MWR Galerkin Fem

FromWeightedResidualMethods
toFiniteElementMethods
LarsErikLindgren
2009
TABLE OF CONTENT
1 INTRODUCTION
2 SOMEDEFINITIONS
3 SHORTFINITEELEMENTCOURSE
4 WEIGHTEDRESIDUALMETHODS
4.1 SUBDOMAINMETHOD
4.2 COLLOCATIONMETHOD
4.3 LEASTSQUARESMETHOD
4.4 METHODOFMOMENTS
4.5 GALERKINANDRITZMETHODS
4.5.1 RELATIONBETWEENTHEGALERKINANDRITZMETHODS
4.6 PETROVGALERKINMETHOD
4.7 COMPARISONOFWRMMETHODS
4.7.1 PROBLEMDEFINITIONANDEXACTSOLUTION
4.7.2 SUBDOMAINEXAMPLE
4.7.3 COLLOCATIONEXAMPLE
4.7.4 LEASTSQUARESMETHOD
4.7.5 METHODOFMOMENTSEXAMPLE
4.7.6 GALERKINEXAMPLE
4.7.7 SUMMARYOFCOMPARISONS
8
8
8
8
8
9
10
10
10
11
12
13
14
15
16
5 CLASSICALANDCOMPUTATIONALGALERKINMETHODS
16
6 FINITEELEMENTMETHODS
17
6.1 GLOBALWEIGHTANDTRIALFUNCTIONS
6.2 NODALBASEDTRIALANDWEIGHTFUNCTIONS
6.3 ELEMENTBASEDTRIALANDWEIGHTFUNCTIONS
18
20
23
7 NUMERICALINTEGRATION
30
8 BEAMELEMENTS
32
8.1 BERNOULLIBEAM
8.2 TIMOSHENKOBEAM
8.3 CANTILEVERBEAMPROBLEM
32
35
38
9 ISOPARAMETRICMAPPINGINTWODIMENSIONS
42
10 AFOURNODEPLANESTRESSELEMENT
45
10.1
10.2
10.3
10.4
47
48
49
49
DISPLACEMENTSANDGEOMETRYOFFOURNODEPLANEELEMENT
STRAINSOFFOURNODEPLANEELEMENT
CONSTITUTIVEMODELFORPLANESTRESS
FLOWCHARTFORCALCULATIONOFFOURNODE,PLANESTRESSELEMENT
11 CONVECTIVEHEATTRANSFER
50
12 APPENDIX:BEAMTHEORIES
12.1 BERNOULLIBEAM
12.2 TIMOSHENKOBEAM
1
4
Lars-Erik Lindgren
Page i
2009-03-31
12.3 ONEDIMENSIONALELEMENTFORHEATCONDUCTION
13 REFERENCES
Page ii
1 Introduction
The finite element method is a general method for solving partial differential equations of
different types. It has become a standard method in industry for analysing thermo-mechanical
problems of varying types. It has to a large extent replaced experiments and testing for quick
evaluation of different design options.
This text is supplementary material in an undergraduate course about modelling in
multiphysics. The aim is to show how a finite element formulation of a given mathematical
problem can be done. Naturally, the focus is on simple, linear equations but some discussions
of more complex equations with convective terms are also included.
2 Some definitions
The most important definition is model a symbolic device built to simulate and predict
aspects of behaviour of a system. The word aspects indicates that there is a limited, specific
purpose for which the model is created. It is the scope of the model. Determining the scope is
the most important step in the modelling process. What information is wanted? Why should
the analysis be done? The scope determines what tool and model can be used. The scope
determines together with when the analysis is done what accuracy is needed. When is
when is it applied in the design process? Less is known at early design phases and therefore
less accurate models are needed. Other useful definitions are:
Verification is the process of assuring that the equations are solved correctly. Numerical
results are compared with known solutions. Verification is not discussed in this text. There
exist several benchmark cases for checking finite element codes. A user should be aware that
some unusual combinations may trigger problems that have not checked for by the code
developer and no code is ever free of programming errors. Validation is when it is assured
that the correct equations are solved. The analysis results are compared with reality.
Qualification is when it is assured that the conceptual model is relevant for the physical
problem. The idealisation should be as large as possible but not larger.
Sufficient valid and accurate solution is what the modelling process should result in.
Sufficient denotes that it must be related to the context the model is used in. For example,
how accurate is loading known. It is no use to refine the model more than what is known
about the real life problem. Then more must be found out about loading, material properties
etc before improving the model.
Prediction is the final phase of where a simulation or analysis of a specific case that is
different from validated case is done.
Uncertainty is of two types in the current context. Those can be removed by further
investigations and those that cannot. This is related to variability which here denotes the
variations that can not be removed. This may be, for example, variation of material properties
for different batches of nominally the same material or fluctuations of loading.
Simulation an imitation of the internal processes and not merely the results of the system
being simulated. This word is less precisely used in this text. Here the word is usually used
when computing the evolution of a problem during a time interval. The word analysis is
sometimes use to compute the results at one instant of time.
Page 3
3 Short finite element course

The Finite Element Method is a numerical method for the approximate solution of most
problems that can be formulated as a system of partial differential equations. There exist
variants of the steps below that are needed in some cases. For the basic theory of the finite
element see [1] and see [2] for its application for nonlinear mechanical problems. The finite
element method belongs to the family of weighted residual methods.
A short version of the basic steps can be described as below.
1. Make a guess (trial function) that has a number of unknown parameters. This is written as,
for a mechanical problem with one unknown displacement field as u(x),
(3.1)
where the functions
often are polynomials. This is a displacement based formulation
which is the standard approach in finite element formulation of mechanical problems. It will
be temperature in the case of thermal problems. The trial functions are set up in such a way
that the N unknown coefficients (parameters) are the field value at some point (node).
2. The trial function is inserted into the partial differential equation and the boundary
conditions of the problem. The variant leading to a standard displacement based finite element
formulation assumes that the essential boundary conditions are fulfilled. The remaining
equations will not be fulfilled. Small errors residuals are obtained. This will be in our fictive
case
(3.2)
where L denotes the differential equation of the problem and is defined over a the domain of
the problem.
denotes the essential boundary conditions that are fulfilled and
are the
natural boundary conditions that will be approximated.
3. Make a weighted average of the errors to be zero. Therefore the name weighted residual
method (WRM). Use the same functions as the trial functions as weighting functions. This
variant of WRM is called a Galerkin method. This step generates the same number of
equations as number of unknowns.
(3.3)
4. Some manipulations leads to a system of coupled equations for the unknown parameters
now in the array . This is the approximate solution.
(3.4)
where is called a stiffness matrix in mechanical problems and
is a load vector due to
different kind of loadings including possible natural boundary conditions.
5. The method has theorems that promise convergence. Thus an improved guess with more
parameters will give a more accurate solution.
6. Derived quantities like strain and stress that are derivatives of the approximate solution has
a larger error than the primary variables in .
Many textbooks formulate the finite element method for mechanical, elastic problems using
the theorem of minimum potential energy. This theorem also requires that the essential
boundary conditions of the problem are fulfilled like we introduced in step 3 above.
Page 4
1. Make a guess (trial function) where a number of unknown parameters, this is the same as
for the WRM approach,
(3.5)
u ( x ) = ui i ( x )
i=1
2. The trial function is set into the expression for the total potential energy, which is
integrated over the domain like the first term in Eq. (3.3). This can be written in matrix form
as
(3.6)
3. The potential energy is stationary w.r.t to the parameters leading to
(3.7)
4. The method has theorems that promise convergence. Thus an improved guess with more
parameters will give a more accurate solution.
6. Derived quantities like strain and stress that are obtained from derivatives of the
approximate solution have a larger error than the primary variables in .
The difference between the approaches is in step 3. The energy method needs fewer
manipulations at this step. However, it cannot be applied for nonlinear problems like those
involving plasticity. Then the WRM approach can still be applied. The principle of virtual
power or work is used in many textbooks to derive the finite element method for nonlinear
mechanical problems. A comparison would show that it is the same as WRM.
Eq. (3.1) and (3.5) used nodal values as unknowns multiplying trial functions. The latter are
usually defined over local regions, elements. Sometimes the trial functions are called
interpolations functions as the field is interpolated between the nodal values using these
functions. The interpolation within one single element is written as
(3.8)
where N is a matrix with interpolation functions, also often called shape functions as they
determine the shape of the possible displacement field on element can describe. u is a vector
with the nodal displacements of the element. nnode is the number of nodes in one element.
The analysed geometry is split into elements. The elements are connected at the nodes as
shown in Figure 3.1. The approximated field is interpolated over the elements from the nodal
values. The elements must be combined so that there is no mismatch between the
displacement fields along common boundaries of elements. The most crucial step in the finite
element modelling process is the choice of elements and the discretisation of the domain.
Figure 3.1. Discretised domain consisting of three and four node elements.
Different physics have different mathematical formulations but share some basic features.
They are illustrated in Figure 3.2 for a static, mechanical problem. The equations are
Page 5
discussed later but we summarise them already here. The left side of the diagram are the
kinematic variables describing the motion, u, and the gradient of it, i.e. the strain .
Therefore, the matrix B contains derivatives of the interpolations functions N, i.e. the shape
functions.
The constitutive equation, E, relates strain to stresses, for example Hookes law. In other
problems it is also common that some kind of gradient is related to some kind of flux. For
example, the gradient of the temperature gives the heat flux in thermal problems. The model
for this is Fouriers heat conduction law. The stresses are to be in equilibrium with applied
forces. The line from stress, , to the box symbolising the equilibrium equations is dashed.
This means that this equation is approximated and that is why it is an integral. The equation is
only fulfilled at the nodes with nodal equilibrium for the system written as
(3.9)
This is a more general form than in Eq. (3.7).
Figure 3.2. A Tonti diagram illustrating the basic finite element relations in mechanics.
4 Weighted Residual Methods

The Weighted Residual Method is illustrated on a simple one-dimensional problem. First the
problem is given a general mathematical form that is relevant for any differential equation. It
is assumed that a problem is governed by the differential equation
(4.1)
It is to be solved over a given domain. The solution is subject to the initial conditions
(4.2)
and boundary conditions
(4.3)
Page 6
L, I and B denote operators on u. This can be derivatives and any kind of operations so they
represent all kinds of mathematical problems. An approximate solution is inserted in to
these relations giving residuals, errors;
(4.4)
(4.5)
(4.6)
The approximate solution can be structured so that;
i)
ii)
. Then it is called a boundary method.

. Then it is called an interior method This requirement may be violated for some
boundary conditions reducing the efficiency1 of the method.
iii) Else it is a mixed method.

Boundary element methods2 are example of boundary methods. Green functions3 can be used
as trial functions. Then only the boundary of the domain need be discretised. This will result
in small, but full, matrices when the surface-volume ratio is small.
We focus on interior in the following examples of weighted residual methods. The
approximate solution is taken as
(4.7)
where
are analytic functions called the trial functions.
must satisfy initial and
boundary conditions as exact as possible. However, we will later see cases where the trial
functions are used for this purpose also and no separate
is used. The trial functions
should be linearly independent and the first N members of the chosen set should be used.
Notice that the parameters to be determined,
, are chosen to be a function of time and the
trial functions is only dependent on space. This is the most common approach although not
necessary. One exception is space-time finite elements.
The parameters
are determined by setting the weighted average of the residual over the
computational domain to zero
(4.8)
Additional terms may be included if the requirement on fulfilling all the boundary conditions
is relaxed. The functions
are called weight functions. N independent equations are
needed to determine the coefficients. Therefore, N independent weight functions are needed.
If
, then the residual will become zero in the mean, provided the initial and boundary
conditions were fulfilled exactly, and thereby the approximate solution will converge to the
exact solution in the mean
(4.9)
Efficiency in terms of needed number of terms to obtain a given accuracy. However, other advantages may be
gained motivating a relaxing this requirement as will be shown later.
http://www.boundary-element-method.com/intro.htm
http://en.wikipedia.org/wiki/Green%27s_function
Page 7
We will illustrate some variants of WRM methods in the following.

4.1 Subdomain method
The domain is split into subdomains,
, which may overlap. The weight function is then

(4.10)
One example of this is the finite-volume method4. Each element is surrounding its associated
node. Conservation equations then relates changes within this volume with fluxes over its
boundaries.
4.2 Collocation method
The weight function is given by
(4.11)
5
where is the Dirac delta function . Thus the residual is forced to be zero at specific
locations.
4.3 Least-squares method
The weight function is given by
(4.12)
where are the coefficients in the approximate solution, Eq. (4.7). This makes the Eq. (4.8)
corresponding to
(4.13)
This in turn is the stationary value of
(4.14)
thereby motivating the name of the method.
4.4 Method of moments
The weight function is in this case given by
(4.15)
4.5 Galerkin and Ritz methods
The weight function is chosen from the same family of functions as the trial functions in Eq.
(4.7).
(4.16)
The trial (or test) functions are taken from the first N members of a complete set of functions
in order to guarantee convergence when increasing N.
The Galerkin method is the same as the principle of virtual work or power6 used in mechanics
when formulating the finite element method. The weight functions correspond to virtual
4
http://en.wikipedia.org/wiki/Finite_volume_method
http://en.wikipedia.org/wiki/Dirac_delta_function
Page 8
displacements or velocities in this approach. For elastic problems, it also corresponds to the
principle of minimum total energy7,8. This method can be a starting point for formulating
approximate solutions. It is sometimes called Ritz method. It is commonly used in basic
courses about the finite element method in mechanics. However, it is not valid in cases like
plasticity. Then the principle of virtual work is used.
4.5.1 Relation between the Galerkin and Ritz methods
Thus the Galerkin method is more general than Ritz method, in the same way as the principle
of virtual work is more general than the principle of minimum total potential energy.
The relation between the Galerkin method and Ritz method can be described as follows.
Assume that u is the solution to the differential equation
(4.17)
where A is a positive definite operator. This property means that
(4.18)
Then the solution to Eq. (4.17) can be shown to be equivalent to finding the minimum of the
functional
(4.19)
An approximate solution like in Eq. (4.7) is used but now it only fulfils the essential boundary
conditions. Assuming that this fulfilment can be done by fixing appropriate coefficients
leads to
(4.20)
This can be written in matrix form as
(4.21)
where the coefficients of the matrix and vector are
(4.22)
and
(4.23)
The best choice of parameters is the set that minimise this functional. A condition for this is
(4.24)
This leads to
(4.25)
This is the same as for the Galerkin method as can be seen in section 4.7.6. The convergence
properties for the Ritz method states that increasing N makes the functional
go towards the
6
http://en.wikipedia.org/wiki/Virtual_work
http://en.wikiversity.org/wiki/Introduction_to_Elasticity/Principle_of_minimum_potential_energy
http://en.wikipedia.org/wiki/Minimum_total_potential_energy_principle
Page 9
true minimum. Thus an approximate solution will not reach down to the true minimum but we
will have convergence from above in terms of the norm. This norm can be interpreted as an
energy norm in mechanical problems.
4.6 Petrov-Galerkin method
The weight function is represented by
(4.26)
where
is functions similar to the test functions
but with additional terms to impose
some additional requirements on the solution. Typically, terms to improve the solution of
problems with convection like in convection-dominated fluid flow problems.
4.7
Comparison of WRM methods
4.7.1 Problem definition and exact solution

The different weighted residual methods will be applied on the problem
(4.27)
It is assumed that the function u is written in non-dimensional form, ie it does not have any
dimensions.
(4.28)
or
(4.29)
We will limit our discussion to the particular case of a=1 [m2] and b=1 [m] and
. The
constants have units that will not be visible in later discussions and thereby it may seem that
the units are not consistent between different terms but they are!
The boundary conditions have been named e=essential and n=natural for reasons shown later
in the finite element formulation, chapter 6. The exact solution is the same for both boundary
conditions but the first variant of Eq. (4.29) is easier to implement as it gives directly a
condition for the value of one coefficient. The other variant gives a relation between the
unknown coefficients that can be implemented in different ways. The exact solution to Eq.s
(4.27)-(4.29) is
(4.30)
We choose an approximate solution given by
(4.31)
This gives the residuals
(4.32)
(4.33)
or
(4.34)
Fulfilling the essential boundary condition does not require the extra term u0 in Eq. (4.7) but
is achieved by setting
Page 10
(4.35)
The natural boundary condition can also be used to impose conditions on the parameters. Eq.
(4.29) gives directly
(4.36)
4.7.2 Subdomain example
The subdomain method, section 4.1, splits the unit interval into N domains. We make them
equal sized,
, and thus Eq. (4.8) becomes
(4.37)
where
(4.38)
This can be integrated giving, for each k,
(4.39)
This leads to a system of equations with
9
(4.40)
where the right hand side is

and the matrix on the left hand side becomes
and if j >2
The table below shows the results for the subdomain method and its condition number and the
implementation is shown in Table 4.2. A high condition number and indicates that the
solution is sensitive to round off error. It may even be impossible to invert the matrix.
Table 4.1. L2-error and condition number of matrix for different subdomain solutions.
Number of
terms (N)
3
4
5
6
7
8
Error
1.05
0.18
0.04
2.5e-3
3.2e-4
1.3e-5
This has same meaning as
Condition
number
1.0000
27.8
2.9e2
2.0e3
1,5e4
1.0e5
Number of
terms (N)
9
10
11
12
13
14
Error
1.2e-6
3.3e-8
2.3e-9
1.4e-10
2.3e-9
6.9e-9
Condition
number
7.7e5
5.3e6
4.0e7
2,9e8
2.2e9
1.6e10
where it is implied a summation over the repeated index j.
Page 11
Table 4.2. Excerpt from Matlab code for the subdomain method.
for k=1:N
F(k)=dx;
xk=k*dx;
xk_1=xk-dx;
for j=1:N
K(k,j)=(xk^j-xk_1^j)/j;
if j>2
K(k,j)=K(k,j)+(j-1)*(xk^(j-2)-xk_1^(j-2));
end
end
end
% We impose the condition a1=1 that is multiplying first column of K
a1=1;
Fmod=K(:,1)*a1;
% Move this to right hand side
F=F-Fmod;
% The first equation for a1 is not needed any more
F(1)=[];K(1,:)=[];K(:,1)=[];
a2=1/cos(1);
Fmod=K(:,1)*a2;
F=F-Fmod;
% The current first equation for a2 is not needed any more
F(1)=[];K(1,:)=[];K(:,1)=[];
4.7.3 Collocation example

The collocation method, section 4.2, requires the residual to be zero at specific locations. We
make specify these points to be at centre of domain of equal length. Thus N locations cause
these points to be at
(4.41)
Then Eq. (4.8) becomes
(4.42)
This leads to a system of equations with
(4.43)
where the right hand side is
And the matrix on the left hand side becomes
and if j >2
The table below shows the results for the subdomain method and the implementation is
shown in Table 4.4
Table 4.3. L2-error and condition number of matrix for collocation method.
Number of
terms (N)
3
4
5
6
Error
1.0536
0.1793
0.0420
0.0026
Condition
number
1.0
27.7
283.
2.e3
Number of
terms (N)
7
8
9
10
Error
3.4e-4
1.4e-5
1.3e-6
3.8e-8
Page 12
Condition
number
1.5e4
1.0e5
7.5e6
5.3e7
Table 4.4. Excerpt from Matlab code for the point collocation method.
for k=1:N
F(k)=1;
xk=(k-0.5)*dx;
for j=1:N
K(k,j)=xk^(j-1);
if j>2
K(k,j)=K(k,j)+(j-2)*(j-1)*xk^(j-3);
end
end
end
a1=1;
Fmod=K(:,1)*a1;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
a2=1/cos(1);
Fmod=K(:,1)*a2;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
4.7.4 Least squares method

The least squares method, section 4.3, makes Eq. (4.8)
(4.44)
where the first term is only present when k>2. Further elaboration gives leads to a system of
equations with
(4.45)
where the right hand side, coming from -1-term in R, is
and if k>2
The matrix, that is symmetric, becomes
and if k > 2
and if j >2
and if k,j >2
The results are shown in Table 4.5 and the implementation in Table 4.6. It can be noted that
the error increases at N=10. This is due to the high condition number. This trend goes on until
N=14 where the solution procedure fails completely when using Matlab.
Table 4.5. L2-error and condition number of matrix for least squares solution.
Number of
terms (N)
Error
Condition
number
Number of
terms (N)
Error
3
4
5
6
0.56
0.009
6.4e-4
1.1e-5
1.0000
37.4
736.
1.7e4
7
8
9
10
6.3e-7
1.0e-8
5.46e-10
3.5e-9
Page 13
Reciprocal
condition
number
4.5e5
1.3e6
4.0e8
1.1e10
Table 4.6. Excerpt from Matlab code for the least squares method.
for k=1:N
k1k2=(k-1)*(k-2);
F(k)=1/k;
if k>2
F(k)=F(k)+k-1;
end
for j=1:N
j1j2=(j-1)*(j-2);
K(k,j)=1/(j+k-1);
if k>2 && j> 2
K(k,j)=K(k,j)+(j1j2+k1k2)/(j+k-3)+j1j2*k1k2/(k+j-5);
elseif k>2
K(k,j)=K(k,j)+k1k2/(j+k-3);
elseif j>2
K(k,j)=K(k,j)+k1k2/(j+k-3);
end
end
end
a1=1;
Fmod=K(:,1)*a1;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
a2=1/cos(1);
Fmod=K(:,1)*a2;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
4.7.5 Method of moments example

The method of moments, section 4.4, makes Eq. (4.8)
(4.46)
Further elaboration gives leads to a system of equations with
(4.47)
The matrix on the left hand side becomes

if j >2
The results are shown in Table 4.7 and the implementation in Table 4.8. The error increases
when N=10 and fails completely at N=13.
Table 4.7. L2-error and condition number of matrix for method of moments solution.
Number of
terms (N)
3
4
5
Error
0.99
0.142
0.026
Condition
number
1.0000
236.
1.7e4
Number of
terms (N)
7
8
9
Error
1.3e-4
3.8e-6
3.28e-7
Page 14
Condition
number
2.9e6
1.1e7
4.2e8
1.3e-3
8.0e5
10
2.63e-6
1.4e9
Table 4.8. Excerpt from Matlab code for method of moments.

for k=1:N
F(k)=1/(k+1);
for j=1:N
K(k,j)=1/(j+k);
if j>2
K(k,j)=K(k,j)+(j-1)*(j-2)/(j+k-2);
end
end
end
a1=1;
Fmod=K(:,1)*a1;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
a2=1/cos(1);
Fmod=K(:,1)*a2;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
4.7.6 Galerkin example

The Galerkin method, section 4.5, will be quite similar to the method of moments, section 4.4,
for this particular choice of trial functions. Eq. (4.8) becomes
(4.48)
The notation A was introduced for comparison with the Ritz method discussed in the context
of Galerkin method in section 4.5. The operator A is on the approximate solution
is
. This leads to
(4.49)
Further elaboration gives leads to a system of equations with

(4.50)

and if j >2
The results shown below have the same accuracy as the method of moments due to the special
choice of trial/weight functions. The results are shown in Table 4.7 and the implementation in
Table 4.8.
Page 15
Table 4.9. L2-error and condition number of matrix for Galerkin method.
Number of
terms (N)
3
4
5
6
Error
0.99
0.142
0.026
1.3e-3
Condition
number
1.0000
236.
1.7e5
8.0e6
Number of
terms (N)
7
8
9
10
Error
1.3e-4
3.8e-6
3.28e-7
2.63e-6
Condition
number
2.9e7
1.1e8
4.2e8
1.4e9
Table 4.10. Excerpt from Matlab code for Galerkins method.

for k=1:N
F(k)=1/(k+1);
for j=1:N
K(k,j)=1/(j+k);
if j>2
K(k,j)=K(k,j)+(j-1)*(j-2)/(j+k-2);
end
end
end
a1=1;
Fmod=K(:,1)*a1;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
a2=1/cos(1);
Fmod=K(:,1)*a2;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
4.7.7 Summary of comparisons

The above comparisons illustrate the difference between the methods but are too limited to
draw general conclusions. The table below is from Fletcher [3] with subjective comparisons
of the some of the methods.
Table 4.11. Subjective comparisons of different weighted residual methods, from Fletcher [3].
MWR
Galerkin
Least-squares
Subdomain
Collocation
Accuracy
Very high
Very high
High
Moderate
Ease of formulation
Moderate
Poor
Good
Very good
Additional remarks
Equivalent to Ritz
method where
applicable.
Not suited to
eigenvalue or
evolutionary
problems.
Equivalent to finitevolume method;

suited to
conservation
formulation.
Orthogonal
collocation gives
high accuracy.
5 Classical and computational Galerkin methods

The classical Galerkin methods were applied before computers were commonplace. Thus
there was a need to obtain high accuracy with few unknowns. The method used global test
Page 16
functions. The use of orthogonal10 test functions further reduced the calculations needed.
Naturally, the use of global functions also made it difficult to solve problems with irregular
boundaries. The advent of computers made it possible to solve problems with greatly
increased number of parameters. Today the results system of equations can have millions of
unknowns.
The global test functions become less and less unique with increasing number. For example,
going from a polynomial of x19 to x20 does not add much new information into Eq. (4.7). Thus
adding more terms in the approximate solution will make the contribution from higher order
terms smaller and smaller for larger N. This will then lead to that the system of equations that
is to be solved in order to determine the coefficients aj will be ill-conditioned, i.e. sensitive to
round-off and truncation errors.
Therefore the trial functions in computational Galerkin methods are chosen in order to reduce
this problem. The use of spectral methods reduces this problem due to the orthogonal property
of these functions. Finite Element Methods shown next are based on the use of local trial
functions instead. Increasing the number of coefficients is done by increase the number of
domains, elements, they are defined over. The way these domains, elements, are described is
also a key to one of the strong points of Finite Element Methods, their ability to solve
problems with complex boundaries.
6 Finite Element Methods

A Finite Element Method (FEM) solution of the same problem as earlier will be formulated
with the use of the Galerkin method but now allowing the trial solution to approximate the
natural boundary condition. Furthermore, partial integration will be used in order to create a
symmetric formulation with the same derivatives on the trial and weight functions. Global
functions and local functions, the latter in the spirit of the Finite Element Method, will be
used now. Two approaches for the FE-formulation will be shown. The first case is based on
the use of nodal based trial functions whereas the second variant is based on element based
definitions. The latter is the more effective way to implement the method. Iso-parametric
formulation will be introduced in this context.
The problem given in section 4.7 will be solved but the natural boundary condition will be
prescribed at the right end and will be approximated in the solution11. The equations are
repeated below.
(6.1)
(6.2)
(6.3)
The exact solution to Eq.s (6.1)-(6.3) is still
(6.4)
10
http://en.wikipedia.org/wiki/Orthogonal_functions
11
We cannot set it at x=0 in the current approach as we will later set the weight function to zero where we have
the essential boundary condition, which also is at x=0.
Page 17
The natural boundary condition is now included in the residual by extending Eq. (4.8) with
the error in RB from Eq. (4.6) giving
(6.5)
where
is the part of the boundary where the natural boundary conditions are prescribed (in
this case x=1). We have chosen to use same weight functions for
. The reason for the
choice of minus-sign will be obvious later as this will make some terms cancel each other.
Then we get
(6.6)
There is a second derivative on the trial functions but no on the weight functions. We perform
a partial integration12 leading to
(6.7)
The approximate solution must fulfil the essential boundary condition at x=0. Thus we can
freely choose the weight function to be zero along this part of the boundary
or in our case
. Then we get
(6.8)
Now we can see that the first term and the corresponding part of the last term cancel, as we
introduced the minus sign in Eq. (6.5) giving
(6.9)
or
(6.10)
If it was a mechanical problem, then the first term would be called a stiffness matrix. The
second term would be nodal forces due to distributed loads and the term of the right hand side
due to force on the boundary.
The formulation is now symmetric with respect to the weight and trial functions. This will
lead to symmetric matrices giving computational efficiency. We will go through the details in
the following using a global defined trial function and a locally defined as in FEM.
6.1 Global weight and trial functions
Eq. (6.10) is first used with the same global functions as used earlier before the finite element
version is shown in the next section. The approximate solution, and corresponding weight
functions, are
12
Page 18
(6.11)
Inserted into Eq. (6.10) gives
where the k,j >1 for the first term. Notice the relaxed criterion on number of derivatives that
must be possible to define, compared to Eq. (4.50) where j starts with 3 (j>2) in the
summation. This reduction is due to the partial integration performed.
(6.12)
A comparison with Eq. (4.50) shows that the problem is now symmetric with respect to
indexes j and k. We has now
(6.13)
with

and if j,k >1
Imposing the essential boundary condition is done like in earlier cases by prescribing a1=1.
The results shown are shown in Table 6.1 and the implementation in Table 6.2. It can be
noted that if the error in the natural boundary conditions would have been included in the
Galerkin method in Eq. (4.48) then the errors in Table 4.9 and Table 6.1 would have been the
same as the partial integration only makes the matrix symmetric but will bring in any new
assumptions. The error increases from N=9 and the method fails completely at N=13 due to
the ill-condition system of equations.
Table 6.1. L2-error and condition number of matrix for Ritz method.
Number of
terms (N)
3
4
5
6
Error
0.072
0.002
1.8e-4
4.1e-6
Condition
number
18.3
375.
1.0e4
2.8e5
Number of
terms (N)
7
8
9
10
Error
2.6e-7
4.8e-9
4.3e-8
9.1e-7
Page 19
Condition
number
8.5e6
2.7e8
8.4e9
2.5e11
Table 6.2. Excerpt from Matlab code for Ritz method case, i.e. a symmetric variant of the
Galerkins method.
for k=1:N
F(k)=1-1/k;
for j=1:N
K(k,j)=-1/(j+k-1);
if j>1 && k >1
K(k,j)=K(k,j)+(j-1)*(k-1)/(j+k-3);
end
end
end
a1=1;
Fmod=K(:,1)*a1;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
a2=1/cos(1);
Fmod=K(:,1)*a2;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
6.2 Nodal based trial and weight functions

Now we use locally defined trial, and weight, functions of the type that are used in the finite
element method. They are defined as
(6.14)
where
(6.15)
is the length of each segment and assumed constant for simplicity in the following.
A typical function is shown in Figure 6.1. Note that the very first node (j=1) has only the right
half of the function and the final node (j=N) has only the left part of the function.
Figure 6.1. Finite element type of trial functions or shape functions with local base.
The approximate solution is then written as
(6.16)
We change the notation now and use
that have the property
instead of
. This is reasonable as we use functions

(6.17)
Page 20
This means that it has unit value at the coordinate corresponding to j=k and zero at all other
nodal coordinates. Then the coefficients in Eq. (6.18) will correspond to the nodal value at the
specified node
This motivates the change in from a:s to u:s in the notations. Note also the introduction of the
concept node. The trial functions
are associated with a point at
that now will be called
a node. For simplicity we have assumed that the distance between each node is equal in the
example we are discussing. However, this is not necessary. FEM is very flexible with respect
to geometry. This will be particular clear later when discussing element based functions
where the concept of an element is introduced. The discussion of isoparametric element
formulation is the top of the line in the finite element method giving its ultimate flexibility.
Now back on track!
Insertion of Eq. (6.14)into Eq. (6.10) gives
(6.18)
The integral can be split into separate contributions as the different functions only overlap in
certain intervals. Thus for a given k, there is an overlap with right half of the function
for
j=k-1 and left half for j=k +1 according to the gray area in Figure 6.2.
Figure 6.2. Overlap between neighbours where the trial functions contribute to integrals to be
evaluated. Thus equation k will have contributions from three trial functions j=k-1, j=k and
j=k+1.
Then we can limit the evaluation of the integral for given k by
(6.19)
Notice the exception for k=1 and k=N as they are obtained from
(6.20)
(6.21)
The derivatives
and
are the slopes of the functions with the values
where
is assumed to be constant as we limit to constant distance between the points. The

integrals with combinations of
are polynomials that can be integrated analytically. This
gives
Page 21
(6.22)
with
with the exception of
with the exception of
Imposing the essential boundary condition is done principally in the same way as for the
earlier cases by prescribing u1=1. The results shown are shown in Table 6.3. One can note that
it converges slower than when using global trial functions. See also the discussion in chapter
5. However, the condition number increases much slower using the FE-approach. The
condition number is 3.4e8 and the error 2.6e-7 is when using N=10000.
Table 6.3. L2-error and condition number of matrix for Finite Element Method.
Number of
nodes (N)
2
3
4
5
Error
0.77
0.232
0.107
0.061
Condition
number
1.0000
13.9
39.1
66.1
Number of
terms (N)
6
7
8
9
Error
0.039
0.027
0.020
0.015
Page 22
Condition
number
99.8
140.
188.
242
Table 6.4. Excerpt from Matlab code for the Finite Element Method.
F(1)=-dx*0.5;
F(N)=1-dx*0.5;
invdx=1/dx;
K(1,1)=invdx-dx/3;
K(N,N)=K(1,1);
kdiag=2*K(1,1);
koffdiag=-1/dx-dx/6;
K(1,2)=koffdiag;
K(2,1)=koffdiag;
for k=2:N-1
F(k)=-dx;
K(k,k)=kdiag;
K(k,k+1)=koffdiag;
K(k+1,k)=koffdiag;
end
a1=1;
Fmod=K(:,1)*a1;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
a2=1/cos(1);
Fmod=K(:,1)*a2;
F=F-Fmod;
F(1)=[];K(1,:)=[];K(:,1)=[];
6.3 Element based trial and weight functions

We will make some observations on the previous example before we make a more efficient
variant of the finite element formulation.
The previous example showed that the matrix K was banded. It was filled along the diagonal
and the nearest off-diagonals. The general property of finite element matrices is that they are
not full. However, they need not have the nice structure as in the example above. The number
at each position (k,j) can be interpreted in terms of the physics of the problem. One obvious
interpretation is that this number give the contribution from the j:te degree of freedom, uj, to
the k:th equation. In mechanics it is understood as the contribution from the displacement uj
to the nodal equilibrium for the k:th degree of freedom. Then the field u(x) is a displacement
field and the equation we solve is an equilibrium equation expressed in the displacements.
The left end has a given value (displacement in mechanics) whereas the right end has a given
flux. Therefore, it enters naturally into the equilibrium equation as a term, see the right hand
side of Eq. (6.22). That is why it is called a natural boundary condition.
Different ordering of the node numbers, or rather equation numbers13, will change where
different values will be placed in the matrix. The diagonal number couples the response at a
given node with the input to that node. The use of local functions leads to that only nodes that
have functions that overlap will contribute to the integrals and thereby to the numbers at
specific positions in the matrix K. The two examples in Figure 6.3 fill the matrix K
differently. The top example in the figure gives
13
Each node has only one unknown value and therefore nodal numbers and equation numbers can be the same.
Page 23
(6.23)
whereas the lower case gives
(6.24)
Figure 6.3. Two examples of numbering of nodes.

All coefficients in the matrix and vector of Eq. (6.22) that correspond to nodes in the interior
of the domain obtained contributions from an integral over the domains surrounding this
node. One can note that we then integrate twice over the same region. First to give a
contribution to a node on the left side of the segment and then for giving contribution to the
node on the right side. This observation is the basis for the common way to implement the
finite element method based on the concepts of elements. An element corresponds to a region
in which the unknown field is interpolated from connected nodes. The previous FE-example
gives u(x) from linear interpolation between the nodes at the left and right end of the segment.
Based on this observation we make the element the primary entity and describe the total field
by splitting the domain into elements and writing
(6.25)
Nelem is the number of elements used to solve the problem. The field within an element is
written as
(6.26)
where Nnode is the number of nodes associated with the element. Now we change the
notation once more for the trial functions as we switch to element description. Eq. (6.34) can
be written for our earlier example as
Page 24
(6.27)
where
is the element length. This is the same as our dx used earlier. N is the shape
function matrix and Ni is the shape function associated with local node number i of the
element. The shape functions are the most important property of an element. The element is
shown in Figure 6.4. Notice that we have introduced local node numbers, 1 and 2, denoting in
this case the left and right node, respectively. Now it will be very important to keep two
systems of notations apart, the large structure with its numbers and relations etc versus the
local element variables and relations. We will use small letters for local element matrices and
vectors and capital for those associated with the analysed problem. The two node element is a
linear element as it has only two degrees of freedoms for the element (u1 and u2) and thereby
one can only describe a linear variation of the field inside the element. Therefore, it can be
called a linear element in this respect (although it can be used to solve nonlinear problems).
Figure 6.4. Two node element and its linear shape functions N1 and N2.
We will rewrite the element by introducing a local coordinate system also. This is not
necessary for this simple one-dimensional element but extremely important in the general
case. Then the element integrals can be integrated over a simple region, which enables the use
of elements in two and three dimensions without having to stay with rectangular shapes. At
first sight this complicates the element formulation as this variable change makes it
impossible to solve the integrals analytically in the general case. However, the need for
numerical integration to evaluate these integrals is not only very effective but also gives an
additional possibility to develop elements with different properties. Thus this seemingly
drawback is also a possibility in element formulation.
We start with the element formulation in the local coordinate system and will in the end
describe the overall logic corresponding to the nodal based approach in section 6.2.
The element is described in a local coordinate system, s, as shown in Figure 6.5.
(6.28)
This variable change can be described in the same way
(6.29)
This is an example of an isoparametric element as the same parameters used for interpolation
the unknown field is used to interpolate the geometry between the nodes. There exists also
super- and sub-parametric elements but isoparametric is the most common.
Page 25
Figure 6.5. Two node element and its linear shape functions N1 and N2 with a local coordinate
system, s.
The integrals in Eq. (6.19) will now be integrated elementwise as
(6.30)
has the derivatives of both shape functions of the element, one for the left node and one
for the right node. Thus Eq. (6.30) includes the contribution from the trial and weight
functions of these two nodes to the overall system of equations needed. We introduce a
specific notation for the
as this kind of expression is common. We denote it by

(6.31)
We can also move the u-terms outside the integrals as the nodal values do not depend on the
coordinates. Thus we have
(6.32)
Notice that we contribute to two k:s at the same time as
and
has the weight functions
for both nodes of the element. What k:s depend on which global node number the local node
number 1 has and same for local node number 2. Furthermore, the integral only contributes to
two j:s corresponding to u1 and u2 and therefore we do not set the equation above equal to
zero. The j:s thus are the same as the k:s!
We need to extract some information about the variable change before embarking on the
equation above. The equation requires the derivative w.r.t x and not s. We must change the
variable limits and the integration parameter dx. All this has to do with the change of scale
when changing coordinate system. The chain rule gives
(6.33)
Eq. (6.29) gives
(6.34)
J is the Jacobian of the mapping (or variable change) that is the change in scale. Inverting this
gives
Page 26
(6.35)
Now we can write
(6.36)
This answer is what anticipated. The shape functions change from 0 to 1 over the length .
Now we are ready to rewrite our element integrals Eq. (6.32) in the local coordinate system.
We get
(6.37)
The relation above can be expressed in local, element matrices and vectors as
(6.38)
where
Giving
(6.39)
(6.40)
These are the same results as in the previous section expressed in another format. We use
small letters to indicate that these are element matrices. They will give the same resulting
system of equations after what is called the assembly procedure. The latter is a step that was
not needed in the nodal based approach as we did directly form the global matrices and no
element variables where we have a local numbering system, in this case from 1 to 2. As stated
earlier- we need to identify for each element which j,k our nodes correspond to. We will
illustrate this in the Box below by making an assembly of the using the upper model in Figure
6.3. This is demonstrated for the lower model in Figure 6.3 in Box 6.2 but in a way that is
another step towards what is done in an actual FE-software.
Box 6.1. Assembly procedure.
Logic for element assembly
Zero global matrix K, a 4x4 matrix, and global vector
Page 27
, a 4x1 vector.
Element 1 has node number 1 as local node number 1 and node number 2 as right node. Thus
we add the matrix k according to Eq (6.39) to K and vector
to
. Then we get
Notice the change in sign as we had

hand side.
on left hand side in Eq. (6.39) but
is on the right
Element 2 has node number 2 as left node 1 and node number 3 as local node number 2. Thus
we add the matrix k according to Eq. (6.39) to K and vector
to
. Then we get
Element 3 has node number 3 as left node 1 and node number 4 as right node. Thus we add
the matrix k according to Eq. (6.39) to K and vector
to
. Then we get
This is the same as would be obtained using the formula in Eq. (6.22)
Box 6.2. Automated assembly procedure.
Logic for element assembly
Zero global matrix K, a 4x4 matrix, and global vector
, a 4x1 vector.
Element 1
Set up identification vector where the i:th position gives the global equation number of the
degree of freedom that has local number i. It is
Page 28
This is used to automatic locate where corresponding rows/columns of local matrices/vectors

are in the global ones. Thus we use place this as transpose above the matrix to identify global
column numbers and to the right to identify global row numbers where corresponding values
should be assembled. Thus element one with global node numbers 1 and 2 has
Then we get
Notice the change in sign as we had

hand side.
on left hand side in Eq. (6.39) but
Element 2
with
Giving
Element 3
with
Page 29
is on the right
Giving
7 Numerical integration
The previous chapter brought us to isoparametric element but only in the one-dimensional
context. Next chapter generalises this to multidimensional problems where it will be clear that
we cannot form an analytic solution of the element integrals. Thus we prepare by introducing
the concept of numerical integration14.
A one-dimensional integral over the domain [-1,1] of a polynomial, f, can be integrated
exactly by the formula
(7.1)
where are the integration points,
are their weights and nint are the number of integration
points. The most common rule is Gauss rule15. It is the most effective for one-dimensional
integrals. There are other rules, for example the Lobatto rules that have points at end of
interval. However, we discuss only the Gauss integration rule here. It is the most common and
therefore one often uses the notation Gausspoints about the sampling point for the integrand f.
The higher the polynomial to integrate, the more integration points are needed. The Gauss
integration rule is shown in Table 7.1. The relation between the degree of the polynomial, n,
and number of integration points is
(7.2)
Table 7.1. Gauss integration rule.
Nint
Integration points
weights
1,1
14
http://en.wikipedia.org/wiki/Numerical_integration
15
http://en.wikipedia.org/wiki/Gaussian_quadrature
Page 30
3
The rule in Table 7.1 is used to integrate a cubic polynomial. The analytic solution is
(7.3)
The needed two point rule gives
(7.4)
This rule is applied cross-wise for two- and tree-dimensional problems by
(7.5)
or
(7.6)
The exact integration of the element integrals is the basic idea in the finite element
formulation. However, it will be obvious that we do only have an estimate of the degree of the
polynomial to be integrated. However, it is a good estimate provided the shape of the element
does not deviate too much from a square. This is what is called an exact integration in the
finite element context. The discussion of the relation between Galerkin and Ritz method in
section 0 noted that it is possible to state how the finite element method converges when the
number of degree of freedoms is increased. The solution converges from above down to the
true minimum of some kind of function16. However, this is provided the integrals are
integrated exactly. This opens up a possibility to improve the finite element methods.
Underintegration may make it possible to converge faster, but not necessarily from above.
This is illustrated in Figure 7.1. The numerical integration is sometimes reduced on some
terms of the matrices in order to improve elements. The technique is called selective reduced
integration, see section 8.2.
Thus the numerical integration process is not only efficient but can also be used as a trick to
improve elements in different respects. They can be underintegrated or selective
underintegrated when only some specific terms of the integrals are underintegrated. However,
too few points may lead to failure as it may occur that the contribution from certain
combination of element variables may be lacking and lead to a matrix with a too high
condition number. In terms of mechanical problems it is called zero-energy deformation. The
sampling of strain and stress that occurs at the integration points did not discover any strain
due to these modes. This must be avoided or prevented.
16
It is the total potential energy in case of elasticity problems.
Page 31
Figure 7.1. Convergence in terms of total potential energy in case of elastic problem.
8 Beam elements
Ritz method is illustrated for a one-dimensional mechanical problem, the beam element, in
the current chapter. Ritz method corresponds in this context an energy method, minimisation
of the total potential energy of the system. Two variants based on slightly different theories,
one for the classical beam theory and one for less slender beams, Timoshenko beam theory,
are formulated. Selective reduced integration, chapter 7, will be demonstrated on the latter
element.
8.1 Bernoulli beam
The theory for the bending deformation of the Bernoulli beam theory or sometimes called
technical beam theory, applicable to slender beams is given in section 12.1.
The solution of beam problems should make the total potential energy, see Eq. (12.10),
minimum. Thus we want to find the minimum of this expression. This is the Ritz method, see
section 4.5. It is a special case of the Galerkin method. The approximate solution should fulfil
essential boundary conditions and minimise the energy expression
(8.1)
Fulfilling essential boundary conditions are done by assigning appropriate nodal values to the
model as will be seen in section 8.3. The element stiffness matrix and load vector will be
formulated by applying Eq. (8.1) to one element. A few notes are at place first.
The derivatives in the functional are higher than in the example in section 6.2. The
fundamental equilibrium equation expressed in displacement, Eq. (12.9), is a fourth order
differential equation. The approximate solution must at least be able to represent constant
curvature
. Thus it needs to be at least cubic. The function must give continuous
derivatives/slope between elements. Otherwise there is a risk that the split of the integral Eq.
(8.1) into elements may lose energy in the boundaries between these elements. But if the
slope is continuous between elements, then the curvature must at least be finite and then no
Page 32
energy is lost at the point joining two elements. The fulfilment of the required C1-continuity17
of the deformation w between the elements and the essential boundary conditions is simplified
by the use of displacement and slope/rotation as nodal variables. Then a two node beam
element will have two degree of freedoms per element, see Figure 8.1. They can define the
four coefficients of a third order polynomial. The element field is written as
(8.2)
The shape functions are the polynomials shown below.
(8.3)
These functions have the properties
(8.4)
and
(8.5)
The cubic equation is the exact solution to a beam bending problem where only nodal loads
are applied, see Appendix 12.1.
Figure 8.1. Bernoulli beam element.

We need the second derivatives, i.e. the curvature, of the beam in Eq. (8.1). This is
(8.6)
with.
17
The previous case fulfilled C0-continuity. The notation Cn-continuity of a function means that the n:th
derivative of the function is continuous.
Page 33
(8.7)
Then Eq. (8.1) is written as, assuming only one element in the model,
(8.8)
(8.9)
(8.10)
Now we can identify element stiffness matrix as
(8.11)
and consistent nodal load vector due to the distributed load q
(8.12)
The word consistent means that the nodal forces and moments in
give the same
contribution to the potential of the loads as the original distributed load would have given in
Eq. (8.1).
The stiffness matrix becomes
(8.13)
That can be solved analytically or numerically. The latter would require, see chapter 7, two
integration points to for this 2nd order polynomial. We would get
(8.14)
and for the load, assuming q=constant along the beam
(8.15)
Page 34
One can check with elementary cases for a beam that the above edge loads on a beam gives
the same displacements and rotations as the distributed load would have done. However, the
bending between the nodes will be different.
An analysis of a beam problem is shown in section 8.3.
8.2 Timoshenko beam
The Timoshenko beam element in Figure 8.1 is formulated directly by use of the Tonti
diagram in Figure 8.2. See appendix, section 12.2, for more about the Timoshenko beam
theory.
Figure 8.2. Tonti diagram for a two node Timoshenko beam element.
We use an isoparametric approach by
(8.16)
with
(8.17)
The rotation and displacements are interpolated independently as
(8.18)
This is a large difference to the Bernoulli beam case in Eq. (8.3). This gives
(8.19)
We can compute local derivatives, with respect to the coordinate s. However, the global
derivatives are needed. The isoparametric mapping is the same as in Eq.s (6.34). The Jacobian
of the mapping is
(8.20)
where
is the length of the element. The B-matrix becomes then
(8.21)
Page 35
The total potential energy18, section 12.2, is

(8.22)
The elastic stored energy in the first term is leads to the stiffness matrix in the same way as in
for the Bernoulli beam. The element stiffness matrix is
(8.23)
A split is done before solving the integral above. The stress-strain relation matrix is
diagonal as the strain energies due to shear and normal straining are uncoupled. The stiffness
matrix is related to the stored elastic energy as in Eq. (8.10) and can therefore also be
uncoupled18. It is written as
(8.24)
where
(8.25)
and
18
The stored energy can also be written as
showing that the energies can be
added separately.
Page 36
(8.26)
Exact integration gives for the latter
(8.27)
We will in the next chapter apply this exactly integrated element stiffness matrix and find
problem. We will then also use a version where the highest order terms in Eq. (8.26) are
underintegrated. We use one integration point formula according to Table 7.1 and get
(8.28)
and same value for
. The other term to be approximated is

(8.29)
Then we have
(8.30)
The consistent nodal load vector due to the distributed load q is
(8.31)
Notice the extension of q to a vector with two parts. One is related to the deflection of the
beam and the other to its rotation. Originally, Eq. (8.31) comes from the potential of the loads
that is the loads multiplying the deformations in Eq. (8.22). Then we had only qw in the
product. Now we have, see Eq. (8.18), deflection and rotation as fields describing the
deformation independently. The potential of the loads is therefore both force multiplying
deflection w and distributed moment multiplying rotations. Thus motivating Eq. (8.31).
Integration of above for constant distributed loads gives simply
Page 37
(8.32)
8.3 Cantilever beam problem

We will compare finite element solutions with the analytic solution according to Bernoulli
beam theory in section 12.1. A model with one element will be used to solve the cantilever
beam with a point load, upper part in Figure 8.3. The reaction force and moment are replacing
the wall that was drawn in Figure 12.5. They are unknown and their magnitudes comes from
the fixed displacement and rotation at the wall, x=0.
Figure 8.3. Cantilever beam with point load and a one element model.
The problem is shown and solved analytically in section 12.1. The solution is
The solution is thus
(8.33)
The maximum displacement (downwards) is
The one element model is shown in the lower part of Figure 8.3. The fulfilment of the
essential boundary conditions at x=0 requires the rotation and displacement at left end to be
zero. The corresponding unknown reaction force and moment are included in the systems of
equations below.
The Bernoulli beam gives, Eq. (8.14),
Page 38
The unknown displacement and rotation at the free end can be obtained from the system of
equations above but first the essential boundary conditions must be inserted, the two zeros in
the vector with unknown nodal degree of freedoms. The in is only necessary to use the two
last equations for our wanted deformation. The first two can be used afterwards to determine
reaction force and moment at the left end. Thus the system to be solved is
(8.34)
giving
(8.35)
This is the exact solution as expected as our element has cubic variation in the displacement,
which is sufficient to represent the theoretical solution. Particularly we check the deflection at
the end with the analytic solution in Eq.
(12.32)
We get
(8.36)
The fully integrated Timoshenko beam, Eq. (8.25) and Eq. (8.27), gives,
This leads to
(8.37)
A ratio between shear and bending is introduced order to simplify the expression above
(8.38)
Eq. (8.34) can be written as
(8.39)
Page 39
The algebraic solution is
giving
(8.40)
Assuming a square cross-section of the beam gives
(8.41)
Insertion of above into Eq. (8.38) gives
(8.42)
This makes the deflection of the edge of the beam, Eq.(8.40),
However, there is a finite precision in the computer when solving the system of equations and
for slender beams (H<<L) then
Then the computer will experience Eq. (8.39) as

(8.43)
This gives the solution
Leading to
This will be a very small number! Thus the solution is very bad. Before discussing this
phenomenon, lets evaluate the results from the underintegrated Timoshenko beam, Eq. (8.30).
That formulation gives
Page 40
(8.44)
The solution becomes
leading to
(8.45)
This result will not suffer from any truncation error as the fully integrated version does.
The underintegrated Timoshenko beam element handles the slender beam case much better
than the fully integrated version. The table below summarise the result for the two
Timoshenko element formulations and the Bernoulli beam element. The Timoshenko beam
element has a linear interpolation of the rotation and the deflection. The deformation of a
slender beam, i.e. a case where the Bernoulli beam theory approximation is good, is such that
the rotation is the first derivative of the deflection. This cannot be handled well when the two
fields are both linear polynomials. The fully integrated element locks in order to resolved
this issue. Thus the deflection and rotation becomes very small in an attempt to remove the
shearing from the deformation of the beam. The underintegration is performed on the terms
that have to do with the rotation field. Thus the element will, despite the use of a linear shape
function, only experience one value for the rotation as this is sampled at the centre of the
element. Therefore, this element performs better.
The above explanation of the phenomenon underlying the truncation problem of the fully
integrated element can also be discussed in terms of energy. The fully integrated element will
lock in order to reduce the strain energy due to the shearing. The underintegration of this part
of the element stiffness matrix makes the element lose some of that strain energy and thereby
softens the element.
Table 8.1. Computed maximum deflection of cantilever beam. Finite element results are
normalised versus theoretical result. The beam has a thickness/length ratio of 0.01.
Number of
elements
Normalised results
Bernoulli beam element
1
2
4
8
16
0.9999
0.9999
0.9999
0-9999
0.9999
Underintegrated Timoshenko
beam element
0.750
0.938
0.984
0.9961
0.9990
Page 41
Fully integrated Timoshenko

beam element
0.0003
0.0012
0.0049
0.0192
0.0726
9 Isoparametric mapping in two dimensions

The concept of isoparametric mapping was introduced in section 6.3 in the context of onedimensional elements. However, it was unnecessary for this simple type of elements as the
element integrals always have simple boundaries, left and right end of element. The
motivation for showing the formulation was that is simple to show in this context. It was a
prelude to chapter 10 where the isoparametric mapping is a necessary step in the finite
element formulation. There it is shown for a two-dimensional element.
Chapter Error! Reference source not found. described the numerical integration technique.
This gives added possibilities to tailor elements as was described in the case of the
Timoshenko beam in section 8.2. However, the original motivation for applying numerical
integration was that use of isoparametric mapping caused the element integral to be too
complex to be subjected to analytic integration when applied formulated for two- and threedimensional elements.
All the above will be brought together in the next chapter. There all the steps in formulating
an isoparametric element will be demonstrated. The application to a two-dimensional element
is sufficient general to expose all steps needed to form any element. Mastering these steps
together with the basic equations in the relevant field enables formulation of a wide range of
finite elements for linear problems as long as long as the physics does not include additional
complexities like nonsymmetric influences. The latter problem is introduced in chapter 11.
The details of the consequences of the isoparametric mapping on the element integrals will
shown in the current chapter before going to the element formulation chapter 10. The change
of variables will require the determination of the change in scale. This is needed for obtaining
derivatives w.r.t the global coordinate system from the derivatives w.r.t the local coordinate
system. This information is also needed when evaluating the integrals as the area element
dA=dxdy in the integrals is changed.
A two dimensional elements with four nodes is taken as an example. Its geometry is defined
in a finite element model with respect to a coordinate system (x,y). The element integrals are
evaluated after a change of variables to a simple coordinate system (s,t) where the element
integrals have simple boundaries. This mapping, or variable change, is shown in Figure 9.1.
Page 42
Figure 9.1. Isoparametric mapping of a four node element in two dimensions. The element is
a square between (-1,-1) to (1,1) in the local coordinate system.
The same shape functions are used for the interpolating of geometry between the nodes as
interpolating the displacements field. The specific shape functions for this element are given
in the next chapter but not needed now. The derived formulas will be valid for any kind of
element and are easily expanded to the case of three-dimensional integrals. The isoparametric
mapping is written as
(9.1)
It can be written in matrix form as
(9.2)
The interpolation of the displacement fields is written in the same format, see Eq. (10.1) in
next chapter. The derivatives in the equations to be solved are expressed in the global
coordinate system. However, the fundamental variables are given with respect to the local
coordinates system as the elements shape functions Ni are so defined. The previous, Eq.
(9.1), relation between the two coordinate systems is used together with the chain rule. The
chain rule states that if we have a function h that is an explicit function of s and s in turn is
some kind of function of x. Then it is possible to write
(9.3)
Page 43
This can be written as the operator below for the partial derivatives of a function of two
coordinates
or in matrix form
(9.4)
However, we can not compute the derivatives
as we have only x(s,t) and
y(s,t) and not the other way around. Then we set up the derivation in the opposite direction to
Eq. (9.4)
(9.5)
The terms in the Jacobian matrix, J, can be computed from the isoparametric mapping in Eq.
(9.1) as
(9.6)
Then it is possible to invert the expression to give Eq. (9.5).

(9.7)
where
(9.8)
The above derivations are shown in Figure 9.2 where the information in the derivatives and
the determinant of the Jacobian are hopefully clearer. It shows how an infinitesimal area
element dA=dsdt becomes a parallelogram. The mapping of a small area element is always a
linear mapping.
Page 44
Figure 9.2. Change in scale due to change of variables.
10 A four node plane stress element

The purpose of this chapter, as stated already in the introduction of the previous chapter, is to
bring together all the previous developments into the formulation of a two-dimensional
element. An element applicable for solving two-dimensional plane stress19 problems will be
derived. The basic continuum mechanics20 relations are not given and the reader can consult
any standard textbook or book with formulas for solid mechanics or the links below. The
formulation steps used are general enough to enable the formulation of elements for several
types of linear problems. These general formulas are listed below and thereafter they are
detailed for a four node element. The basic relations are also summarised in Figure 10.1
19
http://en.wikipedia.org/wiki/Stress_(physics)#Plane_stress
20
http://en.wikipedia.org/wiki/Strain_(materials_science),
http://en.wikipedia.org/wiki/Strain_tensor#Infinitesimal_strain_tensor,
http://en.wikipedia.org/wiki/Hooke%27s_law
Page 45
Figure 10.1. The shape functions determines the basic properties of an element. They may be
modified by use of underintegration. The diagram shows the formulation steps that follow.
The displacement field within an element is written as
(10.1)
where the shape functions in the matrix N fulfils some requirements listed below.
1. The Kronecker delta property. The shape function associated with node J has the property
(10.2)
2. The partition of unit property. The summation of the shape functions that overlap at a
coordinate x must be unity as given below
(10.3)
3. The shape functions must be sufficient continuous. If the fundamental equation to be
solved has an 2m:th derivative, then the WRM formulation after appropriate partial
integration21 will have a derivative of m. Thus it must be possible to take the derivative of the
functions at least m times. Furthermore, it must be m-1 continuous between elements, Cm-1continuity.
4. The functions must include all lower terms of the family of functions that are used to
generate them. This means that for polynomials, all lower order terms must be present before
including higher order terms. Furthermore, the different coordinates should be equal present
so that the elements properties will not depend on which direction its local coordinate
systems has.
The strains needed in the plane stress problem are computed as combinations of derivatives
from the displacements fields and this leads to
(10.4)
Furthermore, the stresses are related to the strains by Hookes law in case of linear, elastic
problems
21
This is called geometric compatibility. Some incompatible elements exists but then this incompatibility must
disappear as the elements are made smaller in order to guarantee convergence of the solution.
Page 46
(10.5)
Combining these, see Figure 10.1, leads to a general formula for the stiffness matrix
(10.6)
and the consistent load vector
(10.7)
These general relations are shown in detail in the following sections for a four node plane
stress element.
10.1 Displacements and geometry of four-node plane element
The local coordinate system of the element is shown in Figure 10.2. This is the basis for
formulating the element and where the element integrals, Eq. (10.6) and Eq. (10.7) are solved.
The shape functions are
1
N1 = (1 s)(1 t )
4
1
N 2 = (1+ s)(1 t )
1
4
N i = (1+ si s)(1+ t i t )
(10.8)
1
4
N 3 = (1+ s)(1+ t )
4
1
N 4 = (1 s)(1+ t )
4
where the local coordinate ( si ,t i ) is the coordinate of corner i.
Figure 10.2. Local coordinate system of plane, four node element.

The isoparametric mapping from the local to the global coordinate system, see Figure 9.1, can
be written in some variants
Page 47
(10.9)
The first variant with a sum over nnode sub-matrices is a convenient way to programme the
logic for an element with arbitrarily number of nodes. Then follows the specific form for a
four node element and finally the general matrix form is given. c is a vector with nodal
coordinates of the element.
Thus the same form (isoparametric) is used to describe the interpolation of the displacements
as for the geometry. It is written as
(10.10)
10.2 Strains of four-node plane element

The strain components needed for plane stress are
(10.11)
Insertion of the shape functions, Eq. (10.10), for the four node element gives
Page 48
(10.12)
It can be written in the node summation style as
(10.13)
The determination of global derivative of the shape functions are done as shown in chapter 9.
The local derivatives of the shape functions are set up first. The Jacobian is determined by Eq.
(9.6). Notice that it depends on the coordinates of the element as expected. Thereafter Eq.
(9.7) is used to calculate the global derivatives needed for Eq. (10.12). The strains are
explicitly expressed in the local coordinates. If the strain in needed ad a coordinate x, then the
local coordinate must be solved first from Eq. (10.9) and thereafter the found local coordinate
is inserted into Eq. (10.12). Eq. (10.9) can in the general case not be solved analytically but
must be solved numerically.
10.3 Constitutive model for plane stress
The relation between strains and stresses, Eq. (10.5), is
(10.14)
where E is the modulus of elasticity and is Poissons ratio.
10.4 Flow chart for calculation of four node, plane stress element
All terms for the stiffness matrix and consistent load vectors are now available. The logic for
the solving the element integrals is given in Box 10.1. It is assumed that the material
properties are constant over the element. The B matrix in the integral for the element stiffness
is nearly constant, with exception for one term. The st-term in the shape functions, Eq. (10.8),
gives one linear term. Thus most of the terms in the integral for the stiffness matrix, Eq.
(10.6), are constant or linear. However, the highest term is quadratic. This requires nint =2*2
integration points according to Eq. (7.2).
Page 49
Box 10.1. Procedure for the computation of element stiffness matrix.

Set up constitutive matrix E, Eq. (10.14).
For igaus from 1 to nint integration points
Look up the location and weight of current point integration point sigaus, tigaus and wigaus22
Evaluate the shape functions
and their local derivatives
with inode from 1 to nnode.

Compute the Jacobian matrix, Eq. (9.6), and its determinant, Eq. (9.8), at the integration point
Compute the global derivatives of the shape functions according to Eq. (9.4)
Place the global derivatives at appropriate positions in the B-matrix. Evaluate the contribution
to the element integrals from the current integration points
and sum this into k.

Next integration point
11 Convective heat transfer

The previous equations are quite straightforward to solve using the finite element method.
There are other equations that, even in the linear case, need special precautions in order to be
tractable. One example discussed below is the inclusion of a convective term in the heat
conduction equation. This equation describes the combined conduction and convection of
heat in a medium with a given flow where the material moves with respect to the coordinate
system. The equation is also obtained when introducing a moving coordinate system in the
heat conduction equation. Then the coordinate system is moving with respect to the material.
The one-dimensional heat conduction equation23 in absence of heat sinks is written as
(11.1)
22
This is the weight produced by multiplication of two appropriate weights according to Eq. (7.5).
23
http://en.wikipedia.org/wiki/Heat_conduction
Page 50
where is the coordinate, is the density, c is the heat capacity and is the heat
conductivity. Then we switch to a moving coordinate system, x, by
where v is the velocity. This transforms24 Eq. (11.1) to
(11.2)
(11.3)
Furthermore, the solution is assumed to be steady-state w.r.t. this moving coordinate system
leading to
(11.4)
The analytic solution of this problem with the boundary conditions
(11.5)
(11.6)
is
(11.7)
This solution can be expressed by introducing the Peclet number. It measures the relative
strength of convection and conduction of heat. It requires a reference length and we use the
length of an element for this. The number becomes
(11.8)
The analytic solution can now be written as
(11.9)
The finite element solution leads to a global system of equations

(11.10)
KT = 0
The standard contribution to K from the for conductivity matrix using a two-node linear
element is, see section 12.3,
1 1
(11.11)
l e 1 1
v (e )
The additional contribution due to the convection part can be derived by comparing the
convective term with the u-term in Eq. (6.1) or by comparing it with the q-term in Eq. (8.1).
Using the latter leads to an expression for consistent nodal loads shown in Eq. (8.12). Thus q
k cond =
T
th
B th dv =
is repleced by
24
in this equation. This gives
http://en.wikipedia.org/wiki/Total_derivative
Page 51
(11.12)
The gradient term is
(11.13)
Combining the last two equations gives
1
cv 1 1 e
(11.14)
T
2 1 1
1
Thus it will also contribute to the K-matrix. Summing the two matrices, Eq. (11.13) and Eq.
(11.14), and using Eq. (11.8) gives the combined element matrix
T
cvN BJdsT
= k convTe =
1 1 1 Pe 1+ Pe
1 1
(11.15)
+ Pe
= e
e
l 1 1
1 1 l 1 Pe 1+ Pe
to be assembled into Eq (11.10). This formulation becomes unstable when Pe=1 as can be
seen in the example below. This numerical problem can be reduced by enhancing the
conductivity so that the Peclet number is modified to
according to the equation below.
k=
(11.16)
The classical formulation is compared with the stabilised formulation and theoretical
solutions for Pe=1 in Figure 11.1 and Pe=3 in Figure 11.2.
Figure 11.1. Temperature in fluid flow with given inlet and outlet temperatures for Pe=1.
Page 52
Figure 11.2. Temperature in fluid flow with given inlet and outlet temperatures for Pe=3.
The heat transfer in a material moving with the velocity v in positive direction and a point
source at the coordinate xc is
(11.17)
where v is now interpreted as the material flow relative to the heat source. Thus the heat
source is moving towards the left in the figure below. The boundary conditions and the finite
element formulations are taken as the same as in the previous example. It is assumed that the
point source is applied on a node. Then there will be one contribution to the load vector and
Eq. (11.10) becomes
(11.18)
The enhancement of the heat conductivity that gives
in Eq. (11.16) is
(11.19)
where
(11.20)
The value
is placed at the appropriate position in
corresponding to the node
at xc. The classical formulation is compared with the stabilised formulation and theoretical
solutions for Pe=1 in Figure 11.3 and Pe=3 in Figure 11.4.
The improved formulation using an enhanced convection is a prelude to Stream Upwind
Petrov-Galerking (SUPG) formulations. This it is can be derived using a Petrov-Galerkin
approach.
Page 53
Figure 11.3. Temperature due to a point heat source moving to the left, Pe=0.5.
Figure 11.4. Temperature due to a point heat source moving to the left, Pe=1.5. Wiggles at
end of analytic solution due to inaccuracy in its evaluation.
Page 54
12 Appendix : Beam theories

12.1 Bernoulli beam
The classical beam theory assumes that a line orthogonal to the centre line of the beam
remains straight and orthogonal as shown in Figure 12.1. Then the axial displacement due to a
deflection of the centre line is
(12.1)
This gives a linear variation of the axial strain as
(12.2)
All other strains are zero due to the assumed deformation in Eq. ((12.1). The bending moment
corresponding to the linear axial stress distribution, see Figure 12.2, becomes
(12.3)
The last integral is the definition of area moment of inertia giving
(12.4)
The curvature has been introduced above as
(12.5)
Figure 12.1. Assumed deformation of beam cross-section according to Bernoulli beam theory.
Figure 12.2. Axial stress distribution and equivalent bending moment.
Moment equilibrium for an infinitesimal element, Figure 12.3, gives
leading to
(12.6)
Transverse force equilibrium gives
leading to
(12.7)
Combining Eq. (12.6) and Eq. (12.7) gives
(12.8)
Figure 12.3. Equilibrium of infinitesimal piece of beam.

Insertion of Eq.(12.4) gives
(12.9)
This is the equilibrium equation for a beam expressed in the displacement field w(x). The
basic relations are summarised in Figure 12.4. Notice that there are transverse forces playing a
role in the equilibrium equations. They are related to shear stresses on the cross-section. BUT
the assumption about the deformation of the cross-section states that there is no shear
straining. This is implied in the statement that the line orthogonal to the centre line remains
orthogonal.
There exist corresponding formulations for plates and shells. Assumptions about the
deformation behaviour with respect to a reference surface reduce the dimension of the
problem but raise the order of the differential equation.
The boundary conditions for the beam are of the following types;
Essential boundary conditions at one of the ends are prescribed deflection w or

rotation
Natural boundary conditions are prescribed end moment or transverse force. They
correspond to
or
according to Eq.s (12.4) and (12.6).
Page 2
Figure 12.4. Tonti diagram for beam relations

The total potential energy of the beam has a contribution from axial stresses and external
loading. Assuming that we have only distributed load q gives
Hookes law, =E, was used in the last step above. There is no need to include shear stresses
as the corresponding strains are zero and therefore do not contribute to the stored elastic
energy in the first integral. Insertion of Eq. (12.4) gives
Some quantities does only vary with the x-coordinate
Use of the definition of area moment of inertia gives finally

(12.10)
A cantilever beam problem with a point load at the end as shown in Figure 12.5 is defined by
Eq. (12.9) and the appropriate boundary conditions;
no displacement near wall
(12.11)
and no slope near wall
(12.12)
No moment at right edge
(12.13)
Given transverse force at right edge
Page 3
(12.14)
Figure 12.5. Cantilever beam with point load.

Integration of Eq. (12.9) four times with q=0 gives
(12.15)
Thus the exact solution for a beam loaded only at its edges is no higher than a cubic
polynomial. This observation can be related to the beam element formulation in chapter
Figure 8.1. The two first boundary condition at x=0 gives A=B=0.
The boundary condition Eq. (12.14) gives
(12.16)
and the use of Eq. (12.13) leads to
(12.17)
The solution is thus
(12.18)
The maximum displacement (downwards) is
(12.19)
12.2 Timoshenko beam
This beam theory is an extension of the previously described Bernoulli beam theory. It
assumes that a line orthogonal to the centre line of the beam remains straight but not
orthogonal as shown in Figure 12.1 but now as in Figure 12.6. Thus we allow for a constant
shear strain over the thickness of the beam. This is a step forward. Previously it was noted
that the transverse force corresponds to shear stresses. Shear stresses are obtained from shear
strains by the shear modulus G as
(12.20)
However, analysis not included above shows that the shear stresses are largest in the interior
of the beam and zero at the top and bottom surfaces. Thus the shear strains are also varying
with z. The Timoshenko beam theory includes shear, , but assumes it to be constant over the
height of the beam. Then the axial displacement due to a deflection of the centre line is
(12.21)
This gives a linear variation of the axial strain as
Page 4
(12.22)
The axial strain cannot be related to the second derivative w anymore. Accordingly the
formula for the cross-sectional moment, see Figure 12.2, becomes somewhat different than
for the Bernoulli beam. It becomes
(12.23)
Now the relation between shear stresses and transverse force, see Figure 12.7, is needed.
(12.24)
The parameter was introduced for enabling the simplification of constant shear over the
cross-section. It gives an effective shear area
(12.25)
The value depends on the shape of the cross-section of the beam and is tabulated.
Figure 12.6. Assumed deformation of beam cross-section according to Timoshenko beam

theory.
Figure 12.7. Transverse force and shear stresses.

The equilibrium relations for an infinitesimal element, Figure 12.3, are the same as for the
Bernoulli beam giving
(12.26)
and
(12.27)
Combining them gives
Page 5
(12.28)
There exist corresponding formulations for plates and shells and then the theories are called
thick plates/thick shells. Sometimes the plate theory is called Mindlin plate theory. The basic
relations for the Timoshenko beam are shown in Figure 12.8.
We will not formulate the equilibrium equation corresponding to Eq. (12.9) but proceed
directly to the expression for the potential energy. The elastic energy is now a sum of energy
due to shear and normal deformations.
e =
2 + 2 dAdx
(12.29)
Insertion of Eq. (12.22) and Eq. (12.23) in the integral above and integrating over the crosssection gives
(12.30)
The total potential energy of the beam has a contribution from axial stresses and external
loading. Assuming that we have only distributed load q gives
(12.31)
The basic relations are summarised below. The vectors
formulating a finite element in chapter 8.2.
and
will be introduced when
Figure 12.8. Tonti diagram for Timoshenko beam theory.

The solution according to Timoshenko beam theory of the deflection of the beam in Figure
12.5 with a square cross-section is
Page 6
where
and for
This gives the maximum deflection
(12.32)
12.3 One-dimensional element for heat conduction
The one-dimensional heat conduction equation23 is written as
(12.33)
where is the density, c is the heat capacity and is the heat conductivity. A Tonti diagram
for the basic relations is shown in Figure 12.9.
Figure 12.9. Tonti diagram for heat conduction equation.

The finite element equations corresponding to Figure 12.9 are shown in Figure 12.11. They
can be derived using the Weighted Residual Method following the same steps as in chapter 6.
The temperature in an element is interpolated by
(12.34)
where the finite element shape functions for a two node element, see Figure 12.10, are the
same as in section 6.3
(12.35)
The element conductivity matrix is
Page 7
(12.36)
This gives
(12.37)
The element heat capacity matrix is
(12.38)
The use of a lumped hat capacity matrix can be advantageous sometimes. It is
(12.39)
Figure 12.10. Two node element and its linear shape functions N1 and N2.
Figure 12.11. Tonti diagram for finite element relations for heat conduction.
13 References
1.
Hutton, D., Fundamentals of Finite Element Analysis. 2004: McGraw-Hill.
2.
Belytschko, T., W.K. Liu, and B. Moran, Nonlinear Finite Elements for
Continua and Structures. 2000: John Wiley & Sons. 650.
Page 8
3.
Fletcher, C., Computational Galerkin Methods. Springer Series in
Computational Physics, ed. H. Cabannes, et al. 1984, New-York: Springer-Verlag. 320.
4.
Onate, E., J. Miquel, and G. Hauke, Stabilized formulation for the advectiondiffusion-absorption equation using finite calculus and linear finite elements. Computer
Methods in Applied Mechanics and Engineering, 2006. 195: p. 3926-3946.
5.
Levitas, V., et al., Numerical modelling of martensitic growth in an elastoplastic
material. Philosophical Magazine A, 2002. 82(3): p. 429-462.
6.
Codina, R., Finite Element Formulation for the Numerical Solution of the
Convection-Diffusion Equation. Vol. 14. 1993: CIMNE.
Page 9

MWR Galerkin Fem

Uploaded by

Copyright:

Available Formats

MWR Galerkin Fem

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MWR Galerkin Fem

Uploaded by

Copyright:

Available Formats

FromWeightedResidualMethods

3 Short finite element course

4 Weighted Residual Methods

. Then it is called a boundary method.

iii) Else it is a mixed method.

We will illustrate some variants of WRM methods in the following.

, which may overlap. The weight function is then

Comparison of WRM methods

4.7.1 Problem definition and exact solution

where the right hand side is

This has same meaning as

where it is implied a summation over the repeated index j.

4.7.3 Collocation example

4.7.4 Least squares method

4.7.5 Method of moments example

The matrix on the left hand side becomes

Table 4.8. Excerpt from Matlab code for method of moments.

4.7.6 Galerkin example

Further elaboration gives leads to a system of equations with

The matrix on the left hand side becomes

Table 4.10. Excerpt from Matlab code for Galerkins method.

4.7.7 Summary of comparisons

Equivalent to finitevolume method;

5 Classical and computational Galerkin methods

6 Finite Element Methods

The matrix on the left hand side becomes

6.2 Nodal based trial and weight functions

. This is reasonable as we use functions

are the slopes of the functions with the values

is assumed to be constant as we limit to constant distance between the points. The

The matrix on the left hand side becomes

with the exception of

6.3 Element based trial and weight functions

Figure 6.3. Two examples of numbering of nodes.

as this kind of expression is common. We denote it by

Notice the change in sign as we had

on left hand side in Eq. (6.39) but

This is used to automatic locate where corresponding rows/columns of local matrices/vectors

Notice the change in sign as we had

on left hand side in Eq. (6.39) but

It is the total potential energy in case of elasticity problems.

. Thus it needs to be at least cubic. The function must give continuous

These functions have the properties

Figure 8.1. Bernoulli beam element.

is the length of the element. The B-matrix becomes then

The total potential energy18, section 12.2, is

The stored energy can also be written as

showing that the energies can be

Exact integration gives for the latter

. The other term to be approximated is

The consistent nodal load vector due to the distributed load q is

8.3 Cantilever beam problem

The algebraic solution is

Then the computer will experience Eq. (8.39) as

Fully integrated Timoshenko

9 Isoparametric mapping in two dimensions

However, we can not compute the derivatives

as we have only x(s,t) and