Chapter 6 Poisson’s Equation: FEM Approximation

In this chapter we use the two-dimensional finite element function to obtain

approximate solutions to boundary value problems described by Poisson’s and
Laplace’s equations. These equations are adequate to describe a large number of
applied problems in engineering, some of which are described in the next chapter.

6.1 Governing Equations

We begin with a description of the partial differential equation (or strong form) and
then develop the weak form suitable for finite element analysis.
6.1.1 Strong Form
The following equation and boundary conditions (along with Figure 6.1) define the
two-dimensional problem we wish to solve by the finite element method.
       
+ = −Q( x, y )
x  x  y  y 
kx ky in V

 
kx nx + k y n y = q*
x y on S
 = *
where
nx, ny are the x and y components of the outward unit normal vector to S
q*, Φ* are the boundary values of the quantities shown

Figure 6.1. Domain of analysis.

A comment about notation: We have chosen to use V (for volume) and S (for surface)
to indicate the domain of analysis and its boundary. Clearly, for a two-dimensional
analysis, A (for area) might appear more appropriate. However, for most of the
applied problems that we consider, the domain of analysis will be a volume of unit
thickness. Hence, in order to emphasize this dimensionality, we use the symbols V and
S.
A comment about boundary conditions: Eq. 6.2 defines two quantities associated with
the bounsary. For the problems we consider in the next chapter, one or the other, but
not both, will be specufied at all points on the boundary. It is customary to specify
boundaries according to which one of the two conditions is specified: SΦ for
boundaries where Φ* is specified, and Sq for boundaries where q* is specified. Keep in
mind that it is not a question of which of the two quantities is known. There are some
problems (although it is not usually the case) where the analyst might know the value
of both quantities at given point. It is a question of which quantity is specified in the
mathe matical formulation of the particular problem under investigation. At this time,
however, it is neither nor desirable to make that distinction.

6.1.2 Weak From

To obtain the weak form, note that
        
  x k
V
x  + k y 
x  y  y 
+ Q( x, y)dV = 0

must be true for any δΦ. Integration of this equation by parts gives
              
V  x 
 k x + 
x  y 
k y  − 
y   x
k x
x
+
y
k y 
y 
+ Q ( x, y ) dV = 0

Application of Green’s theorem,
 P Q 
  x + y dV =  Pdy − Qdx 
V S

to the first two terms in Eq. 6.5 gives

         
 k
S
x
x
nx + k y n y dS −  
y  V
 x
kx
x
+
y
ky
y
− Q dV = 0.0

From the boundary condition expressed in Eq. 6.2, we can write
     
 
V x
kx
x
+
y
ky
y
− Q dV −  q*dS = 0.0
 S

or, in the more usual form

     
 
V x
kx
x
+
y
ky 
y 
dV =  QdV +  q*dS
V S

It is this form of the equation that we will use to obtain our finite element
approximations.
Before proceeding, we show that Eq. 6.8 can be used to obtain the corresponding
functional associated with Eqs. 6.1 and 6.2. To do so, we note that
1        
2
   
2

kx + ky =  k x   + k y   
x x y y 2   x   y  
 
and
 ( )Q =  (Q )
 ( )q* =  (q* )
Hence, we can write
  1    2 1    2  
   k x   + k y   − Q  dV −  q *dS  = 0
  2  x  2  y   
V S

which is the variational principle associated with Poisson’s equation. We will, on

occasion, use the symbol J for this functional and state Eq. 6.8 as
J = 0

6.2 Finite Element Approximation

We now consider the finite element approximation of the solution of our governing
equation. We begin by placing Eq.6.8 in the following matrix form:

  
    k x 0   x 
     dV − V QdV − S q dS = 0
*
 
V x y   0 k y  
 y 

or, in abbreviated notation,

 RdV −  QdV −  qdS = 0

V V S

where

  
 
 =  x 
 
 y 

Recall from Chapter 5 that within a single element,

( x, y) = N 

and that

    N I N J N K    
   x    
I

 =  x  =  Nx x

N J N K   J 
   I  
 y   y y y   K 

abbreviated as
 = N 
Also recall that
 = N  and  = N 
We now substitute the above FEM approximations into Eq. 6.14 and integrate. We
must do so, of course, element by element. We begin with the volume integral
associate with the stiffness term, where, for our generic element e, we have
 RdV = V N  RN dV
T
Ve e

=  K e 

In expanded notation the element stiffness matrix is
 N I N I 
 x y   N I N J N K 
 0   x
K e = V  N J N J  k x x x  dV
  N K 
e x y   0 k y   N I N J

 N N K   y y y 
 K 
 x y 
a 3×3 matrix. The other volume integral is

 QdV =   N T Q( x, y )dV =  Qe

Ve V e

where
 NI 
Qe = V  N J Q( x, y)dV
e
N 
 K
The total contribution from element e to the volume integration in Eq.6.8 is

K e − Qe

where the Φ terms are those associated with the given element.
We now evaluate the surface integral in Eq. 6.14. Figure 6.2 illustrates a finite
element approximation of Φ along a boundary. Consider a typical segment of S, Ss,
between nodal point m and nodal point n. Along m-n, Φ is a linear function of S. We
may therefore use the shape functions developed for our one-dimensional problems
and write

Figure 6.2. Segment of boundary.

  S  S   m 
 ( S ) = 1 −      = N S e
 lS   lS    n 
where lS is the length of the segment.
The surface integral along SS can now be written as

 q*dS =   N S q( S )dS

SS S S

which gives us

 q *dS =  QS

SS

where
 S 
1 − 
Qm 
QS =   = S  S q(S )dS
l
Qn  S
  S  
  lS  
The governing interpretation for Qm and Qn is that they equal the first moment of area
of q(S) about their respective nodal point, divided by the length of the segment. Hence,
we place a point Q at each node that gives the same moment as the distributed q. Also,
note that
Qm + Qn =  qdS
SS

Figure 6.3. Integration of surface integral.

We now have the contributions to Eq.6.8 from a typical volume element e and a
typical surface segment s. We write these as
J e = K e + Qe
and
J S = −QS
To obtain the total variation, we need to sum all such contributions; hence,
J = J e + J S
J = K e − Q
where [K] is the matrix sum of all the [K]e, and {Q} is the sum of all the {Q}e and
{Q}s. We now seek values of {Φ} for which δJ = 0 for all {δΦ}. This requires
K e  = Q
which is our final governing finite element equation.

6.3 Programming Preliminaries

Before we look at the program to create and solve Eq.6.33, several programming
details must be discussed. Other details will given with the listing of the code, and yet
others you will be able to obtain from the example data and functions included with
the test problem.

6.3.1 Element Identification

As in the codes for one-dimensional analyses, the core of our new code will be the
integration of the stiffness matrix, element by element. To accomplish this, we must
identify the location and geometry of each element. This is done by identifying the
three nodes associated with each element numbers are circled. The nodes associated
with each of these elements can be identified using the following two-dimensional
array:

Figure 6.4. A simple two-dimensional mesh.

The nodal point coordinates will be located in two arrays, XORD and YORD.
6.3.2 Integration of [K]e and {Q}e matrices
Because our shape functions N  are linear in x and y, the gradient terms in N 

are constants. If we also assume that the coefficients kx and ky, in [R] are constant, we
can write
K e = V N  RN dV
T

= N  R N  dV
T
Ve

= N  R N V
T

Note again the notation we have chosen to use for our region of integration. In the
above, V is actually the area of the element. The notation indicates that we are
assuming it is an area associated with a volume of unit thickness.
The derivatives of the shape functions will be stored in the arrays DNDX and DNDY,
where, for example;
N 2
DNDX(2) =
x
The element stiffness matrix, [K], will be stored in the 3×3 array S(J, K), defined in
terms of the DNDX and DNDY arrays as
S(J, K) = (DNDX(J) * RX * DNDX(K)
+ DNDY(J) * RY * DNDY(K)) * VOL
where
VOL = ((XJ * YK) - (XK * YJ))/2.0

The expression for {Q}e is

Qe = V N QdV
e

If we assume Q constant within each element, it can be shown that all three
components of {Q}e are equal and have the value
Q * VOL/3.0 
Qe = Q * VOL/3.0 
Q * VOL/3.0 
 

6.3.3 The Global Stiffness Matrix

Placement of an element S matrix into the global SK matrix is best explained by an
example. Let
 Se II Se IJ Se IK 
S e =  Se JI Se JJ Se JK 
Se KI Se KJ Se KK 
be a typical stiffness matrix. If, say, it is for element in Fig.6.4, the nodes {I, J, K}
correspond to the nodal numbers {6, 5, 2} as obtained from the NP array. Thus, the
nine terms of this matrix would appear in the global stiffness matrix as:

The matrix is symmetric and will also be banded; thus, appropriate storage will be
used.

6.3.4 User-Provided INCLUDE Codes

As our finite element codes become more complex, it is necessary to use more
user-written INCLUDE codes to define specific problems. In the code to follow,
poisson.m, there are two such INCLUDE codes: INITIAL.m and COEF.m.
INITIAL.m
we now consider two-dimensional problems, it becomes more difficult to define the
initial values of the variables this way. It is much more convenient to initialize
variables using an INCLUDE code. This will appear in poisson.m after the mesh data
has been loaded and after the general initialization has taken place. In INITIAL.m the
user must define NPBC, PHI, and Q for all nodes that differ from the values given in
the general initialization (all zeros). This will be accomplished through the use of a
nodal point code defined in the MESH data file exactly as the NPBC array was
defined in wire.m and ode.m. However, in poisson.m the NPcode array, as it will be
called, is not used. Only in INITIAL.m will it be used to identify which nodes should
have which boundary conditions. The NPcode values are assigned by the user and
entered with the mesh data. These values can then be used in the user’s own
INITIAL.m code to assign specific boundary conditions. An example of both a data
file and an INITIAL.m code will be given with the test problem.

COEF.m
The three parameters, kx(x, y), ky(x, y), and Q(x, y), are defined for each element with
the COEF.m INCLUDE. Available for the definitions will be the coordinates of the
centroid of the element, XC and YC, as well as the element number, the NP array, etc.
The parameters are assumed constant within the element for the integrations. This
provides a good approximation for the integration of the stiffness matrix where all
other terms in the integrand are constant. However, it is not as accurate a procedure
for integration of the right-hand side. Care must be taken to make elements small
enough that Q(x, y) is approximately constant within each element, or at least
approximately linear.

6.3.5 The MESHo, NODES, and NP Data files

Three files define the mesh used for the analysis. The first, MESHo, contains three
separate lines:
NUMNP = Number of nodal points
NUMEL = Number of elements
NNPE = Number of nodes per element
The number of nodes per element for poisson.m is 3. For codes to be considered later,
higher-order elements will be used for which the number of nodes per element will be
greater than 3.
The second file, NODES, contains one line per node. Each line gives for its particular
node.
XORD YORD NPcode
That is, the x and y coordinates of the node and the node identification number.
The third file, NP, contains the connectivity array for each element explained earlier.
An example of such an array was given in Section 6.3.1.
One of the problems encountered in using finite element codes for two-dimensional
analyses is how to create the mesh data. It is not unusual for there to be several
thousand nodes and elements. Obviously, the user will not want to create the
corresponding input files by hand. For this purpose there are mesh-generating codes
that create this data and require the user only to enter data related to boundary
locations and number of elements desired. One such program, mesh.m, is presented in
Appendix D. It creates all three of the data files just described. In addition, it provides
the user with an easy method for assigning NPcode values to all boundary nodes.

6.3.6 Bandwidth and Bandwidth Reduction

A problem encountered with two-dimensional analyses is the numbering of the nodes
to minimize the storage needed for the stiffness matrix. For our previous
one-dimensional problems, the bandwidth was either 2 (for symmetric matrices) or 3
(for nonsymmetric matrices). With two-dimensional problems, the needed bandwidth
will depend on the way the nodes have been numbered. Consider the two schemes
shown in Fig. 6.5. For the first, the largest numerical difference between any two
nodes associated with a single element is 3. This would require a bandwidth of 4 if the
matrix were symmetric. For the second numbering scheme, the largest numerical
difference is 8, which would require a bandwidth of 9. You might think that (1) the
difference is not large and (2) it is easy to see how to number the nodes to get the
smaller bandwidth. However, for complex meshes, (1) the difference between a good
numbering scheme and a poor numbering scheme can result in very large difference in
bandwidth requirements, and (2) determining how to number the nodes to minimize
the bandwidth can be difficult.

Figure 6.5. Two numbering schemes.

The optimal numbering scheme depends on the connectivity of the nodes, and this is
completely specified by the NP array. Hence, computer codes are written that use this
array as input to determine a new numbering scheme that will provide a narrow
bandwidth. One such code, newnum.m, is described and given in Appendix D. The
output from this code is the data file NWLD, which stands for new(old), that is, the
new node number in terms of the old number. This array of numbers can be used as
input to any finite element code to create a new numbering system that will produce a
stiffness matrix having a narrow bandwidth. Whether program newnum.m or another
code is used, NWLD must be in the working directory. The last entry in this code
must be the bandwidth that the new numbering creates (see the example problem).

6.4 Program poisson.m

Program poisson.m is the finite element code for solving Eq. 6.1, Poisson’s equation.
In addition to poisson.m, you will need the following files in your working directory:
MESHo is the data file that defines NUMNP, NUMEL, and NNPE. NODES is the
data file that defines, for each node, its XORD, YORD, and NPcode values. NP is the
connectivity array. NWLD gives the new nodal numbering for bandwidth reduction.
INITIAL.m is the user’s INCLUDE code for the initialization of variables to define
the specific problem being analyzed. COEF.m is the user’s INCLUDE code to define
the parameters (coefficients) kx, ky, and Q. sGAUSS.m is the supplied function for
solving the banded, symmetric equations by Gauss elimination.

6.4.1 Flow Chart

6.4.2 Code

6.4.3 Test Problem

For our test problem, we again use the fact that whenever our finite element
approximation can duplicate an exact solution, it will do so. The problem we have
selected is defined as follows
where Φ satisfies
       
+ =0
x  x  y  y 
5 5

The exact solution to this problem is

=9 − 0.75x
where x is measured from the left side of the rectangle. Clearly, our piecewise linear
approximation can duplicate this function; thus, poisson.m should produce the exact
solution.
The first step in obtaining our finite element solution is to select a mesh. We choose a
mesh with 16 elements and 15 evenly spaced nodes. The mesh and the corresponding
data are as follows

Here we have defined our NPcode values to indicate the side of the mesh on which
the node appears, and only on the sides where there is need to specify a boundary
condition that differs from the default values. We have chosen to number the sides
counterclockwise, starting with the lower boundary. Thus, nodes 11, 6, and 1 are
given NPcode numbers equal to 4, and nodes 15, 10, and 5 are given NPcode numbers
equal to 2. However, there is no need to indicate that nodes 2, 3, and 4 on side 1 or
that nodes 12, 13, and 14 are on side 3, because there is no need to change any of the
default values associated with these nodes. The important concept to understand is
that the NPcode values are for your use and only your use in your INITIAL.m code.
Use them there to identify the nodes for which you wish to assign boundary
conditions that differ from the default boundary conditions. See the following
example for the INITIAL.m code. Note, however, if you wish to assign boundary
values in INITIAL.m without reference to the NPcode values, that is your option; you
can simply give all nodes an NPcode value of zero and use, for example, the nodal
coordinates to determine what boundary conditions to assign.

The NWLD values have been assigned to make the numbering begin at the upper left
corner and increase download, similar to the example given earlier. Sketch the mesh
with the new numbers and verify the NWLD data as well as the IB value shown. Keep
in mind that only poisson.m uses the new numbers, not the user. In all data files, and
in INITIAL.m, the original node numbers should be used.

The user INCLUDE codes for the test problem are: INITIAL.m and COEF.m

In INITIAL.m, we search through all NPcode values to identify those equal to 2 (node
is on side 2) and equal to 4 ( node is on side 4). We again emphasize that these
numbers simply point out to the user wants them to point out. We could have used 3
and 9 to indicate the specified PHI values for these nodes and have accomplished the
very same purpose.
The test problem has conastant coeffcients; hence, our COEF.m code is fairly simple.
In problems with more complicated coefficients, for example, those that depend on
the x and y coordinates, functions can be written in this code. Keep in mind when you
write this code that you have use of all variables that have been calculated in
poisson.m up to the line where COEF appears. This includes the coordinates of the
element’s centroid, XC and YC, as well as the gradients of the shape functions and
other quantities.
At the conclusion of the analysis, the resulting Φ values are saved in your working
directory in a file named PHI. For the test problem, it will contain the exact answers,
which are
9.0 7.5 6.0 4.5 3.0
9.0 7.5 6.0 4.5 3.0
9.0 7.5 6.0 4.5 3.0
shwon in the order corresponding to the mesh nodes.
Three auxiliary codes can be used in conjunction with poisson.m to help prepare input
data and to help analyze the solution. They can be found in Appendix D and are as
follows
mesh.m Generate mesh files:
MESHo.txt, NODES.txt, NP.txt
newnum.m Generate data file NWLD.txt
topo.m Creates plots of PHI.txt values