Вы находитесь на странице: 1из 76

Non-linear Finite Element

Method for Solids


Lecture notes for the course 4K440

Varvara Kouznetsova
Chapter 1

Introduction

1.1 Non-linear finite elements and virtual


manufacturing
Non-linear finite element analysis is nowadays an essential component of com-
puter-aided design. Experimental testing of prototypes is increasingly being
replaced by non-linear finite element simulations (so-called virtual prototyp-
ing) because this provides a more rapid and less expensive way to evaluate
design concepts. The modern simulation technology enables companies to op-
timize key factors affecting profitability of their manufactured products. These
include manufacturability, reducing cost of production, material usage, final
shape, residual stress level and product durability. For example, in the field
of automotive design, simulation of crashes is replacing full-scale tests at both
the early and final design stages to meet crashworthiness criteria, Figure 1.1.
Selection of interior and exterior materials and geometries, placement of sen-
sors for airbag deployment etc., are all evaluated using numerical simulation
prior to building and testing expensive prototypes. In many industries, simu-
lation of manufacturing processes such as sheet metal forming, forging, casting
and extrusion allows to significantly speed up the design process, improve cost
efficiency and quality of the products. Examples of deep drawing and extru-
sion simulations are shown in Figure 1.2. Product durability is also being
actively evaluated using the numerical analysis. For example, in electronics
industry, drop-test and thermo-mechanical fatigue tests are being replaced by
the simulation.
In the core of computer-aided design and virtual manufacturing lies non-
linear finite element technology. In the fourties aeronautical engineers used
first arrangements of linear elastic bars and later arrangements of small dis-
crete pieces of a linear elastic continuum for airplane strength calculations.
The name finite elements, however, did not appear until 1960. By that time,
the linear finite element method was already used in many other industries
and fields of engineering science, such as fluid mechanics, heat transfer, elec-
tromagnetic wave propagation, etc. Applied mathematicians showed that the
method converges to the correct result and the relation to the mathemati-

1
CHAPTER 1. INTRODUCTION 2

Figure 1.1: Non-linear finite element simulations of car crashes;


courtesy of ARUP (using LSDYNA).

(a) (b)

Figure 1.2: Non-linear finite element simulations of manufactur-


ing processes: (a) deep drawing of an s-shaped rail,
courtesy of MSC.Software (using MS.Marc); (b) alu-
minium extrusion, from PhD thesis by Bas van Rens,
TUE, 1999.

cal publications (e.g. on weighted residuals, Galerkin’s method, variational


methods etc.) from the beginning of the century and earlier was established.
About the same time, mid-sixties, the non-linear finite element method was
developed, first for non-linear small strain problems; large strains and dynam-
ics followed soon. Since then, in the past 40 years, the finite element analysis
(FEA), and in particular the non-linear FEA, has been successfully applied in
all major industries, including aerospace, automotive, energy, manufacturing,
chemical, electronics, consumer and medical. Moreover, it is a major tool in
the engineering sciences research.
Even at the time when the non-linear finite element method still was in
its “child age”, the commercial software packages for non-linear FEA started
to appear. The program MARC (presently MSC.Marc) appeared in 1969 and
was one of the first commercial non-linear finite element programs. Soon after,
several other non-linear programs followed. At about the same time other
CHAPTER 1. INTRODUCTION 3

major players on the market were born, including ABAQUS, LSDYNA and
ANSYS. In addition to these general purpose programs, specialized software
packages have been developed, for example for modelling of forging, extrusion,
fatigue etc. Most of the modern FEA programs have an interface with CAD
systems, which facilitates the virtual manufacturing even further. Almost all
big and small engineering companies nowadays use one or more of the above
packages, in-house finite element solutions are also often developed.
A natural question, therefore, arises: Why does an engineer-analyst need
to study the non-linear finite element method, if so many ready-to-use soft-
ware packages are available? The answer is rather simple. A finite element
program is a black box that provides simulations. However, a non-linear finite
element analysis confronts an engineer with many choices and pitfalls. Se-
lections made in the model formulation, mesh discretizations, element types,
the solution procedure etc. may have a tremendous effect on the final results.
In numerical non-linear simulations, it is possible to obtain solutions that are
not physically stable and therefore quite meaningless. Solutions of non-linear
problems are often sensitive to imperfections of material and load parameters;
in some cases there is sensitivity to the mesh employed and the element used.
A knowledgeable user of non-linear finite element software must be aware of
these characteristics and pitfalls, must understand what is going on behind
the interface screen. Otherwise, the results obtained by elaborate computer
simulations may be quite misleading and result in incorrect design decisions.

1.2 Linearity versus non-linearity


Consider the simplest example of a linear system: a spring with stiffness K
loaded at the free end by a force f , see Figure 1.3(a). According to Hooke’s
law, a simple linear relationship between the force and deflection u exists
f = Ku (1.1)
This linear relation is schematically shown in Figure 1.3(b). Thus, for any
given force f , the deflection can be easily calculated by u = f /K.
If the applied force is two times larger, the spring will deflect twice as
much. Simple. Unfortunately, the nature is not that simple. If we try to do
this experiment with a spring in reality, after applying a certain force some
part of the deflection will not be recovered upon removal of the force, and if
we continue to increase the force, ultimately the spring will break. So, if we
plot the measured force-deflection curve of the spring it may look like the one
in Figure 1.4. The reality clearly deviates from the linearity!
Before we go any further in comparing linearity and non-linearity, let us
introduce the notion of equilibrium path, which is graphically represented by
a response diagram. Using the response diagram many key concepts can
be illustrated and interpreted in physical, mathematical and computational
terms. The gross or overall static behaviour of a deforming body or structure
can be characterized by a load-deflection or force-displacement response dia-
gram. Above we have already used such load-deflection diagrams shown in
CHAPTER 1. INTRODUCTION 4

f goes on forever

linear (Hooke’s) law

K
f u
u 0
(a) (b)

Figure 1.3: A simple spring (a) and its linear response according
to Hooke’s linear elastic law (b).

f linear (Hooke’s) law

failure

non-linear response

u
0
Figure 1.4: A possible non-linear force-deflection relation.

Figure 1.3(b) and Figure 1.4 to discuss the behaviour of a spring. In that case
the choice of the load and deflection was obvious. For complex cases, a choice
of a “representative” force and a “representative” displacement quantity may
not be so obvious and even not unique. In general, a response diagram depicts
the relationships between inputs and outputs, or in physical terms, between
what is applied and what is measured.
The behaviour of a body or structure may be defined non-linear, if its
response plot is non-linear. A linear structure is characterized by a linear
response diagram for all possible choices of load and deflection variables. The
consequences of such behaviour are not difficult to foresee:

• A linear structure can sustain any load whatsoever and undergo any
displacement magnitude.

• Removing all loads returns the structure to the origin of the response
plot (zero load, zero displacement).

• Response to different load systems can be obtained by superposition.

For these conditions to hold, the following requirements should be satisfied:


• perfect linear elasticity for any deformation;
CHAPTER 1. INTRODUCTION 5

• infinitesimally small deformations;

• infinite strength.

Clearly, these assumptions are not only physically unrealistic but mutually
contradictory. For example, if the deformation to remain infinitesimal for any
load, the body must be rigid, rather than elastic, which contradicts the first
assumption. Thus a linear structure never exists in reality. It is just a model
which has its limits of validity. Despite these limitations, the linear model can
be a good approximation of portions of a non-linear response.
Now that we know that the nature is non-linear, let us look at implications
of this for numerical computations. Clearly, if a spring exhibits the non-
linear response as shown in Figure 1.4, instead of the linear one shown in
Figure 1.3(b), the deflection u for any given force f can not be simply found
by using formula (1.1). Some numerical methods are needed to solve the
problem, e.g. iterative methods. The most widely used in the non-linear finite
element analysis is Newton’s method and its variations. Now assume that
instead of one spring we have a structure consisting of 1000 interconnected
springs. If each spring behaves according to the linear model, the behaviour of
the whole structure can be found by solving a system of 1000 linear equations.
Under normal circumstances a linear system has only one solution, which can
be readily calculated using modern computers. In contrast to that, if each
spring behaves non-linearly, say according to a cubic relation, a system of
1000 cubic equations has 31000 ≈ 10450 solutions in the complex plane. This
number is actually larger than the number of atoms in the Universe, which
is approximately 1050 ! Suppose, just several billions or millions of these are
real solutions. Which solution(s) have physical meaning? And how do you
compute those solutions without wasting time on the others?
To answer these questions we again use the notion of equilibrium path. For
static equilibrium problems, each point on the response diagram represents an
equilibrium state or equilibrium configuration of the structure. Therefore, if
we start from an easily computable solution, say the linear solution still in the
linear regime of the structure, we then try to follow the equilibrium path of the
system as actions (loads) applied are changed by small steps, called increments.
For each new increment the previous solution is used as a starting point for
the iterative solution search procedure. The smaller the steps we take, the
more chances we have that the first iteratively found solution also lies on the
equilibrium path, and thus it is the solution that we are interested in. Such a
solution procedure is called incremental-iterative and forms the basis for the
solution procedure in the non-linear finite element method.

1.3 Sources of non-linearities


In the previous section a non-linearity in the behaviour of a structure has
been identified on a response diagram, as it could have been observed by
conducting a physical experiment. The response diagram, however, represents
CHAPTER 1. INTRODUCTION 6

only the overall behaviour of the structure, but where does the non-linearity
come from? If we want to model such overall non-linear behaviour with a
mathematical or computational model, where exactly in the model should it
be introduced? To be able to answer these questions, further insight into the
sources of non-linearity is required.
Before identifying the sources of non-linearities, it is useful to recall the
mathematical formulation of a linear elastic problem. The static linear elastic
boundary value problem is governed by the following field equations:

• equilibrium equation
~ · σ + ~q = ~0
∇ (1.2)
~ the gradient operator, σ the stress tensor and ~q the body forces;
with ∇

• kinematic equation, strain-displacement relation


  
ε = 21 ∇~ ~u T
~ u + ∇~ (1.3)

with ε the (infinitesimal) strain tensor and ~u the displacement field;

• linear elastic constitutive law, the stress-strain relation

σ = 4C : ε (1.4)

with 4C the tensor of elastic material constants;

• natural boundary conditions, surface tractions are prescribed on the St


part of the boundary

~t = ~t∗ , on St (1.5)

• essential boundary conditions, displacements are prescribed on the Su


part of the boundary

~u = ~u∗ , on Su (1.6)

Note, that each of the above relations is linear. Any of these relations, however,
may be non-linear, thus giving rise to different types of non-linearities. Tracing
this fact back to physics, four types of non-linearities are commonly distin-
guished in solid mechanics: geometric, material, force boundary conditions and
displacement boundary conditions, as schematically shown in Figure 1.5. In
the following sections, each of these sources of non-linearity is discussed in
more detail.
CHAPTER 1. INTRODUCTION 7

equilibrium strain-displacement constitutive prescribed prescribed


equation relation law tractions displacements

geometric material force b.c. displacement b.c.


non-linearity non-linearity non-linearity non-linearity

! large strains ! non-linear elasticity ! pressure loads ! contact


! large deformations ! plasticity ! aerodynamic loads
and/or rotations ! viscoelasticity ! hydrodynamic loads
! viscoplasticity
! creep
! damage

Figure 1.5: Sources of nonlinearities

1.3.1 Geometric non-linearity


In some cases deformation (or displacements) of the body may be such that the
changes in the geometry may not be neglected. Examples include deformation
of rubber structures (tires, membranes), metal and plastic forming, cables,
springs, bars etc. If the change in geometry is important, then the linear
relation between the displacement and strain (1.3) does not hold any more and
the equilibrium equation (1.2) has to be carefully considered, since it includes
the divergence operator with respect to the changing geometry. Geometric
non-linearities may come from several physical sources:

• Large strain. The strains themselves may be large, say over 5%, such
as in rubber structures. These large strains are often associated with
material non-linearities.

• Small strains, but large displacements and/or rotations. This may take
place in slender structures, such as cables, springs, arches, bars etc. The
deformation strains in this case may still be treated as infinitesimal, but
finite displacements and rotations lead to geometrical non-linearities.

Combinations of large strains, large rotations and large displacements may


also be found in practice, for instance in forming operations.

1.3.2 Material non-linearity


The constitutive equations (1.4) that relate stresses and strains may become
non-linear. In addition, a material behaviour may depend not only on the cur-
rent deformation state, but possibly on the past history of the deformation.
It may also be influenced by other constitutive variables, such as initial stress,
temperature, time, electro-magnetic fields, moisture etc. Examples of non-
linear material behaviour include non-linear elasticity, typical for rubber and
elastomers, plasticity (metals), viscoelasticity and viscoplasticity (polymers
and plastics), creep (metals at high temperatures and concrete) and damage,
which occurs in all materials but has different mechanisms and thus requires
CHAPTER 1. INTRODUCTION 8

different models. Material non-linearities may give rise to very complex phe-
nomena such as path dependence, hysteresis, localization, fatigue etc. Addi-
tionally, in non-linear regime many materials behave in a nearly incompressible
manner. For example, elastically non-linear rubber materials are incompress-
ible. Another example includes the incompressible plastic behaviour of von
Mises elastic-plastic materials. Applying the standard finite element technol-
ogy to modelling incompressible and nearly incompressible material behaviour
typically result in numerical difficulties, e.g. such as volumetric locking, lead-
ing to erroneous results. Therefore the ability to treat incompressible material
behaviour correctly is important for non-linear finite element analysis.

1.3.3 Force boundary conditions non-linearity


The applied forces, the prescribed surface tractions ~t∗ and/or body forces ~q,
may depend of the deformation of the body. Mathematically, this may be
transformed into dependence of tractions and body forces on displacements
(which are usually unknown before solving the problem)

~t∗ = ~t∗ (~u), q=~


~ q (~u) (1.7)

The most important engineering applications of non-linearity in force bound-


ary conditions concern pressure loads. These include hydrostatic loads on sub-
merged bodies, aerodynamic and hydrodynamic loads (e.g wind loads, wave
loads, drag forces). More exotic examples are gyroscopic forces.

1.3.4 Displacement boundary conditions non-linearity


Displacement boundary conditions (1.6) may also depend on the deformation
of the body. The most important application is the contact problem, in which
no-interpenetration conditions are enforced on flexible bodies, while the con-
tact area is unknown. In this case, the determination of the essential boundary
conditions may be a part of the solution process. Other examples of displace-
ment boundary conditions non-linearity may be found in melting/solidifing
problems, phase changes and other cases when the boundary of a body is not
known beforehand.

1.4 Outline of this course


It is the objective of this course to provide understanding of some basic con-
cepts of the non-linear finite element analysis for solids. The course will be
mostly concerned with treatment of geometrical non-linearities and force dis-
placement boundary conditions non-linearities. Material non-linearities will be
only slightly addressed on an example of non-linear elasticity and by consider-
ing some of the techniques for numerical treatment of the incompressible and
slightly compressible materials. The exclusion of the material non-linearities
CHAPTER 1. INTRODUCTION 9

does not imply that they are less important than the others, quite the con-
trary. The material non-linearities, as well as their treatment in non-linear
finite element framework, are discussed in other courses1 .
First, in chapter 2 the linear finite element method is briefly reviewed on
the example of a linear elastic static problem. The key steps of the finite
element technique are underlined. These include the strong and weak form of
the governing equations, discretization, numerical integration, assembly, parti-
tioning and solving. In chapter 3 a geometrically non-linear elastic problem is
treated using the total Lagrange procedure. The total Lagrange formulation
uses the initial undeformed configuration as the reference configuration, to
which all the equations are transformed. The non-linear system of discretized
equations is first derived, followed by the summary of Newton’s method for
the solution of non-linear equations and systems. The linearized form of the
equations is derived for the modelling problem considered. It should also be
emphasized, that linearization is an essential step in solving many non-linear
problems, and thus its deep understanding is a key to success. Finally, some
aspects of implementation are discussed. The chapter concludes with an out-
line of the incremental procedure, which has been already briefly introduced
above in section 1.2. Chapter 4 considers the same geometrically non-linear
elastic problem, but uses the updated Lagrange solution procedure, in which
the last known configuration is used as the reference. Linearization of the
non-linear discretized equations is again discussed in detail. At the end of
the chapter the total and updated Lagrange frameworks are compared and
their equivalence is emphasized, while some recommendations are given with
respect to convenience of using one or another. In chapter 5 force bound-
ary conditions non-linearity is considered. Only surface tractions dependent
of the deformation (pressure loads) are discussed. Both, total and updated
Lagrange, formulations for deformation dependent forces are elaborated. Fi-
nally, in chapter 6 some computational problems caused by incompressible and
nearly incompressible material behaviour are addressed and several numerical
techniques for these types of material behaviour are discussed.

1
4K060 Damage mechanics
4H100 Constitutive modelling of solids
4K620 Computational material models
Chapter 2

Finite element method for


linear elastic problems

2.1 Static linear elasticity boundary value problem


Consider a linear elastic body that occupies the volume V and has the bound-
ary S, as schematically depicted in Figure 2.1. The body is in a state of static
equilibrium. This is mathematically reflected by the equilibrium equation
~ · σ(~x) + ~
∇ q (~x) = ~0 (2.1)

where ~x is the position vector of a point P within the body, ∇ ~ denotes the
gradient operator, σ the Cauchy stress tensor, and ~q(~x) are given body forces.
~t∗ (~x)

St
P
~q(~x) S
~x
0 V
Su
~u∗ (~x)

Figure 2.1: A linear elastic body with the applied boundary con-
ditions.

The body is subjected to boundary conditions. Two types of boundary


conditions are typically distinguished: essential (or Dirichlet) and natural (or
Neumann). For static equilibrium boundary value problems these boundary
conditions correspond to prescribed displacements and prescribed tractions,
respectively. Assume that the displacement field ~u is specified along a certain
part of the boundary Su ⊂ S

~u(~x) = ~u∗ (~x) on Su (2.2)

10
CHAPTER 2. LINEAR FEM 11

with ~u∗ (~x) a given function. On the remaining part of the boundary, St =
S\Su tractions ~t∗ (~x) are prescribed

~t(~x) = ~n(~x) · σ(~x) = ~t∗ (~x) on St (2.3)

where ~n(~x) denotes the unit outward normal vector on S.


To complete the system of equations (2.1)–(2.3) constitutive equations that
relate stress and strain for each material point are required. Constitutive
relations for a linear elastic material may be written in a general format as

σ(~x) = 4C(~x) : ε(~x) (2.4)

where 4C(~x) is a fourth order elasticity tensor, containing the material con-
stants and ε is the infinitesimal strain tensor given by
 
~ u + (∇~
ε = 21 ∇~ ~ u)T (2.5)

For an isotropic linear elastic material, Hooke’s law is widely used, for which
4C(~
x) is written as
4
C(~x) = κ(~x)II + 2G(~x)(4 IS − 13 II) (2.6)

where κ(~x) is the compressibility (bulk) modulus and G(~x) the shear modulus.
The dependence on the position vector ~x is here explicitly specified, since for
complex structures, e.g. made of several different materials, these material
properties may vary at different points. In equation (2.6) I denotes the second
order unit tensor; II is the dyadic product of two second-order unit tensors,
which gives the fourth-order tensor such that for an arbitrary second order
tensor Z we have II : Z = Itr(Z). The fourth-order symmetrization tensor
4 IS is defined by 4 IS : Z = ZS = 1 (Z + ZT ). (See appendix A for more
2
information on special second- and forth-order unit tensors). It can be easily
verified that definition (2.6) is equivalent to the more usual form of Hooke’s
law (with explicit dependence on the coordinate ~x omitted for clarity)

σ = κtr(ε)I + 2Gεd (2.7)

The aim is to find the displacement field ~u(~x) that satisfies equations (2.1)
and (2.4) subjected to boundary conditions (2.2) and (2.3). Additionally, the
strain field ε(~x) and the stress field σ(~x) may be calculated using (2.5) and
(2.4), respectively.
Unfortunately, in most cases (and almost all cases encountered in engi-
neering practice) the solution of a boundary value problem can not be found
analytically. This is mostly due to the complexity of the geometry and some-
times due to the complexity of the elasticity tensor. Therefore, one has to
rely on an approximate, numerical solution of the problem. One of the most
versatile and the most widely used numerical methods is the finite element
method. In the following, the finite element solution procedure of the problem
summarized above will be considered in detail.
CHAPTER 2. LINEAR FEM 12

2.2 Weak formulation


The finite element discretization of the boundary value problem is based on
a weak form of the respective differential equation. One way to derive the
weak form of a partial differential equation is by using the weighted residuals
approach. The weighted residuals form of equation (2.1) is obtained by taking
the scalar product with a vector-valued test function φ(~ ~ x) (also sometimes
called weight function) followed by integration over the domain V
Z  
~ x) · ∇
φ(~ ~ · σ(~x) + ~
q (~x) dV = 0 (2.8)
V

By application of the chain rule, it can be shown that1

∇ ~ =φ
~ · (σ · φ) ~ · (∇
~ · σ) + (∇ ~ T:σ
~ φ) (2.9)

Use of this result for the first term in (2.8) yields


Z   Z Z
T
~ ~
∇· σ(~x) · φ(~x) dV − ~ ~
∇φ(~x) : σ(~x)dV + φ(~ ~ x)·~q(~x)dV = 0 (2.10)
V V V

According to the divergence theorem


Z Z
~ · ~a(~x)dV = ~n(~x) · ~a(~x)dS
∇ (2.11)
V S

where ~a(~x) is a sufficiently smooth vector function on domain V and ~n(~x) is


the unit outward normal to the boundary S of the domain V .
Application of the divergence theorem to the first term of (2.10) gives after
rearrangement of terms
Z Z Z
~ x) T : σ(~x)dV = ~n(~x) · σ(~x) · φ(~
~ φ(~ ~ x)dS + φ(~~ x) · ~q(~x)dV (2.12)


V S V

On the surface S the stress vector ~t(~x) is defined as

~t(~x) = ~n(~x) · σ(~x) (2.13)

These tractions are either prescribed (on the part St of the boundary) or
unknown at the parts where displacements are prescribed (Su ). For a moment,
1
Using index notation the proof is given by

~ = ∂ ~ei · (σjk~ej ~ek · φm~em ) = ∂ ~ei · (σjk φk~ej )


~ · (σ · φ)

∂xi ∂xi
∂σjk ∂φk ∂σik ∂φk
= δij φk + σjk δij = φk + σik
∂xi ∂xi ∂xi ∂xi
=φ~ · (∇
~ · σ) + (∇ ~ T:σ
~ φ)
CHAPTER 2. LINEAR FEM 13

however, we will disregard this distinction between the essential and natural
parts of S. Then (2.12) takes the form
Z Z Z
T
~ ~ ~ ~
∇φ(~x) : σ(~x)dV = φ(~x) · t(~x)dS + φ(~ ~ x) · ~q(~x)dV (2.14)
V S V

~ x). Relation (2.14)


which must be satisfied for all possible test functions φ(~
is called the weak form of the differential equation (2.1). In contrast, (2.1) is
called the strong form.
The main rationale for rewriting the strong form into a weak form is to
reduce the continuity requirements imposed on the unknown field ~u(~x). Sub-
stitution of (2.4) and (2.5) into (2.1) (assuming that the fourth order tensor
4C(~x) possesses the required symmetry) yields
 
~ · 4C(~x) : ∇~
∇ ~ u(~x) + q~(~x) = ~0 (2.15)

from where it is immediately obvious that the strong form of the linear elastic
equilibrium problem involves the second derivatives of ~u(~x), thus requiring
~u(~x) to be twice differentiable. On the other hand, the weak form involves
only the first derivatives of ~u(~x)
Z Z Z
  
~ x) T : 4C(~x) : ∇~
~ φ(~
∇ ~ u(~x) dV = φ(~ ~ x)·~q(~x)dV (2.16)
~ x)·~t(~x)dS + φ(~
V S V

~ u appears only in the integrand, the continuity requirements


Moreover, since ∇~
on ~u(~x) can be relaxed even further: ~u(~x) must be continuous and needs to
be only piecewise differentiable. The price that has been paid for reducing
the conditions on ~u(~x) is a slightly stronger continuity requirement on the test
~ x) compared to (2.8): these functions must now be continuous and
functions φ(~
piecewise differentiable as well. However, this price is quite affordable because
~ x) is still no more stringent than that on ~u(~x).
this requirement on φ(~

2.3 Finite element discretization


2.3.1 Discretization of the problem domain
Now that the strong form of the linear elastic equilibrium problem (2.1), has
been replaced by the weak form (2.16) or (2.14), we can start constructing an
approximate solution of the boundary value problem. To this end, the domain
V is first subdivided into a number of finite-sized elements, as schematically
shown in Figure 2.2. Typical element shapes in two dimensions are trian-
gles and quadrilaterals and in three dimensions – tetrahedra and hexahedra
(bricks). The discretized domain is denoted by V h . Likewise, the discretized
boundary S h replaces S. For simple geometries V and V h may be identical,
but for more complex shapes this discretization involves some approximation.
CHAPTER 2. LINEAR FEM 14

V \V h

Sh S
V h

Figure 2.2: A quadrilateral finite element discretization V h of a


problem domain V .

An unknown field, in our case it is the displacement field ~u(~x), is approxi-


mated by
n
X
~uh (~x) = Ni (~x)~ui (2.17)
i=1

where n is the total number of nodes and Ni (~x) are the scalar-valued functions,
called the shape functions, which are used to interpolate discretized fields on
the basis of nodal values ~ui . Each node i has a corresponding shape function
Ni (~x). The shape functions should satisfy the property that they equal one in
the node to which they correspond and vanish in all the other nodes

Ni (~xj ) = δij , i, j = 1, . . . , n (2.18)

with δij the Kronecker delta.


In the following a more compact column/matrix notation will be used.
Storing the nodal displacement vectors ~ui and the shape functions Ni (~x) in
column matrices
h iT
~u = ~u1 ~u2 ~u3 . . . ~un
˜ h iT
N(~x) = N1 (~x) N2 (~x) N3 (~x) . . . Nn (~x) (2.19)
˜
relation (2.17) can be rewritten as

~uh (~x) = NT(~x) ~u (2.20)


˜ ˜
2.3.2 Discretization of the weak form
Now the weak form (2.16) can be rewritten on the discretized domain V h
rather than the exact domain V with the unknown field ~u(~x) replaced by the
approximation ~uh (~x) according to (2.20). Furthermore, following the Galerkin
approach the test functions are approximated in the same form as the solution,
thus

φ ~
~ h (~x) = NT(~x) φ (2.21)
˜ ˜
CHAPTER 2. LINEAR FEM 15

where the shape function column N(~x) is identical to that in (2.20) and the
column φ ˜
~ contains n vector constants ~ i . Since the nodal values contained
φ
˜~ are constants, the gradients of the approximate displacement field
in ~u and φ
˜
and of the˜ shape functions are simply calculated as

~ uh (~x) = ∇N~ (~x) T ~u, ~φ ~ (~x) T φ


~ h (~x) = ∇N ~
 
∇~ ∇ (2.22)
˜ ˜ ˜ ˜
~
where ∇N(~x) stands for a column of vectors which is defined as
˜
h iT
~ (~x) = ∇N
∇N ~ 1 (~x) ∇N
~ 2 (~x) . . . ∇N~ n (~x) (2.23)
˜
Substitution of (2.21) and (2.22) into the weak form (2.16) gives
Z
~ T ∇N ~ (~x) T~u dV
~ (~x) : 4 C(~x) : ∇N

φ
˜ ˜ ˜ ˜
Vh
Z Z (2.24)
~
= N (~x)φ · t(~x) dS + NT (~x)φ
T ~ ~ · ~q(~x) dV
˜ ˜ ˜ ˜
S h h V

Since columns ~u and φ ~ contain constants (nodal values), they can be taken
˜ ˜
out of the integrals, thus we have
Z 
T T
~
φ · ~ 4 ~
∇N(~x) · C(~x) · ∇N(~x) dV · ~u
˜ ˜ ˜ ˜
Vh
Z Z (2.25)
T T
~ ~ ~
=φ · N(~x)t(~x) dS + φ · N(~x)~q(~x) dV
˜ ˜ ˜ ˜
Sh Vh

Now we should recall that the weak form (2.16) had to hold for all test func-
~ x). Since the test functions φ
tions φ(~ ~ h (~x) are fully determined by the coef-
~
ficients φ, requiring that the weak form is satisfied for all φ(~ ~ x) = φ~ h (~x) is
˜ to requiring that equation (2.25) holds for all φ
equivalent ~ . This, in turn, is
only possible if and only if ˜
Z !
~ (~x) · 4 C(~x) · ∇N T
~ (~x) dV · ~u

∇N
˜ ˜ ˜
Vh (2.26)
Z Z
= N(~x)~t(~x) dS + N(~x)~q(~x) dV
˜ ˜
Sh Vh

The integral on the left-hand side of this equation forms an n × n matrix of


second-order tensors. If we denote this tensor matrix by K
¯
Z
~ (~x) T dV
~ (~x) · 4 C(~x) · ∇N

K= ∇N (2.27)
¯ ˜ ˜
Vh
CHAPTER 2. LINEAR FEM 16

and furthermore define the vector columns


Z
~q = N(~x)~
q (~x) dV (2.28)
˜ ˜
Vh
Z
~f = N(~x)~t(~x) dS (2.29)
˜ ˜
Sh

then equation (2.26) may be rewritten as


K · ~u = ~q + ~f (2.30)
¯ ˜ ˜ ˜
The matrix K is called system matrix or stiffness matrix. The sum ~q + ~f is
¯ ˜ a
usually referred to as the right-hand side of the system. We thus end up˜ with
linear system of equations with the column of unknown vectors ~u, which can be
readily solved, after incorporation of boundary conditions (as ˜discussed later
in section 2.3.6). However before doing so, we shall look for efficient ways to
numerically calculate the integrals (2.27)–(2.29).

2.3.3 Element coordinates and isoparametric mapping


In practice, finite elements are usually developed with shape functions ex-
pressed in terms of master (also called parent) element coordinates. The most
widely used master element coordinates are triangular coordinates and isopara-
metric coordinates. As an example, a square, four-node isoparametric master
element is shown in Figure 2.3(a). Associated with this element is a normalised
coordinate vector ξ, ~ with Cartesian components ξ1 and ξ2 that range from −1
to 1. The column of shape functions of the four-node element in Figure 2.3(a)
can be written as
 
1
(1 − ξ 1 )(1 − ξ 2 )
 14 
e~
 4 (1 + ξ1 )(1 − ξ2 )
N (ξ) =  1
  (2.31)
˜ 4 (1 + ξ 1 )(1 + ξ )
2 

1
4 (1 − ξ1 )(1 + ξ2 )

It is easy to verify that the shape functions (2.31) indeed satisfy condition
(2.18) and that they provide a bilinear interpolation in the interior of the
element. Examples of some other most widely used elements together with
the corresponding shape functions can be found in [2].
Unlike the master element, the elements that are derived from it need not
to be square, but may take other quadrilateral shapes that depend on the
true position of the nodes associated with them. An example of the physical
shape of the four-node quadrilateral element is given in Figure 2.3(b). The
relation between the parent domain of an element and the physical domain of
the element is established by the isoparametric mapping, given by
~x(ξ) ~ xe
~ = NeT(ξ)~ (2.32)
˜ ˜
where the vector column ~xe contains the real (physical, spatial) position vectors
of the element nodes. ˜
CHAPTER 2. LINEAR FEM 17

ξ2 3
4 3 4
1

-1 0 1 ξ1
2
-1
1 2 1
(a) (b)

Figure 2.3: Bilinear quadrilateral master element (a) and its pos-
sible physical shape (b).

2.3.4 Isoparametric transformation


Computationally it is more efficient to evaluate the integrals in the expres-
sions (2.27)–(2.29) not directly, but element by element and in terms of the
normalised element coordinates associated with each element. Using isopara-
metric mapping (2.32), the gradient ∇ ~ with respect to ~x, which appears in the
system matrix and the right-hand side, can be rewritten as a gradient with
~ which we denote by ∇
respect to ξ, ~ ξ . In a three-dimensional Cartesian basis
this gradient is defined as

~ ξ = ~e1 ∂ + ~e2 ∂ + ~e3 ∂ = ~ei ∂


∇ (2.33)
∂ξ1 ∂ξ2 ∂ξ3 ∂ξi
where Einstein summation convention has been used.
~ ξ and ∇
The relation between the operators ∇ ~ is obtained by applying the
chain rule
~ ~
~ ξ = ~ei ∂ = ~ei ∂xj (ξ) ∂ = ~ei ∂xj (ξ) ~ej · ~ek ∂ = ∇
∇ ~ ·∇
~ ξ ~x(ξ) ~ ·∇
~ = J (ξ) ~
∂ξi ∂ξi ∂xj ∂ξi ∂xk
(2.34)
~ has been defined as
where the Jacobian tensor J (ξ)
~ =∇
J (ξ) ~ =∇
~ ξ ~x(ξ) ~ ~xe
~ ξ NeT(ξ) (2.35)
˜ ˜
where the isoparametric mapping (2.32) has been substituted. Using the above
~ can be easily calculated for any given ξ.
relation, the Jacobian tensor J(ξ) ~ Pre-
~
multiplying equation (2.34) by the inverse of J (ξ) yields the transformation
of gradients in the opposite direction

∇ ~ ∇
~ = J −1(ξ)· ~ξ (2.36)

The above expression can be used to compute the gradients of the shape
functions with respect to ~x which appear in the definition for the system matrix
CHAPTER 2. LINEAR FEM 18

(2.27). For the relevant shape functions on element e we have:


∇N ~ ∇
~ e = J −1(ξ)· ~ ξ Ne (2.37)
˜ ˜
in which ∇~ ξ Ne follows simply by differentiation of the local shape functions
e ~ ˜
N (ξ). The remaining factor 4 C(~x) in the integrand of (2.27) can be written
˜ terms of ξ~ by directly using the mapping (2.32) on each element. The
in
contribution of element e to the integral can now be transformed to an integral
on a cube Q = (−1, 1) × (−1, 1) × (−1, 1) as
Z Z
e ~ e4 ~ eT ~ ·∇N
~ e· 4 C ~x(ξ) ~ dQ (2.38)
~ eT det J(ξ)
 
K = ∇N · C(~x)·∇N dV = ∇N
¯ ˜ ˜ ˜ ˜
Ve Q

where the factor det J(ξ) appears as a consequence of the coordinate trans-
formation. This can be seen by writing the infinitesimal volume dV as
     
dV = d~x1 × d~x2 · d~x3 = dξ~1 · J (ξ)
~ × dξ~2 · J(ξ)
~ · dξ~3 · J(ξ)
~

~ dξ~1 × dξ~2 · dξ~3 = det J (ξ)


~ dQ
 
= det J (ξ) (2.39)
The contribution of element e to the right-hand side term ~q given by (2.28),
˜
which is related to the body forces ~q(~x), can be written analogically to the
above as
Z Z
e ~ q (~x(ξ))
q (~x) dV = Ne (ξ)~ ~ det J(ξ)
~ dQ

~q = N(~x)~ (2.40)
˜ ˜ ˜
eV Q

The contribution which stems from the influence of the boundary trac-
tions, ~f (2.29), requires a slightly different treatment since it involves a surface
integral˜ rather than a volume integral. This implies that elements in the in-
terior of the domain do not contribute and only one layer of elements at the
surface of the domain needs to be considered. Furthermore, on the part of the
boundary on which an essential boundary condition is acting, Suh , the tractions
~t(~x) are yet unknown. Indeed, these entries will be a result of the computa-
tion after partitioning and solving the reduced system, as will be discussed in
section 2.3.6.
A useful property of isoparametric element families with respect to the cal-
culation of surface integrals is that the shape functions of a three-dimensional
element on each of its faces (i.e. for one component of ξ~ equal to plus or mi-
nus one) reduce to the shape functions of the corresponding two-dimensional
element. Similarly, the shape functions of a two-dimensional element reduce
to those of the corresponding one-dimensional element on each of its edges.
This can easily be seen for the bilinear master element of Figure 2.3(a), with
the shape functions given by (2.31).
The contribution of an element e to the boundary integral (2.29) is then
calculated by
Z Z
~f e = Ne(~x) ~t(~x) dS = Ne(ξ~′ ) ~t(~x(ξ~′ )) det J ′ (ξ~′ ) dQ′

(2.41)
˜ ˜ ˜
Se Q′
CHAPTER 2. LINEAR FEM 19

where the prime denotes local coordinates and Jacobian associated with the
lower-dimensional boundary element. Note, that in practice zero tractions
~t∗ = ~0 are often prescribed on (a part of) St (so called traction-free boundary).
In this case the corresponding contribution to ~f will be, of course, simply zero.
˜
2.3.5 Numerical integration
Evaluating the integrals in (2.38), (2.40) and (2.41) analytically is usually
cumbersome, if not impossible. This is not only due to the function 4C(~x),
which may in principle be arbitrary, but also to the determinant of the Ja-
cobian, which very easily gets a complicated form for arbitrarily shaped ele-
ments. For this reason, numerical integration, also called quadrature rule, is
generally used. Quadrature rules replace the integration by a weighted sum
of the integrand values in a number of points, the so-called integration points.
The most widely applied numerical integration rule in finite element analyses
is Gauss quadrature. The Gauss quadrature selects the integration points ξ~k
(called in this case Gauss points) in such a way that together with the corre-
sponding weight factors wk they provide an exact integration for polynomials
of the highest possible degree for a given number of points q. Positions of
Gauss points and weight factors for some most often used element types can
be found, for example, in [2].
The numerical integration of (2.38) yields
q
~ e · 4 C ~x(ξ~k ) · ∇N
~ eT det J (ξ~k )
X
e
 
K = wk ∇N (2.42)
¯ ˜ ξ=ξk~ ~ ˜ ξ=ξk ~ ~
k=1

Applying the quadrature rule to (2.40) gives


q
wk Ne(ξ~k )~
q ~x(ξ~k ) det J (ξ~k )
X
e
 
~q = (2.43)
˜ k=1
˜

Finally, using the appropriate quadrature rules for the corresponding surface
element (or line element in two dimensions), the integral in (2.41) can be
approximated by
q′
e
~f = wk′ Ne(ξ~k′ ) ~t ~x(ξ~k′ ) det J ′ (ξ~k′ )
X  
(2.44)
˜ k=1
˜

2.3.6 Assembly, partitioning and solving


e
Having computed the element contributions Ke , ~qe and ~f for each element e as
¯ ˜ ˜ number of elements):
described above, the global system can be assembled (m
m m m
e
K=
¯ A
e=1
Ke
¯
~q =
˜
A ~qe
e=1 ˜
~f =
˜ A~˜f
e=1
(2.45)

During this assembly, for example, the second-order tensor Keij in the element
¯
matrix of an element e must be added to the global matrix K in row i, column
¯
CHAPTER 2. LINEAR FEM 20

j, where i is the global node number corresponding to the first node of the ele-
ment and j corresponds to the second node. Storing all element contributions
at the appropriate positions in the global system thus requires some knowledge
on which nodes are connected to which element, i.e. on the mesh connectivity.
This bookkeeping, however, can easily be dealt with by a computer code.
Note that the above assembly operations act on tensorial and vectorial
components. Indeed, using modern programming methods all the computa-
tions and operations discussed so far, as well as the subsequent solution of the
system can be performed directly in terms of vectorial and tensorial quantities.
However, in most of the finite element software such an option is not available,
and the system of equations thus has to be rewritten in scalar format. These
scalar equations can be obtained by defining a basis and writing the tensors
and vectors in terms of this basis. The components with respect to the basis
will then be the scalar variables of the problem.
Finally, before solving the global assembled system of equations, the essen-
tial boundary conditions have to be taken into account. This is usually done
by partitioning of the system2 . Here we assume that in a given boundary
node i either the displacement vector ~ui or the boundary traction vector ~ti is
entirely prescribed. Then the global assembled system of equations (2.30) can
be partitioned as follows
" # " # " # " #
Kuu Kuf ~u ~q ~f
¯ ¯ · ˜u = ˜u + ˜u (2.46)
Kfu Kff ~uf ~qf ~f
f
¯ ¯ ˜ ˜ ˜
where ~uu contains those displacements, which are given by the essential bound-
˜
ary conditions, and ~uf the remaining degrees of freedom which are still “free”.
On the other hand,˜~f u contains the unknown boundary tractions associated
˜
with the prescribed values of ~uu , whereas ~f f is fully known. Note that rewrit-
ing the original system in this˜form may require ˜ some reordering of rows and
columns of K. Also note that the matrices Kuu and Kff will always be square,
¯ ¯ ¯
because an unknown ~ti is always associated with a prescribed ~ui and vice versa.
Now the partitioned system of equations can be split into two separate
matrix equations:

Kuu · ~uu + Kuf · ~uf = ~qu + ~f u (2.47)


¯ ˜ ¯ ˜ ˜ ˜
Kfu · ~uu + Kff · ~uf = ~qf + ~f f (2.48)
¯ ˜ ¯ ˜ ˜ ˜
In the second of these equations the only unknowns are the components of
~uf . This means that this equation can be rewritten in the form of a standard
˜
linear system

Kff · ~uf = ~qf + ~f f − Kfu · ~uu (2.49)


¯ ˜ ˜ ˜ ¯ ˜
in which the right-hand side is fully known and the matrix Kff is square. This
¯
system can thus be solved for the unknown displacements ~uf . Combining this
˜
2
Other ways include adding the essential boundary conditions in forms of constraints, that
are enforced by, for example, the penalty method or Lagrange multipliers.
CHAPTER 2. LINEAR FEM 21

solution with the already known ~uu according to


˜
" #
~u
~u = ˜u (2.50)
˜ ~uf
˜
the entire set of degrees of freedom is known and the approximate solution
~uh (x) thus is fully determined. If necessary, the unknown reaction forces ~f u
˜
can be retrieved from the top part of the partitioned system (equation (2.47))
according to:
~f = K · ~u + K · ~u − ~q (2.51)
u
˜ ¯ uu ˜u ¯ uf ˜f ˜u
In practice, so-called mixed boundary conditions are often encountered.
This means that, for example in a two dimensional case, in a node one com-
ponent of the displacement vector (with respect to some vector basis) and
one component of traction vector are prescribed. Another possibility is that
a component normal to the surface of, say, the traction vector and a tangen-
tial component of the displacement vector may be specified. These boundary
conditions can be prescribed in a similar manner as discussed above, except
this has to be done on the system expanded from the vector formulation into
scalars.
Chapter 3

Total Lagrange procedure for


geometrically non-linear
problems

3.1 Geometrically non-linear boundary value prob-


lem
Consider a body that in the reference (undeformed) configuration B0 occupied
volume V0 and had boundary S0 and after deformation, in the current con-
figuration B, the body became of volume V with boundary S, see Figure 3.1.
In small deformation theories it is assumed that the reference and current
configurations are so close together that they may be taken to coincide. If
the deformation of the body is such that this assumption may not be made,
then it is said that the body undergoes large deformations, and special large
deformation theories and numerical frameworks are required.
ψ
S0 S V
V0
P
P0
B0 ~u B

~0
X ~x
t=0 t
0

Figure 3.1: The non-linear deformation map.

Let X~ 0 denote the position of a particle in the reference configuration. At


time t the position of the same particle in the current configuration is given
by ~x. The transformation between these two states may be described by the
so-called non-linear deformation map ~x = ψ(X ~ 0 , t). The non-linear deforma-
tion map is unique, i.e. a one-to-one correspondence exists between particles

22
CHAPTER 3. TOTAL LAGRANGE 23

with position vector ~x at time t and material points X ~ 0 in the undeformed


state; and continuous, i.e. neigbourhoods remain neighbourhoods, no cracks
occur; mathematically this means that the non-linear deformation map is dif-
ferentiable. These properties of the non-linear deformation map imply that
both the material (or Lagrangian) coordinates (X01 , X02 , X03 ) of a particle in
the reference configuration or the spatial (or Eulerian) coordinates (x1 , x2 , x3 )
in the current configuration could be used as independent variables to de-
scribe the deformation. If the material coordinates (X01 , X02 , X03 ) are taken
as independent variables, attention is focused on a definite material particle
that moves in space. This is the Lagrangian description. If the coordinates
(x1 , x2 , x3 ) are taken as independent variables, one fixes in space a geometrical
point (x1 , x2 , x3 ) and investigates which particle, at a certain time t, passes
through that point. This is the Eulerian description. Although use of one
or another description leads to mathematically different formulations, these
two descriptions are equivalent. Theoretically, for every problem any of the
descriptions may be used and a particular choice is mostly dictated by the
convenience. For this reason, in solid mechanics the Lagrangian description
is mostly used, while in fluid mechanics the Eulerian description is usually
adopted. Since in this course the attention is focused on non-linear finite el-
ements for solids, in the following the Lagrangian treatment will be applied.
The Eulerian approach in finite element formulations is considered in other
courses.
An infinitesimal material line element d~x in the current configuration is
related to the infinitesimal material line element dX ~ 0 in the reference config-
uration by (see Figure 3.2)
∂ψ T
 
d~x = · dX~ 0 = F · dX
~0 (3.1)
∂X~0
where the deformation gradient tensor F is defined as
~ 0 ~x)T
F = (∇ (3.2)
~ 0 the gradient operator with respect to the reference configuration.
with ∇

S0 F S V
V0

~0
dX d~x
B0 B

t=0 t

Figure 3.2: The linearized deformation map – the deformation


gradient tensor F.

In absence of body forces the equilibrium problem is described by


~ · σ(~x) = ~0
∇ (3.3)
CHAPTER 3. TOTAL LAGRANGE 24

where σ is the Cauchy stress tensor and ∇ ~ denotes the gradient operator with
respect to the current configuration. The equilibrium equation (3.3) should
hold on the current domain V and should be supplemented by appropriate
boundary conditions. As usually, these may be the essential (Dirichlet) or
natural (Neumann) boundary conditions. Assume that the displacement field
~u is specified along a certain part of the boundary Su ⊂ S
~u(~x) = ~u∗ (~x) on Su (3.4)
On the remaining part of the boundary, St = S\Su tractions ~t∗ (~x) are pre-
scribed
~t(~x) = ~n(~x) · σ(~x) = ~t∗ (~x) on St (3.5)
where ~n(~x) denotes the unit outward normal vector on the current surface S.
Note that in a large deformation framework the boundary conditions can be
defined with respect to the reference or the current configuration. One has to
be extremely careful as to in which configuration the boundary conditions are
defined, since in some problems this may lead to erroneous results.
To complete the above system of equations relations between a stress mea-
sure (e.g. the Cauchy stress tensor σ) and a kinematical quantity (e.g. the
deformation gradient tensor F) are required. This relations are called consti-
tutive equations and describe the (thermo-)mechanical response of a particular
material when subjected to a (non-isothermal) deformation history. In general
(omitting thermal effects), a constitutive relation may be written in the form
σ(~x, t) = F{F(~x, τ ); τ ≤ t} (3.6)
Constitutive equations have to satisfy a number of physical requirements,
known as the fundamental principles of constitutive equations. These princi-
ples include the principles of determinism, local action and frame-indifference.
The principle of frame-indifference, also called the principle of objectivity, is
especially important for a correct formulation of constitutive equations for
large deformation problems. This principle states that a constitutive equa-
tion must be invariant under changes of the coordinate system, to which the
stresses are referred. In other words, the state of stress in a material does not
change if the body undergoes rigid translation or rotation. For more details on
the fundamental principles of constitutive equations see [1] (chapter 6). For
example, it is easy to show that the small strain elastic constitutive relation
(2.7) does not satisfy the principle of objectivity, and thus leads to unphysical
results when used for large deformations. For instance, it predicts a non-zero
stress state in a body undergoing rigid rotation.
In these chapters the following elastic constitutive law will be used
τ = κ(J − 1)I + G(B − I) (3.7)
where κ is the compressibility (bulk) modulus and G the shear modulus; J =
det(F) is the volume change ratio; τ is the Kirchhoff stress tensor related to
the Cauchy stress tensor σ by
τ = Jσ (3.8)
CHAPTER 3. TOTAL LAGRANGE 25

and B is the Finger tensor defined as


B = F · FT (3.9)
It is easy to show that the constitutive relation (3.7) is objective, and thus
does not suffer from the same drawbacks at large deformations as the small
strain relation (2.4).
~ 0 , t) that satisfies the
The aim is now to find the position vector field ~x(X
equations (3.3) and (3.7) subjected to the boundary conditions (3.4) and (3.5).
Note, that the time parameter t does not necessarily describe a physical time,
but may, for example, correspond to a loading parameter. In general, it may be
associated with each configuration, through which the body passes while being
deformed from its initial configuration to its final configuration. Additional
difficulty in the problem at hand arises from the fact that the equilibrium
equation (3.3) includes the divergence operator with respect to the current
configuration, which is not yet known. Moreover, the constitutive equations
are now formulated as non-linear functions of the unknown (gradient of) the
position vector field.

3.2 Weak formulation


As it has been already established for the linear finite element solution pro-
cedure, we start by writing a weighted residuals form of equation (3.3). This
form is obtained by taking the scalar product of (3.3) with a vector-valued
test (weight) function φ(~ X ~ 0 ) followed by integration over the current volume
V of the body. Note, that the test functions are defined with respect to po-
sition vectors X ~ 0 in the initial configuration and are not functions of time.
Therefore, we have
Z  
~ X
φ( ~ 0) · ∇~ · σ(~x) dV = 0 (3.10)
V
~ X
This equation must be satisfied for all admissible test functions φ( ~ 0 ), then
the weighted residuals problem (3.10) is equivalent to (3.3). Following the
same procedure as in section 2.2, applying the chain rule and the divergence
theorem, leads to the weak form of the differential equation (3.3)
Z Z
T
~ ~ ~
∇φ(X0 ) : σ(~x) dV = φ( ~ X~ 0 ) · ~t(~x) dS (3.11)
V S

where the traction vector ~t = ~n · σ has been defined on the current surface S
of the body.

3.3 Transformation to the undeformed configura-


tion
The weak form of the equilibrium equations, (3.11), is formulated in the cur-
rent, deformed state. Namely, the integration is carried out on the deformed
CHAPTER 3. TOTAL LAGRANGE 26

configuration V and the gradient of the test function ∇ ~ is defined with re-

spect to the position vector ~x in the deformed state. However, this deformed
configuration is unknown and yet to be determined from the solution of the
problem. Therefore, it is convenient to rewrite the left-hand side of (3.11)
in terms of the (known) reference configuration V0 . In the general case of
a distributed load along the boundary S the right-hand side must also be
transformed to the undeformed configuration, but here it is left unchanged for
simplicity and interpreted later on, in chapter 5.
The infinitesimal volume in the current configuration dV is related to the
infinitesimal volume in the reference configuration dV0 by the volume ratio
J = det(F) (see [1], chapter 2)

dV = JdV0 (3.12)

Therefore, integration over the current, unknown, domain V of an arbitrary


function f (~x) may be transformed into an integration over the reference do-
main V0 by means of
Z Z
f (~x) dV = f (~x)J(~x) dV0 (3.13)
V V0

~ with respect to the current configuration is related to


The gradient operator ∇
the gradient operator with respect to the reference undeformed configuration
~ 0 by the deformation gradient tensor F according to ([1], chapter 2)

~ = F−T · ∇
∇ ~0 (3.14)

Using this expression it follows that


 
~φ~ T : σ = F−T · ∇ ~ T:σ= ∇
~ 0φ ~ T · F−1 : σ = ∇
~ 0φ ~ T : F−1 · σ
~ 0φ
    

(3.15)

Therefore, the weak form (3.11) can be rewritten in the undeformed con-
figuration as
Z Z
T
~ ~ ~ −1 ~ X~ 0 ) · ~t(~x) dS

∇0 φ(X0 ) : F (~x) · σ(~x) J(~x) dV0 = φ( (3.16)
V0 S

3.4 Finite element discretization


Now the weak form (3.16) can be rewritten on the discretized domain V0h
rather than the exact domain V0 . The unknown field of the position vectors
~x(X~ 0 , t) in the current configuration is approximated by

n
~ 0 , t) = ~0 )~x(t)
~ 0 )~xi (t) = NT(X
X
~xh (X Ni (X (3.17)
i=1
˜ ˜
CHAPTER 3. TOTAL LAGRANGE 27

where the column/matrix notation introduced in section 2.3.1 has been used.
Here Ni (X ~ 0 ) is the scalar-valued shape function at node i; ~xi (t) the position
vector of this node in the current configuration at time t; n the total number
of nodes; N(X ~0 ) is the column containing the shape functions and ~x(t) is the
˜
column of the nodal position vectors. Here it is important to notice ˜ that
the shape functions N(X ~0 ) depend exclusively on the coordinates X ~ 0 in the
undeformed configuration ˜ (material coordinate), while the dependence on time
is fully concentrated in the positions of the nodes. If we write (3.17) for a node
j with the initial position X ~ 0j , we have
n
~ 0j , t) = ~ 0j )~xi (t) = ~xj (t)
X
h
~x (X Ni (X (3.18)
i=1

where the property (2.18) of the shape functions has been used. From the
above relation we see that node j always (in each deformed configuration)
corresponds to the same material point X~ 0j . This is an important property of
Lagrangian meshes.
Following the Galerkin approach, the test functions are approximated with
the same interpolation (shape) functions as the solution, thus
~ h (X
φ ~ 0 ) = NT(X ~
~ 0) φ (3.19)
˜ ˜
where the shape function column N(X ~ 0 ) is identical to that in (3.17) and the
column φ ~ contains n vector constants φ ˜ ~ i . Note once more, that the test func-
˜
tions are not functions of the time. Now the gradients of the shape functions
are simply calculated as
~ 0φ ~ h (~x) = ∇ ~ 0 N(X ~ 0) T φ
~

∇ (3.20)
˜ ˜
where ∇ ~ 0 N(X ~ 0 ) stands for a column of vectors which is defined as
˜ h iT
~ 0 N(X
∇ ~ 0) = ∇ ~ 0 N1 (X~ 0) ∇~ 0 N2 (X
~ 0) . . . ∇ ~ 0 Nn (X
~ 0) (3.21)
˜
Substitution of (3.19) and (3.20) into the weak form (3.16) gives
Z Z
T
~ ~ ~
φ ∇0 N(X0 ) : F (~x) · σ(~x) J(~x) dV0 = NT (X
−1 ~ · ~t(~x) dS (3.22)
~ 0 )φ

˜ ˜ ˜ ˜
V0h hS

Taking the column with constants φ~ out of the integrals gives


Z ˜  Z
T T
~ ~ ~ −1 ~ ~ 0 )~t(~x) dS (3.23)

φ · ∇0 N(X0 )· F (~x) · σ(~x) J(~x) dV0 = φ · N(X
˜ ˜ ˜ ˜
V0h Sh

The requirement that the weak form (3.11) must be satisfied for all test func-
~ X
tions φ( ~ . This
~ 0 ) of the form (3.19) implies that (3.23) must be valid for all φ
is only possible if ˜
Z Z
~ 0 N(X ~ 0 ) · F−1 (~x) · σ(~x) J(~x) dV0 = N(X~ 0 )~t(~x) dS

∇ (3.24)
˜ ˜
V0h Sh
CHAPTER 3. TOTAL LAGRANGE 28

This equation forms a set of non-linear algebraic equations in terms of


nodal positions ~x(t) (explicit dependence on time will be omitted in the fol-
˜
lowing for brevity)
~f (~x) = ~f (3.25)
int ext
˜ ˜ ˜
with
Z
~f (~x) = ~ 0 N(X
~ 0 ) · F−1 (~x) · σ(~x) J(~x) dV0

int ∇ (3.26)
˜ ˜ ˜
V0h
Z
~f ext = ~ 0 )~t(~x) dS
N(X (3.27)
˜ ˜
Sh

The right-hand side of (3.25), ~f ext , can be interpreted as the column with
external nodal forces, i.e. as the ˜discrete representation of the boundary trac-
tions acting at the surface S of the deformed configuration. If relevant, body
forces can also be incorporated in this term. Note again, that here it has been
assumed that on the boundary St , where the traction boundary conditions are
specified, forces prescribed on the boundary are fixed and do not depend on
the deformation of the boundary, so that
~t∗ dS0 = ~t∗ dS (3.28)

holds. For example, for a traction-free boundary (zero tractions on the bound-
ary) the above condition holds trivially. Therefore, here the external nodal
forces ~f ext do not depend on the solution ~x. If this assumption can not be made,
˜ force boundary conditions non-linearity
i.e. if the ˜ is important, then a special
treatment of this term is required; it will be considered later in chapter 5.
The term ~f int (~x), on the left-hand side of (3.25), represents the set of
internal forces˜in the ˜ material as a result of the deformation process. These
internal forces must balance the external forces.
Since here the constitutive law (3.7) is given in terms of the Kirchhoff
tensor τ , it is convenient to rewrite (3.26) as
Z
~f (~x) = ~ 0 N(X
∇ ~ 0 ) · F−1 (~x) · τ (~x) dV0 (3.29)
int
˜ ˜ ˜
V0h

where relation (3.8) between the Cauchy stress tensor σ and the Kirchhoff
stress tensor τ has been used. Note, that this step is done here merely for the
convenience of further derivations.
Solving the set of algebraic equations (3.25) yields a solution ~x(t) in terms
of nodal position vectors, and thus ~x(X ~ 0 , t), which satisfies the˜ equilibrium
equations in a weak sense, i.e which satisfies the weak form (3.11) of the
equilibrium problem.
If we define residual ~r, which represents out-of-balance forces, as
˜
~r(~x) = ~f int (~x) − ~f ext (3.30)
˜˜ ˜ ˜ ˜
CHAPTER 3. TOTAL LAGRANGE 29

then equation (3.25) can be rewritten as

~r(~x) = ~0 (3.31)
˜˜ ˜
Therefore, in the equilibrium state the residual (out-of-balance forces) should
vanish.
The most widely used method for the solution of non-linear algebraic equa-
tions of the type (3.31), and perhaps one of the most robust methods, is New-
ton’s method. In computational mechanics and particularly in finite element
environments it is often called the Newton-Raphson method. In the follow-
ing first Newton’s method will be discussed in somewhat more detail; then it
will be applied to the finite element solution of large deformation elasticity
equilibrium problems.

3.5 Newton’s method for solution of non-linear equa-


tions
3.5.1 Non-linear equation in one dimension
First Newton’s method will be illustrated on one single non-linear equation
with one unknown. Consider a non-linear equation

r(x) = 0 (3.32)

where r is a differentiable function of x. The goal is to find a solution x∗ , such


that the above equation (3.32) is satisfied.
Unlike linear equations, most non-linear equations cannot be solved in a
finite number of steps. Thus, one must usually resort to an iterative method
that produces increasingly accurate approximations to a solution, and the
iterations are terminated when the result is sufficiently accurate. Let xi denote
an approximate to the solution x∗ at iteration i. Newton’s iterative method is
based on a Taylor expansion of the function around the current iterative value
of the argument and setting the result to zero

0 ≈ r(xi+1 ) = r(xi ) + r ′ (xi )δx + O(δx2 ) (3.33)

where
dr
r ′ (xi ) = (3.34)
dx x=xi

and

δx = xi+1 − xi (3.35)

If the terms higher order than linear in δx are dropped, then (3.33) gives a
linear equation for δx (approximation)

r(xi ) + r ′ (xi )δx = 0 (3.36)


CHAPTER 3. TOTAL LAGRANGE 30

The above is called a linear model or linearized model of the non-linear equa-
tion. The linear model is the tangent to the non-linear function. The process
of obtaining the linear model is called linearization.
Assuming r ′ (xi ) 6= 0, this linear model can be easily solved for δx (which
is often called iterative correction)
−1
δx = − r ′ (xi ) r(xi ) (3.37)

The new iterative value of the unknown is then obtained by rewriting (3.35)
as

xi+1 = xi + δx (3.38)

In Newton’s algorithm, the solution of the non-linear equation is obtained


by iteratively solving a sequence of linear models (3.37)–(3.38). To begin
the iterative procedure, a starting value for the unknown x0 must be chosen.
The process is continued until a solution is obtained with the desired level of
accuracy. The Newton’s procedure is schematically illustrated in Figure 3.3.

r(x)

r(xi )
linear model
(tangent)

x∗ x
xi xi+1 xi+2

Figure 3.3: Newton’s method for solving the non-linear equation


r(x) = 0.

3.5.2 System of non-linear equations


Newton’s method is easily generalized to a system of non-linear equations by
replacing the above scalar equations by matrix equations. Consider a system
of n non-linear equations with n unknowns

r(x) = 0 (3.39)
˜˜ ˜
Neglecting terms higher order than linear in δx, the truncated Taylor series
˜
of the function r around the current iterative approximate xi is written as (cf.
(3.36)) ˜ ˜

r(xi ) + K(xi )δx = 0 (3.40)


˜˜ ¯ ˜ ˜ ˜
CHAPTER 3. TOTAL LAGRANGE 31

where the matrix K is called the system Jacobian matrix and is defined as
¯
∂r ∂rs
K = ˜ with components Kst = (3.41)
¯ ∂x ∂xt
˜
Relation (3.40) is the linear model of the non-linear equations. Newton’s
method replaces a system of non-linear equations by a sequence of systems
of linear equations. The iterative correction δx is obtained by solving a sys-
tem of linear algebraic equations ˜

K(xi )δx = −r(xi ) (3.42)


¯ ˜ ˜ ˜˜
Once the above system of equations has been solved (by any method for the
solution of linear equations), the new iterative values of the unknowns are
updated as

xi+1 = xi + δx (3.43)
˜ ˜ ˜
Note, that (3.40) may also be written as the value of the function at iter-
ation i, ri ≡ r(xi ), plus some iterative correction δr
˜ ˜˜ ˜
ri + δr = 0 (3.44)
˜ ˜ ˜
with the iterative correction δr related to the iterative correction of the un-
knowns δx by the Jacobian matrix ˜
˜
δr = K(xi )δx (3.45)
˜ ¯ ˜ ˜
This reasoning leads to exactly the same system of linear equations (3.42), but
is more often adopted in linearization of non-linear finite element models.

3.5.3 Convergence rate of Newton’s method


The rate of the convergence of the iterations in Newton’s method is quadratic
when the Jacobian matrix K satisfies certain conditions. These conditions
¯
may roughly be summarized as follows:

• The Jacobian K should be a sufficiently smooth function of x.


¯ ˜
• The Jacobian K should be non-singular (invertable) and well-conditioned
¯
in the entire domain of x where iterations take place.
˜
Quadratic convergence means that the ℓ2 norm (Euclidean norm) of the
difference between the solution x∗ and the iterate xi decreases quadratically
in each iteration ˜ ˜

kxi+1 − x∗ k ≤ ckxi − x∗ k2 (3.46)


˜ ˜ ˜ ˜
where c is a constant that depends on the non-linearity of the problem. Thus
the convergence of Newton’s algorithm is quite rapid when K meets the above
¯
conditions. Interpretation of quadratic convergence is given in Table 3.1, where
CHAPTER 3. TOTAL LAGRANGE 32

Table 3.1: Illustration of quadratic convergence

i ∆i is ≤
1 0.81
2 0.66
3 0.43
4 0.19
5 0.034
6 0.0012
7 0.000014

formula (3.46) has been used to estimate the error at each iteration ∆i =
kxi − x∗ k, assuming ∆0 = 0.9 and c = 1. Note, that the number of zeros after
the ˜
˜ floating point doubles at each iteration (after iteration 5 in the example);
this is a typical feature of quadratic convergence.
It should be remembered, however, that quadratic convergence of Newton’s
method is only local, and Newton’s method may converge much slower or not
converge at all unless started close enough to the solution. Thus, the choice
of the initial guess x0 is extremely important.
˜
3.5.4 Convergence criteria
The termination of iterations in Newton’s algorithm is determined by a con-
vergence criterion. There are several convergence criteria that can be applied.
Three types are typically used:
1. Residual convergence test. Since the residual r(xi ) at an approximate
˜ ˜ solution x∗ for which
solution xi measures how far we are from the exact
˜ ˜ be to re-
r(x∗ ) = 0, one of the appropriate convergence criteria would
˜quire
˜ that ˜ the residual, measured in some norm, does not exceed a given
absolute tolerance ǫ. For example, using the ℓ2 (Euclidean) norm
p
krkℓ2 = rT r ≤ ǫ (3.47)
˜ ˜ ˜
Since r usually has physical dimensions, so does necessarily ǫ. For a
general˜ purpose implementation this dependency on physical units is
undesirable and it is more convenient to work with ratios that render ǫ
dimensionless. For example
krkℓ2
˜ ≤ǫ (3.48)
kr0 kℓ2
˜
where r0 = r(x0 ) is the residual at the initial guess (provided kr0 kℓ2 6= 0).
In finite ˜ ˜ implementations it is more customary to check
˜ element ˜ the ratio
of the residual to the external forces (defined for example by (3.27))
krkℓ2
˜ ≤ǫ (3.49)
kf ext kℓ2
˜
CHAPTER 3. TOTAL LAGRANGE 33

2. Convergence test on the unknown. The value of the last correction δx of


the unknown x may enable another appropriate convergence test ˜
˜
p
kδxkℓ2 = δxT δx ≤ ǫ (3.50)
˜ ˜ ˜
or, to be able to use a dimensionless tolerance parameter, the ratio of
the iterative correction to the initial guess x0 or to the last obtained
approximation xi+1 may be used ˜
˜
kδxkℓ2 kδxkℓ2
˜0 ≤ǫ or ˜ ≤ǫ (3.51)
kx kℓ2 kxi+1 kℓ2
˜ ˜
Depending on the specific characteristics of the function r(x) each of
the above two criteria may signal the stop of iterations very ˜far
˜ from the
solution, as is schematically indicated in Figure 3.4 for a one dimensional
example. Therefore, these tests may also be used in an “and” manner,
requiring both of them to be satisfied for iterations to stop. It is also
possible to combine them in one criterion, as described below.

r(x) r(x)

r(xi )
x∗ x∗ xi
x
x
2ǫ r(xi )
xi

(a) (b)

Figure 3.4: Illustration of the cases when convergence criteria


may lead to the stop of iterations far from the so-
lution; (a) the residual convergence criterion, (b) the
convergence criterion on the unknown.

3. Energy convergence test. One may also require that

|rT δx| ≤ ǫ (3.52)


˜ ˜
For a mechanical problem this may be interpreted as a criterion on the
energy flow to the system resulting from the residual, thus the name.
This criterion may also be made dimensionless as in the previous cases.

The error tolerance ǫ determines with which precision the problem is


solved. For example, with the convergence test (3.51) and ǫ = 10−3 the mean
error in the unknown will be in the third significant digit. The convergence
CHAPTER 3. TOTAL LAGRANGE 34

tolerance determines the speed and accuracy of a calculation. If the criterion


is too coarse, the solution may be quite inaccurate. On the other hand, a
criterion which is too tight results in unnecessary computations.
Finally, it is mentioned that instead of the ℓ2 norm other norms may be
used. The ℓ2 norm is appropriate when the mean error over all degrees of
freedom is to be controlled. If one wishes to limit the maximum error at any
degree of freedom, the maximum norm kxk∞ = maxk |xk | may be used.
˜

3.6 Linearization
Now the Newton method will be applied to solve the system of non-linear
equations (3.31), which requires the residual ~r between the internal forces ~f int
˜
and the external forces ~f ext to be zero with unknowns ˜
being the nodal position
vectors ~x ˜
˜
int ext
~r(~x) = ~f (~x) − ~f = ~0 (3.53)
˜˜ ˜ ˜ ˜ ˜
As it has been stated in the previous section, Newton’s method replaces a
non-linear problem by a sequence of linear models. The process of obtaining a
linear model was called linearization. The linearization of the non-linear finite
element model (3.53) will be considered in this section.
Let ~xi denote an approximate solution obtained at iteration i. At the next
iteration,˜ i + 1, the approximate solution may be written as ~xi+1 = ~xi + δ~x,
˜
where δ~x is an (yet unknown) iterative correction. At the iteration i ˜+ 1 the˜
residual˜is
int ext
~r(~xi+1 ) = ~f (~xi+1 ) − ~f = ~0 (3.54)
˜˜ ˜ ˜ ˜ ˜
The internal forces (given by (3.29)) can be further elaborated at the iteration
i + 1 as
Z
~f (~xi+1 ) = ~ 0 N(X
∇ ~ 0 ) · F−1 (~xi+1 ) · τ (~xi+1 ) dV0 (3.55)
int
˜ ˜ ˜
V0h

Recall, that if the nodal position vectors at an iteration, say i, are known
(or estimated), then the approximate field of position vectors (~xh )i is also
immediately known according to (3.17).
The Kirchhoff stress tensor τ (~xi+1 ), the deformation gradient tensor F(~xi+1 )
and the inverse of the deformation gradient tensor F−1 (~xi+1 ) at iteration i+ 1,
in turn, may be written as a sum of their respective values at the previous
iteration and iterative corrections

τ (~xi+1 ) = τ (~xi ) + δτ ≡ τ i + δτ (3.56)


F(~xi+1 ) = F(~xi ) + δF ≡ Fi + δF (3.57)
−1 i+1 −1 i −1 i −1 −1
F (~x )=F (~x ) + δF ≡ (F ) + δF (3.58)
CHAPTER 3. TOTAL LAGRANGE 35

where the notation Zi ≡ Z(~xi ) has been introduced. Note that in (3.58) and
throughout the text δF−1 denotes the iterative correction of F−1 and not the
inverse of the iterative correction δF
−1
δF−1 ≡ δ F−1 6= δF

(3.59)

Substituting (3.56) and (3.58) into (3.55) and neglecting the terms of an order
higher than linear in the iterative correction (e.g. term δF−1 · δτ ) leads to
Z  
~f (~xi+1 ) = ~ ~ i −1 i −1 i i −1
 
int ∇ 0 N (X 0 )· F · τ + δF · τ + F · δτ dV0 (3.60)
˜ ˜ ˜
V0h

Inserting (3.60) into (3.53) and rearranging the terms gives


Z
~ 0 ) · δF−1 · τ i + Fi −1 · δτ dV0 = ~f ext − ~f i
 
~ 0 N(X

∇ int (3.61)
˜ ˜ ˜
V0h

with
Z
~f i =
−1
int
~ 0 N(X
∇ ~ 0 ) · Fi · τ i dV0 (3.62)
˜ ˜
V0h

Now it remains to linearize δτ and δF−1 with respect to δ~x. First, each of
˜
these terms will be rewritten in the format 4 A : δFT . The rationale for doing
so will become clear later on.
The iterative correction δF−1 is elaborated first. Consider the identity
−1   −1 
Fi+1 · Fi+1 = Fi + δF · Fi + δF−1 = I (3.63)
−1
Expanding this relation with account for Fi · Fi = I and neglecting the
term δF · δF−1 leads to
−1
Fi · δF−1 + δF · Fi =0 (3.64)

From here it immediately follows that


−1 −1
δF−1 = − Fi · δF · Fi (3.65)

In the following active use will be made of special fourth-order unit tensors
defined as (applied to an arbitrary second-order tensor Z), see also appendix A,
4
I : Z = Z : 4I = Z (3.66)
4 RT T
I :Z=Z (3.67)
4 S 4
1
I + 4I RT 1 T
= ZS
 
I :Z= 2 :Z= 2 Z+Z (3.68)
CHAPTER 3. TOTAL LAGRANGE 36

Using the fourth-order unit tensor 4 IRT , relation (3.65) and the symmetry
of the Kirchhoff stress tensor, the first term in (3.61) can be rearranged as
follows
−1 −1 i
δF−1 · τ i = − Fi · δF · Fi ·τ
   
−1 T −T
= − Fi · 4 IRT : τ i · Fi · δFT

(3.69)
 
−1 4 RT −T
= − Fi · I · τ i · Fi : δFT
 

To obtain the last result the use has been made of the identity
4
A : B · C = 4A · B : C
 
(3.70)

that is valid for an arbitrary forth-order tensor 4 A and arbitrary second-order


tensors B and C.
Note, that until now we did not use any constitutive equation, so the results
obtained so far are valid for any material. From here on, the constitutive law
(3.7) will be used. From (3.7) it follows for δτ

δτ = κδJI + GδB (3.71)

Using the definition of the Finger tensor (3.9), the iterative correction δB is
expressed in terms of δFT as
T
δB = δ F · FT ≈ δF · Fi + Fi · δFT


=4 IRT : Fi · δFT + 4 I : Fi · δFT (3.72)


 

=24 IS : Fi · δFT = 24 IS · Fi : δFT


 

The derivation of δJ is somewhat more elaborate. Consider


 −1  i 
J i+1 = det(Fi+1 ) = det(Fi + δF) = det I + δF · Fi ·F
 −1   −1  i (3.73)
= det I + δF · Fi det Fi = det I + δF · Fi

J

It can easily be shown that for a small second-order tensor δZ

det(I + δZ) ≈ 1 + tr(δZ) (3.74)

is valid. Using this result in (3.73) gives


  −1  i −1
J i+1 ≈ 1 + tr δF · Fi J = J i + J i δF : Fi
−T (3.75)
=J i + J i Fi : δFT
Therefore
−T
δJ = J i Fi : δFT (3.76)

Substitution of the relations (3.72) and (3.76) into (3.71) gives for δτ
 −T 
δτ = κJ i I Fi + 2G 4 IS · Fi : δFT (3.77)
CHAPTER 3. TOTAL LAGRANGE 37

Combining (3.69) and (3.77), we have for the sum in the left-hand side of
(3.60)
−1 4
δF−1 · τ i + Fi Kigeo + 4 Kimat : δFT

· δτ = (3.78)

where
4
−1 4 RT −T
Kigeo = − Fi · I · τ i · Fi (3.79)
4 i
−1  i −T 
Kmat = Fi · κJ I Fi + 2G 4 IS · Fi (3.80)

The separation into 4 Kimat and 4 Kigeo has been explicitly made to reflect the
influences of the material (physical) non-linearity and the geometrical non-
linearity, respectively. The tensor 4 Kimat is the result of the linearization of
the constitutive relation, while 4 Kigeo originates from the geometrical non-
linearity.
Linearization of δFT with respect to δ~x is straightforward
˜
 
δFT = δ ∇ ~ 0 ~x = ∇
~ 0 δ~x ≈ ∇
~ 0 δ~xh = ∇ ~ 0 ) T δ~x
~ 0 N(X

(3.81)
˜ ˜
Finally, substitution of (3.78) and (3.81) into (3.61) and taking into ac-
count that the column δ~x contains nodal values and thus can be taken out of
the integral, gives ˜

Z !
T i
~ 0 N(X~ 0) · K
4 i 4 i ~ ~ dV0 · δ~x = ~f ext −~f int (3.82)

∇ mat + Kgeo · ∇0 N(X0 )
˜ ˜ ˜ ˜ ˜
V0h

The above equation can be written as a standard linear system


i
Ki · δ~x = ~f ext − ~f int (3.83)
¯ ˜ ˜ ˜
with
Z
i ~ 0 N(X
~ 0) · 4 ~ 0 ) T dV0
~ 0 N(X
Kimat + 4 Kigeo · ∇
 
K = ∇ (3.84)
¯ ˜ ˜
V0h

This tensor matrix Ki in computational solid mechanics is called the tangential


¯
stiffness matrix. Note that this is in fact the system Jacobian matrix (as
has been introduced in the section on Newton’s method, equation (3.41)) for
the non-linear system of equations (3.53) resulting from the finite element
formulation.
It is instructive to compare the obtained linearized system of equations
(3.83) with the linear system of equations (2.30) resulting from the linear fi-
nite element analysis. The mathematical structure of both systems is rather
similar. There are, however, some significant differences. The linearized sys-
tem matrix in non-linear analysis is calculated at every iteration and it depends
on two contributions: material non-linearity and geometrical non-linearity. In
CHAPTER 3. TOTAL LAGRANGE 38

the linear analysis, it is possible to calculate the stiffness matrix “once and
for all” and it depends only on the linear material tensor. The right-hand
side of both systems includes external forces. In the non-linear case also the
i
contribution of the out-of-balance internal forces ~f int is present. Furthermore,
in (3.84) the gradients of the shape functions are ˜ taken with respect to the
reference (undeformed) configuration and the integration is performed over
the reference domain. Note, however, that this is a characteristic feature of
the total Lagrange formulation. In the next chapter another formulation will
be presented.
Solving the set of linear equations (3.83) gives a correction δ~x to the nodal
position vectors, which can be used to update the positions according˜ to

~xi+1 = ~xi + δ~x (3.85)


˜ ˜ ˜
The linear system can then be reassembled and solved for this new approxi-
mation. This process is repeated until the convergence criterion is satisfied (as
discussed in section 3.5.4) and a sufficiently accurate solution has thus been
obtained.
For the purposes of implementation it is more convenient to rewrite the
relation for the iterative update of the solution (3.85) in terms of the displace-
ment vector ~u rather than the position vector ~x. Using the definition of the
displacement˜vector ~xi+1 = X ~˜ 0 + ~ui and substituting these
~ 0 + ~ui+1 and ~xi = X
˜
relations into (3.85) ˜gives another ˜ iterative ˜
˜ update ˜of the solution, which is
equivalent to (3.85)

~ui+1 = ~ui + δ~u (3.86)


˜ ˜ ˜
where δ~x has been exchanged for δ~u, which are logically the same δ~u ≡ δ~x.
Thus the ˜ linearized system of equations
˜ ˜ ˜
(3.83) can be equivalently rewritten
in terms of δ~u
˜
i
K · δ~u = ~f ext − ~f int
i
(3.87)
¯ ˜ ˜ ˜

3.7 Isoparametric transformation


As in the linear finite element method, the integrals in the expressions (3.84)
and (3.62) are usually not determined analytically, but approximated by nu-
merical quadrature. This approximation is done on an element-by-element
basis. The integration at the element level is carried out by rewriting the
integrals in terms of a suitably chosen element reference frame.
When discussing the linear finite elements we have already introduced the
notion of a master (or parent) element (see section 2.3.3). When a geometri-
cally non-linear problem is considered, we are concerned with three domains
that correspond to an element:
1. the master element domain Q;

2. the initial (reference) element domain V0e ;


CHAPTER 3. TOTAL LAGRANGE 39

3. the current element domain V e = V e (t).

The following maps can be constructed to relate the above element domains:

1. initial configuration to current configuration, i.e. the deformation map


~ 0 , t) ≡ ψ(X
~x = ~x(X ~ 0 , t);

~ t);
2. master domain to current configuration: ~x = ~x(ξ,
~0 = X
3. master domain to initial configuration: X ~
~ 0 (ξ).

These domains and maps are schematically illustrated in Figure 3.5 for a
two-dimensional quadrilateral element. Note, that in the linear finite element
analysis no distinction was made between the current and reference configu-
rations, therefore only the map between the master domain and the physical
domain of an element needed to be considered.
x2 (t) current configuration
4

3 ~ t)
~x(ξ,
Ve (-1,1) ξ2 (1,1)
4 3
1
Q
2
x1 (t) ξ1
~ 0 , t)
~x(X
1 2
X02 (-1,-1) (1,-1)
3
master element
4 ~
~ 0 (ξ)
X
V0e
2

1 X01
initial configuration

Figure 3.5: Initial and current configurations of a quadrilateral


element and their relations to the master element.

With these maps at hand, the relation (3.17) can be rewritten (for one
element) as
~ t = NeT X
~ 0 (ξ),
~xhe X ~0 (ξ) ~ xe (t)
~ ~xe (t) = NeT(ξ)~
 
(3.88)
˜ ˜ ˜ ˜
~ 0) is chosen to correspond to X
and since the map ~x = ~x(ξ, ~0 = X ~ we
~ 0 (ξ)
obtain
~ h e (ξ)
X ~ X
~ = NeT(ξ) ~e (3.89)
0 0
˜ ˜
CHAPTER 3. TOTAL LAGRANGE 40

The above maps are also useful to transform integrals and gradient opera-
tors with respect to the reference or current physical configurations to integrals
and gradient operators with respect to the master element domain. By apply-
ing the chain rule the relation between the operators ∇ ~ ξ and ∇
~ 0 can be easily
established
~
~ ξ = ~ei ∂ = ~ei ∂X0j (ξ) ~ej · ~ek ∂ = ∇
∇ ~ ξX ~ ·∇
~ 0 (ξ) ~ ·∇
~ 0 = J 0 (ξ) ~ 0 (3.90)
∂ξi ∂ξi ∂X0k

~ ξ and ∇
Analogously the relation between ∇ ~ is obtained

~
~ ξ = ~ei ∂ = ~ei ∂xj (ξ, t) ~ej · ~ek ∂ = ∇
∇ ~ t) · ∇
~ ξ ~x(ξ, ~ t) · ∇
~ = J x (ξ, ~ (3.91)
∂ξi ∂ξi ∂xk
In these relations the Jacobian tensors have been defined as
~ =∇
J 0 (ξ) ~ ξX ~ =∇
~ 0 (ξ) ~ ξ NeT(ξ) ~e
~ X (3.92)
0
˜ ˜
~ t) = ∇
J x (ξ, ~ t) = ∇
~ ξ ~x(ξ, ~ xe (t)
~ ξ NeT(ξ)~ (3.93)
˜ ˜
Using the inverse of the relations (3.90) and (3.91) the gradient of the shape
functions on an element e with respect to X ~ 0 or ~x can be easily computed

∇ ~ ·∇
~ 0 Ne = J −1 (ξ) ~ ξ Ne (3.94)
0
˜ ˜
~ e = J −1
∇N ~ ~ e
x (ξ, t) · ∇ξ N (3.95)
˜ ˜
Now the contributions of the element e to the integrals in (3.84) and (3.62)
can be transformed to integrals over the element domain Q as
Z
i e ~ 0 NeT dV0
~ 0 Ne · 4 Kimat + 4 Kigeo · ∇
 
K = ∇
¯ ˜ ˜
V0e
Z (3.96)
~ e 4 i 4 i ~ eT

= ∇0 N · Kmat + Kgeo · ∇0 N det(J 0 ) dQ
˜ ˜
Q

 i e Z
~f ~ 0 Ne · Fi −1 · τ i dV0

int = ∇
˜ ˜
V0e
Z (3.97)
~ 0 Ne · Fi −1 · τ i det(J 0 ) dQ

= ∇
˜
Q

where the factor det(J 0 ) appears as a consequence of the coordinate transfor-


mation dV0 = det(J 0 )dQ.
To evaluate the integral (3.27) for ~f ext that incorporates the boundary
˜
tractions ~t(~x) the property of isoparametric elements to lead to the shape
functions of the corresponding low-dimensional element on the boundary (as
discussed in section 2.3.4) is used. Thus, with account for the assumption
CHAPTER 3. TOTAL LAGRANGE 41

(3.28) that the prescribed forces do not depend on the change of geometry, we
have
 e Z Z
~f = N (~x) t(~x) dS = Ne(ξ)
e ~ ~ ~t(~x(ξ~′ )) det J ′ (ξ~′ ) dQ′

ext 0 (3.98)
˜ ˜ ˜
Se Q′

where the prime again denotes local coordinates and shape functions associ-
ated with the lower-dimensional boundary element.

3.8 Numerical integration, assembly, partitioning


and solving
The steps of numerical integration, assembly, partitioning and solving are
performed in exactly the same way as for the linear finite element method,
except that these steps are repeated at each iteration.
The numerical integration of (3.96) – (3.98) yields
q  
i e ~k ) + 4 Ki (ξ~k ) · ∇
~ 0 Ne ~ 0 NeT det J 0 (ξ~k )
X
4 i
 
K = wk ∇ · K mat (ξ geo
¯ ˜ ξ=ξk~ ~ ˜ ξ=ξk ~ ~
k=1
(3.99)
 i e X q  −1
~f ~ e i ~
· τ i (ξ~k ) det J 0 (ξ~k )

int = w ∇
k 0 N ~ ~ · F (ξ k ) (3.100)
˜ k=1
˜ ξ=ξk

 e X q′
~f ext = wk′ Ne(ξ~k′ ) ~t(~x(ξ~k′ )) det J ′0 (ξ~k′ )

(3.101)
˜ k=1
˜

where ξ~k are the Gauss quadrature integration points and wk are the corre-
sponding weights.
After the numerical integration, the element contributions can be assem-
bled to the global system in a standard manner
m m  i e m  e
~f i =
i
A i e
A ~f ~f =
A ~f

K = K int int ext ext (3.102)
¯ e=1
¯ ˜ e=1 ˜ ˜ e=1 ˜

Finally, some words need to be said with respect to application of the


essential boundary conditions in a non-linear formulation. Assuming that in a
given boundary node k either the displacement vector ~uk = ~u∗k or the boundary
traction vector ~tk = ~t∗k is prescribed, the assembled system of equations at
iteration i + 1 can be partitioned as
" # " # " i  # " i  #
Kiuu Kiuf δ~ui+1 ~f ~f
¯i u
¯ i · ˜i+1 = ˜ext u − ˜int i 
u (3.103)
Kfu Kff δ~uf ~f ~
f
ext f int
¯ ¯ ˜ ˜ ˜ f

where the correction at iteration i + 1 has been explicitly marked by the


iteration index: ~ui+1 = ~ui + δ~ui+1 . In (3.103) the subscript “u” refers to
˜ ˜ ˜
CHAPTER 3. TOTAL LAGRANGE 42

those nodes, where the displacements are prescribed; the subscript “f” de-
notes all the remaining (free) nodes. Since the components of ~uu are known
(prescribed), no iterative corrections are required for these degrees ˜ of free-
i+1
dom. Therefore, in practice, the components of δ~uu at the first iteration are
usually set to the prescribed values ~u∗u and to ~0 at ˜ the subsequent iterations
˜ ˜
(

~u , for (i + 1) = 1
δ~ui+1
u = ˜u (3.104)
˜ ~0, for (i + 1) > 1
˜
The second equation to be extracted from (3.103) can then be used to calculate
the unknown corrections δ~ui+1
˜f
i 
Kiff · δ~ui+1 = ~f ext f − ~f int f − Kifu · δ~ui+1

f u (3.105)
¯ ˜ ˜ ˜ ¯ ˜
Combining the solution δ~ui+1 f of this system of equations with the already
i+1
known δ~uu according to ˜
˜
" #
i+1 δ~ui+1
u
δ~u = ˜i+1 (3.106)
˜ δ~uf
˜
gives the entire set of iterative corrections. After this, the convergence crite-
rion can be checked. If the convergence criterion is not satisfied, a new system
of linear equations can be assembled to obtain the next iterative correction.
If necessary, the unknown boundary tractions ~tu (reaction forces) can be re-
~
 ˜
trieved from f ext u , which, in turn, can be calculated from the top part of
˜
the partitioned system (3.103) as

~f i = Kiuu · δ~ui+1 i
ui+1
i
+ ~f int
 
ext u u + Kuf · δ~ f u
(3.107)
˜ ¯ ˜ ¯ ˜ ˜

3.9 Incremental approach


In order to start the Newton process as discussed in the previous sections, an
initial estimate ~u0 of the solution is needed. A natural choice is to assume that
˜
the current configuration coincides with the undeformed, reference, configura-
tion and thus ~u0 = ~0. However, if the response is strongly non-linear and if
˜ ˜
the current configuration differs significantly from the reference configuration,
the iteration process may fail to converge because of a relatively poor initial
estimate.
To avoid this problem, the loading may be applied in a number of incre-
ments. This means that only a fraction of the entire load (or of the prescribed
displacements) is first applied. The discrete non-linear system (3.31) is solved
for this reduced load (or displacement) by the iterative process as discussed
above. The loading is then increased by a fraction of the total load and the
discrete non-linear problem is again solved iteratively for this new increment
etc. At each new increment n + 1 the solution obtained at the previous incre-
ment is used as a starting value for the iteration process. If the increments
CHAPTER 3. TOTAL LAGRANGE 43

are selected sufficiently small, the initial guess is sufficiently close to the ac-
tual solution for the iterative process to converge. Such approach is called the
incremental-iterative procedure.
Denoting at increment n + 1 the increment of the prescribed displacements
by ∆~un+1 u and the increment of prescribed forces by ∆~f ext,n+1 f , the par-


˜ system (3.103) at iteration i + 1 of increment n + ˜1 is written as
titioned

~f i ~f i
"  # " i+1  # " # "  #
Kin+1 uu Kin+1 uf
 
δ~un+1 u int,n+1 u
¯ ¯  · ˜  = ˜ext,n+1 u  − ˜
Kin+1 fu Kin+1 ff δ~ui+1 ~f ~ ~f i
 
n+1 f ext,n f + ∆f ext,n+1 f int,n+1 f
¯ ¯ ˜ ˜ ˜ ˜
(3.108)

with
(
∆~u∗n+1 u , for (i + 1) = 1

δ~ui+1

n+1 u = ˜ (3.109)
˜ ~0, for (i + 1) > 1
˜
and with the starting value for the unknown displacements taken equal to the
solution obtained at the previous increment

~u0n+1 f = ~un f
 
(3.110)
˜ ˜
An incremental approach is always necessary for inelastic constitutive ma-
terial models, since the response of these models depends on the loading path.
This loading path must therefore be approximated with sufficient accuracy,
i.e. with a sufficient number of increments in order to correctly describe the
deformation history process. Path dependent material models are, however,
beyond the scope of this course.
Chapter 4

Updated Lagrange procedure


for geometrically non-linear
problems

4.1 Transformation to the last known configuration


The geometrically non-linear finite element formulation discussed in the pre-
vious chapter was based on the transformation of the governing equations to
the initial, undeformed configuration. Instead of this, one may choose to use
the last configuration where the solution is known as a reference configuration.
This is the main idea behind the updated Lagrange procedure.
As in the previous chapter, we consider the same geometrically non-linear
elastic problem given by the equilibrium equation in the absence of body forces
(3.3), the constitutive relation (3.7) and boundary conditions (3.4) and (3.5).
As has already been derived in the previous chapter, the discretization of
the weak form of this problem leads to the set of non-linear equations (3.25)
expressing the balans between external and internal forces. The linearization
of this non-linear system of equations around the previous solution estimate
at iteration i is given by relation (3.61)
Z
~ 0 ) · δF−1 · τ i + Fi −1 · δτ dV0 = ~f ext − ~f i
 
~ 0 N(X

∇ int (4.1)
˜ ˜ ˜
V0h

with
Z
~f i =
−1
int
~ 0 N(X
∇ ~ 0 ) · Fi · τ i dV0 (4.2)
˜ ˜
V0h

and ~f ext given by (3.27). As it was the case in the previous chapter, here
it is ˜assumed for simplicity that the forces prescribed on the boundary do
not depend on the deformation of the body and thus the boundary integral
expressing the external forces can be left unchanged.

44
CHAPTER 4. UPDATED LAGRANGE 45

In the above relations, the initial undeformed geometry of the body B0


has been chosen as the reference configuration to compute integrals and gra-
dients. On the other hand, the solution estimate of the unknown position
vector field ~xi obtained at the previous iteration i can be considered as the
last known configuration, see Figure 4.1. Note, that this is a fictitious com-
putational configuration, which in most cases doesn’t have physical meaning
since it corresponds to a non-converged solution and thus does not describe
an equilibrium state of the system. From a numerical point of view, however,
this is the last known configuration that can be chosen as the reference.

B
V
unknown
dx
t
i+1
F B
i+1
V
dx i+1 current
V0 iteration i+1
dF unknown
F i t
B0 dX0
Vi
i i previous
t=0 B dx
iteration i
t known

F i + dF
Figure 4.1: Initial configuration, unknown configuration and the
approximate configurations at the iterations i and
i + 1.

Using the deformation gradient tensor Fi = (∇ ~ 0 ~xi )T , the infinitesimal


volume in the undeformed configuration dV0 can be related to the infinitesimal
volume dV i in the last iterative configuration (B i ) as (compare to equation
(3.12))
1 1
dV0 = i
dV i = i dV i (4.3)
det(F ) J

The gradient operator ∇ ~ 0 with respect to the undeformed configuration is


related to the gradient operator ∇~ i with respect to the last iterative configu-
ration B i according to (compare to (3.14))
~ 0 = (Fi )T · ∇
∇ ~i=∇
~ i · Fi (4.4)

Then the gradient of the shape functions with respect to the undeformed
configuration can be rewritten as

~ 0 N(X
∇ ~ 0 ) = (Fi )T · ∇
~ i N(X
~ 0) = ∇
~ i N(X
~ 0 ) · Fi (4.5)
˜ ˜ ˜
CHAPTER 4. UPDATED LAGRANGE 46

Applying the above relations to transform the integrals in (4.1) and (4.2) gives

~ 0 ) · Fi · δF−1 · τ i + δτ 1 dV i = ~f ext − ~f i
Z
~ i N(X

∇ int (4.6)
˜ Ji ˜ ˜
Vih

with

~ 0 ) · τ i 1 dV i
Z
~f i = ~ i N(X
∇ (4.7)
int
˜ ˜ Ji
Vih

Next, the second-order tensor Lδ is introduced1 as


−1
Lδ = δF · Fi (4.8)

Using relation (4.4) it is easy to show that


−1  i −T T   T
−T ~
Lδ =δF · Fi = F · δFT = Fi · ∇0 δ~x
 T (4.9)
= ∇~ i δ~x

Then, with account for (3.65), (4.8) and the symmetry of the Kirchhoff stress
tensor, we have for the first term in (4.6)
−1 −1
Fi · δF−1 · τ i = − Fi · Fi · δF · Fi · τi
(4.10)
= − Lδ · τ i = −4 IRT · τ i : LTδ


Until now no particular constitutive relation was used in the derivations of


this chapter, so the results obtained so far are valid for any material. To elab-
orate the second term in (4.6), a constitutive law is needed. The linearization
of the constitutive equation (3.7) with respect to δFT has already been done
in the previous chapter (see section 3.6, equations (3.71)–(3.77)) and reads
 −T 
δτ = κJ i I Fi + 2G 4 IS · Fi : δFT (4.11)

Noting that δFT = (Fi )T · LTδ , which follows directly from (4.8), δτ can be
related to LTδ as
    
−T T
δτ = κJ i I Fi + 2G 4 IS · Fi : Fi · LTδ
!
 
i i −T 4 S i i T
: LTδ
 
= κJ I F + 2G I · F · F (4.12)
 
= κJ i II + 2G 4 IS · Bi : LTδ
1
Note the similarity between the tensor Lδ and the velocity gradient tensor L = Ḟ · F−1 used
in continuum mechanics.
CHAPTER 4. UPDATED LAGRANGE 47

Next, results (4.10) and (4.12) are substituted into (4.6)

T 1
Z
~ i N(X
~ 0 ) · 4 Ki + 4 Ki i ~ ~i

∇ geo mat : Lδ i dV = f ext − f int (4.13)
˜ J ˜ ˜
Vih

i
where ~f int is given by (4.7) and
˜
4 i
Kgeo = −4 IRT · τ i (4.14)
4
Kimat = κJ i II + 2G 4 IS · Bi (4.15)

These fourth-order tensors have the same origins as the respective tensors
for the total Lagrange framework: 4 Kigeo originates from geometrical non-
linearities and 4 Kimat is the result of the non-linearity of the constitutive law.
The expressions for these tensors, however, differ from those for the total La-
grange framework, even though we consider exactly the same problem. (Com-
pare (4.14)–(4.15) with (3.79)–(3.80)).
Using relation (4.9), LTδ can be easily written in terms of the column of
the iterative corrections to the position vectors of the nodes δ~x as
˜
T
T ~ i ~ i h ~ i
Lδ = ∇ δ~x ≈ ∇ δ~x = ∇ N(X0 ) δ~x~ (4.16)
˜ ˜
Substitution of (4.16) into (4.13) with account for the fact that the column
δ~x contains constants and thus can be taken out of the integral, finally gives
˜ linearized model of the non-linear system (3.25) in the updated Lagrange
the
framework
!
T 1
Z
i
~ i ~ 4 i 4 i ~ i ~ dV ·δ~x = ~f ext −~f int (4.17)
i

∇ N(X0 )· Kgeo + Kmat · ∇ N(X0 ) i
˜ ˜ J ˜ ˜ ˜
Vih

or introducing the tensor matrix Ki (the stiffness matrix) according to


¯
~ 0 ) T 1 dV i
Z
Ki = ~ i N(X
~ 0 ) · 4 Kigeo + 4 Kimat · ∇
~ i N(X
 
∇ (4.18)
¯ ˜ ˜ Ji
Vih

equation (4.17) can be written as a standard linear system of equations


i
Ki · δ~x = ~f ext − ~f int (4.19)
¯ ˜ ˜ ˜
Recalling that the iterative correction for the position vectors δ~x may be
˜ (4.19)
replaced by the iterative correction for the displacements δ~u, relation
can be equivalently written as ˜

i
Ki · δ~u = ~f ext − ~f int (4.20)
¯ ˜ ˜ ˜
CHAPTER 4. UPDATED LAGRANGE 48

4.2 Isoparametric transformation, numerical inte-


gration, assembly, partitioning and solving
The calculation of integrals in (4.18) and (4.7) is, as usually, carried out on an
element-by-element basis with respect to an element domain Q. In section 3.7
we already saw, that a map exists between the physical domain of an element
in the current configuration and the master element domain. Using this map,
the gradients of the element shape functions with respect to the configuration
at the last iteration, that appear in (4.18) and (4.7), may be conveniently
calculated as (utilizing relations (3.95) and (3.93))
 −1
∇ ~ t)
~ i Ne = J i (ξ, ~ ξ Ne
·∇ (4.21)
˜ ˜
with the Jacobian tensor
~ t) = ∇
J i (ξ, ~ t) i = ∇
~ ξ ~x(ξ, ~ ~xe (t) i
~ ξ NeT(ξ)
 
(4.22)
˜ ˜
Now the contributions of the element e to the integrals in (4.18) and (4.7)
are transformed to integrals over the element domain Q as

~ i e T 1 det(J i ) dQ
Z
i e ~ i Ne · 4 Ki + 4 Ki
  
K = ∇ mat geo · ∇ N (4.23)
¯ ˜ ˜ Ji
Q

~ i Ne · τ i 1 det(J i ) dQ
 i e Z
~f = ∇ (4.24)
int
˜ ˜ Ji
Q

where the relation dV i = det(J i )dQ has been used.


The numerical integration of the above expressions, assembly, partitioning
and solving are performed in a standard manner, as has been outlined in
the previous chapter. Application of the essential boundary conditions again
has to be done carefully. The prescribed incrementally updated values of
the displacements ~u∗u should be given only at the first iteration, while at the
˜ they should be set to ~0.
subsequent iterations
˜

4.3 Total Lagrange versus updated Lagrange


formulations
Let us now compare the total and updated Lagrange formulations. The lin-
earized model (4.17) obtained here for the updated Lagrange framework looks
similar to the one obtained in the previous chapter for the total Lagrange
approach, however the expressions for the system matrix Ki and the inter-
i ¯
nal forces ~f int differ. Compare (4.18) with (3.84) and (4.7) with (3.62). The
˜
most important difference is clearly that in the total Lagrange approach the
gradient operators and the integrals are calculated with respect to the initial
undeformed configuration B0 , while in the updated Lagrange framework the
CHAPTER 4. UPDATED LAGRANGE 49

last approximate configuration obtained at the last iteration B i is used: the


integrals are taken over the volume V i and gradient operators ∇ ~ i are calcu-
lated. Also other quantities that enter the expressions are calculated with
respect to the initial undeformed or the the last known configuration, respec-
tively. For these reasons the total Lagrange formulation is sometimes called
the reference configuration formulation and the updated Lagrange – the cur-
rent configuration formulation. It is, however, important to realize that these
two frameworks are equivalent. This fact can easily be proved by transforming
the expressions of one of the formulations into those of another using trans-
formations of tensors and mappings of configurations. And the solution of a
problem obtained within each of the frameworks will be, of course, the same!
However the question still remains which of the formulations should one
use in practice? The choice of the particular framework is largely a matter of
preference, mostly dictated by convenience. Each problem can be solved us-
ing any of the two formulations, although usually one of the formulations will
lead to simpler derivations and implementation than another. This is mostly
related to the type of a constitutive law used. In continuum mechanics sev-
eral kinematical tensors and stress tensors are introduced [1]. These tensors
are defined in the undeformed configuration, the current configuration or in
both, as schematically shown in Figure 4.2. For example, the right Cauchy-
Green deformation tensor C, the Green-Lagrange strain tensor E, the right
stretch tensor U and the second Piola-Kirchhoff stress tensor S are all defined
in the reference, undeformed configuration (for the definitions of these ten-
sors and the tensors discussed below see [1]); the left Cauchy-Green (Finger)
deformation tensor B, the Euler-Almansi strain tensor e, the left stretch ten-
sor V and the Cauchy stress tensor σ and the Kirchhoff stress τ are defined
in the current configuration; the deformation gradient tensor F and the first
Piola-Kirchhoff stress tensor P are defined in both configurations, since they
map vectors of one configuration on to another. In the previous paragraph
we have discussed that the total Lagrange formulation in fact can be viewed
as the reference configuration formulation, while the updated Lagrange – as
the current configuration formulation. Based on this reasoning it is natural
to expect that the constitutive laws that are formulated in terms of the ten-
sors defined in the reference configuration (e.g. S = F(E)) will be easier to
implement in the total Lagrange framework, while the constitutive laws de-
fined in the current configuration (e.g σ = F(B)) are easier to incorporate
into the updated Lagrange formulation. The constitutive relations defined in
terms of the first Piola-Kirchhoff stress tensor P lead to the simplest form of
the total Lagrange expressions, although these constitutive laws are not often
found in practice since the first Piola-Kirchhoff stress tensor lacks a physical
meaning and is moreover not objective (thus the constitutive equation has to
be carefully formulated to maintain objectivity).
Furthermore, constitutive relations may be formulated in a rate form. The
objective constitutive laws of a rate-type typically use an objective rate of the

Cauchy stress tensor, e.g. the Lie (Truesdell) derivative σ, the Cotter-Rivlin
CHAPTER 4. UPDATED LAGRANGE 50

Eulerian
Kinematical tensors F F P Stress tensors deformed
(current)
ufacturing
Ma n J = det(F) configuration
C Kinematical tensors s t
C U E Stress tensors

B V e
B Kinematical tensors
S t
Stress tensors
t + dt

Lagrangian
undeformed Kinematical (rate) tensors L D W
(reference)
D o
configuration Objective stress derivatives
sÑ s s
Figure 4.2: Kinematical and stress tensors defined in the refer-
ence and current configurations.

∆ ◦
derivative σ or the Jaumann derivative σ and kinematical rate quantities, e.g.
the velocity gradient tensor L, the deformation rate tensor D and the spin
tensor Ω, see Figure 4.2. Definitions of these derivatives and tensors may be
found, for example, in reference [1]. Calculation of these rate-type quantities
at time t + dt clearly requires knowledge of the relevant tensors at the last
iteration at time t. Thus, implementation of these models becomes easier if
the updated Lagrange formulation is adopted.
Finally, it should be remarked that, in the literature the term updated
Lagrange is sometimes also used for the finite element formulation in which
the discretized system of equations is written with respect to the configuration
Bn obtained at the previous increment n (rather than at the previous iteration
i, as discussed above). In this case the expressions for the stiffness matrix and
the right-hand side are, in fact, the same as those derived for the total Lagrange
formulation, except that the undeformed volume V0 is replaced by the volume
at obtained at the previous increment Vn and the gradient operator ∇ ~ 0 is
~
replaced by ∇n .
Chapter 5

Forces dependent on
deformation

5.1 Introduction
In the previous chapters it was assumed that the applied forces (tractions)
~t∗ were themselves not dependent on the deformation. In some cases this is
not true. For instance, pressure loads on a deforming structure are in this
category. As a schematic example, consider a rectangular-shaped structure
loaded on one side by a pressure load, as shown in Figure 5.1. Assume that
under the applied pressure or due to other loads the structure was deformed to
a trapezoidal shape. Then the normal to the surface has changed its direction
and so did the pressure vector, which always acts in the direction of the normal.
One can also imagine that the pressure vector follows the deformation of the
body, always staying normal to the boundary. This is why these deformation
dependent loads are sometimes also called follower forces.
~n0 ~t∗0 = −p~n0
~t∗ = −p~n
~n

S0
ψ S

V0 V
t=0 t
Figure 5.1: A deformation dependent pressure load.

In general, a non-linear finite element formulation leads to the system of


non-linear equations expressing the balance between the internal and external
nodal forces
~f (~x) = ~f (~x) (5.1)
int ext
˜ ˜ ˜ ˜
If the externally applied forces vary with displacement, then the column of

51
CHAPTER 5. FORCES DEPENDENT ON DEFORMATION 52

the external forces


Z
~f (~x) = N(X ~ 0 )~t(~x) dS (5.2)
ext
˜ ˜ ˜
Sh

depends on the yet unknown nodal position vectors ~x and the integral in the
˜
above expression has to be calculated over the yet unknown current surface
S h of the body. Therefore, the integral in (5.2) has to be first transformed to
some known configuration. After that, the external forces ~f ext (~x) have to be
˜ chapter
linearized with respect to the iterative correction δ~x. In this ˜ we will
˜
derive linearized relations for the deformation dependent external loading in
both the total Lagrange and updated Lagrange frameworks. We will consider
the case of pressure loading, which remains normal to the surface of the body
throughout the deformation history. Thus the prescribed tractions on the St
part of the boundary are given by
~t∗ (~x) = −p~n(~x) (5.3)
where ~n(~x) is the unit outward normal to the surface of the body and p is a
given pressure, which itself may depend on the position ~x and thus also on the
deformation. Here, however, we will take it a constant.

5.2 Relation between surface normals in the refer-


ence and current configurations
We begin by deriving the relation between the oriented surface element in
the reference (undeformed) configuration ~n0 dS0 and the same oriented surface
element ~ndS in the current (deformed) configuration. Consider an elementary
material volume. In its reference configuration this volume is denoted by dV0
and it is spanned by three vectors ~a0 , ~b0 and ~c0 . In the current configuration,
this volume is deformed to dV , which is now spanned by the vectors ~a, ~b and
~c, see Figure 5.2.

dV0 dV

Figure 5.2: Characterization of a deformed volume and surface.

Let F be the deformation gradient tensor that describes the deformation


of the elementary volume dV0 to dV . Then the following relations hold
~a = F · ~a0 , ~b = F · ~b0 , ~c = F · ~c0 (5.4)
CHAPTER 5. FORCES DEPENDENT ON DEFORMATION 53

Let the undeformed area dS0 under consideration be given by the area
spanned by vectors ~a0 and ~b0 . The normal on this area is given by ~n0 . Likewise,
the area dS will be spanned by the vectors ~a and ~b. The normal in this
deformed area dS is denoted by ~n. Using the property of the cross (vector)
product we can write

~n0 dS0 = ~a0 × ~b0 , ~ndS = ~a × ~b (5.5)

Here it should be mentioned, that comparing (5.5) with (5.4) it immediately


follows that

~n 6= F · ~n0 (5.6)

Using the property of the triple product and (5.5), the infinitesimal volumes
are obtained as
dV0 = (~a0 × ~b0 ) · ~c0 = (~n0 dS0 ) · ~c0
dV = (~a × ~b) · ~c = (~ndS) · ~c (5.7)
The change of volume from the reference state to the current state is given by
J = det(F). This then yields
dV = JdV0
(~ndS) · ~c = J(~n0 dS0 ) · ~c0
(~ndS) · F · ~c0 = J(~n0 dS0 ) · ~c0 (5.8)
which must hold for all possible dV0 and thus all ~c0 . Hence, we finally obtain

~ndS = J~n0 · F−1 dS0 (5.9)

5.3 Total Lagrange formulation


As a point of departure, we take the expression for the discretized external
forces (5.2), in which we substitute the pressure load (5.3)
Z Z
~f (~x) = N(X ~ ~ x) dS = − N(X
0 ) t(~
~ 0 ) p ~n(~x) dS (5.10)
ext
˜ ˜ ˜ ˜
Sh Sh

Now the external forces are formulated in the current, deformed state: ~n(~x) is
the normal to the current surface of the body and the integral has to be taken
over the current boundary of the body. This deformed state is, however,
not yet known, therefore we need to transform this expression to a known
configuration, e.g. the reference undeformed configuration. So, we need to
make a trick similar to the one in section 3.3, except that now we have to
work with surface integrals instead of volume integrals. Using formula (5.9)
we immediately obtain
Z
~f (~x) = − N p J(~x) ~n0 · F−1 (~x) dS0 (5.11)
ext
˜ ˜ ˜
S0h
CHAPTER 5. FORCES DEPENDENT ON DEFORMATION 54

This is clearly a non-linear function of the solution ~x. Thus, to facilitate the
solution of the non-linear system by Newton’s method ˜ it has to be linearized
with respect to the iterative correction δ~x. Writing relation (5.11) for the
approximate solution ~xi+1 = ~xi + δ~x at the˜ iteration i + 1 gives
Z˜ ˜ ˜
~f (~xi+1 ) = − N p J(~xi+1 ) ~n0 · F−1 (~xi+1 ) dS0
ext
˜ ˜ ˜
S0h
Z (5.12)
i
= ~f ext − N p δJ ~n0 · (Fi )−1 + J i ~n0 · δF−1 dS0

˜ ˜
S0h

where
Z
~f i = − N p J i ~n0 · (Fi )−1 dS0 (5.13)
ext
˜ ˜
S0h

The iterative corrections δF−1 and δJ are given by (3.65) and (3.76), respec-
tively
−1 −1  −1 4 RT −T 
δF−1 = − Fi · δF · Fi = − Fi · I · Fi : δFT
(5.14)
i −T
δJ = J i F : δFT

(5.15)

Substitution of these relations into (5.12), after rearrangement of terms, yields


Z
~f (~xi+1 ) = ~f i − N ~n0 · p J i (Fi )−1 · II − 4 IRT · Fi −T : δFT dS0
   
ext ext
˜ ˜ ˜ ˜
S0h
Z
i
= ~f ext + N ~n0 · 4 Kiload : δFT dS0
˜ ˜
S0h
(5.16)

where
4
−T
Kiload = −p J i (Fi )−1 · II − 4 IRT · Fi

(5.17)

Substitution of relation (3.81)

δFT ≈ ∇ ~ 0 N(X ~ 0 ) T δ~x



(5.18)
˜ ˜
into (5.16), with account for the fact that the column δ~x contains constants
and thus can be taken out of the integral, leads to ˜
Z !
i
~f (~xi+1 ) = ~f + ~ 0 ) ~n0 · 4 Ki · ∇ ~ 0 ) T dS0 · δ~x
~ 0 N(X

ext ext N(X load
˜ ˜ ˜ ˜ ˜ ˜
Sh (5.19)
0
i
= ~f ext + KiL · δ~x
˜ ¯ ˜
CHAPTER 5. FORCES DEPENDENT ON DEFORMATION 55

Here the matrix of the second-order tensors KiL has been introduced as
¯
Z
KiL = N(X ~ 0 ) ~n0 · 4Ki · ∇ ~ 0 ) T dS0
~ 0 N(X

load (5.20)
¯ ˜ ˜
S0h

This matrix is the result of the linearization of the external forces with respect
to the iterative correction δ~x and thus it is sometimes called loading stiffness
or loading correction matrix.˜
Combining these results with the results for the linearization of internal
forces as obtained in section 3.6, finally gives the linearized system of equations
with account for the external loading dependent on deformation
i i
Ki − KiL · δ~x = ~f ext − ~f int

(5.21)
¯ ¯ ˜ ˜ ˜
i i
with Ki given by (3.84), KiL by (5.20), ~f ext by (5.13) and ~f int by (3.62).
¯ ¯
The above system of equations is as ˜usually assembled ˜on the element-by-
element basis with integrals evaluated using numerical quadrature. Terms Ki
i ¯
and ~f int involve the volume integrals and are dealt with in the same manner
˜
as discussed in the previous chapters. Here we will outline computation of
i
the surface integrals that appear in KiL and ~f ext . Using arguments similar to
¯ ˜ the following transformation
those used in section 5.2, it is easy to show that
holds between the oriented surface ~n0 dS0 of an element in the undeformed
configuration and the same oriented surface of a master element ~ne dQ′ , with
~ne the outward unit normal to the corresponding face of the master element

~n0 dS0 = det(J 0 ) ~ne · J −T


0 dQ

(5.22)

where J 0 is the Jacobian tensor as has been introduced in section 3.7 (equa-
tion (3.92))
e
~ =∇
J 0 (ξ) ~ ξX ~ =∇
~ 0 (ξ) ~ X
~ ξ NeT(ξ) ~0 (5.23)
˜ ˜
Using (5.22), relations (5.20) and (5.13) can be transformed to integrals over
the surface of the master element domain (so, in fact, over a lower dimensional
boundary element, which is denoted by the prime)
Z
i e
−T 4 i
KL = Ne det(J 0 ) ~ne · J 0 ~ 0 NeT dQ′

· Kload · ∇ (5.24)
¯ ˜ ˜
Q ′
 i e Z
−T −1
~f = − Ne p J i det(J 0 ) ~ne · J 0 · Fi dQ′ (5.25)
ext
˜ ˜
Q′

~ 0 Ne calculated according to (3.94) as


with ∇
˜
~ −1 · ∇
~ 0 Ne = J 0 (ξ) ~ ξ Ne

∇ (5.26)
˜ ˜
CHAPTER 5. FORCES DEPENDENT ON DEFORMATION 56

Finally, applying the appropriate numerical quadrature for the lower dimen-
sional boundary element yields
q ′
e −T 4 i
wk′ Ne (ξ~k′ ) det J 0 (ξ~k′ ) ~ne · J 0 (ξ~k′ ) · Kload (ξ~k′ ) · ∇
~ 0 NeT
X
KiL

=
¯ ˜ ˜ ξ= ~ ξ~k
k=1
(5.27)
 i e q′
−T −1
~f wk′ Ne (ξ~k′ ) p(ξ~k′ ) J i (ξ~k′ ) det J 0 (ξ~k′ ) ~ne · J 0 (ξ~k′ ) · Fi (ξ~k′ )
X 
ext =−
˜ k=1
˜
(5.28)

5.4 Updated Lagrange formulation


To derive the linearized relations for the deformation dependent load in the
updated Lagrange framework, the linearized relations (5.16) and (5.13) have
to be transformed to the last known configuration, i.e. the configuration ob-
tained at the previous iteration i. From (5.9) the relation between the oriented
surface element in the undeformed configuration ~n0 dS0 and in the last iterative
configuration ~ni dS i follows as
1 i i i
~n0 dS0 = ~n · F dS (5.29)
Ji
Applying the above relation to transform integrals in (5.16) and (5.13) gives
Z
~f i+1 = ~f i − N ~ni · p II − 4 IRT · Fi −T : δFT dS i
   
ext ext (5.30)
˜ ˜ ˜
Sih

with
Z
~f i = − N p ~ni dS i (5.31)
ext
˜ ˜
Sih

Making use of the tensor Lδ defined as (see (4.8))


−1
Lδ = δF · Fi (5.32)

relation (5.30) can further be simplified as


Z
~f i+1 = ~f i − N ~ni · p II − 4 IRT : LT dS i

ext ext δ (5.33)
˜ ˜ ˜
Sih

or introducing notation
4
Kiload = −p II − 4 IRT

(5.34)
Z
~f i+1 = ~f i + N ~ni · 4 Ki : LT dS i (5.35)
ext ext load δ
˜ ˜ ˜
Sih
CHAPTER 5. FORCES DEPENDENT ON DEFORMATION 57

Finally, after substitution of (4.16)

LTδ ≈ ∇ ~ i N T δ~x

(5.36)
˜ ˜
into (5.35) and taking the column of the iterative corrections to nodal position
vectors δ~x out of the integral, we arrive at the linearized expression for the
deformation ˜ dependent load in the updated Lagrange framework

Z !
i+1 i T
~f ~ ~ i N dS i · δ~x
N ~ni · 4 Kiload · ∇

ext = f ext +
˜ ˜ ˜ ˜ ˜
Sh (5.37)
i
i
=~f ext
+ KiL
· δ~x
˜ ¯ ˜
where the load correction matrix for the updated Lagrange formulation is
defined as
Z
~ i N T dS i
KiL = N ~ni · 4 Kiload · ∇

(5.38)
¯ ˜ ˜
Sih

i
and ~f ext is given by (5.31).
˜
This result is to be combined with other linearization results for the up-
dated Lagrange formulation, as described in section 4.1, to yield the linear
system of equations
i i
Ki − KiL · δ~x = ~f ext − ~f int

(5.39)
¯ ¯ ˜ ˜ ˜
Element transformation and numerical integration of integrals (5.31) and
(5.38) is done along the same lines as has been described above for the total
Lagrange formulation, except that now the Jacobian J i (given by (4.22))

~ t) = ∇
J i (ξ, ~ t) i = ∇
~ ξ ~x(ξ, ~ ~xe (t) i
~ ξ NeT(ξ)
 
(5.40)
˜ ˜
is used for the transformation between the current physical domain of an
element and the master element. This gives
Z
i e ~ i Ne T dQ′
−T 4 i
KL = Ne det(J i ) ~ne · J i
 
· Kload · ∇ (5.41)
¯ ˜ ˜
Q ′
 i e Z
−T
~f = − Ne p det(J i ) ~ne · J i dQ′ (5.42)
ext
˜ ˜
Q′

with
 −1
∇ ~ t)
~ i Ne = J i (ξ, ~ ξ Ne
·∇ (5.43)
˜ ˜
Finally, it is instructive to notice that the linearized expressions for the
deformation dependent load turned out to be much simpler in the updated
CHAPTER 5. FORCES DEPENDENT ON DEFORMATION 58

Lagrange framework, than in the total Lagrange, compare (5.34) with (5.17)
and (5.31) with (5.13). This is not surprising, since the loading has been
defined with respect to the current configuration and the updated Lagrange
framework is, in fact, the current configuration formulation.
Note also that here for simplicity we have assumed the magnitude of the
pressure to be a constant. If the pressure magnitude is also a function of the
coordinate, i.e. p(~x), then this term also has to be linearized with respect
to δ~x. Additionally, the variation of the pressure over the surface may be
˜
described using an interpolation
T
p(ξ~′ ) = N′ (ξ~′ )p (5.44)
˜ ˜
where p are the values of the known pressure at the nodes.
˜
Chapter 6

Incompressible and slightly


compressible materials

6.1 Introduction
Although problems involving incompressible behaviour are relatively rare in
linear stress analysis, many materials behave in a nearly incompressible man-
ner in the non-linear regime. The typical example is rubber, that is an intrinsi-
cally incompressible material, combined with the non-linear elastic behaviour
up to very large strain levels. Another example is the von Mises elasto-plastic
material behaviour, representative for metals. Although metals are elastically
compressible, the plastic deformation is typically volume preserving, and thus
in the plastic regime these materials behave as nearly incompressible. Other
examples include undrained soils, incompressible fluids and in general mate-
rials that are much stiffer in volumetric deformation compared to the shear
deformation.
The standard displacement finite element formulations, as derived in the
previous chapters, lead to poor element performance for slightly compress-
ible and incompressible materials resulting in erroneous solutions. Perhaps
the largest trouble is in this case caused by the so-called volumetric lock-
ing. When volumetric locking occurs, the displacements are underpredicted
by large factors, in some cases by several orders of magnitude.
Despite the fact that in the engineering practice the incompressible and
nearly incompressible material behaviour is mostly dealt with in non-linear
regime, in this chapter the problems of incompressibility and related numerical
solution techniques will be considered on the example of small strain linear
elasticity. This significantly simplifies the derivations, while still perfectly
illustrates the main points. The same techniques are also applicable for use
within the non-linear finite element formulations discussed in the previous
chapters.

59
CHAPTER 6. INCOMPRESSIBILITY 60

6.2 Incompressible linear elasticity formulation


As already mentioned, the incompressibility is observed when a material is
infinitely stiff in volumetric deformation compared to deformation in shear. In
isotropic linear elasticity this means that in this case the ratio of the material
bulk modulus κ to the shear modulus G approaches infinity

κ 2(1 + ν)
= →∞ (6.1)
G 3(1 − 2ν)

where ν is the Poisson’s ratio and the expressions of the bulk and shear moduli
in terms of the elasticity modulus and the Poisson’s ratio have been used (see,
for example, [10]). From (6.1) it is immediately clear that the condition of
incompressibility in isotropic linear elasticity may be formulated in terms of
the Poisson’s ratio ν; as ν → 12 , the ratio between the bulk and shear moduli
approaches infinity. The limiting value of ν = 21 thus represents incompress-
ibility.
This limit creates problems in the equations of (compressible) elasticity.
Recall that the constitutive equation in the isotropic case may be written as
(see equation (2.7))

σ = κtr(ε)I + 2Gεd (6.2)

Since κ approaches infinity in the incompressible limit, the compressible elas-


ticity law (6.2) will result in infinitely large stresses in this case. Therefore an
alternative formulation of the theory necessary.
In the incompressible isotropic linear elastic formulation the constitutive
equation is written as

σ = −pI + 2Gεd (6.3)

where p = p(~x) is the hydrostatic pressure. The pressure must be determined


as part of the solution of the boundary value problem and thus represents an
additional unknown. Therefore an additional equation is necessary to have a
well-defined system of equations. This additional equation is the kinematic
condition of incompressibility, which in the geometrically linear case can be
written as
~ u) = ∇
tr(ε) = tr(∇~ ~ · ~u = 0 (6.4)

6.3 Mixed pressure-displacement formulation for in-


compressible and compressible materials
6.3.1 Strong form
From the practical point of view, it is desirable to have a formulation of
isotropic elasticity that is valid for compressible, slightly compressible and
incompressible behaviour. To this end we introduce a pair of constitutive
CHAPTER 6. INCOMPRESSIBILITY 61

equations where, in addition the the displacement ~u, also the pressure p is
viewed as independent unknown
~u
σ = −pI + 2Gεd = −pI + 4 Cd : ∇~ (6.5)
~ · ~u + p = 0
∇ (6.6)
κ
where the material tensor 4 Cd is given by
4
Cd = 2G(4 IS − 13 II) (6.7)

Relation (6.6) is a modified version of the incompressibility condition (6.4). If


the material is incompressible, ν = 21 , κ = ∞, and thus (6.6) becomes exactly
the incompressibility condition (6.4) and p is the hydrostatic pressure as in
the previous section. If ν < 12 , then by substituting (6.6) into (6.5), p may
be eliminated to result in the standard constitutive equation of compressible
elasticity (6.2). Thus, relations (6.5) and (6.6) describe both compressible and
incompressible cases.
Note, that in the formulation given here p may directly be interpreted as
the hydrostatic pressure defined as σh = −tr(σ)/3. Relation (6.5) gives

σh = − 13 tr(σ) = −κtr(ε) (6.8)

whereas from (6.6) it follows that


~ · ~u = −κtr(ε)
p = −κ∇ (6.9)

and thus it follows that p = σh .


To complete the strong form, the relations (6.5) and (6.6) have to be
complemented by the equilibrium equation (2.1) (where we will omit the body
forces for simplicity) and the boundary conditions (2.2) and (2.3). Thus, the
strong form can be summarized as follows: for given ~u∗ and ~t∗ , find ~u(~x) and
p(~x) such that
~ · σ = ~0
∇ in V (6.10)
∇~ · ~u + p = 0 in V (6.11)
κ
~u = ~u∗ on Su (6.12)
~t = ~n · σ = ~t∗ on St (6.13)

with σ given by
~u
σ = −pI + 4 Cd : ∇~ (6.14)

6.3.2 Weak form


To obtain the finite element formulation of the problem, the first step is to
recast the above strong form into a weak form by means of a weighted residuals
approach. In chapter 2 this was done for the compressible linear elasticity by
CHAPTER 6. INCOMPRESSIBILITY 62

taking the scalar product of the equilibrium equation with a vector valued test
~ x) followed by the integration over the domain V . For the
(weight) function φ(~
formulation considered here we also need to introduce a term that implies
satisfaction of the incompressibility condition (6.11). For this, in addition to
~ x), we introduce another set of (pressure)
the (displacement) test function φ(~
test functions ψ(~x). Note that these are the scalar valued functions, since
equation (6.11) is a scalar equation. Multiplication of the equations (6.10) and
(6.11) with their respective test functions and integration over the domain V
gives

~ · ~u + p dV = 0,
Z   Z  
φ~· ∇~ · σ dV − ψ ∇ ~ x), ψ(~x)
∀φ(~ (6.15)
κ
V V

where the minus sign before the last term has been chosen purely for the
convenience of the further derivations. The weighted residuals form (6.15)
~ x) and ψ(~x), to ensure its
must be satisfied for all admissible test functions φ(~
equivalence with the strong form.
Application of the chain rule and the divergence theorem to the first term
of (6.15) in the same way as has been done in chapter 2 (see equations (2.8)
through (2.14)) leads to the weak form of the problem (6.10)-(6.13)
p
Z Z  Z
T
~ ~
∇φ : σ dV − ψ ∇ · ~u + ~ dV = φ ~ · ~t dS (6.16)
κ
V V S

Substitution of the constitutive law (6.14) into (6.16) results in

~ · ~u + p dV = φ
Z Z Z   Z
~φ~ T : 4 Cd : ∇~
~ u dV − ~ p dV −
~ ·φ ~ · ~t dS
 
∇ ∇ ψ ∇
κ
V V V S
(6.17)
~ T : I = tr(∇
~ φ)
where the identity (∇ ~ =∇
~ φ) ~ has been used.
~ ·φ

6.3.3 Discretization of the weak form


Next, the exact domain V is replaced by the discretized domain V h , as usually.
Now, however, we have two unknown fields, the displacement ~u(~x) and the
pressure p(~x), which are approximated as
nu
X
~uh (~x) = Nui (~x) ~ui = NTu(~
x) ~u (6.18)
i=1
˜ ˜
np
X
ph (~x) = Npi (~x) pi = NTp (~
x) p (6.19)
i=1
˜ ˜

where nu and np are the total number of nodes used to approximate the dis-
placement and the pressure fields, respectively; ~ui and pi are the nodal values
of the discretized fields and Nui (~x) and Npi (~x) are the shape functions used to
CHAPTER 6. INCOMPRESSIBILITY 63

interpolate the displacement and the pressure fields, respectively. In the above
relations the column/matrix notation as previously introduced has again been
used. Note, that each of the unknown fields is approximated separately, using
“own” set of shape functions on its “own” set of nodes. The particular choice
of these discretizations will be discussed in the next section. Here we will only
assume that the respective shape functions satisfy all necessary continuity
requirements.
Again following the Galerkin approach the test functions are approximated
in the same form as the corresponding unknown fields
φ~ h (~x) = NT(~x) φ~ (6.20)
u
h
˜ T ˜
ψ (~x) = Np (~x) ψ (6.21)
˜ ˜
From the approximations (6.18) and (6.20), the gradient and the divergence
of the displacement field and the displacement test functions that enter (6.17)
are simply obtained as
~ u (~x) T ~u,
~ uh (~x) = ∇N ~φ~ h (~x) = ∇N ~ u (~x) T φ ~
 
∇~ ∇ (6.22)
˜ T
˜ ˜ ˜
T
~ · ~uh (~x) = ∇N~ u (~x) · ~u, ~ ·φ~ h (~x) = ∇N ~ u (~x) · φ ~
 
∇ ∇ (6.23)
˜ ˜ ˜ ˜
Substitution of approximations (6.18)-(6.23) into the weak form (6.17)
gives
Z Z
T
~ T · ∇N
T
~ ~ 4 d ~
φ ∇Nu : C : ∇Nu ~u dV − φ ~ u NT p p dV
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
V h V h
  (6.24)
1
Z Z
T
 T
~ u · ~u + Np p dV = NT
T ~ · ~t dS
− ψ Np (∇N u φ
˜ ˜ ˜ ˜ κ˜ ˜ ˜ ˜
V h h S

The columns ~u, p, φ ~ and ψ contain constants (nodal values) and thus can be
˜
taken out of the ˜integrals,
˜ ˜leading to
Z  Z 
~T · ~ u · 4 Cd · ∇N ~T ·
~ u T dV · ~u − φ ~ u NT

φ ∇N ∇N p dV p
˜ ˜ ˜ ˜ ˜ ˜ ˜ ˜
Vh Vh
Z  Z 
1
Z
−ψ T ~ u T dV · ~u − ψ T Np NT ~T Nu~t dS

Np (∇N p dV p = φ ·
˜ ˜ ˜ ˜ ˜ κ˜ ˜ ˜ ˜ ˜
h
V h V h S
(6.25)
Taking into account that the above equation must be satisfied for all admissible
~ x) and ψ(~x), and therefore for all φ
test functions φ(~ ~ and ψ , results in the
following system of equations ˜ ˜
Z  Z  Z
~ u T dV · ~u −
~ u · 4 Cd · ∇N ~ u NT dV p = Nu~t dS

∇N ∇N p
˜ ˜ ˜ ˜ ˜ ˜ ˜
Vh Vh Sh
(6.26)
Z  Z 
~ u
T 1
− Np (∇N dV · ~u − N NT dV p = 0 (6.27)
˜ ˜ ˜ κ ˜p ˜p ˜ ˜
V h V h
CHAPTER 6. INCOMPRESSIBILITY 64

Introducing notation
Z
~ u T dV
~ u · 4 Cd · ∇N

Kuu = ∇N (6.28)
¯ ˜ ˜
Vh
Z Z T
T T
~ up = −
K ~ u NT
∇N p dV = −
~ u
Np (∇N dV ~ pu
=K (6.29)
¯ ˜ ˜ ˜ ˜ ¯
Vh Vh
1
Z
Kpp = − N NT dV (6.30)
¯ κ ˜p ˜p
h
Z V
~f = N ~t dS (6.31)
u
˜ ˜
Sh

allows to rewrite the above system of equations in a compact form as


Kuu · ~u + K ~ up p = ~f (6.32)
¯ ˜ ¯ ˜ ˜
~T
K · ~u + Kpp p = 0 (6.33)
¯ up ˜ ¯ ˜ ˜
After evaluation of the integrals (6.28)-(6.31) and application of the bound-
ary conditions, the above system of equations can be readily solved for the
unknown nodal displacements ~u and nodal pressures p. There are several pos-
sible ways of solving this system˜ of equations. It may ˜be solved simultaneously
for both unknowns, the displacement and the pressure. Note, however, that in
practice this might result in necessity of solving rather large system of equa-
tions. Alternatively, if the compressible case is considered, and therefore the
matrix Kpp 6= 0, then by solving (6.33) the pressures can be expressed as
¯ ¯
−1 ~ T
p = −Kpp Kup · ~u (6.34)
˜ ¯ ¯ ˜
and substituting this result into (6.32) gives

(Kuu − K ~ T u = ~f
~ up K−1 K
pp up ) · ~ (6.35)
¯ ¯ ¯ ¯ ˜ ˜
Equation (6.35) has the usual format (2.30) with the stiffness matrix

K = Kuu − K ~T
~ up K−1 K (6.36)
pp up
¯ ¯ ¯ ¯ ¯
After solving (6.35) for displacements ~u, the pressure p can be determined
from (6.34). ˜ ˜
If the incompressible case is dealt with, Kpp = 0, the preceding elimination
¯ ¯
of pressure can not be performed. The alternative is to solve (6.32) for ~u and
employ (6.33) to obtain the equation for pressure ˜
 T
~ T · K−1 · ~f

~ · K−1 · K
K ~ up p = K (6.37)
up uu
¯ ¯ ¯ ˜ ¯ up ¯ uu ˜
After p is obtained by solving (6.37), the displacements ~u can be calculated
˜
using (6.32). ˜
Finally, it is remarked that the mixed pressure-displacement formulation
presented here was first proposed by Herrmann [11]. Therefore sometimes it
is also called Herrmann formulation.
CHAPTER 6. INCOMPRESSIBILITY 65

6.3.4 Element selection


In order to compute the integrals (6.28)-(6.31) providing the system matrices,
a specific discretization for the unknown fields and approximation (shape)
functions Nu and Np have to be chosen. In other words, an element has to
be chosen.˜It turns˜ out, that the element selection for mixed formulations in
general, and the pressure-displacement formulation in particular, is a rather
delicate task, as will be discussed in this section.
Before formulating possible elements for the mixed formulation, let us be-
gin by considering the continuity requirements imposed on the unknown fields
~u and p by the weak form (6.17). The first-order derivatives (gradient and
~
divergence) of the displacement field ~u and the corresponding test function φ
appear in the integrand. Therefore, just as before, the discretized displace-
ment field ~uh has to be at least continuous and piecewise differentiable. On
the other hand, no derivatives of the pressure field p and the test function
ψ are present in (6.17). Therefore, the approximation to the pressure field
ph may be discontinuous across element boundaries. Hence, a wider range
of interpolations is permissible for pressure than for displacements. One may
think of a large number of possible combinations for the interpolations of the
two unknown fields. Some possibilities are illustrated in Figure 6.1. It must
however be emphasized that arbitrary combinations of interpolations may lead
to poor numerical performance of the elements and ultimately to erroneous
results, as will be discussed in the following.

(a) (b)
continuous bilinear pressure discontinuous constant pressure
continuous bilinear displacement continuous bilinear displacement

(c) (d)
continuous bilinear pressure discontinuous bilinear pressure
continuous biquadratic displacement continuous biquadratic displacement

displacement nodes pressure nodes

Figure 6.1: Examples of possible mixed pressure-displacement el-


ements in two dimensions. Note: not all of these
elements work well in practice.

The advantage of using elements with discontinuous pressure approxima-


CHAPTER 6. INCOMPRESSIBILITY 66

tion is that in the compressible case the pressure degrees of freedom may be
eliminated at the element level following the procedure similar to (6.34)–(6.36)
but applied to the element matrices instead of the total assembled matrices:

pe = −(Kepp )−1 (K ~ eup )T · ~ue (6.38)


˜ ¯ ¯ ˜ 
~ e (Ke )−1 (K ~ e )T · ~ue = ~f e

Keuu − K (6.39)
¯ ¯ up ¯ pp ¯ up ˜ ˜
e e −1 ~ e T
e e ~
K = Kuu − Kup (Kpp ) (Kup ) (6.40)
¯ ¯
m
¯ ¯ m ¯
e
K=
¯ A K¯ ,
e=1
e ~f =
˜ A~˜f
e=1
(6.41)

Since in this case only the operations on small matrices are involved in the
elimination process and no extra memory needs be allocated for the extra
degrees of freedom and the related matrices at the global level, this may lead to
significant reduction of the computational efforts especially for large problems.
A classical example [6] of one of the fundamental difficulties related to
the element selection for the incompressible case is illustrated in Figure 6.2.
Consider a two-dimensional mesh consisting of linear displacement – contant
pressure triangles. The boundary conditions are such that the left-hand side
and bottom edges of the mesh are fixed, i.e the displacements are identically
zero, see Figure 6.2. From the weak form and with application of the Galerkin
approach, the discretized incompressibility condition follows in the form
Z   m Z  
~ · ~uh dV = ~ · ~uh dV e = 0
X
ψh ∇ ψh ∇ (6.42)
e=1V e
Vh

where m is the number of elements in the mesh. Since the elements considered
use the constant approximation for the pressure field within the element, then
by the Galerkin approach ψ h is also an arbitrary constant within each trian-
gle. Therefore, it follows from (6.42), that the incompressibility condition is
satisfied if within each element holds
Z
∇~ · ~uh dV e = 0, ∀ e = 1, ..., m (6.43)
Ve

Thus the volume (area in 2D) of each triangle must necessarily remain con-
stant.
Now let us examine the kinematical constraints enforced on the displace-
ments of the element nodes by the condition (6.43). Consider element 1. With
account for the boundary conditions, it follows that the constant volume re-
quirement permits the displacement of node A in the horizontal direction
only. Now consider element 2. For this element the constat volume condition
prohibits the horizontal displacement of node A and only allows the vertical
displacement at the node. Taken together, the displacement of node A must
be identically zero, ~uA = ~0. Repeating the above arguments for the elements
3 and 4 leads to the conclusion that the displacement at node B also has to
CHAPTER 6. INCOMPRESSIBILITY 67

Load

A ~uA = ~0

2 A

A B
2 4
1 3 possible displacements 1
of element 2

possible displacements
of element 1

Figure 6.2: Mesh for which incompressibility results in locking


(zero displacements).

be zero, ~uB = ~0. Repeating this for all elements of the mesh leads to the
conclusion that all nodes must have zero displacement. This result holds no
matter how many elements are present in the mesh in each direction. This
phenomenon is often referred to as mesh locking. Note, that in the nearly
incompressible case, the same phenomenon occurs, only this time ~u ≈ ~0. Thus
introducing slight incompressibility does not make the problem go˜ away. ˜ The
mesh locking is just one of the difficulties related to the problems of incom-
pressibility and slight compressibility.
To assess if a mixed element will perform well for the formulation consid-
ered here, a rather complex mathematical inf-sup criterion has been formulated
(see for example [7]). In the literature this criterion is better known under the
names Babuška-Brezzi condition, or BB condition, or LBB condition. To es-
tablish whether or not this condition is satisfied for a given element is not
a trivial task and involves some advanced mathematics. Therefore, it is de-
sirable to have a simple procedure to evaluate if a given element will lock or
not. For this purpose, a very simple method of constraint count provides a
necessary condition for stability of mixed elements [3]. It should be empha-
sized that this is a heuristic approach and provides only a necessary, but not
sufficient, condition for mixed element stability. Despite this, the constraint
count method has proven to be rather efficient tool in indicating the elements
that will definitely fail and the elements that may perform satisfactory, but
require further testing.
Consider a standard mesh, which is illustrated in Figure 6.3 for a two
dimensional case. Let nu denote the total number of scalar displacement
equations after boundary conditions have been imposed (i.e. the total number
of scalar components after expanding the vector ~u in (6.33) to the component
˜
form) and let np denote the total number of incompressibility constraints,
which for the problem considered here equals to the number of pressure equa-
tions (i.e the length of the column p in (6.33)). The constrain ratio r is defined
˜
CHAPTER 6. INCOMPRESSIBILITY 68

as
nu
r= (6.44)
np

The conjecture is that, for a typical mesh in Figure 6.3 with the number of
elements per side approaching infinity, the constrain ratio r should mimic the
behaviour of the continuum problem formulation, for which the number of
equilibrium equations equals 2 in two dimensions and 3 in three dimensions
and the incompressibility constraint is described by one scalar equation. Thus,
for a continuum problem in 2D the constraint ratio is 2 and in 3D – 3. To
perform the constraint count test for a standard mesh consisting of given
element type, r is computed according to (6.44) and compared to the values
of r corresponding to the continuum case. In 2D, a value of r less than 2
would indicate tendency to lock. If r ≤ 1, there are more constraints, than
there are displacement degrees of freedom available and thus severe locking
would be expected. A value of r much greater than 2 indicates that there
are not enough incompressibility constraints present, so the incompressibility
constraint might be poorly approximated. These ideas are summarized below

r>2 too few incompressibility constraints


r=2 optimal
r<2 too many incompressibility constraints, tendency to lock
r≤1 number of constraints exceeds number of unknowns, locking

nel

nel
fixed displacement
~u = ~0

Figure 6.3: A standard two dimensional finite element mesh with


the number of elements along each direction nel . Dis-
placements of all nodes at two sides are fixed.

Values of constraint ratio r for some two-dimensional elements with discon-


tinuous pressure are indicated in Figure 6.4. Consider the linear displacement
– constant pressure triangle, Figure 6.4a, which has also been used previously
to illustrate the locking phenomenon. For this element nu = 2 and np = 2,
thus r = nu /np = 1, which indicates severe locking, as has been already ob-
served before. Note that the same number for r is obtained if a bigger mesh
would be used, e.g. the one shown in Figure 6.3. Now consider a triangle
element with quadratic displacement – constant pressure, Figure 6.4b. For
this element nu = 8, np = 2 and r = 4, which indicates no locking, but not
CHAPTER 6. INCOMPRESSIBILITY 69

optimal approximation of the incompressibility constraint. If instead of con-


stant pressure, linear approximation for pressure is used (Figure 6.4c), then
nu = 8, np = 6 and r = 4/3 < 2, thus indicating too many constraints and
tendency to lock. To “fix” this, one more displacement node can be added
in the center of triangle (Figure 6.4d), resulting in nu = 12, np = 6, r = 2
and thus the optimal constraint ratio is attained. Note, that the elements
in Figure 6.4b and Figure 6.4d also satisfy the mathematical BB condition.
Some examples of quadrilateral elements with discontinuous pressure are also
shown in Figure 6.4e,f,g, where the constraint ratio is indicated. It is left for
the reader to verify these results.

r=1 r=4 r = 4/3

(a) (b) (c)

r=2 r=2 r = 3/2

(d) (e) (f)

r=2
displacement nodes

pressure nodes

(g)

Figure 6.4: Discontinuous pressure-field elements. The con-


straint ratio r is indicated. Elements (b) and (d)
satisfy the Babuška-Brezzi condition, the rest violate
this condition.

Next, the constraint count test is illustrated for a few elements with con-
tinuous pressure, Figure 6.5. The elements in Figure 6.5a,b both have the
constraint ration r = 2, which should be optimal, however these elements do
not satisfy the mathematical BB condition and indeed have been observed to
yield oscillatory solution for pressure field. This appears to be a general draw-
back affecting the elements possessing the identical displacement and pressure
interpolations (bilinear – bilinear or biquadratic – biquadratic etc.) This de-
CHAPTER 6. INCOMPRESSIBILITY 70

ficiencies can be corrected by lowering the pressure interpolation as has been


done for the elements shown in Figure 6.5c,d. Both of these elements have
rather high constraint ratio r = 8, indicating poor approximation of the in-
compressibility constant. Despite this, these elements are often used in the
incompressible analysis. They also can be shown to satisfy the BB condition.

r=2 r=2 r=8

(a) (b) (c)


r=8
displacement nodes

pressure nodes

(d)
Figure 6.5: Continuous pressure-field elements. The constraint
ratio r is indicated. Elements (c) and (d) satisfy
the Babuška-Brezzi condition, elements (a) and (b)
violate this condition.

6.4 Selective reduced integration for slightly com-


pressible materials
Consider now a slightly compressible material. As has been already discussed
above, dealing with slight compressibility instead of full incompressibility does
not change the situation with respect to the element performance, e.g. patho-
logical mesh locking. Therefore, for slightly compressible materials, the ele-
ments that perform well in the incompressible analysis, as discussed in the
previous section, have to be used.
Alternatively, for slightly compressible materials another approach to im-
prove the element behaviour has been prosed, namely the selective reduced
integration. This approach is based on slight modification of the usual dis-
placement formulation, discussed in Chapter 2. For this, the fourth-order
elasticity tensor 4 C is explicitly split into the volumetric 4 Cv and deviatoric
4 Cd parts as

4
C = 4 Cv + 4 Cd (6.45)
CHAPTER 6. INCOMPRESSIBILITY 71

with
4 4
Cv = κII and Cd = 2G(4 IS − 31 II) (6.46)

Note, that for nearly incompressible materials the bulk modulus κ is very large
compared to the shear modulus G, κ/G ≫ 1, and in the incompressible limit
approaches infinity, see (6.1). Substitution of (6.45) into the expression (2.27)
for the linear elastic stiffness matrix K gives
Z Z ¯
~ · 4 Cv · ∇N~ T dV + ∇N ~ T dV = Kv + Kd (6.47)
~ · 4 Cd · ∇N
 
K= ∇N
¯ ˜ ˜ ˜ ˜ ¯ ¯
Vh Vh

The idea of the selective reduced integration method is to “underintegrate”


the volumetric part of the stiffness matrix Kv . This means that the order
¯
of the numerical quadrature employed to evaluate Kv is reduced compared
¯
to the one “normally” used for a given element. The deviatoric part Kd is
¯
evaluated using the standard quadrature. For example, for a four-node bilinear
quadrilateral, for which four integration points are normally used, the selective
reduced integration means that the deviatoric part of the stiffness is evaluated
using four integration points, while for the volumetric part only one integration
point located at the center of the element is used, thus (cf. (2.42))

(1) ~ e ~
(K ) = w ∇N ~ · C ~x(0) · ∇N ~ det J (~0)
v e 4 v ~ eT
 
(6.48)
¯ ˜ ξ=~0 ˜ ξ=~0
4
(4) ~ e
~ ~ · 4 C d ~x(ξ~k ) · ∇N
~ eT det J(ξ~k )
X
d e
 
(K ) = wk ∇N (6.49)
¯ ˜ ξ=ξk ~
˜ ξ=ξk ~
k=1
Ke = (Kv )e + (Kd )e (6.50)
¯ ¯ ¯
(4)
where w(1) and wk are the weight factors for one and four point 2D Gauss
quadratures, respectively. For a nine-node biquadratic quadrilateral this pro-
cedure would mean that the deviatoric part is integrated using nine integration
points, while for the volumetric part only 4 integration points are used, etc.
Although the selective reduced integration approach has initially been de-
veloped based on purely heuristic arguments, later the equivalence of many
mixed elements with discontinuous pressure (with the pressure nodes located
at the locations of the Gauss integration points) and selectively integrated el-
ements has been established in the so-called Malkus-Hughes equivalence the-
orem [6, 3]. For example, it has been shown that the element stiffness matrix
of the selectively integrated four-node bilinear quadrilateral discussed above
is identical to the element stiffness of the mixed element with bilinear dis-
placement and discontinuous contant pressure (Figure 6.4e), after elimination
of the pressure degrees of freedom. Since the element stiffnesses are iden-
tical, consequently the computed displacements are equivalent as well, and
hence the two formulations are fully equivalent. The selectively integrated
nine-node biquadratic quadrilateral is equivalent to the mixed element with
biquadratic displacement and discontinuous bilinear pressure with the pres-
sure nodes located in the four-point Gauss quadrature points, see Figure 6.4g.
CHAPTER 6. INCOMPRESSIBILITY 72

Therefore, it can be concluded that the selective reduced integration proce-


dures are very simple ways of attaining the performance of mixed formulation
without necessity to deal with additional complications. For this reason the
selective reduced integration approach is especially often used for elasto-plastic
material behaviour with von Mises plasticity law, where due to the purely de-
viatoric character of the plastic deformation, the overall material behaviour
in the plastic regime is nearly incompressible, even though the material is
elastically compressible. Note, however, that the selective reduced integration
is only suitable for compressible and slightly incompressible materials, in the
fully incompressible case the volumetric part of the stiffness Kv becomes in-
¯
finite and thus no sensible solution can be obtained. In this case the mixed
formulation has to be used.
In addition to the selective reduced integration, also the full reduced inte-
gration or simply reduced integration is sometimes used. In this case both terms
in the stiffness matrix are “underintegrated”, using less integration points than
is standard for given element. Obvious advantage of such approach is the econ-
omy of the computer time, since less computations per element have to be
made. This becomes especially advantageous in large scale non-linear analy-
sis. For three dimensions, cost reductions on the order of 8 have been achieved
through underintegration compared to full integration. The disadvantage of
the full reduced integration is that the rank of the element stiffness may be
reduced, resulting in singularity of the global system matrix, which may lead
to erroneous solution (with the so-called singular modes or hourglass mode)
or no solution at all. To avoid these problems the reduced integrated elements
have to be stabilized to exclude the singular deformation modes. Such stabi-
lization methods, also known in the literature as hourglass control, are beyond
the scope of this course. The interested reader is referred for example to [5].
References and Bibliography

[1] M. G. D. Geers, F. P. T. Baaijens and P. J. G. Schreurs. Continuum


Mechanics for Advanced Manufacturing Technologies. Lecture notes for
the course 4C600. Eindhoven University of Technology, Eindhoven, 2004.

[2] R. Peerlings. Finite Element Method. Lecture notes for the course 4A700.
Eindhoven University of Technology, Eindhoven, 2004.

[3] O. C. Zienkiewicz and R. L. Taylor. The Finite Element Method. Vol-


ume 1: The Basics. Butterworth Heinemann, 5th edition, 2000.

[4] O. C. Zienkiewicz and R. L. Taylor. The Finite Element Method. Vol-


ume 2: Solid Mechanics. Butterworth Heinemann, 5th edition, 2000.

[5] T. Belytschko, W. K. Liu and B. Moran. Nonlinear Finite Elements for


Continua and Structures. Wiley, Chichester, 2000.

[6] T. J. R. Hughes. The Finite Element Method : Linear Static and Dynamic
Finite Element Analysis. Prentice-Hall, 1987.

[7] K.-J. Bathe. Finite Element Procedures. Prentice-Hall, 1996.

[8] M. A. Crisfield. Non-linear Finite Element Analysis of Solids and Struc-


tures. Volume 1. Wiley, 1991.

[9] M. A. Crisfield. Non-linear Finite Element Analysis of Solids and Struc-


tures. Volume 2: Advanced Topics. Wiley, 1997.

[10] M. G. D. Geers Applied Elasticicty in Engineering. Lecture notes for the


course 4A450. Eindhoven University of Technology, Eindhoven, 2004.

[11] L. R. Herrmann Elasticity equations for nearly incompressible materials


by a variational theorem AIAA Journal, vol. 3, pp. 1896–1900, 1965

73
Appendix A

Second- and fourth-order unit


tensors

Second-order unit tensor


The second-order unit tensor I is defined as the tensor that maps a vector on
itself

I · ~a = ~a · I = ~a (A.1)

for an arbitrary vector ~a. With respect to a Cartesian basis {~e1 , ~e2 , ~e3 } the
second-order unit tensor can be written as

I = ~e1~e1 + ~e2~e2 + ~e3~e3 (A.2)

or in index notation

I =⇒ Iij = δij (A.3)

with δij the Kronecker delta. Below are some useful relations involving the
second-order unit tensor and an arbitrary second-order tensor A

I·A =A·I=A (A.4)


I : A = A : I = tr(A) (A.5)
−1 −1
A·A =A ·A =I (A.6)

Special fourth-order unit tensors


The fourth-order unit tensor 4 I maps a second-order tensor A on itself
4
I : A = A : 4I = A (A.7)

This tensor is given in index notation by


4
I =⇒ Iijkl = δil δjk (A.8)

74
APPENDIX A. SECOND- AND FOURTH-ORDER UNIT TENSORS 75

and with respect to Cartesian basis


4
I =~e1~e1~e1~e1 + ~e1~e2~e2~e1 + ~e1~e3~e3~e1
~e2~e1~e1~e2 + ~e2~e2~e2~e2 + ~e2~e3~e3~e2 (A.9)
~e3~e1~e1~e3 + ~e3~e2~e2~e3 + ~e3~e3~e3~e3

The right-transpose of the tensor 4 I maps a second-order unit tensor A on


the transpose of A, denoted AT :
4 RT
I : A = A : 4 IRT = AT (A.10)

In index notation 4 IRT is given by


4 RT
I =⇒ Iijkl = δik δjl (A.11)

Using the index notation it is easy to prove (A.10)

4 RT
I : A = δik δjl~ei~ej ~ek~el : Amn~em~en = δik δjl Amn δkn δlm~ei~ej = Aji~ei~ej = AT
A : 4 IRT = Amn~em~en : δik δjl~ei~ej ~ek ~el = Amn δmj δni δik δjl~ek~el = Alk ~el~ek = AT
(A.12)

The symmetric (also called symmetrization) fourth-order tensor 4 IS is de-


fined by
4 S
I = 21 4 I + 4 IRT

(A.13)

This tensor maps a second-order unit tensor A on its symmetric part, denoted
AS :
4 S
I : A = 21 4 I + 4 IRT : A = 21 A + AT = AS
 
(A.14)
A : 4 IS = A : 12 4 I + 4 IRT = 21 A + AT = AS
 
(A.15)

Another useful fourth-order tensor II is given by the dyadic (outer) product


of two second-order unit tensors I. Using (A.5) it follows

II : A = A : II = I tr(A) (A.16)

Вам также может понравиться