CI3-321 Computational Engineering Analysis

Finite Element Analysis

a.y. 2015-16
Contents

1. Introduction to Finite Element Analysis ........ 1

1.1. Basics of Finite Element Analysis ....... 1

1.2. Assembly and Solution of Discrete Systems ........ 2

Example E 1.1. ...... 4

Example E 1.2. ...... 7

1.3. Formulation of Finite Elements ........ 8

Example E 1.3. ...... 9

1.4. Characteristics of Finite Element Predictions ..... 11

Example E 1.4. ..... 11

2. Plane Stress/Strain Elements ......... 13

2.1. Plane Stress vs. Plane Strain ....... 13

2.2. Continuity Requirements of Shape Functions ....... 14

2.3. Constant Strain Triangular Element ...... 15

2.3.1 Local and global reference systems ...15

2.3.2 Shape functions ..... 15

2.3.3 Strains .... 17

2.3.4 Local element response 17

2.3.5 Global element response ... 18

2.3.6 Accuracy of triangular element . 19

2.4. Basic Rectangular Element ....... 20

2.4.1 Shape functions ..... 20

 2.4.2 Strains .... 21

2.4.3 Local element response 22

2.4.4 Benefit of rectangular element . 22

3. Higher Order Elements and Error Estimation 23

3.1. Higher Order Shape Functions ........... 23

3.2. Error Estimation for Finite Element Analysis ....... 25

3.2.1 Sources of Error ....25

3.2.2 Discretisation errors ...... 25

Appendix .. 27

A. Mixing of Two-Dimensional Elements of Different Orders .................. 27

B. Accuracy of Finite Element Prediction Based on Stress/Strain Smoothness .. 27

C. Determination of Individual Terms of [k] .......... 28

ii

1. Introduction to Finite Element

Analysis

1.1. Basics of Finite Element Analysis

The finite element method is a numerical technique for solving boundary value problems
in numerous fields of engineering, including structures, geotechnics and fluid mechanics.
This method of analysis is generally approximate, but is typically applied to real engineering
problems which have no exact analytical solutions. It involves the discretisation of the
continuum problem, shifting the focus from the infinite to the finite. In this respect, the
displacement-based finite element method, which is applicable to structural and geotechnical
analysis, is concerned with determining unknown displacements at a finite number of nodes
(while the exact analytical solution involves the determination of one or more displacement
functions from an infinity of possibilities). In order to discretise the continuum problem,
shape functions are used to relate the displacement fields within the problem domain to nodal
displacements. Considerable computational benefits arise if the shape functions are chosen
such that the displacements at a node influence the immediate region surrounding the node,
leading to the idea of an element which is only influenced by its associated nodes.

Equivalent Equivalent
nodal forces nodal forces

Bounday value problem Mesh of 3-noded Mesh of 6-noded

triangular elements triangular and 8-noded
rectangular elements

CI3-321: Finite Element Analysis 1/28

The boundary value problem involves known boundary conditions:
Displacements at the support boundary (essential BCs)
Forces at the remaining boundary and within the domain (natural BCs)
When the problem is discretised using finite elements, the known conditions become:
Nodal displacements at the support boundary (essential BCs)
Equivalent nodal forces elsewhere (natural BCs)
In the original continuum problem, the displacement fields are normally the primary
unknowns, and once established these can be used to determine secondary unknowns such as
strains and stresses: compatibility conditions relate the strain and displacement fields,
whereas the governing constitutive law relates the stress and strain fields. Upon discretisation
into finite elements, the nodal displacements become the primary unknowns, and the
displacement fields become secondary unknowns which are related to the nodal
displacements using pre-defined shape functions.
The accuracy and convergence characteristics of a finite element are largely determined by
the order of its shape functions, which in turn depends on the number of element nodes: the
more nodes an element has the more sophisticated its shape functions and the better its
approximation over a specific region become. Nevertheless, even very basic finite elements
possess the important characteristic of convergence to the exact analytical solution with mesh
refinement. This fact, along with the ease of computer implementation and application of the
finite element method, is largely responsible for the widespread use of finite element analysis
in engineering practice for a variety of problems with different degrees of complexity.

1.2. Assembly and Solution of Discrete Systems

Because of the complexity of real engineering systems, their analysis is most naturally
performed through subdividing the actual system into a number of components whose
behaviour is individually understood, and then reassembling the whole system to study its
response under prescribed boundary conditions. When the subdivision process results in a
finite number of components or elements, the corresponding model is called a discrete
system. If the subdivision process is continued until the problem is expressed infinitesimally,
such as through differential equations, the corresponding model is called a continuous system.
A finite element model represents a discrete system which approaches the continuous system
in the limit as the number of elements is increased. Hereafter, the characteristics of a discrete
system are discussed with reference to the displacement-based finite element (FE) method.

The response of a FE discrete system is defined by a set of nodal displacement parameters

{U } , representing the primary problem unknowns. These unknowns are determined from a
corresponding set of equations representing the nodal equilibrium conditions:
{P} {R} (1.1)
where {P} is the directly applied nodal forces, and {R} is the system nodal resistance forces.
For a linear response, {R} can be related to {U } as follows:

CI3-321: Finite Element Analysis 2/28

 {R} [ K ]{U } {Pp } (1.2)
in which [ K ] is a constant stiffness matrix, and {Pp } represents equivalent nodal forces due
to distributed loading, initial lack of fit, initial thermal strains, etc..
Therefore, considering (1.1) and (1.2), {U } can be determined from the solution of linear
simultaneous equations:
[ K ]{U } {P} {Pp } (1.3)
Considering now that the nodal resistance for an individual element e can be expressed as:
{R}e [K]e {U}e {Pp }e (1.4)
and that the overall system nodal resistance is simply the assembly of element contributions:
{R} {R}e (1.5)
e
the overall system [ K ] and {Pp } , as defined in (1.2), can be assembled from element
contributions:
[ K ] [K]e (1.6.a)
{Pp } {Pp }
e e
(1.6.b)
e
It is noted that the element nodal displacements {U}e are a subset of the system nodal
displacements {U } , and hence the assembly in (1.6) is not a direct summation, since [ K ] and
{Pp } are of different respective sizes to [K]e and {Pp }e . Assuming that a system includes m
elements, that element e has n e displacement parameters, and that the incidence vector {}e
denotes the system parameter numbers associated with each of the n e element parameters,
then the following assembly procedure can be used:
Initialise all terms of [ K ] and {Pp } to zero.
Loop over all elements using element counter e:
e
Form the element stiffness matrix [K] and equivalent nodal force vector
{Pp } with dimensions n n and n , respectively.
e e e e

Loop over rows of [K]e and {Pp }e using counter i:

o i {}ie , {Pp }i {Pp }i {Pp }ie
o For the current row (i), loop over columns of using counter j:
j {}ej , [ K ]i, j [ K ]i, j [K]i,e j
o Continue the j-loop until ( j n e ), increasing j by 1.
Continue the i-loop until ( i n e ), increasing i by 1.
Continue the e-loop until ( e m ), increasing e by 1.
Since nodes represent points over the system domain where displacement parameters are
defined, the assembly procedure can also be described on a nodal level, provided that all
e
nodes are associated with the same number of parameters, and that the element [K] and
{Pp }e account for all parameters of the element nodes. Therefore, if p represents the number
of parameters per node and n e is redefined as the number of nodes for element e, then the
sizes of [K] and {Pp }e must be pn e pn e and pn e , respectively. In the nodal assembly
e

procedure, {} is redefined to denote the system node numbers associated with each of the
e

CI3-321: Finite Element Analysis 3/28

n e element nodes. Counters i and j become associated with nodes, and therefore two further
nested sub-loops are required to consider the parameters of each node.
When [ K ] and {Pp } have been assembled, the equilibrium conditions in (1.3) can be solved
for {U } . Often, however, the assembled terms refer to all system parameters, regardless of
whether natural (equilibrium) or essential (support) boundary conditions apply for such
parameters. Since the natural and essential boundary conditions are mutually exclusive in
relation to specific parameters, there is always an equal number of natural boundary
conditions and unknown parameters, and the problem can be solved by considering only
those equations of (1.3) related to the natural boundary conditions. In the context of a
computer-based procedure, this is most conveniently performed by modifying [ K ] and
{P} {Pp } such that the essential boundary conditions replace the unwanted natural
boundary conditions as follows:
Loop over all parameters using counter i:
If parameter i is associated with an essential boundary condition {U }i a i :
o Set [ K ]i,i 1 and all other terms of the ith row of [ K ] equal to zero.
o Set {P}i {Pp }i a i .
Continue the i-loop until all parameters are considered, increasing i by 1.
Following the above modifications, the system of equations represented by (1.3) can be
solved for all parameters {U } . An alternative approach which is more appropriate for manual
calculation is to assemble only those rows of [ K ] and {Pp } associated with the natural
boundary conditions, including the columns of [ K ] associated with the unknown parameters
and the non-zero known parameters of {U } .

Example E.1.1
Consider an axial member with a varying
cross-sectional area (A) subjected to a
concentrated axial force (P) and a uniformly
distributed axial loading (p). The variation of
the cross-sectional area is assumed to be
quadratic as given by:
2
x
A A0 A L A0 (E.1)
L
where A 0 and A L are the cross-sectional
area at ( x 0 ) and ( x L ), respectively.
Continuous model
The cross-sectional axial force (F) can be expressed in terms of the first derivative of u with
respect to x:
F E A u (E.2)
Equilibrium of an infinitesimal length (dx) of the member can be expressed as:

CI3-321: Finite Element Analysis 4/28

 dF
(F dF) F p dx 0 p 0
dx
E A u A u p 0 (E.3)

The continuous model can therefore be represented by the second-order differential equation
in (E.3) subject to the two boundary conditions:
P
u (0) , u(L) 0 (E.4)
E A0
A first-order differential equation can also be established, either by integration of (E.3) or by
considering equilibrium at x:
F P px 0
E A u P p x 0 (E.5)
The continuous model can therefore also be represented by the first-order differential
equation in (E.5) subject to the boundary condition:
u(L) 0 (E.6)
Taking the special case of a member with ( AL 2A0 ) and ( p 2P / L ), the exact analytical
solution to (E.5) subject to (E.6) is as follows:

u(x)
PL tan 1 x log 2
(E.7)
E A0 4 L 1 x / L 2

Values of u(x) at ( x 0 ) and ( x L / 2 ) become:
PL PL
u 0 1.47855 , u L / 2 0.791754 (E.8)
E A0 E A0
Discrete model
Assume that the previous system is discretised into two elements, each employing two
degrees of freedom, as shown below.
R1 U1
R1e U1e
R R2 , U U 2 e ,
e

e e
R U
R U R2 U1
3 3

CI3-321: Finite Element Analysis 5/28

The response of each element can be represented by a relationship between the end forces
and the corresponding displacements:

E A1e Ae2
R1e
2Le
U e
1
Ue2
pLe
2
(E.9.a)

E A1e Ae2
R e2
2Le
U1e Ue2 pLe
2
(E.9.b)

Note that the above relationship is approximate for each of the discrete elements, although a
similar but more accurate relationship can be derived. Matrix notation can be used to express
such relationships, which for (E.9) leads to:
{R}e [K]e {U}e {Pp }e (E.10.a)

where,

[K]

E A1e Ae2 1 1
e
1 1 (E.10.b)
2Le
pLe 1
{Pp }e (E.10.c)
2 1
For the special case of a member with ( AL 2A0 ) and ( p 2P / L ), the response of each
element can be determined in terms of the system parameters.
Element 1: A11 A0 , A12 1.25A0 , L1 L / 2, U11 U1 , U12 U 2
2.25E A0 P
R11 U1 U 2 (E.11.a)
L 2
2.25E A0 P
R12 U1 U 2 (E.11.b)
L 2
Element 2: A12 1.25A0 , A22 2A0 , L2 L / 2, U12 U 2 , U22 U3
3.25E A0 P
R12 U 2 U 3 (E.12.a)
L 2
3.25E A0 P
R 22 U 2 U 3 (E.12.b)
L 2
The overall system response can be assembled from element contributions as follows:
2.25E A0 P
R1 R11 U1 U 2 (E.13.a)
L 2
E A0
R2 R12 R12 2.25U1 5.5U 2 3.25U3 P (E.13.b)
L
3.25E A0 P
R3 R 22 U 2 U 3 (E.13.c)
L 2

CI3-321: Finite Element Analysis 6/28

which in matrix notation is identical to (1.2) with:
2.25 2.25 0 P / 2
E A0
[K ] 2.25 5.5 3.25 , {Pp } P (E.14)
L
0 3.25 3.25 P / 2
The assembly procedure described in the previous section leads to an identical result
considering the individual element contributions with the following incidence vectors:
1 2
{}1 , {}2 (E.15)
2 3
Now considering the two natural and one essential boundary conditions:
P natural U 1

{P} 0 natural , {U } U 2 (E.16)
P 0 essential
3
the two equilibrium equations ( R1 P, R2 0 ) with the support condition ( U 3 0 ) can be
used to obtain:
PL PL
U1 1.4359 , U 2 0.7692 (E.17)
E A0 E A0
which are only 3% smaller than the corresponding exact values obtained from the continuous
model.

Example E.1.2
Consider the adjacent system discretised into a
triangular element, a quadrilateral element and
two line elements. The various stages of
assembly are shown below, where it is
assumed that each node is associated with only
one parameter (i.e. p 1 ). For the case of
( p 1), each term in {Pp } should be replaced by a p sub-vector from {Pp }e corresponding to
node i, and each term in [ K ] should be replaced by p p sub-matrix from [K]e
corresponding to nodes i and j.

0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0
P
p 0 K 0 0 0 0 0 0 P
p 0 K 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 0 0 0 0

Initialisation Assembly of element (1)

CI3-321: Finite Element Analysis 7/28

 0 0 0 0 0 0
0 0
0 0 0
P
p 0 K 0 0 0 0 0 0 P
p K 0 0 0 0
0 0
0 0 0 0

Assembly of element (2) Assembly of element (3)

0 0 0 0 0 0
0 0
0 0
P
p K 0 0 0 P
p K 0 0 0
0 0
0 0

Assembly of element (4) End of assembly

1.3. Formulation of Finite Elements

Consider a system, defined by a geometric domain (), which is subjected to applied
loading {p}, and which sustains internal stresses {}. The equilibrium conditions for such a
system can be expressed using the principle of virtual work as follows:
Equilibrium {u}T {p}d {}T {}d {u}, {} (1.7)

where {u}, {} represent a compatible and admissible set of infinitesimal virtual

displacements and strains. The above equation can be restated as:
A system is in equilibrium if the external work performed by the applied loading over
any possible infinitesimal displacement mode is equal to the internal work performed
by the material stresses over the corresponding compatible infinitesimal strains.
The above principle can be applied on the element level to formulate displacement-based
finite elements. The first step in any finite element formulation is to relate the displacement
field(s) {u} within the element domain ( e ) to the nodal element displacement {U}e , using a
matrix of shape functions [N]:
{u} [N]{U}e (1.8)
For geometrically linear analysis, the strain field(s) {} are related to {u} through linear
differential operators, leading to the following relationship to {U}e :
{} [L]{u} [L][N]{U}e [B]{U}e (1.9)

CI3-321: Finite Element Analysis 8/28

The stresses {} are in turn related to the strains {} according to the material constitutive
law. Considering an element with nodal resistance forces {R}e , the application of the
principle of virtual work using the element displacement modes defined in (1.8) leads to:

{U}eT e [B]T
{}d e
{U}eT


{R}e
e [N]T
{p}d e
{U}

e
(1.10)

{R}e [B] {}de [N] {p}de
T T
(1.11)
e
e

The above expression is applicable to both linear and nonlinear material responses. When the
material constitutive law is linear, the stress-strain relationship may be expressed as:
{} [D] {} {o } (1.12)

where {o } represents the initial strains at zero stresses.

Substituting (1.9) into (1.12) and then into (1.11) leads to the following relationship between
the nodal resistance forces and the corresponding nodal displacements:

{R}e [B]T [D][B]de {U}e [N]T {p}de [B]T [D]{o }de (1.13)
e
e e

Comparing (1.13) to (1.4), the linear element response is determined by the following
stiffness matrix and equivalent nodal forces due to distributed loading and initial strains:
[K]e [B] [D][B]de
T
(1.14)
e

{Pp }e [N] {p}de [B] [D]{o }de

T T
(1.15)
e
e

Example E.1.3
Consider an axial bar element with a cross-sectional area that R1e U1e
R e , U e
e e
varies according to a quadratic function:
R2 U1
xe
A(x e ) A1e (4A em 3A1e A e2 )
Le
2 (E.18)
xe
2(2A em A1e A e2 ) e

L
Considering two axial degrees of freedom for the element,
linear shape functions can be used to approximate u(x e ) :
xe
u(x e ) U1e (U e2 U1e ) (E.19)
Le
which, with reference to (1.8), leads to a 12 [N] matrix:

CI3-321: Finite Element Analysis 9/28

 xe xe
[N] 1 e (E.20)
L Le
The axial strain is the first derivative of u(x e ) with respect to x e :
u (Ue2 U1e )
(x e ) (E.21)
x e Le
which, with reference to (1.9), leads to a 12 [B] matrix:
1 1
[B] e (E.22)
L Le
Note that [B] may also be obtained directly as [L][N] where:

[L] e (E.23)
x
With the 11 [D] representing the linear generalised constitutive response of the cross-
section:
[D] EA(x e ) (E.24)

the stiffness matrix is obtained according to (1.14) as:

Le
EAeeq 1 1
[K] [B] [D][B]dx e
e T

e
(E.25.a)
0 L 1 1
where,
4Aem A1e Ae2
Aeeq Aav
e
(E.25.b)
6
Considering a uniformly distributed load (p), the equivalent nodal forces are obtained from
(1.15) as:
Le
pLe 1
{Pp } [N]T {p}dx e
e
(E.25.c)
0
2 1

Clearly, the above [K]e and {Pp }e compare well with the previous corresponding entities
given in (E.10), which are determined from a crude application of structural mechanics
e
principles. The only difference is that [K] in (E.25) is based on the true average area over
the element length instead of an approximate average based on the nodal areas. Importantly,
the finite element formulation procedure has the benefit that it relies mainly on the
determination of the approximation shape functions [N], everything else arising automatically
from the application of the principle of virtual work.

CI3-321: Finite Element Analysis 10/28

1.4. Characteristics of Finite Element Predictions
For linear analysis, finite element predictions can be shown to be associated with a stiffer
response in comparison with the exact analytical solution, expressed in terms of the work
done by the applied loads:

W FE
({P} {Pp })T {U } Wexact (1.16)

Clearly, if there is a single point load applied to the system, then the displacement associated
with the load is typically smaller for the finite element solution than for the exact solution.
Given that the finite element approximation improves as the number of elements is increased,
convergence to the exact solution occurs from below when considering the work done by the
loads.
Another important characteristic of finite element analysis is that the approximation of
displacements is much better than that of displacement derivatives. Therefore, strains and
stresses are not as accurately predicted as displacements for a specific mesh of elements.
However, displacements as well as strains/stresses converge to the exact analytical solution
as the mesh is refined, though convergence to within a specific tolerance occurs earlier for
displacements than for strains/stresses.

Example E.1.4
Consider the problem of example E.1 which is now discretised using two finite elements of
equal length according to the formulation derived in example E.3. The average areas for the
two elements are given by ( A1av 1.0833A0 , Aav
2
1.5833A0 ), leading to new values for
[ K ] and {Pp } which replace (E.14):
2.1667 2.1667 0 P / 2
E A0
[K ] 2.1667 5.3333 3.1667 , {Pp } P (E.26)
L
0 3.1667 3.1667 P / 2
Applying the boundary conditions, as before, leads to the following solution:
PL PL
U1 1.4818 , U 2 0.7895 (E.27)
E A0 E A0
where the error in the nodal displacements is now less than 0.3%. In this case, the stiffer
response of the finite element solution is manifested in only one of the nodal displacements,
namely U 2 , being smaller than the exact solution, and is demonstrated in a smaller work
done by the loads:
P2 L
WFE ({P} {Pp }) {U } 3.0121
T
(E.28.a)
EA0
L
P2 L
Wexact P u(0) p u(x) dx 3.0301 (E.28.b)
0
EA0

CI3-321: Finite Element Analysis 11/28

A comparison between the finite element
solution (dotted line) and exact solution
(solid line) is shown in the adjacent figure for
the normalised displacement along the
member length ( d u EA0 /(PL) ), where it
is clear that the finite element solution with
only two elements is fairly accurate for the
displacement prediction. On the other hand,
the predictions for the normalised strain
( e EA0 / P ) and axial force ( f F/ P ) are
not as accurate, thus requiring a finer mesh
of elements especially in the left half of the
member.

CI3-321: Finite Element Analysis 12/28

2. Plane Stress/Strain Elements

2.1. Plane Stress vs. Plane Strain

Consider the prismatic solid shown below:

The plane stress condition is appropriate for relatively thin solids which are loaded in plane,
such that the out-of-plane direct and shear stresses are negligible. It is defined as:
u u(x, y), v v(x, y) (independent of z)
z xz yz 0 (2.1)

On the other hand, the plane strain condition is appropriate for solids that are loaded in plane
and restrained in the out-of-plane direction. Typically, this condition is applied to a slice of a
longitudinally uniform 3D body which is restrained at its two ends, for example a dam
structure. The plane strain condition is defined as:
u u(x, y), v v(x, y) (independent of z)

w 0 z xz yz 0

(2.2)

For both conditions, the displacement fields u(x, y) and v(x, y) represent the primary
unknowns of the exact mathematical model, while the planar strains (x , y , xy ) and
corresponding stresses (x , y , xy ) are secondary unknowns. Therefore, both conditions fall

CI3-321: Finite Element Analysis 13/28

within the class of two-dimensional problems, with the problem discretisation involving the
approximation of u(x, y) and v(x, y) over the two-dimensional x-y domain. Importantly, the
out-of-plane stresses and strains are not involved in the problem formulation for plane stress
and plane strain analysis, since the corresponding internal work is zero due to the fact that the
out-of-plane stresses are zero for plane stress problems, while the out-of-plane strains are
zero for plane strain problems.
The linear constitutive law for the planar stress components can be expressed as:



x


x


xo


{} [D]({} {o }), {} y , {} y , {o } yo (2.3)

xy
xy
xyo

The only distinction between plane stress and plane strain problems is reflected in the [D]
matrix, which can be derived in each case from the general triaxial constitutive matrix
considering the conditions of (2.1) and (2.2). For isotropic material:

1 0

(plane stress)
E
[D] 1 0
1 2

0 1
0
2
(2.4)

1 0


[D]
E 1 0 (plane strain)
(1 2)(1 )
1 2
0 0
2
For typical engineering materials for which ( 0) , it can be shown that the plane strain
condition is typically associated with a stiffer material response in comparison with the plane
stress condition.

2.2. Continuity Requirements of Shape Functions

A key requirement for valid discretisation is that the strain fields must be defined over the
whole system domain. For plane stress and plane strain problems the material strains are first
derivatives of the displacement fields, and hence the approximated displacement fields must
be continuous over the whole domain. This is referred to as a C0 continuity requirement1,
and it is guaranteed if the chosen shape functions satisfy the following conditions:
0
1. C continuity within the element domain, and

1
Cn refers to continuity in the nth derivative.

CI3-321: Finite Element Analysis 14/28

 0
2. C continuity along the edges between adjacent elements. It follows that the
displacements of each edge should be uniquely defined by the displacements of the
nodes on that edge alone. Hence, displacements of interior nodes must not cause any
displacements on the element edges.
C0 continuity violated C0 continuity violated

4 elements

External nodes should not cause Internal nodes should not cause
displacement of opposite edges displacement of element edges

2.3. Constant Strain Triangular Element

2.3.1 Local and global reference systems

The triangular element is derived in a local reference system, with transformations applied to
establish the element response characteristics in the global system.

Global reference system Local reference system of element (II)

2.3.2 Shape functions

In the local system, the tr iangular element has 6 displacement parameters and corresponding
resistance forces:

CI3-321: Finite Element Analysis 15/28

 u1 p1
v q
1 1
u 2 p
{d} , {f} 2 (2.5)
v2 q 2
u3 p3

v3 q3

The displacement fields within the element and the distributed applied forces are denoted by:
u(x, y) p(x, y)
{u} , {p} (2.6)
v(x, y) q(x, y)
In approximating u(x, y) and v(x, y) over the element domain in terms of the nodal
parameters, the following conditions must be satisfied:
u(0,0) u1 , u(x 2 ,0) u 2 , u(x3 , y3 ) u3 (2.7.a)
v(0,0) v1 , v(x 2 ,0) v2 , v(x 3 , y3 ) v3 (2.7.b)
Since there are three conditions on each of u(x, y) and v(x, y) , complete linear functions can
be used for their approximation: Complete
1 linear
u(x, y) a 0 a1x a 2 y (2.8.a)
x y
v(x, y) b0 b1x b2 y (2.8.b) x xy y2
2

These terms represent the first two rows of Pascal's polynomial x3 x2y xy2 y3
triangle.
Solving for (a 0 ,a1 ,a 2 ) in terms of (u1 , u 2 , u 3 ) , using (2.7.a) and (2.8.a), and for (b0 , b1 , b2 )
in terms of (v1 , v2 , v3 ) , using (2.7.b) and (2.8.b), leads to the approximation of u(x, y) and
v(x, y) in terms of the element nodal parameters:
u(x, y) 1 (x, y) u1 2 (x, y) u 2 3 (x, y) u 3 (2.9.a)
v(x, y) 1 (x, y) v1 2 (x, y) v2 3 (x, y) v3 (2.9.b)
or,
u1
v
1
u(x, y) 1 0 2 0 3 0 u 2
{u} [N]{d} (2.10)
v(x, y) 0 1 0 2 0 3 v 2
u3

v3

CI3-321: Finite Element Analysis 16/28

where,
x 2 y3 y3 x (x 2 x 3 )y
1 (x, y)
x 2 y3
y3 x x 3 y
2 (x, y) (2.11)
x 2 y3

y
3 (x, y)
y3
When plotted on an axis perpendicular to the x-y plane, the
shape functions represent planes formed by the
corresponding node and the opposite edge, as illustrated in
the adjacent figure for 1 (x, y) .
Note that the displacement of a specific node does not cause
displacements at the opposite edge, since the associated
shape function has zero values along the opposite edge.

2.3.3 Strains

The material strains are first derivatives of the displacement fields:

u
0
x x
v
x
u(x, y)
{} y 0 [L]{u} [L][N]{d} [B]{d} (2.12)
y y v(x, y)
xy u v

y x y x
where,
1 1
0 0 0 0
x2 x2
x3 x 2 x3 1
[B] 0 0 0 (2.13)
x 2 y3 x 2 y3 y3
x x 1 x3 1 1
3 2
0
x 2 y3 x2 x 2 y3 x2 y3

Note that [B] is independent of x and y. Therefore, for a specific displacement configuration
of the triangular element the strains are constant within the element, hence the name constant
strain triangle.

CI3-321: Finite Element Analysis 17/28

2.3.4 Local element response

The local element stiffness and equivalent nodal forces can now be determined from the
principle of virtual work as:
{f} [k]{d} {f p } (2.14)

[k] [B]T [D][B]t dA (2.15)

{f p } [N]T {p}t dA [B]T [D]{0 }t dA (2.16)

where t is the element thickness, and [D] is the constitutive matrix defined in (2.4), both of
which can vary over the element domain in terms of (x,y).
For plane strain problems, and for plane stress problems in which the thickness is constant
over the full system domain, the response of all elements can be determined and assembled
for a unit thickness, in which case thickness specification is not necessary.
When t and [D] are constant over the element domain, the integrand matrix in (2.15)
becomes constant since [B] is constant, and the stiffness matrix simply becomes:
x 2 y3
[k] [B]T [D][B]t A, A (2.17)
2
The equivalent nodal forces {f p } for uniform body and edge forces, assuming a constant
element thickness, are illustrated in the following table.

Uniform body force Uniform edge force

Actual load

Equivalent
nodal forces

CI3-321: Finite Element Analysis 18/28

2.3.5 Global element response

The global element response can be obtained from transformation of the local response
considering the element orientation in the global system:
{d} [T]{U}, {R} [T]T {f} (2.18)
where for the triangular element:
c s 0 0 0 0
s c 0 0 0 0

0 0 c s 0 0
[T] , c cos(), s sin() (2.19)
0 0 s c 0 0
0 0 0 0 c s

0 0 0 0 s c
in which is the angle between local x-axis and global X-axis.
Considering (2.14) and (2.18), the global element response is obtained as:
{R} [K]{U} {Pp } (2.20)

where,
[K] [T]T [k][T], {Pp } [T]T {f p } (2.21)

2.3.6 Accuracy of triangular element

The basic triangular element assumes constant strains within the element domain. Hence, it is
not very suitable for modelling plane stress and plane strain problems with high variations in
strain due to non-uniform loading or geometric irregularities. However, such problems can
still be modelled with the constant strain triangular element if a finer mesh is used in regions
of high strain variations. A good measure of analysis accuracy for problems in which the
strains are continuous over the system domain is the smoothness of the predicted strain
solution between adjacent elements.

Higher strain variation finer mesh required

CI3-321: Finite Element Analysis 19/28

Notwithstanding the above shortcomings, an
important use of the triangular element is to form a
transitional mesh between fine and coarse meshes of
quadrilateral elements, as illustrated in the adjacent
figure.

2.4. Basic Rectangular Element

The rectangular element is considered hereafter only in the local system, since its
characteristics in the global system can be obtained through a transformation similar to that of
the triangular element.

2.4.1 Shape functions

For simplicity, the shape functions will be derived in a local system based on natural
coordinates () instead of (x,y). The origin is taken at the centre of the element, and the
natural coordinates are assumed to extend between 1 and +1.
Hence, the real coordinates of any material point (x,y) are related to its natural coordinates
() as follows:
x a , y b (2.22)
The 4-noded rectangular element employs 8 local degrees of freedom and corresponding
resistance forces:

u1 p1
v q
1 1
u 2 p2

v2 q
{d} , {f} 2 (2.23)
u3 p3
v3 q3

u 4 p4
v q
4 4
Since there are four conditions on u(x, y) and v(x, y) , bilinear functions can be used for
their approximation:

u(x, y) a 0 a1x a 2 y a 3 x y (2.24.a) 1

x y
v(x, y) b0 b1x b2 y b3 x y (2.24.b) x xy y2
2

representing the shown terms in Pascal's triangle. x3 x2y xy2 y3

The shape functions can be derived easily in terms of the natural coordinates () instead of
(x,y):

CI3-321: Finite Element Analysis 20/28

 u1
v
1
u 2

u(, ) 1 0 2 0 3 0 4 0 v 2
[N]{d} (2.25)
v(, ) 0 1 0 2 0 3 0 4 u 3
v3

u 4
v
4
where:
1
1 (, ) (1 )(1 )
4
1
2 (, ) (1 )(1 )
4 (2.26)
1
3 (, ) (1 )(1 )
4
1
4 (, ) (1 )(1 )
4
The above shape functions can be written generally as:
1
i (, ) (1 i )(1 i ) (2.27)
4
where (i , i ) are the natural coordinates of node (i). Linear

Curvilinear
These shape functions satisfy the C0
continuity requirements, since the
displacements of each edge are uniquely
defined by the nodal displacements of that
edge alone.

2.4.2 Strains

Since the shape functions are obtained in terms of the natural coordinates, the material strains
must be derived as a function of displacement derivatives with respect to ():
u 1
0
x a


x
v 1 u(, )
{} y 0 [L]{u} [L][N]{d} [B]{d} (2.28)
y b v(, )
xy
u v 1 1

y x b a

CI3-321: Finite Element Analysis 21/28

where:
1 1 1 1
a 0 0 0 0
a a a

1 1 1 1
[B] 0
1
0 0 0 (2.29)
4 b b b b

1 1 1 1 1 1 1 1
b a b a b a b a
Note that [B] is a linear function of (), hence, unlike the previous triangular element, the
strains are not constant within the element. The strain x is constant with (or x), but can
vary linearly with (or y). Similarly, y is constant with (or y), but can vary linearly (or
x). The shear strain xy can vary linearly in both directions.

2.4.3 Local element response

The same general expressions used for the triangular element, (2.14) to (2.16), are also
applicable to the rectangular. Given that the infinitesimal area can be expressed as:
dA dx dy a bd d (2.30)
the integration of the stiffness matrix and equivalent nodal forces can be performed using the
natural coordinates ():
1 1
[k] [B] [D][B](a b t) d d
T
(2.31)
1 1

1 1 1 1
{f p } [N] {p}(a b t) d d [B] [D]{o }(a b t) d d
T T
(2.32)
1 1 1 1

2.4.4 Benefit of rectangular element

The use of one rectangular instead of two triangular elements for modelling a rectangular
sub-domain is particularly beneficial for representing bending modes, since two triangular
elements lead in such cases to discontinuous strain distributions.

CI3-321: Finite Element Analysis 22/28

3. Higher Order Elements and
Error Estimation

3.1. Higher Order Shape Functions

It was noted previously, for plane stress/strain analysis, that the basic 4-noded
rectangular element typically offers an improved performance over the 3-noded triangular
element, since it can represent continuous variations in the strains for bending types of
mode, involving the linear variation of x with y or y with x. However, the rectangular
element can only represent constant x with x and y with y, and hence for problems
dominated by such variations its performance is not much better than that of the triangular
element. Although the shape functions of the rectangular element contain a quadratic (x y)
term, it requires complete quadratic shape functions, including all (x 2 , x y, y2 ) terms, to
represent linear variations of all strains in both x and y directions. Higher order shape
functions of various orders (quadratic, cubic, etc..) may be derived, where the critical factor
influencing element accuracy and convergence rate is the degree of complete polynomial that
such functions achieve. Consideration is given hereafter to quadratic shape functions derived
for rectangular elements, though other higher order shape functions may also be established
for rectangular and triangular elements.
Lagrangian shape functions are derived for rectangular elements that employ a perfect grid of
nodes, including nodes within the element domain. Quadratic shape functions may be readily
derived for the 9-noded element shown below:

1
x y
x xy y2
2

x3 x2y xy2 y3
x x y x2y2 xy3 y3
4 3

where the shape function associated with a node k located at (x n , ym ) is easily obtained as:
(x x i ) (y y j )
k (x, y) (3.1)
i 1,3: i n (x n x i ) j1,3: jm (y m y j )

CI3-321: Finite Element Analysis 23/28

For example, the shape function for node 2 of the 9-noded element is:
(x x1 )(x x 3 ) (y y 2 )(y y3 )
2 (x, y) (3.2)
(x 2 x1 )(x 2 x 3 ) (y1 y2 )(y1 y3 )
which clearly satisfies the nodal boundary conditions (values of 1 at the corresponding node
and 0 at the remaining nodes) as well as continuity requirements.
Serendipity shape functions achieve a complete quadratic polynomial for rectangular
elements without the need for an interior element node.

1
x y
x xy y2
2

x3 x2y xy2 y3
The shape functions associated with the mid-side nodes are easily established as quadratic
along the corresponding side, and varying in the orthogonal direction linearly to zero at the
opposite side. For example, the shape functions associated with mid-side nodes 6 and 7 are
expressed in the natural coordinates system as:
1 1
6 (, ) ( )(1 )(1 ), 7 (, ) ( )(1 )(1 ) (3.3)
2 2
The shape functions associated with the corner nodes are obtained from those of the lower
order bilinear element, adjusting the mid-side values to zero through combination with the
mid-side shape functions. For example, the shape function associated with corner node 3 is
given by:
1 1 1
3 (, ) 3L (, ) 6 (, ) 7 (, ), 3L (, ) (1 )(1 )
2 2 4
1
3 (, ) (1 )(1 )( 1) (3.4)
4

CI3-321: Finite Element Analysis 24/28

3.2. Error Estimation for Finite Element Analysis
3.2.1 Sources of Error

The first (implicit) step involved in structural analysis is building a mathematical model of
the physical problem. Errors can be introduced at this stage if the assumptions of the
mathematical model do not reflect accurately the physical problem, such as the incorrect
choice of material properties, or the assumption of a plane strain model for a 3D problem
which is only approximately uniform in the longitudinal direction. These errors are termed
modelling errors.
Finite element analysis involves the discretisation of the mathematical model into a finite
number of degrees of freedom, and the subsequent numerical solution of the discrete system
of equilibrium equations. Round-off errors are generated due to the inability of computers to
store real numbers with an infinite degree of precision. Consequently, the addition of a very
small number to a very large number is often not performed and represented accurately by the
computer. Round-off errors can lead to considerable inaccuracies in the numerical solution of
the discrete system equations, particularly in the presence of an ill-conditioned overall
stiffness matrix for the finite element model. Discretisation errors reflect the inability of a
finite element model with a finite number of degrees of freedom to predict the exact solution
of the corresponding mathematical model, irrespective of the presence of round-off errors.

3.2.2 Discretisation errors

The finite element method is based on approximating the structural response in terms of a set
of discrete parameters through the use of shape functions. In linear analysis, the use of
conventional finite elements guarantees the following relationship between the discrete
system solution and the exact solution:

W FE
({P} {Pp })T {U } Wexact (3.5)

provided that (i) the geometric domain and material characteristics are represented accurately,
(ii) continuity requirements of the shape functions are satisfied, (iii) the integration of the
stiffness matrix and equivalent nodal forces is accurate, and (iv) all essential boundary
conditions are associated with zero values. The relationship in (3.5) expresses the stiffer-
response characteristic of finite element approximation. Since better accuracy can be
achieved with more elements, a finer mesh of finite elements is generally associated with a
more flexible response, provided the previous four conditions are satisfied.
It can be shown that the discretisation errors associated with a finite element mesh are of the
following orders:
Displacement error O(h p1 ), Stress/strain error O(h p1m ) (3.6)
where h is the characteristic element size, p is the order of complete polynomial used for the
displacement shape functions in terms of (x,y), and m is the order of differentiation relating

CI3-321: Finite Element Analysis 25/28

the strain to displacement fields. Thus, for the linear triangular and rectangular plane
stress/strain elements, the following orders of discretisation error are obtained:
Displacement error O(h 2 ), Stress/strain error O(h1 ) (3.7)
Therefore for these elements, halving the element size leads approximately to a 75%
reduction in error for displacements and to a 50% reduction in error for strains and stresses.
Such information can be used with the results from two related finite element meshes to
estimate the exact solution for displacements and strains/stresses, but not at points of
singularity in the solution (e.g. at positions of concentrated loading).
Improvement in accuracy can be achieved using p- or h-refinement. In p-refinement, higher
order elements of the same shape are used to replace lower order elements, such as using
8-noded quadrilateral elements to replace 4-noded quadrilateral elements 2 . Replacing two
linear triangular elements by one basic rectangular element is not considered to be a
p-refinement, since the shape functions of the rectangular element do not subsume those of
two triangular elements. This means that whilst the response of one basic rectangular element
is usually better than that of two triangular elements, the improvement in accuracy cannot be
guaranteed for all types of loading and for any distribution of material properties.
To ensure improvement in accuracy with h-refinement, the new mesh must subsume the
previous mesh in terms of the distribution of displacements. The safest procedure is to
replace each element by two or more elements of the same type, which usually ensures that
the new mesh is more accurate than the previous mesh. Considering the plane stress analysis
example shown below, with five meshes employing linear triangular and rectangular
elements, we can guarantee that mesh (b) is less stiff (and more accurate) than mesh (a), and
that mesh (e) is less stiff than mesh (d), provided that the four conditions for the stiffer
response of finite element approximation are satisfied. No other direct relationship between
the response of the five meshes can be guaranteed, although we would normally expect mesh
(e) to be the most accurate.

2
Note that mixing basic and higher order elements within the same mesh violates continuity requirements.

CI3-321: Finite Element Analysis 26/28

Appendix

A. Mixing of Two-Dimensional Elements of Different Orders

Consider the shown mesh which includes 4-noded bilinear and 8-noded quadratic elements.
Clearly, a mode which consists of a displacement at any of the edges linking the bilinear and
quadratic elements leads to violation of C0 continuity requirements, due to the different
orders of shape function used along the edges. Therefore, the mixing of elements of different
orders should be avoided. Of course, such an issue does not arise with meshes using 3-noded
triangular and 4-noded quadrilateral elements, since the shapes functions along the edges are
linear in both cases.
C0 violated

Linear Quadratic

B. Accuracy of Finite Element Prediction Based on Stress/Strain

Smoothness
For problems in which the exact analytical solution for stresses and strains is smooth, the FE
solution smoothness can be used to assess the solution accuracy. One common technique is
based on averaging the element stress/strain values at a specific nodal location. These
averaged nodal values are typically more accurate than the individual element values3. The
maximum difference between the element values at a specific node and the averaged nodal
value normally reflects the approximation involved in the finite element solution.

Average value
Element values

3
ANSYS provides element and nodal (averaged) values for stresses and strains. The contour plots of element
stresses typically show discontinuity between elements when the solution is not accurate.

CI3-321: Finite Element Analysis 27/28

C. Determination of Individual Terms of [k]
There is occasionally a need to determine individual terms of the [k] matrix, particularly in
manual calculations when only a few elements of [k] are required. In general, a specific (i,j)
element of [k] may be obtained using the ith and jth columns of the [B] matrix as follows:
[k]i, j [B]i T [D][B] j de

For example, for the constant strain triangular element with uniform thickness and properties:
[k]i, j [B]iT [D][B] j t A

CI3-321: Finite Element Analysis 28/28