Anda di halaman 1dari 26




This chapter provides a methodology and corresponding nomenclature for the

development of the finite element method as a means for solving the governing
equations of structural systems that are susceptible to instabilities. This develop-
ment, presented in Sections 20.1 and 20.2, uses principles and notation that are
not typically encountered in elementary finite element courses but are nonethe-
less of great importance as an introduction to the conventions that are typically
encountered within modern literature on the subject.
With this overview, Section 20.3 presents more practical aspects of stability
analysis by the finite element method. After providing a survey of the dominant for-
mulations behind the finite element buckling procedures employed in commercial
software, comparisons in predicted capacities are made and differences discussed.
Additionally, observations regarding the nature of the first eigenmodes (as obtained
from finite element buckling analyses) are made and their usefulness as devices for
defining seed (initial geometric) imperfections when using incremental nonlinear
finite element analysis methods to study structural stability problems is assessed.


This discussion on the finite element method, as it relates to the evaluation of

stability of equilibrium, takes as its point of departure the integral statement for
the governing equations of motion. While, historically, it has been more common
to employ the principles of virtual work , or equivalently stationarity in the total
potential , when developing the finite element equations for solids and structures
(Bathe, 1996; Cook et al., 2002; McGuire et al., 2000; Zienkiewicz and Taylor
2000a), this treatment focuses on the Galerkin method , as it applies to the weak
form of the governing equations in a structural problem (Reddy, 1993). This
approach is taken in order to equip the reader with concepts that will be useful
in understanding modern research advances concerning the finite element method
(Ciarlet and Lions, 1991; Chapelle and Bathe, 2003; Duarte and Oden, 1996;
Guide to Stability Design Criteria for Metal Structures, Sixth Edition Edited by Ronald D. Ziemian 933
Copyright © 2010 John Wiley & Sons, Inc.

Babuska and Melenk, 1997; Melenk and Babuska, 1996; Belytschko and Black,
1997; Farhat et al., 2003) as well as enhancing an understanding of the method’s
convergence properties (Reddy, 1993; Langtangen, 2000).
It is ultimately the goal of the finite element method to idealize a continuous
system in a discrete form that preserves salient physical responses while at the
same time admitting insight into convergence properties and accuracy. With this in
mind, consider the statement of some homogeneous and continuous model problem
of the form
L (u(x )) = 0 ∀x ∈  (20.1)

where L is a general linear differential operator, u (x ) is the solution of this model

problem, and  is some open set defining the problem domain. In a discrete
numerical solution to the problem posed in Eq. 20.1, an approximate solution of
the form û (x ) will be obtained with1

u ≈ û = ui Ni (x ) (20.2)

subject to the Fourier requirement (Reddy, 1993; Kreyszig, 1978) that u − ûi  →
0 as i → ∞. Given that û is an approximation, it is not reasonable to expect
that Eq. 20.1 will be satisfied by Eq. 20.2 In general L (û (x )) = 0, and thus it is
appropriate to define a residual for the model problem as

R = L (û (x )) (20.3)

In any discrete solution of the model problem, it can only be ensured that the
residual, R, is arbitrarily small, and it is desired that R → 0 ⇒ u − ûi  → 0.
Considering now exclusively the residual , the goal is to minimize approximation
error. With Eq. 20.4 below as an alternate form of Eq. 20.3, an important observa-
tion can be made: because R varies in space, the requirement of a minimum can
only be imposed in an average or integral sense:

R (ui , x ) = L (û (x )) (20.4)

One approach to providing such a minimization involves the least squares

method , wherein the objective is to minimize the average square of the residual,
which can be expressed by
min R2 d  w.r.t. ui (20.5)

Because Eq. 20.5 describes a discrete system, the minimization may be obtained
using partial derivatives:  
∂ ∂R
R d  = 2R
d  =0 (20.6)
∂ui   ∂ui

1 Einstein’s summation convention is used throughout this discussion.


It is actually instructive to recast the results of Eq. 20.6 into the slightly
different form
2 R d  =0 (20.7)

From Eq. 20.7, it is now more obvious that the least squares method is a type
of weighted residual method with a canonical form expressed as

wi R d  = 0 (20.8)

Thus, in the case of the least squares method, Eq. 20.7 implies that
wi = 2 (∂R/∂ui ). In general, the weighting functions, wi , are somewhat arbitrary,
although they do need to satisfy certain requirements, including:

• They are linearly independent functions.

• They vanish on the part of the boundary, ∂E , where essential (Dirichlet)
boundary conditions are specified.

These requirements may be stated more formally as

wi ∈ Hom () (20.9)

where Hom is the Sobolev space of order m, where 2m refers to the order of the
highest derivative present in the linear differential operator L.

Example For illustrative purposes, these somewhat formal concepts may be

applied to a simple problem. Figure 20.1 depicts a bar of length L that is loaded by
end forces and a distributed body force that takes the form of a traction over the
entire length of the member. Considering now the equilibrium of an isolated rod
segment of differential length, forces may be summed as (consistent with Fig. 20.2)
 Fx = 0 = − σ dA + (σ + d σ ) dA + β dx (20.10)


dx x, u

FIGURE 20.1 Rod of length L loaded by end forces P and a distributed body force β.


s + ds
x, u


FIGURE 20.2 Free-body diagram of differential length of rod.

Equation 20.10 simplifies to

A +β =0 (20.11)

Subsequent consideration of linear kinematics and constitutive relations,

ε = du/dx and σ = E ε, respectively, yields the governing differential equation
(strong-form statement) of the problem,
d du
EA +β =0 (20.12)
dx dx

and thus provides

d du
L (u (x )) = EA +β =0 (20.13)
dx dx

over the domain  = {x |x ∈ (0, L)}

This continuous problem may be discretized by replacing the unknown solution,
u (x ), with an approximation in the form of that given by Eq. 20.2. The quantities
Ni are frequently referred to as shape functions in the literature on the finite element
method or as trial functions within the context of the classical applications of the
Rayleigh–Ritz method . Given that these individual functions, Ni , satisfy compact
support (Reddy, 1993) with respect to individual finite elements, their superposition
leads to the creation of a Schauder basis (Kreyszig, 1978) that spans the Sobolev
space of dimension i . As i → ∞, the approximation space approaches the space
of continuous functions where u (x ) occurs.
Using û (x ) in L yields the residual per-unit length of prismatic bar of homo-
geneous material,

d 2 (û (x ))
L (û (x )) = EA +β =R (20.14)
dx 2

Upon substituting Eq. 20.2 into Eq. 20.14, the following expression for the residual
per unit length is obtained:

R = (EA) ui Ni

(x ) + β (20.15)

where the primes signify spatial differentiation.

The least squares method may now be used to identify a suitable weighting
function for use in a weighted residual statement (i.e., weak form) of this example
∂R ∂
wi = 2 = (2EA) ui Ni

(x ) = (2EA) Nj

(x ) (20.16)
∂uj ∂uj

Note that the multiplier of 2EA may be dropped, as it is merely a constant and
thus does not impact the linear independence or satisfaction of essential (Dirichlet)
boundary conditions.
Continuing with the discussion of the least squares method as a weighted integral
form, the goal is to minimize the residual, R, as
wi R d  = (EA) ui Ni

(x ) + β Nj

(x ) dx = 0 (20.17)

The contents of Eq. 20.17 may be placed in slightly more familiar form as
  L   L
− EA Ni

(x ) Nj

(x ) d x ui = βNj

(x ) dx (20.18)
0 0

thus leading to a matrix representation of

[A] {u} = {b} (20.19)

where [A] is a stiffness matrix having entries

Aij = −EA Ni

(x ) Nj

(x ) d x (20.20)

and the force vector {b} is of the form


(x ) dx (20.21)

Galerkin Method A similar approach to the foregoing replaces the least squares
method with an approach known as the Galerkin method . This latter approach is
more common in the finite element analysis of solids and structures as a result of

the usual positive definiteness and strong ellipticity (Naylor and Sell, 1982) of the
governing functionals in structural mechanics. One important benefit in using the
Galerkin method is that the required order for continuity in the shape functions,
Ni , is reduced, as will be seen below.
In the Galerkin method, the weighting functions, wi , are selected to be identical
to the shape functions, Ni , underpinning the approximate solution ûi , as outlined
in Eq. 20.2. It should be noted that the foregoing approach is sometimes referred
to as the Bubanov–Galerkin method , in contrast to the Petrov–Galerkin method ,
in which the weighting functions and shape functions are not of the same form.
This distinction is pointed out to avoid confusion when working with the finite
element literature. For the purposes of the current discussion, reference to the
Bubanov–Galerkin method will be made simply as the Galerkin method .
Reconsidering the above example problem in terms of the Galerkin method, the
following weighted integral statement (i.e., weak-form statement) is obtained:
d 2u
wi R d  = Nj EA 2 + β dx = 0 (20.22)
 0 dx

In addition, it is instructive to include a set of practically useful boundary con-


• End fixity, u (0) = 0

• Loading applied to the free end, EA [du (L)/dx ] = P

In this example, the linear differential operator, L, is of order 2, and thus m =

1 (see Eq. 20.9), and the weighting functions wi in general, and Ni in particular,
must be admissible, stated formally as

Ni ∈ H0L (0, 1) (20.23)

Standard integration-by-parts approaches from elementary differential calcu-

lus may be applied to Eq. 20.22 using the following integration and substitution
dNj d 2u du
u d v = uv − v du u = Nj du = d v = EA 2 v = EA
dx dx dx

leading to
    L  L
du L dNj du
uv − v du = Nj EA − EA dx + Nj β dx = 0 (20.25)
dx 0 0 dx dx 0


that Nj (0) = 0 (because Nj ∈ H0m ), Nj (L) = 1, and
EA du(L)/dx = P, it is seen that the natural (Neumann) boundary conditions

are “automatically” recovered and Eq. 20.25 simplifies to

 L  L
dNj du
P − EA dx + Nj β dx = 0 (20.26)
0 dx dx 0

The approximation described by Eq. 20.2 may now be employed within Eq.
20.26 by replacing u by û, and rearranging terms yields
 L  L
dNj dNi
EA ui dx = P + Nj β dx (20.27)
0 dx dx 0

The left-hand terms in Eq. 20.27 lead to the stiffness matrix, expressing the
parameters associated with the response of the structural system under external
actions. The terms on the right lead directly to the generalized force vector, express-
ing the external actions on a structural system.
The benefit of using the Galerkin method is now apparent when comparing Eqs.
20.18 and 20.27. In Eq 20.18, the approach based on the least squares method
yielded an integral statement wherein the shape functions of the approximate
solution needed to be twice differentiable, and able to satisfy both the essential
(Dirichlet) and natural (Neumann) boundary conditions. In contrast, the outcome
of the Galerkin method is that the application of integration by parts results in a
shifting of the derivative from one term to another in Eq. 20.22, thus weakening the
continuity requirements on the shape functions used in forming the approximate
solution. Furthermore, it is observed (see Eq. 20.26) that the natural (Neumann)
boundary conditions are implied within the integral statement (Eq. 20.22) of the
example problem. As a result of their implied sense, the natural boundary condi-
tions did not require overt consideration during the selection of the shape functions
used in the approximate solution, û.
The foregoing consideration of the weighted integral approach to developing
a finite element representation of a structural problem may be brought into con-
sonance with other common formulation approaches (Bathe, 1996; Cook et al.,
2002; Zienkiewicz and Taylor, 2000a) through further consideration of the example
depicted in Fig. 20.1. When considering Eq. 20.27, the form of the so-called
strain–displacement matrix, commonly given as [B] in the finite element litera-
ture (Bathe, 1996; Belytschko et al., 2000; Cook et al., 2002; Crisfield, 1991a,b;
Hughes, 2000; Zienkiewicz and Taylor, 2000a), may be developed as
du d ûi dNi
ε= ≈ = ui → [B] {u} (20.28)
dx dx dx

Substituting the foregoing result into Eq. 20.27 yields (in matrix form)
  L   L
EA [B] [B] dx {u} = P +
{N }T β dx (20.29)
0 0

Equation 20.29 lends itself to being cast in the familiar general form of the
equilibrium statement

[K ] {u} = {F } (20.30)

where [K ] is the stiffness matrix formed through integration of the product of the
strain–displacements terms and {F } the force vector representation (in this case,
the aggregate effect of concentrated loads and distributed body forces).


Nonlinearities in structural analysis may arise from three primary sources: boundary
conditions that change during the problem solution (e.g., hard contact between
bodies); nonlinear material response (e.g., yielding of steel in a structural member);
and geometric nonlinearity (e.g., elastic column buckling). It is the last of these,
geometric nonlinearity, which is most closely aligned with the notion of quality
of equilibrium and thus is the focus of the present discussion. This is not to say
that the other two sources do not pertain to stability problems; each of these can
lead to manifestations of instability in their own right. The effects of nonlinear
material and boundary conditions, however, will not be treated herein due to space
limitations and instead reference is given to Simo and Hughes (1991), Wriggers
(2002), Bathe (1996), and Belytschko et al. (2000).
A convenient starting point to a discussion on ways of admitting geometric
nonlinearity into the finite element formulation is through a consideration of large
displacements as well as large strains. Since instability problems invariably require
the formulation of equilibrium on some deformed configuration of the structural
system being investigated, it appears as a truism that, at the very least, large dis-
placements must be admitted. In the case of the present discussion, large strains
are also included for completeness. With this in mind, a new strain measure that
is suitable for large displacement and large strain response is now introduced.
The Green–Lagrange strain is commonly encountered in treatments of the
nonlinear finite element method (Bathe, 1996; Belytschko et al., 2000; Crisfield,
1991; Zienkiewicz and Taylor, 2000b), but important foundation theory can be
found within relevant texts on continuum mechanics (Malvern, 1969; Fung, 1994;
Ogden, 1984; Novozhilov, 1953).
In general, the Green–Lagrange strain is a second-order tensor describing the
deformation of a body and whose tensorial components are
1 dui duj duk duk
Eij = + + (20.31)
2 dxj dxi dxi dxj

One very important property of the Green–Lagrange strain, as compared with

the usual linear engineering strain, is its invariance under rigid-body rotations, a
critically important feature with respect to stability-related finite element analyses
that involve large displacements (rotations).


FIGURE 20.3 Large rotation of rigid rod.

Consider a simple unit-dimensional example problem depicted in Fig. 20.3.

This example will highlight two important ideas that are at the heart of the present

• The unsuitability of the engineering strain in cases of large rotation

• The invariance of the Green–Lagrange strain in such cases

Clearly, a strain measure that yields nonzero terms when a body merely rotates as
a rigid body (i.e., no deformation) would prove disastrous for applications involving
geometric nonlinearity. Such a rigid-body rotation condition is depicted in Fig. 20.3,
wherein a bar that is initially aligned with the x axis of a two-dimensional Cartesian
reference frame rotates 45◦ about its left end. The displacement components of the
right end may then be given as
√ √
2 2
u1 = − L − L and u2 = L
2 2

Considering these displacements, the response of the element according to engi-

neering strain is given by
√ √
du1 ( 2/2)L − L 2
ε= = = − 1 = 0 (20.32)
dx L 2

Clearly the strains obtained according to linear engineering strain theory (Eq.
20.32) are incorrect because there can be no deformation of the structural element
under a pure rigid-body rotation.
Consider now the Green–Lagrange strain of this problem. The components of
the tensor that measure deformation along the x axis yield
1 ∂u1 ∂u1 ∂u1 ∂u1 ∂u2 ∂u2
E11 = + + +
2 ∂x1 ∂x1 ∂x1 ∂x1 ∂x1 ∂x1
 2  2
∂u1 1 ∂u1 1 ∂u2
= + +
∂x1 2 ∂x1 2 ∂x1

√ √ 2 √ 2
2 1 2 2
= −1 + −1 +
2 2 2 2
2 1 1 √ 1
= −1 + − 2+1 + =0 (20.33)
2 2 2 4

Thus, the Green–Lagrange strain remains invariant (i.e., null) during the
rigid-body rotation depicted in Fig. 20.3.
In light of the foregoing discussion on the Green–Lagrange strain, an extension
of the equilibrium relation of Eq. 20.29 to include large deformations may now be
defined. Such an extension will permit the consideration of geometric nonlinearity,
a condition that is at the heart of essentially all structural stability problems.
Toward this end, consider an extension of the notion of a strain–displacement
matrix [B] to account for geometrically nonlinear effects. Recognizing that the
first two terms on the right-hand side of Eq. 20.31 correspond to the usual, linear
engineering strain, a matrix designation of [B L ] can be applied to this portion of the
strain tensor. The remaining terms, which are responsible for the consideration of
large deformations, can be included in a second strain–displacement relation, [BNL ].
Using this approach, the Green–Lagrange strain tensor now may be represented
(in matrix form) as

[E ] = [BL ] {u} + [BNL ] {u} = [BGNL ] {u} (20.34)

This result may now be used to extend the equilibrium relation (Eq. 20.29),
for the example problem of Fig 20.2, for applications that include geometrically
nonlinear behavior. This extension takes the straightforward form
  L   L
EA [BGNL ] [BGNL ] dx {u} = P +
{N }T β dx (20.35)
0 0

Evaluation of the left-hand side of Eq. 20.35 results in the tangent stiffness
matrix , a linear expansion of the equilibrium path about a particular point in the
configuration space of the particular problem. The right-hand side, in this case,
embodies the loading effects related to the specific example problem. In the incre-
mental nonlinear solution of Eq. 20.35, within a Lagrangian reference frame, the
treatment of the solution of an unknown equilibrium configuration, using a previ-
ously solved-for configuration, leads naturally to the identification of the system
internal force vector that arises in order to balance applied loadings from the
right-hand side. A more detailed treatment of these concepts is provided in Chapter
6 of Bathe (1996).
To facilitate the discussion to follow, the tangent stiffness matrix is repre-
sented as

[KT ] = [Ko ] + [Kσ ] (20.36)


in which the linear elastic stiffness matrix is

[Ko ] = EA [BL ]T [BL ] dx (20.37)

and the initial stress matrix is

[Kσ ] = EA T
[BNL ] [BNL ] dx (20.38)

In all subsequent discussions, a total Lagrangian reference frame is assumed,

and thus all formulations take the initially undeformed configuration of the
structure as the reference frame from which to measure kinematic relations (Bathe,
1996). It is pointed out that the initial state from the total Lagrangian perspective
is used to formulate nonlinear finite element equations of motion. This initial
state has no bearing on the baseline and characteristic equilibrium states that are
subsequently introduced in the discussions pertaining to linearized eigenvalue
buckling approaches.


There is an increasing availability of commercial finite element software that

permits the consideration of the effects of geometric nonlinearity in structural
analysis. Oftentimes these software systems will have the capability to treat
stability problems through eigenvalue extraction routines applied to the global
system stiffness matrix, an approach often referred to as buckling, eigenvalue
buckling, or linearized eigenvalue buckling analysis.
This type of a stability analysis is attractive from the standpoint that it is not
computationally expensive. As compared with a more general incremental analy-
sis that traces the entire nonlinear equilibrium path of the structural system, the
eigenvalue buckling approach concerns itself with only one or two points on the
equilibrium path. In addition, results obtained from eigenvalue buckling analyses,
when applied to stability problems exhibiting bifurcation instability, are usually
quite accurate (Earls, 2007), and this accuracy is obtained without much interven-
tion on the part of the software user. Care must be taken, however, when applying
this technique to problem types exhibiting other manifestations of instability (e.g.,
limit point instability).
In practice, situations may arise in the design office where the application of
eigenvalue buckling may seem attractive for problems involving elastic beam
buckling (e.g., lateral–torsional buckling of a beam or planar truss) or elastic
snap-through buckling (e.g., lattice dome, arch, or shallow truss assembly). In
the domain of stability research, linearized eigenvlaue buckling is also oftentimes
attractive as a means for identifying an initial or seed imperfection for application
within a more detailed incremental nonlinear finite element analysis of a beam or

framework, for instance. In all of the foregoing, there are finite structural defor-
mations prior to the onset of instability that are additive to the governing buckling
mode (as compared to a bifurcation instability where the pre- and postbuckling
deformations may be thought of as being orthogonal to one another). This fact
creates an inconsistency with regard to assumptions made in the formulation of the
linearized eigenvalue buckling procedure itself.
The present discussion examines the underlying assumptions within the for-
mulation of the eigenvalue buckling method in order to highlight the problem
types that most readily lend themselves to solution by this method. In addition,
problems presenting responses that violate these fundamental assumptions are also
examined. In this latter case, it becomes very important to understand the nature of
the implementation of eigenvalue buckling in the given software system (example
problems, solved by MASTAN2, ADINA, ANSYS, and ABAQUS, are included
in this discussion); certain implementations will make application of eigenvalue
buckling to other than bifurcation problems extremely problematic. The present
discussion related to the various finite element buckling formulations employs a
single, standardized notation to allow for a transparent comparison of underlying

20.3.1 Overview of Finite Element Buckling Analysis Approaches

The literature adopts the term buckling analysis when referring to a family of finite
element techniques applied to structural systems for the identification of critical
load levels through the solution of an eigenproblem arising out of assumptions
made relative to changes in structural stiffness and associated applied loadings.
While the technique is applicable to structures that exhibit critical responses aris-
ing from the limit point as a well as bifurcation of equilibrium, the term buckling is
nonetheless universally applied. While this may seem inconsistent, since buckling
is normally associated with the condition of bifurcation in the equilibrium path
only, the nomenclature is defensible nonetheless as a result of the fact that the
eigenproblem posed within the finite element context resembles the case where the
vanishing of the determinant of the stiffness matrix is associated with a form of the
Sturm–Liouville problem (Reddy, 1993; Boyce and DiPrima, 1986). All formula-
tions within the present discussion will be presented in a standardized notation (to
facilitate comparison) that is defined subsequently.
Because the analyses considered here are strictly static, time will be used to
signify an equilibrium point within the configuration space of a given structure
corresponding to certain load level. Definitions for several left subscripts are now
introduced to facilitate subsequent discussions. A generic place holder term, ,
will be used to show the relative locations of the subscripts:
o  ≡ generic quantity, , evaluated at the equilibrium configuration
associated with the trivial case of no external actions.
t  ≡ generic quantity, , evaluated at an intermediate equilibrium
configuration occurring between the unloaded and critical configurations.

t  ≡ denotes the change in a generic quantity, , occurring as a result of

movement from location t to t + t on the equilibrium path.
t+t  ≡ generic quantity, , evaluated at an intermediate equilibrium
configuration occurring between the unloaded and critical configurations,
that is, arbitrarily close to t .
In subsequent discussions, it will also be helpful to define three applied loading
conditions that are used to surmise an assumed characteristic change in the system
stiffness. In general, the applied loading will be denoted with P.
{Pbaseline } ≡ loading condition used to bring the structure to a point in
configuration space associated with the left subscript t.
{Pcharacteristic } ≡ loading condition used to bring the structure to a point in
configuration space associated with the left subscript t + t.
{Pcr } ≡ the critical load associated with the equilibrium configuration at
incipient instability.
Structural stiffness will be denoted by the usual quantity [K ], more specifically
defined by:
[o Ko ] ≡ linear elastic stiffness matrix whose elements are independent of the
current structural configuration.
[τ Kσ ] ≡ initial stress matrix dependent on the state of stress at an arbitrary time,
t. This matrix is populated with terms that include both linear and
quadratic dependencies on the current displacement field.

The sum of the foregoing two stiffness matrices is typically what is referred to
as the tangent stiffness matrix , and associated with a specific equilibrium point in
configuration space. Some readers may be more familiar with the notion of the tan-
gent stiffness being associated with the linear terms of a Taylor series expansion of
the internal force vector about the current configuration during the solution; while
others may recognize it as emanating from the stationarity of the total potential
functional with an internal energy term that includes the influence of finite strains.
While other options exist for the population of the tangent stiffness matrix (Wood
and Schrefler, 1978; Holzer et al., 1990; Chang and Chen, 1986), the former def-
inition has emerged as the most popular to date. In particular, the instantaneous
stiffness of the structure arrived at by retaining only the linear terms in a Tay-
lor series expansion of the load–deflection response of the structure about the
point in configuration space may be defined according to the applied loading as
[Kbaseline ] ≡ instantaneous stiffness corresponding to the applied loading
{Pbaseline }
[Kcharacteristic ] ≡ instantaneous stiffness corresponding to the applied loading
{Pcharacteristic }

Classical Formulation The initial treatment of the finite element buckling anal-
ysis appeared in the literature prior to the formal naming of the finite element

method (Gallagher et al., 1967); this earliest reference identified the approach as
being based on the discrete element procedure. In light of the foregoing, and based
on a survey of the literature, it appears that in the most commonly held defini-
tion of the classical formulation for finite element buckling analysis, the following
problem is solved (Cook et al., 2002; Holzer et al., 1990; Chang and Chen, 1986;
Brendel and Ramm, 1980):

det ([o Ko ] + [t Kσ ]) = 0 (20.39)

It is frequently assumed in the literature that the equilibrium point at time t is

very close to the initial configuration at time 0, but this is not a strict requirement.
The subsequent buckling load is computed as

{Pcr } = λ {Pbaseline } (20.40)

Secant Formulation The present discussion adopts the name secant formula-
tion to describe the variation of the finite element buckling problem that is referred
to variously as the secant formulation (Bathe and Dvorkin, 1983; ADINA, 2007)
and the linear and nonlinear analysis (Holzer et al., 1990). This problem is posed as

det ([Kbaseline ] + λ ([Kcharacteristic ] − [Kbaseline ])) = 0 (20.41)

and the subsequent buckling load is computed as

{Pcr } = {Pbaseline } + λ ({Pcharacteristic } − {Pbaseline }) (20.42)

20.3.2 Example Problems

In the following examples, comparisons of results from the two approaches to
finite element buckling analyses are considered (i.e., those involving software pack-
ages employing buckling approaches characterized by Eqs. 20.39 and 20.40 and
20.41and 20.42, respectively). In some instances the results from closed-form solu-
tions are also presented. In addition, several cases also include results from manual
implementation of the finite element buckling approaches as encapsulated in Eqs.
20.39 and 20.40.
This discussion begins by distinguishing between bifurcation and limit point
instability in a formal way. Consider the classical form of the finite element state-
ment of the incremental equilibrium equations:

([o Ko ] + [t Kσ ]) {t+t u} = {t+t R} (20.43)

where {u} are the incremental nodal displacements and {R} is the residual vector
representing the imbalance between the internal forces at time t and the desired
load levels associated with some set of external forces, {P}. It is pointed out that
this residual differs in context from the quantity sharing the same name and defined

in Eq. 20.3. The standard form of the eigenvalue problem will be used to compute
eigenvalues, ωi , and eigenvectors, {φ}i , for the tangent stiffness matrix according to

([o Ko ] + [t Kσ ]) {φ}i = ωi {φ}i (20.44)

These eigenpairs lead to the spectral representation of the tangent stiffness


[KT ] = ωi {φ}i {φ}Ti (20.45)
These same eigenvectors may also be used to define a projection operator that
takes the original displacement and load vectors and projects them onto a new
vector space spanned by the eigenvectors, {φ}i , such that

{t+t u} = αi {φ}i (20.46a)

{t+t P} = ρi {φ}i (20.46b)

where {P} is the externally applied load vector on the structure and ρi = {φ}Ti {P}.
The transformations embodied in Eqs. 20.45 and 20.46 effectively diagonalize Eq.
20.43 and result in the following transformation (Pecknold et al. 1985):

ωi αi = λρi (20.47)

In an elastic structure, the first critical point occurs when the equilibrium
equations become singular or, in other words, when the tangent stiffness matrix is
no longer positive definite. Well-ordered eigenpairs result from this being one case
of the Sturm–Liouville problem, and thus it can be recognized that loss of positive
definiteness of the stiffness occurs when ω1 = 0, at which point the condition
λρ1 = 0 occurs. Based on this fact, it can be recognized that a distinction between
bifurcation and limit point instability exists (Pecknold et al., 1985). If ρ1 = 0,
then it is clear that the eigenvector is not orthogonal to the externally applied
loading vector {P}, and thus λ will have to be zero in order for λρ1 = 0 to be
true. In this case, a limit point instability condition exists (this point will become
more clear subsequently). Conversely, ρ1 = 0 when the loading is orthogonal to
the eigenmode and buckling is occurring. This may be summarized as:

• Limit point instability: ω1 = 0 and {φ}T1 {P} = 0

1 {P} = 0
• Bifurcation instability: ω1 = 0 and {φ}T

These results lead to the following interpretation. In the neighborhood of limit

points on the equilibrium path in configuration space, there is no increasing load,
and the eigenvector is not orthogonal to the external load vector. Conversely, if an
increase in loading is possible in the neighborhood of the critical point, and the
eigenvector is orthogonal to the load vector, then bifurcation instability is present.

Bifurcation Instability Two classical examples of bifurcation instability are

considered next. In these problems, prebuckling deformations are small, and thus
the assumptions in this regard, which underlie the linearized eigenvalue buckling
approach, are preserved. The first problem considers a one-dimensional structural
idealization of a three-dimensional problem, while the second example treats a
two-dimensional idealization of a three-dimensional case.
A classical Euler column example is depicted in Fig. 20.4. This figure displays
the problem geometry, along with boundary condition, cross section, and material
response information. The well-known solution of the example problem is given by

π 2 EI
Pcr = (20.48)
which for the given problem results in a theoretically “exact” answer of Pcr = 238
kips. Clearly this agrees well with both linear eigenvalue buckling solutions (also
shown in Fig. 20.4), irrespective of baseline loading, Pbaseline .
The second bifurcation instability example is given in Fig. 20.5. This
plate-buckling problem also has a well-known solution, which is given by

π 2E n
σcr =k   (20.49)
12 1 − ν 2 (b/t)2

Substitution of the specific properties and boundary conditions described in

Fig. 20.5 yields a critical elastic stress σcr = 1.114 ksi. As might be expected,
finite element results for both buckling approaches under discussion are in close
agreement with this result.

Boundary Conditions:
Z, u3 b = 1" Free? u 1 u 2 u 3 ur 1 ur 2 ur 3
L = 10"
1" I = 0.0833 roller N Y N Y N N
Y, u2 E = 29,000 ksi
pin N N N Y N N Pcr


Linearized Eigenvalue Buckling Results:

(5 ADINA Hermitian beam elements)

Pbaseline Pcr Method

200 k 235 Classical
100 k 235.8 Classical
10 k 236.5 Classical
1k 236.6 Classical
200 k 235 Secant
100 k 235.8 Secant
10 k 236.5 Secant
1k 236.6 Secant

FIGURE 20.4 Euler column example.




a = 50" Linearized Eigenvalue Buckling Results:

b (10 ADINA MITC4 four node shell elements
b = 10"
t = 0.1" across width)
E = 29,000 ksi

Pbaseline scr Method

n = 0.3

a 0.001 k/in 1.222 ksi Classical

since a > b 0.01 k/in 1.217 ksi Classical
Y, u2 and b >> t 0.04 k/in 1.217 ksi Classical
then 0.05 k/in 1.217 ksi Classical
k = 0.425 0.12 k/in NA Classical
X, u 1 0.001 k/in NA Secant
pinY h=1 0.01 k/in 1.217 ksi Secant
0.04 k/in 1.217 ksi Secant
Boundary Conditions: 0.05 k/in 1.217 ksi Secant
Free? u 1 u2 u 3 ur1 ur 2 ur 3 0.12 k/in NA Secant

rollerY Y Y N Y N N
pinY Y N N Y N N
simplex N Y N N Y N
free Y Y Y Y Y N

FIGURE 20.5 Plate-buckling example.

Based on the two simple bifurcation buckling examples presented, it may seem
as though linearized eigenvalue buckling approaches satisfactorily predict critical
loads in this type of problem. This, however, is not always the case (Earls, 2007)
and caution should be taken against merely accepting the buckling loads from
finite element software, even under the favorable conditions of bifurcation buckling.
For example, it is noted here that in the case of the ADINA two-point buckling
formulation (known as the secant formulation in the ADINA literature), it is not
possible to know what the load level of Pcharacteristic is because the information
contained in the output files are incomplete in this regard. This is considered to be
a critical shortcoming in ADINA in terms of reliance on the secant formulation for
finite element buckling analysis. As will be seen subsequently, it is not possible to
take full advantage of the secant method when using ADINA, because sufficient
information regarding the underlying solution process is not available to the user,
and thus critical judgment related to the specification of Pbaseline , or even simply
the interpretation of results, is compromised.

Limit Point Instability The remaining cases for discussion all exhibit limit point
(snap-through) instability, and thus prebuckling deformations tend to be finite.
Strictly speaking, this may be viewed as a violation of the underlying assumptions
used in the formulations of Eqs. 20.39 and 20.42. Thus, it might be reasonably con-
cluded that finite element buckling analyses are, technically, not applicable to such
instances. Engineers, however, do employ this type of approach to cases that are

not strictly in consonance with the underlying assumptions of the formulation, and
thus it is important to consider this class of problems within the present discussion.
In addition, it is not possible to escape the fact that, under certain circumstances,
accurate solutions are obtained when comparing finite element buckling results with
the results of more exact methods of analysis.
In the subsequent discussion, it will be useful to refer to a class of diagrams
known as eigenvalue plots (Brendel and Ramm, 1980; Holzer et al., 1990). As
shown in Fig. 20.6, such a plot depicts a graph of Pcritical versus Pbaseline (both
normalized by dividing by the “exact” critical load). The depiction of the eigenvlaue
plot in Fig. 20.6 is useful to consider in the case of limit point instabilities, because
it highlights the dependence of the finite element buckling solution on the selection
of a reasonable baseline loading, Pbaseline . The predicted critical loading from the
finite element buckling solution is arrived at by multiplying the eigenvalues by the
baseline loads, as described in Eqs. 20.40 and 20.42, respectively, for the classical
and secant approaches. The results presented in Fig. 20.6 are consistent with what
is expected, from the standpoint that the approximate finite element buckling loads
improve in accuracy as the magnitude of the baseline load increases.
Next, consider the truss arch shown in Fig. 20.7, with several height-to-span
ratios studied. In all cases, the span length is held constant at 20 in., and the height
is varied from 2 to 17 in. The truss arch is a particularly useful example, because
it affords the opportunity to easily obtain results using hand calculations. In the
case of the truss arch shown Fig. 20.7, with a height of 2 in. (i.e., shallow case), an
energy formulation involving the stationarity of the total potential yields a critical
load of 85.4 kips. For this particular truss arch geometry, finite element buckling
results were obtained using MASTAN2, ANSYS, and ADINA. In the case of the
first two software packages, only a very small axial force is considered in the

Pcritical versus Pbaseline (Normalized by exactcritical load)




lsecant Pbaseline


0.4 Base Load

0.2 Classical
Pbaseline Secant
0 0.2 0.4 0.6 0.8 1

FIGURE 20.6 Representative eigenvalue plot.


b = 1" Boundary Conditions:

Span = 20" Pcr Free? u1 u2 u3
Height = 2" to 17"
E = 29,000 ksi pin N N N
Y, u2 1" brace Y Y N

q° X, u1

FIGURE 20.7 Truss arch example.

formulation of the tangent stiffness matrix used in the classical approach from Eq.
20.39. These results, along with a hand calculation meant to parallel the classical
formulation, as presented in Eqs. 20.39 and 20.41, appear in Fig. 20.8. It is clear
that MASTAN2, ADINA, ANSYS, and the hand calculation all agree reasonably
well for small values of Pbaseline . As Pbaseline is increased, MASTAN2 and ANSYS
remain constant in their predictions, because their implementation of the classical
method does not admit the possibility of a varying Pbaseline . In addition, while the
hand calculations and ADINA both permit a variation in Pbaseline , their agreement at
high levels is less favorable than at low values of Pbaseline . This may be a result of
subtle differences in the way the classical formulation is implemented in ADINA;
but as an unfortunate byproduct of a lack of inclusion (within the output files of
ADINA) of intermediate values in the solution process, it is very difficult to test
any theories aimed at understanding the nature of the observed differences.

Pcritical versus Pbaseline (Normalized by exact critical load)


Base Load
3 ADINA - classical
ADINA - secant
Hand Calculation

0 0.2 0.4 0.6 0.8 1

FIGURE 20.8 Truss arch example results for shallow case.


Pcritical versus Pbaseline (Normalized by exact critical load)

Base Load
4 ADINA - classical
ADINA - secant
3.5 Hand Calculation




0 0.2 0.4 0.6 0.8 1

FIGURE 20.9 Truss arch example results for deep case.

In the case of the 17-in. high truss arch (i.e., deep case), we see a difference
in the trend of the response observed in the eigenvalue plots of Fig. 20.9. While
the MASTAN2 and hand calculations agree well with each other at low loads
and produce a reasonable estimate for the critical load, the same is not true for
the ADINA results. The ADINA secant results are clearly diverging from the
correct solution (13,000 kips; obtained from an incremental nonlinear finite element
analysis) while the ADINA classical result begins at a point inexplicably far away
from the other two classical implementations.

20.3.3 Linearized Eigenvalue Buckling for Imperfection Seeding in

Incremental Nonlinear Finite Element Analysis
Linearized eigenvalue buckling analysis techniques are sometimes employed as a
means for obtaining a seed (initial) imperfection to the structural geometry in order
to permit a realistic incremental nonlinear finite element analysis (e.g., to facilitate
equilibrium branch switching in the study of a bifurcation problem). In such an
approach, it is the components of the eigenvector (which represent the relative
displacement values at each degree of freedom associated with all nodes) that are
scaled to define an initial displacement field that is assigned (as a perturbation) to
the ideal node locations of a mesh employed in an incremental nonlinear analysis.
Such an approach is predicated on the notion of repeatability across commercial
codes (i.e., commercial codes should all yield similar mode shapes from models
possessing identical parameters). It has been clearly illustrated, however, that each
of the dominant formulations for eigenvalue analysis solves a slightly different
eigenproblem, as seen in Eqs. 40.39 through 40.42, and thus complete uniformity

in predictions across commercial codes is unlikley. In order to further analyze the

significance of this variation in formulation, eigenvector results are quantified to
facilitate comparison.
The eigenvectors computed with the secant and classical formulations for a
given model may be very similar or, unfortunately, drastically different. This dif-
ference or similarity in results can accompany minor changes in geometry and/or
loading conditions as well as changes in eigenproblem formulation. Finite element
postprocessing software most often scales any reported eigenvectors in order to
facilitate visualization of the mode shapes. This scaling can cause problems when
comparing results from different commercial programs. In order to more precisely
compare the eigenvectors from different formulations (applied to models with iden-
tical geometries and material parameters), a sampling of eigenvector values along
the centerline of two example beams (Fig. 20.10) is taken. A discrete Fourier anal-
ysis is then conducted with these data using the position along the beam as the
independent variable.
Nominal dimensions for the two cases depicted in Fig. 20.10 are presented in
Table 20.1. Twist is prevented at the ends of the beams by guiding the rectangu-
lar stiffener plates to resist out-of-plane deformation. Both beams are modeled in
ADINA and ABAQUS using shell elements. Although the shell element formula-
tions are not the same for ADINA and ABAQUS, any differences in formulation
are not significant to this comparison because the observed behavior is within the
elastic, small-strain-response region (all element meshes were of equivalent den-
sity using linear, four node shells; the meshes were also seen to be converged).
Two ADINA analyses were performed: one using the classical formulation, and the
other with the secant formulation. The ABAQUS analyses only employed a classi-
cal type of formulation because secant buckling analysis can only be approximated
through the use of a “restart” option.

FIGURE 20.10 Beam imperfection example cases.


TABLE 20.1 Nominal Dimensions of Example Beams

Case 1 Case 2
Measure in. mm in. mm

Length, L 250 6350 29 736

Depth, d 19.00 483 8.25 210
Flange width, bf 6 152.4 4 101.6
Web thickness, tw 1.00 25.4 0.20 5.1
Flange thickness, tf 1.00 25.4 0.25 6.4

The first example (case 1) is a steel I-beam that is loaded with a uniformly
distributed pressure on the top flange. Figure 20.11 shows that the resulting mode
shapes are very similar for both programs. The eigenvectors obtained from ADINA
and ABAQUS postprocessing have been normalized (scaled), to enable comparison,
using the relation

φi  = (20.50)

where φ represents the eigenvector of interest, with the subscripts corresponding

to the discrete nodal degree of freedom displacement components of the vector.
When using eigenvector data corresponding to the minor axis (x -direction) lat-
eral displacement component, from nodes located along the centerline of the web,
it becomes apparent (Fig. 20.12) that the first-mode eigenvectors for the different
formulations employed are only slightly different. This difference in mode shape
becomes more definable once a discrete Fourier transform of the minor axis direc-
tion eigenvector displacement data is carried out. Figure 20.12 displays the results
from this approach, highlighting that the first mode shape resulting from the ADINA
classical formulation, the ADINA secant formulation, and ABAQUS formulation
are related. They are obviously all harmonic, and their peaks occur at the same
frequency. The frequency of the Fourier transform corresponds to the number of
cycles of the harmonic function, which for this particular beam example describes
the deformation per length of the member. The peak amplitude of the ABAQUS
Fourier analysis is 8% less than that obtained using both ADINA formulations; the

FIGURE 20.11 Case 1 buckling mode predictions.


Mode 1 Web Centerline Deformation Deformation Spectra

1 ADINA Two-Point 12 ADINA Two-Point

Normalized X-Dir. Deformation

Normalized X-Dir. Deformation

ADINA Classical ADINA Classical
0.8 ABAQUS Classical 10 ABAQUS Classical

0.2 2

0 0
0 50 100 150 200 250 4 3.5 3 2.5 2 1.5 1 0.5 0
Position (in) Frequency Content (1/ in) × 10−3

(a) (b)

FIGURE 20.12 Case 1 results: (a) eigenvector plots along web centerline; (b) discrete
Fourier transformation of these eigenvectors.

difference in peak frequency is only 1% between the ADINA classical and secant
formulation results. In the second mode the difference in Fourier peaks between
ABAQUS and ADINA results is 30%.
The second beam example (case 2, Fig. 20.10, with nominal dimensions listed in
Table 20.1) incorporates an extra stiffener plate at the center of the steel I-beam and
includes a midspan concentrated load, which is distributed linearly along the upper
edge of the center stiffener plates. The first buckling modes for this model, obtained
using ADINA and ABAQUS, are shown in Fig. 20.13. The results vary to such an
extent that the discrepancies are clearly not an artifact of magnification or scaling.
Normalized eigenvector values along the centerline nodes are plotted in Fig. 20.14.
The discrete Fourier transform of the minor axis (x -direction) component of the
eigenvector data, which is also presented in Fig. 20.14, indicates not only that
there is a difference in the peak height of the results but also that the peaks are
also located at different points along the member longitudinal axis. The secant
formulation solution has a much more pronounced harmonic displacement field
when compared to the classical formulation. In this case, the ABAQUS results are
nearly equivalent to those of the ADINA classical formulation. ADINA classical
and secant formulations differ significantly in that the major peaks do not coincide
in height or location.

FIGURE 20.13 Case 2 buckling mode predictions.


Mode 1 Web Centerline Deformation Deformation Spectra

1 4
Normalized X-Dir. Deformation

ADINA Two-Point

Normalized X-Dir Deformation

3.5 ADINA Classical
ABAQUS Classical
0.5 3
0 2
–0.5 ADINA Two-Point 1
ADINA Classical 0.5
ABAQUS Classical
–1 0
0 5 10 15 20 25 30 0.06 0.05 0.04 0.03 0.02 0.01 0
Position (in) Frequency Content (1/in)
(a) (b)

FIGURE 20.14 Case 2 results: (a) eigenvector plots along web centerline; (b) discrete
Fourier transformation of these eigenvectors.

20.3.4 Observations on Linearized Eigenvalue Buckling

Finite element buckling analysis results should be interpreted with great care. The
results of ostensibly identical formulations (e.g., as described in theory manuals)
within various software packages frequently lead to estimates of critical loads that
vary significantly for identical structural configurations.
It seems reasonable to avoid using such finite element buckling approaches for
all but the simplest cases of bifurcation buckling, but even then care must be taken
when considering the validity of the results.
It also appears that the approach of using eignenmodes, obtained from linearized
eignevalue buckling analyses, as an imperfection seed within a more general incre-
mental nonlinear finite element analysis should be approached with caution; as
models with identical geometries and materials properties may result in very dif-
ferent imperfection fields, depending on the finite element software that is employed
and with what formulation options the analysis is carried out.


ADINA (2007), Theory Manual , ADINA R and D, Watertown, MA.

Babuska, I., and Melenk, J. M. (1997), “The Partition of Unity Method,” Int. J. Numer.
Methods Eng., Vol. 40, pp. 727–758.
Bathe, K. J. (1996), Finite Element Procedures, Prentice Hall, Englewood Cliffs, NJ.
Bathe, K. J., and Dvorkin, E. N. (1983), “On the Automatic Solution of Nonlinear Finite
Element Equations,” Comput. Struct., Vol. 17, Nos. 5–6, pp. 871–879.
Belytschko, T., and Black, T. (1997), “Elastic Crack Growth in Finite Elements with Minimal
Remeshing,” Int. J. Numer. Methods Eng., Vol. 45, pp. 601–620.
Belytschko, T., Liu, W. K., and Moran, B. (2000), Nonlinear Finite Elements for Continua
and Structures, Wiley, New York.
Boyce, W. E., and DiPrima, R. R. (1986), Elementary Differential Equations and Boundary
Value Problems, 4th ed., Wiley, New York.

Brendel, B., and Ramm, E. (1980), “Linear and Nonlinear Stability Analysis of Cylindrical
Shells,” Comput. Struct., Vol. 12, pp. 549–558
Chang, S. C., and Chen, J. J. (1986), “Effectiveness of Linear Bifurcation Analysis for
Predicting the Nonlinear Stability Limits of Structures,” Int. J. Numer. Methods Eng.,
Vol. 23, pp. 831–846.
Chapelle, D., and Bathe, K. J. (2003), The Finite Element Analysis of Shells—Fundamentals,
Springer, New York.
Ciarlet, P. G., and Lions, J. L. (1991), Handbook of Numerical Analysis, Vol. II: Finite
Element Methods (Part 1), North-Holland, Amsterdam.
Cook, R. D., Malkus, D. S., Plesha, M. E., and Witt, R. J. (2002), Concepts and Applications
of Finite Element Analysis, 4th ed., Wiley, Hoboken, NJ.
Crisfield, M. A. (1991a), Non-Linear Finite Element Analysis of Solids and Structures, Vol. 1,
Wiley, New York.
Crisfield, M.A. (1991b), Non-Linear Finite Element Analysis of Solids and Structures, Vol. 2,
Wiley, New York.
Duarte, C. A., and Oden, J. T. (1996), “An H-P Adaptive Method Using Clouds,” Comput.
Methods Appl. Mech. Eng., Vol. 139, pp. 237–262.
Earls, C. J. (2007), “Observations on Eigenvalue Buckling Analysis Within a Finite Ele-
ment Context,” Proceedings of the Structural Stability Research Council, Annual Stability
Conference, New Orleans, LA.
Farhat, C., Harari, I., and Hetmaniuk, U. (2003), “The Discontinuous Enrichment Method
for Multiscale Analysis,” Comput. Methods Appl. Mech. Eng., Vol. 192, pp. 3195–3209.
Fung, Y. C. (1994), A First Course in Continuum Mechanics, Prentice Hall, Englewood
Cliffs, NJ.
Gallagher, R. H., Gellatly, R. A., Padlog, J., and Mallett, R. H. (1967), “A Discrete Element
Procedure for Thin-Shell Instability Analysis,” AIAA J., Vol. 5, No. 1, pp. 138–145.
Holzer, S. M., Davalos, J. F., and Huang, C. Y. (1990), “A Review of Finite Element Stability
Investigations of Spatial Wood Structures,” Bull. Int. Assoc. Shell Spatial Struct., Vol. 31,
No. 2, pp. 161–171.
Hughes, T. J. R. (2000), The Finite Element Method: Linear Static and Dynamic Finite
Element Analysis, Dover, New York.
Kreyszig, E. (1978), Introductory Functional Analysis with Applications, Wiley, New York.
Langtangen, H. P. (2000), Computational Partial Differential Equations, Springer, New
Malvern, L. E. (1969), Introduction to the Mechanics of a Continuous Medium, Prentice-Hall,
Englewood Cliffs, NJ, 1969.
McGuire, W., Gallagher, R. H., and Ziemian, R. D. (2000), Matrix Structural Analysis, 2nd
ed., Wiley, New York.
Melenk, J. M., and Babuska, I. (1996) “The Partition of Unity Finite Element Method: Basic
Theory and Applications,” Comput. Methods Appl. Mech., Eng., Vol. 139, pp. 289–314.
Naylor, A. W., and Sell, G. R. (1982), Linear Operator Theory in Engineering and Science,
Springer. New York.
Novozhilov, V. V. (1953), Foundations of the Nonlinear Theory of Elasticity, Dover,
New York.
Ogden, R. W. (1984), Non-Linear Elastic Deformations, Dover, New York.

Pecknold, D. A., Ghaboussi, J., and Healey, T. J. (1985), “Snap-Through and Bifurcation in
a Simple Structure,”. J Eng. Mech., Vol. 111, No. 7, pp. 909–922.
Reddy, B. D. (1998), Introductory Functional Analysis, Springer, New York.
Reddy, J. N. (1993), An Introduction to the Finite Element Method, McGraw-Hill,
New York.
Simo, J. C., and Hughes, T. J. R. (1991), Computational Inelasticity, Springer, New York.
Wood, R. D., and Schrefler, B. (1978), “Geometrically Non-linear Analysis—A Correlation
of Finite Element Notations,” Int. J. Numer. Methods Eng., Vol. 12, pp. 635–642.
Wriggers, P. (2002), Computational Contact Mechanics, Wiley, Hoboken, NJ.
Zienkiewicz, O. C., and Taylor, R. L. (2000a) The Finite Element Method , Vol. 1: The
Basics, Butterworth-Heinemann, Oxford, UK
Zienkiewicz, O. C., and Taylor, R. L. (2000b), The Finite Element Method , Vol. 2: Solid
Mechanics, Butterworth-Heinemann, Oxford, UK.