0521830540 cambridge university press non linear modeling and analysis of solids and structures aug 2009

This page intentionally left blank


Non-linear Modeling and
Analysis of Solids and
Structures
Steen Krenk


CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore,
São Paulo, Delhi, Dubai, Tokyo
Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521830546
© Cambridge University Press 2009
This publication is in copyright. Subject to statutory exception and to the
provision of relevant collective licensing agreements, no reproduction of any part
may take place without the written permission of Cambridge University Press.
First published in print format 2009

ISBN-13

978-0-511-60413-3

eBook (EBL)

ISBN-13

978-0-521-83054-6

Hardback

Cambridge University Press has no responsibility for the persistence or accuracy
of urls for external or third-party internet websites referred to in this publication,
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.


To Jette



Contents

Preface
1
1.1

page ix

1.3
1.4

Introduction
A simple non-linear problem
1.1.1 Equilibrium

1.1.2 Virtual work and potential energy
Simple non-linear solution methods
1.2.1 Explicit incremental method
1.2.2 Newton–Raphson method
1.2.3 Modified Newton–Raphson method
Summary and outlook
Exercises

1
2
3
6
7
8
9
13
14
15

2
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8

Non-linear bar elements

Deformation and strain
Equilibrium and virtual work
Tangent stiffness matrix
Use of shape functions
Assembly of global stiffness and forces
Total or updated Lagrangian formulation
Summing up the principles
Exercises

17
18
20
24
26
31
36
39
43

3
3.1
3.2
3.3
3.4
3.5

Finite rotations
The rotation tensor
Rotation of a vector into a specified direction
The increment of the rotation variation

Parameter representation of an incremental rotation
Quaternion parameter representation
3.5.1 Representation of the rotation tensor

47
49
53
55
60
63
64

1.2

v


vi

3.6
3.7
3.8
4
4.1
4.2
4.3

4.4

4.5

4.6
5
5.1

5.2

5.3
5.4
6
6.1

6.2

Contents

3.5.2 Addition of two rotations
3.5.3 Incremental rotation from quaternion parameters
3.5.4 Mean and difference of two rotations
Alternative representation of the rotation tensor
Summary of rotations and their virtual work
Exercises

65
67
68
69
72
73

Finite rotation beam theory

Equilibrium equations
Virtual work, strain and curvature
Increment of the virtual work equation
4.3.1 Constitutive stiffness
4.3.2 Geometric stiffness
4.3.3 The load increments
Finite element implementation
4.4.1 Element stiffness matrix
4.4.2 Loads and internal forces
4.4.3 Shear locking
Summary of ‘elastica’ beam theory
Exercises

76
77
78
81
82
83
85
86
87
89
91
98
99

Co-rotating beam elements
Co-rotating beams in two dimensions
5.1.1 Co-rotation form of the tangent stiffness

5.1.2 Element deformation stiffness
5.1.3 Total tangent stiffness
5.1.4 Finite element implementation
Co-rotating beams in three dimensions
5.2.1 Co-rotation form of the tangent stiffness
5.2.2 Element deformation stiffness
5.2.3 Total tangent stiffness
5.2.4 Finite element implementation
Summary and extensions
Exercises

100
101
104
107
110
112
117
120
127
130
133
139
141

Deformation and equilibrium of solids
Deformation and strain
6.1.1 Non-linear strain
6.1.2 Decomposition into deformation and rigid body motion
Virtual work and stresses

6.2.1 Piola–Kirchhoff stress
6.2.2 Cauchy and Kirchhoff stresses

145
146
148
151
154
155
158


Contents

6.3

6.4

6.5
6.6
7
7.1

7.2

7.3

7.4

7.5

7.6
7.7
7.8
8
8.1

vii

6.2.3 Stress rates
Total Lagrangian formulation
6.3.1 Equilibrium and residual forces
6.3.2 Tangent stiffness
6.3.3 Finite element implementation
Updated Lagrangian formulation
6.4.1 Transformation from total to updated format
6.4.2 Virtual work in the current configuration
6.4.3 Finite element implementation
Summary of non-linear motion of solids
Exercises

160
165
166
167
170
174
174
176
180

185
186

Elasto-plastic solids
Elastic solids
7.1.1 Stress invariants
7.1.2 Strain invariants and small strain elasticity
7.1.3 Isotropic elasticity at finite strain
General plasticity theory
7.2.1 Reversible deformation
7.2.2 Maximum plastic dissipation rate
7.2.3 Evolution equations
7.2.4 Isotropic and kinematic hardening
Von Mises plasticity models
7.3.1 Yield surface and flow potential
7.3.2 Explicit integration
7.3.3 Radial return algorithm
General aspects of plasticity models
7.4.1 Combined isotropic and kinematic hardening
7.4.2 Internal variables and non-associated flow
7.4.3 General computational procedure
Models for granular materials
7.5.1 Flow potential and yield surface
7.5.2 Elasticity and hardening
Finite strain plasticity
Summary
Exercises

189
190

192
198
200
203
204
207
212
216
218
219
222
225
229
230
234
237
241
242
247
249
252
253

Numerical solution techniques
Iterative solution of equilibrium equations
8.1.1 Non-linear iteration strategies

256
257
259

viii

8.2
8.3

8.4
8.5
8.6
9
9.1

9.2
9.3

9.4

9.5
9.6

Contents

8.1.2 Direction and step-size control
Orthogonal residual method
Arc-length methods
8.3.1 General constraint formulation
8.3.2 Hyperplane constraints
8.3.3 Hypersphere constraint
Quasi-Newton methods

Summary
Exercises

260
263
270
272
274
278
283
287
288

Dynamic effects and time integration
Newmark algorithm for linear systems
9.1.1 Energy balance and stability
9.1.2 Numerical accuracy and damping
Non-linear Newmark algorithm
Energy-conserving integration
9.3.1 State-space formulation
9.3.2 Non-linear kinematics for Green strain
9.3.3 Energy-conserving algorithm
Algorithmic energy dissipation
9.4.1 Spectral analysis of linear systems
9.4.2 Linear algorithm with energy dissipation
9.4.3 Non-linear algorithm with energy dissipation
Summary and outlook
Exercises

290

292
295
300
304
309
310
311
315
323
323
325
327
331
333

References
Index

336
345


Preface

The aim of this book is to take the reader on a concentrated tour of some
of the central issues of non-linear modeling and analysis of structures and
solids. Traditionally, the non-linear theories of solids have been treated in
books on continuum mechanics, while the questions of analysis have formed
the focus of books on finite element techniques. The idea of the present
book is to place the emphasis on modeling with a view to its numerical

implementation right from the outset. Two guiding principles have determined the main style of the book: the story should be told in the form of
concentrated chapters, each giving the central ideas of a specific aspect such
as ‘finite rotations’ or ‘elasto-plastic solids’, and the reader should have the
possibility of getting a feel for the numerical implementation by access and
use of simple high-level implementations of the basic algorithms. A text
based on these principles cannot provide exhaustive coverage, but aims at
giving an interesting introduction to the basic ideas, which can then be studied elsewhere in greater detail as needed. It is hoped that the combination
of a concise theoretical presentation in plain language supported by specific
algorithms will make the text of interest to graduate students as well as
professionals.
The book contains nine chapters: a brief introductory chapter setting the
scene by use of elementary arguments, four chapters on structures, two chapters on non-linear deformation and material behavior of solids, and finally
two chapters on numerical techniques for non-linear quasi-static and dynamic analysis. The theory is combined with demonstrations and exercises
using a small Matlab toolbox FemFiles providing routines for creation
and assembly of element matrices and permitting the solution of non-linear
finite element problems in a fairly simple script file format. The toolbox
FemFiles is available from the author via the internet. Exercises that require the use of a high-level program like FemFiles are marked ∗.
ix


x

Preface

The text started as a draft manuscript prepared for a short introductory
course on non-linear aspects of the finite element method at Aalborg University in the fall of 1992. A visit to Lund Institute of Technology sponsored
by NorFA provided an opportunity to include additional material on the
numerical aspects. The text was later extended with material on finite rotations, co-rotating formulation of elements, potential theory of plasticity
theory and plasticity models for geotechnical materials, and conservation
algorithms for numerical integration of dynamic problems. Several parts of

this work have been sponsored by the Danish Technical Research Council.
The work on bringing it all together was initiated during a visiting appointment as Melchor Professor at the University of Notre Dame, Indiana, in the
fall of 2001. The final stage has been combined with courses at Helsinki University of Technology 2004, and at Aalborg University and Lund Institute
of Technology 2005.


1
Introduction

Many problems of practical interest involve non-linear behavior of solids
and structures. In the present context a solid means a body with a firm
shape, as opposed to a fluid, while a structure refers to a solid composed
of slender elements such as beams, plates and shells. Typical problems
are the motion of robots, collapse scenarios of structures, metal forming
processes in industrial production, and material deformation and failure in
geotechnical engineering. These problems typically involve a considerable
change of shape, often accompanied by non-linear material behavior.
The finite element method is an important tool for the analysis of nonlinear problems, such as geometrical and material non-linear behavior of
solids and structures. The solution of non-linear problems by the finite
element method involves modeling, leading to the formulation of an appropriate set of non-linear equations describing the problem, followed by an
appropriate strategy for the numerical solution of these equations. In contrast to linear problems, where the solution strategy reduces to the solution
of a system of linear equations, the solution phase in a non-linear problem
typically involves an iterative procedure.
Non-linear modeling and analysis is a very active research area with many
engineering applications. The many different aspects involved are not covered in any single text. However, some central references to general texts
should be given here. A brief introduction to some of the basic problems
of non-linear finite element analysis of solids and structures is included in
the book by Cook et al. (1989). A general state-of-the-art presentation of
the finite element method, including the non-linear aspects of solids, structures and fluids, has been given in Zienkiewicz and Taylor (2000). A presentation with main emphasis on incremental formulation of geometrically
non-linear problems, including details of implementation, has been given by

Bathe (1996). The books by Crisfield (1991, 1997) and Belytschko et al.
1


2

Introduction

(2000) are entirely devoted to non-linear analysis of solids and structures,
combining illustrative examples with specific finite element procedures.
The present text is an introduction to some of the central ideas of nonlinear modeling and finite element analysis. It covers theoretical aspects of
geometric and material non-linearity and associated numerical techniques.
The text proceeds from the elementary level to a fairly rigorous presentation of ideas used in current research. Only the main ideas can be covered,
and the references should be consulted according to need. This first chapter gives an illustration of geometric non-linear behavior with reference to
a simple two-element truss model. The example serves as a vehicle for an
informal introduction to a non-linear load–displacement relation, the tangent stiffness, and the relation between the equilibrium and the virtual work
approach to the problem. The example also provides a simple realistic nonlinear equation on which to try different variants of the Newton–Raphson
solution technique.

1.1 A simple non-linear problem
The simple two-element truss model shown in Fig. 1.1 has often been used
to illustrate some basic features of geometric non-linear behavior, see e.g.
Bathe (1996, p. 494) and Crisfield (1991, pp. 2–13). The structure consists
of two identical truss elements, loaded with a vertical force f at the center
and simply supported at the other ends. The vertical displacement at the
center is called u. In the initial configuration the length of the bars is l0 .

Fig. 1.1. Two-element truss model.

Application of the load leads to a deformed state with vertical displacement u of the central node, Fig. 1.2. The structure is assumed to be shallow,

i.e. a
b. This permits series expansion of the square roots defining the
original bar length l0 and the bar length l corresponding to the current


1.1 A simple non-linear problem

3

deformed state:
l0 =

b2 + a2

l =

b2 + (a + u)2

b 1+

1 a2
,
2 b2
1 a+u
b 1+
2
b

(1.1)
2

.

(1.2)

Fig. 1.2. Initial length l0 and current length l.

The deformation of the bars is described by their elongation. A nondimensional measure of deformation is the engineering strain, defined as the
elongation relative to the original length,
ε =

l − l0
l0

au
1 u
+
l0 l0
2 l0

2

.

(1.3)

The first term is the linear part of the strain, while the second term is nonlinear. A true measure of deformation must not be influenced by any rigid
body motion of the bar, and thus a true deformation measure must be a nonlinear function of the displacement component(s). If the displacement u is
small relative to all characteristic lengths of the geometry – l0 and a – the
linear term will constitute a fair approximation, but if this approximation is

used, some of the characteristic non-linear features of the problem are lost.
1.1.1 Equilibrium
The two bars are assumed to be linear elastic with axial stiffness EA, where
E is the elastic modulus and A is the cross-section area. Thus, the axial
force in each bar is expressed in terms of the strain as
N = EA ε

EA

au
1 u
+
l0 l0
2 l0

2

.

(1.4)

Equilibrium of the central node in the deformed state requires that the
external force f is equal to the internal force g(u) generated by deformation
of the structure. Projection of the normal force gives
g(u) = 2 N

a+u
l

2EA

au + 12 u2 (a + u).
l03

(1.5)


4

Introduction

In non-dimensional form this is
g(u) = 2 EA

a
l0

3

u 3 u
+
a 2 a

2

+

1 u
2 a

3

,

(1.6)

where the normalized displacement is u/a. The load–displacement relation
(1.6) is shown in Fig. 1.3 corresponding to a downward load.

Fig. 1.3. Load–displacement curve for two-element truss.

From the unloaded state A an increasing downward load leads to a local
maximum B. In this state the structure cannot support a further increase
of the load. Thus, further increase of the load from B would lead to snapthrough to F . The snap-through is an unstable dynamic process, and thus
the load–displacement curve in Fig. 1.3 is not fully representative. Alternatively the structure may be loaded in displacement control, in which the
central node is given a controlled downward displacement −u. This would
require an increasing load from A to B, and then a decreasing load from B
to C, where u = −a and the two bars form a straight line. An upward force
is now required to proceed to D and E, where the structure is stress-free,
forming an angle symmetric to the original configuration with respect to the
base line. Further downward load leads through F with increasing stiffness
of the structure.
For a structure with one degree of freedom, the stiffness is a measure of
the change in force for a given change in displacement. Thus, the tangent
stiffness K is defined as the stiffness corresponding to infinitesimal changes
in u and g:
K =

dg
.
du

(1.7)

In the present case the tangent stiffness K follows from straightforward


1.1 A simple non-linear problem

5

differentiation of (1.6):
K =

2EA a
l0
l0

2

1+3

3 u
u
+
a
2 a

2

.

(1.8)

Although this expression defines the tangent stiffness K, it does not convey
the physics of the problem very clearly. This is better accomplished by
differentiation of the equilibrium equation (1.5):
K =

a+u
d
2N
du
l0

= 2

EA a + u
l0
l0

2

+2

N
.
l0

(1.9)

Here a + u is the height of the structure in the current state, while N is
the current value of the axial force. The first term is due to changes in the
normal force N , while the second term is due to changes in the geometric
configuration with constant normal force N . Sometimes the first term is
separated into a constant corresponding to u = 0 and the rest, whereby
(1.9) takes the form
K = 2
=

EA
l0
K0

a
l0

2

+2
+

EA 2au + u2
N
+2
2
l0
l0
l0
Ku

+ Kσ ,

(1.10)

where K0 is the linear stiffness, Ku is the initial displacement stiffness, and
Kσ is the initial stress stiffness. In an incremental procedure, where the
geometry is updated, the current value of u is absorbed in the updated
value of a, and in that case the initial displacement stiffness Ku vanishes.

Fig. 1.4. Load–displacement curve for two-element truss with spring.

A family of load–displacement curves with different degrees of non-linearity
can be obtained by introducing a vertical linear elastic spring with stiffness


6

Introduction

k at the central node of the structure. The load–displacement relation (1.6)
is changed to
a
l0

g(u) = 2EA

3

u 3 u
+

a 2 a

2

+

1 u
2 a

3

+ ku

(1.11)

and the tangent stiffness (1.8) to
K =

2EA a
l0
l0

2

1+3

3 u
u
+
a

2 a

2

+ k.

(1.12)

Figure 1.4 shows the load–displacement curve for different values of the
spring stiffness k. For k ≥ EAa2 /l03 the variation of load with displacement
is monotonic, corresponding to K ≥ 0.
1.1.2 Virtual work and potential energy
The load–displacement relations (1.6) and (1.11) were obtained from equilibrium of the center node. For structures with more degrees of freedom or
more complicated elements it is often convenient to make use of the principle
of virtual work. Essentially, the principle of virtual work is a restatement
of a set of equilibrium equations, where each equation is multiplied by a
corresponding infinitesimal virtual displacement component. With an appropriate definition of the force and displacement components summation
of their products forms a scalar invariant, known as the virtual work.
In the particular example of the two-element truss with an elastic spring
the equilibrium equation can be written as
a+u
2N
+ ku − f = 0.
(1.13)
l
Multiplication by a virtual displacement δu gives the virtual work equation
a+u
δu + (ku)δu − f δu = 0.
(1.14)
δV = 2 N

l
The displacement factor in the first term is similar to the first variation of
the strain (1.3):
δε =

∂ a u
1 u
+
∂u l0 l0
2 l0

2

δu =

a + u δu
.
l0 l0

(1.15)

If, for the time being, the difference between l0 and l is neglected, the virtual
work equation (1.14) can now be written as
l0

δV

2

N δε ds + (ku) δu − f δu = 0.

(1.16)

0

The integral is the internal virtual work of the bar elements, the second term


1.2 Simple non-linear solution methods

7

is the virtual work of the elastic spring, while the last term is the external
virtual work.
Apart from the factor l0 /l that is somehow missing, the use of virtual work
in the present case where δε is constant within the elements is almost trivial.
However, for more general problems with more degrees of freedom and nontrivial displacement fields within the elements, the principle of virtual work
is an important tool for establishing the balance equations of the discretized
model. The question of the factor l0 /l is discussed in Chapter 2, where the
theory of non-linear bar elements is discussed more rigorously. Here, the
relation between virtual work and potential energy is discussed briefly before
turning to elementary numerical solution methods for non-linear equilibrium
equations.
When the internal forces such as the axial force N and the spring force
ku are functions of the state of displacement given by u, and the external
load is also a function of u, the virtual work δV can be considered as the
differential of an energy function Φ(u) – the potential energy. In the present
case (1.16) is written as
l0

δΦ(u) = 2

EA εδε ds + ku δu − f δu.

(1.17)

0

This relation can be integrated with respect to the displacement u, giving
the following expression for the potential energy:
l0

Φ(u) = 2
0

2
1
2 EAε

ds + 12 k u2 − f u.

(1.18)

The potential energy is the internal strain energy of the structure, including
the spring, minus the external work represented by f u. For linear elastic structures it may be simpler to derive the equilibrium equations from
the potential energy by considering an incremental change δu of the displacements. However, the principle of virtual work is valid irrespective of
the specific material behavior, and thus the principle of virtual work has
become the method of choice for setting up equilibrium equations.

1.2 Simple non-linear solution methods

For a system with only one degree of freedom non-linear behavior can often
be described explicitly as a function of the displacement u, and the problem
may then be considered as one of displacement control. However, in the case
of several degrees of freedom the use of displacement control is non-trivial,
and most problems are formulated in terms of a load history, for which


8

Introduction

the corresponding displacement history is to be calculated. This requires
the solution of a system of non-linear equations. Here some of the simpler
methods for solving non-linear equations are briefly introduced, leaving more
specialized techniques to Chapter 8. The methods are illustrated for a single
degree of freedom and then generalized to matrix form.
1.2.1 Explicit incremental method
An explicit incremental method, often called the Euler explicit method, is
obtained by replacing the differentials in the definition (1.7) of the tangent
stiffness with finite increments ∆f and ∆u:
∆u = K −1 ∆f.

(1.19)

The load–displacement history is described by a number of increments ∆fn ,
∆un , n = 1, 2, . . . defining the states
fn = fn−1 + ∆fn ,

un = un−1 + ∆un ,

n = 1, 2, . . .

(1.20)

In the explicit incremental method the tangent stiffness K corresponds to the
state at the beginning of the increment. Thus, the precise form of (1.19) is
∆un = K −1 (un−1 ) ∆fn ,

n = 1, 2, . . .

(1.21)

This procedure is illustrated in Fig. 1.5.

Fig. 1.5. Explicit incremental method.

It is seen that the computed states deviate more and more from the exact load–displacement curve. There are two reasons for this: the tangent
stiffness of each increment is taken at the left end-point and in this particular case overestimates the stiffness, and deviations from the exact curve are


1.2 Simple non-linear solution methods

9

added to a cumulative error. While it is difficult to use an exact representation for the stiffness corresponding to the full increment, the problem of
increasing deviations can be countered by introducing equilibrium iterations
as discussed in the following.
The explicit incremental method is easily generalized to multi-degree of
freedom systems. Let the displacement vector be u and the corresponding
load vector f . The tangent stiffness matrix K is then defined by

df = K(u) du.

(1.22)

The corresponding explicit incremental method is
∆un = K−1 (un−1 ) ∆fn ,

n = 1, 2, . . .

(1.23)

The use of the inverse matrix K−1 in (1.23) should not be taken literally.
In practice the matrix K is factored and the product K−1 ∆f found by back
substitution.
1.2.2 Newton–Raphson method
In order to avoid accumulating errors in each additional load step, equilibrium iterations may be used to establish equilibrium to a desired degree
of accuracy at each load step. This procedure is a special instance of the
Newton–Raphson method, well known from numerical analysis. In principle, the method works by applying two steps intermittently: (i) check if
equilibrium is satisfied to within the desired accuracy; (ii) if not, make a
suitable adjustment of the state of deformation.
The first step consists in checking the equilibrium equation. This is done
by forming the difference between the external load f and internal force
g(u),
r(u, f ) = f − g(u) = 0,

(1.24)

where r(u, f ) is called the residual force. In a state of equilibrium the internal force g(u) is equal to the external load f , and thus the residual vanishes.
In practice, lack of equilibrium will be produced at the beginning of each
load increment, where the load f is increased, while no new displacement

estimate u is yet available. Thus, the need arises for obtaining an improved
estimate of the state of displacement u.
In the absence of equilibrium an improved estimate of the displacement
u is obtained from a linearized form of the residual r(u + δu, f ) around the
known residual r(u, f ),
r(u + δu, f ) = r(u, f ) + δr(u, f ) + · · · = 0.

(1.25)


10

Introduction

The dots indicate higher-order terms, because δr is only a linearized form of
the increment of the residual. In the classic form of equilibrium iterations
the load f is assumed fixed within the given load step, and thus the increment of the residual only depends on the internal force g(u). The linearized
increment is then given by the first derivative of the internal force as
dg(u)
δu = −K(u) δu.
(1.26)
du
Here the tangent stiffness K, introduced in (1.7), has been introduced. The
displacement increment is now obtained from the linearized form of (1.25)
by substitution of the tangent stiffness relation (1.26). When rearranging
the terms in (1.25), the linearized equation becomes
δr = −

K(u) δu = r.

(1.27)

In this equation the residual r(u, f ) is known, as it relates to the current
state of load f and displacement u. The tangent stiffness K(u) at the
current displacement state u can also be calculated. Thus, this equation
permits determination of the displacement increment δu,
δu = K −1 (u) r.

(1.28)

Once the displacement increment δu is determined, the current displacement
state is updated to
ui = ui−1 + δui .

(1.29)

In this equation the superscript is used to indicate that the iteration i
changes the estimated displacement from ui−1 to ui . In a computer program the iteration superscript i is not needed, as the register containing
ui−1 is simply overwritten by the new value ui according to the assignment
statement
u : = u + δu.

(1.30)

Here, : = is the assignment operator, implying that the variable u is assigned
a new value. In this book many of the algorithms are presented in the form
of pseudocode – i.e. a code format that appears like high-level programs such
as Matlab. In the pseudocode presented here assignments are indicated by
the normal equality sign, as all equalities are assignment statements.
The Newton–Raphson equilibrium iteration procedure is illustrated in

Fig. 1.6. The figure shows load step n. This load step starts from a state of
equilibrium already established at the previous load fn−1 with displacement
un−1 . The load step is initiated by increasing the load by ∆fn to fn . This
generates the first residual rn1 = ∆fn . This residual and the tangent stiffness


1.2 Simple non-linear solution methods

11

Fig. 1.6. Newton–Raphson equilibrium iterations.

K(un−1 ) lead to the displacement increment δu1n , shown in the figure. At the
new displacement un−1 +δu1n , the internal force g – represented by the curve
– is still smaller than the imposed load. The difference forms the residual
rn2 , and the procedure is continued. It should be noted that the use of suband superscripts to indicate load step and iteration number, respectively, is
merely for illustration in relation to the figure. These indices are not needed
when programming the algorithm.
The iteration process needs a termination criterion. This may be taken
as the requirement that the current residual force rn should be less than a
prescribed fraction of the load increment ∆f of the present load step,
|r| <

|∆f |.

(1.31)

The value of could be on the order of say 10−4 –10−6 . For structures
developing very small stiffness, the criterion (1.31) may be supplemented by
the displacement criterion

|δu| <

|∆u|,

(1.32)

where ∆u is the total displacement increment accumulated in the present
load step.
In the corresponding multi-component problem with displacement vector
u and load vector f , the residual force vector is
r(u, f ) = f − g(u).

(1.33)

The tangent stiffness matrix is defined by the incremental change of the


12

Introduction
Algorithm 1.1. Newton–Raphson method.
Load steps

n = 1, 2, . . . , nmax

fn = fn−1 + ∆fn
un = un−1
Iterations i = 1, 2, . . . , imax
dg(un )
du

= fn − g(un )

Kn =
rn

δun = K−1
n rn
un = un + δun
Stop iteration when

rn <

∆fn

End of load step

internal forces,
K(u) =

dg(u)
.
du

(1.34)

The tangent stiffness is now introduced into a linearized form of the equilibrium condition, whereby the following vector equation is obtained for the
displacement sub-increment δu:
K(u) δu = r.

(1.35)

In contrast to the one-dimensional case, the solution of these equations may
be a non-trivial part of the procedure. The termination criteria will typically
make use of the ‘length’ of the corresponding vectors, whereby
(rTn rn )1/2 <
(δuT δu)1/2 <

(∆f T ∆f )1/2 ,

(1.36)

(∆uT ∆u)1/2 .

(1.37)

The Newton–Raphson procedure is summarized as Algorithm 1.1. Note that
in the iteration loop the computer overwrites quantities like rn by their new
value in the same register. Therefore, the superscript i does not appear
explicitly in the algorithm. Similarly, the load step subscript n is only used
to avoid the explicit indication of storing the result of each completed load
step. The actual algorithm is conveniently programmed without the use of
indexed variables in the iteration loops.


1.2 Simple non-linear solution methods

13

1.2.3 Modified Newton–Raphson method
In the original Newton–Raphson method the current tangent stiffness matrix

K(u) is computed and factored in each iteration. For non-linear problems
with a single or a few degrees of freedom this is usually not a problem, but for
problems with many degrees of freedom the computational cost involved in
forming the stiffness matrix K(u) and solving the corresponding equations
for δu in each iteration may be considerable. It is seen from Fig. 1.6 and Algorithm 1.1 that Kn appears within the inner loop. A simple modification of
the Newton–Raphson method consists in moving the stiffness matrix K outside the iteration loop. Then K = Kn−1 is only computed and factored once
for each load step on the basis of the previous state of displacement un−1 ,
dg(un−1 )
.
(1.38)
du
This simplifies the iteration loop as shown in Fig. 1.7 and Algorithm 1.2.
Kn−1 =

Fig. 1.7. The modified Newton–Raphson method.

The asymptotic convergence of the modified Newton–Raphson method is
slower than that of the Newton–Raphson method, and this may offset some
of its computational efficiency. A different, more refined type of modification makes use of a secant approximation of K. This requires a non-trivial
generalization of the secant concept to multi-degree of freedom systems.
The corresponding methods, called quasi-Newton methods, are described in
Chapter 8.
The classic Newton methods encounter problems at a load maximum.
Several methods have been developed to deal with this problem. A common
feature of these methods is that the load increment is also subject to changes
during the iterations, e.g. by linking load and displacement increments. This