Finite Element Method(FEM) for Two Dimensional Laplace

Equation with Dirichlet Boundary Conditions

April 9, 2007

1 Variational Formulation of the Laplace Equation

The problem is to solve the Laplace equation
r2u = 0 (1)
in domain
subject to Dirichlet boundary conditions on @
. We know from our study of the
uniqueness of the solution of the Laplace equation that nding the solution is equivalent to nding
u that minimizes
W=
1 Z jjrujj2d (2)
2

subject to the same boundary conditions. Here the dierential d denotes the volume dierential
and stands for dxdy for a plane region. W has interpretations such as stored energy or dissipated
power in various problems.

2 Meshing
First we approximate the boundary of
by polygons. Then
can be divided into small triangles
called triangular elements. There is a great deal of exibility in this division process. The term
meshing is used for this division. For the resulting FEM matrices to be well-conditioned it is
important that the triangles produced by meshing should not have angles which are too small.

At the end of the meshing process the following quantities are created.
 Nv : number of vertices or nodes.
 Nv  2 array of real numbers holding the x and y coordinates of the vertices.
 Ne: number of triangular elements.

1
  Ne  3 array of integers holding the vertices of the triangular elements.
 Nvf : Number of vertices on which the u values are not specied or free.
 Nvf  1 array of integers holding free vertex indices.
 Nvp : Number of vertices on which the u values are specied or prescribed.
 Nvp  1 array of integers holding prescribed vertex indices.
 Nvp  1 array of real numbers holding prescribed u values.
Data structures holding adjacency information for the vertices, edges, and the triangular elements
are also generated by sophisticated FEM meshing subroutines.

3 Planar Approximation over a Triangle


u = a + bx + cy 
@R@  U

 U
U : (x ; y )
 : (x ; y ) H 
HH 
HHH : (x ; y )
 

Let us consider a triangular element whose node numbers are ,  , and . The node coordinates
are (x; y), (x ; y ), (x ; y ). On this triangle u(x; y) is assumed to have a planar variation:
2 3
h i6 a
7
u(x; y) = a + bx + cy = 1 x y 4 b 5 (3)
c
First a, b, c are to be determined in terms of the values of u(x; y) at the nodes U, U , U , using
the three equations
u(xj ; yj ) = Uj ; j 2 f; ; g (4)
In matrix form the equations are
2 32 3 2 3
1 x y a U
6 76 7 6 7
4 1 x y  5 4 b 5 = 4 U  5 (5)
1 x y c U
So
2 3 2 3 12 3
h i6 a 7 h i 6 1 x y 7 6 U 7
u(x; y) = 1 x y 4 b 5 = 1 x y 4 1 x y 5 4 U 5
c 1 x y U
2 32 3
h i 1 6 x y x y x y x y x y x y 7 6 U 7
= 1 x y 2A 4 y  y y y y y 5 4 U 5
x x x x x x U
= U (x; y) + U  (x; y) + U (x; y) (6)
where
2A = x y x y + x y xy + xy x y (7)
and
= 21A [x y x y + (y y )x + (x x )y]
 (x; y ) (8)
etc. 2A is twice the area of the triangle  . The functions j are interpolatory in nature.
j (xk ; yk ) = jk ; j; k 2 f; ; g (9)

  : (x ; y )
 : (x ; y )  
HH 
HHH PAA 
HHH AA 
HHHA 
HA  : (x ; y )

The functions are also known as natural coordinates, simplex coordinates or areal coordinates.
Let P be the point (x; y). Then
Area of triangle P 
 (x; y ) = (10)
Area of triangle 
Exercises:
 Show that (x; y) +  (x; y) + (x; y) = 1.
 At which point is 2 (x; y) + 2 (x; y) + 2 (x; y) minimum? What is the minimum value?
Since the original domain is now approximated as a union of small triangular elements, the total
W corresponding to stored energy or power dissipation can be expressed as a sum of element W 's.
For the  element Z
(e) 1 2
W =
2 ( ) jjrujj d (11)
Now
ru = Ur  + U r  + U r (12)
It should be noted that due to the planar variation of u(x; y)
r (x; y) = 21A [(y y )^x + (x x )^y] (13)
etc. are constant over the entire element. So
Z
W =(e) 1 jjr ujj2 d =
1 X X U S (e)U (14)
2 ( ) 2 j=;; k=;; j jk k
where Z
(e) =
Sjk (r j )
(r k )d; j; k 2 f; ; g (15)
( )
Let 2 3
U
U (e) = 64 U 7
5 (16)
U
Then
1 T
W (e) = U (e) S (e) U (e) (17)
2
S (e) is called the element Dirichlet matrix. Show that,
(e) = 1 [(y y )(y y ) + (x x )(x x )]
S (18)
2A    

etc. The Dirichlet matrix is symmetric, positive semi-denite. The expression for element energy is
a quadratic form in the nodal values U, U , U . Since W is expressed as a sum of all the element
energies, it follows that W is given by a positive semi-denite quadratic form in the nodal values
U1 , U2 , . . . UN .
1
W = U T SU (19)
2
where 2 3
U1
6 U2 7
6 7
U = 66 ... 7
7 (20)
4 5
UN
and the Dirichlet matrix S has contributions from all element S (e)'s. In the program developed
(e) contribution is added to
here, S is initialized to zeros and as one loops over all elements any S
S . This is in contrast to other FEM programs which actually generate 3  3 element matrices
and then combine all of them during element assembly.

4 Minimization of W
P of the N U values are prescribed. N P are free to vary. For simplicity of presentation we assume
that the prescribed values come after the free values. (MATLAB and modern matrix libraries allow
sub-matrices to be selected based on arbitrary index sets. So there is no need to actually number the
prescribed nodes after the free nodes.) Then the matrices U and S may be partitioned as follows:
" #
U= Uf
Up (21)
" #
S = SSff SSfp (22)
pf pp
Note that due to the symmetry of S , Sfp = SpfT . In terms of the partitioned matrices W may be
expressed as
W =
1 U T SU (23)
2 " #" #
1 h
= 2 Uf Up T T
i S
ff S fp U f
(24)
Spf Spp Up
= 12 UfT Sff Uf + 21 UpT Spf Uf + 12 UfT SfpUp + 12 UpT SppUp (25)
But since UpT Spf Uf is a scalar it equals its own transpose.
T U = UT S U
UpT Spf Uf = (UpT Spf Uf )T = UfT Spf p f fp p (26)
The last equality is due to the fact that Sfp = SpfT . So by combining the two equal middle terms of
(25) we get
1 1
W = UfT Sff Uf + UfT Sfp Up + UpT Spp Up (27)
2 2
Now this W is to be minimized with respect to Uf . At the minimum the gradient of W with respect
to Uf equals zero.
rUf W = Sff Uf + SfpUp = 0 (28)
So the solution is
Uf = Sff1 Sfp Up (29)
Once Uf is known, U is known, and one computes W = 21 U T SU . In physical problems, quantities
related to W are usually important.
In eigenvalue problems
R
like the HelmholtzR equation, and loading problems likeR the Poisson
equation, not only 2
jjrujj2d , butR also 12
u2d is of importance. Just as 12
jjrujj2d is
1
expressed as 12 U T SU , the quantity 12
u2d can be expressed as 12 U T T U , where the T matrix is
called the metric matrix. Like the Dirichlet matrix S , T can be constructed from individual element
contributions.
1Z u2 d =
1 X X U T (e)U (30)
2
( ) 2 j=;; k=;; j jk k
where Z
(e)
Tjk = j (x; y ) k (x; y )d; j; k 2 f; ; g (31)

( )
First we evaluate Tjj(e) =
( ) j2(x; y)d . We consider a dierential strip parallel to the side
R
opposite node j in the triangle  If j changes by d j on the strip, its width is hd j and the
length of the strip is (1 j )b, where b is the length of the side opposite to node j and h is the
height of the perpendicular from node j to the opposite side. So
Z Z
Tjj(e) = 2 d = 1 2 (1 j )bhd j =
bh A

( ) j
0 j 12 = 6 (32)
(e), T (e),
where A is the area of triangle  . Thus the diagonal terms of the element T matrix, T 
(e)
and T , are equal to A=6. What about the o-diagonal terms?
Z Z Z
(e) + T (e) + T (e) = 2+
T    +  d =  (  +  + )d =  d (33)

( )
( )
( )
since +  + = 1. So
Z Z 1
(e) (e) (e)
T + T + T = = = bh A
 d  (1  )bhd  = (34)


( ) 0 6 3
(e) = A=6. So
But T
(e) + T (e) =
T
A A
= A
(35)
3 6 6
Similarly we also have
(e) + T (e) =
T
A
(36)
 6
 (e) + T (e) =
T 
A
(37)
 6
But T  (e) = T (e) etc. by the symmetry of the element T matrix. So adding the above three
equations and dividing by 2 we see that
(e) + T (e) + T 
T (e) = A (38)
 4
It it then seen that each o-diagonal element of the element T matrix is equal to A=12.

5 Example Code
The example code available on the web site solves two problems.
 Solution of Laplace equation on a rectangle: On three sides of the rectangle u = 0, while on
the other side u = 1.
 Capacitance of a cable with elliptic cross sections for both the conductors.
Please download the code and run these examples.