1 2kiz6.textbook - Of.finite - Element.Analysis - by.P..Seshu PDF

Textbook of
Finite Element
Analysis
P. Seshu
Textbook of
Finite Element Analysis
P. Seshu
Professor
Department of Mechanical Engineering
Indian Institute of Technology Bombay
Mumbai
NEW DELHI
2012
Rs. 250.00
TEXTBOOK OF FINITE ELEMENT ANALYSIS

P. Seshu
2003 by PHI Learning Private Limited, New Delhi. All rights reserved. No part of this book may be
reproduced in any form, by mimeograph or any other means, without permission in writing from the
publisher.
ISBN-978-81-203-2315-5
The export rights of this book are vested solely with the publisher.
Tenth Printing L L January, 2012
Published by Asoke K. Ghosh, PHI Learning Private Limited, M-97, Connaught Circus,
New Delhi-110001 and Printed by Mohan Makhijani at Rekha Printers Private Limited,
New Delhi-110020.
Respectfully Dedicated to
My Parents and Teachers
Contents
Preface ix
1. Introduction 115
1.1 Typical Application Examples 4
1.1.1 Automotive Applications 4
1.1.2 Manufacturing Process Simulation 7
1.1.3 Electrical and Electronics Engineering Applications 8
1.1.4 Aerospace Applications 14
Summary 14
2. Finite Element Formulation Starting from Governing Differential

Equations 1665
2.1 Weighted Residual MethodUse of a Single Continuous Trial
Function 16
2.2 The General Weighted Residual (WR) Statement 28
2.3 Weak (Variational) Form of the Weighted Residual Statement 33
2.4 Comparison of Differential Equation, Weighted Residual and Weak
Forms 36
2.5 Piece-wise Continuous Trial Function Solution of the Weak Form 41
2.6 One-dimensional Bar Finite Element 48
2.7 One-dimensional Heat Transfer Element 57
Summary 61
Problems 61
3. Finite Element Formulation Based on Stationarity of a Functional 66 88

3.1 Introduction 66
3.2 Functional and Differential Equation Forms 67
3.3 Principle of Stationary Total Potential (PSTP) 73
3.3.1 RayleighRitz Method 75
v
vi Contents
3.4 Piece-wise Continuous Trial FunctionsFinite Element Method 81

3.4.1 Bar Element Formulated from the Stationarity of a Functional 81
3.4.2 One-dimensional Heat Transfer Element Based on the Stationarity of a
Functional 83
3.5 Meaning of Finite Element Equations 84
Summary 87
Problems 88
4. One-dimensional Finite Element Analysis 89144

4.1 General Form of the Total Potential for 1-d 89
4.2 Generic Form of Finite Element Equations 90
4.3 The Linear Bar Finite Element 93
4.4 The Quadratic Bar Element 101
4.4.1 Determination of Shape Functions 101
4.4.2 Element Matrices 102
4.5 Beam Element 117
4.5.1 Selection of Nodal d.o.f. 117
4.5.2 Determination of Shape Functions 118
4.5.3 Element Matrices 119
4.6 Frame Element 125
4.7 One-dimensional Heat Transfer 132
Summary 138
Problems 138
5. Two-dimensional Finite Element Analysis 145231

5.1 IntroductionDimensionality of a Problem 145
5.2 Approximation of Geometry and Field Variable 148
5.2.1 Simple Three-noded Triangular Element 149
5.2.2 Four-noded Rectangular Element 152
5.2.3 Six-noded Triangular Element 153
5.3 Natural Coordinates and Coordinate Transformation 156
5.3.1 Alternate Methods of Deriving Shape Functions 157
5.3.2 Natural CoordinatesQuadrilateral Elements 159
5.3.3 Natural CoordinatesTriangular Elements 164
5.4 2-d Elements for Structural Mechanics 167
5.4.1 Generic Relations 167
5.4.2 Three-noded Triangular Element 171
5.4.3 Four-noded Rectangular Element 179
5.4.4 Compatibility of Displacements 181
5.4.5 Four-node Quadrilateral Element 183
5.4.6 Eight-node Quadrilateral Element 188
5.4.7 Nine-node Quadrilateral Element 190
5.4.8 Six-node Triangular Element 192
Contents vii
5.5 Numerical Integration 194

5.5.1 Trapezoidal Rule 195
5.5.2 Simpsons 1/3 Rule 196
5.5.3 NewtonCotes Formula 197
5.5.4 Gauss Quadrature Formula 198
5.5.4 Gauss Quadrature in Two Dimensions 201
5.6 Incorporation of Boundary Conditions 205
5.7 Solution of Static Equilibrium Equations 206
5.8 2-d Fluid Flow 220
Summary 225
Problems 226
6. Dynamic Analysis Using Finite Elements 232294

6.1 Introduction 232
6.2 Vibration Problems 232
6.3 Equations of Motion Based on Weak Form 235
6.3.1 Axial Vibration of a Rod 235
6.3.2 Transverse Vibration of a Beam 237
6.4 Equations of Motion Using Lagranges Approach 240
6.4.1 Formulation of Finite Element Equations 242
6.4.2 Consistent Mass Matrices for Various Elements 245
6.5 Consistent and Lumped Mass Matrices 246
6.5.1 HRZ Lumping Scheme 247
6.6 Form of Finite Element Equations for Vibration Problems 253
6.7 Some Properties of Eigenpairs 255
6.8 Solution of Eigenvalue Problems 257
6.8.1 Transformation Based Methods 258
6.8.2 Vector Iteration Methods 264
6.9 Transient Vibration Analysis 272
6.9.1 Modelling of Damping 272
6.9.2 The Mode Superposition Scheme 275
6.9.3 Direct Integration Methods 279
6.10 Thermal TransientsUnsteady Heat Transfer in a Pin-Fin 289
Summary 293
Problems 293
7. Application Examples 295307

7.1 Finite Element Analysis of Crankshaft Torsional Vibrations 295
7.1.1 Beam Element Model of Crankshaft Assembly 296
7.1.2 Results and Discussion 299
7.1.3 Dynamic Response Analysis 301
7.2 Axisymmetric Finite Element Analysis of a Pressure Vessel 303
7.2.1 Finite Element Formulation for Axisymmetric Loads 304
7.2.2 Stress Analysis of a Pressure Vessel 305
viii Contents
Appendix ASuggested Mini-Project Topics 309320

Project 1: Thermal Analysis of a Pressure Vessel 309
Project 2: Structural Dynamic Analysis of a Pressure Vessel 310
Project 3: Dynamics of a Scooter Frame 312
Project 4: Automotive Chassis Dynamics 313
Project 5: Analysis of a Turbine Disk 316
Project 6: Dynamic Analysis of a Building 317
Project 7: Thermal Analysis of an IC Engine Cylinder 318
Project 8: Stress Concentration 319
Project 9: Dynamics of a Hard Disk Drive Read/Write Head Assembly 319
Appendix BReview of Preliminaries 321323
B1.1 Matrix Algebra 321
B1.2 Interpolation 322
Appendix CTypical Finite Element Program 324 328
Index 329330
Preface
Many excellent books have been written on Finite Element Analysis, for example, books
authored by R.D. Cook, D.S. Malkus and M.E. Plesha; J.N. Reddy; K.J. Bathe; O.C. Zienkiewicz,
K. Morgan, et al. I have immensely benefitted from these books, both as a student and as a
teacher. Most books, however, present finite element method (FEM) primarily as an extension
of matrix methods of structural analysis (except, for example, the ones by J.N. Reddy and
Huebner). Present-day applications of FEM, however, range from structures to bio-mechanics
to electromagnetics. Thus the primary aim of the present book, based on several years of
teaching the course at Indian Institute of Technology (IIT) Bombay, is to present FEM as a
general tool to find approximate solutions to differential equations. This approach should give
the student a better perspective on the technique and its wide range of applications. Finite
element formulation, based on stationarity of a functional, is also discussed, and these two
forms are used throughout the book.
Over the past few years, several universities and engineering institutes in India have
introduced a one-semester course on finite element method at the senior undergraduate level.
The material presented in this book covers the syllabus of most such courses. Several worked-
out examples, drawn from the fields of structural mechanics, heat transfer and fluid flow,
illustrate the important concepts. Some of the issues are in fact brought out through example
problems. At the same time, some problems to investigate given at the end of each chapter
encourage the student to think beyond what has been presented in the book.
FEM is a technique (numerical tool), and the various nuances are best mastered by
attempting challenging real-life problems. While teaching the course at IIT Bombay, I have
successfully used two types of term projects (running through the semester)one where a
near-real-life problem is modelled using a commercial finite element analysis software and the
other where the student attempts to develop his own code and verify the same on simple text-
book type problems. Some topics for the former are suggested in Appendix A and a sample
code is given in Appendix C for isoparametric, nine-node quadrilateral element.
I am indebted to IIT Bombay and my colleagues (particularly, Profs. C. Amarnath,
Bharat Seth, and Kurien Issac) in the Mechanical Engineering Department for providing a
congenial work environment. The financial support provided by the Quality Improvement
ix
x Preface
Contents
Programme (QIP) at IIT Bombay for the preparation of the manuscript is gratefully
acknowledged.
My students Anurag Ganguli, Vignesh Raja, Gaurav Sharma, Kartik Srinivasan and V.K. Gupta
did a splendid job by studiously going through the entire manuscript and offering their critical
and incisive comments. I am also thankful to my student Pankaj Langote for assisting me in
getting the diagrams drawn and to Mrs. Padma Amin for keying in the whole manuscript. I
would like to thank my wife Uma and my little jewels Soumya and Saket for their continuous
support and encouragement. Preparation of the manuscript took longer than I expected and I
am grateful to my wife for helping me stay focussed till the very end. Finally, I wish to thank
the Publishers, PHI Learningtheir sales, editorial and production teamfor patiently pursuing the
scheduling of the manuscript and for its careful processing.
P. Seshu
CHAPTER 1
Introduction
Finite element analysis has now become an integral part of Computer Aided Engineering (CAE)
and is being extensively used in the analysis and design of many complex real-life systems.
While it started off as an extension of matrix methods of structural analysis and was initially
perceived as a tool for structural analysis alone, its applications now range from structures to
bio-mechanics to electromagnetic field problems. Simple linear static problems as well as highly
complex nonlinear transient dynamic problems are effectively solved using the finite element
method. The field of finite element analysis has matured and now rests on rigorous
mathematical foundation. Many powerful commercial software packages are now available,
enabling its widespread use in several industries.
Classical analytical methods consider a differential element and develop the governing
equations, usually in the form of partial differential equations. When applied to real-life problem
situations, it is often difficult to obtain an exact solution to these equations in view of complex
geometry and boundary conditions. The finite element method (FEM) can be viewed simply as
a method of finding approximate solutions for partial differential equations or as a tool to
transform partial differential equations into algebraic equations, which are then easily solved.
Some of the key ideas used in finite element formulation are now summarised:
Since the solution for the field variable satisfying both the boundary conditions and
the differential equation is unknown, we begin with an assumed trial solution. The
trial solution is chosen such that the boundary conditions are satisfied.
The trial solution assumed, in general, does not satisfy the differential equation exactly
and leaves a domain residual defined as the error in satisfying the differential
equation.
In general, the domain residual varies from point to point within the domain and
cannot be exactly reduced to zero everywhere. We can choose to make it vanish at
select points within the domain, but we prefer to render the residual very small, in
some measure, over the entire domain. Thus, the weighted sum of the domain
residual computed over the entire domain is rendered zero.
1
2 Introduction
The accuracy of the assumed trial solution can be improved by taking additional,
higher order terms, but the computations become tedious and do not readily render
themselves for automation. Also, for complex real-life problems, choosing a single
continuous trial function valid over the entire domain satisfying the boundary
conditions is not a trivial task. We therefore prefer to discretise the domain into
several segments (called finite elements) and use several piece-wise continuous
trial functions, each valid within a segment (finite element).
Trial functions used in each segment (finite element) are known as element level
shape functions. These are defined in the form of interpolation functions used to
interpolate the value of the field variable at an interior point within the element from
its value at certain key points (called the nodes) in the element. Typical elements
commonly used in finite element analysis are shown in Figure 1.1.
w2
v2 u2
j2
q2
y2
w1
v1
j1
q1
u1
y1
(a) General frame element (Six d.o.f. Per node)
(b) Common 2-d elements

Introduction 3
W
W y j v
j q
y v
u
q
u
(d) Plate and shell elements
(e) Common 3-d elements

Fig. 1.1 Typical finite elements
With these shape functions, the weighted sum of the domain residual is computed
for each element and then summed up over all the elements to obtain the
weighted sum for the entire domain.
For all elements using the same shape functions, the computations will be identical
and, thus, for each type of element we have element level characteristic matrices.
These characteristic matrices for several types of elements are derived a priori and
programmed into a finite element software such as ANSYS, NASTRAN, IDEAS, etc.
The user can choose to discretise (model) his domain with a variety of different finite
elements. The computer program sets up the characteristic matrices for each element
and then sums them all up for the entire finite element mesh to set up and solve the
system level equations.
The basic steps of finite element analysis, as outlined above, are quite generic and can
be applied to any problembe it from the field of structural mechanics or heat transfer, or fluid
flow or electromagnetic fields, given the appropriate differential equation and boundary
conditions. In view of the similarity in the form of governing differential equations, the finite
element formulation for a particular type of differential equation can be used to solve a class
of problems. For example, a differential equation of the type
d2 f
AC + q = 0
dx2
4 Introduction
describes axial deformation of a rod when we use the connotation that f represents the axial
deformation, q represents the load, and A, C stand for cross-sectional area and Youngs
modulus, respectively. The same equation, when interpreted with the connotation that f stands
for temperature, q represents internal heat source and A, C stand for cross-sectional area and
coefficient of thermal conductivity, respectively will be the governing equation for one-
dimensional heat conduction. Thus, a finite element formulation developed for the above
differential equation can be readily used to solve either of the physical problems.
Sometimes, the governing equations are more readily available in the form of
minimization of a functional. For example, in problems of structural mechanics, the
equilibrium configuration is the one that minimizes the total potential of the system. Finite
element formulation can be developed readily for a problem described by a functional, rather
than a differential equation. When both the forms are available for a given problem, the
differential equation and functional forms are equivalent and can be derived from each other.
The finite element method essentially grew up as a tool for structural mechanics
problems, as an extension of the matrix methods of structural analysis. While such an approach
towards the study of finite element formulation enables easy visualisation in the form of lumped
springs, masses, etc., the approach outlined above highlights the generic nature of the method,
applicable for a variety of problems belonging to widely varying physical domains. It is felt that
this approach gives a proper perspective on the entire field of finite element analysis. In the
chapters that follow, we elaborate on the various basic steps outlined above for one- and two-
dimensional, static and dynamic problems.
We now present several examples of application of finite element analysis to real-life
problems, to give an overview of the capabilities of the method. Our application examples
are drawn from the fields of structural mechanics, aerospace, manufacturing processes,
electromagnetics, etc.
1.1 Typical Application Examples
1.1.1 Automotive Applications

In a vehicle having monocoque construction, the body itself is connected to the suspension.
Therefore, the body panels are subjected to road loads. Hence, stresses and strains in these
body panels are of interest. Figure 1.2 shows a FE mesh of a floor panel from the rear end
of the vehicle. Provision for spare wheel as well as the various depressions used as stiffeners
can be seen in the figure. A total of about 13,000 quadrilateral and triangular shell elements have
been used to perform modal analysis, torsional stiffness analysis, and service load analysis. The
same finite element mesh is also used for crash analysis using LS-DYNA software.
An automotive engine cylinder block experiences severe pressures and temperature
gradients and other transient loads. It is essential to predict accurately the stresses and the
vibration levels for further correlation with noise predictions. Figure 1.3 shows a typical finite
element (shell element) model of a four cylinder in-line diesel engine cylinder block. Such a
model is used to predict the system natural frequencies and mode shapes, response to
combustion gas pressure, etc.
Typical Application Examples 5
Fig. 1.2 Finite element model (MSC/NASTRAN) of the floor panel of an automobile.
(Courtesy: TELCO, Pune.)
Fig. 1.3 Finite element model of an automotive engine cylinder block.

(Courtesy: Mahindra & Mahindra Ltd., Nasik.)
Figures 1.4 1.7 show representative finite element models of various components of a
driveline where the gears have been modelled as friction wheels. Such a model can be used
for studying the dynamic response of the entire driveline.
6 Introduction
Fig. 1.4 3-d Finite element model of a crankshaft. Fig. 1.5 3-d Finite element model of a gearbox.
(Courtesy: Mahindra & Mahindra Ltd., Mumbai.) (Courtesy: Mahindra & Mahindra Ltd., Mumbai.)
X
Y
Fig. 1.6 3-d Finite element model of a differential.

(Courtesy: Mahindra & Mahindra Ltd., Mumbai.)
Y
X
Fig. 1.7 3-d Finite element model of a rear axle.

(Courtesy: Mahindra & Mahindra Ltd., Mumbai.)
Friction f1 Friction f2 Friction f3

Fig. 1.8 Finite element modelling of cold heading process on a cylindrical slug.
(Courtesy: TRDDC, Pune.)
1.1.2 Manufacturing Process Simulation

Figure 1.8 is an illustrative example in which finite element tools have been used for inverse
analysis. A cylindrical slug is cold headed into the form of a near spherical ball. The coefficient
of friction plays an important role in the final ball shape. However, it is difficult to measure the
value of coefficient of friction. To overcome this difficulty, the operation is simulated for
various values of coefficient of friction and the pole diameter (diameter of near flat portion near
the poles) is compared with the measured pole diameter in the actual heading process for one
case. This value is used for further analysis and optimization of the process.
Figure 1.9 shows stress distribution in a fusion cast ceramic block. A transformation
behaviour of one of the components of the material mixture causes anomalous expansion during
cooling in a particular temperature range. This leads to tensile stress build-up and failure on the
cast block. A 3-d thermomechanical FE analysis is used to study the solidification, thermal field
and evolution of stress, and the cause for failure. This information is further used to change
the processing conditions so as to eliminate these high tensile stresses.
8 Introduction
Central core
Narrow face
Broad face
Fig. 1.9 Finite element model of fusion cast ceramic block.

(Courtesy: TRDDC, Pune.)
1.1.3 Electrical and Electronics Engineering Applications

FEA can be used for reliability enhancement and optimization of insulation design in high voltage
equipment by finding accurately the voltage stresses and corresponding withstands. For
complex configuration of electrodes and dielectric insulating materials, analytical formulations
are inaccurate and extremely difficult, if not impossible. The FEA can be effectively used in
such cases. A typical equipotential field plot for a high voltage transformer obtained from finite
element simulation is shown in Figure 1.10. An analysis of eddy currents in structural
conducting parts and minimization of stray losses in electrical machines is possible using FEM.
A typical mesh for one such analysis is depicted in Figure 1.11, where it is aimed at estimating
the eddy current losses in structural conducting plate supporting high current terminations.
Shield
LV winding HV winding
Fig. 1.10 High voltage insulation design of transformer using FEA.

(Courtesy: Electrical Engg. Dept., Indian Institute of Technology Bombay.)
Nonmagnetic inserts
B phase Y phase R phase
Fig. 1.11 Analysis of eddy currents using FEM.

(Courtesy: Electrical Engg. Dept., Indian Institute of Technology Bombay.)
Thermosonic wire bonding is one of the preferred processes for completing an electrical
interconnection between a semiconductor chip and a leadframe by a thin metal wire. The wire
bonder machine (solid model shown in Figure 1.12) consists of a linear motor driven precision
XY-table on which a precision Z-axis assembly integrated with an ultrasonic transducer is
mounted. The silicon chip is mounted on a hot plate called heater block assembly which is
maintained at 200C. The ultrasonic transducer helps in welding the gold wire to the silicon
chip. The combined motion of the X, Y and Z-axis helps achieve a curvilinear profile to produce
the required shape to the bonded wire.
10 Introduction
XX
and
&Y motors
Y motors
XY Table
Table &
& BH
assembly
Assembly
Heater block
Heater Block
assembly Elevator
Assembly Elevator
Base
Base
Work Holder
Work holder
Isolator
Power supply
Power supply Isolator
Fig. 1.12 Solid model of a wire bonder machine.

(Courtesy: ASM International Pte Ltd., Singapore.)
The present day needs of the wirebonder technology are quite challenging and have to
meet stringent process specifications such as small pitch distance of 35 mm between the
successive bonds, stringent bond placement accuracy of 2 mm, and loop height consistency
of 5 mm within a bonding area of 70 mm 70 mm. Additionally, to meet higher output, a
bonding rate of 6 12 wires per second, depending on the wire length range of 2 to 8 mm and
wire diameter of 20 to 75 mm, is necessary.
Since the bonding rate is high, the drive system experiences a peak acceleration of 13 g
(about 130 m/s2) in the XY-table. To achieve a position accuracy of within 2 m, the residual
vibration during such a motion should be well within 0.1 micron. Hence the design should be
aimed at higher dynamic stiffness and lower mass and inertia. The main objective of the finite
element simulation (a typical mesh is shown in Figure 1.13) is to analyse the dynamic rigidity
of the system. A typical bending mode of the machine is depicted in Figure 1.14.
The next example relates to the dynamic analysis of a one-axis linear motor. The one-axis
table (the FE model in Figure 1.15) consists of a three-phase linear motor that directly drives
a 5 kg mass in rectilinear motion. The table is designed for a peak acceleration of about
130 m/s2. Further, the design would be used in a typical semiconductor packaging machine
such as the wire bonder (described above) which involves bonding gold wires at a high speed
of 12 wire interconnects per second. Hence the finite element analysis aims at finding the
resonant frequencies and the mode shapes very accurately. This will help in achieving a high
dynamic rigidity with low mass. Also, knowledge of the mode shape helps in illustrating the
Fig. 1.13 Finite element model of a wire bonder machine.

Undeformed
shape
Deformed
shape
Fig. 1.14 Second mode shape (114 Hz) of wire bonder machine.
12 Introduction
Dummy load
Z
Y X
Table
Coil bracket
Coil
Base
Fig. 1.15 Finite element model of a one-axis table.

weakness in the design for the various resonant frequencies. The model consists of a coil
bracket containing the three coils and directly connected to the dummy mass by four screws.
The table is guided by roller bearings which are modelled as equivalent springs with appropriate
spring constant which corresponds to the applied preload in the bearings. A typical mode shape
is shown in Figure 1.16.
Z
Y X
Undeformed model
Undeformed model
Deformed
Deformedmodel
model
Mode 1
Bending Mode Mode 2
Edge-wise
bending Mode
Deformed model
Undeformed model
Deformed model
Undeformed model
Fn = 296 Hz Fn = 584 Hz
F =296 Hz F =584 Hz
Fig. n1.16 Natural frequencies and mode shapes ofn one-axis table.
Encapsulation of integrated circuits (IC) is commonly made using the transfer moulding
process where a thermoset epoxy is heated and transferred under pressure so as to encapsulate
the silicon IC chip. The moulding machine consists of a top and a bottom mould mounted on
large platens. The top platen is supported by four steel columns, while the bottom platen guided
by the four steel columns moves in a straight path using a motor-driven four bar mechanism.
A very high pressure (equivalent load of about 100 tons) is maintained between the top and the
bottom mould faces, so as to avoid liquid mould compound to leak at the interface. Heaters
are placed in the moulds to produce the heat necessary for maintaining a constant uniform
mould surface temperature of 175C. Such a uniform temperature enables uniform melting of
epoxy and its smooth flow into the mould cavities.
A typical finite element model of the complete system is illustrated in Figure 1.17. The
top and bottom mould surfaces are coated with surface contact elements in order to readily
determine the contact pressure. Further, thermal analysis of the model is conducted to predict
the temperature distribution on the mould surfaces. Results also include the heat losses to the
surroundings in the form of convection and radiation heat transfer as well as conduction heat
transfer to the base of the machine. Knowledge of these heat losses helps in deciding the
number of heaters and their locations as well as their corresponding wattage.
Top Platen
Contact surface
between mold faces Top Mold
Guide Column Bottom Mold
Bottom Platen
Fig. 1.17 Finite element model of a moulding machine.

14 Introduction
1.1.4 Aerospace Applications

In typical aerospace applications, finite element analysis is used for several purposes, viz.,
structural analysis for natural frequencies, mode shapes, response analysis, aero-servo-elastic
studies, and aerodynamics. For example, Figure 1.18 shows a typical combat aircraft designed
Fig. 1.18 Typical combat aircraft.

(Courtesy: Aeronautical Development Agency, Bangalore.)
and developed in India. Figure 1.19 depicts a typical deformed shape of the aircraft as predicted
using finite element tools. To model the complete aircraft, various types of elements (viz., rod,
shear panel, plate/shell, etc.) are used.
SUMMARY
As may be evident from the above examples of real-life application of finite element analysis,
present day engineering design based on CAE tools involves extensive use of finite elements in
a wide variety of fields. The question that naturally arises is how a single technique can be
applied to such a wide variety of application domains. Once we recognize the commonality in
the mathematical representation of these various problems (for example, the partial differential
equations or the functional expression), we realize that a tool to solve a type of differential
equation can be used to solve a class of problems. Knowledge of finite element principles is
therefore crucial in two significant ways: to aid intelligent use of commercial software, and
to lay a strong foundation for further research in this field. Thus this text focusses on
emphasising the fundamental principles used in the formulation of finite element method in a
lucid manner.
Summary 15
Fig. 1.19 Finite element simulation of deformation of a typical combat aircraft.

(Courtesy: Aeronautical Development Agency, Bangalore.)
CHAPTER 2
Finite Element Formulation
Starting from Governing
Differential Equations
As discussed in Chapter 1, the finite element method has been applied to a wide variety of
problems in various fields of application. This has been possible because of the common
features in the mathematical formulation of these seemingly different problems, e.g. many
problems of engineering analysis can be represented by partial differential equations. In some
cases, the same type of partial differential equation (e.g. the two-dimensional Laplace/Poisson
equation) can represent a large number of physical problems (e.g. ground water seepage,
torsion of bars, and heat flow). Thus, to gain a proper perspective of the method of finite
elements, we would like to present it in this chapter as a method for finding an approximate
solution to differential equations.
The Weighted Residual (WR) method is a powerful way of finding approximate solutions
to differential equations. In particular, The Galerkin Weighted Residual formulation is the most
popular from the finite element point of view. Piece-wise trial function approximation of the
weak form of the Galerkin weighted residual technique forms the basis of the finite element
method. In what follows, we will first introduce the general weighted residual technique and
the Galerkin form of the weighted residual technique, using a set of trial functions, each of
which is valid over the entire solution domain. We will then introduce the weak form of the
same. Finally, we will present the piece-wise trial function approximation concept, wherein
each of the trial functions used is valid only over a small part of the domain. This leads us to
the formulation of the finite element method.
2.1 Weighted Residual MethodUse of a Single Continuous

Trial Function
Let us consider a general problem of engineering analysis described in the form of a differential
equation (to be valid within a particular domain W), while satisfying the prescribed boundary
conditions on the boundary G. Our scheme of finding approximate solution to differential
equations consists of the following steps:
16
Weighted Residual MethodUse of a Single Continuous Trial Function 17
Assume a guess (or trial) solution to the problem. For example, for a one-dimensional
problem, we may choose a trial solution as
f (x) = c0 + c1x + c2x2 + (2.1)
In general, the function so assumed will satisfy neither the differential equation within
the domain (W) nor the boundary conditions (on G). By substituting the assumed
function in the differential equation and the boundary conditions of the problem, find
the error in satisfying these (we will call these domain residual and boundary
residual).
Determine the unknown parameters (c0, c1, c2, ) in the assumed trial function in
such a way as to make these residuals as low as possible.
In the process, if we can make the domain and boundary residuals identical to zero
everywhere, we will get the exact solution to the problem itself. In general, we expect to get
a reasonably accurate solution to the problem at hand. In the context of the finite element
method, we will limit our discussion to trial solutions that satisfy the applicable
boundary conditions and hence, only domain residual remains. The choice of trial
solutions that implicitly satisfy the differential equation but not the boundary conditions (thus
resulting in nonzero boundary residual) leads to the boundary element method. As a detailed
discussion on the boundary element method is beyond the scope of this text, the interested
reader may refer to standard texts for details of this technique.
We will use the following simple example to illustrate the above method of finding
approximate solutions to differential equations.
Example 2.1. Consider a uniform rod subjected to a uniform axial load as illustrated in
Figure 2.1. It can be readily shown that the deformation of the bar is governed by the
differential equation
L
q0
Fig. 2.1 Rod subjected to axial load.
d 2u
AE q0 0 (2.2)
dx2
du
with the boundary conditions u(0) = 0, = 0.
dx x=L
18 Finite Element Formulation Starting from Governing Differential Equations
Let us now find an approximate solution to this problem using the method just discussed.
Step 1: Assume a trial or guess solution. Let
u(x) (x) = c0 + c1x + c2x2 (2.3)
where the constants c0, c1, c2 are yet to be determined. In order to satisfy the first boundary
condition that (0) = 0, we have c0 = 0. To satisfy the second boundary condition, we have
c1 = 2c2L. Thus we now have, for our trial solution,
(x) = c2(x2 2Lx) (2.4)
Since the trial solution contains only one free parameter c2, it is often referred to as a one-
parameter solution.
Step 2: Find the domain residual. Substituting in the governing differential equation
d 2 u
Rd = AE + q0 = AE(2c2) + q0 (2.5)
dx 2
Step 3: Minimise the residual. Since there is one residual to be minimised and one
parameter to be determined, we can readily solve for the undetermined coefficient by setting
the residual to zero, i.e., Rd = 0, yielding
q0
c2 = (2.6)
2 AE
Thus our final solution is
q
u( x) = 0 (2 xL x 2 ) (2.7)
2 AE
For this simple example, since we could make the residual identically zero everywhere,
our final solution tallies with the exact solution.
Example 2.2. The governing equation for a fully developed steady laminar flow of a
Newtonian viscous fluid on an inclined flat surface (see Figure 2.2) is given by
y
LL
x
v(x)
v(x)
Gravity
g
q
Fig. 2.2 Laminar flow on an inclined surface.

d 2v
N + S g cos R = 0 (2.8)
dx 2
where
m = coefficient of viscosity,
v = fluid velocity,
r = density,
g = acceleration due to gravity,
q = angle between the inclined surface and the vertical.
The boundary conditions are given by
dv
= 0 (zero shear stress) (2.9)
dx x =0
v(L) = 0 (no slip)

Let us find the velocity distribution v(x) using the weighted residual method.
Step 1: Assume a trial solution. Let
v(x) v (x) = c0 + c1x + c2x2 (2.10)
Hence,
dv
= c1 + 2c2 x (2.11)
dx
From the boundary conditions, c1 = 0, c0 = c2L2. Therefore,
v (x) = c2(x2 L2) (2.12)
Step 2: Find the domain residual
Rd = m (2c2) + r g cos q (2.13)
Step 3: Minimise the residual. Rd is a constant and can therefore be set to zero. Hence,
S g cos R (2.14)
c2 =
2N
Therefore,
S g cos R 2
v( x ) = (L x2 ) (2.15)
2N
It is readily verified that our solution matches the exact solution as we can make the
domain residual identically zero.
Example 2.3. Consider the problem of a cantilever beam under uniformly distributed load
q0 as shown in Figure 2.3. The governing differential equation is given by
d 4v
q0 = 0
EI (2.16)
dx 4
and the boundary conditions are given by
dv
v(0) = 0, (0) = 0
dx
d 2v d 3v
( L) = 0, ( L) = 0 (2.17)
dx 2 dx3
where the first two boundary conditions enforce zero displacement and slope at the fixed
end and the last two conditions prescribe zero bending moment and shear force at the free end.
q0
x
L
Fig. 2.3 Cantilever beam under load.
Step 1: Assume a trial or guess solution. We observe that it is not easy to select a trial function
that satisfies all the boundary conditions. Let us choose v (x) = c0 + c1x + c2x2 + c3 x3 + c4x4.
From the boundary conditions that v (0) = 0 and d v /dx (0) = 0, we have c0 = 0 = c1.
In order to satisfy the boundary conditions at x = L, we should have
c2 = 3c3L 6c4L2, c3 = 4c4L (2.18)
Substituting and rearranging the terms in the trial solution, we get
v (x) = c4(x4 4Lx3 + 6x2L2) (2.19)
We thus observe that finding trial solution functions that satisfy all the boundary
conditions could, in general, be cumbersome.
Step 2: Find the domain residual. Substituting in the differential equation, we get the domain
residual as
Rd (x) = 24EIc4 q0 (2.20)
Step 3: Minimise the residual. Since there is one residual to be minimised and one parameter
to be determined, we can readily solve for the undetermined coefficient c4 by setting the
residual to zero, i.e., Rd = 0, yielding thereby
c4 = q0/(24EI) (2.21)
Thus our trial solution is
v (x) = q0/(24EI) [x4 4Lx3 + 6x2L2]
which can be readily verified to be the exact solution itself. This is to be expected since we
were able to make the residual identically zero within the entire domain.
Example 2.4. Let us consider the example of a simply supported beam under uniformly
distributed load as shown in Figure 2.4. The governing differential equation and the boundary
conditions are given by
q0
x
L
Fig. 2.4 Beam on simple supports.
d 4v
EI q0 = 0
dx 4
d 2v
v(0) = 0, (0) = 0 (2.22)
dx 2
d 2v
v (L) = 0, ( L) = 0
dx 2
Step 1: Assume a trial solution. We can use the approach of Example 2.3 and find a
polynomial trial solution satisfying all boundary conditions. However, in view of the special
boundary conditions of simple supports, we can make the process simple by choosing
trigonometric functions. Let
v(x) v (x) = c1 sin (p x/L) (2.23)
This one-parameter trial solution satisfies all boundary conditions.
Step 2: Find the domain residual. Substituting the trial solution v (x) in the governing
differential equation, the domain residual is obtained as
Rd = c1(p /L)4(EI) sin (p x/L) q0 (2.24)
Step 3: Minimise the residual. We observe that, unlike in the previous examples, the domain
residual is now varying from point to point within the domain (0 < x < L). Since we have only
one coefficient to be determined, we can set the residual zero only at any one point of our
choice within the domain if we follow the approach of previous examples. This technique is
called the point collocation technique, wherein we set the residual (in general, a function
of x) to zero at chosen points within the domainthe number of points being equal to the
number of coefficients in the trial function that need to be determined. We will illustrate this
procedure now. In this procedure, however, there is a danger that the residual might be unduly
large at some other points within the domain. We may thus want to minimise the residual
in an overall sense over the entire domain rather than setting it identically zero at only few
selected points. This will be discussed in the next section.
Solution by point collocation. Let us make Rd = 0 at x = L/4, i.e.
c1(p /L)4 (EI) sin (p /4) q0 = 0 (2.25)
yielding thereby
2 q0 L4 (2.26)
c1 =
Q 4 EI
and the resulting trial solution is
2 q0 L4 Qx
v( x) = sin (2.27)
Q EI
4
L
We may repeat this process by setting the domain residual zero at other axial locations.
Figure 2.5 compares the solutions v (x) obtained in this manner against the classical exact
16
R d (L/4) = 0
Rd(L/4)=0
14 R d (L/3) = 0
Rd(L/3)=0
Transverse Deflection
R d (L/2) = 0
Rd(L/2)=0
deflection
12 SOM SOM
10
8
Transverse
6
4
2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalisedaxial
Normalised Axial location
location on beam
on the the beam
q0L4
Fig. 2.5 Simply supported beam deflection. = 1000
EI
solution. Figure 2.6 shows a plot of the domain residual in each case. In this way, we are able
to generate different approximate solutions to the problem but each of these solutions deviates
appreciably from the exact solution. We can improve our trial solution by adding one more term
to the Sine series. Since the problem at hand is symmetric about x = L/2, we modify our trial
solution as follows:
v(x) v (x) = c1 sin (p x/L) + c3 sin (3p x/L) (2.28)
0.6
Residual
Residual R d (L/4) = 0
Rd(L/4)=0
Residual
Rd(L/3)=0
0.4
Residual Rd(L/2)=0
0.2
0
Residual
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

residual
0.2
-0.2
0.4
-0.4
0.6
-0.6
0.8
-0.8
-11
Normalised x-location
normalised x location
Fig. 2.6 Plot of residual-simply supported beam (Example 2.4).
With this two-term trial function, we obtain the domain residual as

Rd = c1(p /L)4(EI) sin (p x/L) + c3(3p /L)4(EI) sin (3p x/L) q0 (2.29)
Since we have two constants to be determined, we can set the residual zero at two
selected points. Let us set
L L
Rd = Rd = 0 (2.30)
4 3
i.e.,
4 4
Q 1 3Q 1
c1 EI + c3 EI q0 = 0 (2.31)
L 2 L 2
Q
4
3 3Q
4
c1 EI + c3 EI (0 ) q0 = 0 (2.32)
L 2 L
Solving these equations, we obtain
2 q0 L4
c1 = (2.33)
3 Q 4 EI
q0 L4
c3 = 0.00003289 (2.34)
EI
Figure 2.7 compares the solution v (x) obtained in this manner against the classical exact
solution. Figure 2.8 shows a plot of the domain residual. We observe that this approximates
the solution perhaps better than the earlier one-parameter solution, but it still deviates
appreciably from the exact solution.
14
TWO TERM
SOM
12
Transverse deflection
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Normalised axial location
Fig. 2.7 Simply supported beam deflectionTwo-parameter solution.
0.2
0.2
Residual
ResidualRd Rd(L/4
(L/4 and& L/3) =0
L/3)=0
00
0 0.1
0.1 0.2 0.3
0.3 0.4 0.5
0.5 0.6 0.7
0.7 0.8 0.9
0.9 1
0.2
-0.2
Residual
0.4
-0.4
-0.6
0.6
-0.8
0.8
1
-1
Normalised axial location
Fig. 2.8 Domain residualTwo-parameter solution Rd (L/4 and L/3) = 0.
Rather than trying to set the residual exactly zero at a few selected points and having no
control over the residual at all the other points in the domain, we will now try to minimise the
residual in an overall sense. This technique is called the weighted residual technique.
Solution by weighted residual technique. For the problem at hand, we will formulate this technique
as
L
0
Wi ( x) Rd ( x) dx = 0 (2.35)
where Wi (x) are appropriately chosen weighting functions, helping us to achieve the task of
minimising the residual over the entire domain. We choose as many weighting functions as
necessary to generate the required number of equations for the solution of the undetermined
coefficients in the trial function. For example, for a two-term trial solution, we would take two
different weighting functions to generate two equations necessary to solve for the two
coefficients. While the choice of weighting functions W(x) is entirely arbitrary (as long as they
are nonzero and integrable), Galerkin (1915) introduced the idea of letting W(x) to be same
as the trial functions themselves. Thus, in this example, we take the weighting function to
be sin (p x/L), sin (3p x/L), etc. Let us illustrate the procedure of the Galerkin weighted residual
technique for the one-term trial function, i.e., let
v(x) v (x) = c1 sin (p x/L) (2.36)

The domain residual

Rd = c1(p /L)4(EI) sin (p x/L) q0 (2.37)
By the Galerkin procedure, we set
Qx Q
4
L Qx
0
sin

c1 EI sin
L
L L
q0 dx = 0

(2.38)
w ( x)
Rd ( x )
We have chosen only one weighting function because there is only one constant c1 to be
determined. Thus we have
4
Q L
2 Qx
L Qx
EIc1 sin
L 0 L
dx =
0
q0 sin
L
dx (2.39)
yielding c1 = 0.013071q0L4/EI. Therefore,

q L4 Qx
v( x) = 0.013071 0 sin (2.40)
EI L
Figure 2.9 compares the solution v (x) obtained in this manner against the classical exact
solution. It is observed that the one-term solution is significantly improved by the Galerkin
14
12 SOM
SOM
Galerkin Solution
Galerkin Solution
10
Deflection
deflection
0
0 0.1 0.2 0.40.3 0.5 0.6 0.7 0.8 0.9 1
Normalised x-location
Fig. 2.9 Strength of material (SOM) and Galerkin WR solutionsSimply supported beam (Example 2.4).
weighted residual minimisation process. The error in the deflection at the mid-span is just
0.38%. Figure 2.10 shows a plot of the domain residual obtained by the Galerkin process.
We have thus discussed a very powerful tool for determining approximate solutions to
boundary value problems, given the governing differential equation and the boundary conditions.
This method requires us to choose trial solutions that satisfy the boundary conditions of the
problem. We then generate sufficient number of equations to solve for the undetermined
parameters by setting the weighted integral of the residual to zero.
0.4
Galerkin Residual
0.2
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
residual
-0.2
0.2
Residual
-0.4
0.4
-0.6
0.6
-0.8
0.8
-1
1 Normalised axial location
Fig. 2.10 Domain residual obtained by Galerkin method (Example 2.4).
We can, in principle, choose any nonzero, integrable weighting functions to generate

the necessary equations. It is easily realised that if we choose the weighting functions,
to be DiracDelta functions, i.e., W(x) = d (x xp), the weighted residual process will reduce
to the point collocation technique discussed earlier, with the collocation points being x = xp. In
the Galerkin method, we choose the weighting functions to be the trial functions themselves.
To be precise, if the trial solution is assumed to be
f (x) = f (x) + S ci Ni (x) (2.41)
where f (x) and Ni (x) are fully known functions of x, pre-selected such that the boundary
conditions are satisfied by f (x), then we choose the weighting functions to be Wi (x) = f/ci
= Ni (x).
We observe that our method is not limited to problems of structural mechanics alone but
can be used to solve any problem given the appropriate differential equation and the boundary
conditions. We will now work out a few more example problems to gain further insight into
the Galerkin Weighted Residual technique. Our next example illustrates the application of this
technique to heat transfer problems.
Example 2.5. Consider the problem of a long cylinder of radius R with uniformly distributed
heat source q0. The governing differential equation is given by
d 2T 1 dT q
2
+ + 0 = 0 (2.42)
dr r dr k
and the boundary conditions are given by

T(R) = Tw, the wall temperature
dT
q0Q R 2 L = ( k )(2Q RL) (2.43)
dr r=R
The second boundary condition implies that heat generated = heat lost.
We wish to determine the temperature distribution T as a function of radial location, r.
Step 1: Assume a trial solution. We observe that it may be convenient to choose the trial
solution in terms of (r R) rather than r.
Let us choose the trial solution as
T = c0 + c1(r R) + c2(r R)2 (2.44)
From the boundary conditions, we find that c0 = Tw and c1 = (q0R/2k).
Thus we have
T = [Tw (q0R/2k)(r R)] + c2(r R)2 (2.45)
In our notation introduced above, we can write T = f (r) + c2N(r) where N(r) = (r R)2
will be our weighting function W(r).
Step 2: Find the domain residual. The domain residual
Rd = 2c2 + q0/k + (1/r)[(q0R/2k) + 2c2(r R)] (2.46)
Step 3: Minimise the residual. We observe that the residual is actually minimised over the
whole volume of the domain and in radial coordinates, an elemental volume is given by
(r dr dq dz). Considering the fact that temperature does not vary in the circumferential direction
(q ) or the axial direction (z) for this problem, explicit integration in these two directions leads
to the constant (2p L), and the weighted residual statement is given by
R
2Q L (r R) 2 [Rd ]r dr = 0 (2.47)
0
q0
Solving, we get c2 = . Therefore, our approximation to the temperature distribution is
4k
q
T ( r ) Tw = 0 ( R 2 r 2 ) (2.48)
4k
Thus we observe that our procedure can be systematically applied to any problem, given
the differential equation and the boundary conditions.
We will now illustrate the Galerkin Weighted Residual procedure on a two-term trial
function through the following example.
Example 2.6. Consider the uniform bar problem discussed in Example 2.1. Let
the bar be subjected to a linearly varying load q = ax. The governing differential equation is
given by
d 2u
AE + ax = 0 (2.49)
dx 2
du
with the boundary conditions u(0) = 0, AE = 0.
dx x=L
Step 1: Assume a trial or guess solution. In view of the load term, we expect that our trial
solution should be a higher degree polynomial than that used in Example 2.1. Let
u(x) (x) = c0 + c1x + c2x2 + c3x3 (2.50)
From the boundary conditions, we get
c0 = 0, c1 = (2c2L + 3c3L2) (2.51)
Thus
(x) = c2(x2 2Lx) + c3(x3 3L2x) (2.52)
Step 2: Find the domain residual
Rd = AE(2c2 + 6c3x) + ax (2.53)
Step 3: Minimise the residual. The weighting functions to be used in the Galerkin technique
are the trial functions themselves, i.e., W1(x) = (x2 2Lx) and W2(x) = (x3 3L2x).
Thus we have the weighted residual statements
L
0
( x 2 2 Lx)[ AE (2c2 + 6c3 x) + ax] dx = 0 (2.54)
L
0
( x3 3L2 x)[ AE (2c2 + 6c3 x) + ax] dx = 0 (2.55)
Solving for c2 and c3, we obtain
c2 = 0, c3 = (a/(6AE)) (2.56)
Thus, (x) = (a/(6AE))(3L2x x3), which is readily verified to tally with the exact solution.
2.2 The General Weighted Residual (WR) Statement

Having understood the basic technique and successfully solved a few problems, we can now
attempt to write down the general weighted residual statement.
For the unknown field variable u, we assume an approximate solution of the form
n
u u = G + c N
i =1
i i
(2.57)
where ci are the independent coefficients to be determined by the WR process and the
functions f and Ni are preselected such that satisfies all the prescribed boundary conditions.
Let Rd (x, y, z) be the domain residual. The WR statement can then be written as
:W R
i d d: = 0 for i = 1, 2,..., n (2.58)
where Wi = Ni.
It is observed that this criterion is implicitly satisfied by the exact solution for any and
every weighting function since Rd is trivially zero everywhere. Thus the principal idea of
The General Weighted Residual (WR) Statement 29
weighted residual approach is that, if we are able to satisfy this criterion for a sufficiently large
number of independent weighting functions, then it is likely that the assumed solution will be
reasonably close to the exact solution. More strongly, if we have taken a series representation
(e.g. polynomial, trigonometric) for the trial function (and, therefore, W), we expect that our
result will get better as we include more terms in the series. Thus we expect good convergence
properties.
Example 2.7. To show the application of the Galerkin technique to a more complex
situation, consider the simply supported rectangular plate subjected to uniform load as shown
in Figure 2.11. The governing differential equation is given by
Eh3 4 w 4 w 4 w
+ 2 + q0 = 0 (2.59)
12(1 v 2 ) x 4 x 2 y 2 y 4
q0
o
0
x
a bb
All edges
All edges simply
simp
supported
supported
Fig. 2.11 A simply supported plate.
and the boundary conditions by

w(0, y) = 0 = w(a, y), 2w/x2 = 0 for x = 0 and a
w(x, 0) = 0 = w(x, b) 2w/y2 = 0 for y = 0 and b
Step 1: Assume a trial or guess solution. In view of the boundary conditions for simple
supports, we will choose a trigonometric function for trial solution.
Let us choose a one-term trial function given by
w(x, y) w (x, y) = c11 sin (p x/a) sin (p y/b) (2.60)
It is easy to see that our trial function satisfies all the boundary conditions of the problem.
Step 2: Find the domain residual. The domain residual is obtained by substituting w (x, y) in
the differential equation as follows:
2
Eh3 Q 2 Q
2
Qx Qy
Rd = c11 + sin sin q0 (2.61)
12(1 v 2 ) a b a b
Step 3: Minimise the residual. The weighting function is W(x, y) = sin (p x/a) sin (p y/b). The
weighted residual statement will now be
b a Q x Q y Eh3 2 Qx Qy
0 0
sin
a
sin
b 12(1 v )
2
c11 (Q /a )2 + (Q /b)2 sin
a
sin
b
q0 dx dy = 0 (2.62)

Solving for c11, we get
16q0 12(1 v 2 )
2
1
c11 = 2 (2.63)
Q Eh3 (Q /a)2 + (Q / b)2
If a = b, i.e., for a square plate, we have the deflection at the centre of the plate as given
by (4/p 6) [12(1 v2)/(Eh3)] (q0a4).
Example 2.8. Consider a 1 mm diameter, 50 mm long aluminium pin-fin (as shown in

Figure 2.12) used to enhance the heat transfer from a surface wall maintained at 300C. The
governing differential equation and the boundary conditions are given by
h T

w
Wall
Wall
Tw
L L
Fig. 2.12 A pin-fin.
d 2T Ph
k 2
= (T T )
dx Ac
T(0) = Tw = 300C
dT (insulated tip) (2.64)
( L) = 0
dx
The General Weighted Residual (WR) Statement 31
where
k = coefficient of thermal conductivity,
P = perimeter,
Ac = cross-sectional area,
h = convective heat transfer coefficient,
Tw = wall temperature,
T = ambient temperature.
Let k = 200 W/m/C for aluminium, h = 20 W/m2C, T = 30C. Estimate the temperature
distribution in the fin using the Galerkin weighted residual method.
With the given numerical values, we have
d 2T
= 400(T 30) (2.65)
dx 2
dT
with T (0) = 300 C, ( L) = 0.
dx
T(x) T (x) = c0 + c1x + c2x2
From the boundary conditions, c0 = 300, c1 = 2c2L. Therefore,
T (x) = 300 + c2(x2 2Lx) (2.66)
Step 2: Compute the residual. Substituting T (x) in the differential equation, we get
Rd (x) = 2c2 400[270 + c2(x2 2Lx)] (2.67)
Step 3: Minimise the residual. The weight function W(x) = x2 2Lx. Thus we have
L
0
( x 2 2 Lx) Rd ( x) dx = 0 (2.68)
Solving for c2, we get c2 = 38,751.43. Hence,
T (x) = 300 + 38,751.43(x2 2Lx) (2.69)

We can readily obtain the exact solution as
T T
T ( x) exact = T + w cosh [m( L x)]
cosh (mL)
where
hP
m2 = (2.70)
kAc
We observe that the exact solution is an exponential function. The approximate solution
obtained just now is quadratic. We can improve our approximation by taking higher degree
terms in the polynomial trial function. Let
T (x) = c0 + c1x + c2x2 + c3x3 (2.71)
From the boundary conditions, c0 = 300C.

c1 = 2c2L 3c3L2 (2.72)
Thus,
T (x) = 300 + c2(x2 2Lx) + c3(x3 3L2x) (2.73)
The weighted residual equations can be developed using the two weighting functions
(x2 2Lx) and (x3 3L2x). We can get the solution as
c2 = 48,860.4, c3 = 109,229 (2.74)
Thus,
T (x) = 300 + 48,860.4(x2 2Lx) 109,229(x3 3L2x) (2.75)
We can further improve upon our solution by taking one more term in the series, i.e.
T (x) = c0 + c1x + c2x2 + c3x3 + c4x4 (2.76)

From the boundary conditions we have
c0 = 300, c1 = 2c2L 3c3L2 4c4L3 (2.77)
Thus,
T (x) = 300 + c2(x2 2Lx) + c3(x3 3L2x) + c4(x4 4L3x) (2.78)
We can develop the three weighted residual equations using the weighting functions
(x 2Lx), (x3 3L2x) and (x4 4L3x). We can get the solution as
2
c2 = 53,702.8, c3 = 255,668, c4 = 1.32 106

Therefore,
T (x) = 300 + 53,702.8 (x2 2Lx) 255,668 (x3 3L2x) + 1.32 106 (x4 4L3x) (2.79)
Table 2.1 compares the various approximate solutions with the exact solution. It is
observed that the accuracy of the Galerkin approximate solutions can be systematically
improved by taking more and more terms in the series solution. However, the mathematical
Table 2.1 Comparison of Various Solutions for Temperature in a Fin
Axial Quadratic Cubic Quartic Exact

location solution solution solution solution
0 300 300 300 300
0.005 281.59 280.87 280.75 280.75
0.01 265.12 264.11 264.00 264.02
0.015 250.59 249.62 249.60 249.62
0.02 238.00 237.33 237.39 237.43
0.025 227.34 227.16 227.27 227.31
0.03 218.62 219.02 219.12 219.16
0.035 211.84 212.83 212.86 212.91
0.04 207.00 208.51 208.43 208.49
0.045 204.09 205.98 205.79 205.85
0.05 203.12 205.16 204.91 204.97
Weak (Variational) Form of the Weighted Residual Statement 33
manipulations become very tedious. We will instead try to improve the solution while working
with simpler, lower order polynomials. Towards this purpose, we will first introduce the weak
form of the weighted residual statement.
2.3 Weak (Variational) Form of the Weighted Residual Statement

Considering the general weighted residual statement (Eq. 2.58), it is possible to carry out
integration by parts and recast the general WR statement. This will reduce the continuity
requirement on the trial function assumed in the solution. We will thus have a much wider
choice of trial functions. Let us illustrate this with the following example.
Example 2.9. Let us reconsider Example 2.6 in which we had the governing differential
equation
d 2u
AE + ax = 0 (2.80)
dx 2
du
with the boundary conditions u(0) = 0, AE = 0.
dx x=L
Let be the trial solution assumed. Substituting in the differential equation, we find the
domain residual as Rd = AE (d2/dx2) + ax.
With W(x) as the weighting function, we have the weighted residual statement
L d 2 u
0
W ( x) AE 2 + ax dx = 0
dx
(2.81)
i.e.
L d 2u L

0
W ( x) AE
dx 2
dx +
0
W ( x) ax dx = 0 (2.82)
or
L du L
0
W ( x) d AE
dx
+ 0
W ( x) ax dx = 0 (2.83)
Recall the standard formula for integration by parts (for a definite integral)
C C C C
B u dv = [uv ]B B v du = (uv) |C (uv) |B + B v du (2.84)
du
We observe that, in our case, u = W(x) and v = AE . Integrating the first term in Eq. (2.83)
dx
by parts, we have
L
du L du dW L
W ( x) AE dx
0
0
AE

dx dx
dx +
0
W ( x) ax dx = 0 (2.85)
Writing AE (du/dx) = P, the axial force at the section, we can rewrite the above equation as
L du dW L
W ( L) PL W (0) P0
0
AE
dx dx
dx + 0
W ( x) ax dx = 0 (2.86)
We observe that the continuity demanded on u(x) has gone down. In the original weighted
residual statement, Eq. (2.81), we had the term d2/dx2 while in Eq. (2.86) we have only
d/dx. Thus our choice of (x) should at least be quadratic for the original weighted residual
statement, while with the modified form, even linear trial functions are permissible. Also, the
force (natural) boundary conditions have been explicitly brought out in the WR statement itself.
We can actually substitute for known natural boundary conditions on P0 or PL. (Thus the trial
function assumed need only satisfy the essential boundary condition at x = 0, i.e. u(0) = 0.)
This form of the WR statement is called the weak (or variational) form of the weighted residual
statement. It is referred to as the weak form because of the weaker continuity demand on the
trial solution. From now on, we will use only this form of the WR technique.
du
From the prescribed natural boundary condition, AE = PL = 0. Also, in view of the
dx
prescribed essential boundary condition at x = 0 on u(0), we require that the weighting function
W(x) be such that W(0) = 0. Thus the weak form becomes
L du dW L
0
AE
dx dx
dx =
0
W ( x) ax dx (2.87)
subject to u(0) = 0, W(0) = 0

Let us clearly state at this stage that we now have three equivalent ways of formulating
the problem as shown in Table 2.2.
Table 2.2 Equivalent Representation of a Problem Statement
Differential Weighted residual Weak form

equation statement
d 2u L d 2u L du dW L
AE
dx 2
+ ax = 0
0
W AE 2 + ax dx = 0
dx 0
AE
dx dx
dx =

0
W ax dx
subject to u(0) = 0 subject to (0) = 0 subject to (0) = 0

du du
AE ( L) = 0 AE ( L) = 0 W(0) = 0
dx dx
To illustrate the solution of the weak form statement of the above problem, let
u(x) (x) = c1x + c2x2 (2.88)
Then we have
d/dx = c1 + 2c2x, W1 = x, W2 = x2
dW1/dx = 1, dW2/dx = 2x
We observe that (0) = 0, W1(0) = 0 and W2(0) = 0, as required.
Weak (Variational) Form of the Weighted Residual Statement 35
Weak form w.r.t. W1

From Eq. (2.87),
L L
0
( AE )(c1 + 2c2 x)(1) dx = 0
( x)(ax) dx
i.e.,
AE (c1L + c2 L2 ) = aL3 /3 (2.89)
Weak form w.r.t, W2

L L
0
( AE )(c1 + 2c2 x)(2 x) dx = 0
( x2 )(ax) dx
i.e.,
c L2 + 4c2 L3 aL4
AE 1 = (2.90)
3 4
Rearranging the two equations, we have
aL2
c1 + c2 L =
3 AE
(2.91)
4 aL2
c1 + c2 L =
3 4 AE
yielding
7aL2 aL
c1 = , c2 =
12 AE 4 AE
Thus the deflection (x) is obtained as
aL
u( x) = (7 xL 3 x 2 ) (2.92)
12 AE
It is easily verified that the exact solution is
a
u ( x) = (3L2 x x3 )
6 AE
Thus our approximate solution tallies with the exact solution only at both the ends but is
otherwise deficient (being only quadratic). From Example 2.6, it is observed that we needed
a cubic trial function to get the exact solution using the Galerkin weighted residual statement.
It is left as an exercise for the reader to show that if we were to take the same quadratic
trial solution, i.e. (x) = c0 + c1x + c2x2, but work with the original Galerkin weighted residual
form, we would get (x) = (5aL/(16AE))(2Lx x2). Figure 2.13 compares the two quadratic
approximations with the exact solution. It is observed that the weak form approxima-tion is
much better than the WR approximation.
We will now review the three different mathematical statements of the same physical
problem as discussed above and compare their features.
0.35 QuadraticWeighted residual

Quadratic - Weighted Residual
QuadraticWeak
Quadratic - Weakform
Form
0.3
Exact Solution
EXACT SOLUTION
0.25
Axial deformation
Axial deformation
0.2
0.15
0.1
0.05
0
0 0.2 0.4 0.6 0.8 1
Axial
Axial Location
location
Fig. 2.13 Comparison of approximate solutions for a rod under load (Example 2.9).
2.4 Comparison of Differential Equation, Weighted Residual

and Weak Forms
Consider the example of the axial deformation of a rod, described by the differential equation
d 2u
AE + q = 0 (2.93)
dx 2
and the boundary conditions
u(0) = u0
du (2.94)
AE = PL
dx x= L
In the context of finding approximate solutions, we have developed the weighted residual
statement as follows:
d 2u L
0 dx
W ( x) AE 2 + q dx = 0 (2.95)
subject to the boundary conditions

(0) = u0
du (2.96)
AE = PL
dx x= L
Comparison of Differential Equation, Weighted Residual and Weak Forms 37
We make the following observations:

1. If the function (x) is the exact solution, then the weighted residual form is implicitly
satisfied for any and every weighting function W(x) as the domain residual term given
within the parentheses is identically zero; otherwise, the error (or residual) is
minimised over the entire domain in a weighted integral sense.
2. For any given differential equation (linear/nonlinear, ordinary/partial, etc.), we can
write the weighted residual statement.
3. The weighted residual statement is equivalent only to the differential equation and does
not account for boundary conditions. The trial solution is expected to satisfy all the
boundary conditions and it should be differentiable as many times as required in the
original differential equation.
4. The weighting functions W(x) can, in principle, be any nonzero, integrable function.
By choosing n different weighting functions, we generate the necessary n equations
for the determination of coefficients ci(i = 1, 2, K, n) in the trial solution field
assumed.
5. In general, the weighting function W(x) is subject to less stringent continuity
requirement than the dependent field variable u. For example, in the above case,
the trial solution (x) has to be at least twice differentiable (because of the term
d 2/dx2), but no such restriction applies for the weighting function. The objective of
performing integration by parts and developing the weak form is to distribute the
continuity demand uniformly between the trial solution (x) and the weighting
function W(x).
Upon performing integration by parts, we obtain the weak form statement of the problem
as
L
du L du dW L
W ( x) AE dx
0
0
AE
dx dx
dx + 0
W ( x)q dx = 0 (2.97)
Substituting the prescribed natural (force) boundary condition at x = L, i.e.

du
AE = PL , and requiring that W(0) = 0 in view of prescribed essential boundary
dx x= L
condition u(0) = u0, we get the weak form statement as
L du dw L
0
AE
dx dx
dx = 0
W ( x)q dx + W ( L) PL (2.98)
subject to (0) = u0, W(0) = 0.
Observations on the Weak Form

1. In the weak form obtained by integration by parts, the continuity demand on trial
function u has gone down and that on the weighting function W has increased.
2. For even order differential equations (the only type we will be dealing with, in this
book), we will be able to distribute the continuity demand (or the differentiation)
equally between the trial solution function, and the weighting function by performing
integration by parts sufficient number of times.
3. Since the natural boundary condition explicitly arises when we perform integration by
parts, we can directly substitute the prescribed natural boundary conditions. Only the
remaining essential boundary conditions have to be satisfied by the trial solution of the
weak form equation.
4. In view of the above observations, a much wider choice of trial functions (e.g.
significantly lower order polynomials than is possible with weighted residual forms)
is available.
5. The weak form can be developed for the weighted residual form of any given
second- and higher-order differential equation or its corresponding weighted residual
form.
6. In view of the prescribed essential boundary conditions, we require that the weighting
function be zero at all those points on the boundary where such conditions are
specified, i.e., the weighting function should satisfy the homogeneous part of the
prescribed essential boundary conditions. Though at this stage this appears to be an
arbitrary requirement, it is not really so. The weight function can be interpreted as a
virtual displacement (see Section 3.2) and, therefore, to be consistent with the
supports, it must vanish at all those points where a prescribed displacement boundary
condition exits.
We will now illustrate the intricacies of the weak form solution through the following
interesting example.
Example 2.10. Let us revisit Example 2.4, where we considered the deflection of a simply
supported beam under uniformly distributed load. The governing differential equation was given
by
d 4v
EI q = 0 (2.99)
dx 4
and the boundary conditions
Essential: v(0) = 0, v(L) = 0 (zero displacements)
2 2
d v d v
Natural: (0) = 0 , ( L) = 0 (zero moments)
dx 2 dx 2
Let v( x) be the assumed trial solution and let W(x) be the weighting function. The
weighted residual statement can be written as
L d 4v

0
W ( x) EI 4 q dx = 0
dx
(2.100)
Comparison of Differential Equation, Weighted Residual and Weak Forms 39
i.e.
d 3v
L L
0
W ( x)d EI 3 =
dx 0
Wq dx (2.101)
On integration by parts, we get

L
d 3v L d 3v dW L
W ( x) EI 3
dx 0 0
EI
dx3 dx
dx =
0
Wq dx (2.102)
Integrating by parts once again, we obtain
d 3v
L dW ( x) d 2 v
L
L d 2 v d 2W ( x) L
W ( x) EI 3
dx 0 dx
EI 2
dx 0 0
EI 2
dx
dx 2
dx

=
0
Wq dx (2.103)
Thus we now have equal order of differentiation on both the dependent variable v (x) and
the weighting function W(x). Substituting for the prescribed natural boundary conditions at
either end that d 2 v /dx2 be zero, we have the resulting weak form:
L
L d 2 v d 2W ( x) d 3v L
0
EI 2
dx dx 2
dx +

W ( x ) EI
dx 0
3 =
0
Wq dx (2.104)
subject to the essential boundary condition that v (0) = 0, v (L) = 0. In view of these prescribed
essential boundary conditions, we require that the weighting function be such that W(0) = 0,
W(L) = 0. Thus the weak form becomes
L d 2 v d 2W ( x) L
0
EI
dx 2 dx2
dx = 0
Wq dx (2.105)
subject to
v (0) = 0 = v (L)
W(0) = 0 = W(L)
We observe that our trial solution v (x) for Eq. (2.105) need only be twice differentiable while
a trial solution for Eq. (2.100) should be differentiable at least four times. Thus we can now
take a simple quadratic trial function.
Let v (x) = c0(x)(L x) which satisfies the essential boundary condition that the
displacement be zero at either end. Also, following the Galerkin procedure, we choose the
weighting function to be the trial function itself and thus we have W(x) = x(L x), which also
satisfies the required conditions on W(x). For simplicity, let q(x) = q0, a uniformly distributed
load. Substituting these in the weak form, we obtain
L L
0
EI (c0 )( 2)(2) dx = 0
( x)( L x)q0 dx (2.106)
i.e.
q0 L3
(4 EIL)c0 =
6
or
q0 L2
c0 =
24 EI
Thus our approximate solution to the given problem is
q0 L2
v( x ) = ( x )( L x ) (2.107)
24 EI
The exact solution can be readily obtained as
q0 x 3
v( x) = ( L 2 Lx 2 + x3 ) (2.108)
24 EI
Our approximation to the maximum deflection of the beam at the mid-span is seen to be
20% in error. However, it is interesting to see that the slopes at either end tally with the exact
solution! The reader is urged to ponder over this result.
Observations
1. We observe that in the form of the WR statement of Example 2.4, we could not have
considered this trial function, whereas with the weaker continuity demanded on v (x)
now, we are able to consider lower order polynomials.
2. While we obtained a not-so-bad estimate to the field variable with such a lower order
polynomial trial function, it would be interesting to study the behaviour of the
derivatives of the solution obtained in Example 2.4 and the present one. Figure 2.14
shows a comparison of the nature of the function (i.e. deflection here) in both the
0.35
0.35
Quadratic
Quadratic(Eq. (2.107))
Sinusoidal
Sinusoidal(Eq. (2.40))
Exact
Exact(Eq. (2.108))
Deflection
00
00 0.1
0.1 0.2
0.2 0.3
0.3 0.4
0.4 0.5
0.5 0.6
0.6 0.7
0.7 0.8
0.8 0.9
0.9 11
Axial location
Fig. 2.14 Comparison of approximate solutions for deflection of a simply supported beam (Example 2.10).
Piece-wise Continuous Trial Function Solution of the Weak Form 41
cases. The first and second derivatives are compared in Figures 2.15(a) and 2.15(b).
Thus we see that a trial function may approximate the field variable itself reasonably
well but the derivative quantities may be very poorly approximated. The derivatives
yield us information regarding physical variables of interest in the problem, e.g.
strain/ stress, heat flux, etc. It is important to predict not only the dependent
variable (e.g. displacements, temperatures, etc.) but also its derivatives fairly
accurately.
1
Quadratic
Quadratic
Sinusoidal
Sinusoidal
Exact
Exact
First derivative of deflection
00 0.1
0.1 0.2
0.2 0.3
0.3 0.4
0.4 0.5
0.5 0.6
0.6 0.7
0.7 0.8
0.8 0.9
0.9 11
-1
Axial location
Fig. 2.15(a) Comparison of first derivatives of deflection (Example 2.10).
3. We can use a higher order polynomial to improve our approximation. However, one
of the primary purposes in converting the weighted residual form into the weak form
is to be able to work with lower order polynomials. Thus we would like to improve
our approximation by retaining the lower order polynomials. We propose to achieve
this by using many such polynomials over the domain. We now describe the technique
of approximating the solution in a piece-wise manner, rather than using a single
continuous trial function for the whole domain.
2.5 Piece-wise Continuous Trial Function Solution of the

Weak Form
We have seen that the general method of weighted residual technique consisted in assuming a
trial function solution and minimising the residual in an overall sense. The Galerkin method gave
us a way of choosing the appropriate weighting functions. We have so far used different trial
functions such as polynomial and trigonometric series. However, in each case the trial function
Axial location
00
00 0.1
0.1 0.2
0.2 0.3
0.3 0.4
0.4 0.5
0.5 0.6
0.6 0.7
0.7 0.8
0.8 0.9
0.9 1
Second derivative
Quadratic
Quadratic
Sinusoidal
Sinusoidal
Exact
Exact
3.5
-3.5
Fig. 2.15(b) Comparison of second derivatives of deflection (Example 2.10).
chosen was a single composite function, valid over the entire solution domain. For example,
c1 sin (p x/L) used in Example 2.4 was valid over the entire solution domain 0 < x < L, i.e.,
the entire beam in this case. Similarly, the trial function c11 sin (p x/a) sin (p y/b) used in
Example 2.7 was valid over the entire solution domain 0 < x < a, 0 < y < b, i.e., the entire plate.
When we consider the essence of the WR method, i.e., assuming a trial function solution
and matching it as closely as possible to the exact solution of the given differential equation
and the boundary conditions, we realise that this is essentially a process of curve fitting. It
is well known that curve fitting is best done piece-wise; the more the number of pieces,
the better the fit. Figure 2.16(a) illustrates this basic idea on a function f (x) = sin (p x/L) being
approximated using straight line segments. Since a straight line can be drawn through any two
points, we can generate one such approximation by drawing a line through the function values
at x = 0 and x = L/2. Clearly, this is a poor approximation of the function for x > L/2. If we
were to use two straight line segments instead of only one, we could draw two line segments
one through the function values at x = 0 and x = L/2; and the other through the function values
at x = L/2 and x = L. Each of these is our approximation to the function within that piece of
the solution domainthe former segment in the sub-domain 0 < x < L/2, and the latter in the
sub-domain L/2 < x < L. Clearly, this is a better approximation to the function. It can be further
improved by taking four-line segment approximation as shown in Figure 2.16(b).
This is the essential idea of piece-wise continuous trial function approximation for the
weak form. Of course, when we consider piece-wise trial functions, we need to ensure
continuity of the field variable and its derivatives at the junctions. While the exact solution will
ensure continuity of the field variable and ALL the derivatives, we expect to be able to satisfy
this to derivatives of only desired degree with our approximate solution. This leads to an
2
f (x)
One-segment
Two-segment
1
f (x)
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Fig. 2.16(a) One and two-line segment approximation of a function.
1
Exact
0.9 Four-segment
0.8
0.7
0.6
f (x)
0.5
0.4
0.3
0.2
0.1
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x
Fig. 2.16(b) Four-line segment approximation of a function.
inherent error in our solution. We aim, however, to obtain a reasonably accurate solution using
several convenient trial functions (e.g. simple polynomials) defined in a piece-wise manner over
the entire solution domain.
Let us define these trial functions, each valid in its own sub-domain, in terms of the
function values at the ends of the sub-domain. For example, for Figure 2.16, we define the trial
functions in each case as follows: for the one-line segment (Figure 2.16(a)) approximation,
f (0.5) 0
f (X) X = f (0.5) * 2X, 0 < X < 1 (2.109)
0.5 0
for the two-line segment (Figure 2.16(a)) approximation,
f (X) f (0.5) * 2 X, 0 < X < 0.5 (2.110)
0 f (0.5)
f (X) f (0.5) + (X 0.5), 0.5 < X < 1 (2.111)
1 0.5
for the four-line segment (Figure 2.16(b)) approximation,
f (X) f (0.25) * 4X, 0 < X < 0.25 (2.112)
f (0.5) f (0.25)
f (X) f (0.25) + (X 0.25), 0.25 < X < 0.5 (2.113)
0.5 0.25
f (0.75) f (0.5)
f (X) f (0.5) + (X 0.5), 0.5 < X < 0.75 (2.114)
0.75 0.5
f (1) f (0.75)
f (X) f (0.75) + (X 0.75), 0.75 < X < 1 (2.115)
1 0.75
For two-line and four-line segmentation, we can, in fact, more conveniently rewrite these
trial functions if we define a local coordinate x with the origin fixed at the left end of each sub-
domain as follows (ref. Figure 2.16(c)):
For the two-line segment (Figure 2.16(c)) approximation (with l = L/2 = 1/2),
First sub-domain f (x) [1 (x/l)] f (0) + [x/l] f (0.5), 0 < x < l (2.116)
Second sub-domain f(x) [1 (x/l)] f (0.5) + [x/l] f (1), 0 < x < l (2.117)
For the four-line segment (Figure 2.16(c)) approximation (with l = L/4 = 1/4),
First sub-domain f (x) [1 (x/l)] f (0) + [x/l] f (0.25), 0 < x < l (2.118)
Second sub-domain f(x) [1 (x/l)] f (0.25) + [x/l] f (0.5), 0 < x < l (2.119)
Third sub-domain f(x) [1 (x/l)] f (0.5) + [x/l] f (0.75), 0 < x < l (2.120)
Fourth sub-domain f(x) [1 (x/l)] f (0.75) + [x/l] f (1), 0 < x < l (2.121)
xx == 00 xx ==ll
x
1 2
x
xx == 0 xx == ll
L/2
L/2 L/2 L/2
x = 0
x =0 xx == ll x = 0
x=0 xx ==ll
x x
1 2 3 4
x
x
xx== 00 xx ==ll x = 0
x=0 xx == l
X
X
X = 0 X = L
X =0
L/4 L/4 L/4 L/4

L/4 L/4 L/4 L/4
Fig. 2.16(c) Use of local coordinate frames.
We observe that we have used the characteristic functions [1 (x/l)] and [x/l] to
interpolate the value of the function at any point within the sub-domain from the values of the
function at the ends of the sub-domain. We can thus generate very conveniently any desired
number of such line segments (say, n) to approximate the function. For the kth line segment
we have
f (x) [1 (x/l)] fk1 + [x/l] fk, 0 < x < l = L/n (2.122)
These characteristic functions [1 (x/l)] and [x/l] used in our interpolation are called
interpolation functions or shape functions since they determine the shape of variation of the
function within each sub-domain. We will use the notation Ni(i = 1, 2, in this case) to designate
the shape functions.
These characteristic functions dictate the contribution of a given fk to the value of the
function at any point P within the domain 0 < X < L. Figure 2.17 shows these interpolation
functions for the two-line segment case discussed earlier. Figure 2.17(a) shows the interpola-
tion functions separately for the two sub-domains while the Figure 2.17(b) shows them more
compactly for the whole domain 0 < X < L. Figure 2.18 shows these shape functions for the
four-line segment case.
x x x x
1 1
A A A A
1.0 1.0 1.0 1.0

1
x= 0 x=l x= 0 2 x=l
1
(a) Interpolation function within each sub-domain
N2
N1 N3
1.0 1.0 1.0
1 x=l 2
(b) Compact representation of shape function
Fig. 2.17 Linear interpolation functions.
We observe that we have as many shape functions Nk as there are function values fk used
in the interpolation. Since we have chosen linear shape functions, each function Nk ramps up
and down within each sub-domain. Each Nk takes on the value of unity at the point x = xk and
goes to zero at all other end-points x = xi(i k). This is to be expected since, at x = xk, the
full contribution to the value of the function comes from fk alone and none else. Equally
importantly, the nature of the interpolation we have used is such that each fk contributes to the
value of the function only within the sub-domains on its either side or, in other words, only
in those sub-domains to which it is connected.
With respect to the solution of differential equations, the solution to be determined can
itself be considered to be the function f (x) which is being approximated in a piece-wise manner
as discussed above. Thus, all the values of the function fk, k = 1, 2, 3, , n, replace the
coefficients c1, c2, K, which were earlier determined satisfying the weak form. Considering
the computation of the integrals over the entire domain, when we used a single trial function
we achieved this task by evaluating the integral Wi(x)Rd (x) dx over the domain (0 < X < L).
With the use of interpolation functions as explained above, we need to rewrite this as
S Wi(x) Rd (x) dx, where the integration is carried out over each sub-domain and the
summation is carried out over all the sub-domains. By interpolating the value of the
function at an interior point from the function values at the ends of the sub-domain,
we gain two advantagesfirst, continuity of the function at the junction of two sub-
domains is assured and, secondly, the prescribed essential boundary conditions can be
readily imposed. Since the weak form is equivalent to the differential equation and the natural
boundary condition, the essential boundary condition has to be satisfied by the trial function and
thus the second advantage comes in handy.
We reiterate the fact that the unknown coefficients to be determined by the weak form
are now the values of the field variable fk at the ends of the sub-domains. If we discretise the
physical domain 0 < X < L into n sub-domains, we thus have n + 1 unknown values of the
function to be determined through satisfying the weak form. We expect to generate n + 1
algebraic equations, compactly written in matrix form as [A] {f } = {b}. The n + 1 weighting
functions to generate the necessary equations are the interpolation functions themselves,
following the Galerkin procedure. These interpolation or weighting functions are the same as
those shown in Figures 2.17 and 2.18. It is observed that these functions simply repeat
themselves from one sub-domain to another. Thus, it is convenient for us to do our
1.0
N1
1 2 3 4
1.0
N2
1 2 3 4
1.0
F
N3
1 2 3 4
1.0
N4
1 2 3 4
1.0
N5
1 2 3 4
Fig. 2.18 Interpolation functions for four-segment approximation.

computations such as evaluation of the integral Wi(x)Rd(x) dx or its weak form, just once for
the kth sub-domain, and merely reuse the same result for all other sub-domains when we
perform the summation over all the sub-domains. We note that the contributions from each
sub-domain to the elements of the matrix [A] above will essentially be the same except that the
corresponding coefficients they multiply will keep shifting from one sub-domain to another as
given by fk and fk+1. Thus the process of evaluating the total weighted residual or its weak form
over the entire domain is greatly simplified and is summarised now.
Evaluate the sub-domain level contributions to the weighted residual by merely
computing the integral Wi(x)Rd(x) dx or its weak form, just once for the kth sub-
domain.
Build up the entire coefficient matrices [A] and {b} by appropriately placing these sub-
domain level contributions in the appropriate rows and columns.
Solve the (n + 1) algebraic equations to determine the unknowns, viz., function values
fk at the ends of the sub-domains.
Once the function values fk at the ends of the sub-domains have been determined in this
manner, the value of the function at any interior point can be obtained using the appropriate
interpolation functions.
This is the essence of the finite element method. Each of the sub-domains is called a finite
elementto be distinguished from the differential element used in continuum mechanics. The
ends of the sub-domain are referred to as the nodes of the element. We will later on discuss
elements with nodes not necessarily located at only the ends, e.g. an element can have mid-
side nodes, internal nodes, etc. The unknown function values fk at the ends of the sub-domains
are known as the nodal degrees of freedom (d.o.f.). A general finite element can admit the
function values as well as its derivatives as nodal d.o.f. The sub-domain level contributions to
the weak form are typically referred to as element level equations. The process of building
up the entire coefficient matrices [A] and {b} is known as the process of assembly, i.e.,
assembling or appropriately placing the individual element equations to generate the system level
equations.
We will now illustrate the finite element formulation based on the weak form through a
simple example. Through this example we will be formulating the one-dimensional bar finite
element which is a line element with two nodes and the axial displacement u at each node
as the nodal d.o.f. This is a very useful element to solve problems of complex network of
trusses. When combined with a one-dimensional beam finite element, it can be used for solving
any complicated problem involving frame structures, automobile chassis, etc.
2.6 One-dimensional Bar Finite Element

Let us reconsider Example 2.9 where we solved the problem using the weak form, but this time
we will use piece-wise defined interpolation functions. From Eq. (2.85), the weak form of the
differential equation can be written as
L
du L du dW L
W ( X ) AE dX
0
0
AE
dX dX
dX + 0
W ( X )q ( X ) dX = 0 (2.123)
One-dimensional Bar Finite Element 49
We use the symbol X for global coordinate running along the length of the entire
bar (0 to L), and the symbol x for representing the local coordinate within each element
(0 to l). Rewriting Eq. (2.123), we have
L
L du dW L du
0
AE

dX dX
dX =
0
W ( X ) q ( X ) dX + W ( X ) AE
dX 0
(2.124)
Let us discretise the bar into n finite elements, not necessarily of the same length, as
shown in Figure 2.19. Figure 2.20 shows a typical bar element of length l, two nodes and axial
displacement u as the nodal d.o.f.
x = 0 x = l
x
11 22 k n
X
X = L
X = 0 Fig. 2.19 Mesh of n bar elements.
k k + 1
k
x = 0 x = l
uk Nodal d.o.f. uk+1
Fk Nodal forces Fk+1

Fig. 2.20 Typical bar element.
For this n-element mesh, let us rewrite the above equation as

n n n A
A du dW A du

k =1
0
AE
dx dx
dx =
k =1
0
W ( x)q( x) dx + k =1
W ( x) AE dx
0
(2.125)
where we have replaced the integral over 0 to L, by summation over all the finite elements.
Let (x) within each element be given by the interpolation functions introduced earlier, i.e., for
a kth element,
x x
u ( x) = 1 uk + uk +1 = [N] {}k (2.126)
A A
where we have introduced the generic representation of interpolation for the field variable within
an elementvalid for any element, with any number of nodes or nodal d.o.f.for generalising
our method of formulation. It is observed that the shape function array [N] and the nodal d.o.f.
array {}k are of appropriate size depending on the element under consideration. For the present
case,
x x uk
[ N ] = 1 A , {u}k = (2.127)
A uk +1
du 1 1 uk 1 uk
uk uk 1 [ B]{u}k (2.128)
dx A A A
where
1 1
[B] =
A A
Let us now compute the contribution of the kth element to the LHS of Eq. (2.125) above,
using
W1(x) = (1 x/l ), dW1/dx = 1/l
W2(x) = (x/l ), dW2/dx = 1/l (2.129)
Contribution to LHS of Eq. (2.125)

Now,
A du dW1 A uk +1 uk 1
0
AE
dx dx
dx = 0
AE
A
dx
A
AE
= (uk uk +1)
A
AE uk
= [1 1] (2.130)
A uk +1
Similarly,
A du dW2 A uk +1 uk 1
0
AE
dx dx
dx = 0
AE
A
dx
A
AE
= (uk+1 uk )
A
AE uk
= [1 1] (2.131)
A uk +1
Contribution to RHS of Eq. (2.125)

We observe that the first term on the RHS of Eq. (2.125) requires specifying the nature of
externally applied force. Assuming q(x) = q0 to be a constant, we can compute the RHS terms
as
A A x q0 A
W (x)q( x) dx =
0
1
0
1 q0 dx =
A 2
(2.132)
A A x q0 A
0
W2 ( x)q( x) dx =
0
q0 dx =
A 2
For any other specified q(x), similar computation can be readily performed. Contributions
to the second term of RHS can be computed as follows:
A A
du x
W1 ( x) AE dx = 1 A P = P0
0 0
(2.133)
A A
du x
W2 ( x ) AE dx = A P = PA
0 0
where we have used the fact that the term AE(d/dx) stands for the axial force P in the bar
at that section. Thus, P0 and Pl stand for the forces at either end of the element. It is observed
that these designate the algebraic sum of the externally applied forces F at these nodes and the
internal reaction forces R (i.e., those forces that are exerted on this element by the adjoining
elements), that get exposed when we consider this element alone as a free body.
Combining Eqs. (2.130) (2.134), we can write the contributions of the kth element to
the overall system equations given in Eq. (2.125) as
LHS RHS
AE 1 1 uk q0 A /2 P0
+ (2.134)
A 1 1 uk +1 q0 A /2 PA
These two sets of terms constitute the characteristic matrices or characteristic
equations for the bar element. We observe that (AE/l) represents the axial stiffness of a rod;
the first force vector represents the nodal forces equivalent to the distributed force q0, and the
element equations are simply the nodal force equilibrium equations. In view of the fact that
these equations represent the force-deflection relations for the bar element, the LHS matrix is
termed as the element stiffness matrix and the RHS vectors as the element nodal force vectors.
Now we need to sum up all the element level contributions to generate the system level
equations. As already mentioned, this process would simply involve repetitive use of the generic
element matrices given above. To illustrate how the element matrices get simply placed
appropriately in system level equations, we show the detailed computations for a two-element
mesh. Based on this illustration, we provide a general method of assembling element equations
for later use.
Example 2.11. Illustration of assembly. Consider a rod of total length L = 2l, discretised
into two elements, as shown in Figure 2.21. The weak form of the differential equation is
given by
L
L du dW L du
0
AE
dx dx
dx = 0
Wq dx + WAE
dx 0
(2.135)
We observe that the weighting functions can be depicted as shown in Figure 2.22, similar
1 1 2 2 3
1 , 2 Element nos.
1, 2, 3 Node nos.
l = L/2
Local node Local node

number number
1 2 1 2
1 1 2 2 2 3
Global node Global node

number number
Fig. 2.21 Two-element discretisation of a rod.
1.0
N1,W1
1 2
1.0
N2,W2
1 2
1.0
N3,W3
1 2
Fig. 2.22 Weight (shape) functions for two linear bar elements.
to Figures 2.172.18 given earlier. In other words, we can write the weighting function as
follows:
W1(x) = 1 x/l, 0 < x < l in the 1st element
= 0 in the 2nd element
W2(x) = x/l, 0 < x < l in the 1st element
= 1 x/l, 0 < x < l in the 2nd element (2.136)
W3(x) = 0 in the 1st element
= x/l, 0 < x < l in the 2nd element
We have, in all, three field variables u1, u2 and u3, and we have three weighting functions
required to generate the three equations necessary to solve for the three field variables.
Substituting these in the weak form, we get the necessary equations as follows:
LHS of Eq. (2.124)
2
du dW1 A du dW1

L
0
AE
dX dX
dX =
1
0
AE
dx dx
dx
A u u1 1
= 0
AE 2
A
dx + 0
A
AE
= (u1 u2 ) (2.137)
A
2
L du dW2 A du dW2
0 AE dX = 0 AE dx
dX dX 1 dx dx
A u2 u1 1 A u3 u2 1
0 AE A
dx
A 0 AE A
dx
A
AE (2.138)
[( u1 u2 ) + (u2 u3 )]
A
2
du dW3 A du dW3

L
0
AE
dX dX
dX =
1
0
AE
dx dx
dx
A u u2 1
= 0+ 0
AE 3
A
dx
A (2.139)
AE
= [(u3 u2 )
A
RHS first term of Eq. (2.124)
2
A A x q0 A
Wq
L

0
W1q0 dX =
1
0
1 0 dx =
0
1 q0 dx + 0 =
A 2
(2.140)
2
A A x A x q0 A qA

L

0
W2 q0 dX =
1
0
W2 q0 dx =
0
q0 dx +
A
0
1 q0 dx =
A 2
+ 0
2
(2.141)
2
A A x q0 A

L
0
W3 q0 dX =
1
0
W3 q0 dx = 0 + 0
q0 dx =
A 2
(2.142)
RHS second term of Eq. (2.124)
L 2 A A
du du x
W1 ( X ) AE dX =
0
1
W1 ( x) AE dx = 1 A P + 0 = P0
0 0
(1)
(2.143)
Note: The superscript on the force term (e.g. P0(1) ) represents element number and the
subscript refers to x = 0 or l.
L 2 A
du du
W2 ( X ) AE dX =
0
1
W2 ( x) AE dx
0
A A
x x
= P + 1 P
A 0 A 0
= PA(1) P0(2) (2.144)
L 2 A A
du du x
W3 ( X ) AE dX =
0

1
W3 ( x) AE dx = 0 + A P = PA
0 0
(2) (2.145)
Thus the system equations are (from Eqs. (2.137)(2.145))
1 1 0 u1 q0 A /2 P0(1)
AE
1 1 + 1 1 u2 = q0 A /2 + q0 A /2 + PA(1) P0(2)
A 0 1 1 u3 q0 A /2 (2)
PA
q0 A /2 F1
= q0 A /2 + q0 A /2 + F2 (2.146)
q0 A /2 F3
With respect to the system of equations given here, we can make the following
observations:
1. The system level equations can be readily built up from the element level equations by
appropriately placing them in the system matrices.
2. Since each bar element can be viewed as a spring of stiffness AE/l, the bar element
mesh can be viewed as an assemblage of springs connected together at the junction
nodes. Since they share the same nodal displacement, they act as springs in parallel
at the junction node, and hence their individual stiffness gets added up (for example,
at junction node 2 here).
3. The first term on the RHS represents the equivalent nodal force for the distributed
force and, at the junction nodes, the forces too get added up.
4. We may recall that the second term on RHS represents net axial force at the section,
consisting of any concentrated external forces (F) applied at that section as well as
the internal reaction forces (R) that get exposed when we view a single element as
a free body. However, when we assemble the elements together to get the system level
equations, the internal reaction forces annul each other (based on Newtons third law),
leaving out the externally applied concentrated nodal forces.
5. Each of the equations represents the nodal force equilibrium equation. Thus, by
viewing the structure as being made up of a set of discrete finite elements connected
together at only their nodes, we have effectively replaced the differential equations of
equilibrium of continuum mechanics by the present set of algebraic equilibrium
equations applicable only at the nodes. This leads us to the interesting question of
whether or not equilibrium equations are satisfied at interior points in a finite element.
It can be shown that we cannot guarantee that the equilibrium equations are satisfied
at interior points in a general finite element.
6. We may recollect that the weak form was equivalent to only the differential equation
and the natural boundary conditions. Thus the finite element equations given above,
derived from the weak form, cannot be solved until we substitute for the essential
boundary conditions also. In the present context, the essential boundary conditions
refer to prescribed values of the displacement u.
We illustrate the solution of finite element equations using the following two examples.
Example 2.12. A tip loaded rod. Consider a 1 m long steel rod (3 10 mm cross-section)
held fixed at its left end and subjected to a concentrated force of 100 N at its right end. Let
us use the two-element mesh of Example 2.11 to model this problem. Setting
A = 30 mm2 = 30 106 m2, E = 2 1011 N/m2,
q0 = 0, F1 = R1 (the support reaction force)
F2 = 0, F3 = 100 N
We have the finite element equations as follows:
1 1 0 u1 R1
6
(30 10 )(2 10 )
11
1 2 1 u2 = 0
(0.5) (2.147)
0 1 1 u3 100
At the left end, the rod is clamped and thus the essential boundary condition is that
u1 = 0. It is observed that the unknowns in these equations are therefore the nodal
displacements u2, u3 and the ground reaction force R1. We therefore make a mental picture of
partitioning the system matrices given in Eq. (2.147) above and use the last two equations for
determining u2, u3 and the first equation for obtaining the reaction force R1. Solving, we get
u2 = 0.00834 mm, u3 = 0.01667 mm
R1 = 100 N (2.148)
F L
We observe that the tip deflection tallies with the exact value = 3 .
AE
Example 2.13. Rod under distributed and concentrated forces. Consider the same rod as
in Example 2.12 but subjected to forces as shown in Figure 2.23.
q = 100 N/m
100 N
1 2 3 4 5
1 2 3 4
l = 0.25 l l l
Fig. 2.23 Rod under distributed and concentrated forces (Example 2.13).
We will use a four-element mesh depicted in the figure. We have the element level equations
as
(30 106 )(2 1011 ) 1 1 u1 12.5 P0(1)

= +
(0.25) 1 1 u2 12.5 PA(1)
(30 106 )(2 1011 ) 1 1 u2 12.5 P0

(2)
= + (2)
(0.25) 1 1 u3 12.5 PA
(30 106 )(2 1011 ) 1 1 u3 0 P0(3)

= +
(0.25) 1 1 u4 0 PA(3)
(30 106 )(2 1011 ) 1 1 u4 0 P0(4)

= + (2.149)
(0.25) 1 1 u5 0 PA
(4)
One-dimensional Heat Transfer Element 57
Assembling together all the element level equations, we get the system level equations as
1 1 0 0 0 u1 12.5 R1

1 2 1 0 0 u2 25 0
(30 106 )(2 1011 )
0 1 2 1 0 u3 = 12.5 + 0 (2.150)
(0.25)
0 0 1 2 1 u4 0 100

0 0 0 1 1 u
5 0 0
Here, R1 = P0(1) represents the reaction force to be exerted by the support. Since there is no
externally applied concentrated force at node 2 or 3, we have PA(1) = P0(2) and PA(2) = P0(3). At
node 4 we have PA(3) P0(4) = 100 , and node 5 being free edge, PA(4) = 0 .
Substituting for the essential boundary condition that u1 = 0, we can solve the equations
to get the unknown nodal deflections and the support reaction force as
u2 = 0.00573 mm, u3 = 0.0104 mm, u4 = 0.0146 mm, u5 = 0.0146 mm
R1 = 150 N (2.151)
We observe that the axial deflection of nodes 4 and 5 is the same as expected (rigid body
motion). The reaction force at support equals the total external force applied on the structure.
We have demonstrated how to develop the finite element equations, starting from the
weak form of the differential equation, for the axial deformation of a bar. We will now show
that our method is quite generic and can be applied to any given differential equation. We
demonstrate that exactly the same procedure can be used to formulate, for example, the heat
transfer element.
2.7 One-dimensional Heat Transfer Element

Consider a plane wall with a uniformly distributed heat source. The governing differential
equation for the steady state one-dimensional conduction heat transfer is given by
d 2T
k + q0 = 0 (2.152)
dX 2
where k is the coefficient of thermal conductivity of the material, T represents the temperature,
and q0 the internal heat source per unit volume. We assume that k is not a function of
temperature.
The weighted residual statement can be written as
L d 2T
0
W k
dX 2
+ q0 dX = 0

(2.153)
Integrating by parts, the weak form of the differential equation can be obtained as
L
dT L dW dT L
Wk dX
0
0
k
dX dX
dX +
0
Wq0 dX = 0 (2.154)
i.e.
L
L dW dT L dT
0
k
dX dX
dX = 0
Wq0 dX + Wk
dX 0
(2.155)
The weak form for a typical mesh of n finite elements can be written as
n
A dW dT
n A dT
A
k =1
0 k dx dx dx =
k =1

0
Wq0 dx + Wk

dx 0
(2.156)
The attention of the reader is drawn to the similarity of the weak form in the above
equation and that in Eq. (2.125) for the bar problem. We now propose to develop the one-
dimensional heat transfer element following the procedure we used while developing the
equations for the bar element. Figure 2.24 shows a typical element, with its nodes and nodal
degree of freedom (viz., temperature). Following the similarity of the bar and heat transfer
Element k Tk+1
Tk
Node k Node k + 1
x = 0 x=l
Fig. 2.24 Typical linear, heat conduction element.
element, we simply use the same shape functions for interpolating the temperature within the
element from the nodal temperatures. Thus,
x x
T ( x) = 1 Tk + Tk +1 (2.157)
A A
dT T Tk
= k +1 (2.158)
dx A
x dW1 1 x dW2 1
W1 = 1 , = , W2 = , = (2.159)
A dx A A dx A
We now compute the element level contributions to the weak form (Eq. (2.156)) above.
LHS first term of Eq. (2.156). Similar to the LHS of the bar element equation, k
replaces AE and T replaces u. Thus we have
k 1 1 Tk
(2.160)
A 1 1 Tk +1
One-dimensional Heat Transfer Element 59
RHS first term of Eq. (2.156). Since it is similar to the RHS first term of the bar
element equation, we can write
A /2
( q0 ) (2.161)
A /2
RHS second term of Eq. (2.156). Since it is similar to the RHS second term of bar
element, we have
Q0
(2.162)
QA
where Q0, Ql represent the net heat flux at the ends of the element.
Thus the element level equations can be written as
k 1 1 Tk A/2 Q0
= (q0 ) +
A (2.163)
1 1 Tk +1 A/2 QA
The LHS may be considered to be the element conductance matrix and the first term
on the RHS is the equivalent nodal heat flux corresponding to the distributed heat source/sink.
The second vector on the RHS refers to the net heat flux at the nodes and permits us to
incorporate prescribed natural boundary conditions. When the individual elements are assembled
together, these Qs algebraically sum up to the resultant externally applied heat flux at the
section.
The attention of the reader is drawn to the similarity of finite element equations of a bar
and the present heat transfer element, and we note that this is due to the similarity of the
governing differential equations and the shape functions used. Thus, essentially the same
equations can be used for bar and heat transfer problems except that the variables have different
physical meaning in their respective context. It is due to this fact that a single commercial finite
element software package is able to solve a wide variety of physical problems.
We will illustrate the use of this heat transfer element with the following example.
Example 2.14. Plane wall with uniformly distributed heat source. Consider a four-element
mesh for a plane wall of thickness 2L, as shown in Figure 2.25. The length of each element
is given as l = 2L/4 = L/2. The temperatures at the left and the right ends of the wall are
prescribed as Tw. The assembled equations are
q0 A
1 1 0 0 0 Tw 2 Qleft

1 2 1 0 0 T2 q A 0
0

k
0 1 2 1 0 T3 = q0 A + 0 (2.164)
A
0 0 1 2 1 T4 q0 A 0

0 0 0 1 1 T A Q
w q0 right

2
Wall temperature Wall temperature

Tw Tw
T1 T2 T3 T4 T5
1 2 3 4 5
1 2 3 4
l l l l
2L
Fig. 2.25 Plane wall with uniform heat source (Example 2.14).
The unknowns are the temperatures at the interior nodes and the heat fluxes at the left
and right ends of the wall. We visualise an imaginary partition of the above equation, using the
first and the last rows for determination of heat fluxes and the middle three rows for the
determination of temperatures. Rearranging the three equations for the three temperature
variables, we have
2 1 0 T2 q0 A Tw
k

1 2 1 T3 = q0 A + 0
A (2.165)
0 1 2 T4 q0 A Tw
Solving, we get (using l = L/2)

3q0 L2
T2 = T4 = + Tw
8k
(2.166)
q L2
T3 = 0 + Tw
2k
It is easily verified that these results tally with the exact solution.
Problems 61
SUMMARY
We have described the formulation of the finite element equations, assuming the weak form of
any differential equation. We have also illustrated the procedure through two typical one-
dimensional examplesaxial deformation of a rod and the temperature distribution in a wall. We
have also demonstrated how to incorporate the boundary conditions and obtain a solution with
a given finite element mesh. We will now summarise the procedure for finite element analysis
given any differential equation. Our methodology will not be restricted to problems of structural
mechanics alone but can equally be applied to other fields.
Procedure for finite element analysis starting from a given differential equation
1. Write down the Weighted Residual statement.
2. Perform integration by parts and develop the weak form of the WR statement. The
integration by parts is to be carried sufficient number of times to evenly distribute the
differentiation between the field variable and the weighting function. It is observed
that the weak form is equivalent to the differential equation and the natural boundary
conditions.
3. Rewrite the weak form as a summation over n elements.
4. Define the finite element i.e. the geometry of the element, its nodes, nodal d.o.f.
5. Derive the shape or interpolation functions. Use these as the weighting functions also.
6. Compute the element level equations by substituting these in the weak form.
7. For a given topology of finite element mesh, build up the system equations by
assembling together element level equations.
8. Substitute the prescribed boundary conditions and solve for the unknowns.
PROBLEMS
2.1 Solve the following equation using a two-parameter trial solution by (a) the point-
collocation method (Rd = 0 at x = 1/3 and x = 2/3); (b) the Galerkin method. Then,
compare the two solutions with the exact solution.
dy
+ y = 0, 0 x 1
dx
y(0) = 1
2.2 Solve the following equation using the Galerkins method (Use at least a two-parameter
solution):
dy
= 50(1 + cos x) 0.05 y, 0 x 2p
dx
y(0) = 200
2.3 Determine a three-parameter solution of the following using the Galerkins method and
compare it with the exact solution:
d2y
= cos Q x, 0 x 1
dx 2
u(0) = 0, u(1) = 0
2.4 Determine a two-parameter Galerkin approximation solution of the differential equation:
d 2u
= f0 , 0 < x < L
dx2
du
= 0 at x = 0 and L
dx
2.5 Give a one-parameter Galerkin solution of the following equation, for the two domains
shown in Fig. P2.5:
2u 2u
2 + =1
x y 2
(0,1) u = 0 (1,1)
60
u = 0 u = 0 u = 0
u = 0 W
60 60
(0,0) (1,0)
u = 0 u = 0
Fig. P2.5 (a) Square domain and (b) Triangular domain.
2.6 Develop the weak form and find the solution (Fig. P2.6).
d 4v
EI q0 = 0
dx 4
dv q0
v(0) = 0; (0) = 0 ,
dx M
2
d v L
EI (I ) = M ,
dx 2 P
3
d v Fig. P2.6 Cantilever beam.
EI (I ) = P
dx3
Problems 63
2.7 For fully developed fluid flow in an annular pipe (Fig. P2.7), the governing
equation is
1 d du
rN = k
r dr dr
ur = a = 0, u r =b = 0
Find the solution using the weak form.
r = a
r = b
Fig. P2.7 Flow through annular pipe.
2.8 Solve the following by making use of weak formulation (Fig. P2.8):
d 4v
EI q0 = 0
dx 4 q0
dv
v(0) = 0, (0) = 0
dx
dv L
v( L) = 0, ( L) = 0
dx
Fig. P2.8 Fixed-fixed beam.
2.9 For heat conduction through an annulus region (Fig. P2.9), the governing equation is
given below along with the boundary conditions. Derive the weak form and find the
solution
d 2T 1 dT q
2
+ + 0 = 0
dr r dr k
T(ri) = Ti, T(ro) = To
ri
Ti To
ro
Fig. P2.9 Heat conduction in a disc.
2.10 For the problem given below, find a two-parameter Galerkin approximate solution that
contains trigonometric trial functions.
d2 y Qx
cos 10 = 0
dx 2
A
dy
= 0 at x = 0 and L
dx
2.11 Determine the Galerkin approximation solution of the differential equation
d 2u du
A 2
+ B + C = 0
dx dx
u(0) = u(L) = 0
2.12 A certain problem of one-dimensional steady heat transfer with a distributed heat
source is governed by equation
d 2G
+ G + 1 = 0
dx 2
f = 0 at x = 0
dG
= G at x = 1
dx
Find the Galerkin approximation solution of the above differential equation.
2.13 Determine a two-parameter Galerkin approximation solution of the differential equation
d 2u
= fc , 0 < x < L
dx 2
du at x = 0; x = L
=0
dx
Problems 65
Problem to Investigate
2.14 The coupled, nonlinear governing equations for large deflections of a beam are as
follows:
d du 1 dv
2
AE + + qa = 0
dx dx 2 dx
d dv du 1 dv
2
d 4v
EI AE + qt = 0
dx 4 dx dx dx 2 dx
where A, E, qa, qt are known constants; u and v refer to axial and transverse
deflections. Develop the weak form statement for this nonlinear problem.
CHAPTER 3
Finite Element Formulation
Based on Stationarity
of a Functional
3.1 Introduction
In Chapter 2, we discussed the formulation of the finite element equations from the weak form
of the Galerkin weighted residual statement. Our discussion typically started with the
specification of the governing differential equation and the boundary conditions for the problem
at hand. While the formulation of the finite element method starting from the governing
differential equation is quite general, in many situations of practical importance, the governing
equations can also be given, perhaps more conveniently, in the form of stationarity of a
functional. The term functional is used to represent a function of a function, e.g. an
expression such as
1 du
2

I = AE (q )(u ) dx
2 dx
(3.1)
is a function of u, which itself is a function of x. We state that the problem of finding u(x)
that makes I stationary with respect to small, admissible variations in u(x) is equivalent
to the problem of finding u(x) that satisfies the governing differential equation for this
problem. For example, for most problems of structural mechanics, such a functional
expression represents the total potential of the system, a quantity with direct physical
significance for the structural analyst. The first term in the above expression for I signifies the
strain energy stored in the structure (in this case, just a rod undergoing axial deformation) and
the second term represents the potential of external forces. The Principle of Stationary Total
Potential (PSTP) characterising the equilibrium configuration of a structure is a well-known
principle of structural mechanics. For conservative systems (i.e. no dissipative forces), it can
be stated as
Among all admissible displacement fields, the equilibrium configuration of the system is
that, which makes the total potential of the system stationary with respect to small,
admissible variations of displacement.
66
Functional and Differential Equation Forms 67
Here we use the term admissible to describe the displacement field that satisfies the
essential boundary conditions and does not lead to any voids/openings/kinks, etc. in the
structure. For long, it has been common practice to use energy methods in structural
mechanics and thus for most problems of structural mechanics, it may be convenient to use
the PSTP to formulate the governing equations. The finite element equations can be developed
starting from such a principle governing the stationarity of a functional. In fact, the finite
element method had its origins in the late 50s and early 60s in structural mechanics where it
was purely formulated as a technique to solve problems of structural analysis. Subsequently,
it was realised that the method belonged to the general class of weighted residual methods and
its capability to model a variety of problems in engineering science was fully appreciated. At
present it is used in a wide variety of applications, viz., structural mechanics, heat transfer, fluid
mechanics, biomechanics, electromagnetic field problems, etc.
In what follows, we will first discuss some basic ideas in variational calculus and the
extrema of a functional and show that the governing differential equations (EulerLagrange
equations) can be derived from this. Thus we establish the equivalence of differential equation
representation and the functional representation. We will then discuss the method of finding
approximate solutions to the problem of minimisation of a functional, e.g., u(x) that minimises
I in the example cited above, using a single composite trial function valid over the entire solution
domain. We will then introduce piece-wise continuous trial functions leading to the formulation
of the finite element equations starting from a given functional. We will show that the resulting
finite element formulation is identical to that obtained, based on the weak form of the governing
differential equation.
It must be appreciated that the finite element equations can be derived systematically,
with equal ease, starting from either form of representation of governing equationsdifferential
or functional. Given a functional, we can always find the corresponding differential equation
for the problem, but some differential equation forms may not have a corresponding functional
form. Unlike the total potential of a structure, in many cases the functional (even when it exists)
may not have a direct physical meaning. While traditionally the functional form (via PSTP for
structures) is used to introduce the finite element method, it must be appreciated that the finite
element formulation itself does not require the existence of a functional. In the subsequent
chapters, we will use both these forms interchangeably for developing the necessary finite
elements.
3.2 Functional and Differential Equation Forms

It is beyond the scope of the present discussion to provide the general theory of variational
calculus. The interested reader is advised to refer to the standard texts on the subject. We will
present a brief introduction to the relevant concepts of variational calculus.
It is well known that for a function of single variable, viz., f(x), the condition for an
extremum is given by df/dx = 0. A function of several variables f (x1, x2, , xn) will have an
extremum at a given (x1, x2, , xn) if
f f f
= = " = = 0 (3.2)
x1 x2 xn
68 Finite Element Formulation Based on Stationarity of a Functional
Now consider a functional (i.e. function of functions) such as

b du
I (u ) = a
F u,
dx
, x dx

(3.3)
We wish to find u(x) which makes the functional stationary, subject to the end conditions
u(a) = ua and u(b) = ub. As an example, if F = (1/2)AE(du/dx)2 (q)(u), and the prescribed
end conditions are u(0) = 0; u(L) = 0, then this represents the problem of a uniform bar
clamped at both ends and subjected to a distributed load q(x). In Figure 3.1, we represent the
possible set of u (x) as also the as yet unknown solution u(x). We observe that all u (x) satisfy
u(x)
ub
u(x)
u(x)
ua
x
x=a x=b
Fig. 3.1 Admissible solutions u(x).
the prescribed end conditions, and thus we are dealing with only fixed end variations. We wish
to study what happens to I, if u(x) is slightly changed to u (x), i.e.,
u ( x) = u(x) + e v(x) (3.4)

where e is a small parameter and v(a) = 0 = v(b).
The difference between u(x) and u(x) is termed the variation in u(x), and we denote
this by d u(x). Thus,
d u(x) = u (x) u(x) = e v(x) (3.5)
Figure 3.2 shows the distinction between du and d u at a given position x, the variation
d u refers to the difference between u (x) and u(x) while du refers to the differential change
in u(x) as x is changed to x + dx.
u(x)
u(x)
u(x)
du
du
dx
x
x x + dx
Fig. 3.2 Distinction between d u, du.
We define
d (u) = difference in slope of u (x) and u(x)
= u (x) u(x) = u(x) + e v(x) u(x)
= e v(x) = [d u] (3.6)
where ( )' designates differentiation with respect to x. Thus we observe that the process of
variation and differentiation are permutable. Now, for a given x, as we move from u(x) to u (x),
we have
DF = F( u , u , x) F(u, u, x) = F(u + d u, u + d u, x) F(u, u, x) (3.7)
Expanding the first term in Taylors series, we have
F F
F(u + d u, u + d u, x) = F(u, u, x) + Eu + E u
u u
1 2 F 2 2 2 F 2 F
+ 2 Eu + E u E u + E u 2 + " (3.8)
2! u u u u 2

Thus,
F F 1 2 F 2 2 2 F 2F
'F = Eu + E u + 2 Eu + E u E u + E u 2 + " (3.9)
u u 2 ! u u u E u 2

The first variation of F is defined as
F F
EF = Eu + E u (3.10)
u u
and the second variation of F as
2 F 2 2 2 F 2 F
E 2 F = E (E F ) = E u + E u E u + E u 2 (3.11)
u 2
u u u 2
Therefore,
1 2
'F = E F +
E F + " (3.12)
2!
Now let us study what happens to I in the neighbourhood of u(x), i.e.
D I = I( u , u , x) I(u, u, x)
b b
=
a
F (u , u , x) dx a
F (u, u, x) dx
b
= a
'F dx
b 1 2
=
E F +

a 2!
E F + " dx

(3.13)
The first variation of I is defined as

b
EI = E F dx
a
(3.14)
and the second variation of I as
b
E 2I = E
2
F dx (3.15)
a
Therefore,
1
DI = d I + 2 ! d 2I + L (3.16)
If I represents the total potential of a structure and we look for a stable equilibrium
configuration, then we wish to find u(x) that minimises I. Since u(x) minimises I, DI 0. As
e is reduced, DI approaches zero and when u (x) = u(x), I attains a minimum and DI 0.
By studying the relative orders of magnitude of the various terms, it is possible to show
that d I is of the order O(e) while d 2I is O(e 2), etc. When e is sufficiently small, d 2I and higher
variations become negligible compared to d I, and thus the condition for I to be stationary
becomes d I = 0. We propose to show that the variational form given by d I = 0 reduces to
the governing differential equation of the problem. From Eqs. (3.10)(3.14), we have
b F F
EI = a

u
Eu +
u
E u dx

(3.17)
From Eq. (3.6),

d
E (u ) = (E u ) = (E u ) (3.18)
dx
Thus,
b F b F
a u
E u dx = a u
d (E u ) (3.19)
Performing integration by parts, we obtain
b
b F F b d F

a u
d (E u ) = Eu
u a
a
(E u ) dx
dx u
(3.20)
Substituting Eq. (3.20) in Eq. (3.17), we get
b
b F d F F
EI = a

u

dx
E u dx + E u = 0
u u a
(3.21)
Since all the trial functions u(x) satisfy the end conditions at x = a and b, we have
d u(a) = 0 = d u(b) (3.22)
For arbitrary d u, we therefore have, from Eq. (3.21), the relation
F d F
= 0 (3.23)
u dx u
This is known as the EulerLagrange equation. For example, if
2
1 du
F = AE ( q)(u )
2 dx
then we have
F
= q
u
(3.24)
F du
= AE
u dx
Substituting in Eq. (3.23), we get
d 2u
AE + q = 0 (3.25)
dx 2
which is the governing differential equation for a bar subjected to distributed loading q(x).
Let us now rewrite Eq. (3.21) for the specific case of axial deformation of a rod.
We have
b d 2u du
L

a
q AE
dx 2
E u dx + AE E u
dx 0
= 0 (3.26)
with d u(0) = 0; d u(L) = 0 in view of prescribed boundary conditions at x = 0, L.

Recapitulating the weighted residual statement for this problem, we have
d 2u b

a
q + AE 2 W dx = 0
dx
(3.27)
While developing the weak form, we demanded that W(x) be zero at those points where
u is prescribed. Thus, W(0) = 0 = W(L).
Comparing the form of the equations obtained using the weak form and the present
discussion, we observe that the weighting function W(x) used earlier in weighted residual
methods has the connotation of a variation in u(x), i.e. d u(x) here. Since all the variations
of u(x) being considered satisfy the essential boundary conditions, d u(x) at x = a and b is
zero. In the weak form, recall that we required the weighting function to be zero at these
points. Further, for structural mechanics problems, if u(x) denotes the equilibrium
configuration, d u(x) (and hence W) has the connotation of a virtual displacement consistent
with the boundary conditions (i.e., d u(a) = 0 = d u(b)). d I = 0, in essence, is a statement
of the familiar virtual work principle.
We will leave it as an exercise to the reader to show that if

b
I =
a
F (u, u u , x) dx (3.28)
the corresponding EulerLagrange equation is
F d F d 2 F
+ = 0 (3.29)
u dx u dx 2 u
2
d 2w
As an example, let F = ( EI )
1
(q )( w) , which is the case for a beam subjected
2 dx 2
to distributed transverse loading q(x). The resulting EulerLagrange equation is the familiar form
of the governing differential equation of a beam.
Similarly, for a one-dimensional heat transfer problem satisfying the differential equation
d 2T
k + q0 = 0 (3.30)
dx 2
subject to T(0) = T0, q x = L = h(TL T) is equivalent to minimising the functional
2
L 1 dT L 1
I =
0
k dx
2 dx 0
q0T dx +
2
h(T T ) 2 (3.31)
subject to T (0) = T0.

Principle of Stationary Total Potential (PSTP) 73
Similarly, for two-dimensional heat conduction, satisfying the differential equation

2T 2T
k + k + q0 = 0 (3.32)
x 2 y 2
with the boundary conditions
(i) T = T0 on ST,
(ii) qn = q on Sq,
(iii) qh = h(T T) on Sc
is equivalent to minimising the functional
1 T 2 T
2 1
I =
2
A
k
x

+
y
2 q T
0 dA +

S1
q T ds +
Sc 2
h(T T )2 ds
subject to
T = T0 on ST (3.33)
Thus, given a functional I, using the variational principle d I = 0, we can derive the
governing differential equation. Therefore, both forms (differential and functional) are
equivalent.
We will now discuss an approach to find approximate solutions to the problem of
minimisation of a functional, e.g., u(x) that minimises I in the example cited above. We will use,
for our discussion, the popular principle of structural mechanics, viz., the Principle of
Stationary Total Potential, and present the RayleighRitz method of finding approximate
solutions based on this principle. We will initially use a single composite trial function valid over
the entire solution domain, and then introduce piece-wise continuous trial functions leading to
the formulation of the finite element equations starting from a given functional.
3.3 Principle of Stationary Total Potential (PSTP)

Consider a spring-mass system shown in Figure 3.3. Let the position of free length of the
spring be at a height H from the datum. Now, when the mass is attached to the spring, it causes
extension. We wish to illustrate the application of the Principle of Stationary Total Potential in
determining the equilibrium position of the spring-mass system. At any extension of the spring
given by x, the strain energy stored in the spring is (1/2)(kx2). The gravitational potential energy
of the mass with respect to the datum is mg(H x). So we define the total potential of the
system in this configuration to be
1
3 p = ( kx2 ) + mg ( H x) (3.34)
2
Being the total potential of the system, this quantity represents the total amount of energy
that the system can give to an external agency. The principle of stationary total potential
demands that the equilibrium value of x be such that p is stationary with respect to x in that
configuration, i.e.,
k
Position corresponding to
free length of spring
x
g m
H
Datum
Fig. 3.3 Typical spring-mass system.
d3p
= 0 (3.35)
dx
i.e.,
(k)(xeq) (mg) = 0 (3.36)
Thus, xeq is given by (mg/k), as expected. This is a trivial example to introduce the basic
idea. We observe that the choice of the datum is arbitrary and in no way affects the final result.
We thus choose to drop the datum altogether and rewrite p as
1
3 p = (kx 2 ) mgx (3.37)
2
This quantity represents the work to be done by an external observer in bringing the system
to this configuration from x = 0. The weight in this example can as well be any external
(conservative) force P, and we write the general expression for the total potential of a system
as
p = Internal strain energy in the structure + Potential of external forces (3.38)
The potential of external forces is, by convention, defined in such a way that its derivative
(with a negative sign) yields the force. If d represents the generic displacement, then we have
3 p
= 0 (3.39)
E
Thus, for a general multidegree of freedom structure, if d1, d2, d3, K, dn are the
displacements, then we have
3 p
= 0, i = 1, 2, , n (3.40)
E i
Thus, based on this principle, we can formulate a method of finding an approximate
solution to the problem at hand as follows. Just as we did in the weighted residual method
discussed in Chapter 2, we begin with an assumed trial function solution to the problem. For
the deformation of the structure as per this assumed displacement field, we determine the total
potential p. Based on the PSTP, we set up the necessary equations to solve for the unknown
coefficients. We will now discuss the details of the RayleighRitz method, which is popularly
used in structural mechanics.
3.3.1 RayleighRitz Method

The RayleighRitz (RR) method consists of three basic steps:
Step 1: Assume a displacement field. Let the displacement field be given by {f(x) + ciNi},
i = 1, 2, , n, where Ni are the shape functions and ci are the as yet undetermined coefficients.
It is observed that the assumed displacement field should satisfy both the essential boundary
conditions of the problem and internal compatibility, i.e. there should be no kinks, voids, etc.
within the structure.
Step 2: Evaluation of the total potential. For the system under consideration, evaluate the
total potential p consistent with the assumed displacement field in Step 1 above.
Step 3: Set up and solve the system of equations. By virtue of the PSTP, the total potential
will be stationary with respect to small variations in the displacement field. The variations in
the displacement field in our case are attained by small variations in the coefficients ci. Thus
we have
3 p
= 0, i = 1, 2, , n (3.41)
ci
which will yield the necessary equations to be solved for the coefficients ci.
This method was first formulated by Rayleigh (1877) and later refined and generalised by
Ritz (1908). Rayleigh worked with just one term, viz., c1N1 while Ritz extended the technique
to an n-term approximation. However, both of them used shape functions N which were single
composite functions valid over the entire domain of the problem. Further extension of their
technique by using piece-wise defined shape functions Ni will lead us to the variational
formulation of the finite element method.
We will now illustrate the basic scheme of RR method with some simple example
problems.
Example 3.1. A bar under uniform load. Consider a bar clamped at one end and left free
at the other end and subjected to a uniform axial load q0 as shown in Figure 3.4. The governing
differential equation is given by
q0
x = 0 x= l
Fig. 3.4 Rod under axial load (Example 3.1).

d 2u
AE + q = 0 (3.42)
dx 2
du
with the boundary conditions u(0) = 0; = 0.
dx x=L
We discussed the solution using the Galerkin weighted residual method in Example 3.1.
We will now illustrate the solution using the RR method. For a general deformation u(x),
L 1 du
2
Strain energy stored in the bar, U =
0
AE dx
2 dx
(3.43)
L
Potential of the external forces, V = 0
q0 u dx PL u ( L) (3.44)
du
where, for the sake of generality, a tip load PL = AE has been included in the potential
dx x=L
of external forces. In view of the second boundary condition, the expression for the potential
of external forces reduces (for this example) to the following:
L
V = 0
q0 u dx (3.45)
Thus we need to find u(x) that minimises the total potential of the system given by
L 1 du
2
3P = 0
AE q0u dx
2 dx
(3.46)
subject to the essential boundary condition that u(0) = 0. It is to be observed that the force
boundary condition at the free end ( du / dx | x = L = 0) has been incorporated in the expression
for the functional itself and thus the trial function assumed need only satisfy the essential
boundary condition. Also, we observe that the continuity demanded on u(x) is lower (i.e. need
to be differentiable only once) compared to the differential equation form.
Let us now solve for the displacement field using the RR method.
Step 1: Assume a displacement field. Let us assume that
u(x) c1x + c2x2 (3.47)
This satisfies the essential boundary condition that u(0) = 0. We have
du
= c1 + 2c2 x (3.48)
dx
Step 2: Evaluation of the total potential. The total potential of the system is given as
L AE 2
3p = 2 (c1 + 2c2 x) q0 (c1 x + c2 x ) dx
2
0
AE 2 4c22 2 c1 L2 L3
= c1 L + + 2c1c2 L q0 q0 c2 (3.49)
2 3L3 2 3
Step 3: Set up and solve the system of equations. From the principle of stationary total
potential (PSTP), we have
3p
= 0, i = 1, 2 (3.50)
ci
Therefore,
3 p AE q L2
= 0 (2c1 L + 2c2 L2 ) 0 = 0
ci 2 2
3 p AE q L3
= 0 (8c2 L3 /3 + 2c1 L2 ) 0 = 0 (3.51)
c2 2 3
Solving, we obtain
q0 L q0
c1 = , c2 =
AE 2 AE
Thus,
q0 q0
u ( x) = x( L x /2) = (2 Lx x 2 ) (3.52)
AE 2 AE
which is observed to be the same as that obtained in Example 2.1.
Example 3.2. A simply supported beam under uniform load. Consider a simply supported
beam under uniformly distributed load q0 as shown in Figure 3.5. For a deformation v(x), we
have
q0
x = 0 x = L
Fig. 3.5 Simply supported beam under load (Example 3.2).

The strain energy

2
L 1 d 2v
U =

0
EI
2 dx2
dx (3.53)
The potential of the external forces is

L
V =
0
q0 v dx (3.54)
Thus we have the total potential
L EI d 2 v 2
3p = 0

2 dx 2
q0 v dx

(3.55)
Step 1: Assume a displacement field. Let us assume v(x) c1 sin (p x/L). This satisfies the
essential boundary conditions v(0) = 0 = v(L). We have
2
d 2v Q Qx
= c1 sin (3.56)
dx 2
L L
Step 2: Evaluation of the total potential. The total potential of the system is given by
2
2
EI c Q sin Q x dx q c sin Q x dx
L
3p = 0 2 1 L L
0 1
L
Q 4 EI 2q0 L (3.57)
= c12 c1
4 L3 Q
Step 3: Set up and solve the system of equations. From the principle of stationary total
potential, we have
3 p
= 0 (3.58)
c1
Therefore,
Q 4 EI 2q0 L
c1 = 0
2L 3 Q
i.e.
q0 L4
c1 = 0.01307 (3.59)
EI
Thus the final solution is
q0 L4 Qx
v(x) = 0.01307 sin (3.60)
EI L
We observe that this solution is identical to the solution obtained from the Galerkin method
(ref. Example 2.4). In the Galerkin method, the original Weighted Residual statement was
equivalent to only the differential equation and, therefore, required that the trial solution assumed
should satisfy all the boundary conditions of the problem (essential and natural). The RR
method is based on the functional in which the external forces, if any, applied at the boundaries
are taken into account through the potential of the external forces. Thus the trial solution for
the RR method need only satisfy the essential boundary conditions. In general, if the
differential equation form and the functional form of a given problem are both available, then
the Galerkin method and the RR method yield identical solutions when the problem involves
only essential boundary conditions and when they both use the same shape functions.
Example 3.3. Temperature distribution in a pin-fin. Consider a 1 mm diameter, 50 mm long

aluminium pin-fin as shown in Figure 3.6, used to enhance the heat transfer from a surface
wall maintained at 300C. Let k = 200 W/m/C for aluminium; h = 20 W/m2/C, T = 30C.
T
h
Wall
Tw
Fig. 3.6 A pin-fin (Example 3.3).
The temperature distribution in the fin was obtained using the Galerkin method in Example 2.8
based on the governing differential equation. The governing differential equation and the
boundary conditions we used are repeated here for convenience:
d 2T Ph
k 2
= (T T )
dx Ac
T (0) = Tw = 300 C (3.61)
dT
qL = kAc ( L) = 0 (insulated tip)
dx
where k is the coefficient of thermal conductivity, P is the perimeter, Ac is the cross-sectional
area, h is the convective heat transfer coefficient, Tw is the wall temperature, and T the
ambient temperature. The equivalent functional representation is given by
2
L 1 dT L 1 Ph
3 = 0
k dx +
2 dx
0 2 Ac
(T T )2 dx qLTL (3.62)
where, in general, the heat flux qL is included. However, in view of the insulated tip boundary
condition, the equivalent statement of the problem is to find T(x) that minimises the functional
2
L 1 dT L 1 Ph
3 = 0
k dx +
2 dx 0 2 Ac
(T T )2 dx (3.63)
subject to the boundary condition that T(0) = Tw = 300C.

T(x) T(x) = c0 + c1x + c2x2

From the boundary conditions, c0 = 300. Therefore,
T (x) = 300 + c1x + c2x2 (3.64)

Step 2: Compute the functional
L 1 L Ph
k [c1 + 2c2 x ] dx +
2
270 + c1 x + c2 x dx
2
3 =
2
(3.65)
0 2 0
2 Ac
Step 3: Minimise the functional
3 3
= 0, = 0 (3.66)
c1 c2
From Eq. (3.65),
Ph 2 L3 c
k (2c1 L + 2c2 L2 ) + c1 + 270 L2 + 2 L4 = 0
Ac 3 2
(3.67)
8 Ph 2 5 c2 4
k c2 L3 + 2c1 L2 + L c2 + L + 180 L3 = 0
3 Ac 5 2
On substituting the numerical values and solving for the two coefficients c1 and c2, we get
c1 = 3923.36, c2 = 40,498.44 (3.68)
Thus our approximate solution for temperature, based on the minimisation of the
functional, is given as follows:
T(x) = 300 3923.36x + 40,498.44x2 (3.69)
The Galerkin WR solution obtained in Example 2.8 is reproduced here from Eq. (2.69)
for ready reference:
T(x) = 300 + 38,751.43(x2 2Lx) (3.70)
Piece-wise Continuous Trial FunctionsFinite Element Method 81
The two approximations are compared with the exact solution in Table 3.1. Even though
both trial solutions are quadratic, the boundary conditions are different. The RR trial solution
was c1x + c2x2 (with two independent parameters c1 and c2) while the original Galerkin WR
method used c2(x2 2Lx) after satisfying both the boundary conditions.
Table 3.1 Comparison of Galerkin WR and RR Solutions (Example 3.3)
Axial location Exact solution Galerkin WR RR solution

solution
0 300 300 300
0.005 280.75 281.59 281.40
0.01 264.02 265.12 264.82
0.015 249.62 250.59 250.26
0.02 237.43 238.00 237.73
0.025 227.31 227.34 227.23
0.03 219.16 218.62 218.75
0.035 212.91 211.84 212.29
0.04 208.49 207.00 207.86
0.045 205.85 204.09 205.46
0.05 204.97 203.12 205.08
3.4 Piece-wise Continuous Trial FunctionsFinite Element

Method
In the examples so far, we have used a single composite trial function solution for the entire
domain. When we consider the essence of the RR method, i.e., assuming a trial function
solution and matching it as closely as possible to the exact solution, we realise that this is
essentially a process of curve fitting. It is well known that curve fitting is best done piece-
wisethe more the number of pieces, the better the fit. We can thus benefit by using piece-
wise defined trial functions rather than a set of functions valid over the entire domain of the
problem. We will now discuss this technique leading to the formulation of the finite element
method starting from a variational principle. We recall that the evaluation of the total potential
p involved integration over the entire domain of the problem. With the discretisation of the
structure into sub-domains and use of piece-wise defined trial functions, we will evaluate these
integrals over each sub-domain (finite element) and sum up over all the elements. We will
illustrate the procedure on the one-dimensional bar problem and demonstrate that we end up
with exactly the same set of equations as in Chapter 2.
3.4.1 Bar Element Formulated from the Stationarity of a Functional

Consider a typical bar element as shown in Figure 3.7 with two nodes and the axial
displacement u as the nodal d.o.f. We will use the same interpolation functions as used in
Chapter 2, i.e., at any point within the element we have
x x
u = 1 u1 + u2 (3.71)
A A
x = 0 x = l
x
u1 u2
1 2
F1 F2
Thus the strain energy stored within the element is given by

2
A AE du AE (u2 u1 ) 2
Ue = 0
dx =
2 dx 2 A
(3.72)
If there is a distributed force q0 acting at each point on the element and concentrated
forces F at the nodes, the potential of the external forces is given by
A A
Ve =
0
q0 u dx F1u1 F2 u2 = q0 (u1 + u2 ) F1u1 F2 u2
2
(3.73)
Thus the total potential for the kth element is given by
AE (u2 u1 ) 2 qA
3 ep = U e + V e = 0 (u1 + u2) F1u1 F2u2 (3.74)
2 A 2
For the entire structure, summing up the contributions of all elements, we have
n
3p = 3
k =1
e
p (3.75)
By virtue of the PSTP, the total potential will be stationary with respect to small variations
in the displacement field. The variations in the displacement field in our case are attained by
small variations in the nodal d.o.f., ui. Thus,
3 p
= 0, i = 1, 2, , n (3.76)
ui
When applied to Eq. (3.75) above, this will lead to a set of simultaneous linear algebraic
equations in the nodal d.o.f., in general, designated as [K]{d} = { f }. The contributions of the
kth element to these system level equations can be obtained by applying Eq. (3.76) to the
element total potential as given by Eq. (3.74). Thus we have the element level contributions
as
3ep AE qA
= 0 (u1 u2 ) = 0 + F1
u1 A 2
(3.77)
3ep AE qA
= 0 (u2 u1 ) = 0 + F2
u2 A 2
Piece-wise Continuous Trial FunctionsFinite Element Method 83
In matrix form, we write the element level equations as follows:
AE 1 1 u1 q0 A /2 F1
= + (3.78)
A 1 1 u2 q0 A /2 F2
We observe that these are identical to those equations obtained by the weak form of the
Galerkin weighted residual statement, starting from the governing differential equation of the
problem. (Please note that the sense of F1 is opposite to that of P0 in Eq. (2.134).) The process
of assembling together individual element contributions is identical to the assembly procedure
we used earlier.
3.4.2 One-dimensional Heat Transfer Element Based on the Stationarity of a

Functional
Consider a typical one-dimensional heat conduction element as shown in Figure 3.8 with two
nodes and the temperature T as the nodal d.o.f. The governing functional is given by
T1 T2
1 2
Q1 Q2
x = 0 x = l
Fig. 3.8 Typical heat transfer element.
2
1 A dT
3 =
2
0
k dx
dx q T dx
0 Q1T1 Q2T2 (3.79)
where q0 is internal heat source and Q1, Q2 represent heat flux at the nodes. We will use the
same interpolation functions as we used previously, i.e., at any point within the element we have
x x
T ( x) = 1 T1 + T2 (3.80)
A A
Thus the element level functional is given by
1k 2 q0 A
3e = (T2 T1 ) (T1 + T2 ) Q1T1 Q2T2 (3.81)
2 A 2
For the entire system, summing up the contributions of all elements, we have
n
3 = 3
k =1
e
(3.82)
The functional will be made stationary with respect to small variations in the temperature field.
The variations in the temperature field in our case are attained by small variations in the nodal
d.o.f. Ti., Thus,
3 i = 1, 2, K, n (3.83)
= 0,
Ti
When applied to Eq. (3.82) above, this will lead to a set of simultaneous linear algebraic
equations in the nodal d.o.f., in general designated as [K]{d} = { f }. The contributions of the
kth element to these system level equations can be obtained by applying Eq. (3.83) to the
element level functional as given by Eq. (3.81). Thus we have the element level contributions as
k 1 1 T1 q0 A /2 Q1
= + (3.84)
A 1 1 T2 q0 A /2 Q2
We observe that these contributions are identical to those obtained by the weak form of
the Galerkin weighted residual statement, starting from the governing differential equation of the
problem (please note that the sense of Q1 here is opposite to that of Q0 in Eq. (2.162)).
3.5 Meaning of Finite Element Equations

Let us look at the element level equations for a bar element. We have, from Eq. (3.78), with
q0 = 0,
AE 1 1 u1 F1
= (3.85)
A 1 1 u2 F2
i.e.
[K]{u} = {F}
The individual elements of the element stiffness matrix have a direct physical meaning.
Suppose we consider a deformation state for the element wherein we have u1 = 1 and
u2 = 0. From elementary mechanics of materials we know that the force required to cause such
a deformation will be AE/L to be applied at node 1 and, since node 2 is to be held at rest, an
equal and opposite reaction force develops at node 2. From the above equation if we solve for
the forces F1, F2 for this prescribed displacement, we get the same result. We observe that
the elements of the first column of the stiffness matrix are identical to these forces. Similarly,
if we consider a deformation state wherein u2 = 1 and u1 = 0, the forces required will be given
in the second column of the stiffness matrix. Thus elements of each column of a stiffness
matrix actually represent the forces required to cause a certain deformation pattern
the ith column standing for a deformation pattern, where the ith d.o.f. is given unit
displacement (translational or rotational) and all other d.o.f. are held zero.
We can actually attempt to derive the finite element equations based on this direct physical
correlation, called the Direct method of formulation of finite element equations. We will
illustrate the usefulness of such an approach for a beam finite element (to be discussed in detail
in Section 4.5). Our beam element (Figure 3.9) is a line element having two nodes just like the
Meaning of Finite Element Equations 85
v1 v2
F1 F2
M1 M2
1 2
dv
R1 = R2 =
dv
dx x =0 dx x =A
x
x = 0 x = l
Fig. 3.9 Typical beam element.
bar element, but the d.o.f. permitted at each node are the transverse displacement v, and the
slope (rotational d.o.f.) dv/dx. We thus expect to derive the beam element equations as
k11 k12 k13 k14 v1 F1

k21 k22 k23 k24 R1 M1
= (3.86)
k31 k32 k33 k34 v2 F2

k41 k42 k43 k44 R 2 M 2

Based on the physical interpretation of the columns of the stiffness matrix above, we
observe that the columns of the stiffness matrix should correspond to the forces (moments)
required to cause the deformation states shown in Figure 3.10. For example, the first
deformation state requires that the beam act as a clamped-free beam (x = L being the clamped
end) with a force F and a moment M applied at the end x = 0 such that it undergoes unit
transverse deformation but zero rotation. From elementary mechanics of materials involving the
EulerBernoulli beam theory, we find the forces (moments) for causing these deformation
states as shown in Figure 3.10. Thus the beam element stiffness matrix derived using the
direct method is found to be
12 EI
A3 Symmetric

6 EI 4 EI
2
A A
[k ] =
12 EI 6 EI 12 EI
A3 (3.87)
A2 A3

6 EI 2 EI 6 EI 4 EI
2
A A A2 A
k11
k21
v1 = 1.0, q1 = 0
v2 = 0, q2 = 0
1.0 k41
k31
v1 = 0, q1 = 1.0
v2 = 0, q2 = 0
k42
1.0
k22
k32
k12 v1 = 0, q1 = 0
v2 = 1.0, q2 = 0
k33
k43
1.0
k23
v1 = 0, q1 = 0
v2 = 0, q2 = 1.0
k13
k44 1.0
k24
k14 k34
Fig. 3.10 Characteristic deformation shapes for a beam element.
We will derive the beam element equations from the PSTP and show that we indeed get
the same equations (see Section 4.5). Even though the direct physical meaning of the individual
columns of the stiffness matrix provides useful insight, its use in actually deriving the element
matrices is limited to only simple cases. When we discuss two- and three-dimensional finite
elements and higher order elements with internal nodes, etc., it is not possible to derive the
stiffness matrices based on this interpretation alone. We need a systematic procedure based on
Summary 87
either the Galerkin WR or the Variational method, to formulate these finite elements. We have
demonstrated that both differential and functional forms yield identical finite element equations.
We shall now modify our general procedure for finite element analysis given in Chapter 2
and recast it as follows:
General Procedure for Finite Element Analysis

Step 1
Using the governing differential equation, write down the weighted residual form.
Perform integration by parts sufficient number of times and develop the weak form
of the differential equation. Rewrite the weak form as a summation over n elements.
OR
Write the functional representation as a summation over n elements.
Step 2
Define the finite element, i.e. the geometry of the element, its nodes, and nodal d.o.f.
Derive the shape or interpolation functions.
Step 3
Compute the element level equations by substituting these shape functions in the
weak form or the functional representation.
Step 4
Discretise the given domain into a finite element mesh. For a given topology of finite
element mesh, build up the system of equations by assembling together element level
equations.
Substitute the prescribed boundary conditions and solve for the unknowns.
SUMMARY
We have discussed the formulation of the finite element equations starting from either the weak
form of the Galerkin weighted residual statement or the condition of stationarity for a functional
and demonstrated that they yield identical results. For long, it has been common practice to use
energy methods in structural mechanics and, thus, for most problems of structural mechanics,
it may be convenient to use the PSTP and formulate the necessary finite element equations. The
variational principle dealing with the strain energy and the work potential of external forces has
direct physical meaning for the structural analyst. However, given the functional for any
problem (such as the heat transfer example above), we will be able to develop the necessary
finite element equations. The functional for some problems, such as some problems of heat
transfer and fluid flow, may not have any direct physical meaning, and so it may be more
convenient to state the problem in terms of the governing differential equations. It must be
appreciated that the finite element equations can be derived systematically with equal ease,
starting from either form of representation of governing equationsdifferential or functional.
In the subsequent chapters of the text, we will use both these forms interchangeably for
developing the necessary finite elements.
PROBLEMS
3.1 Redo Example 3.1 with a tip load P.
[Hint: Use V = Pu(L)].
3.2 Redo Example 3.1 with a linearly varying load q = ax. Do we still get an exact
solution?
3.3 Redo Example 3.2, using v(x) c1(x)(L x).
3.4 Repeat Example 3.2, using v(x) c0 + c1x + c2x2 + c3x3 + c4x4.
(Note: Observe that this field does not automatically satisfy the essential boundary
conditions of the problem. Thus we need to set v(0) = 0 = v(L), yielding two
relations amongst ci. We then get the remaining three equations from the PSTP.)
3.5 Solve for the displacement field of a cantilever beam under uniformly distributed
load q0. Use v(x) c1[1 cos (p x/2L)].
Problems to Investigate
3.6 The functional, for a one-dimensional problem in structural mechanics, is given by
2
EI L d2y
I =
2
0
2 dx Py ( L)
dx
dy
y (0) = (0) = 0
dx
By performing the variation d I = 0, derive the governing differential equation of the

problem and the boundary conditions at x = L. What is the physical situation
represented by this functional?
3.7 For the functional given below, obtain the corresponding differential (Euler
Lagrange) equation and boundary conditions:
1 L dy 2
I = B C y 2 + ryx 2 dx y ( L)
2 0 dx
subject to y(0) = 0. Assume a, b to be known constants.
CHAPTER 4
One-dimensional Finite
Element Analysis
We have already discussed three different methods of deriving finite element equations, viz.,
direct method, weak form, and stationarity of a functional. The direct method was useful for
very simple structural elements only, and we illustrated the equivalence of other two methods
with the example of a one-dimensional bar element and heat transfer element. In this chapter
we will study the finite element analysis of one-dimensional problems in greater detail. We will
use the stationarity of a functional approach (i.e., the Principle of Stationary Total Potential) to
develop the finite element equations for structural mechanics problems. We will first recast the
earlier procedure in a generic form and then work out more example problems using the bar
element. Subsequently, we will discuss the beam and frame elements. For heat transfer
problems, we will use the weak form equation.
4.1 General Form of the Total Potential for 1-d

A one-dimensional structure (such as a bar or a beam) can, in general, be subjected to
distributed loading in the form of body forces (such as gravity), forces distributed over a
surface (such as a pressure) or concentrated forces lumped at certain points. The first two
types of forces for one-dimensional problems can be reduced to force per unit length, q. It may
also be subjected to initial strains and stresses (such as those due to pre-loading, and thermal
loading). Under the action of all these loads, the structure, when properly supported (so as to
prevent rigid body motion), undergoes deformation and stores internal strain energy. We will
now develop an expression for the total potential of such a structure assuming linear elastic
isotropic material obeying Hookes law. We also assume that all loads and other parameters are
time-invariant.
Stress sx = E(ex ex0) + sx0 (4.1)
where ex0 is the initial strain and sx0 is the initial stress. This stress (sx) does work in causing
an increment in strain (dex), which is further stored in the elastic body as an increment in strain
89
90 One-dimensional Finite Element Analysis
energy. Consider a cube of unit edge length in the body, i.e. unit volume of material. The
increment in strain energy in this unit volume, dU is given by the incremental work which
equals the product of force (F) and incremental displacement (da). Here, a stands for the
generic displacement e.g. axial displacement u for a bar. Thus
dU = [F][da] = [(sx)(1)(1)][(dex)(1)] = sx dex (4.2)
Substituting for stress from Eq. (4.1) and integrating with respect to strain yields the strain
energy in this unit volume of material. This, when integrated over the whole structure, yields
the total strain energy. Thus,
1 2
U = EF x A dx
2 F EF
x x0 A dx + F T
x x0 A dx (4.3)
The potential V of an external force in any generalised coordinate a is defined, by convention,

in such a way that the generalised force in that coordinate is given as Fa = (V/a). The
forces on a one-dimensional structure may be distributed (q per unit length) or concentrated
(P). Therefore, the potential of external forces is given by
V = ( B q dx + B P ) i i (4.4)
where Pi is the concentrated force acting at a point i , ai is its displacement and the summation
is taken over all the points where concentrated forces are present. The total potential of the
structure is obtained as
3p = U + V (4.5)
We will use this generic form in our further discussion while attempting to derive the finite
element equations in their generic form.
4.2 Generic Form of Finite Element Equations

Let the displacement a at any point P within the element be interpolated from the nodal values
of displacement {d }e using the shape functions N as follows:
a = [N]{d }e (4.6)
The interpretation of Eq. (4.6) is as follows: For a one-dimensional problem, we have one
possible displacement at each point (axial displacement u or transverse displacement v), and
{d }e is a column vector of nodal d.o.f. of the element and is of size equal to the product of
number of nodes per element (NNOEL) and the d.o.f. per node (NFREE).
For a bar element (Figure 4.1),
NNOEL = 2
NFREE = 1
u1
{E }e = (4.7)
u2
Generic Form of Finite Element Equations 91
x = 0 x = l
x (local coordinate frame)
11 22
Nodal
u1 u2 displacement
F1 F2 Nodal forces
Fig. 4.1 Linear bar element.
For a beam element

NNOEL = 2
NFREE = 2
v1

dv1 / dx
{E } =
e

v2 (4.8)

dv2 / dx
The shape functions [N] represent the interpolation formulae used to obtain the displace-
ment field within the element, and the size of the matrix [N] will be consistent with the size
of {d }e. For example, for the bar element, [N] is given by
x x
[N] = 1 (4.9)
A A
The nonzero strain at a point P within the element (i.e., ex) is given in terms of the
derivatives of the displacement field and is again represented in terms of nodal d.o.f. as
ex = [](a) = [][N]{d}e = [B]{d}e (4.10)
where [] is a differential operator and [B] = [][N], representing the derivatives of shape
1 1
functions. For a bar element, for example, [] is simply d/dx and [B] = .
A A
We can now rewrite the generic expression for the total potential as follows from
Eq. (4.3). For an element,
1
{E } {E }
T
eT eT (4.11)
Ue = {E }e [ B ]T [ B ]{E }e EA dx [ B ]T E F x 0 A dx + [ B ]T T x 0 A dx
2
Again from Eq. (4.4),
Ve = {E }
eT
[ N ]T q dx {E } eT
[ N ]Ti Pi (4.12)
Thus,
1
2 {E } {E }
eT eT
3ep = [B]T [ B]{E }e EA dx [ B]T E F x 0 A dx
(4.13)

T T T
+ {E }e [ B ]T T x 0 A dx {E }e [ N ]T q dx {E }e [ N ]Ti Pi
where [N]i is the matrix of shape function values at the location i.

Since {d }e is a vector of nodal d.o.f., and is not a function of the spatial coordinate x,
it can be taken outside the integration. Thus we get
([ B] [B]EA dx ) {E }
1
[ B] [ B] T
T T T
3ep = {E }e T e
{E }e T
EF x0 A dx + {E }e T
x 0 A dx
2
(4.14)
[N ]
T T
{E }e [ N ]T q dx {E }e T
i Pi
Defining the element stiffness matrix [k]e and the load vector { f }e as
[ k ]e = [ B] [ B]EA dx
T
(4.15)
{ f }e = [ B]
T
EF x 0 A dx [ B] T
T
x0 A dx + [N ]
T
q dx + [N ] T
i Pi (4.16)
we can rewrite Eq. (4.14) in the form
1 T T
3ep = {E }e [k ]e {E }e {E }e { f }e (4.17)
2
Potential Q ep being a scalar quantity, the total potential for a mesh of finite elements can be
obtained by simple addition. The total potential of the system is given as
NOELEM

1 T
3p = 3ep = {E } [ K ]{E } {E }T {F } (4.18)
1
2
wherein the global stiffness matrix of the structure [K] and the global load vector {F} are
obtained as
NOELEM
[K ] =
n =1
[k ]e (4.19)
NOELEM
{F } = 1
{ f }e (4.20)
NOELEM represents the number of elements in the mesh and {d } contains all the nodal d.o.f.
variables for the entire finite element mesh. The summations indicating assembly imply that the
individual element matrices have been appropriately placed in the global matrices following the
standard procedure of assembly.
The Linear Bar Finite Element 93
Using the PSTP, we set the total potential stationary with respect to small variations in
the nodal d.o.f., i.e.*,
3 p
=0 (4.21)
{E }T
Thus the system level equations are given by

[ K ]{E } = {F } (4.22)
We will now illustrate the details of the procedure through the familiar bar element and
then derive the element matrices for a beam element. We will follow this generic notation in
our derivation of two-dimensional elements also. Thus the reader is advised to become fully
conversant with this notation, which will be extensively used in the rest of the book.
4.3 The Linear Bar Finite Element

We recall that our linear bar element (Figure 4.1) is a one-dimensional line element with two
nodes. Each point P within the element is permitted to move only along the axis of the element
and the displacement of the entire cross-section is assumed to be same. Therefore, we have
Displacement field {a} = u (4.23)
u1
Nodal d.o.f. vector {d}e = (4.24)
u2
x x
Shape functions [N] = [N1 N2 ] = 1 (4.25)
A A
du
Strain {e} = F x = (4.26)
dx
Thus the differential operator [] is simply d/dx. Hence,
dN dN 2 1 1
[ B] = 1 = (4.27)
dx dx A A
Thus the element stiffness matrix is obtained as
A
[ k ]e = 0
[ B ]T [ B ] EA dx (4.28)
l 1/ A
=
0
[1/ A
1/ A
1/ A]EA dx (4.29)
AE 1 1
= (4.30)
A 1 1
*Though this equation is written in compact matrix notation, it simply states that the total potential is
stationary with respect to small variations in each and every nodal d.o.f.
which is the same as the result obtained earlier (see Eq. 3.78). In the examples that follow, we
will illustrate the use of the bar element in several situations.
Example 4.1. A bar subjected to self-weight. Consider a vertically hanging rod (Figure 4.2)
of length L, uniform cross-sectional area A, density r, and Youngs modulus E. We wish to
find the state of deformation and stress in the rod subjected to gravity load, using the bar
element just described. Through this example, we wish to bring out several insights into the
nature of finite element approximate solutions.
For a bar element, the stiffness matrix is given by Eq. (4.30). Thus,
F
g
1
Fig. 4.2 A vertically hanging rod (Example 4.1).
AE 1 1
[k ]e = (4.31)
l 1 1
For the present example, the body force (i.e., self-weight) expressed as q per unit length is
given by (rAg). In the absence of any other force, the element nodal force vector is given by
Eq. (4.16):
l 1 x/ l
{ f }e = [ N ]T q dx = ( S g ) ( A) dx (4.32)
0 x / l
S Alg /2
= (4.33)
S Alg /2
Since (rAlg) represents the weight of the element, the equivalent nodal force vector just
obtained is thus simply half the weight of the element distributed at either node. We will now
use one, two, four, eight and 16 finite elements and, in each case, compare our solution with
the exact solution. The exact solution is readily verified to be given by
Sg x2
Displacement u(x) = Lx (4.34)
E 2
Stress s (x) = rg(L x) (4.35)
We observe that the exact solution corresponds to linear variation of stress and quadratic
variation of displacement, whereas within each bar element we permit only linear displacement
field (Eq. 4.25) and, consequently, constant stress. Thus we expect our solution to be poor
when the number of elements is very small and should become more and more accurate as we
increase the number of elements.
One-element solution. We use a single bar element to model the entire length of the rod
as shown in Figure 4.3(a). We have
1 1 1
L
1 A =
4
1
L 2
A=
2
2
l = L
1 2 3
3
2
4
2 3 5
(a) One element (a) Two elements (a) Four elements
Fig. 4.3 Finite element mesh for bar under self-weight (Example 4.1).
AE 1 1 u1 F1
= (4.36)
l 1 1 u2 F2
Using l = L and substituting for the boundary condition u1 = 0 and for the known force
F2 = rAgL/2, we get
S AgL L S gL2
u2 = = (4.37)
2 AE 2E
Thus,
S gL
u(x) = x (4.38)
2E
S gL
Strain ex = (4.39)
2E
S gL
Stress sx = EF x = (4.40)
2
Two-element solution. Let us discretise the bar into two equal length elements (l = L/2)
as shown in Figure 4.3(b). For the first element, we have
AE 1 1 u1 f1
(1)
= (4.41)
L/2 1 1 u2 f 2(1)
For the second element,
AE 1 1 u2 f1
(2)
= (4.42)
L /2 1 1 u3 f 2(2)
Assembling the two individual element level equations, we obtain
1 1 0 u1 f1(1)

AE
1 1 + 1 1 u2 = f 2(1) + f1(2)
L /2 (4.43)
0 1 1 u3 f 2(2)
Substituting the boundary condition, u1= 0, and the known forces
S AgL /2 S AgL/2 S AgL/2

f 2(1) = , f1 =
(2)
, f 2(2) =
2 2 2
we get
AE 2 1 u2 S AgL/2
= (4.44)
L/2 1 1 u3 S AgL/4
Solving, we obtain
u2 S gL2 3/2
= (4.45)
u3 4 E 2
The element level strains and stresses are computed as
(1) 0
u1 1 1 3 S gL
F (1)
= [ B] = 3S gL2 =
x
L /2 L /2 4 E (4.46)
u2
8E
(2)
u1 1 1 3S gL /8E
2
1 S gL
F x(2) = [ B] = = (4.47)
u2 L/2 L /2 S gL2 /2 E 4 E

3S gL
T x(1) = EF (1)
x = (4.48)
4
S gL
T (2)
x = EF (2)
x = (4.49)
4
Four-element solution. We further subdivide each of the two elements used earlier by
half Figure 4.3(c), i.e., the total rod is discretised into four elements, each of length (L/4). For
each element, we have
AE 1 1 u1 F1
= (4.50)
L / 4 1 1 u2 F2
Assembling all the individual element matrices and substituting for the boundary condition and
the known forces, we obtain
1 + 1 1 0 0 u 2 1/2 + 1/2

AE 1 1 + 1 1 0 u3 1/2 + 1/2 S ALg
=
L /4 0 1 1 + 1 1 u4 1/2 + 1/2 4 (4.51)

0 0 1 1 u5 1/2
Solving, we get
7 S gL2 12 S gL2
u2 = , u3 = ,
16 2 E 16 2 E
(4.52)
15 S gL2 S gL2
u4 = , u5 =
16 2 E 2E
The corresponding stresses in individual elements are obtained as
T (1) = S gL, T (2) = S gL ,

7 5
32 32
T (3) = S gL , T (4) = S gL
3 1
(4.53)
32 32
On similar lines, we can obtain the solution with eight and 16 bar elements, each time
subdividing the existing elements into two. For a one metre long steel rod (E = 2 1011 N/m2,
r = 7,800 kg/m3) of cross-sectional area 5 cm2, Table 4.1 (Figures 4.4 and 4.5) compares the
displacements (stresses) obtained using the various finite element models against the exact
solution.
Table 4.1 Rod under Self-weightcomparison of deflections (deflections are in micrometers)
x 1 element 2 elements 4 elements 8 elements 16 elements Exact
0 0 0 0 0 0 0
0.0625 0.0231 0.0231
0.125 0.0448 0.0448 0.0448
0.1875 0.0649 0.0649
0.25 0.0836 0.0836 0.0836 0.0836
0.3125 0.101 0.101
0.375 0.116 0.116 0.116
0.4375 0.131 0.131
0.5 0.143 0.143 0.143 0.143 0.143
0.5625 0.155 0.155
0.625 0.164 0.164 0.164
0.6875 0.172 0.172
0.75 0.179 0.179 0.179 0.179
0.8125 0.184 0.184
0.875 0.188 0.188 0.188
0.9375 0.190 0.190
1 0.191 0.191 0.191 0.191 0.191 0.191
We observe some very significant pointsfor example, no matter how few the elements
we take are, our solution for the displacement at the nodes is exact! Even with just one element
(i.e., a single straight line variation of displacement within the entire rod), we get the exact
value for tip displacement, whereas the exact solution is clearly a quadratic variation of
displacement!* However, we must note here that the nodal solution is not exact for most finite
element solutions.
A point of immediate interest is With nodal displacements being exact, how could the
stresses be inaccurate?. The calculation of stress depends not just on the nodal values but on
the variation of displacement within the element. Since the element used permits only linear
*The interested reader is referred to P. Tong, Exact solution of certain problems by the finite element method,
AIAA Journal, 7, 1, 178180, 1969.
Exact solution
4 Element solution
2 Element solution
1 Element solution
x
Fig. 4.4 Rod under self-weight Comparison of stresses (1, 2 and 4 element solutions).
Exact solution
8 Element solution
16 Element solution
x
Fig. 4.5 Rod under self-weightComparison of stresses (with 8 and 16 elements).
variation of displacement and the exact solution is quadratic, we expect differences in the stress
result. To illustrate this, let us compute the value of displacement at an interior point as
predicted by our finite element solution. Let us choose to compute the displacement at
x = 0.33L, which is not a node in any of the finite element solutions we obtained. For this
purpose, we first determine on which element this point P (corresponding to x = 0.33L) lies,
and use the regular interpolation formula, i.e.,
x p x p ui
u p = 1 (4.54)
A A u j

where ui and uj are the nodal displacements of that element and xp in this equation refers
to the distance from the ith node of the element to the point P (recall that we used a local
coordinate system for each element with its local origin at the ith node, see Figure 4.1).
Table 4.2 compares the exact solution and the finite element solution obtained with different
number of elements. We observe that our finite element solution is in great error initially but
our solution improves with increasing number of elements. We have achieved excellent
convergence as we increase the number of elements.
Table 4.2 Convergence of Displacement at an Interior Point
No. of Normalised displacement at x = 0.33

elements (FE displacement /Exact solution)
1 0.598
2 0.898
4 0.975
8 0.993
16 0.998
Since this element uses linear interpolation functions for displacement, it is a constant
stress element. This is reflected in our finite element results for stresses as shown in
Figures 4.4 and 4.5. While the exact solution for stress is linear, our finite element prediction
approximates the same using step function. Our approximation of course improves with
increasing number of elements. Since stress is constant within each element, we observe that
at each node we have two values of stressone for either element. A significant result is
that the average value of stress at any node tallies with the exact solution! It is common
practice in plotting stress contours of a complex finite element mesh to perform stress
averaging at nodes; this leads to smooth contours and also gives a better estimate of stress at
a node. We also observe that the finite element values of stress at the centre of each element
tally with the exact solution. A significant generalisation of this result is that there exist certain
points within a finite element where the error in the estimation of stresses is the least. It is
common practice to evaluate stresses within an element at these points and then extrapolate to
the desired points.
The Quadratic Bar Element 101
A summary of our observations based on the above discussion is now given.

Depending on the variation of the field variable (displacement in this case) and the
shape functions used in the finite element, we may require many elements to model
the field accurately, especially at the interior points.
The field variable itself may be reasonably well captured, but the error in the
derivatives (e.g. stress here) could be considerable.
Stress averaging will be necessary to get a more realistic value of stress at a node,
jointly shared by many elements.
The bar element is capable of modelling only a linear variation of displacement, and hence
we require many such elements to model a physical system accurately. We can improve this
element by permitting quadratic variation of displacement within the element. We will now
discuss a typical higher order element, viz., quadratic bar element. We cannot achieve this (i.e.,
a complete quadratic polynomial) with just two nodes and so we have to add another node to
the element.
It must be observed that we could have used two nodes only but permitted an
incomplete quadratic polynomial shape function, e.g. u(x) = c0 + c2x2. For such an element,
ex = du/dx = 2c2x. Thus, strain and stress will always be zero at one end of each element
(at x = 0)!
4.4 The Quadratic Bar Element

4.4.1 Determination of Shape Functions
The three noded quadratic bar element is shown in Figure 4.6. We number the two end nodes
as 1 and 2 and the middle node is given number 3. The displacement at any point within the
x = 0 x = l/2 x = l
x
1 2 3
u1 u3 u2
Fig. 4.6 Quadratic bar element.
element is now interpolated from the three nodal displacements using the shape functions as
follows:
u(x) = N1(x)u1 + N2(x)u2 + N3(x)u3 (4.55)
The shape functions Ni vary quadratically within the element and we now discuss a
systematic procedure to derive such shape functions. Let u(x) be given by the complete
quadratic polynomial
u(x) = c0 + c1x + c2x2 (4.56)
We know that u(0) = u1, u(l) = u2 and u(l/2) = u3. Hence, from Eq. (4.56),
u1 = c0, u2 = c0 + c1l + c2l2
2
A A
u3 = c0 + c1 + c2 (4.57)
2 2
Solving for ci, we obtain
c0 = u1
c1 = (4u3 u2 3u1)/l (4.58)
c2 = (2u1 + 2u2 4u3)/l2
Thus we have
4u3 u 2 3u1 2u1 + 2u 2 4u3 2
u(x) = u1 + x + x (4.59)
A A2
Rearranging the terms, we get
3x 2 x2 x 2x2 4x 4x2
u ( x) = u1 1 + + u 2 + + u3 (4.60)
A A2 A A2 A A2
Comparing Eqs. (4.55) and (4.60), we obtain
3x 2 x2 x 2 x2 4x 4 x2
N1 = 1 + 2 , N2 = + , N3 = (4.61)
A A A A2 A A2
This is a general technique for determining the necessary shape functions Ni. We first
assume a general polynomial field (Eq. (4.56)). Substituting the nodal coordinates and requiring
that the expression must reduce to nodal d.o.f., we generate sufficient number of equations
(Eq. (4.57)) to solve for coefficients ci in terms of nodal d.o.f. We then substitute for ci
(Eq. (4.59)) and rearrange the terms to obtain the desired shape functions (Eq. (4.60)). Later
on we will discuss other strategies for effective generation of the desired shape function.
4.4.2 Element Matrices

Given the shape functions, we get the strain-displacement relation matrix [B] as
du d
Fx = = [ N ]{E }e = [ B ]{E }e (4.62)
dx dx
where
dN dN 2 dN3 2 3A 2 A 4
[B] = 1 = 2 2x 2 2 x 2 (2 x A) (4.63)
dx dx dx l 2 A 2 A
Hence we have the element stiffness matrix given by from Eq. (4.15) and Eq. (4.63):
[k ]e = [B] [B]EA dx
T
(4.64)
2 3A
2 2x
A 2
2 3A 2 A 4
2 2 x A 2 2 x
l
= 0 A
2
2 A 2
2
A
2x
2 A 2
(2 x A) EA dx (4.65)

4
2 (2 x A )
A
On evaluating all the integrals, we obtain the element stiffness matrix as
7 1 8
AE
[k ]e = 1 7 8
3A (4.66)
8 8 16
We now redo Example 4.1 with this quadratic element and compare the results obtained
using typical lower order (linear bar) and higher order (quadratic bar) finite elements.
Example 4.2. A vertically hanging rod (Example 4.1 Reworked). For the present example,
the body force due to gravity expressed as force q per unit length is given by (rAg). Thus,
the element nodal force vector (in the absence of any other force) is given by
3x 2x2
1 + 2
A A

A A
x 2x2
{ f }e =
0
[ N ]T q dx =
0 A
+ 2
A
S Ag dx (4.67)

4x 4x2
2
A A
1/6

= S Agl 1/6
(4.68)
4/6
Using just one element spanning the entire length of the rod, as shown in Figure 4.7, we
obtain the equations
1 3 g
Fig. 4.7 Rod under self-weight Quadratic element solution (Example 4.2).
AE 7 8 u2 1/6
= S Ag A (4.69)
3A 8 16 u3 4/6
where the boundary condition u1 = 0, and the known forces have already been substituted and
l = L.
Solving, we get
S gL2 , 3S gL2
u2 = u3 = (4.70)
2E 8E
Substituting these equations in Eq. (4.60), we obtain
x 2x2 S gA2 4x 4 x 2 3S g A 2
u ( x) = + 2 + 2
A A 2E A A 8E
S gA2 2x x2 Sg x2 (4.71)
= = Lx
2E A A2 E 2
Comparing with Eq. (4.34), we observe that our one-element solution (l = L) tallies with
the exact solution completely (i.e. not just at the nodes but interior points also). It is left as
an exercise for the reader to show that the stresses also match the exact solution.
We thus observe that we could get the exact solution with just one quadratic bar element
where as we needed several linear bar elements to get a reasonably accurate solution.
Conclusion
A few higher order elements are far superior to several lower order elements.
Example 4.3. Effect of node numbering on assembled stiffness matrix. We consider the rod
shown in Figure 4.8 to highlight an important aspect of typical finite element analysis. We use
six bar elements each of length l but in the two cases considered, the node numbering is
different. For simplicity, let us assume AE/l = 1. Thus the element stiffness matrix is simply
given as
1 1
[k ]e = (4.72)
1 1
For the two finite element meshes shown in Figure 4.8, we now perform the assembly
and obtain the global stiffness matrix. In either case, since there are seven nodes, we expect
1 2 3 4 5 6
1 2 3 4 5 6 7
(a) Mesh 1
1 2 3 4 5 6
1 7 2 6 3 5 4
(b) Mesh 2
Fig. 4.8 Bar element meshDifferent node numbering (Example 4.3).
to get a (7 7) global stiffness matrix for the entire structure. For the sake of convenience,
the individual element matrices; the local node numbers, and the global node numbers indicating
the appropriate locations for the coefficients of the matrices in the assembled matrix are all
indicated.
Mesh 1 Mesh 2
Element 1
Local Global
Global [1] [3] [1] [7]
Local [1] [2] [1] [2]
1 1 [1] [1] 1 1 [1] [1]
1 1 [2] [2] 1 1 [2] [7]
Element 2
Global [2] [3] [7] [2]
Local [1] [2] [1] [2]
1 1 [1] [2] 1 1 [1] [7]
1 1 [2] [3] 1 1 [2] [2]
Element 3
Global [3] [4] [2] [6]
Local [1] [2] [1] [2]
1 1 [1] [3] 1 1 [1] [2]
1 1 [2] [4] 1 1 [2] [6]
Element 4
Global [4] [5] [6] [3]
Local [1] [2] [1] [2]
1 1 [1] [4] 1 1 [1] [6]
1 1 [2] [5] 1 1 [2] [3]
Element 5
Global [5] [6] [3] [5]
Local [1] [2] [1] [2]
1 1 [1] [5] 1 1 [1] [3]
1 1 [2] [6] 1 1 [2] [5]
Element 6
Global [6] [7] [5] [4]
Local [1] [2] [1] [2]
1 1 [1] [6] 1 1 [1] [5]
1 1 [2] [7] 1 1 [2] [4] (4.73)
For the two cases of node numbering, the global assembled matrices obtained are shown
in Figures 4.9 and 4.10.
We observe the following with respect to the assembled stiffness matrix of the
structure:
1 1 0 0 0 0 0

1 2 1 0 0 0 0

0 1 2 1 0 0 0

0 0 1 2 1 0 0

0 0 0 1 2 1 0

0 0 0 0 1 2 1

0 0 0 0 0 1 1
Fig. 4.9 Assembled stiffness matrix [K] for Mesh 1
1 0 0 0 0 0 1

0 2 0 0 0 1 1

0 0 2 0 1 1 0

0 0 0 1 1 0 0

0 0 1 1 2 0 0

0 1 1 0 0 2 0

1 1 0 0 0 0 2
Fig. 4.10 Assembled stiffness matrix [K] for Mesh 2.
Many elements in the assembled stiffness matrix [K] are zero. Recall our interpretation
for the elements of a column of an element stiffness matrix [k]ethe ith column represents
the forces required to cause a deformation state such that ui = 1, and all other nodal d.o.f. are
zero. Thus a zero element implies that no force is required to be applied at that node. With
respect to the node numbering in case (1), for example, since elements are connected only at
their nodes, it is easy to see that no force need be applied at node 2 (for example) to cause
a unit deformation at node 6 with all other d.o.f. arrested.
The assembled matrix [K] in both the cases is symmetric. Thus we need only to store
half the matrix, i.e. either the upper or the lower triangle, saving valuable computer memory.
All the nonzero coefficients in case 1 (Figure 4.9) are clustered around the leading diagonal,
while they are all well spread out in case 2 (Figure 4.10). If we employ typical Gauss elimination
to solve for the deflections from [K]{d} = {F}, most of these zero coefficients in [K ] could
remain zero throughout the solution process. Thus there is no reason to even store them, and
we look for efficient schemes of storage for [K ] to optimally use computer memory. For
example, it will be sufficient to simply store the following as the stiffness matrix for Mesh 1
1 1

2 1

2 1

2 1 (4.74)

2 1

2 1

1 0
It is usually a simple computer program that relates these stiffness coefficients to their
appropriate locations in the full, global stiffness matrix.
The node numbering employed in case (1) leads to minimal size arrays for storage of [K ].
Node numbering, therefore, has a significant effect on the distribution of zero and nonzero
coefficients in the global matrix [K], and efficient node numbering schemes are an integral part
of any pre-processor used for modelling complex real-life finite element problems.
Example 4.4 A bar subjected to thermal stresses. Consider a uniform rod of length L, area
of cross-section A, Youngs modulus E, and the coefficient of thermal expansion a, heated
uniformly through a temperature rise of T C. Determine the thermal stresses in the bar for the
end conditions shown in Figure 4.11(a).
(a) Fixed-free bar (b) Fixed-fixed bar
a LT
(c) Conceptualization of Pushed back (d) Partial constraints

constraint forces by support
Fig. 4.11 Thermal stress analysis (Example 4.4).
This example problem is indicative of how problems of initial strain {e0} are solved. The
initial strain in this example is due to heating of the bar. It could as well be due to any other
reasons such as bolt pretensioning. While we work out this example on a bar element, the
general methodology remains identical even when we consider situations involving more
complex elements.
Before we analyse the finite element solution to the problem, we will briefly review the
concept of thermal stresses. When a rod is heated through a temperature rise T, the rod would
tend to expand by an amount (aLT). This corresponds to an initial strain e0 = (aT). If the
support restraints on the rod are such that this expansion is possible, then the rod freely
undergoes the thermal expansion (aLT), and thus the mechanical strain in the rod is given by
e = aLT/L = (aT). This is the case in Figure 4.11(a). As there are no forces restraining this
thermal expansion, there will be no internal stresses developed. This is reflected in the material
constitutive equation given by
s = E(e e0) = E(aT aT) = 0 (4.75)
However, as in Figure 4.11(b), if the rod is entirely prevented from undergoing the free
thermal expansion as dictated by the initial strain e0, then the mechanical strain in the rod is
e = 0. Thus the rod is subjected to a compressive stress as
s = E(e e0) = E(0 aT) = EaT (4.76)
This can be visualised as shown in Figure 4.11(c), i.e., the rod can be imagined to have
expanded through the support by the amount (a LT), and the support develops reaction forces
which push the rod back (i.e. compress) all the way to its original position. If, in general, the
support is a little distance away from the end of the rod as shown in Figure 4.11(d) so that
the rod is allowed partial free expansion, then internal stresses arise only corresponding to the
part of initial strain that was not actually realised. The structure depicted in Figure 4.11(d) is
for the purpose of illustration only. In real structures, when a part of the structure undergoes
thermal expansion, it will be against the elasticity of the other parts. Therefore, only a part of
the total thermal expansion will be permitted.
Thus our general method of modelling initial strain will be as follows: we first determine
the mechanical loads corresponding to the initial strain, i.e. the loads that would be required in
order to cause a mechanical strain equal to the initial strain. Under these loads, we then find
the actual deformation of the structure consistent with its support restraint conditions. From
these actual deformations, we find the actual mechanical strain. The difference between the
actual mechanical strain and the originally imposed initial strain leads to the internal stresses.
From our general expression for nodal force vector given in Eq. (4.16), the nodal loads
due to initial strain are given by
{ f }e = v
[ B]T EF x 0 A dx (4.77)
For the linear bar element,
1 1
[B] = (4.78)
l l
ex0 = a T (4.79)
Thus the nodal force vector corresponding to initial strain is given by
l 1/ l AEBT
{ f }e = B ETA dx = (4.80)
0 1/ l AEBT
We will use, for the purpose of illustration, two linear bar elements, each of length
(L/2) for modelling the entire rod. For element 1, we have, in the absence of any other load,
AEB T
(1)
AE 1 1 u1 f1(1)
= = (4.81)
L /2 1 1 u2 f 2(1) AEB T
For element 2, we have
AEB T
(2)
AE 1 1 u1 f1(2)
= = (4.82)
L /2 1 1 u 2 f 2(2) AEB T
It is readily seen that these forces, acting on a free bar element, will cause strains
corresponding to the initial strain e0 = aT. Assembling the element matrices, we get
1 1 0 u1 f1(1)

AE
1 1 + 1 1 u2 = f 2(1) + f1(2)
L/2 (4.83)
0 1 1 u3 f 2(2)
If the end x = L is free, then we have, with u1 = 0,
AE 2 1 u2 AEB T AEB T 0
= = (4.84)
L / 2 1 1 u3 AEB T AEB T
Solving, we get
u2 = LaT/2, u3 = LaT (4.85)
For element 1, the actual mechanical strain is
1 0
{F } = [ B]{E }e = 1 = BT
l l LBT/2 (4.86)
The internal thermal stress in element 1 is given by
{s} = E(e e0) = 0 (4.87)
Similarly, for element 2, we can show that the internal stresses will be zero. It is left as an
exercise for the reader to show that internal stresses equal to (aT) will be developed if the
end of the rod x = L is also clamped.
Summary: Finite Element Modelling of Initial Strains

Step 1: Compute the element nodal force vector corresponding to the initial strain {e0}.
Step 2: Solve for the deformations of the structure when subjected to the loads due to
initial strain along with any other loads on the structure, consistent with the
boundary conditions.
Step 3: From the deformations obtained in Step 2 above, evaluate the element level strains
{e} using {e} = [B]{d}e.
Step 4: Evaluate the internal stresses developed using the constitutive equation of the
material, viz., {s} = E({e} {eo}).
Example 4.5. A three-bar truss network. Let us consider the statically indeterminate
problem of three trusses shown in Figure 4.12. Unlike all the previous examples discussed so
far, in this case the members of the structure are not all in one line. We use this example to
Fig. 4.12 A three-bar truss (Example 4.5).

bring out the issue of transforming element level equations in local coordinates into a
common global reference frame before assembly. Since the structure is subjected to only a
concentrated tip load, we expect the deflection to vary linearly within each member, and hence
we expect that one linear bar element would suffice for each truss member. Figure 4.13(a)
1 2 4
Y
2
1 3
3
P
X
X
Fig. 4.13(a) Three-element model of indeterminate truss (Example 4.5).
shows a three-element model of the structure indicating the element numbers and the node
numbers. We also show a generic truss element in Figure 4.13(b), indicating its own local axis
and the nodal d.o.f. along its axis, which are not necessarily along the global axes.
x x
uj Fj
ujY Fj sin q
j j
ujX Fj cos q
i q i
Y Fi sin q
ui Fi
Fi cos q
0 X
Fig. 4.13(b) Bar element in local and global frames.
For each truss element in its own local coordinate frame, we have the force-deflection
relations, viz.
AE 1 1 ui Fi
= (4.88)
1 1 u j
l F j
In our particular example, the length of the central element is L while that of the inclined
rods is L/cos a. Thus the individual element equations will be essentially similar to Eq. (4.88)
above except for the appropriate value of element length. Considering the components of
deflection as shown in Figure 4.13(b), we can write
ui = uiX cos q + uiY sin q
(4.89)
uj = ujX cos q + ujY sin q
Thus a point can now move along the X, Y directions as given by uX, uY, but these are
not independentthe resultant deformation can only be along the axis of the element since
that is the only deformation permitted for a truss. The relationship between the nodal d.o.f. in
the local coordinate frame of the element and the generic d.o.f. in the global frame can be
written as
uiX

ui cos R sin R 0 0 uiY
= (4.90)
u j 0 0 cos R sin R u jX

u jY
Thus we have
{E }eA = [T ]{E }eg (4.91)
where the subscript l stands for local and the subscript g stands for global coordinate frame.
If we write the element level equations as
[k ]eA {E }Ae = {F }eA (4.92)
then
[k ]eA [T ]{E }eg = {F }Ae (4.93)
T
Premultiplying by [T] , we get
([T ]T [k ]e [T ]){E }eg
A = [T ]T {F }Ae (4.94)
The RHS can be expanded as
cos R 0 Fi cos R

sin R 0 Fi Fi sin R

[T ]T {F }Ae = = (4.95)
0 cos R F j F j cos R

0 sin R F j sin R
From Figure 4.13(b), we see that these are generic forces at the nodes in the global
directions. Thus we have
[T ]T {F }A = {F }g
e e
(4.96)
If we define the element stiffness matrix in global reference frame as
[k ]eg = [T ]T [ k ]Ae [T ] (4.97)
then Eq. (4.94) can be rewritten as
[ k ]eg [E ]eg = {F }g
e
(4.98)
Figure 4.14 contrasts the element level equations in local and global coordinates for a
generic element. Equation (4.97), though derived in the context of this specific example, is a
generic equation, and can be used to transform any finite element stiffness matrix from local
to global coordinate frame.
We can derive Eq. (4.97) in another way, based on the equivalence of work done. We
observe that the nodal forces acting on the nodal deflection must perform the same amount of
work whether we consider them in their own local coordinate frame or a global coordinate
system. Hence the corresponding strain energy in the structure must also be the same and be
independent of the coordinate reference frame. Thus we have
Ulocal = Uglobal (4.99)
1
({E }eA ) [k ]eA {E }eA =
1
( ) [k ] {E }
T T
{E }eg e
g
e
g (4.100)
2 2
Fj
uj
Fi
AE 1 1 ui
=
L 1 1 u j F j

Fi i
ui FjY ujY
ujX
j
FjX

cos R
2
Symmetric uiX FiX

F
AE sin R cosR sin R iY
2 u
FiY
=
uiY iY

L u jX FjX
cos R sin R cosR cos2 R
2
uiX
u jY FjY
i sin R cosR sin R
2
sin R cos R sin R
2

FiX
Fig. 4.14 Bar element equations in local and global frames.
Using the transformation

{E }eA = [T ]{E }eg (4.101)
we have
1
( ) ([T ] [k ] [T ]){E } 1
( ) [k ] {E }
T T
{E }eg T e
A
e
g = {E }eg e
g
e
g (4.102)
2 2
Therefore,
{k}eg = [T ]T [ k ]Ae [T ] (4.103)
which is the same as the result obtained earlier.
We will now use this result to transform the element stiffness matrices in our example.
Element 1. The element stiffness matrix in its own coordinate frame (xy) is given by
AE 1 1
[k ]le = (4.104)
L / cos B 1 1
In the global reference frame (XY), we have, with q = a,

Global (1X) (1Y) (3X) (3Y)
Local (iX) (iY) (jX) ( jY) Local Global
cos2 B sin B cos B cos2 B sin B cos B (iX ) (1X )

AE cos B sin 2 B sin B cos B sin 2 B (iY ) (1Y )
[k ](1)
g
(4.105)
L cos2 B sin B cos B ( jX ) (3 X )

Symmetric sin B ( jY ) (3Y )
2

Element 2. The local frame (xy) and global frame (XY) are identical and hence there is
no transformation involved. We have
(2) (3)
(i) ( j)
AE 1 1 (i ) (2)
[k ](2)
g = (4.106)
L 1 1 ( j ) (3)
Equation (4.106), when rewritten as a (4 4) matrix, gets modified as

Global (2X) (2Y) (3X) (3Y)
Local (iX) (iY) ( jX) ( jY) Local Global
1 0 1 0 (iX ) (2 X )

AE 0 0 0 0 (iY ) (2Y )
[k ](2)
g =
L 1 0 1 0 ( jX ) (3 X ) (4.107)

0 0 0 0 ( jY ) (3Y )
Element 3. The transformation matrix would be [T(a)], i.e.,
cos B sin B 0 0
[T ] = (4.108)
0 0 cos B sin B
Therefore,
(4X) (4Y) (3X) (3Y)
(iX) (iY) (jX) (jY)
cos2 B sin B cosB cos2 B sin B cos B (iX ) (4 X )

AE sin 2 B sin B cos B sin 2 B (iY ) (4Y )
[k ](3) = (4.109)
L / cos B
g
cos2 B sin B cos B ( jX ) (3X )

Symmetric sin B ( jY ) (3Y )
2

Having obtained the element matrices in a common global frame XY, we are ready to
assemble the equations at the structure level. For the convenience of assembly, we have marked
the element level and structure level locations of the matrix coefficients. Since we have four
nodes and uX and uY at each node, we have eight variables, and our global stiffness matrix is
of size (8 8). The assembled equations are given in Figure 4.15.
cos3 B sin B cos 2 B 0 0 cos3 B sin B cos 2 B 0 0 u1 X 0

sin 2 B cos B 0 0 sin B cos 2
B sin 2
B cos B 0 0 u1Y 0

1 0 1 0 0 0 u 2 X 0

0 0 0 0 0 u 0
AE 2Y
=
L u P
1 + 2 cos 3
B 0 cos3 B sin B cos B 3 X
2

2 sin 2 B cos B sin B cos 2 B sin 2 B cos B u3Y 0

cos3 B sin B cos 2 B u4 X 0

Symmetric sin 2 B cos B u 4Y 0
Fig. 4.15 Assembled equations (Example 4.5).
By virtue of the boundary conditions due to supports at nodes 1, 2 and 3, we have

u1X = 0 = u1Y, u2X = 0 = u2Y, u4X = 0 = u4Y (4.110)
Thus from the 5th and 6th equations of Figure 4.15,
AE 1 + 2 cos B u3 X P
3
0
= (4.111)
L
0 2 sin 2 B cos B u3Y 0
Solving, we get
PL 1 u3Y = 0 (4.112)
u3 X = ,
AE 1 + 2 cos3 B
which tallies with the well-known result.
Beam Element 117
Having obtained the deflection of the structure, we can use the first six equations of
Figure 4.15 to find the support reactions at nodes 13, which ensure that the displacements
at these nodes are zero. We now summarise the general procedure for assembling element level
equations.
Summary: General Method of Assembly

Step 1: Identify the local (or element) level and global (on structure) level coordinate frames
of reference.
Step 2: Obtain the element matrices in the local reference frame.
Step 3: Obtain the coordinate transformation matrix [T] between the local and global frames
of reference.
Step 4: Transform the element matrices into the common global reference frame.
Step 5: Identify the global locations of the individual coefficients of the element matrices
based on local and global node numbers.
Step 6: Assemble the element matrices by placing the coefficients of the element matrices
in their appropriate places as identified in Step 5.
4.5 Beam Element

4.5.1 Selection of Nodal d.o.f.
A truss or rod undergoes only axial deformation and we assumed that the entire cross-section
undergoes the same displacement u. Thus the bar element was just a line element with two
nodes and each node having u as the nodal d.o.f. A beam, on the other hand, undergoes
transverse deflection denoted by v. Figure 4.16 shows a typical beam section undergoing
transverse deflection. We assume that the cross-section is doubly symmetric and the bending
Fig. 4.16 Typical beam section and its deformation.

takes place in a plane of symmetry. According to the classical theory of beam bending
(EulerBernoulli theory), the entire cross-section has the same transverse deflection v as the
neutral axis; sections originally perpendicular to the neutral axis remain so after bending. Hence
the deflections are small and we assume that the rotation of each section is the same as the
slope of the deflection curve at that point, i.e., (dv/dx). Therefore, the axial deformation of any
point P away from the neutral axis (due to bending alone) is given by uP = (z)(dv/dx). Thus,
if v(x) is determined, then the entire state of deformation of the body is completely determined.
Our beam element is therefore also a simple line element, representing the neutral axis of
the beam. However, in order to ensure the continuity of deformation at any point (i.e. those
on the neutral axis as well as away from it), we have to ensure that v and (dv/dx) are
continuous. We achieve this by taking two nodal d.o.f., viz., v and q (= dv/dx). Thus when
two beam elements are joined together, they share a common node and so have the same
transverse deflection and slope. If we did not have the slope d.o.f., then we would only have
been able to ensure that transverse deflection is continuous across elements, but not guarantee
that the slope also would be continuous. Thus the points away from the neutral axis could
actually separate (thereby creating a void) or overlap (creating a kink). We recall that for this
problem (fourth order differential equation in v), the essential boundary conditions consist of
both v and dv/dx. Thus, taking these nodal d.o.f. permits us to readily assign prescribed
essential boundary conditions (e.g. fixed end v = 0 = dv/dx). Moreover, a prescribed value of
moment load can readily be taken into account with the rotational d.o.f q. The EulerBernoulli
beam element is shown in Figure 4.17.
vi vj
qi qj
x = L
x = 0
Fig. 4.17 The EulerBernoulli beam element.
4.5.2 Determination of Shape Functions

The displacement field v(x) assumed for the beam element should be such that it takes on the
values of deflection and the slope at either end as given by the nodal values vi, qi, vj, qj. Let
v(x) be given by
v(x) = c0 + c1x + c2x2 + c3x3 (4.113)
Differentiating with respect to x, we have
dv/dx = c1 + 2c2x + 3c3x2 (4.114)
At x = 0 and L, we have
vi = c0, qi = c1, vj = c0 + c1L + c2L2 + c3L3, qj = c1 + 2c2L + 3c3L2 (4.115)
We can solve for the coefficients c0, c1, c2, c3 in terms of the nodal d.o.f. vi, qi, vj, qj
and substitute in Eq. (4.113) to get
Beam Element 119
v(x) = N1vi + N2qi + N3vj + N4qj (4.116)

where the shape functions N are given by
N1 = 1 3x2/L2 + 2x3/L3, N2 = x 2x2/L + x3/L2
(4.117)
N3 = 3x2/L2 2x3/L3, N4 = x2/L + x3/L2
(Note: The reader is advised to carry out the steps to gain familiarity with this method of
deriving the shape functions.)
Rewriting Eq. (4.11), we get
vi

R i

v(x) = [N]{d}e = [N1 N2 N3 N4] (4.118)
v j

R j
4.5.3 Element Matrices

The strain at any point in the beam is given by
d dv d 2v
{F } = F x = du = z = z 2
dx dx dx dx (4.119)
From Eq. (4.118), we have
d 2 N1 d 2 N 2 d 2 N3 d 2 N 4
Fx = z 2
{E }e
dx 2
dx 2
dx 2
dx
6 12 x 4 6 x 6 12 x 2 6 x (4.120)
+ 2 {E }
e
= z 2 + 3 + 2 2 3
L L L L L L L L
Thus the straindisplacement relation matrix [B] is given by
12 x 6 6x 4 6 12 x 6 x 2
[B] = z 3 2 2 3 2
(4.121)
L L L L L2
L L L
We can obtain the element stiffness matrix as
[k ]e =
v
[ B]T [ B]E dv
12 x 6
L3
L2
6x 4
2
l L L 12 x 6 6x 4 6 12 x 6 x 2
=
A 0
( z )
6
( E )( z ) 3 2 2
12 x L L L

L L2
3 2
L L
(dA) dx
L
L2 3
L

6x 2

L2 L
(4.122)
Noting that A
z 2 dA = I , the second moment of area of the cross-section and carrying
out the integration with respect to x, we get
12 6L 12 6L

EI 6L 4 L2 6 L 2 L2
[k ]e = 3 (4.123)
L 12 6 L 12 6 L

6 L 2 L2 6 L 4 L2
The element nodal force vectors are now obtained for specific cases of loading.
Gravity loading
Gravity loading is a typical body force and is given by (rg) per unit volume or (rAg) per unit
length, where r is the mass density of the material. The equivalent nodal force vector for the
distributed body force can be obtained as
A/2

L2
/12
[ N ]T ( S g ) A dx = (S Ag )
L
v
{ f }e = [ N ]T ( S g ) dv =
0 A/2
(4.124)

L2 /12
Any other distributed load on a beam specified in the form of q per unit length
can be addressed in a similar manner. For a specified q(x), one can perform the integration
[N]Tq(x) dx and obtain { f }e. Thus, for a uniformly distributed force q0, we can write the
equivalent nodal force vector as
Beam Element 121
q0 L/2

L q0 L2 /12
{ f }e =
0 [ N ]T q0 dx =
q0 L/2
(4.125)

q0 L2 /12
We observe that our equivalent nodal forces include nodal moments! We term this as
equivalent to the distributed force in the sense that these nodal forces acting through the nodal
displacements do the same amount of work as the distributed force.* Hence such a force
vector is termed consistent nodal force vector. We could have simply replaced the entire
distributed force on the element (q0L) by lumping half each at the two nodes. Such a nodal
force vector is termed lumped nodal force vector and is given by
q0 L/2

0
{ f }e = (4.126)
q0 L/2

0
Since the lumped force vector does not guarantee work equivalence, in general, we expect
the results to be inferior. Moreover, for more complex two- and three-dimensional elements
with internal nodes, an appropriate scheme of lumping is not entirely intuitive. Thus, normally,
we use only the consistent nodal force vector. The positive sign convention for the forces and
moments is shown in Figure 4.18, where the consistent and lumped nodal loads are illustrated
for two typical external load cases.
Fi Fj
Mi Mj
(a) Positive forces and moments
Fig. 4.18 Beam element forces. (cont.)
*Recall that work done = (force) (displacement). Thus work done by the distributed force
A A
(
(q dx){E } eT
[ N ] q dx = {E}
T
eT
q dx) v = [ N ]T = {E }e T
{ f }consistent
0 0
elemental force
qL qL
2 2
q
L qL2 qL2
12 12
P P
2 2
P
L/2 PL PL
8 8
(b) Consistent nodal forces
Fig. 4.18 Beam element forces.
Example 4.6. Cantilever beam under transverse load. Consider the cantilever beam loaded
as shown in Figure 4.19. Using one-beam element, we get
12 6L 12 6 L v1 F1

EI 6 L 4 L2 6 L 2 L2 R1 M 1
=
L3 12 6L 12 6 L v2 F2 (4.127)

6 L 2 L2 6 L 4 L2 R 2 M 2
P
q0
Fig. 4.19 Cantilever beam (Example 4.6).
At the clamped end, the boundary condition stipulates that v = q = 0. At x = L, taking

into account concentrated force/moment as well as nodal forces equivalent to the distributed
forces, we have
q0 L
F2 = P (4.128)
2
+ q0 L2
M2 = M (4.129)
12
Beam Element 123
From the last two rows of Eq. (4.127),
q0 L
P
EI 12 6 L 2
v 2
= (4.130)
3
L 6 L 4 L2 R 2 + q0 L
2

12 M
Solving, we obtain the following tip deflection and slope:
3q0 L3
v2 L 4 PL2 6 ML
= 2 (4.131)
R
2 12 EI
2q0 L 6 PL 12M
2
Thus the displacement field v(x) over the element is given by

v(x) = [N]{d}e = (N1)(0) + (N2)(0) + (N3)(v2) + (N4)(q2)
3x 2 2 x3 3q0 L3 L
= 2 3 4PL2 6ML
12EI
L L 2
x3 x2 L
+ 2
L
(
2q0 L 6PL 12M
2

12EI
)
L
P Mx2 q0
= ( x3 3 x 2 L ) + (2 x3 L 5x2 L2 ) (4.132)
6 EI 2EI 24EI
From standard text books on the Mechanics of Materials, we know that
P Mx2 q0
v( x ) = ( x 3 3x 2 L ) + ( x4 + 4 x3 L 6 x2 L2 ) (4.133)
6EI 2EI 24EI
The beam element admits cubic variation of transverse deflection and thus we can get the
exact solution for the concentrated load P and the moment M. However, the uniformly
distributed force q0 causes a quadratic curve and our one-element solution is in error. We
expect that our solution will improve with more number of elements.
Example 4.7. A continuous beam. Figure 4.20 shows a typical continuous beam. We wish
to obtain the deflection of the beam using the beam element just described. For simplicity we
assume EI = 1.
We will take a five-element mesh as shown in the figure. Subdividing the span AB into
two elements with a node at the load point has the advantage that we can very easily specify
the nodal forces. On the other hand, if we had not placed a node at the load point, then it would
have been necessary to find equivalent nodal forces. Since we have just now seen in the
P = 1 kN q = 12 kN/m
A B C
D
1m 1m 1m 2m
(a) Continuous beam
1 kN 6 kN 12 kN 6 kN
11 22 3
3 33 44 44 5 55 6
1
22 1 kN-m 1 kN-m
(b) Finite element model

Fig. 4.20 Typical continuous beam (Example 4.7).
previous example that one beam element is inadequate for a uniformly distributed force, we try
with two elements in the span CD. Our meshing has also ensured that all elements are of
uniform size, for easy hand calculation. Following the standard procedure, we obtain the global
stiffness matrix and force vector as
[ K ]1212 {E }121 = {F }121 (4.134)
12 6 12 6 0 0 0 0 0 0 0 0
4 6 2 0 0 0 0 0 0 0 0

12 + 12 6 + 6 12 6 0 0 0 0 0 0

4+4 6 2 0 0 0 0 0 0
12 + 12 6 + 6 12 6 0 0 0 0

4+4 6 2 0 0 0 0
[K ] = (4.135)
12 + 12 6 + 6 12 6 0 0

4+4 6 2 0 0

12 + 12 6 + 6 12 6
4 + 4 6 2

12 6
Symmetric 4

{F} ^R1 M1 1000 0 R3 0 (R4 600) 1000 12000 0 ( R6 6000) 1000`

T
(4.136)
Frame Element 125
Since there are six nodes and two d.o.f. per node, the global stiffness matrix [K] is of
size (12 12) and {F} is a column vector of size (12). The boundary conditions stipulate that
the slope and deflection be zero at node 1, and the vertical deflection be zero at nodes 3, 4
and 6. Thus the reduced set of equations involving unknown nodal d.o.f. is obtained in matrix
form as
24 0 6 0 0 0 0 v2 1000

8 2 0 0 0 0 R2 0

8 2 0 0 0 R3 0

8 6 2 0 R4 = 1000 (4.137)

24 0 6 v5 12, 000

8 2 R5 0

Symmetric
4 R6 1000
Solving, we get
v2 = 0.1570, v5 = 1.4720
q2 = 0.1153, q5 = 0.3427 (4.138)
q3 = 0.4612, q6 = 2.6293
q4 = 1.2586
Using these values in the system equations (Eqs. (4.134)(4.136)), we can obtain the
support reaction forces. Following standard mechanics of materials approaches, we can obtain
the exact solution as
v2 exact
= 0.1570
(4.139)
v5 exact
= 1.4720
From the element stiffness matrices and the nodal displacements, we can obtain the
element level forces as
{ f }e = [ K ]e {E }e (4.140)
These are the forces on the element at its nodes when each element is viewed as a free body.
The element level forces are shown in Figure 4.21. When the individual elements are put
together to build up the entire beam, the forces are algebraically summed up as shown in the
figure. At each node, the internal reaction forces annul each other and the resultant forces are
the externally applied forces. If a node is supported, the resultant force is the support reaction.
4.6 Frame Element

The bar element can permit only axial deformation and the beam element can permit only
transverse deflection. By combining the bar and the beam elements, we obtain the frame
1.1918
1.1918 1.1918
1.1918
-1.1918 0.1918
0.1918 0.1918
0.1918
-0.1918
Element
Element 1 Element
Element 2
0.7112
0.7112 0.4806
0.4806
0.4806
0.4806 0.6724
0.6724
4.7845
4.7845 4.7845
4.7875 8.056
8.056 8.056
8.056
Element
Element 3 Element
Element 4
0.6724
0.6724 4.1121
4.1121 3.1121 4.944
4.944
3.944
3.944 3.944
3.944
Element
Element 5
4.944
4.944 1.0
1.0
1.0
1.0 4.9763
4.5927 12
12
0.7112
0.7112 1.0
1.0 1.0
1.0
1.1918
1.1918 12.8405
12.8405 3.944
3.944
Fig. 4.21 Element and system level forces (Example 4.7).
element which enables us to model typical problems of framed structures which in general
involve both types of deformation. A typical planar frame element is shown in Figure 4.22, and
the corresponding element equations are
[k]e{d}e = {f}e (4.141)
vi vj
ui uj
qi i qj j
x
x= l
x = 0
Fig. 4.22 Plane frame element.
Frame Element 127
where
{ui }
T
{E }e = vi R i u j v j R j (4.142)
and the element stiffness matrix is
AE / L 0 0 AE / L 0 0

0 12 EI / L3 6 EI / L2 0 12 EI / L3 6 EI / L2

0 6 EI / L2 4 EI / L 0 6 EI / L2 2 EI / L
[ k ]e =
AE / L (4.143)
0 0 AE / L 0 0

0 12 EI / L3 6 EI / L2 0 12 EI / L3 6 EI / L2

0 6 EI / L2 6 EI / L2 4 EI / L
2 EI / L 0
Example 4.8. A frame structure. Consider the simple framed structure made of steel shown
in Figure 4.23(a). We wish to obtain the deflections using the frame element just described.
20 mm dia.
Steel rod
4m
Steel I-beam
A = 2000 mm2
I = 106 mm4
3m
10 kN
Fig. 4.23(a) Structure details (Example 4.8).
Through this example, we will illustrate the procedure for transformation of frame element
matrices from local to global reference frame. We will also study how to model pin joints in
a structure. We will use two elements as shown in Figure 4.23(b). Using the expression for
3
X
q
y2 x2
x2 20 mm dia.
Y Steel rod
Steel I-beam
A = 2000 mm2
I = 106 mm4
1 x1 2 4
1
y1
10 kN
Fig. 4.23(b) Finite element model (Example 4.8).
the element matrix given above, we get the element stiffness matrices as described now.
Element 1. L = 3 m, A = 2,000 mm2, E = 200,000 N/mm2, I = 106 mm4. The element

stiffness matrix in its local coordinate frame is given in the form
0.133 106

0 0.889 102 Symmetric

0 0.133 106 0.267 109
[k ](1) =
0.133 106 0 0 0.133 106

0 0.889 102 0.133 106 0 0.889 102
9
0 0.133 106 0.133 106 0 0.133 106 0.267 10
(4.144)
Frame Element 129
The local coordinate frame of element 1 is identical to the global frame and hence no
transformation is required.
Element 2. L = 5 m, A = 314.16 mm2, E = 200,000 N/mm2, I = 7,854 mm4. The element

stiffness matrix in its local coordinate frame is obtained in the form
0.126 105 Symmetric

0 0.151

0 0.377 103 0.126 107
(2 )
[k ]A =
0.126 105 0 0 0.126 105

0 0.151 0.377 103 0 0.151
7
0 0.377 103 0.628 106 0 0.377 103 0.126 10
(4.145)
The local coordinate frame is inclined with respect to global reference frame and hence
we need to transform the element stiffness matrix into global frame. For any orientation q, the
transformation matrix relating local and global d.o.f. can be readily shown to be given by
cos R sin R 0 0 0 0

sin R cos R 0 0 0 0

0 0 1 0 0 0
[T ] = (4.146)
0 0 0 cos R sin R 0

0 0 0 sin R cos R 0

0 1
0 0 0 0
Thus the element stiffness matrix in the global frame is written in the form
[k ](2)
g = [T ]T [ k ]l(2) [T ]
0.452 104

0.603 10 4
0.804 10 4
Symmetric

0.302 103 0.226 103 0.126 107
=
0.452 104 0.603 104 0.302 103 0.452 104

0.603 104 0.804 104 0.226 103 0.603 104 0.804 104

0.302 103 0.226 103 0.628 106 0.302 103 0.226 103 0.126 107
(4.147)
Assembly
We observe that, at the pin joint between elements 1 and 2, we use two coincident nodes. These
two nodes undergo the same translation but can have relative rotation because of the pin joint.
This is reflected in assembly as the translations having the same equation numbers (rows in
the matrix) and the rotational d.o.f. having different equation numbers (rows in the matrix). The
boundary conditions require that at node 1, all the d.o.f. be arrested and at node 3, the
translations be arrested. Thus the reduced system of equations is given as follows:
u2 , u4 v2 , v4 R2 R3 R4
u2 0
0.138 106
v2 10, 000
0.603 10 0.813 10
4 4
Symmetric
R 2 = 0 (4.148)
0 0.133 106 0.267 109
R 3 0
0.302 103 0.226 103 0 0.126 107
R 4 0
0.302 103 0.226 103 7
0.628 10 0.126 10
6
0
Solving, we get
u2 = 0.12 mm, v2 = 2.564 mm
q2 = 0.0013 rad, q3 = q4 = 3.26 104 rad (4.149)
These results can be easily verified based on classical mechanics. We observe that the
rotations at node 3 and 4 are the same for element 2, in view of the pin joints and it remains
a two-force member as expected.
Example 4.9. Simply supported beam under combined axial and transverse loads. Consider
the simply supported beam shown in Figure 4.24 subjected to an axial force Pa and a transverse
force Pt. We wish to use the frame element just developed to model this problem. We will use
Pt
L
2
Pa
(a) Axial and transverse loads on a beam

Pt
11 11 22 3
2 Pa
(b) Finite element model

Fig. 4.24 Beam under combined loads.
Frame Element 131
just two elements. For simplicity, let EI = 1, L = 2, AE = 1. Incorporating the boundary

conditions at nodes 1 and 3, we get the following equations for the unknown nodal d.o.f.:
4 0 6 2 0 0 R1 0

1 0 0 1 0 u2 0

24 0 0 6 v2 Pt
= (4.150)
8 0 2 R 2 0

1 0 u3 Pa

Symmetric 4 R3 0

We observe that the solution of these equations is the same as the solution of the following
equations separately:
4 6 2 0 R1 0

24 0 6 v2 Pt
=
8 2 R 2 0 (4.151)

Symmetric 4 R 3 0
1 1 u2 0
= (4.152)
Symmetric 1 u3 Pa
We observe that our solution is the same as the one we would have obtained if the loads
were individually applied. Thus there is no effect of the axial force on the transverse deflection
in our finite element solution whereas the exact solution for this problem is given as

L P L3 tan C C
Vmax = V = t
2 1 3 (4.153)
48 EI C
3
where
Pa L2
C2 = (4.154)
4 EI
This is to be expected since in Eq. (4.144) there is no coupling between axial and
transverse deflections (note the zeroes in the stiffness matrix [k]e). More fundamentally, we
have simply placed together the individual bar and beam elements and not really formulated a
frame element starting from first principles. Therefore, we should have started with an
appropriate expression for strain energy with the nonlinear coupling explicitly taken into
account. Thus, even though our frame element has the nodal d.o.f. u, v and q, it cannot model
phenomena such as buckling or stiffening due to centrifugal forces on a rotating beam (such
as a fan blade).
Conclusion
The finite element solution cannot capture any more physics of the problem than has been
taken into account in the functional or the weak form.
We have thus far studied various aspects of one-dimensional finite element analysis.
Though we used problems from the field of structural mechanics for the purpose of illustration,
many of the key ideas are equally applicable to other fields too. We will now study some heat
transfer problems. In Section 2.7, we studied the one dimensional heat conduction problem. We
will extend it to include convection also. We will use the weak form approach, which is perhaps
more pertinent for such problems.
4.7 One-dimensional Heat Transfer

The governing differential equation for the steady state one-dimensional conduction heat
transfer with convective heat loss from lateral surfaces is given by
d 2T P
k 2
+ q = h (T T ) (4.155)
dx Ac
where
k = coefficient of thermal conductivity of the material,
T = temperature,
q = internal heat source per unit volume,
P = perimeter,
Ac = the cross-sectional area,
h = convective heat transfer coefficient, and
T = ambient temperature.
The weighted residual statement can be written as
d 2T L P
0
W k 2 + q
dx

h(T T ) dx = 0
Ac
(4.156)
By performing integration by parts, the weak form of the differential equation can be
obtained as
L
dT L dW dT L P
L
Wk dx
0 0
k
dx dx
dx + 0
Wq dx
0
W h(T T ) dx = 0 (4.157)
Ac
i.e.
L
L dW dT L P L L P dT

0
k
dx dx
dx + 0
W hT dx =
Ac
0
Wq dx + 0
W
Ac
h(T ) dx + Wk
dx 0
(4.158)
One-dimensional Heat Transfer 133
The weak form, for a typical mesh of n finite elements, can be written as
n
A A n A dT
A
W q
dW dT P P
k =1
0
k
dx dx
dx + 0
W
Ac
hT dx =
k =1
0
+
Ac
hT dx + Wk

dx 0 (4.159)
The attention of the reader is drawn to the similarity of the weak form in the above
equation and that in Eq. (2.125) for the bar problem. We now propose to develop the one-
dimensional heat transfer element following the procedure we used while developing the
equations for the bar element. Figure 4.25 shows a typical heat transfer element, with its nodes
T1 T2
Node 1 Node 2
x
x=l
x = 0
Fig. 4.25 Typical heat transfer element.
and nodal d.o.f. Following the similarity of the bar and heat transfer element, we simply use
the same shape functions for interpolating the temperature within the element from the nodal
temperature.
x x
T ( x) = 1 T1 + T2 (4.160)
A A
dT T T1
= 2 (4.161)
dx A
x dW1 1
W1 = 1 , =
A dx A
(4.162)
x dW2 1
W2 = , =
A dx A
We now compute the element level contributions to Eq. (4.159) above.
LHS first term (Eq. (4.159)). Similar to the LHS of bar element (Eq. (2.125)), k
replaces AE and T replaces u. Thus we have
k 1 1 T1

A 1 1 T2
LHS second term (Eq. (4.159)). With W1,
A x P x x
1
0

A Ac
h 1 T1 + T2 dx
A A
Ph x
2
A x x
A
=
Ac

0
1 dx T1 +
A
0A
1 T2 dx
A
Phl
=
6 Ac
[2 T1 + T2 ]
With W2,
x P
A x x Phl

h 1 T1 + T2 dx = [T1 + 2T2 ]
A Ac
0 A A 6 Ac
We can now write the second term on the LHS in matrix notation as
Phl 2 1 T1

6 Ac 1 2 T2
RHS first term. Similar to the RHS first term of the bar element equation replacing q0
PhT
by q0 + , we can write
Ac
P A /2
q0 + hT
Ac A /2
RHS second term. Similar to RHS second term of the bar element (Eq. (2.125), we have
Q0

Ql
where Q0, Ql represent the heat flux at the ends of the element (nodes). Thus the element level
equations can be written as
k 1 1 Phl 2 1 T1 Ph A /2 Q0
+ = q0 + T +
A 1 1 6 Ac 1 2 T2 (4.163)
Ac
A /2

Ql
The LHS may be considered to be the element conductance matrix, and the first term
on the RHS is the equivalent nodal heat flux corresponding to the distributed heat source/sink.
The second vector on the RHS refers to the net heat flux at the nodes and permits us to
incorporate prescribed natural boundary conditions. When the individual elements are assembled
together, these Qs algebraically sum up to the resultant externally applied heat flux at the
section. Comparing Eq. (2.163) and (4.163), we observe that the latter contains certain
additional terms corresponding to convention.
We now illustrate the use of this element through the following example.
Example 4.10. Temperature distribution in a pin-fin. Consider a 1 mm diameter, 50 mm

long aluminium pin-fin as shown in Figure 4.26 used to enhance the heat transfer from a
surface wall maintained at 300C. Use k = 200 W/m/C for aluminium, h = 20 W/m2 C,
T = 30C. We obtained the temperature distribution in the fin using the Galerkin weighted
residual method in Example 2.8. We will now estimate the temperature in the pin using one, two,
and four equal size elements, assuming that the tip of the fin is insulated. There is no internal
heat source.
h T
Wall
TW
L
Fig. 4.26 A pin-fin.
One-element solution. Substituting the numerical values in Eq. (4.163), we get the
element level equations as
200 1 1 (Q )(0.001)(20)(0.05) 2 1 T1
+
0.05 1 1 (6)(Q )(0.0005)2 1 2 T2

(Q )(0.001)(20) 0.025 Qwall

= (30) + (4.164)
(Q )(0.0005) 2
0.025 Qtip
From the boundary conditions, we have T1 = 300C, Qtip = 0. Thus we get T2 = 198.75oC.
Two-element solution. With two equal elements, we have l = 0.025 m. Assembling the
two-element level equations in the same way as we did for the two bar elements, we get the
system level equations as
1 1 0 2 1 0 T1
(Q )(0.001)(20)(0.025)
200 1 2 1 + 1 4 1 T2
0.025 (6)(Q )(0.0005) 2
0 1 1 0 1 2 T3

0.0125 Qwall
(Q )(0.001)(20)
= (30) 0.025 + 0 (4.165)
(Q )(0.0005) 2
0.0125 Qtip
From the boundary conditions we have T1 = 300C, Qtip = 0. Substituting and solving
the equations, we get
T2 = 226.185C, T3 = 203.548C (4.166)
Four-element solution. With four equal elements, we have l = 0.0125 m. The

assembled equations can be obtained as
1 1 0 0 0 2 1 0 0 0 T1

1 2 1 0 0 1 4 1 0 0 T2

200 (Q ) (0.001) (20) (0.0125)
0 1 2 1 0 + 0 1 4 1 0 T3
0.0125 (6) (Q ) (0.0005)2
0 0 1 2 1 0 0 1 4 1 T4

1 1 0 0 0 1 2 T5
0 0 0
0.00625 Qwall

0.0125 0
(Q ) (0.001) (20)
= (30) 0.0125 + 0 (4.167)
(Q ) (0.0005) 2

0.0125 0

0.00625 Qtip
Substituting the boundary conditions and solving the equations, we get
T2 = 256.37C, T3 = 227.03C
(4.168)
T4 = 210.14C, T5 = 204.63C
The exact solution for this problem is given by
cosh m ( L x )
T ( x ) = T + (Tw T ) (4.169)
cosh mL
hP
where m = .
kAc
Figure 4.27 compares the exact solution with the various finite element solutions. It is
observed that the exact solution is an exponential function while the finite element permits only
linear variation in temperature within the element. The finite element approximation improves
with increasing number of elements. A 10-element solution is shown to match very closely with
the exact solution in Figure 4.28.
As observed with respect to the problem of a rod under self-weight (Example 4.1 and
4.2), we expect that fewer number of higher order elements, e.g. quadratic heat transfer
element, may give much better solution than the linear element used here.
Exact
One-element
Two-element
Four-element
T(X)
X
Fig. 4.27 Comparison of exact solution with finite element solutions (Example 4.10).
Exact
Ten-element
T (X )
X
Fig. 4.28 Comparison of exact solution with 10-element solution (Example 4.10).
SUMMARY
In this chapter, we have discussed the general formulation of 1-d finite element analysis. We
have solved several example problems from the fields of structural mechanics and heat transfer.
We have brought out several key issues, e.g. lower/higher order finite elements, consistent
nodal loads, generic procedure for assembly, modelling of thermal stresses, etc. In Chapter 5,
we will extend our studies to typical 2-d problems.
PROBLEMS
4.1 Consider a bar element whose area of cross-section varies linearly along the
longitudinal axis (Figure P4.1). Derive its stiffness matrix. How will this compare
with the stiffness matrix obtained assuming that the bar is of uniform c/s area equal
to that at its mid-length?
A1 1 2 A2 x
Fig. P4.1 Bar element with varying cross-section.
4.2 The fixed-fixed bar shown in Figure P4.2 has axial forces applied at L/3 and 2L/
3. Use the finite element method to compute the axial deflections and support
reactions.
3P P
L/3 L/3 L/3
Fig. P4.2 Fixed-fixed bar.
4.3 For the bar shown in Figure P4.3, give the assembled stiffness matrix for the two
node numbering schemes indicated. What effect does node numbering have on the
assembled matrix? Area of c/s = A; Youngs modulus E.
Problems 139
3 5 1 2 4 Scheme 2
1 2 3 4 5 Scheme 1
F
11 22 33 44
l l l l
Fig. P4.3 Effect of node numbering.
4.4 Consider the spring mounted bar shown in Figure P4.4. Solve for the displacements
of points P and Q using bar elements (assume AE = constant).
L
k
P Q 100 N
Fig. P4.4 Spring mounted bar.
4.5 The planar structure shown in Figure P4.5 is hinged at the supports A and B. The
elements of the truss have identical cross-sectional properties. Find out the
displacement of the node C under the action of the given loading.
2P
P
C
A, E
A, E, l
A B
A, E, l
Fig. P4.5 A planar structure.
4.6 Determine the deflections for the truss structure shown in Figure P4.6.
P
A, E, l
A, E, l
60 60 k
Fig. P4.6 A truss structure.
4.7 Consider a rod rotating about a fixed axis as shown in Figure P4.7 at a constant
speed w rad/s. Derive the element nodal force vector due to the centrifugal force
field. Consider both linear and quadratic bar elements. Let the density be r and
uniform cross-sectional area A.
w rad/s
Dia, d
ri
r0
Fig. P4.7 Rotating rod.

Problems 141
4.8 Consider a truss element arbitrarily oriented in space. If the coordinates of the two
ends of an element are given by (x1, y1, z1) and (x2, y2, z2), we can get the direction
cosines of the line vector as
'x 'y
A = , m =
('x) + ('y) + ('z )
2 2 2
('x) + ('y)2 + ( 'z )2
2
'z
n =
( ' x) + ( ' y ) 2 + ( ' z ) 2
2
where Dx = x2 x1; Dy = y2 y1; Dz = z2 z1. Show that the transformation matrix

[T] relating the nodal displacements in local and global coordinate frames is given by
l m n 0 0 0
[T ] =
0 0 0 l m n
Derive the element stiffness matrix in the global frame XYZ.
4.9 For a uniform circular c/s bar (diameter d, length L, modulus of rigidity G) under
twist, determine the stiffness equations, i.e.
T1 k11 k12 G1
=
T2 k21 k22 G2
Use this to solve for the angle of twist f at points A and B in Figure P4.9
Dia, 2d
T Dia, d
A B
L L
Fig. P4.9 Torsion of a stepped shaft.
4.10 A crude approximation of a dam is shown in Figure P4.10. Using one beam element
and consistent nodal load vector, find the tip deflection. Using this result find the
deflection of the point P.
4.11 For the beam problem shown in Figure P4.11 determine the tip deflection and slope
at the roller support.
4.12 A beam element is loaded by a moment at the mid-span. Find the consistent nodal
force vector. How would the result change if the moment were applied at one-third
the span rather than at the centre.
L/3
EI = 2 108 N/m2
P 2L/3
L/3
Fig. P4.10 A cantilever beam under hydrostatic load.
q = 1000 N/m 1000 N

M
2m
K = 200 N/m
3m
Fig. P4.11 Propped cantilever beam.
4.13 Describe how you would implement the boundary condition for the beam shown in
Figure P4.13.
4.14 For the discretisation of beam elements shown in Figure P4.14, number the nodes
so as to minimise the bandwidth of the assembled stiffness matrix [K].
4.15 The elements of a row or column of the stiffness matrix of a bar element sum up
to zero, but not so for a beam element. Explain why this is so.
4.16 The plane wall shown in Figure P4.16 is 0.5 m thick. The left surface of the wall
is maintained at a constant temperature of 200C, and the right surface is insulated.
Problems 143
q0 P
k2
k1
q
Fig. P4.13 Beam on spring supports.
7 9 11
5 12
44 13
6 8 10
33 14
15
2
1 16
Fig. P4.14 Node numbering.
0.5 m
Insulated
T = 200C
Q = 400 W/m3
k = 25 W/mC
Fig. P4.16 Heat conduction in a slab.

The thermal conductivity k = 25 W/mC, and there is a uniform heat generation

inside the wall of Q = 400 W/m3. Determine the temperature distribution through
the wall thickness using linear elements.
4.17 For a fully developed laminar flow between two long parallel plates (reparated by
a distance 2L), the governing equation is
d 2u dP
N =
dy 2 dx
where m is the viscosity of the fluid.
Derive the finite element form and hence determine the velocity distribution uy for
a given constant pressure gradient dP/dx. The boundary conditions are that the fluid
velocity is zero at the surface of the plates i.e. u(L) = 0; u(L) = 0.
4.18 Imagine that a uniform bar element is to have two DOF per node, namely, axial
displacement u and axial strain ex. How should we proceed to derive the element
characteristic matrices?
4.19 How would you formulate a frame element that would enable us to model a buckling
problem?
4.20 In general, a thin long structural member such as a beam can bend in two planes,
undergo extensional as well as twisting deformation along its axis. Thus we would
like to formulate our general one-dimensional element for structural analysis with
two nodes and six d.o.f. per node, viz., u, v, w, qx, qy, qz. Formulate such an element
and derive its matrices.
4.21 Develop a computer program for finite element analysis using the finite elements
discussed in this chapter. You may use the following steps:
Allow the user to choose the type of analysis (structural/heat transfer, etc.) and the
type of element (truss/beam/frame, etc., linear/quadratic, and so on).
Accept appropriate input data, orientation of the element, local and global node number
correspondence, boundary condition details, etc.
Generate element matrices and transform to global reference frame.
Perform assembly incorporating boundary conditions.
Solve for unknown variables.
From the nodal variables of each element, compute the quantities of interest, e.g.
stresses, heat flux.
CHAPTER 5
Two-dimensional Finite
Element Analysis
5.1 IntroductionDimensionality of a Problem

In Chapter 4 we discussed one-dimensional finite element analysis in general and applied the
technique to problems of structural mechanics and heat transfer. We formulated the truss,
torsional bar and beam elements and then combined all of them into a general 1-d element for
structural analysis. We discussed the general one-dimensional heat transfer element with
conduction as well as convection. While we solved many example problems, we did not
emphasise how a real-life problem gets itself posed as a one-dimensional problem in the first
place. We now briefly discuss this important issue.
Let us consider a space frame or space truss structure with arbitrary configuration in
3-d space. We consider this to be a typical one-dimensional problem. The nodal d.o.f. vary only
along one direction, viz., along the length of the element (say x). In view of the fact that the
members are long and slender, we can make suitable assumptions on the variation of
displacements in the cross-sectional plane. For example, for a slender bar of uniform cross-
section, we can make a reasonable assumption that the deflection varies in a known way across
the cross-section (e.g. uniform, linear). Many a time, for complex real-life structures such as
a gear tooth or a thick pressure vessel or a crank shaft, the unknown field variables vary
independently along two or even three directions and we cannot make any suitable assumptions
about their variation. For a block of material of comparable dimensions, for example, the heat
transfer is more closely approximated as 3-d rather than as 1-d or 2-d. Thus we seek to
determine the entire field of variation of the unknown variable using the Finite element method,
and such problems are said to be two- or three-dimensional problems. While every real-life
problem can be modelled as a three-dimensional problem, it may be computationally very
expensive to perform such an analysis. In a design situation, typically, several iterations are
carried out before a final design is frozen. Thus, repeated 3-d analysis can be very cumbersome
and we look for simpler but reasonably accurate one/two-dimensional models. The detailed
discussion of three-dimensional finite element analysis is beyond the scope of this text. We will
discuss in detail several two-dimensional elements and related issues in this chapter.
A structural mechanics problem can be considered to be two dimensional if we are able
to make reasonable assumption about the variation of unknown parameters in the third
dimension. Three practically important situations of this type, arising in structural mechanics,
are shown in Figure 5.1. In the first case [Fig. 5.1a], the structure is very thin as compared
145
146 Two-dimensional Finite Element Analysis
(a) Thin disk under plane stress
z
T x
(b) Long, prismatic shaft under plane strain
(c) Thick cylinder under axisymmetric loads
Fig. 5.1 Typical 2-d problems in stress analysis.

Introduction Dimensionality of a Problem 147
to its other two dimensions and is loaded in its own plane, e.g. a rotating impeller wheel. Thus
the top and bottom faces are free from any normal stress sz and since the thickness is small,
we assume sz is zero throughout. So also the shearing stresses txz, tyz. Thus the only nonzero
stresses are sx, sy, txy. The stress-strain constitutive relation for a linear, elastic, isotropic
material under such plane-stress condition is given by

T x 1 O 0 Fx

E F 0
T y = O 1 0 {}
F
1 O2 y (5.1)
U xy 1O H
0 0 xy
2
where {e}0 represents an initial strain.
The problem situation depicted in the second case (Fig. 5.1b) is the other extreme of the
first case, where the length of the member is so long that it is reasonable to assume the
longitudinal strain ez to be zero, e.g. torsion of a long uniform shaft. So also the shearing strains
gxz, gyz. Thus the only nonzero strains are ex, ey, gxy. The stress-strain constitutive relation for
a linear, elastic, isotropic material under such plane-strain condition is given by

T x 1 O O 0 Fx

E F
T y = O 1O 0 {F}
0
(5.2)
(1 + O )(1 2O ) y
U 1 2O H
xy 0 0 xy
2
It is to be observed that the normal stress sz is nonzero but can be readily obtained from
the condition that the longitudinal strain ez be zero.
The problem situation depicted in the third case (Fig. 5.1c) represents an important class
of problems, commonly described as axisymmetric problems, e.g. pressure vessels. If the
geometry of the structure, material properties, support conditions and loading are all
axisymmetric, then it is reasonable to assume that the resulting deformation will also be
axisymmetric, i.e., it does not vary along the circumferential direction. Thus, it is sufficient to
determine the variation of the unknown field variable (e.g. deflection, temperature) in any one
plane as shown in the figure. In practical pressure vessels, there may be minor geometrical
features which cause deviation from perfect axial symmetry such as helical screw threads on
an end-cap. It may be worthwhile to ignore the small helix angle and model the problem as
purely 2-d axisymmetric. However, sometimes a perfect axisymmetric structure may be
subjected to nonaxisymmetric loading such as earthquake loading. It may still be useful to
modify a two-dimensional axisymmetric analysis rather than going for a full-fledged three-
dimensional analysis. The non-axisymmetric loading, expressed as a function of circumferential
coordinate q, has a periodicity 2p. It can be expressed as a Fourier series. To each harmonic
of load, the resulting deflection also varies as a sine or cosine function and it can still be
analysed as a two-dimensional problem. The Fourier harmonics of deflection are superposed
to find the total response. This technique is known as semi-analytical FEM. The nonzero
stresses and strains in a perfect axisymmetric situation are related as follows:
O O
1 0
1 O 1O
T r Fr

O
T R (1 O ) E 1 0 FR
= 1O {F } (5.3)
0
T z (1 + O )(1 2O ) Fz
1 0
U rz H
1 2O rz
Sysmmetric 2(1 O )

Sometimes, a one-dimensional problem may have to be modelled as a two-dimensional

problem, e.g. a beam with a transverse load, and we wish to determine the stresses under the
point of application of load. Thus, if we are interested in detailed modelling of local effects such
as concentrated loads, and sudden changes in geometry (such as a stepped sheft), a one-
dimensional system may have to be modelled using two-dimensional finite elements.
5.2 Approximation of Geometry and Field Variable

A typical two-dimensional problem domain is shown in Figure 5.2, where we depict a bracket
with an internal hole subjected to loads. We immediately observe that unlike in one-dimensional
Fig. 5.2 Typical 2-d problems: Bracket with an internal hole.

Approximation of Geometry and Field Variable 149
finite element analysis given in Chapter 4, here we need to model the complex two-dimensional
geometry of the system as well as the variation of the unknown field variable. To represent the
two-dimensional geometry, we have to formulate two-dimensional elements, and most common
elements consist of triangles and quadrilaterals as shown in Figure 5.3. Each node is capable
of independent motion along the x and y axes and, typically, we have u and v as the independent
nodal degree of freedom (d.o.f.) for plane elasticity problems. However, there is only one
temperature d.o.f. for 2-d heat transfer problems. Simple lower order elements have straight
sides and nodes only at the vertices. More complex higher order elements permit curved edges
(so as to model curved geometric feature in the domain such as a circular hole) and mid-side
or even internal nodes. 3
3
5
6
2
4
1 2 1
3
7
3
6
4 4
8 2
2
1 5
2 1
Fig. 5.3 Typical 2-d finite elements.
We will now begin our discussion of two-dimensional elements. We will first discuss a
few typical elements and derive their shape functions. Subsequently, we will discuss the
formulation of 2-d structural mechanics and heat transfer/fluid flow problems using these
elements. We will now begin our discussion with the simplest element, viz., the three-noded
triangular element. We will keep in mind that each node of the elements has just one d.o.f. (viz.,
temperature T ) when applied to heat transfer problems and two displacements, u, v along the
coordinate axes for structural mechanics problems. The flow of an ideal fluid in 2-d can be
represented in terms of a scalar potential function and so we have only one d.o.f. per node.
5.2.1 Simple Three-noded Triangular Element

A typical element is shown in Figure 5.4. It has straight sides and three nodes, one at each
vertex. The nodes have coordinates (x1, y1), (x2, y2), (x3, y3) in the global Cartesian coordinate
Y
v3
(x3, y3) 3 u3
v2
u2
2 (x2, y2)
v1
(x1, y1) 1 u1
O X
Fig. 5.4 Simple three-noded triangular element.
frame OXY as shown in the figure. Each node has just temperature d.o.f. T or two d.o.f., viz.,
u and v, the translations along global X and Y axes, respectively. In what follows, we derive
the shape (interpolation) functions for the temperature field and use the same for the
displacement field.
We assume that the temperature field over the element is given by
T(x, y) = c0 + c1x + c2y (5.4)
Considering that this expression should reduce to the nodal temperature at the nodal
points, we have
T1 = c0 + c1x1 + c2 y1
14243
T2 = c0 + c1x2 + c2 y2 (5.5)
T3 = c0 + c1x3 + c2 y3
Solving for c0, c1, c2, we get
1
c0 1 x1 y1 T1

c1 = 1 x2 y2 T2 (5.6)

c2 1 x3 y3 T3
Substituting these in Eq. (5.5) and rewriting, we finally obtain
B + C1 x + H 1y B 2 + C2 x + H 2y B 3 + C3 x + H 3y
T ( x, y ) = 1 T1 + T2 + T3 (5.7)
2 2 2
where
a1 = x2y3 x3y2, b1 = y2 y3, g1 = x3 x2 (5.8)
Other coefficients (a2, b2, g2) and (a3, b3, g3) can be obtained by a simple cyclic
permutation of subscripts 1, 2, 3. Also,
1 x1 y1
2 = 1 x2 y2 = 2 (Area of triangle 123) (5.9)

1 x3 y3
In our standard finite element notation, we write this as

T(x, y) = N1T1 + N2T2 + N3T3 = [N]{T}e (5.10)
Thus we have the shape functions Ni as follows:
1
Ni = {B i + Ci x + H i y}, i = 1, 2, 3 (5.11)
2
If this element is used to model structural mechanics problems, each node will have two
d.o.f., viz., u and v, and we can write the displacements at any interior point in terms of the
nodal displacements using the same shape functions as
u(x, y) = N1u1 + N2u2 + N3u3 (5.12)
v(x, y) = N1v1 + N2v2 + N3v3 (5.13)
In our standard finite element notation, the displacement field at an interior point in the
element is written in terms of the nodal d.o.f. as
u1

v1

u N1 0 N2 0 N3 0 u2
= = [ N ]{E }e (5.14)
v 0 N1 0 N2 0 N 3 v2

u3

v3
We observe that this element may only permit a linear variation of the unknown field
variable within the element and thus derivatives such as strains and heat flux will be constant
throughout the element. In structural mechanics, this element is therefore popularly known as
the constant strain triangle (CST). We now discuss a typical rectangular element.
5.2.2 Four-noded Rectangular Element

A typical element is shown in Figure 5.5. It is rectangular and has four nodes, one at each
vertex. The nodes have coordinates (X1, Y1), (X2, Y2), (X3, Y3) and (X4, Y4) in the global
Cartesian coordinate frame (OXY) as shown in the figure. For ease of derivation, we define
4 3
y h
o
1 x 2
l
O X
Fig. 5.5 A four-noded rectangular element.
a local coordinate frame (oxy). Each node has temperature d.o.f. T or two d.o.f., viz., u and
v, the translations along the global X and Y axes, respectively. In what follows, we will derive
the shape (interpolation) functions for the temperature field and use the same for the
displacement field.
We assume that the temperature field over the element is given by
T(x, y) = c0 + c1x + c2y + c3xy (5.15)
We observe that as we have four nodes now, we are able to take a fourth term in our
polynomial field. Considering that this expression should reduce to the nodal temperatures at
the nodal points we have, for a rectangular element of size (l h) as shown in Figure 5.5,
T1 = c0, T2 = c0 + c1l, (5.16)
T3 = c0 + c1l + c2h + c3lh, T4 = c0 + c2h
Solving for c0, c1, c2 and c3, we get
T2 T1 , T4 T1 T3 + T1 T2 T4
c0 = T1 , c1 = c2 = , c3 = (5.17)
A h Ah
Substituting these in Eq. (5.15) and rewriting, we obtain
x y xy x xy xy y xy
T(x, y) = 1 + T1 + T2 + T3 + T4 (5.18)
A h Ah A Ah Ah h A h
In our standard finite element notation, we write this as
T1

T2
T ( x, y ) = [N1 N2 N3 N4 ] (5.19)
T3

T4
Thus we obtain the shape functions Ni:
x y xy x xy
N1 = 1 + , N2 =
A h Ah A Ah
xy y xy (5.20)
N3 = , N4 =
Ah h Ah
If this element is used to model structural mechanics problems, each node will have two
d.o.f., viz., u and v, and we can write the displacements at any interior point in terms of the
nodal displacements using the same shape functions as follows:
u1

v1

u2

u N1 0 N2 0 N3 0 N4 0 v2
= = [ N ]{E }
e

v 0 N1 0 N2 0 N3 0 N 4 u3
(5.21)
v3

u4

v4
We observe that this element permits a linear variation of unknown field variable along
x = constant or y = constant lines, and thus it is known as the bilinear element. Strain and
heat flux are not necessarily constant within the element.
We will now discuss a typical higher order element.
5.2.3 Six-noded Triangular Element

A typical higher order element is shown in Figure 5.6. It has six nodes, three at the vertices
and three on the edges. While usually the edge nodes are located at the mid-point, it is not
necessary. By convention, we number the nodes as shown in the figure. The nodes have
3
5
6 2
O X
Fig. 5.6 A six-noded triangular element.
coordinates (x1, y1), (x2, y2), K, (x6, y6) in the global Cartesian coordinate frame OXY as shown
in the figure. Each node has just temperature d.o.f. T or two d.o.f., viz., u and v, the
translations along global X and Y axes, respectively.
Since we have six nodes, we can fit a higher degree polynomial for the variation of the
unknown field variable. To be precise, following our procedure, we can determine up to six
coefficients ci(i = 0, 1, K, 5). Thus we assume that the temperature field over the element
is given by the complete quadratic polynomial
T(x, y) = c0 + c1x + c2y + c3x2 + c4xy + c5y2 (5.22)
Considering that this expression should reduce to the nodal temperature at the nodal
points, we have
T1 = c0 + c1 x1 + c2 y1 + c3 x12 + c4 x1 y1 + c5 y12
1444442444443
T2 = c0 + c1 x2 + c2 y2 + c3 x22 + c4 x2 y2 + c5 y22
T3 = c0 + c1 x3 + c2 y3 + c3 x32 + c4 x3 y3 + c5 y32
(5.23)
T4 = c0 + c1 x4 + c2 y4 + c3 x42 + c4 x4 y4 + c5 y42
T5 = c0 + c1 x5 + c2 y5 + c3 x52 + c4 x5 y5 + c5 y52
T6 = c0 + c1 x6 + c2 y6 + c3 x62 + c4 x6 y6 + c5 y62
We can, in principle, solve for coefficients c0 c5 in terms of nodal coordinates (x1, y1),
(x2, y2), K, (x6, y6), and nodal temperatures T1 T6 and substitute in Eq. (5.22) above and
rearrange the terms to get the desired shape functions. However, it would be very tedious to
get explicit expressions for the shape functions in this manner. Further computations, in
particular evaluation of integrals, involved in deriving the element matrices, would also turn out
to be quite intensive.
Our elements so far were limited to only straight-sided triangular and rectangular shape.
To model typical structural geometries, we would like our elements to possess, in general,
curved edges. We would also like to have alternate methods of deriving the shape functions and
a way by which the computations can be performed inside a computer program through
methods such as numerical integration.
In order to achieve these goals, we now introduce dimensionless natural coordinates,
viz., x-h for typical quadrilateral elements. We propose that we have a parent rectangular
element in natural coordinates, which is transformed to the child element in the real x-y
coordinates as shown in Figure 5.7. The natural coordinates vary from 1 to +1, and the size
of the parent element will be 2 2 units, irrespective of the actual element size in real Cartesian
coordinates x-y. Depending on the transformation used between the (x-h) and the (x-y) frames,
we observe that a straight-sided rectangular parent element can be transformed into an element
with curved edges in the x-y coordinates. We observe that even when we use natural
h
4 3 3 (x3, y3)
(1, 1) (1, 1)
(1,1)
(x4, y4) 4
O x
(x2, y2)
2
(1, 1) (1, 1) 1
1 2 (x1, y1)
O
X
Fig. 5.7 A general four-noded quadrilateral child element with the parent square element.
coordinates for the formulation, the nodal d.o.f. (such as displacements) are along the physical
x-y coordinates and not along (x-h). We will now discuss the formulation of the finite elements
using natural coordinates. We will see that the formulation now involves coordinate
transformation and seem to be formidable but, in fact, permits us to develop a powerful generic
structure for finite element development. We will begin our discussion with an introduction to
the natural coordinates in one dimension/two dimensions, and then take up the formulation of
typical finite elements.
5.3 Natural Coordinates and Coordinate Transformation

We will designate the natural coordinate used for one-dimensional case by x. Consider a child
line shown in Figure 5.8 with two end points defined by their coordinates A(x1) and B(x2). The
Q P Q P
X
A x B B
A
(x = 1) (x = 1) (x = x1) (x = x2)
Fig. 5.8 Natural coordinates in 1-d.
parent line indicating x varying from A(1) to B(+1) is also shown in the figure. An
appropriate linear transformation between x and x is given by
x + x2 x2 x1
x(Y ) = 1 + Y
2 2
1 Y 1 + Y
= x1 + x2

2 2
(5.24)

N1 N2
= N1 x1 + N 2 x2
where Ni represent the interpolation functions used for coordinate transformation. Here we
interpolate the coordinates of an interior point (within the element) from the coordinates of the
nodal points in the same way as we interpolated the unknown field variable. If, in fact, the same
shape functions are used for interpolating the unknown field variable and geometry
transformation, we call them isoparametric elements. If the shape functions used for coordinate
transformation are of a lower degree than those for the unknown field variable, then we call
them sub-parametric elements. If, on the other hand, the shape functions used for coordinate
transformation are of a higher degree than those for the unknown field variable, then we call
them Superparametric elements.
Rewriting the above equation in our standard notation, we may write
x1
x = [N1 N 2 ] (5.25)
x2
For a point P , x = 0.5, correspondingly,
x1 + 3 x2
x =
4
which is the point P.
For a point Q,
x2 + x1
x = x1 +
4
correspondingly, x = 0.5 which is the point Q.
Natural Coordinates and Coordinate Transformation 157
X
A C x B
A B
C
(x = -1) (x = 0) (x = 1) (x = x ) (x = x3) (x = x2)
1
Fig. 5.9 Quadratic transformation between x and x.
In order to fit a quadratic transformation between x and x, we need one more point C
as shown in Figure 5.9. The corresponding parent line (A C B) is also shown in the figure.
We observe that C is always located at x = 0 while the corresponding point C can be located
at any desired position on the line AB in the Cartesian frame.
We can write the transformation as
x (x ) = Ni xi = N1x1 + N2x2 + N3x3 (5.26)
We will now discuss different ways of deriving the shape functions Ni in natural
coordinates.
5.3.1 Alternate Methods of Deriving Shape Functions

Serendipity approach
An easy way of writing down the necessary shape or interpolation functions by inspection is
as follows: If we observe the characteristic of shape functions (e.g. ref. Eq. (5.24)), we see
that each Ni takes a value of unity at the point i and goes to zero at all other points j i. This
has to be so in order to satisfy the relation x = Ni xi, K. We can use this fact to write down
the shape functions very easily. For example, we can now write the shape function N1 in
Eq. (5.26) based on the above observation that N1 should be unity at point 1 and zero at points
2 and 3. Since the coordinates of points 2 and 3 are x = 1 and 0, respectively, the functions
which become zero at the points 2 and 3 are simply (x 0) and (x 1). Thus we can write
the required shape function as:
N1 = (SF )(x 0)(x 1) (5.27)
where the scale factor (SF) is determined from the requirement that N1 be unity at point 1, i.e.
(x = 1). We readily find that SF = 0.5. Thus the required shape function is
N1 = 0.5(x 0) (x 1) (5.28)
On similar lines, we can directly write the other two shape functions as
N2 = 0.5(x 0)(x + 1) (5.29)
N3 = (x + 1)(x 1) (5.30)
In this way, we can also write down very easily the shape functions for any complex
finite element. Considering the example of quadratic coordinate transformation once again, we
get, using Eqs. (5.28)(5.30) in Eq. (5.26), the required transformation
x(x ) = N1x1 + N2x2 + N3x3 = 0.5(x )(x 1) x1 + 0.5(x )(x + 1) x2 (x + 1)(x 1)x3 (5.31)
If point 3 in the physical space x-y were at the middle of the line, i.e. x3 = (0.5)
(x1 + x2), then we observe from the above equation that the quadratic transformation just
obtained reduces to the simple linear transformation obtained earlier. However, if point 3 is
located away from the mid-point, say for example, x3 = x1 + (0.25)(x2 x1), then we require
that a quarter point in the real Cartesian plane be transformed to the mid-point in the x plane.
This cannot be achieved through a linear transformation and Eq. (5.31) becomes (for simplicity,
let x1 = 0, x2 = 4, x3 = 1),
x(x ) = 0.5(x )(x 1)(0) + 0.5(x )(x + 1)(4) (x + 1)(x 1)(1)
= (1 + x )2 (5.32)
Thus we can fit a nonlinear transformation between x and x. Such a transformation can
be used to formulate finite elements which possess curved edges so that we can model curved
structural geometry.
Lagranges interpolation
If we observe our process of interpolation of either the coordinates or the field variable from
the nodal values, we notice that it is simply to find a polynomial curve fit passing through the
prescribed function values at specified points (nodes). It is possible to achieve this through the
classical Lagranges interpolation formula. If we are given, in general, a set of data points
x1, x2, K, xn and the corresponding function values at these data points as f1, f2, K, fn, we
can then write
f (x) L1 f1 + L2 f2 + L + Ln fn (5.33)
where Li are the Lagrange polynomials given by
( x1 x)( x2 x) " ( xi 1 x)( xi +1 x) " ( xn x)
Li = (5.34)
( x1 xi )( x2 xi ) " ( xi 1 xi )( xi +1 xi ) " ( xn xi )
We observe that the Li given above are simply the required shape functions Ni. Thus we
can write the shape functions for the quadratic coordinate transformation above as follows:
Given the data points (x1 = 1), (x2 = 1) and (x3 = 0) and the corresponding function values
x1, x2, and x3, we can write the required Lagrange interpolation polynomials as
(Y 2 Y )(Y3 Y ) (1 Y )(0 Y ) Y (Y 1)
N1 = = = (5.35)
(Y 2 Y1 )(Y3 Y1 ) (1 + 1)(0 + 1) 2
(Y1 Y )(Y3 Y ) ( 1 Y )(0 Y ) Y (1 + Y )

N2 = = = (5.36)
(Y1 Y2 )(Y3 Y2 ) ( 1 1)(0 1) 2
(Y1 Y )(Y 2 Y ) (1 Y )(1 Y )

N3 = = = 1 Y2 (5.37)
(Y1 Y3 )(Y 2 Y3 ) (1)(1)
We observe that, in general, the shape functions derived following either approach, viz.,
Serendipity or Lagranges, will be different. Since the shape function is the most fundamental
property of the finite elements, elements formulated based on these shape functions also exhibit
different characteristics. Elements formulated based on Serendipity (Lagrange) shape functions
are popularly known as Serendipity (Lagrange) elements. In general, Lagrange elements tend to
have more internal nodes and admit more, higher degree terms into the polynomial shape function.
5.3.2 Natural CoordinatesQuadrilateral Elements

We will designate the natural coordinates used for two-dimensional domains by x and h. Thus
we are looking for transformation between (x-y) and (x-h). Consider the general quadrilateral
ABCD shown in Figure 5.10. The parent square ABCD indicating x, h varying from 1 to
+1 is also shown in the figure. If the element in the (x-y) frame were rectangular (size l h),
h
(1, 1) (1, 1) 3
(x3, y3)
4 3
D C C
4
(x4, y4)
D
O x
B (x2, y2)
2
A
A B 1
1 2 (x1, y1)
(1, 1) (1, 1)
O X
Fig. 5.10 A general four-noded quadrilateral element.
it is easy to see that the coordinate transformation would be simply a scale factor and is given
by
2x
Y = (5.38)
A
2y
I = (5.39)
h
For a general quadrilateral element, we have to obtain the transformation
(x, y) f (x, h) (5.40)
This transformation, which yields the coordinates of a point P(x, y) within the element in terms
of the nodal coordinates (just as we interpolate the displacement/temperature field), can be
written as
xP = Ni xi, yP = Ni yi (5.41)
where i = 1, 2, K, NNOEL which stands for the number of nodes per element (= 4 here).
Considering that Ni must be unity at node i and zero at all other nodes, the shape functions
Ni can be obtained as
N1 = (SF) (Equation of line 23) (Equation of line 3 4)
= (SF) (1 x ) (1 h ) (5.42)
where the scale factor (SF) is to be chosen such that N1 is unity at node 1 and the product
of equation of lines 2 3 and 3 4 has been taken to ensure that N1 is zero at nodes 2, 3
and 4.
At node 1, x = h = 1 and, therefore, SF = 0.25. Thus our shape function is given by
1
N1 = (1 Y ) (1 I ) (5.43)
4
On similar lines, we can obtain the other shape functions as
1
N 2 = (1 + Y )(1 I ) (5.44)
4
1
N3 = (1 + Y )(1 + I ) (5.45)
4
1
N 4 = (1 Y ) (1 + I ) (5.46)
4
While we have derived the shape functions by the Serendipity approach, the reader is
advised to obtain the same using Lagranges approach. Figure 5.11 shows a plot of these shape
functions.
4 3 4 3
1.0 1.0
1 2 1 2
(a) N1 (b) N2
1.0 1.0
4 3 4 3
1 2 1 2
(c) N3 (d) N4
Fig. 5.11 Shape functions Ni (i = 1, 2, 3, 4) for Quad4 element.
Using this coordinate transformation, we will be able to map a parent square element in
(x-h) space into a general quadrilateral element in physical (x-y) coordinate space. Consider, for
example, the general quadrilateral shown in Figure 5.12. Using the nodal coordinate data given,
we can write
x = S Nixi = N1 + N3 (5.47)
y = S Niyi = 5N2 + 5N4 (5.48)
Thus we have
1
x =
4
[(1 + Y )(1 + I ) (1 Y )(1 I )] (5.49)
5
y =
4
[(1 Y )(1 + I ) (1 + Y )(1 I )] (5.50)
Using this transformation relation, we can make certain interesting observations. If we
consider the line x = 0 in the (x-h) space, its mapping onto the (x-y) space is given by
x = (1/2)h, y = (5/2)h (5.51)
Therefore, the equation of the line in the (x-y) space is y = 5x.
If we consider the line h = 0 in the (x-h) space, its mapping onto the (x-y) space is given
by
x = (1/2)x, y = (5/2)x (5.52)
Thus the equation of the line in the (x-y) space is y = 5x. These two lines are indicated in
Figure 5.12. We observe that the x-h axis lines, when transformed into the physical (x-y) space,
need not be parallel to the x and y axes. The transformed lines need not even be orthogonal
in the x-y space. We reiterate the fact that our nodal d.o.f. (such as displacements) are along
real, physical x-y axes, and not along x-h.
Y
4 (0, 5)
x = 0 line ( y = 5x)
(1, 0) 1 3 (1, 0)
X
h = 0 line (y = 5x)
2 (0, 5)
Fig. 5.12 Natural and Cartesian coordinates for a general quadrilateral element.
For the four-noded element, with the above linear shape functions, we have been able to
transform a parent square element in (x-h) space into a general quadrilateral child element in
the physical (x-y) space. If we need general curved edge element, we require a higher order
transformation and we will now discuss the eight-noded element for this purpose.
Consider the general quadrilateral element shown in Figure 5.13 with eight nodes. The
parent eight-noded element in (x-h) space has also been shown in the figure. The attention of
Y
h 3 (x3, y3)
(x7, y7) 7
7 (0,1)
(0, 1)
(1, 1) 4 3 (1, 1) 6 (x6, y6)
(x4, y4) 4
(1, 0) 8 6 (1, 0) 2 (x2, y2)

x (x8, y8) 8
5 (x5, y5)
(1, 1) 1 2 (1, 1) 1 (x1, y1)
5 (0, 1)
O X
Fig. 5.13 A general eight-noded quadrilateral child element with its parent.
the reader is drawn to the convention followed in numbering the nodes. The shape functions
Ni can be obtained as
N5 = (SF)(Eq. of line 263)(Eq. of line 374)(Eq. of line 481)
= (SF)(1 x )(1 h)(1 + x ) (5.53)
of equation of lines 263, 374 and 481 has been taken to ensure that N5 is zero at all
other nodes.
At node 5, x = 1 and h = 0 and, therefore, SF = 0.5. Thus our shape function is given
by
1
(
N 5 = 1 Y 2 (1 I )
2
) (5.54)
On similar lines, we can obtain the other shape functions as

1
(
N 6 = (1 + Y ) 1 I 2
2
) (5.55)
1
( )
N 7 = 1 Y 2 (1 + I )
2
(5.56)
1
N8 = (1 Y ) 1 I2
2
( ) (5.57)
We observe, for example, that N5 is quadratic in x but linear in h. Similarly, N8 is quadratic

in h but linear in x.
Referring to the plot of N1 for the four-noded element (Fig. 5.11(a)), we can infer that
we can readily modify that shape function to obtain the shape function N1 for the present eight-
noded element by making it vanish at nodes 5 and 8 also. Thus,
1 1
N1 8 Node N1 4 Node N 5 N8
2 2
1 1
(1 Y )(1 I ) (1 Y )(1 I )(2 Y I ) (5.58)
4 4
1
(1 Y )(1 I )( 1 Y I )
4
On similar lines we can obtain the shape functions for other nodes as
1 1
N2 8 Node
= N2 4 Node
N5 N6
2 2
1 (5.59)
= (1 + Y )(1 I)(1 + Y I)
4
1 1
N 3 8 Node = N3 4 Node
N5 N7
2 2
1
= (1 + Y )(1 + I )( 1 + Y + I ) (5.60)
4
1 1
N 4 8 Node = N 4 4 Node N 7 N8
2 2
1
= (1 Y )(1 + I)( 1 Y + I) (5.61)
4
4 7
3
8
6
1
5 2
Fig. 5.14 N1 function of Quad-8 element.

As an illustration, shape function N1 is plotted in Figure 5.14. Using these shape functions,
it is possible to map a straight line in the (xh) space to a curve in the physical (xy) space.
Consider edge 152 (h = 1). On this edge, we have N3 = N4 = N6 = N7 = N8 = 0. Thus we
have, for this edge, the relations
x = N1x1 + N2x2 + N5x5 (5.62)
y = N1y1 + N2y2 + N5y5 (5.63)
Thus,
Y (1 Y ) Y (1 + Y )
x = x1 + x2 + (1 Y 2 ) x5 (5.64)
2 2
Y (1 Y ) Y (1 + Y )
y = y1 + y2 + (1 Y 2 ) y5 (5.65)
2 2
For the general values of nodal coordinates (x1, y1), (x2, y2) and (x5, y5), this would
represent a nonlinear transformation. For example, if the nodal coordinates are given by
(0, 0), (3, 9) and (2, 4), then we have
3 1
x = (1 + Y ) + (1 Y 2 ) (5.66)
2 2
9 1
y = (1 + Y ) (1 Y 2 ) (5.67)
2 2
Therefore, the equation representing the edge 152 in the Cartesian space is (eliminating x
from Eqs. (5.66)(5.67))
x2 + y2 + 2xy + 42x 30y = 0 (5.68)
Thus it is possible, in general, to map a straight line in the (xh) space to a curve in the physical
(xy) space, and we can thus develop elements with curved edges so as to be able to model
curved domains. In the special case, when node 5 is located at the mid-point of edge 12 in
the physical space, i.e., if
x + x2
x5 = 1 (5.69)
2
y + y2
y5 = 1 (5.70)
2
then the transformation (as given in Eq. (5.64))(5.65) and reduces to the simple case of
mapping a straight edge 152 in the (xh) space to another straight edge in the physical
(xy) space. Thus, by making the nodes 5, 6, 7 or 8 move away from the middle of their edges
in the physical (xy) space, we can achieve curved edge elements. It is observed that,
irrespective of their locations in the physical (xy) space, nodes, 5, 6, 7, 8 are at the middle
of their edges in the xh frame.
5.3.3 Natural CoordinatesTriangular Elements

The natural coordinates for triangular elements are conveniently defined as shown in
Figure 5.15. The three natural coordinates are defined by the following expressions:
A1 Area of ' P 23
L1 = = (5.71)
A Area of '123
P (L1, L2, L3)

1
O X
Fig. 5.15 Natural coordinates for triangular elements.
A2 Area of '1P3
L2 = = (5.72)
A Area of '123
A3 Area of 'P12
L3 = = (5.73)
A Area of '123
It is observed that in the parent equilateral triangle, each of the natural coordinates varies
from 0 to 1. We observe that the three natural coordinates for a point are not independent and
should sum up to unity. Thus,
L1 + L2 + L3 = 1 (5.74)
For a simple three-noded triangle, the required linear transformation can readily be written
as
x = L1x1 + L2x2 + L3x3 (5.75)
y = L1y1 + L2y2 + L3y3 (5.76)
In our standard finite element notation, the coordinate mapping is written as
x1

y1

x N1 0 N2 0 N3 0
x2
= (5.77)
y 0 N1 0 N2 0 N3 y2

x3

y3
where the shape functions are simply obtained as

N1 = L1, N2 = L2, N3 = L3 (5.78)
Comparing with the shape functions obtained for the three-noded triangular element in
section 5.2.1, it is observed that we can very easily write down the required shape functions
in terms of the natural coordinates just described. If we require a higher order transformation,
for example, to model curved domains, we can then use a six-noded triangular element as
shown in Figure 5.16. The natural coordinates of the nodes are indicated in the figure.
Following our earlier procedure, the shape functions can be obtained as
N1 = (SF) (Equation of line 253) (Equation of line 46)
= (SF) (L1) (2L1 1) (5.79)
(0,0,1)
3
1 1
5 0, ,
2 2
1 1
,0, 6
2 2 2 (0,1,0)
4 1 , 1 ,0
1
2 2
(1,0,0)
O X
of the equations of lines 253 and 46 has been taken to ensure that N1 is zero at all these
nodes. Equation of line 253 is L1 = 0 and that of line 46 is L1 = 1/2, i.e., 2L1 1 = 0.
At node 1, L1 = 1 and, therefore, SF = 1. Thus the desired shape function is readily obtained
as
N1 = (L1) (2L1 1) (5.80)
Similarly, other shape functions can be obtained as
N2 = (L2) (2L2 1), N3 = (L3)(2L3 1)

(5.81)
N4 = 4L1L2, N5 = 4L2L3, N6 = 4L3L1
2-d Elements for Structural Mechanics 167
Using these shape functions, we can write
x1

y1

x2

y2

x
3
x N1 0 N2 0 N3 0 N4 0 N5 0 N6 0 y3
= (5.82)
y 0 N1 0 N2 0 N3 0 N4 0 N5 0 N 6 x4

y4

x5

y5

x
6
y
6
Comparing with our discussion in section 5.2.3, we readily observe that the natural
coordinates as defined here for the triangular coordinates significantly simplify the derivation
of shape functions. We will now discuss the formulation of the element level equations for these
elements, viz., three- and six-noded triangles, four- and eight-noded quadrilaterals, etc. for both
structural mechanics problems and fluid flow problems.
5.4 2-d Elements for Structural Mechanics

5.4.1 Generic Relations
We will now summarise certain generic relations which will be useful for all the 2-d finite
elements for structural mechanics. For a two-dimensional structural problem, each point on the
structure may have two independent displacements, viz., u and v, along the two Cartesian
coordinates X and Y, respectively. The strain-displacement relations are given by
u
Fx = (5.83)
x
v
Fy = (5.84)
y
u v
H xy = + (5.85)
y x
For axisymmetric problems, we can consider the r-z plane (analogous to x-y plane for
plane elasticity) as shown in Figure 5.17. We observe that, even though the deformation is
Axis of
symmetry
Axis of
symmetry
r
i
Fig. 5.17 An axisymmetric ring element.
axisymmetric, there can be four independent, nonzero strains (as a change in radius
automatically leads to circumferential strain e0):
u
Fr = (5.86)
r
u
FR = (5.87)
r
v
Fz = (5.88)
z
u v
H rz = + (5.89)
z r
In our standard finite element notation, we write the strain-displacement relations as
{e} = [B]{d }e (5.90)
where the size of [B] and {d } will be dictated by the element and [B] will contain the
derivatives of the shape functions. The stress-strain relations for plane stress specialised from
the 3-d Hookes law, are given by
T x 1 O 0 Fx

E
T y = O 1 0 Fy {F}0 + {T }0
1 O2 (5.91)
0 0 1 O 2 /2 H xy
U xy
where {e}0 and {s}0 refer to the initial strain and stress, respectively.
2-d Elements for Structural Machanics 169
The stress-strain relations for plane strain are given by
T x 1 O O 0 Fx

E
T y = O 1 O 0 Fy {F}0 + {T }0 (5.92)
(1 + O )(1 2O )
U xy 0 0 1 2O /2 H xy

For axisymmetric problems, we have four independent nonzero stresses given by
T r 1 O /1 O O /1 O 0 Fr

T R (1 O ) E 1 O /1 O 0 FR
= {F} + {T }0
0
T z (1 + O )(1 2O ) 1 0 Fz

U rz Symmetric 1 2O /2(1 O ) H rz

(5.93)
In our standard finite element notation, we write the generic stress-strain relation as
{s} = [D]({e} {e}0) + {s}0 (5.94)
where the elements of the [D] matrix depend on plane stress/strain/axisymmetric situation and
are taken from Eqs. (5.91)(5.93).
A structure can, in general, be subjected to distributed loading in the form of body
(volume) forces {qv} such as gravity, forces distributed over a surface {qs} such as a pressure
or concentrated forces lumped at certain points {Pi}. It may also be subjected to initial strains
and stresses (such as those due to preloading and thermal loading). Under the action of all these
loads, the structure, when properly supported (so as to prevent rigid body motion), undergoes
deformation and stores internal strain energy. The strain energy expression can be written as
U =
1
2 {F}v
T
{T } dv =
1
2 {F}
v
T
([D] ( {F} {F} ) + {T }) dv
0 0
1 (5.95)
= O {F } [ D ] {F }
T
{F}T [ D] {F }0 + {F}T {T 0} dv
2
The derivation of the above expression follows closely our discussion in Section 4.1
(Eq. (4.3)) except that we introduce the matrix notation to account for two-dimensional state
of stress/strain.
From Eqs. (5.90) and (5.95), element level strain energy is given as
1 eT eT
2 {E } [ B]T [ D][ B] {E } dv {E }
e
Ue = [ B]T [ D ] {F}0 d v
eT
+ {E } [ B]T {T }0 d v (5.96)
The potential of external forces is given by
Ve
v
Ê `T ^qv ` dv
s
Ê `T ^qs ` ds Ê i `
T
^Pi ` (5.97)
where {d}i is the displacement of the point i on which a concentrated force Pi is acting and
the summation is taken over all such points. It is observed that these displacements {d } and
forces {q} have both X- and Y-components. Hence, typically at a point,
u
{E } = (5.98)
v
q x
{q} = (5.99)
q y
From Eq. (5.97), using the notation of Eq. (5.14) and (5.21), we get
eT eT eT
Ve = v {E } [ N ]T {qv }dv {E }
S
[ N ]T {QS } ds {E } [ N ]Ti {Pi } (5.100)
where [N]i is the value of shape functions at the location i. Thus the total potential of an
element can be written as
1 eT eT
2 {E } [ B]T [ D][ B] {E } dv {E }
e
ep = [ B]T [ D]{F}0 dv
T T T
(5.101)
eT
{E } v {E } S {E } [ N ]T {QS } ds {E } [ N ]Ti {Pi }
e e e
+ [ B]T {T }0 dv [ N ]T {qv } dv
Since {d}e is a vector of nodal d.o.f., it can be taken outside the integration. Thus we get
([B] [ D][ B] dv ){E }

1 T T T
ep = {E }e T e
{E }e [ B] [ D]{F} dv
T 0
+ {E }e [ B] {T } dv
T 0
2 v v v
eT eT eT
{E } [ N ] {q } dv
T
v {E } [ N ] {q }ds {E }
T
s [ N ]Ti {Pi } (5.102)
v s
Defining the element stiffness matrix [k]e and the load vector {f}e as
[k ]e = [ B] [ D][ B] dv
T
v
(5.103)
{ f }e = v [ B]T [ D]{F}0 dv [B] {T} dv

T 0
+
v
[ N ]T {qv }dv + s
[N ]T {qs } ds + [ N ]Ti {Pi } (5.104)
v
we can rewrite Eq. (5.102) as
1
ep = {E }eT [k ]e {E }e {E }eT { f }e (5.105)
2
Once all the element matrices are assembled together, we obtain the total potential of
the system as
1 T
p = ep = {E } [ K ]{E } {E }T {F } (5.106)
2
where the global stiffness matrix of the structure [K] and the global load vector {F} are given
by
NOELEM
[K ] =
n =1
[k ]e (5.107)
NOELEM
{F } = 1
{ f }e (5.108)
NOELEM refers to the number of elements and {d } contains all the nodal d.o.f. variables for
the entire finite element mesh. The summations indicating assembly imply that the individual
element matrices have been appropriately placed in the global matrices following the standard
procedure of assembly.
Using the Principle of stationary total potential we set the total potential stationary with
respect to small variations in the nodal d.o.f., i.e.*
3 p
= 0 (5.109)
{E }T
Thus we have the system level equations given by

[ K ]{E } = {F } (5.110)
We will now discuss each individual 2-d element and describe how the element stiffness
matrix and nodal load vectors can be obtained.
5.4.2 Three-noded Triangular Element

This is one of the simplest and historically one of the first finite elements to be formulated. A
typical element is shown in Figure 5.18. As discussed in section 5.2.1, the shape functions
are given by (from Eq. (5.11)).
1
N1 = (B1 + C1 x + H 1 y) (5.111)
2'
1
N2 = (B 2 + C2 x + H 2 y) (5.112)
2'
1
N3 = (B 3 + C3 x + H 3 y) (5.113)
2'
Since it has been possible to explicitly derive the shape functions in physical Cartesian
coordinates, we shall use these directly rather than using the natural coordinates.
* Though this equation is written in compact matrix notation, it is to be observed that it simply states that
the total potential is stationary with respect to small variations in each and every nodal d.o.f.
Y
v3
(x3, y3) 3 u3
v2
u2
v1 2 (x2, y2)
(x1, y1) 1 u1
O X
Fig. 5.18 Simple three-noded triangular element.
Strains and stresses

The independent nonzero strains for plane stress/strain are given by
u N1 N 2 N 3 u1
0 0 0
1
v
Fx x x x x

v N1 N 2 N3 u2

Fy = = 0 0 0 (5.114)
y y y y v2
H xy
u v N1 N1 N 2 N 2 N 3 N3 u3
+
y x y x y x y x v
3
Thus the strain-displacement relation matrix [B] is given by
N1 N 2 N 3
0 0 0
x x x

N1 N 2 N3
[ B] = 0 0 0 (5.115)
y y y

N1 N1 N 2 N 2 N 3 N3

y x y x y x
Thus
C1 0 C2 0 C3 0
1
[B] = 0 H1 0 H 2 0 H3 (5.116)
2'
H 1 C1 H 2 C2 H3 C3
We observe that the displacement field being linear, the strains are constant (elements of
[B] are all constants) within an element and thus this element is known as a Constant Strain
Triangle (CST). This may be considered the two-dimensional analogue of the simple linear truss
element, and we expect that it will not be able to accurately model complex problems involving
significant strain/stress gradients. The stress-strain relation matrix [D] is given in Eq. (5.91) for
plane stress and in Eq. (5.92) for plane strain cases.
Element matrices
Following Eq. (5.103), the element stiffness matrix is given by
[k ]e = [ B]
T
v
63 [ D ]33 [ B]36 dv (5.117)
The size of the stiffness matrix is (6 6) and, even after exploiting the symmetry of the
stiffness matrix, we still need to evaluate 21 integrals! Since the arguments are all constants,
the integration can be readily carried out and the element stiffness matrix obtained as
[k]e = [B]T[D][B](t)(A) (5.118)
where t is the uniform thickness and A the area of the element.
For more complex, higher order elements, the elements of [B] matrix will vary from point
to point within an element, and explicit integration of expressions in the above equation becomes
very tedious. Also, the size of the matrix itself will be larger and we need to evaluate many
integrals. Thus we look for an efficient way of computing these integrals numerically within
a computer program, which is readily accomplished using natural coordinates.
External loads can be concentrated forces acting at the nodes along global X and/or Y
axes, and they may be distributed over the element edges as shown in Figure 5.19. Two
schemes of finding equivalent nodal loads exist. In the first scheme, we generate a lumped
nodal force vector by appropriately lumping the total force. In the second scheme, we compute
the consistent nodal force vector which does the same amount of work as the distributed force.
The element nodal force vector { f }e given in Eq. (5.104) is a consistent nodal force vector.
We will illustrate the use of this element with a simple example.
Example 5.1. Illustration of the use of CST element. We wish to determine the deflection
of a thin plate subjected to extensional loads as shown in Figure 5.20. We will use two CST
elements to solve this problem. From the coordinates of the nodes, we first compute ai, b i,
gi, etc. as follows:
Fy
Fx
Fy
Concentrated forces
Distributed at nodes
forces Fx
O X
Fig. 5.19 Forces on a CST element.
2 4
3 cm 1 N/cm
1 3
4 cm
Fig. 5.20 Thin plate under uniform tension (Example 5.1).
Element 1: Shape function coefficients

ai = 12, bi = 0, gi = 4
aj = 0, bj = 3, gj = 0 (5.119)
ak = 0, bk = 3, gk = 4
Element 2: Shape function coefficients

ai = 12, bi = 3, gi = 0
aj = 0, bj = 3, gj = 4 (5.120)
ak = 0, bk = 0, gk = 4
We observe that the area of both the elements is 6 cm2 and we let the thickness to be
unity for plane stress. The strain-displacement matrix [B] is now obtained as given below.
Element 1: Strain-displacement matrix
0.25 0 0.25 0 0 0

[ B](1) = 0 0.333 0 0 0 0.333 (5.121)

0.333 0.25 0 0.25 0.333 0
Element 2: Strain-displacement matrix
0.25 0 0.25 0 0 0

[ B](2) = 0 0.333 0 0 0 0.333 (5.122)

0.333 0.25 0 0.25 0.333 0
Assuming E = 2 107 N/cm2 and n = 0.3 for steel, we have the stress-strain relation
matrix [D] as
0.22 108 0.659 107 0

[ D ] = 0.659 107 0.22 108 0 (5.123)

0 0 7
0.769 10

The element stiffness matrix for each element is next obtained as
[k]e = [B]T[D][B](t)(A) (5.124)
The numerical values of [k]e for the two elements are obtained as given below.
Element 1: Stiffness matrix. The element stiffness matrix is given by
1.34

0.714 1.75 Symmetric

0.824 0.33 0.824
[k ](1) = 107
(5.125)
0.385 0.288 0 0.288

0.513 0.385 0 0.385 0.513

0.330 1.47 1.47
0.33 0 0
1 2 3 4 5 6 [Local]
1 2 5 6 3 4 [Global]
Element 2: Stiffness matrix. The element stiffness matrix is given by
1.34


0.824 0.33 0.824
[k ]( ) = 107
2
0.385 0.288 0 0.288
(5.126)
0.513 0.385 0 0.385 0.513

0.330 1.47 1.47
0.33 0 0
1 2 3 4 5 6 [Local]
7 8 3 4 5 6 [Global]
Since the two elements are essentially identical, we observe that our stiffness matrices
are also identical.
For ease of assembly, the local and global destinations of the elements of the stiffness
matrices have also been indicated above. Following the standard procedure of assembly and
incorporating the boundary conditions that u1 = 0 = v1 = u2, we get the final set of equations
as follows:
1.75 v2 0

0.714 1.34 Symmetric u3 1.5

107 0 0 1.75 3
v = 0 (5.127)

0.385 0.513 0.33 1.34 u4 1.5

1.75
0.288 0.385 1.47 0.714 v4 0
The RHS force vector is equivalent nodal forces for the distributed force on the edge
34, viz., half the total force at each node in the X-direction. Solving these equations, we get
the nodal deflections as
v2 = 4.5 108 cm
u3 = 2 107 cm
v3 = 0 (5.128)
u4 = 2 107 cm
v4 = 4.5 108 cm
From the nodal deflections, we can compute element level forces as described now:
Element 1: Forces
1.34


0.824 0.33 0.824
{ f }(1) = [k](1){d}(1) = 107
0.385 0.288 0 0.288

0.513 0.385 0 0.385 0.513

0.330 1.47 1.47
0.33 0 0
0
1.5
0
0
2 107
1.5
= (5.129)
0 0

0 0

4.5 108 0
Element 2: Forces
1.34


0.824 0.33 0.824
{ f}(2) = [k](2){d }(2) = 107
0.385 0.288 0 0.288

0.513 0.385 0 0.385 0.513

0.330 1.47 1.47
0.33 0 0
2 107
1.5
4.5 108
0

0 1.5 (5.130)
=
4.5 108 0

2 107 0

0 0
These forces are along the d.o.f. From the nodal deflections, the strains in elements 1 and
2 can be computed as follows:
{e}e = [B]e{d }e (5.131)
Element 1: Strains
F x 0.25 0 0.25 0 0 0

{F }
(1)
= FY = 0 0.333 0 0 0 0.333

H xy 0.333 0.25 0 0.25 0.333 0
0

0
5 108
2 107

=

1.5 10 8
(5.132)
0
0
0

4.5 108
Element 2: Strains
F x 0.25 0 0.25 0 0 0

{F }(2) = F Y = 0 0.333 0 0 0 0.333

H xy 0.333 0.25 0 0.25 0.333 0
2 107

4.5 108
5 108
0
8
= 1.5 10 (5.133)
4.5 108
0

2 107

4.5 108
It is observed that our solution tallies perfectly with the expected solution, viz.,
d = (PL/AE) and ex and ey are in the ratio of n. The element stresses can be readily obtained
from these strains.
While we can get excellent results for this simple problem with the CST element, in
2-D Elements for Structural Mechanics 179
general, we cannot expect such a good performance because it is only a constant strain
element. In problems involving significant gradients of stress (e.g., stress concentration near
a hole or any sudden geometry change), the CST element will not give good results.
5.4.3 Four-noded Rectangular Element

This is the simplest element of the family of general quadrilateral elements and is depicted in
Figure 5.21. It has four nodes, straight sides, and each node admits two d.o.f., viz., u and v
translations along global the X and Y axes. Our formulation of the element follows standard
procedure and is described now.
4 3
y
h
1 x 2
l
O X
Fig. 5.21 A four-noded rectangular element.
(i) Displacement field and shape functions

From Eq. (5.20), we have the following shape functions:
x y xy
N1 = 1 +
A h Ah
x xy
N2 = (5.134)
A Ah
xy
N3 =
Ah
y xy
N4 =
h Ah
(ii) Strains and stresses

The independent nonzero strains for plane stress/strain are given by
u1

u N1 N 2 N 3 N 4 v1

0 0 0 0
Fx x x x x x u
2

v N1 N 2 N 3 N 4 v2
{F}e = Fy = = 0 0 0 0
y y y y y u3
H xy
u + v N1 N1 N 2 N 2 N 3 N 3 N 4 N 4 v3
y x y x y x y x y

x 38 u
4

v4 81
(5.135)
Thus the strain-displacement relation matrix [B] is given by
y 1 1 y y y
0 A Ah 0 0 0
Ah A
Ah

Ah

x 1 x x 1 x
[B] = 0 Ah h 0 Ah 0 Ah 0
h Ah (5.136)

x 1 y 1 x 1 y
x y 1 x
y

A Ah Ah Ah h Ah Ah
Ah h Ah A Ah
We observe that on x = constant line, the strain varies linearly with y and vice versa. This
element is known as a bilinear rectangular element since its shape functions involve linear
polynomials in both the x and y directions. The stress-strain relation matrix [D] is given in
Eq. (5.91) for plane stress and in Eq. (5.92) for plane strain.
(iii) Element matrices

The element stiffness matrix, following Eq. (5.103), is given by
[k]e = [B]T[D][B] dv (5.137)
The size of the stiffness matrix is (8 8) since there are four nodes and two d.o.f. per node
and even after exploiting the symmetry of the stiffness matrix, we still need to evaluate 36
integrals! Also, the elements of [B] matrix now vary from point to point within an element and
explicit integration of expressions in the above equation becomes very tedious.
External loads can be concentrated forces acting at the nodes along the global X and/or
Y axes (or) they may be distributed over the element edges as shown in Figure 5.22. Two
schemes of finding equivalent nodal loads exist. In the first scheme, we generate a lumped
nodal force vector by appropriately lumping the total force. In the second scheme, we compute
the consistent nodal force vector which does the same amount of work as the distributed force.
The element nodal force vector {f}e given in Eq. (5.104) is a consistent nodal force vector.
Fy
Fy
2 Fx 4
Fx
Fy Fy
Fx
Fx
1 3
O X
Fig. 5.22 Forces on RECT4 element.
5.4.4 Compatibility of Displacements

In the one-dimensional elements described in Chapter 4, elements shared only common nodes
and the nodal d.o.f. were such that across elements, displacements were compatible, i.e., in
a deformed position, there would be no openings or kinks. However, in the two-dimensional
case, the elements share common edges and we need to ensure that the deformation of a point
lying on such a common edge would be the same whether it is assumed to belong to one
element or the other. Let us consider the bilinear rectangular element just described. Consider
two adjacent elements as shown in Figure 5.23. We see that they share common nodes and
displacements are therefore the same at these nodal points for both elements. Now consider
1 2
2 3
P
y
x
1 2
Fig. 5.23 Continuity of displacement across two elements.
an intermediate point P on edge 23 whose deformation can in general be interpolated from

nodal deflections using the shape functions. On edge 23, we can substitute for x = l and have
y y
N1 = 0, N2 = 1 , N3 = 1 , N4 = 0 (5.138)
h h
u1

v1

u2

u P N1 0 N2 0 N3 0 N4 0 v2 N 2 u2 + N 3u3
= = (5.139)
vP 0 N1 0 N2 0 N3 0 N 4 u3 N 2 v2 + N 3v3

v3

u4

v4
Thus we see that uP and vP, the deflections at the interior point P, are functions of only
the nodes on that edge, viz., 2 3. Since nodal deflections are common for both the elements,
it can be concluded that the deflection of an interior point on the edge will also be the same
for both the elements. Thus we have compatibility of displacements along the entire edge. One
of the major limitations of the bilinear element as formulated above is that it can only be
rectangular in shape. We will now illustrate that once this element becomes a general
quadrilateral, we cannot guarantee compatibility of displacements across elements. Let us
consider a four-node quadrilateral element as shown in Figure 5.24 and let the displacements
be given as
u (x, y) = c0 + c1x + c2y + c3xy (5.140)
v(x, y) = c4 + c5x + c6y + c7xy (5.141)
Since these expressions must reduce to nodal displacements when we substitute nodal
coordinates, we get the following equations for u-displacement:
u1 = c0
u2 = c0 + c1a
u3 = c0 + 2ac1 + bc2 + 2abc3 (5.142)
u4 = c0 + c2b
Solving for the coefficients and rearranging the terms, we get
u(x, y) = N1u1 + N2u2 + N3u3 + N4u4 (5.143)

(0, b) 4 3 (2a, b)
(0, 0) 1 X
2 (a, 0)
Fig. 5.24 A quadrilateral element.
where
x y xy x xy
N1 = 1 + , N2 = (5.144)
a b ab a ab
xy y xy
N3 = , N4 = (5.145)
2ab b 2ab
At node 2, N2 = 1 and N1 = N3 = N4 = 0. Similarly, at node 3, N3 = 1, N1 = N2 = N4
y x
= 0. However, for any intermediate point on the edge 23, = 1 we see that the shape
b a
functions N1 and N4 do not vanish. So, the displacement of a point on edge 23, in general,
is given as
u = N1u1 + N2u2 + N3u3 + N4u4 (5.146)
Since this is not dependent only on u2 and u3 (the common nodes between two adjacent
elements sharing this edge), we cannot guarantee that on edge 23 every interior point will have
the same u-displacement for both the elements. While some incompatible finite elements are
used in practice, we will limit our discussion in this text only to those elements where
displacements are compatible across inter-element boundaries. We will now formulate the
four-node quadrilateral element using the natural coordinates and show that it is possible to
achieve full compatibility.
5.4.5 Four-node Quadrilateral Element

Consider the parent four-node square element in xh frame as well as the general quadrilateral
element in the physical x-y space as shown in Figure 5.25. The coordinates of any point P(x, y)
are interpolated from nodal coordinates as follows:
4
x = N x
i =1
i i (5.147)
4
y = N y
i =1
i i (5.148)
where Ni are as given in Eqs. (5.43)(5.46).

Y
h
(1, 1) 3 3
(x3, y3)
4 (1, 1)
4
(x4, y4)
x
O (x2, y2)
2
1
1 2 (x1, y1)
(1, 1) (1, 1)
O X
Fig. 5.25 A general four-noded quadrilateral element.
The displacements of the interior point P are also interpolated from nodal deflections using
the same shape functions for this element, which is therefore called an isoparametric element.
The displacements can be written as
u1

v1

u2

N1 0
u 0 N2 0 N3 0 N4 v2
= (5.149)
v 0 N1 0 N2 0 N3 0 N 4 u3

v3

u4

v4
Let us now show that the displacements in this element are compatible across inter-
element boundaries. Along edge 23 (x = 1), for example, the shape functions N1 = 0 = N4.
Thus, irrespective of the shape of the element in the physical x-y space (i.e. even for a general
quadrilateral), the displacements of any point on edge 23 are given as
u = N2u2 + N3u3
(5.150)
v = N2v2 + N3v3
Therefore, we observe that even for a general quadrilateral shape, the isoparametric element
guarantees the continuity of displacements across inter-element boundary edges.
Recall our observation that the nodal d.o.f. are in X-Y directions and not along xh. Also
the x-h axes, as transformed into real physical xy space, may not be orthogonal. Thus we
write the strain displacement relations in an orthogonal, Cartesian frame of reference, as
u

F x x

v
{F} = F y = (5.151)
y
H xy
u +
v
y x
We observe that u and v are given in terms of shape functions Ni which are expressed
in the xh coordinates rather than xy. For a general function f, using the chain rule of
differentiation, we obtain the equations
f f Y f I
= +
x Y x I x
f f Y f I
= + (5.152)
y Y y I y
However, from Eqs. (5.147)(5.148), it is easier to write down derivatives of (xy) in
terms of (xh) rather than the other way round. Thus we write
f f x f y
= +
Y x Y y Y
f f x f y
= + (5.153)
I x I y I
Thus, in general, we can write
x y
Y Y Y x

= (5.154)

x y
I y
I I
i.e.

Y x

= [J ]
(5.155)
I y
where
x y
Y Y
[J ] =
(5.156)

x y
I I
From Eqs. (5.147)(5.148),
N i N i
Y xi Y yi
[J ] = (5.157)

N i N i

I
xi I
yi

Here, [J] is known as the Jacobian matrix relating the derivatives in two coordinate
frames. This generic expression for [J] is true for all 2-d elements, but the actual coefficients
in the matrix will depend on the shape functions being employed in a given element and the
nodal coordinates.
For this particular element, for example, the Jacobian is
1 I 1+I 1 I 1+I
( x2 x1 ) + 4 ( x3 x4 ) 4 ( y2 y1 ) + 4 ( y3 y4 )
4
[J ] =
(5.158)
1 Y ( x x ) + 1 + Y ( x x ) 1 Y
( y y ) +
1+ Y
( y y )
4 4 1 4 3

2 4

4 1 4

3 2
From Eq. (5.155),

x
Y 1 J 22 J12 Y
= [J ]1 = (5.159)
J J 21 J11
y I I
We can now obtain the required strain displacement relations.
u
u x

x 1 0 0 0 u
v
y
{F } = = 0 0 0 1 (5.160)
y v
0 1 1 0 x
u + v

y x [B1 ] v
y

However, from Eq. (5.159),

J 22 J
u J 12 0 0 u
x J Y

u J 21
0 u
J11
0

y I
J J

=
J v
v 0 J 22
x 0 12
J J Y
(5.161)
v

J 21 J11 u
y 0 0 I
J J

[B 2 ]
From Eq. (5.149),
1
N 3 u
u N1 N 2 N 4
0 0 0 0
Y
Y Y Y Y 1
v

u u
N1 N 2 N 3 N 4
0
2
0 0 0
I I I I I v2
= (5.162)
v N1 N 2 N 3 N 4 u3
Y 0 0 0 0
Y Y Y Y v3

v
N1 N 2 N 3 N 4 u4
I 0 I
0
I
0
I
0
I v
4
= [B3] {d}e
For this element, for example, the shape function derivatives in the x-h frame are as
follows:
N1 (1 I) ,
N1 (1 I)
= =
Y 4 Y 4
N 2 1 I, N 2 (1 + Y )
= =
Y 4 I 4
(5.163)
N 3 1 + I N 3 (1 + Y )
= , =
Y 4 I 4
N 4 (1 + I) N 4 (1 Y )
= , =
Y 4 I 4
Thus, finally, combining Eqs. (5.160)(5.162), the strain-displacement relations can be written
as
F x

{F } = F y = [ B ]{E }
e
(5.164)
F xy
where [B] = [B1][B2][B3].
The stress-strain relation matrix [D] remains as given earlier (see, e.g. Eq. (5.91)). Thus
we have the element stiffness matrix
[k ]8e8 = v
[ B]T [ D][ B] dv = [B] [D][B] t dx dy
T
1 1 (5.165)
=
1 1
[ B]T [ D][ B]t J dY dI
where the coordinate transformation has been taken into account and the integrals need now
be evaluated over the parent square element in natural coordinates. Since [B] in general varies
from point to point within the element, we look for ways of evaluating these integrals
numerically within a computer program rather than attempting to derive them explicitly. We will
discuss numerical integration schemes in Section 5.5.
5.4.6 Eight-node Quadrilateral Element

Consider the parent eight-node rectangular element in x- h frame as well as the general
quadrilateral element in the physical x-y space as shown in Figure 5.26. The coordinates of any
point P(x, y) are interpolated from nodal coordinates as follows:
x = N x , i i y = N y i i (5.166)
h (x7, y7) 7 3 (x3, y3)

7 (0, 1)
(1, 1) 4 3 (1, 1) (x4, y4) 6 (x6, y6)
4
(1, 0) 8 6 (1,0)
(1, 0)
x (x8, y8) 8
2 (x2, y2)
(x5, y5)
1 5
2 (1, 1) (x1, y1)
(1, 1) 1 5(0,-1)
(0, 1)
O X
Fig. 5.26 A general eight-noded quadrilateral element.
where we may take coordinates of just the four vertices and follow a simple linear interpolation
for coordinate transformation (or) use the full eight nodes to permit curved edges. The
displacements of the interior point P are also interpolated from nodal deflections using the same
shape functions. Three possibilities arise as shown below:
Coordinate Interpolation Displacement Interpolation
Linear (i.e. four vertex nodes) Quadratic (i.e. all eight nodes)
Quadratic Quadratic
Quadratic Linear
The first element belongs to the category of sub-parametric elements and is useful when
the structural geometry is simple polygonal, but unknown field variation may involve sharp
variations. The second element is an isoparametric element and permits curved edge modelling
as well as quadratic variations in displacements. If the geometry has certain curved features
but is in a low stress region, we can use the third element belonging to the superparametric
category. The displacements can, in general, be written as
u1

v1
u N1 0 N2 0
= u2 (5.167)
v 2 1 0 N1 0 N2 2 16
v2

# 16 1
where the shape functions are appropriately taken from Eqs. (5.54)(5.61) and sizes of the
matrices shown are valid for the sub- and isoparametric elements.
The corresponding strain-displacement relation, following Eq. (5.164), is written as
{F}31 = [ B1 ]34 [ B2 ]44 [ B3 ]416 {E }16 1

e

(5.168)
= [B ]316 {E }16e 1
For the eight-noded isoparametric element, for example, the Jacobian can be obtained as
8 N i
8
N i

i =1 Y
xi Y
i =1
yi

[J ] = (5.169)
8 N i
8
N i

i =1 I
xi i =1
yi
I
The necessary shape function derivatives in the x-h frame are summarised now from
Eqs. (5.54) (5.61):
N1 (1 Y )(Y + 2I )
N1 (1 I )(I + 2Y ) , =
= I 4
Y 4
N 2 (1 I )( 2Y I ) , N 2 (1 + Y )(2I Y )
= =
Y 4 I 4
N 3 (1 + I )( 2Y + I ) , N 3 (1 + Y )(2I + Y )
= =
Y 4 I 4
N 4 (1 + I )( 2Y I ) , N 4 (1 Y )(2I Y )
= =
Y 4 I 4
N 5 N5 (1 Y 2 )
= Y (1 I ) , = (5.170)
Y I 2
N 6
N 6 1 I2 , = I (1 + Y )
= I
Y 2
N 7 N 7 (1 Y 2 )
= Y (1 + I ) , =
Y I 2
N8 (1 I2 ) , N 8
= = I (1 Y )
Y 2 I
We can write the expression for the element stiffness matrix as
[k ]e = v
[ B]T [ D][ B] dv
= [ B] [D][B] t dx dy
T
(5.171)
1 1
=
1 1
[ B]T [ D][ B]t J d Y dI
which can be evaluated by using suitable numerical integration schemes.
5.4.7 Nine-node Quadrilateral Element

This element, as shown in Figure 5.27, is essentially similar to the eight-node element discussed
in section 5.4.6 except that we have added one more node in the centre (x = h = 0) of the
element. The shape functions for the eight-node element were derived based on the serendipity
approach. The basic idea behind adding the centre node here is that we can use the product
of one-dimensional Lagrange polynomials. For node 1, the quadratic, one-dimensional Lagrange
polynomial in x can be written as
(Y Y5 )(Y Y2 ) (Y 0)(Y 1) 1
L1 (Y ) = = = Y (Y 1) (5.172)
(Y1 Y5 )(Y1 Y2 ) ( 1 0)( 1 1) 2
h (x7, y7) 7 3 (x3, y3)

7 (0, 1)
(1, 1) 4 3 (1, 1) (x4, y4) 6 (x6, y6)
4
9(x9, y9)
(1, 0) 8 9 (0, 0) 6 (1,0)
(1, 0)
x (x8, y8) 8
2 (x2, y2)
(x5, y5)
1 5
2 (1, 1) (x1, y1)
(1, 1) 1 5(0,-1)
(0, 1)
O X
Fig. 5.27 A general nine-noded quadrilateral element.
and in h it can be written as
(I I8 )(I I4 ) (I 0)(I 1) 1
L1 (I) = = = I (I 1) (5.173)
(I1 I8 )(I1 I4 ) ( 1 0)( 1 1) 2
The complete shape function for node 1 can now be obtained as product of the above
polynomials, in the form
1 YI
N1 = L1 (Y ) L1 (I ) = YI (Y 1) (I 1) = (1 Y )(1 I ) (5.174)
4 4
The other shape functions can similarly be obtained as
YI YI YI
N2 = (1 + Y )(1 I), N3 = (1 + Y )(1 + I), N4 = (1 + I)(1 Y )
4 4 4
I Y I
N5 = (1 I)(1 Y 2 ), N6 = (1 + Y )(1 I2 ), N7 = (1 + I)(1 Y 2 )
2 2 2
Y
N8 = (1 Y )(1 I2 ), N 9 = (1 Y 2 )(1 I2 ) (5.175)
2
The Jacobian matrix of coordinate transformation is given by
9 N i 9
N i

i =1 Y
xi Y
i =1
yi

[J ] = (5.176)
9 N 9
N i

i =1 I
i
xi
i =1
yi
I
The strain-displacement matrix can be written as
{F }31 = [ B1 ]34 [ B2 ]44 [ B3 ]418 {E }181 =

e
[B ]318 {E }18e 1 (5.177)
Using the standard stress-strain matrix, we can write the element stiffness matrix as
[k ]e = [B] [D][B] dv = [B] [D][B]t dx dy

v
T T
1 1 (5.178)
=
1 1
[ B]T [D][ B]t J d Y dI
The integrals can be evaluated using a suitable numerical integration scheme. The basic
difference between the eight-node serendipity and the nine-node Lagrange element comes in the
shape functions because of the addition of a node. It is observed that the serendipity shape
functions involves the following eight polynomial terms:
1, x, h, x 2, xh, h2, x 2h, xh 2 (5.179)
The Lagrange shape functions for the nine-node element include an additional ninth
polynomial term, viz.
x 2h 2 (5.180)
Thus, both the elements adopt more terms than necessary for the complete quadratic
polynomial in x h, viz., (1, x, h, x 2, xh, h 2). In general, serendipity elements have fewer
extra terms. These two elements are commonly used in two-dimensional analysis and have
certain crucial variations in their behaviour, a detailed discussion of which is beyond the scope
of the present discussion.
5.4.8 Six-node Triangular Element

A typical six-node triangular element is shown in Figure 5.28. The quadratic shape functions
in natural coordinates are obtained from Eqs. (5.80) and (5.81) as follows:
N1 = L1(2L1 1), N2 = L2(2L2 1), N3 = L3(2L3 1)
(5.181)
N4 = 4L1L2, N5 = 4L2L3, N6 = 4L3L1
We recall that L1, L2, L3 are not independent:
L1 + L2 + L3 = 1
Hence,
L 3 = 1 L1 L 2 (5.182)
So, in terms of L1, L2, we rewrite the shape functions as
N3 = (1 L1 L2)(1 2L1 2L2), N5 = 4L2(1 L1 L2), N6 = 4L1 (1 L1 L2)

(5.183)
3
5
6
2
4
1
O X
The derivatives of the shape function are given by
N1 N1
= 4 L1 1, = 0
L1 L2
N 2 N 2
= 0, = 4 L2 1
L1 L2
N 3 N 3
= 3 + 4 L1 + 4 L2 , = 3 + 4 L1 + 4 L2
L1 L2
N 4 N 4 (5.184)
= 4 L2 , = 4 L1
L1 L2
N 5 N 5
= 4 L2 , = 4 4 L1 8 L2
L1 L2
N 6 N 6
= 4 4 L2 8L1 , = 4 L1
L1 L2
The Jacobian for the isoparametric element is given by
6 N i
6
N i

i =1 L1
xi L
i =1 1
yi

[J ] = (5.185)
6 N i
6
N i

i =1 L2
xi
i =1
yi
L2
The strain-displacement relation [B] can be written as
{F }31 = [B ]312 {E }12e 1 (5.186)

where
[B ]312 = [B1 ]34 [B2 ]44 [B3 ]412 (5.187)
where [B1] and [B2] are the same as those given in Eqs. (5.160)(5.161), and [B3] is obtained
as
N1 N 2 N 3 N 4 N 5 N 6
L 0 0 0 0 0 0
L1 L1 L1 L1 L1
1

N1 0
N 2
0
N 3
0
N 4
0
N 5
0
N 6
0
L2 L2 L2 L2 L2 L2
[B3 ]312 =
N1 N 2 N 3 N 4 N 5 N 6
0 0 0 0 0 0
L1 L1 L1 L1 L1 L1

N1 N 2 N 3 N 4 N 5 N 6
0 L2
0
L2
0
L2
0
L2
0
L2
0
L2 412

(5.188)
The element stiffness matrix can be expressed as
1 1 L1
[k ]e =
v
[ B]T [ D ][ B] dv =
0 0
[ B]T [ D ][ B ]t J dL1 dL2 (5.189)
where again, the evaluation of the integrals can be done using suitable integration rule in
triangular coordinates. We observe that L1, L2 are similar to x, h except that range of L1 is
(0, 1) rather than (1, 1) and, in view of the triangular domain, L2 varies between 0 and
(1 L1).
We have discussed the formulation of several two-dimensional elements for structural
analysis purpose. Using natural coordinates and isoparametric formulation, the procedure for
derivation of element characteristic matrices is fairly systematic and amenable to
computerisation. While we have discussed linear and quadratic elements that are most
commonly used in practice, we can extend the same procedure to formulate cubic or quartic
or other higher order elements. Explicit derivation of element level matrices is quite tedious and
we look for appropriate schemes for numerical integration, which will now be discussed.
5.5 Numerical Integration

We will first discuss numerical integration in one dimension and then extend the scheme to two
dimensions. Consider the evaluation of the following integral:
x2
I = x1
f ( x) dx (5.190)
Numerical Integration 195
The geometrical interpretation of the integration is that we are trying to determine the area under
the curve f (x) between the limits x1 and x2. Let the function be known at discrete points xi
in this interval. Two classical schemes of numerical integration are the Trapezoidal rule and the
Simpsons rule. These are now summarised.
5.5.1 Trapezoidal Rule

Let a discrete set of uniformly spaced points xi and the corresponding values of the function
fi be given. In this scheme, we assume that the function is varying linearly within each segment
as shown in Figure 5.29. We know that the area of the trapezium ABCD is given by
f (x)
C
fi
f i+1
D
f i1
A B
xi1 xi xi+1 X
d d
Fig. 5.29 Trapezoidal rule for numerical integration.
f i + f i 1
Area ABCD = d (5.191)
2
The area under the curve is then approximated as the sum of the areas of the trapezia
as indicated in the figure. Thus we have
n
fi + fi 1 f + f
f ( x) dx i =1
2
d = d 1 n + ( f2 + f3 + " + fn1 )
2
(5.192)
For later use, we now rewrite this formula as

n
d d
f ( x) dx f1 + (d ) f 2 + (d ) f 2 + " + (d ) f n1 + f n =
2 2
W f
i =1
i i (5.193)
where Wi are the weights corresponding to fi.

5.5.2 Simpsons 1/3 Rule

In this scheme, instead of linear approximation, we use a parabolic approximation for the
function within each segment as shown in Figure 5.30. We can find the area under the curve
in one such segment as follows:
f (x)
Q
B
fi
P
C
f i1
A
O M
X
xi1 xi xi+1
d d
Fig. 5.30 Simpsons one-third rule for numerical integration.
Area under the curve ABC = Area OACM + Area ABCA (5.194)
f + fi +1
Area OACM = i 1 (2 d ) = d ( f i 1 + f i +1 ) (5.195)
2
2 2 f i 1 + f i +1
Area ABCA = (area APBQCA) = fi (2d ) (5.196)
3 3 2
Thus
d
Area of one segment = ( fi1 + 4fi + fi+1) (5.197)
3
n
3 (fi1 + 4fi + fi+1)

d
Total area under the curve =
i=2
d 5d 6d 5d d
= f1 + f 2 + 3 f3 + " + 3 f n 1 + 3 f n
3 3
(5.198)
Thus our procedure for numerical integration consists of two simple steps, viz., first do
a curve fit (straight line, parabola, etc.) for the function within each segment and next evaluate
the total area, and express the result as a generic formula of the type
n
f ( x) dx = W f
i =1
i i (5.199)
The numerical integration scheme popularly used in finite element computations is known
as the Gauss quadrature, which is essentially based on similar lines. We will first discuss the
NewtonCotes formula and then present the Gauss quadrature rule, followed by a set of
examples. In view of the intended application in finite element computations using natural
coordinates, we shall take, the limits of integration to be 1 and 1.
5.5.3 NewtonCotes Formula

Let us consider the evaluation of the integral
1
I ( x) = 1
f ( x) dx (5.200)
where we assume that the function values fi at some arbitrarily spaced xi(i = 1, 2, , n) are
known. We then approximate the function f (x) using a polynomial g(x) as
f (x) g(x) = c0 + c1x + c2x2 + L + cn1xn1 (5.201)
Considering that at x = xi, g(xi) = fi, we get equations
c0 + c1 x1 + c2 x12 + " + cn 1 x1n 1 = f1

# # # # # (5.202)
c0 + c1 xn + c2 xn + " + cn 1 xnn 1 =
2
fn
If x1, x2, K, xn are given, then Eqs. (5.202) are simple linear algebraic equations which
can be readily solved using, for example, the Gauss elimination. We thus have n equations in
n unknowns, which can be used to solve for the coefficients c0, c1, K, cn1 in terms of the
function values fi. Assuming that these coefficients have been determined, we can evaluate the
integral as
1 1
I = 1
f ( x) dx 1
g ( x) dx (5.203)
However,
1 1
1
g ( x) dx =
1
(c0 + c1 x + c2 x 2 + " + cn 1 x n 1 ) dx
1
x2 x3 xn
= c0 x + c1 + c2 + " + cn 1
2 3 n 1
2 2
= 2c0 + 0 + c2 + 0 + c4 + " (5.204)
3 5
If we substitute the values of the coefficients in this expression, we obtain a formula of

the type
I W 1 f1 + W 2 f2 + L + W n fn (5.205)
Using this formula, called the NewtonCotes formula, our evaluation of the integral will
be exact if the original function f(x) were any polynomial up to degree n 1. Here we have
assumed that at given locations (sampling points) xi, the function values fi are known. How
do we know which are the good locations for sampling the function? The Gauss quadrature
formula attempts to find the locations of the sampling points xi as well as the weight factors
Wi. We will now discuss the details of this scheme which is commonly used in finite element
computations.
5.5.4 Gauss Quadrature Formula

Let the function be approximated by a polynomial of degree p, which is also to be found, i.e.,
let
f(x) g(x) = c0 + c1x + c2x2 + L + cp x p (5.206)
The evaluation of the integral can be done as follows:
1 1 2 2 cp
I =
1
f ( x) dx 1
g ( x) dx = 2c0 +
3
c2 + c4 + " +
5 p +1
[1 ( 1) p +1 ] (5.207)
As per our generic formula, we attempt to express the integral in the following way:
I W 1 f1 + W 2 f2 + L + W n fn (5.208)
Using Eq. (5.206), we can write
I W1 (c0 + c1 x1 + c2 x12 + " + c p x1p )
+ W2 (c0 + c1 x2 + c2 x22 + " + c p x2p )

+"
(5.209)
+"
+ Wn (c0 + c1 xn + c2 xn2 + " + c p xnp )
Comparing the coefficients of c0, c1, c2, in Eqs. (5.207) and (5.209), we can write
W1 + W2 + " + Wn = 2
W1 x1 + W2 x2 + " + Wn xn = 0
2
W1 x12 + W2 x22 + " + Wn xn2 =
3
# (5.210)
1
W1 x1p + W2 x2p + " + Wn xnp = [1 (1) p +1 ]
p + 1
These equations can be solved for the locations of the sampling points xi(i = 1, 2, , n)
and the corresponding weight factors Wi(i = 1, 2, , n). It is easy to see that our number
of unknowns (viz., 2n) and equations available ( p + 1) match exactly when the polynomial
degree p is such that
p + 1 = 2n
or
p = 2n 1 (5.211)
In other words, using n Gauss sampling points as determined above, we can evaluate an
integral exactly if the integrand function f (x) is a polynomial of degree less than or equal to
(2n 1). Thus, with one-point rule, we can integrate a linear polynomial exactly, while with
two-point rule we can integrate up to cubic polynomial exactly. In practical use, we observe
the degree of the highest terms in the integrand and then decide the order of Gauss quadrature
rule. As compared to the NewtonCotes formula, we observe that, for a given number of
sampling points, the Gauss quadrature rule can exactly integrate much higher degree polynomial.
We now list the sampling points and the weight factors for Gauss quadrature for one-
dimensional case in Table 5.1.
Table 5.1 Sampling Points and Weights for Gauss Quadrature
n Location of sampling point xi Weight factor Wi

1 0 2
1
2 + 1
3
1
1
3
3 + 0.6 5/9
0 8/9
0.6 5/9
4 +0.8611363 0.347854845
+0.3399810 0.652145155
0.3399810 0.652145155
0.8611363 0.347854845
It is easy to observe that the exact numerical values of the sampling points and the weight
factors, obtained as a solution of Eq. (5.210), depend on the integration interval. Thus the
formula is standardised for the interval 1 to 1, keeping in mind the finite element computations
using natural coordinates. If the interval is different, we need to perform an appropriate change
of coordinates such that the limits tally with 1 and 1. We will now illustrate the use of the
formula through the following examples.
Example 5.2. Evaluate the integral
(2 )
1
I = + x + x 2 dx (5.212)
1
It is easy to verify that we need at least a two-point integration rule since the integrand contains
a quadratic term. We will obtain our approximation to the integral with one and two-point Gauss
quadrature rules and show that the two-point rule result matches with the exact solution.
For a one-point rule, from Table 5.1, we have
x1 = 0, W1 = 2
Hence,
I 2(2 + 0 + 0) = 4 (5.213)
For a two-point rule, from Table 5.1 we have
1
x1 = + , W1 = 1
3
1
x2 = , W2 = 1
3
Therefore,
1 1 1 1
I (1) 2 + + + (1) 2 + = 4.6667 (5.214)
3 3 3 3
The exact solution is easily verified to be
(2 x + x 2 + x 3/3)1 = 4.6667
1 1
I =
1
(2 + x + x 2 ) dx = (5.215)
Example 5.3. Evaluate the integral

1 Qx
I =
1
cos
2
dx (5.216)
The exact solution can be readily obtained as follows:

1
Qx Q
I = sin = 1.273239544 (5.217)
2 2 1
For the one-point Gauss rule,
(Q )(0)
I 2 cos = 2 (5.218)
2
For the two-point Gauss rule,
Q 1 Q 1
I 1 cos + 1 cos (5.219)
2 3 2 3
Q
= 2 cos = 1.232381 (5.220)
2 3
For the three-point Gauss rule,

Q Q Q
I
5
9
cos
2

0.6 +

8
9
cos 0 +
2
5
9 2

cos 0.6

( )
= 1.274123754 (5.221)
With a four-point Gauss rule, we get
Q
I (0.347854845) cos (0.8611363)
2
Q
+ (0.652145155 ) cos (0.3399810)
2
Q
+ (0.652145155) cos ( 0.3399810)
2
Q
+ (0.347854845) cos ( 0.8611363)
2
= 1.273229508 (5.222)
Since the integrand is not an exact polynomial, our numerical scheme can never give the
exact solution but approximates fairly well when sufficient number of points are taken.
5.5.4 Gauss Quadrature in Two Dimensions

We shall first discuss two-dimensional rules appropriate for quadrilateral finite elements and
then present integration rules for triangular elements.
For two-dimensional integration, we can write
1 1
I = 1 1
f (Y ,I ) d Y dI

W
1

1

i
i f (Y i , I ) dI

(5.223)

j
Wj

i
Wi f (Y i , I j )

WW
i j
i j f (Y i , I j )
Thus we can obtain the necessary Gauss quadrature rule simply by taking the
product of appropriate one-dimensional rules. Thus, a (2 2) Gauss rule would be as shown
in Figure 5.31.
1 1 1 1 1 1
I (1)(1) f , + (1)(1) f , + (1)(1) f ,
3 3 3 3 3 3
1 1 (5.224)
+ (1)(1) f ,
3 3
1 3
I = 4
3
x
1
I=
3 1 2
1 1
Y= Y =
3 3
Fig. 5.31 Sampling points for a 2 2 Gauss quadrature rule.
Similarly, a (3 3) rule would be as shown in Figure 5.32. Thus,
5 5 8 5 5 5
I 9 f1 + 9 9 f2 + 9 9 f3
9
5 8 8 8 5 8
+ f 4 + f5 + f6
9 9 9 9 9 9 (5.225)
5 5 8 5 5 5
+ 9 f7 + 9 9 f8 + 9 9 f9
9
7 8 9
I = 0 .6
I=0 x
4 5 6
I = 0.6
1 2 3
Y=0
Y = 0.6 Y= 0.6

It is not necessary that we take the same number of points along both x and h directions.
For example, a (2 3) rule would be as shown in Figure 5.33.
5 5 8 8 5 5
I (1) f1 + (1) f 2 + (1) f 3 + (1) f 4 + (1) f5 + (1) f 6 (5.226)
9 9 9 9 9 9
I = 0.6
5 6
I=0 x
3 4
I = 0.6
1 2
1 1
Y= Y =
3 3
The Gauss quadrature rules can be effectively used for finite element computations.
Consider that we need to evaluate the following:
1 1
[k ]e =
1 1
[ B]T [ D ][ B] t J d Y d I (5.227)
Using Gauss quadrature rule we rewrite this equation as
[k ]e = ([B] [ D][ B] t J ) W W
i j
T
i j
(5.228)
where all the terms within the parentheses are evaluated at each (xi, hj). Once we select the
appropriate number of Gauss points needed, based on the highest polynomial term expected in
the integrand, we can readily evaluate the integral. The integrals corresponding to element nodal
force vector are evaluated in a similar manner. Numerical integration for triangular domains
follows similar pattern. However, we do not take the product of Gauss quadrature rule in two
directions. Instead, we write
1 1 L1
I 0 1 f (L1, L2) dL1 dL2 Wi fi (5.229)
The locations of the sampling points and the corresponding weights are given in
Table 5.2 and are depicted in Figure 5.34.
3
(1/3, 1/3, 1/3)
1 2
(a) One-point rule
3 3
(0.2, 0.2, 0.6)
(1/2, 0, 1/2) (0, 1/2, 1/2) (0.6, 0.2, 0.2) (0.2, 0.6, 0.2)
(1/3, 1/3, 1/3)
1 2 1 2
(1/2, 1/2, 0)
(b) Three-point rule (c) Four-point rule
Fig. 5.34 Sampling points for numerical integration over triangular domain (Table 5.2).
Table 5.2 Gauss Quadrature Points and Weights for Triangular Domains
n, No. of points Location of sampling points Weight factor W

1 L1 = 1/3
L2 = 1/3 1
L3 = 1/3
3 (1/2, 1/2, 0) 1/3
(1/2, 0, 1/2) 1/3
(0, 1/2, 1/2) 1/3
4 (1/3, 1/3, 1/3) 27/48
(0.6, 0.2, 0.2) 25/48
(0.2, 0.6, 0.2) 25/48
(0.2, 0.2, 0.6) 25/48
We have so far discussed, in detail, the formulation of element level equations and
evaluation of element characteristic matrices through numerical integration. These element
matrices can be assembled together to form the global system level matrices, in a manner
Incorporation of Boundary Conditions 205
similar to the assembly procedure used for one-dimensional problems. Before we can actually
use these elements to solve some example problems, we need to discuss two more aspects,
viz., incorporating the prescribed boundary conditions and, secondly, the solution of system
level equations. In the sections that follow, we will discuss these two issues.
5.6 Incorporation of Boundary Conditions

Let the system level assembled equations be given as
[ A]nn {x}n1 = {b}n1 (5.230)
In general, a d.o.f. xi may be prescribed to be zero (e.g. a fixed support on a structure) or

nonzero (e.g. a prescribed wall temperature). An effective way of implementing either boundary
condition is now described. Retaining all other coefficients of [A] and {b} the same, we modify
aii and bi as
aii = aii + a p (5.231)
bi = a p xi (5.232)
where ap is called a penalty number and xi is the prescribed value for xi ( xi may or may not
be zero). Thus our method is to simply replace existing bi by ap xi and add ap to the existing
diagonal coefficient aii. If ap >> aij ( j = 1, 2, K, n), this equation will essentially give
xi xi . The larger the value of ap, the closer our enforcement of the constraint xi = xi .
Typically, ap = 103 106 aii. This method is easy to implement in a computer program; it can
take care of prescribed zero/nonzero constraints; does not involve any rearrangement of
coefficients in [A] or {b}. However, a large ap may lead to numerical difficulties in solution.
A simpler alternative to take care of prescribed zero boundary condition is to incorporate it at
the time of assembly itself, i.e., if a d.o.f. is prescribed to be zero, do not assemble that
equation itself (i.e. row and column in the matrices)! A more cumbersome way of implementing
the general nonzero prescribed boundary condition is as follows. Consider three equations in
three unknowns as
a11 x1 + a12 x2 + a13 x3 = b1
a21 x1 + a22 x2 + a23 x3 = b2
a31 x1 + a32 x2 + a33 x3 = b3
and let x1 = x1 . We can rewrite these equations as
x1 = x1
a22 x2 + a23 x3 = b2 a21 x1
a32 x2 + a33 x3 = b3 a31 x1
Thus, our general strategy is to make aii = 1 and all other coefficients in that row and column
to be zero. Also, the RHS {b} is modified as
{(bj }
a ji xi ) , j = 1, 2, !, n (5.233)
The ith element in {b} is simply taken as xi . The key idea of this method is that the ith
equation is made to read simply as xi = xi ; the value xi is plugged in all other equations and
taken to the RHS. The enforcement of the constraint xi = xi is exact and not approximate as
in the penalty method.
5.7 Solution of Static Equilibrium Equations

The general form of the finite element equations in linear static analysis is given by
[ A]nn {x}n1 = {b}n1 (5.234)
In typical finite element models of complex, real-life problems, a large number of elements are
required. Hence the size of the matrices (viz. n) is usually very large. The finite element models
with several thousand d.o.f. are routinely used. Thus, computationally efficient solution
algorithms are a crucial part of a total finite element software package. Two approaches to the
solution of Eq. (5.234) are possibledirect method and the iterative methods. The direct
solution methods are essentially based on traditional Gauss elimination strategy. Among the
iterative methods, the GaussSiedel and the conjugate gradient method (with preconditioning)
are popular. We will discuss here only Gauss elimination based direct methods. It may be
recalled that, in the basic Gauss elimination process, through a series of operations on the
matrices [A] and {b}, we aim to transform Eq. (5.234) into the following form:
a11 a11 ! ! ! a1n

0 a22 ! ! ! a2 n

0 0 a33 ! ! a3n x b
= (5.235)
%

an 1,n 1 an 1,n

0 ann
0 ! ! !
The coefficient matrix in this form is a simple upper triangular matrix. The last row in
Eq. (5.235) reads
ann xn = bn (5.236)
Thus,
xn = bn / ann (5.237)
From the penultimate row, we have
an 1, n 1 xn 1 + an 1, n xn = bn 1 (5.238)
Solution of Static Equilibrium Equations 207
Substituting for xn from Eq. (5.237) into Eq. (5.238), we can find xn1. Using xn and
xn1, we can find xn2, and so on through the process of backward substitution.
With respect to the finite element equations, we observe that the nonzero elements in [A]
are clustered around the principal diagonal as shown below.
h
dt

wi
0
nd
Ba
h
dt
wi
et
rp
Ca
The elements of [A] beyond the carpet width are zero and remain zero throughout the
Gauss elimination solution process. Thus it is unnecessary to store and operate on these
elements.
A further characteristic of the finite element coefficient matrices in typical structural
mechanics applications is symmetry and positive definiteness (i.e. all the eigenvalues of [A] are
positive). For a symmetric, banded [A], we need only store and operate on the nonzero
elements contained within the bandwidth, as shown below.
Bandwidth
0

0 n n 0 n b
[A] [Ab]
Since b << n, the saving in computer storage space and also further computations could be
tremendous. However, this calls for a (usually small) computer program to keep track of the
correspondence of elements in the original [A] and [Ab] matrices. The key ideas in such a
program are:
The principal diagonal in [A] is the first column in [Ab], i.e. a(i, i) = ab(i, 1) for
i = 1, n.
Similarly, for j = i to (i + bandwidth 1), a(i, j) = ab (i, l), where l is an index that
goes from 1 to bandwidth.
From symmetry, a( j,i) = a(i, j).
For symmetric, positive definite, banded matrix [A], the Gauss elimination solution can
be effectively implemented through Cholesky factorisation which is now discussed.
Cholesky factorisation
Here we aim to factorise the given coefficient matrix as
[ A] = [ L][ L]T (5.239)
where [L] is a lower triangular matrix, i.e.,
A11

A12 A 22 0
[ L] = (5.240)
# %

A1n ! A nn

If we perform the multiplication [L] [L]T and compare coefficient by coefficient with
elements in [A], we get the necessary equations to find lij .
A11
2
= a11 A11 = a11
a1 j
A11A1 j = a1 j A1 j = , j = 2, n
A11
A12
2
+ A 222 = a22 A 22 = a22 A12
2
a2 j A12 A1 j
A12 A1 j + A 22 A 2 j = a2 j A 2 j = , j = 3, n
A 22
A13
2
+ A 223 + A 233 = a33 A 33 = a33 (A13
2
+ A 223 )
a3 j (A13A1 j + A 23 A 2 j )
A13A1 j + A 23A 2 j + A 33A 3 j = a3 j A 3 j = , j = 4, n (5.241)
A 33
Thus, all the coefficients in [L] can be determined. We have

[ A]{x} = {b} (5.242)
or
[ L][ L]T {x} = {b} (5.243)
let
[ L]T {x} = { y} (5.244)
Hence
[ L]{ y} = {b} (5.245)
i.e.
A11 y1 b1

A12 A 22 0 y2 b2
= (5.246)
# % # #

A1n ! A nn yn
bn
The first row gives

b1
A11 y1 = b1 y1 = (5.247)
A11
and the second row yields
A12 y1 + A 22 y2 = b2 (5.248)
Substituting for y1 from Eq. (5.247), we can find y2 and so on till yn through the process
of forward substitution. Once {y} is found, we use (from Eq. (5.244)).
A11 A12 A1n x1 y1

A 22 x2 y2
= (5.249)
0 % # #

A xn
nn yn
Equation (5.249) gives xn and we find xn1, xn2 K, x1 through backward substitution.
Thus the Cholesky scheme of solution is in three steps:
Step 1: Cholesky factorisation [ A] = [ L][ L]T .
Step 2: Forward substitution [ L]{ y} = {b}.
Step 3: Backward substitution [ L]T {x} = { y} .
In the actual implementation on a banded matrix, the coefficients (lij) outside the
bandwidth are zero and are therefore never computed. For a matrix [ Ab ]nb size, [L] is also
of size (n b). It can be shown that the total number of floating point operations involved in
the above three steps is approximately (nb 2 / 2) + 2nb , where the Cholesky factorisation
takes (nb2/2) flops and each of steps 2 and 3 takes (nb) flops. For a given mesh (i.e. n is
given), the node numbering scheme dictates bandwidth b and the computational effort depends
significantly on b. Hence an efficient solution strategy will require that bandwidth b be as
small as possible. Thus efficient node numbering schemes, which optimize the resulting
bandwidth, are an integral part of any pre-processor in a typical finite element software
package. We will now present a few example problems.
Example 5.4. Illustration of use of bilinear rectangular element. We wish to redo the
problem of Example 5.1, but this time, using a single rectangular bilinear element as shown in
Figure 5.35. The computations presented here are based on a 2 2 Gauss quadrature. At each
integration point, the Jacobian and [B] are given as follows:
Integration Point 1:
x = 1/3, h = 1/3
2 0
[J ] = , J = 3.0
0 1.5
0.197 0 0.197 0 0.0528 0 0.0528 0

[ B] = 0 0.263 0 0.0704 0 0.0704 0 0.263 (5.250)

0.263 0.197 0.0704 0.197 0.0704 0.0528 0.263 0.0528
4 3
3 cm 1 N/m
1 2
4 cm
Fig. 5.35 Thin plate under uniform tension (Example 5.4).

x = 1/3, h = 1/3
2 0
[J ] = , J = 3.0
0 1.5
0.0528 0 0.0528 0 0.197 0 0.197 0

[ B] = 0 0.263 0 0.0704 0 0.0704 0 0.263 (5.251)

0.263 0.0528 0.0704 0.0528 0.0704 0.197 0.263 0.197
x = 1/3, h = 1/3
2 0
[J ] = , J = 3.0
0 1.5
0.197 0 0.197 0 0.0528 0 0.0528 0

[B] = 0 0.0704 0 0.263 0 0.263 0 0.0704 (5.252)

0.0704 0.197 0.263 0.197 0.263 0.0528 0.0704 0.0528
x = 1/3, h = 1/3
2 0
[J ] = , J = 3.0
0 1.5
0.0528 0 0.0528 0 0.197 0 0.197 0

[B] = 0 0.0704 0 0.263 0 0.263 0 0.0704 (5.253)

0.0704 0.0528 0.263 0.0528 0.263 0.197 0.0704 0.197
The stress-strain matrix [D] remains the same as before (see Eq. (5.123)). The element
stiffness matrix is evaluated as
2 2
[k ]e = [B] [D][B]t J WW
i =1 j =1
T
i j (5.254)
After appropriate summation over all the integration points, we get the element stiffness
matrix as follows:
Element stiffness matrix
0.891


0.379 0.0275 0.891

0.0275 0.296 0.357 1.17
[k ]e = 107
0.446 0.357 0.0672 0.0275 0.891 (5.255)

0.357 0.585 0.0275 0.881 0.357 1.17

0.0672 0.0275 0.446 0.357 0.379 0.0275 0.891

0.0275 0.881 0.357 0.585 0.0275 0.296 0.357 1.17
Upon incorporating the boundary conditions and using equivalent nodal forces, we obtain
the final set of equations for the unknown deflections, as
0.891 u2 1.5

0.357 1.17 Symmetric v2 0

107 0.0672 0.0275 0.881 u3 = 1.5 (5.256)

0.0275 0.881 0.357 1.17 v3 0

0.357 0.585 0.0275 0.296 v
1.17 4 0
Solving, we get the nodal deflection as

u2 = 2 107 cm
v2 = 0
u3 = 2 107 cm (5.257)
v3 = 4.5 108 cm
v4 = 4.5 108 cm
Example 5.5. Effect of element shape on the Jacobian. Consider the four-node quadrilateral
element shown in Figure 5.36(a). The corresponding parent element in the x-h frame is shown
in Figure 5.36(b) along with the sampling points for 2 2 Gauss quadrature rule. For each
Gauss point, we will now compute the mapping point in the real x-y space and also the
Jacobian. We wish to study what happens to [J] as the element loses its rectangular shape.
y h
4 (0, 1) (1, 1) 4 3 (1, 1)
h= 0.6
S R
(1, 0) 1 3 (1, 0)
x x
O
h = 0.6 P Q
2 (0, 1) (1, 1) 1 2 (1, 1)
x = 1/3 x = 1/3
(a) A Quad4 element (b) Parent element in x-h
Fig. 5.36 A regular quadrilateral element (Example 5.5).
For x1 = 1/3, h1 = 1/3, we have, from Eqs. (5.43) (5.46), the relations
x p = N1 x1 + N 2 x2 + N 3 x3 + N 4 x4
= N1 ( 1) + 0 + N 3 (1) + 0
= N1 + N 3
(5.258)
1
( )( ) (
= 1 + 1/ 1 + 1/ + 1 1/ 1 1/
4 )( )
= 1/
y p = N1 y1 + N 2 y2 + N3 y3 + N 4 y4
= N1 (0) + N 2 (1) + N 3 (0) + N 4 (1)
= N4 N2 (5.259)
=
1
4
( )( ) ( )(
1 + 1/ 3 1 1/ 3 1 1/ 3 1 + 1/ 3 = 0
)
From Eq. (5.158), we have
1 + 1/ 3 1 1/ 3 1 1/ 3 1 + 1/ 3
(0 + 1) + (1 0) ( 1 0) + (0 1)
4 4 4 4
[J ] =
1 1/ 3 1 1/ 3 1 + 1/ 3 1 1/ 3
(1 + 0) + (1 0) (1 0) + (1 + 0)
4 4 4 4
Thus,
0.5 0.5
[J ] = , J = 0.5 (5.260)
0.5 0.5
Similarly, for other quadrature points, we obtain
Y2 = 1/ 3, I2 = 1/3, xQ = 0, yQ = 1/ 3
0.5 0.5
[J ] = , J = 0.5 (5.261)
0.5 0.5
Y3 = 1/3 , I3 = 1/3 , xR = 0 , yR = 1/3
0.5 0.5
[J ] = , J = 0.5 (5.262)
0.5 0.5
Y4 = 1/3 , I4 = 1/3 , xS = 1/3 , yS = 0
0.5 0.5
[J ] = , J = 0.5 (5.263)
0.5 0.5
1
J = 0.5 + 0.5 + 0.5 + 0.5 = 2 (5.264)
Thus, sum of the determinants of Jacobian equals the area of the element.
We observe that | J | is positive and uniform at all quadrature points. Let us now consider
a distorted four-node quadrilateral as shown in Figure 5.37. The corresponding computations
are as follows:
(0, 1) 4
1 3 (1, 0)
x
(1/2, 0)
2 (0, 1)
Fig. 5.37 A distorted quadrilateral element (Example 5.5).
Y1 = 1/3 , I1 = 1/3 , xP = 0.35566 , yP = 0
0.0915 0.5
[J ] = , J = 0.0915 (5.265)
0.0915 0.5
We observe that (xP, yP) falls outside the element and | J | is negative.
Y2 = 1/3 , I2 = 1/3 , xQ = 0.25 , yQ = 1/ 3
0.3415 0.5
[J ] = , J = 0.125 (5.266)
0.0915 0.5
Y3 = 1/3 , I3 = 1/ 3 , xR = 0.25 , yR = 1/3
0.0915 0.5
[J ] = , J = 0.125 (5.267)
0.3415 0.5
Y4 = 1/3 , I4 = 1/3 , xS = 0.64434 , yS = 0
0.3415 0.5
[J ] = , J = 0.3415 (5.268)
0.3415 0.5
1
J = 0.0915 + 0.125 + 0.125 + 0.3415 = 0.5 , area of the element (5.269)
We observe that, as the element is distorted from its rectangular shape, | J | may vary
from point to point and may become even negative. Though the isoparametric formulation in
natural coordinates permits general quadrilateral elements and even elements with curved edges,
etc., all such elements lose accuracy when distorted from a rectangular shape. Included angles
at the vertices, variation of | J | from one Gauss point to another, ratio of length of the largest
to the smallest side (aspect ratio), etc. can all be used as indices of distortion. Such measures
of quality of shape of an element are incorporated in all the automatic mesh generation
algorithms available in most commercial FE software packages.
Example 5.6. Use of plane elements for bending. In this example, we will consider the
simple problem of a cantilever beam (EI = 0.1667 1010 N-cm2) subjected to shear load at the
tip. We attempt to use the 2-d plane elements to model the bending deformation. Consider the
cantilever beam shown in Figure 5.38. Under the tip load, the expected tip deflection and slope
are readily obtained as
P = 100 N
1 cm
10 cm
Fig. 5.38 A cantilever loaded at the tip (Example 5.6) (5 2 mesh of Quad4 element is shown).
PL3
E tip = = 0.02 cm (5.270)
3EI
P L2
R tip = = 0.00 3 (5.271)
2 EI
We will first use the plane four-node quadrilateral element to model this problem. The
load at the tip is distributed between the nodes using the consistent load vector, viz.,
[ N ] q dv. The plane element discussed here does not permit rotational d.o.f. at the nodes.
T
Thus we compute the slope at the tip as (utop ubottom)/h, where utop and ubottom correspond
to the top and bottom nodes and h is the height of the beam. The results obtained with different
meshes are listed in Table 5.3.
Table 5.3 Bending of a Cantilever (Example 5.6)
Mesh of Quad4 elements Tip deflection (cm) Tip slope

1 element 0.00051 0.000076
5 1 mesh 0.0076 0.00114
(5 elements along length; 1 element in height)
5 2 mesh 0.0078 0.00117
10 1 mesh 0.0136 0.002
20 1 mesh 0.0168 0.0025
40 1 mesh 0.018 0.0027
We observe that the plane Quad4 element is too stiff (deflection is underestimated) for
modelling such bending deformation and is not to be used for such problems. We find the tip
deflection to be 10% in error even with 40 elements! From the shape functions, we observe
that on any vertical line the variation of u-displacement is linear, which is the case in Euler
Bernoulli beam bending. Thus it is somewhat surprising as to why the Quad4 elements
performance is so poor. The interested reader is advised to compute the nonzero shear strain
in this element under a typical pure bending moment (where the shear strain is expected to be
zero). The spurious shear strain energy (parasitic shear) results in poor performance of the
element.
If we model the problem with the eight-noded serendipity element, the results obtained are
listed in Table 5.4.
Table 5.4 Bending of a Cantilever (Example 5.6)
Mesh of Quad8 elements Tip deflection (cm) Tip slope

1 element 0.0188 0.00287
2 1 mesh 0.0194 0.0029
In fact, the Quad9 element has been found to be even better than Quad8 element in terms of
sensitivity to shape distortion, ability to model bending deformation, and so on.
Example 5.7. Stress concentration around a circular hole. In this example, we wish to
illustrate the use of typical 2-d elements for modelling the problem of stress concentration in
a thin plate with a circular hole. We also wish to discuss exploiting symmetry in typical FE
analysis. Consider a plate subjected to in-plane uni-axial tension as shown in Figure 5.39. For
X
Y Y
Fig. 5.39 Thin plate with a circular hole under uniform tension (Example 5.7).
a homogeneous, isotropic material, perfect symmetry exists about the X- and Y-axes shown in
the figure. Thus only one quadrant as shown in Figure 5.40 needs to be modelled for finite
X
Fig. 5.40 Quadrant of plate (Example 5.7).

element analysis. For all nodes that lie on the symmetry planes, we need to stipulate symmetry
boundary conditions. By virtue of symmetry, the points on the X-axis can only move along the
X-axis. Similarly, for points on the Y-axis. A typical mesh for the quadrant of the structure is
shown in Figure 5.41. The tensile loading on the edge is modelled as consistent load vector for
Fig. 5.41 Typical finite element mesh (Example 5.7).
each element. For example, for a Quad 4 element,
N1 0

0 N1

N2 0

T 0 N 2 qx
{f } [N ] q dv v
e
= = ds
s N3 0 q y
0 N 3

N 0
4
0 N 4

where qx, qy represent the distributed force per unit length on the edge; shape functions
Ni are evaluated on the edge, and s is the coordinate used for integration along the edge
(e.g., x or h). In this example problem, only those element nodes on the loaded edge will have
non-zero nodal forces. For a typical boundary (edge) element, qx = 0 and x = 1 on the edge.
Thus, for a 4-node element, N2 = 0 = N3 and with qx = 0, the integrals for nodal force vector
simply involve N1 (Y 1)q y dy and N 4 (Y 1)q y d y . The consistent nodal forces for
Quad 4 and Quad 9 elements are depicted below.
qy h
2
qy h
qy h
2
qy h
6
4q y h
qy h
6
qy h
6
From the classical solution, it is well known that the stress concentration factor is 3.0
for this problem. The FE estimate of stress concentration along the vertical (y = 0) edge is
shown in Figure 5.42 for a 10 15 mesh (10 elements along radial and 15 along circumferential
0.917
0.917 (0.907)
(0.907)
10
10
0.925
0.925 (0.923)
(0.923)
0.928 (0.923)
0.928 (0.923)
99
0.93
0.93 (0.930)
(0.930)
0.935 (0.930)
(0.930)
88
0.936 (0.938)
0.936
0.944 (0.938)
(0.938)
0.944 (0.938)
77
0.947
0.947 (0.949)
(0.949)
0.959 (0.949)
0.959 (0.949)
66
0.966
0.966 (0.970)
(0.970)
0.988 (0.970)
(0.970)
55
1.003 (1.011)
1.003
1.045 (1.011)
(1.011)
1.045 (1.011)
44
1.083
1.083 (1.095)
(1.095)
1.168 (1.096)
33 1.168 (1.096)
1.265
1.265 (1.282)
(1.282)
(1.285)
1.447 (1.285)
22 0.447
1.70 (1.720)
1.70 (1.720)
(1.727)
11 2.12 (1.727)
2.813 (2.795)
2.813 (2.795)
6
Quad4 Quad9
Quad4 Quad9
elements elements
elements elements
Fig. 5.42 Stress concentration factor obtained using FEM (Example 5.7).
direction) of Quad4 and Quad9 elements. The discretisation is done in geometric progression
in the radial direction (so that elements nearer the hole are smaller in size) but arithmetic
progression in circumferential direction. The two numbers on either side of a node are the
estimates of stress concentration factor obtained by treating the node as belonging to one
element or the other. It is observed that the FE solution underestimates the stress concentration
factor. The stresses at a common node are significantly different for the Quad4 mesh while
they are very much uniform for the Quad9 mesh. The variation of stress at a common node
when computed of all the adjoining elements (that share this node) is an index of the quality
of the mesh. Adaptive refinement of mesh, implemented in many commercial FE software
packages, uses such indices to selectively refine the mesh where needed. Use of more elements
of the same type (e.g. a 15 15 mesh of Quad4 elements here) is referred to as h-type
refinement. The h here refers to the element size. This example illustrates p-type refinement
where we refine the polynomial shape function used (Quad4 to Quad9 here).
We have so far discussed two-dimensional finite element analysis for structural
mechanics. As discussed earlier, many of the concepts discussed are applicable to other fields
also. We will now analyse the finite element solution of 2-d fluid flow problems. In contrast
to the structural mechanics problems, we wish to start from the basic governing differential
equation and develop the finite element equations through the weak form statement.
5.8 2-d Fluid Flow

We now discuss the finite element solution of a simple, two-dimensional fluid flow problem,
viz., the steady-state irrotational flow of an ideal fluid over a cylinder, confined between two
parallel plates as shown in Figure 5.43. We assume that, at the inlet, the velocity is uniform,
say V0. We wish to determine the flow velocities near the cylinder. We will formulate the
V0
Fig. 5.43 Ideal fluid flow over a cylinder.
problem in terms of stream function Y. A stream line is a line that is tangent to the velocity
vector. By definition, therefore, there is no flow across a stream line. The stream function is
constant for a stream line defined in such a way that the velocity components are given by
Z
u = Z
y , v = (5.272)
x
2-d Fluid Flow 221
The governing differential equation, in terms of the stream function, is given by

2Z 2Z
+ = 0 (5.273)
x 2 y 2
Considering the symmetry in the problem, only one quadrant as shown in Figure 5.44
needs to be modelled. By virtue of symmetry, AB becomes a stream line (no flow across it)
E
D
V0 = 1 H
C
r=a
Y
A B
L
Fig. 5.44 Domain used for finite element modelling.
and y should be a constant along AB. Since the flow velocities are only derivatives of y, any
arbitrary datum for y can be chosen. Let y = 0 on the line AB.
The curved boundary BC is again a stream line (and, therefore, ABC becomes a single
continuous stream line) and y = 0 on this. In view of the uniform inlet velocity V0, y = V0x.
on the line AE. On the line DE, therefore, y = V0 H. By virtue of symmetry, on the line CD,
Z / y = 0 . We can write the weighted residual statement as
2Z 2Z
A
W 2 +
x

y 2
dA = 0

(5.274)
Recall that, integrating by parts, we get
d 2Z dZ u dW dZ
x
xu xu
xA
W
dx 2
dx = W
dx xA xA dx dx
dx (5.275)
On similar lines, we write (using the GreenGauss theorem)
2Z Z u W Z
x
yu
A
W
x 2
dA = yA
W dy
x xA
A x x
dA (5.276)
The general interpretation of xl, xu and yl, yu is shown in Figure 5.45. Therefore,
dy = A x ds (5.277)
ds n
xu
dy
xl
ds
n
Y
yu yl
Fig. 5.45 Integration limits.
where lx is the direction cosine of the outward normal nG and is introduced to take care of
both ends appropriately. Thus,
Z u Z
x
yu
yA W x dy =
xA v Ws x
A x ds (5.278)
Substituting in Eqs. (5.276), we get

2Z W Z Z
A
W
x 2
dA = A x x
dA + v W
s x
A x ds (5.279)
where s refers to the boundary curve. Similarly, we have

2Z W Z Z
A
W
y 2
dA = A y y
dA + v Ws y
A y ds (5.280)
Substituting Eqs. (5.279)(5.280) in Eq. (5.274), we have
2Z 2Z W Z W Z
A
W 2 +
x
dA =
y 2
A

x x
+ dA
y y
Z Z
+ v W x A
s
x +
y
A y ds

(5.281)
= 0
Therefore,
W Z W Z Z Z Z
A

x x
+ dA =
y y v W x A
s
x +
y
A y ds =

v W n ds
s
(5.282)
y
where Z / n represents flow velocity in the negative s-direction (i.e. Vs = ).
n
2-d Fluid Flow 223
For a typical element, each node has one d.o.f., viz., the stream function value y. In our
standard finite element notation, for a general 2-d element, we have
y = [N ]{y}e (5.283)
e
where [N], the shape function matrix and {y} , the nodal values of stream function take
appropriate size depending on the element. For a Quad4 element, the sizes are {y}e41, [N]14.
Z
x

= [ B]{Z }
e

Z (5.284)
y

where [B] contains the derivatives of shape functions. In the Galerkin formulation, we use the
same shape functions as the weight functions also (i.e., W = N). Thus, Eq. (5.282) for an
element can be rewritten as
Z
( [B] [B] dA){Z }
A
T e
= v [ N ]
s
T
ds
n
(5.285)
i.e.
[K]e{y}e = {f}e (5.286)
where
Z
[ K ]e = ( [B] [B] dA),
A
T
{f }e = v [ N ]
s
T
n ds

(5.287)
For the entire finite element mesh, we have
Z
NELEM NELEM
[B] [ B] dA = v [ N ]
1
A
T
1
s
T
n ds

(5.288)
i.e.
[K]{y} = {F} (5.289)
where
NELEM
[K ] = 1
{K }e (5.290)
NELEM
{F } = 1
{ f }e (5.291)
and {y} contains all the nodal d.o.f. of stream function values for the entire mesh and the
summation over the elements is interpreted as per standard assembly procedure in FEM.
The term on the RHS in Eq. (5.288) is evaluated along the element boundary. For every
edge of an element that is shared with another element, this term will be evaluated twice-once
clockwise (for one element) and the other anti-clockwise (for the other element). Hence, when
summed up for all the elements, these terms on internal boundaries of individual elements
vanish, leaving out only the outer boundary of the entire domain. In the present problem, no
velocity parallel to any boundary (recall that Vs = Z / n ) has been specified and hence this
term will be zero.
Following Eq. (5.164), for a typical element we write
y = [N]{y}e (5.292)
Z
x

= [B ]{Z }e = [B1 ]22 [B2 ]22 [B3 ]2NDF {Z }eNDF1
Z (5.293)
y
21
where, for the present case,
1 0
[B1 ] = (5.294)
0 1 22
J 22 J12
J J

[B2 ] = (5.295)
J 21 J11
J J
2 2
N1 N 2
Y !
Y
[B3 ] = (5.296)

N1 N 2
!
I I 2NDF
where NDF is the total number of d.o.f. for the element (e.g. 4 for Quad4, 9 for Quad9, etc.).
Thus the solution of the fluid flow over cylinder is very much similar to the stress
concentration problem discussed earlier (i.e., we can consider [D] to be identity matrix), and
we use the same mesh for both the problems.
The governing equations are finally obtained as
[K]{y} = {0} (5.297)
where, for all nodes falling on the line AB or BC (Figure 5.44) y = 0; for nodes on the line
DE, y = V0H; for each node on the line AE, y = V0x, where x represents the x-coordinate
of that node. For these prescribed boundary conditions, we solve Eq. (5.297) for the remaining
unknown y values. Once the y values are known, the flow velocities at any point in the
element can be calculated from [B]{y}e for that element. Results obtained with a 10 15 mesh
(10 elements along radial and 15 along circumferential direction) of Quad4 and Quad9 elements
are shown in Figure 5.46. It can be easily verified that the finite element solution tallies very
well with the exact solution.
Summary 225
1.0177 1.017
10
1.022 1.018
1.0177 1.019
9
1.030 1.0252
1.022 1.0253
8
1.0454 1.0367
1.030 1.0367
7
1.0706 1.0563
1.0454 1.0562
6
1.1124 1.0887
1.0706 1.0886
5
1.1813 1.1424
1.1124 1.1422
4
1.2846 1.2308
1.1813 1.2305
3
1.4812 1.3765
1.2846 1.3759
2
1.7878 1.6165
1.4812 1.6160
1
1.7878 2.012
6
Quad4 Quad9
elements elements
Fig. 5.46 Flow velocity obtained using FEM.
SUMMARY
We have discussed various 2-d finite elements in this chapter. Our basic procedure for finite
element formulation has been the same as that used for 1-d elements. However, to allow for
general curved edge elements, we introduced the natural coordinates and presented the use of
numerical integration to evaluate the element matrices in a typical FE program. We also talked
about incorporation of prescribed boundary conditions and Gauss elimination based algorithms
for solution of equations. We studied typical structural mechanics and ideal fluid flow problems.
Extension to three-dimensional problems is possible, i.e., triangular elements become
tetrahedral, etc., and numerical integration will have to be carried out in 3-d. So far, however,
we have discussed only steady state problems. In Chapter 6, we will discuss the finite element
solution of an important class of problems, viz., 1d and 2d transient dynamic problems.
PROBLEMS
5.1 For the triangular element shown in Fig. P5.1, the nodal values of displacements are:
u1 = 2.0, u2 = 3.0, u3 = 5.0
v1 = 1.0, v2 = 2.0, v3 = 3.0
Obtain the displacements (i.e. u, v) of point P(2, 2) within the element.
3 (2, 3)
P (2, 2)
(1, 1) 1 2 (3, 1)
O X
Fig. P5.1 Displacement of a CST element.
5.2 Determine the consistent load vector for the CST element under the action of the loading
shown in Fig. P5.2.
3
py = 100 N/m2
0.5 m
px = 200 N/m2 2
0.5 m
1
1m
Fig. P5.2 Consistent nodal load vector for CST element (thickness = 0.1 m).
5.3 Determine the consistent load vector for the CST element under the action of gravity
loading, acting in the plane of the element.
Problems 227
5.4 Find the consistent nodal force vector for the six-noded triangular element loaded as
shown in Fig. P5.4.
3
6
y y
2
py = po 1
L
5
x
1
2
Fig. P5.4 Consistent nodal load vector for LST element (edge length = L; thickness = 1).
5.5 Obtain the consistent nodal load vector for the element shown in Fig. P5.5.
4 (5, 5) 3 (5, 5)
p0
1 (5, 5) 2 (5, 5)
Fig. P5.5 Consistent nodal load vector for Quad4 element (thickness = 1).
5.6 For an eight-noded quadrilateral element, shown in Fig. P5.6, determine the consistent
nodal load vector.
y
7 3
4
p2
6
8 x
p1
1 5 2
Fig. P5.6 Consistent nodal load vector for Quad8 element (edge length = L; thickness = 1).
5.7 Evaluate the following integral by a 2 2 Gauss quadrature rule:
1 1 2 + x

1 1 3 + xy
dx dy
5.8 Determine the hatched area in Fig. P5.8 by 1-pt Gauss quadrature.
5.9 Check what order of Gauss quadrature (i.e. No. of points) would exactly integrate the
following:
(a) 2 + 3x + 5x3 + 8x6;
(b) (2 x2)/(2 + x2);
(c) (x2 + y2);
(d) 2 + 3x + 5x3;
(e) (2 x)/(2 + x).
(0, 1) (1, 1)
0.3
0.3 0.3
0.3
(0, 0) (1, 0)
Fig. P5.8 Area of a figure.
5.10 Use a 2 2 Gauss quadrature rule to evaluate I.

3 + x2
I = 2 + y2
dx dy , x (0, 6), y (0, 4)
5.11 Explain why the weight factors in 1-d Gaussian quadrature, sum up to 2.
5.12 Determine the area bounded by y2 = 4x, y = 0 and x = 4 using a single-point Gauss
quadrature. What is the percentage error?
5.13 Starting from the two-dimensional heat conduction equation, obtain the element nodal heat
flow-temperature relationship for a three-node triangular element through Galerkin
formulation.
5.14 For the two-dimensional heat transfer in a square slab shown in Fig. P5.14, each element
ijk has the following equations:
Problems 229
4 3
i k
k j
3
j 5
4 2
j k
k
1
i i
i j
1 2
Fig. P5.14 A square slab.
25 1 21 Ti 300

1 18 7 T j = 500

21 7 41 Tk 800
All the edges are maintained at T = 0oC. Obtain the final equation(s) after assembling
element matrices and incorporating the boundary conditions.
5.15 For the three-noded triangular element shown in Fig. P5.15, calculate the temperature at
point P, given the nodal temperatures as T1 = 100C, T2 = 200C, and T3 = 300C.
3 (6, 8)
P (5.5, 5.5)
2 (7, 6)
Y
(5, 5) 1
O X
Fig. P5.15 Temperature at an interior point.
5.16 In a particular discretisation consisting of all six-node triangles for stress analysis, the
temperature over an element can be taken to be constant, say, Ts. The material properties
are E and v. Determine the element nodal load vector in the absence of any other loads.
5.17 The temperature at the four corner nodes of a four-noded rectangle are T1, T2, T3, and
T4. Determine the consistent load vector for a 2-d analysis, aimed to determine the thermal
stresses.
5.18 Number the nodes in the following discretisation (Fig. P5.18) using eight-noded
quadrilateral/six-noded triangular elements, so as to minimise the bandwidth. What is the
bandwidth for such numbering?
Fig. P5.18 Node numbering.
5.19 The Gauss elimination method (or Cholesky factorisation) broke down while solving a
system of equations arising from FEM. Explain why this may be so.
5.20 Consider the uniform strength beam shown in Fig. P5.20. It is aimed at using two-
dimensional finite elements to study the variation of strain across the depth of such a
P
h0 h x
x h 2 = h02
l
Fig. P5.20 Uniform strength beam.

Problems 231
beam in order to verify the accuracy of the EulerBernoulli assumptions. Show clearly
how the curved edge may be modelled using isoparametric elements, through appropriate
coordinate transformation.
5.21 Write all the shape functions for the elements shown in Fig. P5.21.
h h
4 3
4 3
x
x
1 5 2 1 5 6 2
( 31 , 1)( 13 , 1)
Fig. P5.21 Transition elements.
5.22 Give an approximate estimate of the total number of multiplications (n.o.m.) involved in
the Gaussian elimination solution of [k]nn{d}n1 = { f }n1. Show that the n.o.m. involved
in Cholesky factorising a banded, symmetric matrix [k]nb is nb2/2.
5.23 What are the conditions for convergence in the displacement finite element formulation?
Show that the completeness requirements are met if Ni = 1, where Nis are the element
displacement shape functions.
5.24 For a six-noded linear strain triangular element used for 2-d stress analysis:
(a) Show whether the equilibrium equations (ignoring body forces) are satisfied or not
within the element, i.e.
T x U xy U xy T y
+ = 0, + = 0
x y x y
(b) Discuss whether the equilibrium is satisfied across inter-element boundaries.
(c) Comment on the compatibility across element boundaries as well as within the element
(d) State whether or not the equilibrium of nodal forces and moments is satisfied.
CHAPTER 6
Dynamic Analysis Using
Finite Elements
6.1 Introduction
So far we have discussed situations where the external excitation as well as the response of
the system were time invariant. However, in many practical situations, such steady-state
conditions are reached after a period of time in which the external disturbances cause the
system response to fluctuate with time (transient period). For example, when a certain
temperature boundary condition is suddenly prescribed, thermal transients are set up in the
system. The specific heat of the material is the property that resists variation of temperature
with time and needs to be accounted for. In a vibrating system, the acceleration and
deceleration of the structural parts are resisted by the inertia of the system and certain
dissipative forces are also developed (such as viscous friction proportional to the velocity of
the moving parts).
When analysing such unsteady state or transient dynamic problems, we are interested in
finding the response of the system (i.e., temperature, displacement, etc.) as a function of time
given the external disturbances. In this chapter we discuss the formulation and solution of such
problems using the finite element method. We will first discuss the vibration problems and then
transient heat transfer problems. We can develop the formulation using the governing
differential equations (as in Chapter 2) or using certain energy principle (as in Chapter 3), and
we will show that the formulation follows on very similar lines. The additional terms in the
governing equations due to inertia, dissipation, specific heat, etc. lead to very interesting and
often complex behaviour of the system. We will analyse the solution of the equations in
considerable detail.
6.2 Vibration Problems

When a structure is excited by forces which vary with time, the response of the structure is
also time varying, and inertia/dissipation properties of the structure affect the response. As we
will be discussing in this chapter, a complete dynamic response analysis is usually much more
complex than a static (time invariant) analysis. Thus it is useful to assess the need for a
232
Vibration Problems 233
dynamic analysiswhen the excitation forces are varying slowly with time, we term the
response quasi-static, and a static analysis is sufficient. If the highest frequency component in
the excitation is less than about one-third the lowest (fundamental) natural frequency of the
structure, a static analysis is usually assumed to be sufficient.
This can be argued from the typical response of a single d.o.f. system to a harmonic
excitation (Figure 6.1). The magnification factor (MF), (i.e. the factor by which the dynamic
response is more than the static response), is given as
1
MF =
(6.1)
( /Xn )
2
1
when
=
1
Xn , MF = 1.125 (6.2)
3
| MF | = x/E st
k
m 1.0
x
F0 sin Wt W/w n
1.0
(a) Simple spring mass system (b) Dynamic magnification factor
Fig. 6.1 Single d.o.f. system and its dynamic response.
Thus the dynamic response amplitude will only be 12.5% more than the static response
for W = wn/3. The typical response of a multi-d.o.f. system (with several natural frequencies)
is indicated in Figure 6.2. Static analysis is considered sufficient if the excitation frequency is
less than about one-third the lowest (fundamental) natural frequency of the system.
Amplitude
Frequency
Fig. 6.2 Dynamic response of a multi d.o.f. system.
234 Dynamic Analysis Using Finite Elements
We need to distinguish two types of dynamic analysis problems, called the wave
propagation and structural dynamics problems depicted in Figure 6.3. In the wave propagation
problems, the excitation is usually an impact or blast force usually lasting for a fraction of a
second or so. The entire structure does not instantaneously know that it has been hit. The time
scales involved are comparable to the time taken for this information to traverse (i.e. the stress
wave to propagate at the speed of sound in the medium) the entire structure. Crash analysis
of a car shown in Figure 6.3(a) is a typical problem of this type. Impact analysis of a missile
on a target structure is another typical wave propagation problem. In the structural dynamics
type of problems, the entire structure simultaneously participates in the response, and the time
scales involved are often several seconds. An automotive crank shaft vibration shown in
Figure 6.3(b) is a typical problem of this type. A typical earthquake excitation contains
frequencies up to, say 25 Hz, and is essentially a low frequency structural dynamic problem.
So also the problem of response of off-shore structures to wave loading.
(a) Automobile crash
(b) Crank shaft vibration

Fig. 6.3 Types of dynamic problems.
Equations of Motion Based on Weak Form 235
Our discussion in this book will primarily centre on finite element solution of structural
dynamics problems. We will analyse problems of free vibration (i.e. how the structure will
vibrate when subjected to some initial conditions and left free to vibrate on its own) as well
as forced vibration (i.e. response of structure under time varying forces). Free vibration
problems are eigenvalue problems whose solution gives us the resonance frequencies and mode
shapes of the structure. It is not only of interest for the designer to know the resonance
frequencies (to be isolated from operating frequencies), but some of the solution strategies for
the forced vibration problems require this information (e.g. mode superposition technique). So
we will first discuss the free vibration problems and then take up determination of transient
dynamic response of structures subjected to time dependent loads.
Returning to the simple example of a single d.o.f. system (Figure 6.1), we observe that
the damped natural frequency is given as
Xd = Xn 1 Y2 (6.3)
where x is the damping factor. Normally, x is very low,x 2 << 1. For example, if x = 0.1 (i.e.
10% critical damping), wd = 0.995wn. So, for many practical systems, wd wn. In other
words, we can ignore damping in the evaluation of natural frequencies. However, the response
near a resonance is critically determined by the damping present in the system and we will
include damping while estimating the dynamic response.
Finite element equations for dynamics can be derived in a manner entirely analogous to
static equations discussed in earlier chapters (e.g. Chapter 2). We start with the typical partial
differential equation, write down its weak form, and develop the finite element form of the
weak form equation. Alternatively, we can use an energy based method (similar to Chapter 3)
to derive the finite element dynamic equations. We will discuss both the methods for typical
elements.
6.3 Equations of Motion Based on Weak Form

6.3.1 Axial Vibration of a Rod
The governing equation for free axial vibration of a rod (considering the dynamic equilibrium
of a differential element) is given by
2u 2u
AE = S A (6.4)
x 2 t 2
Using the technique of separation of variables and assuming harmonic vibration, we have
u(x,t) = U(x)ei w t (6.5)
Substituting from Eq. (6.5) in Eq. (6.4), we obtain
d 2U
AE + S AX 2U = 0 (6.6)
dx 2
The Weighted-Residual (WR) statement can be written as
L d 2U
0
W ( x) AE 2 + S AX 2U dx = 0
dx
(6.7)
Integrating by parts, the weak form of the WR statement can be rewritten as

L
dU L dU dW L
W ( x) AE dx
0 0
AE
dx dx
dx + 0
W ( x)S AX 2U ( x) dx = 0 (6.8)
We observe that the first two terms are identical to Eq. (2.97) and we have an additional
term involving the mass density (inertia) effects. For a finite element mesh, the integrals are
evaluated over each element and then summation is done over all elements.
We will now develop the necessary equations for a single element. For a typical bar
element (Figure 6.4), we have two nodes and axial deformation d.o.f. at each node. The
interpolation functions are given by Eq. (2.126).
x x
U ( x ) = 1 U1 + U 2 (6.9)
A A
x= 0
x=l
u1 u2
1 2
F1 F2
Following the Galerkin formulation, the weight functions are the same as the shape
(interpolation) functions. So we have
x x
W1 ( x) = 1 , W2 ( x) = (6.10)
A A
Writing the weak form equations with respect to W1 and W2, we obtain (from Eq. (6.8)).
U U1 1 x 2 x x
A A
P0 AE 2
A A dx + 1 A S AX 1 A U1 + A U 2 dx = 0
0 0
(6.11)
U U1 1 x 2 x x
A A
PA AE 2
A A dx + A S AX 1 A U1 + A U 2 dx = 0 (6.12)
0 0
where P stands for AE (dU/dx).

We can rewrite this in matrix form as
A x x A x x
AE 1 1 U1 P0
1
0 A

1 dx
A 1
0 A dx
A U1
2
= + S AX U
A 1 1 U 2 PA A x x A x x 2

0
1 dx
A
A
0 A A
dx

(6.13)
On evaluating the integrals, we can write Eq. (6.13) as
AE 1 1 U1 P0 2 1 S ALX 2 U1
= + (6.14)
A 1 1 U 2 PA 1 2 6 U 2
We have designated the element stiffness matrix as
AE 1 1
[ k ]e = (6.15)
A 1 1
We now designate the element mass matrix as
S AA 2 1
[ m]e = (6.16)
6 1 2
We observe that the sum of all the coefficients in this mass matrix equals rAl, the mass
of the element. Equation (6.14) can be rewritten as
P0
[ k ]e {E }e = + [ m ] X {E }
e 2 e
(6.17)
A
P
where {d }e contains the nodal d.o.f.
The element mass matrices can be assembled following exactly the same procedure as
for the element stiffness matrices. Since free vibration problems do not involve external forces,
the column of forces vanishes upon assembly and we obtain the typical free vibration equation
[K ]{U} = w2[M ]{U} (6.18)
where [K ] and [M ] represent the global, assembled stiffness and mass matrices, respectively
and {U } contains all the nodal d.o.f. This is a typical eigenvalue problem, and we will discuss
the solution of such equations in Section 6.8.
6.3.2 Transverse Vibration of a Beam

The governing equation for free transverse vibration of a beam based on the EulerBernoulli
theory (considering the dynamic equilibrium of a differential element) is given by
4v 2v
+ S
EIA = 0 (6.19)
x 4 t 2
Using the technique of separation of variables and assuming harmonic vibration, we have
V(x,t) = V(x)eiwt (6.20)
Substituting in Eq. (6.19), we obtain
d 4V
S AX 2V = 0
EI (6.21)
dx 4
The WR statement can be written as
d 4V
L
0
W ( x) EI 4 S AX 2V dx = 0
dx
(6.22)
On integrating by parts, we obtain

L
d 3V d 3V dW
S AX 2W ( x)V ( x) dx = 0
L L
W ( x) EI 3
dx 0 0
EI
dx3 dx
dx 0
(6.23)
Once again performing integration by parts, we get

L L
d 3V dW d 2V L d 2V d 2W
W ( x) EI 3
dx 0 dx
EI 2 +
dx 0 0
EI
dx2 dx2
dx
(6.24)
L
S AX W ( x)V ( x) dx = 0
2
0
For a typical EulerBernoulli beam element (Figure 6.5), we have

V(x) = N1V1 + N2q1 + N3V2 + N4q2 (6.25)
V1 V2
q1 q2
x = l
x = 0
Fig. 6.5 Typical EulerBernoulli beam element.
where N1, N2, N3 and N4 are given in Eq. (4.117), and are reproduced here for easy reference:
N1 = 1 3x2/l2 + 2x3/l3, N2 = x 2x2/l + x3/l2
N3 = 3x2/l2 2x3/l3, N4 = x2/l + x3/l2 (6.26)
Following the Galerkin procedure, we use these same shape functions as the weight
functions. Substituting in the weak form [Eq. (6.24)] and evaluating the integrals with respect
to each of the weighting functions one at a time, we obtain the governing equations. However,
we observe that the first three terms in [Eq. (6.24)] are similar to those obtained in the static
case [Eq. (2.103)]and we will therefore present here only the additional mass matrix terms. The
mass matrix terms will correspond to the following integrals:
A A 3x 2 2 x3 3x 2 2 x3
0
S AW1 ( x)V ( x) dx =
0
1 2 + 3
A A
S A

1 2 + 3
A A
V1

2 x2 x3 3x 2 2 x3
+ x + 2 R1 + 2 3 V2
A A A A
x2 x3
+ + 2 R 2 dx
A A
S AA
=
420
[156V1 + 22AR1 + 54V2 13AR 2 ] (6.27)
A A 2 x2 x3 3x 2 2 x3
0
S AW2 ( x)V ( x) dx = 0
x
A
+ 2
A
S A

1
A2
+
A3
V1

2 x2 x3 3x 2 2 x3
+ x + 2 R
1 + 2 V2
A A A A3
x2 x3
+ + 2 R 2 dx
A A
S AA
=
420
(
22A V1 + 4A2R1 + 13A V2 13A R 2 ) (6.28)
A A 3x 2 2 x3 3x 2 2 x3

0
S AW3 ( x)V ( x) dx = 0
2 3 S A
A A
1
A2
+ V1
A3
2 x2 x3 3x 2 2 x3
+ x + 2 R1 + 2 3 V2
A A A A
x2 x3
+ + 2 R 2 dx
A A
S AA
=
420
[54V1 + 13AR1 + 156V2 22AR 2 ] (6.29)
A x2
A x3 3x 2 2 x3
0
S AW4 ( x)V ( x) dx =
0

A
+ 2 S A 1
A A2
+ V1
A3
2 x 2 x3 3x 2 2 x3
+ x + 2 R1 + 2 3 V2
A A A A
x 2 x3
+ + 2 R 2 dx
A A
S AA
= 13A V1 3A 2R1 22A V2 + 4A 2R 2 (6.30)
420
Rewriting the terms in matrix form, we obtain the EulerBernoulli beam element mass
matrix as
156 Symmetric

S AA 22A 4A2
[m]e = (6.31)
420 54 13A 156

13A 3A2 22A 4A2
We observe that the coefficients of the mass matrix corresponding to translational d.o.f.
S AA
viz. 420 (156 + 54 + 54 + 156) sum up to rAl, the mass of the element.

If we need to write the mass matrix for a plane frame element, we can combine the bar
element and beam element mass matrices in just the same way as we combined their stiffness
matrices in Section 4.6.
We have illustrated the formulation of finite element equations starting from the governing
differential equations. We shall now illustrate an energy based approach based on Lagranges
equations of motion.
6.4 Equations of Motion Using Lagranges Approach

The equations of motion [e.g. Eqs. (6.4) and (6.19)] for the axial motion of a rod and the
transverse motion of a beam, can be readily derived based on Newtons Second Law of motion.
Many a time it is convenient to use an energy based method and Lagranges equations* of
motion are commonly used. If T represents the kinetic energy of a system and p represents
*Excellent discussions of the theory behind Lagranges equations of motion are available in standard texts such
as Elements of Vibration Analysis, 2nd ed., by L. Meirovitch, McGraw-Hill, New York, 1986 and Lagrangian
Dynamics by D.A. Wells, Schaum Series, McGraw-Hill, New York, 1967.
Equations of Motion Using Lagranges Approach 241
potential energy, then Lagranges equations of motion in the independent generalised coordinates
q are given by
d T T Q
+ = Fq (6.32)
dt q q q
where F |q represents the generalised force in the coordinate q and q = dq/dt.

We will illustrate the use of Lagranges approach in formulating the equations of motion
through the following simple example.
Example 6.1. Consider a two-d.o.f. spring mass system shown in Figure 6.6; x1(t) and x2(t)
are the independent generalised coordinates. The kinetic energy of the system is
1 1
T = m1 x12 + m2 x22 (6.33)
2 2
x1 x2
k1 k2 k3
m1 m2
Fig. 6.6 2-d.o.f. system (Example 6.1).
The potential energy of the system is given by

1 1 1
Q = k1 x12 + k2 ( x1 x2 ) 2 + k3 x22 (6.34)
2 2 2
For further use in Lagranges equations, we now obtain the derivatives of T and p w.r.t.
x1, x2, x1 and x2.
T d T
= m1 x1 , = m1
x1
x1 dt x1
T d T
= m2 x2 , = m2
x2 (6.35)
x2 dt x2
T T
= 0, = 0
x1 x2
Q
= k1 x1 + k2 ( x1 x2 )
x1
Q
= k3 x2 k2 ( x1 x2 )
x2
Substituting from Eq. (6.35) in Eq. (6.32), we obtain the required equations of motion as
m1
x1 + k1 x1 + k 2 ( x1 x2 ) = F1 (t )
(6.36)
m2
x2 + k3 x2 k 2 ( x1 x2 ) = F2 (t )
or, in matrix form,
(k1 + k2 ) k2 x1 m1 0 x1 F1 (t )
+ = (6.37)
k2 ( k2 + k3 ) x2 0 m2
x2 F2 (t )
6.4.1 Formulation of Finite Element Equations

Using our standard shape functions, for a typical finite element, the displacement of an interior
point can be written, in terms of the nodal d.o.f., as
{d} = [N]{d}e (6.38)
Differentiating with time, the velocity at the point is given by
{d} = [ N ]{E}e (6.39)
where we have used the fact that the shape functions are only functions of spatial coordinates
and are therefore time-invariant, and {E}e represents nodal velocities.
For the simple truss element,
x x u1
{d } = u = 1 (6.40)
A A u2
x x u1 (6.41)
{d} = u = 1
A A u2
For the beam element,
v1

R1
{d } = v = [N1 N2 N3 N 4 ] (6.42)
v2

R 2
v1

R1
{d} = v = [N1 N2 N3 N 4 ] (6.43)
v2

R2
For a two-dimensional element, each point can have u and v displacements and, therefore,
u1

v1
u (6.44)
{ d } = = [ N ] u2
v
v2

#
u1

v1
u

{d} = = [ N ] u2 (6.45)
v
v2

#
where the size of [N ] and {d}e depend on the particular element (e.g. Quad4, Quad8, etc.),
as discussed in Chapter 5. The kinetic energy of an elemental mass dm ( = r dV) within the
element is given by
1 1
dT e = (dm) (velocity)2 = {d}T {d}S dV (6.46)
2 2
Using Eq. (6.39), we can write
1 eT
dT e = {E } [ N ]T [ N ]{E}e S dV (6.47)
2
For the whole element, the kinetic energy is obtained by integrating over the element as
follows:
1 eT
Te = {E } [ N ]T [ N ]{E}eS dV (6.48)
v 2
Since nodal point velocities {E }e do not vary from point to point within the element, these
can be taken out of the integral, and hence
Te =
1 eT
2
{E } (v
)
S[ N ]T [ N ] dv {E}e (6.49)
1 eT
Te = {E } [ m]e {E}e (6.50)
2
where [m]e = ( S[N ] [N ] dv) is the consistent mass matrix for the element.
v
T
For the whole body (i.e. mesh of finite elements), we have
NOELEM
T = 1
Te (6.51)
where NOELEM is the number of elements.

The total potential energy of the system has already been obtained in Chapter 5 and is
reproduced here from Eq. (5.1055.106):
NOELEM
1 e
3p = 1
eT e e eT
2 {E } [ k ] {E } {E } { f }

(6.52)
Interpreting the summation over all the elements as per standard assembly of finite
elements, we can write
1 T
T = {E } [ M ]{E} (6.53)
2
1 T
Q = {E } [ K ]{E } {E }T {F} (6.54)
2
where {d } and {E } contain the displacement and velocities for all the nodes of the entire
structure; [k] and [M] are the assembled global stiffness and mass matrices; and {F} is the
assembled global nodal force vector.
The required derivatives
T T d T
= 0, = [ M ]{E}, = [ M ]{E} (6.55)
{E } {E} dt {E}
can be employed for further use in Lagranges equations,
Q
= [ K ]{E } {F} (6.56)
{E }
Substituting in Lagranges equations of motion, we obtain
[ M ]{E} + [ K ]{E } {F } = {0} (6.57)

i.e.
[ M ]{E} + [ K ]{E } = {F } (6.58)
where the global mass matrix, stiffness matrix and force vectors are obtained following the
usual assembly of individual element level matrices. The stiffness matrix and force vectors have
already been dealt with in the previous chapters. For dynamic problems we see that the inertia
effects result in mass matrix. Since we have used the same shape functions as for the stiffness
matrix, the mass matrix in Eq. (6.58) is a consistent mass matrix. The consistent mass matrices
for a few elements are discussed in the following section.
6.4.2 Consistent Mass Matrices for Various Elements

Bar element
x
1 A x x
1
A
[m]e = S[ N ] [ N ] dv = SA
T
dx
v 0 x A A

A
(6.59)
S AA 2 1
=
6 1 2
This equation is the same as that given by Eq. (6.16) which was obtained using the weak form.
Beam element
N1

A N2
[m]e = S [ N ]T [ N ] dv = S A [N1
N2 N3 N 4 ] dx
0 N
v
3 (6.60)

N4

where the shape functions N1, N2, N3 and N4 are given in Eq. (6.26).
After performing all the integrations, the element mass matrix is obtained as
156 Symmetric

S AA 22A 4A
2
[ m] =
e

420 54 13A 156 (6.61)

13A 3A 2 22A 4A 2
which is the same as the mass matrix given in Eq. (6.31) obtained using the weak form.
Two-dimensional elements
In general, it is tedious to evaluate, in closed form, the expression for the element mass matrix
using
[m]e = v
S[ N ]T [ N ] dv (6.62)
For example, for a Quad8 element we have [N]216. Thus, [N]T[N] will be a (16 16) matrix.
Even if we use symmetry, we still need to evaluate 136 integrals! Invariably, these computations
are performed numerically inside a computer program. While performing numerical integration,
it must be observed that those integrals involve [N ]T[N ] type of terms. The terms in the
integrals for a stiffness matrix [see Eq. (5.103)] are of the type [B]T[D][B], where coefficients
in [B] are spatial derivatives of [N ]. Therefore, the highest degree polynomial term in the mass
matrix expression is always of a degree higher than that for stiffness matrix. Thus the same
rule of Gauss quadrature may not be adequate, and an appropriate rule must be carefully selected.
6.5 Consistent and Lumped Mass Matrices

The mass matrices derived above are termed consistent mass matrices since we used the same
shape functions as for their stiffness matrices. It is, however, possible to lump the mass of
the entire element at its nodes and come up with a lumped-mass equivalent representation. Such
an intuitive lumping is depicted for bar and beam elements in Figure 6.7.
l
i j
ui uj
rAl rAl
2
2
vi vj
l
qi qj
rAl rAl
2 2
Fig. 6.7 Mass lumping for bar and beam elements.
Mathematically, the corresponding lumped mass matrices can be written as
S AA
2 0
[me ]lumped = for a bar element (6.63)
S AA
0
2
S AA
2 0 0 0

0 0 0 0
[me ]lumped = for a beam element (6.64)
S AA
0 0 0
2

0 0 0 0
Consistent and Lumped Mass Matrices 247
From the beam element lumped mass matrix, we observe that we have assigned no
lumped rotary inertia to the nodes. The mass lumping at the nodes has resulted in a diagonal
mass matrix which is computationally very advantageous. The eigenvalue problem solvers for
free vibration problems are typically iterative (see Section 6.8) and dynamic response
calculations involve time marching schemes with small increments in time Dt (see Section 6.9),
and hence are computationally intensive. Thus a diagonal mass matrix will be useful, and several
researchers have proposed efficient schemes to arrive at the lumped mass matrix rather than
the intuitive, ad hoc lumping above. We will briefly describe one such scheme, popularly used
in many finite element calculations.
6.5.1 HRZ Lumping Scheme

The essential idea in this scheme is to simply use only the diagonal elements of the consistent
mass matrix but to scale them in such a way that the total mass of the element is preserved.
For example, for the bar element, the HRZ lumped mass matrix can be written as
S AA 1 0
[me ]HRZ = (6.65)
2 0 1
With reference to the beam element lumped mass matrix, we first write the diagonal
elements of the consistent mass matrix (Eq. 6.31) as follows:
156 0 0 0

S AA 0 4 A 2
0 0
[me ]diag. = (6.66)
420 0 0 156 0

0 0 0 4A 2
The total mass (as represented in this model) is obtained by summing all the diagonal
S AA
elements corresponding to translational d.o.f. in one direction. Thus we get (156 + 156)
420
312
= S AA . Since the total mass of the element is S AA , the scaling factor is 420 . We scale
420 312
all the diagonal elements with this scale factor to generate the HRZ lumped mass matrix for
a beam as
1/2 0 0 0

0 A 2 /78 0 0
[me ]HRZ = S AA (6.67)
0 0 1/2 0

0 0 0 A /78
2
Simple intuitive lumping and HRZ lumping schemes can be readily applied to two-
dimensional elements also. For example, for the Quad8 element, the HRZ mass distribution will
yield the mass lumping at the nodes as shown in Figure 6.8.
m 8m m
36 36 36
8m 8m
36 36
m 8m m
36 36 36
Fig. 6.8 HRZ lumped mass matrix for Quad8 element (based on 2 2 Gauss quadrature)
(m = total mass of the element).
We will now solve a few simple free vibration problems. In the following example, we
compare the first few natural frequencies obtained using both consistent and lumped mass
matrices.
Example 6.2. Consider a uniform cross-section bar (Figure 6.9) of length L made up of a
material whose Youngs modulus and density are given by E and r. Estimate the natural
frequencies of axial vibration of the bar using both consistent and lumped mass matrices.
Fig. 6.9 A uniform bar (Example 6.2).
Using just one element for the entire rod (i.e. l = L) and using lumped mass matrix, we have
1
2 0
AE 1 1 u1 u
1
= X lump
2
S AA (6.68)
A 1 1 u2 1 u2
0
2
In view of the boundary condition at node 1 (u1 = 0), we have
AE S AL
u2 = X lump
2
u2 (6.69)
L 2
Hence,
2E 1.414 E
X lump = = (6.70)
SL2
L S
With one element and consistent mass matrix, we have
AE 1 1 u1 S AL2 1 u1
= X cons.
2
(6.71)
L 1 1 u2 6 1 2 u2
and with u1 = 0,
AE S AL
u2 = Xcons. 6 (2u2 ) (6.72)
2
L
Therefore,
3E 1.732 E
X cons. = = (6.73)
SL2
L S
The exact solution for the fundamental frequency of a fixed-free bar is given as
1.571 E/S . Thus we see that the consistent mass matrix overestimates, and the lumped mass
matrix underestimates, the natural frequency.
If we now use a two-element mesh, i.e. l = L/2, see Figure 6.10, we get the following
results with lumped and consistent mass matrices.
1 2
1 2 3
Fig. 6.10 Two-element model of a bar (Example 6.2).
Results with lumped mass matrix

The assembled equations can be readily verified to be given by
1 1 0 u1 1/2 0 0 u1
AE L
1 1 + 1 1 u2 = S A X 2 0 1/2 + 1/2 0 u2 (6.74)
L 2

2 0 1 1 u3 0 0 1/ 2 u3
With u1 = 0 (boundary condition), we have
2 AE 2 1 u2 S AL 2 1 0 u2
= X (6.75)
L 1 1 u3 2 0 1/2 u3
i.e.
2 M 1 u2 0
= (6.76)
1 1 ( M /2) u3 0
S L2
where M = X 2.
4E
For a nontrivial solution, we have
2M 1
= 0 yielding l1 = 0.586, l2 = 3.414 (6.77)
1 1 ( M /2)
Thus the natural frequencies are
1.531 E
X1 = X2 =
3.695 E
L S, L S
(6.78)
Results with consistent mass matrix

Using consistent mass matrices, we obtain
1 1 0 u1 2 1 0 u1
AE S ALX 2
1 2 1 u2 = 1 4 1 u2 (6.79)
L/2 12
0 1 1 u3 0 1 2 u3
Using u1 = 0 (boundary condition) and M =

X 2 S L2 , for nontrivial solution, we obtain the
24 E
equation
2 4M 1 M
= 0 (6.80)
1 M 1 2M
On solving, we get l1 = 0.108 and l2 = 1.32. Thus the natural frequencies are
X1 =
1.61 E , X2 =
5.63 E
(6.81)
L S L S
The exact solution can be readily verified to be
iQ E
Xi = , i = 1, 3, 5, K, (6.82)
2L S
Thus,
Q E , 3Q E , 5Q E , ... (6.83)
X1 = X2 = X3 =
2L S 2L S 2L S
We observe that the given rod is a continuous system and hence has infinitely many d.o.f.
and therefore infinitely many natural frequencies. Our finite element model with one element has
just one independent d.o.f., and that with two elements has just two independent d.o.f. Thus
we can predict only one or two natural frequencies, for the one-element and two-element
model, respectively. As we refine the mesh with more and more finite elements, we will actually
be admitting more and more d.o.f., and therefore, predict higher natural frequencies also. The
accuracy of predicted natural frequencies also improves.
Table 6.1 shows, for the axial vibrations of a rod, the predicted natural frequencies as we
refine the mesh.*
Table 6.1 Natural Frequencies of a Fixed-free Bar

(L = 1 m, E = 2 1011 N/m2, r = 7800 kg/m3, A = 30 106 m2)
No. of elements 1 2 3 4 8 16 Exact
Mode
1 1140.0 1234.0 1252.0 1258.0 1264.0 1265.0
1396.0 1299.0 1280.0 1274.0 1268.0 1266.0 1265.9
2 2978.0 3420.0 3582.0 3743.0 3784.0
4537.0 4188.0 4019.0 3853.0 3812.0 3797.8
3 4670.0 5366.0 6078.0 6266.0
7597.0 7301.0 6586.0 6393.0 6329.6
4 6319.0 8180.0 8688.0
10,560.0 9563.0 9037.0 8861.5
5 10,000.0 11,030.0
12,850.0 11,770.0 11,393.3
In each case, the upper row indicates the frequencies obtained with lumped mass and the lower
row, the frequencies obtained with consistent mass matrices. We observe that typically these
two provide the lower and upper bounds on the frequencies and the exact frequency is in
between. It is conceivable that a mass matrix [m] which is an average of the two can provide
greater accuracy than either of them!
Example 6.3. Consider the simply supported beam, shown in Figure 6.11. Let the length
L = 1.0 m, E = 2 1011 N/m2; area of cross-section, A = 30 cm2; moment of inertia I = 100 mm4;
density r = 7800 kg/m3. We will obtain the first five natural frequencies using the three types
of mass matrices, viz., simple lumped, HRZ lumped, and consistent mass matrices.
Table 6.2 gives the results.
*
See, for example, L. Meirovitch, Elements of Vibration Analysis, 2nd ed., McGraw-Hill, New York, 1986.
Fig. 6.11 Simply supported beam (Example 6.3).
Table 6.2 Natural Frequencies (Hz) of a Simply Supported Beam

(For each mode, the first row gives the results obtained with simple lumping,
the second row with HRZ lumping, and the third row with consistent mass matrix.)
No. of elements 2 3 4 8 Exact

Mode
1 14.42 14.52 14.52 14.52
14.21 14.46 14.51 14.52 14.52
14.58 14.53 14.52 14.52
2 57.67 58.07 58.09
104.3 56.84 57.84 58.03 58.11
64.47 58.32 58.11 58.09
3 122.4 130.5 130.7
149.2 120.2 129.3 130.4 130.75
162.1 133.1 130.9 130.7
4 230.7 232.3
180.0 416.2 227.4 231.4 232.45
295.5 257.9 233.3 232.4
5 354.7 362.8
481.3 348.0 360.6 363.20
409.9 366.4 363.3
We observe that the simple lumped mass matrix assigns zero rotary inertia, and hence has
fewer nonzero mass matrix elements. Therefore, it predicts fewer natural frequencies of the
system. The HRZ lumping scheme corrects this by assigning appropriate rotational inertia at the
nodes and accurately predicts as many frequencies as the full consistent mass matrix. A good
mass lumping scheme, e.g. the HRZ scheme, is computationally advantageous, excellent in
accuracy and is, therefore, popularly used.
The exact mode shapes for the simply supported beam are given by
2 iQ x
vi ( x) = sin , i = 1, 2, 3, ... (6.84)
S AL L
The fifth mode shape, for example, is plotted in Figure 6.12. Within each beam element,
we recall that the interpolation functions used permit cubic variation. Thus, even with eight
elements, our approximation to the mode shape is inaccurate and we commit approximately
4% error in the natural frequency using the HRZ lumped mass. The anticipated mode shape
vis--vis the individual element shape/interpolation functions must always be borne in mind
while deciding the fineness of finite element mesh required for dynamics problems.
Form of Finite Element Equations for Vibration Problems 253
5Q x
G ( x) = sin
L
Fig. 6.12 Fifth mode shape of a simply supported beam.
6.6 Form of Finite Element Equations for Vibration Problems

A typical finite element formulation would yield the governing equations as
[M ]{ X} + [C ]{ X } + [ K ]{ X } = {F (t )} (6.85)
In the case of free vibration, there is no external loading {F} = 0 and damping has negligible
role. Therefore, for undamped free vibration problems,
[M ]{ X} + [ K ]{ X } = 0 (6.86)
Assuming harmonic vibration at a frequency wi, {Xi} = {Ui} sin w it, we have
w i2[M ]{Ui} + [K ]{Ui} = 0 (6.87)
or
[ K ]nn{U i }n1 = Xi2 [M ]nn{U i }n1 , i = 1, 2, L, n (6.88)
This is the basic equation, in matrix form, governing the undamped free vibration of the
structure. Given the stiffness and mass matrices [K ] and [M ], we have to find the natural
frequencies wi and mode shapes {Ui}, i =1, 2, L, n. These represent certain characteristic
states of free vibration of the system, each such state characterised by its eigenvalue wi and
eigenvector {Ui}.
Equation (6.88) is a set of n equations in (n + 1) unknowns, viz., w i and n elements of
{Ui}. So a unique solution cannot be obtained. We choose to determine w i, and the ratios of
elements of {Ui}. Thus the mode shapes are defined only within a multiple of themselves, i.e.
if we have an eigenvector {U}, then a{U } (a 0) is also an eigenvector.
It is to be understood that there exist n pairs of natural frequencies w and mode shape
{U }. If the structure is excited at the frequency wi, it would vibrate such that the various
points will have relative amplitudes as given in the mode shape eigenvector {Ui}. For
example, the first mode of a simply supported beam would be as shown in Figure 6.13.
Qx
G ( x ) = sin
L
Fig. 6.13 First mode shape of a simply supported beam.
If the beam is modelled with four beam elements, then the finite element nodal points will
vibrate as shown in the following equation:
v1 0

R1 1

v2 0.707

R2 0.707

v
3 1
{U1 } = =
R3 0

v4 0.707 (6.89)

R4 0.707

v5 0

R 1
5
The mode shape only indicates the relative amplitudes of vibration (viz., overall shape
of vibration) and may readily be scaled to any amplitude. Thus we could write {U1} assuming
v3 = 10 units, and all the d.o.f. will be also scaled by the same factor. Thus there is no unique
amplitude of free vibration. There are certain commonly accepted norms for scaling, which we
will discuss later on.
Another way of interpreting the natural frequency and mode shape is that, if the structure
were given an initial displacement to all its d.o.f. according to the relative amplitudes given in
the mode shape, and left free to vibrate on its own, it will vibrate at the natural frequency w,
always maintaining these relative amplitudes.
Since the finite element mesh of the structure has n d.o.f., our model would yield n
pairs of natural frequency wi and mode shape {Ui} (i = 1, 2, , n). It is to be appreciated
that the structure, being a continuous system, has infinitely many d.o.f., and hence infinitely
many natural frequencies and mode shapes. Thus it is expected that our finite element model
will become more and more accurate as we refine the mesh. The general form of the governing
equations for the undamped free vibration is given by
[ K ]nn {U i }n1 = Xi2 [M ]nn {Ui }n1 (6.90)
We may rewrite Eq. (6.90) as
[M ]1 [ K ]{U i } = X i2 {U i } (6.91)
i.e.
[ A]{U i } = Mi {U i } (6.92)
Some Properties of Eigenpairs 255
where [ A] = [M ] [ K ] and M i = Xi2. We can also rewrite Eq. (6.90) as

1
1
{U i } = [ K ]1[ M ]{U i } (6.93)
Xi2
i.e.
li{Ui} = [A]{Ui} (6.94)
where [A] = [K]1[M] and li = 1/w 2i .
The form of representation
[A]{U} = l{U} (6.95)
is known as the standard form of eigenvalue problem, whereas the form [K ]{U} = l [M ]{U}
is known as the nonstandard form. With the understanding that these two forms can be
converted from one to the other easily, we will use both the forms in our subsequent
discussion. However, it must be appreciated that, in typical finite element computations,
[M]1 may not be possible to compute (e.g. because of diagonal zeros in a lumped mass
matrix), or [K]1 may not be desirable (e.g. heavy computational overhead for large size
matrices and also bandedness will be lost). Thus we will normally avoid any explicit
computation of an inverse.
An important aspect to be observed here is the size of the problem. A typical finite element
mesh for complex real-life problems may involve several thousand d.o.f. and thus [K ] and [M ]
matrices are of this size. Hence we can potentially determine several thousand natural
frequencies of the system. However, in most practical systems, the excitation frequencies are
in a limited range (mostly less than a few kHz) and only those modes of the system that are
near this range participate significantly in the response. Thus we are interested in determining
a few (may be 10, 20 or 50) natural frequencies that lie in the frequency range of interest to
us. Another aspect specific to finite element models of typical structural eigenvalue problems
is the symmetry of [K ] and [M ] matrices. While several general purpose algorithms are
available for solution of matrix eigenvalue problems, we will discuss here a few that have been
popularly used in typical finite element analysis.
It is helpful to understand some of the basic properties of the eigenvalues and
eigenvectors before we proceed to discuss the algorithms for their actual determination. In fact,
these properties are extensively used in the actual determination of the eigenpairs.
6.7 Some Properties of Eigenpairs

Basis vectors
The eigenvectors constitute the least number of linearly independent mode shapes of the
system. Any deflected shape of the structure can thus be represented as a linear combination
of these eigenvectors {Ui}, i.e.,
{X } = c1{U1} + c2{U2} + c3{U3} + L + cn{Un} (6.96)

Orthogonality of eigenvectors (for symmetric matrices)

Consider two distinct eigenpairs (wi, {Ui}) and (wj, {Uj}):
[K ]{Ui} = Xi2 [M ]{Ui} (6.97)
[K ]{Uj} = X 2j [M ]{Uj} (6.98)
Premultiplying Eq. (6.97) by {Uj}T, we obtain
{U j }T [ K ]{U i } = X i2 {U j }T [ M ]{U i } (6.99)
Taking the transpose of Eq. (6.98) and postmultiplying by {Ui}, we get
{Uj }T [ K ]T {U i } = X 2j {Uj }T [M ]T {U i } (6.100)
If [K ] and [M ] are symmetric, which is normally the case for a finite element mesh, we can
subtract Eq. (6.100) from Eq. (6.99) and obtain
0 = (i 2j ){Uj}T[M ]{Ui}
2
(6.101)
Since they are distinct eigenpairs, w i wj. Therefore,

{Ui}T[M ]{Uj} = 0 (i j) (6.102)
i.e. the eigenvectors are mutually orthogonal with respect to the mass matrix [M ].
Eigenvalue shifting
[K ]{U} = w2[M ]{U} (6.103)
Using a scalar m, we can write
([K ] m[M ]){U} = (w 2 m)[M ]{U} (6.104)
i.e.
[ K ] {U } = l[M]{U} (6.105)
Thus we see that by effecting a shift in the eigenvalues of the system, we are not
changing the eigenvectors. This property is used in typical computer procedures, to avoid zero
eigenvalues (i.e. rigid body modes) as also to improve the convergence of two close
eigenvalues.
Sturm sequence property

We have
[K]{U } = l [M]{U} (6.106)
Let us use a shift m yielding
[[K ] m[M ]]{U} = (l m)[M ]{U} (6.107)
Solution of Eigenvalue Problems 257
If we now Cholesky factorise such that

[K] m[M ] = [L][D][L]T (6.108)
where [L] is a lower triangular matrix and [D] is a diagonal matrix, then an important property
of [D] is that the number of negative elements in [D] equals the number of eigenvalues smaller
than m. This property is typically used to check our eigenvalue solution. If we have obtained,
say m eigenvalues, we can use a trial value (say, m) slightly higher than the highest eigenvalue
obtained and perform a Sturm sequence check. A Sturm check fail is indicated if the number
of eigenvalues below m turns out to be more than m. This means that our eigenvalue solver
missed out some eigenvalues, indicating that we need to redo the analysis.
6.8 Solution of Eigenvalue Problems

There are essentially three groups of methods of solution of eigenvalue problems:
1. Determinant based methods
2. Transformation based methods
3. Vector iteration based methods
The determinant based methods are primarily based on the following:
[A]{U} = l{U} (6.109)
[[A] l[I]]{U} = {0} (6.110)
Hence,
[ A] M[ I ] = 0 for a nontrivial {U} (6.111)
Thus, in principle, we can take a trial value of l and compute the determinant |[A] l[I]|.
With several trial values, we can generate a plot such as the one shown in Figure 6.14. By
suitably monitoring the sign changes in the value of the determinant, we can iterate towards
the eigenvalues l.
Such a scheme is, however, never implemented in practice because of the heavy
computational cost. Evaluation of each determinant of size (n n) requires of the order of n3
| A - mI |
m
0 l1 l2 l3 l4
Fig. 6.14 Determinant-based method.

floating point operations, and several iterations may be required to determine all the eigenvalues
of interest. We will therefore not discuss these methods in this book.
The second group of methods, viz., the transformation based methods, attempt to
transform the eigenvalue problem as follows: Given
[A]{U } = l{U } (6.112)
Transform [A] into a diagonal matrix or tridiagonal matrix, using a series of matrix
transformations of the type [ A] = [T ]T [A][T ], where [T ], the transformation matrix, is
usually an orthogonal matrix, i.e. [T ]T = [T ]1. If we are able to transform [A] completely into
a diagonal matrix, then the elements on the diagonal themselves are the required eigenvalues.
Among the well known methods of this type are the Givens method, Householders method,
Jacobi method, and the Lanczos method. We will discuss the Jacobi method in this text, which
is commonly implemented in many commercial finite element software packages.
The final group of methods, viz., vector iteration based methods, involve assuming a
trial eigenvector and performing repeated matrix manipulations to converge to the desired
eigenvector. We will discuss the basis of vector iteration methods, followed by the forward and
inverse iteration methods, and simultaneous and subspace iteration methods. Vector iteration
methods are normally available in many commercial finite element software packages.
6.8.1 Transformation Based Methods

The central idea behind these methods is based on the orthogonality property of the
eigenvectors. We recall that, for a symmetric eigenvalue problem given by
[K ]{U} = w2[M ]{U} (6.113)
We have
{Ui}T[M ]{Uj} = 0 for i j (6.114)
If we scale the eigenvectors such that {Ui}T[M ]{Ui} = 1 and we arrange together
individual eigenvectors as columns of a matrix, we can write
[F] = [{U1} L {Un}] (6.115)
M1 0

M2
[F]T[K ][F] = (6.116)
%

0 Mn

1 0

1
[F]T[M ][F] = (6.117)
%

0 1
where li are the eigenvalues.
Thus we can try to construct the eigenvector set [F] through a series of matrix mani-
pulations through which we seek to transform [K ] and [M ] into diagonal matrices. We shall
explain the basic procedure with respect to the eigenvalue problem in its standard form, i.e.
[A]{U} = l{U} (6.118)
Let us denote the transformation matrices used by [T ]. The iterative procedure can be
outlined as follows:
Denote [A1] = [A] (6.119)
Compute [A2] = [T1]T[A1][T1] (6.120)
where [T1] has been so chosen as to make any one off-diagonal element (say, a12) zero. Compute
[A3] = [T2]T[A2][T2] = ([T1][T2])T [A]([T1][T2]) (6.121)
where [T2] is formed in such a way as to make another off-diagonal element (say, a13) zero.
It is to be appreciated at this stage that in this process the (1, 2)-element may have once again
become nonzero! Thus we need to perform several iterations. Compute
[F] = [T1][T2] (6.122)
Repeat this process till we reach the (n 1, n) off-diagonal element and make it zero. At this
stage we have typically completed one sweep of all the off-diagonal elements. Now we repeat
the entire cycle.
After several such sweeps, we would have effectively diagonalised the matrix [A]. We
talk of effective diagonalisation because, all computations being numerical, we will never be able
to achieve exact diagonalisation. We can define a threshold value (e.g. 1% of the smallest
diagonal element) and, if an off-diagonal element is less than this threshold, we can assume
it to be zero! Our estimate of the eigenvector set at any stage is clearly given by the product
of the transformation matrices
[F] = [T1][T2] L (6.123)
and the diagonal elements themselves are our approximation of the eigenvalues.
Jacobi method:
If an off-diagonal element aij is to be reduced to zero, the corresponding matrix [T] is chosen
as
1 0 0 0
1
%

1
cos R 0 sin R i th row (6.124)
[T ] = 0 1 #
# #

sin R cos R j th row
1
%

0 1

ith col. j th col.
where
2aij
tan 2R = for aii ajj,
aii a jj
(6.125)
Q
R= if aii = ajj
4
In a typical computer implementation, we actually do not have to evaluate [T]T[A][T]
through complete matrix multiplication. It is just the linear combination of the corresponding
two rows and columns, while the rest of the matrix remains the same.
When we have both [K] and [M] as in the typical FEM, we cannot directly apply the
Jacobi procedure outlined above, even if [M] is a diagonal lumped mass matrix. If [T ] has been
chosen to make a kij vanish, then [T]T[M][T] will remain diagonal only if mii = mjj. We can,
however, extend the basic Jacobi scheme to the nonstandard eigenvalue problem.
[K ]{U } = w 2[M ]{U} (6.126)
Let us say that we wish to make kij and mij vanish. Let us write the transformation matrix
as
1 0
1 B ith row
[T ] = (6.127)
C 1 j th row

0 0 0 1

ith jth
and let us write col. col.
[ K ] = [T ]T [ K ][T ], [ M ] = [T ]T [ M ][T ] (6.128)
It is easily verified that
kij = kij + B kii + C kjj + BC kij (6.129)
mij = mij + B mii + C mjj + BC mij (6.130)
a and b can be obtained in such a way that kij = 0, mij = 0. The solution for a and b can
*
be obtained as follows: Let
N1 = kii mij mii kij
N2 = k jj mij m jj kij
N3 = kii m jj mii k jj
*K.J. Bathe, Finite Element Procedures, Prentice-Hall of India, New Delhi, 2001.
N3 N
2
N4 = + sign ( N3 ) 3 + N1 N2 (6.131)
2 2
then
N2 N1
B = , C = (6.132)
N4 N4
In the special case that [M] is a diagonal lumped mass matrix, we have
N1 = mii kij , N2 = m jj kij (6.133)
We will illustrate the generalised Jacobi method using the following example.
Example 6.4. Consider the three-element model of the fixed-free bar (shown in Figure 6.15)
undergoing axial vibrations. Let L = 1 m, A = 30 106 m2, E = 2 1011 N/m2, r = 7800
1 1 2 2 3 3 4
L/3 L/3 L/3
Fig. 6.15 Three-element model of a fixed-free bar (Example 6.4).
kg/m3. Using consistent mass matrix, the assembled stiffness and mass matrices are readily
obtained as
0.360 108 0.180 108 0

[ K1 ] = 0.180 108 0.360 108 0.180 108 ,

0 0.180 108 0.180 108

0.052 0.013 0
(6.134)
[ M 1 ] = 0.013 0.052 0.013

0 0.013 0.026
where the fixed-end boundary condition is already incorporated (i.e. the corresponding row and
column have not been assembled).
We will obtain the eigenpairs of this system using the generalised Jacobi method. One
sweep will consist of zeroing the elements (1, 2), (1, 3) and (2, 3) in [K ] and [M ] matrices.
First sweep i = 1, j = 2. From Eqs. (6.131) (6.132),
m1 = 140,4000.0, m2 = 140,4000.0, m3 = 0, m4 = 140,4000.0

1 1 0

[T1 ] = 1 1 0 (6.135)

0 0 1
1.08 0 0.180

[ K 2 ] = [T1 ]T [ K1 ][T1 ] = 108 0 0.360 0.18 (6.136)

0.18 0.18 0.18
0.078 0 0.013

[ M 2 ] = [T1 ] [ M 1 ][T1 ] = 0
T
0.13 0.013 (6.137)

0.013 0.013 0.026
First sweep i = 1, j = 3.
1 0 0.309

[T2 ] = 0 1 0 (6.138)

1.24 0 1
1.80 0.222 0

[ K 3 ] = [T2 ] [ K 2 ][T2 ] = 10 0.222
T 8
0.36 0.18 (6.139)

0 0.18 0.172
0.0856 0.0161 0

[ M 3 ] = [T2 ]T [ M 2 ][T2 ] = 0.0161 0.130 0.013 (6.140)

0 0.013 0.0415
First sweep i = 2, j = 3.
1 0 0

[T3 ] = 0 1 0.47
(6.141)
0 1.36 1
1.80 0.222 0.105

[ K 4 ] = [T3 ]T [ K 3 ][T3 ] = 108 0.222 0.188 0 (6.142)

0.105 0 0.421
0.0856 0.0161 0.00756

[ M 4 ] = [T3 ]T [ M 3 ][T3 ] = 0.0161 0.242 0 (6.143)

0.00756 0 0.058
We observe that the off-diagonal elements, once zeroed, do not remain zero. Also, these
are not negligible yet. We therefore need to perform another sweep from (1, 2) to (2, 3) element.
We now present the results after the third such sweep has been completed. So,
1 0 0

[T9 ] = 0 1 0.132 106 (6.144)
6
0 0.373 10 1
1.96 0.216 1010 0.284 1017

[ K10 ] = [T9 ]T [ K9 ][T9 ] = 108 0.216 1010 0.161 0.222 1022 (6.145)
16

0.416 10 0.421 1016 0.424
0.0859 0.156 1011 0.204 1018

[ M 10 ] = [T9 ]T [ M 9 ][T9 ] = 0.156 1011 0.248 0 (6.146)
18

0.195 10 0.694 1017 0.0612
We conclude that the off-diagonal elements in [K10] and [M10] are negligibly small. Our
estimates of the eigenvalues of the bar are, therefore,
1 k11 1 1.96 108

f1 = = = 7597 Hz (6.147)
2Q m11 2Q 0.0859
1 k22 1 0.161 108

f2 = = = 1280.43 Hz (6.148)
2Q m22 2Q 0.248
1 k33 1 0.424 108

f3 = = = 4187.64 Hz (6.149)
2Q m33 2Q 0.0612
which are the same as those given against consistent mass in Table 6.1 for three-element case,
though the order is interchanged. Our estimate of the eigenvectors is given by
[' ] = [T1 ][T2 ] " [T9 ] (6.150)

0.697 0.745 0.886

[' ] = 1.21 1.29 0 (6.151)

1.39 1.49 0.886
However, these approximate mode shapes have not been scaled such that [' ]T [ M ] [ ] = [ I ].
We can extract the [M]-orthonormalised mode shapes [F] from [' ] , following the procedure
discussed later in the section. We then obtain
2.38 1.50 3.58

[' ] = 4.13 2.58 0.005 (6.152)

4.74 3.01 3.59
We observe that the original matrices were banded (semi-bandwidth = 2). During Jacobi
transformations, bandedness is not guaranteed. So, in a practical implementation of the scheme
on a several thousand d.o.f. finite element mesh (though usually bandwidth may be much
smalla couple of hundred at most), we need to store the entire (n n) stiffness and mass
matrices. It can be readily shown that even a diagonal lumped mass matrix will not remain
diagonal after Jacobi transformation. Moreover, numerical round-off errors slowly creep into
the computations. Thus the most common implementation of the generalised Jacobi scheme has
been to solve the reduced eigenvalue problem of the subspace iteration technique discussed
towards the end of this section.
6.8.2 Vector Iteration Methods

Basis of vector iteration method
Consider the standard eigenvalue problem [A]{U} = l{U}. Let its eigenpairs be (l1{U1}),
(l2{U2}), L (ln{Un}). Since any deflected shape can be represented as a linear combination
of these eigenvectors, a trial vector {X 1} can be written as
{X 1} = c1{U1} + c2{U2} + c3{U3} + L + cn{Un} (6.153)
Compute
{X2} = [A]{X1} = c1[A]{U1} + c2[A]{U2} + L + cn[A]{Un}
= c1l1{U1} + c2l2 {U2} + L + cnln{Un} (6.154)
After performing premultiplications with [A] m times, we obtain
{ X m +1} = c1M1m {U1} + c2 M2m {U 2 } + " + cn Mnm {U n }

M
m
M
m
= c1M1m {U1} + c2 2 {U 2 } + " + cn n {U n } (6.155)

M1 M1

If l1 > l2 > l3, L > ln , i.e. ( M2 /M1 < 1, !, Mn /M1 < 1) , then after several iterations,
m
Mi
0 (i = 2, 3, 4, K, n), or
M1
{ X m+1} c1M1m {U1} (6.156)
For example, if an eigenvalue is one-fifth the highest eigenvalue, after 10 iterations, we have
(1/5)10 = 0.0000001024, which is negligibly small. If the individual eigenvalues are well
separated, the trial vector {X m+1} will essentially contain components of the mode shape {U1}
while all other terms would vanish. We say that we have achieved convergence to the
eigenvector {U1}. Since mode shapes are only relative amplitudes, the scale factor c1lm is of
( )
little consequence. If we now compute { X }T [ A]{ X }/{ X }T { X } , we get an approximate
estimate of l1. It should be observed that, irrespective of our choice for trial vector {X1},
convergence is always to the largest eigenvalue and the corresponding eigenvector.*
The rate of convergence (i.e. how quickly we can estimate the mode shape components)
depends on how well separated the eigenvalues are. If an eigenvalue is close to the highest
eigenvalue, say, the ratio is 0.9, then, even after several iterations, its contribution in {X} will
not vanish. To improve the rate of convergence, we need to separate the eigenvalues and,
for this purpose, we can successfully exploit the eigenvalue shift property. For example, let
l1 = 100 and l2 = 90, yielding l2/l1 = 0.9. If we use a shift of m = 85, i.e. we operate on
the matrix [ A] = [A] m [I ], then we will effectively have M 1 = 100 85 = 15 and M 2 = 90
85 = 5 yielding M 2 / M 1 = 0.33. After 10 iterations, (l2/l1)10 = (90/100)10 = 0.3486 whereas
( M2/M1)10 = (5/15)10 = 0.000015!
Forward and inverse iteration

In typical finite element computations, we have the nonstandard eigenvalue problem
[K ]{U} = w2[M ]{U } (6.157)
When we extend the basic vector iteration scheme to this nonstandard form, we have two
options [K ]{U } = w2[M ]{U} can be written as
1/w 2{U } = [K]1[M]{U} [M]1[K]{U} = w 2{U}

\l{U} = [A]{U} [A]{U} = l{U}
Assume {U} and premultiply by [A] to Assume {U} and premultiply by [A] to
converge to the largest l, i.e. the lowest w2 converge to the largest l, i.e. the largest w 2
INVERSE ITERATION FORWARD ITERATION
Since we are usually interested in the lowest eigenvalue (i.e. the first or the lowest critical speed
of a rotor, etc.), the inverse iteration is commonly employed.
In each iteration an estimate of the eigenvalue can be obtained by computing a quotient
called Rayleigh Quotient (RQ), defined as
{ X }T [ K ]{ X }
RQ = (6.158)
{ X }T [ M ]{ X }
*This does not hold good in the degenerate case when our trial vector {X} is perfectly orthogonal to {U1}.
This is based on the fact that

[K ]{U } = w2[M ]{U} (6.159)
Therefore,
{U}T[K ]{U } = w2{U }T[M ]{U } (6.160)
Thus,
{U }T [ K ]{U }
X2 (6.161)
{U }T [ M ]{U }
This is, in fact, a measure of the potential energy/kinetic energy. We have already noted that
{U } is a set of relative amplitudes. To fix the magnitude of the vector {U }, we usually
normalise {U}T[M ]{U } = 1. So, in each iteration simply compute the RQ, and continue the
iterative process till the eigenvalue converges within the preset tolerance limit.
Algorithm for inverse iteration scheme

Step 1: Formulate the global [K] and [M] for the structure, i.e. form [K ]{U} = w2[M ]{U}.
Step 2: Assume a trial vector {X1}.
Step 3: Compute {R} = [M ]{X1}.
Step 4: Solve [K ] { X } = {R}.
Thus, this iteration scheme is similar to premultiplying the trial vector {X 1} by [K ]1[M ].
Instead of explicit computation of [K ]1, we have effective algorithms for solving [K ] { X } =
{R}, as discussed in Chapter 5 (see section 5.7).
Step 5: Obtain {X 2} from { X } such that
{ X 2 }T [M ]{X2} = 1
i.e.
1
{X 2} = {X } (6.162)
T
{ X } [ M ]{ X }
Step 6: Compute l = { X 2 }T [K]{X 2}.
Repeat steps (3) (6) till l converges to within a preset tolerance.

The inverse iteration scheme forms one of the most powerful tools for the determination
of the first eigenpair of a system. Since several modes usually participate in a vibration
response, we are usually interested in determining a few (and not just the lowest) eigenpairs.
It is reiterated that our trial vector always converges to the largest eigenpair. Thus, if we need
three eigenpairs and we take three distinct trial vectors, they will all converge to the largest
eigenpair only. Recall that if it so happens that the trial vector is exactly orthogonal to the actual
eigenvector, then, it can never converge to this eigenvector. This leads us to a scheme of
obtaining higher eigenpairs from this inverse iteration process.
GramSchmidt deflation
We have
[K ]{U } = w2[M ]{U} (6.163)
Let l1{U1}, the lowest eigenpair, be already known by, for example, inverse iteration. Then
from any trial vector we choose {X2} to get {U2}, subtract b1{U1}, where
b1 = {U1}T[M]{X2} (6.164)
{ X 2 } = {X2} b1{U1} (6.165)
Then, after proper iteration, { X 2 } will converge to {U2}. Let us see how this happens.
Let
{X2} = c1{U1} + c2{U2} + c3{U3} + L + cn{Un} (6.166)
Hence,
{U1}T[M ]{X2} = c1{U1}T[M ]{U1} + c2{U1}T[M ]{U2}
+ L + cn{U1}T[M ]{Un} (6.167)
Since the eigenvectors are [M ] orthogonal,
{U1}T[M ]{Uj} = 0 for j = 2, 3, L, n (6.168)
{U1}T[M]{U1} = 1 (6.169)
Therefore,
b1 = {U1}T[M ]{X2} = c1 (6.170)
Thus,
{ X 2 } = {X2} b1{U1}
= c1{U1} + c2{U2} + c3{U3} + L + cn{Un} c1{U1}
= c2{U2} + c3{U3} + L + cn{Un} (6.171)
We have thus eliminated {U1} component from the trial vector { X 2 } . Therefore, { X 2 } .
Therefore will actually be orthogonal to, and will not converge to, {U1}. In its implementation
in inverse iteration scheme, we need to perform this GramSchmidt deflation in every iteration.
Our trial vector will then converge to the eigenpair l2 and {U2}. After determining l2 and {U2}
to determine l3 and {U3}, we will use a deflated trial vector given by
{ X } = {X} b1{U1} b2{U2} (6.172)

where b2 is similarly determined.
Algorithm for inverse iteration with GramSchmidt deflation
Step 1: Formulate the assembled global [K] and [M] for the entire structure.
Step 2: Assume a trial vector {X1}.
If iterating for the lowest eigenpair, go to step 4. Else, go to step 3.

Step 3: Perform deflation, i.e.
i.e. m
{ X 1} = { X 1} C {U }
i =1
i i (6.173)
where bi = {Ui}T[M ]{X1} and m denotes the No. of eigenpairs already found.
Step 4: Compute {R} = [M ] { X 1}.
Step 5: Solve [K ] { X 1 } = {R} (similar to solution of static problem).
Step 6: Obtain {X2} from { X 1 } such that
{X 2 }T [M ]{X2} = 1 (6.174)
1
{X 2} = T (6.175)
{ X } [ M ]{ X }
Step 7: Compute the (m + 1)th eigenvalue, i.e. lm+1 = {X2}T[K]{X2}.

Repeat steps (3) (7) till a satisfactory convergence is achieved.
Simultaneous iteration method

In the GramSchmidt deflation technique, each trial vector was made orthogonal to all the
eigenvectors obtained earlier. Thus we can ensure that convergence was obtained to the
eigenpairs in a sequential manner, starting from the first till the mth desired vector. In the
present method, we start with a set of trial vectors and iterate on them simultaneously.
This iteration scheme is similar to the basic vector iteration method, but with multiple trial
vectors. The basic idea is to ensure that the individual trial vectors are mutually orthogonal in
each iteration so that they will converge to distinct eigenvectors. We have
[K ]{U } = w 2[M ]{U } (6.176)
The basic steps in this method are now outlined.
Step 1: Choose the trial vector set { X1}nm . For any arbitrary trial vector set, individual vectors
may not be orthogonal to one another. Let us assume that by some procedure, we are able to
extract an orthonormal set of vectors { X 1}nm from { X1}nm , i.e.,
{X 1}T [ M ]{ X1} = [ I ] (6.177)
Step 2: Compute {R}nm = [ M ]nn { X 1}nm.

Step 3: Solve for { X 2 } from [ K ]nn {X 2 }nm = {R}nm .
Step 4: Compute [M ]mm = { X 1}T [ K ]{ X1} . [l] is now an (m m) matrix, the diagonal
elements of which are our current approximation to the eigenvalues. The iterations are
continued till [l] converges fully. When converged, [l] will be of the form
M1 0

M2

% (6.178)

0 Mn
and the corresponding {X } will have the eigenvectors.

We see that these basic steps are identical to those of inverse iteration except that we
operate on multiple trial vectors. The significant deviation is in the procedure to get
orthonormalised vectors {X } from original { X } , which is now described*.
To get {X } from { X } , we use the following condition:
{ X }Tm n [M ]nn { X }nm = [ I ]mm (6.179)
Let
{ X }nm = { X }nm [ S ]mm (6.180)
Therefore,
{ X }T [M ]{ X } = [ S ]T { X }T [ M ]{ X }[ S ] = [ S ]T [ D][ S ] = [ I ] (6.181)
where
[ D] = { X }T [M ]{ X } (6.182)
Cholesky factorise (see Section 5.7)
[D] = [L][L]T (6.183)
Then, from Eq. (6.181),
[S ] = [L]T (6.184)
Thus, our method of finding the orthonormal vector set {X} from { X } involves use of
the inverse of the lower triangular matrix [L] obtained from Cholesky factorisation of
{ X }T [M ]{ X } . We therefore write
{ X } = { X }[ L]T (6.185)
T
Postmultiplying with [L] , we get
{ X }[ L]T = { X } (6.186)
Avoiding the explicit computation of the inverse [ L]1 , we prefer to obtain {X} as a solution
of Eq. (6.186) by backward substitution (see Section 5.7).
*The interested reader should refer V. Ramamurti, Computer Aided Mechanical Design and Analysis, McGraw-
Hill, New York, 1998.
Algorithm for simultaneous iteration scheme

Step 1: Formulate the assembled global stiffness and mass matrices [K] and [M].
Step 2: Choose a trial vector set { X1}nm .
Step 3: Compute [ D]mm = { X 1}Tmn [M ]nn { X1}nm .
Step 4: Cholesky factorise [ D]mm = [ L]mm [ L]Tmm .

Step 5: Solve for { X 1}nm from { X1}nm [ L]Tnm = { X1}nm (by back substitution).
Step 6: Compute [R ]nm = [ M ]nn { X 1}nm .
Step 7: Solve for { X 2 }nm from [ K ]nn { X 2 }nm = {R}nm .
Step 8: Compute the Rayleigh quotient matrix [l] = { X 1}mm [ K ]nn { X1}nm .
T
Repeat steps 38 till satisfactory convergence is obtained in [l], i.e. the off-diagonal
elements must be negligibly small, and the change in the diagonal elements from one iteration
to the next must be negligibly small.
We will now illustrate the simultaneous vector iteration procedure using the following
example.
Example 6.5. Axial vibrations of a fixed-free bar. Consider the 3-element model of a
uniform rod shown in Figure 6.15. Let L = 1 m, A = 30 106 m2, E = 2 1011 N/m2, r = 7800
kg/m3. Using consistent mass matrix, the assembled stiffness and mass matrices are readily
obtained as
0.360 0.180 0

[ K ] = 10 0.180
8
0.360 0.180 (6.187)

0 0.180 0.180
0.052 0.013 0

[ M ] = 0.013 0.052 0.013 (6.188)

0 0.013 0.026
which are the same as those in Example 6.4. Let us obtain the three eigenpairs of this system
using the simultaneous vector iteration method.
Iteration 1 Let
1 0 0

{ X 1}33 = 0 1 0 (6.189)

0 0 1
0.052 0.013 0

[ D ] = { X 1} [ M ]{ X 1} = 0.013 0.052 0.013
T
(6.190)

0 0.013 0.026
The orthonormalised trial vectors are given by
4.39 1.13 0.444

{ X 1} = 0 4.53 1.78 (6.191)

0 0 6.66
0.228 0 0

[ R ] = [ M ]{ X 1} = 0.057 0.221 0 (6.192)

0 0.0589 0.15
Solving for { X 2 } from [ K ] { X 2 } = [ R] , we have
0.158 0.155 0.0834

7

{ X 2 } = 10 0.190 0.311 0.167 (6.193)

0.190 0.343 0.250
Extracting the orthonormal set of vectors {X2}, from { X 2 }, we proceed to the second
iteration. After five iterations, our approximation to mode shapes is obtained as
1.50 3.59 2.37

{G} = 2.59 0.008 4.12
(6.194)
2.99 3.57 4.76
The natural frequencies are obtained from the diagonal elements of the Rayleigh quotient
matrix as
f1 = 1280 Hz, f2 = 4188 Hz, f3 = 7597 Hz (6.195)
The mode shapes corresponding to each frequency (i.e. respective column of [f]) match
with those obtained from generalised Jacobi scheme in Example 6.4.
Subspace iteration technique

This technique can essentially be viewed as being similar to the simultaneous iteration
technique, varying only in the way the orthonormalisation of trial vectors is done, i.e. how do
we get {X} from { X } ? In this scheme we evaluate
[ K ]mm = { X }Tmn [ K ]nn { X }nm (6.196)
[M ]mm = { X }Tmn [ M ]nn { X }nm (6.197)

Let us consider the eigenvalue problem
[ K ]mm {G}mm = [M ]mm [ M ]mm {G}mm (6.198)

This is said to constitute the subspace (m-dimensional) of the vector basis of the parent
problem (n-dimensional). Since the typical finite element mesh d.o.f. (n) is very large and we
are usually interested in a few eigenpairs only, m << n. The typical values may be n = 10,000
and m = 20! The size of the subspace eigenvalue problem is thus much smaller than the parent
problem.
Now, we solve for all the eigenpairs {M , G} , using, for example, the generalised Jacobi
transformation scheme. Obtain {X } as
{ X }nm = { X }nm {G}mm (6.199)
This is because
{ X }T [M ]{ X } = {G}T { X }T [M ]{ X }{G} = {G}T [ M ]{G} = [ I ] (6.200)
since the eigenvectors {f} are [ M ] orthogonal.
Thus, instead of using {X } = { X }[L]T, as in simultaneous vector iteration, we use here
{X } = { X }{f}. [f] has to be obtained as a solution of the subspace eigenvalue problem.
While a rigorous discussion on the convergence characteristics of various algorithms is
beyond the scope of the book, a useful guideline is that it is common practice to take 2m or
(m + 10), whichever is less, number of trial vectors, where m is the desired number of eigenpairs.
6.9 Transient Vibration Analysis

6.9.1 Modelling of Damping
We have so far discussed free vibration problems, where we studied the formulation of element
mass matrices as also several solution algorithms. We observed that damping had little effect,
in most cases, on the natural frequencies, and hence we studied undamped free vibration
problems. The results of such an analysis gave us insights into the behaviour of the system
when left free to vibrate on its own. In most practical situations, however, time dependent
excitation forces are always present and it is of interest to find the actual dynamic response
of the structure to such an excitation.
We also observed that a particular mode of vibration participates significantly in the
dynamic response if that natural frequency closely matches any frequency component in the
excitation. For such a situation, the dynamic response is limited essentially by the damping
present in the system. In a typical structure, there exist many sources of dissipation such as
viscous friction (for example, a shock absorber in an automobile), Coulombic friction, internal
material damping (hysterisis), air drag, etc. In most cases, these various mechanisms cannot
be reliably modelled and, quite often, an experimental estimate of overall damping present in the
system is our only guideline. Thus, an equivalent viscous damping model is commonly used
viscous friction force, being linearly proportional to the relative velocity, offers significant
Transient Vibration Analysis 273
modelling advantage. Analogous to the stiffness and mass matrices, we can use a damping
matrix and write the generalised governing equations as
[M ]{ X} + [C ]{ X } + [ K ]{ X } = {F (t )} (6.201)
While we have discussed, in detail, the derivation of stiffness and mass matrices, it is not
entirely obvious at this stage how the individual elements of the damping matrix [C]nn can be
obtained. We will address this issue a little later. Given the loading and the system parameters,
we are interested in determining the response {X(t)}. There are generally two categories of
methods of solving Eq. (6.201), viz., direct solution using a numerical integration scheme
(called direct integration methods) and, secondly, transform these equations into a reduced set
of uncoupled equations using the mode shapes prior to the numerical integration (called mode
superposition method). We will discuss both these methods in this chapter.
We have already noted that any deflected shape of the structure can be represented as
a linear combination of its eigenvectors, i.e.
{ X } = c1{U1} + c2 {U 2 } + " + cn {U n } (6.202)
or, since it is normally observed that a few modes (in the range of excitation frequencies)
predominantly contribute to the structural response,
X 1 (t ) u 1
u1
2
u1
m p1 (t )
1
X 2 (t ) u2 u2 u2 p2 (t )
= "
# # # # # (6.203)

X n (t ) un un un pm (t ) m1
nm
or, simply,
{X(t)} = [U ]{ p(t)} (6.204)
The pi are the mode participation factors indicating how much each mode contributes to
the response and these factors vary with time. {X(t)} contains all the nodal d.o.f. values for
the entire finite element mesh. Substituting from Eq. (6.204) in Eq. (6.201), we obtain
[ M ]{U }{
p} + [C ]{U }{ p } + [ K ]{U }{ p} = {F (t )} (6.205)
T
Premultiplying by {U} , we get
{U }T [ M ]{U }{
p} + {U }T [C ]{U }{ p } + {U }T [ K ]{U }{ p} = {U }T {F (t )} (6.206)
From the [M ] orthonormality property of eigenvectors,
{U}T[M ]{U } = [I ] (6.207)
From the Rayleigh quotient,
3 0

{U }T[K ]{U } = X2 (6.208)

0 2

Thus, {U }T[M ]{U } and {U }T[K ]{U } are both diagonal, and Eq. (6.206) can be rewritten as
3 0

{
p} + {U }T [C ]{U }{ p } + X2 { p} = {U } {F (t )}
T
(6.209)

0 2
If [C] is such that {U}T[C]{U} is also diagonal, we will have effectively decoupled the
equations of motion. The independent equations so obtained are typical second order, ordinary
differential equations (representative of spring-mass-damper system) for which the solution is
well known. For a single d.o.f. spring-mass-damper system, we have
mx + cx + kx = f (t ) (6.210)
Equation (6.210) can be rewritten as
f (t )

x + 2YX x + X 2 x = (6.211)
m
where x is the damping factor given by Y = c /2 (km) .
The m equations in modal coordinates contained in Eq. (6.209) can be written, in a
manner analogous to Eq. (6.211), as

pi + 2Yi Xi p i + Xi 2 pi = f (t ), i = 1, 2, ..., m (6.212)
The natural frequencies w and the mode shapes {f} required to obtain the right-hand side
transformed forces have already been determined during the solution of the undamped free
vibration problem. The modal damping factors can be directly determined experimentally and
plugged into Eq. (6.212), and these equations can be used to determine the dynamic response.
Such a technique is known as the mode superposition technique. Thus we see that we need
not form an explicit damping matrix [C]. However, if we do desire to determine the damping
matrix [C] we can, in principle, proceed in the following way*:
3 0

{U }T[C ]{U } = [c] = 2YX (6.213)

0 2
Let us assume that {U } is a square matrix containing all the n mode shapes of the system.
Thus,
[C ] = [U ]T[c][U ]1 (6.214)
If the entire set of mode shapes is known, we can then determine the desired damping matrix
[C ] from the experimentally measured modal damping factors. However, in practice, this is not
directly implemented but we use the relation
[U ]T[M ][U ] = [I] (6.215)
*The interested reader should refer to Dynamics of Structures, R.W. Clough and J. Penzien, McGraw-Hill,
New York, 1982.
Postmultiplying both sides of Eq. (6.215) with [U ]1, we get

[U ]1 = [U ]T[M ] (6.216)
T
Substituting in Eq. (6.214), (since [M ] = [M ]), we get
[C ] = [M ][U ][c][U ]T[M ] (6.217)
If m modes are likely to participate in the response and we have experimental estimates
of modal damping factors for these m modes, then we can write Eq. (6.217) as
[C ]nn = [M ]nn[U ]nm[c]mm[U ]Tmn[M ]nn (6.218)
Such an explicit formulation of the system damping matrix is needed when we solve for the
dynamic response {X(t)} from Eq. (6.201) using direct integration methods, as will be
discussed in Section 6.9.3. A system damping matrix [C] satisfying the orthogonality with
respect to the mode shapes [as in Eq. (6.213)] is referred to as Rayleigh (Proportional)
Damping. We shall use only this form of damping in the rest of our discussion.
Having addressed the issue of modelling system damping, we will now proceed to the
discussion of solution techniques for transient response determination. We will first discuss the
Mode Superposition technique and subsequently the direct integration methods.
6.9.2 The Mode Superposition Scheme

Using the orthogonality property of the mode shapes and assuming Rayleigh (proportional)
damping, we can transform the original equations of motion given in Eq. (6.201) into a set of
uncoupled single degree of freedom oscillator equations in modal coordinates as follows:

p1 + 2Y1X1 p1 + X12 p1 = f1 (t )

p2 + 2Y 2X 2 p 2 + X 22 p2 = f 2 (t )
(6.219)
#

pm + 2Y m X m p m + X m2 pm = f m (t )
where f1 (t ) , f 2 (t ) , etc. are obtained from {U }T {F (t )} . It must be observed here that we have
used the mode shapes obtained for the undamped system, assuming that the effect of damping
on the mode shapes is negligible.
These are a set of independent second order differential equations which can be easily
solved. For example, consider any one of the equations in Eq. (6.219) above:

pi + 2Yi Xi p i + X i2 pi = fi (t ) (6.220)
This is a second order, ordinary differential equation representing the equilibrium of a single
d.o.f. oscillator. Its solution can be calculated using the well-known Duhamel integral
1
fi (t ) sin X i (t U ) d U + B i sin X i t + Ci cos X i t
t
p i (t ) =
Xi 0
(6.221)
where ai, bi are determined from the initial conditions on pi and p i .

The initial conditions are, however, known in terms of physical coordinates {X0} and
{ X 0}. We need to estimate {p0} and { p 0 } from the known {X0} and { X0}. We have
{X } = [U ]{p} (6.222)
Therefore,
[U ]T[M ]{X } = [U ]T[M ][U ]{p} (6.223)
Since [U]T[M ][U ] = [I], we have

{p0} = [U]T[M ]{X0} (6.224)
{ p 0 } = [U ]T[M ] { X 0 } (6.225)
The integral in Eq. (6.221) may have to be evaluated numerically. Also, any other
convenient strategy can be adopted for the solution of Eq. (6.220).
Algorithm for mode superposition

Step 1: Find all the m eigenpairs wi, {Ui}, i = 1, 2, K, m.
Step 2: Evaluate {U }Tmn {F }n1 = { f }m1.
Step 3: Transform the initial conditions {X0} and { X 0 } into initial conditions in modal
coordinates { p0} and { p 0 }.
Step 4: Use the initial displacement and velocity conditions and the force { f (t )} to evaluate
{ p (t )}m1 , { p (t )}m1 .
Step 5: Use the transformations
{ X (t )}n1 = {U }nm { p}m1
{ X (t )}n1 = {U }nm { p }m1
to obtain actual displacement and velocity at all points of the structure.
Step 6: From X(t), obtain strains and then stresses as a function of time. Thus the complete
structural response is available.
Most of the computational effort usually goes into formulating the global [K] and [M] and
computing the eigenpairs. The number of modes (m) participating in the response is usually
very small, e.g. m = 10 or 20 while n, the d.o.f. of the finite element mesh, may be several
thousands. Thus mode superposition itself consumes very little extra computational effort and
hence is a popular scheme for evaluation of structural response.
We will now illustrate the mode superposition technique through the following simple
examples.
Example 6.6. Consider the undamped, 2-d.o.f. system shown in Figure 6.16. Find the
response of the system when the first mass alone is given an initial displacement of unity and
released from rest.
x1 x2
k k k
m m
Fig. 6.16 Two-d.o.f. system (Example 6.6).
The mathematical representation of the system for free, harmonic vibration is given by
2k k x1 m 0 x1
= X2 (6.226)
k 2k x2 0 m x2
For a nontrivial solution,
2k mX 2 k
=0 (6.227)
k 2k mX 2
yielding
3k
X1 =
k
, X2 = (6.228)
m m
1 1
The corresponding mode shapes are and .
1 1
If we organise the two mode shapes as columns of a matrix, [U ], then we have
1 1
[U ] = (6.229)
1 1
1 1 m 0 1 1 2m 0
[U ]T [ M ][U ] = = (6.230)
1 1 0 m 1 1 0 2m
Orthonormalising the mode shapes with respect to the mass matrix such that [U ]T[M ][U] = [I],
the identity matrix, we have the orthonormalised mode shapes as
1 1

2m 2m
[U ] = (6.231)

1 1

2m 2m
Using these shapes to transform the physical coordinates into modal coordinates, we write
1 1

x1 (t ) 2m 2m p1 (t )
{x} = = (6.232)

x2 (t ) 1 1 p2 (t )

2m 2m
and the modal equations are given by
k 3k

p1 + p1 = f1 (t ) ,
p2 + p 2 = f 2 (t ) (6.233)
m m
Since the external forces are zero, f1 (t ) = 0 = f2 (t ). Let us convert the initial conditions
in real physical coordinates x1 and x2 into those on modal coordinates p1 and p2. The initial
conditions in modal coordinates are given by
1 1 m
m 0
p10 2m 2m 1 2
= =
(6.234)
p20 1 1 0 m 0 m

2m 2m 2
Since x1 (0) = 0 = x2 (0) , p10 = 0 = p 20 . The solution for p1(t) can be written as
p1(t) = A cos w1t + B sin w1t (6.235)
from the initial conditions, p1 (0) = m / 2 and p1 (0) = 0, we have
A = m/ 2 , B = 0 (6.236)
Therefore,
m k
p1 (t ) = cos t (6.237)
2 m
Similarly,
m 3k
p2 (t ) = cos t (6.238)
2 m
Thus,
x1 (t )
{x(t )} = = [U ]{ p}
x2 (t )
1 1 m k
cos t
2m 2m 2 m
=
1
1 m 3k
cos t
2m 2m 2 m
1 k 1 3k
cos t + cos t
2 m 2 m
(6.239)
=
1 k 1 3k
cos t cos t
2 m 2 m
6.9.3 Direct Integration Methods

The governing equation for the transient dynamic response of a structure is given by
[M ] { X} + [C ] { X } + [K ]{X} = {F(t)} (6.240)

Direct integration methods, as the name suggests, attempt to directly integrate equation
(6.240) in time without any transformation of coordinates (e.g. to modal coordinates as done
in the mode superposition method). The key ideas in such a method of direct numerical
integration are as follows:
1. The above equation of motion is assumed to hold good at particular, discrete instants
of time tn, i.e. we write
[M ] { Xn } + [C ] { X n } + [K ]{Xn} = {F(tn)} (6.241)

2. The dynamic response is obtained by solution of the above equation.
3. The solution obtained in the above step holds good till we write the equation of motion
again at time tn+1 = tn+ Dt.
4. The time increment Dt is chosen to be sufficiently small to ensure accuracy and
stability of the solution.
5. Some solution methods require that Dt be less than a specific limit, i.e.
Dt Dtcrit (6.242)
without which the resulting solution could actually grow unbounded. These are
referred to as conditionally stable methods. It must be appreciated that this is purely
a numerical phenomenon. The actual dynamic response of the structure may be
perfectly normal.
Unconditionally stable methods are usually preferred. In such methods, Dt needs to be

chosen purely from accuracy point of view, which may permit much larger Dt to be
used.
6. Since we use a single Dt, it must be observed that all the individual equations
contained in Eq. (6.241) are integrated with the same Dt. Thus it is easy to see that
a uniform Dt may result in very poor integration of some of the high frequency (i.e.
low Tp) terms. On the other hand, Dt required to represent the highest frequency term
accurately may be infeasibly small. Moreover, as evident from Tables 6.1 and 6.2, for
a given finite element mesh, the highest frequencies may not be accurate after all.
Also, depending on the frequency content of the excitation, only a few modes are
predominant in the response. Thus the correct choice of Dt, keeping in mind all these
factors becomes crucial.
7. Since we attempt to directly integrate the equation of motion, we do not require that
the eigenvalues and eigenvectors be first determined.
Let us now study a simple method of solving the equation
[M ] { Xn } + [C ] { X n } + [K ]{Xn} = F(tn)} (6.243)
Central difference method
In this method, the time derivatives of {X }, viz., { X } and { X}, are approximated as
1
{ X n } = {{ X n +1} { X n 1}} (6.244)
2( 't )
1
{ Xn } = {{ X n +1} 2{ X n } + { X n 1}} (6.245)
('t )2
where we have used the standard Newtons central difference formulae. Substituting in
Eq. (6.243), we get
1 1
[ M ]{{ X n +1} 2{ X n } + { X n 1}} + [C ]{{ X n +1} { X n 1}}
( 't ) 2
2( 't ) 2
+ [ K ]{ X n } = {F (tn )} (6.246)
Rearranging the terms, we obtain
1 1 2
[M ] + [C ] { X n +1 } = {F (tn )} [ K ] [ M ] { X n }
('t ) 2( ' t ) ( 't )
2 2 2

1 1
[M ] [C ] { X n 1 } (6.247)
('t ) 2( 't )
2 2

For an undamped system,
1 2 1
[M ]{ X n+1} = {F (tn )} [ K ] [ M ] { X n } [ M ]{ X n 1} (6.248)
( 't ) 2
( 't ) 2
( 't ) 2
For a lumped mass matrix with no nonzero diagonal terms, the solution for {Xn+1} can
be directly written as
{ X n +1} = [ M ]1 ({F (tn )} [ K ]{ X n })( 't )2 + 2{ X n } { X n 1} (6.249)
where
3 0

1
[ M ]1 =
mii
0 2

For diagonal [M] and [C], the solution of Eq. (6.247) for {Xn+1} is straightforward.
Otherwise, the solution of Eq. (6.247) at each time step may be computationally very time
consuming.
Once we have {Xn+1}, we can proceed in exactly the same manner to find {Xn+2}, etc.
using the recurrence relation given in Eq. (6.247). If required, velocities and accelerations can
be readily obtained using Eqs. (6.244)(6.245).
We observe from Eq. (6.247) that the solution for {Xn+1} requires knowledge of
{Xn} as well as {Xn1}. When we start off the solution process at time t = 0, we know the
initial conditions. Thus to get the response at t = Dt, we need {X0} as also {XDt}. Obviously,
{XD t} has no physical meaning but can be computed from the known initial conditions {X0}
and { X 0} as
1
{ X 0 } = ({ X 't } { X 't }) (6.250)
2't
1
{ X0 } = ({ X 't } 2{ X 0 } + { X 't }) (6.251)
( 't ) 2
Eliminating {XDt} from these two equations, we get
( 't )2 { X0 } 2( 't ){ X 0 } = 2{ X 0 } + 2{ X 't } (6.252)

Thus,
( 't )2
{ X 't } = { X 0 } ( 't ){ X 0 } + {X 0 } (6.253)
2
It must be observed that the initial conditions are specified only on position and velocity
and { X0 } is not an initial condition but must be obtained from solving the equation of motion
at time t = 0, i.e.
[M ]{ X0 } + [C ]{ X 0 } + [ K ]{ X 0 } = {F (0)} (6.254)

Thus,
(
{ X0 } = [ M ]1 {F (0)} [C ]{ X 0 } [ K ]{ X 0 } ) (6.255)
We can now summarise our solution procedure.
Algorithm for central difference method

Step 1: Form assembled [K], [M] and [C ].
Step 2: From the prescribed {X0} and { X 0 }, find { X0 } using Eq. (6.255).
Step 3: Using Eq. (6.253), find {XDt}.

Step 4: Solve for {XDt}, {X2Dt}, {X3Dt}, etc. by repeated use of Eq. (6.247).
The major drawback of central difference method is that it is only conditionally stable,
2
i.e. ' t < , where wmax is the maximum frequency of [K]{U} = w2[M]{U}. If Dt chosen
X m ax
is larger than this, the solution obtained will be erroneous.
Since the central difference method discussed here finds {Xn+1} purely based on historical
information, i.e., {Xn}, {Xn1}, it is known as an explicit method. Implicit methods, on the other
hand, require the knowledge of time derivatives of {Xn+1} for the determination of {Xn+1}. We
will now discuss some popular implicit methods. Most of those methods are unconditionally
stable and place no restriction on Dt except as required for accuracy.
Average acceleration method (trapezoidal rule)

Let { Xn } and { Xn +1} be the accelerations at time tn and tn+1, respectively. In the interval
(tn, tn+1), the acceleration is assumed to be constant at 1/2({ Xn } + { Xn+1}), as indicated in
Figure 6.17. Therefore, we can write*
1
(
{ X n +1} = { X n } + { Xn } + { Xn +1}
2
) 't (6.256)
1 1
{ X n +1} = { X n } + { X n }'t +
2 2
({ X n } + { Xn +1} ) ('t )2 (6.257)
*Using the standard relations v = u + at and s = ut + at2.

Acceleration
xn+1
1
( + )
2 x n x n+1
xn
Time
0 tn tn+1
Fig. 6.17 Average acceleration scheme.
Substituting Eq. (6.256) in Eq. (6.257), we obtain
{ X n+1} = { X n } +
2
(
1
)
{ X n } + { X n+1} 't (6.258)
Rearranging the terms in Eq. (6.258), we get
2
{ X n +1} =
't
[{ X n+1} { X n }] { X n } (6.259)
From Eq. (6.257),
4 4
2 [
{ X n +1} = { X n +1} { X n }] { X n } { Xn } (6.260)
('t ) 't
Writing the equation of motion [Eq. (6.243)] at time tn+1, we have
[M ]{ Xn+1} + [C ]{ X n +1} + [ K ]{ X n +1} = {F (tn +1 )} (6.261)
Substituting from Eq. (6.259) and Eq. (6.260) in Eq. (6.261), we get
[K ]{ X n +1} = {Fn +1} (6.262)

where
4 2
[K ] = [M ] + [C ] + [ K ] (6.263)
( 't ) 2
't
4 4
{Fn +1} = {F (tn +1 )} + [ M ] {X n } + { X n } + { Xn }
( 't )
2
't
2
+ [C ] {X n } + { X n } (6.264)
't
We can readily solve Eq. (6.262) for {Xn+1}, and we now summarise the solution
procedure.
Algorithm for average acceleration method

Step 1: Form assembled [K ], [M ] and [C ].
Step 2: From the prescribed initial conditions {X0} and {X 0 }, find {X0 } using Eq. (6.255).
Step 3: Form [ K ] of Eq. (6.263).
Step 4: Form { F } and solve for {X} from Eq. (6.262) repeatedly for t = Dt, 2Dt, 3Dt, L
We observe that, for linear problems, [ K ] does not change with time. If the problem
involves material nonlinearity, [K ], and hence [ K ] would depend on {X}. If we are using
Cholesky factorisation scheme for solution of [ K ] {X } = [ F ] , for linear problems we need
to factorise [ K ] only once. Repeated solution for [ X ] is therefore fairly straightforward. We
also observe that [ K ] has off-diagonal elements whether or not [M ] is diagonal. Therefore,
there is no particular advantage in using diagonal, lumped mass matrix. Thus the more accurate
consistent mass matrices are commonly used with implicit integration methods. On the other
hand, the central difference method is computationally attractive if we use a diagonal lumped
mass matrix. This method does not require, unlike the central difference method, any special
starting procedure. It is unconditionally stable and, therefore, Dt can be chosen such that
acceptable accuracy is achieved.
Linear acceleration method

The average acceleration method assumed acceleration to be uniform in the interval (tn, tn+1).
Instead, we may assume that the acceleration varies linearly from { Xn } to { Xn +1} as indicated
in Figure 6.18.
Thus we have, for any time tn + t, the relations
U
{ Xtn +U } = { Xn } +
({ Xn +1} { Xn }) (6.265)
't
Integrating Eq. (6.265) with respect to time, we get
U2
{ X tn +U } = { X n } + { Xn }U + ({ Xn +1} { Xn }) (6.266)
2( 't )
U2 U3
{ X tn +U } = { X n } + { X n }U + { Xn } + ({ Xn +1} { Xn }) (6.267)
2 6( 't )
Acceleration
xn+1
xn
Time
0 tn tn+1
Fig. 6.18 Linear acceleration scheme.
At time tn+1 we, therefore, have

't
{ X n +1} = { X n } + { Xn } 't + ({ Xn +1} { Xn }) (6.268)
2
( 't ) 2 ( 't ) 2
{ X n +1} = { X n } + { X n } 't + {X n } + ({ Xn +1} { Xn }) (6.269)
2 6
Equation (6.269) can be rewritten as
6 }'t ( 't ) { X }

2
{ Xn +1} = { Xn } + { X n +1 } { X n } { X n n
( 't ) 2 2
6 6 6
= { X n +1} {X n } { X n } 2{ Xn } (6.270)
( 't ) 2
( 't ) 2 't
Using Eq. (6.270), Eq. (6.268) can be rewritten as
' t
{ X n +1} = { X n } + { Xn } ' t {X n }
2
't 6 6 6
+ { X n +1} {X n } { X n } 2{ Xn }
2 ('t ) 2
('t ) 2
't
3 3 ' t
= { X n +1} { X n } 2{ X n } {X n} (6.271)
't 't 2
The equation of motion at time tn+1 is given as
[M ]{ Xn+1} + [C ]{ X n +1} + [ K ]{ X n +1} = {F (tn +1 )} (6.272)
Substituting from Eqs. (6.270) (6.271) in Eq. (6.272), we get
[ K ]{ X n+1} = {F } (6.273)
where
6 3
[K ] = [K ] + [M ] + [C ] (6.274)
('t ) 2
't
6 6
{F } = {F (tn +1 )} + [ M ] { X n } + { X n } + 2{ Xn }
( 't )
2
't
3 't
+ [C ] { X n } + 2{ X n } + {X n } (6.275)
't 2
A commonly used variant of the linear acceleration method is the Wilson-q method, where
the acceleration is assumed to be linear, not just within the interval (t, t + Dt) but over an
extended time interval (t, t + qDt), where q 1, as shown in Figure 6.19. It can be shown
that if q 1.37, unconditional stability can be ensured. Thus, normally, q = 1.4 is used.
We proceed on exactly similar lines as for the simple linear acceleration scheme, except
that we consider the equations of motion at time tn + q Dt, rather than at tn + Dt. Thus we
use an extrapolated (often linearly) force vector at time tn + q Dt. The final equations are
obtained as
[ K ]{ X tn +R't } = {F } (6.276)
where
6 3
[K ] = [K ] + [M ] + [C ] (6.277)
(R't ) 2
R 't
(
{F } = {F (tn )} + R {Ftn + 't } {Ftn } )
6 6
+ [M ] {X n } + { X n } + 2{ Xn }
(R't )
2
R' t
(6.278)
3 R't
+ [C ] { X n } + 2{ X n } + { X n }
R' t 2
The solution steps in the Wilson-q method are identical to steps 14 of average
acceleration method except that we use Eqs. (6.277) (6.278) for [ K ] and [ F ] .
Acceleration
xn + q Dt
xn+1
xn
tn tn+1 Time
0 tn + q Dt
Fig. 6.19 Wilson-q linear acceleration scheme.
Newmark family of methods

The Newmark method can be viewed as a generalisation of the average acceleration method
and linear acceleration method. Let us now reproduce the basic equations used in these two
methods.
Average acceleration method

From Eqs. (6.256) (6.257),
1 1
{ X n +1} = { X n } + ( ' t ) { Xn } + { Xn +1} (6.279)
2 2
( 't )2 1 1
{X n +1} = { X n } + ('t ){ X n } + { X n } + { Xn +1} (6.280)
2 2 2
Linear acceleration method
From Eqs. (6.268) and (6.269),
1 1
{ X n +1} = { X n } + ( ' t ) 1 { X n } + { X n +1} (6.281)
2 2
( ' t ) 2 1 1
{ X n +1} = { X n } + ( 't ){ X n } + 1 { X n } + { Xn +1} (6.282)
2 3 3
In the Newmark method, we write
{ X n +1} = { X n } + ( 't ) (1 E ){ Xn } + (E ){ Xn +1} (6.283)
( 't ) 2
{ X n +1} = { X n } + ( 't ){ X n } + (1 2B ){ Xn } + (2B ){ Xn +1} (6.284)
2
where a and d are parameters that control integration accuracy and stability. It is readily seen
that
1 1
for E = and B = , Newmark method reduces to average acceleration method
2 4
1 1
for E = and B = , Newmark method reduces to linear acceleration method
2 6
In the Newmark method, we also use the equations of motion at time tn + Dt, i.e.
[M ]{ Xn +1} + [C ]{ X n +1} + [ K ]{ X n +1} = {F (tn + 't)} (6.285)

From Eqs. (6.283) (6.284), we can write
1 2B 1
{ Xn +1} = {X n} + { X } { X n } ( ' t ){ X n } (6.286)
2B B ( ' t )2 n +1
1 2B
{ X n +1} = { X n } + ( 't ) (1 E ){ Xn } + (E ) {X n }
2B
1 (6.287)
+ { X n +1} { X n } ( 't ){ X n }
2
B ( 't )
Substituting these in Eq. (6.285), we can write
[ K ]{ X n+1} = {F } (6.288)
where
1 E
[K ] = [K ] + [M ] + [C ] (6.289)
B ( 't ) 2
B ('t )
1 1
[ F ] = {Ftn + 't } + [M ] {X n } +
1
B ('t )
{ }
X n + { X n }
2B 1
B ('t )
2

E E ('t ) (6.290)
+ [C ] {X n } + {X n } + {Xn }
B ('t ) B 1 2 (E /B 1)

The solution steps are identical to those given earlier for average acceleration method
except for the fact that we use Eqs. (6.289)(6.290) for [ K ] and { F }.
Newmark originally proposed the unconditionally stable scheme with E 12 and B 14
(i.e. the constant average acceleration scheme). For general, a, d, it can be shown that the
method is unconditionally stable only if 2B E 1/2.
We will now illustrate the use of direct integration methods using the example of thermal
transients, viz., unsteady heat transfer in a fin.
Thermal TransientsUnsteady Head Transfer in a Pin-Fin 289
6.10 Thermal TransientsUnsteady Heat Transfer in a Pin-Fin

In Section 4.7, we discussed the solution of steady-state heat conduction problems with
convection from lateral surfaces. We studied the particular example of a pin-fin attached to a
wall. We shall now consider the situation when the temperature of the wall to which the fin
is attached is suddenly raised to and maintained at Tw, while the rest of the fin is initially at
ambient temperature. We expect that the temperature of the fin will also gradually increase with
time and we wish to determine the temperature in the fin as a function of time. Obviously, the
temperature in the fin stabilises over a period of time and we reach the steady-state solution
obtained in Section 4.7. The differential equation governing this unsteady heat transfer problem
is given by
2T P T
+ q = h(T T ) + Sc
k (6.291)
x 2
Ac t
where k = coefficient of thermal conductivity of the material, T = temperature, q = internal heat
source per unit volume, P = perimeter, Ac = cross-sectional area, h = convective heat transfer
coefficient, T = ambient temperature, r = density, and c = specific heat.
The weighted-residual statement can be written as
L 2T P T
0
W k 2 + q h(T T ) S c
x
Ac
dx = 0
t
(6.292)
Let us assume that there is no internal heat source (i.e. q = 0). The weak form of the
differential equation can be obtained by performing integration by parts as
L
T L W T L P L T
Wk x
0
0
k
x x
dx 0
W h(T T ) dx
Ac
0
WS c
t
dx = 0 (6.293)
i.e.
L dW T L P L T
0
k
dx x
dx + 0
W
Ac
hT dx +

0
WS c
t
dx
L
L P T
= 0
W
Ac
hT dx + Wk
x 0
(6.294)
The weak form for a typical mesh of n finite elements can be written as
n
A dW T A A T

P
1
0
k
dx x
dx + WA
0
c
hT dx +
0
WSc
t
dx

n A PhT T
A
= W
dx +

Wk
x 0
(6.295)
Ac
0
1
We observe that all the terms are the same as used earlier [see Eq. (4.159)] except for
the additional term involving the specific heat. For the standard linear element, we have
x x
T = 1 Ti + T j (6.296)
A A
Differentiating with respect to time, we get
x x
T = 1 Ti + Tj (6.297)
A A
The weight functions being the same as the shape functions, we have
x
W1 = 1 (6.298)
A
x
W2 = (6.299)
A
We can now perform the necessary integration for the additional specific heat term as
follows:
A T A x x x
W Sc t
0
1 dx = 0
Sc 1

1 Ti + T j dx
A A A
(6.300)
A T A x x x
W Sc t
2 dx = Sc 1 Ti + Tj dx (6.301)
0 0 A A A
Carrying out the integrations and rewriting the results in matrix form, we get the element
level specific heat matrix as
ScA 2 1 Ti
[c ] e
= (6.302)
6 1 2 Tj
The matrix [c]e is observed to be analogous to the element consistent mass matrix. Here too,
it is possible to use a lumped matrix similar to the mass lumping we discussed earlier.
Now the final element level equations are given by
k 1 1 PhA 2 1 Ti ScA 2 1 Ti Ph A /2 Q0

+ + = T +
A 1 1
6 Ac
1 2 T j 6 1 2 Tj Ac A /2 QA
(6.303)
Upon assembly, for a mesh of finite elements, we get
[ K ]{T } + [C}{T} = {F (t )} (6.304)
Thermal TransientsUnsteady Head Transfer in a Pin-Fin 291
For example, using the trapezoidal rule, the final equations can be readily obtained as
2 2
[ K ] + ' t [C ] {Tn +1 } = {Fn +1 } + {Fn } [ K ] ' t [C ] {Tn } (6.305)

We will illustrate the solution of this problem using the following example.
Example 6.7. Temperature distribution in a pin-fin. Consider a 1 mm diameter, 50 mm long
aluminium pin-fin as shown in Figure 6.20 used to enhance the heat transfer from a surface
wall maintained at 300C. Use k = 200 W/m/C, r = 2700 kg/m3, c = 900 J/kgC for
h T
Wall
Tw
L
Fig. 6.20 Temperature distrbution in a pin-fin (Example 6.7).
aluminium; h = 20 W/m2 C; and T = 30C. Assume that the tip of the fin is insulated. There
is no internal heat source. We will now study how the temperature in the fin varies with time
before settling to the steady state solution obtained in Example 4.10. We will use four equal
length elements to model the fin.
The assembled matrices can be obtained as follows:
1 1 0 0 0 2 1 0 0 0

1 2 1 0 0 1 4 1 0 0

200 (Q ) (0.001) (20) (0.0125)
[K ] = 0 1 2 1 0 + 0 1 4 1 0
0.0125 (6) (Q ) (0.0005)2
0 0 1 2 1 0 0 1 4 1

0 1 1 0 0 0 1 2
0 0
2 1 0 0 0

1 4 1 0 0
(2700) (900) (0.0125)
[C ] = 0 1 4 1 0
(6)
0 0 1 4 1

0 0 0 1 2

0.00625

0.0125
(Q ) (0.001) (20)
{F } = (30) 0.0125
(Q ) (0.0005) 2

0.0125

0.00625
Qwall

0

{Q} = 0 (6.306)

0

0
With Dt = 1 s, the nodal temperatures for the first few time steps are given in Table 6.3,
and the full solution is given in Figure 6.21.
Table 6.3 Nodal Temperatures in a Pin-Fin
Time (s) T1 T2 T3 T4 T5
1 300 147.58 39.23 30.73 30.11

2 300 166.95 84.25 37.76 31.74
3 300 186.34 101.46 57.04 39.91
4 300 195.88 118.83 70.77 56.61
5 300 204 130.8 86.35 70.57
6 300 209.98 141.96 99.23 85.13
7 300 215.3 151.21 111.43 97.73
8 300 219.81 159.6 122.06 109.4
9 300 223.84 166.93 131.69 119.69
10 300 227.39 173.51 140.2 128.95
Problems 293
300
200
Temperature
T1
T2
T3
T4
T5
100
0
0 10 20 30 40 50 60
Time (s)
Fig. 6.21 Transient temperature distribution in a pin-fin (Four-element model of Example 6.7).
SUMMARY
In this chapter, we have discussed the formulation and solution of transient dynamic problems
using FEM. We observed that the solution of transient dynamics problems is more time
consuming than that of steady state (static) problems. An effective scheme (HRZ scheme) for
generating lumped mass matrices has been discussed. Also, modelling of damping has been
addressed. Several algorithms for solution of dynamic equilibrium equations have been covered
in detail. Through the examples of structural dynamics and thermal transients, we have
demonstrated that finite element formulation can be carried out essentially on the same lines as
those of steady-state problems.
PROBLEMS
6.1 Consider the system shown in Fig. P6.1.
(a) Write down the stiffness and mass matrices of the structure.
2 6 6 6
2 2 3
Fig. P6.1 A 3-d.o.f. spring-mass system.

(b) It is known that the natural frequencies of the system are 1, 2 and 3 rad/s. Now
use the Sturm sequence method to estimate the second natural frequency. Use three
different trial values, less than, equal to, and greater than the exact value and show
what happens to the [D] matrix.
6.2 It is aimed at determining the largest eigenvalue and eigenvector of the matrix given below.
Perform the necessary calculations following the inverse vector iteration method.
20 10 0

20 10

Symmetric 20
6.3 For the system shown in Fig. P6.1, obtain the highest natural frequency by vector
iteration method.
6.4 Explain what would happen if the lowest eigenvalue of a system is zero and the inverse
iteration technique is applied. How will you overcome the difficulty?
6.5 Determine the two eigenvalues and eigenvectors corresponding to the two nonzero
masses, using the method of subspace iteration.
2 1 0 0 2 0 0 0

1 2 1 0 0 1 0 0
[K ] = , [M ] =
0 1 2 0 0 0 0 0

0 0 1 1 0 0 0 0
6.6 For the system shown in Fig. P6.6, determine the transient response over the interval 0
to t0 when a force Pc(t) acts at point C as shown.
Pc
2A0
A0
A B C P0
Pc
L L
t0 t
Fig. P6.6 Transient loading of a stepped bar.

CHAPTER 7
Application Examples
In this chapter, we present some examples of practical application of the various finite element
concepts discussed in the earlier chapters. We will first discuss the use of beam elements in
modelling the vibrations of an automotive crankshaft assembly. As an example of a two-
dimensional finite element analysis, we will then discuss axisymmetric finite element analysis
of pressure vessels.
7.1 Finite Element Analysis of Crankshaft Torsional Vibrations

Automotive crankshafts are subjected to fluctuating torques due to periodic strokes in the
cylinders. Torsional vibrations result from the twisting reaction created in rotating shafts due
to a fluctuating torque. Torsional vibration analysis is usually carried out by modelling the
crankshaft and driveline components as a set of lumped masses and springs. Each throw of
crankshaft is typically modelled as a set of masses and springs. Sometimes, these models
account for clutch hysteresis and universal joint disturbances. Journal bearings are modelled as
a set of linear springs and dashpots. The pulley and flywheel are modelled as lumped masses.
These models are usually restricted to analyse torsional vibrations only. Generally, torsion
modes are coupled with the bending modes in practical multicylinder engine crankshafts due
to their complex geometry. Three-dimensional vibration analysis of crankshafts thus becomes
necessary. Although 3d models could be used for analysis and design of real-life crankshafts,
a lot of time would be consumed in the preprocessing as well as repeated analyses in a design
cycle. Thus it is useful to develop a reasonably accurate one-dimensional model based on simple
beam elements.
We will now describe one such model of an automotive crankshaft. Generalised beam
elements (with six degrees of freedom per node) have been used to model the crankshaft.
Crankshaft main bearings have been modelled as linear springs and viscous dampers for
dynamic analysis. The details of the models are described in the following section.
295
296 Application Examples
7.1.1 Beam Element Model of Crankshaft Assembly

A typical crankshaft used in automobiles is shown in Figure 7.1. The web can be modelled
as rectangular blocks as shown in Figure 7.2. The dimensions of the rectangular blocks are
Side view Front view

Fig. 7.1 Typical automotive crankshaft.
Front view Side view Front view Side view
(a) Actual cheek (b) Simplified model (c) Finite element model
Fig. 7.2 Modelling of the cheek as a beam.
chosen such that the total area of these blocks equals the area of the web profile. This
simplification enables us to use beam elements with uniform cross-section. More generally, we
can estimate the various stiffnesses of the actual cross-section (viz., bending, torsional, etc.),
and use these values for the beam elements. The crankpins ( journals) are modelled using beam
elements whose cross-sectional area and the moment of inertia are taken to be equal to the
original crankpins ( journals). The flywheel and pulley are assumed to be cylindrical in shape
and are modelled as lumped masses.
The crankshaft main bearings are modelled as three sets of linear springs and dashpots
in the directions normal to the crankshaft axis. They are attached at the middle of the respective
Finite Element Analysis of Crankshaft Torsional Vibrations 297
crank-journals. It is assumed that the central set of spring-dashpot system supports 50% of
the bearing load, the remaining two support 25% each because of the parabolic distribution of
the oil film pressure along the crankshaft axis. The gas force and inertia force for a given
cylinder are assumed to be equally supported by the adjacent crank-journals. The spring
stiffness and damping coefficients are calculated using standard formulae*. A representative
beam finite element model of the system is shown in Figures 7.3 and 7.4. For assessing the
accuracy of the beam element model, a three-dimensional finite element model has also been
developed as shown in Figure 7.5.
Pulley end Flywheel end
(a) Actual crankshaft
(b) Simplified beam model
(c) FE model
Fig. 7.3 Modelling of the crankshaft as beams.
*J.S. Rao, Rotor Dynamics, New Age International, New Delhi, 1996.
Pulley end
Journal bearings
Ft/ N
Flywheel end
N: No. of nodes on
Fr/ N the crankpin
Forces for dynamic
analysis
Fig. 7.4 Representative beam element model of crankshaft.
No. of elements 8260

No. of nodes 12,355
Total d.o.f. 37,065
Flywheel
Pulley
Journal bearing
Fig. 7.5 Representative 3d-finite element model.

7.1.2 Results and Discussion

Free vibration analysis
The beam and 3-d element models are used to analyse the following five systems for their free
vibration behaviour:
System 1: A simple crankshaft (free-free) without flywheel, pulley and bearings.
System 2: System 1 is modified by including the journal bearings.
System 3: System 1 is modified by incorporating the flywheel and pulley.
System 4: System 3 is modified by modelling the journal bearings as linear springs.
System 5: System 4 is modified by modelling the reciprocating masses of the pistons.
In both the 3d model and the beam model, the mass is lumped on the
crankpins. This system is further used in dynamic response analysis.
The above systems are analysed in order to study the effect of journal bearings,
pulley and flywheel on the natural frequencies and mode shapes of the crankshaft system. The
results of free vibration analysis for these cases are compared in Tables 7.17.3. Conventional
Table 7.1 Comparison of Torsional Frequencies (Hz) for Beam and 3d Models
System 1 System 2
Mode 3d Beam Per cent 3d Beam Per cent
No. model model deviation model model deviation
1 764 693 9.3 790 710 10.3
2 1206 1291 7.0 1280 1370 7.03
3 1941 1612 17.0 1968 1646 16.4
4 2129 1673 21.4 2132 1673 21.5
5 3023 2335 22.8 3027 2338 22.8
System 3 System 4
Mode 3d Beam Per cent 3d Beam Per cent
No. model model deviation model model deviation
1 487 562 15.4 516 606 17.5
2 1093 1172 7.2 1132 1213 7.15
3 1487 1505 1.2 1491 1513 1.5
4 1892 2207 16.7 1922 2216 15.3
5 2924 3469 18.6 2966 3472 17.1
System 5
Mode 3d Beam Per cent
No. model model deviation
1 484 346 28.5
2 1097 1286 17.2
3 1423 1487 4.5
4 1828 1951 6.7
5 2896 3228 11.5
analysis of crankshaft vibrations considers pure torsional modes. The mode shape plots
(Figures 7.6 7.8), obtained using the beam element model show both the torsional and bending
deflections in the model.
Fig. 7.6 First mode shape of crankshaft.
Fig. 7.7 Second mode shape of crankshaft.

Fig. 7.8 Third mode shape of crankshaft.
Addition of the flywheel and pulley to the crankshaft significantly reduce the natural
frequencies of the system, as can be observed from the results of Systems 1 and 3. The
reduction in the natural frequency could be as much as 18%. It can be observed from the mode
shape plots that, with the introduction of pulley and flywheel, the nodes get shifted towards
the pulley, but the mode shapes do not change significantly. Incorporation of the reciprocating
mass of the pistons significantly reduces the natural frequencies of System 4. The reduction
in the natural frequencies is around 16% for beam model and 6% for the 3d model. In view
of the coupling between the torsion and the bending modes, it is expected that the bearings will
also affect the natural frequencies. From a comparison of results obtained for systems 1 and
2 (or 3 and 4), we observe that the natural frequencies may increase by as much as 6 7%.
It can be observed from the mode shape plots that for a given mode shape, the bending mode
shape alters due to the introduction of bearings while the torsion part remains unaffected.
Hence, for accurate study of crankshaft vibrations, bearings should also be taken into
consideration.
It can be observed that the beam element model overestimates (or) underestimates the
resonances by as much as (1620%). The 3-d finite element model, though requiring greater
computational resources, provides accurate estimate of the resonance frequencies and also
provides better insight into the coupled vibration behaviour.
7.1.3 Dynamic Response Analysis

The gas forces and the inertial forces due to the reciprocating masses will contribute to the
excitation forces on the crankshaft system. Fourier analysis is performed to calculate the first
20 harmonics of the excitation forces (Figure 7.9). The radial and tangential components of the
forces are applied on the crankpins as shown in Figure 7.4. These components are then used
to perform harmonic analysis of crankshaft to calculate the steady state vibration behaviour.
Dynamic response to individual harmonics is first obtained and then superposed to estimate the
60,000 9,000
50,000 8,000
7,000
Tangential force
40,000
Fourier coeff. (N)

6,000
30,000 5,000
20,000 4,000
3,000
10,000
2,000
0
1,000
10,000 0
0 2 4 6 8 10 12 14 0 1 2 3 4 5 6 7 8 9 10
Rotation angle (rad) Order
Fig. 7.9 Dynamic forces on crankshaft and their Fourier decomposition.
total response. Viscous damping in the torsional damper (2000 N s/m) as well as structural
damping in the material (damping ratio = 0.01) has been taken into account while calculating
the steady-state response.
The results for a given excitation order are presented in the form of stress as a function
of the speed of rotation of the crankshaft. The von Mises stress at the crankpin is depicted
in Figures 7.10 and 7.11.
107
2.5
von Mises equivalent stress (Pa)
1.5
0.5
0
12,000
10,000
10
Rot 8000 8
atio
n s 6000 6
pee
d ( 4000 4 ion
rpm citat
) 2000 2
d e r of ex
0 O r
0
Fig. 7.10 Dynamic stresses on crankpin (beam element model).

Axisymmetric Finite Element Analysis of a Pressure Vessel 303
107
6
von Mises equivalent stress (Pa)
0
12,000
10,000
10
Ro 8000 8
tati
on 6000 6
spe n
ed 4000 4 x c itatio
(rp
m) 2000 2 e r of e
0 0 Ord
Fig. 7.11 Dynamic stresses on crankpin (3d model).
The CPU time consumed for complete dynamic analysis of 3d model on Sun Ultrasparc
workstation was approximately 12 hours, while that for the beam model was approximately
12 minutes. However, it is observed that the beam element model predictions are inaccurate for
the stresses even in noncritical regions such as crankpin. It is also incapable of estimating
stresses at a critical location such as a fillet. Thus, even though a beam element model provides
an initial guess at the natural frequencies and dynamic stresses, a full 3-d finite element model
yields a much better insight into the complex mode shapes and critical stress regions of
practical automotive crankshafts.
7.2 Axisymmetric Finite Element Analysis of a Pressure Vessel

Some of the commonly found structures such as pressure vessels, rocket casings, domes,
submarine hulls, cooling towers, etc., possess near axisymmetric geometry. If we neglect small
holes, attachments, and so on, their geometry can be viewed as perfectly axisymmetric, i.e.,
these are solids of revolution. They are usually made of isotropic materials and are supported
all around the periphery. Thus the boundary conditions and material properties can also be
approximated as axisymmetric. When subjected to an axisymmetric loading such as internal
pressure/thermal gradient, the resulting deformation and stresses are also perfectly
axisymmetric. In their design, however, it is crucial to consider certain non-axisymmetric
loading cases such as those due to ground motion caused by earthquakes. Determination of
response may then become quite complex, but can be simplified by invoking the Fourier series.
The non-axisymmetric loading is first decomposed into several Fourier harmonics; the response
to each harmonic is then determined and, finally, the individual harmonic responses are
superimposed to get the complete system response. We now briefly discuss the formulation of
axisymmetric finite elements and their application to the stress analysis of a typical pressure
vessel.
7.2.1 Finite Element Formulation for Axisymmetric Loads

We assume that the body is axisymmetric with respect to the z-axis (e.g. a cylindrical vessel),
and each finite element is a circular ring as shown in Figure 7.12. Thus, all the various 2-d
Axis of
symmetry
Axis of
symmetry
(a) An axisymmetric structure (b) Typical ring finite element
Fig. 7.12 A typical pressure vessel.
finite elements discussed in Chapter 5, viz., three-noded triangle, six-noded triangle, four-noded
quadrilateral, eight-noded quadrilateral, etc. can be formulated as axisymmetric ring elements.
Each node is, in fact, a nodal circle. A typical isoparametric four node quadrilateral ring element
is shown in Figure 7.12. Using standard shape functions, the displacements of any point P
within the element are given, in terms of nodal displacements, as
u = S Niui, w = S Niwi (7.1)
where u and w correspond to radial and axial displacements respectively. The shape functions
are the same as those given in Chapter 5 [see Eqs. (5.43) (5.46)]. As a radial displacement
cannot occur without causing a circumferential strain, we observe that there will be four
nonzero independent strains unlike the case of plane stress/strain. The independent strains for
axisymmetric deformation are:
u , u v
Fr = FR = , Fz =
r r z
u v
Hr z = +
z r (7.2)
Using the shape function derivatives, we can express these strains in terms of nodal
displacements as follows:
{F} = [ B]{E }e (7.3)
The stresses are given by
v v
1 0
1 v 1 v
T r F r
v
T R E (1 v) 1
1 v
0 FR
=
T z (1 + v)(1 2v)
1
F
0 z

U rz H rz
Symmetric 1 2v
2(1 v)

i.e.,
{T } = [ D]{F} (7.4)
The element stiffness matrix is given by
[k]e = [B]T[D][B] dv = [B]T[D][B] (2p r) dr dz (7.5)
Typically, these are evaluated using appropriate order of Gauss quadrature rule.
7.2.2 Stress Analysis of a Pressure Vessel

A typical heavy duty cylindrical pressure vessel with threaded end closures (Figure 7.13)
is undertaken for linear, elastic, static stress analysis under an internal pressure of 2000 bar
(i.e., approx. 200 MPa). Axisymmetric ring elements of triangular cross-section are used.
Ignoring the helix angle of the thread, the problem is treated as axisymmetric.
For the typical discretisation shown in Figure 7.14 of the entire vessel, a total of about
5000 elements have been used with nearly 2750 nodes. An exploded view of the discretisation
is depicted in Figure 7.15. The resulting number of equations to be solved was about 5500
(= n), with a bandwidth (= b) of about 100 for the stiffness matrix. The Cholesky scheme
(LLT decomposition) has been used to solve the equations. Typical stress (sz, sr, sq) contours
are shown in Figure. 7.16.
Threaded end closure
Cylindrical vessel
Fig. 7.13 Structural details of a typical pressure vessel.
Threaded end closure Cylindrical vessel
Fig. 7.14 FE discretisation of pressure vessel.

Fig. 7.15 Discretisation details at the thread.
Fig. 7.16 Typical stress contours under internal pressure loading.

APPENDIX A
Suggested Mini-Project Topics
The mini-projects suggested here are aimed at giving the student a feel for the
practical applications of finite element analysis. We have observed in our
courses that such a term project helps significantly in reinforcing the students
understanding of the subject matter and helps build a proper perspective on
the subject. These projects are expected to be carried out using any typical
commercial finite element analysis software.
Project 1: Thermal Analysis of a Pressure Vessel

This project aims at conducting a thermal analysis of the pressure vessel shown in Figure MP.1.
You may wish to organise your tasks in the following manner:
Do the geometric modelling of the given pressure vessel. Model the pressure vessel
as an axisymmetric two-dimensional structure. Also model the threads.
At steady state, the wall temperatures inside and outside are Tin = 300C and
Tout = 50C. Find the temperature distribution in the walls of the vessel and the
threads.
Solve for the temperature distribution when the gas inside the pressure vessel is at
a temperature of 450C. Inside the vessel assume free convection and the convective
heat transfer coefficient to be 15 W/m2K. Outside, a coolant liquid is being forced
and the convective heat transfer coefficient is 100 W/m2K.
The coefficient of thermal expansion for the vessel is 11 106/C, and for the end
cover material is 19 106/C. Find the thermal stresses at steady state in the vessel.
Initially the system is at ambient temperature and suddenly the gas inside the pressure
vessel is charged to 450C. Find the temperature distribution in the vessel as a
function of time.
For the material of the pressure vessel, assume Youngs modulus Y = 2.1 1011 N/m2,
Poissons ratio = 0.3, and density = 7800 kg/m3. Assume ambient temperature as 25C.
309
310 Appendix ASuggested Mini-Project Topics
Cylindrical vessel 300 750
Threaded end
closure
f 850
f 90
2400
2600
Thread details
Cylindrical vessel
14
23
f 887
f 914 f 904 f 897
f 890
Fig. MP.1 Thermal analysis of a pressure vessel.
Project 2: Structural Dynamic Analysis of a Pressure Vessel

The objective of this project is to perform static and dynamic stress analysis of a pressure
vessel. You may organise your task as follows:
Do the geometric modelling of the given pressure vessel as axisymmetric structure.
Let the gas pressure inside the vessel be 200 MPa. Find the stresses in the wall and
the threads. How do the threads share the load?
An explosion occurs inside the vessel. The variation of pressure with time due to this
explosion is given in Figure MP.2. Perform transient analysis for the given pressure
profile.
The vessel needs to be certified for earthquake resistance. Consider a typical
earthquake ground motion record and determine the structural response.
300 750
Cylindrical vessel
Threaded end
closure
f 850
f 90
2400
2600
13.8
Pressure (MN/m2)
6.9
0 100 200
Time(ms)
Fig. MP.2 Structural analysis of a pressure vessel.
Project 3: Dynamics of a Scooter Frame

Figure MP.3 shows a simplified model of a typical tubular frame of a two-wheeler (Scooter).
Model the system using spring, rigid mass and general beam elements, and determine the first
few natural frequencies and mode shapes. Assume that the vehicle goes on a road bump as
shown in the figure with a forward speed of 25 kmph. Determine the response of the system.
0
20
0
20
500
700
400
500 45
45
1055 NN/m
k fkf ==10 /m
5
105 N
kkrr = 10 /m
N/m
100
300
All dimensions are in mm

Tubular steel members;
dia. 50 and wall thickness 2
Fig. MP.3 Dynamics of scooter frame.
Project 4: Automotive Chassis Dynamics

A typical automotive chassis is shown in Figure MP.4(a). Although an actual chassis is
complicated, for the present project, we can model all the structural members of the chassis
using the same cross-section.
Typical Bimensions
Side width (Rear) = 1.5 m Curb weight = 600 kg
Side width (Front) = 1 m This mass is the sum of two discrete masses:
Wheel base = 2.5 m Engine weight = 300 kg
End-to-end length = 4 m (This weight is assumed to act at point A)
Distance between Body frame weight = 300 kg
mid-supports = 1 m (This weight is assumed to act at point B,
located 0.75 m from the rear axle.)
W
Curb
weight
W=5 cm
H H=10 cm
t=1 cm
Section
1.5
m
Ra
L
Fa Rb
A2 B Rc R
A1
Fb
1m
0.75 m
F 1m
2.5 m
4m
Fig. MP.4(a) Typical ChassisThe front axle is located at F and the rear axle at R. The CG of the
vehicle is at B.
Static Loads
Rear passenger weights (3 nos.@75 kg): Assumed to act at points Ra, Rb, and Rc.
Front passenger weights (2 nos.@75 kg): Assumed to act at points Fa and Fb.
Seat-spring stiffness Ksf = 90 kN/m. The luggage is modelled as a lumped mass (70 kg) acting
at L.
Suspension
The chassis is supported over four tyres through suspension arrangements. The suspension
system can be modelled as shown in Figure MP.4(b).
v
Chassis Front suspension
Rear Front Kfs = 19,000 N/m
Cfs = 70 N/m/s
Krs Shock Rear suspension

ar Kfs af
absorbers Krs = 20,000 N/m
Crs = 100 N/m/s
Unsprung mass at each of the

Unsprung mass four tyre points =
(Tyre mass) + (Axle mass/2)
Mass of axle = 100 kg

Tyres
Mass of each tyre = 35 kg
Ktyre = 100 N/m
Ctyre = 200 N/m/s
Road excitation
Fig. MP.4(b) Typical chassisSuspension details.
You may organise your project tasks as follows: Model the chassis, lumped masses (engine
mass, body-frame mass, passenger mass, axle-tyre mass and luggage). Determine the natural
frequencies and mode shapes of the chassis structure along with all the masses. For the first
five modes, list the natural frequencies and plot the corresponding mode shapes.
Subject the (chassis structure + suspension) to deterministic road excitations.
Figure MP.4(c) shows typical ride oscillation modes of the vehicle, as it navigates different
road-profiles.
In order to simulate the excitation which the tyres receive from the road, we apply four
displacement profiles, one each at the four tyres.
The displacements can be approximated as follows:

TFL = A sin (w t + q1) [displacement at front-left tyre]
T FR = A sin (w t + q2) [displacement at front-right tyre]
TRL = A sin (w t + q3) [displacement at rear-left tyre]
T RR = A sin (w t + q4) [displacement at rear-right tyre]
Circular frequency w = (2p /l) (vehicle speed)
Vehicle speed = 60 km/hr
Wavelength l = 2.5 m (same as wheel base)
Amplitude of road bumps A = 0.05 m
Front
Bounce
Roll
Pitch
Rear
Bounce
Fig. MP.4(c)
All-Four-Bounce Pitch
All q angles = 0. q1 = 0, q2 = 0
Both the axles heave up and down in phase. q3 = 90, q4 = 90
Roll
a a
b c
c d
d
a
a - c b- d b - d
b b
c dd
q1 = 0, q2 = 90, q3 = 0, q4 =90
For the above scenarios, determine
(i) the acceleration levels and response at the passenger locations
(ii) chassis roll
(iii) chassis pitch
Project 5: Analysis of a Turbine Disk

A typical disk used in gas turbines is shown in Figure MP.5. Perform axisymmetric finite
element analysis of the disk, under centrifugal forces when the disk rotates at a steady speed
N. rpm.
45
7.5
15
5
10
25 20
R35
R30
R35
R10 300
150
140
105
100
65
Fig. MP.5 Typical turbo machinery disk.

Project 6: Dynamic Analysis of a Building

Consider a simplified model of a typical building as shown in Figure MP.6. Properties of
concrete can be taken as follows: Youngs modulus = 1.4 1010 N/m2; Poissons ratio = 0.3;
density = 2240 kg/m3; assume proportional damping (0.2 [K] + 0.4 [M]).
Model the system using beam/plate elements and determine the first few natural
frequencies and the corresponding mode shapes.
There are two rotating machinery operating on the second floor. The specifications
are as given below: weight of each machine = 5000 N; support spring stiffness
= 800 kN/m; unbalanced mass on the rotor = 1 kg at an eccentricity of 0.1 m;
rotational speed = 3000 rpm. Determine the steady state response of the system when
these machines are in operation.
4m
0.25 m
0.3 0.3 m2 4m
4m
8m
12 m
Fig. MP.6 Dynamic analysis of a typical building structure.

Project 7: Thermal Analysis of an IC Engine Cylinder

Consider a simplified model of an IC engine cylinder shown in Figure MP.7.
Model the system using finite elements and find the steady state temperature
distribution given that Ti = 200C and T = 30C. Assume k = 45 W/mC and
h = 20 W/m2C.
Use the typical temperature distribution over a cycle as shown in the figure and
perform a transient dynamic analysis for temperature in the walls of the cylinder.
Design appropriate fins on the cylinder block outer surface so as to enhance heat
transfer. The number of fins per surface must be limited to 3, and each fin may be
of thickness 3 mm and length not more than 20 mm.
5
180 5
f
50
370
f 80

Cyclic piston surface temperature

194
192
Surface temp. C
190
188
186
184
182
0 100 200 300 400
Angular timing (in degrees)
Fig. MP.7 Thermal analysis of a typical IC engine cylinder.
Project 8: Stress Concentration

Consider the stress distribution around a small circular hole in a large plate subjected to uni-
axial tension. It is well known that the theoretical stress concentration factor for this case is
3. When we consider an elliptical hole, theoretical studies show that stress concentration gets
worse if we have the major axis perpendicular to the load. If the major axis is however aligned
with the direction of loading, then there is significant improvement in stress distribution.
However, circular holes are easier to make compared to elliptical holes. Thus, in practice, one
can think of approximating the ellipse with smaller circular holes on either side of the original
circular hole as shown in Figure MP.8.
Model the plate with the central circular hole using standard two-dimensional
elements (such as 8/9-noded quadrilateral; 6-noded triangle) and determine the
stresses around the hole.
Consider additional circular holes as indicated in the figure and redo the problem. Find
(if necessary, by trial and error) appropriate size/location/number of the holes required
to get near-uniform stress distribution.
Fig. MP.8 Plate with holes.
Project 9: Dynamics of a Hard Disk Drive Read/

Write Head Assembly
Figure MP.9 shows a simplified version of a typical Read/Write Head assembly used in
computer hard disk drives. The motor is at the left end of the system while the R/W head itself
is mounted at the right end. The assembly is rotated about an axis perpendicular to the plane
of the paper through an appropriate angle to reach the desired location (track) on the disk. The
track density, i.e. Number of tracks per mm is ever increasing and the response time desired
is ever decreasing. It is therefore important that, when the assembly is moved to a desired
location, the residual vibrations at the location of the R/W head be as low as possible. The
system is actually very complex, with natural modes of vibration in two planes of bending and
twisting, all coupled. Model the system using appropriate finite elements and determine the first
few natural frequencies. Next, assume that, under a command to seek a particular track, the
R/W head assembly is rotated by 60 in 10 ms through uniform acceleration/deceleration
motion (i.e., triangular velocity profile). Determine the residual vibrations at the tip, carrying the
R/W head. For the purpose of this analysis, assume that in the 60 position the system behaves
as if its left end is clamped.
Al (thickness 1 mm) Steel (thickness 0.3 mm)
f 15 f5
f5
R5 f1
R 10
20 2
R 10
3
10 10 5
5
1
f3
75
100
All dimensions are in mm
Fig. MP.9 Hard disc drive R/W head assembly.

APPENDIX B
Review of Preliminaries
B1.1 Matrix Algebra

A system of linear algebraic equations given by
a11x1 + a12x2 + L + a1nxn = b1
M M M
an1x1 + an2x2 + L + annxn = bn
can be represented in compact matrix notation as
[A]nn {X }n1 = {b}n1
FEM helps us to convert governing differential equations into algebraic equations and hence we
use matrix notation extensively. The system of equations corresponding to a finite element mesh
are typically sparse, i.e. many coefficients aij are zero and, in particular, the equations are banded,
i.e. nonzero coefficients are clustered around the principal diagonal as indicated in Figure B1.1.
0 d
an t h
BBand
i d
w
width
Carpet width
Carpet wid th
Fig. B1.1 Sparse, banded matrix.

321
322 Appendix BReview of Preliminaries
The rank of a matrix [A] is defined as the order of the largest nonzero determinant (or)
the maximum number of linearly independent columns in [A]. Differentiation and integration of
a matrix [A] refers to these operators being applied on each and every element aij of the matrix.
Thus,
a11 a1n
x " x
[ A]
= # #
x
an1 ann
x " x

a dx " dx

11 a 1n

[ A] dx = # #

a dx " ann dx

n1
A matrix eigenvalue problem is represented as
[A]nn{X}n1 = l{X}n1
or
([A] l[I]){X} = {0}
The nontrivial solution set consists of n paris of li, {Xi}, where li, are referred to as the
eigenvalues of [A] and {Xi} are the corresponding eigenvectors. If [A] is real and symmetric,
li will be real. If, in addition, [A] is positive definite, all li,s are positive. The corresponding
eigenvectors {Xi} are orthogonal i.e., {Xi}T{Xj} = 0 (i j).
B1.2 Interpolation
Understanding the process of interpolation is very useful in finite element method. We assume
that at certain xi, the function value, fi = f (xi) are known and we wish to find an approximate
estimate of the function value at an intermediate x. Typically, a polynomial pn(x) of degree n
is fit to the data given such that
pn(xi) = fi, i = 1, 2, L, n +1
Polynomial fit is normally chosen because it is convenient to differentiate/integrate a polynomial.
Also, they can approximate any continuous function to the desired accuracy. A polynomial fit
pn as above is unique, which implies that pn obtained by any method will be the same. The
various methods differ in their particular forms. One standard method is the Lagrange method.
The basic idea is to write
n +1
pn ( x) = i =0
fi Li ( x)
Appendix BReview of Preliminaries 323
where Li are known as the Lagrange polynomials given by
( x x1 )( x x2 ) " ( x xi 1 )( x xi +1 ) " ( x xn +1 )
Li ( x) =
( xi x1 )( xi x2 ) " ( xi xi 1 )( xi xi +1 ) " ( xi xn +1 )
It would perhaps be intuitive that to improve the accuracy of fit, we need only increase
n, the degree of the polynomial, used to fit the data. However, this is not generally true.
Numerical instabilities may cause the polynomial fit pn(x) to oscillate significantly between the
1
points xi. As an example, let us attempt to fit the function f ( x) = in the interval
1 + x2
(5, 5). The curve fit obtained using Lagranges method, as described above, is given in
Figure B1.2. The original f (x) and the polynomial curve fit pn(x)(n = 6 and 12) are compared
in the figure. It is observed that significant oscillations are present in p(x).
Fig. B1.2 Curve fit using a single polynomial.
Instead of a single, high degree polynomial curve fit in the entire interval a x b,
we may subdivide the interval into many segments and fit several low degree polynomials
(each valid in one segment). One such method uses cubic splines where a cubic curve is fit
within each segment. Such a piecewise curve fitting is far superior to fitting a single, high
degree polynomial for the entire region, as this is not subject to numerical instabilities (or)
oscillations.
APPENDIX C
Typical Finite Element Program
Here we illustrate the computer implementation of typical isoparametric finite

element formulation through the example of a nine-noded quadrilateral
element for 2-d problems. It is assumed that a certain main program is
available which interfaces with the subroutines given here. Clarity of
implementation has been our focus while writing this code in Fortran-90.
Variable name Explanation

Ngauss No. of Gauss Points
Xl, Yl X, Y Coordinates of Nodes in Global Frame
Thc Thickness of the Element
Place Array of Locations of Gauss Points
Wgt Array of Weights of Gauss Points
Jac Jacobean
En Shape Functions
B Strain-displacement Relation Matrix
D Stress-strain Relation Matrix
Ek Element Stiffness Matrix
Subroutine Quad9(Ngauss,Xl,Yl,Thc,Place,Wgt)
Implicit Double Precision (A-h, O-z)
Double Precision Jac
Common/Q9/En(9),Jac(2,2),B(3,18),D(3,3),Ek(18,18)
Dimension Xl(9),Yl(9), Btdb(18,18),Place(3,3),Wgt(3,3)
C
C INITIALISE THE MATRICES
C
DO 10 K=1,18
DO 10 L=1,18
EK(K,L)=0.0
324
Appendix CTypical Finite Element Program 325
10 CONTINUE
DO 20 I=1,3
DO 20 J=1,3
PLACE(I,J)=0.0
20 WGT(I,J)=0.0
PLACE(1,2) = -0.5773502692
PLACE(2,2) = 0.5773502692
PLACE(1,3) = -0.7745966692
PLACE(3,3) = 0.7745966692
WGT(1,1) = 2.0
WGT(1,2) = 1.0
WGT(1,3) = 5.0/9.0
WGT(2,2) = 1.0
WGT(2,3) = 8.0/9.0
WGT(3,3) = 5.0/9.0
C
C GAUSS QUADRATURE LOOP
C
DO 90 NA=1,NGAUSS
PXI=PLACE(NA,NGAUSS)
DO 80 NB=1,NGAUSS
PET=PLACE(NB,NGAUSS)
C
C WE HAVE NOW PICKED UP ONE GAUSS POINT WHOSE COORDINATES ARE
C x (PXI), h(PET)
C AT THIS GAUSS POINT, WE EVALUATE [J], [B] MATRICES USING SHAPE
C ROUTINE
C
CALL SHAPE(PXI,PET,XL,YL,DETJAC)
DV=WGT(NA,NGAUSS)*WGT(NB,NGAUSS)*THC*DETJAC
CALL TRAMUL(B,D,B,3,18,18,BTDB)
DO 70 I=1,18
DO 70 J=1,18
EK(I,J)=EK(I,J)+BTDB(I,J)*DV
80 CONTINUE
90 CONTINUE
C
C WE HAVE SUMMED UP OVER ALL THE GAUSS POINTS AND OBTAINED
C ELEMENT STIFFNESS MATRIX [EK]
C
RETURN
END
SUBROUTINE SHAPE(PXI,PET,XL,YL,DETJAC)
IMPLICIT DOUBLE PRECISION (A-H, O-Z)
DOUBLE PRECISION NXI(9),NET(9),JAC,JACINV(2,2)
DIMENSION XL(9),YL(9),B1(3,4),B2(4,4),B3(4,18)
COMMON/Q9/EN(9),JAC(2,2),B(3,18),D(3,3),EK(18,18)
326 Appendix CTypical Finite Element Program
C
C THE SHAPE FUNCTIONS EVALUATED AT x (PXI) and h (PET)
C
EN(9)=(1.0-PXI*PXI)*(1.0-PET*PET)
EN(5)=0.5*(1.0-PXI*PXI)*(1.0-PET)-0.5*EN(9)
EN(6)=0.5*(1.0+PXI)*(1.0-PET*PET)-0.5*EN(9)
EN(7)=0.5*(1.0-PXI*PXI)*(1.0+PET)-0.5*EN(9)
EN(8)=0.5*(1.0-PXI)*(1.0-PET*PET)-0.5*EN(9)
EN(1)=0.25*(1.0-PXI)*(1.0-PET)-0.5*(EN(8)+EN(5))-0.25*EN(9)
EN(2)=0.25*(1.0+PXI)*(1.0-PET)-0.5*(EN(6)+EN(5))-0.25*EN(9)
EN(3)=0.25*(1.0+PXI)*(1.0+PET)-0.5*(EN(6)+EN(7))-0.25*EN(9)
EN(4)=0.25*(1.0-PXI)*(1.0+PET)-0.5*(EN(8)+EN(7))-0.25*EN(9)
C
C THE DERIVATIVES OF SHAPE FUNCTIONS; NXI REPRESENTS DERIVATIVE WITH x
C NET REPRESENTS DERIVATIVE WITH h
C
NXI(9)=-2.0*PXI*(1.0-PET*PET)
NET(9)=-2.0*PET*(1.0-PXI*PXI)
NXI(5)=-PXI*(1.0-PET)-0.5*NXI(9)
NXI(6)=0.5*(1.0-PET*PET)-0.5*NXI(9)
NXI(7)=-PXI*(1.0+PET)-0.5*NXI(9)
NXI(8)=-0.5*(1.0-PET*PET)-0.5*NXI(9)
NET(5)=-0.5*(1.0-PXI*PXI)-0.5*NET(9)
NET(6)=-PET*(1.0+PXI)-0.5*NET(9)
NET(7)=0.5*(1.0-PXI*PXI)-0.5*NET(9)
NET(8)=-PET*(1.0-PXI)-0.5*NET(9)
NXI(1)=-0.25*(1.0-PET)-0.5*(NXI(8)+NXI(5))-0.25*NXI(9)
NXI(2)= 0.25*(1.0-PET)-0.5*(NXI(6)+NXI(5))-0.25*NXI(9)
NXI(3)= 0.25*(1.0+PET)-0.5*(NXI(6)+NXI(7))-0.25*NXI(9)
NXI(4)=-0.25*(1.0+PET)-0.5*(NXI(8)+NXI(7))-0.25*NXI(9)
NET(1)=-0.25*(1.0-PXI)-0.5*(NET(8)+NET(5))-0.25*NET(9)
NET(2)=-0.25*(1.0+PXI)-0.5*(NET(6)+NET(5))-0.25*NET(9)
NET(3)= 0.25*(1.0+PXI)-0.5*(NET(6)+NET(7))-0.25*NET(9)
NET(4)= 0.25*(1.0-PXI)-0.5*(NET(8)+NET(7))-0.25*NET(9)
C
C INITIALISE [JAC] and [B]
C
DO 10 I=1,2
DO 10 J=1,2
10 JAC(I,J)=0.0
DO 20 I=1,3
DO 20 J=1,18
Appendix CTypical Finite Element Program 327
20 B(I,J)=0.0
C
C FORMING THE JACOBIAN
C
DO 30 L=1,9
JAC(1,1)=JAC(1,1)+NXI(L)*XL(L)
JAC(1,2)=JAC(1,2)+NXI(L)*YL(L)
JAC(2,1)=JAC(2,1)+NET(L)*XL(L)
JAC(2,2)=JAC(2,2)+NET(L)*YL(L)
30 CONTINUE
C
C FINDING THE DETERMINANT OF JACOBEAN AND INVERSE OF JACOBEAN
C
DETJAC = JAC(1,1)*JAC(2,2)-JAC(1,2)*JAC(2,1)
C
DUM1 = JAC(1,1)/DETJAC
JAC(1,1)= JAC(2,2)/DETJAC
JAC(1,2)=-JAC(1,2)/DETJAC
JAC(2,1)=-JAC(2,1)/DETJAC
JAC(2,2)= DUM1
C
C
DO 40 I=1,3
DO 40 J=1,4
40 B1(I,J) = 0
B1(1,1)=1.0
B1(2,4)=1.0
B1(3,2)=1.0
B1(3,3)=1.0
DO 50 I=1,4
DO 50 J=1,4
50 B2(I,J)=0
DO 60 I=1,2
DO 60 J=1,2
60 B2(I,J)=JAC(I,J)
DO 70 I=3,4
DO 70 J=3,4
70 B2(I,J)=JAC(I,J)
DO 80 I=1,4
DO 80 J=1,18
80 B3(I,J)=0
DO 90 J=1,9
L=2*J
K=L-1
B3(1,K)=NXI(J)
328 Appendix CTypical Finite Element Program
B3(2,k)=NET(J)
B3(3,L)=NXI(J)
B3(4,L)=NET(J)
90 CONTINUE
CALL MATMUL(B1,B2,3,4,4,B1B2)
CALL MATMUL(B1B2,B3,3,4,18,B)
RETURN
END
C
C
SUBROUTINE TRAMUL(A,B,C,L,M,N,D)
IMPLICIT DOUBLE PRECISION(A-H,O-Z)
DIMENSION A(L,M),B(L,L),C(L,N),D(M,N),AT(20,20),ATB(20,20)
C
C PROGRAM TO CALCULATE (D)=(A)T*(B)*(C)
C
DO 10 I=1,L
DO 10 J=1,M
10 AT(J,I)=A(I,J)
DO 20 I=1,M
DO 20 K=1,L
ATB(I,K)=0.0
DO 20 J=1,L
20 ATB(I,K)=ATB(I,K)+AT(I,J)*B(J,K)
DO 30 I=1,M
DO 30 K=1,N
D(I,K)=0.0
DO 30 J=1,L
D(I,K)=D(I,K)+ATB(I,J)*C(J,K)
30 CONTINUE
RETURN
END
SUBROUTINE MATMUL(A,B,L,M,N,C)
IMPLICIT DOUBLE PRECISION(A-H,O-Z)
DIMENSION A(L,M),B(M,N),C(L,N)
C
C PROGRAM TO CALCULATE [C] = [A]*[B]
c
DO 10 I=1,L
DO 10 K=1,N
C(I,K)=0.0
DO 10 J=1,M
10 C(I,K)=C(I,K)+A(I,J)*B(J,K)
RETURN
END
Index
Assembly, 48, 117 Eigenvector, 256, 259, 265, 273

Average acceleration method, 282, 287, 288 Element
Axisymmetry, 146, 147, 168169, 303, 305 bar, 90, 91, 93, 101
beam, 91, 117, 118, 238
bilinear, 153, 180182, 210
Bandwidth, 207 conductance matrix, 59
Bar element, 90, 91 CST, 151, 173, 174, 178, 226
linear, 93 equations, 48
quadratic, 101 force vectors, 51
Beam element, 91, 117 four-noded rectangular, 152, 179
Boundary condition, frame, 125
essential, 34, 38, 39 isoparametric, 156, 184, 185
natural, 34, 38, 39 quad 4, 159, 160, 182188, 216
Boundary element method, 17 quad 8, 162, 163, 188190, 217
quad 9, 190192, 217
six-noded triangular, 153155, 166, 192194
Central difference method, 280 stiffness matrix, 51, 119
Cholesky factorisation, 208, 230, 305 sub-parametric, 156
Conditionally stable method, 279 super-parametric, 156
Constant strain triangle (CST), 151, 173, 174, 178, 226 three-noded triangular, 149, 166, 171, 172
EulerLagrange equation, 71
Damping
Functional, 66, 68
factor, 235
minimization of, 73
modelling of, 272
stationarity of, 66, 81, 83
Rayleigh (proportional), 275
Degrees of freedom, nodal, 48, 149
Direct integration methods, 273, 279 Gauss elimination, 107
explicit methods, 282 GramSchmidt deflation, 267
implicit methods, 282
Discretisation 2, 3, 49, 306, 307
Displacement field, 75, 118 Interpolation function (see Shape function)
Duhamel integral, 275 Iteration
forward, 265
inverse, 265267
Eigenvalue, 235, 256, 259, 263265 simulataneous, 268271
shifting, 256 subspace, 271, 272
329
330 Index
Jacobi method, 259264 Rayleigh quotient, 266, 271

Jacobian, 186, 189, 191, 193, 212, 214 RayleighRitz method, 73, 75
Residual,
boundary, 17
Lagrange domain, 1, 17
elements, 158, 192 Refinement
equations, 240, 241, 244 h-type, 220
polynomial, 190 p-type, 220
Linear acceleration method, 284287, 287, 288
Load vector, 92
global, 92 Semi-analytical FEM, 148
Local coordinate frame, 152 Serendipity elements, 158, 192
Shape function, 2, 45, 75
Stationary total potential, principle of, 66, 73, 75
Magnification factor (MF), 233 Stiffness matrix,
Mass matrix element, 92, 102, 113, 129
consistent, 246, 250 direct method, 85
lumped, 246, 249 Strain
Mode superposition method, 273, 275279 initial, 169, 108, 111
plane, 146, 169, 172
Strain-displacement relation matrix, 102
Stress
Newmark methods, 287, 288
initial, 169
Nodal force vector, 120121 plane, 146, 169, 172, 173, 175
consistent, 121 thermal, 108
lumped, 121 Sturm sequence property, 256, 294
Nodal heat flux, 59 Substitution,
backward, 209
forward, 209
One-parameter solution, 18
Transformation matrix, 115, 129

Parasitic shear, 216 Trial function, 17
Point collocation technique, 21, 26 piece-wise continuous, 42, 81
collocation points, 26 Trial solution (see Trial function)
Quadrature,
Weighted residual,
NewtonCotes, 197, 198
Galerkin, 16
Gauss, 197204, 246
general weighted residual statement, 2833, 236
Simpson, 196 weak form of, 16, 33-36, 4148, 236
trapezoidal, 195 Weighting function, 24
Quasi-static, 233 Wilson-q method, 286
Textbook of
Finite Element Analysis
P. Seshu
Designed for a one-semester course in Finite Element Method, this compact
and well-organized text presents FEM as a tool to find approximate solutions
to differential equations. This provides the student a better perspective on the
technique and its wide range of applications. This approach reflects the current
trend as the present-day applications range from structures to biomechanics to
electromagnetics, unlike in conventional texts that view FEM primarily as an
extension of matrix methods of structural analysis.
After an introduction and a review of mathematical preliminaries, the book gives
a detailed discussion on FEM as a technique for solving differential equations and
variational formulation of FEM. This is followed by a lucid presentation of one-dimensional and two-dimensional finite
elements and finite element formulation for dynamics. The book concludes with some case studies that focus on
industrial problems and Appendices that include mini-project topics based on near-real-life problems.
Postgraduate/Senior undergraduate students of civil, mechanical and aeronautical engineering will find this text
extremely useful; it will also appeal to the practicing engineers and the teaching community.
KEY FEATURES ABOUT THE AUTHOR

u
Clear and concise presentation of concepts P. SESHU, Ph.D., is Professor, Department of
Mechanical Engineering, Indian Institute of Technology
u
Large number of worked-out examples Bombay. Earlier (198485), he worked as a Scientist at
u
Flow charts of key algorithms the Defence Research and Development Laboratory,
Hyderabad; he was also a visiting scientist to academic
u
Problems including Problems to Investigate at end institutions in USA, UK and Germany.
of chapters
An outstanding student throughout his studies and a
u
Large number of term-project topics distinguished teacher, Dr. Seshu has contributed
extensively to reputed journals in India and abroad,
including International Journal of Composite Structures,
Communications in Applied Numerical Methods, Journal
of Sound and Vibration, and Journal of Aeronautical
Society of India.
ISBN:978-81-203-2315-5
Rs. 250.00
9 788120 323155
www.phindia.com

1 2kiz6.textbook - Of.finite - Element.Analysis - by.P..Seshu PDF

Uploaded by

Copyright:

Available Formats

1 2kiz6.textbook - Of.finite - Element.Analysis - by.P..Seshu PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 2kiz6.textbook - Of.finite - Element.Analysis - by.P..Seshu PDF

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

Who is the author of the book?

Who is the author of the book?

Textbook of

TEXTBOOK OF FINITE ELEMENT ANALYSIS

Tenth Printing L L January, 2012

2. Finite Element Formulation Starting from Governing Differential

3. Finite Element Formulation Based on Stationarity of a Functional 66 88

3.4 Piece-wise Continuous Trial FunctionsFinite Element Method 81

4. One-dimensional Finite Element Analysis 89144

5. Two-dimensional Finite Element Analysis 145231

5.5 Numerical Integration 194

6. Dynamic Analysis Using Finite Elements 232294

7. Application Examples 295307

Appendix ASuggested Mini-Project Topics 309320

(b) Common 2-d elements

(d) Plate and shell elements

(e) Common 3-d elements

1.1 Typical Application Examples

1.1.1 Automotive Applications

Fig. 1.3 Finite element model of an automotive engine cylinder block.

Fig. 1.6 3-d Finite element model of a differential.

Fig. 1.7 3-d Finite element model of a rear axle.

Friction f1 Friction f2 Friction f3

1.1.2 Manufacturing Process Simulation

Fig. 1.9 Finite element model of fusion cast ceramic block.

1.1.3 Electrical and Electronics Engineering Applications

Fig. 1.10 High voltage insulation design of transformer using FEA.

B phase Y phase R phase

Fig. 1.11 Analysis of eddy currents using FEM.

Fig. 1.12 Solid model of a wire bonder machine.

Fig. 1.13 Finite element model of a wire bonder machine.

Fig. 1.15 Finite element model of a one-axis table.

Guide Column Bottom Mold

Fig. 1.17 Finite element model of a moulding machine.

1.1.4 Aerospace Applications

Fig. 1.18 Typical combat aircraft.

Fig. 1.19 Finite element simulation of deformation of a typical combat aircraft.

2.1 Weighted Residual MethodUse of a Single Continuous

Fig. 2.1 Rod subjected to axial load.

Fig. 2.2 Laminar flow on an inclined surface.

v(L) = 0 (no slip)

Fig. 2.3 Cantilever beam under load.

Fig. 2.4 Beam on simple supports.

and the resulting trial solution is

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

With this two-term trial function, we obtain the domain residual as

v(x) v (x) = c1 sin (p x/L) (2.36)

The domain residual

yielding c1 = 0.013071q0L4/EI. Therefore,

We can, in principle, choose any nonzero, integrable weighting functions to generate

and the boundary conditions are given by

2.2 The General Weighted Residual (WR) Statement

Fig. 2.11 A simply supported plate.

and the boundary conditions by

Example 2.8. Consider a 1 mm diameter, 50 mm long aluminium pin-fin (as shown in

Fig. 2.12 A pin-fin.

Solving for c2, we get c2 = 38,751.43. Hence,