backward difference operators and consistent predictors for
linear hardening elastic-plastic constitutive laws. SM Archives, 13(2):65–102,
1988.
[76] H. Petryk. The energy criteria of instability in time-independent inelastic
solids. Arch. Mech., 43:519–545, 1991.
[77] H. Petryk. Theory of bifurcation and instability in time-independent plasticity.
In Q.S. Nguyen, editor, CISM Courses and Lectures No. 327—Bifurcation and
Stability of Dissipative Systems, pages 95–152. Springer-Verlag, 1993.
[78] H. Petryk. Instability of plastic deformation processes. In T. Tatsumi et Al.,
editor, Theoretical and Applied Mechanics, Proc.XIXth IUATAM Congress,
Kyoto, pages 497–516. Elsevier, Amsterdam, 1997.
[79] H. Petryk and K. Thermann. Second-order bifurcation in elastic-plastic solids.
J. Mech. Phys. Solids, 33(6):577–593, 1985.
[80] H. Petryk and K. Thermann. On discretized plasticity problems with bifurca-
tions. Int. J. Solids Structures, 29(6):745–765, 1992.
[81] H. Petryk and K. Thermann. On post-bifurcation paths in plastic deforma-
tion process. In D.R.J. Owen, editor, Computational Plasticity. Fundamentals
and Applications—Part I. Proceedings of the Third Int. Conf., pages 593–604.
Pineridge Press, 1992.
[82] Y.L. Pi and N.S. Trahair. Nonlinear inelastic analysis of steel beam-columns:
II: Applications. Journal of Structural Engineering, 120(7):2062–2085, 1994.
[83] C.C. Rankin and F.A. Brogan. An element-independent co-rotational proce-
dure for the treatment of large rotations. ASME J. Pressure Vessel Tchn.,
108:165–174, 1986.
[84] C.C. Rankin and B. Nour-Omid. The use of projectors to improve finite
element performance. Computers and Structures, 30:257–267, 1988.
[85] M. Ristinmaa and J. Tryding. Exact integration of constitutive equations in
elasto-plasticity. Int. J. Num. Meth. Engng., 36:2525–2544, 1993.
[86] R. Sauer. Eine Einheitliche Finite-Element-Formulierung Für Stab- und Scha-
lentragwerke mit Endlichen Rotationen. PhD thesis, Bericht Nr. 4, Institut
für Baustatik, Universität Karlsruhe, 1998. In German.
[87] M.J. Sewell. A Survey of Plastic Buckling. Solid Mechanics Division, Univer-
sity of Waterloo, Waterloo, Ontario, Canada, 1972.

170
 [88] F.R. Shanley. Inelastic column theory. J. Aero. Science, 14:261–267, 1947.
[89] J.C. Simo, J.G. Kennedy, and S. Govindjee. Non-smooth multisurface plas-
ticity and viscoplasticity. Loading/unloading conditions and numerical algo-
rithms. Int. J. Num. Meth. Engng., 26:2161–2185, 1988.
[90] J.C. Simo and R.L. Taylor. Consistent tangent operators for rate-independent
elastoplasticity. Comput. Methods Appl. Mech. Engrg., 48:101–118, 1985.
[91] J.C. Simo and R.L. Taylor. A return mapping algorithm for plane stress
elastoplasticity. Int. J. Num. Meth. Engng., 22:649–670, 1986.
[92] J.C. Simo and L. Vu-Quoc. A three-dimensional finite-strain rod model. part
II: Computational aspects. Comput. Methods Appl. Mech. Engrg., 58:79–116,
1986.
[93] J.C. Simo and L. Vu-Quoc. A geometrically-exact rod model incorporating
shear and torsion-warping deformation. Int. J. Solids Structures, 27(3):371–
393, 1991.
[94] X.M. Su and W. Lu. Postbuckling and imperfection sensitivity analysis of
structures in the plastic range—part 1: Model analysis. Thin-Walled Struc-
tures, 10:263–275, 1990.
[95] J.G. Teigen. Nonlinear Analysis of Concrete Structures Based on a 3D Shear-
Beam Element Formulation. PhD thesis, Department of Mathematics, Me-
chanics Division, University of Oslo, Norway, 1994.
[96] J.M.T. Thompson and G.W. Hunt. Elastic Instability Phenomena. Wiley,
Chichester, 1984.
[97] S.P. Timoshenko and J.M. Gere. Theory of Elastic Stability. Second Edition,
McGraw-Hill, Tokyo, 1961.
[98] N. Triantafyllidis. On the bifurcation and postbifurcation analysis of elastic-
plastic solids under general prebifurcation conditions. J. Mech. Phys. Solids,
31(6):499–510, 1983.
[99] A.M.A. van der Heijden. A study of Hutchinson’s plastic buckling model. J.
Mech. Phys. Solids, 27:441–464, 1979.
[100] H. van der Veen. The Significance and Use of Eigenvalues and Eigenvectors
in the Numerical Analysis of Elasto-Plastic Soils. PhD thesis, Faculty of Civil
Engineering and Geosciences, Delft University of Technology, Netherlands,
1998.
[101] H. van der Veen, K. Vuik, and R. de Borst. Branch switching techniques
for bifurcation in soil deformation. Comput. Methods Appl. Mech. Engrg.,
190:707–719, 2000.
[102] T. von Karman. Untersuchungen über Knickfestigkeit. In Verein Deutscher
Ingenieure, editor, Mitteilungen über Forschungsarbeiten, 1910. In German.

171
[103] W. Wagner and F. Gruttmann. Finite element analysis of Saint-Venant torsion
problem with exact integration of the elastic-plastic constitutive equations.
Comput. Methods Appl. Mech. Engrg., 190:3831–3848, 2001.

[104] P. Wriggers and J.C. Simo. A general procedure for the direct computation of
turning and bifurcation points. Int. J. Num. Meth. Engrg., 30:155–176, 1990.

[105] P. Wriggers, W. Wagner, and C. Miehe. A quadratically convergent proce-

dure for the calculation of stability points in finite element analysis. Comput.
Methods Appl. Mech. Engrg., 70:329–347, 1988.

172
Appendix A

3D elastic local formulation

This appendix presents the Maple procedures necessary in order to generate the
Fortran code for the 3D local elastic elements. These elements have been developed
in Section 4.3.

A.1 t3d element

with(linalg):
du:=u12/L:
t1:=(t11+t12)/2:
t2:=(t21+t22)/2:
t3:=(t31+t32)/2:
dt1:=(t12-t11)/L:
dt2:=(t22-t21)/L:
dt3:=(t32-t31)/L:
g12:=-t3:
g13:=t2:
k1:=dt1:
k2:=-dt3:
k3:=dt2:
Phi1:=A*du^2+I22*k2^2+I33*k3^2+Irr/4*dt1^4+Io*du*dt1^2:
Phi2:=A*(g12^2+g13^2)+J*k1^2:
Phi:=1/2*L*(E*Phi1+G*Phi2):
fl:=grad(Phi,[u,t11,t21,t31,t12,t22,t32]):
kl:=hessian(Phi,[u,t11,t21,t31,t12,t22,t32]):
fortran(fl,optimized);
fortran(kl,optimized);

173
A.2 tw3d element

with(linalg):
du:=u12/L:
a:=(a1+a2)/2:
da:=(a2-a1)/L:
t1:=(t11+t12)/2:
t2:=(t21+t22)/2:
t3:=(t31+t32)/2:
dt1:=(t12-t11)/L:
dt2:=(t22-t21)/L:
dt3:=(t32-t31)/L:
g12:=-t3:
g13:=t2:
k1:=dt1:
k2:=-dt3+c3*da:
k3:=dt2-c2*da:
Phi1:=A*du^2+I22*k2^2+I33*k3^2+Irr/4*dt1^4+Iw*da^2+Io*du*dt1^2:
Phi2:=2*I23*k2*k3+I2r*k2*dt1^2+I3r*k3*dt1^2+Iwr*da*dt1^2:
Phi3:=A*(g12^2+g13^2)+Io*k1^2+(Io-J)*a^2-2*(Io-J)*k1*a:
Phi:=1/2*L*(E*Phi1+E*Phi2+G*Phi3):
fl:=grad(Phi,[u,t11,t21,t31,t12,t22,t32,a1,a2]):
kl:=hessian(Phi,[u,t11,t21,t31,t12,t22,t32,a1,a2]):
fortran(fl,optimized);
fortran(kl,optimized);

A.3 b3d element

with(linalg):
u21:=0:
u31:=0:
u22:=0:
u32:=0:
f1:=1-3*(x/L)^2+2*(x/L)^3:
f2:=x*(1-x/L)^2:
f3:=1-f1:
f4:=x^2*(x/L-1)/L:
u2:=f1*u21+f2*t31+f3*u22+f4*t32:
du2:=diff(u2,x):
ddu2:=diff(du2,x):
u3:=f1*u31-f2*t21+f3*u32-f4*t22:
du3:=diff(u3,x):
ddu3:=diff(du3,x):
t1:=(1-x/L)*t11+x/L*t12:

174
dt1:=diff(t1,x):
k1:=dt1:
k2:=-ddu2:
k3:=-ddu3:
eav:=u12/L+1/2/L*int(du2^2+du3^2+Io/A*dt1^2,x=0..L):
Phi1:=A*eav^2+I22*k2^2+I33*k3^2+1/4*(Irr-Io^2/A)*dt1^4:
Phi:=1/2*int(E*Phi1+G*J*k1^2,x=0..L):
fl:=grad(Phi,[u,t11,t21,t31,t12,t22,t32]):
kl:=hessian(Phi,[u,t11,t21,t31,t12,t22,t32]):
fortran(fl,optimized);
fortran(kl,optimized);

A.4 bw3d element

with(linalg):
u21:=0:
u31:=0:
u22:=0:
u32:=0:
f1:=1-3*(x/L)^2+2*(x/L)^3:
f2:=x*(1-x/L)^2:
f3:=1-f1:
f4:=x^2*(x/L-1)/L:
u2:=f1*u21+f2*t31+f3*u22+f4*t32:
du2:=diff(u2,x):
ddu2:=diff(du2,x):
u3:=f1*u31-f2*t21+f3*u32-f4*t22:
du3:=diff(u3,x):
ddu3:=diff(du3,x):
t1:=f1*t11+f2*a1+f3*t12+f4*a2:
dt1:=diff(t1,x):
ddt1:=diff(dt1,x):
k1:=dt1:
k2:=-ddu2+c3*ddt1:
k3:=-ddu3-c2*ddt1:
eav:=u12/L+1/2/L*int(du2^2+du3^2+Io/A*dt1^2,x=0..L):
Phi1:=A*eav^2+I22*k2^2+I33*k3^2+1/4*(Irr-Io^2/A)*dt1^4+Iw*ddt1^2:
Phi2:=2*I23*k2*k3+I2r*k2*dt1^2+I3r*k3*dt1^2+Iwr*ddt1*dt1^2:
Phi3:=G*J*k1^2:
Phi:=1/2*int(E*Phi1+E*Phi2+G*Phi3,x=0..L):
fl:=grad(Phi,[u,t11,t21,t31,t12,t22,t32,a1,a2]):
kl:=hessian(Phi,[u,t11,t21,t31,t12,t22,t32,a1,a2]):
fortran(fl,optimized);
fortran(kl,optimized);

175
Appendix B

Warping function

The purpose of this appendix is to describe the theoretical and numerical aspects
related to the computation of the warping function ω̄ and cross-section quantities
introduced in equation (4.130). The theoretical aspects are taken from Batoz and
Dhatt [3] and Gruttmann et al. [39]. Concerning the finite element implementation,
the description of isoparametric quadratic plane elements can be found e.g. in [21].

B.1 Basic equations

x3

x3
C
M1 c3
x2 x1 x2
x1 c2
G

Figure B.1: Prismatic cantilever beam.

A prismatic cantilever beam, see Figure B.1 with reference axis x1 and section
coordinates x2 and x3 is considered. Following the hypotheses of Section 4.3.1, x2
and x3 intersect at the centroid G and are not necessarily directed along the principal
axes of the cross-section. The displacements of a current point are noted u1 , u2 , u3 .
The linear strain-displacement relations are
ε11 = u1,1 γ12 = u1,2 + u2,1 γ13 = u1,3 + u3,1 (B.1)
Assuming a linear elastic material, the stress-strain relations read
σ11 = E ε11 τ12 = G γ12 τ13 = G γ13 (B.2)

177
Finally, the stress resultants are defined by
 
N= σ11 dA M1 = (τ13 x2 − τ12 x3 ) dA (B.3a)
A A
 
M2 = σ11 x3 dA M3 = − σ11 x2 dA (B.3b)
A A

B.2 Saint-Venant torsion

A torque M1 is applied at the free end of the cantilever, as shown in Figure B.1.
Assuming the torsion angle θ1 small, the displacements of a current point are

u1 = ω θ1,1 u2 = −(x3 − c3 ) θ1 u3 = (x2 − c2 ) θ1 (B.4)

where the warping function ω refers to the shear center C of coordinates c2 , c3 . Using
equations (B.1), the strains can be rewritten as

ε11 = ω θ1,11 γ12 = (ω,2 −x3 + c3 ) θ1,1 γ13 = (ω,3 +x2 − c2 ) θ1,1 (B.5)

The cantilever is free of normal forces and bending moments. This is inserted into
equations (B.3), which gives the normality conditions
  
ω dA = 0 ω x2 dA = 0 ω x3 dA = 0 (B.6)
A A A

B.3 Determination of the warping function

Since the position of the shear center is a priori unknown, the following transforma-
tion is applied

ω̄ = ω − c2 x3 + c3 x2 (B.7)

Introducing the above equation into the first of equations (B.6) gives the constraint

ω̄ dA = 0 (B.8)
A

The shear deformations in equations (B.5) can then be rewritten as

γ12 = (ω̄,2 −x3 ) θ1,1 γ13 = (ω̄,3 +x2 ) θ1,1 (B.9)

The problem is free of external volume or surface forces. The principle of virtual
work, per unit length, can be written as

(τ12 δγ12 + τ13 δγ13 ) dA = 0 (B.10)
A

178
with

δγ12 = θ1,1 δ ω̄,2 δγ13 = θ1,1 δ ω̄,3 (B.11)

By introducing equations (B.2), (B.9) and (B.11) into (B.10), it is obtained

 
(ω̄,2 δ ω̄,2 + ω̄,3 δ ω̄,3 ) dA = (x3 δ ω̄,2 − x2 δ ω̄,3 ) dA (B.12)
A A

This equation is solved numerically by a finite element approach. Isoparametric

quadratic plane elements, see Figure B.2, are used to discretise the cross-section.
This means that the coordinates x2 , x3 and the unknown warping function ω̄ are
interpolated within an element using the same shape functions

8
8
8
x2 = NI (ξ, η) x2 I x3 = NI (ξ, η) x3 I ω̄ = NI (ξ, η) ω̄ I (B.13)
I=1 I=1 I=1

Figure B.2: Isoparametric quadratic plane element.

The element stiffness matrix and load vector are given by

 
KIJ = (NI,2 NJ,2 + NI,3 NJ,3 ) dAe fI = (NI,2 x3 − NI,3 x2 ) dAe (B.14)
Ae Ae

Using the isoparametric transformation, the integrals over each element are per-
formed numerically with four Gauss points. After assembly, the system

Kd = f (B.15)

is solved by setting arbitrarily a nodal value of ω̄ to zero. This provides the vector
d which contains the nodal values of ω̄. The vector d is next transformed using

1
ω̄ := ω̄ − ω̄ dA (B.16)
A A

in order to fulfill the constraint (B.8).

179
B.4 Cross-section quantities

The cross-section quantities defined in equations (4.130), and those introduced in

the present section, are integrated numerically using 4 Gauss points for each isopara-
metric elements.
The coordinates of the shear center are derived from the two last (B.6) equations,
which gives after some work
Iω̄3 I22 − Iω̄2 I23 Iω̄2 I33 − Iω̄3 I23
c2 = − c3 = (B.17)
I22 I33 − I23
2
I22 I33 − I23
2

with
 
Iω̄2 = ω̄ x2 dA Iω̄3 = ω̄ x3 dA (B.18)
A A

The Saint-Venant torsion modulus, defined in equation (4.129), can be calculated

directly by

J = I22 + I33 − dT f (B.19)

with d and f defined in (B.15).

Finally, Iω and Iωr are rewritten with the help of equation (B.7) as

Iω = ω̄ 2 dA − I33 c22 − I22 c32 + 2 I23 c2 c3 (B.20a)
A
Iωr = ω̄r2 dA + I3r c2 − I2r c3 (B.20b)
A

180