You are on page 1of 279

3D7 - Finite Element Methods 2016/17

Chapter 1

Introduction to Two and Three Dimensional FE

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 2

1.1 Background

Many physical phenomena in engineering and science can be described


in terms of partial differential equations. In general, solving these equa-
tions by classical analytical methods for arbitrary shapes is impossible.
Numerical methods are hence employed to solve these partial differen-
tial equations approximately with the help of computers. Engineering
problems are such as stress analysis, heat transfer, fluid flow and elec-
tromagnetics, to name a few.

Consider a plate with a hole for which we wish to find the temperature
distribution. It is straightforward to write a heat balance equation for
each point in the plate. However, the solution of the resulting PDE
for a complicated geometry, such as an engine block, is impossible by
February 8, 2018 3D7 - Finite Element Methods 2016/17 3

classic methods like separation of variables. The basic idea of FEM


is to divide the body into finite elements, connected by nodes, and
obtain an approximate values on these nodes by solving a system of
linear equations; the number of unknowns (which are the nodal tem-
peratures) is equal to the number of nodes. To obtain a reasonable
accurate solution, thousand of nodes are usually needed, so computers
are essential for solving these equations. Generally, the accuracy of the
solution improves as the number of elements (and nodes) increases, but
so does the computer time, and hence the cost.
February 8, 2018 3D7 - Finite Element Methods 2016/17 4

Geometry, loads and finite element meshes.


February 8, 2018 3D7 - Finite Element Methods 2016/17 5

1.2 Course outline

This part of the module consists of eight lectures and will focus on
finite elements for heat conduction and elasticity in two and three
dimensions. More specifically, you will learn

• Strong and weak formulations for heat conduction


• Shape functions for two and three dimensional elements
• Isoparametric mapping and numerical integration
• Application of boundary conditions
• Assembly of element matrices and vectors
• Generalisation to elasticity
February 8, 2018 3D7 - Finite Element Methods 2016/17 6

1.3 The big picture

One important task for engineers is to create mathematical models of


physical systems and to make predictions about the systems’ behaviour.

The models are based on physical laws, such as energy conservation or


Newton’s second law, etc., and are described in form of partial differen-
tial equations. These equations are often very complex and are posed
on geometrically complicated regions.

Over the last three decades or so, it has become possible to solve increas-
ingly complex engineering models using computers. Many numerical
methods enable us to create computational models in order to solve
engineering models approximately. Popular methods include
February 8, 2018 3D7 - Finite Element Methods 2016/17 7

• Finite Difference Methods (solving differential equations by ap-


proximating them with difference equations, in which finite differ-
ences approximate the derivatives). For example,
df f (t + ∆t) − f (t)

dt ∆t
• Finite Volume Methods (based on conservation laws. Volume in-
tegrals in a partial differential equation that contain a divergence
term are converted to surface integrals, popular in Fluid Mechan-
ics, especially with hyperbolic equations).
Z Z
∇ · qdV = qdS
Ω ∂Ω
The method has the important physical property that certain con-
servation laws are maintained.
February 8, 2018 3D7 - Finite Element Methods 2016/17 8

• Finite Element Methods (popular in Solid Mechanics, good for el-


liptical and parabolic equations, less successful with hyperbolic
equations). The fundamental idea is the replacement of continuous
functions by piecewise approximations, usually polynomials.
• Spectral Methods (high order of accuracy).
• Boundary Integral Methods (solving linear partial differential equa-
tions which have been formulated as integral equations (i.e. in
boundary integral form). It can be applied in many areas of engi-
neering and science including fluid mechanics, acoustics, electromag-
netics, and fracture mechanics).
• Lattice Boltzmann Methods (discrete Boltzmann equation is solved
to simulate the flow of a Newtonian fluid with collision models).
February 8, 2018 3D7 - Finite Element Methods 2016/17 9

Flowchart of computational problem solving.

The first step of a numerical method is to divide the domain of a PDE


into small cells. The set of these small cells is called a mesh.

Structured meshes are identified by regular connectivity. The possible


cell choices are quadrilateral in 2D and hexahedra in 3D. This type of
mesh is highly space efficient, i.e. since the neighborhood relationships
are defined by storage arrangement. But it is difficult to generate high
quality of mesh for industrial problems due to the highly complicated
February 8, 2018 3D7 - Finite Element Methods 2016/17 10

geometry involved in general.

Unstructured meshes are identified by irregular connectivity. It cannot


easily be expressed as a two-dimensional or three-dimensional array in
computer memory. These meshes typically employ triangles in 2D and
tetrahedra in 3D. Compared to structured meshes, this model can be
highly space inefficient since it calls for explicit storage of neighborhood
relationships. But they are more flexible and can fit the boundaries of
complicated geometry hence the treatment of boundary conditions can
be very simple and accurate.
February 8, 2018 3D7 - Finite Element Methods 2016/17 11

Types of mesh.
February 8, 2018 3D7 - Finite Element Methods 2016/17 12

Other important issues related to numerical methods are accuracy,


stability, adaptivity, and parallelization, which are beyond the scope
of this course.

The rise of the numerical methods starting end of 1960s is directly


linked to the rise of computers. To obtain a reasonably accurate solu-
tion, equations with thousands of unknowns are necessary. On today’s
computers it is possible to solve problems with about 106 equations, and
in special applications with 109 equations. The increase of the computer
speed is dictated by the Moore’s law, which says that the number of
transistors on integrated circuits doubles about every two years.

Colossus was a set of computers developed by British code-breakers


in 1943-1945 to help in the cryptanalysis of the Lorenz cipher. It used
February 8, 2018 3D7 - Finite Element Methods 2016/17 13

vacuum tubes to perform Boolean and counting operations, and is thus


regarded as the world’s first programmable, electronic, digital computer.
As of June 2016, ranked number one in the TOP500 list as the fastest
supercomputer in the world is the Chinese supercomputer Sunway Tai-
hulight, with a LINPACK benchmark rating of 93 petaflops (1015 flops).
February 8, 2018 3D7 - Finite Element Methods 2016/17 14

Evolution of computer performance in Mflops (106 floating point


operations per second).
February 8, 2018 3D7 - Finite Element Methods 2016/17 15

1.4 Why learn finite elements?

• The knowledge of how the method works will enhance your analysis
skills and provide a better understanding of numerical analysis in
general.
• The intelligent use of commercial finite element software and the
correct interpretation of the output requires a basic knowledge of
the method. (You probably heard the notion of garbage in, garbage
out.)
• It is possible to extend commercial software packages by imple-
menting user defined subroutines. Understanding the finite element
method and basic programming can help you to extend commercial
software.
February 8, 2018 3D7 - Finite Element Methods 2016/17 16

• The finite element method is very cost effective and fast compared to
physical testing. It enables to study the influence of various design
choices (geometry, material parameters or loads) and helps to devise
better designs.
• Unlike the finite difference and finite volume methods, the finite
element method works very well on unstructured meshes, making it
a method very suitable for real industrial problems.

In engineering, the finite element method is a widely used approach for


obtaining approximate solutions to complex problems. Many industries
such as civil, automobile and aerospace are moving towards fully virtual
product development of which numerical simulation using the finite
element method is the cornerstone.
February 8, 2018 3D7 - Finite Element Methods 2016/17 17

1.5 Sample finite element applications

The range of applications of finite elements is too large to list. Just to


give you an idea I have included three snapshots.

Structural engineering: Stress distribution in a bridge cross


section (Picture courtesy of DIANA).
February 8, 2018 3D7 - Finite Element Methods 2016/17 18

Mechanical engineering: Virtual crash test of a van (Picture


courtesy of LS-Dyna).
February 8, 2018 3D7 - Finite Element Methods 2016/17 19

Biomedical engineering: Fluid-structure coupled simulation of a


three leaflet valve.
February 8, 2018 3D7 - Finite Element Methods 2016/17 20

1.6 Proprietary Softwares

• Nastran (NASA STRucture ANalysis) is a finite element analysis


(FEA) program that was originally developed for NASA. NASTRAN
source code is integrated in a number of different software packages,
which are distributed by a range of companies.
• ANSYS (ANalysis SYStems) was founded in 1970. Its primary pur-
pose was to develop and market finite element analysis software
for structural physics that could simulate static (stationary), dy-
namic (moving) and thermal (heat transfer) problems. Ansys now
publishes engineering analysis software across a range of disciplines
including finite element analysis, structural analysis, computational
fluid dynamics, explicit and implicit methods, and heat transfer.
February 8, 2018 3D7 - Finite Element Methods 2016/17 21

• ABAQUS is a software suite for finite element analysis and computer-


aided engineering, originally released in 1978. The name and logo
of this software are based on the abacus calculation tool. It was
acquired by Dassault Systèmes in 2005.
February 8, 2018 3D7 - Finite Element Methods 2016/17 22

1.7 Open Source Softwares

Popular softwares in the public domain:

• FEniCS provides scientific computing tools for working with compu-


tational meshes, finite element variational formulations of ordinary
and partial differential equations, and numerical linear algebra. One
of the main developer is Dr. Garth Wells.
• OpenFoam is a C++ toolbox for the development of customized
numerical solvers for the solution of continuum mechanics problems,
including computational fluid dynamics (CFD).
• Gerris is a Free Software for the solution of the PDEs describing
fluid flows under the Free Software GPL license.
February 8, 2018 3D7 - Finite Element Methods 2016/17 23

1.8 Literature

Meanwhile there are over five hundred books on finite elements. I mainly
used the following three books for preparing this handout.

• J Fish and T Belytschko, A First Course in Finite Elements, John


Wiley & Sons, 2007.
• JN Reddy, An Introduction to the Finite Element Method, McGraw-
Hill, 2005.
• RD Cook, Finite Element Modeling for Stress Analysis, John Wiley
& Sons, 1995.
3D7 - Finite Element Methods 2016/17

Chapter 2

Strong and weak forms for single variable problems

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 25

In the next few chapters we will retrace the basic steps introduced for
one-dimensional problems. First we will consider two dimensional prob-
lems for which the unknown variable is a scalar, such as temperature.
Specifically, we will consider the heat equation as a representative prob-
lem.
February 8, 2018 3D7 - Finite Element Methods 2016/17 26

2.1 Review of Vector Calculus

In 2D variables such as heat flux or displacements are vectors. Re-


member that vectors are physical quantities that have a magnitude and
direction.

A vector is expressed in terms of its components and the unit vectors


in the x- and y-directions.

q = qx i + qy j

where qx is the x-component and qy is the y-component and i and j


are basis vectors. Alternatively in matrix notation a vector is written
February 8, 2018 3D7 - Finite Element Methods 2016/17 27

as a column vector  
qx
q=
qy

Vector q and its components.


February 8, 2018 3D7 - Finite Element Methods 2016/17 28

A frequently occurring vector operation is the scalar product


 
 rx
q · r = q T r = qx

qy = qxrx + qy ry
ry

A useful operator in connection with vectors is the del operator (repre-


sented by the symbol ∇).
 

 ∂x 
∇= 
∂
∂y

If the del operator acts on a scalar field, say temperature T (x, y), it
February 8, 2018 3D7 - Finite Element Methods 2016/17 29

produces a vector that points in the direction of the steepest slope.


 
∂T
 ∂x 
∇T =  
 ∂T 
∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 30

Contour map for a terrain (left) and the associated


three-dimensional model (right). If T is interpreted as the height,
the vector ∇T points in the direction of the steepest slope.
February 8, 2018 3D7 - Finite Element Methods 2016/17 31

The scalar product of the del operator with a vector field q gives the
divergence
∂qx ∂qy
div q = ∇ · q = +
∂x ∂y

Notice the divergence of a vector field is a scalar.

Finally, we can combine the del and divergence operator to form the
Laplace operator

∂ 2T ∂ 2T
∇2T = ∇ · ∇T = 2
+ 2
∂x ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 32

2.1.1 Divergence Theorem

The two-dimensional equations will be derived for a body of arbitrary


shape. If we mean arbitrary shape, we draw in two dimensions some-
thing which resembles a potato.

One-dimensional and two-dimensional domains. The domain is


denoted with Ω and the boundary with Γ.
February 8, 2018 3D7 - Finite Element Methods 2016/17 33

First, we look at the definition of integration in one dimension. For a


domain Ω from x = 0 to x = l we have
Z
dT
dx = T |x=l − T |x=0 = (T n)|Γ in one dimension

dx

Here n is the unit normal to the one dimensional domain. The unit
normal always points away from the domain of interest and has the
length 1.
February 8, 2018 3D7 - Finite Element Methods 2016/17 34

Two-dimensional integration.

The preceding integration formula is generalised to two dimensions by


February 8, 2018 3D7 - Finite Element Methods 2016/17 35

considering each coordinate direction independently


Z Z d Z xr  Z d
∂T ∂T
dΩ = dx dy = T (xr , y) − T (xl , y)dy

∂x c xl
∂x c

I I
= T dy = T nx dΓ (2.1)
Γ Γ

Z I I
∂T
dΩ = T dx = T ny dΓ (2.2)

∂y Γ Γ

Using these two equations we develop the divergence theorem. To that


February 8, 2018 3D7 - Finite Element Methods 2016/17 36

purpose we consider a vector


 
qx
q=
qy
We first insert the qx component into equation 2.1 and the qy component
into equation 2.2 and then sum the equations.
Z Z I I
∂qx ∂qy
dΩ + dΩ = qxnx dΓ + qy ny dΓ

∂x Ω
∂y Γ Γ
Z   I  
∂ ∂
 qx   nx
∂x ∂y dΩ = qx qy dΓ
Ω qy ny
Z IΓ
∇ · q dΩ = q · n dΓ
Ω Γ
February 8, 2018 3D7 - Finite Element Methods 2016/17 37

Z I
div q dΩ = q · n dΓ
Ω Γ
The divergence theorem has been derived in Part I Vector Calculus and
has been used in Fluid Mechanics.
February 8, 2018 3D7 - Finite Element Methods 2016/17 38

Illustrative Example We consider the divergence theorem on a tri-


angular domain

Triangular problem domain.


February 8, 2018 3D7 - Finite Element Methods 2016/17 39

with a given vector field q = [c 0]T , where c is a constant. The


divergence of q is
Z
∂c ∂0
div q = + =0 ⇒ div q dΩ = 0
∂x ∂y Ω

Next, we look at the integral over the boundary


I Z Z Z
q · n dΓ = q · n dΓ + q · n dΓ + q · n dΓ
AB BC CA
" #
√1
Z Z Z
  5
= 0 dΓ + c 0 dΓ + −c dΓ
√2
AB BC 5 CA
c √
= √ 5−c·1=0
5
February 8, 2018 3D7 - Finite Element Methods 2016/17 40

2.2 Strong Form of Heat Equation

We consider a plate of unit thickness in steady state condition (i.e.,


the temperature is not dependent on time). The plate contains a heat
source s(x, y) (energy per unit volume and time).
February 8, 2018 3D7 - Finite Element Methods 2016/17 41

To derive the strong form we look at the energy balance for an in-
finitesimal control volume.

Problem definition and heat fluxes in and out of an infinitesimal


control volume.
February 8, 2018 3D7 - Finite Element Methods 2016/17 42

The energy balance requires that the heat flux q (energy per unit area
and time) through the boundaries equals the heat generated s.
∂qx ∂qy
   
qxdy− qx + dx dy+qy dx− qy + dy dx+s(x, y)dxdy = 0
∂x ∂y
Dividing by dxdy, the energy balance equation is written as
∂qx ∂qy
+ −s=0 (or, div q − s = 0)
∂x ∂y

The first two terms are the divergence of the heat flux, which is the
heat flowing out from the point. This is equal to the heat source s.

Next, recall Fourier’s law in one dimension


dT
q = −k
dx
February 8, 2018 3D7 - Finite Element Methods 2016/17 43

Fourier’s law can be generalised for an isotropic material (such as steel


or concrete) to two dimensions with
 
∂T
    ∂x
qx k 0  
=−  
qy 0 k  ∂T 
∂y
The energy balance equation combined with Fourier’s law leads to
∂ ∂T ∂ ∂T
   
k + k +s=0
∂x ∂x ∂y ∂y

If the heat conductivity is constant throughout the domain, this gives


 
∂ 2T ∂ 2T
k 2
+ 2 + s = k∇2T + s = 0 (2.3)
∂x ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 44

For an anisotropic material, such as wood or fibre composite, Fourier’s


law is generalised to
 
∂T
    ∂x
qx kxx kxy  
=−   (or, q = −D∇T )
qy kyx kyy  ∂T 
|{z} | {z }
q D ∂y }
| {z
∇T
The physical implication of this equation is that, for example, a given
temperature gradient in the x-direction leads to a heat flux in x as well
as y directions.
February 8, 2018 3D7 - Finite Element Methods 2016/17 45

Finally, the heat equation for a problem with an anisotropic and non-
constant heat conductivity is written in matrix notation as:

∇·(D∇T ) + s = 0
February 8, 2018 3D7 - Finite Element Methods 2016/17 46

Boundary conditions The partial differential equation describes the


heat flow in the domain and to solve it we need boundary conditions.
There are two different types of boundary conditions possible; either
the temperature T or the heat flux q can be prescribed.

Two possible types of boundary conditions.


February 8, 2018 3D7 - Finite Element Methods 2016/17 47

On the part of boundary ΓT temperature T̄ is prescribed and on the


part of boundary Γq the normal flux q̄ is prescribed.

T = T̄

ΓT
(q · n) = q̄

Γq

Using Fourier’s law the heat flux boundary condition can be written in
terms of the temperature gradient.

(q · n) = −k∇T · n = q̄

Γq Γq

The boundary conditions on ΓT are called Dirichlet boundary condi-


tions and the ones on Γq as Neumann boundary conditions.
February 8, 2018 3D7 - Finite Element Methods 2016/17 48

2.3 Weak Form of Heat Equations

In preparation for the finite element method we consider next the weak
form of the heat equation. In the weak form the heat equation is
restated in the form of an integral. The weak form is equivalent to the
partial differential equation for heat (strong form).

In some engineering disciplines the weak form has a specific name; in


structural mechanics it is called the principle of virtual work.
February 8, 2018 3D7 - Finite Element Methods 2016/17 49

To derive the weak form of the heat equation, first the energy balance
equation is multiplied with a function w(x, y), and integrated over the
domain Ω Z
w (∇ · q − s) dΩ = 0 (2.4)

The function w(x, y) is the weight function (sometimes called the test
function). The arbitrariness of w(x, y) is important if the weak form
should be equal to the strong form of the heat equation.

The first term in the above equation is modified first using integration
February 8, 2018 3D7 - Finite Element Methods 2016/17 50

by parts
Z Z  
∂qx ∂qy ∂(wqx) ∂(wqy )
 
w + dΩ = + dΩ

∂x ∂y ∂x ∂y
ZΩ 
∂w ∂w

− qx + qy dΩ

∂x ∂y

or in matrix notation
Z Z Z
w∇ · q dΩ = ∇ · (wq) dΩ − ∇w · q dΩ
Ω Ω Ω
Next we apply the divergence theorem to the first integral on the right
hand side Z Z
∇ · (wq) dΩ = wq · n dΓ (2.5)
Ω Γ
February 8, 2018 3D7 - Finite Element Methods 2016/17 51

Furthermore, the boundary integral can be split into prescribed tem-


perature and prescribed flux boundaries
Z Z Z
wq · n dΓ = wq dΓ + wq · n dΓ (2.6)
Γ Γq ΓT

Inserting the last two equations 2.6 and 2.5 into the energy balance
equation 2.4 yields
Z Z Z Z
− ∇w · q dΩ + wq dΓ + wq · ndΓ − ws dΩ = 0
Ω Γq ΓT Ω

Next, the test functions are chosen so that they are zero on Dirichlet
boundaries, i.e. w0(x, y) = 0 on ΓT . Therefore, the integral on ΓT
February 8, 2018 3D7 - Finite Element Methods 2016/17 52

vanishes and the weak form is given by


Z Z Z
∇w0 · q dΩ + w0s dΩ = w0q dΩ
Ω Ω Γq

Introducing Fourier’s law, the weak form can be written in terms of


the temperature
Z Z Z
(∇w0)T D∇T dΩ = w0s dΩ − w0q dΓ
Ω Ω Γq

In the coming lectures this equation will be the starting point for de-
riving the finite element approach in two dimensions.

We note that while in the strong form we require second derivatives,


in the weak formulation only first derivatives are needed.
February 8, 2018 3D7 - Finite Element Methods 2016/17 53

2.4 Appendix: Laplace and Poisson Equations

The heat equation


 
∂ 2T ∂ 2T s
+ + =0
∂x2 ∂y 2 k

is in mathematics known as the Laplace equation if the domain term


s is zero and as the Poisson equation if the domain term s is nonzero.
The Poisson and Laplace equations appear in various physical contexts.
February 8, 2018 3D7 - Finite Element Methods 2016/17 54

Field variable Material constant Flux variable


Heat conduction temperature heat conductivity heat flux
Structural
membrane transversal deflection tension in membrane -
Seepage piezometric head permeability seepage veloci
Potential flow stream function density fluid velocity
Electrostatics voltage dielectric constant electric field
···

Consequently, any numerical method for the Laplace/Poisson equation


will be applicable to any of the problems listed in the above table.
February 8, 2018 3D7 - Finite Element Methods 2016/17 55

2.5 Appendix: Elementary Matrix Algebra

In the following the basic matrix operations needed for this module
are summarised.
3D7 - Finite Element Methods 2016/17

Chapter 3

Shape Functions for Triangular Elements

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 22

As you learned for one dimensional problems, in FEM the weight func-
tions and the trial solutions are constructed by subdividing the problem
domain into elements and considering shape functions within each ele-
ment.

In this chapter we will consider the construction of shape functions for


two-dimensional problems.
February 8, 2018 3D7 - Finite Element Methods 2016/17 23

3.1 Basic Requirements for Shape Functions

Properly constructed shape functions are necessary to ensure that the


finite element solution becomes more accurate as the number of elements
increases. As the element size goes to zero, the finite element solution
should converge to the exact solution.
February 8, 2018 3D7 - Finite Element Methods 2016/17 24

The main requirement for the shape functions is that the approximation
of the temperature T and the weight function w is continuous (since
the weak form contains the first order derivatives of both).
February 8, 2018 3D7 - Finite Element Methods 2016/17 25

3.2 Shape Functions for a Three-Node Triangular Element

For two-dimensional problems the three-node triangular element is the


most versatile and simplest finite element.
February 8, 2018 3D7 - Finite Element Methods 2016/17 26

The first step is to rewrite the integral in the weak formulation on the
whole domain as a sum of the integrals on each individual element:
Z XZ
(...) dxdy = (...) dxdy .
Ω Ωe

Now we just need consider one element e of the mesh above.


February 8, 2018 3D7 - Finite Element Methods 2016/17 27

One triangular three-noded element.

It is conventional to number the nodes counterclockwise from 1 to 3.


We could also use a clockwise orientation. The same orientation is used
February 8, 2018 3D7 - Finite Element Methods 2016/17 28

throughout all derivations.

The nodal coordinates of the element e are denoted (xe1, y1e), (xe2, y2e)
and (xe3, y3e). The temperature is only given at the three nodes. The
temperature within each triangular element is approximated by a linear
function.
February 8, 2018 3D7 - Finite Element Methods 2016/17 29

T e(x, y) = α0e + α1ex + α2ey (3.1)

Here the three α’s are parameters which we will determine. Denote the
nodal values of the temperature by T1e, T2e and T3e, to give the following
February 8, 2018 3D7 - Finite Element Methods 2016/17 30

three equations for determining the α’s:

T1e = α0e + α1exe1 + α2ey1e


T2e = α0e + α1exe2 + α2ey2e
T3e = α0e + α1exe3 + α2ey3e

or in matrix notation
    e
1 xe1 y1e α0e T1
1 xe ye αe = T e
2 2 1 2
1 xe3 y3e α2e T3e

This linear system of equations can be symbolically solved for the α’s
February 8, 2018 3D7 - Finite Element Methods 2016/17 31

using software, like Matlab, Mathematica or Maple.


 e  e e
−1  e
α0 1 x1 y 1 T1
αe = 1 xe2 y2e T2e
1
α2e 1 xe3 y3e T3e

Substituting the α’s into the equation 3.1 yields an expression for the
temperature depending only on the nodal values.

T e(x, y) = N1(x, y)T1e + N2(x, y)T2e + N3(x, y)T3e (3.2)

where N1(x, y), N2(x, y) and N3(x, y) are the shape functions (we
will see that in chapter 5 there is a clever way to deduce these functions
February 8, 2018 3D7 - Finite Element Methods 2016/17 32

through isoparametric mappings).

e 1
xe2y3e − xe3y2e + (y2e − y3e)x + (xe3 − xe2)y

N1 (x, y) =
2Ae
1
N2e(x, y) = xe3y1e − xe1y3e + (y3e − y1e)x + (xe1 − xe3)y

2Ae
e 1 e e e e e e e e

N3 (x, y) = x1y2 − x2y1 + (y1 − y2 )x + (x2 − x1)y
2Ae
where Ae is the area of the element

2Ae = (xe2y3e − xe3y2e) − (xe1y3e − xe3y1e) + (xe1y2e − xe2y1e)


February 8, 2018 3D7 - Finite Element Methods 2016/17 33

We can test if these shape equations are correct by substituting the


nodal coordinates and making sure that the values are correct at the
nodes.

N1e(xe1, y1e) = 1 N2e(xe1, y1e) = 0 N3e(xe1, y1e) = 0


N1e(xe2, y2e) = 0 N2e(xe2, y2e) = 1 N3e(xe2, y2e) = 0
N1e(xe3, y3e) = 0 N2e(xe3, y3e) = 0 N3e(xe3, y3e) = 1

Each shape function vanishes at all nodes except one. This can be
confirmed by the next figure showing the three shape functions. Notice
the shape functions are linear along the triangle edges.
February 8, 2018 3D7 - Finite Element Methods 2016/17 34

Three-node triangular element and its three shape functions.

For subsequent derivations, it is convenient to write equation 3.2 in


matrix notation:
February 8, 2018 3D7 - Finite Element Methods 2016/17 35

 e
T1
T e = N eae with N e = N1e N2e N3 ae = T2e
 e


T3e
February 8, 2018 3D7 - Finite Element Methods 2016/17 36

3.2.1 Example for Three-Node Triangular Element

We will compute the shape functions for the following triangular element

Example.
February 8, 2018 3D7 - Finite Element Methods 2016/17 37

To determine the shape functions we use the following linear polyno-


mial  e
α0
T e(x, y) = α0e + α1ex + α2ey = 1 x y α1e
 

α2e

Evaluating this at the three nodes gives


 e    e
T1 1 2 1 α0
T e = 1 5 3 αe
2 1
T3e 1 3 4 α2e
February 8, 2018 3D7 - Finite Element Methods 2016/17 38

To solve this system of equations the inverse of the matrix is computed


 −1  
1 2 1 11 −5 1
1
1 5 3 = −1 3 −2
7
1 3 4 −2 −1 3

Substituting this into the expression for T e gives


   e
11 −5 1 T1
e
  1
T = 1 x y −1 3 −2 T2e
7
−2 −1 3 T3e
February 8, 2018 3D7 - Finite Element Methods 2016/17 39

 e
1  T1e
= (11 − x − 2y) (−5 + 3x − y) (1 − 2x + 3y) T2 
7
T3e
 e
  T1
= N1e N2e N3e T2e
T3e
February 8, 2018 3D7 - Finite Element Methods 2016/17 40

Alternatively, the shape functions can be directly computed from the


closed form expressions derived in this section. The area of the triangle
is
2Ae = (5 · 4 − 3 · 3) − (2 · 4 − 3 · 1) + (2 · 3 − 5 · 1) = 7

and the shape functions are

e 1 1
N1 = (5 · 4 − 3 · 3 + (3 − 4)x + (3 − 5)y) = (11 − x − 2y)
7 7
e 1 1
N2 = (3 · 1 − 2 · 4 + (4 − 1)x + (2 − 3)y) = (−5 + 3x − y)
7 7
e 1 1
N3 = (2 · 3 − 5 · 1 + (1 − 3)x + (5 − 2)y) = (1 − 2x + 3y)
7 7
February 8, 2018 3D7 - Finite Element Methods 2016/17 41

3.3 Global Approximation

So far we have looked at the approximation of the temperature and


weight function locally on each element. The global approximation
over the entire domain is obtained by gathering the contributions from
individual elements.
February 8, 2018 3D7 - Finite Element Methods 2016/17 42

To explain the construction of the global shape functions, we consider


a four element mesh:

Finite element mesh with four elements and local and global node
numberings.
February 8, 2018 3D7 - Finite Element Methods 2016/17 43

The temperatures at the nodes can be identified either with the global
or local numbering. For example, the temperature at node 4 is denoted
by T4 with respect to the global node numbering and by T31, T22, T23 or
T14 with respect to the local node numberings for elements 1, 2, 3 and
4, respectively.

These equivalencies can be expressed in the form of simple mappings.


For example, the temperatures at the six nodes map to three nodes of
element 1 as follows:
February 8, 2018 3D7 - Finite Element Methods 2016/17 44

 
T1
 1   T2
T1 0 1 0 0 0 0 T3

T 1 = 0 0 0 0 1 0  
2 T4
T31 0 0 0 1 0 0 T5

T6

or in matrix notation
a1 = L1a
February 8, 2018 3D7 - Finite Element Methods 2016/17 45

After establishing these mappings for all the elements it is possible to


represent the temperature T h over the entire mesh. The superscript h
indicates that it is a discretized representation.
nel
X nel
X
Th = N eae = N e Le a
e=1 e=1

where nel is the number of the elements. As an illustration, the first


two terms in the summation for the shown mesh are

N 1L1 = 0 N11 0 N31 N21 0


 
1 1 2 2
 2 
N L + N L = N1 N11 N32 N31 + N22 N21 0
N 1 L1 + N 2 L2 + N 3 L3 = · · ·
February 8, 2018 3D7 - Finite Element Methods 2016/17 46

This motivates the definition of global shape functions in the form


nel
X
N= N e Le
e=1

Notice there are as many global shape functions as there are nodes in
the mesh. Using the global shape functions the temperature T h over
the entire domain is computed with
 
T1
nnp  T2 
X  
Th = Na =

NI aI = N1 N2 N3 ··· N6  T
  3

I=1 · · ·
T6
February 8, 2018 3D7 - Finite Element Methods 2016/17 47

where nnp is the number of nodes in the mesh. For the sample mesh
with four elements, two global shape functions N4 and N5 are shown
below:

Finite element mesh with four elements and the global shape
functions N5 and N4.
February 8, 2018 3D7 - Finite Element Methods 2016/17 48

N5 is one at node 5 and zero at all the other nodes and N4 is one at
node 4 and zero at all the other nodes. The global shape functions are
also known as tent functions.

Importantly, the approximation introduced leads to a globally contin-


uous representation of T h. This means that along element edges the
interpolation for T h on the left and right of the edge match up exactly.
February 8, 2018 3D7 - Finite Element Methods 2016/17 49

3.4 Derivatives of Shape Functions (”B-matrix”)

In evaluating the weak form it is necessary to compute the gradient of


T h. To this end first the gradient of the temperature within an element
e is computed.
February 8, 2018 3D7 - Finite Element Methods 2016/17 50

 
∂N1e(x, y) e ∂N2e(x, y) e ∂N3e(x, y) e
 ∂x T1 + ∂x
T2 +
∂x
T3 
∇T e(x, y) =
 
 
 ∂N e(x, y) ∂N e(x, y) ∂N3e(x, y) e
1 2
T1e + T2e + T3
∂y ∂y ∂y
  
∂N1e(x, y) ∂N2e(x, y) ∂N3e(x, y) T1e
 ∂x ∂x ∂x  
= 
  
  e
 ∂N e(x, y) ∂N e(x, y) ∂N e(x, y)  T2
1 2 3 e
∂y ∂y ∂y T3 }
| {z
| {z e
} ae
B
February 8, 2018 3D7 - Finite Element Methods 2016/17 51

For the triangular three-noded element the B e matrix is computed


with:
e − y e) (y e − y e) (y e − y e)

1 (y
Be = 2 3 3 1 1 2
2Ae (xe3 − xe2) (xe1 − xe3) (xe2 − xe1)

Note that the B e matrix is constant in each element and only depends
on the nodal coordinates of the element. Thus, the gradient of the tem-
perature is constant over each element and is in general discontinuous
over the element edges.
February 8, 2018 3D7 - Finite Element Methods 2016/17 52

3.5 Higher Order Triangular Elements

In deriving the shape functions for the three-noded triangular element a


linear distribution of the temperature within each element was assumed.
It is also possible to derive shape functions which give a quadratic tem-
perature distribution within each element.

T e(x, y) = α0e + α1ex + α2ey + α3ex2 + α4exy + α5ey 2


February 8, 2018 3D7 - Finite Element Methods 2016/17 53

To determine the six α parameters the element has to have six nodes
at which temperature is given. In addition to the three corner nodes a
node along the middle of each side is introduced.

Six node triangular element.


February 8, 2018 3D7 - Finite Element Methods 2016/17 54

Two representative shape functions of the six node triangular


element.

It is possible to create elements of arbitrary polynomial order using the


approach sketched. However, elements of very high order (≈ higher
than cubic) are seldom used because the resulting stiffness matrix is
February 8, 2018 3D7 - Finite Element Methods 2016/17 55

numerically not ”well behaved”.


3D7 - Finite Element Methods 2016/17

Chapter 4

Triangular Finite Elements for Scalar Field Problems

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 57

In this section we will substitute the approximated temperature into


the weak form of the heat equation.

Before we start it is instructive to remember the main steps of a finite


element analysis.

1. Generation of a mesh (preprocessing )


2. Formulation of the discrete finite element equations
3. Solving the discrete equations to obtain the nodal temperatures
4. Computing and displaying the temperature and fluxes over the entire
domain (postprocessing)
February 8, 2018 3D7 - Finite Element Methods 2016/17 58

Finite element mesh of a domain Ω. (Not all elements drawn.)


February 8, 2018 3D7 - Finite Element Methods 2016/17 59

4.1 Finite Element Formulation of the Heat Equation

The weak form of the heat equation introduced in Chapter 2 is the


starting point for the finite element formulation
Z Z Z
(∇w0)T D∇T dΩ = − w0q dΓ + w0s dΩ (4.1)
Ω Γq Ω

with  
∂T
 ∂x 
 
kxx kxy
∇T =   and D =
 ∂T  kyx kyy
∂y
The zero in the subscript w0 indicates that the test functions are chosen
to be zero on prescribed temperature (Dirichlet) boundaries.
February 8, 2018 3D7 - Finite Element Methods 2016/17 60

The integrals in the weak form can be computed by summing up the


element contributions. The area of each element is denoted by Ωe
nel Z Z Z !
X
(∇we)T D∇T dΩ + wq dΓ − ws dΩ =0
e=1 Ωe Γeq Ωe

where nel is the total number of elements in the mesh.

Next the weight function w0 and the temperature T are approximated


by interpolated values. The finite element approximation for the tem-
perature and the weight function within each element is
February 8, 2018 3D7 - Finite Element Methods 2016/17 61

nen
X
T e = N eae = NIeTIe
I=1
nen
X
w e = N ew e = NIewIe
I=1

where nen is the number of element nodes, N e are element shape


functions, T e is the element nodal temperature vector and we is the
element nodal weight function vector.
February 8, 2018 3D7 - Finite Element Methods 2016/17 62

The gradient of the temperature is


 
e ∂N1e ∂N2e e
 
∂T ∂Nen T1 e
···
e
  ∂x
 ∂x ∂x ∂x  T e
∇T = 2  = B eae
 ∂T e  = 
  
···
 ∂N1e ∂N2e e
∂Nen 
··· T e
∂y ∂y ∂y ∂y en
| {ze }
B
Similarly, the gradient of the weight function is

∇we = B ewe

Substituting the gradient of the temperature and the weight functions


February 8, 2018 3D7 - Finite Element Methods 2016/17 63

into the weak form 4.1 yields


nel Z Z Z !
X T
w (B e)T DB eae dΩ +
e N eweq dΓ − N ewes dΩ =0
e=1 Ωe Γeq Ωe
(4.2)
Recall (B ewe)T = weT B eT .

As introduced in Chapter 3, the element nodal temperature and weight


function vectors are related to the global temperature and global weight
function vectors via the scatter matrix Le

ae = Lea w e = Le w

The matrix Le maps global node numbers to local node numbers and
February 8, 2018 3D7 - Finite Element Methods 2016/17 64

consists of ones and zeros. Le introduced into the equation 4.2 gives
" nel Z Z Z !#
X T T T T
wT L e B DB e dΩ Lea +
e e
N q dΓ − e
N s dΩ =
e=1 Ωe Γeq Ωe

Here the first matrix is defined to be the element conductance matrix


Z
T
Ke = B DB e dΩ
e
Ωe

the second and third vectors are the element flux vector
Z Z
T T
fe = − e
N q dΓ + e
N s dΩ
Γeq | Ωe {z }
| {ze } e
fΓ fΩ
February 8, 2018 3D7 - Finite Element Methods 2016/17 65

where f eΓ is the element boundary flux and f eΩ is the element source


flux. Using these definitions the weak form can be written as
" nel nel
#
X T
X T
wT L K e Le a −
e L fe
e =0
e=1 e=1
Now the global conductance matrix is defined with
nel
X T
K= L K e Le
e

e=1

and the global flux vector with


nel
X T
f= L fe
e

e=1
February 8, 2018 3D7 - Finite Element Methods 2016/17 66

which leads to
wT [Ka − f ] = 0

For the left hand side to be zero, the expression within the bracket has
to be zero because the test function vector w is arbitrary. This leads
to the final finite element equation

Ka = f

As for the one dimensional case, it is not necessary to explicitly per-


form the multiplications with Le. The element conductance matrices
are added to the proper positions in the global conductance matrix
based on the correspondence of local and global node numberings. This
procedure is also referred to as assembly. The assembly of the element
February 8, 2018 3D7 - Finite Element Methods 2016/17 67

flux vectors proceeds along the same lines.

4.2 Illustrative Example

Consider the heat equation on a rectangular domain.


February 8, 2018 3D7 - Finite Element Methods 2016/17 68

A constant heat source with s = 3 is present throughout the domain.


The heat fluxes are prescribed with q = 0 on the top and bottom
domain boundaries and with q = 20 on the right domain boundary.
The temperature is prescribed with T = 0 on the left domain boundary.
The conductivity is isotropic with
 
5.0 0.0
D=
0.0 5.0
February 8, 2018 3D7 - Finite Element Methods 2016/17 69

The domain is discretized with two three-noded triangular elements

Finite element mesh with two elements.


February 8, 2018 3D7 - Finite Element Methods 2016/17 70

4.2.1 Global conductance matrix (”Stiffness matrix”)

First the element conductance matrices


Z
T
Ke = e
B DB e dΩ
Ωe
will be computed and later assembled to a global conductance matrix.
To compute the element conductivity matrices the matrix B e is needed.
For the three-noded element B e was
 
1 (y2e − y3e) (y3e − y1e) (y1e − y2e)
Be =
2Ae (xe3 − xe2) (xe1 − xe3) (xe2 − xe1)
where Ae is the element area. For the first element the B 1 matrix is
 
1 1 −1.0 1.0 0.0
B =
2 −2.0 0.0 2.0
February 8, 2018 3D7 - Finite Element Methods 2016/17 71

which leads to the conductance matrix


 
−1.0 −2.0   
1 1 1 T
1 1 5.0 0.0 −1.0 1.0 0.0
K = A B DB = 1.0 0.0 
4 0.0 5.0 −2.0 0.0 2.0
0.0 2.0
 
6.25 −1.25 −5.0
= −1.25 1.25 0.0
−5.0 0.0 5.0

Note the matrix K is symmetrical!!!. The B 2 matrix for the second


element is  
2 1 0 1 −1
B =
2 −2 2 0
February 8, 2018 3D7 - Finite Element Methods 2016/17 72

which leads to the conductance matrix


 
5.0 −5.0 0.0
K 2 = −5.0 6.25 −1.25
0.0 −1.25 1.25
Note the matrix K is symmetrical!!!. To assemble the global con-
ductance matrix the mapping between the local and global node num-
bers is established. Looking at the figures 4.2 and ?? the global node
numbering corresponding to the local numbering can be established as
follows:
   
6.25 −1.25 −5.0 1 5.0 −5.0 0.0 2
K 1 = −1.25 1.25 0.0 2 K 2 = −5.0 6.25 −1.25 4
−5.0 0.0 5.0 3 0.0 −1.25 1.25 3
1 2 3 2 4 3
February 8, 2018 3D7 - Finite Element Methods 2016/17 73

Note the matrix K is symmetrical!!!. The two element matrices


added to the global conductance matrix give
 
6.25 −1.25 −5.00 0.0
 
 
−1.25 1.25 + 5.0 0.0 + 0.0 −5.00
 
K=



−5.00 0.0 + 0.0 5.00 + 1.25 −1.25
 
 
0.0 −5.00 −1.25 6.25

Note that global nodes 2 and 3 receive contributions from both ele-
ments as they are attached to both elements. Note the matrix K is
February 8, 2018 3D7 - Finite Element Methods 2016/17 74

symmetrical!!!.

4.2.2 Global Source Vector

The element source vectors are


Z
T
f eΩ = e
N s dΩ
Ωe
The shape functions for a three-noded triangular element are given with
e 1 e e e e e e e e

N1 (x, y) = x2y3 − x3y2 + (y2 − y3 )x + (x3 − x2)y
2Ae
1
N2e(x, y) = xe3y1e − xe1y3e + (y3e − y1e)x + (xe1 − xe3)y

2Ae
e 1 e e e e e e e e

N3 (x, y) = x1y2 − x2y1 + (y1 − y2 )x + (x2 − x1)y
2Ae
February 8, 2018 3D7 - Finite Element Methods 2016/17 75

For s constant the element source vector is


 
Z 1.0
e e T sAe
fΩ = s N dΩ = 1.0
Ωe
3
1.0

R
where we have used the relation Ωe NIe dΩ = Ae/3. The element
source vectors are
   
1.0 1 1.0 2
f 1Ω = 1.0 2 f 2Ω = 1.0 4
1.0 3 1.0 3
February 8, 2018 3D7 - Finite Element Methods 2016/17 76

The assembly of the two element vectors leads to the global source
vector  
1.0
1.0 + 1.0
fΩ =  
1.0 + 1.0
1.0
February 8, 2018 3D7 - Finite Element Methods 2016/17 77

4.2.3 Boundary Flux Vector

Only element 2 has an edge on which a non-zero heat flux is prescribed


and hence a non-zero boundary flux vector. To compute the boundary
flux vector of element 2 we consider the line integral over the element
edges, of which only edge 1-2 has a non-zero contribution
Z Z
T T
f 2Γ = −q 2
N dΓ = −q 2
N dΓ
Γ2q edge 1−2
February 8, 2018 3D7 - Finite Element Methods 2016/17 78

For computing this line integral, the edge between nodes 1 and 2 is
considered as an one dimensional element. If we recall the visualisation
of the two-dimensional shape functions, it is clear that we have on edge
1-2 the one dimensional shape functions.

One dimensional shape functions on edge 1-2.

N12(t) = 1 − t N22(t) = t
February 8, 2018 3D7 - Finite Element Methods 2016/17 79

where t is a local coordinate system along the edge.


   
Z 1−t −10.0
f 2Γ = −q  t  dt = −10.0
edge 1−2 0 0.0

The last row in the above equation is zero because the shape function
belonging to node 3 is identical to zero along edge 1-2.
February 8, 2018 3D7 - Finite Element Methods 2016/17 80

To compose the global boundary flux vector we consider the mapping


between the local and global node numberings
 
−10.0 2
f 2Γ = −10.0 4
0.0 3

Hence, the global boundary flux vector is


 
0.0
−10.0
fΓ =  
 0.0 
−10.0
February 8, 2018 3D7 - Finite Element Methods 2016/17 81

4.2.4 Solution of Ka = f

Finally, the right hand side vector is


 
1.0
−8.0
f = fΩ + fΓ =  
 2.0
−9.0

The resulting global system of equations is


    
6.25 −1.25 −5.00 0.00 T1 1.0
−1.25 6.25 0.00 −5.00 T2 −8.0
   =  
−5.00 0.00 6.25 −1.25 T3  2.0
0.00 −5.00 −1.25 6.25 T4 −9.0
February 8, 2018 3D7 - Finite Element Methods 2016/17 82

T1 and T3 are zero because of the zero temperature boundary condition


at the left domain boundary.
    
6.25 −1.25 −5.00 0.00 0.0 1.0
−1.25 6.25 0.00 −5.00  T2  −8.0
   =  
−5.00 0.00 6.25 −1.25 0.0  2.0
0.00 −5.00 −1.25 6.25 T4 −9.0

From this linear equation system, the first row and column and the third
row and column can be removed because of the temperature boundary
conditions.     
6.25 −5.00 T2 −8.0
=
−5.00 6.25 T4 −9.0
February 8, 2018 3D7 - Finite Element Methods 2016/17 83

Finally, solving this equation system gives

T2 = −6.76
T4 = −6.84

2.0
2

3 1
−1.0 0

Rotation of element 1. The conductance matrix is unchanged.


February 8, 2018 3D7 - Finite Element Methods 2016/17 84

Invariance of K e under a solid-body transformation In this ex-


ample, Z Z
T T
Ke = B DB e dΩ = 5
e B B e dΩ
e
Ωe Ωe
Let a solid-body transformation be defined as

x̄ = x0 + RT x, or x = x̄0 + Rx̄, RT R = I ,

where R is an orthogonal matrix. A triangle Ωe is transferred to Ω̄e.


Then
N̄ i(x̄) = N i(x), B̄ = ∇¯ N̄ e = RB

Z Z
eT e eT
K̄ e = 5 B̄ B̄ dΩ̄ = 5 (RB) (RB)e dΩ̄
Ω̄e Ω̄e
February 8, 2018 3D7 - Finite Element Methods 2016/17 85

Z Z
eT eT
=5 (B) RT R (B)e dΩ̄ = 5 (B) (B)e |RT | dΩ = K e ,
Ω̄e Ω̄e

where |RT | is the determinant of RT which is 1. Here we used the


chain rule and integration by substitution (Jacobian) As an example,
the triangle element shown in the above figure is obtained by rotating
element 1 anticlockwise 90 degree, the element conductance matrix is
unchanged.
February 8, 2018 3D7 - Finite Element Methods 2016/17 86

4.2.5 Postprocessing

With the computed nodal temperatures, the temperature, its gradient


and the flux vector at any point in the element are computed

T e(x, y) = N e(x, y)ae


∇T e(x, y) = B e(x, y)ae
q(x, y) = −D∇T e(x, y)

As discussed, the gradient of the temperature will be discontinuous


across the edge between elements 1 and 2. Moreover, the computed
heat flux along the flux boundary (element 2, edge 1-2) will be different
from the applied flux q.

The jumps in the heat fluxes across the element boundaries as well as at
February 8, 2018 3D7 - Finite Element Methods 2016/17 87

the flux boundary are physically not reasonable (the discrete solution
violates energy conservation!).

The absolute value of the jumps in the gradients of T will tend to zero
if successively finer meshes (with more and more elements) are used. At
the same time the discrete solution converges to the analytic solution.
The jumps in the gradients can be used as a crude error measure to
assess the quality of the finite element solution.
February 8, 2018 3D7 - Finite Element Methods 2016/17 88

4.3 Appendix: Consistent Units in Finite Element Analysis

Many finite element software expects all the input as numbers without
the units. It is the user’s responsibility to make sure that the set of
units used is ”consistent”.

The easiest way to make things consistent is to use the International


System of Units (SI) base units, i.e., meters (m) for length, kilograms
(kg) for mass, kelvin (K) for temperature and seconds (s) for time.
There are in total three more base units of which not all are relevant
for this module.
February 8, 2018 3D7 - Finite Element Methods 2016/17 89

If the SI base units are used, the corresponding derived units have to
be used, i.e., newtons (N ) for force, joules (J) for energy, watts (W )
for power, watts per square meter (W/m2) and so on.

A basic mix up with units led to the destruction of the Mars Climate
Orbiter in 1999. According to wikipedia, the costs for this mission were
£225 million, of which £63 millions were launch costs.
3D7 - Finite Element Methods 2016/17

Chapter 5

Isoparametric Shape Functions

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 44

Isoparametric mapping is one of the most important finite element con-


cepts. It enables us to construct shape functions over quadrilaterals
and elements with curved boundaries, which are crucial for practical
engineering structures.
February 8, 2018 3D7 - Finite Element Methods 2016/17 45

5.1 Isoparametric Mapping in One Dimension

Although isoparametric mapping is not particularly useful in one di-


mension, it is very helpful for understanding the general approach.

In the first set of four lectures, the following shape functions for two-
noded elements were introduced.

Shape functions for a two-noded element.


February 8, 2018 3D7 - Finite Element Methods 2016/17 46

x − xe1 x − xe1
N1(x) = 1 − e N2(x) = e
x2 − xe1 x2 − xe1
Next, we consider a coordinate transformation which transforms (maps)
the coordinate x into a local (element specific) coordinate ξ

Global coordinate x and local coordinate ξ

The coordinate ξ fulfils the relationships


x = xe1 at ξ = −1 and x = xe2 and ξ=1
February 8, 2018 3D7 - Finite Element Methods 2016/17 47

After some algebra we can compute the following ”stretch” transforma-


tion
1 e
e
x(ξ) = x1 + (x2 − xe1)(1 + ξ)
2
x1 + xe2 xe2 − xe1
e
= + ξ
2 2

The shape functions N1e and N2e can also be expressed using ξ

e 1 e 1
N1 (ξ) = (1 − ξ) N2 (ξ) = (1 + ξ)
2 2

The key idea of the isoparametric concept is to use these shape functions
for writing the coordinate transformation between x and ξ

x(ξ) = N1e(ξ)xe1 + N2e(ξ)xe2 (5.1)


February 8, 2018 3D7 - Finite Element Methods 2016/17 48

We can show that this gives the same result as equation 5.1
1 1
x(ξ) = (1 − ξ)x1 + (1 + ξ)xe2
e
2 2
xe1 + xe2 xe2 − xe1
= + ξ
2 2
This coordinate transformation can be visualised as a mapping from a
parent element in the ξ-coordinate system to the physical element in
the x-coordinate system.

Mapping of the parent element onto the physical element.


February 8, 2018 3D7 - Finite Element Methods 2016/17 49

The coordinate ξ is usually called the natural coordinate and always


lies by definition between −1 and +1. The parent element is solely for
numerical purposes. The finite element analysis is still performed over
the physical domain.

In an isoparametric element the field variable, like temperature, is


approximated with the same set of shape functions as those used for
the coordinate transformation.

T (ξ) = N1(ξ)T1e + N2(ξ)T2e


February 8, 2018 3D7 - Finite Element Methods 2016/17 50

To compute the derivatives which appear in the weak form the chain
rule is used
dT dT dξ
=
dx dξ dx

The derivative dξ/dx is determined from the mapping between ξ and


x (eq. 5.1).

It is evident that isoparametric mapping is also suitable for higher order,


e.g. quadratic, shape functions.
February 8, 2018 3D7 - Finite Element Methods 2016/17 51

5.2 Isoparametric Quadrilateral Elements


February 8, 2018 3D7 - Finite Element Methods 2016/17 52

5.2.1 Four-Node Rectangular Element

Before formulating the shape functions for quadrilateral elements, we


will determine the shape functions for a four-node rectangular element.

Four-node rectangular element. The nodes are by definition


numbered counter-clockwise.
February 8, 2018 3D7 - Finite Element Methods 2016/17 53

As the element has four nodes, it is necessary to start with a polynomial


that has four parameters.

T e = α0e + α1ex + α2ey + α3exy

As shown for the triangular element, it is possible to express (α0e, α1e, α2e, α3e)
in terms of the nodal temperature values (T1e, T2e, T3e, T4e). A deriva-
tion along the lines of the three-node element is tedious as it is necessary
to invert a 4 × 4 matrix.
February 8, 2018 3D7 - Finite Element Methods 2016/17 54

An alternative and more elegant approach is to construct the shape


functions by the tensor product method. This is based on taking
products of one-dimensional shape functions.

Construction of two dimensional shape functions.


February 8, 2018 3D7 - Finite Element Methods 2016/17 55

For example, N2e, which has to have the value one at node 2 and
zero at the other nodes, is obtained by taking the product of the one-
e,1d e,1d
dimensional shape functions N2 (x) and N1 (y).
e,1d e,1d
N2e = N2 (x) × N1 (y)

e,1D e,1D
As visible in the figure above the product N2 (x) × N1 (y) has
the value one at node 2 and is zero at nodes 1, 3 and 4.

Hence, the four shape functions, also called bilinear shape functions, for
February 8, 2018 3D7 - Finite Element Methods 2016/17 56

the rectangular element are


x − xe2 y − y4e 1
N1e(x, y) = = e (x − xe2)(y − y4e)
xe1 − xe2 y1e − y4e A
x − xe1 y − y4e 1
N2e(x, y) = = − e (x − xe1)(y − y4e)
xe2 − xe1 y1e − y4e A
(5.2)
x − xe1 y − y1e 1
N3e(x, y) = = e (x − xe1)(y − y1e)
xe2 − xe1 y4e − y1e A
x − xe2 y − y1e 1
N4e(x, y) = = − e (x − xe2)(y − y1e)
xe1 − xe2 y4e − y1e A

where Ae is the area of the element. The four shape functions are
plotted in the following figure:
February 8, 2018 3D7 - Finite Element Methods 2016/17 57

Four shape functions of the rectangular element (on [0, 2] × [0, 2]).

The computed shape functions are suitable for rectangles and could be
used with meshes consisting only of rectangles, but they are not suitable
for arbitrary quadrilaterals.
February 8, 2018 3D7 - Finite Element Methods 2016/17 58

5.2.2 Isoparametric Quadrilateral Element

The idea of isoparametric mapping is used for deriving shape functions


for arbitrary quadrilateral elements

Mapping of the bi-unit parent element onto the quadrilateral


element in the physical space.
February 8, 2018 3D7 - Finite Element Methods 2016/17 59

The bi-unit square is the parent domain and ξ and η are its natural co-
ordinates. To map points from the parent domain onto the quadrilateral
in the physical domain the four nodal shape functions are used

x(ξ, η) = N 4Q(ξ, η)xe y(ξ, η) = N 4Q(ξ, η)y e

where N 4Q(ξ, η) are the four-node element shape functions in the


natural coordinates and xe and y e are the vectors of the element coor-
dinates  e  e
x1 y1
e
xe e
ye
x =  2e y =  2e
x  y 
3 3
xe4 y4e
Above, the notation has been changed from N e to N 4Q to emphasize
February 8, 2018 3D7 - Finite Element Methods 2016/17 60

that the shape functions are no longer shape functions of the element
coordinates. They are identical for every quadrilateral element.

As the parent element is a bi-unit square its shape functions are identical
to those of the rectangular element expressed in ξ and η coordinates.
4Q 1
N1 (ξ, η) = (1 − ξ)(1 − η)
4
4Q 1
N2 (ξ, η) = (1 + ξ)(1 − η)
4
4Q 1
N3 (ξ, η) = (1 + ξ)(1 + η)
4
4Q 1
N4 (ξ, η) = (1 − ξ)(1 + η)
4

The temperature will be approximated with the same shape functions

T e = N 4Q(ξ, η)ae
February 8, 2018 3D7 - Finite Element Methods 2016/17 61

The element is called isoparametric because the temperature approxi-


mation and the mapping of the geometry is accomplished with the same
shape functions.
February 8, 2018 3D7 - Finite Element Methods 2016/17 62

Derivatives of Isoparametric Shape Functions

The procedure for computing the derivatives of the isoparametric el-


ements is slightly more involved because the shape functions are ex-
pressed in natural coordinates ξ and η.

The gradient of the temperature for the four-node (isoparametric) quadri-


lateral element is
∇T = B eae
February 8, 2018 3D7 - Finite Element Methods 2016/17 63

with  
4Q 4Q 4Q 4Q
∂N1 ∂N2 ∂N3 ∂N4
 ∂x ∂x ∂x ∂x
 

Be =  
 ∂N14Q ∂N24Q ∂N34Q 4Q 
 
∂N4
∂y ∂y ∂y ∂y
To compute shape function derivatives the chain rule will be used
4Q 4Q 4Q
∂NI ∂NI ∂x ∂NI ∂y
= +
∂ξ ∂x ∂ξ ∂y ∂ξ
4Q 4Q 4Q
∂NI ∂NI ∂x ∂NI ∂y
= +
∂η ∂x ∂η ∂y ∂η
February 8, 2018 3D7 - Finite Element Methods 2016/17 64

written as matrices and vectors this becomes


   
4Q ∂x ∂y  ∂N 4Q
∂NI
I
 ∂ξ  ∂ξ ∂ξ   ∂x 
 
  
 =  
 4Q     4Q 
 ∂N  ∂x ∂y  ∂N 
I I
∂η{z∂η } ∂y
∂η | e
J

J e is the Jacobian which contains the derivatives of the physical coordi-


nates with respect to the natural coordinates. The derivatives required
February 8, 2018 3D7 - Finite Element Methods 2016/17 65

for the weak form are computed by inverting the above expression
   
4Q 4Q
∂NI ∂NI
 ∂ξ 
 ∂x 
 
 = (J e)−1 
 

 4Q 

 4Q  (5.3)
 ∂N   ∂N 
I I
∂y ∂η
The inverse of J e is
 ∂y ∂y 

1 ∂η ∂ξ 
(J e)−1 = e 


|J |  
∂x ∂x

∂η ∂ξ
where |J e| is the determinant of the Jacobian. Remember that the de-
terminant of the Jacobian is the ratio of an area element in the physical
February 8, 2018 3D7 - Finite Element Methods 2016/17 66

domain to the corresponding area element in the parent domain.

The isoparametric mapping is used to compute the Jacobian

x(ξ, η) = N 4Q(ξ, η)xe y(ξ, η) = N 4Q(ξ, η)y e

which leads to
 
4Q 4Q 4Q 4Q
∂N1 ∂N2 ∂N3
e ∂N4  
 ∂ξ  1x y1e
 ∂ξ  x∂ξ
e ∂ξ y2e
Je =   2
e

 4Q  x
 ∂N ∂N 4Q ∂N 4Q ∂N 4Q  3e
 y3e
1 2 3 4 x4 y4e
∂η ∂η ∂η ∂η
February 8, 2018 3D7 - Finite Element Methods 2016/17 67

5.2.3 Higher Order Quadrilateral Element

Higher order quadrilateral elements provide the ability to model curved


edges. The advantage of curved edges is that fewer elements can be
used around holes and other curved surfaces than with straight-sided
elements.
February 8, 2018 3D7 - Finite Element Methods 2016/17 68

As an example consider the following nine noded element.

Nine-node ispoarametric quadrilateral in parameter (left) and


physical space (right).

The nine-node isoparametric element is constructed as a tensor product


of the one-dimensional quadratic shape functions.
February 8, 2018 3D7 - Finite Element Methods 2016/17 69

5.2.4 Illustrative Example

We will investigate the effect of the element shape on the isoparametric


mapping.
February 8, 2018 3D7 - Finite Element Methods 2016/17 70

First, consider the following isoparametric mapping for a four-node


quadrilateral element
February 8, 2018 3D7 - Finite Element Methods 2016/17 71

The Jacobian of this mapping


 ∂x ∂y 
 ∂ξ ∂ξ 
Je =  
 
∂x ∂y
∂η ∂η
is computed from
4Q 4Q 4Q 4Q
x = 0.0N1 + 5.0N2 + 3.0N3 + 0.0N4
1 1
= 2ξ − η − ξη + 2
2 2

4Q 4Q 4Q 4Q
y = 0.0N1 + 0.0N2 + 3.0N3 + 5.0N4
1 1
= − ξ + 2η − ξη + 2
2 2
February 8, 2018 3D7 - Finite Element Methods 2016/17 72

After some algebra we obtain


 
η 1 1
2− − − η
 2 2 2 
Je =  
 1 1 1

− − ξ 2− ξ
2 2 2
The Jacobian has to be invertible for computing the derivatives of the
shape functions with respect to the physical coordinates (see eq. 5.3).

For the mapping to be invertible, the determinant of the Jacobian has


to be larger than zero over the entire element

e 5
det J = (3 − ξ − η) > 0
4
which is the case for this mapping.
February 8, 2018 3D7 - Finite Element Methods 2016/17 73

In contrast to the previous mapping, it can be for the following mapping


shown

that the determinant of the Jacobian is zero or negative close to the


non-convex corner.

Notice that some region of the parent element close to node 3 is mapped
February 8, 2018 3D7 - Finite Element Methods 2016/17 74

outside the physical domain. If such non-convex elements are present


in the finite element mesh, the results of the finite element computation
will be useless.

5.2.5 Isoparametric Triangular Elements

The isoparametric idea also applies to triangular elements. The parent


domains for a three noded triangular element is shown below
February 8, 2018 3D7 - Finite Element Methods 2016/17 75

η y
3 (0,1) 3
2

1
x
2 (1,0)
1 (0,0) ξ

Mapping of the three-noded parent element onto the three-noded


element in the physical space.
February 8, 2018 3D7 - Finite Element Methods 2016/17 76

The shape functions for the three-node parent triangular element is


constructed easily.
N13T = 1 − ξ − η
N23T = ξ
N33T = η

To map the parent domain to the triangle of the physical domain, the
three nodal shape functions are used

x = x1 + ξ(x2 − x1) + η(x3 − x1)


y = y1 + ξ(y2 − y1) + η(y3 − y1)

Note that the edges of the element in the physical domain are all
straight. From the above linear equations, it is easy to represent ξ
February 8, 2018 3D7 - Finite Element Methods 2016/17 77

and η in terms of x and y, hence the shape functions of the triangular


element in the physical domain, as NIe(x, y) = NIe(ξ, η) The B e of the
parent element is a constant matrix:
 
−1 1 0
Be =
−1 0 1

And the Jacobian J e is also a constant matrix:


 
x2 − x1 x3 − x1
Je =
y2 − y1 y3 − y1

These constant matrices make the calculation with three-noded trian-


gular elements simple.
February 8, 2018 3D7 - Finite Element Methods 2016/17 78

Six-noded triangular elements enables us to model curved edges. The


parent domains for a six noded triangle is shown below

η y
3 (0,1) 3 5
6
2
5 (0.5,0.5)
6 (0,0.5)
1 4

2 (1,0)
x
1 (0,0) 4 (0.5,0) ξ

Mapping of the six-noded parent element onto the six-noded element


February 8, 2018 3D7 - Finite Element Methods 2016/17 79

in the physical space.

The shape functions for the six-node parent triangle are constructed
easily from those of three-node parent triangle.

N16T = (1 − ξ − η)(1 − 2ξ − 2η)


N26T = ξ(2ξ − 1)
N36T = η(2η − 1)
N46T = 4ξ(1 − ξ − η)
N56T = 4ξη
N66T = 4η(1 − ξ − η)

In accordance with the isoparametric idea, these shape functions are


used for the mapping between the parent and physical domains. Note
that we only need curved edge elements on the physical boundary of
February 8, 2018 3D7 - Finite Element Methods 2016/17 80

the domain. It is computationally more efficient to use straight edge


elements in the interior of the physical domain. In this case, if the nodes
4, 5 and 6 are the middle points of the relevant edge, the map of the six-
noded element is degenerated into the one of the three-node element.
The linearity of the three-node element makes FEM calculation much
simpler. It is a common usage to use six-noded elements to present
the physical field (temperature field) to gain accuracy but three-noded
elements for the mapping between the parent element and that of the
physical domain for computational efficiency.
February 8, 2018 3D7 - Finite Element Methods 2016/17 81

5.3 Appendix: Linear and Quadratic Isoparametric Mapping in


One Dimension

The two shape functions two noded element on the reference interval
[0, 1] are
N12 = 1 − ξ , N22 = ξ .

The linear mapping from [0, 1] to interval [a, b] is

x = N12(ξ)a + N22(ξ)b = a + ξ(b − a) .

The three shape functions three noded element on the reference interval
[0, 1] are

N13 = (1 − ξ)(1 − 2ξ) , N32 = ξ(2ξ − 1) , N33 = 4ξ(1 − ξ) .


February 8, 2018 3D7 - Finite Element Methods 2016/17 82

The quadratic mapping from [0, 1] to interval [a, b] with an intermedi-


ate node of coordinate c is

x = N13(ξ)a + bN23(ξ)b + N33(ξ)c .

When c = (a + b)/2, the quadratic mapping is degenerated into the


linear mapping:
a+b
 
x = (1 − ξ)(1 − 2ξ)a + ξ(2ξ − 1)b + 4ξ(1 − ξ) = a + ξ(b − a) .
2
3D7 - Finite Element Methods 2016/17

Chapter 6

Isoparametric Finite Elements for Scalar Field Problems

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 84

In this chapter we will develop the finite element analysis for arbitrar-
ily shaped quadrilateral and triangular elements. Isoparametric shape
functions developed in the previous chapter will be used.

6.1 Numerical Integration (Quadrature)

During the finite element solution procedure, it is necessary to integrate


various quantities, for instance the conductance matrix or source vec-
tors. For isoparametric elements these quantities must in general be
integrated numerically due to the presence of the Jacobian.

Although there are many different numerical integration techniques, in


finite elements Gauss integration is prefered.
February 8, 2018 3D7 - Finite Element Methods 2016/17 85

6.1.1 Review of Basic Integration Rules

Consider the following integral which is to be integrated numerically


Z b
f (x)dx
a

The integration domain [a, b] is first split into N equidistant subintervals


b−a
xI = a + Ih with h = for I = 0, 1, · · · , N
N

There are various schemes for approximating the integral using the
function values at N positions
February 8, 2018 3D7 - Finite Element Methods 2016/17 86

• Rectangle rule (constant approximation)

Integration of f (x) with the rectangle rule between x = a and


x = b.
Z b
f (x)dx ≈ h (f (x0) + f (x1) + f (x2) · · · + f (xN −1))
a
February 8, 2018 3D7 - Finite Element Methods 2016/17 87

• Trapezoidal rule (linear approximation)

Integration of f (x) with the trapezoidal rule between x = a and


x = b.
Z b
h
f (x)dx ≈ (f (x0) + 2f (x1) + 2f (x2) + · · · + 2f (xN −1) + f (xN ))
a
2
February 8, 2018 3D7 - Finite Element Methods 2016/17 88

• Simpson’s rule (quadratic approximation)


Z b
1
f (x)dx ≈ h(f (x0)+4f (x1)+2f (x2)+4f (x3) · · ·+4f (xN −1)+f (xN ))
a
3

Notice that all these integration rules can be written in the following
form
Z b N
X
f (x)dx ≈ wI f (xI )
a I=0

where wI are the integration weights and xI are the integration points.
February 8, 2018 3D7 - Finite Element Methods 2016/17 89

6.1.2 Gauss Integration in One-dimension

Gauss integration is a very efficient technique for integrating functions


that are (almost) polynomials, like finite element shape functions.

Gauss integration needs fewer subintervals to provide the same accuracy


as the integration rules discussed in the previous section.

The number of integration points is of great importance for practical


computations since the fewer the integration points the faster the finite
element analysis will be.
February 8, 2018 3D7 - Finite Element Methods 2016/17 90

The Gauss integration formulas are always given over the parent domain
[−1, 1]
Z +1 N
X
f (ξ)dξ ≈ wI f (ξI )
−1 I=1

For example, in case of two integration points (Gauss points) the pre-
viously considered integral is approximated with
February 8, 2018 3D7 - Finite Element Methods 2016/17 91

Integration of f (ξ) between ξ = −1 and ξ = 1 using two Gauss


points.
Z +1
f (ξ)dξ ≈ 1.0f (− √1 ) + 1.0f ( √1 )
3 3
−1
February 8, 2018 3D7 - Finite Element Methods 2016/17 92

The Gauss integration weights and the positions for up to three inte-
gration points are tabulated below.

N location ξI weight wI
1 0 2

2 −1/ 3 1


1/ 3 1

3 − 0.6 5/9

0 8/9


0.6 5/9
February 8, 2018 3D7 - Finite Element Methods 2016/17 93

Notice the distances between the integration points are not constant as
in conventional schemes.

It can be shown that with N Gauss points the integration is


exact up to polynomial order (2N − 1). For example, using two
integration points (N = 2) linear, quadratic and cubic functions are
exactly integrated.
February 8, 2018 3D7 - Finite Element Methods 2016/17 94

6.1.3 Integration over Quadrilateral Elements

Gaussian integration in two dimensions is constructed by using one-


dimensional Gaussian integration in each coordinate direction in turn.
For the integration over the bi-unit parent element we have
!
Z +1 Z +1 Z +1 Z +1
f (ξ, η) dξdη = f (ξ, η) dξ dη
−1 −1 −1 −1
M N
!
X X
= f (ξI , ηJ )wI wJ
J=1 I=1
February 8, 2018 3D7 - Finite Element Methods 2016/17 95

This is illustrated in the following figure

Four-point (left) and nine-point (right) integration over bi-unit


square.
February 8, 2018 3D7 - Finite Element Methods 2016/17 96

6.1.4 Integration over Triangular Elements

For triangles the integration rules are different than those for quadri-
laterals. We will not attempt to derive them. For completeness the
following integration formula
Z N
X
f (ξ, η) dξdη = f (ξI , ηI ) wI
Ωe I=1

and the related integration points and weights are given


February 8, 2018 3D7 - Finite Element Methods 2016/17 97

One point (left) and three point (right) integration over the unit
triangle.
February 8, 2018 3D7 - Finite Element Methods 2016/17 98

6.2 Evaluation of Finite Element Integrals

The integration domain for finite element matrices and vectors is the
physical element domain. Therefore, we need to consider the isopara-
metric mapping from the the parent domain ([−1, +1] × [−1, +1]) to
the physical element domain Ωe.

For example, the conductance matrix for isoparametric elements is com-


puted with
Z Z
T T
Ke = B (x, y)DB e(x, y) dΩ =
e B (x, y)DB e(x, y) dx dy
e

ZΩ+1
e
Z +1
Ωe
T
= e
B (ξ, η)DB e(ξ, η) |J e(ξ, η)| dξ dη
−1 −1
February 8, 2018 3D7 - Finite Element Methods 2016/17 99

where |J e| is the Jacobian of the isoparametric mapping and takes care


of the mapping of infinitesimal area elements from parent to the physical
domain.
dx dy = |J e|dξ dη

The conductance matrix is evaluated using Gauss quadrature


N X
X M
T
Ke = B (ξI , ηJ )DB e(ξI , ηJ ) |J e(ξI , ηJ )|wI wJ
e

I=1 J=1
February 8, 2018 3D7 - Finite Element Methods 2016/17 100

6.3 Illustrative Example

Consider the four-noded isoparametric quadrilateral element


February 8, 2018 3D7 - Finite Element Methods 2016/17 101

The heat conductivity is isotropic with


 
5 0
D=
0 5

First, we compute the Jacobian of the mapping from the parent to the
physical domain. The Jacobian is given by (see section 5.2.2)
 
4Q 4Q 4Q 4Q
∂N1 ∂N2 ∂N3 ∂N4 e
 e

 ∂ξ x 1 y1
∂ξ ∂ξ ∂ξ  xe

 y2e
Je =   2
e

 4Q  x
 ∂N ∂N 4Q ∂N 4Q ∂N 4Q  3e
 y3e
1 2 3 4 x4 y4e
∂η ∂η ∂η ∂η
February 8, 2018 3D7 - Finite Element Methods 2016/17 102

4Q
where NI are the four-node shape functions in natural coordinates.
Introducing nodal element coordinates gives
 
  0 1  
e 1 η − 1 1 − η 1 + η −η − 1  0 0  0 0.125η − 0.375
J =   =
4 ξ − 1 −ξ − 1 1 + ξ 1 − ξ 2 0.5 1 0.125ξ + 0.125
2 1
The B e matrix contains the derivatives of the shape functions with
respect to physical coordinates
 
4Q 4Q 4Q 4Q
∂N1 ∂N2 ∂N3 ∂N4
 ∂x ∂x ∂x ∂x
 

Be =  
 ∂N14Q ∂N24Q ∂N34Q 4Q 
 
∂N4
∂y ∂y ∂y ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 103

Hence, the 4 × 4 conductance matrix is given by


4Q 4Q
 
∂N1 ∂N1
 ∂x
 4Q ∂y4Q 
 
4Q 4Q 4Q 4Q

∂N1 ∂N2 ∂N3 ∂N4
Z  ∂N2 ∂N2    
∂x ∂x ∂x ∂x 
  
5 0
 ∂x4Q ∂y4Q 

Ke =  dΩ
  
Ωe ∂N ∂N 0 5  4Q 4Q 4Q 4Q

 3 3   ∂N1 ∂N2 ∂N3 ∂N4 
 
 ∂x ∂y 
 ∂N 4Q ∂N 4Q  ∂y ∂y ∂y ∂y
4 4
∂x ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 104

e component of the element


In the following we will compute the K11
conductance matrix by numerical integration with 2 × 2 Gauss points
 
4Q
Z +1 Z +1  4Q    ∂N1
4Q 5 0 
e ∂N 1 ∂N 1 ∂x
 e
 4Q  |J |dξ dη
K11 =
∂x ∂y 0 5  ∂N1 
−1 −1
∂y
 !2 !2
Z +1 Z +1 4Q 4Q
∂N1 ∂N1
= 5 +  |J e|dξdη
−1 −1
∂x ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 105

whereby the gradient of the shape functions are computed from


   
4Q 4Q
∂N1 ∂N1
 ∂ξ
 ∂x 
  
−1
e
 
  = (J )  
 4Q   4Q 
 ∂N1   ∂N 
1
∂y ∂η
February 8, 2018 3D7 - Finite Element Methods 2016/17 106

e is computed as
The contribution of each Gauss point to K11

√ √
Gauss point 1 ξ = −1/ 3 η = −1/ 3
   
0.0 −0.447 −1 0.1182 1.0
Je = |J e| = 0.4472 J e =
1.0 0.0524 −2.24 0.0
 
4Q
∂N1
 ∂x 
       
  = J e−1 1 η − 1 = J e−1 −0.394 = −0.441
 4Q  4 ξ−1 −0.394 0.882
 ∂N1 
∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 107

√ √
Gauss point 2 ξ = 1/ 3 η = −1/ 3
   
0.0 −0.447 −1 0.441 1.0
Je = |J e| = 0.4472 J e =
1.0 0.197 −2.24 0.0
 
4Q
∂N1
 ∂x
     
 = J e−1 −0.394 = −0.280


 4Q 
 ∂N1  −0.106 0.882

∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 108

√ √
Gauss point 3 ξ = 1/ 3 η = 1/ 3
   
0.0 −0.303 −1 0.651 1.0
Je = |J e| = 0.3028 J e =
1.0 0.197 −3.30 0.0
 
4Q
∂N1
 ∂x
     
 = J e−1 −0.106 = −0.174


 4Q 
 ∂N1  −0.106 0.349

∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 109

√ √
Gauss point 4 ξ = −1/ 3 η = 1/ 3
   
0.0 −0.303 −1 0.174 1.0
Je = |J e| = 0.3028 J e =
1.0 0.0528 −3.30 0.0
 
4Q
∂N1
 ∂x
     
 = J e−1 −0.106 = −0.413


 4Q 
 ∂N1  −0.394 0.349

∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 110

e is computed by summing up the contributions from the four Gauss


K11
points
e = 5[(0.4412 + 0.8822)0.4472]
K11
+ 5[(0.2802 + 0.8822)0.4472]
+ 5[(0.1742 + 0.3492)0.3028]
+ 5[(0.4132 + 0.3492)0.3028] = 4.76
February 8, 2018 3D7 - Finite Element Methods 2016/17 111

6.4 Number of Integration Points to Use

If a high number of integration points is used, finite element integrals are


evaluated very accurately. On the other hand with too few integration
points the finite element integrals may be evaluated poorly.

In particular the conductance matrix integrated with too few points can
cause rank-deficiency and can render the problem unsolvable.

As a rule, the number of integration points is chosen so that the matrices


and vectors are accurately computed for an undistorted isoparametric
element (with constant Jacobian).
February 8, 2018 3D7 - Finite Element Methods 2016/17 112

For example, for the conductance matrix of the quadrilateral element


the following integration rules are to be used.

Element Number of integration points


four-node 2×2
nine-node 3×3
February 8, 2018 3D7 - Finite Element Methods 2016/17 113

6.5 Modelling Considerations

6.5.1 Element Geometries

As discussed in the previous chapter, if det(J e) is zero, an area element


in the parent element is mapped into a zero area in the physical element,
which is not acceptable.

Similarly, if elements are excessively distorted, an area element in the


parent element is mapped into a nearly zero area. To ensure that
det(J )e is safely larger than zero, certain severely distorted element
shapes must be avoided.
February 8, 2018 3D7 - Finite Element Methods 2016/17 114

Quadrilateral element geometries to be avoided.

Triangular element geometries to be avoided.

Notice for higher order elements, like the nine-noded one, the position
of the mid-nodes contribute to the element distortion. Therefore, they
must lie at a certain distance from the corner nodes.
February 8, 2018 3D7 - Finite Element Methods 2016/17 115

Range of acceptable positions for the midnodes of a nine-node


quadrilateral element.
February 8, 2018 3D7 - Finite Element Methods 2016/17 116

6.6 Appendix: Gauss Integration

In contrast to standard integration schemes, e.g. trapezoidal or Simp-


son’s rule, the position of the integration points is chosen so that max-
imum accuracy is achieved.

The basic idea is demonstrated by way of an example. Assume that we


want to integrate an arbitrary cubic polynomial over the parent domain
[−1, 1]
f (ξ) = α1 + α2ξ + α3ξ 2 + α4ξ 3

where α1, α2, α3 and α4 are constants. The integral will be approxi-
mated with two integration points
February 8, 2018 3D7 - Finite Element Methods 2016/17 117

Integration of f (ξ) between ξ = −1 and ξ = 1 using two Gauss


points.
February 8, 2018 3D7 - Finite Element Methods 2016/17 118

Z +1
f (ξ) dξ = w1f (ξ1) + w2f (ξ2)
−1
= w1(α1 + α2ξ1 + α3ξ12 + α4ξ13)
+ w2(α1 + α2ξ1 + α3ξ12 + α4ξ23)

Alternatively, the integration can be performed analytically


Z +1 Z +1
α1 + α2ξ + α3ξ 2 + α4ξ 3 dx

f (ξ) dξ =
−1
−1 +1
ξ2 ξ3 ξ 4
= α1ξ + α2 + α3 + α4
2 3 4
−1
2
= 2α1 + 0 + α3 + 0 (6.1)
3
February 8, 2018 3D7 - Finite Element Methods 2016/17 119

Comparing the coefficients of the constants in equations 6.1 and 6.1 we


obtain the following four equations

w1 + w2 = 2 (6.2)
w1ξ1 + w2ξ2 = 0
2 2 2
w1ξ1 + w2ξ2 = (6.3)
3
w1ξ13 + w3ξ22 = 0

This is a set of four nonlinear equations for determining the values of


w1, w2, ξ1 and ξ2.
February 8, 2018 3D7 - Finite Element Methods 2016/17 120

Symmetry considerations require that w1 = w2 and ξ1 = −ξ2, and


equation 6.2 is solved to obtain

w1 = w2 = 1

Subsequently we obtain from equation (6.3)


1 1
ξ1 = − √ and ξ2 = √
3 3
3D7 - Finite Element Methods 2016/17

Chapter 7

Finite Elements for Planar Linear Elasticity

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 122

The displacements and stresses in solids are determined by solving the


linear elasticity equations. Linear elastic analysis is usually referred to
as stress analysis and is one of the most common applications of the
finite element method.
February 8, 2018 3D7 - Finite Element Methods 2016/17 123

In this chapter we derive the finite element formulation for linear elastic-
ity following a similar approach to the one used for the heat equation.
The essential difference is that the governing equations now have as
variables the two components of the displacement vector (instead
of the one scalar temperature).

7.1 Kinematics

The displacement vector for planar elasticity is a vector with two com-
ponents  
ux
u=
uy
February 8, 2018 3D7 - Finite Element Methods 2016/17 124

The strains are defined based on displacements of an infinitesimal area


element.

The strains xx and yy are a measure of the extension of infinitesimal
line segments in the x− and y−directions, respectively
∂ux ∂uy
xx = yy =
∂x ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 125

The shear strain, γxy , is a measure of the change of the right angle
during the deformation
 ∂u
1 1 ∂ux

y
xy = γxy = +
2 2 ∂x ∂y

Note that this definition implies xy = yx.


February 8, 2018 3D7 - Finite Element Methods 2016/17 126

In computer implementations of the finite element method, it is more


efficient to arrange the strains in a column vector 

 
0
    ∂x
  
xx
 0 ∂  ux
 
 = yy = 
 
 ∂y 
 uy
2xy  
∂ ∂
∂y ∂x
February 8, 2018 3D7 - Finite Element Methods 2016/17 127

This equation is written in a more compact form by defining a new


gradient operator ∇S , which will be referred to as the symmetric
gradient operator

 
0
  ∂x
 
0 ∂
 
 = ∇S u with ∇S = 
 ∂y 

 
∂ ∂
∂y ∂x
February 8, 2018 3D7 - Finite Element Methods 2016/17 128

In contrast to ∇S the gradient operator which we used for single vari-


able problems was  

 ∂x 
∇= 
∂
∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 129

7.2 Stresses and Tractions

The stress state in an infinitesimal area element is visualised as shown


below
February 8, 2018 3D7 - Finite Element Methods 2016/17 130

Positive stress components act in the positive direction on a positive


face. A positive face has its normal pointing in the positive direction.
Furthermore, from moment equilibrium in a unit square, we know that
σyx = σxy .

In finite elements, the stress components are arranged in a vector


 
σxx
σ =  σyy 
σxy
February 8, 2018 3D7 - Finite Element Methods 2016/17 131

If the stresses of a body are known, the corresponding tractions t (force


per area) on the surface of the body can be computed.

An infinitesimal small triangle cut out from the boundary of the


solid. Shown are its dimensions and normal vector.
February 8, 2018 3D7 - Finite Element Methods 2016/17 132

Consider force equilibrium of a triangular element with the unit normal


n = [sin α cos α]T
February 8, 2018 3D7 - Finite Element Methods 2016/17 133

which requires that

tx · 1 = σxx sin α + σyx cos α


ty · 1 = σyy cos α + σxy sin α

This also can be written as a matrix vector multiplication


    
tx σxx σyx nx
= ⇒ t = σn
ty σxy σyy ny

This equation (Cauchy’s stress theorem) is fundamental as it provides


the crucial link between the stresses in a body and the tractions on the
surface.
February 8, 2018 3D7 - Finite Element Methods 2016/17 134

7.3 Force Equilibrium

To derive the equilibrium equations, the tractions for an infinitesimally


small area element are considered.
February 8, 2018 3D7 - Finite Element Methods 2016/17 135

The force balance (Newton’s second law) requires


∂σxx ∂σxy
+ + bx = 0
∂x ∂y
∂σyx ∂σyy
+ + by = 0
∂x ∂y
which can be written
 ∂ as ∂ 
0
 
∂x ∂y σ xx   
 σyy  + bx = 0
 


∂ ∂
 by 0
0 σxy
∂y ∂x

where b = [bx by ]T is a body force vector (per unit volume).

Using the symmetric gradient operator the equilibrium equations in


February 8, 2018 3D7 - Finite Element Methods 2016/17 136

matrix notation read


∇TS σ + b = 0

7.4 Hookes Law (Constitutive equations)

The strains and stresses are related to each other through Hooke’s law.

σ = D

The matrix D depends on whether plane stress or plain strain condi-


tions are assumed.
February 8, 2018 3D7 - Finite Element Methods 2016/17 137

A plane strain model assumes that the body is thick relative to the
xy-plane, like the following earth dam

Elasticity problem under plain strain condition.


February 8, 2018 3D7 - Finite Element Methods 2016/17 138

The displacements in the z-direction are zero and consequently zz


is zero. The D matrix for an isotropic material under plane strain
conditions is
 
1−ν ν 0
E  ν 1−ν
D= 0 
(1 + ν)(1 − 2ν)
0 0 (1 − 2ν)/2

The two independent material parameters are the Young’s modulus E


and the Poisson’s ratio ν.

Notice for ν → 0.5, which corresponds to an incompressible material,


the matrix becomes infinite. Therefore, special elements (which will
be not developed in this module) need to be used for incompressible
materials.
February 8, 2018 3D7 - Finite Element Methods 2016/17 139

A plane stress model is appropriate when the structure is relatively


thin in comparison to its dimensions in the xy-plane.

Elasticity problem under plain stress condition.

It can be assumed that the through-thickness stress σzz is zero. The


February 8, 2018 3D7 - Finite Element Methods 2016/17 140

D matrix for an isotropic material under plane stress conditions is


 
1 ν 0
E 
D= ν 1 0 
1 − ν2
0 0 (1 − ν)/2
February 8, 2018 3D7 - Finite Element Methods 2016/17 141

7.5 Strong Form of Planar Elasticity Equations

The strong form of the elasticity equation is deduced from the equations
derived above.

Equilibrium equation
∇TS σ + b = 0

Strain-displacement relation

 = ∇S u
February 8, 2018 3D7 - Finite Element Methods 2016/17 142

Stress-strain relation
σ = D

These equations can be combined into one

∇TS (D∇S u) + b = 0
February 8, 2018 3D7 - Finite Element Methods 2016/17 143

As boundary conditions either displacements or tractions are prescribed.

Two possible types of boundary conditions.


February 8, 2018 3D7 - Finite Element Methods 2016/17 144

On parts of the boundary Γu (Dirichlet boundary) the displacements


are prescribed and on the parts Γt (Neumann boundary) the tractions
are prescribed

u = u on Γu
σn = t on Γt

7.6 Weak Form of Planar Elasticity Equations

The weak form of the elasticity equations is the principle of virtual


work which you already know from Part I of your studies.
February 8, 2018 3D7 - Finite Element Methods 2016/17 145

In the following an alternative approach for deriving the weak form


along the lines of the weak form derivation for the heat equation (section
2.3) is sketched.

To derive the weak form the equilibrium equations are first multiplied
 T
with a test function vector w = wx wy and then integrated over
the domain Ω Z
wT (∇TS σ + b) dΩ = 0

After integrating the first term by parts and applying the divergence
theorem we obtain the equilibrium equations in weak form
Z Z Z
(∇S w)T σ dΩ = wT t dΓ + wT b dΩ
Ω Γt Ω
Introducing the stress-strain and strain-displacement relationships the
February 8, 2018 3D7 - Finite Element Methods 2016/17 146

weak form can be written in terms of the displacements


Z Z Z
(∇S w)T D∇S u dΩ = wT t dΓ + wT b dΩ
Ω Γt Ω

7.7 Finite Element Discretization

The displacement components ux and uy are two independent unknown


variables and will be approximated using the same shape functions.
There are two-degrees of freedom per finite element node corresponding
to the two displacement components.
February 8, 2018 3D7 - Finite Element Methods 2016/17 147

Nodal degrees of freedom for the three-node triangle and four-node


quadrilateral element.
February 8, 2018 3D7 - Finite Element Methods 2016/17 148

Over one finite element, the displacements and test functions are ap-
proximated with the element shape functions and the corresponding
nodal values.
uex(x, y) = N1e(x, y)uex1 + N2e(x, y)uex2 + · · · + Nneen (x, y)uexnen
uey (x, y) = N1e(x, y)uey1 + N2e(x, y)uey2 + · · · + Nneen (x, y)ueynen

wxe (x, y) = N1e(x, y)wx1


e + N e(x, y)w e + · · · + N e (x, y)w e
2 x2 nen xnen
wye (x, y) = N1e(x, y)wy1
e + N e(x, y)w e + · · · + N e (x, y)w e
2 y2 nen ynen
where nen is the number of nodes of one element. After introducing
the shape function matrix
 
N1e0 N2e 0 · · · Nneen 0
Ne =
0 N1e 0 N2e · · · 0 Nneen
February 8, 2018 3D7 - Finite Element Methods 2016/17 149

and the nodal vectors


e
 e e e e e e
T
a = ux1 uy1 ux2 uy2 · · · uxnen uynen
e
 e e e e e e
T
w = wx1 wy1 wx2 wy2 · · · wxnen wynen

the previous approximation equations can be written in a more compact


form

ue(x, y) = N e(x, y)ae


we(x, y) = N e(x, y)we
February 8, 2018 3D7 - Finite Element Methods 2016/17 150

Next, the domain and boundary integrals in the weak form are com-
puted by summing up the element contributions.
nel Z Z Z !
X T T
(∇S we)T D∇S ue dΩ − e
w t dΓ − e
w b dΩ =0
e=1 Ωe Γet Ωe

In terms of the shape functions and the nodal displacements the strains
are
e = ∇S ue = ∇S N eae = B eae
February 8, 2018 3D7 - Finite Element Methods 2016/17 151

The strain-displacement matrix B e is defined as



 
0
 ∂x

  
e e e
 0 ∂  N1 0 N2 0 · · · Nnen 0
 
e e
B = ∇S N = 
∂y  0 N1e 0 N2e · · · 0 Nneen


 
∂ ∂
∂y ∂x
February 8, 2018 3D7 - Finite Element Methods 2016/17 152

∂N1e ∂N2e ∂Nneen


 
 ∂x 0 ∂x 0 · · · ∂x 0

 
e e e
 
 ∂N1 ∂N2 ∂Nnen 
= 0 0 ··· 0 
 ∂y ∂y ∂y 
 
 
 ∂N e ∂N e ∂N e ∂N e ∂N e ∂N e 
1 1 2 2 nen nen
···
∂y ∂x ∂y ∂x ∂y ∂x
The strains corresponding to the weight functions are
T T T T
e e e e
(∇S w ) = (B w ) = w B e

Introducing the discretized strains and weight functions into the weak
February 8, 2018 3D7 - Finite Element Methods 2016/17 153

form leads to
nel Z Z Z !
X T T T T T T
e
w B DB eae dΩ −
e e e
w N t dΓ − e e
w N b dΩ =
e=1 Ωe Γet Ωe

Next recall the mapping between the global degrees of freedom and
local degrees of freedom (see Section 4.1)
T T
ae = Lea e T
w =w L e

and the discrete weak form can be written as


" nel Z Z Z !#
X T T T T
wT L e B DB e dΩ Lea −
e e
N t dΓ − e
N b dΩ =
e=1 Ωe Γet Ωe
February 8, 2018 3D7 - Finite Element Methods 2016/17 154

The element stiffness matrix is given by


Z
T
Ke = B D eB e dΩ
e
Ωe

and the element external force vector is


Z Z
T T
fe = e
N b dΩ + e
N t dΓ
| Ωe
{ze } | Γet
{ze }
fΩ fΓ
February 8, 2018 3D7 - Finite Element Methods 2016/17 155

Hence, the weak form can be rewritten as


 
 nel ! n
!
el

 X T X T 
T
w  L e K e L e a − L e f e =0
 
| e=1 {z } | e=1 {z }
K f
or
wT (Ka − f ) = 0

As the nodal values of the test functions (in other terms the virtual dis-
placements) are arbitrary, this leads to the final finite element equation

Ka = f
February 8, 2018 3D7 - Finite Element Methods 2016/17 156

7.8 Three-Node Triangular Element

For a three-node triangular element there are six nodal displacements


per element so that
T
ae = uex1 uey1 uex2 uey2 uex3 uey3


The displacement field within the element is expressed with


   
ux N1e 0 N2e 0 N3e 0  T
= uex1 uey1 uex2 uey2 uex3 uey3
uy 0 N1e 0 N2e 0 N3e
February 8, 2018 3D7 - Finite Element Methods 2016/17 157

The B e matrix for this element is computed using the shape functions
given on the Data Sheet
 
(y2e − y3e) 0 (y3e − y1e) 0 (y1e − y2e) 0
e 1  e − xe ) e − xe ) e − xe ) 
B = 0 (x 0 (x 0 (x
2Ae 3 2 1 3 2 1
(xe3 − xe2) (y2e − y3e) (xe1 − xe3) (y3e − y1e) (xe2 − xe1) (y1e − y2e)

Since B e is constant within one element, the element stiffness matrix


is readily integrated

K e = Ae(B e)T DB e
February 8, 2018 3D7 - Finite Element Methods 2016/17 158

7.9 Appendix: Derivation of the Weak Form of Planar Elasticity


Equations

The starting point of the derivations are the equilibrium equations in


component form
∂σxx ∂σxy
+ + bx = 0
∂x ∂y
∂σyx ∂σyy
+ + by = 0
∂x ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 159

First, the equilibrium equation in the x-direction is premultiplied with a


test function wx, which is zero on prescribed displacement boundaries,
and integrated over the entire domain
Z
∂σxx ∂σxy
 
wx + + bx dΩ = 0 (7.1)

∂x ∂y

We then apply integration by parts to the two first terms


Z h Z h
∂ ∂ ∂σxx ∂σxy
i  i
(wxσxx) + (wxσxy ) dΩ = wx + dΩ

∂x ∂y Ω
∂x ∂y
Z h
∂wx ∂wx
i
+ σxx + σxy dΩ

∂x ∂y
(7.2)
February 8, 2018 3D7 - Finite Element Methods 2016/17 160

and apply the divergence theorem to the integral on the left hand side
Z h Z
∂ ∂
i 
(wxσxx) + (wxσxy ) dΩ = wx σxxnx + σxy ny dΓ

∂x ∂y Γ | {z }
tx
(7.3)
Inserting equations 7.3 and 7.2 into equation 7.1 and introducing the
traction boundary conditions gives
Z h Z Z
∂wx ∂wx
i
σxx + σxy dΩ = wxbx dΩ + wxtx dΓ (7.4)

∂x ∂y Ω Γt
Similarly, we can multiply the equilibrium equation in the y-direction
with a second test function wy and obtain
Z h Z Z
∂wy ∂wy
i
σyx + σyy dΩ = wy by dΩ + wy ty dΓ (7.5)

∂x ∂y Ω Γt
February 8, 2018 3D7 - Finite Element Methods 2016/17 161

Adding the equations 7.4 and 7.5 gives the weak form in the component
notation
 
Z h σxx Z Z
∂wy ∂wy ∂wx ∂wy
  i
+ σyy  dΩ = wxbx dΩ + wxtx dΓ

∂x ∂y ∂y ∂x Ω Γt
σxy
Z Z
+ wy by dΩ + wy ty dΓ
Ω Γt
3D7 - Finite Element Methods 2016/17

Chapter 8

Finite Elements for Planar Linear Elasticity - Implementation


and Modelling

Annotated version
February 8, 2018 3D7 - Finite Element Methods 2016/17 163

In the first part of this chapter, we will consider the practical aspects
of the discretization of the elasticity equations.

In the second part we will look at numerical case studies to discuss a


few common pitfalls of linear elastic finite element analysis.
February 8, 2018 3D7 - Finite Element Methods 2016/17 164

8.1 Illustrative Example

Consider the problem of a cantilever plate

Problem domain and boundary conditions.


February 8, 2018 3D7 - Finite Element Methods 2016/17 165

The Young’s modulus is E = 70 · 109 P a, the Poisson’s ratio is (ν =


0.33) and plane stress is assumed
   
1 ν 0 1 0.33 0
E   = 7.86 · 1010 0.33 1 0 
D= ν 1 0
1 − ν2
0 0 1−ν
2 0 0 0.34
February 8, 2018 3D7 - Finite Element Methods 2016/17 166

The domain is discretized with two three-noded triangular elements.

Finite element mesh with two three-noded elements.


February 8, 2018 3D7 - Finite Element Methods 2016/17 167

The node numberings and degrees of freedom (in brackets) for the
two elements. The local ones are indicated within the elements
and the global ones outside the elements.
February 8, 2018 3D7 - Finite Element Methods 2016/17 168

8.1.1 Global Stiffness Matrix

First we compute the element stiffness matrices. The strain-displacement


matrix B e for a three-node triangle was given in Section 7.8
 
xx
 =  yy  = B eae
2xy

with nodal displacements

ae = uex1 uey1 uex2 uey2 uex3 uey3


 

and
February 8, 2018 3D7 - Finite Element Methods 2016/17 169

 
(y2e − y3e)
0 0 (y3e − y1e) 0 (y1e − y2e)
e 1  e − xe ) e − xe ) e − xe ) 
B = 0 (x 0 (x 0 (x
2Ae 3 2 1 3 2 1
(xe3 − xe2) (y2e − y3e) (xe1 − xe3) (y3e − y1e) (xe2 − xe1) (y1e − y2e)

The coordinates of the first element inserted into the above expression
gives  
0 0 1.0 0 −1.0 0
1 1
B = 0 −2.0 0 0.0 0 2.0 
2
−2.0 0.0 0.0 1.0 2.0 −1.0
As B 1 is constant the stiffness matrix is easily integrated

K 1 = A1(B 1)T DB 1
February 8, 2018 3D7 - Finite Element Methods 2016/17 170

The stiffness matrix for element 1 is as follows (for clarity the global
degrees of freedom have been written next to and below the matrix)
 
2.63 0.00 0.00 −1.32 −2.63 1.32 1
 0.00 7.86 −1.30 0.00 1.30 −7.86 2
 
1 9
 0.00 −1.30 1.97 0.00 −1.97 1.30 7
K = 10  
−1.32 0.00 0.00 0.66 1.32 −0.66 8
 
−2.63 1.30 −1.97 1.32 4.60 −2.61 5
1.32 −7.86 1.30 −0.66 −2.61 8.51 6
1 2 7 8 5 6

Note that in the element stiffness matrix the first two rows (and also
the first two columns) correspond to the two degrees of freedom (1, 2)
of the node 1, i.e. u1x1 and u1y1, and so on.
February 8, 2018 3D7 - Finite Element Methods 2016/17 171

The stiffness matrix for element 2 and the correspondence between the
local and global degrees of freedom is
 
1.97 0.00 −1.97 1.30 0.00 −1.30 1
 0.00 0.66 1.32 −0.66 −1.32 0.00 2
 
2 9
−1.97 1.32 4.60 −2.61 −2.63 1.30  3
K = 10  
 1.30 −0.66 −2.61 8.52 1.32 −7.86 4

 0.00 −1.32 −2.63 1.32 2.63 0.00 7
−1.30 0.00 1.30 −7.86 0.00 7.86 8
1 2 3 4 7 8
February 8, 2018 3D7 - Finite Element Methods 2016/17 172

The assembly procedure of the global stiffness matrix is equivalent to


the single degree of freedom per node problems.

In the present example there are four nodes so that there are eight
degrees of freedom. The rows and columns corresponding to global
degrees of freedom (1, 2) and (5, 6) can be deleted because of the zero
displacement boundary conditions
 
4.60 −2.61 −2.63 1.30 3
9
−2.61 8.52 1.32 −7.86 4
K = 10  
−2.63 1.32 4.60 0.00 7
1.30 −7.86 0.00 8.51 8
3 4 7 8
February 8, 2018 3D7 - Finite Element Methods 2016/17 173

8.1.2 External Force Vector

Only element 2 has an edge on which a non-zero traction is prescribed.


The external force vector for element 2 is
Z
T
f 2Γ = 2
N t dΓ
Γ2t

where N 2 is the shape function matrix


 
N12 0 N22 0 N32 0
N2 =
0 N12 0 N22 0 N32
February 8, 2018 3D7 - Finite Element Methods 2016/17 174

To compute the line integral we evaluate these shape functions on the


edge 2-3 (in local numbering).

As introduced for scalar problems in Section 4.2.3, to compute the line


integral we consider the edge between nodes 2 and 3 as a one dimensional
element.

One dimensional shape functions on edge 2-3.


February 8, 2018 3D7 - Finite Element Methods 2016/17 175

   
0 0
 0  0
Z 1   
2
(1 − t) · 10 5
fΓ =   dt =  
(1 − t) · 20 10
0    
 t · 10  5
t · 20 10

The first two rows in the above vector are zero because the shape
function belonging to node 1 is identical to zero along edge 2-3.

It is evident that the computed external nodal forces are equivalent to


the integral of the external traction vector.
February 8, 2018 3D7 - Finite Element Methods 2016/17 176

The global external force vector is assembled using the mapping between
the local and global degrees of freedom.
 
5
10
f = 
5
10
February 8, 2018 3D7 - Finite Element Methods 2016/17 177

8.1.3 Solution of Ka = f

The resulting global system of equations is given by


    
4.60 −2.61 −2.63 1.30 ux2 5
−2.61 8.52 1.32 −7.86   uy2
 10 
109    =  
−2.63 1.32 4.60 0.00 ux4  5 
1.30 −7.86 0.00 8.51 uy4 10
Solving this equation system gives
   
ux2 0.82
uy2  2.61
  = 10−9 
 

ux4 −0.17
uy4 2.41
February 8, 2018 3D7 - Finite Element Methods 2016/17 178

8.2 Stiffness Matrix Integration of Quadrilaterals

The number of integration points is chosen so that the element stiffness


matrix of an undistorted element (with constant Jacobian) is accurately
integrated. For a quadrilateral element the following rules were given

Element Number of integration points


four-noded 2×2
nine-noded 3×3

However if these rules are used, the stiffness of the discretized structure
is often overestimated compared to the real stiffness of the structure
(i.e. the computed displacements are much smaller than the true dis-
placements).
February 8, 2018 3D7 - Finite Element Methods 2016/17 179

Slender plate in pure bending.

Since the plate is slender, simple beam theory can be used to compute
its stresses.
M
σxx = y
I
February 8, 2018 3D7 - Finite Element Methods 2016/17 180

We discretize the slender plate with five four-noded elements with one
element across it’s height.

Slender plate in pure bending discretized with five four-noded


elements.
February 8, 2018 3D7 - Finite Element Methods 2016/17 181

The four-noded element is not able to reproduce the stress state of pure
bending. The shape functions of the four-noded element allow only for
displacements of the form

Deformation of a quadrilateral element in pure bending.

This deformation implies, in addition to correct linear variation of the


strains xx and stresses σxx, spurious shear strains xy and stresses
February 8, 2018 3D7 - Finite Element Methods 2016/17 182

σxy .

One remedy (which is used in commercial software packages) is to use


only one integration point instead of 2 × 2 integration points.

Quadrilateral element in pure bending with one integration point


(left) and 2 × 2 integration points (right).
February 8, 2018 3D7 - Finite Element Methods 2016/17 183

(From here non-examinable) On the other hand, the use of too few in-
tegration points (or so-called reduced integration) leads to numerical
instabilities due to mechanisms in the discretized structure. Mecha-
nisms occur if there are deformation modes possible which display zero
stress at all Gauss points, like the following hourglass instability

Undeformed plate (left) and a possible mechanism for one point


integration.
February 8, 2018 3D7 - Finite Element Methods 2016/17 184

The discretized structure can be deformed into this state without ap-
plying any external loading. In practice these modes may appear su-
perposed on ’realistic’ deformations but the results are unusable.

Example computed with 2 × 2 integration points (left) and one


integration point (right).

In commercial software packages, which use one-point integration, so-


called stabilised elements are used to avoid hourglass instabilities.
3D7 - Finite Element Methods 2016/17

Chapter 9

Assembly of Finite Elements (revisit)

The finite element method is best illustrated with the solution of a


simple elliptic Partial Differential Equation in a two-dimensional space.
Consider the Poisson’s equation
 
∂ 2u ∂ 2u
− 2
+ 2 = f in Ω (9.1)
∂x ∂y
February 8, 2018 3D7 - Finite Element Methods 2016/17 186

with the Dirichlet boundary condition

u = 0 on Γ = ∂Ω (9.2)

where Ω is a bounded open domain. The weak formulation is


Z Z
∇u · ∇vdx = f vdx , (9.3)
Ω Ω
where v is a weight function.

The finite element method consists of approximating the weak solution


by a function in a subspace of functions Vh that are defined as low-
degree polynomials on small pieces (elements) of the original domain.

Consider a region Ω in the plane which is triangulated as shown in


Fig. 9.1. The original domain is thus approximated by the union Ωh of
February 8, 2018 3D7 - Finite Element Methods 2016/17 187

i k
j

Figure 9.1: Finite Element triangulation of a domain.


February 8, 2018 3D7 - Finite Element Methods 2016/17 188

m triangles Ki,
m
[
Ωh = Ki (9.4)
i=1
Then the finite dimensional space Vh is defined as the space of all func-
tion which are piecewise linear and continuous on the polygonal re-
gion Ωh, and which vanish on the boundary Γ. More specifically,

Vh = {φ | φ|Ωh continuous, φ|Γh = 0, φ|Ki linear} (9.5)

If xi, i = 1, ... , n are the nodes of the triangulation, then a function


φi in Vh can be associated with each node xi, so that the family of
functions φi’s satisfies the following condition:

1 if i = j,
φi(xj ) = δij = (9.6)
0 if i 6= j.
February 8, 2018 3D7 - Finite Element Methods 2016/17 189

These conditions define φi, i = 1, ... , n uniquely. In addition, the φi’s


form a basis of the space Vh. Each function of Vh can be expressed as
n
X
φ(x) = Tiφi(x) (9.7)
i=1
where Ti is the nodal value of φ at the vertex xi. The finite element
approximation consists of writing the the weak formulation for for func-
tions in Vh. This defines the approximate problem:
Z Z
Find uh ∈ Vh such that ∇uh ·∇vhdx = f vhdx , vh ∈ Vh.
Ω Ω
(9.8)
Since uh is in Vh, there are n degrees of freedom. By the linearity with
respect to v, it is only necessary to choose vh = φi for i = 1, ... , n.
This results in n equations.
February 8, 2018 3D7 - Finite Element Methods 2016/17 190

Writing the desired solution u in the basis {φj } as


n
X
u= Tjφj(x) (9.9)
j=1

and substituting in (9.8) gives the linear problem


n
X
Aij Tj = bi (9.10)
j=1

where Z Z
Aij = ∇φi∇φj dx, bi = f φidx (9.11)
Ω Ω
The above equations form a linear system of equations

Ax = b (9.12)
February 8, 2018 3D7 - Finite Element Methods 2016/17 191

in which the coefficients of A are Aij ’s; those of b are the bi’s. In
addition, A is a Symmetric Positive Definite matrix.

Another important observation is that the matrix A is also sparse.


Indeed, Aij is nonzero only when the two basis functions φi and φj
have common support triangles, or equivalently when nodes i and j
are the vertices of a common triangle. Specifically, for a given node i
(marked by a full circle in the figure), the coefficient Aij will be nonzero
only when the node j (marked by empty circles) is one of the nodes of
a triangle that is adjacent to node i.

In practice, the matrix is built by summing up the contributions of all


February 8, 2018 3D7 - Finite Element Methods 2016/17 192

triangles, by applying the formula


Z X
∇φi · ∇φj dx = aK (φi, φj ) (9.13)
Ω K
in which the sum is over all the triangles K and
Z
aK (φi, φj ) = ∇φi · ∇φj dx (9.14)
K
Note that aK (φi, φj ) is zero unless the nodes i and j are both vertices
of K. Thus, a triangle contributes to its three vertices from the above
formula. The 3 × 3 matrix
 
aK (φi, φi) aK (φi, φj ) aK (φi, φk )
AK =  aK (φj , φi) aK (φj , φj ) aK (φj , φk )  (9.15)
aK (φk , φi) aK (φk , φj ) aK (φk , φk )
February 8, 2018 3D7 - Finite Element Methods 2016/17 193

associated with the triangle K(i, j, k) with vertices i, j, k is called


element stiffness matrix In order to form matrix A, it is necessary to
sum up all the contributions aK (φi, φj ) to the position i, j position of
the matrix. This process is called an assembly process. In the assembly,
the matrix is computed as
nel
X
A= A[e] (9.16)
e=1

in which nel is the number of the elements. Each matrix A[e] is of the
form
A[e] = P eAKe P Te (9.17)
where AKe is the element matrix for the element Ke as defined above.
Also P e (is the transpose of the scatter matrix Le in previous chapters)
February 8, 2018 3D7 - Finite Element Methods 2016/17 194

is an n × 3 Boolean connectivity matrix which maps the coordinates of


the 3 × 3 matrix AKe into the coordinates of the full matrix A.

The assembly process can be illustrated with a very simple example.


Consider the finite element mesh shown in Fig 9.2. The four elements
are numbered from bottom to top as indicated by labels located at their
centers. There are six nodes in this mesh and their labeling is indicated
in the circled numbers. The four matrices A[e] associated with these
elements are shown in Fig. 9.3. Thus the first element will contribute
to the nodes 1, 2, 3, the second nodes 2, 3, 5, the third 2, 4, 5, and the
fourth to nodes 4, 5, 6.

In fact there are two different ways to represent and use the matrix
A. We can form all the element matrix one by one and then store
February 8, 2018 3D7 - Finite Element Methods 2016/17 195

them, e.g., in an nel × 3 × 3 rectangular array. This representation is


often called the unassembled form of A. Then the matrix A may be
assembled if it is needed. However, element stiffness matrices can also
be used in different ways without having to assemble the matrix.

A more common and somewhat more appealing technique is to perform


the alert assembly of the matrix. All the elements are scanned one by
one and the nine associated contributions aK (φk , φm), k, m ∈ {i, j, k}
added to the corresponding positions in the global “stiffness” matrix.
The assembled matrix must now be stored but the element matrices
may be discarded. The structure of the assembled matrix depends on
the ordering of the nodes. To facilitate the computations, a widely used
strategy transforms all triangles into a reference triangle with vertices
(0, 0), (1, 0) and (0, 1). The area of the triangle is then simply the
February 8, 2018 3D7 - Finite Element Methods 2016/17 196

Finite element mesh


Assembled matrix
6
111
000
000
111 000
111
000
111 000
111
000
111
4
000
111
000
111 000
111
000
111 000
111
000
111 000
111
000
111 000
111
000
111
4 5
000
111 000
111 000
111 000
111
3
000 111
111 000
000
111 000
111 000
111 000
111
000
111 000
111
2 000
111
000
111 000
111 000
111
000
111 000
111
000
111 000
111
000
111
2 3
000 111
111 000 111
000
000
111 000
111
000
111 000
111
000 111
000
1
000 111
111 000
111
1

Figure 9.2: A simple finite element mesh and the pattern of the corre-
sponding assembled matrix

determinant of the Jacobian of the transformation that allows passage


from one set of axes to the other.

Simple boundary conditions such as Neumann or Dirichlet conditions


do not cause any difficulty.
February 8, 2018 3D7 - Finite Element Methods 2016/17 197

[1] [2] [3] [4]


A A A A
00
11000
11111000
111
00111
11000
000000
111000
111 111
000
000111
111000
000
111 111
000
000
111 111
000
000
111 111
000
000111
111000
000
111
00111
11000111
000 000111
111000 000
111 000
111 000
111000
111 111
000
000
111 000
111000
111 000 11001100111
111 000
000
111
111 11001100
000 000
111
000
111 000
111 000
111
000111
111000
000
111 111
000
000
111 0011000 111
000 11001100111
111 000
000
111

Figure 9.3: The element matrix A[e], e = 1, ..., 4 for the finite element
mesh in Fig. 9.2.