Professional Documents
Culture Documents
NUMERICAL
METHODS IN
STRUCTURAL
MECHANICS
This book provides a clear understanding of the nature and theoretical basis of
the most widely used numerical methods—the finite element method (FEM) and
the boundary element method (BEM)—while at the same time presenting the
most promising directions for future developments. Attention is paid mainly to
those methods that have proven to be the most reliable and efficient, as well as
those methods currently under rapid development. Examples were selected
either to illustrate various computational algorithms and compare their accuracy
and efficacy or to elucidate the mechanical processes under investigation, while
traditional examples that are already covered by standard textbooks have been
deliberately omitted. Emphasis is placed on the understanding of basic princi-
ples, rather than on the details of individual numerical algorithms. The booK cov-
ers all topics essential for students of elementary and intermediate courses on
numerical methods in solid mechanics, and it also serves as a useful reference
for researchers and other professionals. This book was recently translated from
the highly regarded, original Czech edition.
Bittnar, Zdenek.
Numerical methods in structural mechanics / Zdenek Bittnar, Jiri Sejnoha.
p. cm.
Includes bibliographical references.
ISBN 0-7844-0170-5
1. Structural analysis (Engineering) 2. Numerical analysis. I. Sejnoha, Jirf. II.
Title.
TA645.B59 1996 96-14306
624.17—dc20 CIP
The material presented in this publication has been prepared in accordance with
generally recognized engineering principles and practices, and is for general
information only. This information should not be used without first securing com-
petent advice with respect to its suitability for any general or specific application.
The contents of this publication are not intended to be and should not be con-
strued to be a standard of the American Society of Civil Engineers (ASCE) and
are not intended for use as a reference in purchase specifications, contracts, reg-
ulations, statutes, or any other legal document.
No reference made in this publication to any specific method, product, process,
or service constitutes or implies an endorsement, recommendation, or warranty
thereof by ASCE.
ASCE makes no representation or warranty of any kind, whether express or
implied, concerning the accuracy, completeness, suitability, or utility of any infor-
mation, apparatus, product, or process discussed in this publication, and
assumes no liability therefore.
Anyone utilizing this information assumes all liability arising from such use, includ-
ing but not limited to infringement of any patent or patents.
Part I can be studied by readers who have acquired basic knowledge in elementary
courses such as Strength of Materials, or Structural Analysis. It consists of five chapters.
Chapter 1 is a review of basic notions, relations and principles of solid mechanics. It
should not only facilitate further reading but also make the reader aware of new trends
in nonlinear material modeling. Problems related to damage localization, size effect, etc.,
are so important that, despite the limited scope of this book, the authors at least briefly
explain their essence and give the appropriate references.
Chapter 2 is devoted to skeletal structures (trusses, frames and grillages) with special
attention to soil-structure interaction. It presents a consistent derivation of the stiffness
matrix of an elastic foundation based on the Winkler-Pasternak model, which is later used
in linear stability and vibration analysis. Attention is also paid to curved beam elements
based on the principle of decomposition of membrane and bending effects. The chapter
is concluded by remarks on static condensation and on coordinate transformation.
Chapter 3 represents the core of the part devoted to linear problems. After an ini-
tial introduction to isoparametric elements, a thin-walled beam element based on the
Umanski-Mindlin-Reissner hypothesis is derived. The next section presents elements for
plane problems (plane stress or plane strain analysis) with several useful modifications,
which can be exploited when analyzing deep beams, when combining in-plane loaded
plates with frames, and when constructing efficient shell elements. Elements for plate
bending (optionally supported by an elastic Winkler-Pasternak foundation) are derived
from Kirchhoff theory, and from Mindlin-Reissner theory. The curved beam element
based on the principle of decomposition from Chapter 2 is generalized to a shell element.
The last portion of the chapter deals with special elements for subgrade modeling in soil-
structure interaction analysis. Chapter 4 generalizes plane elements to three-dimensional
solid elements.
Chapter 5 is devoted to linear stability and vibration analysis. Aside from stan-
dard methods (Rayleigh-Ritz method, inverse iteration, Jacobi method, subspace itera-
tion method), Lanczos method is thoroughly discussed. Forced vibrations are analyzed by
eigenmode decomposition (with special emphasis on alternate models for damping), and
by direct integration (central difference scheme, Newmark method and Wilson method).
The latter approach is applicable to linear as well as nonlinear equations of motion. Two
methods of finding a periodic response to a harmonic excitation (the solution in complex
numbers and the eigenmode decomposition) are then explained, and their applicability to
models with proportional and nonproportional damping is discussed.
Part II has been designed for readers who are already familiar with methods of linear
finite element analysis. It consists of six chapters covering three main subjects: special
linear problems solved by FEM (Chapters 6 and 7) and BEM (Chapter 8), nonlinear
problems (Chapter 9), and some modern topics (adaptivity in Chapter 10 and probabilistic
approach in Chapter 11).
Chapter 6 presents semianalytical solutions based on Fourier expansion in one direction
and finite element discretization in the other (orthogonal) direction. Methods of this kind
(finite strip methods) are applicable, e.g., to curved box girders.
Chapter 7 deals with other special applications of FEM. Analysis of warping torsion
is followed by diffusion problems (heat conduction and moisture transport). A similar
numerical approach is applied in the analysis of deformation of soils and other porous
materials. Some problems of linear elastic fracture mechanics are also included, and they
are supplemented by comments on nonlinear fracture mechanics.
Chapter 8 explains basic ideas of the boundary element method and its modifications.
It tackles both static and dynamic problems with special emphasis on recent developments
leading to a symmetric version of BEM, which has important advantages when combining
BEM with FEM.
Chapter 9 shifts the focus to nonlinear problems. It addresses both geometric and ma-
terial nonlinearities. Geometrically nonlinear effects are demonstrated by an elementary
example of a truss element. The basic notions are then generalized for a continuum, and
the Total Lagrangian and Updated Lagrangian formulations using the incremental form of
the principle of virtual displacements are explained. The discretization procedure is then
generalized to isoparametric elements of an arbitrary shape and supplemented by com-
ments on discretization of a degenerate continuum (arches and shells). Special attention
is paid to modern solution methods for sets of nonlinear equations. Besides being very
efficient, these methods are applicable even to problems for which the standard Newton-
Raphson technique with load control fails (snap-through, snap-back). This section also
includes basic facts on stability analysis of individual branches of the equilibrium dia-
gram. BEM has some advantages when applied to problems with material nonlinearity.
Dual formulations based on initial strain and initial stress concepts are presented and
discussed.
Chapter 10 is devoted to the currently very popular area of adaptive meshes, especially
to hierarchical elements and the p-version of FEM. The mathematical theory of FEM has
provided reliable error estimators. Based on an error estimate, the mesh can be modified so
that the error is approximately uniform. Applications of artificial intelligence to adaptive
remeshing are briefly discussed and illustrated by an example.
Chapter 11 gives an overview of probabilistic methods used in combination with FEM
or BEM, which include statistical methods (Monte Carlo simulation, stratification LHS
method) and nonstatistical methods (probabilistic FEM).
The book is appropriate for undergraduate students on senior level (Volume I) and
for graduate students (both parts). In the authors' opinion, it provides material for up
to four courses—fundamentals of linear FEM, dynamic analysis, nonlinear problems and
special topics.
Contents
I 7
1 Basic Notions, Equations and Principles 9
1.1 Basic equations of elasticity 9
1.2 Linear elastic materials 10
1.2.1 Constitutive equations for anisotropic materials 10
1.2.2 Transformation of constitutive equations for orthotropic materials . 13
1.2.3 Tensorial form of elasticity equations 15
1.3 Elastoplastic materials 16
1.3.1 Yield criterion and yield function 16
1.3.2 Constitutive equations for elastoplastic materials 20
1.4 Damage theory 23
1.4.1 Model of brittle damage 24
1.4.2 Strain localization in softening media 27
1.4.3 Discontinuum modeling 31
1.4.4 Enhanced continuum approach 35
1.5 Viscoplastic materials 38
1.5.1 Constitutive equations for uniaxial stress 39
1.5.2 Incremental constitutive equations—uniaxial stress 41
1.5.3 Incremental constitutive equations—triaxial stress 42
1.6 Principle of virtual work and variational principles 43
1.6.1 Principle of virtual work (PVW) 43
1.6.2 Variational principles 45
1.6.3 Modified variational principles 48
1.6.4 Ritz method 51
1.7 Convergence criteria 54
1.8 Variational principles in anisotropic and nonhomogeneous elasticity 55
1.8.1 Variational principle for body with prescribed surface displacements 55
1.8.2 Dual variational principle for body with prescribed surface tractions 57
1.9 Variational formulation of rate boundary value problem including softening 59
1.10 Nonlinear systems and stability criteria 60
2 Skeletal Structures 64
2.1 Basic relations for beams 64
2.1.1 Transformation of elasticity equations 64
2.1.2 Beam on elastic foundation 67
2.2 Truss and beam elements 73
2.2.1 Force approach 73
2.2.2 Displacement approach 77
2.3 Curved beam element 83
2.4 Grillage element 85
2.4.1 Analogy between axial and torsional deformation 85
2.4.2 Grillage element on Winkler-Pasternak foundation 86
2.5 Static condensation 88
2.6 Coordinate transformation 94
3 Plates and Shells 100
3.1 Basic relations for isoparametric elements 100
3.1.1 Nature of isoparametric elements 100
3.1.2 Approximation functions on a quadrilateral 101
3.2 Basic relations for triangular elements 104
3.2.1 Area coordinates on a triangle 104
3.2.2 Approximation functions on a triangle 106
3.3 Tension-compression bar 106
3.4 Thin-walled elements 107
3.5 Elements for plane problems 109
3.5.1 Triangular element 111
3.5.2 Isoparametric bilinear quadrilateral element 114
3.5.3 Modified quadrilateral element 115
3.5.4 Plane element with rotational degrees of freedom 119
3.6 Plate elements 128
3.6.1 Mindlin theory of thick plates 129
3.6.2 Triangular element DKT (Discrete Kirchhoff Theory) 136
3.6.3 Constant Curvature Triangle (CCT) 138
3.6.4 Quadrilateral plate element on elastic foundation 141
3.6.5 Modified quadrilateral plate element 144
3.7 Shell elements 147
3.7.1 Curved triangle in a local coordinate system 147
3.7.2 Transformation of the shell element into global coordinates 151
3.8 Interaction between structure and foundation 151
3.8.1 Noninteracting foundation structures 152
3.8.2 Interaction of foundation structures 153
3.9 Patch test 157
4 Solids 160
4.1 Tetrahedra 160
4.2 Bricks 162
4.3 Brick with rotational degrees of freedom 163
4.4 Axisymmetric continuum 166
II 213
6 Semianalytical Methods 215
6.1 Energy-based beam analysis by Fourier series 215
6.2 Finite strip method 218
6.2.1 Finite strip method for thick plates 218
6.2.2 Interpolation functions and numerical integration 221
6.3 Curved box girders 222
6.3.1 Approximation of unknown functions. Strip stiffness matrix 225
6.3.2 Axisymmetric shells 226
6.3.3 Transformation of coordinates 226
6.4 Plane strip with rotational degrees of freedom 227
7 FE Solution of Special Problems 230
7.1 Torsion of bars 232
7.1.1 Stiffness approach 232
7.1.2 Flexibility approach 234
7.1.3 Calculation of stiffness moment in free torsion 236
7.2 FE solution of diffusion equation 238
7.3 Deformation of soils and other porous materials 240
7.3.1 Basic notions and relations. Concept of effective stress 240
7.3.2 Deformation of solid skeleton 241
7.3.3 Equation of continuity and equations of equilibrium 242
7.3.4 Variational formulation and FE solution 243
7.4 FEM in fracture mechanics 246
7.4.1 Stress intensity factor. K-concept 248
7.4.2 Energy criteria of fracture 251
7.4.3 Effect of plasticity on crack stability analysis 254
8 Boundary Element Method 258
8.1 Somigliana's formulae 259
8.2 Direct version of BEM 265
8.2.1 Formulae for a boundary point 265
8.2.2 Boundary element discretization 266
8.2.3 Evaluation of matrices H and G 269
8.3 Symmetric version of BEM 272
8.4 Transformation field analysis using BEM 277
8.4.1 Body with prescribed surface displacements 278
8.4.2 Body with prescribed boundary tractions 280
8.4.3 Optimization problem 282
8.5 Solution of dynamic problems by BEM 283
8.5.1 Alternative BEM formulation 283
8.5.2 Symmetric version in dynamic problems 285
8.6 Plate analysis by BEM 288
8.6.1 Static analysis of thin plates—direct version of BEM 288
8.6.2 Dynamic analysis of thin plates—indirect version of BEM 294
Index
417
This page intentionally left blank
Part I
This page intentionally left blank
Chapter 1
9
10 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
New symbols denote the prescribed boundary tractions p on Yp and the prescribed dis-
placements u on Fu. The matrix of direction cosines n x ,n y and nz (components of the
unit outward normal to the surface F) has a structure similar to the matrix d:
The stress field <r and the displacement field u are coupled by a useful integral
relation—the so-called divergence theorem or Clapeyron theorem,
which immediately follows from the Gauss integral theorem. Clapeyron theorem can be
interpreted as an equality between the internal work (left-hand side) and the external
work of surface tractions and body forces (right-hand side).
where D is a symmetric material stiffness matrix of type (6,6). For general anisotropic
materials, this matrix has 21 independent elements (elastic constants). The vector of
initial strain e0 = {£oz,£oy,£oz>0,0>0} T represents the effects of temperature changes,
shrinkage, etc. For dilation due to temperature variation we have
where T is the temperature variation [K], and o>x,ayiaz are the coefficients of thermal
expansion [-K""1]. For common structural materials (e.g., steel or concrete) we can set
ax = OLy = az = 0,000012.
Combining the first equation from (1.3) with (1.10) we get the constitutive equations
of linear elasticity in the matrix form
Fully general anisotropy occurs only for materials arranged in the triclinic system. A
less general case, important for the engineering practice, is the rhombic anisotropy with
three orthogonal planes of elastic symmetry, which is referred to as orthotropy. Using the
technical constants E, v and G, the material compliance matrix is expressed as
The matrix contains only nine independent constants, because the elements of the left
upper block are linked by three symmetry conditions
Denoting
we can write
Elastic constants of the left upper block are given by (1.18) and (1.19). The inverse
relation has the form
where
The plane stress description is based on a reduction of the matrix (1.15), after which
the constitutive equations read
where
Material orthotropy is typical, e.g., for orthogonally reinforced concrete. Due to the
lack of experimental data, the main difficulty is usually the determination of the shear
stiffness. A common practice is to determine the modulus Gxy from a supplementary
condition of invariance with respect to a rotation of coordinates [52].
Introducing the equivalent Poisson's ratio
we get from (1.24) and (1.25) a three-parametric constitutive relation of the form
1.2. LINEAR ELASTIC MATERIALS 13
Table 1.1: Material compliance and stiffness matrices for plane stress and plane strain
where
Suppose that we know the matrix D1 defined with respect to the local coordinate
system, and we search for the matrix D related to the global coordinate system. We will
restrict our attention to the planar description of an orthotropic material schematically
shown in Fig. 1.1 a.
14 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
The components of the strain tensor are transformed according to the well-known
formula
from which
where
1.2. LINEAR ELASTIC MATERIALS 15
The material stiffness matrix is a sum of two matrices. If du = d\i = 0, then the second
matrix in (1.32) vanishes and the first matrix corresponds to the stiffness of the material
damaged by cracks in direction y* = 2 (c?22 ^ 0) tnat s^ transfer some shear. The
shear stiffness G\i ^ 0 must be reduced with respect to the shear stiffness of the basic
material. The dependence of the shear retention factor on the crack opening width is
given in Section 1.4.1.
where Dijki is the material stiffness tensor. For isotropic materials, this tensor is given by
The so-called isotropic tensor 6ij (Kronecker delta) assumes values 1 (for i = j) and
0 (for i ^ j). In the following, the summation symbol will be omitted, and summation
over repeated subscripts will be implied.
Introducing the compliance tensor of an isotropic material,
We must distinguish between the tensorial shear strain component e\i and the "engineer-
ing" shear angle jxy = 712 = 2£i2, etc. It is useful to combine the constitutive equations
(1.33) with the strain-displacement equation (1.37). Assuming that e^i = 0 we get
This equation also holds for plane strain, for which the summation indices vary from 1 to
2.
The tensorial equation for plane stress is obtained from the above relation after re-
placing v by V = v/(l -f i/). A simple manipulation leads to
16 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
where the direction cosines rij are the components of the unit normal to the boundary.
The components of vector k = {ki, k2, ...}T are certain material constants.
One of the most useful yield conditions is the one due to Drucker and Prager,
where
is the mean stress, proportional to the first invariant of the stress tensor cr^-, and
is the second invariant of the stress deviator s^ = a^- — crv6ij (6ij is Kronecker delta). Fi-
nally, \£ is an empirical, monotonically increasing function (often defined as \£ = CK/I), and
k and a are positive material constants. Equation (1.43) corresponds to an axisymmetric
surface in the stress space, with the axis of symmetry coinciding with the hydrostatic
axis <TI = 02 = <73. The dependence on <jv is important for materials with different yield
1.3. ELASTOPLASTIC MATERIALS 17
limits in tension and in compression. Therefore, the Drucker-Prager condition has found
its application in soil mechanics and in modeling of concrete and other porous materials.
For many materials, most notably steel, the dependence on av is negligible, and equa-
tion (1.43) can be simplified to the von Mises condition
This condition is visualized by a circular cylinder in the stress space (Fig. 1.3). Let us
which in a uniaxial stress state reduces to the applied stress. Comparing (1.46) with
(1.47) we immediately see that \/3 k = Ry (yield stress).
The hexagonal prism corresponding to the Tresca condition of maximum shear stress
is inscribed into the cylinder visualizing the von Mises condition. Fig. 1.3b shows the
intersection of the yield surfaces with the plane <73 = 0. The yield curves (ellipse, hexagon)
determine the boundary between the elastic and plastic regions for plane stress (cr3 = 0).
If a material exhibits hardening, the surface described by equation (1.42) expands
depending on the loading history. This process is described by the hardening parameters1
k = k(t). Therefore, the yield function / is also called the loading function. Two loading
curves (yield curves) corresponding to two different time instants2 t\,ti are plotted in
Fig. 1,4. The inequality /(cr, k) < 0 corresponds either to initial elastic loading, or to
elastic unloading from a previously reached plastic state.
In a time interval during which the material remains in a plastic state, equation (1.42)
is satisfied [recall that /(cr, k) > 0 is not allowed according to the definition of the yield
function]. Differentiating the yield condition we arrive at the consistency condition
1
Various types of hardening are discussed, e.g., in [91, 137, 152, 142],
2
Time instants t\ and t^ only label different stress states. Elastoplastic material properties are inde-
pendent of real time.
18 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
where
As follows from the form of equation (1.43), the signs of the material constants can be
chosen such that for loading we always have (df/dk)Tdk < 0. The consistency condition
then provides us with the following loading criterion:
plastic loading
neutral loading
elastic unloading.
The criterion is written in terms of the scalar product of vectors do- and df/dcr, which
in Fig. 1.4 subtend angles c*i < ?r/2 , a2 = ^/2 , a3 > Tr/2.
A useful illustration of the given formulae is the Chen condition used in concrete
plasticity (see [43]). The yield function is composed of two parts: cf is valid in the
compressive region while tf describes the onset of yielding under combined tension and
compression (for a general stress state). The partitioning of the biaxial stress states into
the two regions is clear from Fig. 1.5a. The partitioning of general triaxial stress states
is best visualized in the plane (/i, \f~J~<i)—Fig. 1.5b. Individual regions are defined by the
following conditions:
a) compression-compression
b) compression-tension
c) tension-tension
d) tension-compression
1.3. ELASTOPLASTIC MATERIALS 19
In the compressive region defined by (1.5la), the initial yield surface is described by
Material constants A, k in equations (1.52) and (1.53) can be extracted from experimen-
tally measured yield limits in uniaxial tension, Ry, in uniaxial compression, RyC, and in
biaxial compression, Rybc- The constants are determined separately for the basic regions:
compression-compression
An expansion of the initial yield surface (Fig. 1.5) given by equations (1.52) and (1.53)
leads to subsequent loading surfaces described as follows:
In the compressive region defined by conditions (1.5la),
Constants A and k have been defined by formulae (1.54). Constants Au and ku can be
expressed by the same formulae after replacing the yield limits by the uniaxial tensile and
compressive strengths, R t and RC1 and by the biaxial compressive strength, Rbc.
20 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
The hardening parameter K is determined by the equivalent plastic strain, which will
be discussed in detail in the subsequent section [formula (1.73)]. Constants a and /3 have
been chosen such that for K = k the loading functions (1.55) and (1.56) reduce to the
initial yield functions (1.52) and (1,53).
Substituting K = ku, the functions (1.55) and (1.56) become the failure conditions
for the compressive zone,
The ultimate loading functions are represented by the failure envelope in Fig. 1.5. If the
end point of the stress vector a reaches the failure envelope in the compressive zone,
concrete fails by crushing, while in the tensile and tensile-compressive zone concrete fails
by cracking (separation). A more appropriate formulation of the failure conditions is the
one based on energy criteria (Section 1.4).
Equation (1.60) is also called the normality rule because it shows that the plastic strain
increment vector is orthogonal to the surface / = 0, Thus the yield function has the
meaning of a plastic potential
Parameter dX can be eliminated using the consistency condition (1.48), which is for
this purpose written in Melan's form
1.3. ELASTOPLASTIC MATERIALS 21
where
where
In this case the material does not have to satisfy Drucker's postulate of stability.
Relation (1.66) holds also in the strain-softening region where H < 0. Problems of strain
softening will be discussed in the subsequent section.
Let us present the final form of the elastoplastic stiffness matrix for the Drucker-Prager
yield condition (1.43) and for the plastic potential given by
which differs from the yield function only in the term expressing the influence of av.
Following [21], we denote3
where (3 and f l are the parameters characterizing material dilatancy and internal friction.
The final expression reads
3
Replacing / by g we get from (1.62) H
22 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
where r = \fj^ is the stress intensity, G is the shear modulus of elasticity defined by
(1.28), and
is for uniaxial stress equal to the plastic strain in the direction of applied stress. Setting
€yp = ezp = —(1/2) exp1 we get eeqp = exp (we consider the coefficient of lateral con-
traction corresponding to fully plastic response). The hardening modulus is according to
(1.62) given by
Taking into account (1.60) and the above definition of the hardening parameter «, we can
further transform this expression to
and so
The derived formula is useful when evaluating the hardening modulus from a uniaxial
test.
Using (1.74) we transform (1.72) into the form
where
where
where
dimensions) or bands (in two dimensions). In static problems, this so-called localization is
related to the loss of elipticity of the incremental equations of equilibrium, to the existence
of a bifurcation from the homogeneous into an inhomogeneous strain state, and to the
occurence of multiple equilibrium paths (cf. [145, 21, 26]). Ductile materials tend to fail
in shear, after large plastic strains have been accummulated in narrow localization zones.
A number of constitutive models for the description of strain softening, both phe-
nomenological ones and micromechanical ones, are listed in [21]. As far as phenomeno-
logical models are concerned, we have already mentioned nonassociated plasticity (1.66).
The entire evolution of damage (growth of existing microcracks and nucleation of new
ones), both in the stable and the unstable stage, can be described using the framework
of damage mechanics, which, combined with fracture mechanics* is becoming a powerful
tool of materials engineering. A self-contained presentation of the history, current state
of the art, and future goals of damage mechanics can be found in [98].
Damage is in the constitutive equations represented by one or more internal variables
o;n, which are in classical damage mechanics related to a material point of the body (the
quantity un corresponds to the microcrack density on an elementary area with normal
n,o; n €<0,l».
The need to take into account the interaction between two material points has led to
the concept of nonlocal continuum (cf. [21]). Strains entering the constitutive relations
are averaged with a certain weight over a domain, the size of which is a material constant.
Averaging ensures that the size of the localization zone cannot decrease below a certain
minimum size dictated by the structure of the material.
This property has important implications for the FEM. For the local model, the size
of the localization zone is determined by the size of the element, which represents the
smallest macrocomponent of the system. If the size of the localization zone was not
limited by an appropriate tool, the width of the localization zone would tend to zero as
the mesh is refined. Consequently, the total energy dissipation in the localization zone
would tend to zero as well. This is unacceptable from the physical point of view because
damage must result in nonzero energy dissipation. From the numerical point of view, local
models are characterized by spurious sensitivity of the results to the mesh size. Moreover,
the shape of the mesh causes certain directions (along the element sides or diagonals) to
be preferred as directions of localization bands.
where p is the mass density, E is the Young's modulus, e is the elastic strain and LJ is the
damage parameter. It is obvious that for u = 0 we get the strain energy density of an
elastic material, W = (1/2) Ee2.
The expression for stress is obtained by taking the the derivative of (1.80) with respect
to e [cf. equation (1.3)],
4
Fracture mechanics will be presented separately in Section 7.4.
1.4. DAMAGE THEORY 25
The generalized thermodynamic force associated with the internal variable LJ is the damage
energy release rate [98]
It remains to set up the evolution equation defining the dependence of a; on a; and e. For
general stress states it is derived from the assumption that the damage rate is orthogonal
to an experimentally determined damage surface. The general form of the evolution
equation for uniaxial stress is u = f(e,w). For concrete, a simple rule recommended
in [21] reads
in which the local damage threshold Y\ has different values for tension and for compression,
and b and n are positive material constants. The graph of g is schematically shown in
Fig. 1.7.
Damage grows only during loading. During unloading, and during reloading after
previous unloading, we have u — 0, and so the loading criterion can be written in terms
of the loading function F = F(u) as follows:
Note that both conditions are consistent with the dissipation inequality (1.83). The
loading function usually assumes the form
The initial value of the softening parameter k is zero, and it progressively grows such that
it always equals the maximum value of uj ever reached before the current state.
By differentiating equation (1.81) we get
where
According to (1.87), the first term in (1.89) includes the damage accumulated during the
previous deformation history. The second term appears only during loading (Fig. 1.8a).
The evolution function A depends on the form of the evolution equation (1.84).
The constitutive relation for strain softening can be best explained by decomposing
the stress increment according to Fig. 1.8b. Using the notation E t = Dw < 0, we can
from which
where DI and £>2 are the material stiffness matrices, the meaning of which is clear from
the schematic plot in Fig. 1.8a. For general stress states it is further necessary to account
for the kinematic aspect of microcrack growth, i.e., to consider the changes of microcrack
shape and orientation induced by deformation of the material. The final form of the
constitutive equation for the brittle damage model is
a) The more recent version of the model assumes that the cracks rotate such that their
direction always coincides with the normal to the direction of maximum principal stress
07. This implies that no shear stresses exist on the crack faces. Based on this assumption,
Bazant and Lin [22] derived a constitutive equation that in tensor notation reads
which is the inverse of (1.81). Assuming that the increase of strain due to dam-
age takes place in the direction of the maximum principal stress 07, we can write
£jjr = (crf/E')uj/(l — u). The tensor expression (1.95) is obtained by expressing
£ij,fr = niUjSijr, a/ = njferijo^, where Hi are the components of a unit vector in the
direction of 07. Elastic strain is in (1.95) taken into account by the elastic compliance
tensor Cijki given by (1.35). E' is set equal to E for plane stress while for plane strain
and for the general triaxial case it equals E/(l - z/2). It is easy to rewrite (1.95) in the
matrix notation.
b) In the classical model, microcracks start forming in the direction orthogonal to the
maximum principal stress 07 as soon as the strength Rt is reached. It is assumed that
the orientation of the cracks does not change, and so shear stresses are generated on the
crack faces. The initial shear modulus G is reduced to rs G where, according to Kolmarov
fcf. [471V
Here, en is the normal strain orthogonal to the crack direction (crack opening), and GI and
C2 are constants dependent on the reinforcement ratio. This variant typically uses a simple
constitutive relation for an orthotropic material (orthotropy is induced by cracking; see
Section 1.2.1). Further details can be found in [47].
where
This description is valid before the peak of the diagram in Fig. L8a, and it would remain
valid after the peak if localization did not take place. Localization is a bifurcation from
the homogeneous strain state into a nonhomogeneous one. The jump from the former
state into the latter one is triggered by an imperfection in the stress field [produced, e.g.,
by an added equilibrated set of forces ±8F (Fig. 1.9b), or by a reduction of the sectional
area in the central part by 6A] slightly before the stress reaches the peak value a = Rt. In
the central part of the bar is softening and the corresponding strain grows. On the other
hand, strain in the remaining parts decreases and the material is elastically unloaded.
Stress along the entire bar must be uniform due to the condition of equilibrium. The
nonhomogeneous strain field shown in Fig. 1.9b is described by
where
yields
implies
i.e., by
A critical state occurs for KI + Kn = 0. This state is followed by the so-called
snapback. The relation between the force F = a A and the imposead displacement 2 is u
shown in Fig. 1.12. The dash-dot line corresponds to the homogeneous deformation from
Fig.1.8 behind the bifurcation point (B.B.). Further details will be given in Chapter 9.
We will now generalize the preceding conclusions related to the material stability in
conjunction with the strain-softening constitutive relations within the classical continuum
[34].
A material is stable if [74]
This condition coincides with the singularity of the symmetric part of £>,
For a symmetric tangent stiffness matrix D = DT, the loss of material stability coincides
with the limit point where dcr = Ddcr = 0 and with the loss of uniqueness. For a
nonsymmetric tangent stiffness matrix, the loss of material stability can be encountered
prior to the limit point and loss of uniqueness [176],
The incremental material stiffness matrix D can be derived based on a classical soften-
ing plasticity model resulting into expression (1.66). In the post-peak regime the "harden-
ing" modulus becomes nonpositive, (-E) < H < 0. Combining formulae (1.66), (1.107),
1.4. DAMAGE THEORY 31
It can be shown that for associated plasticity (/ = g) the loss of material stability is
possible when the hardening modulus H is zero or negative. For a nonassociated law
(g ^ /) the loss of material stability may occur even if H > 0 (see [148]).
In order to investigate strain localization, we have to admit a discontinuity of the
deformation gradient across a plane with normal n«, while the continuity of displacements
and the equilibrium conditions are preserved pointwise. Under these assumptions, the
jump of the displacement gradient has the tensorial form [148]
Substituting equation (1.107) into (1.113), assuming that the materials on both sides of
the discontinuity surface are the same, and exploiting the (minor) symmetry property
Dijki = Dijik, we arrive at the condition
where Qjk = niDijkiHi is the so-called acoustic tensor. This equation has a nontrivial
solution if
The singularity of Qjk and formation of a discontinuity correspond to the loss of ellipticity
of incremental (rate) equilibrium equations. This phenomenon may be encountered within
all the classical continuum models including softening plasticity, continuous damage and
smeared cracking.
In the presence of strain softening and localization the governing partial differential
equations must be regularized in order to restore ellipticity, guaranteeing the continuity
of the solution. Regularization introduces an internal length scale that defines the size of
the localization zone.
There exist two phenomenological approaches ensuring well-posedness. They are ref-
ered to as discontinuum modeling and enhanced continuum modeling.
where the stiffness matrix Dcr reflects the softening behavior and may be related to the
energy released during fracture. In the fundamental Mode I crack opening the fracture
energy required to open a unit crack area is expressed as
On the other hand, it is convenient to remain within the continuum description with the
classical quantities of stress and strain. Calculations based on the standard continuum
model are in a good agreement with experiments if the width of the localization zone /
corresponds to a certain minimum width w given by the structure of the material. Bazant
and Wu [21] recommend setting wc = 3da (da is the maximum aggregate size in concrete),
or wc = 5dg (dg is the grain size in rock).
Without any loss of generality let us consider a bar of a unit cross-sectional area
under tension. The energy dissipated in densely spaced microcracks is given by the area
under the descending branch of the stress-strain diagram with softening, multiplied by
the localization width / = wc. This value increased by the energy released by the material
between the cracks gives the fracture energy Fig. 1.13,
Energy Qf related to a unit area of the fracture surface is a material constant. Equation
(1.118) is applicable even in more general loading cases such as the one shown in Fig. 1.14.
Figure 1.14: Body with a localization zone (divided into finite elements)
If we know the energy Qf needed for fracture, we can formulate the Irwin-Orowan Q-
criterion of a "driving force." If this criterion is satisfied, the fracture process zone grows
by Aa (Fig. 1.15 shows one half of the specimen with a central softening zone). The
1.4. DAMAGE THEORY 33
driving force is the energy released by the body as the process zone grows, expressed per
unit area of the fracture surface. We can express this energy as the decrease of the total
potential energy -AH/fcAa = - [II (a 4- A a) - H(a)] /6Aa > 0, where b is the thickness
of the specimen. More information on potential energy follows in Section 1.6. Here we
only recall that II = Ei -f Ee is the sum of the strain energy Ei and the external work Ee.
In the limit as Aa ->• 0 we get
It is usually assumed that the edges of the specimen are fixed, and so the external
work is zero (cf. [21]) and
where
for plane stress,
for plane strain,
and AAint = Aa wc + 2/ca Aa is the increment of the area of the stress-relief zone charac-
terized by an empirical parameter k.
The increment of area is vertically hatched in Fig. 1.15. The horizontally hatched area
including the area of the process zone awc is related to the energy released as the process
zone grows to length a.
Combining (1.119) to (1.121) we find
34 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
where
is the critical value of the stress intensity factor for Mode I fracture. The critical value
determined at plane strain conditions is called the fracture toughness.
It remains to be noted that a similar size effect to the one shown in Fig. 1.16 appears
when the crack tip is blunted by a small plastic region (Fig. 1.17).
The Dugdale-Barenblatt model often used for metals is based on the assumption that
the plastic region has the shape of a small triangle (Fig. 1.17a) with a homogeneous stress
1.4. DAMAGE THEORY 35
state at the yield limit Ry. For a central crack of length 2a (Fig. 1.15) with wc = 0, and
assuming that an = GOO -C Ry, the crack opening is approximately given by
The opening 8 is usually taken as a fracture parameter. It is assumed that the crack
loses stability if the crack opening reaches a critical value 6C. The Crack Opening Dis-
placement (COD) criterion can formally be written as
1. Nonlocal continuum.
Nonlocal (integral) localization limiters replace some quantities by their weighted
averages taken over a certain neighborhood of the material point under consider-
ation. Differential localization limiters add terms with higher-order derivatives of
strain (or stress) into the constitutive equations.
2. Gradient-dependent softening plasticity theory.
3. Cosserat (micropolar) continuum.
4. Rate-dependent localization limiters include time derivatives.
Localization limiters
a) Nonlocal (integral) localization limiters
The constitutive equation for a strain-softening material (cf. [21]) will be written as
where
Fig. 1.18 shows examples of two weight functions giving the same effective length
Xr(x) = A for points x whose distance from the boundary is at least pQ\. In the boundary
regions the length A r varies. Setting r = \x — s\ and introducing a material constant A,
we can write
The bell function (and also the Gauss normal distribution function) satisfies certain
complementary conditions that rule out the danger of instability due to averaging while
the piecewise constant function does not meet these conditions.
As recommended by Bazant and Wu [21], a numerically convenient nonlocal damage
model is obtained if the damage parameter u is replaced by its nonlocal average while the
elastic strain remains local. In relations (1.81) to (1.86), u is replaced by
where r = s - x, assuming that Xr(x) = A, and taking into account the symmetry of a
with respect to r, we get from (1.130) the relation
where
5
The increments of stress, strain and damage are denoted by A to avoid confusion with the differential
of the coordinate x.
1.4. DAMAGE THEORY 37
Expression (1.135) can be substituted into the incremental equation [see (1.90)]
where
Both models lead to a relatively simple evaluation of the strain profile in the local-
ization zone. Model a) gives a continuous profile (Fig. 1.19a) while model b) produces a
discontinuity of strain at the boundary between the two zones (Fig. 1.19b). It is assumed
that the initial homogeneous strain sr0 is given, and the corresponding initial damage OJQ
can then be found from (1.84).
Numerical calculations have shown that the length of the zone / is directly proportional
to the averaging length A. This implies that localization in a local continuum (A -» 0)
takes place in an infinitely small region.
In order to make use of the foregoing ideas within the gradient-dependent softening
plasticity theory in 3D formulation, we now write the assumed yield condition (1.42) in
the enhanced form (see [130])
or
Superimposed dots denote the derivatives with respect to time. Equation (1.146) describes
the evolution of the plastic process in the strong sense. It is a nonhomogeneous partial
differential equation of Helmholtz type, an analytical solution of which has not been
derived yet [130]. Hence a weak form of this equation based on a variational principle
seems to be very useful (see Section 1.9).
The compliance function of a linear viscoelastic material represents the strain at time t
due to a unit stress o — 1 applied at time r and kept constant. This function is often
written as
The inverse relation to (1.147) can also be written using the principle of superposition
as
The relaxation function R has the meaning of stress at time t due to a unit strain e—e° = 1
applied at time r and kept constant afterward.
If the stress a is continuous in time, we can set da(r) = [d(j(r)/dr\ dr and transform
the Stieltjes integral in (1.147) into a Riemann integral. Equation (1.147) is then differ-
entiate, and the integral constitutive relations can be transformed into differential ones.
This step requires a suitable representation of the kernels J(t,r) and R(t,r). From the
numerical point of view, the most convenient choice are the degenerate kernels (cf. [20])
where y^t) = (t/Q^)^. The coefficient q^ < 1 is introduced in order to reduce the number
of terms in the expansions, and for concrete it is usually set to q^ « 2/3. Functions D^
and EH can be evaluated by the method of least squares. Explicit expressions for D^ (see
[24]) are available for the specific functions J frequently used for concrete (double-power
law, or logarithmic law).
Retardation times 0^ for J (or relaxation times for R) must satisfy certain rules
necessary for the success of the calculation: 0 X is set to a small value (in the order
of 10~9 days), so that the first term of the Dirichlet series (1.152) is sufficiently close
to I/DI(T), where DI(T) = E(r), and it thus realistically expresses the instantaneous
compliance. The following terms should be uniformly distributed on a logarithmic time
scale, i.e., 0M = lO1/^©^, for p = 2,3,...,M. Finally, 0M > tmax/2, where tmax is
an upper bound on the time interval in which we intend to analyze the response of the
structure. If £mt-n is the lower bound on this interval, we have to check the condition
02 < 3tmin.
Once the degenerated kernels J and R are specified, we can substitute them into
equations (1.147) and (1.151). Differentiating equation (1.151) with respect to t we get
according to [20] the expression
If we set y^(t) = •#/*(£)/77/i(£), the preceding two relations describe the Maxwell chain
in Fig. 1.22a. The chain models a solid body (i.e., not a fluid) only if the function J has
a finite limit as t -4 r -> oo. This requirement is obviously satisfied if 0M -» oo, i.e.,
1.5. VISCOPLASTIC MATERIALS 41
y -» 0 and r]M ->> oo. In Fig. 1.22a the limit case corresponds to removing the damper in
the last element because the strain in this damper is zero.
Equation (1.147) can be processed in a similar fashion. Substituting from (1.152),
taking the second derivative with respect to t and introducing model variables r]^(t) =
DM)/yM), EM) = DM) - DM)/yM), we arrive at equations (cf. [20])
describing the Kelvin- Voigt chain in Fig. 1.22b. In the limit case (0i —> 0, y\ —> oo, 771 -»
0) the damper of the first element does not carry any stress, and so it must be removed.
the stress c7 M (^_x) for each element of the chain. Equation (1.155), which is simpler than
(1.157), can be numerically integrated assuming that in this interval y^ = const., e-e° =
const., and E^=^ = E^fa - A^/2). The result is
Denoting
42 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
After substituting the last result into (1.154) we get the final expression for the incremental
constitutive relation,
where
We did not mark explicitly that the coefficients after the summation symbol are evaluated
for the interval < ti-l.ti >. The sequence of these coefficients can be computed in
advance.
If we had to take into account also the strain component due to damage, £/r, we would
replace the matrix C in formula (1.160) by the matrix C-f N(l - i/2)o;/(l -a;), where N
is the matrix representation of the tensor Nijki = n^njUkHi introduced in equation (1.95).
Equation (1.151) can be generalized in a similar way. The relaxation function R
defined by (1.153) is replaced by the matrix kernel (for u = 0)
Finally, the incremental constitutive equation based on Dirichlet series (1.158) will be
written as
where the stiffness Ei for the z'th interval < tt_i,£, > is determined by (1.159), and
The left-hand side represents the virtual work of internal forces (stresses) while the right-
hand side corresponds to the virtual work of external forces.
Virtual fields 6e and Su must be kinematically admissible. This means that
• virtual displacements 6u must satisfy the kinematic boundary conditions,
44 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
• and virtual strains 6e must be linked to the virtual displacements by the strain-
displacement relations [see (1.2)],
This equation is satisfied for arbitrary virtual displacements 8u only if the conditions
of equilibrium hold, i.e., if
• Cauchy equations (1.1) are satisfied in ft, and
• static boundary conditions (1.6) are satisfied on Tp.
Equations (1.1) and (1.6) follow from the PVD, which is thus the
general principle of equilibrium.
The PVD can easily be extended to dynamic problems. According to the d'Alembert
principle, we can treat the inertia forces, QU, as externally applied body forces (u denotes
the second partial derivative with respect to time; g is the mass density). Equation (1.168)
is then transformed to
The left-hand side represents the complementary virtual work of internal forces while
the right-hand side corresponds to the complementary virtual work of external forces.
Virtual fields 8a , 8X and 8p must be statically admissible. For 8X = O in ft and
8p = O on Tp, the equilibrium conditions include
• homogeneous Cauchy equations,
This equation is satisfied for arbitrary virtual stresses 8cr(5p — n5<r ^ O on Tu ) only
if the kinematic equations hold,6 i.e.,
• kinematic relations (1.2) are satisfied in H, and
• kinematic boundary conditions (1.7) are satisfied on Fw.
Equations (1.2) and (1.7) follow from the PVF, which is thus the
principle of continuity.
6
Strictly speaking, the virtual stresses are not independent because they must satisfy three Cauchy
equations (1.170). The principle (1.172) is thus equivalent only to three independent equations—the
conditions of compatibility.
1.6. PRINCIPLE OF VIRTUAL WORK AND VARIATIONAL PRINCIPLES 45
45
• The principle of virtual forces (PVF) leads to the Castigliano variational principle.
The Lagrange principle can be formulated as the principle of minimum potential energy:
Among all kinematically admissible states of an elastic body, the actual state minimizes
the total potential energy,
where
• displacements that are continuous and have piecewise continuous derivatives in the
solution domain and that satisfy the kinematic boundary conditions on Fu, and
• strains that are derived from the displacements using the kinematic (strain-
displacement) equations.
and taking into account the first form of the constitutive equations (1.3), we obtain
equation (1.165). This leads to the following important conclusion:
Both the PVD and the principle of minimum potential energy lead to the same equa-
tions, namely the Cauchy equations and the static boundary conditions.
Supplementing (1.176) by the virtual work of inertia forces, the d'Alembert principle
assumes the form
Postulating the assumption that the displacement variations 6u at the initial and final
time instants ti and t^ are equal to zero, we can eliminate the integral in the brackets.
The second term on the right-hand side of (1.179) is the variation of the kinetic energy
K,
where
A statically admissible state of stress must satisfy the equilibrium conditions inside the
body (nonhomogeneous Cauchy equations) and on a part of its boundary (static boundary
conditions on Tp).
Setting the variation of the functional II* equal to zero,
and taking into account the second form of the constitutive equations (1.3), we obtain
equation (1.169). This leads to the following important conclusion:
1.6. PRINCIPLE OF VIRTUAL WORK AND VARIATIONAL PRINCIPLES 47
Both the PVF and the principle of minimum complementary energy lead to the same
equations, namely the strain-displacement equations (or, after the elimination of the dis-
placements, the compatibility equations) and the kinematic boundary conditions.
The PVD and the Lagrange variational principle provide a basis for the
displacement methods in structural analysis. Similarly, the PVF and the Castigliano
variational principle provide a basis for the force methods.
Mixed methods, which deal at the same time with displacements u and stresses a,
can be based on the Hellinger-Reissner principle, which is mathematically written as
where
The dual formulation of the H-R principle, which can be derived using (1.9), assumes the
form
where
The fields u and cr are independent. Equations (1.187) and (1.189) lead to the fol-
lowing stationarity conditions (the H-R principle assumes that the constitutive equations
are satisfied a priori):
• Cauchy equations and stress-displacement relations, and
• static boundary conditions on Tp and kinematic boundary conditions on Tu.
The general Hu-Washizu variational principle does not require any a priori assump-
tions, i.e., the unknown fields u, <r and e are totally independent. The principle is written
as
where
Taking the variation of the functional (1.192) and transforming it using the Clapeyron
theorem (1.9), we can express the condition (1.191) as
48
48 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
The stationarity conditions following from the first line in (1.193) can be identified as
equations (1.1) to (1.3) while those following from the second line are the boundary
conditions (1.6) and (1.7).
As when deriving the Hamilton principle, we can enrich (1.193) by the virtual work
of inertia forces and derive the equation
where
The stationarity conditions are the same as for the static problem, only the equations of
equilibrium (1.1) are augmented by the effect of inertia forces:
The general variational principle is applied in FEM when developing special types of
elements (e.g., elements with an embedded localization zone), in the probabilistic version
of FEM, and in the symmetric formulation of BEM.
or, equivalently,
The functional TIR can be transformed using relation (1.9) in which we set $7 =
ft(1> + ft(2) and T = r*1) + r<2) + r<12>. This leads to the functional
2. In the second (dual) version, we write the stress continuity condition (1.198) as
The new variable represents the interface tractions related to the interface boundary
of element H^. Transformations analogous to the previous case lead to the modified
functional
Exploiting relations (1.197) to (1.199) and (1.202), we can verify that the functionals
(1.200) and (1.204), or (1.201) and (1.203), are equivalent. In addition to the displace-
ments -u(1) and n(2) and stresses tr(1) and a(2), the multipliers A and /x are also subject
to variation.
The modified Hellinger-Reissner principle provides the basis for the so-called mixed
models in structural analysis.
The modified Lagrange principle can be derived from (1.204). Assuming that only the
kinematic relations (1.2) are satisfied a priori, and incorporating the kinematic boundary
conditions (1.7) into the variational principle, we arrive at the functional
50 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
In addition to the stresses cr(1) and <7(2), also the multipliers /x = u^ = n(2) on T(12^ and
u on Tp are subject to variation. The modified Castigliano principle provides the basis
for the so-called equilibrated hybrid models in structural analysis.
The relationship between the variational principles and the corresponding computa-
tional models is clear from Table 1.2. The compatible model is based on a displacement
PV displacements PV forces
Lagrange v. p. Castigliano v.p. compatible equilibrated
Modified
Hellinger-Reissner mixed
principle
Functions ^0,^0 and WQ are constructed such that they satisfy the prescribed kinematic
boundary conditions (1.7), i.e.,
Functions (pk,^k,Xk are linearly independent and satisfy the homogeneous kinematic
boundary conditions, i.e.,
Basic unknowns in this approximate solution are the coefficients of linear combination.
In the matrix form, (1.207) can be written as
The above form of the constitutive equations is not needed when expressing the total
potential energy but it will be exploited when evaluating the stress vector a after the
calculation of the displacements from the Lagrange principle. Substituting approximation
(1.209) into (1.10) we obtain
Using (1.210) and (1.211), the functional (1.173) can further be expanded as
52 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
Applying this condition to the above expression for II we obtain a set of linear algebraic
equations,
This equation represents the generalized equilibrium conditions of the discretized sys-
tem. Individual symbols have the following meaning:
The extension to dynamic problems is easy. After adding the inertia forces to the
second integral in (1.214) we get
This leads to the generalized conditions of dynamic equilibrium of the discretized system,
where
entire solution domain, and so it is quite difficult, or even impossible, to construct them
for domains of a complicated shape, or for problems with complicated boundary conditions
(supports). On the other hand, the FEM deals with very simple shape functions, each of
which is nonzero only in a small neighborhood of the corresponding node; cf. Fig. 1.25.
In order to increase the accuracy of the Ritz method, we have to add additional linearly
independent base functions. In FEM we proceed in a similar manner. The shape functions
are added simply by dividing the solution domain into more elements. The same domain
is covered by a larger number of smaller "hills," which corresponds to an increase of the
number of (global) shape functions.
The function a is selected such that it satisfies the Cauchy equations (1.1) and the static
boundary conditions (1.6) prescribed on Tp, The functions o"i are linearly independent
and they satisfy the homogeneous Cauchy equations (1.1) and the homogeneous static
boundary conditions (1.6). The coefficients of linear combination $ are to be found. The
above approximation can symbolically be written as
Using relation (1.13) we construct the expression for the complementary energy density
Similarly to the Lagrange principle we see that, after the integration, II* becomes a
function of the coefficients /3 yet to be determined. The necessary condition of a minimum
has the form
54
54 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
Substituting the expression for II*, and expressing p according to (1.220) combined with
(1.6), we obtain
where
• the parameters of the mesh (number of elements and approximation of the bound-
ary), and
• the type of approximation of the unknown functions across the element.
An accurate solution can be obtained only if the approximation functions satisfy the
convergence criteria expressed by the conditions of
• continuity and
• completeness.
the requirements for interpolation functions depend on the order of governing differential
equations.
Completeness will be defined with regard to the solution of elasticity problems. The
approximation functions satisfy the condition of completeness if
a) they can represent the displacement of the element as a rigid body, i.e., with zero
strain, and
b) they can represent a state of constant strain. This condition is easy to understand
if we consider a series of meshes with the element size decreasing to zero. If the strain
state of every element is constant, it is possible to approximate any global strain state
with an arbitrary accuracy.
If the approximation functions satisfy the conditions of continuity and completeness,
we call the corresponding element conforming. The convergence to the exact solution is in
such a case monotonic. Of course, these conditions do not ensure stress continuity across
the interelement boundaries, nor do they imply a certain convergence rate. The rate of
convergence depends on the degree of the approximation polynomials. It is desirable to
use complete polynomials of a given degree.
In some cases, the continuity condition is not satisfied but the solution still converges
to the exact solution of the idealized structure. The convergence is in such cases non-
monotonic. A number of elements violating the continuity condition have been proposed,
especially for the solution of plates and shells. At any rate, the element must always
satisfy the condition of completeness. If the approximation functions are continuous but
not complete, the solution converges (sometimes even monotonically) to a wrong result.
For nonconforming elements satisfying the condition of completeness but not that of
continuity, it is necessary to check that the condition of completeness is satisfied by an
assembly of elements. This is usually done by the patch test, which analyzes a simple
patch consisting of a few elements. The patch is subjected to given loads or prescribed
displacements that correspond to a state of constant strain. The element passes the patch
test if the numerical results exactly agree with the theoretical ones. Details are given in
Chapter 3.
strain and stress fields, respectively. The stresses <r° and small strains e° are related by
the linear homogeneous isotropic Hooke's law,
In the second step, a geometrically identical body with the same prescribed surface dis-
placements is considered, which is anisotropic and nonhomogenous. Displacements u,
strains e and stresses cr are unknown, and generalized Hooke's law (1.12) including the
eigenstresses A = {Ax, A y , A 2 , Ay2, Xzx, Xxy}T can be written as
Similarly to the classical Hashin-Shtrikman theorem, let us define the symmetric stress
polarization tensor r by
Also define
and
Our aim is to obtain an appropriate relation between strains and eigenstresses and the
variational principle describing the behavior of the nonhomogenous and anisotropic body
underloading by eigenstresses and prescribed boundary displacements. As both fields a
and cr0 are statically admissible, the following equations have to be satisfied in the sense
of distributions:
Setting
1.8. VARIATIONAL PRINCIPLES IN ANISOTROPIC AND NONHOMOGENEOUS ELASTICITY 57
we find that equation (1.232) is one of the stationarity conditions of ET while the second
condition, equation (1.231), yields after recasting the remaining terms in the brackets.
Finally, it can be proved that the stationary value II* of the functional Ur is equal to
the actual potential energy stored in the anisotropic and heterogeneous body,
and
is the vector of eigenstrains. The functional II attains its maximum (82U < 0) if [D] is
positive definite and its minimum if [D] is negative definite.
where C° = (D0)-1.
In the second step, a geometrically identical body with the same prescribed surface
tractions is considered, which is anisotropic and nonhomogenous. The strains and stresses
are unknown and are denoted by e, cr. The generalized Hooke's law including the eigen-
strains n can be written as
where C = D~l.
Define the symmetric strain polarization tensor 7 by
58 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
Further define
and
Our aim is to obtain an appropriate relation between stresses and eigenstrains and the
variational principle describing the behavior of the nonhomogenous and anisotropic body
under loading by eigenstrains and prescribed surface tractions. As both fields e and e°
are kinematically admissible, the following equations have to be satisfied in the sense of
distributions (for convenience, a tensorial notation is used for the compatibility equation):
It can be proved that the stationary value II*5 of the functional II* is equal to the
actual complementary energy stored in the anisotropic and heterogeneous body,
The functional attains its absolute maximum (82l~l* < 0) if [C] is positive definite, and
into equation (1.255) we obtain the weak equilibrium condition for the elastoplastic body
The body contains two regions —an elastic region Oe where A = 0, / < 0, and a plastic
region A > 0, / = 0 (Fig. 1.28).
After substitution of equation (1.256) into (1.258) and integration of the term con-
taining (7V2A) we arrive by applying the Green formula at
where nep defines the outward normal to the boundary Tep of the plastic region tip.
The boundary integral vanishes if
60 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
represent the fact that the internal forces7 F are in equilibrium with the external forces
R. If the behavior of the system is nonlinear, the internal forces are nonlinear functions
of displacements F = F(ri,r 2 , ...) = F(r). A small increment of the external forces by
dR results into an increment of the internal forces dFi = ^(dFi/dr^drj, i.e.,8
where Kt = (d FT/dr)T is the tangential stiffness matrix. Its elements are determined
by the formula
The term on the right-hand side represents the out-of-balance forces, i.e., the difference
between the external forces dR+R and the actual internal forces F' , which is due to the
absence of the higher-order terms. The numerical solution reduces the unbalanced forces
by equilibrium iteration. Details are given in Section 9.4.
Let us explore the conditions of stability at a point A on the fundamental path OAB
(Fig. 1.29). Assuming that the external forces R are fixed, we consider a perturbation
of the system into a neighboring state induced by small displacements Sr. The potential
7
By "internal forces" we mean nodal forces equivalent to stresses.
8
The derivative of a scalar function / with respect to the vector r is constructed by applying the column
operator (d/dr) on this function. The result is a column matrix {df/dr} = {df/dri,...ydf/drn}T. A
similar approach leads to the derivative of the vector FT with respect to the vector r.
1.10. NONLINEAR SYSTEMS AND STABILITY CRITERIA 61
energy II changes by
The above result can be generalized based on the first and the second laws of thermo-
dynamics. A systematic explanation has been given by Bazant and Wu in [21]. We will
only summarize the main results.
According to the first law of thermodynamics (conservation of energy) the increment
of internal energy, which will be denoted here by U instead of E{, is equal to the sum of
the heat AQ accepted by the body from the environment, and of the increment of work
done by the external forces:
where T is the absolute temperature. The first term depends on the influx of heat into
the body while A5Z- is the internal entropy increment. The second law of thermodynamics
states that a process for which AS; < 0 cannot occur, a process for which A5,- = 0 can
occur (reversible process preserving thermodynamic equilibrium), and a process for which
A5, > 0 must occur (irreversible process). Strictly speaking, any real deformation process
dissipates energy (is irreversible).
62 CHAPTER 1. BASIC NOTIONS, EQUATIONS AND PRINCIPLES
According to the above conclusions, a state of equilibrium is stable if the system cannot
change its state by itself. The criterion (1.268) then assumes a more general form. A state
of equilibrium is
The stability criterion is usually expressed in terms of state functions. The fundamen-
tal state function is the internal energy, the increment of which is according to (1.269)
and (1.270) given by
The above formula is suitable if AS = 0, which in the reversible case (AS* = 0) corre-
sponds to an adiabatic process AQ = TAS = 0 (for very fast deformation processes the
change of energy due to heat transfer can be neglected).
The so-called Legendre transform
transforms the internal energy into the Helmholtz free energy, the increment of which is
This state function is useful for isothermal processes (for very slow deformations the
temperature remains almost uniform).
A generalization of the potential energy H leads to the state functions J- or IA, for
which
The parameter v — dr/(drTdr)1/2 indicates that the state functions U and F, as well as
the tangential stiffness matrices Ks and KT, are in general path-dependent. For some
materials, e.g., for an elastoplastic material with strain hardening, the matrices K can be
independent of v as long as the vector v stays within a certain sector (e.g., cone) of the
multidimensional space with coordinates dri,dr 2 ,.... For the sector shown in Fig. 1.30,
the matrices K are independent of the direction v on paths 1 and 2.
The thermodynamic criterion of stability can be interpreted as follows:
If the structure becomes unstable, the energy TAS; > 0 sets it into motion (kinetic
energy). For dissipative processes such as viscosity, plasticity, internal friction, damage
and fracture, this energy irreversibly changes into heat. If the structure is in a stable
state of equilibrium, the energy — TA5j > 0 represents the work of external forces that
must be supplied in order to disturb the equilibrium.
2.10. NONLINEAR SYSTEMS AND STABILITY CRITERIA 63
In this chapter, we will explain the application of the two basic forms of the principle of
virtual work (or, alternatively, of the corresponding variational principles) to the analysis
of skeletal structures such as trusses, beams or grillages. In structural analysis, energy
functionals are usually expressed in terms of the internal forces rather than stresses.
Therefore, we will first review the most important relations governing the distribution of
internal forces. Practical applications often deal with beams supported (along their entire
length) by a continuous elastic foundation. As an example of a simple yet sufficiently
realistic model, we will study the Winkler-Pasternak model of an elastic foundation.
Based on the conditions of equivalence, the internal forces can be expressed in terms
of the stresses as
As shown in Fig. 2.2, shear stresses cause warping of the cross section. With the exception
of points A and B, at which rxz = 0 and thus the shear strain jxz = rxz/G = 0, the tangent
64
2.1. BASIC RELATIONS FOR BEAMS 65
to the deformed cross section deviates from the normal to the deformed beam axis (called
the theoretical normal).
A practical method for calculating the shear distortion is based on the assumption of
a constant shear stress across the section (bottom part of Fig. 2.2).
The average shear stress r corresponds to the average shear strain 7 given by the angle
between the theoretical normal and the pseudonormal AB' . The constitutive equation
relating 7 and r can be written as
The constant k can be determined from the condition that the work of the actual stresses
(we omit the multiplier 1/2)
in terms of the displacements and rotations. We have used the usual assumption that the
vertical displacements do not vary along the height of the beam and thus w(x, z) = w(x).
Combining (2.5) and (2.6) with the constitutive relations, we get
66 CHAPTER 2. SKELETAL STRUCTURES
Taking into account that the y-axis passes through the centroid Cg, (2.1) yields after
integration
where
The forces denoted by a bar would exist in a structure prevented from deforming
(dus/dx = 0 , d(py/dx = 0). Note that £0 and T must be linear functions of z in order to
be consistent with the assumption that the cross sections remain planar.
The assumption that the effect of shear on deformation is negligible,
results in the following relation between the rotation and lateral deflection:
2.1. BASIC RELATIONS FOR BEAMS 67
Equation (2.12) corresponds to Bernoulli's (Kirchhoff's) assumption that the cross sections
remain perpendicular to the deformed beam axis. The second formula from (2.8) then
takes the form
The internal forces must satisfy three equilibrium conditions for an infinitesimal beam
element (cf. [164])
Using the third equation, the shear force Qz can be eliminated from the second one, which
gives
The foregoing relations will be used in analysis of frame structures. For grillages,
torsion must be taken into account in addition to bending. The torque Mx is proportional
to the relative angle of twist d(px/dx as follows:
where Glk is the torsional stiffness of the cross section in free torsion. For sections whose
shape at least approaches a circle, Ik can be estimated as
where Ip is the polar moment of inertia. For more complicated cross sections, Ik can be
calculated using FEM as described in Section 7.1.
The differential equation of equilibrium for the torque reads
where w = w(x,y,0).
Taking into account the assumptions (2.19), we can write the formulae for stresses
where
The integration domain H is shown in Fig. 2.3. The new material parameters are defined
in terms of the input parameters E and v by
2.1. BA SIC RELATIONS FOR BEA MS 69
Detailed analysis of the layer (see [105]) shows that the function ^ depends on the
input constants E and z/ and on the ratio b/h (Fig. 2.3). This dependence is graphically
presented in Fig. 2.4, which can be used for easy determination of the constants C\ and
<?2.
The expression (2.25) shows that the virtual work of the shear stresses TXZ a ryz, which
vary across the thickness of the layer, can be represented by the work done by two newly
defined equivalent shear forces
it down. Looking only at the separated part, /0 can be thought of as an external force
doing virtual work (per unit width)
By solving equation (2.28), we get the shape of the shear depression shown in Fig. 2.5.
The solution reads
Formula (2.30) is very important for practical applications as it gives an easy estimate of
the length Le of the shear depression. The subscript e denotes a sufficiently small number,
2.1. BASIC RELATIONS FOR BE A MS 71
defining the point at which the surface displacement is negligible (see Fig. 2.5). Solving
the equation
we get
The foregoing results will now be generalized. Consider a segment of the subgrade
below the beam of length / (Fig. 2.6). The foundation beam exerts a distributed load —fz
(fz is the load acting on the beam; see Fig. 2.1), distributed moment loads —mx, —my, and
end forces [—Qz, —Mx, — My]^o . The associated quantities describing the displacement
of the contact surface are WQ, (px and (py = —dwQ/dx. Applying once again the principle
of virtual work, we can eliminate the dependence on the variable y. To this end, let us
express the vertical displacement w — w(x,y) at an arbitrary point using formula (2.30):
or
The last term in the integral can be integrated by parts. This easy step does not need to
be written down explicitly. The equation is satisfied for arbitrary independent variations
6wo,8tpx,6ipy if the coefficients at these variations vanish, and thus
where
Note that the width of the foundation is denoted as 26. The expression in the brackets in
(2.37) is the reaction of the elastic foundation. A similar modification can be performed
for the differential equation (2.18):
where
The expression in the brackets is again the reaction of the elastic foundation.
2.2. TRUSS AND BEAM ELEMENTS 73
equilibrium, assuming that the distributed loads fx and fz are uniform. Among six gen
eralized end forces, three quantities (the axial force fa and the moments fa, ^3, collecte
to a vector (3 = {fa, fa, /3^}T) are independent and the remaining ones are determined b
three equilibrium conditions of the member.
The equilibrated internal forces are expressed by the following formulae (x = £/):
Normal force
bending moment
The deformations corresponding to the internal forces can be expressed from formulae
(2.8). We will deal with the initial deformations £0 that are caused by a uniform change
of temperature T and by a nonuniform change (Td — 7^) varying linearly across the beam
depth. The temperature is assumed to be constant along the beam. The axial extension
is caused by the normal force and by the uniform change of temperature:
The change of curvature is caused by the bending moment and by the nonuniform change
of temperature:
In formulae (2.44) and (2.45), T = \(Td 4- Th) is the change of temperature at the beam
axis and subscripts d and h refer to bottom and top fibers of the beam.
We will use the matrix form of the Castigliano principle from Section 1.6.4 and in-
troduce a similar notation for relations (2.41) to (2.46). However, the matrix symbols
will have a somewhat different meaning because we deal with internal forces rather than
stresses. The integration over the volume will be reduced to integration over the beam
length. Equations (2.41) to (2.43) read in matrix form
or
In the same vein, we will rewrite equations (2.44) to (2.46) describing the material
properties as
where
is the vector of initial deformation due to temperature changes. The remaining symbols
are self-explaining.
The equation of compatibility is given by (1.225). The element flexibility matrix can
be calculated as
where
2.2. TRUSS AND BEAM ELEMENTS 75
This result has been obtained by performing the matrix multiplication and integrating
each entry of the resulting matrix in the limits from 0 to 1. The limits correspond to the
transformation x = £/, which changes the element of length / into a unit element.
Formally similar derivation gives the vector of three deformations (extension A and
rotations of the end sections on a simply supported beam $1,^2) due to the external
loading and temperature changes (Fig. 2.8)
where
When commenting on the FEM mathematical models at the end of Section 1.6.3, we
mentioned the useful transition from the force approach to the displacement approach.
This transition can be illustrated using the present element. The aim is to set up the basic
equations of the displacement approach (1.215) for one element. As the element is cut
out of the structure, the vector of end forces R representing the action of the surrounding
structure on the element must be added to the vector of transformed loads Rp. Both
force vectors R and Rp have the character of external loads when looking at the cut-out
element. The basic equation of the displacement approach thus reads
76 CHAPTER 2. SKELETAL STRUCTURES
The transition from (2.54) and (2.56) to (2.57) can be based on the relations between
the vector of three parameters of pure deformation A and the vector of six generalized
displacements
which include three degrees of freedom of the element as a rigid body. At the same time,
we must know the relation between the vector of three independent force parameters (3
and the vector of six generalized end forces
which are subject to three conditions of equilibrium. These relations are provided by the
following equations derived from Fig. 2.8:
This is the formal description of Krohn Theorem, which states that the matrix in the
kinematic equation (2.60) is the transpose of the matrix in the static equation (2.61).
According to (2.54) and (2.56), the vector (3 is the sum of two terms
Recall that the coefficient « was defined by formula (2.52). A further result is
where
The vector of transformed loads Up contains generalized forces acting on the supports
of the beam. Their opposite values (-Rp) represent the end forces on the fixed beam.
Therefore, the forth, third and sixth entry of (-Rp) corresponds to the entries of the vector
(3p calculated from formula (2.54).
Note that the external work includes the work done by the end forces (2.59), representing
the action of the surrounding structure.
We will use the same assumptions as in Section 2.2.1, i.e., the external load fx and
fz and the temperature change T = \(Td + Th) are constant along the element and the
temperature variation across the depth is linear.
In the Ritz method, the approximations of us, (py and w must be chosen such that the
displacements are continuous. The compatible model does not require the forces to satisfy
the equilibrium conditions (2.14) a priori, and thus we can start from the deformation of
the element at zero applied load, due solely to applied displacements and rotations at the
end sections, i.e. by the components of the vector (2.58).
Let us therefore approximate the longitudinal displacements by
which represents the exact solution under the assumptions made before.
The shape functions corresponding to individual parameters MI, $1, iu2 and $2 are
graphically presented in Fig. 2.9. The parameters $1 and $2 are the rotations of the
tangent to the beam axis at the end cross sections and they are positive in the anticlockwise
direction (same as the rotations of the end sections (piy and (p2y). This implies that
To complete the formulation, we have to find the approximation of the function (py. At
zero load /2, the shear force is constant along the element [this follows from the second
formula of (2.14)1. According to the third formula of (2.8), we then have
2.2. TRUSS AND BEAM ELEMENTS 79
The parameter 7 can be eliminated using the condition that the approximations w and
(py satisfy the moment equilibrium condition (2.14):
we get
The final approximation formulae can be obtained by using (2.74) in (2.70) and (2.75):
80 CHAPTER 2. SKELETAL STRUCTURES
It can be easily checked that if K — 0, formulae (2.76) reduce to (2.70), which verifies the
derived expressions. The sectional rotation (py must be for 7 = 0 equal to the rotation of
the tangent to the beam axis
Note that this verification is incomplete as we have not checked the terms that include 7.
Let us write the derived relations in a compact matrix form. The nodal displacements
and rotations will be collected into the vector r defined in (2.58). The approximations of
the unknown functions u = {us, (py and w}T can be written in the form given by (1.208),
where we set UQ = O. The resulting matrix
contains the shape functions hi corresponding to the individual nodal parameters. They
are shown along with their first derivatives in Table 2.1. The strain field represented by
should be approximated in the form suggested by (1.209). It is easy to derive the corre-
sponding matrix
In addition, we need formulae to be used for the calculation of the transformed loads in
(1.215). The assumption UQ = O implies RU = O. Furthermore, the loading acts only in
the plane ( x , y ) and the weight of the element can be included in / z , i.e.,
The zero term in the middle represents the distributed moment load (absent according to
the present assumptions). The effect of temperature is represented by the force NT and
the moment Mr, which are given by (2.65). In the matrix notation, this corresponds to
the product DeQ, given by
Now we have all the data needed to calculate the stiffness matrix K and the vector
of transformed loads Rp. In principle, the calculation can be performed
numerically
2.2. TRUSS AND BEAM ELEMENTS 81
10
• analytically.
The numerical calculation consists in applying a suitable integration formula to the
integrals in (1.215). This approach is more versatile than the analytical derivation as we
could apply it to elements with variable cross sections or with complicated load distribu-
tions. Integration in a closed form can be used for the prismatic element considered here.
This approach is not only more elegant but also faster when programmed on a computer.
We can also verify that the derived formulae are modifications of the well-known relations
used in the slope-deflection method.
First, let us calculate the vector of transformed loads for a uniform load p and for a
temperature change. The calculation is easy since it only requires multiplication of the
matrices and integration of each individual entry of the resulting matrix (vector). The
result
We did not have to deal with these terms before as they were multiplied by zero entries
of the vector (2.82). Integrating the definition of the stiffness matrix
we get the same result as derived before and given by formula (2.63).
To check the results, we will verify that the approximations of u3l (py and w meet
the basic criteria from Section 1.7, i.e., that they can describe rigid-body motions of the
element and the state of constant strain (constant curvature).
1. Rigid-body motions of the element can be described by three parameters UQ, WQ
and <PQ in such a way that the vector (2.58) has the form
Substituting the displacement parameters into the relations (2.69), (2.77) and (2.76), we
can see that the formulation is correct as
Of course, it is not surprising that the required criteria are met as the approximation
has been based on the exact functions w s , (py and w for an element loaded only by imposed
displacements and rotations at its end sections.
The decomposition consists in expressing the change of distance between the end
sections A as a sum of two terms
The second term can be easily calculated using the principle of virtual forces. Assuming
that the arch is sufficiently flat, we start from the work of the moments M = If (caused
by the unit force T) on the curvature change (-d2w/dx2):
84 CHAPTER 2. SKELETAL STRUCTURES
where
The derivation of the stiffness matrix of a curved beam element will be based on a
special approach that will be generalized in Section 3.7 when constructing the stiffness
matrix of a shell element. The pure stiffness matrix of a straight element (2.55) can be
presented in the form
where the scalar term K\ — EA/l represents axial stiffness, and the matrix KI with
two rows and two columns represents bending stiffness. As this matrix reflects also the
effect of shear we should modify formula (2.92) and derive generalized constants c*, c£
corresponding to the case K, ^ 0. The details of this derivation (which makes use of the
approximations from Section 2.2.2) are left to the reader.
The pure stiffness matrix of the curved element Kf (subscript / refers to the initial
rise of the element) can be obtained by comparing two expressions for the strain energy.
We can write either
where
or
The complete element stiffness matrix with six rows and columns includes the degrees
of freedom corresponding to rigid-body motions. According to formula (2.63), it can be
obtained bv the transformation
and let us treat the entries RI and R± of the vector of end forces R as torques at the end
sections. Comparing the above-mentioned constitutive equations, and further comparing
the vector (2.58) with (2.100) and the vector (2.59) with the present force vector H, we
find out that the stiffness matrix can be obtained by the following modification of formula
(2.63):
where
The vector of transformed loads for an element loaded by constant moments mx can
be obtained using the analogy with (2.64):
86 CHAPTERS. SKELETAL STRUCTURES
Combining the stiffness matrices and vectors of transformed loads of a beam element
with those of a grillage element, we can derive the corresponding relations for a general
three-dimensional beam element.
We will adopt the approximations known from Section 2.2.2 for the displacements and
rotations:
where
The interpolation functions are given in Table 2.1. Substituting the approximations
(2.103) into (2.102), we get
where
In order to get a result consistent with the stiffness matrix of the grillage element (2.101)
we would have to rearrange the rows and columns accordingly. We will leave out the
details of integration of (2.105) and (2.106). Using direct integration, we would get
Example 2.1
The effect of the shear stiffness of the foundation as reflected by the two-parametric
model, as well as the properties of the grillage element supported by a subgrade, will
be demonstrated by an example of a foundation beam in Fig. 2.11. The beam of
length / = 12m, width 26 = 1 m and height d = 0.5m is loaded by three forces
FI = 1MN, F2 = 1.5MN, F3 = 0.25 MN. The material characteristics of the beam
are E = 30 MPa, G = 10 MPa. Five different sets of the material characteristics of the
foundation given in Table 2.2 will be studied.
88 CHAPTER 2. SKELETAL STRUCTURES
set
Ci (MNm~3) 10 10 50 100 100
C2 (MNm-1) 1 5 25 10 50
variant
E (MPa) | 8 | 20 | 125 I 100 | 25
Table 2.3: Elastic moduli of the subgrade
The points corresponding to set 6 are marked by asterisks in the graph in Fig. 2.4.
It is worthwhile to compare the sets c and d. A larger value of the elastic modulus in set
c results in a substantial increase of the shear stiffness (constant C2) while the stiffness C\ is
even smaller than for set d. Although the graph gives only rough estimates of the stiffnesses
Ci,C2, it can be expected that especially the shear stiffness of the foundation (72 will have an
important effect on the stress and strain state in the beam.
FEM solution has been performed using 24 elements with a cubic approximation of the
displacements. The effect of shear has been taken into account. The results are presented in the
following figures.
Fig. 2.12 shows the deflections for the five sets of material characteristics. The distributions
of the contact stress are plotted in Fig. 2.13. It is obvious that a stiffer foundation results in
stress concentrations below the applied forces and in stress relaxation in the intermediate parts.
The distribution of shear forces in Fig. 2.14 clearly shows the effect of the shear stiffness
parameter C<i> For set a, the shear stiffness C<i is relatively small and the shear forces at the end
sections are close to the applied forces FI = 1MN and F3 = 0.25 MN. For the remaining sets,
the effect of shear is larger and the differences between the applied force and the shear force are
due to the shear force in the subgrade, which acts on the end sections of the beam against the
direction of the applied forces.
The distributions of bending moments shown in Fig. 2.15 are very useful in design. The
reduction of bending moments between the applied forces is substantially larger than below the
force FI.
Consider the frame in Fig. 2.16. It is a typical example of a steel structure as it has a
number of internal hinges. However, the fundamental beam is continuous across all the
columns and the main columns have no internal hinges, either. Let us look at the close-up
in Fig. 2.17. The joint denoted as i has 3 degrees of freedom—HI, Wi, (piy = ^ (subscripts
y will be dropped). This joint connects four members denoted as a, 6, c, d.
The displacements and rotations of the end sections of members b and c are uniquely
defined by u^, Wi and </?;. The displacements of the end sections of members a and d are
also given by ut-, w^ but the rotations are different from ^.
There are two ways to deal with this problem: First, we can introduce additional
unknown rotations at joint i so that we work with three independent rotations—^, (pia,
Vid (<Pia is the rotation of the right end section of member a; (pid is the rotation of the
left end section of member d). Each of this parameters is associated with an independent
equation of equilibrium. The conditions of equilibrium associated with the parameters ^a,
(pid require that the moments at the ends of members a and d attached to joint i be equal
to zero. But why should we include these simple conditions, which are related to only
one member each, among the global equilibrium conditions for the entire structure. We
thus discover the second way of solving the problem. The conditions of zero end moments
can be used to modify the element stiffness matrix of the corresponding member and the
global equilibrium conditions will then contain forces on a beam fixed only at the opposite
end.
We will show how to get the modified stiffness matrix and the vector of transformed
90 CHAPTER 2. SKELETAL STRUCTURES
loads. Let us look at a specific member, say member a. For the sake of simplicity, we
assume that the displacements u^ u2 are equal to zero. Even if they are nonzero, they
affect neither the shear forces nor the bending moments.
The matrix manipulations will be easier if we arrange the unknown parameters so
that the parameter (p2 (to be eliminated using the condition M2 = 0) is the last one.
The generalized displacement vector r consists of {wi, <pi,w2, (p2}T and the corresponding
stiffness matrix and vector of transformed loads look as follows:
The condition of zero moment at the right end of the element follows from the last
row of the matrix K and the last entry of the vector Rp'.
2.5. STATIC CONDENSATION 91
Recall that the relation Kr = Rp + R holds even for a single element. The dependence
of (p-2 on the remaining parameters wi, (p\ and w^ can be taken into account simply by
substituting for (p2 into each of the first three equations from Kr = Rp + R. We thus
92 CHAPTER 2. SKELETAL STRUCTURES
get the modified stiffness matrix and the modified load vector:
where
Setting K, = 0, i.e., neglecting the effect of shear, we can derive from the second row
the well-known formula for the moment reaction on a unilaterally fixed beam,
We have neglected the effect of temperature and introduced the beam stiffness k = 2EIy/l
(this is the notation used in structural analysis but note that we have been using the
symbol k to denote a certain coefficient related to the effect of shear stress distribution on
the value of 7). The derivation of the formulae for vertical reactions is left to the reader.
Let us turn attention to the general concept of static condensation. The example
given above is instructive but it lacks formal simplicity and versatility necessary to set
up an effective solution algorithm. Suppose that we know the stiffness matrix K and
the vector of transformed loads Rp corresponding to a certain substructure. In principle,
the substructure can consist of any number of elements. A simple substructure is often
encountered when using triangular elements to analyze two-dimensional structures. The
2.5. STATIC CONDENSATION 93
input of data can be simplified if we deal with quadrilateral elements formed by assembling
four triangular elements as shown in Fig. 2.18 for plate elements. In this case, each node
has three displacement parameters:
• deflection w,
The parameters w, (px, (py associated with the node inside the quadrilateral element will
be referred to as the internal parameters because the corresponding equilibrium conditions
contain only the parameters of this element. This is similar to what we have seen for the
member fixed only at one end. We will again arrange the displacement parameters so that
the ones to be eliminated by condensation follow after the remaining ones. The stiffness
matrix, nodal displacement vector and vector of transformed loads will be partitioned into
blocks. The parameters associated with the internal degrees of freedom will be denoted
with the subscript i (internal), the others by e. (external). Using this notation, we can
rewrite the stiffness matrix and the vectors as follows:
The equilibrium conditions on the substructure taken as a free body then read
The vector Ri contains the external forces acting directly on the internal node z, and the
vector Re contains the generalized forces acting on the substructure from the surrounding
structure. We now express r; from the second equation of (2.115):
This expression can be substituted into the first equation (2.115) to yield
where
Example 2.2
Consider a unilaterally fixed beam element as a special case of a substructure. As only
one parameter is subject to condensation, the block KH is represented by one scalar entry
and thus can be easily inverted.
Solution:
Symmetry of the stiffness matrix implies that Ke{ = K?e. The individual blocks Kee, Kei, KH
and the vectors Rpe and Rpi have the form
The only element of the vector Ri would be the external moment applied at the hinge (it usually
equals zero). The details of the calculation following (2.119) are left to the reader. The result
must agree with formulae (2.111) and (2.112).
The example showing the elimination of the unknown rotation at one end section of a
beam element has illustrated the general approach to condensation described by (2.119).
Whenever possible, we try to avoid inversion. The matrix K* and the vector R* can
be obtained by applying Gauss elimination to the matrix K and the vector -R. This
algorithm is very simple, numerically stable and it preserves the banded character of the
stiffness matrix if the substructure is large.
The transition from the stiffness matrix expressed in the local coordinate system to the
one expressed in the global coordinate system will again be demonstrated by the example
of a beam element.
Let us start by geometrical considerations. The origin of the local and global coor-
dinates can be chosen at the same point, as shown in Fig. 2.19. The coordinates of a
generic point P in the global coordinate system will be denoted as X and Z and in the
local coordinate system as x and z. The transformation formulae follow directly from the
figure:
where
As the displacements are transformed in the same manner as position vectors and the
rotations (py are not affected by in-plane coordinate transformations, the relation between
the unknown parameters r/ in the local coordinates and the parameters rg in the global
coordinates can be described by
Let us emphasize again that all the relations derived so far have been related to the
local coordinate system with the x-axis identical to the beam axis. The equation
in local coordinates can be transformed by substituting from (2.121) and from the similar
equation for the force vector /fy:
96 CHAPTER 2. SKELETAL STRUCTURES
Equation (2.124) can be multiplied from the left by AT, and as ArA = I we get
This relation is the global counterpart of (2.123). The matrix multiplying rg is the stiffness
matrix in global coordinates,gK . It can be calculated from the local stiffness matrix as
The load vector related to the global coordinate system can be expressed by multiply-
ing the basic relation RI = ARg from the left by AT. This yields
The vector p{ usually does not have to be transformed. It contains, e.g., the lateral
and axial force on a beam element. Only the displacement vector in (2.129) is then
transformed according to (2.121) and we get
The block structure of the matrix A shows that there is no geometric coupling between
the displacement components at node 1 and at node 2. The coordinate systems used at
the nodes thus do not have to be aligned. However, all the elements connected to the
same node must use the same coordinate system at this node. It clearly makes no sense
to choose a different referential system at every node just for fun. If there are no special
reasons for not doing so, we relate all the nodal parameters to a single global coordinate
system.
When dealing with frames, the special reason for choosing a referential coordinate
system unaligned with the global one might be a certain type of supports. Consider the
structure shown in Fig. 2.20. The global coordinate system (X, Z) is suitable for all the
joints except the one at the right bottom. Due to the inclined support, it is convenient to
2.6. COORDINATE TRANSFORMATION 97 97
introduce another referential coordinate system (X1, Z'}. The kinematic boundary condi-
tion can be easily described by setting Wj = 0. As either of the end sections on the beam
/ — J has a different referential coordinate system, the transformation matrix A contains
two different diagonal blocks a, a7.
Example 2.3
The preceding paragraphs dealt with elements for the analysis of skeletal systems. The
solution was based on Mindlin's assumptions about the deformation of beam elements.
To get at least some information on the accuracy of such an approach, we will compare
the results of a solution according to Mindlin's theory with the results obtained by an-
alyzing a two-dimensional finite element model. Two-dimensional problems are in detail
discussed in Chapter 3. In the present example, we restrict our attention to comparing
the deflection at the middle cross section of a fixed beam under uniform loading (Fig.
2.21). The loading is applied at the upper and lower surfaces of the beam. The solution
by the beam theory uses the sum of the magnitudes of loadings on both surfaces. The
figure also shows the material characteristics.
Solution:
We first calculate the middle deflection according to the beam theory. The calculation is based
on formulae (2.8) written as
As the problem is symmetric about the vertical axis, shear deformations do not affect the
distribution of internal forces. It thus holds
Substituting into the relation between the bending moment and curvature, we get
As the cross section at x = 0 is fixed we have C — 0. Using this result, (py can be eliminated
from the second equation to yield
Let us substitute numerical values according to Fig. 2.21. First consider the shape coefficient
(2.3) to be k = 1:
This result implies that, for a rectangular cross section and the ratio h/l = 1/3, the effect of
shear deformation on the middle deflection of a fixed beam is 88.8% if k = 1. For k = 5/6, this
effect further increases to 106.7%. The middle deflection for k = 1 is given by
Now let us analyze the beam using two-dimensional finite elements. In this case, the deflec-
tion varies across the depth due to lateral compressibility of the beam. As shown in the figure,
we have applied one half of the loading at the top surface and the other half at the bottom
surface of the beam. This results in a skew-symmetric response relative to the beam axis. The
horizontal reactions at the supports are thus zero. This is very important as the horizontal reac-
tions would produce the arch effect and the results could not be compared to those obtained by
the beam theory. The calculation shows that the deflection at the centroid of the middle cross
section is w = 3.57.10"5 m, and at the top or bottom it is w = 3.678.10"5 m. This perfectly
agrees with the solution by beam theory.
The beam solutions (with k = 1 and k = 5/6) have given the average deflection, which for
k = 5/6 lies inside the interval < 3.57.10"5; 3.678.10~5 >. It can thus be expected that the real
shear stress distribution across the depth is closer to a parabola (k = 5/6) than to a rectangle
(k = i).
Fig. 2.22, showing the deformed shape of the structure, reveals several interesting facts.
First, it can be observed that the shear deformation close to the extreme fibers is small. Second,
the straight line corresponding to the rotation of a rigid cross section in the beam solution clearly
verifies the assumption that warping can be neglected. Fig. 2.23 shows the result of another
calculation approximately simulating Kirchhoff's assumptions by setting G == 1.109 kNm~2 (very
high shear stiffness). The figure indicates that the cross sections remain normal to the deflected
beam axis.
2.6. COORDINATE TRANSFORMATION 99 99
Figure 2.22: Beam deformation modeled by bilinear isoparametric elements (Section 3.1)
Using the interpolation functions NI, 7V2 the previous equation can be recast as
Where
100
3.1. BASIC RELATIONS FOR ISOPARAMETRIC ELEMENTS 101 101
The above reasoning is not new. We have been using an isoparametric interpolation
in the previous chapter. The dimensionless parameter was £ = x/L An important step
when translating from the coordinate system (x, y, z) into the natural coordinates
is the derivation of the relationship between differentials. For one-dimensional elements
we have
where J is the Jacobian of the transformation, which is constant over the bar,
because
The basic property of isoparametric interpolation functions Ni is that they are equal
to 1 at the ith point, and that they are zero at all other nodal points. Let us consider
quadratic interpolation for a one-dimensional element with three nodes. The individual
interpolation functions are visualized in Fig. 3.2.
The formulation as shown in Fig. 3.2 is due to Bathe [17]. Elements with a variable
number of nodal points can be easily constructed. Let us note that leaving out node 2
and the quadratic terms in the interpolation functions, the linear interpolation functions
of 3.2 are obtained (they are framed by dashed lines in Fig. 3.2).
z=5 z=6 2 = 7 z= 8
tfi = 0.25(1+0(1 + 17) -0.5 7V5 -0.5 AT8
#2 = 0.25(1 -0(1 + 17) -0.5 A^5 -0.5 N6
#3 = 0.25(1-0(1-17) -0.5 AT6 -0.5 AT7
AT4 = 0.25(1 +0(1-??) -0.5 JV7 -0.5 A/s
N5 = 0.5(1 -£ 2 ) (1 + 7?)
#6 = 0.5(1 -77 2 ) (1-0
jv7= 0.5(1 -a (i-r?)
N8 = 0.5(1 -r? 2 ) (1 + 0
Fig. 3.5 shows the shape of the element as well as the natural coordinates. The
geometry of the element may be described by appropriate natural coordinates. We express
the physical coordinates x, y as
The essential step is the reformulation of the problem from the coordinates (x, y) into
the coordinates (£,77). Let us consider an arbitrary function / = /(z,y), which might be
thought of as a representation of displacement, temperature, pore pressure, etc., and let
us approximate the function isoparametrically, i.e.,
The functional of the FEM contain not only the values of the sought function but also
its derivatives and its derivatives — and —. Therefore, the dependencies between the partial derivatives
and -^—, -77— and -^—, -77— must be known. Using the chain rule we can write
Equation (3.8) can be viewed as a system of two linear equations from which the
inverse relationship can be obtained
In the above
is the Jacobian of the transformation. In order to compute the Jacobian for the bilinear
interpolation at an arbitrary point, (3.6) must be substituted into (3.9). Then
where
where
Let us collect the values of the function / at the nodal points into a vector rT =
{/i» /2, /3, A} and let us rewrite equation (3.9). In analogy to the computation of
the Jacobian we find
or, alternatively,
Similarly,
The derivatives of / at the point (£, 77) can be expressed by simple relations
The most important formulae needed to compute the stiffness matrix of an isoparametric
element are thus ready for future use.
The cubic approximation with 10 parameters is rarely used. The nodal points are shown
in Fig. 3.6, along with the corresponding hierarchy of approximation polynomials. Let us
note that in all three cases the polynomials are complete.
It is convenient to work with the triangular elements in the so-called area coordinates
I/i, Z/2, 1/3. It holds
Each triple (Z/i, 1/2, I/s) corresponds unambiguously to the pair (x, y) and vice versa.
Values of I/i, 1/2, £3 are not independent, as they are coupled through the relation I/i +
1/2 + Z/3 = 1.
In view of the linearity of the relationship between the Cartesian and area coordinates,
the equations
describe straight lines parallel to the sides of the triangle—compare with Fig. 3.7.
The denotation area coordinates is due to their geometric interpretation. It is easy to
show that, e.g., LI of point P from Fig. 3.7 is the ratio of the hatched area to the area of
the whole triangle.
Besides equation (3.17) it is necessary to have also the inverse relation, which takes the
form
where
106 CHAPTERS. PLATES AND SHELLS
and
For the quadratic approximation the interpolation functions corresponding to the corner
nodes are
Similarly,
When computing the matrices of stiffness, mass or the transformed load, integrals of
functions dependent on LI, L2, L3 must be evaluated. This is greatly facilitated by the
explicit formula
is the displacement along the axis x, i.e., u(x). The displacement approximation along
the bar length is expressed by (3.3), which can be written in matrix form as
where
du
The strain ex — — may be computed from
ax
where
The stiffness matrix of the tension-compression bar is computed from the formula K =
Jo BTDB dx. The matrix D has the only one term, DU = EA, and so we have
The structural element shown may be a part of a vertical support system of a high-
rise building. The length / of the element corresponds in such a case to the story height,
and the local axes y, z might be oriented parallel to the global Cartesian axes Y, Z. The
function u* describes unit warping of the cross section. It is expressed with respect to
the pole P and to the origin SQ of the coordinate s and usually is determined from the
assumption of a free torsion. It corresponds to the sector coordinate u on the open
108 CHAPTERS. PLATES AND SHELLS
branches of the cross section? The cross section of the element possesses seven degrees of
freedom: displacements of the section as a rigid body—u s , vs, w3, (px, (py, (pz—and the
warping measure x- Using the notation of Fig. 3.8 we can express the displacements of a
generic point HI and u2 in the direction of re-axis and in the direction tangential to the
center line of section 5 as
where
Using the equivalence conditions and Hooke's law, we can evaluate the internal forces,
i.e., the normal force Nx, shearing forces Qy, QZ1 torque Mx, bending moments My, Mz
and the bi-moment B. After some manipulation we have [cf. equation (2.8)]
Derivatives with respect to x are denoted by primes. Newly introduced matrices are
defined as:
Further it holds
Derivatives with respect to s are marked by dots. Using the properties of the coordinate
functions y, z, u* we get an equivalent expression
Let
be the vector of generalized displacements of the end sections. Superscripts h and d mark
the parameters corresponding to the top and bottom cross section, respectively. The
stiffness matrix will be of type (14,14) and follows from
It remains to approximate the unknowns us, q±, q2 in (3.36). We restrict further consid-
erations to the linear isoparametric approximation1 and we substitute
where jxz is the shear deformation of the cross section in the plane (x, z] and jxy is the
shear deformation of the cross section in the plane ( x , y ) .
Application of (3.40) leads to all terms in (3.36) being constant and the integration
simplifies to matrix multiplication. The same effect can be achieved by using one-point
Gauss quadrature to integrate (3.36). The terms with derivatives u'8, q'lt q'2 are integrated
exactly, while for the terms with q^ q2 this means a reduced integration. It can be
shown that this approach (also called selective integration) is perfectly consistent with
the assumption (3.40).
The obtained stiffness matrix is summarized in Table 3.2. The symbol OT denotes in
Table 3.2 a row matrix of type (1,3) with all elements equal to zero.
loads) must be independent of z too. Let us summarize the most important relations with
respect to the notions of plane strain and plane stress from the first chapter.
If the component ez of the strain tensor is equal to zero (in addition to the above
conditions), the problem at hand is plane strain. The components 7yz,72X of the strain
tensor are also zero. The nonzero components of the stress tensor are vx,Oy,Txy,Oz- The
solution to the problem must satisfy all equations for a general three-dimensional stress
state. Some equations are identically satisfied because of the above reduction. Plane
strain problems appear naturally for solids whose dimension in the z is much larger than
in the directions of x and y. A classical example is a dam.
If the component az of the stress tensor is equal to zero (in addition to the above
conditions), the problem at hand is plane stress. The nonzero components of the stress
tensor are crx, cry, rxy. Nonzero is also the ez component of the strain tensor. The shear
stresses ryz, rzx and the corresponding strains jyz, 7ZX are zero. A typical example is a
strength analysis of a thin web, loaded only in its plane2 (Fig. 3.9).
The unknown functions—the displacements u and v—are independent of z in both
above cases. The static and geometric equations are identical for both problem types.
The constitutive equations (the material stiffness), on the other hand, differ. For a ho-
2
The assumption of plane stress are satisfied only in the limit of infinitely thin sheet t -»• 0 or for the
so-called generalized plane stress.
3.5. ELEMENTS FOR PLANE PROBLEMS 111
Plane strain
Plane stress
The solution of the two-dimensional problem of the theory of elasticity was the first
successful application of the finite element method (see [30]). In both the plane stress
and plane strain cases there are three nonzero components of stress and strain tensor. In
the case of plane stress the remaining stresses are zero and do not contribute to the work
of internal forces. In the case of plane strain the remaining stress tensor component is
nonzero. The conjugated strain is zero, however, and it follows that the contribution to
the work of internal forces is zero.
In what follows, the fundamental relations for stiffness determination of a triangular
element with constant strain fields will be derived first. Further, our attention will be
devoted to the isoparametric quadrilateral element and some useful modifications to this
element will be described. Some generalized variational principles will be applied to that
end.
The displacements inside the element must therefore be uniquely described by these six
parameters:
and
or, alternatively,
e = Br.
It may be noted in equation (3.46) that the components of the vector e are constant
element-wise. Therefore, the element is conforming as both the conditions of continuity
and constant deformation are satisfied.
The stiffness matrix was derived in the previous chapter as
where the integration domain £) is the element volume. Assuming that the material
stiffness D does not change within the element area, the integration of (3.47) gives (t is
the thickness)
K = BTDBAt.
3.5. ELEMENTS FOR PLANE PROBLEMS 113
A formula for the stiffness matrix can be derived using tensor calculus. The notation
Ui (i — 1,2) will be used for displacements of the point (zi,^) with wf denoting their
values at node /3 = 1, 2,3.
Following the previous reasoning, we will first compute derivatives of the displace-
ments. As it holds
we have
Expansion of (1.39) by the preceding formula renders the stress tensor for plane stress
as
where 7 = (1 — ^)/2.
The stiffness matrix follows from [compare with (3.38)]
Let us rewrite the first row that corresponds to the virtual nodal displacement 6u\,
which is identical to 6u\ in matrix notation. In (3.52) we set i = 1, a = 1 (summation
takes place over over the remaining indices). A simple transformation based on (3.50)
gives for the first row (compare with [118]):
where
The above solution was derived from the assumption of a linearly elastic, isotropic
material. Other material models discussed in the first chapter can be introduced into the
computation through their tensors of material stiffness.
114 CHAPTERS. PLATES AND SHELLS
To simplify the typography, the nodal displacements will be collected into two vectors
The geometric equations can be written for the bilinear approximation using (3.16)
e = Br,
where
2 x 2 2 x 2 2 x 2 3 x 3
2 x 2 3 x 3 3 x 3 I* x A
load vector. It will be assumed for simplicity that the loads fx, fy acts along the side 1
- 2 of the element (see Fig. 3.11). For the displacements along this side we have
The corresponding transformed load vector is computed from (1.215). After substitu-
tion it. rpaHs
The modified matrix B can subsequently be applied in the same algorithm as the original
one.
We cannot analyze the theoretical aspects of the reduced integration with respect to
the extended Lagrange principle in full details here. Let us only show the identity of these
two approaches for the bilinear elements. Let the following notation be introduced to this
end
The FEM formulation is derived from the condition of zero variation, which reduces
to the equation dU/dr = O for the discretized problem. Expanding this condition we
arrive at
3.5. ELEMENTS FOR PLANE PROBLEMS 117
The stiffness matrix is obtained by integration of the term in brackets. If the reduced
integration is used, the first integral is computed by Gauss integration rule with 2 x 2 ,
or 3 x 3 points. The second term is, on the other hand, evaluated only by one-point
integration:
and let us assume that 7 is constant element-wise. The displacements u, v are approxi-
mated bilinearly, in the same manner as with the original element. The new functional is
of the form
Using the second equation (3.62) the function 7 can be eliminated. Let us turn our
attention to the integral of B7. This function is linear in both arguments. To get the
exact value of the integral, it is sufficient to apply one-point Gauss integration. Therefore,
it holds
By substituting this result into the first equation of (3.62), the equilibrium condition for
the element is obtained as
118 CHAPTERS. PLATES AND SHELLS
We have arrived at the same equilibrium conditions as in (3.58) and (3.59). It follows
that for this special case, constant shear and bilinear displacements u, v, the selective
integration is identical to the procedure derived from the generalized variational principle.3
The presented formulation is not invariant with respect to the rotation of the coordi-
nate axes. To eliminate this phenomenon the stiffness is evaluated in the element local
coordinates (z',2/)> which are shown in Fig. 3.12. The orientation of the axes x',y' is
derived from axes £, 77 so that the angle subtended by the axes x' and £ equals the angle
subtended by y' and rj. The stiffness matrix thus computed is transformed in the usual
manner into the global coordinates.
The second modification relies on the introduction of additional degrees of freedom,
which are obtained through a hierarchical bubble function. A similar approach will be
applied when modifying the quadrilateral plate element of Section 3.6.5. The hierarchical
functions are dealt with in depth in chapter 10. The bubble function is expressed in
coordinates (£, 77) as (1 — £2)(1 — r?2). It attains value 1 at the arithmetical center £ = 0,
77 = 0 of the element and it vanishes on its edges. Using this function we can enhance the
approximation of u, v as
Let us denote
Then
3
This is not true in general for higher-order elements. The selective integration is also less efficient for
higher-order isoparametric elements (quadratic, etc.).
3.5. ELEMENTS FOR PLANE PROBLEMS 119
Introducing both modifications discussed so far at the same time, the matrix B is of the
form
The stiffness matrix changes its type from (8,8) to (10,10). However, it is obvious that
the Aw and At; are internal degrees of freedom and can be eliminated by condensation—
cf. Section 2.5. Because of this property the described modification does not inflict upon
the element any "cosmetic" flaws.
The third modification consists of introduction of incompatible interpolation functions,
which violate compatibility requirements on the inter-element boundaries. It can be
proved that this model converges to the correct solution, albeit nonmonotonically. The
incompatible functions are plotted in Fig. 3.13. An incompatible interpolation function
(mode) is introduced as a quadratic both in the £- and //-direction for each of the functions
u, v. It can be shown that the quadrilateral with incompatible modes passes the classical
patch test if it is rectangular. The incompatible element gives very good results in model-
ing of bending, as can be seen from Table 3.4. Reference [80] describes a modification that
ensures that the patch test is passed even by a general quadrilateral. Our own experience
seems to indicate that the performance of this element is not fully satisfactory, however.
membrane element leads to only two additional degrees of freedom. The resulting stiffness
is therefore undefined for the rotation about the normal to the center surface. However,
most finite element packages require six degrees of freedom per node—three displacements
and three rotations. The problem can be solved by relating the degrees of freedom to a
local coordinate system (1,2,3), with axes 1,2 in the plane tangent to the mid-surface—cf.
Fig. 3.14.
Figure 3.14: Local coordinate system (1,2,3), with axes 1,2 in the plane tangent to the
center surface
This solution may lead to a reduction in the total number of unknowns as the rotation
about normal may be eliminated.4 On the other hand, some topological problems may
arise.
The combined membrane-plate elements show an unfortunate behavior—the bending
is often described more adequately than the membrane action, while a well-designed
shell structure carries the loading predominantly by the membrane internal forces. The
approximation of the membrane forces is of critical importance especially with respect
to the stability of equilibrium (cf. Chapter 9). It is interesting to note, that really
serious struggle to remove the above flaw started as late as the eighties [1, 112, 30, 81],
when membrane elements using rotational degrees of freedom appeared. The elements
mentioned here are based
Free formulation
The formulation described in what follows was proposed by Bergan and Felippa [30],
and it represents an alternative approach to the FEM. It starts from the fact that it
is possible to propose a membrane element with rotational degrees of freedom, which
satisfies a priori the modified patch test and performs well in in-plane bending problems.
The free formulation differs from approaches based on variational principles in that full
compatibility on the inter-element boundary is not required to achieve convergence.
The patch test is a standard approach to test the convergence of incompatible
elements—cf. Irons' paper[90]. Bergan and Hanson have proposed an approach that
4
This approach is perfectly correct for folded plates, but it tends to stiffen flat shells due to the
eliminated rotation degree of freedom.
3.5. ELEMENTS FOR PLANE PROBLEMS 121
is equivalent to the patch test in [30]. The test is formulated as a system of linear con-
straints applied to the tested stiffness matrix. The displacement fields corresponding to
the rigid body displacement and to constant strain state can be stated as
where
Nr is a complete system of linearly independent shape functions
(modes), representing the rigid body displacement of the element,
Nc is a complete system of linearly independent modes,
representing states of constant strain,
</r, qc are the corresponding generalized coordinates.
The relationship between the generalized coordinates gr, qc and the nodal displacements
r is of the form
where G>, Gc are obtained by substituting the coordinates of the nodal points into
Nr,Nc.
The conditions (3.71) are applied to the stiffness matrix of the isolated element at
hand. The first condition of the patch test requires the modes qr be stressless. Therefore,
where Pr is a matrix, collecting in columns the nodal forces from the individual rigid
body modes.5
The requirement of constant strain modes leads to a constant strain over the whole
element, i.e., also along its edges. If the stresses are replaced by (statically, energetically)
equivalent nodal forces pc, the second requirement of the patch test can be formulated as
a constraint posed on the stiffness matrix. Consequently,
where Pc is a matrix, whose columns represent nodal forces excited by the individual
constant strain modes. The constraints (3.72), (3.73) are then equivalent to the original
patch test
The free formulation requires the patch test to be a priori passed. It leads to a stiffness
matrix that is composed of two parts—a basic stiffness matrix Kb and a higher-order
stiffness matrix Kh,
The goal is to obtain a membrane element possessing two displacement and one rota-
tional degree of freedom per node. The rotation about the normal to the element plane
is defined as6
5
The terms of the vector of the generalized coordinates have this meaning: qri is the displacement
along the axis x, qr2 is the displacement along the axis y, and qr3 is the rotation about the element
barycenter.
6
Detailed discussion will be presented in the following Section.
122 CHAPTER 3. PLATES AND SHELLS
The rotation thus defined is invariant with respect to the orientation of the coordinate
system used, but it is not identical to the rotation of the element edges at the nodes.
Therefore, the relation between these two quantities is expressed through a parameter a
It turned out that the bending behavior is considerably improved for a > 1. It is
necessary to note that a. must be set to the same value in all elements of the model so
that the second requirement of the patch test is not violated.
The interpolation function matrices for the triangular element of Fig. 3.15 are given
by
The basic stiffness matrix can be computed directly using the material stiffness matrix
D and a transformation matrix L, which transforms constant stress along the element
edges to equivalent nodal forces. Thus, it holds
The stiffness matrix of the regular constant strain triangle (CST) is obtained by in-
serting a = 0 into (3.76).
The higher-order matrix is derived from the interpolation (3.75), extended as
3.5. ELEMENTS FOR PLANE PROBLEMS 123
The interpolation functions NM represent pure bending in the coordinate system £,777,
where & is measured along the triangle median corresponding to i th nodal point, and rjl
is orthogonal to & —cf. Fig. 3.15.
In the local coordinate system (£i,rjj) we have
The parameter is based on the fact that the free formulation satisfies the patch test with
an arbitrary higher-order stiffness matrix. It is only necessary to generate correct rank
stiffness matrix (this is essential for ensuring that there will be no energy-free mode in
the stiffness matrix K). Optimal choices of parameters a, fi lead to very good results.
Recommended values of parameters a, /3 are
Details, including a very readable program listing, can be found in [30]. Comparison with
other membrane elements is presented below.
In contrast to the formulation discussed in Section 1.6, symmetry of the stress tensor
is not assumed a priori. As the matrix notation is natural in FEM, we introduce matrices
where the terms of the vector r are components of the skew-symmetric part of the stress
tensor, and cr collects components of the symmetric part of the stress tensor. Operator
d is defined by (1.5).
Assuming a priori validity of the constitutive equation (the components of <r are not
independent), the governing functional can be cast as
where 7 is a preset quasi-material constant influencing the material stiffness. The func-
tional (3.87) resulted as an extension of the classical energy functional by the energy of r
(third term on the right-hand side) and by the work of r on the dislocations (Vti — u>),
that are the result of the independence of the fields u and cj. Using the stationarity
condition of the functional
the first three equations in (3.84) are obtained as the Euler equations. Due to the fact
that both displacements and stresses appear in (3.87), the discretization leads to a mixed
element.
If r were eliminated by using the quasi-constitutive equation
3.5. ELEMENTS FOR PLANE PROBLEMS 125
leading to a displacement formulation.8 The two functional above are in analogy to the
introduction of the independent shear jxy in the isoparametric membrane element of the
previous section.
The constant 7, appearing in equations (3,87) to (3.89), is problem dependent. Hughes
and Brezzi have shown in [81], that an isotropic material in a planar problem leads to the
optimal choice 7 = G (shear modulus of elasticity). Numerical experiments have shown
that the above formulation is relatively insensitive to the choice of 7. It is recommended
to use 7 = G even for orthotropic materials.
The stiffness matrix will be derived from functional (3.87). The individual functions
in (3.87) will be interpolated as
The functions Ni...NB were given in Table 3.1, where the functions NI...NI are specified
in the bold framing and N5...N8 are taken without the coefficient 0.5. The hierarchical
degrees of freedom Aw and Av are condensed out in analogy to the second modification
above.
The indices are obtained from «/, K FORTRAN-like expressions
J = / - 4 , K =mod(/,4) + l.
The second term in the relations for u and v was taken from Allman, who extracted it
while looking for a solution to the bending of a beam (the derivation was presented in
Section 3.6.5 dealing with the modified plate element). According to Fig. 3.16, IJK stands
for the length of the element edge JK and OLJK is the angle subtended by the normal of
this edge and the x-axis. Let us denote
8
The first integral in (3.89) can be rewritten in the usual way:
The discrete form corresponds to the stiffness matrix of the classical isoparametric element.
126 CHAPTERS. PLATES AND SHELLS
where
The matrix B was specified in Section 3.5.3. The terms of the 7th column of matrix G
are written as:
Af = I + 4, L = M - l + 4 aint(I/I), K = mod(M,4) + 1, J = L - 4.
Let us further note that the expression (Vtx — w) reduces for the membrane problem to
where
where
Similarly, the discrete formulation can be derived from the functional (3.89). We get
where
Example 3.1
To compare the quality of the elements discussed, the results achieved for one partic-
ular problem (clamped beam of Fig. 3.17) are listed here.
Solution:
• Bergan free-formulation element (BEFE) (one quadrilateral being composed of two trian-
gles), and
concerning these elements can be found in [100]. The elements were providing good
results while keeping the number of unknowns down to earth. Their main disadvantages
were degrees of freedom corresponding to higher derivatives (complicated transformation
properties). Also, the influence of shear deformation was not included.
The problem of Cl continuity required by thin plate solutions was relieved by Her-
rmann [72] through application of Hellinger-Reissner principle. Simplest element for thin
plates with one degree of freedom per node (deflection) and one along the side (bending
moment Mn) resulted from this approach (both rectangular and triangular elements can
be constructed in this manner). Details can be found in [7]. The approach is a typical
representative of the mixed formulation of the FEM. Despite that fact that the elements
give good results, they cannot be found usually in general-purpose finite element packages
as they require nondisplacement degrees of freedom to be used.
There is a whole series of plate elements which were developed using the modified
variational principles [133]. These elements are called hybrid and are frequently used,
especially in linear problems. Elements were developed for both Kirchhoff and Mindlin
plate theories.
The following sections discuss in detail three representatives of the modern approach
to plate analysis. The selection was, of course, subjective. Their good features include
robustness, generality and simplicity.
3.6. PLATE ELEMENTS 129
Plate elements can be studied from two viewpoints. From the viewpoint of element
geometry the situation is quite similar to the plane problem: Triangular and isoparametric
quadrilateral elements are the most frequently used elements. The main advantages of
triangular elements can be recognized in combined membrane/plate and shell problems,
especially in the nonlinear regime. Isoparametric elements of higher order are of interest
in particular for plates with a curved boundary, as it is by now a well-established fact that
the inaccuracies due to approximation of the boundary geometry are of the same order
as inaccuracies stemming from the basic approximation of the unknowns.
Looking at the elements from the viewpoint of deformation type, two theories come
into the picture: (i) the Kirchhoff theory of thin plates, and (ii) the Mindlin-Reissner
theory of thick (shear-deformable) plates. It is of interest to note that the Kirchhoff theory
(with the discrete Kirchhoff variant) is a special case of the latter, and consequently, the
Mindlin-Reissner theory will be given a larger share in subsequent discussion.
• The normals to the mid-plane remain straight after the deformation, but they are no
longer perpendicular to the deflected mid-plane surface (that is why they are called
pseudo-normals). The theory therefore neglects deplanation of cuts along the plate
thickness — cf. Fig. 3.19.
• The normal stress az is negligible in comparison to stresses &x, cry and can be left
out of energy considerations.
The first two assumptions are of a purely geometric nature. The first assumption yields
The second assumption can be interpreted as describing the motion of the pseudo-
normals as that of a rigid body in three-dimensional space. The position of such a body
can be specified by six degrees of freedom: location of an arbitrarily picked point, and
three orientation parameters. Taking account of the u and v displacements of any point
being zero, the number of degrees of freedom reduces to four. But pseudo-normals are
130 CHAPTERS. PLATES AND SHELLS
special "bodies," with no "cross section." Therefore, the drill rotation (pz (rotation of the
pseudo-normal about itself) is immaterial. There remain only three degrees of freedom,
i.e., deflection of the mid-surface along z, denoted as w, and the rotations of the pseudo-
normal about x and y, which are denoted as <px and (py. The positive sense of these
parameters (which are functions of x,y) is shown in the Fig. 3.19. It follows that the
motion of the pseudo-normal is described by
Using (3.103), the nonzero components of the strain tensor will be written
where «x, Ky, Kxy are pseudo-curvatures of the mid-surface. The nonzero components of
the stress tensor crx, <7y, r xy , rxz and ryz are obtained from constitutive equations.
Let us imagine the plate cut to thin layers parallel to the mid-plane. The third assump-
tion (negligibility of crz) implies that the stress state in each layer may be approximated
as plane stress by components <72, ay, rxy. The material stiffness for the plane stress is
given by (3.41). Using (3.104) we can proceed by (thermal effects were also included at
this point)
3.6. PLATE ELEMENTS 131
Similarly to the analysis of beams, the plate theory also relies heavily on stress re-
sultants, i.e., bending moments, torques, and shear forces. These are computed per unit
length of section through the plate and are consequently called intensities of internal
forces. To emphasize this property, we will denote them by lowercase letters.
The bending moments mx and my and the torsional moment mxy are defined as
Equation (3.105) can be substituted into (3.106) and (3.107). After integration in the
z direction, the resulting relationship between stress resultants and functions w, <px, (py
reads
where
The equations for qx and qy were derived from the assumption that the shear stresses rxz
and Tyz averaged over the plate thickness [compare to(2.1.])] are energetically equivalent
to the parabolic variation of these stresses along z.
We can derive rather simple formula for m^. If the basic geometric assumption con-
cerning the pseudo-normals is to hold, only linear variation of the temperature along z
can be admitted. Denoting the temperature of the upper surface z = -| by Th and the
temperature of the lower surface z = | as T^, then we have for T
As the first term in the expression for the temperature generates membrane forces only,
we find by substitution into (3.108)
132 CHAPTER 3. PLATES AND SHELLS
will be satisfied in an integral form by including them as constraints into the energy
functional.
Let us first write the expression for the potential energy. Let us assume that the
plate is loaded by a continuous loading p only. In addition, the bending moment mn,
the torsional moment mnt and/or the shear force qn can be prescribed at the boundary
of the plate. The system plate-subgrade is shown in Fig. 3.20. Fig. 3.21 depicts the
stress resultants along the boundary of the plate, when viewed from below, i.e., against
the 2-axis. The potential energy can be written as
Figure 3.21: The stress resultants along the boundary of the plate
3.6. PLATE ELEMENTS 133
where qz = JQ Tznijj dz is the energetically equivalent stress resultant of the shear stresses
per unit length of the boundary of the elastic subgrade.
The functional (3.111) will be applied to the solution of plate structures by the FEM,
however, the resulting stiffness matrix will be computed by selective integration. We
could show in obvious analogy to the relations derived for the beam that the bilinear
approximation of <px, (py and w leads through selective integration to the same results
as a generalized variational statement. This variational formulation, originating from an
extension of the Lagrange principle by supplementary constraints that leads to approxi-
mate satisfaction of geometrical equations (3.109), will be handled in detail below, as it is
of general usefulness. Especially higher-order isoparametric plate elements rely on it. A
typical representative is the biquadratic element with eight nodal points. As the functions
7xy, Txz and jyz are independent of functions (px, (py and w, some of the equations (3.108)
change to
To verify that the stationarity conditions of the above functional give all necessary equa-
tions, let us compute its variation
The Gauss integral theorem helps to remove derivatives in the functions subject to vari-
ation.
To further modify the above functional, we need the transformation properties of the
pseudo-normal rotation on the boundary
as well as the transformation rules of the stress resultants on the boundary (see Fig. 3.21)
where a is the angle subtended by the normal to the boundary F with the z-axis.
Using the constitutive equations (3.108) and (3.112) we finally arrive at the variation
of nm as listed in Table 3.5.
3.6. PLATE ELEMENTS 135
geometric equations
equilibrium condition
for the subgrade
boundary conditions
on the plate
boundary conditions
on the subgrade boundary.
Table 3.5: Euler equations derived from the stationarity condition of the energy functional
136 CHAPTERS. PLATES AND SHELLS
where (pxi, (pyi are rotations at the nodal points and at the mid-points of the edges.
The interpolation functions Ni were listed in Section 3.2.1. The element is depicted
in Fig. 3.22, in which the plate is viewed against the positive z-axis (from below).
3.6. PLATE ELEMENTS 137
• On top of specifying the Kirchhoff conditions at the nodes, we enforce them also at
the mid-points of the edges (nodes 4,5,6), where
where k denotes the mid-point of the edge and lk is the length of the side ij (opposite
to vertex k).
• The rotation of the normal (pak about the element side is computed from the as-
sumption of linear variation of (ps along this side
where
The functions #x4, #x5, HxQ, Hy^ Hy$ and HyQ can be obtained from (3.121) by
substituting N% for NI and by replacing indices 6 by 4, and 5 by 6. Similarly, the functions
HX7, Hx8, HxQ, Hy7, Hys and Hy^ can be obtained by substituting AT3 for NI and by
replacing indices 6 by 5, and 5 by 4. The functions N\ through AT6 were detailed in the
section on area coordinates. The remaining coefficients in (3.121) are given as (do not
sum over k)
where k = 4,5,6 for sides 23, 31, 12, respectively. The rest of the derivation follows
standard tracks. The matrix J3, coupling curvatures K and nodal displacements r, by
K — B r, can be derived from (3.104). The terms of the stiffness matrix can be expressed
as / B1D0 B dA, where the matrix D0 is given as
JA
The computational details regarding matrix B can be found in [18]. The matrix D0
is proportional to the isotropic elasticity matrix for plane stress. Extension to orthotropic
materials is given in the Section 1.2.
Let us follow the rotation of the pseudo-normal at an arbitrary point on side 23 about
this side:
The variation of the function (p\ — (p\(£) is obvious from Fig. 3.23b, where (p^\ = <p2. fii.
The nondimensional coordinate £ varies from 0 to 1 between points 23. Let us express (pi
as
The first term corresponds to the skew-symmetric deformation of the element side with
zero deflections and constant shear (see Fig.3.23c). The second term corresponds to
symmetric deformation (pure bending)—cf. Fig. 3.23d. The bending deformation wsyrn\23
can be obtained by integration of
from which
140 CHAPTERS. PLATES AND SHELLS
Analogous results hold for the remaining element sides. If we consider that the LI = (1—£)
and Z/3 = £ on side 23, the bending deformation which ensures displacement compatibility
along inter-element boundary can be put down as
Let us verify that the curvatures derived from (3.126) are identical to those computed from
(3.124). First, the rotations of the reference plane normal are computed from Fig. 3.23
(positive deflections are measured upward)
The projection of the vector of rotation <£fc - c^; into the direction of Hi is expressed by
the dot product
Once the approximations w,wsyrn,(px,(py,(pxy are known, the shear distortions can be
evaluated from
As follows from (3.127), both (pxsym and (pysym vary linearly on the element.
If the shear strains ((px - (pxsym) and ((py - (pysym) are computed by one-point quadra-
ture, the actual linear variation of the shear strains is replaced by its average value. This
can be simply evaluated by computing the strain at point L\ = LI — L$ — —.
o
The computation of the stiffness matrix in the case of the CCT element is thus reduced
to matrix multiplication, as both K and 7 are element-wise constant vectors.
Analogously to the plane stress case, the nodal displacements will be collected in a vector
The form of the functional allows for decomposition of the stiffness matrix into plate
stiffness and subgrade stiffness. These matrices will be simply summed up for elements
representing plate on an elastic foundation.
The pseudo-curvatures Kx,Ky and Kxy and the shears jxz and jyz will be collected in
a vector
Let us express the individual components of the vector K in dependence on the generalized
nodal displacements
142 CHAPTER 3. PLATES AND SHELLS
Evaluation of the matrix B is given next. When the reduced integration will be used,
the third, fourth and fifth line of the matrix B will be computed only at point £ = 77 = 0.
Consequently, the notation introduced before will be applied with this additional definition
The matrix B
Using the above notation, the plate element stiffness can be written as
The formula for the stiffness computation using Gaussian quadrature reads (for 2 x 2
integration selectively reduced to a one-point quadrature)
3.6. PLATE ELEMENTS 143
In order to derive the relations for the subgrade stiffness matrix the derivatives below
must be known:
where
where
It remains to derive the vector of transformed loading. The formulation contains three
kinds of loads:
• distributed load, normal to the plane of the plate, characterized by the intensity JJ;
• temperature change loads, characterized by the temperature of the lower and upper
plate surfaces;
Let us assume that p and ra^ is constant over the element. The vector Rp can be computed
as
where
The deflection is approximated in the same way, only the bubble function is added (details
in Chapter 10)
where
It is obvious that the last term is also at the foundation of the Allman's complement
of approximation functions for membrane quadrilateral element with rotational degrees
of freedom (Section 3.5.4).
3.6. PLATE ELEMENTS 145
The plate quadrilateral element is an analogy to the beam solution. The rotation of the
pseudo-normal to the reference surface is approximated by usual isoparametric functions
The two-dimensional interpolation functions Ni = Ni(£, 77) are specified in Table 3.1 (bold
frame). The deflection is analogously to (3.90) expressed as:
where the functions Nj are adopted from Table 3.1 without the coefficients 0.5., IJK is
the length of side JK and OLJK is the angle between the outside normal to the side JK
and the rr-axis, where
where
for
146 CHAPTERS. PLATES AND SHELLS
Thin plates and shells suffer from shear locking, therefore a correction of the shear ap-
proximation is in order to avoid it. Additional constraints of type (2.71) provide constant
shear along the sides only, however. It is in general nonconstant in the interior. Let us
modify the expression for 7:
where 7C is unknown parameter, which will be viewed as initial deformation. The first
term in (3.145) can be constant (zero), but the second term gives in general a nonzero
value of 7 for pure bending. Consequently, it is necessary that
This condition cannot be exactly satisfied at all interior points, and so a weaker condition
will be adopted instead. The energy functional will be augmented by
It follows that
where Db is the (3,3) material stiffness for bending and twist and D3 is the (2,2) shear
stiffness of a section of unit width. Matrix Bb is constructed by regrouping the first three
rows of matrix B in (3.136). It holds for orthotropic materials:
The terms of the material matrices are defined by (1.25). It is possible to use these
formulae for the description of geometric orthotropy (resulting from different structural
arrangments in two perpendicular directions).
Reference [88] presents comparison with yet another quadrilateral plate element with
12 degrees of freedom—element Tl designed by Hughes and Tezduyar [82]. Tl is identical
to MITC4 due to Bathe and Dvorkin [14]. It can be shown that all the above-mentioned
elements possess the same properties. The element presented here is logically analogous
to the CCT triangle.
Example 3.1
A general comparison is very difficult. To aid the reader, we present a sample solution to
3.7. SHELL ELEMENTS 147
a plate problem. The structure from reference [15] has been chosen for this purpose. It is
a skewed cantilever plate, which has been analyzed both numerically and experimentally.
The structure is shown in Fig. 3.24 together with finite element meshes.
Solution:
The solution is presented for the DKT and CCT triangles, for the modified quadrilateral MQ
and for the HSM element, which is a hybrid element for thin plates with linear approximation
of moments and cubic approximation of deflections along element edges. The plate is thin
Figure 3.24: Skewed cantilever plate. E = 1.05 x 107, h = 0.125, v = 0.3,p = 0.26066
(h/L = 0.0104). Consequently, it is obvious from liable 3.6 that the best results have been
obtained with elements based on Kirchhoff theory—DKT, HSM. However, even the elements
CCT and MQ perform with acceptable accuracy. It should be noted that the latter elements
are more universally applicable. It is also worth while to note that the triangulation type shows
a nonnegligible influence on accuracy for the CCT element.
measures changes in distance of the nodes along the edges (Fig. 3.26a). The vector
components
are the pure rotations of the normals to the reference surface at the nodes about the
local x— and y—axes. To parallel the approach of Section 2.3 as closely as possible, the
components of the vector $; shown in 3.26a will be expressed as
Here, (rijxi n^), and (tjx, tjy) are components of the unit vectors normal and parallel to
the side opposite node j.
With the help of (3.149) a generalization of equation (2.93) in the form
is obtained, where
3.7. SHELL ELEMENTS 149
The constants c{ and c^ are computed from (2.92) for length lj and angles a{ and o^
which correspond to the arc opposite node j.
In the membrane/plate element the reduced matrices
Using the first row in the transformation matrix (1.30) we get the matrix relation between
the edge extension and the Cartesian components of the strain tensor
or
The matrix K\ follows from the equality of the alternative expressions for the potential
deformation energy
The complete vector of generalized nodal displacements in the local coordinate system
which includes also the rigid body motion can be written as
150 CHAPTER 3. PLATES AND SHELLS
The following transformation must be known [compare with (2.60)] in order to augment
the stiffness matrix
or
The derivation of the matrix T\ is easy. It is sufficient to project the nodal displace-
ments w,-, Vi onto the edge directions (the extension of an edge is given by the difference
in projection lengths):
To derive the matrix T2, it must be taken into account that (compare with Fig. 2.8)
where t/>x and ^y are angles by which the element rotates as a rigid body about the local
axes x and y, while
The area coordinates Li on a triangle were discussed in detail in Section 3.2.1. Using
relations (3.22) and (3.23) we have
The extended stiffness matrix of the shell element in the local coordinate system is of
type (15, 15) and it is expressed as [compare with (2.99)]
where K'l = T^XiTi is the membrane stiffness matrix of type (6, 6) corresponding to
the vector of nodal displacements r\ and K2 — T^K2T2 is the bending stiffness matrix
of type (9, 9^ corresponding to the vector of nodal displacements r2. It is obvious that
the matrix K2 need not be computed.
3.8. INTERACTION BETWEEN STRUCTURE AND FOUNDATION 151
An element with these properties has been presented earlier in the sections on various
modifications of the plane stress element. This adjustment is of minor impact: the matrix
B will simply have nine rows instead of six—the additional rows will be all zeros.
The local stiffness matrix is in our case of type (18,18). The rows and columns needs to
be reordered so that the nodal displacements follow the ordering of nodes on the element
The relationship between the local and the global vector is given by the transformation
matrix A (see Section 2.6)
The transformation matrix consists of six diagonal submatrices A* of type (3,3), which
contains the cosines of angles between the local axes x, y and z and global axes X, Y and
Z:
When the classical membrane element with stiffness matrix K\ of type (6,6) is used,
the degrees of freedom (piz (i = 1,2,3) are associated to zero stiffness. Consequently,
the extended matrix KI of dimensions (18,18) will have three zero diagonal terms. After
the matrix is transformed into global coordinate system according to (3.167), the zero
diagonal terms usually become nonzero. In that case, the element can be applicable.
The other component of the interacting systems are the foundations. Flexible plates
on an elastic subgrade were discussed in Section 3.6.4. Stiff foundation beams on two-
parametric subgrades will be the subject of an investigation below. Also the interaction
between adjacent foundation beams will be addressed.
The stiffness matrix of the beam element II (i.e., end-point element) can be obtained
from matrix K/ by adding a complementing matrix
which relates the shear stiffness of the regions A and B—see Fig. 3.27. The derivation was
based on unit displacement at node 2 and the approximation below was used to describe
the deformed surface.
region A
(3.170)
region B
Element III may be applied in models of walls with openings. It corresponds to the
bending stiffness of a beam on an elastic foundation. The stiffness matrix is a sum of the
3.8. INTERACTION BETWEEN STRUCTURE AND FOUNDATION 153
stiffness of the beam [torsional stiffness neglected in (2.101)], and the stiffness matrix of
the subgrade from (2.110).
Element IV corresponds to places where the beam is interrupted (i.e., there are gaps
between the foundation elements). Numerical experiments have shown that from the
viewpoint of soil-structure interaction there is no substantial difference between the use
of formula (2.110) and (3.168). Additional options will be discussed in the next section.
We have assumed so far that parallel foundation beams do not interact. On the other
hand, the interaction of crossing beams cannot be neglected. The correction at the inner
corner will be done under the simplifying assumption that the displacement does not
include rotation of the contact surface about the beam axis. The displacement of the
subgrade at the inner corner corersponding to a unit displacement is expressed as
The correction (negative) force at the inner corner of region C will be computed from
the principle of virtual forces from the formula
The force RQ would appear at the inner corner of the region C, if the interaction of
crossing orthogonal beams was neglected (additionally, the surfaces of the subgrade were
assumed to be cylindrical). Because of the first two terms in the approximation (3.171),
the force RQ also included in the double integral (3.172). Both terms thus cancel and the
integration of the remaining two terms gives
The inclusion of the derived force in the equilibrium conditions (in the stiffness matrix)
needs no comments.
The shear flexibility of the fields A, B and C and its influence on the interaction can
be captured in two ways:
(a) The classical variant considers each field to be a substructure, which is covered by
isoparametric elements. Quadrilaterals with bilinear approximation of vertical deflections
perform satisfactorily (see field B). After elimination of the internal degrees of freedom,
a condensed stiffness matrix is obtained, which needs to be transformed so that the
continuity conditions on the boundary are met. The transformation matrices are obtained
from the kinematical relations between vertical nodal deflections a, 6, ...,ra, (Fig. 3.29)
and the deformation parameters of the upper part of the building (compare with the
Section 3.4), which are the primary unknowns in the JADRO program.
(b) The second variant enables one to obtain the stiffness matrix of the fields A,B,C
directly from the special choice of the basis functions, which are derived from the differen-
tial equation of the two-parametric subgrade model and which respect the linear variation
of the displacements on the boundary. This variant, which is preferred in the program
JADRO, uses mapping on the unit reference element (Fig. 3.30). The vertical deflections
The basis functions for the three element types from Fig. 3.29 have been summarized in
Table 3.7.
Example 3.3
The adequacy of the described model will be shown on an example of stress calculation for
the spatial wall structure from Fig. 3.31, which is supported by foundation beams and is
loaded vertically and horizontally (by wind forces) in the directions of the Y- and Z-axes.
3.8. INTERACTION BETWEEN STRUCTURE AND FOUNDATION 155
The plan shows the most important foundation structures (III—continuous beam, IV—
interrupted beam) and "subgrade" elements (A, B, C) capturing the interaction between
the foundation beams.
Solution:
The computation was done in three variants
a) Winkler model:
(3) Two-parametric model:
7) Two-parametric model:
The results are presented in the following figures. Figure 3.32 shows in isometry the normal
stresses at the contact surface between the foundation structures and the subgrade for the
vertical load. Only one quarter of the structure is shown because of symmetry. A peculiarity
of the two-parametric model is the stress concentration at the end-points of the interrupted
beams, which shows in the computation as a concentrated shear force. It is an analogy to the
stress concentration under a rigid punch on an elastic foundation. This phenomenon is taken
into account in the JADRO program for the purpose of stress calculation by substituting linear
3.9. PATCH TEST 157
equivalent stresses for nodal forces. The concentration shows partially also for a transition
between a rigid beam into a flexible one (type III), even though it is much less pronounced than
in the case of an interrupted beam (type IV). The effect can also be (even more pronounced)
traced in the cross beams over openings 1,4,6,7 (Fig. 3.33).
The influence of the subgrade is less significant for the horizontal load. It does not affect
substantially neither the contact stresses, nor the shear forces in the cross beams.
applied at the external nodes. The displacements at the interior nodes are solved for nodes
by FEM. The classical Irons patch test requires that not only the displacements but also
strains (stresses) be computed exactly for an arbitrary size and shape of elements.
Figure 3.35 shows results for two patch tests for the plane stress isoparametric element.
The displacements in Fig. 3.35a correspond to constant tension, and those in Fig. 3.35b
correspond to constant shear. In both cases the FEM gave exact values for displacements
and stresses.
Recently, two questions have been posed: Is the patch test a necessary condition for
convergence? Practical experience shows that some elements do not pass the patch test,
158 CHAPTER 3. PLATES AND SHELLS
and still converge to the correct solution. One example has been published in [166],
another in [28]. There is a tendency to formulate the patch test in a weaker form. The
first attempt has been presented in [166] by introducing patch test, which requires that
the error be of order 0(/i) or higher with diminishing patch size h. The weak patch test
is described by Fig. 3.36. It was, however, shown by Belytschko and Lasry in [28] that
a patch test formulated in this way is passed not only by the sound elements, but also
by elements that do not converge to the correct solution. Another form of the patch test
was proposed in the same paper, which is based not on diminishing patch size, but on a
diminishing size of elements within the patch—compare with Fig. 3.37. The patch is a
square and point Ag is placed off the center. The points ^45, A$, AT, AS are located at the
mid-points of the sides. Points B\ to B$ are obtained by bilinear mapping of the square
AiA2A$A4 onto the quadrilateral BiB^B^B^. A sequence of subdivisions results (see Fig.
3.38). As long as the subdivision yields converging results, weak consistency is satisfied
3.9. PATCH TEST 159
and the tested element is applicable. An example is the isoparametric plate element from
Section 3.6.4, which does not in general pass the classical patch test. Starting from a
certain ratio of thickness to span, it passes the weak patch test according to Belytschko
and Lasry [28] and consequently it is applicable to plates on an elastic foundation as long
as they are not too thin (approx. h/l > 0.05),
Chapter 4
Solids
The analysis of solids was slightly neglected for a long time. Not that it was not necessary
to investigate three-dimensional bodies, but the demands on computer resources were
excessive. The situation has recently changed. Current engineering workstations and
parallel computers allow for solutions even of large problems of complex three-dimensional
solids. As far as the FEM analysis is concerned, the formulations of the planar problems
are simply extended into the three-dimensional space. Let us adopt this approach for the
description of the fundamental finite element groups:
The simplest representatives of the individual groups are depicted in Fig. 4.1.
4.1 Tetrahedra
The simplest tetrahedron has four nodes. Each of the nodes is associated to three degrees
of freedom—displacements u, v and w along axes x,y and z of the global coordinate
system. Consequently, the element possesses 4 x 3 = 12 degrees of freedom. The dis-
placements inside the element are approximated by linear functions, and the strains (and
stresses) are thus constant within the element. The interpolation functions can be advan-
tageously expressed in volume coordinates. The definition of the volume coordinates is
clear from Fig. 4.2a. For instance, coordinate LI is computed as
160
4.1. TETRAHEDRA 161
The four coordinates Z/x, L2, £3 and Z/4 are tied together by the relation Li+L2+L3+L4 =
1. A hierarchy of tetrahedral elements can be generated by increasing the degrees of the
approximation polynomials. The quadratic and cubic interpolation elements are shown
in Fig. 4.2b,c. The following interpolation functions are used.
4-noded element (linear)
Vertices
Mid-points of edges
Mid-points of faces
The geometry of the discussed elements is described by
The advantage of the tetrahedron becomes apparent especially when solving linear prob-
lems, as the necessary matrices (stiffness, mass) can be obtained in closed form
The integration cannot be done using formula (4.2); it must be performed numerically.
162 CHAPTERS SOLIDS
4.2 Bricks
The bricks are, similarly to quadrilaterals in planar problems, very popular. Commercial
FE programs usually use linear and quadratic elements. It is of advantage to introduce
isoparametric coordinates (£, 77, £) (analogous to the two-dimensional case), which map
an actual brick onto a cube with side of length 2 as shown in Fig. 4.4. Interpolation
functions are defined as follows.
To compute the stiffness matrix it is necessary to know the derivatives of the inter-
polation functions with respect to x, y and z, the Jacobian matrix of the transformation
between (x,y,z) and (£, 77, C):
and the determinant of this matrix (Jacobian) to be able to transform the elemental
volume:
identical to the linear brick. It has six degrees of freedom at each nodal point. It is a
new approach, so it might be too adventurous to judge its quality. Once its good perfor-
mance can be verified, it will be possible to conclude that a very effective, uniform means
for modeling of structures composed of solid, surface and line elements is available. The
element geometry is described by relations of type (4.3). The interpolation functions for
the displacements were derived from those for the 20-noded brick
where Ati; are the hierarchical displacements and N{ are given by (4.4) to (4.6): formula
(4.4) holds for / = 17,18,19,20, (4.5) applies to / = 9,11,13,15, and (4.6) applies to / =
10,12,14,16. An approach similar to the planar element leads to the hierarchical degrees
of freedom expressed by displacements and rotations at the nodes. The modification is
most easily explained, e.g., on the interpolation function for edge 1-2. Let us introduce
a local coordinate system, which is represented by vectors l\i,m\i and n\i- Vector l\i
points in the direction 1-2, and vectors 77112 and n12 lie in the plane normal to the vector
164 CHAPTERS SOLIDS
Ii2 (see Fig. 4.5)1. These vectors can in general be defined for arbitrary edge defined by
vertices J and K. We have in matrix notation
The hierarchical displacements in the direction rhjK and nJK (normal to the edge) are
eliminated in the same manner as in the planar case; hierarchical displacement along the
edge (in the direction of I JK) is eliminated from the condition that the total displacement
in the direction 1JK is given by the average of the displacements at the nodes J and K.
This condition is consistent with the requirement of linear variation of the displacements
along the edge. After introduction of these conditions into (4.7) we get
where
IJK is the length of the edge J-K. Rotation of the infinitesimal neighborhood u?r =
{ux, Ljy,uz} of an internal point of the element is a function a priori independent of u and
is approximated by
As can be seen from Fig. 4.5, the resulting element has 48 degrees of freedom. In
case the stiffness matrix was derived from the functional (3.87), a mixed model would be
obtained and an approximation to the skew-symmetric part of the stress tensor r would
have to be introduced. As T appears in (3.87) only in the function values, there is no
need to ensure continuity on the inter-element boundary. Therefore, let us introduce
where
where
1
Vector rai2 is parallel to the plane (x,y).
4.3. BRICK WITH ROTATIONAL DEGREES OF FREEDOM 165
where the summation is performed over all edges meeting at node i. The matrix Gi in
the form defined above would lead to a shear locking element. It is therefore necessary to
modify it analogously to the procedure introduced for the plate element.
It is also necessary to compute
where
The functional (3.87) contains the expression Vu — u>, whose approximation over the
element is
Substitution from (4.10), (4.11), (4.9), (4.12) and (4.13) into the functional (3.87)
gives the matrix
where
166 CHAPTER 4. SOLIDS
and R is the vector of the transformed load. As a includes the internal factors only, we
can apply static condensation and arrive at
Static condensation can be avoided by starting from functional (3.89), which leads to a
displacement model. Substituting into (3.89) we get an expression which differs slightly
from (4.16):
where
While the matrices K and H in (4.14) and (4.17) are computed by 14-point quadra-
ture, proposed by Irons in [89], the matrix P is integrated by 2x2x2 Gaussian quadrature.
The parameter 7 was introduced in Section 3.5.4.
The constitutive equations are given by the material stiffness matrix. For an isotropic
material we have
Similarly to the planar case, isoparametric elements are very popular. As far as stiff-
ness matrix and transformed load is concerned, the resulting expressions are similar to
those above, with the exception of the integration, which is carried out on a unit sector:
So far, we have dealt with structures under loads which were changing so slowly that
the inertial forces could be neglected. However, there are many loads in the engineering
practice, which can cause vibration. These loads can be subdivided into
• loads due to rotating machinery,
• moving loads,
• impact loads,
• fluid flow loads, especially wind loads,
• earthquake loads, and
• blast loads.
In order to assess the influence of these loads, a serious dynamic analysis is essential.
This can be split into two basic tasks:
• eigenvibration analysis, and
• forced vibration analysis.
Linear analysis of the onset of instability is a problem analogous in formulation to the
problem of eigenvibration: Given a spatial stress distribution, a scalar factor A is sought
which gives the critical load when the given stress distribution is scaled by this factor.
168
5.1. BASIC NOTIONS AND RELATIONS 169
to sum up to the total element mass. M denotes the total element mass and ma are the
diagonal terms of the mass matrix, which correspond to translation. This approach is
advantageous especially for isoparametric elements. The coefficient 6 is taken as 8 = 1
for line elements, 6 = 2 for two-dimensional elements and 6 — 3 for three-dimensional
elements.
and which are due to the inclined orientation of the normal force N.
The initial stress matrix will be derived from the virtual work equality according to
Fiff. 5.1. We have
1
1£ the element possesses also rotational degrees of freedom, the correction is carried out for the
translational degrees of freedom only.
170 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
Analogously to the mass matrix, the initial stress matrix for the classical beam element
is expressed as a multiple of matrix Kb, of (2.110).
The initial stress matrix for a plate is obtained in an analogous fashion from
where u is the circular frequency of undamped vibration and y is the mode of eigenvibra-
tion. These quantities are discussed in detail in Section 5.1.5. The above homogeneous
system of equations describes the eigenvibration of a structure with initial static stresses.
Let us now consider a proportional loading R = XRv, where RQ describes the spatial
properties of the load, and A is the proportionality parameter. If the matrix Kff cor-
responds to the load RQ, the matrix \Ka corresponds to the load R. Given a spatial
distribution of load, the eigenvibration frequency can be zero for some parameter A. This
corresponds to the case of linear instability,
The problem (5.4) can be solved by any of the methods described in Section 5.2. It
is possible to compute n eigenvalues A; from (5.4). The most significant is the lowest
eigenvalue, A 1? which corresponds to the lowest value of the critical load.
Computing r and substituting into (5.5) the basic equation of undamped eigenvibration
is obtained,
Equation (5.7) represents the generalized mathematical problem of finding a set of eigen-
values for matrices M and K. It is well known that (5.7) possesses nontrivial solutions
if
172 172 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
The matrix K is always banded. The mass matrix can be, according to the way in which
the inertial forces are approximated, either banded (consistent mass matrix) or diagonal
(concentrated mass). The consistent mass matrix M is always positive definite. The
diagonal mass matrix is positive definite only if ma > 0 for alH = 1.,. n. If some term
ma = 0, the matrix M is positive semi-definite.
If both K and M are positive definite, all eigenfrequencies are positive. If M is
diagonal with m zero terms on the diagonal (M positive semi-definite), m eigenfrequencies
flnnrnarVi nlns infinit.v
Multiplying (5.11) from the left by the vector -yj and equation (5.10) by yf, we get
and the eigenmodes are orthogonal with respect to the mass matrix.
Let us first multiply the equation (5.12) by 1/cj2, and the equation (5.12) by
and then sum them. After some manipulation we have
Thus, the eigenmodes are also orthogonal with respect to the stiffness matrix.
Let us normalize the eigenmodes with respect to the mass matrix.2 We require
2
The advantage of the normalization by the mass matrix becomes apparent especially when the re-
sponse of a vibrating structure is sought by the expansion into eigenmodes.
5.1. BASIC NOTIONS AND RELATIONS 173
Whenever, the eigenmodes are used in the following, it is always assumed that they satisfy
(5.17), i.e., that they are orthonormal with respect to the mass matrix. We can therefore
drop the subscript norm. Substituting (5.17) into (5.12), the relations
This means that among all vectors t/, the first eigenmode is the vector minimizing the
Rayleigh quotient. This theorem is called the Rayleigh principle.
The Rayleigh principle can be generalized to higher eigenmodes. It holds that (Jl =
p(yi)-
5.1.8Spectral decomposition of the stiffness matrix
If the eigenmodes, y^ and the eigenfrequencies, a;?, are available, then
The correctness of this relation can be easily proved. Let us solve the static problem
K r = R by an expansion into the eigenmodes. Then
3
The vector y{ is given by (5.5) not only by its direction, but also by its magnitude. However, its sign
is indeterminate. The condition in (5.17) holds also for -y{.
174 CHAPTERS. LINEAR DYNAMICS AND STABILITY
Substituting (5.25) into Kr = R and multiplying the resulting equation from the left by
yj we get
Let us substitute for r from (5.25), where first # is replaced from (5.26). We get
so that the correctness of (5.24) has been proved for arbitrary -R.4
As can be seen from the preceding exposition, the eigenvibration of structures leads to
generalized eigenvalue problem (5.7) when the finite element method is applied. There-
fore, before the classical methods can be used, the generalized problem (5.7) must be
transformed into a standard one. This can be done easily if M is diagonal with all terms
nonzero. If M is not diagonal (consistent mass), the transformation to (5.29) is more
complicated and also more sensitive to numerical accuracy. The details can be found in
[17]. As we are convinced that the eigenvibration problem is best solved in the general-
ized form, we will not discuss the transformation to the standard form. A special place
among the classical methods (from the algorithmical point of view) belongs to the Jacobi
method of rotations. The opportune property of bandedness of K and M should be
taken advantage of to save computer resources.
and massless degrees of freedom, yb. The matrices K and M are split into submatrices
corresponding to the two groups of degrees of freedom:
The vector yb can be expressed from the second equation (5.31) in dependence on ya as
Substituting (5,32) into the first equation (5.31), we get the reduced problem of eigenvi-
bration
It is obvious from (5.33) why the number of mass degrees of freedom should have
been kept low: The matrix Ka is in general full, even though K is banded. To solve
the reduced problem (5.33) efficiently, it is necessary to fit the whole matrix Ka into the
fast part of the computer memory. The main application of this method is for structures
with large concentrated masses, which lend themselves to a natural selection of the mass
degrees of freedom.5
5
The numerical computation of Ka is based on Gaussian elimination, not (5.33), to avoid inversion
ofKbb.
176 CHAPTERS. LINEAR DYNAMICS AND STABILITY
The numerator and denominator in (5.35) are scalars. As usual in the Ritz method, the
coefficients c are unknown. Therefore, p(<p) is a function of Q, i.e.,
The necessary condition for a minimum is that the first partial derivatives with respect
to Ci vanish,
To avoid writing (5.35) as a summation, let us introduce vector 9p(<p)/9c, whose terms are
partial derivatives dp(<p)/dci. Let us adopt the notation K = &TK& and M = *FTM&.
The expression (5.35) can be put down as
Because dp/dc = O, we get from (5.39) a condition for c in the form of a system of linear
homogeneous equations
That means that by the Ritz method we have reduced the original problem with n degrees
of freedom into an eigenvalue problem with p degrees of freedom. The reduction was
effected at the price of some loss of precision. The eigenvalues of (5.40) pi,p2»".,Pp are
the (upper) approximations of the squares to the eigenfrequencies, i.e.,
Let us note that if the eigenmodes GI will be normalized with respect to M, tpi will be
normalized with respect to M.
Rayleigh-Ritz method is approximate. If we select for the basis ^ the correct eigen-
mode, say y^ then the Rayleigh-Ritz method finds this vector as the best approximation,
and we have ujj = pj.
The accuracy of the higher frequencies and eigenmodes as computed by the Rayleigh-
Ritz method is reduced with respect to the lower modes.
which is analogous to (5.34). This transformation can be written for the whole structure,
substructure, or for an isolated element. Thus, using (5.43), the matrices K and M will
be transformed from the generalized coordinates t/ a , yb into the generalized coordinate
ya. We get
Substitution of (5.43) into the first relation of (5.44) leads to K* = Ka [Ka from equation
(5.33)].
The numerical implementation of the condensation process can be advantageously
based on these two alternative ways:
1. Condensation of the mass matrix is carried out by the Gaussian elimination. The
term —K^Kba in the transformation (5.43) is replaced by the Gauss multipliers.
where t is a vector of Gauss multipliers, which are needed for zeroing of a single
column.
Substituting (5.45) into the second relation (5.44) we get
Mtb is a scalar for the condensation of the single degree of freedom. The described
approach is sometimes called the Guyan reduction. As the algorithm of mass matrix con-
densation is practically very useful, the source code to a subprogram for the concurrent
condensation of the stiffness and mass matrices is given below. The degrees of freedom
that are to be condensed out must be the last in the list.
IMPLICIT REAL*8 (A-H.O-Z)
c
c TEST
c CLAMPED BEAM
c RIGHT ROTATION CONDENSED
c
DIMENSION S(4,4), H(4,4)
DATA S/12, -6, -12, -6,
* -6, 4, 6, 2,
* -12, 6, 12, 6,
* "6, 2, 6, 4/,
* H/156, -22, 54, 13,
* -22, 4, -13, -3,
* 54, -13, 156, 22,
* 13, -3, 22, 4/
c
CALL CONDEN(S,H,4,1)
c
WRITE(*,100) ((S(I,J),J=1,4),I=1,4)
WRITE (*, 100) ((H(I,J),J=1,4),I=1,4)
100 FORMAT (4F7.1)
178 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
C
STOP
END
C
SUBROUTINE CONDEN(S,H,N,M)
IMPLICIT REAL*8 (A-H.O-Z)
DIMENSION S(N,N),H(N,N)
C
C SUBROUTINE FOR CONSISTENT CONDENSATION OF MASS
C OR INITIAL STRESS MATRICES
C (CONDENSED DOFS HAVE TO BE PLACED AT THE LAST
C POSITIONS OF NODAL DISPLACEMENT VECTOR)
C
C S(N,N) - STIFFNESS MATRIX
C H(N.N) - MASS OR INITIAL STRESS MATRIX
C N - ORDER OF MATRICES
C M - NUMBER OF CONDENSED DOFS
C
C PROGRAMMED BY Z.BITTNAR
C JULY 1992
C
C STIFFNESS MATRIX CONDENSATION
C
DO 10 K=1,M
L = N-K
MM = L+l
DO 20 11=1, L
C = -S(II,MM)/S(MM,MM)
DO 30 J=1,L
30 S(II,J) = S(II,J) + C*S(MM,J)
CONTINUE
20 S(II,MM) = C
C
C MASS MATRIX CONDENSATION
C
DO 40 11=1, L
DO 40 J=1,L
40 H(II,J) = H(II,J) + S(II,MM)*H(MM,J)
* + H(II,MM)*S(J,MM)
* + H(MM,MM)*S(II,MM)*S(J,MM)
C
10 CONTINUE
C
RETURN
END
Example 5.1
Let us compute the consistent mass matrix of a beam clamped at one end (Fig. 5.2)
by a condensation of the mass matrix of a fully clamped beam.
Solution:
5.2. METHODS OF EIGENVIBRATION ANALYSIS 179
Figure 5.2: Beam clamped at one end only, and fully clamped
Let us adopt pAl = 420, EJ/l3 = 1, / = 1 for simplicity. The stiffness and mass matrices
[corresponding to rT = {^1,^1,^2,^2}] can tnen be written as
The elements of t are obtained from the condition that the end moment M^i on the fully clamped
beam is zero. It follows
Let us note that the condensation includes the influence of shear flexibility if this influence is
given in the stiffness matrix.
which holds for y being an eigenmode with u the corresponding eigenfrequency. Because
neither the mode nor the frequency are known, we have selected an approximation of the
180 CHAPTERS. LINEAR DYNAMICS AND STABILITY
mode xi. The amplitude of the inertial forces (up to scaling a;2, which does not influence
the mode, because it scales homogeneously the whole vector of inertial forces) will be
computed as
The structure is loaded by these forces, and the displacement vector x2 that represents a
new approximation of the first eigenmode t/1, is computed from
The next step consists in the repeated computation of the inertial force amplitude
due to displacement x2, etc. The cycle is repeated until the vectors x^ and x^+i are
sufficiently close. Thus, an eigenmode is obtained. As long as Xi is not orthogonal to
the first eigenmode yl5 i.e., x[ Myl ^ 0, then Xk+i -» yx. The approximate frequency is
obtained from the Rayleigh quotient
Vector xjfc+i can be normalized, e.g., by dividing all its components by the absolutely
largest one. However, as the most convenient to work with are the vectors normalized
with respect to M, we set
The formulae (5.51) and (5.52) describe the iteration cycle for the inverse iteration. We
can compute the Rayleigh quotient from (5.50) to get an approximation to the eigen-
frequency. This is convenient especially for convergence tests. If the eigenfrequency is
desired with a precision of s digits, it must hold
In [17], Bathe and Wilson have proposed the following algorithm of inverse itera-
tion, which minimizes the necessary arithmetical operations: Let us select as the initial
approximation the vector Zi = Mx^ The iteration cycle is described by the formulae
5.2. METHODS OF EIGEN VIBRATION A NA LYSIS 181
The above approach leads to Zk+i -> Myl and p(xk+i) — > u\, if y\z\ 7^ 0.
The coefficient ci is computed from the condition that XQ be orthogonal to y1? i.e.,
= 0. Let us multiply (5.55) from the left by y,M. Then
The approach can be generalized. If the first m eigerimodes are known the initial approx-
imation XQ for the computation of the (ra -f l)th eigenmode must be removed from the
first m modes. We have
It can be shown (see [32]) that the inverse iteration converges in such a case to the mode
with frequency ujj closest to 7.
The original iterative process was proposed by Jacobi already in 1846. The method has
gained in popularity with the appearance of modern computers, especially because of its
simplicity (and its easy programming). It is also very advantageous from the engineering
viewpoint, because it solves the generalized eigenvalue problem. As the method has been
frequently used in many programs, we will discuss it in some detail.
Let us consider the standard eigenvalue problem
Let us assume that A is positive definite, and A is the eigenvalue. The Jacobi method, as
mentioned above, is an iterative method. The basic idea rests with the transformation of
A to a diagonal form, or, more precisely, to an almost-diagonal matrix, as only a finite
number of iterations will be done. In the case of the standard problem, the diagonal
terms are equal to the eigenvalues. At a given iteration step we transform by the rotation
matrix
ith row
?th row
It can be easily shown that this matrix represents a rotation of a plane given by the ith
and jth basis vector (and consequently of the whole space) by the angle <£, with cos (p = c,
sin (p = s. Such matrices appear also in the transformation of end forces of a beam from
the local into the global coordinate system. The whole algorithm resembles a calculation
of principal stress directions or principal directions of the tensor of inertia in a plane.
The matrix T is orthogonal, and thus T~l = TT (see, e.g., [61]). The iterative process is
based on the construction of a sequence of transformed matrices
where
The eigenvalues will be obtained from the diagonal terms of A, as already mentioned
above. The eigenvectors will be computed from the product of transformation matrices
After the principles of the transformation (5.61) have been given, the strategy for the
selection of the pair of indices (i,j) remains to be specified. Jacobi proposed a selection
based on the largest nondiagonal term. This is unsuitable for computer implementation,
however, as the search for the largest term could easily be more expensive than the trans-
formation itself. Alternatives have been sought, which are more suited for computation.
Practical experience has shown that cyclic processes are very suitable. We choose (i,j) of
the zeroed term cyclically, basing our decision on a pre-specified order. The computation
travels usually along rows or columns. A disadvantage is that small terms get zeroed even
though large numbers are still present at off-diagonal locations. This disadvantage can
be removed by cyclic process with limiters.
Let us consider a sequence of numbers 61,62,..., monotonically approaching zero. The
pair (i,j) is selected according to a pre-specified order (e.g., by rows). Before the trans-
formation (5.61) is realized, we check whether the term a$ is smaller than 61. If so, the
transformation is not performed. If all the off-diagonal terms are smaller than 6 lt the
limiter is reduced from b\ to 62 and the process is repeated. The convergence of the algo-
rithm described is quadratic, as in the original Jacobi strategy of the largest off-diagonal
term. Quadratic convergence means that if all off-diagonal terms are smaller than £, the
diagonal terms are given with error smaller than e2. The cyclic process with limiters leads
also to easier programming, as the difficulties due to d being close to zero in (5.65) are
avoided, and c and s can be computed accurately.
We have listed practically important facts about the Jacobi method, but so far only for
the standard problem. Structural dynamics leads to the generalized formulation (5.40).
The advantage of Jacobi method consists in that, after a certain extension, it can accom-
modate the direct solution of the general eigenvalue problem. The basic idea is derived
184 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
from the analogy of matrices (2,2). Let us consider matrices A and B. It is easy to find
C such that CT AC and CTBC are diagonal. It is sufficient to take
with
To compute a and ft it is sufficient to exchange i for 1 and j for 2 in (5.68) and (5.69).
The problem (5.40) will be solved by concurrent diagonalization of K and M. We list
below formulae analogous to (5.62). The terms of the transformed matrix K or M will
be given in the notation
When both K and M are diagonalized, i.e., all the off-diagonal terms of K an d M are
sufficiently small, the squares of the eigenfrequencies can be computed as kk^ /m^ . The
7
If h = 0, then a. = 0 and (3 =
5.2. METHODS OF EIGEN VIBRATION ANALYSIS 185
and
CL11U.
i.e., by quantities which give the coupling between the degrees of freedom i and j.
is applied, will converge to the matrix composed of q times repeated first mode of the
problem (5.9). It can be shown, however, that the subspace corresponding to Xk converges
to Vq°°, whose basis is given by q first eigenmodes yl, t/ 2 > ••• » yq- To explain this, consider
a geometrical example. Let Vn be the space of all three-dimensional vectors. The formula
(5.9) fixes in this space a triple of vectors that are mutually orthogonal in the generalized
sense. Let us select q = 2. The first two eigenmodes, corresponding to the two lowest
frequencies, define a plane, which in this example represents a subspace F2°° —see Fig.
5.3. Let us consider two arbitrary linearly independent vectors x}, x\, which will be
adopted for the basis of V2. When iterating by (5.72) the two vectors rotate so that,
after a sufficiently large number of steps, they point in almost the same direction—that
of the first eigenmode. The vectors xf, x\ define a subspace VJf, which converges to V^00
with growing k. In one word, the plane defined by xf , x\ converges to the plane given by
the basis t/ x , t/2. This explains the seemingly paradoxical statement that V* converges
to V^°°. At the same time, it is clear that as long as vector y2 is to be given with arbitrary
precision by a linear combination of x{, x*, it is necessary to have unlimited precision of
computation at our disposal.
Holding on to the geometric analogy, it is easy to arrive at an efficient combination
of the simultaneous inverse iteration and the Rayleigh-Ritz method. The latter means
geometrically that we search in the plane VQk for two vectors normal to each other, which
are closest to the vectors y l 5 y2 in the sense of the Rayleigh quotient. Thus, if at each step
of the inverse iteration we find by the Rayleigh-Ritz method new orthogonal vectors in
Vf, the coalescing of vectors xf , x\ is prevented, and also difficulties with the computer
precision are alleviated. Let us assume that x\ and x2 are selected as a linear combination
186 CHAPTERS. LINEAR DYNAMICS AND STABILITY
°f 2/i» 2/2- Such vectors xj, x^ belong to the subspace V9°°. The Rayleigh-Ritz method
then finds the correct solution in a single step. The convergence rate of all p requested
modes can be enhanced simply by iterating on a subspace of dimension q > p. It was
recommended in [16] to use the heuristic formula q = min(2p,p + 8). One consequence
of an increased rate is a higher number of arithmetic operations within one iteration
step. The operations per one step of the subspace iteration method (i.e., of the transition
B. Subspace iteration
a
b) Computation of matrices M£, Kk
e) Repeat step B. a)
C. Sturm check
from V* to Vq*1) are summarized in Table 5.1 taken over from Bathe and Wilson [16].
It remains to answer one important question— how to choose the starting vectors Xi —
which influences strongly the computation time. One such algorithm has been proposed
in [16]. A vector with unit displacements at all mass degrees of freedom is chosen as the
starting vector Xi. The remaining (g - 1) vectors have only one nonzero component—
in the direction of the displacement with the largest ratio ma/ku where ma, ka are the
5.2. METHODS OF EIGENVIBRATION ANALYSIS 187
diagonal terms of the mass and stiffness matrices, respectively. This algorithm guarantees
that M in the Rayleigh-Ritz method will be positive definite.
which is in general not orthogonal to g^g2 , ••• 9j> but includes in addition to the
orthogonal vectors Zj also contributions of all preceding vectors g^g2 ••• • 9j- This can
be mathematically written as:
It will be shown that all the coefficients in (5.74), starting with 7^, are zero. Further, we
will derive relations for otj, fy and show that the (j + l)th Lanczos vector can be expressed
as
It is obvious from (5.75) that the vector Zj is orthogonal to all Lanczos vectors as
long as it is orthogonal to the last two (# 7 ,0 J _ 1 ). As the proof of this fact gives at the
same time a hint how to compute a;, /?,, it will be given in detail below. To compute the
coefficients 0^/^,7^ the orthogonality conditions will be used. We multiply (5.74) from
the left by gjM, so that
The first term on the right-hand side is zero. This follows from the definition Zj (it is
orthogonal to all preceding Lanczos vectors). The third term and all the following ones
8
This is a basis of vectors normalized with respect to the mass matrix M.
188 CHA PTER 5. LINE A R DYNA MICS A ND STA BILITY
are also zero because of the orthogonality gi,g% ••• 9j, and the second term is, due to
M-orthonormality, equal to otj. Consequently,
The relation for fa can be derived similarly if (5.74) is multiplied from the left by gJ^M.
In that case all terms on the right-hand side vanish with the exception of the third term.
It follows that PJ = gJ^M'Zj. It is possible to substitute for ~Zj from (5.73), which gives
PJ = gJ^MK^Mgj. It follows from (5.73) that zj^ = gJ^MK'1 and we have
Finally, zj_l can be expanded following (5.74) and substituted into (5.78). Then
All terms on the right-hand side are zero with the exception of the first term. As we have
we obtain after substitution into (5.79) and after simple modifications the relation for Pj
[compare with (5.78)],
Numerical experiments have shown that formula (5.81) is more appropriate than (5.78).
Lanczos vectors are sufficiently normalized even if ^J+1M^J_1 is not exactly zero.
It can be shown in the same manner that 7^ — g?_2M~Zj and an analogous expression
to (5.79) can be constructed,
It is obvious that 7j = 0 because of orthogonality. It can be shown that all the other
terms in (5.74) are also zero. It follows that it is sufficient to orthogonalize the Lanczos
vectors only with respect to the last two vectors, QED.
The algorithm of the Lanczos method is summarized in Table (5.2). The algorithm
requires in addition to the memory to accommodate K and M also storage of flf^, Qj, Rj
and Zj.
where e^ = [0,0, ...1] and Gm are matrices (m, m), whose columns are given by the
Lanczos vectors ^. Tm is a tri-diagonal square matrix
Relation (5.84) can be used to reduce the number of degrees of freedom when solving
for the lowest eigenfrequencies. We carry out the transformation of equation (5.7). The
original degrees of freedom y are replaced by generalized coordinates c,
190 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
where c is the vector of coefficients of the linear combination of the Lanczos vectors.
Substitution into (5.7) and multiplication from the left G^MK'1 gives
Using (5.84) and taking into account the M-orthogonality of Gm we arrive at the reduced
eigenvalue problem
Convergence criterion
The solution of the reduced problem (5.86) gives approximate eigenfrequencies ^c^,
which converge with growing ra to the exact values. To be able to compute a given number
of frequencies with a desired accuracy, it is necessary to have suitable error indicators. As
discussed above, the formula (5.53) is often used to evaluate the accuracy of the solution.
It has been demonstrated in [80] that this criterion sometimes fails and a more reliable
guess (estimate of the accuracy of the ith eigenfrequency after m steps) is proposed which
is based on
The relations (5.87) and (5.88) can be further modified by the transformation (5.85). Let
us multiply (5.82) by the eigenvector ^GI of the reduced problem (5.86). We obtain
From (5.86) it follows that C;. We obtain a simple formula for the
(m)
residue Si by substituting from (5.85). We have
It is obvious that it is simpler to compute the norm of the residue (m ^Sj from the right-
hand side of equation (5.90) than directly from (5.88). At the same time, e^ (m ^Cj is a
dot product, and its value is given by the last term of the zth normalized eigenvector of
Tm, which will be denoted as (i- The norm of the residue is therefore given by
where f3m+i is the value computed within Lanczos algorithm. Let us denote by ^Pi the
norm of the residue ^Si. It then holds
Loss of orthogonality
The derived algorithm of the Lanczos method requires theoretically to orthogonalize at
each step only with respect to the last two Lanczos vectors. However, limited precision
arithmetics leads to a loss of orthogonality. This can be detected by the matrix
where the ijth term of the matrix Hm is 77^ = g^Mg^. Absolute precision would make
Hm a unit matrix. Finite precision arithmetics gives errors (77^1, which we require for
i ^ j to be of the order of the so-called unit truncation error e.9 A recurrent formula can
be derived for the columns of the matrix Hrn, analogous to (5.75) for the construction of
the Lanczos vectors. We get (details in [80], [156])
Another check for loss of orthogonality was proposed by Paige (see, e.g., [80]). Instead of
testing mutual orthogonality between the Lanczos vectors, the orthogonality of the new
vector with respect to the previously computed eigenmodes is tested:
Exact arithmetics should give zero. Paige has demonstrated that finite precision arith-
metics gives
where (j) means that the quantities were obtained for the jth Lanczos step, and 7^,- is a
number approximately equal to one. It follows from (5.96) that sudden drop in /3j+i\£i\
signals a substantial loss of orthogonality. As the expression j3j+i\^\ is a measure of the
accuracy of the eigenvalue (5.91), it is obvious that convergence to the eigenfrequency
leads to a loss of orthogonality.
It holds theoretically that if two following Lanczos vectors are orthogonal to an eigen-
mode t/j, all subsequent Lanczos vectors will be also orthogonal to the same eigenmode.
This statement is not valid in finite precision arithmetics, which produces a phenomenon
discussed already in relation to the inverse iteration.10 The loss of orthogonality is there-
fore to be checked at each step. It is therefore necessary to find an inexpensive and
accurate estimate of TJ — ^yjMgj. This can be done by using (5.82), from which the
recurrent formula
can be derived. The value of TJ+I must be computed for all converged eigenmodes.
9
The unit truncation error e is defined as the smallest number represented in the computer for which
it holds that
10
It was demonstrated in the Gramm-Schmidt orthogonaJization that when computing higher eigen-
frequencies by inverse iteration it is theoretically sufficient to remove the lower modes from the initial
approximation. It was shown in [32], however, that finite precision arithmetics would force convergence
to the first eigenmode.
192 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
Reconstruction of orthogonality
The simplest way of ensuring orthogonality of the newly computed vector gf,-+1 is the
removal of the preceding Lanczos vectors by the Gramm-Schmidt orthogonalization. This
approach is called full re-orthogonalization. If only a few eigenvalues are needed (the
number of Lanczos steps is smaller than half the bandwidth of the stiffness matrix K),
this approach can be applied successfully. The full re-orthogonalization satisfies
This means that the re-orthogonalization need not be done with respect to all Lanczos
vectors, but only with respect to those for which condition (5.99) does not hold. There
are two approaches by which the selective re-orthogonalization can be realized:
1) Orthogonalization with respect to computed eigenvectors
The values of TJ are computed for all computed eigenfrequencies from (5.97). As soon
as \TJ\ > i/e, re-orthogonalization must be carried out, and the detected eigenmode y^
must be removed from QJ+I.
2) Orthogonalization with respect to preceding Lanczos vectors
The measure of loss of orthogonality are the terms of vectors hj+i. As soon as the absolute
value of any term of the vector hj+i is larger than v/6, orthogonality must be reconstructed
between QJ+I and gk.
In [80], Hughes recommends a strategy which is a combination of the two above ap-
proaches. When choosing between 1) and 2) the reasons for the loss of orthogonality are
found. They are either
a) convergence to an eigenfrequency, or
It is recommended that the approach ad 2) be used in case a), and the approach ad 1) is
prefered in case b).11
We compared the Lanczos method and the subspace iteration in an example of a
framed structure of Fig. 5.4. The first 30 eigenmodes were computed. The problem has
1,200 degrees of freedom with half bandwidth of 306. The full re-orthogonalization was
used in the Lanczos method. It turned out that the subspace iteration was faster in this
case. The solution has been obtained on an IBM RS/6000 model 320 workstation in 55
seconds.
11
It is appropriate to note here that the coefficients TJ, r)ij require about one half of the computer time
which is necessary for full re-orthogonalization. It follows that the selective re-orthogonalization proves
advantageous for large problems.
5.2. METHODS OF EIGEN VIBRATION A NA LYSIS 193
The solution of this homogeneous differential equation will be sought in the form r(t) =
ext y. Substitution into (5.100) leads to a characteristic equation
where A and y are the eigenvalue and eigenvector of the system; A and y are in general
complex. This quadratic eigenvalue problem may be transformed to a linear one by
doubling the dimension of (5.101). We obtain
where
The matrices A, B are symmetric, but not positive definite; A is the eigenvalue (in general
complex). The imaginary part of A represents the circular frequency of damped vibration;
the real part is the damping measure.
To solve (5.102) the Lanczos method can be applied with a modification as formulated
in [44]. It consists in extracting the Krylov sequence from
The vectors Zj can be orthogonalized similarly to Section 5.2.10. Coefficients a;- and /3j
are given by
where
194 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
The time-dependence will not be written explicitly to simplify the notation. Substi-
tution of (5.106) into (5.2) gives (the initial stress matrix is not considered)
5.3. FORCED VIBRATION OF LINEAR SYSTEMS 195
As the eigenmodes are orthogonal in the sense of (5.20) and (5.21), the equation (5.108)
can be transformed into
To be able to compute the vector qr 0 , the first equation in (5.110) will be multiplied from
the left bv YTM< so that
The choices for r* are wide. The deformation variant of FEM uses, e.g., all generalized
displacements of support nodes equal to zero, with the exception of those with prescribed
5.3. FORCED VIBRATION OF LINEAR SYSTEMS 197
Function 77(t) gives the time variation of the support motion. Vector rs is chosen such
that
Then r3 expresses static deformation of the structure due to static motion of the supports.
Equation (5.119) is solved by the method of expansion into eigenmodes. It holds for
rv that
Substitution of (5.122) into (5.119) and multiplication of (5.119) from the left by YT gives
Further, we apply to (5.123) the conditions (5.20) and (5.21). The expression YTCr* will
be neglected as a damping of the nonperiodical part of the motion. Finally, the expression
YTKr* is due to (5.121) equal to zero. The equation (5.123) will be rewritten by using
(5.119) into
Equation (5.120) introduced the limiting condition for the motion of the supports. It
could happen occasionally that the time variation of the motion is different for different
supports. In that case the principle of decomposition may be applied. The response to
the external loading and to the motion of all the supports is computed separately, and
the results are summed.
The integration constants a,-, k are computed from the initial conditions. The Duhamel
integral for nonperiodical loading is as a rule computed numerically. The reason is that
either it is not feasible to obtain the integral for the given function fa = /,(r) in closed
form, or function /,- is given by a table. To obtain a complete solution to (5.2), all n
eigenmodes must be used in (5.106). This means solving n equations (5.125). If we
use the same integration method, there is no difference between the method of mode
198 CHAPTERS. LINEAR DYNAMICS AND STABILITY
decomposition and the direct integration of (5.2). We assume, though, that the algorithms
give the same numerical error.
So why should use the method of mode decomposition at all? The reason is that
in most cases it is sufficient to use only a few dozen of eigenmodes, especially for beam
and plated structures. The necessary number of modes in (5.106) depends, however, also
on the frequency characteristics of the loading. The general wisdom is (see [17]) that t
follow the wave character of the solution, the expansion should be based at least on |n
eigenmodes. It is obvious that to solve problems of this type, one of the direct integration
methods is preferable. On the other hand, the mode decomposition method is useful
especially for earthquake, wind and other low-frequency loadings.
where uJ; = w^l - $ is the frequency of the damped vibration. The approach is similar
to that of the preceding section.
To solve the problem of nonproportional damping (the matrix YTCY is nondiagonal),
four approaches can be used (we assume that it is sufficient to include only p <C n
eigenmodes):
(A) Weighted coefficients of relative damping
This approach is based on a diagonalization of the matrix YTCY', which is com-
puted from the weighted coefficients of relative damping, and where the off-diagonal
terms are neglected. The technique for the computation of the weighted coefficients
£i is discussed in [129], which includes examples of calculations showing that valu-
able results can be obtained from realistic values of &. However, this approach is in
general not very useful, as it masks the fundamental difference between nonpropor-
tional and proportional damping.
(B) Diagonalization due to complex eigenmodes
It is possible to diagonalize the matrix YTCY even for nonproportional damping if
the complex eigenvalues are used. To compute the complex eigenvalues the method
of inverse iteration or the Lanczos method can be used (after some modification).
The importance of complex modes lies in the fact that they supply two kinds of
5.3. FORCED VIBRATION OF LINEAR SYSTEMS 199
information: firstly about the shape, and secondly about the phase shift. To better
understand the difference between proportional and nonproportional damping, an
analysis of a structure loaded by support motion is carried out here. We monitor
only one component of the vibration, corresponding to the eigenfrequency u^. Pro-
portional damping gives the same mode of vibration, with amplitude diminishing
exponentially with time. The nodes of the modes do not move (they are stationary).
All points of the structure vibrate in the same phase. The situation is different for
nonproportional damping, where not only the amplitude decays, but also the in-
dividual points vibrate with different phase shifts. Consequently, the nodes of the
eigenmodes move (they are now nonstationary). Similarly to proportional damping,
the solution can be obtained by including only a relatively small number of modes.
The disadvantage is that we have to solve the eigenvalue problem
where 17, Q , Y are matrices, whose terms are complex numbers. This means
higher memory requirements, and also the number of arithmetic operations grows
with respect to solution of (5.9). How to simplify the solution for steady vibration
is shown in the next section.
Equation (5.130) can be solved by the same direct integration methods as the orig-
inal problem (5.2). Details are given in Section 5.3.3. This approach can be com-
petitive if only a few eigenmodes are included in (5.106). Interaction of structure
and subgrade under earthquake loading is a typical example.
Vector equation (5.132) is solved by iteration following the recipe (i is the iteratior
number)
with ql — O. The advantage lies in the fact that we solve only a system of ordinary
differential equations of second order.
200 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
where vector / describes the spatial distribution of the loading, and the function g = g(t)
determines the time-dependency. Further we assume that we have solved for the first p
vibration eigenmodes, and that the approximate solution of (5.2) can thus be obtained
by eigenmode expansion. The fact that only a limited number of eigenmodes is used
in (5.2) (corresponding to the lowest eigenfrequencies) can be interpreted that an exact
solution to a modified loading has been obtained.12 This situation can be mathematically
described as
which shows correctness of (5.134). It enables us to quantify the error in the loading Re
at the same time. It holds that
The error in the response due to only the first p eigenmodes being used can be computed
from
To solve for the response of the structure re to the loading Re by expanding into the first
p eigenmodes is not possible, as Re is orthogonal to these modes. If the p eigenmodes
cover well the load spectrum, it is possible to neglect re and re when computing the
response to Re. Then re will be computed approximately from
Substitution into (5.137) and multiplication from the left by rs gives an ordinary differ-
ential equation
WHERE
12
The spatial variation of this load can be obtained as a linear combination of the lowest p eigenmodes.
5.3. FORCED VIBRATION OF LINEAR SYSTEMS 201
Vector re obtained by coupling (5.140) with the formula (5.139) is called the
dynamic correction.
Extension of the method of mode decomposition by the corrections of this paragraph
makes the solutions considerably more accurate, and especially so for concentrated loads.
The approach is easily implemented in extant programs for dynamic analysis of structures.
is called the integration step. Important part of equation (5.2) constitute the initial
conditions. Let us consider the initial time t = 0. Then it holds that
Let us assume that we know approximate solutions at time instants to, t\, ti,..., tk-i* Using
this information, we shall establish the solution for tk. If successful, the approach defines
a method of direct integration, as the time tk can be any of the sequence t0, ti,..., tm. Let
us denote the typical time instant and step as t + At, t, At instead of tjt+1, t*, At*. It may
seem superfluous to deal with methods of direct integration, when numerical mathematics
has been concerned with these methods for a long time. The reason is similar to that given
when discussing the eigenvalue problem: Equation (5.2) is a special case of a system of
differential equations. The FEM solves very large systems of these equations. Therefore,
it is advantageous to formulate specialized methods, which are economical and fast.
We will discuss several methods of numerical integration which are suitable to solutions
of linear and nonlinear equations (5.2). All of them are one-step methods (with the
exception of the central difference method), i.e., the characteristics of the motion r,f,f
are computed from the known quantities at time t. One of the advantages of one-step
methods is that no special approach is needed to start the integration.
Let us introduce a certain classification of the direct integration methods before we
delve into details. As a rule, three types of methods are recognized:
• explicit,
• implicit, and
• predictor-corrector.
While the first two methods are really basic types, the third method is in fact a variation
on an implicit method. It is listed as a separate type because it assumes a special role in
nonlinear problems. An integration method is either explicit or implicit according to the
time instant at which equation (5.2) is applied.
In an explicit method, the vectors rt+Atj^t+At^t+At are computed from the known
motion characteristics r, r, r in the interval < t, t -f At > by using the equation of motion
202 CHAPTERS. LINEAR DYNAMICS AND STABILITY
(5.2) at time t. No factorization of the stiffness matrix is done in explicit methods, but
the mass matrix is inverted. Sometimes the stiffness matrix is not assembled at all. That
is the reason for using explicit methods predominantly in combination with a diagonal
mass matrix. There are certain difficulties associated to massless degrees of freedom in
that case, so they need to be eliminated either by static condensation, or by converting
them into mass degrees of freedom by the addition of very small masses.
On the other hand, implicit methods use equation (5.2) at time t + At. They are
advantageous especially for linear equation (5.2) with a consistent mass matrix.
As we mark the methods as stable or unstable, let us define briefly the notion of
stability. The solution is stable if it does not grow for arbitrary initial conditions to
infinity. If this condition is satisfied for arbitrary At/Tn, the method is unconditionally
stable; otherwise, if it is satisfied for limited At/T n , the method is conditionally stable
(Tn is the shortest period of eigenvibration).
where Tn is the shortest period of vibration. There are thus two contradictory require-
ments: We have to
• use a diagonal mass matrix with concentrated masses, and
• satisfy condition (5.147).
One of the consequences of the concentrated masses being used to model the inertial
properties are the massless degrees of freedom, i.e., zero terms in the diagonal mass
5.3. FORCED VIBRATION OF LINEAR SYSTEMS 203
matrix M. Even when these terms get replaced by small numbers, very short periods
corresponding to the massless degrees of freedom result, and the time step is much too
short. This in turn makes the method overly expensive. Another disadvantage is the
need to use a special starting procedure. In that respect the method deviates from the
scheme of one-step methods. Despite these facts the method deserves attention, as it was
successfully used under different guises in a number of computations (especially in wave-
propagation problems, and for nonlinear problems in general), as all the disadvantages
are balanced bv the fact that the stiffness matrix needs not be assembled.
Newmark method
This method is implicit. It was published in the original version by Newmark in [123] as
the "average acceleration" method. The basic formulae of the Newmark method, which
specify the relation between displacements, velocities and accelerations at time t and
t + At, are of the form
Now we can substitute (5.148) and (5.149) into (5.150), so that we obtain a system of
linear algebraic equations for
As soon as r f +At is known, we can substitute back into (5.148) and (5.149), from which
r*M-At and r t +At can be obtained. These steps can be repeated m times and the result
would be an approximate solution to the equation (5.2) on the interval < 0, tm >. It is
clear from (5.151) that to compute rt+At the matrix
Wilson 0-method
Similarly to the Newmark method, the Wilson method is implicit. The original version
of the method, which is from today's viewpoint identical to the Newmark method with
(5.148) and (5.149) for 8 = 1/2 and a = 1/6, has been through a long development. It
started as a method of linear acceleration, which is a special case of the so-called Wilson
0-method (9 = 2). The basic idea is a linear variation of the acceleration on the interval
< t, t + O&t >, as shown in Fig. 5.5. For an arbitrary time instant it holds that
We obtain by integration
The equations (5.156), (5.157) are analogous to (5.148), (5.149) in the Newmark method.
We could proceed similarly, i.e., substitute (5.156) and (5.157) into the equation of motion
at time t + 0At and compute from here r t +0At- We can easily obtain rt+0At, **t+0At, f*t+0At
by using (5.153), (5.154) and (5.155) after substitution of r = At (we assume Rayleigh
damping).
The described algorithm requires again storage of large matrices. This can be removed
(similarly to the Newmark method) by an appropriate substitution. Additionally, it is
possible to arrange the algorithm so that it is identical up to the constants a 0 ,ai,...,a 10
to the algorithm of the Newmark method. This enables the programmer to easily include
both methods in one program. The detailed flowchart taken over from Bathe and Wilson
[16] is given in Table 5.3 for both methods. The stability is dependent on the coefficient
9. To make the method stable, it is necessary to have 9 > 1.37. Detailed derivation of
formulae of Table 5.3 can be found in [32].
5.3. FORCED VIBRATION OF LINEAR SYSTEMS 205
Wilson method
b) Compute r* from
where fm represents rath term of the vector YTMr3r}(i). Let us denote fm(t) = Rmij(t}
and let us write the solution of (5.158) by a Duhamel integral. We obtain
The time variation of the response corresponding to the rath mode is given by the integral
in (5.159). This integral is of dimension meter/second, and its maximum is called the
spectral velocity and denoted as Sv:
Plotting the dependency of Sv on the frequency u and the relative damping £ for a
particular accelerogram, we get so-called response spectra. These can be used to compute
the maximum of the response corresponding to the rath eigenmode
The values qmtmax can be applied to compute maxima of arbitrary mechanical quan-
tities Sm,maz (deflection, stress), corresponding to the rath eigenmode according to the
relation
where Sm is the value of 5, corresponding to the rath eigenmode. The maxima for the
individual modes do not occur at the same time. Simple addition of Sm>max would be
much too conservative. Therefore, to estimate the maximum, the following formula is
used:
This is very simple, and it gives good results as long as the structure does not exhibit
very closely spaced eigenfrequencies, and as long as all supports vibrate in approximately
the same phase. DerKiureghian has derived a more general formula,
where
5.4. RESPONSE TO HARMONIC EXCITATION 207
This formulation is based on the theory of random processes. It includes the correlation
of the individual eigenmodes, and it can be used even for structures with closely clustered
frequencies. The research report of the Earthquake Engineering Research Center of the
University of California #08 of 1991 presents further generalizations that include also
the influence of local geological conditions in addition to the correlation of the individual
modes.
• by mode decomposition.
In the case of a steady harmonic vibration the usual parlance uses amplitude and
phase shift to describe the load. In that case the nth component of the vector R(t) can
be expressed by using amplitude an and the phase shift (pn in the following manner (see
Fig. 5.6):
The relation between an, (pn and the components of the vectors RI and jR2 is given by
As the solution is sought for a steady state, in which the eigenvibration has already
vanished from the response, the time-dependency of r can be written as
208 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
Vectors HI, R%, r l 7 r 2 are time-independent. Substitution of (5.165) and (5.168) into the
equation of motion (5.2) leads to two simultaneous matrix equations for the unknowns r\
and r2,
The monitored quantities in a steady harmonic vibration are the resonance phenomena.
It is therefore essential to introduce damping into the computation. As a consequence,
the response of the structure r always possesses both phases, even when the applied loads
act in phase. In that case it only holds that R2 = O.
where
One can verify the correctness of the first relation (5.170) by substituting from (5.171),
We obtain
After modification,
which is identical to (5.165). Correctness of the second relation (5.170) can be proved in
a similar fashion^
If the vector R is shown in the complex plane, the relation between the amplitude an,
the phase shift (pn and the components Rin, R<2n becomes obvious. It is clear that they
represent two different ways to write a complex number (see Fig, 5.6).
Figure 5.6: The relationship between amplitude, phase shift and parts of a complex
number
It is sufficient to satisfy (5.2) at each time instant for both R(t) and r(t), to make R
and r comply with
5.4. RESPONSE TO HARMONIC EXCITATION 209
The equation (5.174) is formally identical to the equilibrium condition in a static solution
However, the matrix K is replaced by the complex matrix
which is called the dynamic stiffness. The algorithm of the solution is the same as in
statics, but all operations must be done in complex arithmetics.13
where q is the vector of unknown complex coefficients. Substitution of (5.177) into (5.174)
gives
Multiplication of (5.178) from the left by YT leads (with the orthogonality conditions
applied) to the system of linear algebraic equations with complex coefficients
where J?2 is a diagonal matrix, whose terms are the squares of the circular eigenfrequen-
cies.
Proportional damping
The system obtained for a general damping matrix C consists of simultaneous equations,
and the following section is devoted to its solution. In many practical cases, the damping
can be considered proportional to mass arid stiffness, so that
13
It is appropriate to note that the matrix of dynamic stiffness K is a function of the forcing frequency
u. It is necessary to repeat the computation for each frequency.
210 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
Using (5.177) and (5.183) the relationship between the vector of output quantities
(displacements, internal forces, stresses, etc.) and the load vector can be formally writter
with matrix G in the form
Equation (5.184) holds also for more than one right-hand side; only the vectors R and
v are replaced by matrices R and V.
The only matrix that changes with different frequencies u in equation (5.184) is the
matrix Q, whose terms can be for different u recomputed from (5.185). For the real and
imaginary part of Q we have
Equation (5.184) can then be rewritten without using complex numbers as two equa-
tions for two phases of the output quantities
Vectors (or matrices, for more right-hand sides) YT R\ and YTR2 can be computed
beforehand, independently of given frequencies.
The algorithm for the solution of forced vibration by mode decomposition is therefore
very simple for the case of proportional damping. It can be included in any computer
program which is able to compute the eigenmodes normalized with respect to the mass
matrix.14
Nonproportional damping
Interaction of the structure and the subgrade is a typical example of a nonproportional
damping. The damping in the upper part of the system differs considerably from the
damping of the foundation. This is due partly to different material properties, partly
to the dissipation of energy into the semi-infinite half-space. Therefore, it is not always
possible to accept the damping model of (5.180). There are two options with respect to
the mode decomposition method:
14
The algorithm can also be used for an analytic solution by the Kolousek exact deflection method.
The eigenmodes are normalized with respect to the mass /i (mass per unit length of the beam), which
means J* [iWiWjds = 6ij.
5.4. RESPONSE TO HARMONIC EXCITATION 211
• Consider the mode decomposition method only as a means for reducing the dimen-
sion of the problem, and solve (5.179) according to Section 5.4.1, or
The first alternative does not need a detailed explanation. The second one, on the
other hand, requires solution to a problem of damped eigenvibration. This is rather a
demanding task for larger systems, and the majority of programs does not include this al-
gorithm. There is a simple approximate solution based on the condensation by undamped
eigenmodes Y, besides the Lanczos method (Section 5.2.11). The nonproportional damp-
ing leads to the homogeneous system of algebraic equations
where
The vectors qt are characterized not only by amplitude, but also by phase for the
nonproportional damping case. If p modes of undamped eigenvibration are used to reduce
the dimension to an approximate system (5.190), is is necessary to solve 2p equations with
real coefficients. Each eigenmode is thus characterized by the vector of amplitudes and the
vector of phases. This is the consequence of the fact that the nonproportionally damped
structure has no stationary nodes of vibration. The method has been proposed in [83]. It
is based on the simple idea that the p original equations are complemented by additional
p equations
The terms of the square matrix of (5.194) are real, but the matrix is nonsymmetric.
If the damping is subcritical (which is the case for building structures), the solution of
(5.194) gives 2p eigenmodes with eigenvalues A. The eigenvalues are complex conjugate
with a negative real part. The imaginary part represents the circular frequencies of the
damped vibration. The numerical solution of the eigenvalue problem (5.194) can be
obtained by using standard algorithms from scientific subroutine libraries. The authors
have used subprograms of the library SSP (Subroutine Scientific Eackage), marketed by
IBM. The system matrix (5.194) is full; its size, however, permits the whole computation
to be done in-core.
212 CHAPTER 5. LINEAR DYNAMICS AND STABILITY
where
Now, the transformation into eigenmodes is repeated once again. The transformation
can be written as
where A is the matrix of coefficients of the damped vibration eigenmodes and c is the
vector of coefficients of the linear combination. Substitution of (5.198) into (5.197) and
multiplication from the left by AT gives
where
As the eigenmodes of the damped system are orthogonal, we have that the matrices A~>
13 are diagonal. Therefore (5.199) is a system of linearly independent algebraic equations
with complex coefficients. Components of c are given by
If the solution produces not only modes for generalized displacements, but also for
stresses (internal forces), then the expression Re[/\cetcjt] can be used to compute the
corresponding mechanical quantities.
Part II
This page intentionally left blank
Chapter 6
Semianalytical Methods
Engineering structures often have constant geometric and material properties along a
certain direction. Typical examples are prismatic plates and box girders, frequently used
in bridge engineering. From the numerical point of view, bridges also have favorable
boundary conditions—they are simply supported at both ends, and box structures are
usually stiffened by end diaphragms. In addition to line structures, axially symmetric
structures also fall within the category of structures that can be efficiently solved by a
combination of the FEM with Fourier series expansions. Such an approach is called a
semianalytical method. It was first applied by Grafton and Strom in [65] to the solution
of axially symmetric shells. The semianalytical approach was later extended to axially
symmetric bodies in [178], and to prismatic folded plates in [45] and [46]. A general
formulation was given in [75] and [127]. In this chapter, we derive the relations neede
for the solution of rectangular plates, and we briefly describe the general formulation for
curved folded plates.
The semianalytical method transforms the solution of a two-dimensional problem into
the solution of a sequence of one-dimensional problems, and the solution of a three-
dimensional problem into the solution of a sequence of two-dimensional problems. Ex-
amples of two-dimensional structures given in Fig. 6.1 show that the structure is not
divided into elements but into strips. This is the origin of the frequently used term
finite strip method.
215
216 6, SEMIANALYTICAL METHODS
Same as the displacement function to, the load function / can also be represented by a
Fourier series
The evaluation of the coefficients nf is made easy by the orthogonality property (6.2).
Multiplying expression (6.3) by sin —-— and integrating from 0 to I (scalar product) we
get
The coefficients nw can now be determined from the condition of minimum potential
energy, which gives
6.1. ENERGY-BASED BEAM ANALYSIS 217
Example 6.1
Determine the deflection curve and the bending moment diagram of a simply supported
beam for two loading cases:
Solution:
The deflection and the bending moment are expressed as
ad a)
ad b)
Both loading cases are symmetric, and so it suffices to take into account the odd terms.
In practical applications we can deal only with a finite number of terms in the series. The
rate of convergence for the above cases is illustrated in Fig. 6.3. As we can see,
• convergence is much faster for the uniform load than for the concentrated force, and
• in both loading cases, the deflections converge much faster than the moments.
218 6. SEMIANALYTICAL METHODS
where D is the plate stiffness and p is the intensity of the surface pressure.
Functions w and p can be expanded into Fourier series in the y-direction (Fig. 6.4) as
where
and NF is the number of Fourier terms. Substituting (6.5) into (6.4) we get
Multiplying the last equation by sin —— , integrating from 0 to 6, and taking into
account the orthogonality condition (6.2), we arrive at a decoupled set of ordinary dif-
ferential equations for nw(x). The sum equals to zero only if the term in the brackets
vanishes for every n. The obtained equations are supplemented by boundary conditions
at x = 0, x = a.
An alternative approach to the solution of plate bending is the variational method
combined with Fourier expansion. This technique is more general because it is not limited
to rectangular domains and it can be extended to the solution of circular plates and curved
folded plates, as will be shown later.
The formulation of the plate bending problem will be based on the Mindlin theory
of thick plates because the solution based on the Kirchhoff theory has been discussed in
detail in [103].
arbitrary but uniform along the entire edge. Functions w, (px, (py can then be represented
by Fourier series
The next step of the finite strip method is an approximation in the x-direction. Am-
plitudes of the displacements and rotations are interpolated in the same way as in the
FEM, i.e., the functions on a strip are given by
where NU is the number of nodes,1 nw^ n(pxi, H(Pyi are the amplitude values on the
ith nodal line, and Ni(x) is the shape function (interpolation function) of the zth node.2
Denoting
where
we have
where nB? = [n-B^, nB^ is a matrix that determines the pseudocurvatures and shear
distortions corresponding to the generalized displacements at node number i and to the
Fourier term number n,
1
Strictly speaking, we deal with nodal lines rather than nodal points.
2
It is clear from (6.6) that we deal with isoparametric interpolation functions. They will be described
in detail in Section 6.2.2.
220 6. SEMIANALYTICAL METHODS
The stiffness matrix and the load vector can be derived from the PVD. It is assumed
that the load is defined by the expansion
h is the plate thickness, and k reflects the influence of the shear stress distribution across
the thickness (k = 5/6 for a homogeneous plate).
After integration we get
for n — m
for all n ^ m
for all n, m
we derive from 811 — 0 the equation for a generic term of the Fourier expansion,3
3
We use the simplified notation
6.2. FINITE STRIP METHOD 221
where
for ra = n
for m ^ n
Linear interpolation
quadratic interpolation
cubic interpolation
Using the above definitions of NI to A^4 we can evaluate B? and substitute into the
integrals in (6.9). Integration can be performed numerically according to the Gauss
integration formula. Same as for all elements based on the Mindlin theory of thick plates,
we have to handle shear locking. Its adverse effect can be alleviated (for a certain range of
222 6. SEMI ANALYTICAL METHODS
where
224 6. SEM/ANALYT/CAL METHODS
When applying the finite strip method to the analysis of curved box girders we have
to divide the structure into curved strips (Fig. 6.7) and introduce an approximation of
the unknown functions on the strip. The general approximation formula is given by (6.7)
where we now have
The vectors of generalized strains and corresponding internal forces are given by
where
consists of three diagonal blocks—the membrane stiffness, Dm, the material stiffness in
bending, Db, and the material stiffness in shear, Ds.
6.3. CURVED BOX GIRDERS 225
where
The angle /3 is defined in Fig. 6.7. Based on the chosen approximation we can write
where
The very same procedure as the one used in plate analysis leads to the set of indepen-
dent equations (6,9) for each term of the Fourier expansion. The only difference is in the
expressions for the matrices nmKij. Here we have
226 6. SEMI ANALYTICAL METHODS
where n r^ are the nodal parameters corresponding to the symmetric deformation, and n fj
are the nodal parameters corresponding to^the skew-symmetric deformation. Matrices
n
Ni are given by (6.12) while matrices nNi have a similar structure with Sn and Cn
mutually exchanged. Matrices HHKij are given by
Matrices nKij have to be computed using matrices nBi, which are derived from
n
Bmi, nBbi, nBsi by replacing n by -n.
is then the same at all nodes (note that we allow only strips that are not curved in the s
direction). The matrix nK\j related to the global coordinate system is expressed as usual
by
where
Special attention is needed whenever two neighboring strips are coplanar. In such a
case, no stiffness is associated with the rotation about the local z'-axis, and the resulting
stiffness matrix is singular. Before performing the transformation (6.14) we have to replace
the zero term on the diagonal by a positive one. This modification does not affect the
results as the corresponding equation is decoupled from the other equations described by
the stiffness matrix.
However, there exists a more elegant remedy to the above problem: The transformation
to the global coordinate system is not performed at all, and the rotation about the local z'-
axis is excluded from the vector of unknowns r. An additional advantage of this approach
is that it reduces the number of unknowns.
6.4. PLANE STRIP WITH ROTATIONAL DEGREES OF FREEDOM 227
The transformation of the coordinate system can be introduced already on the level
of the B matrix by introducing
Then we have
This is advantageous from the practical point of view because the product DnB+ can
be exploited when calculating the internal forces related to the local coordinates from the
global parameters nr.
We restrict our attention to plane problems and to a strip with two nodes. An extension
to curved strips and axisymmetric structures is given in [31]. The approximation on a
two-node strip reads
When we construct hierarchical elements in Chapter 10, the present interpolation func-
tions Ni,N2,N3 will be denoted by /i,/2,/ii. Their graphical representation is shown in
Fig. 10.2. Starting from (6.16), we can write the kinematic equations as
where
228 6. SEM/ANALYT7CAL METHODS
The functional (3.87) further contains the expression (Vu — u>), which, according to
(6.16), can be approximated by
where
Substituting the derived expressions into (3.87) and evaluating the integrals we get
where4
The parameter TQ can be eliminated from the second equation (6.17) by static conden-
sation, after which we obtain for each term of the Fourier expansion and for an isolated
element an equation of the form
Example 6.2
Evaluate the deflection curve and the stress distribution on a simply supported beam
from Fig. 6.8 by the finite strip method.
Solution:
The mid-point deflections calculated with one to three strips are listed in Table 6.2. The solution
according to the beam theory taking into account the effect of shear gives the value of deflection
7.980.
4
In contrast to the original matrices, matrices with an overbar do not contain the sine and cosine
terms.
6.4. PLANE STRIP WITH ROTATIONAL DEGREES OF FREEDOM
229
The preceding chapters were concerned with the application of FEM to the solution of
classical problems in solid mechanics. We will now show that FEM can be applied to
a variety of problems in mathematical physics closely related to mechanics, which are
analytically described by Laplace or Poisson equations or their generalizations.
This class of problems includes especially the diffusion equation
where / is the unknown function, t is time, arid a, 6 are material constants. This equa-
tion describes nonstationary heat convection (/ = T is the temperature), transport of
moisture (/ = h is the moisture) or transport of pore water in the uncoupled problem of
consolidation (/ = p is the pore water pressure), etc.
If the solution is time-independent (stationary transfer of a medium), (7.1) reduces to
where pz is the distributed load (per unit area) acting on the elastic foundation, has to
be combined with the kinematic boundary condition
1
Boundary conditions assumed to be fulfilled a priori are called essential in variational calculus. In the
problem under consideration, this is the case for the kinematic boundary condition. Boundary conditions
implied by the minimum (stationarity) of a functional are called natural This is the case for the static
boundary condition. The above-mentioned definitions can be found in the classic book by Collatz [48].
230
231
or
Recall that the variation of a functional can be expressed in two alternative forms. We
can write either
The functional (7.6) has the meaning of the total potential energy. Note that we have
used the physical meaning to construct the functional directly. In a similar way, one can
get the energy functionals in the mathematical theory of torsion.
FEM can be based on the functional (7.6) as well as on its variation (7.7). The latter
form can be obtained indirectly, based on the differential equation of the problem [(7.3) for
the present problem] and on the static (natural) boundary conditions [(7.5) for the present
problem]. The approximate solution does not satisfy the equations exactly. We thus
require that they be satisfied at least on the average, with a weight 6w, which is described
by (7.8). This equation is the basis of the Galerkin method and it can also be understood
as the requirement that the virtual work (on the kinematically admissible displacements
6w) done by the unbalanced forces acting inside the region £2, i.e., (C\w — C^w — pz),
and by those acting on the part F2 of the boundary, i.e., (Cidw/dn—!}^, must equal zero.
232 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
Assuming that the torque Mx is prescribed, the total potential energy can be, according
to (7.12) and (7.14), expressed as
The variation can be taken with respect to the two unknown functions 0, ip, and so the
minimum condition has the form
As 0 and -0 are independent, the condition (7.16) uncouples into two equations. The first
equation,
where
is the stiffness moment in free torsion, will be solved after the distribution of the function
V> across the section has been calculated. This can be done by applying FEM to the
second equation,
where
The variational equation (7.21) yields a set of linear algebraic equations for the un-
known nodal values of tp:
234 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
After (7.24) has been solved, we can calculate an approximate value of the sectional
characteristic Ik from (7.18). Making use of the approximation (7.20) and the relations
(7.22) through (7.24), we easily get an approximate value
The procedure is concluded by the calculation of the relative twist angle from (7.17)
and of the stresses based on (7.14) and (7.20):
Let us emphasize that the boundary values of the function i/> are not subject to any
conditions.
Relations (7.27) guarantee that the equilibrium conditions (7.9) are satisfied. The static
boundary condition (7.10) takes the form (Fig. 7.2a)
or
where C is a constant.
Applying Gauss formula to (7.11), we can find two alternative expressions for the
torque:
If the cross section is hollow, its boundary F consists of two parts F = Fe U F» (see
Fig. 7.2b). On the outer part we choose </>e = Ce = 0 while on the inner one we set
7.1. TORSION OF BA RS 235
<^ = d = const. It can easily be shown that the integral along the boundary curve in
(7.29) is equal to minus twice the area bounded by the curve F; (Fig. 7.2b). We can thus
supplement the collection of basic formulae by
To implement the energy approach, let us assume that the relative twist angle G is
given. The complementary energy of the system is then
with the dimension of (m2). The stiffness moment in free torsion is according to (7.30)
Similarly to the stiffness approach, we approximate the function 0* over each element
by
where
236 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
The column matrix N? has zero entries except for the entry corresponding to the bound-
ary value <£*, which is set to one.
The matrix P is formally given by the same expression as K. While K has the
character of a sectional stiffness matrix (after multiplication by Or), the matrix P (after
division by G) has the character of a sectional flexibility matrix.
Having solved the system (7.37), we can calculate the approximate value of the moment
Ik. From (7.34) we get an approximation
The stresses can be recovered from (7.27). Combining (7.33) with (7.36), we get
The proof of the left inequality is simple. It suffices to realize that Ik is only an approxi-
mation of the minimum of Ik. To prove the right inequality, we compare the approximate
value of the functional E*// to the exact value. From (7.31) we can derive the expression
for the approximate value
which must be no less than the exact value -(1/2)G624. Comparison of both results
leads to Ik < h-
We have again verified the well-known fact that a compatible model (here based on
the warping function V) is stiffer than the actual structure while an equilibrium model
(here based on the stress function 0) is more flexible than the actual structure.
Example 7.1
Verify the inequalities (7.42) by calculating the stiffness moment Ik of a square cross
section (Fig. 7.3).
Solution:
The stiffness approach does not pose any a priori condition on the boundary values of the
7.1. TORSION OF BARS 237
warping function. The entire section will be considered as one element and the displacement
function will be approximated according to (7.20) by one term in the form
where
The values of the chosen function N along the boundary F are depicted in the right part of
Fig. 7.3, showing also the points at which N has the extreme values TV = ±1.
We will follow the algorithm from Section 7.1.1 and calculate the sectional stiffness K (the
only element of the stiffness matrix) from formula (7.22) as
Finally, the approximate value Ik is given by (7.25) which after substituting Ip = 2(1/12) (2a)4 =
(8/3) a4 for the polar moment of inertia yields
In the flexibility approach, we will also use a single element and approximate the stress function
by
The unknown parameter j3 will be determined from the minimum condition (7.37) in which
238 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
It thus holds
which corresponds to Fourier law i — — q on the boundary with the outward normal n.
In the case of heat conduction, the constant a is the conductivity2 and q is the heat flux
(rate of heat transfer per unit area of the boundary). In the case of moisture transport,
a is the diffusivity. We must not forget the initial condition specifying the distribution of
the function / inside the body at time t = tQ.
Using the analogy with (7.8), we will apply the Galerkin method to equation (7.43)
and its boundary condition (7.45):
To derive the semidiscrete equations of FEM, we first transform the first integral using
Gauss theorem:
2
In the nonstationary heat conduction problem, a = A is the coefficient of thermal conductivity with
the dimension (Js~1.m~1.K~1) and b = pc, where p is the mass density (kg.m~3), and c (J.kg^.K"1)
is the specific heat. The heat conduction equation is sometimes used in a modified form with a = A/pc,
6 = 1. The constant a is then called the coefficient of thermal diffusivity. In such a case, it is necessary
to modify the boundary condition (7.45) appropriately.
7.2. FE SOLUTION OF DIFFUSION EQUATION 239
On the part FI of the boundary we have according to (7.44) 6f = 0, and so the underlined
term equals zero.
The function / (temperature, moisture) is in FEM approximated on individual ele-
ments by
where
The matrix K is called the conductance matrix and C is the capacitance matrix.
The matrices K, C and the vector q can be assembled from the corresponding element
matrices. Note that the boundary values of the vector r must a priori satisfy the boundary
condition (7.44) prescribed on the part FX of the boundary and that the given function
/ in general varies in time. This fact as well as the initial condition must be taken into
account when equation (7.49) is to be numerically integrated.
Consider the time interval At = t{ — ti-\ and suppose that the solution T{,\ at time
ti-i is known. The vector r = r(t) will be approximated by a linear function
and substitute it along with the approximation (7.53) into equation (7.49):
If we choose a fixed value of r, (7.55) becomes a set of linear algebraic equations for the
unknown components of the vector r,. The terms on the right-hand side are known. To
ensure numerical stability, T must be chosen such that 1/2 < r < 1. The most common
choice is r = 1/2.
240 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
where Ar = TI — r2 is the surface traction on the interface between the liquid and gas
phase of neighboring volumes.
Combining (7.56) and (7.57) we find
7.3. DEFORMATION OF SOILS AND OTHER POROUS MATERIALS 241
The dots denote differentiation with respect to time, Ds is the tangential stiffness matrix
of the porous skeleton, and
is the strain rate of the bulk of the material due to changes of the pore pressure; Km is
the bulk modulus of the matrix of the solid (see [98]).
The rate of total stress is according to (7.61) through (7.63) expressed as
where
For a material without any pores, Ks = Km. For cohesive soils, Ka < Km and a w l .
The described solution is also applicable to long-term deformation of rocks, for which
ot < 0.5, and this fact strongly affects equation (7.64).
To be able to describe the motion of pore liquid, we have to assess the rate of pore
volume changes, 0. Without going into details, which can be found in [181], we simply
list the factors affecting the value of 6:
• The grains increase their volume at the expense of the pores due to the increments
of the effective stress (a?/ > 0).
After superposition of all the four effects and some manipulations, the relation for 6 reads
where v = HI — ii is the vector of relative velocity of the liquid with respect to the skeleton.
According to Darcy 's law, the main driving force is the pressure gradient and the main
resisting force is proportional to the velocity of the liquid (cf. [180]):
where
where p is the soil density. As g = O and the stress rate is given by (7.64) we can write
(7.74) as
Loading and supports of the body define a decomposition of the boundary surface
into two parts F = F3 UF 4 . This decomposition in general differs from the decomposition
F = FiUF2 corresponding to convection. The boundary condition will again be expressed
in a rate form. They are represented by the kinematic boundary condition
The equation of equilibrium (7.75) with its static boundary condition (7.77) is equivalent
to
which can be transformed by using (1.9) and taking into account that transposition doei
not change a scalar expression. Furthermore, the symmetry of the matrix Da is considered
The transformed equation has the form
Equations (7.79) and (7.81) constitute the theoretical foundation for the solution of
the problem by FEM. Let us adopt the usual approximation of the displacement field
To get a consistent solution of the problem [see equations (7.75) and (7.81)], we have
to approximate the strains e and the pore pressure p by polynomials of the same degree.
Strain-displacement equations (1.2) then imply that the displacements are to be approx-
imated by a polynomial one order higher than the pore pressure. If we use, e.g., a linear
approximation of the pore pressure, the displacements require a quadratic one.
The approximation of the strain field can be written as
The approximations (7.82) through (7.84) will be substituted into the integrals in
equations (7.79) and (7.81). Defining matrices
and vectors
Due to the independence of the vectors 6rp and 6ru, the following differential equations
of the first order must hold:
The equations can be integrated numerically. Using formulae (7.53) and (7.54), we trans-
form (7.87) into a sequence of linear algebraic equations
To ensure numerical stability of the solution, the parameter r must be chosen such that
1/2 < T < 1. The most common choice is r = 1/2.
Example 7.2
The behavior of the present model will be demonstrated in a simple case study of interac-
tion between a creeping concrete structure and a consolidating foundation (Fig. 7.5). The
upper part of the system can be thought of as a simple bridge structure. The calculation
has been performed with the following geometric and material characteristics:
246 CHAPTER 7. FB SOLUTION OF SPECIAL PROBLEMS
Upper structure: Its stiffness, defined as the vertical force at the interface with the foun-
dation per unit vertical displacement, equals 10 MNm"1. Creep is modeled by the aging
theory (a single element in Maxwell's chain in Fig, 1.22a).
Foundation:
Eoed = (l~~v)E/{(l+v)(l->2v)} = 26,9 MPa, G = 7.70 MPa, coefficient of consolidation
c = 5.10~3m2day~4.
All boundary surfaces of the soil block except for the interface can be permeated by the
liquid. The displacements of the soil at the bottom surface are prescribed as zero.
Solution:
The solution was based on the Hellinger-Reissner principle. The calculation was performed
under the simplifying assumption that the horizontal displacements were negligible compared
to the vertical displacements in the foundation. For the of consistency, the displacements
should have been approximated by polynomials one degree higher than the pore pressure. In
the present solution, we approximated the vertical displacements as well as the pore pressure
by linear isoparametric functions. Therefore, selective one-point Gauss quadrature was applied
to all matrices containing the pore pressure, i.e., the average values were used in each element.
The results are visualized in several figures:
Fig, 7.6 compares the evolution of the vertical displacements at nodes a and b lying in the
contact plane with the load evolution F = F(t). The integration step At = 0.5 day was selected
according to the loading rate and was kept constant throughout the entire, very short interval
of interest equal to 24 days. For most of the materials such as clays, consolidation takes place
for several years and the integration step has to be progressively increased (cf. Section 1.5).
The evolution of the pore pressure under the footing is visualized in Fig, 7.7. Fig, 7.8 depicts
the growth of the reaction X in time and it shows that, in this case, consolidation affects the
foundation stiffness much more than creep affects the stiffness of the upper structure. Finally,
Fig, 7.9 shows the distribution of the vertical displacements and pore pressure in a column of
soil under the footing.
formed from a flaw of the material structure. This determines the local character of
fracture.
The most frequent reason for decohesion of the material is fatigue fracture which is
caused by degradation processes related to alternate plastic yielding under time-dependent
forces. This section concentrates on brittle fracture which is less frequent but more
catastrophic.3
Fracture mechanics studies a continuum with an a priori crack. Instability and crack
growth can be described by comparing integral fracture parameters with their experimen-
tally determined critical values. The fundamental parameters are Irwin's stress intensity
factor K and Rice J-integral.
FEM is a powerful tool for obtaining these parameters.
3
Practical experience has shown that the probability of brittle fracture increases at low temperatures,
under impact loading and at higher loading rates. Susceptibility to brittle fracture is also related to the
presence of residual stresses and it increases for large structures. A common feature of all failures is
that brittle fracture usually starts at points of high stress concentration due to various types of material
flaws, cracks or structural notches. These so-called stress concentrators eventually lead to the formation
of macroscopic cracks. Such cracks can become unstable, i.e., the material can be abruptly split by
fracture propagating at a velocity close to the speed of sound (in metals in the order of 103 ms"1). Let
us emphasize that an unstable crack does not need any external energy supply for its propagation.
248 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
Three types of stress intensity factors KI, KU, Km correspond to three modes of fracture:
Mode I. - tensile fracture (splitting), governed by
tensile stress a22 = &y
Mode II. - in-plane shear fracture, governed by
shear stress o\i — rxy
Mode III. - antiplane shear fracture, governed by
shear stress <723 — ryz-
Mode I is characterized by the stress intensity factor Kj and nonzero stress functions
7.4. FEM IN FRACTURE MECHANICS 249
Mode II is characterized by the stress intensity factor KU and nonzero stress functions
In the foregoing formulae, v — v for plane strain and v — z//(l + v) for plane stress. The
lateral stress
for plane strain
for plane stress.
We will restrict our attention to a combination of modes I and II. FEM provides us with
approximations of the stress field as well as of the displacement field in the vicinity of
the crack tip. The following approach can be proposed to calculate the factors Kj a KU:
From the three formulae (7.89) for a\i — <7X, 0*22 = oy and 0*12 = Txy select, e.g., the
first two, and substitute the stresses crx, ay calculated by FEM for a fixed value of fi and
different values of r. The resulting set of two equations gives approximate values KJ and
Kff for any r. The correct values KI, KU can be obtained in the limit as
from which
250 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
from which
easily be captured by the boundary element method (Chapter 8). It can thus be expected
that the displacements obtained by the boundary element method lead to more accurate
values of the stress intensity factors KI, KJJ than FEM when used in (7.97) and (7.98).
A substantial improvement of FEM performance can be achieved by introducing singular
dements around the crack tip as shown in Fig. 7.12. In contrast to the standard isopara-
metric element (Fig. 7.12a) whose nodes 2 and 8 are at midpoints of the element sides
the corresponding nodes of the singular element have been shifted to the quarters of the
sides (Fig. 7.12b). Due to this modification, the displacements along the side 1-2-3 are
approximated as
Comparing this result with formula (7.90), we can see that the underlined term directly
describes the y/f singularity. Therefore, the singular element gives better results for the
displacements4 as r -> 0 (Fig. 7.11). In a general direction fl G (0;7r/2), the ^/r-singular
term is not present in the quadrilateral element. Only both extreme cases $ = 0 and
fl = 7T/2 are an exception. To remedy this deficiency, it suffices to replace the quadrilateral
4
If we deal with a large plastic zone ahead of the crack tip (nonlinear fracture mechanics), it makes
no sense to use singular elements.
7.4. FEM IN FRACTURE MECHANICS 251
element by a degenerate triangular element in which nodes 7 and 8 are identical with node
1.
The concept sketched above suggests an alternative calculation of the stress intensity
factors KI and KIf. Substituting (7.99) into (7.97) and (7.98) and taking the limit for
r -> 0, the following formulae5 can be derived:
The nodal displacements vt- can thus be obtained by FEM. Let us point out that the error
of this calculation can be large if the element at the crack tip is too small.
Powerful software packages based on FEM sometimes use elements with an embedded
crack (Fig. 7.12c). There exists a variety of elements of this kind, for which Kj and KJJ
are a direct result of FEM calculations.
With regard to adaptive techniques, which will be discussed in detail in Chapter 10,
we note that FEM yields a very accurate description of the stress and strain fields in
the vicinity of the crack tip even when triangular elements with constant strain are used.
It is, however, necessary to use a large number of elements in a graded mesh, in which
the size of the elements decreases in a geometric series with decreasing distance from the
crack tip. Fig. 7.13 shows a possible mesh arrangement used by Malone, Plunkett and
Hodge to analyze a plate with a central crack subject to tension (see [114]). The mesh in
the close vicinity of point G has been formed by dividing an inserted rectangular region.
Beyond this region, the nodes of the triangular elements lie at the intersections of radial
lines from G with circumferential lines. The calculation led to the following conclusion:
An increase of the number of circumferential layers results in higher accuracy in the
vicinity of the crack tip while an increase of the number of radial lines results in a better
description of the stress state in the remaining part of the plate. By increasing solely the
number of the layers (refining the mesh only in the vicinity of the crack tip), one can
improve the accuracy in the outer region r > lOOrx by less than 0.1% (7*1 characterizes
the thickness of the innermost layer).
Knowing the values KI, KU, or Km, we can assess the stability of the crack. Ac-
cording to the so-called /^-concept, the crack becomes unstable if a certain function
K = K(KI,KH,KHI) assumes its critical value Kc- If one fracture mode dominates,
K = Ki and the constant Kc = Kic, where i = /, //, III.6
• Irwin-Orowan ^-criterion.
• Criteria related to Rice ^-integral.
where
where Rk and Kkk are the terms on the right-hand side and the diagonal coefficients of the
factorized stiffness matrix after forward Gauss elimination. The summation is performed
over all degrees of discretization labeled by the subscript k.
The relationship between the Tf-concept and the ^-concept is for mode I described by
(1.125). This formula can be generalized to
where
for plane strain
for plane stress.
Formula (7.104) has been derived for a crack inside the plate. For a three-dimensional
crack approaching the surface, G. P. Cherepanov has proposed the following correction of
the material stiffness:
This modification should capture the real behavior between plane stress and plane strain.
Criterion of jT-integral
In his paper [144], J. R. Rice introduced an integral quantity J, which is for a two-
dimensional continuum defined as (Fig. 7.14a)
where W is the density of strain energy (in linear elasticity, 2W and are the
components of surface tractions acting on the boundary F.
Rice J'-integral is independent of the integration path F, if:
the temperature is constant throughout the body,
254 CHAPTER 7. FE SOLUTION OF SPECIAL PROBLEMS
Numerical calculations have shown that, assuming monotonic loading (without un-
loading) and not too large a plastic zone, the ^-integral is virtually independent of the
integration path even for incremental plasticity. In FEM (as well as in BEM) calculations,
the integration path usually passes through the elastic part of the material. In displace-
ment FEM models, the integration path passes through the Gauss quadrature points while
in some stress (equilibrated) and hybrid models, it can pass along the element boundaries.
A close relationship between the j7-concept and the 5-concept can be detected by
choosing the integration path in formula (7.106) as a circle with radius r. Taking the
limit for r -» 0, we get under the assumption of linear elasticity the equality8
The relationship between Q and J can be exploited to assess the critical value in the
stability criterion
where Ry is the yield limit and a is a coefficient depending on the shape of the specimen.
Irwin estimated for an infinite plate with a central crack of length 2a that 2a = I/TT and
Dugdale derived an improved estimate 2a = Tr/8. It is also known that the value of 2a
is about three times larger for plane stress than for plane strain. When applying the
X-criterion, under SSY, the existence of the plastic zone can be taken into account by
8
The physical meaning of the J'-integral can be explained in the following way: The potential energy
needed for crack propagation by A a enters the crack tip across a fictitious surface F6, which is moving
along with the crack tip and, in an infinitesimal volume dy b Aa (crossed-hatched in Fig. 7.14b), it equals
where 6 is the thickness of the body. The second addend represents the negative work of the boundary
tractions on displacement increments AUJ = (dui/dx) Aa.
7.4. FEM IN FRACTURE MECHANICS 255
adding rp/2 to the real half crack length a. The stress intensity factor is then computed
for the effective length
The analytical solution due to Hutchinson, Rice and Rosengren (HRR) is based on the
deformation theory of plasticity and starts from the assumption that, under small-scale
yielding, the stress singularity term proportional to r~ 1/2 remains dominant in a small
region around the crack tip. The equivalent stress at a point given by polar coordinates
(r,$) is expressed as (cf. [114])
In the last formula, Kj is the stress intensity factor corresponding to an elastic material
An important characteristic used to evaluate the accuracy of the analytical solution given
by (7.110) is the ratio A = &¥qRR/aeqKP' Numerical calculations have revealed that A decreases
with the distance from the crack tip, and does so for all load levels. As the error caused by
the approximate nature of FEM stays safely below 20% (see Table 7.1) we can consider the
analytical solution HRR to be invalid for A < 0.8. The dependence of A on r/a is plotted in
Fig. 7.18 for two load levels: / = 0.3 and / = 0.6. At / = 0.3, the formula is obviously invalid
at distances r/a > 2.10~4 while the relative size of the plastic zone determined by FEM is
rp/a = 0.15. At load level / = 0.6, the assumptions of LEFM under SSY are not valid at all,
among other reasons because the plastic zone is too large to be considered as embedded in the
region surrounding the crack tip.
Figure 7.17: Character of the strain field around the crack tip ([112])
7.4. FEM IN FRACTURE MECHANICS 257
uses COD criteria according to formula (1.127) for which the experimental assessment
of the critical value 6C in the blunted crack tip is quite difficult, and the jT-concept is
applicable only to monotonic loading.
To evaluate crack stability under large-scale yielding, the crack stability criterion must
be supplemented by a criterion of stable crack growth. Crack stability is expressed by the
familiar inequality X < Xc and the condition of stable crack growth can be written as
The function XR = XR(Aa) corresponds to the so-called resistance curve, which can
be constructed if the unstable crack propagation is preceded by some measurable stable
crack growth10 (Fig. 7.19). The parameter X can be identified either with Q or with J.
The usage of the jT-integral in condition (7.112) is questionable as the crack opening
leads inevitably to unloading and the third assumption essential for the path independence
of J is violated. The jT-integral is totally useless for nonproportional loading. Therefore,
the so-called T*-integral has been introduced for this purpose.
10
Geometrically similar specimens of different sizes lose stability at different points of the resistance
curve. Using the size effect method (cf. [21]), the resistance curve can be constructed as an envelope to
the equilibrium curves X = J\T(Aa) in Fig. 7.19.
Chapter 8
In the preceding chapter we have shown that the FEM can be conceived as a certain
variational method solving partial differential equations. The fundamental functional
to be discretized by the FEM was constructed such that the differential equation and
its natural boundary conditions were satisfied on the average with a certain weight. For
energy functionals, the weight function was selected as the variation of the basic unknown
function.
Such an approach can be interpreted as a special case of the so-called weighted residual
method. The case is special in the sense that, in general, the weight functions can be
chosen arbitrarily; special choices lead to the FEM, to the finite difference method, to the
collocation method, and also to the boundary element method (BEM) ([37], [39]). From
the viewpoint of structural mechanics, the BEM can be derived from Betti's reciprocity
theorem. We will take this engineering approach because, according to our teaching
experience, it is easier to follow.
The traditional direct version of the BEM, which is closely related to the collocation
method, leads to a nonsymmetric set of linear algebraic equations for unknown nodal
values at the discretized boundary. Combined with a suitable variational principle, or
with a suitable energy-based method such as the Galerkin method, the BEM can provide
a symmetric set of linear equations, which can be converted to a form that corresponds
to the condensed FEM (after elimination of all internal degrees of freedom).
Sirtori [157], and Maier, Novati and Sirtori [113], among others, applied the Galerkin
method to the solution of static problems of elasticity and plasticity. The solution of static
and dynamic problems of elasticity by the BEM based on a general variational principle
was given by Polizzotto and Zito [135].
For time-dependent problems, such as nonstationary heat conduction, or dynamic
problems, the algebraic equations of the BEM turn into Volterra integral equations. In
the context of the direct version of BEM, the so-called alternative formulation proposed
by Brebbia and Nardini [38] proved to be successful. An interesting feature of this for
mulation is that it fully retains its " boundary character" even though the expressions for,
e.g., inertia forces requires an integration over the entire domain Q. Besides the above-
mentioned solutions in the time domain, there exist a number of techniques known from
classical mechanics or from the FEM (e.g., the Laplace or Fourier transform), that convert
the time-dependent problem into a repeated solution in the transformed domain, with a
sequence of values of the transformation parameter. This step is followed by an inverse
numerical transformation that provides the physical quantities of interest in the real time
domain. As shown by Sladek and Sladek [158], this algorithm consumes a large amount
of computer time because the boundary integrals must be recomputed for each value of
258
8.1. SOMIGLIANA'S FORMULAE 259
Combining the equations from (8.1) we get the Larne equations in the matrix form
The BEM, whether interpreted as a special case of the weighted residual method, or
as a consequence of Betti's theorem, makes use of the so-called fundamental solution U*
that satisfies the Lame equation
In the above, 17^ is an unbounded domain, / is the unit matrix, and 6 is the Dirac delta
function, which has a zero value at all points x except for the origin x = O where it tends
to infinity, and for which
The term IS(x) represents three independent loading cases that correspond to unit forces
applied at the origin x — O in the direction of the coordinate axes x\ = x, x2 = y, and
x3 = z, respectively. The first loading case jF\* = T is shown in Fig. 8.1. The corresponding
displacements ^1X2X3 are elements of the first row of £/*, which is a (3,3)-matrix. A
generic element of this matrix is given by the so-called Kelvin's solution
260 CHAPTER 8. BOUNDARY ELEMENT METHOD
in which 8ij is the Kronecker delta (<5y = 1 for i = j, 6ij — 0 for i / j).
Note that in the local coordinate system in Fig. 8.1 we have r+ = x,, and so dr/dxi =
Xj/r = r*j/r. These relations will be exploited when evaluating the stress state at a point
described by the position vector r due to a unit force F? = T. We start from (1.38) and
replace the subscripts ij by jk:
Let us imagine a fictitious surface at the end-point of the position vector r with an
outward normal n (Fig. 8.2). The action of the outer part of the body on the inner part
separated by the fictitious surface is expressed by the components of the traction vector.
Using the second formula from (8.2) we get
from which
In the case of plane stress we replace v in the above formulae by v = vj(\. -f ^).
The matrix representation 17* of the third-order tensor a\^k must be consistent with th
representation of the second-order tensor 0jk as a column matrix. For a three-dimensional
problem we set up a matrix 17* of type (3,6). Its ith row will be populated by the element
of the vector a* = {&iu,v*22i ^33»am, at*3i» aii2}T- For a plane problem, matrix 17* is of
type (2,3).
Matrices C7*, 17* and P* are two-point matrices in the sense that they are linked to
the starting point and to the end-point of the position vector r. In the local coordinate
system in Figs. 8.1 and 8.2, the starting point has coordinates O and the end-point has
coordinates x. The local coordinate system has important advantages when deriving the
BEM but it somewhat hides the two-point character of the matrices. It is more instructive
to present the problem in the global coordinate system in Fig. 8.3. Here, the point of
application of the unit force F* = 1 is denoted by x = (xi,x<2,Xz), and the point where
we look at the response is denoted by £ = (£1,6, £3).
Let us proceed to the derivation of boundary integral equations that are the basis of
the BEM. Fig. 8.3 presents two states. The solid boundary corresponds to the actual state
while the dashed boundary shows a fictitious state (the domain £1 has been extracted from
fioo)-
According to Betti's theorem, the work of the actual forces on the fictitious displace-
ments equals the sum of two terms—the work of the fictitious forces on the actual dis-
placements, and the work of the fictitious stresses a^k on the initial strains (—£07*;).
262 CHAPTERS. BOUNDARY ELEMENT METHOD
where
Formula (8.13) represents three equations for the unknown displacements Ui(i — 1,2,3)
at point x. The displacements uu(x) are uniquely determined by (8.14). If we knew the
boundary tractions PJ(£) on Fu (reactions at supports) and the boundary displacements
Uj(£) on Tp [—Uj(£) are called the distortions], the displacement field in fl would be
determined as well.
Let us postpone the evaluation of PJ on Tu and Uj on Tp to the next section, and turn
our attention to the final stage of analysis—the evaluation of the stress field. We start
from formula (1.38) and supplement the terms resulting from initial strains:
Displacements can be expressed according to (8.13) and (8.14). We skip the tedious
derivation and present directly the final formula
1
Strictly speaking we should write dT(£) and cK7(£).
8.1. SOMIGLIA NA 'S FORM ULA E 263
It is easy to show that d*;A.(#,£) = -<jj;^(x,£). We simply compare (8.7) with the
relation
that is obtained by using (8.13) in (8.15) and taking into account the symmetry u*j = z^;
cf. (8.6). The summation subscript j from (8.13) has been changed to k. The sign change
results from the transformation from local to global coordinates, in which dri/dxk =
d(£i — X[)/dxk = —5ik- Swapping the arguments x,£ we finally obtain a symmetric
relation between both tensors,
The expression for the tensor s*jk is similar. On the right-hand side of the preceding
equation we write p*lk instead of u*lk, etc. The result is
The underlined term deserves a deeper analysis. The tensor da*kl/dxj has a singularity
at x = £, and so the order of differentiation and integration cannot be exchanged.2 The
exchange would produce an additional term that follows from the relation
We will focus on the plane problem on a circular domain £lp of radius p -» 0. The
complementary term will be transformed using the Gauss theorem. In equation (8.10) we
take into account that on the boundary of the circle we have r = p and r» = pi, and so
Uj — pj/p. Using formulae
2
The singularity becomes transparent if we rewrite Lame equations (8.4) in the Cauchy form; cf.
(1.40):
264 CHAPTER 8. BO UNDARY ELEMENT METHOD
where
Formulae (8.21) and (8.22) have been derived under the assumption of plain strain. After
replacing Poisson's ratio byv = is /(I + z/) they become applicable to plane stress. Certain
complications stemming from the local term in (8.21) can be avoided in two ways:
One possible approach ignores equation (8.16), and the derivatives of the displacements
needed in (8.15) are evaluated numerically using (8.13). An alternative approach is based
on the fact that the local term disappears if the integration over points £ precedes the
differentiation with respect to the components of x. Some algorithmic aspects of this
approach are discussed in Section 9.7.3, which touches upon the application of the BEM
to problems with material nonlinearities.
The elements of tensors d^k and s*;A. will be stored in matrices D* and S*, which
are of type (6,3) [or (3,2) for plane problems]. The storage scheme is similar to the one
exploited when representing the tensor a^k by the matrix 17*. The only difference is that
rows are replaced by columns and vice versa. The elements of the tensor t^kl will be
stored in a square matrix T* of type (6, 6) [or (3, 3)], again taking into account the order
of elements in vectors cr and e"o.
3
The tensor t*-kl can formally be written as
8.2. DIRECT VERSION OF BEM 265
Integral equations (8.13) and (8.16), which represent the fundamental relations of
the direct version of BEM, are usually called Somigliana's formulae. For the sake of
conciseness we rewrite them in the matrix form
This is the starting point for the calculation of unknown displacements u on the part
of the boundary Tp with prescribed tractions, and of the unknown tractions p (support
reactions) on the part of the boundary Fu with prescribed displacements.
The coefficient c(x) depends on the shape of the boundary around point x. To un-
derstand its meaning we have to realize that the term c(x)ui(x) represents the work of
a fictitious body force X% concentrated around point x, acting in the direction xiy and
giving a resultant F* = I. The coefficient c(x) arises due to the simple fact that, in the
actual (bounded) body, we have to consider only the portion of the body force that is
applied inside the domain Q.
In the illustrative Fig. 8.4 we restrict ourselves to the plane problem. In a small
circular neighborhood of the point x the body force X* = l/(ne2) satisfies the condition
266 CHA PTER 8. BO UNDAKY ELEMENT METHOD
The work of the body force X* on the displacements Ui is done only in the cross-hatched
domain 0^, corresponding to the angle if). Provided that the displacement field is suffi-
ciently smooth in this domain, we can treat it as constant in a small neighborhood of x
(i.e., in the limit e —> 0), and then
Consequently, we get
Of course, the angle V> is substituted in radians. The values of c(x) corresponding to
nodes 1 to 6 in Fig. 8.5 are listed in Table 8.1.
node i 1 2 3 4 5 6
Cfc) V>!/(27r) 1/2 &/(2ff) 1/4 1/2 1/4
one, which is considered to be one of its main advantages. As the BEM had been preceded
by a fast development of the FEM, it was possible to make use of the discretization
techniques originally proposed for finite elements, e.g., of the isoparametric machinery
explained in Chapter 3.
Fig. 8.6 presents an example of a discretized boundary of a two-dimensional domain.
Empty circles denote nodes between elements while filled circles represent internal nodes.
where Np is the matrix of shape functions (interpolation functions), and rp is the vector
of nodal boundary loads.
The vector of boundary displacement is approximated in a similar fashion:
Fig. 8.6 shows two quadratic shape functions. One of them corresponds to a node
between two elements, the other to an internal node (dotted area). The degree of ap-
proximation polynomials Np can be different from the degree of Nu. Considering that
the stress field, and thus also the surface tractions, are according to (8.1) and (8.2)
proportional to the first derivative of the displacement field, we are led to the idea of
approximating the displacements by polynomials one degree higher than those used for
the boundary tractions (cf. Section 7.3.4).
The ordering of vectors ru and rp depends on the type of boundary discretization.
Suppose that the part Fw with prescribed displacements contains k nodes, and the part
Tp with prescribed tractions contains / nodes. Using the same approximations for the
prescribed and the unknown quantities, we obtain the following structure of the vectors:
Symbols with bars denote nodal values of prescribed quantities. The above definitions
can be rewritten in the compact matrix form
body forces and initial strains, which affect the vector tx/(x). The corresponding integrals
have to be evaluated by numerical integration on cells covering the domain Q. In contrast
to the FEM, internal nodes affect only the right-hand side but they do not increase
the number of unknowns. If the corresponding terms vanish, equation (8.26) can be
transformed, using the first formula from (8.24), into the algorithmically clean form
Formula (8.31) can be exploited when constructing a set of linear algebraic equations
for the unknown elements of vectors urp and pru. This requires that the point x travel
through all nodes of the boundary F = Tu U Fp. For a plane problem, the matrices U*
and P* are of type (2,2), and so each node produces two unknowns (tractions px, py on
Fu, displacements u, v on Fp) and two equations.
To make use of the structure of vectors (8.30) we have to decompose the fundamental
equations of BEM accordingly. We obtain two subsets of equations
The first line in (8.32) represents k equations (8.31) for nodes x situated on the boundary
curve Fu, which is referred to by the first left superscript at the submatrices **G and
*J'Jf. The second line in (8.32) represents / equations (8.31) for nodes x located on Tp.
The second left superscript refers to the subvector rp or ru from (8.30) that is multiplied
by the matrix. The coefficients c(x) are assembled on the diagonal of uuH, or of ^H,
depending on the location of the node x (on Fw, or on Tp).
Each diagonal coefficient of the above-mentioned matrices is a sum of two terms. We
will describe a simple algorithm of evaluation of these diagonal elements without having
to determine the values of the coefficient c(x). First, we rewrite (8.32) in the compact
form
which is the typical BEM formulation of problems with zero body forces and initial strains.
For plane problems, matrices G and H are of type [2(fc + /), 2(k + /)].
Equations (8.33) must satisfy certain conditions for rigid-body motions. If we prescribe
the same displacement u = v — 1 at all nodes, the vector ru = iu consists of unit
elements, and we have
This means that the diagonal element of H (and thus also of UUH and PPH) is given by
In other words, the diagonal element equals to the negative sum of all out-of-diagonal
elements in the same row of H. Evaluation of these elements as well as of the elements
of G will be explained in the next paragraph.
8.2. DIRECT VERSION OF BEM 269
Regrouping the terms in (8.32) in order to move unknown quantities to the left-hand
side and known quantities to the right-hand side, we obtain the set of BEM equations in
the form
The matrix on the left-hand side is nonsymmetric and regular while the right-hand side
is a known vector. Consequently, the unknown elements of vector urp and pru can be
computed by Gauss elimination.
The calculation of boundary parameters is followed by the stress recovery at points x
of the domain fi. This is done by substituting approximations (8.28) and (8.29) into the
second Somigliana's formula (8.23). The resulting expression reads
where, assuming again zero body forces and initial strains, we have according to
The ordering of elements in matrices 17*,£>* and 5*, which are given by the tensorial
formulae (8.7), (8.10), (8.17) and (8.18), must be consistent with the ordering of elements
in the vector &. For plane problems we have
it follows that matrices H and G? consist (for a plane problem) of (2,2)-submatrices. The
out-of-diagonal submatrices are
The domain of integration reduces to the part of the boundary around the point x on
which the shape functions Nus and Npa have nonzero values.
The evaluation technique is illustrated by Fig. 8.7, in which the boundary has been
divided into linear isoparametric elements. For convenient evaluation of formulae (8.9)
and (8.11) we put the origin of Cartesian coordinates x l5 x 2 into the node r at which two
rows of the BEM equations are to be assembled. Consequently, we have xr = O. The
variable point f on the element (s,5+1) is in the Cartesian coordinate system given by the
position vector r with components r\ = x\ and r2 = x2. Alternatively, the position of the
point is specified by the nondimensional coordinate £, which equals zero at the center of
the element. The outward normal vector ns has components (ni a , n25) = (sin (pa, - cos (pa).
The derivative of r in the normal direction can be expressed as
The angle (pa is fixed by the boundary shape while the angle (p is variable.
The integration domain for the shape function Ns in (8.41) and (8.42) is formed by the
part of the boundary between nodes 5 — 1 and s -f 1. The shape function N3 is described
8.2. DIRECT VERSION OF BEM 271
by
on element 5 — 1,5
on element 5, 5 -f 1.
Finally, we have to express the differential of the boundary. From Fig. 8.7 we conclude
that
on element 5 — 1,5
on element 5, 5 -j-1
The integrals in formulae (8.41) and (8.42) are usually evaluated numerically by Gauss
integration known from the FEM. The diagonal elements of H can be computed from
the out-of-diagonal elements according to formula (8.35). In addition to the rigid-body
displacement u = v = 1, from which we have derived the formula, we can also consider
displacement states u = l , v = 0, o r u = 0, v = l , and check the results.
Some elements of matrix G can contain singular integrals that have to be evaluated
analytically. Using the notation from Fig. 8.8, we can represent the shape functions as
is the solution of the problem, because equations (8.1) and (8.2) are stationarity conditions
of IIw.
8.3. SYMMETRIC VERSION OF BEM 273
When discretizing the problem by the BEM we approximate the unknown boundary
tractions and displacements by polynomials of the form
where UNP and PNU are submatrices of the matrices Np and Nu introduced in(8.28) and
(8.29).
Using the approximations (8.54) in (8.52) we get a modified functional (cf. [135])
where
The stationarity conditions of U'w are equations (8.1) and the generalized (global) bound-
ary conditions
The displacement field u(x) and the stress field cr(x) expressed by relations (8.23)
and (8.24) satisfy equations (8.1) but do not satisfy the generalized boundary conditions
(8.57). These are the very conditions exploited by the symmetric version of BEM to
determine the unknown tractions urp on Tu and displacements pru on Tp. In order to
express them, we have to know the tractions p, which follow from a combination of
the relation p(x) — n(x) cr(x) with the second equation (8.23). This completes the
calculation of Somigliana's formulae.
To emphasize a number of symmetries exhibited by these formulae, some of which
were apparent already in (8.25), we rewrite Somigliana's formulae according to [135] and
[113] as4
The symbol F~ denotes a set of internal points infinitely close to the boundary F. The
first two terms on the right-hand sides represent the response of the domain ^ to the
excitation by arbitrary tractions p on Tu and distortions (—u) on Yp. The last terms
express according to (8.23) and (8.24) the response of 0^ to prescribed distortions (—u)
on Tu and tractions p on Fp. We describe them by simplified formulae that do not
4
Two-point matrix kernels Ghk(x,£) in equations (8.58) and (8.59) are denoted by the same letter as
the constant-valued matrices from (8.32). However, note that they have a different meaning and are of a
different type. To avoid confusion, we have changed the superscripts into subscripts.
274 CHA PTER 8. BO UNDARY ELEMENT METHOD
As already explained in Section 8.1, the two-point kernels represent the fundamental
solution of the problem. The function GM(X,£) is the /ith effect at point x due to a
concentrated source at point £. The effect is a displacement (h — ti), a traction (h = p)
or a stress component (h = a). The subscript k corresponds to the source that is work-
associated with the respective symbol, i.e., a unit force (k — u), a unit displacement
(k = p) or a unit strain (k = a).
Let us mention two important properties of the kernels:
a) Symmetry with respect to x and £ for h = k. If h ^ k and x ^ £ we have
The left-hand sides of the inequalities are obviously twice the strain energy corresponding
to tractions p, or to displacements u. The equality sign in formula (8.62) is valid for
rigid-body motions.
Now it is easy to derive the symmetric version of BEM. It suffices to use the first two
formulae from (8.58) in (8.57), and substitute approximations (8.54). This leads to a set
of equations
where
8.3. SYMMETRIC VERSION OF BEM 275
The reader has probably realized that the derivation of (8.63) can be understood as
an application of the Galerkin method to equations (8.58). We simply multiply the first
equation by the matrix PN^, integrate over F u , and take into account the boundary
condition u = u. This leads to the first equation from (8.63). The second equation can
be obtained in a similar fashion.
The system (8.63) is symmetric, the matrix Kuu is positive definite due to (8.61), and
Kpp is negative definite due to (8.62). It is therefore possible to eliminate the traction
factors urp on Tu from (8.63), and transform the system into a form known from the
FEM:
where
Making use of the above-mentioned properties we can easily see that the stiffness matrix
of the body H is symmetric and positive definite, as might have been expected.
The double integration consumes more computer time but on the other hand it
smoothes the singularities in the kernels of the integral equations, and so it admits an
effective usage of the so-called hypersingular fundamental solutions, which cannot be ex-
ploited otherwise. The evaluation of the double integrals does not bring any noteworthy
complications. In most cases, numerical integration is sufficient. Some exceptions are
mentioned in [135].
There exists a formally simpler algorithm leading to the final expression (8.66). It
requires two subsequent simple integrations, and so it is sensitive to round-off errors.
Consequently, this approach does not guarantee a perfect symmetry of the resulting stiff-
ness matrix (cf. [11]). Let us start from the set of equations (8.33) with added equilibrium
conditions for the entire body fi,
If we use only two force conditions of equilibrium for a plane problem, the matrix Q
has two rows. An additional row appears if we add the moment condition. The global
conditions of equilibrium guarantee that the relation between the traction parameters rp
and displacements ru obtained by elimination of the augmented system
Using approximations (8.28) and (8.29) we transform the preceding formula into
276 CHAPTER 8. BOUNDARY ELEMENT METHOD
from which
In contrast to K from equation (8.66), the matrix K' is positive semidefinite because it
contains rigid-body modes.
The algorithms just described have certain advantages when solving complex structural
systems by the method of substructures. We have reported on their application to static
and dynamic analysis of tall buildings [163]. Of course, it is also possible to combine the
standard BEM equations (8.33) with the FEM equations. This approach is algorithmically
less convenient, and so we will not discuss it here.
Example 8.1.
Fig. 8.9 shows a stiffening wall of a 10-story structure weakened by openings. First, the
entire structure was modeled by beam elements described in Section 3.3, which are the
basic building blocks of the JADRO package (see also Section 3.8). Next, we considered
the part of the wall on the second floor as a substructure. The example should illustrate
the potential of these approaches.
Solution:
The first model was based on the BEM with an element-wise constant approximation of both
displacements and tractions. The stiffness matrix of the substructure was obtained from formula
(8.71). To improve symmetry we set
The second model was constructed from biquadratic isoparametric finite elements. The stiffness
matrix corresponding to relation (a) was obtained by static condensation, i.e., by eliminating
the internal degrees of freedom.
The only remaining degrees of freedom are the displacements 1*1,1/2 of the nodes at the
interface between the substructure and the surrounding structure. The stiffness matrix had to
be transformed such that it corresponded to the parameters used by the simplified model—two
displacements and one rotation of the cross section. The transformation matrix T in
Fig. 8.10a shows the vertical displacements while Fig. 8.1 Ob presents the normal stress dis-
tribution. The shear stresses rxy exhibit a larger difference between both models. However, note
that we are comparing a very detailed FEM model with a very crude BEM model. A linear
interpolation of displacements and tractions leads to a substantial performance improvement
of the BEM model. The effect of the model type on the structure stiffness is clear from Fig.
8.11, in which we have plotted the horizontal displacements at the left edge obtained by three
approaches.
applied integral equations as a basic tool. In this section we concentrate our attention on
homogenization of nonhomogeneous bodies by means of a special treatment proposed by
Dvorak [57], and apply it to the solution of composite structures. The main idea consists
in separating the mutual effect of eigenstrains (eigenstresses) from one inclusion (internal
cell) to another. As was already pointed out, one of the most suitable techniques dealing
with homogenization is the boundary element method. This method provides many ad-
vantageous features in comparison to the finite element method. The influence functions
may be computed at each point of the internal cells with high accuracy, nonlinearities in
inclusions and matrix can easily be introduced and are effectively computed.
According to [57], the influence functions enable one to solve the elastic-plastic and
viscoelastic composite systems based on a change of eigenstrains or eigenstresses, while
the other quantities remain unchanged during the iteration process. Procedures of this
kind suit very effectively to applications of BEM.
Some difficulties arise when applying the BEM to nonhomogeneous bodies. To avoid
them, the stress and strain decomposition proposed by Hashin and Shtrikman is used.
There are two variants of transformation field analysis, corresponding to the two aug-
mented variational principles discussed in Section 1.8.
where r^\ is the stress polarization tensor, u\k is a known kernel, and
Equation (8.72) has been simplified by taking into account the homogeneous kinematic
boundary condition (1.229).
Differentiating equation (8.72) with respect to Xj we arrive at the expression
where h*jk(x,£) = — ej^(£,x). We leave the derivation of the formula for -0*^ to the
reader [see (8.22)].
The converted term in 3D formulation,
arises at the internal point x 6 H due to the exchange of the order of integration and
differentiation when deriving (8.74) from (8.72).
Note that an important property of the kernel 7/>*jA.( is its symmetry with respect to
indices i, j and k,l. The same result holds for h^kl with respect to i and j.
8.4. TRANSFORMATION FIELD ANALYSIS USING BEM 279
The operator on the right-hand side of (8.77) was proved to be positive definite and
symmetric integral operator [177].
Levin [109] used another assumption leading to the vanishing boundary terms in (8.74).
He considered Green's tensor of homogeneous media vanishing on the boundary. This is
possible only for some particular unbounded domains in three dimensions (not in 2D).
Our goal now is to derive the relation between the strains and the eigenstresses of the
form
where A and F are the influence function matrices (A is mostly referred to as the me-
chanical concentration function matrix). Note that, once computed, these matrices do
not change their values during iteration processes for nonlinear solution of plasticity, op-
timization, etc. They depend exclusively on the shape of the body under study and on
its material properties in the initial state.
Let us concentrate our attention on a special case of material properties. The starting
elastic homogeneous isotropic body possesses a stiffness D° while the body under study
is divided into m sub-domains fii,...,fim representing inclusions with stiffnesses D^i =
1, ...,ra, and the rest of the body is the matrix with stiffness D°. The tensor D admits
the representation
where
and Ki is a characteristic function of the ith inclusion, being equal to one inside and to
zero outside the ith inclusion. Substituting (8.79) into (8.77) we arrive at the
Now we describe the procedure leading to the influence functions. Let us assume that
both the original problem involving Hooke's law according to (1.226) and the problem
involving the polarization tensor r have the same geometry as well as the same geometrical
boundary conditions. Placing the point x on the boundary, (8.72) yields
Note that x in (8.80) remains an internal point in Q. After discretization of the boundary
and after discretization of the domain £1 into internal cells, both (8.82) and (8.74) take
the form
where U is a square matrix of type (3N,3N), 3N is the number of degrees of freedom on the
boundary, p1 is the vector of discretized tractions at nodal points of boundary F, E, E' are
(3N,6M)-matrices of influences of the strains and eigenstrains in the discretized domain,
H is a (6M,3N)-matrix and, finally, i/> and t// are square matrices of type (6M,6M).
As the regular matrix U may be inverted, elimination of p' from (8.83) and (8.84) gives
where
Obviously, S is a regular (6M,6M)-matrix, as for a given AJJ, (i,j being fixed) a unique
response e may be expected. The influence function matrix F is equal to S~1T while
S~l is the mechanical concentration function matrix A in (8.78).
where the homogeneous condition (1.245) has been employed. Hence, using the kinematic
and the constitutive equations, we get
where the standard quantities are known kernels. The convected term in 3D formulation
arises at the internal point £ € H by the exchange of the order of integration and differ-
entiation when deriving (8.88), and includes a term corresponding to the last addend in
equation (8.15).
Note that an important property of the kernel a^kl is its symmetry with respect to
indices i, j and fc, l\ the same result holds for s*jjfc with respect to i and j.
Several useful conclusions follow from (8.88). As s*jk is decaying as r~ 2 in 2D and
as r~ 3 in 3D, the distribution of u'k obviously tends to zero for supp 7^- (closure of a set
8.4. TRANSFORMATION FIELD ANALYSIS USING BEM 281
of nonvanishing values of 7^) far from the boundary F. This occurs, for example, when
Cfjkl represents the compliance of a composite matrix, with the fibers being distributed
far enough from the boundary (assumption of Eshelby, Mori and Tanaka, etc.). Then
(8.88) is simplified as
Let us concentrate our attention on a special case of material properties of the body
described by the domain H. The starting elastic homogeneous isotropic body possesses a
compliance C° while the actual body is divided into m subdomains fi1?..., fim representing
a division of the domain 0, with the compliances Ci, i = 1, ...,ra. The tensor M admits
the representation
where
and Ki is a characteristic function of the zth inclusion, being equal to one inside and zero
outside the ith inclusion. Substituting (8.92) into (8.90), we get
Now we describe the procedure leading to the influence functions. Let us assume that
both the original problem involving Hooke's law according to (1.242) and the problem
involving the polarization tensor 7 have the same geometry as well as the same geometrical
boundary conditions. Placing the point x on the smooth boundary, (8.87) yields
Note that x in (8.88) remains an internal point of fi. After discretization of the boundary
into boundary elements as well as of the domain into internal cells with constant stress
and eigenstrain distributions, both (8.95) and (8.88) take the form (the components of
both stress and eigenstrain tensors are collected to vectors in the standard way)
282 CHA PTER 8. BO UNDARY ELEMENT METHOD
where P is a square matrix of type (3N, 37V), 3N is the number of degrees of freedom on
the boundary, u' is a vector of discretizecl displacements at nodal points of boundary F,
S, S' are (3,/V, 6M)-matrices of influences of the strains and eigenstrains in the discretized
domain, M is the number of cells, B is a (3Af, 37V)-matrix and, finally, U and JC' are
square matrices of type (6M, 6M).
As the regular matrix P may be inverted, elimination of u' from (8.96) and (8.97)
gives
where
Obviously, V is a regular (6M, 6M)-matrix, as for a given jj+j (z, j being fixed) a unique
response a may be expected. The influence function matrix is equal to V~1Z while V~l
is the mechanical concentration function matrix.
where B = V~l and G = V~1Z. The submatrices Bkl and Gkl relate the influences of
stresses (a®)1 and eigenstrains (p,j)1 in the cell / of the homogeneous, isotropic body and
the stresses in the cell k. The stresses (0f )* are known quantities.
A natural requirement is to assure that the stresses be small as possible. This means
that the variance
should be minimized.
Differentiation of / with respect to leads to the following system of linear alge-
braic equations for unknown (p>j)1'.
where
It can be proved that, for n < m — 1, the system (8.101) is uniquely solvable. In other
cases, the optimization attains more minimum points, or even may not be solvable.
8.5. SOLUTION OF DYNAMIC PROBLEMS BY BEM 283
Mass density p (kg x m~ 3 ) can be taken out of the integral because we assume that the
body inside the domain 0 is homogeneous. The overdots in (8.102) denote differentiation
with respect to time.
We do not want the unknown u to appear at points outside the boundary F, and there-
fore we express the displacements inside fi by a linear combination of linearly independent
basis functions fq, q = 1, 2, ..., m,
where m is the number of boundary nodes, and aq is an unknown vector of the same type
as u.
After £ has run through all nodes CP of the boundary F, i.e.,
we can invert the matrix F(miTn) consisting of elements /9(CP), and we obtain
284 CHA PTER 8. BO UNDARY ELEMENT METHOD
where
Evaluation of the mass matrix according to (8.106) can be done by cells covering the
domain ft. An elegant substitution that replaces the evaluation of (8.106) by integration
over the boundary F has been described in [158].
Let the point x approach the boundary point £. In the first Somigliana's formula we
replace the work of fictitious body forces X*(i = 1,2,3) in the domain £1^ by the work of
equivalent boundary tractions on F^ (see the schematic Fig. 8.4), we can rewrite equation
(8.102) as
representing the distance between the point £ in £1 and the boundary node £6. After the
point C in equation (8.107) has run through all the boundary nodes £6, we obtain a set of
3m (or, for plane problems, a set of 2m) equations for the unknown boundary tractions
and displacements. Depending on the type of boundary conditions we have to regroup
the terms such that known quantities are moved to the right-hand side; cf. (8.36). If we
construct a vector X consisting of the unknowns (traction parameters and displacements),
we can write the system in the form [158]
Figure 8.13: Comparison of results obtained by FEM and two versions of BEM (Ref. 84)
The solution of equations (8.110) with initial conditions (8.112) can be expressed by the
Wheeler-Stemberg formulae [135], which are the dynamic counterpart of Somigliana's
formulae (8.58):
The star denotes the convolution of functions, which replaces the simple product from
static equations. The definition of convolution is given below:
The matrix Ghk(x, £, t - r) represents the impulse response of QQQ, i.e., the effect h at
point x and at time t due to a unit impulse of the work-conjugate quantity of k applied
at point £ and at time r. They possess the symmetry properties described by (8,60).
The vector functions < u/,p / ,or / can be computed from relations (8.59) after replacing
the products of matrix functions by a convolution according to (8.116). It remains to
express the response of ^1^ to the given initial conditions [135]:
Integral relations (8.115) do not satisfy the generalized boundary conditions (8.57)
that are in the dynamic case functions of time. Combining them in a similar manner as
in the static solution, we arrive at a system of Volterra integral equations
where
8.5. SOLUTION OF DYNAMIC PROBLEMS BY BEM 287
Before attempting to solve the derived system, we rewrite it in a more compact form
The vector
contains all unknown traction parameters and distortions on the boundary Fw and Tp
(superscripts). The other matrices in (8.120) have been formed in a similar fashion.
The numerical solution in time is facilitated by the following transformation: Let
K(t) = J(t) and J(0) = O. The matrix J(t) is ojbtained from formulae (8.64) where
wejreplace Ghk(x,£jt^by their primitive functions G/^(x,£, £), for which Ghk(x,£,t) =
dGhk(x, £, t)/dt and Ghk(&, £, 0) = O. Polizzotto and Zito [135] presented expressions for
the components of the matrix Guu, from which we can construct the remaining primitive
unit solutions by a standard approach. The function
describes the displacement u*- at point x 6 OQO due to a suddenly applied unit force acting
at point £ 6 fi^ in the direction xk. H denotes the step function (Heaviside function), 8jk
is the Kronecker symbol. The variables a = r/c^ and (3 = T/CT (CL,CT are longitudinal
and transversal wave speeds) have the dimension of time.
Equation (8.120) will be rewritten as
The matrix
has to be evaluated only once. It is symmetric and definite [cf. (8.61 and 8.62)], which
are favorable conditions for the success of the numerical algorithm. The vector Rn-i(tn)
is computed from formulae (8.117) and (8.119), in which we replace UQ and UQ by tx n _i
and un-i>
on clamped edges:
on simply supported edges:
and on free edges:
where m n ,m nt and qn respectively stand for the bending moment, twisting moment and
shear force on the boundary F. Specific moments are computed as
8.6. PLATE A NA LYSIS B Y BEM 289
The direct version of BEM will be formulated for two unknown functions, w = w(x, y)
and $ = Au>(rr,7/), for which we have to set up and discretize two integral equations. We
start from the fundamental solutions
where S is the Dirac function of the distance between a general point £ = (£, 77) and the
point x = (x, y) at which we search for the values of w and <£.
To avoid dealing with a singularity at point x, we apply the second Green's theorem
for the Laplace operator to the domain fi — fi£, which excludes a small neighborhood of
point x (Fig. 8.14). The final result is obtained by taking the limit e -» 0. It is easy
to see that, according to (8.136), the first integral on the left-hand side of the original
Green's formula
is equal to zero. Now we consider that d$*/dn = d$*/dr = l/e on re, which leads us to
290 8.6. PLATE A NA LYSIS B Y BEM
The meaning of the coefficient c(x) has been explained in detail in Section 8.2.1. If x
is an internal point of the domain £7, we set c(x) = 1. If we move the point x to the
boundary F, the value of c(x) can be taken from Table 8.1.
The last integral on the right-hand side of (8.138) vanishes in the limit e —> 0, because
c(x) 27T£ln£ tends to zero. Finally, considering that
The second integral equation for the unknown w can also be derived from Green's
theorem (8.138), in which we replace $ by w and <£* by Aw*:
Relation (8.135) defines the function w* in polar coordinates under the assumption of
axial symmetry, and so
Due to (8.137), the first integral on the left-hand side of (8.142) vanishes. The second
integral will be transformed using Green's theorem:
Similar reasoning as before [when deriving (8.141)] leads to the conclusion that the inte-
grals of the underlined terms in (8.142) and (8.145) vanish in the limit e -> 0. Combining
8.6. PLATE ANA LYSIS B Y BEM 291
(8.142) with (8.145) and taking into account (8.140) we arrive at the second integral
equation
Numerical solution
Integral equations (8.141) and (8.146) have to be solved numerically. The simplest ap-
proach is to divide the boundary F into J segments (boundary elements), and take the
unknown boundary functions w and <£ and their derivatives in the direction normal t
F, viz. dw/dn and d$/dn, as constant in each element. As we will have to integra t
the applied load p we divide the domain £7 into K cells. In each cell A;, the intensi ty 0
p is replaced by its mean value pk. This is indicated in Fig. 8.15. Let <£j, (d$/dn)j anc
Wj,(dw/dn)j be the values of the quantities under consideration on the ^'th boundary
element. In equations (8.141) and (8.146) we replace integration by summation, and we
obtain for the ith element (or for the boundary point at the center of this element) the
equations
292 CHA PTER 8. BO UNDARY ELEMENT METHOD
for i = 1,2,..,, J. In agreement with formulae (8.134) and (8.135) and with Fig. 8.15, we
have denoted
and the derivatives along the direction of the normal d <b*/d n, d w*/d n and d (Aw*)/9n.
Denoting the components of the unit outward normal by nx = cos nx, ny = cos ny, we
can write
and so
where Xji = Xj — X{, y^ = yj — yi and n xj , n yj are the components of the unit outwarc
normal at node j (Fig.
refo8.13).
In a similar fashion we arrive at
Curve integrals in equations (8.141) and (8.146), which define the coefficients %• to /#,
are evaluated numerically in an element-by-element fashion. For i = j it is possible to
derive closed-form formulae for the evaluation of coefficients an to f a . Two algebraic
equations (8.147) and (8.148) for four unknowns <£j, (d$/dri)j, Wj, (dw/dri)j at each
boundary node j must be supplemented by two boundary conditions that depend on the
type of support.
On a simply supported edge we set
The derivatives d2 w/ds2 and d3 w/ds2 dn could be evaluated exactly from (8.146). Usu-
ally it is sufficient to use their finite difference approximations and write (see Fig. 8.15)
where ht is the finite difference spacing in the tangential direction to the straight part of
the boundary. After the elimination of $j and (d$/dn)j, the only remaining unknowns
in equations (8.147) and (8.148) are Wj and (dw/dri)j.
When evaluating the bending and twisting moments mx,myimxy we have to know
the derivatives d2 w/dx2, d2 w/dy2, d2 w/dx dy. They can be computed from equation
(8.146), in which we set c(x) = 1 for internal points of the domain ft. As an alternative
to analytical formulae, which are easy to derive, we could use numerical differentiation.
Example 8.3
Gospodinov and Ljutskanov [64] presented the results obtained by the above described
method for various support schemes and boundary discretizations, and compared them
to the exact solutions. As an illustration we show the case of a plate simply supported
at its corners and subjected to a uniformly distributed load according to Fig. 8.16, with
v = 1/6. The number of elements along one edge is N.
Solution
294 CHA PTER 8. BO UNDARY ELEMENT METHOD
exact
point x pl*/D JV = 3 A% N =b A%
1 0.02594 0.02250 14.4 0.02564 2.5
2 0.02413 0.02114 12.4 0.02394 1.2
3 0.01720 0.01631 5.2 0.01693 1.5
4 0.00973 0.01055 8.4 0.00991 1.8
exact
point x pl*/D AT = 9 A% N = 15 A%
1 0.02594 0.02599 1.2 0.02617 0.5
2 0.02413 0.02420 0.3 0.02438 1.0
3 0.01720 0.01714 0.3 0.01727 0.4
4 0.00973 0.00990 1.7 0.01002 2.9
exact
point xpl2 JV = 3 A% JV = 5 A%
1 0.1079 0.0879 18.5 0.1082 0.3
2 0.0887 0.0602 32.1 0.0893 0.7
3 0 0 0
4 0.0515 0.0403 21.7 0.0606 17.6
exact
point xp/2 AT = 9 A% N = 15 A%
1 0.1079 0.1086 0.6 0.1088 0.8
2 0.0887 0.0897 1.1 0.0894 0.8
3 0 0 0
4 0.0515 0.0565 9.7 0.0546 6.0
It is clear from Table 8.2 that the displacements are calculated with a satisfactory accuracy
even for a very coarse discretization (N = 5). On the other hand, Table 8.3 indicates that
the bending moments in the vicinity of the corners (point 4) exhibit a large error A even for
a relatively fine discretization (TV = 15). A substantially better accuracy could be achieved by
using a higher-order approximation (linear or quadratic approximation of the unknown functions
on each element).5
the problem in the frequency domain, and we look for the fundamental solution which
represents the response of an unbounded plate fioo to a harmonic excitation by a force
with a unit amplitude and with circular frequency u applied at point x (Fig. 8.14).
The equation of motion in polar coordinates reads
where 7 is the mass density, h is the plate thickness, 6 is the Dirac function, and t is time.
After separation of variables
which is satisfied by
The first two Bessel functions JQ,!Q grow beyond any bounds as p —> oo, and so the
integration constants A\ = A2 = 0. The values of A3,A^ can be determined from the
conditions at point r —» p —> 0, and we get
of magnitude I/An acting in the direction n at distance An. Exploiting the principle of
superposition we obtain
296 CHA PTER 8. BO UNDARY ELEMENT METHOD
Numerical solution
The functions W* and W* describe the response of an unbounded domain QOQ. In the
indirect version of BEM we can use the response of a plate structure of an arbitrary shape
that surrounds the plate fi to be analyzed (Fig. 8.18). A convenient choice is a clamped
circular plate of a sufficiently large diameter. Its response to a unit excitation is described
by Green's function that also reflects the type of support (clamped boundary).6
There are a number of techniques that, using Green's function on a bounded or un-
bounded domain QO, lead to expressions for the solution of a given problem defined on
the domain Q. They all consist in approximating the boundary conditions on F, and are
in general classified as an indirect version of BEM. In this paragraph we will explain the
nature of these approaches under rather general assumptions. We will make use of the
previously derived functions W* and W^.
Let us assume that the domain Q is embedded in £7o = O^ (Fig, 8.19). Further let us
assume that, in addition to the external loading p, an unknown force loading / acts on fio
along a chosen curve FI, and an unknown moment loading m acts along another chosen
curve F2. If p is the amplitude of a given harmonic loading, then / and m are amplitudes
of the unknown boundary forces. In the indirect version of BEM we have to determine
the functions / and m such that the total loading, consisting of p, / and rn, produces a
state in which the boundary conditions on the actual boundary F are satisfied.7
The amplitude of deflection at point f of the boundary F follows from the relation
6
Green's function for a statically loaded clamped circular plate is given, e.g., in [40].
7
When using a circular domain, both boundaries FI and 1^ can coincide with the circular boundary
8.6. PLATE ANA LYSIS BY BEM 297
The point x runs through the domain £7, the points Xi,x2 run along the curves FI,^.
From equation (8.171) we can calculate all the derivatives needed when expressing the
boundary conditions at point £. For example, if the plate is clamped along the entire
boundary F, we have to know the derivative in the direction of the normal n to the
boundary F,
It should be emphasized that while the normal derivative in equation (8.172) is related to
the boundary F and variable £, the normal derivative in (8.170) is related to the boundary
F2 and variable x2.
For the purpose of eigenanalysis of a clamped plate, we divide the curves FI and F2
into J segments. On the jth segment we assume the forces fj and rrij to be constant. The
boundary conditions w(^) = 0 and dw^^/dn = 0 are to be satisfied at j points of the
boundary F. Approximating integration by summation and replacing the specific forces
fj,m,j by their partial resultants acting on the jth segment, we obtain a discretized set
of equations of the indirect version of BEM in the form (p = 0)
for 2 = 1,2,..., J.
The circular eigenfrequencies u corresponding to individual eigenmodes can be com-
puted from the condition that the determinant of the system of equations for unknowns
FJ and Mj must vanish.
Chapter 9
There are two viewpoints from which we can look at the nonlinear mechanics problems:
material behavior, and geometry of the body (or structure). The nonlinearity due to
the material response is called material (physical), and the nonlinearity reflected in the
geometric equations is denoted as a geometric nonlinearity.
We have already discussed various types of the material nonlinearity in the first chap-
ter. Nonlinear constitutive equations are commonly used with linear geometric equations.
In other words, it is assumed that the change of the shape of the body due to deformation
is negligible when considering its equilibrium.
Geometrically nonlinear response of slender structures is often simplified as well. For
instance, it is in many cases assumed that the strains are (infinitesimally) small while the
overall displacements are of arbitrary magnitude. To fix ideas, let us consider the "pure"
deformation of the element in Fig. 9.1 caused by stretching (a'b' w ab) and by bending
($a,$6 "C uab)- In such a case, we might combine nonlinear geometric equations with lin-
earized constitutive equations (valid for small strains). Thus we would be able to model
not only elastic, but also elastoplastic, viscoelastic, viscoplastic and internal-damage re-
sponse of bodies. Note, however, that it is necessary to monitor whether the assumption
of small strains is really substantiated, especially for inelastic materials (Section 9.2.5).
Materials such as rubber or the hypothetical Mooney-Rivlin material are able to bear
very large strains (on the order of unity). The problem at hand is then a geometric
nonlinearity with large strains and arbitrary displacements. Fortunately, the material
deforms nearly elastically (it is nonlinearly elastic, hyperelastic), and the constitutive
equations can be derived from the expressions for potential energy density [compare with
(1.3)].
When observing geometrically nonlinear deformation processes in bodies we are able
to perceive a continuous change of the shape. On the other hand, numerical solutions
298
299
generalization to the continuum mechanics then follows in Section 9.2, and the discretiza-
tion issues are taken up in the Section 9.3. Attention is then given to the numerical
algorithms, both for the finite element and for the boundary element method (Section
9.4). We shall concentrate on the geometric nonlinearity in the FEM; The numerical
recipes are nevertheless directly applicable also to the material nonlinearity and to the
combination of both types of nonlinearity. The material nonlinearity is taken into account
through the tangential material stiffness, which is for our purpose assumed to be known.
We cannot go into the details associated to the material nonlinearity such as the piercing
of the plasticity surface by the loading path, but on the other hand, we consider essential
to show the importance of the selection of the reference configuration, and of an adequate
constitutive equation related to the transformation of the material stiffness matrix. The
algorithms of the arc-length type described in Section 9.4 do not supply information on
critical points, and they sometimes fail close to bifurcation points. Hence, special algo-
rithms based on extended systems of problem equations have been developed by Simo
and Wriggers (Section 9.5). A finite element approach to problems including material
softening and strain localization is dealt with in Section 9.6. Although application of
the BEM in geometrically nonlinear problems is rather rare, some authors (e.g., Atluri)
regard it as promising. The BEM has a number of ad vantages, for solutions to material
nonlinearity problems. Introductory remarks are given in Section 9.5.
Algorithms based on a perturbation technique applied to the expressions for the po-
tential energy of the system are currently gaining acceptance. It found its application,
e.g., in investigations of instabilities in elastic structures. Some interesting and motivating
issues concerning especially imperfect systems can be found in [134].
This chapter is restricted to static problems. The methods of investigation of dynamic
systems were discussed already in Chapter 5.
where TV is the tensile force in the rod. The last modification was based on the assumption
that the angle /3 is very small.
The deformation of the bar is obtained by a simple application of the Pythagorean
theorem
or
The order of approximation of both (9.2) and (9.3) is the same, and the static equation
(9.2) and the geometric equation (9.3) are consistent.
The constitutive equation for the linearly elastic rod can be written as
Combining (9.2) and (9.4) we obtain the relationship between the force F and the dis-
placement w of the top joint:
Let us consider the two-bar strut loaded by the force (-F). The relationship between
this force and the vertical displacement is shown in Fig. 9.4.
If the load arrives at the point A on the loading path ABC, a phenomenon called
snap-through takes place, manifesting itself by a sudden jump of the center node to the
point C of the loading path. The driving agent of this jump is the kinetic energy given
by the area ABC. The points A and B are called the limit points, since a displacement
increment at these points does not produce an increment of the reaction force.
Many structures respond to some loadings in a manner similar to the response of
the shallow strut—shallow arches and shells, for instance. Their response curves are, of
course, more complicated, and in particular in addition to the limit points there also
appear so-called bifurcation points, which correspond to intersections of several different
loading paths.
The example of the two-bar strut is especially suitable for demonstration of the tangent
stiffness of the structure, since it is reduced to just a single scalar for single-degree-of-
freedom structures. As it is obvious from (9.5) we have
302 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
The same result could have been obtained by combining (9.2) and (9.4) and by differen-
tiation
In general, the scalar expression for the tangent stiffnesst Kin (9.6) corresponds in
the FEM algorithms to three matrices, i.e.,
• Linear stiffness matrix, which is given by the initial geometry of the structure [com-
pare the first term on the right-hand side of equation (9.6)],
• Initial deformation matrix, which is dependent on the current displacements, and
expresses the influence of the change of geometry on the current stiffness at the
beginning of the load increment [compare the second term in (9.6) with (9.3)], and
• Initial stress matrix, which is dependent on the current stresses (here represented
by the normal force N).
The described decomposition gives a hint how to solve the geometrically nonlinear
problems in the framework of a Lagrangian formulation: We start from the initial unde-
formed configuration of the body, and when using the total formulation (TL), we are work-
ing with all terms in (9.6). If the geometry is at the beginning of the loading step updated
by the displacements, i.e., in the case of the two-bar strut, if we replace dimensions L, H
by their current counterparts I, h = H + IL>, then in agreement with (9.7) the second term
on the right-hand side of (9.6) drops out. This idea underlies the updated formulation
(UL), which is based on the use of the current configuration to obtain the linearization
in the given increment.
These two views should be taken into account when defining the various deformation
and stress measures. These measures are conjugate in the sense of the principle of the
virtual work. The derivation will be based here on the deformation measure, and the
9.1. NOTATION AND BASIC EXPRESSIONS FOR NONLINEAR BEAMS 303
Since SB is referred to the undeformed configuration [the denominator in (9.8) is the initial
length L], we express also the principle of virtual work in this configuration
where A is the area of the undeformed cross section of the rod. By equation (9.2) we get
Thus
To compute the conjugate stress measure we should start from the varying configura-
tion of volume v = l'a . Assuming incompressibility of the element LA = l'a = la, the
configuration becomes immaterial and, for instance, from
we have
304 304 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
Hence
or
Obviously, for v = 1/2 (incompressibility), we get al = AL, and the volume is preserved
as expected. In general, we have
The case of i/ = 0 satisfies exactly the usual assumption of constant (unchanging) cross-
sectional area.
It remains to recall yet another notion, which will be generalized in the section dealing
with a continuum. It is the logarithmic strain rate (we denote differentiation with respect
to time by a superimposed dot)
The attentive reader will notice the formal similarity of the quantities CLN and ^LAT [see
formula (9.15)].
c) Green-Lagrange strain is defined by
the formula
The stress GG is not an actual stress as it is attributed to the unit area of the undeformed
rod, while it acts on the deformed rod. The denotation pseudo-stress or second Piola-
Kirchhoff stress is commonly used for this quantity. Comparing (9.17) and (9.25) we may
relate the pseudo-stress and the true and nominal stress as
There is, interestingly, another way to arrive at the pseudo-stress. Let us consider
the stretching of the rod from L to /. The unit vector E connecting the end sections is
stretched to
While the true stress a is referred to the actual cross-sectional area, a, and to the unit
vector E, the stress OQ is referred to the initial area, A, and to the vector of the deformed
basis e. Writing the equilibrium conditions
we have
and
the formula
306 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
The stress conjugate to the Almansi strain CA is therefore not the true (Cauchy) stress,
but
We have discussed various measures of stress and strain. What remains to be clarified
are the physical (material) relationships between these measures. If the simplified assump-
tion that the stress measure (&E,&LN — O,GG) is directly proportional to the conjugate
strain (€E,£LN,£G)<> with E as the proportionality factor in all cases, is adopted, we find
that by equation (9.12), (9.20) and (9.25) the same displacement w generates different
forces F in dependence on the strain selected. In other words, we obtain differing load
paths as in Fig. 9.4. We assume that the experimentally determined value of the elasticity
modulus is given by E = (?E/£E (i.e., nominal stress/engineering strain). In that case EE
is given by (9.3), and the formula (9.12) transforms into (9.5). The other variants of the
relation F = F(w) are obtained analogously, with the formula (9.9) being used to express
the current length /. We leave the derivations to the reader and list here only the results
of Crisfield [50]. We can see from (9.5) that the model based on the Green-Lagrange
strain is very flexible, while the models based on the logarithmic strain are stiffer.
If one attempts to achieve comparable loading paths, the elasticity moduli need to be
modified. Setting the stresses equal
or
where the (Einstein) summation rule has been applied (summation is performed over K).
To describe the change of the distance between two infinitesimally close points PQ ->
pq, the relationship between the two vectors dX and dx must be derived. It holds that
where
The finite line segments / and L from the theory of tension-compression bars corre-
spond to infinitesimal quantities ds = (dx - dx)1/2 and dS — (dX • dX)1/2. We can
generalize (9.22) as
Using the formulae (9.36) and (9.37) we can rewrite the left-hand side of (9.38) as
which yields the sought relationship for the Green- Lagrange strain tensor
This tensorial formula is often expressed in matrix form. We order the differen-
tials of the coordinates into the column matrices dx = {dxi.dx^^dx^}7 and dX =
{dXi,dX<2,dX$}T, and we express their relationship as
Here2
deformation gradient matrix,
matrix of the derivatives of the displacement vector components.
Both matrices are of type (3,3). We store the components of the tensor EJJ in a two-
dimensional field E2 (hence the index "2"). Applying (9.41) to express the difference of
the squares alternatively as
2
The matrix form of the deformation gradient is commonly denoted by F. We have used H to avoid
confusion with the force vector.
308 308 CHAPTER 9. PROBLEMS OF aNONLINEAR MECHANICS
The components of the infinitesimal vectors are both in dx, and dX referred to the
reference basis EI, If the components dx are referred to the rotated basis e/, which can
be regarded as local to the original global basis, the transformation is effected by the
well-known formula
or
where T is the transformation matrix. Its terms are simply the cosines of the angles
subtended by the rotated (local) and original (global) axes. The matrix dxi0k is in the
3
The Cauchy-Green deformation tensor proves rather useful in nonlinear mechanics (see, e.g., [12])
because it allows for easy formulation of the relationships between lengths and angles before and after
deformation. Adopting the matrix notation, and assuming that the deformation transforms dX = dSN
of Fig. 9.2 into the vector dx = dsn, where N and n are matrices corresponding to unit vectors, we can
easily derive that
Let us consider further two vectors dX = dSN arid dX = dS'N , which subtend the an^le 0 before
deformation, and which are transformed by the deformation into dx = dsn and dx = dsn subtending
the angle tf. We can show that
We already know that rigid-body rotation generates C = H T H = TTT = I. Preceding derivations now
can be used to show that also lengths and angles are preserved as expected, ds — dS, and ds = dS'.
9.2. FUNDAMENTALS OF GEOMETRICALLY NONLINEAR CONTINUUM THEORY 309
same relationship to the rotated basis as the matrix dX to the global basis, and we have
dxi0k = dX, Comparing (9.44) and (9.46) the relation between the rotation and the
transformation matrix becomes clear:
We have from (9.41) that H = R for rigid-body rotation, which gives after substitution
into (9.43) the identity 2E2 = (RTR-I) = (TTT - /) = O. We can thus conclude that
the Green-Lagrange tensor is not affected by a rigid-body rotation.
Starting from G = R - /, but neglecting the nonlinear term in (9.43), we arrive
nonzero tensor whose components are proportional to 02. These are negligible only in
the case of very small 0.
The resulting deformation of an infinitesimal parallelepiped can be expressed as a com-
bination of the pure deformation—stretch in the principal directions TV/, (/ = 1,2,3)—
followed by the rotation of the material neighborhood into the final orientation. Let
be the change of the vector dX caused by pure stretch. The subsequent rotation leads to
dxr being moved into [see formula (9.46)]
If A/ = (ds/dS}i, (I — 1,2,3), are the relative stretches in the principal directions, the
matrix of the pure deformations is given by a formula known from the theory of elasticity
where
where
where
It remains to note that the von Kdrmdn theory of large deflections for plates and
shallow shells can be regarded as a specialized use of the Green-Lagrange strain tensor:
9.2. FUNDAMENTALS OF GEOMETRICALLY NONLINEAR CONTINUUM THEORY 311
Here Z is the coordinate of the mid-surface of the shallow shell measured from the plane
XY. The displacements u and v are small. On the other hand the deflection w is com-
parable in magnitude to the thickness of the shell, and, correspondingly, it is considered
"large." In analogy to the shallow two-bar strut, the rotations are limited, and are called
"moderate."
Let us note one peculiarity. Although the internal forces act on the deformed element of
the body, their intensities—the pseudo-stresses Sfj—are referred to the unit area of the
undeformed body. It is clear from what we know from the preceding discussion that the
intensities are components of the second Piola-Kirchhoff stress tensor.
The volume forces acting on the deformed element, but referred to the undeformed
configuration, will be denoted F = ~FKEK> The deformed element is therefore acted
upon by the elementary force vector FAA'iAAr2AX3. The well-known elasticity theory
approach will be used to derive the vector expressions for the static equations (balance of
momentum on the parallelepiped pmim^m^}
These are accompanied by three force boundary conditions, written in vector form
312 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
where Nj are the components of the unit normal to the undeformed surface F of the body
Q, and P = PK&K is the vector of distributed boundary forces prescribed on the part
rP.
Let us recall in addition that the balance of moment of momentum on the deformed
parallelepiped pmim2m3 yields the symmetry of the second Piola-Kirchhoff stress tensor
By using (9.37) in (9.59) we obtain from (9.60) the component form of the static
equations
and similarly from (9.61) we obtain the component form of the force boundary conditions:
Using the Gauss integral theorem and taking into account the relation
derived from (9.37), we transform the preceding equation into
The first integrand in (9.66) can be with respect to (9.40), (9.59) and (9.62) modified as
9.2. FUNDAMENTALS OF GEOMETRICALLY NONLINEAR CONTINUUM THEORY 31
3
This provides a confirmation of the well-known fact that the second Piola-Kirchhoff stress
tensor is conjugate to the Green-Lagrange strain tensor through the principle of virtual
work. The component form of the principle of virtual displacements is trivial:
We have discussed various measures of strain and stress in Section 9.1. The Cauchy
stress tensor (Tij is a generalization of the Cauchy stress (or true stress) of (9.18). Thes
e
stresses do work in the current volume v on the virtual strains
which are a generalization of the formula (9.15). The expression (9.69) is a variation fo
the linear part of the Almansi-Hamel strain tensor &ij defined by [compare (9.38)
]
Matching the right-hand sides of (9.38) and (9.70) we arrive at the transformatio
n
e
formula
The transformation rule for the strain tensors need to be completed by an analogou s
relation between the stress tensors. This can be derived from the equality of the virtua
l
work of the internal forces expressed both in the initial configuration V, and in th e
deformed configuration v
Where
Inserting the above into (9.73) we find the sought relationship between the Cauchy stress
tensor and the second Piola-Kirchhoff stress tensor4
4
The stress tensor Ja^ is called the Kirchhoff tensor or the nominal stress tensor
.
314 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
Collecting the components of the tensors into two-dimensional arrays we get an equiv-
alent form of (9.75)
Or
The strain increment shall be decomposed into two parts. The first is linear in the terms
d&uj/dXj, the second includes their products and is thus nonlinear. We can write
where
According to the principle of virtual displacements, the work of the external forces (8L =
—5Ee) is equal to the work of the internal forces (6Ei) at any time instant, and in particular
at the end of the increment:
The last modification is based on the fact that the strain EU and the displacements it/
are known at the beginning of the increment, and thus 6Eu — 0 and 8ui = 0.
On the condition that neither the magnitude nor the direction of the external forces are
dependent on the configuration of the body (i.e., they are conservative), we can express
the virtual work of the external forces in the TL formulation as
9.2. FUNDAMENTALS OF GEOMETRICALLY NONLINEAR CONTINUUM THEORY 315
Combining equations (9.81) and (9.82) we have after some reordering the following rep-
resentation of the principle of virtual work:
In addition to dilation due to temperature changes, and to shrinking, the term AEQKL
includes also another action, which should riot be omitted in the linearization (compare
with Section 1.5). The resulting form of the principle of virtual work is arrived at by
using the symmetry of the tensor 5/y:
The right-hand side represents the equilibrium mismatch between the external and the
internal forces in the configuration v, which is also partly due to the adopted linearization.
In the frame of the incremental process we remove the imbalance by iteration.5
UL formulation
The solution is referred to the configuration v. The computational quantities assume the
values
5
On principle, equilibrium should be checked at the end of the increment. The resulting mismatch
of the solution of (9.86) is given by the difference between the right- and left-hand side of (9.81). The
imbalance will be removed by iteratively transferring the work 6(&L) in (9.86) to the right-hand side and
replacing the last term of this equation by the right-hand side of (9.81). The unknown displacements are
computed at the fcth iteration [the left-hand side of the modified equation (9.86)] for the right-hand side
obtained at the (k - l)th iteration. These details are discussed in Section 9.4.
316 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
where
Using the incremental constitutive equation, the symmetry of cr^-, and linearization, we
obtain the final form:
which represents tensorially the rate of the logarithmic strain [compare with (9.21)], and
Truesdell stress flucf
6
In computational practice of large strain problems one often meets (under certain assumptions)
instead of (9.93) the Jaumann stress flvx, which is also conjugate to the tensor d»j, and which can be
defined as:
Where
If the tensors EU and 5/j are referred to the initial configuration V, their rates can be
expressed from the preceding tensors by the transformation rules
If the configuration v is the reference configuration, the preceding equations can be sim-
plified to
Let us start from the principle of virtual work in the form of formula (9.89). If the
configuration v is to be equilibrated, the imbalance term on the right-hand side must be
equal to zero. Dividing by (A£)2 and the taking limit A£ —> dt yield
To be able to carry out the computation we need to know the appropriately linearized
constitutive equation <7y = f (frij, dij).
Analogously, for most inelastic material models under the small strain hypothesis (Sec-
tions 1.2 to 1.5) the above methodology is applicable for arbitrarily large rotations and
displacements. However, for materials with significant proportion of inelastic straining
it is necessary to check the magnitude of the equivalent plastic deformation of formula
(1.73) (approximately 2% is a suitable limit).
If the material exhibits aside from large displacements and rotations also large strains,
but remains elastic, the nonlinear constitutive equation is derived from the density of the
strain energy W = W(/i,/2> ^3)* where / l5 l^.h are the invariants of the tensor EJJ. We
can write in agreement with the first formula (1.3)
2. When looking for a suitable constitutive law for the UL formulation, we have to
consider that we have 5^ = GIJ for the configuration v, so that the material tensor
needs to be reformulated. We start from (9.75) and we set
We assume &EQKL = 0 for simplicity. With regard to (9.72) and to the symmetry of the
material tensor, we get for A COM = 0
where
is the material tensor for the UL formulation, and Ac*/ is the increment of the Almansi-
Hamel tensor. Also the inverse relationship
where
The material constants G,v are the same in formulae (9.100) and (9.106) (compare
with (9.34) for 1 < eE)-
It is advantageous to use the formulation of equation (9.98) for large strains as the
rate-of-strain tensor is well suited for this purpose mainly because of its relation to the
9.3. FEM DISCRETIZATION OF GEOMETRICALLY NONLINEAR STRUCTURES 319
by using (9.93). If the Jaumann flux is used instead of cr^-, the algorithm is analogous. It
is, of course, necessary to keep in the numerical solution the imbalance terms of equation
(9.98).
A detailed description of nonlinear solutions in the range of very large strains can be
found in references [153] and [154].
Let
320 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
be the vector of nodal displacements. Using the interpolation functions of (3.2) we can
write (9.108) as
where
where
where
9.3. FEM DISCRETIZATION OF GEOMETRICALLY NONLINEAR STRUCTURES 321
Finally, it is necessary to establish the meaning of the individual terms in the governing
variational principle (9.86) with appropriate adjustment for notation:
Considering the independence of the virtual displacements, we obtain the matrix form of
the equilibrium of the rod:
where
On the right-hand side of (9.127), we are interested in the vector of internal nodal forces
F, derived from the stresses in the element. To satisfy balance of momentum it must be
equal to the vector jR, by which the surrounding structure acts on the element
Here ^ is the angle between the element axis and the horizontal coordinate axis (Fig.
9.9). The last modification was based on (9.26).
The tangent stiffness matrix is, in agreement with (1.263) and (9.6), given by
This result agrees with the tangential stiffness matrix expressed in the formula (9.127).
The solution is referred to the updated (but fixed) configuration v given by the coordinates
324 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
and it is governed by the variational principle (9.91), which will be for clarity rewritten
as
Here a is the Cauchy true stress, and E' is the transformed modulus of elasticity. Since
r = O, the formula (9.119) is simplified to
where
Since we have by (9.26) (era//) = (acA/L), we ascertain that the initial stress matrix
Kff is the same in both Lagrangian formulations. For very small deformations and large
rotations we can set
Corotational formulation
In difference to the preceding approaches with fixed coordinate systems (X, Y) and (z, y)
respectively, we refer the solution to the local coordinate system (zj,yj), whose axes
corotate with the element (Fig. 9.10).
The vector of local displacements
9.3. FEM DISCRETIZATION OF GEOMETRICALLY NONLINEAR STRUCTURES 325
is transformed into the global vector, r, given by (9.110) using the well-known formula
where we use the notation c = cos^ = x 2 i//, s — sin-0 = z^\/l to write the transforma-
tion matrix
The vectors of the nodal forces are transformed by the contragredient transformation
Where
The rotated engineering strain £# with the conjugated nominal stress cr# = E£Q, are
perfectly natural to the corotational formulation, and the tensile force can be expressed
as N = aEA,
Having established this, we derive
Comparing to the UL formulation, we find that the factor (E'a/P) of the UL formulation
is replaced in the corotational formulation by the factor (EA/Ll2). The two approaches
yield practically identical solutions for very small deformations.
To compute the initial stress matrix, we need a suitable approximation of the differ-
ential of the rotation dt/j = (dijj/dr)Tdr, With regard to Fig. 9.11 we set
The corotational formulation is well suited not only for tension/compression rods, but
it is applicable to bending elements as well.
We present an interesting modification of this formulation which consists of an alter-
native expression for the vector of global nodal forces in the form given by (9.145):
where
and we have
The decomposition of the tangent stiffness matrix testifies that that nodal forces come
from two sources. We can see from (9.156) that the forces Ki(ip)dr cause elongation of
the element, while Ka(ip)dr are additional forces caused by the rigid-body rotation dij) of
the element. Let us inspect this detail more closely. We rewrite the initial stress matrix
as
The rendering of the initial stress matrix in the local coordinate system makes its
physical meaning more accessible. Let us suppose that the element of Fig. 9.12 transmits
the force N — OE&, and rotates from the configuration v by the angle d^ = dw^i/L
To satisfy the equilibrium conditions also after the rotation (the resultant must pass
through the nodes) the force ±N must be completed by transverse forces given by
where
formulation starts from the configuration v (Fig. 9.13a), the TL formulation from V (Fig.
9.13b). The additional forces follow from the similarity of the cross-hatched triangles. As
we have
it is obvious that we obtain the same initial stress matrix in both formulations.
9.3.2 Beam
The bending elements shall be investigated both in a fixed coordinate system (we restrict
ourselves to the TL formulation), and in a corotated coordinate system with local axes
following the motion of the element. Finally, we draw the reader's attention to some
peculiarities of initially curved beams.
7
An arch is considered to be shallow (in the engineering practice) if the ratio of its rise to the span
is under 1/5. If the element can be considered shallow we can introduce some simplifying assumptions
when constructing the shallow arch element.
9.3. FEM DISCRETIZATION OF GEOMETRICALLY NONLINEAR STRUCTURES 329
The Kirchhoff theory of shallow arches rests on the basic geometric hypothesis
where ua is the horizontal displacement of the mid-curve point and w = dw/dX is the
derivative of the deflection. Using (9.166) in (9.165), we get for the relative stretch of a
generic fiber
where
are column matrices of unknownsw)(r and prescribed nodal displacements (z). We have
introduced the notation
The normal force is very important in shallow arches. To capture its variation well, a
quadratic hierarchic approximation is recommended
where
is a column matrix, which, in addition to nodal displacements Ui, contains also the hier-
archical degrees of freedom a.
Following the approach of the preceding paragraph, we can first expand the addends
in (9.167):
where, with regard to d/dX = (I/L)d/d£, the row matrices can be expressed as:
330 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
We can compute the strain increment now. Introducing the scalar factor,
The individual terms of the stiffness matrix and of the vector of internal forces can be
computed by using the formulae (9.123) to (9.125). They will be modified in that we first
perform the integration over the thickness of the beam. We transpose the computation
from the stress space to the internal-force space
The modified formula (9.123) yields the linear stiffness matrix &ud the initial displacement
matrix, which is composed of
The tangential stiffness matrix can be assembled by regular localization from submatrices.
9.3. FEM DISCRETIZATION OF GEOMETRICALLY NONLINEAR STRUCTURES 331
The angle u represents (in general finite) rigid-body rotation of the element. The angles
$1 and $2 represent the rotations of the end-section's due to the change of shape of the
element. The total rotations (p\y and (p%y include also the rotation of the element
In agreement with the approach used in the formulation of the tension/compression bar,
we first express the dependence of the strain parameters in a differential form with respect
to the local coordinate system (A is the change of the end-section's distance)
Application of (9.185) and (9.186) yields easily the basic formula of the corotational
formulation
332 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
or
For curved elements, we encounter certain problems for the local incremental relation
We have dealt with this issue already in Section 2.3, where we have drawn the reader's
attention to the fact that the change of curvature of the element affects also the distance
between the end cross sections, and accordingly, the correct expression for the normal
force reads
Similarly, the change in the distance of the end-points must affect the curvature (otherwise
the symmetry of the stiffness matrix would be lost). Having this in mind, and using (2.98),
we can write8
where
This formula includes also the influence of shear, so that the constants c* and c£ need
to take this fact into account (Section 2.3). We set K = 0 for slender beams, and corre-
spondingly c\ = GI, c£ = c2. We can set cx = c2 = 0 for straight beams.
The transformation into the global coordinates can be obtained by differentiation of
(9.189)
8
Note that AC stands in (9.193) for the section shape factor.
9.3. FEM DISCRETIZATION OF GEOMETRICALLY NONLINEAR STRUCTURES 333
The initial stress matrix is dependent on all forces, N, Miy, and M^y. The fact that
(I//) is also subject to differentiation needs to be properly taken into account. From the
viewpoint of practical computation, it is advantageous to separate the membrane and
bending effects, and, by using the structure of the matrix B and the formulae (9.185) and
(9.186), write
The last modification highlights the symmetry of the initial stress matrix, whose first
part is just an extension of (9.156) by two zero rows and columns. The rows and columns
have been reordered to reflect the ordering of the vector (9.184), though.
where we set hi — H^ The values at the reference line are marked by an over-bar.
334 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
The basis functions for the shell element of Fig. 9.16b are constructed in a similar
manner.
Subsequent derivation is well known from the theory of isoparametric elements. We
restrict ourselves to a two-dimensional continuum and we will regard the presentation as
a generalization of the approach of Section 9.3.1 (tension/compression rod).
We start from the displacement approximation
The linear part of the strain tensor increment can be written in the compact form
[compare with (9.119)]
The newly introduced matrix (B\ -t- B2) is composed of row matrices of (9.198). The
nonlinear part of the increment we leave in component form.
9.4. METHODS FOR SYSTEMS OF NONLINEAR EQUATIONS 335
The following expressions for the tangential stiffness matrix of the isoparametric el-
ement are simply generalizations of (9.123) and (9.124). We will adopt the notation of
Section 9.2.4, and write in the TL framework:
where Kt is the tangential stiffness matrix, and A is the coefficient of proportionality of the
loading, as an incremental form of equilibrium conditions for the the structure as a whole.
The distribution of the load over the structure is expressed by the vector H, which is
prescribed. Equation (9.202) can be solved only approximately by a suitable numerical
method.
which will be called the imbalance (or out-of-balance) force vector. If the classical Newton-
Raphson method (NRM) is used, we assemble Kt = Kt(r + Ari) and compute the new
displacement increment from
336 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
This step is repeated as long as gi is larger in a suitable norm than some prescribed bound.
The convergence criterion will be discussed later. The NRM algorithm is illustrated in
Fig. 9.17b. Note that the NRM requires an assembly and factorization of the tangential
stiffness matrix Kt for each iteration. This is of course rather demanding in terms of
computer time. Therefore, the modified Newton-Raphson method has been proposed
(MNRM), which differs from the full Newton-Raphson method in that the iteration matrix
is identical to the tangential stiffness matrix at the beginning of the current increment
and is kept constant for all iteration steps. The algorithm is summarized in Fig. 9.17c.
points. Nonlinear investigations of instabilities often lead to loading paths with both
bifurcation and snap-through/snap-back points. One example is shown in Fig. 9.18b.
The solutions by the NRM or the MNRM are driven by load increments. It is obvious
that it is not possible to trace the whole loading path in that case. If the case at hand is a
path with snap-throughs, it is possible to drive the solution by displacement increments.
Practical implementations of the algorithm drive the computation by prescribing the
displacement of a selected node of the structure. However, this approach fails for snap-
backs.9
9
It is possible to analyze arbitrary paths by combining the load- and displacement-driven approaches
and switching between them as necessary. This was used, e.g., in the ASKA package.
9.4. METHODS FOR SYSTEMS OF NONLINEAR EQUATIONS 337
To circumvent these difficulties, Wempner [175] and Riks [146] have independently
proposed a method, which is called the arc-length method (ALM). Our starting point is
the imbalance vector
where A is the scalar factor which changes the load level. We deal with the so-called
proportional loading here. If the factor A is generated by the solution the computation is
driven by load increments. That means that on a certain level the (hyper)plane A = const,
no longer intersects the loading path, and no solution can be obtained. Wempner and
Riks therefore introduced the length of the loading path, s = f ds, as the controlling
parameter. The differential of this length can be written as
where A/ is the "radius" of the spherical hyper-surface in the space (A, r) and a is the
square of the "mismatched" arc-length of the loading path. The basic difference between
the classical NRM and ALM is in that the parameter A is variable in the ALM. This means
that the problem is now reformulated with (n + 1) unknowns. The extended system of
equations consists of n equations (9.205), and of the equation (9.207), which governs the
step-length along the loading path.
Using the NRM or the MNRM, we start from the Taylor expansion10 of g, and a
10
The meaning of the symbols is quite clear from Fig. 9.19. However, note the difference between
the differentials d\, dr in (9.206) and the quantities 8X, 6r from (9.208), which assume the character of a
differential, but are computed at the point (A0, r0). These are in addition different from the increments
AA,Ar.
338 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
The advantage of (9.210) consists of that fact that the system matrix is not singular even
if Kt is. That means that (9.210) may be used even in close neighborhood of limit (snap-
through) points, and on the unstable portions of the loading path. One disadvantage is
that the matrix is nonsymmetric and without a banded structure.
Spherical ALM
The solution of the nonsymmetric system of (9.210) has been replaced by Crisfield by an
approach which is based on the method of Batoz and Dhatt [19] for nonlinear systems
with displacement control. The basic idea is to decompose the increment 5r into two
parts. We have for the yet unknown load level A = A0 + 6X
The iterative increment 8r can be split into a part caused by internal forces, and a
part caused by the load. It holds
This equation introduces yet unknown quantity 6X. We compute it from (9.207), where
we insert Ar from (9.212) and AA = A0 + 6X. After some algebra we get
where
We obtain <5A from here, and then we solve for 8r from (9.211). Thus we have derived
relations needed in one iteration step. Let us note that we compute 8rt and Kt only
once in the MNRM, at the beginning of the increment, and then we do not change them
during the iteration. Contrary to (9.210), the spherical ALM works with the matrix Kt,
which need not necessarily be regular. Experiences show, however, that the probability
of getting too close to a singular (limit or bifurcation) point is rather small.
9.4. METHODS FOR SYSTEMS OF NONLINEAR EQUATIONS 339
The steps of a single increment are summarized in Table 9.1. Here i denotes the
iteration. The formula (A) in Table 9.1 follows from the geometric dependencies
Cylindrical ALM
It is shown in [25] on a number of examples that the CALM is a robust method for
nonlinear mechanics problems. The algorithm of CALM is presented in Table 9.1. It
remains to supply some details concerning the solution of the quadratic equation (9.213).
Two cases are possible:
For the imaginary roots it is necessary to return at the beginning of the increment, and
restart with shorter step A/. The case of real roots leads to two iterative corrections, Ari
and Ar2. The decision of which correction to adopt is based on the angle between the
current correction vector and the preceding one, Ar0. The explanation is best provided
by inspecting the angle $ subtended by the vectors Ar0 and Ari, or Aro and Ar2,
respectively. The cosine of this angle is computed from
where
Linearized ALM
The linearized step-length constraint is given with regard to (9.209) by
A formula which was used for the first time by Ramm (see [140]) results if do is
neglected in (9.216). This approach11 is illustrated in Fig. 9.20a. It is obvious that the
The original Wempner-Riks approach differed in that the iterative change (Ar 0 ,£A 0 )
was perpendicular to the tangent vector at the beginning of the step, as shown in Fig.
9.20b.
The linearized ALM's are simpler than the Crisfield's spherical or the cylindrical ALM;
especially so, as they avoid the difficulties with the selection of the root from (9.213). On
the other hand, they are reportedly not as robust as full ALM's.
where &W is the prescribed work increment (compare with A/ in the ACM), and A is
the load level at the end of the previous increment. We set for the subsequent iteration
&W = 0. We get
The parameter 6X appears in (9.219) in the second power. Therefore, similarly to the
spherical and cylindrical ALM, we need to solve a quadratic equation. We always obtain
real roots, however.
The Bergan current stiffness parameter is defined as the ratio between the current 'k1
to the initial 'k'Q
The step-length control strategy constitutes very important part of any serious nu-
merical implementation of the methods described above. The recommendations found in
12
The increment of the external work is composed of two parts:
1. The force AR does not change during the increment and
2. The force AAR depends on AA in the same manner as Ar, and therefore
342 CllAFTER 9. PROBLEMS OF NONLINEAR MECHANICS
the literature are mostly empirical, and sometimes differ radically. For instance, Crisfield
originally advocated the step-length control as
where A/o is the preceding step-length, 70 is the number of iterations needed to achieve
the prescribed precision, and In is the allowed number of iterations in the current step.
The above strategy was later abandoned in favor of Ramm's formula
where AC is the prescribed change of the Bergan parameter C and Co, C_i are the values
of this parameter in the preceding steps.
Let us note further that the Bergan parameter can be used in the overall control of
the computation. It is only fair to remark here that all of the above criteria are known
to fail in some cases.
Basing our recommendation on extensive numerical experiments of our PhD students
L. Jendele [93] and P. Krysl, we are inclined to regard the formula (9.222) as the most
robust choice, as it tends to cut down the step-length A/ in high-curvature portions of
the loading path, and preserves the number of iterations needed for convergence.
It seems that most authors recommend to start the solution by a load-controlled
MNRM, and to set the step-length by
13
The formula (14) is in [25] falsely ascribed to Crisfield.
9.4. METHODS FOR SYSTEMS OF NONLINEAR EQUATIONS 343
To remove the dependency on the dimensionality of the individual components of the load
(force) vector, the following modification is recommended:
where S is the weight matrix, for instance, according to Crisfield (see [50]) S =
K*t-l(K*t=DizgKt).
The precision achieved can be also measured by the work done by the imbalance forces
where Kt and r 0 are referred to the beginning of the increment. The step-length 77 i
determined from the condition of zero work of the imbalance forces on the displacements
Sr. It should hold that14
The solution of (9.230) can be carried out, e.g., by the regula falsi method. The equation
need not be solved exactly. The efficiency of the technique is measured by the ratio |^^|,
where S(TJ) = 8rTg(rj). It is sufficient to achieve \%j&\ < plim = 0.8. Details can be found
in [50].
14
The condition (9.230) can be also obtained from the minimum potential energy principle.
344 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
With regard to the balance equation (9.202), we can write so-called "quasi-Newton
equations"
Where
The algorithm was proposed originally by Davidon and later modified by Fletcher
and Powell. This approach, denoted DFP after the contributors, constructs the inverse
stiffness matrix as
Inversion of (9.234) yields the matrix ®K~l. This technique constitutes the founda-
tion of the Broyden-Fletcher-Goldfarb-Shanno technique (BFGS). The modification of
the stiffness matrix can be written in the form of a transformation. It holds that
where
It is a well-known fact from linear algebra that the inverse matrix to a banded matrix is
in general full. Equations (9.235) and (9.236) are typical examples. Therefore, a direct
application of the described approach is not desirable in the FEM. We show a modification,
however, which does not need the inverse (i) K~l to compute ^8r. We write:
and we can compute ^tfr, as the second equation of (9.238) can be solved, e.g., by
ordinary Gauss elimination. It is also possible to use a line search and evaluate T/I. The
total displacement vector for the first iteration is then .
The indirect solution has the disadvantage of large storage requirements to save a number
of vectors. On the other hand, the BGFS is in many cases a very efficient tool, and it has
been used, e.g., in one of the largest FEM packages, MSC/NASTRAN.
The research done in the MacNeal-Schwendler Corporation was also concentrating on
when the matrix K~l needs to be modified during the iteration. The quantities
were found useful for the monitoring of the iteration process. Here W^ is the eigenvalue
of WA, and WE is the divergence radius. Lee presents in [106] five examples where the
modification of K~l is not desirable:
1) WE > 1, i.e., when the solution diverges,
2) W^ -+ o, since WK~I becomes singular,
3) WE -» 1 and W^ _» QQ, as WK becomes singular,
4) (*-1)JrT^7 —> 0, as WK becomes singular,
5) A -» 1, as the solution has converged, or it is at least locally linear problem.
where
where
Hat her remarkable results can be achieved by neglecting also the second term
in (9.242). The simplest iterative update of K~l results; only a single assembly and
factorization of the stiffness matrix is needed in each step.
If
The eigenvalue problem is of great practical value. Solving it for each increment z,
we arrive at a critical load of the nonlinear problem (see Fig. 9.21). In addition, it
where y is the instability (buckling) eigenmode computed from the eigenproblem at the
critical point. When the critical point has been found from (9.250), it can be classified
348 CHAPTER 9. PROBLEMS OF NONLINEARMECHANICS
according to the following criteria (compare with [159], [179]), where the notation of
(9.202) has been used:
Limit points
Bifurcation points
A deeper analysis of the equilibrium at the critical point is possible, if the four scalar
quantities computed from the directional derivatives of the stiffness matrix are known:15
The types of the individual bifurcation points can be easily determined from the signs
of certain scalar quantities, which are functions of the directional derivatives (limit or
bifurcation point, quadratic or cubic limit point, symmetric or nonsymmetric bifurcation
point). As it will be shown in Section 9.5.2, the directional derivatives (9.252) are essential
in the formulation of extended systems for a direct search for critical points. Therefore,
additional constraints serving deeper analysis of quality of equilibria cause no difficulties.
Details can be found in Spence and Jepson[159], brief summary is provided in [179].
Example 9.1
We investigate the nonlinear response of a shallow circular arch of Fig. 9.22. The numer-
ical values of the geometrical characteristics are specified without units. Similarly, the
elastic modulus, E = 100,000.
Solution:
We have stressed in Section 9.3.2 the importance of the approximation of us for shallow arches.
Experience tells us that a linear approximation of us is not good enough. A considerable im-
provement can be achieved by a quadratic approximation with a hierarchical function of (9.173).
Cubic approximation yields another moderate improvement. Bakalikova has shown in [10] that
four elements with a cubic approximation for us and w (Fig. 9.23) can achieve higher accuracy
than 16 elements with a linear approximation of ua and a cubic one for w. The element is
is The operators in (9.252) simplify the notation used so far
9.5. CRITICAL (INSTABILITY) POINT ON THE LOADING PATH 349
and taking into account the boundary conditions w(Q) — w(L) = 0 and the relation (1/R) =
(-Z") we find by per partes integration of the last integral an identitical expression.
Let us investigate the fundamental properties of the loading paths (Fig. 9.23) of a perfect
structure. The bifurcation points (B.P.) separate the path into a stable part (solid line) and
an unstable part (dashed line). The classical load-driven Newton- Rap hson method fails in the
limit point (L.B.). We have the displacement-driven variant at our disposal (controlling, e.g.,
the deflection wc). If we want to investigate the properties of the path between the limit points,
we start by displacement control, and then transfer to load control with the matrix Kt in the
basic form. We conclude that D > 0, Dk < 0 on this part of the path.
The unstable branch corresponding to the anti-symmetric buckling mode is drawn in dash-
dotted line, and it holds here that D < 0.
Introduction of a slight anti-symmetric imperfection according to Fig. 9.22 enables us to
excite additional anti-symmetric modes. We see that the bifurcation point degenerates into a
limit point. Koiter has shown in his seminal work of 1945 [99] that the limit points of weakly
imperfect systems lie in the vicinity of a bifurcation point on the dotted line in Fig. 9.23. This
was later confirmed by a perturbation technique applied to the potential energy of the system
(compare with [84], [134]). It is also possible to excite the imperfections by an anti-symmetric
load.
The primary path of the pinned, uniformly loaded shallow arch is of relatively simple shape.
If, for instance, a symmetric snap-through of a deep circular arch was to be investigated, a very
complicated loading path would result. The phenomenon is caused by the alternating sign of
the load and the displacement.
350 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
Example 9.2
The structure of Fig. 9.24 is often used to test a numerical algorithm for nonlinear me-
chanics of shells. The structure is of dimensions L = 254 mm, R = 2,540 mm , / = 12.70
mm , and the isotropic elastic material has characteristics E = 3103 MPa , if = 0.3.
Solution:
The problem appeared for the first time in a paper by Sabir and Lock, and it was solved by
a combination of displacement- and load-driven computation. Two loading paths are shown
in Fig. 9.24 for two thicknesses H = 6.35 mm and H = 12.7 mm. The load # is represented by
a multiple of a concentrated force F = 0.1 kN. The symmetry conditions were applied reducing
the model to one quarter of the shell. The anti-symmetric bifurcation modes were thus excluded
from consideration, and the loading path consequently does not display any bifurcation points.
The response was either a snap-through (for thickness 12.7 mm) or a combined snap-through
and snap-back (for thickness 6.35 mm). The structure was modeled by shallow shell elements
(membrane/plate combination with constant stress distribution). The full NRM was used. The
computation was done by P. Krysl on an INTERGRAPH 6000 workstation. Note that to track
the loading path the algorithm was restricted to make unnecessarily small steps.
This example shows that a change in the geometric or material parameters may cause a
qualitative change in the response character. As it is in general very difficult to estimate the
trickiness of the computational problem at hand, it always pays off nicely to have a robust
numerical algorithm at our disposal.
The next version is rather better suited for FEM applications, and is based on the
constraint (9.250) in the form Kty = O, completed by the scaling functional l(y).
The extended system reads
and has been used in structural mechanics by a number of authors (Seydel, Moore,
Spence, Werner et al.).
16
In agreement with (9.254), the condition J(y) = yT R - 1 = 0 restricts the solution to limit points.
352 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
where
The system (9.257) is solved by the extended bordering algorithm, which is usually
used for systems of type arc-length. It will be discussed to some detail in Section
9.5.3.
The linearized system uses the Hessian Kt. Progressively worsening condition num-
ber of the Hessian near a critical point is usually circumvented by sacrificing the
symmetry of Kt. The modification proposed by Simo and Wriggers is free of this
disadvantage. The basic idea is similar to Felippa's penalty method which introduces
a penalty function in the form ef r — /i = 0 with the unknown scalar IJL into the
energy functional.
The Euler-Lagrange equations of this modified energy functional
The first equation of (9.259) was produced by a variation of (9.258) with respect to r.
The extension of the equilibrium equations by the penalty term can be interpreted
as a stiffening of the structure by discrete springs. The second equation of (9.254)
was extended analogously. The weight coefficient 7 has the meaning of stiffness.
The numerical experiments show that its choice does not affect the convergence
rate very significantly.
17
The methods of the arc-length type with the constraint equation in the form /(A, r) = 0 lead to an
extended system
We rewrite the linearized system derived from (9.259) to arrive at this lucid form
where g = g(X, n, r, y) is the vector g at the starting point (A, //, r, y) of the given
iteration step and
is the updated extended tangent stiffness matrix. As the same matrix constraint is
used both for the displacements r and for the instability mode y, they are associated
to the same matrix K^- This fact affects the directional derivative computation
Defining
we obtain
where
If we know the load increments, we can compute the corrected r and y by using
(9.263) and (9.267).
The factorization (triangulation) needs to be done only for the matrix Kt^. The
additional operations can be done element-by-element, and the algorithm is conse-
quently well suited for large structural systems.
Using the symmetry of the second derivative g we can write the derivative of Kt y in the
direction 6r is the following equivalent form
Example 9.3
Wriggers and Simo applied the above algorithm to the classical bifurcation problem of a
circular cylindrical bar under tension of length 2L = 106.668 mm, and radius R — 6.413
9.6. FEM APPROACH PROBLEMS INCLUDING SOFTENING AND LOCALIZATION 355
mm. The rod switches from a homogeneous strain state into a necking configuration.
The problem was discussed already in Section 1.4.2 in terms of strain localization (see
Fig. 1.12). Here we display the results for an elastoplastic, large strain solution. The
algorithms used were described in Simo's papers [153, 154].
The example uses the elastoplastic material Huber-von Mises-Hencky model with non-
linear hardening, G = 80.1939 GPa, K = 164.206 GPa, the initial yield stress CTO = 0.45
GPa, the residual plastic stress cr^ = 0.715 GPa, H' = 0.12924 GPa, nonlinear hardening
exponent a = 16.93.
Solution:
One quarter of the axisymmetric structure is discretized. The subdivision into elements is obvi-
ous from Fig. 9.25. The nonlinear hardening is described by
The primary loading path exhibits a limit point LI and two bifurcation points B\ and #2- As
the elastoplastic response is irreversible, the extended system solution requires special attention.
In particular, unloading must be taken into account in the bifurcation analysis. The shape of
the primary and secondary path emanating from B\ is obvious from Fig. 9.25, where we also
show some loading paths for an imperfect rod. The imperfections were introduced by a slight
reduction of the radius in the middle of the bar. The necking is more pronounced for larger
imperfections. The post-bifurcation (post-limit) branch is essentially the same for all paths. It is
worth noting that the secondary paths does not display a snap-back due to the plastic straining.
where c?/ is expanded by formula (1.142), the incremental form of equation (1.259) may
be written as
We omit all the considerations concerning the assessment of the vector <TJ and the decision
whether an elastic point enters the plastic region or whether a plastic point starts unload-
ing. We refer the reader to [130] regarding this point. We just emphasize that a Cl—
continuous interpolation is necessary for the approximation of A, while C°—continuity is
sufficient for the approximation of the displacement field u.
Here, rw and r\ are respective nodal vectors, N are C°—continuous interpolation func-
tions, and functions h satisfy the C fl -continuity requirements.
Substituting the functions introduced by equations (9.275) and (9.276) into (9.272)
and (9.274), we get
9.6. FEM APPROACH PROBLEMS INCLUDING SOFTENING AND LOCALIZATION 357
In the last equation, we have substituted the integration region Q for Qp, considering that
A equals to zero in the elastic region fie = &> - ^V Equations (9.277) and (9.278) may be
Y*/vnt7TM'f•for» infrv + no rtr\Yf~\T\£ir>t mcifrTV Tr\r*m
The constraint (9.283) must be included into the formulation either by means of a La-
grangian multiplier or by the penalty approach:
To compute Aeo, we need a formula for the stress increment [compare with (8.23)]:
The underlined terms have been treated in Section 8.1. The initial strain increment is
dependent on the character of the material nonlinearity, especially for a two-dimensional
reduction of the problem. We focus attention on the elastoplastic material, where Ae0 =
A£D. Assuming incompressibility of the plastic material we have (compare with [117])
where
Equation (9.289) follows from the material properties and it manifests itself for a general
stress state onlv bv the reduction of the constitutive equation
The fictitious stress is nonzero for plane strain for an increment of the plastic defor-
mation Ae033 [see (8.10)]
9.7. PHYSICALLY NONLINEAR AND TIME-DEPENDENT BEM 359
The underlined term in (9.287) needs to be extended by the work of a^Aej^. The
consequence is that the matrix 17*, whose elements constitute the components of the
tensor cr*^, is replaced by JC populated by the components of the tensor
WHERE
For the plane stress is a*33 = 0, and the additional term in (9.292) drops out. In
agreement with (8.21), the matrix T* defined by the tensor
remains valid.
The tensor e*jk can be obtained directly by combining (8.9) with (1.37)
where
Additional details can be found, e.g., in [167]. Let us only remark that the additional
term for a plane stress/strain problem will be written for an elastoplastic material in the
form
directly following from a combination of (9.290) and (9.293). Note that caution is required
when applying formulae for an elastoplastic analysis as it needs to be checked whether or
not the plastic incompressibility constraint is applicable (9.289).
360 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
The second term on the right-hand side is a discrete version of the underlined term in
(9.287). To stress this fact we adopt the same notation for both equations. The system
(9.301) can be reordered by using (8.36) after the boundary conditions have been enforced.
The vector of unknowns is A# = { u ArJ, pAr£}T. The upper indices denote the unknown
forces on Fw, and displacements on Fp. The new reordering can be written as
The vector Ar = {"Arf, pArJ"}r combines in agreement with (8.36), the effects of
displacements on Fu and of traction on Tp.
Equation (9.288) can be handled similarly [see equation (8.37)]
We have used formally identical notation for the last term to stress that it is a discrete
counterpart of the underlined term in (9.288). In agreement with (9.293) we can write
T as a sum of two matrices. The first represents the work of the fictitious forces on the
initial strain increments, and is expressed by a cell-by-cell integral [compare with (8.21)].
The second matrix corresponds to the local term.
Equations (9.302) and (9.303) constitute the governing system of the nonlinear prob-
lem. The solution is effected in steps:
• The first increment is chosen so as to achieve the limit state in at least one cell (the
limit state is given by the initial plasticity surface for an elastoplastic material).
When triangular cells are used, the test is performed either at the barycenter (the
increment is considered constant over the cell), or at the vertices (the initial strain
increment is interpolated linearly over the cell). The second option is mentioned at
the end of the paragraph.
• If the subsequent loading starts from a zero increment in each cell, the steps of the
iterative process correspond to the modified Newton-Raphson method:
1. The prescribed load increment Ar determines the increment of the unknown vector
and we update the vector of the sought boundary displacements ru and tractions rp by
incrementation.
2. Using; (9.303) and considering Fig. 9.26b
we obtain A<re and we construct a list of cells for the stress cr + A<re where the limit
state was reached (plasticity condition for an elastoplastic material). We correct the stress
9.7. PHYSICALLY NONLINEAR AND TIME-DEPENDENT DEM 361
Figure 9.26: Initial strain method (a), initial stress method (b)
increment A(T by using the tangential stiffness matrix; first in a computation of the strain
increment (see Fig. 9.26b)
and subsequently
The material parameters (hardening, damage, etc.) are then determined. At the same
time, certain corrections to the load level can be performed to adjust the stress a to the
loading surface. It is obvious that eventual un-symmetry of the tangential stiffness matrix
does not complicate the computation.
4. In the second iteration step, we return to step 1, and we distribute the nonlinear
boundary contribution Ar' = J7Aeo by using (9.302). Solution is followed by an update
of the boundary vector.
Even though the external forces do not change during the iteration, some stress com-
ponents grow and some diminish.
The iteration process (Fig. 9.27) is concluded by checking a suitable stopping criterion,
which is usually presented as
The increment of equivalent plastic strain defined by (1.73) is a suitable norm for an
elastoplastic material.
Let us discuss the computation of the underlined terms in (9.301), (9.287), and (9.303),
(9.288) in some detail. The initial strain field A£O is suitably approximated on triangular
cells (Fig. 9.28) in the form
362 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
where pi,p2 are the triangular polar coordinates, pi,p2 € (0; 1), and Aejf is the initial
strain increment at, the node cv. The coordinate transformation
assigns to each point Pi,p 2 of the unit square the point X\, X2 of the triangle such that
the sides of the square P'iP2, P2P'z, P^P^ correspond to the sides P\P2Pz of the triangle.
The main advantage of the triangular polar coordinate, introduced by Li et al. in
1984, lies in that a linear change in r corresponds to a linear change in p\ as we have
There is no singularity in Qs and numerical integration does not involve any difficulties.
9.7. PHYSICALLY NONLINEAR AND TIME-DEPENDENT BEM 363
where
The integration is performed over all the triangles of 05. Matsumoto and Yuuki have
derived some closed-form formula for the integrals (see appendix of [117]) including the
correction of (9.313) when d^k is exchanged for a\jk.
The preceding results can be used also to compute the underlined terms in (9.288). If
the order of differentiation and integration is not changed in the singular domain £ls we
get in agreement with (8.15)
The first integral on the right-hand side is regular and can be computed numerically. The
second term is a summation over the cells of the singular domain. The derivatives of Jj
can be obtained numerically from two close points x1 or analytically. As the analytic
computation is rather involved, Matsumoto and Yuuki recommend the use of a symbolic
manipulation system REDUCE, which can produce FORTRAN source code.
Example 9.4.
The efficiency and accuracy of the elastoplastic BEM model (for Aeofcfc = 0) is usually
tested on a thick-walled cylinder loaded by an internal pressure. The exact solution is
known for plane strain and the Tresca plasticity condition.
Solution
The BEM results of Matsumoto arid Yuuki are shown in Fig. 9.29. The data used: Ideal von
Figure 9.29: Elastoplastic solution for the pressurized thick-walled cylinder ([117])
Mises elastoplastic material with yield stress aQ = 237.6 MPa, E - 2.059 x 105 MPa, v = 0.3.
364 CHAPTER 9. PROBLEMS OF NONLINEAR MECHANICS
The load is a uniform internal pressure. The figure shows the adopted discretization into bound-
ary elements and internal cells and the dependency of the radial displacement on the internal
pressure. The agreement of the numerical solution with the exact values is remarkable.
Chapter 10
Adaptive FE Techniques
The goals of an adaptive technique are to achieve more accurate results without spending
more time in the preparation of the computational models. The underlying idea consists
in an automatic optimization of the model. The classical FE solution uses a model
assembled from a certain number of elements of different types. If the results are not
accurate enough, another model is created by using a larger number of elements of the
same type. It is then assumed that more accurate results have been achieved. If the
mesh is uniform (or almost uniform), the larger the number of elements, the smaller the
characteristic dimension h of an element. As long as the necessary convergence criteria
are satisfied, the results achieved for a finer mesh are more accurate than those for a
coarser one. This process is usually denoted as an h-convergence. There is an alternative
approach, however, which consists in keeping the number of elements fixed, and increasing
the order of the polynomial approximation p within an element. The number of unknowns
per element naturally grows, and also the element stiffness matrix is getting larger. This
approach is denoted as a, p—convergence. The notation "/i—version" and "p—version" of
the FEM is sometimes used.1 The most promising approach is currently a combination
of the h— and p—version, the so-called hp—version [4].
Thanks to the progress of the theory of a posteriori error estimates, it was possible
to formulate an adaptive algorithm, which can be formulated as follows. After an ini-
tial computation is performed on a given basic mesh, we can compute the element-wise
discretization error estimate from the given (material constants, loads), and computed
(approximate solution) quantities. In the next step, we consider only those elements, on
which the discretization error is larger than some prescribed value. We then either refine
the elements (in the /i—version), or increase the order of the polynomial approximation
(p—version). Finally, we solve for a new solution on the modified mesh. This is repeated
until the error is below a prescribed bound (within each element).
365
366 CHAPTER 10. ADAPTIVE FE TECHNIQUES
made more precise here. The hierarchical displacement-based element takes off from
the simplest displacement element for the given problem—two-noded element for a one-
dimensional problem, and general four-noded quadrilateral for a two-dimensional problem
(see Fig. 10.1). The difference between the isoparametric and the hierarchical element can
be seen clearly on the example of a one-dimensional element. Fig. 10.2a shows the basis
functions for the quadratic isoparametric element and Fig. 10.2b depicts those for the
quadratic hierarchical element. The displacement approximation is given by Table 10.1.
Figure 10.2: Basis functions for the quadratic isoparametric and hierarchical element
For the isoparametric element, the functions Ni are the interpolation (serendipity)
functions associated to the nodal displacements ^1,^2 and u3. On the other hand, for the
hierarchical element, /i,/2 are the interpolation functions for the basic element, and hi
is the first hierarchical function. It can be seen that u\,u<2 are displacements of the end
nodes, and a\ is the amplitude of the hierarchical function (without a physical meaning).
We have for higher hierarchical functions (on the interval < —1,1 > )
Comparing the stiffness matrices of the quadratic isoparametric element (Ki) and the
hierarchical element (Kh)
it becomes clear that the stiffness matrix of the hierarchical element contains explicitly
the stiffness matrix of its basic (two-noded linear) element.2
In general, it does not hold that the stiffness matrix terms associated to the hierarchical
functions are nonzero only on the diagonal. It holds, however, that the resulting stiffness
matrix is a basic stiffness matrix bordered by the extending terms. This property is at
the foundation of the p-version of the FEM.
The p-version of the FEM has been so far applied mostly for two-dimensional field
problems, elasticity theory, Mindlin plates and shells, i.e., in problems where a CQ con-
tinuity is required. The basic element is the bilinear quadrilateral of Section 3.1. The
two-dimensional hierarchical functions are formulated as tensor products in the natural
coordinates £, 77. We distinguish between three types of basis functions
• basic functions—usually the bilinear shape functions of the four-noded element,
• edge functions—product of a hierarchical function in one direction and a basic func-
tion in the other direction,
• bubble functions—product of two hierarchical functions.
The basic functions guarantee continuity of the function values between adjacent el-
ements. The bubble functions vanish on the boundary, and continuity is not affected.
On the other hand, the edge functions have to meet the consistency condition: Elements
sharing an edge must use the same hierarchical functions on this common edge. Some ex-
amples of all the above types of functions are shown in Fig. 10.4 [a) basic shape function,
b) edge shape function, c) bubble shape function].
possible cost. The basic norm used is the energy norm. The issues were studied from a
mathematical viewpoint by Babuska [5, 97].
Most of the error estimates were initially proposed for plane elasticity problems. The
estimates were based on residual forces due to stress discontinuities on the element bound-
aries, and due to imperfect equilibrium in the element interior. For instance, we have for
the bilinear element
where /, are the element edges, r2 is the residual from the equilibrium conditions, J is
the stress jump across the edge shared by two elements, and
hi is the characteristic dimension of the element, and NE is the number of elements in the
mesh. The indicators (10.1) describe the achieved accuracy in all stress components over
the whole mesh. The rate of convergence for the p-version depends on the smoothness
of the solution. The convergence is much faster for smooth (regular) solutions. If the
solution is singular at some points of the domain, the optimal rate of convergence can
be achieved by combining the p- and h-version. The philosophy of the hp-version is best
explained on a concrete example (after [78]). It is a plane stress problem on an L-shaped
10.2. ADAPTIVE TECHNIQUE OF ZIENKIEWICZ AND ZHU 369
domain. Three different subdivisions into elements are shown in Fig. 10.5. The solution
has been obtained for different polynomial orders of the approximation. The error is
expressed in the energy norm \\e\\ [see (10.2). Fig. 10.6 shows the convergence for the
individual finite element meshes (N is the number of degrees of freedom)]. It can be
Figure 10.6: Dependence of the error on the total number of degrees of freedom and on
the polynomial order ([77])
seen that for a given subdivision the increase in the hierarchical order of approximation
leads to higher convergence rate (pre-asymptotic convergence). For this case, the rate
drops, however, starting with certain p. It is possible to derive from Fig. 10.6 an optimal
strategy for this case. The goal is to obtain an accurate solution with minimal effort. In
our case it means to start with the mesh (a) and increase the polynomial order up to
p = 3. Then refine, and solve on mesh (b) for p — 3 and 4, and, finally, refine and solve on
mesh (c) for p = 4 and p = 5. The envelope of the convergence curves (shown in dashed
line) indicates very fast convergence for the /ip-version.
The described example serves only an illustrative purpose. Unfortunately, it cannot
be used to design general rules for an adaptive process. A practical guide (using methods
of artificial intelligence) can be found in [6, 141] and in Section 10.3.
The Z/2 norm is simpler. For example, for the displacement u the L2 norm of the error
eu reads
Although the integral measures are related to the whole domain, it is obvious that they
can be computed element-by-element
WHERE
For a one-dimensional problem, it can be proved that (10.5) converges to the cor-
rect error.4 The usefulness of cr* has been demonstrated for two- and three-dimensional
problems, and also for higher-order functions. Later it was mathematically justified. It
turned out that the approach could be further generalized in the sense that the approx-
imation cr* can be different from that of u. The chosen approximation should satisfy
traction boundary conditions on the inter-element boundaries, and on the boundary with
prescribed tractions.
As mentioned above, the classical approaches require knowledge of the jumps in
stresses at the shared element boundaries. Interestingly, Hank has shown that the
Zienkiewicz-Zhu estimate is for bilinear two-dimensional elements equivalent to the clas-
sical one. In addition, the Zienkiewicz-Zhu estimate provides more accurate stresses.
The same authors improved the recovery process very significantly([ 1 8 5 ] ) .They
proposed Superconvergent patch recovery, which is more efficient mainly for quadratic
and higher-order elements. The idea is to find points at which a the stress obtained by
FEM is superconvergent. For the one-dimensional case such points are represented by
Gauss-Legendre integration points (see Appendix B). Zienkiewicz and Zhu ( [ 1 8 5 ] )
discussed the problem in more detail and they also addressed the problem of boundaries
and interfaces with discontinuites.
To assess the error estimate quality, we can apply the so-called effectivity index
Index 0 denotes prediction, fl is thus formulated as a ratio between the norms of predicted
and actual error.
4
It has been also proved for a two-dimensional problem on the condition that or* is continuous and
the natural boundary conditions are satisfied.
372 CHAPTER 10. ADAPTIVE FE TECHNIQUES
It was discovered empirically that the above estimate needs to be modified with regard
to the element type. The correction multiplier of 77°, assumes values:
• 1,1 (bilinear quadrilateral elements)
• 1,3 (linear triangular elements)
• 1,6 (biquadratic nine-noded isoparametric elements)
• 1,4 (quadratic triangular elements)
for the whole problem domain. If we assume that the error is uniformly distributed over
all elements, an error bound can be established to limit the error on each element
the mesh should be refined at the location of the element. The £t- yields additionally also
a refinement rule. If the characteristic dimension of the element is /i« before refinement,
it should be changed to
where p is the degree of polynomial approximation. The above relation does not hold
near to singularities. The mesh should be graded rather strongly toward the singularity.
The criteria based on a global energy error norm provide rather inaccurate information
about stresses. Therefore, the stress error measure ACT (defined in the remark of Section
10.2.1") orovides often valuable insight. It is to be met, locally, on each element
Figure 10.8: Initial and final mesh for an adaptive computation. Required accuracy 5%
([182])
The ES enables the user to obtain the desired solution with minimal costs. The ES
is supported during the computation by a p-version FE program, and by an automatic
mesh generator.
Expert system are generally regarded as a very promising step in the evolution of
computation. A typical ES consists of (i) the inference engine (problem independent),
and (ii) knowledge base (problem dependent).
Up to now, the expert systems found application in many fields in which it was usually
not possible to formulate clear algorithms—predictions, diagnostics, monitoring. These
were the so-called first generation expert systems. The next generation was based on a
combination of heuristic knowledge with exact algorithms (FEM, error estimation). There
is a set of general rules which should be met to apply an ES successfully:
• Experts are able to solve the problem much better than laymen.
be established, with the final goal at the top. To achieve a subgoal means to solve a
partial problem. This in turn means either application of an algorithm, or response to
a question (e.g., "Are there any singularities in the exact solution?"). The latter is an
internal task, the former is an external task. The external task may be result in
• success (logical status TRUE),
• failure (logical status FALSE ), or
• not yet solved (logical status UNKNOWN ).
The internal tasks result only in an answer "yes" or "no." This results in a completion
of a subgoal, or in a change of the logical status of current subgoal.
Detailed discussion is beyond the frame of this work, mainly because of the need to
define quite a few additional quantities. The reader may find additional details, e.g.,
in [141].
The problem of how to find an optimal solution by the /ip-version of the FEM is
usually formulated as: With the constraint that the total number of degrees of freedom
is to be less than N0, find p and Mi (i is the total number of singular points, Mi is the
number of layers of elements around the ith singular point) such that the error in a given
norm is minimized.
An essential part of an /ip-version of the FEM is the error prediction. It is the deep-
knowledge portion of the knowledge base. To find an optimal p and M,-, a number of
10.3. ARTIFICIAL INTELLIGENCE METHODS IN AN HP-VERSION OF THE FEM 377
meshes need to be tested. However, realize that it is not necessary to run a computation
on the mesh. The full-scale computation is run only for the final mesh with optimal p
and M{.
The efficiency of the hp-version using the ES is best seen from the following example.
A sample problem was solved with a mesh of 3,000 degrees of freedom, achieving 2%
accuracy (in the energy norm). To obtain the same level of precision with a uniform
subdivision would require w 107 degrees of freedom, an /i-adaptive mesh would have
34,000 degrees of freedom, and approximately the same number of degrees of freedom
would be required also for a p-adaptive computation.
The machine time consumed in the feedback (optimization) is negligible in comparison
with the time needed on the final mesh. What is yet more important, the time spent by
the analyst is independent of the prescribed accuracy.
Hank and Babuska have in the example mentioned above tested the difference between
the predicted error and the actually attained error. The structure is shown in Fig. 10.10.
To be able to compare the data, they ran a computation for all the tested meshes. Figure
10.1 la displays the error prediction, and Fig. 10.lib depicts the actual errors. The
agreement is striking. A larger difference can be noticed only for p = 3, when the actual
error is larger than the predicted error. One possible explanation is that the error in the
smooth part of the solution has been neglected.
To conclude, let us notice that the "marriage" of the hp-version of the FEM with an
ES represents an extremely effective tool for solutions of linear problems of mechanics.
At present, primarily two-dimensional elasticity problems have been addressed, but an
application to Mindlin plates has already been published in [78],
The category of deep knowledge includes, in addition to mathematics, also mechanics.
For instance, we have mentioned the shear locking phenomena when dealing with plates.
We discussed polynomial approximation of w higher than that of tpxi (py as a possible cure
of this effect. The details can be found in [78].
378 CHAPTER 10. ADAPTIVE FE TECHNIQUES
where K is the stiffness matrix, or one of its variants. With some exceptions, the stiffness
matrix is symmetric and positive definite. When resulting from adaptive techniques, the
stiffness is sometimes not well conditioned.
Historically, the most popular solution methods of (10.7) were in the fifties iterative
(Gauss-Seidel, relaxation) methods. The computers were much slower than today, and
since the convergence was because of badly conditioned matrices slow, the direct (finite)
methods took over (decomposition LTDL by the Gauss elimination, or the Choleski
decomposition LTL). The elimination algorithms have the advantage that it is possible
accurately predict the time needed to solve the system. A number of approaches were
designed to reduce the residual errors in the elimination solution by iteration. As the
computers were becoming more powerful, users solved larger and larger systems. The
matrix K is, howver, very sparse when it is large, and the elimination processes lack
the ability to efficiently take advantage of this property. The attention of the analysts
turned therefore again toward the iterative solvers. An idea to couple iterative and direct
methods appeared. Thus, a method called multi-grid came into being.
The solution error can be split into
• high-frequency error, and
• low-frequency error.
The high-frequency error6 is significant, but is of local character when viewed from the
spatial point of view. This error can be reduced easily by an iteration (relaxation). The
low-frequency error is, on the other hand, rather smooth, but affects practically the whole
domain. The low-frequency error is effectively reduced by an elimination method, as the
iterative methods often fail here. It remains to notice that the low-frequency error needs
not be computed on a fine mesh, due to its smoothness, but a coarse mesh is appropriate.
To describe the algorithm in more detail, let us assume that we have two meshes at
our disposal. The quantities on the finer mesh are denoted /, and quantities on the coarse
mesh by c. A single cycle of the multi-grid can be described as:
• We start from the current approximation ^r/, we perform n\ iterative or relaxation
cycles with the matrix Kf, and we obtain an approximate solution (A:)r/. The
number n\ is to be selected so as to sufficiently reduce the high-frequency error
(usually n\ = 2 — 4). The iterative process results in a smooth error.
• We transfer the residuals from the fine mesh to the coarse mesh. Formally:
8
Let us assume that we have a spectral decomposition of the stiffness matrix at our disposal. The
high-frequency error is associated to the high-frequency eigenvalues.
10.4. MULTI-GRID METHODS FOR THE SOLUTION OF SYSTEMS OF LINEAR EQUATIONS 379
where Kc is the stiffness matrix associated to the coarse mesh. Since the error
due to Wrf is smooth, A^V C is a good correction. The solution is done by an
elimination method.
We transform A (fc) r c from the coarse mesh to the fine one by interpolation. Formally:
is passed, where ||.|| is the Euclidean norm of the vector, and e is the required relative
accuracy.
It is appropriate to mention that the multi-grid methods did not inherit only the good
properties of their predecessors. Convergence difficulties were, for example, observed on
meshes with high-aspect ratio elements.
Chapter 11
380
11.1. RANDOM PROPERTIES OF A STRUCTURE 381
which of the approaches described in the following is optimal for a given problem.
and the problem with random functions is then transformed to a problem with random
vectors
In FEM, the number of random variables fcjis usually smaller than the number of degrees
of freedom, and so the shape functions $j are defined on superelements, each of which
extends over several elements.
The basic information on a random vector is in the second order moment theory given
by two characteristics:
mean value of a random vector
where
For random quantities which do not have the same distribution, [63] recommends the
Spearman coefficient of correlation
in which a-"' is the difference of order of the elements in ordered sets and n is the size of
statistical sets 6* and bj. The coefficient R(bi,bj) always lies in the interval < —1,1 >.
Note that the random vector b in formula (11.1) has a specific meaning. Its elements
bi = b(xi) are values of the random function 6 at discrete points x,. The spatial correla-
tion between the elements of the random vector can thus be obtained by calculating the
autocorrelation function R = R(d), where d is the distance between the correlated points.
The most widely used correlation function has an exponential form
Relation (11.5) shows that the covariance matrix C with entries Cov(bi,bj) and the cor-
relation matrix R with entries R(bi,bj) have a similar meaning. If the entries of the
random vector 6* are statistically independent, the corresponding covariance matrix C*
is diagonal and R* is a unit matrix. Assuming equal probability of both vectors, i.e.,
f(b) db = f*(b*) db*, we can transform the vector b to the vector b* and make use of the
fact that the number of operations with the diagonal matrix C* is substantially smaller
than the number of operations with the fully populated matrix C. Two basic approaches
are possible:
• Transformation using the eigenvectors of the matrix C
Let Y be the matrix of normalized eigenvectors and A be the diagonal matrix of
eigenvalues of the matrix C, so that
which can be obtained using the method of roots [55] and in which A is a lower
triangular matrix. The transformation
If all the elements of the vector b had the same characteristic deviation cr, i.e.,
C = cr2R = cr2BBT, the vector 6* determined by the transformation
would have a unit correlation matrix. Formula (11.13) implies that in this case
C* = a1! = a2R*. In a general case, the transformation (11.19) based on the
decomposition
x
lf the random vector b does not have the normal distribution, it can be transformed using the
log-normal transformation. The deviations from the exact solution are usually small.
11.2. BASIC STATISTICAL METHODS 385
The number of addends P is usually taken between 8 and 12. There exist also more
sophisticated formulae with a minimum number of addends between 2 and 5. _
2. Each sample e^ of the auxiliary vector yields a sample of b given by b^ = Ae^+b.
The matrix A follows from the decomposition of the covariance matrix C calculated for
the vector b according to (11.16), and b = E[b] is the vector of the mean values.
3. A problem of (linear or nonlinear) mechanics described by the generated parameters
b^ is solved by the standard methods applicable to the deterministic case.
4. Steps 1 to 3 are repeated / times, where the number of samples / is on the order
of several thousand.
5. The obtained statistical sets of the results (displacements, strains, stresses) are
statistically processed.
we find the mid-points of N layers. If N is sufficiently large, we can assume that all the
values bk from the layer number n have the same probability density as the value bk.
Then it suffices to randomly choose the nth layer of each input variable such that it is
selected only once in each sequence of TV samples. In this process, the layers of different
input variables must be randomly combined. This can be done in the following way:
Generate N random numbers Rk ' , ..., Rk } uniformly distributed in the interval [0, 1]
for each input variable bk, k — 1, 2, ..., q. Then determine the order n of the layer to be
chosen according to [23] as the number of all the numbers Rk < R^ increased by one.
This results in the following assignment:
random numbers
order -n1 1
UIUcI
cjfc according to Fig. 11.2. The management of this process is facilitated by the table of
random permutations.
So far we have assumed that the elements of the random vector b = {61,62,..., bq}T are
independent. This must be reflected by the independence of the statistical sets forming
the columns of Table 11.1. Statistical independence of Table 11.1 can be verified using
formula (11.7) or (11.8). Spearman's coefficient of correlation is more appropriate for this
purpose. If the coefficient of correlation between each two columns is smaller than an
a priori chosen small value (i.e., the correlation matrix T corresponding to Table 11.1
is sufficiently close to the unit matrix), the table is accepted; otherwise a new table is
generated and the procedure repeated (cf. [63]).
The input variables are often statistically dependent (e.g., the strength and Young's
modulus of a material). This is described in a nondimensional form by the correlation
matrix R corresponding to the vector b. The same correlation matrix should correspond
to the vector c = {ci,C2, ...,cq}T, whose elements Ck are represented by the statistical
sets 4n) given in the columns of Table 11.1. The correlation matrix T corresponding
to the randomly generated table differs from the given matrix R and the table must be
modified. According to [63], it is recommended to transform the table first to a matrix
H with elements
where G is the distribution function of the standard normal distribution. The matrix H
has the same order of elements and the same correlation matrix as the table. Let us use
the procedure from Section 11.1 and decompose the given correlation matrix R according
to formula (11.20):
A similar decomposition can be found for the matrix corresponding to the table:
Recall that B and Q are lower triangular matrices. According to the second formula from
(11.19), the transformation2
yields a correlation matrix approximately equal to the unit matrix and a subsequent
transformation according to the first formula from (11.19) gives a matrix
whose correlation matrix is close to the given matrix R. The procedure can be repeated
until a satisfactory result is obtained. The order in Table 11.1 is then expressed using
positive integers. Note that the transformation (11.30) can be used to "clean the table"
from spurious correlations if the elements of the vector b are essentially independent.3
3
Many objections can be posed against the procedure used in the LHS method. Mathematicians
criticize especially the reduction of the choice from each layer to its mid-point. This objection is valid
particularly for the boundary layers, which have a major influence on the coefficient of skewness and
excess of the generated distribution. The selection in these layers should be definitely made using the
procedure of the Monte Carlo method. An important advantage of stratification methods (stratum =
layer) is the fact that no range of admissible values of the simulated variable can be omitted.
388 CHAPTER 11. SYSTEMS WITH RANDOM FIELDS
where
where the first variation e&bi = e(bi-bi) describes the fluctuation of the random variable
bi around its mean value 6; and e2A6^ A6j is the second variation. The small parameter e
will be in the final expressions set equal to one (if the term is considered) or zero (if the
term is neglected).
The coefficients in expansion (11.34) have the following meaning:
They can be obtained either directly from formulae (11.35), if the relation D =
D(bi,bi,...,bq} is known,4 or from experimentally acquired data using the least square
method. The coefficients in the expansion of the nodal force vector can be obtained in a
similar way.
4
Consider the dependence of D on a single random function b—e.g., on the modulus of elasticity.
11.3. PROBABILISTIC FINITE ELEMENT METHOD (PFEM) 389
Special attention should be paid to the expansion of the Jacobian J and of the operator
matrix dT, corresponding to the random shape of the domain ft. Consider the domain
ft as the quadrilateral element in Fig. 11.3. The coordinates of its nodes, arranged in
vectors x = {xij,Xi+itj,Xi+ij+i,Xij+i}T and y = {yt^y*Kj»yt+ij+i»yij+i} T > will be
considered as random numbers, which are directly elements of the vector 6, or depend on
certain geometric parameters from 6. The Jacobian of the isoparametric transformation
is according to formula (3.10) expressed as
Using matrix notation, we get after some manipulations the following relations:
These formulae were used by the authors in analysis of the earth body in Fig. 11.4,
with a random shape of the internal boundary separating two subdomains fti,ft2 with
different material properties. The "mean shape" of the internal boundary was considered
Using (11.1), we get at the point x
as a straight line. The shape fluctuation around this line is denoted as Ad(y). The ele-
ments of the vector b were taken as the random lengths dj.
Example 11.1
Derive the expansions of the Jacobian J and of the operator matrix d1', assuming that
the "mean shape" of the domain Q is a rectangle with sides d,v (Fig. 11.5). The random
input variables (elements of the vector 6) are the horizontal and vertical lengths of the
mesh dni vm.
Solution:
The domain H can be divided into rectangular elements with sides lx = rf/M, ly — v/N. The
Jacobian on the cross-hatched element is given by
This result follows from the general formula (11.38), as well as from the simplified form of the
transformation formulae
from which
Simple algebraic manipulations of (11.39) yield
11.3. PROBABILISTIC FINITE ELEMENT METHOD (PFEM) 391
To simplify the calculation according to formula (11.40), we will evaluate the expression on the
right-hand side at four nodes of the element and replace its value at an internal point f, 77 by
the isoparametric approximation. This approach is consistent with the approximation of the
displacement field in FEM. After some manipulations we get
For a rectangular mesh, the coefficients in the expansion can be obtained more easily from
Taylor's formula for J = vd/(4MN):
A similar shortcut can be used to derive the expansion of the operator matrix dT , or of the
matrix B resulting from the application of the operator matrix 8T on the displacement vector
u. Let us start from the transformation formulae and form the operator expansions
To simplify the notation, let us assume that the vector of nodal displacements r starts by
four horizontal displacements followed by four vertical displacements. The matrix of the shape
functions is then
WHERE
Using the previously derived expansions of scalar operators, we obtain the final relation
Note that the functions d and v are to be replaced by their approximations according to (11.1).
i.e., the constants T^r^r^, are not known. The vector r is obtained by solving the
zero-order equations, the vectors r fei , (i = 1,2,..., q) follow from the system of q equations
of the second order. The set of q2 vectors r^bj is in the theory of second-order moments
transformed into a single vector r<2, which will be solved from a single matrix equation
obtained by averaging q2 equations of the second order. Totally, we get (q + 2) vectors
r, r^,..., rfe q ,r 2 , which can be used to evaluate the statistics of the derived quantities e a
(7.
To keep the calculation simple, we will restrict our attention to two types of random
factors, related to the material and to the loading.5
5
The information from the preceding section can help the reader to include random geometry (i.e.,
expansions of B and J) as well.
11.3. PROBABILISTIC FINITE ELEMENT METHOD (PFEM) 393
where
The equation is satisfied for arbitrary variations if the expression in the parentheses equals
zero for any i:
Second-order equations
An approach similar to the first-order equations would lead to a large set of q2 equations.
Averaging can be used as an efficient tool for reducing the size of the problem. The
preceding equation is multiplied by the probability density / = /(&) and integrated over
the domain of definition of the vector 6. Considering that
where
As
we can use the transformation (11.47) and set e = 1 (second-order accuracy) to get
Covariances between the elements of the vector r can be expressed with first-order
accuracy [terms with e2 are neglected in expansion (11.42)]. The calculation is easier to
follow if we replace summation in the expansion by a matrix product, so that
The matrix G has q columns storing the vectors ^,(1 = 1,2, ...,<?). The covariance
matrix of the vector r can thus be expressed using (11.52) as
where
The derived formulae indicate that when the number of random parameters in the vector
r is large and the parameters are strongly statistically dependent, the calculations will be
enormously time consuming. It is, therefore, recommended to transform the vector 6 to
a vector c with a diagonal covariance matrix. Two possible transformations of this kind
were described in Section 11.1.
This page intentionally left blank
Bibliography
[1] Allman, D. J.: A Compatible Triangular Element Including Vertex Rotations for
Plane Elasticity Problems, Computers & Structures, 19, 1-8, 1984.
[2] Allman, D. J.: A Quadrilateral Finite Element Including Vertex Rotations for Plane
Elasticity Problems, Int. J. Num. Meth. Engng., 26, 717-739, 1988.
[3] Altiero, N. J., Sikarskie, D. L.: A Boundary Integral Method Applied to Plates of
Arbitrary Plan Form. Computers and Structures, 9, 1978.
[4] Babuska, L, Door, M. R.: Error Estimates for the Combined h and p-Version of the
Finite Element Method, Numer. Math., 34, 41-62, 1980.
[5] Babuska, L, Miller, A.: The Post-Processing Approach in the Finite Element
Method—Part 3: A Posteriori Error Estimate and Adaptive Mesh Selection, Int. J.
Numer. Meth. Engng., 20, 2311-2324, 1984.
[6] Babuska, L, Rank, E.: An Expert-System Like Feedback Approach in the hp-Version
of the Finite Element Method, FE Anal. Des., 3, 127-147, 1987.
[7] Backlund, J.: Finite Element Analysis of Nonlinear Structures (PhD Dissertation),
Goteborg 1973.
[8] Baehmann, P. L., Shepard, M. S.: A Posteriori Error Estimation for Triangular
and Tetrahedral Quadratic Elements Using Interior Residuals. SCOREC Report
14-1990, RPI, Troy, 1990.
[9] Baehmann, P. L., Wittchen, S. L., Shepard, M. S., Grice, K. R., Yerry, M. A.:
Robust, Geometrically Based Two-Dimensional Mesh Generation, Int. J. Numer.
Meth. Engng., 24, 1043-1078, 1987.
[10] Bakalikova, D.: Stability of Ribbed Shells, (PhD Dissertation), FSv TU, Bratislava,
1988.
[11] Banerjee, P. K., Butterfield, R.: Boundary Elements in Engineering Science. Mc-
Graw Hill, London 1981.
[12] Bathe, K. J.: Finite Element Procedures in Engineering Analysis. Prentice-Hall,
Englewood Cliffs, NJ, 1982.
[13] Bathe, K. J., Dvorkin, E. N.: On the Automatic Solution of Nonlinear Finite Ele-
ment Equations, Computers & Structures, 17, 871-879, 1983.
[14] Bathe, K. J., Dvorkin, E. N.: A Formulation of General Shell Elements—the Use
of Mixed Interpolation of Tensorial Components, Int. J. Num. Meth. in Engng., 22,
697-722, 1986.
397
398 BIBLIOGRAPHY
[15] Bathe, K. J., Dvorkin, E. N.: A Four-Node Plate Bending Element Based on
Mindlin-Reissner Plate Theory and a Mixed Interpolation, Int. J. Num. Meth. in
Engng., 21, 367-3383, 1986.
[16] Bathe, K. J., Wilson, E. L.: Large Eigenvalue Problems in Dynamic Analysis, J.
Engng. Mech. Div. ASCE, 1972.
[17] Bathe, K. J., Wilson, E. L.: Numerical Methods in Finite Element Analysis.
Prentice-Hall, Englewood Cliffs, NJ, 1976.
[18] Batoz, J. L., Bathe, K. J., Ho, L. W.: A Study of Three-Node Triangular Plate
Bending Elements, Int. J. Num. Meth. in Engng., 15, 1771-1812, 1986. 1980.
[19] Batoz, J. L., Dhatt, G.: Incremental Displacement Algorithmus for Non-linear Prob-
lems, Int. J. Num. Meth. Engng., 14, 1262-1266, 1979.
[20] Bazant, Z. P. ed.: Mathematical Modeling of Creep and Shrinkage of Concrete.
John Wiley & Sons Ltd., New York, 1988.
[21] Bazant, Z., P., Cedolin, L.: Stability of Structures: Elastic, Inelastic, Fracture and
Damage Theories. Oxford University Press, New York, 1991.
[22] Bazant, Z. P., Lin, F. B.: Nonlocal Smeared Cracking Model for Concrete Fracture,
J. Struct. Engng. (ASCE), 114 (11), (Sec. 3.10), 1988.
[23] Bazant, Z. P., Liu, K. L.: Random Creep and Shrinkage in Structures: Sampling,
J. Struct. Engng., 5, 1113-1134, 1985.
[24] Bazant, Z. P., Wu, S. T.: Dirichlet Series Creep Function for Aging Concrete, J.
Engng. Mech. Div., ASCE, 99 (EM2), 1973.
[25] Bellini, P. X., Chulya, A.: An Improved Automatic Incremental Algorithm for the
Efficient Solution of Nonlinear Finite Element Equations, Computers & Structures,
26, 99-110, 1987.
[26] Belytschko, T., Fish, J., Engelmann, B. E.: A Finite Element with Embedded
Localization Zones. In: Computer Methods in Applied Mechanics and Engineering
70, North-Holland, Amsterdam, 1988.
[27] Belytschko, T., Lasry, D.: Nonmonotonic Stress-Strain Laws: Bizarre Behavior and
its Repercussions on Numerical Solutions. Transactions of the Sixth Army Confer-
ence on Applied Mathematics and Computing, ARD Report 89-1.
[28] Belytschko, T., Lasry, D.: A Fractal Patch Test, Int. J. Num. Meth. Engng., 26,
2199-2210, 1988.
[29] Bergan, P. G.: Solution Algorithms for Non-linear Structural Problems, Computers
& Structures, 12, 497-509, 1980.
[30] Bergan, P. G., Felippa, C. A.: A triangular membrane element with rotational
degrees of freedom, Comp. Methods Appl. Mech. Engng., 50, 25-69, 1985.
[31] Bittnar, Z., Patzak, B.: Dynamic Analysis of Curved Folded Plates using Finite
Strip with Independent Rotation Fields. In: Structural Dynamics—Eurodyn'93,
eds. Moan et al, Rotterdam, 569-572, 1990.
BIBLIOGRAPHY 399
[32] Bittnar, Z., Reficha, P.: Finite Element Method in Structural Dynamics (in Czech).
SNTL, Praha, 1981.
[33] Bittnarova, J.: State of Stress in Shear-Wall Structural Systems and Their Interac-
tion with Subgrade (in Czech). Report Faculty of CE, CTU, Praha 1979.
[34] de Borst, R.: Non-linear Analysis of Frictional Materials (PhD Dissertation), Delft
University of Technology, Delft, 1986.
[35] de Borst, R.: A Generalization of J2-Flow Theory for Polar Continua, Comp. Meth.
Appl. Mech. Eng, 103, 347-362, 1993.
[36] de Borst, R., Miihlhaus, H. B.: Gradient-Dependent Plasticity. Formulation and
Algorithmic Aspects, Int. J. Num. Meth. Eng., 35, 521-539, 1992.
[37] Brebbia, C. A.: The Boundary Element Method for Engineers. Pentech Press, Lon-
don, 1978.
[38] Brebbia, C. A., Nardini, D.: Dynamic Analysis in Solid Mechanics by Alternative
Boundary Element Approach, Int. J. Soil Dynamics and Earthquake Engng., 2,
1983.
[39] Brebbia, C. A., Telles, J. C. F., Wrobel, L. C.: Boundary Element Techniques.
Springer, Berlin, 1984.
[40] Broz, P., Prochazka, P.: Boundary Element Method in Engineering Practice (in
Czech). SNTL, Praha, 1987.
[41] Broz, P., Prochazka, P.: Solution of Nonlinear Problems of Mechanics Using the
Boundary Element Method (in Czech). Grada Publishing, Praha, 1995.
[42] Carlslaw, H. S., Jaeger, J. C.: Conduction of Heat in Solids, 2nd ed. Oxford Uni-
versity Press, London, 1959.
[43] Chen, W. F.: Plasticity in Reinforced Concrete. McGraw-Hill, New York, 1982.
[44] Chen, H. C., Taylor, R. Z.: Solution of Viscously Damped Linear Systems Using
a Set of Load-Dependent Vectors, Earthquake Engng. Struct. Dyn., 19, 653-665,
1990.
[45] Cheung, Y. K.: Finite Strip Method Analysis of Elastic Slabs, Proc. A.S.C.E., 94,
1365-1378, 1968.
[46] Cheung, Y. K.: Folded Plate Structures by Finite Strip Method, Proc. A.S.C.E.,
96, 2963-2979, 1969.
[47] Cervenka, V., Eligehausen, R., Pukl, R.: SBETA, Computer Program for Nonlin-
ear Finite Element Analysis of Reinforced Concrete Structures. Mitteilungen IWB,
Stuttgart, 1990/1.
[48] Collatz, L.: Eigenvalue Problems with Technical Applications (Czech edition).
SNTL, Praha, 1965.
[49] Crisfield, M. A.: Accelerating and Damping the Modified Newton-Raphson method,
Compters & Structures, 18, 395-407, 1984.
400 BIBLIOGRAPHY
[50] Crisfield, M. A.: Non-linear Finite Element Analysis of Solids and Structures. John
Wiley & Sons, London, 1991.
[51] Crisfield, M. A., Wills, J.: Solution Strategies and Softening Materials, Comp. Meth.
Appl. Mech. & Engng., 66, 267-289, 1988.
[52] Darwin, D., Pecknold, D. A. W.: Analysis of Cyclic Loading of Plane RIG Struc-
tures, Computers & Structures, 7, 137-147, 1977.
[53] Decker, D. W., Keller, H. B.: Solution Branching—A Constructive Technique. In:
New Approaches to Nonlinear Problems in Dynamics (ed. P. J. Holmes), SIAM,
Philadelphia, PA, 1980.
[54] Delpak, R., Peshkam, V.: A Study of the Influence of Hierarchical Modes on the
Performance of Selected Parametric Elements, Int. J. Num. Meth. Engng., 22, 153-
171, 1986.
[56] Dhatt, G.: Numerical Analysis of Thin Shells by Curved Triangular Elements Based
on Discrete Kirchhoff Hypothesis. Proc. ASCE, Symp. on Applications of FEM in
Civil Engineering, Vanderbilt Univ., Nashville, Tenn., 13-14, 1969.
[58] Dvorak, G. J., Prochazka, P.: Thick-Walled Composite Cylinders With Optimal
Fiber Prestress. Submitted to Composite Engineering.
[59] Eringen, A. C.: Nonlinear Theory of Continuous Media. McGraw-Hill, New York,
1962.
[60] Eshelby, J. D.: The Determination of the Elastic Field of an Ellipsoidal Inclusion,
and Related Problems. J. Mech. Phys. Solids, A 241, 376-396, 1957.
[61] Fadeev, D. K., Fadeeva, V. N.: Vycislitelnye Metody Lineinoi Algebry. Fizmatgiz,
Moskva, 1960.
[63] Florian, A.: Statistical Dependence between Input Quantities in Statistical Analysis
of Computational Models (in Czech). Building J., 37, 12, 895-906, 1989.
[64] Gospodinov, G., Ljutskanov, D.: The Boundary Element Method Applied to Plates.
Appl. Math. Modelling, Vol. 6, August 1982.
[65] Grafton, P. E., Strome, D. R.: Analysis of Axisymmetric Shells by the Direct Stiff-
ness Method, J.A.I.A.A., 2342-2347, 1963.
[66] Green, A. E., Zerna, W.: Theoretical Elasticity. Oxford University Press, London,
1968.
BIBLIOGRAPHY 401
[67] Gui, W., Babuska, I.: The /i, p and h - p Versions of the Finite Element Method
in 1 Dimension: Part 1: The Error Analysis of the p-Version, Numer. Math., 49,
577-612, 1986.
[68] Gui, W., Babuska, L: The /i, p and h - p Versions of the Finite Element Method
in 1 Dimension: Part 2: The Error Analysis of the h and h — p Versions, Numer.
Math., 49, 659-683, 1986.
[69] Hanuska, A.: On the Westergaard Theory of Reinforced Subgrade (in Czech), Build-
ing J., 39, 2, 1989.
[70] Hashin, Z., Shtrikman, S.: On Some Variational Principles in Anisotropic and Non-
homogeneous Elasticity, J. Mech. Phys. Solids, 10, 335-342, 1963.
[71] Hashin, Z., Shtrikman, S.: A Variational Approach to the Theory of the Elastic
Behavior of Multi-phase Materials, J. Mech. Phys. Solids, 11, 127-140, 1963.
[72] Herrmann, L. R.: Finite Element Bending Analysis for Plates, J. Engng. Mech. Div.
ASCE, EMS, 1967.
[73] Hill, R.: The Mathematical Theory of Plasticity, Oxford University Press, New
York, 1950.
[74] Hill, R.: A General Theory of Uniqueness and Stability in Elastic-Plastic Solids, J.
Meth. Phys. Solids, 6, 236-249, 1958.
[75] Hinton, E., Owen, D. R. J.: Finite Element Software for Plates and Shells. Pineridge
Press, Swansea, 1984.
[76] Hinton, E., Rock, T., Zienkiewicz, 0. C.: A Note on Mass Lumping and Related
Processes in the Finite Element Method, Int. J. Earthq. Engng. Struct. Dyn., 4,
245-249, 1976.
[77] Holicky, M.: Fuzzy Concept of Serviceability Limit States. In: Symp. Workshop on
Serviceability of Buildings, Canada-Ottawa, 19-31, 1988.
[78] Holzer, S., Rank, E., Werner, H.: An Implementation of the hp-Version of the Finite
Element Method for Reissner-Mindlin Plate Problems, Int. J . Num. Meth. Engng.,
30, 459-471, 1990.
[79] Huang, H. C.: Elastic and Elasto-Plastic Analysis of Shell Structures Using the
Assumed Strain Elements, Computers & Structures, 33, 327-335, 1989.
[80] Hughes, T. J. R.: The Finite Element Method. Prentice-Hall, Englewood Cliffs, NJ,
1987.
[81] Hughes, T. J. R., Brezzi, F.: On Drilling Degrees of Freedom, Comp. Methods Appl.
Mech. Engng., 72, 105-121, 1989.
[82] Hughes, T. J. R., Tezduyar, T. E.: Finite Elements Based Upon Mindlin Plate
Theory with Particular Reference to the Four-Node Bilinear Isoparametric Element,
J. Appl. Mech. ASME, 46, 587-596, 1981.
[83] Hurty, W., Rubinstein, M. F.: Dynamics of Structures. Prentice-Hall, Englewood
Cliffs, NJ, 1964.
402 BIBLIOGRAPHY
[84] Huseyin, K.: Nonlinear Theory of Elastic Stability. Noordhoff, Leyden, 1975.
[85] Ibrahimbegovic, A., Chen, H. C., Taylor, L. R., Wilson, E. L.: Ritz Method for Dy-
namic Analysis of Large Discrete Linear Systems with Non-proportional Damping,
Earthquake Engng. Struct. Dyn., 19, 877-889, 1990.
[86] Ibrahimbegovic, A., Taylor, R. L., Wilson, E. L.: A Robust Membrane Quadrilateral
Element with Drilling Degrees of Freedom, Int. J. Num. Meth. Engng., 30, 445-457,
1990.
[87] Ibrahimbegovic, A., Wilson, E. L.: A Methodology for Dynamic Analysis of Lin-
ear Structure-Foundation Systems with Local Non-linearities, Earthquake Engng.
Struct. Dyn., 19, 1197-1208, 1990.
[88] Ibrahimbegovic, A., Wilson, E. L.: Thick Shell and Solid Finite Elements with
Independent Rotation Fields, Int. J. Num. Meth. Engng., 31, 1393-1414, 1991.
[89] Irons, B. M.: Quadrature Rules for Brick Based Finite Elements, Int. J. Num. Meth.
Engng, 3, 293-294, 1971.
[90] Irons, B. M, Razzaque, A.: Experience with the Patch Test for Convergence of Fi-
nite Element Methods, In: Mathematical Foundations of the Finite Element Method
(ed. A. K. Aziz), Academic Press, New York, 1982.
[91] Ishlinskii, I. U.: General Theory of Plasticity with Linear Strain Hardening (in
Russian), Ukr. Mat. Z, 6, 1954.
[93] Jendele, L.: Nonlinear Analysis of Reinforced Concrete Structures (in Czech, PhD
Dissertation), KI CTU, Praha, 1989.
[94] Jennings, A.: Matrix Computations for Engineers and Scientists. John Wiley, Lon-
don, 1977.
[96] Kafka, V.: Inelastic Mezomechanics. World Scientific Publishing, Volume 5, Singa-
pore, 1987.
[99] Koiter, W. T.: On the Stability of Elastic Equilibrium (Engl. transl). AFFDL.1970.
[100] Kolaf, V. et al.: Berechnung von Flachen- und Raumtragwerken nach der Methode
der finiten Elernente. Springer Verlag, New York, 1975.
BIBLIOGRAPHY 403
[101] Kolaf, V., Nemec, I.: Modelling of Soil-Structure Interaction. Academia, Elsevier,
Amsterdam, 1990.
[102] Kolaf, V., Nemec, L: Contact Stress and Settlement in the Structure-Soil Interface.
Academia Praha, 1991.
[103] Kfistek, V.: Theory of Box Girders. John Wiley & Sons., Chichester, 1979.
[104] Krohn, R,: Berechnung Statisch Unbestirnmter Fachwerktraeger, Zeitschrift Arch,
und Ing. Ver., Hannover, 1984.
[105] Kuklik, P.: A Contribution to the Solution of a Layered Subgrade (in Czech),
Pozemni stavby, 7, 1984.
[106] Lee, S. H.: Rudimentary Considerations for Effective Quasi-Newton Updates in
Nonlinear Finite Element Analysis, Computers & Structures, 33, 463-476, 1989.
[107] Lekhnitski, S. G.: Theory of Elasticity of an Anisotropic Elastic Body. Holden Day,
San Francisco, 1963.
[108] Lekhnitski, S. G.: Anisotropic Plates. Gordon and Breach Science Publishers, New
York, 1968.
[109] Levin, I. M.: Determination of Composite Material Elastic and Thermoelastic Con-
stants, Izd. AN SSSR, Mekhanika Tverdogo Tela, 11, 6, 137-145, 1976.
[110] Liu, W. K., Belytschko, T., Mani, A.: Random Field Finite Elements, Int. J. Num.
Meth. Engiig., 23, 1831-1845, 1986.
[Ill] Liu, W. K., Besterfield, G. H., Belytschko, T.: Variational Approach to Probabilistic
Finite Elements, J. Engng. Mech., Vol.114, 12, 2115-2133, 1988.
[112] MacNeal, R. H., Harder, R. L.: A Refined Four-Noded Membrane Element with
Rotational Degrees of Freedom, Computers & Structures, 18, 75-84, 1988.
[113] Maier, G., Novati, G., Sirtori, S.: Symmetric Formulation of an Indirect Boundary
Element Method for Elastic-plastic Analysis and Relevant Extremurh Properties.
Pergamon Press, Toronto, 1988.
[114] Malone, J. G., Plunkett, R., Hodge, P. G.(Jr.): An Elastic-Plastic Finite Element
Solution for a Cracked Plate. In: Finite Elements in Analysis and Design 2, Elsevier
Science Publishers B.V., North-Holland, Amsterdam, 1986.
[115] Malvern, L. E.: Introduction to the Mechanics of Continuous Medium. Prentice-
Hall, Englewood Cliffs, NJ, 1969.
[116] Marquis, D.: Modelisation et Identification de 1'Ecrouissage Anisotrope des Metaux,
These, Universite Paris, 1979.
[117] Matsumoto, T., Yuuki, R.: Accurate Boundary Element Analysis of Two-
dimensional Elasto-plastic Problems. In: Boundary Element Methods in Applied
Mechanics. Pergamon Press, Toronto, 1988.
[118] Melosh, R. J.: Structural Engineering Analysis by Finite Elements. Prentice-Hall,
London, 1990.
404 BIBLIOGRAPHY
[119] Mittelman, H. D., Weber, H.: Numerical Methods for Bifurcation Problems-a Sur-
vey and Classification. In: Bifurcation Problems and Their Numerical Solution (eds.
H. D. Mittelman and H. Weber), Birkhauser, Basel, 1980.
[120] Moreau, J. I., Panagiotopoulos, P. D., Strang, G.: Topics in Nonsmooth Mechanics.
Birhauser Verlag, Berlin, 1988.
[121] Miihlhaus, H. B., Aifantis, E. C.: A Variational Principle for Gradient Plasticity,
Int. J. Solids Structures, 28, 845-857, 1991.
[122] Muskhelishvili, N. I.: Some Basic Problems of the Mathematical Theory of Elastic-
ity. Noordhoff, Croningen,1963.
[123] Newmark, N. M.: A Method of Computation for Structural Dynamics, J. Engng.
Mech. Div. ASCE, 85, 67-94, 1959.
[124] Nour-Ornid, B., Clough, R. W.: Dynamic Analysis of Structures Using Lanczos
Coordinates, Earthquake Engng. Struct. Dyn., 12, 566-577, 1984.
[125] Novak, D., Teply, B. Estimation of Structural Failure Probability. In: European
Conference on New Advances in Computational Structural Mechanics. France-Giens
(Var), 1991.
[126] Oden, J. T., Reddy, J. N.: Variational Methods in Theoretical Mechanics. Springer
Verlag, Berlin, 1976.
[127] Onate, E., Suarez, B.: A Unified Approach for the Analysis of Bridges, Plates
and Axisymmetric Shells Using the Linear Mindlin Strip Elements, Computers &;
Structures, 17, 407-426, 1983.
[128] Onate, E., Suarez, B.: A Comparison of the Linear Quadratic and Cubic Mindlin
Strip Elements for the Analysis of Thick and Thin Plates, Computers fe Structures,
17, 427-439, 1983.
[129] Ondracek, E., Valentik, V.: Dynamical States in Complicated Bodies, Colloquium
on Solid States Properties at High Loading Rates, Brno, 1978.
[131] Park, K. C.: A Family of Solution Algoritms for Nonlinear Structural Analysis
Based on the Relaxation Equations, Int. J. Num. Meth. Engng., 18, 1337-1347,
1982.
[133] Pian, T. H. H.: Finite Element Methods bz Variational Principles with Relaxed
Continuity Requirements. In: Variational Methods in Engineering, Vol.1, 3/1-3/24,
Southampton University Press, 1973.
[134] Pluhaf, M.: Nonlinear Analysis of Slender Structures Using Perturbation Method
(in Czech), Building J., 38, 12, 1990.
BIBLIOGRAPHY 405
[135] Polizzotto, C., Zito, M.: A Variational Approach to Boundary Element Methods.
In: Boundary Element Methods in Applied Mechanics, Pergamon Press, Toronto,
1988.
[136] Prager, W.: Recent Developments in the Mathematical Theory of Plasticity, J.
Appl. Mech, 20, 1949.
[137] Prager, W.: The Theory of Plasticity: A Survey of Recent Achievements, Proc.
Instn. Mech. Engrs. 169, 1955.
[138] Prochazka, P., Sejnoha, J.: Behavior of Composites on Bounded Domain (to be
published in Boundary Element Communication).
[139] Prochazka, P., Sejnoha, J.: Optimization of Structures Using Transformation Field
Analysis. Structural Optimization (Opti' 95), eds. Hernandez, El-Sayed, Brebbia.
Comp. Mech. Publ., Southampton, Boston, 27-34, 1995.
[140] Ramm, E.: The Riks/Wempner Approach—an Extension of the Displacement Con-
trol Method in Non-linear analysis. In: Non-linear Computational Mechanics (ed.
E. Hinton et al.), Pineridge Press, Swansea, 1982.
[141] Rank, E., Babuska, I.: An Expert System for the Optimal Mesh Design in the hp-
Version of the Finite Element Method, Int. J. Num. Meth. Engng., 24, 2087-2106,
1987.
[142] Rees, D. W. A.: On Isotropy and Anisotropy in the Theory of Plasticity, Proc. R.
Soc. Lond., A383, 1982.
[143] Reissner, E.: A Note on Variational Principles in Elasticity, Int. J. Solids Struct.,
1, 93-95, 1965.
[144] Rice, J. R.: A Path-Independent Integral and the Appropriate Analysis of Strain
Concentration by Notches and Cracks, J. Appl. Mech., 35, 1968.
[145] Rice, J. R.: The Localization of Plastic Deformation in Theoretical and Applied
Mechanics. In: 14th Congr. Theoret. Appl. Mech.(ed. W. T. Koiter), North-Holland,
Amsterdam, 1977.
[146] Riks, E.: The Application of Newton's Method to the Problem of Elastic Stability,
J. Appl. Mech., 39, 1060-1066, 1972.
[147] Reficha, P.: A Shell Element for Transient Dynamic Loading (in Czech). Report
Klokner's Institute CTU 314/75, Praha, 1975.
[148] Rudnicki, J. W., Rice, J. R.: Conditions for the Localization of Deformation in
Pressure-Sensitive Dilatant Materials, J. Mech. Phys. Solids, 13, 371-394, 1975.
[149] Scheaffer, R. L., McClave, J. T.: Probability and Statistics for Engineers, 2nd ed.,
Duxbury Press, Boston, 1986.
[150] Schueller, G. I., Bucher, C. G., Bourgund, U., Quypornprasert, W.: On Efficient
Computational Schemes to Calculate Structural Failure Probabilities, Probabilistic
Engineering Mechanics, 4(1), 10-18, 1989.
406 BIBLIOGRAPHY
[151] Shen, C. H., Roeck, G., Laethem, M., Geyskens, P.: Multi-level Substructuring
and an Experimental Self-adaptive Newton-Raphson Method for Two-dimensional
Nonlinear Analysis, Computers & Structures, 33, 489-497, 1989.
[152] Shield, R. T., Ziegler, H.: On Prager's Hardening Rule, Z. Ang. Math, and Phys.,
9a, 1958.
[153] Simo, J. C.: A Framework for Finite Strain Elastoplasticity Based on the Multi-
plicative Decomposition and Hyperelastic Relations; Part I: Formulation, Comp.
Methods Appl. Mech. Engng., 66, 199-219, 1988.
[154] Simo, J. C.: A Framework for Finite Strain Elastoplasticity Based on the Multiplica-
tive Decomposition and Hyperelastic Relations; Part II: Computational Aspects,
Comp. Methods Appl. Mech. Engng., 66, 1-31, 1988.
[155] Simo, J. C., Wriggers, P., Schweizerhof, K. H., Taylor, R. L.: Finite Deformation
Postbuckling Analysis Involving Inelasticity and Contact Constraints. In: Innovative
Methods for Nonlinear Problems (ed. W. K. Liu et al.), Pineridge Press, Swansea,
365-388, 1984.
[156] Simon, H. D.: The Lanczos Algorithm with Partial Reorthogonalization, Math.
Comput. 42, 115-142, 1984.
[157] Sirtori, S.: General Stress Analysis Methods by Means of Integral Equations and
Boundary Elements, Meccanica, 14, 210-218, 1979.
[158] Sladek, V., Sladek, J.: Two Approaches to Solution of Elastodynamic Problems
Using Boundary Element Formulations, Acta Technica SAV, 32, 1, 1987.
[159] Spence, A., Jepson, A. D.: Folds in the Solution of Two Parameter Systems and
Their Calculation. Part I, SIAM, J. Numer. Anal., 22, 347-368, 1985.
[160] Stolarski, H., Belytschko, T., Carpenter, N., Kennedy, J. M.: A Simple Triangular
Curved Shell Element, Eng. Comput., Vol. 1, September 1985.
[161] Sejnoha, J.: Vibration of Box Structures (in Czech), Building J., 6, 1970.
[162] Sejnoha, J.: A Stochastic Model of Shear-Wall Systems (in Czech), Acta Polytech-
nica CTU, 18, 7, 35-41, 1982.
[163] Sejnoha, J., Bittnarova, J., Blazek, V., Kalouskova, M.: Mathematical Models for
a Dynamic Analysis of Tall Buildings, Building Journal, 38, 2, 1990.
[164] Sejnoha, J., Jirasek, M.: Generalized Theory of Thin-Walled Beams. Proc. of the
International Conference on Lightweight Structures in CE, 414-420, Warsaw, 1995.
[165] Sejnoha, J., Prochazka, P., Feraidon, A.: Reliability-based Design of Composite
Cylinders and Plates. Int. Conf. on Lightweight Structures in Civil Engng., Warsaw,
1995.
[166] Taylor, R., Simo, J. C., Zienkiewicz, O. C., Chan, A.: The Patch Test—a Condition
for Assessing FEM Convergence, Int. J. Num. Meth. Engng., 22, 39-62, 1986.
[167] Telles, J. C. F., Brebbia, C. A.: Elastoplastic Boundary Element Analysis in Struc-
tural Mechanics, Ruhr-Universitat Bochum, 1980.
BIBLIOGRAPHY 407
[168] Telles, J. C. F., Brebbia, C. A. On the Application of the Boundary Element Method
to plasyicity, appi. math modelling, 3 466-4701981
[169] Timoshenko, S., Goodier, J. N.: Theory of Elasticity (2nd ed.). McGraw-Hill, New
York, 1959.
[170] Tottenham, H.: The Boundary Method for Plates and Shells. In: Developments in
BEM. Red. Banerjee P. K., Butterfield R. London, Applied Science Publishers Ltd.
1979.
[171] Truesdell, C.: The Mathematical Foundations of Elasticity and Fluid Dynamics, J.
Rational Mech. and Analysis, 1, 3, 1952.
[172] Truesdell, C.: The Mathematical Foundations of Elasticity and Fluid Mechanics.
Gordon and Breach Science Publishers, Inc., New York, 1966.
[173] Turner, M. J., Clough, R. W., Martin, H.C., Topp, L. J.: Stiffness and Deflection
Analysis of Complex Structures, J. Aeronaut. Sci, 23, 1956.
[174] Washizu, K.: Variational Methods in Elasticity and Plasticity. Pergamon Press,
New York, 1975.
[175] Wempner, G. A.: Discrete Approximations Related to Nonlinear Theories of Solids,
Int. J. Solids & Structs., 7, 1581-1599, 1971.
[176] Willam, K. J., Etse, G.: Failure Assessment of the Extended Leon Model for Plain
Concrete, Proc. Second Int. Conf. Computer Aided Analysis and Design of Concrete
Structures, Eds. Bicanic et al, Pineridge Press, Swansea, 851-870, 1990.
[177] Willis, J. R.: Bounds and Self-consistent Estimates for the Overall Properties of
Anisotropic Composites. J. Mech. Phys. Solids, 25, 185-202, 1977.
[178] Wilson, E. L.: Structural Analysis of Axisymmetric Solids, J.A.I.A.A., 3, 2269-2274,
1965.
[179] Wriggers, P., Simo, J. C.: A General Procedure for the Direct Computation of
Turning and Bifurcation Points, Int. J. for Numer. Meth. Engng., 30, 155-176,
1990.
[180] Zareckij, lu. K.: Teoria uprugosti gruntov. Nauka, Moskva, 1967.
[181] Zienkiewicz, O. C.: Basic Formulation of Static and Dynamic Behaviour of Soil and
Other Porous Media. Institute for Numerical Methods in Engineering, University
College of Swansea, 1983.
[182] Zienkiewicz, 0. C., Taylor, R. L.: The Finite Element Method-4th ed. Vol.1: Basic
Formulation and Linear Problems. McGraw-Hill, London, 1994.
[183] Zienkiewicz, O. C., Taylor, R. L.: The finite element method-4th ed. Vol.2: Solid
and fluid mechanics. Dynamics and non-linearity. McGraw-Hill, London, 1994.
[184] Zienkiewicz, O. C., Zhu, J. Z.: A Simple Error Estimator and Adaptive Procedure
for Practical Engineering Analysis, Int. J. Numer. Meth. Engng., 24, 337-357, 1987.
[185] Zienkiewicz, O. C., Zhu, J. Z.: Error Estimates and Adaptive Refinement for Plate
Bending Problems, Int. J. Numer. Meth. Engng., 28, 2839-2853, 1989.
Appendix A
Matrix Formulation of Gauss
Elimination
Elements li+jj are Gauss factors required to zero the columns below the main diagonal, and
superscript (i) means that the elements A: are taken from the matrix L~\ - - -L^1L\1K.
The inverse of L~l is constructed by a simple sign reversal at the elements below the
diagonal. Based on this we can write
where D is a diagonal matrix with elements equal to the diagonal elements of S, and S
is an upper triangular matrix with unit elements on its diagonal. After substitution from
(A.4) into (A.3) we get
408
APIWD/.X 1. MATRIX PORMl !IATIOi\ ()/•' CAl XS' MJMIKATIOX 409
Transposition gives
This decomposition can be exploited when solving the set of linear equations. The solution
is performed in two steps. First, we introduce an auxiliary vector v for which
Let us multiply (A. 10) from the left, first by L'1, and afterward by D"1. We obtain
This set of linear algebraic equations is easy to solve because LT is an upper triangular
matrix with unit elements on the diagonal. The last element of the vector D~ v is equal
to the unknown r n . The remaining elements of r are evaluated by backward substitution.
Expressing from (A.9)
we can see that v is obtained by a simple modification of the right-hand side (vector of
transformed loads) using the same Gauss factors as those producing the decomposition
of the stiffness matrix.
A clear and simple algorithm, which also takes advantage of the banded structure of
the stiffness matrix, is given in [17].
Appendix B
Numerical Integration
where <5jj is the Kronecker delta. The value of <p(£) at an arbitrary point of the interval
< — 1, 1 > is computed as
The subscripts run from zero because a polynomial of the nth degree has also a zero-degree
term.
410
APPENDIX 411 NIMERIC INTERTARION 411
The simplest way to select the location of the evaluation points is to divide the interval
< —1,1 > into n equal segments. This leads to the points
#n denotes the integration error, /j are the values at points £ = &, and <7f are the
Newton-Cotes constants for numerical integration with (n + 1) points. The values of the
constants taken from [17] are listed in Table B.I
The constants from this table can be used for integration on a general interval < a, b >.
The integration formula is generalized to1
1
f11, fiv , /v 7, /v 7// are respectively the second, fourth, sixth and eighth derivative of / with respect
tof.
412 APPENDIX B. NUMERICAL INTEGRATION
For n = 1 this formula gives the trapezoidal rule, and for n = 2 we get the well-known
Simpson rule.
The table shows that the upper bound on the integration error is the same for n = 2
as for n = 3. A similar situation arises for n = 4 and n = 5. This implies that it is not a
good idea to use an odd number of segments.
As already stated, the accuracy can be increased by refining the discretization of the
integration domain. Here we have two options—either to use a more accurate formula,
or to repeat a simple formula on several subintervals. Engineering applications almost
exclusively opt for the latter approach, because it is simpler and more universal. It is
applicable to functions that are not very smooth, which is a frequent case. An example
of such function is given in Fig. B.2.
If we select 1 and 2 as the evaluation points and apply the trapezoidal rule on each
of the three subintervals, we obtain the exact result. However, if we use even a very
accurate Newton-Cotes formula, we may get a good approximation but never the exact
result. The reason is that we try to approximate a function with a discontinuous first
derivative by a function with all derivatives continuous. On the other hand, if the function
to be integrated is sufficiently smooth, a higher-order formula gives better results than a
repeated lower-order formula with the same number of points.
So far we have discussed the case in which the domain of integration is divided into
equal segments. Let us now explore the potential improvement resulting from a more
sophisticated distribution of the evaluation points. We will derive the Gauss integration
formula, which is the standard integration technique for isoparametric elements.
The Gauss integration formula approximates the value of the integral BY
where the weights a\, a^,..., a.n as well as the coordinates of the evaluation points £1,
£2 v» £,n are to be determined. We want to select these parameters such that the formula
gives the most accurate results. In contrast to the previous approach, in which only
the Newton-Cotes constants were free parameters, we now have a double number of free
parameters with the same number of evaluation points. Intuitively it may be expected
that we can construct a more accurate integration formula.
Now let us construct the approximation functions. The general form of the approxi-
mation is the same as for the Newton-Cotes formulae. This means that the function /(£)
is approximated by a function <^(£) defined as
A!>/>/j\/)/\ n. MJMMtICA L li\TM;KA TION 413
The coordinates of n points at which we fit the approximation to the actual values are
yet to be determined. The approximation function is a polynomial of degree (n - 1).
However, as we have 2n parameters at our disposal, the approximation function can be a
polynomial of degree up to (2ra - 1). Let us look for the second part of the approximation
that increases the order of the polynomial but vanishes at all n evaluation points. Such
a formula is easily constructed as a product of the function
The general polynomial in (B.7) in fact represents the Taylor series, and so the sum
can approximate the function /(£) with an arbitrary accuracy. Moreover,
we know that convergence of the partial sums to /(£) is monotonic. For the purpose of
approximation, it suffices to truncate the polynomial (B0 -f BI£ 4- B^2 -f ....) to the first
n terms and write
Now we have two options. We could find suitable conditions for the determination of
the coefficients Bi and perform the integration. However, this would not lead to any con-
ditions for the location of the evaluation points and we could distribute them arbitrarily.
A better approach is to select the coordinates & such that all n integrals
vanish. This condition leads to a set of nonlinear equations of the form
Solving this system we obtain the values f i , f 2 i — > f n , which are the coordinates of the
evaluation points. It is known that Legendre polynomials have the property (B.10), and
so the coordinates & are the roots of Legendre polynomials. Once we know &, it is easy
to evaluate
With the present choice of & Bi it is not necessary to compute the coefficients BI, even
though they are nonzero. The function /(£) is approximated by a polynomial of degree
(2n-l).
To illustrate the development of an integration formula we will demonstrate the eval-
uation of & and on for n = 1,2,3. First, note that the points & must be located symmet-
rically with respect to the origin because there is no reason why one half of the interval
< -1,1 > should be treated differently from the other one. Consequently, for an odd
value of n we always place one of the evaluation points into the origin. This reduces the
number of nonlinear equations to be solved.
The simplest case is n = 1. As 1 is an odd number, we have & = 0. The corresponding
interpolation function is hi = 1. This implies c*i = fLi 1 df - 2. For n = 2 we have to
414 APPMDIX It. NUMKIUCAI. IXTMHATION
place two points symmetrically with respect to the origin, i.e., fi = —£2- We have to
satisfy the conditions
Due to the relation £1 = —£2, the second condition is fulfilled by default. The first
condition yields
Finally, for n = 3 we have to locate three points, knowing that f2 = 0 and fi = —£3.
Three conditions have to be satisfied:
The reader can easily verify that the first and the third condition are satisfied identically
if £2 = 0, £1 = —£3. The second condition yields
1 0,000000000000000 2,000000000000000
2 ±0,577350269189626 1,000000000000000
3 ±0,774596669241483 0,555555555555555
0,000000000000000 0,888888888888888
4 ±0,861136311594053 0,347854845137454
±0,339981043584856 0,652145154862546
5 ±0,906179845938664 0,236926885056189
±0,538469310105683 0,478628670499366
0,000000000000000 0,568888888888889
6 ±0,932469514203152 0,171324492379170
±0,661209386466265 0,360761573048139
±0,238619186083197 0,467913934572691
n 6 »?• Oii
417
Cylindrical shell 350 59 Fatigue fracture 247
d'Alembert principle 44, 45, Dugdale-Barenblatt model 34 FEM. See Finite element
283 Duhamel integral 197 method
Damage mechanics 24 Dynamic correction 200-201 Finite element method 15, 48,
Damage theory 23-38 Dynamic problems 283-288 246; adaptive techniques
Damped eigenvibration 193, Effective stress 240-241 369-373; algorithms 302;
211 Eigenmodes, orthogonality of alternative approach 120-
Damped vibration eigenmodes 172-173 123; convergence criteria
212 Eigenstrains 55-59 54-55, 367-369; diffusion
Damping matrices 195-196, Eignestresses 55-59 equation 238-239; dis-
198-199 Eigenvibration analysis 168, cretization 319-335; dis-
Dams 110 174-193 placement based formula-
Darcy's law 242 Eigenvibrations of linear sys- tion 249; fracture mechanics
Decomposition of displace- tems 171-172, 190 246-257; mixed formulation
ments 148 Elastic foundation 64, 67-72, 128; p-version 365-369;
Decomposition, principle of 141-144,230 shape function 53
83 Elastic-plastic body 59 Finite strip method 215, 218-
Deflection curve 217, 228-229 Elastically supported plate 132 222, 224, 227
Deflections of the beam 89 Elasticity, basic equations 9- Finite strips, order of integra-
Deformation 73, 153; soils 10 tion 222
and porous materials 240- Elasticity equations, tensorial First law of thermodynamics
246; solid skeleton 241-242 form 15-16 61
Deformation theory of plastic- Elasticity equations, transfor- Fixed coordinate system 328-
ity 255 mation of 64-67 330
Deformed beam, equivalence Elastoplastic materials 16-23, Flexibility approach 234-236,
offerees on 170 358-359; constitutive equa- 237
Degenerate continuum 333 tions 20-23 Flexibility matrix 50
Degree of saturation 240 Element with curved boundary Force and displacement para-
Degrees of freedom, reduction 102 meters 76
of 188-190 Elements with rotational Force approach 73-77; transi-
Diagonalization 198-199 degrees of freedom 1 60 tion to displacement
Differential localization lim- Energy criteria of fracture approach 75
iters 36-37 251-254 Force methods 47, 50
Diffusion equation 238-239 Energy density 10-11 Forced vibration 194-207
Direct integration 199, 201- Energy-based beam analysis Foundation beams 67-72, 152-
205 215-217 153, 156; interaction 154
Discontinuum modeling 31-35 Enhanced continuum approach Foundation structures, interac-
Discrete Kirchhoff theory 35-38 tion of 153-157
(DKT) 136-138, 147 Equation of continuity 242- Foundation structures, nonin-
Displacement approach 77-83 243 teracting 152-153
Displacement methods 47, 50 Equation of motion 168-169, Fourier law 238
Displacement velocity 316 170 Fourier series 215-217
Displacements and loading, Equilibrium, equations of 59, Fracture energy 32
time history of 246 242-243; general principle Fracture mechanics 24, 246-
Displacements and pore pres- of 44 257
sure, vertical distribution of Error estimates 369-373 Fracture toughness 34
248 Euler equations 135,335-336 Framed structure 192
Divergence theorem 10 Euler-Lagrange equations 352 Free energy density 24
DKT. See Discrete Kirchhoff Eulerian description 299 Free formulation 120-123
theory Evolution equation 25 Free torsion 236-238
Drucker's postulate of stabili- Expert systems 373-377 Galerkin method 231, 275
ty 20-21 Explicit method 202 Gauss elimination 408-409
Dual variational principle 57- Extended systems 350-353 Gauss formula 234
418
Gauss integration rule 1 17 Infinitesimal vector 308 Kirchhoff theory of thin plates
Gauss multipliers 177 Initial deformation matrix 302 128, 129, 136-138
Gauss theorem 10, 134, 238- Initial displacement matrix Kirchoff s assumptions 67, 98,
239,243-244,312 322, 330 99
Gaussian quadrature 114, 142- Initial strain problem 358-359, Knowledge base 375-377
143, 236, 246 361 Krohn theorem 77, 332
Geometric nonlinearity 298, Initial stress matrix 169-170, Kronecker delta 15
299, 300-306, 333-335; con- 302, 322, 327-328, 330 Krylov series 187
tinuum theory 306-319 Initial stress problem 359-361 Lagrange principle 45-46, 47,
Geometrically nonlinear struc- Initial yield surface 19 53,77, 116-117, 128,387;
tures 319-335 Integral fracture parameters modified 49-50
Generalized variational princi- 247 Lagrangian description of
ple 123-125, 126 Integration, recommended deformation 306-3 1 1
Global coordinate system 13, order 115 Lagrangian formulation 299,
94-99, 151 Internal energy 62 3 14; stress state 3 11-3 12.
Global stiffness matrix 226 Interpolation functions 101, See also Total Lagrangian
Gradient-dependent softening 221-222; incompatible 119; (TL) formulation; Updated
plasticity theory 37-38, 355- quadrilateral element 125- Lagrangian (UL) formula-
356 127; two-dimensional ele- tion
Gramm-Schmidt orthogonal- ment 102 Lanczos method 187-192;
ization 181, 187 Inverse iteration 179-181 algorithm of 189; applica-
Green's formula 59, 289, 290, Inverse power. See Inverse tion to damped eigenvibra-
296 iteration tions 193
Green-Lagrange strain tensor Irwin-Orowan Gamma-criteri- Lanczos vectors, construction
304-305,308-311,313,314, on 32, 252-253 of 187-188, 192
317,319,320,334 Irwin's stress intensity factor Laplace equations 230
Grillages 64, 67, 85-88 247 Latin Hypercube Sampling
Guyan reduction 177 Isochrones 39 (LHS) 380, 385-387
Hamilton principle 46, 48 Isoparametric bilinear quadri- LEFM. See linear elastic frac-
Hardening 21-22 lateral element 114-115 ture mechanics
Harmonic excitation 207-212 Isoparametric discretization Linear elastic fracture
Hashin-Shtrikman theorem 56, 333-335 mechanics 34, 254
57 Isoparametric elements, basic Linear elastic materials 10-16
Hellinger-Reissner principle relations 100-104 Linear elasticity 259
47, 128, 246; modified 48- Isothermal processes 62 Linear hardening 255
49 Isotropic medium 13 Linear interpolation 22 1
Helmholtz free energy 62 Isotropic tensor 15 Linear stability 169 170-171
Hierarchical bubble function Isotropic nonhomogenous Linear stiffness matirx 302,
118 bodies 55-59 322, 330
Hierarchical elements 365-369 Iterative solution 199 Linear systems, eigenvibra-
Hooke's law 56-57, 280 Jacobi method of rotations tions of 17 1-172; forced
Horizontal displacements 277 181-185 vibration of 194-207
Hu-Washizu variational prin- Jacobian 101, 103-104,313 Linear tetrahedron 161
ciple 47, 272 JADRO program 151, 154, Linearized arc-length mehtod
Hutchinson, Rice and 156, 276 340-341
Rosengren solution 255 257 Jaumann flux 319 Loading functions 17
Implicit method 203 Kantorovich method 218 Loading surfaces 18
Incompatible approximating Kelvin's solution 259-260 Local coordinate system 118,
functions 119 Kelvin- Voigt chain 41 147-150,226
Incremental constitutive equa- Kinematic boundary condi- Localization 24, 27-31, 355-
tion 41 -43 tions 10, 47, 230 357; definition of 28; stabil-
Incremental formulation 355- Kirchoff theory of shallow ity aspects 29-3 1
356 arches 329 Localization limiters 35-37
419
Localization, stability aspects Nonlinear equations 335-346 Pore pressure and loading,
29 Nonlinear fracture mechanics history of 247
Localization-induced bifurca- 256-257 Porous materials 240-246
tion 30 Nonlinear response 298 Potential energy 132
Mass matrix 52, 168-169 Nonlinear systems 60-63 Potential energy density 57
Material nonlinearity 298 Nonlocal continuum 24 Probabilistic finite element
Material stiffness matrix. See Nonlocal (integral) localiza- method (PFEM) 380, 387-
Stiffness matrix tion limiters 35-36 395
Matrix formulation 408-409 Nonperiodical loading, struc- Proportional damping 209-
Maxwell chain 40-41 tural response 194-199, 210
Membrane elements, compar- 201-205 Pseudo-stress 305
ison of 128 Nonproportional damping Quadratic elements 366
Mesh arrangement 252 198-199,210-212 Quadratic interpolation 221
Method of fictitious loads Nonstatistical methods 380 Quadratic tetrahedron 161
259 Numerical integration 221- Quadrilateral, approximation
Mindlin-Reissner theory. See 222, 410-415 functions 101-104; map-
Mindlin theory of thick Optimal step-length 343 ping of 103
plates Orthogonality, loss of 191 Quadrilateral element 114-
Mindlin assumptions 97, 227; Orthogonality of eigenmodes 115; interpolation functions
verification of 97 172-173 125- 127; modified 115-119
Mindlin hypothesis 65-66, Orthogonality, reconstruction Quadrilateral plate element
138, 144 of 192 141-144; modified 144-147
Mindlin theory of bending Orthotropic materials 13-15 Quasi-Newton methods 344-
222 Orthotropy 1 1 346
Mindlin theory of thick plates Patch test 55, 120-121, 157- Random properties 381-383
129-135, 136,218 159 Rate boundary value problem
Minimum complementary Physical nonlinearity 358-359 59-60
energy, principle of 47 Piola-Kirchhoff stress tnesor Rayleigh quotient 173
Minimum potential energy 45 311-313,317,319 Rayleigh-Ritz method 175-
Mixed models in structural Plane elements, rotational 179, 185-187
analysis 49 degrees of freedom 119- Reaction and loading, history
Mode decomposition 194- 127 of 247
199,209-212 Plane problems, elements Rectangular cross section 237
Modified patch test 120 109- 128; examples of 110 Relative damping, weighted
Modified quadrilateral ele- Plane strain 11-12, 13, 15,23, coefficients of 198
ment 115-119 110 Resistance curve 257
Modified quadrilateral plate Plane stress 12, 13, 15,23, Response spectrum 206-207
element 144-147 110,261-264 Rhombic anisotropy 1 1
Monte Carlo simulation 380, Plane strip 227-229 Rice Tau-integral 252, 253-
384-385 Plastic hardening modulus 2 1 254
Mooney-Rivlin material 298 Plastic zone, propagation of Rigid body motions 82
Motion of supports 196-197 256 Rigid body rotation 308-31 1
Multi-grid methods 378-379 Plasticity, effect of 254-257 Ritz method 78; compared to
Natural coordinates 100 Plate analysis 128, 288-297 finite element method 52-
Navier solution 218 Plate boundary, stress resul- 53; displacement version
Newmark method 203, 204 tants 132 51-53; force version 53-54
Newton-Raphson method Plate deformations, assump- Rotational degrees of free-
335-336, 345-346; extend- tions 129 dom 119-127, 160, 163-
ed 35 1-352 Plate elements 128-147 166, 227-229
Nonhomogeneous elasticity Poisson equations 230 Second law of thermodynam-
55-59 Poisson's ratio 12, 13 ics 61
Noninteracting foundation Polar decomposition theorem Sector of path independence
structures 152-153 309 63
420
Seismic effects 206-207 127; directional derivative Thin-walled beam, cross sec-
Shallow circular arch 348- 354-355; spectral decompo- tion 107
349 sition of 173-174 Thin- walled elements 107-
Shape functions and deriva- Stiffness moment, calculation 109
tives 78, 79, 81 of 236-238 Time dependence 258
Sharp crack 248 Stiffness tensor 15 Timoshenko beam solution
Shear distortion 65, 73 Strain field 256 144
Shear forces, beams 90, 157 Strain localization 27-3 1 Torsion of bars 232-238
Shear locking 146 Strain softening 23-24, 26, 30 Total Lagrangian (TL) formu-
Shear strains 232 Strain tensor 14 lation 299, 302, 314-315,
Shell elements 147-151,334; Strain-displacement equations 317,318,319,327-328,
transformation into global 9 328-330
coordinates 151 Stress distribution 228-229, Transformation field analysis
Shell structures 336 277 277-282
Simply supported beam 216, Stress intensity factor 34, Tresca plasticity condition
228-229 248-251 363
Simulation 380 Stress relief zone 33 Triangle, approximation func-
Size effect 33, 34 Stress state 16 tions 106
Skeletal structures 64-99 Strip stiffness matrix 225 Triangle, area coordinates
Skewed cantilever plate 147 Structural model 245 104-105
Small-scale yielding 254 Structural response, nonperi- Triangular elements 15-16,
Snap-through 301, 336 odical loading 194-199, 93, 111-113, 136-138; basic
Softening 59-60, 355-357 201-205; harmonic excita- relations 104-106; types of
Softening media 27-31 tion 207-212 105
Soil-structure interaction 151 Structure and foundation, Triaxial stress 42-43
Soils, deformation of 240-246 interaction between 151- Triclinic system 1 1
Solid skeleton, deformation 157 Truesdale stress flux 316
of 241 -242 Subgrade below beam 7 1 Truss and beam elements 73-
Somigliana's formulae 259- Subgrade elements, basic 83
265, 273, 283, 286 functions for 155 Trusses 64
Spectral decomposition 173- Subspace iteration 185-187 Two-dimensional structures,
174 Subspace iteration 192 examples of 215
Spherical arc-length method Substructure 93 Uniaxial stress, constitutive
338-339 Superconvergent patch recov- equations 39-42
SSY. See Small scale yielding ery 371 Unit loading case 260 261
Stability, conditions of 60 6 1 Superposition principle 39-40 Unknown functions 225
Stability criteria 60-63 Surface displacements 55-57, Updated Lagrangian (UL)
Static analysis 288 278-280 formulation 299, 302, 315-
Static boundary conditions Surface tractions 57-59 317,318,323-324,326,
10, 47, 230 Tangent stiffness 301-302, 327-328
Static boundary condition 244 322, 330 Variational formulation 59-
Static condensation 88-94, Temperature 131 60, 243-246
174-175, 176- 179; and Tension-compression bar 100, Variational principles 43, 45-
Rayleigh-Ritz method, 106-107,319-328 48, 55-59; compared with
combined 176 Tensor notation 15 finite element method mod-
Static correction 200-201 Tetrahedra 160-161 els 50; generalized 123-
Statistical methods 380 Thermodynamics 61-62 125, 126; modified 48-50
Steel frame 89, 91 Thick plates 2 18-220 Vertical displacements 277
Stiff foundation beams 152 Thick- walled cylinder 166, Vector of internal forces 330
Stiffening wall 276 363-364 Vibration 168, 194-207
Stiffness approach 232-234 Thin plate analysis 288-297; Virtual displacements, princi-
Stiffness matrix 11, 13, 15, compatible elements 128 ple of 43-44, 45, 47, 312-
30, 50; computation of 125- Thin plate bending 2 18 317
421
Virtual forces, principle of
44, 45, 47, 83-84
Virtual work, principle of 43-
44
Viscoplastic materials 38-43
Volterra integral equation 40,
258
von Karman theory 310, 328
Walled structure 156
Warping 23 1-238
Weak formulation 355, 356-
357
Weak patch test 158- 159
Weighted residual method
258
Wheeler-Sternberg formulae
286
Wilson method 204-205
Winkler-Pasternak model 64,
67-72,86-88, 132,230-231
Yield conditions 17
Yield criterion 16-20
Yield function 16-20
Yield surfaces 18
Zienkiewicz-Zhu technique
369-373
422