CV7001_L1.

docx 25/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 1

Introduction to the course

Lectures
39 Hrs Lectures
3 Hrs Lectures/Hands-on sessions per week

Lecturers
Assoc. Prof. Lee Chi King
Subject coordinator
Lecturer for the theory part
Address: Block N1-1c-101
Telephone extension: 6790-5293
E-mail: [email protected]

Assoc. Prof. Zhao Zhiye

Lecturer for the application and non-linear part
Address: Block N1-1b-55
Telephone extension: 6790-5255
E-mail: [email protected]

Edventure site: Check the NTU Edventure site for all

course documents and announcements.

L1-1
 Course Outline for CV7001

Wk Lecture Contents Reference TS/AS

No.
1 L1 Introduction to the course and outline. Text and
reference. Assessment. Course website and other
information.
L2 Modeling of physical problems. Mathematical Ch. 1.1-1.2
models of the physical problem. Finite element
solution procedures.
L3 Review of basic concepts: Part I, Equilibrium Ch. 3
equations in 2D and 3D. Strain-displacement
relations. Plane strain and plane stress cases
2 L4 Review of basic concepts: Part II, Scalars, vectors, Appendices
and matrix. Linear equations system. Eigenvalues A-C
and eigenvectors
L5 Potential energy of an elastic body. Principle of Ch. 4.2-4.4 TS1
stationary potential energy.
L6 The Rayleigh-Ritz method. Ch. 4.5-4.6 TS1
3 L7 Formulation of truss problem. Ch. 2.4-2.7 TS1
L8 Coordinate transformation. Frame problem. Ch. 2.4-2.7 TS1
Ch. 8.1-8.4
L9 Assembly of element stiffness matrix and load Ch. 2.5
vectors. TS1

4 HOS 1 Finite element code introduction (VisualFEA) – VisualFEA

basic procedure and modelling features, truss manual
analysis.
5 L10 Inclined support. Elements with nonuniform cross Ch. 8.1-8.5 TS1
sections.
L11 Initial and thermal strain effects. Releases of end Ch. 2.10, TS1
moments. 2.13, 6.6, 6.7
L12 The model problem for linear solid mechanics. 2D Ch. 3, 4.8 TS2
and 3D elasticity problems. Solution of linear solid
mechanics problems by minimization of total
potential energy.
6 L13 Displacement finite element solution to linear solid Ch. 1, 3 TS2
mechanics problems. Finite element discretization.
Assembly of global equations and force vectors.
L14 Properties of interpolation functions. Ch. 3.2-3.4 TS2
L15 Interpolation of geometry and solutions. Coordinate Ch. 3.5- 3.9, TS2
transformation and mapping. Ch. 6.1-6.4
7 L16 Isoparametric finite elements for 1D and 2D Ch. 6.1-6.4 TS2
problems. Lagrangian and serendipity families.
Techniques of numerical integration. Appropriate Ch. 6.8, 6.14 TS3
L17 order of numerical integration. Ch. 7
L18 2D plane strain and plane stress formulations. Ch. 6.1-6.4 TS3
Axisymmetric formulation and applications. Ch. 7, Ch. 14

L1-2
Wk Lecture Contents Reference TS/AS
No.
Recess
8 - 10 HOS 2-4 Computer applications: 2D plane problems and VisualFEA AS1
project assignment. VisualFEA demo on dynamics manual
and seepage analyses. Introduction to nonlinear Lect notes
FEM and plasticity modelling in VisualFEA.
11 L19 3D continuum elements. Ch. 6.5, 6.15. TS3
L20-L21 Basic modeling techniques of finite element Ch. 6.11. TS3
methods. Selection of finite elements. Modelling of Ch. 6.12,
domain geometry and boundary conditions. Mesh Ch. 10
refinement. Interpretation of results.
12 L(NL)1-3 Introduction to geometrical nonlinearity and Ch. 17.1 Ch. AS1
material nonlinearity in finite element method. 17.9
Lect notes
13 L(NL)4-6 Basic nonlinear material models. Basic solution Ch. 17.2 – AS1
methods for nonlinear problems: pure incremental 17.3
method, direct iterative method, Newton-Raphson Lect notes
method.

Legends:
Lectures by Assoc. Prof. Lee Chi King
Lecture by Assoc. Prof. Zhao Zhiye

TS1-TS3: Tutorial sheets.

AS1: Computer application assignments.

Text and References:

Text

T1. Robert D. Cook, David S. Malkus, Michael E. Plesha

and Robert J. Witt. “Concepts and Applications of
Finite Element Analysis”. Fourth Edition, New York,
NY: Wiley, c2002, TA646.C771 2002
(Library 2, Level B2, Reserve Book Room)
ISBN0-471-35605-0

L1-3
References:

R1. Chandrupatla, Tirupathi R., “Introduction to finite

elements in engineering”, Prentice Hall, 2002. Third
Edition.
ISBN: 0-1306-1591-9

R2. O. C. Zienkiewicz, R. L. Taylor and J. Z. Zhu, “The

Finite Element Method: Its Basis and Fundamentals”,
Elsevier Butterworth Heinemann, 2005. Sixth Edition.
ISBN: 0-7506-6320-0

Website for the software VisualFEA/CBT

http://www.visualfea.com/

L1-4
Assessment
• Examination, application exercises and Quiz

Application exercises and Quiz

• Given during the hands-on sessions by A/P Zhao
• 30% of the total marks

Examination
• 70% of total marks.

Some remarks on the lecture note/tutorial sheets

and use of text/references

• Lecture notes are designed to bring out the essential

ideas and concepts

• Summary of the theories and do not contain all the

details of the course materials

• Only essential and important equations are normally

given in the notes (due to limited space)

• For complete details and calculations/formulae, one

must read (at least) the text and you are strongly advised
to read the references as well.

• Tutorial sheets contain selected (good) questions to

strengthen the concepts and calculation details: Always
try to attempt the questions yourself.

L1-5
 CV7001_L2.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 2

2.1 Modelling of Physical Problems

Change of
1. Physical Problem and criteria Physical model

2. Basic Mathematical Problem Refined

(Often too complicated) mathematical
model
3 Simplified Mathematical Problem
(Mathematical Model)

4 Solution of the mathematical

model by computational analysis
(e.g. Matrix/Stiffness Method, Finite
Element Methods)
Refinement
5. Interpretation of results and of model
physical conclusions. details
Improvement of
design/optimization

Fig. 2.1 General process for modelling physical problems

L2-1
• The area of numerical analysis is very board and only a
small part (numerical structural analysis) will be
considered in this course.

• In the area of numerical structural analysis, many

different methods exist. Only the most commonly used
(and important) method, namely the finite element
method will be descriped with some details. (Shaded
area in the figure above).

2.2 Mathematical Model of the Physical Problem

• Physical problems with engineering interests can come

with many forms and with very wide range of
complexity.
• Two main methods of solution of physical
problems: Experimental and mathematical solution
methods.

Physical Problems

Experimental Model: Mathematical Model:

Full scale analysis Analytical solutions
Field analysis measurements Numerical solutions/Analysis
Model analysis Semi-analytical/numerical methods

Combination of Experimental model

and Mathematical/numerical model

L2-2
 Fig. 2.2. Modelling methods for solving physical
problems.

• Very often, both modelling techniques will be used

together for better understanding/solution of the
physical problem. For example, numerical analysis
supplemented by (more expensive) scaled down model
or even full scale analysis.

• Some advantages of mathematical/numerical modelling:

(i) faster modelling (what-if analysis) and ease for
modifications
(ii) less expensive compare with experimental
modelling
(iii) allow engineers to have good insight of the
problems
(iv) can be effectively integrated with design and
optimalization process

• Some disadvantages of mathematical/numerical

modelling:
(i) very often can only give approximated/incomplete
solutions of the problems.
(ii) engineers must understand the assumptions made in
the underlying mathematical model, the nature of
the problem and significance of the solutions.
Hence, a correct interpretation of the
mathematical/numerical solution is very important.
(iii) use of wrong models (either mathematical or
numerical model) may lead to incorrect solutions
and interpretations.

L2-3
• Only the numerical model of the mathematical model
and its used in structural analysis will be discussed is
this course.

• The mathematical model transforms the physical model

(reality) into information which is of direct interest and
does not add anything new (In fact, it nearly always
leads to some loss in details).

• In structural analysis, the mathematical model can

always be described by partial differential equations
(PDEs) of different
complexities: Linear/Nonlinear, Time
independent/dependent.

• Normally the mathematical model must be simplified

on details of geometry, kinematics, material laws and
boundary conditions etc.

• The mathematical model must has reasonable

mathematical properties, e.g. the existence of solution.
Note that the ‘obvious’ existence of solutions of the
physical problem does not necessarily mean that the
solution of the mathematical model exists. Furthermore,
solution of the mathematical model should give
importance insight to the reality.

• For practical engineering problems, it is very often

(almost always) that even the simplified mathematical
model is still too complicated for exact analytical
solution (i.e. the governing PDE cannot be solved

L2-4
 exactly) and therefore numerical methods are used to
obtain approximated solutions in most situations.
2.3 Solution of mathematical model by numerical
analysis

• In numerical analysis, only the mathematical model, and

not the reality is (and can be) analyzed. Hence, one can
never obtain more information from the computational
analysis in the prediction of reality than the information
contained in the mathematical model.

• Many different numerical methods exist (e.g. finite

element, boundary element, finite difference methods).

• In almost all practical applications, the numerical

method can only provide approximated solutions to the
mathematical problem.

• In general, more computational effort is needed if one

would like to increase the accuracy of the solutions of
the computational analysis.

• With the rapid advancement in the speed and capacity

of digital computers (and reduction in cost at the same
time), numerical analysis becomes more and more
popular and has been extended to many different areas
of engineering problems.

• Many people think that numerical analysis≡computer

analysis. However, as all numerical analysis procedures
must based on the mathematical model, engineers must

L2-5
 have essential knowledge on the mathematics theories
of the numerical methods rather than just the
information/skills of using commercial analysis
packages.

2.4 Example

Consider the structural analysis of common multi-storey

RC building (physical problems)

Fig. 2.3. A typical concrete building

Factors that may affect the structural behaviours of the

buildings:

L2-6
(1) Material: concrete, steel reinforcements
(2) Structural design and fabrication, workmanship,
quality control etc.
(3) Loading and other "boundary conditions": Static
loading (live and dead loads), dynamic (e.g. wind,
seismic, temperature effect.....), foundation support
conditions (i.e. settlement, uplifting ....) and other
service conditions.

Not possible to consider all factors in one time and usually

to build an experimental model (either full scale or scaled
down model) will be too expensive and time consuming
except for some very special and important structures.

Hence, we must use mathematical model to study the

problem. For most structural engineers, they only
interested in the structural responses (displacement,
strains, stresses, force, moments) of the structure under
different combination of work (dead + live + wind)
loadings (and in some areas seismic loading). (simplified
mathematical model)

Most often, static linear elastic analysis (further

simpled model) will be carried out in design stages as
other more completed models (e.g. nonlinear response,
collapse/shake-down analysis) may be either too
complicated or not necessary.

However, even for static linear elastic analysis, we still

have many options with different complexities

L2-7
(i) consider the whole building as a 3D continuum
(almost always too complicated)
(ii) consider the building to be made up of different
"structural elements" of walls, slabs, beam and
columns. Identify the key elements which will resist
the loading under different loadings. For examples,
when under wind loading, the model will consist of
shear wall + beam + columns elements. (Still too
complicated to many engineers) However, this model
will often lead to more effective design in many
situations (e.g. tall buildings).

Beam
elements

column
element
"wall"
element

Fig. 2.4 A model with beam, column and wall elements

L2-8
(iii) Only consider the building as a frame (beams +
columns) for structural analysis (most commonly
used model. Only acceptable for building of
moderately height).

Hence, we need more than just beam and column elements

for many structural analysis problems. The most
commonly used elements are walls (2D elements), slabs
(plates), shell (curved plate) and even 3D elements (3D
continuum solid).

To carry out analysis including 2D elements, plates, shell

and 3D elements, we need to used the Finite Element
Method in which the problem domain is discretized the
into a number of "finite elements" for analysis. In fact, the
commonly used truss, beam and column elements are only
three particular types of "finite elements".

L2-9
 2.5 Finite Element Solution Procedure
Mesh
Finite Element solution
Selection of the following ingredients:
1. Finite Elements (order/ method of
approximation)
2. Mesh density and pattern
3. Solution parameters
4. Representation of loading
5. Representation of boundary conditions

Finite Element Error estimation

Assessment of the discretization error of
the finite element solutions

Finite element model refinement:

1. Mesh refinement
2. Use of higher order elements
3. Use both mesh refinement and higher
order elements

Fig. 2.5

Note: Only the basic formulation of finite element analysis

for structural problem (shaded) will be considered in this
course.

L2-10
 CV7001_L3.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001

FINITE ELEMENT METHODS

Lecture 3

Review of basic concepts: Part I

3.1 Introduction

Structural analysis => analysis of displacements, strains,

stresses and hence, forces and moments

Hence, essential ingredients of the course

(1) Linear elasticity solid mechanics

(2) Vectors and matrix algebra
(3) Elementary differential calculus (coordinate
transformation and mapping)

3.2 Displacements, strains, stresses and

equilibrium equations

Displacements in 2D and 3D
Column vectors with 2 and 3 rows

2D case u(x,y)=[u(x,y),v(x,y)]T (3.1a)

L3-1
 CV7001_L3.docx 24/06/2009

3D case u(x,y,z)=[u(x,y,z),v(x,y,z),w(x,y,z)]T (3.1b)

Sometime (e.g. in a beam or a plate ) rotation parameters

are also involved.

u(x,y)=[w, θx, θy] (3.1c)

Strain components in 2D and 3D

Fig. 3.1 Definition of strain

2D strain components

Normal strain (ratio of change in length to original length)

L 0'2' − L 02 [dx + (u + u , x dx) - u] - dx ∂u

εx = = = u,x = (3.2a)
L 02 dx ∂x

L3-2
 CV7001_L3.docx 24/06/2009

Similarly, y-direction norm strain, εy

∂v
ε y = v, y = (3.2b)
∂y

Shear strain, γxy is the amount of change in right angle (0-

1-2 => 0'-1'-2') Assumed that with small values of β1 and
β2, β1≈tanβ1 and β2≈tanβ2 and

(u + u , y dy) - u (v + v, x dx) - v
γ xy = β1 + β 2 = + = u , y + v , x (3.2c)
dy dx

or ε x = u , x , ε y = v, y , γ xy = u , y + v, x (3.3)

3D strain components

With the presence of displacement component in the z-

direction, w, the corresponding strains are

ε z = w , z , γ yz = v, z + w , y , γ xz = u , z + w , x (3.4)

Eqns. (3.3) and (3.4) can be expressed in matrix form as

⎡∂ ⎤
⎢ 0⎥
⎧ ε x ⎫ ⎢ ∂x ⎥
⎪ ⎪ ⎢ ∂ ⎥ ⎧u ⎫
⎨ εy ⎬ = 0 ⎨ ⎬ (3.5a)
⎪γ ⎪ ⎢ ∂y ⎥ ⎩ v ⎭
⎩ xy ⎭ ⎢ ∂ ∂⎥
⎢ ⎥
⎣ ∂y ∂x ⎦

L3-3
 CV7001_L3.docx 24/06/2009

⎧ ε x ⎫ ⎡∂ / ∂x 0 0 ⎤
⎪ε ⎪ ⎢ 0 ∂ / ∂y 0 ⎥
⎪ ⎪ ⎢
y ⎥⎧ u ⎫
⎪ εz ⎪ ⎢ 0 0 ∂ / ∂z ⎥ ⎪ ⎪
⎨ ⎬ ⎢ = ⎥⎨ v ⎬ (3.5b)
⎪γ xy ⎪ ⎢∂ / ∂y ∂ / ∂x 0 ⎥⎪ ⎪
⎪ γ yz ⎪ ⎢ 0 ⎩w ⎭
∂ / ∂z ∂ / ∂y ⎥
⎪ ⎪ ⎢ ⎥
γ
⎩ zx ⎭ ⎣ ∂ / ∂z 0 ∂ / ∂x ⎦

or in both cases ε = Lu (3.5c)

Note the L is sometime called the strain operator.

Compatibility

When a body is deformed without breaking (i.e. no crack

appears in stretching or no kink in bending) or no part
overlaps other. The compatibility conditions have to be
satisfied and the displacement function will be continuous
and single valued.

Compatibility conditions in 2D

ε x, yy + ε y,xx = γ xy,xy (3.6)

For 3D cases, totally six compatibility equations can be

obtained.

L3-4
 CV7001_L3.docx 24/06/2009

Stress and deformation

Fig. 3.2 An elementary area in 2D

2D and 3D equilibrium equations

Consider the elementary area shown in Fig. 3.2, for

balance of forces in x direction gives

- σ x tdy - τ xy tdx + (σ x + σ x, x dx)tdy +

(3.7a)
(τ xy + τ xy, y dy)tdx + Fx tdxdy = 0

or σ x, x + τ xy, y + Fx = 0 (3.7b)

In y-direction τ xy,x + σ y, y + Fy = 0 (3.7c)

⎡∂ ∂ ⎤⎧ σ ⎫
⎢ ∂x 0
∂y ⎥ ⎪ ⎪ ⎧Fx ⎫ ⎧0⎫
x
or ⎢ ⎥ σ + = (3.8a)
⎢0 ∂ ∂ ⎥ ⎨ y ⎬ ⎨⎩Fy ⎬⎭ ⎨⎩0⎬⎭
⎪τ ⎪
⎣⎢ ∂y ∂x ⎦⎥ ⎩ xy ⎭

L3-5
 CV7001_L3.docx 24/06/2009

i.e. LT σ + Fb = 0 (3.8b)

Note: Fx and Fy are known as the body forces in x and y

directions respectively. (e. g. weight of the material).

3D equilibrium equations

By considering the equilibrium conditions of an

elementary volume in 3D, we can again obtain the
following equilibrium equations for 3D problems.

σ x, x + τ xy, y + τ zx, z + Fx = 0
τ xy, x + σ y, y + τ yz,z + Fy = 0 (3.9a)
τ zx, x + τ yz, y + σ z, z + Fz = 0

or, again, LT σ + Fb = 0 (3.9b)

Boundary tractions

Fig. 3.3 Boundary tractions

L3-6
 CV7001_L3.docx 24/06/2009

Boundary tractions are surface pressure usually applied

along the surface of the structures (in contrast to body
forces, which act inside and throughout a volume). The
surface tractions Φx and Φy along the x and y directions
respectively (Fig. 3.3) are related to the stress inside the
structures as

x direction: Φx=lσx+mτxy (3.10a)

y direction: Φy=lτxy+mσy (3.10b)

Note: l and m are the direction cosines along the boundary

such that

l=dy/ds, m=dx/ds (3.11)

In 3D case:

x direction: Φx=lσx+mτxy+nτzx (3.12a)

y direction: Φy=lτxy+mσy+nτyz (3.12b)

z direction: Φz=lτzx+mτyz+nσz (3.12c)

with, l, m and n are the direction cosines in 3D.

L3-7
 CV7001_L3.docx 24/06/2009

Note that Eqn. 3.12 can also be written as

⎡σ xx ⎤
⎢σ ⎥
⎡ l 0 0 m 0 n ⎤ ⎢ ⎥ ⎡Φ x ⎤
yy

⎢0 m 0 l n 0 ⎥ ⎢ σ zz ⎥ = ⎢Φ ⎥ (3.13a)
⎢ ⎥ ⎢ τ xy ⎥ ⎢ y ⎥
⎢⎣0 0 n 0 m l ⎥⎦ ⎢ ⎥ ⎢⎣ Φ z ⎥⎦
⎢ τ zx ⎥
⎢ ⎥
⎢⎣ τ yz ⎥⎦

or Gσ=Φ (3.13b)

Note: Surface tractions Φx, Φy and Φz have units of stress

(pressure).

3.3 Stress-Strain relations

Since we will eventually (after the discussion of trusses

and frames) deals with problems of general linear
elasticity with general material properties, a section will
be devoted for revision of stress-strain relations.

3D Problems

From Eqn. (3.5b), there are six components in the strain

vector

ε=[εx, εy, εz, γxy, γyz, γzx]T (3.14a)

L3-8
 CV7001_L3.docx 24/06/2009

In general, the strain vector can be related to the stress

vector, σ, by a symmetric matrix C

σ=[σx, σy, σz, τxy, τyz, τzx]T (3.14b)

and ε=Cσ or σ=Eε and C=E-1 (3.14c)

The matrix C is called the material compliance matrix

while its inverse E is known as material stiffness matrix or
more commonly the constitutive matrix. Eqn. 3.14c is
known as the generalized Hooke's Law.

Note that:

C and E are symmetric and positive definite and are 6 by 6

matrices.

In the most general case (3D anisotropic case), C and E

will have 21 independent coefficients.

For orthotropic materials (materials display different

stiffness in mutually perpendicular directions e.g. wood
cut from a log), there are only 9 independent coefficients.

For isotropic materials (i.e. material with the same

properties in any direction e.g. steel and most metals),
only two independent coefficients exist and they are,
namely, the Young's (Elastic) modulus, E, and the Poisson
ratio, ν, such that

L3-9
 CV7001_L3.docx 24/06/2009

⎡1 - ν ν ν 0 0 0 ⎤
⎢ ν 1- ν ν 0 0 0 ⎥
⎢ ⎥
E ⎢ ν ν 1- ν 0 0 0 ⎥
E= ⎢ ⎥
(1 + ν)(1 - 2ν) ⎢ 0 0 0 1/2 - ν 0 0 ⎥
⎢ 0 0 0 0 1/2 - ν 0 ⎥
⎢ ⎥
⎣ 0 0 0 0 0 1/2 - ν ⎦
(3.15)

Plane strain and Plane Stress cases

Plane stress and plane strain problems are particular cases

of 3D elasticity problems in which some stress and strain
components are assumed to be identically equal to zero in
the whole problem domain.

Plane stress problems

The stress components σz, τxz and τyz are all equal to (or
assumed to be) zero due to the reason that the thickness of
the problem domain in the z direction is very thin. An
example of plane stress analysis is a shear wall under
lateral loading.

Eliminate σz, τxz and τyz from Eqn. (3.15) gives

⎡1 ν 0 ⎤
E ⎢ ⎥
E= ν 1 0 (3.16)
1- ν2 ⎢ ⎥
⎢⎣ 0 0 (1 - ν)/2⎥⎦

L3-10
 CV7001_L3.docx 24/06/2009

such that

[
σ = σ x , σ y , τ xy ]T and ε = [ε x , ε y , γ xy ]T and σ=Eε (3.17)

60m

Lateral loading
(e.g. wind loading

Thickness of wall=0.15m<< height=60m

2m
3m

6m
2m 6m

Fig. 3.4 Shear wall under lateral loading

Plane Strain Problems

In plane strain problems, the simplification of the stress

components comes from another extreme when the
dimension of the body in the z direction is very large and
displacement in the z direction (w) are equal to zero. For
example, a retaining wall under lateral pressure (Fig. 3.5a)
and a tunnel or an underground structure (Fig. 3.5b). It is
assumed that (i) all cross sections are in the same
condition and (ii) the end sections are confined between
fixed smooth rigid planes so that the displacement in the z
direction is prevented.

L3-11
 CV7001_L3.docx 24/06/2009

(a) (b)

Fig. 3.5 Plane strain problems

Normally, only a slice with unit thickness is considered (as

the loading does not vary along the length and conditions
are same at all cross sections), Sine w=0 and u,v are
independent of z,

∂v ∂w ∂u ∂w ∂u
γ yz = + = 0, γ xz = + = 0, ε z = =0 (3.18)
∂z ∂y ∂z ∂x ∂z

Thus,

σz=ν(σx+σy) and τxz=τyz=0 (3.19)

Eliminate σz, τxz and τyz from Eqn. (3.15) again will give

⎡ ⎤
⎢ 1 ν (1 - ν ) 0 ⎥
⎢ ⎥
E(1 - ν) ⎢
E= ν (1 - ν) 1 0 ⎥ (3.20)
(1 + ν)(1 - 2ν) ⎢ ⎥
⎢ (1 - 2ν) ⎥
⎢ 0 0 ⎥
⎣ 2(1 - ν ) ⎦

L3-12
 CV7001_L3.docx 24/06/2009

and again

[
σ = σ x , σ y , τ xy ]T and ε = [ε x , ε y , γ xy ]T and σ=Eε (3.21)

(Note that the E matrix of Eqn. 3.17 is not the same as in

Eqn. 3.21.)

More details (e.g. initial stresses and strain, thermal

effects) related to stress-strain relations will be given in
the section related to finite element formulations of 3D,
plane stresses and strains problems.

L3-13
 CV7001_L4.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001

FINITE ELEMENT METHODS

Lecture 4

Review of basic concepts: Part II

4.1 Introduction

A brief introduction/revision of vector and matrix algebra

is given here before the discussion finite element analysis
as knowledge of them is essential for this subject.

4.2 Scalars and Vectors

Scalars

A scalar is a quantity having magnitude but no direction.

E.g. mass, time, temperature
Often denoted as lower case characters: α, a, b, c

Vectors

A vector is a quantity having both magnitude and

direction. E.g. force, displacement, acceleration.
Often denoted as bolded lower case characters a, v or
enclosed with braces {d}, {D}.

L4-1
 CV7001_L4.docx 24/06/2009

Vector Operations

Dot product: a⋅b = a b cosθ, 0≤θ≤π

a⋅b = b⋅a, a⋅(b+c)=a⋅b+a⋅c etc.
Cross product: a×b
a×b=- b×a a×(b+c)= a×b+ a×c etc.

Linear Dependence and Independence vectors

A set of vectors {u1,u2,....,us} is said to be linear

dependent if there exist numbers α1, α2, .... αs which are
not all zero such that

α1u1 + α2u2 + ..... + αsus=0 (4.1)

If the vectors are not linearly dependent, they are called

linearly independent vectors.

Two vectors a and b are orthogonal if a⋅b=0. Furthermore,

if a = b = 1, they are orthonormal.

4.3 Matrix

A matrix is usually denoted as upper case bold characters

e. g. A matrix A is enclosed in square bracket [A]

[ ]m×n a ij ∈ ℜ
A m×n = a ij (4.2)

(Note: ℜ denote the space of real number.)

L4-2
 CV7001_L4.docx 24/06/2009

A matrix with one element m=n=1 is a scalar.

A m×1 matrix is a column vector and a 1×n matrix is a

row vector

⎡ a1 ⎤
⎢ ⋅ ⎥
⎢ ⎥
column vector : ⎢ ⋅ ⎥ row vector : [a1 ⋅ ⋅ ⋅ a n ] (4.3)
⎢ ⎥
⎢ ⋅ ⎥
⎢⎣a m ⎥⎦

Two matrices A and B are equal iff (if and only if)

(1) A and B have the same number of rows and columns.

(2) All corresponding elements are equal. i.e. aij=bij ∀i,j

(∀ means “for all”)

Matrix Operations

Transpose

C=AT, Cij=Aji (4.4)

The matrix A is symmetric if A=AT.

L4-3
 CV7001_L4.docx 24/06/2009

Addition

C=A+B cij=aij+bij (4.5)

Multiplication

n
Cm×p=Am×n × Bn×p cij = ∑ a ir b rj (4.6a)
r =1

C(A+B)=CA+CB, AB≠BA (4.6b)

Note: The dot product of two column vectors u and v can

be represented as uTv .

(What will happen if we write uvT if both u and v are

column vectors?)

Inverse of matrix

The inverse of A is denoted as A-1 such that

A-1A=A A-1=I (4.7)

A matrix that possesses an inverse is said to be non-

singular, otherwise it is a singular matrix.

Matrix partitioning

A matrix can always be partitioned into submatrices. A

submatrix of a matrix is itself a matrix that contains
elements from certain rows and columns of the original
matrix. For example

L4-4
 CV7001_L4.docx 24/06/2009

⎡ a11 a12 a13 a14 ⎤

⎡A A12 A13 ⎤
A = ⎢a 21 a 22 a 23 a 24 ⎥ = ⎢ 11 (4.8a)
⎢ ⎥ ⎣ A 21 A 22 A 23 ⎥⎦
⎢⎣a 31 a 32 a 33 a 34 ⎥⎦

⎡a ⎤ ⎡a a13 ⎤ ⎡ a14 ⎤
where A11 = ⎢ 11 ⎥, A12 = ⎢ 12 ⎥ , A13 = ⎢ ⎥ etc. (4.8b)
⎣a 21 ⎦ ⎣a 22 a 23 ⎦ ⎣a 24 ⎦

⎡ A11 A12 ⎤ ⎡ B11 ⎤

and if A = ⎢ and B = ⎢ ⎥ then
⎣ A 21 A 22 ⎥⎦ ⎣B 21 ⎦

⎡ A B + A12B 21 ⎤
AB = ⎢ 11 11 ⎥ (4.8c)
⎣ A 21B11 + A 22B 21 ⎦

Determinant of a square matrix

The determinant of a square matrix A, Det(A) can be

defined as the determinants of submatrices of A and the
determinant of a 1×1 matrix is simply the element of the
matrix. I.e. if A=[a11], Det(A)=a11 and

n
Det(A) = ∑ (-1)1+ j a ijDet(A1j ) (4.9)
j=1

where A1j is the (n-1)×(n-1) matrix obtained by

eliminating the 1st row and jth column from A.

L4-5
 CV7001_L4.docx 24/06/2009

Rank of matrix

The rank, r, of a m×n matrix A is defined as the order of

the largest nonzero determinant in A. An equivalent
definition of rank is the maximum number of linearly
independent rows (or columns) in A. Furthermore, n-r is
the dimension of the solution space of the linear equations
Ax=b where b is any column vector.

Examples
⎡2 1 1 ⎤ ⎡1⎤
(1) Let A = ⎢ ⎥ and b = ⎢ ⎥ . The rank of A is
⎣0 1 − 1⎦ ⎣0 ⎦
equal to 2. Hence, the dimension of the solution space
of the equations system Ax=b is equal 1.
(What is the solution set for this equations set?)
⎡7 4 1 ⎤
(2) Let A = ⎢4 2 0⎥ and rank of A=3. Thus, the
⎢ ⎥
⎢⎣1 0 3⎥⎦
dimension of the solution set for any equation of the
form Ax=b is a uniquely defined point is given by
x=A-1b.

Positive-definite

A n×n matrix A is positive-definite if and only if

x T Ax ≥ 0 and x T Ax = 0 iff x = 0, ∀x ∈ ℜ n (4.10a)

A is semi-positive-definite if
x T Ax = 0 (4.10b)

L4-6
 CV7001_L4.docx 24/06/2009

for a given vector x≠0. (The form xT Ax is commonly

called the quadratic form of A.)

Warning: Symmetry and positive-definite are two different

properties.

Examples
⎡1 2⎤
Consider the matrix A = ⎢ ⎥ . It is positive-definite
⎣0 2 ⎦
(but not symmetric) since ∀x=[x1,x2]T∈ℜ2.

x T Ax = x12 + 2x1x 2 + 2x 22 = (x1 + x 2 ) 2 + x 22 ≥ 0 (4.11a)

⎡ - 1 - 2⎤
The matrix B = ⎢ ⎥ is symmetric but not positive-
⎣ - 2 0 ⎦
definite

⎡ − 1 − 2⎤ ⎡1⎤
[1 0]⎢ ⎥ ⎢0⎥ = -1 < 0 (4.11b)
⎣ − 2 0 ⎦⎣ ⎦

System of linear equations

Consider a linear equations system of the form

Ax=b A=[aij]n×n and x,b∈ℜn (4.12)

(Note: ℜn denote the n dimensional real vector space, its

elements can always be expressed as a n×1 column
vector.)
Then

L4-7
 CV7001_L4.docx 24/06/2009

(1) The system is called consistent if there exists at least

one solution.

(2) Any vector xc∈Ax=0 is called the complement

solution of the system (Note: ∈ means such that).

(3) Any vector xp∈Axp=b is called a particular solution of

the system.

(4) The vector x=xp+xc is again a solution of the system.

(5) If Det(A)≠0, then the system will have a unique

solution equal to A-1b.

Let A be a n×n square matrix, then if A is non-singular.

Then

(1) Det(A)≠0

(2) Column (row) rank of A=n and all the columns (rows)
are independent.

(3) The only solution for Ax=0 is the trivial solution x=0.

(4) A-1 exists and Ax=b has a unique solution x= A-1b.

(5) In practice, to solve a system of linear equation, there

is no need to find the inverse explicitly. The technique
of Gaussian Elimination is often used instead as
demonstrated below.

L4-8
 CV7001_L4.docx 24/06/2009

Example: Solve the following system

⎡2 4 7 3 ⎤ ⎡ α1 ⎤ ⎡1 ⎤
⎢ 6 9 2 1 ⎥ ⎢α ⎥ ⎢ 2 ⎥
⎢ ⎥⎢ 2 ⎥ = ⎢ ⎥ (4.13a)
⎢3 8 20 2 ⎥ ⎢α 3 ⎥ ⎢3⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣ 5 11 2 10 ⎦ ⎣α 4 ⎦ ⎣ 4 ⎦

Eliminate the first column/row:

⎡2 4 7 3 ⎤ ⎡ α1 ⎤ ⎡ 1 ⎤
⎢0 − 3 − 19 − 8 ⎥ ⎢α ⎥ ⎢ - 1 ⎥
⎢ ⎥⎢ 2 ⎥ = ⎢ ⎥ (4.13b)
⎢0 2 9.5 − 2.5⎥ ⎢α 3 ⎥ ⎢1.5⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣0 1 15.5 2.5 ⎦ ⎣α 4 ⎦ ⎣1.5⎦

Second column/row:

⎡2 4 7 3 ⎤ ⎡ α1 ⎤ ⎡ 1 ⎤
⎢0 − 3 − 19 − 8 ⎥ ⎢α 2 ⎥ ⎢ - 1 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥ (4.13c)
⎢0 0 − 3.1667 − 7.83333⎥ ⎢α 3 ⎥ ⎢0.8333⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣ 0 0 − 21.8333 − 0.16667 α
⎦⎣ 4 ⎦ ⎣ 1.1667 ⎦

Third column/row

⎡2 4 7 3 ⎤ ⎡ α1 ⎤ ⎡ 1 ⎤
⎢0 − 3 − 19 − 8 ⎥ ⎢α 2 ⎥ ⎢ - 1 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥ (4.13d)
⎢0 0 − 3.1667 − 7.83333⎥ ⎢α 3 ⎥ ⎢ 0.8333 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣0 0 0 53.8421 ⎦ ⎣α 4 ⎦ ⎣− 4.579⎦

L4-9
 CV7001_L4.docx 24/06/2009

After back substitution

⎡ α1 ⎤ ⎡ - 0.9765 ⎤
⎢α ⎥ ⎢ 0.8944 ⎥
⎢ 2 ⎥=⎢ ⎥ (4.13e)
⎢ α 3 ⎥ ⎢ - 0.052786 ⎥
⎢ ⎥ ⎢ ⎥
⎣α 4 ⎦ ⎣− 0.085044⎦

Eigenvalues and eigenvectors

Consider a linear equations system of the form

(A-λI)x=0 A=[aij]n×n and x, b∈ℜn (4.14)

where I is a n×n identity matrix. The scalar λ which

satisfies Eqn. 4.14 is known as an eigenvalue of the matrix
A. Furthermore, Eqn. 4.14 is known as the characteristic
equation of A.

It can be shown that if λ is an eignvalue of A then

det(A-λI)=0 (4.15)

Furthermore, if λi is an eignvalue of A, the non-trivial

solution, xi to the equation

(A-λiI)xi=0 (4.16)

is the eignvector corresponding to λi. In general, for the

n×n matrix A, there exist at most n distinct eignvectors.

L4-10
 CV7001_L4.docx 24/06/2009

Let λi, i=1,….,n be the eignvalues of A and xi, i=1,…, n

their corresponding eignvecotors, then

(1) The sum of the n eigenvalues equals the sum of the n

diagonal entries of A. The sum is called the trace of A.

(2) The product of the n eignvalues equal to det(A).

(3) If the eignvectors xi, i=1,…,n corresponding to

different (distinct) eignvalues λi, i=1,…., n, then xi,
i=1,…,n are linearly independents.

(5) If A has n linearly independents eigenvectors. Then if

the eignvectors are chosen to be the columns of a
matrix S, then A can be written as

A=SΛS-1 (4.17a)

⎡M M M M ⎤
⎢M M M M ⎥
⎢ ⎥
where S = ⎢x1 x 2 L L x n −1 xn ⎥ (4.17b)
⎢ ⎥
⎢M M M M ⎥
⎢⎣ M M M M ⎥⎦

⎡λ1 ⎤
⎢ λ2 ⎥
⎢ ⎥
and Λ=⎢ O ⎥ (4.17c)
⎢ ⎥
⎢ O ⎥
⎢⎣ λ n ⎥⎦

L4-11
 CV7001_L4.docx 24/06/2009

(6) If A is real and symmetric, all it eignvalues are real.

(7) If A is real and symmetric and all its eignvalues are

distinct. All its eignvectors can be chosen
orthonormal. That is

⎧1 if i = j
xi ⋅ x j = ⎨ (4.18)
⎩0 if i ≠ j

(8) If A is positive definite, then all its eignvalues λi > 0.

L4-12
 CV7001_L5.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering
CV7001
FINITE ELEMENT METHODS
Lecture 5

Principle of stationary potential energy

5.1 Potential energy of an elastic body

Consider a general 3D elastic body (Fig. 5.1) under the

action of external loading (surface traction force). In
general, the body deforms under the loading and we
denote the deformation as u. Note that u=[u,v,w]T is a
vector function of position (i.e. the value of u depends on
the position) and in general not equal to zero. However,
we note that u≡0 along the support surface where the
displacement is fixed to zero.

New position of surface traction Surface traction force acting on

the body, Φ
deformed shape of the body

domain of the body=Ω

Body force=Fb

Γu, Support surface where u=0

Fig. 5.1 A 3D solid under the action of external load.

L5-1
 CV7001_L5.docx 24/06/2009

Elastic strain energy

In general, the displacement field (or deformation of the

body) u will induce strain inside the body (via Eqn. 3.5)
and hence creates stress (Eqns. 3.14 and 3.15) and store up
energy. This energy contained in elastic distortion (or
deformation) is known as the elastic strain energy and is
often denoted as U and is given by

1 T 1 T
U= ∫2 ( ε σ )d Ω = ∫ 2 (ε Εε )dΩ (5.1a)
Ω Ω

Use Eqn. 3.5c

1 T 1 T
U= ∫2 ( ε Εε )d Ω = ∫2 ( Lu ΕLu )dΩ (5.1b)
Ω Ω

Hence, if one can find out the exact expression of u (not at

one or at a few points, but everywhere inside the problem
domain Ω), one can compute the strain energy term U.

Note:
(1) In this moment, the effect of initial stress and strain is
ignored.

(2) Suppose that if

⎡b ⎤
u = ⎢0 ⎥ (5.2)
⎢ ⎥
⎢⎣0 ⎥⎦

L5-2
 CV7001_L5.docx 24/06/2009

where b is a constant. Then obviously, U=0. Since all

strain component become identically zero everywhere
in the problem domain Ω. The displacement field in
Eqn. 5.2 is called rigid body motion (or mode). There
are total six independent rigid body modes for 3D
objects (3 translations and 3 rotations) and Eqn. 5.2
is, in fact, translation in x-direction.

(3) In many cases, the magnitudes of some strain and

stress components are much greater than other
components. We can then simplify Eqn. (5.1a) by
ignoring components with small magnitude without
introducing much error in the energy expression.

For example, consider the bar shown in Fig. 5.2.

The strain component εxx (and σxx) will dominate the strain
energy expression when comparing with all other
components.

y Area=A,
Young's Modulus=E
L

x
P
Fig. 5.2 Bar under tension force P.

L5-3
 CV7001_L5.docx 24/06/2009

Hence,

1 T 1 L1
U= ∫2 ( ε σ )dΩ = ∫ 2 xx xx
ε σ dΩ = ∫0 2 ε xx σ xx Adx (5.3a)
Ω Ω

Since, σ=P/A and ε=P/EA, then

L1 L1 P P P2L
U = ∫0 ε xx σ xx dx = ∫0 ( )( )Adx = (5.3b)
2 2 EA A 2EA

So what about the cases for a (i) beam (ii) column?

Work done by body force and surface tractions

As the body deforms, the body force Fb and the surface

tractions Φ will move as well => Work done (dot product
between forces and displacements) by the body force and
external traction, W, is equal to

− ∫ u T FbdΩ − ∫ u T ΦdΓ (5.4)

Ω ΓΦ
Note:
ΓΦ is the surface on which the traction Φ is applied and
hence the second term in Eqn. 5.4 is a surface (or
boundary) integral.

A negative sign is used to denote that the work is done by

external sources (compare with the internal strain energy).

L5-4
 CV7001_L5.docx 24/06/2009

Total potential energy

The total potential energy of the whole system shown in

Fig. 5.1 is give by

1
Πp=U+W= ∫ (ε T σ )dΩ − ∫ u T FbdΩ − ∫ u T ΦdΓ (5.5)
Ω2 Ω ΓΦ

Note:
In Eqn. 5.5, we assume that
(i) The problem (system) is essentially static and no
kinetic energy of any form (e.g. vibration, acceleration
etc) is considered.

(ii) No other form of energy (e.g. thermal, sound) is

considered.

(iii) The system is conservative in the sense that there is

no change in total potential energy before and after
the loading is applied and the structure is deformed. It
is also path independent (i.e. the work done, W, is
independent of the path the body force and boundary
tractions taken from the initial position to the final
deformed positions).

(iv) From (iii), if the system is initial stress free, before

deformation, Πp=0 and since the potential energy will
not be changed. We have Πp=0 after the deformation
or

L5-5
 CV7001_L5.docx 24/06/2009

1 T T T
∫ 2 (ε σ )dΩ = ∫ u FbdΩ + ∫ u ΦdΓ (5.6a)
Ω Ω ΓΦ
or

U=W (5.6b)

i.e. Total strain energy = work done by body force

and surface tractions.

(v) The concept and expressions of Πp, U and W are very

important in structural mechanics/finite element
methods as they are the starting point of the solution
methods and many energy theorems (including Betti's
and Maxwell, Castigliano theorems).

5.2 Principle of stationary potential energy

The principle of stationary potential energy can be applied

to solve many structural/continuum mechanics problems
in which the solutions are (i) time and path independent
and (ii) conservative.

Some definitions

System: The physical structures (e.g. a building or a dam)

and the loads applied to it.

Configuration: The configuration of a system is the set of

positions of all particles of the structure. The original
configuration normally refers as the positions of all
particles before the action of the applied loading while the

L5-6
 CV7001_L5.docx 24/06/2009

deformed configuration is the positions of all particles

under the influence of the loading.

Deformation: It refers to the displacements of the position

of all particles in the structures.

Boundary conditions: These refer to the prescribed values

of displacements or traction forces along the boundary of
the structures.

Essential boundary conditions: In structural mechanics,

these simply are the prescribed displacements at the
support of the structures. (Hence, they are essential as
without them the structure will move or collapse). The
essential boundary conditions are also referred as the
geometric or kinematics boundary conditions.

Natural boundary conditions: In structural mechanics,

these are simply the prescribed boundary traction forces
(loadings).

Admissible configuration: Any configuration of the system

that satisfies internal compatibility (Eqn. 3.6) and the
essential boundary conditions.

Fig. 5.3

L5-7
 CV7001_L5.docx 24/06/2009

For example, for the beam in Fig. 5.3a, the essential

boundary conditions are that at x=0, w=0, w,x=0 (slope=0).
Then among the four configurations in Fig. 5.3b,
configuration A and B are not admissible (why?) while the
reminding two are admissible.

Note:
(1) For any system => An infinite number of possible
configurations.

Among all the configurations => An infinite number

of admissible configurations.

Among all admissible configurations => Only a

unique (one and only one) true (real, physical)
configuration (which will, of course, satisfy the
equilibrium condition) exists.

How to find out the true configuration => Using the

Principle of stationary potential energy

Principle of stationary potential energy

For conservative systems, of all admissible displacement
fields, those corresponding to equilibrium extremize the
total potential energy. If the extremum condition is a
minimum, the equilibrium state is stable and
corresponding to the real physical configuration.

Note:
(1) There may be more than one admissible configurations
that satisfy the equilibrium but among them only one

L5-8
 CV7001_L5.docx 24/06/2009

will extremize (minimize) the potential energy of the

system and is stable.

For example, consider the case of a column under

compressive force=> under any compressive forces,
two admissible configurations => (I) buckling mode
and (II) column reminds straight.

If the loading is below the Euler load => (II) is the real
configuration and it will minimize the potential
energy.

If compressive force increase > Euler load => column

buckle and (I) is the true configuration.

(2) Assume that a system has attended the true

configuration so that the total potential energy of the
system is minimized. If now we perturb the system by
applying a small admissible variation of displacements
to the true configuration, we can conclude that

(I) the new configuration after the perturbation will

no longer be the true configuration and
(II) the new configuration will thus has greater
potential energy than the true configuration.

Hence the stationary principle can be rephrased as

Among all admissible configurations of a conservative

system, those that satisfy the equations of equilibrium
make the potential energy stationary with respect to small
admissible variations of displacement.

L5-9
 CV7001_L5.docx 24/06/2009

Example

Reconsider the bar in Fig. 5.4, we would like to know the

displacement, u(x), of the bar at any cross section at a
distance x from the support.

y Area=A,
x
Young's Modulus=E
L L

x D
P
P

Fig. 5.4 Deformation of a bar

Assumed that D=displacement of the bar at the free end.

For a point x from the support (essential boundary
condition), the displacement u(x) is

u(x)=Dx/L (5.7)

stress and strain are then equal to

∂u D DE
ε= = , σ= εE = (5.8)
∂x L L

L5-10
 CV7001_L5.docx 24/06/2009

L1 DE D
Π P (D) = U + W = ∫ Adx - PD
0 2 L L
Hence, (5.9)
2
1 D AE
= − PD
2 L

If D is the true displacement (a constant) at the free end,

we now perturb it by a small value equal to δD (a
variable). The corresponding potential energy will become

1 (D + δD) 2 AE
Π P (D + δD) = − P(D + δD) (5.10a)
2 L

∂Π P
We must have = 0 when δD=0 (Why?) and
∂δD

∂Π P AE
=0⇒ (D + δD) - P = 0
∂δD δD = 0 L
(5.10b)
PL
⇒D=
AE

and the displacement along the bar is (using Eqn. 5.7)

u(x)=Px/AE (5.11)

Some important remarks:

(1) Instead of finding u(x), in the first step we express

u(x) in terms of (i) the displacement at a point (the free
end) of the bar and (ii) a function of x (x/L).

L5-11
 CV7001_L5.docx 24/06/2009

(2) In general, we select the displacement at some

convenient points call a node to represent the
displacement function. This point displacements are
commonly referred as nodal unknowns or nodal
degree of freedoms (dof) and the number of them is
called the numbers of degree of freedom (i.e. dof=1 in
the above example).

(3) The function used will govern the form or variation of

the final displacement solution. We used a linear
function x/L in the above example and we will never
obtain anything higher order then linear function (i.e.
the final displacement function will never of the form
u=bx+cx2). This function which interpolates the
solution is commonly called the interpolation or shape
function.

(4) After the nodal dof and the shape function are
selected, the solution procedure becomes quite
straightforward:

=> Express the strain component in terms of nodal

dofs and shape functions.
=> Obtain the expression of stress component using
the generalize Hooke's law (Eqn. 3.14c).
=> Write down the expression for the total potential
energy => it will depend on the nodal unknowns.
=> Using the principle of stationary potential energy
to obtain the nodal unknowns.
=> The displacement functions are obtained by
substituting the nodal unknowns back into the
original assumed solution form.

L5-12
 CV7001_L6.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001

FINITE ELEMENT METHODS

Lecture 6

The Rayleigh-Ritz Method

6.1 The Rayleigh-Ritz Method

The Rayleigh-Ritz method can be considered as a

generalization of the procedure described in Lecture 5 for
solving elasticity problems involving continuum.

Assume the general case of 3D problems, the

displacement solutions is described by three functions
u(x,y,z), v(x,y,z) and w(x,y,z) (Eqn. 3.1b). Since in
practice, the exact solution form could be very
complicated, this method tries to obtain an approximating
solution of the problem by assuming that the solution can
be written in the form
l
u = ∑ a i f i (x, y, z)
i =1
m
v = ∑ a i fi (x, y, z) (6.1)
i = l +1
n
w= ∑ a i f i (x, y, z)
i = l + m +1

L6-1
 CV7001_L6.docx 24/06/2009

The functions fi(x,y,z) i=1,....,n are a set of admissible

shape functions that satisfy the essential (support)
boundary conditions (c.f. x/L in Eqn. 5.7) and ai is a set of
generalized coordinates (c.f. D in Eqn. 5.7).

To obtain ai, Eqn. 6.1 is substituted into the expression of

ΠP to replace the exact displacement and by using the
stationary conditions that

∂Π p
=0 for i=1,....,n (6.2)
∂a i
Remarks

(1) The functions fi(x,y,z) must be admissible, otherwise,

the solution obtained will not be an admissible
solution. It is not required that they will satisfy the
natural boundary conditions. (However, if they
satisfy the natural boundary condition, the solution
will be more accurate.)

(2) The dof of the problem are clearly equal to n.

However, in general, ai are not always (thought often)
corresponding to nodal dof.

(3) The selection of the form of fi(x,y,z) is entirely up to

the analyst. However, it may find that for some
problems, some particular forms of fi(x,y,z) may lead
to more accurate results. Usually, but not necessarily,
fi are polynomials. However, one important point is
that fi must be linear independent. That is, it is not
allow to have the case that

L6-2
 CV7001_L6.docx 24/06/2009

k
f j = ∑ bi f i for some k, i ≤n (6.3)
i =1

(4) The analyst must decide how many terms are needed
in each series in order to achieve the accuracy
required. In general, the more the number of terms,
the more accurate the solution but the higher the
computational cost. (Why?)

(5) The solution obtained by the Rayleigh-Ritz Method is

obviously only an approximated solution. (Why?)

(6) Some particular choices of ai (nodal value) and fi

(finite element shape functions) will lead to the
commonly referred "finite element method". Hence,
the finite element method (in this course's level) is a
subset of Rayleigh-Ritz Method.

6.2 Example

Consider the bar under axial load as shown in Fig. 6.1 (T1
page 147)

Fig. 6.1 Bar under axial load (Fig. 4.5-1 of T1)

L6-3
 CV7001_L6.docx 24/06/2009

For the axial loaded bar, the boundary condition are

Essential boundary condition: x=0, u=0.0

Natural boundary condition: x=LT, σx=0.

The only strain and stress components that are considered

are εx and σx.

The total potential energy

L 1
Π P = ∫0 T E (u , x ) Adx − ∫0 T u(cx)dx
2 L
(6.4)
2

Use polynomial functions as our trial or shape functions

such that

u = a1x + a 2 x 2 + a 3 x 3 + ...... + a n x n (6.5)

(Why there is no constant term in Eqn. 6.5?)

Use one term such that u=a1x gives

ΠP=AELT(a1)2/2-c(LT)3a1/3 (6.6a)

dΠP/da1=0 yields a1=c(LT)2/3AE (6.6b)

cL2T x cL2T cL2T

and u= and ε x = and σ x = (6.6c)
3AE 3AE 3A

L6-4
 CV7001_L6.docx 24/06/2009

The results are plotted in Fig. 6.1b.

Use two terms such that u=a1x+a2x2 and use ∂ΠP/∂ai=0,

i=1,2 gives

⎡1 L T ⎤ ⎧ a1 ⎫ cL T ⎧ 4 ⎫
AELT ⎢ ⎨ ⎬= ⎨ ⎬ (6.7a)
⎣L T 4L T / 3⎥⎦ ⎩a 2 ⎭ 12 3L
⎩ T⎭

or
⎧ a1 ⎫ cL T ⎧ 7 L T ⎫
⎨ ⎬= ⎨ ⎬ (6.7b)
a
⎩ 2⎭ 12AE ⎩ − 3L T⎭

Therefore,

cLT
u= (7LT x - 3x 2 )
12AE (6.7c)
cL
σ x = T (7LT - 6x)
12A

As plotted again in Fig. 6.1, by using two terms, the

answer obtained is closer to the exact solution

c
(3x (LT ) - x 3 )
2
u= (6.8)
6AE

and again we can see that the solutions for strain and stress
are not so accurate when comparing with the solution of
displacement.

L6-5
 CV7001_L6.docx 24/06/2009

Remarks:
(1) From Fig. 6.1 it can be seen for both the cases of one
and two terms, the solutions of u are better than σx.
(Why?)

(2) We can see that the approximated solution obtained

satisfied the essential boundary condition (u=0 at
x=0) but not the natural boundary condition.

(3) What can you see for the matrix in Eqn. 6.7a?

(4) What will be the solution if three terms are used (i.e.
u=a1x+a2x2+a3x3)? Or in general, when the selected
shape functions are capable of representing the exact
field? (Please read T1 pages 148-149 for more
details)

(5) The terms x, x2, x3, ..... are called the bases of the
shape function and they should be independent of
each and other. In addition, one shall able to obtain
the same results if the two sets of shape functions
used have the same bases. E.g. if one try to solve the
problem using the shape functions

u=b1x+b2(x+2x2) (6.9)

The results obtained will be the same as Eqn. 6.7c.

(Why?)
In fact, only different selected positions of ai and the
bases of the shape function of fi will lead to different
approximated solutions.

L6-6
 CV7001_L6.docx 24/06/2009

(6) Try this example again using the sine series of shape
functions

u=a1sin(πx/LT)+ a2sin(2πx/LT)+a3sin(3πx/LT)+ ....... (6.10)

What are the solutions, will they be exact?

Summaries of the Rayleigh-Ritz method

(1) It is an approximated, numerical methods used with

the principle of stationary potential energy.

(2) It is in fact a general method and the so called finite

element method is only a subset of it.

(3) The most important part of the formulation of

Rayleigh-Ritz method is the selections of generalized
coordinates (or dof of the system) and their
bases/shape functions. The generalized coordinates
(dof) and the shape functions will define an
approximated solution of the problem everywhere in
the problem domain.

(4) We will see that for the cases of trusses and frames,
the selection of the generalized coordinates (dof) and
shape functions are predefined in most cases and
usually the analyst has very few choices. However,
for continuum problems, there will be a lot more
options (in fact infinite number of choices) for the
analyst to select the generalized coordinates and
shape functions.

L6-7
 CV7001_L6.docx 24/06/2009

(5) It will lead to a system of simultaneous equations

with the generalized coordinate as the unknowns.
Furthermore, the matrix of the simultaneous equation
is symmetric.

(6) In general, the solutions of displacements are more

accurate than the solutions obtained for strain and
stress. Furthermore, the solutions obtained will only
be an approximated solution unless the shape
functions used are capable of representing the exact
field.

L6-8
 CV7001_L7.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 7

Formulation of the truss problem

7.1 Formulation for truss problem

In this chapter, the principle of stationary potential energy

and the Rayleigh-Ritz method will be applied for the
solution of truss (2D and 3D cases) problems. We will see
that for these types of problems, the generalized
coordinates and shape functions used are almost "fixed" in
advance. This is due to (i) the geometry of the structures
and (ii) the form of loading on the structures is almost
fixed.

The truss problem

Consider Fig. 7.1 which shows a continuum body and a

2D truss

L7-1
 CV7001_L7.docx 24/06/2009

surface traction force

acting along the surface Problem domain made-up of bar members
ΓΦ

loading at nodal point

domain of the body=Ω

P
Γu, Support surface where u=0 Supports where u,v=0 or u=0,

Fig. 7.1 Comparing a truss with a continuum.

One can consider that the truss is only a special case of

continuum problem with

(1) The domain, Ω, is replaced by a number of distinct

bar members. The union of these finite numbers of
bar members (or elements) makes up the whole
problem domain.

(2) The truss is supported only at the end points of the

bar element (or at the nodes of the bar elements).

(3) Loading (point loading) only act at the end points of

the bar elements.

(4) From elementary static, we conclude that only axial

force is acting on the bar elements.

L7-2
 CV7001_L7.docx 24/06/2009

Remarks
(1) Even only axial force are present in the bar elements,
we should not forget that all six stresses/strains
components are in fact always present in the bar (Fig.
7.2) as the bar itself is a 3D solid. However, we know
that the axial stress along the bar will be much greater
than all other stress components (Why?).

Fig. 7.2 Axial force acting on a bar

(2) The total strain energy of the whole truss can be

computed by summing up the strain energy of
individual bar (since energy is only a scalar). Hence,

NE
U = ∑ Ui (7.1)
i =1

where NE is the number of bar element in the truss.

(3) The total work done by the external force can also be
summing up as

L7-3
 CV7001_L7.docx 24/06/2009

NP
W = − ∑ Piu i (7.2)
i =1

where NP is the number of point loading and ui is the

displacement at the loading point i.

Therefore,

NE NP
Π p = ∑ U i − ∑ Piu i (7.3)
i =1 i =1

Selection of generalized coordinates and shape function

for truss problem

For truss problem, since the problem domain is already

"discretized" into a finite number of bar elements and all
loading and support conditions are applied at the end
points of these bar elements (nodal point), the most
convenient choice of the generalized coordinates (dofs)
are the displacements at the nodal points of the truss.
Thus, if the global numbering of nodes of the truss is as
shown in Fig. 7.3, then the unknowns of the problem
defined at the five nodal points will be given by

[u1,v1, u2,v2, u3,v3, u4,v4, u5,v5]T (7.4)

Now, we need to represent Eqn. 7.3 in terms of the

unknowns in Eqn. 7.4 together with other inputs
parameters including (i) material properties (E) and
dimension of the truss (e.g. bar length and cross section
area).

L7-4
 CV7001_L7.docx 24/06/2009

3 4 4
y, v

7
1
2 5

1 5
3 2 6
x, u
P

Fig. 7.3 Numbering of truss, NE=7, NP=1, NN=5

From Fig. 7.3, we see that our example contains (i) 7

elements (NE=7), five nodal points (NN=5) and one
loading point (NP=1).

For the selection of shape functions, it is clear that even

through all the bar elements may not be the same (e.g. will
have different length and other properties), they are
similar in the sense that they can be described by the same
set of parameters (e.g. length of the bar, L, Young's
modulus, E and cross sectional area, A).

Hence, we can focus on a typical bar elements and

obtained its strain energy expression. The results can be
repeatedly applied to all the members of the truss so that
we can compute Eqn. 7.3.

Consider a general case for a bar element shown in Fig.

7.4. For convenient we set up a local coordinate x' along
the bar and a local numbering of the end nodes of the
elements 1' and 2'.

L7-5
 CV7001_L7.docx 24/06/2009

Note that in this case, as only axial force is present, we

only need to consider the axial displacement of the bar.

Now, denote the displacements at the nodal point 1' and 2'
of the bars with respect to the local coordinate system x'
as u1' and u2' respectively. In addition, let L, E and A be,
respectively the length, Young's modulus and cross
sectional area of the bar.

f1'2'
y,v
node 2', (x2,y2), u2', x'=L
x',u'

node 1', (x1,y1), u1', x'=0,

f1'2'
x,u

Fig. 7.4 A bar element

As shown in Fig. 7.4, at node 1', x'=0 and at node 2', x'=L
and the corresponding global coordinates of node 1' and
node 2' are (x1,y1) and (x2,y2) respectively.

Hence, for node 1' and node 2', only one dof of
displacement per node (in x' direction) is present. In fact,
along any point 0≤x'≤L on the bar, only one displacement
field, u', are needed to be found. Hence, we only need to
select shape function for u'. The most simplest valid shape
function is linear variation along the x' direction (why
constant shape function is not valid?) such that

L7-6
 CV7001_L7.docx 24/06/2009

u'=a+bx' (7.5a)

Note that at x'=0, u'=u1' and x'=L, u'=u2', so we can write

u1'=a and u2'=a+bL (7.5b)

Solves for a and b gives

a=u1' and b=(u2'-u1')/L (7.5c)

Hence,

u'=u1'+x'(u2'-u1')/L (7.5d)

or u'=(1-x'/L)u1'+(x'/L)u2' (7.5e)

or u'=N1(x')u1'+N2(x')u2' (7.5f)

with N1(x')=(1-x'/L) and N2(x')=x'/L (7.5g)

Even though Eqn. 7.5d is identical (in what sense?) with

Eqns. 7.5e and 7.5g, the latter two expressions are more
useful as they express the variation of unknown functions
in terms of the nodal degree of freedoms.

Furthermore,

N1(0)=1, N1(L)=0 and N2(0)=0, N2(L)=1 (7.6)

so that the shape functions N1 and N2 assume a value of

unity (1.0) at its own nodal point and vanish (=0) at the
other nodal point. It will be shown that this property is a

L7-7
 CV7001_L7.docx 24/06/2009

very useful property in creating standard shape functions

in finite element analysis.

Note that we can also consider Eqn. 7.5f as the product of

a matrix (in this case a vector) and the displacement vector
as

⎡ u1' ⎤
u'=[N1(x'), N2(x')] ⎢ ⎥ =N'd' (7.7)
⎣u 2' ⎦

With the shape functions selected, we can carry on to

express the strain energy in terms of the nodal dofs.

In addition, one can also write N1 and N2 in the from

N1=(1-ξ), N2=ξ, ξ=x'/L (7.8)

The normalize coordinate ξ is called the parent (natrual)

coordinate with range [0,1]. By using the natural
coordinate, it is now possible to normalize all shape
functions so that the nodal dof can be expressed in terms
independent of the geometry parametric of the bar.

Only axial strain/stress are considered

∂u' ∂N1 ∂N 2
εx'= = u1' + u 2' = B' d' =(u2'-u1')/L (7.9a)
∂x ′ ∂x' ∂x'

⎡ ∂N ∂N ⎤
where B'= ⎢ 1 , 2 ⎥ = [-1/L, 1/L] (7.9b)
⎣ ∂x' ∂x' ⎦

L7-8
 CV7001_L7.docx 24/06/2009

The matrix (in this case a column vector) B' relates the
nodal displacement vector d' and the strain of the element
is often called the strain-displacement matrix.

To obtain the stress in the bar elements, all we need to do

is to multiply the strain with the constitutive matrix (Eqn.
3.14c) (in this case just a constant, E)

σx'=Eεx'=EB'd'=E(u2'-u1')/L (7.10)

Hence, both stress and strain are constant within the

element and the bar element is a constant strain element.
Furthermore, from Eqns. 7.9a and 7.10, it can be seen that
the expressions for strain and stress are exactly
corresponding to the simple bar expression model used in
structural analysis. This justifies that use of the linear
shape function in Eqn. 7.5f.

Could we use more terms in the shape functions? For

example

u'=a+bx'+c(x')2 (7.11)

If not, why?

Come back to the total potential energy of the bar

elements, using Eqn. 5.3a

L1 L1
U= ∫0 σ x'ε x'Adx' = ∫0 d'T (B'T EB' )d' Adx' (7.12a)
2 2

L7-9
 CV7001_L7.docx 24/06/2009

(Note for the arrangements for the matrix B' and vector
d'.)

Using Eqns. 7.7 and 7.9, Eqn. 7.12a can be simplified to

EA
U= (u 2' − u1' ) 2 (7.12b)
2L

and the total potential energy

EA
ΠP = (u 2' − u1' ) 2 - (f1'2'u1' + f1'2'u 2' ) (7.13)
2L

Using the stationary principle give

EA(u1'-u2')/L=f1'2' (7.14a)

EA(u2'-u1')/L=f1'2' (7.14b)

EA ⎡ 1 − 1⎤ ⎡ u1' ⎤ ⎡f1'2' ⎤
or = (7.14c)
L ⎢⎣− 1 1 ⎥⎦ ⎢⎣u 2' ⎥⎦ ⎢⎣f1'2' ⎥⎦

or k'd'=f' (7.14d)

which is the familiar stiffness equation of truss which

relates the nodal displacements and the nodal forces of a
bar element. The matrix k' is commonly called the
stiffness matrix (or in finite element => element stiffness
matrix).

L7-10
 CV7001_L7.docx 24/06/2009

Remarks
(1) Note that the force f1'2' is not the external loading, it
is the axial force acting on the bar member and Eqn.
7.12b is the strain energy expression of on bar
member only.

(2) The expressions for the matrix B' and k' and vectors
d' and f' are written with respect to the local
coordinate x'. They are required to be transformed to
the global coordinate system (x,y) before combining
with the results from other bar elements in the
system.

(3) Reconsider Eqn. 7.12a, the term

L1 T T
∫0 d' (B' EB' )d' Adx' can be re-written as
2

1 T
2
d' [∫ (B' EB')Adx']d'
L
0
T
(7.15a)

since d' does not depend on x'. Furthermore, the term

inside the square bracket in fact equals to k'. That is

[ L
][
k'= ∫0 (B'T EB' )Adx' = ∫Ω (B'T EB' )dΩ ] (7.15b)

and Eqn. 7.15b in fact is a general expression for many

element stiffness matrices.

L7-11
 CV7001_L7.docx 24/06/2009

(4) Thus, we can write the expression of potential energy

as

1
Π P = d'T k ' d'-(d') f '
T
(7.16)
2

(Prove that by taking the vector differentiation,

∂Π P
= 0 , one can also obtain Eqn. 7.14c.)
∂d'

(5) Since we use linear shape functions for the

displacement, the strain and stress of the bar are
constant within the element. Therefore, we can expect
that the numerical solutions obtained by using this
element will be exact only when the exact strain is
also constant within the bar element. This is normal
the case when (i) loadings are only applied at the
nodal points of the truss and (ii) the cross sectional
properties of the bar members are constant along the
bar elements.

L7-12
 CV7001_L8.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 8

Coordinate transformation and frame problem

8.1 Element stiffness matrix with respect to the

global coordinates

In the last lecture, by using the stationary principle and the

Ritz method, we obtained the element stiffness equations
in the local, element coordinate system (x'). Obviously, we
must try to express the element stiffness matrix in terms of
the global coordinates (x,y) so that we can combine the
contributions of all elements.

Note that the local displacement u' must be transformed to

two displacement components [u,v] in the global (x,y)
coordinate system (Figs. 7.4 and Eqn. 8.1) such that

⎡u ⎤
u' = [c s]⎢ ⎥ (8.1a)
⎣v⎦

(Prove Eqn. 8.1a yourself!)

where c=cosβ and s=sinβ (8.1b)

L8-1
 CV7001_L8.docx 24/06/2009

β is the angle between the bar element and the x axis.

f1'2'
y,v
node 2', (x2,y2), u2', x'=L
x',u'

node 1', (x1,y1), u1', x'=0,

f1'2'
x,u

Fig. 7.4 A bar element

y, v (x2,y2),
global: u2,v2
Local: u2'

β
(x1,y1)
global: u1,v1
Local: u1'
x,u

Fig. 8.1 Coordinate transformation

Obviously

c=(x2-x1)/L and s=(y2-y1)/L (8.1c)

⎡ u1 ⎤
⎢ ⎥
⎡ u1' ⎤ ⎡ c s 0 0⎤ ⎢ v1 ⎥
and ⎢u ⎥ = ⎢0 0 c s ⎥ ⎢u ⎥ or d' = [T]d (8.2a)
⎣ 2' ⎦ ⎣ ⎦ 2
⎢ ⎥
⎣v2 ⎦

L8-2
 CV7001_L8.docx 24/06/2009

[T] = ⎡⎢
c s 0 0⎤
such that ⎥ (8.2b)
⎣0 0 c s ⎦

or in matrix form: d'=Td (8.2c)

By using Eqn. 8.2a, Eqn. 7.16 becomes

Π P = (Td )T k ' (Td ) - (Td ) f '

1 T
(8.3a)
2

1
or Π P = d T (T T k ' T)d - d T (T T f ' ) (8.3b)
2

1
Hence Π P = d T kd - d T f (8.3c)
2

where k=TTk'T and f=TTf' (8.3d)

The term f=TTf'=[f1x, f1y, f2x, f2y]T is the end forces acting
on the nodes of the bar (Fig. 8.2).
f2y
y, v f2x

f1y

f1x

x,u

Fig. 8.2 End forces acting on the bar element

L8-3
 CV7001_L8.docx 24/06/2009

(Prove that f=TTf' is correct with respect to Fig. 8.2

yourself!)

∂Π P
Upon using the stationary principle = 0 gives
∂d

kd=f (8.4)

and the element stiffness matrix in the global coordinate

system is given by

⎡ c2 cs - c 2 - cs ⎤
⎢ ⎥
EA ⎢ cs s 2 - cs - s 2 ⎥
k= (8.5)
L ⎢- c 2 - cs c 2 cs ⎥
⎢ 2 2 ⎥
⎣ - cs - s cs s ⎦

(Prove that k is semi-positive definite and why it must be

semi-positive definite?)

Note that Eqn. 8.5 is, of course, exactly the same as the
stiffness matrix obtained from structural mechanic theory.

3D formulation

For the case of 3D truss problem, the expression of the bar

element stiffness matrix can also be obtained in a similar
way.

First consider the local coordinate system x' again as

shown in Fig. 8.3, it will not be difficult to re-obtain Eqns.
7.14 and 7.16 again.

L8-4
 CV7001_L8.docx 24/06/2009

For coordinate transformation, we need to define

c=(x2-x1)/L and s=(y2-y1)/L and e=(z2-z1)/L (8.6)

f2z
f1'2'
z,w
f2y
f2x
node 2', (x2,y2,z2), u2', x'=L
x',u'
f1z
f1'2'
node 1', (x1,y1,z1), u1', x'=0,
f1y
f1x y,v
x,u

Fig. 8.3 3D truss problem

and the relationship between u' and [u,v,w] is

⎡u⎤
u′ = [c s e]⎢ v ⎥ (8.7)
⎢ ⎥
⎢⎣ w ⎥⎦

Then the corresponding results for 3D cases are:

Transformation matrix

[T] = ⎡⎢
c s e 0 0 0⎤
⎥ (8.8)
⎣0 0 0 c s e ⎦

Finally, the stiffness matrix is given by

L8-5
 CV7001_L8.docx 24/06/2009

⎡ c2 cs ce - c 2 - cs - ce ⎤
⎢ ⎥
⎢ cs s2 se - cs s 2 - se ⎥
EA ⎢ ce se e 2 - ce - se - e 2 ⎥
k= ⎢ ⎥ (8.9)
L ⎢ - c2 - cs - ce c 2 cs ce ⎥
⎢ - cs s 2 - se cs s 2 se ⎥
⎢ ⎥
⎢⎣− cf - se - e 2
ce se e ⎥⎦2

(Derive Eqn. 8.9 from the first principle yourself!).

8.2 Formulation of frame problem

2D frame formulation

For 2D frame problem, we first consider the case without

any axial force (Fig. 8.4).
q
S2'

M2'
z',w'
y,v
x',u' node 2', (x2,y2), x'=0, w2', θ2'

S1'
M1'
node 1', (x1,y1), x'=0, w1', θ1'

x,u

Fig. 8.4 2D frame problem

L8-6
 CV7001_L8.docx 24/06/2009

We should remember that again the "true" beam again is a

3D solid but we only consider the most
important/dominating strain/stress components =>
Bending strain/stress => generalized coordinates=> end
vertical displacements and rotations.

Assumed that the cross sectional properties of the member

are constant along the length of the element. From
elementary beam theory, we can write

M' σ M'
σ x' x' = z' ε x' x' = x' x' = z' (8.10a)
I E EI

and if the beam rotation is small

dw'
θ'≈ (8.10b)
dx'

d 2 w' M'
2
= (8.10c)
dx' EI

Therefore,

d 2 w' d 2 w'
σ x' x' = E 2 z' and ε x'x' = 2
z′ (8.10d)
dx' dx'

Again, to obtain the strain energy of the beam, we need to

carry out integration

L8-7
 CV7001_L8.docx 24/06/2009

2
⎛ d 2 w' ⎞
U = ∫ ∫ σ x' x'ε x' x'dAdx' = ∫ ∫ E⎜ 2 ⎟ (z')2 dAdx'
1 L 1 L
2 0 A 2 0 A ⎜⎝ dx' ⎟⎠
(8.11)

( )
2 2
1 L ⎛⎜ d 2 w' ⎞⎟ 1 L ⎛⎜ d 2 w' ⎞⎟
= ∫ E ∫A (z') dA dx' = 2 ∫0 EI⎜ dx'2 ⎟ dx'
2
2 0 ⎜⎝ dx'2 ⎟⎠ ⎝ ⎠

For the work done by the external distributed load, q(x'),

and the end shear forces (S1', S2') and end moments (M1',
M2'):

W = -⎛⎜ ∫ q(x)w' dx + S1' w1' + M1'θ1' + S2' w 2' + M 2'θ2' ⎞⎟

L
⎝ 0 ⎠
(8.12a)
= -⎛⎜ ∫ q(x)w' dx + (d')T f ' ⎞⎟
L
⎝ 0 ⎠

⎡ w1' ⎤ ⎡ S1' ⎤
⎢θ ⎥ ⎢M ⎥
where d' = ⎢ 1' ⎥ and f ' = ⎢ 1' ⎥ (8.12b)
⎢ w 2' ⎥ ⎢ S2' ⎥
⎢ ⎥ ⎢ ⎥
⎣ θ2' ⎦ ⎣M 2' ⎦

Note that now we have four local dofs for the element.
Thus, the total potential energy is equal to

2
1 L ⎛ d 2 w' ⎞
Π P = ∫0 EI⎜⎜ 2 ⎟⎟ dx - ∫0 q(x' )w' dx - (d') f ' (8.13)
L T
2 ⎝ dx' ⎠

L8-8
 CV7001_L8.docx 24/06/2009

Selection of shape functions

As we have four dofs, the expression of w' should has four

terms, again use polynomial and use the parent coordinate
ξ=x'/L

w'=a+bξ+cξ2+dξ3 (8.14)

dw' dw' dξ 1 dw'

Thus, from Eqn. 8.10b and = =
dx' dξ dx' L dξ

θ'=(b+2cξ+3dξ2)(1/L) (8.15)

(Why we cannot express θ' in terms of other parameter

and must use Eqn. 8.15?)

Using the conditions that at ξ=0=>w'=w1', θ'=θ1' and at

ξ=1=>w'=w2', θ'=θ2'
w1'=a, θ1'L=b, w2'=a+b+c+d, θ2'L=b+2c+3d (8.16)

Solving gives

w'=w1'+θ1'Lξ+ξ2(3w2'-3w1'-2θ1'L-θ2'L)+ξ3(2w1'-2w2'+θ1'L+θ2'L)
(8.17a)
or

w'=w1'(1-3ξ2+2ξ3)+θ1'L(ξ-2ξ2+ξ3)+w2'(3ξ2-2ξ3)+ θ2'L(ξ3-ξ2)
(8.17b)

Again we rewrite (Eqn. 8.17b) in the form

L8-9
 CV7001_L8.docx 24/06/2009

w'=N1(ξ)w1'+LN2(ξ)θ1'+N3(ξ)w2'+LN4(ξ)θ2' (8.17c)

or w'=N'd' (8.17d)

The functions Ni(ξ) (sometimes denoted as Hi(ξ)) are

known as the Hermite shape functions.

Using Eqn. 8.17b, the strain-displacement B matrix is

given by

d 2 w' ⎡ 6 12x' 4 6x' 6 12x' 2 6x' ⎤

= ⎢ - + , - + , − , - + 2 ⎥d' = Bd'
⎣ L L L ⎦
2 2 3 2 2 3
dx' L L L L L
(8.18)

d 2 w'
(Note that for frame/beam element, the curvature 2
is
dx'
considered as the strain.)
Thus, Eqn. 8.13 can be written as

ΠP =
1
2
( )
(d')T ∫0L BT EIBdx' d'-(d')T (∫ N' q(x')dx )- (d') f '
L
0
T T

(8.19a)
1
= (d') k ' d'-(d') fe '-(d') f '
T T T
2
L L
with k ' = ∫0 B T EIBdx' and fe ' = ∫0 N'T q(x)dx' (8.19b)

and upon using the stationary principle

k'd'=fe'+f' (8.20)

L8-10
 CV7001_L8.docx 24/06/2009

such that the element stiffness matrix is given by

⎡ 12 6L − 12 6L ⎤
⎢ 2 ⎥
EI ⎢ 6L 4L − 6L 2L ⎥
2
k' = 3 (8.21)
L ⎢− 12 − 6L 12 − 6L⎥
⎢ 2 ⎥
⎣ 6L 2L − 6L 4L ⎦
2

which is, of course, exactly equal to the stiffness matrix

obtained by other structural mechanics approach (e.g.
slope deflection method).

If the distributed loading is constant such that q(x')=q,

then the load vector fe' will equal to

⎡ qL/2 ⎤
⎢ qL2 / 12 ⎥
f e′ = ⎢ ⎥ (8.22)
⎢ qL/2 ⎥
⎢ 2 ⎥
⎣- qL / 12⎦

Frame with axial force

Now consider a frame with axial force as shown in Fig.

8.5, one can consider it as the "sum" of a truss element and
a frame element without axial force when using the
stationary principle as the energy terms corresponding to
axial force and bending moment can be summed up
directly.

L8-11
 CV7001_L8.docx 24/06/2009

q
f2'
S 2'
M2'
z',w'
y,v node 2', (x2,y2), x'=L, u2', w2', θ2'
x',u'

M1'
f1' S 1'
node 1', (x1,y1), x'=0, u1',w 1', θ1'

θ
x,u

Fig. 8.5 Frame with axial force

Using a similar procedure (derive it once again yourself!),

it can be shown that the element stiffness matrix and the
load vectors are given by

d'=[ u1', w1', θ1', u2', w2', θ2']T (8.23a)

f'=[ f1', S1', M1', f2', S2', M2']T (8.23b)

⎡ EA EA ⎤
⎢ L 0 0 - 0 0 ⎥
L
⎢ 12EI 6EI 12EI 6EI ⎥
⎢ 3 2
0 - 3 ⎥
⎢ L L L L2 ⎥
⎢ 4EI
0
6EI
- 2
2EI ⎥
⎢ L ⎥
k′ = ⎢ L L
⎥ (8.24)
EA
⎢ 0 0 ⎥
⎢ L ⎥
⎢ 12EI 6EI ⎥
sym. - 2
⎢ L3 L ⎥
⎢ 4EI ⎥
⎢⎣ L ⎥⎦

L8-12
 CV7001_L8.docx 24/06/2009

8.3 Transformation to global coordinate system

Similar to the truss problem, the displacement in the local

and global coordinate system can be related as

⎡ u' ⎤ ⎡ c s 0⎤ ⎡u ⎤
⎢ w'⎥ = ⎢- s c 0⎥ ⎢ v ⎥ (8.25)
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢⎣ θ' ⎥⎦ ⎢⎣ 0 0 1⎥⎦ ⎢⎣ θ ⎥⎦

(Prove Eqn. 8.25 yourself!)

Hence the transformation matrix will equal to

⎡c s 0 0 0 0⎤
⎢- s c 0 0 0 0⎥
⎢ ⎥
⎢0 0 1 0 0 0⎥
T=⎢ ⎥ (8.26a)
⎢0 0 0 c s 0⎥
⎢0 0 0 -s c 0⎥
⎢ ⎥
⎣0 0 0 0 0 1⎦

such that

k=TTk'T (8.26b)

and

f=TTf' and fe=TTfe' (8.26c)

where f=[f1x, f1y, M1, f2x, f2y, M2]T (8.26d)

L8-13
 CV7001_L8.docx 24/06/2009

as shown in Fig. 8.6

f2y

y,v f2x
M2
f1y

f1x
M1
θ
x,u

Fig. 8.6 Forces action at the end of a beam element

L8-14
 CV7001_L9.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 9

Assembly of element stiffness matrix and load

vector

9.1 Assembly of element stiffness matrix

In the last lecture, it is shown that by using the principle of

stationary potential energy and the Rayleigh-Ritz method,
and after appropriate coordinate transformation, we can
always obtain the element stiffness equations of the form

kd=f (9.1)

where k is called the element stiffness matrix, d is a set of

element nodal displacements with respect to the global
coordinate system. f is a set of element nodal force (not
the external loading force P) acting on the element. For
2D and 3D truss problems, k is a 4×4 and a 6×6 matrix
respectively while d and f are 1×4 and 1×6 vectors (Eqns.
8.5 and 8.9)

(What will be the size of k, d and f for 2D and 3D frame

problems?)

L9-1
 CV7001_L9.docx 24/06/2009

As Eqn. 9.1 is the general expression of a typical element,

we can repeatedly apply it to all the elements (truss or
frame members) in the structure. However, one can never
solves the element stiffness equations (Eqn. 9.1) one-by-
one as one can be observed that

(1) Both d (nodal displacements) and f (member force)

are, in general, unknowns.

(2) Even when f is given, it can be shown easily that the

matrix k is singular.

(What are the corresponding physical meanings for the

above two observations?)

Thus, in order to solve for the values of d, we must

(1) Assembly all the element stiffness matrices together

to form the global stiffness matrix of the whole
structure.

(2) Assembly all the element nodal force vectors together

and apply the external loading to form the global
loading or right hand size (RHS) vector acting on the
structures.

(3) Impose the boundary support conditions on the

structures.

(What are the corresponding physical meanings of the

above three actions?)

L9-2
 CV7001_L9.docx 24/06/2009

9.2 Assembly regarded as satisfying equilibrium

To most engineers (or engineering students!), the easiest

way to interpret/understand the assembly process is to
regard it as a process to restore/achieve the equilibrium
status of the structures.

Reconsider the truss example used in Lecture 7 (Fig. 7.3)

as shown below again. Now, let us consider joint 3 of the
structures. As three members, 1, 2 and 4 are connected to
joint 3, the assembly process is started by written down
the element stiffness equations for these three members.

3 4 4
y, v

7
1
2 5

1 5
3 2 6
x, u
P

Fig. 7.3 Numbering of truss, NE=7, NP=1, NN=5

For member 1, if we denoted it as 1-3, we have, as shown

in Fig. 9.1
k 1d1 = f1
⎡ c12 c1s1 - c12 - c1s1 ⎤ ⎡ u1 ⎤ ⎡f 11x ⎤
⎢ ⎥⎢ ⎥ ⎢ ⎥
E 1A 1 ⎢ c1s1 s12 - c1s1 - s12 ⎥ ⎢ v1 ⎥ ⎢f 11y ⎥ (9.2)
=
L1 ⎢ - c12 - c1s1 c12 c1s1 ⎥ ⎢ u 3 ⎥ ⎢f 13x ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣- c1s1 - s12 c1s1 s12 ⎦ ⎣ v 3 ⎦ ⎣f 13y ⎦

L9-3
 CV7001_L9.docx 24/06/2009

where subscript 1 denote parameters corresponding to

element 1.
f3y

f3x
y, v u3,v3
1

f1y

f1x
u1,v1
x,u

Fig. 9.1 Member 1

For simplicity, from now on, we rewrite Eqn. 9.2 (and

subsequently equations corresponding to elements 2 and
4) as

⎡11x1x 11x1y 11x3x 11x3y ⎤ ⎡ u1 ⎤ ⎡f11x ⎤

⎢1 11y1y 11y3x 11y3y ⎥ ⎢ v1 ⎥ ⎢f11y ⎥
⎢ 1y1x ⎥⎢ ⎥ = ⎢ ⎥ (9.3a)
⎢13x1x 13x1y 13x3x 13x3y ⎥ ⎢u 3 ⎥ ⎢f13x ⎥
⎢ ⎥ ⎢ ⎥
⎣13y1x 13y1y 13y3x 13y3y ⎦ ⎢⎣ v3 ⎥⎦ ⎣f13y ⎦

Similarly, we can write down two sets of element stiffness

equations for members 2 (3-2) and 4 (3-4)

⎡ 23x3x 23x3y 23x2x 23x2y ⎤ ⎡ u 3 ⎤ ⎡f 23x ⎤

⎢2 23y3y 23y2x 23y2y ⎥ ⎢ v3 ⎥ ⎢f 23y ⎥
⎢ 3y3x ⎥⎢ ⎥ = ⎢ ⎥ (9.3b)
⎢22x3x 22x3y 22x2x 23x2y ⎥ ⎢u 2 ⎥ ⎢f 22x ⎥
⎢ ⎥ ⎢ ⎥
⎣ 22y3x 22y3y 22y2x 23y2y ⎦ ⎢⎣ v 2 ⎥⎦ ⎣f 22y ⎦

L9-4
 CV7001_L9.docx 24/06/2009

and

⎡43x3x 43x3y 43x4x 43x4y ⎤ ⎡ u 3 ⎤ ⎡f 43x ⎤

⎢4 43y3y 43y4x 43y4y ⎥ ⎢ v3 ⎥ ⎢f 43y ⎥
⎢ 3y3x ⎥⎢ ⎥ = ⎢ ⎥ (9.3c)
⎢44x3x 44x3y 44x4x 44x4y ⎥ ⎢u 4 ⎥ ⎢f 44x ⎥
⎢ ⎥ ⎢ ⎥
⎣44y3x 44y3y 44y4x 44y4y ⎦ ⎢⎣ v 4 ⎥⎦ ⎣f 44y ⎦

Note the relationship between the coefficients in Eqns.

9.3a-9.3c and the local node numberings (i.e. 1-3, 3-2 and
3-4) of the elements.

Using the equilibrium conditions at joint 3, we have

-(f13x+ f23x+f43x)=0 and -(f13y+ f23y+f43y)=0 (9.4)

Eqn. 9.4 simply means to sum up the third and fourth

equations in Eqn. 9.3a with the first and second equations
of Eqns. 9.3b and 9.3c.

Similarly, by consider the equilibrium conditions at each

of the joints, one can always write down 2×NN numbers
of equations which described the equilibrium conditions of
the structures. In our example, we know that we will
finally obtain a matrix of size 10×10 (since NN=5).

In practice it is always more convenient to consider the

above joint equilibrium conditions are obtained by a
process of matrix summation (or assembly). Imagine that
one always starts with a 10×10 global stiffness matrix with
all entries equal to zero, the 4×4 element stiffness matrices

L9-5
 CV7001_L9.docx 24/06/2009

of the structures are then added or assembled one-by-one

to this 10×10 global stiffness matrix.

For example, if element 1 (Eqn. 9.3a) is the first element

to be assembled to the global stiffness matrix, then after it
is added to the system, the global stiffness matrix will
assume the form shown in Eqn. 9.5a. Note the relationship
between the indices of the entries and the row and column
they occupied.

⎡11x1x 11x1y 0 0 11x3x 11x3y 0 0 0 0⎤ ⎡ u1 ⎤ ⎡f 11x ⎤

⎢1 11y1y 0 0 11y3x 11y3y 0 0 0 0⎥ ⎢ v1 ⎥ ⎢f 11y ⎥
⎢ 1y1x ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ v 2 ⎥ ⎢ 0 ⎥
⎢13x1x 13x1y 0 0 13x3x 13x3y 0 0 0 0⎥ ⎢ u 3 ⎥ ⎢f 13x ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢13y1x 13y1y 0 0 13y3x 13y3y 0 0 0 0⎥ ⎢ v 3 ⎥ ⎢f 13y ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 0 0 0 v
0 0 0 0⎥ ⎢ 4 ⎥ ⎢ 0 ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 0 0 0 0 0 0 0 0⎥⎦ ⎢⎣ v 5 ⎥⎦ ⎢⎣ 0 ⎥⎦
(9.5a)

Eqn. 9.5b. shows the global stiffness equations system

after element 4 is assembled into the global system. (Note
that the element stiffness matrices need not be assembled
sequentially since matrix addition is commutative.)

L9-6
 CV7001_L9.docx 24/06/2009

⎡11x1x 11x1y 0 0 11x3x 11x3y 0 0 0 0⎤ ⎡ u1 ⎤ ⎡ f 11x ⎤

⎢1 11y1y 0 0 11y3x 11y3y 0 0 0 0⎥ ⎢ v1 ⎥⎥ ⎢ f 11y ⎥
⎢ 1y1x ⎥⎢ ⎢ ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ v 2 ⎥ ⎢ 0 ⎥
⎢13x1x 13x1y 0 0 13x3x + 43x3x 13x3y + 43x3y 43x4x 43x4y 0 0⎥ ⎢ u 3 ⎥ ⎢f 13x + f 43x ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢13y1x 13y1y 0 0 13y3x + 43y3x 13y3y + 43y3y 43y4x 43y4y 0 0⎥ ⎢ v 3 ⎥ ⎢f 13y + f 43y ⎥
⎢ 0 0 0 0 44x3x 44x3y 44x4x 44x4y 0 0⎥ ⎢u 4 ⎥ ⎢ f 44x ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 0 44y3x 44y3y 44y4x 44y4y 0 0⎥ ⎢ v 4 ⎥ ⎢ f 44y ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 0 0 0 0 0 0 0 0⎥⎦ ⎣⎢ v 5 ⎦⎥ ⎢⎣ 0 ⎥⎦

(Eqn. 9.5b)

L9-7
 CV7001_L9.docx 24/06/2009

⎡11x1x 11x1y 0 0 11x3x 11x3y 0 0 0 0⎤ ⎡ u1 ⎤ ⎡f11x ⎤

⎢1 11y1y 0 0 11y3x 11y3y 0 0 0 0⎥ ⎢ v1 ⎥ ⎢ f11y ⎥
⎢ 1y1x ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 22x2x 22x2y 22x3x 22x3y 0 0 0 0⎥ ⎢u 2 ⎥ ⎢f 22x ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 22y2x 22y2y 22y3x 22y3y 0 0 0 0⎥ ⎢ v 2 ⎥ ⎢f 22y ⎥
⎢13x1x 13x1y 23x2x 23x2y 13x3x + 43x3x + 23x3x 13x3y + 43x3y + 23x3y 43x4x 43x4y 0 0⎥ ⎢ u 3 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢13y1x 13y1y 23y2x 23y2y 13y3x + 43y3x + 23y3x 13y3y + 43y3y + 23y3y 43y4x 43y4y 0 0⎥ ⎢ v 3 ⎥ ⎢ 0 ⎥
⎢ 0 0 0 0 44x3x 44x3y 44x4x 44x4y 0 0⎥ ⎢u 4 ⎥ ⎢f 44x ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢ 0 0 0 0 44y3x 44y3y 44y4x 44y4y 0 0⎥ ⎢ v 4 ⎥ ⎢f 44y ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥ ⎢ u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 0 0 0 0 0 0 0 0⎥⎦ ⎢⎣ v5 ⎥⎦ ⎢⎣ 0 ⎥⎦

(Eqn. 9.5c)

L9-8
 CV7001_L9.docx 24/06/2009

Eqn. 9.5c shows the resulting global stiffness equations

system after element 2 is added to the global system. Note
that the assembly process is solely dictated by the global
node numberings of nodes in the elements. It is not
affected by the local node numbering (3-2) of the
elements. Even node 3 appears first in the element level
(as in element 3-2), the position of its corresponding terms
in the global stiffness matrix is always dictated by it global
numbering. In the case of truss and frame, the local
numbering of node within an element is immaterial. That
is element 4 can be labelled as 3-2 or 2-3. However, in
general finite element analysis, the local numbering of
nodes in an element must follow some predefined rules
(e.g. numbering in anti-closkwise manner). More details
will be given when we discuss 2D problems.

From Eqn. 9.5c, we can see that the equilibrium equations

corresponding to node 3 has been fully assembed (row 3
and 4 in Eqn. 9.5c) and the RHS vectors are already
summed to zero. Of course, the solution for u3 and v3
cannot be solved as the two unknowns are coupled with
other unknowns.

Repeat the above assembly procedures for all the seven

elements in the structure will eventually lead to the finial
global stiffness equations system of the form shown in
Eqn. 9.6a. In Eqn. 9.6a, an X indicates that the entry of the
stiffness matrix is non-zero.

(Try to establish and validate the form of the stiffness

matrix yourself! Also note where you will find the zero
entries)

L9-9
 CV7001_L9.docx 24/06/2009

⎡X X X X X X 0 0 0 0 ⎤ ⎡ u1 ⎤ ⎡ R 1x ⎤
⎢ ⎥
⎢X X X X X X 0 0 0 0 ⎥ ⎢ v1 ⎥ ⎢ R 1y ⎥
⎢ ⎥⎢ ⎥
⎢X X X X X X X X X X ⎥ ⎢u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎢X
⎢ X X X X X X X X X ⎥⎥ ⎢ v 2 ⎥ ⎢ - P ⎥
⎢X X X X X X X X 0 0 ⎥ ⎢u 3 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢X X X X X X X X 0 0 ⎥ ⎢ v3 ⎥ ⎢ 0 ⎥
⎢0 0 X X X X X X X X ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 X X X X X X X X⎥ ⎢v4 ⎥ ⎢ 0 ⎥
⎢0 0 X X 0 0 X X X X⎥ ⎢u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣⎢ 0 0 X X 0 0 X X X X ⎦⎥ ⎣⎢ v5 ⎦⎥ ⎢⎣R 5y ⎥⎦
(9.6a)

or KD=F (9.6b)

where K and F are the global stiffness matrix and the

global force vector respectively.
Now consider another set of global numbering scheme as
shown in Fig. 9.2. Show that it will lead to the pattern
shown in Eqn. 9.6c.

4 4 5
y, v

7
1
2 5

1 3
3 2 6
x, u
P
Fig. 9.2 An alternative numbering scheme of the truss

L9-10
 CV7001_L9.docx 24/06/2009

⎡X X X X 0 0 X X 0 0 ⎤ ⎡ u1 ⎤ ⎡ R 1x ⎤
⎢X X X X 0 0 X X 0 0 ⎥ ⎢ v1 ⎥ ⎢ R 1y ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X X X X X X ⎥ ⎢u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X X X X X X⎥ ⎢v2 ⎥ ⎢ - P ⎥
⎢0 0 X X X X 0 0 X X⎥ ⎢u 3 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢0 0 X X X X 0 0 X X ⎥ ⎢ v 3 ⎥ ⎢R 3y ⎥
⎢X X X X 0 0 X X X X ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X 0 0 X X X X⎥ ⎢v4 ⎥ ⎢ 0 ⎥
⎢0 0 X X X X X X X X ⎥ ⎢u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 X X X X X X X X ⎥⎦ ⎢⎣ v 5 ⎥⎦ ⎢⎣ 0 ⎥⎦
(9.6c)

From Eqns. 9.6a to 9.6c, we can conclude that

(1) The pattern of non-zero entries of the global stiffness

matrix is closely related to the node numbering and
connectivity of the elements (but not the numbering of
the elements!).

(2) The matrices are symmetric (prove it yourself!) and is

not full.

(3) The external loading -P is added to the equation

corresponding to v2.

(4) For Eqn. 9.6a, in equations corresponding to

unknowns u1, v1 and v5, the corresponding terms in the
RHS vector are the reactions from the supports (R1x,
R1y and R5y). (What about the case Eqn. 9.6c?) In

L9-11
 CV7001_L9.docx 24/06/2009

general, the reactions from the supports are not known

in advance (as the structure is very likely to be
statically indeterminate) and can be computed only
when the displacements (ui, vi, i=1,...,5) are founded.

(5) If one checks carefully, it can be easily shown that the

matrix in Eqn. 9.6 will be singular and cannot be
inverted to solve the displacements (why?).

So it is not enough just to assemble the stiffness matrix

(the structure) alone, we must also impose the boundary
conditions (provide supports to the structures) on the
global equations system.

As we note in advance that for the numbering scheme

shown in Fig. 7.3, u1=v1=v5=0. By reconsidering Eqn. 9.6a
again, we can discover that they are corresponding to the
unknown reactions on the RHS vector.

Now, using that conditions that

Xu1=1×u1=0, Xv1=1×v1=0, Xv5=1×v5=0 (9.7a)

for any coefficient X. We can modify Eqn. 9.6a to the

form shown in Eqn. 9.8a so that

(1) the unknown reactions (R1x, R1y and R5y) no longer

exist in the system,

(2) the size of the stiffness matrix system reminds

unchanged.

L9-12
 CV7001_L9.docx 24/06/2009

Note that Eqn. 9.8a is equivalent to cross out all the rows
and columns corresponding to the unknowns u1,v1 and v5
(or the essential boundary conditions) and is equivalent to
solve the system shown in Eqn. 9.8b.

⎡1 0 0 0 0 0 0 0 0 0⎤ ⎡ u1 ⎤ ⎡ 0 ⎤
⎢0 1 0 0 0 0 0 0 0 0⎥ ⎢ v1 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 X X X X X X X 0⎥ ⎢ u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 X X X X X X X 0 ⎥ ⎢ v 2 ⎥ ⎢- P ⎥
⎢0 0 X X X X X X 0 0⎥ ⎢ u 3 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢0 0 X X X X X X 0 0⎥ ⎢ v 3 ⎥ ⎢ 0 ⎥
⎢0 0 X X X X X X X 0⎥ ⎢ u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 X X X X X X X 0⎥ ⎢ v 4 ⎥ ⎢ 0 ⎥
⎢0 0 X X 0 0 X X X 0⎥ ⎢ u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣0 0 0 0 0 0 0 0 0 1⎥⎦ ⎢⎣ v 5 ⎥⎦ ⎢⎣ 0 ⎥⎦
(9.8a)

⎡X X X X X X X ⎤ ⎡u 2 ⎤ ⎡ 0 ⎤
⎢X X X X X X X ⎥ ⎢ v 2 ⎥ ⎢- P ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X X 0 ⎥ ⎢u 3 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X X 0 ⎥ ⎢v2 ⎥ = ⎢ 0 ⎥ (9.8b)
⎢X X X X X X X ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X X X⎥ ⎢v 2 ⎥ ⎢ 0 ⎥
⎢⎣X X 0 0 X X X ⎥⎦ ⎣⎢ u 5 ⎥⎦ ⎢⎣ 0 ⎥⎦

(Work out the corresponding forms for Eqns. 9.8a and

9.8b for numbering scheme shown in Fig. 9.2.)

L9-13
 CV7001_L9.docx 24/06/2009

Note that in practice (computer implementation), we

normally prefer to use Eqn. 9.8a rather than Eqn. 9.8b.
The reason is that during the assembly process, we already
reserve memory for storing the 10×10 matrix
corresponding to the total dofs of the system. It is
definitely more convenient to adjust the values of the
entries of the global stiffness matrix rather than changing
the matrix size after all the coefficients have been
computed and allocated in the computer memory.

9.3 Assembly regarded as stationary of potential

energy

In the last section, we established the global stiffness

matrix equations system by considering the equilibrium
conditions at the joint of the structure. In this section, it
will be shown that the same set of equations can also be
obtained using the stationary principle.

For simplicity, we use the truss example again. Recall

Eqn. 8.3c which express the total potential energy of a bar
member in terms of the element stiffness matrix, k, and
element nodal displacements, d

1
Π P = d Tkd - d Tf (8.3c)
2

Obviously, the total strain energy of the ith element, Ui, is

given by
1
U i = diTk idi (9.9)
2

L9-14
 CV7001_L9.docx 24/06/2009

where suffix i denote the element number.

As strain energy is a scalar, the total strain energy of the

structure can be obtained by summing up (this time is true
summation of scalar!) the contribution from individual
elements as

NE 1
U system = ∑ diTk id i (9.10)
i =1 2

Now, we known that ki is a 4×4 matrix and di is a 4×1

vector (Eqn. 9.3a) and Eqn. 9.10 will remind true if we
expand ki to become a 10×10 matrix and replace di by D
(the 10×1 vector shown in Eqn. 9.6b) such that, for
element 1

⎡11x1x 11x1y 0 0 11x3x 11x3y 0 0 0 0⎤

⎢1 11y1y 0 0 11y3x 11y3y 0 0 0 0⎥
⎢ 1y1x ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥
⎢ ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥
⎢13x1x 13x1y 0 0 13x3x 13x3y 0 0 0 0⎥
k1 = ⎢ ⎥
⎢13y1x 13y1y 0 0 13y3x 13y3y 0 0 0 0⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥
⎢ ⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥
⎢ 0 0 0 0 0 0 0 0 0 0⎥
⎢ ⎥
⎣⎢ 0 0 0 0 0 0 0 0 0 0⎦⎥
(9.11a)

L9-15
 CV7001_L9.docx 24/06/2009

and

D = [u1 v5 ] (9.11b)
T
v1 u 2 v2 u3 v3 u4 v4 u5

By doing so for all the ki, Eqn. 9.10 can be written as

1 T⎛ ⎞
NE
1 T
U system = 2
D ⎜ ∑ i k ⎟ D =
2
D KD (9.12)
⎝ i =1 ⎠

where K, of course, will exactly equal to the one shown in

Eqn. 9.6.

Eventually, the total potential energy of the system can be

expressed as

1
Π system = DT KD - DT F (9.13)
2

∂Π
and upon setting = 0 , we obtain, again
∂D

KD=F (9.14)

Remarks
(1) The exact forms of K, D and F will depend on the
global numbering of the nodes (e.g. compare Eqn.
9.6a with Eqn. 9.6b). However, only a unique set of
solution will be obtained as we know that the system
can has only one set of solution. Different numbering
schemes will simply lead to different permutations of
the unknowns in the vector D.

L9-16
 CV7001_L9.docx 24/06/2009

(2) The reactions R1x, R1y …. etc. do not appear in the

expression of the total potential energy (why?).

(3) After D is found, we can always use Eqns. 9.9 and

9.12 to compute the strain energy of a given member
and the whole structures respectively.

(4) From Eqn. 9.6, we can see that one simplest way to
compute the support reactions is to compute the
product KD.

(5) In this Lecture, for simplicity, the 2D truss problem is

selected to demonstrate the assembly procedure. The
same assembly procedure can be applied to 3D truss,
2D and 3D frame problems and in fact, almost to all
finite element formulations.

L9-17
 CV7001_L10.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001

FINITE ELEMENT METHODS

Lecture 10

Support settlements and inclined supports,

elements with non-uniform cross sections

10.1 Supports settlements and inclined supports

for truss and frame

In the last two lectures, the general expressions of the

element stiffness matrices for truss and frame problems
are given. In addition, the treatment for simply boundary
conditions with zero displacement at support is also given.
In this lecture, some more details regarding the treatment
of general boundary conditions (non-zero and inclined
supports) which frequently occurred in practice will be
given.

Non-zero boundary conditions

Recall our simple truss example in Lecture 7 (Fig. 7.3),

assume now that for some reasons, the roller support at the
RHS settles for a (small) distance equal to δ3 (downward)
as shown in Fig. 10.1.

L10-1
 CV7001_L10.docx 24/06/2009

4 4 5
y, v

7
1
2 5

1 3
3 2 6
x, u δ3
P
Fig. 10.1 A truss with support settlement

Now the question is how to obtain again the nodal

displacements?

Obviously, if the settlement at the support is small enough,

it will not affect the structural/material properties of the
truss structures (e.g. the members will be equally stiff
without yielding nor damaged). Thus, we can conclude
that the final assembled global stiffness matrix will remind
the same.

In this moment, let we denote the global system (based on

the node numbering of Fig. 10.1) as (c.f. Eqn. 9.6c)

L10-2
 CV7001_L10.docx 24/06/2009

⎡X X X X 0 0 X X 0 0 ⎤ ⎡ u1 ⎤ ⎡ R1x ⎤
⎢X X X X 0 0 X X 0 0 ⎥ ⎢ v1 ⎥ ⎢ R1y ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X X 2x3y X X X X ⎥ ⎢u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥
⎢X X X X X X 2y3y X X X X⎥ ⎢v2 ⎥ ⎢ - P ⎥
⎢ ⎥ ⎢ ⎥
⎢0 0 X X X X 3x3y 0 0 X X⎥ ⎢u 3 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢0 0 X X X X 3y3y 0 0 X X ⎥ ⎢ v 3 ⎥ ⎢R 3y ⎥
⎢X X X X 0 0 X X X X ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X 0 0 X X X X⎥ ⎢v4 ⎥ ⎢ 0 ⎥
⎢0 0 X X X X 5x3y X X X X⎥ ⎢u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 X X X X 5y3y X X X X ⎥⎦ ⎣⎢ v5 ⎦⎥ ⎢⎣ 0 ⎥⎦
(10.1a)

Note that the reactions at support 1 and 3 will now assume

different values as compared with Eqn. 9.6c. In additions,
we also label the entries located at the column
corresponding to v3.

Obviously, as we know that v3=-δ3 and we can substitute

v3=-δ3 in Eqn. 10.1a and thus gives

L10-3
 CV7001_L10.docx 24/06/2009

⎡X X X X 0 0 X X 0 0 ⎤ ⎡ u1 ⎤ ⎡ R 1x ⎤
⎢X X X X 0 0 X X 0 0 ⎥ ⎢ v1 ⎥ ⎢ R 1y ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X X 2x3y X X X X⎥ ⎢ u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥
⎢X X X X X X 2y3y X X X X⎥ ⎢ v2 ⎥ ⎢ - P ⎥
⎢ ⎥ ⎢ ⎥
⎢0 0 X X X X 3x3y 0 0 X X⎥ ⎢ u 3 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢0 0 X X X X 3y3y 0 0 X X ⎥ ⎢- δ3 ⎥ ⎢R 3y ⎥
⎢X X X X 0 0 X X X X⎥ ⎢ u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X 0 0 X X X X⎥ ⎢ v4 ⎥ ⎢ 0 ⎥
⎢0 0 X X X X 5x3y X X X X⎥ ⎢ u 5 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 X X X X 5y3y X X X X ⎥⎦ ⎣⎢ v 5 ⎦⎥ ⎢⎣ 0 ⎥⎦
(10.1b)

Eqn. 10.1b is clearly very clumsy to use (in computer

implementation) as the displacement vector D now
contains both unknowns and known values. Thus, we need
to modify the global stiffness matrix in a way similar to
Eqn. 9.8a. However, one should remember now that δ3≠0.
Therefore, we cannot just simply set all the off diagonal
entries of the sixth column to zero. In fact, we have to

(i) multiply these entries (which some of them may

equal to zero) with the settlement value (-δ3) and
move the products to the RHS of the equations
system.

(ii) set the diagonal term to 1.0 and set the corresponding
RHS equal to -δ3 such that after the solution of the
system, we can obtain v3=-δ3.

L10-4
 CV7001_L10.docx 24/06/2009

By doing so, the global equations system will assume the

form shown in Eqn. 10.1c.

Note that the treatment shown in Eqn. 9.8 is just a special

case of the above treatment with δ3=0.

⎡X X X X 0 0 X X 0 0 ⎤ ⎡ u1 ⎤ ⎡ R 1x ⎤
⎢ ⎥
⎢X X X X 0 0 X X 0 0 ⎥ ⎢ v1 ⎥ ⎢ R1y
⎥
⎢ ⎥⎢ ⎥
⎢X X X X X 0 X X X X ⎥ ⎢u 2 ⎥ ⎢ - X 2y3y (-δ3 ) ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X X 0 X X X X ⎥ ⎢ v 2 ⎥ ⎢- P - X 2y3y (-δ3 )⎥
⎢0 0 X X X 0 0 0 X X ⎥ ⎢ u 3 ⎥ ⎢ - X 3x3y (-δ3 ) ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢0 0 0 0 0 1 0 0 0 0 ⎥ ⎢ v3 ⎥ ⎢ - δ3 ⎥
⎢X X X X 0 0 X X X X ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢X X X X 0 0 X X X X⎥ ⎢v4 ⎥ ⎢ 0 ⎥
⎢0 0 X X X 0 X X X X ⎥ ⎢ u 5 ⎥ ⎢ - X 5x3y (-δ3 ) ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 X X X 0 X X X X ⎥⎦ ⎢⎣ v5 ⎥⎦ ⎢⎣ - X 5y3y (-δ 3 ) ⎥⎦

(10.1c)

Finally, we can treat the support at node 1 as usual since

u1=v1=0 and the final system is shown in Eqn. 10.1d.

L10-5
 CV7001_L10.docx 24/06/2009

⎡1 0 0 0 0 0 0 0 0 0 ⎤ ⎡ u1 ⎤ ⎡ R 1x ⎤
⎢ ⎥
⎢0 1 0 0 0 0 0 0 0 0 ⎥ ⎢ v1 ⎥ ⎢ R 1y
⎥
⎢ ⎥⎢ ⎥
⎢0 0 X X X 0 X X X X ⎥ ⎢u 2 ⎥ ⎢ X 2y3yδ3 ⎥
⎢ ⎥ ⎢ ⎥
⎢0
⎢ 0 X X X 0 X X X X ⎥⎥ ⎢ v 2 ⎥ ⎢- P + X 2y3yδ3 ⎥
⎢0 0 X X X 0 0 0 X X ⎥ ⎢ u 3 ⎥ ⎢ X 3x3y δ3 ⎥
⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢0 0 0 0 0 1 0 0 0 0 ⎥ ⎢ v3 ⎥ ⎢ - δ3 ⎥
⎢0 0 X X 0 0 X X X X ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 X X 0 0 X X X X⎥ ⎢v4 ⎥ ⎢ 0 ⎥
⎢0 0 X X X 0 X X X X ⎥ ⎢ u 5 ⎥ ⎢ X 5x3y δ3 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣⎢0 0 X X X 0 X X X X ⎦⎥ ⎣⎢ v 5 ⎦⎥ ⎢⎣ X 5y3yδ3 ⎥⎦

(10.1d)

We can see that the treatments for boundary conditions are

applied at equation level after the assembly of the global
stiffness matrix. Hence, it can be applied to many problem
types (truss, frame and later, continuum problems) in
general.

Inclined supports

Now consider the same truss problem as shown in Fig.

10.2 with the roller support rotated in such a way is it no
longer parallel to the global axes (Fig. 10.2).

L10-6
 CV7001_L10.docx 24/06/2009

4 4 5
y, v
y', v'
7
y', v' 1
2 5

1 3

3 2 6
θ x', u'
x, u
x', u' P

Fig. 10.2. A truss with inclined supported.

From Fig. 10.2, the support boundary condition at node 3

will become

v3=-u3tanθ (10.2)

where θ is the angle between the global x axis and the

local x' axis. Obviously, it is inconvenient to impose Eqn.
10.2 as the boundary condition to the global stiffness
system. However, if the local x' and y' axes are employed
to express the constraint boundary condition, one will has
the simple form that

v'=0 (10.3)

where (x',y') is the coordinate system defined by support

orientation at node 3 and is neither the local coordinate
axis of elements 6 nor 7.

The relationship between these pair of coordinate systems

is defined as (c.f. Eqn. 8.1a)

L10-7
 CV7001_L10.docx 24/06/2009

⎡u ⎤ ⎡ cosθ sinθ ⎤ ⎡u ′⎤
⎢ v ⎥ = ⎢- sinθ cosθ⎥ ⎢ v′⎥ (10.4)
⎣ ⎦ ⎣ ⎦⎣ ⎦

Thus, the simplest way to treat the inclined support

condition is to first transform the element stiffness matrix
so that it will be valid for the local displacement u' and v'.

Now consider element 6 which is connected to node 3,

using the notations used in Lecture 9 (Eqn. 9.3), the
element stiffness equation can be expressed as

⎡62x2x 62x2y 62x3x 62x3y ⎤ ⎡u 2 ⎤ ⎡f62x ⎤

⎢6 62y2y 62y3x 62y3y ⎥ ⎢ v 2 ⎥ ⎢f62y ⎥
⎢ 2y2x ⎥⎢ ⎥ = ⎢ ⎥ (10.5)
⎢63x2x 63x2y 63x3x 63x3y ⎥ ⎢ u 3 ⎥ ⎢f63x ⎥
⎢ ⎥ ⎢ ⎥
⎣63y2x 63y2y 63y3x 63y3y ⎦ ⎢⎣ v3 ⎥⎦ ⎣f63y ⎦

or kd=f (10.6)

Now consider the relation between the two vectors

d=[u2,v2,u3,v3]T and d'=[u2,v2,u'3,v'3]T, from Eqn. 10.4

⎡ u 2 ⎤ ⎡1 0 0 0 ⎤⎡u2 ⎤
⎢ v ⎥ ⎢0 1 0 0 ⎥ ⎢ v2 ⎥
⎢ 2⎥ = ⎢ ⎥ ⎢ ⎥ or d = T' d' (10.7a)
⎢ u 3 ⎥ ⎢0 0 sinθ ⎥ ⎢u'3 ⎥
cosθ
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎣ v 3 ⎦ ⎣0 0 - sinθ cosθ⎦ ⎣ v'3 ⎦

and similarly f=T'f' (10.7b)

where f=[f62x,f62y,f63x,f63y]T and f'=[f62x,f62y,f'63x,f'63y]T (10.7c)

L10-8
 CV7001_L10.docx 24/06/2009

Substitute Eqn. 10.7 into Eqn. 10.6, we have

kT'd'=T'f' (10.8a)

Note that (T')-1=(T')T and therefore,

(T')TkT'd'=f' (10.8b)

Hence, using the transformation matrix T', we can

transform the element stiffness matrix and load vector to
the support local coordinate system. Note that after the
transformation, the element stiffness matrix will use d' as
unknowns.

We can then using Eqn. 10.8b to transform all elements

that are connected to support 3 (i.e. elements 6 and 7) so
that the finial global equations system will only contain u'3
and v'3 but not u3 and v3.

(Why we only need to transform those elements that are

connected to support node 3 only?)

For example, for the system shown in Fig. 10.2, the global
unknown vector is given by

D=[u1,v1, u2,v2, u'3,v'3, u4,v4, u5,v5]T (10.9)

Thus, we can apply the simple boundary condition v'3=0 to

the global system. After D is obtained, the value of u3 and
v3 can be obtained by using Eqn. 10.4.
Clearly, the above procedure can be applied to all the
problem types discussed (2D, 3D truss and frame) so far.

L10-9
 CV7001_L10.docx 24/06/2009

10.2 Elements with non-uniform cross section

So far we only discuss cases in which the member cross

sections are uniform (i.e. prismatic members). However,
in real structures, members with non-uniform cross
sections are frequently encountered. By using the principle
of stationary of potential energy, it is found that the
problem of elements with non-uniform cross sections and
non-uniform material properties can be easily solved.

Consider the expression of element stiffness matrix (w.r.t.

the element local coordinate system) of a 2D truss,
recalling Eqn. 7.15b it is given by

[ L
][
k'= ∫0 (B'T EB' )Adx' = ∫Ω (B'T EB' )dΩ ] (7.15b)

Note that in Eqn. 7.15b, the B' matrix is independent of

the cross section and the material properties of the
member. In case that the cross sectional area, A, and the
Young's modulus vary along the member, the stiffness
matrix expression can be obtained by slightly modifying
Eqn. 7.15b to

L
k'= ∫0 (B'T E(x' )B' )A(x' )dx' (10.10)

That is, provide that we know the variation of E(x') and

A(x') along the member, the local element stiffness matrix
expression can be obtained by integrating Eqn. 10.10. In
the case that E(x') and A(x') are "simple" functions (e.g.
polynomials) then Eqn. 10.10 can be integrated in exactly.

L10-10
 CV7001_L10.docx 24/06/2009

However, in some cases, the exact form of E(x') and A(x')

may not be exactly known, in such a situation, we still can
integrate Eqn. 10.10 approximately to obtain some useful
results.

A simple case of a 2D truss with non-uniform cross

sectional area will be used as an example to demonstrate
the basic idea.

Consider a bar element in which its width, b, is constant

while the depth of the bar is changing linearly from h1 at
one end to h2 at the order end (Fig. 10.3). In addition, we
also assume that the Young's modulus of the bar is
constant and equal to E.

From Lecture 7, we know that the B' matrix is given by

(Eqn. 7.9b)

B'=[-1/L, 1/L] (10.11)

b
h1
h2 x'
L

Fig. 10.3 A bar with non-uniform cross sectional area

For the cross sectional area, A(x'), of the member, if we

use again the parent coordinate ξ=x'/L (Eqn. 7.8), it can be
expressed as

L10-11
 CV7001_L10.docx 24/06/2009

⎡h ⎤
A(x')=b[(1-ξ)h1+ξh2]=b[N1(ξ), N2(ξ)] ⎢ 1 ⎥ (10.12a)
⎣h 2 ⎦
or

⎡h ⎤
A(x')=b[(1-x'/L), x'/L] ⎢ 1 ⎥ (10.12b)
⎣h 2 ⎦

Put Eqns. 10.11 and 10.12b into Eqn. 10.10 and since E is
a constant along x', we have

L ⎡- 1/L ⎤
k ' = ∫0 ⎢
1/L ⎥ Eb[- 1/L 1 /L]((1 - x' /L)h1 + (x' /L)h 2 )dx'
⎣ ⎦
Eb ⎡ 1 - 1⎤ L ⎛ (h 2 − h1 )x' ⎞
= 2⎢ ∫ 1
L ⎣- 1 1 ⎥⎦ 0 ⎝
⎜ h +
L
⎟dx'
⎠
(10.13a)
Eb ⎡ 1 - 1⎤ (h 2 - h1 )L
= 2⎢ (h L + )
L ⎣- 1 1 ⎥⎦
1
2
Eb(h 2 + h1 ) ⎡ 1 - 1⎤
= ⎢- 1 1 ⎥
2L ⎣ ⎦

Similar procedure, of course, can be used to deal with the

case when E is a function of x'. After, k' is computed, k
can be obtained by coordinate transformation.

As an exercise, obtain the matrix k' the same element

shown in Fig. 10.3 if it is considered as a 2D frame
element with axial force.

L10-12
 CV7001_L10.docx 24/06/2009

Remarks
(1) Note that while the solution shown in Eqn. 10.13a is
obtained from exact integration of Eqn. 10.10, it is not
the exact expression of the element stiffness matrix.
The reason is that in Eqns. 10.10 and 7.15b, it is
assumed that the axial strain of the member is constant
(hence the B' is constant). This assumption is true
only when the cross section of the bar is uniform.

(2) If the cross section of the bar is not uniform, the axial
strain of the member will not be uniform and thus B'
should not be constant along the bar. However, it can
be shown that the error due to this assumption is
normally quite small. For the case of h2=0.5h1, the
error is about 4%. However, greater error will occur if
say, h2 << h1.

Questions:
(1) For the case of the 2D frame elements, the axial force
terms will obviously have the same error. What about
the terms corresponding to the bending moment and
shear forces? Will they again be only an
approximated solution?

(2) If you encounter a case that for a frame member, h2

<< h1, (say h2=0.1h1) how can you reduce the error
but keep using Eqn. 10.13a?

L10-13
 CV7001_L11.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS
Lecture 11

Initial and thermal strain effects,

releases of end moments

11.1 Initial and thermal strain effects on bar and

frame

By using the principle of stationary potential energy, we

can also easily handle with problems with initial strain due
to different causes (temperature (thermal) effects, lack-of-
fit etc). All we need to do is to obtain the expression of the
initial strain components, ε0, induced and include its effect
in the strain energy term U of the total potential energy
expression. In this section, the general case will be
introduced first and then it will be followed by the
particular cases of bar and frame problems.

Initial strain effect in general situation

Consider the case that for some reasons, initial strain is

induced in the material of the structures without causing
any initial stress. In such a situation, the stress-strain
relationship (law) of the material will assume a form
shown in Fig. 11.1

L11-1
 CV7001_L11.docx 24/06/2009

From Fig. 11.1, it can be seen that with the presence of

initial strain, the stress-strain relationship (Hooke's law)
should be modified to

σ=E(ε-ε0) (11.1)

σ=E(ε-ε0)

ε
ε0

Fig. 11.1 Stress-strain relationship with initial strain

Thus, in order to include the effect of initial strain, all we

need to use is to employ the new stress-strain relationship
stated in Eqn. 11.11 in the total potential energy
expression (Eqn. 5.5). From Fig. 11.1, the strain energy
per unit volume will be given by

1
(ε − ε 0 )T E(ε − ε 0 ) (11.12)
2

After using Eqn. 11.2, Eqn. 5.5 will become

1
Πp= ∫ (ε − ε 0 )T E(ε − ε 0 ))dΩ − ∫ u T FbdΩ − ∫ u T ΦdΓ (11.3a)
Ω2 Ω ΓΦ

L11-2
 CV7001_L11.docx 24/06/2009

After simplification, we have

1 T
Πp= ∫ (ε T Eε + ε 0 Eε 0 ) − ε T Eε 0dΩ
Ω2
− ∫ u T FbdΩ − ∫ u T ΦdΓ
Ω ΓΦ
(11.3b)

T
The term ε 0 Eε 0 can be consider as the strain energy
stored up during the initial strain formulation and thus it
can be dropped from Eqn. 11.3b without affecting the total
potential energy expression since we only consider the
energy balance after the initial strain occurred. Thus, Eqn.
11.3b can be written as

1
Πp= ∫ (ε T Eε ) − ε T Eε 0dΩ − ∫ u T FbdΩ − ∫ u T ΦdΓ (11.3c)
Ω2 Ω ΓΦ

Eqn. 11.3c is now the starting point to obtain the stiffness

equations with initial strain effect.

In the cases of bar and frame elements, we know that we

can always
(1) approximate the displacement u in terms of some
shape functions and the nodal displacement d (Eqns.
7.7 and 8.17d)

u=Nd (11.4a)

L11-3
 CV7001_L11.docx 24/06/2009

(2) relate the strain ε and nodal displacements of the

element d by a strain-displacement matrix, B (Eqn.
7.9b and 8.18). That is

ε=Bd (11.4b)

Thus, after we applied Eqn. 11.4 to Eqn. 11.3b,

1 T T T T T
Πp = ∫2 d ( B EB ) d + ε 0 E ε 0 ) − d B E ε 0 dΩ
Ω
(11.5)
T T T T
− ∫ d N FbdΩ − ∫ d N ΦdΓ
Ω ΓΦ

∂Π P
Upon using the condition =0
∂d

⎡ T ⎤ T T T
⎢∫ B EB d Ω ⎥d = ∫ B Eε 0dΩ + ∫ N FbdΩ + ∫ N ΦdΓ (11.6)
⎣Ω ⎦ Ω Ω ΓΦ

Thus, we can see that the initial strain will not affect the
expression of the element stiffness matrix. It will,
however, lead to an additional loading term (first term on
the RHS of Eqn. 11.6) for the element load vector. We
also see that provide that we can obtain the expression for
the initial strain ε0, we can obtain the corresponding load
vector, Fε 0 through the expression

Fε 0 = ∫ B T Eε 0dΩ (11.7)
Ω

L11-4
 CV7001_L11.docx 24/06/2009

Initial strain due to thermal effect

Bar element

For the one-dimensional bar element discussed in Lecture

7, if the bar is subject to an average temperature change of
ΔT degree and the coefficient of thermal expansion is
equal to α, the change in length of the bar is equal to

αΔTL (11.8a)

Hence, the average strain of the bar element is equal to

αΔT (11.8b)

Thus, the additional term due to initial thermal strain is

given by

L ⎡- 1/L ⎤ ⎡- 1⎤
Fε 0 = ∫ B T EαΔTdΩ = ∫0 ⎢ ⎥ Eα ΔTAdx' = Eα ΔTA ⎢ 1 ⎥ (11.9)
Ω ⎣ 1/L ⎦ ⎣ ⎦

Note that
(1) ΔT is positive for an increase in temperature.
(2) We assumed that the initial strain is constant within
an element, in case that there is a temperature
gradient along the bar, the above expression will not
be valid. (Then, how to handle the case that there is a
linear variation in temperature along the bar?)

L11-5
 CV7001_L11.docx 24/06/2009

Beam element

For the case of a beam element, a common cause of

thermal strain effect will be due to a thermal gradient
through the depth of beam as shown in Fig. 11.2. As
shown in Fig. 11.2, the depth of the beam element is equal
to L and it is subjected to a linear temperature gradient of
ΔT=Tl-Tu along the whole length of the element. (Note
that, in this case, a positive ΔT refers to the case that
Tl>Tu).

Tu

Length=L, EI, α constant

Tl

Fig. 11.2 Thermal gradient through a beam

In such a case, we can expect that the thermal gradient will

lead to a constant curvature of the beam and it can be
proved that the curvature is given by

d 2 w' αΔT
ε0 = 2 = (11.10)
dx h

Hence, the addition load vector is given by

L11-6
 CV7001_L11.docx 24/06/2009

⎡ - 6 12x' ⎤
⎢ L2 + L3 ⎥
⎢ - 4 6x' ⎥ ⎡0⎤
⎢ + 2 ⎥ ⎢- 1⎥
αΔT L L L ⎥ EI α Δ T EIα ΔT ⎢ ⎥
T
Fε 0 = ∫ B EI dΩ = ∫0 ⎢ dx' =
Ω h ⎢ 6 − 12x' ⎥ h h ⎢0⎥
⎢ L2 L3 ⎥ ⎢ ⎥
⎢ - 2 6x' ⎥ ⎣1⎦
⎢ + 2 ⎥
⎣ L L ⎦

(11.11)

Note that
(1) The temperature gradient does not induce any
additional loading term corresponding to the end
shear force.

(Why? How can you relate this fact to the physical

behaviour of the beam under the thermal gradient?)

(2) We have assumed that EI, α, and the temperature

gradient is uniform along the length of the beam.
Hence, the initial curvature (thus the initial strain) is
again constant along the beam.

L11-7
 CV7001_L11.docx 24/06/2009

11.2 Releases of end moments

Simple end release without rearrangement of stiffness

matrix

In order to demonstrate the treatment for end releases,

examples involve 2D frame elements with 3 DOFs (u, v,
θ) at each end will be used. Treatment for the case of 3D
frame can be implemented in a similar way.

Consider a 2D frame element AB as shown in Fig. 11.3

Suppose that the moment at end B is released by adding a
hinge there. It is now required to obtain the modified form
of the stiffness matrix and load vector for this element.

B A

Fig. 11.3 Release of moment at end A for a member AB

To start with, let k and f be the original element stiffness

matrix and RHS load vector for element AB at the element
level, before the moment at end B is released, we have

kd = f (11.12a)

where k is a 6×6 symmetric matrix

L11-8
 CV7001_L11.docx 24/06/2009

⎡k u A u A k u A vA k u AθA kuAuB k u A vB k u AθB ⎤

⎢k k vA vA k vAθA k vAu B k vA vB k vAθB ⎥
⎢ u A vA ⎥
⎢k u θ k vAθ A kθ Aθ A kθAu B kθA vB kθAθB ⎥
k=⎢ A A ⎥ (11.12b)
⎢k u A u B k vAu B kθAu B k u Bu B k u BvB k u Bθ B ⎥
⎢k u A vB k vA vB kθA vB k uBvB k vBvB k v Bθ B ⎥
⎢ ⎥
⎢⎣ k u A θ B k vAθB kθ Aθ B k u Bθ B k v Bθ B k θ Bθ B ⎥⎦

The corresponding displacement, d and RHS load vector f

are given by

⎡u A ⎤ ⎡ FA ⎤
⎢v ⎥ ⎢S ⎥
⎢ A⎥ ⎢ A⎥
⎢θ ⎥ ⎢M ⎥
d = ⎢ A ⎥ and f = ⎢ A ⎥ (11.12c)
⎢u B ⎥ ⎢ FB ⎥
⎢ vB ⎥ ⎢ SB ⎥
⎢ ⎥ ⎢ ⎥
⎣ θB ⎦ ⎣MB ⎦

In Eqn. 11.12c, FA, SA and MA are the axial force, shear

force and moment at end A (before assembly or the fixed
end forces) respectively and so on. Note that, in here, it is
assumed that all components of f are nonzero.

The effect of the hinge release effectively set the final

moment at end B to zero. (However, please note that in
general, MB≠0 and it is obvious that θB≠0.) From Eqns.
11.12a and 11.12b, one can obtain θB by using the
condition that

L11-9
 CV7001_L11.docx 24/06/2009

(k u θ )u A + (k v θ )vA + (k θ θ )θA +
A B A B A B
(Eqn. 11.13a)
(k u θ )u B + (k v θ )vB + (k θ θ )θB = M B
B B B B B B

or

⎡(k u A θ B )u A + (k v A θ B )v A + (k θ A θ B )θ A ⎤
θB = −⎢ ⎥ / (k θ Bθ B )
⎢⎣ + (k u Bθ B )u B + (k v Bθ B )v B - M B ⎥⎦
(Eqn. 11.13b)

By using Eqn. 11.13b, we could eliminate θB from Eqn.

11.12 and then redefine a 5×5 matrix for the element.
However, using Eqn. 11.13 for direct elimination is not
convenient in computer implementation, especially when
more than one DOF is released at one end (e.g. releases of
both MB and SB at end B). In practice, the elimination
procedure can be carried out more efficiently using
the matrix condensation technique given below.

Reconsider Eqns. 11.12b and 11.12c again, k can be

partitioned into four sub-matrices while d and f can be
considered as consisting of two components such that

⎡ k k ar ⎤ ⎡d a ⎤ ⎡fa ⎤
k = ⎢ aa T , d = and f = (11.14a)
⎣(k ar ) k rr ⎥⎦ ⎢d ⎥
⎣ r⎦
⎢f ⎥
⎣ r⎦

where

L11-10
 CV7001_L11.docx 24/06/2009

⎡k u A u A k u A vA k u AθA kuAuB k u A vB ⎤
⎢k k vA vA k vAθA k vAu B k vA vB ⎥
⎢ u A vA ⎥
k aa = ⎢k u AθA k vAθA kθAθA kθAu B kθA vB ⎥
⎢ ⎥
⎢k u A u B k vA u B kθAu B k u Bu B k u BvB ⎥
⎢⎣ k u v k vA vB kθA vB k u BvB k v B v B ⎥⎦
A B

⎡k u A θ B ⎤
⎢k ⎥
⎢ A B⎥
v θ
k ar = ⎢ k θ A θ B ⎥ and k rr = k θ B θ B [ ]
⎢ ⎥
k
⎢ u BθB ⎥
⎢⎣ k v θ ⎥⎦
B B

(11.14b)

and

⎡u A ⎤ ⎡ FA ⎤
⎢v ⎥ ⎢S ⎥
⎢ ⎥ A ⎢ A⎥
d a = ⎢ θA ⎥, d r = [θB ], fa = ⎢M A ⎥ and f r = [M B ] (11.14c)
⎢ ⎥ ⎢ ⎥
u
⎢ B⎥ F
⎢ B⎥
⎢⎣ v B ⎥⎦ ⎣⎢ SB ⎥⎦

By using Eqn. 11.14, Eqn. 11.12 can be rewritten as

⎡ k aa k ar ⎤ ⎡d a ⎤ ⎡fa ⎤
⎢(k )T ⎥ ⎢ ⎥ =⎢ ⎥ (11.15a)
⎣ ar k rr ⎦ ⎣ d r ⎦ ⎣ f r ⎦

and can be arranged as

L11-11
 CV7001_L11.docx 24/06/2009

k aa d a + k ard r = fa
(11.15b)
(k ar ) T
d a + k rrd r = f r

The second equation in Eqn. 11.15b gives

-1
(
d r = (k rr ) f r - (k ar ) d a
T
) (11.15c)

Back substitute Eqn. 11.15c into the first equation of Eqn.

11.15b will eliminate out dr such that

-1
(
k aa d a + k ar (d rr ) f r - (k ar ) d a = fa
T
) (11.15d)

and after re-arrangement gives the final form

(k aa
-1 T
)
− k ar (k rr ) (k ar ) d a = fa - k ar (k rr ) f r (11.15e)
-1

Eqn. 11.15e can be written as

~ ~
kd a = f
~
(
k = k aa − k ar (k rr ) (k ar )
-1 T
) (11.15f)
~
f = fa - k ar (k rr ) f r
-1

~ ~
In Eqn. 11.15f, k and f are the modified stiffness matrix
and RHS loading vector corresponding to the release
element and could be assembled into the global stiffness
matrix and load vector.
~ ~
From Eqn. 11.15f, it could be seen that k and f are a 5×5
matrix and a 5×1 vector respectively. In practice, for easier

L11-12
 CV7001_L11.docx 24/06/2009

implementation, the condensed matrix may be “expanded”

back to become a 6×6 matrix and additional zero entries
are added at the last row/columns. That is, if after
~ ~
condensation, k and f are given by
~ ~ ~ ~ ~ ~
⎡ku A u A ku A vA ku Aθ A ku A u B ku A vB ⎤ ⎡ FA ⎤
⎢~ ~ ~ ~ ~ ⎥ ⎢~ ⎥
⎢ku A vA kvA vA k vAθ A kvAu B kvA vB ⎥ ⎢ SA ⎥
~ ~ ~ ~ ~ ~ ~ ~ ⎥
k = ⎢ ku Aθ A k vAθ A kθ A θ A kθ A u B kθ A v B ⎥, f = ⎢M A
⎢~ ~ ~ ~ ~ ⎥ ⎢~ ⎥
⎢ ku A u B kvAu B kθ A u B ku Bu B ku BvB ⎥ ⎢ FB ⎥
⎢~ ~ ~ ~ ~
k v B v B ⎥⎦
~ ⎥
⎢⎣ S
⎣ ku A vB kvA vB kθ A v B ku BvB B ⎦
(11.16a)
~ ~
then, the “expanded” forms of k and f for assembly are
given by
~ ~ ~ ~ ~ ~
⎡ku A u A ku A vA ku Aθ A ku Au B ku A vB 0 ⎤ ⎡ FA ⎤
⎢~ ~ ~ ~ ~ ⎥ ⎢~ ⎥
⎢ku A vA kvA vA k v Aθ A kvAu B kvA vB 0 ⎥ ⎢ SA ⎥
~ ~ ~ ~ ~ ~
~ ⎢ ku Aθ A k v Aθ A kθ A θ A kθ A u B kθ A v B 0 ⎥ ~ ⎢M A ⎥
k = ⎢~ ~ ~ ~ ~ ⎥ and f = ⎢ ~ ⎥
⎢ ku A u B kvAu B kθ A u B ku Bu B kvBu B 0 ⎥ ⎢ FB ⎥
⎢~k
~
kvAvB
~
kθ A v B
~
kvBu B
~
kvBvB 0 ⎥ ⎢S~ ⎥
⎢ u A vB ⎥ ⎢
B
⎥
⎣⎢0 0 0 0 0 0 ⎦⎥ ⎣ 0 ⎦
(Eqn. 11.16b)

Note in Eqns. 11.16a and 11.16b, obviously,

~ ~
k u A u A ≠ k u A u A and FA ≠ FA .

L11-13
 CV7001_L11.docx 24/06/2009

Remarks:
(1) After the global stiffness matrix is solved and the
value of da is known, the value of dr (i.e. θB) for
element AB at end B can be obtained by using Eqn.
11.15c.

(2) From Eqn. 11.15f, one can see that only when fr=0,
~
then f =fa=f.

(3) The above matrix condensation technique can be

formulated (and implemented) as a general procedure
for the case of more than one release at the end such as
the releases of both MB and SB at end B (Fig. 11.4). In
this case, Eqn. 11.14a, and all equations in Eqn. 11.15
will remain valid and it is only required to redefine
Eqns. 11.14b and 11.14c such that

⎡k u A u A ku A vA k u AθA kuAuB ⎤
⎢k k vAvA k v Aθ A k vAu B ⎥
=⎢ A A ⎥
u v
k aa
⎢k u Aθ A k v Aθ A kθ Aθ A kθAu B ⎥
⎢ ⎥
⎣k u A u B k vAu B kθAu B k vBu B ⎦
(11.17a)
⎡k u A v B k u Aθ B ⎤
⎢k k v Aθ B ⎥
⎥ and k = ⎡ v B v B
k k v Bθ B ⎤
k ar = ⎢ A B
v v
⎢kθA vB kθ Aθ B ⎥ rr ⎢k k θ Bθ B ⎥⎦
⎣ v Bθ B
⎢ ⎥
⎣k u B v B k u Bθ B ⎦

and

L11-14
 CV7001_L11.docx 24/06/2009

⎡u A ⎤ ⎡ FA ⎤
⎢v ⎥ ⎢S ⎥
⎡ vB ⎤ ⎡S ⎤
d a = ⎢ ⎥, d r = ⎢ ⎥, fa = ⎢ A ⎥ and f r = ⎢ B ⎥ (11.17b)
A
⎢ θA ⎥ ⎣ θB ⎦ ⎢M A ⎥ ⎣M B ⎦
⎢ ⎥ ⎢ ⎥
⎣uB ⎦ ⎣ FB ⎦

Hence, the matrix condensation technique is very suitable

for a general implementation for the treatment of end
releases.

B A

Fig. 11.4. Releases of MB and SB at end A for a member

AB

End releases with rearrangement of stiffness matrix

Consider that situation that both the moments at ends A

and B of the element are released (Fig. 11.5).

B A

Fig. 11.5. Release of moments at ends A and B for a

member AB

L11-15
 CV7001_L11.docx 24/06/2009

In such case, the condition of MA=MB=0 shall be applied.

However, in order to carry out the matrix condensation
technique, all the released DOFs are required to
be concentrated at the bottom of the stiffness matrix.
Such condition can be achieved by a simple re-
arrangement of the DOFs of the elements. In the case of
Figure 3, one can rearrange the components in u and F
such that (c.f. Eqn. 11.12c)

⎡u A ⎤ ⎡ FA ⎤
⎢v ⎥ ⎢S ⎥
⎢ ⎥ A ⎢ A⎥
⎢u B ⎥ ⎢ FB ⎥
d = ⎢ ⎥ and f = ⎢ ⎥ (Eqn. 11.18a)
v
⎢ ⎥ B ⎢ S B ⎥
⎢θ A ⎥ ⎢M A ⎥
⎢ ⎥ ⎢ ⎥
⎣ θB ⎦ M
⎣ B⎦

In order to reflect such re-arrangement of displacement

and force vector components, the element stiffness matrix
now should be arranged so that (c.f. Eqn. 11.12b)

⎡k u A u A kuAvA kuAuB ku AvB k u Aθ A k u Aθ B ⎤

⎢k k vAvA k vAu B k vA vB k v Aθ A k v Aθ B ⎥
⎢ u AvA ⎥
⎢k u A u B k vAu B k u Bu B k u BvB k u Bθ A k u Bθ B ⎥
k=⎢ ⎥
⎢k u A vB k vA vB k u BvB k vBvB k v Bθ A k v Bθ B ⎥
⎢k u Aθ A k v Aθ A k u Bθ A kθA vB kθ Aθ A kθAθB ⎥
⎢ ⎥
⎣⎢ k u A θ B k vAθB k u Bθ B k v Bθ B kθ Aθ B k θ Bθ B ⎦⎥
(11.18b)

L11-16
 CV7001_L11.docx 24/06/2009

After all the released DOFs are re-arranged at the bottom

of the stiffness matrix, matrix condensation can be carried
out as usual with

⎡k u A u A k u A vA kuAuB k u A vB ⎤
⎢k k vA vA k vAu B k vAvB ⎥
=⎢ A A ⎥
u v
k aa
⎢k u A u B k vAu B k u Bu B k u BvB ⎥
⎢ ⎥
⎣k u A vB k vAvB ku BvB k vBvB ⎦
(11.19a)
⎡k u A θ A k u Aθ B ⎤
⎢k k v Aθ B ⎥
⎥ and k = ⎡ θ A θ A
k k θA θB ⎤
k ar = ⎢ A A
v θ
⎢ k u Bθ A k u Bθ B ⎥ rr ⎢k k θ B θ B ⎥⎦
⎣ θAθB
⎢ ⎥
⎣ k v Bθ A k v Bθ B ⎦

and

⎡u A ⎤ ⎡ FA ⎤
⎢v ⎥ θA ⎤ ⎢S ⎥
⎡ ⎡M ⎤
d a = ⎢ ⎥, d r = ⎢ ⎥, fa = ⎢ A ⎥ and f r = ⎢ A ⎥ (11.19b)
A
⎢u A ⎥ ⎣ θB ⎦ ⎢ FA ⎥ ⎣MB ⎦
⎢ ⎥ ⎢ ⎥
⎣uB ⎦ ⎣ FB ⎦

Finally, after condensation, two rows and columns of zero

vectors will be added to the corresponding rows and
columns for θA and θB of the condensation matrix for the
easy of assembly (c.f. Eqn. 11.16b).

L11-17
 CV7001_L11.docx 24/06/2009

~ ~ ~ ~ ~
⎡ku A u A ku A vA 0 ku A u B ku A vB 0⎤ ⎡ FA ⎤
⎢~ ~ ~ ~ ⎥ ⎢~ ⎥
⎢ku A vA kvA vA 0 kvAu B kvAvB 0⎥ ⎢ SA ⎥
~ ⎢ 0 0 0 0 0 0⎥ ~ ⎢0 ⎥
k = ⎢~ ~ ~ ~ ⎥ and f = ⎢~ ⎥
⎢ ku A u B kvAu B 0 ku Bu B ku BvB 0⎥ ⎢ FB ⎥
⎢~ ~ ~ ~ ~ ⎥
k kvAvB 0 ku B vB kvBvB 0⎥ ⎢S B
⎢ uAuB ⎥ ⎢ ⎥
⎣ 0 0 0 0 0 0⎦ ⎣ 0⎦
(11.19c)

Remarks:
Eqn. 11.18b can be considered as the results of the
following steps of transformations which “push” the
entries corresponding to the DOF θA towards the bottom
of the matrix and will be very suitable for implementation
in general case.
Step 1, Starting from matrix shown in Eqn. 11.18a, swap
column 3 with column 4:

⎡k u A u A k u A vA k u AθA k uAuB k u A vB k u Aθ B ⎤ ⎡ k u A u A k u A vA k u AuB k u Aθ A k u A vB k u Aθ B ⎤

⎢ ⎥ ⎢ ⎥
⎢k u A vA k vA vA k vAθA k vA u B k vA vB k v Aθ B ⎥ ⎢ k u A v A k vA vA k vA u B k v Aθ A k vA vB k v Aθ B ⎥
⎢k u θ k vAθA k θAθA k θA u B k θA vB k θ Aθ B ⎥ ⎢ k u Aθ A k v Aθ A k θAu B k θ Aθ A k θA vB k θ Aθ B ⎥
⎢ AA ⎥→⎢ ⎥ (11.20a)
⎢k uAuB k vA u B k θA u B k u Bu B k u BvB k u BθB ⎥ ⎢ k u A u B k vA u B k u Bu B k θAu B k u BvB k u Bθ B ⎥
⎢k k vA vB k θA vB k u BvB k vBvB k vBθB ⎥ ⎢ k u A vB k vA vB k u BvB k θA vB k v BvB k vBθ B ⎥
⎢ u A vB ⎥ ⎢ ⎥
⎢⎣ k u AθB k v Aθ B k θ Aθ B k u BθB k vBθB k θBθB ⎥⎦ ⎢⎣ k u AθB k v Aθ B k u Bθ B k θ Aθ B k vBθ B k θBθ B ⎥⎦

Step 2, swap row 3 with row 4:

⎡k u A u A k u A vA k u Au B k u Aθ A k u A vB k u Aθ B ⎤ ⎡ k u A u A k u A vA k u AuB k u Aθ A k u A vB k u Aθ B ⎤
⎢ ⎥ ⎢ ⎥
⎢k u A vA k vA vA k vA u B k v Aθ A k vA vB k v Aθ B ⎥ ⎢ k u A v A k vA vA k vAu B k v Aθ A k vA vB k v Aθ B ⎥
⎢k u θ k vAθ A k θA u B k θ Aθ A k θA vB k θ Aθ B ⎥ ⎢ k u A u B k vA u B k u Bu B k u Bθ A k u BvB k u Bθ B ⎥
⎢ AA ⎥→⎢ ⎥ (11.20b)
⎢k uAu B k vA u B k u Bu B k θA u B k u BvB k u BθB ⎥ ⎢ k u Aθ A k v Aθ A k u BθA k θ Aθ A k θA vB k θ Aθ B ⎥
⎢k k vA vB k u BvB k θA vB k vBvB k vBθB ⎥ ⎢ k u A vB k vA vB k u BvB k θA vB k v BvB k vBθ B ⎥
⎢ u A vB ⎥ ⎢ ⎥
⎢⎣ k u AθB k v Aθ B k u BθB k θ Aθ B k vBθB k θBθB ⎥⎦ ⎢⎣ k u AθB k v Aθ B k u Bθ B k θ Aθ B k vBθ B k θBθ B ⎥⎦

L11-18
 CV7001_L11.docx 24/06/2009

Step 3, swap column 4 with column 5:

⎡k u A u A k u A vA k uAuB k u Aθ A k u A vB k u Aθ B ⎤ ⎡ k u A u A k u A vA k u AuB k u A vB k u Aθ A k u Aθ B ⎤
⎢ ⎥ ⎢ ⎥
⎢k u A vA k vA vA k vA u B k v Aθ A k vA vB k v Aθ B ⎥ ⎢ k u A v A k vA vA k vAu B k vA vB k v Aθ A k v Aθ B ⎥
⎢k u u k vA u B k u Bu B k u BθA k u BvB k u BθB ⎥ ⎢ k u A u B k vA u B k u Bu B k u BvB k u Bθ A k u Bθ B ⎥
⎢ A B ⎥→⎢ ⎥ (11.20c)
⎢ k u Aθ A k vAθ A k u BθA k θ Aθ A k θA vB k θ Aθ B ⎥ ⎢ k u A θ A k v Aθ A k u Bθ A k θA vB k θ Aθ A k θ Aθ B ⎥
⎢k k vA vB k u BvB k θA vB k vBvB k vBθ B ⎥ ⎢ k u A vB k vA vB k u BvB k vBvB k θA vB k vBθ B ⎥
⎢ u A vB ⎥ ⎢ ⎥
⎢⎣ k u Aθ B k v Aθ B k u Bθ B k θ Aθ B k vBθ B k θ Bθ B ⎥⎦ ⎢⎣ k u AθB k v Aθ B k u Bθ B k vBθ B k θ Aθ B k θBθ B ⎥⎦

Step 5, swap row 4 with row 5

⎡k u A u A k u A vA k uAuB k u A vB k u AθA k u Aθ B ⎤ ⎡ k u A u A k u A vA k u AuB k u A vB k u Aθ A k u Aθ B ⎤
⎢ ⎥ ⎢ ⎥
⎢k u A vA k vA vA k vA u B k vA vB k vAθA k v Aθ B ⎥ ⎢ k u A v A k vA vA k vA u B k vA vB k v Aθ A k v Aθ B ⎥
⎢k u u k vA u B k u Bu B k u BvB k u Bθ A k u BθB ⎥ ⎢ k u A u B k vA u B k u Bu B k u BvB k u Bθ A k u Bθ B ⎥
⎢ A B ⎥→⎢ ⎥ (11.20d)
⎢ k u Aθ A k vAθA k u Bθ A k θA vB k θ Aθ A k θ Aθ B ⎥ ⎢ k u A v B k vA vB k u BvB k vBvB k vBθ A k vBθ B ⎥
⎢k k vA vB k u BvB k vBvB k θA vB k vBθB ⎥ ⎢ k u Aθ A k v Aθ A k u Bθ A k vBθA k θ Aθ A k θ Aθ B ⎥
⎢ u A vB ⎥ ⎢ ⎥
⎢⎣ k u AθB k v Aθ B k u BθB k vBθB k θ Aθ B k θBθB ⎥⎦ ⎢⎣ k u AθB k v Aθ B k u Bθ B k vBθ B k θ Aθ B k θBθ B ⎥⎦

Summary for actual implementation

From the above discussions, it can be seen that in actual

implementation, the following steps could be used:

(i) Compute the “original” element stiffness matrix and

the RHS element load vector as usual.

(ii) Identify the DOFs corresponding to the releases at the

ends of the element.

(iii) Re-arrange the entries of the element stiffness matrix

and RHS load vectors (Eqns. 11.18 – 11.20)
according to the DOFs released.

(iv) Condense the element stiffness matrix and the RHS

vector (Eqn. 11.15).

L11-19
 CV7001_L11.docx 24/06/2009

(v) Expand the condensed element stiffness matrix and

assemble it to the global stiffness matrix.

(vi) Solve the global system of equations as usual.

(vii) Resolve the condensed DOFs from the solution using

Eqn. 11.15c.

L11-20
 CV7001_L12.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 12

Model problem for linear solid mechanics

12.1 The model problem for linear solid

mechanics for 2D and 3D elasticity problems

After focused on truss and frame problems and their

solutions using (i) the minimization of potential energy,
and then (ii) the Ritz method, in the next few lectures, we
will turn our attention to the more general 2D/3D elasticity
problems which, in fact, embrace, near all static
structural analysis problems. We will see very soon that
the formulations and natures of these (2D/3D elasticity
problems) in fact is very similar to the truss and frame
problems. That is, we can again obtain an approximated
solution by using (i) the minimization of potential energy
and then (ii) by the Ritz method. However, we will see
that the solution procedures are (i) more complicated (but
essentially the same) and (ii) unlike the truss and frame
problems, the problem domain interested is a continuum
and we need to discretize it (by ourselves) into a finite
number of finite elements for analysis. Hence, the method
we used is commonly called the finite element method.

L12-1
 CV7001_L12.docx 24/06/2009

12.1.1 The model problem and governing

differential equations

In this section, we consider a general 3D continuum body

subject to the action of external loading as shown in Fig.
12.1. (Note: In order to cover a wide range of problems, in
here we assumed that the model problem is a 3D problem
and also give no specific descriptions on the geometry of
the body: It can be a building, a car or even one of our
teeth! The only common point is that we would like to
know what will be the responses of the body under the
influence of the external loading.)

Φ
ΓΦ
z, w
n
Ω∈ℜ3

Γu
Body force =Fb

x, u y, v

Fig. 12.1 The model problem of 3D elasticity.

In Fig. 12.1, the problem domain is denoted as Ω and

while its boundary is denoted as Γ (or ∂Ω). It is loaded by
some external tractions forces Φ along part of the domain

L12-2
 CV7001_L12.docx 24/06/2009

boundary ΓΦ (Natural boundary conditions). At the same

time, it is supported along another part of the domain
boundary Γu (essential boundary conditions). The body
force of the domain is denoted as Fb.

From Lecture 3, we know that the equilibrium conditions

(of the whole body or every particles of the body) are

σ x, x + τ xy, y + τzx, z + Fx = 0
τ xy,x + σ y, y + τ yz,z + Fy = 0 (3.9a)
τzx, x + τ yz, y + σ z, z + Fz = 0

or LT σ + Fb = 0 (3.9b)

where Fb=[Fx, Fy, Fz]T and the stress components are

related to the strain component by the equations

ε=Cσ or σ=Eε and C=E-1 (3.14c)

and finally, the strain and displacement relationship is

given by

ε = Lu (3.5c)

or we can write

σ=ELu (12.1)

As the essential boundary conditions are prescribed in

term of the displacement vectors u, it is more convenient

L12-3
 CV7001_L12.docx 24/06/2009

to eliminate the stress term σ from Eqn. 3.9b. Using Eqn.

12.1 we have

LT ELu + Fb = 0 (12.2)

To complete the formulation, we have to express the

support and tractions boundary conditions in terms of
some mathematical equations. For the essential boundary
conditions, as only u is prescribed we can simply write

~ along Γu
u=u (12.3a)

where u~ is the prescribed value of displacement (may or

may not be equal to zero) along the support.

For the natural boundary conditions, we can write

Gσ = Φ along ΓΦ (12.3b)

where G is the normal operator defined in Eqn. 3.13b.

Using, Eqn. 12.1 again, Eqn. 12.3b can be written as

GELu=Φ along ΓΦ (12.3c)

Thus, the general 3D elasticity problem can be

summarized as

Inside the problem domain, Ω

LT ELu + Fb = 0 (12.2)

L12-4
 CV7001_L12.docx 24/06/2009

with boundary conditions

Essential: ~ along Γu
u=u (12.3a)

Natural: GELu=Φ along ΓΦ (12.3c)

with Γu∪ΓΦ=Γ and Γu∩ΓΦ=∅ (12.4)

Example

Consider a 2D simplified case of a deep beam shown in

Fig. 12.2. If the depth (y direction) and length (x direction)
of the beam is much larger than its thickness, the problem
can be considered as a plan stress problem with

u(x,y)=[u(x,y),v(x,y)]T (3.1a)

y
q
A D

Ω h

x
B C
L

Fig. 12.2 A deep beam

L12-5
 CV7001_L12.docx 24/06/2009

⎧ εx ⎫ ⎧ σx ⎫
⎪ ⎪ ⎪ ⎪ ⎡Fx ⎤ ⎡Φ x ⎤
ε = ⎨ ε y ⎬, σ = ⎨ σ y ⎬, Fb = ⎢ ⎥, Φ = ⎢ ⎥ (12.5a)
⎪γ ⎪ ⎪τ ⎪ ⎣ Fy ⎦ ⎣Φ y ⎦
⎩ xy ⎭ ⎩ xy ⎭

⎡∂ ⎤
⎢ 0⎥
⎢ ∂x ⎥
∂⎥ ⎡ l 0 m⎤
L=⎢ 0 , G=⎢ ⎥ (12.5b)
⎢ ∂y ⎥ ⎣0 m l ⎦
⎢∂ ∂⎥
⎢ ⎥
⎣ ∂y ∂x ⎦

If the material of the beam is isotropic,

⎡1 ν 0 ⎤
E ⎢
E= 2 ⎢
ν 1 0 ⎥ (3.16)
1- ν ⎥
⎢⎣ 0 0 (1 - ν)/2⎥⎦

For boundary conditions,

Γ=AB∪BC∪CD∪DA (12.6a)

and
Γu=AB, ΓΦ= BC∪CD∪DA (12.6b)

Essential boundary condition

~ ⎡0 ⎤
Along AB: u=0 or u = ⎢ ⎥ (12.6c)
⎣0 ⎦

L12-6
 CV7001_L12.docx 24/06/2009

Natural Boundary condition

Along BC: l=0, m=-1 → σy=τxy=0 (12.7a)

Along CD: l=1, m=0, → σx=τxy=0 (12.7b)

Along DA l=0, m=1, → σy=-q and τxy=0 (12.7c)

Remarks
(1) Note that L is a first order linear differential operator.
Hence, the governing partial differential equation
(Eqn. 12.2) is second order in u.

(2) Eqn. 12.2 does not assume that E (and Fb) is constant
within the problem domain. In fact, in practice, it is
universal that E will vary within the problem domain.

(3) The problem is considered to be solved only when we

can find a function u(x,y) which (i) satisfies Eqn.
12.2 at every point inside Ω and (ii) satisfies all the
boundary conditions (Eqn. 12.3).

(4) If such a solution is found, the strain and stress at any

point inside Ω can be computed from Eqns. 3.5c and
12.1 respectively.

Of course, now, the main problem is how to solve the

PDE and find the function u(x,y)?

L12-7
 CV7001_L12.docx 24/06/2009

12.1.2 Total potential energy in a linear elasticity

body

As mentions in the last lecture, the total potential energy

of an elasticity body with initial strain can be written as

1
Πp= ∫ (ε T Eε ) − ε T Eε 0dΩ − ∫ u T FbdΩ − ∫ u T ΦdΓ (11.3c)
Ω2 Ω ΓΦ

Now we consider the case that initial stress, σ0, is present

in the material of the elasticity body before the body is
deformed. (Fig. 12.3)

σ σ=E(ε-ε0)+σ0

σ0

ε
ε0

Fig. 12.3 Stress-strain relationship with initial strain and

stress

L12-8
 CV7001_L12.docx 24/06/2009

In this case, the stress-strain relation can be written as

σ=E(ε-ε0)+σ0 (12.8)

while energy per unit volume of the material will be given

by

1
(ε − ε 0 )T E(ε − ε 0 )+ ε T σ 0 (12.9)
2

Using Eqn. 12.9, it can be shown that the total potential

energy expression is given by

1
Πp= ∫ (ε T Eε ) − ε T Eε 0 + ε T σ 0dΩ − ∫ u T FbdΩ − ∫ u T ΦdΓ (12.10a)
Ω2 Ω ΓΦ

Note that Eqn. 12.10a dose not include the case that there
are point loadings applied to the body.

Let P the external point loadings vector acting on the body

and the D is the corresponding displacement vector at
these point loading points. The work done by the point
loading set will then equal to DTP.

As a result, the total potential energy expression including

the effect of point loadings will be given by

1
Πp= ∫ (ε T Eε ) − ε T Eε 0 + ε T σ 0dΩ − ∫ u T FbdΩ − ∫ u T ΦdΓ -DTP
Ω2 Ω ΓΦ
(12.10b)

L12-9
 CV7001_L12.docx 24/06/2009

12.2 Solution of linear elasticity problems by

minimization of total potential energy

In general, the exact analytical solution for the PDE Eqns.

12.2 and 12.3 are not easy to find since the problem
domain and the boundary conditions can be very
complicated. In fact, in many cases, it is difficult even to
write down explicitly the exact mathematical expressions
for Eqn. 12.3. Thus, the analytical solutions for linear
elasticity problems are only limited in some simple
"classical" cases. In real life, only approximated solutions
are available through some approximated methods.

By using the principle of stationary potential energy, we

can prove that

(i) If u is the exact solution of Eqns. 12.2 and 12.3, it

will also minimize the total potential energy of the
system (Eqn. 12.10b).

(ii) In opposite, if we can find an admissible function u

that will minimize Eqn. 12.10b, it will be the exact
solution of Eqns. 12.2 and 12.3.

However, in general, to search for a function u that will

minimize Eqn. 12.10b will be as difficult as solving Eqns.
12.2 and 12.3. Hence, we need to resolve to some
numerical method such as the Ritz method to obtain an
approximated function which will "almost" minimize Eqn.
12.10b. This function will then be a reasonable (or good)
approximated solution of Eqns. 12.2 and 12.3.

L12-10
 CV7001_L12.docx 24/06/2009

The basic idea of applying the minimization principle and

Ritz method for solving the problem is shown
schematically in Fig. 12.4.

Stationary principle Problem of minimization

Problem casted in
form of PDE of potential energy, ΠP

Rayleigh-Ritz method
or other methods

Approximated solution Admissible function that

of PDE "almost" minimizes ΠP

Fig. 12.4 Use of stationary principle and Ritz method for

solution

From Fig. 12.4, it should be emphasized that the Ritz

method is only one of the many possible methods that can
be used to find the approximated function that will
minimize the total potential energy of the system.

Now, we consider the minimization problem, since our

solution must be admissible, the minimization problem
can be stated as

Given

1
Πp= ∫ (ε T Eε ) − ε T Eε 0 + ε T σ 0dΩ − ∫ u T FbdΩ − ∫ u T ΦdΓ -DTP
Ω2 Ω ΓΦ
(12.10b)

L12-11
 CV7001_L12.docx 24/06/2009

Find a function u such that

~ along ∂Ωu
u=u (12.3a)

which will minimize Πp (or makes Πp stationary).

For convenience and simplicity, from now on, when there

is no ambiguity we denote again the approximated
solution as u. By using the Rayleigh-Ritz method, we
know that we will express our trial (approximated)
solution in the form
n n n
u = ∑ N xi u i , v = ∑ N vi vi , w = ∑ N zi w i (12.11a)
i =1 i =1 i =1

where [ui, vi. wi]T are the generalized coordinates for the
displacement components in the x, y and z directions
respectively. Nxi, Nyi and Nzi are respectively the
admissible trial functions for the approximation. If put in
matrix from, Eqn. 12.11a can be written as
⎡ u1 ⎤
⎢v ⎥
⎢ 1⎥
⎢ w1 ⎥
⎢ ⎥
⎡ u ⎤ ⎡ N x1 0 0 . . . N xn 0 0 ⎤⎢ . ⎥
u = ⎢v⎥ = ⎢ 0 N y1 0 . . . 0 N yn 0 ⎥⎢ . ⎥
⎢ ⎥ ⎢ ⎥⎢ ⎥
⎢⎣ w ⎥⎦ ⎢⎣ 0 0 N z1 . . . 0 0 N zn ⎥⎦ ⎢ . ⎥
⎢ un ⎥
⎢ ⎥
⎢ vn ⎥
⎢⎣ w ⎥⎦
n
(12.11b)

L12-12
 CV7001_L12.docx 24/06/2009

or, in matrix form

u=ND (12.11c)

Note that in general, Nxi≠Nyi≠Nzi but in many cases, the

same function is used.

Using Eqn. 3.5c, our approximated strain, ε will be given

by

ε=LND=BD (12.12a)

where

B=LN (12.12b)

is (again!) the strain displacement matrix which related the

generalized coordinates with the approximated strain.

Substitute Eqns. 12.11c and Eqn. 12.12b into Eqn. 12.10b,

we have

1 T T T
Πp = ∫ 2 (BD) E(BD) − (BD) Eε 0 + (BD) σ 0dΩ
Ω
(12.13a)
T T T
− ∫ (ND) FbdΩ − ∫ (ND) Φ dΓ - D P
Ω ΓΦ

L12-13
 CV7001_L12.docx 24/06/2009

Since D is not a function of (x,y), we have

⎡ 1 ⎤ ⎧ ⎫ ⎧ ⎫
Π p = DT ⎢ ∫ B T EBdΩ ⎥ D − DT ⎨ ∫ B T Eε 0dΩ ⎬ + DT ⎨ ∫ B T σ 0dΩ⎬
⎣Ω 2 ⎦ ⎩Ω ⎭ ⎩Ω ⎭
T⎧ T ⎫ ⎧
T⎪ T
⎫⎪ T
− D ⎨ ∫ N FbdΩ ⎬ − D ⎨ ∫ N ΦdΓ ⎬ - D P
⎩Ω ⎭ ⎪⎩ΓΦ ⎪⎭
(12.13b)

As in the cases of truss and beam, to obtain D, all we set

∂Π P
= 0 and obtain
∂D

⎡ T ⎤ ⎧ T ⎫
⎢ ∫ B EBdΩ ⎥ D − ⎨ ∫ B Eε 0dΩ⎬ +
⎣Ω ⎦ ⎩Ω ⎭
(12.14a)
⎧ T ⎫ ⎧ T ⎫ ⎧⎪ T
⎫⎪
⎨ ∫ B σ 0dΩ ⎬ − ⎨ ∫ N FbdΩ ⎬ − ⎨ ∫ N ΦdΓ ⎬ - P = 0
⎩Ω ⎭ ⎩Ω ⎭ ⎪⎩ΓΦ ⎪⎭

⎡ T ⎤ ⎧ T ⎫ ⎧⎪ T
⎫⎪
⎢ ∫ B EBdΩ ⎥ D = ⎨ ∫ N FbdΩ ⎬ + ⎨ ∫ N ΦdΓ ⎬ + P
⎣Ω ⎦ ⎩Ω ⎭ ⎪⎩ΓΦ ⎪⎭
or (12.14b)
⎧ T ⎫ ⎧ T ⎫
+ ⎨ ∫ B E ε 0 dΩ ⎬ − ⎨ ∫ B σ 0 dΩ ⎬
⎩Ω ⎭ ⎩Ω ⎭

Apparently, the matrix on the LHS of Eqn. 12.14b which

assume a form similar to the element/global stiffness
matrix for truss and beam problem (Eqns. 8.4 and 8.26), is
called the stiffness matrix, K, of the system.

L12-14
 CV7001_L12.docx 24/06/2009

For the vectors on the RHS of Eqn. 12.14b, we can see

that they are corresponding to the contributions from the
body force, boundary traction, point loadings, initial strain
and initial stress. We denote the sum of them as F.

Thus, by using the stationary principle and the Ritz

method we finally have

KD=F (12.15a)

where

⎡ T ⎤
K = ⎢ ∫ B EBdΩ ⎥ (12.15b)
⎣⎢Ω ⎥⎦

and

⎧ T ⎫ ⎧⎪ T
⎫⎪ ⎧ T ⎫ ⎧ T ⎫
F = ⎨ ∫ N FbdΩ ⎬ + ⎨ ∫ N ΦdΓ ⎬ + P + ⎨ ∫ B Eε 0 dΩ ⎬ − ⎨ ∫ B σ 0 dΩ ⎬
⎩Ω ⎭ ⎪⎩ΓΦ ⎪⎭ ⎩Ω ⎭ ⎩Ω ⎭

(12.15c)

After D is solved, we can then obtained the approximated

displacement functions u as

u=ND (12.11c)

and the approximated strain field

ε=LND=BD (12.12a)

L12-15
 CV7001_L12.docx 24/06/2009

and finally, the approximated stress field

σ=E(ε-ε0)+σ0 (12.16)

Note that we select the generalize coordinates D, we

assumed that it is always coincidence with the point
loading (Eqn. 12.13a) as we can always include the
displacements at the point loadings points as the unknown
in D.

L12-16
 CV7001_L13.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 13

Displacement finite element solutions to linear

elasticity problems

13.1 Displacement finite element formulation

In the last lecture, a brief outlined was given for the

solution procedure for linear elasticity problems using the
stationary principle and Ritz method. In this lecture, some
more details regarding the solution procedure will be
given.

Recalled that in the pervious lectures, we know that if the

approximated solution is written in the form

u=ND (12.11c)

then the generalized coordinates D can be obtained by

solving the system

KD=F (12.15a)
where

L13-1
 CV7001_L13.docx 24/06/2009

⎡ ⎤
K = ⎢ ∫ B T EBdΩ⎥ (12.15b)
⎣Ω ⎦
and

⎧ T ⎫ ⎧⎪ T ⎫⎪ ⎧ T ⎫ ⎧ T ⎫
F = ⎨ ∫ N FbdΩ⎬ + ⎨ ∫ N ΦdΓ ⎬ + P + ⎨ ∫ B Eε 0dΩ⎬ − ⎨ ∫ B σ 0dΩ ⎬
⎩Ω ⎭ ⎪⎩ΓΦ ⎪⎭ ⎩Ω ⎭ ⎩Ω ⎭

(12.15c)

As mentioned in Lecture 6 and from Eqns. 12.15, we can

expect that the accuracy of the solution will largely
depended on the selection of N and D.

In order to demonstrate the basic idea, we consider the

deep beam problem in Lecture 12 (Fig. 12.2).

y
q
A D

Ω h

x
B C
L

Fig. 12.2 A deep beam

In order to construct the trial solution u, the most obvious

choice is to use polynomial. I.e.

L13-2
 CV7001_L13.docx 24/06/2009

For any point inside Ω, we write

u = a1x + a 2 y + a 3 x 2 + a 4 xy + .......
(13.1)
2
v = b1x + b 2 y + b3 x + b 4 xy + .......

(Why we do not use constant terms in Eqn. 13.1?)

We can expect that by increasing the number of terms

used in Eqn. 13.1, the accuracy of the solution will be
improved. In addition, we can also see that the trial
functions in Eqn. 13.1 are global functions which are non-
zero over the entire problem domain.

However, for a deep beam shown in Eqn. 12.2, we can

expect that the displacement field will be a very
complicated function of position. We also know that, in
general, the use of a global polynomial function may not
be the best approach for approximating the exact solution.
As better way is to construct a number of local solutions
which are defined only over a small part of the problem
domain to approximate the behaviour of the exact solution
locally. The overall picture of the global solution is then
obtained by joining all the local solutions together.

For example, to display complex patterns/images on a TV

or a computer screen, we divide the screen into a finite
number of pixels and display different colours over
individual pixel. Provide that the number of pixels are
large enough (say 1024 × 1024), even if we only have a
finite number of choices of colours in each pixel (e.g. only

L13-3
 CV7001_L13.docx 24/06/2009

black and white), the screen will be able to present

complex image and pattern vividly.

Hence, instead of treating the whole problem domain at

one piece, a better approach is to first divide or
"discretize" the problem domain into a finite number of
small areas or elements as shown in Fig. 13.1.

In general, we denote the domain of a typical element as

Ωi and within Ωi, we now assume that the displacements
will assume a simple form

u = c1 + c 2 x + c3 y
or u=Nd (13.2)
v = g1 + g 2 x + g 3 y

y
q
A D
F
2 4
10
1 3
x
B E C
5
10

Fig. 13.1 Discretized the beam into NE=4 finite elements

(Why we can use constant terms c1 and g1 in Eqn. 13.2?)

In Eqn. 13.2, N is the local trial function and d is the local

displacement parameter.

L13-4
 CV7001_L13.docx 24/06/2009

It must be stressed that Eqn. 13.2 is referred to a typical

element Ωi and each element will has its own sets of ci and
gi, N and d.

Obviously, the total potential energy of the system is the

same of the potential energy of all the elements. That is

NE ⎛ ⎞
⎜ 1 T T ⎟
Π p = ∑ ∫ (ε Eε ) − ε Eε 0 + ε σ 0dΩ − ∫ u FbdΩ − ∫ u ΦdΓ - d p
T T T T
⎜
i =1 Ω i 2
⎟
⎝ Ωi ΓΦ ∩ ΓΩi ⎠
NE
= ∑ Π pΩ i
i =1
(13.3)

Where in Eqn. 13.3, Γ is the boundary of the element

Ωi

Ωi. p is the point loading acting at the nodal point of the

element. The number of elements in the whole domain is
denoted as NE. Π pΩ i , of course, is the total potential
energy of element Ωi.

In order to obtain the local, element stiffness equations,

we can apply exactly the same method described in
Lecture 12 to individual elements. First we substitute the
assumed displacement field (Eqn. 13.2) into the
expression of Π pΩ i . Then upon using the stationary
principle, we can finally obtain the following element
stiffness equation (c.f. Eqn. 12.14b)

L13-5
 CV7001_L13.docx 24/06/2009

⎡ T ⎤ ⎧⎪ T ⎫⎪ ⎧⎪ T
⎫⎪
⎢ ∫ B EBdΩ ⎥d = ⎨ ∫ N FbdΩ ⎬ + ⎨ ∫ N ΦdΓ ⎬ + p
⎣⎢Ω i ⎦⎥ ⎪⎩Ω i ⎪⎭ ⎪⎩ΓΦ ∩ΓΩi ⎪⎭
(13.4a)
⎧⎪ T ⎫⎪ ⎧⎪ T ⎫⎪
+ ⎨ ∫ B E ε 0 d Ω ⎬ − ⎨ ∫ B σ 0 dΩ ⎬
⎪⎩Ω i ⎪⎭ ⎪⎩Ω i ⎪⎭
(9.4a)

or kd=f (13.4b)

(Remember that N and d in Eqn. 13.4 are defined locally

for element Ωi)

where (c.f. Eqn. 12.15b)

⎡ T ⎤
k = ⎢ ∫ B EBdΩ⎥ (13.5a)
⎣⎢Ω i ⎥⎦

is the element stiffness matrix, and

⎧⎪ T ⎫⎪ ⎧⎪ ⎫⎪ ⎧⎪ T ⎫⎪ ⎧⎪ T ⎫⎪
T
f = ⎨ ∫ N FbdΩ⎬ + ⎨ ∫ N ΦdΓ ⎬ + p + ⎨ ∫ B Eε 0dΩ⎬ − ⎨ ∫ B σ 0dΩ⎬
⎪⎩Ω i ⎪⎭ ⎪⎩ΓΦ ∩ΓΩi ⎪⎭ ⎪⎩Ω i ⎪⎭ ⎪⎩Ω i ⎪⎭
(13.5b)

is called the element load vector (c.f. Eqn. 12.15c).

In Eqns. 13.4 and 13.5, the matrix B is the strain-

displacement matrix again defined as (c.f. Eqn. 12.12b)

B=LN (13.6)

L13-6
 CV7001_L13.docx 24/06/2009

Eqns 13.4 to 13.6 can be applied to all the elements in the

domain to obtained NE sets of local stiffness equations.
These equations sets are then can be assembled together to
form the global stiffness matrix as in the cases of truss
and frame problem. In fact, the assembled process can
again be regarded as either (i) a process satisfying
equilibrium or (ii) the summation of total potential energy.
After the global stiffness matrix is formed, boundary
conditions can then be applied in a similar manner as in
the truss and frame problems. Finally, the global
displacement can be solved and the strain and stress in the
problem domain can be computed.

Remarks
(1) Note the main difference between the procedures
described in this lecture and the one described in
Lecture 12 is the assumed solution (trial) form. While
in Lecture 12, the approximation is in general a global
approximation (Eqn. 12.11a), in here we assumed that
the approximation (or interpolation) of unknowns is
carried out in a local, element-by-element manner.
Hence, in Lecture 12, we may not need to partition the
domain into finite elements. The term "finite element
methods" is normally referred to any solution scheme
in which the approximated solution is constructed in a
local, element-by-element manner. Hence, we can see
that the truss and frame problems can be considered as
two particular cases of finite element method.

(2) The most obvious difference between the case of

continuum problem and truss/frame problem is that a
truss/frame is already discretized into well defined

L13-7
 CV7001_L13.docx 24/06/2009

"elements" of bar/beam/column members due to the

geometry of the structural form. However, in
continuum problems, no such "predefined"
discretization exists. In fact, it is all up to the analyst
to design how to partition the domain in to finite
element mesh. The number of nodes and elements,
shapes and forms of elements in the mesh are not fixed
and can be changed according to the needs in practical
situations. All the analyst need is to follow some
(obvious and easy to remember) basic rules when
preparing the mesh.

(3) The existence of a width range of freedom for

choosing the form and design of the finite element
mesh should be regarded as an advantage rather than a
shortcoming of the method. In fact, it is the main
reason why the finite element method is so popular in
the solutions for engineering and scientific problems.
Well designed finite element meshes (and sometime
even special formulation of finite elements) can be
used to suit the special need for the problems (e.g. due
to the complex geometry of the problem domain or
special strain-stress distributions).

(4) Normally, the following terminology is commonly

used when describing the partition of the domain.

The whole domain is partitioned in to a finite number

of elements (denoted as NE) and the finite elements
make up the finite element mesh. Each element in the
mesh will associate with a finite number of nodal
points. The number of nodal points per element will

L13-8
 CV7001_L13.docx 24/06/2009

depend on the form of the element. For example, for

triangular elements, the elements will have three
corner nodes while for quadrilateral elements, the
elements will have four corner nodes. Each of the
nodal point in the mesh will associate with a certain
number of degree of freedom (dof). The total degree of
freedom of the mesh is simply the product of the
number of degree of freedoms per node and the
number of nodes (denoted as NN) in the mesh.

Usually, the most important question is how to select

(construct) the local approximating (interpolating)
functions N and the local generalized coordinates d. From
the experience in truss/frame problem, we can see that it
will be easier for us to select the generalized coordinates
first. Also from the experience in truss/frame problem, we
can conclude that the displacements at the nodal points of
the elements will be a convenient choice. By selecting the
nodal displacements as the generalized coordinates, the
user can pick up some critical points in the structures and
obtained there displacements. Furthermore, using the
nodal points will largely simplify the assembly procedure
of the global stiffness matrix and the imposition of
boundary conditions.

13.2 Example

To demonstrate the procedure, we consider element 1

shown in Fig. 13.1 as an example (Fig. 13.2). Assume
plane stress conditions are applied with E=96000 and
ν=0.2.

L13-9
 CV7001_L13.docx 24/06/2009

Recalling Eqn. 13.2, we have

u = c1 + c 2 x + c3 y
(13.2)
v = g1 + g 2 x + g 3 y

y
A: (0,10)
B: (0,0)
E: (5,0)
A
E=96000
ν=0.2
Plane stress conditions
No initial strain or stress
1 Fb=[0,-1]T
x
B E

Fig. 13.2 Element 1

At point A(10,0), u=[uA,vA]T, at point B (0,0), u=[uB,vB]T

and at point E (5,0), u=[uE,vE]T. Thus,

⎡u A ⎤ ⎡c1 + 10c3 ⎤ ⎡1 0 10⎤ ⎡ c1 ⎤

⎢u ⎥ = ⎢ c ⎥ = ⎢1 0 0 ⎥ ⎢c ⎥ (13.7)
⎢ ⎥ ⎢B 1 ⎥ ⎢ ⎥⎢ 2 ⎥
⎢⎣ u E ⎥⎦ ⎢⎣ c1 + 5c 2 ⎥⎦ ⎢⎣1 5 0 ⎥⎦ ⎢⎣c3 ⎥⎦

It can be shown that then matrix in Eqn. 13.7 is not

singular and thus we can solve for the value of ci, i=1,...,3
and obtain

c1=uB, c2=(uE-uB)/5, c3=(uA-uB)/10 (13.8)

Hence,

L13-10
 CV7001_L13.docx 24/06/2009

u=uB+x(uE-uB)/5+y(uA-uB)/10 (13.9a)

Similarly,

v=vB+x(vE-vB)/5+y(vA-vB)/10 (13.9b)

It will be more useful to rewrite Eqn. 13.9a in the form

u=uA(y/10)+uB(1-x/5-y/10)+uE(x/5)=uANA+uBNB+uENE (13.9c)

such that

NA=y/10, NB=(1-x/5-y/10), NE=x/5 (13.9d)

Similarly, for the expression of v,

v=vANA+vBNB+vENE (13.9e)

or in matrix form

⎡u A ⎤
⎢v ⎥
⎢ A⎥
⎡u ⎤ ⎡ N 0 NB 0 NE 0 ⎤⎢u B ⎥
u=⎢ ⎥=⎢ A ⎥ ⎢ ⎥ = Nd (13.10)
⎣ v⎦ ⎣ 0 NA 0 NB 0 NE ⎦ ⎢ vB ⎥
⎢u E ⎥
⎢ ⎥
⎣ vE ⎦

(Note: c.f. Eqn. 13.10 with Eqn. 12.11b)

L13-11
 CV7001_L13.docx 24/06/2009

In Eqns. 13.9 and 13.10, the functions NA, NB and Nc are

commonly called the element shape functions while d is
the element displacement vector.

With the assumed displacement defined, we can proceed

on to compute the B matrix as

⎡∂ ⎤
⎢ 0⎥
⎢ ∂x ⎥
∂ ⎥⎡NA 0 NB 0 NE 0 ⎤
B=⎢ 0
⎢ ∂y ⎥ ⎢⎣ 0 NA 0 NB 0 N E ⎥⎦
⎢∂ ∂⎥
⎢ ⎥
⎣ ∂y ∂x ⎦
⎡ ∂N A ∂N B ∂N E ⎤
⎢ 0 0 0 ⎥
⎢ ∂x ∂x ∂x ⎥
∂N A ∂N B ∂N E ⎥
=⎢ 0 0 0
⎢ ∂y ∂y ∂y ⎥
⎢ ∂N ∂N A ∂N B ∂N B ∂N E ∂N E ⎥
⎢ A ⎥
⎣ ∂y ∂x ∂y ∂x ∂y ∂x ⎦
⎡0 0 - 0.2 0 0.2 0 ⎤
= ⎢ 0 0.1 0 - 0.1 0 0⎥
⎢ ⎥
⎢⎣0.1 0 - 0.1 - 0.2 0 0.2⎥⎦
(13.11)

Note that this simply means that the displacement-strain

matrix is constant within the element.

(Hence: What about the assumed strain/stress within the

element?)

L13-12
 CV7001_L13.docx 24/06/2009

The E matrix is given by

⎡5 1 0 ⎤
E = 20000⎢1 5 0⎥ (13.12)
⎢ ⎥
⎢⎣0 0 2⎥⎦

Therefore, the product BTEB is given by

⎡ 400 0 − 400 − 800 0 800 ⎤

⎢ 0 1000 − 400 − 1000 400 0 ⎥
⎢ ⎥
⎢− 400 − 400 4400 1200 − 4000 − 800 ⎥
⎢ ⎥ (13.13)
⎢ − 800 − 1000 1200 2600 − 400 − 1600 ⎥
⎢ 0 400 − 4000 − 400 4000 0 ⎥
⎢ ⎥
⎣ 800 0 − 800 − 1600 0 1600 ⎦

Since BTEB is constant, the element stiffness matrix is

simply the product of the matrix shown in Eqn. 13.13 and
the area of the element=25. Thus,

⎡ 400 0 − 400 − 800 0 800 ⎤

⎢ 0 1000 − 400 − 1000 400 0 ⎥
⎢ ⎥
⎢− 400 − 400 4400 1200 − 4000 − 800 ⎥
k = 25⎢ ⎥
⎢ − 800 − 1000 1200 2600 − 400 − 1600 ⎥
⎢ 0 400 − 4000 − 400 4000 0 ⎥
⎢ ⎥
⎣ 800 0 − 800 − 1600 0 1600 ⎦
(13.14)

L13-13
 CV7001_L13.docx 24/06/2009

For the element load vector, since there is no initial strain

or stress, the last two terms in Eqn. 13.5b will equal to
zero. In addition, as no point loading is acting on the
element, the third terms of Eqn. 13.5b will also be equal to
zero.

For the second term in Eqn. 13.5b, as the boundary of the

element is AB∪BE∪EA. Its intersection with the domain
boundary is simply equal to AB∪BE. It is obvious that
AB=Γu, the boundary for the essential boundary condition.
Hence, the only part of ΓΩ1 intersect with ΓΦ is BE.
However, from Eqn. 12.7a, the natural boundary condition
is

Along BC: l=0, m=-1 →τxy=0 (12.7a)

Thus, the second term in Eqn. 13.5b is also equal to zero.

For the first term in Eqn. 13.5b, it is given by

L13-14
 CV7001_L13.docx 24/06/2009

⎡NA 0 ⎤ ⎡ 0 ⎤
⎢ 0 N ⎥ ⎢- N ⎥
⎢ A ⎥ ⎢ A⎥
T ⎢ NB 0 ⎥⎡ 0 ⎤ ⎢ 0 ⎥
∫ bN F d Ω = ∫ ⎢ 0 N ⎥ ⎢− 1⎥ d Ω = ∫ ⎢ - N ⎥ dΩ
Ωi Ωi ⎢ B ⎥⎣ ⎦ Ωi ⎢ B⎥
⎢ NE 0 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎣ 0 N E⎦ -
⎣ E⎦N
⎡ 0 ⎤ ⎡ 0 ⎤
⎢ - y/10 ⎥ ⎢- 25/3⎥
⎢ ⎥ ⎢ ⎥
5 10 (1- x/5)
⎢ 0 ⎥ ⎢ 0 ⎥
=∫ ∫ ⎢ ⎥ dydx = ⎢ ⎥
0 0 ⎢ - (1 - x/5 - y/10) ⎥ ⎢ - 25/3 ⎥
⎢ 0 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
⎣ - x/5 ⎦ ⎣- 25/3⎦
(13.14)

Thus, the element stiffness equation will be given by

⎡ 400 0 − 400 − 800 0 800 ⎤ ⎡u A ⎤ ⎡ 0 ⎤

⎢ 0 1000 − 400 − 1000 400 0 ⎥ ⎢ v A ⎥ ⎢- 25/3⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢− 400 − 400 4400 1200 − 4000 − 800 ⎥ ⎢ u B ⎥ ⎢ 0 ⎥
25⎢ ⎥⎢ ⎥ = ⎢ ⎥
⎢ − 800 − 1000 1200 2600 − 400 − 1600 v
⎥⎢ ⎥ ⎢
B - 25/3⎥
⎢ 0 400 − 4000 − 400 4000 0 ⎥⎢u E ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎣ 800 0 − 800 − 1600 0 1600 ⎦ ⎣ v E ⎦ ⎣- 25/3⎦
(13.15)

Similar procedure can be used to compute the element

stiffness equations for elements 2 to 4. (Note that for
elements 2 and 4, the loading terms corresponding to the

L13-15
 CV7001_L13.docx 24/06/2009

boundary traction vector will not be equal to zero). These

element stiffness equations obtained can then be
assembled to form the global stiffness matrix. For
example, if the node numbering scheme shown in Fig.
13.3 is used, the corresponding global stiffness matrix
pattern will assume the form shown in Eqn. 13.16a.
y
q
4 6
5
2 4
10
1 3
x
1 2 3
5
10

Fig. 13.3 A numbering scheme for the deep beam

problem.

⎡x x x x 0 0 x x 0 0 0 0 ⎤ ⎡ u1 ⎤ ⎡ 0 ⎤
⎢x x x x 0 0 x x 0 0 0 0 ⎥ ⎢ v1 ⎥ ⎢ x ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢x x x x x x x x x x 0 0 ⎥ ⎢u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢x x x x x x x x x x 0 0 ⎥ ⎢v2 ⎥ ⎢x ⎥
⎢0 0 x x x x 0 0 x x x x ⎥ ⎢u3 ⎥ ⎢0⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 x x x x 0 0 x x x x ⎥ ⎢ v3 ⎥ ⎢ x ⎥
=
⎢x x x x 0 0 x x x x 0 0 ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢x x x x 0 0 x x x x 0 0 ⎥ ⎢v4 ⎥ ⎢x ⎥
⎢0 0 x x x x x x x x x x ⎥ ⎢u 5 ⎥ ⎢0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 x x x x x x x x x x ⎥ ⎢ v5 ⎥ ⎢ x ⎥
⎢0 0 0 0 x x 0 0 x x x x ⎥ ⎢u 6 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣ 0 0 0 0 x x 0 0 x x x x ⎥⎦ ⎢⎣ v 6 ⎥⎦ ⎢⎣ x ⎥⎦
(13.16a)

L13-16
 CV7001_L13.docx 24/06/2009

And upon the imposition of boundary conditions such that

u1=v1=u4=v4=0.

⎡1 0 0 0 0 0 0 0 0 0 0 0 ⎤ ⎡ u1 ⎤ ⎡ 0 ⎤
⎢0 1 0 0 0 0 0 0 0 0 0 0 ⎥ ⎢ v1 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 x x x x 0 0 x x 0 0 ⎥ ⎢u 2 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 x x x x 0 0 x x 0 0 ⎥ ⎢v2 ⎥ ⎢x ⎥
⎢0 0 x x x x 0 0 x x x x ⎥ ⎢u3 ⎥ ⎢0⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 x x x x 0 0 x x x x ⎥ ⎢ v3 ⎥ ⎢ x ⎥
=
⎢0 0 0 0 0 0 1 0 0 0 0 0 ⎥ ⎢u 4 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 0 0 0 0 0 1 0 0 0 0⎥ ⎢v4 ⎥ ⎢0⎥
⎢0 0 x x 0 0 0 0 x x x x ⎥ ⎢u5 ⎥ ⎢0⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢0 0 x x 0 0 0 0 x x x x ⎥ ⎢ v5 ⎥ ⎢ x ⎥
⎢0 0 0 0 x x 0 0 x x x x ⎥ ⎢u 6 ⎥ ⎢ 0 ⎥
⎢ ⎥⎢ ⎥ ⎢ ⎥
⎢⎣0 0 0 0 x x 0 0 x x x x ⎥⎦ ⎣⎢ v6 ⎦⎥ ⎢⎣ x ⎥⎦
(13.16b)

and the global displacement vector D can then be solved.

Remarks
(1) The above formulation which uses the nodal
displacements as the unknowns is commonly referred
as the "displacement finite element formulation".

(2) After the vector D is found, to compute the

displacement at any point (x,y) in the problem
domain, it is only required to (i) locate the element
which contains the point, (ii) use Eqn. 13.9 (or Eqn.
13.10).
(What happened if one wants to compute the
displacement at a point located exactly on the
interface two elements?)

L13-17
 CV7001_L13.docx 24/06/2009

(3) For strain and stress, one needs to use the equation
(c.f. Eqn. 12.12a)

ε=LNd=Bd (13.17a)

and

σ=E(ε-ε0)+σ0 (12.16)

(4) If we denote

NDN= No. of displacement unknowns per node.

NNE = No. of nodes per element
NSC = No. of stress/strain components of the
problem.
NN = No. of nodes of the finite element mesh

then enclosed below is a summary of sizes of

matrices and vectors appeared in the above
formulation
Fb, Φ = NDN×1
ε0, σ0 = NSC×1
E= NSC×NSC
N = NDN×(NNE×NDN)
B = NSC×(NNE×NDN)
f, p,d = (NNE×NDN)×1
k = (NNE×NDN)×(NNE×NDN)
F = (NN×NDN)×1
K = (NN×NDN)× (NN×NDN)
(Note: In our example: NDN=2, NSC=3, NNE=3, NN=6,
What about the case of full 3D problem?

L13-18
 CV7001_L14.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 14

Properties of interpolation functions

14.1 Properties element interpolation (shape)

functions

In the last lecture the basic computing procedures for the

displacement finite element formulation were given. It can
be seen that after the displacement shape functions and the
nodal displacements are selected, the procedures for the
formation of element stiffness matrix and element loading
vector are very similar to the truss and frame problems. In
addition, the assembly procedure for the global stiffness
equations follows exactly the same way as in the usual
matrix structural analysis processes. Finally, the
imposition of boundary conditions again follows a
straightforward way.

Hence, we can conclude that the procedure to carry out

finite element analysis is in fact very similar to the way we
carry out matrix structural analysis. However, it should be
reminded that there are two major differences

(1) We need to create our only finite element mesh (or

discretization) which will determine implicitly the

L14-1
 CV7001_L14.docx 24/06/2009

unknowns in the global displacement vector. That is,

the design and creation of finite element mesh will
also fix the generalized coordinates used in the
approximated solutions.

(2) We need to select the trial functions used in the

construction of the approximated solution (Eqn.
13.2).

We note that the accuracy of the finite element solutions

will be closely depended on the selection of (1) and (2).
For the design of finite element mesh, more details will be
given in Lecture 20, while in this lecture, concentration
will be focused on the construction of the local trial
functions and their properties.

To begin with, let us reconsider the simple 3-node

triangular element used in Lecture 13 (Fig. 13.2) as shown
below. Note that in here we relabel the nodes of the
element as 1-2-3 instead of ABE.

y
A: (0,10)
B: (0,0)
E: (5,0)
1
E=96000
ν=0.2
Plane stress conditions
No initial strain or stress
1 Fb=[0,-1]T
x
2 3

Fig. 14.1 A simple 3-node triangular element

L14-2
 CV7001_L14.docx 24/06/2009

Recall that the approximation of u within the element is

given by

u=u1(y/10)+u2(1-x/5-y/10)+u3(x/5)=u1N1+u2N2+u3N3 (13.9c)

such that

N1(x,y)=y/10, N2(x,y)=(1-x/5-y/10), N3(x,y)=x/5 (13.9d)

and if we denote the coordinates of node i as (xi,yi) with

(x1,y1)=(0,10), (x2,y2)=(0,0) and (x3,y3)=(5,0). Then, it will
not be difficult to find that

For N1: N1(x1,y1)=1.0, N1(x2,y2)=0 and N1(x3,y3)=0 (14.1a)

For N2: N2(x1,y1)=0, N2(x2,y2)=1 and N2(x3,y3)=0 (14.1b)

For N3: N3(x1,y1)=0, N3(x2,y2)=0 and N3(x3,y3)=1 (14.1c)

or

Ni(xj,yj)=δij (14.2)

where in Eqn. 14.2, the symbol δij is called the Kronecker

delta symbol such that

⎧0 if i ≠ j
δij = ⎨ (14.3)
⎩1 if i = j

and Eqn. 14.2 is known as the interpolant condition.

L14-3
 CV7001_L14.docx 24/06/2009

That is, the shape functions always assume a value of

unity (1.0) at its own node but vanish at all the other node.
In addition, if one plot out the value of shape functions
(shown in Fig. 14.2), one will see that the shape functions,
Ni will also vanish at the edge opposite to its own node
and vary linearly along the other two edges.

1.0
1 1
1

1.0

2 2 1.0
2
3 3
3
N1 N2
N3

Fig. 14.2 Shape functions for the 3-node triangular

element

From Eqn. 13.9c, the assumed displacement field u is a

linear combination of these three shape functions and will
also vary linearly along the edges of the element as shown
in Fig. 14.3.

More importantly, since the value of shape function at the

node opposite to an edge will not affect the value of u
along that edge, we can conclude that the assumed
displacement field will be continuous along the inter-
element boundary and therefore will be continuous (C0)
over the whole problem domain.

L14-4
 CV7001_L14.docx 24/06/2009

u within an element along the interelement boundary

Fig. 14.3 u inside the element and along the interface

between adjacent elements

Finally, at any point inside the element, we find that

3
∑ Ni ≡ 1.0 (14.4).
i =1

In fact, the above observed properties are the minimum

requirements need for a valid approximation of the
problem using the finite element method for the reason
that

(1) Assumed that our shape functions do not satisfy the

interpolation conditions. For example, if N3(x2,y2)≠0
and N3(x1,y1)≠0. Now for element 1 shown in Fig.
14.1, along the edge 1-2, we know that u=v=0. Recall
that the expression of u is

u(x,y)=u1N1(x,y)+u2N2(x,y)+u3N3(x,y) (14.5)

setting u1=u2=0, x=x1,y=y1 and u3≠0 we have

L14-5
 CV7001_L14.docx 24/06/2009

u(x,y)=u3N3(x1,y1)≠0 (14.6)

Therefore, if the shape functions do not satisfy the

interpolant condition, the finite element solution will
not satisfy the essential boundary condition.

(2) Assume that now we refine the mesh gradually by

reducing the element size by increasing the number
of elements in the mesh. At the limited state we will
have the sizes of all the elements are reduced to zero
and become a point. Thus, as we reduce the size of
the element, the approximated displacement solution
shall converge to a constant value (why?) such that
u1=u2=u3=α. Therefore,

3 3 3
u = ∑ Ni u i = ∑ Niα = α∑ Ni = α (14.7)
i =1 i =1 i =1

Thus, condition shown in Eqn. 14.4 is essential for

the element achieve a limit state of the solution if the
mesh is refined. In addition, the condition will also
imply that the approximated solution will able to
represent the rigid body motion.

(3) It is obvious that the exact solution of displacement is

continuous and hence the C0 property of the
approximated solution is also required.

L14-6
 CV7001_L14.docx 24/06/2009

14.2 Different types of finite elements

The following terms are frequently used to describe a

finite element

(1) The number of nodes in an element (NNE)

(2) The Dimension of an element. I.e. 1D, 2D or 3D

elements.

(3) The shape of an element. I.e. line (1D), triangular or

quadrilateral (2D), tetrahedral or prism or brick
element (3D). See Fig. 14.4.

(4) The order of an element, p, is referred to the highest

complete polynomial order of the element shape
functions (linear, quadratic, cubic etc). An element is
an order p element if and only if its shape functions
are complete up to order p. For example, for the 3-
node triangular element we discussed in the last
section. It is obviously the shape functions Ni,
i=1,...,3 contain that terms

{1,x,y} (14.8)

and the shape functions are said to be completed up

to first order terms. For 2D problems, the shape
functions of a second order element should contain
the terms

{1,x,y,x2,xy,y2} (14.9a)

while for a second order 3D element

L14-7
 CV7001_L14.docx 24/06/2009

{1,x,y,z,xy,xz,yz,x2,y2,z2} (14.9b)

In general, for 2D problems, a pth order element

should contain at least (p+1)(p+2)/2 while for 3D
element, a pth order element should contain at least
(p+2)(p+3)(p+4)/6 terms. Furthermore, the number of
terms in the shape functions is exactly equal to the
number of nodes of the element. (Why?)
1D Elements

Linear element (C0) Beam element (C1) quadratic element (C0)

2D Elements

T3, linear triangle T6, quadratic triangle T10, cubic triangle

L9, quadratic Lagangian Q12, cubic Lagangian

Q4, linear quadrilateral
quadrilateral quadrilateral

Fig. 14.4a Some commonly used finite elements,

1D and 2D

L14-8
 CV7001_L14.docx 24/06/2009

3D Elements

T4 linear tetrahedral element T10 quadratic tetrahedral element P6 linear prism element

P15 quadratic prism element H8 linear brick element H20 quadratic brick element

Fig. 14.4b Some commonly used finite elements, 3D

(5) The order of continuity of the element

Note that for displacement type elements, the shape
function Ni, i=1,....,NEN is differentiable up to any
order within the element. However, in general, along
the inter-element boundary, the shape functions are
only C0 functions (Fig. 14.5) unless special
considerations are taken (e.g. the beam element
discussed in Lecture 8). As a result, displacement
type elements are also known as C0 continuous
elements.

C0 continuous displacement discontinuous first order derivative

finite element solution of finite element solution

Fig. 14.5 C0 continuous element

L14-9
 CV7001_L15.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 15

Coordinate transformation and mapping of

elements

15.1 Construction of element shape functions:

coordinate transformation and mapping

From previous lectures, we see that the following

procedures are used for the construction of element shape
functions.

(1) Base on the number of nodes in the element, select

the polynomial expression of the assumed functions.
(Eqn. 13.2).

(2) Correlate the coefficients of the polynomial

expression with the nodal displacements of the
elements (Eqn. 13.7). From this, we can solve for the
coefficients in terms of the nodal displacements and
the coordinates of the nodal points of the element. In
this step, we always need to invert a matrix of size
equal to NNE (Eqn. 13.7). This is the reason why the
number of coefficients in the original expression must
exactly equal to the number of nodal points of the
elements.

L15-1
 CV7001_L15.docx 24/06/2009

(3) After solving the coefficients, the shape functions of

the element are obtained by rearrange the terms in the
original expression of displacement (Eqn. 13.9c)

This is not difficult to see that the above procedure has a

number of shortcomings. First of all, a matrix is required
to be inverted for the construction of the shape functions.
Worst still, we find that the shape functions will depend
on the nodal coordinates of the elements. This simply
means that given two different elements located at
different positions of the structure, even though their
element shapes and sizes are the same, we still need to
invert two different matrices for the construction of their
shape functions.

Thus, a better idea is to construct a set of element shape

functions which are independent of the nodal positions.
The solution turns out is quite simple and only involve
some elementary coordinate transformations. To illustrate
the basic idea, a 1D problem is first considered.

1D problem

In Lecture 7, we know that the shape function for the bar

element is given by (Eqn. 7.5g)

N1(x)=(1-x/L) and N2(x)=x/L (15.1)

and we can "normalize" it to the form

N1=(1-ξ), N2=ξ (15.2)

by defining ξ=x/L.

L15-2
 CV7001_L15.docx 24/06/2009

The above normalization procedure can be regarded as a

mapping or coordinate transformation where the physical
space, x (the bar axis) is mapped to a
normalized/parametric space which is simply the straight
line [0,1] with normalized or parent or parametric
coordinate ξ (Fig. 15.1).

x
x=0

x=L

ξ=0 ξ=1

Fig. 15.1 Mapping of 1D bar element

By using the same transformation, any point with parent

coordinate ξ0 will be mapped to a point ξ0L in the physical
space. In general, if x1 and x2 are the coordinates of end
points of the bar. Then for a point with parent coordinate
equal to ξ, its corresponding point x in the physical space
will be given by

x=N1(ξ)x1+N2(ξ)x2 (15.3a)

or
⎡ x1 ⎤
x=[N1(ξ), N2(ξ)] ⎢ ⎥ (15.3b)
⎣x 2 ⎦

L15-3
 CV7001_L15.docx 24/06/2009

Now it is clear that Eqn. 15.3b defines the mapping

relationship between the physical space and the parametric
space.

Note that Eqn. 15.3b takes a similar form as the

interpolation of the axial displacement (Eqn. 7.7) with
only u replaced by x.

⎡u ⎤
u=[N1(ξ), N2(ξ)] ⎢ 1 ⎥ (15.4)
⎣u 2 ⎦

Thus, the process of shape functions construction can be

carried out without first considering the geometry of the
element. We only need to first construct some appropriate
shape functions (which satisfy the conditions discussed in
Lecture 14). These shape functions can then be used in the
interpolation of both the geometry of the element and the
unknown function (displacements). Obviously, the order
of the element, p, will be equal to the order of the
interpolation function used.

From Eqns. 15.3 and 15.4, we can see that the same set of
shape functions N1 and N2 are now interpolate both the
displacements (unknown function) and the geometry of
the bar element. Finite elements constructed in this way,
are known as isoparametric finite elements (iso=the
same).

Hence, the bar element discussed in Lecture 7 is an

isoparametric element. However, the beam element is not
an isoparametric element. The reason is that the shape
functions used for the displacement interpolation are cubic
(the Hermite functions, Eqn. 8.17b)
L15-4
 CV7001_L15.docx 24/06/2009

w'=w1'(1-3ξ2+2ξ3)+θ1'L(ξ-2ξ2+ξ3)+w2'(3ξ2-2ξ3)+θ2'L(ξ3-ξ2)
(8.17b)

while only the linear shape functions (Eqn. 15.3b) are used
for geometry interpolation. This kind of finite element in
which the order of interpolation of displacement is higher
than the order of interpolation of geometry is called
superparametric element (super=higher).

Clearly, it is possible that we can have the case that the

order of interpolation of displacement is lower than the
order of interpolation of geometry. This kind of element is
called the subparametric element.

In this course, only the isoparametric elements will be

discussed for 2D and 3D problems.

2D case

For 2D case, the same procedure can again be used to

construct the interpolation functions for displacement and
geometry. The only difference is that the mapping
procedure is slightly more complicated.

Consider the simplest case of a linear triangle with

arbitrary shape as shown in Fig. 15.2.

L15-5
 CV7001_L15.docx 24/06/2009

(x3,y3)
3

η
2
(0,1) 3
1 (x2,y2)

(x1,y1)

1 2
ξ
(0,0) (1,0)

Fig. 15.2 Isoparametric mapping for a linear triangle

In this case, as the shape of the element in the physical

space is a triangle, the most suitable choice of shape of the
parent element is obviously again a triangle and in
particular, a right angle triangle with unit length as shown
again in Fig. 15.2.

The element has three nodes so we need three shape

functions for both displacement and geometry. Obviously,
we would like to map node 1 in the parametric (ξ-η) space
to node 1 in the physical space (x1,y1) and so on. With
some simple calculations, the shape functions are found to
be equal to

N1(ξ,η)=1-ξ-η N2(ξ,η)=ξ, and N3(ξ,η)=η (15.5)

and it can be shown that the shape functions in Eqn. 15.5

satisfy all the requirements mentioned in the Lecture 14.

Now, we can write the interpolation of both the

displacement and geometry in the form

L15-6
 CV7001_L15.docx 24/06/2009

⎡u ⎤ ⎡u ⎤ ⎡u ⎤ ⎡u ⎤
u = ⎢ ⎥ = N1 (ξ, η) ⎢ 1 ⎥ + N 2 (ξ, η) ⎢ 2 ⎥ + N3 (ξ, η) ⎢ 3 ⎥
⎣v⎦ ⎣ v1 ⎦ ⎣v2 ⎦ ⎣ v3 ⎦
(15.6a)
3 ⎡u i ⎤ 3
= ∑ N i (ξ, η) ⎢ ⎥ = ∑ N i (ξ, η)ui
i =1 ⎣ vi ⎦ i =1

and

⎡x ⎤ ⎡x ⎤ ⎡x ⎤ ⎡x ⎤
x = ⎢ ⎥ = N1 (ξ, η) ⎢ 1 ⎥ + N 2 (ξ, η) ⎢ 2 ⎥ + N 3 (ξ, η) ⎢ 3 ⎥
⎣ y⎦ ⎣ y1 ⎦ ⎣y2 ⎦ ⎣ y3 ⎦
(15.6b)
3 ⎡x i ⎤ 3
= ∑ N i (ξ, η) ⎢ ⎥ = ∑ N i (ξ, η)x i
i =1 ⎣ yi ⎦ i =1

Provide that we know how to construct the shape

functions, it is obvious that Eqn. 15.6 can be generalized
to any 1D, 2D and 3D element. If NNE is the number of
nodes in the element, the generalized expression of Eqn.
15.6 will equal to

1D elements

NNE NNE
u= ∑ Ni (ξ, η)u i , x= ∑ Ni (ξ, η)x i (15.7a)
i =1 i =1

2D/3D elements:

NNE NNE
u= ∑ Ni (ξ, η)ui , x= ∑ Ni (ξ, η)xi (15.7b)
i =1 i =1

L15-7
 CV7001_L15.docx 24/06/2009

For 2D elements

u=[u,v]T, ui=[ui, vi]T and x=[x,y]T, xi=[xi, yi]T (15.8a)

For 3D elements

u=[u,v,w]T, ui=[ui, vi, wi]T and x=[x,y,z]T, xi=[xi, yi, zi]T (15.8b)

15.2 Coordinate transformation of isoparametric

elements

The governing equations of the physical system are (Eqn.

12.2) written in the global coordinate system and involves
∂u ∂v
unknowns u and their derivatives , etc. However,
∂x ∂y
for parametric elements, the interpolation functions are
defined in terms of the parent coordinates ξ=(ξ,η).
Therefore, in order to compute the element stiffness
matrices and the load vectors for parametric elements, it is
necessary to establish some means to express the global
∂u ∂v
derivatives , etc. in terms of the local derivatives
∂x ∂y
∂u ∂v
, etc.
∂ξ ∂η

Consider the case of 2D problems for which

NNE NNE
u= ∑ Ni (ξ, η)ui , x= ∑ Ni (ξ, η)xi (15.7b)
i =1 i =1

which implies

L15-8
 CV7001_L15.docx 24/06/2009

∂u NNE ∂N i (ξ, η) ∂u NNE ∂N i (ξ, η)

= ∑ ui , = ∑ u i (15.9)
∂x i =1 ∂x ∂y i =1 ∂y

∂u ∂v ∂N i ∂N i
Therefore, in order to compute , etc., , are
∂x ∂y ∂x ∂y
required. Now from the chain rule of differentiation, for
2D problems,

⎡ ∂N i ⎤ ⎡ ∂x ∂y ⎤ ⎡ ∂N i ⎤ ⎡ ∂N i ⎤
⎢ ∂ξ ⎥ ⎢ ∂ξ ∂ξ ⎥ ⎢ ∂x ⎥ ⎢ ∂x ⎥
⎢ ⎥=⎢ ⎥ = J ⎢ ∂N ⎥ (15.10a)
∂
⎢ i ⎥ ⎢ ∂x
N ∂y ⎥ ⎢ ∂N i ⎥
⎢ ⎥ ⎢ i⎥
⎢⎣ ∂η ⎥⎦ ⎢⎣ ∂η ∂η ⎥⎦ ⎣ ∂y ⎦ ⎣ ∂y ⎦

−1
⎡ ∂N i ⎤ ⎡ ∂x ∂y ⎤ ⎡ ∂N i ⎤ ⎡ ∂N i ⎤
⎢ ∂x ⎥ ⎢ ∂ξ ∂ξ ⎥ ⎢ ∂ξ ⎥ ⎢
−1 ∂ξ
⎥
or ⎢ ∂N ⎥ = ⎢ ∂x ∂y ⎥ ⎢ ⎥=J ⎢ ⎥ (15.10b)
∂N
⎢ i⎥ ∂N
⎢ i⎥
⎢ i⎥ ⎢ ⎥
⎣ ∂y ⎦ ⎢⎣ ∂η ∂η ⎥⎦ ⎣⎢ ∂η ⎥⎦ ⎣⎢ ∂η ⎥⎦

x, y
where the Jacobian J= J ( ) =[Jkl], k,l=1,...,2 can be
ξ, η
computed by the geometric interpolation as

∂x NNE ∂N i ∂y NNE ∂N i
J11 = = ∑ x i , J12 = = ∑ yi
∂ξ i =1 ∂ξ ∂ξ i =1 ∂ξ
(15.11a)
∂x NNE
∂N i ∂y NNE
∂N i
J 21 =
∂η
= ∑ ∂η
x i , J 22 =
∂η
= ∑ ∂η
yi
i =1 i =1

The determinant of the Jacobian Det(J) is simply equal to

J11J22-J12J21 and

L15-9
 CV7001_L15.docx 24/06/2009

1 ⎡ J 22 - J12 ⎤
J −1 = (15.11b)
Det(J ) ⎢⎣- J 21 J11 ⎥⎦

Similarly for 3D problems

∂N i x, y, z ∂N i ∂N i x, y, z −1 ∂N i
= J( ) and = J( ) (15.12a)
∂ξ ξ, η, ζ ∂x ∂x ξ, η, ζ ∂ξ

where

∂x NNE ∂Ni ∂y NNE ∂Ni ∂z NNE ∂Ni

J11 = = ∑ x i , J12 = = ∑ yi , J13 = = ∑ zi
∂ξ i =1 ∂ξ ∂ξ i =1 ∂ξ ∂ξ i =1 ∂ξ
∂x NNE ∂Ni ∂y NNE ∂N i ∂z NNE ∂N i
J 21 = = ∑ x i , J 22 = = ∑ yi , J 23 = = ∑ zi
∂η i =1 ∂η ∂η i =1 ∂η ∂η i =1 ∂η
∂x NNE ∂N i ∂y NNE ∂N i ∂z NNE ∂N i
J 31 = = ∑ x i , J 32 = = ∑ yi , J 33 = = ∑ zi
∂ζ i =1 ∂ζ ∂ζ i =1 ∂ζ ∂ζ i =1 ∂ζ
(15.12b)

such that ξ=[ξ,η,ζ]T are the parametric coordinates for the

3D parametric element.

The determinant of the Jacobian, Det(J), is an important

parameter in coordinate transformation since that the
volume of the element, V can be expressed as

V = ∫ dΩ = ∫ Det(J (ξ))dΩξ (15.13)

Ω Ωξ

L15-10
 CV7001_L15.docx 24/06/2009

where Ω ξ is the domain of the parent element and

whenever Det(J)≤0, it implies that the mapping is not
unique.

With the definition of the above mapping procedure, it is

now more convenient to evaluate the element stiffness
matrix using the natural coordinate system since all the
shape functions are defined in terms of the natural
coordinates. The element stiffness matrix, Eqn. 12.15a
should now be written as

k = ∫Ω B T EBDet(J (ξ))dΩξ (15.14)

ξ

In order to compute the terms B in Eqn. 15.14, one should

first use Eqn. 15.10b or Eqn. 15.12a to compute the
derivatives of the shape functions with respect to the
global coordinates. In addition, the domain of integration
should also be defined with respect to the parametric
coordinate system. For example, for the 3-node triangle
shown in Fig. 15.3, the domain of integration should
change to

1 1− ξ
k=∫ ∫ BT EBDet(J (ξ))t(ξ, η)dηdξ (15.15)
0 0

Since in 2D case, the elementary volume of dΩ ξ can be

simplified to

dΩ ξ = t(ξ, η)dηdξ (15.16)

L15-11
 CV7001_L15.docx 24/06/2009

where t(ξ, η) is the thickness of the 2D element.

L15-12
 CV7001_L16.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS
Lecture 16

Isoparametric elements in 1D and 2D problems

In the last lecture, an overview was given for the

formulation of isoparametric elements. In this lecture,
details regarding to the constructions of different
isoparametric elements in 1D and 2D problems will be
given.

16.1 Isoparametric elements in 1D

In 1D problems the domain of the physical space is in

general a curve while that for the parametric space is a
straight line [0,1]1 (Fig. 16.1).

1 2 1 2 1 2
3 3 4
ξ=0 ξ= 1 ξ=0 ξ=0.5 ξ=1 ξ=0 ξ= 1/3 ξ=2/3 ξ= 1

(a) Linear element (b) Quadratic element (c) Cubic element

Fig. 16.1 Isoparametric elements in 1D

1
In some cases, the domain of the parametric space may be selected as
[-1,1].

L16-1
 CV7001_L16.docx 24/06/2009

Let ξ be the parent coordinate in the parametric space. In

order to ensure that the shape functions can satisfy the
conditions stated in Section 14.1, the 1D Lagrangian
polynomials are commonly used. For an 1D element with
NNE nodes, the shape function at node i can be taken as
the NNE-1 order Lagrangian polynomial correspond to
that node, LNNE
i
-1
, and is defined as

(ξ - ξ1 )....(ξ - ξi -1 )(ξ - ξi +1 )....(ξ - ξ NNE )

LNNE
i
-1
= (16.1)
(ξi - ξ1 )....(ξi - ξi -1 )(ξi - ξi +1 )....(ξi - ξ NNE )

where ξi, i=1,...,NNE are the parent coordinates at the

nodes such that ξ1=0 and ξ NNE =1.

For example, NNE=2,

(ξ - ξ 2 ) (ξ - ξ1 )
L11 = and L12 = (16.2a)
(ξ1 - ξ 2 ) (ξ2 - ξ1 )

with ξ1=0 and ξ2=1 (16.2b)

we have L11 = 1 - ξ and L12 = ξ (16.2c)

which are the shape functions used in Eqn. 15.2.

The shape functions LNNE i

-1
is C∞ (continuous up to any
order of differentiation) within the element and is C0
across the inter-element boundary. In addition, they also
satisfy the interpolant and conditions. Furthermore, the set
{ LNNE
i
-1
, i=1,...., NNE} is complete up to order NNE-1.

L16-2
 CV7001_L16.docx 24/06/2009

In here it is assumed that for higher order elements (p>1),

the interior nodes of the element are uniformly distributed
within the elements as shown in Fig. 16.1. In general, one
can construct elements with non-uniformly spaced nodes
for special applications but this will not be discussed in
this course.

16.2 Isoparametric elements in 2D

2D isoparametric elements are widely used in many 2D

problems such as plane stress and plan strain formulations
as well as axisymmetic problems. A 2D element can
assume two different shapes: triangle and quadrilateral.

16.2.1 Isoparametric triangles

The isoparametric triangles are also known as complete

triangular family since the shape functions for these
elements contain complete polynomial terms. The domain
of the normalized element for this element family is a right
angle triangle with unit length and straight edges as shown
in Fig. 16.2 2 . In Fig. 16.2, (ξ,η) are the parametric
coordinates of the parent element while the parameter
ζ=1-ξ-η is used for the cyclic expressions of the element
shape functions.

2
For the isoparametric triangles, an alternative method to
construct the shape functions is to use the so-called area
coordinates. For more details, please refer to pages 259-267 of
T1.
L16-3
 CV7001_L16.docx 24/06/2009

η
(0,1)

ξ
(0,0) (1,0)
ζ ζ=1-ξ-η

Fig. 16.2 Parametric domain for isoparametric triangular

element

The details of the coordinate transformation for the linear

T3, the quadratic T6 and the cubic T10 elements are
shown in Fig. 16.3, Fig. 16.4 and Fig. 16.5 respectively.
Tables 16.1 to 16.3 summarize the local node numberings
and the shape functions of these elements.

ζ ξ
1 2

Fig. 16.3 Linear T3 isoparametric triangle

(Constant strain triangle)

L16-4
 CV7001_L16.docx 24/06/2009

Node Ni ξi ηi
1 ζ 0 0
2 ξ 1 0
3 η 0 1

Table 16.1 Shapes function for T3 element

5 4

ζ ξ
1 6 2

Fig. 16.4 Quadratic T6 isoparametric element

Node Ni ξi ηi
1 ζ(2ζ-1) 0 0
2 ξ(2ξ-1) 1 0
3 η(2η-1) 0 1
4 4ξη 1/2 1/2
5 4ηζ 0 1/2
6 4ξζ 1/2 0

Table 16.2 Shape functions for T6 element

L16-5
 CV7001_L16.docx 24/06/2009

8 4

10
5 7

ζ ξ
1 9 6 2

Fig. 16.5 Cubic T10 isoparametric triangle

Node Ni ξi ηi
1 ζ(3ζ-1)(3ζ-2)/2 0 0
2 ξ(3ξ-1)(3ξ-2)/2 1 0
3 η(3η-1)(3η-2)/2 0 1
4 9ξη(3η-1)/2 1/3 2/3
5 9ηζ(3ζ-1)/2 0 1/3
6 9ζξ(3ξ-1)/2 2/3 0
7 9ηξ(3ξ-1)/2 2/3 1/3
8 9ζη(3η-1)/2 0 2/3
9 9ξζ(3ζ-1)/2 1/3 0
10 27ξηζ 1/3 1/3

Table 16.3 Shape functions for T10 element

For an order p isoparametric triangular element, the shape

functions contain complete polynomial terms in the
parametric coordinates up to order p as shown in Fig. 16.6.

L16-6
 CV7001_L16.docx 24/06/2009

1
ξ η T3
ξ 2 ξη η2 T6
ξ 3 ξ 2 η η2 ξ η3 T10

Fig. 16.6 Polynomial terms for isoparametric triangular

elements

The number of node of the element NNE is related to p as

NNE=(p+1)(p+2)/2 (16.3)

Note that the T10 element contains one interior nodes

while the T3 and T6 elements contain only corner and
mid-side nodes.

16.2.2 Lagrangian quadrilaterals

The Lagrangian quadrilateral family is in fact an extension

of the 1D Lagrangian elements to 2D. The domain of the
normalized element for this element family is a square [-
1,1]×[-1,1] with area equal to 4 units as shown in Fig.
16.7. The interpolation functions of the elements in this
family are formed by taking tensor products between
Lagrangian polynomial (with domain of parent coordinate
[-1, 1]) shape functions in the 1D. That is, the shape
function corresponding to the ith node in the ξ direction
and the jth node in the η direction is defined by

Lpi (ξ)Lpj (η) (16.4)

L16-7
 CV7001_L16.docx 24/06/2009

where the relationship between p and NEN is

NNE=(p+1)2 (16.5)

The details of the coordinate transformation for the linear

Q4, the quadratic L9 and the cubic L16 elements are
shown in Fig.16.8, Fig. 16.9 and Fig. 16.10 respectively.
Tables 16.4 to 16.6 summarize the local node numberings
and the shape functions for these elements.

η
(-1,1) (1,1)

(0,0) ξ

(-1,-1) (1,-1)

Fig. 16.7 Parametric domain for Lagrangian quadrilateral

element

η
4 3

1 2
Fig.
16.8 Linear Q4 quadrilateral

L16-8
 CV7001_L16.docx 24/06/2009

Node Ni ξi ηi
1 (1-ξ)(1-η)/4 -1 -1
2 (1+ξ)(1-η)/4 1 -1
3 (1+ξ)(1+η)/4 1 1
4 (1-ξ)(1+η)/4 -1 1

Table 16.4 Shape functions for Q4 element

η
4 7 3

8 ξ
9 6

1 5 2
Fig.
16.9 Quadratic L9 quadrilateral

Node Ni ξi ηi
1 ξη(1-ξ)(1-η)/4 -1 -1
2 -ξη(1+ξ)(1-η)/4 1 -1
3 ξη(1+ξ)(1+η)/4 1 1
4 -ξη(1-ξ)(1+η)/4 -1 1
5 -η(1-ξ2)(1-η)/2 0 -1
6 ξ(1+ξ)(1-η2)/2 1 0
7 η(1-ξ2)(1+η)/2 0 1
8 -ξ(1-ξ)(1-η2)/2 -1 0
9 (1-ξ2)(1-η2) 0 0

Table 16.5 Shape functions for L9 element

L16-9
 CV7001_L16.docx 24/06/2009

η
4 10 9 3

16 15
11 8

ξ
12 13 14 7

1 5 6 2
Fig. 16.10 Cubic L16 quadrilateral

Node Ni ξi ηi
1 L1(ξ)L1(η) -1 -1
2 L4(ξ)L1(η) 1 -1
3 L4(ξ)L4(η) 1 1
4 L1(ξ)L4(η) -1 1
5 L2(ξ)L1(η) -1/3 -1
6 L3(ξ)L1(η) 1/3 -1
7 L4(ξ)L2(η) 1 -1/3
8 L4(ξ)L3(η) 1 1/3
9 L3(ξ)L4(η) 1/3 1
10 L2(ξ)L4(η) -1/3 1
11 L1(ξ)L3(η) -1 1/3
12 L1(ξ)L2(η) -1 -1/3
13 L2(ξ)L2(η) -1/3 -1/3
14 L3(ξ)L2(η) 1/3 -1/3
15 L3(ξ)L3(η) 1/3 1/3
16 L2(ξ)L3(η) -1/3 1/3
L1 (ϕ) = −(1 - ϕ)(1 - 9ϕ 2 )/16, L 2 (ϕ) = 9(1 - ϕ 2 )(1 - 3ϕ)/16
L 3 (ϕ) = 9(1 - ϕ 2 )(1 + 3ϕ)/16, L 4 (ϕ) = −(1 + ϕ)(1 - 9ϕ 2 )/16
Table 16.6 Shape functions for L16 element

L16-10
 CV7001_L16.docx 24/06/2009

For an order p Lagrangian quadrilateral element, the shape

functions contain more terms than a complete polynomial
of order p as shown in Fig. 16.11.

Q4
1
L9
ξ η
L16
2 2
ξ ξη η
ξ3 ξ 2η η2ξ η3
Parasitic terms ξ3η ξ 2η2 η3ξ
ξ3η2 ξ 2η3
ξ3η3

Fig. 16.11 Polynomial terms for Lagrangian quadrilateral

elements

The L9 and the L16 elements contain one and four interior
nodes respectively and they are not connected to any other
node outside the element. These interior nodes contribute
the terms inside the red boxes shown in Fig. 16.11. These
four terms together with the terms ξ3η and ξη3 are called
the parasitic terms (terms inside the blue box) since they
are extra terms that do not increase the complete
polynomial order of the interpolations.

16.2.3 Serendipity quadrilaterals

Since it may be more convenient to make shape functions

dependent only on nodal values placed on the element
boundary, the Serendipity family rather than the

L16-11
 CV7001_L16.docx 24/06/2009

Lagrangian family, is sometimes preferred in finite

element analysis. Serendipity elements are derived by
eliminating all the interior nodes from the corresponding
Lagrangian elements. The quadratic Q8 element is
obtained by eliminating the interior node 9 from the L9
element while the cubic Q12 element is obtained by
eliminating the interior nodes 13, 14, 15 and 16 from the
L16 element. After the elimination procedure, the Q8
element is still complete up to second order terms in the
parametric coordinate and similarly the Q12 element is
still complete up to third order terms.

The details of the coordinate transformation for the

quadratic Q8 and the cubic Q12 elements are shown in
Fig. 16.12 and Fig. 16.13 respectively. Tables 16.7 and
16.8 summarize the local node numbering and the shape
functions used.

η
4 7 3

8 ξ
6

1 5 2

Fig. 16.12 Quadratic Q8 quadrilateral

L16-12
 CV7001_L16.docx 24/06/2009

Node Ni ξi ηi
1 -(1+ξ+η)(1-ξ)(1-η)/4 -1 -1
2 -(1-ξ+η)(1+ξ)(1-η)/4 1 -1
3 -(1-ξ-η)(1+ξ)(1+η)/4 1 1
4 -(1+ξ-η)(1-ξ)(1+η)/4 -1 1
5 (1-ξ2)(1-η)/2 0 -1
6 (1+ξ)(1-η2)/2 1 0
7 (1-ξ2)(1+η)/2 0 1
8 (1-ξ)(1-η2)/2 -1 0

Table 16.7 Shape functions for Q8 element

η
4 10 9 3

11 8

ξ
12 7

1 5 6 2

Fig. 16.13 Cubic Q12 quadrilateral

L16-13
 CV7001_L16.docx 24/06/2009

Node Ni ξi ηi
1 9(1-ξ)(1-η)(ξ2+η2-10/9)/32 -1 -1
2 9(1+ξ)(1-η)(ξ2+η2-10/9)/32 1 -1
3 9(1+ξ)(1+η)(ξ2+η2-10/9)/32 1 1
4 9(1-ξ)(1+η)(ξ2+η2-10/9)/32 -1 1
5 9(1-3ξ)(1-ξ2)(1-η)/32 -1/3 -1
6 9(1+3ξ)(1-ξ2)(1-η)/32 1/3 -1
7 9(1+ξ)(1-3η)(1-η2)/32 1 -1/3
8 9(1+ξ)(1+3η)(1-η2)/32 1 1/3
9 9(1+3ξ)(1-ξ2)(1+η)/32 1/3 1
10 9(1-3ξ)(1-ξ2)(1+η)/32 -1/3 1
11 9(1-ξ)(1+3η)(1-η2)/32 -1 1/3
12 9(1-ξ)(1-3η)(1-η2)/32 -1 -1/3

Table 16.8 Shape functions for Q12 element

After the elimination of the shape functions corresponding

to the interior nodes, the Q12 element will contain only
two parasitic terms as shown in Fig. 16.14.

Q4
1
Q8
ξ η
ξ2 ξη η 2 Q12

ξ 3 ξ 2 η η 2 ξ η3
Parasitic terms ξ 3η η3 ξ

Fig. 16.14 Polynomial terms for

Serendipity quadrilateral elements

L16-14
 CV7001_L16.docx 24/06/2009

Fig. 16.15 Systematic generation of Serendipity shape

functions for Q8 element (Fig. 6.4-2 of T1)

The shape functions for the Serendipity family can be

derived in a systematic way as shown in Fig. 16.15 for the
case of a Q8 element. First of all, it can be seen that for the
mid-side nodes (5,6,7 and 8), the shape functions can be
obtained by the tensor product of the quadratic Lagrangian
shape function in the first direction with the linear
Lagrangian shape function in the second direction (Fig.
16.15a and 16.15b). In order to construct the shape
function for the four corner nodes (1,2,3 and 4), one can
start with the 2D linear Lagrangian shape functions for a
Q4 element and note that the values of the corner shape
functions are equal to 1/2 at the mid-side nodes connected
to them. Thus, in order to enforce that the corner shape
functions will equal to zero at the mid-side nodes, one can

L16-15
 CV7001_L16.docx 24/06/2009

subtract the corner function by one half times the two mid-
side shape functions as shown in Fig. 16.15c. The
resulting functions will then take a value of zero at the two
mid-side nodes. Repeat the above procedure for the other
three corners nodes will give the shapes functions shown
in Table 16.7. The shape functions listed in Table 16.8 for
the Q12 element can also be constructed in a similar
manner.

Remarks
(1) From Figs. 16.3 to 16.5, Figs. 16.8 to 16.10 and Figs.
16.12 and 16.13, one can see that the local numbering
of the nodes in the element follows the sequence of
corner, edge and then interior nodes. Such a
hierarchical numbering procedure will facilitate the
derivation of the element shape functions as well as
leading to a more efficient implementation of the
elements.

(2) A more careful inspection of the shape functions

(Tables 16.1 to 16.8) shows that the basis of them are
all symmetric with respect to the cyclic permutations
of the parent coordinates (ξ,η). This implies that the
element will be invariant with respect to its local node
numbering and orientation of the element.

L16-16
 CV7001_L17.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 17

Techniques of numerical integration

17.1 Numerical integration

In the previous lectures, it is shown that when using

parametric formulation for the construction of element
shape functions, the most natural way to compute the
element stiffness matrix and the element load vector is by
integration over the parametric co-ordinates. Consider the
computation of an element stiffness matrix of the form

k = ∫Ω B T EBDet(J (ξ))dΩ ξ
ξ
(17.1)
T
= ∫Ω (LN ) ELNDet(J (ξ))dΩξ
ξ

Since B will contain differential operator with respect to

∂
the global coordinate system (i.e. it contains terms like
∂x
∂
and ) while the shape functions N are expressed in
∂y
terms of the parent coordinates, coordinate transformation
is needed to compute the terms LN. However, from
Lecture 15, it is shown that

L17-1
 CV7001_L17.docx 24/06/2009

∂ x, y, z −1 ∂ 1 ∂
= J( ) = Ψ (ξ ) (17.2)
∂x ξ, η, ζ ∂ξ Det(J ) ∂ξ

where Ψ(ξ) is polynomial function of ξ. Hence,

1
LN = Λ ( ξ, x ) (17.3)
Det(J )

where Λ (ξ, x) is another polynomial function of ξ and the

global coordinates of nodes of the element. Substitute Eqn.
17.3 into Eqn. 17.1 gives

1 1
k = ∫Ω ( Λ (ξ, x))T E( Λ (ξ, x)) Det(J (ξ))dΩξ
ζ Det(J ) Det(J )
(17.4)
1
= ∫Ω Λ (ξ, x)T EΛ (ξ, x)dΩ ξ
ζ Det( J )

It can be seen from Eqn. 17.3 that the integrand of the

1
element stiffness matrix will contain the term .
Det(J )

For n-dimensional (n=1,2 or 3) problem, if an order p

element is used and q is the order of the highest order
terms in the shape functions such that q≥p1, then it can be

1
q=p when the element used only contains complete polynomial terms. For
example, for a T6 element q=2 corresponding to the terms ξ2, ξη and η2
while for a L9 element q=6 since in the shape function the highest order
term is ξ3η3.

L17-2
 CV7001_L17.docx 24/06/2009

shown that Det(J) is in general a polynomial of ξ of order

equal to

O(n(q-1)) (17.5)

For example, for a T3 element, q=1 and n=2 hence

O(Det(J))= O(0)=constant while for a Q4 element q=2 and
n=2 and therefore O(Det(J))= O(2).

From Eqn. 12.4, it can be seen that the closed form (exact)
solutions of the element stiffness matrices are difficult to
1
obtain due to the present of the rational term . The
Det(J )
only exception is the linear triangular element (T3) in 2D
case and the linear tetrahedral element (T4) in 3D case,
where p=1 and hence O(Det(J))= O(0)=constant. Other
situations for which closed form solution of the element
stiffness matrices exist are when the elements assume
some trivial shapes such as a square or a rectangle or when
all the edges of the element are straight lines.

In order to obtain the element stiffness matrix, numerical

integration is widely used in the finite element method to
obtain an approximation of the exact global stiffness
matrix. Even though by using numerical integration errors
are introduced in the computation of the element stiffness
matrix, it is interesting to know that in solid mechanics,
numerical integration frequently compensates the
discretization error due to the use of infinite dimension
base functions and lead to overall beneficial results.

L17-3
 CV7001_L17.docx 24/06/2009

The basic idea of numerical integration is straightforward.

Suppose that now it is required to integrate a function f(x)
over the domain Ω. The integral is approximated by a
linear combination of the values of the integrand. That is

NIP
∫ f(x)dΩ ≈ ∑ w if(xi ) (17.5)
Ω i =1

where NIP is the number of sampling (or integration or

quadrature) points used and xi are the coordinates of the
ith sampling points inside Ω. wi is the weight for the ith
sample point. Depending on the ways to select the position
of xi and wi, different integration schemes can be derived.

17.2 One-dimensional numerical integration,

Newton-Cotes Quadrature and Gauss
Quadrature

As in one-dimensional case it is always possible to

transform the problem domain into some normalize
intervals (say [0,1] or [-1,1]), it is only necessary to
consider the approximation of the form

1 NIP
I = ∫ f(ξ)dξ ≈ ∑ w if(ξi ) where ξi = [-1,1] (17.6)
-1 i =1

The two most popular ways to carry out the approximation

are namely, the Newton-Cotes Quadrature and the Gauss
Quadrature.

L17-4
 CV7001_L17.docx 24/06/2009

Newton-Cotes Quadrature

In the Newton-Cotes Quadrature, the positions of the

sampling points are determined in a priori at equal
intervals. That is, if NIP=1, ξ1=0, if NIP=2, ξ1=-1 and
ξ2=1 and so on. The values of the weights wi are obtained
by forcing the approximation to be exact for polynomials
up to order NIP-1. For example, if NIP=2, ξ1=-1 and ξ2=1,
the quadrature must be exact when f(ξ) is a constant and a
linear polynomial in ξ. That is

1 1

∫1dξ = w1 + w 2 = 2 and ∫ ξdξ = w1 (-1) + w 2 (1) = 0 (17.7)

-1 -1

Solving Eqn. 17.7 gives w1=w2=1 which gives the well-

known trapezoidal rule. Similarly, by selecting the points
{-1,0,1} as the sampling points, one can obtain the
Simpson one-third rule as

1
I = (f(-1) + 4f(0) + f(1)) (17.8)
3

In general, if NIP sampling points are used, a system of

NIP linear equations will be generated for the
determination of the weights. Newton-Cotes Quadrature
can exactly integrate polynomials up to order NIP-1 or the
error is of order O(hNIP) where h is the biggest interval
size.

L17-5
 CV7001_L17.docx 24/06/2009

Gauss Quadrature

If instead of fixing the positions of the sampling points in

a priori but rather treating them as some undetermined
parameters as the weight wi, Eqn. 17.6 will now has
totally 2NIP unknowns to be determine. These unknowns
can be obtained by forcing the approximation to be exact
for polynomials up to order 2NIP-1. For example, if
NIP=1, then

1 1

∫1dξ = w1 (1) = 2 and ∫ ξdξ = w1ξ1 = 0 (17.9)

-1 -1

Hence, w1=2 and ξ1=0.

Note that the 2NIP equations generated will be nonlinear.

For example, if NIP=2, the equations are

1 1

∫1dξ = w1 (1) + w 2 (1) = 2 and ∫ ξdξ = w1ξ1 + w 2ξ2 = 0

-1 -1
1 1
2
∫ ξ dξ = = and ∫ ξ3dξ = w1ξ13 + w 2ξ32 = 0
2
w1ξ12 + w 2ξ 22
-1 3 -1
(17.10)

and are more difficult to solve. In fact, it can be proved

that the positions of the sampling points are the root of the
Legendre Polynomial (Proof will not be given here). The
solutions for NIP=1 to 4 are listed in Table 17.1.

L17-6
 CV7001_L17.docx 24/06/2009

NIP wi ±ξi
1 2.000 000 000 000 000 0.000 000 000 000 000
2 1.000 000 000 000 000 0.577 350 269 189 626
3 0.555 555 555 555 556 0.774 596 669 241 483
0.888 888 888 888 889 0.000 000 000 000 000
4 0.347 854 845 137 454 0.861 136 311 594 953
0.652 145 154 862 546 0.339 981 043 584 856

Table 17.1 Abscissas and weights of the

Gaussian Quadrature formula

17.3 Integration in 2D

In 2D, the numerical integration rules used will depend on

the shape of the element (rectangular or triangular) under
consideration.

Rectangular regions

In the 2D case the integral will appear in the form

1 1
I = ∫ ∫ f(ξ, η)dξdη (17.11)
-1 −1

and the most easiest way is to first evaluate the inner

integral and keeping η constant. That is

1 NIPξ

∫ f(ξ, η)dξ = ∑ w if(ξi , η) = g(η) (17.12a)

-1 i =1

L17-7
 CV7001_L17.docx 24/06/2009

where NIPξ is the number of integration points in the ξ

direction. The outer integral can be evaluated in a similar
manner.

1 NIPη
I = ∫ g(η)dη = ∑ w jg(η j )
-1 j=1
(17.12b)
NIPη NIPξ NIPη NIPξ
= ∑ w j ∑ w if(ξi , η j ) = ∑ ∑ w i w jf(ξi , η j )
j=1 i =1 j=1 i =1

In practice, Gaussian Quadrature is used in both directions

and it is quite often that NIPξ=NIPη. However, this is not
necessary and on occasion it may be of advantage to use
different numbers of sampling points in each direction.

Obviously, if NIP Gaussian sample points are used in each

direction, the method will be exact for polynomial of
2NIP-1 order in each direction. For example, if 2 × 2
sampling points are used, then the scheme will be exact
for polynomial up to order 3 in each direction. That is, the
product of {1,ξ,ξ2,ξ3} and {1,η,η2,η3} which is complete
up to order 3 even through the product itself contains 16
terms (Fig. 16.11). Thus, an alternative approach by
tackling the problem directly may lead to a more
economical integration scheme. Suppose that one would
like to exactly integrate a third order polynomial in two
directions (totally 10 terms). By tackling the problem
directly, at any sampling point k there are two coordinates
(ξk,ηk) and one weight wk to be determined by the
equation

L17-8
 CV7001_L17.docx 24/06/2009

1 1 NIP
∫ ∫ f(ξ, η)dξdη ≈ ∑ w k f(ξk , ηk ) (17.13)
-1 −1 i =1

From Eqn. 17.13, four points will be sufficient to integrate

exactly a polynomial of order 3 in two directions (10
terms: {1,ξ,η,ξ2,ξη,η2,ξ3,ξ2η,ξη2, η3}). Similarly, it can
be shown that only seven points are sufficient to integrate
a complete fifth order polynomial (21 terms) exactly in
two directions. Some commonly used direct schemes for
rectangle are listed in Table 17.2.

Triangular regions

For triangular elements, as the domain of the parametric

space is a right-angle triangle, the integrals are of the form

1 1− ξ

∫ ∫ f(ξ, η)dξdη (17.14)

0 0

Once again one could use NIP× NIP Gaussian points to

approximate Eqn. 17.14. However, as the limits of the
integration now involve the natural coordinates themselves
and to make the direct use of the Gauss quadrature
formulae, it is required to transform the integration
domain into a square. Thus, it is much more desirable for
both aesthetic and economic considerations to use special
formulae in which no bias is given to any of the natural
coordinates. Such formulae are list in Table 17.3.

L17-9
 CV7001_L17.docx 24/06/2009

Order Number of Coordinates Weight

m* points ξi ηi s
wi
2 3 2/3 0 4/3
− 1/ 6 ± 1/ 2 4/3

2 3 1 1 4/7
-5/9 1/3 27/14
1/3 -5/9 3/2
3 4 ±1 0 2/3
0 ± 1/ 2 4/3

3 4 ± 2/3 0 1
0 ± 2/3 1

5 7 0 0 8/7
0 ± 14 / 15 20/63
± 3/ 5 ± 3/ 5 20/36
5 7 0 0 8/7
r r 25/42
r= 7 / 15 s= -r -r 25/42
(7 + 24 )/15 s -t 25/42
t= (7 − 24 )/15
-s t 5/12
t -s 5/12
-t s 5/12

Table 17.2 Direct numerical integration schemes for

2D rectangular elements
*These schemes are exact up to order m for ξiηj such that
i+j≤m.

L17-10
 CV7001_L17.docx 24/06/2009

Order Number of points Coord Weights

m* inates wi
ξi ηi
1 1 1/3 1/3 1/2

2 3 1/2 1/2
0 1/2 1/6
1/2 0
2 3 1/6 1/6
2/3 1/6 1/6
1/60 2/3
3 4 1/3 1/3 -9/32
1/5 1/5 25/96
3/5 1/5 25/96
1/5 3/5 25/96
4 6 a a 0.111 690 794 839 005
1-2a a 0.111 690 794 839 005
a=0.445 948 490 915 965 a 1-2a 0.111 690 794 839 005
b=0.091 576 213 509 771 b b 0.054 975 871 827 661
1-2b b 0.054 975 871 827 661
b 1-2b 0.054 975 871 827 661

L17-11
 CV7001_L17.docx 24/06/2009

5 7 1/3 1/3 9/80

a a 0.066 197 076 394 252
a=0.470 142 064 105 115 1-2a a 0.066 197 076 394 252
b=0.101 286 507 323 456 a 1-2a 0.066 197 076 394 252
b b 0.062 969 590 272 414
1-2b b 0.062 969 590 272 414
b 1-2b 0.062 969 590 272 414
7 12 a a 0.025 422 453 185 103
1-2a a 0.025 422 453 185 103
a=0.063 089 014 491 502 a 1-2a 0.025 422 453 185 103
b=0.249 286 745 170 910 b b 0.058 393 137 863 189
c=0.310 352 451 033 785 1-2b b 0.058 393 137 863 189
d=0.053 145 049 844 816 b 1-2b 0.058 393 137 863 189
c d 0.041 425 537 809 187
d c 0.041 425 537 809 187
1- c 0.041 425 537 809 187
(c+d) d 0.041 425 537 809 187
1- 1-(c+d) 0.041 425 537 809 187
(c+d) 1-(c+d) 0.041 425 537 809 187
c
d

Table 17.3 Direct numerical integration schemes for 2D triangular elements

*These schemes are exact up to order m for ξiηj such that i+j≤m.
L17-12
 CV7001_L17.docx 24/06/2009

17.4 Appropriate order of numerical

integration

With the use of numerical integration to replace exact

integration, an additional error is introduced into the
approximation and the first impression is that this should
be reduced as much as possible. The cost of numerical
integration could be quite significant and indeed for some
elements it will take up a comparable amount of computer
time to the subsequent solution of the global stiffness
equations. Thus, it will be important to determine (a) the
minimum integration requirement permitting convergence
and (b) the integration requirement need to preserve the
rate of convergence which would result if exact integration
were used. In practice, it is found that the use of higher
orders integration than those actually need will normally
lead to disadvantage. The reason as that inexact
integration quite often cancels out some of the errors due
to the finite element discretization.

17.4.1 Minimum order of integration for convergence

Consider the global stiffness equations (Eqn. 12.14b)

⎡ T ⎤ ⎧ T ⎫ ⎧⎪ T
⎫⎪
⎢ ∫ B EBdΩ ⎥ D = ⎨ ∫ N FbdΩ ⎬ + ⎨ ∫ N Φd ⎬ + P
⎣Ω ⎦ ⎩Ω ⎭ ⎪⎩ΓΦ ⎪⎭
(12.14b)
⎧ ⎫ ⎧ ⎫
+ ⎨ ∫ B T E ε 0 dΩ ⎬ − ⎨ ∫ B T σ 0 dΩ ⎬
⎩Ω ⎭ ⎩Ω ⎭

It is not difficult to see that in order to obtain a non-trivial

solution, the minimum order of the shape function N must

L17-13
 CV7001_L17.docx 24/06/2009

equal to 1 (i.e. linear shape function). In this case, the

integrand for the integral of the stiffness matrix will be of
order zero (i.e. a constant). If it is expressed in the natural
coordinate system

T
∫Ω (LΝ ) E(L Ν)dΩ = ∫ kDet(ξ)dΩ ξ = kV (17.15)
Ωξ

where k is a constant and V is the volume of the element.

Thus, in order to ensure convergence to occur, the
numerical integration scheme used should at least
integrate the volume of the element exactly.

17.4.2 Order of integration for no loss of convergence

rate

For linear elasticity problems, it can be proved that the

total potential energy the finite element solution was exact
up to order 2(p-1) where p is the order of the element
(order of complete polynomial terms with respect to the
global coordinate system). If h is the element size
parameter. Hence, providing the numerical integration for
the element stiffness matrix is exact to the order 2(p-1) or
shows an error of O(h2(p-1)+1)= O(h2p-1), then no loss of
convergence order will occur. Hence, for linear elements
(T3, Q4) the integration scheme should integrate a linear
function exactly and hence one-point integration is
sufficient. For quadratic elements, one has O(h3) and
therefore for 2D elements, 3-point or 2×2 are sufficient.

Note that in finite element literature, terms like reduced

integration, standard integration and over integration are
frequently used to describe a given numerical integration

L17-14
 CV7001_L17.docx 24/06/2009

scheme. By standard integration it means that the

integration scheme can exactly integrate an element in its
undistorted shape (e.g. square for quadrilateral elements).
Reduced integration means that the order of the integration
scheme used is lower (usually one) than the standard one.
Over integration means that an integration scheme higher
than the standard one is used.

17.4.3 Matrix singularity due to numerical integration

As the final outcome of a finite element analysis in linear

problems is always a linear equations system of the form

KD=F (12.15a)

in which the boundary conditions have been applied. Eqn.

12.15a should give an approximated solution of the
problem. If a unique solution exists, the matrix A must be
non-singular. In general, this can be assumed that this is
the case with exact integration. However, with the use of
numerical integration, singular matrix may arise for low
order integration (i.e. insufficient number of integration
points). In general, it is quite easy to identify the situations
for which a singular matrix must arise, but it is much
harder to prove that it will not. Thus, in here, attention will
be paid on the former case.

With the numerical integration the integral is replaced by a

weighed sum of independent linear relations between the
nodal parameters. These linear relations are the only
information form which the stiffness matrix is constructed.
Hence, if the number of unknown exceeds the number of
independent relations supplied at all the integration points

L17-15
 CV7001_L17.docx 24/06/2009

(for the whole mesh), then the stiffness matrix must be

singular. The above argument can be better illustrated by
the following example.

Consider the 2D elasticity problem using the Q4 and the

Q8 elements with one and four-point quadratures
respectively. Here at each integration point three
independent strain relations are used and therefore, the
total number of independent relations equals to 3×number
of integration points (in the whole mesh). The number of
unknowns is, of course, equal to 2×number of nodes less
restrained degrees of freedom. Now consider Fig. 17.1a
and 17.1b which shows a single element and an assembly
of two elements by a minimum of constraints to eliminate
rigid body motion. Simple calculations show that only the
stiffness matrix for the quadratic element assembly will be
non-singular. In Fig. 17.1c, a well-support block of both
kinds of element is considered and here for both element
types non-singular matrices may arise although local, near
singularity may still lead to unsatisfactory results. The
assembly of linear elements with a single integration point
could be singular while the quadratic one will usually be
well behaved. For this reason, linear one-point integrated
elements are seldom used whereas the four-point
quadrature is almost universal for the Q8 element.
Finally, some recommended Gauss numerical integration
schemes for 2D isoparametric quadrilateral elements are
given in Table 17.4.

L17-16
 CV7001_L17.docx 24/06/2009

Element type Integration scheme

Q4 2×2
Q8 7-point or 3×3
L9 3×3
Q12, L16 4×4
T3 1-point
T6 4-point
T10 6-point
Table 17.4 Suggested integration schemes for some
commonly use isoparametric elements.

Only v fixed
(a)
Both u,v fixed

Node u,v
(b)
Integration point

(c)

(a) Q4: 4×2-3=5 > 1×3=3 (singular) Q8: 2×8-3=13 > 4×3=12 (singular)
(b) Q4: 6×2-3=9 > 2×3=6 (singular) Q8: 13×2-3=23 < 8×3=24
(c) Q4: 25×2-18=32 < 16×3=48 Q8: 48×2=96 < 64×3=192
Fig. 17.1 Singular stiffness matrices due to insufficient
number of integration points

L17-17
 CV7001_L18.DOCX 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 18

Plane strain, plane stress and axisymmetric

problems

18.1 Plane strain and plane stress formulations

As mentioned in Lecture 3, the plane strain and plane

stress formulations are particular cases of 3D problems.
These two formulations are the most frequently employed
finite element formulations in structural mechanics
applications. In this lecture, some more details regarding
these two formulations will be given.

From Lecture 3, it is shown tat for both the plane strain

and plane stress problem, only two displacement
components and three stress/strain components are needed
to be considered.

u(x,y)=[u(x,y),v(x,y)]T (3.1a)

[
σ = σ x , σ y , τ xy ]T and ε = [εx , ε y , γ xy ]T and σ=Eε (3.17)

L18-1
 CV7001_L18.DOCX 24/06/2009

In the finite element solution for these problems the

formulae used are the very similar, except that the material
matrix E used are different.

⎡1 ν 0 ⎤
E ⎢ ⎥
Plane stress: E= ν 1 0 (3.16)
1- ν2 ⎢ ⎥
⎢⎣ 0 0 (1 - ν)/2⎥⎦

Plane strain:
⎡ ⎤
⎢ 1 ν (1 - ν ) 0 ⎥
⎢ ⎥
E(1 - ν ) ⎢
E= ν (1 - ν) 1 0 ⎥ (3.20)
(1 + ν)(1 - 2ν ) ⎢ ⎥
⎢ (1 - 2ν) ⎥
⎢ 0 0 ⎥
⎣ 2(1 - ν ) ⎦

For the stiffness matrix and RHS load vector, treatment for
these two types of problem are very simply. Re-call Eqn.
12.15 for the general expression of the K and F.

KD=F (12.15a)

⎡ ⎤
K = ⎢ ∫ B T EBdΩ⎥ (12.15b)
⎣Ω ⎦
and

⎧ T ⎫ ⎧⎪ T ⎫⎪ ⎧ ⎫ ⎧ ⎫
F = ⎨ ∫ N FbdΩ⎬ + ⎨ ∫ N ΦdΓ ⎬ + P + ⎨ ∫ B T Eε 0dΩ⎬ − ⎨ ∫ B T σ 0dΩ ⎬
⎩Ω ⎭ ⎪⎩ΓΦ ⎪⎭ ⎩Ω ⎭ ⎩Ω ⎭

(12.15c)

L18-2
 CV7001_L18.DOCX 24/06/2009

Plane stress problem

For the case of plan stress problem, since elementary

volume dΩ in Eqn. 12.15 can be expressed as

dΩ=tdA (18.1)

where dA is the elementary area of the problem domain.

Hence, Eqns. 12.15b and 12.15c can be written as

⎡ T ⎤
K = ⎢ ∫ B EBtdA ⎥ (18.2a)
⎣A ⎦
and

⎧ T ⎫ ⎧⎪ T ⎫⎪ ⎧ T ⎫ ⎧ T ⎫
F = ⎨ ∫ N Fb tdA ⎬ + ⎨ ∫ N Φ tdS⎬ + P + ⎨ ∫ B Eε 0 tdA ⎬ − ⎨ ∫ B σ 0 tdA ⎬
⎩A ⎭ ⎪⎩SΦ ⎪⎭ ⎩A ⎭ ⎩A ⎭
(18.2b)

where SΦ and dS are the natural boundary curve and the

elementary boundary of the domain area respectively.

Note that in general, t, the thickness may vary inside the

problem domain. However, if t is constant everywhere, it
can be taken out from all the integral and divided from
both sides of Eqn. 12.15a. Eventually,

⎡ ⎤
K = ⎢ ∫ B T EBdA ⎥ (18.3a)
⎣A ⎦
and

L18-3
 CV7001_L18.DOCX 24/06/2009

⎧ T ⎫ ⎧⎪ T ⎫⎪ ⎧ ⎫ ⎧ ⎫
F = ⎨ ∫ N FbdA ⎬ + ⎨ ∫ N ΦdS⎬ + P/t + ⎨ ∫ B T Eε 0dA ⎬ − ⎨ ∫ B T σ 0dA ⎬
⎩A ⎭ ⎪⎩SΦ ⎪⎭ ⎩A ⎭ ⎩A ⎭
(18.3b)

Note that the point loading P is the total loading acting at

the point. Hence the term P/t in Eqn. 18.3b is, in fact,
equal to the loading per unit thickness acting on the
loading point.

Plane strain problem

For plane strain problem, exactly the same treatment can

be applied. In this case, one can simply set t=1 for Eqn.
18.3.

Note that Eqn. 18.3 is the expression for the global

stiffness matrix and the global RHS load vector.
Calculation of the element stiffness matrix and load vector
can be done in exactly the same way by restricting the
integration domain to the area of the element and applying
the numerical integration technique mentioned in Lecture
17.

L18-4
 CV7001_L18.DOCX 24/06/2009

18.1 Rigid body modes and zero energy modes

for 2D elements

Rigid body mode

Consider the four-node Q4 plane stress element shown in

Fig. 18.1.
4 3

1 2

Fig. 18.1 A Q4 plane stress element

If one computes the element stiffness of the element, k,

exactly and expressed it in terms of its eigenvalues λi and
eigenvectors vi, i=1,...,8. That is

⎡λ1 ⎤
⎢ λ2 ⎥
k = [v1 , v 2 ,...., v8 ]⎢ ⎥[v1 , v 2 ,...., v8 ]T = QΛQ T
⎢ ⋅ ⎥
⎢ ⎥
⎣ λ8 ⎦
(18.4)

and arranges the eigenvalues in the order that

λ1≤λ2≤....≤λ8 (18.5)

then one will find that for an unsupported element, the

first three eigenvalues λ1,λ2,λ3 will exactly equal to zero
L18-5
 CV7001_L18.DOCX 24/06/2009

while the first three eignvectors will represent the three

rigid body motions of the elements (Fig. 18.2).

Transition in y direction Transition in x direction Rigid body rotation

Fig. 18.2 Rigid body modes for a Q4 element

These three eigenvectors (displacement vectors) are

commonly known as the rigid body modes of the elements.
The physical meaning of the rigid body modes is that if
the element is not properly constrained (supported), rigid
body motions can take place and the element stiffness
matrix is itself a singular matrix. Obviously, any 2D plane
stress and plane strain element will possess only three
rigid body modes if the element stiffness matrix is
computed exactly.

Zero energy modes for 2D elements

For any 2D finite element stiffness matrix k and any nodal

displacement vector uˆ the product

1 T
û kû (18.6)
2

will equal to the strain energy of the element due to the

displacement field
L18-6
 CV7001_L18.DOCX 24/06/2009

u = Nû (18.7)

where N are the element shape function matrix.

Now suppose we set uˆ =vi, i=1,2,3, i.e. the eigenvectors

that are corresponding with the rigid body modes, then
Eqn. 18.6 will become, by using Eqn. 18.4

1 T 1
v i kvi = v iTQΛQ T v i = 0 (18.8)
2 2

Since vivj=δij (see Eqn. 4.18) and λj=0 for i=1,2 and 3.
This simply confirms that the rigid body modes will not
cause any deformation and strain. Hence, no strain energy
is introduced to the element.

Suppose now for a given element, one can find a non-

trivial (zero) nodal displacement vector uˆ such that û ≠vi
for i=1,2,3 and

1 T
û kû =0 (18.9)
2

then the corresponding displacement field u = Nû is called

the zero energy mode (or the spurious energy mode or the
hourglass mode) of the element since it produces no strain
energy to the element. From Eqn. 18.4, Eqn. 18.9 implies
that there exist an eigenvalue λi, i>3 of k such that

λi=0 (18.10)

and uˆ is the corresponding eigenvector for λi.

L18-7
 CV7001_L18.DOCX 24/06/2009

Thus, a simple way to determine whether an element

possesses any zero energy mode is to compute the number
of zero eigenvalues of the element stiffness matrix. For 2D
elements, if the number of zero eigenvalues for an element
(or a patch of elements) is more than three, it means that
zero energy mode exist for that element and the element
should be used with great care.

Obviously, the existence of zero energy mode is not a

desirable property for a good finite element since in real
situation strain energy is required to deform the element.
Zero energy modes are quite often the results of inexact
numerical integration. Since in numerical integration the
stress strain relationship is only used in a finite number of
sampling points. If it turns out that the number of
integration points is insufficient, it is possible to deform
the element in such a way that no deformation is produced
at those integration points and hence lead to the existence
of zero energy mode as shown in Fig. 18.3 and Fig. 18.4

Fig. 18.3 Zero energy mode for a patch of 4-node elements

L18-8
 CV7001_L18.DOCX 24/06/2009

Fig. 18.4 Zero energy mode for a 8-node/9-node elements

Another common cause of zero energy mode (or near zero

energy node) is the present of extremely different material
properties for two adjacent elements as shown in Fig. 18.5.
Note that in this case the spurious energy mode can
actually propagate from a single element to the whole
patch of elements.

Fig. 18.5 Near mechanisms due to reduced integration.

Note that the zero energy mode problem can usually be

controlled by using a proper integration scheme. For
example, instead of using the reduced 2×2 scheme, a 3×3
scheme can often eliminate the present of zero energy
mode in the L9 elements.

L18-9
 CV7001_L18.DOCX 24/06/2009

18.3 Axisymmetric problem

Axisymmetric stress analysis is also one of the most

frequently encountered formulations in elasticity
problems. In the axisymmetric stress analysis, the domain
of interest is a body of revolution (i.e. with rotationally
symmetric about an axis). Two components of
displacements (u,v) in any plane section of the body along
the axis of symmetry (r=0) will be sufficient to define the
state of strain and stress of the body (Fig. 18.6).

z,v,

r, u

Fig. 18.6 Axisymmetric problem

In Fig. 18.6, r and z denote respectively the radial and

axial coordinates of a point while u and v are the
corresponding displacements and hence u=[u,v]T.

For axisymmetric stress analysis, there are totally four

stress components as shown in Fig. 18.7

L18-10
 CV7001_L18.DOCX 24/06/2009

Fig. 18.7 Stress components for axisymmetric analysis

The displacement-strain relationship can be defined as

⎡∂ ⎤
⎢ ∂r 0⎥
⎡ εr ⎤ ⎢ ∂⎥
⎢ε ⎥ ⎢ 0 ⎥ ⎡u ⎤
⎢ z⎥=⎢ ∂z ⎥ (18.11a)
⎢ γ rz ⎥ ⎢ ∂ ∂ ⎥ ⎢⎣ v ⎥⎦
⎢ ⎥ ⎢ ∂z ∂r ⎥
⎣ εθ ⎦ ⎢ 1 ⎥
⎢ 0⎥
⎣r ⎦

Since the component εθ does not involve differentiation, it

can be taken out from Eqn. 18.11a and re-computed after
the solution u is found. Therefore, Eqn. 18.11a can be
written in a form similar to the cases of plane stress and
plane strain formulation as discussed before (Eqn. 3.5a).
That is

L18-11
 CV7001_L18.DOCX 24/06/2009

⎡∂ ⎤
⎢ ∂r 0⎥
⎢ ∂ ⎥ ⎡u ⎤
ε = Lu = ⎢ 0 ⎥⎢ ⎥ (18.11b)
⎢ ∂z ⎥ ⎣ v ⎦
⎢∂ ∂⎥
⎢⎣ ∂z ∂r ⎥⎦

and the corresponding stress, σ is defined as

σ = Eε (18.12)

For isotropic material, the matrix E is defined by

⎡ ν ⎤
⎢ 1 (1 - ν)
0⎥
⎢ ⎥
E(1 - ν) ⎢ ν
E= 1 0⎥ (18.13)
(1 + ν)(1 - 2ν) ⎢ (1 - ν) ⎥
⎢ 0 0 1⎥
⎢ ⎥
⎣ ⎦

The stress terms σθ can be computed as

E
σθ = εθ (18.14)
2(1 + ν)

Similar to the case of plane stress and plane strain

problems, the initial strains can be written as

⎡ ε r0 ⎤
ε 0 = ⎢ ε z0 ⎥ (18.15)
⎢ ⎥
⎢⎣τrz0 ⎥⎦

L18-12
 CV7001_L18.DOCX 24/06/2009

Follow the usual finite element discretization procedure

and use the stationary principle, the stiffness matrix and
the load vectors are defined by

K = ∫Ω Β T EBrdrdz (18.16a)

F = ∫ N T Fb rdrdz + ∫ N T ΦdΓ + P + ∫ B T Eε 0 rdrdz - ∫ B T σ 0 rdrdz

Ω ΓΦ Ω Ω
(18.16b)

Again, in general, numerical integration is required to

compute the element stiffness matrix and load vectors
terms. For axisymmetric elements, only two rigid body
modes (which two?) are present which are corresponding
to the transitions in the r and z directions.

L18-13
 CV7001_L19.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
FINITE ELEMENT METHODS

Lecture 19

3D problems

The most general finite element for elasticity problems is

the full 3D formulation which embraces clearly all
practical cases. However, when comparing with the 2D
formulations, full 3D formulation is clearly less popular
and less frequently used. The major difficulties hindering
the extensive use of fully 3D finite element analysis are (i)
the huge amount of computational cost involved for the
solution of the large system of linear equations formed
during a full 3D analysis and, (ii) the difficulties of
generating a finite element mesh for general 3D objects.

19.1 3D isoparametric elements

In 3D analysis, there exists a number of possible element

shapes and these include the simplest tetrahedron, the
prism, the cube or brick elements as well as the less
frequently used pyramid element (Fig. 19.1). In this
section, the properties of some commonly used 3D
isoparametric elements will be described.

L19-1
 CV7001_L19.docx 24/06/2009

Fig. 19.1 The pyramid element

The Complete tetrahedral family

The complete tetrahedral family is the simplest 3D

element family. The order p tetrahedral element will
contain complete polynomial terms up to order p. (i.e.
(p+1)(p+2)(p+3)/6 terms). The domain of the parent
element is a right-angle tetrahedron with volume equal to
1/6. The local node numbering for the T4 and the T10
elements are shown in Figs. 19.2 and 19.3 respectively
while the shape functions for these elements are listed in
Tables 19.1 and 19.2.

ζ
Face 4, ξ+η+ζ=1
4

Face 3, ξ=0
Face 1, η=0

2
1
η
ξ Face 2, ζ=0
Fig. 19.2 The T4 element

L19-2
 CV7001_L19.docx 24/06/2009

Node ξ η ζ Ni
1 1 0 0 ξ
2 0 1 0 η
3 0 0 0 ϕ
4 0 0 1 ζ
ϕ=1-ξ-η-ζ
Table 19.1 Shape functions for T4 element.
ζ

10 9
8

7 3 6

2
1 5
η
ξ
Fig. 19.3 The T10 element

Node ξ η ζ Ni
1 1 0 0 ξ(2ξ-1)
2 0 1 0 η(2η-1)
3 0 0 0 ϕ(2ϕ-1)
4 0 0 1 ζ(2ζ-1)
5 0.5 0.5 0 4ξη
6 0 0.5 0 4ηϕ
7 0.5 0 0 4ξϕ
8 0.5 0 0.5 4ξζ
9 0 0.5 0.5 4ηζ
10 0 0 0.5 4ζϕ
ϕ=1-ξ-η-ζ
Table 19.2 Shape functions for the T10 element

L19-3
 CV7001_L19.docx 24/06/2009

(You should now know how to obtain the shape functions

for the next element, the cubic T20 elements in this series.
As an exercise obtain its shape functions yourself.)

The Lagrangian hexahedral family

Elements in the Lagrangian hexahedral family contain

Lagrangian polynomial terms in three-dimension. An
order p element in this family contains (p+1)3 terms. The
domain of the parent element is a cube [-1,1]×[-1,1]×[-1,1]
with volume equal to 8 units. The local node numbering
for the H8 and the H27 elements are shown in Figs. 19.4
and 19.5 respectively while the shape functions are listed
in Tables 19.3 and 19.4.

5 ζ

8
6

1
4
2 η
ξ
3

Fig. 19.4 The H8 element

All these shape functions can be easily derived by forming

tensor products of 1D Lagrangian shape functions.

L19-4
 CV7001_L19.docx 24/06/2009

Node ξ η ζ
1 -1 -1 -1
2 1 -1 -1
3 1 1 -1
4 -1 1 -1
5 -1 -1 1
6 1 -1 1
7 1 1 1
8 -1 1 1
The shape functions can be computed by using the formula:
1
N i = (1 + ξi ξ)(1 + ηi η)(1 + ζ i ζ )
8
Table 19.3 Shape functions for the H8 element
12 25
1 4 13 16

η η
9 21 22 27
11 24

2 3 ζ=-1 14 15 ζ=0
10 23

ξ ξ
20
5 8

η
17 26
19

7 ζ=1
6
18

Fig. 19.5 The H27 element.

The H20 element is formed by the removal of
all face nodes (nodes 21-26) and the interior node 27.

Node ξ η ζ Ni
L19-5
 CV7001_L19.docx 24/06/2009

1 -1 -1 -1 f1(ξ)f1(η)f1(ζ)
2 1 -1 -1 f3(ξ)f1(η)f1(ζ)
3 1 1 -1 f3(ξ)f3(η)f1(ζ)
4 -1 1 -1 f1(ξ)f3(η)f1(ζ)
5 -1 -1 1 f1(ξ)f1(η)f3(ζ)
6 1 -1 1 f3(ξ)f1(η)f3(ζ)
7 1 1 1 f3(ξ)f3(η)f3(ζ)
8 -1 1 1 f1(ξ)f3(η)f3(ζ)
9 0 -1 -1 f2(ξ)f1(η)f1(ζ)
10 1 0 -1 f3(ξ)f2(η)f1(ζ)
11 0 1 -1 f2(ξ)f3(η)f1(ζ)
12 -1 0 -1 f1(ξ)f2(η)f1(ζ)
13 -1 -1 0 f1(ξ)f1(η)f2(ζ)
14 1 -1 0 f3(ξ)f1(η)f2(ζ)
15 1 1 0 f3(ξ)f3(η)f2(ζ)
16 -1 1 0 f1(ξ)f3(η)f2(ζ)
17 0 -1 1 f2(ξ)f1(η)f3(ζ)
18 1 0 1 f3(ξ)f2(η)f3(ζ)
19 0 1 1 f2(ξ)f3(η)f3(ζ)
20 -1 0 1 f1(ξ)f2(η)f3(ζ)
21 0 0 -1 f2(ξ)f2(η)f1(ζ)
22 0 -1 0 f2(ξ)f1(η)f2(ζ)
23 1 0 0 f3(ξ)f2(η)f2(ζ)
24 0 1 0 f2(ξ)f3(η)f2(ζ)
25 -1 0 0 f1(ξ)f2(η)f2(ζ)
26 0 0 1 f2(ξ)f2(η)f3(ζ)
27 0 0 0 f2(ξ)f2(η)f2(ζ)
1 1
f1 (ϕ) = ϕ(ϕ − 1), f 2 (ϕ) = 1 - ϕ 2 , f 3 (ϕ) = ϕ(ϕ + 1)
2 2

Table 19.4 The H27 element

The Serendipity hexahedral family

L19-6
 CV7001_L19.docx 24/06/2009

Serendipity hexahedral elements, the H20 element can be

obtained by eliminating all the face and interior nodes
from the H27 element as shown in Figs. 19.5. Note that
the numbering scheme in Fig. 19.5 is designed in such a
way that the numbering of all face nodes and interior
nodes are greater than that for the corner and the edge
nodes. The shape functions for the H20 element and can
be derived in a similar manner as described in Section
16.2.3.

(As an exercise obtain the shape functions of the H20

elements and the cubic H64 elements.)

The Prism family

The prism family is important in 3D analysis because they

are essential for the connection of the brick elements with
the tetrahedral elements. An order p prism element
contains complete p order polynomial terms in 3D space
((p+1)2(p+2)/2 terms). The domain of the elements in the
parametric space is a right-angle prism. The local node
numbering for the P6 and the P18 elements are shown in
Figs. 19.6 and 19.7 respectively while the shape functions
for these elements are listed in Tables 19.5 and 19.6. The
shape functions can be obtained by forming tensor
products between the complete triangular elements shape
functions with the 1D Lagrangian shape functions.

L19-7
 CV7001_L19.docx 24/06/2009

5
4

3
η
ξ
2
1

Fig. 19.6 The P6 prism element

Node ξ η ζ Ni
1 1 0 -1 ξ(1-ζ)/2
2 0 1 -1 η(1-ζ)/2
3 0 0 -1 (1-ξ-η)(1-ζ)/2
4 1 0 1 ξ(1+ζ)/2
5 0 1 1 η(1+ζ)/2
6 0 0 1 (1-ξ-η)(1+ζ)/2

Table 19.5 Shape functions for the P6 element

ζ

6
12 11
5
4 10
18 15 17
14
13 16
3 η
9 8
ξ
2
1 7 4

Fig. 19.7 The P18 prism element

L19-8
 CV7001_L19.docx 24/06/2009

Node ξ η ζ Ni Node ξ η ζ Ni
1 1 0 -1 f2(ξ,η)g1(ζ) 10 0.5 0.5 1 f4(ξ,η)g3(ζ)
2 0 1 -1 f3(ξ,η)g1(ζ) 11 0 0.5 1 f5(ξ,η)g3(ζ)
3 0 0 -1 f1(ξ,η)g1(ζ) 12 0.5 0 1 f6(ξ,η)g3(ζ)
4 1 0 1 f2(ξ,η)g3(ζ) 13 1 0 0 f2(ξ,η)g2(ζ)
5 0 1 1 f3(ξ,η)g3(ζ) 14 0 1 0 f3(ξ,η)g2(ζ)
6 0 0 1 f1(ξ,η)g3(ζ) 15 0 0 0 f1(ξ,η)g2(ζ)
7 0.5 0.5 -1 f4(ξ,η)g1(ζ) 16 0.5 0.5 0 f4(ξ,η)g2(ζ)
8 0 0.5 -1 f5(ξ,η)g1(ζ) 17 0 0.5 0 f5(ξ,η)g2(ζ)
9 0.5 0 -1 f6(ξ,η)g1(ζ) 18 0.5 0 0 f6(ξ,η)g2(ζ)

f1 (ξ, η) = (1 - ξ - η)(1 - 2ξ - 2η), f 2 (ξ, η) = ξ(2ξ − 1)

f 3 (ξ, η) = η(2η − 1), f 4 (ξ, η) = 4ξη
f 5 (ξ, η) = 4η(1 − ξ - η), f 6 (ξ, η) = 4ξ(1 − ξ - η)
1 1
g1 (ϕ) = ϕ(ϕ − 1), g 2 (ϕ) = 1 - ϕ 2 , g 3 (ϕ) = ϕ(ϕ + 1)
2 2

Table 19.6 Shape functions for the P18 element

L19-9
 CV7001_L19.docx 24/06/2009

19.2 Numerical integration, rigid body and zero

energy modes for 3D elements

For 3D elements, if exact integration is carried out, the

element stiffness matrix will possess exact six rigid body
modes as shown in Fig. 19.8.

x
Transition in z direction Transition in y direction Transition in x direction

Rotation about z axis Rotation about y axis Rotation about x axis

Fig. 19.8 Rigid body modes for 3D elements

Similar to the case of 2D analysis, with the only exception

of the linear T4 elements, numerical integration is
generally required for the evaluation of the element
stiffness matrices and the load vectors. Again, sufficient

L19-10
 CV7001_L19.docx 24/06/2009

numbers of integration points are needed to ensure

convergence and to avoid the present of singular global
stiffness matrix and zero energy modes.

For 3D case, the integration formulae for the hexahedral

elements will be in the form shown in Eqn. 19.1 and can
be integrated using NIP×NIP×NIP points Gaussian
Quadrature in the ξ,η and the ζ directions.

1 1 1 NIPζ NIPη NIPζ

I = ∫ ∫ ∫ f(ξ, η, ζ )dξdηdζ ≈ ∑ ∑ ∑ w i w jw k f(ξi , η j , ζ k ) (19.1)
-1 −1−1 i =1 j=1 k =1

Instead of using NIP×NIP×NIP points Gaussian

Quadrature in the ξ, η and the ζ directions, direct
integration of the form

1 1 1 NIP
∫ ∫ ∫ f(ξ, η, ζ )dξdηdζ ≈ ∑ w k f(ξk , ηk , ζ k ) (19.2)
-1 -1 −1 i =1

could be used to reduce the computational cost of the

numerical integration procedure in 3D problems. Table
19.7 shows some of the commonly used direct integration
schemes for a cube. For example, in order to exactly
integrate a fifth order polynomial exactly, the direct
method only required 14 sample points which is nearly
one half of the 3×3×3 Gaussian quadrature scheme.

For tetrahedral elements, as the domain of the parametric

space is a right angle tetrahedron, the integrals are of the
form

L19-11
 CV7001_L19.docx 24/06/2009

1 1− ξ 1- ξ - η

∫ ∫ ∫ f(ξ, ηζ )dξdηdζ (19.3)

0 0 0

Once again one could use NIP×NIP×NIP Gaussian points

for integration. However, as in the 2D case, direct
integration formulae are preferred and they are shown in
Table 19.8.

L19-12
 CV7001_L19.docx 24/06/2009

Order Number of points Coordinates Weights

m* ξi ηi ζi wi
2 4 0 ± 2/3 − 1/ 3
± 2/3 0 1/ 3 2
3 6 1/ 6 ± 1/ 2 − 1/ 3
− 1/ 6 ± 1/ 2 1/ 3
− 2/3 0 − 1/ 3 4/3
2/3 0 1/ 3
3 6 ±1 0 0
0 ±1 0 4/3
0 0 ±1
5 14 ±a 0 0 320/361
0 ±a 0 320/361
a= 19 / 30 0 0 ±a 320/361
b= 19 / 33 ±b ±b ±b 121/361
7 34 ±a 0 0 0.295 747 599 451 303
0 ±a 0 0.295 747 599 451 303
a=0.925 820 099 772 552 0 0 ±a 0.295 747 599 451 303
b=0.330 814 963 699 288 ±a ±a 0 0.094 101 508 916 324
c= 0.734 112 528 752 115 0 ±a ±a 0.094 101 508 916 324
±a 0 ±a 0.094 101 508 916 324
±b ±b ±b 0.412 333 862 271 436
±c ±c ±c 0.224 703 174 765 601
Table 19.7 Direct numerical integration schemes for 3D hexahedral elements
*These schemes are exact up to order m for ξiηjζk such that i+j+k≤m.
L19-13
 CV7001_L19.docx 24/06/2009

Order Number of points Coordinates Weights

m* ξi ηi ζi wi
1 1 1/4 1/4 1/4 1/6
2 4 a a a
a= (5 − 5 )/20 a a b
b= (5 + 3 5 )/20 a b a 1/24
b a a
3 5 a a a -2/15
a=1/4 b b b 3/40
b=1/6 b b c 3/40
b c b 3/40
c b b 3/40
Table 19.8 Direct numerical integration schemes for 3D tetrahedral elements
*These schemes are exact up to order m for ξiηjζk such that i+j+k≤m.

L19-14
 CV7001_L19.docx 24/06/2009

Order Number of points Coordinates Weights

m* ξi ηi ζi wi
5 15 a a a 8/405

a=1/4 b1 b1 b1 0.011 511 367 104 540

b1 b1 c1 0.011 511 367 104 540
b1= (7 + 15 )/34 b1 c1 b1 0.011 511 367 104 540
c1= (13 − 3 15 )/34 c1 b1 b1 0.011 511 367 104 540

b2= (7 − 15 )/34 b2 b2 b2 0.011 989 513 963 170

c2= (13 + 3 15 )/34 b2 b2 c2 0.011 989 513 963 170
b2 c2 b2 0.011 989 513 963 170
d= (5 − 15 )/20
c2 b2 b2 0.011 989 513 963 170
e= (5 + 15 )/20
d d e
d e d
e d d 5/567
d e e
e d e
e e d
Table 19.8 (Continued) Direct numerical integration schemes for 3D tetrahedral elements
*These schemes are exact up to order m for ξiηjζk such that i+j+k≤m.

L19-15
 CV7001_L20.docx 24/06/2009

Nanyang Technological University

School of Civil and Environmental Engineering

CV7001
NUMERICAL STRUCTURAL ANALYSIS

Lecture 20

Finite element Modeling Technique

20.1 Introduction

In this lecture, some basic techniques for finite element

modeling will be discussed. Modeling is an art based on
the ability to visualize physical interactions. All basic and
applied knowledge of physical problems, finite element,
and solution algorithms contributes to model expertise.
Hence, only guidelines will be given here. In fact, there is
no golden rule for modeling except the following two
statements:

"All models are wrong, but some are useful" George Box

"Garbage in, Garbage out!"

L20-1
 CV7001_L20.docx 24/06/2009

20.2 General aspect for using commercial

package

• Many many commercial software packages have been

written and marketed. Initial, many of them are
originated from structural engineering.

• The "finite element methods" have been applied to

many different areas of engineering applications and no
longer limited to structural engineering.

• Large scale structural (dynamic and non-linear) analysis

of buildings. (Figs. 20.1 and 20.2) and bridges.

Fig. 20.1 Seismic assessment of buildings

L20-2
 CV7001_L20.docx 24/06/2009

Fig. 20.2 Analysis of bridge deck.

• Large scale complex impact analysis of vehicles (Fig.

20.3).

Fig. 20.3 Impact analysis of vehicle

L20-3
 CV7001_L20.docx 24/06/2009

• Themo-mechanical analysis of car engine (Fig. 20.4).

Fig. 20.4 Themo-mechanical analysis of car engine.

• Large scale computational fluid dynamics (CFD)

analysis of space shuttle (Fig. 20.5).

Fig. 20.5 CFD analysis of space shuttle.

L20-4
 CV7001_L20.docx 24/06/2009

• Biomechanics analysis of human blood vessels (Fig.

20.6).

Fig. 20.6 Analysis of human blood vessels.

• No matter what kind of software package you are using,

understanding of the physical actions of your model is
the most important part of modeling.

• Computer software (or computer) are only as smart as

the user!

• Understanding of the capacity/limitation of the software

is also important

L20-5
 CV7001_L20.docx 24/06/2009

• A well-developed software package normal has three

"modules"
1. Pre-processing (geometrical modelling, mesh
generation, loading and boundary conditions)

2. Solver (finite element analysis)

3. Post-processing (displacements and stresses plots,

deformation, animations, data exchange files (DXF)
etc)

• Always read the manual before you use the programs

• Always check your input data

• Always ask yourself whether the results make sense or

not

• MANY well-known examples of failure of buildings are

due to wrong interpretation/misuse of computer
software!

L20-6
 CV7001_L20.docx 24/06/2009

20.3 Cost and Dimensionality

• What model to use 1D, 2D or 3D?

• Cost increase with higher dimensions but more details

are available.

Let N be the number of elements in the model.

For 1D => Number of dofs = N
Semi-bandwidth=> Constant
Solution time => = c2×N = O(N)
(Solution time ∝ semi-banwidth2×Number of dofs)

For 2D => Number dofs = N2

Semi-bandwidth=> N
Solution time => N2×N2=O(N4)

For 2D => Numb dofs = N3

Semi-bandwidth=> N2
Solution time => N4×N3=O(N7)

So ratios of complexities of the problems

N:N4:N7 (20.1)

• As a rule-of-thumb, if you use a PC (say 3GHz+ CPU

with 4Gb of RAM), all 1D and most 2D problems can
be solved but not for large models of 3D problems.

L20-7
 CV7001_L20.docx 24/06/2009

20.4 Which element to use?

• Performance of elements is normally problem

dependent.

• Lagrangian elements are more accurate than Serendipity

elements.

• Quadrilaterals (if not distorted) are better than triangles

(Why?)

• High order elements are more accurate than low order

elements.

• But one must always remember the cost-benefit effect

• Normally, quadratic elements (L9, T6, T10, H27) are

the best and linear elements are not suggested to use!

20.5 Creation of Model

• Always start with simple (but not too simple!) model

that will give you desired answer.

• Always keep in mind what you wish to obtain

• What kind of mathematics model you need to use?

• What are the key unknowns you wish to find out?

• What are the key components of your structures?

L20-8
 CV7001_L20.docx 24/06/2009

• How accurate you wish your results will be?

• How much time you have for solving the problem?

• Remember that in reality, you need to compromise the

amount of details you included in the model with the
time/resources you have (Figs 20.7 and 20.8)

• What is the mesh density you need to use to model (i)

the geometry and (ii) the solution? Is there any high
stress gradient region in the model?

Fig. 20.7 Selection of model

Fig. 20.8 Simplification of models

L20-9
 CV7001_L20.docx 24/06/2009

20.6 Support and loading conditions

• The support conditions can largely change the results of

the model (Fig. 20.9).

• To many beginners => can easily make (fatal!) errors

especially without good understanding of the nature of
the problem.

• Use of point loading can sometime lead to convergence

problems.

Fig. 20.9 Modelling of supports

20.7 Element shapes, Connection and Grading

• Elements perform best when undistorted and regular

(square, equilateral triangles).

• Mesh performs best when it has good grading of

elements sizes according to the exact solutions.

• Difficult to set out universal rule in general. But some

"obvious" good practice can be highlighted (Fig. 20.10).

L20-10
 CV7001_L20.docx 24/06/2009

• Avoid elements with large aspect ratio or largely

distorted elements.
• Avoid elements with highly curved edges/displacement
edge/interior nodes.

Fig. 20.10

• Placement and alignment of elements are also

important. Ensure the conformal property the elements
(Fig. 20.11)

Fig. 20.11 Alignment of elements

• When graded mesh is need, pay attention on how to

achieve a good grading while causing minimum
distortion to the elements in the mesh (Fig. 20.12).
L20-11
 CV7001_L20.docx 24/06/2009

Fig. 20.12 Grading of elements

20.8 Checking and refinement

• Always check your input data!

• Good software normally has some checking procedures

for input data but always NOT 100% fool-proved.

• Always check your results carefully. The pre-processing

modules are designed to help you to do this.

• Always a good idea to compare your results with some

simplified models, if possible.

• For stress calculation, always evaluate the stress at the

Gaussian points of the elements.

• Not only check the results at only one or a few points,

results at there may be misleading. Good to plot out the
global solutions using the post-processing module of the
software (Fig. 20.13)

L20-12
 CV7001_L20.docx 24/06/2009

• Refinement will help you to identify the problems. Ask

yourself why convergence or divergence occurs.

Fig. 20.13 Stress contour

20.9 Ill-conditioning, Locking and Instability

• These are problems caused by misuse of numerical

technique (i.e. your model is correctly but the results are
not good/no results due to wrong computational
approaches).

• Quite difficult to list all the possibilities but some

common sources can be highlighted.

• Great stiffness discrepancies between adjacent elements

• Poor choices (to low or to high) of numerical

integration rules

• Possion ratio near 0.5 for plane strain or 3D problems

(Why?)

L20-13
 CV7001_L20.docx 24/06/2009

• Support "stiff" elements by flexible elements.

• Let 3D elements become very thin => plate

20.10 Additional resources

The most convenient, cheapest additional resources can be

obtained form, of course, the Internet.

If one tries to search for the keyword "Finite Element

Methods" on the Internet, one will obtain thousands of
websites. There is no way to quote all these websites.
However, the following two sites are good starting points.

The International Association for the Engineering

Analysis Community (NAFEMS)

"To promote the safe and reliable use of finite element and
related technology"

Website: http://www.nafems.org

Also publishes the BENCHmark magazine for general

educational and latest news in finite element modelling.

Some useful articles:

"Building Better Products with Finite Element Analysis

(Common misconceptions about FEA)" by Vince Adams
and Abraham Askenazi, January 1999

"Back to Basics: The effect of a Point Load", By Dr. John

Smart, January 1999

L20-14
 CV7001_L20.docx 24/06/2009

The Internet Finite Element Resources page (from Roger

Young and Ian McPhedron) describes and provides access
to FE software via the Internet.

http://www.engr.usask.ca/~macphed/finite/fe_resources/fe
_resources.html

L20-15