1

Overview

1–1
Chapter 1: OVERVIEW

TABLE OF CONTENTS

Page
§1.1 Book Scope . . . . . . . . . . . . . . . . . . . . . 1–3
§1.2 Where the Material Fits . . . . . . . . . . . . . . . . 1–3
§1.2.1 Top Level Classification . . . . . . . . . . . . . . 1–3
§1.2.2 Computational Mechanics . . . . . . . . . . . . . 1–3
§1.2.3 Statics versus Dynamics . . . . . . . . . . . . . . 1–5
§1.2.4 Linear versus Nonlinear . . . . . . . . . . . . . 1–5
§1.2.5 Discretization Methods . . . . . . . . . . . . . . 1–5
§1.2.6 FEM Formulation Levels . . . . . . . . . . . . . 1–6
§1.2.7 FEM Choices . . . . . . . . . . . . . . . . . 1–7
§1.2.8 Finally: What The Book Is About . . . . . . . . . . 1–7
§1.3 What Does a Finite Element Look Like? . . . . . . . . . . . 1–7
§1.4 The FEM Analysis Process . . . . . . . . . . . . . . . 1–9
§1.4.1 The Physical FEM . . . . . . . . . . . . . . . . 1–9
§1.4.2 The Mathematical FEM . . . . . . . . . . . . . 1–11
§1.4.3 Synergy of Physical and Mathematical FEM . . . . . . . 1–11
§1.4.4 Streamlined Idealization and Discretization . . . . . . . 1–13
§1.5 Method Interpretations . . . . . . . . . . . . . . . . . 1–13
§1.5.1 Physical Interpretation . . . . . . . . . . . . . . 1–13
§1.5.2 Mathematical Interpretation . . . . . . . . . . . . 1–14
§1.6 Keeping the Course . . . . . . . . . . . . . . . . . . 1–15
§1.7 *What is Not Covered . . . . . . . . . . . . . . . . . 1–15
§1.8 The Origins of the Finite Element Method . . . . . . . . . . 1–16
§1.9 Recommended Books for Linear FEM . . . . . . . . . . . . 1–16
§1.9.1 Hasta la Vista, Fortran . . . . . . . . . . . . . . 1–16
§1. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 1–17
§1. References . . . . . . . . . . . . . . . . . . . . . . 1–18
§1. Exercises . . . . . . . . . . . . . . . . . . . . . . 1–19

1–2
 §1.2 WHERE THE MATERIAL FITS

§1.1. Book Scope

This is a textbook about linear structural analysis using the Finite Element Method (FEM) as a
discretization tool. It is intended to support an introductory course at the first-year level of graduate
studies in Aerospace, Mechanical, or Civil Engineering.
Basic prerequisites to understanding the material covered here are: (1) a working knowledge of
matrix algebra, and (2) an undergraduate structures course at the Materials of Mechanics level.
Helpful but not required are previous courses in continuum mechanics and advanced structures.
This Chapter presents an overview of what the book covers, and what finite elements are.

§1.2. Where the Material Fits

This Section outlines where the book material fits within the vast scope of Mechanics. In the
ensuing multilevel classification, topics addressed in some depth in this book are emphasized in
bold typeface.

§1.2.1. Top Level Classification

Definitions of Mechanics in dictionaries usually state two flavors:
• The branch of Physics that studies the effect of forces and energy on physical bodies.1
• The practical application of that science to the design, construction or operation of material
systems or devices, such as machines, vehicles or structures.
These flavors are science and engineering oriented, respectively. But dictionaries are notoriously
archaic. For our objectives it will be convenient to distinguish four flavors:


 Theoretical

Applied
Mechanics (1.1)

 Computational
Experimental

Theoretical mechanics deals with fundamental laws and principles studied for their intrinsic sci-
entific value. Applied mechanics transfers this theoretical knowledge to scientific and engineering
applications, especially as regards the construction of mathematical models of physical phenomena.
Computational mechanics solves specific problems by model-based simulation through numerical
methods implemented on digital computers. Experimental mechanics subjects the knowledge de-
rived from theory, application and simulation to the ultimate test of observation.

Remark 1.1. Paraphrasing an old joke about mathematicians, one may define a computational mechanician
as a person who searches for solutions to given problems, an applied mechanician as a person who searches
for problems that fit given solutions, and a theoretical mechanician as a person who can prove the existence
of problems and solutions. As regards experimentalists, make up your own joke.

1 Here the term “bodies” includes all forms of matter, whether solid, liquid or gaseous; as well as all physical scales, from
subatomic through cosmic.

1–3
Chapter 1: OVERVIEW

COMPUTATIONAL MECHANICS

Finite Theoretical
Element and Applied
Methods Mechanics

Applied
Computer &
Mathematics
Information
& Numerical
Sciences
Analysis

Figure 1.1. The “pizza slide:” Computational Mechanics

integrates aspects of four disciplines.

§1.2.2. Computational Mechanics

Computational Mechanics represents the integration of several disciplines, as depicted in the “pizza
slice” Figure 1.1 Several branches of computational mechanics can be distinguished according to
the physical scale of the focus of attention:



Nanomechanics




 Micromechanics  Solids and Structures
Computational Mechanics (1.2)

 Continuum mechanics Fluids



 Multiphysics

Systems

Nanomechanics deals with phenomena at the molecular and atomic levels. As such, it is closely
related to particle physics and chemistry. At the atomic scale it transitions to quantum mechanics.
Micromechanics looks primarily at the crystallographic and granular levels of matter. Its main
technological application is the design and fabrication of materials and microdevices.
Continuum mechanics studies bodies at the macroscopic level, using continuum models in which the
microstructure is homogenized by phenomenological averaging. The two traditional areas of appli-
cation are solid and fluid mechanics. Structural mechanics is a conjoint branch of solid mechanics,
since structures, for obvious reasons, are fabricated with solids. Computational solid mechan-
ics favors a applied-sciences approach, whereas computational structural mechanics emphasizes
technological applications to the analysis and design of structures.
Computational fluid mechanics deals with problems that involve the equilibrium and motion of
liquid and gases. Well developed related subareas are hydrodynamics, aerodynamics, atmospheric
physics, propulsion, and combustion.
Multiphysics is a more recent newcomer.2 This area is meant to include mechanical systems that
transcend the classical boundaries of solid and fluid mechanics. A key example is interaction

2 This unifying term is in fact missing from most dictionaries, as it was introduced by computational mechanicians in the
1970s. Several multiphysics problems, however, are older. For example, aircraft aeroelasticity emerged in the 1920s.

1–4
 §1.2 WHERE THE MATERIAL FITS

between fluids and structures, which has important application subareas such as aeroelasticity and
hydroelasticity. Phase change problems such as ice melting and metal solidification fit into this
category, as do the interaction of control, mechanical and electromagnetic systems.
Finally, system identifies mechanical objects, whether natural or artificial, that perform a distin-
guishable function. Examples of man-made systems are airplanes, building, bridges, engines, cars,
microchips, radio telescopes, robots, roller skates and garden sprinklers. Biological systems, such
as a whale, amoeba, virus or pine tree are included if studied from the viewpoint of biomechanics.
Ecological, astronomical and cosmological entities also form systems.3
In the progression of (1.2), system is the most general concept. Systems are studied by decompo-
sition: its behavior is that of its components plus the interaction between the components. Com-
ponents are broken down into subcomponents and so on. As this hierarchical process continues
the individual components become simple enough to be treated by individual disciplines, but their
interactions may get more complex. Thus there are tradeoff skills in deciding where to stop.4

§1.2.3. Statics versus Dynamics

Continuum mechanics problems may be subdivided according to whether inertial effects are taken
into account or not:
 
 Time Invariant
Statics
Continuum mechanics Quasi-static (1.3)

Dynamics

In statics inertial forces are ignored or neglected. These problems may be subclassified into time
invariant and quasi-static. For the former time need not be considered explicitly; any time-like
response-ordering parameter (should one be needed) will do. In quasi-static problems such as
foundation settlements, creep flow, rate-dependent plasticity or fatigue cycling, a more realistic
estimation of time is required but inertial forces are ignored as long as motions remain slow.
In dynamics the time dependence is explicitly considered because the calculation of inertial (and/or
damping) forces requires derivatives respect to actual time to be taken.

§1.2.4. Linear versus Nonlinear

A classification of static problems that is particularly relevant to this book is

Linear
Statics (1.4)
Nonlinear
Linear static analysis deals with static problems in which the response is linear in the cause-and-
effect sense. For example: if the applied forces are doubled, the displacements and internal stresses
also double. Problems outside this domain are classified as nonlinear.
3 Except that their function may not be clear to us. “What is it that breathes fire into the equations and makes a universe
for them to describe? The usual approach of science of constructing a mathematical model cannot answer the questions
of why there should be a universe for the model to describe. Why does the universe go to all the bother of existing?”
(Stephen Hawking).
4 Thus in breaking down a car engine, say, the decomposition does not usually proceed beyond the components that may
be bought at a automotive shop.

1–5
Chapter 1: OVERVIEW

§1.2.5. Discretization Methods

A final classification of computational solid and structural mechanics (CSSM) for static analysis is
based on the discretization method by which the continuum mathematical model is discretized in
space, i.e., converted to a discrete model of finite number of degrees of freedom:



Finite Element Method (FEM)




Boundary Element Method (BEM)
Finite Difference Method (FDM)
CSSM spatial discretization (1.5)

 Finite Volume Method (FVM)



 Spectral Method
Mesh-Free Method

For linear problems finite element methods currently dominate the scene, with boundary element
methods posting a strong second choice in selected application areas. For nonlinear problems the
dominance of finite element methods is overwhelming.
Classical finite difference methods in solid and structural mechanics have virtually disappeared
from practical use. This statement is not true, however, for fluid mechanics, where finite dif-
ference discretization methods are still important although their dominance has diminished over
time. Finite-volume methods, which focus on the direct discretization of conservation laws, are
favored in highly nonlinear problems of fluid mechanics. Spectral methods are based on global
transformations, based on eigendecomposition of the governing equations, that map the physical
computational domain to transform spaces where the problem can be efficiently solved.
A recent newcomer to the scene are the mesh-free methods. These are finite different methods on
arbitrary grids constructed using a subset of finite element techniques

§1.2.6. FEM Formulation Levels

The term Finite Element Method actually identifies a broad spectrum of techniques that share
common features. Since its emergence in the framework of the Direct Stiffness Method (DSM)
over 1956–1964, [765,768] FEM has expanded like a tsunami, surging from its origins in aerospace
structures to cover a wide range of nonstructural applications, notably thermomechanics, fluid
dynaamics, and electromagnetics. The continuously expanding range makes taxology difficult.
Restricting ourselves to applications in computational solid and structural mechanics (CSSM), one
classificaation of particular relevance to this book is


 Mechanics of Materials (MoM) Formulation

Conventional Variational Formulation
FEM-CSSM Formulation Level

 Advanced Variational Formulation
Template Formulation
(1.6)
The MoM formulation is applicable to simple structural elements such as bars and beams, and does
not require any knowledge of variational methods. This level is accesible to undergraduate students,
as only require some elementary knowledge of linear algebra and makes no use of variational
calculus. The second level is characterized by two features: the use of standard work and energy

1–6
 §1.3 WHAT DOES A FINITE ELEMENT LOOK LIKE?

methods (such as the Total Potential Energy principle), and focus on full compliance with the
requirements of the classical Ritz-Galerkin direct variational methods (for example, interelement
continuity). It is approriate for first year (master level) graduate students with basic exposure to
variational methods. The two lower levels were well established by 1970, with no major changes
since, and are those used in the present book.
The next two levels are covered in the Advanced Finite Element Methods book [255]. The third
one requires a deeper exposure to variational methods in mechanics, notably multifield and hybrid
principles. The last level (templates) is the pinnacle “where the rivers of our wisdom flow into
one another.” Reaching it requires both mastery of advanced variational principles, as well as the
confidence and fortitude to discard them along the way to the top.

§1.2.7. FEM Choices

A more down to earth classification considers two key selection attributes: Primary Unknown
Variable(s), or PUV, and solution method:5


 Displacement (a.k.a. Primal) 
 Stiffness
Force (a.k.a. Dual or Equilibrium)
PUV Choice Solution Choice Flexibility (1.7)

 Mixed (a.k.a. Primal-Dual) Combined
Hybrid

Here PUV Choice governs the variational framework chosen to develop the discrete equations; if
one works at the two middle levels of (1.6). It is possible, however, to develop those completely
outside a variational framework, as noted there. The solution choice is normally dictated by the
PUV, but exceptions are possible.

§1.2.8. Finally: What The Book Is About

Using the classification of (1.1) through (1.5) we can now state the book topic more precisely:

The model-based simulation of linear static structures discretized by FEM,

formulated at the two lowest levels of (1.6). (1.8)

Of the FEM variants listed in (1.7) emphasis will be placed on the displacement PUV choice and
stiffness solution, just like in [257]. This particular combination is called the Direct Stiffness Method
or DSM.

5 The alternative PUV terms: primal, dual or primal-dual, are those used in FEM non-structural applications, as well as in
more general computational methods.

1–7
Chapter 1: OVERVIEW

(a) (b) 3 (c) (d)

4 2 4

r 2r sin(π/n)
d 5 1 5 i j
2π/n
r
6 8
7

Figure 1.2. The “find π” problem treated with FEM concepts: (a) continuum object, (b) a discrete
approximation by inscribed regular polygons, (c) disconnected element, (d) generic element.

§1.3. What Does a Finite Element Look Like?

The subject of this book is FEM. But what is a finite element? As discussed later, the term admits
of two interpretations: physical and mathematical. For now the underlying concept will be partly
illustrated through a truly ancient problem: find the perimeter L of a circle of diameter d. Since
L = π d, this is equivalent to obtaining a numerical value for π.
Draw a circle of radius r and diameter d = 2r as in Figure 1.2(a). Inscribe a regular polygon of
n sides, where n = 8 in Figure 1.2(b). Rename polygon sides as elements and vertices as nodes.
Label nodes with integers 1, . . . 8. Extract a typical element, say that joining nodes 4–5, as shown in
Figure 1.2(c). This is an instance of the generic element i– j pictured in Figure 1.2(d). The element
length is L i j = 2r sin(π/n). Since all elements have the same length, the polygon perimeter is
L n = n L i j , whence the approximation to π is πn = L n /d = n sin(π/n).

Table 1.1. Rectification of Circle by Inscribed Polygons (“Archimedes FEM”)

n πn = n sin(π/n) Extrapolated by Wynn- Exact π to 16 places

1 0.000000000000000
2 2.000000000000000
4 2.828427124746190 3.414213562373096
8 3.061467458920718
16 3.121445152258052 3.141418327933211
32 3.136548490545939
64 3.140331156954753 3.141592658918053
128 3.141277250932773
256 3.141513801144301 3.141592653589786 3.141592653589793

Values of πn obtained for n = 1, 2, 4, . . . 256 and r = 1 are listed in the second column of
Table 1.1. As can be seen the convergence to π is fairly slow. However, the sequence can be
transformed by Wynn’s  algorithm6 into that shown in the third column. The last value displays
15-place accuracy.

6 A widely used lozenge extrapolation algorithm that speeds up the convergence of many sequences. See, e.g, [812].

1–8
 §1.4 THE FEM ANALYSIS PROCESS

Some key ideas behind the FEM can be identified in this example. The circle, viewed as a source
mathematical object, is replaced by polygons. These are discrete approximations to the circle.
The sides, renamed as elements, are specified by their end nodes. Elements can be separated by
disconnecting nodes, a process called disassembly in the FEM. Upon disassembly a generic element
can be defined, independently of the original circle, by the segment that connects two nodes i
and j. The relevant element property: side length L i j , can be computed in the generic element
independently of the others, a property called local support in the FEM. The target property: polygon
perimeter, is obtained by reconnecting n elements and adding up their length; the corresponding
steps in the FEM being assembly and solution, respectively. There is of course nothing magic about
the circle; the same technique can be be used to rectify any smooth plane curve.7
This example has been offered in the FEM literature, e.g. in [476], to aduce that finite element ideas
can be traced to Egyptian mathematicians from circa 1800 B.C., as well as Archimedes’ famous
studies on circle rectification by 250 B.C. But comparison with the modern FEM, as covered in
following Chapters, shows this to be a stretch. The example does not illustrate the concept of degrees
of freedom, conjugate quantities and local-global coordinates. It is guilty of circular reasoning: the
compact formula π = limn→∞ n sin(π/n) uses the unknown π in the right hand side.8 Reasonable
people would argue that a circle is a simpler object than, say, a 128-sided polygon. Despite these
flaws the example is useful in one respect: showing a fielder’s choice in the replacement of one
mathematical object by another. This is at the root of the simulation process described next.

§1.4. The FEM Analysis Process

Processes that use FEM involve carrying out a sequence of steps in some way. Those sequences
take two canonical configurations, depending on (i) the environment in which FEM is used and (ii)
the main objective: model-based simulation of physical systems, or numerical approximation to
mathematical problems. Both are reviewed below to introduce terminology used in the sequel.

§1.4.1. The Physical FEM

A canonical use of FEM is simulation of physical systems. This requires models of such systems.
Consequenty the methodology is often called model-based simulation.
The process is illustrated in Figure 1.3. The centerpiece is the physical system to be modeled.
Accordingly, this configuration is called the Physical FEM. The processes of idealization and
discretization are carried out concurrently to produce the discrete model. The solution step is
handled by an equation solver often customized to FEM, which delivers a discrete solution (or
solutions).
Figure 1.3 also shows an ideal mathematical model. This may be presented as a continuum limit or
“continuification” of the discrete model. For some physical systems, notably those well modeled
by continuum fields, this step is useful. For others, such as complex engineering systems (say, a
flying aircraft) it makes no sense. Indeed Physical FEM discretizations may be constructed and
adjusted without reference to mathematical models, simply from experimental measurements.
7 A similar limit process, however, may fail in three dimensions for evaluation of surface areas.
8
objection
The circularity

is bypassed if n is advanced as a power of two, as in Table 1.1, by using the half-angle recursion
√
2 sin α = 1 − 1 − sin2 2α, started from 2α = π for which sin π = −1.

1–9
Chapter 1: OVERVIEW

occasionally
Ideal relevant
Mathematical
model

CONTINUIFICATION
SOLUTION

Physical FEM Discrete Discrete

system model solution

IDEALIZATION & VERIFICATION

DISCRETIZATION
solution error
simulation error: modeling & solution error
VALIDATION

Figure 1.3. The Physical FEM. The physical system (left box) is the source
of the simulation process. The ideal mathematical model (should one go to the
trouble of constructing it) is inessential.

The concept of error arises in the Physical FEM in two ways. These are known as verification and
validation, respectively.9 Verification is done by replacing the discrete solution into the discrete
model to get the solution error. This error is not generally important. Substitution in the ideal
mathematical model in principle provides the discretization error. This step is rarely useful in
complex engineering systems, however, because there is no reason to expect that the continuum
model exists, and even if it does, that it is more physically relevant than the discrete model.
Validation tries to compare the discrete solution against observation by computing the simulation
error, which combines modeling and solution errors. As the latter is typically unimportant, the
simulation error in practice can be identified with the modeling error. In real-life applications this
error overwhelms the others.10
One way to adjust the discrete model so that it represents the physics better is called model updating.
The discrete model is given free parameters. These are determined by comparing the discrete
solution against experiments, as illustrated in Figure 1.4. Inasmuch as the minimization conditions
are generally nonlinear (even if the model is linear) the updating process is inherently iterative.

Physical Experimental FEM Parametrized Discrete

system discrete solution
database
model

EXPERIMENTS simulation error

Figure 1.4. Model updating process in the Physical FEM.

9 Programming analogs: static and dynamic testing are called verification and validation, respectively. Static testing is
carried at the source level (e.g., code walkthroughs, compilation) whereas dynamic testing is done by running the code.
10 “All models are wrong; some are useful” (George Box)

1–10
 §1.4 THE FEM ANALYSIS PROCESS

Mathematical Discretization & solution error

model

IDEALIZATION REALIZATION FEM VERIFICATION

SOLUTION

Ideal Discrete Discrete

physical model solution
system

IDEALIZATION &
DISCRETIZATION VERIFICATION
solution error
ocassionally relevant

Figure 1.5. The Physical FEM. The physical system (left box) is the source
of the simulation process. The ideal mathematical model (should one go to the
trouble of constructing it) is inessential.

§1.4.2. The Mathematical FEM

The other canonical way of using FEM focuses on the mathematics. The process steps are illustrated
in Figure 1.5. The spotlight now falls on the mathematical model. This is often an ordinary
differential equation (ODE), or a partial differential equation (PDE) in space and time. A discrete
finite element model is generated from a variational or weak form of the mathematical model.11
This is the discretization step. The FEM equations are solved as described for the Physical FEM.
On the left, Figure 1.5 shows an ideal physical system. This may be presented as a realization of
the mathematical model. Conversely, the mathematical model is said to be an idealization of this
system. E.g., if the mathematical model is the Poisson’s PDE, realizations may be heat conduction
or an electrostatic charge-distribution problem. This step is inessential and may be left out. Indeed
Mathematical FEM discretizations may be constructed without any reference to physics.
The concept of error arises when the discrete solution is substituted in the “model” boxes. This
replacement is generically called verification. As in the Physical FEM, the solution error is the
amount by which the discrete solution fails to satisfy the discrete equations. This error is relatively
unimportant when using computers, and in particular direct linear equation solvers, for the solution
step. More relevant is the discretization error, which is the amount by which the discrete solution
fails to satisfy the mathematical model.12 Replacing into the ideal physical system would in principle
quantify modeling errors. In the Mathematical FEM this is largely irrelevant, however, because the
ideal physical system is merely that: a figment of the imagination.
§1.4.3. Synergy of Physical and Mathematical FEM
The foregoing canonical sequences are not exclusive but complementary. This synergy13 is one of
the reasons behind the power and acceptance of the method. Historically the Physical FEM was the
11 The distinction between strong, weak and variational forms is discussed in advanced FEM courses. In the present book
such forms will be largely stated (and used) as recipes.
12 This error can be computed in several ways, the details of which are of no importance here.
13 Such interplay is not exactly a new idea: “The men of experiment are like the ant, they only collect and use; the reasoners
resemble spiders, who make cobwebs out of their own substance. But the bee takes the middle course: it gathers its
material from the flowers of the garden and field, but transforms and digests it by a power of its own.” (Francis Bacon).

1–11
Chapter 1: OVERVIEW

NT
P ONE
atic
a l COMEVEL
hem L
Matmodel
ent
ponns
ry Com o
ibra a ti
L t equ
en
FEM Com
pon e
et TEM
SYS EL
s c r
di del
mo
LEV e
plet
Comution
so l
em
Systrete
disc del
l mo
sica
Phy tem
s y s

Figure 1.6. Combining physical and mathematical modeling through multilevel

FEM. Only two levels (system and component) are shown for simplicity.

first one to be developed to model complex physical systems such as aircraft, as narrated in §1.7.
The Mathematical FEM came later and, among other things, provided the necessary theoretical
underpinnings to extend FEM beyond structural analysis.
A glance at the schematics of a commercial jet aircraft makes obvious the reasons behind the Physical
FEM. There is no simple differential equation that captures, at a continuum mechanics level,14 the
structure, avionics, fuel, propulsion, cargo, and passengers eating dinner. There is no reason for
despair, however. The time honored divide and conquer strategy, coupled with abstraction, comes
to the rescue.
First, separate the structure out and view the rest as masses and forces. Second, consider the aircraft
structure as built up of substructures (a part of a structure devoted to a specific function): wings,
fuselage, stabilizers, engines, landing gears, and so on.
Take each substructure, and continue to break it down into components: rings, ribs, spars, cover
plates, actuators, etc. Continue through as many levels as necessary. Eventually those components
become sufficiently simple in geometry and connectivity that they can be reasonably well described
by the mathematical models provided, for instance, by Mechanics of Materials or the Theory of
Elasticity. At that point, stop. The component level discrete equations are obtained from a FEM
library based on the mathematical model.
The system model is obtained by going through the reverse process: from component equations
to substructure equations, and from those to the equations of the complete aircraft. This system
14 Of course at the (sub)atomic level quantum mechanics works for everything, from landing gears to passengers. But
it would be slightly impractical to represent the aircraft by, say, 1036 interacting particles modeled by the Schrödinger
equations. More seriously, Truesdell and Toupin correctly note that “Newtonian mechanics, while not appropriate to the
corpuscles making up a body, agrees with experience when applied to the body as a whole, except for certain phenomena
of astronomical scale” [759, p. 228].

1–12
 §1.5 METHOD INTERPRETATIONS

member
Physical System
support

joint

IDEALIZATION
Idealized and
Discrete System
;;
;

;;
;
Figure 1.7. The idealization process for a simple structure. The physical system — here
a conventional roof truss — is directly idealized by the mathematical model: a pin-jointed
bar assembly. For this particular structure idealized and discrete models coalesce.

assembly process is governed by the classical principles of Newtonian mechanics, which provide
the necessary inter-component “glue.” The multilevel decomposition process is diagramed in
Figure 1.6, in which intermediate levels are omitted for simplicity

Remark 1.2. More intermediate decomposition levels are used in systems such as offshore and ship structures,
which are characterized by a modular fabrication process. In that case multilevel decomposition mimics the
way the system is actually fabricated. The general technique, called superelements, is discussed in Chapter 10.

Remark 1.3. There is no point in practice in going beyond a certain component level while considering the
complete system. The reason is that the level of detail can become overwhelming without adding relevant
information. Usually that point is reached when uncertainty impedes further progress. Further refinement
of specific components is done by the so-called global-local analysis technique outlined in Chapter 10. This
technique is an instance of multiscale analysis.

§1.4.4. Streamlined Idealization and Discretization

For sufficiently simple structures, passing to a discrete model is carried out in a single idealization
and discretization step, as illustrated for the truss roof structure shown in Figure 1.7. Other levels
are unnecessary in such cases. Of course the truss may be viewed as a substructure of the roof, and
the roof as a a substructure of a building. If so the multilevel process would be more appropriate.

§1.5. Method Interpretations

Just like there are two complementary ways of using the FEM, there are two complementary
interpretations for explaining it, a choice that obviously impacts teaching. One interpretation
stresses the physical significance and is aligned with the Physical FEM. The other focuses on the
mathematical context, and is aligned with the Mathematical FEM. They are outlined next.

1–13
Chapter 1: OVERVIEW

§1.5.1. Physical Interpretation

The physical interpretation focuses on the flowchart of Figure 1.3. This interpretation has been
shaped by the discovery and extensive use of the method in the field of structural mechanics. The
historical connection is reflected in the use of structural terms such as “stiffness matrix”, “force
vector” and “degrees of freedom,” a terminology that carries over to non-structural applications.
The basic concept in the physical interpretation is the breakdown (≡ disassembly, tearing, partition,
separation, decomposition) of a complex mechanical system into simpler, disjoint components
called finite elements, or simply elements. The mechanical response of an element is characterized
in terms of a finite number of degrees of freedom. These degrees of freedoms are represented as
the values of the unknown functions as a set of node points. The element response is defined by
algebraic equations constructed from mathematical or experimental arguments. The response of
the original system is considered to be approximated by that of the discrete model constructed by
connecting or assembling the collection of all elements.
The breakdown-assembly concept occurs naturally when an engineer considers many artificial and
natural systems. For example, it is easy and natural to visualize an engine, bridge, aircraft or
skeleton as being fabricated from simpler parts.
As discussed in §1.4.3, the underlying theme is divide and conquer. If the behavior of a system
is too complex, the recipe is to divide it into more manageable subsystems. If these subsystems
are still too complex the subdivision process is continued until the behavior of each subsystem is
simple enough to fit a mathematical model that represents well the knowledge level the analyst
is interested in. In the finite element method such “primitive pieces” are called elements. The
behavior of the total system is that of the individual elements plus their interaction. A key factor in
the initial acceptance of the FEM was that the element interaction could be physically interpreted
and understood in terms that were eminently familiar to structural engineers.

§1.5.2. Mathematical Interpretation

This interpretation is closely aligned with the flowchart of Figure 1.5. The FEM is viewed as
a procedure for obtaining numerical approximations to the solution of boundary value problems
(BVPs) posed over a domain
. This domain is replaced by the union ∪ of disjoint subdomains

(e) called finite elements. In general the geometry of

is only approximated by that of ∪
(e) .
The unknown function (or functions) is locally approximated over each element by an interpolation
formula expressed in terms of values taken by the function(s), and possibly their derivatives, at a
set of node points generally located on the element boundaries. The states of the assumed unknown
function(s) determined by unit node values are called shape functions. The union of shape functions
“patched” over adjacent elements form a trial function basis for which the node values represent the
generalized coordinates. The trial function space may be inserted into the governing equations and
the unknown node values determined by the Ritz method (if the solution extremizes a variational
principle) or by the Galerkin, least-squares or other weighted-residual minimization methods if the
problem cannot be expressed in a standard variational form.

Remark 1.4. In the mathematical interpretation the emphasis is on the concept of local (piecewise) approx-
imation. The concept of element-by-element breakdown and assembly, while convenient in the computer
implementation, is not theoretically necessary. The mathematical interpretation permits a general approach

1–14
 §1.7 *WHAT IS NOT COVERED

to the questions of convergence, error bounds, trial and shape function requirements, etc., which the physical
approach leaves unanswered. It also facilitates the application of FEM to classes of problems that are not so
readily amenable to physical visualization as structures; for example electromagnetics and heat conduction.

Remark 1.5. It is interesting to note some similarities in the development of Heaviside’s operational meth-
ods, Dirac’s delta-function calculus, and the FEM. These three methods appeared as ad-hoc computational
devices created by engineers and physicists to deal with problems posed by new science and technology
(electricity, quantum mechanics, and delta-wing aircraft, respectively) with little help from the mathematical
establishment.15 Only some time after the success of the new techniques became apparent were new branches
of mathematics (operational calculus, distribution theory and piecewise-approximation theory, respectively)
constructed to justify that success. In the case of the finite element method, the development of a formal
mathematical theory started in the late 1960s, and much of it is still in the making.

§1.6. Keeping the Course

The first Part of this book, covered in Chapters 2 through 10, stresses the physical interpretation
of FEM within the framework of the Direct Stiffness Method (DSM). This is done on account of
its instructional advantages. Furthermore the computer implementation becomes more transparent
because the sequence of operations can be placed in close correspondence with the DSM steps.
Chapters 11 through 19 deal specifically with element formulations. Ingredients of the mathematical
interpretation are called upon whenever it is felt proper and convenient to do so. Nonetheless
excessive entanglement with the mathematical theory is avoided if it may obfuscate the physics.
In Chapters 2 and 3 the time is frozen at about 1965, and the DSM presented as an aerospace
engineer of that time would have understood it. This is not done for sentimental reasons, although
that happens to be the year in which the writer began thesis work on FEM under Ray Clough.
Virtually all FEM commercial codes are now based on the DSM and the computer implementation
has not essentially changed since the late 1960s.16 What has greatly improved since is “marketing
sugar”: user interaction and visualization.

§1.7. *What is Not Covered

The following topics are not covered in this book:
1. Elements based on equilibrium, mixed and hybrid variational formulations.
2. Flexibility and mixed solution methods.
3. Plate and shell elements.
4. Variational methods in mechanics.
5. General mathematical theory of finite elements.
6. Buckling and stability analysis.
7. General nonlinear response analysis.
8. Structural optimization.

15 Oliver Heaviside took heavy criticism from the lotus eaters, which he returned with gusto. His legacy is a living proof
that “England is the paradise of individuality, eccentricity, heresy, anomalies, hobbies and humors” (George Santayana).
Paul Dirac was luckier: he was shielded as member of the physics establishment and eventually received a Nobel Prize.
Gilbert Strang, the first mathematician to dwelve in the real FEM (the one created by engineers) was kind to the founders.
16 With the gradual disappearance of Fortran as a “live” programming language, noted in §1.7.7, changes at the implemen-
tation level have recently accelerated. E.g., C++, Python, Java and Matlab “wrappers” are becoming more common.

1–15
Chapter 1: OVERVIEW

9. Error estimates and problem-adaptive discretizations.

10. Non-structural and multiphysics applications of FEM.
11. Designing and building production-level FEM software and use of special hardware (e.g. vector and
parallel computers)
Topics 1–5 belong to what may be called “Advanced Linear FEM”, which is covered in the book [255].
Topics 6–7 pertain to “Nonlinear FEM”, which is covered in the book [258]. Topics 8–10 fall into advanced
applications, covrede in other books in preparation, whereas 11 is an interdisciplinary topic that interweaves
with computer science.

§1.8. The Origins of the Finite Element Method

This section moved to Appendix O to facilitate further expansion.

§1.9. Recommended Books for Linear FEM

The literature is voluminous: over 200 textbooks and monographs have appeared since 1967. Some
recommendations for readers interested in further studies within linear FEM are offered below.
Basic level (reference): Zienkiewicz and Taylor [837]. This two-volume set is a comprehensive
upgrade of the previous edition [835]. Primarily an encyclopœdic reference work that gives a
panoramic coverage of FEM applications, as well as a comprehensive list of references. Not a
textbook or monograph. Prior editions suffered from loose mathematics, largely fixed in this one.
A three-volume fifth edition has appeared recently.
Basic level (textbook): Cook, Malkus and Plesha [149]. The third edition is comprehensive in
scope although the coverage is more superficial than Zienkiewicz and Taylor. A fourth edition has
appeared recently.
Intermediate level: Hughes [389]. It requires substantial mathematical expertise on the part of the
reader. Recently (2000) reprinted as Dover edition.
Mathematically oriented: Strang and Fix [705]. Still the most readable mathematical treatment for
engineers, although outdated in several subjects. Out of print.
Best value for the $$$: Przemieniecki’s Dover edition [603], list price $15.95 (2003). A reprint of a
1966 McGraw-Hill book. Although woefully outdated in many respects (the word “finite element”
does not appear except in post-1960 references), it is a valuable reference for programming simple
elements. Contains a fairly detailed coverage of substructuring, a practical topic missing from the
other books. Comprehensive bibliography in Matrix Structural Analysis up to 1966.
Most fun (if you appreciate British “humor”): Irons and Ahmad [401]. Out of print.
For buying out-of-print books through web services, check the metasearch engine in www3.addall.com
(most comprehensive; not a bookseller) as well as that of www.amazon.com. A newcomer is
www.campusi.com

§1.9.1. Hasta la Vista, Fortran

Most FEM books that include programming samples or even complete programs use Fortran. Those
face an uncertain future. Since the mid-1990s, Fortran is gradually disappearing as a programming
language taught in USA engineering undergraduate programs. (It still survives in some Physics and

1–16
 §1. Notes and Bibliography

Chemistry departments because of large amounts of legacy code.) So one end of the pipeline is
drying up. Low-level scientific programming17 is moving to C and C++, mid-level to Java, Perl and
Python, high-level to Matlab, Mathematica and their free-source Linux equivalents. How attractive
can a book teaching in a dead language be?
To support this argument with some numbers, here is a September-2003 snapshot of ongoing open
source software projects listed in http://freshmeat.net. This conveys the relative importance
of various languages (a mixed bag of newcomers, going-strongs, have-beens and never-was) in the
present environment.
Lang Projects Perc Lang Projects Perc Lang Projects Perc
Ada 38 0.20% APL 3 0.02% ASP 25 0.13%
Assembly 170 0.89% Awk 40 0.21% Basic 15 0.08%
C 5447 28.55% C# 41 0.21% C++ 2443 12.80%
Cold Fusion 10 0.05% Common Lisp 27 0.14% Delphi 49 0.26%
Dylan 2 0.01% Eiffel 20 0.10% Emacs-Lisp 33 0.17%
Erlang 11 0.06% Euler 1 0.01% Euphoria 2 0.01%
Forth 15 0.08% Fortran 45 0.24% Haskell 28 0.15%
Java 2332 12.22% JavaScript 236 1.24% Lisp 64 0.34%
Logo 2 0.01% ML 26 0.14% Modula 7 0.04%
Object Pascal 9 0.05% Objective C 131 0.69% Ocaml 20 0.10%
Other 160 0.84% Other Scripting Engines 82 0.43%
Pascal 38 0.20% Perl 2752 14.42% PHP 2020 10.59%
Pike 3 0.02% PL/SQL 58 0.30% Pliant 1 0.01%
PROGRESS 2 0.01% Prolog 8 0.04% Python 1171 6.14%
Rexx 7 0.04% Ruby 127 0.67% Scheme 76 0.40%
Simula 1 0.01% Smalltalk 20 0.10% SQL 294 1.54%
Tcl 356 1.87% Unix Shell 550 2.88% Vis Basic 15 0.08%
Xbasic 1 0.01% YACC 11 0.06% Zope 34 0.18%
Total Projects: 19079

Notes and Bibliography

Here is Ray Clough’s personal account of how FEM and DSM emerged at Boeing in the early 1950s. (For
further historical details, the interested reader may consult Appendices H and O.)
“ My involvement with the FEM began when I was employed by the Boeing Airplane Company in Seattle during
summer 1952 as a member of their summer faculty program. When I had joined the civil engineering faculty at
Berkeley in 1949, I decided to take advantage of my MIT structural dynamics background by taking up the field
of Earthquake Engineering. So because the Boeing summer faculty program offered positions with their structural
dynamics unit, I seized on that as the best means of advancing my preparation for the earthquake engineering field.
I was particularly fortunate in this choice of summer work at Boeing because the head of their structural dynamics
unit was Mr. M. J. Turner — a very capable man in dealing with problems of structural vibrations and flutter.
When I arrived for the summer of 1952, Jon Turner asked me to work on the vibration analysis of a delta wing
structure. Because of its triangular plan form, this problem could not be solved by procedures based on standard
beam theory; so I spent the summer of 1952 trying to formulate a delta wing model built up as an assemblage of
one-dimensional beams and struts. However, the results of deflection analyses based on this type of mathematical
model were in very poor agreement with data obtained from laboratory tests of a scale model of a delta wing. My
final conclusion was that my summer’s work was a total failure—however, at least I learned what did not work.

17 “A programming language is low level when its programs require attention to the irrelevant” (Alan Perlis).

1–17
Chapter 1: OVERVIEW

Spurred by this disappointment, I decided to return to Boeing for the summer faculty program in 1953. During
the winter, I stayed in touch with Jon Turner so I was able to rejoin the structural dynamics unit in June. The most
important development during the winter was that Jon suggested we try to formulate the stiffness property of the
wing by assembling plane stress plates of either triangular or rectangular shapes. So I developed stiffness matrices
for plates of both shapes, but I decided the triangular form was much more useful because such plates could be
assembled to approximate structures of any configuration. Moreover, the stiffness properties of the individual
triangular plates could be calculated easily based on assumptions of uniform states of normal stress in the X and
the Y directions combined with an uniform state of shear stress. Then the stiffness of the complete structure was
obtained by appropriate addition of the contributions from the individual pieces. The Boeing group called this
procedure the direct stiffness method.
The remainder of the summer of 1953 was spent in demonstrating that deflections calculated for structures formed
as assemblages of triangular elements agreed well with laboratory measurements on the actual physical models.
Also, it became apparent that the precision of the calculated results could be improved asymptotically by continued
refinement of the finite element mesh. The conclusions drawn from that summer’s work were presented in a paper
given by Jon Turner at the annual meeting of the Institute of Aeronautical Sciences in January 1954. However, for
reasons I never understood Jon did not submit the paper for publication until many months later. So this paper,
which often is considered to be the first published description of the FEM, was not published until September 1956
— more than two years after the verbal presentation.
It is important to note that the basic purpose of the work done by Jon Turner’s structural dynamics unit was vibration
and flutter analysis. They were not concerned with stress analysis because that was the responsibility of the stress
analysis unit. However, it was apparent that the model formed by the direct stiffness method could be used for
stress analysis as well as for vibration analysis, and I made plans to investigate this stress analysis application as
soon as possible. However, because of my other research responsibilities, I was not able to spend any significant
time on the stress analysis question until I went on my sabbatical leave to Trondheim, Norway in September 1956.
Then, when I arrived in Norway all I could do was to outline the procedures for carrying out the analysis, and to
do calculations for very small systems using a desk calculator because the Norwegian Institute of Technology did
not yet have an automatic digital computer.
The presentation of the paper to the Institute of Aeronautical Sciences was the first introduction of the principles
of the FEM to a technical audience; although some of the basic concepts of the method were stated a short time
later in a series of articles published in Aircraft Engineering by Dr. John H. Argyris during October 1954 to May
1955. However, the rectangular element presented in those articles is only a minor part of that contribution. The
Argyris work came to my attention during my sabbatical leave in Norway, and I considered it then (as I still do
now) to be the most important series of papers ever published in the field of Structural Mechanics. I credit that
work for extending the scope of my understanding of structural theory to the level it eventually attained.
From my personal point of view, the next important event in finite element history was the coining of the name
FEM. My purpose in choosing that name was to distinguish clearly the relatively large size pieces of the structure
that make up a finite element assemblage as contrasted with the infinitesimal contributions that go into evaluation of
the displacements of a structure in a typical virtual work analysis. The name first appeared in a publication that was
written to demonstrate the finite element procedure for the civil engineering profession. A much more significant
application of the method was presented at the Symposium on the use of Computers in Civil Engineering, held in
Lisbon, Portugal in 1962, where it was used to evaluate the stress concentrations developed in a gravity dam that
had cracked at its mid-section.”

References
Referenced items have been moved to Appendix R.

1–18
 Exercises

Homework Exercises for Chapter 1

Overview

EXERCISE 1.1 [A:15] Work out Archimedes’ problem using a circumscribed regular polygon, with n =
1, 2, 4, . . . 256. Does the sequence converge any faster?

EXERCISE 1.2 [D:20] Select one of the following vehicles: truck, car, motorcycle, or bicycle. Draw a two
level decomposition of the structure into substructures, and of selected components of some substructures.

EXERCISE 1.3 [D:30] In one of the earliest articles on the FEM, Clough [139] writes:
“When idealized as an assemblage of appropriately shaped two- and three-dimensional elements in this manner,
an elastic continuum can be analyzed by standard methods of structural analysis. It should be noted that the
approximation which is employed in this case is of physical nature; a modified structural system is substituted
for the actual continuum. There need be no approximation in the mathematical analysis of this structural
system. This feature distinguishes the finite element technique from finite difference methods, in which the
exact equations of the actual physical system are solved by approximate mathematical procedures.”
Discuss critically the contents of this paragraph while placing it in the context of time of writing (early 1960s).
Is the last sentence accurate?

1–19
 2
The Direct
Stiffness Method I

2–1
Chapter 2: THE DIRECT STIFFNESS METHOD I

TABLE OF CONTENTS

Page
§2.1 Foreword . . . . . . . . . . . . . . . . . . . . . . 2–3
§2.2 Why A Plane Truss? . . . . . . . . . . . . . . . . . 2–3
§2.3 Truss Structures . . . . . . . . . . . . . . . . . . . . 2–3
§2.4 Idealization . . . . . . . . . . . . . . . . . . . . . 2–4
§2.5 The Example Truss . . . . . . . . . . . . . . . . . . . 2–5
§2.6 Members, Joints, Forces and Displacements . . . . . . . . . 2–6
§2.7 The Master Stiffness Equations . . . . . . . . . . . . . . 2–7
§2.8 The DSM Steps . . . . . . . . . . . . . . . . . . . 2–8
§2.9 Breakdown Stage . . . . . . . . . . . . . . . . . . . 2–10
§2.9.1 Disconnection . . . . . . . . . . . . . . . . 2–10
§2.9.2 Localization . . . . . . . . . . . . . . . . . . 2–11
§2.9.3 Member Stiffness Equations . . . . . . . . . . . . 2–11
§2.10 Assembly and Solution Stage: Globalization . . . . . . . . . . 2–12
§2.10.1 Displacement and Force Transformations . . . . . . . . 2–12
§2.10.2 Global Member Stiffness Equations . . . . . . . . . . 2–14
§2. Notes and Bibliography . . . . . . . . . . . . . . . . . 2–15
§2. References . . . . . . . . . . . . . . . . . . . . . . 2–15
§2. Exercises . . . . . . . . . . . . . . . . . . . . . . 2–16

2–2
 §2.3 TRUSS STRUCTURES

§2.1. Foreword

This Chapter begins the exposition of the Direct Stiffness Method (DSM) of structural analysis.
The DSM is by far the most common implementation of the Finite Element Method (FEM). In
particular, all major commercial FEM codes are based on the DSM.
The exposition is done by following the DSM steps applied to a simple plane truss structure. The
method has two major stages: breakdown, and assembly+solution. This Chapter covers primarily
the breakdown stage.

§2.2. Why A Plane Truss?

The simplest structural finite element is the two-node bar (also called linear spring) element, which
is illustrated in Figure 2.1(a). A six-node triangle that models thin plates, shown in Figure 2.1(b)
displays intermediate complexity. Perhaps the most geometrically complex finite element (at least
as regards number of degrees of freedom) is the curved, three-dimensional, 64-node “brick” element
depicted in Figure 2.1(c).
Yet the remarkable fact is that, in the DSM, all elements, regardless of complexity, are treated alike!
To illustrate the basic steps of this democratic method, it makes educational sense to keep it simple
and use a structure composed of bar elements.

(a) (b) (c)

Figure 2.1. From the simplest through progressively more complex structural finite elements:
(a) two-node bar element for trusses, (b) six-node triangle for thin plates, (b) 64-node tricubic,
“brick” element for three-dimensional solid analysis.

A simple yet nontrivial structure is the pin-jointed plane truss, whose members may be modeled as
two-node bars.1 Using a plane truss to teach the stiffness method offers two additional advantages:

(a) Computations can be entirely done by hand as long as the structure contains just a few elements.
This allows various steps of the solution procedure to be carefully examined and understood
(learning by doing) before passing to the computer implementation. Doing hand computations
on more complex finite element systems rapidly becomes impossible.
(b) The computer implementation on any programming language is relatively simple and can be
assigned as preparatory computer homework before reaching Part III.

1 A one dimensional bar assembly would be even simpler. That kind of structure would not adequately illustrate some of
the DSM steps, however, notably the back-and-forth transformations from global to local coordinates.

2–3
Chapter 2: THE DIRECT STIFFNESS METHOD I

member

support

joint

Figure 2.2. An actual plane truss structure. That shown is typical of a roof
truss used in building construction for rather wide spans, say, over 10 meters. For
shorter spans, as in residential buildings, trusses are simpler, with fewer bays.

§2.3. Truss Structures

Plane trusses, such as the one depicted in Figure 2.2, are often used in construction, particularly
for roofing of residential and commercial buildings, and in short-span bridges. Trusses, whether
two or three dimensional, belong to the class of skeletal structures. These structures consist of
elongated structural components called members, connected at joints. Another important subclass
of skeletal structures are frame structures or frameworks, which are common in reinforced concrete
construction of buildings and bridges.
Skeletal structures can be analyzed by a variety of hand-oriented methods of structural analysis
taught in beginning Mechanics of Materials courses: the Displacement and Force methods. They
can also be analyzed by the computer-oriented FEM. That versatility makes those structures a good
choice to illustrate the transition from the hand-calculation methods taught in undergraduate courses,
to the fully automated finite element analysis procedures available in commercial programs.

§2.4. Idealization

The first analysis step carried out by a structural engineer is to replace the actual physical structure
by a mathematical model. This model represents an idealization of the actual structure. For truss
structures, by far the most common idealization is the pin-jointed truss, which directly maps to a
FEM model. See Figure 2.3.
The replacement of true by idealized is at the core of the physical interpretation of the finite
element method discussed in Chapter 1. The axially-carrying-load members and frictionless pins
of the pin-jointed truss are only an approximation of the physical one. For example, building and
bridge trusses usually have members joined to each other through the use of gusset plates, which
are attached by nails, bolts, rivets or welds. Consequently members will carry some bending as
well as direct axial loading.
Experience has shown, however, that stresses and deformations calculated using the simple idealized
model will often be satisfactory for preliminary design purposes; for example to select the cross
section of the members. Hence the engineer turns to the pin-jointed assemblage of axial-force-
carrying elements and uses it to perform the structural analysis.

2–4
 §2.5 THE EXAMPLE TRUSS

(a) Physical System member

support

joint

IDEALIZATION
(b) Idealized Sytem:
FEM-Discretized
Mathematical Model
;;
;;
;;

;;
;;
;;
Figure 2.3. Idealization of roof truss: (a) physical system, (b) idealization
as FEM discretized mathematical model.

In this and following Chapter we will go over the basic steps of the DSM in a “hand-computer” cal-
culation mode. This means that although the steps are done by hand, whenever there is a procedural
choice we shall either adopt the way that is better suited towards the computer implementation, or
explain the difference between hand and computer computations. The actual computer implemen-
tation using a high-level programming language is presented in Chapter 4.

§2.5. The Example Truss

To keep hand computations manageable we will use just about the simplest structure that can be
called a plane truss, namely the three-member truss illustrated in Figure 2.4(a). The idealized
model of this physical truss as a pin-jointed assemblage of bars as shown in Figure 2.4(b), In this
idealization truss members carry only axial loads, have no bending resistance, and are connected
by frictionless pins.
Geometric, material and fabrication properties of the idealized truss are given in Figure 2.4(c),2
while idealized loads and support conditions are provided in Figure 2.4(d).
It should be noted that as a practical structure the example truss is not particularly useful — that
shown in Figure 2.2 is more common in construction. But with the example truss we can go over
the basic DSM steps without getting mired into too many members, joints and degrees of freedom.

2 Member mass densities are given in Figure 2.4(c) so that this truss can be used later for examples in vibration and dyamics
analysis.

2–5
Chapter 2: THE DIRECT STIFFNESS METHOD I

(a) (b) 3

Idealization as Pin-Jointed Truss

and FEM Discretization

;;
;;
;;
;;
1 2

fy3 , uy3 f y3 = 1
fx3 , ux3 f x3 = 2
(c) 3(10,10) (d) 3

E (3) = 100, A(3) = 

2 2, 45o
(3)  2, ρ(3) = 3/20
L = 10 E (2) = 50, A(2) = 1,
L = 10, ρ(2) = 1/5
(2)

y (3) (2) y (3) (2)

fx1 , ux1 45 o x (1) fx2 , ux2 x (1)

1 2

;
;
;
;
1(0,0) 2(10,0)
fy1 , uy1 E (1) = 50, A(1) = 2, fy2 , uy2
(1)
L = 10, ρ(1) = 1/5

Figure 2.4. The three-member example truss: (a) physical structure; (b) idealization as a pin-jointed bar
assemblage; (c) geometric, material and fabrication properties; (d) support conditions and applied loads.

§2.6. Members, Joints, Forces and Displacements

The pin-jointed idealization of the example truss, pictured in Figure 2.4(b,c,d), has three joints,
which are labeled 1, 2 and 3, and three members, which are labeled (1), (2) and (3). Those members
connect joints 1–2, 2–3, and 1–3, respectively. The member lengths are denoted by L (1) , L (2) and
L (3) , their elastic moduli by E (1) , E (2) and E (3) , and their cross-sectional areas by A(1) , A(2) and
A(3) . Note that an element number superscript is enclosed in parenthesis to avoid confusion with
exponents. Both E and A are assumed to be constant along each member.
Members are generically identified by index e (because of their close relation to finite elements,
as explained below). This index is placed as superscript of member properties. For example, the
cross-section area of a generic member is Ae . The member superscript is not enclosed in parentheses
in this case because no confusion with exponents can arise. But the area of member 3 is written
A(3) and not A3 .
Joints are generically identified by indices such as i, j or n. In the general FEM, the names “joint”
and “member” are replaced by node and element, respectively. This dual nomenclature is used in
the initial Chapters to stress the physical interpretation of the FEM.
The geometry of the structure is referred to a common Cartesian coordinate system {x, y}, which
will be called the global coordinate system. Other names for it in the literature are structure
coordinate system and overall coordinate system. For the example truss its origin is at joint 1.

2–6
 §2.7 THE MASTER STIFFNESS EQUATIONS

The key ingredients of the stiffness method of analysis are the forces and displacements at the joints.
In a idealized pin-jointed truss, externally applied forces as well as reactions can act only at the
joints. All member axial forces can be characterized by the x and y components of these forces,
denoted by f x and f y , respectively. The components at joint i will be identified as f xi and f yi ,
respectively. The set of all joint forces can be arranged as a 6-component column vector called f.
The other key ingredient is the displacement field. Classical structural mechanics tells us that the
displacements of the truss are completely defined by the displacements of the joints. This statement
is a particular case of the more general finite element theory. The x and y displacement components
will be denoted by u x and u y , respectively. The values of u x and u y at joint i will be called u xi
and u yi . Like joint forces, they are arranged into a 6-component vector called u. Here are the two
vectors of nodal forces and nodal displacements, shown side by side:
   
f x1 u x1
 f y1   u y1 
   
 f x2   u x2 
f= , u= . (2.1)
 f y2   u y2 
   
f x3 u x3
f y3 u y3
In the DSM these six displacements are the primary unknowns. They are also called the degrees of
freedom or state variables of the system.3
How about the displacement boundary conditions, popularly called support conditions? This data
will tell us which components of f and u are actual unknowns and which ones are known a priori. In
pre-computer structural analysis such information was used immediately by the analyst to discard
unnecessary variables and thus reduce the amount of hand-carried bookkeeping.
The computer oriented philosophy is radically different: boundary conditions can wait until the
last moment. This may seem strange, but on the computer the sheer volume of data may not be so
important as the efficiency with which the data is organized, accessed and processed. The strategy
“save the boundary conditions for last” will be followed here also for the hand computations.
Remark 2.1. Often column vectors such as (2.1) will be displayed in row form to save space, with a transpose
symbol at the end. For example, f = [ f x1 f y1 f x2 f y2 f x3 f y3 ]T and u = [ u x1 u y1 u x2 u y2 u x3 u y3 ]T .

§2.7. The Master Stiffness Equations

The master stiffness equations relate the joint forces f of the complete structure to the joint dis-
placements u of the complete structure before specification of support conditions.
Because the assumed behavior of the truss is linear, these equations must be linear relations that
connect the components of the two vectors. Furthermore it will be assumed that if all displace-
ments vanish, so do the forces.4 If both assumptions hold the relation must be homogeneous and
3 Primary unknowns is the correct mathematical term whereas degrees of freedom has a mechanics flavor: “any of a limited
number of ways in which a body may move or in which a dynamic system may change” (Merrian-Webster). The term
state variables is used more often in nonlinear analysis, material sciences and statistics.
4 This assumption implies that the so-called initial strain effects, also known as prestress or initial stress effects, are
neglected. Such effects are produced by actions such as temperature changes or lack-of-fit fabrication, and are studied
in Chapter 29.

2–7
Chapter 2: THE DIRECT STIFFNESS METHOD I

Globalization
Disconnection Merge
Breakdown Assembly &
Localization Application of BCs
(Section 2.9) Solution
Element Formation Solution
(Secs 2.10, 3.2-3.4)
Recovery of Derived Quantities

conceptual processing post-processing

steps steps steps

Figure 2.5. The Direct Stiffness Method steps.

expressable in component form as

    
f x1 K x1x1 K x1y1 K x1x2 K x1y2 K x1x3 K x1y3 u x1
 f y1   K y1x1 K y1y1 K y1x2 K y1y2 K y1x3 K y1y3   u y1 
    
 f x2   K x2x1 K x2y1 K x2x2 K x2y2 K x2x3 K x2y3   u x2 
 =  . (2.2)
 f y2   K y2x1 K y2y1 K y2x2 K y2y2 K y2x3 K y2y3   u y2 
    
f x3 K x3x1 K x3y1 K x3x2 K x3y2 K x3x3 K x3y3 u x3
f y3 K y3x1 K y3y1 K y3x2 K y3y2 K y3x3 K y3y3 u y3
In matrix notation:
f = K u. (2.3)

Here K is the master stiffness matrix, also called global stiffness matrix, assembled stiffness matrix,
or overall stiffness matrix. It is a 6 × 6 square matrix that happens to be symmetric, although this
attribute has not been emphasized in the written-out form (2.2). The entries of the stiffness matrix
are often called stiffness coefficients and have a physical interpretation discussed below.
The qualifiers (“master”, “global”, “assembled” and “overall”) convey the impression that there
is another level of stiffness equations lurking underneath. And indeed there is a member level or
element level, into which we plunge in the Breakdown section.
Remark 2.2. Interpretation of Stiffness Coefficients. The following interpretation of the entries of K is valuable
for visualization and checking. Choose a displacement vector u such that all components are zero except the
i th one, which is one. Then f is simply the i th column of K. For instance if in (2.3) we choose u x2 as unit
displacement,
u = [ 0 0 1 0 0 0 ]T , f = [ K x1x2 K y1x2 K x2x2 K y2x2 K x3x2 K y3x2 ]T . (2.4)
Thus K y1x2 , say, represents the y-force at joint 1 that would arise on prescribing a unit x-displacement at
joint 2, while all other displacements vanish. In structural mechanics this property is called interpretation of
stiffness coefficients as displacement influence coefficients. It extends unchanged to the general finite element
method.

§2.8. The DSM Steps

The DSM steps, major and minor, are summarized in Figure 2.5 for the convenience of the reader.
The two major stages are Breakdown, followed by Assembly & Solution. A postprocessing step
may follow, although this is not part of the DSM proper.

2–8
 §2.8 THE DSM STEPS

FEM model:

;;

;;
Remove loads
& supports:

Disassemble:

longerons
Localize: battens
diagonals
longerons

Generic element:

Figure 2.6. Visualization of the DSM breakdown stage.

longerons
Form battens
elements: diagonals
longerons

Globalize:

Merge:

Apply loads
and supports:
;
;; ;

;
;; ;

Solve for joint

displacements:
;;

;;

Figure 2.7. Visualization of the DSM assembly-and-solution stage.

2–9
Chapter 2: THE DIRECT STIFFNESS METHOD I

The first three DSM steps are: (1) disconnection, (2) localization, and (3) computation of member
stiffness equations. Collectively these form the breakdown stage. The first two are flagged as con-
ceptual in Figure 2.5 because they are not actually programmed as such: they are implicitly carried
out either through the user-provided problem definition, or produced by separate preprocessing pro-
grams such as CAD front ends. DSM processing actually begins at the element-stiffness-equation
forming step.
Before starting with the detailed, step by step description of the DSM, it is convenient to exhibit the
whole process through a graphic sequence. This is done in Figures 2.6 and 2.7, which are shown
to students as animated slides. The sequence starts with the FEM model of the typical roof truss
displayed in Figure 2.3(b). Thus the idealization step pictured there is assumed to have been carried
out.5

§2.9. Breakdown Stage

The three breadown steps: disconnection, localization and formation of the element stiffness equa-
tions, are covered next.

(a) 3
3 (b)
Remove loads and supports,
and disconnect pins _ _
x (3)
y (3)
_
x (2)
(3) (2)
y _
_ y (2)
y (1)
_
1 2 1 x x (1)
;
;
;

2
;
;

(1)

Figure 2.8. Disconnection step: (a) idealized example truss; (b) removal of loads and support,
disconnection into members (1), (2) and (3), and selection of local coordinate systems. The
latter are drawn offset from member axes for visualization convenience.

§2.9.1. Disconnection
To carry out the first breakdown step we begin by discarding all loads and supports (the so-called
boundary conditions). Next we disconnect or disassemble the structure into its components. For a
pin-jointed truss disconnection can be visualized as removing the pin connectors. This is illustrated
for the example truss in Figure 2.8. To each member e = 1, 2, 3 assign a Cartesian system {x̄ e , ȳ e }.
Axis x̄ e is aligned along the axis of the eth member. Actually x̄ e runs along the member longitudinal
axis; it is drawn offset in Figure 2.8 (b) for clarity.
By convention the positive direction of x̄ e runs from joint i to joint j, where i < j. The angle
formed by x̄ e and x is the orientation angle ϕ e , positive CCW. The axes origin is arbitrary and may
be placed at the member midpoint or at one of the end joints for convenience.

5 The sequence depicted in Figures 2.6 and 2.7 uses the typical roof truss of Figures 2.2 and 2.3, rather than the 3-member
example truss. Reason: those pictures are more representative of actual truss structures, as seen often by students.

2–10
 §2.9 BREAKDOWN STAGE

Equivalent spring stiffness k s = EA / L

_ _ _ _
(a) fyi, u yi _ fyj, u yj (b)
y
_ _ _ _ _ i j
fxi , uxi x fxj , uxj
−F F
i j
L d

Figure 2.9. Generic truss member referred to its local coordinate system {x̄, ȳ}: (a) idealization as 2-node bar
element, (b) interpretation as equivalent spring. Element identification number e dropped to reduce clutter.

Systems {x̄ e , ȳ e } are called local coordinate systems or member-attached coordinate systems. In
the general finite element method they also receive the name element coordinate systems.
§2.9.2. Localization

To reduce clutter we drop the member identifier e so we are effectively dealing with a generic
truss member, as illustrated in Figure 2.9(a). The local coordinate system is {x̄, ȳ}. The two end
joints are labelled i and j. As shown in that figure, a generic plane truss member has four joint
force components and four joint displacement components (the member degrees of freedom). The
member properties are length L, elastic modulus E and cross-section area A.
§2.9.3. Member Stiffness Equations

The force and displacement components of the generic truss member shown in Figure 2.9(a) are
linked by the member stiffness relations

f̄ = K ū, (2.5)
which written out in full become
 ¯   K̄ xi xi K̄ xi yi K̄ xi x j

K̄ xi y j  ū xi 
f xi
 f¯yi   K̄ yi xi K̄ yi yi K̄ yi x j K̄ yi y j  ū yi 
 ¯ =    . (2.6)
fx j K̄ x j xi K̄ x j yi K̄ x j x j K̄ x j y j ū x j
f¯y j K̄ y j xi K̄ y j yi K̄ y j x j K̄ y j y j ū y j

Vectors f̄ and ū are called the member joint forces and member joint displacements, respectively,
whereas K̄ is the member stiffness matrix or local stiffness matrix. When these relations are
interpreted from the standpoint of the general FEM, “member” is replaced by “element” and “joint”
by ”node.”
There are several ways to construct the stiffness matrix K̄ in terms of L, E and A. The most
straightforward technique relies on the Mechanics of Materials approach covered in undergraduate
courses. Think of the truss member in Figure 2.9(a) as a linear spring of equivalent stiffness ks , an

2–11
Chapter 2: THE DIRECT STIFFNESS METHOD I

interpretation illustrated in Figure 2.9(b). If the member properties are uniform along its length,
Mechanics of Materials bar theory tells us that6

EA
ks = , (2.7)
L
Consequently the force-displacement equation is

EA
F = ks d = d, (2.8)
L
where F is the internal axial force and d the relative axial displacement, which physically is the
bar elongation. The axial force and elongation can be immediately expressed in terms of the joint
forces and displacements as

F = f¯x j = − f¯xi , d = ū x j − ū xi , (2.9)

which express force equilibrium7 and kinematic compatibility, respectively. Combining (2.8) and
(2.9) we obtain the matrix relation8
 ¯    
f xi 1 0 −1 0 ū xi
 f¯yi  EA  0 0 0 0   ū yi 
f̄ =  ¯  =    = K̄ ū, (2.10)
fx j L −1 0 1 0 ū x j
f¯y j 0 0 0 0 ū y j

Hence
 
1 0 −1 0
EA  0 0 0 0
K̄ =  . (2.11)
L −1 0 1 0
0 0 0 0
This is the truss stiffness matrix in local coordinates.
Two other methods for deriving the local force-displacement relation (2.8) are covered in Exercises
2.6 and 2.7.

§2.10. Assembly and Solution Stage: Globalization

The first step in the assembly & solution stage, as shown in Figure 2.5, is globalization. This
operation is done member by member. It refers the member stiffness equations to the global system
{x, y} so it can be merged into the master stiffness. Before entering into details we must establish
relations that connect joint displacements and forces in the global and local coordinate systems.
These are given in terms of transformation matrices.

6 See for example, Chapter 2 of [68].

7 Equations F = f¯x j = − f¯xi follow by considering the free body diagram (FBD) of each joint. For example, take joint i
as a FBD. Equilibrium along x requires −F − f¯xi = 0 whence F = − f¯xi . Doing the same on joint j yields F = f¯x j .
8 The matrix derivation of (2.10) is the subject of Exercise 2.3.

2–12
 §2.10 ASSEMBLY AND SOLUTION STAGE: GLOBALIZATION

(a) Displacement (b) Force _

_ fyj
transformation uyj uyj transformation fyj _
_
uxj fxj
_ _ _ _
y uxj y fxj
x j x j
_ fyi
_ u yi
uyi y fyi _
_
uxi ϕ fxi ϕ

uxi fxi
x i
i
Figure 2.10. The transformation of node displacement and force
components from the local system {x̄, ȳ} to the global system {x, y}.

§2.10.1. Displacement and Force Transformations

The necessary transformations are easily obtained by inspection of Figure 2.10. For the displace-
ments
ū xi = u xi c + u yi s, ū yi = −u xi s + u yi c,
(2.12)
ū x j = u x j c + u y j s, ū y j = −u x j s + u y j c,
in which c = cos ϕ, s = sin ϕ and ϕ is the angle formed by x̄ and x, measured positive CCW
from x. The matrix form that collects these relations is
    
ū xi c s 0 0 u xi
 ū yi   −s c 0 0   u yi 
 =  . (2.13)
ū x j 0 0 c s ux j
ū y j 0 0 −s c uyj

The 4 × 4 matrix that appears above is called a displacement transformation matrix and is denoted9
by T. The node forces transform as f xi = f¯xi c − f¯yi s, etc., which in matrix form become
    f¯xi

f xi c −s 0 0
 yi   s
f c 0 0  f¯yi 
 =  . (2.14)
fx j 0 0 c −s  f¯x j 
fyj 0 0 s c f¯y j

The 4 × 4 matrix that appears above is called a force transformation matrix. A comparison of (2.13)
and (2.14) reveals that the force transformation matrix is the transpose TT of the displacement
transformation matrix T. This relation is not accidental and can be proved to hold generally.10

9 This matrix will be called Td when its association with displacements is to be emphasized, as in Exercise 2.5.
10 A simple proof that relies on the invariance of external work is given in Exercise 2.5. However this invariance was only
checked by explicit computation for a truss member in Exercise 2.4. The general proof relies on the Principle of Virtual
Work, which is discussed later.

2–13
Chapter 2: THE DIRECT STIFFNESS METHOD I

Remark 2.3. Note that in (2.13) the local system (barred) quantities appear on the left-hand side, whereas in
(2.14) they show up on the right-hand side. The expressions (2.13) and and (2.14) are discrete counterparts
of what are called covariant and contravariant transformations, respectively, in continuum mechanics. The
continuum counterpart of the transposition relation is called adjointness. Colectively these relations, whether
discrete or continuous, pertain to the subject of duality.

Remark 2.4. For this particular structural element T is square and orthogonal, that is, TT = T−1 . But
this property does not extend to more general elements. Furthermore in the general case T is not even a
square matrix, and consequently does not possess an ordinary inverse. However the congruent transformation
relations (2.15)–(2.17) given below do hold generally.

§2.10.2. Global Member Stiffness Equations

From now on we reintroduce the member (element) index, e. The member stiffness equations in
global coordinates will be written
f e = K e ue . (2.15)
The compact form of (2.13) and (2.14) for the eth member is
e
ūe = Te ue , fe = (Te )T f̄ . (2.16)
e e
Inserting these matrix expressions into f̄ = K ūe and comparing with (2.15) we find that the
e
member stiffness in the global system {x, y} can be computed from the member stiffness K̄ in the
local system {x̄, ȳ} through the congruent transformation11

e
Ke = (Te )T K̄ Te . (2.17)

Carrying out the matrix multiplications in closed form (Exercise 2.8) we get
 
c2 sc −c2 −sc
e e
E A  sc s2 −sc −s 2 
Ke =  2 , (2.18)
Le −c −sc c2 sc
−sc −s 2 sc s2
in which c = cos ϕ e , s = sin ϕ e , with e superscripts of c and s suppressed to reduce clutter. If the
orientation angle ϕ e is zero we recover (2.10), as may be expected. Ke is called a member stiffness
matrix in global coordinates. The proof of (2.17) and verification of (2.18) is left as Exercise 2.8.
The globalized member stiffness equations for the example truss can now be easily obtained by
inserting appropriate values provided in Figure 2.4(c). into (2.18).
For member (1), with end joints 1–2, E (1) A(1) = 100, L (1) = 10, and ϕ (1) = 0◦ :
 f (1)     u (1) 
x1 1 0 −1 0 x1
(1)
 f y1


  0 0 0 0 u (1) 

y1 
 f (1)  = 10  −1 0 1 0  u x2 
(1) . (2.19)
x2
(1)
f y2 0 0 0 0 u (1)
y2

11 Also known as congruential transformation and congruence transformation in linear algebra books.

2–14
 §2. Notes and Bibliography

For member (2), with end joints 2–3, E (2) A(2) = 50, L (2) = 10, and ϕ (2) = 90◦ :
 f (2)     u (2) 
x2 0 0 0 0 x2
(2)
 f y2 
  0 1 0 −1   u (2) 
 f (2)  = 5  0   (2) 
 . (2.20)
y2

x3
0 0 0 u x3 
(2)
f y3 0 −1 0 1 u (2)
y3
√ √
For member (3), with end joints 1–3, E (3) A(3) = 200 2, L (3) = 10 2, and ϕ (3) = 45◦ :
 f (3)     u (3) 
x1 0.5 0.5 −0.5 −0.5 x1
(3)
 f y1


 = 20  0.5 0.5 −0.5 −0.5  
 u (3)
y1 

 f (3)  
−0.5 −0.5 0.5

0.5  u (3) . (2.21)
x3 x3
(3)
f y3 −0.5 −0.5 0.5 0.5 u (3)
y3

In the following Chapter we complete the DSM steps by putting the truss back together through the
merge step, and solving for the unknown forces and displacements.
Notes and Bibliography
The Direct Stiffness Method has been the dominant FEM version since the mid-1960s, and is the procedure
followed by all major commercial codes in current use. The general DSM was developed at Boeing in the
mid and late 1950s, through the leadership of Jon Turner [766,768], and had defeated its main competitor, the
Force Method, by 1970 [240].
All applications-oriented FEM books cover the DSM, although the procedural steps are sometimes not clearly
delineated. In particular, the textbooks recommended in §1.9 offer adequate expositions.
Trusses, also called bar assemblies, are usually the first structures treated in Mechanics of Materials books
written for undergraduate courses in Aerospace, Civil and Mechanical Engineering. Two widely used textbooks
at this level are [68] and [595].
Steps in the derivation of stiffness matrices for truss elements are well covered in a number of early treatment
of finite element books, of which Chapter 5 of Przemieniecki [603] is a good example.
Force and displacement transformation matrices for structural analysis were introduced by G. Kron [428].

References
Referenced items have been moved to Appendix R.

2–15
Chapter 2: THE DIRECT STIFFNESS METHOD I

Homework Exercises for Chapter 2

The Direct Stiffness Method I

EXERCISE 2.1 [D:10] Explain why arbitrarily oriented mechanical loads on an idealized pin-jointed truss
structure must be applied at the joints. [Hint: idealized truss members have no bending resistance.] How
about actual trusses: can they take loads applied between joints?
EXERCISE 2.2 [A:15] Show that the sum of the entries of each row of the master stiffness matrix K of any
plane truss, before application of any support conditions, must be zero. [Hint: apply translational rigid body
motions at nodes.] Does the property hold also for the columns of that matrix?
EXERCISE 2.3 [A:15] Using matrix algebra derive (2.10) from (2.8) and (2.9). Note: Place all equations in
matrix form first and eliminate d and F by matrix multiplication. Deriving the final form with scalar algebra
and rewriting it in matrix form gets no credit.
T
EXERCISE 2.4 [A:15] By direct multiplication verify that for the truss member of Figure 2.9(a), f̄ ū = F d.
Intepret this result physically. (Hint: what is a force times displacement in the direction of the force?)
EXERCISE 2.5 [A:20] The transformation equations between the 1-DOF spring and the 4-DOF generic truss
member may be written in compact matrix form as
d = Td ū, f̄ = F T f , (E2.1)
T
where Td is 1 × 4 and T f is 4 × 1. Starting from the identity f̄ ū = F d proven in the previous exercise, and
using compact matrix notation, show that T f = TdT . Or in words: the displacement transformation matrix and
the force transformation matrix are the transpose of each other. (This can be extended to general systems)
EXERCISE 2.6 [A:20] Derive the equivalent spring formula F = (E A/L) d of (2.8) by the Theory of
Elasticity relations e = d ū(x̄)/d x̄ (strain-displacement equation), σ = Ee (Hooke’s law) and F = Aσ (axial
force definition). Here e is the axial strain (independent of x̄) and σ the axial stress (also independent of x̄).
Finally, ū(x̄) denotes the axial displacement of the cross section at a distance x̄ from node i, which is linearly
interpolated as 
x̄ x̄
ū(x̄) = ū xi 1 − + ū x j (E2.2)
L L
Justify that (E2.2) is correct since the bar differential equilibrium equation: d[A(dσ/d x̄)]/d x̄ = 0, is verified
for all x̄ if A is constant along the bar.
EXERCISE 2.7 [A:20] Derive the equivalent spring formula F = (E A/L) d of (2.8) by the principle of
Minimum Potential Energy (MPE). In Mechanics of Materials it is shown that the total potential energy of the
axially loaded bar is

L

= 1
2
A σ e d x̄ − Fd, (E2.3)
0
where symbols have the same meaning as the previous Exercise. Use the displacement interpolation (E2.2),
the strain-displacement equation e = d ū/d x̄ and Hooke’s law σ = Ee to express
as a function
(d) of the
relative displacement d only. Then apply MPE by requiring that ∂
/∂d = 0.
e e
EXERCISE 2.8 [A:20] Derive (2.17) from K̄ ūe = f̄ , (2.15) and (2.17). (Hint: premultiply both sides of
e e
K̄ ūe = f̄ by an appropriate matrix). Then check by hand that using that formula you get (2.18). Falk’s
scheme is recommended for the multiplications.12

12 This scheme is useful to do matrix multiplication by hand. It is explained in §B.3.2 of Appendix B.

2–16
 Exercises

EXERCISE 2.9 [D:5] Why are disconnection and localization labeled as “conceptual steps” in Figure 2.5?

EXERCISE 2.10 [C:20] (Requires thinking) Notice that the expression (2.18) of the globalized bar stiffness
matrix may be factored as
   
c2 sc −c2 −sc −c
E A  sc
e e
s2 −sc −s 2   −s  E e Ae
Ke =  −c2 = [ −c −s (E2.4)
sc   c  L e
c s]
Le −sc c2
−sc −s 2 sc s2 s

Interpret this relation physically as a chain of global-to-local-to-global matrix operations: global displacements
→ axial strain, axial strain → axial force, and axial force → global node forces.

2–17
 3
The Direct
Stiffness Method II

3–1
Chapter 3: THE DIRECT STIFFNESS METHOD II

TABLE OF CONTENTS

Page
§3.1 The Remaining DSM Steps . . . . . . . . . . . . . . . . 3–3
§3.2 Assembly: Merge . . . . . . . . . . . . . . . . . . . 3–3
§3.2.1 Governing Rules . . . . . . . . . . . . . . . . 3–3
§3.2.2 Assembly by Hand Using Augment-and-Add . . . . . . 3–4
§3.3 Solution . . . . . . . . . . . . . . . . . . . . . . . 3–6
§3.3.1 Applying Boundary Conditions by Reduction . . . . . . 3–6
§3.3.2 Solving for Displacements . . . . . . . . . . . . . 3–7
§3.4 PostProcessing . . . . . . . . . . . . . . . . . . . . 3–7
§3.4.1 Recovery of Reaction Forces . . . . . . . . . . . . 3–7
§3.4.2 Recovery of Internal Forces and Stresses . . . . . . . . 3–8
§3.4.3 *Reaction Recovery: General Case . . . . . . . . . . 3–9
§3.5 *Computer Oriented Assembly and Solution . . . . . . . . . 3–10
§3.5.1 *Assembly by Freedom Pointers . . . . . . . . . . . 3–10
§3.5.2 *Applying DBC by Modification . . . . . . . . . . 3–11
§3.6 Prescribed Nonzero Displacements . . . . . . . . . . . . . 3–11
§3.6.1 Application of Nonzero-DBCs by Reduction . . . . . . 3–11
§3.6.2 *Application of Nonzero-DBCs by Modification . . . . . . 3–13
§3.6.3 *Matrix Forms of Nonzero-DBC Application Methods . . . 3–14
§3. Notes and Bibliography . . . . . . . . . . . . . . . . . 3–14
§3. References . . . . . . . . . . . . . . . . . . . . . 3–15
§3. Exercises . . . . . . . . . . . . . . . . . . . . . . 3–16

3–2
 §3.2 ASSEMBLY: MERGE

§3.1. The Remaining DSM Steps

Chapter 2 covered the initial steps of the DSM. The three breakdown steps: disconnection, local-
ization and formation of member stiffness take us down all the way to the generic truss element: the
highest level of fragmentation. This is followed by the assembly and solution stage.
Assembly involves merging the stiffness equations of each member into the global stiffness equa-
tions. For this to make sense, the member equations must be referred to a common coordinate
system, which for a plane truss is the global Cartesian system {x, y}. This is done through the
globalization process covered in §2.10. On the computer the formation, globalization and merge
steps are done concurrently, member by member. After all members are processed we have the
free-free master stiffness equations.
Next comes the solution. This process embodies two steps: applying boundary conditions (BC),
and solving for the unknown joint displacements. Application of BC is done by modifying the free-
free master stiffness equations to take into account which components of the joint displacements
and forces are given and which are unknown.
The modified equations are submitted to a linear equation solver, which returns the unknown joint
(node) displacements. As discussed under Notes and Bibliography, on some FEM implemen-
tations — especially programs written in the 1960s and 1970s — one or more of the foregoing
operations are done concurrently.
The solution step completes the DSM proper. Postprocessing steps may follow, in which derived
quantities such as internal forces and stresses are recovered from the displacement solution.

§3.2. Assembly: Merge

3 (b)
(a) 3

Reconnect the pins

(3) (2) y
y

1 x 1 x 2
2
(1)
Figure 3.1. Physical meaning of merge operation: (a) disconnected example truss after
globalization; (b) reconnected truss with the pins put back into the joints.

§3.2.1. Governing Rules

The key operation of the assembly process is the “placement” of the contribution of each member
to the master stiffness equations. The process is technically called merge of individual members.
The merge operation can be physically interpreted as reconnecting that member in the process of

3–3
Chapter 3: THE DIRECT STIFFNESS METHOD II

(a) f3 (b)
3 3
f(3)
− f(3) − f(2)
3
3
3 (3) f(2)
3

(2)

Figure 3.2. The force equilibrium of joint 3 of the example truss, depicted as an FBD in (a). Here
f3 is the known external joint force applied on the joint. Internal forces f(2) (3)
3 and f3 are applied by
the joint on the members, as illustrated in (b). Thus the forces applied by the members on the joint
are −f(2) (3)
3 and −f3 . These forces would act in the directions shown in (a) if members (2) and (3)
were in tension. The free-body equilibrium statement is f3 − f(2) (3) (2) (3)
3 − f3 = 0 or f3 = f3 + f3 . This
(2) (3) (2) (3)
translates into the two component equations: f x3 = f x3 + f x3 and f y3 = f y3 + f y3 , of (3.2).

fabricating the complete structure. For a pin-jointed (model of a) truss structure, reconnection
means inserting the pins back into the joints. See Figure 3.1.
Merge logic is mathematically governed by two rules of structural mechanics:

1. Compatibility of displacements: The displacements of all members that meet

at a joint are the same.
2. Force equilibrium: The sum of internal forces exerted by all members that
meet at a joint balances the external force applied to that joint. (3.1)

The first rule is physically obvious: reconnected joints must move as one entity. The second one
can be visualized by doing the free body diagram (FBD) of the joint, although some care is required
in the separation of external and internal forces, and their signs. Notational conventions to this
effect are explained in Figure 3.2 for joint 3 of the example truss, at which members (2) and (3)
meet. Application of the foregoing rules at that particular joint gives

Rule 1: u (2) (3)

x3 = u x3 , u (2) (3)
y3 = u y3 .
(2) (3) (1) (2) (3) (2) (3) (1) (2) (3)
Rule 2: f x3 = f x3 + f x3 = f x3 + f x3 + f x3 , f y3 = f y3 + f y3 = f y3 + f y3 + f y3 .
(3.2)
(1) (2) (3) (1) (2) (3)
The addition of f x3 to f x3 + f x3 and of f y3 to f y3 + f y3 , respectively, changes nothing because
member (1) is not connected to joint 3. We are simply adding zeros. But appending those terms
enables us to write the compact matrix relation for the complete structure:

f = f(1) + f(2) + f(3) . (3.3)

3–4
 §3.2 ASSEMBLY: MERGE

§3.2.2. Assembly by Hand Using Augment-and-Add

To directly visualize how the two rules (3.1) translate to merging logic, we first augment the global
stiffness equations listed in §2.10.2 by adding zero rows and columns as appropriate to complete
the force and displacement vectors.
For member (1):  (1)   
f x1   u (1)
(1)  10 0 −10 0 0 0 x1
 f y1   0  (1) 
 0 0 0 0 0   u y1 
 (1)     (1) 
 f x2   −10 0 10 0 0 0   u x2 
 (1)  =   . (3.4)
 f y2   0 0 0 0 0 0   u (1) 
     (1) 
y2
 (1) 
f x3 0 0 0 0 0 0 u 
x3
(1) 0 0 0 0 0 0
f y3 u (1)
y3
For member (2):  (2)   (2) 
f x1   u x1
(2)  0 0 0 0 0 0
 f y1   0  (2) 
 0 0 0 0 0   u y1 
 (2)     (2) 
 f x2  0 0 0 0 0 0   u x2 
 (2)  =   . (3.5)
 f y2  0 0 0 5 0 −5   u (2) 
     (2) 
y2
 (2) 
f x3 0 0 0 0 0 0 u 
x3
(2) 0 0 0 −5 0 5
f y3 u (2)
y3
For member (3):
 (3)   (3) 
f x1   u x1
(3)  10 10 0 0 −10 −10
 f y1  (3) 
   10 10 0 0 −10 −10   u y1 
 f (3)     (3) 
 x2   0 0 0 0 0 0   u x2 
 (3)   =  . (3.6)
 f y2   0 0 0 0 0 0   u (3) 
 (3)    y2 
f  −10 −10 0 0 10 10  u (3) 
x3 x3
(3) −10 −10 0 0 10 10
f y3 u (3)
y3
According to the first rule, we can drop the member identifier in the displacement vectors that
appear in the foregoing matrix equations. Hence the reconnected member equations are
 (1) 
f x1   
(1)  10 0 −10 0 0 0 u x1
 f y1
   0 0 0 0 0 0   u y1 
 f (1)    
 x2   −10 0 10 0 0 0   u x2 
 (1)  =   , (3.7)
 f y2   0 0 0 0 0 0   u y2 
 (1)    
f  0 0 0 0 0 0 u x3
x3
(1) 0 0 0 0 0 0 u y3
f y3
 (2) 
f x1   
(2) 0 0 0 0 0 0 u x1
 f y1 
   0 0 0 0 0 0   u y1 
 f (2)    
 x2   0 0 0 0 0 0   u x2 
 (2)  =   , (3.8)
 f y2   0 0 0 5 0 −5   u y2 
 (2)    
f  0 0 0 0 0 0 u x3
x3
(2) 0 0 0 −5 0 5 u y3
f y3

3–5
Chapter 3: THE DIRECT STIFFNESS METHOD II
 (3) 
f x1   
(3)  10 10 0 0 −10 −10 u x1
 f y1
   10 10 0 0 −10 −10   u y1 
 f (3)    
 x2   0 0 0 0 0 0   u x2 
 (3)  =   . (3.9)
 f y2   0 0 0 0 0 0   u y2 
 (3)    
f  −10 −10 0 0 10 10 u x3
x3
(3) −10 −10 0 0 10 10 u y3
f y3
These three equations can be represented in compact matrix notation as
f(1) = K(1) u, f(2) = K(2) u, f(3) = K(3) u. (3.10)

According to the second rule, expressed in matrix form as (3.3), we have


f = f(1) + f(2) + f(3) = K(1) + K(2) + K(3) u = K u, (3.11)

so all we have to do is add the three stiffness matrices that appear in (3.7) through (3.9), and we
arrive at the master stiffness equations:
    
f x1 20 10 −10 0 −10 −10 u x1
 f y1   10 10 0 0 −10 −10   u y1 
    
 f x2   −10 0 10 0 0 0   u x2 
 =
   . (3.12)
 f y2   0 0 0 5 0 −5   u y2 
    
f x3 −10 −10 0 0 10 10 u x3
f y3 −10 −10 0 −5 10 15 u y3
Using this technique member merging becomes simply matrix addition.
This explanation of the assembly process is conceptually the easiest to follow and understand. It
is virtually foolproof for hand computations. However, this is not the way the process is carried
out on the computer because it would be enormously wasteful of storage for large systems. A
computer-oriented procedure is discussed in §3.5.
§3.3. Solution
Having formed the master stiffness equations we can proceed to the solution phase. To prepare the
equations for a linear solver we need to separate known and unknown components of f and u. In
this Section a technique suitable for hand computation is described.
§3.3.1. Applying Boundary Conditions by Reduction
If one attempts to solve the system (3.12) numerically for the displacements, surprise! The solution
“blows up” because the coefficient matrix (the master stiffness matrix) is singular. The mathematical
interpretation of this behavior is that rows and columns of K are linear combinations of each other
(see Remark 3.1 below). The physical interpretation of singularity is that there are unsuppressed
rigid body motions: the truss still “floats” in the {x, y} plane.
To eliminate rigid body motions and render the system nonsingular we must apply the physical
support conditions as displacement boundary conditions. From Figure 2.4(d) we observe that the
support conditions for the example truss are
u x1 = u y1 = u y2 = 0, (3.13)

3–6
 §3.4 POSTPROCESSING

whereas the known applied forces are

f x2 = 0, f x3 = 2, f y3 = 1. (3.14)

When solving the overall stiffness equations by hand, the simplest way to account for support
conditions is to remove equations associated with known zero joint displacements from the master
system. To apply (3.13) we have to remove equations 1, 2 and 4. This can be systematically
accomplished by deleting or “striking out” rows and columns number 1, 2 and 4 from K and the
corresponding components from f and u. The reduced three-equation system is





10 0 0 u x2 f x2 0
0 10 10 u x3 = f x3 = 2 . (3.15)
0 10 15 u y3 f y3 1

Equation (3.15) is called the reduced master stiffness system. The coefficient matrix of this system
is no longer singular.

Remark 3.1. In mathematical terms, the free-free master stiffness matrix K in (3.12) has order N = 6, rank
r = 3 and a rank deficiency of d = N − r = 6 − 3 = 3 (these concepts are summarized in Appendix C.) The
dimension of the null space of K is d = 3. This space is spanned by three independent rigid body motions:
the two rigid translations along x and y and the rigid rotation about z.

Remark 3.2. Conditions (3.13) represent the simplest type of support conditions, namely specified zero
displacements. More general constraint forms, such as prescribed nonzero displacements and multifreedom
constraints, are handled as described in §3.6 and Chapters 8–9, respectively.

§3.3.2. Solving for Displacements

Solving the reduced system by hand (for example, via Gauss elimination) yields



u x2 0
u x3 = 0.4 . (3.16)
u y3 −0.2

This is called a partial displacement solution (also reduced displacement solution) because it
excludes known displacement components. This solution vector is expanded to six components by
including the three specified values (3.13) in the appropiate slots:
   
u x1 0
 u y1   0 
   
u   0 
u =  x2  =  . (3.17)
 u y2   0 
   
u x3 0.4
u y3 −0.2

This is the complete displacement solution, or simply the displacement solution.

3–7
Chapter 3: THE DIRECT STIFFNESS METHOD II

§3.4. PostProcessing
The last processing step of the DSM is the solution for joint displacements. But often the analyst
needs information on other mechanical quantities; for example the reaction forces at the supports, or
the internal member forces. Such quantities are said to be derived because they are recovered from
the displacement solution. The recovery of derived quantities is part of the so-called postprocessing
steps of the DSM. Two such steps are described below.
§3.4.1. Recovery of Reaction Forces
Premultiplying the complete displacement solution (3.17) by K we get
    
20 10 −10 0 −10 −10 0 −2
 10 10 0 0 −10 −10   0   −2 
    
 −10 0 10 0 0 0 0   0
f = Ku =   =  (3.18)
 0 0 0 5 0 −5   0   1 
    
−10 −10 0 0 10 10 0.4 2
−10 −10 0 −5 10 15 −0.2 1
This vector recovers the known applied forces (3.14), as can be expected. Furthermore we get
three reaction forces: f x1 = f y1 = −2 and f y2 = 1 that are associated with the support conditions
(3.13). It is easy to check that the complete force system is in self equilibrium for the free-free
structure; this is the topic of Exercise 3.1. For a deeper look at reaction recovery, study §3.4.3.
§3.4.2. Recovery of Internal Forces and Stresses
Often the structural engineer is not so much interested in displacements as in internal forces and
stresses. These are in fact the most important quantities for preliminary structural design. In pin-
jointed trusses the only internal forces are the axial member forces. For the example truss these
forces, denoted by F (1) , F (2) and F (3) , are depicted in Figure 3.3. The average axial stress σ e is
obtained on dividing F e by the cross-sectional area of the member.
The axial force F e in member e can be obtained as follows. Extract the displacements of member
e from the complete displacement solution u to form ue . Then recover local joint displacements
from ūe = Te ue .
Compute the member elongation d e (relative axial displacement) and recover the axial force from
the equivalent spring constitutive relation:
E e Ae e
d e = ū ex j − ū exi , Fe = d . (3.19)
Le
Note that ū eyi and ū ey j are not needed in computing d e .

Example 3.1. Recover F (2) in example truss. Member (2) goes from node 2 to node 3 and ϕ (2) = 90◦ . Extract
the global displacements of the member from (3.17): u(2) = [ u x2 u y2 u x3 u y3 ]T = [ 0 0 0.4 −0.2 ]T .
Convert to local displacements using ū(2) = T(2) u(2) :
         
ū x2 cos 90◦ sin 90◦ 0 0 u x2 0 1 0 0 0 0
 ū y2   − sin 90◦ cos 90◦ 0 0   u y2   −1 0 0 0 0   0 
 ū  =  0 0
=
cos 90◦ sin 90◦   u x3   0 0 0
=
1   0.4   −0.2 
.
x3
ū y3 0 0 − sin 90◦ cos 90◦ u y3 0 0 −1 0 −0.2 −0.4
(3.20)

3–8
 §3.4 POSTPROCESSING

uy3 =−0.2
ux3 =0.4
(a) (b)
3 3

(3) (2)
(2)
F (3)
F (2) F (2)
Disconnect

0
1 F (1) 2 0
2
(1)

Figure 3.3. Internal force recovery for example truss: (a) member axial forces F (1) , F (2) and
F (3) , with arrow directions pertaining to tension; (b) details of computation for member (2).

The member elongation is d (2) = ū x3 − ū x2 = −0.2 − 0 = −0.2, whence F (2) = (50/10) × (−0.2) = −1, a
compressive axial force.
√
Follwing his procedure for all members provides F (1) = 0, F (2) = −1, and F (3) = 2 2 = 2.82843.

Remark 3.3. An alternative interpretation of (3.19) is to regard ee = d e /L e as the (average) member axial
strain, σ e = E e ee as (average) axial stress, and F e = Ae σ e as the axial force. This is more in tune with the
Theory of Elasticity viewpoint discussed in Exercise 2.6.

§3.4.3. *Reaction Recovery: General Case

Node forces at supports recovered from f = Ku, where u is the complete displacement solution, were called
reactions in §3.4.1. Although the statement is correct for the example truss, it oversimplifies the general case.
To cover it, consider f as the superposition of applied and reaction forces:

K u = f = fa + fr. (3.21)

Here f a collects applied forces, which are known before solving, whereas f r collects unknown reaction forces
to be recovered in post-processing. Entries of f r that are not constrained are set to zero. For the example truss,
 f  0  f x1
  −2  0  −2 
x1
 y1   0  
f f y1   −2   0   −2 
 0  0  0   0  0  0 
f=      f=     
upon assembly upon recovery
⇒  f y2  =  0  +  ⇒  1  = 0 +  1 . (3.22)
f y2 
           
2 2 0 2 2 0
1 1 0 1 1 0
There is a clean separation in (3.22). Every nonzero entry in f comes from either f a or f r . This allows us to
interpret f x1 = f x1
r
, f y1 = f y1
r
and f y2 = f y2
r
as reactions. If nonzero applied forces act directly on supported
freedoms, however, a reinterpretation is in order. This often occurs when distributed loads such as pressure or
own weight are lumped to the nodes. The adjustment can be more easily understood by following the simple
example illustrated in Figure 3.4.
The fixed-free prismatic bar pictured in Figure 3.4(a) is subjected to a uniformly line load q per unit length.
The bar has length L, elastic modulus E and cross-section area A. It is discretized by two equal-size elements
as shown in Figure 3.4(b). The three x node displacements u 1 = u x1 , u 2 = u x2 and u 3 = u x3 are taken as

3–9
Chapter 3: THE DIRECT STIFFNESS METHOD II

;; (a) E, A constant (b) (c)

;;

;;
q (uniform) 1 qL/4 2 qL/2 3 qL/4 qL 1 qL/4 2 qL/2 3 qL/4
x 1 2
L L/2 L/2

Figure 3.4. A simple problem to illustrate reaction recovery of support reactions when nonzero
applied loads act on supports. (a) bar under distributed load; (b) two-element FEM idealization;
(c) free body diagram showing applied node forces in blue and support reaction in red.

degrees of freedom. The line load is converted to node forces f 1a = 14 q L, f 2a = 12 q L and f 3a = 14 q L at

nodes 1, 2 and 3, respectively, using the EbE method discussed in Chapter 7. The master stiffness equations
configured as per (3.21) are





2E A 1 −1 0 u1 qL 1 f 1r
−1 2 −1 u2 = f = fa + fr = 2 + 0 . (3.23)
L 0 −1 1 u3 4 1 0

Applying the displacement BC u 1 = 0 and solving gives u 2 = 3q L 2 /(8E A) and u 3 = q L 2 /(2E A). Force
recovery yields





qL −3 qL −3 qL 1 −q L
f = Ku = 2 , fr = f − fa = 2 − 2 = 0 . (3.24)
4 1 4 1 4 1 0
The fixed-end reaction emerges as f 1r = −q L, the correctness of which may be verified on examining the FBD
of Figure 3.4(c). Note that taking f 1 = −3q L/4 as reaction would be in error by 25%. This general recovery
procedure should always be followed when reaction values are used in the design of structural supports.

§3.5. *Computer Oriented Assembly and Solution

§3.5.1. *Assembly by Freedom Pointers
The practical computer implementation of the DSM assembly process departs significantly from the “augment
and add” technique described in §3.2.2. There are two major differences:
(I) Member stiffness matrices are not expanded. Their entries are directly merged into those of K through
the use of a “freedom pointer array” called the Element Freedom Table or EFT.
(II) The master stiffness matrix K is stored using a special format that takes advantage of symmetry and
sparseness.
Difference (II) is a more advanced topic that is deferred to the last part of the book. For simplicity we shall
assume here that K is stored as a full square matrix, and study only (I). For the example truss the freedom-pointer
technique expresses the entries of K as the sum


3
K pq = K iej for i = 1, . . . 4, j = 1, . . . 4, p = EFTe (i), q = EFTe ( j). (3.25)
e=1

Here K iej denote the entries of the 4 × 4 globalized member stiffness matrices in (3.7) through (3.9). Entries
K pq that do not get any contributions from the right hand side remain zero. EFTe denotes the Element Freedom
Table for member e. For the example truss these tables are

EFT(1) = {1, 2, 3, 4}, EFT(2) = {3, 4, 5, 6}, EFT(3) = {1, 2, 5, 6}. (3.26)

3–10
 §3.6 PRESCRIBED NONZERO DISPLACEMENTS

Physically these tables map local freedom indices to global ones. For example, freedom number 3 of member
(2) is u x3 , which is number 5 in the master equations; consequently EFT(2) (3) = 5. Note that (3.25) involves
three nested loops: over e (outermost), over i, and over j. The ordering of the last two is irrelevant. Advantage
may be taken of the symmetry of Ke and K to roughly halve the number of additions. Exercise 3.5 follows
the scheme (3.25) by hand.
The assembly process for general structures using this technique is studied in Chapter 25.

§3.5.2. *Applying DBC by Modification

In §3.3.1 the support conditions (3.13) were applied by reducing (3.12) to (3.15). Reduction is convenient for
hand computations because it cuts down on the number of equations to solve. But it has a serious flaw for
computer implementation: the equations must be rearranged. It was previously noted that on the computer the
number of equations is not the only important consideration. Rearrangement can be as or more expensive than
solving the equations, particularly if the coefficient matrix is stored in sparse form or on secondary storage.1
To apply support conditions without rearranging the equations we clear (set to zero) rows and columns
corresponding to prescribed zero displacements as well as the corresponding force components, and place
ones on the diagonal to maintain non-singularity. The resulting system is called the modified set of master
stiffness equations. For the example truss this approach yields
1 0 0 0 0 0
 
u x1
 
0
0 1 0 0 0 0   u y1   0 
 0 0 10 0 0     
 0   u x2  =  0  , (3.27)
0 0 0 1 0 0  u y2   0 

    
0 0 0 0 10 10 u x3 2
0 0 0 0 10 15 u y3 1

in which rows and columns for equations 1, 2 and 4 have been cleared. Solving this modified system produces
the complete displacement solution (3.17) directly.

Remark 3.4. In a “smart” stiffness equation solver the modified system need not be explicitly constructed by
storing zeros and ones. It is sufficient to mark the equations that correspond to displacement BCs. The solver
is then programmed to skip those equations. However, if one is using a standard solver from, say, a library
of scientific routines or a commercial program such as Matlab or Mathematica, such intelligence cannot be
expected, and the modified system must be set up explicitly.

§3.6. Prescribed Nonzero Displacements

The support conditions considered in the example truss resulted in the specification of zero dis-
placement components; for example u y2 = 0. There are cases, however, where the known value is
nonzero. This happens, for example, in the study of settlement of foundations of ground structures
such as buildings and bridges, and in the analysis of motion-driven machinery components.
Mathematically these are called non-homogenous boundary conditions. The treatment of this
generalization of the FEM equations is studied in the following subsections.

1 On most modern computers, reading a floating-point number from memory at a random address takes 100 to 1000 times
as long as performing a floating-point arithmetic operation on numbers that are already in registers.

3–11
Chapter 3: THE DIRECT STIFFNESS METHOD II

fy3 = 1
3
fx3 = 2

u y1 = −0.5 u y2 = +0.4
going down going up
2

;;
;;
;;
;;
1 ux1 = 0
no horizontal motion
Figure 3.5. The example truss with prescribed nonzero vertical displacements at joints 1 and 2.

§3.6.1. Application of Nonzero-DBCs by Reduction

We describe first a matrix reduction technique, analogous to that used in §3.3.1, which is suitable
for hand computations. Recall the master stiffness equations (3.12) for the example truss:
    
20 10 −10 0 −10 −10 u x1 f x1
 10 10 0 0 −10 −10   u y1   f y1 
    
 −10 0 10 0 0 0   u x2   f x2 
  =  (3.28)
 0 0 0 5 0 −5   u y2   f y2 
    
−10 −10 0 0 10 10 u x3 f x3
−10 −10 0 −5 10 15 u y3 f y3
Suppose that the applied forces are again (3.14) but the prescribed displacements change to
u x1 = 0, u y1 = −0.5, u y2 = 0.4 (3.29)
This means that joint 1 goes down vertically whereas joint 2 goes up vertically, as depicted in
Figure 3.5. Inserting the known data into (3.28) we get
    
20 10 −10 0 −10 −10 0 f x1
 10 10 0 0 −10 −10   −0.5   f y1 
    
 −10 0 10 0 0 0   u x2   0 
  =
   (3.30)
 0 0 0 5 0 −5   0.4   f y2 
    
−10 −10 0 0 10 10 u x3 2
−10 −10 0 −5 10 15 u y3 1
The first, second and fourth rows of (3.30) are removed, leaving only
 
0

  −0.5 

−10 0 10 0 0 0   0
 u 
−10 −10 0 0 10 10  x2  = 2 (3.31)
 0.4 
−10 −10 0 −5 10 15   1
u x3
u y3

3–12
 §3.6 PRESCRIBED NONZERO DISPLACEMENTS

Columns 1, 2 and 4 are removed by transferring all known terms from the left to the right hand side:






10 0 0 u x2 0 (−10) × 0 + 0 × (−0.5) + 0 × 0.4 0
0 10 10 u x3 = 2 − (−10) × 0 + (−10) × (−0.5) + 0 × 0.4 = −3 .
0 10 15 u y3 1 (−10) × 0 + (−10) × (−0.5) + (−5) × 0.4 −2
(3.32)
These are the reduced stiffness equations. Note that its coefficient matrix of (3.32) is exactly the
same as in the reduced system (3.15) for prescribed zero displacements. The right hand side,
however, is different. It consists of the applied joint forces modified by the effect of known nonzero
displacements. These are called the modified node forces or effective node forces. Solving the
reduced system yields


u x2 0
u x3 = −0.5 . (3.33)
u y3 0.2
Filling the missing entries with the known values (3.29) yields the complete displacement solution
(listed as row vector to save space):
u = [0 −0.5 0 0.4 −0.5 0.2 ]T . (3.34)
Taking the solution (3.34) and going through the postprocessing steps discussed in §3.4, we can
find that reaction forces and internal member forces do not change. This is a consequence of the
fact that the example truss is statically determinate. The force systems (internal and external) in
such structures are insensitive to movements such as foundation settlements.
§3.6.2. *Application of Nonzero-DBCs by Modification
The computer-oriented modification approach follows the same idea outlined in §3.5.2. As there, the main
objective is to avoid rearranging the master stiffness equations. To understand the process it is useful to think
of being done in two stages. First equations 1, 2 and 4 are modified so that they become trivial equations, as
illustrated for the example truss and the displacement boundary conditions (3.29):
 1 0 0 0 0 0
u x1
  0

 0 1 0 0 0 0   u x2   −0.5 
 −10 0    
 0 10 0 0   u x2  =  0  (3.35)
 0 0 0 1 0 0   u y2   0.4 
    
−10 −10 0 0 10 10 u x3 2
−10 −10 0 −5 10 15 u y3 1
The solution of this system recovers (3.30) by construction (for example, the fourth equation is simply 1×u y2 =
0.4). In the next stage, columns 1, 2 and 4 of the coefficient matrix are cleared by transferring all known terms
to the right hand side, following the same procedure explained in (3.33). We thus arrive at
1 0 0 0 0 0
 
u x1
 0

0 1 0 0 0 0   u x2   −0.5 
 0 0 10 0 0     
 0   u x2  =  0  (3.36)
0 0 0 1 0 0     
   u y2   0.4 
0 0 0 0 10 10 u x3 −3
0 0 0 0 10 15 u y3 −2
As before, these are called the modified master stiffness equations. Note that (3.36) retains the original number
and order as well as matrix symmetry. Solving this system yields the complete displacement solution (3.34).
If all prescribed displacements are zero, forces on the right hand side are not modified, and one would get
(3.27) as may be expected.

3–13
Chapter 3: THE DIRECT STIFFNESS METHOD II

Remark 3.5. The modification is not actually programmed as discussed above. First the applied forces
in the right-hand side are modified for the effect of nonzero prescribed displacements, and the prescribed
displacements stored in the reaction-force slots. This is called the force modification step. Second, rows and
columns of the stiffness matrix are cleared as appropriate and ones stored in the diagonal positions. This is
called the stiffness modification step. It is essential that the procedural steps be executed in the indicated order,
because stiffness terms must be used to modify forces before they are zeroed out.

§3.6.3. *Matrix Forms of Nonzero-DBC Application Methods

The reduction and modification techniques for applying DBCs can be presented in compact matrix form. First,
the free-free master stiffness equations Ku = f are partitioned as follows:
  
K11 K12 u1 f1
= . (3.37)
K21 K22 u2 f2

In this matrix equation, subvectors u2 and f1 collect displacement and force components, respectively, that are
known, given or prescribed. Subvectors u1 and f2 collect force and displacement components, respectively,
that are unknown. Forces in f2 represent reactions on supports; consequently f2 is called the reaction vector.
On transferring the known terms to the right hand side the first matrix equation becomes

K11 u1 = f1 − K12 u2 . (3.38)

This is the reduced master equation system. If the support B.C.s are homogeneous (that is, all prescribed
displacements are zero), u2 = 0, and we do not need to change the right-hand side:

K11 u1 = f1 . (3.39)

Examples that illustrate (3.38) and (3.39) are (3.32) and (3.27), respectively.
The computer-oriented modification technique retains the same joint displacement vector as in (3.38) through
the following rearrangement:
  
K11 0 u1 f − K12 u2
= 1 . (3.40)
0 I u2 u2
This modified system is simply the reduced equation (3.38) augmented by the trivial equation Iu2 = u2 . This
system is often denoted as

Ku = f̂. (3.41)

Solving (3.41) yields the complete displacement solution, including the specified displacements u2 .
For the computer implementation it is important to note that the partitioned form (3.37) is only used to allow
use of compact matrix notation. In actual programming the equations are not explicitly rearranged: they retain
their original numbers. For instance, in the example truss





u x1 DOF #1 u x2 DOF #3
u1 = u y1 ≡ DOF #2 , u2 = u x3 ≡ DOF #5 . (3.42)
u y2 DOF #4 u y3 DOF #6

The example shows that u1 and u2 are generally interspersed throughout u. Thus, matrix operations such as
K12 u2 involve indirect (pointer) addressing so as to avoid explicit array rearrangement.

3–14
 §3. Notes and Bibliography

Notes and Bibliography

The coverage of the assembly and solution steps of the DSM, along with globalization and application of BCs,
is not uniform across the wide spectrum of FEM books. Authors have introduced “quirks” although the overall
concepts are not affected. The most common variations arise in two contexts:
(1) Some treatments apply support conditions during merge, explicitly eliminating known displacement
freedoms as the elements are processed and merged into K. The output of the assembly process is what
is called here a reduced stiffness matrix.2
(2) In the frontal solution method of Irons [399,401], assembly and solution are done concurrently. More
precisely, as elements are formed and merged, displacement boundary conditions are applied, and Gauss
elimination and reduction of the right hand side starts once the assembler senses (by tracking an “element
wavefront”) that no more elements contribute to a certain node.
Both variants appeared in FEM programs written during the 1960s and 1970s. They were motivated by
computer resource limitations of the time: memory was scarce and computing time expensive.3 On the
negative side, interweaving leads to unmodular programming (which easily becomes “spaghetti code” in low-
level languages such as Fortran). Since a frontal solver has to access the element library, which is typically
the largest component of a general-purpose FEM program, it has to know how to pass and receive information
about each element. A minor change deep down the element library can propagate and break the solver.
Squeezing storage and CPU savings on present computers is of less significance. Modularity, which simplifies
scripting in higher order languages such as Matlab is desirable because it increases “plug-in” operational
flexibility, allows the use of built-in solvers, and reduces the chance for errors. These priority changes reflect
economic reality: human time is nowadays far more expensive than computer time.
A side benefit of modular assembly-solution separation is that often the master stiffness must be used in a
different way than just solving Ku = f; for example in dynamics, vibration or stability analysis. Or as input
to a model reduction process. In those cases the solution stage can wait.
Both the hand-oriented and computer-oriented application of boundary conditions have been presented here,
although the latter is still considered an advanced topic. While hand computations become unfeasible beyond
fairly trivial models, they are important from a instructional standpoint.
The augment-and-add procedure for hand assembly of the master stiffness matrix is due to H. Martin [474].
The general-case recovery of reactions, as described in §3.4.3, is not covered in any FEM textbook.

References
Referenced items have been moved to Appendix R.

2 For the example truss, the coefficient matrix in (3.15) is a reduced stiffness whereas that in (3.27) is a modified one.
3 As an illustration, the first computer used by the writer, the “classical mainframe” IBM 7094, had a magnetic-core
memory of 32,768 36-bit words (≈ 0.2 MB), and was as fast as an IBM PC of the mid 1980s. One mainframe, with
the processing power of an iPhone, served the whole Berkeley campus. Ph.D. students were allocated 2 CPU hours per
semester. Getting a moderately complex FE model through involved heavy use of slower secondary storage such as disk
(or magnetic tape) in batch jobs.

3–15
Chapter 3: THE DIRECT STIFFNESS METHOD II

Homework Exercises for Chapter 3

The Direct Stiffness Method II

EXERCISE 3.1 [A:15] Draw a free body diagram of the nodal forces (3.18) acting on the free-free truss
structure, and verify that this force system satisfies translational and rotational (moment) equilibrium.

EXERCISE 3.2 [A:15] Using the method presented √ in §3.4.2 compute the axial forces in the three members
of the example truss. Partial answer: F (3) = 2 2.

EXERCISE 3.3 [A:20] Describe an alternative method that recovers the axial member forces of the example
truss from consideration of joint equilibrium, without going through the computation of member deformations.
Can this method be extended to arbitrary trusses?

EXERCISE 3.4 [A:20] Suppose that the third support condition in (3.13) is u x2 = 0 instead of u y2 = 0.
Rederive the reduced system (3.15) for this case. Verify that this system cannot be solved for the joint
displacements u y2 , u x3 and u y3 because the reduced stiffness matrix is singular.4 Offer a physical interpretation
of this failure.

EXERCISE 3.5 [N:20] Construct by hand the free-free master stiffness matrix of (3.12) using the freedom-
pointer technique (3.25). Note: start from K initialized to the null matrix, then cycle over e = 1, 2, 3.

f y2 = 0

u y2
2 u x2 f x2 = P

E, A(1) (1)
y
(2)
E, A(2)
H

;; ;;
1 x 3

S/2 S/2
S
Figure E3.1. Truss structure for Exercises 3.6 and 3.7.

EXERCISE 3.6 [N:25] Consider the two-member arch-truss structure shown in Figure E3.1. Take span
S = 8, height H = 3, elastic modulus E = 1000, cross section areas A(1) = 2 and A(2) = 4, and horizontal
crown force P = f x2 = 12. Using the DSM carry out the following steps:
(a) Assemble the master stiffness equations. Any method: augment-and-add, or the more advanced “freedom
pointer” technique explained in §3.5.1, is acceptable.
(b) Apply the displacement BCs and solve the reduced system for the crown displacements u x2 and u y2 .
Partial result: u x2 = 9/512 = 0.01758.
(c) Recover the node forces at all joints including reactions. Verify that overall force equilibrium (x forces,
y forces, and moments about any point) is satisfied.
(d) Recover the axial forces in the two members. Result should be F (1) = −F (2) = 15/2.

4 A matrix is singular if its determinant is zero; cf. §C.2 of Appendix C for a “refresher” in that topic.

3–16
 Exercises

EXERCISE 3.7 [N:20] Resolve items (a) through (c) —

omitting (d) — of the problem of Exercise 3.6 if the vertical y

;;
;;
;;
right support “sinks” so that the displacement u y3 is now
prescribed to be −0.5. Everything else is the same. Use the x
matrix reduction scheme of §3.6.1 to apply the displacement 2 3 4
BCs. (2)
(1) α α (3)
EXERCISE 3.8 [A/C:25] Consider the truss problem L
defined in Figure E3.2. All geometric and material E and A same for
properties: L, α, E and A, as well as the applied forces all three bars
P and H , are to be kept as variables. This truss has 8 1 H
degrees of freedom, with six of them removable by the
fixed-displacement conditions at nodes 2, 3 and 4. Unlike P
previous examples, this structure is statically indeterminate
as long as α = 0. Figure E3.2. Truss structure for Exercise 3.8.

(a) Show that the master stiffness equations are

    
2cs 2 0 −cs 2 c2 s 0 0 −cs 2 −c2 s u x1 H
 1 + 2c3 c2 s −c3 0 −1 −c2 s −c   u y1   −P 
3
 −c2 s 0     
 cs 2 0 0 0   u x2   0 
EA 
 c3 0 0 0 0   u y2   0 
   

L  0 0 0 0   u x3  =  0  , (E3.1)
    
 1 0 0   u y3   0 
 cs 2 c2 s
  u x4
  0

symm c3 u y4 0
in which c = cos α and s = sin α. Explain from physics why the 5th row and column contain only zeros.
(b) Apply the BCs and show the 2-equation modified stiffness system.
(c) Solve for the displacements u x1 and u y1 . Check that the solution makes physical sense for the limit cases
α → 0 and α → π/2. Why does u x1 “blow up” if H = 0 and α → 0?
(d) Recover the axial forces in the three members. Partial answer: F (3) = −H/(2s) + Pc2 /(1 + 2c3 ). Why
do F (1) and F (3) “blow up” if H = 0 and α → 0?

3–17
 4
Analysis
of Example Truss
by a CAS

4–1
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

TABLE OF CONTENTS

Page
§4.1 Computer Algebra Systems . . . . . . . . . . . . . . . . 4–3
§4.1.1 Why Mathematica? . . . . . . . . . . . . . . . 4–3
§4.1.2 How to Get It . . . . . . . . . . . . . . . . . 4–3
§4.1.3 Programming Style and Prerequisites . . . . . . . . . 4–4
§4.1.4 Class Demo Scripts . . . . . . . . . . . . . . . 4–5
§4.2 Program Organization . . . . . . . . . . . . . . . . . 4–6
§4.3 The Element Stiffness Module . . . . . . . . . . . . . . . 4–7
§4.3.1 Module Description . . . . . . . . . . . . . . . 4–7
§4.3.2 Programming Remarks . . . . . . . . . . . . . . 4–8
§4.3.3 Case Sensitivity . . . . . . . . . . . . . . . . 4–9
§4.3.4 Testing the Member Stiffness Module . . . . . . . . . 4–9
§4.4 Merging a Member into the Master Stiffness . . . . . . . . . 4–9
§4.5 Assembling the Master Stiffness . . . . . . . . . . . . . . 4–11
§4.6 Modifying the Master System . . . . . . . . . . . . . . 4–11
§4.7 Recovering Internal Forces . . . . . . . . . . . . . . . . 4–13
§4.8 Putting the Pieces Together . . . . . . . . . . . . . . . 4–14
§4.8.1 The Driver Script . . . . . . . . . . . . . . . . 4–14
§4.8.2 Is All of This Worthwhile? . . . . . . . . . . . . 4–15
§4. Notes and Bibliography . . . . . . . . . . . . . . . . . 4–17
§4. References . . . . . . . . . . . . . . . . . . . . . 4–17
§4. Exercises . . . . . . . . . . . . . . . . . . . . . . 4–18

4–2
 §4.1 COMPUTER ALGEBRA SYSTEMS

§4.1. Computer Algebra Systems

Computer algebra systems, known by the acronym CAS, are programs designed to perform sym-
bolic and numeric manipulations following the rules of mathematics.1 The development of such
programs began in the mid 1960s. The first comprehensive system — the “granddaddy” of them
all, called Macsyma (an acronym for Project Mac Symbolic Manipulator) — was developed using
the programming language Lisp at MIT’s famous Artificial Intelligence Laboratory over the period
1967 to 1980.
The number and quality of symbolic-manipulation programs has expanded dramatically since the
availability of graphical workstations and personal computers has encouraged interactive and ex-
perimental programming. As of this writing the leading general-purpose contenders are Maple and
Mathematica.2 In addition there are a dozen or so more specialized programs, some of which are
available free or at very reasonable cost. See Notes and Bibliography at the end of the Chapter.
§4.1.1. Why Mathematica?

In the present book Mathematica will be used for Chapters and Exercises that develop symbolic
and numerical computation for matrix structural analysis and FEM implementations. Mathematica
is a commercial product developed by Wolfram Research, web site: http://www.wolfram.com.
The version used to construct the code fragments presented in this Chapter is 4.1, which was
commercially released in 2001. (Update: The latest version is 8, released in 2011.) The main
advantages of Mathematica for technical computing are:
1. Availability on a wide range of platforms that range from PCs and Macs through Unix work-
stations. At CU Boulder, it is available through a free campus license (see §4.1).
2. Up-to-date user interface. On all machines Mathematica offers a graphics user interface
called the Notebook front-end. This is mandatory for serious work. It provides professional
typesetting for results.
3. A powerful programming language.
4. Good documentation and abundance of application books at all levels.
One common disadvantage of CAS, and Mathematica is no exception, is computational inefficiency
in numerical calculations compared with a low-level implementation in, for instance, C or Fortran.
The relative penalty can reach several orders of magnitude. For instructional use, however, the
penalty is acceptable when compared to human efficiency. This means the ability to get FEM
programs up and running in very short time, with capabilities for symbolic manipulation and
graphics as a bonus.

1 Some CAS vendors call that kind of activity “doing mathematics by computer.” It is more appropriate to regard such
programs as enabling tools that help humans with complicated and error-prone manipulations. Mettle and metal. As of
now, only humans can do mathematics.
2 Another commonly used program for engineering computations: Matlab, does only numerical computations although a
[poorly done] interface to Maple can be purchased as a toolbox. Historical notes: Macsyma died as a commercial product
in 1999, although the Lisp source code of some versions is freely available. The Maple developer company Waterloo
Maple Inc., also known as Maplesoft, was purchased in 2009 by the Japanese software retailer Cybernet Systems. As a
result, the Matlab symbolic toolbox is likely to be replaced.

4–3
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

Eligibility: students, faculty, staff and departments of all CU campuses

Platforms: Mac OSX, Windows, Linux, Solaris, AIX, HP-UX

How to Get It

1) Download and install software as per instructions at

http://oit.colorado.edu/software-hardware/site-licenses/mathematica
or request an installation CD from the Site Licensing office:
[email protected], 303-492-8995

2) Register your copy online at http://register.wolfram.com/ using

campus license number L2437-5121 plus your computer MathID number.
Be sure to use your CU mail address.

3) A password (unique to your computer) will be forwarded to you via

email from Site Licensing

ITS Support

Question & tech support problems: send email to [email protected]

Periodic hands-on workshops are available for those new to the software,
click on Workshops on the Web page given above.
Details (for example: what is a MathID?) about licensing & support are
posted at the above web site.

Figure 4.1. Instructions to get Mathematica for free from CU-OIT (Office of Information
Technology). One correction: the CD version is no longer available.

§4.1.2. How to Get It

Starting 1 August 2007, a free one-year license from CU’s Information Technology Services (ITS)
is available. Students may renew this license as long as they are registered. See Figure 4.1 for the
“How to Get It” instructions.
If you plan to keep Mathematica for a longer time, an academic version is available.3
One word of caution about the campus-licensed free copy: once installed on your laptop or desktop, it
cannot be transferred to another computer since the license is forever keyed to the disk identification.
(For example, it won’t launch if the disk is erased or replaced.)
§4.1.3. Programming Style and Prerequisites
The following material assumes that you are a moderately experienced user of Mathematica, or
are willing to learn to be one. See Notes and Bibliography for a brief discussion of tutorial and
reference materials in case you are interested.
The best way to learn it from scratch is trying it as a calculator and using online help as needed.
Practice with the program until you reach the level of writing functions, modules and scripts with
relative ease. With the Notebook interface and a good primer it takes only a few hours. When
approaching that level you may notice that functions in Mathematica display many aspects similar

3 This was written in 2001; this option does not seem to exist anymore.

4–4
 §4.1 COMPUTER ALGEBRA SYSTEMS

to C.4 You can exploit this similarity if you are proficient in that language. But Mathematica
functions do have some unique aspects, such as matching arguments by pattern, and the fact that
internal variables are global unless otherwise made local.5
Modification of function arguments should be avoided because it may be difficult to trace side
effects. The programming style enforced here outlaws output arguments and a function can only
return its name. But since the name can be a list of arbitrary objects the restriction is not serious.6
Our objective is to develop a symbolic program written in Mathematica that solves the example
plane truss as well as some symbolic versions thereof. The program will rely heavily on the
development and use of functions implemented using the Module construct of Mathematica. Thus
the style will be one of procedural programming.7 The program will not be particularly modular
(in the computer science sense) because Mathematica is not suitable for that programming style.8
The code presented in §4.2 through §4.8 uses a few language constructs that may be deemed as
advanced, and these are briefly noted in the text so that appropriate reference to the Mathematica
reference manual can be made.
§4.1.4. Class Demo Scripts
The cell scripts shown in Figures 4.2 and 4.3 will be used to illustrate the organization of a Note-
book file and the “look and feel” of some basic Mathematica commands. These scripts will be
demonstrated in class from a laptop.

4 Simple functions can be implemented in Mathematica directly, for instance DotProduct[x ,y ]:=x.y; more com-
plicated functions are handled by the Module construct. These constructs are called rules by computer scientists.
5 In Mathematica everything is a function, including programming constructs. Example: in C for is a loop-opening
keyword, whereas in Mathematica For is a function that runs a loop according to its arguments.
6 Such restrictions on arguments and function returns are closer in spirit to C than Fortran although you can of course
modify C-function arguments using pointers — exceedingly dangerous but often unavoidable.
7 The name Module should not be taken too seriously: it is far away from the concept of modules in Ada, Modula, Oberon
or Fortran 90. But such precise levels of interface control are rarely needed in symbolic languages.
8 Indeed none of the CAS packages in popular use is designed for strong modularity because of historical and interactivity
constraints.

4–5
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

Integration example

f[x_,α_,β_]:=(1+β*x^2)/(1+α*x+x^2);
F=Integrate[f[x,-1,2],{x,0,5}];
F=Simplify[F];
Print[F]; Print[N[F]];
F=NIntegrate[f[x,-1,2],{x,0,5}];
Print["F=",F//InputForm];

10 + Log[21]
13.0445
F=13.044522437723455

Figure 4.2. Example cell for class demo.

Fa=Integrate[f[z,a,b],{z,0,5}]; Fa=Simplify[Fa]; Print["Fa=",Fa];

Plot3D[Fa,{a,-1.5,1.5},{b,-10,10},ViewPoint->{-1,-1,1}];
Fa=FullSimplify[Fa]; (* very slow but you get *) Print["Fa=",Fa];

Result after Simplify[ ..]

50
0
-50
10
1
5

0 0

-5
-1

-10
Result after FullSimplify[ .. ]

Figure 4.3. Another example cell for class demo. (Note: displayed results were obtained
with Mathematica version 4.2. Integration answers from versions 5 and up are quite different.)

4–6
 §4.3 THE ELEMENT STIFFNESS MODULE

(a) Problem (b) Driver script (not a

User prepared module)
Driver
script Cell 7

Stiffness Built in Internal AssembleMasterStiffOfExampleTruss ModifiedMasterStiffForDBC Built in IntForcesOfExampleTruss

Application of Equation
Assembler Equation Force Recovery Cell 3 Solver Cell 6
Support BCs Cell 4
Solver

Merge of MergeElemIntoMasterStiff

Bar Stiffness Cell 2

Derived in
Bar Chapter 2 Bar Internal ElemStiff2DTwoNodeBar IntForces2DTwoNodeBar

Stiffness Force Cell 1 Cell 5

Element Library Element Library

Figure 4.4. Example truss program: (a) organization by function; (b) organization by cell and module names.

§4.2. Program Organization

The overall organization of the example truss program is flowcharted in Figure 4.4. Figure 4.4(a)
illustrates the program as divided into functional tasks. For example the driver program calls the
stiffness assembler, which in turn calls two functions: form bar stiffness and merge into master
stiffness. This kind of functional division would be provided by any programming language.
On the other hand, Figure 4.4(b) is Mathematica specific. It displays the names of the module that
implement the functional tasks and the cells where they reside. (Those program objects: modules
and cells, are described in the sections below.)
The following sections describe the code segments of Figure 4.4(b), one by one. For tutorial
purposes this is done in a “bottom up” fashion, that is, going cell by cell, from left to right and
bottom to top.

§4.3. The Element Stiffness Module

As our first FEM code segment, the top box of Figure 4.5 shows a module that evaluates and returns
the 4 × 4 stiffness matrix of a plane truss member (two-node bar) in global coordinates. The text in
that box of that figure is supposed to be placed on a Notebook cell. Executing the cell, by clicking
on it and hitting an appropriate key (<Enter> on a Mac), gives the output shown in the bottom box.
The contents of the figure is described in further detail below.
§4.3.1. Module Description
The stiffness module is called ElemStiff2DTwoNodeBar. Such descriptive names are permitted
by the language. This reduces the need for detailed comments.
The module takes two arguments:
{ { x1,y1 },{ x2,y2 } } A two-level list9 containing the {x, y} coordinates of the bar end

9 A level-one list is a sequence of items enclosed in curly braces. For example: { x1,y1 } is a list of two items. A level-two
list is a list of level-one lists. An important example of a level-two list is a matrix.

4–7
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

ElemStiff2DTwoNodeBar[{{x1_,y1_},{x2_,y2_}},{Em_,A_}] :=
Module[{c,s,dx=x2-x1,dy=y2-y1,L,Ke},
L=Sqrt[dx^2+dy^2]; c=dx/L; s=dy/L;
Ke=(Em*A/L)* {{ c^2, c*s,-c^2,-c*s},
{ c*s, s^2,-s*c,-s^2},
{-c^2,-s*c, c^2, s*c},
{-s*c,-s^2, s*c, s^2}};
Return[Ke]
];
Ke= ElemStiff2DTwoNodeBar[{{0,0},{10,10}},{100,2*Sqrt[2]}];
Print["Numerical elem stiff matrix:"]; Print[Ke//MatrixForm];
Ke= ElemStiff2DTwoNodeBar[{{0,0},{L,L}},{Em,A}];
Ke=Simplify[Ke,L>0];
Print["Symbolic elem stiff matrix:"]; Print[Ke//MatrixForm];

Numerical elem stiff matrix:

10 10 −10 −10
10 10 −10 −10
−10 −10 10 10
−10 −10 10 10
Symbolic elem stiff matrix:
A Em A Em A Em A Em
2 2 L 2 2 L 2 2 L 2 2 L
A Em A Em A Em A Em
2 2 L 2 2 L 2 2 L 2 2 L
A Em A Em A Em A Em
2 2 L 2 2 L 2 2 L 2 2 L
A Em A Em A Em A Em
2 2 L 2 2 L 2 2 L 2 2 L

Figure 4.5. Module ElemStiff2DTwoNodeBar to form the element stiffness of a 2D

2-node truss element in global coordinates. Test statements (in blue) and test output.

nodes labelled as 1 and 2.10

{ Em,A } A one-level list containing the bar elastic modulus, E and the member
cross section area, A. See §4.3.3 as to why name E cannot be used.
The use of the underscore after argument item names in the declaration of the Module is a require-
ment for pattern-matching in Mathematica. If, as recommended, you have learned functions and
modules this language-specific feature should not come as a surprise.
The module name returns the 4 × 4 member stiffness matrix internally called Ke. The logic that
leads to the formation of that matrix is straightforward and need not be explained in detail. Note,
however, the elegant direct declaration of the matrix Ke as a level-two list, which eliminates the
fiddling around with array indices typical of low-level programming languages. The specification
format in fact closely matches the mathematical expression given as (2.18) in Chapter 2.
§4.3.2. Programming Remarks
The function in Figure 4.5 uses several intermediate variables with short names: dx, dy, s, c and
L. It is strongly advisable to make these symbols local to avoid potential names clashes somewhere
else.11 In the Module[ ...] construct this is done by listing those names in a list immediately

10 These are called the local node numbers, and replace the i, j of previous Chapters. This is a common FEM programming
practice.
11 The “global by default” choice is the worst one, but we must live with the rules of the language.

4–8
 §4.4 MERGING A MEMBER INTO THE MASTER STIFFNESS

after the opening bracket. Local variables may be initialized when they are constants or simple
functions of the argument items; for example on entry to the module dx=x2-x1 initializes variable
dx to be the difference of x node coordinates, namely x = x2 − x1 .
The Return statement fulfills the same purpose as in C or Fortran 90. Mathematica guides and
textbooks advise against the use of that and other C-like constructs. The writer strongly disagrees:
the Return statement makes clear what the Module gives back to its invoker and is self-documenting.
§4.3.3. Case Sensitivity
Mathematica, like most recent computer languages, is case sensitive so that for instance E is not the
same as e. This is fine. But the language designer decided that names of system-defined objects
such as built-in functions and constants must begin with a capital letter. Consequently the liberal
use of names beginning with a capital letter may run into clashes. For example you cannot use E
because of its built-in meaning as the base of natural logarithms.12
In the code fragments presented throughout this book, identifiers beginning with upper case are used
for objects such as stiffness matrices, modulus of elasticity, and cross section area. This follows
established usage in Mechanics. When there is danger of clashing with a protected system symbol,
additional lower case letters are used. For example, Em is used for the elastic modulus instead of E
because (as noted above) the latter is a reserved symbol.
§4.3.4. Testing the Member Stiffness Module
Following the definition of ElemStiff2DTwoNodeBar in Figure 4.5 there are several statements
that constitute the module test script. This script calls the module and prints returned results. Two
cases are tested. First, the stiffness of member (3) of the example truss, using all-numerical values.
Next, some of the input arguments for the same member are given symbolic names so they stand
for variables. For example, the elastic module is given as Em instead of 100 as in the foregoing test.
The print output of the test is shown in the lower portion of Figure 4.5.
The first test returns the member stiffness matrix (?), as may be expected. The second test returns a
symbolic form in which three symbols appear: the coordinates of end node 2, which is taken to be
located at { L,L } instead of {10, 10}, A, which is the cross-section area and Em, which is the elastic
modulus. Note that the returning matrix Ke is subject to a Simplify step before printing it, the
reason for which is the subject of an Exercise. The ability to carry along variables is of course a
fundamental capability of any CAS, and the main reason for which such programs are used.
§4.4. Merging a Member into the Master Stiffness
The next code fragment, listed in Figure 4.6, is used in the assembly step of the DSM.
Module MergeElemIntoMasterStiff receives the 4 × 4 element stiffness matrix formed by
FormElemStiff2DNodeBar and “merges” it into the master stiffness. It takes three arguments:
Ke The 4 × 4 member stiffness matrix to be merged. This is a level-two list.
eftab The column of the Element Freedom Table or EFT, defined in §3.5.1, appropriate
to the member being merged. Recall that the EFT lists the global equation numbers
12 In retrospect this appears to have been a highly questionable decision. System defined names should have been identified
by a reserved prefix or postfix to avoid surprises, as done in Macsyma or Maple. Mathematica issues a warning message,
however, if an attempt to redefine a “protected symbol” is made.

4–9
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

MergeElemIntoMasterStiff[Ke_,eftab_,Kin_]:=Module[
{i,j,ii,jj,K=Kin},
For [i=1, i<=4, i++, ii=eftab[[i]];
For [j=i, j<=4, j++, jj=eftab[[j]];
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]]
]
]; Return[K]
];
K=Table[0,{6},{6}];
Print["Initialized master stiffness matrix:"];
Print[K//MatrixForm]
Ke=ElemStiff2DTwoNodeBar[{{0,0},{10,10}},{100,2*Sqrt[2]}];
Print["Member stiffness matrix:"]; Print[Ke//MatrixForm];
K=MergeElemIntoMasterStiff[Ke,{1,2,5,6},K];
Print["Master stiffness after member merge:"];
Print[K//MatrixForm];

Initialized master stiffness matrix:

0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
0 0 0 0 0 0
Member stiffness matrix:
10 10 −10 −10
10 10 −10 −10
−10 −10 10 10
−10 −10 10 10
Master stiffness after member merge:
10 10 0 0 −10 −10
10 10 0 0 −10 −10
0 0 0 0 0 0
0 0 0 0 0 0
−10 −10 0 0 10 10
−10 −10 0 0 10 10

Figure 4.6. Module MergeElemIntoMasterStiff to merge a 4 × 4 bar element

stiffness into the master stiffness matrix. Test statements (in blue) and test output.

for the four member degrees of freedom, cf. (3.26) in Chapter 3. This is a level-one
list consisting of 4 integers.
Kinp The incoming 6 × 6 master stiffness matrix. This is a level-two list.
MergeElemIntoMasterStiff returns, as module name, the updated master stiffness matrix inter-
nally called K with the member stiffness merged in. We encounter here a novelty: an input-output
argument. Because a formal argument cannot be modified, the situation is handled by copying the
incoming Kinp into K (a local variable declared in the list that follows the Module name) on entry.
That is the copy updated and returned via the Return statement. The implementation has a strong
C flavor with two nested For loops. Because the iterators are very simple, nested Do loops could
have been used as well.
The statements after the module provide a simple test. Before the first call to this function, the
master stiffness matrix must be initialized to a zero 6×6 array. This is done in the first test statement
using the Table function. The test member stiffness matrix is that of member (3) of the example
truss, and is obtained by calling ElemStiff2DTwoNodeBar. The EFT is { 1,2,5,6 } since element
freedoms 1,2,3,4 map into global freedoms 1,2,5,6. Running the test statements yields the listing

4–10
 §4.6 MODIFYING THE MASTER SYSTEM

AssembleMasterStiffOfExampleTruss[]:=
Module[{Ke,K=Table[0,{6},{6}]},
Ke=ElemStiff2DTwoNodeBar[{{0,0},{10,0}},{100,1}];
K= MergeElemIntoMasterStiff[Ke,{1,2,3,4},K];
Ke=ElemStiff2DTwoNodeBar[{{10,0},{10,10}},{100,1/2}];
K= MergeElemIntoMasterStiff[Ke,{3,4,5,6},K];
Ke=ElemStiff2DTwoNodeBar[{{0,0},{10,10}},{100,2*Sqrt[2]}];
K= MergeElemIntoMasterStiff[Ke,{1,2,5,6},K];
Return[K]
];
K=AssembleMasterStiffOfExampleTruss[];
Print["Master stiffness of example truss:"]; Print[K//MatrixForm];

Master stiffness of example truss:

20 10 −10 0 −10 −10
10 10 0 0 −10 −10
−10 0 10 0 0 0
0 0 0 5 0 −5
−10 −10 0 0 10 10
−10 −10 0 −5 10 15

Figure 4.7. Module AssembleMasterStiffOfExampleTruss that forms the 6 × 6

master stiffness matrix of the example truss. Test statements (in blue) and test output.

given in Figure 4.6. The output is as expected.

§4.5. Assembling the Master Stiffness

The module AssembleMasterStiffOfExampleTruss, listed in the top box of Figure 4.7, makes
use of the foregoing two modules: ElemStiff2DTwoNodeBar and MergeElemIntoMasterStiff,
to form the master stiffness matrix of the example truss. The initialization of the stiffness matrix
array in K to zero is done by the Table function of Mathematica, which is handy for initializing
lists. The remaining statements are self explanatory. The module is similar in style to C functions
with no arguments. All the example truss data is “wired in.”
The output from the test script is shown in the lower box of Figure 4.7. The output stiffness matches
that in Equation (3.12), as can be expected if all code fragments used so far work correctly.

§4.6. Modifying the Master System

Following the assembly process the master stiffness equations Ku = f must be modified to account
for single-freedom displacement boundary conditions. This is done through the computer-oriented
equation modification process outlined in §3.5.2.
Module ModifiedMasterStiffForDBC and ModifiedMasterForcesForDBC carry out this
modification for K and f, respectively. These two modules are listed in the top box of Figure 4.8,
along with test statements. The logic of both functions is considerably simplified by assuming that
all prescribed displacements are zero. That is, the BCs are homogeneous. (The more general case
of nonzero prescribed values is treated in Part III of the book.) The module has two arguments:
pdof A list of the prescribed degrees of freedom identified by their global number. For
the example truss this list contains three entries: {1, 2, 4}.
K The master stiffness matrix K produced by the assembly process.

4–11
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

ModifiedMasterStiffForDBC[pdof_,K_] := Module[
{i,j,k,nk=Length[K],np=Length[pdof],Kmod=K},
For [k=1,k<=np,k++, i=pdof[[k]];
For [j=1,j<=nk,j++, Kmod[[i,j]]=Kmod[[j,i]]=0];
Kmod[[i,i]]=1];
Return[Kmod]
];
ModifiedMasterForcesForDBC[pdof_,f_] := Module[
{i,k,np=Length[pdof],fmod=f},
For [k=1,k<=np,k++, i=pdof[[k]]; fmod[[i]]=0];
Return[fmod]
];
K=Array[Kij,{6,6}]; Print["Assembled master stiffness:"];
Print[K//MatrixForm];
K=ModifiedMasterStiffForDBC[{1,2,4},K];
Print["Master stiffness modified for displacement B.C.:"];
Print[K//MatrixForm];
f=Array[fi,{6}]; Print["Force vector:"]; Print[f];
f=ModifiedMasterForcesForDBC[{1,2,4},f];
Print["Force vector modified for displacement B.C.:"]; Print[f];

Assembled master stiffness:

Kij 1, 1 Kij 1, 2 Kij 1, 3 Kij 1, 4 Kij 1, 5 Kij 1, 6
Kij 2, 1 Kij 2, 2 Kij 2, 3 Kij 2, 4 Kij 2, 5 Kij 2, 6
Kij 3, 1 Kij 3, 2 Kij 3, 3 Kij 3, 4 Kij 3, 5 Kij 3, 6
Kij 4, 1 Kij 4, 2 Kij 4, 3 Kij 4, 4 Kij 4, 5 Kij 4, 6
Kij 5, 1 Kij 5, 2 Kij 5, 3 Kij 5, 4 Kij 5, 5 Kij 5, 6
Kij 6, 1 Kij 6, 2 Kij 6, 3 Kij 6, 4 Kij 6, 5 Kij 6, 6
Master stiffness modified for displacement B.C.:
1 0 0 0 0 0
0 1 0 0 0 0
0 0 Kij 3, 3 0 Kij 3, 5 Kij 3, 6
0 0 0 1 0 0
0 0 Kij 5, 3 0 Kij 5, 5 Kij 5, 6
0 0 Kij 6, 3 0 Kij 6, 5 Kij 6, 6
Force vector:
fi 1 , fi 2 , fi 3 , fi 4 , fi 5 , fi 6
Force vector modified for displacement B.C.:
0, 0, fi 3 , 0, fi 5 , fi 6

Figure 4.8. Modules ModifiedMasterStiffForDBC and ModifiedMasterForceForDBC

that modify the master stiffness matrix and force vector of a truss to impose displacement BCs.
Test statements (in blue) and test output.

The function clears appropriate rows and columns of K, places ones on the diagonal, and re-
turns the modified K as function value. The only slightly fancy thing in this module is the use
of the Mathematica function Length to extract the number of prescribed displacement compo-
nents: Length[pdof] here will return the value 3, which is the length of the list pdof. Similarly
nk=Length[K] assigns 6 to nk, which is the order of matrix K. Although for the example truss
these values are known a priori, the use of Length serves to illustrate a technique that is heavily
used in more general code.
Module ModifiedMasterForcesForDBC has similar structure and logic and need not be described
in detail. It is important to note that for homogeneous BCs the modules are independent of each other
and may be called in any order. On the other hand, if there were nonzero prescribed displacements
present the force modification must be done before the stiffness modification. This is because
stiffness coefficients that are cleared in the latter are needed for modifying the force vector.

4–12
 §4.7 RECOVERING INTERNAL FORCES

IntForce2DTwoNodeBar[{{x1_,y1_},{x2_,y2_}},{Em_,A_},eftab_,u_]:=
Module[ {c,s,dx=x2-x1,dy=y2-y1,L,ix,iy,jx,jy,ubar,e},
L=Sqrt[dx^2+dy^2]; c=dx/L; s=dy/L; {ix,iy,jx,jy}=eftab;
ubar={c*u[[ix]]+s*u[[iy]],-s*u[[ix]]+c*u[[iy]],
c*u[[jx]]+s*u[[jy]],-s*u[[jx]]+c*u[[jy]]};
e=(ubar[[3]]-ubar[[1]])/L; Return[Em*A*e]
];
p =IntForce2DTwoNodeBar[{{0,0},{10,10}},{100,2*Sqrt[2]},
{1,2,5,6},{0,0,0,0,0.4,-0.2}];
Print["Member int force (numerical):"]; Print[N[p]];
p =IntForce2DTwoNodeBar[{{0,0},{L,L}},{Em,A},
{1,2,5,6},{0,0,0,0,ux3,uy3}];
Print["Member int force (symbolic):"]; Print[Simplify[p]];

Member int force (numerical):

2.82843
Member int force (symbolic):
A Em ( ux3 + uy3)
2L

Figure 4.9. Module IntForce2DTwoNodeBar for computing the internal force in a

bar element. Test statements (in blue) and test output.

The test statements are purposedly chosen to illustrate another feature of Mathematica: the use of
the Array function to generate subscripted symbolic arrays of one and two dimensions. The test
output is shown in the bottom box of Figure 4.8, which should be self explanatory. The force vector
and its modified form are printed as row vectors to save space.

§4.7. Recovering Internal Forces

Since Mathematica provides built-in matrix operations for solving a linear system of equations and
multiplying matrices by vectors, we do not need to write application functions for the solution of
the modified stiffness equations, and for the recovery of nodal forces as f = Ku. Consequently, the
last application functions we need are those for internal force recovery.
Function IntForce2DTwoNodeBar listed in the top box of Figure 4.9 computes the internal
force in an individual bar element. It is somewhat similar in argument sequence and logic to
ElemStiff2DTwoNodeBar of Figure 4.5. The first two arguments are identical. Argument eftab
provides the Element Freedom Table array for the element. The last argument, u, is the vector of
computed node displacements.
The logic of IntForce2DTwoNodeBar is straightforward and follows the method outlined in §3.2.1.
Member joint displacements ū(e) in local coordinates {x̄, ȳ} are recovered in array ubar, then the
longitudinal strain e = (ū x j − ū xi )/L and the internal (axial) force p = E Ae is returned as function
value. As coded the function contains redundant operations because entries 2 and 4 of ubar (that
is, components ū yi and ū y j ) are not actually needed to get p, but were kept to illustrate the general
backtransformation of global to local displacements.
Running this function with the test statements shown after the module produces the output shown
in the bottom box of Figure 4.9. The first test is for member (3) of the example truss using the
actual nodal displacements (3.17). It also illustrates the use of the Mathematica built in function
N to produce output in floating-point form. The second test does a symbolic calculation in which
several argument values are fed in variable form.

4–13
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

IntForcesOfExampleTruss[u_]:= Module[{f=Table[0,{3}]},
f[[1]]=IntForce2DTwoNodeBar[{{0,0},{10,0}},{100,1},{1,2,3,4},u];
f[[2]]=IntForce2DTwoNodeBar[{{10,0},{10,10}},{100,1/2},{3,4,5,6},u];
f[[3]]=IntForce2DTwoNodeBar[{{0,0},{10,10}},{100,2*Sqrt[2]},
{1,2,5,6},u];
Return[f]
];
f=IntForcesOfExampleTruss[{0,0,0,0,0.4,-0.2}];
Print["Internal member forces in example truss:"];Print[N[f]];

Internal member forces in example truss:

{0., -1., 2.82843}

Figure 4.10. Module IntForceOfExampleTruss that computes internal forces

in the 3 members of the example truss. Test statements (in blue) and test output.

The top box of Figure 4.10 lists a higher-level function, IntForceOfExampleTruss, which has
a single argument: u. This is the complete 6-vector of joint displacements u. This function calls
IntForce2DTwoNodeBar three times, once for each member of the example truss, and returns the
three member internal forces thus computed as a 3-component list.
The test statements listed after IntForcesOfExampleTruss feed the displacement solution (3.24)
to the module. Running the test produces the output √
shown in the lower box of Figure 4.10. The
(1) (2) (3)
internal forces are F = 0, F = −1 and F = 2 2 = 2.82843, in agreement with the values
found in §3.4.2.

§4.8. Putting the Pieces Together

After all this development and testing effort, documented in Figures 4.5 through 4.10, we are ready
to make use of all these bits and pieces of code to analyze the example plane truss. This is actually
done with the logic shown in Figure 4.11. This particular piece of code is called the driver script.
Note that it is not a Module. It uses the seven previously described modules

ElemStiff2DTwoNodeBar
MergeElemIntoMasterStiff
AssembleMasterStiffOfExampleTruss
ModifiedMasterStiffForDBC
ModifiedMasterForcesForDBC
IntForce2DTwoNodeTruss
IntForcesOfExampleTruss (4.1)

These functions must have been defined (”compiled”) at the time the driver scripts described below
are run. A simple way to making sure that all of them are defined is to put all these functions in
the same Notebook file and to mark them as initialization cells. These cells may be executed by
picking up Kernel → Initialize → Execute Initialization.13 (An even simpler scheme would to
group them all in one cell, but that would make placing separate test statements messy.)
For a hierarchical version of (4.1), see the last CAETE slide.

13 In Mathematica versions 6 and higher, there is an initialization shortcut.

4–14
 §4.8 PUTTING THE PIECES TOGETHER

f={0,0,0,0,2,1};
K=AssembleMasterStiffOfExampleTruss[];
Kmod=ModifiedMasterStiffForDBC[{1,2,4},K];
fmod=ModifiedMasterForcesForDBC[{1,2,4},f];
u=Simplify[Inverse[Kmod].fmod];
Print["Computed nodal displacements:"]; Print[u];
f=Simplify[K.u];
Print["External node forces including reactions:"]; Print[f];
p=Simplify[IntForcesOfExampleTruss[u]];
Print["Internal member forces:"]; Print[p];

Computed nodal displacements:

{0, 0, 0, 0, 25 ,− 51 }
External node forces including reactions:
{-2, -2, 0, 1, 2, 1}
Internal member forces:
{0, -1, 2 2}

Figure 4.11. Driver script for numerical analysis of example truss and its output.

§4.8.1. The Driver Script

The code listed in the top box of Figure 4.11 first assembles the master stiffness matrix through
AssembleMasterStiffOfExampleTruss. Next, it applies the displacement boundary condi-
tions through ModifiedMasterStiffForDBC and ModifiedMasterForcesForDBC. Note that
the modified stiffness matrix is placed into Kmod rather than K to save the original form of the
master stiffness for the reaction force recovery later. The complete displacement vector is obtained
by the matrix calculation
u=Inverse[Kmod].fmod; (4.2)
This statement takes advantage of two built-in Mathematica functions. Inverse returns the inverse
of its matrix argument.14 The dot operator denotes matrix multiplication (here, matrix-vector
multiply.) The enclosing Simplify function in Figure 4.11 simplifies the expression of vector u
in case of symbolic calculations; it is actually redundant if all computations are numerical.
The remaining statements recover the node vector including reactions via the matrix-vector multiply
f = K.u (recall that K contains the unmodified master stiffness matrix) and the member internal
forces p through IntForcesOfExampleTruss. The code prints u, f and p as row vectors to save
space.
Running the script of the top box of Figure 4.11 produces the output shown in the bottom box of
that figure. The results confirm the hand calculations of Chapter 3.
§4.8.2. Is All of This Worthwhile?
At this point you may wonder whether all of this work is worth the trouble. After all, a hand
calculation (typically helped by a programable calculator) would be quicker in terms of flow time.
Typing and debugging the Mathematica fragments displayed here took the writer about six hours
(although about two thirds of this was spent in editing and getting the fragment listings into the
Chapter.) For larger problems, however, Mathematica would certainly beat hand-plus-calculator

14 This is a highly inefficient way to solve Ku = f if the linear system becomes large. It is done here for simplicity.

4–15
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

f={0,0,0,0,fx3,fy3};
K=AssembleMasterStiffOfExampleTruss[];
Kmod=ModifiedMasterStiffForDBC[{1,2,4},K];
fmod=ModifiedMasterForcesForDBC[{1,2,4},f];
u=Simplify[Inverse[Kmod].fmod];
Print["Computed nodal displacements:"]; Print[u];
f=Simplify[K.u];
Print["External node forces including reactions:"]; Print[f];
p=Simplify[IntForcesOfExampleTruss[u]];
Print["Internal member forces:"]; Print[p];

Computed nodal displacements:

1 1
{0, 0, 0, 0, 10 (3 fx3 - 2 fy3),
5
(-fx3 + fy3)}
External node forces including reactions:
{-fx3, -fx3, 0, fx3 - fy3, fx3, fy3}
Internal member forces:
{0, -fx3 + fy3, 2 fx3}

Figure 4.12. Driver script for symbolic analysis of example truss and its output.

computations, the cross-over typically appearing for 10 to 20 equations. For up to about 500
equations and using floating-point arithmetic, Mathematica gives answers within minutes on a fast
PC or Mac with sufficient memory but eventually runs out of steam at about 1000 equations. For
a range of 1000 to about 50000 equations, Matlab, using built-in sparse solvers, would be the best
compromise between human and computer flow time. Beyond 50000 equations a program in a
low-level language, such as C or Fortran, would be most efficient in terms of computer time.15
One distinct advantage of computer algebra systems emerges when you need to parametrize a
small problem by leaving one or more problem quantities as variables. For example suppose that
the applied forces on node 3 are to be left as f x3 and f y3 . You replace the last two components of array
p as shown in the top box of Figure 4.12, execute the cell and shortly get the symbolic answer shown
in the bottom box of that figure. This is the answer to an infinite number of numerical problems.
Although one may try to undertake such studies by hand, the likelyhood of errors grows rapidly with
the complexity of the system. Symbolic manipulation systems can amplify human abilities in this
regard, as long as the algebra “does not explode” because of combinatorial complexity. Examples
of such nontrivial calculations will appear throughout the following Chapters.
Remark 4.1. The “combinatorial explosion” danger of symbolic computations should be always
kept in mind. For example, the numerical inversion of a N × N matrix is a O(N 3 ) process,
whereas symbolic inversion goes as O(N !). For N = 48 the floating-point numerical inverse
will be typically done in a fraction of a second. But the symbolic adjoint will have 48! =
12413915592536072670862289047373375038521486354677760000000000 terms, or O(1061 ). There may
be enough electrons in our Universe to store that, but barely ...

15 The current record for FEM structural applications is about 100 million equations, done on a massively parallel super-
computer (ASCI Red at SNL). Fluid mechanics problems with over 500 million equations have been solved.

4–16
 §4. Notes and Bibliography

Notes and Bibliography

The hefty Mathematica Book [814] is a reference manual. Since the contents are available online (click on
Help in topbar), buying the printed book makes little sense.16
The best available tutorial to learn the basic (procedural) level of Mathematica is Schaum’s Outlines: Mathe-
matica, McGraw-Hill, New York, second edition 2009. Covers up to version 7.0. Contains 750 exercises with
answers. Can be inexpensively purchased from Amazon at www.amazon.com. Learning that level, which is
sufficient for this course, should take only a few hours.
There is a another nice tutorial by Glynn and Gray [310], list: $35, dated 1999. (Theodore Gray invented
the Notebook front-end that appeared in version 2.2.) It has the disadvantage of covering only up to fersion
3.0 and may be difficult to find as used book. More up to date and comprehensive is the recent appeared
Mathematica Navigator in two volumes [657,658]; list: $69.95 but used copies are discounted, down to about
$20. The recently appeared Mathematica Cookbook by S. Mangano, published by O’Reilly, is being reviewed
as of this writing; if useful, it will be placed in the reference list. This one covers more advaanced programmin
techniques that are not needed for this course.
Beyond these, there are many books at all levels that expound on the use of Mathematica for various applications
ranging from pure mathematics to physics and engineering. A web search on September 2003 done on
www3.addall.com hit 150 book titles containing Mathematica, compared to 111 for Maple and 148 for
Matlab. A google search (August 2005) hits 3,820,000 pages containing Mathematica, but here Matlab wins
with 4,130,000.
Wolfram Research hosts the MathWorld web site at http://mathworld.wolfram.com, maintained by Eric
W. Weisstein. It is essentially a hyperlinked, on-line version of his Encyclopœdia of Mathematics [796].
To close the topic of symbolic versus numerical computation, here is a nice summary by A. Grozin, posted on
the Usenet:
“Computer Algebra Systems (CASs) are programs [that] operate with formulas. Mathematica is a
powerful CAS (though quite expensive). Other CASs are, e.g., Maple, REDUCE, MuPAD <...>.
There are also quite powerful free CASs: Maxima and Axiom. In all of these systems, it is possible
to do some numerical calculations (e.g., to evaluate the formula you have derived at some numerical
values of all parameters). But it is a very bad idea to do large-scale numerical work in such systems:
performance will suffer. In some special cases (e.g., numerical calculations with very high precision,
impossible at the double-precision level), you can use Mathematica to do what you need, but there are
other, faster ways to do such things.
There are a number of programs to do numerical calculations with usual double-precision numbers.
One example is Matlab; there are similar free programs, e.g., Octave, Scilab, R, ... Matlab is very
good and fast in doing numerical linear algebra: if you want to solve a system of 100 linear equations
whose coefficients are all numbers, use Matlab; if coefficients contain letters (symbolic quantities)
and you want the solution as formulas, use Mathematica or some other CAS. Matlab can do a limited
amount of formula manipulations using its “symbolic toolbox,” which is an interface to a cut-down
Maple. It’s a pain to use this interface: if you want Maple, just use Maple.”

References
Referenced items have been moved to Appendix R.

16 The fifth edition, covering version 5, lists for $49.95 but older editions are heavily discounted on the web, some under
$2. There are no printed manual versions for versions 6 and higher.

4–17
Chapter 4: ANALYSIS OF EXAMPLE TRUSS BY A CAS

Homework Exercises for Chapter 4

Analysis of Example Truss by a CAS

Before doing any of these Exercises, download the Mathematica Notebook file ExampleTruss.nb from the
course web site. (Go to Chapter 4 Index and click on the link). Open this Notebook file using version 4.1
or later one. The first six cells contain the modules and test statements listed in the top boxes of Figures
4.5 through 4.10. These six are marked as initialization cells. Before running driver scripts, they should be
executed by picking up Kernel → Evaluation → Execute Initialization (or equivalent in newer versions).
Verify that the output of those six cells agrees with that shown in the bottom boxes of those figures. Then
execute the driver script in Cell 7 by clicking on the cell and pressing the appropriate key: <Enter> on a Mac,
<Shift-Enter> on a Windows PC. Compare the output with that shown in 4.11. If the output checks out, you
may proceed to the Exercises.

EXERCISE 4.1 [C:10] Explain why the Simplify command in the test statements of Figure 4.5 says L>0.
(One way to figure this out is to just say Ke=Simplify[Ke] and look at the output. Related question: why
does Mathematica refuse to simplify Sqrt[Lexp2] to L unless one specifies the sign of L in the Simplify
command?

EXERCISE 4.2 [C:10] Explain the logic of the For loops in the merge function MergeElemIntoMasterStiff
of Figure 4.6. What does the operator += do? (If you are a C programmer, all of this should be easy.)

EXERCISE 4.3 [C:10] Explain the reason behind the use of Length in the modules of Figure 4.6. Why not
simply set nk and np to 6 and 3, respectively?

EXERCISE 4.4 [C:15] Of the seven modules listed in Figures 4.5 through 4.10, with names collected in (4.1),
two can be used only for the example truss, three can be used for any plane truss, and two can be used for other
structures analyzed by the DSM. Identify which ones, and briefly state the reasons for your classification.

EXERCISE 4.5 [C:20] Modify the modules AssembleMasterStiffOfExampleTruss,

IntForcesOfExampleTruss, and the driver script of Figure 4.11, to solve numerically the three-node, two-
member truss of Exercise 3.6. Verify that the output reproduces the solution given for that problem. Procedural
recommendation: modify cells but keep a copy of the original Notebook handy in case things go wrong.

EXERCISE 4.6 [C:25] Expand the logic of ModifiedMasterForcesForDBC to permit specified nonzero
displacements. Specify these in a second argument called pval, which contains a list of prescribed values
paired with pdof.

xynode={{0,0},{10,0},{10,10}}; elenod={{1,2},{2,3},{3,1}};
unode={{0,0},{0,0},{2/5,-1/5}}; amp=5; p={};
For [t=0,t<=1,t=t+1/5,
For [e=1,e<=Length[elenod],e++, {i,j}=elenod[[e]];
xyi=xynode[[i]];ui=unode[[i]];xyj=xynode[[j]];uj=unode[[j]];
p=AppendTo[p,Graphics[Line[{xyi+amp*t*ui,xyj+amp*t*uj}]]];
];
];
Show[p,Axes->False,AspectRatio->Automatic];

Figure E4.1. Mystery scriptfor Exercise 4.7.

4–18
 Exercises

EXERCISE 4.7 [C:20] Explain what the program of Figure E4.1 does, and the logic behind what it does.
(You may want to put it in a cell and execute it.) What modifications would be needed so it can be used for
any plane struss?

4–19
 5
Constructing
MoM Members

5–1
Chapter 5: CONSTRUCTING MOM MEMBERS

TABLE OF CONTENTS

Page
§5.1 Introduction . . . . . . . . . . . . . . . . . . . . . 5–3
§5.2 Formulation of MoM Members . . . . . . . . . . . . . . 5–3
§5.2.1 What They Look Like . . . . . . . . . . . . . . 5–3
§5.2.2 End Quantities, Degrees of Freedom, Joint Forces . . . . . 5–4
§5.2.3 Internal Quantities . . . . . . . . . . . . . . . . 5–4
§5.2.4 Discrete Field Equations, Tonti Diagram . . . . . . . . 5–5
§5.3 Simplex MoM Members . . . . . . . . . . . . . . . . . 5–6
§5.3.1 The Bar Element Revisited . . . . . . . . . . . . 5–7
§5.3.2 The Spar Element . . . . . . . . . . . . . . . . 5–8
§5.3.3 The Shaft Element . . . . . . . . . . . . . . . 5–9
§5.4 *Non-Simplex MoM Members . . . . . . . . . . . . . . . 5–11
§5.4.1 *Formulation Rules . . . . . . . . . . . . . . . 5–11
§5.4.2 *Examples . . . . . . . . . . . . . . . . . . 5–12
§5. Notes and Bibliography . . . . . . . . . . . . . . . . . 5–13
§5. References . . . . . . . . . . . . . . . . . . . . . . 5–13
§5. Exercises . . . . . . . . . . . . . . . . . . . . . . 5–14

5–2
 §5.2 FORMULATION OF MOM MEMBERS

§5.1. Introduction

The truss member used as example in Chapters 2–4 is an instance of a structural element. Such ele-
ments may be formulated directly using concepts and modeling techniques developed in Mechanics
of Materials (MoM).1 The construction does not involve more advanced tools, such as variational
principles, which are required for the continuum-based finite elements that appear in Part II.
This Chapter presents an overview of the technique to construct the element stiffness equations of
“MoM members” using simple matrix operations. These simplified equations come in handy for
a surprisingly large number of applications, particularly in skeletal structures. Focus is on simplex
elements, which may be formed directly as a sequence of matrix operations. Non-simplex elements
are presented as a recipe because their proper formulation requires work theorems not yet studied.
The physical interpretation of the FEM is still emphasized. Consequently we continue to speak of
structures built up of members (elements) connected at joints (nodes).

§5.2. Formulation of MoM Members

§5.2.1. What They Look Like

MoM-based formulations are largely restricted to intrinsically one-dimensional members. These

are structural components one of whose dimensions, called longitudinal, is significantly larger than
the other two, called the transverse dimensions. Such members are amenable to the simplified
structural theories developed in MoM textbooks. This Chapter covers only straight members with
geometry defined by the two end joints.2 The member cross sections are defined by the intersection
of planes normal to the longitudinal dimension with the member. See Figure 5.1. Note that although
the individual member will be idealized as being one-dimensional in its intrinsic or local coordinate
system, it often functions as component of a two- or three-dimensional structure.
This class of structural components embodies bars, beams, beam-columns, shafts and spars. Al-
though geometrically similar, the names distinguish the main type of internal forces the member
resists and transmits: axial forces for bars, bending and shear forces for beams, axial compression
and bending for beam-columns, torsion forces for shafts, and shear forces for spars.
Members are connected at their end joints by displacement degrees of freedom. For truss (bar) and
spar members those freedoms are translational components of the joint displacements. For other
types, notably beams and shafts, nodal rotations are chosen as additional degrees of freedom.
Structures fabricated with MoM members are generally three-dimensional. Their geometry is de-
fined with respect to a global Cartesian coordinate system {x, y, z}. Two-dimensional idealizations
are useful simplifications should the nature of the geometry and loading allow the reduction of the
structural model to one plane of symmetry, which is chosen to be the {x, y} plane. Plane trusses
and plane frameworks are examples of such simplifications.

1 Mechanics of Materials was called Strength of Materials in older texts. It covers bars, beams, shafts, arches, thin plates
and shells, but only one-dimensional models are considered in introductory undergraduate courses. MoM involves ab
initio phenomenological assumptions such as “plane sections remain plane” or “shear effects can be neglected in thin
beams.” These came about as the byproduct of two centuries of structural engineering practice, justified by success.
A similar acronym (MOM) is used in Electrical Engineering for something completely different: the Method of Moments.
2 Advanced Mechanics of Materials includes curved members. Plane arch elements are studied in Chapter 13.

5–3
Chapter 5: CONSTRUCTING MOM MEMBERS

z ȳ
x̄

z̄
x y
Figure 5.1. A Mechanics of Materials (MoM) member is a structural element
one of whose dimensions (the longitudinal dimension) is significantly larger
than the other two. Local axes {x̄, ȳ, z̄} are chosen as indicated. Although
the depicted member is prismatic, some applications utilize tapered or stepped
members, the cross section of which varies as a function of x̄.

In this Chapter we study generic structural members that fit the preceding class. An individual
member is identified by e but this superscript will be usually suppressed in the equations below to
reduce clutter. The local axes are denoted by {x̄, ȳ, z̄}, with x̄ along the longitudinal direction. See
Figure 5.1.
The mathematical model of a MoM member is obtained by an idealization process. The model
represents the member as a line segment that connects the two end joints, as depicted in Figure 5.2.
§5.2.2. End Quantities, Degrees of Freedom, Joint Forces
The set of mathematical variables used to link members are called end quantities or connectors. In
the Direct Stiffness Method (DSM) these are joint displacements (the degrees of freedom) and the
joint forces. These quantities are related by the member stiffness equations.
The degrees of freedom at the end joints i and j are collected in the joint displacement vector ū.
This may include translations only, rotations only, or a combination of translations and rotations.
The vector of joint forces f̄ groups components in one to one correspondence with ū. Component
pairs must be conjugate in the sense of the Principle of Virtual Work. For example if the x̄-translation
at joint i: ū xi appears in ū, the corresponding entry in f̄ is the x̄-force f¯xi at i. If the rotation about
z̄ at joint j: θ̄z j appears in ū, the corresponding entry in f̄ is the z-moment m̄ z j .
§5.2.3. Internal Quantities
Internal quantities are mechanical actions that take place within the member. Those actions involve
stresses and deformations. Accordingly two types of internal quantities appear:
Internal member forces form a finite set of stress-resultant quantities collected in an array p. They
are obtained by integrating stresses over each cross section, and thus are also called generalized
stresses in structural mechanics. This set characterizes the forces resisted by the material. Stresses
at any point in a section may be recovered if p is known.
Member deformations form a finite set of quantities, chosen in one-to one correspondence with
internal member forces, and collected in an array v. This set characterizes the deformations expe-
rienced by the material. They are also called generalized strains in structural theory. Strains at any
point in the member can be recovered if v is known.
As in the case of end quantities, internal forces and deformations are paired in one to one corre-
spondence. For example, the axial force in a bar member must be paired either with an average

5–4
 §5.2 FORMULATION OF MOM MEMBERS

End quantities (connectors)

are defined at the joints i, j
ȳ
z (e) x̄
j
i

z̄ Internal quantities are

y defined within the member
x
and may depend on x. ¯

Figure 5.2. The FE mathematical idealization of a MoM member. The

model is one-dimensional in x̄. The two end joints are the site of end quantities:
joint forces and displacements, that interconnect members. Internal quantities
characterize internal forces, stresses and deformations in the member.

axial deformation, or with the total elongation. Pairs that mutually correspond in the sense of the
Principle of Virtual Work are called conjugate. Unlike the case of end quantities, conjugacy of
internal quantities is not a mandatory requirement although it simplifies some expressions.
§5.2.4. Discrete Field Equations, Tonti Diagram
The matrix equations that connect ū, v, p and f̄ are called the discrete field equations. There are
three of them.
The member deformations v are linked to the joint displacements ū by the kinematic compatibility
conditions, also called the deformation-displacement or strain-displacement equations:

v = B ū. (5.1)
The internal member forces are linked to the member deformations by the constitutive equations.
In the absence of initial strain effects those equations are homogeneous:

p = S v. (5.2)

Finally, the internal member forces are linked to the joint forces by the equilibrium equations. If
the internal forces p are constant over the member, the relation is simply

f̄ = AT p. (5.3)

In (5.3) the transpose of A is used for convenience.3

The foregoing equations can be presented graphically as shown in Figure 5.3. This is a discrete
version of the so-called Tonti diagrams, which represent governing equations as arrows linking
boxes containing kinematic and static quantities. Tonti diagrams for field (continuum) equations
are introduced in Chapter 11.

3 If p is a function of x̄ the relation is of differential type. This form is studied in §5.4.

5–5
Chapter 5: CONSTRUCTING MOM MEMBERS

Stiffness
¯f = AT S B u = K
¯ u¯
ū f̄

Kinematic Equilibrium
v = B u¯ f̄ = AT p

p=Sv
v p
Constitutive
Figure 5.3. Tonti diagram of the three discrete field equations (5.1)–(5.3) and the stiffness
equation (5.4) fora simplex MoM member. Internal and end quantities appear inside the orange
and yellow boxes, respectively.

Matrices B, S and A receive the following names in the literature:

A Equilibrium, leverage
S Rigidity, material, constitutive4
B Compatibility, deformation-displacement, strain-displacement
If the element is sufficiently simple, the determination of these three matrices can be carried out
through MoM techniques. If the construction requires more advanced tools, however, recourse to
the general methodology of finite elements and variational principles is necessary.
§5.3. Simplex MoM Members
Throughout this section we assume that the internal quantities are constant over the member length.
Such members are called simplex elements. If so the matrices A, B and S are independent of member
cross section. For simplex elements the derivation of the element stiffness equations reduces to a
straightforward sequence of matrix multiplications.
Under the constancy-along-x̄ assumption, elimination of the interior quantities p and v from (5.1)
through (5.3) yields the element stiffness relation
f̄ = AT S B ū = K̄ ū, (5.4)
whence the element stiffness matrix is
K̄ = AT S B. (5.5)

If both pairs: {p, v} and {f̄, ū}, are conjugate in the sense of the Principle of Virtual Work, it can be
shown that A = B and that S is symmetric. In that case

K̄ = BT SB. (5.6)
is a symmetric matrix. Symmetry is computationally desirable for reasons outlined in Part III.

4 The name rigidity matrix for S is preferable. It is a member integrated version of the cross section constitutive equations.
The latter are usually denoted by symbol R, as in §5.4.

5–6
 §5.3 SIMPLEX MOM MEMBERS

(a) (b)
ȳ Axial rigidity EA, length L
z −F ȳ
EA
f¯xi , ūxi i (e) j f¯xj , ūxj
x̄
z̄ F
x̄ L
x y

Figure 5.4. The prismatic bar (also called truss) member: (a) individual member shown
in 3D space, (b) idealization as generic member.

Remark 5.1. If f̄ and ū are conjugate (as required in §5..2.2) but p and v are not, K̄ must come out to be
symmetric even if S is unsymmetric and A = B. However there are more opportunities to go wrong.

§5.3.1. The Bar Element Revisited

The simplest MoM element is the prismatic bar or truss member already derived in Chapter 2.
See Figure 5.4. This qualifies as simplex because all internal quantities are constant. One minor
difference in the derivation below is that the joint displacements and forces in the ȳ direction are
omitted in the generic element because they contribute nothing to the stiffness equations. In the
FEM terminology, freedoms associated with zero stiffness are called inactive. Three choices for
internal deformation and force variables are considered next. The results confirm that the element
stiffness equations coalesce, as expected, since the end quantities f̄ and ū stay the same.
Derivation Using Axial Elongation and Axial Force. The member axial elongation d is taken as
deformation measure, and the member axial force F as internal force measure. Hence v and p
reduce to the scalars ≡ d and p ≡ F, respectively. The chain of discrete field equations is easily
constructed:
     
ū xi EA f¯xi −1
d = [ −1 1 ] = Bū, F= d = S d, f̄ = ¯ = F = AT F.
ū x j L fx j 1
(5.7)
Consequently
   
−1 EA 1 −1
K̄ = A SB = SB B = S [ −1 1 ]
T T
= . (5.8)
1 L −1 1

Note that A = B because F and d are conjugate: F d is work. The foregoing equations can be
represented graphyically with the discrete Tonti diagram of Figure 5.5.
Derivation Using Mean Axial Strain and Axial Force. Instead of d we may use the mean axial strain
ē = d/L as deformation measure whereas F is kept as internal force measure. The only change
is that B becomes [ −1 1 ] /L whereas S becomes E A. Matrix A does not change. The product
AT SB gives the same K̄ as in (5.8), as can be expected. Now AT is not equal to B because F and
ē are not conjugate, but they differ only by a factor 1/L.
Derivation Using Mean Axial Strain and Axial Stress. We keep the mean axial strain ē = d/L
as deformation measure but take the mean axial stress σ̄ = F/A (which is not conjugate to ē) as

5–7
Chapter 5: CONSTRUCTING MOM MEMBERS

Stiffness
_
EA 1 −1 _
f= u
_ L −1 1 _
u f
Kinematic Equilibrium
ū xi f¯xi −1
d= −1 1 = Bū f̄ = = F = AT F
ū x j f¯x j 1
EA
F= d = Sd
d L
F
Constitutive

Figure 5.5. Tonti diagram for the bar element discrete equations (5.7)–(5.8).

internal force measure. Now B = [ −1 1 ] /L, S = E and AT = A [ −1 1 ]. The product AT SB

gives again the same K̄. as shown in (5.8).
Transformation to Global Coordinates. Since ū yi and ū y j are not part of (5.7) and (5.8) the
displacement transformation matrix from local to global {x, y} coordinates is 2 × 4, instead of 4 × 4
as in §2.8.1. On restoring the element identifier e the appropriate local-to-global transformation is
 
u exi
   
ū exi c s 0 0  e 
 u yi  = Te ue ,
ūe = = (5.9)
ū ex j 0 0 c s  u ex j 
u ey j

in which c = cos ϕ e , s = sin ϕ e , and ϕ e is the angle from x to x̄, cf. Figure 2.10 of Chapter 2. The
e
4 × 4 globalized element stiffness matrix Ke = (Te )T K̄ Te agrees with (2.18). The extension of
(5.9) to 3D is handled as an Exercise.

§5.3.2. The Spar Element

The spar or shear-web member has two joints (end nodes): i and j. This member can only resist
and transmit a constant shear force V in the plane of the web, which is chosen to be the {x̄, ȳ} plane.
See Figure 5.6. It is often used for modeling high-aspect aircraft wing structures, as illustrated in
Figure 5.7. We consider here only prismatic spar members of uniform material and constant cross
section, which thus qualify as simplex.
The active degrees of freedom for a generic spar member of length L, as depicted in Figure 5.6(b),
are ū yi and ū y j . Let G be the shear modulus and As the effective shear area. The latter is a concept
developed in Mechanics of Materials; for a narrow rectangular cross section, As = 5A/6.
The shear rigidity is G As . As deformation measure the mean shear strain γ = V /(G As ) is chosen.
The kinematic, constitutive, and equilibrium equations are
     
1 ū yi f¯yi −1
γ = [ −1 1] = Bū, V = G As γ = Sγ , f̄ = = V = AT V,
L ū y j f¯y j 1
(5.10)

5–8
 §5.3 SIMPLEX MOM MEMBERS

V
(a) ȳ (b)
x̄ ȳ
z
f¯yi , ūyi GA f¯yj , ū yj
s
z̄ i (e) j
x̄
Shear rigidity GA s , length L
x y −V L

Figure 5.6. The prismatic spar (also called shear-web) member: (a) individual member
shown in 3D space, (b) idealization as generic member in local system.

Note that A = B because V and γ are not work-conjugate. (This difference is easily adjusted for,
however; see Exercise 5.1.) The local stiffness equations follow as
    
f¯yi G As 1 −1 ū yi
f̄ = = A S Bū =
T
= K̄ ū. (5.11)
f¯y j L −1 1 ū y j

If the spar member is used in a two dimensional context, the displacement transformation from
local to global coordinates {x, y} is

 
u exi
   
ū eyi −s c 0 0  e 
 u yi  = Te ue ,
ū =
e
= (5.12)
ū ey j 0 0 −s c  u ex j 
u ey j
in which c = cos ϕ e , s = sin ϕ e , and ϕ e is the angle from x to x̄, cf. Figure 2.10 of Chapter 2. The
e
4 × 4 globalized spar stiffness matrix is then Ke = (Te )T K̄ Te . The full expression is worked out
in Exercise 5.2.
More often, however, the spar member will be a component in a three-dimensional structural model,
e.g. the aircraft wing shown in Figure 5.7. If so the determination of the 2 × 6 transformation matrix
is more involved, as an “orientation node” is required. This is the topic of Exercise 5.5.
§5.3.3. The Shaft Element

The shaft, also called torque member, has two joints (end nodes): i and j. A shaft can only resist
and transmit a constant torque or twisting moment T along its longitudinal axis x̄, as pictured in
Figure 5.8(a).
We consider here only prismatic shaft members with uniform material and constant cross section,
which thus qualify as simplex. The active degrees of freedom of a generic shaft member of length L,
depicted in Figure 5.8(b), are θ̄xi and θ̄x j . These are the infinitesimal end rotations about x̄, positive
according to the right-hand rule. The associated joint (node) forces are end moments denoted as
m̄ xi and m̄ x j .
The only internal force is the torque T , positive if acting as pictured in Figure 5.8(a). Let G be the

5–9
Chapter 5: CONSTRUCTING MOM MEMBERS

COVER
PLATES

SPAR
RIB

Figure 5.7. Spar members in aircraft wing (Piper Cherokee). For more
impressive aircraft structures see CAETE slides.

(a) Torsional rigidity GJ, length L

ȳ x̄
(b)
_ _
z T m xj , θxj
ȳ GJ
_ _ T
m xi , θxi m̄xi, θ¯xi i (e) j mxj, θ¯xj x̄

x y z̄ L

Figure 5.8. The prismatic shaft (also called torque member): (a) individual member
shown in 3D space, (b) idealization as generic member in the local system.

shear modulus and G J the effective torsional rigidity.5 As deformation measure pick the relative
twist angle φ = θ̄x j − θ̄xi . The kinematic, constitutive, and equilibrium equations provided by
Mechanics of Materials are
     
θ̄xi GJ m̄ xi −1
φ = [ −1 1 ] = Bū, T = φ = Sγ , f̄ = = T = BT T. (5.13)
θ̄x j L m̄ x j 1

From these the local stiffness equations follow as

    
m̄ xi GJ 1 −1 θ̄xi
f̄ = = B S Bū =
T
= K̄ ū. (5.14)
m̄ x j L −1 1 θ̄x j

If the shaft is used in a two-dimensional context, the displacement transformation to global coordi-

5 J has dimension of (length)4 . For a circular or annular cross section it reduces to the polar moment of inertia about x̄.
The determination of J for noncircular cross sections is covered in Mechanics of Materials textbooks.

5–10
 §5.4 *NON-SIMPLEX MOM MEMBERS

nates {x, y} within the framework of infinitesimal rotations, is

 e 
θxi
 e   
θ̄ 
c s 0 0  θ yie   = Te θe ,
ūe = xi = (5.15)
θ̄xej 0 0 c s  θxej 
θ yej
where as usual c = cos ϕ e , s = sin ϕ e , and ϕ e is the angle from x to x̄. Note that θe collects only
global node rotations components. This operation is elaborated further in Exercise 5.3.
§5.4. *Non-Simplex MoM Members
The straightforward formulation of simplex MoM elements does not immediately carry over to the case in
which internal quantities p and v vary over the member; that is, depend on x̄. The dependence may be due to
element type, varying cross section, or both. As a result, one or more of the matrices A, B and S depend on
x̄. Such members are called non-simplex.
e
The matrix multiplication rule (5.4) cannot be used to construct the element stiffness matrix K̄ of non-simplex
e
members. This can be grasped by observing that A(x̄)T S(x̄)B(x̄) would depend on x̄. On the other hand, K̄
must be independent of x̄ because it relates the end quantities ū and f̄.
§5.4.1. *Formulation Rules
The derivation of non-simplex MoM elements requires use of the work principles of mechanics, for example
the Principle of Virtual Work or PVW. Thus, more care must be exercised in the choice of conjugate internal
quantities. The following rules can be justified through the arguments presented in Part II. They are stated
here as recipe, and apply only to displacement-assumed elements.
Rule 1. Select internal deformations v(x̄) and internal forces p(x̄) that are conjugate in the PVW sense. Link
deformations to node displacements by v(x̄) = B(x̄)u.
Rule 2. From the PVW it may be shown (see Remark 5.2 below) that the force equilibrium equation exists
only in a differential sense:
BT dp = d f̄. (5.16)
Here d in dp denotes differentiation with respect to x̄. The meaning of dp is simply p(x̄) d x̄. That is, the
differential of internal forces as one passes from cross-section x̄ to a neighboring one x̄ +d x̄. The interpretation
of d f̄ is less immediate because f̄ is not a function of x̄. It actually means the contribution of that member
slice to the building of the node force vector f̄. See (5.18) and (5.19) below.
Rule 3. The constitutive relation is
p = Rv, (5.17)
in which R, which may depend on x̄, must be symmetric. Note that symbol R in (5.17) replaces the S of (5.2).
Matrix R pertains to a specific cross section whereas S applies to the entire member. This distinction is further
elaborated in Exercise 5.9.
The discrete relations supplied by the foregoing rules are displayed in the discrete Tonti diagram of Figure 5.9.
Internal quantities are now eliminated starting from the differential equilibrium relation (5.16):

d f̄ = BT dp = BT p d x̄ = BT R v d x̄ = BT R B ū d x̄ = BT R B d x̄ ū. (5.18)
Integrating both sides over the member length L yields

L
L
f̄ = d f̄ = BT R B d x̄ ū = K̄ ū, (5.19)
0 0

5–11
Chapter 5: CONSTRUCTING MOM MEMBERS

Stiffness
L
f¯ = BT R B dx¯ u
¯
0
ū f̄

Kinematic Equilibrium
(at each x-section) (at each x-section)
v = B u¯ d f¯ = BT dp

p=Rv
v p
Constitutive
(at each x-section)
Figure 5.9. Discrete Tonti diagram of the equations for a non-simplex MoM
member.

because ū does not depend on x̄. Consequently the local element stiffness matrix is

L
K̄ = BT R B d x̄ (5.20)
0

The recipe (5.20) will be justified in Part II through energy methods. It will be seen that it generalizes to
arbitrary displacement-assumed finite elements in any number of space dimensions. It is used in the derivation
of the stiffness equations of the plane beam element in Chapter 12. The reduction of (5.20) to (5.6) when the
dependence on x̄ disappears is the subject of Exercise 5.8.

Remark 5.2. The proof of (5.16) follows by equating expressions of the virtual work of a slice of length d x̄
T
undergoing virtual node displacements δ ū and associated deformations δv: d f̄ δ ū = dpT δv = dpT (B ū) =
(B dp)T δ ū. Since δ ū is arbitrary, B dp = d f̄.
T T

§5.4.2. *Examples
Example 5.1. A two-node bar element has constant elastic modulus E but a continuously varying area: Ai ,
A j and Am at i, j and m, respectively, where m is the midpoint between end joints i and j. This variation can
be fitted by
A(x̄) = Ai Ni (x̄) + A j N j (x̄) + Am Nm (x̄). (5.21)
Here Ni (x̄) = − 12 ξ(1 − ξ ), N j (x̄) = 12 ξ(1 + ξ ) and Nm (x̄) = 1 − ξ 2 , with ξ = 2x/L − 1, are interpolating
polynomials further studied in Part II as “element shape functions.”
As internal quantities take the strain e and the axial force p = E Ae, which are conjugate quantities. Assuming
the strain e to be uniform over the element (this is characteristic of a displacement assumed element and is
justified through the method of shape functions explained in Part II.) the MoM equations are
1
e = Bū, p = E A(x̄) e = R(x̄) e, d f̄ = BT dp, B= [ −1 1]. (5.22)
L
Inserting into (5.20) and carrying out the integration yields

E Ā  1 −1

K̄ = , with Ā = 16 (Ai + A j ) + 23 Am . (5.23)
L −1 1

5–12
 §5. Notes and Bibliography

Example 5.2. Same as in the previous case but now the strain e is taken to be e = p/(E A), whereas the axial
force p is constant and defined by p = − f¯xi = f¯x j . The integrals become rational functions of x̄ and are best
evaluated through Mathematica. The completion of this Example is the matter of an Exercise.

Notes and Bibliography

The derivation of MoM elements using straightforward matrix algebra is typical of pre-1962 Matrix Structural
Analysis (MSA). The excellent book of Pestel and Leckie [578], unfortunately out of print, epitomizes that
approach. Historically this idea interweaved with Generation 1 of FEM, as outlined in Appendix O. By 1970
simplified derivations had fallen out of favor as yokelish. But these elements do not need improvement. They
still work fine: a bar or beam stiffness today is the same as 40 years ago.6
The Mechanics of Materials books by Beer-Johnston [68] and Popov [595] may be cited as being widely used
in US undergraduate courses. But they are not the only ones. A September 2003 in-print book search through
www3.addall.com on “Mechanics of Materials” returns 99 hits whereas one on “Strength of Materials” (the
older name) compiles 112. Folding multiple editions and hardback/paperback variants one gets about 60
books; by all accounts an impressive number.
Spar members are discussed only in MoM books that focus on aircraft structures, since they are primarily used
in modeling shear web action. On the other hand, bars, shafts and beams are standard fare.
The framework presented here is a tiny part of MSA. A panoramic view, including linkage to continuum
formulations from the MSA viewpoint, is presented in [232].
The source of Tonti diagrams is discussed in Chapter 11.

References
Referenced items have been moved to Appendix R.

6 The chief technical difference is the heavier use of differential equations prior to 1962, as opposed to the energy methods
in vogue today. The end result for simple one-dimensional models is the same.

5–13
Chapter 5: CONSTRUCTING MOM MEMBERS

Homework Exercises for Chapter 5

Constructing MoM Members

=== check (5.14)

EXERCISE 5.1 [A:10] Explain how to select the deformation variable v (paired to V ) of the spar member
formulated in §5.3.2, so that A = B. Draw the Tonti diagram with the discrete equations for that choice of v
and p, using Figure 5.5 as guide (that is, with the actual matrix equations along the arrows).

EXERCISE 5.2 [A:15] Obtain the 4 × 4 global element stiffness matrix of a prismatic spar member in a two
dimensional Cartesian system {x, y}. Start from (5.11). Indicate where the transformation (5.12) comes from
e
(Hint: read §2.8). Evaluate Ke = (Te )T K̄ Te in closed form.

EXERCISE 5.3 [A:15] Obtain the 4 × 4 global element stiffness matrix of a prismatic shaft element in a
two dimensional Cartesian system {x, y}. Include only node rotation freedoms in the global displacement
vector. Start from (5.14). Justify the transformation (5.15) (Hint: infinitesimal rotations transform as vectors).
e
Evaluate Ke = (Te )T K̄ Te in closed form.

j (x j , y j , zj ) −
uxj
(a)
E, A x− (b)
−
x
j
−
y
i (x i , yi , z i ) −
uxi L −y
L
y −
z i

z x

Figure E5.1. Bar element in 3D for Exercise 5.4.

EXERCISE 5.4 [A+N:15(10+5)] A bar element moving in three dimensional space is completely defined
by the global coordinates {xi , yi , z i }, {x j , y j , z j } of its end nodes i and j, as illustrated in Figure E5.1.
The 2 × 6 displacement transformation matrix T, with superscript e dropped for brevity, links ūe = Tue .
Here ūe contains the two local displacements ū xi and ū x j whereas ue contains the six global displacements
u xi , u yi , u zi , u x j , u y j , u z j .
(a) From vector mechanics show that

1  x ji y ji z ji 0 0 0
 
cx ji c y ji cz ji 0 0 0

T= = (E5.1)
L 0 0 0 x ji y ji z ji 0 0 0 cx ji c y ji cz ji

in which L is the element length, x ji = x j − xi , etc., and cx ji = x ji /L, etc., are the direction cosines of
the vector going from i to j.
(b) Evaluate T for a bar going from node i at {1, 2, 3} to node j at {3, 8, 6}.

EXERCISE 5.5 [A+N:30(10+15+5)] A spar element in three dimensional space is only partially defined by
the global coordinates {xi , yi , z i }, {x j , y j , z j } of its end nodes i and j, as illustrated in Figure E5.2. The
problem is that axis ȳ, which defines the direction of shear force transmission, is not uniquely defined by i

5–14
 Exercises

Orientation node k defines plane (b)

(a) − − as passing through {i,j,k}
{x,y} k
k (xk , yk , z k) −
uyj j (x j , y j , z j) x−
a
G, A s x− j
h β
b
u−yi
−
y y− α m L−c
i (x i , yi , z i) −
uxi L
c
− i
y z

z x

Figure E5.2. Spar element in 3D for Exercise 5.5.

and j.7 Most FEM programs use the orientation node method to complete the definition. A third node k, not
colinear with i and j, is provided by the user. Nodes {i, j, k} define the {x̄, ȳ} plane and thus z̄. The projection
of k on line i j is point m. The distance h > 0 from m to k is called h as shown in Figure E5.2(b). The
2 × 6 displacement transformation matrix T, with superscript e omitted to reduce clutter, relates ūe = Tue .
Here ūe contains the local transverse displacements ū yi and ū y j whereas ue contains the global displacements
u xi , u yi , u zi , u x j , u y j , u z j .
(a) Show that

1  xkm ykm z km 0 0 0
 
cxkm c ykm czkm 0 0 0

T= = (E5.2)
h 0 0 0 xkm ykm z km 0 0 0 cxkm c ykm czkm

in which xkm = xk − xm , etc., and cxkm = xkm / h, etc., are the direction cosines of the vector going from
m to k. (Assume that the position of m is known. That computation is carried out in the next item.)
(b) Work out the formulas to compute the coordinates of point m in terms of the coordinates of {i, j, k}.
Assume a, b and L are computed immediately from the input data. Using
the notation of Figure E5.2(b)
and elementary trigonometry, show that h = 2A/L, where A = p( p − a)( p − b)( p − L) with
p = 2 (L + a + b) (Heron’s formula), cos α = (L + b − a )/(2bL), cos β = (L 2 + a 2 − b2 )/(2a L),
1 2 2 2

c = b cos α, L − c = a cos β, xm = xi (L − c)/L + x j c/L, etc.8

(c) Evaluate T for a spar member going from node i at {1, 2, 3} to node j at {3, 8, 6}. with k at {4, 5, 6}.

EXERCISE 5.6 [A:20] Explain how thermal effects can be generally incorporated in the constitutive equation
(5.2) to produce an initial force vector for a simplex element.

EXERCISE 5.7 [A:15] Draw the discrete Tonti diagram for the prismatic shaft element. Use 5.5 as a guide
(that is, with the actual matrix equations along the arrows).

EXERCISE 5.8 [A:15] If the matrices B and R are constant over the element length L, show that expression
(5.20) of the element stiffness matrix for a non-simplex member reduces to (5.6), in which S = LR.

7 The same ambiguity arises in beam elements in 3D space. These elements are covered in Part III.
8 An alternative and more elegant procedure, found by a student in 1999, can be sketched as follows. From Figure E5.2(b)
obviously the two subtriangles imk and jkm are right-angled at m and share side km of length h. Apply Pythagoras’
theorem twice, and subtract so as to cancel out h 2 and c2 , getting a linear equation for c that can be solved directly.

5–15
Chapter 5: CONSTRUCTING MOM MEMBERS

EXERCISE 5.9 [A:20] Explain in detail the quickie derivation of footnote 6. (Knowledge of the Principle of
Virtual Work is required to do this exercise.)

EXERCISE 5.10 [A:25(10+5+10)] Consider a non-simplex element in which R varies with x̄ but B = A is
constant.
(a) From (5.20) prove that

L
1
K̄ = LB R̄ B, T
with R̄ = R(x̄) d x̄ (E5.3)
L 0

(b) Apply (E5.3) to obtain K̄ for a tapered bar with area defined by the linear law A = Ai (1− x̄/L)+ A j x̄/L,
where Ai and A j are the end areas at i and j, respectively. Take B = [ −1 1 ] /L.
(c) Apply (E5.3) to verify the result (5.23) for a bar with parabolically varying cross section.

EXERCISE 5.11 [A/C+N:30(25+5)] A prismatic bar element in 3D space is referred to a global coordinate
system {x, y, z}, as in Figure E5.1. The end nodes are located at {x1 , y1 , z 1 } and {x2 , y2 , z 2 }.9 The elastic
modulus E and the cross section area A are constant along the length. Denote x21 = x2 − x1 , y21 = y2 − y1 ,
z 21 = z 2 − z 1 and L = x21
2
+ y21
2
+ z 21
2
.
(a) Show that the element stiffness matrix in global coordinates can be compactly written10 A plodding way
is to start from the local stiffness (5.8) and transform to global using (?)

EA T
Ke = B B, in which B = [ −x21 −y21 −z 21 x21 y21 z 21 ] . (E5.4)
L3

(b) Compute Ke if the nodes are at {1, 2, 3} and {3, 8, 6}, with elastic modulus E = 343 and cross section
area A = 1. Note: the computation can be either done by hand or with the help of a program such as the
following Mathematica module, which is used in Part III:

Stiffness3DBar[ncoor_,mprop_,fprop_,opt_]:= Module[
{x1,x2,y1,y2,z1,z2,x21,y21,z21,Em,Gm,rho,alpha,A,
num,L,LL,LLL,B,Ke}, {{x1,y1,z1},{x2,y2,z2}}=ncoor;
{x21,y21,z21}={x2-x1,y2-y1,z2-z1};
{Em,Gm,rho,alpha}=mprop; {A}=fprop; {num}=opt;
If [num,{x21,y21,z21,Em,A}=N[{x21,y21,z21,Em,A}]];
LL=x21^2+y21^2+z21^2; L=PowerExpand[Sqrt[LL]];
LLL=Simplify[LL*L]; B={{-x21,-y21,-z21,x21,y21,z21}};
Ke=(Em*A/LLL)*Transpose[B].B;

9 End nodes are labeled 1 and 2 instead of i and j to agree with the code listed below.
10 There are several ways of arriving at this result. Some are faster and more elegant than others. Here is a sketch of one of
the ways. Denote by L 0 and L the lengths of the bar in the undeformed and deformed configurations, respectively. Then

2 (L − L 20 ) = x21 (u x2 − u x1 ) + y21 (u y2 − u y1 ) + z 21 (u z2 − u z1 ) + Q ≈ Bu,

1 2

in which Q is a quadratic function of node displacements which is therefore dropped in the small-displacement linear
theory. Also on account of small displacements

2 (L − L 20 ) = 12 (L + L 0 )(L − L 0 ) ≈ L L .
1 2

Hence the small axial strain is e = L/L = (1/L 2 )Bue , which begins the Tonti diagram. Next is F = E A e. Finally
you must show that force equilibrium at nodes requires fe = (1/L)BT F. Multiplying through gives (E5.4).

5–16
 Exercises

Return[Ke]];

ClearAll[Em,A]; Em=343; A=1;

ncoor={{0,0,0},{2,6,3}}; mprop={Em,0,0,0}; fprop={A}; opt={False};
Ke=Stiffness3DBar[ncoor,mprop,fprop,opt];
Print["Stiffness of 3D Bar Element:"];
Print[Ke//MatrixForm];
Print["eigs of Ke: ",Eigenvalues[Ke]];

As a check, the six eigenvalues of this particular Ke should be 98 and five zeros.

EXERCISE 5.12 [A/C:25] Complete Example 5.2.

5–17
 6
FEM Modeling:
Introduction

6–1
Chapter 6: FEM MODELING: INTRODUCTION

TABLE OF CONTENTS

Page
§6.1. Introduction 6–3
§6.2. FEM Terminology 6–3
§6.3. Idealization 6–4
§6.3.1. Models . . . . . . . . . . . . . . . . . . . 6–4
§6.3.2. Mathematical Models . . . . . . . . . . . . . . 6–5
§6.3.3. Implicit vs. Explicit Modeling . . . . . . . . . . . . 6–6
§6.4. Discretization 6–7
§6.4.1. Analytical or Numerical? . . . . . . . . . . . . . 6–7
§6.4.2. Error Sources and Approximation . . . . . . . . . . 6–7
§6.4.3. Other Discretization Methods . . . . . . . . . . . 6–8
§6.5. The Finite Element Method 6–8
§6.6. Element Attributes 6–9
§6.6.1. Element Dimensionality . . . . . . . . . . . . . . 6–9
§6.6.2. Element Nodes . . . . . . . . . . . . . . . . 6–10
§6.6.3. Element Geometry . . . . . . . . . . . . . . . 6–10
§6.6.4. Element Degrees of Freedom . . . . . . . . . . . . 6–10
§6.6.5. Nodal Forces . . . . . . . . . . . . . . . . . 6–10
§6.6.6. Element Constitutive Properties . . . . . . . . . . . 6–10
§6.6.7. Element Fabrication Properties . . . . . . . . . . . 6–10
§6.7. Classification of Mechanical Elements 6–11
§6.7.1. Primitive Structural Elements . . . . . . . . . . . 6–11
§6.7.2. Continuum Elements . . . . . . . . . . . . . . . 6–11
§6.7.3. Special Elements . . . . . . . . . . . . . . . . 6–12
§6.7.4. Macroelements . . . . . . . . . . . . . . . . . 6–12
§6.7.5. Substructures . . . . . . . . . . . . . . . . . 6–13
§6.8. Assembly 6–13
§6.9. Boundary Conditions 6–13
§6.9.1. Essential and Natural B.C. . . . . . . . . . . . . . 6–13
§6.9.2. B.C. in Structural Problems . . . . . . . . . . . . 6–14
§6. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 6–14
§6. References . . . . . . . . . . . . . . . . . . . . . . 6–15

6–2
 §6.2 FEM TERMINOLOGY

§6.1. Introduction

Chapters 2 through 5 cover material technically known as Matrix Structural Analysis or MSA. As
chronicled in Appendix H, this is a subject that historically preceded the Finite Element Method
(FEM), although it contributed substantially to it.
This Chapter begins covering the FEM proper. This is distinguished from MSA by three traits:
• The ability to go beyond structures.
• The increasing importance of continuum models in two and three space dimensions.
• The key role of variational mathematics.
This Chapter introduces terminology used in FEM modeling, and surveys attributes and types of
finite elements used in structural mechanics. The next Chapter goes over more specific rules for
defining meshes and boundary conditions.

§6.2. FEM Terminology

The ubiquitous term “degrees of freedom,” often abbreviated to either freedom or DOF, has figured
prominently in previous Chapters. This term, as well as “stiffness matrix” and “force vector,”
originated in structural mechanics, the application for which FEM was invented. These names have
carried over to non-structural applications. This “terminology overspill” is discussed next.
Classical analytical mechanics is that invented by Euler and Lagrange in the XVIII century and
further developed by Hamilton, Jacobi and Poincaré as a systematic formulation of Newtonian
mechanics. Its objects of attention are models of mechanical systems ranging from material particles
composed of sufficiently large number of molecules, through airplanes, to the Solar System.1 The
spatial configuration of any such system is described by its degrees of freedom or DOF. These are
also called generalized coordinates. The terms state variables and primary variables are also used,
particularly in mathematically oriented treatments.
If the number of degrees of freedom is finite, the model is called discrete, and continuous otherwise.
Because FEM is a discretization method, the number of DOF of a FEM model is necessarily finite.
They are collected in a column vector called u. This vector is called the DOF vector or state vector.
The term nodal displacement vector for u is reserved to mechanical applications.
In analytical mechanics, each degree of freedom has a corresponding “conjugate” or “dual” term,
which represents a generalized force.2 In non-mechanical applications, there is a similar set of
conjugate quantities, which for want of a better term are also called forces or forcing terms. They
are the agents of change. These forces are collected in a column vector called f. The inner product
fT u has the meaning of external energy or work.3
Just as in the truss problem, the relation between u and f is assumed to be of linear and homogeneous.
The last assumption means that if u vanishes so does f. The relation is then expressed by the master

1 For cosmological scales, such as galaxy clusters or black holes, the general theory of relativity is necessary. For the
atomic and sub-atomic world, quantum mechanics is appropriate.
2 In variational mathematics this is called a duality pairing.
3 Energy is the capacity to do work. Thus energy and work potentials are the same function (or functional), but with signs
reversed.

6–3
Chapter 6: FEM MODELING: INTRODUCTION

Table 6.1. Significance of u and f in Miscellaneous FEM Applications

Application State (DOF) vector u Conjugate vector f

Problem represents represents

Structures and solid mechanics Displacement Mechanical force

Heat conduction Temperature Heat flux
Acoustic fluid Displacement potential Particle velocity
Potential flows Pressure Particle velocity
General flows Velocity Fluxes
Electrostatics Electric potential Charge density
Magnetostatics Magnetic potential Magnetic intensity

stiffness equations:
Ku = f. (6.1)

K is universally called the stiffness matrix even in non-structural applications because no consensus
has emerged on different names.
The physical significance of the vectors u and f varies according to the application being modeled,
as illustrated in Table 6.1.
If the relation between forces and displacements is linear but not homogeneous, equation (6.1)
generalizes to
Ku = f M + f I . (6.2)

Here f I is the initial node force vector introduced in Chapter 29 for effects such as temperature
changes, and f M is the vector of mechanical forces.
The basic steps of FEM are discussed below in more generality. Although attention is focused on
structural problems, most of the steps translate to other applications problems as noted above. The
role of FEM in numerical simulation is schematized in Figure 6.1, which is a merged simplification
of Figures 1.2 and 1.3. Although this diagram oversimplifies the way FEM is actually used, it serves
to illustrate terminology. The three key simulation steps shown are: idealization, discretization and
solution. Each step is a source of errors. For example, the discretization error is the discrepancy
that appears when the discrete solution is substituted in the mathematical model. The reverse steps:
continuification and realization, are far more difficult and (generally) ill-posed problems.
The idealization and discretization steps, briefly mentioned in Chapter 1, deserve further discussion.
The solution step is dealt with in more detail in Part III of this book.

§6.3. Idealization

Idealization passes from the physical system to a mathematical model. This is the most important
step in engineering practice, because it cannot be “canned.” It must be done by a human.

6–4
 §6.3 IDEALIZATION

IDEALIZATION DISCRETIZATION SOLUTION

FEM
Physical Mathematical Discrete Discrete
system model model solution

REALIZATION & CONTINUIFICATION Solution error

IDENTIFICATION
Discretization + solution error
Modeling + discretization + solution error

Figure 6.1. A simplified view of the physical simulation process, reproduced

from Chapter 2 to illustrate modeling terminology.

§6.3.1. Models
The word “model” has the traditional meaning of a scaled copy or representation of an object. And
that is precisely how most dictionaries define it. We use here the term in a more modern sense,
which has become increasingly common since the advent of computers:

A model is a symbolic device built to simulate and predict aspects of behavior of a system.
(6.3)
Note the distinction made between behavior and aspects of behavior. To predict everything, in all
physical scales, you must deal with the actual system. A model abstracts aspects of interest to the
modeler.4 The qualifier symbolic means that a model represents a system in terms of the symbols
and language of another discipline. For example, engineering systems may be (and are) modeled
with the symbols of mathematics and/or computer sciences.5

§6.3.2. Mathematical Models

Mathematical modeling, or idealization, is a process by which an engineer or scientist passes from
the actual physical system under study, to a mathematical model of the system, where the term
model is understood in the sense of (6.3).
The process is called idealization because the mathematical model is necessarily an abstraction of
the physical reality — note the phrase aspects of behavior in (6.3). The analytical or numerical
results produced by the mathematical model are physically re-interpreted only for those aspects.6
To give an example of the choices that an engineer may face, suppose that the structure is a flat plate
structure subjected to transverse loading. Here is a non-exhaustive list of four possible mathematical
models:
1. A very thin plate model based on Von Karman’s coupled membrane-bending theory.
2. A thin plate model, such as the classical Kirchhoff’s plate theory.

4 “All models are wrong, some are useful.” (George Box)

5 A problem-definition input file, a digitized earthquake record, or a stress plot are examples of the latter.
6 Whereas idealization can be reasonably taught in advanced design courses, the converse process of “realization” or
“identification” — see Figure 6.1 — generally requires considerable physical understanding and maturity that can only
be gained through professional experience.

6–5
Chapter 6: FEM MODELING: INTRODUCTION

EN T
M PON
al CO EVE L
atic L
hem
Matmodel
ent
ponns
ary Com atio
M Libr ent
equ
E
F pon
Com ete TEM
SYS EL
r
disc del
mo
LEV e
plet
Comution
sol
em
Systrete
disc del
m o
l
sica
Phy tem
sys

Figure 6.2. A reproduction of Figure 1.5 with some relabeling. Illustrates implicit modeling:
picking elements from an existing FEM code consents to an idealization. This has professional as
well as legal implications.

3. A moderately thick plate model, for example that of Mindlin-Reissner plate theory.
4. A very thick plate model based on three-dimensional elasticity.
The person responsible for this kind of decision is supposed to be familiar with the advantages,
disadvantages, and range of applicability of each model. Furthermore the decision may be different
in static analysis than in dynamics.
Why is the mathematical model an abstraction of reality? Engineering systems, particularly in
Aerospace and Mechanical, tend to be highly complex. For simulation it is necessary to reduce that
complexity to manageable proportions. Mathematical modeling is an abstraction tool by which
complexity can be tamed.
Complexity control is achieved by “filtering out” physical details that are not relevant to the design
and analysis process. For example, a continuum material model filters out the aggregate, crystal,
molecular and atomic levels of matter. Engineers are typically interested in a few integrated
quantities, such as the maximum deflection of a bridge or the fundamental periods of an airplane.
Although to a physicist this is the result of the interaction of billions and billions of molecules, such
details are weeded out by the modeling process. Consequently, picking a mathematical model is
equivalent to choosing an information filter.

§6.3.3. Implicit vs. Explicit Modeling

As noted the diagram of Figure 6.1 is an oversimplification of engineering practice. The more
common scenario is that pictured in Figures 1.2, 1.4 and 1.5 of Chapter 1. The latter is reproduced
in Figure 6.2 for convenience.
A common scenario in industry is: you have to analyze a structure or portion(s) of one, and at your
disposal is a “black box” general-purpose finite element program. Those programs offer a catalog
of element types; for example, bars, beams, plates, shells, axisymmetric solids, general 3D solids,
and so on. The moment you choose specific elements from the catalog you automatically accept
the mathematical models on which the elements are based. This is implicit modeling. Ideally you

6–6
 §6.4 DISCRETIZATION

should be fully aware of the implications of your choice. Providing such “finite element literacy” is
one of the objective of this book. Unfortunately many users of commercial programs are unaware
of the implied-consent aspect of implicit modeling and their legal implications.7
The other extreme happens when you select a mathematical model of the physical problem with
your eyes wide open and then either shop around for a finite element program that implements
that model, or write the program yourself. This is explicit modeling. It requires far more technical
expertise, resources, experience and maturity than implicit modeling. But for problems that fall out
of the ordinary it could be the right thing to do.
In practice a combination of implicit and explicit modeling is common. The physical problem
to be simulated is broken down into subproblems. Those subproblems that are conventional and
fit available programs may be treated with implicit modeling, whereas those that require special
handling may only submit to explicit modeling.

§6.4. Discretization
Mathematical modeling is a simplifying step. But models of physical systems are not necessarily
simple to solve. They often involve coupled partial differential equations in space and time subject to
boundary and/or interface conditions. Such models have an infinite number of degrees of freedom.

§6.4.1. Analytical or Numerical?

At this point one faces the choice of going for analytical or numerical solutions. Analytical solutions,
also called “closed form solutions,” are more intellectually satisfying, particularly if they apply to
a wide class of problems, so that particular instances may be obtained by substituting the values of
free parameters. Unfortunately they tend to be restricted to regular geometries and simple boundary
conditions. Moreover some closed-form solutions, expressed for example as inverses of integral
transforms, may have to be anyway numerically evaluated to be useful.
Most problems faced by the engineer either do not yield to analytical treatment or doing so would
require a disproportionate amount of effort.8 The practical way out is numerical simulation. Here
is where finite element methods enter the scene.
To make numerical simulations practical it is necessary to reduce the number of degrees of freedom
to a finite number. The reduction is called discretization. The product of the discretization process
is the discrete model. As discussed in Chapter 1, for complex engineering systems this model is
the product of a multilevel decomposition.
Discretization can proceed in space dimensions as well as in the time dimension. Because the
present book deals with static problems, we need not consider the time dimension and are free to
focus on spatial discretization.
7 Legal problems arise when something goes wrong. Say, a structure collapses under service (non emergency) conditions,
and there is loss of life. Who is to blame? It should be noted that vendors of commercial FEM codes are generally
immune to lawsuits through “accept use” clauses inserted by company lawyers.
8 This statement has to be tempered in two respects. First, the wider availability and growing power of computer algebra
systems, outlined in Chapter 4, has widened the realm of analytical solutions than can be obtained within a practical time
frame. Second, a combination of analytical and numerical techniques is often effective in reducing the dimensionality
of the problem to facilitate parameter studies. Important examples are provided by Fourier analysis, perturbation and
boundary-element methods.

6–7
Chapter 6: FEM MODELING: INTRODUCTION

§6.4.2. Error Sources and Approximation

Figure 6.1 conveys graphically that each simulation step introduces a source of error. In engineering
practice modeling errors are by far the most important. But they are difficult and expensive to
evaluate, because model validation requires access to and comparison with experimental results.
These may be either scarce, or unavailable in the case of a new product in the design stage.
Next in order of importance is the discretization error. Even if solution errors are ignored — and
usually they can — the computed solution of the discrete model is in general only an approximation
in some sense to the exact solution of the mathematical model. A quantitative measurement of
this discrepancy is called the discretization error. The characterization and study of this error is
addressed by a branch of numerical mathematics called approximation theory.
Intuitively one might suspect that the accuracy of the discrete model solution would improve as
the number of degrees of freedom is increased, and that the discretization error goes to zero as that
number goes to infinity. This loosely worded statement describes the convergence requirement of
discrete approximations. One of the key goals of approximation theory is to make the statement as
precise as it can be expected from a branch of mathematics.9

§6.4.3. Other Discretization Methods

It was stated in Chapter 1 that the most popular discretization techniques in structural mechanics are
finite element methods and boundary element methods. The finite element method (FEM) is by far
the most widely used. The boundary element method (BEM) has gained in popularity for special
types of problems, particularly those involving infinite domains, but remains a distant second, and
seems to have reached its natural limits.
In non-structural application areas such as fluid mechanics and electromagnetics, the finite element
method is gradually making up ground but faces stiff competition from both the classical and energy-
based finite difference methods. Finite difference and finite volume methods are particularly well
entrenched in computational fluid dynamics spanning moderate to high Reynolds numbers.

§6.5. The Finite Element Method

The finite element method (FEM) is the dominant discretization technique in structural mechanics.
As discussed in Chapter 1, the FEM can be interpreted from either a physical or mathematical
viewpoint. The treatment in Chapters 1–10 emphasizes the former.
The basic concept in the physical FEM is the subdivision of the mathematical model into disjoint
(non-overlapping) components of simple geometry called finite elements or elements for short. The
response of each element is expressed in terms of a finite number of degrees of freedom characterized
as the value of an unknown function, or functions, at a set of nodal points. The response of the
mathematical model is then considered to be approximated by that of the discrete model obtained
by connecting or assembling the collection of all elements.

9 The discretization error is often overhyped in the FEM literature, since it provides an inexhaustible source of publishable
poppycock. If the mathematical model is way off, reducing the discretization error buys nothing; just a more accurate
answer to the wrong problem.

6–8
 §6.6 ELEMENT ATTRIBUTES

1D

2D

2D

3D

Figure 6.3. Typical finite element geometries in one through three dimensions.

The disconnection-assembly concept occurs naturally when examining many artificial and natural
systems. For example, it is easy to visualize an engine, bridge, building, airplane, or skeleton as
fabricated from simpler components.
Unlike finite difference models, finite elements do not overlap in space. In the mathematical
interpretation of the FEM, this property goes by the name disjoint support or local support.

§6.6. Element Attributes

Just like members in the truss example, one can take finite elements of any kind one at a time. Their
local properties can be developed by considering them in isolation, as individual entities. This is
the key to the modular programming of element libraries.
In the Direct Stiffness Method, elements are isolated by the disconnection and localization steps,
which were described for the truss example in Chapter 2. The procedure involves the separation of
elements from their neighbors by disconnecting the nodes, followed by referral of the element to a
convenient local coordinate system.10 After that we can consider generic elements: a bar element,
a beam element, and so on. From the standpoint of the computer implementation, it means that
you can write one subroutine or module that constructs, by suitable parametrization, all elements
of one type, instead of writing a new one for each element instance.
Following is a summary of the data associated with an individual finite element. This data is used
in finite element programs to carry out element level calculations.

§6.6.1. Element Dimensionality

10 Both steps are only carried out in the modeler’s mind. They are placed as part of the DSM for instructional convenience.
In practice, processing begins directly at the element level.

6–9
Chapter 6: FEM MODELING: INTRODUCTION

Elements can have intrinsic dimensionality of one, two or three space dimensions.11 There are also
special elements with zero dimensionality, such as lumped springs or point masses. The intrinsic
dimensionality can be expanded as necessary by use of kinematic transformations. For example a
1D element such as a bar, spar or beam may be used to build a model in 2D or 3D space.
§6.6.2. Element Nodes
Each element possesses a set of distinguishing points called nodal points or nodes for short. Nodes
serve a dual purpose: definition of element geometry, and home for degrees of freedom. When a
distinction is necessary we call the former geometric nodes and the latter connection nodes. For
most elements studied here, geometric and connector nodes coalesce.
Nodes are usually located at the corners or end points of elements, as illustrated in Figure 6.3. In
the so-called refined or higher-order elements nodes are also placed on sides or faces, as well as
possibly the interior of the element.
Remark 6.1. In some elements geometric and connection nodes may be at different locations. This is illustrated
by the Veubeke equilibrium triangle described in Chapter 15. Some elements have purely geometric nodes,
also called orientation nodes to complete the definition of certain geometric attributes. An example is the spar
element in 3D shown in Figure E5.2, in which a third geometric node is used to define a local plane.

§6.6.3. Element Geometry

The geometry of the element is defined by the placement of the geometric nodal points. Most
elements used in practice have fairly simple geometries. In one-dimension, elements are usually
straight lines or curved segments. In two dimensions they are of triangular or quadrilateral shape.
In three dimensions the most common shapes are tetrahedra, pentahedra (also called wedges or
prisms), and hexahedra (also called cuboids or “bricks”). See Figure 6.3.
§6.6.4. Element Degrees of Freedom
The element degrees of freedom (DOF) specify the state of the element. They also function as
“handles” through which adjacent elements are connected. DOFs are defined as the values (and
possibly derivatives) of a primary field variable at connector node points. The actual selection
depends on criteria studied at length in Part II. Here we simply note that the key factor is the way
in which the primary variable appears in the mathematical model. For mechanical elements, the
primary variable is the displacement field and the DOF for many (but not all) elements are the
displacement components at the nodes.
§6.6.5. Nodal Forces
There is always a set of nodal forces in a one-to-one correspondence with degrees of freedom. In
mechanical elements the correspondence is established through energy arguments.
§6.6.6. Element Constitutive Properties
For a mechanical element these are relations that specify the material behavior. For example, in a
linear elastic bar element it is sufficient to specify the elastic modulus E and the thermal coefficient
of expansion α.
11 In dynamic analysis, time may appear as an additional dimension.

6–10
 §6.7 CLASSIFICATION OF MECHANICAL ELEMENTS

Physical Mathematical Finite Element

Structural Model Name Idealization
Component

bar

beam

tube, pipe

spar (web)

shear panel
(2D version of above)

Figure 6.4. Examples of primitive structural elements.

§6.6.7. Element Fabrication Properties

For mechanical elements these are fabrication properties which have been integrated out from the
element dimensionality. Examples are cross sectional properties of MoM elements such as bars,
beams and shafts, as well as the thickness of a plate or shell element.
For computer implementation the foregoing data sets are organized into data structures. These are
used by element generation modules to compute element stiffness relations in the local system.

§6.7. Classification of Mechanical Elements

The following classification of finite elements in structural mechanics is loosely based on the
“closeness” of the element with respect to the original physical structure. It is given here because
it clarifies points that recur in subsequent sections, as well as providing insight into advanced
modeling techniques such as hierarchical breakdown and global-local analysis.

§6.7.1. Primitive Structural Elements

These resemble fabricated structural components. They are often drawn as such; see Figure 6.4.
The qualifier primitive distinguishes them from macroelements, which is another element class
described below. Primitive means that they are not decomposable into simpler elements.
These elements are usually derived from Mechanics-of-Materials simplified theories and are better
understood from a physical, rather than mathematical, standpoint. Examples are the elements
discussed in Chapter 5: bars, cables, beams, shafts, spars.

6–11
Chapter 6: FEM MODELING: INTRODUCTION

Physical Finite element Physical Finite element

idealization idealization

plates 3D solids

Figure 6.5. Continuum element examples.

§6.7.2. Continuum Elements

These do not resemble fabricated structural components at all. They result from the subdivision of
“blobs” of continua, or of structural components viewed as continua.
Unlike structural elements, continuum elements are better understood in terms of their mathematical
interpretation. Examples: plates, slices, shells, axisymmetric solids, general solids. See Figure 6.5.

§6.7.3. Special Elements

Special elements partake of the characteristics of structural and continuum elements. They are
derived from a continuum mechanics standpoint but include features closely related to the physics
of the problem. Examples: crack elements for fracture mechanics applications, shear panels, infinite
and semi-infinite elements, contact and penalty elements, rigid-body elements. See Figure 6.6.

Infinity
double node

Crack Infinite Honeycomb

element element panel

Figure 6.6. Special element examples.

§6.7.4. Macroelements
Macroelements are also called mesh units and superelements, although the latter term overlaps with
substructures (defined below). These often resemble structural components, but are fabricated with
simpler elements. See Figure 6.7.
The main reason for introducing macroelements is to simplify preprocessing tasks. For example, it
may be simpler to define a regular 2D mesh using quadrilaterals rather than triangles. The fact that,
behind the scene, the quadrilateral is actually a macroelement may not be important to most users.

6–12
 §6.9 BOUNDARY CONDITIONS

Figure 6.7. Macroelement examples.

Similarly a box macroelement can save modeling times for structures that are built by such com-
ponents; for example box-girder bridges

§6.7.5. Substructures
Also called structural modules and superelements. These are sets of elements with a well defined
structural function, typically obtained by cutting the complete structure into functional components.
Examples: the wings and fuselage of an airplane; the towers, deck and cables of a suspension bridge.
The distinction between substructures and macroelements is not clear-cut. The main conceptual
distinction is that substructures are defined “top down” as parts of a complete structure, whereas
macroelements are built “bottom up” from primitive elements. The term superelement is often used
in a collective sense to embrace element groupings that range from macroelements to substructures.
This topic is further covered in Chapter 10.

§6.8. Assembly

The assembly procedure of the Direct Stiffness Method for a general finite element model follows
rules identical in principle to those discussed for the truss example. As in that case the processs
involves two basic steps:
Globalization. The element equations are transformed to a common global coordinate system, if
necessary.
Merge. The element stiffness equations are merged into the master stiffness equations by appropriate
indexing and matrix-entry addition.
The hand calculations for the example truss conceal, however, the implementation complexity.
The master stiffness equations in practical applications may involve thousands or even millions of
freedoms, and programming can become involved. The topic is elaborated upon in Chapter 25.

§6.9. Boundary Conditions

A key strength of the FEM is the ease and elegance with which it handles arbitrary boundary
and interface conditions. This power, however, has a down side. A big hurdle faced by FEM
newcomers is the understanding and proper handling of boundary conditions. Below is a simple
recipe for treating boundary conditions. The following Chapter provides more specific rules and
examples.

6–13
Chapter 6: FEM MODELING: INTRODUCTION

§6.9.1. Essential and Natural B.C.

The key thing to remember is that boundary conditions (BCs) come in two basic flavors: essential
and natural.
Essential BCs directly affect DOFs, and are imposed on the left-hand side vector u.
Natural BCs do not directly affect DOFs and are imposed on the right-hand side vector f.
The mathematical justification for this distinction requires use of concepts from variational calculus,
and is consequently relegated to Part II. For the moment, the basic recipe is:

1. If a boundary condition involves one or more degrees of freedom in a direct way, it

is essential. An example is a prescribed node displacement.
2. Otherwise it is natural.

The term “direct” is meant to exclude derivatives of the primary function, unless those derivatives
also appear as degrees of freedom, such as rotations in beams and plates.

§6.9.2. B.C. in Structural Problems

Essential boundary conditions in mechanical problems involve displacements (but not strain-type
displacement derivatives). Support conditions for a building or bridge problem furnish a particularly
simple example. But there are more general boundary conditions that occur in practice. A structural
engineer must be familiar with displacement B.C. of the following types.

Ground or support constraints. Directly restraint the structure against rigid body motions.
Symmetry conditions. To impose symmetry or antisymmetry restraints at certain points, lines or
planes of structural symmetry. This allows the discretization to proceed only over part of the
structure with a consequent savings in modeling effort and number of equations to be solved.
Ignorable freedoms. To suppress displacements that are irrelevant to the problem.12 Even ex-
perienced users of finite element programs are sometimes baffled by this kind. An example are
rotational degrees of freedom normal to smooth shell surfaces.
Connection constraints. To provide connectivity to adjoining structures or substructures, or to
specify relations between degrees of freedom. Many conditions of this type can be subsumed under
the label multipoint constraints or multifreedom constraints. These can be notoriously difficult to
handle from a numerical standpoint, and are covered in Chapters 8–9.

12 In classical dynamics these are called ignorable coordinates.

6–14
 §6. References

Notes and Bibliography

Most FEM textbooks do not provide a systematic treatment of modeling. This is no accident: few academic
authors have experience with complex engineering systems. Good engineers are too busy (and in demand) to
have time for writing books. This gap has been particularly acute since FEM came on the scene because of
generational gaps: “real engineers” tend to mistrust the computer, and often for good reason. The notion of
explicit versus implicit modeling, which has legal and professional implications, is rarely mentioned.
FEM terminology is by now standard, and so is a majority of the notation. But that is not so in early publications.
E.g. K is universally used13 for stiffness matrix in virtually all post-1960 books. There are a few exceptions:
the often cited book of Przemieniecki [575] uses S. There is less unanimity on u and f for node displacement
and force vectors, respectively; some books such as Zienkiewicz and Taylor [794] still use different symbols.
The element classification given here attempts to systematize dispersed references. In particular, the distinction
between macroelements, substructures and superelements is an ongoing source of confusion, particularly
since massively parallel computation popularized the notion of “domain decomposition” in the computer
science community. The all-encompassing term “superelement” emerged in Norway by 1968 as part of the
implementation of the computer program SESAM. Additional historical details are provided in Chapter 10.
The topic of BC classification and handling is a crucial one in practice. More modeling mistakes are done in
this aspect of FEM application than anywhere else.

References
Referenced items have been moved to Appendix R.

13 A symbol derived from the “spring constant” k that measures the stiffness of a mechanical spring.

6–15
 7
FEM Modeling:
Mesh, Loads
and BCs

7–1
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

TABLE OF CONTENTS

Page
§7.1. Introduction 7–3
§7.2. General Recommendations 7–3
§7.3. Guidelines on Element Layout 7–3
§7.3.1. Mesh Refinement . . . . . . . . . . . . . . . . 7–3
§7.3.2. Element Aspect Ratios . . . . . . . . . . . . . . 7–4
§7.3.3. Physical Interfaces . . . . . . . . . . . . . . . 7–5
§7.3.4. Preferred Shapes . . . . . . . . . . . . . . . . 7–5
§7.4. Direct Lumping of Distributed Loads 7–5
§7.4.1. Node by Node Lumping . . . . . . . . . . . . . . 7–6
§7.4.2. Element by Element Lumping . . . . . . . . . . . 7–6
§7.4.3. *Weighted Lumping . . . . . . . . . . . . . . . 7–9
§7.4.4. *Energy Consistent Lumping . . . . . . . . . . . 7–9
§7.5. Boundary Conditions 7–10
§7.6. Support Conditions 7–11
§7.6.1. Supporting Two Dimensional Bodies . . . . . . . . . 7–11
§7.6.2. Supporting Three Dimensional Bodies . . . . . . . . 7–12
§7.7. Symmetry and Antisymmetry Conditions 7–12
§7.7.1. Symmetry and Antisymmetry Visualization . . . . . . . 7–12
§7.7.2. Effect of Loading Patterns . . . . . . . . . . . . . 7–13
§7. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 7–15
§7. References . . . . . . . . . . . . . . . . . . . . . . 7–15
§7. Exercises . . . . . . . . . . . . . . . . . . . . . . 7–16

7–2
 §7.3 GUIDELINES ON ELEMENT LAYOUT

§7.1. Introduction
This Chapter continues the exposition of finite element modeling principles. After some general
recommendations, it gives guidelines on layout of finite element meshes, conversion of distributed
loads to node forces, and handling the simplest forms of support boundary conditions. The next two
Chapters deal with more complicated forms of boundary conditions called multifreedom constraints.
The presentation is “recipe oriented” and illustrated by specific examples from structural mechanics.
Most examples are two-dimensional. No attempt is made at rigorous justification of rules and
recommendations, because that would require mathematical tools beyond the scope of this course.

§7.2. General Recommendations

The general rules that should guide you in the use of commercial or public FEM packages, are1

• Use the simplest type of finite element that will do the job.
• Never, never, never mess around with complicated or special elements, unless you
are absolutely sure of what you are doing.
• Use the coarsest mesh you think will capture the dominant physical behavior of the
physical system, particularly in design applications.

Three word summary: keep it simple. Initial FE models may have to be substantially revised to
accommodate design changes. There is little point in using complicated models that will not survive
design iterations. The time for refinement is when the design has stabilized and you have a better
picture of the underlying physics, possibly reinforced by experiments or observation.

§7.3. Guidelines on Element Layout

The following guidelines are stated for structural applications. As noted above, they will be often
illustrated for two-dimensional meshes of continuum elements for ease of visualization.
§7.3.1. Mesh Refinement
Use a relatively fine (coarse) discretization in regions where you expect a high (low) gradient of
strains and/or stresses.2 Regions to watch out for high gradients are:
• Near entrant corners (see Remark below) or sharply curved edges.
• In the vicinity of concentrated (point) loads, concentrated reactions, cracks and cutouts.
• In the interior of structures with abrupt changes in thickness, material properties or cross
sectional areas.
The examples in Figure 7.1 illustrate some of these “danger regions.” Away from such regions one
can use a fairly coarse discretization within constraints imposed by the need of representing the
structural geometry, loading and support conditions reasonably well.

1 Paraphrasing the Bellman in The Hunting of the Snark: “what I say three times is true.”
2 Gradient is the key word. High gradient means rapid variation. A high value by itself means nothing in this context.

7–3
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

entrant corners
(cf. Remark 7.1)
Cutouts Cracks Vicinity of concentrated (point)
loads, and sharp contact areas
weld

Load transfer
(bonded joints,
Abrupt thickness
welds, anchors, Material
changes
reinforcing bars, etc.) interfaces

Figure 7.1. Some situations where a locally refined finite element

discretization (in the red-shaded areas) is recommended.

Remark 7.1. The first bullet above mentions “entrant corner.” That is a region where isostatics (principal
stress trajectories) “bunch up.” For a two-dimensional problem mathematically posed on a singly-connected
interior domain, they can be recognized as follows. Traverse the boundary CCW so the body or structure is
on your left. When hitting a sharp or rounded corner, look at the angle (positive CCW) formed by the exterior
normals (the normal going toward your right) before and after, measured from before to after, and always
taking the positive value.
If the angle exceeds 180◦ , it is an entrant corner. [If it is close to 360◦ , it is a crack tip.] For exterior problems
or multiple-connected domains, the definition must be appropriately adjusted.

§7.3.2. Element Aspect Ratios

Good Bad
When discretizing two and three dimensional prob-
lems, try to avoid finite elements of high aspect ratios:
elongated or “skinny” elements, such as the ones
illustrated on the right of Figure 7.2. (The aspect
ratio of a two- or three-dimensional element is the ratio
between its largest and smallest dimension.)
As a rough guideline, elements with aspect ratios
exceeding 3 should be viewed with caution and those
exceeding 10 with alarm. Such elements will not
necessarily produce bad results — that depends on the Figure 7.2. Elements with good and bad
loading and boundary conditions of the problem — but aspect ratios.
do introduce the potential for trouble.
Remark 7.2. In many “thin” structures modeled as continuous bodies the appearance of “skinny” elements is
inevitable on account of computational economy reasons. An example is provided by the three-dimensional
modeling of layered composites in aerospace and mechanical engineering problems.

7–4
 §7.4 DIRECT LUMPING OF DISTRIBUTED LOADS

No OK

Physical interface

Figure 7.3. llustration of the rule that elements should not cross material interfaces.

§7.3.3. Physical Interfaces

A physical interface, resulting from example from a change in material, should also be an interele-
ment boundary. That is, elements must not cross interfaces. See Figure 7.3.

§7.3.4. Preferred Shapes

In 2D FE modeling, if you have a choice between triangles and quadrilaterals with similar nodal
arrangement, prefer quadrilaterals. Triangles are quite convenient for mesh generation, mesh
transitions, rounding up corners, and the like. But sometimes triangles can be avoided altogether
with some thought. One of the homework exercises is oriented along these lines.
In 3D FE modeling, prefer strongly bricks over wedges, and wedges over tetrahedra. The latter
should be used only if there is no viable alternative.3 The main problem with tetrahedra and wedges
is that they can produce wrong stress results even if the displacement solution looks reasonable.

§7.4. Direct Lumping of Distributed Loads

In practical structural problems, distributed loads are more common than concentrated (point)
loads.4 Distributed loads may be of surface or volume type.
Distributed surface loads (called surface tractions in continuum mechanics) are associated with
actions such as wind or water pressure, snow weight on roofs, lift in airplanes, live loads on bridges,
and the like. They are measured in force per unit area.
Volume loads (called body forces in continuum mechanics) are associated with own weight (gravity),
inertial, centrifugal, thermal, prestress or electromagnetic effects. They are measured in force per
unit volume.
A derived type: line loads, result from the integration of surface loads along one transverse direction,
such as a beam or plate thickness, or of volume loads along two transverse directions, such as a bar
or beam area. Line loads are measured in force per unit length.
Whatever their nature or source, distributed loads must be converted to consistent nodal forces for
FEM analysis. These forces end up in the right-hand side of the master stiffness equations.

3 Unfortunately, many existing space-filling automatic mesh generators in three dimensions produce tetrahedral meshes.
There are generators that try to produce bricks, but these often fail in geometrically complicated regions.
4 In fact, one of the objectives of a good structural design is to avoid or alleviate stress concentrations produced by
concentrated forces.

7–5
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

Nodal force f3 at 3 is set to P, the

magnitude of the crosshatched area
under the load curve. This area
extends halfway over adjacent
;;
;;
Distributed load
intensity (load acts
downward on boundary)

;;
element sides

;;
f3 = P
Boundary
1 2 3 4 5 6

Finite element
mesh

Figure 7.4. NbN direct lumping of distributed line load, illustrated for a 2D problem.

The meaning of “consistent” can be made precise through variational arguments, by requiring that
the distributed loads and the nodal forces produce the same external work. Since this requires
the introduction of external work functionals, the topic is deferred to Part II. However, a simpler
approach called direct load lumping, or simply load lumping, is often used by structural engineers
in lieu of the mathematically impeccable but complicated variational approach. Two variants of
this technique are described below for distributed surface loads.

§7.4.1. Node by Node Lumping

The node by node (NbN) lumping method is graphically explained in Figure 7.4. This example
shows a distributed surface loading acting normal to the straight boundary of a two-dimensional
FE mesh. (The load is assumed to have been integrated through the thickness normal to the figure,
so it is actually a line load measured as force per unit length.)
The procedure is also called tributary region or contributing region method. For the example of
Figure 7.4, each boundary node is assigned a tributary region around it that extends halfway to
adjacent nodes. The force contribution P of the cross-hatched area is directly assigned to node 3.
This method has the advantage of not requiring the error-prone computation of centroids, as needed
in the EbE technique discussed below. For this reason it is often preferred in hand computations.5
It can be extended to three-dimensional meshes as well as volume loads.6 It should be avoided,
however, when the applied forces vary rapidly (within element length scales) or act only over
portions of the tributary regions.

§7.4.2. Element by Element Lumping

In this variant the distributed loads are divided over element domains. The resultant load is assigned
to the centroid of the load diagram, and apportioned to the element nodes by statics. A node force
is obtained by adding the contributions from all elements meeting at that node. The procedure is
illustrated in Figure 7.5, which shows details of the computation over 2–3. The total force at node
3, for instance, would be that contributed by segments 2–3 and 3–4. For the frequent case in which

5 It has been extensively used in the aircraft industry for smooth-varying pressure loads computations.
6 The computation of tributary areas and volumes for general 2D and 3D regions can be done through the so-called Voronoi
diagrams. This is an advanced topic in computational geometry (see, e.g., [194]) and thus not treated here.

7–6
 §7.4 DIRECT LUMPING OF DISTRIBUTED LOADS

;;;
Distributed load
Force P has magnitude of intensity (load acts C
f 2e = (b/Le )P

;;;
crosshatched area under load downward on boundary) f3e = (a/Le)P
curve and acts at its centroid

;;;
e centroid C of P
crosshatched area 2 3
a b

;;;
f2e P f3e
Boundary Le=a+b
1 2 3 4 5 6 Details of element-level
force computations

Finite element
mesh

Figure 7.5. EbE direct lumping of distributed line load, illustrated for a 2D problem.

;;;;
the variation of the load over the element is linear (so the area under the load is a trapezoid) the
node forces can be computed directly by the formulas given in Figure 7.6.
If applicable, EbE is more accurate

;;;;
;;;;
than NbN lumping. In fact it agrees
with consistent node lumping for sim- qj

;;;;
ple elements that possess only corner qi e
e e
nodes. In those cases it is not affected f ie = L (2qi +qj ) f je = L6 (qi +2qj )
6
by sharpness of the load variation and i j
can be even used for point loads that are Le
not applied at the nodes. Figure 7.6. EbE lumping of linearly varying line load over element.

The EbE procedure is not applicable if the centroidal resultant load cannot be apportioned by statics.
This happens if the element has midside faces or internal nodes in addition to corner nodes, or if
it has rotational degrees of freedom. For those elements the variational-based consistent approach
covered in Part II and briefly outlined in §7.3.3, is preferable.
Example 7.1. Figures 7.7(a,b) show web-downloaded pictures of the Norfork Dam, a 220-ft high, concrete-
built gravity dam. Its typical cross section is shown in 7.7(c). The section is discretized by triangular elements
as illustrated in Figure 7.7(d). The dam has a length of 2624 ft and was constructed over the White River in
Arkansas over 1941–44.7 The structure is assumed to be in plane strain. Accordingly the FEM model shown
in 7.7(d) is a typical 1-ft slice of the dam and near-field soil.
In the analysis of dam and marine structures, a “wet node” of a FEM discretization is one in contact with the
water. The effect of hydrostatic pressure is applied to the structure through nodal forces on wet nodes. The wet
nodes for a water head of 180 ft over the riverbed are shown in Figure 7.8(b). Nodes are numbered 1 through
9 for convenience. The pressure in psf (lbs per sq-ft) is p = 62.4 d where d is the depth in ft. Compute the
horizontal hydrostatic nodal forces f x1 and f x2 using NbN and EbE, assuming that the wet face AB is vertical
for simplicity.
Node by Node. For node 1 go halfway to 2, a distance of (180 − 136)/2 = 22 ft. The tributary load area
is a triangle extending 22 ft along y with bottom pressure 62.4 × 22 = 1372.8 psf and unit width normal to

7 As discussed in Notes and Bibliography, this example has historical significance, as the first realistic Civil Engineering
structure modeled ca. 1960 by the Finite Element Method, which until then had been largely confined to aerospace.

7–7
 ;;;;
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

(a)

;;;;
MAX WATER EL. 584'
SPILLWAY CREST EL. 552'

;;;;
(c)
Concrete
y dam

;;;;
;;
;;
200'
A
EL. 364' x

spillway crest

;;
(b) Water 188' over
C riverbed
B D

(d)
Soil

Figure 7.7. Norfork Dam: (a,b) pictures; (c) cross section of dam above foundation (line inside dam is a thermally
induced crack considered in the 1960 study); (d) coarse mesh including foundation and soil but not crack [137].

A 1 y =180 22 1 1
(a) 1 (b) (c) 44
2 y =136 70 2 2
d 2 120 96
Water 144
180 ft 3 y = 84 162 3 3
3 180
p=62.4 d 4 4 4
(lbs/sq-ft) y4 = 36
C B y5 = 0
9 8 7 5 6 5 5
x5 =0
x =−70 All dimensions in ft.
x =−210 6
x8 =−350 7 Plane strain model, slice
x 9=−490 1 ft thick normal to section

Figure 7.8. Norfork Dam example: (a) computation of wet node forces due to hydrostatic pressure;
(b) NbN tributary regions; (c) EbE regions.

paper, giving f x1 = 12 22 × 1372.8 = 15101 lbs. For node 2 go up 22 ft and down (136 − 84)/2 = 26 ft. The
tributary load area is a trapezoid extending 22 + 26 = 48 ft vertically, with pressures 1372.8 psf at the top and
62.4 × 70 = 4368.0 psf at the bottom. This gives f x2 = 48 × (1372.8 + 4368.0)/2 = 137779 lbs.
Element by Element. It is convenient to pre-compute hydro pressures at wet node levels: p1 = 0, p2 = 62.4 ×
44 = 2745.6 psf, p3 = 62.4 × 96 = 5990.4. Because the variation of p is linear, the formulas of Figure 7.6
(1) (1)
can be applied directly: f x1 = (44/6)(2 × 0 + 2745.6) = 20134 lbs, f x2 = (44/6)(0 + 2 × 2745.6) =
(2)
40269 lbs and f x2 = (52/6)(2 × 2745.6 + 5990.4) = 99507 lbs. Adding contributions to nodes 1 and 2:
(1) (1) (2)
f x1 = f x1 = 20134 lbs and f x2 = f x2 + f x2 = 40269 + 99507 = 139776 lbs.
The computations for wet nodes 3 through 9 are left as an exercise.

Example 7.2. Figure 7.9 shows the mesh for the y > 0 portion of the gravity dam of the previous example.

7–8
 §7.4 DIRECT LUMPING OF DISTRIBUTED LOADS

(The node numbering is different). The specific weight for concrete of γ = 200 pcf (pounds per cubic foot).
Compute the node force f y11 due to own weight.
For this calculation NbN is unwieldy because computa-
tion of the nodal tributary region requires construction 1 (1)
2 3
of a Voronoi diagram. (Furthermore for constant (2) (3) (4)
4 6 Node x y
specific weight it gives the same answer as EbE.) To (5) 5 (7)
(6)
7 *.00 152.00
apply EbE, select the elements that contribute to 11: 7
8 (8)
9 8 *.00 152.00
the six triangles (9), (10), (11), (15), (16) and (17). 9 *.00 152.00
(10) (12) 10 *.00 94.00
The area of a triangle with corners at {x1 , y1 }, {x2 , y2 } 10 (9) 11 (11) 12 (13) 13 11
12
*.00
*.00
94.00
94.00
and {x3 , y3 }} is given by A = (x2 y3 − x3 y2 ) + (x3 y1 − (15) (17) (19) 13 *.00 94.00
14 *.00 40.00
x1 y3 )+(x1 y2 − x2 y1 ). Applying this to the geometry of 14 (14) 15 (16) 16 (18) 17 (20) 18 (29) 15 *.00 40.00
the figure one finds that the areas are A(9) =, A(10) =, (21) (23) (25) (27) 19
16
17
*.00
*.00
40.00
40.00
(22) (26) (28)
A(11) =, A(15) =, A(16) =, and A(17) =. The weight
(24)
18 *.00 40.00
20 21 22 23 24
forces on each element are W (9) = γ h, W (10) = γ h,
W (11) = γ h, A(15) = γ h, A(16) = γ h, and A(17) = γ h, Figure 7.9. Computation of weight force at node
where h is the thickness normal to the figure (1 ft here). 11 of gravity dam example of Figure 7.7(d).

For uniform element thickness and specific weight, one third of each force goes to each element corner. Thus
node 11 receives
f y11 = 13 (W (9) + W (10) + W (11) + W (17) + W (18) ) = (7.1)

(example incomplete, TBF)

§7.4.3. *Weighted Lumping

The NbN and EbE methods are restricted to simple elements, specifically those with corner nodes only. We
outline here a general method that works for more complicated models. The mathematical justification requires
energy theorems covered in Part II. Thus at this stage the technique is merely presented as a recipe.
To fix the ideas consider again the 2D situation depicted in Figures 7.4 and 7.5. Denote the distributed load
by q(x). We want to find the lumped force f n at an interior node n of coordinate xn . Let the adjacent nodes
be n − 1 and n + 1, with coordinates xn−1 and xn+1 , respectively. Introduce a weight function Wn (x) with
properties to be specified below. The lumped force is given by
 xn+1
fn = Wn (x) q(x) d x (7.2)
xn−1

For this formula to make sense, the weight function must satisfy several properties:
1. Unit value at node n: Wn (xn ) = 1.
2. Vanishes over any element not pertaining to n: Wn = 0 if x ≤ xn−1 or x ≥ xn+1
3. Gives the same results as NbN or EbE for constant q over elements with corner nodes only.
Both NbN and EbE are special cases of (7.2). For NbN pick
Wn (x) = 1 if 1
2
(xn−1 + xn ) ≤ x ≤ 12 (xn + xn+1 ), wn (x) = 0 otherwise. (7.3)
For EbE pick  x−x
 1 − xn − xn−1
n−1
if xn−1 ≤ x ≤ xn ,
Wn (x) = x − x (7.4)
 1 − xn+1 − xn if xn ≤ x ≤ xn+1 ,
n

0 otherwise.
These particular weight functions are depicted in Figure 7.10.

7–9
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

(a) halfway Wn = 1 halfway

between between (b) Wn
n−1 and n 1 n and n+1 1

n−1 n n+1 n−1 n n+1

xn xn

Figure 7.10. Weight functions corresponding to: (a) NbN lumping, and (b) EbE lumping.

q h (per unit
§7.4.4. *Energy Consistent Lumping of x length)

The rule (7.2) can be justified from the W1 = − ξ(1−ξ)/2 = N1

standpoint of the Principle of Virtual f1 =? f2 =? 3 f 3 =? 1
1 2
Work. Let δu n be a virtual node displace- 2 3
1
ment paired to f n . Take δu n Wn (x) to W2 = 1−ξ = N2
2
be the associated displacement variation.
1
 external virtual work of q(x) is
The
q(x)Wn (x) δu n d x extended over the 1 2 3
portion where Wn = 0. Equating this
to f n δu n and cancelling δu n from both
W3= ξ(1+ξ)/2 = N3 1
sides yields (7.2).
1 2 3
If Wn is the trial displacement function
actually used for the development of ξ = 1 − 2x/a
element equations the lumping is called a x
energy consistent, or consistent for short.
Figure 7.11. Example of consistent load lumping.

(As will be seen later, trial functions are the union of shape functions over the patch of all elements connected
to node n.) This important technique is studied in Part II of the course after energy methods are introduced.

Example 7.3. Conside the mesh of 9-node quadrilaterals shown in Figure 7.11. (This is later used as a
benchmark problem in Chapter 27.) The upper plate edge 1–3 is subject to a uniform normal load qh per unit
length, where h is the plate thickness. The problem is to compute the node forces f 1 , f 2 and f 3 . If 1–2 and
2–3 were on two different elements, both NbN and EbE would give f 1 = f 3 = 14 qha and f 2 = 12 qha. But
this lumping is wrong for an element with midside nodes. Instead, pick the weight functions Wi (i = 1, 2, 3)
shown on the right of Figure 7.11.
Since there is only one element on the loaded edge, the Wi are actually the quadratic shape functions Ni for
a 3-node line element, developed in later Chapters. The dimensionless variable ξ is called an isoparametric
natural coordinate. Applying the rule (7.2) we get

 a  1  1
dx
f1 = W1 (x) q h d x = W1 (ξ ) q h dξ = − 12 ξ(1 − ξ )q h ( 12 a) dξ = 16 qha. (7.5)
0 −1 dξ −1

Similarly f 2 = 23 qha and f 3 = f 1 . As a check, f 1 + f 2 + f 3 = qha, which is the total load acting on the
plate edge.

7–10
 §7.6 SUPPORT CONDITIONS

;;;;
(a) (b) y (c) y

;;
;;
;;;
; ;
;;;;; ;;;; ;
; B B B

;;;
; ;;
A
A x A x

Figure 7.12. Suppressing two-dimensional rigid body motions.

§7.5. Boundary Conditions

The key distinction between essential and natural boundary conditions (BC) was introduced in
the previous Chapter. The distinction is explained in Part II from a variational standpoint. In
this section we discuss the simplest essential boundary conditions in structural mechanics from
a physical standpoint. This makes them relevant to problems with which a structural engineer is
familiar. Because of the informal setting, the ensuing discussion relies heavily on examples.
In structural problems formulated by the DSM, the recipe of §6.7.1 that distinguishes between
essential and natural BC is: if it directly involves the nodal freedoms, such as displacements or
rotations, it is essential. Otherwise it is natural. Conditions involving applied loads are natural.
Essential BCs take precedence over natural BCs.
The simplest essential boundary conditions are support and symmetry conditions. These appear
in many practical problems. More exotic types, such as multifreedom constraints, require more
advanced mathematical tools and are covered in the next two Chapters.

§7.6. Support Conditions

Supports are used to restrain structures against relative rigid body motions. This is done by attaching
them to Earth ground (through foundations, anchors or similar devices), or to a “ground structure”
which is viewed as the external environment.8 The resulting boundary conditions are often called
motion constraints. In what follows we analyze two- and three-dimensional motions separately.

§7.6.1. Supporting Two Dimensional Bodies

Figure 7.12 shows two-dimensional bodies that move in the plane of the paper. If a body is not
restrained, an applied load will cause infinite displacements. Regardless of loading conditions, the
body must be restrained against two translations along x and y, and one rotation about z. Thus the
minimum number of constraints that has to be imposed in two dimensions is three.
In Figure 7.12, support A provides translational restraint, whereas support B, together with A,
provides rotational restraint. In finite element terminology, we say that we delete (fix, remove,
preclude) all translational displacements at point A, and that we delete the translational degree of

8 For example, the engine of a car is attached to the vehicle frame through mounts. The car frame becomes the “ground
structure,” which moves with respect to Earth ground, as Earth rotates and moves through space, etc.

7–11
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

freedom directed along the normal to the AB direction at point B. This body is free to distort in any
manner without the supports imposing any deformation constraints.
Engineers call A and B reaction-to-ground points. This means that if the supports are conceptually
removed, applied loads are automatically balanced by reactive forces at A and B, in accordance
with Newton’s third law. Additional freedoms may be precluded to model greater restraint by the
environment. However, Figure 7.12(a) does illustrate the minimal number of constraints.
Figure 7.12(b) is a simplification of Figure 7.12(a). Here the line AB is parallel to the global y
axis. We simply delete the x and y translations at point A, and the x translation at point B. If the
roller support at B is modified as in Figure 7.12(c), however, it becomes ineffective in constraining
the infinitesimal rotational motion about point A because the rolling direction is normal to AB.
The configuration of Figure 7.12(c) is called a kinematic mechanism, and will result in a singular
modified stiffness matrix.

;
§7.6.2. Supporting Three Dimensional Bodies

;
y
Figure 7.13 illustrates the extension of the freedom-
restraining concept to three dimensions. The minimal

;;;;;
number of freedoms that have to be constrained is now B

; ;;
;;;;;
six and many combinations are possible. In the example

;
;
D
of Figure 7.13, all three degrees of freedom at point A
have been fixed. This prevents all rigid body translations, A C x
and leaves three rotations to be taken care of. The x z
displacement component at point B is deleted to prevent
rotation about z, the z component is deleted at point C to Figure 7.13. Suppressing rigid body
prevent rotation about y, and the y component is deleted motions in a three-dimensional body.
at point D to prevent rotation about x.

§7.7. Symmetry and Antisymmetry Conditions

Engineers doing finite element analysis should be on the lookout for conditions of symmetry or
antisymmetry. Judicious use of these conditions allows only a portion of the structure to be analyzed,
with a consequent saving in data preparation and computer processing time.9

§7.7.1. Symmetry and Antisymmetry Visualization

Recognition of symmetry and antisymmetry conditions can be done by either visualization of the
displacement field, or by imagining certain rotational ot reflection motions. Both techniques are
illustrated for the two-dimensional case.
A symmetry line in two-dimensional motion can be recognized by remembering the “mirror” dis-
placement pattern shown in Figure 7.14(a). Alternatively, a 180◦ rotation of the body about the
symmetry line reproduces exactly the original problem.
An antisymmetry line can be recognized by the displacement pattern illustrated in Figure 7.14(b).

9 Even if symmetry or antisymmetry are not explicitly applied through boundary conditions, they provide valuable checks
on the computed solution.

7–12
 §7.7 SYMMETRY AND ANTISYMMETRY CONDITIONS

Alternatively, a 180◦ rotation of the body

Symmetry Antisymmetry
about the antisymmetry line reproduces line line
exactly the original problem except that (a) (b)
all applied loads are reversed.
Similar recognition patterns can be drawn
in three dimensions to help visualization A' A" A' A"
A A
of planes of symmetry or antisymme- loads
try. More complex regular patterns displacement
associated with sectorial symmetry (also vectors
called harmonic symmetry) as well as
rotational symmetry can be treated in a
similar manner, but will not be discussed
here. Figure 7.14. Visualizing symmetry and antisymmetry lines.

§7.7.2. Effect of Loading Patterns

Although the structure may look symmetric in shape, it must be kept in mind that model reduction
can be used only if the loading conditions are also symmetric or antisymmetric.

y
(a) (b)
;; A B

;
A B

C D x C D

;;
;;
;;
Figure 7.15. A doubly symmetric structure under symmetric loading.

Consider the plate structure shown in Figure 7.15(a). This structure is symmetrically loaded on the
x-y plane. Applying the recognition patterns stated above one concludes that the structure is doubly
symmetric in both geometry and loading. It is evident that no displacements in the x-direction are
possible for any point on the y-axis, and that no y displacements are possible for points on the x
axis. A finite element model of this structure may look like that shown in Figure 7.15(b).
On the other hand if the loading is antisymmetric, as illustrated in Figure 7.16(a), the x axis becomes
an antisymmetry line as none of the y = 0 points can move along the x direction. The boundary
conditions to be imposed on the FE model are also different, as shown in Figure 7.16(b).
Remark 7.3. For the case shown in Figure 7.16(b) note that all rollers slide in the same direction. Thus
the vertical rigid body motion along y is not precluded. To do that, one node has to be constrained in the y
direction. If there are no actual physical supports, the choice is arbitrary and amounts only to an adjustment
on the overall (rigid-body) vertical motion. In Figure 7.16(b) the center point C has been so chosen. But any
other node could be selected as well; for example A or D. The important thing is not to overconstrain the
structure by applying more than one y constraint.

7–13
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

;;
(a) (b)
A B A B

C
D
x
;;;;
C
D

Vertical (y) motion of C (or

another node) should be constrained
to eliminate the y rigid motion

Figure 7.16. A doubly symmetric structure under antisymmetric loading.

Remark 7.4. Point loads acting at nodes located on symmetry or antisymmetry lines require special care. For
example, consider the doubly symmetric plate structure of Figure 7.15 under the two point loads of magnitude
P, as pictured in Figure 7.17(a). If the structure is broken down into 4 quadrants as in Figure 7.17(b), P must
be halved as indicated in Figure 7.17(c). The same idea applies to point loads on antisymmetry lines, but there
the process is trickier, as illustrated in Figure 7.18. The load must not be applied if the node is fixed against
motion, since then the node force will appear as a reaction.

;;
(a) (b) (c)

;;
A B
P P/2 P/2

;;
C D P
P/2

;;;
P/2 P/2

;
;;
;
Figure 7.17. Breaking up a point load acting on a symmetry line node.

P
A B
P ;; ;;
;;
C D
; ;;
;;;;

;;
;;
;;
;;

;;;;

P/2
;

; ;

P/2
2P
Figure 7.18. Breaking up a point load acting on a antisymmetry line node.

Distributed loads should not be divided when the structure is broken down into pieces, since the lumping
process will take care of the necessary apportionment to nodes.

7–14
 §7. References

Notes and Bibliography

FEM modeling rules in most textbooks are often diffuse, if given at all. In those that focus on the mathematical
interpretation of FEM they are altogether lacking; the emphasis being on academic boundary value problems.
The rule collection at the start of this Chapter attempts to collect important recommendations in one place.
The treatment of boundary conditions, particularly symmetry and antisymmetry, tends to be also flaky. A
notable exception is [397], which is understandable since Irons worked in industry (at Rolls-Royce Aerospace
Division) before moving to academia.
The Norfork Dam used in Examples 7.1 and 7.2 (and two Exercises) for hydrostatic load-lumping calculations
was the first realistic Civil Engineering structure analyzed by FEM. It greatly contributed to the acceptance
of the method beyond the aerospace industry where it had originated. How this seminal event came to pass is
narrated by Wilson in [803], from which the following fragment is taken. Annotations are inserted in squared
brackets.
“On the recommendation from Dr. Roy Carlson, a consultant to the Little Rock District of the Corps
of Engineers, Clough [then a Professor at UC Berkeley] submitted [in 1960] a proposal to perform a
finite element analysis of Norfork Dam, a gravity dam that had a temperature induced vertical crack
near the center of the section. The proposal contained a coarse mesh solution that was produced by
the new program [a matrix code developed by E. L. Wilson, then a doctoral student under Clough’s
supervision; the mesh is that shown in Figure 7.7(d)] and clearly indicated the ability of the new
method to model structures of arbitrary geometry with different orthotropic properties within the dam
and foundation. The finite element proposal was accepted by the Corps over an analog computer
proposal submitted by Professor Richard MacNeal of CalTech [who later directed the development
of NASTRAN under a NASA contract in the late 1960s], which at that time was considered as the
state-of-the-art method for solving such problems.
The Norfork Dam project provided an opportunity to improve the numerical methods used within the
program and to extend the finite element method to the nonlinear solution of the crack closing due to
hydrostatic loading. Wilson and a new graduate student, Ian King, conducted the detailed analyses
that were required by the study. The significant engineering results of the project indicated that the
cracked dam was safe [since the crack would be closed as the reservoir was filled].”

References
Referenced items moved to Appendix R.

7–15
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

Homework Exercises for Chapters 6 and 7

FEM Modeling: Mesh, Loads and BCs

P
C E
A D
B F G

;; ;;
N M J I
O H
L K

Figure E7.1. Inplane bent plate for Exercise 7.1.

EXERCISE 7.1 [D:10] The plate structure shown in Figure E7.1 is loaded and deforms in the plane of the
paper. The applied load at D and the supports at I and N extend over a fairly narrow area. List what you
think are the likely “trouble spots” that would require a locally finer finite element mesh to capture high stress
gradients. Identify those spots by its letter and a reason. For example, D: vicinity of point load.

A B

D C

Figure E7.2. Transition meshing for Exercise 7.2.

EXERCISE 7.2 [D:15] Part of a two-dimensional FE mesh has been set up as indicated in Figure E7.2.
Region ABC D is still unmeshed. Draw a transition mesh within that region that correctly merges with the
regular grids shown, uses 4-node quadrilateral elements (quadrilaterals with corner nodes only), and avoids
triangles. Note: There are several (equally acceptable) solutions.

EXERCISE 7.3 [A:15] A rectangular plate of constant thickness h and inplane dimensions 8a and 6a is
meshed with 8 rectangular elements as shown in Figure E7.3(a). The plate specific weight is γ and acts along
the −y axis direction.
(a) Compute the node forces due to plate weight at nodes 1 through 15, using the NbN method. Obtain the
node-tributary regions as sketched in Figure E7.3(b), which shows each element divided by the medians
drawn as dashed lines (the tributary region of node 7 is shown in yellow). Partial answer: f y1 = −2a 2 γ h.
Check that adding up all y forces at the 15 nodes one gets W = −48a 2 γ h.
(b) Repeat the computations using the EbE method. For this, take the total weight force on each element,
and assign one quarter to each corner node. (This agrees with consistent energy lumping for 4-node
rectangular elements.) Do the results agree with NbN lumping?

7–16
 Exercises

Plate thickness h,
(a) specific weight γ (b)
1 2 3 4 5 1 2 3 4 5
y

x (1) (2) (3) (4)

6a
6 7 8 9 10 6 7 8 9 10

2a (5) (6) (7) (8)

11 12 13 14 15 11 12 13 14 15
8a
Figure E7.3. (a) Mesh layout for Exercise 7.3. (b) shows tributary area for node 7.

EXERCISE 7.4 [N/C:20] Complete the computation of hydrostatic node forces on the Norfork Dam under a
water head of 180 ft, initiated in Example 7.1, using the data of Figure 7.8, and either the NbN or EbE method
(pick one). Assume face AB is vertical. Do two checks: sum of horizontal (x) forces on nodes 1 through 5 is
1
2
1802 × 62.4 lbs, and sum of vertical (y) forces on nodes 5 through 9 is −180 × 62.4 × 490 lbs.

;;;;
EXERCISE 7.5 [N/C:25] Complete the computation of own weight nodal forces of the coarse mesh of the
Norfork dam proper, initiated in Example 7.2, reusing the data of Figure 7.9. Use the EbE method. Recover
coordinates, and write a computer program to compute node forces. Compute the total weight of the dam slice

;;;;
in lbs.

EXERCISE 7.6 [A:10] Figure E7.4 depicts

two instances of a pull test. In (a) a stiffer Load Steel

;;;;
material (steel rod) is pulled out of a softer (a) transfer B P
A
one (concrete block); in this case the load
transfer diagram shows a rapid variation Concrete

;;;;
near the inner end of the rod. In (b) a softer
material (plastic rod) is pulled out of a stiffer
Load

;;;;
one (concrete rod), and the load transfer transfer Plastic
diagram is reversed. (b) A B P
If the test is to be simulated by a finite
element model, indicate for (a) and (b) Concrete
where a finer mesh would be desirable.
Explain. Figure E7.4. Pull tests for Exercise 7.6.

EXERCISE 7.7 [D:20] Identify the symmetry and antisymmetry lines in the two-dimensional problems
illustrated in Figure E7.5. They are: (a) a circular disk under two diametrically opposite point forces (the
famous “Brazilian test” for concrete); (b) the same disk under two diametrically opposite force pairs; (c) a
clamped semiannulus under a force pair oriented as shown; (d) a stretched rectangular plate with a central
circular hole. Finally (e) and (f) are half-planes under concentrated loads.10
Having identified those symmetry/antisymmetry lines, state whether it is possible to cut the complete structure
to one half or one quarter before laying out a finite element mesh. Then draw a coarse FE mesh indicating, with

10 Note that (e) is the famous Flamant’s problem, which is important in the 2D design of foundations of civil structures.
The analytical solution of (e) and (f) may be found, for instance, in Timoshenko-Goodier’s Theory of Elasticity, 2nd
Edition, page 85ff.

7–17
Chapter 7: FEM MODELING: MESH, LOADS AND BCS

P P
(a) (b)
P P

P P

(c)
;; ;; P
45 o
(d) (uniform)

fixed
q center q
hole
45o

(e) P (f) P P

infinite half plane

Figure E7.5. Problems for Exercise 7.7.

rollers or fixed supports, which kind of displacement BCs you would specify on the symmetry or antisymmetry
lines. Note: Do all sketches on your paper, not on the printed figures.

EXERCISE 7.8 [D:20] You (a finite element guru) pass away and come back to the next life as an intelligent
but hungry bird. Looking around, you notice a succulent big worm taking a peek at the weather. You grab one
end and pull for dinner; see Figure E7.6.
After a long struggle, however, the worm wins. While hungrily looking for a smaller one your thoughts wonder
to FEM and how the worm extraction process might be modeled so you can pull it out more efficiently. Then
you wake up to face this homework question. Try your hand at the following “worm modeling” points.
(a) The worm is simply modeled as a string of one-dimensional (bar) elements. The “worm axial force” is
of course constant from the beak B to ground level G, then decreases rapidly because of soil friction
(which varies roughly as plotted in the figure above) and drops to nearly zero over D E. Sketch how a
good “worm-element mesh” should look like to capture the axial force well.
(b) On the above model, how would you represent boundary conditions, applied forces and friction forces?
(c) Next you want a more refined anaysis of the worm that distinguishes skin and insides. What type of
finite element model would be appropriate?
(d) (Advanced) Finally, point out what need to be added to the model of (c) to include the soil as an elastic
medium.

7–18
 Exercises

;; ; ; ;
;;;;;;;;; ; ;
G

friction force on
worm surface
D

Figure E7.6. The hungry bird.

Briefly explain your decisions. Dont write equations.

EXERCISE 7.9 [D:15, 5 each] Prove from kinematics:

(a) Two symmetry lines in 2D cannot cross at a finite point, unless that point is fixed.
(b) Two antisymmetry lines in 2D cannot cross at a finite point, unless that point is fixed.
(c) A symmetry line and an antisymmetry line must cross at right angles, unless the cross point is fixed.
Note: proofs of (a,b,c) are very similar; just draw vectors at alleged intersections.

EXERCISE 7.10 [A/D:15] A 2D body has n > 1 symmetry lines passing through a point C and spanning an
angle π/n from each other. This is called sectorial symmetry if n ≥ 3. Draw a picture for n = 5, say for a car
wheel. Explain why point C is fixed.

EXERCISE 7.11 [A/D:25, 5 each] A body is in 3D space. The analogs of symmetry and antisymmetry lines
are symmetry and antisymmetry planes, respectively. The former are also called mirror planes.
(a) State the kinematic properties of symmetry and antisymmetric planes, and how they can be identified.
(b) Two symmetry planes intersect. State the kinematic properties of the intersection line.
(c) A symmetry plane and an antisymmetry plane planes intersect. State the kinematic properties of the
intersection line. Can the angle between the planes be arbitrary?
(d) Can two antisymmetry planes intersect?
(e) Three symmetry planes intersect. State the kinematic properties of the intersection point.

EXERCISE 7.12 [A:25] A 2D problem is called periodic in the x direction if all fields, in particular displace-
ments, repeat upon moving over a distance a > 0: u x (x + a, y) = u x (x, y) and u y (x + a, y) = u y (x, y). Can
this situation be treated by symmetry and/or antisymmetry lines?

EXERCISE 7.13 [A:25] Extend the previous exercise to antiperiodicity, in which u x (x + a, y) = u x (x, y)
and u y (x + a, y) = −u y (x, y).

EXERCISE 7.14 [A:20] Prove that EbE and energy consistent lumping agree if the element shape functions
are piecewise linear.

EXERCISE 7.15 [A:40] If the world were spatially n-dimensional (meaning it has elliptic metric), how many
independent rigid body modes would a body have? (Prove by induction)

7–19
 8
MultiFreedom
Constraints I

8–1
Chapter 8: MULTIFREEDOM CONSTRAINTS I

TABLE OF CONTENTS

Page
§8.1. Classification of Constraint Conditions 8–3
§8.1.1. MultiFreedom Constraints . . . . . . . . . . . . . 8–3
§8.1.2. Methods for Imposing Multifreedom Constraints . . . . . 8–4
§8.1.3. *MFC Matrix Forms . . . . . . . . . . . . . . . 8–5
§8.2. The Example Structure 8–6
§8.3. The Master-Slave Method 8–7
§8.3.1. A One-Constraint Example . . . . . . . . . . . . 8–7
§8.3.2. Multiple Homogeneous MFCs . . . . . . . . . . . . 8–8
§8.3.3. Nonhomogeneous MFCs . . . . . . . . . . . . . 8–9
§8.3.4. *The General Case . . . . . . . . . . . . . . . 8–10
§8.3.5. *Retaining the Original Freedoms . . . . . . . . . . 8–10
§8.4. Model Reduction by Kinematic Constraints 8–11
§8.5. Assessment of the Master-Slave Method 8–13
§8. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 8–14
§8. References . . . . . . . . . . . . . . . . . . . . . . 8–14
§8. Exercises . . . . . . . . . . . . . . . . . . . . . . 8–15

8–2
 §8.1 CLASSIFICATION OF CONSTRAINT CONDITIONS

§8.1. Classification of Constraint Conditions

In previous Chapters we have considered structural support conditions that are mathematically
expressable as constraints on individual degrees of freedom:

nodal displacement component = prescribed value. (8.1)

These are called single-freedom constraints. Chapter 3 explains how to incorporate constraints of
this form into the master stiffness equations, using hand- or computer-oriented techniques. The
displacement boundary conditions studied in Chapter 7, which include modeling of symmetry and
antisymmetry, lead to constraints of this form.
For example:
u x4 = 0, u y9 = 0.6. (8.2)

These are two single-freedom constraints. The first one is homogeneous while the second one is
non-homogeneous. These attributes are defined below.

§8.1.1. MultiFreedom Constraints

The next step up in complexity involves multifreedom equality constraints, or multifreedom con-
straints for short, the last name being acronymed to MFC. These are functional equations that
connect two or more displacement components:

F(nodal displacement components) = prescribed value, (8.3)

where function F vanishes if all its nodal displacement arguments do. Equation (8.3), in which all
displacement components are in the left-hand side, is called the canonical form of the constraint.
An MFC of this form is called multipoint or multinode if it involves displacement components at
different nodes. The constraint is called linear if all displacement components appear linearly on
the left-hand-side, and nonlinear otherwise.
The constraint is called homogeneous if, upon transfering all terms that depend on displacement
components to the left-hand side, the right-hand side — the “prescribed value” in (8.3) —is zero.
It is called non-homogeneous otherwise.
In this and next Chapter only linear constraints will be studied. Furthermore more attention is
devoted to the homogeneous case, because it arises more frequently in practice.

Remark 8.1. The most general constraint class is that of inequality constraints, such as u y5 − 2u x2 ≥ 0.5.
These constraints are relatively infrequent in linear structural analysis, except in problems that involve contact
conditions. They are of paramount importance, however, in other fields such as optimization and control.

8–3
Chapter 8: MULTIFREEDOM CONSTRAINTS I

Unmodified master stiffness

equations K u = f before applying MFCs

master-slave
Apply MFCs by penalty functions
Lagrange multipliers

^ ^ ^
Modified stiffness equations K u =f

^
Equation solver returns u

Recover u if necessary

Figure 8.1. Schematics of MFC application.

Example 8.1. Here are three instances of MFCs:

u x2 = 12 u y2 , u x2 − 2u x4 + u x6 = 14 , (x5 + u x5 − x3 − u x3 )2 + (y5 + u y5 − y3 − u y3 )2 = 0. (8.4)

The first one is linear and homogeneous. It is not a multipoint constraint because it involves the displacement
components of one node: 2.
The second one is multipoint because it involves three nodes: 2, 4 and 6. It is linear and non-homogeneous.
The last one is multipoint, nonlinear and homogeneous. Geometrically it expresses that the distance between
nodes 3 and 5 in two-dimensional motions on the {x, y} plane remains constant. This kind of constraint appears
in geometrically nonlinear analysis of structures, which is a topic beyond the scope of this book.

§8.1.2. Methods for Imposing Multifreedom Constraints

Accounting for multifreedom constraints is done — at least conceptually — by changing the
assembled master stiffness equations to produce a modified system of equations:
MFC
Ku = f ⇒ K̂ û = f̂. (8.5)

The modification process (8.5) is also called constraint application or constraint imposition. The
modified system is that submitted to the equation solver, which returns û.
The procedure is flowcharted in Figure 8.1. The sequence of operations sketched therein applies to
all methods outlined below.
Three methods for treating MFCs are discussed in this and the next Chapter:
1. Master-Slave Elimination. The degrees of freedom involved in each MFC are separated into
master and slave freedoms. The slave freedoms are then explicitly eliminated. The modified
equations do not contain the slave freedoms.
2. Penalty Augmentation. Also called the penalty function method. Each MFC is viewed as
the presence of a fictitious elastic structural element called penalty element that enforces it
approximately. This element is parametrized by a numerical weight. The exact constraint

8–4
 §8.1 CLASSIFICATION OF CONSTRAINT CONDITIONS

is recovered if the weight goes to infinity. The MFCs are imposed by augmenting the finite
element model with the penalty elements.
3. Lagrange Multiplier Adjunction. For each MFC an additional unknown is adjoined to the
master stiffness equations. Physically this set of unknowns represent constraint forces that
would enforce the constraints exactly should they be applied to the unconstrained system.
For each method the exposition tries to give first the basic flavor by working out the same example
for each method. The general technique is subsequently presented in matrix form for completeness
but is considered an advanced topic.
Conceptually, imposing MFCs is not different from the procedure discussed in Chapter 3 for single-
freedom constraints. The master stiffness equations are assembled ignoring all constraints. Then
the MFCs are imposed by appropriate modification of those equations. There are, however, two
important practical differences:
1. The modification process is not unique because there are alternative constraint imposition
methods, namely those listed above. These methods offer tradeoffs in generality, programming
implementation complexity, computational effort, numerical accuracy and stability.
2. In the implementation of some of these methods — notably penalty augmentation — constraint
imposition and assembly are carried out simultaneously. In that case the framework “first
assemble, then modify,” is not strictly respected in the actual implementation.

Remark 8.2. The three methods are also applicable, as can be expected, to the simpler case of a single-freedom
constraint such as (8.2). For most situations, however, the generality afforded by the penalty function and
Lagrange multiplier methods are not warranted. The hand-oriented reduction process discussed in Chapters 3
and 4 is in fact a special case of the master-slave elimination method in which “there is no master.”

Remark 8.3. Often both multifreedom and single-freedom constraints are prescribed. The modification
process then involves two stages: apply multifreedom constraints and apply single freedom constraints. The
order in which these are carried out is implementation dependent. Most implementations do the MFCs first,
either after the assembly is completed or during the assembly process. The reason is practical: single-freedom
constraints are often automatically taken care of by the equation solver itself.

§8.1.3. *MFC Matrix Forms

Matrix forms of linear MFCs are often convenient for compact notation. An individual constraint such as the
second one in (8.4) may be written
 
u x2
[1 −2 1] u x4 = 0.25. (8.6)
u x6

In direct matrix notation:

āi ūi = gi , (no sum on i) (8.7)
in which index i (i = 1, 2, . . .) identifies the constraint, āi is a row vector, ūi collects the set of degrees of
freedom that participate in the constraint, and gi is the right hand side scalar (0.25 in the foregoinf example).
The bars over a and u distinguishes (8.7) from the expanded form (8.9) discussed below.
For method description and general proofs it is often convenient to expand matrix forms so that they embody all
degrees of freedom. For example, if (8.6) is part of a two-dimensional finite element model with 12 freedoms:

8–5
Chapter 8: MULTIFREEDOM CONSTRAINTS I

u 1 , f1 u 2 , f2 u 3 , f3 u4 , f4 u5 , f5 u6 , f 6 u7 , f7
(1) (2) (3) (4) (5) (6)
x
1 2 3 4 5 6 7
Figure 8.2. A one-dimensional problem discretized with six bar finite
elements. The seven nodes may move only along the x direction. Subscript
x is omitted from the u’s and f ’s to reduce clutter.

u x1 , u y1 , . . . u y6 , the left-hand side row vector may be expanded with nine zeros as follows
u 
x1

 u y1 
 
[ 0 0 1 0 0 0 −2 0 0 0 1 0 ]  u x2  = 0.25, (8.8)
 .. 
.
u y6
in which case the matrix notation
ai u = gi (8.9)
is used. Finally, all multifreedom constraints expressed as (8.9) may be collected into a single matrix relation:

A u = g, (8.10)

in which rectangular matrix A is formed by stacking the ai ’s as rows and column vector g is formed by stacking
the gi s as entries. If there are 12 degrees of freedom in u and 5 multifreedom constraints then A will be 5 × 12.

§8.2. The Example Structure

The one-dimensional finite element discretization shown in Figure 8.2 will be used throughout
Chapters 8 and 9 to illustrate the three MFC application methods. This structure consists of six bar
elements connected by seven nodes that can only displace in the x direction.
Before imposing various multifreedom constraints discussed below, the master stiffness equations
for this problem are assumed to be
    
K 11 K 12 0 0 0 0 0 u1 f1
 12
K K 22 K 23 0 0 0 0   u2   f2 
    
 0 K 23 K 33 K 34 0 0 0   u3   f3 
    
 0 0 K 34 K 44 K 45 0 0   u4  =  f4  , (8.11)
    
 0 0 0 K 45 K 55 K 56 0   u5   f5 
    
0 0 0 0 K 56 K 66 K 67 u6 f6
0 0 0 0 0 K 67 K 77 u7 f7
or
Ku = f. (8.12)
The nonzero stiffness coefficients K i j in (8.11) depend on the bar rigidity properties. For example, if
E e Ae /L e = 100 for each element e = 1, . . . , 6, then K 11 = K 77 = 100, K 22 = . . . = K 66 = 200,
K 12 = K 23 = . . . = K 67 = −100. However, for the purposes of the following treatment the

8–6
 §8.3 THE MASTER-SLAVE METHOD

coefficients may be kept arbitrary. The component index x in the nodal displacements u and nodal
forces f has been omitted for brevity.
Now let us specify a multifreedom constraint that states that nodes 2 and 6 must move by the same
amount:
u2 = u6. (8.13)
Passing all node displacements to the right hand side gives the canonical form:

u 2 − u 6 = 0. (8.14)

Constraint conditions of this type are sometimes called rigid links because they can be mechanically
interpreted as forcing node points 2 and 6 to move together as if they were tied by a rigid member.1
We now study the imposition of constraint (8.14) on the master equations (8.11) by the methods
mentioned above. In this Chapter the master-slave method is treated. The other two methods:
penalty augmentation and Lagrange multiplier adjunction, are discussed in the following Chapter.

§8.3. The Master-Slave Method

To apply this method by hand, the MFCs are taken one at a time. For each constraint a slave degree
of freedom is chosen. The freedoms remaining in that constraint are labeled master. A new set
of degrees of freedom û is established by removing all slave freedoms from u. This new vector
contains master freedoms as well as those that do not appear in the MFCs. A matrix transformation
equation that relates u to û is generated. This equation is used to apply a congruent transformation
to the master stiffness equations. This procedure yields a set of modified stiffness equations that
are expressed in terms of the new freedom set û. Because the modified system does not contain the
slave freedoms, these have been effectively eliminated.

§8.3.1. A One-Constraint Example

The mechanics of the process is best seen by going through an example. To impose (8.14) pick u 6
as slave and u 2 as master. Relate the original unknowns u 1 , . . . u 7 to the new set in which u 6 is
missing:
   
u1 1 0 0 0 0 0  
u1
 u2   0 1 0 0 0 0 
     u
 u3   0 0 1 0 0 0   2 
   u 
 u4  =  0 0 0 1 0 0   3  , (8.15)
     u4 
 5 
u 0 0 0 0 1 0  
    u5
u6 0 1 0 0 0 0
u7
u7 0 0 0 0 0 1
This is the required transformation relation. In compact form:

u = Tû . (8.16)

1 This physical interpretation is exploited in the penalty method described in the next Chapter. In two and three dimensions
rigid link constraints are more complicated.

8–7
Chapter 8: MULTIFREEDOM CONSTRAINTS I

Replacing (8.15) into (8.12) and premultiplying by TT yields the modified system

K̂ û = f̂, in which K̂ = TT K T, f̂ = TT f. (8.17)

Carrying out the indicated matrix multiplications yields

    
K 11 K 12 0 0 0 0 u1 f1
 K 12 K 22 + K 66 K 23 0 K 56 K 67   u 2   f 2 + f 6 
    
 0 K 23 K 33 K 34 0 0   u 3   f3 
   =  , (8.18)
 0 0 K 34 K 44 K 45 0   u 4   f4 
    
0 K 56 0 K 45 K 55 0 u5 f5
0 K 67 0 0 0 K 77 u7 f7
Equation (8.18) is a new linear system containing 6 equations in the remaining 6 unknowns: u 1 ,
u 2 , u 3 , u 4 , u 5 and u 7 . Upon solving it, u 6 is recovered from the constraint (8.13).
Remark 8.4. The form of modified system (8.17) can be remembered by a simple mnemonic rule: premultiply
both sides of T û = u by TT K, and replace K u by f on the right hand side.

Remark 8.5. For a simple freedom constraint such as u 4 = 0 the only possible choice of slave is of course
u 4 and there is no master. The congruent transformation is then nothing more than the elimination of u 4 by
striking out rows and columns from the master stiffness equations.

Remark 8.6. For a simple MFC such as u 2 = u 6 , it does not matter which degree of freedom is chosen as
master or unknown. Choosing u 2 as slave produces a system of equations in which now u 2 is missing:
K 0 0 0 K 12 0
 u1
  f1

11
 0 K 33 K 34 0 K 23 0   u 3   f3 
 0 0     
 K 34 K 44 K 45 0   u 4  =  f4  . (8.19)
 0 0 K 45 K 55 K 56 0   u 5   f5 
    
K 12 K 23 0 K 56 K 22 + K 66 K 67 u6 f2 + f6
0 0 0 0 K 67 K 77 u7 f7
Although (8.18) and (8.19) are algebraically equivalent, the latter would be processed faster if a skyline solver
(Part III of course) is used for the modified equations.

§8.3.2. Multiple Homogeneous MFCs

The matrix equation (8.17) in fact holds for the general case of multiple homogeneous linear
constraints. Direct establishment of the transformation equation, however, is more complicated if
slave freedoms in one constraint appear as masters in another. To illustrate this point, suppose that
for the example system we have three homogeneous multifreedom constraints:

u 2 − u 6 = 0, u 1 + 4u 4 = 0, 2u 3 + u 4 + u 5 = 0, (8.20)

Picking as slave freedoms u 6 , u 4 and u 3 from the first, second and third constraint, respectively, we
can solve for them as

u6 = u2, u 4 = − 14 u 1 , u 3 = − 12 (u 4 + u 5 ) = 18 u 1 − 12 u 5 . (8.21)

8–8
 §8.3 THE MASTER-SLAVE METHOD

Observe that solving for u 3 from the third constraint brings u 4 to the right-hand side. But because
u 4 is also a slave freedom (it was chosen as such for the second constraint) it is replaced in favor of
u 1 using u 4 = − 14 u 1 . The matrix form of the transformation (8.21) is
   1 0 0 0
u1
 u2   0 1 0 0 
   
0  u1
 u3   0 − 12
1
   81    u2 
 u4  =  − 4
   0 0 0  u5  , (8.22)
 u5   0 
   0 1 0  u7
u6 
0 1 0 0
u7 0 0 0 1
The modified system is now formed through the congruent transformation (8.17). Note that the
slave freedoms selected from each constraint must be distinct; for example the choice u 6 , u 4 , u 4
would be inadmissible as long as the constraints are independent. This rule is easy to enforce when
slave freedoms are chosen by hand, but can lead to implementation and numerical difficulties when
it is programmed as an automated procedure, as further discussed later.
Remark 8.7. The three MFCs (8.20) with u 6 , u 4 and u 2 chosen as slaves and u 1 , u 2 and u 5 chosen as masters,
may be presented in the partitioned matrix form:
     
0 0 1 u3 0 1 0 u1
0 4 0 u4 = −1 0 0 u2 (8.23)
2 1 0 u6 0 0 −1 u5

This may be compactly written As us + Am um = 0. Solving for the slave freedoms gives us = −A−1 s A m um .
Expanding with zeros to fill out u and û produces (8.22). This general matrix form is considered in §8.4.4.
Note that non-singularity of As is essential for this method to work.

§8.3.3. Nonhomogeneous MFCs

Extension to non-homogeneous constraints is immediate. In this case he transformation equation
becomes non-homogeneous. For example suppose that (8.14) has a nonzero prescribed value:

u 2 − u 6 = 0.2 (8.24)

Nonzero RHS values such as 0.2 in (8.24) may be often interpreted physically as “gaps” (thus the
use of the symbol g in the matrix form). Chose u 6 again as slave: u 6 = u 2 − 0.2, and build the
transformation      
u1 1 0 0 0 0 0   0
u1
 u2   0 1 0 0 0 0   0 
   u   
 u3   0 0 1 0 0 0   2   0 
     u3   
 u4  =  0 0 0 1 0 0    +  0  . (8.25)
     u4   
 5 
u 0 0 0 0 1 0    0 
    u5  
u6 0 1 0 0 0 0 −0.2
u7
u7 0 0 0 0 0 1 0
In compact matrix notation,
u = T û + g. (8.26)

8–9
Chapter 8: MULTIFREEDOM CONSTRAINTS I

Here the constraint gap vector g is nonzero and T is the same as before. To get the modified system
applying the shortcut rule of Remark 8.4, premultiply both sides of (8.26) by TT K, replace Ku by
f, and pass the data to the RHS:

K̂ û = f̂, in which K̂ = TT K T, f̂ = TT (f − K g). (8.27)

Upon solving (8.27) for û, the complete displacement vector is recovered from (8.26). For the MFC
(8.24) this technique gives the system
    
K 11 K 12 0 0 0 0 u1 f1
 K 12 K 22 + K 66 K 23 0 K 56 K 67   u 2   f 2 + f 6 − 0.2K 66 
    
 0 K 23 K 33 K 34 0 0   u3   f3 
   =  . (8.28)
 0 0 K 34 K 44 K 45 0   u4   f4 
    
0 K 56 0 K 45 K 55 0 u5 f 5 − 0.2K 56
0 K 67 0 0 0 K 77 u7 f 7 − 0.2K 67

See Exercise 8.2 for multiple non-homogeneous MFCs.

§8.3.4. *The General Case

For implementation in general-purpose programs the master-slave method can be described as follows. The
degrees of freedoms in u are classified into three types: independent or unconstrained, masters and slaves.
(The unconstrained freedoms are those that do not appear in any MFC.) Label these sets as uu , um and us ,
respectively, and partition the stiffness equations accordingly:
    
Kuu Kum Kus uu fu
T
Kum Kmm Kms um = fm (8.29)
T T
Kus Kms Kss us fs

The MFCs may be written in matrix form as

A m u m + A s us = g A , (8.30)

where As is assumed square and nonsingular. If so we can solve for the slave freedoms:

us = −A−1 −1 def
s Am um + As g A = T um + g, (8.31)

Inserting into the partitioned stiffness equations (8.30) and symmetrizing yields




Kuu Kum + Kus T uu fu − Kus g
= (8.32)
symm Kmm + T Kms
T T
+ Kms T + TT Kss T um fm − Kms g

It is seen that the misleading simplicity of the handworked examples is gone.

§8.3.5. *Retaining the Original Freedoms

A potential disadvantage of the master-slave method in computer work is that it requires a rearrangement of
the original stiffness equations because û is a subset of u. The disadvantage can be annoying when sparse
matrix storage schemes are used for the stiffness matrix, and becomes intolerable if secondary storage is used
for that purpose.

8–10
 §8.4 MODEL REDUCTION BY KINEMATIC CONSTRAINTS

u 1 , f1 u 2 , f2 u 3 , f3 u4 , f4 u5 , f5 u6 , f 6 u7 , f7
(1) (2) (3) (4) (5) (6)
x
1 2 3 4 5 6 7
u 1, f1 2 master DOFs to be retained u 7, f7

x
1 7
5 slave DOFs to be eliminated
Master Master
u 1, f1 u 7, f7
Reduced model
x
1 7
Figure 8.3. Model reduction of the example structure of Figure 8.2 to the end freedoms.

With a bit of trickery it is possible to maintain the original freedom ordering. Let us display it for the example
problem under (8.14). Instead of (8.15), use the square transformation
    
u1 1 0 0 0 0 0 0 u1
 2  0
u 1 0 0 0 0 0   u2 
 u3   0 0  
   0 1 0 0 0   u3 
 u4  =  0 0
 
   0 0 1 0 0  u4  , (8.33)
 u5   0 0 0 0 1 0 0   u5 
    
u6 0 1 0 0 0 0 0 ũ 6
u7 0 0 0 0 0 0 1 u7
in which ũ 6 is a placeholder for the slave freedom u 6 . The modified equations are
    
K 11 K 12 0 0 0 0 0 0 u1 f1
 12
K K 22 + K 66 K 23 0 0 K 56 0 K 67   u 2   f 2 + f 6 
 0 0     
 K 23 K 33 K 34 0 0 0   u 3   f3 
 0 0   u 4  =  f4 
   
 0 K 34 K 44 K 45 0 0 , (8.34)
 0 K 56 0 K 45 K 55 0 0 0   u 5   f5 
    
0 0 0 0 0 0 0 0 ũ 6 0
0 K 67 0 0 0 0 0 K 77 u7 f7
which are submitted to the equation solver. If the solver is not trained to skip zero rows and columns, a one
should be placed in the diagonal entry for the ũ 6 (sixth) equation. The solver will return ũ 6 = 0, and this
placeholder value is replaced by u 2 . Note several points in common with the computer-oriented placeholder
technique described in §3.4 to handle single-freedom constraints.

§8.4. Model Reduction by Kinematic Constraints

The congruent transformation equations (8.17) and (8.27) have additional applications beyond the
master-slave method. An important one is model reduction by kinematic constraints. Through
this procedure the number of DOF of a static or dynamic FEM model is reduced by a significant
number, typically to 1% – 10% of the original number. This is done by taking a lot of slaves and a
few masters. Only the masters are left after the transformation. The reduced model is commonly

8–11
Chapter 8: MULTIFREEDOM CONSTRAINTS I

(* Model Reduction Example *)

ClearAll[K11,K12,K22,K23,K33,K34,K44,K45,K55,K56,K66,
f1,f2,f3,f4,f5,f6];
K={{K11,K12,0,0,0,0,0},{K12,K22,K23,0,0,0,0},
{0,K23,K33,K34,0,0,0},{0,0,K34,K44,K45,0,0},
{0,0,0,K45,K55,K56,0},{0,0,0,0,K56,K66,K67},
{0,0,0,0,0,K67,K77}}; Print["K=",K//MatrixForm];
f={f1,f2,f3,f4,f5,f6,f7}; Print["f=",f];
T={{6,0},{5,1},{4,2},{3,3},{2,4},{1,5},{0,6}}/6;
Print["T (transposed to save space)=",Transpose[T]//MatrixForm];
Khat=Simplify[Transpose[T].K.T];
fhat=Simplify[Transpose[T].f];
Print["Modified Stiffness:"];
Print["Khat(1,1)=",Khat[[1,1]],"\nKhat(1,2)=",Khat[[1,2]],
"\nKhat(2,2)=",Khat[[2,2]] ];
Print["Modified Force:"];
Print["fhat(1)=",fhat[[1]]," fhat(2)=",fhat[[2]] ];

Figure 8.4. Mathematica script for the model reduction example of Figure 8.3.

used in subsequent calculations as component of a larger system, particularly during design or in

parameter identification.

Example 8.2. Consider the bar assembly of Figure 8.2. Assume that the only masters are the end motions u 1
and u 7 , as illustrated in Figure 8.2, and interpolate all freedoms linearly:

   
u1 1 0
 2 
u 5/6 1/6 
 u 3   4/6 2/6 
     u1
 u 4  =  3/6 3/6 
    u , or u = T û. (8.35)
 u 5   2/6 4/6  7
   
u6 1/6 5/6
u7 0 1

8–12
 §8.5 ASSESSMENT OF THE MASTER-SLAVE METHOD

The reduced-order-model (ROM) equations are K̂û = TT KTû = TT f = f̂, or in detail

   ˆ
K̂ 11 K̂ 17 u1 f1
= , (8.36)
K̂ 17 K̂ 77 u7 fˆ7
in which
K̂ 11 = 1
36
(36K 11 +60K 12 +25K 22 +40K 23 +16K 33 +24K 34 +9K 44 +12K 45 +4K 55 +4K 56 +K 66 ),
K̂ 17 = 1
(6K 12 +5K 22 +14K 23 +8K 33 +18K 34 +9K 44 +18K 45 +8K 55 +14K 56 +5K 66 +6K 67 ),
36
(8.37)
K̂ 77 = 1
36
(K 22 +4K 23 +4K 33 +12K 34 +9K 44 +24K 45 +16K 55 +40K 56 +25K 66 +60K 67 +36K 77 ),
fˆ1 = 1
6
(6 f 1 +5 f 2 +4 f 3 +3 f 4 +2 f 5 + f 6 ), fˆ7 = 16 ( f 2 +2 f 3 +3 f 4 +4 f 5 +5 f 6 +6 f 7 ).
This reduces the order of the FEM model from 7 to 2. A Mathematica script to do the reduction is shown in
Figure 8.4. The key feature is that the masters are picked a priori, as the freedoms to be retained in the model
for further use.

Remark 8.8. Model reduction can also be done by the static condensation method explained in Chapter 10.
As its name indicates, condensation is restricted to static analysis. On the other hand, for such problems it is
exact whereas model reduction by kinematic constraints generally introduces approximations.

§8.5. Assessment of the Master-Slave Method

What are the good and bad points of this constraint application method? It enjoys the advantage of
being exact (except for inevitable solution errors) and of reducing the number of unknowns. The
concept is also easy to explain and learn. The main implementation drawback is the complexity of
the general case as can be seen by studying (8.29) through (8.32). The complexity is due to three
factors:
1. The equations may have to be rearranged because of the disappearance of the slave freedoms.
This drawback can be alleviated, however, through the placeholder trick outlined in §8.3.5.
2. An auxiliary linear system, namely (8.31), has to be assembled and solved to produce the
transformation matrix T and vector g.
3. The transformation process may generate many additional matrix terms. If a sparse matrix
storage scheme is used for K, the logic for allocating memory and storing these entries can be
difficult and expensive.
The level of complexity depends on the generality allowed as well as on programming decisions. If
K is stored as full matrix and slave freedom coupling in the MFCs is disallowed the logic is simple.2
On the other hand, if arbitrary couplings are permitted and K is placed in secondary (disk) storage
according to some sparse scheme, the complexity can become overwhelming.
Another, more subtle, drawback of this method is that it requires decisions as to which degrees
of freedom are to be treated as slaves. This can lead to implementation and numerical stability
problems. Although for disjointed constraints the process can be programmmed in reliable form, in
more general cases of coupled constraint equations it can lead to incorrect decisions. For example,
suppose that in the example problem you have the following two MFCs:
1
u
6 2
+ 12 u 4 = u 6 , u 3 + 6u 6 = u 7 . (8.38)

2 This is the case in model reduction, since each slave freedom appears in one and only one MFC.

8–13
Chapter 8: MULTIFREEDOM CONSTRAINTS I

For numerical stability reasons it is usually better to pick as slaves the freedoms with larger co-
efficients. If this is done, the program would select u 6 as slave freedoms from both constraints.
This leads to a contradiction because having two constraints we must eliminate two slave degrees
of freedom, not just one. The resulting modified system would in fact be inconsistent. Although
this defect can be easily fixed by the program logic in this case, one can imagine the complexity
burden if faced with hundreds or thousands of MFCs.
Serious numerical problems can arise if the MFCs are not independent. For example:
1
u
6 2
= u6, 1
u
5 3
+ 6u 6 = u 7 , u 2 + u 3 − u 7 = 0. (8.39)

The last constraint is an exact linear combination of the first two. If the program blindly choses
u 2 , u 3 and u 7 as slaves, the modified system is incorrect because we eliminate three equations
when in fact there are only two independent constraints. Exact linear dependence, as in (8.39), can
be recognized by a rank analysis of the As matrix defined in (8.30). In floating-point arithmetic,
however, such detection may fail because that kind of computation is inexact by nature.3
The complexity of slave selection is in fact equivalent to that of automatically selecting kinematic
redundancies in the Force Method of structural analysis. It has led implementors of programs that
use this method to require masters and slaves be prescribed in the input data, thus transfering the
burden to users.
The method is not generally extendible to nonlinear constraints without case by case programming.
In conclusion, the master-slave method is useful when a few simple linear constraints are imposed
by hand. As a general purpose technique for finite element analysis it suffers from complexity and
lack of robustness. It is worth learning, however, because of the great importance of congruent
transformations in model reduction for static and dynamic problems.

Notes and Bibliography

Multifreedom constraints are treated in several of the FEM books recommended in §1.7.5, notably Zienkiewicz
and Taylor [830]. The master-slave method was incorporated to treat MFCs as part of the DSM developed at
Boeing during the 1950s. It is first summarily described in the DSM-overview by Turner, Martin and Weikel
[761, p. 212]. The implementation differs, however, from the one described here because the relation of FEM
to energy methods was not clear at the time.
The master-slave method became popular through its adoption by the general-purpose NASTRAN code de-
veloped in the late 1960s [20] and early assessments of its potential [745]. The implementation unfortunately
relied on user inputs to identify slave DOFs. Through this serious blunder the method gained a reputation for
unreliability that persists to the present day.
The important application of master-slave to model reduction, which by itself justifies teaching the method, is
rarely mentioned in FEM textbooks.

References
Referenced items have been moved to Appendix R.

3 The safest technique to identify dependencies is to do a singular-value decomposition (SVD) of As . This can be, however,
prohibitively expensive if one is dealing with hundreds or thousands of constraints.

8–14
 Exercises

Homework Exercises for Chapter 8

MultiFreedom Constraints I

EXERCISE 8.1 [C+N:20] The example structure of Figure 8.1 has E e Ae /L e = 100 for each element
e = 1, . . . , 6. Consequently K 11 = K 77 = 100, K 22 = . . . = K 66 = 200, K 12 = K 23 = . . . = K 67 = −100.
The applied node forces are taken to be f 1 = 1, f 2 = 2, f 3 = 3, f 4 = 4, f 5 = 5, f 6 = 6 and f 7 = 7, which
are easy to remember. The structure is subjected to one support condition: u 1 = 0 (a fixed left end), and to
one MFC: u 2 − u 6 = 1/5.
Solve this problem using the master-slave method to process the MFC, taking u 6 as slave. Upon forming the
modified system (8.27) apply the left-end support u 1 = 0 using the placeholder method of §3.4. Solve the
equations and verify that the displacement solution and the recovered node forces including reactions are

u = [ 0 0.270 0.275 0.250 0.185 0.070 0.140 ]T

(E8.1)
Ku = [ −27 26.5 3 4 5 −18.5 7 ]T

Use Mathematica or Matlab to do the algebra is recommended. For example, the Mathematica script of
Figure E8.1 solves this Exercise.

(* Exercise 8.1 - Master-Slave Method *)

(* MFC: u2-u6 = 1/5 - slave: u6 *)
MasterStiffnessOfSixElementBar[kbar_]:=Module[
{K=Table[0,{7},{7}]}, K[[1,1]]=K[[7,7]]=kbar;
For [i=2,i<=6,i++,K[[i,i]]=2*kbar];
For [i=1,i<=6,i++,K[[i,i+1]]=K[[i+1,i]]=-kbar];
Return[K]];
FixLeftEndOfSixElementBar[Khat_,fhat_]:=Module[
{Kmod=Khat,fmod=fhat}, fmod[[1]]=0; Kmod[[1,1]]=1;
Kmod[[1,2]]=Kmod[[2,1]]=0; Return[{Kmod,fmod}]];

K=MasterStiffnessOfSixElementBar[100];
Print["Stiffness K=",K//MatrixForm];
f={1,2,3,4,5,6,7}; Print["Applied forces=",f];
T={{1,0,0,0,0,0},{0,1,0,0,0,0},{0,0,1,0,0,0},
{0,0,0,1,0,0},{0,0,0,0,1,0},{0,1,0,0,0,0},
{0,0,0,0,0,1}};
Print["Transformation matrix T=",T//MatrixForm];
g={0,0,0,0,0,-1/5,0};
Print["Constraint gap vector g=",g];
Khat=Simplify[Transpose[T].K.T]; fhat=Simplify[Transpose[T].(f-K.g)];
{Kmod,fmod}=FixLeftEndOfSixElementBar[Khat,fhat]; (* fix left end *)
Print["Modified Stiffness upon fixing node 1:",Kmod//MatrixForm];
Print["Modified RHS upon fixing node 1:",fmod];
umod=LinearSolve[Kmod,fmod];
Print["Computed umod (lacks slave u6)=",umod];
u=T.umod+g; Print["Complete solution u=",u];
Print["Numerical u=",N[u]];
fu=K.u; Print["Recovered forces K.u with reactions=",fu];
Print["Numerical K.u=",N[fu]];

Figure E8.1. Script for solving Exercise 8.1.

8–15
Chapter 8: MULTIFREEDOM CONSTRAINTS I

EXERCISE 8.2 [C+N:25] As in the previous Exercise but applying the following three MFCs, two of which
are non-homogeneous:

u 2 − u 6 = 1/5, u 3 + 2u 4 = −2/3, 2u 3 − u 4 + u 5 = 0. (E8.2)

Hint. Chose u 4 , u 5 and u 6 as slaves. Much of the script shown for Exercise 8.1 can be reused. The main
changes are in the formation of T and g. If you are a Mathematica wizard (or willing to be one) those can be
automatically formed by the statements listed in Figure E8.2.

sol=Simplify[Solve[{u2-u6==1/5, u3+2*u4==-2/3,2*u3-u4+u5==0},{u4,u5,u6}]];
ums={u1,u2,u3,u4,u5,u6,u7}/.sol[[1]]; um={u1,u2,u3,u7};
T=Table[Coefficient[ums[[i]],um[[j]]],{i,1,7},{j,1,4}];
g=ums/.{u1->0,u2->0,u3->0,u4->0,u5->0,u6->0,u7->0};
Print["Transformation matrix T=",T//MatrixForm];
Print["Gap vector g=",g];

Figure E8.2. Script for partially solving Exercise 8.2.

If you do this, explain what it does and why it works. Otherwise form and enter T and g by hand. The
numerical results (shown to 5 places) should be

u = [ 0. 0.043072 −0.075033 −0.29582 −0.14575 −0.15693 −0.086928 ]T ,

(E8.3)
Ku = [ −4.3072 16.118 10.268 −37.085 16.124 −8.1176 7. ]T .

EXERCISE 8.3 [A:25] Can the MFCs be pre-processed to make sure that no slave freedom in a MFC appears
as master in another?

EXERCISE 8.4 [A:25] In the general case discussed in §8.4.4, under which condition is the matrix As of
(8.30) diagonal and thus trivially invertible?

EXERCISE 8.5 [A:25] Work out the general technique by which the unknowns need not be rearranged, that
is, u and û are the same. Use “placeholders” for the slave freedoms. (Hint: use ideas of §3.4).

EXERCISE 8.6 [A/C:35] Is it possible to establish a slave selection strategy that makes As diagonal or
triangular? (This requires knowledge of matrix techniques such as pivoting.)

EXERCISE 8.7 [A/C:40] Work out a strategy that produces a well conditioned As by selecting new slaves as
linear combinations of finite element freedoms if necessary. (Requires background in numerical analysis and
advanced programming experience in matrix algebra).

8–16
 9
MultiFreedom
Constraints II

9–1
Chapter 9: MULTIFREEDOM CONSTRAINTS II

TABLE OF CONTENTS

Page
§9.1 Introduction . . . . . . . . . . . . . . . . . . . . . 9–3
§9.2 The Penalty Method . . . . . . . . . . . . . . . . . . 9–3
§9.2.1 Physical Interpretation of Penalty Method . . . . . . . . 9–3
§9.2.2 Choosing the Penalty Weight . . . . . . . . . . . . 9–3
§9.2.3 The Square Root Rule . . . . . . . . . . . . . . 9–4
§9.2.4 Penalty Elements for General MFCs . . . . . . . . . 9–5
§9.2.5 *The Theory Behind the Penalty Method . . . . . . . . 9–6
§9.2.6 Assessment of the Penalty Method . . . . . . . . . . 9–7
§9.3 Lagrange Multiplier Adjunction . . . . . . . . . . . . . . 9–8
§9.3.1 Physical Interpretation . . . . . . . . . . . . . . 9–8
§9.3.2 Lagrange Multipliers for General MFCs . . . . . . . . 9–9
§9.3.3 *The Theory Behind Lagrange Multipliers . . . . . . . 9–10
§9.3.4 Assessment of the Lagrange Multiplier Method . . . . . . 9–10
§9.4 *The Augmented Lagrangian Method . . . . . . . . . . . 9–10
§9.5 Summary . . . . . . . . . . . . . . . . . . . . . . 9–11
§9. Notes and Bibliography . . . . . . . . . . . . . . . . . 9–11
§9. References . . . . . . . . . . . . . . . . . . . . . . 9–12
§9. Exercises . . . . . . . . . . . . . . . . . . . . . . 9–13

9–2
 §9.2 THE PENALTY METHOD

§9.1. Introduction

In this Chapter we continue the discussion of methods to treat multifreedom constraints (MFCs). The
master-slave method described previously was easy to explain, but exhibits serious shortcomings for
treating arbitrary constraints (although the method has important applications to model reduction).
We now pass to the study of two other methods: penalty augmentation and Lagrange multiplier
adjunction. Both techniques are better suited to general implementations of the Finite Element
Method, whether linear or nonlinear.

§9.2. The Penalty Method

u 1 , f1 u 2 , f2 u 3 , f3 u4 , f4 u5 , f5 u6 , f 6 u7 , f7
(1) (2) (3) (4) (5) (6)
x
1 2 3 4 5 6 7
Figure 9.1. The example structure of Chapter 8, repeated for convenience.

§9.2.1. Physical Interpretation of Penalty Method

The penalty method will be first presented using a physical argument, leaving the mathematical
formulation to a subsequent section. Consider again the example structure of Chapter 8, which
is reproduced in Figure 9.1 for convenience. To impose u 2 = u 6 imagine that nodes 2 and 6
are connected with a “fat” bar of axial stiffness w, labeled with element number 7, as shown in
Figure 9.2. This bar is called a penalty element and w is its penalty weight.
Such an element, albeit fictitious, can be treated exactly like another bar element insofar as con-
tinuing the assembly of the master stiffness equations. The penalty element stiffness equations,
K(7) u(7) = f(7) , are1     
1 −1 u2 0
w = (9.1)
−1 1 u6 0
Because there is one freedom per node, the two local element freedoms map into global freedoms 2
and 6, respectively. Using the assembly rules of Chapter 3 we obtain the following modified master
û = f̂, which shown in detail are
stiffness equations: K
    
K 11 K 12 0 0 0 0 0 u1 f1
 K 12 K 22 + w K 23 0 0 −w 0   u 2   f2 
    
 0 K 23 K 33 K 34 0 0 0   u 3   f3 
    
 0 0 K 34 K 44 K 45 0 0   u 4  =  f4  . (9.2)
    
 0 0 0 K 45 K 55 K 56 0  5  5
u f
    
0 −w 0 0 K 56 K 66 + w K 67 u6 f6
0 0 0 0 0 K 67 K 77 u7 f7

This system can now be submitted to the equation solver. Note that û ≡ u, and only K has changed.

1 The general method to construct these equations is described in §9.1.4.

9–3
Chapter 9: MULTIFREEDOM CONSTRAINTS II

u 1 , f1 u 2 , f2 u 3 , f3 u4 , f4 u5 , f5 u6 , f 6 u7 , f7
(1) (2) (3) (4) (5) (6)
x
1 2 3 4 5 6 7

(7)
penalty element of axial rigidity w
Figure 9.2. Adjunction of a fictitious penalty bar of axial stiffness w,
identified as element 7,to enforce the MFC u 2 = u 6 .

§9.2.2. Choosing the Penalty Weight

What happens when (9.2) is solved numerically? If a finite weight w is chosen the constraint
u 2 = u 6 is approximately satisfied in the sense that one gets u 2 − u 6 = eg , where eg = 0. The “gap
error” eg is called the constraint violation. The magnitude |eg | of this violation depends on the
weight: the larger w, the smaller the violation. More precisely, it can be shown that |eg | becomes
proportional to 1/w as w gets to be sufficiently large (see Exercises). For example, raising w
from, say, 106 to 107 can be expected to cut the constraint violation by roughly 10 if the physical
stiffnesses are small compared to w.
Therefore it seems as if the proper strategy should be: try to make w as large as possible while
respecting computer overflow limits. However, this is misleading. As the penalty weight w tends
to ∞ the modified stiffness matrix in (9.2) becomes more and more ill-conditioned with respect to
inversion.
To make this point clear, suppose for definiteness that the rigidities E e Ae /L e of the actual bars
e = 1, . . . 6 are unity, that w >> 1, and that the computer solving the stiffness equations has a
floating-point precision of 16 decimal places. Numerical analysts characterize such precision by
saying that f = O(10−16 ), where | f | is the smallest power of 10 that perceptibly adds to 1 in
floating-point arithmetic.2 The modified stiffness matrix of (9.2) becomes
 
1 −1 0 0 0 0 0
 −1 2 + w −1 0 0 −w 0 
 
 0 −1 2 −1 0 0 0 
=
K  0 0 −1 2 −1 0 0 

(9.3)
 
 0 0 0 −1 2 −1 0 
 
0 −w 0 0 −1 2 + w −1
0 0 0 0 0 −1 1
As w → ∞ rows 2 and 6, as well as columns 2 and 6, tend to become linearly dependent; in
fact the negative of each other. But linear dependency means singularity. Therefore K approaches
singularity as w → ∞. In fact, if w exceeds 1/ f = 1016 the computer will not be able to
 from an exactly singular matrix. If w << 1016 but w >> 1, the effect will be seen in
distinguish K
increasing solution errors affecting the computed displacements û returned by the equation solver.
These errors, however, tend to be more of a random nature than the constraint violation error.
2 Such definitions are more rigurously done by working with binary numbers and base-2 arithmetic but for the present
discussion the use of decimal powers is sufficient.

9–4
 §9.2 THE PENALTY METHOD

§9.2.3. The Square Root Rule

Plainly we have two effects at odds with each other. Making w larger reduces the constraint violation
error but increases the solution error. The best w is that which makes both errors roughly equal
in absolute value. This tradeoff value is difficult to find aside of systematically running numerical
experiments. In practice the heuristic square root rule is often followed.
This rule can be stated as follows. Suppose that the largest stiffness coefficient, before adding
penalty elements, is of the order of 10k and that the working machine precision is p digits.3 Then
choose penalty weights to be of order 10k+ p/2 with the proviso that such a choice would not cause
arithmetic overflow.4
For the above example in which k ≈ 0 and p ≈ 16, the optimal w given by this rule would be
w ≈ 108 . This w would yield a constraint violation and a solution error of order 10−8 . Note that
there is no simple way to do better than this accuracy aside from using extended (e.g., quad) floating-
point precision. This is not easy to do when using standard low-level programming languages.
√
The name “square root” arises because the recommended w is in fact 10k 10 p . It is seen that
picking the weight by this rule requires knowledge of both stiffness magnitudes and floating-point
hardware properties of the computer used, as well as the precision selected by the program.
§9.2.4. Penalty Elements for General MFCs

For the constraint u 2 = u 6 the physical interpretation of the penalty element is clear. Nodal points 2
and 6 must move in lockstep long x, which can be approximately enforced by the heavy bar device
shown in Figure 9.2. But how about 3u 3 + u 5 − 4u 6 = 1? Or just u 2 = −u 6 ?
The treatment of more general constraints is linked to the theory of Courant penalty functions,
which in turn is a topic in variational calculus. Because the necessary theory given in §9.1.5 is
viewed as an advanced topic, the procedure used for constructing a penalty element is stated here
as a recipe. Consider the homogeneous constraint

3u 3 + u 5 − 4u 6 = 0. (9.4)

Rewrite this equation in matrix form

 
u3
[3 1 −4 ] u 5 = 0, (9.5)
u6

and premultiply both sides by the transpose of the coefficient matrix:

        
3 u3 9 3 −12 u3 0
e e
1 [3 1 −4 ] u 5 = 3 1 −4 u5 = K̄ u = 0 . (9.6)
−4 u6 −12 −4 16 u6 0

3 Such order-of-magnitude estimates can be readily found by scanning the diagonal of K because the largest stiffness
coefficient of the actual structure is usually a diagonal entry.
4 If overflows occurs, the master stiffness should be scaled throughout or a better choice of physical units made.

9–5
Chapter 9: MULTIFREEDOM CONSTRAINTS II

e
Here K̄ is the unscaled stiffness matrix of the penalty element. This is now multiplied by the
penalty weight w and assembled into the master stiffness matrix following the usual rules. For the
example problem, augmenting (9.2) with the w-scaled penalty element (9.6) yields
    
K 11 K 12 0 0 0 0 0 u1 f1
 12
K K 22 K 23 0 0 0 0   u2   f2 
    
 0 K 23 K 33 + 9w K 34 3w −12w 0   u3   f3 
    
 0 0 K 34 K 44 K 45 0 0   u4  =  f4  . (9.7)
    
 0 0 3w K 45 K 55 + w K 56 − 4w 0   u5   f5 
    
0 0 −12w 0 K 56 − 4w K 66 + 16w K 67 u6 f6
0 0 0 0 0 K 67 K 77 u7 f7

If the constraint is nonhomogeneous the force vector is also modified. To illustrate this effect,
consider the MFC: 3u 3 + u 5 − 4u 6 = 1. Rewrite in matrix form as
 
u3
[3 1 −4 ] u 5 = 1. (9.8)
u6

Premultiply both sides by the transpose of the coefficient matrix:

    
9 3 −12 u3 3
3 1 −4 u5 = 1 . (9.9)
−12 −4 16 u6 −4

Scaling by w and assembling yields

    
K 11 K 12 0 0 0 0 0 u1 f1
 K 12 K 22 K 23 0 0 0 0   u2   f2 
    
 0 K 23 K 33 + 9w K 34 3w −12w 0   u 3   f 3 + 3w 
    
 0 0 K 34 K 44 K 45 0 0   u4  =  f4  .
    
 0 0 3w K 45 K 55 + w K 56 − 4w 0   u 5   f5 + w 
    
0 0 −12w 0 K 56 − 4w K 66 + 16w K 67 u6 f 6 − 4w
0 0 0 0 0 K 67 K 77 u7 f7
(9.10)
§9.2.5. *The Theory Behind the Penalty Method
The rule comes from the following mathematical theory. Suppose we have a set of m linear MFCs. Using the
matrix notation introduced in §8.1.3, these will be stated as

apu = bp, p = 1, . . . m (9.11)

where u contains all degrees of freedom and each a p is a row vector with same length as u. To incorporate
the MFCs into the FEM model one selects a weight w p > 0 for each constraints and constructs the so-called
Courant quadratic penalty function or “penalty energy”

m
1 
P= Pp , with Pp = u T 2
aTp a p u − w p aTp b p = 12 uT K( p) u − uT f( p) , (9.12)
p=1

9–6
 §9.2 THE PENALTY METHOD

where we have called K( p) = w p aTp a p and f( p) = w p aT bi . P is added to the potential energy function
 = 12 uT Ku − uT f to form the augmented potential energy a =  + P. Minimization of a with respect
to u yields

m

m
Ku + K( p) u = f + f( p) . (9.13)
p=1 p=1

Each term of the sum on p, which derives from term Pp in (9.12), may be viewed as contributed by a penalty
element with globalized stiffness matrix K( p) = w p aTp a p and globalized added force term f( p) = w p aTp b p .
To use a even more compact form we may write the set of multifreedom constraints as Au = b. Then the
penalty augmented system can be written compactly as

(K + AT WA) u = f + WAT b, (9.14)

where W is a diagonal matrix of penalty weights. This compact form, however, conceals the configuration of
the penalty elements.

§9.2.6. Assessment of the Penalty Method

The main advantage of the penalty function method is its straightforward computer implementation.
Looking at modified systems such as (9.2), (9.7) or (9.10) it is obvious that the master equations
need not be rearranged. That is, u and û are the same. Constraints may be programmed as “penalty
elements,” and stiffness and force contributions of these elements merged through the standard
assembler. In fact using this method there is no need to distinguish between unconstrained and
constrained equations! Once all elements — regular and penalty — are assembled, the system can
be passed to the equation solver.5
An important advantage with respect to the master-slave (elimination) method is its lack of sensitivity
with respect to whether constraints are linearly dependent. To give a simplistic example, suppose
that the constraint u 2 = u 6 appears twice. Then two penalty elements connecting 2 and 6 will be
inserted, doubling the intended weight but not otherwise causing undue harm.
An advantage with respect to the Lagrange multiplier method described in §9.2 is that positive
definiteness is not lost. Such loss can affect the performance of certain numerical processes.6
Finally, it is worth noting that the penalty method is easily extendible to nonlinear constraints
although such extension falls outside the scope of this book.
The main disadvantage, however, is a serious one: the choice of weight values that balance solution
accuracy with the violation of constraint conditions. For simple cases the square root rule previously
described often works, although its effective use calls for knowledge of the magnitude of stiffness
coefficients. Such knowledge may be difficult to extract from a general purpose “black box”
program. For difficult cases selection of appropriate weights may require extensive numerical
experimentation, wasting the user time with numerical games that have no bearing on the actual
objective, which is getting a solution.
The deterioration of the condition number of the penalty-augmented stiffness matrix can have
serious side effects in some solution procedures such as eigenvalue extraction or iterative solvers.

5 Single freedom constraints, such as those encountered in Chapter 3, are usually processed separately for efficiency.
6 For example, solving the master stiffness equations by Cholesky factorization or conjugate-gradients.

9–7
Chapter 9: MULTIFREEDOM CONSTRAINTS II

u 1 , f1 u 2 , f2 u 3 , f3 u4 , f4 u5 , f5 u6 , f 6 u7 , f7
(1) (2) (3) (4) (5) (6)
x
1 2 3 4 5 6 7
−λ λ
Figure 9.3. Physical interpretation of Lagrange multiplier
adjunction to enforce the MFC u 2 = u 6 .

Finally, even if optimal weights are selected, the combined solution error cannot be lowered beyond
a threshold value.
From this assessment it is evident that penalty augmentation, although superior to the master-slave
method from the standpoint of generality and ease of implementation, is no panacea.

§9.3. Lagrange Multiplier Adjunction

§9.3.1. Physical Interpretation
As in the case of the penalty function method, the method of Lagrange multipliers can be given a
rigorous justification within the framework of variational calculus. But in the same spirit it will be
introduced for the example structure from a physical standpoint that is particularly illuminating.
Consider again the constraint u 2 = u 6 . Borrowing some ideas from the penalty method, imagine
that nodes 2 and 6 are connected now by a rigid link rather than a flexible one. Thus the constraint
is imposed exactly. But of course the penalty method with an infinite weight would “blow up.”
We may remove the link if it is replaced by an appropriate reaction force pair (−λ, +λ), as illustrated
in Figure 9.3. These are called the constraint forces. Incorporating these forces into the original
stiffness equations (8.10) we get
    
K 11 K 12 0 0 0 0 0 u1 f1
 K 12 K 22 K 23 0 0 0 0   u 2   f2 − λ 
    
 0 K 23 K 33 K 34 0 0 0   u 3   f3 
    
 0 0 K 34 K 44 K 45 0 0   u 4  =  f4  . (9.15)
    
 0 0 0 K 45 K 55 K 56 0   u 5   f5 
    
0 0 0 0 K 56 K 66 K 67 u6 f6 + λ
0 0 0 0 0 K 67 K 77 u7 f7

This λ is called a Lagrange multiplier. Because λ is an unknown, let us transfer it to the left hand
side by appending it to the vector of unknowns:
 
  u1  
K 11 K 12 0 0 0 0 0 0 f1
 u2 
 K 12 K 22 K 23 0 0 0 0 1     f2 
 u   
 0 K 23 K 33 K 34 0 0 0 0   3   f3 
 u   
 0 0 K 34 K 44 K 45 0 0 0   4  =  f4  . (9.16)
   u5   
 0 0 0 K 45 K 55 K 56 0 0    5 f
 u   
0 0 0 0 K 56 K 66 K 67 −1  6  f6
u7
0 0 0 0 0 K 67 K 77 0 f7
λ

9–8
 §9.3 LAGRANGE MULTIPLIER ADJUNCTION

But now we have 7 equations in 8 unknowns. To render the system determinate, the constraint
condition u 2 − u 6 = 0 is appended as eighth equation:

    
K 11 K 12 0 0 0 0 0 0 u1 f1
 K 12 K 22 K 23 0 0 0 0 1   u2   f2 
    
 0 K 23 K 33 K 34 0 0 0 0   u3   f3 
    
 0 0 K 34 K 44 K 45 0 0 0   u4   f4 
   =  , (9.17)
 0 0 0 K 45 K 55 K 56 0 0   u5   f5 
    
 0 0 0 0 K 56 K 66 K 67 −1   u 6   f6 
    
0 0 0 0 0 K 67 K 77 0 u7 f7
0 1 0 0 0 −1 0 0 λ 0

This is called the multiplier-augmented system. Its coefficient matrix, which is symmetric, is
called the bordered stiffness matrix. The process by which λ is appended to the vector of original
unknowns is called adjunction. Solving this system provides the desired solution for the degrees
of freedom while also characterizing the constraint forces through λ.

§9.3.2. Lagrange Multipliers for General MFCs

The general procedure will be stated first as a recipe. Suppose that we want to solve the example
structure subjected to three MFCs

u 2 − u 6 = 0, 5u 2 − 8u 7 = 3, 3u 3 + u 5 − 4u 6 = 1, (9.18)

Adjoin these MFCs as the eighth, nineth and tenth equations:

   
K 11 K 12 0 0 0 0 0 f1
 K 12 K 22 K 23 0 0 0 0    f2 
  u1  
 0 K 23 K 33 K 34 0 0 0   f3 
   u2   
 0 0 K 34 K 44 K 45 0 0    f4 
 u   
 0 0 0 K 45 K 55 K 56 0  3  f5 
 u  =  , (9.19)
 0 0 0 0 K 56 K 66 K 67   4   f6 
 u   
 0 0 0 0 0 K 67 K 77   5   f7 
  u6  
 0 1 0 0 0 −1 0   0
  u7  
0 5 0 0 0 0 −8 3
0 0 3 0 1 −4 0 1

Three Lagrange multipliers: λ1 , λ2 and λ3 , are required to take care of three MFCs. Adjoin those
unknowns to the nodal displacement vector. Symmetrize the coefficient matrix by appending 3
columns that are the transpose of the 3 last rows in (9.19), and filling the bottom right-hand corner

9–9
Chapter 9: MULTIFREEDOM CONSTRAINTS II

with a 3 × 3 zero matrix:

    
K 11 K 12 0 0 0 0 0 0 0 0 u1 f1
 K 12 K 22 K 23 0 0 0 0 1 5 0   u2   f2 
    
 0 K 23 K 33 K 34 0 0 0 0 0 3   u3   f3 
    
 0 0 K 34 K 44 K 45 0 0 0 0 0   u4   f4 
    
 0 0 0 K 45 K 55 K 56 0 0 0 1   u5   f5 
   =  . (9.20)
 0 0 0 0 K 56 K 66 K 67 −1 0 −4   u 6   f6 
    
 0 0 0 0 0 K 67 K 77 0 −8 0   u7   f7 
    
 0 1 0 0 0 −1 0 0 0 0   λ1   0
    
0 5 0 0 0 0 −8 0 0 0 λ2 3
0 0 3 0 1 −4 0 0 0 0 λ3 1

§9.3.3. *The Theory Behind Lagrange Multipliers

The recipe illustrated by (9.20) comes from a well known technique of variational calculus. Using the matrix
notation introduced in §8.1.3, compactly denote the set of m MFCs by Au = b, where A is m × n. The
potential energy of the unconstrained finite element model is  = 12 uT Ku − uT f. To impose the constraint,
adjoin m Lagrange multipliers collected in vector λ and form the Lagrangian

L(u, λ) =  + λT (Au − b) = 12 uT Ku − uT f + λT (Au − b). (9.21)

Extremizing L with respect to u and λ yields the multiplier-augmented form

    
K AT u f
= . (9.22)
A 0 λ b

The master stiffness matrix K in (9.22) is said to be bordered with A and AT . Solving this system provides u
and λ. The latter can be interpreted as forces of constraint in the following sense: a removed constraint can
be replaced by a system of forces characterized by λ multiplied by the constraint coefficients. More precisely,
the constraint forces are −AT λ.
§9.3.4. Assessment of the Lagrange Multiplier Method
In contrast to the penalty method, the method of Lagrange multipliers has the advantage of being
exact (aside from computational errors due to finite precision arithmetic). It provides directly the
constraint forces, which are of interest in many applications. It does not require guesses as regards
weights. As the penalty method, it can be extended without difficulty to nonlinear constraints.
It is not free of disadvantages. It introduces additional unknowns, requiring expansion of the original
stiffness method, and more complicated storage allocation procedures. It renders the augmented
stiffness matrix indefinite, an effect that may cause grief with some linear equation solving methods
that rely on positive definiteness. Finally, as the master-slave method, it is sensitive to the degree
of linear independence of the constraints: if the constraint u 2 = u 6 is specified twice, the bordered
stiffness is obviously singular.
On the whole this method appears to be the most elegant one for a general-purpose finite element
program that is supposed to work as a “black box” by minimizing guesses and choices from its users.
Its implementation, however, is not simple. Special care must be exercised to detect singularities
due to constraint dependency and to account for the effect of loss of positive definiteness of the
bordered stiffness on equation solvers.

9–10
 §9. Notes and Bibliography

§9.4. *The Augmented Lagrangian Method

The general matrix forms of the penalty function and Lagrangian multiplier methods are given by expressions
(9.13) and (9.22), respectively. A useful connection between these methods can be established as follows.
Because the lower diagonal block of the bordered stiffness matrix in (9.22) is null, it is not possible to directly
eliminate λ. To make this possible, replace this block by S−1 , where S is a constraint-scaling diagonal matrix
of appropriate order and is a small number. The reciprocal of is a large number called w = 1/ . To
maintain exactness of the second equation, S−1 λ is added to the right-hand side:
    
K AT u f
= (9.23)
A S−1 λ −1 P
S λ

Here superscript P (for “predicted value”) is attached to the λ on the right-hand side as a “tracer.” We can
now formally solve for λ and subsequently for u. The results may be presented as

(K + wAT SA) u = f + wAT Sb − AT λ P ,

(9.24)
λ = λ P + wS(b − Au),

Setting λ P = 0 in the first matrix equation yields

(K + wAT SA) u = f + wAT Sb. (9.25)

On taking W = wS, the general matrix equation (9.13) of the penalty method is recovered.
This relation suggests the construction of iterative procedures in which one tries to improve the accuracy
of the penalty function method while w is kept constant [?]. This strategy circumvents the aforementioned
ill-conditioning problems when the weight w is gradually increased. One such method is easily constructed
by inspecting (9.24). Using superscript k as an iteration index and keeping w fixed, solve equations (9.24) in
tandem as follows:
(K + AT WA) uk = f + AT Wb − AT λk ,
(9.26)
λk+1 = λk + W(b − Auk ),

for k = 0, 1, . . . , beginning with λ0 = 0. Then u0 is the penalty solution. If the process converges one
recovers the exact Lagrangian solution without having to solve the Lagrangian system (9.23) directly.
The family of iterative procedures that may be precipitated from (9.24) collectively pertains to the class of
augmented Lagrangian methods.

§9.5. Summary

The treatment of linear MFCs in finite element systems can be carried out by several methods. Three
of these: master-slave elimination, penalty augmentation and Lagrange multiplier adjunction, have
been discussed. It is emphasized that no method is uniformly satisfactory in terms of generality,
robustness, numerical behavior and simplicity of implementation.
Figure 9.4 gives an assessment of the three techniques in terms of seven attributes.
For a general purpose program that tries to attain “black box” behavior (that is, minimal decisions
on the part of users) the method of Lagrange multipliers has the edge. This edge is unfortunately
blunted by a fairly complex computer implementation and by the loss of positive definiteness in the
bordered stiffness matrix.

9–11
Chapter 9: MULTIFREEDOM CONSTRAINTS II

Master-Slave Penalty Lagrange

Elimination Function Multipliers

Generality fair excellent excellent

Ease of implementation poor to fair good fair

Sensitivity to user decisions high high small to none

Accuracy variable mediocre excellent

Sensitivity as regards high none high

constraint dependence

Retains positive definiteness yes yes no

Modifies unknown vector yes no yes

Figure 9.4. Assessment summary of three MFC application methods.

Notes and Bibliography

A form of the penalty function method, quite close to that described in §9.1.5, was first proposed by Courant
in the early 1940s [154]. It entered the FEM through the work of researchers in the 1960s. There is a good
description in the book by Zienkiewicz and Taylor [837].
The Lagrange Multiplier method is much older. Multipliers (called initially “coefficients”) were described
by Lagrange in his famous Mécanique Analytique monograph [435], as part of the procedure for forming the
function now called the Lagrangian. Its use in FEM is more recent than penalty methods.
Augmented Lagrangian methods have received much attention since the late 1960s, when they originated in
the field of constrained optimization. The original papers are by Hestenes [360] and Powell [598]. The use
of the Augmented Lagrangian Multiplier method for FEM kinematic constraints is first discussed in [212],
wherein the iterative algorithm (9.26) for the master stiffness equations is derived.

References
Referenced items have been moved to Appendix R.

9–12
 Exercises

Homework Exercises for Chapter 9

MultiFreedom Constraints II

EXERCISE 9.1 [C+N:20] This is identical to Exercise 8.1, except that the MFC u 2 − u 6 = 1/5 is to be
treated by the penalty function method. Take the weight w to be 10k , in which k varies as k = 3, 4, 5, . . . 16.
For each sample w compute the Euclidean-norm solution error e(w) = ||u p (w) − uex ||2 , where u p is the
computed solution and uex is the exact solution listed in (E8.1). Plot k = log10 w versus log10 e and report for
which weight e attains a minimum. (See Slide #5 for a check). Does it roughly agree with the square root rule
(§9.1.3) if the computations carry 16 digits of precision?
As in Exercise 8.1, use Mathematica, Matlab (or similar) to do the algebra. For example, the following
Mathematica script solves this Exercise:

(* Exercise 9.1 - Penalty Method *)

(* MFC: u2-u6=1/5 variable w *)
K=MasterStiffnessOfSixElementBar[100];
Print["Stiffness K=",K//MatrixForm];
f={1,2,3,4,5,6,7}; Print["Applied forces=",f];
uexact= {0,0.27,0.275,0.25,0.185,0.07,0.14}; ew={};
For [w=100, w<=10^16, w=10*w; (* increase w by 10 every pass *)
Khat=K; fhat=f;
Khat[[2,2]]+=w; Khat[[6,6]]+=w; Khat[[6,2]]=Khat[[2,6]]-=w;
fhat[[2]]+=(1/5)*w; fhat[[6]]-=(1/5)*w; (*insert penalty *)
{Kmod,fmod}=FixLeftEndOfSixElementBar[Khat,fhat];
u=LinearSolve[N[Kmod],N[fmod]];
Print["Weight w=",N[w]//ScientificForm," u=",u//InputForm];
e=Sqrt[(u-uexact).(u-uexact)];
(*Print["L2 solution error=",e//ScientificForm]; *)
AppendTo[ew,{Log[10,w],Log[10,e]}];
];
ListPlot[ew,AxesOrigin->{5,-8},Frame->True, PlotStyle->
{AbsolutePointSize[4],AbsoluteThickness[2],RGBColor[1,0,0]},
PlotJoined->True,AxesLabel->{"Log10(w)","Log10(u error)"}];

Here MasterStiffnessOfSixElementBar and FixLeftEndOfSixElementBar are the same modules listed

in Exercise 8.1.
Note: If you run the above program, you may get several beeps from Mathematica as it is processing some of
the systems with very large weights. Don’t be alarmed: those are only warnings. The LinearSolve function
is alerting you that the coefficient matrices K̂ for weights of order 1012 or bigger are ill-conditioned.

EXERCISE 9.2 [C+N:15] Again identical to Exercise 8.1, except that the MFC u 2 − u 6 = 1/5 is to be
treated by the Lagrange multiplier method. The results for the computed u and the recovered force vector Ku
should agree with (E8.1). Use Mathematica, Matlab (or similar) to do the algebra. For example, the following
Mathematica script solves this Exercise:

(* Exercise 9.2 - Lagrange Multiplier Method *)

(* MFC: u2-u6=1/5 *)
K=MasterStiffnessOfSixElementBar[100];
Khat=Table[0,{8},{8}]; f={1,2,3,4,5,6,7}; fhat=AppendTo[f,0];
For [i=1,i<=7,i++, For[j=1,j<=7,j++, Khat[[i,j]]=K[[i,j]] ]];
{Kmod,fmod}=FixLeftEndOfSixElementBar[Khat,fhat];

9–13
Chapter 9: MULTIFREEDOM CONSTRAINTS II

Kmod[[2,8]]=Kmod[[8,2]]= 1;
Kmod[[6,8]]=Kmod[[8,6]]=-1; fmod[[8]]=1/5;
Print["Kmod=",Kmod//MatrixForm];
Print["fmod=",fmod];
umod=LinearSolve[N[Kmod],N[fmod]]; u=Take[umod,7];
Print["Solution u=",u ,", lambda=",umod[[8]]];
Print["Recovered node forces=",K.u];

Here MasterStiffnessOfSixElementBar and FixLeftEndOfSixElementBar are the same modules listed

in Exercise 8.1.
Does the computed solution agree with (E8.1)?

EXERCISE 9.3 [A:10] For the example structure, show which penalty elements would implement the fol-
lowing MFCs:
(a) u 2 + u 6 = 0,
(E9.1)
(b) u 2 − 3u 6 = 1/3.
As answer, show the stiffness equations of those two elements in a manner similar to (9.1).

EXERCISE 9.4 [A/C+N:15+15+10] Suppose that the assembled stiffness equations for a one-dimensional
finite element model before imposing constraints are
    
2 −1 0 u1 1
−1 2 −1 u2 = 0 . (E9.2)
0 −1 2 u3 2

This system is to be solved subject to the multipoint constraint

u1 = u3. (E9.3)

(a) Impose the constraint (E9.3) by the master-slave method taking u 1 as master, and solve the resulting
2 × 2 system of equations by hand.
(b) Impose the constraint (E9.3) by the penalty function method, leaving the weight w as a free parameter.
Solve the equations by hand or CAS (Cramer’s rule is recommended) and verify analytically that as
w → ∞ the solution approaches that found in (a). Tabulate the values of u 1 , u 2 , u 3 for w = 0, 1, 10, 100.
Hint 1: the value of u 2 should not change. Hint 2: the solution for u 1 should be (6w + 5)/(4w + 4).
(c) Impose the constraint (E9.3) by the Lagrange multiplier method. Show the 4 × 4 multiplier-augmented
system of equations analogous to (9.13) and solve it by computer or calculator.

EXERCISE 9.5 [A/C:10+15+10] The left end of the cantilevered beam-column member illustrated in Fig-
ure E9.1 rests on a skew-roller that forms a 45◦ angle with the horizontal axis x. The member is loaded axially
by a force P as shown. The finite element equations upon removing the fixed right end freedoms {u x2 , u y2 , θ2 },
but before imposing the skew-roller MFC, are
    
E A/L 0 0 u x1 P
0 12E I /L 3 6E I /L 2 u y1 = 0 , (E9.4)
0 6E I /L 2 4E I /L θ1 0

where E, A, and I = Izz are given member properties, θ1 is the left end rotation, and L is the member length.7

7 The stiffness equations for a beam column are derived in Part III of this book. For now consider (E9.4) as a recipe.

9–14
 Exercises

To simplify the calculations set P = α E A, and I = β AL 2 , in which α and β are dimensionless parameters,
and express the following solutions in terms of α and β.
(a) Apply the skew-roller constraint by the master-slave

;
method (make u y1 slave) and solve for u x1 and θ1 in terms
of L, α and β. This may be done by hand or a CAS. Partial beam-column E, A, I

;
member
solution: u x1 = αL/(1 + 3β). y

;;;
θ x P 1 2
(b) Apply the skew-roller constraint with the penalty method

;
;
z
by inserting a penalty element at node 1. Follow the rule
of §9.1.4 to construct the 2 × 2 penalty stiffness. Com- L
pute u x1 from the modified equations (Cramer’s rule is 45
o

recommended if solved by hand). Verify that as w → ∞

the answer obtained in (a) is recovered. Partial solution: Figure E9.1. Cantilevered beam-column
u x1 = αL(3E Aβ +wL)/(3E Aβ +wL(1+3β)). Can the on skew-roller for Exercise 9.5.
penalty stiffness be physically interpreted in some way?
(c) Apply the skew roller constraint by Lagrangian multiplier adjunction, and solve the resulting 4×4 system
of equations using a CAS (by hand it will take long). Verify that you get the same solution as in (a).

EXERCISE 9.6 [A:5+5+10+10+5] A cantiveler beam-column is to be joined to a plane stress plate mesh as
depicted in Figure E9.2.8 Both pieces move in the plane {x, y}. Plane stress elements have two degrees of
freedom per node: two translations u x and u y along x and y, respectively, whereas a beam-column element
has three: two translations u x and u y along x and y, and one rotation (positive CCW) θz about z. To connect
the cantilever beam to the mesh, the following “gluing” conditions are applied:
3 6 9
(1) The horizontal (u x ) and vertical (u y ) dis-
placements of the beam at their common plane beam H/2
node (2 of beam, 4 of plate) are the same. 1 4 7 10
2 H
(2) The beam end rotation θ2 and the mean ro- nodes 2 and 4 H/2
tation of the plate edge 3–5 are the same. occupy same position
y 5 8 11
For infinitesimal displacements and rota- θ plane stress mesh
avg
tions the latter is θ35 = (u x5 − u x3 )/H . x
z
Questions:
Figure E9.2. Beam linked to plate in plane stress
(a) Write down the three MFC conditions: two for Exercise 9.6. Beam shown slightly separate from
from (1) and one from (2), and state whether plate for visualization convenience: nodes 2 and 4
they are linear and homogeneous. actually are at the same location.
avg
(b) Where does the above expression of θ35 come from? (Geometric interpretation is OK.) Can it be made
more accurate9 by including u x4 ?
(c) Write down the master-slave transformation matrix if {u x2 , u y2 , θ2 } are picked as slaves. It is sufficient
to write down the transformation for the DOFs of nodes 2, 3, 4, and 5, which gives a T of order 9 × 6,
since the transformations for the other freedoms are trivial.
(d) If the penalty method is used, write down the stiffness equations of the three penalty elements assuming
the same weight w is used. Their stiffness matrices are of order 2 × 2, 2 × 2 and 3 × 3, respectively.
(Do not proceed further)

8 This is extracted from a question previously given in the Aerospace Ph. D. Preliminary Exam. Technically it is not
difficult once the student understand what is being asked. This can take some time, but a HW is more relaxed.
9 To answer the second question, observe that the displacements along 3–4 and 4–5 vary linearly. Thus the angle of rotation
about z is constant for each of them, and (for infinitesimal displacements) may be set equal to the tangent.

9–15
Chapter 9: MULTIFREEDOM CONSTRAINTS II

(e) If Lagrange multiplier adjunction is used, how many Lagrange multipliers will you need to append? (Do
not proceed further).

EXERCISE 9.7 [A:30] Show that the master-slave transformation method u = Tû can be written down as a
special form of the method of Lagrange multipliers. Start from the augmented functional

 M S = 12 uT Ku − uT f + λT (u − Tû) (E9.5)

and write down the stationarity conditions of  M S with respect to u, λ and û in matrix form.

EXERCISE 9.8 [A:35] Check the matrix equations (9.23) through (9.26) quoted for the Augmented La-
grangian method.

EXERCISE 9.9 [A:40] (Advanced, close to a research paper). Show that the master-slave transformation
method u = Tû can be expressed as a limit of the penalty function method as the weights go to infinity. Start
from the augmented functional

 P = 12 uT Ku − uT f + 12 w(u − Tû)T (u − Tû) (E9.6)

Write down the matrix stationarity conditions with respect to to u and û and take the limit w → ∞. Hint:
using Woodbury’s formula (Appendix C, §C.5.2)

(K + wTT ST)−1 = K−1 − K−1 TT (K + w−1 S−1 )−1 T K−1 . (E9.7)

show that
−1
K = TK−1 TT . (E9.8)

9–16
 10
Superelements
and Global-Local
Analysis

10–1
Chapter 10: SUPERELEMENTS AND GLOBAL-LOCAL ANALYSIS

TABLE OF CONTENTS

Page
§10.1 Superelement Concept . . . . . . . . . . . . . . . . . 10–3
§10.1.1 Where Does the Idea Come From? . . . . . . . . . . 10–3
§10.1.2 Subdomains . . . . . . . . . . . . . . . . . . 10–5
§10.1.3 *Mathematical Requirements . . . . . . . . . . . 10–5
§10.2 Static Condensation . . . . . . . . . . . . . . . . . . 10–5
§10.2.1 Condensation by Explicit Matrix Operations . . . . . . 10–5
§10.2.2 Condensation by Symmetric Gauss Elimination . . . . . . 10–6
§10.2.3 Recovery of Internal Freedoms . . . . . . . . . . . 10–8
§10.3 Global-Local Analysis . . . . . . . . . . . . . . . . . . 10–8
§10. Notes and Bibliography . . . . . . . . . . . . . . . . . 10–10
§10. References . . . . . . . . . . . . . . . . . . . . . . 10–10
§10. Exercises . . . . . . . . . . . . . . . . . . . . . . 10–11

10–2
 §10.1 SUPERELEMENT CONCEPT

§10.1. Superelement Concept

A superelement is a grouping of finite elements that, upon assembly, may be regarded as an individual
element for computational purposes. These purposes may be driven by modeling or processing
needs.
A random assortment of elements does not necessarily make up a superelement. To be considered as
such, a grouping must meet certain conditions. Informally we can say that it must form a structural
component on its own. This imposes certain conditions stated mathematically in §10.1.3. Inasmuch
as these conditions involve advanced concepts such as rank sufficiency, which are introduced in
later Chapters, the restrictions are not dwelled upon here.
As noted in Chapter 6, superelements may originate from two overlapping contexts: “bottom up”
or “top down.” In a bottom up context one thinks of superelements as built from simpler elements.
In a top-down context, superelements may be thought as being large pieces of a complete structure.
This dual viewpoint motivates the following classification:
Macroelements. These are superelements assembled with a few primitive elements. Also called
mesh units when they are presented to program users as individual elements.
Substructures. Complex assemblies of elements that result on breaking up a structure into distin-
guishable portions.
When does a substructure becomes a macroelement or vice-versa? There are no precise rules. In
fact the generic term superelement was coined to cover the entire spectrum, ranging from individual
elements to complete structures. This universality is helped by common processing features.
Both macroelements and substructures are treated exactly the same way as regards matrix pro-
cessing. The basic rule is that associated with condensation of internal degrees of freedom. The
technique is illustrated in the following section with a simple example. The reader should note,
however, that condensation applies to any superelement, whether composed of two or a million
elements.1
§10.1.1. Where Does the Idea Come From?
Substructuring was invented by aerospace engineers in the early 1960s2 to carry out a first-level
breakdown of complex systems such as a complete airplane, as depicted in Figure 10.1. The
decomposition may continue hierarchically through additional levels as illustrated in Figure 10.2.
The concept is also natural for space vehicles operating in stages, such as the Apollo short stack
depicted in Figure 10.3.
Three original motivating factors for substructuring can be cited.

1. Facilitate division of labor. Substructures with different functions are done by separate
design groups with specialized knowledge and experience. For instance an aircraft company
may set up a fuselage group, a wing group, a landing-gear group, etc. These groups are thus
protected from “hurry up and wait” constraints. More specifically: a wing design group can

1 Of course the computer implementation becomes totally different as one goes from macroelements to substructures,
because efficient processing for large matrix systems requires exploitation of sparsity.
2 See Notes and Bibliography at the end of this Chapter.

10–3
Chapter 10: SUPERELEMENTS AND GLOBAL-LOCAL ANALYSIS

S4 S6
S2

S5

S1
S3

Figure 10.1. Complete airplane broken down into six level one
substructures identified as S1 through S6 .

level two substructure

level one substructure

individual
element

Figure 10.2. Further breakdown of wing structure. The decomposition

process may continue down to the individual element level.

keep on working on refinements, improvements and experimental model verification as long

as the interface information (the wing-fuselage intersection) stays sensibly unchanged.
2. Take advantage of repetition. Often structures are built COMMAND MODULE

of several identical or nearly identical units. For instance,

the wing substructures S2 and S3 of Figure 10.1 are mirror SERVICE MODULE
images on reflection about the fuselage midplane, and so
are the stabilizers S4 and S5 . Even if the loading is not
ADAPTER
symmetric about that midplane, recognizing repetitions
saves model preparation time. LUNAR MODULE

3. Overcome computer limitations. The computers of the

1960s operated under serious memory limitations. (For
INSTRUMENT UNIT
example, the first supercomputer: the Control Data 6600,
had a total high-speed memory of 131072 60-bit words or THIRD STAGE
1.31 MB; that machine cost $10M in 1966 dollars.) It was SIV-B

difficult to fit a complex structure such as an airplane as

one entity. Substructuring permitted the complete anal-
ysis to be carried out in stages through use of auxiliary Figure 10.3. The Apollo short stack.
storage devices such as tapes or disks.
Of the three motivations, the first two still hold today. The third one has moved to a different plane:

10–4
 §10.2 STATIC CONDENSATION

parallel processing, as noted in §10.1.2 below.

In the late 1960s the idea was picked up and developed extensively by the offshore and shipbuilding
industries, the products of which tend to be modular and repetitive to reduce fabrication costs. As
noted above, repetition favors the use of substructuring techniques.
At the other end of the superelement spectrum, the mesh units herein called macroelements appeared
in the mid 1960s. They were motivated by user convenience. For example, in hand preparation of
models, quadrilateral and bricks involve less human labor than triangles and tetrahedra, respectively.
It was therefore natural to combine the latter to assemble the former. Going a step further one can
assemble components such as “box elements” for applications such as box-girder bridges.
§10.1.2. Subdomains

Applied mathematicians working on solution procedures for parallel computation have developed
the concept of subdomains. These are groupings of finite elements that are entirely motivated by
computational considerations. They are subdivisions of the finite element model done more or less
automatically by a program called domain decomposer.
Although the concepts of substructures and subdomains overlap in many respects, it is better to
keep the two separate. The common underlying theme is divide and conquer but the motivation is
different.
§10.1.3. *Mathematical Requirements

A superelement is said to be rank-sufficient if its only zero-energy modes are rigid-body modes. Equivalently,
the superelement does not possess spurious kinematic mechanisms.
Verification of the rank-sufficient condition guarantees that the static condensation procedure described below
will work properly.

§10.2. Static Condensation

Degrees of freedom of a superelement are classified into two groups:

Internal Freedoms. Those that are not connected to the freedoms of another superelement. Nodes
whose freedoms are internal are called internal nodes.
Boundary Freedoms. These are connected to at least another superelement. They usually reside at
boundary nodes placed on the periphery of the superelement. See Figure 10.4.
The objective is to get rid of all displacement degrees of freedom associated with internal freedoms.
This elimination process is called static condensation, or simply condensation.
Condensation may be presented in terms of explicit matrix operations, as shown in the next sub-
section. A more practical technique based on symmetric Gauss elimination is discussed later.
§10.2.1. Condensation by Explicit Matrix Operations

To carry out the condensation process, the assembled stiffness equations of the superelement are
partitioned as follows:
    
Kbb Kbi ub f
= b . (10.1)
Kib Kii ui fi

10–5
Chapter 10: SUPERELEMENTS AND GLOBAL-LOCAL ANALYSIS

b
b
(a)
(b)
i
i
b b
b
b b b
b b

Figure 10.4. Classification of superelement freedoms into boundary and internal. (a) shows the vertical
stabilizer substructure S6 of Figure 10.2. (The FE mesh is pictured as two-dimensional for illustrative purposes;
for an actual aircraft it will be three dimensional.) Boundary freedoms are those associated to the boundary
nodes labeled b (shown in red), which are connected to the fuselage substructure. (b) shows a quadrilateral
macroelement mesh-unit fabricated with 4 triangles: it has one interior and four boundary nodes.

where subvectors ub and ui collect boundary and interior degrees of freedom, respectively. Take
the second matrix equation:
Kib ub + Kii ui = fi , (10.2)
If Kii is nonsingular we can solve for the interior freedoms:

ui = Kii−1 (fi − Kib ub ), (10.3)

Replacing into the first matrix equation of (10.1) yields the condensed stiffness equations

K̃bb ub = f̃b . (10.4)

In this equation,
K̃bb = Kbb − Kbi Kii−1 Kib , f̃b = fb − Kbi Kii−1 fi , (10.5)
are called the condensed stiffness matrix and force vector, respectively, of the substructure.
From this point onward, the condensed superelement may be viewed, from the standpoint of further
bb
operations, as an individual element whose element stiffness matrix and nodal force vector are K
and f̃b , respectively.
Often each superelement has its own “local” coordinate system. A transformation of (10.5) to an
overall global coordinate system is necessary upon condensation. In the case of multiple levels,
the transformation is done with respect to the next-level superelement coordinate system. This
coordinate transformation procedure automates the processing of repeated portions.
Remark 10.1. The feasibility of the condensation process (10.3)–(10.5) rests on the non-singularity of Kii .
This matrix is nonsingular if the superelement is rank-sufficient in the sense stated in §10.1.3, and if fixing
the boundary freedoms precludes all rigid body motions. If the former condition is verified but not the latter,
the superelement is called floating. Processing floating superelements demands more advanced computational
techniques, among which one may cite the use of projectors and generalized inverses [?].

10–6
 §10.2 STATIC CONDENSATION

CondenseLastFreedom[K_,f_]:=Module[{pivot,c,Kc,fc,
n=Length[K]}, If [n<=1,Return[{K,f}]];
Kc=Table[0,{n-1},{n-1}]; fc=Table[0,{n-1}];
pivot=K[[n,n]]; If [pivot==0, Print["CondenseLastFreedom:",
" Singular Matrix"]; Return[{K,f}]];
For [i=1,i<=n-1,i++, c=K[[i,n]]/pivot;
fc[[i]]=f[[i]]-c*f[[n]];
For [j=1,j<=i,j++,
Kc[[j,i]]=Kc[[i,j]]=K[[i,j]]-c*K[[n,j]]
];
];
Return[Simplify[{Kc,fc}]]
];

K={{6,-2,-1,-3},{ -2,5,-2,-1},{ -1,-2,7,-4},{-3,-1,-4,8}};

f={3,6,4,0};
Print["Before condensation:"," K=",K//MatrixForm," f=",f//MatrixForm];
{K,f}=CondenseLastFreedom[K,f];Print["Upon condensing DOF 4:",
" K=",K//MatrixForm," f=",f//MatrixForm];
{K,f}=CondenseLastFreedom[K,f];Print["Upon condensing DOF 3:",
" K=",K//MatrixForm," f=",f//MatrixForm];

Figure 10.5. Mathematica module to condense the last degree of freedom from a stiffness
matrix and force vector. The test statements carry out the example (10.6)–(10.10).

§10.2.2. Condensation by Symmetric Gauss Elimination

In the computer implementation of the the static condensation process, calculations are not carried
out as outlined above. There are two major differences. The equations of the substructure are not
actually rearranged, and the explicit calculation of the inverse of Kii is avoided. The procedure
may be in fact coded as a variant of symmetric Gauss elimination. To convey the flavor of this
technique, consider the following stiffness equations of a superelement:
    
6 −2 −1 −3 u1 3
 −2 5 −2 −1  2 6
u
   =  . (10.6)
−1 −2 7 −4 u3 4
−3 −1 −4 8 u4 0
Suppose that the last two displacement freedoms: u 3 and u 4 , are classified as interior and are to be
statically condensed out. To eliminate u 4 , perform symmetric Gauss elimination of the fourth row
and column:
    
6 − (−3)×(−3)
8
−2 − (−1)×(−3)
8
−1 − (−4)×(−3)
8 u1 3 − 0×(−3)
8
   
 −2 −
(−3)×(−1)
8
5 − (−1)×(−1) −2 − (−4)×(−1)   u 2  =  6 − 0×(−1)  ,
8 8 8
(10.7)
(−3)×(−4) (−1)×(−4) (−4)×(−4)
−1 − 8
−2 − 8
7− 8
u3 4− 0×(−4)
8
or  39
8
− 19
8
− 52   u 1   3 
 − 19 39
− 5   u2  =  6  . (10.8)
8 8 2
− 52 − 52
5 u3 4
Repeat the process for the third row and column to eliminate u 3 :
 39 (−5/2)×(−5/2)    
− − 19 − (−5/2)×(−5/2) u1 3− 4×(−5/2)
8 5
(−5/2)×(−5/2)
8 5
(−5/2)×(−5/2) = 5
, (10.9)
− −
19
8 5
39
− 8 5
u2 6− 4×(−5/2)
5

10–7
Chapter 10: SUPERELEMENTS AND GLOBAL-LOCAL ANALYSIS

or
    
29
− 29 u1 5
8 8
= . (10.10)
− 29
8
29
8
u2 8

These are the condensed stiffness equations. Figure 10.5 shows a Mathematica program that carries
out the foregoing steps. Module CondenseLastFreedom condenses the last freedom of a stiffness
matrix K and a force vector f. It is invoked as { Kc,fc }=CondenseLastFreedom[K,f]. It returns
the condensed stiffness Kc and force vector fc as new arrays. To do the example (10.6)–(10.10),
the module is called twice, as illustrated in the test statements of Figure 10.5.
Obviously this procedure is much simpler than going through the explicit matrix inverse. Another
important advantage of Gauss elimination is that equation rearrangement is not required even if the
condensed degrees of freedom do not appear sequentially. For example, suppose that the assembled
superelement contains originally eight degrees of freedom and that the freedoms to be condensed
out are numbered 1, 4, 5, 6 and 8. Then Gauss elimination is carried out over those equations only,
and the condensed (3 × 3) stiffness and (3 × 1) force vector extracted from rows and columns 2, 3
and 7. An implementation of this process is considered in Exercise 10.2.

Remark 10.2. The symmetric Gauss elimination procedure, as illustrated in steps (10.6)–(10.10), is primarily
useful for macroelements and mesh units, since the number of stiffness equations for those typically does not
exceed a few hundreds. This permits the use of full matrix storage. For substructures containing thousands
or millions of degrees of freedom — such as in the airplane example — the elimination is carried out using
more sophisticated sparse matrix algorithms; for example that described in [211].

Remark 10.3. The static condensation process is a matrix operation called “partial inversion” or “partial
elimination” that appears in many disciplines. Here is the general form. Suppose the linear system Ax = y,
where A is n × n square and x and y are n-vectors, is partitioned as

    
A11 A12 x1 y1
= . (10.11)
A21 A22 x2 y2

Assuming the appropriate inverses to exist, then the following are easily verified matrix identities:
         
A−1 −A−1
11 A12 y1 x1 A11 − A12 A−1
22 A21 A12 A22
−1
x1 y1
11
= , = . (10.12)
A21 A−111 A 22 − A −1
21 A11 A12 x2 y2 −A−1
22 A21 A−122 y2 x2

We say that x1 has been eliminated or “condensed out” in the left identity and x2 in the right one. In FEM
applications, it is conventional to condense out the bottom vector x2 , so the right identity is relevant. If A is
symmetric, to retain symmetry in (10.12) it is necessary to change the sign of one of the subvectors.

§10.2.3. Recovery of Internal Freedoms

After the boundary freedoms are obtained by the solution process, the interior freedoms can be
recovered directly from (10.3). This process may be carried out either directly through a sequence
of matrix operations, or equation by equation as a backsubstitution process.

10–8
 §10.3 GLOBAL-LOCAL ANALYSIS

coarse mesh finer meshes

Figure 10.6. Left: example panel structure for global-local analysis. Right: a FEM mesh for a one-shot
analysis.

§10.3. Global-Local Analysis

As discussed in the first Chapter, complex engineering systems are often modeled in a multilevel
fashion following the divide and conquer approach. The superelement technique is a practical
realization of that approach.
A related, but not identical, technique is multiscale analysis. The whole system is first analyzed as
a global entity, discarding or passing over details deemed not to affect its overall behavior. Local
details are then analyzed using the results of the global analysis as boundary conditions. The
process can be continued into the analysis of further details of local models. And so on. When this
procedure is restricted to two stages and applied in the context of finite element analysis, it is called
global-local analysis in the FEM literature.
In the global stage the behavior of the entire structure is simulated with a finite element model that
necessarily ignores details such as cutouts or joints. These details do not affect the overall behavior
of the structure, but may have a bearing on safety. Such details are a posteriori incorporated in a
series of local analyses.
The gist of the global-local approach is explained in the example illustrated in Figures 10.6 and 10.7.
Although the structure is admittedly too simple to merit the application of global-local analysis, it
serves to illustrate the basic ideas. Suppose one is faced with the analysis of the rectangular panel
shown on the top of Figure 10.6, which contains three small holes. The bottom of that figure shows
a standard (one-stage) FEM treatment using a largely regular mesh that is refined near the holes.
Connecting the coarse and fine meshes usually involves using multifreedom constraints because
the nodes at mesh boundaries do not match, as depicted in that figure.
Figure 10.7 illustrates the global-local analysis procedure. The global analysis is done with a coarse
but regular FEM mesh that ignores the effect of the holes. This is followed by local analysis of
the region near the holes using refined finite element meshes. The key ingredient for the local
analyses is the application of boundary conditions (BCs) on the finer mesh boundaries. These BCs
may be of displacement (essential) or of force (natural) type. If the former, the applied boundary
displacements are interpolated from the global mesh solution. If the latter, the internal forces
or stresses obtained from the global calculation are converted to nodal forces on the fine meshes

10–9
Chapter 10: SUPERELEMENTS AND GLOBAL-LOCAL ANALYSIS

Global analysis with a coarse mesh, ignoring holes, followed

bylocal analysis of the vicinity of the holes with finer meshes:

Figure 10.7. Global-local analysis of problem of Figure 10.6.

through a lumping process.

The BC choice noted above gives rise to two basic variations of the global-local approach. Expe-
rience accumulated over several decades3 has shown that the stress-BC approach generally gives
more reliable answers.
The global-local technique can be extended to more than two levels, in which case it receives the more
encompassing name multiscale analysis. Although this generalization is still largely in the realm
of research, it is receiving increasing attention from various science and engineering communities
for complex products such as the thermomechanical analysis of microelectronic components.
Notes and Bibliography
Substructuring was invented by aerospace engineers in the early 1960s. Przemieniecki’s book [603] contains
a fairly complete bibliography of early work. Most of this was in the form of largely inaccessible internal
company or lab reports and so the actual history is difficult to trace. Macroelements appeared simultaneously
in many of the early FEM codes. Quadrilateral macroelements fabricated with triangles are described in [205].
For a survey of uses of the static condensation algorithm in FEM, see [809].
The generic term superelement was coined in the late 1960s by the SESAM group at DNV Veritas [196].
The matrix form of static condensation for a complete structure is presented in [23], as a scheme to eliminate
unloaded DOF in the displacement method. It is unclear when this idea was first applied to substructures or
macroelements. The first application for reduced dynamical models is by Guyan [326].
The application of domain decomposition to parallel FEM solvers has produced an enormous and highly
specialized literature. Procedures for handling floating superelements using generalized inverse methods are
discussed in [234,242,571].
The global-local analysis procedure described here is also primarily used in industry and as such it is rarely
mentioned in academic textbooks.
References
Referenced items have been moved to Appendix R.

3 Particularly in the aerospace industry, in which the global-local technique has been used since the early 1960s.

10–10
 Exercises

Homework Exercises for Chapter 10

Superelements and Global-Local Analysis

EXERCISE 10.1 [N:15] The free-free stiffness equations of a superelement are

    
88 −44 −44 0 u1 5
 −44 132 −44 −44   u 2   10 
 −44 −44 = . (E10.1)
176 −44   u 3   15 
0 −44 −44 220 u4 20

Eliminate u 2 and u 3 from (E10.1) by static condensation, and show (but do not solve) the condensed equation
system. Use either the explicit matrix inverse formulas 10.5 or the symmetric Gauss elimination process
explained in §10.2.2. Hint: regardless of method the result should be
    
52 −36 u1 15
= . (E10.2)
−36 184 u4 30

EXERCISE 10.2 [C+N:20] Generalize the program of Figure 10.5 to a module CondenseFreedom[K,f,k]
that is able to condense the kth degree of freedom from K and f, which is not necessarily the last one. That
is, k may range from 1 to n, where n is the number of freedoms in Ku = f. Apply that program to solve the
previous Exercise.
Hint: here is a possible way of organizing the inner loop:
ii=0; For [i=1,i<=n,i++, If [i==k, Continue[]]; ii++;
c=K[[i,k]]/pivot; fc[[ii]]=f[[i]]-c*f[[k]]; jj=0;
For [j=1,j<=n,j++, If [j==k,Continue[]]; jj++;
Kc[[ii,jj]]=K[[i,j]]-c*K[[k,j]]
];
];
Return[{Kc,fc}]

EXERCISE 10.3 [D:15] Explain the similarities and differences between superelement analysis and global-
local FEM analysis.

EXERCISE 10.4 [A:20] If the superelement stiffness K is symmetric, the static condensation process can be
viewed as a special case of the master-slave transformation method discussed in Chapter 8. To prove this, take
exterior freedoms ub as masters and interior freedoms ui as slaves. Assume the master-slave transformation
relation
     
ub I 0 def
u= = [ u ] − = Tub − g. (E10.3)
ui −Kii−1 Kib b
−Kii−1 fi

Work out K̂ = TT KT and f̂ = TT (f − Kg), and show that K̂û = f̂ coalesces with the condensed stiffness
equations (10.4)–(10.5) if ub ≡ û. (Take advantage of the symmetry properties Kbi
T
= Kib , (Kii−1 )T = Kii−1 .)

EXERCISE 10.5 [D:30] (Requires thinking) Explain the conceptual and operational differences between
one-stage FEM analysis and global-local analysis of a problem such as that illustrated in Figures 10.6 and
10.7. Are the answers the same? What is gained, if any, by the global-local approach over the one-stage FEM
analysis?

10–11
Chapter 10: SUPERELEMENTS AND GLOBAL-LOCAL ANALYSIS

EXERCISE 10.6 [N:20] The widely used Guyan’s scheme [326] for dynamic model reduction applies the
static-condensation relation (E10.3) as master-slave transformation to both the stiffness and the mass matrix
of the superelement. Use this procedure to eliminate the second and third DOF of the mass-stiffness system:
   
2 1 0 0 1 −1 0 0
1 4 1 0  −1 2 −1 0
M= , K= . (E10.4)
0 1 4 1 0 −1 2 −1 
0 0 1 2 0 0 −1 1

Compute and compare the vibration frequencies of the eigensystems (M − ωi2 K)vi = 0 and (M̂ − ω̂i2 K̂)v̂i = 0
before and after reduction.
Hints. The four original squared frequencies are the eigenvalues of M−1 K, which may be obtained, for
example, with the Matlab eig function. The largest is 2, lowest 0. To perform the Guyan reduction it is
convenient to reorder M and K so that rows and columns 2 and 3 become 3 and 4, respectively:
     
2 0 1 0 v1 1 0 −1 0 v1
 0 2 0 1   v4   0 1 0 −1   v4 
ω2  = , (E10.5)
1 0 4 1   v2   −1 0 2 −1   v2 
0 1 1 4 v3 0 −1 −1 2 v3

which in partitioned matrix form is

     
Mbb Mbi vb Kbb Kbi vb
ω2 = . (E10.6)
Mib Mii vi Kib Kii vi
Next show that static condensation of K is equivalent to a master-slave transformation
     
1 0 v1 1 0
   
I  0 1   v4   0 1  v1
T= =  v  =  2/3 . (E10.7)
1/3  
relating
−Kii−1 Kib 2/3 2 1/3 v4
1/3 2/3 v3 1/3 2/3

The rest is easy. Form M̂ = TT MT and K̂ = TT KT, find the two squared frequencies of the condensed
eigensystem and verify that they approximate the two lowest original squared frequencies.

EXERCISE 10.7 [A:20] Two beam elements: 1–2 and 2–3, each of length L and rigidity E I are connected at
node 2. The macroelement has 6 degrees of freedom: {v1 , θ1 , v2 , θ2 , v3 , θ3 }. Eliminate the two DOF of node 2
by condensation. Is the condensed stiffness the same as that of a beam element of length 2L and rigidity E I ?
(For expressions of the beam stiffness matrices, see Chapter 12.)

10–12
 11
Variational
Formulation of
Bar Element

11–1
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

TABLE OF CONTENTS

Page
§11.1. A New Beginning 11–3
§11.2. Definition of Bar Member 11–3
§11.3. Variational Formulation 11–4
§11.3.1. The Total Potential Energy Functional . . . . . . . . . 11–4
§11.3.2. Admissible Variations . . . . . . . . . . . . . . 11–6
§11.3.3. The Minimum Total Potential Energy Principle . . . . . . 11–6
§11.3.4. TPE Discretization . . . . . . . . . . . . . . . 11–7
§11.3.5. Bar Element Discretization . . . . . . . . . . . . . 11–8
§11.3.6. Interpolation by Shape Functions . . . . . . . . . . 11–9
§11.3.7. The Strain-Displacement Matrix . . . . . . . . . . . 11–9
§11.3.8. *Trial Basis Functions . . . . . . . . . . . . . . 11–9
§11.4. The Finite Element Equations 11–10
§11.4.1. The Stiffness Matrix . . . . . . . . . . . . . . . 11–10
§11.4.2. The Consistent Node Force Vector . . . . . . . . . . 11–11
§11.5. Weak Forms 11–13
§11.5.1. From Strong to Weak . . . . . . . . . . . . . . . 11–13
§11.5.2. Weak Form Based Approximation Example . . . . . . . 11–14
§11.5.3. Balanced Weak Forms . . . . . . . . . . . . . . 11–15
§11.5.4. Principle of Virtual Work as Balanced Weak Form . . . . . 11–15
§11.5.5. *Weighted Residual Methods . . . . . . . . . . . . 11–16
§11.6. *Accuracy Analysis 11–16
§11.6.1. *Nodal Exactness and Superconvergence . . . . . . . . 11–16
§11.6.2. *Fourier Patch Analysis . . . . . . . . . . . . . . 11–17
§11.6.3. *Robin Boundary Conditions . . . . . . . . . . . 11–18
§11. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 11–19
§11. References . . . . . . . . . . . . . . . . . . . . . . 11–20
§11. Exercises . . . . . . . . . . . . . . . . . . . . . . 11–21

11–2
 §11.2 DEFINITION OF BAR MEMBER

§11.1. A New Beginning

This Chapter begins Part II of the course. This Part focuses on the construction of structural and
continuum finite elements using a variational formulation based on the Total Potential Energy. Why
only elements? Because the other synthesis steps of the DSM: globalization, merge, BC application
and solution, remain the same as in Part I. Those operations are not element dependent.
Individual elements are constructed in this Part beginning with the simplest ones and progressing
to more complicated ones. The formulation of 2D finite elements from a variational standpoint is
discussed in Chapters 14 and following. Although the scope of that formulation is broad, exceeding
structural mechanics, it is better understood by going through specific elements first.
From a geometrical standpoint the simplest finite elements are one-dimensional or line elements.
This means that the intrinsic dimensionality is one, although these elements may be used in one,
two or three space dimensions upon transformation to global coordinates as appropriate. The
simplest one-dimensional structural element is the two-node bar element, which we have already
encountered in Chapters 2, 3 and 5 as the truss member.
In this Chapter the bar stiffness equations are rederived using the variational formulation. For
uniform properties the resulting equations are the same as those found previously using the physical
or Mechanics of Materials approach. The variational method has the advantage of being readily
extendible to more complicated situations, such as variable cross section or more than two nodes.

(a) (b)
y

;;
x
cross section axial rigidity EA

;;
u(x)
;;
;;

z q(x) P
P
x cross
section
Longitudinal axis L

Figure 11.1. A fixed-free bar member: (a) 3D view showing reference frame; (b) 2D view on {x, y} plane
highlighting some quantities that are important in bar analysis.

§11.2. Definition of Bar Member

In structural mechanics a bar is a structural component characterized by two properties:

(1) One preferred dimension: the longitudinal dimension or axial dimension is much larger that the
other two dimensions, which are collectively known as transverse dimensions. The intersection
of a plane normal to the longitudinal dimension and the bar defines the cross sections. The
longitudinal dimension defines the longitudinal axis. See Figure 11.1(a).
(2) The bar resists an internal axial force along its longitudinal dimension.

11–3
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

Table 11.1 Nomenclature for Mathematical Model of Axially Loaded Bar

Quantity Meaning
x Longitudinal bar axis∗
(.) d(.)/d x
u(x) Axial displacement
q(x) Distributed axial force, given per unit of bar length
L Total bar length
E Elastic modulus
A Cross section area; may vary with x
EA Axial rigidity
e = du/d x = u  Infinitesimal axial strain
σ = Ee = Eu  Axial stress
F = Aσ = E A e = E Au  Internal axial force
P Prescribed end load
∗
x is used in this Chapter instead of x̄ (as in Chapters 2–3) to simplify the notation.

In addition to trusses, bar elements are used to model cables, chains and ropes. They are also used
as fictitious elements in penalty function methods, as discussed in Chapter 9.
We will consider here only straight bars, although their cross section may vary. Our one-dimensional
mathematical model assumes that the bar material is linearly elastic obeying Hooke’s law, and that
displacements and strains are infinitesimal. Figure 11.1(b) pictures some relevant quantities for a
fixed-free bar. Table 11.1 collects the necessary terminology for the governing equations.
Figure 11.2 displays the governing equations of the bar in a graphic format called a Tonti diagram.
The formal similarity with the diagrams used in Chapter 5 to explain MoM elements should be
noted, although the diagram of Figure 11.2 pertains to the continuum bar model rather than to the
discrete one. (The qualifier “strong form” is explained in the next Chapter.)

§11.3. Variational Formulation

To illustrate the variational formulation, the finite element equations of the bar will be derived from
the Minimum Potential Energy principle.

§11.3.1. The Total Potential Energy Functional

In Mechanics of Materials it is shown that the internal energy density at a point of a linear-elastic
material subjected to a one-dimensional state of stress σ and strain e is U = 12 σ (x)e(x), where
σ is to be regarded as linked to the displacement u through Hooke’s law σ = Ee and the strain-
displacement relation e = u  = du/d x. This U is also called the strain energy density. Integration
over the volume of the bar gives the total internal energy
  L  L  L
 
U= 1
2
σ e dV = 1
2
Fe d x = 1
2
(E Au )u d x = 1
2
u  E A u  d x. (11.1)
V 0 0 0

11–4
 §11.3 VARIATIONAL FORMULATION

Displacement
Prescribed Axial Distributed
end BCs displacement axial load
displacement u(x) q(x)

Kinematic e=u' F'+q=0 Equilibrium

Axial F = EA e Axial Force BCs Prescribed

strain force end load
e(x) Constitutive F(x)

Figure 11.2. Strong-form Tonti diagram for the continuum model of a bar member. Field equations and BCs
are represented as lines connecting the boxes. Yellow (brown) boxes contain unknown (given) quantities.

All integrand quantities in (11.1) may depend on x.

The external work potential is the work performed by applied mechanical loads working on the
bar displacements. This potential is denoted by W . (The external energy V is the negative of the
work potential: V = −W . In the ensuing derivations W will be used instead of V .) It collects
contributions from two sources:
1. The distributed load q(x). This contributes a cross-section density of q(x)u(x) because q is
assumed to be already integrated over the section.
2. Any specified axial point load(s). For the fixed-free example of Figure 11.1 the end load P
would contribute P u(L).
The second source may be folded into the first by conventionally writing any point load P acting at
a cross section x = a as a contribution P δ(a) to q(x), in which δ(a) denotes the one-dimensional
Dirac delta function at x = a. If this is done the external energy can be concisely expressed as
 L
W = q u d x. (11.2)
0

The total potential energy of the bar is given by

=U −W (11.3)

Mathematically is a functional, called the Total Potential Energy functional or TPE. It depends
only on the axial displacement u(x). In Variational Calculus u(x) is called the primary variable
of the functional. When the dependence of on u needs to be emphasized we shall write [u] =
U [u]−W [u], with brackets enclosing the primary variable. To display both primary and independent
variables we write, for example, [u(x)] = U [u(x)] − W [u(x)].

Remark 11.1. According to the rules of Variational Calculus, the Euler-Lagrange equation for is

∂ d ∂
E= − = −q − (E A u  ) (11.4)
∂u d x ∂u 

11–5
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

δu(x)
u(x)+δu(x) δu(L)
u
u(L)
u(x)
x
u(0) = 0
x
Figure 11.3. Concept of admissible variation of the axial displacement function u(x). For convenience u(x) is
plotted normal to the longitudinal axis. Both u(x) and u(x) + δu(x) shown above are kinematically admissible,
and so is the variation δu(x). Note that the variation δu(L) is not zero because the BC at x = L is natural.

The stationary condition for is E = 0, or

(E A u  ) + q = 0 (11.5)
This is the strong (pointwise) equation of equilibrium in terms of the axial displacement, which reduces to
E A u  + q = 0 if E A is constant. This equation is not explicitly used in the FEM development. It is instead
replaced by δ = 0, with the variation restricted over the class of finite element interpolation functions.

§11.3.2. Admissible Variations

The concept of admissible variation is fundamental in both variational calculus and the variationally
formulated FEM. Only the primary variable(s) of a functional may be varied. For the TPE functional
(11.3) this is the axial displacement u(x). Suppose that u(x) is changed to u(x) + δ u(x).1 This is
illustrated in Figure 11.3, where for convenience u(x) is plotted normal to x. The TPE functional
changes accordingly as
= [u] ⇒ + δ = [u + δu]. (11.6)
The function δ u(x) and the scalar δ are called the variations of u(x) and , respectively. The
variation δ u(x) should not be confused with the ordinary differential du(x) = u  (x) d x since on
taking the variation the independent variable x is frozen; that is, δx = 0.
A displacement variation δu(x) is said to be admissible when both u(x) and u(x) + δ u(x) are
kinematically admissible in the sense of the Principle of Virtual Work (PVW). This agrees with the
conditions of classical variational calculus, and are restated next.
A kinematically admissible axial displacement u(x) obeys two conditions:
(i) It is continuous over the bar length, that is, u(x) ∈ C 0 in x ∈ [0, L].
(ii) It satisfies exactly any displacement boundary condition, such as the fixed-end specification
u(0) = 0 of Figure 11.1. See of Figure 11.3.
The variation δ u(x) pictured in Figure 11.3 is kinematically admissible because both u(x) and
u(x) + δ u(x) satisfy the foregoing conditions. Note that the variation δu(L) at the free end x = L
is not necessarily zero because that boundary condition is natural; that is, not specified directly in
terms of the displacement u(L). On the other hand, δ(0) = 0.
The physical meaning of conditions (i)–(ii) is the subject of Exercise 11.1.
1 The symbol δ not immediately followed by a parenthesis is not a delta function but instead denotes variation with respect
to the variable that follows.

11–6
 §11.3 VARIATIONAL FORMULATION

u 1, f1 u 2, f2 u 3, f3 u 4, f4 u 5, f5

(1) (2) (3) (4)

1 2 3 4 5
u u(x)
u5
u2 u3 u4
x
u1 = 0
Figure 11.4. FEM discretization of bar member. A piecewise- linear admissible displacement trial
function u(x) is drawn underneath the mesh. It is assumed that the left end is fixed; thus u 1 = 0.

§11.3.3. The Minimum Total Potential Energy Principle

The Minimum Total Potential Energy (MTPE) principle states that the actual displacement solution
u ∗ (x) that satisfies the governing equations is that which renders stationary:2

δ = δU − δW = 0 iff u = u∗ (11.7)

with respect to admissible variations u = u ∗ + δu of the exact displacement field u ∗ (x).

Remark 11.2. Using standard techniques of variational calculus3 it can be shown that if E A > 0 and kinematic
boundary conditions weed out any rigid motions, the solution u ∗ (x) of (11.7) exists, is unique, and renders
[u] a minimum over the class of kinematically admissible displacements. The last attribute explains the
“minimum” in the name of the principle.

§11.3.4. TPE Discretization

To apply the TPE functional (11.3) to the derivation of FEM equations we replace the contin-
uum mathematical model by a discrete one consisting of a union of bar elements. For example,
Figure 11.4 illustrates the subdivision of a fixed-free bar member into four two-node elements.
Functionals are scalars. Therefore, for a discretization such as that shown in Figure 11.4, the TPE
functional (11.3) may be decomposed into a sum of contributions of individual elements:

= (1) + (2) + . . . + (Ne ) (11.8)

in which Ne denotes the number of elements. The same decomposition applies to both its internal
energy and external work potential components:

δU = δU (1) + . . . + δU (Ne ) = 0, δW = δW (1) + . . . + δW (Ne ) = 0, (11.9)

as well as to the stationarity condition (11.7):

δ = δ (1) + δ (2) + . . . + δ (Ne ) = 0. (11.10)

2 The symbol “iff” in (11.7) is an abbreviation for “if and only if”.
3 See references in Notes and Bibliography at the end of Chapter.

11–7
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

_
(a) u e (x) (b1) (b2)
u2 u2 u2

u1 u1 u1
= +
_
1 (e) 2 x
_
x = xγ − x1 (b3)
= Le
= u1 1 + 1 u2
(c) _
1−x/
1 N1e
0
_
x/
N2e 1
0

Figure 11.5. A two-node, TPE-based bar element: (a) element configuration and axial displacement
variation (plotted normal to element axis for better visibility); (b1,b2,b3) displacement interpolation
expressed in terms of linear shape functions; (c) element shape functions.

Using the fundamental lemma of variational calculus,4 it can be shown that (11.10) implies that for
a generic element e we may write

δ e = δU e − δW e = 0. (11.11)
This variational equation is the basis for the derivation of element stiffness equations once the
displacement field has been discretized over the element, as described next.
Remark 11.3. In mathematics (11.11) is called a first variation form. It is a special case of a more general
expression called a weak form, which is covered in more detail later. In mechanics it states the Principle of
Virtual Work or PVW for each element: δU e = δW e , which says that the virtual work of internal and external
forces on admissible displacement variations is equal if the element is in equilibrium [588].

§11.3.5. Bar Element Discretization

Figure 11.5(a) depicts a generic bar element e. It has two nodes, which are labeled 1 and 2. These
are called the local node numbers.5 The element is referred to its local axis x̄ = x − x1 , which
measures the distance from its left end. The two degrees of freedom are u e1 and u e2 . (Bars are not
necessary since the directions of x̄ and x are the same.) The element length is = L e .
The mathematical concept of bar finite elements is based on approximating axial displacement u(x)
over the element. The exact displacement u ∗ is replaced by an approximate displacement
u ∗ (x) ≈ u e (x) (11.12)
4 See, e.g., Chapter II of Gelfand and Fomin [297].
5 Note the notational change from the labels i and j of Part I. This will facilitate transition to multidimensional elements
in Chapters 14ff.

11–8
 §11.3 VARIATIONAL FORMULATION

over the finite element mesh. This approximate displacement, u e (x), taken over all elements e =
1, 2, . . . N e , is called the finite element trial expansion or simply trial expansion. See Figure 11.4.
This FE trial expansion must belong to the class of kinematically admissible displacements defined
in ?. Consequently, it must be C 0 continuous over and between elements. The most common
choices fpr u e are polynomials in x, as in the development that follows.

§11.3.6. Interpolation by Shape Functions

In a two-node bar element the only possible polynomial choice of the displacement u e that satisfies
the interelement continuity requirement is linear. It can be expressed by the following interpolation
formula, which is graphically developed in Figure 11.5(b1,b2,b3):
 
ue
u (x) =
e
N1e u e1 + N2e u e2 =[ N1e N2e ] 1e = Ne ue . (11.13)
u2

The functions N1e and N2e that multiply the node displacements u 1 and u 2 are called shape functions,
while N is called the shape function matrix. In this case Ne reduces to a row vector.
The shape functions interpolate the internal displacement u e directly from the node values. They
are pictured in Figure 11.5(c). For this element, with x̄ = x − x1 measuring the axial distance from
the left node i, the shape functions are

x̄ x̄
N1e = 1 − = 1 − ζ, N2e = = ζ. (11.14)

Here
x − x1 x̄
ζ = = , (11.15)

is a dimensionless coordinate, also known as a natural coordinate, that varies from 0 through 1
over the element. Note that d x = dζ and dζ = d x/ . The shape function N1e has the value 1 at
node 1 and 0 at node 2. Conversely, shape function N2e has the value 0 at node 1 and 1 at node 2.
This is a general property of shape functions. It follows from the fact that element displacement
interpolations such as (11.13) are based on physical node values.

§11.3.7. The Strain-Displacement Matrix

The axial strain associated with the trial function u e is

du e  dNe dNe   e  1  
e  u1 ue
e= = (u ) = 1 2
e = [ −1 1 ] 1e = B ue , (11.16)
dx dx dx u 2 u2

in which
1
B= [ −1 1], (11.17)

is called the strain-displacement matrix. Note that B is constant over the element.

11–9
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

(a) (b)
u N3
(thick line)
(1) (2) (3) (4) N3(2) N3(3) x
1 2 3 4 5
Figure 11.6. Trial basis function (a.k.a. hat function) for node 3 of a four-element bar discretization.

§11.3.8. *Trial Basis Functions

Shape functions are associated with elements. A trial basis function, or simply basis function, is associated
with a node. Suppose node i of a bar discretization connects elements (e1) and (e2). The trial basis function
Ni is defined as  (e1)
Ni if x ∈ element (e1)
Ni (x) = N (e2) if x ∈ element (e2) (11.18)
i
0 otherwise
For a piecewise linear discretization, such as that used in the two-node bar, this function has the shape of a
hat. Thus it is also called a hat function or chapeau function. See Figure 11.6, in which i = 3, e1 = 2, and
e2 = 3. The concept is important in the variational interpretation of FEM as a Rayleigh-Ritz method.

Remark 11.4. In addition to continuity, shape and trial functions must satisfy a completeness requirement
with respect to the governing variational principle. This condition is stated and discussed in later Chapters.
Suffices for now to say that the shape functions (11.14), as well as the associated trial functions, do satisfy this
requirement.

§11.4. The Finite Element Equations

In linear FEM the discretization process based on the TPE functional leads to the following algebraic
form in the node displacements
def 1 e T def
e = U e − W e , in which U e = 2
(u ) Ke ue and W e = (ue )T fe . (11.19)
Here Ke and fe are called the element stiffness matrix and the element consistent nodal force vector,
respectively. The three scalars e , U e and W e are only function of the node displacements ue . (This
is a consequence of displacements being the only primary variable of the TPE functional.) Note
that U e and W e depend quadratically and linearly, respectively, on those displacements. Taking
the variation of e with respect to the node displacements gives6
∂ e   e T
e e

e T

δ = δu e
= δu K u − f e
= 0. (11.20)
∂ue
Because the variations δue can be arbitrary, the bracketed expression must vanish, which yields

Ke ue = fe . (11.21)
These are the familiar element stiffness equations. Hence the foregoing names given to Ke and fe
are justified a posteriori.
6 The 12 factor disappears on taking the variation because U e is quadratic in the node displacements. For a review on the
calculus of discrete quadratic forms, see Appendix D.

11–10
 §11.4 THE FINITE ELEMENT EQUATIONS

§11.4.1. The Stiffness Matrix

We now apply the foregoing expressions to the two-node bar element. Its internal energy U e is
 x2  1
U =2
e 1
e E A e dx = 2
1
e E A e dζ. (11.22)
x1 0

Note that the integration variable x has been changed to the natural coordinate ζ defined in (11.15)
that varies from 0 through 1, whence d x = dζ . This form is symmetrically expanded using the
strain-displacement matrix relation (11.16), by inserting e = e T = (ue )T BT and e = B ue into the
first and second e of (11.22), respectively, to get
 1  1    e
1 −1 1 u
U =2
e 1
(u ) B E A B u dζ = 2
e T T e 1
[ u e1 u e2 ] E A [ −1 1 ] 1e dζ.
0 0 1 u 2
(11.23)
The nodal displacements do not depend on position and can be moved out of the integral. Also
BT E A B = E A BT B since E A is a scalar:
 1  1    e
EA 1 −1 u1
U = 2 (u )
e 1 e T
E A B B dζ u = 2 [ u 1 u 2 ]
T e 1 e e
dζ e . (11.24)
0 0 2 −1 1 u 2
 T
By (11.19) this is expressible as 12 ue Ke ue . Since ue is arbitrary, Ke is extracted as
 1  1     1
EA 1 −1 1 1 −1
K =
e
E A B B dζ =
T
dζ = E A dζ.
0 0 2 −1 1 −1 1 0
(11.25)
This is the bar element stiffness matrix. For a homogeneous and prismatic bar of constant rigidity,
1
E A can be moved outside the integral, 0 dζ = 1 and (11.25) collapses to
 
EA 1 −1
K = e
. (11.26)
−1 1

This is the same element stiffness matrix of the prismatic truss member derived in Chapters 2 and
5 by a Mechanics of Materials approach, but now obtained through a variational argument.

§11.4.2. The Consistent Node Force Vector

The consistent node force vector fe defined in (11.19) comes from the element contribution to the
external work potential W :
 x2  1   
 e T 1 1−ζ def  T
W =
e
q u dx = q N u dζ = u
T e
q dζ = ue fe , (11.27)
x1 0 0 ζ

Since ue is arbitrary,
 x2    1  
1−ζ 1−ζ
f =e
q dx = q dζ. (11.28)
x1 ζ 0 ζ

11–11
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

;;;;;;;
(b) Load case I: point load P at x=L
2 P
1 1
(a)
;;; (c) Load case II: q(x) varies linearly
constant EA
from q1 at x=0 through q2 at x=L
x 1 1 2
L
(d) Load case III: q(x)=q0 (constant)
from x=0 through x=L/2, else 0
1 1 2

Figure 11.7. Fixed-free, prismatic bar example: (a) configuration; (b,c,d) FEM discretization and load cases.

in which ζ is defined by (11.15). If q is constant over the element, it may be taken out of the
integral:
 1 
1−ζ
f =q
e
dζ. (11.29)
0 ζ

This gives the same results as with the EbE lumping method of Chapter 7. See Exercise 11.3.

Example 11.1. The two-node bar element is tested on the benchmark problem defined in Figure 11.7. A fixed-
free, homogeneous, prismatic bar of length L, elastic modulus E and cross section area A has the configuration
illustrated in Figure 11.7(a). It is discretized with a single element as shown in Figure 11.7(b,c,d), and subjected
to the three load cases pictured there. Case I involves a point load P at the free end, which may be formally
represented as
q I (x) = P δ(L) (11.30)
where δ() denotes the delta function with argument x.
Case II involves a distributed axial load that varies linearly from q1 = q(0) at the fixed end through q2 = q(L)
at the free end:
q I I (x) = q1 (1 − ζ ) + q2 ζ, (11.31)
in which ζ = 1 − x/L. Case III involves a “box” distributed load q(x) that is constant and equal to q0 from
the fixed end x = 0 through midspan x = L/2, and zero otherwise:

q I I I (x) = q0 H (x) − H (x − 12 L) , (11.32)

in which H () denotes the Heaviside unit step function with argument x. The master stiffness equations
constructed using the prismatic stiffness matrix (11.26) with = L and x̄ → x are

EA  1 −1

u m1
 
f 1m

= = fm . (11.33)
L −1 1 u m2 f 2m

Here supercript m identifies the load case. The consistent node forces computed from (11.28) with = L and
x̄ → x are  
0 L  2q1 + q2  q0 L  3 
fI = , fI I = , fI I I = . (11.34)
P 6 q1 + 2q2 8 1

11–12
 §11.5 WEAK FORMS

On applying the fixed end support condition u m1 = 0 and solving for u m2 , the free end deflections are

PL (q1 + 2q2 )L 2 q0 L 2
u 2I = , u 2I I = , u 2I I I = . (11.35)
EA 6E A 8E A
The analytical solutions for u(x), obtained on integrating the ODE E Au  + q = 0 with boundary conditions
u(0) = 0, F(L) = E A u  (L) = P for case I and F(L) = E A u  (L) = 0 for cases II and III, are
Px x [3(q1 +q2 )L 2 −3q1 L x+(q1 −q2 )x 2 ] q0 
u I (x) = , u I I (x) = , u I I I (x) = L x−x 2 +x− 12 L2 .
EA 6E A 2E A
(11.36)
In the expression of u I I I (x), x − 12 L2 means (x − 12 L)2 if x ≥ 12 L, else zero (Macauley’s bracket notation
for discontinuity functions). Evaluating (11.36) at x = L and comparing to (11.35), one finds that the three
computed end deflections are exact.
For case I this agreement is no surprise: the exact u I (x) is linear in x, which is contained in the span of
the linear shape functions. But for II and III this is far from obvious since the exact solutions are cubic and
piecewise quadratic, respectively, in x. The fact that the exact solution is verified at the free end node is an
instance of the nodal exactness property discussed in §11.6.1.
Note that in cases II and III the FEM displacement solutions inside the element, which vary linearly, will
not agree pointwise with the exact solutions, which do not. For example the exact midspan displacement is
u I I I ( 12 L) = q0 L 2 /(8E A) = u I I I (L), whereas the FEM interpolation would give q0 L 2 /(16E A) there, in error
by 100%. To reduce such internal discrepancies the member may be divided into more elements.

§11.5. Weak Forms

Weak forms are expressions notoriously difficult to explain to newcomers. They occupy an inter-
mediate position between differential equations and functionals. There are so many variants and
procedural kinks, however, that their position in the mathematical food chain is fuzzy. Confusion
is compounded by the use of disparate terminology, some generic, some application oriented. To
shed some sunlight into this murky swamp, we go through a specific example: the bar member.

§11.5.1. From Strong to Weak

The
 governing
 differential equation for a bar member in terms of axial displacements is
E A u (x) + q(x) = 0, or E A u  (x) + q(x) = 0 if the rigidity E A is constant. Replace the


zero by r (x), which stands for residual, and move it to the left-hand side:
 
r (x) = E A u  (x) + q(x), or if E A is constant: r (x) = E A u  (x) + q. (11.37)

The governing ODE may be compactly stated as r (x) = 0. This must hold at each point over the
member span, say x ∈ [0, L]. Hence the term strong form (SF) used for this kind of mathematical
model. No ambiguity so far. But suppose that insisting on r (x) = 0 everywhere is too demanding.
We would like to relax that condition so it is satisfied only in an average sense. To accomplish that,
multiply the residual by a function v(x), integrate over the problem domain, and set the result to
zero:  L
J= r (x) v(x) d x = 0. (11.38)
0
Here v(x) is supposed to be sufficiently well behaved for the integral to exist. Ignoring boundary
conditions for now, (11.38) is called a weak form, which is often abbreviated to WF in the sequel.

11–13
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

Function v(x) receives two names in the literature: test function in a general mathematical context,
and weight function (also weighting function) in the context of approximation methods based on
weak forms. In what follows both terms will be used within the appropriate context.
§11.5.2. Weak Form Based Approximation Example
To show how weak forms can be used to generate approximate solutions, consider again a fixed-free,
prismatic, homogeneous bar member (that is, E A is constant), under uniform load q(x) = q0 along
its length and zero load at the free end. The WF (11.38) becomes
 L

J= E A u  (x) + q0 v(x) d x = 0. (11.39)
0

subject to the end conditions

u(0) = 0, F(L) = E A u  (L) = 0. (11.40)
We will restrict both u(x) and v(x) to be quadratic polynomials:
u(x) = a0 + a1 x + a2 x 2 , v(x) = b0 + b1 x + b2 x 2 . (11.41)
in which ai and bi are numerical coefficients, real in this case. Once assumptions such as those
in (11.41) are made, more terminology kicks in. The assumed u(x) is now called a trial function,
which is spanned by the linear-space basis {1, x, x 2 } of dimension 3. The assumed v(x) is called
a weight function, which is spanned by exactly the same basis. There is a special name for the
scenario when the trial and weight function bases coalesce: the Galerkin method.7 . We will call
the end result a Galerkin solution. Replacing (11.41) into (11.39) we get
L
J= (6b0 + 3b1 L + 2b2 L 2 ) (2E A a2 + q0 ). (11.42)
6
Now J must vanish for any arbitrary value of {b0 , b1 , b2 }. On extracting the expressions that multiply
those coefficients we obtain the same equation thrice: 2E A a2 + q0 = 0. Thus a2 = −q0 /(2E A),
whereas a0 and a1 remain arbitrary. Consequently the Galerkin solution before BC is
q0 2
u(x) = a0 + a1 x − x . (11.43)
2E A
ODE aficionados would recognize this as the general solution of E Au  + q0 = 0 so Uncle Boris
has done the job. Applying the end conditions (11.40) gives a0 = 0 and a1 = q0 /(E A) whence the
final solution is q0
u(x) = x(2L − x). (11.44)
2E A
Replacing into (11.37) and (11.40) it may be verified that this is the exact analytical solution.
Instead of applying the end conditions a posteriori we may try to incorporate them a priori into
the trial function assumption. On enforcing (11.40) into the assumed u(x) of (11.41) we find that
a0 = 0 and a1 = −2a2 L. The trial function becomes
u(x) = a2 x (x − 2L), (11.45)
7 Introduced by Boris Galerkin in 1912. For a brief account of the general methodology, see Notes and Bibliography

11–14
 §11.5 WEAK FORMS

and only one free coefficient remains. Accordingly only one weight basis function is needed: either
1, x or x 2 does the job, and the exact solution (11.44) is obtained again.8
What happens if the load q(x) varies, say, linearly and the same quadratic polynomial assumptions
(11.41) are used? Then Galerkin goes gaga. See Exercise 11.8.
Even for this trivial example, several procedural choices are apparent. If we allow the trial and
weight function spaces to differ, volatility zooms up. Furthermore, we can apply transformations
to the residual integral as done in the next subsection. Compared to the well ordered world of
variational-based FEM, confusion reigns supreme.
§11.5.3. Balanced Weak Forms
Some method in the madness can be injected by balancing. A look at (11.39) reveals an unpleasant
asymmetry. Second derivatives of u(x) appear, but none of v(x). This places unequal restrictions
on smoothness of the trial and test function spaces. Integration by parts restores derivative order
L L L
balance. Replacing 0 E A u  v d x = − 0 E A u  v  d x + (E Au  )v 0 and rearranging terms yields
 L  L
 L
J=  
E A u (x) v (x) d x − q(x) v(x) d x − E A u  (x) v(x) 0 . (11.46)
0 0

This will be called a balanced-derivative weak form, or simply a balanced weak form (BWF). It
displays obvious advantages: (i) same smoothness requirements for assumed u and v, and (ii) end
BC appear explicitly in the non-integral term, neatly factored into essential and natural. A minor
flaw is that the original residual is no longer clearly visible.
For a bar with variable axial rigidity replace E A u  by (E A u  ) in the first integrand.
On repeating the Galerkin procedure of the previous subsection with the assumptions (11.41) one
finds an identical J , as may be expected, and the same final solution. Again one has the choice of
pre- or post-imposing the end conditions (11.40). Generally the latter choice is far more convenient
in a computer implementation.
§11.5.4. Principle of Virtual Work as Balanced Weak Form
There is a close relationship between the BWF (11.46) and one of the fundamental tools of Analytical
Mechanics: the Principle of Virtual Work (PVW). To exhibit it, set the test function to be an
admissible variation of u(x): v(x) = δ u(x), in which δu(x) strongly satisfies all essential BC.
Then assume that J is the first variation of a functional :
 L  L
 L def
J=  
E A u (x) δu (x) d x − q(x) δu(x) d x − E A u  (x) δu(x) 0 = δ . (11.47)
0 0

Indeed this is the first variation of the TPE functional:

 L  L
 
=U −W = 2 1
u (x) E A u (x) d x − q(x) u(x) d x (11.48)
0 0

8 Some early works covering weighted residual methods, for example Crandall [159], proclaim that the trial function must
satisfy all BC ab initio. Later ones, e.g., [260,261], relax that rule to BC of essential type (in Galerkin methods, this rule
applies to both trial and test functions since the spaces coalesce). In practice this rule can be often relaxed further, as in
the example of §11.5.2, applying essential BCs at the last moment.

11–15
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

Hence J = 0 is the same as δ = 0 or δU = δW , which is the PVW for an elastic bar member.
This relationship can be used to prove an important property: Galerkin method is equivalent to a
variational formulation if the residual is the Euler-Lagrange equation of a functional.
 L
Remark 11.5. Where does the boundary term E A u  (x) δu(x) in (11.47) go? Actually, into δW . This
0
immersion is a bit tricky, and depends on redefining q(x) to include prescribed end point forces such as
N (L) = E A u  (L) = P through delta functions. This is the subject of Exercise 11.9.

§11.5.5. *Weighted Residual Methods

Galerkin method is widely used in computational mechanics, but does not exhaust all possibilities of using a
weak form as source for obtaining numerical solutions. The main generalization consist of allowing trial and
test (weight) functions to be different. This leads to a rich class of approximation methods unified under the
name Method of Weighted Residuals or MWR.
The key idea is as follows. Both u(x) and v(x) are restricted to belong to linear function spaces of finite
dimension Nu and Nv . These are the trial function space and the test function space, respectively. which are
spanned by basis functions φi (x) and ψ(x), respectively:
   
u(x) = span φi (x), 1 ≤ i ≤ Nu , v(x) = span ψi (x), 1 ≤ i ≤ Nv (11.49)

in which usually Nu = Nv . Since the spaces are linear, any u(x) and v(x) can be represented as linear
combination of the basis functions:

Nu 
Nv
u(x) = ai φi (x), v(x) = bi ψi (x). (11.50)
i=1 i=1

Here ai and bi are scalar coefficients, which may be real or complex depending on the nature of the problem.
Insert these into the weak form, perform the necessary integrations, and extract the Nv expressions that are
coefficients of the bi . Solve these equations for the coefficients ai , and replace in the first of (11.50) to get the
approximate solution u(x).
The MWR methodology is of course not restricted to one space dimension. It also extends to time-dependent
problems. It can be merged smoothly with the FEM concept of piecewise approximation using shape functions.
Some references are provided under Notes and Bibliography.

§11.6. *Accuracy Analysis

Low order 1D elements may give surprisingly high accuracy. In particular the lowly two-node bar element
can display infinite accuracy under some conditions. This phenomenon is studied in this advanced section as
it provides an introduction to modified equation methods and Fourier analysis along the way.

§11.6.1. *Nodal Exactness and Superconvergence

Suppose that the following two conditions are satisfied:
1. The bar properties are constant along the length (prismatic member).
2. The distributed load q(x) is zero between nodes. The only applied loads are point forces at the nodes.
If so, a linear axial displacement u(x) as defined by (11.13) and (11.14) is the exact solution over each element
since constant strain and stress satisfy, element by element, all of the governing equations listed in Figure 11.2.9

9 The internal equilibrium equation p  + q = E A u  + q = 0 is trivially verified because q = 0 from the second
assumption, and u  = 0 because of shape function linearity.

11–16
 §11.6 *ACCURACY ANALYSIS

(a)
(b)
q(x)
Trial basis function Nj
for node j
xi = xj −
i j k 1 − (xj −x)/ = 1−ψ 1 + (x j−x)/ = 1+ψ
EA = const 1
xj Two-element
xk = xj + patch ijk
L i j k

Figure 11.8. Superconvergence patch analysis: (a) lattice of bar elements; (b) two element patch.

It follows that if the foregoing conditions are verified the FEM solution is exact; that is, it agrees with the
analytical solution of the mathematical model.10 Adding extra elements and nodes would not change the
solution. That is the reason behind the truss discretizations used in Chapters 2–3: one element per member is
enough if they are prismatic and loads are applied to joints. Such models are called nodally exact.
What happens if the foregoing assumptions are not met? Exactness is then generally lost, and several elements
per member may be beneficial if spurious mechanisms are avoided.11 For a 1D lattice of equal-length, prismatic
two-node bar elements, an interesting and more difficult result is: the solution is nodally exact for any loading
if consistent node forces are used. This is proven in the subsection below. This result underlies the importance
of computing node forces correctly.
If conditions such as equal-length are relaxed, the solution is no longer nodally exact but convergence at the
nodes is extremely rapid (faster than could be expected by standard error analysis) as long as consistent node
forces are used. This phenomenon is called superconvergence in the FEM literature.

§11.6.2. *Fourier Patch Analysis

The following analysis is based on the modified differential equation (MoDE) method of Warming and Hyett
[?] combined with the Fourier patch analysis approach of Park and Flaggs [553,554]. Consider a lattice of
two-node prismatic bar elements of constant rigidity E A and equal length , as illustrated in Figure 11.8. The
total length of the lattice is L. The system is subject to an arbitrary axial load q(x). The only requirement on
q(x) is that it has a convergent Fourier series in the space direction.
From the lattice extract a patch12 of two elements connecting nodes xi , x j and xk as shown in Figure 11.8. The
FEM patch equations at node j are
 
EA ui
[ −1 2 −1 ] uj = fj, (11.51)
uk
in which the node force f j is obtained by consistent lumping:
 xk  0  1
fj = q(x)N j (x) d x = q(x j + ψ )(1 + ψ) dψ + q(x j + ψ )(1 − ψ) dψ. (11.52)
xi −1 0

Here N j (x) is the “hat” trial basis function for node j, depicted in Figure 11.8, and ψ = (x − x j )/ is a
dimensionless coordinate that takes the values −1, 0 and 1 at nodes i, j and k, respectively. If q(x) is expanded

10 In variational language: the Green function of the u  = 0 problem is included in the FEM trial space.
11 These can happen when transforming such elements for 2D and 3D trusses. See Exercise E11.7.
12 A patch is the set of all elements connected to a node; in this case j.

11–17
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

in Fourier series

M
q(x) = qm eiβm x , βm = mπ/L , (11.53)
m=1

(the term m = 0 requires special handling) the exact solution of the continuum equation E A u  + q = 0 is


M
qm eiβm x
u ∗ (x) = u ∗m eiβm x , u ∗m = . (11.54)
m=1
E Aβm2

Evaluation of the consistent force using (11.52) gives


M
sin2 ( 12 βm )
fj = f jm , f jm = qm eiβm x2 . (11.55)
m=1
β
1 2 2
4 m

To construct a modified differential equation (MoDE), expand the displacement by Taylor series centered at
node j. Evaluate at i and k: u i = u j − u j + 2 u j /2! − 3 u j /3! + 4 u ivj /4! + . . . and u k = u j + u j +
2 u j /2 + 3 u j /3! + 4 u ivj /4! + . . .. Replace these series into (11.51) to get
 
1  2 iv 4 vi
−2E A u + u + uj + ... = fj. (11.56)
2! j 4! j 6!

This is an ODE of infinite order. It can be reduced to an algebraic equation by assuming that the response of
(11.56) to qm eiβm x is harmonic: u jm eiβm x . If so u jm = −βm2 u jm , u ivjm = βm4 u jm , etc, and the MoDE becomes
 
1 β 2 2 β 4 4 sin2 ( 1 βm )
2E A βm2 − m + m − . . . u jm = 4E A sin2 ( 12 βm ) u jm = f jm = qm 1 22 2 eiβm x j .
2! 4! 6! β
4 m
(11.57)
Solving gives u jm = qm eiβm x j /(E Aβm2 ), which compared with (11.54) shows that u jm = u ∗m for any m > 0.
Consequently u j = u ∗j . In other words, the MoDE (11.56) and the original ODE: E Au  + q = 0 have the
same value at x = x j for any load q(x) developable as (11.53). This proves nodal exactness. In between
nodes the two solutions will not agree.13
The case m = 0 has to be treated separately since the foregoing expressions become 0/0. The response to a
uniform q = q0 is a quadratic in x, and it is not difficult to prove nodal exactness.

§11.6.3. *Robin Boundary Conditions

Suppose that for a bar of length L one has the following end conditions: u  (0) = au(0) + b at x = 0 and
u  (L) = au(L) + b at x = L, in which a and b are given coefficients. Those are called Robin BCs in the
literature. Adjoining them as Courant penalty terms gives the functional
 L
F(u) = [ 12 E A (u  )2 − q u] d x + 12 [u  (0) − au(0) − b]2 + 12 [u  (L) + au(L) + b]2 . (11.58)
0

Divide [0,L] into Ne elements and N = Ne + 1 nodes. Do C 0 linear interpolation over each element, insert
into F(u) to get Fd (u) = 12 uT Ku − fT v, in which u is the vector of node values, K the master stiffness matrix
and f the master force vector. Coefficients a and b will affect both K and f.

13 The FEM solution varies linearly between nodes whereas the exact one is generally trigonometric.

11–18
 §11. Notes and Bibliography

Vanishing of the first variation: δ Fd = 0 yields the FEM equations Ku = f to be solved for u. The Robin BCs
at x = 0 and x = L will affect the stiffness and force contributions of the first and last elements, but not those
of interior elements.
This kind of boundary value problem (i.e., with Robin BCs) is common in heat conduction and heat transfer
with convection given over cooling surfaces. In that case the heat flux is proportional to the difference of the
(unknown) surface temperature and that of the cooling fluid. Elements that ”touch” the convention boundary
are affected.
Notes and Bibliography
The foregoing account pertains to the simplest structural finite element: the two-node bar element. For bar
members these developments may be generalized in several directions, three of which are mentioned next.
Refined bar models. Adding internal nodes we can pass from linear to quadratic and cubic shape functions.
These elements are rarely useful on their own right, but as accessories to 2D and 3D high order continuum
elements (for example, to model edge reinforcements.) For that reason they are not considered here. The
3-node bar element is developed in exercises assigned in Chapter 16.
Use in 2D and 3D truss structures. The only additional ingredients are the local-to-global transformations
discussed in Chapters 3 and 6.
Curved bar elements. These can be derived using isoparametric mapping, a device introduced later.
Matrices for straight bar elements are available in any finite element book; for example Przemieniecki [596].
Tonti diagrams were introduced in the 1970s in papers now difficult to access, for example [749]. Scanned
images are available, howewer, from http://www.dic.units.it/perspage/discretephysics
The fundamentals of Variational Calculus may be studied in the excellent textbook [297], which is now
available in an inexpensive Dover edition. The proof of the MPE principle can be found in texts on variational
methods in mechanics. For example: Langhaar [435], which is the most readable “old fashioned” treatment
of the energy principles of structural mechanics, with a clear treatment of virtual work. (Out of print but
used copies may be found via the web engines cited in §1.5.2.) The elegant treatment by Lanczos [434] is
recommended as reading material although it is more oriented to physics than structural mechanics.
It was noted that weak forms occupy an intermediate position between two older classical areas: differential
equations (introduced in the XVII Century by the Calculus founders) and variational forms (introduced by
Euler in the XVIII Century). Some weak forms in disguise are also ancient; e.g., the PVW was placed on
firm mathematical grounds by Lagrange in the late XVIII Century [430]. But their rapid development as tools
for producing approximate solutions of ODEs and PDEs took place in the early XIX Century. Five important
variants are: Galerkin (1915), subdomain (1923), least squares (1928), moments (1932), and collocation
(1937). These, as well as a few others of less importance, were unified in 1956 under the label Method of
Weighted Residuals or MWR, by Crandall [159]. Other attempts at unification during this period may be found
in [19,147]. The use of MWR methods, especially Galerkin’s, as enabling devices to generate finite element
equations developed rapidly following the pioneer paper [820]. The chief motivation was to accommodate
application problems where a classical variational formulation does not exist, or is inconvenient to use.
The first accuracy study of FEM discretizations using modified equation methods is by Waltz et. al. [780];
however their procedures were faulty, which led to incorrect conclusions. The first correct derivation of
modified equations appeared in [783]. The topic has recently attracted interest from applied mathematicians
because modified equations provide a systematic tool for backward error analysis of differential equations: the
discrete solution is the exact solution of the modified problem. This is particularly important for the study of
long term behavior of discrete dynamical systems, whether deterministic or chaotic. Recommended references
along these lines are [318,327,709].
Nodal exactness of bar models for point node loads is a particular case of a theorem by Tong [746]. For
arbitrary loads it was proven by Park and Flaggs [553,554], who followed a variant of the scheme of §11.6.2.

11–19
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

A different technique is used in Exercise 11.10. The budding concept of superconvergence, which emerged
in the late 1960s, is outlined in the book of Strang and Fix [698]. There is a monograph [781] devoted to the
subject; it covers only Poisson problems but provides a comprehensive reference list until 1995.
References
Referenced items moved to Appendix R.

11–20
 Exercises

Homework Exercises for Chapter 11

Variational Formulation of Bar Element

EXERCISE 11.1 [D:10] Explain the kinematic admissibility requirements stated in ? in terms of physics,
namely ruling out the possibility of gaps or interpenetration as the bar material deforms.

EXERCISE 11.2 [A/C:15] Using (11.25), derive the stiffness matrix for a tapered bar element in which the
cross section area varies linearly along the element length:

A = Ai (1 − ζ ) + A j ζ, (E11.1)

where Ai and A j are the areas at the end nodes, and ζ = x e / is the dimensionless coordinate defined in
§11.3.6. Show that this yields the same answer as that of a stiffness of a constant-area bar with cross section
1
2
(Ai + A j ). Note: the following Mathematica script may be used to solve this exercise:14

ClearAll[Le,x,Em,A,Ai,Aj];
Be={{-1,1}}/Le; ζ =x/Le; A=Ai*(1-ζ )+Aj*ζ ;
Ke=Integrate[Em*A*Transpose[Be].Be,{x,0,Le}];
Ke=Simplify[Ke];
Print["Ke for varying cross section bar: ",Ke//MatrixForm];

In this and following scripts Le stands for .

EXERCISE 11.3 [A:10] Find the consistent load vector fe for a bar of constant area A subject to a uniform
axial force q = ρg A per unit length along the element. Show that this vector is the same as that obtained with
the element-by-element (EbE) “lumping” method of §8.4, which simply assigns half of the total load: 12 ρg A ,
1
to each node. Hint: use (11.29) and 0
ζ dζ = /.

EXERCISE 11.4 [A/C:15] Repeat the previous calculation for the tapered bar element subject to a force
q = ρg A per unit length, in which A varies according to (E11.1) whereas ρ and g are constant. Check that if
Ai = A j one recovers f i = f j = 12 ρg A . Note: the following Mathematica script may be used to solve this
exercise:15

ClearAll[q,A,Ai,Aj,ρ,g,Le,x];
ζ =x/Le; Ne={{1-ζ ,ζ }}; A=Ai*(1-ζ )+Aj*ζ ; q=ρ*g*A;
fe=Integrate[q*Ne,{x,0,Le}];
fe=Simplify[fe];
Print["fe for uniform load q: ",fe//MatrixForm];
ClearAll[A];
Print["fe check: ",Simplify[fe/.{Ai->A,Aj->A}]//MatrixForm];

EXERCISE 11.5 [A/C:20] A tapered bar element of length , end areas Ai and A j with A interpolated as
per (E11.1), and constant density ρ, rotates on a plane at uniform angular velocity ω (rad/sec) about node i.
Taking axis x along the rotating bar with origin at node i, the centrifugal axial force is q(x) = ρ Aω2 x along
the length, in which x ≡ x e . Find the consistent node forces as functions of ρ, Ai , A j , ω and , and specialize
the result to the prismatic bar A = Ai = A j . Partial result check: f j = 13 ρω2 A 2 for A = Ai = A j .

14 The ClearAll[...] at the start of the script is recommended programming practice to initialize variables and avoid
“cell crosstalk.” In a Module this is done by listing the local variables after the Module keyword.
15 The ClearAll[A] before the last statement is essential; else A would retain the previous assignation.

11–21
Chapter 11: VARIATIONAL FORMULATION OF BAR ELEMENT

EXERCISE 11.6 [A:15] (Requires knowledge of Dirac’s delta function properties.) Find the consistent load
vector fe if the bar is subjected to a concentrated axial force Q at a distance x = a from its left end. Use
(11.28), with q(x) = Q δ(a), in which δ(a) is the one-dimensional Dirac’s delta function at x = a. Note: the
following script does it by Mathematica, but it is overkill:

ClearAll[Le,q,Q,a,x];
ζ =x/Le; Ne={{1-ζ ,ζ }}; q=Q*DiracDelta[x-a];
fe=Simplify[ Integrate[q*Ne,{x,-Infinity,Infinity}] ];
Print["fe for point load Q at x=a: ",fe//MatrixForm];

EXERCISE 11.7 [C+D:20] In a learned paper, Dr. I. M. Clueless proposes “improving” the result for the
example truss by putting three extra nodes, 4, 5 and 6, at the midpoint of members 1–2, 2–3 and 1–3, respectively.
His “reasoning” is that more is better. Try Dr. C.’s suggestion using the Mathematica implementation of Chapter
4 and verify that the solution “blows up” because the modified master stiffness is singular. Explain physically
what happens.

EXERCISE 11.8 [C+D:15] This exercise illustrates “Galerkin surprises.” Take up again the example of
§11.5.2, but suppose now that the axial load varies linearly, as in (11.31). The trial and weight function
assumptions are the quadratic polynomials (11.41). Show that the integral (11.39) is given by
  
12 J/L = b0 24E A a2 + 6(q1 +q2 ) + b1 12E A a2 + 2(q1 +2q2 ) + b2 8E A a2 + (q1 +3q2 ) , (E11.2)

and that the resulting 3 equations for a2 are inconsistent unless q1 = q2 . Only one weight function gives
the correct solution at x = L; which one? Note that the Galerkin method is generally viewed as the “most
reliable” member of the MWR tribe. But unforeseen surprises have a silver lining: more papers can be written
to explain them. Here is a partial fix: make the test function satisfy the essential BC a priori.

EXERCISE 11.9 [A:20]. Prove that (11.47) is the first variation of (11.48), thus linking the PVW with the
TPE functional. See Remark 11.5 for a hint on how to treat the boundary term in (11.47).

EXERCISE 11.10 [A:35, close to research paper level]. Prove nodal exactness of the two-node bar element
for arbitrary but Taylor expandable loading without using the Fourier series approach. Hints: expand q(x) =
q(x j ) + ( ψ)q  (x j ) + ( ψ)2 q  (x j )/2! + . . ., where ψ = x − x j is the distance to node j, compute the
consistent force f j (x) from (11.52), and differentiate the MoDE (11.56) repeatedly in x while truncating all
derivatives to a maximum order n ≥ 2. Show that the original ODE: E Au  + q = 0, emerges as an identity
regardless of how many derivatives are kept.

11–22
 12
Variational
Formulation of
Plane Beam
Element

12–1
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

TABLE OF CONTENTS

Page
§12.1 Introduction . . . . . . . . . . . . . . . . . . . . . 12–3
§12.2 What is a Beam? . . . . . . . . . . . . . . . . . . . 12–3
§12.2.1 Terminology . . . . . . . . . . . . . . . . . 12–3
§12.2.2 Mathematical Models . . . . . . . . . . . . . . 12–3
§12.2.3 Assumptions of Classical Beam Theory . . . . . . . . . 12–4
§12.3 The Bernoulli-Euler Beam Theory . . . . . . . . . . . . . 12–4
§12.3.1 Element Coordinate Systems . . . . . . . . . . . . 12–4
§12.3.2 Kinematics . . . . . . . . . . . . . . . . . 12–5
§12.3.3 Loading . . . . . . . . . . . . . . . . . . . 12–5
§12.3.4 Support Conditions . . . . . . . . . . . . . . . 12–5
§12.3.5 Strains, Stresses and Bending Moments . . . . . . . . . 12–5
§12.4 Total Potential Energy Functional . . . . . . . . . . . . . 12–6
§12.5 Beam Finite Elements . . . . . . . . . . . . . . . . . . 12–7
§12.5.1 Finite Element Trial Functions . . . . . . . . . . . 12–8
§12.5.2 Shape Functions . . . . . . . . . . . . . . . . 12–8
§12.6 The Finite Element Equations . . . . . . . . . . . . . . 12–9
§12.6.1 The Stiffness Matrix of a Prismatic Beam . . . . . . . . 12–10
§12.6.2 Consistent Nodal Force Vector for Uniform Load . . . . . 12–11
§12. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 12–15
§12. References . . . . . . . . . . . . . . . . . . . . . . 12–15
§12. Exercises . . . . . . . . . . . . . . . . . . . . . . 12–16

12–2
 §12.2 WHAT IS A BEAM?

§12.1. Introduction

The previous Chapter introduced the TPE-based variational formulation of finite elements, which
was illustrated for the bar element. This Chapter applies that technique to a more complicated
one-dimensional element: the plane beam described by engineering beam theory.
Mathematically, the main difference of beams with respect to bars is the increased order of conti-
nuity required for the assumed transverse-displacement functions to be admissible. Not only must
these functions be continuous but they must possess continuous x first derivatives. To meet this
requirement both deflections and slopes are matched at nodal points. Slopes may be viewed as
rotational degrees of freedom in the small-displacement assumptions used here.

§12.2. What is a Beam?

Beams are the most common type of structural component, particularly in Civil and Mechanical
Engineering. A beam is a bar-like structural member whose primary function is to support transverse
loading and carry it to the supports. See Figure 12.1.
By “bar-like” it is meant that one of the dimen-
sions is considerably larger than the other two.
This dimension is called the longitudinal dimen-
sion or beam axis. The intersection of planes nor-
mal to the longitudinal dimension with the beam
Figure 12.1. A beam is a structural member
member are called cross sections. A longitudinal
designed to resist transverse loads.
plane is one that passes through the beam axis.

A beam resists transverse loads mainly through bending action, Bending produces compressive
longitudinal stresses in one side of the beam and tensile stresses in the other.
The two regions are separated by a neutral Neutral surface Compressive stress
surface of zero stress. The combination of
tensile and compressive stresses produces
an internal bending moment. This moment
is the primary mechanism that transports
Tensile stress
loads to the supports. The mechanism is
illustrated in Figure 12.2. Figure 12.2. Beam transverse loads are
primarily resisted by bending action.
§12.2.1. Terminology

A general beam is a bar-like member designed to resist a combination of loading actions such as
biaxial bending, transverse shears, axial stretching or compression, and possibly torsion. If the
internal axial force is compressive, the beam has also to be designed to resist buckling. If the
beam is subject primarily to bending and axial forces, it is called a beam-column. If it is subjected
primarily to bending forces, it is called simply a beam. A beam is straight if its longitudinal axis
is straight. It is prismatic if its cross section is constant.
A spatial beam supports transverse loads that can act on arbitrary directions along the cross section.
A plane beam resists primarily transverse loading on a preferred longitudinal plane. This Chapter
considers only plane beams.

12–3
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

§12.2.2. Mathematical Models

One-dimensional mathematical models of structural beams are constructed on the basis of beam
theories. Because beams are actually three-dimensional bodies, all models necessarily involve some
form of approximation to the underlying physics. The simplest and best known models for straight,
prismatic beams are based on the Bernoulli-Euler beam theory (also called classical beam theory
and engineering beam theory), and the Timoshenko beam theory. The Bernoulli-Euler theory is that
taught in introductory Mechanics of Materials courses, and is the one emphasized in this Chapter.
The Timoshenko beam model is presented in Chapter 13, which collects advanced material.
Both models can be used to formulate beam finite elements. The Bernoulli-Euler beam theory leads
to the so-called Hermitian beam elements.1 These are also known as C 1 elements for the reason
explained in §12.5.1. This model neglects the effect of transverse shear deformations on the internal
energy. Elements based on Timoshenko beam theory, also known as C 0 elements, incorporate a
first order correction for transverse shear effects. This model assumes additional importance in
dynamics and vibration.
§12.2.3. Assumptions of Classical Beam Theory

The Bernoulli-Euler or classical beam theory for plane beams rests on the following assumptions:
1. Planar symmetry. The longitudinal axis is straight and the cross section of the beam has a
longitudinal plane of symmetry. The resultant of the transverse loads acting on each section
lies on that plane. The support conditions are also symmetric about this plane.
2. Cross section variation. The cross section is either constant or varies smoothly.
3. Normality. Plane sections originally normal to the longitudinal axis of the beam remain plane
and normal to the deformed longitudinal axis upon bending.
4. Strain energy. The internal strain energy of the member accounts only for bending moment
deformations. All other contributions, notably transverse shear and axial force, are ignored.
5. Linearization. Transverse deflections, rotations and deformations are considered so small that
the assumptions of infinitesimal deformations apply.
6. Material model. The material is assumed to be elastic and isotropic. Heterogeneous beams
fabricated with several isotropic materials, such as reinforced concrete, are not excluded.

§12.3. The Bernoulli-Euler Beam Theory

§12.3.1. Element Coordinate Systems

Under transverse loading one of the top surfaces shortens while the other elongates; see Figure 12.2.
Therefore a neutral surface that undergoes no axial strain exists between the top and the bottom.
The intersection of this surface with each cross section defines the neutral axis of that cross section.2

1 The qualifier “Hermitian” relates to the use of a transverse-displacement interpolation formula studied by the French
mathematician Hermite. The term has nothing to do with the mathematical model used.
2 If the beam is homogenous, the neutral axis passes through the centroid of the cross section. If the beam is fabricated
of different materials — for example, a reinforced concrete beam — the neutral axes passes through the centroid of an
“equivalent” cross section. This topic is covered in Mechanics of Materials textbooks; for example Popov [595].

12–4
 §12.3 THE BERNOULLI-EULER BEAM THEORY

y, v q(x)
y,v
x, u z Neutral axis
_
zc Beam
cross Centroid
Neutral surface L section Symmetry plane

Figure 12.3. Terminology and choice of axes for Bernoulli-Euler model of plane beam.

The Cartesian axes for plane beam analysis are chosen as shown in Figure 12.3. Axis x lies along
the longitudinal beam axis, at neutral axis height. Axis y lies in the symmetry plane and points
upwards. Axis z is directed along the neutral axis, forming a RHS system with x and y. The origin
is placed at the leftmost section. The total length (or span) of the beam member is called L.
§12.3.2. Kinematics
The motion under loading of a plane beam member in the x, y plane is described by the two
dimensional displacement field  
u(x, y)
, (12.1)
v(x, y)
where u and v are the axial and transverse displacement components, respectively, of an arbitrary
beam material point. The motion in the z direction, which is primarity due to Poisson’s ratio
effects, is of no interest. The normality assumption of the Bernoulli-Euler model can be represented
mathematically as
∂v(x)
u(x, y) = −y = −yv  = −yθ, v(x, y) = v(x). (12.2)
∂x
Note that the slope v  = ∂v/∂ x = dv/d x of the deflection curve has been identified with the
rotation symbol θ . This is permissible because θ represents to first order, according to the kinematic
assumptions of this model, the rotation of a cross section about z positive CCW.
§12.3.3. Loading
The transverse force per unit length that acts on the beam in the +y direction is denoted by q(x), as
illustrated in Figure 12.3. Concentrated loads and moments acting on isolated beam sections can
be represented by the delta function and its derivative. For example, if a transverse point load F
acts at x = a, it contributes Fδ(a) to q(x). If the concentrated moment C acts at x = b, positive
CCW, it contributes Cδ  (b) to q(x), where δ  denotes a doublet acting at x = b.
§12.3.4. Support Conditions
Support conditions for beams exhibit far more variety than for bar members. Two canonical cases
are often encountered in engineering practice: simple support and cantilever support. These are
illustrated in Figures 12.4 and 12.5, respectively. Beams often appear as components of skeletal
structures called frameworks, in which case the support conditions are of more complex type.

12–5
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

;;;;;;;
;;;;;;;
;;;;;;;
;;;;;;;
Figure 12.4. A simply supported beam has end
supports that preclude transverse displacements but
permit end rotations.
Figure 12.5. A cantilever beam is clamped at one end
and free at the other. Airplane wings and stabilizers are
examples of this configuration.

§12.3.5. Strains, Stresses and Bending Moments

The Bernoulli-Euler or classical model assumes that the internal energy of beam member is entirely
due to bending strains and stresses. Bending produces axial stresses σx x , which will be abbreviated
to σ , and axial strains ex x , which will be abbreviated to e. The strains can be linked to the
displacements by differentiating the axial displacement u(x) of (12.2):

∂u ∂ 2v d 2v
e= = −y 2 = −y 2 = −yv  = −yκ. (12.3)
∂x ∂x dx

Here κ denotes the deformed beam axis curvature, which to first order is κ ≈ d 2 v/d x 2 = v  . The
bending stress σ = σx x is linked to e through the one-dimensional Hooke’s law

d 2v
σ = Ee = −E y = −E yκ, (12.4)
dx2
where E is the longitudinal elastic modulus. The most important stress resultant in classical beam
theory is the bending moment M, which is defined as the cross section integral
 
d 2v
M= −yσ d A = E 2 y 2 d A = E I κ. (12.5)
A dx A

 y
Here I ≡ Izz denotes the moment of inertia A y 2 d A of
the cross section with respect to the z (neutral) axis. The M
bending moment M is considered positive if it compresses
the upper portion: y > 0, of the beam cross section, as illus- z
x

trated in Figure 12.6. This convention explains the negative

sign of y in the integral (12.5). The product E I is called the V
bending rigidity of the beam with respect to flexure about Figure 12.6. Positive sign
the z axis. convention for M and V .

The governing equations of the Bernoulli-Euler beam model are summarized in the Tonti diagram
of Figure 12.7.

12–6
 §12.5 BEAM FINITE ELEMENTS

Prescribed Displacement Distributed

BCs Transverse
end displacements transverse load
displacements v(x) q(x)

Kinematic κ = v'' M''=q Equilibrium

Curvature M = EI κ Bending
moment
Force BCs Prescribed
end loads
κ(x) Constitutive M(x)

Figure 12.7. The Tonti diagram for the governing equations of the Bernoulli-Euler beam model.

§12.4. Total Potential Energy Functional

The total potential energy of the beam is

=U −W (12.6)

where as usual U and W denote the internal and external energies, respectively. As previously
explained, in the Bernoulli-Euler model U includes only the bending energy:
    
L L L  2 L
U= 1
2
σe dV = 1
2
Mκ d x = 1
2
E Iκ dx =
2 1
2
E I v  d x = 1
2
v  E I v  d x.
V 0 0 0 0
(12.7)
The external work W accounts for the applied transverse force:
 L
W = qv d x. (12.8)
0

The three functionals , U and W must be regarded as depending on the transverse displacement
v(x). When this dependence needs to be emphasized we write [v], U [v] and W [v].
Note that [v] includes up to second derivatives in v, because v  = κ appears in U . This number
is called the variational index. Variational calculus tells us that since the index is 2, admissible
displacements v(x) must be continuous, have continuous first derivatives (slopes or rotations), and
satisfy the displacement BCs exactly. This continuity requirement can be succintly stated by saying
that admissible displacements must be C 1 continuous. This condition guides the construction of
beam finite elements described below.
Remark 12.1. If there is an applied distributed moment m(x) per unit of beam length, the external energy
L
(12.8) must be augmented with a 0 m(x)θ(x) d x term. This is further elaborated in Exercises 12.4 and 12.5.
Such kind of distributed loading is uncommon in practice although in framework analysis occasionally the
need arises for treating a concentrated moment between nodes.

12–7
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

§12.5. Beam Finite Elements

Beam finite elements are obtained by subdivid-

ing beam members longitudinally. The simplest
Bernoulli-Euler plane beam element has two end P'(x+u,y+v) θ2
nodes: 1 and 2, and four degrees of freedom (DOF). y, v
These are collected in the node displacement vector
v2
θ1 v1 E, I
ue = [ v1 θ1 v2 θ2 ]T . (12.9) x, u
1 2
The element is shown in Figure 12.8, which pictures x
the undeformed and deformed configurations.  P(x,y)

Figure 12.8. The two-node Bernoulli-Euler

plane beam element with four DOFs.

§12.5.1. Finite Element Trial Functions

The freedoms (12.9) are used to define uniquely the variation of the transverse displacement v e (x)
over the element. The C 1 continuity requirement says that both v(x) and the slope θ = v  (x) =
dv(x)/d x must be continuous over the entire member, and in particular between beam elements.
C 1 continuity can be trivially met within each element by choosing polynomial interpolation shape
functions as shown below, because polynomials are C ∞ continuous. Matching nodal displacements
and rotations with adjacent elements enforces the necessary interelement continuity.

interpenetration gap

(a) (b)

v(x) v(x)

Figure 12.9. Deflection of a clamped-SS beam discretized with four elements, grossly
exaggerated for visibility. (a) Cubic deflection elements; (b) linear deflection elements. The
latter maintains only C 0 continuity, leading to unacceptable material gap and interpenetration at
nodes.

Remark 12.2. The physical reason for C 1 continuity is illustrated in Figure 12.9, in which the lateral deflection
curve v(x) is grossly exaggerated for visibility. The left figure shows the approximation of v(x) by four cubic
functions, which maintain the required continuity. The right figure shows an attempt to approximate v(x) by
four piecewise linear functions that maintain only C 0 continuity. In this case material gap and interpenetration
occur at the nodes, as well as at the clamped left end, because section rotations jump between elements.

12–8
 §12.6 THE FINITE ELEMENT EQUATIONS

§12.5.2. Shape Functions

The simplest shape functions that meet the C 1 continuity requirement for the nodal DOF configu-
ration (12.9) are called the Hermitian cubic shape functions. The interpolation formula based on
these functions is  
v1
θ 
v e = [ Nv1
e
Nθe1 Nv2
e
Nθe2 ]  1  = Ne ue . (12.10)
v2
θ2
These shape functions are conveniently expressed in terms of the dimensionless “natural” coordinate

2x
ξ= − 1, (12.11)

where  is the element length. Coordinate ξ varies v1 = 1 Nv1e (ξ)
from ξ = −1 at node 1 (x = 0) to ξ = +1
at node 2 (x = ). Note that d x/dξ = 12  and θ1 = 1
dξ/d x = 2/. The shape functions in terms of ξ
Nθ1e (ξ)
are
e
Nv1 = 14 (1 − ξ )2 (2 + ξ ), v2 = 1
Nv2e (ξ)
Nθe1 = 1
(1 − ξ )2 (1 + ξ ),
8
(12.12)
e
Nv2 = 1
(1 + ξ )2 (2 − ξ ),
4
θ2 = 1
Nθe2 = − 18 (1 + ξ )2 (1 − ξ ).
Nθ2e (ξ)
ξ = −1 ξ=1
These four functions are depicted in Figure 12.10.
Figure 12.10. Cubic shape functions of
plane beam element.

The curvature κ that appears in U can be expressed in terms of the nodal displacements by differ-
entiating twice with respect to x:

d 2 v e (x) 4 d 2 v e (ξ ) 4 dNe e
κ= = 2 = 2 2 u = B ue = N ue . (12.13)
dx 2  dξ 2  dξ
Here B = N is the 1 × 4 curvature-displacement matrix

1 ξ ξ 
B= 6 3ξ − 1 −6  3ξ + 1 . (12.14)


Remark 12.3. The 4/2 factor in (12.13) comes from the differentiation chain rule. If f (x) is a function of x,
and ξ = 2x/ − 1, noting that d(2/)/d x = 0 one gets

0  
d f (x) d f (ξ ) dξ 2 d f (ξ ) d 2 f (x) d(2/) d f (ξ ) 2 d d f (ξ ) 4 d 2 f (ξ )
= = , = + = . (12.15)
dx dξ d x  dξ dx2 dx dξ  dx dξ 2 dξ 2

12–9
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

ClearAll[EI,l,ξ ];
Be={{6*ξ ,(3*ξ -1)*l,-6*ξ ,(3*ξ +1)*l}}/l^2;
Ke=(EI*l/2)*Integrate[Transpose[Be].Be,{ξ ,-1,1}];
Ke=Simplify[Ke]; Print["Ke for prismatic beam:"];
Print[Ke//MatrixForm];
Print[Simplify[Ke*l^2/EI]//MatrixForm];
Ke for prismatic beam:
6 EI _ 12 EI
ClearAll[q,l,ξ];
12 EI 6 EI Ne={{2*(1-ξ)^2*(2+ξ), (1-ξ)^2*(1+ξ)*l,
l3 l l3 l2 2*(1+ξ)^2*(2-ξ),−(1+ξ)^2*(1-ξ)*l}}/8;
6 EI 4 EI _ 6 EI 2 EI fe=(q*l/2)*Integrate[Ne,{ξ,-1,1}]; fe=Simplify[fe];
l2 l l2 l Print["fe^T for uniform load q:\n",fe//MatrixForm];
_ 12 EI _ 6 EI 12 EI _ 6 EI fe^T for uniform load q:
l3 l2 l3 l2 l 2q l q _ l 2q
2 EI _ 6 EI
lq
6 EI 4 EI
l2 l l2 l 2 12 2 12

Figure 12.11. Using Mathematica to form Ke for a Figure 12.12. Using Mathematica to form fe for
prismatic beam element. uniform transverse load q.

§12.6. The Finite Element Equations

Insertion of (12.12) and (12.14) into the TPE functional specialized to this element, yields the
quadratic form in the nodal displacements
e = 12 (ue )T Ke ue − (ue )T fe , (12.16)
where
   1
K =
e
E I B B dx =
T
E I BT B 12  dξ, (12.17)
0 −1

is the element stiffness matrix and

   1
f =
e
N q dx =
T
NT q 12  dξ, (12.18)
0 −1

is the consistent element node force vector. The calculation of the entries of Ke and fe for prismatic
beams and uniform load q is studied next. More complex cases are treated in the Exercises.
§12.6.1. The Stiffness Matrix of a Prismatic Beam
If the bending rigidity E I is constant over the element it can be moved out of the ξ -integral in
(12.17):
 6ξ 
 1  1  
E I  3ξ − 1   6ξ −6ξ 
Ke = 12 E I  BT B dξ =  −6ξ  3ξ − 1 3ξ + 1 dξ. (12.19)
−1 2 −1    

3ξ + 1
Expanding and integrating over the element yields
   
36ξ 2 6ξ(3ξ −1) −36ξ 2 6ξ(3ξ +1) 12 6 −12 6
 1
 (3ξ −1)  −6ξ(3ξ −1) (9ξ −1) 
2 2 2 2
42 −6 22 
Ke = 3
EI   dξ = E I 
 
2 −1  36ξ 2 −6ξ(3ξ +1)  3 12 −6
symm (3ξ +1)2 2 symm 42
(12.20)

12–10
 §12.6 THE FINITE ELEMENT EQUATIONS

Although the foregoing integrals can be easily carried out by hand, it is equally expedient to use a
CAS such as Mathematica or Maple. For example the Mathematica script listed in the top box of
Figure 12.11 processes (12.20) using the Integrate function. The output, shown in the bottom
box, corroborates the hand integration result.

§12.6.2. Consistent Nodal Force Vector for Uniform Load

If q does not depend on x it can be moved out of (12.18), giving
 1   
4
(1 − ξ )2 (2 + ξ ) 1
 1  1 1 
 8 (1 − ξ ) (1 + ξ ) 
2  1 
f = 2 q
e 1
N dξ = 2 q
T 1
 1  dξ = 1
q   6 . (12.21)
−1 −1  4 (1 + ξ ) (2 − ξ ) 
2 2  1 
− 18 (1 + ξ )2 (1 − ξ ) − 16 

This shows that a uniform load q over the beam element maps to two transverse node loads q/2, as
may be expected, plus two nodal moments ±q2 /12. The latter are called the fixed-end moments in
the structural mechanics literature.3 The hand result (12.21) can be verified with the Mathematica
script of Figure 12.12, in which fe is printed as a row vector to save space.
;;
;;;;;;
(b) Load case I M
(a) 1 x 1 2
y,v EI constant
;;

;;

(c) Load case II P

A x B 1 x 1 2
L (d) Load case III q uniform
;;

1 x 1 2

Figure 12.13. Cantilever beam problem for Example 12.1: (a) structure,
(b-c): one-element FEM idealizations for three load cases.

Example 12.1. To see the beam element in action consider the cantilever illustrated in Figure 12.13(a). The
beam is prismatic with constant rigidity E I and span L. It is discretized with a single element as shown
in Figure 12.13(b,c,d), and subjected to the three load cases pictured there. Case I involves an applied end
moment M, case II a transverse end force P, and case III a uniformly distributed load q over the entire beam.
The FEM equations are constructed using the stiffness matrix (12.20) with  = L.
For the first two load cases, forces at end node 2 are directly set up from the given loads since no lumping is
needed. Applying the support conditions v1 = θ1 = 0 gives the reduced stiffness equations
E I  12 −6L

v2I
 
0
 E I  12 −6L

v2I I
 
P

= , = , (12.22)
L 3 −6L 4L 2 θ2I M L 3 −6L 4L 2 θ2I I 0

3 Introduced by Hardy Cross in 1930 (long before FEM) as a key ingredient for his moment distribution method. Indeed
the title of his famous paper [174] is “Analysis of continuous frames by distributing fixed-end moments.”

12–11
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

EI v(x)/(qL4 ) EI θ(x)/(qL 3 ) M(x)/(qL2 )

0.2
Section rotation Bending moment
Lateral deflection 0.3 Analytical 0.5 Analytical
Analytical 0.25 FEM, β = 0 FEM, β = 0
0.15 FEM, β = 0 FEM, β = 1 0.4
FEM, β = 1 0.2 FEM, β = 1
0.1 0.3
0.15
0.2
0.1
0.05
0.05 0.1
x/L x/L x/L
0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1

Figure 12.14. FEM versus analytical solutions for load case III of Example 12.1.

for load cases I and II, respectively. Solving gives the tip deflections v2I = M L 2 /(2E I ) and v2I I = P L 3 /(3E I ),
and the tip rotations θ2I = M L/E I and θ2I I = P L 2 /(2E I ). These agree with the analytical values provided
by Bernoulli-Euler beam theory. Thus a one-element idealization is sufficient for exactness. The reason is that
the analytical deflection profiles v(x) are quadratic and cubic polynomials in x for cases I and II, respectively.
Both are included in the span of the element shape functions. Displacements v(x), rotations θ (x) and moments
M(x) expressed as functions of x also agree with the analytical solution, as may be expected.

The results for load case III are more interesting since now the exact deflection is a quartic polynomial, which
lies beyond the span of the FEM shape functions. A dimensionless parameter 0 ≤ β ≤ 1 is introduced in the
reduced stiffness equations to study the effect of load lumping method on the solution:

E I  12 −6L

v2I I I
 
1

= 1
qL . (12.23)
L 3 −6L 4L 2 θ2I I I 2 − 16 β L

Setting β = 1 gives the energy consistent load lumping (12.21) whereas β = 0 gives the EbE (here same
as NbN) load lumping f 2I I I = 12 q L with zero fixed-end moments. The solution of (12.23) is v2I I I =
q L 4 (4 − β)/(24 E I ) and θ2I I I = q L 3 (3 − β)/(12 E I ). From this one recovers the displacement, rotation
and bending moment over the beam as

L(6−β) − 2x L(6−β) − 3x q L 
v I I I (x) = q L 2 x 2 , θ I I I (x) = q L x , M I I I (x) = L(6−β) − 6x .
24 E I 12 E I 12
(12.24)
The analytical (exact) solution is

q x 2 (3L 2 −3L x+x 2 ) q x (6L 2 −4L x+x 2 )

vex
III
(x) = , θexI I I (x) = , MexI I I (x) = 1
2
q(L − x)2 . (12.25)
24 E I 6 EI

The FEM and analytical solutions (12.24)-(12.25) are graphically compared in Figure 12.14. Deflections and
rotations obtained with the consistent load lumping β = 1 agree better with the analytical solution. In addition
the nodal values are exact (a superconvergence result further commented upon in the next Example). For the
bending moment the values provided by the EbE lumping β = 0 are nodally exact but over the entire beam
the β = 1 solution gives a better linear fit to the parabolic function MexI I I (x).

Example 12.2. The second example involves a simply supported beam under uniform line load q, depicted in
Figure 12.15(a). It is prismatic with constant rigidity E I , span L, and discretized with two elements of length
L 1 = L(/ + α) and L 2 = L − L 1 = L(/ − α), respectively. (Ordinarily two elements of the same length
/ L would be used; the scalar α ∈ (−/, /) is introduced to study the effect of unequal element sizes.)

12–12
 §12.6 THE FINITE ELEMENT EQUATIONS

(a) (b)
y,v EI constant y,v
q (uniform)

A x B 1 x 1 2 2 3

;
;;

;;
L1=L(12 +α) L2=L(12 −α)
L
Figure 12.15. SS beam problem for Example 12.2: (a) structure,
(b) two-element FEM idealization.

Using (12.20) and (12.21) to form the stiffness and consistent forces for both elements, assembling and applying
the support conditions v1 = v3 = 0, provides the reduced stiffness equations
 8L 2 −24L 4L 2 
0  L(1+2α)2 

1+2α (1+2α)2 1+2α
θ 
 −24L 192(1+12α 2 ) 192Lα 24L  1  24 
EI 
 (1+2α)
2
(1−4α 2 )3 (1−4α 2 )2 2 v 
(1−2α)  q L 

1 
.
θ  = (12.26)
2
3 
L  4L 2
192Lα 16L 2 4L 2   Lα
− 3 
 2 2  
 1+2α (1−4α 2 )2 1−4α 2 1−2α  θ3 2
L(1−2α)
0 24L 4L 2 8L 2 − 24
(1−2α)2 1−2α 1−2α
Solving for the lateral displacement of node 2 gives v2 = q L 4 (5 − 24α 2 + 16α 4 )/(384E I ). The exact
deflection is v(x) = q L 4 (ζ − 2ζ 3 + ζ 4 )/(24E I ) with ζ = x/L. Replacing x = L 1 = L(/ + α) yields
v2exact = q L 4 (5 − 24α 2 + 16α 4 )/(384E I ), which is the same as the FEM result. Likewise θ2 is exact.
The result seems prima facie surprising. First, since the analytical solution is a quartic polynomial in x we
have no reason to think that a cubic element will be exact. Second, one would expect accuracy deterioration
as the element sizes differ more and more with increasing α. The fact that the solution at nodes is exact for
any combination of element lengths is an illustration of superconvergence, a phenomenon already discussed in
§11.5. A general proof of nodal exactness is given in §13.7, but it does require advanced mathematical tools.
Note that displacements and rotations inside elements will not agree with the exact one; this can be observed
in Figure 12.14(a,b) for load case III of the previous example.

y,v (a)
EI constant q(x)=−w constant

A x B C D E
;
;
;
;
;

L1= 12 Lα L2= 12 L(1−α) L3= 12 L(1−α) L4= 12 Lα

L/2 L/2
L
y,v (b) _ (c)
;;
;;

w W
1.5
1.25
1 x 1 2 2 3
;;
;;

1
0.75
0.5
L1= 12 Lα L2= 12 L(1−α) 0.25 α
L/2 0.1 0.2 0.3 0.4 0.5

Figure 12.16. Continuum beam problem for Example 12.3,

(a): structure, (b) two-element FEM model of half beam, (c)
scaled external energy of FEM model as function of α.

12–13
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

Example 12.3. (Adapted from a driven-tank experiment by Patrick Weidman). This example displays the
advantages of symbolic computation for solving a problem in geometric design: optimal location of supports.
The prismatic continuous beam shown in Figure 12.16(a) is free at ends A and E, and simply supported at B,
C and D. The beam has total span L and constant bending rigidity E I . It is loaded by a uniform distributed
load q(x) = −w. Support C is at midspan whereas B and D are at distances L 1 = L 4 = 12 Lα from the left
and right free ends, respectively. Here 0 ≤ α < 1 is a design parameter to be determined as discussed later.
Since the problem is symmetric about midspan C only one half of the structure, say AC, need to be discretized.
The finite element model of this portion is shown in Figure 12.16(b). It has two beam elements and three nodes
placed at A, B and C, respectively. Element lengths depend on the design parameter α, which is carried along
as a variable. The six degrees of freedom are collected in u = [ v1 θ1 v2 θ2 v3 θ3 ]T . The master stiffness
equations are
 24 6L 
− 243 6L 0 0
α3 α2 α α2
      
 6L 2L 2 − 6L L2  v1 −α
 α2 α α 0 0  0
 α 
2
 θ   −Lα 2   0 
  1     
 24 6L 24 1−3α α̂ − 6L(1−2α) − 24 6L
4E I  − α 3 − α 2
  v  wL  −1 12
  fr 
α 3 α̂ 3 α 2 α̂ 2 α̂ 3 α̂ 2  2     2
   =  +
L 3  6L   θ2  4  L(2α−1)   0 
  
2 6L(1−2α) 2 2
 2 L − 2L − 6L2 L    12
 α α α α̂
2 2 α α̂ α̂ α̂  v3  −α̂   f 3r 
 
 0 0 − 324 6L
− 2 24 − 6L2  θ3 L α̂ 2 m r3
 α̂ α̂ α̂ 3 α̂  12
2 2
0 0 6L L 6L
− 2 2L
α̂ 2 α̂ α̂ α̂
(12.27)
in which α̂ = 1 − α. Note that reaction forces are carefully segregated in (12.27) to simplify application of
the general recovery technique discussed in §3.4.3. The support BCs are v2 = v3 = θ3 = 0, where the latter
comes from the symmetry condition at C. Removing those freedoms provides the reduced stiffness equations
 24 6L 6L 
   
α3 α2 α2 v −α
  1
4E I  6L 2L 2 L 2  wL  −Lα 2 
 α α   θ  =  . (12.28)
L3  α
2 1 12
 4
6L L 2 2L 2 θ2 L(2α−1)
α2 α α α̂ 12

Solving yields

wL 4   wL 3   wL 3
v1 = − α (1 + α)3 − 2 , θ1 = (1 + α)3 − 2 , θ2 = α̂ (1 − 2α − 5α 2 ). (12.29)
768 E I 384 E I 384 E I
The complete solution is u = [ v1 θ1 0 θ2 0 0 ]T . Inserting into (12.27) and solving for reactions gives

wL 3 + 2α + α 2 wL 5 − 10α − α 2 wL 2
fr 2 = , fr 3 = , mr 3 = − (1 − 2α − α 2 ). (12.30)
16 α̂ 16 α̂ 32
whence the support reactions follow as R B = fr 2 and RC = 2 fr 3 . It remains to find the best α. Of course
“best” depends on the optimality criterion. Four choices are examined below.
Minimum External Energy. The external energy at equilibrium is W (α) = fT u = w 2 L 5 W̄ (α)/(18432 E I ),
in which W̄ (α) = 1 − 5α − 2α 2 + 26α 3 + 5α 4 + 3α 5 . Minimizing W with respect to α may be interpreted
as finding the stiffest structure (in the energy sense) under the given load vector f. A plot of W̄ (α) over
0 ≤ α ≤ 12 clearly displays a minimum at α ≈ 0.27 as shown in Figure 12.16(c). Solving the quartic equation
d W̄ /dα = 0 gives one positive real root in the range α ∈ [0, 1), which to 5 places is αbest = 0.26817.

12–14
 §12. Notes and Bibliography

Equal Reactions. A second choice is to require that supports at B and C take the same load: R B = RC (note
that, because of symmetry, R D = R B ). Setting fr 2 = 2 fr 3 with their expressions taken from (12.30),√yields
3 + 2α + α 2 = 10 − 20α − 2α 2 , or 7 − 22α − 3α 2 = 0. This quadratic has the roots α = 13 (−11 ± 142).
The positive real root αbest = 0.30546 makes R B = RC = R D = wL/3, as may be expected.
Minimum Relative Deflection. Consider two sections located at xi and x j , in which {xi , x j } ∈ [0, 12 L], with
lateral displacements vi = v(xi ) and v j = v(x j ), respectively. The maximum relative deflection is defined as
ji (α) = max |v j − vi | for a fixed α. We seek the α ∈ [0, 1) that minimizes v ji (α). The computations are
v max max

far more complex than for the previous two criteria and are the subject of Exercise 12.11. Result: the best α
is the positive real root of 4 + 11α − 81α 2 − 49α 3 − 47α 4 = 0, which to 5 places is αbest = 0.26681. If this
value is adopted, the relative deflection does not exceed vimax
j < wL 4 /(67674E I ).
Minimum Absolute Moment. Let M(x, α) denote the bending moment function recovered from the FEM
solution for a fixed α. The maximum absolute moment is M max (α) = max |M(x, α)| for x ∈ [0, 12 L]. We
seek an α ∈ [0, 1) that minimizes it. This is the topic of Exercise 12.12. This problem is less well posed than
the previous one because M(x, α) varies linearly over each element, is nonzero at node 1 and discontinous at
node 2. On the other hand, the exact bending moment varies parabolically, is zero at node 1 and continuous
at node 2. Result: using the FEM-recovered M(x, α) and taking the average M at node 2, one finds that the
best α is the positive root of 2 − 4α − 15α 2 = 0, or αbest = 0.25540, for which M max < wL 2 /589. The
optimal solution using the exact moment distribution, however, is quite different. This is an intrinsic weakness
of displacement-based FEM since internal forces are obtained by differentiation, which boosts errors. To get
a better result a finer mesh would be needed.
In summary, the optimal α from the foregoing criteria varies between 0.255 to 0.306. As a reasonable
compromise an engineer could pick αbest ≈ 0.28.
Notes and Bibliography
The Bernoulli-Euler (BE) beam model synthesizes pioneer work by Jacob and Daniel Bernoulli as well as that
of Leonhard Euler in the XVIII Century. Although the model was first enunciated by 1750, it was not applied
in structural design and analysis until the second half of the XIX Century. While Galileo Galilei is credited
with first attempts at a theory, recent studies [43] argue that Leonardo da Vinci made crucial observations a
century before Galileo. However, da Vinci lacked Hooke’s law and calculus to complete the theory.
A comprehensive source of stiffness and mass matrices of plane and spatial beams is the book by Przemieniecki
[603]. The derivation of stiffness matrices is carried out there using differential equilibrium equations rather
than energy methods. This was in fact the common practice before 1962, as influenced by the use of transfer
matrix methods [578] on the limited memory computers of the time. Results for prismatic elements, however,
are identical.
Energy derivations were popularized by Archer [35,36], Martin [473] and Melosh [490,491].

References
Referenced items have been moved to Appendix R.

12–15
Chapter 12: VARIATIONAL FORMULATION OF PLANE BEAM ELEMENT

Homework Exercises for Chapter 12

Variational Formulation of Plane Beam Element

EXERCISE 12.1 [A/C:20] Use (12.17) to derive the element stiffness matrix Ke of a Hermitian beam element
of variable bending rigidity given by the inertia law
x x
I (x) = I1 (1 − ) + I2 = I1 12 (1 − ξ ) + I2 12 (1 + ξ ). (E12.1)
 
Use of Mathematica or similar CAS tool is recommended since the integrals are time consuming and error
prone. Mathematica hint: write

EI = EI1*(1-ξ )/2 + EI2*(1+ξ )/2; (E12.2)

and keep EI inside the argument of Integrate. Check whether you get back (12.20) if EI=EI1=EI2. If you
use Mathematica, this check can be simply done after you got and printed the tapered beam Ke, by writing
ClearAll[EI]; Ke=Simplify[ Ke/.{EI1->EI,EI2->EI}]; and printing this matrix.4
EXERCISE 12.2 [A/C:20] Use (12.18) to derive the consistent node force vector fe for a Hermitian beam
element under linearly varying transverse load q defined by
x x
q(x) = q1 (1 − ) + q2 = q1 12 (1 − ξ ) + q2 12 (1 + ξ ). (E12.3)
 
Again use of a CAS is recommended, particularly since the polynomials to be integrated are quartic in ξ , and
hand computations are error prone. Mathematica hint: write

q = q1*(1-ξ )/2 + q2*(1+ξ )/2; (E12.4)

and keep q inside the argument of Integrate. Check whether you get back (12.21) if q1 = q2 = q (See
previous Exercise for Mathematica procedural hints).
EXERCISE 12.3 [A:20] Obtain the consistent node force vector fe of a Hermitian beam element subject to
a transverse point load P at abscissa x = a where 0 ≤ a ≤ . Use the Dirac’s delta function expression

q(x) = P δ(a) and the fact that for any continuous function f (x), 0 f (x) δ(a) d x = f (a) if 0 ≤ a ≤ .
Check the special cases a = 0 and a = .
EXERCISE 12.4 [A:25] Derive the consistent node force vector fe of a Hermitian beam element subject to a
linearly varying z-moment m per unit length, positive CCW, defined by the law m(x) = m 1 (1 − ξ )/2 + m 2 (1 +
ξ )/2. Use the fact that the external work per unit length is m(x)θ(x) = m(x) v  (x) = (ue )T (dN/d x)T m(x).
For arbitrary m(x) show that this gives
   1  1
∂NT ∂NT 2 1
f =
e
m dx = m  dξ = NξT m dξ, (E12.5)
0 ∂x −1 ∂ξ  2 −1

where NξT denote the column vectors of beam shape function derivatives with respect to ξ . Can you see a
shortcut that avoids the integral altogether if m is constant?
EXERCISE 12.5 [A:20] Obtain the consistent node force vector fe of a Hermitian beam element subject to
a concentrated moment (“point moment”, positive CCW) C applied at x = a. Use the Concentrated moment
load on beam element expression (E12.5) in which m(x) = C δ(a), where δ(a) denotes the Dirac’s delta
function at x = a. Check the special cases a = 0, a =  and a = /2.

4 ClearAll[EI] discards the previous definition (E12.2) of EI; the same effect can be achieved by writing EI=. (dot).

12–16
 Exercises

EXERCISE 12.6 [A/C:25] Consider the one-dimensional Gauss integration rules.5

 1
.
One point : f (ξ ) dξ = 2 f (0). (E12.6)
−1

 1
. √ √
Two points: f (ξ ) dξ = f (−1/ 3) + f (1/ 3). (E12.7)
−1
  5 
1
. 5 8
Three points: f (ξ ) dξ = f (− 3/5) + f (0) + f ( 3/5). (E12.8)
−1 9 9 9
Try each rule on the monomial integrals
 1  1  1
dξ, ξ dξ, ξ 2 dξ, . . . (E12.9)
−1 −1 −1

until the rule fails. In this way verify that rules (E12.6), (E12.7) and (E12.8) are exact for polynomials of
degree up to 1, 3 and 5, respectively. (Labor-saving hint: for odd monomial degree no computations need to
be done; why?).

EXERCISE 12.7 [A/C:25] Repeat the derivation of Exercise 12.1 using the two-point Gauss rule (E12.7) to
evaluate integrals in ξ . A CAS is recommended. If using Mathematica you may use a function definition to
1
save typing. For example to evaluate −1 f (ξ ) dξ in which f (ξ ) = 6ξ 4 − 3ξ 2 + 7, by the 3-point Gauss rule
(E12.8), say
f[ξ ]:=6ξ ^4-3ξ ^2+7; int=Simplify[(5/9)*(f[-Sqrt[3/5]]+f[Sqrt[3/5]])+(8/9)*f[0]];
and print int. To form an element by Gauss integration define matrix functions in terms of ξ , for example
Be[ξ ], or use the substitution operator /., whatever you prefer. Check whether one obtains the same answers
as with analytical integration, and explain why there is agreement or disagreement. Hint for the explanation:
consider the order of the ξ polynomials you are integrating over the element.

EXERCISE 12.8 [A/C:25] As above but for Exercise 12.2.

EXERCISE 12.9 [A/C:30] Derive the Bernoulli-Euler beam stiffness matrix (12.20) using the method of
differential equations. To do this integrate the homogeneous differential equation E I v  = 0 four times over a
cantilever beam clamped at node 1 over x ∈ [0, ] to get v(x). The process yields four constants of integration
C1 through C4 , which are determined by matching the two zero-displacement BCs at node 1 and the two force
BCs at node 2. This provides a 2 × 2 flexibility matrix relating forces and displacements at node j. Invert to
get a deformational stiffness, and expand to 4 × 4 by letting node 1 translate and rotate.

EXERCISE 12.10 [C:20] Using Mathematica, repeat Example 12.2 but using EbE lumping of the distributed
force q. (It is sufficient to set the nodal moments on the RHS of (12.26) to zero.) Is v2 the same as the exact
analytical solution? If not, study the ratio v2 /v2exact as function of α, and draw conclusions.

EXERCISE 12.11 [C:25] For the continuous beam of Example 12.3, verify the results given there for the
optimal α that minimizes the maximum relative deflection. Plot the deflection profile when α = αbest .

EXERCISE 12.12 [C:25] For the continuous beam of Example 12.3, verify the results given there for the
optimal α that minimizes the absolute bending moment. Plot the moment diagram when α = αbest .

5 Gauss integration is studied further in Chapter 17.

12–17
 13
Advanced
One-Dimensional
Elements

13–1
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

TABLE OF CONTENTS

Page
§13.1. Introduction 13–3
§13.2. Generalized Interpolation 13–3
§13.2.1. Legendre Polynomials . . . . . . . . . . . . . . 13–3
§13.2.2. Generalized Stiffnesses . . . . . . . . . . . . . 13–4
§13.2.3. Transforming to Physical Freedoms: BE Model . . . . . . 13–4
§13.2.4. Transforming to Physical Freedoms: Shear-Flexible Model . . 13–5
§13.2.5. Hinged Plane Beam Element . . . . . . . . . . . . 13–5
§13.2.6. Timoshenko Plane Beam Element . . . . . . . . . . 13–6
§13.2.7. Shear-Curvature Recovery . . . . . . . . . . . . . 13–7
§13.2.8. Beam on Elastic Supports . . . . . . . . . . . . . 13–9
§13.3. Interpolation with Homogeneous ODE Solutions 13–11
§13.3.1. Exact Winkler/BE-Beam Stiffness . . . . . . . . . . 13–11
§13.4. Equilibrium Theorems 13–16
§13.4.1. Self-Equilibrated Force System . . . . . . . . . . . 13–16
§13.4.2. Handling Applied Forces . . . . . . . . . . . . . 13–16
§13.4.3. Flexibility Equations . . . . . . . . . . . . . . 13–17
§13.4.4. Rigid Motion Injection . . . . . . . . . . . . . . 13–19
§13.4.5. Applications . . . . . . . . . . . . . . . . . 13–19
§13.5. Flexibility Based Derivations 13–20
§13.5.1. Timoshenko Plane Beam-Column . . . . . . . . . . 13–20
§13.5.2. Plane Circular Arch in Local System . . . . . . . . . 13–21
§13.5.3. Plane Circular Arch in Global System . . . . . . . . . 13–24
§13.6. *Accuracy Analysis 13–27
§13.6.1. *Accuracy of Bernoulli-Euler Beam Element . . . . . . 13–27
§13.6.2. *Accuracy of Timoshenko Beam Element . . . . . . . . 13–29
§13. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 13–30
§13. References . . . . . . . . . . . . . . . . . . . . . . 13–31
§13. Exercises . . . . . . . . . . . . . . . . . . . . . . 13–32

13–2
 §13.2 GENERALIZED INTERPOLATION

§13.1. Introduction
This Chapter develops special one-dimensional elements, such as thick beams, arches and beams on elastic
foundations, that require mathematical and modeling resources beyond those presented in Chapters 11–12.
The techniques used are less elementary,1 and may be found in books on Advanced Mechanics of Materials,
e.g. [91,100]. Readers are expected to be familiar with ordinary differential equations and energy methods.
The Chapter concludes with beam accuracy analysis based on the modified equation method.
All of this Chapter material would be normally bypassed in an introductory finite element course. It is primarily
provided for offerings at an intermediate level, for example a first graduate FEM course in Civil Engineering.
Such courses may skip most of Part I as being undergraduate material.

§13.2. Generalized Interpolation

For derivation of special and C 0 beam elements it is convenient to use a transverse-displacement cubic inter-
polation in which the nodal freedoms v1 , v2 , θ1 and θ2 are replaced by generalized coordinates c1 to c4 :

v(ξ ) = Nc1 c1 + Nc2 c2 + Nc3 c3 + Nc4 c4 = Nc c. (13.1)

Here Nci (ξ ) are generalized shape functions that satisfy the completeness requirement discussed in Chapter 19.
Nc is a 1×4 matrix whereas c is a column 4-vector. Formula (13.1) is a generalized interpolation. It includes
the Hermite interpolation (12.10–12.12) as an instance when c1 = v1 , c2 = v1 , c3 = v2 and c4 = v2 .

§13.2.1. Legendre Polynomials

An obvious generalized interpolation is the ordinary cubic polynomial v(ξ ) = c1 + c2 ξ + c3 ξ 2 + c4 ξ 3 , but this
turns out not to be particularly useful. A more seminal expression is

v(ξ ) = L 1 c1 + L 2 c2 + L 3 c3 + L 4 c4 = L c, (13.2)

where the L i are the first four Legendre polynomials

L 1 (ξ ) = 1, L 2 (ξ ) = ξ, L 3 (ξ ) = 12 (3ξ 2 − 1), L 4 (ξ ) = 12 (5ξ 3 − 3ξ ). (13.3)

Here c1 through c4 have dimension of length. Functions (13.3) and their first two ξ -derivatives are plotted in
Figure 13.1. Unlike the shape functions (12.12), the L i have a clear physical meaning:
L1 Translational rigid body mode.
L2 Rotational rigid body mode.
L3 Constant-curvature deformation mode, symmetric with respect to ξ = 0.
L4 Linear-curvature deformation mode, antisymmetric with respect to ξ = 0.
These properties are also shared by the standard polynomial c1 + c2 ξ + c3 ξ 2 + c4 ξ 3 . What distinguishes the
set (13.3) are the orthogonality properties
 
48
Q0 = L L d x = diag [ 1 1/3 1/5 1/7 ] ,
T
Q2 = (L )T L d x = diag [ 0 0 3 25 ] ,
0 0 3

14400 d(.)
Q3 = (L )T L d x = diag [ 0 0 0 1 ] , in which (.) ≡ .
0 5 dx
(13.4)
Qn is called the covariance matrix for the n th derivative of the Legendre polynomial interpolation. The

first-derivative covariance Q1 = 0 (L )T L d x is not diagonal, but this matrix is not used here.

1 They do not reach, however, the capstone level of Advanced Finite Element Methods [?].

13–3
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

15
1 6 dL4 /dξ d 2L4 /dξ 2
L1 10
L4 4 dL3 /dξ d 2L3 /dξ 2
L3 0.5 L2 5
dL 2 /dξ
2 −1 −0.5 0.5 1
−1 −0.5 0.5 1 −5
−0.5 −1 −0.5 0.5 1 −10 d 2L1 /dξ2
dL1 /dξ −2 and
−15
−1 d 2L 2 /dξ2

Figure 13.1. The Legendre polynomials and their first two ξ -derivatives shown over ξ ∈[−1, 1]. Those
interpretable as beam rigid body modes (L 1 and L 2 ) in black; deformational modes (L 3 and L 4 ) in color.

Remark 13.1. The notation (13.2)–(13.3) is FEM oriented. L 1 through L 4 are called P0 through P3 in the mathematical
literature; e.g. Chapter 22 of the handbook [2]. The general definition for n = 0, 1 . . . is
n  
   n  2
  n 2n−2k 
n/2

n −n−1 1−ξ k 1 n 1
L n+1 (ξ ) ≡ Pn (ξ ) = = n (ξ −1) n−k
(ξ +1) = n
k
ξ n−2k .
k k 2 2 k 2 k n
k=0 k=0 k=0
n
(13.5)
where k is the binomial coefficient. Legendre polynomials are normalized by Pn (1) = 1, Pn (−1) = (−1)n . They can
also be indirectly defined by generating functions such as

∞
1 
∞
1 
Pn (ξ ) z = n
 , or alternatively Pn (ξ ) z n = e x z J0 (z 1 − ξ 2 ). (13.6)
1 − 2ξ z + z 2 n!
k=0 k=0

They can also be defined through a 3-term recurrence relation: (n + 2)Pn+2 (ξ ) − (2n + 3) Pn+1 (ξ ) + (n + 1) Pn (ξ ) = 0
started with P0 (ξ ) = 1 and P1 (ξ ) = ξ . One important application of these polynomials in numerical analysis is the
construction of one-dimensional Gauss integration rules: the abscissas of the n-point rule are the zeros of L n+1 (ξ ) = Pn (ξ ).

§13.2.2. Generalized Stiffnesses

The beam stiffness matrix expressed in terms of the ci is called a generalized stiffness. Denote the beam
bending and shear rigidities by R B and R S , respectively. Then Kc = KcB + KcS , where KcB comes from the
bending energy and KcS from the shear energy. For the latter is its assumed that the mean shear distortion γ at
a cross section is γ = ϒ 2 v  , where ϒ is a dimensionless coefficient that depends on the mean-shear model
used. Then  
KcB = R B (L )T L d x, KcS = R S ϒ 2 4 (L )T L d x. (13.7)
0 0

In the case of a Bernoulli-Euler (BE) beam, the shear contribution is dropped: Kc = KcB . Furthermore if
the element is prismatic, R B = E I is constant. If so KcS = R B and KcB = E I Q2 , where Q2 is the second
diagonal matrix in (13.4). With view to future use it is convenient to differentiate between symmetric and
antisymmetric bending rigidities R Bs and R Ba , which are associated with the responses to modes L 3 and L 4 ,
respectively. Assuming R Bs and R Ba to be uniform along the element we get
144R Bs 1200R Ba
Kc = KcBs + KcBa , KcBs = diag [ 0 0 1 0 ] , KcBa = diag [ 0 0 0 1 ] , (13.8)
3 3
If shear flexibility is accounted for, the contribution KcS of (13.7) is kept. Assuming R S to be constant over
the element, Kc is split into 3 contributions (two bending and one shear):
14400R S ϒ 2
Kc = KcBs + KcBa + KcS , with KcS = R S ϒ 2 4 Q3 = diag [ 0 0 0 1 ] . (13.9)

13–4
 §13.2 GENERALIZED INTERPOLATION

§13.2.3. Transforming to Physical Freedoms: BE Model

For a BE beam model, the generalized coordinates ci of (13.2) can be connected to the physical DOFs by
         
v1 1 −1 1 −1 c1 c1 30 5 30 −5 v1
 θ1   0 2/ −6/ 12/   c2   c2  1  −36 −3 36 −3   θ1 
v  = 1 1 1 1   c3 
, c  =  0 −5 0 5   v2 
. (13.10)
2 3 60
θ2 0 2/ 6/ 12/ c4 c4 6 3 −6 3 θ2

In compact form: ue = G B c and c = H B ue , with H B = G−1  

B . Here θ1 ≡ v1 and θ2 ≡ v2 , which reflects the
fundamental “plane sections remain plane” kinematic assumption of the BE model. The physical stiffness is
 
12Ra 6Ra −12Ra 6Ra

1  6Ra (3Ra + Rs ) 2 −6Ra (3Ra − Rs ) 2 
Ke = HTB KcBs + KcBa H B = 3  . (13.11)
−12Ra −6Ra 12Ra −6Ra 
6Ra (3Ra − Rs ) 2 −6Ra (3Ra + Rs ) 2
If Rs = Ra = E I the well known stiffness matrix (12.20) is recovered, as can be expected. The additional
freedom conferred by (13.11) is exhibited later in two unconventional applications.

§13.2.4. Transforming to Physical Freedoms: Shear-Flexible Model

A shear flexible beam has mean shear distortion γ = ϒ 2 v  . If ϒ is constant and v(ξ ) interpolated by (13.2),
v  = 120 c4 / 3 . Thus γ = 120ϒc4 / is constant over the element. The end rotational freedoms become
θ1 = v1 + γ and θ2 = v2 + γ . Using = 12ϒ to simplify the algebra, the transformations (13.10) change to
         
v1 1 −1 1 −1 c1 c1 30 5 30 −5 v
3 36+30 − 3  1
 θ1   − 6 12+10  c2   c2  1  − 36+30
2 −
0 
1+  θ1 
v = 
  ,   = 
 
1+ 1+ 1+
 
2 1 1 1 1 c3 c3 60 0 −5 0 5  v2
θ2 0 2 6 12+10 c4 c4 6
− 1+ 3 6 3 θ2
1+ 1+ 1+
(13.12)
In compact form, ue = G S c and c = G−1 S u e
= H S u e
. Transforming Kc of (13.9) to physical freedoms yields
the stiffness used to construct the Timoshenko beam element in §13.2.6:
   
0 0 0 0 4 2 −4 2
R Bs  0 1 0 −1  12R Ba + R S 2 2  2 2 −2 2 
K = H S Kc H S = 0 + . (13.13)
0  −4 −2 
e T
0 0 4 3 (1 + )2 −2 4
0 −1 0 1 2 2 −2 2
y
§13.2.5. Hinged Plane Beam Element
The two-node prismatic plane BE beam element depicted in z
Figure 13.2 has a mechanical hinge at midspan (ξ = 0). The
cross sections on both sides of the hinge can rotate respect
to each other. The top figure also sketches a fabrication y
method sometimes used in short-span pedestrian bridges. hinge
x
Gaps on either side of the hinged section cuts are filled with
a bituminous material that permits slow relative rotations. 1 2 x
Both the curvature κ and the bending moment M must vanish /2 /2
at midspan. But in a element built via cubic interpolation of
v(x), κ = v  must vary linearly in both ξ and x. Figure 13.2. Beam element with hinge located
at midspan. The top figure sketches a hinge
fabrication method.

13–5
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

Consequently the mean curvature, which is controlled by the Legendre function L 2 (shown in blue on Fig-
ure 13.1) must be zero. The kinematic constraint of zero mean curvature is enforced by setting the symmetric
bending rigidity R Bs = 0 whereas the antisymmetric bending rigidity is the normal one: R Ba = E I .
Plugging into (13.11) yields
   
4 2 −4 2 2
3E I  2 2 −2 2  3E I  
Ke = 3  = 3 [ 2 −2 ]  . (13.14)
−4 −2 4 −2  −2 
2 2 −2 2

This matrix has rank one, as it can be expected from the last (dyadic) expression in (13.14). Ke has one
nonzero eigenvalue: 6E I (4 + 2 )/ 3 , and three zero eigenvalues. The eigenvector associated with the nonzero
eigenvalue pertains2 Matrix (13.14) can be derived by more sophisticated methods (e.g., mixed variational
principles) but the present technique is the most expedient one.

Example 13.1. y,v (a)

EI constant q0

;;
This example deals with the prismatic continuous beam shown
in Figure 13.3(a). This has two spans with lengths αL and L,
respectively, where α is a design parameter, and is subjected
1 x 2 3

;
;
to uniform load q0 . The beam is free, simple supported and (1) hinge (2)
fixed at nodes 1, 2 and 3, respectively. There is a hinge at the αL L/2 L/2
center of the 2–3 span. Of interest is the design question: for L
which α > 0 is the tip deflection at end 1 zero? (b)
The beam is discretized with two elements: (1) and (2), going 1 2 3
from 1 to 2 and 2 to 3, respectively, as shown in the figure. α = 0.362
The stiffnesses for elements (1) and (2) are those of (12.20)
and (13.14), respectively, whereas (12.21) is used to build the Figure 13.3. Beam problem for Example
consistent node forces for both elements. 13.1. (a): beam problem, (b) deflection profile
when α = 0.362.
Assembling and applying the support conditions v2 = v3 = θ3 = 0, provides the reduced stiffness equations
 12/α 3 6L/α 2 6L/α 2 v   α 
EI 1 q0 L
6L/α 2 4L 2 /α 2L 2 /α θ1 = Lα 2 /6 . (13.15)
L3 6L/α 2 2L 2 /α L (4 + 3α)/α
2 θ2 2 L(1 − α 2 )/6

Solving for the node displacements gives v1 = q0 L 4 α(12α 2 + 9α 3 − 2)/(72E I ), θ1 = q0 L 3 (1 − 6α 2 − 6α 3 )/(36E I )

and θ2 = q0 L 3 (1 − 6α 2 )/(36E I ). The equation v1 = 0 is quartic in α and has four roots, which to 8 places are
α1 = −1.17137898, α2 = −0.52399776, α3 = 0, and α4 = 0.3620434. Since the latter is the only positive root, the
solution is α = 0.362. The deflection profile for this value is pictured in Figure 13.3(b).

§13.2.6. Timoshenko Plane Beam Element

As observed in §12.2.2, the Timoshenko beam model [735] includes a first order correction for transverse
shear flexibility. The key kinematic assumption changes to “plane sections remain plane but not necessarily
normal to the deformed neutral surface.” This is illustrated in Figure 13.4(a) for a 2-node plane beam element.
The cross section rotation θ differs from v  by γ . Ignoring axial forces, the displacement field is analogous to
that of the Bernoulli-Euler model (12.2) but with a shear correction:
∂v V
u(x, y) = −yθ, v(x, y) = v(x), with θ= + γ = v + γ , γ = . (13.16)
∂x G As

2 Compare the vector in the last expression in (13.14) to the last row of H B in (13.10) to the antisymmetric deformational
mode L 3 .

13–6
 §13.2 GENERALIZED INTERPOLATION

(a) θ2
y, v γ (b) (c) γ
θ1 Deformed γ2 y
cross
section M
γ1 v'2 =|dv/dx|2
V(+)
v2 z x
v1 v(x)
x, u V A positive transverse
v'1 =|dv/dx|1 shear force V produces
1 2 Positive M, V a CCW rotation (+γ) of
conventions the beam cross section

Figure 13.4. Two-node Timoshenko plane beam element: (a) kinematics (when developed with cubic shape
functions, γ1 = γ2 = γ ); (b) M and V sign conventions; (c) concurrence of sign conventions for V and γ .

Here V is the transverse shear force, γ the “shear rotation” (positive CCW) averaged over the cross section, G
the shear modulus and As the effective shear area.3 The product R S = G As is the shear rigidity. To correlate
def
with the notation of §13.2.4, note that V = E I v  , γ = ϒ 2 v  = V /(G As ), so ϒ = E I /(G As 2 ) and

12E I
= 12ϒ = (13.17)
G A s 2

This dimensionless ratio characterizes the “shear slenderness” of the beam element.4 It is not an intrinsic
beam property because it involves the element length. As → 0 the Timoshenko model reduces to the BE
model. Replacing R Bs = R Ba = E I and R S = G As = 12E I /( 2 ) into (13.13) yields the Timoshenko
beam stiffness  
12 6 −12 6
EI  6 2 (4 + ) −6 2 (2 − ) 
Ke = 3  (13.18)
(1 + ) −12 −6 12 −6 
6 2 (2 − ) −6 2 (4 + )
If = 0 this reduces to (12.20). The Mathematica module TimoshenkoBeamStiffness[Le,EI, ], listed
in Figure 13.5, implements (13.18).

TimoshenkoBeamStiffness[Le_,EI_,Φ_]:=Module[{Ke},
Ke=EI/(Le*(1+Φ))*{{ 12/Le^2, 6/Le,-12/Le^2, 6/Le },
{ 6/Le , 4+Φ, -6/Le , 2-Φ },
{-12/Le^2, -6/Le, 12/Le^2,-6/Le},
{ 6/Le , 2-Φ, -6/Le , 4+Φ }};
Return[Ke]];

Figure 13.5. Module to produce stiffness matrix for Timoshenko beam element.

The calculation of the consistent node forces for uniform transverse load is covered in Exercise 13.2. A hinged
Timoshenko beam is constructed in Exercise 13.3.

3 A concept defined in Mechanics of Materials; see e.g. Chapter 10 of Popov [588] or Chapter 12 of Timoshenko and
Goodier [737]. As is calculated by equating the internal shear energy 12 V γ = 12 V 2 /(G As ) to that produced by the shear
stress distribution over the cross section. For a thin rectangular cross section and zero Poisson’s ratio, As = 5A/6.
4 Note that in (13.8)–(13.9), 1200R Bs / 3 + 14400 R S ϒ 2 / = 1200 E I (1 + )/ 3 , giving a simple interpretation for .

13–7
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

EIzz and GA s constant along beam

y,v P P

;;
;;
1 x 1 2 2 3 (a) Finite element
discretization
L/2 L/2
L
Mz = PL −
(b) Transverse
2 Vy =−P shear force Vy

+ (c) Bending
moment Mz

Figure 13.6. Example: cantilever beam discretized with two Timoshenko beam elements.

§13.2.7. Shear-Curvature Recovery

When using the Timoshenko beam model, the following arises during postprocessing. Suppose that the
element node displacement vector ue = ueT is given following the solution process. Recover the mean shear
distorsion γ e and the curvature κ e over the element on the way to internal forces and stresses. The problem
is not trivial because γ e is part of the rotational freedoms. The recovery process can be effectively done
by passing first to generalized coordinates: ce = H S ue , and then to Bernoulli-Euler node displacements:
ueB E = G B ce = G B H S ue = TBTe
ueT , in which
 
0 0 0 0
 L1e 1 1 1

T BT = −G B H S = I − 0 (13.19)
0
2 Le 2
1+ 0 0
1 1 1 1
Le 2 Le 2
def
Subtracting: uγ = ueT − ueB R = (I − G B H S )ueT = [ 0 γ e 0 γ e ]T , Explicitly
1 e
γe = (v − v2e ) + 12 (θ2e + θ1e ) (13.20)
Le 1
The curvature is obtained from the Bernoulli-Euler vector: κ = Be ueB E , where the curvature displacement
matrix Be is that given in the previous Chapter.

Example 13.2. Consider the prismatic cantilever beam of length L pictured in Figure 13.6(a). It is subject to two point
loads as shown. Shear flexibility is to be accounted for using the Timoshenko model. The bending and shear rigidities E Izz
and G As are constant along the span. The objective is to find deflections, curvatures and shear distortions and associated
bending moments and shear forces.
It is sufficient to discretize the beam with two Timoshenko beam elements oflength L/2 as shown in the figure. The stiffness
matrices for both elements are given by (13.18), in which L e = L/2 and = 12 E Izz /(G As (L/2)2 ) = 48 E Izz /(G As L 2 ).
The master stiffness equations are
 48 12L −48 12L 0 0
 v1 0
  
2 (4+ ) −12L 2 (2− ) θ
 12L L L 0 0  1   0 
2E Izz  −48 −12L 96 0 −48 12L   v2   −P 

L 3 (1+ )  12L L (2− ) 2L 2 (4+ ) −12L L 2 (2− )   = . (13.21)
  θ2   0 
2 0
0 0 −48 −12L 48 −12L v3 P
0 0 12L L 2 (2− ) −12L L 2 (4+ ) θ3 0
Setting the displacement B.C. v1 = θ1 = 0 and solving yields
P L2  L L (22 + ) 3
T
u= 0 0 1 . (13.22)
4E Izz 4 24 2

13–8
 §13.2 GENERALIZED INTERPOLATION

The mean element shear distortions are calculated from (13.22) using (13.20). This gives
P L2 P
γ (1) = 0, γ (2) = − =− (13.23)
48E Izz G As
The element-level Bernoulli-Euler node displacements are obtained from (13.22) on subtracting the shear distortions (13.23)
from the rotations:
P L2  T P L2 L T
u(1)
B = 0 0 L 1 , u(2)
B = 1+ L (22 + ) 3+ . (13.24)
4E Izz 4 4E Izz 4 12 24 2 12
(1) (2)
Note that θ B2 = 1 = θ B2 = 1 + 121
, the “kink” being due to the shear distortion jump at node 2. The curvatures are now
recovered as
PL PL
κ (1) = B(1) u(1)
B = 2E I , κ (2) = B(2) u(2)
B = 4E I (1 − ξ ),
(2)
(13.25)
zz zz
The transverse shear force resultant and bending moment are easily recovered as Vye = G As γ (e) and Mze = E Izz κ e ,
respectively. The results are drawn in Figure 13.6(b,c).

§13.2.8. Beam on Elastic Supports

Sometimes beams, as well as other structural members, may be supported elastically along their span. Two
common configurations that occur in structural engineering are:
(i) Beam resting on a continuum medium such as soil. This is the case in foundations.
(ii) Beam supported by discrete but closely spaced flexible supports, as in the “bed of springs” pictured in
Figure 13.7. This occurs in railbeds (structurally rails are beams supported by crossties) and some types
of grillworks.
The Winkler foundation is a simplified elastic-support model. It is an approximation for (i) because it ignores
multidimensional elasticity effects as well as friction. It is a simplification of (ii) because the discrete nature
of supports is smeared out. It is nonetheless popular, particularly in foundation and railway engineering, when
the presence of physical uncertainties would not justify a more complicated model. Such uncertainties are
inherent in soil mechanics.
The Winkler model may be viewed as a “continuification” of case (ii). Take a beam slice going from x to
x + d x. The spring-reaction force acting on the beam is taken to be d f F = −k F v(x) d x. Here v(x) is the
transverse deflection and k F the Winkler foundation stiffness, which has dimension of force per length-squared.
Force d f F has the opposite sign of v(x), pushing up if the beam moves down and pulling down if it moves up.
Beam-foundation separation effects that may occur in case (i) are ignored here because that would lead to a
nonlinear contact problem.
The internal energy stored in the d x slice of Winkler springs is −/ v d f F = / k F v 2 d x. Consequently the
effect of elastic supports is to modify the internal energy U Be of the beam element so that it becomes

y, v q(x)

x beam

;;;;;;;;;
bed of
springs

Figure 13.7. A beam supported by a bed of springs.

Continuification of this configuration leads to the
Winkler foundation model treated in this subsection.

13–9
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS


U =
e
U Be + U Fe , with U Fe = 1
2
k F v 2 d x. (13.26)
0

Therefore the total stiffness of the element is computed by adding the foundation stiffness to the beam stiffness.
Care must be taken, however, that the same set of nodal freedoms is used. This is best handled by doing the
generalized stiffness KcF first, and then using the appropriate generalized-to-physical transformation. If the
transverse deflection v is interpolated with (13.2) as v = L c, the generalized Winkler foundation stiffness for
constant k F is 
KcF = k F L T L d x = k F Q0 , (13.27)
0

where Q0 is the first diagonal matrix in (13.4). This holds regardless of beam model. Now if the member resting
on the foundation is modeled as a BE beam, one picks H B of (13.10) as generalized-to-physical transformation
matrix to get  
156 22 54 −13
k F  22 4 2 13 −3 2 
KeF = k F HTB Q0 H B =  54 , (13.28)
420 13 156 −22 
−13 −3 −22 4
2 2

If instead the supported member is modeled as a Timoshenko beam, one picks H S of (13.12) to get

KeF = k F HTS Q0 H S
 
4(78+147 +70 2 ) (44+77 +35 2 ) 4(27+63 +35 2 ) − (26+63 +35 2 )
kF  (44+77 +35 2 ) 2 (8+14 +7 2 ) (26+63 +35 2 ) − 2 (6+14 +7 2 ) 
=  4(27+63 +35 2 )
840(1+ )2 (26+63 +35 2 ) 4(78+147 +70 2 ) − (44+77 +35 2 ) 
− (26+63 +35 2 ) − 2 (6+14 +7 2 )− (44+77 +35 2 ) 2 (8+14 +7 2 )
(13.29)
The module TimoshenkoWinklerStiffness[Le,kF, ] listed in Figure 13.8 implements the stiffness
(13.29). To get the BE-beam Winkler stiffness (13.28), invoke with = 0. Examples of use of this module
are provided in §13.3.1.

TimoshenkoWinklerStiffness[Le_,kF_,Φ_]:=Module[{KeW},
KeW={{4*(78+147*Φ+70*Φ^2), Le*(44+77*Φ+35*Φ^2),
4*(27+63*Φ+35*Φ^2), -Le*(26+63*Φ+35*Φ^2)},
{Le*(44+77*Φ+35*Φ^2), Le^2*(8+14*Φ+7*Φ^2),
Le*(26+63*Φ+35*Φ^2), -Le^2*(6+14*Φ+7*Φ^2)},
{4*(27+63*Φ+35*Φ^2), Le*(26+63*Φ+35*Φ^2),
4*(78+147*Φ+70*Φ^2), -Le*(44+77*Φ+35*Φ^2)},
{-Le*(26+63*Φ+35*Φ^2),-Le^2*(6+14*Φ+7*Φ^2),
-Le*(44+77*Φ+35*Φ^2), Le^2*(8+14*Φ+7*Φ^2)}}*
kF*Le/(840*(1+Φ)^2); Return[KeW]];

Figure 13.8. Stiffness matrix module for a Winkler foundation supporting a Timoshenko beam element.

13–10
 §13.3 INTERPOLATION WITH HOMOGENEOUS ODE SOLUTIONS

§13.3. Interpolation with Homogeneous ODE Solutions

For both BE and Timoshenko beam models, the Legendre polynomials L 1 (ξ ) through L 4 (ξ ) are exact solutions
of the homogeneous, prismatic, plane beam equilibrium equation E I d 4 v/d x 4 = 0. When used as shape
functions in the generalized interpolation (13.2), the resulting stiffness matrix is exact if the FEM model is
loaded at the nodes, as further discussed in §13.6. The technique can be extended to more complicated one-
dimensional problems. It can be used to derive exact stiffness matrices if homogeneous solutions are available
in closed form, and are sufficiently simple to be amenable to analytical integration. The following subsection
illustrates the method for a BE beam resting on a Winkler elastic foundation.

§13.3.1. Exact Winkler/BE-Beam Stiffness

Consider again a prismatic, plane BE beam element resting on a Winkler foundation of stiffness k F , as pictured
in Figure 13.7. The governing equilibrium equation for constant E I > 0 and k F > 0 is E I d 4 v/d x 4 + k F v =
q(x). The general homogeneous solution over an element of length going from x = 0 to x = is

ζ

−ζ

4 kF
v(x) = e c1 sin ζ + c2 cos ζ + e c3 sin ζ + c4 cos ζ , with ζ = χ x/ and χ = . (13.30)
4E I

Here the ci are four integration constants to be determined from four end conditions: the nodal degrees of
freedom v1 , v1 , v2 and v2 . These constants are treated as generalized coordinates and as before collected into
vector c = [ c1 c2 c3 c4 ]T . The solution (13.30) is used as generalized interpolation with eζ sin ζ through
e−ζ cos ζ as the four shape functions. Differentiating twice gives v  = dv/d x and v  = d 2 v/d x 2 . The TPE
functional of the element in terms of the generalized coordinates can be expressed as

1

ec = 2
E I (v  )2 + 12 k F v 2 − q0 v d x = 12 cT (KcB + KcF ) c − cT fc . (13.31)
0

This defines KcB and KcF as generalized element stiffnesses due to beam bending and foundation springs,
respectively, whereas fc is the generalized force associated with a transverse load q(x). The nodal freedoms
are linked to generalized coordinates by
    
v1 0 1 0 1 c1
 θ1   χ / χ / χ / −χ /   c2 
v  =  eχ sin χ eχ cos χ e−χ sin χ e−χ cos χ c .
1 χ χ −χ −χ 3
θ1 χ e (cos χ+ sin χ ) χ e (cos χ − sin χ ) χ e (cos χ − sin χ ) −χe (cos χ + sin χ ) c4
(13.32)
In compact form this is ue = G F c. Inverting gives c = H F ue with H F = G−1 F . The physical stiffness is
Ke = KeB + KeF with KeB = HTF KcB H F and KeF = HTF KcF H F . The consistent force vector is fe = HTF fc .
Computation with transcendental functions by hand is unwieldy and error-prone, and at this point it is better
to leave that task to a CAS. The Mathematica script listed in Figure 13.9 is designed to produce KeB , KeF and
fe for constant E I and k F , and uniform transverse load q(x) = q0 . The script gives

     
B1 B2 B5 −B4 F1 F2 F5 −F4 f1
E I χ  B B B  k F  F F F  q 0  f 
KeB = 3 2  3 4 6
 , KeF = 
3 4 6
 , fe = 2  2  ,
4 g B1 −B2 16χ g 3 2 F1 −F2 χ g f1
symm B3 symm F3 − f2
(13.33)

13–11
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

ClearAll[EI,kF,Le,χ,q0,x]; g=2-Cos[2*χ]-Cosh[2*χ];
Nf={Exp[ χ*x/Le]*Sin[χ*x/Le],Exp[ χ*x/Le]*Cos[χ*x/Le],
Exp[-χ*x/Le]*Sin[χ*x/Le],Exp[-χ*x/Le]*Cos[χ*x/Le]};
Nfd=D[Nf,x]; Nfdd=D[Nfd,x];
KgF=kF*Integrate[Transpose[{Nf}].{Nf},{x,0,Le}];
KgB=EI*Integrate[Transpose[{Nfdd}].{Nfdd},{x,0,Le}];
fg= q0*Integrate[Transpose[{Nf}],{x,0,Le}];
{KgF,KgB,fg}=Simplify[{KgF,KgB,fg}];
Print["KgF=",KgF//MatrixForm]; Print["KgB=",KgB//MatrixForm];
GF=Simplify[{Nf/.x->0,Nfd/.x->0,Nf/.x->Le,Nfd/.x->Le}];
HF=Simplify[Inverse[GF]]; HFT=Transpose[HF];
Print["GF=",GF//MatrixForm]; Print["HF=",HF//MatrixForm];
facB=(EI*χ/Le^3)/(4*g^2); facF=(kF*Le)/(16*χ^3*g^2);
KeB=Simplify[HFT.KgB.HF]; KeBfac=FullSimplify[KeB/facB];
Print["KeB=",facB," * ",KeBfac//MatrixForm]
KeF=Simplify[HFT.KgF.HF]; KeFfac=FullSimplify[KeF/facF];
Print["KeF=",facF," * ",KeFfac//MatrixForm];
facf=(q0*Le)/(χ^2*g); fe=Simplify[ExpToTrig[HFT.fg]];
fefac=Simplify[fe/facf]; Print["fe=",facf," * ",fefac];

Figure 13.9. Script to produce the exact Winkler-BE beam stiffness matrix and consistent force vector.

BEBeamWinklerExactStiffness[Le_,EI_,kF_,q0_]:=Module[{B1,B2,B3,B4,B5,B6,
F1,F2,F3,F4,F5,F6,f1,f2,facB,facF,facf,KeB,KeF,fe,χ},
χ=PowerExpand[Le*((kF/(4*EI))^(1/4))];
B1 =2*χ^2*(-4*Sin[2*χ]+Sin[4*χ]+4*Sin[χ]*(Cos[χ]*Cosh[2*χ]+
8*χ*Sin[χ]*Sinh[χ]^2)+2*(Cos[2*χ]-2)*Sinh[2*χ]+Sinh[4*χ]);
B2 =2*Le*χ*(4*Cos[2*χ]-Cos[4*χ]-4*Cosh[2*χ]+ Cosh[4*χ]-
8*χ*Sin[2*χ]*Sinh[χ]^2+8*χ*Sin[χ]^2*Sinh[2*χ]);
B3 =-(Le^2*(8*χ*Cos[2*χ]-12*Sin[2*χ]+Cosh[2*χ]*(6*Sin[2*χ]-8*χ)+3*Sin[4*χ]+
2*(6-3*Cos[2*χ]+4*χ*Sin[2*χ])*Sinh[2*χ]-3*Sinh[4*χ]));
B4 =-4*Le*χ*(χ*Cosh[3*χ]*Sin[χ]-χ*Cosh[χ]*(-2*Sin[χ]+Sin[3*χ])+(χ*(Cos[χ]+
Cos[3*χ])+Cosh[2*χ]*(-2*χ*Cos[χ]+4*Sin[χ])+2*(-5*Sin[χ]+Sin[3*χ]))*Sinh[χ]);
B5 =-4*χ^2*(2*Cos[χ]*(-2+Cos[2*χ]+Cosh[2*χ])*Sinh[χ]+Sin[3*χ]*
(Cosh[χ]-2*χ*Sinh[χ])+Sin[χ]*(-4*Cosh[χ]+Cosh[3*χ]+2*χ*Sinh[3*χ]));
B6 =2*Le^2*(Cosh[3*χ]*(-2*χ*Cos[χ]+3*Sin[χ])+Cosh[χ]*(2*χ*Cos[3*χ]+3*(Sin[3*χ]-
4*Sin[χ]))+(9*Cos[χ]-3*Cos[3*χ]-6*Cos[χ]*Cosh[2*χ]+16*χ*Sin[χ])*Sinh[χ]);
F1 =2*χ^2*(-32*χ*Sin[χ]^2*Sinh[χ]^2+6*(-2+Cos[2*χ])*
(Sin[2*χ]+Sinh[2*χ])+6*Cosh[2*χ]*(Sin[2*χ]+Sinh[2*χ]));
F2 =2*Le*χ*(4*Cos[2*χ]-Cos[4*χ]-4*Cosh[2*χ]+Cosh[4*χ]+
8*χ*Sin[2*χ]*Sinh[χ]^2-8*χ*Sin[χ]^2*Sinh[2*χ]);
F3 =Le^2*(8*χ*Cos[2*χ]+4*Sin[2*χ]-2*Cosh[2*χ]*(4*χ+Sin[2*χ])-Sin[4*χ]+
2*(Cos[2*χ]+4*χ*Sin[2*χ]-2)*Sinh[2*χ]+Sinh[4*χ]);
F4 =4*Le*χ*(χ*Cosh[3*χ]*Sin[χ]-χ*Cosh[χ]*(-2*Sin[χ]+Sin[3*χ])+(χ*Cos[χ]+
χ*Cos[3*χ]+10*Sin[χ]-2*Cosh[2*χ]*(χ*Cos[χ]+2*Sin[χ])-2*Sin[3*χ])*Sinh[χ]);
F5 =-4*χ^2*(6*Cos[χ]*(-2+Cos[2*χ]+Cosh[2*χ])*Sinh[χ]+Sin[3*χ]*
(3*Cosh[χ]+2*χ*Sinh[χ])+Sin[χ]*(-12*Cosh[χ]+3*Cosh[3*χ]-2*χ*Sinh[3*χ]));
F6 =-2* Le^2*(-(Cosh[3*χ]*(2*χ*Cos[χ]+Sin[χ]))+Cosh[χ]*(2*χ*Cos[3*χ]+
4*Sin[χ]-Sin[3*χ])+(Cos[3*χ]+Cos[χ]*(2*Cosh[2*χ]-3)+16*χ*Sin[χ])*Sinh[χ]);
f1=2*χ*(Cosh[χ]-Cos[χ])*(Sin[χ]-Sinh[χ]); f2=-(Le*(Sin[χ]-Sinh[χ])^2);
g=2-Cos[2*χ]-Cosh[2*χ]; facf=(q0*Le)/(χ^2*g);
facB=(EI*χ/Le^3)/(4*g^2); facF=(kF*Le)/(16*χ^3*g^2);
KeB=facB*{{B1,B2,B5,-B4},{ B2,B3,B4,B6},{B5,B4,B1,-B2},{-B4,B6,-B2,B3}};
KeF=facF*{{F1,F2,F5,-F4},{ F2,F3,F4,F6},{F5,F4,F1,-F2},{-F4,F6,-F2,F3}};
fe=facf*{f1,f2,f1,-f2}; Return[{KeB,KeF,fe}]];

Figure 13.10. Module to get the exact BE-Winkler stiffness and consistent load vector.

13–12
 §13.3 INTERPOLATION WITH HOMOGENEOUS ODE SOLUTIONS

;;;;;;;;;;;
y, v EI constant
P : Load case (I)
q0 : Load case (II)

;;;;;;;;;;;
A C

;;;;;;;;;;;
x B

;;;;;;;;;;;
k F constant

L L
2L
Figure 13.11. Example: fixed-fixed beam on Winkler elastic foundation.

in which
g = 2 − cos 2χ − cosh 2χ,
B1 = 2χ 2 (−4 sin 2χ + sin 4χ +4 sin χ (cos χ cosh 2χ +8χ sin χ sinh2 χ )+2(cos 2χ −2) sinh 2χ + sinh 4χ ),
B2 = 2 χ (4 cos 2χ − cos 4χ −4 cosh 2χ + cosh 4χ−8χ sin 2χ sinh2 χ +8χ sin2 χ sinh 2χ ),
B3 = −( 2 (8χ cos 2χ−12 sin 2χ + cosh 2χ (6 sin 2χ −8χ )+3 sin 4χ +
2(6−3 cos 2χ +4χ sin 2χ) sinh 2χ −3 sinh 4χ )),
B4 = −4 χ (χ cosh 3χ sin χ−χ cosh χ (−2 sin χ + sin 3χ )+(χ (cos χ + cos 3χ )
+ cosh 2χ (−2χ cos χ +4 sin χ )+2(−5 sin χ+ sin 3χ )) sinh χ ),
B5 = −4χ 2 (2 cos χ (−2+ cos 2χ + cosh 2χ ) sinh χ + sin 3χ (cosh χ−2χ sinh χ )
+ sin χ(−4 cosh χ+ cosh 3χ +2χ sinh 3χ ))
B6 = 2 2 (cosh 3χ (−2χ cos χ +3 sin χ )+ cosh χ (2χ cos 3χ +3(−4 sin χ+ sin 3χ ))
+(9 cos χ −3 cos 3χ −6 cos χ cosh 2χ +16χ sin χ ) sinh χ )
F1 = 2χ 2 (−32χ sin2 χ sinh2 χ +6(−2+ cos 2χ )(sin 2χ + sinh 2χ )+6 cosh 2χ (sin 2χ + sinh 2χ)),
F2 = 2 χ (4 cos 2χ − cos 4χ −4 cosh 2χ + cosh 4χ+8χ sin 2χ sinh2 χ −8χ sin2 χ sinh 2χ ),
F3 = 2 (8χ cos 2χ+4 sin 2χ −2 cosh 2χ (4χ + sin 2χ )− sin 4χ +2(cos 2χ +4χ sin 2χ −2) sinh 2χ + sinh 4χ ),
F4 = 4 χ (χ cosh 3χ sin χ −χ cosh χ (sin 3χ −2 sin χ )+(χ cos χ +χ cos 3χ +10 sin χ
−2 cosh 2χ (χ cos χ+2 sin χ )−2 sin 3χ ) sinh χ ),
F5 = −4χ 2 (6 cos χ (−2+ cos 2χ + cosh 2χ ) sinh χ + sin 3χ (3 cosh χ+2χ sinh χ ),
+ sin χ(−12 cosh χ +3 cosh 3χ −2χ sinh 3χ ))
F6 = −2 2 (−(cosh 3χ(2χ cos χ + sin χ ))+ cosh χ (2χ cos 3χ +4 sin χ − sin 3χ )
+(cos 3χ + cos χ(2 cosh 2χ−3)+16χ sin χ ) sinh χ ),
f 1 = 2χ(cosh χ − cos χ )(sin χ − sinh χ ), f 2 = − (sin χ − sinh χ )2 .
(13.34)
These expressions are used to code module BEBeamWinklerExactStiffness[Le,EI,kF,q0], which is
listed in Figure 13.10.

Example 13.3. A fixed-fixed BE beam rests on a Winkler foundation as shown in Figure 13.11. The beam has span 2L, and
constant E I . The Winkler foundation coefficient k F is constant. As usual in foundation engineering we set

k F = E I λ4 /L 4 , (13.35)

13–13
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

Table 13.1 - Results for Example of Figure 13.11 at Selected λ Values

Load case (I): Central Point Load Load case (II): Line Load Over Right Half
λ exact C I Ne = 2 Ne = 4 Ne = 8 exact C I I Ne = 2 Ne = 4 Ne = 8
0.1 0.999997 0.999997 0.999997 0.999997 0.999997 0.999997 0.999997 0.999997
1 0.970005 0.969977 0.970003 0.970005 0.968661 0.969977 0.968742 0.968666
2 0.672185 0.668790 0.671893 0.672167 0.657708 0.668790 0.658316 0.657746
5 0.067652 0.049152 0.065315 0.067483 0.041321 0.049152 0.041254 0.041317
10 0.008485 0.003220 0.006648 0.008191 0.002394 0.003220 0.002393 0.002395
100 8.48×10−6 3.23×10−7 8.03×10−7 1.63×10−6 2.40×10−7 3.23×10−7 2.62×10−7 2.42×10−7

where λ is a dimensionless rigidity to be kept as parameter.5 The beam is subjected to two load cases: (I) a central point
load P at x = L, and (II) a uniform line load q0 over the right half x ≥ L. See Figure 13.11.
All quantities are kept symbolic. The focus of interest is the deflection vC at midspan C (x = L). For convenience this
is rendered dimensionless by taking vCI (λ) = C I (λ)vC0
I and v I I (λ) = C (λ)v I I for load cases (I) and (II)), respectively.
C II C0
Here vC0 = −P L /(24E I ) and vC0 = −q0 L /(48E I ) are the midspan deflections of cases (I) and (II) for λ = 0, that is,
I 3 I I 4

k F = 0 (no foundation). The exact deflection factors for this model are
√ √ √
6 2 cos 2λ + cosh 2λ − 2 13 4 137 8
C I (λ) = 3 √ √ =1− λ + λ ...
λ sin 2λ + sinh 2λ 420 138600
√ √ √ √ (13.36)
48 (cos λ/ 2− cosh λ/ 2)(sin λ/ 2− sinh λ/ 2) 163 4 20641 8
C I I (λ) = 4 √ √ =1− λ + λ + ...
λ sin 2λ + sinh 2λ 5040 19958400

Both load cases were symbolically solved with two exact elements of length L produced by the module of Figure 13.10.
As can be expected, the answers reproduce the exact solutions (13.36). Using any number of those elements would match
(13.36) as long as the midspan section C is at a node. Then both cases were solved with 2, 4, and 8 elements with the stiffness
(13.28) produced by cubic polynomials. The results are shown in a log-log plot in Figure 13.12. Results for selected values
of λ are presented in Table 13.1.
As can be seen, for a “soft” foundation characterized by λ < 1, the cubic-polynomial elements gave satisfactory results and
converged quickly to the exact answers, especially in load case (II). As λ grows over one, the deflections become rapidly
smaller, and the polynomial FEM results exhibit higher relative errors. On the other hand, the absolute errors remain small.
The conclusion is that exact elements are only worthwhile in highly rigid foundations (say λ > 5) and then only if results
with small relative error are of interest.

0 0
log10 CI log10 CII
−1 −1
Load Case I: Load Case II:
−2 Central Point Load −2 Uniform Line Load
2 exact elements
Over Right Half 2 exact elements
−3 −3

−4 2 poly elements −4 2 poly elements

4 poly elements 4 poly elements
−5 −5 8 poly elements
8 poly elements
−6 −6
log 10 λ log 10 λ
−1 −0.5 0 0.5 1 1.5 2 −1 −0.5 0 0.5 1 1.5 2

Figure 13.12. Log-log plots of C I (λ) and C I I (λ) for Example of Figure 13.11 over range λ ∈ [0.1, 100].

5 Note that λ is a true physical parameter, whereas χ is discretization dependant because it involves the element length.

13–14
 §13.3 INTERPOLATION WITH HOMOGENEOUS ODE SOLUTIONS

Remark 13.2. To correlate the exact stiffness and consistent forces with those obtained with polynomial shape functions

it is illuminating to expand (13.33) as power series in χ. The rationale is that as the element size gets smaller, χ =
4
k F /(4E I ) goes to zero for fixed E I and k F . Mathematica gives the expansions

KeB = K B0 + χ 8 K B8 + χ 12 K B12 + . . . , KeF = K F0 + χ 4 K F4 + χ 8 K F8 + . . . , fe = f0 + χ 4 f4 + . . . , (13.37)

in which
 
25488 5352 23022 −5043
 5352 1136 5043 −1097  ,
E I 2 2
KeB0 = eqn (12.20) , KeB8 =
4365900 3 23022 5043 25488 −5352
−5043 −1097 2 −5352 1136 2
 
528960 113504 522090 −112631
(13.38)
 113504 24384 112631 −24273  , KeF0 = eqn (13.28),
EI 2 2
KeB12 =−
5959453500 3 522090 112631 528960 −113504
−112631 −24273 2 −113504 24384 2
k F 4 e 3k F 4 e q0 q0
KeF4 = − K B8 , KeF8 = − K , fe = [ 6 6 − ]T , fe4 = − [ 14 3 14 −3 ]T .
2E I 8E I B12 0 12 5040
Thus as χ → 0 we recover the stiffness matrices and force vector derived with polynomial shape functions, as can be
expected. Note that K B0 and K F0 decouple, which allows them to be coded as separate modules. On the other hand the
exact stiffnesses are coupled if χ > 0. The foregoing expansions indicate that exactness makes little difference if χ < 1.

13–15
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

§13.4. Equilibrium Theorems

One way to get high performance mechanical elements is to use equilibrium conditions whenever possible.
These lead to flexibility methods. Taking advantage of equilibrium is fairly easy in one space dimension. It is
more difficult in two and three, because it requires advanced variational methods that are beyond the scope of
this book. This section surveys theorems that provide the theoretical basis for flexibility methods. These are
applied to 1D element construction in §13.5.

§13.4.1. Self-Equilibrated Force System

First we establish a useful theorem that links displacement and force transformations. Consider a FEM dis-
cretized body such as that pictured in Figure 13.13(a). The generic potato intends to symbolize any discretized
material body: an element, an element assembly or a complete structure. Partition its degrees of freedom into
two types: r and s. The s freedoms (s stands for suppressed or supported) are associated with a minimal set of
supports that control rigid body motions or RBMs. The r freedoms (r is for released) collect the rest. In the
figure those freedoms are shown collected at invididual points Ps and Pr for visualization convenience. Node
forces, displacements and virtual displacements associated with those freedoms are partitioned accordingly.
Thus      
f us δus
f= s , u= , δu = . (13.39)
fr ur δur
The dimension n s of fs , us and δus is 1, 3 and 6 in one-, two- and three-dimensional space, respectively.
Figure 13.13(b) shows the force system {fs , fr } undergoing virtual displacements, which are exaggerated for
visibility.6 Consider now the “rigid + deformational” displacement decomposition u = Gus + d, in which
matrix G (of appropriate order) represents a rigid motion and d are deformational displacements. Evaluating
this at the r freedoms gives
ur = Gr us + dr , δur = Gr δus + δdr , (13.40)
The first decomposition in (13.40), being linear in the actual displacements, is only valid only in geometrically
linear analysis. That for virtual displacements is valid for a much broader class of problems.
If the supported freedom motion vanishes: us = 0, then ur = dr . Thus dr represents a relative displacement
of the unsupported freedoms with respect to the rigid motion Gus , and likewise for the virtual displacements.
Because a relative motion is necessarily associated with deformations, the alternative name deformational
displacements is justified.
The external virtual work is δW = δWs + δWr = δusT fs + δurT fr . If the force system in Figure 13.13(a) is in
self equilibrium and the virtual displacements are imparted by rigid motions δdr = 0 and δur = Gr δus , the
virtual work must vanish: δW = δusT fs + δusT GrT fr = δusT (fs + GrT fr ) = 0. Because the δus are arbitrary,
it follows that
fs + GrT fr = 0, fs = −GrT fr . (13.41)
These are the overall static equilibrium equations of a discrete mechanical system in self equilibrium. Some-
times it is useful to express the foregoing expressions in the complete-vector form
               
us I 0 δus I 0 f −GrT
u= = us + , δu = = δus + , f= s = fr . (13.42)
ur Gr dr δur Gr δdr fr I

Relations in (13.42) are said to be reciprocal.7

6 Under virtual displacements the forces are frozen for application of the Principle of Virtual Work.
7 If the model is geometrically nonlinear, the first form in (13.42) does not hold.

13–16
 §13.4 EQUILIBRIUM THEOREMS

(a) (b)

δus δur
fs Pr fr fs Pr fr
Ps Ps

Figure 13.13. Body to illustrate equilibrium theorems. Nodal freedoms classified into supported
(s) and released (r ), each lumped to a point to simplify diagram. (a) Self equilibrated node force
system. (b) Force system of (a) undergoing virtual displacements; grossly exaggerated for visibility.

Remark 13.3. The freedoms in us are virtual supports, chosen for convenience in flexibility derivations. They should not
be confused with actual or physical supports. For instance Civil Engineering structures tend to have redundant physical
supports, whereas aircraft or orbiting satellites have none.

§13.4.2. Handling Applied Forces

Consider now a generalization of the previous scenario. An externally applied load system of surface or body forces, not
necessarily in self equilibrium, acts on the body. For example, the surface tractions pictured in Figure 13.14(a). To bring
this under the framework of equilibrium analysis, a series of steps are required.
First, the force system is replaced by a single resultant q, as pictured in Figure 13.14(b).8 The point of application is Pq .
Equilibrium is restored by introducing node forces qr and qs at the appropriate freedoms. The overall equilibrium condition
is obtained by putting the system {q, qr , qs } through rigid-motion virtual displacements, as pictured in Figure 13.14(c).
Point Pq moves through δuq , and G evaluated at Pq is Gq . The virtual work is δW = δusT (qs + GrT qr + GqT q) = 0
whence
qs + GrT qr + GqT q = 0. (13.43)

If (13.43) is sufficient to determine qs and qr , the load system of Figure 13.14(a) can be effectively replaced by the nodal
forces −qs and −qr , as depicted in Figure 13.14(d). These are called the equivalent node forces. But in general (13.43)
is insufficient to fully determine qs and qr . The remaining equations to construct the equivalent forces must come from a
theorem that accounts for the internal energy, as discussed in §13.4.3.
Adding (13.41) and (13.43) gives the general overall equilibrium condition

fs + qs + GrT (fr + qr ) + GqT q = 0, (13.44)

which is applied in §13.4.4 to the recovery of supported freedoms.

Remark 13.4. The replacement of the applied force by a resultant is not strictly necessary, as it is always possible to write
out the virtual work by appropriately integrating distributed effects. The resultant is primarily useful as an instructional
tool, because matrix Gq is not position dependent.
Remark 13.5. Conditions (13.41) and (13.43), which were derived through the PVW, hold for general mechanical systems
under mild reversibility requirements [752, §231], including geometric nonlinearities. From now on we restrict attention to
systems linear in the actual displacements.

§13.4.3. Flexibility Equations

The first step in FEM equilibrium analysis is obtaining discrete flexibility equations. The stiffness equations
introduced in Chapter 2 relate forces to displacements. At the element level they are fe = Ke ue . By definition,
flexibility equations relate displacements to forces: ue = Fe fe , where Fe is the element flexibility matrix. So
the expectation is that the flexibility can be obtained as the inverse of the stiffness: Fe = (Ke )−1 . Right?

8 Although the figure shows a resultant point force, in general it may include a point moment that is not shown for simplicity.
See also Remark 13.3.

13–17
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

(a) (b)
qs qr
Ps Pr
Ps Pr

Pq
t(x) q
(c)
δus qs δur
(d)
qr −qs −q r
Ps Pr Ps Pr
δuq
Pq
q

Figure 13.14. Processing non-self-equilibrated applied loads with flexibility methods. (a) Body
under applied distributed load. (b) Substitution by resultant and self equilibration. (c) Deriving overall
equilibrium conditions through the PVW. (d) Replacing the applied loads by equivalent nodal forces.

Wrong. Recall that Ke for a disconnected free-free element is singular. Its ordinary inverse does not exist.
Expectations go up in smoke. The same difficulty holds for a superelement or complete structure.
To get a conventional flexibility matrix9 it is necessary to remove all rigid body motions in advance. This can
be done through the virtual supports introduced in §13.4.1. The support motions us are fixed, say us = 0.
Flexibility equations are sought between what is left of the kinematics. Dropping the element superscript for
brevity, for a linear problem one gets
Frr fr = dr . (13.45)
Note that ur does not appear: only the deformational or relative displacements. To recover ur it is necessary
to release the supports, but if that is naively done Frr ceases to exist. This difficulty is overcome in §13.4.4.
There is another key difference with stiffness methods. The DSM assembly procedure covered in Chapter 3
(and extended in Chapter 27 to general structures) does not translate into a similar technique for flexibility
methods. Although it is possible to assemble flexibilities of MoM elements, the technique is neither simple
nor elegant. And it becomes dauntingly complex when tried on continuum-based elements [233].
So one of the main uses of flexibility equations today is as a stepping stone on the way to derive element
stiffness equations, starting from (13.45). The procedural steps are explained in §13.4.4. But how should
(13.45) be derived? There are several methods but only one, based on the Total Complementary Potential
Energy (TCPE) principle of variational mechanics is described here.
To apply TCPE, the complementary energy ∗ of the body must be be expressed as a function of the nodal
forces fr . For fixed supports (us = 0) and a linear system, the functional can be expressed as

∗ (fr ) = U ∗ (fr ) − frT dr = 12 frT Frr fr + frT br − frT dr + ∗0 . (13.46)

Here U ∗ is the internal complementary energy, also called the stress energy by many authors, e.g., [321], br is a
term resulting from loading actions such as as thermal effects, body or surface forces, and ∗0 is independent of
fr . Calculation of U ∗ in 1D elements involves expressing the internal forces (axial force, shear forces, bending

9 In the FEM literature it is often called simply the flexibility. The reason is that for a long time it was believed that getting
a flexibility matrix required a supported structure. With the recent advent of the free-free flexibility (see Notes and
Bibliography) it becomes necessary to introduce a “deformational” or “conventional” qualifier.

13–18
 §13.4 EQUILIBRIUM THEOREMS

moments, torque, etc.) in terms of fr from statics. Application examples are given in the next section.10 The
TCPE principle states that ∗ is stationary with respect to variations in fr when kinematic compatibility is
satisfied:
∂∗
= Frr fr + br − dr = 0, whence dr = Frr fr + br . (13.47)
∂fr
By hypothesis the deformational flexibility Frr is nonsingular. Solving for fr gives the deformational stiffness
equations
−1
fr = Krr dr − qr , with Krr = Frr and qr = Krr br . (13.48)
The matrix Krr is the deformational stiffness matrix, whereas qr is the equivalent load vector.

§13.4.4. Rigid Motion Injection

Suppose that Frr and qr of (13.48) have been found, for example from the TPCE principle (13.47). The goal
is to arrive at the free-free stiffness equations, which are partitioned in accordance with (13.39) as
      
fs Kss Ksr us qs
= − , (13.49)
fr Kr s Krr ur qr
To justify the presence of Krr and qr here, set us = 0, whence ur = dr . Consequently the second equation
reduces to fr = Krr dr − qr , which matches (13.48). Inserting fs and fr into (13.44) yields
 
(Kss + GrT Kr s )us + (Ksr + GrT Krr )ur + qs + GrT qr + GqT q = 0, (13.50)
and replacing ur = Gr us + dr ,
     
Kss + GrT Kr s + Ksr Gr + GrT Krr Gr us + Ksr + GrT Krr dr + qs + GrT qr + GqT q = 0. (13.51)
Because us , dr and q can be arbitrarily varied, each bracket in (13.51) must vanish identically, giving
qs = −GrT qr − GqT q, Ksr = −GrT Krr , Kr s = Ksr
T
= −Krr Gr ,
(13.52)
Kss = −GrT Kr s − Ksr Gr − GrT Krr Gr = GrT Krr Gr + GrT Krr Gr − GrT Krr Gr = GrT Krr Gr .
Inserting these into (13.49) yields
      
fs GrT Krr Gr −GrT Krr us GrT qr − GqT q
= − . (13.53)
fr −Krr Gr Krr ur qr
This can be put in the more compact form
         
fs −GrT u −GrT Gq
= Krr [ −Gr I ] s − qr + q = TT Krr T u − TT qr + TqT q,
fr I ur I 0 (13.54)
with T = [ −Gr I ] and Tq = [ Gq 0 ] .
−1
The end result is that the free-free stiffness is TT Frr T. Alternatively (13.54) may be derived by plugging
dr = ur − Gr us into (13.48) and then into (13.44).
Remark 13.6. Let S be a n s × n s nonsingular matrix. The matrix R built by the prescription
 
S
R= (13.55)
GS
is called a rigid body motion matrix or simply RBM matrix. The columns of R represent nodal values of rigid motions,
hence the name. The scaling provided by S may be adjusted to make R simpler. The key property is T R = 0 and thus
K R = TT Krr T R = 0. Other properties are studied in [?].

10 For 2D and 3D elements the process is more delicate and demands techniques, such as hybrid variational principles, that
lie beyond the scope of this material.

13–19
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

§13.4.5. Applications
Stiffness Equilibrium Tests. If one injects ur = Gus and qr = 0 into (13.54) the result is fr = 0 and fs = 0.
That is, all node forces must vanish for arbitrary us . This test is useful at any level (element, superelement, full
structure) to verify that a directly generated K (that is, a K constructed independently of overall equilibrium)
is “clean” as regards rigid body modes.
Element Stiffness from Flexibility. Here Frr is constructed at the supported element level, inverted to get Krr
and rigid motions injected through (13.54). Applications to element construction are illustrated in §13.5.
Experimental Stiffness from Flexibility. In this case Frr is obtained through experimental measurements on
a supported structure or substructure.11 To insert this as a “user defined superelement” in a DSM code, it is
necessary to produce a stiffness matrix. This is done again by inversion and RBM injection.

§13.5. Flexibility Based Derivations

The equilibrium theorems of the foregoing section are applied to the flexibility derivation of several one-
dimensional elements.

§13.5.1. Timoshenko Plane Beam-Column

A beam-column member combines axial and bending effects. A 2-node, straight beam-column has three
DOFs at each node: the axial displacement, the transverse displacement and a rotation. If the cross section is
doubly symmetric, axial and bending effects are decoupled. A prismatic, plane element of this kind is shown
in Figure 13.15(a). End nodes are 1–2. The bending component is modeled as a Timoshenko beam. The
element is subjected to the six node forces shown, and to a uniformly distributed load q0 . To suppress rigid
motions node 1 is fixed as shown in Figure 13.15(b), making the beam a cantilever. Following the notation of
§13.4.1–§13.4.2,
           
u x1 u x2 f x1 f x2 0
1 0 0
us = , dr = ur =
u y1 , fs =
u y2 , fr = f y1 , q= , G(x) = 0 1 x ,
f y2 q
θ1 θ2 m1 m2 0
0 0 0
(13.56)
Further, Gr = G( ) and Gq = G( /2). The internal forces are the axial force F(x), the transverse shear
V (x) and the bending moment M(x). These are directly obtained from statics by doing a free-body diagram
at distance x from the left end as illustrated in Figure 13.15(c). With the positive convention as shown we get

N (x) = − f x2 , V (x) = − f y2 − q0 ( − x), M(x) = m 2 + f y2 ( − x) + 12 q0 ( − x)2 . (13.57)

Useful check: d M/d x = V . Assuming a doubly symmetric section so that N and M are decoupled, the
element TCPE functional is
  
∗ N2 M2 V2
 = 1
2
+ + d x − frT dr = 12 frT Frr fr + frT br − frT dr + ∗0 ,
0 EA EI G As
   
E A 0 0 0 (13.58)
∂ 2 ∗  
3 (4 + ) 2  , b = q  (4 + )  .
3
in which Frr = =
 0 0 12E I 
∂fr ∂fr 24E I 2E I  r

2
2
0 2E I EI 6E I

11 The classical static tests on an airplane wing are performed by applying transverse forces and torques to the wing tip
with the airplane safely on the ground. These experimental influence coefficients can be used for model validation.

13–20
 §13.5 FLEXIBILITY BASED DERIVATIONS

y,v (b) q0 fy2

;;
E,I,G,As constant m2 fx2
(a) fy2
fy1 q0 1 2
m1 m2 f
x2
x 2
fx1 1 +V(x) +M(x) q0 fy2
(c) m2 f
x2
+F(x) 2
x
Figure 13.15. Flexibility derivation of Timoshenko plane beam-column stiffness: (a) element and node
forces, (b) removal of RBMs by fixing left node, (c) FBD that gives internal forces at varying x.

Term ∗0 is inconsequential, since it disappears on differentiation.

Applying the TCPE principle yields Frr fr = br − dr . This is inverted to produce the deformational stiffness
relation fr = Krr dr + qr , in which
 EA 
0 0  
 0  0
12E I − 2 6E I
−1
Krr = Frr =
 3 (1 + )

(1 + )  , qr = q0 Krr br =  q0 /2  . (13.59)
E I (4 + ) −q0 2 /12
0 − 2 6E I
(1 + ) (1 + )
To use (13.54) the following transformation matrices are required:
 −1 0 0
 1 0 0

 0 −1 0  0 1 0  
   0  
− −1  0 1
0
−GrT /2
= , = ,
Gq
TT =  1 Tq = q= q0 . (13.60)
0 0 0  0 0 0 0
    0
0 1 0 0 0 0
0 0 1 0 0 0
Injecting the rigid body modes from (13.54), Ke = TT Krr T and fe = TT qr , yields
 1 0 −10 0 0
 0
 0 0 0 0 0

 0 0 0 0 0 0 0 12 6 0 −12 6 
EA 
 0 0 0 0 0 0+ E I 0
 6 2 (4 + ) 0 −6 2 (2 − ) 
,
K =
e

 −1 0 1 0 0 0  3 
 (1 + )  0 0 0 0 0 0 

0 0 0 0 0 0 0 −12 −6 0 12 −6
0 0 0 0 0 0 0 6 2 (2 − ) 0 −6 2 (4 + )
f = q0 [ 0 1/2 /12 0
e
1/2 − /12 ] .
T

(13.61)
The bending component is the same stiffness found previously in §13.2.6; compare with (13.18). The node
force vector is the same as the consistent one constructed in an Exercise. A useful verification technique is to
support the beam element at end 2 and recompute Ke and fe . This should reproduce (13.61).
All of the foregoing computations were carried out by the Mathematica script shown in Figure 13.16.

§13.5.2. Plane Circular Arch in Local System

In this and next subsection, the flexibility method is used to construct the stiffness matrix of a curved, prismatic,
plane beam-column element with circular profile, pictured in Figure 13.17(a). The local system {x, y} is defined

13–21
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

ClearAll[Le,EI,GAs,Φ,q0,fx2,fy2,m2]; GAs=12*EI/(Φ*Le^2);
F=fx2; V=-fy2-q0*(Le-x); M=m2+fy2*(Le-x)+(1/2)*q0*(Le-x)^2;
Print["check dM/dx=V: ",Simplify[D[M,x]-V]];
Ucd=F^2/(2*EA)+M^2/(2*EI)+V^2/(2*GAs);
Uc=Simplify[Integrate[Ucd,{x,0,Le}]]; Print["Uc=",Uc];
u2=D[Uc,fx2]; v2=D[Uc,fy2]; θ2=D[Uc,m2];
Frr={{ D[u2,fx2], D[u2,fy2], D[u2,m2]},
{ D[v2,fx2], D[v2,fy2], D[v2,m2]},
{ D[θ2,fx2], D[θ2,fy2], D[θ2,m2]}};
br={D[Uc,fx2],D[Uc,fy2],D[Uc,m2]}/.{fx2->0,fy2->0,m2->0}; Print["br=",br];
Frr=Simplify[Frr]; Print["Frr=",Frr//MatrixForm];
Krr=Simplify[Inverse[Frr]]; Print["Krr=",Krr//MatrixForm];
qr=Simplify[-Krr.br]; Print["qr=",qr];
TT={{-1,0,0},{0,-1,0},{0,-Le,-1},{1,0,0},{0,1,0},{0,0,1}};
T=Transpose[TT]; Simplify[Ke=TT.Krr.T]; Print["Ke=",Ke//MatrixForm];
GrT={{1,0,0},{0,1,0},{0,Le,1}}; Gr=Transpose[GrT];
GqT={{1,0,0},{0,1,0},{0,Le/2,1}}; Gq=Transpose[GqT];
Print["Gr=",Gr//MatrixForm," Gq=",Gq//MatrixForm];
qv={0,q0*Le,0}; Print["qs=",Simplify[-GrT.qr-GqT.qv]];

Figure 13.16. Script to derive the stiffness matrix and consistent load vector of the prismatic,
plane Timoshenko beam element of Figure 13.15 by flexibility methods.

y, v
(a) E,I,G,As constant (b) (c) +M(ψ)
+V(ψ)
m1 fy1 fy2 fy2 fy2
+F(ψ)
fx2 fx2 fx2
fx1 1 ψ
1 x, u |R| 2 2 2
|2φ| m2 m2 m2

Figure 13.17. Flexibility derivation of plane circular arch element: (a) element and node forces,
(b) removal of RBMs by fixing left node, (c) free body diagram of varying cross section.
O O
+ψ +ψ
y C +ds y +2φ C +ds −2φ
+R −R
x +H x −H
1 2 1 −H 2 2 x 1 2 +H x 1
−2φ −R +ds +R +2φ y +ds y
C C
+ψ +ψ
O O
2S 2S 2S 2S

Figure 13.18. Sign conventions in derivation of circular arch element.

as shown there: x is a “chord axis” that passes through end nodes 1–2, and goes from 1 to 2. Axis y is placed
at +90◦ from x. No load acts between nodes.
In a curved plane element of this nature, axial extension and bending are intrinsically coupled. Thus consid-
eration of three freedoms per node is mandatory. These are the translations along x and y, and the rotation θ
about z. This element can be applied to the analysis of plane arches and ring stiffeners (as in airplane fuselages

13–22
 §13.5 FLEXIBILITY BASED DERIVATIONS

and submarine pressure hulls). If the arch curvature varies along the member, it should be subdivided into
sufficiently small elements over each of which the radius magnitude is sensibly constant.
Care must be taken as regards sign conventions to ensure correct results when the arch convexity and node
numbering changes. Various cases are pictured in Figure 13.18. The conventions are as follows:
(1) The local node numbers define a positive arclength traversal along the element midline as going from
1 to 2. The curved length (not shown in figure) and the (chord) spanlength 2S are always positive.
(2) The arch rise H , the angular span 2φ and the arch radius R are signed quantities as illustrated in
Figure 13.18. The rise is the distance from chord midpoint to arch crown C: it has the sign of its
projection on y. The angular span 2φ is that subtended by the arch on moving from 1 to 2: it is positive
if CCW. Finally, the radius R has the sign of φ so = 2Rφ is always positive.
Some signed geometric relations:

S = R sin φ, H = R(cos φ − 1) (13.62)

The location of an arch section is defined by the “tilt angle” ψ measured from the circle center-to-crown
symmetry line OC, positive CCW. See Figure 13.18. The differential arclength ds = R dψ always points in
the positive traversal sense 1 → 2.
The rigid motions are removed by fixing the left end as shown in in Figure 13.17(b). The internal forces F(ψ),
V (ψ) and M(ψ) at an arbitrary cross section are obtained from the FBD of Figure 13.17(c) to be

F = f x2 cos ψ + f y2 sin ψ, V = f x2 sin ψ − f y2 cos ψ, M = m 2 + f x2 R(cos φ− cos ψ)+ f y2 R(sin ψ− sin φ).
(13.63)
For typical straight beam-column members there are only two practically useful models: Bernoulli-Euler (BE)
and Timoshenko. For curved members there are many more. These range from simple corrections to BE
through theory-of-elasticity-based models. The model selected here is one of intermediate complexity. It is
defined by the internal complementary energy functional
    φ  
∗ (F − M/R)2 M2 (F − M/R)2 M2
U = + ds = + R dψ. (13.64)
0 2E A 2E I −φ 2E A 2E I

The assumptions enbodies in this formula are: (1) the shear energy density V 2 /(2G As ) is neglected; (2) the
cross section area A and moment of inertia I are unchanged with √ respect of those of the straight member.
These assumptions are reasonable if |R| > 10 r , where r = + I /A is the radius of gyration of the cross
section. Further corrections are treated in Exercises. To simplify the ensuing formulas it is convenient to take

EI 4E I 2r I
EA = = , or = with r 2 = . (13.65)
 2 R2φ2  2 2 A

This defines  as a dimensionless geometric parameter. Note that this is not an intrinsic measure of arch
slenderness, because it involves the element length. (In that respect it is similar to of the Timoshenko
beam element, defined by (13.17).) The necessary calculations are carried out by the Mathematica script
of Figure 13.21, which has been pared down to essentials to save space. The deformational flexibility and
stiffness computed are
   
F11 F12 F13 K 11 K 12 K 13
−1
Frr = F22 F23 , Krr = Frr = K 22 K 23 ,
symm F33 symm K 33

13–23
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

in which (first expression is exact value, second a Taylor series expansion in φ)

3
3

F11 = 2φ(2 + φ 
2 2
) + 2φ(1 + φ 
2 2
) cos 2φ − 3 sin 2φ = 15 2 + φ 2 (2 − 15 2 ) + O(φ 4 )
16E I φ 3 60E I
sin φ
3
φ
3

F12 = sin φ − φ(1 + φ 2  2 ) cos φ = 1 − 3 2 + O(φ 3 ),

4E I φ 3 12E I
2
3 φ

F13 = sin φ − φ(1 + φ 

2 2
) cos φ = 1 − 3 2
+ O(φ 3 ),
2E I φ 2 6E I
3
3

F22 = 2φ(2 + φ 
2 2
) − 2φ(1 + φ 
2 2
) cos 2φ − sin 2φ = 20 + 3φ 2 (5 2 − 2) + O(φ 4 ),
16E I φ 3 60E I
sin φ
2
2

F23 = (1 + φ 2  2 ) = 6 + φ 2 (6 2 − 1) + O(φ 4 ), F33 = (1 + φ 2  2 ).
2E I φ 12E I EI
(13.66)
Introduce d1 = φ(1 + φ 2  2 )(φ + sin φ cos φ) − 2 sin2 φ and d2 = φ − sin φ cos φ. Then

4E I φ 4 EI

K 11 = (1 + φ 2  2 ) = 45 2 + φ 2 (15 2 + 45 4 − 1) + O(φ 4 ), K 12 = 0,

d1
3 45 3
2
6E I φ
4E I φ
K 13 = φ(1 + φ 
2 2
) cos φ − sin φ = 2 2 (3 2 − 1) + O(φ 3 ),
2 d1 
8E I φ 3 12E I sin φ EI
K 22 = = (5 + φ 2 ) + O(φ 4 ), K 23 = − K 22 = − 3 (30 + φ 2 ) + O(φ 4 ),
d2
3 5 3 2φ 5
EIφ

K 33 = 8φ 2 (3 + 2φ 2  2 ) − 9 + 16 cos 2φ − 7 cos 4φ − 8φ sin 2φ 2 + (1 + φ 2  2 ) cos 2φ

8 d1 d2
EI

= 180 2
+ φ 2
(5 − 48 2
) + O(φ 4 ),
45  2
(13.67)
The constraint K 23 = −K 22 sin φ/(2φ) must be verified by any arch stiffness, regardless of the TCPE form
used. The necessary transformation matrix to inject the rigid body modes
   
−1 0 0 1 0 0 −1 0 0 1 0 0
T = [ −G I ] = 0 −1 −2R sin φ 0 1 0 = 0 −1 − sin φ/φ 0 1 0 . (13.68)
0 0 −1 0 0 1 0 0 −1 0 0 1
Applying the congruent transformation gives the free-free stiffness
 K 0 K 13 −K 11 0 −K 13

11
 0 K 22 −K 23 0 −K 22 −K 23 
 K 13 −K 23 K 33 −K 13 K 36 
Ke = TT Krr T =  
K 23
 −K 11 (13.69)
0 −K 13 K 11 0 K 13 
 
0 −K 22 K 23 0 K 22 K 23
−K 13 −K 23 K 36 K 13 K 23 K 33
The only new entry not in (13.67) is
EI
K 36 = −K 33 − K 23 sin φ/φ = (90 2 + φ 2 (12ψ 2 − 5) + O(φ 4 )
45
As a check, if φ → 0 the entries reduce to that of the beam-column modeled with Bernoulli-Euler. For
example K 11 → 4E I  2 / 3 = E A/ , K 22 → 12E I / 3 , K 33 → 4E I / , K 36 → 2E I / , etc.

13–24
 §13.5 FLEXIBILITY BASED DERIVATIONS

ClearAll[φ,ψ,Ψ,Γ,R,F,M,V,Le,EA,EI];
EA=4*EI/(Ψ^2*Le^2);
V=-fy2*Cos[ψ]+fx2*Sin[ψ]; F=fx2*Cos[ψ]+fy2*Sin[ψ];
M=m2+fx2*R*(Cos[ψ]-Cos[φ])+fy2*R*(Sin[ψ]+Sin[φ]);
Ucd=(F-M/R)^2/(2*EA)+ M^2/(2*EI);
Uc=Simplify[Integrate[Ucd*R,{ψ,-φ,φ}]]; Print["Uc=",Uc];
u2=D[Uc,fx2]; v2=D[Uc,fy2]; θ2=D[Uc,m2];
Frr=Simplify[{{ D[u2,fx2], D[u2,fy2], D[u2,m2]},
{ D[v2,fx2], D[v2,fy2], D[v2,m2]},
{ D[θ2,fx2], D[θ2,fy2], D[θ2,m2]}}];
Frr=FullSimplify[Frr/.{R->Le/(2*φ)}]; Print["Frr=",Frr//MatrixForm];
Krr=FullSimplify[Inverse[Frr]]; Print["Krr=",Krr//MatrixForm];
TT={{-1,0,0},{0,-1,0},{0,-2*R*Sin[φ],-1},{1,0,0},{0,1,0},{0,0, 1}};
TT=TT/.{R->Le/(2*φ)}; T=Transpose[TT];
Ke=Simplify[TT.Krr.T]; Print["Ke=",Ke//MatrixForm];

Figure 13.19. Script to produce circular arch element stiffness in local coordinates.

_
(a) x (b) (c)
2 (x2 ,y2) 2 2
b
3 (x3 ,y3 ) 3 R 3
H
_
y ϕ 2φ a 2S
y
1 (x 1 ,y1) 1 1
x
Figure 13.20. Plane circular arch element in global coordinates: (a) geometric id (b) intrinsic
geometry recovery.

Figure 13.21. Module to produce plane circular arch element stiffness in global coordinates.
(To be done)

13–25
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

§13.5.3. Plane Circular Arch in Global System

To use the circular arch element in a 2D finite element program it is necessary to specify its geometry in the
{x, y} plane and then to transform the stiffness (13.69) to global coordinates. The first requirement can be
handled by providing the coordinates of three nodes: {xi , yi }, i = 1, 2, 3. Node 3 (see Figure) is a geometric
node that serves to define the element mean curvature but has no associated freedoms.
(Section to be completed).

13–26
 §13.6 *ACCURACY ANALYSIS

q(x) vj = 1
Two-element Trial functions
i j k θj = 1 at node j
xi = xj − EI = const
patch ijk
xj
xk = xj + i (1) j (2) k
L

Figure 13.22. Repeating beam lattice for accuracy analysis.

§13.6. *Accuracy Analysis

This section presents the accuracy analysis of a repeating lattice of beam elements, analogous to that done for
the bar element in §12.5. The analysis uses the method of modified differential equations (MoDE) mentioned
in the Notes and Bibliography of Chapter 12. It is performed using a repeated differentiations scheme similar
to that used in solving Exercise 12.8. Only the case of the Bernoulli-Euler model is worked out in detail.

§13.6.1. *Accuracy of Bernoulli-Euler Beam Element

Consider a lattice of repeating two-node, prismatic, plane Bernoulli-Euler beam elements of rigidity E I and
length , as illustrated in Figure 13.22. The system is subject to an arbitrary lateral load q(x), which is assumed
infinitely differentiable in x. From the lattice extract a patch of two elements: (1) and (2), connecting nodes
i– j and j–k, respectively, as shown in Figure 13.22.
The FEM patch equations at node j are
v 
i
 θi   
E I  −12 −6 24 0 −12 6   
 vj  = f j (13.70)
3 6 2 0 8 −6 2  θ j 
2 2 2
mj
 
vk
θk

Expand v(x), θ(x) and q(x) in Taylor series about x = x j , truncating at n +1 terms for v and θ, and m +1 terms
for q. Using ψ = (x − x j )/ , the series are v(x) = v j + ψ v j + (ψ 2 2 /2!)v j + . . . + (ψ n n /n!)v [n]
j , θ(x) =
   
θ j +ψ θ j +(ψ /2!)θ j +. . .+(ψ /n!)θ j , and q(x) = q j +ψ q j +(ψ /2!)q j +. . .+(ψ /m!)q [m]
2 2 n n [n] 2 2 m m
j .
Here v j , etc., is an abbreviation for d v(x j )/d x . Evaluate the v(x) and θ (x) series at i and k by setting
[n] n n 12

ψ = ±1, and insert in (13.70). Use the q(x) series evaluated over elements (1) and (2), to compute the
consistent forces f j and m j as
 1  1   1  1 

fj = N3(1) q (1) dξ (1)
+ N1(2) q (2) dξ (2)
, mj = N4(1) q (1) dξ (1)
+ N2(2) q (2) dξ (2)
.
2 −1 −1 2 −1 −1
(13.71)
Here N3(1) = 14 (1 + ξ (1) )2 (2 − ξ (1) ), N4(1) = − 8 (1 + ξ (1) )2 (1 − ξ (1) ) N1(2) = 14 (1 − ξ (2) )2 (2 + ξ (2) ), and
N2(2) = 8 (1 − ξ (2) )2 (1 + ξ (2) ) are the Hermitian shape functions components of the j node trial function,
whereas q (1) = q(ψ (1) ), ψ (1) = − 12 (1 − ξ (1) ) and q (2) = q(ψ (2) ), ψ (2) = 12 (1 + ξ (2) ) denote the lateral loads.

12 Brackets are used instead of parentheses to avoid confusion with element superscripts. If derivatives are indexed by
primes or roman numerals the brackets are omitted.

13–27
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

To show the resulting system in compact matrix form it is convenient to collect the derivatives at node j into
vectors:

v j = [ v j v j v j . . . v [n]
j ] ,
T
θ j = [ θ j θ j θ j . . . θ [n]
j ] ,
T
q j = [ q j q j q j . . . q [n]
j ] .
T
(13.72)

The resulting differential system can be compactly written

    
Svv Svθ vj Pv
= qj. (13.73)
Sθv Sθθ θj Pθ

Here Svv , Svθ , Sθv and Sθθ are triangular Toeplitz matrices of order (n+1) × (n+1) whereas Pv and Pθ are
generally rectangular matrices of order (n+1) × (m+1). Here is the expression of these matrices for n = 8,
m = 4:
 −12  0 3 5
0 
− 0 − −
3 5 12
0 0
0 30
0 1680
0 2 0 10 0 420
3 5
0 0 0 −12
0 − 0 −
3
0  0 0 12
0 2 0 10 0 
 30  
 420
0 0 0 0 −12
0 − 0 30  − 3  0 0 0 12
0 2 0 3
0 
  10 
0 0 0 0 0 −12 0 − 0  0 0 0 0 12
0 2 3 
0 10
   
Svv = EI 0 0 0 0 0 0 −12
0 −  , Svθ = E I  0 0 0 0 0 12 0 2 0 
   
0 0 0 0 0 0 0 −12
0  0 0 0 0 0 0 12 0 2 
  
0 0 0 0 0 0 0 0 −12  0 0 0 0 0 0 0 12
0 
  
12

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
0  12 
0  0 6 0
−12
3 5 7
0 − − 5
3

0 −2 10
0 420
0 2 180
0 10080
0 0 −12
0 −2 0 − 3
0  − 5 0 12
0 2 0 6
3
0 5
0 
 10
 420  180 
0 −12
0 −2 − 3  0 0 12
0 2 0 3
0 5 

0 0
0 0
  6 180 
0 −12
10
 0 0 6
3 
0 −
3 12
 0 0 0 0 −2 10   0 0
0 2 0 
Sθv = E I  0

−12  , S = E I  3 
 0 0 0 0 0 −2 0  θθ 0 0 0 12
0 0 2 0 
0

−12
0 −2 
 6

 0 0 0 0 0  0 0 0 0 0 12 0 2 0 
0 −12   
 0 0 0 0 0 0
0
 0 0 0 0 0 0 12

0 2 
0 0 0 0 0 0 0 0 −12 0 0 0 0 0 0 0 12 0 

0 0 0 0 0 0 0 0 0 12
0 0 0 0 0 0 0 0
(13.74)
 0 3 5   3 5 
15
0 560
0 15
0 315
0
0 3
0  0 0 3
0 5 

0 15   15 315 
0 0 3  0 3 
 0
  0 0 15
0 
0 0
15
 0 3 
0 0  0 0 0 
Pv = 
0 0 0 0
,
 Pθ =  0 15
 (13.75)
0 0   0 0 0 0 
 0 0 0  0 0 0 0 0 
0 0   
 0 0 0  0 0 0 0 0 
0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0
Elimination of θ j gives the condensed system

Sv j = Pq j , with S = Svv − Svθ S−1

θθ Sθv , P = Pv − Svθ S−1
θv Pθ . (13.76)

13–28
 §13.6 *ACCURACY ANALYSIS
0 0 0 0 0 0 0 − 5   0 0 0 − 5 
720 720
0 0 0 0 0 0 0 0  0 0 0 
0
0 0 0 0 0 0 0 0  0 0 0 
0
   
0 0 0 0 0 0 0 0  0 0 0 0 

   
S = Svv − Svθ S−1 S
θθ θv = E I 0 0 0 0 0 0 0 0 , P = Pv − Svθ S−1 P
θv θ = 0 0 0 
0
0 0  0 0 0 
 0 0 0 0 0 0 0   0 0 
0 0 0 0 0 0 0 0 0  0 0 0 0 
0
   
0 0 0 0 0 0 0 0 0 0 0 0 00
0 0 0 0 0 0 0 0 0 0 0 0 00
(13.77)
4 viii
The nontrivial differential relations13 given by S v j = P q j are E I (v iv j − v j /720) = q j − q j /720,
4 iv

v  vi  vii  viii

E I v j = q j , E I v j = q j , E I v j = q j , and E I v j = q j . The first one is a truncation of the infinite order
iv

MoDE. Elimination of all v j derivatives but v iv j yields

E I v iv
j = qj, (13.78)

exactly. That this is not a fluke can be confirmed by increasing n and while taking m = n − 4. The first 4
columns and last 4 rows of S, which are always zero, are removed to get Ŝ. The last 4 rows of P, which are
also zero, are removed to get P̂. With m = n−4 both matrices are of order n−3 × n−3. Then Ŝ = P̂ for any
n, which leads immediately to (13.78). This was confirmed with Mathematica running n up to 24.
The foregoing analysis shows that the BE cubic element is nodally exact for any smooth load over a repeating
lattice if consistent load computation is used. Exercise 13.18 verifies that the property is lost if elements are
not identical. Numerical experiments confirm these conclusions. The Laplace transform method only works
part way: it gives a different infinite order MoDE but recovers (13.78) as n → ∞.
These accuracy properties are not widely known. If all beam elements are prismatic and subjected only to
point loads at nodes, overall exactness follows from a theorem by Tong [746], which is not surprising since
the exact solution is contained in the FEM approximation. For general distributed loads the widespread belief
is that the cubic element incurs O( 4 ) errors. The first study of this nature by Waltz et. al. [780] gave the
modified differential equation (MoDE) for a uniform load q as

4 viii q
v iv − v + ... = . (13.79)
720 EI
The above terms are correct. In fact a more complete expression, obtained in this study, is
 
4 viii 6 x 7 8 4 iv 6 vi 7 8
EI v −iv
v + v − v xii + . . . =q−
q + q − q viii + . . .
720 3024 259200 720 3024 259200
(13.80)
But the conclusion that “the principal error term” is of order 4 [780,p. 1009] is incorrect. The misinterpretation
is due to (13.80) being an ODE of infinite order. Truncation is fine if followed by elimination of higher
derivatives. If this is done, the finite order MoDE (13.78) emerges regardless of where (13.80) is truncated; an
obvious clue being the repetition of coefficients in both sides. The moral is that conclusions based on infinite
order ODEs should be viewed with caution, unless corroborated by independent means.

13 Derivatives of order 4 and higher are indicated by Roman numeral superscripts instead of primes.

13–29
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

§13.6.2. *Accuracy of Timoshenko Beam Element

Following the same procedure it can be shown that the infinite order MoDE for a Timoshenko beam element
repeating lattice with the stiffness (13.18) and consistent node forces is
 
2 vi 4 (1 + 5 − 5 2 ) viii 6 (20 + 7 − 70 2 + 35 3 ) x
EI v iv
j + v − vj + vj + . . .
12 j 720 60480
4 (1 + 5 ) iv 6 (20 + 14 − 35 2 ) vi
= qj − qj + qj ) + . . . ,
in which = 12E I /(G As 2 ).
720 60480
(13.81)
This reduces to (13.80) if = 0. Elimination of higher order derivatives gives the finite order MoDE (aka
FOMoDE)
2  E I 
E I v iv
j = qj − q = qj − q . (13.82)
12 j G As j
which repeats for any n > 8. This happens to be the exact governing differential equation for a statically
loaded Timoshenko beam [263, p. 23]. Consequently the Timoshenko beam is nodally exact under the same
conditions previously stated for the Bernoulii-Euler model.
Notes and Bibliography
The material in this Chapter interwines the very old and the very new. Before energy methods came to the
FEM forefront by 1960 (see historical sketch in §1.7), ordinary differential equations (ODE) and flexibility
methods were essential part of the toolbox repertoire of the professional structural engineer. Traces of that
dominance may be found in the books by Przemieniecki [596], Pestel and Leckie [572], and the survey by
Gallagher [288]. Energy derivations were popularized by Archer [34,35], Martin [467], and Melosh [484,485].
For one-dimensional elements, however, results are often identical. This can provide a valuable crosscheck.
The Legendre interpolation (13.2) was introduced in [?] [236] to study optimal mass-stiffness combinations
for beam elements in the context of finite element templates [228]. The diagonal covariance matrices Qn
given in (13.4) play a key role in model customization. The hinged element stiffness (13.14) is rederived in
the Advanced FEM Lecture Notes [253]. using a mixed variational principle. The separation of uncoupled
rigidity effects in stiffness forms such as (13.8) and (13.9) is suggested by template theory [243].
The Timoshenko beam model was originally proposed in [735]. Timoshenko cleverly packaged the model
with miscellaneous ingredients introduced earlier by Bresse and Hencky. It has become important as a tool for
transient response and control simulations because its dynamic form is strictly hyperbolic.14 The Timoshenko
beam element stiffness (13.18) first appeared in [758] in the guise of a spar element for use in aircraft structures;
the end node freedoms of that element differring from the classical set used here. The particular form (13.18)
is derived in Section 5.6 of [596] using ODEs. This beam model pertains to the class of “C 0 elements”
that have been extensively studied in the FEM literature after 1968. The book of Hughes [385] provides a
comprehensive treatment of such methods for beams and plates.
The classical work on beams on elastic foundations is by Hetenyi [357]. Useful solutions are tabulated in
Roark-Young’s handbook [642].
That the use of homogeneous solutions of governing differential equations yields nodally-exact stiffness
equations was first proven generally by Tong [746] in a Galerkin context. This derivation procedure, however,
was rarely used after the 1960s. Two obstacles: (1) it is largely restricted to either one dimensional elements,
or to problems with special symmetries that can be modeled with ODEs;15 (2) rapidly increasing solution

14 The Bernoulli-Euler beam dynamic model is parabolic and thus exhibits an infinite transverse wave speed. Such a model
is unsuitable for wave propagation problems.
15 For example, symmetrically loaded circular plates or shells of revolution.

13–30
 §13. References

complexity in complicated problems discouraged hand derivations. Whereas the first limitation still holds, the
increasing availability of CAS allows timely consideration of more difficult problems. The construction of the
exact beam-on-Winkler-foundation element in §13.3 offers a case in point. Using Mathematica the complete
derivation, checking and fully-symbolic testing took about 6 hours, whereas a hand derivation, coding and
numerical testing would likely take weeks or months. The main application of exact elements appears to be
a priori error estimation: how many simpler elements are needed to do the job of an exact one?
The construction of stiffness matrices from flexibility information was historically one of the first techniques
by which stiffness equations of MoM members were derived. The rigid body injection method of §13.4.4
largely follows Section 6.6 of [596]. The presentation of discrete-system equilibrium theorems in §13.4.2
includes a new ingredient missing from previous work: handling non-self-equilibrated loading systems. This
extension removes the 40-year-old objection that flexibility methods (or more generally, schemes based on the
TCPE principle) are unable to produce equivalent or consistent node forces.
The use of equilibrium methods for multidimensional finite elements was pioneered by Fraeijs de Veubeke in
the 1960s and early 1970s. His obsession with solution bounding got these methods seriously stuck, however,
because of difficulties in interelement connections that maintain system-level equilibrium, as well as avoidance
of spurious modes. More practical extensions lead to the so-called Trefftz and equilibrium hybrid methods.
These are presently the topic of active research16 but require advanced variational techniques beyond the scope
of this book. Another recent advance is the discovery of the free-free flexibility as the true dual of the free-free
stiffness [232,240,233]. This extension relies heavily on projection operators.
That Hermitian BE element models are nodally exact if consistent loads are used is stated in [397, Sec. 8.3] as
the “beam theorem.” Despite the name, no proof is given; only anecdotal evidence — it was likely discovered
by numerical experimentation. A proof based on Fourier series appears in [263].
As the study in §13.6 illustrates, modified equation methods in boundary value problems17 are delicate and
should be used with care. Their intricacies baffled Strang and Fix who, upon doing Fourier analysis of a cubic
beam element, incorrectly stated [698, p. 171] that only one of the discrete equations — that for v(x) — is
consistent “and the others are completely inconsistent.” The alternative is the variational approach to error
analysis. Although more robust and forgiving, predictions are often so conservative as to be of little value.

References
Referenced items have been moved to Appendix R.

16 Along with discontinuous Galerkin methods, a reinvention of Fraeijs de Veubeke’s weakly diffusive models.
17 They are more forgiving in initial value problems.

13–31
Chapter 13: ADVANCED ONE-DIMENSIONAL ELEMENTS

Homework Exercises for Chapter 13

Advanced One-Dimensional Elements

EXERCISE 13.1 [A:15] Evaluate the strain and stress fields associated with the Timoshenko beam displace-
ment field (13.16).

EXERCISE 13.2 [A/C:20] Find the consistent node forces for a Timoshenko beam element under a uniform
transverse load q. Hint: find fc for the Legendre interpolation, then premultiply by HTS .

EXERCISE 13.3 [A/C:20] Construct the stiffness matrix a Timoshenko plane beam element with a hinge at
the center. Hint: set R Bs = 0, R Ba = E I and R S = 12E I /( 2 ) in (13.13).

EXERCISE 13.4 [A/C:15=5+10] Consider a cantilever beam of constant bending rigidity E I , constant shear
rigidity G As and length L, which is modeled with one Timoshenko element. The beam is end loaded by a
transverse force P. The support conditions are v1 = θ1 = 0.
(a) Find the end displacement v2 and end rotation θ2 in terms of P, E, G, I , As and L. Compare with the
analytical values P L 3 /(3E I ) + P L/(G As ) and P L 2 /(2E I ), respectively.
(b) Why does the finite element model provides the exact answer with one element?

EXERCISE 13.5 [A:25] (Requires math ability). Discuss what happens in (13.18) if → ∞. Is the result
useful for a shear-only “spar” element? Hint: eliminate θ1 and θ2 by a master-slave MFC.

EXERCISE 13.6 [A:10] For a given number of elements N e of length = 2L/N e , relate χ and λ in Example
13.2.

EXERCISE 13.7 [A/C:40] (research paper level). Derive an exact Timoshenko-beam-on-Winkler-foundation

equation method.element using the differential equation method.

EXERCISE 13.8 [C:30] Write Mathematica code to verify the nodal exactness conclusion of §13.6.1 using
the repeated differentiation approach.

EXERCISE 13.9 [C:30] As above, but using the Laplace transform. Show that this only does half the job.

EXERCISE 13.10 [A/C:35] Find the general symbolic expression of the terms in

EXERCISE 13.11 [A/C:40] (research paper level) Analyze nodal accuracy if the length of the beam elements
(1 ± α) , where 0 ≤ α ≤ 12 .

EXERCISE 13.12 [A/C:35] Using Mathematica, verify the results (13.81) and (13.82) in §13.6.2.

13–32
 14
The Plane Stress
Problem

14–1
Chapter 14: THE PLANE STRESS PROBLEM

TABLE OF CONTENTS

Page
§14.1 Introduction . . . . . . . . . . . . . . . . . . . . . 14–3
§14.2 Plate in Plane Stress . . . . . . . . . . . . . . . . . . 14–3
§14.2.1 Behavioral Assumptions . . . . . . . . . . . . . . 14–3
§14.2.2 Mathematical Model . . . . . . . . . . . . . . 14–4
§14.2.3 Problem Data . . . . . . . . . . . . . . . . . 14–5
§14.2.4 Problem Unknowns . . . . . . . . . . . . . . . 14–5
§14.3 Plane Stress Governing Equations . . . . . . . . . . . . . 14–6
§14.3.1 Governing Equations . . . . . . . . . . . . . . 14–6
§14.3.2 Boundary Conditions . . . . . . . . . . . . . . . 14–7
§14.3.3 Weak versus Strong Form . . . . . . . . . . . . . 14–8
§14.3.4 Total Potential Energy . . . . . . . . . . . . . . 14–9
§14.4 Finite Element Equations . . . . . . . . . . . . . . . . 14–10
§14.4.1 Displacement Interpolation . . . . . . . . . . . . . 14–10
§14.4.2 Element Energy . . . . . . . . . . . . . . . . 14–11
§14.4.3 Element Stiffness Equations . . . . . . . . . . . . 14–12
§14. Notes and Bibliography . . . . . . . . . . . . . . . . . 14–12
§14. References . . . . . . . . . . . . . . . . . . . . . . 14–12
§14. Exercises . . . . . . . . . . . . . . . . . . . . . . 14–13

14–2
 §14.2 PLATE IN PLANE STRESS

§14.1. Introduction

We now pass to the variational formulation of two-dimensional continuum finite elements. The
problem of plane stress will serve as the vehicle for illustrating such formulations. As narrated
in Appendix O, continuum-based structural finite elements were invented in the aircraft industry
(at Boeing during the early 1950s) to solve this kind of problem when it arose in the design and
analysis of delta wing panels [765].
The problem is presented here within the framework of the linear theory of elasticity.

§14.2. Plate in Plane Stress

In structural mechanics, a flat thin sheet of material is called a plate.1 The distance between the
plate faces is the thickness, denoted by h. The midplane lies halfway between the two faces.
The direction normal to the midplane is the transverse direction. Directions parallel to the midplane
are called in-plane directions. The global axis z is oriented along the transverse direction. Axes x
and y are placed in the midplane, forming a right-handed Rectangular Cartesian Coordinate (RCC)
system. Thus the equation of the midplane is z = 0. The +z axis conventionally defines the top
surface of the plate as the one that it intersects, whereas the opposite surface is called the bottom
surface. See Figure 14.1(a).

(a) (b) (c) y

z Referral to Mathematical
Top surface Midplane
midplane idealization

Plate Γ x

y
x Ω

Figure 14.1. A plate structure in plane stress: (a) configuration; (b) referral to its midplane;
(c) 2D mathematical idealization as boundary value problem.

§14.2.1. Behavioral Assumptions

A plate loaded in its midplane is said to be in a state of plane stress, or a membrane state, if the
following assumptions hold:
1. All loads applied to the plate act in the midplane direction, and are symmetric with respect to
the midplane.
2. All support conditions are symmetric about the midplane.
3. In-plane displacements, strains and stresses can be taken to be uniform through the thickness.
4. The normal and shear stress components in the z direction are zero or negligible.

1 If it is relatively thick, as in concrete pavements or Argentinian beefsteaks, the term slab is also used but not usually for
plane stress conditions.

14–3
Chapter 14: THE PLANE STRESS PROBLEM

In-plane internal forces y + sign conventions

for internal forces,
dx dy stresses and strains
h pyy
Thin plate in plane stress y x
pxx pxy dy
z x dx

dx dy In-plane stresses
In-plane body forces
y dx dy
dx dy
h
σyy y h
x y
x σxx σxy = σ yx by
bx
x

In-plane strains In-plane displacements

dx dy dx dy
h h
eyy y uy y
x e xx e xy = eyx x ux

Figure 14.2. Notational conventions for in-plane stresses, strains,

displacements and internal forces of a thin plate in plane stress.

The last two assumptions are not necessarily consequences of the first two. For the latter to hold,
the thickness h should be small, typically 10% or less, than the shortest in-plane dimension. If the
plate thickness varies it should do so gradually. Finally, the plate fabrication must exhibit symmetry
with respect to the midplane.
To these four assumptions we add the following restriction:
5. The plate is fabricated of the same material through the thickness. Such plates are called
transversely homogeneous or (in aerospace) monocoque plates.
The last assumption excludes wall constructions of importance in aerospace, in particular composite
and honeycomb sandwich plates. The development of mathematical models for such configurations
requires a more complicated integration over the thickness as well as the ability to handle coupled
bending and stretching effects, and will not be considered here.
Remark 14.1. Selective relaxation from assumption 4 leads to the so-called generalized plane stress state, in
which z stresses are accepted. The plane strain state is obtained if strains in the z direction are precluded.
Although the construction of finite element models for those states has many common points with plane stress,
we shall not consider those models here. For isotropic materials the plane stress and plane strain problems
can be mapped into each other through a fictitious-property technique; see Exercise 14.1.

Remark 14.2. Transverse loading on a plate produces plate bending, which is associated with a more complex
configuration of internal forces and deformations. This subject is studied in [255].

§14.2.2. Mathematical Model

The mathematical model of the plate in plane stress is set up as a two-dimensional boundary value
problem (BVP), in which the plate is projected onto its midplane; see Figure 14.1(b). This allows
to formulate the BVP over a plane domain  with a boundary , as illustrated in Figure 14.1(c).

14–4
 §14.2 PLATE IN PLANE STRESS

In this idealization the third dimension is represented as functions of x and y that are integrated
through the plate thickness. Engineers often work with internal plate forces, which result from
integrating the in-plane stresses through the thickness. See Figure 14.2.
§14.2.3. Problem Data

The following summarizes the givens in the plate stress problem.

Domain geometry. This is defined by the boundary  illustrated in Figure 14.1(c).
Thickness. Most plates used as structural components have constant thickness. If the thickness
does vary, in which case h = h(x, y), it should do so gradually to maintain the plane stress state.
Sudden changes in thickness may lead to stress concentrations.
Material data. This is defined by the constitutive equations. Here we shall assume that the plate
material is linearly elastic but not necessarily isotropic.
Specified Interior Forces. These are known forces that act in the interior  of the plate. There
are of two types. Body forces or volume forces are forces specified per unit of plate volume; for
example the plate weight. Face forces act tangentially to the plate faces and are transported to the
midplane. For example, the friction or drag force on an airplane skin is of this type if the skin is
modeled to be in plane stress.
Specified Surface Forces. These are known forces that act on the boundary  of the plate. In
elasticity they are called surface tractions. In actual applications it is important to know whether
these forces are specified per unit of surface area or per unit length. The former may be converted
to the latter by multiplying through the appropriate thickness value.
Displacement Boundary Conditions. These specify how the plate is supported. Points subject
to support conditions may be fixed, allowed to move in one direction, or subject to multipoint
constraints. Also symmetry and antisymmetry lines may be identified as discussed in Chapter 8 of
IFEM [257].
If no displacement boundary conditions are imposed, the plate is said to be free-free or floating.
§14.2.4. Problem Unknowns

The unknown fields are displacements, strains and stresses. Because of the assumed wall fabrication
homogeneity the in-plane components are assumed to be uniform through the plate thickness. Thus
the dependence on z disappears and all such components become functions of x and y only.
Displacements. The in-plane displacement field is defined by two components:
 
u x (x, y)
u(x, y) = (14.1)
u y (x, y)

The transverse displacement component u z (x, y, z) component is generally nonzero because of

Poisson’s ratio effects, and depends on z. However, this displacement does not appear in the
governing equations.
Strains. The in-plane strain field forms a tensor defined by three independent components: ex x ,
e yy and ex y . To allow stating the FE equations in matrix form, these components are cast to form a

14–5
Chapter 14: THE PLANE STRESS PROBLEM

Prescribed Displacement
BCs Displacements Body forces
displacements ^
^
u u = u u b
on Γu Γ
Ω

e = D u Kinematic Equilibrium DTσ + b = 0

in Ω (aka Balance) in Ω

Constitutive Stresses Force BCs Prescribed

Strains
σ tractions ^
t or
e σ=Ee σT n = ^t ^
forces q
or e = Cσ or pT n = q^
in Ω on Γt

Figure 14.3. The Strong Form of the plane stress equations of linear elastostatics displayed as a
Tonti diagram. Yellow boxes identify prescribed fields whereas orange boxes denote unknown fields.
The distinction between Strong and Weak Forms is explained in §14.3.3.

3-component “strain vector”  

ex x (x, y)
e(x, y) = e yy (x, y) (14.2)
2ex y (x, y)
The factor of 2 in ex y shortens strain energy expressions. The shear strain components ex z and e yz
vanish. The transverse normal strain ezz is generally nonzero because of Poisson’s ratio effects.
This strain does not enter the governing equations as unknown, however, because the associated
stress σzz is zero. This eliminates the contribution of σzz ezz to the internal energy.
Stresses. The in-plane stress field forms a tensor defined by three independent components: σx x ,
σ yy and σx y . As in the case of strains, to allow stating the FE equations in matrix form, these
components are cast to form a 3-component “stress vector”
 
σx x (x, y)
σ(x, y) = σ yy (x, y) (14.3)
σx y (x, y)
The remaining three stress components: σzz , σx z and σ yz , are assumed to vanish.
The plate internal forces are obtained on integrating the stresses through the thickness. Under the
assumption of uniform stress distribution,

px x = σx x h, p yy = σ yy h, px y = σx y h. (14.4)

These p’s also form a tensor. They are called membrane forces in the literature. See Figure 14.2.

§14.3. Plane Stress Governing Equations

We shall develop plane stress finite elements in the framework of classical linear elasticity. The
necessary governing equations are presented below. They are graphically represented in the Strong
Form Tonti diagram of Figure 14.3.

14–6
 §14.3 PLANE STRESS GOVERNING EQUATIONS

§14.3.1. Governing Equations

The three internal fields: displacements, strains and stresses (14.1)–(14.3) are connected by three
field equations: kinematic, constitutive and internal-equilibrium equations. If initial strain effects
are ignored, these equations read
    
ex x ∂/∂ x 0
ux
e yy = 0 ∂/∂ y ,
uy
2ex y ∂/∂ y ∂/∂ x
    
σx x E 11 E 12 E 13 ex x
σ yy = E 12 E 22 E 23 e yy , (14.5)
σx y E 13 E 23 E 33 2ex y
   σx x     
∂/∂ x 0 ∂/∂ y b 0
σ yy + x = .
0 ∂/∂ y ∂/∂ x by 0
σx y
The compact matrix version of (14.5) is

e = D u, σ = E e, DT σ + b = 0, (14.6)

Here E = ET is the 3 × 3 stress-strain matrix of plane stress elastic moduli, D is the 3 × 2

symmetric-gradient operator and its transpose the 2 × 3 tensor-divergence operator.2
If the plate material is isotropic with elastic modulus E and Poisson’s ratio ν, the moduli in the
constitutive matrix E reduce to E 11 = E 22 = E/(1 − ν 2 ), E 33 = 12 E/(1 + ν) = G, E 12 = ν E 11
and E 13 = E 23 = 0. See also Exercise 14.1.
§14.3.2. Boundary Conditions
Boundary conditions prescribed on  may be of two types: displacement BC or force BC (the
latter is also called stress BC or traction BC). To write down those conditions it is conceptually
convenient to break up  into two subsets: u and t , over which displacements and force or
stresses, respectively, are specified. See Figure 14.4.
Displacement boundary conditions are prescribed on u in the form

u = û. (14.7)

Here û are prescribed displacements. Often û = 0. This happens in fixed portions of the boundary,
as the ones illustrated in Figure 14.4.
Force boundary conditions (also called stress BCs and traction BCs in the literature) are specified
on t . They take the form
σn = t̂. (14.8)

Here t̂ are prescribed surface tractions specified as a force per unit area (that is, not integrated
through the thickness), and σn is the stress vector shown in Figure 14.4.

2 The dependence on (x, y) has been omitted to reduce clutter.

14–7
Chapter 14: THE PLANE STRESS PROBLEM

;;;;
; t

;;;;
;;
;
t^
Γu + Γt n (unit exterior

;;;
;;
; ;
^tn
σnt ^tt normal)

;;;
;;;
^t
σn
u^ = 0 σ nn
Boundary displacements u ^ Boundary tractions ^t or Stress BC details
are prescribed on Γu boundary forces q^ (decomposition of forces
(figure depicts fixity condition) are prescribed on Γt q^ would be similar)

Figure 14.4. Displacement and force (stress, traction) boundary conditions for the plane stress problem.

An alternative form of (14.8) uses the internal plate forces:

pn = q̂. (14.9)

Here pn = σn h and q̂ = t̂ h. This form is used more often than (14.8) in structural design,
particularly when the plate wall construction is inhomogeneous.
The components of σn in Cartesian coordinates follow from Cauchy’s stress transformation formula

     σx x 
σx x n x + σx y n y nx 0 ny
σn = = σ yy , (14.10)
σx y n x + σ yy n y 0 ny nx
σx y

in which n x and n y denote the Cartesian components of the unit normal vector ne (also called
the direction cosines of the normal). Thus (14.8) splits into two scalar conditions: tˆx = σnx and
tˆy = σny . The derivation of (14.10) is the subject of Exercise 14.4.
It is sometimes convenient to write the condition (14.8) in terms of normal n and tangential t
directions:
σnn = tˆn , σnt = tˆt (14.11)

in which σnn = σnx n x + σny n y and σnt = −σnx n y + σny n x . See Figure 14.4.

Remark 14.3. The separation of  into u and t is useful for conciseness in the mathematical formulation,
such as the energy integrals presented below. It does not exhaust, however, all BC possibilities. Frequently
at points of  one specifies a displacement in one direction and a force (or stress) in the other. An example
of these are roller and sliding conditions as well as lines of symmetry and antisymmetry. These are called
mixed displaceent-traction BC. To cover these situations one needs either a generalization of the boundary
split, in which u and t are permitted to overlap, or to define another portion m for‘mixed conditions. Such
generalizations will not be presented here, as they become unimportant once the FE discretization is done.

14–8
 §14.3 PLANE STRESS GOVERNING EQUATIONS

Prescribed Displacement
BCs Displacements Body forces
displacements ^
^
u u = u u b
on Γu Γ
Ω

e = D u Kinematic Equilibrium δΠ= 0

in Ω (weak) in Ω

Force BCs
Constitutive Stresses (weak) Prescribed
Strains
σ tractions ^
t or
e σ=Ee δΠ = 0 ^
forces q
or e = Cσ on Γt
in Ω

Figure 14.5. The TPE-based Weak Form of the plane stress equations of linear elastostatics.
Weak links are marked with grey lines.

§14.3.3. Weak versus Strong Form

We introduce now some further terminology from variational calculus. The Tonti diagram of Figure
14.3 is said to display the Strong Form of the governing equations because all relations are verified
point by point. These relations, called strong links, are shown in the diagram with black lines.
A Weak Form is obtained by relaxing one or more strong links, as brifley described in Chapter 11.
Those are replaced by weak links, which enforce relations in an average or integral sense rather
than point by point. The weak links are then provided by the variational formulation chosen for the
problem. Because in general many variational forms of the same problem are possible, there are
many possible Weak Forms. On the other hand the Strong Form is unique.
The Weak Form associated with the Total Potential Energy (TPE) variational form is illustrated
in Figure 14.5. The internal equilibrium equations and stress BC become weak links, which are
drawn by gray lines. These equations are given by the variational statement δ = 0, where the
TPE functional is given in the next subsection. The FEM displacement formulation discussed
below is based on this particular Weak Form.
§14.3.4. Total Potential Energy
The Total Potential Energy functional for the plane stress problem is given by

= U − W. (14.12)

The internal energy can be expressed in terms of the strains only as

 
U= 1
2
h σ e d =
T 1
2
h eT E e d. (14.13)
 

in which 12 eT Ee is the strain energy density. The derivation details are relegated to Exercise 14.5,
The external energy (potential of the applied forces) is the sum of contributions from the given

14–9
Chapter 14: THE PLANE STRESS PROBLEM

(a) (b)
(c)
Γ Ωe Γe

Figure 14.6. Finite element discretization and extraction of generic element.

interior (body) and exterior (boundary) forces:

 
W = h u b d +
T
h uT t̂ d. (14.14)
 t

Note that the boundary integral over  is taken only over t . That is, the portion of the boundary
over which tractions or forces are specified.

§14.4. Finite Element Equations

The necessary equations to apply the finite element method to the plane stress problem are collected
here and expressed in matrix form. The domain of Figure 14.6(a) is discretized by a finite element
mesh as illustrated in Figure 14.6(b). From this mesh we extract a generic element labeled e with
n ≥ 3 node points. In subsequent derivations the number n is kept arbitrary. Therefore, the
formulation is applicable to arbitrary two-dimensional elements, for example those sketched in
Figure 14.7.
To comfortably accommodate general element types, the node points will be labeled 1 through n.
These are called local node numbers. Numbering will always start with corners.
The element domain and boundary are denoted by e and  e , respectively. The element has 2n
degrees of freedom. These are collected in the element node displacement vector in a node by node
arrangement:
ue = [ u x1 u y1 u x2 . . . u xn u yn ]T . (14.15)

§14.4.1. Displacement Interpolation

The displacement field ue (x, y) over the element is interpolated from the node displacements. We
shall assume that the same interpolation functions are used for both displacement components.3
Thus
n 
n
u x (x, y) = Ni (x, y) u xi ,
e
u y (x, y) = Nie (x, y) u yi , (14.16)
i=1 i=1

3 This is the so called element isotropy condition, which is studied and justified in advanced FEM courses.

14–10
 §14.4 FINITE ELEMENT EQUATIONS

3 4
3 3 8
4 9 3
5
2 10 12 7
6
2 11 6
1
1
1 2 1 4 5 2
n=3 n=4 n=6 n = 12

Figure 14.7. Example plane stress finite elements, characterized by their number of nodes
n.

in which Nie (x, y) are the element shape functions. In matrix form:
   e 
u x (x, y) N1 0 N2e 0 ... Nne 0
u(x, y) = = ue = N ue . (14.17)
u y (x, y) 0 N1e 0 N2e ... 0 Nne

This N (with superscript e omitted to reduce clutter) is called the shape function matrix. It has
dimensions 2 × 2n. For example, if the element has 4 nodes, N is 2 × 8.
The interpolation condition on the element shape function Nie (x, y) states that it must take the value
one at the i th node and zero at all others. This ensures that the interpolation (14.17) is correct at
the nodes. Additional requirements on the shape functions are stated in later Chapters.
Differentiating the finite element displacement field yields the strain-displacement relations:
 ∂Ne ∂ N2e ∂ Nne 
1 0 0 . . . 0
 ∂x ∂x ∂x 
 ∂ N e
∂ N e
∂ N e 
e(x, y) =  0 ∂y
1 0 ∂y
2 ... 0 n  e
∂y  u = B u .
e
(14.18)
 e 
∂ N1 ∂ N1 ∂ N2 ∂ N2
e e e e
∂N e
∂ Nn
∂y ∂x ∂y ∂x . . . ∂ yn ∂x
This B = D N is called the strain-displacement matrix. It is dimensioned 3 × 2n. For example, if
the element has 6 nodes, B is 3 × 12. The stresses are given in terms of strains and displacements
by σ = E e = EBue , which is assumed to hold at all points of the element.
§14.4.2. Element Energy
To obtain finite element stiffness equations, the variation of the TPE functional is decomposed into
contributions from individual elements:
δ e = δU e − δW e = 0. (14.19)
in which  
U =
e 1
2
h σ e d =
T e 1
2
h eT Ee de (14.20)
e e

and  
W =
e
h u b d +
T e
h uT t̂ d e (14.21)
e e

14–11
Chapter 14: THE PLANE STRESS PROBLEM

Note that in (14.21) te has been taken equal to the complete boundary  e of the element. This is
a consequence of the fact that displacement boundary conditions are applied after assembly, to a
free-free structure. Consequently it does not harm to assume that all boundary conditions are of
stress type insofar as forming the element equations.
§14.4.3. Element Stiffness Equations
Inserting the relations u = Nue , e = Bue and σ = Ee into e yields the quadratic form in the
nodal displacements
e = 12 ue T Ke ue − ue T fe . (14.22)
Here the element stiffness matrix is

K = e
h BT EB de , (14.23)
e

and the consistent element nodal force vector is

 
f =
e
h N b d +
T e
h NT t̂ d e . (14.24)
e e

In the second integral of (14.24) the matrix N is evaluated on the element boundary only.
The calculation of the entries of Ke and fe for several elements of historical or practical interest is
described in subsequent Chapters.
Notes and Bibliography
The plane stress problem is well suited for introducing continuum finite elements, from both historical and
technical standpoints. Some books use the Poisson equation for this purpose, but problems such as heat
conduction cannot illustrate features such as vector-mixed boundary conditions and shear effects.
The first continuum structural finite elements were developed at Boeing in the early 1950s to model delta-wing
skin panels [146,765]. A plane stress model was naturally chosen for the panels. The paper that gave the
method its name [137] used the plane stress problem as application driver.
The technical aspects of plane stress can be found in any book on elasticity. A particularly readable one is the
excellent textbook by Fung [289], which is unfortunately out of print.

References
Referenced items have been moved to Appendix R.

14–12
 Exercises

Homework Exercises for Chapter 14

The Plane Stress Problem

EXERCISE 14.1 [A+C:15] Suppose that the structural material is isotropic, with elastic modulus E and
Poisson’s ratio ν. The in-plane stress-strain relations for plane stress (σzz = σx z = σ yz = 0) and plane strain
(ezz = ex z = e yz = 0) as given in any textbook on elasticity, are
    
σx x E 0 1 ν ex x
plane stress: σ yy = 0 ν 1 e yy ,
σx y 1 − ν2 1−ν 0 0 2ex y
2 (E14.1)
    
σx x E 1−ν ν 0 ex x
plane strain: σ yy = ν 1−ν 0 e yy .
σx y (1 + ν)(1 − 2ν) 0 0 1
(1 − 2ν) 2ex y
2

Show that the constitutive matrix of plane strain can be formally obtained by replacing E by a fictitious
modulus E ∗ and ν by a fictitious Poisson’s ratio ν ∗ in the plane stress constitutive matrix. Find the expression
of E ∗ and ν ∗ in terms of E and ν.
You may also chose to answer this exercise by doing the inverse process: go from plane strain to plain stress
by replacing a fictitious modulus and Poisson’s ratio in the plane strain constitutive matrix.
This device permits “reusing” a plane stress FEM program to do plane strain, or vice-versa, as long as the
material is isotropic.
Partial answer to go from plane stress to plane strain: ν ∗ = ν/(1 − ν).

EXERCISE 14.2 [A:25] In the finite element formulation of near incompressible isotropic materials (as well
as plasticity and viscoelasticity) it is convenient to use the so-called Lamé constants λ and µ instead of E and
ν in the constitutive equations. Both λ and µ have the physical dimension of stress and are related to E and ν
by
νE E
λ= , µ=G= . (E14.2)
(1 + ν)(1 − 2ν) 2(1 + ν)
Conversely
µ(3λ + 2µ) λ
E= , ν= . (E14.3)
λ+µ 2(λ + µ)
Substitute (E14.3) into both of (E14.1) to express the two stress-strain matrices in terms of λ and µ. Then split
the stress-strain matrix E of plane strain as

E = E µ + Eλ (E14.4)

in which Eµ and Eλ contain only µ and λ, respectively, with Eµ diagonal and E λ33 = 0. This is the Lamé or
{λ, µ} splitting of the plane strain constitutive equations, which leads to the so-called B-bar formulation of
near-incompressible finite elements.4 Express Eµ and Eλ also in terms of E and ν.
For the plane stress case perform a similar splitting in which where Eλ contains only λ̄ = 2λµ/(λ + 2µ) with
E λ33 = 0, and Eµ is a diagonal matrix function of µ and λ̄.5 Express Eµ and Eλ also in terms of E and ν.

4 Equation (E14.4) is sometimes referred to as the deviatoric+volumetric splitting of the stress-strain law, on account of
its physical meaning in plane strain. That interpretation is not fully accurate, however, for plane stress.
5 For the physical significance of λ̄ see [688, pp. 254ff].

14–13
Chapter 14: THE PLANE STRESS PROBLEM

EXERCISE 14.3 [A:20] Include thermoelastic effects in the plane stress constitutive field equations, assuming
a thermally isotropic material with coefficient of linear expansion α. Hint: start from the two-dimensional
Hooke’s law including temperature:
1 1
ex x = (σx x − νσ yy ) + α T, e yy = (σ yy − νσx x ) + α T, 2ex y = σx y /G, (E14.5)
E E
in which T = T (x, y) and G = 12 E/(1 + ν). Solve for stresses and collect effects of T in one vector
of “thermal stresses.”

EXERCISE 14.4 [A:15] Derive the Cauchy stress- ty n(nx =dx/ds, ny =dy/ds)
to-traction equations (14.10) using force equilibrium y
along x and y and the geometric relations shown σx x
x dy ds tx
in Figure E14.1. (This is the “wedge method” in dx
Mechanics of Materials.) σx y = σy x
σyy
Hint: tx ds = σx x dy + σx y d x, etc.
Figure E14.1. Geometry for deriving (?).

EXERCISE 14.5 [A:25=5+5+15] A linearly elastic plate is in plane stress. It is shown in courses in elasticity
that the internal strain energy density stored per unit volume of the plate expressed in terms of stresses and
strains is the bilinear form

U = 12 (σx x ex x + σ yy e yy + σx y ex y + σ yx e yx ) = 12 (σx x ex x + σ yy e yy + 2σx y ex y ) = 1

2
σT e. (E14.6)

(a) Show that (E14.6) can be written in terms of strains only as

U= 1
2
eT E e, (E14.7)

thus justifying the strain energy density expression given in (14.13) for the plane stress problem.
(b) Show that (E14.6) can be written in terms of stresses only as

U= 1
2
σT C σ, (E14.8)

in which C = E−1 is the elastic compliance (strain-stress) matrix.

(c) Suppose you want to write (E14.6) in terms of the extensional strains {ex x , e yy } and of the shear stress
σx y = σ yx . This is known as a mixed representation, which is used in finite elements formulated with
mixed variational principles. Show that
 T   
ex x A11 A12 A13 ex x
U= 1
2
e yy A12 A22 A23 e yy , (E14.9)
σx y A13 A23 A33 σx y

and explain how the entries Ai j of the kernel matrix A that appears in (E14.9) can be calculated6 in terms
of the elastic moduli E i j .
Hint. Parts (a,b) are easy. Part (c) is more difficult. It can be symbolically done by the Mathematica script

6 The process of computing A is an instance of “partial inversion” of the elasticity matrix E. It is closely related to the
Schur complement concept covered in Appendix P.

14–14
 Exercises

ClearAll[exx,eyy,gxy,sxx,syy,sxy,E11,E22,E33,E12,E13,E23];
Emat={{E11,E12,E13},{E12,E22,E23},{E13,E23,E33}};
s={sxx,syy,sxy}; e={exx,eyy,gxy}; m={exx,eyy,sxy};
eqs={sxx==E11*exx+E12*eyy+E13*gxy,syy==E12*exx+E22*eyy+E23*gxy,
sxy==E13*exx+E23*eyy+E33*gxy};
sol=Simplify[Simplify[Solve[eqs,{sxx,syy,gxy}]]];
Print[sol]; U=Simplify[(e.Emat.e/2)/.sol[[1]]];
fac[i_,j_]:=If[i==j,1,1/2];
A=Table[fac[i,j]*Coefficient[U,m[[i]]*m[[j]]],{i,1,3},{j,1,3}];
Print["A=",A//MatrixForm];

If you use this solution, make sure to explain what is going on.

Note: the following Table list relations between commonly used moduli for isotropic linear elastic material.
Here K is the bulk modulus whereas M is the P-wave modulus used in seismology. Tha table is useful for
Exercise 14.2.

(λ, µ) (E, µ) (K , λ) (K , µ) (λ, ν) (µ, ν) (E, ν) (K , ν) (K , E)

2µ Eµ 2µ(1+ν)
K = λ+ λ 1+ν E
3 3(3µ−E) 3ν 3(1−2ν) 3(1−2ν)
9K µ λ(1+ν)(1−2ν)
9K K −λ
3λ+2µ
E = µ λ+ν ν 2µ(1+ν) 3K (1−2ν)
3K −λ 3K +µ
E−2µ 2µ 2µν Eν 3K 3K (3K −E)
λ= µ K−
3µ−E 3 1−2ν (1+ν)(1−2ν) 1+ν 9K −E
9K µ λ(1−2ν) (1−2ν)
µ=G= µ K −λ λ K −λ E 3K
2 3K −λ 3K +µ 2ν 2(1+ν) 2(1+ν)
λ E λ 3K −2µ 3K −E
ν= −1
2(λ+µ) 2µ 3K −λ 2(3K +µ) 6K
M= λ+2µ µ
4µ−E
3K −2λ K+
4µ
λ 1−ν µ 2−2ν
E(1−ν)
3K 1−ν 3K 3K +E
3µ−E 3 ν 1−2ν (1+ν)(1−2ν) 1+ν 9K −E
(E14.10)

14–15
 15
Three-Node
Plane Stress
Triangles

15–1
Section 15: THREE-NODE PLANE STRESS TRIANGLES

TABLE OF CONTENTS

Page
§15.1 Introduction . . . . . . . . . . . . . . . . . . . . . 15–3
§15.2 Background . . . . . . . . . . . . . . . . . . . . . 15–3
§15.2.1 Parametric Representation of Functions . . . . . . . . . 15–3
§15.2.2 Geometry . . . . . . . . . . . . . . . . . . 15–4
§15.2.3 Triangular Coordinates . . . . . . . . . . . . . . 15–4
§15.2.4 Linear Interpolation . . . . . . . . . . . . . . . 15–5
§15.2.5 Coordinate Transformations . . . . . . . . . . . . 15–5
§15.2.6 Partial Derivatives . . . . . . . . . . . . . . . 15–6
§15.2.7 *Homogeneous Polynomials in Triangular Coordinates . . . 15–7
§15.2.8 *Interesting Points and Lines . . . . . . . . . . . . 15–7
§15.3 The Turner Triangle . . . . . . . . . . . . . . . . . . 15–8
§15.3.1 Strain-Displacement Equations . . . . . . . . . . . 15–8
§15.3.2 Stress-Strain Equations . . . . . . . . . . . . . . 15–8
§15.3.3 The Stiffness Matrix . . . . . . . . . . . . . . 15–9
§15.3.4 The Consistent Nodal Force Vector . . . . . . . . . . 15–9
§15.3.5 Implementation . . . . . . . . . . . . . . . . 15–10
§15.3.6 *Consistency Verification . . . . . . . . . . . . . 15–11
§15.3.7 *Checking Continuity . . . . . . . . . . . . . . 15–11
§15.3.8 *Checking Completeness . . . . . . . . . . . . . 15–12
§15.3.9 *Tonti Matrix Diagram . . . . . . . . . . . . . . 15–12
§15.4 *Derivation Using Natural Strains and Stresses . . . . . . . . 15–12
§15.4.1 *Natural Strains and Stresses . . . . . . . . . . . . 15–12
§15.4.2 *Covariant Node Displacements . . . . . . . . . . . 15–13
§15.4.3 *The Natural Stiffness Matrix . . . . . . . . . . . 15–14
§15.5 *The Veubeke Equilibrium Triangle . . . . . . . . . . . . . 15–14
§15.5.1 *Kinematic Relations . . . . . . . . . . . . . . 15–14
§15.5.2 *Stiffness Matrix . . . . . . . . . . . . . . . . 15–15
§15.5.3 *Implementation . . . . . . . . . . . . . . . . 15–15
§15.5.4 *Spurious Kinematic Modes . . . . . . . . . . . . 15–16
§15.6 *Shear Locking in Turner Triangles . . . . . . . . . . . . 15–17
§15.6.1 *The Inplane Bending Test . . . . . . . . . . . . . 15–18
§15.6.2 *Energy Ratios . . . . . . . . . . . . . . . . 15–18
§15.6.3 *Convergence as Mesh is Refined . . . . . . . . . . . 15–20
§15. Notes and Bibliography . . . . . . . . . . . . . . . . . 15–20
§15. References . . . . . . . . . . . . . . . . . . . . . . 15–21
§15. Exercises . . . . . . . . . . . . . . . . . . . . . . 15–22
§15. Solutions to Exercises . . . . . . . . . . . . . . . . . . 15–27

15–2
 §15.2 BACKGROUND

§15.1. Introduction

This Chapter derives element stiffness equations of three-node triangles constructed with linear
displacements for the plane stress problem formulated in Chapter 14. These elements have six
displacement degrees of freedom, which are placed at the connection nodes. There are two main
versions that differ on where the connection nodes are located:
1. The Turner triangle has connection nodes located at the corners.
2. The Veubeke equilibrium triangle has connection nodes located at the side midpoints.
The triangle geometry is defined by the corner locations or geometric nodes in both cases. Of the
two versions, the Turner triangle is by far the most practically important one in solid and structural
mechanics.1 Thus most of the material in this Chapter is devoted to it. It enjoys several important
properties:
(i) It belongs to both the isoparametric and subparametric element families, which are introduced
in the next Chapter.
(ii) It allows closed form derivations for the stiffness matrix and consistent force vector without
need for numerical integration.
(iii) It cannot be improved by the addition of internal degrees of freedom.
Properties (ii) and (iii) are shared by the Veubeke equilibrium triangle. Since this model is rarely
used in structural applications it is covered only as advanced material in §15.5.
The Turner triangle is not a good performer for structural stress analysis. It is still used in problems
that do not require high accuracy, as well as in non-structural applications such as thermal and
electromagnetic analysis. One reason is that triangular meshes are easily generated over arbitrary
two-dimensional domains using techniques such as Delaunay triangulation.

§15.2. Background
§15.2.1. Parametric Representation of Functions
The concept of parametric representation of functions is crucial in modern FEM. Together with
multidimensional numerical integration, it is a key enabling tool for developing elements in two
and three space dimensions.2 Without these tools the developer would become lost in an algebraic
maze as element geometry and shape functions get more complicated. The essentials of parametric
representation can be illustrated through a simple example. Consider the following alternative
representations of the unit-circle function, x 2 + y 2 = 1:

(I) y = 1 − x 2 , (II) x = cos θ and y = sin θ. (15.1)

The direct representation (I) fits the conventional function notation, i.e., y = f (x). Given a value of
x, it returns one or more y. On the other hand, the parametric representation (II) is indirect: both x

1 The triangle was one of the two plane-stress continuum elements presented by Turner, Clough, Martin and Topp in their
1956 paper [786]. This publication is widely regarded as the start of the present FEM. The derivation was not done,
however, with assumed displacements. See Notes and Bibliography at the end of this Chapter.
2 Numerical integration is not useful for the triangular elements covered here, but essential in the more complicated iso-P
models covered in Chapters 16ff.

15–3
Section 15: THREE-NODE PLANE STRESS TRIANGLES

and y are given in terms of one parameter, the angle θ . Elimination of θ through the trigonometric
identity cos2 θ + sin2 θ = 1 recovers x 2 + y 2 = 1. But there are situations in which working with
the parametric form throughout the development is more convenient. Continuum finite elements
provide a striking illustration of this point.
§15.2.2. Geometry
The geometry of the 3-node triangle shown in
Figure 15.1(a) is specified by the location of its (a) (b)
3 (x3 ,y3) 3
three corner nodes on the {x, y} plane. Nodes
are labelled 1, 2, 3 while traversing the sides Area A > 0
in counterclockwise fashion. Their location is
defined by their Cartesian coordinates: {xi , yi } 2 (x2 ,y2) 2
for i = 1, 2, 3.
y
The Turner triangle has six degrees of freedom,
1 (x 1 ,y1) 1
defined by the six corner displacement compo-
nents { u xi , u yi }, for i = 1, 2, 3. The interpo- x
lation of the internal displacements { u x , u y } z up, toward you
from these six values is studied in §15.3, after Figure 15.1. The three-node, linear-displacement
triangular coordinates are introduced. The plane stress triangular element: (a) geometry; (b) area
triangle area can be obtained as and positive boundary traversal.
 
1 1 1
2A = det x1 x2 x3 = (x2 y3 − x3 y2 ) + (x3 y1 − x1 y3 ) + (x1 y2 − x2 y1 ). (15.2)
y1 y2 y3
The area given by (15.2) is a signed quantity. It is positive if the corners are numbered in cyclic
counterclockwise order (when looking down from the +z axis), as illustrated in Figure 15.1(b).
This convention is followed in the sequel.
§15.2.3. Triangular Coordinates
Points of the triangle may also be located in terms of a parametric coordinate system:
ζ1 , ζ 2 , ζ3 . (15.3)
In the literature these 3 parameters receive an astonishing number of names, as the list collected
in Table 15.1 shows. In the sequel the name triangular coordinates will be used to emphasize the
close association with this particular geometry.
Equations
ζi = constant (15.4)

represent a set of straight lines parallel to the side opposite to the i th corner, as depicted in Figure 15.2.
The equations of sides 2–3, 3–1 and 1–2 are ζ1 = 0, ζ2 = 0 and ζ3 = 0, respectively. The
three corners have coordinates (1,0,0), (0,1,0) and (0,0,1). The three midpoints of the sides have
coordinates ( 12 , 12 , 0), (0, 12 , 12 ) and ( 12 , 0, 12 ), the centroid has coordinates ( 13 , 13 , 13 ), and so on. The
coordinates are not independent because their sum is unity:
ζ1 + ζ2 + ζ3 = 1. (15.5)

15–4
 §15.2 BACKGROUND

3 ζ1 = 0 3 3 ζ3 = 1
/
ζ2 = 0 ζ2 = 1
/ /
/ / /
2 / / 2 2
/
1 1 1 ζ3 = 0
ζ1 = 1

Figure 15.2. Triangular coordinates ζ1 , ζ2 , ζ3 .

Table 15.1 Names of element parametric coordinates

Name Applicable to
natural coordinates all elements
isoparametric coordinates isoparametric elements
shape function coordinates isoparametric elements
barycentric coordinates simplices (triangles, tetrahedra, ...)
Möbius coordinates triangles
triangular coordinates all triangles
area (also written “areal”) coordinates straight-sided triangles
Triangular coordinates normalized as per ζ1 + ζ2 + ζ3 = 1 are often
qualified as “homogeneous” in the mathematical literature.

Remark 15.1. In pre-1970 FEM publications, triangular coordinates were often called area coordinates, and
occasionally areal coordinates. This comes from the following interpretation: ζi = A jk /A, where A jk is the
area subtended by the subtriangle formed by the point P and corners j and k, in which j and k are 3-cyclic
permutations of i. Historically this was the way coordinates were defined in 1960s papers. However this
relation does not carry over to general isoparametric triangles with curved sides and thus it is not used here.

§15.2.4. Linear Interpolation

Consider a function f (x, y) that varies linearly over the triangle domain. In terms of Cartesian
coordinates it may be expressed as
f (x, y) = a0 + a1 x + a2 y, (15.6)
where a0 , a1 and a2 are coefficients to be determined from three conditions. In finite element work
such conditions are often the nodal values taken by f at the corners:
f1, f2, f3. (15.7)
The expression in triangular coordinates makes direct use of those three values:
   
ζ1 f1
f (ζ1 , ζ2 , ζ3 ) = f 1 ζ1 + f 2 ζ2 + f 3 ζ3 = [ f 1 f 2 f 3 ] ζ2 = [ ζ1 ζ2 ζ3 ] f 2 . (15.8)
ζ3 f3
Formula (15.8) is called a linear interpolant for f .

15–5
Section 15: THREE-NODE PLANE STRESS TRIANGLES

§15.2.5. Coordinate Transformations

Quantities that are closely linked with the element geometry are best expressed in triangular co-
ordinates. On the other hand, quantities such as displacements, strains and stresses are usually
expressed in the Cartesian system {x, y}. Thus we need transformation equations through which it
is possible to pass from one coordinate system to the other.
Cartesian and triangular coordinates are linked by the relation

    
1 1 1 1 ζ1
x = x1 x2 x3 ζ2 . (15.9)
y y1 y2 y3 ζ3

The first equation says that the sum of the three coordinates is one. The next two express x and
y linearly as homogeneous forms in the triangular coordinates. These are obtained by applying
the linear interpolant (15.8) to the Cartesian coordinates: x = x1 ζ1 + x2 ζ2 + x3 ζ3 and y =
y1 ζ1 + y2 ζ2 + y3 ζ3 . Assuming A = 0, inversion of (15.9) yields
       
ζ1 1 x2 y3 − x3 y2 y2 − y3 x3 − x2 1 1 2A23 y23
1 x32
ζ2 = x3 y1 − x1 y3 y3 − y1 x1 − x3 x = 2A31 x .
y31 x13
ζ3 2A x1 y2 − x2 y1 y1 − y2 x2 − x1 y 2A 2A12 y12
y x21
(15.10)
Here x jk = x j − xk , y jk = y j − yk , A is the triangle area given by (15.2) and A jk denotes the area
subtended by corners j, k and the origin of the x–y system. If this origin is taken at the centroid of
the triangle, A23 = A31 = A12 = A/3.

§15.2.6. Partial Derivatives

From equations (15.9) and (15.10) we immediately obtain the following relations between partial
derivatives:
∂x ∂y
= xi , = yi , (15.11)
∂ζi ∂ζi

∂ζi ∂ζi
2A = y jk , 2A = xk j . (15.12)
∂x ∂y

In (15.12) j and k denote the 3-cyclic permutations of i. For example, if i = 2, then j = 3 and
k = 1. The derivatives of a function f (ζ1 , ζ2 , ζ3 ) with respect to x or y follow immediately from
(15.12) and application of the chain rule:

 
∂f 1 ∂f ∂f ∂f
= y + y + y
∂x 2A ∂ζ1 23 ∂ζ2 31 ∂ζ3 12
  (15.13)
∂f 1 ∂f ∂f ∂f
= x + x + x
∂y 2A ∂ζ1 32 ∂ζ2 13 ∂ζ3 21

15–6
 §15.2 BACKGROUND

(a) 3 (b) 3 (c) 3

Μ1 RC Μ1
Η1
Μ2 Η2
Μ2 OC
C 2 Η 2 2

Μ3 Η3 Μ3

1 1 1

Figure 15.3. Interesting points and lines of a triangle.

which in matrix form is

 ∂f 
 ∂f   ∂ζ 
 1 
 ∂x  1 y23 y31 y12  ∂ f 
 ∂f  =  . (15.14)
2A x32 x13 x21 
 ∂ζ2


∂y  ∂f 
∂ζ3
With these mathematical ingredients in place we are now in a position to handle the derivation of
straight-sided triangular elements, and in particular the Turner and Veubeke triangles.
§15.2.7. *Homogeneous Polynomials in Triangular Coordinates
Because ζ1 , ζ2 and ζ3 are not independent, polynomial functions in those variables are not unique. For example
3 − 2ζ1 + ζ2 − 3ζ3 and ζ1 + 4ζ2 are identical, since they differ by 3 − 3(ζ1 + ζ2 + ζ3 )=0. To achieve uniqueness
it is necessary to write the function as a homogeneous polynomial, as in the second form of this example.
To reduce the general linear polynomial c000 + c100 ζ1 + c010 ζ2 + c001 ζ3 to homogeneous form, subtract c000 (1 −
ζ1 − ζ2 − ζ3 ), which is zero, to get P1 = (c100 − c000 )ζ1 + (c010 − c000 )ζ2 + (c001 − c000 )ζ3 .
To reduce the general quadratic polynomial c000 +c100 ζ1 +c010 ζ2 +c001 ζ3 +c200 ζ12 +c020 ζ22 +c002 ζ32 +c110 ζ1 ζ2 +
c011 ζ2 ζ3 + c101 ζ3 ζ1 to homogeneous form, subtract (c000 + c100 ζ1 + c010 ζ2 + c001 ζ3 )(1 − ζ1 − ζ2 − ζ3 ).
And so on. All polynomial expressions used in this book for triangles are expressed in homogeneous form.
§15.2.8. *Interesting Points and Lines
Some distinguished lines and points of a straight-sided triangle are briefly described here for use in other
developments as well as in Exercises. The triangle medians are three lines that join the corners to the
midpoints of the opposite sides, as pictured in Figure 15.3(a). The midpoint opposite corner i is labeled Mi .
The medians 1–M1 , 2–M2 and 3–M3 have equations ζ2 = ζ3 , ζ3 = ζ1 and ζ1 = ζ2 , respectively, in triangular
coordinates. They intersect at the centroid C of coordinates { 13 , 13 , 13 }. Other names for the centroid are
barycenter and center of gravity. If you make a real triangle out of cardboard, you can balance the triangle at
this point. It can be shown that the centroid trisects the medians, that is to say, the distance from a corner to
the centroid is twice the distance from the centroid to the opposite side of the triangle.
The altitudes are three lines that connect each corner with their projections onto the opposing sides, as
depicted in Figure 15.3(b). The projection of corner i is identified Hi , so the altitudes are 1–H1 , 2–H2 and
3–H3 . Locations Hi are called altitude feets. The altitudes intersect at the triangle orthocenter H . The lengths
of those segments are the triangle heights. The triangular coordinates of Hi and H , as well as the altitude
equations, are worked out in an Exercise.

15–7
Section 15: THREE-NODE PLANE STRESS TRIANGLES

Another interesting point is the center OC of the circumscribed circle, or circumcircle. This is the unique
circle that passes through the three corners, as shown in Figure 15.3(c). It can be geometrically constructed by
drawing the normal to each side at the midpoints. Those three lines, called the perpendicular side bisectors,
intersect at OC . A famous theorem by Euler asserts that the centroid, the orthocenter and the circumcircle
center fall on a straight line, called the Euler line. Furthermore, C lies between OC and H , and the distance
OC –H is three times the distance H –C.

§15.3. The Turner Triangle

The simplest triangular element for plane stress (and in general, for 2D problems of variational
index m = 1) is the three-node triangle with linear shape functions, with degrees of freedom
located at the corners. The shape functions are simply the triangular coordinates. That is, Nie = ζi
for i = 1, 2, 3. When applied to the plane stress problem, this element is called the Turner triangle.
For the plane stress problem we select the linear interpolation (15.8) for the displacement compo-
nents u x and u y at an arbitrary point P(ζ1 , ζ2 , ζ3 ):

u x = u x1 ζ1 + u x2 ζ2 + u x3 ζ3 , u y = u y1 ζ1 + u y2 ζ2 + u y3 ζ3 . (15.15)

The interpolation is illustrated in Figure 15.4. The two

expressions in (15.15) can be combined in a matrix form
}
u x by linear
that befits the expression (14.17) for an arbitrary plane u y3 P(ζ1 ,ζ 2 ,ζ 3 )
u y interpolation
stress element: 3
  ux3
u x1
uy uy2
  u 
 y1 
ux ζ 0 ζ2 0 ζ3 0  u x2  P ux2
= 1   = N ue , ux
uy 0 ζ1 0 ζ2 0 ζ3  u y2  uy1 2
 
u x3
u y3 1 ux1
(15.16)
where N is the matrix of shape functions. Figure 15.4. Displacement
interpolation over triangle.

§15.3.1. Strain-Displacement Equations

The strains within the elements are obtained by differentiating the shape functions with respect to
x and y. Using (15.14), (15.16) and the general form (14.18) we get
 
u x1
 u 
y23 0 y31 0 y12 0  y1 
1 u 
e = D N ue = 0 x32 0 x13 0 x21  x2  = B ue , (15.17)
2A x  u y2 
32 y23 x 13 y 31 x 21 y 12  
u x3
u y3
in which D denotes the symbolic strain-to-displacement differentiation operator given in (14.6),
and B is the strain-displacement matrix. Note that the strains are constant over the element. This is
the origin of the name constant strain triangle (CST) given it in many finite element publications.

15–8
 §15.3 THE TURNER TRIANGLE

§15.3.2. Stress-Strain Equations

The stress field σ is related to the strain field by the elastic constitutive equation in (14.5), which is
repeated here for convenience:
    
σx x E 11 E 12 E 13 ex x
σ = σ yy = E 12 E 22 E 23 e yy = E e, (15.18)
σx y E 13 E 23 E 33 2ex y
where E i j are plane stress elastic moduli. The constitutive matrix E will be assumed to be constant
over the element. Because the strains are constant, so are the stresses.
§15.3.3. The Stiffness Matrix
The element stiffness matrix is given by the general formula (14.23), which is repeated here

K =
e
h BT EB d, (15.19)
e
where  is the triangle domain, and h the plate thickness that appears in the plane stress problem.
e

Since B and E are constant, they can be taken out of the integral:

K = B EB
e T
h d (15.20)
e
If h is uniform over the element the remaining integral in (15.20) is simply h A, and we obtain the
closed form
 
y23 0 x32
 0 x32 y23    
  E 11 E 12 E 13 y23 0 y31 0 y12 0
h  31
y 0 x 13 
Ke = A h B T E B =   E E E 0 x32 0 x13 0 x21 .
4A  0 x13 y31  E 12 E 22 E 23
  13 23 33 x32 y23 x13 y31 x21 y12
y12 0 x21
0 x21 y12
(15.21)
Exercise 15.1 deals with the case of a linearly varying plate thickness.
§15.3.4. The Consistent Nodal Force Vector
For simplicity we consider here only internal body forces3 defined by the vector field

bx
b= (15.22)
by
which is specified per unit of volume. The consistent nodal force vector fe is given by the general
formula (14.23) of the previous Chapter:
 
ζ1 0
   0 ζ1 
 
ζ 0
fe = h NT b d = h 2  b d. (15.23)
e e  0 ζ2 
 
ζ3 0
0 ζ3
3 For consistent force computations corresponding to distributed boundary loads over a side, see Exercise 15.4.

15–9
Section 15: THREE-NODE PLANE STRESS TRIANGLES

Trig3TurnerMembraneStiffness[ncoor_,Emat_,h_,numer_]:=Module[{
x1,x2,x3,y1,y2,y3,x21,x32,x13,y12,y23,y31,A,Be,Ke},
{{x1,y1},{x2,y2},{x3,y3}}=ncoor;
A=Simplify[(x2*y3-x3*y2+(x3*y1-x1*y3)+(x1*y2-x2*y1))/2];
{x21,x32,x13,y12,y23,y31}={x2-x1,x3-x2,x1-x3,y1-y2,y2-y3,y3-y1};
Be={{y23,0,y31,0,y12,0},{0,x32,0,x13,0,x21},
{x32,y23,x13,y31,x21,y12}}/(2*A);
If [numer, Be=N[Be]]; Ke=A*h*Transpose[Be].Emat.Be;
Return[Ke]];

Figure 15.5. Implementation of Turner triangle stiffness matrix calculation as a Mathematica module.

The simplest case is when the body force components (15.22) as well as the thickness h are constant
over the element. Then we need the integrals
  
ζ1 d = ζ2 d = ζ3 d = 13 A (15.24)
e e e

which replaced into (15.23) gives

Ah
fe = [ bx by bx by bx b y ]T . (15.25)
3

This agrees with the simple element-by-element force-lumping procedure, which assigns one third
of the total force along the {x, y} directions: Ahbx and Ahb y , to each corner.
Remark 15.2. The integrals (15.24) are particular cases of the general integration formula of monomials in
triangular coordinates:

1 j i! j! k!
ζ1i ζ2 ζ3k d = , i ≥ 0, j ≥ 0, k ≥ 0. (15.26)
2A e (i + j + k + 2)!

which can be derived through the Beta function. Here i, j, k are integer exponents. This formula only holds
for triangles with straight sides, and thus does not apply for higher order elements with curved sides. Formulas
(15.24) are obtained by setting exponents i = 1, j = k = 0 in (15.26), and permuting {i, j, k} cyclically.

§15.3.5. Implementation
The implementation of the Turner triangle in any programming language is very simple. A Mathe-
matica module that returns Ke is shown in Figure 15.5. The module needs only 8 lines of code. It
is invoked as

Ke=Trig3TurnerMembraneStiffness[ncoor,Emat,h,numer]; (15.27)

The arguments are

ncoor Element node coordinates, arranged as a list: { { x1,y1 },{ x2,y2 },{ x3,y3 } }.
Emat A two-dimensional list storing the 3 × 3 plane stress matrix of elastic moduli as
{ { E11,E12,E13 },{ E12,E22,E23 },{ E13,E23,E33 } }.
h Plate thickness, assumed uniform over the triangle.

15–10
 §15.3 THE TURNER TRIANGLE

ncoor={{0,0},{3,1},{2,2}}; Emat=8*{{8,2,0},{2,8,0},{0,0,3}};
Ke=Trig3TurnerMembraneStiffness[ncoor,Emat,1,False];
Print["Ke=",Ke//MatrixForm];
Print["eigs of Ke=",Chop[Eigenvalues[N[Ke]]]];
Show[Graphics[RGBColor[1,0,0]],Graphics[AbsoluteThickness[2]],
Graphics[Polygon[ncoor]],Axes->True];

11 5 − 10 −2 −1 −3 2
5 11 2 10 −7 − 21
1.5
− 10 2 44 − 20 − 34 18
Ke =
−2 10 − 20 44 22 − 54 1
−1 −7 − 34 22 35 − 15
0.5
−3 − 21 18 − 54 − 15 75
eigs of Ke = {139.33, 60., 20.6704, 0, 0, 0} 0.5 1 1.5 2 2.5 3

Figure 15.6. Test statements to exercise the module of Figure 15.5, and outputs.

numer A logical flag: True to request floating-point computation, else False.

This module is exercised by the statements listed at the top of Figure 15.6, which form a triangle
with corner coordinates { { 0,0 },{ 3,1 },{ 2,2 } }, isotropic material matrix with E 11 = E 22 = 64,
E 12 = 16, E 33 = 24, others zero, (that is, E = 60 and ν = 14 ) and unit thickness. The results are
shown at the bottom of Figure 15.6. The computation of stiffness matrix eigenvalues is always a
good programming test, since 3 eigenvalues must be exactly zero and the other 3 real and positive,
as explained in Chapter 19. The last test statement draws the triangle (this plot was moved to the
right of the numeric output to save space.)
§15.3.6. *Consistency Verification
It remains to check whether the interpolation (15.15) for element displacements meets the completeness and
continuity criteria studied in Chapter 19 for finite element trial functions. Such consistency conditions are
sufficient to insure convergence toward the exact solution of the mathematical model as the mesh is refined.
The variational index for the plane stress problem is m = 1. According to the rules stated in §19.3, the trial
functions should be 1-complete, C 0 continuous, and C 1 piecewise differentiable.
§15.3.7. *Checking Continuity
Along any triangle side, the variation of u x and u y is linear and uniquely determined by the value at the nodes
on that side. For example, over side 1–2 of an individual triangle, which has equation ζ3 = 0:

The variation of u x and u y over

u x = u x1 ζ1 + u x2 ζ2 + u x3 ζ3 = u x1 ζ1 + u x2 ζ2 , side 1-2 depends only on the nodal
(15.28) values ux1, ux2 , u y1 and u y2.
u y = u y1 ζ1 + u y2 ζ2 + u y3 ζ3 = u y1 ζ1 + u y2 ζ2 .
An identical argument holds for that side when it belongs
to an adjacent triangle, such as elements (e1) and (e2) (e1)
shown in Figure 15.7. Since the node values on all 2
elements that meet at a node are the same, u x and u y match
along the side, and the trial function is C 0 interelement 1 (e2)
continuous. Because the functions are continuous inside
the elements, it follows that the continuity requirement is
met.
15–11
Section 15: THREE-NODE PLANE STRESS TRIANGLES

Figure 15.7. Interelement continuity check.

§15.3.8. *Checking Completeness

The completeness condition for variational order m = 1 requires that the shape functions Ni = ζi be able to
represent exactly any linear displacement field:

u x = α0 + α1 x + α2 y, u y = β0 + β1 x + β1 y. (15.29)

To check this we obtain the nodal values associated with the motion (15.29): u xi = α0 + α1 xi + α2 yi and
u yi = β0 + β1 xi + β2 yi for i = 1, 2, 3. Replace these in (15.16) and see if (15.29) is recovered. Here are the
detailed calculations for component u x :
  
ux = u xi ζi = (α0 + α1 xi + α2 yi )ζi = (α0 ζi + α1 xi ζi + α2 yi ζi )
  
i i i
(15.30)
= α0 ζi + α1 (xi ζi ) + α2 (yi ζi ) = α0 + α1 x + α2 y.
i i i

Component u y can be similarly verified. Consequently (15.16) satisfies the completeness requirement for
the plane stress problem — and in general, for any problem of variational index 1. Finally, a piecewise
linear trial function is obviously C 1 piecewise differentiable and consequently has finite energy. Thus the two
completeness requirements are satisfied.
Stiffness
§15.3.9. *Tonti Matrix Diagram u f
For further developments covered in more advanced f = V B TE B u = K u
courses, it is convenient to split the governing equations
of the element. In the case of the Turner triangle they are, Kinematic Equilibrium
omitting element superscripts: T
e=Bu f=VB σ
e = Bu, σ = Ee, f = AT σ = V BT σ. (15.31) Constitutive
e σ
Here V = h m A is the volume of the element, h m being the σ=Ee
mean thickness. The equations (15.31) may be graphically
Figure 15.8. Tonti matrix diagram for
represented with the diagram shown in Figure 15.8. This
Turner triangle.
is a discrete Tonti diagram similar to those of Chapter 6.

§15.4. *Derivation Using Natural Strains and Stresses

The foregoing derivation of the Turner triangle uses Carte- 3 3
sian strains and stresses, as well as {x, y} displacements. (a) (b)
The only intrinsic quantities are the triangle coordinates. ε32 = ε1 τ32 = τ1
This advanced section examines the derivation of the ε13= ε2 2
τ13= τ2 2
element stiffness matrix through natural strains, natural
stresses and covariant displacements. τ21= τ3
ε21= ε3
Although the procedure does not offer obvious shortcuts 1 1
over the previous derivation, it becomes important in
the construction of more complicated high performance Figure 15.9. Geometry-intrinsic fields for
elements. It also helps reading recent literature in assumed the Turner triangle: (a) natural strains i , (b)
natural stresses τi .
strain elements.

15–12
 §15.4 *DERIVATION USING NATURAL STRAINS AND STRESSES

(a) 3 (b) 3 (c) d6

φ1 3
L1 = L32 d5

L2 = L13
2 d3 d4
2
φ2 φ3 2

L3 = L21 d1
x 1
1 1 d2

Figure 15.10. Additional quantities appearing in natural strain and stress calculations:
(a) side lengths, (b) side directions, (c) covariant node displacements.

§15.4.1. *Natural Strains and Stresses

Natural strains are extensional strains directed parallel to the triangle sides, as shown in Figure 15.10(a).
Natural strains are denoted by 21 ≡ 3 , 32 ≡ 1 , and 13 ≡ 2 .
Similarly, natural stresses are normal stresses directed parallel to the triangle sides, as shown in Figure 15.10(b).
Natural stresses are denoted by τ21 ≡ τ3 , τ32 ≡ τ1 , and τ13 ≡ τ2 .
Because both natural stresses and strains are constant over the triangle, no node value association is needed.
The natural strains can be related to Cartesian strains by the following tensor transformation4
   
1 c12 s12 s1 c1  ex x 
= 2 =  c22 s22 s2 c2  e yy = T−1
e e. (15.32)
3 c32 s32 s3 c3 2e xy

Here c1 = x32 /L 1 , s1 = y32 /L 1 , c2 = x13 /L 2 , s2 = y13 /L 2 , c3 = x21 /L 3 , and s3 = y21 /L 3 , are sines and
cosines of the side directions with respect to {x, y}, as illustrated in Figure ?(a,b). The inverse of this relation
is
     
ex x y31 y21 L 21 y12 y32 L 22 y23 y13 L 23 1
1
e= e yy =  x31 x21 L 1 2
x12 x32 L 2 2
x23 x13 L 3 2  2 = Te .
4A2 3
2ex y (y31 x12 + x13 y21 )L 21 (y12 x23 + x21 y32 )L 22 (y23 x31 + x32 y13 )L 23
(15.33)
Note that Te is constant over the triangle. From the invariance of the strain energy density σ e = τT  it T

follows that the stresses transform as τ = Te σ and σ = T−1 e τ. That strain energy density may be expressed as

U = 12 eT Ee = 12 T En , En = TeT ETe . (15.34)

Here En is a stress-strain matrix that relates natural stresses to natural strains as τ = En . It may be therefore
called the natural constitutive matrix.

§15.4.2. *Covariant Node Displacements

Covariant node displacements di are directed along the side directions, as shown in Figure ?(c), which defines

4 This is the “straingage rosette” transformation studied in Mechanics of Materials books.

15–13
Section 15: THREE-NODE PLANE STRESS TRIANGLES

the notation used for them. They are related to the Cartesian node displacements by
d  c s3 0 0 0 0

u x1

1 3
 d2   c2 s2 0 0 0 0   u y1 
 d3   0 0  
d=     u x2  = Td u.
0 c1 s1 0
 d4  =  0 0 c3 s3 0 0   u y2 
(15.35)
    
d5 0 0 0 0 c2 s2 u x3
d6 0 0 0 0 c1 s1 u y3
The inverse relation is
u  L y L 2 y21 0 0 0 0

d1

x1 3 31
 u y1   L 3 x13 L 2 x12 0 0 0 0   d2 
 u x2  1  0 0   
u=     d3  = T−1
0 L 1 y12 L 3 y32 0
 u y2  = 2A  0 0 L 1 x21 L 3 x23 0 0   d4  d d. (15.36)
    
u x3 0 0 0 0 L 2 y23 L 1 y13 d5
u y3 0 0 0 0 L 2 x32 L 1 x31 d6
The natural strains are evidently given by the relations 1 = (d6 − d3 )/L 1 , 2 = (d2 − d5 )/L 2 and 3 =
(d4 − d1 )/L 3 . Collecting these in matrix form:
d 
1
    d2 
1 0 0 −1/L 1 0 0 1/L 1 
 d3 
= 2 = 0 1/L 2 0 0 −1/L 2 0   = B d. (15.37)
 d4 
3 −1/L 3 0 0 1/L 3 0 0  
d5
d6

§15.4.3. *The Natural Stiffness Matrix

The natural stiffness matrix for constant thickness h is
Ken = (Ah) BT En B , En = TeT E Te . (15.38)
The Cartesian stiffness matrix is
Ke = TdT Kn Td . (15.39)
Comparing with K = (Ah) B E B we see that
e T

B = T e B Td , B = T−1 −1
e BTd . (15.40)

§15.5. *The Veubeke Equilibrium Triangle

The Veubeke equilibrium triangle5 differs from the Turner triangle in the degree-of-freedom configuration.
As illustrated in Figure 15.11, those are moved to the midpoints {4, 5, 6} while the corner nodes {1, 2, 3} still
define the geometry of the element. In the FEM terminology introduced in Chapter 6, the geometric nodes
{1, 2, 3} and the connection nodes {4, 5, 6} no longer coincide. The node displacement vector collects the
freedoms shown in Figure 15.11(b):
ue = [ u x4 u y4 u x5 u y5 u x6 u y6 ]T . (15.41)
The quickest way to formulate the stiffness matrix of this element is to relate 15.41 to the node displacements
of the Turner triangle, renamed for convenience as
ueT = [ u x1 u y1 u x2 u y2 u x3 u y3 ]T . (15.42)

5 The qualifier equilibrium distinguishes this element from others created by Fraeijs de Veubeke, including the 6-node
plane stress comforming triangle. See Notes and Bibliography for the original derivation from an equilibrium field.

15–14
 §15.5 *THE VEUBEKE EQUILIBRIUM TRIANGLE

(a) 3 (x3 ,y3) (b) (c)

uy5
5 u y6 5 ux5

6 6 ux6 uy4
2 (x2 ,y2)
y
4 4 ux4
1 (x 1 ,y1)
x

Figure 15.11. The Veubeke equilibrium triangle: (a) geometric definition; (b) degree-
of-freedom configuration; (c) element patch showing how triangles are connected at the
midpoints.

§15.5.1. *Kinematic Relations

The node freedom vectors 15.41 and 15.42 are easily related since by linear interpolation along the sides one
obviously has u x4 = 12 (u x1 + u x2 ), u y4 = 12 (u y1 + u y2 ), etc. Expressing those links in matrix form gives
u  1 0 1 0 0 0

u x1
 u   1 0 −1 0 1 0 u x4
 
x4 x1
 u y4  0 1 0 1 0 0   u y1   u y1   0 1 0 −1 0 1   u y4 
 u x5  1  0 0 0    u x2   1 0 1 0 −1 0   
 =  1 0 1   u x2  ,  =   u x5  .
 u y5  2  0 0 0 1 0 1  u y2 
  u y2   0 1 0 1 0 −1  u y5 

         
u x6 1 0 0 0 1 0 u x3 u x3 −1 0 1 0 1 0 u x6
u y6 0 1 0 0 0 1 u y3 u y3 0 −1 0 1 0 1 u y6
(15.43)
In compact form: ue = TV T ueT and ueT = TT V ue , with TV T = T−1
TV . The shape functions are

N4 = ζ1 + ζ2 − ζ3 , N5 = −ζ1 + ζ2 + ζ3 , N6 = ζ1 − ζ2 + ζ3 . (15.44)

Renaming the Turner triangle strain-displacement matrix of (15.17) as BT , the corresponding matrix that
relates e = B ue in the Veubeke equilibrium triangle becomes
 
1 y21 0 y32 0 y13 0
B = BT TT V = 0 x12 0 x23 0 x31 (15.45)
A x12 y21 x23 y32 x31 y13

§15.5.2. *Stiffness Matrix

The element stiffness matrix is given by the general formula (14.23). For constant plate thickness h one obtains
the closed form
y 0 x12

21
 0 x12 y21    
h  y32 x23 
E 11 E 12 E 13 y21 0 y32 0 y13 0
Ke = A h BT E B =  0  E 12 E 22 E 23 0 x12 0 x23 0 x31 . (15.46)
A  0 x23 y32 
  E 13 E 23 E 33 x12 y21 x23 y32 x31 y13
y13 0 x31
0 x31 y13

The computation of consistent body forces is left as an Exercise.

15–15
Section 15: THREE-NODE PLANE STRESS TRIANGLES

Trig3VeubekeMembraneStiffness[ncoor_,Emat_,h_,numer_]:=Module[{
x1,x2,x3,y1,y2,y3,x12,x23,x31,y21,y32,y13,A,Be,Te,Ke},
{{x1,y1},{x2,y2},{x3,y3}}=ncoor;
A=Simplify[(x2*y3-x3*y2+(x3*y1-x1*y3)+(x1*y2-x2*y1))/2];
{x12,x23,x31,y21,y32,y13}={x1-x2,x2-x3,x3-x1,y2-y1,y3-y2,y1-y3};
Be={{y21,0,y32,0,y13,0}, {0,x12,0,x23,0,x31},
{x12,y21,x23,y32,x31,y13}}/A;
If [numer,Be=N[Be]]; Ke=A*h*Transpose[Be].Emat.Be;
Return[Ke]];

Figure 15.12. Implementation of Veubeke equilibrium triangle stiffness matrix as a Mathematica module.

§15.5.3. *Implementation
The implementation of the Veubeke equilibrium triangle as a Mathematica module that returns Ke is shown
in Figure 15.12. It needs only 8 lines of code. It is invoked as

Ke=Trig3VeubekeMembraneStiffness[ncoor,Emat,h,numer]; (15.47)

The arguments have the same meaning as those of the module Trig3TurnerMembraneStiffness described
in §15.3.6.

ncoor={{0,0},{3,1},{2,2}}; Emat=8*{{8,2,0},{2,8,0},{0,0,3}};
Ke=Trig3VeubekeMembraneStiffness[ncoor,Emat,1,False];
Print["Ke=",Ke//MatrixForm];
Print["eigs of Ke=",Chop[Eigenvalues[N[Ke]]]];

140 − 60 −4 − 28 − 136 88
− 60 300 − 12 − 84 72 − 216
−4 − 12 44 20 − 40 −8
Ke=
− 28 − 84 20 44 8 40
− 136 72 − 40 8 176 − 80
88 − 216 −8 40 − 80 176

eigs of Ke={557.318, 240., 82.6816, 0, 0, 0}

Figure 15.13. Test statements to exercise the module of Figure 15.12, and outputs.

This module is exercised by the statements listed at the top of Figure 15.13, which form a triangle with corner
coordinates { { 0,0 },{ 3,1 },{ 2,2 } }, isotropic material matrix with E 11 = E 22 = 64, E 12 = 16, E 33 = 24,
others zero, and unit thickness. The results are shown at the bottom of Figure 15.13. This is the same triangle
used to test module Trig3TurnerMembraneStiffness in §15.3.6. Note that the element is rank sufficient.
§15.5.4. *Spurious Kinematic Modes
Although an individual Veubeke equilibrium triangle is rank sufficient, assemblies are prone to the appearance
of spurious mechanisms. That is, kinematic modes that produce no strain energy although they are not rigid
body modes. These will be illustrated by studying the three macroelements pictured in Figure 15.14. For
simplicity the macroelements are of rectangular shape, but the conclusions apply to more general geometries.
Type I macroelement is built with two triangles. It has four geometric nodes: 1–4, five connection nodes: 5–9,
and 10 degrees of freedom. The eigenvalue analysis of the assembled stiffness K is given as an Exercise. It
shows that K has 4 zero eigenvalues. Since there are 3 rigid body modes in 2D, one is spurious. It is easily

15–16
 §15.6 *SHEAR LOCKING IN TURNER TRIANGLES

4 7 3
Type I 8 9 6 Spurious mode:
9
1 5 2

Thickness h/2
4 7 3 4 7 3 4 7 3
9,10
Type II 8 9 6 + 8 10 6 = 8 6
1 5 2 1 5 2 1 5 2

8 8
4 3
13 5 12 Spurious mode: 9 7
Type III 9
10 11
7

1 6 2 6

Figure 15.14. Three macroelement assemblies fabricated with Veubeke equilibrium

triangles to investigate spurious kinematic modes. Red-filled and white-filled circles
mark geometric and connection nodes, respectively.

shown that the spurious mode corresponds to the relative rotation of the two triangles with center node 9 as
pivot, as pictured to the right of the macroelement.
Type II macroelement is built with four crisscrossed triangles of thickness h/2 as illustrated in the Figure. It
has four geometric nodes: 1–4, six connection nodes: 5–10, and 12 degrees of freedom. (Note that although
9 and 10 occupy the same location for this geometry, they should be considered as two separate nodes.) The
eigenvalue analysis of the assembled stiffness K is given as an Exercise. It shows that K has 3 zero eigenvalues
and therefore this macroelement has no spurious modes.
Type III macroelement is of Union-Jack type and is built with 4 triangles. It has five geometric nodes: 1–5,
eight connection nodes: 6–13, and 16 degrees of freedom. The eigenvalue analysis of the assembled stiffness
K is given as an Exercise. It shows that K has 4 zero eigenvalues and consequently one spurious mode.
This correspond to the triangles rotating about the midpoints 6–9 as pivots, as pictured to the right of the
macroelement.
These examples show that this element, when used in a stiffness code, is prone to spurious pivot modes where
sides of adjacent triangles rotate relatively from each other about the midpoint connector. This is a consequence
of the element being nonconforming: full determination of linearly varying side displacements requires two
nodes over that side, and there is only one. Even if a rank sufficiently macroelement mesh unit such as Type II
of Figure 15.14 can be constructed, there is no guarantee that spurious pivot modes will not occur when those
mesh units are connected. For this reason this element is rarely used in DSM-based structural programs, but
acquires importance in applications where flux conservation is important.

§15.6. *Shear Locking in Turner Triangles

A well known deficiency of the 3-node Turner triangle is inability to follow rapidly varying stress fields. This
is understandable since stresses within the element, for uniform material properties, are constant. But its 1D
counterpart: the 2-node bar element, is nodally exact for displacements under some mild assumptions stated
in Chapter 11, and correctly solves loaded-at-joints trusses with one element per member. On the other hand,
the triangle can be arbitrarily way off under unhappy combinations of loads, geometry and meshing.

15–17
 ;
Section 15: THREE-NODE PLANE STRESS TRIANGLES

;
y
M 4 3 M
y z
b = a/γ b
Cross
1 2 section
x a h
z Thickness h/2
4 34 3 4 3
5
Type I: Crisscrossed + Type II: Union-Jack
1 1 2
2 1 2

Figure 15.15. The bending test with two macroelement types.

What happens in going from 1D to 2D? New effects emerge, notably shear energy and inplane bending. These
two can combine to produce shear locking: elongated triangles can become extraordinarily stiff under inplane
bending because of spurious shear energy.6 The bad news for engineers is that wrong answers caused by
locking are non-conservative: deflections and stresses can be so grossly underestimated that safety margins
are overwhelmed.
To characterize shear locking quantitatively it is convenient to use macroelements in which triangles are
combined to form a 4-node rectangle. This simplifies repetition to form regular meshes. The rectangle
response under in-plane bending is compared to that of a Bernoulli-Euler beam segment. It is well known that
the latter is exact under constant moment. The response ratio of macroelement to beam is a good measure of
triangle performance under bending. Such benchmarks are technically called higher order patch tests. Test
results can be summarized by one number: the energy ratio, which gives a scalar measure of relative stiffness.

§15.6.1. *The Inplane Bending Test

The test is defined in Figure 15.15. A Bernoulli-Euler plane beam of thin rectangular cross-section of height
b and thickness h is bent under applied end moments M. The beam is fabricated of isotropic material with
elastic modulus E and Poisson’s ratio ν. Except for possible end effects the exact solution of the beam problem
(from both the theory-of-elasticity and beam-theory standpoints) is a constant bending moment M(x) = M
along the span. The associated curvature is κ = M/(E Izz ) = 12M/(Eb3 h). The exact energy taken by a
beam segment of length a is Ubeam = 12 Mκa = 6M 2 a/(Eb3 h) = 241 Eb3 hκ 2 a = 241 Eb3 hθa2 /a. In the latter
θa = κa is the relative rotation of two cross sections separated by a.
To study the bending performance of triangles the beam is modeled with one layer of identical rectangular
macroelements dimensioned a ×b and made up of triangles, as illustrated in Figure 15.15. The rectangle aspect
ratio is γ = a/b. All rectangles undergo the same deformations and thus it is enough to study a individual
macroelement 1-2-3-4. Two types are considered here:
Crisscrossed (CC). Formed by overlaying triangles 1-2-4, 3-4-2, 2-3-1 and 4-1-2, each with thickness h/2.
Using 4 triangles instead of 2 makes the macroelement geometrically and physically symmetric since 2 triangles
are attached to each corner.
Union-Jack (UJ). Formed by placing a fifth node at the center and dividing the rectangle into 4 triangles: 1-2-5,
2-3-5, 3-4-5, 4-1-5. By construction this element is also geometrically and physically symmetric.

6 The deterioration can be even more pronounced for its spatial counterpart: the 4-node tetrahedron element, because shear
effects are even more important in three dimensions.

15–18
 §15.6 *SHEAR LOCKING IN TURNER TRIANGLES
−θa/2 θa/2
§15.6.2. *Energy Ratios
P 4 3 P
The assembled macroelement stiffnesses are KCC and
KU+ J , of orders 8 × 8 and 10 × 10, respectively. For the b = a/γ
1 2
latter the internal node 5 is statically condensed producing P P
a
an 8 × 8 stiffness KU . To test performance we apply four
alternating corner loads as shown in Figure 15.16. The Figure 15.16. Bending a macroelement by
resultant bending moment is M = Pb. applying a relative edge rotation.

Although triangles cannot copy curvatures pointwise,7 macroelement edges can rotate since constituent tri-
angles can expand or contract. Because of symmetries, the rotations of sides 1-2 and 3-4 are −θa /2 and
θa /2, as illustrated in Figure 15.16. The corresponding corner x displacements are ±bθa /4 whereas the y
displacements are zero. Assemble these into a node displacement 8-vector u M .

u M = 14 bθa [ −1 0 1 0 −1 0 1 0 ]T (15.48)

The internal energy taken by a macroelement of 8 × 8 stiffness K M under (15.48) is U M = 12 uTM K M u M , which
can be expressed as a function of E, ν, a, b, h and θa .8

ClearAll[a,b,Em,h,γ ];
b=a/γ ; Iz=h*b^3/12; Ubeam=Simplify[(1/2)*Em*Iz*θa^2/a];
Emat=Em*{{1,0,0},{0,1,0},{0,0,1/2}};
nc={{-a,-b},{a,-b},{a,b},{-a,b},{0,0}}/2;
enCC={{1,2,4},{3,4,2},{2,3,1},{4,1,3}};
enUJ={{1,2,5},{2,3,5},{3,4,5},{4,1,5}}; r={0,0};
For [m=1,m<=2,m++, mtype={"CC","UJ"}[[m]];
nF={8,10}[[m]]; K=Table[0,{nF},{nF}]; f=Table[0,{nF}];
For [e=1,e<=4,e++,
If [mtype=="CC", enl=enCC[[e]], enl=enUJ[[e]]];
{n1,n2,n3}=enl; encoor={nc[[n1]],nc[[n2]],nc[[n3]]};
ht=h; If [mtype=="CC", ht=h/2];
Ke=Trig3TurnerMembraneStiffness[encoor,Emat,ht,False];
eft={2*n1-1,2*n1,2*n2-1,2*n2,2*n3-1,2*n3};
For [i=1,i<=6,i++, For [j=1,j<=6,j++, ii=eft[[i]];
jj=eft[[j]]; K[[ii,jj]]+=Ke[[i,j]] ]];
]; KM=K=Simplify[K];
If [mtype=="UJ",
{K,f}= Simplify[CondenseLastFreedom[K,f]];
{KM,f}=Simplify[CondenseLastFreedom[K,f]]];
Print["KM=",KM//MatrixForm];
uM={1,0,-1,0,1,0,-1,0}*θa*b/4;
UM=uM.KM.uM/2; rM=Simplify[UM/Ubeam];
Print["rM=",rM]; r[[m]]=rM;
];
Plot[Evaluate[r],{γ ,0,10}];

Figure 15.17. Script to compute energy ratios for the two macroelements of Figure 15.15.

The ratio r M = U M /Ubeam is called the energy ratio. If r M > 1 the macroelement is stiffer than the beam
because it take more energy to bend it to conform to the same edge rotations, and the 2D model is said to be
overstiff. Results for zero Poisson’s ratio, computed with the script of Figure 15.17, are

3 3(1 + γ 2 )2
rCC = 3 + γ 2 , rU J = . (15.49)
2 2 + 4γ 2

7 That is the reason why they can be so stiff under bending.

8 The load P could be recovered via K M u M , but this value is not needed to compute energy ratios.

15–19
Section 15: THREE-NODE PLANE STRESS TRIANGLES

If for example γ = a/b = 10, which is an elongated rectangular shape of 10:1 aspect ratio, rCC = 153
and the crisscrossed macroelement is 153 times stiffer than the beam. For the Union-Jack configuration
rU J = 10201/134 = 76.13; about twice better but still way overstiff. If γ = 1, rCC = 4.5 and rU J = 2:
overstiff but not dramatically so. The effect of a nonzero Poisson’s ratio is studied in Exercise 15.10.
§15.6.3. *Convergence as Mesh is Refined
Note that if γ = a/b → 0, rCC → 3 and rU J → 1.5. So even if the beam of Figure 15.15 is divided into
an infinite number of macroelements along x the solution will not converge. It is necessary to subdivide also
along the height. If 2n (n ≥ 1) identical macroelement layers are placed along the beam height while γ is
kept fixed, the energy ratio becomes

22n − 1 + r (1) r (1) − 1

r (2n) = = 1 + , (15.50)
22n 22n
where r (1) is the ratio (15.49) for one layer. If r (1) = 1, r (2n) = 1 for all n ≥ 1, so bending exactness is
maintained as expected. If n = 1 (two layers), r (2) = (3+r (1) )/4 and if n = 2 (four layers), r (4) = (7+r (1) )/8.
If n → ∞, r (2n) → 1, but convergence can be slow. For example, suppose that γ = 1 (unit aspect ratio a = b)
and that r (1) = rCC = 4.5. To get within 1% of the exact solution, 1 + 3.5/22n < 1.01. This is satisfed if
n ≥ 5, meaning 10 layers of elements along y. If the beam span is 10 times the height, 1000 macroelements
or 4000 triangles are needed for this simple problem, which is exactly solvable by one beam element.
The stress accuracy of triangles is examined in Chapter 28.

Notes and Bibliography

As a plane stress structural element, the Turner triangle was first developed in the 1956 paper by Turner
et. al. [786]. The target application was modeling of delta wing skin panels. Arbitrary quadrilaterals were
formed by assembling triangles as macroelements. Because of its geometric flexibility, the element was soon
adopted in aircraft structural analysis codes in the late 1950’s. It moved to Civil Engineering applications
through the research and teaching at Berkeley of Ray Clough, who gave the method its name in [138].
The derivation method of [786] would look unfamiliar to present FEM practicioners used to the displacement
method. It was based on assumed stress modes. More precisely: the element, referred to a local Cartesian
system {x, y}, is put under three constant stress states: σx x , σ yy and σx y , collected in array σ. Lumping the
stress field to the nodes gives the node forces: f = Lσ. The strain field computed from stresses is e = E−1 σ.
This is integrated to get a deformation-displacement field, to which 3 rigid-body modes are added as integration
constants. Evaluating at the nodes produces e = Au, and the stiffness matrix follows on eliminating σ and
e: K = LEA. For constant thickness and material properties it happens that L = V AT and so K = V AT EA
happily turned out to be symmetric. This A is the B of (15.17) times 2A, so in the end the stiffness matrix (for
constant plate thickness) turns out to be the same as (15.21).
The derivation from assumed displacements evolved later. It is not clear who worked it out first, although
it is mentioned in [138,830]. The equivalence of the two forms, through energy principles, had been noted
by Gallagher [297]. Early displacement derivations typically started from linear polynomials in Cartesian
coordinates. For example Przemieniecki [619] begins with

u x = c1 x + c2 y + c3 , u y = c4 x + c5 y + c6 . (15.51)

Here the ci play the role of generalized coordinates, which have to be eventually eliminated in favor of node
displacements. The same approach is used by Clough in a widely disseminated 1965 article [140]. Even
for this simple element the approach is unnecessarily complicated and leads to long hand computations. The
elegant derivation in triangular coordinates was popularized by Argyris [28].

15–20
 §15. Notes and Bibliography

The idea of using piecewise linear interpolation over a triangular mesh actually precedes [786] by 13 years.
As noted in Chapter 1, it appears in an article by Courant [156], where it is applied to a Poisson’s equation
modeling St. Venant’s torsion. The idea did not influence early work in FEM, however, since as noted above
the derivation in [786] was not based on displacement interpolation.
The Veubeke equilibrium triangle appears in [283, p. 170] and is further elaborated in [284, p. 176]. It
is constructed there as an equilibrium element, that is, the stress field inside the triangle is assumed to be
σx x = β1 , σ yy = β2 and σx y = β3 , where {β1 , β2 , β3 } are stress parameters. (A field of constant stresses
satisfies identically the plane-stress differential equilibrium equations for zero body forces.) Stress parameters
can be uniquely expressed in terms of generalized edge loads, which turn out to be virtual-work conjugate to
midside displacements.9 The direct displacement derivation given here as a “Turner triangle mapping” is new.
As previously noted, this element is rarely used in structural mechanics because of the danger of spurious
kinematic modes discussed in §15.5.4. It has importance, however, in some non-structural applications.
The completeness check worked out in §15.4.2 is a specialization case of a general proof developed by Irons
in the mid 1960s (see [411, §3.9] and references therein) for general isoparametric elements. The check works
because the Turner triangle is isoparametric.
What are here called triangular coordinates were introduced by Möbius in his 1827 book [512].10 They are
often called barycentric coordinates on account on the interpretation discussed in [158]. Other names are
listed in Table 15.1. Triangles possess many fascinating geometric properties studied even before Euclid. An
exhaustive development can be found, in the form of solved exercises, in [711].
It is unclear when the monomial integration formula (15.26) was first derived. As an expression for integrands
expressed in triangular coordinates it was first stated in [211].
The natural strain derivation of §15.4 is patterned after that developed for the so-called ANDES (Assumed
Natural Deviatoric Strain) elements [509]. For the Turner triangle it provides nothing new aside of fancy
terminology. Energy ratios of the form used in §15.6 were introduced in [89] as a way to tune up the stiffness
of Free-Formulation elements.

References
Referenced items have been moved to Appendix R.

9 The initial step of assuming stresses exactly mimics that of [786] a decade earlier. What is fundamentally different in
Fraeijs de Veubeke’s derivation is the use of energy theorems (in this case, PVW) to pass from generalized edge loads to
mean edge displacements. The approach is characteristic of FEM Generation 2.
10 He is better remembered for the “Möbius strip” or “Möbius band,” the first one-sided 3D surface in mathematics.

15–21
Section 15: THREE-NODE PLANE STRESS TRIANGLES

Homework Exercises for Chapter 15

The Linear Plane Stress Triangle

EXERCISE 15.1 [A:15] Assume that the 3-node plane stress triangle has variable thickness defined over the
element by the linear interpolation formula
h(ζ1 , ζ2 , ζ3 ) = h 1 ζ1 + h 2 ζ2 + h 3 ζ3 , (E15.1)
where h 1 , h 2 and h 3 are the thicknesses at the corner nodes. Show that the element stiffness matrix is still
given by (15.21) but with h replaced by the mean thickness h m = (h 1 + h 2 + h 3 )/3. Hint: use (15.20) and
(15.26).
EXERCISE 15.2 [A:20] The exact integrals of triangle-coordinate monomials over a straight-sided triangle
are given by the formula (15.26), where A denotes the area of the triangle, and i, j and k are nonnegative
integers. Tabulate the right-hand side for combinations of exponents i, j and k such that i + j + k ≤ 3,
beginning with i = j = k = 0. Remember that 0! = 1. (Labor-saving hint: don’t bother repeating exponent
permutations; for example i = 2, j = 1, k = 0 and i = 1, j = 2, k = 0 are permutations of the same thing.
Hence one needs to tabulate only cases in which i ≥ j ≥ k).
EXERCISE 15.3 [A/C:20] Compute the consistent node force vector fe for body loads over a Turner triangle,
if the element thickness varies as per (E15.1), bx = 0, and b y = b y1 ζ1 + b y2 ζ2 + b y3 ζ3 . Check that for
h 1 = h 2 = h 3 = h and b y1 = b y2 = b y3 = b y you recover (15.25). For area integrals use (15.26). Partial
result: f y1 = (A/60)[b y1 (6h 1 + 2h 2 + 2h 3 ) + b y2 (2h 1 + 2h 2 + h 3 ) + b y3 (2h 1 + h 2 + 2h 3 )].
EXERCISE 15.4 [A/C:20] Derive the formula for the
3
consistent force vector fe of a Turner triangle of constant q y = q y1 (1−ζ 2 ) + q y2 ζ 2
thickness h = 1, if side 1–2 (ζ3 = 0, ζ2 = 1 − ζ1 ), is subject
to a linearly varying boundary force q = h t̂ such that qy2
qx = qx1 ζ1 + qx2 ζ2 = qx1 (1 − ζ2 ) + qx2 ζ2 , 2 qx2
(E15.2) y qy1
q y = q y1 ζ1 + q y2 ζ2 = q y1 (1 − ζ2 ) + q y2 ζ2 .
This “ line boundary force” q has dimension of force per unit
1 q x1 q x = q x1 (1−ζ 2 ) + q x2 ζ 2
of side length.
x
Procedural Hint. Use the last term of the line integral (14.21),
in which t̂ is replaced by q/ h, and show that since the Figure E15.1. Line force on triangle side 1–2
for Exercise 15.4.
contribution of sides 2-3 and 3-1 to the line integral vanish,
  1
W e = (ue )T fe = uT q d e = uT q L 21 dζ2 , (E15.3)
e 0
where L 21 is the length of side 1–2. Replace u x (ζ2 ) = u x1 (1 − ζ2 ) + u x2 ζ2 ; likewise for u y , qx and q y , integrate
and identify with the inner product shown as the second term in (E15.3). Partial result: f x1 = L 21 (2qx1 +qx2 )/6,
f x3 = f y3 = 0.
Note. The following Mathematica script solves this Exercise. If you decide to use it, explain the logic.

ClearAll[ux1,uy1,ux2,uy2,ux3,uy3,z2,L12];
ux=ux1*(1-z2)+ux2*z2; uy=uy1*(1-z2)+uy2*z2;
qx=qx1*(1-z2)+qx2*z2; qy=qy1*(1-z2)+qy2*z2;
We=Simplify[L12*Integrate[qx*ux+qy*uy,{z2,0,1}]];
fe=Table[Coefficient[We,{ux1,uy1,ux2,uy2,ux3,uy3}[[i]]],{i,1,6}];
fe=Simplify[fe]; Print["fe=",fe];

15–22
 Exercises

EXERCISE 15.5 [C+N:15] Compute the entries of Ke for the following plane stress triangle:

x1 = 0, y1 = 0, x2 = 3, y2 = 1, x3 = 2, y3 = 2,
 
100 25 0 (E15.4)
E= 25 100 0 , h = 1.
0 0 50

This may be done by hand (it is a good exercise in matrix multiplication) or (more quickly) using the script of
Figure 15.5. Partial result: K 11 = 18.75, K 66 = 118.75.

EXERCISE 15.6 [A+C:15] Show that the sum of the rows (and columns) 1, 3 and 5 of Ke as well as the sum
of rows (and columns) 2, 4 and 6 must vanish, and explain why. Check it with the foregoing script.

EXERCISE 15.7 [A:10]. Consider two triangles T and T ∗ , both with positive area. The corner coordinates
of T 1 are { { x1 , y1 },{ x2 , y2 },{ x3 , y3 } } and those of T 2 are { { x1∗ , y1∗ },{ x2∗ , y2∗ },{ x3∗ , y3∗ } }. A point P in
T has Cartesian coordinates { x, y } and triangular coordinates { ζ1 , ζ2 , ζ3 }. A point P ∗ in T ∗ has Cartesian
coordinates { x ∗ , y ∗ } and the same triangular coordinates. Show that { x ∗ , y ∗ } and { x, y } are connected by
the affine transformation
    −1  
1 1 1 1 1 1 1 1
x∗ = x1∗ x2∗ x3∗ x1 x2 x3 x (E15.5)
y∗ y1∗ y2∗ y3∗ y1 y2 y3 y

(The indicated inverse exists if T has positive area, as assumed.)

EXERCISE 15.8 [A:15]. Let point P have triangular coordinates 3

(a)
{ζ1P , ζ2P , ζ3P }, as shown in Figure E15.2. Find the distances h P1 , h P2 h3
and h P3 of P to the three triangle sides, and the triangular coordinates P1
hP1
of points P1 , P2 and P3 shown in the Figure (Pi is projection on the hP2
side opposite to corner i.) Show that h Pi = ζ Pi h i = 2ζ Pi A/L k j , P2 P(ζP1,ζP2 ,ζP3 ) 2
for i = 1, 2, 3, j = 2, 3, 1 and k = 3, 1, 2, in which L ji denotes the hP3
length of the side that joins corners i and j and h i is the distance from P3
corner i to the opposite side, as illustrated in Figure E15.2. (Note: L 21
1
the distances {h P1 , h P2 , h P3 } are called the trilinear coordinates of
a point P with respect to the vertices of the triangle. They were
Figure E15.2. Distances of arbi-
introduced by Plücker in 1835. They are essentially scaled versions trary point P to three triangle sides.
of the triangular coordinates.)

EXERCISE 15.9 [A:10]. Express the distances from the triangle centroid to the 3 sides in term of the triangle
area and the side lengths. Answer: 23 A/L 21 , 23 A/L 32 and 23 A/L 13 , where A is the area of the triangle assumed
positive and L ji is the length of side that joins corners i and j, cf. Figure E15.2, Hint: the area of each
subtriangle subtended by the centroid and two corners is 13 A.

EXERCISE 15.10 [A:20] Find the triangular coordinates of the altitude feet points H1 , H2 and H3 pictured
in Figure 15.3. Once these are obtained, find the equations of the altitudes in triangular coordinates, and the
coordinates of the orthocenter H . Answer for H3 : ζ1 = 12 + (L 213 − L 232 )/(2L 221 ), where L ji is the length of
side that joins corners i and j; cf. Figure E15.2.

EXERCISE 15.11 [C+D:20] Let p(ζ1 , ζ2 , ζ3 ) represent a polynomial expression in the natural coordinates.
The integral 
p(ζ1 , ζ2 , ζ3 ) d (E15.6)
e

15–23
Section 15: THREE-NODE PLANE STRESS TRIANGLES

over a straight-sided triangle can be computed symbolically by the following Mathematica module:

IntegrateOverTriangle[expr_,tcoord_,A_,max_]:=Module [{p,i,j,k,z1,z2,z3,c,s=0},
p=Expand[expr]; {z1,z2,z3}=tcoord;
For [i=0,i<=max,i++, For [j=0,j<=max,j++, For [k=0,k<=max,k++,
c=Coefficient[Coefficient[Coefficient[p,z1,i],z2,j],z3,k];
s+=2*c*(i!*j!*k!)/((i+j+k+2)!);
]]];
Return[Simplify[A*s]] ];

This is referenced as int=IntegrateOverTriangle[p,{ z1,z2,z3 },A,max]. Here p is the polynomial to

be integrated, z1, z2 and z3 denote the symbols used for the triangular coordinates, A is the triangle area and max
the highest exponent appearing in a triangular coordinate. The module name returns the integral. For example,
if p=16+5*b*z2^2+z1^3+z2*z3*(z2+z3) the call int=IntegrateOverTriangle[p,{ z1,z2,z3 },A,3]
returns int=A*(97+5*b)/6. Explain how the module works.

EXERCISE 15.12 [C+D:25] Explain the logic of the script listed in Figure 15.17. Then extend it to account
for isotropic material with arbitrary Poisson’s ratio ν. Obtain the macroelement energy ratios as functions of
γ and ν. Discuss whether the effect of a nonzero ν makes much of a difference if γ >> 1.

EXERCISE 15.13 [A/C:25] Verify the conclusions of §15.5.4 as regards rank sufficiency or deficiency of the
three Veubeke macroelement assemblies pictured in Figure 15.14. Carry out tests with rectangular macroele-
ments dimensioned a × b, constant thickness h, elastic modulus E and Poisson’s ratio 0.

EXERCISE 15.14 [C+D:25] To find whether shear is the guilty party in the poor performance of elongated
triangles (as alledged in §15.6) run the script of Figure 15.17 with a zero shear modulus. This can be done by
setting Emat=Em*{ { 1,0,0 },{ 0,1,0 },{ 0,0,0 } } in the third line. Discuss the result. Can Em be subsequently
reduced to a smaller (fictitious) value so that r ≡ 1 for all aspect ratios γ ? Is this practical?

HomogenizedLinTrigCoorFunction[expr_,{ζ 1_,ζ 2_,ζ 3_}]:=Module[

{f=expr,repζ 0,C0}, repζ 0={ζ 1->0,ζ 2->0,ζ 3->0};
C0=Simplify[f/.repζ 0]; f=Simplify[f-C0(1-ζ 1-ζ 2-ζ 3)];
Return[f]];

HomogenizedQuadTrigCoorFunction[expr_,{ζ 1_,ζ 2_,ζ 3_}]:=Module[

{f,repζ 0,C0,C1,C2,C3}, repζ 0={ζ 1->0,ζ 2->0,ζ 3->0};
f=HomogenizedLinTrigCoorFunction[expr,{ζ 1,ζ 2,ζ 3}];
C1=Coefficient[f,ζ 1]/.repζ 0; C2=Coefficient[f,ζ 2]/.repζ 0;
C3=Coefficient[f,ζ 3]/.repζ 0; {C1,C2,C3}=Simplify[{C1,C2,C3}];
f=Simplify[Expand[f-(C1*ζ 1+C2*ζ 2+C3*ζ 3)(1-ζ 1-ζ 2-ζ 3)]];
Return[f]];

Figure E15.3. Two Mathematica modules that homogenize linear and quadratic polynomials
expressed in triaangular coordinates.

EXERCISE 15.15 [C:15] The two Mathematica modules listed in Figure E15.3 homogenize linear and
quadratic polynomials, respectively, expressed in triangular coordinates. Explain their logic.

15–24
 Exercises

EXERCISE 15.16 [C+D:25] Access the file Trig3PlaneStress.nb from the course Web site by clicking
on the appropriate link in Chapter 15 Index. This is a Mathematica Notebook that does plane stress FEM
analysis using the 3-node Turner triangle.
Download the Notebook into your directory. Load into Mathematica. Execute the top 7 input cells (which are
actually initialization cells) so the necessary modules are compiled. Each cell is preceded by a short comment
cell which outlines the purpose of the modules it holds. Notes: (1) the plot-module cell may take a while
to run through its tests; be patient; (2) to get rid of unsightly messages and silly beeps about similar names,
initialize each cell twice.
After you are satisfied everything works fine, run the cantilever beam problem, which is defined in the last
input cell.
After you get a feel of how this code operate, study the source. Prepare a hierarchical diagram of the modules,11
beginning with the main program of the last cell. Note which calls what, and briefly explain the purpose of
each module. Return this diagram as answer to the homework. You do not need to talk about the actual run
and results; those will be discussed in Part III.
Hint: a hierarchical diagram for Trig3PlaneStress.nb begins like

Main program in Cell 8 - drives the FEM analysis

GenerateNodes - generates node coordinates of regular mesh
GenerateTriangles - generate element node lists of regular mesh
........

EXERCISE 15.17 [A:10] Consider the Veubeke triangle with 3 midside nodes 4, 5 and 6. Show that three
possible shape functions are 1 − 2ζ3 , 1 − 2ζ1 and 1 − 2ζ2 , respectively. Show that these functions satisfy the
interpolation and completeness conditions, but fail the compatibility condition.

11 A hierarchical diagram is a list of modules and their purposes, with indentation to show dependence, similar to the table
of contents of a book. For example, if module AAAA calls BBBB and CCCC, and BBBB calld DDDD, the hierarchical diagram
may look like:

AAAA - purpose of AAAA

BBBB - purpose of BBBB
DDDD - purpose of DDDD
CCCC - purpose of CCCC

15–25
Section 15: THREE-NODE PLANE STRESS TRIANGLES

Hint on Exercise 15.3 (added October 19, 2011)

If doing this Exercise by hand, you should process as follows.
First, multiply NT by b: ζ 
1 0
0 ζ1 
 ζ2  
0
NT .b =   0
0 ζ2  b y1 ζ1 + b y2 ζ2 + b y3 ζ3
 
ζ3 0
0 ζ3
to get a 6-vector. Entries 1,3 and 5 are zero. Entry 2 is (b y1 ζ1 + b y2 ζ2 + b y3 ζ3 )ζ1 , and so on for entries 4 and
6. Next, scale this vector by h = h 1 ζ1 + h 2 ζ2 + h 3 ζ3 . Entries 1,3 and 5 remain zero, whereas entries 2, 4 and
6 become cubic polynomials in the ζi . For example, the second entry is

(h 1 ζ1 + h 2 ζ2 + h 3 ζ3 ) (b y1 ζ1 + b y2 ζ2 + b y3 ζ3 )ζ1

Expand these in term of cubic monomials. For example, the expanded second entry becomes

h 1 b y1 ζ 3 + h 1 b y2 ζ12 ζ2 + 7 more terms

Next, collect the ζi monomials that appear in entries 2, 4 and 6. The 10 possible monomials are ζ13 , ζ23 , ζ33 ,
ζ12 ζ2 , ζ12 ζ3 , ζ22 ζ1 , ζ22 ζ3 , ζ32 ζ1 , ζ32 ζ2 , and ζ1 ζ2 ζ3 . Move all monomial coefficients such as b y1 h 1 , etc., outside the
area integral, and apply the formula (15.26) to the monomial integrals. Three cases:
  
A
ζ13 d = ζ23 d = ζ33 d =
e e e 10
  
A
ζ12 ζ2 d = ζ12 ζ3 d = ζ22 ζ1 d = . . . =
e e e 30

A
ζ1 ζ2 ζ3 =
e 60
Finally, collect the common factor A, collect the h factors of the b yi as in (E15.2) and you are done. Well, not
quite. It is instructive to check your results for the special cases h 1 = h 2 = h 3 = h (constant thickness), and
b y1 = b y2 = b y3 = b y (constant body force). If both the thickness h and the body force b y are constant, the
total force on the element, which is then b y h A, should divide equally in 3 for each node. This would agree
with the element-by-element force lumping recipe of Section 7).
If you are good in Mathematica, the result can be obtained in milliseconds, but you need to use the module
listed under Exercise 15.11.

15–26
 16
The Isoparametric
Representation

16–1
Chapter 16: THE ISOPARAMETRIC REPRESENTATION

TABLE OF CONTENTS

Page
§16.1. Introduction 16–3
§16.2. Isoparametric Representation 16–3
§16.2.1. Motivation . . . . . . . . . . . . . . . . . . 16–3
§16.2.2. Equalizing Geometry and Displacements . . . . . . . . 16–4
§16.3. General Isoparametric Formulation 16–5
§16.4. Triangular Elements 16–5
§16.4.1. The Linear Triangle . . . . . . . . . . . . . . . 16–6
§16.4.2. The Quadratic Triangle . . . . . . . . . . . . . 16–6
§16.4.3. *The Cubic Triangle . . . . . . . . . . . . . . . 16–6
§16.5. Quadrilateral Elements 16–6
§16.5.1. Quadrilateral Coordinates and Iso-P Mappings . . . . . . 16–6
§16.5.2. The Bilinear Quadrilateral . . . . . . . . . . . . . 16–7
§16.5.3. The Biquadratic Quadrilateral . . . . . . . . . . . 16–7
§16.6. Completeness Properties of Iso-P Elements 16–8
§16.6.1. *Completeness Analysis . . . . . . . . . . . . . . 16–8
§16.6.2. Completeness Checks . . . . . . . . . . . . . . 16–9
§16.6.3. *Completeness for Higher Variational Index . . . . . . . 16–11
§16.7. Iso-P Elements in One and Three Dimensions 16–11
§16. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 16–11
§16. References . . . . . . . . . . . . . . . . . . . . . . 16–11
§16. Exercises . . . . . . . . . . . . . . . . . . . . . . 16–12

16–2
 §16.2 ISOPARAMETRIC REPRESENTATION

§16.1. Introduction
The procedure used in Chapter 15to formulate the stiffness equations of the linear triangle can
be formally extended to quadrilateral elements as well as higher order triangles. But one quickly
encounters technical difficulties:
1. The construction of shape functions that satisfy consistency requirements for higher order
elements with curved boundaries becomes increasingly complicated.
2. Integrals that appear in the expressions of the element stiffness matrix and consistent nodal
force vector can no longer be evaluated in simple closed form.
These two obstacles can be overcome through the concepts of isoparametric elements and numerical
quadrature, respectively. The combination of these two ideas transformed the field of finite element
methods in the late 1960s. Together they support a good portion of what is presently used in
production finite element programs.
In the present Chapter the concept of isoparametric representation is introduced for two dimen-
sional elements. This representation is illustrated on specific elements. In the next Chapter these
techniques, combined with numerical integration, are applied to quadrilateral elements.

§16.2. Isoparametric Representation

§16.2.1. Motivation
The linear triangle presented in Chapter 115is an isoparametric element although was not originally
derived as such. The two key equations are (15.10), which defines the triangle geometry, and
(15.16), which defines the primary variable, in this case the displacement field. These equations
are reproduced here for convenience:
    
1 1 1 1 ζ1
x = x1 x2 x3 ζ2 , (16.1)
y y1 y2 y3 ζ3
u x = u x1 N1e + u x2 N2e + u x3 N3e = u x1 ζ1 + u x2 ζ2 + u x3 ζ3 ,
(16.2)
u y = u y1 N1e + u y2 N2e + u y3 N3e = u y1 ζ1 + u y2 ζ2 + u y3 ζ3 .
The interpretation of these equations is as follows.
The triangular coordinates define the element geom-
etry via (16.1). The displacement expansion (16.2) Triangular
Geometry
coordinates
is defined by the shape functions, which are in turn 1, x, y
ζ1 , ζ 2 , ζ 3
expressed in terms of the triangular coordinates. For
the linear triangle, shape functions and triangular
coordinates coalesce.
These relations are diagrammed in Figure 16.1. Displacement
Shape
Evidently geometry and displacements are not functions interpolation
treated equally. If we proceed to higher order Ni(e) ux , uy
triangular elements while keeping straight sides,
Figure 16.1. Superparametric rep-
only the displacement expansion is refined whereas
resentation of triangular element.
the geometry definition remains the same.

16–3
Chapter 16: THE ISOPARAMETRIC REPRESENTATION

Geometry
1, x, y

Triangular Shape
coordinates functions
ζ1 , ζ 2 , ζ 3 Ni(e)

Displacement
interpolation
ux , uy

Figure 16.2. Isoparametric representation of triangular elements.

Elements built according to the foregoing prescription are called superparametric, a term that
emphasizes that unequal treatment.

§16.2.2. Equalizing Geometry and Displacements

On first inspection (16.2) and (16.1) do not look alike. Their inherent similarity can be displayed,
however, if the second one is rewritten and adjoined to (16.1) to look as follows:
     
1 1 1 1 1 1 1
   
 x   x1 x2 x3  ζ1  x1 x2 x3  N1e
     
 y  =  y1 y2 y3  ζ2 =  y1 y2 y3  N2e . (16.3)
     
ux u x1 u x2 u y3 ζ3 u x1 u x2 u y3 N3e
uy u y1 u y2 u y3 u y1 u y2 u y3
This form emphasizes that geometry and displacements are given by the same parametric represen-
tation, as shown in Figure 16.2.
The key idea is to use the shape functions to represent both the element geometry and the problem
unknowns, which in structural mechanics are displacements. Hence the name isoparametric element
(“iso” means equal), often abbreviated to iso-P element. This property may be generalized to
arbitrary elements by replacing the term “triangular coordinates” by the more general one “natural
coordinates.” This generalization is illustrated in Figure 16.3.

Geometry
1, x, y

Shape
Natural
functions
coordinates
Ni(e)

Displacement
interpolation
ux , uy

Figure 16.3. Isoparametric representation of arbitrary two-dimensional

elements: triangles or quadrilaterals. For 3D elements, expand the geometry
list to {1, x, y, z} and the displacements to {u x , u y , u z }.

16–4
 §16.4 TRIANGULAR ELEMENTS

Under this generalization, natural coordinates (triangular coordinates for triangles, quadrilateral
coordinates for quadrilaterals) appear as parameters that define the shape functions. The shape
functions connect the geometry with the displacements.
Remark 16.1. The terms isoparametric and superparametric were introduced by Irons and coworkers at
Swansea in 1966. See Notes and Bibliography at the end of this Chapter. There are also subparametric
elements whose geometry is more refined than the displacement expansion.

§16.3. General Isoparametric Formulation

The generalization of (16.3) to an arbitrary two-dimensional element with n nodes is straightforward.
Two set of relations, one for the element geometry and the other for the element displacements, are
required. Both sets exhibit the same interpolation in terms of the shape functions.
Geometric relations:

n
1= Nie , x= xi Nie , y= yi Nie . (16.4)
i=1 i=1 i=1

Displacement interpolation:

n
ux = u xi Nie , uy = u yi Nie . (16.5)
i=1 i=1

These two sets of equations may be combined in matrix form as

     N1e 
1 1 1 ... 1
 x   x1 x2 . . . xn   
N2e 
  
y2 . . . yn  
 .. 

 y  =  y1 . . (16.6)
    
ux u x1 u x2 . . . u xn
uy u y1 u y2 . . . u yn Ne n
The first three scalar equations in (16.6) express the geometry definition, and the last two the
displacement expansion. Note that additional rows may be added to this matrix expression if more
variables are interpolated by the same shape functions. For example, suppose that the thickness h
and a temperature field T are both interpolated from the n node values:
   
1 1 1 ... 1  e
 x   x1 x2 . . . xn  N1e
    N2 
 y   y1 y2 . . . yn  
     .. 
 u x  =  u x1 u x2 . . . u xn   
. . (16.7)
    
 u y   u y1 u y2 . . . u yn 
   
h h1 h2 . . . hn Nne
T T1 T2 . . . Tn
Note that the column of shape functions does not change.
To illustrate the use of the isoparametric concept, we take a look at specific 2D isoparametric
elements that are commonly used in structural and non-structural applications. These are separated
into triangles and quadrilaterals because different natural coordinates are used.

16–5
Chapter 16: THE ISOPARAMETRIC REPRESENTATION

§16.4. Triangular Elements

§16.4.1. The Linear Triangle
The three-noded linear triangle, studied in Chapter 15 and 3
pictured in Figure 16.4, may be presented as an isoparametric
element:
   
1 1 1 1
  2
 x   x1 x2 x3  N1e
   
 y  =  y1 y2 y3  N2e . (16.8)
   
ux u x1 u x2 u x3 N3e 1
uy u y1 u y2 u y3 Figure 16.4. The 3-node linear triangle.

The shape functions are simply the triangular coordinates:

N1e = ζ1 , N2e = ζ2 , N3e = ζ3 . (16.9)

The linear triangle is the only triangular element that is both superparametric and isoparametric.

§16.4.2. The Quadratic Triangle

The six node triangle shown in Figure 16.5 is the next (b)
(a) 3 3
complete-polynomial member of the isoparametric
triangle family. The isoparametric definition is
5 5
 e 6
    N1 6
1 1 1 1 1 1 1  Ne 
 x   x1 x2 x3 x4 x5 x6  
2  2 2
    N3e 
 y  =  y1 y2 y3 y4 y5 y6    4 4
    N4e  1
u x1 u x2 u x3 u x4 u x5 u x6   1
ux  N5e 
uy u y1 u y2 u y3 u y4 u y5 u y6 N6e Figure 16.5. The 6-node quadratic triangle:
(a) the superparametric version, with straight
(16.10) sides and midside nodes at midpoints; (b) the
isoparametric version.

The shape functions are

N1e = ζ1 (2ζ1 − 1), N2e = ζ2 (2ζ2 − 1), N3e = ζ3 (2ζ3 − 1),

(16.11)
N4e = 4ζ1 ζ2 , N5e = 4ζ2 ζ3 , N6e = 4ζ3 ζ1 .

The element may have parabolically curved sides defined by the location of the midnodes 4, 5 and 6.
The triangular coordinates for a curved triangle are no longer straight lines, but form a curvilinear
system as can be observed in Figure 16.5(b).

§16.4.3. *The Cubic Triangle

The cubic triangle has ten nodes. This shape functions of this element are the subject of an Exercise in Chapter
18. The implementation is studied in Chapter 24.

16–6
 §16.5 QUADRILATERAL ELEMENTS

§16.5. Quadrilateral Elements

§16.5.1. Quadrilateral Coordinates and Iso-P Mappings

Before presenting examples of quadrilateral

η η
elements, we must introduce the appropriate
natural coordinate system for that geometry. η=1 η=1
The natural coordinates for a triangular ξ=−1 ξ=1
element are the triangular coordinates ζ1 , ξ ξ
ζ2 and ζ3 . The natural coordinates for a
quadrilateral element are ξ and η, which are ξ=−1 ξ=1
illustrated in Figure 16.6 for both straight η=−1 η=−1
sided and curved side quadrilaterals. These
are called quadrilateral coordinates.
Figure 16.6. Quadrilateral coordinates.

These coordinates vary from −1 on one side to +1 at the other, taking the value zero over the
quadrilateral medians. This particular variation range (instead of taking, say, 0 to 1) was chosen by
Irons and coworkers to facilitate use of the standard Gauss integration formulas. Those formulas
are discussed in the next Chapter.

Remark 16.2. In some FEM derivations it is convenient to visualize the quadrilateral coordinates plotted as
Cartesian coordinates in the {ξ, η} plane. This is called the reference plane. All quadrilateral elements in the
reference plane become a square of side 2, called the reference element, which extends over ξ ∈ [−1, 1], η ∈
[−1, 1]. The transformation between {ξ, η} and {x, y} dictated by the second and third equations of (16.4),
is called the isoparametric mapping. A similar version exists for triangles. An important application of this
mapping is discussed in §16.6; see Figure 16.9 there.

§16.5.2. The Bilinear Quadrilateral η

3
The four-node quadrilateral shown in Figure 16.7 is the 4 η=1
simplest member of the quadrilateral family. It is defined by ξ=1
ξ=−1
    ξ
1 1 1 1 1  e
N1
 x   x1 x2 x3 x4 
     Ne  1
 y  =  y1 y2 y3 y4   2e  . (16.12) η=−1
    N3 2
ux u x1 u x2 u x3 u x4
N4e
uy u y1 u y2 u y3 u y4 Figure 16.7. The 4-node
bilinear quadrilateral.

The shape functions are

N1e = 14 (1 − ξ )(1 − η), N2e = 14 (1 + ξ )(1 − η),

(16.13)
N3e = 14 (1 + ξ )(1 + η), N4e = 14 (1 − ξ )(1 + η).

These functions vary linearly on quadrilateral coordinate lines ξ = const and η = const, but are
not linear polynomials as in the case of the three-node triangle.

16–7
Chapter 16: THE ISOPARAMETRIC REPRESENTATION

η η
(a) 3 (b) 3
4 η=1 7 η=1 7
4

9 6 ξ 6
8 8 ξ
ξ=−1 ξ=1 ξ=−1 ξ=1
1 η=−1 5 η=−1 5
1
2 2

Figure 16.8. Two widely used higher order quadrilaterals: (a) the nine-node biquadratic
quadrilateral; (b) the eight-node “serendipity” quadrilateral.

§16.5.3. The Biquadratic Quadrilateral

The nine-node quadrilateral shown in Figure 16.8(a) is the next complete member of the quadrilateral
family. It has eight external nodes and one internal node. It is defined by
     N1e 
1 1 1 1 1 1 1 1 1 1
 x   x1 x2 x3 x4 x5 x6 x7 x8 x9   
N2e 
   
y9   .. 

 y  =  y1 y2 y3 y4 y5 y6 y7 y8 . . (16.14)
   
ux u x1 u x2 u x3 u x4 u x5 u x6 u x7 u x8 u x9  
uy u y1 u y2 u y3 u y4 u y5 u y6 u y7 u y8 u y9 Ne 9

This element is often referred to as the Lagrangian quadrilateral in the FEM literature, a term
explained in the Notes and Bibliography. Its shape functions are

N1e = 14 (1 − ξ )(1 − η)ξ η, N5e = − 12 (1 − ξ 2 )(1 − η)η,

N2e = − 14 (1 + ξ )(1 − η)ξ η, N6e = 12 (1 + ξ )(1 − η2 )ξ, N9e = (1 − ξ 2 )(1 − η2 ) (16.15)
··· ···

These functions vary quadratically along the coordinate lines ξ = const and η = const. The shape
function associated with the internal node 9 is called a bubble function because of its geometric
shape, which is pictured in §18.4.2.
Figure 16.8(a) depicts a widely used eight-node variant called the “serendipity” quadrilateral. (A
name that originated from circumstances surrounding the element discovery.) The internal node is
eliminated by kinematic constraints as worked out in an Exercise of Chapter 18.

§16.6. Completeness Properties of Iso-P Elements

Some general conclusions as regards the range of applications of isoparametric elements can be
obtained from a completeness analysis. More specifically, whether the general prescription (16.6)
that combines (16.4) and (16.5) satisfies the completeness criterion of finite element trial expansions.
This is one of the conditions for convergence to the analytical solution. The requirement is treated
generally in Chapter 19, and is stated here in recipe form.

16–8
 §16.6 COMPLETENESS PROPERTIES OF ISO-P ELEMENTS

§16.6.1. *Completeness Analysis

The plane stress problem has variational index m = 1. A set of shape functions is complete for this problem
if they can represent exactly any linear displacement motions such as

u x = α0 + α1 x + α2 y, u y = β0 + β1 x + β2 y. (16.16)

To carry out the check, evaluate (16.16) at the nodes

u xi = α0 + α1 xi + α2 yi u yi = β0 + β1 xi + β2 yi , i = 1, . . . n. (16.17)

Insert this into the displacement expansion (16.5) to see whether the linear displacement field (16.16) is
recovered. Here are the computations for the displacement component u x :

ux = (α0 + α1 xi + α2 yi ) Nie = α0 Nie + α1 xi Nie + α2 yi Nie = α0 + α1 x + α2 y. (16.18)

i=1 i i i

For the last step we have used the geometry definition relations (16.4), reproduced here for convenience:

n
1= Nie , x= xi Nie , y= yi Nie . (16.19)
i=1 i=1 i=1

A similar calculation may be made for u y . It appears that the isoparametric displacement expansion represents
(16.18) for any element, and consequently meets the completeness requirement for variational order m = 1.
The derivation carries without essential change to three dimensions.1
Can you detect a flaw in this conclusion? The fly in the ointment is the last replacement step of (16.18),
which assumes that the geometry relations (16.19) are identically satisfied. Indeed they are for all the example
elements presented in the previous sections. But if the new shape functions are constructed directly by the
methods of Chapter 18, a posteriori checks of those identities are necessary.

§16.6.2. Completeness Checks

The first check in (16.19) is easy: the sum of shape functions must be unity. This is also called the
unit sum condition. It can be easily verified by hand for simple elements. Here are two examples.

Example 16.1. Check for the linear triangle: directly from the definition of triangular coordinates,

N1e + N2e + N3e = ζ1 + ζ2 + ζ3 = 1. (16.20)

1 This derivation is due to B. M. Irons. See for example [397, p. 75]. The property was known since the mid 1960s and
contributed substantially to the rapid acceptance of iso-P elements.

16–9
Chapter 16: THE ISOPARAMETRIC REPRESENTATION

ξ−η plane good mapping x−y plane

bad mapping
(compatible) (incompatible)
η

2 (e1) (e1)
(e1) ξ

1 2 1 1
η 2 2
2 (e2) (e2)
y
ξ
(e2)

2 x

Figure 16.9. Good and bad isoparametric mappings of 4-node quadrilateral

from the {ξ, η} reference plane onto the {x, y} physical plane.

Example 16.2. Check for the 4-node bilinear quadrilateral:

N1e + N2e + N3e + N4e = 14 (1 − ξ − η + ξ η) + 14 (1 + ξ − η − ξ η)

(16.21)
+ 14 (1 + ξ + η + ξ η) + 14 (1 − ξ + η − ξ η) = 1

For more complicated elements see Exercises 16.2 and 16.3.

The other two checks are less obvious. For specificity consider the 4-node bilinear quadrilateral.
The geometry definition equations are

4
x= xi Nie (ξ, η), y= yi Nie (ξ, η). (16.22)
i=1 i=1

Given the corner coordinates, {xi , yi } and a point P(x, y) one can try to solve for {ξ, η}. This
solution requires nontrivial work because it involves two coupled quadratics, but can be done.
Reinserting into (16.22) simply gives back x and y, and nothing is gained.2
The correct question to pose is: is the correct geometry of the quadrilateral preserved by the
mapping from {ξ, η} to {x, y}? In particular, are the sides straight lines? Figure 16.9 illustrate
these questions. Two side-two squares: (e1) and (e2), contiguous in the {ξ, η} reference plane, are
mapped to quadrilaterals (e1) and (e2) in the {x, y} physical plane through (16.22). The common
side 1-2 must remain a straight line to preclude interelement gaps or interpenetration.
We are therefore lead to consider geometric compatibility upon mapping. But this is equivalent to the
question of interelement displacement compatibility, which is stipulated as item (C) in §18.1. The
statement “the displacement along a side must be uniquely determined by nodal displacements on
that side” translates to “the coordinates of a side must be uniquely determined by nodal coordinates
on that side.” Summarizing:

2 This tautology is actually a blessing, since finding explicit expressions for the natural coordinates in terms of x and y
rapidly becomes impossible for higher order elements. See, for example, the complications that already arise for the
bilinear quadrilateral in §23.3.

16–10
 §16. References

Unit-sum condition + interelement compatibility → completeness. (16.23)

This subdivision of work significantly reduces the labor involved in element testing.
§16.6.3. *Completeness for Higher Variational Index
The completeness conditions for variational index 2 are far more demanding because they involve quadratic
motions. No simple isoparametric configurations satisfy those conditions. Consequently isoparametric for-
mulations have limited importance in the finite element analysis of plate and shell bending.

§16.7. Iso-P Elements in One and Three Dimensions

The reader should not think that the concept of isoparametric representation is confined to two-
dimensional elements. It applies without conceptual changes to one and three dimensions as long as
the variational index remains one.3 Three-dimensional solid elements are covered in an advanced
course. The use of the isoparametric formulation to construct a 3-node bar element is the topic of
Exercises 16.4 through 16.7.
Notes and Bibliography
A detailed presentation of the isoparametric concept, with annotated references to the original 1960 papers
may be found in the textbook [397].
This matrix representation for isoparametric elements used here was introduced in [204].
The term Lagrangian element in the mathematical FEM literature identifies quadrilateral and hexahedra (brick)
elements that include all polynomial terms ξ i η j (in 2D) or ξ i η j µk (in 3D) with i ≤ n, j ≤ n and k ≤ n,
as part of the shape function interpolation. Such elements have (n + 1)2 nodes in 2D and (n + 1)3 nodes in
3D, and the interpolation is said to be n-bicomplete. For example, if n = 2, the biquadratic quadrilateral with
(2 + 1)2 = 9 nodes is Lagrangian and 2-bicomplete. (The qualifier “Lagrangian” in this context refers to
Lagrange’s interpolation formula, not to Lagrange multipliers.)

References
Referenced items have been moved to Appendix R

3 A limitation explained in §16.6.3.

16–11
Chapter 16: THE ISOPARAMETRIC REPRESENTATION

Homework Exercises for Chapter 16

The Isoparametric Representation

EXERCISE 16.1 [D:10] What is the physical interpretation of the shape-function unit-sum condition discussed
in §16.6? Hint: the element must respond exactly in terms of displacements to rigid-body translations in the
x and y directions.
EXERCISE 16.2 [A:15] Check by algebra that the sum of the shape functions for the six-node quadratic
triangle (16.11) is exactly one regardless of natural coordinates values. Hint: show that the sum is expressable
as 2S12 − S1 , where S1 = ζ1 + ζ2 + ζ3 .
EXERCISE 16.3 [A/C:15] Complete the table of shape functions (16.23) of the nine-node biquadratic quadri-
lateral. Verify that their sum is exactly one.
EXERCISE 16.4 [A:20] Consider a three-node bar element referred to the natural coordinate ξ . The two end
nodes and the midnode are identified as 1, 2 and 3, respectively. The natural coordinates of nodes 1, 2 and 3
are ξ = −1, ξ = 1 and ξ = 0, respectively. The variation of the shape functions N1 (ξ ), N2 (ξ ) and N3 (ξ ) is
sketched in Figure E16.1. These functions must be quadratic polynomials in ξ :
N1e (ξ ) = a0 + a1 ξ + a2 ξ 2 , N2e (ξ ) = b0 + b1 ξ + b2 ξ 2 , N3e (ξ ) = c0 + c1 ξ + c2 ξ 2 . (E16.1)

N2 (ξ)
e N 3e (ξ)
N1 (ξ)
e
1 1 1
3 2 1 3 2 1 3 2
1 ξ=0
ξ=−1 ξ=0 ξ=1 ξ=−1 ξ=0 ξ=1 ξ=−1 ξ=1

Figure E16.1. Isoparametric shape functions for 3-node bar element (sketch). Node 3 has been
drawn at the 1–2 midpoint but it may be moved away from it, as in Exercises E16.5 and E16.6.

Determine the coefficients a0 , through c2 using the node value conditions depicted in Figure E16.1; for example
N1e = 1, 0 and 0 for ξ = −1, 0 and 1 at nodes 1, 3 and 2, respectively. Proceeding this way show that
N1e (ξ ) = − 12 ξ(1 − ξ ), N2e (ξ ) = 12 ξ(1 + ξ ), N3e (ξ ) = 1 − ξ 2 . (E16.2)
Verify that their sum is identically one.
EXERCISE 16.5
[A/C:15+10+15+5] A 3-node straight bar element is defined by 3 nodes: 1, 2 and 3, with axial coordinates
x1 , x2 and x3 , respectively, as illustrated in Figure E16.2. The element has axial rigidity E A and length
 = x2 − x1 . The axial displacement is u(x). The 3 degrees of freedom are the axial node displacements u 1 ,
u 2 and u 3 . The isoparametric definition of the element is
    
1 1 1 1 N1e
x = x1 x2 x3 N2e , (E16.3)
u u1 u2 u3 N3e
in which Nie (ξ ) are the shape functions (E16.2) of the previous Exercise. Node 3 lies between 1 and 2 but is
not necessarily at the midpoint x = 12 . For convenience define
x1 = 0, x2 = , x3 = ( 12 + α), (E16.4)

16–12
 Exercises

axial rigidity EA
x, u
1 (ξ= −1) 3 (ξ=0) 2 (ξ=1)
x1 = 0 x 3 = /2+α x2 = 
 = L (e)

Figure E16.2. The 3-node bar element in its local system.

where − 12 < α < 12 characterizes the location of node 3 with respect to the element center. If α = 0 node 3
is located at the midpoint between 1 and 2. See Figure E16.2.
(a) From (E16.4) and the second equation of (E16.3) get the Jacobian J = d x/dξ in terms of , α and ξ .
Show that: (i) if − 14 < α < 14 then J > 0 over the whole element −1 ≤ ξ ≤ 1; (ii) if α = 0, J = /2
is constant over the element.
(b) Obtain the 1 × 3 strain-displacement matrix B relating e = du/d x = B ue , where ue is the column
3-vector of node displacements u 1 , u 2 and u 3 . The entries of B are functions of , α and ξ . Hint:
B = dN/d x = J −1 dN/dξ , where N = [ N1 N2 N3 ] and J comes from item (a).
(c) Show that the element stiffness matrix is given by
   1
K =
e
E A B B dx =
T
E A BT B J dξ. (E16.5)
0 −1

Evaluate the rightmost integral for arbitrary α but constant E A using the 2-point Gauss quadrature rule
(E13.7). Specialize the result to α = 0, for which you should get K 11 = K 22 = 7E A/(3), K 33 =
16E A/(3), K 12 = E A/(3) and K 13 = K 23 = −8E A/(3), with eigenvalues {8E A/, 2E A/, 0}.
Note: use of a CAS is recommended for this item to save time.
(d) What is the minimum number of Gauss points needed to integrate Ke exactly if α = 0?

EXERCISE 16.6 [A/C:20] This Exercise is a continuation of the foregoing one, and addresses the question of
why Ke was computed by numerical integration in item (c). Why not use exact integration? The answer is that
the exact stiffness for arbitrary α is numerically useless. To see why, try the following script in Mathematica:

ClearAll[EA,L,alpha,xi]; (* Define J and B={{B1,B2,B3}} here *)

Ke=Simplify[Integrate[EA*Transpose[B].B*J,{xi,-1,1},
Assumptions->alpha>0&&alpha<1/4&&EA>0&&L>0]];
Print["exact Ke=",Ke//MatrixForm];
Print["exact Ke for alpha=0",Simplify[Ke/.alpha->0]//MatrixForm];
Keseries=Normal[Series[Ke,{alpha,0,2}]];
Print["Ke series about alpha=0:",Keseries//MatrixForm];
Print["Ke for alpha=0",Simplify[Keseries/.alpha->0]//MatrixForm];

At the start of this script define J and B with the results of items (a) and (b), respectively. Then run the script.
The line Print["exact Ke for alpha=0",Simplify[Ke/.alpha->0]//MatrixForm] will trigger er-
ror messages. Comment on why the exact stiffness cannot be evaluated directly at α = 0 (look at the printed
expression before this one). A Taylor series expansion about α = 0 circumvents these difficulties but the
2-point Gauss integration rule gives the correct answer without the gyrations.

16–13
Chapter 16: THE ISOPARAMETRIC REPRESENTATION

ξ=ξL ξ=ξR
load q
x, u
1 (ξ= −1) 3 (ξ=0) 2 (ξ=1)
x1 = 0 x 3 = /2+α x2 = 
 = L (e)

Figure E16.3. The 3-node bar element under a “box” axial load q.

EXERCISE 16.7 [A/C:20] Construct the consistent force vector for the 3-node bar element of the foregoing
exercise, if the bar is loaded by a uniform axial force q (given per unit of x length) that extends from ξ = ξ L
through ξ = ξ R , and is zero otherwise. Here −1 ≤ ξ L < ξ R ≤ 1. See Figure E16.3. Use
 ξR
f =
e
q NT J dξ, (E16.6)
−ξ L

with the J = d x/dξ found in Exercise 16.5(a) and analytical integration. The answer is quite complicated
and nearly hopeless by hand. Specialize the result to α = 0, ξ L = −1 and ξ R = 1.

16–14
 17
Isoparametric
Quadrilaterals

17–1
Chapter 17: ISOPARAMETRIC QUADRILATERALS

TABLE OF CONTENTS

Page
§17.1. Introduction 17–3
§17.2. Partial Derivative Computation 17–3
§17.2.1. The Jacobian . . . . . . . . . . . . . . . . . 17–3
§17.2.2. Shape Function Derivatives . . . . . . . . . . . . 17–4
§17.2.3. Computing the Jacobian Matrix . . . . . . . . . . . 17–4
§17.2.4. The Strain-Displacement Matrix . . . . . . . . . . 17–5
§17.2.5. *A Shape Function Implementation . . . . . . . . . . 17–5
§17.3. Numerical Integration by Gauss Rules 17–6
§17.3.1. One Dimensional Rules . . . . . . . . . . . . . 17–6
§17.3.2. Implementation of 1D Rules . . . . . . . . . . . . 17–8
§17.3.3. Two Dimensional Rules . . . . . . . . . . . . . 17–8
§17.3.4. Implementation of 2D Gauss Rules . . . . . . . . . . 17–9
§17.4. The Stiffness Matrix 17–10
§17.5. *Integration Variants 17–11
§17.5.1. *Weighted Integration . . . . . . . . . . . . . . 17–11
§17.5.2. *Selective Integration . . . . . . . . . . . . . . . 17–12
§17. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 17–12
§17. References . . . . . . . . . . . . . . . . . . . . . . 17–12
§17. Exercises . . . . . . . . . . . . . . . . . . . . . . 17–13

17–2
 §17.2 PARTIAL DERIVATIVE COMPUTATION

§17.1. Introduction
In this Chapter the isoparametric representation of element geometry and shape functions discussed
in the previous Chapter is used to construct quadrilateral elements for the plane stress problem.
Formulas given in Chapter 14 for the stiffness matrix and consistent load vector of general plane
stress elements are of course applicable to these elements. For a practical implementation, however,
we must go through more specific steps:
1. Construction of shape functions.
2. Computations of shape function derivatives to form the strain-displacement matrix.
3. Numerical integration over the element by Gauss quadrature rules.
The first topic was dealt in the previous Chapter in recipe form, and is systematically covered in
the next one. Assuming the shape functions have been constructed (or readily found in the FEM
literature) the second and third items are combined in an algorithm suitable for programming any
isoparametric quadrilateral. The implementation of the algorithm in the form of element modules
is partly explained in the Exercises of this Chapter, and covered more systematically in Chapter 23.
We shall not deal with isoparametric triangles here to keep the exposition focused. Triangular coor-
dinates, being linked by a constraint, require “special handling” techniques that would complicate
and confuse the exposition. Chapter 24 discusses isoparametric triangular elements in detail.

§17.2. Partial Derivative Computation

Partial derivatives of shape functions with respect to the Cartesian coordinates x and y are required
for the strain and stress calculations. Because shape functions are not directly functions of x and y
but of the natural coordinates ξ and η, the determination of Cartesian partial derivatives is not trivial.
The derivative calculation procedure is presented below for the case of an arbitrary isoparametric
quadrilateral element with n nodes.
§17.2.1. The Jacobian
In quadrilateral element derivations we will need the Jacobian of two-dimensional transformations
that connect the differentials of {x, y} to those of {ξ, η} and vice-versa. Using the chain rule:
 ∂x ∂x   ∂ξ ∂ξ 
           
dx  ∂ξ ∂η  dξ T dξ dξ  ∂x ∂y  dx −T d x
= =J , = =J .
dy ∂ y ∂ y  dη dη dη ∂η ∂η  dy dy
∂ξ ∂η ∂x ∂y
(17.1)
Here J denotes the Jacobian matrix of (x, y) with respect to (ξ, η), whereas J−1 is the Jacobian
matrix of (ξ, η) with respect to (x, y):
 ∂x ∂y   ∂ξ ∂η 
   
∂(x, y)  ∂ξ ∂ξ  J11 J12 −1 ∂(ξ, η)  ∂ x ∂ x  1 J22 −J12
J= = = , J = =  ∂ξ ∂η  = ,
∂(ξ, η) ∂x ∂y  J21 J22 ∂(x, y) J −J21 J11
∂η ∂η ∂y ∂y
(17.2)
where J = |J| = det(J) = J11 J22 − J12 J21 . In FEM work J and J−1 are called simply the Jacobian
and inverse Jacobian, respectively; the fact that it is a matrix being understood. The scalar symbol

17–3
Chapter 17: ISOPARAMETRIC QUADRILATERALS

J is reserved for the determinant of J. In one dimension J and J coalesce. Jacobians play a
crucial role in differential geometry. For the general definition of Jacobian matrix of a differential
transformation, see Appendix D.

Remark 17.1. Observe that the matrices relating the differentials in (17.1) are the transposes of what we
call J and J−1 . The reason is that coordinate differentials transform as contravariant quantities: d x =
(∂ x/∂ξ ) dξ + (∂ x/∂η) dη, etc. But Jacobians are arranged as in (17.2) because of earlier use in covariant
transformations: ∂φ/∂ x = (∂ξ/∂ x)(∂φ/∂ξ ) + (∂η/∂ x)(∂φ/∂η), as in (17.5) below.
The reader is cautioned that notations vary among application areas. As quoted in Appendix D, one author
puts it this way: “When one does matrix calculus, one quickly finds that there are two kinds of people in this
world: those who think the gradient is a row vector, and those who think it is a column vector.”

Remark 17.2. To show that J and J−1 are in fact inverses of each other we form their product:

∂ x ∂ξ ∂ x ∂η ∂y ∂ξ ∂ y ∂η 
∂x ∂y  
∂ξ ∂ x + ∂η ∂ x ∂ξ ∂x + ∂η ∂ x ∂x ∂x 1 0
J−1 J = = ∂y = 0 , (17.3)
∂ x ∂ξ + ∂ x ∂η ∂y ∂ξ ∂ y ∂η ∂x 1
∂ξ ∂ y ∂η ∂ y ∂ξ ∂y + ∂η ∂ y ∂y ∂y
where we have taken into account that x = x(ξ, η), y = y(ξ, η) and the fact that x and y are independent
coordinates. This proof would collapse, however, if instead of {ξ, η} we had the triangular coordinates
{ζ1 , ζ2 , ζ3 } because rectangular matrices have no conventional inverses. This case requires special handling
and is covered in Chapter 24.

§17.2.2. Shape Function Derivatives

The shape functions of a quadrilateral element are expressed in terms of the quadrilateral coordinates
ξ and η introduced in §16.5.1. The derivatives with respect to x and y are given by the chain rule:

∂ Nie ∂ Nie ∂ξ ∂ Nie ∂η ∂ Nie ∂ Nie ∂ξ ∂ Nie ∂η

= + , = + . (17.4)
∂x ∂ξ ∂ x ∂η ∂ x ∂y ∂ξ ∂ y ∂η ∂ y

This can be put in matrix form as

 ∂Ne   ∂ξ ∂η   ∂ Nie   ∂Ne   ∂Ne 
i i i
 ∂x   ∂x ∂ x   ∂ξ  ∂(ξ, η)  ∂ξ  −1  ∂ξ 
 ∂ N e  =  ∂ξ ∂η   ∂ Nie  = ∂(x, y)  ∂ Nie  = J  ∂ Nie  . (17.5)
i
∂y ∂y ∂y ∂η ∂η ∂η

where J−1 is defined in (17.2). The computation of J is addressed in the next subsection.

§17.2.3. Computing the Jacobian Matrix

To compute the entries of J at any quadrilateral location we make use of the last two geometric
relations in (16.4), which are repeated here for convenience:


n 
n
x= xi Nie , y= yi Nie . (17.6)
i=1 i=1

17–4
 §17.2 PARTIAL DERIVATIVE COMPUTATION

Differentiating with respect to the quadrilateral coordinates,

∂x n
∂ Nie ∂y n
∂ Nie ∂x n
∂ Nie ∂y n
∂ Nie
= xi , = yi , = xi , = yi . (17.7)
∂ξ i=1
∂ξ ∂ξ i=1
∂ξ ∂η i=1
∂η ∂η i=1
∂η

because the xi and yi do not depend on ξ and η. In matrix form:

 ∂x  y1 
∂y   ∂Ne ∂ N2e ∂ Nne  x1
  1
...
J11 J12  ∂ξ ∂ξ   ∂ξ ∂ξ ∂ξ   
x2 y2 
J=
J21 J22
=
∂x ∂y 
= PX = 
∂ N1e ∂ N2e 
∂ Nne  ... .. 
. (17.8)
... .
∂η ∂η ∂η ∂η ∂η xn yn
Given a quadrilateral point of coordinates ξ , η we calculate the entries of J using (17.8). The inverse
Jacobian J−1 is then obtained by numerically inverting this 2 × 2 matrix.
Remark 17.3. The symbolic inversion of J for arbitrary ξ , η in general leads to extremely complicated
expressions unless the element has a particularly simple geometry, (for example rectangles as in Exercises
17.1–17.3). This was one of the difficulties that motivated the use of Gaussian numerical quadrature, as
discussed in §17.3 below.

§17.2.4. The Strain-Displacement Matrix

The strain-displacement matrix B that appears in the computation of the element stiffness matrix is
given by the general expression (14.18), which is reproduced here for convenience:
 ∂Ne ∂ N2e ∂ Nne 
  1 0 0 . . . 0
ex x  ∂x ∂x ∂x 
 ∂ N e
∂ N e
∂ N e 
e =  e yy  =  0 ∂y
1 0 ∂y
2 ... 0  e
∂ y  u = Bu .
n e
(17.9)
2ex y  
∂ N1 ∂ N1 ∂ N2 ∂ N2
e e e e
∂N e
∂ Nne

∂y ∂x ∂y ∂x . . . ∂ yn ∂x

The nonzero entries of B are partials of the shape functions with respect to x and y. The calculation
of those partials is done by computing J via (17.8), inverting and using the chain rule (17.5).

Quad4IsoPShapeFunDer[ncoor_,qcoor_]:= Module[
{Nf,dNx,dNy,dNξ ,dNη,i,J11,J12,J21,J22,Jdet,ξ ,η,x,y},
{ξ ,η}=qcoor;
Nf={(1-ξ )*(1-η),(1+ξ )*(1-η),(1+ξ )*(1+η),(1-ξ )*(1+η)}/4;
dNξ ={-(1-η), (1-η),(1+η),-(1+η)}/4;
dNη= {-(1-ξ ),-(1+ξ ),(1+ξ ), (1-ξ )}/4;
x=Table[ncoor[[i,1]],{i,4}]; y=Table[ncoor[[i,2]],{i,4}];
J11=dNξ .x; J12=dNξ .y; J21=dNη.x; J22=dNη.y;
Jdet=Simplify[J11*J22-J12*J21];
dNx= ( J22*dNξ -J12*dNη)/Jdet; dNx=Simplify[dNx];
dNy= (-J21*dNξ +J11*dNη)/Jdet; dNy=Simplify[dNy];
Return[{Nf,dNx,dNy,Jdet}]
];

Figure 17.1. A shape function module for the 4-node bilinear quadrilateral.

17–5
Chapter 17: ISOPARAMETRIC QUADRILATERALS

§17.2.5. *A Shape Function Implementation

To make the foregoing discussion more specific, Figure 17.1 shows the shape function module for the 4-node
bilinear quadrilateral. This is a code fragment that returns the value of the shape functions and their {x, y}
derivatives at a given point of quadrilateral coordinates {ξ, η}. The module is invoked by saying

{ Nf,Nfx,Nfy,Jdet }=Quad4IsoPShapeFunDer[ncoor,qcoor] (17.10)

where the arguments are

ncoor Quadrilateral node coordinates arranged in two-dimensional list form:
{ { x1,y1 },{ x2,y2 },{ x3,y3 },{ x4,y4 } }.
qcoor Quadrilateral coordinates { ξ, η } of the point.
The module returns:
Nf Value of shape functions, arranged as list { Nf1,Nf2,Nf3,Nf4 }.
Nfx Value of x-derivatives of shape functions, arranged as list { Nfx1,Nfx2,Nfx3,Nfx4 }.
Nfy Value of y-derivatives of shape functions, arranged as list { Nfy1,Nfy2,Nfy3,Nfy4 }.
Jdet Jacobian determinant.

Example 17.1. Consider a 4-node bilinear quadrilateral shaped as an axis-aligned 2:1 rectan-
gle, with 2a and a as the x and y dimensions, respectively. The node coordinate array is
ncoor={ { 0,0 },{ 2*a,0 },{ 2*a,a },{ 0,a } }. The shape functions and their {x, y} derivatives are to be
evaluated at the rectangle center ξ = η = 0. The appropiate call is
{ Nf,Nfx,Nfy,Jdet }=Quad4IsoPShapeFunDer[ncoor,{ 0,0 }]
This returns Nf={ 1/8,1/8,3/8,3/8 }, Nfx={ -1/(8*a),1/(8*a),3/(8*a),-3/(8*a) },
Nfy={ -1/(2*a),-1/(2*a),1/(2*a),1/(2*a) } and Jdet=a^2/2.

§17.3. Numerical Integration by Gauss Rules

Numerical integration is essential for practical evaluation of integrals over isoparametric element
domains. The standard practice has been to use Gauss integration because such rules use a minimal
number of sample points to achieve a desired level of accuracy. This economy is important for
efficient element calculations, since a matrix product is evaluated at each sample point. The fact
that the location of the sample points in Gauss rules is usually given by non-rational numbers is of
no concern in digital computation.

§17.3.1. One Dimensional Rules

The classical Gauss integration rules are defined by
 1 
p
F(ξ ) dξ ≈ wi F(ξi ). (17.11)
−1 i=1

Here p ≥ 1 is the number of Gauss integration points (also known as sample points), wi are the
integration weights, and ξi are sample-point abcissae in the interval [−1,1]. The use of the canonical
interval [−1,1] is no restriction, because an integral over another range, say from a to b, can be

17–6
 §17.3 NUMERICAL INTEGRATION BY GAUSS RULES

Table 17.1 - One-Dimensional Gauss Rules with 1 through 5 Sample Points

Points Rule
1
1 −1
F(ξ ) dξ ≈ 2F(0)
1 √ √
2 −1
F(ξ ) dξ ≈ F(−1/ 3) + F(1/ 3)
1 √ √
3 −1
F(ξ ) dξ ≈ 5
9
F(− 3/5) + 8
9
F(0) + 5
9
F( 3/5)
1
4 −1
F(ξ ) dξ ≈ w14 F(ξ14 ) + w24 F(ξ24 ) + w34 F(ξ34 ) + w44 F(ξ44 )
1
5 −1
F(ξ ) dξ ≈ w15 F(ξ15 ) + w25 F(ξ25 ) + w35 F(ξ35 ) + w45 F(ξ45 ) + w55 F(ξ55 )
 √  √
For the 4-point rule,√ξ34 = −ξ24 = (3 − 2 6/5)/7,
√ ξ44 = −ξ14 = (3 + 2 6/5)/7,
w14 = w44 = 12 − 16 5/6, and w24 = w34 = 12 + 16 5/6.
 √  √
For the 5-point rule, ξ55 = −ξ15 = 13 5 + 2 10/7, ξ45 = −ξ35 = 13 5 − 2 10/7, ξ35 = 0,
√ √
w15 = w55 = (322 − 13 70)/900, w25 = w45 = (322 + 13 70)/900 and w35 = 512/900.

ξ = −1 ξ=1
p=1
p=2
p=3
p=4
p=5
Figure 17.2. The first five one-dimensional Gauss rules p = 1, 2, 3, 4, 5 depicted over the line segment
ξ ∈ [−1, +1]. Sample point locations are marked with black circles. The radii of those circles are
proportional to the integration weights.

transformed to [−1, +1] via a simple linear transformation of the independent variable, as shown
in the Remark below.
The first five one-dimensional Gauss rules, illustrated in Figure 17.2, are listed in Table 17.1. These
integrate exactly polynomials in ξ of orders up to 1, 3, 5, 7 and 9, respectively. In general a one-
dimensional Gauss rule with p points integrates exactly polynomials of order up to 2 p − 1. This
is called the degree of the formula.

Remark 17.4. A more general integral, such as F(x) over [a, b] in which = b − a > 0, is transformed
to the canonical interval [−1, 1] through the mapping x = 12 a(1 − ξ ) + 12 b(1 + ξ ) = 12 (a + b) + 12 ξ , or
ξ = (2/ )(x − 12 (a + b)). The Jacobian of this mapping is J = d x/dξ = / . Thus
 b  1  1
F(x) d x = F(ξ ) J dξ = F(ξ ) 12 dξ. (17.12)
a −1 −1

17–7
Chapter 17: ISOPARAMETRIC QUADRILATERALS

LineGaussRuleInfo[{rule_,numer_},point_]:= Module[
{g2={-1,1}/Sqrt[3],w3={5/9,8/9,5/9},
g3={-Sqrt[3/5],0,Sqrt[3/5]},
w4={(1/2)-Sqrt[5/6]/6, (1/2)+Sqrt[5/6]/6,
(1/2)+Sqrt[5/6]/6, (1/2)-Sqrt[5/6]/6},
g4={-Sqrt[(3+2*Sqrt[6/5])/7],-Sqrt[(3-2*Sqrt[6/5])/7],
Sqrt[(3-2*Sqrt[6/5])/7], Sqrt[(3+2*Sqrt[6/5])/7]},
g5={-Sqrt[5+2*Sqrt[10/7]],-Sqrt[5-2*Sqrt[10/7]],0,
Sqrt[5-2*Sqrt[10/7]], Sqrt[5+2*Sqrt[10/7]]}/3,
w5={322-13*Sqrt[70],322+13*Sqrt[70],512,
322+13*Sqrt[70],322-13*Sqrt[70]}/900,
i=point,p=rule,info={{Null,Null},0}},
If [p==1, info={0,2}];
If [p==2, info={g2[[i]],1}];
If [p==3, info={g3[[i]],w3[[i]]}];
If [p==4, info={g4[[i]],w4[[i]]}];
If [p==5, info={g5[[i]],w5[[i]]}];
If [numer, Return[N[info]], Return[Simplify[info]]];
];

Figure 17.3. A Mathematica module that returns the first five one-dimensional Gauss rules.

Remark 17.5. Higher order Gauss rules are tabulated in standard manuals for numerical computation. For
example, the widely used Handbook of Mathematical Functions [2] lists (in Table 25.4) rules with up to 96
points. For p > 6 the abscissas and weights of sample points are not expressible as rational numbers or
radicals, and can only be given as floating-point numbers.

§17.3.2. Implementation of 1D Rules

The Mathematica module shown in Figure 17.3 returns either exact or floating-point information
for the first five unidimensional Gauss rules. To get information for the i th point of the p th rule, in
which 1 ≤ i ≤ p and p = 1, 2, 3, 4, 5, call the module as
{ xii,wi }=LineGaussRuleInfo[{ p,numer },i] (17.13)
Logical flag numer is True to get numerical (floating-point) information, or False to get exact
information. The module returns the sample point abcissa ξi in xii and the weight wi in wi. If p
is not in the implemented range 1 through 5, the module returns { Null,0 }.
Example 17.2. { xi,w }=LineGaussRuleInfo[{ 3,False },2] returns xi=0 and w=8/9, whereas
{ xi,w }=LineGaussRuleInfo[{ 3,True },2] returns (to 16 places) xi=0. and w=0.888888888888889.

§17.3.3. Two Dimensional Rules

The simplest two-dimensional Gauss rules are called product rules. They are obtained by applying
the one-dimensional rules to each independent variable in turn. To apply these rules we must first
reduce the integrand to the canonical form:
 1 1  1  1
F(ξ, η) dξ dη = dη F(ξ, η) dξ. (17.14)
−1 −1 −1 −1

Once this is done we can process numerically each integral in turn:

 1 1  1  1  p1 p2
F(ξ, η) dξ dη = dη F(ξ, η) dξ ≈ wi w j F(ξi , η j ). (17.15)
−1 −1 −1 −1 i=1 j=1

17–8
 §17.3 NUMERICAL INTEGRATION BY GAUSS RULES

p = 1 (1 x 1 rule) p = 2 (2 x 2 rule)

p = 3 (3 x 3 rule) p = 4 (4 x 4 rule)

Figure 17.4. The first four two-dimensional Gauss product rules p = 1, 2, 3, 4

depicted over a straight-sided quadrilateral region. Sample points are marked with
black circles. The areas of these circles are proportional to the integration weights.

where p1 and p2 are the number of Gauss points in the ξ and η directions, respectively. Usually
the same number p = p1 = p2 is chosen if the shape functions are taken to be the same in the ξ
and η directions. This is in fact the case for all quadrilateral elements presented here. The first four
two-dimensional Gauss product rules with p = p1 = p2 are illustrated in Figure 17.4.

§17.3.4. Implementation of 2D Gauss Rules

The Mathematica module listed in Figure 17.5 implements two-dimensional product Gauss rules
having 1 through 5 points in each direction. The number of points in each direction may be the
same or different. If the rule has the same number of points p in both directions the module is
called in either of two ways:

{ { xii,etaj },wij }=QuadGaussRuleInfo[{ p, numer }, { i,j }]

(17.16)
{ { xii,etaj },wij }=QuadGaussRuleInfo[{ p, numer },k ]

The first form is used to get information for point {i, j} of the p × p rule, in which 1 ≤ i ≤ p and
1 ≤ j ≤ p. The second form specifies that point by a “visiting counter” k that runs from 1 through
p 2 ; if so {i, j} are internally extracted1 as j=Floor[(k-1)/p]+1; i=k-p*(j-1).
If the integration rule has p1 points in the ξ direction and p2 points in the η direction, the module
may be called also in two ways:

{ { xii,etaj },wij }=QuadGaussRuleInfo[{ { p1,p2 }, numer },{ i,j }]

(17.17)
{ { xii,etaj },wij }=QuadGaussRuleInfo[{ { p1,p2 }, numer },k ]

The meaning of the second argument is as follows. In the first form i runs from 1 to p1 and j from 1 to
p2 . In the second form k runs from 1 to p1 p2 ; if so i and j are extracted by j=Floor[(k-1)/p1]+1;
1 Indices i and j are denoted by i1 and i2, respectively, inside the module.

17–9
Chapter 17: ISOPARAMETRIC QUADRILATERALS

QuadGaussRuleInfo[{rule_,numer_},point_]:= Module[
{ξ ,η,p1,p2,i,j,w1,w2,m,info={{Null,Null},0}},
If [Length[rule]==2, {p1,p2}=rule, p1=p2=rule];
If [p1<0, Return[QuadNonProductGaussRuleInfo[
{-p1,numer},point]]];
If [Length[point]==2, {i,j}=point, m=point;
j=Floor[(m-1)/p1]+1; i=m-p1*(j-1) ];
{ξ ,w1}= LineGaussRuleInfo[{p1,numer},i];
{η,w2}= LineGaussRuleInfo[{p2,numer},j];
info={{ξ ,η},w1*w2};
If [numer, Return[N[info]], Return[Simplify[info]]];
];

Figure 17.5. A Mathematica module that returns two-dimensional product Gauss rules.

i=k-p1*(i-1). In all four forms, logical flag numer is set to True if numerical information is
desired and to False if exact information is desired.
The module returns ξi and η j in xii and etaj, respectively, and the weight product wi w j in wij.
This code is used in the Exercises at the end of the chapter. If the inputs are not in range, the module
returns { { Null,Null },0 }.

Example 17.3. { { xi,eta },w }=QuadGaussRuleInfo[{ 3,False },{ 2,3 }] returns xi=0, eta=Sqrt[3/5]
and w=40/81.

Example 17.4. { { xi,eta },w }=QuadGaussRuleInfo[{ 3,True },{ 2,3 }] returns (to 16-place precision)
xi=0., eta=0.7745966692414834 and w=0.49382716049382713.

§17.4. The Stiffness Matrix

The stiffness matrix of a general plane stress element is given by the expression (14.23), which is
reproduced here: 
Ke = h BT EB de (17.18)
e

Of the terms that appear in (17.18) the strain-displacement matrix B has been discussed previously.
The thickness h, if variable, may be interpolated via the shape functions. The stress-strain matrix
E is usually constant in elastic problems, but we could in principle interpolate it as appropriate
should it vary over the element. To integrate (17.18) numerically by a two-dimensional product
Gauss rule, we have to reduce it to the canonical form (17.14), that is
 1  1
K =e
F(ξ, η) dξ dη. (17.19)
−1 −1

If ξ and η are the quadrilateral coordinates, everything in (17.19) already fits this form, except the
element of area de .
To complete the reduction we need to express de in terms of the differentials dξ and dη. The
desired relation is (see Remark below)

de = d x d y = det J dξ dη = J dξ dη. (17.20)

17–10
 §17.5 *INTEGRATION VARIANTS

∂x
dη η
∂η
y
C
B
∂y
dη dΩe
∂η ξ
A x
O
∂y
∂x ∂ξ
dξ
dξ
∂ξ

Figure 17.6. Geometric interpretation of the Jacobian-determinant

formula.

We therefore have
F(ξ, η) = h BT EB detJ. (17.21)
This matrix function can be numerically integrated over the domain −1 ≤ ξ ≤ +1, −1 ≤ η ≤ +1
by an appropriate Gauss product rule.
Remark 17.6. To geometrically justify the area transformation formula (17.20), consider the element of area
OACB depicted in Figure 17.6. The area of this differential parallelogram can be computed as

∂x ∂y ∂x ∂y
d A = OB × OA = dξ dη − dη dξ
∂ξ ∂η ∂η ∂ξ
 ∂x ∂x 
  (17.22)
 ∂ξ ∂η 
=  ∂ y ∂ y  dξ dη = |J| dξ dη = det J dξ dη.
 
∂ξ ∂η
This formula can be extended to any number of dimensions, as shown in textbooks on differential geometry;
for example [265,319,708].

§17.5. *Integration Variants

Several deviations from the standard integration schemes described in the foregoing sections are found in the
FEM literature. Two variations are described below and supplemented with motivation Exercises.

§17.5.1. *Weighted Integration

It is sometimes useful to form the element stiffness as a linear combination of stiffnesses produced by two
different integration rules Such schemes are known as weighted integration methods. They are distinguished
from the selective-integration schemes described in the next subsection in that the constitutive properties are
not modified.
For the 4-node bilinear element weighted integration is done by combining the stiffnesses Ke1×1 and Ke2×2
produced by 1×1 and 2×2 Gauss product rules, respectively:
Keβ = (1 − β)Ke1×1 + βKe2×2 . (17.23)
Here β is a scalar in the range [0, 1]. If β = 0 or β = 1 one recovers the element integrated by the 1×1 or
2×2 rule, respectively.2

2 For programming the combination (17.23) may be regarded as a 5-point integration rule with weights
√ w1 = 4(1−β)
√ at
the sample point at ξ = η = 0 and wi = β (i = 2, 3, 4, 5) at the four sample points at ξ = ±1/ 3, η = ±1/ 3.

17–11
Chapter 17: ISOPARAMETRIC QUADRILATERALS

The idea behind (17.23) is that Ke1×1 is rank-deficient and too soft whereas Ke2×2 is rank-sufficient but too stiff.
A combination of too-soft and too-stiff hopefully “balances” the stiffness. An application of this idea to the
mitigation of shear locking for modeling in-plane bending is the subject of Exercise E17.4.

§17.5.2. *Selective Integration

In the FEM literature the term selective integration is used to described a scheme for forming Ke as the sum
of two or more matrices computed with different integration rules and different constitutive properties.3 We
consider here the case of a two-way decomposition. Split the plane stress constitutive matrix E into two:

E = EI + EII (17.24)

This is called a stress-strain splitting. Inserting (17.24) into (17.13) the expression of the stiffness matrix
becomes  
Ke = h BT EI B de + h BT EII B de = KeI + KeII . (17.25)
e e
If these two integrals were done through the same integration rule, the stiffness would be identical to that
obtained by integrating h BT E B de . The trick is to use two different rules: rule (I) for the first integral and
rule (II) for the second.
In practice selective integration is mostly useful for the 4-node bilinear quadrilateral. For this element rules
(I) and (II) are the 1×1 and 2×2 Gauss product rules, respectively. Exercises E17.5–7 investigate stress-strain
splittings (17.24) that improve the in-plane bending performance of rectangular elements.
Notes and Bibliography
The 4-node quadrilateral has a checkered history. It was first derived as a rectangular panel with edge rein-
forcements (not included here) by Argyris in his 1954 Aircraft Engineering series [22, p. 49 in the Butterworths
reprint]. Argyris used bilinear displacement interpolation in Cartesian coordinates.4
After much flailing, a conforming generalization to arbitrary geometry was published in 1964 by Taig and
Kerr [719] using quadrilateral-fitted coordinates already denoted as {ξ, η} but running from 0 to 1. (Reference
[719] cites an 1961 English Electric Aircraft internal report as original source but [397, p. 520] remarks that
the work goes back to 1957.) Bruce Irons, who was aware of Taig’s work while at Rolls Royce, changed the
{ξ, η} range to [−1, 1] to fit Gauss quadrature tables. He proceeded to create the seminal isoparametric family
as a far-reaching extension upon moving to Swansea [64,197,394,397].
Gauss integration is also called Gauss-Legendre quadrature. Gauss presented these rules, derived from first
principles, in 1814; cf. Sec 4.11 of [310]. Legendre’s name is often adjoined because the abcissas of the 1D
sample points turned out to be the zeros of Legendre polynomials. A systematic description is given in [706].
For references in multidimensional numerical integration, see Notes and Bibliography in Chapter 24.
Selective and reduced integration in FEM developed in the early 1970s, and by now there is a huge literature.
An excellent textbook source is [385].

References
Referenced items have been moved to Appendix R.

3 This technique is also called “selective reduced integration” to reflect the fact that one of the rules (the “reduced rule”)
underintegrates the element.
4 This work is probably the first derivation of a continuum-based finite element by assumed displacements. As noted in
§1.7.1, Argyris was aware of the ongoing work in stiffness methods at Turner’s group in Boeing, but the plane stress
models presented in [758] were derived by interelement flux assumptions. Argyris used the unit displacement theorem,
displacing each DOF in turn by one. The resulting displacement pattern is now called a shape function.

17–12
 Exercises

Homework Exercises for Chapter 17

Isoparametric Quadrilaterals

The Mathematica module Quad4IsoPMembraneStiffness listed in Figure E17.1 computes the element
stiffness matrix of the 4-node bilinear quadrilateral. This module is useful as a tool for the Exercises that
follow.

Quad4IsoPMembraneStiffness[ncoor_,Emat_,th_,options_]:=
Module[{i,k,p=2,numer=False,h=th,qcoor,c,w,Nf,
dNx,dNy,Jdet,Be,Ke=Table[0,{8},{8}]},
If [Length[options]==2, {numer,p}=options,{numer}=options];
If [p<1||p>4, Print["p out of range"]; Return[Null]];
For [k=1, k<=p*p, k++,
{qcoor,w}= QuadGaussRuleInfo[{p,numer},k];
{Nf,dNx,dNy,Jdet}=Quad4IsoPShapeFunDer[ncoor,qcoor];
If [Length[th]==4, h=th.Nf]; c=w*Jdet*h;
Be={Flatten[Table[{dNx[[i]], 0},{i,4}]],
Flatten[Table[{0, dNy[[i]]},{i,4}]],
Flatten[Table[{dNy[[i]],dNx[[i]]},{i,4}]]};
Ke+=Simplify[c*Transpose[Be].(Emat.Be)];
]; Return[Simplify[Ke]]
];

Figure E17.1. Mathematica module to compute the stiffness matrix of a 4-node bilinear
quadrilateral in plane stress.

The module makes use of the shape function module Quad4IsoPShapeFunDer listed in Figure 17.1, and of
the Gauss integration modules QuadGaussRuleInfo and (indirectly) LineGaussRuleInfo, listed in Figures
17.5 and are included in the web-posted Notebook Quad4Stiffness.nb.5 The module is invoked as

Ke=Quad4IsoPMembraneStiffness[ncoor,Emat,thick,options] (E17.1)

The arguments are:

ncoor Quadrilateral node coordinates arranged in two-dimensional list form:
{ { x1,y1 },{ x2,y2 },{ x3,y3 },{ x4,y4 } }.
Emat A two-dimensional list storing the 3 × 3 plane stress matrix of elastic moduli:


E 11 E 12 E 13
E= E 12 E 22 E 23 (E17.2)
E 13 E 23 E 33
arranged as { { E11,E12,E33 },{ E12,E22,E23 },{ E13,E23,E33 } }. Must be symmetric. If
the material is isotropic with elastic modulus E and Poisson’s ratio ν, this matrix becomes


E 1 ν 0
E= ν 1 0 (E17.3)
1 − ν2 0 0 1
(1 − ν)
2

thick The plate thickness specified either as a four-entry list: { h1,h2,h3,h4 } or as a scalar: h.

5 This Notebook does not include scripts for doing the Exercises below, although it has some text statements at the bottom
of the cell. You will need to enter the Exercise scripts yourself.

17–13
Chapter 17: ISOPARAMETRIC QUADRILATERALS

The first form is used to specify an element of variable thickness, in which case the entries
are the four corner thicknesses and h is interpolated bilinearly. The second form specifies
uniform thickness.
options Processing options. This list may contain two items: { numer,p } or one: { numer }.
numer is a logical flag with value True or False. If True, the computations are done in
floating point arithmetic. For symbolic or exact arithmetic work set numer to False.6
p specifies the Gauss product rule to have p points in each direction. p may be 1 through 4.
For rank sufficiency, p must be 2 or higher. If p is 1 the element will be rank deficient by
two.7 If omitted p = 2 is assumed.
The module returns Ke as an 8 × 8 symmetric matrix pertaining to the following arrangement of nodal
displacements:
ue = [ u x1 u y1 u x2 u y2 u x3 u y3 u x4 u y4 ]T . (E17.4)

y η
4 3 Uniform thickness h = 1
Isotropic material with elastic
modulus E and Poisson's ratio ν
b = a/γ
ξ

x
1 2
a
Figure E17.2. Element for Exercises 17.1 to 17.3.

For the following three exercises we consider the specialization of the general 4-node bilinear quadrilateral
to a rectangular element dimensioned a and b in the x and y directions, respectively, as depicted in Figure
E17.2. The element has uniform unit thickness h. The material is isotropic with elastic modulus E and
Poisson’s ratio ν and consequently E reduces to (E17.3). The stiffness matrix of this element can be expressed
in closed form.8 For convenience define γ = a/b (rectangle aspect ratio), ψ1 = (1 + ν)γ , ψ2 = (1 − 3ν)γ ,
ψ3 = 2 + (1 − ν)γ 2 , ψ4 = 2γ 2 + (1 − ν), ψ5 = (1 − ν)γ 2 − 4, ψ6 = (1 − ν)γ 2 − 1, ψ7 = 4γ 2 − (1 − ν)
and ψ8 = γ 2 − (1 − ν). Then the stiffness matrix in closed form is
 
4ψ3 3ψ1 2ψ5 −3ψ2 −2ψ3 −3ψ1 −4ψ6 3ψ2
 4ψ4 3ψ2 4ψ8 −3ψ1 −2ψ4 −3ψ2 −2ψ7 
 −3ψ1 −4ψ6 −3ψ2 −2ψ3 3ψ1 
 4ψ3 
Eh  −2ψ7 −2ψ4 
Ke =  4ψ4 3ψ2 3ψ1 . (E17.5)
24γ (1 − ν 2 ) 
 4ψ3 3ψ1 2ψ5 −3ψ2 

 4ψ4 3ψ2 4ψ8 
 4ψ3 −3ψ1

symm 4ψ4

6 The reason for this option is speed. A symbolic or exact computation can take orders of magnitude more time than a
floating-point evaluation. This becomes more pronounced as elements get more complicated.
7 The rank of an element stiffness is discussed in Chapter 19.
8 This closed form can be obtained by either exact integration, or numerical integration with a 2 × 2 or higher Gauss rule.

17–14
 Exercises

EXERCISE 17.1 [C:20] Exercise the Mathematica module of Figure E17.1 with the following script:

ClearAll[Em,nu,a,b,h]; Em=48; h=1; a=4; b=2; nu=0;

ncoor={{0,0},{a,0},{a,b},{0,b}};
Emat=Em/(1-nu^2)*{{1,nu,0},{nu,1,0},{0,0,(1-nu)/2}};
For [p=1, p<=4, p++,
Ke= Quad4IsoPMembraneStiffness[ncoor,Emat,h,{True,p}];
Print["Gauss integration rule: ",p," x ",p];
Print["Ke=",Chop[Ke]//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[N[Ke]]]]
];
Verify that for integration rules p=2,3,4 the stiffness matrix does not change and has three zero eigenvalues,
which correspond to the three two-dimensional rigid body modes. On the other hand, for p = 1 the stiffness
matrix is different and displays five zero eigenvalues, which is physically incorrect. (This phenomenon is
analyzed further in Chapter 19.) Question: why does the stiffness matrix stays exactly the same for p ≥ 2?
Hint: take a look at the entries of the integrand h BT EB J ;for a rectangular geometry are those polynomials
in ξ and η, or rational functions? If the former, of what polynomial order in ξ and η are the entries?

EXERCISE 17.2 [C:20] Check the rectangular element stiffness closed form given in (E17.5). This may be
done by hand (takes a while) or (quicker) running the script of Figure E17.3, which calls the Mathematica
module of Figure E17.1.

ClearAll[Em,Ν,a,b,h,Γ]; b=a/Γ;
ncoor={{0,0},{a,0},{a,b},{0,b}};
Emat=Em/(1-Ν^2)*{{1,Ν,0},{Ν,1,0},{0,0,(1-Ν)/2}};
Ke= Quad4IsoPMembraneStiffness[ncoor,Emat,h,{False,2}];
scaledKe=Simplify[Ke*(24*(1-Ν^2)*Γ/(Em*h))];
Print["Ke=",Em*h/(24*Γ*(1-Ν^2)),"*\n",scaledKe//MatrixForm];

Figure E17.3. Script suggested for Exercise E17.2.

The scaling introduced in the last two lines is for matrix visualization convenience. Verify (E17.5) by printout
inspection and report any typos to instructor.

EXERCISE 17.3 [A/C:25=5+10+10] A Bernoulli-Euler plane beam of thin rectangular cross-section with
span L, height b and thickness h (normal to the plane of the figure) is bent under end moments M as illustrated
in Figure E17.4. The beam is fabricated of isotropic material with elastic modulus E and Poisson’s ratio
ν. The exact solution of the beam problem (from both the theory-of-elasticity and beam-theory standpoints)
is a constant bending moment M along the span. Consequently the beam deforms with uniform curvature
κ = M/(E Iz ), in which Iz = 121
hb3 is the cross-section second moment of inertia about z.
The beam is modeled with one layer of identical 4-node iso-P bilinear quadrilaterals through its height. These
are rectangles with horizontal dimension a; in the Figure a = L/4. The aspect ratio b/a is denoted by γ . By
analogy with the exact solution, all rectangles in the finite element model will undergo the same deformation.
We can therefore isolate a typical element as illustrated in Figure E17.4.
The exact displacement field for the beam segment referred to the {x, y} axes placed at the element center as
shown in the bottom of Figure E17.4, are

u x = −κ x y, u y = 12 κ(x 2 + νy 2 ), (E17.6)

17–15
Chapter 17: ISOPARAMETRIC QUADRILATERALS

M M y

z
b
Cross
section
L h
M M

y
4 3
b=γa x
1 2
a

Figure E17.4. Pure bending of Bernoulli-Euler plane beam of thin rectangular cross section,
for Exercises 17.3–7. The beam is modeled by one layer of 4-node iso-P bilinear quadrilaterals
through its height.

where κ is the deformed beam curvature M/E I . The stiffness equations of the typical rectangular element
are given by the close form expression (E17.5).
The purpose of this Exercise is to compare the in-plane bending response of the 4-node iso-P bilinear rectangle
to that of a Bernoulli-Euler beam element (which would be exact for this configuration). The quadrilateral
element will be called x-bending exact if it reproduces the beam solution for all {γ , ν}. This comparison is
distributed into three items.
(a) Check that (E17.6), as a plane stress 2D elasticity solution, is in full agreement with Bernoulli-Euler beam
theory. This can be done by computing the strains ex x = ∂u x /∂ x, e yy = ∂u y /∂ y and 2ex y = ∂u y /∂ x +
∂u x /∂ y. Then get the stresses σx x , σ yy and σx y through the plane stress constitutive matrix (E17.3) of
an isotropic material. Verify that both σ yy and σx y vanish for any ν, and that σx x = −E κ y = −M y/Iz ,
which agrees with equation (13.4) in Chapter 13.
(b) Compute the strain energy Uquad = 12 (ubeam )T Ke ubeam absorbed by the 4-node element under nodal
displacements ubeam constructed by evaluating (E17.6) at the nodes 1,2,3,4. To simplify this calculation,
it is convenient to decompose that vector as follows:
y
ubeam = ubeam
x
+ ubeam = 14 κab [ −1 0 1 0 −1 0 1 0 ]T
(E17.7)
+ 18 κ(a 2 + νb2 ) [ 0 1 0 1 0 1 0 1 ]T
y
Explain why Ke ubeam must vanish and consequently

Uquad = 12 (ubeam
x
)T Ke ubeam
x
. (E17.8)

This energy can be easily computed by Mathematica by using the first 4 lines of the script of the previous
Exercise, except that here ncoor={ { -a,-b },{ a,-b },{ a,b },{ -a,b } }/2. If vector ubeam x
is formed
in u as a one-dimensional list, Uquad=Simplify[u.Ke.u/2]. This should come out as a function of
M, E, ν, h, a and γ because κ = M/(E Iz ) = 12M/(Eha 3 γ 3 ).
(c) From Mechanics of Materials, or equation (13.7) of Chapter 13, the strain energy absorbed by the
beam segment of length a under a constant bending moment M is Ubeam = 12 Mκa = M 2 a/(2E Iz ) =

17–16
 Exercises

6M 2 /(Eha 2 γ 3 ). Form the energy ratio r = Uquad /Ubeam and show that it is a function of the rectangle
aspect ratio γ = b/a and of Poisson’s ratio ν only:
1 + 2/γ 2 − ν
r = r (γ , ν) = . (E17.9)
(2/γ 2 )(1 − ν 2 )
This happens to be the ratio of the 2D model solution to the exact (beam) solution. Hence r = 1 means
that we get the exact answer, that is the 2D model is x-bending exact. If r > 1 the 2D model is overstiff,
and if r < 1 the 2D model is overflexible. Evidently r > 1 for all γ if 0 ≤ ν ≤ 12 . Moreover if
b << a, r >> 1; for example if a = 10b and ν = 0, r ≈ 50 and the 2D model gives only about 2%
of the correct solution. This phenomenon is referred to in the FEM literature as shear locking, because
overstiffness is due to the bending motion triggering spurious shear energy in the element. Remedies
to shear locking at the element level are studied in advanced FEM courses. Draw conclusions as to the
adequacy or inadequacy of the 2D model to capture inplane bending effects, and comment on how you
might improve results by modifying the discretization of Figure E17.4.9
EXERCISE 17.4 [A+C:20] A naive remedy to shear locking can be attempted with the weighted integration
methodology outlined in §17.6.1. Let Ke1×1 and Ke2×2 denote the element stiffnesses produced by 1×1 and
2×2 Gauss product rules, respectively. Take
Keβ = (1 − β)Ke1×1 + βKe2×2 (E17.10)
where β is adjusted so that shear locking is reduced or eliminated. It is not difficult to find β if the element is
rectangular and isotropic. For the definition of x-bending exact please read the previous Exercise. Inserting
Keβ into the test introduced there verify that
β(1 + 2γ 2 − ν)
r= . (E17.11)
(2/γ 2 )(1 − ν 2 )
Whence show that if
2/γ 2 (1 − ν 2 )
β= , (E17.12)
1 + 2/γ 2 − ν
then r ≡ 1 for all {γ , ν} and the element is x-bending exact. A problem with this idea is that it does not
make it y-bending exact because r (γ ) = r (1/γ ) if γ = 1. Moreover the device is not easily extended to
non-rectangular geometries or non-isotropic material.
EXERCISE 17.5 [A+C:35] (Advanced) To understand this Exercise please begin by reading Exercise 17.3,
and the concept of shear locking. The material is again assumed isotropic with elastic modules E and Poisson’s
ratio ν. The 4-node rectangular element will be said to be bending exact if r = 1 for any {γ , ν} if the bending
test described in Exercise 17.3 is done in both x and y directions. A bending-exact element is completely
shear-lock-free.
The selective integration scheme outlined in §17.6.2 is more effective than weighted integration (covered in
the previous exercise) to fully eliminate shear locking. Let the integration rules (I) and (II) be the 1×1 and
2×2 product rules, respectively. However the latter is generalized so the sample points are located at {−χ, χ},
{χ, −χ }, {χ , χ} and {−χ , χ }, with weight 1.10 Consider the stress-strain splitting




1 ν 0 α β 0 1−α ν−β 0
E E E
E= ν 1 0 = β α 0 + ν−β 1−α 0 = EI + EII , (E17.13)
1−ν 2 0 0 1−ν 1−ν 2 1−ν 1−ν 2 0 0 0
2 0 0 2
9 Note that even if we make a → 0 and γ = b/a → ∞ by taking an infinite number of rectangular elements along x,
the energy ratio r remains greater than one if ν > 0 since r → 1/(1 − ν 2 ). Thus the 2D model would not generally
converge to the correct solution if we keep one layer through the height.
10 For a rectangular
√ geometry these sample points lie on the diagonals. In the case of the standard 2-point Gauss product
rule χ = 1/ 3.

17–17
Chapter 17: ISOPARAMETRIC QUADRILATERALS

where α and β are scalars. Show that if 

1 − ν2
χ= (E17.14)
3(1 − α)
the resulting element stiffness KeI + KeII is bending exact for any {α, β}. As a corollary show that that if α = ν 2 ,
which corresponds to the splitting




1 ν 0 ν2 β 0 1−ν 2 ν−β 0
E E E
E= ν 1 0 = β ν 2
0 + ν−β 1−ν 2 0 = EI + EII , (E17.15)
1−ν 2 0 0 1−ν 1−ν 2 1−ν 2
2 0 0 1−ν
2
0 0 0
√
then χ = 1/ 3 and rule (II) becomes the standard 2×2 Gauss product rule. What are two computationally
convenient settings for β?

EXERCISE 17.6 [A+C:35] (Advanced) A variation on the previous exercise on selective integration to make
the isotropic rectangular 4-node element bending exact. Integration rule (I) is not changed. However rule (II)
has four sample points located at {0, −χ }, {χ , 0}, {0, χ} and {−χ , 0} each with weight 1.11 Show that if one
selects the stress-strain splitting (E17.13) and

2(1 − ν 2 )
χ= (E17.16)
3(1 − α)

the resulting
√ element stiffness KeI + KeII is bending exact for any {α, β}. Discuss which choices of α reduce χ
√
to 1/ 3 and 2/3, respectively.

EXERCISE 17.7 [A+C:40] (Advanced, research paper level, requires a CAS to be tractable) Extend Exercise
17.5 to consider the case of general anisotropic material:


E 11 E 12 E 13
E= E 12 E 22 E 23 (E17.17)
E 13 E 23 E 33
The rules for the selective integration scheme are as described in Exercise 17.5. The appropriate stress-strain
splitting is



E 11 α1 E 12 β E 13 E 11 (1 − α1 ) E 12 (1 − β) 0
E = EI + EII = E 12 β E 22 α2 E 23 + E 12 (1 − β) E 22 (1 − α2 ) 0 (E17.18)
E 13 E 23 E 33 0 0 0
in which β is arbitrary and

|E| 1 |E| 1
1 − α1 = = , 1 − α2 = = ,
11 (E 22 E 33 − E 23 ) 22 (E 11 E 33 − E 13 )
2 2 2 2
3χ 2 E 3χ C11 3χ 2 E 3χ C22
|E| = det(E) = E 11 E 22 E 33 + 2E 12 E 13 E 23 − E 11 E 23
2
− E 22 E 13
2
− E 33 E 12
2
,
C11 = E 11 (E 22 E 33 − E 13
2
)/|E|, C22 = E 22 (E 11 E 33 − E 13
2
)/|E|.
(E17.19)
Show that the√resulting rectangular element is bending exact for any E and χ = 0. (In practice one would
select χ = 1/ 3.)

11 This is called a 4-point median rule, since the four points are located on the quadrilateral medians.

17–18
 18
Shape Function
Magic

18–1
Chapter 18: SHAPE FUNCTION MAGIC

TABLE OF CONTENTS

Page
§18.1. Requirements 18–3
§18.2. Direct Fabrication of Shape Functions 18–3
§18.3. Triangular Element Shape Functions 18–4
§18.3.1. The Three-Node Linear Triangle . . . . . . . . . . . 18–4
§18.3.2. The Six-Node Quadratic Triangle . . . . . . . . . . 18–5
§18.4. Quadrilateral Element Shape Functions 18–6
§18.4.1. The Four-Node Bilinear Quadrilateral . . . . . . . . . 18–6
§18.4.2. The Nine-Node Biquadratic Quadrilateral . . . . . . . 18–7
§18.4.3. The Eight-Node “Serendipity” Quadrilateral . . . . . . . 18–9
§18.5. Does the Magic Wand Always Work? 18–10
§18.5.1. Hierarchical Corrections . . . . . . . . . . . . . 18–10
§18.5.2. Transition Element Example . . . . . . . . . . . . 18–11
§18.6. *Mathematica Modules to Plot Shape Functions 18–12
§18. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 18–14
§18. References . . . . . . . . . . . . . . . . . . . . . . 18–14
§18. Exercises . . . . . . . . . . . . . . . . . . . . . . 18–15

18–2
 §18.2 DIRECT FABRICATION OF SHAPE FUNCTIONS

§18.1. Requirements
This Chapter explains, through a series of examples, how isoparametric shape functions can be
directly constructed by geometric considerations. For a problem of variational index 1, the isopara-
metric shape function Nie associated with node i of element e must satisfy the following conditions:
(A) Interpolation condition. Takes a unit value at node i, and is zero at all other nodes.
(B) Local support condition. Vanishes over any element boundary (a side in 2D, a face in 3D) that
does not include node i.
(C) Interelement compatibility condition. Satisfies C 0 continuity between adjacent elements over
any element boundary that includes node i.
(D) Completeness condition. The interpolation is able to represent exactly any displacement field
which is a linear polynomial in x and y; in particular, a constant value.
Requirement (A) follows directly by interpolation from node values. Conditions (B), (C) and (D)
are consequences of the convergence requirements discussed further in the next Chapter.1 For the
moment these three conditions may be viewed as recipes.
One can readily verify that all isoparametric shape function sets listed in Chapter 16 satisfy the first
two conditions from construction. Direct verification of condition (C) is also straightforward for
those examples. A statement equivalent to (C) is that the value of the shape function over a side
(in 2D) or face (in 3D) common to two elements must uniquely depend only on its nodal values on
that side or face.
Completeness is a property of all element isoparametric shape functions taken together, rather than
of an individual one. If the element satisfies (B) and (C), in view of the discussion in §16.6 it is
sufficient to check that the sum of shape functions is identically one.

§18.2. Direct Fabrication of Shape Functions

Contrary to the what the title of this Chapter implies, the isoparametric shape functions listed in
Chapter 16 did not come out of a magician’s hat. They can be derived systematically by a judicious
inspection process. By “inspection” it is meant that the geometric visualization of shape functions
plays a crucial role.
The method is based on the following observation. In all examples given so far the isoparametric
shape functions are given as products of fairly simple polynomial expressions in the natural coor-
dinates. This is no accident but a direct consequence of the definition of natural coordinates. All
shape functions of Chapter 16 can be expressed as the product of m factors:

Nie = ci L 1 L 2 . . . L m , (18.1)

where
L j = 0, j = 1, . . . m. (18.2)
are the homogeneous equation of lines or curves expressed as linear functions in the natural coor-
dinates, and ci is a normalization coefficient.
1 Convergence means that the discrete FEM solution approaches the exact analytical solution as the mesh is refined.

18–3
Chapter 18: SHAPE FUNCTION MAGIC

(a) 3 (b) 3 (c)

ζ1 = 0
3

1
2 2

1 1 2

Figure 18.1. The three-node linear triangle: (a) element geometry; (b) equation
of side opposite corner 1; (c) perspective view of the shape function N1 = ζ1 .

For two-dimensional isoparametric elements, the ingredients in (18.1) are chosen according to the
following five rules.
R1 Select the L j as the minimal number of lines or curves linear in the natural coordinates that
cross all nodes except the i th node. (A sui generis “cross the dots” game.) Primary choices in
2D are the element sides and medians.
R2 Set coefficient ci so that Nie has the value 1 at the i th node.
R3 Check that Nie vanishes over all element sides that do not contain node i.
R4 Check the polynomial order over each side that contains node i. If the order is n, there must
be exactly n + 1 nodes on the side for compatibility to hold.
R5 If local support (R3) and interelement compatibility (R4) are satisfied, check that the sum of
shape functions is identically one.
The examples that follow show these rules in action for two-dimensional elements. Essentially the
same technique is applicable to one- and three-dimensional elements.

§18.3. Triangular Element Shape Functions

This section illustrates the use of (18.1) in the construction of shape functions for the linear and the
quadratic triangle. The cubic triangle is dealt with in Exercise 18.1.
§18.3.1. The Three-Node Linear Triangle
Figure 18.1 shows the three-node linear triangle that was studied in detail in Chapter 15. The three
shape functions are simply the triangular coordinates: Ni = ζi , for i = 1, 2, 3. Although this result
follows directly from the linear interpolation formula of §15.2.4, it can be also quickly derived from
the present methodology as follows.
The equation of the triangle side opposite to node i is L j -k = ζi = 0, where j and k are the cyclic
permutations of i. Here symbol L j -k denotes the left hand side of the homogeneous equation of
the natural coordinate line that passes through node points j and k. See Figure 18.1(b) for i = 1,
j = 2 and k = 3. Hence the obvious guess is
guess
Nie = ci L i . (18.3)

18–4
 §18.3 TRIANGULAR ELEMENT SHAPE FUNCTIONS

(a) 3 (b) 3 ζ1 = 0
(c) 3 ζ1 = 0

5 5 5
6 6 6

2 2 2
4 ζ1 = 1/2 4 4
ζ2 = 0
1 1 1
Figure 18.2. The six-node quadratic triangle: (a) element geometry; (b) lines
(in red) whose product yields N1e ; (c) lines (in red) whose product yields N4e .

This satisfies conditions (A) and (B) except the unit value at node i; this holds if ci = 1. The
local support condition (B) follows from construction: the value of ζi is zero over side j–k.
Interelement compatibility follows from R4: the variation of ζi along the 2 sides meeting at node
i is linear and that there are two nodes on each side; cf. §15.4.2. Completeness follows since
N1e + N2e + N3e = ζ1 + ζ2 + ζ3 = 1. Figure 18.1(c) depicts N1e = ζ1 , drawn normal to the element
in perspective view.

§18.3.2. The Six-Node Quadratic Triangle

The geometry of the six-node quadratic triangle is shown in Figure 18.2(a). Inspection reveals two
types of nodes: corners (1, 2 and 3) and midside nodes (4, 5 and 6). Consequently we can expect
two types of associated shape functions. We select nodes 1 and 4 as representative cases.
For both cases we try the product of two linear functions in the triangular coordinates because we
expect the shape functions to be quadratic. These functions are illustrated in Figures 18.2(b,c) for
corner node 1 and midside node 4, respectively.
For corner node 1, inspection of Figure 18.2(b) suggests trying
guess
N1e = c1 L 2-3 L 4-6 , (18.4)

Why is (18.4) expected to work? Clearly N1e will vanish over 2-5-3 and 4-6. This makes the function
zero at nodes 2 through 6, as is obvious upon inspection of Figure 18.2(b), while being nonzero at
node 1. This value can be adjusted to be unity if c1 is appropriately chosen. The equations of the
lines that appear in (18.4) are

L 2-3 : ζ1 = 0, L 4- 6 : ζ1 − 1
2
= 0. (18.5)

Replacing into (18.3) we get

N1e = c1 ζ1 (ζ1 − 12 ), (18.6)
To find c1 , evaluate N1e (ζ1 , ζ2 , ζ3 ) at node 1. The triangular coordinates of this node are ζ1 = 1,
ζ2 = ζ3 = 0. We require that it takes a unit value there: N1e (1, 0, 0) = c1 × 1 × 12 = 1 whence
c1 = 2 and finally
N1e = 2ζ1 (ζ1 − 12 ) = ζ1 (2ζ1 − 1), (18.7)

18–5
Chapter 18: SHAPE FUNCTION MAGIC

3
3 6
6 1
1 5
5 4
4
2
2
e
N1 = ζ 1(2ζ1 − 1) N4e = 4ζ1ζ 2

Figure 18.3. Perspective view of shape functions N1e and N4e for the quadratic
triangle. The plot is done over a straight side triangle for programming simplicity.

as listed in §16.5.2. Figure 18.3 shows a perspective view. The other two corner shape functions
follow by cyclic permutations of the corner index.
For midside node 4, inspection of Figure 18.2(c) suggests trying
guess
N4e = c4 L 2-3 L 1-3 (18.8)

Evidently (18.8) satisfies requirements (A) and (B) if c4 is appropriately normalized. The equation
of sides L 2-3 and L 1-3 are ζ1 = 0 and ζ2 = 0, respectively. Therefore N4e (ζ1 , ζ2 , ζ3 ) = c4 ζ1 ζ2 .
To find c4 , evaluate this function at node 4, the triangular coordinates of which are ζ1 = ζ2 = 12 ,
ζ3 = 0. We require that it takes a unit value there: N4e ( 12 , 12 , 0) = c4 × 12 × 12 = 1. Hence c4 = 4,
which gives
N4e = 4ζ1 ζ2 (18.9)
as listed in §16.5.2. Figure 18.3 shows a perspective view of this shape function. The other two
midside shape functions follow by cyclic permutations of the node indices.
It remains to carry out the interelement continuity check. Consider node 1. The boundaries
containing node 1 and common to adjacent elements are 1–2 and 1–3. Over each one the variation
of N1e is quadratic in ζ1 . Therefore the polynomial order over each side is 2. Because there are
three nodes on each boundary, the compatibility condition (C) of §18.1 is verified. A similar check
can be carried out for midside node shape functions. Exercise 16.1 verified that the sum of the Ni
is unity. Therefore the element is complete.

§18.4. Quadrilateral Element Shape Functions

Three quadrilateral elements, with 4, 9 and 8 nodes, respectively, which are commonly used in com-
putational mechanics serve as examples to illustrate the construction of shape functions. Elements
with more nodes, such as the bicubic quadrilateral, are not treated as they are rarely used.

§18.4.1. The Four-Node Bilinear Quadrilateral

The element geometry and natural coordinates are shown in Figure 18.4(a). Only one type of
node (corner) and associated shape function is present. Consider node 1 as typical. Inspection of

18–6
 §18.4 QUADRILATERAL ELEMENT SHAPE FUNCTIONS

η η=1
(a) (c)
3
(b) 3 4
4 4 ξ=1

ξ 3
1
1 1
2 2
2

Figure 18.4. The four-node bilinear quadrilateral: (a) element geometry; (b) sides (in red)
that do not contain corner 1; (c) perspective view of the shape function N1e .

Figure 18.4(b) suggests trying

guess
N1e = c1 L 2-3 L 3-4 (18.10)
This plainly vanishes over nodes 2, 3 and 4, and can be normalized to unity at node 1 by adjusting
c1 . By construction it vanishes over the sides 2–3 and 3–4 that do not belong to 1. The equation of
side 2-3 is ξ = 1, or ξ − 1 = 0. The equation of side 3-4 is η = 1, or η − 1 = 0. Replacing in
(18.10) yields
N1e (ξ, η) = c1 (ξ − 1)(η − 1) = c1 (1 − ξ )(1 − η). (18.11)
To find c1 , evaluate at node 1, the natural coordinates of which are ξ = η = −1:

N1e (−1, −1) = c1 × 2 × 2 = 4c1 = 1. (18.12)

Hence c1 = 1
4
and the shape function is

N1e = 14 (1 − ξ )(1 − η), (18.13)

as listed in §16.6.2. Figure 18.4(c) shows a perspective view.

For the other three nodes the procedure is the same, traversing the element cyclically. It can be
verified that the general expression of the shape functions for this element is

Nie = 14 (1 + ξi ξ )(1 + ηi η). (18.14)

The continuity check proceeds as follows, using N1e as example. Node 1 belongs to interelement
boundaries 1–2 and 1–3. Over side 1–2, η = −1 is constant and N1e is a linear function of ξ . To see
this, replace η = −1 in (18.13). Over side 1–3, ξ = −1 is constant and N1e is a linear function of η.
Consequently the polynomial variation order is 1 over both sides. Because there are two nodes on
each side the compatibility condition is satisfied. The sum of the shape functions is one, as shown
in (16.21); thus the element is complete.

18–7
Chapter 18: SHAPE FUNCTION MAGIC
η
η=1
3
7 η=0 3
4 7 ξ=0
4
9 6
8 ξ 9 6 ξ=1
8
1 5
2 1 5
2
η=1 η=1
η=0 3 3
7 7
4 4
ξ = −1 9 6 ξ=1 ξ = −1 9 6 ξ=1
8 8

1 5 1 5
2 2
η = −1
Figure 18.5. The nine-node biquadratic quadrilateral: (a) element geometry; (b,c,d): lines
(in red) whose product makes up the shape functions N1e , N5e and N9e , respectively.

(a) 4 (b) 4
7 7
8 8
3
3
9 9
1 6 1 6
5
5
2
2
e
N1 = 1
4
(ξ − 1)(η − 1)ξ η N5
e
= 12 (1 − ξ 2 )η(η − 1)

(c) 5 (d)
1 4
7
8 8 9 3
2 9 4

1
6 7 6
5
3 2

N5e = 12 (1 − ξ 2 )η(η − 1) (back view) N9e = (1 − ξ 2 )(1 − η2 )

Figure 18.6. Perspective view of the shape functions for nodes 1, 5 and 9 of the nine-node
biquadratic quadrilateral.

18–8
 §18.4 QUADRILATERAL ELEMENT SHAPE FUNCTIONS

η
η=1 η=1
3
7 3 ξ + η = −1 3
4 7 7
4 4
6
8 ξ 6 ξ = −1 8
6 ξ=1
8

1 5 ξ=1
1 5 1 5
2
2 2

Figure 18.7. The eight-node serendipity quadrilateral: (a) element geometry; (b,c):
lines (in red) whose product make up the shape functions N1e and N5e , respectively.

§18.4.2. The Nine-Node Biquadratic Quadrilateral

The element geometry is shown in Figure 18.5(a). This element has three types of shape functions,
which are associated with corner nodes, midside nodes and center node, respectively.
The lines whose product is used to construct three types of shape functions are illustrated in
Figure 18.5(b,c,d) for nodes 1, 5 and 9, respectively. The technique has been sufficiently illustrated
in previous examples. Here we summarize the calculations for nodes 1, 5 and 9, which are taken
as representatives of the three types:

N1e = c1 L 2-3 L 3-4 L 5-7 L 6-8 = c1 (ξ − 1)(η − 1)ξ η. (18.15)

N5e = c5 L 2-3 L 1-4 L 6-8 L 3-4 = c5 (ξ − 1)(ξ + 1)η(η − 1) = c5 (1 − ξ 2 )η(1 − η). (18.16)
N9e = c9 L 1-2 L 2-3 L 3-4 L 4-1 = c9 (ξ − 1)(η − 1)(ξ + 1)(η + 1) = c9 (1 − ξ 2 )(1 − η2 ) (18.17)
Imposing the normalization conditions we find

c1 = 14 , c5 = − 12 , c9 = 1, (18.18)

and we obtain the shape functions listed in §16.6.3. Perspective views are shown in Figure 18.6.
The remaining Ni ’s are constructed through a similar procedure.
Verification of the interelement continuity condition is immediate: the polynomial variation order
of Nie over any side that belongs to node i is two and there are three nodes on each side. Exercise
16.2 checks that the sum of shape function is unity. Thus the element is complete.

§18.4.3. The Eight-Node “Serendipity” Quadrilateral

This is an eight-node quadrilateral element that results when the center node 9 of the biquadratic
quadrilateral is eliminated by kinematic constraints. The geometry and node configuration is shown
in Figure 18.7(a). This element has been widely used in commercial codes since the 70s for static
problems. It is gradually being phased out in favor of the 9-node quadrilateral for dynamic problems.

18–9
Chapter 18: SHAPE FUNCTION MAGIC

(a) (b) (c) (d)

3 3 3
4 4 3
4

5 6 6
7
1 1 1
2 1
4 2 5 2 2
5

Figure 18.8. Node configurations for which the magic recipe does not work.

The 8-node quadrilateral has two types of shape functions, which are associated with corner nodes
and midside nodes. Lines whose products yields the shape functions for nodes 1 and 5 are shown
in Figure 18.7(b,c).
Here are the calculations for shape functions of nodes 1 and 5, which are taken again as representative
cases.

N1e = c1 L 2-3 L 3-4 L 5-8 = c1 (ξ − 1)(η − 1)(1 + ξ + η) = c1 (1 − ξ )(1 − η)(1 + ξ + η), (18.19)

N5e = c5 L 2-3 L 3-4 L 4-1 = c5 (ξ − 1)(ξ + 1)(η − 1) = c5 (1 − ξ 2 )(1 − η). (18.20)

Imposing the normalization conditions we find

c1 = − 14 , c5 = 1
2
(18.21)

The other shape functions follow by appropriate permutation of nodal indices. The interelement
continuity and completeness verification are similar to that carried out for the nine-node element,
and are relegated to exercises.

§18.5. Does the Magic Wand Always Work?

The “cross the dots” recipe (18.1)–(18.2) is not foolproof. It fails for certain node configurations
although it is a reasonable way to start. It runs into difficulties, for instance, in the problem posed
in Exercise 18.6, which deals with the 5-node quadrilateral depicted in Figure 18.8(a). If for node 1
one tries the product of side 2–3, side 3–4, and the diagonal 2–5–4, the shape function is easily
worked out to be N1e = − 18 (1 − ξ )(1 − η)(ξ + η). This satisfies conditions (A) and (B). However,
it violates (C) along sides 1–2 and 4–1, because it varies quadratically over them with only two
nodes per side.

§18.5.1. Hierarchical Corrections

A more robust technique relies on a correction approach, which employs a combination of terms
such as (18.1). For example, a combination of two patterns, one with m factors and one with n
factors, is
Nie = ci L c1 L c2 . . . L cm + di L d1 L d2 . . . L dn , (18.22)
Here two normalization coefficients: ci and di , appear. In practice trying forms such as (18.22)
from scratch becomes cumbersome. The development is best done hierarchically. The first term is

18–10
 §18.5 DOES THE MAGIC WAND ALWAYS WORK?

taken to be that of a lower order element, called the parent element, for which the one-shot approach
works. The second term is then a corrective shape function that vanishes at the nodes of the parent
element. If this is insufficient one more corrective term is added, and so on.
The technique is best explained through examples. Exercise 18.6 illustrates the procedure for the
element of Figure 18.8(a). The next subsection works out the element of Figure 18.8(b).

§18.5.2. Transition Element Example

The hierarchical correction technique is useful for transition elements, which have corner nodes
but midnodes only over certain sides. Three examples are pictured in Figure 18.8(b,c,d). Shape
functions that work can be derived with one, two and three hierarchical corrections, respectively.
As an example, let us construct the shape function N1e for the 4-node transition triangle shown in
Figure 18.8(b). Candidate lines for the recipe (18.1) are obviously the side 2–3: ζ1 = 0, and the
median 3–4: ζ1 = ζ2 . Accordingly we try
guess
N1e = c1 ζ1 (ζ1 − ζ2 ), N1 (1, 0, 0) = 1 = c1 . (18.23)

This function N1e = ζ1 (ζ1 − ζ2 ) satisfies conditions (A) and (B) but fails compatibility: over side
1–3 of equation ζ2 = 0, because N1e (ζ1 , 0, ζ3 ) = ζ12 . This varies quadratically but there are only 2
nodes on that side. Thus (18.23) is no good.
To proceed hierarchically we start from the shape function for the 3-node linear triangle: N1e = ζ1 .
This will not vanish at node 4, so apply a correction that vanishes at all nodes but 4. From
knowledge of the quadratic triangle midpoint functions, that is obviously ζ1 ζ2 times a coefficient
to be determined. The new guess is
guess
N1e = ζ1 + c1 ζ1 ζ2 . (18.24)

Coefficient c1 is determined by requiring that N1e vanish at 4: N1e ( 12 , 12 , 0) = 1

2
+ c1 41 = 0, whence
c1 = −2 and the shape function is
N1e = ζ1 − 2ζ1 ζ2 . (18.25)
This is easily checked to satisfy compatibility on all sides. The verification of completeness is left
to Exercise 18.8.
Note that since N1e = ζ1 (1 − 2ζ2 ), (18.25) can be constructed as the normalized product of lines
ζ1 = 0 and ζ2 = /. The latter passes through 4 and is parallel to 1–3. As part of the opening moves
in the shape function game this would be a lucky guess indeed. If one goes to a more complicated
element no obvious factorization is possible.

18–11
Chapter 18: SHAPE FUNCTION MAGIC

Cell 18.1 Mathematica Module to Draw a Function over a Triangle Region

PlotTriangleShapeFunction[xytrig_,f_,Nsub_,aspect_]:=Module[
{Ni,line3D={},poly3D={},zc1,zc2,zc3,xyf1,xyf2,xyf3,
xc,yc, x1,x2,x3,y1,y2,y3,z1,z2,z3,iz1,iz2,iz3,d},
{{x1,y1,z1},{x2,y2,z2},{x3,y3,z3}}=Take[xytrig,3];
xc={x1,x2,x3}; yc={y1,y2,y3}; Ni=Nsub*3;
Do [ Do [iz3=Ni-iz1-iz2; If [iz3<=0, Continue[]]; d=0;
If [Mod[iz1+2,3]==0&&Mod[iz2-1,3]==0, d= 1];
If [Mod[iz1-2,3]==0&&Mod[iz2+1,3]==0, d=-1];
If [d==0, Continue[]];
zc1=N[{iz1+d+d,iz2-d,iz3-d}/Ni];
zc2=N[{iz1-d,iz2+d+d,iz3-d}/Ni];
zc3=N[{iz1-d,iz2-d,iz3+d+d}/Ni];
xyf1={xc.zc1,yc.zc1,f[zc1[[1]],zc1[[2]],zc1[[3]]]};
xyf2={xc.zc2,yc.zc2,f[zc2[[1]],zc2[[2]],zc2[[3]]]};
xyf3={xc.zc3,yc.zc3,f[zc3[[1]],zc3[[2]],zc3[[3]]]};
AppendTo[poly3D,Polygon[{xyf1,xyf2,xyf3}]];
AppendTo[line3D,Line[{xyf1,xyf2,xyf3,xyf1}]],
{iz2,1,Ni-iz1}],{iz1,1,Ni}];
Show[ Graphics3D[RGBColor[1,0,0]],Graphics3D[poly3D],
Graphics3D[Thickness[.002]],Graphics3D[line3D],
Graphics3D[RGBColor[0,0,0]],Graphics3D[Thickness[.005]],
Graphics3D[Line[xytrig]],PlotRange->All,
BoxRatios->{1,1,aspect},Boxed->False]
];
ClearAll[f1,f4];
xyc1={0,0,0}; xyc2={3,0,0}; xyc3={Sqrt[3],3/2,0};
xytrig=N[{xyc1,xyc2,xyc3,xyc1}]; Nsub=16;
f1[zeta1_,zeta2_,zeta3_]:=zeta1*(2*zeta1-1);
f4[zeta1_,zeta2_,zeta3_]:=4*zeta1*zeta2;
PlotTriangleShapeFunction[xytrig,f1,Nsub,1/2];
PlotTriangleShapeFunction[xytrig,f4,Nsub,1/2.5];

§18.6. *Mathematica Modules to Plot Shape Functions

A Mathematica module called PlotTriangleShape Functions, listed in Cell 18.1, has been developed to
draw perspective plots of shape functions Ni (ζ1 , ζ2 , ζ3 ) over a triangular region. The region is assumed to
have straight sides to simplify the logic. The test statements that follow the module produce the shape function
plots shown in Figure 18.3 for the 6-node quadratic triangle. Argument Nsub controls the plot resolution while
aspect controls the x yz box aspect ratio. The remaining arguments are self explanatory.
Another Mathematica module called PlotQuadrilateralShape Functions, listed in Cell 18.2, has been
developed to produce perspective plots of shape functions Ni (ξ, η) over a quadrilateral region. The region
is assumed to have straight sides to simplify the logic. The test statements that follow the module produce
the shape function plots shown in Figure 18.6(a,b,d) for the 9-node biquadratic quadrilateral. Argument Nsub

18–12
 §18.6 *MATHEMATICA MODULES TO PLOT SHAPE FUNCTIONS

Cell 18.2 Mathematica Module to Draw a Function over a Quadrilateral Region

PlotQuadrilateralShapeFunction[xyquad_,f_,Nsub_,aspect_]:=Module[
{Ne,Nev,line3D={},poly3D={},xyf1,xyf2,xyf3,i,j,n,ixi,ieta,
xi,eta,x1,x2,x3,x4,y1,y2,y3,y4,z1,z2,z3,z4,xc,yc},
{{x1,y1,z1},{x2,y2,z2},{x3,y3,z3},{x4,y4,z4}}=Take[xyquad,4];
xc={x1,x2,x3,x4}; yc={y1,y2,y3,y4};
Ne[xi_,eta_]:=N[{(1-xi)*(1-eta),(1+xi)*(1-eta),
(1+xi)*(1+eta),(1-xi)*(1+eta)}/4]; n=Nsub;
Do [ Do [ ixi=(2*i-n-1)/n; ieta=(2*j-n-1)/n;
{xi,eta}=N[{ixi-1/n,ieta-1/n}]; Nev=Ne[xi,eta];
xyf1={xc.Nev,yc.Nev,f[xi,eta]};
{xi,eta}=N[{ixi+1/n,ieta-1/n}]; Nev=Ne[xi,eta];
xyf2={xc.Nev,yc.Nev,f[xi,eta]};
{xi,eta}=N[{ixi+1/n,ieta+1/n}]; Nev=Ne[xi,eta];
xyf3={xc.Nev,yc.Nev,f[xi,eta]};
{xi,eta}=N[{ixi-1/n,ieta+1/n}]; Nev=Ne[xi,eta];
xyf4={xc.Nev,yc.Nev,f[xi,eta]};
AppendTo[poly3D,Polygon[{xyf1,xyf2,xyf3,xyf4}]];
AppendTo[line3D,Line[{xyf1,xyf2,xyf3,xyf4,xyf1}]],
{i,1,Nsub}],{j,1,Nsub}];
Show[ Graphics3D[RGBColor[1,0,0]],Graphics3D[poly3D],
Graphics3D[Thickness[.002]],Graphics3D[line3D],
Graphics3D[RGBColor[0,0,0]],Graphics3D[Thickness[.005]],
Graphics3D[Line[xyquad]], PlotRange->All,
BoxRatios->{1,1,aspect},Boxed->False]
];
ClearAll[f1,f5,f9];
xyc1={0,0,0}; xyc2={3,0,0}; xyc3={3,3,0}; xyc4={0,3,0};
xyquad=N[{xyc1,xyc2,xyc3,xyc4,xyc1}]; Nsub=16;
f1[xi_,eta_]:=(1/2)*(xi-1)*(eta-1)*xi*eta;
f5[xi_,eta_]:=(1/2)*(1-xi^2)*eta*(eta-1);
f9[xi_,eta_]:=(1-xi^2)*(1-eta^2);
PlotQuadrilateralShapeFunction[xyquad,f1,Nsub,1/2];
PlotQuadrilateralShapeFunction[xyquad,f5,Nsub,1/2.5];
PlotQuadrilateralShapeFunction[xyquad,f9,Nsub,1/3];

controls the plot resolution while aspect controls the x yz box aspect ratio. The remaining arguments are self
explanatory.

18–13
Chapter 18: SHAPE FUNCTION MAGIC

Notes and Bibliography

The name “shape functions” for interpolation functions directly expressed in terms of physical coordinates
(the node displacements in the case of isoparametric elements) was coined by Irons. The earliest published
reference seems to be the paper [64]. This was presented in 1965 at the first Wright-Patterson conference, the
first all-FEM meeting that strongly influenced the development of computational mechanics in Generation 2.
The key connection to numerical integration was presented in [394], although it is mentioned in prior internal
reports. A comprehensive exposition is given in the textbook by Irons and Ahmad [397].
The quick way of developing shape functions presented here was used in the writer’s 1966 thesis [203] for
triangular elements. The qualifier “magic” arose from the timing for covering this Chapter in a Fall Semester
course: the lecture falls near Halloween.

References
Referenced items have been moved to Appendix R.

18–14
 Exercises

Homework Exercises for Chapter 18

Shape Function Magic

EXERCISE 18.1 [A/C:10+10] The complete cubic triangle for plane stress has 10 nodes located as shown in
Figure E18.1, with their triangular coordinates listed in parentheses.

3(0,0,1) 3

7(0,1/3,2/3)
8 7
8(1/3,0,2/3)
6(0,2/3,1/3)
9 6
9(2/3,0,1/3) 0(1/3,1/3,1/3) 2(0,1,0) 0
2
5(1/3,2/3,0)
4 5
4(2/3,1/3,0)
1(1,0,0) 1

Figure E18.1. Ten-node cubic triangle for Exercise 18.1. The left picture shows the
superparametric element whereas the right one shows the isoparametric version with curved sides.

e
N1e N4e N0

Figure E18.2. Perspective plots of the shape functions N1e , N4e and N0e
for the 10-node cubic triangle.

(a) Construct the cubic shape functions N1e , N4e and N0e for nodes 1, 4, and 0 (the interior node is labeled as
zero, not 10) using the line-product technique. [Hint: each shape function is the product of 3 and only 3
lines.] Perspective plots of those 3 functions are shown in Figure E18.2.
(b) Construct the missing 7 shape functions by appropriate node number permutations, and verify that the
sum of the 10 functions is identically one. For the unit sum check use the fact that ζ1 + ζ2 + ζ3 = 1.

EXERCISE 18.2 [A:15] Find an alternative shape function N1e for corner node 1 of the 9-node quadrilateral
of Figure 18.5(a) by using the diagonal lines 5–8 and 2–9–4 in addition to the sides 2–3 and 3–4. Show that
the resulting shape function violates the compatibility condition (C) stated in §18.1.

EXERCISE 18.3 [A/C:15] Complete the above exercise for all nine nodes. Add the shape functions (use a
CAS and simplify) and verify whether their sum is unity.

EXERCISE 18.4 [A/C:20] Verify that the shape functions N1e and N5e of the eight-node serendipity quadri-
lateral discussed in §18.4.3 satisfy the interelement compatibility condition (C) stated in §18.1. Obtain all 8
shape functions and verify that their sum is unity.

EXERCISE 18.5 [C:15] Plot the shape functions N1e and N5e of the eight-node serendipity quadrilateral studied
in §18.4.3 using the module PlotQuadrilateralShapeFunction listed in Cell 18.2.

18–15
Chapter 18: SHAPE FUNCTION MAGIC

η
3 N1 N5
4

5 ξ

1
2

Figure E18.3. Five node quadrilateral element for Exercise 18.6.

EXERCISE 18.6 [A:15]. A five node quadrilateral element has the nodal configuration shown in Figure E18.3.
Perspective views of N1e and N5e are shown in that Figure.2 Find five shape functions Nie , i = 1, 2, 3, 4, 5 that
satisfy compatibility, and also verify that their sum is unity.
Hint: develop N5 (ξ, η) first for the 5-node quad using the line-product method; then the corner shape functions
N̄i (ξ, η) (i = 1, 2, 3, 4) for the 4-node quad (already given in the Notes); finally combine Ni = N̄i + α N5 ,
determining α so that all Ni vanish at node 5. Check that N1 + N2 + N3 + N4 + N5 = 1 identically.
µ
EXERCISE 18.7 [A:15]. An eight-node “brick” finite ele-
ment for three dimensional analysis has three isoparametric
natural coordinates called ξ , η and µ. These coordinates vary 8 η
from −1 at one face to +1 at the opposite face, as sketched
in Figure E18.4. 5 7

Construct the (trilinear) shape function for node 1 (follow the

z 4 6
node numbering of the figure). The equations of the brick ξ
faces are: 3
1
x 2
1485 : ξ = −1 2376 : ξ = +1 y
1265 : η = −1 4378 : η = +1
1234 : µ = −1 5678 : µ = +1 Figure E18.4. Eight-node isoparametric
“brick” element for Exercise 18.7.

EXERCISE 18.8 [A:15]. Consider the 4-node transition triangular element of Figure 18.8(b). The shape
function for node 1, N1 = ζ1 − 2ζ1 ζ2 was derived in §18.5.2 by the correction method. Show that the others
are N2 = ζ2 − 2ζ1 ζ2 , N3 = ζ3 and N4 = 4ζ1 ζ2 . Check that compatibility and completeness are verified.

EXERCISE 18.9 [A:15]. Construct the six shape functions for the 6-node transition quadrilateral element of
Figure 18.8(c). Hint: for the corner nodes, use two corrections to the shape functions of the 4-node bilinear
quadrilateral. Check compatibility and completeness. Partial result: N1 = 14 (1 − ξ )(1 − η) − 14 (1 − ξ 2 )(1 − η).

EXERCISE 18.10 [A:20]. Consider a 5-node transition triangle in which midnode 6 on side 1–3 is missing.
Show that N1e = ζ1 − 2ζ1 ζ2 − 2ζ2 ζ3 . Can this be expressed as a line product like (18.1)?

2 Although this N1e resembles the N1e of the 4-node quadrilateral depicted in Figure 18.4, they are not the same. That in
Figure E18.3 must vanish at node 5 (ξ = η = 0). On the other hand, the N1e of Figure 18.4 takes the value 14 there.

18–16
 Exercises

4 3
4

Set CL1
Side 2-4 maps to this
parabola; part of triangle
2-3-4 turns "inside out"
4
3 2
4
1 2

4 3
4
1 2
Reference 2
triangular
elements
Set CL2
2
1
2

Figure E18.5. Mapping of reference triangles under sets (E18.1) and (E18.2).
Triangles are slightly separated at the diagonal 2–4 for visualization convenience.

EXERCISE 18.11 [A:30]. The three-node linear triangle is known to be a poor performer for stress analysis.
In an effort to improve it, Dr. I. M. Clueless proposes two sets of quadratic shape functions:
CL1: N1 = ζ12 , N2 = ζ22 , N3 = ζ32 . (E18.1)
CL2: N1 = ζ12 + 2ζ2 ζ3 , N2 = ζ22 + 2ζ3 ζ1 , N3 = ζ32 + 2ζ1 ζ2 . (E18.2)
Dr. C. writes a learned paper claiming that both sets satisfy the interpolation condition, that set CL1 will work
because it is conforming and that set CL2 will work because N1 + N2 + N3 = 1. He provides no numerical
examples. You get the paper for review. Show that the claims are false, and both sets are worthless. Hint:
study §16.6 and Figure E18.5.
EXERCISE 18.12 [A:25]. Another way of constructing shape functions for “incomplete” elements is through
kinematic multifreedom constraints (MFCs) applied to a “parent” element that contains the one to be derived.
Suppose that the 9-node biquadratic quadrilateral is chosen as parent, with shape functions called NiP , i =
1, . . . 9 given in §18.4.2. To construct the shape functions of the 8-node serentipity quadrilateral, the motions
of node 9 are expressed in terms of the motions of the corner and midside nodes by the interpolation formulas
u x9 = α(u x1 + u x2 + u x3 + u x4 ) + β(u x5 + u x6 + u x7 + u x8 ),
(E18.3)
u y9 = α(u y1 + u y2 + u y3 + u y4 ) + β(u y5 + u y6 + u y7 + u y8 ),
where α and β are scalars to be determined. (In the terminology of Chapter 9, u x9 and u y9 are slaves
while boundary DOFs are masters.) Show that the shape functions of the 8-node quadrilateral are then
Ni = NiP + α N9P for i = 1, . . . 4 and Ni = NiP + β N9P for i = 5, . . . 8. Furthermore, show that α and β can
be determined by two conditions:
8
1. The unit sum condition: i=1
Ni = 1, leads to 4α + 4β = 1.
2. Exactness of displacement interpolation for ξ 2 and η2 leads to 2α + β = 0.
Solve these two equations for α and β, and verify that the serendipity shape functions given in §18.4.3 result.
EXERCISE 18.13 [A:25] Construct the 16 shape functions of the bicubic quadrilateral.

18–17
 19
FEM Convergence
Requirements

19–1
Chapter 19: FEM CONVERGENCE REQUIREMENTS

TABLE OF CONTENTS

Page
§19.1 Overview . . . . . . . . . . . . . . . . . . . . . . 19–3
§19.2 The Variational Index . . . . . . . . . . . . . . . . . 19–3
§19.3 Consistency Requirements . . . . . . . . . . . . . . . . 19–4
§19.3.1 Completeness . . . . . . . . . . . . . . . . . 19–4
§19.3.2 Compatibility . . . . . . . . . . . . . . . . . 19–4
§19.3.3 Matching and Non-Matching Meshes . . . . . . . . . 19–6
§19.4 Stability . . . . . . . . . . . . . . . . . . . . . . . 19–7
§19.4.1 Rank Sufficiency . . . . . . . . . . . . . . . . 19–8
§19.4.2 Jacobian Positiveness . . . . . . . . . . . . . . . 19–9
§19. Notes and Bibliography . . . . . . . . . . . . . . . . . 19–11
§19. References . . 19–11
§19. Exercises . . . . . . . . . . . . . . . . . . . . . . 19–13

19–2
 §19.2 THE VARIATIONAL INDEX

§19.1. Overview

Chapters 11 through 18 have discussed, in piecemeal fashion, requirements for shape functions
of isoparametric elements. These are motivated by convergence: as the mesh is refined, the FEM
solution should approach the analytical solution of the mathematical model.1 This attribute is
obviously necessary to instill confidence in FEM results from the standpoint of mathematics.
This Chapter provides unified information on convergence requirements. These requirements can
be grouped into three:
Completeness. The elements must have enough approximation power to capture the analytical
solution in the limit of a mesh refinement process. This intuitive statement is rendered more precise
below.
Compatibility. The shape functions should provide displacement continuity between elements.
Physically these insure that no material gaps appear as the elements deform. As the mesh is refined,
such gaps would multiply and may absorb or release spurious energy.
Stability. The system of finite element equations must satify certain well posedness conditions that
preclude nonphysical zero-energy modes in elements, as well as the absence of excessive element
distortion.
Completeness and compatibility are two aspects of the so-called consistency condition between
the discrete and mathematical models. A finite element model that passes both completeness and
continuity requirements is called consistent. This is the FEM analog of the famous Lax-Wendroff
theorem,2 which says that consistency and stability imply convergence.

Remark 19.1. A deeper mathematical analysis done in more advanced courses shows that completeness is
necessary for convergence whereas failure of the other requirements does not necessarily precludes it. There
are, for example, FEM models in common use that do not satisfy compatibility. Furthermore, numerically
unstable models may be used (with caution) in situations where that property is advantageous, as in the
modeling of local singularities. Nonetheless, the satisfaction of the three criteria guarantees convergence and
may therefore be regarded as a safe choice for the beginner user.

§19.2. The Variational Index

For the mathematical statement of the completeness and continuity conditions, the variational index
alluded to in previous sections plays a fundamental role.
The FEM is based on the direct discretization of an energy functional [u], where u (displacements
for the elements considered in this book) is the primary variable, or (equivalently) the function to
be varied. Let m be the highest spatial derivative order of u that appears in . This m is called the
variational index.

1 Of course FEM convergence does not guarantee the correctness of the mathematical model in capturing the physics. As
discussed in Chapter 1, model verification against experiments is a different and far more difficult problem.
2 Proven originally for classical finite difference discretizations in fluid mechanics. More precisely, it states that a numerical
scheme for the scalar conservation law, du/dt + d f /d x = 0 converges to a unique (weak) solution, if it is consistent,
stable and conservative. There is no equivalent theorem for systems of conservation laws.

19–3
Chapter 19: FEM CONVERGENCE REQUIREMENTS

Example 19.1. In the bar problem discussed in Chapter 11,

 L
1 
[u] = 2
u  E Au  − qu d x. (19.1)
0

The highest derivative of the displacement u(x) is u  = du/d x, which is first order in the space coordinate x.
Consequently m = 1. This is also the case on the plane stress problem studied in Chapter 14, because the
strains are expressed in terms of first order derivatives of the displacements.

Example 19.2. In the plane beam problem discussed in Chapter 12,

 L
1 
[v] = 2
v  E I v  − qv d x. (19.2)
0

The highest derivative of the transverse displacement is the curvature κ = v  = d 2 v/d x 2 , which is of second
order in the space coordinate x. Consequently m = 2.

§19.3. Consistency Requirements

Using the foregoing definition of variational index, we can proceed to state the two key requirements
for finite element shape functions.

§19.3.1. Completeness

The element shape functions must represent exactly all polynomial terms
of order ≤ m in the Cartesian coordinates. A set of shape functions that
satisfies this condition is called m-complete.

Note that this requirement applies at the element level and involves all shape functions of the
element.

Example 19.3. Suppose a displacement-based element is for a plane stress problem, in which m = 1. Then
1-completeness requires that the linear displacement field

u x = α0 + α1 x + α2 y, u y = α0 + α1 x + α2 y (19.3)

be exactly represented for any value of the α coefficients. This is done by evaluating (19.3) at the nodes to
form a displacement vector ue and then checking that u = Ne ue recovers exactly (19.3). Section 16.6 presents
the details of this calculation for an arbitrary isoparametric plane stress element. That analysis shows that
completeness is satisfied if the sum of the shape functions is unity and the element is compatible.

Example 19.4. For the plane beam problem, in which m = 2, the quadratic transverse displacement

v = α0 + α1 x + α2 x 2 (19.4)

must be exactly represented over the element. This is easily verified in for the 2-node beam element developed
in Chapter 13, because the assumed transverse displacement is a complete cubic polynomial in x. A complete
cubic contains the quadratic (19.4) as special case.

19–4
 §19.3 CONSISTENCY REQUIREMENTS

(a) (b) (c)

i i i

2-node bars
Figure 19.1. An element patch is the set of all elements attached to a patch node, labeled i.
(a) illustrates a patch of triangles; (b) a mixture of triangles and quadrilaterals; (c) a mixture of
triangles, quadrilaterals, and bars.

§19.3.2. Compatibility
To state this requirement succintly, it is convenient to introduce the concept of element patch, or
simply patch. This is the set of all elements attached to a given node, called the patch node. The
definition is illustrated in Figure 19.1, which shows three different kind of patches attached to patch
node i in a plane stress problem. The patch of Figure 19.1(a) contains only one type of element:
3-node linear triangles. The patch of Figure 19.1(b) mixes two plane stress element types: 3-node
linear triangles and 4-node bilinear quadrilaterals. The patch of Figure 19.1(c) combines three
element types: 3-node linear triangles, 4-node bilinear quadrilaterals, and 2-node bars.
We define a finite element patch trial function as the union of shape functions activated by setting
a degree of freedom at the patch node to unity, while all other freedoms are zero.
A patch trial function “propagates” only over the patch, and is zero beyond it. This property follows
from the local-support requirement stated in §18.1: a shape function for node i should vanish on
all sides or faces that do not include i.
With the help of these definitions we can enunciate the compatibility requirement as follows.

Patch trial functions must be C (m−1) continuous between interconnected

elements, and C m piecewise differentiable inside each element.

If the variational index is m = 1, the patch trial functions must be C 0 continuous between elements,
and C 1 inside elements.
A set of shape functions that satisfies the first requirement is called conforming. A conforming
expansion that satisfies the second requirement is said to be of finite energy. Note that this condition
applies at two levels: individual element, and element patch. An element endowed with conforming
shape functions is said to be conforming. A conforming element that satisfies the finite energy
requirement is said to be compatible.3
3 The FEM literature is a bit fuzzy as regards these terms. It seems better to leave the qualifier “conforming” to denote
interelement compatibility; informally “an element that gets along with its neighbors.” The qualifier “compatible” is
used in the stricter sense of conforming while possessing sufficient internal smoothness.

19–5
Chapter 19: FEM CONVERGENCE REQUIREMENTS

(a) (b) (c)

3-node bars

(e)
(d) (f)

Figure 19.2. Examples of 2D non-matching meshes. Interelement boundaries that fail matching
conditions are shown offset for visualization convenience. In (a,b,c) some nodes do not match. In
(d,e,f) nodes and DOFs match but some sides do not, leading to violations of C 0 continuity.

Figures 19.1(b,c) illustrates the fact that one needs to check the possible connection of matching
elements of different types and possibly different dimensionality.
§19.3.3. Matching and Non-Matching Meshes
As stated, compatibility refers to the complete finite element mesh because mesh trial functions are
a combination of patch trial functions, which in turn are the union of element shape functions. This
generality poses some logistical difficulties because the condition is necessarily mesh dependent.
Compatibility can be checked at the element level by restricting attention to matching meshes. A
matching mesh is one in which adjacent elements share sides, nodes and degrees of freedom, as in
the patches shown in Figure 19.1.
For a matching mesh it is sufficient to restrict consideration first to a pair of adjacent elements, and
then to the side shared by these elements. Suppose that the variation of a shape function along
that side is controlled by k nodal values. Then a polynomial variation of order up to k − 1 in the
natural coordinate(s) can be specified uniquely over the side. This is sufficient to verify interelement
compatibility for m = 1, implying C 0 continuity, if the shape functions are polynomials.
This simplified criterion is the one used in previous Chapters. Specific 2D examples were given in
Chapters 15 through 18.
Remark 19.2. If the variational index is m = 2 and the problem is multidimensional, as in the case of
plates and shells, the check is far more involved and trickier because continuity of normal derivatives along
a side is involved. This practically important scenario is examined in advanced FEM treatments. The case of
non-polynomial shape functions is, on the other hand, of little practical interest.

19–6
 §19.4 STABILITY

Mesh of
tetrahedra

Common surface

Mesh of
bricks

Figure 19.3. Example of a 3D non-matching mesh. Top portion discretized with

tetrahedra, lower portion with bricks. Nodes and boundary-quad edges and DOFs
match, but element types are different, leading to violation of C 0 continuity.

A mesh that does not satisfy the matching criteria stated above is called a nonmatching mesh.
Several two-dimensional examples are shown in Figure 19.2. As can be seen there is a wide range
of possibilities: nonmatching nodes, matching nodes but different element types, etc. Figure 19.3
depicts a three-dimensional example, in which case even more variety can be expected.
Nonmatching meshes are the rule rather than the exception in contact and impact problems (which,
being geometrically nonlinear, are outside the scope of this book). See Figure 19.4 illustrates what
happens in a problem of slipping contact.
Initial shape Deformed shape

;;;;;;;;;
;;;;;;;;;
;;;;;;;;;
;;;;;;;;;
Figure 19.4. In contact and impact problems, matching meshes are the exception rather than the
rule. Even if the meshes match at initial contact, slipping may produce a nonmatching mesh in
the deformed configuration, as illustrated in the figure.

In multiphysics simulations nonmatching meshes are common, since they are often prepared sepa-
rately for the different physical components, as illustrated in Figure 19.5.

§19.4. Stability

Stability may be informally characterized as ensuring that the finite element model enjoys the same
solution uniqueness properties of the analytical solution of the mathematical model. For example, if
the only motions that produce zero internal energy in the mathematical model are rigid body motions,

19–7
Chapter 19: FEM CONVERGENCE REQUIREMENTS

Fluid

Structure

Figure 19.5. Nonmatching meshes are common in multiphysics problems, as in this example of
fluid-structure interaction (FSI). Two-dimensional model to simulate flow around a thin plate. If
the meshes are independenly prepared node locations will not generally match.

the finite element model must inherit that property. Since FEM can handle arbitrary assemblies of
elements, including individual elements, this property is required to hold at the element level.
In the present outline we are concerned with stability at the element level. Stability is not a property
of shape functions per se but of the implementation of the element as well as its geometrical
definition. It involves two subordinate requirements: rank sufficiency, and Jacobian positiveness.
Of these, rank sufficiency is the most important one.
§19.4.1. Rank Sufficiency
The element stiffness matrix must not possess any zero-energy kinematic mode other than rigid
body modes.
This can be mathematically expressed as follows. Let n F be the number of element degrees of
freedom, and n R be the number of independent rigid body modes. Let r denote the rank of Ke . The
element is called rank sufficient if r = n F − n R and rank deficient if r < n F − n R . In the latter
case, the rank deficiency is defined by
d = (n F − n R ) − r (19.5)
If an isoparametric element is numerically integrated, let n G be the number of Gauss points, while
n E denotes the order of the stress-strain matrix E. Two additional assumptions are made:
(i) The element shape functions satisfy completeness in the sense that the rigid body modes are
exactly captured by them.
(ii) Matrix E is of full rank.
Then each Gauss point adds n E to the rank of Ke , up to a maximum of n F − n R . Hence the rank
of Ke will be
r = min(n F − n R , n E n G ) (19.6)
To attain rank sufficiency, n E n G must equal or exceed n F − n R :
n E nG ≥ n F − n R (19.7)
from which the appropriate Gauss integration rule can be selected.
In the plane stress problem, n E = 3 because E is a 3 × 3 matrix of elastic moduli; see equation
(14.5)2 . Also n R = 3. Consequently r = min(n F − 3, 3n G ) and 3n G ≥ n F − 3.

19–8
 §19.4 STABILITY

Table 19.1 Rank-sufficient Gauss Rules for Some Plane Stress Elements

Element n nF nF − 3 Min n G Recommended rule

3-node triangle 3 6 3 1 centroid∗

6-node triangle 6 12 9 3 3-point rules∗
10-node triangle 10 20 17 6 6-point rule∗
4-node quadrilateral 4 8 5 2 2x2
8-node quadrilateral 8 16 13 5 3x3
9-node quadrilateral 9 18 15 5 3x3
16-node quadrilateral 16 32 29 10 4x4
∗
These triangle integration rules are introduced in §24.2.

3 3 3 3 3

4 4 4 4 4

1 2 1 2 1 2 1 2 1 2

Figure 19.6. Effect of displacing node 4 of the four-node bilinear quadrilateral shown on the
leftmost picture, to the right.

Remark 19.3. The fact that each Gauss point adds n E n G to the rank can be proven considering the following
property. Let B be a n E ×n F rectangular real matrix with rank r B ≤ n E , and E an n E ×n E positive-definite (p.d.)
symmetric matrix. Then the rank of BT E B is r B . Proof: let u
= 0 be a non-null n F -vector. If BT E B u = 0
then 0 = uT BT E B u = ||E1/2 B u||. Therefore B u = 0. Identify now B and E with the strain-displacement
and stress-strain (constitutive) matrix, respectively. In the plane stress case n E = 3, n F = 2n > 3 is the
number of element freedoms. Thus B has rank 3 and a fortiori BT E B must also have rank 3 since E is p.d.
At each Gauss point i a contribution of wi BT E B, which has rank 3 if wi > 0, is added to Ke . By a theorem
of linear algebra, the rank of Ke increases by 3 until it reaches n F − n R .

Example 19.5. Consider a plane stress 6-node quadratic triangle. Then n F = 2 × 6 = 12. To attain the proper
rank of 12 − n R = 12 − 3 = 9, n G ≥ 3. A 3-point Gauss rule, such as the midpoint rule defined in §24.2,
makes the element rank sufficient.

Example 19.6. Consider a plane stress 9-node biquadratic quadrilateral. Then n F = 2 × 9 = 18. To attain
the proper rank of 18 − n R = 18 − 3 = 15, n G ≥ 5. The 2 × 2 product Gauss rule is insufficient because
n G = 4. Hence a 3 × 3 rule, which yields n G = 9, is required to attain rank sufficiency.

Table 19.1 collects rank-sufficient Gauss integration rules for some widely used plane stress elements
with n nodes and n F = 2n freedoms.

19–9
Chapter 19: FEM CONVERGENCE REQUIREMENTS

§19.4.2. Jacobian Positiveness

The geometry of the element should be such that the determinant J = det J of the Jacobian matrix
defined4 in §17.2, is positive everywhere. As illustrated in Equation (17.20), J characterizes the
local metric of the element natural coordinates.

3 3 3

7 7 7

4 6 4 6 4 6
9 9 9
8 8 8

1 5 2 1 5 2 1 5 2
3 3 3

7 7 7

4 6 4 6 4 6
9 9 9
8 8 8

1 5 2 1 5 2 1 5=2

Figure 19.7. Effect of moving midpoint 5 of a 9-node biquadratic quadrilateral

tangentially toward corner 2.

For a three-node triangle J is constant and in fact equal to 2A. The requirement J > 0 is equivalent
to saying that corner nodes must be positioned and numbered so that a positive area A > 0 results.
This is called a convexity condition. It is easily checked by a finite element program.
But for 2D elements with more than 3 nodes distortions may render portions of the element metric
negative. This is illustrated in Figure 19.6 for a 4-node quadrilateral in which node 4 is gradually
moved to the right. The quadrilateral gradually morphs from a convex figure into a nonconvex
one. The center figure is a triangle; note that the metric near node 4 is badly distorted (in fact
J = 0 there) rendering the element unacceptable. This clearly contradicts the erroneous advice of
some FE books, which state that quadrilaterals can be reduced to triangles as special cases, thereby
rendering triangular elements unnecessary.
For higher order elements proper location of corner nodes is not enough. The non-corner nodes
(midside, interior, etc.) must be placed sufficiently close to their natural locations (midpoints,
4 This definition applies to quadrilateral elements. The Jacobian determinant of an arbitrary triangular element is defined
in §24.2.

19–10
 §19. Notes and Bibliography

3 3 3 3

6 5
6 5
6 5
6 5
4
4 1 2
1 2 1 2 1 4 2 4

6 5
3 3
6 5 3
6 5

1 2 1 2
1 2

Figure 19.8. Effect of displacing midpoints 4, 5 and 6 of an equilateral 6-node triangle along
the midpoint normals. Motion is inwards in first two top frames, outwards in the last four. In
the lower leftmost picture nodes 1 through 6 lie on a circle.

centroids, etc.) to avoid violent local distortions. The effect of midpoint motions in quadratic
elements is illustrated in Figures 19.7 and 19.8.
Figure 19.7 depicts the effect of moving midside node 5 tangentially in a 9-node quadrilateral
element while keeping all other 8 nodes fixed. When the location of 5 reaches the quarter-point of
side 1-2, the metric at corner 2 becomes singular in the sense that J = 0 there. Although this is
disastrous in ordinary FE work, it has applications in the construction of special “crack” elements
for linear fracture mechanics.
Displacing midside nodes normally to the sides is comparatively more forgiving, as illustrated in
Figure 19.8. This depicts a 6-node equilateral triangle in which midside nodes 4, 5 and 6 are moved
inwards and outwards along the normals to the midpoint location. As shown in the lower left
picture, the element may be even morphed into a “parabolic circle” (meaning that nodes 1 through
6 lie on a circle) without the metric breaking down.

Notes and Bibliography

The literature on the mathematics of finite element methods has grown exponentially since the monograph of
Strang and Fix [705]. This is very readable but out of print. A more up-to-date exposition is the textbook by
Szabo and Babuska [721]. The subjects collected in this Chapter tend to be dispersed in recent monographs
and obscured by overuse of functional analysis.

19–11
Chapter 19: FEM CONVERGENCE REQUIREMENTS

References
Referenced items have been moved to Appendix R.

19–12
 Exercises

Homework Exercises for Chapter 19

FEM Convergence Requirements

EXERCISE 19.1 [D:20] Explain why the two-dimensional meshes pictured in Figure 19.2(d,e,f) fail interele-
ment compatibility although nodes and DOFs match.

EXERCISE 19.2 [A:20] The isoparametric definition of the straight 3-node bar element in its local system x̄
is     
1 1 1 1 N1e (ξ )
x̄ = x̄1 x̄2 x̄3 N2e (ξ ) . (E19.1)
v̄ ū 1 ū 2 ū 3 N3e (ξ )
Here ξ is the isoparametric coordinate that takes the values −1, 1 and 0 at nodes 1, 2 and 3, respectively, while
N1e , N2e and N3e are the shape functions found in Exercise 16.3 and listed in (E16.2).
For simplicity, take x̄1 = 0, x̄2 = L, x̄3 = 12 L + αL. Here L is the bar length and α a parameter that
characterizes how far node 3 is away from the midpoint location x̄ = 12 L. Show that the minimum α’s
(minimal in absolute value sense) for which J = d x̄/dξ vanishes at a point in the element are ±1/4 (the
quarter-points). Interpret this result as a singularity by showing that the axial strain becomes infinite at a an
end point. (This result has application in fracture mechanics modeling.)

EXERCISE 19.3 [A:15] Consider one dimensional bar-like elements with n nodes and 1 degree of freedom
per node so n F = n. The correct number of rigid body modes is 1. Each Gauss integration point adds 1 to
the rank; that is N E = 1. By applying (19.7), find the minimal rank-preserving Gauss integration rules with
p points in the longitudinal direction if the number of node points is n = 2, 3 or 4.

EXERCISE 19.4 [A:20] Consider three dimensional solid “brick” elements with n nodes and 3 degrees of
freedom per node so n F = 3n. The correct number of rigid body modes is 6. Each Gauss integration point
adds 6 to the rank; that is, N E = 6. By applying (19.7), find the minimal rank-preserving Gauss integration
rules with p points in each direction (that is, 1×1×1, 2×2×2, etc) if the number of node points is n = 8, 20,
27, or 64. Partial answer: for n = 27 the minimal rank preserving rule is 3 × 3 × 3.

EXERCISE 19.5 [A/C:35] (Requires use of a CAS help to be tractable). Repeat Exercise 19.2 for a 9-node
plane stress element. The element is initially a perfect square, nodes 5,6,7,8 are at the midpoint of the sides
1–2, 2–3, 3–4 and 4–1, respectively, and 9 at the center of the square. Displace 5 tangentially towards 2 until
the Jacobian determinant at 2 vanishes. This result is important in the construction of “singular elements” for
fracture mechanics.

EXERCISE 19.6 [A/C:35] Repeat Exercise 19.5 but moving node 5 along the normal to the side. Discuss
the range of motion for which det J > 0 within the element.

EXERCISE 19.7 [A:20] Discuss whether the deVeubeke triangle presented in Chapter 15 satisfies complete-
ness and interelement-compatbility requirements.

19–13
 20
Implementation of
One-Dimensional
Elements

20–1
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

TABLE OF CONTENTS

Page
§20.1 The Plane Bar Element . . . . . . . . . . . . . . . . . 20–3
§20.1.1 Plane Bar Stiffness Matrix . . . . . . . . . . . . 20–3
§20.1.2 Plane Bar Stiffness Module . . . . . . . . . . . . . 20–3
§20.1.3 Testing the Plane Bar Module . . . . . . . . . . . 20–4
§20.2 The Space Bar Element . . . . . . . . . . . . . . . . . 20–5
§20.2.1 SpaceBarStiffness Matrix . . . . . . . . . . . . . 20–6
§20.2.2 Space Bar Stiffness Module . . . . . . . . . . . . 20–6
§20.2.3 Testing the Space Bar Module . . . . . . . . . . . 20–7
§20.3 The Plane Beam-Column Element . . . . . . . . . . . . . 20–8
§20.3.1 Plane Beam-Column Stiffness Matrix . . . . . . . . . 20–8
§20.3.2 Plane Beam Column Stiffness Module . . . . . . . . . 20–9
§20.3.3 Testing the Plane Beam-Column Module . . . . . . . . 20–9
§20.4 *Plane Beam With Offset Nodes . . . . . . . . . . . . . . 20–11
§20.4.1 Plate Reinforced With Edge Beams . . . . . . . . . 20–11
§20.4.2 Rigid Link Transformation . . . . . . . . . . . . . 20–12
§20.4.3 Direct Fabrication of Offset Beam Stiffness . . . . . . . 20–13
§20.5 *Layered Beam Configurations . . . . . . . . . . . . . . 20–14
§20.5.1 Layered Beam With No Interlayer Slip . . . . . . . . 20–15
§20.5.2 Beam Stacks Allowing Interlayer Slip . . . . . . . . . 20–17
§20.6 The Space Beam Element . . . . . . . . . . . . . . . . 20–18
§20.6.1 Space Beam Stiffness Matrix . . . . . . . . . . . . 20–18
§20.6.2 Space Beam Stiffness Module . . . . . . . . . . . 20–20
§20. Notes and Bibliography . . . . . . . . . . . . . . . . . 20–20
§20. Exercises . . . . . . . . . . . . . . . . . . . . . . 20–21

20–2
 §20.1 THE PLANE BAR ELEMENT

This Chapter begins Part III of the course. This Part deals with the computer implementation of
the Finite Element Method for static analysis. It is organized in “bottom up” fashion. It begins
with simple topics, such as programming of bar and beam elements, and gradually builds up toward
more complex models and calculations.
Specific examples of this Chapter illustrate the programming of one-dimensional elements: bars
and beams, using Mathematica as implementation language.

§20.1. The Plane Bar Element

x−
The two-node, prismatic, two-dimensional bar element 2 (x 2 , y2 )
e
was studied in Chapters 2-3 for modeling plane trusses.  =L
It is reproduced in Figure 20.1 for conveniency. It has y−
E, A constant
two nodes and four degrees of freedom. The element ϕ
node displacements and conjugate forces are y
    1 (x 1 , y1 )
u x1 f x1
u  f 
x
ue =  y1  , fe =  y1  . (20.1)
u x2 f x2 Figure 20.1. Plane bar element.
u y2 f y2
The element geometry is described by the coordinates {xi , yi }, i = 1, 2 of the two end nodes. For
stiffness computations, the only material and fabrication properties required are the modulus of
elasticity E = E e and the cross section area A = Ae , respectively. Both are taken to be constant
over the element.

§20.1.1. Plane Bar Stiffness Matrix

The element stiffness matrix in global {x, y} coordinates is given by the explicit expression derived
in §3.1:
 2   
c sc −c2 −sc x21 x21 x21 y21 −x21 x21 −x21 y21
E A  sc s 2 −sc −s 2  E A  x21 y21 y21 y21 −x21 y21 −y21 y21 
Ke =  2 = 3   . (20.2)
 −c −sc c 2
sc  −x21 x21 −x21 y21 x21 x21 x21 y21
−sc −s 2 sc s 2 −x21 y21 −y21 y21 x21 y21 y21 y21

Here c = cos ϕ = x21 /, s = sin ϕ = y21 /, in which x21 = x2 −x1 , y21 = y2 −y1 ,  = x21 2
+ y21
2
,
and ϕ is the angle formed by x̄ and x, measured from x positive counterclockwise — see Figure 20.1.
The second expression in (20.2) is preferable in a computer algebra system because it enhances
simplification possibilities when doing symbolic work, and is the one actually implemented in the
module described below.

§20.1.2. Plane Bar Stiffness Module

The computation of the stiffness matrix Ke of the two-node, prismatic plane bar ele-
ment is done by Mathematica module PlaneBar2Stiffness. This is listed in Fig-
ure IFEM:Ch20:fig:PlaneBarStiffMatrixModule. The module is invoked as

Ke = PlaneBar2Stiffness[ncoor, Em, A, options] (20.3)

20–3
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

PlaneBar2Stiffness[ncoor_,Em_,A_,options_]:= Module[
{x1,x2,y1,y2,x21,y21,EA,numer,L,LL,LLL,Ke},
{{x1,y1},{x2,y2}}=ncoor; {x21,y21}={x2-x1,y2-y1};
EA=Em*A; {numer}=options; LL=x21^2+y21^2; L=Sqrt[LL];
If [numer,{x21,y21,EA,LL,L}=N[{x21,y21,EA,LL,L}]];
If [!numer, L=PowerExpand[L]]; LLL=Simplify[LL*L];
Ke=(Em*A/LLL)*{{ x21*x21, x21*y21,-x21*x21,-x21*y21},
{ y21*x21, y21*y21,-y21*x21,-y21*y21},
{-x21*x21,-x21*y21, x21*x21, x21*y21},
{-y21*x21,-y21*y21, y21*x21, y21*y21}};
Return[Ke]];

Figure 20.2. Mathematica stiffness module for a two-node, prismatic plane bar element.

The arguments are

ncoor Node coordinates of element arranged as { { x1,y1 },{ x2,y2 } }.
Em Elastic modulus.
A Cross section area.
options A list of processing options. For this element is has only one entry: { numer }. This
is a logical flag with the value True or False. If True the computations are carried
out in floating-point arithmetic. If False symbolic processing is assumed.
The module returns the 4 × 4 element stiffness matrix as function value.

ClearAll[A,Em,L];
ncoor={{0,0},{30,40}}; Em=1000; A=5;
Ke= PlaneBar2Stiffness[ncoor,Em,A,{True}];
Print["Numerical Elem Stiff Matrix: "];
Print[Ke//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[N[Ke]]]];
Print["Symmetry check=",Simplify[Chop[Transpose[Ke]-Ke]]];
Numerical Elem Stiff Matrix:
36. 48. − 36. − 48.
48. 64. − 48. − 64.
− 36. − 48. 36. 48.
− 48. − 64. 48. 64.
Eigenvalues of Ke = {200., 0, 0, 0}
Symmetry check={{0, 0, 0, 0}, {0, 0, 0, 0}, {0, 0, 0, 0}, {0, 0, 0, 0}}

Figure 20.3. Test of plane bar stiffness module with numerical inputs.

§20.1.3. Testing the Plane Bar Module

The modules are tested by the scripts listed in Figures 20.3 and 20.4. The script shown on the top
of Figure 20.3 tests a numerically defined element with end nodes located at (0, 0) and (30, 40),
with E = 1000, A = 5, and numer set to True. Executing the script produces the results listed in
the bottom of that figure.

20–4
 §20.2 THE SPACE BAR ELEMENT

ClearAll[A,Em,L];
ncoor={{0,0},{L,0}};
Ke= PlaneBar2Stiffness[ncoor,Em,A,{False}];
kfac=Em*A/L; Ke=Simplify[Ke/kfac];
Print["Symbolic Elem Stiff Matrix: "];
Print[kfac," ",Ke//MatrixForm];
Print["Eigenvalues of Ke=",kfac,"*",Eigenvalues[Ke]];
Symbolic Elem Stiff Matrix:
1 0 −1 0
A Em 0 0 0 0
L −1 0 1 0
0 0 0 0
A Em
Eigenvalues of Ke = ∗{0, 0, 0, 2}
L

Figure 20.4. Test of plane bar stiffness module with symbolic inputs.

On return from PlaneBar2Stiffness, the stiffness matrix returned in Ke is printed. Its four
eigenvalues are computed and printed. As expected three eigenvalues, which correspond to the
three independent rigid body motions of the element, are zero. The remaining eigenvalue is positive
and equal to E A/. The symmetry of Ke is checked by printing (Ke )T − Ke upon simplification
and chopping.
The script of Figure 20.4 tests a symbolically defined bar element with end nodes located at (0, 0)
and (L , 0), which is aligned with the x axis. Properties E and A are kept symbolic. Executing the
script shown in the top of Figure 20.4 produces the results shown in the bottom of that figure. One
thing to be noticed is the use of the stiffness scaling factor E A/, called kfac in the script. This is
a symbolic quantity that can be extracted as factor of matrix Ke . The effect is to clean up matrix
and vector output, as can be observed in the printed results.

§20.2. The Space Bar Element

To show how the previous implementation extends easily to three dimensions, this section describes
the implementation of the space bar element.
The two-node, prismatic, space bar
element is pictured in Figure 20.5. −
y
The element has two nodes and six
degrees of freedom. The element node −
z 1 (x1 ,y1 ,z 1 ) E, A constant
displacements and conjugate forces
are arranged as
y
    2 (x 2 ,y 2 ,z 2 )
u x1 f x1
 u y1 
 
 f y1 
   = Le x−
 u   f  x
ue =  z1  , fe =  z1  .
 u x2   f x2  z Global system
   
u y2 f y2
u z2 f z2 Figure 20.5. The space (3D) bar element.
(20.4)

20–5
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

SpaceBar2Stiffness[ncoor_,Em_,A_,options_]:=Module[
{x1,x2,y1,y2,z1,z2,x21,y21,z21,EA,numer,L,LL,LLL,Ke},
{{x1,y1,z1},{x2,y2,z2}}=ncoor;{x21,y21,z21}={x2-x1,y2-y1,z2-z1};
EA=Em*A; {numer}=options; LL=x21^2+y21^2+z21^2; L=Sqrt[LL];
If [numer,{x21,y21,z21,EA,LL,L}=N[{x21,y21,z21,EA,LL,L}]];
If [!numer, L=PowerExpand[L]]; LLL=Simplify[LL*L];
Ke=(Em*A/LLL)*
{{ x21*x21, x21*y21, x21*z21,-x21*x21,-x21*y21,-x21*z21},
{ y21*x21, y21*y21, y21*z21,-y21*x21,-y21*y21,-y21*z21},
{ z21*x21, z21*y21, z21*z21,-z21*x21,-z21*y21,-z21*z21},
{-x21*x21,-x21*y21,-x21*z21, x21*x21, x21*y21, x21*z21},
{-y21*x21,-y21*y21,-y21*z21, y21*x21, y21*y21, y21*z21},
{-z21*x21,-z21*y21,-z21*z21, z21*x21, z21*y21, z21*z21}};
Return[Ke];
];

Figure 20.6. Module to form the stiffness of the space (3D) bar element.

The element geometry is described by the coordinates {xi , yi , z i }, i = 1, 2 of the two end nodes.
As in the case of the plane bar, the two properties required for the stiffness computations are the
modulus of elasticity E and the cross section area A. Both are assumed to be constant over the
element.

§20.2.1. SpaceBarStiffness Matrix

 space bar element, introduce the notation x21 = x2 − x1 , y21 = y2 − y1 , z 21 = z 2 − z 1 and

For the
 = x21 2
+ y21
2
+ z 21
2
. It can be shown1 that the element stiffness matrix in global coordinates is
given by
 
x21 x21 x21 y21 x21 z 21 −x21 x21 −x21 y21 −x21 z 21
 x21 y21 y21 y21 x21 z 21 −x21 y21 −y21 y21 −y21 z 21 
e e  
E A  x21 z 21 y21 z 21 z 21 z 21 −x21 z 21 −y21 z 21 −z 21 z 21 
Ke =  . (20.5)
 3  −x x
21 21 −x 21 y21 −x21 z 21 x21 x21 x21 y21 x21 z 21 
 
−x21 y21 −y21 y21 −x21 z 21 x21 y21 y21 y21 y21 z 21
−x21 z 21 −y21 z 21 −z 21 z 21 x21 z 21 y21 z 21 z 21 z 21
This matrix expression in terms of coordinate differences is useful in symbolic work, because it
enhances simplification possibilities.

§20.2.2. Space Bar Stiffness Module

The computation of the stiffness matrix Ke of the two-node, prismatic space bar element, is done by
Mathematica module SpaceBar2Stiffness. This is listed in Figure 20.6. The module is invoked
as
Ke = SpaceBar2Stiffness[ncoor, Em, A, options] (20.6)
The arguments are
ncoor Node coordinates of element arranged as { { x1,y1,z1 },{ x2,y2,z2 } }.

1 The derivation was the subject of Exercise 6.10.

20–6
 §20.2 THE SPACE BAR ELEMENT

ClearAll[A,Em];
ncoor={{0,0,0},{2,3,6}}; Em=343; A=10;
Ke= SpaceBar2Stiffness[ncoor,Em,A,{True}];
Print["Numerical Elem Stiff Matrix: "];
Print[Ke//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[Ke]]];
Numerical Elem Stiff Matrix:
40. 60. 120. − 40. − 60. − 120.
60. 90. 180. − 60. − 90. − 180.
120. 180. 360. − 120. − 180. − 360.
− 40. − 60. − 120. 40. 60. 120.
− 60. − 90. − 180. 60. 90. 180.
− 120. − 180. − 360. 120. 180. 360.
Eigenvalues of Ke = {980., 0, 0, 0, 0, 0}

Figure 20.7. Testing the space bar stiffness module with numerical inputs.

Em Elastic modulus.
A Cross section area.
options A list of processing options. For this element is has only one entry: { numer }. This
is a logical flag with the value True or False. If True the computations are carried
out in floating-point arithmetic. If False symbolic processing is assumed.
The module returns the 6 × 6 element stiffness matrix as function value.

ClearAll[A,Em,L];
ncoor={{0,0,0},{L,2*L,2*L}/3};
Ke= SpaceBar2Stiffness[ncoor,Em,A,{False}];
kfac=Em*A/(9*L); Ke=Simplify[Ke/kfac];
Print["Symbolic Elem Stiff Matrix: "];
Print[kfac," ",Ke//MatrixForm];
Print["Eigenvalues of Ke=",kfac,"*",Eigenvalues[Ke]];
Symbolic Elem Stiff Matrix:
1 2 2 −1 −2 −2
2 4 4 −2 −4 −4
A Em 2 4 4 −2 −4 −4
9L −1 −2 −2 1 2 2
−2 −4 −4 2 4 4
−2 −4 −4 2 4 4
A Em
Eigenvalues of Ke = ∗ {0, 0, 0, 0, 0, 18}
9L

Figure 20.8. Testing the space bar stiffness module with symbolic inputs.

§20.2.3. Testing the Space Bar Module

The modules are tested by the scripts listed in Figures 20.7 and 20.8. As these are similar to previous
tests done on the plane bar they need not be described in detail.

20–7
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

x−
u−x2
−
y (a) (b)
θz2
2 (x 2 , y2 )

u−x1 θz1
u−y2 y−
u−y1
ϕ
1 2 x− y
e
E, A, I constant 1 (x 1 , y1 )
 =L zz
x
Figure 20.9. Plane beam-column element: (a) in its local system; (b) in the global system.

The script of Figure 20.7 tests a numerically defined space bar with end nodes located at (0, 0, 0)
and (30, 40, 0), with E = 1000, A = 5 and numer set to True. Executing the script produces the
results listed in the bottom of that Figure.
The script of Figure 20.8 tests a symbolically defined bar element with end nodes located at (0, 0, 0)
and (L , 2L , 2L)/3, which has length L and is not aligned with the x axis. The element properties
E and A are kept symbolic. Executing the script produces the results shown in the bottom of that
Figure. Note the use of a stiffness factor kfac of E A/(9) to get cleaner printouts.

§20.3. The Plane Beam-Column Element

Beam-column elements model structural members that resist both axial and bending actions. This
is the case in skeletal structures such as frameworks which are common in steel and reinforced-
concrete building construction. A plane beam-column element is a combination of a plane bar
(such as that considered in §20.1), and a plane beam.
We consider a beam-column element in its local system (x̄, ȳ) as shown in Figure 20.9(a), and then
in the global system (x, y) as shown in Figure 20.9(b). The six degrees of freedom and conjugate
node forces of the elements are:
   ¯     
ū x1 f x1 u x1 f x1
 ū y1   f¯y1   u y1   f y1 
       
 θ  e  m   θ  m 
ūe =  z1  , f̄ =  z1  , ue =  z1  , fe =  z1  . (20.7)
 ū x2   ū x2   u x2   f x2 
       
ū y2 ū y2 u y2 f y2
θz2 m z2 θz2 m z2

The rotation angles θ and the nodal moments m are the same in the local and the global systems
because they are about the z axis, which does not change in passing from local to global.
The element geometry is described by the coordinates {xi , yi }, i = 1, 2 of the two end nodes. The
element length is  = L e . Properties involved in the stiffness calculations are: the modulus of
elasticity E, the cross section area A and the moment of inertia I = Izz about the neutral axis. All
properties are taken to be constant over the element.

§20.3.1. Plane Beam-Column Stiffness Matrix

20–8
 §20.3 THE PLANE BEAM-COLUMN ELEMENT

To obtain the plane beam-column stiffness in the local system we simply add the stiffness matrices
derived in Chapters 11 and 12, respectively, to get
   
1 0 0 −1 0 0 0 0 0 0 0 0
 0 0 0 0 0  12 6 0 −12 6 
  EI  
e E A  0 0 0 0  42 0 −6 22 
K̄ =   +   (20.8)
  1 0 0 3  0 0 0 
   
0 0 12 −6
symm 0 symm 42

The two matrices on the right of (20.8) come from the bar stiffness (12.22) and the Bernoulli-
Euler bending stiffness (13.20), respectively. Before adding them, rows and columns have been
rearranged in accordance with the nodal freedoms (20.7).
The displacement transformation matrix between local and global systems is
    
ū x1 c s 0 0 0 0 u x1
 ū y1   −s c 0 0 0 0   u y1 
    
 θz1   0 0 1 0 0 0   θz1 
ū = 
e
=   = T ue , (20.9)
 u x2   0 0 0 c s 0   u x2 
    
ū y2 0 0 0 −s c 0 u y2
θz2 0 0 0 0 0 1 θz2
in which c = cos ϕ = (x2 − x1 )/, s = sin ϕ = (y2 − y1 )/, and ϕ is the angle between x̄ and x,
measured positive-counterclockwise from x; see Figure IFEM:Ch20:fig:PlaneBeamColElem. The
stiffness matrix in the global system is obtained through the congruent transformation
e
Ke = TT K̄ T. (20.10)

Explicit expressions of the entries of Ke are messy. Unlike the bar, it is better to let the program do
the transformation.

§20.3.2. Plane Beam Column Stiffness Module

The computation of the stiffness matrix Ke of the two-node, prismatic plane beam-column element
is done by Mathematica module PlaneBeamColumn2Stiffness. This is listed in Figure 20.10.
The module is invoked as

Ke = PlaneBeamColumn2Stiffness[ncoor, Em, { A,Izz }, options] (20.11)

The arguments are

ncoor Node coordinates of element arranged as { { x1,y1 },{ x2,y2 } }.
Em Elastic modulus.
A Cross section area.
Izz Moment of inertia of cross section area respect to axis z.
options A list of processing options. For this element is has only one entry: { numer }. This
is a logical flag with the value True or False. If True the computations are carried
out in floating-point arithmetic. If False symbolic processing is assumed.
The module returns the 6 × 6 element stiffness matrix as function value.

20–9
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

PlaneBeamColumn2Stiffness[ncoor_,Em_,{A_,Izz_},options_]:= Module[
{x1,x2,y1,y2,x21,y21,EA,EI,numer,L,LL,LLL,Te,Kebar,Ke},
{{x1,y1},{x2,y2}}=ncoor; {x21,y21}={x2-x1,y2-y1};
EA=Em*A; EI=Em*Izz; {numer}=options;
LL=Simplify[x21^2+y21^2]; L=Sqrt[LL];
If [numer,{x21,y21,EA,EI,LL,L}=N[{x21,y21,EA,EI,LL,L}]];
If [!numer, L=PowerExpand[L]]; LLL=Simplify[LL*L];
Kebar= (EA/L)*{
{ 1,0,0,-1,0,0},{0,0,0,0,0,0},{0,0,0,0,0,0},
{-1,0,0, 1,0,0},{0,0,0,0,0,0},{0,0,0,0,0,0}} +
(2*EI/LLL)*{
{ 0,0,0,0,0,0},{0, 6, 3*L,0,-6, 3*L},{0,3*L,2*LL,0,-3*L, LL},
{ 0,0,0,0,0,0},{0,-6,-3*L,0, 6,-3*L},{0,3*L, LL,0,-3*L,2*LL}};
Te={{x21,y21,0,0,0,0}/L,{-y21,x21,0,0,0,0}/L,{0,0,1,0,0,0},
{0,0,0,x21,y21,0}/L,{0,0,0,-y21,x21,0}/L,{0,0,0,0,0,1}};
Ke=Transpose[Te].Kebar.Te;
Return[Ke] ];

Figure 20.10. Mathematica module to form the stiffness matrix of a two-node,

prismatic plane beam-column element.

ClearAll[L,Em,A,Izz];
ncoor={{0,0},{3,4}}; Em=100; A=125; Izz=250;
Ke= PlaneBeamColumn2Stiffness[ncoor,Em,{A,Izz},{True}];
Print["Numerical Elem Stiff Matrix: "];
Print[Ke//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[Ke]]];
Numerical Elem Stiff Matrix:
2436. 48. −4800. −2436. −48. −4800.
48. 2464. 3600. −48. −2464. 3600.
−4800. 3600. 20000. 4800. −3600. 10000.
−2436. −48. 4800. 2436. 48. 4800.
−48. −2464. −3600. 48. 2464. −3600.
−4800. 3600. 10000. 4800. −3600. 20000.
Eigenvalues of Ke = {34800., 10000., 5000., 0, 0, 0}

Figure 20.11. Test of two-node plane beam-column element with numeric inputs.

§20.3.3. Testing the Plane Beam-Column Module

The beam-column stiffness are tested by the scripts shown in Figures 20.11 and 20.12.
The script at the top of Figure 20.11 tests a numerically defined element of length  = 5 with
end nodes located at (0, 0) and (3, 4), respectively, with E = 100, A = 125 and Izz = 250. The
output is shown at the bottom of that figure. The stiffness matrix returned in Ke is printed. Its
six eigenvalues are computed and printed. As expected three eigenvalues, which correspond to the
three independent rigid body motions of the element, are zero. The remaining three eigenvalues
are positive.
The script at the top of Figure 20.12 tests a plane beam-column of length L with end nodes at (0, 0)
and (3L/5, 4L/5). The properties E, A and Izz are kept in symbolic form. The output is shown
at the bottom of that figure. The printed matrix looks complicated because bar and beam coupling

20–10
 §20.4 *PLANE BEAM WITH OFFSET NODES

occurs when the element is not aligned with the global axes. The eigenvalues are obtained in closed
symbolic form, and their simplicity provides a good check that the transformation matrix (20.9) is
orthogonal. Three eigenvalues are exactly zero; one is associated with the axial (bar) stiffness and
two with the flexural (beam) stiffness.

ClearAll[L,Em,A,Izz];
ncoor={{0,0},{3*L/5,4*L/5}};
Ke= PlaneBeamColumn2Stiffness[ncoor,Em,{A,Izz},{False}];
Print["Symbolic Elem Stiff Matrix:"]; kfac=Em;
Ke=Simplify[Ke/kfac]; Print[kfac," ",Ke//MatrixForm];
Print["Eigenvalues of Ke=",kfac,"*",Eigenvalues[Ke]];
Symbolic Elem Stiff Matrix:
3 ( 64 Izz +3 A L2 ) 12 ( −12 Izz +A L2 ) 3 ( 64 Izz +3 A L2 ) ( −12 Izz +A L2 )
25 L3 25 L3
− 24 Izz
5 L2
− 25 L3
− 12 25 L3
− 24 Izz
5 L2
12 ( −12 Izz +A L2 ) 4 ( 27 Izz +4 A L2 ) 12 ( −12 Izz +A L2 ) Izz +4 A L2 )
25 L3 25 L3
18 Izz
5 L2
− 25 L3
− 4 ( 27 25 L3
18 Izz
5 L2
− 245 LIzz2 18 Izz
5 L2
4 Izz
L
24 Izz
5 L2
− 18 Izz
5 L2
2 Izz
L
Em
3 ( 64 Izz +3 A L2 ) 12 ( −12 Izz +A L2 ) 3 ( 64 Izz +3 A L2 ) 12 ( −12 Izz +A L2 )
− 25 L3
− 25 L3
24 Izz
5 L2 25 L3 25 L3
24 Izz
5 L2

− 12 ( −12 Izz +A L2 ) 4 ( 27 Izz +4 A L2 ) 12 ( −12 Izz +A L2 ) 4 ( 27 Izz +4 A L2 ) −

25 L3
− 25 L3
− 18 Izz
5 L2 25 L3 25 L3
18 Izz
5 L2
− 245 LIzz2 18 Izz
5 L2
2 Izz
L
24 Izz
5 L2
− 185 LIzz2 4 Izz
L
2
Eigenvalues of Ke = Em∗{0, 0, 0, 2A , 2 Izz , 6 (4 Izz +3 Izz L ) }
L L L

Figure 20.12. Test of two-node plane beam-column element with symbolic inputs.

§20.4. *Plane Beam With Offset Nodes

§20.4.1. Plate Reinforced With Edge Beams
Consider a plate reinforced with edge beams, as shown in Figure 20.13(a). The conventional placement of the
nodes is at the plate midsurface and beam longitudinal (centroidal) axis. But those element centered locations
do not coincide. To assemble the structure it is necessary to refer both the plate and beam stiffness equations
to common locations, because such equations are only written at nodes. We assume that those connection
nodes, or simply connectors, will be placed at the plate midsurface, as sketched in Figure 20.13(b). With that
choice there is no need to change the plate equations. The beam connectors have been moved, however, from
their original centroidal positions. For the beam these connectors are also known as offset nodes.
Structural configurations such as that of Figure 20.13(a) are common in aerospace, civil and mechanical
engineering when shells or plates are reinforced with eccentric stiffeners.
The process of moving the beam stiffness equation to the offset nodes is called offsetting. It relies on setting up
multifreedom constraints (MFC) between centered and offset node freedoms, and applying the master-slave
congruential transformation introduced in Chapter 8. For simplicity we discuss only this process assuming
that the beam of Figure 20.13(b) is a plane beam whose freedoms are to be moved upwards by a distance
d, which is positive if going upward from beam centroid. Freedoms at connection and centroidal nodes are
declared to be master and slaves, respectively. They are labeled as shown in Figure 20.13(c,d). The original
stiffness equations referred to centroidal (slave) freedoms are

20–11
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

(a) plate (b)

plate

d
edge beam

edge beams

d slave-to-master rigid-link offset distance for beam

beam centroidal node → slave node
_ _ connection node → master node
yc , ym (c)
u−y2m
u−y1m
1m master nodes 2m _
xm θz1m (d) θ z2m u−x2m
d rigid links d _ u− x1m
xc 1m 2
u−y1c u−y2c
m
1c centroidal (slave) nodes 2c d d
E, A, Izzc constant θ z1c u−x1c θ z2c u−x2c
1c 2c
Le
Figure 20.13. Plane beam with nodes offset for a rigid-link connection to plate.

A 0 0 −A 0 0 
12Izzc 6Izzc 0 − 12Izzc − 6Izzc    ¯ 
  ū x1s f x1s
 (L e )2 (L e )2 (L e )2 (L e )2   ū   f¯ 
   y1s   y1s 
E 
 4Izzc 0 − 6Izzc 2Izzc  θ  m 
Le (L e )2 Le   z1s   z1s 
  = , or Kec uec = fec , (20.12)
Le  A 0 0   ū x2s 
  f¯x2s 

 
 12Izzc − 6Izzc  ū y2s   f¯y2s 
 (L e )2 (L e )2 
θz2s m z2s
symm 4Izzc
Le
in which A is the beam cross section area while Izzc denotes the section moment of inertia with respect to the
centroidal axis z̄ c .

§20.4.2. Rigid Link Transformation

Kinematic constraints between master and centroidal (slave) freedoms are obtained assuming that they are
connected by rigid links as pictured in Figure 20.13(c,d). This gives the centroidal(slave)-to-master transfor-
mation  ū   1 0 d 0 0 0   ū 
x1c x1m
 ū y1c   0 1 0 0 0 0   y1m 
ū
 θz1c   0 0 1 0 0 0   θz1m 
    
 ū x2c  =  0 0 0 1 0 d   ū x2m  or us = Tsm um . (20.13)
e e

    
ū y2c 0 0 0 0 1 0 ū y2m
θz2c 0 0 0 0 0 1 θz2m
The inverse transformation: Tmc = T−1 cm is obtained on replacing d with −d, as is physically obvious. The
T
modified stiffness equations are obtained by the congruential transformation: Tcm Kec Tcm
T
= Tcm fc = fem ,
T e

which yields

20–12
 §20.4 *PLANE BEAM WITH OFFSET NODES

 A 0 d −A 0 −d 
12 I 6 I
   ¯ 
 0 zzc zzc
0 − 12 Ie zzc 6 Izzc  ū x1m f x1m
 (L e )2 Le (L )2 L e
  ū   f¯ 
   y1m   y1m 
 6 Izzc 4 I +A d 2 −d − 6 Izzc 2 
E  d Le zzc
Le
2 I zzc −A d   θz1m   
 =  m z1m 
  (20.14)
L e  −A 0 −d A 0 d   ū x2m   ¯ 
  f x2m 
     f¯y2m 
 0 − e zzc2
12 I − 6 Izzc 0 12 I zzc 6
− zzc I  ū y2m
 (L ) Le (L e )2 Le 
θz2m m z2m
−d 6 Izzc 2 Izzc −A d 2 d − 6 Izzc 4 Izzc +A d 2
Le L e

Note that the modified equations are still referred to the local system {x̄m , ȳm } pictured in Figure 20.13(c).
Prior to assembly they should be transformed to the global system {x, y} prior to assembly.

The foregoing transformation procedure has a flaw: for standard plate elements it will introduce compatibility
errors at the interface between plate and beam. This may cause the beam stiffness to be significantly under-
estimated. See the textbook by Cook et. al. [149] for an explanation, and references therein. The following
subsections describes a different scheme that builds Kem directly and cures that incompatibility.

§20.4.3. Direct Fabrication of Offset Beam Stiffness

This approach directly interpolates displacements and strains from master nodes placed at distance d from the
beam longitudinal (centroidal) axis, as pictured in Figure 20.14. As usual the isoparametric coordinate ξ along
the beam element varies from −1 at node 1 through +1 at node 2. The following cross section geometric
properties are defined for use below:

A= d A, Sz = ȳ d A = A d, Izzc = ȳc2 d A, Izzm = ȳ 2 d A = Izzc + A d 2 , (20.15)
Ae Ae Ae Ae

The inplane displacements are expressed in term of the master freedoms at nodes 1m and 2m . Using the
Bernoulli-Euler model gives

 ū 
x1m

  ∂N ∂N  ū y1m 
− ȳ ∂ x̄uy1 − ȳ ∂ N θ z1
− ȳ ∂ x̄uy2 − ȳ ∂ N θ z2  θz1m 
ū xm
=
Nux1 ∂ x̄ Nux2 ∂ x̄   (20.16)
ū ym  ū x2m 
0 Nuy1 Nθ z1 0 Nuy2 Nθ z2  
ū y2m
θz2m

in which Nux1 = 12 (1 − ξ )/2, Nux2 = 12 (1 + ξ )/2, Nuy1 = 14 (1 − ξ )2 (2 + ξ ), Nθ z1 = 18 (1 − ξ )2 (1 + ξ ),

Nuy2 = 14 (1 + ξ )2 (2 − ξ ), Nθ z2 = − 18 (1 + ξ )2 (1 − ξ ) are the usual bar and beam shape functions, but here
referred to the offset axis x̄m .

The axial strain is ex x = ∂u x /∂ x̄ and the strain energy U e = 1
2 Ve
E ex2 x d V where d V = A d x̄=A. Carrying

20–13
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

_
ym
u−y1m Master node offset line u−y2m _
θz1m u−x1m θ z2m u−x2m xm _
yc
1m master (connector) nodes 2m
d d _ Neutral axis
_ zc
xc
E, A, Izzc constant Beam Cross section
cross centroid
Le section
ξ=1 Symmetry plane
ξ = −1
Figure 20.14. Plane beam fabricated directly from offset master node freedoms.

out the integral and differentiating twice with respect to the degrees of freedom yields the stiffness matrix
 A 0 −A d −A 0 Ad 
 0 12 (Izzc +A d ) 2
6 (Izzc +A d )
2
−12 (Izzc +A d ) 2
6 (Izzc +A d ) 
2

 0 
(L e )2 Le (L e )2 Le
 
 6 (Izzc +A d ) 2
−6 (Izzc +A d ) 2 
E  −A d 4 (Izzc +A d )
2
Ad 2 (Izzc +A d ) 
2
K = e
e

Le Le 

L  −A 0 Ad A 0 −A d 
 −12 (Izzc +A d ) −6 (Izzc +A d )
2 2
12 (Izzc +A d ) 2 2 
−6 (Izzc +A d ) 
 0 0
 (L e )2 Le (L e )2 Le 
6 (Izzc +A d ) 2
−6 (Izzc +A d ) 2
Ad e 2 (Izzc +A d 2 ) −A d 4 (Izzc +A d 2 )
 1 L Le 
0 −d −1 0 d
 0 12 (r G2 + d 2 ) 6 (r G2 + d 2 ) −12 (r G2 + d 2 ) 6 (r G2 + d 2 ) 
 (L ) e 2 e 0
(L ) e 2 Le 
 L 
 6 (r G2 + d 2 ) −6 (r + ) 
 2 (r G2 + d 2 ) 
2 2
d
E A  −d e 4 (r G2 + d 2 ) d G
e 
= e  L L 
L  −1 0 d 1 0 −1 d 
 
 0 −12 (r G2 + d 2 ) −6 (r G2 + d 2 ) 0 12 (r G2 + d 2 ) −6 (r G2 + d 2 ) 
 (L ) e 2 e
(L ) e 2 e 
 L L 
6 (r G2 + d 2 ) −6 (r 2
+ d 2
)
d 2 (r G2 + d 2 ) −d G
4 (r G2 + d 2 )
Le Le
(20.17)
In the second form, r G2 = Izzc /A is the squared radius of gyration of the cross section about z.
Comparing the first form of Ke in (20.17) with (20.14) shows that the bending terms are significantly different
if d
= 0. These differences underscore the limitations of the rigid link assumption.

§20.5. *Layered Beam Configurations

Another application of rigid link constraints is to modeling of layered beams, also called composite beams
as well as beam stacks. These are beam-columns fabricated with beam layers that are linked to operate
collectively as a beam member. In this section we study how to form the stiffness equations of beam stacks
under the following assumptions:
1. The overall cross section of the beam member is rectangular and prismatic (that is, the cross section is
constant along the longitudinal direction).
2. Both the beam stack and each of its constituent layers are prismatic.

20–14
 §20.5 *LAYERED BEAM CONFIGURATIONS

3. The layers are of homogeneous isotropic material. Layer material, however, may vary from layer to
layer.
The key modeling assumption is: is interlayer slip allosed or not? The two cases are studied next.

§20.5.1. Layered Beam With No Interlayer Slip

The main modeling constraint here is: if all layers are of the same material, the stiffness equations should
reduce to those of a homogenous beam. To discuss how to meet this requirement it is convenient to introduce
a beam template that separates the stiffness matrix into basic and higher order components. Consider a
homogeneous, isotropic prismatic beam column element with elastic modulus E, cross section area A and
cross section second moment of inertia Ix xc with respect to its neutral (centroidal) axis. The template form of
the stiffness matrix in the local system is
 K 0 0 −K b1 0 0
 0
 0 0 0 0 0

b1
 0 0 0 0 0 0  0 K h3 K h4 0 −K h3 K h4 
 0 0 −K b2   −K h4 K h5 
K̄ = Keb + Keh =   + βh  0  (20.18)
e 0 K b2 0 K h4 K h5 0
 −K b1 0 0 K b1 0 0  0 0 0 0 0 0 
   
0 0 0 0 0 0 0 −K h3 −K h4 0 K h3 −K h4
0 0 −K b2 0 0 K b2 0 K h4 K h5 0 −K h4 K h5
in which K b1 = E A/L e , K b2 = E Izzc /L e , K h3 = 12E Izzc /(L e )3 , K h4 = 6E Izzc /(L e )2 and K h5 = 3E Izzc /L e .
Here βh is a free parameter that scales the higher order stiffness Kh . If βh = 1 we recover the standard beam
column stiffness (20.8). For a rectangular cross section of height H and width h, A = H h and Izzc = H 3 h/12,
and the template (20.18) becomes
 H 0 0 −H 0 0
 0 0 0 0 0 0

 0 0 0 0 0 0   0 4 2L
e
0 −4 2L e 
Eh  0 121 H 3 0 0 − 121 H 3 
 + βh E H h
3  0 2L e (L e )2 0 −2L e (L e )2 
K̄ = e    (20.19)
e 0
L 
 −H 0 0 H 0 0 
 4(L e )3 0 0
 0 0 0 0 

0 0 0 0 0 0 0 −4 −2L 0 4 −2L e
e

0 0 − 121 H 3 0 0 121 H 3 0 2L e (L e )2 0 −2L e (L e )2

Next, cut the foregoing beam into two identical layers of height Hk = H/2, where k = 1, 2 is used as layer
index. See Figure 20.15(b). The layers have area Ak = Hk h = H h/2 and self inertia Ix xk = Hk3 h/12 =
H 3 h/96. The layer stiffness matrices in template form are
 H 0 0 −Hk 0 0
 0 0 0 0 0 0

k
 0 0 0 0 0 0   0 4 2L e
0 −4 2L e 
Eh  0 1 3
0 0 − 12 Hk3 
1
+ βhk E Hk h
3  0 2L e (L e )2 0 −2L e (L e )2 
K̄k = e   ,
e 0 12 Hk
k = 1, 2.
L 
 −Hk 0 0 Hk 0 0 
 4(L e )3 0 0
 0 0 0 0 

0 0 0 0 0 0 0 −4 −2L 0 4 −2L e
e

0 0 − 12
1 1
Hk3 0 0 12 Hk3 0 2L e (L e )2 0 −2L e (L e )2
(20.20)
Beacuse the layers are identical it is reasonable to assume the same higher order free parameter for both layers,
that is, βh1 = βh2 = βh . The offset distances from each layer to the centroid of the full beam to the centroid
of each layer are d1 = −H/4 and d2 = H/4. The rigid-link transformation matrices for (20.20) are the same
as those found in the previous Section:
 1 0 −H/4 0 0 0
  1 0 H/4 0 0 0

0 1 0 0 0 0  0 1 0 0 0 0 
0 0 0  0 0 0 
T1 =  , T2 =  .
1 0 0 1 0 0
0 0 (20.21)
0 1 0 −H/4  0 0 0 1 0 H/4 
   
0 0 0 0 1 0 0 0 0 0 1 0
0 0 0 0 0 1 0 0 0 0 0 1

20–15
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

E, A, Izzc constant

1 1 2 d1=0 1 H (a) 1 layer

 = Le h

2
d2= H/4 2 H/2
1 2 d1=−H/4
(b) 2 layers
1 1 H/2
 = Le h

3
d3= H/3
3 H/3
1 2 2 d1=−H/3
2 H/3 (c) 3 layers
1 1 H/3
d2=0
 = Le h

Figure 20.15. Plane beam divided into identical “sticking” layers.

e e e
Transforming and adding layers contributions as K̄ = T1T K̄1 T1 + T2T K̄2 T2 gives
 H 0 0 −H 0 0
 0 0 0 0 0 0

 0 0 0 0 0 0   0 4 2L e
0 −4 2L e 
Eh  1
H3 0 0 − 121 H 3 
 + βh E H h
3  0 2L e (L e )2 0 −2L e (L e )2 
K̄ = e    . (20.22)
e 0 0 12
L 
 −H 0 0 H 0 0 
 16(L e )3 0 0
 0 0 0 0 

0 0 0 0 0 0 0 −4 −2L 0 4 −2L e
e

0 0 − 121 H 3 0 0 121 H 3 0 2L e (L e )2 0 −2L e (L e )2

This becomes identical to (20.19) if we set βh = 4.
Carrying out the same exercise for three identical layers of height H/3, as shown in Figure 20.15(c), yields
 H 0 0 −H 0 0
 0 0 0 0 0 0

 0 0 0 0 0 0   0 4 2L e
0 −4 2L e 
Eh  1
H3 0 0 − 121 H 3 
 + βh E H h
3  0 2L e (L e )2 0 −2L e (L e )2 
K̄ = e    , (20.23)
e 0 0 12
L 
 −H 0 0 H 0 0 
 36(L e )3 0 0
 0 0 0 0 

0 0 0 0 0 0 0 −4 −2L 0 4 −2L e
e

0 0 − 121 H 3 0 0 121 H 3 0 2L e (L e )2 0 −2L e (L e )2

which becomes identical to (20.19) if we set βh = 9. It is not difficult to show that if the beam is divided into
N ≥ 2 layers of height H/N , the correct beam stiffness is recovered if we take βh = N 2 .
If is not difficult to prove the following generalization. Suppose that the beam is cut into N layers of heights
N N
Hk = γk H , k = 1, . . . N that satisfy 1 Hk = H or 1 γk = 1. To get the correct stiffness of the layered
beam take
βhk = 1/γk2 . (20.24)
For example, suppose that the but is divided into 3 layers of thicknesses H1 = H3 = H/4 and H2 = H/2.
Then pick βh1 = βh3 = 1/( 14 )2 = 16 and βh2 = 1/( 12 )2 = 4.
What is the interpretation of this boost? A spectral analysis of the combined stiffness shows that taking βhk = 1
lowers the rigidity associated with the antisymmetric bending mode of the element. But this mode is associated
with shear-slippage between layers. Boosting βhk as found above compensates exactly for this rigidity decay.

20–16
 §20.5 *LAYERED BEAM CONFIGURATIONS

§20.5.2. Beam Stacks Allowing Interlayer Slip

Sandwich beam fabrication

longitudinal cut _ Cross _
y section y
tf =2 cm
Metal sheet facing
Aluminum
_ z
honeycomb core
z x c= 8 cm
Metal sheet facing
L = 80 cm tf = 2 cm b = 9 cm

Elastic modulus of facings = E f = 65 GPa

Elastic modulus of core ~ 0

Figure 20.16. Plane sandwich beam.

There are beam fabrications where layers can slip longitudinally past each other. One important example is
the sandwich beam illustrated in Figure 20.16. The beam is divided into three layers: 2 metal sheet facings
and a honeycomb core. The core stiffness can be neglected since its effective elastic modulus is very low
compared to the facings modulus E f . In addition, the facings are not longitudinally bonded since they are
separated by the weaker core. To form the beam stiffness by the rigid link method it is sufficient to form the
faces self-stiffness, which are identical, and the rigid-link transformation matrices:
 
Af 0 0 −A f 0 0
 12 Izz f 6 Izz f 12 I 6 Izz f  1 0 d 
 0 0 − ezz2f  0 0 0
 (L e )2 Le (L ) Le  f
   0 1 0 0 0 0 
 6 Izz f 6I
−d − zze f  0 0 1 
E f  0 4 Izz f 2 Izz f 
 , Tk =   , k = 1, 2.
Le 0 0 0
Kek = e  L
0 0 0 
L  −A f 0 0 Af 0 0   1 0 df

 
 0 12 Izz f
− e 2 − e
6 Izz f 12 Izz f 6 Izz f
− e
 0 0 0 0 1 0
 (L )
0
(L e )2
 0 0 0 0 0 1
 L L 
6 Izz f 6 Izz f
0 2 Izz f 0 − e 4 Izz f
Le L
(20.25)
where, in the notation of Figure 20.16, A f = bt f , Ix x f = bt 3f /12, d1 = −(c + t f )/2 and d2 = (c + t f )/2.
The stiffness of the sandwich beam is Ke = T1T K bold1e T1 + T2T K bold2e T2 into which the numerical values
given in Figure 20.16 may be inserted. There is no need to use here the template form and of adjusting the
higher order stiffness.

20–17
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

§20.6. The Space Beam Element

A second example in 3D is the general beam element shown in Figure 20.17. The element is
prismatic and has two end nodes: 1 and 2, placed at the centroid of the end cross sections.
These define the local x̄ axis as directed from
1 to 2. For simplicity the cross section will 3 (x3 ,y3 ,z3 )
Orientation node
be assumed to be doubly symmetric, as is the −−
defining plane x,y
x−
case in commercial I and double-T profiles.
The principal moments of inertia are defined
by these symmetries. The local ȳ and z̄ axes 2 (x2 ,y2 ,z2 )
are aligned with the symmetry lines of the
cross section forming a RH system with x̄. y−
Consequently the principal moments of in-
ertia are I yy and Izz , the bars being omitted −
z
for convenience. 1 (x ,y ,z ) 1 1 1
The global coordinate system is {x, y, z}. To
define the orientation of { ȳ, z̄} with respect y
to the global system, a third orientation node
z x
3, which must not be colinear with 1–2, is
introduced. See Figure 20.17. Axis ȳ lies in Figure 20.17. The space (3D) beam element.
the 1–2–3 plane and z̄ is normal to 1–2–3.

Six global DOF are defined at each node i: the 3 translations u xi , u yi , u zi and the 3 rotations θxi ,
θ yi , θzi .

§20.6.1. Space Beam Stiffness Matrix

The element global node displacements and conjugate forces are arranged as

ue = [ u x1 u y1 u z1 θx1 θ y1 θz1 u x2 u y2 u z2 θx2 θ y2 θz2 ]T ,

(20.26)
fe = [ f x1 f y1 f z1 m x1 m y1 m z1 f x2 f y2 f z2 m x2 m y2 m z2 ]T .

The beam material is characterized by the elastic modulus E and the shear modulus G (the latter
appears in the torsional stiffness). Four cross section properties are needed: the cross section area
A, the moment of inertia J that characterizes torsional rigidity,2 and the two principal moments of
inertia I yy and Izz taken with respect to ȳ and z̄, respectively. The length of the element is denoted
by L. The Bernoulli-Euler model is used; thus the effect of tranverse shear on the beam stiffness is
neglected.
To simplify the following expressions, define the following “rigidity” combinations by symbols:
R a = E A/L, R t = G J/L, R by3 = E I yy /L 3 , R by2 = E I yy /L 2 , R by = E I yy /L, Rz3
b
= E Izz /L 3 ,
b
Rz2 = E Izz /L 2 , Rzb = E Izz /L. Note that R a is the axial rigidity, R t the torsional rigidity, while

2 For circular and annular cross sections, J is the polar moment of inertia of the cross section wrt x̄. For other sections J
has dimensions of (length)4 but must be calculated according to St. Venant’s theory of torsion, or approximate theories.

20–18
 §20.6 THE SPACE BEAM ELEMENT

SpaceBeamColumn2Stiffness[ncoor_,{Em_,Gm_},{A_,Izz_,Iyy_,Jxx_},
options_]:= Module[
{x1,x2,y1,y2,z1,z2,x21,y21,z21,xm,ym,zm,x0,y0,z0,dx,dy,dz,
EA,EIyy,EIzz,GJ,numer,ra,ry,ry2,ry3,rz,rz2,rz3,rx,
L,LL,LLL,yL,txx,txy,txz,tyx,tyy,tyz,tzx,tzy,tzz,T,Kebar,Ke},
{x1,y1,z1}=ncoor[[1]]; {x2,y2,z2}=ncoor[[2]];
{x0,y0,z0}={xm,ym,zm}={x1+x2,y1+y2,z1+z2}/2;
If [Length[ncoor]<=2,{x0,y0,z0}+={0,1,0}];
If [Length[ncoor]==3,{x0,y0,z0}=ncoor[[3]] ];
{x21,y21,z21}={x2-x1,y2-y1,z2-z1}; {numer}=options;
EA=Em*A; EIzz=Em*Izz; EIyy=Em*Iyy; GJ=Gm*Jxx;
LL=Simplify[x21^2+y21^2+z21^2]; L=Sqrt[LL];
If [numer, {x21,y21,z21,EA,EIyy,EIzz,GJ,LL,L}=
N[{x21,y21,z21,EA,EIyy,EIzz,GJ,LL,L}]];
If [!numer, L=PowerExpand[L]]; LLL=Simplify[LL*L];
ra=EA/L; rx=GJ/L;
ry=2*EIyy/L; ry2=6*EIyy/LL; ry3=12*EIyy/LLL;
rz=2*EIzz/L; rz2=6*EIzz/LL; rz3=12*EIzz/LLL;
Kebar={
{ ra, 0, 0, 0, 0, 0, -ra, 0, 0, 0, 0, 0},
{ 0, rz3, 0, 0, 0, rz2, 0,-rz3, 0, 0, 0, rz2},
{ 0, 0, ry3, 0,-ry2, 0, 0, 0,-ry3, 0,-ry2, 0},
{ 0, 0, 0, rx, 0, 0, 0, 0, 0, -rx, 0, 0},
{ 0, 0,-ry2, 0,2*ry, 0, 0, 0, ry2, 0, ry, 0},
{ 0, rz2, 0, 0, 0,2*rz, 0,-rz2, 0, 0, 0, rz},
{-ra, 0, 0, 0, 0, 0, ra, 0, 0, 0, 0, 0},
{ 0,-rz3, 0, 0, 0,-rz2, 0, rz3, 0, 0, 0,-rz2},
{ 0, 0,-ry3, 0, ry2, 0, 0, 0, ry3, 0, ry2, 0},
{ 0, 0, 0,-rx, 0, 0, 0, 0, 0, rx, 0, 0},
{ 0, 0,-ry2, 0, ry, 0, 0, 0, ry2, 0,2*ry, 0},
{ 0, rz2, 0, 0, 0, rz, 0,-rz2, 0, 0, 0,2*rz}};
{dx,dy,dz}={x0-xm,y0-ym,z0-zm};If[numer,{dx,dy,dz}=N[{dx,dy,dz}]];
tzx=dz*y21-dy*z21; tzy=dx*z21-dz*x21; tzz=dy*x21-dx*y21;
zL=Sqrt[tzx^2+tzy^2+tzz^2];
If [!numer,zL=Simplify[PowerExpand[zL]]];
{tzx,tzy,tzz}={tzx,tzy,tzz}/zL; {txx,txy,txz}={x21,y21,z21}/L;
tyx=tzy*txz-tzz*txy; tyy=tzz*txx-tzx*txz; tyz=tzx*txy-tzy*txx;
Te={{txx,txy,txz, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{tyx,tyy,tyz, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{tzx,tzy,tzz, 0, 0, 0, 0, 0, 0, 0, 0, 0},
{ 0, 0, 0, txx,txy,txz, 0, 0, 0, 0, 0, 0},
{ 0, 0, 0, tyx,tyy,tyz, 0, 0, 0, 0, 0, 0},
{ 0, 0, 0, tzx,tzy,tzz, 0, 0, 0, 0, 0, 0},
{ 0, 0, 0, 0, 0, 0, txx,txy,txz, 0, 0, 0},
{ 0, 0, 0, 0, 0, 0, tyx,tyy,tyz, 0, 0, 0},
{ 0, 0, 0, 0, 0, 0, tzx,tzy,tzz, 0, 0, 0},
{ 0, 0, 0, 0, 0, 0, 0, 0, 0, txx,txy,txz},
{ 0, 0, 0, 0, 0, 0, 0, 0, 0, tyx,tyy,tyz},
{ 0, 0, 0, 0, 0, 0, 0, 0, 0, tzx,tzy,tzz}};
Ke=Transpose[Te].Kebar.Te;
Return[Ke]
];

Figure 20.18. Module to form stiffness of space (3D) beam.

the R b ’s are bending rigities scaled by the length in various ways. Then the 12 × 12 local stiffness
matrix can be written as3

3 Cf. page 79 of Pzremieniecki [603]. The presentation in this book includes transverse shear effects as per Timoshenko’s
beam theory. The form (20.27) results from neglecting those effects.

20–19
Chapter 20: IMPLEMENTATION OF ONE-DIMENSIONAL ELEMENTS

 Ra 0 0 0 0 0 −R a 0 0 0 0 0 
0 12Rz3b
0 0 0 b
6Rz2 0 −12Rz3 b
0 0 0 b
6Rz2
 
 0 0 12R by3 0 −6R by2 0 0 0 −12R y3 0 −6R y2
b b
0 
 
 0 0 0 Rt 0 0 0 0 0 −R t 0 0 
 0 −6R y2 0
b
4R by b b 0 
 0 0 0 0 6R y2 0 2R y 
 0 b
4Rzb −6Rz2b
2Rzb 
K̄ =  
e 6Rz2 0 0 0 0 0 0 0
(20.27)
 −R a 0 0 0 0 0 Ra 0 0 0 0 0 
 0 −12R b b 
 0 0 0 −6Rz2 b
0 b
12Rz3 0 0 0 −6Rz2 
 0 z3
−12R y3 0
b b
12R by3 6R by2 0 
 0 6R y2 0 0 0 0 
 0 −R t Rt 0 
 0 0 0 0 0 0 0 0

0 0 −6R by2 0 2R by 0 0 0 6R by2 0 4R by 0
0 b
6Rz2 0 0 0 2Rzb 0 −6Rz2b
0 0 0 4Rzb

The transformation to the global system is the subject of Exercise 20.8.

§20.6.2. Space Beam Stiffness Module

The computation of the stiffness matrix Ke of the two-node, prismatic space beam-column element
is done by Mathematica module SpaceBeamColumn2Stiffness. This is listed in Figure 20.18.
The module is invoked as

Ke = SpaceBeamColumn2Stiffness[ncoor, { Em,Gm }, { A,Izz,Iyy,Jxx }, options]

(20.28)
The arguments are
ncoor Node coordinates of element arranged as { { x1,y1,z1 },{ x2,y2,z2 },{ x3,y3,z3 } },
in which { x3,y3,z3 } specifies an orientation node 3 that defines the local frame.
See §20.4.1.
Em Elastic modulus.
Gm Shear modulus.
A Cross section area.
Izz Moment of inertia of cross section area respect to axis z̄.
Iyy Moment of inertia of cross section area respect to axis ȳ.
Jxx Inertia with respect to x̄ that appears in torsional rigidity G J .
options A list of processing options. For this element is has only one entry: { numer }. This
is a logical flag with the value True or False. If True the computations are carried
out in floating-point arithmetic. If False symbolic processing is assumed.
The module returns the 12 × 12 element stiffness matrix as function value.
The implementation logic and testing of this element is the subject of Exercises 20.8 and 20.9.
Notes and Bibliography
All elements implemented here are formulated in most books dealing with matrix structural analysis. Przemie-
niecki [603] has been recommended in Chapter 1 on account of being inexpensive. Implementation and testing
procedures are rarely covered.
The use of rigid links for offsetting degrees of freedom is briefly covered in §7.8 of the textbook [149].

20–20
 Exercises

Homework Exercises for Chapter 20

Implementation of One-Dimensional Elements

EXERCISE 20.1 [C:15] Download the plane bar stiffness module and their testers and verify the test results
reported here. Comment on whether the stiffness matrix Ke has the correct rank of 1.
EXERCISE 20.2 [C:15] Download the space bar stiffness module and their testers and verify the test results
reported here. Comment on whether the computed stiffness matrix Ke has the correct rank of 1.
EXERCISE 20.3 [C:15] Download the plane beam-column stiffness module and their testers and verify the
test results reported here. Comment on whether the computed stiffness matrix Ke has the correct rank of 3.
EXERCISE 20.4 [A+C:30] Explain why the space bar element has rank 1 although it has 6 degrees of freedom
and 6 rigid body modes. (According to the formula given in Chapter 19, the correct rank should be 6 − 6 = 0.)
EXERCISE 20.5 [C:25] Implement the plane bar, plane beam-column and space bar stiffness element module
in a lower level programming language and check them by writing a short test driver. [Do not bother about
the mass modules.] Your choices are C, Fortran 77 or Fortran 90. (C++ is overkill for this kind of software).
EXERCISE 20.6 [A:25] Explain why the eigenvalues of Ke of any the elements given here do not change if
the {x, y, z} global axes change.
EXERCISE 20.7 [A+C:30] (Advanced) Implement a 3-node space bar element. Hint: use the results of
Exercise 16.5 and transform the local stiffness to global coordinates via a 3 × 9 transformation matrix. Test
the element and verify that it has two nonzero eigenvalues.
EXERCISE 20.8 [D+A:25] Explain the logic of the space beam module listed in Figure 20.18. Assume that
e
the local stiffness matrix K̄ stored in Kebar is correct (it has been trranscribed from [603]. Instead, focus on
how the local to global transformation is built and applied.
EXERCISE 20.9 [C:25] Test the space beam element of Figure 20.18 using the scripts given in Figures E20.1
and E20.2, and report results. Comment on whether the element exhibits the correct rank of 6.

ClearAll[L,Em,Gm,A,Izz,Iyy,Jxx];
ncoor={{0,0,0},{1,8,4}}; Em=54; Gm=30;
A=18; Izz=36; Iyy=72; Jxx=27;
Ke= SpaceBeamColumn2Stiffness[ncoor,{Em,Gm},{A,Izz,Iyy,Jxx},{True}];
Print["Numerical Elem Stiff Matrix: "];
Print[SetPrecision[Ke,4]//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[Ke]]];

Figure E20.1. Script for numeric testing of the space beam module of Figure 20.18.

ClearAll[L,Em,Gm,A,Izz,Iyy,Jxx];
ncoor={{0,0,0},{2*L,2*L,L}/3};
Ke=SpaceBeamColumn2Stiffness[ncoor,{Em,Gm},{A,Izz,Iyy,Jxx},{False}];
kfac=Em; Ke=Simplify[Ke/kfac];
Print["Numerical Elem Stiff Matrix: "];
Print[kfac," ",Ke//MatrixForm];
Print["Eigenvalues of Ke=",kfac,"*",Eigenvalues[Ke]];

Figure E20.2. Script for symbolic testing of the space beam module of Figure 20.18.

20–21
 21
FEM Program
for Space Trusses

21–1
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

TABLE OF CONTENTS

Page
§21.1 Introduction . . . . . . . . . . . . . . . . . . . . . 21–3
§21.2 Analysis Stages . . . . . . . . . . . . . . . . . . . . 21–3
§21.3 Analysis Support Modules . . . . . . . . . . . . . . . . 21–3
§21.3.1 Defining Boundary Conditions . . . . . . . . . . . 21–5
§21.3.2 Modifying the Master Stiffness Equations . . . . . . . . 21–6
§21.3.3 Displacement Solution and Reaction Recovery . . . . . . 21–8
§21.3.4 Flattening and Partitioning Node-Freedom Vectors . . . . . 21–9
§21.3.5 Internal Force Recovery . . . . . . . . . . . . . 21–10
§21.3.6 The Solution Driver . . . . . . . . . . . . . . . 21–11
§21.4 Utility Print Modules . . . . . . . . . . . . . . . . . 21–12
§21.5 Utility Graphic Modules . . . . . . . . . . . . . . . . . 21–13
§21.5.1 Plot Module Calls . . . . . . . . . . . . . . . 21–13
§21.5.2 Plot View Specification . . . . . . . . . . . . . . 21–15
§21.6 Example 1: Bridge Plane Truss Example . . . . . . . . . . 21–15
§21.7 Example 2: An Orbiting Truss Structure . . . . . . . . . . . 21–19
§21. Notes and Bibliography . . . . . . . . . . . . . . . . . 21–20
§21. Exercises . . . . . . . . . . . . . . . . . . . . . . 21–21

21–2
 §21.3 ANALYSIS SUPPORT MODULES

§21.1. Introduction

This Chapter presents a complete FEM program for analysis of space trusses. Why not start with
plane trusses? Three reasons. First, plane truss code already appeared in Chapter 4, although most
components were tailored to the example truss of Chapters 2–3. Second, the difference between
2D and 3D implementation logic for truss analysis is insignificant. Finally, space trusses are more
interesting in engineering applications, particularly for orbiting space structures.
The overall organization of the space truss analysis program is diagramed in Figure 21.1.
The description is done in “bottom up”
fashion. That means the element level Problem
User prepares
modules are presented first, followed script for each Driver
by midlevel modules, ending with the problem Utilities:
Tabular Printing,
driver program. This means traversing Analysis
Graphics, etc
Driver
the diagram of Figure 21.1 left to right
and bottom to top.
BC Built in Internal
The program includes some simple Assembler
Application Equation Force Recovery
Solver
minded graphics, including animation.
The graphic modules are provided in Presented in
Element Chapter 20 Element Int
the posted Notebook but not described Stiffness Forces
in detail. Data structures are explained Element Library
along the way as they arise.
Figure 21.1. Overall organization of space truss analysis program.

§21.2. Analysis Stages

The analysis of a structure by the Direct Stiffness Method involves three major stages: preprocessing
or model definition, processing, and postprocessing. This is true for toy programs such as the one
presented here, through huge commercial codes. Of course the stages here are very short.
The preprocessing portion of the space truss analysis is done by a driver script, which directly sets
the data structures for the problem at hand.
The processing stage involves three steps:
• Assembly of the master stiffness matrix, with a subordinate element stiffness module.
• Modification of master stiffness matrix and node force vector for displacement boundary
conditions.
• Solution of the modified equations for displacements. For the program presented here the built
in Mathematica function LinearSolve is used.
Upon executing the processing steps, the displacements are available. The following postprocessing
steps follow.
• Recovery of forces including reactions, done through a Ku matrix multiplication.
• Computation of internal (axial) forces and stresses in truss members.
• Plotting deflected shapes and member stress levels.

21–3
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

SpaceTrussMasterStiffness[nodxyz_,elenod_,
elemat_,elefab_,prcopt_]:=Module[
{numele=Length[elenod],numnod=Length[nodxyz],neldof,
e,eftab,ni,nj,i,j,ii,jj,ncoor,Em,A,options,Ke,K},
K=Table[0,{3*numnod},{3*numnod}];
For [e=1, e<=numele, e++, {ni,nj}=elenod[[e]];
eftab={3*ni-2,3*ni-1,3*ni,3*nj-2,3*nj-1,3*nj};
ncoor={nodxyz[[ni]],nodxyz[[nj]]};
Em=elemat[[e]]; A=elefab[[e]]; options=prcopt;
Ke=SpaceBar2Stiffness[ncoor,Em,A,options];
neldof=Length[Ke];
For [i=1, i<=neldof, i++, ii=eftab[[i]];
For [j=i, j<=neldof, j++, jj=eftab[[j]];
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]] ];
];
]; Return[K];
];
SpaceBar2Stiffness[ncoor_,Em_,A_,options_]:=Module[
{x1,x2,y1,y2,z1,z2,x21,y21,z21,EA,numer,L,LL,LLL,Ke},
{{x1,y1,z1},{x2,y2,z2}}=ncoor;{x21,y21,z21}={x2-x1,y2-y1,z2-z1};
EA=Em*A; {numer}=options; LL=x21^2+y21^2+z21^2; L=Sqrt[LL];
If [numer,{x21,y21,z21,EA,LL,L}=N[{x21,y21,z21,EA,LL,L}]];
If [!numer, L=PowerExpand[L]]; LLL=Simplify[LL*L];
Ke=(Em*A/LLL)*
{{ x21*x21, x21*y21, x21*z21,-x21*x21,-x21*y21,-x21*z21},
{ y21*x21, y21*y21, y21*z21,-y21*x21,-y21*y21,-y21*z21},
{ z21*x21, z21*y21, z21*z21,-z21*x21,-z21*y21,-z21*z21},
{-x21*x21,-x21*y21,-x21*z21, x21*x21, x21*y21, x21*z21},
{-y21*x21,-y21*y21,-y21*z21, y21*x21, y21*y21, y21*z21},
{-z21*x21,-z21*y21,-z21*z21, z21*x21, z21*y21, z21*z21}};
Return[Ke]];

Figure 21.2. Master stiffness assembly module for a space truss. The element stiffness
module SpaceBar2Stiffness, already discussed in Chapter 20, is listed for convenience.

§21.3. Analysis Support Modules

We begin by listing here modules that support processing steps. These are put into separate cells
for display and testing convenience.
HAssembling the Master Stiffness
Module SpaceTrussMasterStiffness, listed in Figure 21.2, assembles the master stiffness ma-
trix of a space truss. It uses the element stiffness formation module SpaceBar2Stiffness dis-
cussed in the previous Chapter. That module is also listed in Figure 21.2 for convenience of the
reader. The assembler is invoked by

K = SpaceTrussMasterStiffness[nodxyz,elenod,elemat,elefab,prcopt] (21.1)

The arguments are

nodxyz Nodal coordinates placed in a node-by-node list { { x1,y1,z1 },{ x2,y2,z2 }, . . .
{ xn,yn,zn } }, where n is the total number of nodes of the truss.
elenod Element end nodes placed in an element-by-element list: { { i1,j1 }, { i2,j2 },
. . . { ie,je } }, where e is the total number of elements of the truss.

21–4
 §21.3 ANALYSIS SUPPORT MODULES

ClearAll[nodxyz,elemat,elefab,eleopt];
nodxyz={{0,0,0},{10,0,0},{10,10,0}};
elenod= {{1,2},{2,3},{1,3}};
elemat= Table[100,{3}]; elefab= {1,1/2,2*Sqrt[2]}; prcopt= {False};
K=SpaceTrussMasterStiffness[nodxyz,elenod,elemat,elefab,prcopt];
Print["Master Stiffness of Example Truss in 3D:\n",K//MatrixForm];
Print["eigs of K:",Chop[Eigenvalues[N[K]]]];
Master Stiffness of Example Truss in 3D:
20 10 0 −10 0 0 −10 −10 0
10 10 0 0 0 0 −10 −10 0
0 0 0 0 0 0 0 0 0
−10 0 0 10 0 0 0 0 0
0 0 0 0 5 0 0 −5 0
0 0 0 0 0 0 0 0 0
−10 −10 0 0 0 0 10 10 0
−10 −10 0 0 −5 0 10 15 0
0 0 0 0 0 0 0 0 0
eigs of K: {45.3577, 16.7403, 7.902, 0, 0, 0, 0, 0, 0}

Figure 21.3. Testing the space truss assembler module.

elemat Element material properties. The only such property required for this analysis is
the elastic modulus E of each element. These are put in an element-by-element list:
{ E1, E2, . . . Ee }, where e is the total number of elements of the truss.
elefab Element fabrication properties. The only such property required for this analysis is
the cross section area A of each element. These are put in an element-by-element
list: { A1, A2, . . . Ae }, where e is the total number of elements of the truss.
prcopt Processing option list. Only one option is required in the assembler and so prcopt
is simply { numer }. Here numer is a logical flag set to { True } to tell the assembler
to carry out element stiffness computations in floating-point arithmetic. Else set to
{ False } to keep computations in exact arithmetic or symbolic form.
The module returns the assembled stiffness, stored as a full 3n × 3n matrix, as function value.
Details of the assembly process are discussed in Chapter 25 for more general scenarios. Here we
note that the module uses the freedom-pointer-table technique described in §3.4 for merging each
element stiffness matrix into the master stiffness.
The assembler is tested by the script shown in the top cell of Figure 21.3. The script defines the
√ placed in the {x, y} plane.
example truss of Chapters 2–3 as a 3D structure with its three members
See Figure 21.4. The axial rigidity values E A = 100, 50 and 200 2 for elements 1, 2 and 3,
respectively, have to be untangled because E is placed in elemat whereas A goes to elefab.
To split E A we take E = 100 for the three elements. Thus elemat = { 100,100,100 } =
Table[100,{ 3 }] whereas elefab = { 1,1/2,2*Sqrt[2] }.
Running the assembler in exact arithmetic gives the 9 × 9 master stiffness shown in the bottom
cell of Figure 21.3. Taking its eigenvalues gives 6 zeros, which is the expected number in three
dimensions.

21–5
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

§21.3.1. Defining Boundary Conditions

The modification process described here refers to the application of displacement boundary condi-
tions on the master stiffness equations. These are assumed to be single-freedom constraints, either

;;;
homogeneous such as u x3 = 0, or nonhomogeneous such as u z6 = −0.72.

;;;;
Boundary condition data in FEM programs is usually
specified in two levels: nodes at the first level, and fy3 = 1
freedoms assigned to that node at the second level. (The 3 uz3 = 0
reason is that nodes are highly visible to casual users, fx3 = 2
whereas direct access to freedom numbers is difficult.)
This space truss program is no exemption. This data is (3)

;;;;;
(2)
organized into two lists: node freedom tags and node

;;;
y
freedom values. Their configuration is best specified
through an example. 1

;
(1)
2
Consider again the example truss in 3D shown in z
x

;;
;
Figure 21.4, which has 3 nodes and 9 freedoms. The u x1= u y1 = u z1 = 0
node freedom tag list, internally called nodtag, is u y2 = u z2 = 0
Figure 21.4. The example truss in three
nodtag = { { 1,1,1 },{ 0,1,1 },{ 0,0,1 } } (21.2) dimensions, used as module tester.

Each first-level entry of this list pertains to a node. The second level is associated with freedoms:
displacements in the x, y and z directions. Freedom activity is marked by tag 0 or 1. A 1-tag
means that the displacement is prescribed, whereas a 0-tag indicates that the force is known. Thus
{ 1,1,1 } for node 1 means that the node is fixed in the three directions.
The node freedom value list, internally called nodval, gives the prescribed value of the force or
the displacement. For the example truss it is

nodval = { { 0,0,0 },{ 0,0,0 },{ 2,1,0 } } (21.3)

For node 1, the tag entry is { 1,1,1 } and the value entry is { 0,0,0 }. This says that u x1 = u y1 =
u z1 = 0. For node 2 it says that f x2 = 0 and u y2 = u z2 = 0. For node 3 it says that f x3 = 2,
f y3 = 1 and u z3 = 0. The entries of nodval can be integers, floating point numbers, or symbols,
whereas those in nodtag can only be 0 or 1.1
§21.3.2. Modifying the Master Stiffness Equations
The modification of the master stiffness equations K u = f for displacement BCs produces the
modified system K̂ u = f̂. This is done by the two modules listed in Figure 21.5. These modules
are not restricted to space trusses, and may in fact be used for more general problems.
The stiffness modifier is invoked by

Kmod = ModifiedMasterStiffness[nodtag, K] (21.4)

The arguments are:

1 Other tag values may be implemented in more complicated programs to mark multifreedom constraints, for example.

21–6
 §21.3 ANALYSIS SUPPORT MODULES

ModifiedMasterStiffness[nodtag_,K_] := Module[
{i,j,k,n=Length[K],pdof,np,Kmod=K},
pdof=PrescDisplacementDOFTags[nodtag]; np=Length[pdof];
For [k=1,k<=np,k++, i=pdof[[k]];
For [j=1,j<=n,j++, Kmod[[i,j]]=Kmod[[j,i]]=0];
Kmod[[i,i]]=1];
Return[Kmod]];

ModifiedNodeForces[nodtag_,nodval_,K_,f_]:= Module[
{i,j,k,n=Length[K],pdof,pval,np,d,c,fmod=f},
pdof=PrescDisplacementDOFTags[nodtag]; np=Length[pdof];
pval=PrescDisplacementDOFValues[nodtag,nodval]; c=Table[1,{n}];
For [k=1,k<=np,k++, i=pdof[[k]]; c[[i]]=0];
For [k=1,k<=np,k++, i=pdof[[k]]; d=pval[[k]];
fmod[[i]]=d; If [d==0, Continue[]];
For [j=1,j<=n,j++, fmod[[j]]-=K[[i,j]]*c[[j]]*d];
];
Return[fmod]];

Figure 21.5. Modules to modify the master stiffness matrix and node force vector
to apply the displacement boundary conditions.

nodtag A node by node list of freedom tags, as defined in the previous subsection.
K The master stiffness matrix K produced by the assembler module.
The modified stiffness matrix, which has the same order as K, is returned as function value.
The force modifier is invoked by

fmod = ModifiedNodeForces[pdof, pval, K, f] (21.5)

The arguments are:

nodtag The node freedom tag list defined in the previous subsection.
nodval The node freedom value list defined in the previous subsection.
K The master stiffness matrix K produced by the assembler module (before modifica-
tion). This is only used if at least one of the displacement BCs is non-homogeneous.
f The force vector before application of the displacement BCs.
The modified force vector, which has the same order as f, is returned as function value.
The modules are tested by the script listed in the top cell of Figure 21.6. It uses symbolic master
stiffness equations of order 6. The test illustrates a not well known feature of Mathematica: use of
Array function to generate subscripted symbolic arrays of one and two dimensions. The results,
shown in the bottom cell of Figure 21.6, should be self explanatory.

Remark 21.1. On entry, the modification modules of Figure 21.5 build auxiliary lists pdof and pval to
simplify the modification logic. pdof is a list of the prescribed degrees of freedom identified by their equation
number in the master stiffness equations. For example, if freedoms 4, 7, 9 and 15 are specified, pdof =
{ 4,7,9,15 }. These indices are stored in ascending order. pval is a list of the prescribed displacement
values listed in pdof. These lists are constructed by the modules listed in Figure 21.7. The calls are pdof =
PrescribedDOFTags[nodtag] and pval = PrescribedDOFValues[nodtag,nodval].

21–7
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

ClearAll[K,f,v1,v2,v4]; Km=Array[K,{6,6}];
Print["Master Stiffness: ",Km//MatrixForm];
nodtag={{1,1},{0,1},{0,0}}; nodval={{v1,v2},{0,v4},{0,0}};
Kmod=ModifiedMasterStiffness[nodtag,Km];
Print["Modified Master Stiffness:",Kmod//MatrixForm];
fm=Array[f,{6}]; Print["Master Force Vector:",fm];
fmod=ModifiedNodeForces[nodtag,nodval,Km,fm];
Print["Modified Force Vector:",fmod//MatrixForm];
K[1, 1] K[1, 2] K[1, 3] K[1, 4] K[1, 5] K[1, 6]
K[2, 1] K[2, 2] K[2, 3] K[2, 4] K[2, 5] K[2, 6]
Master Stiffness: K[3, 1] K[3, 2] K[3, 3] K[3, 4] K[3, 5] K[3, 6]
K[4, 1] K[4, 2] K[4, 3] K[4, 4] K[4, 5] K[4, 6]
K[5, 1] K[5, 2] K[5, 3] K[5, 4] K[5, 5] K[5, 6]
K[6, 1] K[6, 2] K[6, 3] K[6, 4] K[6, 5] K[6, 6]
1 0 0 0 0 0
0 1 0 0 0 0
Modified Master Stiffness: 0 0 K[3, 3] 0 K[3, 5] K[3, 6]
0 0 0 1 0 0
0 0 K[5, 3] 0 K[5, 5] K[5, 6]
0 0 K[6, 3] 0 K[6, 5] K[6, 6]
Master Force Vector: { f[1], f[2], f[3], f[4], f[5], f[6] }
v1
v2
Modified Force Vector: f[3] − v1 K[1, 3] − v2 K[2, 3] − v4 K[4, 3]
v4
f[5] − v1 K[1, 5] − v2 K[2, 5] − v4 K[4, 5]
f[6] − v1 K[1, 6] − v2 K[2, 6] − v4 K[4, 6]

Figure 21.6. Testing the displacement boundary condition application modules.

Remark 21.2. The logic of ModifiedMasterStiffness is straightforward. Construct pdof, then clear
appropriate rows and columns of K and place ones on the diagonal. Note the use of the Mathematica function
Length to control loops: np=Length[pdof] sets np to the number of prescribed freedoms. Similarly
n=Length[K] sets n to the order of the master stiffness matrix K, which is used to bound the row and column
clearing loop. These statements may be placed in the list that declares local variables.

Remark 21.3. ModifiedNodalForces has more complicated logic because it accounts for nonhomogeneous
BCs. On entry it constructs pdof and pval. If nonzero values appear in pval, the original entries of f are mod-
ified as described in §4.1.2, and the end result is the effective force vector. Force vector entries corresponding
to the prescribed displacement values are replaced by the latter in accordance with the prescription (4.13). If
there are nonhomomogeneous BCs it is important that the stiffness matrix provided as third argument be the
master stiffness before modification by ModifiedStiffnessMatrix. This is because stiffness coefficients
that are cleared by ModifiedStiffnessMatrix are needed for modifying the force vector.

§21.3.3. Displacement Solution and Reaction Recovery

The linear system K̂ u = f̂, where K̂ and f̂ are the modified stiffness and force matrices, respectively,
is solved for displacements by a built-in linear algebraic solver. In Mathematica LinearSolve is
available for this purpose:
u = LinearSolve[Kmod, fmod] (21.6)

At this point postprocessing begins. The node forces including reactions are obtained from f = K u.

21–8
 §21.3 ANALYSIS SUPPORT MODULES

PrescDisplacementDOFTags[nodtag_]:= Module [
{j,n,numnod=Length[nodtag],pdof={},k=0,m},
For [n=1,n<=numnod,n++, m=Length[nodtag[[n]]];
For [j=1,j<=m,j++, If [nodtag[[n,j]]>0,
AppendTo[pdof,k+j]];
]; k+=m;
]; Return[pdof]];
PrescDisplacementDOFValues[nodtag_,nodval_]:= Module [
{j,n,numnod=Length[nodtag],pval={},k=0,m},
For [n=1,n<=numnod,n++, m=Length[nodtag[[n]]];
For [j=1,j<=m,j++, If [nodtag[[n,j]]>0,
AppendTo[pval,nodval[[n,j]]]];
]; k+=m;
]; Return[pval]];

Figure 21.7. Modules to build auxiliary lists pdof and pval from node-by-node BC data.

This can be done simply as a matrix product:

f = K.u (21.7)

where K is the original master stiffness matrix before modification, and u the displacement vector
computed by LinearSolve.
§21.3.4. Flattening and Partitioning Node-Freedom Vectors
In the analysis process one often needs displacement and force vectors in two different list formats.
For example, the computed node displacement vector produced by (21.6) for the example truss in
3D is
u = { 0,0,0,0,0,0,0.4,-0.2,0 } (21.8)
Following the terminology of Mathematica this will be called the flat form of the displacement
vector. For postprocessing purposes (especially printing and plotting) it is convenient to rearrange
to the node by node form

noddis = { { 0,0,0 },{ 0,0,0 },{ 0.4,-0.2,0 } } (21.9)

This will be called the node-partitioned form of the displacement vector. Similar dual formats exist
for the node force vector. In flat form this is called f and in node-partitioned form nodfor.

FlatNodePartVector[nv_]:=Flatten[nv];

NodePartFlatVector[nfc_,v_]:= Module [
{i,k,m,n,nv={},numnod},
If [Length[nfc]==0, nv=Partition[v,nfc]];
If [Length[nfc]>0, numnod=Length[nfc]; m=0;
nv=Table[0,{numnod}];
For [n=1,n<=numnod,n++, k=nfc[[n]];
nv[[n]]=Table[v[[m+i]],{i,1,k}];
m+=k]];
Return[nv]];

Figure 21.8. Utility modules to flatten and node-partition node-DOF vectors.

21–9
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

SpaceTrussIntForces[nodxyz_,elenod_,elemat_,elefab_,
noddis_,prcopt_]:= Module[{ numnod=Length[nodxyz],
numele=Length[elenod],e,ni,nj,ncoor,Em,A,options,ue,p},
p=Table[0,{numele}];
For [e=1, e<=numele, e++, {ni,nj}=elenod[[e]];
ncoor={nodxyz[[ni]],nodxyz[[nj]]};
ue=Flatten[{ noddis[[ni]],noddis[[nj]] }];
Em=elemat[[e]]; A=elefab[[e]]; options=prcopt;
p[[e]]=SpaceBar2IntForce[ncoor,Em,A,ue,options]
];
Return[p]];

SpaceBar2IntForce[ncoor_,Em_,A_,ue_,options_]:= Module[
{x1,x2,y1,y2,z1,z2,x21,y21,z21,EA,numer,LL,pe},
{{x1,y1,z1},{x2,y2,z2}}=ncoor; {x21,y21,z21}={x2-x1,y2-y1,z2-z1};
EA=Em*A; {numer}=options; LL=x21^2+y21^2+z21^2;
If [numer,{x21,y21,z21,EA,LL}=N[{x21,y21,z21,EA,LL}]];
pe=(EA/LL)*(x21*(ue[[4]]-ue[[1]])+y21*(ue[[5]]-ue[[2]])+
+z21*(ue[[6]]-ue[[3]]));
Return[pe]];

SpaceTrussStresses[elefab_,elefor_,prcopt_]:= Module[
{numele=Length[elefab],e,elesig}, elesig=Table[0,{numele}];
For [e=1, e<=numele, e++, elesig[[e]]=elefor[[e]]/elefab[[e]] ];
Return[elesig]];

Figure 21.9. Modules to compute internal forces and stresses in a space truss.

The utility modules listed in Figure 21.8 can be used to pass from one format to the other. To flatten
the node-partitioned form of a vector, say nv, say

v = FlatNodePartVector[nv] (21.10)

To node-partition a flat vector v say

nv = NodePartFlatVector[nfc,v] (21.11)

where nfc is the number of freedoms per node. For space trusses this is 3 so appropriate conversion
calls are noddis = NodePartFlatVector[3,u] and nodfor = NodePartFlatVector[3,f].

Remark 21.4. FlatNodePartVector can be directly done by the built-in function Flatten whereas
NodePartFlatVector — for a fixed number of freedoms per node — can be done by Partition. The
reason for the “wrappers” is to guide the conversion of Mathematica code to a lower level language such as
C, where such built-in list functions are missing.

Remark 21.5. The additional code in NodePartFlatVector caters to the case where the number of freedoms
can vary from node to node, in which case nfc is a list (not a number) called the node freedom count, hence
the abbreviation. That facility is useful in more advanced courses.

§21.3.5. Internal Force Recovery

The calculation of internal forces and stresses in a space truss involves computing axial forces in
the bar elements. The modules that do those calculations given the displacement solution are listed

21–10
 §21.3 ANALYSIS SUPPORT MODULES

ClearAll[nodxyz,elenod,elemat,elefab,noddis];
nodxyz={{0,0,0},{10,0,0},{10,10,0}}; elenod={{1,2},{2,3},{1,3}};
elemat= Table[100,{3}]; elefab= {1,1/2,2*Sqrt[2]};
noddis={{0,0,0}, {0,0,0}, {4/10,-2/10,0}}; prcopt={False};
elefor=SpaceTrussIntForces[nodxyz,elenod,elemat,elefab,noddis,prcopt];
Print["Int Forces of Example Truss:",elefor];
Print["Stresses:",SpaceTrussStresses[elefab,elefor,prcopt]];

Int Forces of Example Truss: { 0, −1, 2*Sqrt[2] }

Stresses: { 0, −2, 1 }

Figure 21.10. Test of the internal force recovery module.

in Figure 21.9. Module SpaceTrussIntForces computes the internal forces (axial forces) in all
truss members. It is invoked by

elefor = SpaceTrussIntForces[nodxyz,elenod,elemat, elefab,noddis,prcopt]

(21.12)
Five of the arguments: nodxyz, elenod, elemat, elefab and prcopt, are the same used in the
call (21.1) to the stiffness assembler. The additional argument, noddis, contains the computed
node displacements arranged in node-partitioned form

noddis = { { ux1,uy1,uz1 },{ ux2,uy2,uz2 }, . . . { uxn,uyn,uzn } } (21.13)

This form can be obtained from the computed displacement solution through the utility module
(21.11). As function value SpaceTrussIntForces returns a list of element axial forces

{ p1, p2 . . . pe } (21.14)

SpaceTrussIntForces makes use of SpaceBar2IntForce, which computes the internal force

in an individual bar element. This is invoked as SpaceBar2IntForce

p = SpaceBar2IntForces[ncoor,Em,A,ue,options] (21.15)

Arguments ncoor, Em, A and options are the same as in the call to SpaceBar2Stiffness
described in §20.2.2. The additional argument, ue, contains the flat list of the six element node
displacements in the global system arranged as { ux1,uy1,ux1,ux2,ux2,uz2 }. The recovery
equation is the subject of Exercise 21.4.
The last module in Figure 21.9 computes the member stresses simply by dividing the internal forces
by the cross section areas. It is invoked as

elesig = SpaceTrussStresses[elefab,elefor,prcopt] (21.16)

Here elefab and prcopt are as before, and elefor contains the element forces computed by
SpaceTrussIntForces. The element axial stresses are returned as function value.
The statements of the top cell of Figure 21.9 exercise the internal force recovery for the example
truss in 3D, requesting
√ exact calculiations. Array p is printed in the bottom cell. The axial forces
of 0,, −1 and 2 2 agree with those determined in Chapter 3.

21–11
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

§21.3.6. The Solution Driver

It is convenient to package the sequence of operations desribed in the previous subsections, namely
assembly, modification, solution, force and stress recovery, into one module called the solution
driver. This is listed in Figure 21.11. It is invoked by saying
{ noddis,nodfor,elefor,elesig }= SpaceTrussSolution[nodxyz,elenod,
elemat,elefab,nodtag,nodval,prcopt] (21.17)

All arguments: nodxyz, elenod, elemat, elefab, nodtag, nodval and prcopt, have been
described in previous subsections. The module returns four lists:
noddis Computed node displacement in node partitioned format.
nodfor Recovered node forces including reactions in node partitioned format.
elefor Element internal forces.
elesig Element stresses.
Note that the listing of SpaceTrussSolution in Figure 21.11 has two commented out eigenvalue
computations, one for the master stiffness K and one for the modified stiffness Kmod. Decommenting
those commands comes in handy when setting up and running a new problem if errors are detected.

SpaceTrussSolution[nodxyz_,elenod_,elemat_,elefab_,nodtag_,nodval_,
prcopt_]:= Module[{K,Kmod,f,fmod,u,noddis,nodfor,elefor,elesig},
K=SpaceTrussMasterStiffness[nodxyz,elenod,elemat,elefab,prcopt];
(* Print["eigs of K=",Chop[Eigenvalues[N[K]]]]; *)
Kmod=ModifiedMasterStiffness[nodtag,K];
f=FlatNodePartVector[nodval];
fmod=ModifiedNodeForces[nodtag,nodval,K,f];
(* Print["eigs of Kmod=",Chop[Eigenvalues[N[Kmod]]]]; *)
u=LinearSolve[Kmod,fmod]; u=Chop[u]; f=Chop[K.u, 10.0^(-8)];
nodfor=NodePartFlatVector[3,f]; noddis=NodePartFlatVector[3,u];
elefor=Chop[SpaceTrussIntForces[nodxyz,elenod,elemat,elefab,
noddis,prcopt]];
elesig=SpaceTrussStresses[elefab,elefor,prcopt];
Return[{noddis,nodfor,elefor,elesig}];
];

Figure 21.11. The analysis driver module.

§21.4. Utility Print Modules

Utility print modules are used to display input and output data in tabular form. The following six modules are
provided in Cell 6 of the SpaceTruss.nb notebook.
To print the node coordinates in nodxyz:

PrintSpaceTrussNodeCoordinates[nodxyz,title,digits] (21.18)

To print the element nodes in elenod, element materials in elemat and element fabrications in elefab:

PrintSpaceTrussElementData[elenod,elemat,elefab,title,digits] (21.19)

21–12
 §21.5 UTILITY GRAPHIC MODULES

To print the freedom activity data in nodtag and nodval:

PrintSpaceTrussFreedomActivity[nodtag,nodval,title,digits] (21.20)

To print the node displacements in noddis (configured in node-partitioned form):

PrintSpaceTrussNodeDisplacements[noddis,title,digits] (21.21)

To print the node forces in nodfor (configured in node-partitioned form):

PrintSpaceTrussNodeForces[nodfor,title,digits] (21.22)

To print the element internal forces in elefor and element stresses in elesig:

PrintSpaceTrussElemForcesAndStresses[elefor,elesig,title,digits] (21.23)

In all cases, title is an optional character string to be printed as a title before the table; for example "Node
coordinates of bridge truss". To eliminate the title, specify "" (two quote marks together).
The last argument of the print modules: digits, is optional. If set to { d,f } it specifies that floating point
numbers are to be printed with room for at least d digits, with f digits after the decimal point.
If digits is specified as a void list: { }, a preset default is used for d and f.

§21.5. Utility Graphic Modules

Graphic modules that support preprocessing are placed in Cells 4 and 5 of the SpaceTruss.nb notebook.
These display unlabeled elements, elements and nodes with labels, deformed shapes and element stress levels.
§21.5.1. Plot Module Calls
To plot elements only:

PlotSpaceTrussElements[nodxyz,elenod,title,{ view,aspect,{ } }] (21.24)

To plot element and nodes with optional labeling:

PlotSpaceTrussElementsAndNodes[nodxyz,elenod,title,{ view,aspect,labels }] (21.25)

To plot deformed shape of the truss under computed node displacements:

PlotSpaceTrussDeformedShape[nodxyz,elenod,noddis,amplif,box,title,
{ view,aspect,colors }] (21.26)

To plot element axial stress levels using a coloring scheme:

PlotSpaceTrussStresses[nodxyz,elenod,elesig,sigfac,box,title,
{ view,aspect,{ } }] (21.27)

In the foregoing nodxyz, elenod, noddis, noddfor elefor and elesig have been described above. The
other arguments are as follows.
view A list configured as a list of two 3D vectors: { { Vhat1,Vhat2,Vhat3 },{ W1,W2,W3 } }. Vector
W with global components { W1,W2,W3 } specifies the view direction whereas vector V̂, with
global components { V1,V2,V3 }, specifies the up direction for the plot view, as discussed in
§21.5.2. If a void list is provided for the argument, the default is { { 0,1,0 },{ 0,0,1 } }; this
means that view direction is along the −z axis whereas the up direction is the y axis.

21–13
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

aspect Vertical-to-horizontal aspect ratio of the plot as it appears on a Notebook cell. Three possibilities.
If set to −1, the Mathematica default Aspect->Automatic is chosen. If set to zero, an aspect
ratio is computed by the plot module as appropriate. If set to a positive number, the aspect ratio
is set to that number.
labels Only used in PlotSpaceTrussElementsAndNodes. A list with the following configuration:
{ { nlabels,frn,fex,fey }, { elabels,fre },{ fntfam,fntsiz,fntwgt,fntslt } }.
nlabels A logical flag. Set to True to get node labels in plot.
frn Radius of circle drawn around each node expressed as percentage of plot size.
fex Horizontal eccentricity in points of node label from node location.
fex Vertical eccentricity in points of node label from node location.
elabels A logical flag. Set to True to get element labels in plot.
fre Radius of circle drawn around each element number expressed as percentage of
plot size. from node location.
fntfam Font family used for labels, for example "Times"
fntsiz Size in points of font used for labels; usually 10 through 14.
fntwgt Font weight used for element labels. Typical settings are "Plain" or ”Bold”.
Node labels are always drawn in boldface.
fntslt Font slant used for element labels. Typical settings are "Plain", "Italics" or
"Slanted". Node labels are always drawn in Plain.
amplif The displacement amplification factor to be used by PlotSpaceTrussDeformedShape. Usu-
ally a value much larger than one (say 100 or 1000) is necessary to visualize displacements in
actual structures. Becomes a list if the call is to draw several shapes (for example undeformed
and deformed). For example setting amplif to { 0,100 } will draw two shapes: the undeformed
configuration and one with magnification of 100. If more than one shape is to be drawn the
colors specification comes in handy.
box A list of points, specified by their {x, y, z} global coordinates, that forms a box that encloses the
plot. Used in PlotSpaceTrussDeformedShape and PlotSpaceTrussStresses. The box is
converted to a frame by the view projector. This is useful for various purposes, one being to do
realistic animations by drawing a sequence of deformed shapes moving inside this box.
colors Defines element colors to be used by PlotSpaceTrussDeformedShape, specified as lower
case character string. Legal ones are "black", "red", "blue", "green" and ”white”. If the
call is to draw several shapes (for example undeformed and deformed), this argument can be a
list, such as { "black","red" }, in which case the colors are in one to one correspondence with
the amplification values in amplif. If no color is specified, black is assumed.
sigfac A stress scaling factor for PlotSpaceTrussStresses. Normally set to 1.
Because of the confusing way Mathematica handles plots, (some features do scale with plot size while others,
such as font sizes, do not), some interactive experimentation with dimension specs seems inevitable.

21–14
 §21.6 EXAMPLE 1: BRIDGE PLANE TRUSS EXAMPLE

V = ys Horizontal
Up vector, aka screen direction
vertical screen C
direction e
an
pl U = xs
en eye point, aka
re y camera point,
Sc
observer point,
view reference point
O
W = zs x
Global coordinate system (FEM)
View direction z World coordinate system (graphics)

Mathematica cell

Figure 21.12. Plot view in 3D as mapping of world to screen coordinates.

§21.5.2. Plot View Specification

Plotting 3D objects, such as the space trusses considered here, involves mapping the coordinates given in
the FEM global system {x, y, z} into screen coordinates {xs , ys , z s } in which z s = 0. Objects are rendered
using screen coordinates. Construction of this mapping is based on the view specification, which appears as
argument of all plotting routines described in the foregoing subsection.
The viewing ingredients are shown in Figure 21.12. In computer graphics, the 3D space spanned by the FEM
global system {x, y, z} is called the world space for obvious reasons. The screen coordinates {xs , ys , z s } are
defined by two vectors, which in computer graphics are typically identified as W and V: the view direction
and up direction, respectively.
The view direction is the line defined by joining the eye position at C with the origin O of {x, y, z}. The
screen plane passes through O and is normal to W. Screen coordinates xs and ys are in the screen plane and
going along the horizontal and vertical directions, respectively. In computer graphics the xs and ys directions
are called U and V, respectively. When the plot is rendered in a Mathematica cell, U ≡ xs goes horizontally
from left to right. Axes {u ≡ xs , V ≡ ys , W ≡ z s } form a RHS Cartesian coordinate system. Mappings from
screen to pixel coordinates are handled by the plotting system, and need not be discussed here.
In the plot modules used here, the eye point C is assumed to be at infinity.2 Thus only the view direction
W, as specified by three direction numbers, is used. For example, the specification { 1,1,1 } says that
W is the trisector of the {x, y, z} octant. To define the up direction V ≡ ys one has to enter a second
“indication” vector: V̂, which must not be parallel to W. Since V̂ is not necessarily normal to W, V is
T
constructed by orthogonalization: V = V̂ − (V̂ Wn )Wn , where Wn is W normalized to length one. For
example if V̂ and and W are specified by view argument { { 0,1,0 },{ 2,2,1 } }, then V = [0, 1, 0] −
([0, 1, 0]T [2/3, 2/3, 1/3)[2/3, 2/3, 1/3] = [0, 1, 0] − (2/3)[2/3, 2/3, 1/3] = [−4/9, 5/9, −2/9]. Note that
VT W = 0.

§21.6. Example 1: Bridge Plane Truss Example

This example deals with the analysis of the 6-bay bridge truss problem defined in Figure 21.13.
This truss has 12 nodes and 17 elements. It is contained in the {x, y} plane and can only move in
that plane. It is fixed at node 1 and on rollers at node 12.

2 Graphics where the eye point C is at a finite distance produce perspective plots. The Graphics3D system of Mathematica
allows perpective plotting. However the plots described here use only the 2D graphics subset.

21–15
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

(a) Elastic modulus E = 1000

Cross section areas of bottom longerons: 2,
top longerons: 10, battens: 3, diagonals: 1 top joints lie
on parabolic
profile
9

10 10 16 10 10
span: 6 bays @ 10 = 60

(b) (9) 6 (10) 8

4
y (8) (11)
2 (15) 10
(7) (14) (16) (12)
(19) (20)
1 x (13) (18) (21) (17)
12
;;

;;
(1) 3 (2) 5 (3) 7 (4) 9 (5) 11 (6)

Figure 21.13. Six-bay bridge plane truss used as example problem: (a) truss structure
showing supports and applied loads; (b) finite element idealization as pin-jointed truss.

The driver is listed in Figure 21.14. Preprocessing statements appear on top, with light green
background. These define the problem through specification of the following data structures.3
NodeCoordinates Same configuration as nodxyz
ElemNodes Same configuration as elenod
ElemMaterials Same configuration as elemat
ElemFabrications Same configuration as elefab. This list is built up from four repeating
cross sectional areas: Abot, Atop, Abat and Adia, for the areas of bottom
longerons, top longerons, battens and diagonals, respectively.
NodeDOFTags Same configuration as nodtag. Initialized to { 0,0,1 } for all nodes
on creation so as to fix all z displacements. Then the support condi-
tions in the {x, y} plane are specified at supported nodes. For example,
NodeDOFTags[[1]] = { 1,1,1 } says that the three node displacement
components u x1 , u y1 and u z1 of node 1 are specified.
NodeDOFValues Same configuration as nodval. Initialized to { 0,0,0 } on creation for all
nodes. Then the value of nonzero applied loads is set at nodes 3, 5, 7,
9 and 11. For example NodeDOFValues[[7]] = { 0,-16,0 } specifies
f x7 = 0, f y7 = −16 and u z7 = 0.
The input data structures can be shown in tabular form for convenient inspection using print utility
modules. Printed tables are shown on the left of Figure 21.15.
Running the solution module returns computed displacements, node forces including reactions,
internal forces and member stress. These are printed with utility modules. These results are shown

3 Note the use of longer mnemonic names in the problem driver. For example NodeCoordinates instead of nodxyz.
This simplifies the preparation of problem solving assignments since driver scripts are more self-documenting. It also
helps grading returned assignments.

21–16
 §21.6 EXAMPLE 1: BRIDGE PLANE TRUSS EXAMPLE

NodeCoordinates={{0,0,0},{10,5,0},{10,0,0},{20,8,0},{20,0,0},{30,9,0},
{30,0,0},{40,8,0},{40,0,0},{50,5,0},{50,0,0},{60,0,0}};
ElemNodes={{1,3},{3,5},{5,7},{7,9},{9,11},{11,12},
{1,2},{2,4},{4,6},{6,8},{8,10},{10,12},
{2,3},{4,5},{6,7},{8,9},{10,11},
{2,5},{4,7},{7,8},{9,10}};
PrintSpaceTrussNodeCoordinates[NodeCoordinates,"Node coordinates:",{}];
numnod=Length[NodeCoordinates]; numele=Length[ElemNodes];
Em=1000; Abot=2; Atop=10; Abat=3; Adia=1;
ElemMaterials= Table[Em,{numele}];
ElemFabrications={Abot,Abot,Abot,Abot,Abot,Abot,Atop,Atop,Atop,Atop,
Atop,Atop,Abat,Abat,Abat,Abat,Abat,Adia,Adia,Adia,Adia};
PrintSpaceTrussElementData[ElemNodes,ElemMaterials,ElemFabrications,
"Element data:",{}];
ProcessOptions= {True};

view={{0,1,0},{0,0,1}};
labels={{True,0.06,-1.5,1.5},{True,0.12},{"Times",11,"Roman"}};
PlotSpaceTrussElementsAndNodes[NodeCoordinates,ElemNodes,
"bridge mesh",{view,-1,labels}];

NodeDOFTags= Table[{0,0,1},{numnod}];
NodeDOFValues=Table[{0,0,0},{numnod}];
NodeDOFValues[[3]]={0,-10,0}; NodeDOFValues[[5]]= {0,-10,0};
NodeDOFValues[[7]]={0,-16,0};
NodeDOFValues[[9]]={0,-10,0}; NodeDOFValues[[11]]={0,-10,0};
NodeDOFTags[[1]]= {1,1,1}; (* fixed node 1 *)
NodeDOFTags[[numnod]]={0,1,1}; (* hroller @ node 12 *)
PrintSpaceTrussFreedomActivity[NodeDOFTags,NodeDOFValues,
"DOF Activity:",{}];

{NodeDisplacements,NodeForces,ElemForces,ElemStresses}=
SpaceTrussSolution[ NodeCoordinates,ElemNodes,ElemMaterials,
ElemFabrications,NodeDOFTags,NodeDOFValues,ProcessOptions ];

PrintSpaceTrussNodeDisplacements[NodeDisplacements,
"Computed node displacements:",{}];
PrintSpaceTrussNodeForces[NodeForces,
"Node forces including reactions:",{}];
PrintSpaceTrussElemForcesAndStresses[ElemForces,ElemStresses,
"Int Forces and Stresses:",{}];

view={{0,1,0},{0,0,1}}; box={{0,-4,0},{60,-4,0},{60,10,0},{0,10,0}};
PlotSpaceTrussDeformedShape[NodeCoordinates,ElemNodes,NodeDisplacements,
{0,1},box,"deformed shape (unit magnif)",{view,-1,{"black","blue"}}];
PlotSpaceTrussStresses[NodeCoordinates,ElemNodes,ElemStresses,1,box,
"axial stresses in truss members",{view,0,labels}];

Figure 21.14. Driver script for analysis of the 6-bay plane bridge truss. Preprocessing statements
in light green background. Processing and postprocessing statements in light blue.

on the right of Figure 21.15.

Output plot results are collected in Figure 21.16.

21–17
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

Node coordinates : Computed node displacements :

node x−coor y−coor z−coor node x−displ y−displ z−displ
1 0.000000 0.000000 0.000000 1 0.000000 0.000000 0.000000
2 10.000000 5.000000 0.000000 2 0.809536 −1.775600 0.000000
3 10.000000 0.000000 0.000000 3 0.280000 −1.792260 0.000000
4 20.000000 8.000000 0.000000 4 0.899001 −2.291930 0.000000
5 20.000000 0.000000 0.000000 5 0.560000 −2.316600 0.000000
6 30.000000 9.000000 0.000000 6 0.847500 −2.385940 0.000000
7 30.000000 0.000000 0.000000 7 0.847500 −2.421940 0.000000
8 40.000000 8.000000 0.000000 8 0.795999 −2.291930 0.000000
9 40.000000 0.000000 0.000000 9 1.135000 −2.316600 0.000000
10 50.000000 5.000000 0.000000 10 0.885464 −1.775600 0.000000
11 50.000000 0.000000 0.000000 11 1.415000 −1.792260 0.000000
12 60.000000 0.000000 0.000000 12 1.695000 0.000000 0.000000
Element data: Node forces including reactions :
elem nodes modulus area node x−force y−force z−force
1 { 1, 3} 1000.00 2.00 1 0.0000 28.0000 0.0000
2 { 3, 5} 1000.00 2.00 2 0.0000 0.0000 0.0000
3 { 5, 7} 1000.00 2.00 3 0.0000 −10.0000 0.0000
4 { 7, 9} 1000.00 2.00 4 0.0000 0.0000 0.0000
5 { 9, 11} 1000.00 2.00 5 0.0000 −10.0000 0.0000
6 { 11, 12} 1000.00 2.00 6 0.0000 0.0000 0.0000
7 { 1, 2} 1000.00 10.00 7 0.0000 −16.0000 0.0000
8 { 2, 4} 1000.00 10.00 8 0.0000 0.0000 0.0000
9 { 4, 6} 1000.00 10.00 9 0.0000 −10.0000 0.0000
10 { 6, 8} 1000.00 10.00 10 0.0000 0.0000 0.0000
11 { 8, 10} 1000.00 10.00 11 0.0000 −10.0000 0.0000
12 { 10, 12} 1000.00 10.00 12 0.0000 28.0000 0.0000
13 { 2, 3} 1000.00 3.00
14 { 4, 5} 1000.00 3.00 Int Forces and Stresses :
15 { 6, 7} 1000.00 3.00
16 { 8, 9} 1000.00 3.00 elem axial force axial stress
17 { 10, 11} 1000.00 3.00 1 56.0000 28.0000
18 { 2, 5} 1000.00 1.00 2 56.0000 28.0000
19 { 4, 7} 1000.00 1.00 3 57.5000 28.7500
20 { 7, 8} 1000.00 1.00 4 57.5000 28.7500
21 { 9, 10} 1000.00 1.00 5 56.0000 28.0000
6 56.0000 28.0000
DOF Activity : 7 −62.6100 −6.2610
8 −60.0300 −6.0030
node x−tag y−tag z−tag x−value y−value z−value 9 −60.3000 −6.0300
1 1 1 1 0 0 0 10 −60.3000 −6.0300
2 0 0 1 0 0 0 11 −60.0300 −6.0030
3 0 0 1 0 −10 0 12 −62.6100 −6.2610
4 0 0 1 0 0 0 13 10.0000 3.3330
5 0 0 1 0 −10 0 14 9.2500 3.0830
6 0 0 1 0 0 0 15 12.0000 4.0000
7 0 0 1 0 −16 0 16 9.2500 3.0830
8 0 0 1 0 0 0 17 10.0000 3.3330
9 0 0 1 0 −10 0 18 1.6770 1.6770
10 0 0 1 0 0 0 19 3.2020 3.2020
11 0 0 1 0 −10 0 20 3.2020 3.2020
12 0 1 1 0 0 0 21 1.6770 1.6770

Figure 21.15. Bridge truss example: tabular printed output. On the left: node, element and freedom
data. On the right: computed displacements, node forces, internal forces and member stresses.

bridge mesh

bridge mesh with elem & node labels

6
4 9 10 8
2 8 11 10
14 19 15 20 16
7 13 18 21 17 12
1 3 5 7 9 11 12
1 2 3 4 5 6

deformed shape

axial stress level plot

Figure 21.16. Bridge truss example: graphic output collected in one figure.

21–18
 §21.7 EXAMPLE 2: AN ORBITING TRUSS STRUCTURE

§21.7. Example 2: An Orbiting Truss Structure

To be included in the final version of the Chapter.

21–19
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

Notes and Bibliography

The dominant philosophy in FEM implementation is to construct general purpose programs that can solve
a wide range of problems. For example, static and dynamic response of arbitrary structures with linear or
nonlinear behavior. This path was naturally taken once the Direct Stiffness Method became widely accepted
in the mid 1960s. It is reflected in the current crop of commercial FEM programs. Their source code by now
has reached into millions of lines.
These codes do have a place in undergraduate engineering education, starting at the junior level. At this level
students should be taught rudiments of modeling and how to use those black box programs as tools. This
exposure provides also basic knowledge for capstone senior projects that require finite element analysis.
At the graduate level, however, students should understand what goes on behind the scenes. But access to
innards of commercial programs is precluded.4 The philosophy followed here is to use special purpose codes
written in a high level language. These may be collectively called grey level codes. A high level language such
as Mathematica conceals utility operations such as matrix products, linear solvers and graphics, but permits
the application code logic to be seen and studied.
As a result a complete FEM program is tiny (typically a few hundreds lines), it can be built and debugged in a
few hours, and may be understood as a whole by one person. On the down side of course these toy programs
can do only very limited problems, but for instructional use simplicity outweights generality.

4 And even if it were available, the source code would be difficult to follow given its complexity and the use of lower order
languages such as Fortran.

21–20
 Exercises

Homework Exercises for Chapter 21

FEM Program for Space Trusses

EXERCISE 21.1 [D:10] The logic of SpaceTrussMasterStiffness cannot be used for structures other
than space trusses. Justify this assertion.

EXERCISE 21.2 [D:10] The logic of ModifiedMasterStiffness and ModifiedNodeForces is not re-
stricted to space trusses, but can be used for any FEM program that stores K and f as full arrays. Justify this
assertion.

EXERCISE 21.3 [D:20] The logic of PrescDisplacementDOFTags and PrescDisplacementDOFValues

is not restricted to a fixed number of DOF per node. Justify this assertion.

EXERCISE 21.4 [A:15] Show that the longitudinal elongation of a space bar can be computed directly from
the global displacements u x1 , u y1 , . . . u z2 from

d = (x21 u x21 + y21 u y21 + z 21 u z21 )/, (E21.1)

in which x21 = x2 − x1 , u x21 = u x2 − u x1 , etc, and  is the bar length. Hence justify the formula used in
module SpaceBar2IntForce listed in Figure 21.9 to recover the axial force p = (E A/)d.

EXERCISE 21.5 [C:25] Analyze the structure shown in Figure E21.1. This is a pin-jointed truss model of a
200-in-high (5m) transmission tower originally proposed by Fox and Schmit in 1964 [281] as a test for early
automated-synthesis codes based on FEM. It became a standard benchmark for structural optimization.
The truss has 10 joints (nodes) and 25 members (elements). The truss geometry and node numbering are
defined in Figure E21.1(a). Joints 1 and 2 at the top of the tower lie on the {x, z} plane. The truss (but not the
loads) is symmetric about the {y, z} and {x, z} planes. Figure E21.1(b) gives the element numbers.

(a) z
1 2 (b)
Coordinates:
1 (1) 2
node 2 ( 37.5,0,200)
node 1 (−37.5,0,200) (8)
100 in (9) (6)
75 in (14) joins 3-10 (5) (7)
(4) (3)
(20) joins 5-10 (2)
6 3 (10)
4 3 (12)
5 (13) 6 4
75 in (20) (25)
5
(18)
(22) (23) (19)
7 (14)
10 (15)
7 (11)
y
10 (24) (17)
100 in 8 (21)
200 in (16)
9 x 8
200 in 9

Figure E21.1. 25-member space truss model of a transmission tower. (a): Geometry definition and
node numbers; (b) element numbers. For member properties and loads see Tables E21.1 and E21.2.

21–21
Chapter 21: FEM PROGRAM FOR SPACE TRUSSES

Table E21.1 Cross section areas of transmission tower members

Element A (sq in) Element A (sq in) Element A (sq in)

1 0.033 10 0.010 19 1.760
2 2.015 11 0.010 20 1.760
3 2.015 12 0.014 21 1.760
4 2.015 13 0.014 22 2.440
5 2.015 14 0.980 23 2.440
6 2.823 15 0.980 24 2.440
7 2.823 16 0.980 25 2.440
8 2.823 17 0.980
9 2.823 18 1.760

Table E21.2 Applied load case for transmission tower

Node x-load (lb) y-load (lb) z-load (lb)

1 1000 10000 −5000
2 0 10000 −5000
3 500 0 0
6 500 0 0
Applied forces at all other nodes are zero.
Own-weight loads not considered.

The members are aluminum tubes with the cross sections listed in Table E21.1.5 The modulus of elasticity is
E = 107 psi for all members. The specific weight is 0.1 lb/in3 . The applied load case to be studied is given
in Table E21.2.
Analyze the transmission tower using the program provided in the SpaceTruss.nb Notebook (downloadable
from Chapter 21 Index). Results to report: driver program cell, node displacements and element stresses.
(More details on HW assignment sheet.) Note: as a quick check on model preparation, the total weight of the
tower should be 555.18 lb.
A Fix for Version 6. Cell 0 added to Notebook SpaceTruss.nb on November 1, 2008 so that plots will
display correctly under Mathematica Version 6.0 and later versions. If plots dont show up correctly please
notify instructor.

5 Data taken from an optimal design reported by Venkayya, Khot and Reddy in 1968 [801].

21–22
 22
.

FEM Programs
for Plane Trusses
and Frames

22–1
Chapter 22: FEM PROGRAMS FOR PLANE TRUSSES AND FRAMES 22–2

TABLE OF CONTENTS

Page
§22.1. Introduction 22–3
§22.2. Analysis Stages 22–3
§22.3. Analysis Support Modules 22–3
§22.3.1. Assembling the Master Stiffness . . . . . . . . . . . 22–3
§22.3.2. Modifying the Master Stiffness Equations . . . . . . . 22–5
§22.3.3. Internal Force Recovery . . . . . . . . . . . . . . 22–6
§22.3.4. Graphic Modules . . . . . . . . . . . . . . . 22–6
§22.4. A Bridge Truss Example 22–6
§22. Exercises . . . . . . . . . . . . . . . . . . . . . . 22–11

22–2
22–3 §22.3 ANALYSIS SUPPORT MODULES

§22.1. Introduction

This Chapter presents a complete FEM program for analysis of plane trusses, programmed in
Mathematica. The program includes some simple minded graphics, including animation. Exercises
22.2-22.4 discuss a complete FEM program for frame analysis for homework assignments.
The description is done in “bottom up” fashion. That means the basic modules are presented, then
the driver program. Graphic modules are provided in posted Notebooks but not explained.

§22.2. Analysis Stages

As in all FEM programs, the analysis of a structure by the Direct Stiffness Method involves three
major stages: (I) preprocessing or model definition, (II) processing, and (III) postprocessing.
The preprocessing portion of the plane truss analysis is done by the driver program, which directly
sets the data structures.
I.1 Model definition by direct setting of the data structures.
I.2 Plot of the FEM mesh, including nodes and element labels.
The processing stage involves three steps:
II.1 Assembly of the master stiffness matrix, with a subordinate element stiffness module.
II.2 Modification of master stiffness matrix and node force vector for displacement boundary
conditions.
II.3 Solution of the modified equations for displacements. For the programs presented here the
built in Mathematica function LinearSolve is used.
Upon executing these three processing steps, the displacements are available The following post-
processing steps may follow:
III.1 Recovery of forces including reactions, done through a Ku matrix multiplication.
III.2 Computation of internal (axial) forces in truss members.
III.3 Plotting deflected shapes and member stress levels.
These steps will be demonstrated in class from a laptop computer.

§22.3. Analysis Support Modules

We begin by listing here the modules that support various analysis steps, and which are put into
separate cells for testing convenience.

§22.3.1. Assembling the Master Stiffness

The function PlaneTrussMasterStiffness, listed in Figure 22.1, assembles the master stiffness
matrix of a plane truss. It uses the element stiffness formation module PlaneBar2Stiffness
described in the previous Chapter. The statements at the end of the cell test this module by forming
and printing K of the example truss.
The arguments of PlaneTrussMasterStiffness are

22–3
Chapter 22: FEM PROGRAMS FOR PLANE TRUSSES AND FRAMES 22–4

PlaneTrussMasterStiffness[nodcoor_,elenod_,
elemat_,elefab_,eleopt_]:=Module[
{numele=Length[elenod],numnod=Length[nodcoor],
e,eNL,eftab,ni,nj,i,j,ncoor,mprop,fprop,opt,Ke,K},
K=Table[0,{2*numnod},{2*numnod}];
For [e=1, e<=numele, e++,
eNL=elenod[[e]]; {ni,nj}=eNL;
eftab={2*ni-1,2*ni,2*nj-1,2*nj};
ncoor={nodcoor[[ni]],nodcoor[[nj]]};
mprop=elemat[[e]]; fprop=elefab[[e]]; opt=eleopt;
Ke=PlaneBar2Stiffness[ncoor,mprop,fprop,opt];
neldof=Length[Ke];
For [i=1, i<=neldof, i++, ii=eftab[[i]];
For [j=i, j<=neldof, j++, jj=eftab[[j]];
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]]
];
];
]; Return[K];
];
PlaneBar2Stiffness[ncoor_,mprop_,fprop_,opt_]:= Module[
{x1,x2,y1,y2,x21,y21,Em,Gm,rho,alpha,A,numer,L,LL,LLL,Ke},
{{x1,y1},{x2,y2}}=ncoor; {x21,y21}={x2-x1,y2-y1};
{Em,Gm,rho,alpha}=mprop; {A}=fprop; {numer}=opt;
If [numer,{x21,y21,Em,A}=N[{x21,y21,Em,A}]];
LL=x21^2+y21^2; L=PowerExpand[Sqrt[LL]]; LLL=Simplify[LL*L];
Ke=(Em*A/LLL)*{{ x21*x21, x21*y21,-x21*x21,-x21*y21},
{ y21*x21, y21*y21,-y21*x21,-y21*y21},
{-x21*x21,-x21*y21, x21*x21, x21*y21},
{-y21*x21,-y21*y21, y21*x21, y21*y21}};
Return[Ke]
];
nodcoor={{0,0},{10,0},{10,10}};
elenod= {{1,2},{2,3},{1,3}};
elemat= Table[{100,0,0,0},{3}];
elefab= {{1},{1/2},{2*Sqrt[2]}};
eleopt= {True};
K=PlaneTrussMasterStiffness[nodcoor,elenod,
elemat,elefab,eleopt];
Print["Master Stiffness of Example Truss:"];
Print[K//MatrixForm];

Figure 22.1. Master stiffness assembly module, with test statements in red.

nodcoor Nodal coordinates arranged as a two-dimensional list:

{ { x1,y1 },{ x2,y2 }, . . . { xn,yn } },
where n is the total number of nodes.
elenod Element end nodes arranged as a two-dimensional list:
{ { i1,j1 }, { i2,j2 }, . . . { ie,je } },
where e is the total number of elements.
elemat Element material properties arranged as a two-dimensional list:
{ { Em1,Gm1,rho1,alpha1 }, . . . { Eme,Gme,rhoe,alphae } },
where e is the total number of elements. Only the elastic modulus Em is used in this
program. For the other properties zeros may be stored as placeholders.

22–4
22–5 §22.3 ANALYSIS SUPPORT MODULES

ModifiedMasterStiffness[pdof_,K_] := Module[
{i,j,k,n=Length[K],np=Length[pdof],Kmod}, Kmod=K;
For [k=1,k<=np,k++, i=pdof[[k]];
For [j=1,j<=n,j++, Kmod[[i,j]]=Kmod[[j,i]]=0];
Kmod[[i,i]]=1
];
Return[Kmod]
];
ModifiedNodeForces[pdof_,f_] := Module[
{i,k,np=Length[pdof],fmod}, fmod=f;
For [k=1,k<=np,k++, i=pdof[[k]]; fmod[[i]]=0];
Return[fmod]
];
K=Array[Kij,{6,6}];
Print["Assembled Master Stiffness:"];Print[K//MatrixForm];
K=ModifiedMasterStiffness[{1,2,4},K];
Print["Master Stiffness Modified For Displacement B.C.:"];
Print[K//MatrixForm];
f=Array[fi,{6}];
Print["Node Force Vector:"]; Print[f];
f=ModifiedNodeForces[{1,2,4},f];
Print["Node Force Vector Modified For Displacement B.C.:"];
Print[f];

Figure 22.2. Modifying the master stiffness and node force vector for displacement
boundary conditions, with test statements in red.

elefab Element fabrication properties arranged as a two-dimensional list:

{ { A1 }, . . . { Ae } },
where e is the total number of elements, and A the cross section area.
eleopt Element processing option: set to { True } to tell PlaneBar2Stiffness to carry
out element stiffness computations in floating-point arithmetic. Else set to { False }
to keep computations in exact arithmetic or symbolic form.
The assembler uses the freedom-pointer-table technique described in §3.4 for merging the element
stiffness matrix into the master stiffness. The module returns the master stiffness matrix K in list
K, which is stored as a full matrix.
The statements at the end of Figure 22.1 test this module by forming and printing the master stiffness
matrix of the example truss. These statements are executed when the cell is initialized.

§22.3.2. Modifying the Master Stiffness Equations

Following the assembly process the master stiffness equations Ku = f must be modified to account
for displacement boundary conditions. This is done through the computer-oriented equation modi-
fication process described in §3.4.2. Modules that perform this operation are listed in Figure 22.2,
along with test statements.
Module ModifiedMasterStiffness carries out this process for the master stiffness matrix
K, whereas ModifiedNodalForces does this for the nodal force vector f. The logic of
ModifiedNodalForces is considerably simplified by assuming that all prescribed displacements
are zero, that is, the BCs are homogeneous. This is the case in the implementation shown here.
Module ModifiedMasterStiffness receives two arguments:

22–5
Chapter 22: FEM PROGRAMS FOR PLANE TRUSSES AND FRAMES 22–6

pdof A list of the prescribed degrees of freedom identified by their global number. For
the example truss of Chapters 2-3 this list would have three entries: {1, 2, 4}, which
are the freedom numbers of u x1 , u y1 and u y2 in u.
K The master stiffness matrix K produced by the assembler module described in the
previous subsection.
The module clears appropriate rows and columns of K, places ones on the diagonal, and returns the
thus modified K as function value. Note the use of the Mathematica function Length to control
loops: np=Length[pdof] sets np to the number of prescribed freedoms. Similarly n=Length[K]
sets n to the order of the master stiffness matrix K, which is used to bound the row and column
clearing loop. These statements may be placed in the list that declares local variables.
Module ModifiedNodalForces has a very similar structure and logic and need not be described in
detail. It is important to note, however, that for homogeneous BCs the two module are independent
of each other and may be called in any order. On the other hand, if there were nonzero prescribed
displacements the force modification must be done before the stiffness modification. This is because
stiffness coefficients that are cleared in the latter are needed for modifying the force vector.
The test statements at the bottom of Figure 22.2 are chosen to illustrate another feature of Math-
ematica: the use of the Array function to generate subscripted symbolic arrays of one and two
dimensions. The test output should be self explanatory and is not shown here. Both the force vector
and its modified form are printed as row vectors to save space.
§22.3.3. Internal Force Recovery

Module PlaneTrussIntForces listed in Figure 22.3 computes the internal forces (axial forces)
in all truss members. The first five arguments are the same as for the assembler routine described
previously. The last argument, u, contains the computed node displacements arranged as a flat, one
dimensional list:
{ ux1, uy1 . . . uxn, uyn } (22.1)
PlaneTrussIntForces makes use of PlaneBar2IntForce, which computes the internal force
in an individual member. PlaneBar2IntForce is similar in argument sequence and logic to
PlaneBar2Stiffness of Figure 22.1. The first four arguments are identical. The last argument,
ue, contains the list of the four element node displacements in the global system. The logic of the
recovery module is straightforward and follows the method outlined in §3.2.
The statements at the bottom of Figure 22.3 test the internal
√ force recovery for the example truss of
Chapters 2-3, a and should return forces of 0,, −1 and 2 2 for members 1, 2 and 3, respectively.
§22.3.4. Graphic Modules
Graphic modules that support preprocessing are placed in Cells 4A, 4B and 4C of the Mathematica
notebook. These plot unlabeled elements, elements and nodes with labels, and boundary conditions.
Graphic modules that support postprocessing are placed in Cells 5A and 5B of the Notebook. These
plot deformed shapes and axial stress levels in color.
These modules are not listed since they are still undergoing modifications at the time of this writing.
One unresolved problem is to find a way for absolute placement of supported nodes for correct
deflected-shape animations.

22–6
22–7 §22.4 A BRIDGE TRUSS EXAMPLE

PlaneTrussIntForces[nodcoor_,elenod_,elemat_,elefab_,
eleopt_,u_]:= Module[{numele=Length[elenod],
numnod=Length[nodcoor],e,eNL,eftab,ni,nj,i,
ncoor,mprop,fprop,opt,ue,p},
p=Table[0,{numele}]; ue=Table[0,{4}];
For [e=1, e<=numele, e++,
eNL=elenod[[e]]; {ni,nj}=eNL;
eftab={2*ni-1,2*ni,2*nj-1,2*nj};
ncoor={nodcoor[[ni]],nodcoor[[nj]]};
mprop=elemat[[e]]; fprop=elefab[[e]]; opt=eleopt;
For [i=1,i<=4,i++, ii=eftab[[i]]; ue[[i]]=u[[ii]]];
p[[e]]=PlaneBar2IntForce[ncoor,mprop,fprop,opt,ue]
];
Return[p]
];
PlaneBar2IntForce[ncoor_,mprop_,fprop_,opt_,ue_]:= Module[
{x1,x2,y1,y2,x21,y21,Em,Gm,rho,alpha,A,numer,LL,pe},
{{x1,y1},{x2,y2}}=ncoor; {x21,y21}={x2-x1,y2-y1};
{Em,Gm,rho,alpha}=mprop; {A}=fprop; {numer}=opt;
(*If [numer,{x21,y21,Em,A}=N[{x21,y21,Em,A}]];*)
LL=x21^2+y21^2;
pe=Em*A*(x21*(ue[[3]]-ue[[1]])+y21*(ue[[4]]-ue[[2]]))/LL;
Return[pe]
];
nodcoor={{0,0},{10,0},{10,10}}; elenod= {{1,2},{2,3},{1,3}};
elemat= Table[{100,0,0,0},{3}]; elefab= {{1},{1/2},{2*Sqrt[2]}};
eleopt= {True}; u={0,0,0,0,0.4,-0.2};
p=PlaneTrussIntForces[nodcoor,elenod,elemat,elefab,eleopt,u];
Print["Int Forces of Example Truss:"];
Print[p];

Figure 22.3. Calculation of truss internal forces, with test statements in red.

§22.4. A Bridge Truss Example

The driver program in Cell 6 defines and runs the analysis of the 6-bay bridge truss problem defined
in Figure 22.4. This truss has 12 nodes and 17 elements. It is fixed at node 1 and on rollers at node
12.
The driver is listed in Figures 22.5 and 22.6. It begins by defining the problem through specification
of the following data structures:
NodeCoordinates Same configuration as nodcoor
ElemNodeLists Same configuration as elenod
ElemMaterials Same configuration as elemat
ElemFabrication Same configuration as elefab. This list is built up from four repeating cross
sectional areas: Abot, Atop, Abat and Adia, for the areas of bottom longerons, top longerons,
battens and diagonals, respectively.
ProcessOptions Same configuration as eleopt
FreedomValue. This is a two-dimensional list that specifies freedom values, node by node. It is
initialized to zero on creation, then the value of nonzero applied loads is set at nodes 3, 5, 7, 9 and
11. For example FreedomValue[[7]] = { 0,-16 } specifies f x7 = 0 and f y7 = −16.
FreedomTag. This is a two-dimensional list that specifies, for each nodal freedom, whether the value

22–7
Chapter 22: FEM PROGRAMS FOR PLANE TRUSSES AND FRAMES 22–8

(a) Elastic modulus E = 1000

Cross section areas of bottom longerons: 2,
top longerons: 10, battens: 3, diagonals: 1 top joints lie
on parabolic
profile
9

10 10 16 10 10
span: 6 bays @ 10 = 60

(b) (9) 6 (10) 8

4
y (8) (11)
2 (15) 10
(7) (14) (16)
;;
;;

;;
;;
(12)
(13) (18) (19) (20) (21) (17)
1 x 12
(1) 3 (2) 5 (3) 7 (4) 9 (5) 11 (6)

Figure 22.4. Six-bay bridge truss used as example problem: (a) truss structure showing
supports and applied loads; (b) finite element idealization as pin-jointed truss.

of the force (tag 0) or displacement (tag 1) is prescribed. It is initialized to zero on creation, then the
support tags at nodes 1 (fixed) and 12 (horizontal roller) are set. For example, FreedomTag[[1]]
= { 1,1 } says that both node displacement components u x1 and u y1 of node 1 are specified.
Processing commands are listed in Figure 22.6. The master stiffness matrix is assembled by the
module listed in Figure 22.1. The stiffness matrix is placed in K. The applied force vector stored
as a one-dimensional list, is placed in f, which results from the application of built-in function
Flatten to Freedom Value.
The prescribed degree-of-freedom array pdof is constructed by scanning FreedomTag for nonzero
values. For this problem pdof={ 1,2,23 }. The displacement boundary conditions are applied by
the modules of Figure 22.2, which return the modified master stiffness Kmod and the modified node
force vector fmod. Note that the modified stiffness matrix is stored into Kmod rather than K to save
the original form of the master stiffness for the recovery of reaction forces later.
The complete displacement vector is obtained by the matrix calculation

u=LinearSolve[Kmod,fmod] (22.2)

which takes advantage of the built-in linear solver provided by Mathematica.

The remaining calculations recover the node vector including reactions by the matrix-vector multiply
f = K.u (recall that K contains the unmodified master stiffness matrix). The member internal forces
p are obtained through the module listed in Figure 22.3. The program prints u, f and p as row
vectors to conserve space.
Running the program of Figures 22.5–6 produces the output shown in Figure 22.7. Output plot
results are collected in Figure 22.8.

22–8
22–9 §22.4 A BRIDGE TRUSS EXAMPLE

ClearAll[];
NodeCoordinates={{0,0},{10,5},{10,0},{20,8},{20,0},{30,9},
{30,0},{40,8},{40,0},{50,5},{50,0},{60,0}};
ElemNodeLists= {{1,3},{3,5},{5,7},{7,9},{9,11},{11,12},
{1,2},{2,4},{4,6},{6,8},{8,10},{10,12},
{2,3},{4,5},{6,7},{8,9},{10,11},
{2,5},{4,7},{7,8},{9,10}};
numnod=Length[NodeCoordinates];
numele=Length[ElemNodeLists]; numdof=2*numnod;
ElemMaterial= Table[{1000,0,0,0},{numele}];
Abot=2; Atop=10; Abat=3; Adia=1;
ElemFabrication=Join[Table[{Abot},{6}],Table[{Atop},{6}],
Table[{Abat},{5}],Table[{Adia},{4}]];
ProcessOptions= {True}; aspect=0;
PlotLineElements[NodeCoordinates,ElemNodeLists,aspect,
"test mesh"];
PlotLineElementsAndNodes[NodeCoordinates,ElemNodeLists,aspect,
"test mesh with elem & node labels",{True,0.12},{True,0.05}];

FreedomTag=FreedomValue=Table[{0,0},{numnod}];
FreedomValue[[3]]={0,-10}; FreedomValue[[5]]={0,-10};
FreedomValue[[7]]={0,-16};
FreedomValue[[9]]={0,-10}; FreedomValue[[11]]={0,-10};
Print["Applied node forces="]; Print[FreedomValue];
FreedomTag[[1]]= {1,1}; (* fixed node 1 *)
FreedomTag[[numnod]]={0,1}; (* hroller @ node 12 *)

Figure 22.5. Driver for analysis of the bridge truss: preprocessing.

f=Flatten[FreedomValue];
K=PlaneTrussMasterStiffness[NodeCoordinates,
ElemNodeLists,ElemMaterial,ElemFabrication,ProcessOptions];
pdof={}; For[n=1,n<=numnod,n++, For[j=1,j<=2,j++,
If[FreedomTag[[n,j]]>0, AppendTo[pdof,2*(n-1)+j]]]];
Kmod=ModifiedMasterStiffness[pdof,K];
fmod=ModifiedNodeForces [pdof,f];
u=LinearSolve[Kmod,fmod]; u=Chop[u];
Print["Computed Nodal Displacements:"]; Print[u];
f=Simplify[K.u]; f=Chop[f];
Print["External Node Forces Including Reactions:"]; Print[f];
p=PlaneTrussIntForces[NodeCoordinates,ElemNodeLists,
ElemMaterial,ElemFabrication,eleopt,u]; p=Chop[p];
sigma=Table[p[[i]]/ElemFabrication[[i,1]],{i,1,numele}];
Print["Internal Member Forces:"]; Print[p];
PlotTrussDeformedShape[NodeCoordinates,ElemNodeLists,u,
1.0,aspect,"Deformed shape"];
PlotAxialStressLevel[NodeCoordinates,ElemNodeLists,sigma,
1.0,aspect,"Axial stress level"];

Figure 22.6. Driver for analysis of the bridge truss: processing and postprocessing.

22–9
Chapter 22: FEM PROGRAMS FOR PLANE TRUSSES AND FRAMES 22–10

Applied node forces

0, 0,
0, 0,
0, 10,
0, 0,
0, 10,
0, 0,

0, 16,
0, 0,
0, 10,
0, 0,
0, 10,
0, 0
Computed Nodal Displacements:

0, 0, 0.809536, 1.7756, 0.28, 1.79226, 0.899001,
2.29193, 0.56, 2.3166, 0.8475, 2.38594,
0.8475, 2.42194, 0.795999, 2.29193, 1.135, 2.3166,
0.885464, 1.7756, 1.415, 1.79226, 1.695, 0
External Node Forces Including Reactions:

0, 28., 0, 0, 0, 10., 0, 0, 0, 10., 0, 0, 0,
16., 0, 0, 0, 10., 0, 0, 0, 10., 0, 28.
Internal Member Forces:

56., 56., 57.5, 57.5, 56., 56., 62.6099,
60.0318, 60.2993, 60.2993,
60.0318, 62.6099, 10., 9.25, 12., 9.25,
10., 1.67705, 3.20156, 3.20156, 1.67705

Figure 22.7. Printed output from the bridge truss analysis.

test mesh

test mesh with elem & node labels

4 6 8
9 10
2 8 11 10
14 19 15 20 16
7 13 18 21 17 12
1 1 3 2 5 3 7 4 9 5 11 6 12

Deformed shape

test axial stress level plot

Figure 22.8. Graphics output from the bridge truss analysis.

22–10
22–11 Exercises

Homework Exercises for Chapter 22

FEM Programs for Plane Trusses and Frames

EXERCISE 22.1
Placeholder.
EXERCISE 22.2 [C:25] Using the PlaneTruss.nb Notebook posted on the web site as guide, complete
the PlaneFrame.nb Notebook that implements the analysis of an arbitrary plane frame using the plane beam
stiffness presented in Chapter 21. To begin this homework, download the PlaneFrame.nb Notebook file from
the web site.
The modification involves selected Cells of the Notebook, as described below. For additional details refer to
the Notebook; some of the modifications involve only partially filled Modules.
Cell 1: Assembler. The element stiffness routine is the beam-column stiffness presented in Chapter 21. This
module, called PlaneBeamColumn2Stiffness, is included in Cell 1 of the PlaneFrame.nb notebook linked
from the Chapter 22 index.1 In modifying the assembler module, remember that there are three degrees of
freedom per node: u xi , u yi and θi , not just two.
The incomplete assembler is shown in Figure E22.1. The test statements are shown in red after the module.
The lower cell shows the test output produced by a correctly completed module.
Cell 2: BC Applicator. No modifications are necessary. The two modules should work correctly for this
problem since they dont assume anything about freedoms per nodes. The same test statements can be kept.
Cell 3: Internal Force Recovery. Both modules require modifications because the beam has 3 internal forces:
(i) the axial force (which is recovered exactly as for bars), (ii) the bending moment m i = E I κi at end node i,
and (iii) the bending moment m j = E I κ j at end node j.2 Furthermore the element displacement vector has
six degrees of freedom, not just four. Test the modules on a simple beam problem.
The incomplete internal force module is shown in Figure E22.2. The test statements are shown in red after the
module. The lower cell shows the test output produced by a correctly completed module.
Cell 4-5: Graphic Modules. These are now provided ready to use, and should not be touched.
Cell 6: Driver Program: use this for Exercise 22.3. Do not touch for this one.
As solution to this Exercise return a listing of the completed Cells 1 and 3, along with the output of the test
statements for those cells.
EXERCISE 22.3
[C:25] Use the program developed in the previous Exercise to analyze the plane frame shown in Figure E22.3.
The frame is fixed3 at A, B and C. It is loaded by a downward point load P at G and by a horizontal point
load 12 P at D. All members have the same cross section a × a for simplicity, and the material is the same.

1 If you want to extract an individual cell from a Notebook to work on a separate file (a good idea for working on groups)
select the cell, then pick SaveSelectionAs -> Plain Text from the Edit menu, give it a file name in the pop-up dialogue
box, and save it.
2 Two end moments are required because the beam element used here has linearly varying curvatures and hence moments.
To recover the end moments refer to the pertinent equations in Chapter 13. The steps are as follows. For element e
recover the transverse displacements ū yi and ū y j in the local element system x̄ (e) , ȳ (e) . Complete these with rotations
θi and θ j (which do not need to be transformed) to form the 4 × 1 beam local node displacement vector. Then apply the
curvature-displacement relation (13.13) at ξ = −1 and ξ = 1 to get the two end curvatures κi(e) and κ (e) j , respectively.
(e)
Finally multiply those curvatures by E (e) Izz to get the bending moments.
3 For a plane frame, a fixed condition, also known as clamped condition, means that the x, y displacements and the rotation
about z are zero.

22–11
Chapter 22: FEM PROGRAMS FOR PLANE TRUSSES AND FRAMES 22–12

(* PlaneBeamColumn2Stiffness module goes here *)

PlaneFrameMasterStiffness[nodcoor_,elenod_,
elemat_,elefab_,eleopt_]:=Module[
{numele=Length[elenod],numnod=Length[nodcoor],
e,eNL,eftab,ni,nj,i,j,ncoor,mprop,fprop,opt,Ke,K},
K=Table[0,{3*numnod},{3*numnod}];
For [e=1, e<=numele, e++,
(* missing statements *)
Ke=PlaneBeamColumn2Stiffness[ncoor,mprop,fprop,opt];
(* missing statements *)
];
Return[K]
];
nodcoor={{0,0},{10,0},{10,10}};
elenod= {{1,2},{2,3},{1,3}}; elemat= Table[{100,0,0,0},{3}];
elefab= {{1,10},{1/2,10},{2*Sqrt[2],10}}; eleopt= {True};
K=PlaneFrameMasterStiffness[nodcoor,elenod,elemat,elefab,eleopt];
Print["Master Stiffness of Example Frame:"];
Print[K//MatrixForm]; Print[Chop[Eigenvalues[K]]];
Master Stiffness of Example Frame:

22.1213 7.87868
21.2132
10. 0. 0.
12.1213
7.87868
21.2132 







 7.87868 24.1213 81.2132 0.
12. 60.
7.87868
12.1213 21.2132 



 



21.2132 81.2132 682.843 0.
60. 200. 21.2132
21.2132 141.421  










10. 0. 0. 22. 0.
60.
12. 0.
60. 









 0.
12.
60. 0. 17.
60. 0.
5. 0. 









 0. 60. 200.
60.
60. 800. 60. 0. 200. 


 


12.1213
7.87868 21.2132
12.
 0. 60. 24.1213 7.87868 81.2132 











7.87868
12.1213
21.2132 0.
5. 0. 7.87868 17.1213
21.2132 

 

21.2132 21.2132 141.421
60. 0. 200. 81.2132
21.2132 682.843 
1121.7, 555.338, 531.652, 46.6129, 22.4971, 14.366, 0, 0, 0

Figure E22.1. The incomplete assembler module for Exercise 22.2, with test statements in red.
Module PlaneBeamColumn2Stiffness, which goes in the upper portion of the cell is omitted to save
space; that module was listed in Figure 21.9. Output cell results are produced by a correct module.

The SI physical units to be used are: mm for lengths, N for forces, and MPa=N/mm2 for elastic moduli. For
the calculations use the following numerical data: L = 10,000 mm (10 m), H = 6,000 (6 m), a = 500 mm
(0.5 m), P = 4,800 N, E = 35,000 MPa (high strength concrete). The member cross section area is A = a 2 ,
and the flexural moment of inertia about the neutral axis is Izz = a 4 /12.
The recommended finite element discretization of two elements per member is shown in Figure E22.4.
As solution to this exercise list the driver program you used and the displacement and internal force outputs.
The results of primary interest are:
1. The horizontal displacement at D, and the vertical displacement at G (mm).
2. The axial forces (N) in columns AD, B E and C F, with appropriate sign4 .
3. The maximum bending moment (N.mm) over the floor members D E and E F, and the maximum bending
moment in the columns AD, B E and C F.5
Provide deformed shape plot (but not the animation) and the frame stress level plots as part of the homework
results. Note: the Notebook on the web contains some of the actual plots produced by the complete Notebook,

4 Plus for tension, minus for compression

5 The bending moment varies linearly over each element. It should be continuous at all joints except E.

22–12
22–13 Exercises

PlaneFrameIntForces[nodcoor_,elenod_,elemat_,elefab_,
eleopt_,u_]:= Module[{numele=Length[elenod],
numnod=Length[nodcoor],e,eNL,eftab,ni,nj,i,
ncoor,mprop,fprop,opt,ue,p},
p=Table[0,{numele}]; ue=Table[0,{6}];
For [e=1, e<=numele, e++,
eNL=elenod[[e]]; {ni,nj}=eNL;
ncoor={nodcoor[[ni]],nodcoor[[nj]]};
mprop=elemat[[e]]; fprop=elefab[[e]]; opt=eleopt;
eftab={3*ni-2,3*ni-1,3*ni,3*nj-2,3*nj-1,3*nj};
For [i=1,i<=6,i++, ii=eftab[[i]]; ue[[i]]=u[[ii]]];
p[[e]]=PlaneBeamColumn2IntForces[ncoor,mprop,fprop,opt,ue]
];
Return[p]
];
PlaneBeamColumn2IntForces[ncoor_,mprop_,fprop_,opt_,ue_]:=
Module[{x1,x2,y1,y2,x21,y21,Em,Gm,rho,alpha,A,Izz,num,LL,L,
dvy,cv1,cv2,pe=Table[0,{3}]},
{{x1,y1},{x2,y2}}=ncoor; {x21,y21}={x2-x1,y2-y1};
{Em,Gm,rho,alpha}=mprop; {A,Izz}=fprop; {num}=opt;
If [num,{x21,y21,Em,A,Izz}=N[{x21,y21,Em,A,Izz}]];
(* missing statements for Exercise 22.2 *)
Return[pe]
];
ClearAll[L,Em,A1,A2,Izz1,Izz2,Izz3,uz1,uz2,uz2,uz3];
nodcoor={{0,0},{L,0},{L,L}}; elenod= {{1,2},{2,3},{1,3}};
elemat= Table[{Em,0,0,0},{3}];
elefab= {{A1,Izz1},{A2,Izz2},{A3,Izz3}}; eleopt= {False};
u={0,uy1,0, 0,uy2,0, 0,uy3,0};
p=PlaneFrameIntForces[nodcoor,elenod,elemat,elefab,eleopt,u];
Print["Int Forces of Example Frame:"]; Print[p];
Int Forces of Example Frame:
6 Em Izz1

uy1  uy2 6 Em Izz1

uy1  uy2 A2 Em

uy2  uy3
0,  ,

 , 
 , 0, 0,
L2 L2 L
A3 Em

uy1  uy3 3 Em Izz3

uy1  uy3 3 Em Izz3

uy1  uy3

 , 
   ,

   
2L 2 L2 2 L2

Figure E22.2. Figure E22.2. Incomplete element internal force recovery module for
Exercise 22.2. Test statements in red. Results in output cell are those produced by a
correct module.

to be used as targets.
Partial answers for this exercise:
Largest vertical displacement: -0.232 mm (⇓) at node 4
Largest negative bending moment: −6.8 × 106 N.mm, at node 5 of element (4)
Axial forces: −2651 N over (3) and (4)

EXERCISE 22.4 [D:20] Explain why the solution given by the FEM model of Figure E22.3 is exact for the
Bernoulli-Euler bending model; that is, cannot be improved by subdividing each element into more elements.

22–13
Chapter 22: FEM PROGRAMS FOR PLANE TRUSSES AND FRAMES 22–14

P
L /2 y
P/2 cross section
of all members:
D G E H F
a
H a
Origin of x,y
A B x C is arbitrary.
L L

Figure E22.3. Plane frame structure for Exercise 22.3.

P
P/2 3 (3) (4) 5 (5) (6) 7
4 6
(2) (9) (7)

2 10 8
(1) (10) (8)
1 11 9

Figure E22.4. Recommended FEM discretization for the

plane frame of the previous figure.

22–14
 23
.

Implementation of
Iso-P Quadrilateral
Elements

23–1
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–2

TABLE OF CONTENTS

Page
§23.1. Introduction 23–3
§23.2. Bilinear Quadrilateral Stiffness Matrix 23–3
§23.2.1. Gauss Quadrature Rule Information . . . . . . . . . . 23–3
§23.2.2. Shape Function Evaluation . . . . . . . . . . . . 23–5
§23.2.3. Element Stiffness . . . . . . . . . . . . . . . . 23–5
§23.3. Test of Bilinear Quadrilateral 23–7
§23.3.1. Test of Rectangular Geometry . . . . . . . . . . . 23–7
§23.3.2. Test of Trapezoidal Geometry . . . . . . . . . . . . 23–8
§23.4. *Consistent Node Forces for Body Force Field 23–10
§23.5. *Recovery of Corner Stresses 23–10
§23.6. *Quadrilateral Coordinates of Given Point 23–12
§23. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 23–12
§23. References . . . . . . . . . . . . . . . . . . . . . . 23–13
§23. Exercises . . . . . . . . . . . . . . . . . . . . . . 23–14

23–2
23–3 §23.2 BILINEAR QUADRILATERAL STIFFNESS MATRIX

§23.1. Introduction

This Chapter illustrates, through a specific example, the computer implementation of isoparametric
quadrilateral elements for the plane stress problem. Triangles, which present some programming
quirks, are covered in the next Chapter.
The programming example is that of the four-node bilinear quadrilateral. It covers the computation
of the element stiffness matrix, the consistent node force vector for a body forces, and stress recovery
at selected points. The organization of the computations is typical of isoparametric-element modules
in any number of space dimensions.

§23.2. Bilinear Quadrilateral Stiffness Matrix η

We consider here the implementation of the 4-node bilin- η=1 3 (x 3 , y 3 )
ear quadrilateral for plane stress, depicted in Figure 23.1. 4 (x 4 , y 4 )
The element stiffness matrix of simple one-dimensional ξ=1
elements, such as the ones discussed in Chapters 20– ξ
22, can be easily packaged in a simple module. For ξ=−1
multidimensional elements, however, it is convenient to
break up the implementation into application dependent 1 (x 1 , y 1 ) η=−1
and application independent modules, as flowcharted in 2 (x 2 , y 2 )
Figure 23.2. The application independent modules can
Figure 23.1. The 4-node
be “reused” in other FEM applications, for example to
bilinear quadrilateral element.
construct thermal, fluid or electromagnetic elements.
For the bilinear quadrilateral stiffness computations, the separation of Figure 23.2 is done by dividing
the work into three modules:
Quad4IsoPMembraneStiffness. Computes the element stiffness matrix Ke of a four-node
isoparametric quadrilateral element in plane stress.
QuadGuassRuleInfo. Returns two-dimensional Gauss quadrature formula of product type.
Quad4IsoPShapeFunDer. Evaluates the shape functions of a four-node isoparametric quadrilat-
eraland their x/y derivatives, at a specific point.
These modules are described in further detail in the following subsections, in a “bottom up” fashion.

§23.2.1. Gauss Quadrature Rule Information

Recall from §17.3 that Gauss quadrature rules for isoparametric quadrilateral elements have the
canonical form
 1 1  1  1
. 
p1 p2
F(ξ, η) dξ dη = dη F(ξ, η) dξ = wi w j F(ξi , η j ). (23.1)
−1 −1 −1 −1 i=1 j=1

Here F = hBT EB J is the matrix to be integrated, whereas p1 and p2 are the number of Gauss
points in the ξ and η directions, respectively. Often, but not always, the same number p = p1 = p2
is chosen in both directions. A formula with p1 = p2 is called an isotropic integration rule because
directions ξ and η are treated alike.

23–3
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–4

QuadGaussRuleInfo is an application
APPLICATION DEPENDENT
independent module that implements the
two-dimensional product Gauss rules with
Element
1 through 5 points in each direction. Stiffness Module
The number of points in each direction
may be the same or different. Usage of
this module was described in detail in
§17.3.4. For the readers convenience it is
listed, along with its subordinate module Shape Gauss Quadrature
Function Module Information Module
LineGaussRuleInfo, in Figure 23.3.
This module is classified as application
APPLICATION INDEPENDENT
independent since it can reused for any
quadrilateral element. Figure 23.2. Separation of element stiffness modules into
application-dependent and application-independent levels.

QuadGaussRuleInfo[{rule_,numer_},point_]:= Module[
{ξ ,η,p1,p2,i,j,w1,w2,m,info={{Null,Null},0}},
If [Length[rule]==2, {p1,p2}=rule, p1=p2=rule];
If [p1<0, Return[QuadNonProductGaussRuleInfo[
{-p1,numer},point]]];
If [Length[point]==2, {i,j}=point, m=point;
j=Floor[(m-1)/p1]+1; i=m-p1*(j-1) ];
{ξ ,w1}= LineGaussRuleInfo[{p1,numer},i];
{η,w2}= LineGaussRuleInfo[{p2,numer},j];
info={{ξ ,η},w1*w2};
If [numer, Return[N[info]], Return[Simplify[info]]];
];
LineGaussRuleInfo[{rule_,numer_},point_]:= Module[
{g2={-1,1}/Sqrt[3],w3={5/9,8/9,5/9},
g3={-Sqrt[3/5],0,Sqrt[3/5]},
w4={(1/2)-Sqrt[5/6]/6, (1/2)+Sqrt[5/6]/6,
(1/2)+Sqrt[5/6]/6, (1/2)-Sqrt[5/6]/6},
g4={-Sqrt[(3+2*Sqrt[6/5])/7],-Sqrt[(3-2*Sqrt[6/5])/7],
Sqrt[(3-2*Sqrt[6/5])/7], Sqrt[(3+2*Sqrt[6/5])/7]},
g5={-Sqrt[5+2*Sqrt[10/7]],-Sqrt[5-2*Sqrt[10/7]],0,
Sqrt[5-2*Sqrt[10/7]], Sqrt[5+2*Sqrt[10/7]]}/3,
w5={322-13*Sqrt[70],322+13*Sqrt[70],512,
322+13*Sqrt[70],322-13*Sqrt[70]}/900,
i=point,p=rule,info={{Null,Null},0}},
If [p==1, info={0,2}];
If [p==2, info={g2[[i]],1}];
If [p==3, info={g3[[i]],w3[[i]]}];
If [p==4, info={g4[[i]],w4[[i]]}];
If [p==5, info={g5[[i]],w5[[i]]}];
If [numer, Return[N[info]], Return[Simplify[info]]];
];

Figure 23.3. Module to get Gauss-product quadrature

information for a quadrilateral.

23–4
23–5 §23.2 BILINEAR QUADRILATERAL STIFFNESS MATRIX

Quad4IsoPShapeFunDer[ncoor_,qcoor_]:= Module[
{Nf,dNx,dNy,dNξ ,dNη,i,J11,J12,J21,J22,Jdet,ξ ,η,x,y},
{ξ ,η}=qcoor;
Nf={(1-ξ )*(1-η),(1+ξ )*(1-η),(1+ξ )*(1+η),(1-ξ )*(1+η)}/4;
dNξ ={-(1-η), (1-η),(1+η),-(1+η)}/4;
dNη= {-(1-ξ ),-(1+ξ ),(1+ξ ), (1-ξ )}/4;
x=Table[ncoor[[i,1]],{i,4}]; y=Table[ncoor[[i,2]],{i,4}];
J11=dNξ .x; J12=dNξ .y; J21=dNη.x; J22=dNη.y;
Jdet=Simplify[J11*J22-J12*J21];
dNx= ( J22*dNξ -J12*dNη)/Jdet; dNx=Simplify[dNx];
dNy= (-J21*dNξ +J11*dNη)/Jdet; dNy=Simplify[dNy];
Return[{Nf,dNx,dNy,Jdet}]
];

Figure 23.4. Shape function module for 4-node bilinear quadrilateral.

§23.2.2. Shape Function Evaluation

Quad4IsoPShapeFunDer is an application independent module that computes the shape functions

Nie , i = 1, 2, 3, 4 and its x-y partial derivatives at the sample integration points. The logic, listed
in Figure 23.4, is straightforward and follows closely the description of Chapter 17.
The arguments of the module are the {x, y} quadrilateral corner coordinates, which are passed in
ncoor, and the two quadrilateral coordinates {ξ, η}, which are passed in qcoor. The former have
the same configuration as described for the element stiffness module below.
The quadrilateral coordinates define the element location at which the shape functions and their
derivatives are to be evaluated. For the stiffness formation these are Gauss points, but for strain and
stress computations these may be other points, such as corner nodes.
Quad4IsoPShapeFunDer returns the two-level list { Nf,Nx,Ny,Jdet }, in which the first three are
4-entry lists. List Nf collects the shape function values, Nx the shape function x-derivatives, Ny the
shape function y-derivatives, and Jdet is the Jacobian determinant called J in Chapters 17.

§23.2.3. Element Stiffness

Module Quad4IsoPMembraneStiffness computes the stiffness matrix of a four-noded isopara-

metric quadrilateral element in plane stress. The module configuration is typical of isoparametric
elements in any number of dimensions. It follows closely the procedure outlined in Chapter 17.
The module logic is listed in Figure 23.5. The statements at the bottom of the module box (not
shown in that Figure) test it for specific configurations.
The module is invoked as

Ke=Quad4IsoPMembraneStiffness[ncoor,Emat,th,options] (23.2)

The arguments are:

ncoor Quadrilateral node coordinates arranged in two-dimensional list form:
{ { x1,y1 },{ x2,y2 },{ x3,y3 },{ x4,y4 } }.

23–5
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–6

Quad4IsoPMembraneStiffness[ncoor_,Emat_,th_,options_]:=
Module[{i,k,p=2,numer=False,h=th,qcoor,c,w,Nf,
dNx,dNy,Jdet,Be,Ke=Table[0,{8},{8}]},
If [Length[options]==2, {numer,p}=options,{numer}=options];
If [p<1||p>4, Print["p out of range"]; Return[Null]];
For [k=1, k<=p*p, k++,
{qcoor,w}= QuadGaussRuleInfo[{p,numer},k];
{Nf,dNx,dNy,Jdet}=Quad4IsoPShapeFunDer[ncoor,qcoor];
If [Length[th]==4, h=th.Nf]; c=w*Jdet*h;
Be={Flatten[Table[{dNx[[i]], 0},{i,4}]],
Flatten[Table[{0, dNy[[i]]},{i,4}]],
Flatten[Table[{dNy[[i]],dNx[[i]]},{i,4}]]};
Ke+=Simplify[c*Transpose[Be].(Emat.Be)];
]; Return[Simplify[Ke]]
];

Figure 23.5. Element stiffness formation module for 4-node bilinear quadrilateral.

Emat A two-dimensional list storing the 3 × 3 plane stress matrix of elastic moduli:
 
E 11 E 12 E 13
E = E 12 E 22 E 23 (23.3)
E 13 E 23 E 33
arranged as { { E11,E12,E33 },{ E12,E22,E23 },{ E13,E23,E33 } }. This matrix
must be symmetric. If the material is isotropic with elastic modulus E and Poisson’s
ratio ν, it reduces to
 
E 1 ν 0
E= ν 1 0 (23.4)
1 − ν 2 0 0 1 (1 − ν)
2

th The plate thickness specified either as a four-entry list: { h1,h2,h3,h4 } or as a

scalar: h.
The first form is used to specify an element of variable thickness, in which case the
entries are the four corner thicknesses and h is interpolated bilinearly. The second
form specifies uniform thickness.
options Processing options. This list may contain two items: { numer,p } or one: { numer }.
numer is a logical flag with value True or False. If True, the computations are
done in floating point arithmetic. For symbolic or exact arithmetic work set numer
to False.1
p specifies the Gauss product rule to have p points in each direction. p may be 1
through 4. For rank sufficiency, p must be 2 or higher. If p is 1 the element will be
rank deficient by two.2 If omitted p = 2 is assumed.

1 The reason for this option is speed. A symbolic or exact computation can take orders of magnitude more time than a
floating-point evaluation. This becomes more pronounced as elements get more complicated.
2 The rank of an element stiffness is discussed in Chapter 19.

23–6
23–7 §23.3 TEST OF BILINEAR QUADRILATERAL

(a) 4 3 4 a
(b) 3

a a

1 2 1 2
2a 2a

Figure 23.6. Test quadrilateral element geometries.

The module returns Ke as an 8 × 8 symmetric matrix pertaining to the following arrangement of

nodal displacements:

ue = [ u x1 u y1 u x2 u y2 u x3 u y3 u x4 u y4 ]T . (23.5)

§23.3. Test of Bilinear Quadrilateral

The stiffness module is tested on the two quadrilateral geometries shown in Figure 23.6. Both
elements have unit thickness and isotropic material. The left one is a rectangle of base 2a and
height a. The right one is a right trapezoid with base 2a, top width a and height a.
The two geometries will be used to illustrate the effect of the numerical integration rule.

§23.3.1. Test of Rectangular Geometry

The script listed in Figure 23.7 computes and displays the stiffness of the rectangular element shown
in Figure 23.6(a). This is a rectangle of base 2a and height a. The plate has unit thickness and
isotropic material with E = 96 and ν = 1/3, giving the stress-strain constitutive matrix
 
108 36 0
E= 36 108 0 (23.6)
0 0 36

Using a 2 × 2 Gauss integration rule returns the stiffness matrix

 
42 18 −6 0 −21 −18 −15 0
 18 78 0 30 −18 −39 0 −69 
 
 −6 0 42 −18 −15 0 −21 18 
 
 0 30 −18 78 0 −69 18 −39 
K =
e
 (23.7)
 −21 −18 −15 0 42 18 −6 0
 
 −18 −39 0 −69 18 78 0 30 
 
−15 0 −21 18 −6 0 42 −18
0 −69 18 −39 0 30 −18 78

Note that the rectangle dimension a does not appear in (23.7). This is a general property: the stiffness
matrix of plane stress elements is independent of in-plane dimension scalings. This follows from
the fact that entries of the strain-displacement matrix B have dimensions 1/L, where L denotes a

23–7
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–8

ClearAll[Em,nu,a,b,e,h,p,numer]; h=1;
Em=96; nu=1/3; (* isotropic material *)
Emat=Em/(1-nu^2)*{{1,nu,0},{nu,1,0},{0,0,(1-nu)/2}};
Print["Emat=",Emat//MatrixForm];
ncoor={{0,0},{2*a,0},{2*a,a},{0,a}}; (* 2:1 rectangular geometry *)
p=2;(* 2 x 2 Gauss rule *)numer=False;(* exact symbolic arithmetic *)
Ke=Quad4IsoPMembraneStiffness[ncoor,Emat,h,{numer,p}];
Ke=Simplify[Chop[Ke]]; Print["Ke=",Ke//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[N[Ke]],.0000001]];

Figure 23.7. Driver for stiffness calculation of rectangular element of Figure 23.6(a) for 2 × 2 Gauss rule.

characteristic inplane length. Consequently entries of BT B have dimension 1/L 2 . Integration over
the element area cancels out L 2 .
Using a higher order Gauss integration rule, such as 3 × 3 and 4 × 4, reproduces exactly (23.7).
This is a property characteristic of the rectangular geometry, since in that case the entries of B vary
linearly in ξ and η, and J is constant. Therefore the integrand h BT E BJ is at most quadratic in ξ
and η, and 2 Gauss points in each direction suffice to compute the integral exactly. Using a 1 × 1
rule yields a rank-deficiency matrix, a result illustrated in detail in §23.2.2.
The stiffness matrix (23.7) has the eigenvalues

[ 223.64 90 78 46.3603 42 0 0 0] (23.8)

This verifies that Ke has the correct rank of five (8 total DOFs minus 3 rigid body modes).

§23.3.2. Test of Trapezoidal Geometry

ClearAll[Em,nu,h,a,p]; h=1;
Em=48*63*13*107; nu=1/3;
Emat=Em/(1-nu^2)*{{1,nu,0},{nu,1,0},{0,0,(1-nu)/2}};
ncoor={{0,0},{2*a,0},{a,a},{0,a}};
For [p=1,p<=4,p++,
Ke=Quad4IsoPMembraneStiffness[ncoor,Emat,h,{True,p}];
Ke=Rationalize[Ke,0.0000001]; Print["Ke=",Ke//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[N[Ke]],.0000001]]
];

Figure 23.8. Driver for stiffness calculation of trapezoidal element of Figure 23.6(b)
for four Gauss integration rules.

The trapezoidal element geometry of Figure 23.6(b) is used to illustrate the effect of changing the
p × p Gauss integration rule. Unlike the rectangular case, the element stiffness keeps changing as
p is varied from 1 to 4. The element is rank sufficient, however, for p ≥ 2 in agreement with the
analysis of Chapter 19.
The computations are driven with the script shown in Figure 23.8. The value of p is changed in a loop.
The flag numer is set to True to use floating-point computation for speed (see Remark 23.1). The
computed entries of Ke are transformed to the nearest rational number (exact integers in this case)
using the built-in function Rationalize. The strange value of E = 48×63×13×107 = 4206384,

23–8
23–9 §23.3 TEST OF BILINEAR QUADRILATERAL

in conjunction with ν = 1/3, makes all entries of Ke exact integers when computed with the first
4 Gauss rules. This device facilitates visual comparison between the computed stiffness matrices:
 1840293 1051596 −262899 −262899 −1840293 −1051596 262899 262899 
1051596 3417687 −262899 1314495 −1051596 −3417687 262899 −1314495
 −262899 
 −262899 1051596 −525798 262899 262899 −1051596 525798 
 −262899 1314495 −525798 1051596 262899 −1314495 525798 −1051596 
Ke1×1 =  −1840293  (23.9)
 −1051596 262899 262899 1840293 1051596 −262899 −262899 
 −1051596 −3417687 262899 −1314495 1051596 3417687 −262899 1314495 
 
262899 262899 −1051596 525798 −262899 −262899 1051596 −525798
262899 −1314495 525798 −1051596 −262899 1314495 −525798 1051596
 2062746 1092042 −485352 −303345 −1395387 −970704 −182007 182007 
1092042 3761478 −303345 970704 −970704 −2730105 182007 −2002077
 −485352 
 −303345 1274049 −485352 −182007 182007 −606690 606690 
 −303345 −485352 −2002077 −364014 
=  −1395387 
970704 1395387 182007 606690
Ke2×2 (23.10)
 −970704 −182007 182007 2730105 1213380 −1152711 −424683 
 −970704 −2730105 −2002077 −424683 −60669 
 182007 1213380 4792851 
−182007 182007 −606690 606690 −1152711 −424683 1941408 −364014
182007 −2002077 606690 −364014 −424683 −60669 −364014 2426760
 2067026 1093326 −489632 −304629 −1386827 −968136 −190567 179439 
1093326 3764046 −304629 968136 −968136 −2724969 179439 −2007213
 −489632 
 −304629 1278329 −484068 −190567 179439 −598130 609258 
 −304629 −484068 −2007213 −358878 
=  −1386827 
968136 1397955 179439 609258
Ke3×3 (23.11)
 −968136 −190567 179439 2747225 1218516 −1169831 −429819 
 −968136 −2724969 −2007213 −429819 −70941 
 179439 1218516 4803123 
−190567 179439 −598130 609258 −1169831 −429819 1958528 −358878
179439 −2007213 609258 −358878 −429819 −70941 −358878 2437032
 2067156 1093365 −489762 −304668 −1386567 −968058 −190827 179361 
1093365 3764124 −304668 968058 −968058 −2724813 179361 −2007369
 −489762 
 −304668 1278459 −484029 −190827 179361 −597870 609336 
 −304668 −484029 −2007369 −358722 
=  −1386567 
968058 1398033 179361 609336
Ke4×4 (23.12)
 −968058 −190827 179361 2747745 1218672 −1170351 −429975 
 −968058 −2724813 −2007369 −429975 −71253 
 179361 1218672 4803435 
−190827 179361 −597870 609336 −1170351 −429975 1959048 −358722
179361 −2007369 609336 −358722 −429975 −71253 −358722 2437344

As can be seen entries change substantially in going from p = 1 to p = 2, then more slowly. The
eigenvalues of these matrices are:

Rule Eigenvalues (scaled by 10−6 ) of Ke

1×1 8.77276 3.68059 2.26900 0 0 0 0 0
2×2 8.90944 4.09769 3.18565 2.64521 1.54678 0 0 0
3×3 8.91237 4.11571 3.19925 2.66438 1.56155 0 0 0
4×4 8.91246 4.11627 3.19966 2.66496 1.56199 0 0 0 (23.13)

The stiffness matrix computed by the one-point rule is rank deficient by two. The eigenvalues do
not change appreciably after p = 2. Because the nonzero eigenvalues measure the internal energy
taken up by the element in deformation eigenmodes, it can be seen that raising the order of the
integration makes the element stiffer.

23–9
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–10

Quad4IsoPMembraneBodyForces[ncoor_,rho_,th_,options_,bfor_]:=
Module[{i,k,p=2,numer=False,h=th,
bx,by,bx1,by1,bx2,by2,bx3,by3,bx4,by4,bxc,byc,qcoor,
c,w,Nf,dNx,dNy,Jdet,B,qctab,fe=Table[0,{8}]},
If [Length[options]==2, {numer,p}=options, {numer}=options];
If
[Length[bfor]==2,{bx,by}=bfor;bx1=bx2=bx3=bx4=bx;by1=by2=by3=by4=by];
If [Length[bfor]==4,{{bx1,by1},{bx2,by2},{bx3,by3},{bx4,by4}}=bfor];
If [p<1||p>4, Print["p out of range"]; Return[Null]];
bxc={bx1,bx2,bx3,bx4}; byc={by1,by2,by3,by4};
For [k=1, k<=p*p, k++,
{qcoor,w}= QuadGaussRuleInfo[{p,numer},k];
{Nf,dNx,dNy,Jdet}=Quad4IsoPShapeFunDer[ncoor,qcoor];
bx=Nf.bxc; by=Nf.byc; If [Length[th]==4, h=th.Nf];
c=w*Jdet*h;
bk=Flatten[Table[{Nf[[i]]*bx,Nf[[i]]*by},{i,4}]];
fe+=c*bk;
]; Return[fe]
];

Figure 23.9. Module for computation of consistent node forces from a given
body force field.

Remark 23.1.
The formation of the trapezoidal element stiffness using floating-point computation by setting numer=True
took 0.017, 0.083, 0.15 and 0.25 seconds for p = 1, 2, 3, 4, respectively, on a Mac G4/867. Changing
numer=False to do exact computation increases the formation time to 0.033, 1.7, 4.4 and 44.6 seconds,
respectively. (The unusually large value for p = 4 is due to the time spent in the simplification of the highly
complex exact expressions produced by the Gauss quadrature rule.) This underscores the speed advantage of
using floating-point arithmetic when exact symbolic and algebraic calculations are not required.

§23.4. *Consistent Node Forces for Body Force Field

The module Quad4IsoPMembraneBodyForces listed in Figure 23.9 computes the consistent force associated
with a body force field b  = {bx , b y } given over a 4-node iso-P quadrilateral in plane stress. The field is
specified per unit of volume in componentwise form. For example if the element is subjected to a gravity
acceleration field (self-weight) in the −y direction, bx = 0 and b y = −ρg, where ρ is the mass density.
The arguments of the module are are exactly the same as for Quad4IsoPMembraneStiffness except for the
following differences.
mprop Not used; retained as placeholder.
bfor Body forces per unit volume. Specified as a two-item one-dimensional list: { bx,by }, or as
a four-entry two-dimensional list: { bx1,by1 },{ bx2,by2 }, { bx3,by3 },{ bx4,by4 }.
In the first form the body force field is taken to be constant over the element. The second
form assumes body forces to vary over the element and specified by values at the four
corners, from which the field is interpolated bilinearly.
The module returns fe as an 8 × 1 one dimensional array arranged { fx1,fy1,fx2,fy2,fx3,fy3,
fx4,fy4 } to represent the vector

fe = [ f x1 f y1 f x2 f y2 f x3 f y3 f x4 f y4 ]T . (23.14)

23–10
23–11 §23.5 *RECOVERY OF CORNER STRESSES

Quad4IsoPMembraneStresses[ncoor_,Emat_,th_,options_,udis_]:=
Module[{i,k,numer=False,qcoor,Nf,
dNx,dNy,Jdet,Be,qctab,ue=udis,sige=Table[0,{4},{3}]},
qctab={{-1,-1},{1,-1},{1,1},{-1,1}};
numer=options[[1]];
If [Length[udis]==4, ue=Flatten[udis]];
For [k=1, k<=Length[sige], k++,
qcoor=qctab[[k]]; If [numer, qcoor=N[qcoor]];
{Nf,dNx,dNy,Jdet}=Quad4IsoPShapeFunDer[ncoor,qcoor];
Be={ Flatten[Table[{dNx[[i]], 0},{i,4}]],
Flatten[Table[{0, dNy[[i]]},{i,4}]],
Flatten[Table[{dNy[[i]],dNx[[i]]},{i,4}]]};
sige[[k]]=Emat.(Be.ue);
]; Return[sige]
];

Figure 23.10. Module for calculation of corner stresses.

§23.5. *Recovery of Corner Stresses

Akthough the subject of stress recovery is treated in further detail in a later chapter, for completeness a stress
computation module called Quad4IsoPMembraneStresses for the 4-node quad is listed in Figure 23.10.
The arguments of the module are are exactly the same as for Quad4IsoPMembraneStiffness except for the
following differences.
fprop Not used; retained as placeholder.
udis The 8 corner displacements components. these may be specified as a 8-entry one-
dimensional list form:
{ ux1,uy1, ux2,uy2, ux3,uy3, ux4,uy4 },
or as a 4-entry two-dimensional list:
{ ux1,uy1 },{ ux2,uy2 },{ ux3,uy3 },{ ux4,uy4 }.
The module returns the corner stresses stored in a 4-entry, two-dimensional list:
{ { sigxx1,sigyy1,sigxy1 },{ sigxx2,sigyy2,sigxy2 }, { sigxx3,sigyy3,sigxy3 },
{ sigxx4,sigyy4,sigxy4 } } to represent the stress array
 
σx x1 σx x2 σx x3 σx x4
σe = σ yy1 σ yy2 σ yy3 σ yy4 (23.15)
σx y1 σx y2 σx y3 σx y4

The stresses are directly evaluated at the corner points without invoking any smoothing procedure. A more
elaborated recovery scheme is presented in a later Chapter.

23–11
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–12

§23.6. *Quadrilateral Coordinates of Given Point

The following inverse problem arises in some ap-
plications. Given a 4-node quadrilateral, defined
by the Cartesian coordinates {xi , yi } of its cor- η=1
ners, and an arbitrary point P(x P , y P ), find the
quadrilateral coordinates ξ P , η P of P. In answer- 4
3
ing this question it is understood that the quadri- ξ=−1 ξ=1
lateral coordinates can be extended outside the
element, as illustrated in Figure 23.11.
P(ξ?,η?)
The governing equations are x P = x1 N1 +x2 N2 +
2
x3 N3 + x4 N4 and y P = y1 N1 + y2 N2 + y3 N3 +
1 η=−1
y4 N4 , where N1 = 14 (1−ξ P )(1−η P ), etc. These
bilinear equations are to be solved for {ξ P , η P }.
Figure 23.11. Quadrilateral coordinates can be
Elimination of say, ξ P , leads to a quadratic equa-
extended outside the element to answer the problem
tion in η P : aη2P + bη P + c = 0. It can be shown posed in §23.6. The six yellow-filled circles identify
that b2 ≥ 4ac so there are two real roots: η1 and the four corners plus the intersections of the opposite
η2 . These can be back substituted to get ξ1 and sides. This six-point set defines the so-called complete
ξ2 . Of the two solutions: {ξ1 , η1 } and {ξ2 , η2 } the quadrilateral, which is important in projective geometry.
one closest to ξ = 0, η = 0 is to be taken. The evolute of the coordinate lines is a parabola.

Although seemingly straightforward, the process is prone to numerical instabilities. For example, if the
quadrilateral becomes a rectangle or parallelogram, the quadratic equations degenerate to linear, and one of
the roots takes off to ∞. In floating point arithmetic severe cancellation can occur in the other root. A robust
numerical algorithm, which works stably for any geometry, is obtained by eliminating ξ and η in√ turn, getting
the minimum-modulus root of aη2P + bη P + c = 0 with the stable formula.3 ηmin P = b/(b + b2 − 4ac),
forming the other quadratic equation, and computing its minimum-modulus root the same way. In addition, x P
and y P are referred to the quadrilateral center as coordinate origin. The resulting algorithm can be presented
as follows. Given {x1 , y1 , . . . x4 , y4 } and {x P , y P }, compute

xb = x1 − x2 + x3 − x4 , yb = y1 − y2 + y3 − y4 , xcx = x1 + x2 − x3 − x4 , ycx = y1 + y2 − y3 − y4 ,
xce = x1 − x2 − x3 + x4 , yce = y1 − y2 − y3 + y4 , A= 2 ((x 3
1
− x1 )(y4 − y2 ) − (x4 − x2 )(y3 − y1 )),
J1 = (x3 − x4 )(y1 − y2 ) − (x1 − x2 )(y3 − y4 ), J2 = (x2 − x3 )(y1 − y4 ) − (x1 − x4 )(y2 − y3 ),
x0 = 4 (x 1
1
+ x2 + x3 + x4 ), y0 = 4 (y1
1
+ y2 + y3 + y4 ), x P0 = x P − x0 , y P0 = y P − y0 ,
bξ = A − x P0 yb + y P0 xb , bη = −A − x P0 yb + y P0 xb , cξ = x P0 ycx − y P0 xcx ,
2cξ 2cη
cη = x P0 yce − y P0 xce , ξP =  , ηP = .
− bξ2 − 2J1 cξ − bξ bη + 2J2 cη − bη
2

(23.16)
One common application is to find whether P is inside the quadrilateral: if both ξ P and η P are in the range
[−1, 1] the point is inside, else outside. This occurs, for example, in relating experimental data from given
sensor locations4 to an existing FEM mesh.
A Mathematica module that implements (23.16) is listed in Figure 23.12.

3 See Section 5.6 of [320].

4 While at Boeing in 1969 the writer had to solve a collocation problem of this nature, although in three dimensions.
Pressure data measured at a wind tunnel had to be transported to an independently constructed FEM quadrilateral mesh
modeling the wing skin.

23–12
23–13 §23. References

QuadCoordinatesOfPoint[{{x1_,y1_},{x2_,y2_},{x3_,y3_},
{x4_,y4_}},{x_,y_}]:= Module[{A,J0,J1,J2,
xb=x1-x2+x3-x4,yb=y1-y2+y3-y4,xcξ =x1+x2-x3-x4,ycξ =y1+y2-y3-y4,
xcη=x1-x2-x3+x4,ycη=y1-y2-y3+y4,bξ ,bη,cξ ,cη,
x0=(x1+x2+x3+x4)/4,y0=(y1+y2+y3+y4)/4,dx,dy,ξ ,η},
J0=(x3-x1)*(y4-y2)-(x4-x2)*(y3-y1); A=J0/2;
J1=(x3-x4)*(y1-y2)-(x1-x2)*(y3-y4);
J2=(x2-x3)*(y1-y4)-(x1-x4)*(y2-y3);
dx=x-x0; dy=y-y0;
bξ =A-dx*yb+dy*xb; bη=-A-dx*yb+dy*xb;
cξ = dx*ycξ -dy*xcξ ; cη=dx*ycη-dy*xcη;
ξ =2*cξ /(-Sqrt[bξ ^2-2*J1*cξ ]-bξ );
η=2*cη/( Sqrt[bη^2+2*J2*cη]-bη);
Return[{ξ ,η}]];

Figure 23.12. A Mathematica module that implements the algorithm (23.16).

Notes and Bibliography

For an outline of the history of the 4-node quadrilateral, see Notes and Bibliography in Chapter 17. The
element is called the Taig quadrilateral in the early FEM literature, recognizing his developer [371]. This
paper actually uses the exactly integrated stiffness matrix.
Gauss numerical integration for quadrilaterals was advocated by Irons [214,218], who changed the range of
the quadrilateral coordinates to [−1, +1] to fit tabulated Gauss rules.

References
Referenced items moved to Appendix R.

23–13
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–14

Homework Exercises for Chapter 23

Implementation of Iso-P Quadrilateral Elements

EXERCISE 23.1 [C:15] Figures E23.1–2 show the Mathematica implementation of the stiffness modules for
the 5-node, “bilinear+bubble” iso-P quadrilateral of Figure E18.3. Module Quad5IsoPMembraneStiffness
returns the 10 × 10 stiffness matrix whereas module Quad5IsoPShapeFunDer returns shape function values
and Cartesian derivatives. (The Gauss quadrature module is reused.) Both modules follow the style of the
4-node quadrilateral implementation listed in Figures 23.4–5. The only differences in argument lists is that
ncoor has five node coordinates: { { x1,y1 },{ x2,y2 },{ x3,y3 },{ x4,y4 },{ x5,y5 } }, and that a variable
plate thickness in fprop (one of the two possible formats) is specified as { h1,h2,h3,h4,h5 }.

Quad5IsoPMembraneStiffness[ncoor_,Emat_,th_,options_]:=
Module[{i,j,k,p=2,numer=False,h=th,qcoor,c,w,Nf,
dNx,dNy,Jdet,B,Ke=Table[0,{10},{10}]},
If [Length[options]==2, {numer,p}=options, {numer}=options];
If [p<1||p>4, Print["p out of range"]; Return[Null]];
For [k=1, k<=p*p, k++,
{qcoor,w}= QuadGaussRuleInfo[{p,numer},k];
{Nf,dNx,dNy,Jdet}=Quad5IsoPShapeFunDer[ncoor,qcoor];
If [Length[th]>0, h=th.Nf]; c=w*Jdet*h;
B={ Flatten[Table[{dNx[[i]], 0},{i,5}]],
Flatten[Table[{0, dNy[[i]]},{i,5}]],
Flatten[Table[{dNy[[i]],dNx[[i]]},{i,5}]]};
Ke+=Simplify[c*Transpose[B].(Emat.B)];
]; Return[Ke];
];

Figure E23.1. Stiffness module for the 5-node “bilinear+bubble” iso-P quadrilateral.

Quad5IsoPShapeFunDer[ncoor_,qcoor_]:= Module[
{Nf,dNx,dNy,dNξ ,dNη,Nb,dNbξ ,dNbη,J11,J12,J21,J22,Jdet,ξ ,η,x,y},
{ξ ,η}=qcoor; Nb=(1-ξ ^2)*(1-η^2); (* Nb: node-5 "bubble" function *)
dNbξ =2*ξ (η^2-1); dNbη=2*η*(ξ ^2-1);
Nf= { ((1-ξ )*(1-η)-Nb)/4,((1+ξ )*(1-η)-Nb)/4,
((1+ξ )*(1+η)-Nb)/4,((1-ξ )*(1+η)-Nb)/4, Nb};
dNξ ={-(1-η+dNbξ )/4, (1-η-dNbξ )/4,
(1+η-dNbξ )/4,-(1+η+dNbξ )/4, dNbξ };
dNη={-(1-ξ +dNbη)/4,-(1+ξ +dNbη)/4,
(1+ξ -dNbη)/4, (1-ξ -dNbη)/4, dNbη};
x=Table[ncoor[[i,1]],{i,5}]; y=Table[ncoor[[i,2]],{i,5}];
J11=dNξ .x; J12=dNξ .y; J21=dNη.x; J22=dNη.y;
Jdet=Simplify[J11*J22-J12*J21];
dNx= ( J22*dNξ -J12*dNη)/Jdet; dNx=Simplify[dNx];
dNy= (-J21*dNξ +J11*dNη)/Jdet; dNy=Simplify[dNy];
Return[{Nf,dNx,dNy,Jdet}]
];

Figure E23.2. The shape function module for the 5-node “bilinear+bubble” iso-P quadrilateral.

Test Quad5IsoPMembraneStiffness for the 2:1 rectangular element studied in §23.3.1, with node 5 placed
at the element center. Use Gauss rules 1 × 1, 2 × 2 and 3 × 3. Take E = 96×30 = 2880 in lieu of E = 96 to

23–14
23–15 Exercises

Quad5IsoPMembraneCondStiffness[Ke5_]:=
Module[{i,j,k,n,c,Ke=Ke5,Kc=Table[0,{8},{8}]},
For [n=10,n>=9,n--,
For [i=1,i<=n-1,i++, c=Ke[[i,n]]/Ke[[n,n]];
For [j=1,j<=i,j++, Ke[[j,i]]=Ke[[i,j]]=Ke[[i,j]]-c*Ke[[n,j]];
]]];
For [i=1,i<=8,i++, For [j=1,j<=8,j++, Kc[[i,j]]=Ke[[i,j]]]];
Return[Kc]
];

Figure E23.3. A mystery module for Exercise 23.2.

get exact integer entries in Ke for all Gauss rules while keeping ν = 1/3 and h = 1. Report on which rules
give rank sufficiency. Partial result: K 22 = 3380 and 3588 for the 2 × 2 and 3 × 3 rules, respectively.

EXERCISE 23.2 [D:10] Module Quad5IsoPMembraneCondStiffness in Figure E23.3 is designed to re-

ceive, as only argument, the 10 × 10 stiffness Ke computed by Quad5IsoPMembraneStiffness, and returns
a smaller (8 × 8) stiffness matrix. State what the function of the module is but do not describe programming
details.

EXERCISE 23.3 [C:20] Repeat Exercise 17.3 for the problem illustrated in Figure E17.4, but with
the 5-node “bilinear+bubble” iso-P quadrilateral as the 2D element that models the plane beam. Skip
item (a). Use the modules of Figures E23.1–3 to do the following. Form the 10 × 10 stiffness ma-
trix Ke5 using Quad5IsoPMembraneStiffness with p = 2 and numer=False. Insert this Ke5 into
Quad5IsoPMembraneCondStiffness, which returns a 8 × 8 stiffness Ke. Stick this Ke into equations
(E17.6) and (E17.7) to get Uquad . Show that the energy ratio is
 
Uquad γ 2 (1 + ν) 2 + γ 2 (1 − ν)
r= = . (E23.1)
Ubeam (1 + γ 2 )2

Compare this to the energy ratio (E17.8) for γ = 1/10 and ν = 0 to conclude that shear locking has not been
eliminated, or even mitigated, by the injection of the bubble shape functions associated with the interior node.5

EXERCISE 23.4 [C:25] Implement the 9-node biquadratic quadrilateral element for plane stress to get its
18 × 18 stiffness matrix. Follow the style of Figures 23.3–4 or E23.1–2. (The Gauss quadrature module may
be reused without change.) Test it for the 2:1 rectangular element studied in §23.3.1, with nodes 5–8 placed at
the side midpoints, and node 9 at the element center. For the elastic modulus take E = 96×39×11×55×7 =
15855840 instead of E = 96, along with ν = 1/3 and h = 1, so as to get exact integer entries in Ke . Use both
2 × 2 and 3 × 3 Gauss integration rules and show that the 2 × 2 rule produces a rank deficiency of 3 in the
stiffness. (If the computation with num=False takes too long on a slow PC, set num=True and Rationalize
entries as in Figure 23.8.) Partial result: K 11 = 5395390 and 6474468 for the 2×2 and 3×3 rules, respectively.

EXERCISE 23.5 [C:25] An element is said to be distortion insensitive when the discrete solution does not
appreciably change when the mesh is geometrically distorted while keeping the same number of elements,
nodes and degrees of freedom. It is distortion sensitive otherwise. A distortion sensitivity test often found in
the FEM literature for plane stress quadrilateral elements is pictured in Figure E23.4.

5 Even the addition of an infinite number of bubble functions to the 4-node iso-P quadrilateral will not cure shear locking.
This “bubble futility” has been known since the late 1960s. But memories are short. Bubbles have been recently revived
by some FEM authors for other application contexts, such as multiscale modeling.

23–15
Chapter 23: IMPLEMENTATION OF ISO-P QUADRILATERAL ELEMENTS 23–16

;;
y
(a)
A M z
b x B b
Cross
section
a
h

(b)
;;; 1 3 e 5 −M/b

;;; 2 e 4 6
M/b

;;;
(c) 1 4 7 e 10 13 −M/b

;;
2 5 8 11 14

3 6 e 9 12 15
M/b

Figure E23.4. Two-element cantilever model for assessing distortion sensitivity, for
Exercise E23.5.

The cantilever beam of span a, height b and narrow rectangular cross section of thickness h depicted in Figure
E23.4 is under an applied end moment M. If the material is isotropic with elastic modulus E and Poisson ratio
ν = 0 the end vertical deflection given by the Bernoulli-Euler beam theory is vbeam = Ma 2 /(2E Izz ), where
Izz = hb3 /12. This is also the exact solution given by elasticity theory if the surface tractions over the free
end section match the beam stress distribution.6
The problem is discretized with two 4-node isoP bilinear quadrilaterals as illustrated in Figure E23.4(b), which
show appropiate displacement and force boundary conditions. The mesh distortion is parametrized by the
distance e, which defines the slope of interface 3-4. If e = 0 there is no distortion.
Obtain the finite element solution for a = 10, b = 2, h = E = M = 1 and e = 0, 1, 2, 3, 5 and record the end
deflection vquad as the average of the vertical displacements u y5 and u y6 . Define the ratio r (e) = vquad /vbeam
and plot g(e) = r (e)/r (0) as function of e. This function g(e) characterizes the distortion sensitivity. Show
that g(e) < 1 as e increases and thus the mesh stiffness further as a result of the distortion. Conclude that the
4-node bilinear element is distortion sensitive.

EXERCISE 23.6 [C:25] Repeat the steps of Exercise 23.5 for the mesh depicted in Figure E23.4(c). The
cantilever beam is modeled with two 9-node biquadratic quadrilaterals integrated by the 3×3 Gauss product
rule. It is important to keep the side nodes at the midpoints of the sides, and the center node at the crossing of
the medians (that is, the 9-node elements are superparametric). Show that r = 1 for any e < 12 a. Thus not
only this element is exact for the regular mesh, but it is also distortion insensitive.

6 This statement would not be true if ν = 0 since the fixed-displacement BC at the cantilever root would preclude the
lateral expansion or contraction of the cross section. However, the support condition shown in the models (b) and (c)
allow such changes so the results are extendible to nonzero ν.

23–16
23–17 Exercises

EXERCISE 23.7 [C:25] Implement the 8-node Serendipity quadrilateral element for plane stress to get its
16 × 16 stiffness matrix. Call the module Quad8IsoPMembraneStiffness; the subordinate shape-function
and Gauss-quadrature-info modules should be called Quad8IsoPShapeFunDer and QuadGaussRuleInfo,
respectively. Follow the style of Figures 23.3–4 for argument sequence and coding schematics. (The Gauss
quadrature module may be reused without change.) Test it for the 2:1 rectangular element studied in §23.3.1,
with nodes 5, 6, 7, and 8 placed at midpoints of sides 1–2, 2–3, 3–4, and 4–1, respectively. For elastic modulus
take E = 96×39×11×55×7 = 15855840 along with ν = 1/3 and h = 1. The elastic modulus matrix E that
is passed as second argument should be
 
17837820 5945940 0
E= 5945940 17837820 0 . (E23.2)
0 0 5945940

Use both 2 × 2 and 3 × 3 Gauss integration rules and show that the 2 × 2 rule produces a rank deficiency
of 1 in the stiffness. (If the computation with num=False takes too long on a slow PC, set num=True and
Rationalize entries as in Figure 23.8.) Partial results: K 11 = 11561550 and 12024012 for the 2 × 2 and
3 × 3 rules, respectively, whereas K 13 = 4954950 and 5021016 for the 2 × 2 and 3 × 3 rules, respectively.

23–17
 24
.

Implementation of
Iso-P Triangular
Elements

24–1
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–2

TABLE OF CONTENTS

Page
§24.1. Introduction 24–3
§24.2. Gauss Quadrature for Triangles 24–3
§24.2.1. Requirements for Gauss Rules . . . . . . . . . . . . . 24–3
§24.2.2. Superparametric Triangles . . . . . . . . . . . . . 24–4
§24.2.3. Arbitrary Iso-P Triangles . . . . . . . . . . . . . . 24–5
§24.2.4. Implementation . . . . . . . . . . . . . . . . . 24–6
§24.3. Partial Derivative Computation 24–6
§24.3.1. Triangle Coordinate Partials . . . . . . . . . . . . . 24–7
§24.3.2. Solving the Jacobian System . . . . . . . . . . . . . 24–8
§24.4. The Quadratic Triangle 24–9
§24.4.1. Shape Function Module . . . . . . . . . . . . . . . 24–9
§24.4.2. Stiffness Module . . . . . . . . . . . . . . . . . 24–11
§24.4.3. Test on Straight-Sided Triangle . . . . . . . . . . . . 24–12
§24.4.4. Test on Highly Distorted Triangle . . . . . . . . . . . 24–12
§24.5. *The Cubic Triangle 24–14
§24.5.1. *Shape Function Module . . . . . . . . . . . . . . 24–14
§24.5.2. *Stiffness Module . . . . . . . . . . . . . . . . 24–16
§24.5.3. *Test Element . . . . . . . . . . . . . . . . . . 24–17
§24. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . . 24–18
§24. References . . . . . . . . . . . . . . . . . . . . . . . 24–19
§24. Exercises . . . . . . . . . . . . . . . . . . . . . . . 24–20
§24. Solutions to Exercises
. . . . . . . . . . . . . . . . . . . . . . . 24–22

24–2
24–3 §24.2 GAUSS QUADRATURE FOR TRIANGLES

§24.1. Introduction

This Chapter continues with the computer implementation of two-dimensional finite elements. It covers
the programming of isoparametric triangular elements for the plane stress problem. Triangular elements
bring two sui generis implementation quirks with respect to quadrilateral elements:
(1) The numerical integration rules for triangles are not product of one-dimensional Gauss rules, as in
the case of quadrilaterals. They are instead specialized to the triangle geometry.
(2) The computation of x-y partial derivatives and the element-of-area scaling by the Jacobian de-
terminant must account for the fact that the triangular coordinates ζ1 , ζ2 and ζ3 do not form an
independent set.
We deal with these issues in the next two sections.

§24.2. Gauss Quadrature for Triangles

The numerical integration schemes for quadrilaterals introduced in §17.3 and implemented in §23.2
are built as “tensor products” of two one-dimensional Gauss formulas. On the other hand, Gauss rules
for triangles are not derivable from one-dimensional rules, and must be constructed especially for the
triangular geometry.

§24.2.1. Requirements for Gauss Rules

Gauss quadrature rules for triangles must possess triangular symmetry in the following sense:

If the sample point (ζ1 , ζ2 , ζ3 ) is present in a Gauss integration rule with

weight w, then all other points obtainable by permuting the three triangular
coordinates arbitrarily must appear in that rule, and have the same weight. (24.1)

This constraint guarantees that the result of the quadrature process will not depend on element node
numbering.1 If ζ1 , ζ2 , and ζ3 are different, (24.1) forces six equal-weight sample points to be present in
the rule, because 3! = 6. If two triangular coordinates are equal, the six points coalesce to three, and
(24.1) forces three equal-weight sample points to be present. Finally, if the three coordinates are equal
(which can only happen for the centroid ζ1 = ζ2 = ζ3 = 1/3), the six points coalesce to one.2
Additional requirements for a Gauss rule to be numerically acceptable are:

All sample points must be inside the triangle (or on the triangle boundary)
and all weights must be positive. (24.2)

This is called a positivity condition. It insures that the element internal energy evaluated by numerical
quadrature is nonnegative definite.

1 It would disconcerting to users, to say the least, to have the FEM solution depend on how nodes are numbered.
2 It follows that the number of sample points in triangle Gauss quadrature rules that satisfy (24.1) must be of the form 6i +3 j +k,
where i and j are nonnegative integers and k is 0 or 1. Consequently there are no rules with 2, 5 or 8 points.

24–3
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–4

(a) rule=1 (b) rule=3 (c) rule=−3

degree=1 degree=2 degree=2
0.33333
0.33333 0.33333
1.0000

0.33333 0.33333
0.33333

(d) rule=6 (e) rule=7

0.10995 0.12594
degree=4 degree=5

0.22338 0.22338 0.13239 0.13239

0.22500

0.10995 0.22338 0.10995 0.12594 0.13239 0.12594

Figure 24.1. Location of sample points (dark circles) of five Gauss quadrature rules for
straight sided (superparametric) 6-node triangles. Weight written to 5 places near each
sample point; sample-point circle areas are proportional to weight.

A rule is said to be of degree n if it integrates exactly all polynomials in the triangular coordinates of
order n or less when the Jacobian determinant is constant, and there is at least one polynomial of order
n + 1 that is not exactly integrated by the rule.

Remark 24.1. The positivity requirement (24.2) is automatically satisfied in quadrilaterals by using Gauss product
rules, since the points are always inside while weights are positive. Consequently it was not necessary to call
attention to it. On the other hand, for triangles there are Gauss rules with as few as 4 points that violate positivity.

§24.2.2. Superparametric Triangles

We first consider superparametric straight-sided triangles geometry defined by the three corner nodes.
Over such triangles the Jacobian determinant defined below is constant. The five simplest Gauss rules
that satisfy the requirements (24.1) and (24.2) have 1, 3, 3, 6 and 7 points, respectively. The two rules
with 3 points differ in the location of the sample points. The five rules are depicted in Figure 24.1 over
6-node straight-sided triangles; for such triangles to be superparametric the side nodes must be located
at the midpoint of the sides.
One point rule. The simplest Gauss rule for a triangle has one sample point located at the centroid. For
a straight sided triangle, 
1
F(ζ1 , ζ2 , ζ3 ) d ≈ F( 13 , 13 , 13 ), (24.3)
A e
where A is the triangle area

  
1 1 1  
A= d = 2 det x1
1
x2 x3 = 1
2
(x2 y3 − x3 y2 ) + (x3 y1 − x1 y3 ) + (x1 y2 − x2 y1 ) . (24.4)
e y1 y2 y3

This rule is illustrated in Figure 24.1(a). It has degree 1, meaning that it integrates exactly up to linear
polynomials in {ζ1 , ζ2 , ζ3 }. For example, F = 4 − ζ1 + 2ζ2 − ζ3 is exactly integrated by (24.3).

24–4
24–5 §24.2 GAUSS QUADRATURE FOR TRIANGLES

(a) rule=1 (b) rule=3 0.33333 (c) rule=−3 0.33333

1.0000
0.33333
0.33333
0.33333
0.33333

0.10995 0.12594
(d) rule=6 (e) rule=7 0.13239
0.22338

0.22338 0.13239 0.22500

0.22338 0.13239 0.12594

0.10995
0.10995 0.12594

Figure 24.2. Location of sample points (dark circles) of five Gauss quadrature rules
for curved sided 6-node triangles. Weight written to 5 places near each sample point;
sample-point circle areas are proportional to weight.

Three Point Rules. The next two rules in order of simplicity contain three sample points:

1
F(ζ1 , ζ2 , ζ3 ) d ≈ 13 F( 23 , 16 , 16 ) + 13 F( 16 , 23 , 16 ) + 13 F( 16 , 16 , 23 ). (24.5)
A e

1
F(ζ1 , ζ2 , ζ3 ) d ≈ 13 F( 12 , 12 , 0) + 13 F(0, 12 , 12 ) + 13 F( 12 , 0, 12 ). (24.6)
A e
These are depicted in Figures 24.1(b,c). Both rules are of degree 2; that is, they integrate exactly up
to quadratic polynomials in the triangular coordinates. For example, the function F = 6 + ζ1 + 3ζ3 +
ζ22 − ζ32 + 3ζ1 ζ3 is integrated exactly by either rule. Formula (24.6) is called the midpoint rule.
Six and Seven Point Rules. There is a 4-point rule of degree 3, but it has a negative weight and so violates
(24.2). There are no symmetric rules with 5 points. The next useful rules have six and seven points.
There is a 6-point rule of degree 4 and a 7-point rule of degree 5, which integrate exactly up to quartic
and quintic polynomials in {ζ1 , ζ2 , ζ3 }, respectively. The 7-point rule includes the centroid as sample
point. The abcissas and weights are expressable as rational combinations of square roots of integers
and fractions. The expressions are listed in the Mathematica implementation discussed in §24.2.4. The
sample point configurations are depicted in Figure 24.1(d,e).

§24.2.3. Arbitrary Iso-P Triangles

If the triangle has variable metric, as in the curved sided 6-node triangle geometries shown in Figure 24.2,
the foregoing formulas need adjustment because the element of area d becomes a function of position.
Consider the more general case of an isoparametric element with n nodes and shape functions Ni . In
§24.3 it is shown that the differential area element is given by
 
1 1 1

n
∂ Ni
n
∂ Ni
n
∂ Ni 
 x x xi 
 i i 
d = J dζ1 dζ2 dζ3 , J = 12 det  i=1 ∂ζ1 i=1 ∂ζ2 i=1 ∂ζ3  (24.7)
 n 

∂ Ni
n
∂ Ni
n
∂ Ni 
yi yi yi
i=1
∂ζ1 i=1 ∂ζ2 i=1 ∂ζ3

24–5
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–6

TrigGaussRuleInfo[{rule_,numer_},point_]:= Module[
{zeta,p=rule,i=point,g1,g2, info={{Null,Null,Null},0} },
If [p== 1, info={{1/3,1/3,1/3},1}];
If [p== 3, info={{1,1,1}/6,1/3}; info[[1,i]]=2/3];
If [p==-3, info={{1,1,1}/2,1/3}; info[[1,i]]=0 ];
If [p== 6, g1=(8-Sqrt[10]+Sqrt[38-44*Sqrt[2/5]])/18;
g2=(8-Sqrt[10]-Sqrt[38-44*Sqrt[2/5]])/18;
If [i<4, info={{g1,g1,g1},(620+Sqrt[213125-
53320*Sqrt[10]])/3720}; info[[1,i]]=1-2*g1];
If [i>3, info={{g2,g2,g2},(620-Sqrt[213125-
53320*Sqrt[10]])/3720}; info[[1,i-3]]=1-2*g2]];
If [p== 7, g1=(6-Sqrt[15])/21; g2=(6+Sqrt[15])/21;
If [i<4, info={{g1,g1,g1},(155-Sqrt[15])/1200};
info[[1,i]]= 1-2*g1];
If [i>3&&i<7, info={{g2,g2,g2},(155+Sqrt[15])/1200};
info[[1,i-3]]=1-2*g2];
If [i==7, info={{1/3,1/3,1/3},9/40} ]];
If [numer, Return[N[info]], Return[Simplify[info]]];
];

Figure 24.3. Module to get triangle Gauss quadrature rule information.

Here J is the Jacobian determinant, which plays the same role as J in the isoparametric quadrilaterals.
If the metric is simply defined by the 3 corners, as in Figure 24.1, the geometry shape functions are
N1 = ζ1 , N2 = ζ2 and N3 = ζ3 . Then the foregoing determinant reduces to that of (24.4), and J = A
everywhere. But for general (curved) geometries J = J (ζ1 , ζ2 , ζ3 ), and the triangle area A cannot be
factored out of the integration rules. For example the one point rule becomes

F(ζ1 , ζ2 , ζ3 ) d ≈ J ( 13 , 13 , 13 ) F( 13 , 13 , 13 ). (24.8)
e

whereas the midpoint rule becomes


F(ζ1 , ζ2 , ζ3 ) d ≈ 13 J ( 12 , 12 , 0) F( 12 , 12 , 0) + 13 J (0, 12 , 12 ) F(0, 12 , 12 ) + 13 J ( 12 , 0, 12 ) F( 12 , 0, 12 ).
e
(24.9)
These can be expressed more compactly by saying that the Gauss integration rule is applied to J F.

§24.2.4. Implementation
The five rules pictured in Figures 24.1 and 24.2 are implemented in the module TrigGaussRuleInfo
listed in Figure 24.3. The module is invoked as

{ { zeta1,zeta2,zeta3 },w } = TrigGaussRuleInfo[{ rule,numer },point] (24.10)

The module has three arguments: rule, numer and i. The first two are grouped in a two-item list.
Argument rule, which can be 1, 3, −3, 6 or 7, defines the integration formula as follows. Abs[rule]
is the number of sample points. Of the two 3-point choices, if rule is −3 the midpoint rule is picked,
else if +3 the 3-interior point rule is chosen. Logical flag numer is set to True or False to request
floating-point or exact information, respectively
Argument point is the index of the sample point, which may range from 1 through Abs[rule].
The module returns the list {{ζ1 , ζ2 , ζ3 }, w}, where ζ1 , ζ2 , ζ3 are the triangular coordinates of the sample
point, and w is the integration weight. For example, TrigGaussRuleInfo[{ 3,False },1] returns
{ { 2/3,1/6,1/6 },1/3 }. If rule is not implemented, the module returns { { Null,Null,Null },0 }.

24–6
24–7 §24.3 PARTIAL DERIVATIVE COMPUTATION

§24.3. Partial Derivative Computation

The calculation of Cartesian partial derivatives is illustrated in this section for the 6-node triangle shown
in Figure 24.4. The results are applicable, however, to iso-P triangles with any number of nodes.
The element geometry is defined by the corner coordinates
3
{xi , yi }, with i = 1, 2, . . . 6. Corners are numbered 1,2,3
in counterclockwise sense. Side nodes are numbered 4,5,6
opposite to corners 3,1,2, respectively. Side nodes may
5
be arbitrarily located within positive Jacobian constraints as 6
discussed in §19.4.2. The triangular coordinates are as usual
denoted by ζ1 , ζ2 and ζ3 , which satisfy ζ1 + ζ2 + ζ3 = 1. 2
The quadratic displacement field {u x (ζ1 , ζ2 , ζ3 ), u y (ζ1 , ζ2 , ζ3 )} 4
is defined by the 12 node displacements {u xi , u yi }, i = 1
1, 2, . . . , 6, as per the iso-P quadratic interpolation formula Figure 24.4. The 6-node iso-P triangle.
(16.10–11). That formula is repeated here for convenience:
 
    N1e
1 1 1 1 1 1 1
 N2e 
   1
x x x2 x3 x4 x5 x6   e
     N3 
 y  =  y1 y2 y3 y4 y5 y6   e (24.11)
     N4 
ux u x1 u x2 u x3 u x4 u x5 u x6  e
N5
uy u y1 u y2 u y3 u y4 u y5 u y6
N6e
in which the shape functions and their natural derivatives are
 e        
N1 ζ1 (2ζ1 −1) 4ζ1 −1 0 0
 N2e   ζ2 (2ζ2 −1)   0   4ζ2 −1   0 
 e   ∂NT   ∂NT   ∂NT  
 N   ζ (2ζ −1)   0   0   4ζ −1 
NT =  3e  =  3 3 , = , = , =  3 .
 N4   4ζ1 ζ2  ∂ζ1  4ζ2  ∂ζ2  4ζ1  ∂ζ3  0 
 e        
N5 4ζ2 ζ3 0 4ζ3 4ζ2
N6e 4ζ3 ζ1 4ζ3 0 4ζ1
(24.12)

§24.3.1. Triangle Coordinate Partials

In this and following sections the superscript e of shape functions will be omitted for brevity. The bulk
of the shape function logic is concerned with the computation of the partial derivatives of the shape
functions (24.12) with respect to x and y at any point in the element. For this purpose consider a generic
scalar function w(ζ1 , ζ2 , ζ3 ) that is quadratically interpolated over the triangle by

w = w1 N1 + w2 N2 + w3 N3 + w4 N4 + w5 N5 + w6 N6 = [ w1 w2 w3 w4 w5 w6 ] NT . (24.13)

Symbol w may stand for 1, x, y, u x or u y , which are interpolated in the iso-P representation (24.11),
or other element-varying quantities such as thickness, temperature, etc. Taking partials of (24.13) with
respect to x and y and applying the chain rule twice yields
∂w
∂ Ni
 ∂ Ni ∂ζ1 ∂ Ni ∂ζ2 ∂ Ni ∂ζ3

= wi = wi + + ,
∂x ∂x ∂ζ1 ∂ x ∂ζ2 ∂ x ∂ζ3 ∂ x
∂w
∂ Ni
 ∂ Ni ∂ζ1 ∂ Ni ∂ζ2 ∂ Ni ∂ζ3
 (24.14)
= wi = wi + + ,
∂y ∂y ∂ζ1 ∂ y ∂ζ2 ∂ y ∂ζ3 ∂ y

24–7
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–8

where all sums are understood to run over i = 1, . . . 6. In matrix form:

  ∂ Ni 
 ∂w    wi ∂ζ
∂ζ1 ∂ζ2 ∂ζ3 
 ∂ Ni 
1
∂x  ∂ x ∂ x ∂ x   
∂w = ∂ζ1 ∂ζ2 ∂ζ3  wi ∂ζ  . (24.15)
 2 
∂y ∂y ∂y ∂y  ∂ Ni
wi ∂ζ
3
Transposing both sides of (24.15) while exchanging sides yields
 ∂ζ1 ∂ζ1 
∂x ∂y
  
   ∂ζ ∂ζ   ∂w  .
wi ∂∂ζNi wi ∂∂ζNi wi ∂∂ζNi  ∂ x2 ∂ y2  = ∂w ∂x ∂y (24.16)
1 2 3  
∂ζ3 ∂ζ3
∂x ∂y
Now make w ≡ 1, x, y and stack the results row-wise:
  ∂ Ni  ∂ Ni  ∂ Ni   ∂ζ ∂ζ1   ∂1 ∂1 
1
∂ζ1 ∂ζ2 ∂ζ3 ∂x ∂y
     ∂x ∂y 
 
 ∂ ∂ ∂   ∂ζ ∂ζ  ∂x ∂x
 xi ∂ζ i
N N
xi ∂ζ i xi ∂ζ i   ∂ x2 ∂ y2  = 
N
 ∂ x ∂y
.
 (24.17)
 1 2 3  
 ∂ Ni  ∂ Ni  ∂ Ni ∂ζ3 ∂ζ3 ∂y ∂y
yi ∂ζ
1
yi ∂ζ
2
yi ∂ζ
3 ∂x ∂y ∂x ∂y

But ∂ x/∂ x = ∂ y/∂ y = 1 and ∂1/∂ x = ∂1/∂ y = ∂ x/∂ y = ∂ y/∂  x = 0 because x and y are
independent coordinates. It is shown in Remark 24.2 below that, if Ni = 1, the entries of the first
row of the coefficient matrix are equal to a constant C. These entries can be scaled to unity because the
first row of the right-hand side is null. Consequently we arrive at a system of linear equations of order
3 with two right-hand sides:
 
  ∂ζ1 ∂ζ1

1 1 1
 ∂ Ni  ∂ Ni  ∂ x ∂y   0 0 
 x i ∂ Ni xi ∂ζ xi ∂ζ   ∂ζ2 ∂ζ2 
 ∂ζ1 2 3   ∂x ∂y  = 1 0 . (24.18)
 ∂ Ni  ∂ Ni  ∂ Ni  
yi ∂ζ yi ∂ζ yi ∂ζ ∂ζ3 ∂ζ3 0 1
1 2 3 ∂x ∂y

§24.3.2. Solving the Jacobian System

By analogy with quadrilateral elements, the coefficient matrix of (24.18) will be called the Jacobian
matrix and denoted by J. Its determinant scaled by one half is equal to the Jacobian J = 12 det J used
in the expression of the area element introduced in §24.2.3. For compactness (24.18) is rewritten
 
  ∂ζ1 ∂ζ1
1 1 1  ∂x ∂y   0 0 
 ∂ζ ∂ζ 
J P =  Jx1 Jx2 Jx3   ∂ x2 ∂ y2  = 1 0 . (24.19)
 
Jy1 Jy2 Jy3 ∂ζ3 ∂ζ3 0 1
∂x ∂y
If J = 0, solving this system gives
 ∂ζ ∂ζ1 
1
 
 ∂x ∂y 
1 Jy23 Jx32
 ∂ζ2 ∂ζ2 
 ∂x ∂y  = Jy31 Jx13 = P, (24.20)
  2J
∂ζ3 ∂ζ3 J y12 Jx21
∂x ∂y
24–8
24–9 §24.4 THE QUADRATIC TRIANGLE

in which Jx ji = Jx j − Jxi , Jy ji = Jy j − Jyi and J = 12 det J = 12 (Jx21 Jy31 − Jy12 Jx13 ). Substituting
into (24.14) we arrive at

∂w
∂ Ni
wi  ∂ Ni ∂ Ni ∂ Ni

= wi = Jy23 + Jy31 + Jy12 ,
∂x ∂x 2J ∂ζ1 ∂ζ2 ∂ζ3
∂w
∂ Ni
wi  ∂ Ni ∂ Ni ∂ Ni
 (24.21)
= wi = Jx32 + Jx13 + Jx21 .
∂y ∂y 2J ∂ζ1 ∂ζ2 ∂ζ3

In particular, the shape function derivatives are

 
∂ Ni 1 ∂ Ni ∂ Ni ∂ Ni
= Jy23 + Jy31 + Jy12 ,
∂x 2J ∂ζ1 ∂ζ2 ∂ζ3
  (24.22)
∂ Ni 1 ∂ Ni ∂ Ni ∂ Ni
= Jx32 + Jx13 + Jx21 .
∂y 2J ∂ζ1 ∂ζ2 ∂ζ3

in which the natural derivatives ∂ Ni /∂ζ j can be read off (24.12). Using the 3 × 2 P matrix defined in
(24.20) yields finally the compact form
   
∂ Ni ∂ Ni ∂ Ni ∂ Ni ∂ Ni P.
∂x ∂y = ∂ζ1 ∂ζ2 ∂ζ3 (24.23)

Remark 24.2. Here is the proof that each first row entry of the 3 × 3 matrix in (24.17) is a numerical constant, say
C. Suppose the shape functions are polynomials of order n in the triangular coordinates, and let Z = ζ1 + ζ2 + ζ3 .
The completeness identity is

S= Ni = 1 = c1 Z + c2 Z 2 + . . . cn Z n , c1 + c2 + . . . + cn = 1. (24.24)

where the ci are element dependent scalar coefficients. Differentiating S with respect to the ζi ’s and setting Z = 1
yields

∂ Ni
∂ Ni
∂ Ni
C= = = = c1 + 2c2 Z + c3 Z 3 + . . . (n − 1)Z n−1 = c1 + 2c2 + 3c3 + . . . + cn−1
∂ζ1 ∂ζ2 ∂ζ3
= 1 + c2 + 2c3 + . . . + (n − 2)cn ,
(24.25)
which proves the assertion. For the 3-node linear triangle, S = Z and C = 1. For the 6-node quadratic triangle,
S = 2Z 2 − Z and C = 3. For the 10-node cubic triangle, S = 9Z 3 /2 − 9Z 2 /2 + Z and C = 11/2. Because the
first equation in (24.18) is homogeneous, the C’s can be scaled to unity.

§24.4. The Quadratic Triangle

§24.4.1. Shape Function Module

We specialize now the results of §24.3 to obtain the stiffness matrix of the 6-node quadratic triangle in
plane stress. Taking the dot products of the natural-coordinate partials (24.12) with the node coordinates
we obtain the entries of the Jacobian matrix of (24.19):

Jx1 = x1 (4ζ1 −1) + 4(x4 ζ2 + x6 ζ3 ), Jx2 = x2 (4ζ2 −1) + 4(x5 ζ3 + x4 ζ1 ), Jx3 = x3 (4ζ3 −1) + 4(x6 ζ1 + x5 ζ2 ),
Jy1 = y1 (4ζ1 −1) + 4(y4 ζ2 + y6 ζ3 ), Jy2 = y2 (4ζ2 −1) + 4(y5 ζ3 + y4 ζ1 ), Jy3 = y3 (4ζ3 −1) + 4(y6 ζ1 + y5 ζ2 ),
(24.26)

24–9
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–10

Trig6IsoPShapeFunDer[ncoor_,tcoor_]:= Module[
{ζ 1,ζ 2,ζ 3,x1,x2,x3,x4,x5,x6,y1,y2,y3,y4,y5,y6,
dx4,dx5,dx6,dy4,dy5,dy6,Jx21,Jx32,Jx13,Jy12,Jy23,Jy31,
Nf,dNx,dNy,Jdet}, {ζ 1,ζ 2,ζ 3}=tcoor;
{{x1,y1},{x2,y2},{x3,y3},{x4,y4},{x5,y5},{x6,y6}}=ncoor;
dx4=x4-(x1+x2)/2; dx5=x5-(x2+x3)/2; dx6=x6-(x3+x1)/2;
dy4=y4-(y1+y2)/2; dy5=y5-(y2+y3)/2; dy6=y6-(y3+y1)/2;
Nf={ζ 1*(2*ζ 1-1),ζ 2*(2*ζ 2-1),ζ 3*(2*ζ 3-1),4*ζ 1*ζ 2,4*ζ 2*ζ 3,4*ζ 3*ζ 1};
Jx21= x2-x1+4*(dx4*(ζ 1-ζ 2)+(dx5-dx6)*ζ 3);
Jx32= x3-x2+4*(dx5*(ζ 2-ζ 3)+(dx6-dx4)*ζ 1);
Jx13= x1-x3+4*(dx6*(ζ 3-ζ 1)+(dx4-dx5)*ζ 2);
Jy12= y1-y2+4*(dy4*(ζ 2-ζ 1)+(dy6-dy5)*ζ 3);
Jy23= y2-y3+4*(dy5*(ζ 3-ζ 2)+(dy4-dy6)*ζ 1);
Jy31= y3-y1+4*(dy6*(ζ 1-ζ 3)+(dy5-dy4)*ζ 2);
Jdet = Jx21*Jy31-Jy12*Jx13;
dNx= {(4*ζ 1-1)*Jy23,(4*ζ 2-1)*Jy31,(4*ζ 3-1)*Jy12,4*(ζ 2*Jy23+ζ 1*Jy31),
4*(ζ 3*Jy31+ζ 2*Jy12),4*(ζ 1*Jy12+ζ 3*Jy23)}/Jdet;
dNy= {(4*ζ 1-1)*Jx32,(4*ζ 2-1)*Jx13,(4*ζ 3-1)*Jx21,4*(ζ 2*Jx32+ζ 1*Jx13),
4*(ζ 3*Jx13+ζ 2*Jx21),4*(ζ 1*Jx21+ζ 3*Jx32)}/Jdet;
Return[Simplify[{Nf,dNx,dNy,Jdet}]]
];

Figure 24.5. Shape function module for 6-node quadratic triangle.

From these J can be constructed, and shape function partials ∂ Ni /∂ x and ∂ Ni /∂ y explicitly obtained
from (24.22). Somewhat simpler expressions, however, result by using the following “hierarchical”
side node coordinates:

x4 = x4 − 12 (x1 + x2 ), x5 = x5 − 12 (x2 + x3 ), x6 = x6 − 12 (x3 + x1 ),

(24.27)
y4 = y4 − 12 (y1 + y2 ), y5 = y5 − 12 (y2 + y3 ), y6 = y6 − 12 (y3 + y1 ).

Geometrically these represent the deviations from midpoint positions; thus for a superparametric element
x4 = x5 = x6 = y4 = y5 = y6 = 0. The Jacobian coefficients become

Jx21 = x21 + 4(x4 (ζ1 −ζ2 ) + (x5 −x6 )ζ3 ), Jx32 = x32 + 4(x5 (ζ2 −ζ3 ) + (x6 −x4 )ζ1 ),
Jx13 = x13 + 4(x6 (ζ3 −ζ1 ) + (x4 −x5 )ζ2 ), Jy12 = y12 + 4(y4 (ζ2 −ζ1 ) + (y6 −y5 )ζ3 ),
Jy23 = y23 + 4(y5 (ζ3 −ζ2 ) + (y4 −y6 )ζ1 ), Jy31 = y31 + 4(y6 (ζ1 −ζ3 ) + (y5 −y4 )ζ2 ).
(24.28)
(Note that if all midpoint deviations vanish, Jx ji = x ji and Jy ji = y ji .) From this one gets J =
1
2
det J = 12 (Jx21 Jy31 − Jy12 Jx13 ) and
 
1 y23 + 4(y5 (ζ3 −ζ2 ) + (y4 −y6 )ζ1 ) x32 + 4(x5 (ζ2 −ζ3 ) + (x6 −x4 )ζ1 )
P= y31 + 4(y6 (ζ1 −ζ3 ) + (y5 −y4 )ζ2 ) x13 + 4(x6 (ζ3 −ζ1 ) + (x4 −x5 )ζ2 ) .
2J y12 + 4(y4 (ζ2 −ζ1 ) + (y6 −y5 )ζ3 ) x21 + 4(x4 (ζ1 −ζ2 ) + (x5 −x6 )ζ3 )
(24.29)
and the Cartesian derivatives of the shape functions are
   
(4ζ1 − 1)Jy23 (4ζ1 − 1)Jx32
 (4ζ2 − 1)Jy31   (4ζ2 − 1)Jx13 
∂NT 1 
 (4ζ3 − 1)Jy12 
 ∂NT 1 
 (4ζ3 − 1)Jx21 

=  , =  . (24.30)
∂x 2J  4(ζ2 Jy23 + ζ1 Jy31 )  ∂y 2J  4(ζ2 Jx32 + ζ1 Jx13 ) 
   
4(ζ3 Jy31 + ζ2 Jy12 ) 4(ζ3 Jx13 + ζ2 Jx21 )
4(ζ1 Jy12 + ζ3 Jy23 ) 4(ζ1 Jx21 + ζ3 Jx32 )

24–10
24–11 §24.4 THE QUADRATIC TRIANGLE

Trig6IsoPMembraneStiffness[ncoor_,Emat_,th_,options_]:=
Module[{i,k,p=3,numer=False,h=th,tcoor,w,c,
Nf,dNx,dNy,Jdet,Be,Ke=Table[0,{12},{12}]},
If [Length[options]>=1, numer=options[[1]]];
If [Length[options]>=2, p= options[[2]]];
If [!MemberQ[{1,-3,3,6,7},p], Print["Illegal p"]; Return[Null]];
For [k=1, k<=Abs[p], k++,
{tcoor,w}= TrigGaussRuleInfo[{p,numer},k];
{Nf,dNx,dNy,Jdet}= Trig6IsoPShapeFunDer[ncoor,tcoor];
If [numer, {Nf,dNx,dNy,Jdet}=N[{Nf,dNx,dNy,Jdet}]];
If [Length[th]==6, h=th.Nf]; c=w*Jdet*h/2;
Be= {Flatten[Table[{dNx[[i]],0 },{i,6}]],
Flatten[Table[{0, dNy[[i]]},{i,6}]],
Flatten[Table[{dNy[[i]],dNx[[i]]},{i,6}]]};
Ke+=c*Transpose[Be].(Emat.Be);
]; If[!numer,Ke=Simplify[Ke]]; Return[Ke]
];

Figure 24.6. Stiffness matrix module for 6-node plane stress triangle.

A Mathematica shape function module Trig6IsoPShapeFunDer is shown in Figure 24.5. It receives

two arguments: ncoor and tcoor. The first one is the list of {xi , yi } coordinates of the six nodes. The
second is the list of three triangular coordinates { ζ1 , ζ2 , ζ3 } of the location at which the shape functions
and their Cartesian derivatives are to be computed.
The module returns { Nf,dNx,dNy,Jdet } as module value. Here Nf collects the shape function values,
dNx and dNy the x and y shape function derivatives, respectively, and Jdet is the determinant of matrix
J, equal to 2J in the notation used here.

§24.4.2. Stiffness Module

The numerically integrated stiffness matrix is


p
K =e
hBT EB d ≈ wi F(ζ1i , ζ2i , ζ3i ), where F(ζ1 , ζ2 , ζ3 ) = hBT EB J. (24.31)
e i=1

Here p denotes the number of sample points of the Gauss rule being used, wi is the integration weight
for the i th sample point, ζ1i , ζ2i , ζ3i are the sample point triangular coordinates and J = 12 det J. This
data is provided by TrigGaussRuleInfo. For the 6-node triangle (24.31) is implemented in module
Trig6IsoPMembraneStiffness, listed in Figure 24.6. The module is invoked as

Ke = Trig6IsoPMembraneStiffness[ncoor,Emat,th,options] (24.32)

The arguments are:

ncoor The list of node coordinates arranged as { { x1,y1 },{ x2,y2 }, ... { x6,y6 } }.
Emat The plane stress elasticity matrix
 
E 11 E 12 E 13
E= E 12 E 22 E 23 (24.33)
E 13 E 23 E 33
arranged as { { E11,E12,E33 },{ E12,E22,E23 },{ E13,E23,E33 } }.
th Plate thickness specified as either a scalar h or a six-entry list: { h1,h2,h3,h4,h5,h6 }.

24–11
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–12

The one-entry form specifies uniform thickness h. The six-entry form is used to specify
an element of variable thickness, in which case the entries are the six node thicknesses
and h is interpolated quadratically.
options Processing options list. May contain two items: { numer,rule } or one: { numer }.
numer is a logical flag. If True, the computations are forced to proceed in floating-point
arithmetic. For symbolic or exact arithmetic work set numer to False.
rule specifies the triangle Gauss rule as described in §24.2.4; rule may be 1, 3, −3, 6
or 7. For the 6-node element the three point rules are sufficient to get the correct rank.
If omitted rule = 3 is assumed.
The module returns Ke as an 12 × 12 symmetric matrix pertaining to the following arrangement of
nodal displacements:
ue = [ u x1 u y1 u x2 u y2 u x3 u y3 u x4 u y4 u x5 u y5 u x6 u y6 ]T . (24.34)

§24.4.3. Test on Straight-Sided Triangle

The stiffness module of Figure 24.6 is tested on two 3 (4,4)

triangle geometries, one with straight sides and constant
metric and one with curved sides and highly variable 5
metric. The straight sided triangle geometry, shown in
6
Figure 24.7, has the corner nodes placed at (0, 0), (6, 2) 2 (6,2)
and (4, 4) with side nodes 4,5,6 at the midpoints of the
sides. The element has unit plate thickness. The material 4
is isotropic with E = 288 and ν = 1/3.
1 (0,0)
The element is tested with the script shown in Figure
Figure 24.7. Straight sided 6-node triangle
24.8. The stiffness matrix Ke is computed by the three- test element.
interior-point Gauss rule, specified as p=3.
The computed Ke for integration rules with 3 or more points are identical since those rules are exact
for this element if the metric is constant, which is the case here. That stiffness is
 
54 27 18 0 0 9 −72 0 0 0 0 −36
 27 54 0 −18 9 36 0 72 0 0 −36 −144 
 18 216 −108 54 −36 −72 −216 0
 0 0 144 0 
 0 −18 −108 216 −36 90 0 72 144 −360 0 0
 
 0 9 54 −36 162 −81 0 0 −216 144 0 −36 
 
 9 36 −36 90 −81 378 0 0 144 −360 −36 −144 
 −72 0 −72 576 −216 −72 −432
 (24.35)
 0 0 0 0 288 
 0 0 −216 −72 −288 −720 
 72 0 72 0 864 288 
 0 0 −216 144 −216 144 0 −72 576 −216 −144 0
 
 0 0 144 −360 144 −360 −72 −288 −216 864 0 144 
 
0 −36 0 0 0 −36 −432 288 −144 0 576 −216
−36 −144 0 0 −36 −144 288 −720 0 144 −216 864
The eigenvalues are:
[ 1971.66 1416.75 694.82 545.72 367.7 175.23 157.68 57.54 12.899 0 0 0 ] (24.36)
The 3 zero eigenvalues pertain to the three independent rigid-body modes. The 9 other ones are positive.
Consequently the computed Ke has the correct rank of 9.

24–12
24–13 §24.4 THE QUADRATIC TRIANGLE

ClearAll[Em,ν,a,b,e,h]; h=1; Em=288; ν=1/3;

ncoor={{0,0},{6,2},{4,4},{3,1},{5,3},{2,2}};
Emat=Em/(1-ν^2)*{{1,ν,0},{ν,1,0},{0,0,(1-ν)/2}};
Print["Emat=",Emat//MatrixForm]
Ke=Trig6IsoPMembraneStiffness[ncoor,Emat,h,{False,3}];
Ke=Simplify[Ke]; Print[Chop[Ke]//MatrixForm];
Print["eigs of Ke=",Chop[Eigenvalues[N[Ke]]]];

Figure 24.8. Script to form Ke of test triangle of Figure 24.7.

√
3 ( 0, 3/2γ)
§24.4.4. Test on Highly Distorted Triangle
√ √
6 (−1/2, 1/ 3 ) 5 (1/2, 1/ 3 )
A highly distorted test geometry places the 3 corners
at the vertices √ of an equilateral triangle: {−1/2, 0},
{1/2, 0}, {0, √ 3/2} whereas√the 3 side nodes are √ located
at {0, −1/(2 3)}, {1/2, 1/ 3} and {−1/2, 1/ 3}. √ The 2 (1/2,0)
1 (−1/2,0)
result is that the six nodes lie on a circle of radius 1/ 3, as
depicted in Figure 24.9. The element has unit thickness. √
4 ( 0, −1/(2 3) )
The material is isotropic with E = 504 and ν = 0. The
stiffness Ke is evaluated for four rank-sufficient rules: Figure 24.9. Test curved sided quadratic
3,-3,6,7, using the script of Figure 24.10. triangle, with the 6 nodes lying on a circle.

ClearAll[Em,ν,h]; h=1; Em=7*72; ν=0; h=1;

{x1,y1}={-1,0}/2; {x2,y2}={1,0}/2; {x3,y3}={0,Sqrt[3]}/2;
{x4,y4}={0,-1/Sqrt[3]}/2; {x5,y5}={1/2,1/Sqrt[3]};
{x6,y6}={-1/2,1/Sqrt[3]};
ncoor= {{x1,y1},{x2,y2},{x3,y3},{x4,y4},{x5,y5},{x6,y6}};
Emat=Em/(1-ν^2)*{{1,ν,0},{ν,1,0},{0,0,(1-ν)/2}};
For [i=2,i<=5,i++, p={1,-3,3,6,7}[[i]];
Ke=Trig6IsoPMembraneStiffness[ncoor,Emat,h,{True,p}];
Ke=Chop[Simplify[Ke]];
Print["Ke=",SetPrecision[Ke,4]//MatrixForm];
Print["Eigenvalues of Ke=",Chop[Eigenvalues[N[Ke]],.0000001]]
];

Figure 24.10. Script to form Ke for the triangle of Figure 24.9, using four integration rules.

For rule=3 the result (printed to 4 places because of the SetPrecision statement in script) is
 
344.7 75.00 −91.80 21.00 −86.60 −24.00 −124.7 −72.00 −20.78 −36.00 −20.78 36.00
 75.00 258.1 −21.00 −84.87 18.00 −90.07 96.00 0 −36.00 20.78 −132.0 −103.9 
 −91.80 −21.00 −75.00 −86.60 −124.7 −20.78 −36.00 −20.78 36.00 
 344.7 24.00 72.00 
 21.00 −84.87 −75.00 258.1 −18.00 −90.07 −96.00 0 132.0 −103.9 36.00 20.78 
 
 −86.60 18.00 −86.60 −18.00 214.8 0 41.57 0 −41.57 144.0 −41.57 −144.0 
 
 −24.00 −90.07 24.00 −90.07 0 388.0 0 −41.57 −24.00 −83.14 24.00 −83.14 
 −124.7 −124.7 −96.00 −83.14 −72.00 −83.14

 96.00 41.57 0 374.1 0 72.00 
 −72.00 −41.57 −72.00 −166.3 −166.3 
 0 72.00 0 0 0 374.1 72.00 
 −20.78 −36.00 −20.78 132.0 −41.57 −24.00 −83.14 −72.00 374.1 0 −207.8 0
 
 −36.00 20.78 −36.00 −103.9 144.0 −83.14 −72.00 −166.3 0 374.1 0 −41.57 
 
−20.78 −132.0 −20.78 36.00 −41.57 24.00 −83.14 72.00 −207.8 0 374.1 0
36.00 −103.9 36.00 20.78 −144.0 −83.14 72.00 −166.3 0 −41.57 0 374.1
(24.37)

24–13
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–14

For rule=-3:
 
566.4 139.0 129.9 21.00 79.67 8.000 −364.9 −104.0 −205.5 −36.00 −205.5 −28.00
 139.0 405.9 −21.00 62.93 50.00 113.2 64.00 −129.3 −36.00 −164.0 −196.0 −288.7 
 129.9 −21.00 −139.0 −8.000 −364.9 −205.5 −205.5 36.00 
 566.4 79.67 104.0 28.00 
 21.00 62.93 −139.0 405.9 −50.00 113.2 −64.00 −129.3 196.0 −288.7 36.00 −164.0 
 
 79.67 50.00 79.67 −50.00 325.6 0 −143.2 0 −170.9 176.0 −170.9 −176.0 
 
 8.000 113.2 −8.000 113.2 0 646.6 0 −226.3 8.000 −323.3 −8.000 −323.3 
 −364.9 −364.9 −64.00 −143.2 −104.0

 64.00 0 632.8 0 120.1 120.1 104.0 
 −104.0 −129.3 −129.3 −226.3 −104.0 0
 104.0 0 0 485.0 0 104.0 
 −205.5 −36.00 −205.5 196.0 −170.9 8.000 120.1 −104.0 521.9 −64.00 −60.04 0
 
 −36.00 −164.0 28.00 −288.7 176.0 −323.3 −104.0 0 −64.00 595.8 0 180.1 
 
−205.5 −196.0 −205.5 36.00 −170.9 −8.000 120.1 104.0 −60.04 0 521.9 64.00
−28.00 −288.7 36.00 −164.0 −176.0 −323.3 104.0 0 0 180.1 64.00 595.8
(24.38)
The stiffness for rule=6 is omitted to save space. For rule=7:
 
661.9 158.5 141.7 21.00 92.53 7.407 −432.1 −117.2 −190.2 −29.10 −273.8 −40.61
 158.5 478.8 −21.00 76.13 49.41 125.3 50.79 −182.7 −29.10 −156.6 −208.6 −341.0 
 141.7 −21.00 −158.5 −7.407 −432.1 −273.8 −190.2 29.10 
 661.9 92.53 117.2 40.61 
 21.00 76.13 −158.5 478.8 −49.41 125.3 −50.79 −182.7 208.6 −341.0 29.10 −156.6 
 
 92.53 49.41 92.53 −49.41 387.3 0 −139.8 0 −216.3 175.4 −216.3 −175.4 
 
 7.407 125.3 −7.407 125.3 0 753.4 0 −207.0 7.407 −398.5 −7.407 −398.5 
 −432.1 −432.1 −50.79 −139.8 −117.2

 50.79 0 723.6 0 140.2 140.2 117.2 
 −117.2 −182.7 −182.7 −207.0 −117.2 4.884 
 117.2 0 0 562.6 4.884 117.2 
 −190.2 −29.10 −273.8 208.6 −216.3 7.407 140.2 −117.2 602.8 −69.71 −62.78 0
 
 −29.10 −156.6 40.61 −341.0 175.4 −398.5 −117.2 4.884 −69.71 683.3 0 207.9 
 
−273.8 −208.6 −190.2 29.10 −216.3 −7.407 140.2 117.2 −62.78 0 602.8 69.71
−40.61 −341.0 29.10 −156.6 −175.4 −398.5 117.2 4.884 0 207.9 69.71 683.3
(24.39)
The eigenvalues of these matrices are:

Rule Eigenvalues of Ke
3 702.83 665.11 553.472 553.472 481.89 429.721 429.721 118.391 118.391 0 0 0
−3 1489.80 1489.80 702.833 665.108 523.866 523.866 481.890 196.429 196.429 0 0 0
6 1775.53 1775.53 896.833 768.948 533.970 533.970 495.570 321.181 321.181 0 0 0
7 1727.11 1727.11 880.958 760.719 532.750 532.750 494.987 312.123 312.123 0 0 0
(24.40)
Since the metric of this element is highly distorted near its boundary, the stiffness matrix entries and
eigenvalues change significantly as the integration formulas are advanced from 3 to 6 and 7 points.
However as can be seen the matrix remains rank-sufficient.

§24.5. *The Cubic Triangle

The 10-node cubic triangle, depicted in Figure 24.11, is rarely used as such in applications because of the difficulty
of combining it with other elements. Nevertheless the derivation of the element modules is instructive as this
triangle has other and more productive uses as a generator of more practical elements with “drilling” rotational
degrees of freedom at corners for modeling shells. Such transformations are studied in advanced FEM courses.
The geometry of the triangle is defined by the coordinates of the ten nodes. A notational warning: the interior node
is labeled as 0 instead of 10 (cf. Figure 24.8) to avoid confusion with notation such as x12 = x1 − x2 for coordinate
differences.

24–14
24–15 §24.5 *THE CUBIC TRIANGLE

§24.5.1. *Shape Function Module

The shape functions obtained in Exercise 18.1 are N1 = 12 ζ1 (3ζ1 − 1)(3ζ1 − 2), N2 = 12 ζ2 (3ζ2 − 1)(3ζ2 − 2),
N3 = 12 ζ3 (3ζ3 −1)(3ζ3 −2), N4 = 92 ζ1 ζ2 (3ζ1 −1), N5 = 92 ζ1 ζ2 (3ζ2 −1), N6 = 92 ζ2 ζ3 (3ζ2 −1), N7 = 92 ζ2 ζ3 (3ζ3 −1),
N8 = 92 ζ3 ζ1 (3ζ3 − 1), N9 = 92 ζ3 ζ1 (3ζ1 − 1) and N0 = 27ζ1 ζ2 ζ3 . Their natural derivatives are

 2 −2ζ +3ζ 2   0   0 
9 1 1
 0  
2
−2ζ2 +3ζ22   0 
     2
9
 0   0  
2
9
−2ζ 3 +3ζ3 
 ζ2 (6ζ1 −1)   ζ1 (3ζ1 −1)   0 
∂NT
  ∂NT
  ∂NT
 
9  ζ2 (3ζ2 −1)  9  ζ1 (6ζ2 −1)  9  0 
= 2 , = 2 , = 2 . (24.41)
∂ζ1  0  ∂ζ2  ζ3 (6ζ2 −1)  ∂ζ3  ζ2 (3ζ2 −1) 
   ζ3 (3ζ3 −1)   ζ2 (6ζ3 −1) 
 0     
 ζ3 (3ζ3 −1)   0   ζ1 (6ζ3 −1) 
     
ζ3 (6ζ1 −1) 0 ζ1 (3ζ1 −1)
6ζ2 ζ3 6ζ1 ζ3 6ζ1 ζ2
3
As in the case of the 6-node triangle it is convenient to introduce
the deviations from thirdpoints and centroid: x4 = x4 − 8 7
1
3
(2x1 +x2 ), x5 = x5 − 13 (x1 +2x2 ), x6 = x6 − 13 (2x2 +x3 ),
x7 = x7 − 13 (x2 +2x3 ), x8 = x8 − 13 (2x3 +x1 ), x9 = x9 − 9 6
0
1
3
(x3 +2x1 ), x0 = x0 − 13 (x1 +x2 +x3 ), y4 = y4 − 13 (2y1 +y2 ), 2
y5 = y5 − 13 (y1 +2y2 ), y6 = y6 − 13 (2y2 +y3 ), y7 = y7 − 4 5
1
1
3
(y2 +2y3 ), y8 = y8 − 13 (2y3 +y1 ), y9 = y9 − 13 (y3 +2y1 ),
and y0 = y0 − 13 (y1 + y2 + y3 ). Figure 24.11. The 10-node cubic triangle.

Using (24.22) and (24.41), the shape function partial derivatives can be worked out to be
   
Jy23 ( 29 − 2ζ1 + 3ζ12 /2) Jx32 ( 29 − 2ζ1 + 3ζ12 /2)
 Jy31 ( 29 − 2ζ2 + 3ζ22 /2)   Jx13 ( 29 − 2ζ2 + 3ζ22 /2) 
   
 Jy12 ( 9 − 2ζ3 + 3ζ3 /2)
2 2   Jx21 ( 9 − 2ζ3 + 3ζ3 /2)
2 2 
   
 Jy31 ζ1 (3ζ1 − 1) + Jy23 ζ2 (6ζ1 − 1)   Jx13 ζ1 (3ζ1 − 1) + Jx32 ζ2 (6ζ1 − 1) 
∂N T
9  Jy23 ζ2 (3ζ2 − 1) + Jy31 ζ1 (6ζ2 − 1) 

, ∂N T
9  Jx32 ζ2 (3ζ2 − 1) + Jx13 ζ1 (6ζ2 − 1) 

.
= =
∂x 4J 
 Jy12 ζ2 (3ζ2 − 1) + Jy31 ζ3 (6ζ2 − 1) 
 ∂y 4J 
 Jx21 ζ2 (3ζ2 − 1) + Jx13 ζ3 (6ζ2 − 1) 

 Jy31 ζ3 (3ζ3 − 1) + Jy12 ζ2 (6ζ3 − 1)   Jx13 ζ3 (3ζ3 − 1) + Jx21 ζ2 (6ζ3 − 1) 
   
 Jy23 ζ3 (3ζ3 − 1) + Jy12 ζ1 (6ζ3 − 1)   Jx32 ζ3 (3ζ3 − 1) + Jx21 ζ1 (6ζ3 − 1) 
 J ζ (3ζ − 1) + J ζ (6ζ − 1)   J ζ (3ζ − 1) + J ζ (6ζ − 1) 
y12 1 1 y23 3 1 x21 1 1 x32 3 1
6(Jy12 ζ1 ζ2 + Jy31 ζ1 ζ3 + Jy23 ζ2 ζ3 ) 6(Jx21 ζ1 ζ2 + Jx13 ζ1 ζ3 + Jx32 ζ2 ζ3 )
(24.42)
where the expressions of the Jacobian coefficients are
    
Jx21 = x21 + 9
x4 ζ1 (3ζ1 −6ζ2 −1)+ζ2 + x5 ζ2 (1−3ζ2 +6ζ1 )−ζ1 + x6 ζ3 (6ζ2 −1)

2

+ x7 ζ3 (3ζ3 −1) + x8 ζ3 (1−3ζ3 ) + x9 ζ3 (1−6ζ1 ) + 6x0 ζ3 (ζ1 −ζ2 ) ,

  
Jx32 = x32 + 9
x4 ζ1 (1−3ζ1 ) + x5 ζ1 (1−6ζ2 ) + x6 ζ2 (3ζ2 −6ζ3 −1)+ζ3
  
2
(24.43)
+ x7 ζ3 (1−3ζ3 +6ζ2 )−ζ2 + x8 ζ1 (6ζ3 −1) + x9 ζ1 (3ζ1 −1) + 6x0 ζ1 (ζ2 −ζ3 ) ,

Jx13 = x13 + 9
x4 (6ζ1 −1)ζ2 + x5 ζ2 (3ζ2 −1) + x6 ζ2 (1−3ζ2 ) + x7 ζ2 (1−6ζ3 )
    
2

+ x8 ζ3 (3ζ3 −6ζ1 −1)+ζ1 + x9 ζ1 (1−3ζ1 +6ζ3 )−ζ3 + 6x0 ζ2 (ζ3 −ζ1 ) ,

24–15
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–16

Trig10IsoPShapeFunDer[ncoor_,tcoor_]:= Module[
{ζ 1,ζ 2,ζ 3,x1,x2,x3,x4,x5,x6,x7,x8,x9,x0,y1,y2,y3,y4,y5,y6,y7,y8,y9,y0,
dx4,dx5,dx6,dx7,dx8,dx9,dx0,dy4,dy5,dy6,dy7,dy8,dy9,dy0,
Jx21,Jx32,Jx13,Jy12,Jy23,Jy31,Nf,dNx,dNy,Jdet},
{{x1,y1},{x2,y2},{x3,y3},{x4,y4},{x5,y5},{x6,y6},{x7,y7},
{x8,y8},{x9,y9},{x0,y0}}=ncoor; {ζ 1,ζ 2,ζ 3}=tcoor;
dx4=x4-(2*x1+x2)/3; dx5=x5-(x1+2*x2)/3; dx6=x6-(2*x2+x3)/3;
dx7=x7-(x2+2*x3)/3; dx8=x8-(2*x3+x1)/3; dx9=x9-(x3+2*x1)/3;
dy4=y4-(2*y1+y2)/3; dy5=y5-(y1+2*y2)/3; dy6=y6-(2*y2+y3)/3;
dy7=y7-(y2+2*y3)/3; dy8=y8-(2*y3+y1)/3; dy9=y9-(y3+2*y1)/3;
dx0=x0-(x1+x2+x3)/3; dy0=y0-(y1+y2+y3)/3;
Nf={ζ 1*(3*ζ 1-1)*(3*ζ 1-2),ζ 2*(3*ζ 2-1)*(3*ζ 2-2),ζ 3*(3*ζ 3-1)*(3*ζ 3-2),
9*ζ 1*ζ 2*(3*ζ 1-1),9*ζ 1*ζ 2*(3*ζ 2-1),9*ζ 2*ζ 3*(3*ζ 2-1),
9*ζ 2*ζ 3*(3*ζ 3-1),9*ζ 3*ζ 1*(3*ζ 3-1),9*ζ 3*ζ 1*(3*ζ 1-1),54*ζ 1*ζ 2*ζ 3}/2;
Jx21=x2-x1+(9/2)*(dx4*(ζ 1*(3*ζ 1-6*ζ 2-1)+ζ 2)+
dx5*(ζ 2*(1-3*ζ 2+6*ζ 1)-ζ 1)+dx6*ζ 3*(6*ζ 2-1)+dx7*ζ 3*(3*ζ 3-1)+
dx8*ζ 3*(1-3*ζ 3)+dx9*ζ 3*(1-6*ζ 1)+6*dx0*ζ 3*(ζ 1-ζ 2));
Jx32=x3-x2+(9/2)*(dx4*ζ 1*(1-3*ζ 1)+dx5*ζ 1*(1-6*ζ 2)+
dx6*(ζ 2*(3*ζ 2-6*ζ 3-1)+ζ 3)+dx7*(ζ 3*(1-3*ζ 3+6*ζ 2)-ζ 2)+
dx8*ζ 1*(6*ζ 3-1)+dx9*ζ 1*(3*ζ 1-1)+6*dx0*ζ 1*(ζ 2-ζ 3)) ;
Jx13=x1-x3+(9/2)*(dx4*(6*ζ 1-1)*ζ 2+dx5*ζ 2*(3*ζ 2-1)+
dx6*ζ 2*(1-3*ζ 2)+dx7*ζ 2*(1-6*ζ 3)+dx8*(ζ 3*(3*ζ 3-6*ζ 1-1)+ζ 1)+
dx9*(ζ 1*(1-3*ζ 1+6*ζ 3)-ζ 3)+6*dx0*ζ 2*(ζ 3-ζ 1));
Jy12=y1-y2-(9/2)*(dy4*(ζ 1*(3*ζ 1-6*ζ 2-1)+ζ 2)+
dy5*(ζ 2*(1-3*ζ 2+6*ζ 1)-ζ 1)+dy6*ζ 3*(6*ζ 2-1)+dy7*ζ 3*(3*ζ 3-1)+
dy8*ζ 3*(1-3*ζ 3)+dy9*ζ 3*(1-6*ζ 1)+6*dy0*ζ 3*(ζ 1-ζ 2));
Jy23=y2-y3-(9/2)*(dy4*ζ 1*(1-3*ζ 1)+dy5*ζ 1*(1-6*ζ 2)+
dy6*(ζ 2*(3*ζ 2-6*ζ 3-1)+ζ 3)+dy7*(ζ 3*(1-3*ζ 3+6*ζ 2)-ζ 2)+
dy8*ζ 1*(6*ζ 3-1)+dy9*ζ 1*(3*ζ 1-1)+6*dy0*ζ 1*(ζ 2-ζ 3)) ;
Jy31=y3-y1-(9/2)*(dy4*(6*ζ 1-1)*ζ 2+dy5*ζ 2*(3*ζ 2-1)+
dy6*ζ 2*(1-3*ζ 2)+dy7*ζ 2*(1-6*ζ 3)+dy8*(ζ 3*(3*ζ 3-6*ζ 1-1)+ζ 1)+
dy9*(ζ 1*(1-3*ζ 1+6*ζ 3)-ζ 3)+6*dy0*ζ 2*(ζ 3-ζ 1));
Jdet = Jx21*Jy31-Jy12*Jx13;
dNx={Jy23*(2/9-2*ζ 1+3*ζ 1^2),Jy31*(2/9-2*ζ 2+3*ζ 2^2),Jy12*(2/9-2*ζ 3+3*ζ 3^2),
Jy31*ζ 1*(3*ζ 1-1)+Jy23*ζ 2*(6*ζ 1-1),Jy23*ζ 2*(3*ζ 2-1)+Jy31*ζ 1*(6*ζ 2-1),
Jy12*ζ 2*(3*ζ 2-1)+Jy31*ζ 3*(6*ζ 2-1),Jy31*ζ 3*(3*ζ 3-1)+Jy12*ζ 2*(6*ζ 3-1),
Jy23*ζ 3*(3*ζ 3-1)+Jy12*ζ 1*(6*ζ 3-1),Jy12*ζ 1*(3*ζ 1-1)+Jy23*ζ 3*(6*ζ 1-1),
6*(Jy12*ζ 1*ζ 2+Jy31*ζ 1*ζ 3+Jy23*ζ 2*ζ 3)}/(2*Jdet/9);
dNy={Jx32*(2/9-2*ζ 1+3*ζ 1^2),Jx13*(2/9-2*ζ 2+3*ζ 2^2),Jx21*(2/9-2*ζ 3+3*ζ 3^2),
Jx13*ζ 1*(3*ζ 1-1)+Jx32*ζ 2*(6*ζ 1-1),Jx32*ζ 2*(3*ζ 2-1)+Jx13*ζ 1*(6*ζ 2-1),
Jx21*ζ 2*(3*ζ 2-1)+Jx13*ζ 3*(6*ζ 2-1),Jx13*ζ 3*(3*ζ 3-1)+Jx21*ζ 2*(6*ζ 3-1),
Jx32*ζ 3*(3*ζ 3-1)+Jx21*ζ 1*(6*ζ 3-1),Jx21*ζ 1*(3*ζ 1-1)+Jx32*ζ 3*(6*ζ 1-1),
6*(Jx21*ζ 1*ζ 2+Jx13*ζ 1*ζ 3+Jx32*ζ 2*ζ 3)}/(2*Jdet/9);
Return[Simplify[{Nf,dNx,dNy,Jdet}]]
];

Figure 24.12. Shape function module for 10-node cubic triangle.

and
  
Jy12 = y12 − 9
y4 ζ1 (3ζ1 −6ζ2 −1)+ζ2 + y5 (ζ2 (1−3ζ2 +6ζ1 )−ζ1 ) + y6 ζ3 (6ζ2 −1)

2

+ y7 ζ3 (3ζ3 −1) + y8 ζ3 (1−3ζ3 ) + y9 ζ3 (1−6ζ1 ) + 6y0 ζ3 (ζ1 −ζ2 ) ,

  
Jy23 = y23 − 9
y4 ζ1 (1−3ζ1 ) + y5 ζ1 (1−6ζ2 ) + y6 ζ2 (3ζ2 −6ζ3 −1)+ζ3
  
2
(24.44)
+ y7 ζ3 (1−3ζ3 +6ζ2 )−ζ2 + y8 ζ1 (6ζ3 −1) + y9 ζ1 (3ζ1 −1) + 6y0 ζ1 (ζ2 −ζ3 ) ,

Jy31 = y31 − 9
y4 (6ζ1 −1)ζ2 + y5 ζ2 (3ζ2 −1) + y6 ζ2 (1−3ζ2 ) + y7 ζ2 (1−6ζ3 )
    
2

+ y8 ζ3 (3ζ3 −6ζ1 −1)+ζ1 + y9 ζ1 (1−3ζ1 +6ζ3 )−ζ3 + 6y0 ζ2 (ζ3 −ζ1 ) .

The Jacobian determinant is J = 12 (Jx21 Jy31 − Jy12 Jx13 ). The shape function module Trig10IsoPShapeFunDer
that implements these expressions is listed in Figure 24.12. This has the same arguments as Trig6IsoPShapeFunDer.
As there Jdet denotes 2J .

24–16
24–17 §24.5 *THE CUBIC TRIANGLE

Trig10IsoPMembraneStiffness[ncoor_,Emat_,th_,options_]:=
Module[{i,k,l,p=6,numer=False,h=th,tcoor,w,c,
Nf,dNx,dNy,Jdet,Be,Ke=Table[0,{20},{20}]},
If [Length[options]>=1, numer=options[[1]]];
If [Length[options]>=2, p= options[[2]]];
If [!MemberQ[{1,3,-3,6,7},p], Print["Illegal p"]; Return[Null]];
For [k=1, k<=Abs[p], k++,
{tcoor,w}= TrigGaussRuleInfo[{p,numer},k];
{Nf,dNx,dNy,Jdet}= Trig10IsoPShapeFunDer[ncoor,tcoor];
If [numer, {Nf,dNx,dNy,Jdet}=N[{Nf,dNx,dNy,Jdet}]];
If [Length[th]==10, h=th.Nf]; c=w*Jdet*h/2;
Be= {Flatten[Table[{dNx[[i]],0 },{i,10}]],
Flatten[Table[{0, dNy[[i]]},{i,10}]],
Flatten[Table[{dNy[[i]],dNx[[i]]},{i,10}]]};
Ke+=c*Transpose[Be].(Emat.Be);
]; If[!numer,Ke=Simplify[Ke]]; Return[Ke]
];

Figure 24.13. Stiffness module for 10-node cubic triangle.

§24.5.2. *Stiffness Module

Module Trig10IsoPMembraneStiffness, listed in Figure 24.13, implements the computation of the element
stiffness matrix of the 10-node cubic plane stress triangle. The module is invoked as

Ke = Trig10IsoPMembraneStiffness[ncoor,Emat,th,options] (24.45)

this has the same arguments of the quadratic triangle module invoked in (24.32), with the following changes. The
node coordinate list ncoor contains 10 entries: { { x1,y1 },{ x2,y2 },{ x3,y3 }, . . .,{ x9,y9 },{ x0,y0 } }. If the
plate thickness varies, th is a list of 10 entries: { h1,h2,h3,h4,. . .,h9,h0 }. The triangle Gauss rule specified in
options as { numer,rule } should be 6 or 7 to produce a rank sufficient stiffness. If omitted rule = 6 is assumed.
The module returns Ke as an 20 × 20 symmetric matrix pertaining to the following arrangement of nodal displace-
ments:
ue = [ u x1 u y1 u x2 u y2 u x3 u y3 u x4 u y4 . . . u x9 u y9 u x0 u y0 ]T . (24.46)

Remark 24.3. For symbolic work with this element the 7-point rule should be preferred because the exact
expressions of the abcissas and weigths at sample points are simpler than for the 6-point rule, as can be observed
in Figure 24.3. This speeds up algebraic simplification. For numerical work the 6-point rule is slightly faster.

§24.5.3. *Test Element 3 (4,4)

The stiffness module is tested on the superparametric triangle 7

geometry shown in Figure 24.14, which has been already 8 6
used for the quadratic triangle.
The corner nodes are placed at (0, 0), (6, 2) and (4, 4). The 0
2 (6,2)
9
6 side nodes are located at the thirdpoints of the sides and the 5
interior node at the centroid. The plate has unit thickness.
4
The material is isotropic with E = 1920 and ν = 0.
1 (0,0)
The script of Figure 24.15 computes Ke using the 7-point
Gauss rule and exact arithmetic. Figure 24.14. Test 10-node triangle.

The returned stiffness matrix using either 6- or 7-point integration is the same, since for a superparametric element
the integrand is quartic in the triangular coordinates. For the given combination of inputs the entries are exact
integers:

24–17
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–18

 306 102 −42 −42 −21 21 −315 −333 171 153

 102 306 42 42 −63 −105 351 369 −135 −117
 −42 42 1224 −408 −210 42 252 −180 −234 306

 −42 42 −408 1224 126 −294 108 −36 −378 450

 −21 −63 −210 126 1122 −306 −99 27 −99 27
 21 −105 42 −294 −306 −171 −171
 1938 27 27
 −315 −99 5265 −1215 −1539 −243
 351 252 108 27
 −333 369 −180 −36 27 −171 −1215 6885 729 −891

 171 −135 −234 −378 −99 27 −1539 729 5265 −1215

 153 −117 306 450 27 −171 −243 −891 −1215 6885
K =
e
 −27 −9 −1602 990 819 −171 81 81 −405 −405
 −9 −27 306 −2286 −459 −405 −2025
 1179 81 405
 −27 −9 828 −468 −1611 315 81 81 81 81

 −9 −27 −180 1116 999 −2223 81 405 81 405

 99 225 −108 36 −72 144 810 −324 810 −324
 −63 36 −108 −540 −684 −324 −324
 387 1134 1134
 −144 −504 −108 −99 −4050 −324
 36 171 1620 810
 180 −828 36 −108 189 531 1620 −5670 −324 1134

0 0 0 0 0 0 −486 −486 −4860 1944
0 0 0 0 0 0 −486 −2430 1944 −6804

−27 −9 −27 −9 99 −63 −144 180 0 0

−9 −27 −9 −27 225 387 −504 −828 0 0
−1602 306 828 −180 −108 36 −108 36 0 0
990 −2286 −468 1116 36 −108 36 −108 0 0

819 −459 −1611 999 −72 −540 171 189 0 0
−171 1179 315 −2223 144 −684 −99 531 0 0

81 81 81 81 810 −324 −4050 1620 −486 −486 

81 405 81 405 −324 1134 1620 −5670 −486 −2430 

−405 −405 81 81 810 −324 810 −324 −4860 1944 

−405 −2025 81 405 −324 1134 −324 1134 1944 −6804 
5265 −1215 −3483 729 162 0 162 0 −972 0

−1215 6885 1701 −4779 0 −162 0 −162 0 972 

−3483 1701 5265 −1215 −810 0 162 0 −486 −486 

729 −4779 −1215 6885 0 810 0 −162 −486 −2430 

162 0 −810 0 5265 −1215 −1296 −486 −4860 1944 
0 −162 0 810 −1215 6885 486 −2592

1944 −6804 
162 0 162 0 −1296 486 5265 −1215 −972 0
0 −162 0 −162 −486 −2592 −1215 6885 0 972 

−972 0 −486 −486 −4860 1944 −972 0 12636 −2916
0 972 −486 −2430 1944 −6804 0 972 −2916 16524
(24.47)
The eigenvalues are

[ 26397. 16597. 14937. 12285. 8900.7 7626.2 5417.8 4088.8 3466.8 3046.4
(24.48)
1751.3 1721.4 797.70 551.82 313.22 254.00 28.019 0 0 0 ]

The 3 zero eigenvalues pertain to the three independent rigid-body modes in two dimensions. The 17 other
eigenvalues are positive. Consequently the computed Ke has the correct rank of 17.

24–18
24–19 §24. References

ClearAll[Em,Ν,a,b,e,h]; Em=1920; Ν=0; h=1;

ncoor={{0,0},{6,2},{4,4}};
x4=(2*x1+x2)/3; x5=(x1+2*x2)/3; y4=(2*y1+y2)/3; y5=(y1+2*y2)/3;
x6=(2*x2+x3)/3; x7=(x2+2*x3)/3; y6=(2*y2+y3)/3; y7=(y2+2*y3)/3;
x8=(2*x3+x1)/3; x9=(x3+2*x1)/3; y8=(2*y3+y1)/3; y9=(y3+2*y1)/3;
x0=(x1+x2+x3)/3; y0=(y1+y2+y3)/3;
ncoor= {{x1,y1},{x2,y2},{x3,y3},{x4,y4},{x5,y5},{x6,y6},{x7,y7},
{x8,y8},{x9,y9},{x0,y0}};
Emat=Em/(1-Ν^2)*{{1,Ν,0},{Ν,1,0},{0,0,(1-Ν)/2}};
Ke=Trig10IsoPMembraneStiffness[ncoor,Emat,h,{False,7}];
Ke=Simplify[Ke]; Print[Ke//MatrixForm]; ev=Chop[Eigenvalues[N[Ke]]];
Print["eigs of Ke=",ev];

Figure 24.15. Script for testing cubic triangle of Figure 24.14.

Notes and Bibliography

The 3-node, 6-node and 10-node plane stress triangular elements are generated by complete polynomials in two-
dimensions. In order of historical appearance:
1. The three-node linear triangle, also known as Constant Strain Triangle (CST) and Turner triangle, was
developed as “triangular skin panel” by Turner, Clough and Martin in 1951–53 [68,69] using interelement
flux assumptions, and published in 1956 [392]. It is not clear when the assumed-displacement derivation,
which yields the same stiffness matrix, was done first. The displacement derivation is mentioned in passing
by Clough in [61] and worked out in the theses of Melosh [261] and Wilson [414].
2. The six-node quadratic triangle, also known as Linear Strain Triangle (LST) and Veubeke triangle, was
developed by B. M. Fraeijs de Veubeke in 1962–63 [433]; published 1965 [149].
3. The ten-node cubic triangle, also known as Quadratic Strain Triangle (QST), was developed by the writer in
1965; published 1966 [100]. Shape functions for the cubic triangle were presented there but used for plate
bending instead of plane stress.
A version of the cubic triangle with freedoms migrated to corners and recombined to produce a “drilling rotation”
at corners, was used in static and dynamic shell analysis in Carr’s thesis under Ray Clough [57,65]. The drilling-
freedom idea was independently exploited for rectangular and quadrilateral elements, respectively, in the theses of
Abu-Ghazaleh [2] and Willam [413], both under Alex Scordelis. A variant of the Willam quadrilateral, developed
by Bo Almroth at Lockheed, has survived in the nonlinear shell analysis code STAGS as element 410 [326].
Numerical integration came into FEM by the mid 1960s. Five triangle integration rules were tabulated in the
writer’s thesis [100, pp. 38–39]. These were gathered from three sources: two papers by Hammer and Stroud
[178,179] and the 1964 Handbook of Mathematical Functions [1, §25.4]. They were adapted to FEM by converting
Cartesian abscissas to triangle coordinates. The table has been reproduced in Zienkiewicz’ book since the second
edition [431, Table 8.2] and, with corrections and additions, in the monograph of Strang and Fix [360, p. 184].
The monograph by Stroud [363] contains a comprehensive collection of quadrature formulas for multi-
ple integrals. That book gathers most of the formulas known by 1970, as well as references until that
year. (Only a small fraction of Stroud’s tabulated rules, however, are suitable for FEM work.) The col-
lection has been periodically kept up to date by Cools [73–75], who also maintains a dedicated web site:
http://www.cs.kuleuven.ac.be/~nines/research/ecf/ecf.html This site provides rule information in
16- and 32-digit accuracy for many geometries and dimensionalities as well as a linked “index card” of references
to source publications.

References
Referenced items moved to Appendix R.

24–19
Chapter 24: IMPLEMENTATION OF ISO-P TRIANGULAR ELEMENTS 24–20

Homework Exercises for Chapter 24

Implementation of Iso-P Triangular Elements
3
EXERCISE 24.1 [C:20] Write an element stiffness module and a shape
function module for the 4-node “transition” iso-P triangular element with
only one side node: 4, which is located between 1 and 2. See Figure E24.1.
The shape functions are N1 = ζ1 − 2ζ1 ζ2 , N2 = ζ2 − 2ζ1 ζ2 , N3 = ζ3 and
N4 = 4ζ1 ζ2 . Use the three interior point integration rule=3.
2
Test the element for the geometry of the triangle depicted in Figure 24.7, 4
removing nodes 5 and 6, and with E = 2880, ν = 1/3 and h = 1.
Report results for the stiffness matrix Ke and its 8 eigenvalues. Note:
1
Notebook Trig6Stiffness.nb for the 6-node triangle, posted on the
Figure E24.1. The 4-node transi-
index of Chapter 24, may be used as “template” in support of this Exercise.
tion triangle for Exercise 24.1.
Partial results: K 11 = 1980, K 18 = 1440.
EXERCISE 24.2 [A/C:5+20] As in the foregoing Exercise, but now write the module for a five-node triangular
“transition” element that lacks midnode 6 opposite corner 2. Begin by deriving the five shape functions.
EXERCISE 24.3 [A+C:25] Consider the superparametric straight-sided 6-node triangle where side nodes are
located at the midpoints. By setting x4 = 12 (x1 + x2 ), x5 = 12 (x2 + x3 ), x3 = 12 (x3 + x1 ), y4 = 12 (y1 + y2 ),
y5 = 12 (y2 + y3 ), y3 = 12 (y3 + y1 ) in (24.26), deduce that
     
1 1 1 1 1 1 1 1 1
J= 3x1 2x1 + x2 2x1 + x3 ζ1 + x1 + 2x2 3x2 2x2 + x3 ζ2 + x1 + 2x3 x2 + 2x3 3x3 ζ3 . (E24.1)
3y1 2y1 + y2 2y1 + y3 y1 + 2y2 3y2 2y2 + y3 y1 + 2y3 y2 + 2y3 3y3
This contradicts publications that, by mistakingly assuming that the results for the linear triangle (Chapter 15) can
be extended by analogy, take  
1 1 1
J = x1 x2 x3 (E24.2)
y1 y2 y3
Show, however, that
 
1 y23 y32
2J = 2A = det J = x3 y12 + x1 y23 + x2 y31 , P= y31 x13 (E24.3)
2J y x21
12
are the same for both (E24.1) and (E24.2). Thus the mistake has no effect on the computation of derivatives.
EXERCISE 24.4 [A/C:15+15] Consider the superparametric straight-sided 6-node triangle where side nodes are
at the midpoints of the sides and the thickness h is constant. Using the 3-midpoint quadrature rule show that the
element stiffness can be expressed in a closed form obtained in 1966 [100]:
 
Ke = 13 Ah B1T EB1 + B2T EB2 + B3T EB3 (E24.4)
in which A is the triangle area and
 
1 y32 0 y31 0 y12 0 2y23 0 2y32 0 2y23 0
B1 = 0 x23 0 x13 0 x21 0 2x32 0 2x23 0 2x32
2A x23 y32 x13 y31 x21 y12 2x32 2y23 2x23 2y32 2x32 2y23
 
1 y23 0 y13 0 y12 0 2y31 0 2y31 0 2y13 0
B2 = 0 x32 0 x31 0 x21 0 2x13 0 2x13 0 2x31 (E24.5)
2A x32 y23 x31 y13 x21 y12 2x13 2y31 2x13 2y31 2x31 2y13
 
1 y23 0 y31 0 y21 0 2y21 0 2y12 0 2y12 0
B3 = 0 x32 0 x13 0 x12 0 2x12 0 2x21 0 2x21
2A x32 y23 x13 y31 x12 y21 2x12 2y21 2x21 2y12 2x21 2y12

24–20
24–21 §24. References

With this form Ke can be computed in approximately 1000 floating-point operations.

Using next the 3-interior-point quadrature rule, show that the element stiffness can be expressed again as (E24.4)
but with
 
1 5y23 0 y13 0 y21 0 2y21 + 6y31 0 2y32 0 6y12 + 2y13 0
B1 = 0 5x32 0 x31 0 x12 0 2x12 + 6x13 0 2x23 0 6x21 + 2x31
6A 5x32 5y23 x31 y13 x12 y21 2x12 + 6x13 2y21 + 6y31 2x23 2y32 6x21 + 2x31 6y12 + 2y13
 
1 y32 0 5y31 0 y21 0 2y21 + 6y23 0 6y12 + 2y32 0 2y13 0
B2 = 0 x23 0 5x13 0 x12 0 2x12 + 6x32 0 6x21 + 2x23 0 2x31
6A x23 y32 5x13 5y31 x12 y21 2x12 + 6x32 2y21 + 6y23 6x21 + 2x23 6y12 + 2y32 2x31 2y13
 
1 y32 0 y13 0 5y12 0 2y21 0 6y31 + 2y32 0 2y13 + 6y23 0
B3 = 0 x23 0 x31 0 5x21 0 2x12 0 6x13 + 2x23 0 2x31 + 6x32
6A x23 y32 x31 y13 5x21 5y12 2x12 2y21 6x13 + 2x23 6y31 + 2y32 2x31 + 6x32 2y13 + 6y23
(E24.6)
The fact that two very different expressions yield the same Ke explains why sometimes authors rediscover the same
element derived with different methods.
Yet another set that produces the correct stiffness is
 
1 y23 0 y31 0 y21 0 2y21 0 2y12 0 2y12 0
B1 = 0 x32 0 x13 0 x12 0 2x12 0 2x21 0 2x21
6A x32 y23 x13 y31 x12 y21 2x12 2y21 2x21 2y12 2x21 2y12
 
1 y32 0 y31 0 y12 0 2y23 0 2y32 0 2y23 0
B2 = 0 x23 0 x13 0 x21 0 2x32 0 2x23 0 2x32 (E24.7)
6A x23 y32 x13 y31 x21 y12 2x32 2y23 2x23 2y32 2x32 2y23
 
1 y23 0 y13 0 y12 0 2y31 0 2y31 0 2y13 0
B3 = 0 x32 0 x31 0 x21 0 2x13 0 2x13 0 2x31
6A x32 y23 x31 y13 x21 y12 2x13 2y31 2x13 2y31 2x31 2y13

EXERCISE 24.5 [A/C:40] (Research paper level) Characterize the most general form of the Bi (i = 1, 2, 3)
matrices that produce the same stiffness matrix Ke in (E24.4). (Entails solving an algebraic Riccati equation.)

EXERCISE 24.6 [A/C:25]. Derive the 4-node transition triangle by forming the 6-node stiffness and applying
MFCs to eliminate 5 and 6. Prove that this technique only works if sides 1–3 and 2–3 are straight with nodes 5 and
6 initially at the midpoint of those sides.

24–21
 25
The Assembly
Process

25–1
Chapter 25: THE ASSEMBLY PROCESS

TABLE OF CONTENTS

Page
§25.1 Introduction . . . . . . . . . . . . . . . . . . . . . 25–3
§25.2 Simplifications . . . . . . . . . . . . . . . . . . . . 25–3
§25.3 Simplified Assemblers . . . . . . . . . . . . . . . . . . 25–4
§25.3.1 A Plane Truss Example Structure . . . . . . . . . . 25–4
§25.3.2 Simplified Assembler Implementation . . . . . . . . . 25–6
§25.4 MET Assemblers . . . . . . . . . . . . . . . . . . . 25–7
§25.4.1 A Plane Stress Assembler . . . . . . . . . . . . . 25–7
§25.4.2 MET Assembler Implementation . . . . . . . . . . 25–9
§25.5 MET-VFC Assemblers . . . . . . . . . . . . . . . . . 25–10
§25.5.1 Node Freedom Arrangement . . . . . . . . . . . . 25–10
§25.5.2 Node Freedom Signature . . . . . . . . . . . . . 25–11
§25.5.3 The Node Freedom Allocation Table . . . . . . . . . 25–12
§25.5.4 The Node Freedom Map Table . . . . . . . . . . . . 25–12
§25.5.5 The Element Freedom Signature . . . . . . . . . . 25–13
§25.5.6 The Element Freedom Table . . . . . . . . . . . . 25–13
§25.5.7 A Plane Trussed Frame Structure . . . . . . . . . . 25–14
§25.5.8 MET-VFC Assembler Implementation . . . . . . . . . 25–17
§25.6 *Handling MultiFreedom Constraints . . . . . . . . . . . 25–19
§25. Notes and Bibliography . . . . . . . . . . . . . . . . . 25–20
§25. Exercises . . . . . . . . . . . . . . . . . . . . . . 25–21

25–2
 §25.2 SIMPLIFICATIONS

§25.1. Introduction

Chapters 20, 23 and 24 dealt with element level operations. Sandwiched between element processing
and solution there is the assembly process that constructs the master stiffness equations. Assembler
examples for special models were given as recipes in the complete programs of Chapters 21 and
22. In the present chapter assembly will be studied with more generality.
The position of the assembler in a DSM-based code for static analysis is sketched in Figure 25.1. In
most codes the assembler is implemented as a element library loop driver. This means that instead
of forming all elements first and then assembling, the assembler constructs one element at a time
in a loop, and immediately merges it into the master equations. This is the merge loop.

Model definition data:

geometry
element connectivity
material ^
K Modify Eqs K Equation
fabrication Assembler
freedom activity for BCs Solver

merge Some equation solvers apply BCs

loop K e and solve simultaneously

Nodal
Element
displacements
ELEMENT
Stiffness LIBRARY
Matrices
To postprocessor

Figure 25.1. Role of assembler in FEM program.

Assembly is the most complicated stage of a production finite element program in terms of data
flow. The complexity is not due to mathematical operations, which merely involve matrix addition,
but interfacing with a large element library. Forming an element requires access to geometry,
connectivity, material, and fabrication data. (In nonlinear analysis, to the state as well.) Merge
requires access to freedom activity data, as well as knowledge of the sparse matrix format in which
K is stored. As illustrated in Figure 25.1, the assembler sits at the crossroads of this data exchange.
The coverage of the assembly process for a general FEM implementation is beyond the scope of
an introductory course. Instead this Chapter takes advantage of assumptions that lead to coding
simplifications. This allows the basic aspects to be covered without excessive delay.

§25.2. Simplifications

The assembly process is considerably simplified if the FEM implementation has these properties:
• All elements are of the same type. For example: all elements are 2-node plane bars.
• The number and configuration of degrees of freedom at each node is the same.
• There are no “gaps” in the node numbering sequence.
• There are no multifreedom constraints treated by master-slave or Lagrange multiplier methods.
• The master stiffness matrix is stored as a full symmetric matrix.

25–3
Chapter 25: THE ASSEMBLY PROCESS

(a) (b) 1 (1,2) (2)

(1) 2 (3,4) 3 (5,6)
E = 3000 and y
A = 2 for all bars
(3) (4) (5)
1 (1) 2 (2) 3

(4) 4 (7,8)
3 (3) (5)
assembly
4 x
4 4 2 (3,4)
1 (1,2) 3 (5,6)

4 (7,8)

Figure 25.2. Left: Example plane truss structure. (a): model definition; (b) disconnection →
assembly process. Numbers in parentheses written after node numbers are global DOF numbers.

If the first four conditions are met the implementation is simpler because the element freedom table
described below can be constructed “on the fly” from the element node numbers. The last condition
simplifies the indexing to access entries of K.

§25.3. Simplified Assemblers

In this section all simplifying assumptions listed in §25.2 are assumed to hold. This allows us to
focus on local-to-global DOF mapping. The process is illustrated on a plane truss structure.

§25.3.1. A Plane Truss Example Structure

The plane truss shown in Figure 25.2(a) will be used to illustrate the details of the assembly process.
The structure has 5 elements, 4 nodes and 8 degrees of freedom. The disconnection-to-assembly
process is pictured in Figure 25.2(b).
Begin by clearing all entries of the 8 × 8 master stiffness matrix K to zero, so that we effectively
start with the null matrix:
 
0 0 0 0 0 0 0 0 1
0 0 0 0 0 0 0 0 2
 
0 0 0 0 0 0 0 0 3
 
0 0 0 0 0 0 0 0 4
K=  (25.1)
0 0 0 0 0 0 0 0 5
 
0 0 0 0 0 0 0 0 6
 
0 0 0 0 0 0 0 0 7
0 0 0 0 0 0 0 0 8

The numbers written after each row of K are the global DOF numbers.
Element (1) joins nodes 1 and 2. The global DOFs of those nodes are 2 × 1 − 1 = 1, 2 × 1 = 1,
2 × 2 − 1 = 3 and 2 × 2 = 4. See Figure 25.2(b). Those four numbers are collected into an array
called the element freedom table, or EFT for short. The element stiffness matrix K(1) is listed on

25–4
 §25.3 SIMPLIFIED ASSEMBLERS

the left below with the EFT entries annotated after the matrix rows. On the right is K upon merging
the element stiffness:
 
1500 0 −1500 0 0 0 0 0 1
 0 0 0 0 0 0 0 0 2
   
1500 0 −1500 0 1  −1500 0 1500 0 0 0 0 0  3
 
 0 0 0 0 2  0 0 0 0 0 0 0 0 4
    (25.2)
−1500 0 1500 0 3  0 0 0 0 0 0 0 0
 
0 0 0 0 4  0 0 0 0 0 0 0 0
 
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0

Element (2) joins nodes 2 and 3. The global DOFs of those nodes are 2 × 2 − 1 = 3, 2 × 2 = 4,
2 × 3 − 1 = 5 and 2 × 3 = 6; thus the EFT is { 3,4,5,6 }. Matrices K(2) and K upon merge are
 
1500 0 −1500 0 0 0 0 0
 0 0 0 0 0 0 0 0
   
1500 0 −1500 0 3  −1500 0 3000 0 −1500 0 0 0  3
 
 0 0 0 0 4  0 0 0 0 0 0 0 0 4
    (25.3)
−1500 0 1500 0 5  0 0 −1500 0 1500 0 0 0  5
 
0 0 0 0 6  0 0 0 0 0 0 0 0 6
 
0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0

Element (3) joins nodes 1 and 4. Its EFT is { 1,2,7,8 }. Matrices K(3) and K upon merge are
 
2268 −576 −1500 0 0 0 −768 576 1
 −576 432 0 0 0 0 576 −432  2
   
768 −576 −768 576 1  −1500 0 3000 0 −1500 0 0 0 
 
 −576 432 576 −432  2  0 0 0 0 0 0 0 0 
   
−768 576 768 −576 7  0 0 −1500 0 1500 0 0 0 
 
576 −432 −576 432 8  0 0 0 0 0 0 0 0 
 
−768 576 0 0 0 0 768 −576 7
576 −432 0 0 0 0 −576 432 8
(25.4)
Element (4) joins nodes 2 and 4. Its EFT is { 3,4,7,8 }. Matrices K(4) and K upon merge are
 
2268 −576 −1500 0 0 0 −768 576
 −576 432 0 0 0 0 576 −432 
   
0 0 0 0 3  −1500 0 3000 0 −1500 0 0 0 3
 
 0 2000 0 −2000  4  0 0 0 2000 0 0 0 −2000  4
   
0 0 0 0 7  0 0 −1500 0 1500 0 0 0 
 
0 −2000 0 2000 8  0 0 0 0 0 0 0 0 
 
−768 576 0 0 0 0 768 −576 7
576 −432 0 −2000 0 0 −576 2432 8
(25.5)

25–5
Chapter 25: THE ASSEMBLY PROCESS

Finally, element (5) joins nodes 3 and 4. Its EFT is { 5,6,7,8 }. Matrices K(5) and K upon merge are
 
2268 −576 −1500 0 0 0 −768 576
 −576 432 0 0 0 0 576 −432 
   
768 576 −768 −576 5  −1500 0 3000 0 −1500 0 0 0 
 
 576 432 −576 −432  6  0 0 0 2000 0 0 0 −2000 
   
−768 −576 768 576 7  0 0 −1500 0 2268 576 −768 −576  5
 
−576 −432 576 432 8  0 0 0 0 576 432 −576 −432  6
 
−768 576 0 0 −768 −576 1536 0 7
576 −432 0 −2000 −576 −432 0 2864 8
(25.6)
Since all elements have been processed, (25.6) is the master stiffness before application of boundary
conditions.

PlaneTrussMasterStiffness[nodxyz_,elenod_,elemat_,elefab_,
eleopt_]:=Module[{numele=Length[elenod],numnod=Length[nodxyz],
e,ni,nj,eft,i,j,ii,jj,ncoor,Em,A,options,Ke,K},
K=Table[0,{2*numnod},{2*numnod}];
For [e=1, e<=numele, e++, {ni,nj}=elenod[[e]];
eft={2*ni-1,2*ni,2*nj-1,2*nj};
ncoor={nodxyz[[ni]],nodxyz[[nj]]};
Em=elemat[[e]]; A=elefab[[e]]; options=eleopt;
Ke=PlaneBar2Stiffness[ncoor,Em,A,options];
For [i=1, i<=4, i++, ii=eft[[i]];
For [j=i, j<=4, j++, jj=eft[[j]];
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]] ];
];
]; Return[K]
];

Figure 25.3. Plane truss assembler module.

nodxyz={{-4,3},{0,3},{4,3},{0,0}};
elenod= {{1,2},{2,3},{1,4},{2,4},{3,4}};
elemat= Table[3000,{5}]; elefab= Table[2,{5}]; eleopt= {True};
K=PlaneTrussMasterStiffness[nodxyz,elenod,elemat,elefab,eleopt];
Print["Master Stiffness of Plane Truss of Fig 25.2:"];
K=Chop[K]; Print[K//MatrixForm];
Print["Eigs of K=",Chop[Eigenvalues[N[K]]]];
Master Stiffness of Plane Truss of Fig 25.2:
2268. −576. −1500. 0 0 0 −768. 576.
−576. 432. 0 0 0 0 576. −432.
−1500. 0 3000. 0 −1500. 0 0 0
0 0 0 2000. 0 0 0 −2000.
0 0 −1500. 0 2268. 576. −768. −576.
0 0 0 0 576. 432. −576. −432.
−768. 576. 0 0 −768. −576. 1536. 0
576. −432. 0 −2000. −576. −432. 0 2864.
Eigs of K={5007.22, 4743.46, 2356.84, 2228.78, 463.703, 0, 0, 0}

Figure 25.4. Plane truss assembler tester and output results.

25–6
 §25.4 MET ASSEMBLERS

§25.3.2. Simplified Assembler Implementation

Figure 25.3 shows a a Mathematica implementation of the assembly process just described. This
assembler calls the element stiffness module PlaneBar2Stiffness of Figure 20.2. The assembler
is invoked by

K = SpaceTrussMasterStiffness[nodxyz,elenod,elemat,elefab,prcopt] (25.7)

The five arguments: nodxyz, elenod, elemat, elefab, and prcopt have the same function as
those described in §21.1.3 for the SpaceTrussMasterStiffness assembler. In fact, comparing
the code in Figure 25.3 to that of Figure 21.1, the similarities are obvious.
Running the script listed in the top of Figure 25.4 produces the output shown in the bottom of
that figure. The master stiffness (25.6) is reproduced. The eigenvalue analysis verifies that the
assembled stiffness has the correct rank of 8 − 3 = 5.
4 (7,8) 5(9,10)
§25.4. MET Assemblers
y (3) (1)
The next complication occurs when ele- 4 5 in
5
ments of different types are to be assem- (2)
bled, while the number and configuration 5 in trig (1)
1(1,2)
of degrees of freedom at each node bar (3)
6 in (4) 2(3,4) 3(5,6)
remains the same. This scenario is quad (2) assembly
common in special-purpose programs for 1 in 1
4 (7,8) 5(9,10)
analysis of specific problems: plane x
bar (4) 2 4 in 3
stress, plate bending, solids. This will be 1in
called a multiple-element-type assembler, E (plate, bars) = 10000 ksi
ν (plate)= 0.25
or MET assembler for short. h (plate) = 0.3 in
A (bars) = 2 sq in 1(1,2)
The only incremental change with respect
to the simplest assemblers is that an array 2(3,4) 3(5,6)
of element types, denoted here by eletyp,
Figure 25.5. A plane stress structure used to test
has to be provided. a MET assembler.

The assembler calls the appropriate element stiffness module according to type, and builds the local
to global DOF mapping accordingly.

§25.4.1. A Plane Stress Assembler

This kind of assembler will be illustrated using the module PlaneStressMasterStiffness. As
its name indicates, this is suitable for use in plane stress analysis programs. We make the following
assumptions:
1 All nodes in the FEM mesh have 2 DOFs, namely the {u x , u y } displacements. There are no
node gaps or MFCs. The master stiffness is stored as a full matrix.
2 Element types can be: 2-node bars, 3-node linear triangles, ot 4-node bilinear quadrilaterals.
These elements can be freely intermixed. Element types are identified with character strings
"Bar2", "Trig3" and "Quad4", respectively.

25–7
Chapter 25: THE ASSEMBLY PROCESS

As noted above, the chief modification over the simplest assembler is that a different stiffness
module is invoked as per type. Returned stiffness matrices are generally of different order; in our
case 4 × 4, 6 × 6 and 8 × 8 for the bar, triangle and quadrilateral, respectively. But the construction
of the EFT is immediate, since node n maps to global freedoms 2n − 1 and 2n. The technique is
illustrated with the 4-element, 5-node plane stress structure shown in Figure 25.5. It consists of
one triangle, one quadrilateral, and two bars. The assembly process goes as follows.
Element (1) is a 3-node triangle with nodes 3, 5 and 4, whence the EFT is { 5,6,9,10,7,8 }. The
element stiffness is

 
500.0 0 −500.0 −600.0 0 600.0 5
 0 1333.3 −400.0 −1333.3 400.0 0 6
 
 −500.0 −400.0 2420.0 1000.0 −1920.0 −600.0  9
K(1) =  (25.8)
 −600.0 −1333.3 1000.0 2053.3 −400.0 −720.0  10
 
0 400.0 −1920.0 −400.0 1920.0 0 7
600.0 0 −600.0 −720.0 0 720.0 8

Element (2) is a 4-node quad with nodes 2, 3, 4 and 1, whence the EFT is { 3,4,5,6,7,8,1,2 }.
The element stiffness computed with a 2 × 2 Gauss rule is

 
2076.4 395.55 −735.28 −679.11 −331.78 −136.71 −1009.3
420.27 3
 395.55 2703.1 −479.11 −660.61 −136.71 −1191.5 −850.95  4
220.27
 
 −735.28 −479.11 2067.1 215.82 386.36 427.34 −1718.1
−164.05  5
 
 −679.11 −660.61 215.82 852.12 627.34 358.30 −164.05
−549.81  6
K(2) = 
 −331.78 −136.71 386.36 627.34 610.92 178.13 −665.50
−668.76  7
 
 −136.71 −1191.5 427.34 358.30 178.13 1433.4 −468.76
−600.15  8
 
−1009.3 220.27 −1718.1 −164.05 −665.50 −468.76 3393.0
412.54 1
420.27 −850.95 −164.05 −549.81 −668.76 −600.15 412.54
2000.9 2
(25.9)
Element (3) is a 2-node bar with nodes 1 and 4, whence the EFT is { 1,2,7,8 }. The element
stiffness is
 
0 0 0 0 1
 0 4000.0 0 −4000.0  2
K(3) =   (25.10)
0 0 0 0 7
0 −4000.0 0 4000.0 8

Element (4) is a 2-node bar with nodes 1 and 2, whence the EFT is { 1,2,3,4 }. The element
stiffness is
 
7071.1 −7071.1 −7071.1 7071.1 1
 −7071.1 7071.1 7071.1 −7071.1  2
K(4) =   (25.11)
−7071.1 7071.1 7071.1 −7071.1 3
7071.1 −7071.1 −7071.1 7071.1 4

Upon assembly, the master stiffness of the plane stress structure, printed with one digit after the

25–8
 §25.4 MET ASSEMBLERS

PlaneStressMasterStiffness[nodxyz_,eletyp_,elenod_,
elemat_,elefab_,prcopt_]:=Module[{numele=Length[elenod],
numnod=Length[nodxyz],ncoor,type,e,enl,neldof,
i,n,ii,jj,eftab,Emat,th,numer,Ke,K},
K=Table[0,{2*numnod},{2*numnod}]; numer=prcopt[[1]];
For [e=1,e<=numele,e++, type=eletyp[[e]];
If [!MemberQ[{"Bar2","Trig3","Quad4"},type], Print["Illegal type",
" of element e=",e," Assembly interrupted"]; Return[K]];
enl=elenod[[e]]; n=Length[enl];
eftab=Flatten[Table[{2*enl[[i]]-1,2*enl[[i]]},{i,1,n}]];
ncoor=Table[nodxyz[[enl[[i]]]],{i,n}];
If [type=="Bar2", Em=elemat[[e]]; A=elefab[[e]];
Ke=PlaneBar2Stiffness[ncoor,Em,A,{numer}] ];
If [type=="Trig3", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Trig3IsoPMembraneStiffness[ncoor,Emat,th,{numer}] ];
If [type=="Quad4", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Quad4IsoPMembraneStiffness[ncoor,Emat,th,{numer,2}] ];
neldof=Length[Ke];
For [i=1,i<=neldof,i++, ii=eftab[[i]];
For [j=i,j<=neldof,j++, jj=eftab[[j]];
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]] ];
];
]; Return[K];
];

Figure 25.6. Implementation of MET assembler for plane stress.

decimal point, is
 10464.0 −6658.5 −8080.4 7291.3 −1718.1 −164.1 −665.5 −468.8 0 0
 −6658.5 13072.0 7491.3 −7922.0 −164.1 −549.8 −668.8 −4600.2 0 0
 −8080.4 7491.3 9147.5 −6675.5 −735.3 −679.1 −331.8 −136.7 0 0
 
 7291.3 −7922.0 −6675.5 9774.1 −479.1 −660.6 −136.7 −1191.5 0 0
 
 −1718.1 −164.1 −735.3 −479.1 2567.1 215.8 386.4 1027.3 −500.0 −600.0 
 −164.1 −549.8 −679.1 −660.6 −400.0

−1333.3 
 215.8 2185.5 1027.3 358.3
 −665.5 −668.8 −331.8 −136.7 −1920.0 −400.0 
 386.4 1027.3 2530.9 178.1 
 −468.8 −4600.2 −136.7 −1191.5 1027.3 358.3 178.1 6153.4 −600.0 −720.0 
 
0 0 0 0 −500.0 −400.0 −1920.0 −600.0 2420.0 1000.0
0 0 0 0 −600.0 −1333.3 −400.0 −720.0 1000.0 2053.3
(25.12)
The eigenvalues of K are

[ 32883. 10517.7 5439.33 3914.7 3228.99 2253.06 2131.03 0 0 0 ] (25.13)

which displays the correct rank.

§25.4.2. MET Assembler Implementation

An implementation of the MET plane stress assembler as a Mathematica module called
PlaneStressMasterStiffness is shown in Figure 25.6. The assembler is invoked by

K = PlaneStressMasterStiffness[nodxyz,eletyp,elenod,elemat,elefab,prcopt]
(25.14)
The only additional argument is eletyp, which is a list of element types.

25–9
Chapter 25: THE ASSEMBLY PROCESS

Em=10000; ν=1/4; h=3/10; A=2;

Emat=Em/(1-ν^2)*{{1,ν,0},{ν,1,0},{0,0,(1-ν)/2}};
nodxyz= {{0,1},{1,0},{5,0},{0,6},{5,6}};
elenod= {{3,5,4},{2,3,4,1},{1,4},{1,2}};
eletyp= {"Trig3","Quad4","Bar2","Bar2"};
elemat= {Emat,Emat,Em,Em};
elefab= {h,h,A,A}; prcopt={True};
K=PlaneStressMasterStiffness[nodxyz,eletyp,elenod,elemat,elefab,prcopt];
Print["Master Stiffness Matrix of Plane Stress Structure:"];
Print[SetPrecision[Chop[K],5]//MatrixForm];
Print["Eigs of K:",Chop[Eigenvalues[N[K]]]];

Master Stiffness Matrix of Plane Stress Structure:

10464. − 6658.5 − 8080.4 7291.3 − 1718.1 − 164.05 − 665.50 − 468.76 0 0
− 6658.5 13072. 7491.3 − 7922.0 − 164.05 − 549.81 − 668.76 − 4600.2 0 0
− 8080.4 7491.3 9147.5 − 6675.5 − 735.28 − 679.11 − 331.78 − 136.71 0 0
7291.3 − 7922.0 − 6675.5 9774.1 − 479.11 − 660.61 − 136.71 − 1191.5 0 0
− 1718.1 − 164.05 − 735.28 − 479.11 2567.1 215.82 386.36 1027.3 − 500.00 − 600.00
− 164.05 − 549.81 − 679.11 − 660.61 215.82 2185.5 1027.3 358.30 − 400.00 − 1333.3
− 665.50 − 668.76 − 331.78 − 136.71 386.36 1027.3 2530.9 178.13 − 1920.0 − 400.00
− 468.76 − 4600.2 − 136.71 − 1191.5 1027.3 358.30 178.13 6153.4 − 600.00 − 720.00
0 0 0 0 − 500.00 − 400.00 − 1920.0 − 600.00 2420.0 1000.0
0 0 0 0 − 600.00 − 1333.3 − 400.00 − 720.00 1000.0 2053.3
Eigs of K: {32883., 10517.7, 5439.33, 3914.7, 3228.99, 2253.06, 2131.03, 0, 0, 0}

Figure 25.7. Script for assembling the plane stress model of Figure 25.5 and output results.

The script listed in the top of Figure 25.7 runs module PlaneStressMasterStiffness with the
inputs appropriate to the problem defined in Figure 25.5. The output shown in the bottom of the
figure reproduces the master stiffness matrix (25.12) and the eigenvalues (25.13). .

§25.5. MET-VFC Assemblers

The next complication beyond multiple element types is a major one: allowing nodes to have
different freedom configurations. An assembler that handles this feature will be called a MET-VFC
assembler, where VFC is an acronym for Variable Freedom Configuration.
A MET-VFC assembler gets closer to what is actually implemented in production FEM programs.
A assembler of this kind can handle orientation nodes as well as node “gaps” automatically. Only
two advanced attributes remain: allowing treatment of MFCs by Lagrange multipliers, and handling
a sparse matrix storage scheme. The former can be done by an extension of the concept of element.
The latter requires major programming modifications, however, and is not considered in this Chapter.
The implementation of a MET-VFC assembler requires the definition of additional data structures
pertaining to freedoms, at both node and element levels. These are explained in the following
subsections before giving a concrete application example.
§25.5.1. Node Freedom Arrangement
A key decision made by the implementor of a FEM program for structural analysis is: what freedoms
will be allocated to nodes, and how are they arranged? While many answers are possible, we focus

25–10
 §25.5 MET-VFC ASSEMBLERS

PackNodeFreedomSignature[s_]:=FromDigits[s,10];
UnpackNodeFreedomSignature[p_,m_]:=IntegerDigits[p,10,m];
NodeFreedomCount[p_]:=DigitCount[p,10,1];

Figure 25.8. Node freedom manipulation utility functions.

here on the most common arrangement for a linear three-dimensional FEM program.1 At each node
n three displacements {u xn , u yn , u zn } and three infinitesimal rotations {θxn , θ yn , θzn } are chosen as
freedoms and ordered as follows:
u xn , u yn , u zn , θxn , θ yn , θzn (25.15)
This is a Node Freedom Arrangement or NFA. Once the decision is made on a NFA, the position of
a freedom never changes.2 Thus if (25.15) is adopted, position #2 is forever associated to u yn .
The length of the Node Freedom Arrangement is called the NFA length and often denoted (as a
variable) by nfalen. For the common choice (25.15) of 3D structural codes, nfalen is 6.
§25.5.2. Node Freedom Signature
Having picked a NFA such as (25.15) does not mean that all such freedoms need to be present at
a node. Take for example a model, such as a space truss, that only has translational degrees of
freedom. Then only {u xn , u yn , u zn } will be used, and it becomes unnecessary to carry rotations in
the master equations. There are also cases where allocations may vary from node to node.
To handle such scenarios in general terms a Node Freedom Signature or NFS, is introduced. The
NFS is a sequence of zero-one integers. If the jth entry is one, it means that the jth freedom in
the NFA is allocated; whereas if zero that freedom is not in use. For example, suppose that the
underlying NFA is (25.15) and that node 5 uses the three freedoms {u x5 , u y5 , θz5 }. Then its NFS is
{ 1,1,0,0,0,1 } (25.16)
If a node has no allocated freedoms (e.g, an orientation node), or has never been defined, its NFS
contains only zeros.
It is convenient to decimally pack signatures into integers to save storage because allocating a full
integer to hold just 0 or 1 is wasteful. The packed NFS (25.16) would be 110001. This technique
requires utilities for packing and unpacking. Mathematica functions that provide those services are
listed in Figure 25.8. All of them use built-in functions. A brief description follows.
p = PackNodeFreedomSignature[s] Packs the NFS list s into integer p.
s = UnpackNodeFreedomSignature[p,m] Unpacks the NFS p into a list s of length m, where
m is the NFA length. (The second argument is
necessary in case the NFS has leading zeros.)
k = NodeFreedomCount[p] Returns count of ones in p.

1 This is the scheme used by most commercial codes. In a geometrically nonlinear FEM code, however, finite rotation
measures must be used. Consequently the infinitesimal node rotations {θx , θ y , θz } must be replaced by something else.
2 NFA positions #7 and beyond are available for additional freedom types, such as Lagrange multipliers, temperatures,
pressures, fluxes, etc. It is also possible to reserve more NFA positions for structural freedoms.

25–11
Chapter 25: THE ASSEMBLY PROCESS

NodeFreedomMapTable[nodfat_]:=Module[{numnod=Length[nodfat],
i,nodfmt}, nodfmt=Table[0,{numnod}];
For [i=1,i<=numnod-1,i++, nodfmt[[i+1]]=nodfmt[[i]]+
DigitCount[nodfat[[i]],10,1] ];
Return[nodfmt]];

TotalFreedomCount[nodfat_]:=Sum[DigitCount[nodfat[[i]],10,1],
{i,Length[nodfat]}];

Figure 25.9. Modules to construct the NFMT from the NFAT, and to compute the total number of freedoms.

Example 25.1. PackNodeFreedomSignature[{ 1,1,0,0,0,1 }] returns 110001. The inverse is

UnpackNodeFreedomSignature[110001,6], which returns { 1,1,0,0,0,1 }.
NodeFreedomCount[110001] returns 3.

§25.5.3. The Node Freedom Allocation Table

The Node Freedom Allocation Table, or NFAT, is a node by node list of packed node freedom
signatures. In Mathematica the list containing this table is internally identified as nodfat. The
configuration of the NFAT for the plane stress example structure of Figure 25.5 is

nodfat = { 110000,110000,110000,110000,110000 } = Table[110000,{ 5 }]. (25.17)

This says that only two freedoms: {u x , u y }, are used at each of the five nodes.
A zero entry in the NFAT flags a node with no allocated freedoms. This is either an orientation
node, or an undefined one. The latter case happens when there is a node numbering gap, which is a
common occurence when certain mesh generators are used. In the case of a MET-VFC assembler
such gaps are not considered errors.

§25.5.4. The Node Freedom Map Table

The Node Freedom Map Table, or NFMT, is an array with one entry per node. Suppose that node
n has k ≥ 0 allocated freedoms. These have global equation numbers i + j, j = 1, . . . , k. This i
is called the base equation index: it represent the global equation number before the first equation
to which node n contributes. Obviously i = 0 if n = 1. Base equation indices for all nodes are
recorded in the NFMT.3
In Mathematica code the table is internally identified as nodfmt. Figure 25.9 lists a module that
constructs the NFMT given the NFAT. It is invoked as

nodfmt = NodeFreedomMapTable[nodfat] (25.18)

The only argument is the NFAT. The module builds the NFMT by simple incrementation, and
returns it as function value.

3 The “map” qualifier comes from the use of this array in computing element-to-global equation mapping.

25–12
 §25.5 MET-VFC ASSEMBLERS

ElementFreedomTable[enl_,efs_,nodfat_,nodfmt_,m_]:=Module[
{nelnod=Length[enl],eft,i,j,k=0,ix,n,s,es,sx},
eft=Table[0,{Sum[DigitCount[efs[[i]],10,1],{i,1,nelnod}]}];
For [i=1,i<=nelnod,i++, n=enl[[i]];
s=IntegerDigits[nodfat[[n]],10,m]; ix=0;
sx=Table[0,{m}]; Do [If[s[[j]]>0,sx[[j]]=++ix],{j,1,m}];
es=IntegerDigits[efs[[i]],10,m];
For [j=1,j<=m,j++, If [es[[j]]>0, k++;
If [s[[j]]>0, eft[[k]]=nodfmt[[n]]+sx[[j]] ]]];
];
Return[eft]];

Figure 25.10. Module to construct the Element Freedom Table for a MET-VFC assembler.

Example 25.2. The NFAT for the plane truss-frame structure of Figure 25.11 is nodfat = { 110001,110000,
110001,000000,110001 }. The node freedom counts are { 3,2,3,0,3 }. The call

nodfmt = NodeFreedomMapTable[nodfat] (25.19)

returns { 0,3,5,8,8 } in nodfmt. This can be easily verified visually: 0+3=3, 0+3+2=5, etc.
Figure 25.9 also lists a function TotalFreedomCount that returns the total number of freedoms in
the master equations, given the NFAT as argument.

§25.5.5. The Element Freedom Signature

The Element Freedom Signature or EFS is a data structure that shares similarities with the NFAT, but
provides freedom data at the element rather than global level. More specifically, given an element,
it tells to which degrees of freedom it constributes.
The EFS is best described through an example. Suppose that in a general FEM program all defined
structural nodes have the signature 111111, meaning the arrangement (25.15). The program includes
2-node space bar elements, which contribute only to translational degrees of freedom {u x , u y , u z }.
The EFS of any such element will be

efs = { 111000,111000 } (25.20)

For a 2-node space beam element, which contributes to all six nodal freedoms, the EFS will be

efs = { 111111,111111 } (25.21)

For a 3-node space beam element in which the third node is an orientation node:

efs = { 111111,111111,000000 } (25.22)

This information, along with the NFAT and NFMT, is used to build Element Freedom Tables.

25–13
Chapter 25: THE ASSEMBLY PROCESS

§25.5.6. The Element Freedom Table

The Element Freedom Table or EFT has been encountered in all previous assembler examples. It
is a one dimensional list of length equal to the number of element DOF. If the ith element DOF
maps to the kth global DOF, then the ith entry of EFT contains k. This allows to write the merge
loop compactly. In the simpler assemblers discussed in previous sections, the EFT can be built “on
the fly” simply from element node numbers. For a MET-VFC assembler that is no longer the case.
To take care of the VFC feature, the construction of the EFT is best done by a separate module. A
Mathematica implementation is shown in Figure 25.10. Discussion of the logic, which is not trivial,
is relegated to an Exercise, and we simply describe here the interface. The module is invoked as

eft = ElementFreedomTable[enl,efs,nodfat,nodfmt,m] (25.23)

The arguments are

enl The element node list.
efs The Element Freedom Signature (EFS) list.
nodfat The Node Freedom Allocation Table (NFAT) described in 25.5.3
nodfmt The Node Freedom Map Table (NFMT) described in 25.5.4
m The NFA length, often 6.
The module returns the Element Freedom Table as function value. Examples of use of this module
are provided in the plane trussed frame example below.

Remark 25.1. The EFT constructed by ElementFreedomTable is guaranteed to contain only positive entries
if every freedom declared in the EFS matches a globally allocated freedom in the NFAT. But it is possible
for the returned EFT to contain zero entries. This can only happen if an element freedom does not match a
globally allocated one. For example suppose that a 2-node space beam element with the EFS (25.21) is placed
into a program that does not accept rotational freedoms and thus has a NFS of 111000 at all structural nodes.
Then six of the EFT entries, namely those pertaining to the rotational freedoms, would be zero.
The occurrence of a zero entry in a EFT normally flags a logic or input data error. If detected, the assembler
should print an appropriate error message and abort. The module of Figure 25.10 does not make that decision
because it lacks certain information, such as the element number, that should be placed into the message.

§25.5.7. A Plane Trussed Frame Structure

To illustrate the workings of a MET-VFC assembler we will follow the assembly process of the
structure pictured in Figure 25.11(a). The finite element discretization, element disconnection and
assembly process are illustrated in Figure 25.11(b,c,d). Although the structure is chosen to be plane
to facilitate visualization, it will be considered within the context of a general purpose 3D program
with the freedom arrangement (25.15) at each node.
The structure is a trussed frame. It uses two element types: plane (Bernoulli-Euler) beam-columns
and plane bars. Geometric, material and fabrication data are given in Figure 25.11(a). The FEM
idealization has 4 nodes and 5 elements. Nodes numbers are 1, 2, 3 and 5 as shown in Figure 25.11(b).
Node 4 is purposedly left out to illustrate handling of numbering gaps. There are 11 DOFs, three

25–14
 §25.5 MET-VFC ASSEMBLERS

E=30000 MPa, A=0.02 m 2 , E=200000 MPa y

Izz =0.0004 m4 y (b)
A=0.003 m 2
1 Beam-column 3 Beam-column 5
(a) FEM idealization
Bar
3m E=200000 MPa Bar Bar
A=0.001 m 2
x 2
x
E=200000 MPa 4m 4m disassembly (node 4: undefined)
A=0.001 m2

(c) (1) (2) (d)

1(1,2,3) 3(6,7,8) 5(9,10,11)
1(1,2,3) 3(6,7,8) 5(9,10,11)
assembly
(3) (4) (5)

2(4,5)
2(4,5)

Figure 25.11. Trussed frame structure to illustrate a MET-VFC assembler: (a) original structure
showing dimensions, material and fabrication properties; (b) finite element idealization with bars and
beam column elements; (c) conceptual disassembly; (d) assembly. Numbers written in parentheses
after a node number in (c,d) are the global freedom (equation) numbers allocated to that node.

at nodes 1, 3 and 5, and two at node 2. They are ordered as

Global DOF #: 1 2 3 4 5 6 7 8 9 10 11
DOF: u x1 u y1 θz1 u x2 u y2 u x3 u y3 θz3 u x5 u y5 θz5 (25.24)
Node #: 1 1 1 2 2 3 3 3 5 5 5

The NFAT and NFMT can be constructed by inspection of (25.24) to be

nodfat = { 110001,110000,110001,000000,110001 }
(25.25)
nodfmt = { 0,3,5,8,8 }

The element freedom data structures can be also constructed by inspection:

Elem Type Nodes EFS EFT

(1) Beam-column { 1,3 } { 110001,110001 } { 1,2,3,6,7,8 }
(2) Beam-column { 3,5 } { 110001,110001 } { 6,7,8,9,10,11 }
(3) Bar { 1,2 } { 110000,110000 } { 1,2,4,5 }
(4) Bar { 2,3 } { 110000,110000 } { 4,5,6,7 }
(5) Bar { 2,5 } { 110000,110000 } { 4,5,9,10 } (25.26)

Next we list the element stiffness matrices, computed with the modules discussed in Chapter 20.
The EFT entries are annotated as usual.

25–15
Chapter 25: THE ASSEMBLY PROCESS

Element (1): This is a plane beam-column element with stiffness

 
150. 0. 0. −150. 0. 0. 1
 0. 22.5 45. 0. −22.5 45.  2
 
(1)  0. 45. 120. 0. −45. 60.  3
K =  (25.27)
 −150. 0. 0. 150. 0. 0.  6
 
0. −22.5 −45. 0. 22.5 −45. 7
0. 45. 60. 0. −45. 120. 8

Element (2): This is a plane beam-column with element stiffness identical to the previous one
 
150. 0. 0. −150. 0. 0. 6
 0. 22.5 45. 0. −22.5 45.  7
 
(2)  0. 45. 120. 0. −45. 60.  8
K =  (25.28)
 −150. 0. 0. 150. 0. 0.  9
 
0. −22.5 −45. 0. 22.5 −45. 10
0. 45. 60. 0. −45. 120. 11
Element (3): This is a plane bar element with stiffness
 
25.6 −19.2 −25.6 19.2 1
 −19.2 14.4 19.2 −14.4  2
K(3) =   (25.29)
−25.6 19.2 25.6 −19.2 4
19.2 −14.4 −19.2 14.4 5

Element (4): This is a plane bar element with stiffness

 
0 0 0 0 4
 0 200. 0 −200.  5
K(4) =   (25.30)
0 0 0 0 6
0 −200. 0 200. 7

Element (5): This is a plane bar element with stiffness

 
25.6 19.2 −25.6 −19.2 4
 19.2 14.4 −19.2 −14.4  5
K(5) =   (25.31)
−25.6 −19.2 25.6 19.2 9
−19.2 −14.4 19.2 14.4 10

Upon merging the 5 elements the master stiffness matrix becomes

 
175.6 −19.2 0 −25.6 19.2 −150. 0 0 0 0 0 1
 −19.2 36.9 45. 19.2 −14.4 0 −22.5 45. 0 0 0 2
 
 0 45. 120. 0 0 0 −45. 60. 0 0 0 3
 
 −25.6 19.2 0 51.2 0 0 0 0 −25.6 −19.2 0 4
 
 19.2 −14.4 0 0 228.8 0 −200. 0 −19.2 −14.4 0 5
 
K =  −150. 0 0 0 0 300. 0 0 −150. 0 0 6
 
 0 −22.5 −45. 0 −200. 0 245. 0 0 −22.5 45.  7
 
 0 45. 60. 0 0 0 0 240. 0 −45. 60.  8
 
 0 0 0 −25.6 −19.2 −150. 0 0 175.6 19.2 0 9
 
0 0 0 −19.2 −14.4 0 −22.5 −45. 19.2 36.9 −45. 10
0 0 0 0 0 0 45. 60. 0 −45. 120. 11
(25.32)

25–16
 §25.5 MET-VFC ASSEMBLERS

in which global freedom numbers are annotated for convenience. Since all elements have been
processed, (25.32) is the master stiffness matrix. It eigenvalues are

[ 460.456 445.431 306.321 188.661 146.761 82.7415 74.181 25.4476 0 0 0 ] (25.33)

This has the correct rank of 11 − 3 = 8.

Remark 25.2. For storage as a skyline matrix (Chapter 26) the template for (25.32) would look like
 175.6 −19.2 0 −25.6 19.2 −150.

 36.9 45. 19.2 −14.4 0 −22.5 45. 
 −45. 
 120. 0 0 0 60. 
 51.2 0 0 0 0 −25.6 −19.2 
 
 228.8 0 −200. 0 −19.2 −14.4 
 
K= 300. 0 0 −150. 0 
 −22.5 45. 
 245. 0 0 
 −45. 60. 
 240. 0 
 175.6 19.2 0 
 
36.9 −45.
symm 120.
(25.34)
The diagonal location pointers of (25.34) are defined by the list

DLT = { 0,1,3,6,10,15,21,27,34,40,47,52 } (25.35)

Examination of the template (25.34) reveals that some additional zero entries could be removed from the
template; for example K 13 . The fact that those entries are zero is, however, fortuitous. It comes from the fact
that some elements such as the beams are aligned along x, which decouples axial and bending stiffnesses.

§25.5.8. MET-VFC Assembler Implementation

Figure 25.12 lists the Mathematica implementation of a MET-VFC assembler capable of doing the
trussed frame structure of Figure 25.11. The assembler module is invoked as

K = PlaneTrussedFrameMasterStiffness[nodxyz,eletyp,elenod,
elemat,elefab,nodfat,prcopt] (25.36)

The only new argument with respect to the MET assembler of §25.4.2 is
nodfat The Node Freedom Arrangement Table (NFAT).
The module returns the master stiffness matrix as function value.
PlaneTrussedFrameMasterStiffness uses five other modules: the element stiffness modules
PlaneBar2Stiffness and PlaneBeamColumn2Stiffness of Chapter 20, the NFMT construc-
tor NodeFreedomMapTable of Figure 25.9, the total-DOF-counter TotalFreedomCount of Fig-
ure 25.9, and the EFT constructor ElementFreedomTable of Figure 25.10.
The element fabrication list elefab has minor modifications. A plane beam-column element
requires two properties: { A,Izz }, which are the cross section area and the moment of inertia about
the local z axis. A bar element requires only the cross section area: A. For the example trussed frame
structure elefab is { { 0.02,0.004 },{ 0.02,0.004 },0.001,0.003,0.001 } in accordance with

25–17
Chapter 25: THE ASSEMBLY PROCESS

PlaneTrussedFrameMasterStiffness[nodxyz_,eletyp_,
elenod_,elemat_,elefab_,nodfat_,prcopt_]:=Module[
{numele=Length[elenod],numnod=Length[nodxyz],numdof,nodfmt,e,enl,
eftab,n,ni,nj,i,j,k,m,ncoor,Em,A,Izz,options,Ke,K},
nodfmt=NodeFreedomMapTable[nodfat];
numdof=TotalFreedomCount[nodfat]; K=Table[0,{numdof},{numdof}];
For [e=1, e<=numele, e++, enl=elenod[[e]];
If [eletyp[[e]]=="Bar2", {ni,nj}=enl;
ncoor={nodxyz[[ni]],nodxyz[[nj]]};
Em=elemat[[e]]; A=elefab[[e]]; options=prcopt;
eftab=ElementFreedomTable[enl,{110000,110000},nodfat,nodfmt,6];
Ke=PlaneBar2Stiffness[ncoor,Em,A,options] ];
If [eletyp[[e]]=="BeamCol2", {ni,nj}=enl;
ncoor={nodxyz[[ni]],nodxyz[[nj]]};
eftab=ElementFreedomTable[enl,{110001,110001},nodfat,nodfmt,6];
Em=elemat[[e]]; {A,Izz}=elefab[[e]]; options=prcopt;
Ke=PlaneBeamColumn2Stiffness[ncoor,Em,{A,Izz},options] ];
If [MemberQ[eftab,0], Print["Zero entry in eftab for element ",e];
Print["Assembly process aborted"]; Return[K]];
neldof=Length[eftab];
For [i=1, i<=neldof, i++, ii=eftab[[i]];
For [j=i, j<=neldof, j++, jj=eftab[[j]];
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]] ];
];
]; Return[K]];

Figure 25.12. A MET-VFC assembler written for the plane trussed frame structure.

nodxyz={{-4,3},{0,0},{0,3},0,{4,3}};
eletyp= {"BeamCol2","BeamCol2","Bar2","Bar2","Bar2"};
elenod= {{1,3},{3,5},{1,2},{2,3},{2,5}};
elemat= Join[Table[30000,{2}],Table[200000,{3}]];
elefab= {{0.02,0.004},{0.02,0.004},0.001,0.003,0.001};
prcopt= {True};
nodfat={110001,110000,110001,000000,110001};
K=PlaneTrussedFrameMasterStiffness[nodxyz,eletyp,
elenod,elemat,elefab,nodfat,prcopt]; K=Chop[K];
Print["Master Stiffness of Example Trussed Frame:"];
Print[K//MatrixForm];
Print["Eigs of K=",Chop[Eigenvalues[N[K]]]];

Master Stiffness of Example Trussed Frame:

175.6 −19.6 0 −25.6 19.2 −150. 0 0 0 0 0
−19.2 36.9 45. 19.2 −14.4 0 −22.5 45. 0 0 0
0 45. 120. 0 0 0 −45. 60. 0 0 0
−25.6 19.2 0 51.2 0 0 0 0 −25.6 −19.2 0
19.2 −14.4 0 0 228.8 0 −200. 0 −19.2 −14.4 0
−150. 0 0 0 0 300. 0 0 −150. 0 0
0 −22.5 −45. 0 −200. 0 245. 0 0 −22.5 45.
0 45. 60. 0 0 0 0 240. 0 −45. 60.
0 0 0 −25.6 −19.2 −150. 0 0 175.6 19.2 0
0 0 0 19.2 −14.4 0 −22.5 −45. 19.2 36.9 −45.
0 0 0 0 0 0 45. 60. 0 −45. 120.
Eigs of K= {460.456, 445.431, 306.321, 188.661, 146.761, 82.7415, 74.181, 25.4476, 0, 0, 0}

Figure 25.13. Script to test the assembler of Figure 25.12 with the structure of Figure 25.11.

25–18
 §25.6 *HANDLING MULTIFREEDOM CONSTRAINTS

(a) (b)
1 3 5
FEM idealization Beam-column Beam-column
Bar
Bar Bar
Same data as in Figure 25.11
plus MFC u y1 = u y5 2
disassembly (node 4:undefined)
MFC
(6)
(c) (1) (2) (d) MFC
3(6,7,8) 1(1,2,3) 3(6,7,8) 5(9,10,11)
1(1,2,3) 5(9,10,11)
assembly
(3) (4) (5)

2(4,5)
2(4,5)

Figure 25.14. Finite element discretization, disassembly and assembly of

trussed frame example structure with MFC u y1 = u y5 .

the data in Figure 25.11. The element material property list elemat is modified on the account that
the elastic moduli for the beam columns (made of reinforced concrete) and bars made of (steel) are
different.
Running the script listed in the top of Figure 25.13 produces the output shown in the bottom of that
figure. As can be observed this agrees with (25.32) and (25.33).

§25.6. *Handling MultiFreedom Constraints

To see the effect of imposing an MFC through the Lagrange multiplier method on the configuration of the
master stiffness matrix, suppose that the trussed frame example structure of the previous section is subjected
to the constraint that nodes 1 and 5 must move vertically by the same amount. That is,

u y1 = u y5 or u y1 − u y5 = 0. (25.37)

For assembly purposes (25.37) may be viewed as a fictitious element4 labeled as (6). See Figure (25.14).
The degrees of freedom of the assembled structure increase by one to 12. They are ordered as

Global DOF #: 1 2 3 4 5 6 7 8 9 10 11 12
DOF: u x1 u y1 θz1 u x2 u y2 u x3 u y3 θz3 u x5 u y5 θz5 λ(6) (25.38)
Node #: 1 1 1 2 2 3 3 3 5 5 5 none

The assembly of the first five elements proceeds as explained before, and produces the same master stiffness
as (25.34), except for an extra zero row and column. Processing the MFC element (6) yields the 12 × 12

4 This device should not be confused with penalty elements, which have stiffness matrices. The effect of adding a “Lagrange
multiplier element” is to append rows and columns to the master stiffness.

25–19
Chapter 25: THE ASSEMBLY PROCESS

bordered stiffness
 
175.6 −19.2 0 −25.6 19.2 −150. 0 0 0 0 0 0
 −19.2 36.9 45. 19.2 −14.4 0 −22.5 45. 0 0 0 1. 
 0 −45. 60. 0
 0 45. 120. 0 0 0 0 0 
 −25.6 19.2 0 51.2 0 0 0 0 −25.6 −19.2 0 0
 
 19.2 −14.4 0 0 228.8 0 −200. 0 −19.2 −14.4 0 0
 
 −150. 0 0 0 0 300. 0 0 −150. 0 0 0
K=  (25.39)
 0 −22.5 −45. 0 −200. 0 245. 0 0 −22.5 45. 0
 −45. 60. 0
 0 45. 60. 0 0 0 0 240. 0 
 0 0 0 −25.6 −19.2 −150. 0 0 175.6 19.2 0 0
 
 0 0 0 −19.2 −14.4 0 −22.5 −45. 19.2 36.9 −45. −1. 
 
0 0 0 0 0 0 45. 60. 0 −45. 120. 0
0 1. 0 0 0 0 0 0 0 −1. 0 0
in which the coefficients 1 and −1 associated with the MFC (25.37) end up in the last row and column of K.
There are several ways to incorporate multiplier adjunction into an automatic assembly procedure. The most
elegant ones associate the fictitious element with a special freedom signature and a reserved position in the
NFA. Being of advanced nature such schemes are beyond the scope of this Chapter.
Notes and Bibliography
The DSM assembly process for the simplest case of §25.3 is explained in any finite element text. Few texts
cover, however, the complications that arise in more general scenarios. Especially when VFCs are allowed.
Assembler implementation flavors have fluctuated since the DSM became widely accepted as standard. Variants
have been strongly influenced by limits on random access memory (RAM). Those limitations forced the use
of “out-of-core” blocked equation solvers, meaning that heavy use was made of disk auxiliary storage to store
and retrieve the master stiffness matrix in blocks. Only a limited number of blocks could reside on RAM.
Thus a straightforward element-by-element assembly loop (as in the assemblers presented here) was likely to
“miss the target” on the receiving end of the merge, forcing blocks to be read in, modified and saved. On large
problems this swapping was likely to “trash” the system with heavy I/O, bringing processing to a crawl.
One solution favored in programs of the 1965–85 period was to process all elements without assembling, saving
matrices on disk. The assembler then cycled over stiffness blocks and read in the contributing elements. With
smart asynchronous buffering and tuned-up direct access I/O such schemes were able to achieved reasonable
efficiency. However, system dependent logic quickly becomes a maintenance nightmare.
A second way out emerged by 1970: the frontal solvers referenced in Chapter 11. A frontal solver carries out
assembly, BC application and solution concurrently. Element contributions are processed in a special order
that traverses the FEM mesh as a “wavefront.” Application of displacement BCs and factorization can “trail
the wave” once no more elements are detected as contributing to a given equation. As can be expected of
trying to do too much at once, frontal solvers can be extremely sensitive to changes. One tiny alteration in the
element library over which the solver sits, and the whole thing may crumble like a house of cards.
The availability of large amounts of RAM (even on PCs and laptops) since the mid 90s, has had a happy
consequence: interweaved assembly and solver implementations, as well as convoluted matrix blocking, are
no longer needed. The assembler can be modularly separated from the solver. As a result even the most
complex assembler presented here fits in one page of text. Of course it is too late for the large scale FEM
codes that got caught in the limited-RAM survival game decades ago. Changing their assemblers and solvers
incrementally is virtually impossible. The gurus that wrote those thousands of lines of spaghetti Fortran are
long gone. The only practical way out is rewrite the whole shebang from scratch. In a commercial environment
such investment-busting decisions are unlikely.

25–20
 Exercises

Homework Exercises for Chapter 25

The Assembly Process

EXERCISE 25.1 [D:10] Suppose you want to add a six-node plane stress quadratic triangle to the MET
assembler of §25.4. Sketch how you would modify the module of Figure 25.6 for this to happen.

EXERCISE 25.2 [C:10] Exercise the module ElementFreedomTable listed in Figure 25.10 on the trussed
frame structure. Verify that it produces the last column of (25.26).

EXERCISE 25.3 [D:15] Describe the logic of module ElementFreedomTable listed in Figure 25.10. Can it
return zero entries in the EFT? If yes, give a specific example of how this can happen. Hint: read the Chapter
carefully.

EXERCISE 25.4 [A/C:30] The trussed frame structure of Figure 25.4 is reinforced with two triangular steel
plates attached as shown in Figure ?(a). The plate thickness is 1.6 mm= 0.0016 m; the material is isotropic
with E = 240000 MPa and ν = 1/3. Each reinforcing plate is modeled with a single plane stress 3-node
linear triangle. The triangles are numbered (6) and (7), as illustrated in Figure E25.1(c). Compute the master
stiffness matrix K of the structure.5

(b) y
y
1 Beam-column 3 Beam-column 5
(a) FEM idealization
Plate properties: Bar
Bar
E =240000 MPa, ν=1/3 Plate
h =1.6 mm = 0.0016 m, Plate Bar
x 2
x
Other dimensions & properties
same as in the trussed frame of Fig. 25.11 disassembly (node 4:undefined)

(c) (1) (2) (d)

1(1,2,3) 3(6,7,8) 1(1,2,3) 3(6,7,8) 5(9,10,11)
5(9,10,11)
(6) (7) assembly
(3) (4) (5)

2(4,5) 2(4,5)

Figure E25.1. Plate-reinforced trussed frame structure for Exercise 25.4.

This exercise may be done through Mathematica. For this download Notebook ExampleAssemblers.nb from
this Chapter index, and complete Cell 4 by writing the assembler.6 The plate element identifier is "Trig3".
The driver script to run the assembler is also provided in Cell 4 (blue text).

5 This structure would violate the compatibility requirements stated in Chapter 19 unless the beams are allowed to deflect
laterally independently of the plates. This is the fabrication actually sketched in Figure E25.1(a).
6 Cells 1–3 contain the assemblers presented in §25.3, §25.4 and §25.5, respectively, which may be used as guides. The
stiffness modules for the three element types used in this structure are available in Cells 2 and 3 and may be reused for
this Exercise.

25–21
Chapter 25: THE ASSEMBLY PROCESS

When reusing the assembler of Cell 3 as a guide, please do not remove the internal Print commands that show
element information (red text). Those come in handy for debugging. Please keep that printout in the returned
homework to help the grader.
Another debugging hint: check that the master stiffness (25.32) is obtained if the plate thickness, called hplate
in the driver script, is temporarily set to zero.
Target:
 337.6 −19.2 0 −25.6 91.2 −312. −72. 0 0 0 0

 −19.2 90.9 45. 91.2 −14.4 −72. −76.5 45. 0 0 0 
 0 −45. 0 
 45. 120. 0 0 0 60. 0 0 
 −25.6 91.2 0 243.2 0 −192. 0 0 −25.6 −91.2 0 
 
 91.2 −14.4 0 0 804.8 0 −776. 0 −91.2 −14.4 0 
 
K =  −312. −72. 0 −192. 0 816. 0 0 −312. 72. 0  (E25.1)
 −72. −76.5 −45. −776. −76.5 45. 
 0 0 929. 0 72. 
 0 −45. 60. 
 45. 60. 0 0 0 0 240. 0 
 0 0 0 −25.6 −91.2 −312. 72. 0 337.6 19.2 0 
 
0 0 0 −91.2 −14.4 72. −76.5 −45. 19.2 90.9 −45.
0 0 0 0 0 0 45. 60. 0 −45. 120.

EXERCISE 25.5 [A:30] §25.5 does not explain how to construct the NFAT from the input data. (In the trussed
frame example script, nodfat was set up by inspection, which is OK only for small problems.) Explain how
this table could be constructed automatically if the EFS of each element in the model is known. Note: the
logic is far from trivial.

25–22
 26
.

Solving
FEM Equations

26–1
Chapter 26: SOLVING FEM EQUATIONS 26–2

TABLE OF CONTENTS

Page
§26.1. Motivation for Sparse Solvers 26–3
§26.1.1. The Curse of Fullness . . . . . . . . . . . . . . 26–3
§26.1.2. The Advantages of Sparsity . . . . . . . . . . . . 26–4
§26.2. Sparse Solution of Stiffness Equations 26–5
§26.2.1. Skyline Storage Format . . . . . . . . . . . . . . 26–5
§26.2.2. Factorization . . . . . . . . . . . . . . . . . 26–6
§26.2.3. Solution . . . . . . . . . . . . . . . . . . . 26–6
§26.2.4. Treating MFCs with Lagrange Multipliers . . . . . . . 26–6
§26.3. A SkySolver Implementation 26–7
§26.3.1. Skymatrix Representation . . . . . . . . . . . . . 26–7
§26.3.2. *Skymatrix Factorization . . . . . . . . . . . . . 26–8
§26.3.3. *Solving for One or Multiple RHS . . . . . . . . . . 26–11
§26.3.4. *Matrix-Vector Multiply . . . . . . . . . . . . . 26–13
§26.3.5. *Printing and Mapping . . . . . . . . . . . . . . 26–14
§26.3.6. *Reconstruction of SkyMatrix from Factors . . . . . . . 26–15
§26.3.7. *Miscellaneous Utilities . . . . . . . . . . . . . . 26–16
§26. Exercises . . . . . . . . . . . . . . . . . . . . . . 26–20

26–2
26–3 §26.1 MOTIVATION FOR SPARSE SOLVERS

§26.1. Motivation for Sparse Solvers

In the Direct Stiffness Method (DSM) of finite element analysis, the element stiffness matrices and
consistent nodal force vectors are immediately assembled to form the master stiffness matrix and
master force vector, respectively, by the process called merge. The assembly process is described in
Chapter 25. For simplicity the description that follows assumes that no MultiFreedom Constraints
(MFCs) are present. The end result of the assembly process are the master stiffness equations

Ku = f (26.1)

where K is the master stiffness matrix, f the vector of node forces and u the vector or node
displacements. Upon imposing the displacement boundary conditions, the system (26.1) is solved
for the unknown node displacements. The solution concludes the main phase of DSM computations.
In practical applications the order of the stiffness system (26.1) can be quite large. Systems of
order 1000 to 10000 are routinely solved in commercial software. Larger ones (say up to 100000
equations) are not uncommon and even millions of equations are being solved on suoercomputers.
Presently the record is about 50 million equations on parallel computers.1
In linear FEM analysis the cost of solving this system of equations rapidly overwhelms other
computational phases. Much attention has therefore given to matrix processing techniques that
economize storage and solution time by taking advantage of the special structure of the stiffness
matrix.
The master force vector is stored as a conventional one-dimensional array of length equal to the
number N of degrees of freedom. This storage arrangement presents no particular difficulties even
for very large problem sizes.2 Handling the master stiffness matrix, however, presents computational
difficulties.

§26.1.1. The Curse of Fullness

If K is stored and processed as if it were a full matrix, the storage and processing time resources
rapidly becomes prohibitive as N increases. This is illustrated in Table 26.1, which summarizes
the storage and factor-time requirements for orders N = 104 , 105 and 106 .3
As regards memory needs, a full square matrix stored without taking advantage of symmetry,
requires storage for N 2 entries. If each entry is an 8-byte, double precision floating-point number,
the required storage is 8N 2 bytes. Thus, a matrix of order N = 104 would require 8 × 108 bytes or
800 MegaBytes (MB) for storage.
For large N the solution of (26.1) is dominated by the factorization of K, an operation discussed in
§26.2. This operation requires approximately N 3 /6 floating point operation units. [A floating-point
operation unit is conventionally defined as a (multiply,add) pair plus associated indexing and data
movement operations.] Now a fast workstation can typically do 107 of these operations per second,

1 For fluid equations solved by fluid-voulume methods the number is over 100 million.
2 A force of displacement vector of, say, 1M equations uses 8MB for double precision storage. In these days of GB RAMs,
that is a modest amount.
3 The factor times given in that table reflect 1998 computer technology. To update to 2003, divide times by 10 to 20.

26–3
Chapter 26: SOLVING FEM EQUATIONS 26–4

Table 26.1 Storage & Solution Time for a Fully-Stored Stiffness Matrix

Matrix Storage Factor Factor time Factor time

order N (double prec) op.units workstation supercomputer
(FLOPS) (or fast PC)

104 800 MB 1012 /6 3 hrs 2 min

105 80 GB 1015 /6 4 mos 30 hrs
106 8 TB 1018 /6 300 yrs 3 yrs

Table 26.2 Storage & Solution Time for a Skyline

√ Stored Stiffness Matrix
Assuming B = N

Matrix Storage Factor Factor time Factor time

order N (double prec) op.units workstation supercomputer
(FLOPS) (or fast PC)

104 8 MB 108 /2 5 sec 0.05 sec

105 240 MB 1010 /2 8 min 5 sec
106 8 GB 1012 /2 15 hrs 8 min

whereas a supercomputer may be able to sustain 109 or more. These times assume that the entire
matrix is kept in RAM; for otherwise the elapsed time may increase by factors of 10 or more due to
I/O transfer operations. The elaspsed timed estimated given in Table 26.1 illustrate that for present
computer resources, orders above 104 would pose significant computational difficulties.

§26.1.2. The Advantages of Sparsity

Fortunately a very high percentage of the entries of the master stiffness matrix K are zero. Such
matrices are call sparse. There are clever programming techniques that take advantage of sparsity
that fit certain patterns. Although a comprehensive coverage of such techniques is beyond the scope
of this course, we shall concentrate on a particular form of sparse scheme that is widely use in FEM
codes: skyline storage. This scheme is simple to understand, manage and implement, while cutting
storage and processing times by orders of magnitude as the problems get larger.
The skyline storage format is a generalization of its widely used predecessor called the band storage
scheme. A matrix stored in accordance with the skyline format will be called a skymatrix for short.
Only symmetric skymatrices will bve considered here, since the stiffness matrices in linear FEM
are symmetric.
If a skymatrix of order N can be stored in S memory locations, the ratio B = S/N is called the
mean bandwidth. If the entries are, as usual, 8-byte double-precision floating-point numbers, the
storage requirement is 8N B bytes. The factorization of a skymatrix requires approximately 12 N B 2
√
floating-point operation units. In two-dimensional problems B is of the order of N . Under this
assumption, storage requirements and estimated factorization times for N = 104 , N = 105 and

26–4
26–5 §26.2 SPARSE SOLUTION OF STIFFNESS EQUATIONS

N = 106 are reworked in Table 26.2. It is seen that by going from full to skyline storage significant
reductions in computer resources have been achieved. For example, now N = 104 is easy on a
workstation and trivial on a supercomputer. Even a million equations do not look far-fetched on a
supercomputer as long as enough memory is available.4
In preparation for assembling K as a skymatrix one has to set up several auxiliary arrays related to
nodes and elements. These auxiliary arrays are described in the previous Chapter. Knowledge of
that material is useful for understanding the following description.

§26.2. Sparse Solution of Stiffness Equations

§26.2.1. Skyline Storage Format

The skyline storage arrangement for K is best illustrated through a simple example. Consider the
6 × 6 stiffness matrix
 
K 11 0 K 13 0 0 K 16
 K 22 0 K 24 0 0 
 
 K 33 K 34 0 0 
K=  (26.2)
 K 44 0 K 46 
 
K 55 K 56
symm K 66

Since the matrix is symmetric only one half, the upper triangle in the above display, need to be
shown.
Next we define the envelope of K as follows. From each diagonal entry move up the corresponding
column until the last nonzero entry is found. The envelope separates that entry from the rest of the
upper triangle. The remaining zero entries are conventionally removed:
 
K 11 K 13 K 16
 K 22 0 K 24 0 
 
 K 33 K 34 0 
K=  (26.3)
 K 44 K 46 
 
K 55 K 56
symm K 66

What is left constitute the skyline profile of skyline template of the matrix. A sparse matrix that
can be profitably stored in this form is called a skymatrix for brevity. Notice that the skyline profile
may include zero entries. During the factorization step discussed below these zero entries will in
general become nonzero, a phenomenon that receives the name fill-in.

4 On a PC or laptop with a 32-bit processor, there is a RAM “hard barrier” because of hardware addressing limitations.
That is typically 2GB. As PCs migrate to 64-bit processors over the next 10 years, the barrier disappears. For example
the high-end Mac workstation can presently be expanded up to 16GB RAM. The new limits are primarily dictated by
power comsumption, circuitry and cooling considerations. A distributed (Beowulf) network is a practical solution to go
beyond the individual machine limits.

26–5
Chapter 26: SOLVING FEM EQUATIONS 26–6

The key observation is that only the entries in the skyline template need to be stored, because
fill-in in the factorization process will not occur outside the envelope. To store these entries it is
convenient to use a one-dimensional skyline array:

s: [ K 11 , K 22 , K 13 , 0, K 33 , K 24 , K 34 , K 44 , K 55 , K 16 , 0, 0, K 46 , K 56 , K 66 ] (26.4)

This array is complemented by a (N + 1) integer array p that contains addresses of diagonal

locations. The array has N + 1 entries. The (i + 1)th entry of p has the location of the i th diagonal
entry of K in s. For the example matrix:

p: [ 0, 1, 2, 5, 8, 9, 15 ] (26.5)

In the previous Chapter, this array was called the Global Skyline Diagonal Location Table, or GSDLT.
Equations for which the displacement component is prescribed are identified by a negative diagonal
location value. For example if u 3 and u 5 are prescribed displacement components in the test
example, then
p : [ 0, 1, 2, −5, 8, −9, 15 ] (26.6)

Remark 26.1. In Fortran it is convenient to dimension the diagonal location array as p(0:n) so that indexing
begins at zero. In C this is the standard indexing.

§26.2.2. Factorization
The stiffness equations (26.1) are solved by a direct method that involves two basic phases: factor-
ization and solution.
In the first stage, the skyline-stored symmetric stiffness matrix is factored as

K = LDU = LDLT = UT DU, (26.7)

where L is a unit lower triangular matrix, D is a nonsingular diagonal matrix, and U and L are the
transpose of each other. The original matrix is overwritten by the entries of D−1 and U; details may
be followed in the program implementation. No pivoting is used in the factorization process. This
factorization is carried out by Mathematica module SymmSkyMatrixFactor, which is described
later in this Chapter.

§26.2.3. Solution
Once K has been factored, the solution u for a given right hand side f is obtained by carrying out
three stages:

Forward reduction : Lz = f, (26.8)

Diagonal scaling : Dy = z, (26.9)
Back substitution : Uu = y, (26.10)

where y and z are intermediate vectors. These stages are carried out by Mathematica modules
SymmSkyMatrixVectorSolve, which is described later.

26–6
26–7 §26.3 A SKYSOLVER IMPLEMENTATION

§26.2.4. Treating MFCs with Lagrange Multipliers

In Mathematica implementations of FEM, MultiFreedom Constraints (MFCs) are treated with
Lagrange multipliers. There is one multiplier for each constraint. The multipliers are placed at the
end of the solution vector.
Specifically, let the nummul>0 MFCs be represented in matrix form as Cu = g, where C and g are
given, and let the nummul multipliers be collected in a vector λ. The multiplier-augmented master
stiffness equations are   
K CT u f
= (26.11)
C 0 λ g
or
Ax = b. (26.12)
where the symmetric matrix A, called a stiffness-bordered matrix, is of order numdof+nummul.
The stiffness bordered matrix is also stored in skyline form, and the previous solution procedure
applies, as long as the skyline array is properly constructed as described in the previous Chapter.
The main difference with respect to the no-MFC case is that, because of the configuration (26.11),
A can no longer be positive definite. In principle pivoting should be used during the factorization
of A to forestall possible numerical instabilities. Pivoting can have a detrimental effect on solution
efficiency because entries can move outside of the skyline template. However, by placing the λ at
the end such difficulties will not be encountered if K is positive definite, and the constraints are
linearly independent (that is, C has full rank). Thus pivoting is not necessary.
§26.3. A SkySolver Implementation
The remaining sections of this revised Chapter describe a recent implementation of the skyline
solver and related routines in Mathematica. This has been based on similar Fortran codes used
since 1967.

§26.3.1. Skymatrix Representation

In what follows the computer representation in Mathematica of a symmetric skymatrix will be
generally denoted by the symbol S. Such a representation consists of a list of two numeric objects:

S = { p, s } (26.13)

Here p=GSDLT is the Global Skyline Diagonal Location Table introduced in §11.6, and s is the array
of skymatrix entries, arranged as described in the previous section. This array usually consists of
floating-point numbers, but it may also contain exact integers or factions, or even symbolic entries.
For example, suppose that the numerical entries of the 6 × 6 skymatrix (26.10) are actually
 
11 13 16
 22 0 24 0 
 
 33 34 0 
K=  (26.14)
 44 46 
 
55 56
symm 66

26–7
Chapter 26: SOLVING FEM EQUATIONS 26–8

Cell 26.1 Factorization of a Symmetric SkyMatrix

SymmSkyMatrixFactor[S_,tol_]:= Module[
{p,a,fail,i,j,k,l,m,n,ii,ij,jj,jk,jmj,d,s,row,v},
row=SymmSkyMatrixRowLengths[S]; s=Max[row];
{p,a}=S; n=Length[p]-1; v=Table[0,{n}]; fail=0;
Do [jj=p[[j+1]]; If [jj<0|row[[j]]==0, Continue[]]; d=a[[jj]];
jmj=Abs[p[[j]]]; jk=jj-jmj;
Do [i=j-jk+k; v[[k]]=0; ii=p[[i+1]];
If [ii<0, Continue[]]; m=Min[ii-Abs[p[[i]]],k]-1;
ij=jmj+k; v[[k]]=a[[ij]];
v[[k]]-=Take[a,{ii-m,ii-1}].Take[v,{k-m,k-1}];
a[[ij]]=v[[k]]*a[[ii]],
{k,1,jk-1}];
d-=Take[a,{jmj+1,jmj+jk-1}].Take[v,{1,jk-1}];
If [Abs[d]<tol*row[[j]], fail=j; a[[jj]]=Infinity; Break[] ];
a[[jj]]=1/d,
{j,1,n}];
Return[{{p,a},fail}]
];

SymmSkyMatrixRowLengths[S_]:= Module[
{p,a,i,j,n,ii,jj,m,d,row},
{p,a}=S; n=Length[p]-1; row=Table[0,{n}];
Do [ii=p[[i+1]]; If [ii<0, Continue[]]; m=ii-i; row[[i]]=a[[ii]]^2;
Do [If [p[[j+1]]>0, d=a[[m+j]]^2; row[[i]]+=d; row[[j]]+=d],
{j,Max[1,Abs[p[[i]]]-m+1],Min[n,i]-1}],
{i,1,n}]; Return[Sqrt[row]];
];

Its Mathematica representation, using the symbols (26.13) is

p= { 0,1,2,5,8,9,15 };
s= { 11,22,13,0,33,24,34,44,55,16,0,0,46,56,66 }; (26.15)
S= { p, s };

or more directly

S={ { 0,1,2,5,8,9,15 },{ 11,22,13,0,33,24,34,44,55,16,0,0,46,56,66 } }; (26.16)

[The remaining sections on details of skyline processing logic, marked with a *, will not be covered
in class. They are intended for a more advanced course.]

26–8
26–9 §26.3 A SKYSOLVER IMPLEMENTATION

Cell 26.2 Factorization Test Input

ClearAll[n]; n=5; SeedRandom[314159];

p=Table[0,{n+1}]; Do[p[[i+1]]=p[[i]]+
Max[1,Min[i,Round[Random[]*i]]],{i,1,n}];
a=Table[1.,{i,1,p[[n+1]]}];
Print["Mean Band=",N[p[[n+1]]/n]];
S={p,a};
Sr=SymmSkyMatrixLDUReconstruct[S];
Print["Reconstructed SkyMatrix:"]; SymmSkyMatrixLowerTrianglePrint[Sr];
SymmSkyMatrixLowerTriangleMap[Sr];
Print["eigs=",Eigenvalues[SymmSkyMatrixConvertToFull[Sr]]];
x=Table[{N[i],3.,(-1)^i*N[n-i]},{i,1,n}];
Print["Assumed x=",x];
b=SymmSkyMatrixColBlockMultiply[Sr,x];
(*x=Transpose[x]; b=SymmSkyMatrixRowBlockMultiply[Sr,x];*)
Print["b=Ax=",b];
Print[Timing[{F,fail}=SymmSkyMatrixFactor[Sr,10.^(-12)]]];
If [fail!=0, Print["fail=",fail]; Abort[]];
Print["F=",F]; Print["fail=",fail];
Print["Factor:"];
SymmSkyMatrixLowerTrianglePrint[F];
x=SymmSkyMatrixColBlockSolve[F,b];
(*x=SymmSkyMatrixRowBlockSolve[F,b];*)
Print["Computed x=",x//InputForm];

§26.3.2. *Skymatrix Factorization

Module SymmSkyMatrixFactor, listed in Cell 26.1, factors a symmetric skymatrix into the product LDU
where L is the transpose of U. No pivoting is used. The module is invoked as
{ Sf,fail } = SymmSkyMatrixFactor[S,tol]

The input arguments are

S The skymatrix to be factored, stored as the two-object list { p,s }; see previous subsection.
tol Tolerance for singularity test. The appropriate value of tol depends on the kind of skymatrix
entries stored in s.
If the skymatrix entries are floating-point numbers handled by default in double precision
arithmetic, tol should be set to 8× or 10× the machine precision in that kind of arithmetic.
The factorization aborts if, when processing the j-th row, d j ≤ tol ∗ r j , where d j is the
computed j th diagonal entry of D, and r j is the Euclidean norm of the j th skymatrix row.
If the skymatrix entries are exact (integers, fractions or symbols), tol should be set to zero.
In this case exact singularity is detectable, and the factorization aborts only on that condition.
The outputs are:
Sf If fail is zero on exit, Sf is the computed factorization of S. It is a two-object list { p, du

26–9
Chapter 26: SOLVING FEM EQUATIONS 26–10

Cell 26.3 Output from Program of Cells 26.1 and 26.2

Mean Band=1.6
Reconstructed SkyMatrix:

Col 1 Col 2 Col 3 Col 4 Col 5

Row 1 1.0000
Row 2 1.0000
Row 3 1.0000 2.0000
Row 4 1.0000
Row 5 1.0000 1.0000 3.0000

1 2 3 4 5

1 +
2 +
3 + +
4 +
5 + + +
eigs={3.9563, 2.20906, 1., 0.661739, 0.172909}
Assumed x={{1., 3., -4.}, {2., 3., 3.}, {3., 3., -2.}, {4., 3., 1.},

{5., 3., 0}}

b=Ax={{1., 3., -4.}, {5., 6., 1.}, {13., 12., -1.}, {9., 6., 1.},

{22., 15., -1.}}

{0.0666667 Second, {{{0, 1, 2, 4, 5, 8},

{1., 1., 1., 1., 1., 1., 1., 1.}}, 0}}

F={{0, 1, 2, 4, 5, 8}, {1., 1., 1., 1., 1., 1., 1., 1.}}
fail=0
Factor:

Col 1 Col 2 Col 3 Col 4 Col 5

Row 1 1.0000
Row 2 1.0000
Row 3 1.0000 1.0000
Row 4 1.0000
Row 5 1.0000 1.0000 1.0000
Computed x={{1., 3., -4.}, {2., 3., 3.}, {3., 3., -2.}, {4., 3., 1.},

{5., 3., 0.}}

}, where du stores the entries of D−1 in the diagonal locations, and of U in its strict upper
triangle.
fail A singularity detection indicator. A zero value indicates that no singularity was detected. If
fail returns j>0, the factorization was aborted at the j-th row. In this case Sf returns the
aborted factorization with ∞ stored in d j .

26–10
26–11 §26.3 A SKYSOLVER IMPLEMENTATION

Cell 26.4 Solving for a Single RHS

SymmSkyMatrixVectorSolve[S_,b_]:= Module[
{p,a,n,i,j,k,m,ii,jj,bi,x},
{p,a}=S; n=Length[p]-1; x=b;
If [n!=Length[x], Print["Inconsistent matrix dimensions in",
" SymmSkyMatrixVectorSolve"]; Return[Null]];
Do [ii=p[[i+1]];If [ii>=0, Continue[]]; ii=-ii; k=i-ii+Abs[p[[i]]]+1;
bi=x[[i]]; If [bi==0, Continue[]];
Do [jj=p[[j+1]], If [jj<0, Continue[]];
m=j-i; If [m<0, x[[j]]-=a[[ii+m]]*bi; Break[]];
ij=jj-m; If [ij>Abs[p[[j]]], x[[j]]-=a[[ij]]*bi],
{j,k,n}],
{i,1,n}];
Do [ii=p[[i+1]]; If [ii<0, x[[i]]=0; Continue[]];
imi=Abs[p[[i]]]; m=ii-imi-1;
x[[i]]-=Take[a,{imi+1,imi+m}].Take[x,{i-m,i-1}],
{i,1,n}];
Do [ii=Abs[p[[i+1]]]; x[[i]]*=a[[ii]], {i,1,n}];
Do [ii=p[[i+1]]; If [ii<0, x[[i]]=b[[i]]; Continue[]];
m=ii-Abs[p[[i]]]-1;
Do [ x[[i-j]]-=a[[ii-j]]*x[[i]], {j,1,m}],
{i,n,1,-1}];
Return[x]
];

A test of SymmSkyMatrixFactor on the matrix (26.14) is shown in Cells 26.2 and 26.3. The modules that
print and produce skyline maps used in the test program are described later in this Chapter.

§26.3.3. *Solving for One or Multiple RHS

Module SymmSkyMatrixVectorSolve, listed in Cell 26.4, solves the linear system Ax = b for x, following
the factorization of the symmetric skymatrix A by SymmSkyMatrixFactor. The module is invoked as
x = SymmSkyMatrixVectorSolve[Sf,b]
The input arguments are
Sf The factored matrix returned by SymmSkyMatrixFactor.
b The right-hand side vector to be solved for, stored as a single-level (one dimensional) list. If
the i-th entry of x is prescribed, the known value must be supplied in this vector.
The outputs are:
x The computed solution vector, stored as a single-level (one-dimensional) list. Prescribed
solution components return the value of the entry in b.
Sometimes it is necessary to solve linear systems for multiple (m > 1) right hand sides. One way to do that is
to call SymmSkyMatrixVectorSolve repeatedly. Alternatively, if m right hand sides are collected as columns
of a rectangular matrix B, module SymmSkyMatrixColBlockSolve may be invoked as

26–11
Chapter 26: SOLVING FEM EQUATIONS 26–12

Cell 26.5 Solving for a Block of Righ Hand Sides

SymmSkyMatrixColBlockSolve[S_,b_]:= Module[
{p,a,n,nrhs,i,j,k,m,r,ii,jj,bi,x},
{p,a}=S; n=Length[p]-1; x=b;
If [n!=Dimensions[x][[1]], Print["Inconsistent matrix dimensions in",
" SymmSkyMatrixBlockColSolve"]; Return[Null]]; nrhs = Dimensions[x][[2]];
Do [ii=p[[i+1]];If [ii>=0, Continue[]]; ii=-ii; k=i-ii+Abs[p[[i]]]+1;
Do [bi=x[[i,r]]; If [bi==0, Continue[]];
Do [jj=p[[j+1]], If [jj<0, Continue[]];
m=j-i; If [m<0,x[[j,r]]-=a[[ii+m]]*bi; Break[]];
ij=jj-m; If [ij>Abs[p[[j]]], x[[j,r]]-=a[[ij]]*bi],
{j,k,n}],
{r,1,nrhs}],
{i,1,n}];
Do [ii=p[[i+1]]; If [ii<0, Do[x[[i,r]]=0,{r,1,nrhs}];Continue[]];
imi=Abs[p[[i]]]; m=ii-imi-1;
Do [ Do [ x[[i,r]]-=a[[imi+j]]*x[[i-m+j-1,r]], {j,1,m}], {r,1,nrhs}],
{i,1,n}];
Do [ii=Abs[p[[i+1]]]; Do[x[[i,r]]*=a[[ii]], {r,1,nrhs}], {i,1,n}];
Do [ii=p[[i+1]]; If [ii<0, Do[x[[i,r]]=b[[i,r]],{r,1,nrhs}];Continue[]];
m=ii-Abs[p[[i]]]-1;
Do [ Do [ x[[i-j,r]]-=a[[ii-j]]*x[[i,r]], {j,1,m}], {r,1,nrhs}],
{i,n,1,-1}];
Return[x]
];
SymmSkyMatrixRowBlockSolve[S_,b_]:= Module[
{p,a,n,nrhs,i,j,k,m,r,ii,jj,bi,x},
{p,a}=S; n=Length[p]-1; x=b;
If [n!=Dimensions[x][[2]], Print["Inconsistent matrix dimensions in",
" SymmSkyMatrixBlockRowSolve"]; Return[Null]]; nrhs = Dimensions[x][[1]];
Do [ii=p[[i+1]];If [ii>=0, Continue[]]; ii=-ii; k=i-ii+Abs[p[[i]]]+1;
Do [bi=x[[r,i]]; If [bi==0, Continue[]];
Do [jj=p[[j+1]], If [jj<0, Continue[]];
m=j-i; If [m<0,x[[j,r]]-=a[[ii+m]]*bi; Break[]];
ij=jj-m; If [ij>Abs[p[[j]]], x[[r,j]]-=a[[ij]]*bi],
{j,k,n}],
{r,1,nrhs}],
{i,1,n}];
Do [ii=p[[i+1]]; If [ii<0, Do[x[[r,i]]=0,{r,1,nrhs}];Continue[]];
imi=Abs[p[[i]]]; m=ii-imi-1;
Do [ Do [ x[[r,i]]-=a[[imi+j]]*x[[r,i-m+j-1]], {j,1,m}], {r,1,nrhs}],
{i,1,n}];
Do [ii=Abs[p[[i+1]]]; Do[x[[r,i]]*=a[[ii]], {r,1,nrhs}], {i,1,n}];
Do [ii=p[[i+1]]; If [ii<0, Do[x[[r,i]]=b[[r,i]],{r,1,nrhs}];Continue[]];
m=ii-Abs[p[[i]]]-1;
Do [ Do [ x[[r,i-j]]-=a[[ii-j]]*x[[r,i]], {j,1,m}], {r,1,nrhs}],
{i,n,1,-1}];
Return[x]
];

26–12
26–13 §26.3 A SKYSOLVER IMPLEMENTATION

Cell 26.6 Multiplying Skymatrix by Individual Vector

SymmSkyMatrixVectorMultiply[S_,x_]:= Module[
{p,a,n,i,j,k,m,ii,b},
{p,a}=S; n=Length[p]-1;
If [n!=Length[x], Print["Inconsistent matrix dimensions in",
" SymmSkyMatrixVectorMultiply"]; Return[Null]];
b=Table[a[[ Abs[p[[i+1]]] ]]*x[[i]], {i,1,n}];
Do [ii=Abs[p[[i+1]]]; m=ii-Abs[p[[i]]]-1; If [m<=0,Continue[]];
b[[i]]+=Take[a,{ii-m,ii-1}].Take[x,{i-m,i-1}];
Do [b[[i-k]]+=a[[ii-k]]*x[[i]],{k,1,m}],
{i,1,n}];
Return[b]
];

(*
ClearAll[n]; n=10; SeedRandom[314159];
p=Table[0,{n+1}]; Do[p[[i+1]]=p[[i]]+
Max[1,Min[i,Round[Random[]*i]]],{i,1,n}];
a=Table[1.,{i,1,p[[n+1]]}];
Print["Mean Band=",N[p[[n+1]]/n]];
S={p,a};
Sr=SymmSkyMatrixLDUReconstruct[S];
Print["Reconstructed SkyMatrix:"]; SymmSkyMatrixLowerTrianglePrint[Sr];
SymmSkyMatrixLowerTriangleMap[Sr];
x=Table[1.,{i,1,n}];
b=SymmSkyMatrixVectorMultiply[Sr,x];
Print["b=Ax=",b];*)

X = SymmSkyMatrixVectorSolve[Sf,B]
to provide the solution X of SX = B. This module is listed in Cell 26.5. The input arguments and function
returns have the same function as those described for SymmSkyMatrixVectorSolve. The main difference
is that B and X are matrices (two-dimensional lists) with the righ-hand side and solution vectors as columns.
There is a similar module SymmSkyMatrixRowBlockSolve, notlisted here, which solves for multiple right
hand sides stored as rows of a matrix.

§26.3.4. *Matrix-Vector Multiply

For various applications it is necessary to form the matrix-vector product
b = Sx (26.17)
where S is a symmetric skymatrix and x is given.
This is done by module SymmSkyMatrixVectorMultiply, which is listed in Cell 26.6. Its arguments are the
skymatrix S and the vector x. The function returns Sx in b.
Module SymmSkyMatrixColBlockMultiply implements the multiplication by a block of vectors stored as
columns of a rectangular matrix X:
B = SX (26.18)

26–13
Chapter 26: SOLVING FEM EQUATIONS 26–14

Cell 26.7 Multiplying Skymatrix by Vector Block

SymmSkyMatrixColBlockMultiply[S_,x_]:= Module[
{p,a,n,nrhs,i,j,k,m,r,ii,aij,b},
{p,a}=S; n=Length[p]-1;
If [n!=Dimensions[x][[1]], Print["Inconsistent matrix dimensions in",
" SymmSkyMatrixColBlockMultiply"]; Return[Null]];
nrhs = Dimensions[x][[2]]; b=Table[0,{n},{nrhs}];
Do [ii=Abs[p[[i+1]]]; m=ii-Abs[p[[i]]]-1;
Do [b[[i,r]]=a[[ii]]*x[[i,r]], {r,1,nrhs}];
Do [j=i-k; aij=a[[ii-k]]; If [aij==0, Continue[]];
Do [b[[i,r]]+=aij*x[[j,r]]; b[[j,r]]+=aij*x[[i,r]], {r,1,nrhs}],
{k,1,m}],
{i,1,n}];
Return[b]
];

SymmSkyMatrixRowBlockMultiply[S_,x_]:= Module[
{p,a,n,nrhs,i,j,k,m,r,ii,aij,b},
{p,a}=S; n=Length[p]-1;
If [n!=Dimensions[x][[2]], Print["Inconsistent matrix dimensions in",
" SymmSkyMatrixRowBlockMultiply"]; Return[Null]];
nrhs = Dimensions[x][[1]]; b=Table[0,{nrhs},{n}];
Do [ii=Abs[p[[i+1]]]; m=ii-Abs[p[[i]]]-1;
Do [b[[r,i]]=a[[ii]]*x[[r,i]], {r,1,nrhs}];
Do [j=i-k; aij=a[[ii-k]]; If [aij==0, Continue[]];
Do [b[[r,i]]+=aij*x[[r,j]]; b[[r,j]]+=aij*x[[r,i]], {r,1,nrhs}],
{k,1,m}],
{i,1,n}];
Return[b]
];

This module is listed in Cell 26.7. Its arguments are the skymatrix S and the rectangular matrix X. The function
returns SX in B.
There is a similar module SymmSkyMatrixRowBlockMultiply, also listed in Cell 26.7, which postmultiplies
a vector block stored as rows.

§26.3.5. *Printing and Mapping

Module SymmSkyMatrixUpperTrianglePrint, listed in Cell 26.8, prints a symmetric skymatrix in upper

triangle form. Is is invoked as

SymmSkyMatrixUpperTrianglePrint[S]

where S is the skymatrix to be printed. For an example of use see Cells 262-3.
The print format resembles the configuration depicted in Section 26.1. This kind of print is useful for program

26–14
26–15 §26.3 A SKYSOLVER IMPLEMENTATION

development and debugging although of course it should not be attempted with a very large matrix.5

There is a similar module called SymmSkyMatrixLowerTrianglePrint, which displays the skymatrix entries
in lower triangular form. This module is also listed in Cell 26.8.

Sometimes one is not interested in the actual values of the skymatrix entries but only on how the skyline
template looks like. Such displays, called maps, can be done with just one symbol per entry. Module
SymmSkyMatrixUpperTriangleMap, listed in Cell 26.9, produces a map of its argument. It is invoked as
SymmSkyMatrixUpperTriangleMap[S]

The entries within the skyline template are displayed by symbols +, - and 0, depending on whether the value
is positive, negative or zero, respectively. Entries outside the skyline template are blank.
As in the case of the print module, there is module SymmSkyMatrixLowerTriangleMap which is also listed
in Cell 26.9.

§26.3.6. *Reconstruction of SkyMatrix from Factors

In testing factorization and solving modules it is convenient to have modules that perform the “inverse process”
of the factorization. More specifically, suppose that U, D (or D−1 ) are given, and the problem is to reconstruct
the skymatrix that have them as factors:

S = LDU, or S = LD−1 U (26.19)

in which L = U . Modules SymmSkyMatrixLDUReconstruction and
T

SymmSkyMatrixLDinvUReconstruction perform those operations. These modules are listed in Cell 26.10.
Their argument is a factored form of S: U and D in the first case, and U and D−1 in the second case.

5 Computer oriented readers may notice that the code for the printing routine is substantially more complex than those
of the computational modules. This is primarily due to the inadequacies of Mathematica in handling tabular format
output. The corresponding Fortran or C implementations would be simpler because those languages provide much better
control over low-level display.

26–15
Chapter 26: SOLVING FEM EQUATIONS 26–16

Cell 26.8 Skymatrix Printing

SymmSkyMatrixLowerTrianglePrint[S_]:= Module[
{p,a,cycle,i,ii,ij,it,j,jj,j1,j2,jref,jbeg,jend,jt,kcmax,kc,kr,m,n,c,t},
{p,a}=S; n=Dimensions[p][[1]]-1; kcmax=5; jref=0;
Label[cycle]; Print[" "];
jbeg=jref+1; jend=Min[jref+kcmax,n]; kc=jend-jref;
t=Table[" ",{n-jref+1},{kc+1}];
Do [If [p[[j+1]]>0,c=" ",c="*"];
t[[1,j-jref+1]]=StringJoin[c,"Col",ToString[PaddedForm[j,3]]],
{j,jbeg,jend}]; it=1;
Do [ii=Abs[p[[i+1]]]; m=ii-Abs[p[[i]]]-1; j1=Max[i-m,jbeg];j2=Min[i,jend];
kr=j2-j1+1; If [kr<=0, Continue[]]; If [p[[i+1]]>0,c=" ",c="*"];
it++; t[[it,1]]=StringJoin[c,"Row",ToString[PaddedForm[i,3]]];
jt=j1-jbeg+2; ij=j1+ii-i;
Do[t[[it,jt++]]=PaddedForm[a[[ij++]]//FortranForm,{7,4}],{j,1,kr}],
{i,jbeg,n}];
Print[TableForm[Take[t,it],TableAlignments->{Right,Right},
TableDirections->{Column,Row},TableSpacing->{0,2}]];
jref=jend; If[jref<n,Goto[cycle]];
];
SymmSkyMatrixUpperTrianglePrint[S_]:= Module[
{p,a,cycle,i,ij,it,j,j1,j2,jref,jbeg,jend,kcmax,k,kc,m,n,c,t},
{p,a}=S; n=Dimensions[p][[1]]-1; kcmax=5; jref=0;
Label[cycle]; Print[" "];
jbeg=jref+1; jend=Min[jref+kcmax,n]; kc=jend-jref;
t=Table[" ",{jend+1},{kc+1}];
Do [If [p[[j+1]]>0,c=" ",c="*"];
t[[1,j-jref+1]]=StringJoin[c,"Col",ToString[PaddedForm[j,3]]],
{j,jbeg,jend}]; it=1;
Do [it++; If [p[[i+1]]>0,c=" ",c="*"];
t[[it,1]]=StringJoin[c,"Row",ToString[PaddedForm[i,3]]]; j=jref;
Do [j++; If [j<i, Continue[]]; ij=Abs[p[[j+1]]]+i-j;
If [ij<=Abs[p[[j]]], Continue[]];
t[[it,k+1]]=PaddedForm[a[[ij]]//FortranForm,{7,4}],
{k,1,kc}],
{i,1,jend}];
Print[TableForm[Take[t,it],TableAlignments->{Right,Right},
TableDirections->{Column,Row},TableSpacing->{0,2}]];
jref=jend; If[jref<n,Goto[cycle]];
];

Sr={{0, 1, 3, 6}, {1., 2., 7., 4., 23., 97.}};

SymmSkyMatrixLowerTrianglePrint[Sr];SymmSkyMatrixUpperTrianglePrint[Sr];

26–16
26–17 §26.3 A SKYSOLVER IMPLEMENTATION

Cell 26.9 Skymatrix Mapping

SymmSkyMatrixLowerTriangleMap[S_]:=Module[
{p,a,cycle,i,ii,ij,it,itop,j,jj,j1,j2,jref,jbeg,jend,jt,kcmax,kc,kr,m,n,c,t},
{p,a}=S; n=Dimensions[p][[1]]-1; kcmax=40; jref=0;
Label[cycle]; Print[" "];
jbeg=jref+1; jend=Min[jref+kcmax,n]; kc=jend-jref;
itop=2; If[jend>9,itop=3]; If[jend>99,itop=4]; If[jend>999,itop=5];
t=Table[" ",{n-jref+itop},{kc+1}]; it=0;
If [itop>=5, it++; Do [m=Floor[j/1000];
If[m>0,t[[it,j-jref+1]]=ToString[Mod[m,10]]], {j,jbeg,jend}]];
If [itop>=4, it++; Do [m=Floor[j/100];
If[m>0,t[[it,j-jref+1]]=ToString[Mod[m,10]]], {j,jbeg,jend}]];
If [itop>=3, it++; Do [m=Floor[j/10];
If[m>0,t[[it,j-jref+1]]=ToString[Mod[m,10]]], {j,jbeg,jend}]];
it++; Do[t[[it,j-jref+1]]=ToString[Mod[j,10]],{j,jbeg,jend}];
it++; Do[If[p[[j+1]]<0,t[[it,j-jref+1]]="*"],{j,jbeg,jend}];
Do [ii=Abs[p[[i+1]]]; m=ii-Abs[p[[i]]]-1; j1=Max[i-m,jbeg];j2=Min[i,jend];
kr=j2-j1+1; If [kr<=0, Continue[]]; If [p[[i+1]]>0,c=" ",c="*"];
it++; t[[it,1]]=StringJoin[ToString[PaddedForm[i,2]],c];
jt=j1-jbeg+2; ij=j1+ii-i;
Do [ c=" 0"; If[a[[ij]]>0,c=" +"]; If[a[[ij++]]<0,c=" -"];
t[[it,jt++]]=c, {j,1,kr}],
{i,jbeg,n}];
Print[TableForm[Take[t,it],TableAlignments->{Right,Right},
TableDirections->{Column,Row},TableSpacing->{0,0}]];
jref=jend; If[jref<n,Goto[cycle]];
];
SymmSkyMatrixUpperTriangleMap[S_]:=Module[
{p,a,cycle,i,ij,it,itop,j,j1,j2,jref,jbeg,jend,kcmax,k,kc,m,n,c,t},
{p,a}=S; n=Dimensions[p][[1]]-1; kcmax=40; jref=0;
Label[cycle]; Print[" "];
jbeg=jref+1; jend=Min[jref+kcmax,n]; kc=jend-jref;
itop=2; If[jend>9,itop=3]; If[jend>99,itop=4]; If[jend>999,itop=5];
t=Table[" ",{jend+itop},{kc+1}]; it=0;
If [itop>=5, it++; Do [m=Floor[j/1000];
If[m>0,t[[it,j-jref+1]]=ToString[Mod[m,10]]], {j,jbeg,jend}]];
If [itop>=4, it++; Do [m=Floor[j/100];
If[m>0,t[[it,j-jref+1]]=ToString[Mod[m,10]]], {j,jbeg,jend}]];
If [itop>=3, it++; Do [m=Floor[j/10];
If[m>0,t[[it,j-jref+1]]=ToString[Mod[m,10]]], {j,jbeg,jend}]];
it++; Do[t[[it,j-jref+1]]=ToString[Mod[j,10]],{j,jbeg,jend}];
it++; Do[If[p[[j+1]]<0,t[[it,j-jref+1]]="*"],{j,jbeg,jend}];
Do [it++; If [p[[i+1]]>0,c=" ",c="*"];
t[[it,1]]=StringJoin[ToString[PaddedForm[i,2]],c]; j=jref;
Do [j++; If [j<i, Continue[]]; ij=Abs[p[[j+1]]]+i-j;
If [ij<=Abs[p[[j]]], Continue[]]; c=" 0";
If[a[[ij]]>0,c=" +"]; If[a[[ij++]]<0,c=" -"]; t[[it,k+1]]=c,
{k,1,kc}],
{i,1,jend}];
Print[TableForm[Take[t,it],TableAlignments->{Right,Right},
TableDirections->{Column,Row},TableSpacing->{0,0}]];
jref=jend; If[jref<n,Goto[cycle]];
];
26–17
Chapter 26: SOLVING FEM EQUATIONS 26–18

Cell 26.10 Skymatrix Reconstruction from Factors

SymmSkyMatrixLDUReconstruct[S_]:= Module[
{p,ldu,a,v,n,i,ii,ij,j,jj,jk,jmj,k,m},
{p,ldu}=S; a=ldu; n=Length[p]-1; v=Table[0,{n}];
Do [jmj=Abs[p[[j]]]; jj=p[[j+1]]; If [jj<0, Continue[]];
jk=jj-jmj; v[[jk]]=ldu[[jj]];
Do [ij=jmj+k; i=j+ij-jj; ii=p[[i+1]]; If [ii<0, v[[k]]=0; Continue[]];
If [i!=j, v[[k]]=ldu[[ij]]*ldu[[ii]]];
m=Min[ii-Abs[p[[i]]],k]; a[[ij]]= v[[k]];
a[[ij]]+=Take[ldu,{ii-m+1,ii-1}].Take[v,{k-m+1,k-1}],
{k,1,jk}],
{j,1,n}]; Return[{p,a}];
];

SymmSkyMatrixLDinvUReconstruct[S_]:= Module[
{p,ldu,a,v,n,i,ii,ij,j,jj,jk,jmj,k,m},
{p,ldu}=S; a=ldu; n=Length[p]-1; v=Table[0,{n}];
Do [jmj=Abs[p[[j]]]; jj=p[[j+1]]; If [jj<0, Continue[]];
jk=jj-jmj; v[[jk]]=1/ldu[[jj]];
Do [ij=jmj+k; i=j+ij-jj; ii=p[[i+1]]; If [ii<0, v[[k]]=0; Continue[]];
If [i!=j, v[[k]]=ldu[[ij]]/ldu[[ii]]];
m=Min[ii-Abs[p[[i]]],k]; a[[ij]]= v[[k]];
a[[ij]]+=Take[ldu,{ii-m+1,ii-1}].Take[v,{k-m+1,k-1}],
{k,1,jk}],
{j,1,n}]; Return[{p,a}];
];

p={0,1,2,5,8,9,15}; s={11,22,13,0,33,24,34,44,55,16,0,0,46,56,66};
S={p,s};
Sr=SymmSkyMatrixLDinvUReconstruct[S]; Print[Sr//InputForm];
Print[SymmSkyMatrixFactor[Sr,0]];

§26.3.7. *Miscellaneous Utilities

Finally, Cell 26.11 lists three miscellaneous modules. The most useful one is probably
SymmSkyMatrixConvertToFull, which converts its skymatrix argument to a fully stored symmetric matrix.
This is useful for things like a quick and dirty computation of eigenvalues:
Print[Eigenvalues[SymmSkyMatrixConvertToFull[S]]];

because Mathematica built-in eigensolvers require that the matrix be supplied in full storage form.

26–18
26–19 §26.3 A SKYSOLVER IMPLEMENTATION

Cell 26.11 Miscellaneous Skymatrix Utilities

SymmSkyMatrixConvertToFull[S_]:= Module[
{p,a,aa,n,j,jj,jmj,k},
{p,a}=S; n=Length[p]-1; aa=Table[0,{n},{n}];
Do [jmj=Abs[p[[j]]]; jj=Abs[p[[j+1]]]; aa[[j,j]]=a[[jj]];
Do [aa[[j,j-k]]=aa[[j-k,j]]=a[[jj-k]],{k,1,jj-jmj-1}],
{j,1,n}]; Return[aa];
];

SymmSkyMatrixConvertUnitUpperTriangleToFull[S_]:= Module[
{p,ldu,aa,n,j,jj,jmj,k},
{p,ldu}=S; n=Length[p]-1; aa=Table[0,{n},{n}];
Do [jmj=Abs[p[[j]]]; jj=Abs[p[[j+1]]]; aa[[j,j]]=1;
Do [aa[[j-k,j]]=ldu[[jj-k]],{k,1,jj-jmj-1}],
{j,1,n}]; Return[aa];
];

SymmSkyMatrixConvertDiagonalToFull[S_]:= Module[
{p,ldu,aa,n,i,j,jj,jmj,k},
{p,ldu}=S; n=Length[p]-1; aa=Table[0,{n},{n}];
Do [jj=Abs[p[[j+1]]]; aa[[j,j]]=ldu[[jj]],
{j,1,n}]; Return[aa];
];

26–19
Chapter 26: SOLVING FEM EQUATIONS 26–20

Homework Exercises for Chapter 26

Solving FEM Equations

u 1, f1 u 2, f2 u 3, f3 u 4, f4 u 5, f5
(1) (2) (3) (4)

1 2 3 4 5

Figure 26.1. Structure for Exercise 26.1

EXERCISE 26.1 [A/C:10+10+15] Consider the 4-element assembly of bar elements shown in Figure 26.1.
The only degree of freedom at each node is a translation along x. The element stiffness matrix of each element
is

1 −1
K(e) = (E26.1)
−1 1
(a) Assemble the 5 × 5 master stiffness matrix K showing it as a full symmetric matrix. Hint: the diagonal
entries are 1, 2, 2, 2, 1.
(b) Show K stored as a skyline matrix using a representation like illustrated in (26.15). Hint p = { 0, 1,
3, 5, 7, 9}.
(c) Perform the symmetric factorization K = LDLT of (26.7), where K is stored as a full matrix to simplify
hand work.6 Show that the entries of D are 1, 1, 1, 1 and 0, which mean that K is singular. Why?

6 You can do this with Mathematica using the function LUDecomposition, but hand work is as quick.

26–20
 27
.

A Complete
Plane Stress
FEM Program

27–1
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–2

TABLE OF CONTENTS

Page
§27.1. Introduction 27–3
§27.2. Analysis Stages 27–3
§27.3. Model Definition 27–3
§27.3.1. Benchmark Problems . . . . . . . . . . . . . . . 27–4
§27.3.2. Node Coordinates . . . . . . . . . . . . . . . 27–4
§27.3.3. Element Type . . . . . . . . . . . . . . . . . 27–6
§27.3.4. Element Connectivity . . . . . . . . . . . . . . 27–6
§27.3.5. Material Properties . . . . . . . . . . . . . . . 27–8
§27.3.6. Fabrication Properties . . . . . . . . . . . . . . 27–8
§27.3.7. Node Freedom Tags . . . . . . . . . . . . . . . 27–8
§27.3.8. Node Freedom Values . . . . . . . . . . . . . . 27–9
§27.3.9. Processing Options . . . . . . . . . . . . . . . 27–9
§27.3.10. Model Display Utility . . . . . . . . . . . . . . 27–9
§27.3.11. Model Definition Print Utilities . . . . . . . . . . . 27–10
§27.3.12. Model Definition Script Samples . . . . . . . . . . 27–10
§27.4. Processing 27–12
§27.4.1. Processing Tasks . . . . . . . . . . . . . . . . 27–12
§27.5. Postprocessing 27–14
§27.5.1. Result Print Utilities . . . . . . . . . . . . . . 27–14
§27.5.2. Displacement Field Contour Plots . . . . . . . . . . 27–15
§27.5.3. Stress Field Contour Plots . . . . . . . . . . . . . 27–15
§27.5.4. Animation . . . . . . . . . . . . . . . . . . 27–16
§27.6. A Complete Problem Script Cell 27–16
§27. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 27–18

27–2
27–3 §27.3 MODEL DEFINITION

§27.1. Introduction

This Chapter describes a complete finite element program for analysis of plane stress problems.
Unlike the previous chapters the description is top down, i.e. starts from the main driver down to
more specific modules.
The program can support questions given in the take-home final exam, if they pertain to the analysis
of given plane stress problems. Consequently this Chapter serves as an informal users manual.

§27.2. Analysis Stages

As in all DSM-based FEM programs, the analysis of plane stress problems involves three major
stages: (I) preprocessing or model definition, (II) processing, and (III) postprocessing.
The preprocessing portion of the plane stress analysis is done by the first part of the problem script,
driver program, already encountered in Chapters 21-22. The script directly sets the problem data
structures. Preprocessing tasks include:
I.1 Model definition by direct setting of the data structures.
I.2 Plot of the FEM mesh for verification. At the minimum this involves producing a mesh
picture that shows nodes and element labels.
The processing stage involves three tasks:
II.1 Assembly of the master stiffness equations Ku = f. The plane stress assembler is of multiple
element type (MET). This kind of assembler was discussed in §27.4. Element types include
various plane stress Iso-P models as well as bars.
II.2 Application of displacement BC by a modification method that produces K̂u = f̂. The same
modules described in §21.3.3 are used, since those are application problem independent.
II.3 Solution of the modified equations for node displacements u. The built-in Mathematica
function LinearSolve is used for this task.
Upon executing the processing steps, the nodal displacement solution is available. The postpro-
cessing stage embodies three tasks:
III.1 Recovery of node forces including reactions through built-in matrix multiplication f = Ku.
III.2 Recovery of plate stresses and bar forces (if bars are present). The former are subject to
interelement averaging (Chapter 28) to get nodal stresses.
III.3 Print and plotting of results.

§27.3. Model Definition

The model-definition data may be broken down into three sets, which are listed below by order of
appearance:
 Geometry data: node coordinates
Model definition Element data: type, connectivity, material and fabrication
Degree of freedom (DOF) activity data: force and displacement BCs
(27.1)
The element data is broken down into four subsets: type, connnectivity, material and fabrication,
each of which has its own data structure. The degree of freedom data is broken into two subsets:

27–3
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–4

Table 27.1. Plane Stress Model Definition Data Structures

Long name Short name Dimensions Description

NodeCoordinates nodxyz numnod x 2 Node coordinates in global system

ElemTypes elenod numele Element type identifiers
ElemNodes elenod numele x <var> Element node lists
ElemMaterials elemat numele x <var> Element material properties
ElemFabrications elefab numele x <var> Element fabrication properties
NodeDOFTags nodtag numnod x 2 Node freedom tags marking BC type
NodeDOFValues nodval numnod x 2 Node freedom specified values
ProcessOptions prcopt <var> Processing specifications
Notation: numnod: number of nodes (no numbering gaps allowed, each node has two )
displacement DOFs; numele: number of elements; <var> variable dimension.
Long names are used in user-written problem scripts. Short names are used in programmed modules.
Dimensions refer to the first two level Dimensions of the Mathematica list that implements the
data structure. These give a guide as to implementation in a low level language such as C.

tags and values. In addition there are miscellaneous process options, such as the symbolic versus
numeric processing. These options are conveniently collected in a separate data set.
Accordingly, the model-definition input to the plane stress FEM program consists of eight data
structures, which are called NodeCoordinates, ElemTypes, ElemNodes, ElemMaterials,
ElemFabrications, NodeDOFTags, NodeDOFValues and ProcessOptions. These are sum-
marized in Table 27.1.
The configuration of these data structures are described in the following subsections with reference
to the benchmark problems and discretizations shown in Figures 27.1 and 27.2.

§27.3.1. Benchmark Problems

Figure 27.1(a) illustrates a rectangular steel plate in plane stress under uniform uniaxial loading in
the y (vertical) direction. Note that the load q is specified as force per unit area (kips per square
inch); thus q has not been integrated through the thickness h. The exact analytical solution of the
problem is1 σ yy = q, σx x = σx y = 0, u y = qy/E, u x = −νu y = −νq x/E. This problem should
be solved exactly by any finite element mesh as long as the model is consistent and stable. In
particular, the two one-quadrilateral-element models shown in Figure 27.1(b,c).
A similar but more complicated problem is shown in Figure 27.2: a rectangular steel plate dimen-
sioned and loaded as that of Figure 27.1(a) but now with a central circular hole. This problem is
defined in Figure 27.2(a). A FEM solution is to be obtained using the two quadrilateral element
models (with 4-node and 9-nodes, respectively) depicted in Figure 27.2(b,c). The main result
sought is the stress concentration factor on the hole boundary, and comparison of this computed
factor with the exact analytical value.

1 The displacement solution u y = qy/E and u x = −νu y assumes that the plate centerlines do not translate or rotate, a
condition enfoirced in the FEM discretizations shown in Figure 27.1(b,c).

27–4
27–5 §27.3 MODEL DEFINITION

(a) y q = 10 ksi (b) (c)

;; ;;
75 kips 75 kips 25 kips 100 kips 25 kips
A B C B C B C
1 3 1 4 7

1
; 2 5
1
8

;;
;;
;;;
;;
12 in H J D J 2 4 D 6 9 D

;;;
J 3

;;
x

;
;
E = 10000 ksi Model (I):
ν = 0.25 Model (II):
h = 3 in 4 nodes, 8 DOFs, 9 nodes, 18 DOFs,
1 bilinear quad 1 biquadratic quad

G F E Global node numbers shown

q
10 in

Figure 27.1. Rectangular plate under uniform uniaxial loading, and two one-
element FEM discretizations of its upper right quadrant.

§27.3.2. Node Coordinates

The geometry data is specified through NodeCoordinates. This is a list of node coordinates
configured as

NodeCoordinates = { { x1 , y1 },{ x2 , y2 }, . . . { x N , y N } } (27.2)

where N is the number of nodes. Coordinate values should be normally be floating point numbers;2
use the N function to insure that property if necessary. Nodes must be numbered consecutively and
no gaps are permitted.

Example 27.1. For Model (I) of Figure 27.1(b):

NodeCoordinates=N[{ { 0,6 },{ 0,0 },{ 5,6 },{ 5,0 } }];

Example 27.2. For Model (II) of Figure 27.1(c):

NodeCoordinates=N[{ { 0,6 },{ 0,3 },{ 0,0 },{ 5/2,6 },{ 5/2,3 },{ 5/2,0 },{ 5,6 },{ 5,3 },{ 5,0 } }];

Example 27.3. For Model (I) of Figure 27.2(a), using a bit of coordinate generation:
s={1,0.70,0.48,0.30,0.16,0.07,0.0};
xy1={0,6}; xy7={0,1}; xy8={2.5,6}; xy14={Cos[3*Pi/8],Sin[3*Pi/8]};
xy8={2.5,6}; xy21={Cos[Pi/4],Sin[Pi/4]}; xy15={5,6};
xy22={5,2}; xy28={Cos[Pi/8],Sin[Pi/8]}; xy29={5,0}; xy35={1,0};
NodeCoordinates=Table[{0,0},{35}];
Do[NodeCoordinates[[n]]=N[s[[n]] *xy1+(1-s[[n]]) *xy7], {n,1,7}];
Do[NodeCoordinates[[n]]=N[s[[n-7]] *xy8+(1-s[[n-7]]) *xy14],{n,8,14}];
Do[NodeCoordinates[[n]]=N[s[[n-14]]*xy15+(1-s[[n-14]])*xy21],{n,15,21}];

2 Unless one is doing a symbolic or exact-arithmetic analysis. Those are rare at the level of a full FEM analysis.

27–5
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–6

(a) (b) Node 8 is exactly midway between 1 and 15 (c)

y q = 10 ksi 75 kips

; ;;
37.5 kips 37.5 kips 25 kips 100 kips 25 kips

B C 15 B C

;
A B C 8 8 15
1 1
Note: internal point of a
1 9-node quadrilateral is placed

; ;;
2 7 at intersection of the medians
9 2 9

;
2 1 16
3 3

; ;;
J 10 16
8 10
12 in H K D
4
22

;
17 17
R = 1 in x 11
22
4 11
2
5 12 18 23
5 18

; ;
23 12
E = 10000 ksi 6 13 19 24 6 13 19 24
ν = 0.25 7 20
26
25 7 20 25
J 14 21 J 26
h = 3 in 27 14
21 27
28 28 35 34 33 31
35 34 33 32 31 30 29 32 30 29

;;
;
;;
;
;;
;
;;
;
G F E K D K D
q
10 in Model (I): 35 nodes, 70 DOFs, Model (II): 35 nodes, 70 DOFs,
24 bilinear quads 6 biquadratic quads

Figure 27.2. Plate with a circular hole and two FEM discretizations of its upper right quadrant. Only a few
element numbers are shown to reduce clutter. Note: the meshes generated with the example scripts given later
differ slightly from those shown above. Compare, for example, mesh in (b) above with that in Figure 27.3.

Do[NodeCoordinates[[n]]=N[s[[n-21]]*xy22+(1-s[[n-21]])*xy28],{n,22,28}];
Do[NodeCoordinates[[n]]=N[s[[n-28]]*xy29+(1-s[[n-28]])*xy35],{n,29,35}];
The result of this generation is that some of the interior nodes are not in the same positions as sketched in
Figure 27.2(b), but this slight change hardly affects the results.

§27.3.3. Element Type

The element type is a label that specifies the type of model to be used. These labels are placed into
an element type list:

ElemTypes = { type(1) , type(2) , . . . type Ne } (27.3)

Here typee is the type identifier of the e-th element specified as a character string. Ne is the number
of elements; no element numbering gaps are allowed. Legal type identifiers are listed in Table 27.2.
ElemTypes=Table["Quad4",{ numele }];
ElemTypes=Table["Quad9",{ numele }];
Here numele is a variable that contains the number of elements. This can be extracted, for example,
as numele=Length[ElemNodes], where ElemNodes is defined below, if ElemNodes is defined
first as is often the case.

§27.3.4. Element Connectivity

Element connectivity information specifies how the elements are connected.3 This information is

3 Some FEM programs call this the “topology” data.

27–6
27–7 §27.3 MODEL DEFINITION

Table 27.2. Element Type Identifiers Implemented in Plane Stress Program

Identifier Nodes Model for Description

"Bar2" 2 bar 2-node bar element

"Bar3" 3 bar 3-node bar element; may be curved
"Trig3" 3 plate 3-node linear triangle
"Trig6" 6 plate 6-node quadratic triangle (3-interior point Gauss rule)
"Trig6.-3" 6 plate 6-node quadratic triangle (midpoint Gauss rule)
"Trig10" 10 plate 10-node cubic triangle (7 point Gauss rule)
"Quad4" 4 plate RS 4-node iso-P bilinear quad (2 × 2 Gauss rule)
"Quad4.1" 4 plate RD 4-node iso-P bilinear quad (1-point Gauss rule)
"Quad8" 8 plate RS 8-node iso-P serendipity quad (3 × 3 Gauss rule)
"Quad8.2" 8 plate RD 8-node iso-P serendipity quad (2 × 2 Gauss rule)
"Quad9" 9 plate RS 9-node iso-P biquadratic quad (3 × 3 Gauss rule)
"Quad9.2" 9 plate RD 9-node iso-P biquadratic quad (2 × 2 Gauss rule)
RS: rank sufficient, RD: Rank deficient.

stored in ElemNodes, which is a list of element nodelists:

ElemNodes = { { enl(1) }, { enl(2) }, . . . { enl Ne } } (27.4)

Here enle denotes the lists of nodes of the element e (as given by global node numbers) and Ne is
the total number of elements.
Element boundaries must be traversed counterclockwise (CCW) but you can start at any corner.
Numbering elements with midnodes requires more care: begin listing corners CCW, followed by
midpoints also CCW (first midpoint is the one that follows first corner when traversing CCW). If
element has thirdpoints, as in the case of the 10-node triangle, begin listing corners CCW, followed
by thirdpoints CCW (first thirdpoint is the one that follows first corner). When elements have an
interior node, as in the 9-node quadrilateral or 10-node triangle, that node goes last.
Example 27.4. For Model (I) of which has only one 4-node quadrilateral:
ElemNodes={ { 1,2,4,3 } };

Example 27.5. For Model (II) of Figure 27.1(c), which has only one 9-node quadrilateral:
ElemNodes={ { 1,3,9,7,2,6,8,4,5 } };
numbering the elements from top to bottom and from left to right:
ElemNodes=Table[{0,0,0,0},{24}];
ElemNodes[[1]]={1,2,9,8};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,2,6}];
ElemNodes[[7]]=ElemNodes[[6]]+{2,2,2,2};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,8,12}];
ElemNodes[[13]]=ElemNodes[[12]]+{2,2,2,2};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,14,18}];
ElemNodes[[19]]=ElemNodes[[18]]+{2,2,2,2};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,20,24}];

27–7
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–8

Example 27.6. For Model (II) of Figure 27.2(c), numbering the elements from top to bottom and from left to
right:

ElemNodes=Table[{0,0,0,0,0,0,0,0,0},{6}];
ElemNodes[[1]]={1,3,17,15,2,10,16,8,9};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{2,2,2,2,2,2,2,2,2},{e,2,3}];
ElemNodes[[4]]=ElemNodes[[3]]+{10,10,10,10,10,10,10,10,10};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{2,2,2,2,2,2,2,2,2},{e,5,6}];
Since this particular mesh only has 6 elements, it would be indeed faster to write down the six nodelists.

§27.3.5. Material Properties

Data structure ElemMaterials is a list that provides the constitutive properties of the elements:

ElemMaterials = { mprop(1) , mprop(2) , . . . mprop Ne } (27.5)

For a plate element, mprop is the stress-strain matrix of elastic moduli (also known as elasticity
matrix) arranged as { { E11,E12,E33 },{ E12,E22,E23 },{ E13,E23,E33 } }. Note that although
this matrix is symmetric, it must be specified as a full 3 × 3 matrix. For a bar element, mprop is
simply the longitudinal elastic modulus.
A common case in practice is that (i) all elements are plates, (ii) the plate material is uniform and
isotropic. An isotropic elastic material is specified by the elastic modulus E and Poisson’s ratio ν.
Then this list can be generated by a single Table instruction upon building the elastic ity matrix,
as in the example below.

Example 27.7. For all FEM discretizations in Figures 27.1 and 27.2 all elements are plates of the same
isotropic material. Suppose that values of the elastic modulus E and Poisson’s ratio ν are stored in Em and nu,
respectively, which are typically declared at the beginning of the problem script. Let numele give the number
of elements. Then the material data is compactly declared by saying
Emat=Em/(1-nu^2)*{ { 1,nu,0 },{ nu,1,0 },{ 0,0,(1-nu)/2 } };
ElemMaterials=Table[Emat,{ numele }];

§27.3.6. Fabrication Properties

Data structure ElemFabrications is a list that provides the fabrication properties of the elements:

ElemFabrications = { fprop(1) , fprop(2) , . . . fprop Ne } (27.6)

For a plate element, fprop is the thickness h of the plate, assumed constant.4 For a bar element,
fprop is the cross section area.
If all elements are plates with the same thickness, this list can be easily generated by a Table
instruction as in the example below.

4 It is possible also to specify a variable thickness by making fprop a list that contains the thicknesses at the nodes. Since
the variable thickness case is comparatively rare, it will not be described here.

27–8
27–9 §27.3 MODEL DEFINITION

Example 27.8. For all FEM discretizations in Figures 27.1 and 27.2 all elements are plates with the same
thickness h, which is stored in variable th. This is typically declared at the start of the problem script. As
before, numele has the number of elements. Then the fabrication data is compactly declared by saying
ElemFabrications=Table[th,{ numele }]

§27.3.7. Node Freedom Tags

Data structure NodeDOFTags is a list that labels each node DOF as to whether the load or the
displacement is specified. The configuration of this list is similar to that of NodeCoordinates:
NodeDOFTags={ { tagx1 , tag y1 },{ tagx2 , tag y2 }, . . . { tagx N , tag y N } } (27.7)
The tag value is 0 if the force is specified and 1 if the displacement is specified.
When there are a lot of nodes, often the quickest way to specify this list is to create with a Table
command that initializes it to all zeros. Then displacement BCs are inserted appropriately, as in the
example below.
Example 27.9. For Model (I) in Figure 27.2(a):

numnod=Length[NodeCoordinates];
NodeDOFTags=Table[{0,0},{numnod}]; (* create and initialize to zero *)
Do[NodeDOFTags[[n]]={1,0},{n,1,7}]; (* vroller @ nodes 1 through 7 *)
Do[NodeDOFTags[[n]]={0,1},{n,29,35}]; (* hroller @ nodes 29 through 35 *)
This scheme works well because typically the number of supported nodes is small compared to the total
number.

§27.3.8. Node Freedom Values

Data structure NodeDOFValues is a list with the same node by node configuration as NodeDOFTags:
NodeDOFValues={ { valuex1 , value y1 },{ valuex2 , value y2 }, . . . { valuex N , value y N } }
(27.8)
Here value is the specified value of the applied node force component if the corresponding tag is
zero, and of the prescribed displacement component if the tag is one.
Often most of the entries of (27.8) are zero. If so a quick way to build it is to create it with a Table
command that initializes it to zero. Then nonzero values are inserted as in the example below.
Example 27.10. For the model (I) in Figure 27.2(a) only 3 values (for the y forces on nodes 1, 8 and 15) will
be nonzero:

numnod=Length[NodeCoordinates];
NodeDOFValues=Table[{0,0},{numnod}]; (* create and initialize to zero *)
NodeDOFValues[[1]]=NodeDOFValues[[15]]={0,37.5};
NodeDOFValues[[8]]={0,75}; (* y nodal loads *)

§27.3.9. Processing Options

Array ProcessOptions is a list of general processing options that presently contains only the
numer logical flag. This is normally be set to True to specify numeric computations:
ProcessOptions={ True };

27–9
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–10

§27.3.10. Model Display Utility

Only one graphic display utility is presently provided to show

the mesh. Nodes and elements of Model (I) of Figure 27.2(a)
may be plotted by saying
One element mesh - 4 node quad
1 8 15
aspect=6/5;
Plot2DElementsAndNodes[NodeCoordinates,
1 7
ElemNodes,aspect, "Plate with circular
hole - 4-node quad model",True,True]; 2 9 16

Here aspect is the plot frame aspect ratio (y dimension over 2 8

x dimension). The 4th argument is a plot title textstring. The 3 10
17
13

last two True argument values specify that node labels and 3 9
element labels, respectively, be shown. The output of the 4 11 14
18
mesh plot command is shown in Figure 27.3. 4
10 22
5 12 15
5 19 23
6 13 11 16
24
§27.3.11. Model Definition Print Utilities 7 6
14 12
20
17 25 19
21
1827 26 20
28 21
22
24 23
Several print utilities are provided in the plane stress program 35 34 33 32 31 30 29

to print out model definition data in tabular form. They are

invoked as follows.
Figure 27.3. Mesh plot for
Model (I) of Figure 27.2(a).

To print the node coordinates:

PrintPlaneStressNodeCoordinates[NodeCoordinates,title,digits];
To print the element types and nodes:
PrintPlaneStressElementTypeNodes[ElemTypes,EleNodes,title,digits];
To print the element materials and fabrications:
PrintPlaneStressElementMatFab[ElemMaterials,ElemFabrications,title,digits];
To print freedom activity data:
PrintPlaneStressFreedomActivity[NodeDOFTags,NodeDOFValues,title,digits];
In all cases, title is an optional character string to be printed as a title before the table; for example "Node
coordinates". To eliminate the title, specify "" (two quote marks together).
The last argument of the print modules: digits, is optional. If set to { d,f } it specifies that floating point
numbers are to be printed with room for at least d digits, with f digits after the decimal point. If digits is
specified as a void list: { }, a preset default is used for d and f.

§27.3.12. Model Definition Script Samples

As capstone examples, Figures 27.4 and Figures 27.5 list the preprocessing (model definition) parts
of the problem scripts for Model (I) and (II), respectively, of Figure 27.1(b,c).

27–10
27–11 §27.3 MODEL DEFINITION

ClearAll[Em,ν,th];
Em=10000; ν=.25; th=3; aspect=6/5; Nsub=4;
Emat=Em/(1-ν^2)*{{1,ν,0},{ν,1,0},{0,0,(1-ν)/2}};

(* Define FEM model *)

NodeCoordinates=N[{{0,6},{0,0},{5,6},{5,0}}];
PrintPlaneStressNodeCoordinates[NodeCoordinates,"",{6,4}];
ElemNodes= {{1,2,4,3}};
numnod=Length[NodeCoordinates]; numele=Length[ElemNodes];
ElemTypes= Table["Quad4",{numele}];
PrintPlaneStressElementTypeNodes[ElemTypes,ElemNodes,"",{}];
ElemMaterials= Table[Emat, {numele}];
ElemFabrications=Table[th, {numele}];
PrintPlaneStressElementMatFab[ElemMaterials,ElemFabrications,"",{}];
NodeDOFValues=NodeDOFTags=Table[{0,0},{numnod}];
NodeDOFValues[[1]]=NodeDOFValues[[3]]={0,75}; (* nodal loads *)
NodeDOFTags[[1]]={1,0}; (* vroller @ node 1 *)
NodeDOFTags[[2]]={1,1}; (* fixed node 2 *)
NodeDOFTags[[4]]={0,1}; (* hroller @ node 4 *)
PrintPlaneStressFreedomActivity[NodeDOFTags,NodeDOFValues,"",{}];
ProcessOptions={True};
Plot2DElementsAndNodes[NodeCoordinates,ElemNodes,aspect,
"One element mesh - 4-node quad",True,True];

Figure 27.4. Model definition part of Model (I) of Figure 27.1(b).

ClearAll[Em,ν,th];
Em=10000; ν=.25; th=3; aspect=6/5; Nsub=4;
Emat=Em/(1-ν^2)*{{1,ν,0},{ν,1,0},{0,0,(1-ν)/2}};

(* Define FEM model *)

NodeCoordinates=N[{{0,6},{0,3},{0,0},{5/2,6},{5/2,3},
{5/2,0},{5,6},{5,3},{5,0}}];
PrintPlaneStressNodeCoordinates[NodeCoordinates,"",{6,4}];
ElemNodes= {{1,3,9,7,2,6,8,4,5}};
numnod=Length[NodeCoordinates]; numele=Length[ElemNodes];
ElemTypes= Table["Quad9",{numele}];
PrintPlaneStressElementTypeNodes[ElemTypes,ElemNodes,"",{}];
ElemMaterials= Table[Emat, {numele}];
ElemFabrications=Table[th, {numele}];
PrintPlaneStressElementMatFab[ElemMaterials,ElemFabrications,"",{}];
NodeDOFValues=NodeDOFTags=Table[{0,0},{numnod}];
NodeDOFValues[[1]]=NodeDOFValues[[7]]={0,25};
NodeDOFValues[[4]]={0,100}; (* nodal loads *)
NodeDOFTags[[1]]=NodeDOFTags[[2]]={1,0}; (* vroller @ nodes 1,2 *)
NodeDOFTags[[3]]={1,1}; (* fixed node 3 *)
NodeDOFTags[[6]]=NodeDOFTags[[9]]={0,1}; (* hroller @ nodes 6,9 *)
PrintPlaneStressFreedomActivity[NodeDOFTags,NodeDOFValues,"",{}];
ProcessOptions={True};
Plot2DElementsAndNodes[NodeCoordinates,ElemNodes,aspect,
"One element mesh - 9-node quad",True,True];

Figure 27.5. Model definition part of Model (II) of Figure 27.1(c).

27–11
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–12

PlaneStressSolution[nodxyz_,eletyp_,elenod_,elemat_,elefab_,
nodtag_,nodval_,prcopt_]:= Module[{K,Kmod,f,fmod,u,numer=True,
noddis,nodfor,nodpnc,nodsig,barele,barfor},
If [Length[prcopt]>=1, numer=prcopt[[1]]];
K=PlaneStressMasterStiffness[nodxyz,eletyp,elenod,elemat,elefab,prcopt];
If [K==Null,Return[Table[Null,{6}]]];
Kmod=ModifiedMasterStiffness[nodtag,K];
f=FlatNodePartVector[nodval];
fmod=ModifiedNodeForces[nodtag,nodval,K,f];
u=LinearSolve[Kmod,fmod]; If [numer,u=Chop[u]];
f=K.u; If [numer,f=Chop[f]];
nodfor=NodePartFlatVector[2,f]; noddis=NodePartFlatVector[2,u];
{nodpnc,nodsig}=PlaneStressPlateStresses[nodxyz,eletyp,elenod,elemat,
elefab,noddis,prcopt];
{barele,barfor}=PlaneStressBarForces[nodxyz,eletyp,elenod,elemat,
elefab,noddis,prcopt];
ClearAll[K,Kmod];
Return[{noddis,nodfor,nodpnc,nodsig,barele,barfor}];
];

Figure 27.6. Module to drive the analysis of the plane stress problem.

§27.4. Processing
Once the model definition is complete, the plane stress analysis is carried out by calling module
LinearSolutionOfPlaneStressModel listed in Figure 27.6. This module is invoked from the
problem driver as
{NodeDisplacements,NodeForces,NodePlateCounts,NodePlateStresses,
ElemBarNumbers,ElemBarForces}= PlaneStressSolution[NodeCoordinates,
ElemTypes,ElemNodes,ElemMaterials,ElemFabrications,
NodeDOFTags,NodeDOFValues,ProcessOptions];
The module arguments: NodeCoordinates, ElemTypes, ElemNodes , ElemMaterials,
ElemFabrications, NodeDOFTags, NodeDOFValues and ProcessOptions are the data struc-
tures described in the previous section.
The module returns the following:
NodeDisplacements Computed node displacements, in node-partitioned form.
NodeForces Recovered node forces including reactions, in node-partitioned form.
NodePlateCounts For each node, number of plate elements attached to that node. A zero
count means that no plate elements are attached to that node.
NodePlateStresses Averaged stresses at plate nodes with a nonzero NodePlateCounts.
ElemBarNumbers A list of bar elements if any specified, else an empty list.
ElemBarForces A list of bar internal forces if any bar elements were specified, else an
empty list.
§27.4.1. Processing Tasks
PlaneStressSolution carries out the following tasks. It assembles the free-free master stiffness matrix K by
calling the MET assembler PlaneStressMasterStiffness. This module is listed in Figure 27.7. A study

27–12
27–13 §27.4 PROCESSING

PlaneStressMasterStiffness[nodxyz_,eletyp_,elenod_,elemat_,
elefab_,prcopt_]:=Module[{numele=Length[elenod],
numnod=Length[nodxyz],ncoor,type,e,enl,neldof,OKtyp,OKenl,
i,j,n,ii,jj,eft,Emat,th,numer,Ke,K},
OKtyp={"Bar2","Bar3","Quad4","Quad4.1","Quad8","Quad8.2","Quad9",
"Quad9.2","Trig3","Trig6","Trig6.-3","Trig10","Trig10.6"};
OKenl= {2,3,4,4,8,8,9,9,3,6,6,10,10};
K=Table[0,{2*numnod},{2*numnod}]; numer=prcopt[[1]];
For [e=1,e<=numele,e++, type=eletyp[[e]];
If [!MemberQ[OKtyp,type], Print["Illegal type:",type,
" element e=",e," Assembly aborted"]; Return[Null] ];
enl=elenod[[e]]; n=Length[enl]; {{j}}=Position[OKtyp,type];
If [OKenl[[j]]!=n, Print ["Wrong node list length, element=",e,
" Assembly aborted"]; Return[Null] ];
eft=Flatten[Table[{2*enl[[i]]-1,2*enl[[i]]},{i,1,n}]];
ncoor=Table[nodxyz[[enl[[i]]]],{i,1,n}];
If [type=="Bar2", Em=elemat[[e]]; A=elefab[[e]];
Ke=PlaneBar2Stiffness[ncoor,Em,A,{numer}] ];
If [type=="Bar3", Em=elemat[[e]]; A=elefab[[e]];
Ke=PlaneBar3Stiffness[ncoor,Em,A,{numer}] ];
If [type=="Quad4", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Quad4IsoPMembraneStiffness[ncoor,Emat,th,{numer,2}] ];
If [type=="Quad4.1", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Quad4IsoPMembraneStiffness[ncoor,Emat,th,{numer,1}] ];
If [type=="Quad8", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Quad8IsoPMembraneStiffness[ncoor,Emat,th,{numer,3}] ];
If [type=="Quad8.2", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Quad8IsoPMembraneStiffness[ncoor,Emat,th,{numer,2}] ];
If [type=="Quad9", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Quad9IsoPMembraneStiffness[ncoor,Emat,th,{numer,3}] ];
If [type=="Quad9.2", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Quad9IsoPMembraneStiffness[ncoor,Emat,th,{numer,2}] ];
If [type=="Trig3", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Trig3IsoPMembraneStiffness[ncoor,Emat,th,{numer}] ];
If [type=="Trig6", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Trig6IsoPMembraneStiffness[ncoor,Emat,th,{numer,3}] ];
If [type=="Trig6.-3", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Trig6IsoPMembraneStiffness[ncoor,Emat,th,{numer,-3}] ];
If [type=="Trig10", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Trig10IsoPMembraneStiffness[ncoor,Emat,th,{numer,7}] ];
If [type=="Trig10.6", Emat=elemat[[e]]; th=elefab[[e]];
Ke=Trig10IsoPMembraneStiffness[ncoor,Emat,th,{numer,6}] ];
neldof=Length[Ke];
For [i=1,i<=neldof,i++, ii=eft[[i]];
For [j=i,j<=neldof,j++, jj=eft[[j]];
K[[jj,ii]]=K[[ii,jj]]+=Ke[[i,j]] ];
];
]; Return[K];
];

Figure 27.7. Plane stress assembler module.

of its code reveals that it handle the multiple element types listed in Table 27.2. The modules that compute
the element stiffness matrices have been studied in previous Chapters and need not be listed here.
The displacement BCs are applied by ModifiedMasterStiffness and ModifiedNodeForces. These are
the same modules used in Chapter 21 for the space truss program, and thus need not be described further.
The unknown node displacements u are obtained through the built in LinearSolve function, as
u=LinearSolve[Kmod,fmod]. This function is of course restricted to small systems, typically less than
200 equations (it gets extremely slow for something bigger) but it has the advantages of simplicity. The node

27–13
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–14

PlaneStressPlateStresses[nodxyz_,eletyp_,elenod_,elemat_,
elefab_,noddis_,prcopt_]:=Module[{numele=Length[elenod],
numnod=Length[nodxyz],ncoor,type,e,enl,i,k,n,Emat,th,
numer=True,nodpnc,nodsig}, If [Length[prcopt]>0, numer=prcopt[[1]]];
nodpnc=Table[0,{numnod}]; nodsig=Table[{0,0,0},{numnod}];
If [Length[prcopt]>=1, numer=prcopt[[1]]];
For [e=1,e<=numele,e++, type=eletyp[[e]];
enl=elenod[[e]]; k=Length[enl];
ncoor=Table[nodxyz[[enl[[i]]]],{i,1,k}];
ue=Table[{noddis[[enl[[i]],1]],noddis[[enl[[i]],2]]},{i,1,k}];
ue=Flatten[ue];
If [StringTake[type,3]=="Bar", Continue[]];
If [type=="Quad4", Emat=elemat[[e]]; th=elefab[[e]];
sige=Quad4IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Quad4.1", Emat=elemat[[e]]; th=elefab[[e]];
sige=Quad4IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Quad8", Emat=elemat[[e]]; th=elefab[[e]];
sige=Quad8IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Quad8.2", Emat=elemat[[e]]; th=elefab[[e]];
sige=Quad8IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Quad9", Emat=elemat[[e]]; th=elefab[[e]];
sige=Quad9IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Quad9.2", Emat=elemat[[e]]; th=elefab[[e]];
sige=Quad9IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Trig3", Emat=elemat[[e]]; th=elefab[[e]];
sige=Trig3IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Trig6", Emat=elemat[[e]]; th=elefab[[e]];
sige=Trig6IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Trig6.-3", Emat=elemat[[e]]; th=elefab[[e]];
sige=Trig6IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Trig10", Emat=elemat[[e]]; th=elefab[[e]];
sige=Trig10IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
If [type=="Trig10.6", Emat=elemat[[e]]; th=elefab[[e]];
sige=Trig10IsoPMembraneStresses[ncoor,Emat,th,ue,{numer}] ];
Do [n=enl[[i]]; nodpnc[[n]]++; nodsig[[n]]+=sige[[i]], {i,1,k}];
];
Do [k=nodpnc[[n]]; If [k>1,nodsig[[n]]/=k], {n,1,numnod}];
If [numer, nodsig=Chop[nodsig]];
Return[{nodpnc,nodsig}];
];

Figure 27.8. Module that recovers averaged nodal stresses at plate nodes.

forces including reactions are recovered by the matrix multiplication f = K.u, where K is the unmodified
master stiffness.
Finally, plate stresses and bar internal forces are recovered by the modules PlaneStressPlateStresses
and PlaneStressBarForces, respectively. The former is listed in Figure 27.8. This computation is actually
part of the postprocessing stage. It is not described here since stress recovery is treated in more detail in a
subsequent Chapter.
The bar force recovery module is not described here as it has not been used in assigned problems so far.

§27.5. Postprocessing

As noted above, module PlaneStressSolution carries out preprocessing tasks: recover node
forces, plate stresses and bar forces. But those are not under control of the user. Here we describe
result printing and plotting activities that can be specified in the problem script.

27–14
27–15 §27.5 POSTPROCESSING

§27.5.1. Result Print Utilities

Several utilities are provided in the plane stress program to print solution data in tabular form. They are
invoked as follows.
To print computed node displacements:
PrintPlaneStressNodeDisplacements[NodeDisplacements,title,digits];
To print receoverd node forces including reactions:
PrintPlaneStressNodeForces[NodeForces,title,digits];
To print plate node stresses:
PrintPlaneStressPlateNodeStresses[NodePlateCounts,NodePlateStresses,title,digits];
To print node displacements, node forces and node plate stresses in one table:
PrintPlaneStressSolution[NodeDisplacements,NodeForces,
NodePlateCounts,NodePlateStresses,title,digits];
No utility is provided to print bar forces, as problems involving plates and bars had not been assigned.
In all utilities listed above, title is an optional character string to be printed as a title before the table; for
example "Node displacements for plate with a hole". To eliminate the title, specify "" (two quote
marks together).
The last argument of the print modules: digits, is optional. If set to { d,f } it specifies that floating point
numbers are to be printed with room for at least d digits, with f digits after the decimal point. If digits is
specified as a void list: { }, a preset default is used for d and f.

§27.5.2. Displacement Field Contour Plots

Contour plots of displacement components u x and u y (interpolated over elements from the computed node
displacements) may be produced. Displacement magnitudes are shown using a internally set color scheme
based on “hue interpolation”, in which white means zero. Plots can be obtained using a script typified by

ux=Table[NodeDisplacements[[n,1]],{n,numnod}];
uy=Table[NodeDisplacements[[n,2]],{n,numnod}];
{uxmax,uymax}=Abs[{Max[ux],Max[uy]}];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,ux,
uxmax,Nsub,aspect,"Displacement component ux"];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,uy,
uymax,Nsub,aspect,"Displacement component uy"];
The third argument of ContourPlotNodeFuncOver2DMesh is the function to be plotted, specified at nodes.
A function maximum (in absolute value) is supplied as fourth argument. The reason for supplying this from the
outside is that in many cases it is convenient to alter the actual maximum for zooming or animation purposes.
As for the other arguments, Nsub is the number of element subdivisions in each direction when breaking down
the element area into plot polygons. Typically Nsub is set to 4 at the start of the script and is the same for
all plots of a script. Plot smoothness in terms of color grading increases with Nsub, but plot time grows as
Nsub-squared. So an appropriate tradeoff is to use a high Nsub, say 8, for coarse meshes containing few
elements whereas for finer meshes a value of 4 or 2 may work fine. The aspect argument specifies the ratio
between the y (vertical) and x (horizontal) dimensions of the plot frame, and is usaully defined at the script
start as a symbol so it is the same for all plots. The last argument is a plot title.

27–15
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–16

Nodal stress sig- xx Nodal stress sig- yy Nodal stress sig- xy

Figure 27.9. Stress component contour plots for rectangular plate with
a circular central hole, Model (I) of Figure 27.2(b).

§27.5.3. Stress Field Contour Plots

Contour plots of stress components σx x , σ yy and σx y (interpolated over elements from the computed node
displacements) may be produced. Stress magnitudes are shown using a internally set color scheme based on
“hue interpolation”, in which white means zero. Plots can be obtained using a script typified by

sxx=Table[NodePlateStresses[[n,1]],{n,numnod}];
syy=Table[NodePlateStresses[[n,2]],{n,numnod}];
sxy=Table[NodePlateStresses[[n,3]],{n,numnod}];
{sxxmax,syymax,sxymax}=Abs[{Max[sxx],Max[syy],Max[sxy]}];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,
sxx,sxxmax,Nsub,aspect,"Nodal stress sig-xx"];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,
syy,syymax,Nsub,aspect,"Nodal stress sig-yy"];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,
sxy,sxymax,Nsub,aspect,"Nodal stress sig-xy"];
For a description of ContourPlotNodeFuncOver2DMesh arguments, see the previous subsection.
As an example, stress component contour plots for Model (I) of the rectangular plate with a circular central
hole are shown in Figure 27.9.

Remark 27.1. Sometimes it is useful to show contour plots of principal stresses, Von Mises stresses, maximum
shears, etc. To do that, the appropriate function should be constructed first from the Cartesian components
available in NodePlateStresses, and then the plot utility called.

§27.5.4. Animation
Occasionally it is useful to animate results such as displacements or stress fields when a load changes as a
function of a time-like parameter t.
This can be done by performing a sequence of analyses in a For loop. Such sequence produces a series of
plot frames which may be animated by doubly clicking on one of them. The speed of the animation may be
controlled by pressing on the righmost buttons at the bottom of the Mathematica window. The second button
from the left, if pressed, changes the display sequence to back and forth motion.

27–16
27–17 §27.6 A COMPLETE PROBLEM SCRIPT CELL

ClearAll[Em,ν,th,aspect,Nsub];
Em=10000; ν=.25; th=3; aspect=6/5; Nsub=4;
Emat=Em/(1-ν^2)*{{1,ν,0},{ν,1,0},{0,0,(1-ν)/2}};

(* Define FEM model *)

s={1,0.70,0.48,0.30,0.16,0.07,0.0};
xy1={0,6}; xy7={0,1}; xy8={2.5,6}; xy14={Cos[3*Pi/8],Sin[3*Pi/8]};
xy8={2.5,6}; xy21={Cos[Pi/4],Sin[Pi/4]}; xy15={5,6};
xy22={5,2}; xy28={Cos[Pi/8],Sin[Pi/8]}; xy29={5,0}; xy35={1,0};
NodeCoordinates=Table[{0,0},{35}];
Do[NodeCoordinates[[n]]=N[s[[n]] *xy1+(1-s[[n]]) *xy7],{n,1,7}];
Do[NodeCoordinates[[n]]=N[s[[n-7]] *xy8+(1-s[[n-7]]) *xy14],{n,8,14}];
Do[NodeCoordinates[[n]]=N[s[[n-14]]*xy15+(1-s[[n-14]])*xy21],{n,15,21}];
Do[NodeCoordinates[[n]]=N[s[[n-21]]*xy22+(1-s[[n-21]])*xy28],{n,22,28}];
Do[NodeCoordinates[[n]]=N[s[[n-28]]*xy29+(1-s[[n-28]])*xy35],{n,29,35}];
PrintPlaneStressNodeCoordinates[NodeCoordinates,"",{6,4}];
ElemNodes=Table[{0,0,0,0},{24}];
ElemNodes[[1]]={1,2,9,8};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,2,6}];
ElemNodes[[7]]=ElemNodes[[6]]+{2,2,2,2};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,8,12}];
ElemNodes[[13]]=ElemNodes[[12]]+{2,2,2,2};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,14,18}];
ElemNodes[[19]]=ElemNodes[[18]]+{2,2,2,2};
Do [ElemNodes[[e]]=ElemNodes[[e-1]]+{1,1,1,1},{e,20,24}];
numnod=Length[NodeCoordinates]; numele=Length[ElemNodes];

ElemTypes= Table["Quad4",{numele}];
PrintPlaneStressElementTypeNodes[ElemTypes,ElemNodes,"",{}];
ElemMaterials= Table[Emat, {numele}];
ElemFabrications=Table[th, {numele}];
(*PrintPlaneStressElementMatFab[ElemMaterials,ElemFabrications,"",{}];*)
NodeDOFValues=NodeDOFTags=Table[{0,0},{numnod}];
NodeDOFValues[[1]]=NodeDOFValues[[15]]={0,37.5};
NodeDOFValues[[8]]={0,75}; (* nodal loads *)
Do[NodeDOFTags[[n]]={1,0},{n,1,7}]; (* vroller @ nodes 1-7 *)
Do[NodeDOFTags[[n]]={0,1},{n,29,35}]; (* hroller @ node 4 *)
PrintPlaneStressFreedomActivity[NodeDOFTags,NodeDOFValues,"",{}];
ProcessOptions={True};
Plot2DElementsAndNodes[NodeCoordinates,ElemNodes,aspect,
"One element mesh - 4-node quad",True,True];

Figure 27.10. Preprocessing (model definition) script for problem of Figure 27.2(b).

§27.6. A Complete Problem Script Cell

A complete problem script cell (which is Cell 13 in the plane stress Notebook placed in the web
index of this Chapter) is listed in Figures 27.10 through 27.12. This driver cell does Model (I) of
the plate with hole problem of Figure 27.2(b), which uses 4-node quadrilateral elements.
The script is divided into 3 parts for convenience. Figure 27.10 lists the model definition script
followed by a mesh plot command. Figure 27.11 shows the call to PlaneStressSolution analysis
driver and the call to get printout of node displacements, forces and stresses. Finally, Figure 27.12
shows commands to produce contour plots of displacements (skipped) and stresses.
Other driver cell examples may be studied in the PlaneStress.nb Notebook posted on the course
web site. It can be observed that the processing and postprocessing scripts are largely the same.

27–17
Chapter 27: A COMPLETE PLANE STRESS FEM PROGRAM 27–18

(* Solve problem and print results *)

{NodeDisplacements,NodeForces,NodePlateCounts,NodePlateStresses,
ElemBarNumbers,ElemBarForces}= PlaneStressSolution[
NodeCoordinates,ElemTypes,ElemNodes,
ElemMaterials,ElemFabrications,
NodeDOFTags,NodeDOFValues,ProcessOptions];

PrintPlaneStressSolution[NodeDisplacements,NodeForces,NodePlateCounts,
NodePlateStresses,"Computed Solution:",{}];

Figure 27.11. Processing script and solution print for problem of Figure 27.2(b).

(* Plot Displacement Components Distribution - skipped *)

(* ux=Table[NodeDisplacements[[n,1]],{n,numnod}];
uy=Table[NodeDisplacements[[n,2]],{n,numnod}];
{uxmax,uymax}=Abs[{Max[ux],Max[uy]}];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,ux,
uxmax,Nsub,aspect,"Displacement component ux"];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,uy,
uymax,Nsub,aspect,"Displacement component uy"]; *)

(* Plot Averaged Nodal Stresses Distribution *)

sxx=Table[NodePlateStresses[[n,1]],{n,numnod}];
syy=Table[NodePlateStresses[[n,2]],{n,numnod}];
sxy=Table[NodePlateStresses[[n,3]],{n,numnod}];
{sxxmax,syymax,sxymax}={Max[Abs[sxx]],Max[Abs[syy]],Max[Abs[sxy]]};
Print["sxxmax,syymax,sxymax=",{sxxmax,syymax,sxymax}];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,
sxx,sxxmax,Nsub,aspect,"Nodal stress sig-xx"];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,
syy,syymax,Nsub,aspect,"Nodal stress sig-yy"];
ContourPlotNodeFuncOver2DMesh[NodeCoordinates,ElemNodes,
sxy,sxymax,Nsub,aspect,"Nodal stress sig-xy"];

Figure 27.12. Result plotting script for problem of Figure 27.2(b). Note that displacement
contour plots have been skipped by commenting out the code.

What changes is the model definition script portion, which often benefits from ad-hoc node and
element generation constructs.
Notes and Bibliography
Few FEM books show a complete program. More common is the display of snipets of code. These are left
dangling chapter after chapter, with no attempt at coherent interfacing. This historical problem comes from
early reliance on low-level languages such as Fortran. Even the simplest program may run to several thousand
code lines. At 50 lines/page that becomes difficult to display snugly in a textbook while providing a running
commentary.
Another ages-old problem is plotting. Languages such as C or Fortran do not include plotting libraries for
the simple reason that low-level universal plotting standards never existed. The situation changed when
widely used high-level languages like Matlab or Mathematica appeared. The language engine provides the
necessary fit to available computer hardware, concealing system dependent details. Thus plotting scripts
become transportable.

27–18
 28
.

Stress
Recovery

28–1
Chapter 28: STRESS RECOVERY 28–2

TABLE OF CONTENTS

Page
§28.1. Introduction 28–3
§28.2. Calculation of Element Strains and Stresses 28–3
§28.3. Direct Stress Evaluation at Nodes 28–4
§28.4. Extrapolation from Gauss Points 28–4
§28.5. Interelement Averaging 28–6

28–2
28–3 §28.2 CALCULATION OF ELEMENT STRAINS AND STRESSES

§28.1. Introduction

In this lecture we study the recovery of stress values for two-dimensional plane-stress elements.1
This analysis step is sometimes called postprocessing because it happens after the main processing
step — the calculation of nodal displacements — is completed. Stress calculations are of interest
because in structural analysis and design the stresses are often more important to the engineer than
displacements.
In the stiffness method of solution discussed in this course, the stresses are obtained from the
computed displacements, and are thus derived quantities. The accuracy of derived quantities is
generally lower than that of primary quantities (the displacements), an informal statement that may
be mathematically justified in the theory of finite element methods. For example, if the accuracy
level of displacements is 1% that of the stresses may be typically 10% to 20%, and even lower at
boundaries.
It is therefore of interest to develop techniques that enhance the accuracy of the computed stresses.
The goal is to “squeeze” as much accuracy from the computed displacements while keeping the
computational effort reasonable. These procedures receive the generic name stress recovery tech-
niques in the finite element literature. In the following sections we cover the simplest stress recovery
techniques that have been found most useful in practice.

§28.2. Calculation of Element Strains and Stresses

In elastic materials, stresses σ are directly related to strains e at each point through the elastic
constitutive equations σ = Ee. It follows that the stress computation procedure begins with strain
computations, and that the accuracy of stresses depends on that of strains. Strains, however, are
seldom saved or printed.
In the following we focus our attention on two-dimensional isoparametric elements, as the compu-
tation of strains, stresses and axial forces in bar elements is strightforward.
Suppose that we have solved the master stiffness equations

Ku = f, (28.1)

for the node displacements u. To calculate strains and stresses we perform a loop over all defined
elements. Let e be the element index of a specific two-dimensional isoparametric element encoun-
tered during this loop, and u(e) the vector of computed element node displacements. Recall from
§15.3 that the strains at any point in the element may be related to these displacements as

e = Bu(e) , (28.2)

where B is the strain-displacement matrix (14.18) assembled with the x and y derivatives of the
element shape functions evaluated at the point where we are calculating strains. The corresponding
stresses are given by
σ = Ee = EBu (28.3)

1 This Chapter needs rewriting to show the use of Mathematica for stress computation. To be done in the future.

28–3
Chapter 28: STRESS RECOVERY 28–4

Table 28.1 Natural Coordinates of Bilinear Quadrilateral Nodes

Corner ξ η ξ η Gauss ξ η ξ η
node node
√ √ √ √
1 −1 −1 −√3 −√3 1’ −1/√3 −1/√3 −1 −1
2 +1 −1 +√ 3 −√3 2’ +1/√3 −1/√3 +1 −1
3 +1 +1 +√ 3 +√3 3’ +1/√3 +1/√3 +1 +1
4 −1 +1 − 3 + 3 4’ −1/ 3 +1/ 3 −1 +1

Gauss nodes, and coordinates ξ  and η are defined in §28.4 and Fig. 28.1

In the applications it is of interest to evaluate and report these stresses at the element nodal points
located on the corners and possibly midpoints of the element. These are called element nodal point
stresses.
It is important to realize that the stresses computed at the same nodal point from adjacent elements
will not generally be the same, since stresses are not required to be continuous in displacement-
assumed finite elements. This suggests some form of stress averaging can be used to improve the
stress accuracy, and indeed this is part of the stress recovery technique further discussed in §28.5.
The results from this averaging procedure are called nodal point stresses.

For the moment let us see how we can proceed to compute element nodal stresses. Two approaches
are followed in practice:

1. Evaluate directly σ at the element node locations by substituting the natural coordinates of
the nodal points as arguments to the shape function modules. These modules return qx and q y
and direct application of (28.2)-(28.4) yields the strains and stresses at the nodes.

2. Evaluate σ at the Gauss integration points used in the element stiffness integration rule and
then extrapolate to the element node points.

Empirical evidence indicates that the second approach generally delivers better stress values for
quadrilateral elements whose geometry departs substantially from the rectangular shape. This is
backed up by “superconvergence” results in finite element approximation theory. For rectangular
elements there is no difference.
For isoparametric triangles both techniques deliver similar results (identical if the elements are
straight sided with midside nodes at midpoints) and so the advantages of the second one are
marginal. Both approaches are covered in the sequel.

§28.3. Direct Stress Evaluation at Nodes

This approach is straightforward and need not be discussed in detail.

28–4
28–5 §28.4 EXTRAPOLATION FROM GAUSS POINTS

η
(a) (b)
3 3

4 4 η' 3'
4'
(e) ξ (e')
ξ'
1'
1 1 2'
2 2

Figure 28.1. Extrapolation from 4-node quad Gauss points: (a) 2 × 2 rule, (b) Gauss element (e )

§28.4. Extrapolation from Gauss Points

This will again be explained for the four-node bilinear quadrilateral. The normal Gauss integration
rule for element stiffness evaluation is 2 × 2, as illustrated in Figure 28.1.
The stresses are calculated at the Gauss points, which are identified as 1 , 2 , 3 and 4 in Figure 28.1.
Point i  is closest to node i so it is seen that Gauss point numbering essentially follows element
node numbering in the counterclockwise sense. The natural coordinates of these points are listed
in Table 28.1. The stresses are evaluated at these Gauss points by passing these natural coordinates
to the shape function subroutine. Then each stress component is “carried” to the corner nodes 1
through 4 through a bilinear extrapolation based on the computed values at 1 through 4 .
To understand the extrapolation procedure more clearly it is convenient to consider the region
bounded by the Gauss points as an “internal element” or “Gauss element”. This interpretation is
depicted in Figure 28.1(b). The Gauss element, denoted by (e ), is also a four-node quadrilateral.
Its quadrilateral (natural) coordinates are denoted by ξ  and η . These are linked to ξ and η by the
simple relations
√ √ √ √
ξ = ξ  / 3, η = η / 3, ξ  = ξ 3, η = η 3. (28.4)
Any scalar quantity w whose values wi at the Gauss element corners are known can be interpolated
through the usual bilinear shape functions now expressed in terms of ξ  and η :
 (e ) 
N1


    
 (e ) 
  N2  
w(ξ , η ) = [ w1 w2 w3 w4 ]  (e )  , (28.5)
N 3

N4(e )
where (cf. §15.6.2)

N1(e ) = 14 (1 − ξ  )(1 − η ),

N2(e ) = 14 (1 + ξ  )(1 − η ),
 (28.6)
N3(e ) = 14 (1 + ξ  )(1 + η ),

N4(e ) = 14 (1 − ξ  )(1 + η ).

28–5
Chapter 28: STRESS RECOVERY 28–6

(a) (b) 3
3 7 (c) 3
7
7' 3'
7' 3' 4 3' 5
4 4'
4' 6' 6
8' 9 6
8' 9 9' 6' 6
8
8 2' 2'
1' 5' 1' 2
1' 5' 2'
5 2 1
2 1
1 5 4

Figure 28.2. Gauss elements for higher order quadrilaterals and triangles:
(a) 9-node element with 3 × 3 Gauss rule, (b) 8-node element with
3 × 3 Gauss rule, (c) 6-node element with 3-interior point rule.

√
To extrapolate w to corner 1, say, we replace its ξ  and η coordinates, namely ξ  = η = − 3,
into the above formula. Doing that for the four corners we obtain
√ √
  1 + 1 3 − 12 1 − 12 3 − 12
  
w1 2 √ √ w1

 w2   − 2√ 1 + 2 3
1 1
−2 1 
1 − 2 3   w2 
1
 = √   (28.7)
w3 1 − 12 3 − 12 1 + 12 3 − 12  w3
√ √
w4 −1 1− 1 3 −1 1+ 1 3 w4
2 2 2 2

Note that the sum of the coefficients in each row is one, as it should be. For stresses we apply this
formula taking w to be each of the three stress components, σx x , σ yy and τx y , in turn.
Extrapolation in Higher Order Elements
For eight-node and nine-node isoparametric quadrilaterals the usual Gauss integration rule is 3 × 3,
and the Gauss elements are nine-noded quadrilaterals that look as in Figure 28.2(a) and (b) above.
For six-node triangles the usual quadrature is the 3-point rule with internal sampling points, and
the Gauss element is a three-node triangle as shown in Figure 28.2(c).
§28.5. Interelement Averaging
The stresses computed in element-by-element fashion as discussed above, whether by direct evalu-
ation at the nodes or by extrapolation, will generally exhibit jumps between elements. For printing
and plotting purposes it is usually convenient to “smooth out” those jumps by computing averaged
nodal stresses. This averaging may be done in two ways:

(I) Unweighted averaging: assign same weight to all elements that meet at a node;
(II) Weighted averaging: the weight assigned to element contributions depends on the stress
component and the element geometry and possibly the element type.
Several weighted average schemes have been proposed in the finite element literature, but they do
require additional programming.

28–6
 30
Thermomechanical
Effects

30–1
Chapter 30: THERMOMECHANICAL EFFECTS

TABLE OF CONTENTS

Page
§30.1. Introduction 30–3
§30.2. Thermomechanical Behavior 30–3
§30.2.1. Thermomechanical Stiffness Relations . . . . . . . . . 30–4
§30.2.2. Globalization . . . . . . . . . . . . . . . . . 30–6
§30.2.3. Merge . . . . . . . . . . . . . . . . . . . . 30–6
§30.2.4. Solution . . . . . . . . . . . . . . . . . . 30–6
§30.2.5. Postprocessing . . . . . . . . . . . . . . . . . 30–6
§30.2.6. Worked-Out Examples . . . . . . . . . . . . . . 30–7
§30. Notes and Bibliography
. . . . . . . . . . . . . . . . . . . . . . 30–10
§30. References . . . . . . . . . . . . . . . . . . . . . . 30–10
§30. Exercises . . . . . . . . . . . . . . . . . . . . . . 30–11

30–2
 §30.2 THERMOMECHANICAL BEHAVIOR

§30.1. Introduction

The assumptions invoked in Chapters 2–3 for the example truss result in zero external forces under
zero displacements. This is implicit in the linear-homogeneous expression of the master stiffness
equation f = Ku. If u vanishes, so does f. This behavior does not apply, however, if there are initial
force effects.1 If those effects are present, there can be displacements without external forces, and
internal forces without displacements.
A common source of initial force effects are temperature changes. Imagine that a plane truss
structure is unloaded (that is, not subjected to external forces) and is held at a uniform reference
temperature. External displacements are measured from this environment, which is technically
called a reference state. Now suppose that the temperature of some members changes with respect
to the reference temperature while the applied external forces remain zero. Because the length
of members changes on account of thermal expansion or contraction, the joints will displace. If
the structure is statically indeterminate those displacements will induce strains and stresses and
thus internal forces. These are distinguished from mechanical effects by the qualifier “thermal.”
For many structures, particularly in aerospace and mechanical engineering, such effects have to be
considered in the analysis and design.
There are other physical sources of initial force effects, such as moisture (hygrosteric) effects,2
member prestress, residual stresses, or lack of fit. For linear structural models all such sources may
be algebraically treated in the same way as thermal effects. The treatment results in an initial force
vector that has to be added to the applied mechanical forces. This subject is outlined in §30.3 from
a general perspective. However, to describe the main features of the matrix analysis procedure it is
sufficient to consider the case of temperature changes.
In this Section we go over the analysis of a plane truss structure whose members undergo temperature
changes from a reference state. It is assumed that the disconnection and localization steps of the
DSM have been carried out. Therefore we begin with the derivation of the matrix stiffness equations
of a generic truss member.

§30.2. Thermomechanical Behavior

Consider the generic plane-truss member shown in Figure 30.1. The member is prismatic and
uniform. The temperature T is also uniform. To reduce clutter the member identification subscript
will be omitted in the following development until the globalization and assembly steps.
We introduce the concept of reference temperature Tr e f . This is conventionally chosen to be the
temperature throughout the structure at which the displacements, strains and stresses are zero if
no mechanical forces are applied. In structures such as buildings and bridges Tr e f is often taken
to be the mean temperature during the construction period. Those zero displacements, strains and
stresses, together with Tr e f , define the thermomechanical reference state for the structure.
The member temperature variation from that reference state is T = T −Tr e f . This may be positive
or negative. If the member is disassembled or disconnected, under this variation the member length

1 Called initial stress or initial strain effects by many authors. The different names reflect what is viewed as the physical
source of initial force effects at the continuum mechanics level.
2 These are important in composite materials and geomechanics.

30–3
Chapter 30: THERMOMECHANICAL EFFECTS

is free to change from L to L + dT . If the thermoelastic constitutive behavior is linear3 then dT is

proportional to L and T :
dT = αL T. (30.1)

Here α is the coefficient of thermal expansion, which has physical units of one over temperature.
This coefficient will be assumed to be uniform over the generic member. It may, however, vary
from member to member.
The thermal strain is defined as
eT = dT /L = α T. (30.2)

Now suppose that the member is also subject to mechanical forces, more precisely the applied
axial force F shown in Figure 30.1. The member axial stress is σ = F/A. In response to
this stress the length changes by d M . The mechanical strain is e M = d M /L. The total strain
e = d/L = (d M + dT )/L is the sum of the mechanical and the thermal strains:

σ
e = e M + eT = + αT (30.3)
E

This superposition of deformations is the basic assumption made in the thermomechanical analysis.
It is physically obvious for an unconstrained member such as that depicted in Figure 30.1.

(a) f yi, uyi f yj , uyj (b)

ȳ

fxi , uxi (e) x̄ fxj , uxj

i j

k s = EA/L F
α, T F
d = d M + dT
L
L + d M + dT

Figure 30.1. Generic truss member subjected to mechanical and thermal effects:
(a) idealization as bar, (b) idealization as equivalent linear spring.

At the other extreme, suppose that the member is completely blocked against axial elongation; that
is, d = 0 but T = 0. Then e = 0 and e M = −eT . If α > 0 and T > 0 the blocked member
goes in compression because σ = Ee M = −EeT = −Eα T < 0. This thermal stress is further
discussed in Remark 30.2.

3 An assumption justified if the temperature changes are small enough so that α is approximately constant through the
range of interest, and no material phase change effects occur.

30–4
 §30.2 THERMOMECHANICAL BEHAVIOR

F = − fxi p = pM + pT F = f xj

i j

Figure 30.2. Equilibrium of truss member under thermomechanical forces.

§30.2.1. Thermomechanical Stiffness Relations

Because e = d/L and d = ū x j − ū xi , (30.3) can be developed as
ū x j − ū xi σ
= + α T, (30.4)
L E
To pass to internal forces (30.4) is multiplied through by E A:
EA
(ū x j − ū xi ) = Aσ + E A α T = p M + pT = p = F. (30.5)
L
Here p M = Aσ denotes the mechanical axial force, and pT = E A α T , which has the dimension
of a force, is called (not surprisingly) the internal thermal force. The sum p = p M + pT is called
the effective internal force. The last relation in (30.5), F = p = p M + pT follows from free-body
member equilibrium; see Figure 30.2. Passing to matrix form:
 
ū xi
EA  ū 
F= [ −1 0 1 0 ]  yi  . (30.6)
L ū yi
ū y j
Noting that F = f¯x j = − f¯xi while f¯yi = f¯y j = 0, we can relate joint forces to joint displacements
as
 ¯          
f xi −F f¯M xi −1 1 0 −1 0 ū xi
 f¯yi   0   f¯M yi   0  EA  0 0 0 0   ū yi 
 f¯  =  F  =  f¯M x j 
+ E A α T 
1 
=  −1 0 1 0   ū x j 
. (30.7)
xj L
f¯y j 0 f¯M y j 0 0 0 0 0 ū y j

In compact matrix form this is f̄ = f̄ M + f̄T = K̄ū, or

K̄ū = f̄ M + f̄T . (30.8)

Here K̄ is the same member stiffness matrix derived in Chapter 2. The new ingredient that appears
is the vector  
−1
 0 
f̄T = E A α T  , (30.9)
1
0
This is called the vector of thermal joint forces in local coordinates. It is an instance of an initial
force vector at the element level.

30–5
Chapter 30: THERMOMECHANICAL EFFECTS

Remark 30.1. A useful physical interpretation of (30.8) is as follows. Suppose that the member is precluded
from joint (node) motions so that ū = 0. Then f̄ M + f̄T = 0 or f̄ M = −f̄T . It follows that fT contains the
negated joint forces (internal forces) that develop in a heated or cooled bar if joint motions are precluded.
Because for most materials α > 0, rising the temperature of a blocked bar:T > 0, produces an internal
compressive thermal force pT = AσT = −E AαT , in accordance with the expected physics. The quantity
σT = −Eα T is the thermal stress. This stress can cause buckling or cracking in severely heated structural
members that are not allowed to expand or contract. This motivates the use of expansion joints in pavements,
buildings and rails, and roller supports in long bridges.

§30.2.2. Globalization

At this point we restore the member superscript so that the member stiffness equations (30.7) are
rewritten as
e e e
K̄ ūe = f̄ M + f̄T . (30.10)

Use of the transformation rules developed in Chapter 2 to change displacements and forces to the
global system {x, y} yields
Ke ue = feM + feT , (30.11)

where Te is the displacement transformation matrix (3.1), and the transformed quantities are
 T e  T e  T e
Ke = Te K̄ Te , feM = Te f̄ , feT = Te f̄T . (30.12)

These globalized member equations are used to assemble the free-free master stiffness equations
by a member merging process.

§30.2.3. Merge

The merge process is based on the same assembly rules stated in §3.1.3 with only one difference:
thermal forces are added to the right hand side. The member by member merge is carried out
much as described as in §3.1.4, the main difference being that the thermal force vectors feT are also
merged into a master thermal force vector. Force merge can be done by augmentation-and add (for
hand work) or via freedom pointers (for computer work). Illustrative examples are provided below.
Upon completion of the assembly process we arrive at the free-free master stiffness equations

Ku = f M + fT = f. (30.13)

§30.2.4. Solution

The master system (30.13) has formally the same configuration as the master stiffness equations
(2.3). The only difference is that the effective joint force vector f contains a superposition of
mechanical and thermal forces. Displacement boundary conditions can be applied by reduction
or modification of these equations, simply by using effective joint forces in the descriptions of
Chapters 2–3. Processing the reduced or modified system by a linear equation solver yields the
displacement solution u.

30–6
 §30.2 THERMOMECHANICAL BEHAVIOR

§30.2.5. Postprocessing
The postprocessing steps described in Chapter 3 require some modifications because the derived
quantities of interest to the structural engineer are mechanical reaction forces and internal forces.
Effective forces by themselves are of little use in design. Mechanical joint forces including reactions
are recovered from
f M = Ku − fT (30.14)
To recover mechanical internal forces in member e, compute p e by the procedure outlined in Chapter
3, and subtract the thermal component:
p eM = p e − E e Ae α e T e . (30.15)
This equation comes from solving (30.5) for p M . The mechanical axial stress is σ e = p eM /Ae .
§30.2.6. Worked-Out Examples

Example 30.1. The first worked out problem is defined in Figure 30.3. Two truss members are connected
in series as shown and fixed at the ends. Properties E = 1000, A = 5 and α = 0.0005 are common to
both members. The member lengths are 4 and 6. A mechanical load P = 90 acts on the roller node. The
temperature of member (1) increases by T (1) = 25◦ while that of member (2) drops by T (2) = −10◦ . Find
the stress in both members.
Ο
E = 1000, A = 12, α = 0.0005 1/ C for both
Ο
members. ∆T(1) = 25, ∆T(2) = −10

P = 90
ȳ // y 1 2 3
;;
;;
;;
;;
; x̄ // x L
(1)
(1)

=4 L =6
(2)

(2)

Figure 30.3. Structure for worked-out Example 1.

To reduce clutter note that all y motions are suppressed so only the x freedoms are kept: u x1 = u 1 , u x2 = u 2
and u x3 = u 3 . The corresponding node forces are denoted by f x1 = f 1 , f x2 = f 2 and f x3 = f 3 . The thermal
force vectors, stripped to their x̄ ≡ x components, are

f¯(1) 


f¯(2) 


(1) −1 −150 (2) −1 60
f̄T = ¯T(1) 1
= E (1)A(1)α (1) T (1) = , f̄T = ¯T(2) 2
= E (2)A(2)α (2) T (2) = .
fT 2 1 150 fT 3 1 −60
(30.16)
The element stiffness equations are:


ū (1) 
f¯(1) 


ū (2) 
f¯(2) 

1 −1 −150 1 −1 60
3000 1
(1) = ¯M1 (1) + , 2000 2
(2) = ¯M2 (2) + , (30.17)
−1 1 ū 2 f M2 150 −1 1 ū 3 f M3 −60
No globalization is needed because the equations are already in the global system, and thus we can get rid of
the localization marker symbols: f¯ → f , ū → u. Assembling by any method yields
     
3 −3 0 u1 f M1 −150 f M1 −150
1000 −3 5 −2 u2 = f M2 + 150 + 60 = f M2 + 210 . (30.18)
0 −2 2 u3 f M3 −60 f M3 −60

30–7
Chapter 30: THERMOMECHANICAL EFFECTS

The displacement boundary conditions are u 1 = u 3 = 0. The mechanical force boundary condition is f M2 =
90. On removing the first and third equations, the reduced system is 5000 u 2 = f M2 + 210 = 90 + 210 = 300,
which yields u 2 = 300/5000 = +0.06. The mechanical internal forces in the members are recovered from

E (1) A(1)
p (1)
M = (u 2 − u 1 ) − E (1)A(1) α (1) T (1) = 3000 × 0.06 − 12000 × 0.0005 × 25 = 60,
L (1)
E (2) A(2)
p (2)
M = (u 3 − u 2 ) − E (2)A(2) α (2) T (2) = 2000 × (−0.06) − 12000 × 0.0005 × (−10) = −72,
L (2)
(30.19)
whence the stresses are σ (1) = 60/12 = 5 and σ (2) = −72/12 = −6. Member (1) is in tension and member
(2) in compression.

Example 30.2. The second example concerns the example truss of Chapters 2-3. The truss is mecanically
unloaded, that is, f M x2 = f M x3 = f M y3 = 0. However the temperature of members (1) (2) and (3) changes
by T , −T and 3T , respectively, with respect to Tr e f . The thermal expansion coefficient of all three
members is assumed to be α. We will perform the analysis keeping α and T as variables.
The thermal forces for each member in global coordinates are obtained by using (30.10) and the third of
(30.12):
    
0 0 01 −1 −1

1 0 0 0
0  0
f(1) = E (1) A(1) α (1) T (1)  = 100α T  ,
T 0 1 0 1
0 1
0 0 10 0 0
    
−1 0 0 0 −1 0

0 01 0 0   −1 
f(2) = E (2) A(2) α (2) T (2)  = 50α T  , (30.20)
T 0 0 −1   1 
0 0
0 10 0 0 1
    
1 −1 0 0 −1 −1
(3) 1  1 1 0 0 0  −1 
f(3) (3) (3) (3)
= E A α T √  = 200α T  .
T
2 0 0 1 −1   1  1 
0 0 1 1 0 1

Merging the contribution of these 3 members gives the master thermal force vector
 −100 + 0 − 200   −300 
 0 + 0 − 200   −200 
 100 + 0 + 0   
fT = α T 
 0 − 50 + 0
 = α T  100 
  −50  (30.21)
   
0 + 0 + 200 200
0 + 50 + 200 250

The master stiffness matrix K does not change. Consequently the master stiffness equations are
 20 10 −10 0 −10 −10 u x1 = 0
f M x1
    −300 
 10 10 0 0 −10 −10   u y1 = 0   f M y1   −200 
 −10 0      100 
 0 10 0 0   u x2  =  f M x2 = 0  + α T   (30.22)
 0 0 −5       −50 
 0 0 5
  u y2 = 0   f M y2   
−10 −10 0 0 10 10 u x3 f M x3 = 0 200
−10 −10 0 −5 10 15 u y3 f M y3 = 0 250

30–8
 §30.2 THERMOMECHANICAL BEHAVIOR

in which f M x1 , f M y1 and f M y2 are the unknown mechanical reaction forces, and the known forces and dis-
placements have been marked. Since the prescribed displacements are zero, the reduced system is simply
    
10 0 0 u x2 0 100 100
0 10 10 u x3 = 0 + α T 200 = α T 200 . (30.23)
0 10 15 u y3 0 250 250
Solving (30.23) gives u x2 = u x3 = u y3 = 10α T . Completing u with the prescribed zero displacements and
premultiplying by K gives the complete effective force vector:
 20 10 −10 0 −10 −10 0
   −300 
 10 10 0 0 −10 −10   0   −200 
 −10 0    100 
f = Ku =    10  α T = α T  
0 10 0 0
 0 0 0 5 0 −5  0 
  −50  . (30.24)
    
−10 −10 0 0 10 10 10 200
−10 −10 0 −5 10 15 10 250
But this vector is exactly fT . Consequently

f M = Ku − fT = 0. (30.25)

All mechanical joint forces, including reactions, vanish, and so do the internal mechanical forces. This is a
consequence of the example frame being statically determinate.4

Initial Force Effects

As previously noted, a wide spectrum of mechanical and non-mechanical effects can be acommodated under
the umbrella of the initial force concept. The stiffness equations at the local (member) level are
e e e e
K̄ ūe = f̄ M + f̄ I = f̄ , (30.26)

and at the global (assembled structure) level:

Ku = f M + f I = f. (30.27)

In these equations subscripts M and I identify mechanical and initial node forces, respectively. The sum of
the two: f̄ at the local member level and f at the global structure level, are called effective forces.
A physical interpretation of (30.26) can be obtained by considering that the structure is blocked against all
motions: u = 0. Then f M = −f I , and the undeformed structure experiences mechanical forces. These
translate into internal forces and stresses. Engineers also call these prestresses. ‘prestress
Local effects that lead to initial forces at the member level are: temperature changes (studied in §4.2, in which
f I ≡ fT ), moisture diffusion, residual stresses, lack of fit in fabrication, and in-member prestressing. Global
effects include prescribed nonzero joint displacements (studied in §4.1) and multimember prestressing (for
example, by cable pretensioning of concrete structures).
As can be seen there is a wide variety of physical effects, whether natural or artificial, that lead to nonzero
initial forces. The good news is that once the member equations (30.25) are formulated, the remaining DSM
steps (globalization, merge and solution) are identical. This nice property extends to the general Finite Element
Method.
4 For the definition of static determinacy, see any textbook on Mechanics of Materials. Such structures do not develop
thermal stresses under any combination of temperature changes.

30–9
Chapter 30: THERMOMECHANICAL EFFECTS

Pseudo Thermal Inputs

Some commercial FEM programs do not have a way to handle directly effects such as moisture, lack of fit, or
prestress. But all of them can handle temperature variation inputs. Since in linear analysis all such effects can
be treated as initial forces, it is possible (at least for bar elements) to model them as fictitious thermomechanical
effects, by inputting phony temperature changes. The following example indicate that this is done for a bar
element.
Suppose that a prestress force FP is present in a bar. The total elongation is d = d M + d P where d P =
FP L/(E A) is due to prestress. Equate to a thermal elongation: dT = αTP L and solve for TP =
FP /(E Aα). This is input to the program as a fictitious temperature change. If in addition there is a real
temperature change T one would of course specify T + TP .
If this device is used, care should be exercised in interpreting results for internal forces and stresses given by
the program. The trick is not necessary for personal or open-source codes over which you have full control.
Notes and Bibliography
The additional DSM topics treated in this Chapter are covered in virtually all books on Matrix Structural
Analysis, such as the often quoted one by Przemieniecki [575]. Several recent FEM books ignore these topics
as too elementary.
The physics of thermomechanics and the analysis of thermal stresses is covered adequatedly in textbooks such
as Boley and Wiener [95], or manuals such as the widely used by Roark et. al. [620].
For the separate problems of heat conduction and heat transfer, the book by Özişik [516] provides a com-
prehensive classic treatment. There is a vast literature on prestressed structures; search under “prestress” in
http://www3.addall.com.
The concepts of static determinacy and its counterpart: static indeterminacy, are important in skeletal structures
such as trusses and frameworks. The pertinent design tradeoff is: insensitivity to initial force effects versus
redundant safety. A discussion of this topic is beyond the scope of the book. Once going past skeletal structural
systems, however, indeterminacy is the rule.

References
Referenced items have been moved to Appendix R.

30–10
 Exercises

Homework Exercises for Chapter 30

Thermomechanical Effects

EXERCISE 30.1 [N:20] Use the same data of Exercise 3.7, except that P = 0 and consequently there are no
applied mechanical forces. Both members have the same dilatation coefficient α = 10−6 1/◦ F. Find the crown
displacements u x2 and u y2 and the member stresses σ (1) and σ (2) if the temperature of member (1) rises by
T = 120 ◦ F above Tr e f , whereas member (2) stays at Tr e f .
Shortcut: the element stiffnesses and master stiffness matrix are the same as in Exercise 3.7, so if that Exercise
has been previously assigned no stiffness recomputations are necessary.

EXERCISE 30.2 [A:15] Consider the generic truss member of §2.4, reproduced in Figure E30.1 for conve-
nience.

_ _ _ _
fyi, u yi _ fyj, u yj
y
_ _ _ _ _
fxi , uxi (e) x fxj , uxj
i j
Figure E30.1. Generic truss member.

The disconnected member was supposed to have length L, but because of lack of quality control it was
fabricated with length L + δ, where δ is called the “lack of fit.” Determine the initial force vector f̄ I to be used
in (30.26). Hint: find the mechanical forces that would compensate for δ and restore the desired length.

EXERCISE 30.3 [A:10] Show that the lack of fit of the foregoing exercise can be viewed as equivalent to a
prestress force of −(E A/L)δ.

EXERCISE 30.4 [A:20] Show that prescribed nonzero displacements can, albeit somewhat artificially, be
placed under the umbrella of initial force effects. Work this out for the example of §30.1.1. Hint: split node
displacements into u = u H + u N , where u N (the “nonhomogeneous” or “particular” part of the solution)
carries the nonzero displacement values.

EXERCISE 30.5 [A:35]. (Research paper level). Prove that any statically determinate truss structure is free
of thermal stresses.

30–11