Introducing the Finite Element Method in

electromagnetics to undergraduates using

MATLAB
M. K. Haldar
Department of Electrical and Computer Engineering, National University of Singapore,
Singapore
E-mail: [email protected]

Abstract This paper shows how electrical engineering undergraduates can acquire a working
knowledge of the finite element method (FEM) within a short period of time using MATLAB. For
simplicity, only first-order triangular elements are considered. The scalar wave equation for
homogeneous isotropic waveguides is used for introducing the FEM. Simple waveguide problems,
including a design problem, are discussed as examples. It is shown how the knowledge acquired can
be extended to other electromagnetic problems.

Keywords electromagnetic field computation; finite element method; first order triangular element;
MATLAB

The availability of immense and cheap computing power on desktop or laptop com-
puters has made the numerical solution of electromagnetic problems viable, even
for undergraduate students. Many educators have taken two approaches: use com-
mercially available software1 (which may be an expensive option), or design user
interfaces and simulation code2,3 based on existing programmable mathematical
packages. Neither of these constitutes the object of this paper. The aim of this paper
is similar to that reported in Ref. 4, which introduces the method of moments through
MATLAB.5 MATLAB has received worldwide acceptance in the teaching of many
engineering courses in, for example, signal processing and control engineering. It
will not be unreasonable for teachers in electromagnetics to expect students to have
a working knowledge of and access to MATLAB.
The Finite Element Method (FEM) is a well-established technique in electro-
magnetics and still generates considerable research. Several excellent texts and
monographs6–9 are available. They are suitable for researchers or for a complete
course on finite elements. Moreover, these texts often emphasise the rigorous
mathematical basis of various FEM formulations. Some also provide program
code.6,9 A practical approach to the derivation of the FEM and its implementation is
provided in Ref. 8. The object of this paper is to provide a brief hands-on under-
standing of FEM through MATLAB, as MATLAB provides powerful support for
matrix operations and dynamic memory allocation.
The material in this paper can be covered in a two-hour lecture. Students may be
given two weeks to absorb the material and carry out assignments similar to the
examples given in this paper.
It is noted that at least one commercial FEM software package, namely Femlab
3 from Comsol Inc.,10 has a version that runs as an add-on to MATLAB. It is not

International Journal of Electrical Engineering Education 43/3

FEM in electromagnetics using MATLAB 233

suitable for teaching FEM programming. However, as illustrated in Ref. 11, students
can use this software to test variational formulations derived by them as the soft-
ware accepts variational formulations as input. Thus, Ref. 11 and this paper are com-
plementary if Femlab 3 is available.

FEM formulation

Introduction
For initiating students to the FEM through MATLAB, the scalar wave equation for
a homogeneous isotropic medium is chosen. The equation is written as:
∇ 2f + k 2f = 0 (1)
2
where k is the eigenvalue.
The scalar wave equation has many applications. In the context of electromag-
netics, the scalar wave equation can be used to analyse problems such as the
propagation of plane waves, TE and TM modes in waveguides,12 weakly guiding
optical fibers,13 etc. Moreover, by setting k = 0 in the FEM formulation of eqn (1),
one can also illustrate the FEM solution of electrostatic problems.
The FEM solves eqn (1) by minimisation of a corresponding functional given by:
1   ∂f   ∂f  
2 2

F(f ) = ∫∫  +   − k 2f 2  ds (2)
2 s  ∂ x   ∂ y  
where S represents the cross-sectional area of the waveguide, k2 = erk20 − kz2,
k0 = 2p/l0, l0 is the free space wavelength, er is the relative permittivity of
the medium and the field variation along the direction of propagation, z is taken as
e−jkzz. Students can be referred to Ref. 7 for methods for obtaining FEM functionals.

The FEM approximation

For introducing the FEM approximation, the cross-sectional area is considered to
be constituted of small triangles. Each triangle is called an element. Hence we
can write:
1   ∂f e   ∂f e  
Ne 2 2

F(f ) = ∑ ∫∫  +   − k 2f e2  ds (3)
e =1 Ae 2  ∂ x   ∂ y  
where e represents the element (triangle) number, Ne represents the total number of
elements and Ae represents the area of the element e over which the functions are
integrated.
The function fe at a point P(x,y) inside the triangle may be approximated as
f e = a + bx + cy + dx 2 + fy 2 + gxy + . . . .
For simplicity, we consider only linear terms:
f e = a + bx + cy (4)

International Journal of Electrical Engineering Education 43/3

234 M. K. Haldar

φe1
(x 1, y 1)

1 y

x
Element e
P(x, y)
2
(x2, y 2) 3
φe2 (x 3, y 3)
φ e3

Fig. 1 A typical first order triangular element.

Triangles with fe expressed as linear functions are called first-order triangular ele-
ments. Fig. 1 shows a triangle where the values of coordinates and fe at the vertices
are shown.
Using the linear approximation in eqn (4), we can write
f e1 = a + bx1 + cy1 f e 2 = a + bx 2 + cy2 f e 3 = a + bx3 + cy3 .
Solution of these equations gives
3 3 3
a = ∑ aif ei b = ∑ bif ei c = ∑ cif ei
i =1 i =1 i =1

where, ai = (xi1yi2 − xi2yi1)/(2Ae), bi = (yi1 − yi2)/(2Ae), and ci = (xi2 − xi1)/(2Ae). i, i1

and i2 are a cyclic permutation of 1, 2 and 3.
Hence, we can write
3
f e = ∑ uif ei
i =1

where the terms ui = ai + bix + ciy are also known as barycentric or area coordinates.
It may be noted that
3 3
∂f e ∂f e
= ∑ bif ei and = ∑ cif ei .
∂x i =1 ∂y i =1

Using row vectors fe = [fe1 fe2 fe3], u = [u1 u2 u3], b = [b1 b2 b3] and c = [c1 c2 c3],
(3) can be written as
1 Ne (bf et ) bf et + (cf et ) 
t t

F(f ) = ∑ cf t − k 2(uf t )t uf t  ds

2 e=1 ∫∫
Ae  e e e

where the superscript t indicates the transpose.

Hence,
1 Ne
F(f ) = ∑ {fe Pefet − k 2fe Qefet }
2 e=1
(5)

International Journal of Electrical Engineering Education 43/3

FEM in electromagnetics using MATLAB 235

where,
Pe = Ae{b t b + c tc} (6)
and
2 1 1 
Ae 
Qe = 1 2 1 . (7)
12  
1 1 2 
To obtain the matrix Qe, one uses the result provided in Ref. 8:
l! m! n! 2! Ae
∫∫ u u u ds =
l m n
1 2 3
(l + m + n + 2)!
.
Ae

Matrices with subscript e, obtained for triangle number e, are called local
matrices.

Forming local matrices with MATLAB

The first step in the FEM is to generate a triangular mesh over the cross-section, S.
As mesh generation is not the object of this introductory paper, there are several
options:
(i) download free mesh generator programs from the internet;
(ii) use mesh generators of any available FEM software (e.g., civil engineering
FEM software); and
(iii) use pdetool available in MATLAB.
For simplicity, only one domain (region) with one boundary and first-order triangu-
lar elements will be considered. Most mesh generators produce text files as output
which can be edited to yield the following data:

Triangle node numbers

A file element.txt which has three columns containing three node numbers of
each triangle, with rows arranged to correspond to triangle numbers in ascending
order.

Coordinates of nodes
A file coord.txt which has two columns containing x coordinate in the first column
and y coordinate in the second, with rows arranged to correspond to node numbers
in ascending order.

Boundary node numbers

A file bn.txt with one column containing the boundary node numbers in ascending
order.
The function M-file triangle.m contains the following MATLAB program to
compute the local matrices Pe and Qe.

International Journal of Electrical Engineering Education 43/3

236 M. K. Haldar

function [pe,qe] = triangle (x,y)

ae=x(2)*y(3)-x(3)*y(2)+x(3)*y(1)x(1)*y(3)+. . .
x(1)*y(2-x(2)*y(1);
ae=abs(ae)/2;
b=[y(2)-y(3),y(3)-y(1),y(1)-y(2)];
c=[x(3)-x(2),x(1)-x(3),x(2)-x(1)];
b=b/(2*ae);
c=c/(2*ae);
pe=(b.’*b+c.’*c)*ae;
qe=[2,1,1;1,2,1;1,1,2];qe=qe*(ae/12);

Forming global matrices with MATLAB

If the total number of nodes is Nn, eqn (5) can be written as
1 Ne
F(f ) = ∑ { fPegf t − k 2f Qegf t }
2 e=1
(8)

where f = [f1 f2 . . . fNn], and P eg and Qeg are Nn × Nn square matrices, such that if p,
q and r are node numbers (1), (2) and (3) of triangle e, only nine matrix elements
of Peg and Qeg are nonzero: pp, pq, pr, qp, qq, qr, rp, rq and rr.
Equation (8) can be written as:
1
F(f ) =
2
[f P gf t − k 2f Q gf t ] (9)

Ne Ne
where P g = ∑ P eg and Q g = ∑ Qeg .
e =1 e =1
Pg and Qg are called global matrices. The global matrices are thus obtained
by adding the local matrices written as Nn × Nn square matrices. This operation is
easily carried out by the following MATLAB program (explained by comments
following %).

% load the data files

load element.txt;
load coord.txt;
load bn.txt;
% Find the total number of elements and nodes
ne=length(element(:,1));
nn=length(coord(:,1));
% Set up null global matrices
pg=zeros(nn,nn);
qg=zeros(nn,nn);
% Sum over all triangles

International Journal of Electrical Engineering Education 43/3

FEM in electromagnetics using MATLAB 237

for e=1:ne;
% Get the three node numbers of triangle number e
node=[element(e,:)];
% Get the coordinates of each node and form row vectors
x=[coord(node(1),1),coord(node(2),1),coord(node(3),
1)];
y=[coord(node(1),2),coord(node(2),2),coord(node(3),
2)];
% Calculate the local matrix for triangle number e
[pe,qe]=triangle(x,y);
% Add each element of the local matrix to
% the appropriate element of the global matrix
for k=1:3;for m=1:3;
pg(node(k),node(m))=pg(node(k),node(m))+pe(k,m);
qg(node(k),node(m))=qg(node(k),node(m))+qe(k,m);
% Close the three for loops
end;end;end;

Forming the FEM equations

The functional given by eqn (9) can be written as
1 Nn Nn g 1 Nn Nn
F(f ) = ∑ f i ∑ Pijf j − k 2∑ f i ∑ Q ijgf j
2 i =1 j =1 2 i =1 j =1
(10)

The finite element solution is obtained by minimising the functional in (10) with
respect to the nodal values, i.e., by setting
∂ F(f )
= 0 for k = 1, 2, . . . Nn .
∂f k
Thus
1 Nn g 1 Nn
∑
2 i =1
( Pik + Pkig )f i − k 2 ∑ (Q ikg + Q gki )fi = 0.
2 i =1
Since the matrices are symmetrical, this becomes
Nn Nn

∑ Pikgfi − k 2 ∑ Qikgfi = 0
i =1 i =1

or in matrix form,
P gf t − k 2 Q gf t = 0. (11)
Now one has to implement boundary conditions. Only perfectly conducting bound-
aries are considered here. The number of boundary nodes is denoted as Nb.

International Journal of Electrical Engineering Education 43/3

238 M. K. Haldar

Boundary conditions: TE modes

For TE modes, f represents the axial magnetic field, Hz, and the boundary condition
is the Neuman condition, df/dn = 0, where n is the normal to the perfectly con-
ducting boundary. In FEM, this is a natural boundary condition and need not be
imposed. Thus the values at the boundary are considered to be unknown and the
eigenvalue equation (11) is solved using the MATLAB statement ksquare =
eig(pg,qg).

Boundary conditions: TM modes

For TM modes, f represents the axial electric field, Ez, and the boundary condition
for the perfectly conducting boundary is, f = 0. This is the Dirichlet condition, which
must be strictly imposed in the FEM. It requires the following:
(i) Omitting differentiation with respect to the known boundary nodes. This can
be done by simply deleting the rows of Pg and Qg corresponding to the
boundary nodes.
(ii) Deleting the columns of Pg and Qg corresponding to the boundary nodes, as
f = 0 at these nodes.
The deletion of the row and column corresponding to a boundary node is easily
accomplished using Matlab’s dynamic memory allocation feature. The deletion
procedure is implemented as follows:

% Find the total number of boundary nodes,nb

nb=length(bn(:,1));
% Create an auxiliary column vector of
% boundary nodes
bnaux=bn;
for n=1:nb;
%Read the node number k of the boundary node
k=bnaux(n);
% Delete kth row and column of the global
% matrices
pg(k,:)=[];
qg(k,:)=[];
pg(:,k)=[];
qg(:,k)=[];
% After deletion, the row/column numbers
% corresponding to boundary node numbers>k are
% reduced by one. As boundary node numbers are
% in ascending order, all node numbers can be
% reduced with a unit column vector of length nb
bnaux=bnaux – ones(nb,1);
end;

International Journal of Electrical Engineering Education 43/3

FEM in electromagnetics using MATLAB 239

As before, the eigenvalue equation is solved, but this time with the reduced
matrices obtained with the above procedure.

Homogeneous hollow waveguide examples

There are many homogeneous hollow waveguide problems in the available litera-
ture which can be given as assignments to students. Four examples are given here.
In these examples, pdetools was used to get first-order triangular mesh data. All the
examples can be worked out with notebook PCs.

Homogeneous rectangular waveguide

The waveguide (WR-90) dimensions are 2.286 × 1.016 cm. Tables 1 and 2 compare
the FEM and analytical values of k2 for TEmn and TMmn modes respectively. m rep-
resents the number of half wavelengths along the longer side and n represents the
number of half wavelengths along the shorter side. FEM values are obtained with
2688 triangles and 1405 nodes. It may be seen that the error increases for higher
values of k2, i.e., for higher-order modes. This is because for a given triangle size,
the linear approximation inside a triangle can better represent f for lower-order
modes which vary less rapidly in the waveguide cross-section.

Homogeneous square waveguide

This example is used to show degenerate (equal) eigenvalues. It also allows one to
progress towards a physical understanding of why FEM produces degenerate eigen-
values for circular waveguides. The waveguide dimensions are 1.5 × 1.5 cm. For the
sake of brevity, a comparison of FEM and analytical results is shown only for TE
modes in Table 3. FEM values are obtained for 4960 triangles and 2565 nodes. It
may be seen that modes such as TE12 and TE21 have the same eigenvalue. It is noted
that rotating the field profile of the TE12 mode by 90° gives the field profile of the
TE21 mode.

TABLE 1 Values of k2 (in cm−2) for TEmn modes (analytical values in brackets)
n
m 0 1 2

0 No mode 9.574(9.561) 38.459(38.245)

1 1.889(1.889) 11.468(11.450) 40.366(40.133)
2 7.563(7.555) 17.157(17.116) 46.104(45.799)

TABLE 2 Values of k2 (in cm−2) for TMmn modes (analytical values in brackets)
n
m 1 2 3

1 11.468(11.450) 40.362(40.133) 89.024(87.939)

2 17.158(17.116) 46.103(45.799) 94.846(93.605)
3 26.662(26.559) 55.695(55.242) 103.655(103.048)

International Journal of Electrical Engineering Education 43/3

240 M. K. Haldar

TABLE 3 Values of k2 (in cm−2) for TEmn modes (analytical values in brackets)
n
m 0 1 2

0 No mode 4.388(4.386) 17.567(17.546)

1 4.388(4.386) 8.778(8.773) 21.967(21.932)
2 17.569(17.546) 21.967(21.932) 35.184(35.092)

TABLE 4 Values of k2 (in cm−2) for TEmn modes (analytical values in brackets)
n
m 1 2 3

0 6.540(6.526) 22.013(21.877) 46.581(46.005)

1 1.507(1.506) 12.681(12.631) 32.684(32.384)
2 4.151(4.146) 20.102(19.987) 44.289(44.178)

Homogeneous circular waveguide

The radius of the circular waveguide is taken as 1.5 cm. FEM and analytical values
of k2 for TEmn modes are given in Table 4. m represents the degree of rotational sym-
metry of the fields and n represents the nth root of the characteristic equation for k.
FEM values are obtained for 4160 triangles and 2145 nodes. For m ≠ 0, the FEM
produces two equal eigenvalues. As in the square waveguide, the field patterns cor-
responding to these equal eigenvalues are rotated by 90° with respect to one another.
For m = 0, rotation by 90° produces identical patterns. This result is not obvious
from analytical methods.

Design example: single ridge waveguide

Design curves for single ridge waveguides were obtained analytically by Hopfer.14
Figs 2 and 3 compare the FEM solutions, represented by dots, with some of the
design curves given in Figs 5 and 7 of Hopfer’s paper respectively. Good agree-
ment is obtained. Moreover, Fig. 3 illustrates the power of the FEM. In the range,
0.008 < s / a < 0.50, Hopfer was unable to obtain analytical solutions using a single
rectangular waveguide TE mode as the TE30 mode cannot exist by itself but couples
to the TE01 mode. The FEM analysis has no such limitation.
It must be mentioned that for all the waveguide examples considered, a very small
eigenvalue for TE modes was produced by FEM. This value approximates the zero
eigenvalue corresponding to the f = constant mode15 and is ignored.

Extension to other problems

In this section, it is shown that the discussion in Section II can with minor exten-
sions be used to solve many other problems. For the sake of brevity, we consider
only two of them.

International Journal of Electrical Engineering Education 43/3

FEM in electromagnetics using MATLAB 241

Fig. 2 Single ridge waveguide TE10 mode cut-off wavelength a = 2.286 cm and b/a = 0.45.
Dots represent values calculated by FEM.

Fig. 3 Single ridge waveguide TE30 mode cut-off wavelength a = 2.286 cm and b/a = 0.45.
Dots represent values calculated by FEM.

International Journal of Electrical Engineering Education 43/3

242 M. K. Haldar

Electrostatic problems
The functional for Laplace’s equation is:
1  ∂f  2  ∂f  2 
F(f ) =
2 ∫∫
e 0 r
e +   ds
S  ∂ x   ∂ y  
where f represents the electric potential. e0 is the permittivity of free space and er
is the dielectric constant.
For homogeneous isotropic media, it is easy to see that the FEM equation is
Pgf t = 0, the matrix P g being the same as that in Section II.D.
Two boundaries are considered here. The mesh data should have two sets of
boundary nodes. We can take f = 0 at one boundary. For nodes on this boundary,
the corresponding rows and columns of P g are deleted. For the other boundary,
f = V, a numerical value. The rows of P g corresponding to nodes on this boundary
are deleted as before. However the column corresponding to each node is multiplied
by −V and put as a separate matrix. The matrices are added together to form the
matrix R. Next the columns are deleted from P g to get the reduced matrix P.
The FEM equation is now of the form Pf t = R which can be solved for f t using
Matlab.

Wave propagation in inhomogeneous waveguides

For homogeneous waveguides, the refractive index, n, is constant and so k is a
constant. We could find the eigenvalue k2 and then compute the propagation
constant, kz, from k 2z = erk20 − k2 and k0 = 2p/l0 for a given wavelength, l0. For an
inhomogeneous waveguide, eqn (8) is written as:
1 Ne
F(f ) = ∑ {fPeg f t − e re k02fQeg f t + kz2fQeg f t } , where, ere is the relative permit-
2 e=1
tivity of the medium inside element number e.
Hence eqn (9) becomes:
1 1
F(f ) = f R gf t − kz2f Q gf t (12)
2 2
where
 g  2p  2
Ne 
R = ∑  P e −   e re Q eg 
g

e =1 
 l 0  
and
Ne
Q g = − ∑ Q eg .
e =1

Equation (11) now becomes

R gf t − kz2 Q gf t = 0. (13)
The only modification is to generate the mesh with elements of different relative
permittivity, ere, and find Rg for a given wavelength, l0. Unlike for a homogeneous

International Journal of Electrical Engineering Education 43/3

FEM in electromagnetics using MATLAB 243

waveguide we need to determine the eigenvalue, kz2, numerically for each

wavelength.

Conclusion
It is shown how the FEM can be introduced to electrical engineering undergradu-
ates using only a few lines of MATLAB code. The MATLAB codes relate directly
to the FEM operations. Four waveguide examples are given. These examples can be
run on a standard PC. A design example illustrated the advantage of FEM compared
to an analytical method. It was also shown how the concepts acquired can be readily
extended to other problems. The demonstrated problems are 2D in nature. However,
similar programming techniques can be employed for more complex problems, such
as 3D problems presented, for example, by Silvester and Ferrari.6 Institutions with
inadequate financial resources may benefit from the proposed methodology, as only
MATLAB is needed, no add-ons (not even a mesh generator). The material presented
in this paper may also be useful to beginning research students who are uninitiated
in the FEM.

Acknowledgement
The author thanks Mr Duen Hong Ong, an undergraduate, for the results given in
Section III.

References
1 J. Lu and D. V. Thiel, ‘Computational and visual electromagnetics using an integrated program-
ming language for undergraduate engineering students’, IEEE Trans. Magnetics, 36 (2000),
1000–1003.
2 S. Selleri, ‘A Matlab experimental framework for electromagnetic education’, IEEE Antennas and
Propagat. Mag., 45 (2003), 85–90.
3 S. Selleri, ‘A Matlab application programmer interface for educational electromagnetics’ in
Antennas and Propagat. Soc. Symp, 22–27 June, 2003.
4 S. Makarov, ‘MoM antenna simulations with Matlab RWG basis functions’, IEEE Antennas and
Propagat. Mag., 43 (2001), 100–107.
5 MATLAB, The Math Works Inc., http://www.mathworks.com
6 P. P. Silvester and R. L. Ferrari, Finite Elements for Electrical Engineers (Cambridge University
Press, Cambridge, 1990).
7 J. Jin, The Finite Element Method in Electromagnetics (John Wiley and Sons, New York, 1993).
8 S. R. H. Hoole, Computer-Aided Design of Electromagnetic Devices (Elsevier, New York,
1989).
9 T. Itoh, G. Pelosi and P. P. Silvester, Finite Element Software for Microwave Engineering (John
Wiley and Sons, New York, 1996).
10 K. Foster, ‘Retaking the field – an old computational favourite is overhauled’, IEEE Spectrum, July
(2004), 54–55.
11 L. Y. Tio, A. A. P. Gibson, B. M. Dillon and L. E. Davis, ‘Weak form finite element formulation for
Helmholtz equation’, Int. J. Elect. Enging. Educ., 41 (2004), 1–9.
12 S. Ramo, J. R. Whinnery and T. Van Duzer, Fields and Waves in Communication Electronics (John
Wiley, New York, 1994).

International Journal of Electrical Engineering Education 43/3

244 M. K. Haldar

13 A. K. Ghatak and K. Thyagarajan, Optical Electronics (Cambridge University Press, Cambridge,

1989).
14 S. Hopfer, ‘Design of ridged waveguides’, IRE Trans. Microw. Theory. Tech., 3 (1955), 20–29.
15 P. Silvester, ‘A general high-order finite element waveguide analysis program’, IEEE Trans. Microw.
Theory Tech., 17 (1969), 204–210.

International Journal of Electrical Engineering Education 43/3