Anda di halaman 1dari 41

Home

Search

Collections

Journals

About

Contact us

My IOPscience

Application of the finite-element method to electromagnetic and electrical topics

This content has been downloaded from IOPscience. Please scroll down to see the full text.
1995 Rep. Prog. Phys. 58 1673
(http://iopscience.iop.org/0034-4885/58/12/002)
View the table of contents for this issue, or go to the journal homepage for more

Download details:
IP Address: 143.107.252.156
This content was downloaded on 03/12/2014 at 18:03

Please note that terms and conditions apply.

Rep. hog. Phys. 58 (1995) 167>1712. Printed in the UK

Application of the finite-element method to electromagnetic


and electrical topics
J P Webbt
Department of Electrical Engineering,Mffiill University, 3480 Universily Street. Montreal, PQ, H3A 2.47 Canada

Abstract
The application of the finite element method in electromagnetics is reviewed. The emphasis
is on formulations of 3D electromagneticproblems that are suitable for finite-element analysis
and difficulties that have been encountered, such as spurious modes and gauging of vector
potentials. Following an introduction to scalar and vector finite elements and a discussion of
techniques for handling unbounded domains, the three main areas of application are covered:
magnetic fields, electric fields, and electromagnetic waves. In the analysis of magnetic fields,
the various possible potential formulations are explained and compared; magnetic material
models are reviewed; and force calculation techniques are summarized. For electric fields
complications arise when the relationship between electric field and charge movement is
not simple: several different cases are considered. Electromagnetic wave problems are of
two types: eigenvalue (resonant cavities and waveguides) and deterministic (radiation and
scattering). The principle difficulty here is the occurence of spurious (non-physical) modes,
which can affect both types of problem. The reasons for spurious modes and the~available
remedies are outlined.
This review was received in September 1995

t E-mail address: 1W016.3521~compuserve.com


00344885/95/121673+40$59.50

0 1995 IOP Publishing Lld

1673

1674

J P Webb

Contents

1. Introduction
2. Basics
2. I , Finite-element solution of an electrostatic problem
2.2. Vector elements
2.3. Solving the matix equation
2.4. Unbounded problems
3. Magnetic fields
3.1. Potentials
3.2. Connecting to circuits
3.3. Time and frequency-domain solution
3.4. Magnetic material models
35. Forces and torques
4. Electric fields
4.1. Conductivity constant: Ohmic materials
4.2. Conductivity dependent on electric field strength
4.3. Conductivity dependent on charge density
4.4. Conductivity not defined
5 . Waves
5.1. Direct solution and solution via potentials
5.2. Eigenvalue problems
5.3. Deterministic problems
6. Conclusion
Acknowledgment
References

Page
1675
i675
1676
1679
1680
1681
1685
1686
1693
1694
1695
1698
1699
1700
1701
1701
1701
1702
1702
1704
1707
1710
1710
1710

Application of the finite-element method

1675

1. Introduction
The first paper on the application of finite elements to electrical problems appeared in 1968
(S Ahmed). It was soon realized that the finite element method offered the possibility of
a truly general tool for electromagnetic analysis, capable of handling arbitrary shapes and
realistic materials. However, it also became apparent that there were significant obstacles
to its successful application. Over the past quarter of a century, many of these obstacles
have been overcome. We now know how to use the finite element method to calculate the
electromagnetic field reliably and efficiently, under most circumstances. To be sure, there
are still challenging areas, and there is still debate over which are the most reliable and
most efficient ways of getting the job done. But, by and large, good techniques have been
established, and in many cases implemented in commercially-available software packages.
The aim of this paper is to explain the problems that have been encountered and the
methods that overcome them, without going into excessive mathematical detail. The focus
is purely on the electromagnetic: difficulties such as mesh generation and visualization,
that are common to most of finite element analysis, are omitted. So too are the secondary
techniques such as adaptive mesh refinement and optimization, that cannot be contemplated
until there are robust methods of field analysis. The emphasis is mainly on three-dimensional
problems and solutions, because most of the difficulties only emerge in three dimensions.
The organization of the paper is as follows. On the assumption that the reader is not at all
familiar with the finite element method, but at least a little familiar with electromagnetics,
the next section, Basics, introduces the main concepts of the finite element method by
considering the solution of a simple electrostatic problem. It goes on to describe a number
of other important ideas that are common to many areas of finite element analysis. There
follow three sections that cover the main categories of electromagnetics: Magnetic fields,
Electric fields and Waves. To a large extent these three categories are of interest to
distinct groups of researchers working with quite different devices. For this reason, each
section is deliberately constructed to be independent of the others (though not independent
of the Basics section).
2. Basics

There are a number of basic ideas in finite element analysis that are shared by all branches
of electromagnetics and they are described briefly here. More details may be found in the
several books on the subject: (Steele 1987, Hoole 1989, Silvester and Femari 1990, Jin
1993).
In,section 2.1, the finite element method is introduced by means of a simple electrostatic
problem. Though the electrostatic problem is scalar in nature, many electromagnetic
problems involve the calculation of a vector field, and in section 2.2 there is a discussion of
vector elements. The finite element method inevitably produces a system of algebraic
equations-a matrix equation-whose solution consumes much of the computational
resources. Solution techniques commonly used in electromagnetics are discussed in
section 2.3. Finally, though finite elements are well suited to the analysis of problems
with finite boundaries, electric and magnetic fields often extend over all space; section 2.4
contains a survey of techniques for these unbounded problems.

1616

J P Webb

2.1. Finite-element solution of an electrostatic problem


Consider a simple electrostatic problem: a metal sphere is held at a potential of 1 V and is
enclosed in a cubical grounded metal box (figure 1). Find the potential distribution in the
volume V between the sphere and the box.

Metal Box

Figure 1. A simple electmsratic problem.

The equation governing the electrostatic potential, CJ, is the Laplace equation:

v% = 0

in V.

Another way of stating this is that, for all weighting functions


lwV2@dV=0.

(1)
IO

(2)

This is called the weighted-residunl formulation of the problem. What is required is the
potential @ that solves ( 2 ) and takes the value 1 on the surface of the sphere and the value 0
on the inner surface of the box. (We can also constrain w to zero on both surfaces, because
this has no effect on the equivalence of (1) and (Z).)
Equation ( 2 ) could be used as the stating point of a finite element analysis. However,
it still contains second-order derivatives, and it is convenient to reduce these to first-order
derivatives. This is accomplished by transferring a V from CJ to w. using a vector identity.
The additional term thereby produced gives a surface integral when the divergence theorem
is applied to it, and the surface integral vanishes because w is zero on the surfaces. The
weighted-residual formulation becomes
Vu,. VQdV = 0 .

(3)

To solve (3) numerically, the volume V is divided into a number of finite elements. For
example, tetrahedral elements might be used (figure 2). Of course, having flat faces these
elements cannot exactly represent the curved surface on the sphere, but by using enough of
them a sufficiently accurate polyhedral approximation is possible. Inside each tetrahedron
the unknown potential is approximated by a polynomial which is first-order in the space
coordinates ( x , y , z ) . The polynomial can be so written that the four unknown coefficients

Application of the finite-element method

1611

Figure 2. A tetrahedral finite element

are the values of the potential at the four nodes (vertices) of the tetrahedron,
and (Pd:
(P

= (Plffl

(P201Z

e3013

(P4014.

( P I , (P2, (P3

(4)

The quantities 011, 012, 013 and q are polynomials whose dependence on x , y , z and on the
size and shape of the tetrahedron is known explicitly. Polynomial ai takes the value 1 at
node i, vanishes at the other three nodes, and varies linearly in between.
Equation (4) provides a complete and continuous representation for the potential within
the element, while at the same time using physically meaningful quantities (the values at
the four nodes) as unknown parameters. Furthermore, it permits the continuity of potential
to be imposed in a simple way between elements. It is only necessary for two elements
with a common edge or face to agree in their values of (P at their common nodes for (P
to be continuous everywhere over the edge or face. In practice this is done by assigning a
global numbering, 1, . . ., N to all the nodes of the finite elements mesh, such that a node
has a single global number even if it is shared by several tetrahedra. The unknowns in the
problem are then the globally numbered values of potential at the nodes 4;. i = 1, . .., N.
In fact, one may think of the polynomials ai in the same global way: ai is a function that
is 1 at global node i, vanishes at the other N - 1 nodes and is linear in each tetrahedron.
The piecewise linear potential in V is then

The CY( are known as basis or trial functions.


In fact not all N of the Qiare unknown. For a node i that lies on a boundary where the
potential is constrained to be 0 or I ,
is set to that value from the outset. For simplicity
it will be assumed that the first N, nodes are interior, free, nodes and the remainder are
boundary nodes. The problem therefore has only N, unknowns. These are the degrees of
freedom of the potential function.
Since there are a finite number of degrees of freedom with which to represent the
solution, it is not generally possible to satisfy (3) exactly for all possible weighting functions.

1678

J P Webb

In fact, it is desirable to have no more than N , independent weighting functions, so that (3)
becomes N, equations in N, unknowns. Though it is possible to choose any set of
weighting functions, it is convenient to use the basis functions themselves (Galerkin's
method). Substituting (5) into (3) with this set of weighting functions leads to the matrix
equation
where (@I is an N, vector of the unknown potentials; ( b ) is a known N, vector that depends
on the boundary constraint; and [SIis a square symmetric matrix known as the global or
sriffness matrix. Entry i, j is given by
Stj

Vai , Vaj dV

(7)

In general, there will be rather a large number of nodes and so [SI will be large. Even a
modest problem may have 10' nodes. On the other hand, [SI is very sparse. The function
oli is non-zero only in those tetrahedra that share node i . If i and j are not nodes of the
same tetrahedron, it is evident from (7) that Slj will be zero. Typically a row of [SI may
have only 20-30 non-zero entries out of a total of many thousands. It is very important to
exploit this sparsity, both in constructing [SIand in solving (6). For example, the assembly
of [SI is done tetrahedron by tetrahedron, not row by row. For each tetrahedron, a 4 x 4
local matrix is calculated from its four basis functions. The 16 entries of this dense local
matrix are added to the global matrix in the right places, taking into account the global
numbers of the four degrees of freedom,
An alternative way of arriving at (6) is to start with what is known as afrcncrional. a
real-valued function of the function O ( x , y. z):

F=

J,(VQ)*dV.

(8)

This particular functional is related l o energy; when 0 is the correct potential, EOFis twice
the electrostatic stored energy of the system. In fact, it can be shown that the correct potential
is exactly the function that minimizes F , while at the same time satisfying the boundary
conditions on the conductors. This variationalprinciple can be used to find an approximate
solution for O.The potential is approximated by basis functions in each tetrahedron, exactly
as above, and F is expressed in terms of the nodal values, Q?i. The best set of nodal values
is the set that minimizes F . The minimum is found by setting aF/aOi = 0 for each free
node. Since F is quadratic in @[, this leads to a set of simulataneous equations that are
linear in Qi,i.e. (6). In general, variational principles lead to the same matrix equation as
Galerkin's method. However, in some cases working from a variational principle establishes
useful relationships between the computed solution and physical quantities such as energy,
impedance or resonant frequency.
Once the solution (O}is fomd, the potential at any point can be determined from (5).
If the electric field is needed, it can be calculated by taking the gradient of (5).
The above procedures can be extended in a number of ways. First, higher-order finite
elements are possible. Instead of representing O as a first-order function of x , y and z,
second- or higher-order polynomials may be used. This usually gives greater accuracy for
a given number of unknowns, because the approximation is smoother and more realistic.
There is some additional cost though. The matrix [SIbecomes less sparse as the polynomial
order is raised, leading to longer solution times. In addition, the computation of the local
matrix is more complicated and lengthy, though it is considerably helped by the use of
precomputed (or universal) matrices, first proposed by Silvester (1969).

Application of the finite-element method

1619

A second extension is to the boundary conditions. More general boundary conditions


are
4 = 40
on Sd
(9)
a
4
_ -- P ( 4 ) on S,
an
where S, and S, are two surfaces that together make up the entire boundary of V. The
derivative a/an is the directional derivative out of V along a line perpendicular to the
boundary. The function P is a specified function of the potential on S,. For example,
P might he simply zero; such a boundary condition arises on planes of symmetry. Other
examples of P will be introduced later. To solve the Laplace equation subject to these more
general boundary conditions, an extra term must be added to the weighted-residual equation
(2). The extra term enforces the boundary condition a@/an = P ( 4 ) in a weighted sense
(the condition 4 = 40is enforced explicitly by constraining boundary nodes). The result,
after applying the divergence theorem and choosing w appropriately to eliminate certain
terms, is

Finally, this whole procedure can be applied to any of the other partial differential
equations that arise in electromagnetics. The details will differ hut the general features
described above are common to all.
2.2. Vector elements
In many electromagnetic problems it is impossible to reduce the unknown quantity to a
scalar potential such as 4. Instead, a two- or three-component vector field must be found.
The question therefore arises, how should a vector field be represented with finite elements?
The earliest answer to this question was, component-wise. Each Cartesian component of
the vector was treated as a scalar, and the nodal values of the component interpolated over
elements in the same way as it is for the electrostatic potential. However, it is now accepted
that this node-based representation is not always the best. For one thing, it imposes full
continuity on the vector field, and not all fields are evevwhere continuous. For example,
the normal component of an electric field will jump discontinuously across a dielecvioair
interface.
A popular and useful alternative is an edge-based element, such as the Whitney
tetrahedron (figure 3), introduced by Nedelec (1980). The vector field H in this element is
given by:

where the hi are unknown parameters and the v; are known vector basis functions, linear
functions of the coordinates x , y , z in the element. The six degrees of freedom h; are
nor the values of any component of H . However, each is associated with an edge of the
tetrahedron in the following way: the tangential part of H along edge i is determined
solely by hi. In other words, to make H tangentially continuous across an edge shared
by two tetrahedra, it is only necessary to relate the two his for that edge, one from each
tetrahedron. In fact, the two his are simply set equal (with a change of sign if the directions
chosen for the two edges are opposite). In a mesh of Whitney tetrahedra, there is one

1680

J P Webb
1

Figure 3. A vector finite element: the Whitney temhedron.

degree of freedom per edge of the mesh and the resulting field H is tangentially continuous
from one element to the next, across both edges and faces. The normal component is, in
general, discontinuous; it is left to the weighted-residual formulation to impose appropriate
conditions on the normal continuity.
Since the introduction of the Whitney tetrahedron, other elements of this kind have been
invented, tetrahedral and brick-shaped in 3D and triangular and quadrilateral in 2D. They
are now used widely in vector electromagnetics (Webb 1993). The higher-order varieties
have basis functions associated with faces as well as edges, but their common feature is
that they impose only tangential continuity. The most accurate name for them is perhaps
tangentially continuous vector elements. However, it is common (and shorter) to refer to
them all as edge elements.

2.3. Solving the matir equation


In most cases, application of the finite element leads to a set of linear simultaneous equations
which may be written in matrix form:
where ( U } is the column vector of unknowns. The matrices tend to be large particularly in
3D,but are very sparse. Since for large problems most of the computation time is consumed
by the solution of the matrix equation, it is very important to use a technique that takes full
advantage of the sparsity.
It is outside of the scope of this paper to discuss all the methods that might be used to
solve a large sparse set of linear simultaneous equations. However, certain techniques have
become widely adopted for the finite element analysis of electromagnetic devices and these
will be briefly described.
For static problems, such as the electric field problem described above, the
preconditioned conjugate gradient method works extremely well. This is an iterative method,
i.e. it begins with a guess at the solution and iteratively improves the guess until it is
deemed accurate enough. Before the main iteration loop begins, an approximate inverse of
[SI is computed and used to precondition the problem so as to hasten convergence. Thcre
are several ways of computing an approximate inverse, but the most commonly used is an

Application of rhe finite-element method

1681

incomplete factorization of [SIinto a product of lower and upper triangular matrices. Such a
factorization is known as an incomplete Choleski decomposition. A complete decomposition
would, in effect, solve the problem, but it would be expensive. The incomplete (i.e.
approximate) factorization is fast-it takes the same time as a few iterations. A conjugate
gradient method preconditioned in this way is known as the incomplete-Choleski conjugate
gradient (ICCG) method (Kershaw 1978). Typically, if the matrix dimension is NI the
number of iterations needed to obtain a satisfactory solution with ICCG is proportional to the
square root of N, . The memory requirement also is modest: it increases linearly with Nf
because only the non-zero entries in [SIand its preconditioning matrix need to be stored.
In some problems the matrix [SI is unsymmetric, e.g. when computing the magnetic
field in the presence of moving conductors. In such cases the biconjugate gradient (biCG) algorithm can be used instead (Fletcher 1976). An incomplete factorization of the
matrix into lower and upper triangular matrices serves as a preconditioner (e.g. Brussino
and Sonnad 1989).
For time-harmonic problems a phasor analysis is usual, i.e. the vector [ U ) and the
matrix [SIare complex. Jacobs (1986) described a generalization of bi-CG to the complex
case. (For symmetric matrices it amounts to nothing more than CG performed in complex
arithmetic.) Unfortunately, there is no guarantee of convergence for bi-CG. In general, for low
frequency problems, in which the smctures are far smaller than the freespace wavelength,
preconditioned bi-CG works almost as well as ICCG does for static problems. However, at
higher Erequencies, and particularly once the structures are larger than a wavelength, the
performance degrades rapidly. Despite this, the conjugate gradient method is still often
used for large high-frequency problems, partly because of its low memory requirements.
Alternatives to these iterative techniques are the direct methods, i.e. methods based on
Gaussian elimination and its variants, in sparse form (see Duff er ai 1986). The frontal
method and the nested dissection method have both been used to solve high-frequency
problems.
2.4. Unbounded problems

Since it is impossible to fill all space with a finite number of finite elements, some other
means must be found for solving unbounded problems. The general approach is to use
finite elements inside a surface SO that surrounds all materials and sources, and outside to
use some other numerical method (figure 4). A wide range of methods is available. Each

Exterior

air

Figure 4.

Handling unbounded problems with the finite

element method.

1682

J P Webb

can be described as either local or global. A local method is one which does not affect the
sparsity of system matrix [SI,because it only couples nodes that are close together on SO.
Global methods link all the unknowns together and result in a dense submatrix of [SI.The
submatrix has dimension at least equal to the number of unknowns on So,and can cause
the solution times and matrix storage requirements to be excessive.
A further consideration in discussing the methods is the distinction between static and
quasi-static problems on the one hand, and wave problems on the other. For static and quasistatic problems the electromagnetic fields outside So decay smoothly, with no oscillation
spatially and no wave propagation. For wave problems, the field in the exterior takes
the form of a wave moving outwards. Although the wave decreases in amplitude, the
wave characteristics persist to infinity, and must be properly accounted for. Some of the
methods discussed below can handle static and quasi-static problems adequately, but not
wave problems; others can be applied to both.
The following are the more commonly used local methods:
Simple consrruinr. The simplest thing to do with SO is to constrain the field to be zero
there. Since in all cases the field decays to zero at infinity, ifso is sufficiently far away from
the sources, imposing this constraint will not affect the result too much. For quasi-static
problems this approach works quite well. If the diameter of the smallest sphere containing all
the materials and sources is 0, giving So a diameter of 50-100 generally gives acceptable
accuracy. Of course, this large empty volume has to be filled with elements. However, the
elements can be made larger further out, because the field variation decreases with distance
away from sources, so the extra cost is usually modest.
Unfortunately, for wave problems this simple method will not work. First, a boundary
constrained in this way reflects all energy incident on it, and does not allow for any radiation
loss through Sa. Worse, the reflected wave may even be re-focused on the source region,
creating large errors there. Finally, the element size in the large empty region surrounding
the sources cannot be increased to the same extent as in the quasi-static case, because the
elements cannot be larger than a fraction of a wavelength without introducing errors.
Infinire elements. An infinite element in 3D is a brick-shaped (six-faced) element in which
one of the faces is at infinity. A layer of such elements surrounding So, each with one
face on SO and the opposite face at infinity, will, in effect, fill all of the infinite exterior.
Infinite elements work in much the same way as finite elements: basis functions are used
to represent the field and the weighted-residual method reduces the differential equation to
a set of algebraic equations. The key difference is in the basis functions, which are chosen
to have the same asymptotic behaviour at large distances from SO as the w e field. Li er
a1 (1994) reported the use of infinite elements for 3D eddy-current (quasi-static) problems.
Towers et a1 (1993) analysed scattering of waves from 2D objects using infinite elements.
Perhaps the biggest obstacles to their wide acceptance a e difficulties in integrating over
the infinite domain, and the fact that they provide only a relatively simple variation of field
with distance outwards from SO,which means that in some cases, So must still be placed
quite far away from the sources to obtain sufficient accuracy.
Transformation method. By mapping (transforming) the infinite exterior volume into a
finite volume, the problem becomes one with a finite domain-or, rather, two finite domains,
the interior and the (mapped) exterior. Both domains can be subdivided into finite elements
and analysed in the usual way, provided only that the mapping is incorporated into the

Application of the finite-element method

1683

calculation of the global matrix entries. Freeman and Lowther (1988) solved static problems
in this way, choosing SO to be a circle, and mapping the outside of the circle into another
circle (the Kelvin transformation). In this case it turns out that the transformation has no
effect on the calculation of the matrix entries, so the whole procedure can be carried out
rather neatly with an existing finite-element program. Imhoff et a1 (1990) mapped the
infinite domain to a circular or spherical shell surrounding SO. A very general mapping as
described by Stochniol (1992).
The assumption behind the transformation method is that if the elements and basis
functions of the mapped, finite domain were inverse-mapped back to the real, infinite
domain, they would represent the true field there adequately. In the quasi-static case,
this is likely to be so because the field variation decreases with distance from SO and large
elements are permissible. However, it is unlikely that the transformation methods would
work for wave problems.

Absorbing boundary conditions. An absorbing boundary condition is a boundary condition


placed on SO to absorb waves incident on it from within. A simple example for a timeharmonic wave problem for scalar U is the Sommerfeld radiation condition:

where R is the distance from an origin within SO;ko is the free-space wavenumber, w IC; w is
the angular frequency of the wave and c is the velocity of light. Such a boundary condition
is only truly accurate at infinity, but will partially absorb outgoing waves if applied to SO.
.Other, more accurate boundary conditions are possible, all of the form (10).
One way of developing such boundary conditions is to expand the field in the exterior
as an infinite series whose nth term is
e-jkoR

-g"

(15)
R"
where g. is independent of R. An mth-order differential operator can be built which will
annihilate the first m terms of the series. From the operator a boundary condition of the
form (10) can be derived. The boundary condition is satisfied exactly by any wave consisting
of just the first m terms, and can be considered as a generalization of the Sommerfeld
radiation condition. Boundary conditions derived from higher order operators are more
accurate, and permit SOto be reduced in size. In practice the second-order operator (m = 2)
is most often used.
In the basic theory propounded by Bayliss etul (1982), S, is spherical, but there have
been extension which permit other boundary shapes (e.g. Mittra er a1 1989), albeit at the
expense of some loss of accuracy or of matrix symmetry. The extension of the technique
to vector electromagnetic wave problems was presented by Peterson (1988), Webb and
Kanellopoulos (1989) and Chatterjee and Vola!& (1993).

Measured equation of invariance and numerical absorbing boundary condition. Like


absorbing boundary conditions, these are local methods specifically designed to absorb
outgoing waves. When the finite element program is run a linear relationship is established
between the unknown uo at a node on So and the unknowns U,,U I , . . . ,U, at m nearby
nodes (on the surface or in the interior):

1684

J P Webb

The coefficients cj are chosen so that the relationship (16) is satisfied exactly when the
unknowns U, are the nodal values of a chosen function, called a measuring function. A
measuring function is an outgoing wave of some sort. By applying this condtion to m
independent measuring functions, the m coefficients in (16) can be determined.
Once a linear relationship like (16) has been established for each node on So, these
equations are solved together with the usual finite-element equations for the interior.
The equations (16) for all the nodes on So together constitute an absorbing boundary
condition. Since the boundary condition is derived numerically at run time, it has been called
a numerical absorbing boundary condition. The term measuring function, and measured
equation of invariance, are due to Mei et al (1992).
This approach is quite promising but is in the early stages of development. The
appropriate choice of measuring functions in the general case is still the subject of researchsee, for example, Wright and Cangellaris (1994).
The final three methods are global:
Hybrid A hybrid method combines finite elements and boundary elements. In the interior,
finite elements are used to solve the governing partial differential equation. On SO,boundary
elements (triangles or quadrilaterals) are used to solve an integal equation, i.e. an equation
in which the unknowns appear inside an integral. The nature of the integral equation depends
on the type of problem being solved. In electrostatics, it is

for all points T on SO,where G is a known function (a Greens function). This equation can *
be solved numerically using boundary elements to give an approximate relationship between
aQlan and @ on the surface So-in other words, to give a version of the general boundary
condition (10). This boundary condition is then applied to the finite element analysis of the
interior.
The hybrid approach is quite accurate and imposes no restrictions on the shape or
closeness of the boundary SO. It was one of the earlier methods for extending the finiteelement analysis of electromagnetic problems to unbounded domains (Morgan and Mei
1979). but like all the global methods it has fallen out of favour because of the large cost
associated with the dense submatrix. Salon and DAngelo (1988) discussed its application to
electromagnetic problems. Lynch et a/ (1986) used it to analyse unbounded wave problems
that arise in the treatment of hyperthermia.
Series expansion. Using the technique of separation of variables, the electrostatic potential
outside a spherical boundary SO can be expanded in an infinite series of terms, each of
which is an exact solution to the Laplace equation and has the correct behaviour at large
distances from the sources. The coefficientsof these terms are unknown, of course, but since
the potential must be continuous across SO it is possible to relate them to the unknowns
of the finite elements. For example, in the transfinite element method (Lee and Cendes
1987) the potentials at nodes on SO are eliminated by expressing them in terms of the
coefficients.
Once this is done, the (truncated) infinite series can be used to express a@/an on SO
in terms of the unknowns of the problem, i.e. to provide a boundary condition of the form
(IO). The interior finiteelement problem is then solved in the usual way.
The dense submatrix arises because the coefficients multiply functions which are nonzero over the entire surface SO. thereby linking together all the unknowns associated with

Application of the finite-element method

1685

the surface.
Coupled pairs of basis functions. A technique that can partly overcome the difficulty of
the dense submatrix associated with global methods is the following. Choose a set of m
basis functions, Yi, on the surface SO.Impose the boundary condition a @ / & = Yi and find
the resulting potential inside SO with the finite-element method (note that a@fan = \Yi is a
simple case of the general boundary condition (10)). Let Qi be the potential on SOobtained
from the finite-element soltuion. The functions (Vi,Q;) form a coupledpair. The complete
set of coupled pairs can be found by m entirely sparse finite-element analyses. There is, in
addition, a function QOwhich is obtained with the boundary condition aQ/an = 0.
Now, in general, on SO

where the coefficients ci are problem-dependent. On the other hand, the global method
(boundary integral or series expansion) gives another relationship between a@/an and Q
on So. Substituting (18) into this relationship gives a single equation for the m unknowns,
ci; applying the weighted-residual method gives m equations for the m unknowns, and
hence a solution.
Mei (1974) applied the technique to two-dimensional scattering problems, using a seriesexpansion method. Morgan et a1 (1984) tackled the same class of problem using a boundary
integral method.
In as much as one step of the technique is the solution of a dense m x m matrix
problem, it might seem that nothing has been gained over the hybrid or series-expansion
method, in which the dense submatrix is equal in dimension to the number of unknowns
on SO. However, it is found that by a judicious choice of basis functions, the number m
can be made considerably smaller than the number of nodes on SO,so the dense matrix is
relatively small. On the other hand, there is a heavy price to be paid in that not one but
m 1 finite-element solutions must be calculated.

3. Magnetic fields

Maxwells equations say that magnetic field is produced by two types of current: the
conduction current J , which is a flow of electric charge; and the displacement current
a D / a t , which arises whenever the electric flux varies in time. For a large number of
electromagnetic devices, particularly those operating at low frequencies, the displacement
current has a very much smaller contribution to the magnetic field than the conduction
current, and for all practical purposes can be neglected. The study of Maxwells equations
under this assumption is known as mngneto-quasi-statics. It is the study of the magnetic field
and the conduction currents that both generate it and, via Faradays law, are generated by
it. Devices that are amenable to quasi-static analysis include electric motors and generators,
transformers, actuators, magnetic recording heads and induction furnaces.
Under the quasi-static approximation, wave phenomena no longer occur and there is no
radiation to deal with. The magnetic field at each instant of time is just the static field of
the currents at that instant.
If the current density J is known in advance, the problem is just to find the magnetic
field produced by those currents, i.e. to solve the static equations:

VxH=J

1686

J P Webb

V.B=O
together with a constitutive relation, such as
B=pH

(20)

(21)

where H is the magnetic field, B is the magnetic flux density and /I is the permeability. This
problem arises surprisingly often, and will be referred to henceforth as the magnetostatic
case. The currents can always be found in advance when they do not vary in time (perhaps
by a prior electric field analysis-see section 4). Even timevarying currents have a known
distribution in stranded conductors (the coils of fine wire often used to generate magnetic
fields), because each strand carries the same current. In addition, the practice of laminating
iron cores to reduce induced currents often allows the approximation J = 0 to be made in
such cores.
However, when there are solid conductors present, and the conductors are either moving
with velocity U or are in the vicinity of time-varying currents, there will be currents induced
in them. This is the eddy-current case. The currents have a magnitude and distribution that
cannot be predicted in advance; they are subject to the following equations:
aB

V x E = - - ar

(22)

V-J=O

(23)

together with another constitutive relation, usually Ohms law:

J = o(E+ U x B)

(24)

where E is the electric field and U is the conductivity. The problem can then only be solved
by tackling equations (19)-(24) simultaneously. (For simplicity, the velocity term in (24)
will be omitted henceforth.)
The quasi-static approximation simplifies Maxwells equations and thereby offers the
possibility of more efficient finite-element solutions that can be obtained from a treatment
of the full equations. However, this possibility has led to a proliferation of alternative
methods of solution, based on different potentials to represent the magnetic field and the
current density. The first papers appeared in the early seventies: Silvester and Chari (1970)
and Chari (1973). In section 3.1 an attempt is made to describe the main themes of the
continuing work on potential theory.
Magnetic devices are often driven by sources in external circuits, and coupling the
magnetic field analysis to electric circuits is discussed in section 3.2. The finite-element
method reduces the problem to a set of ordinary differential equations in time, and the
solution of these is the subject of section 3.3. One of the complications of magnetic field
analysis is the existence of materials with non-linear and even multi-valued constitutive
relations, and this subject is explored in section 3.4. Section 3.5 deals with post-processing
the computed field lo extract net loading forces and torques.

3.1. Potentials
The tendency in quasi-statics has been to express the magnetic field or the current density
in terms of potentials and then to solve for the potentials rather than the fields themselves.
There are two major reasons for this. First, potentials are often continuous at abrupt material
interfaces where one or more components of the magnetic field or the current density is
discontinuous. Using nodal elements to represent vector fields, it is difficult to handle such
discontinuities. Before the invention of edge elements, potentials provided the answer. The

Application of the jinite-element method

1687

other main reason for using potentials remains as valid today as it ever d i d a scalar potential
or a vector potential with only two components can be used in some regions instead of a
three-component vector. The reduction in number of components translates intp a reduction
in the size of the system mauix, i.e. less memory and smaller computation times.
There are two ways to introduce potentials in quasi-statics:

(i) From equation (20) it follows that


(25)

B = V x A

where A is the magnetic vector potential. Substituting this into (22) leads to

where @ is the electric scalar potential.


(ii) From equation (23) it follows that
(27)

J=VxT

where T is the electric vectorpotential. Substituting this into (19) leads to

H = -VSt+T
where S2 is the magnetic scalar potential.

(28)

A - @ Methods. A good review of these methods was provided by Bk6 and Preis (1989).
Expressing B and E in terms of A and @ in (19) leads to a partial differential equation
for the potentials
V x WVx A -

(-U?

-uVO) =0

The remaining equation, (23), is implied by this and adds no additional constraint on the
potentials.
The term in brackets is the current density, J . As explained above, in many problems
the current density is known in advance, In these problems the term in brackets is simply
replaced by the known current density, J, and the electric scalar potential disappears-the
unknown is just A . This is the megneto static case.
A difficulty that arises in looking for A and @ is that A is not fully defined, Only its
curl is specified (by equation (25)) and that is not enough. If ( A , O) is any one solution
to (29), so too is ( A V g . @ - ag/at), for any scalar function g. This uncertainty does not
matter from the point of view of the flux density, because all solutions lead to the same B .
However, it makes the problem difficult to solve numerically. For this reason, equation (29)
is usually supplemented with another condition, called a gauge condition, to fix A and 0
uniquely. A convenient classification of the methods of solution is based on the choice of
gauge.

No gauge. Biddlecombe et al (1982) explained how it is possible to find A and 0 with


no gauge-at least, no explicit gauge. The pariicular elements they used in fact provide an
implicit gauge which leads to a unique solution.
The magnetostatic case can also be solved with no gauge if a suitable mahix-equation
solver is used. Preis et al (1992) reported some success using the ICCG method to solve
the matix equation, with edge elements representing the vector potential. However, though
better than the corresponding nodal-element method, this approach is generally unreliable.

1688

J P Webb

Coulomb gauge. The classical answer to the uncertainty of A is to specify its divergence.
The choice V A = 0 is referred to as the Coulomb gauge, and can be applied by adding
to equation (29) the term -VuV. A. Once this term is added, however, equation (23) is
no longer implied by (29) and the potentials must be made to satisfy this also:

v.(-uaa-uv'4)=o.
at

(30)

Equations (29) (with the extra term) and equation (30) together have a unique solution. The
method can also be applied to the magnetostatic case, though in that case '4 does not appear
and the supplementary equation (30) is unnecessary.
The Coulomb gauge can only be used with finite elements which impose the continuity
of the normal component of A across interfaces, e.g. nodal elements. Preis et a1 (1991)
suggested that this continuity, combined with the gauge itself, can lead to significant
numerical errors at iron-air interfaces where the tangential components of the flux density
change abruptly.

Modified vector potential. The gauge

v.

(*g)
=0

has the interesting effect of decoupling A and '4. Substituted into (30), it produces an
equation for '4 alone, which can be computed in advance from the known voltages applied
to the ends of the conductors or net currents through them. From '4 is obtained part
of the current density, -uV@, often called the source current and given the symbol Js.
Equation (29) is then an equation for A alone, with J, as the driving term. The resulting
A is called the 'modified' vector potential (Emson and Simkin, 1983).
An important point to note is that this gauge can only be applied when both U and
a/at are non-zero. When either vanishes, the modified vector potential becomes ungauged;
if either becomes small, the numerical problem becomes ill-conditioned. As a result, the
gauge cannot be used in magnetostatics, and even in eddy-current problems it must be
combined with other techniques that can handle the non-conducting regions. Despite these
restrictions, the modified vector potential is used widely.
One of the implications of the gauge is that, at an interface between two regions of
different conductivity, the normal component of A is discontinuous. Nodal elements do
not permit this, and therefore cannot handle abrupt changes of conductivity. The obvious
solution is to use edge elements for A, and this has been done (Kameari 1990).
Low-frequency Lorentz gauge. The condition

V .A = -/LU@

(32)

is known as the low-frequency Lorentz gauge. As explained by Bryant e f af (1990), it can


be used to eliminate '4 from (29) and obtain a differential equation in A alone. This yields
the magnetic field, but if the current density is needed it is also necessary to find '4 in the
conductors, by solving a second differential equation, (equation (30)).
In non-conducting regions the gauge reduces to the Coulomb gauge, which can be
applied aby adding -VuV.A to (29), as mentioned above. Therefore, unlike (31), the lowfrequency Lorentz gauge can be used in all regions of an eddy-current problem. Furthermore,
even when there are abrupt changes in conductivity A is fully continuous, so nodal elements
can be used.

Application of the finite-element method

1689

Tree gauge. In the magnetostatic case, one way of gauging A is to make use of certain
properties of edge elements. Consider a simple mesh of Whitney triangles (figure 5). Each
degree of freedom of A is associated with an edge: there are 12 of them in this mesh. The
trouble is, 12 is too many. There are many fields A on the mesh which have the same curl
and therefore correspond to the same flux density. We can reduce the number of degrees
of freedom by picking some of the edges and setting to zero the variables associated with
them. But which ones do we pick?

t Spanning tree edge


Figure 5. Tree and CO-tree edges used to impose a gauge on
the vector potential A.

The answer lies in the spanning tree of the mesh. A tree is just a set of edges which
does not contain any closed loops (cycles). A spanning tree is a tree that spans the whole
mesh, i.e. every node in the mesh is connected to at least one of the edges of the tree. It
turns out that by picking the edges of a spanning tree and setting their variables to zero,
A is constrained in exactly the right way, i.e. no two fields have the same curl (Albanese
and Rubinacci 1988). In our example, this reduces the original 12 degrees of freedom to 6,
each of which is associated with a non-tree (or co-free) edge.
This strategy can also be viewed as a way of reducing a three-component A to two
components (Albanese and Rubinacci 1990).
Preis er al (1992) reported that for some choices of spanning tree, the resulting system
matrix is ill-conditioned, leading to a large number of steps of the conjugate gradient method.
Albanese and Rubinacci (1992) discuss ways of choosing a 'good' spanning tree.

T - C2 Methods. Expressing J and H in terms of T and Q in (22) leads to a partial


differential equation for the potentials:
1
v x -V
U

XT+

za ( -,VQ)=O.
~ ~

(33)

The remaining equation (ZO), is usually implied by this and adds no additional constraint
on the potentials. The obvious exception is magnetostatics, where a / & = 0, but in this
case a considerable simplification is possible. Since the current density, Js,is known, the
vector T can be obtained directly from it by inverting (27). There are a number of ways
of doing this which are discussed below. The resulting electric vector potential acts as a
source of magnetic field, and is designated T,. The magnetic scalar potential can then be
found by solving (20), which becomes

v . (pT, - pvn) = 0.

(34)

Returning to the general case, equation (33) does not have a unique solution unless a gauge
is imposed on T ;without it T is not fully specified. Once again, the various formulations
may be classified according to the choice of gauge, and are discussed below.

1690

J P Webb

Finding the source field T,. The problem is: given a current density J, , find a field Ts
whose curl is equal to JS.There are infinitely many such fields and any one of them will
do.
The earliest approach to this problem was to throw away the magnetic materials and find
the DC magnetic field in free space due to J, using direct integration (the Biot-Savart law).
The resulting field has the correct curl and is therefore a suitable T,. However, it became
apparent that source fields produced in this way give rise to unacceptably high levels of error
in highly-permeable regions. The reason is that the net magnetic field in these regions is
usually very small, much smaller than its components Tsand -Vi2 (equation (28)). Adding
two large terms to give a very small term can produce significant cancellation errors. For
this reason, Simkin and Trowbridge (1979) devised an alternative method in which the magnetic field in highly permeable regions is represented by a total potential, Ilr: H = -VY.
The S2 and Y regions are stitched together by appropriate interface conditions.
While this approach has been successful, it has certain drawbacks. Consider a closed
torus of iron with a current-carrying wire passing through the hole in the middle. The total
potential cannot be used throughout the iron because doing so sets to zero the circulation of
magnetic field around a path in the iron encircling the wire, which contradicts Ampbres law.
Another drawback is that the Biot-Savart integration can be time consuming, particularly
for complicated coil shapes.
To reduce the complexity of the two-potential approach, Mayergoyz et a1 (1987) proposed that T, should be the magnetic field that results when the iron is replaced not by air
but by a material of infinite permeability. In practice this is done by excluding the iron regions from the finite-element mesh and imposing the boundary condition that the tangential
part of T, vanishes on the surface of the iron. Since there is no iron in this problem, it can
be solved with a single potential without any difficulty. The computed field, T,. is then
taken as the source field for the second finite-element problem, in which a single potential
is again used in both air and iron. There is no cancellation error in the iron because, by
construction, TIvanishes there.
Gyimesi et ai (1993) pointed out that this approach still does not overcome the torus
of iron problem. Their solution is to generate an additional magnetic field for the iron
regions alone. The source field TI then consists of the computed solution for the non-iron
parts, plus the computed solution for the iron parts.
The methods so far have all been based on the idea that T,is an approximation to a
physically correct magnetic field. In fact, though this may be convenient, it is by no means
necessary. The source field may be quite non-physical, just as long as its curl is equal to
the specified current density. The method of Webb and Forghani (1989) makes use of this.
The source field is represented by Whitney tetrahedra, i.e. there is one paramater per edge
of the mesh. The parameters are determined by an iterative procedure based on the satisfaction of Amperes circuital law around each element face. The scalar S2 is represented by
first-order tetrahedra that coincide with Whitney elements. With these choices of elements,
it is possible for VR to exactly match the variation of T, over a current-free region; as a
result, there is no cancellation error in iron.
The calculation of the source field by this technique is quite efficient; for complicated
coil shapes, it is much faster than applying the Biot-Savart law. Kladas and Tegopoulos
(1992) showed that for some coils TIcan even be obtained analytically.

If R is set to zero, T becomes the magnetic field itself


and is therefore unique. The obvious objection to this gauge is that the magnetic field is

i2 = 0: the magneticfieidmethod

Application of the finite-element method

1691

discontinuous whenever the permeability changes abruptly; if nodal elements are used to
represent T,the discontinuity cannot easily be modelled. The solution is equally obvious:
use edge elements (Bossavit and V6rit6 1982).
Like the modified vector potential A, this T becomes ungauged when a p t is zero, and
the numerical method is ill-conditioned when a / a t (or the frequency) is small. Webb and
Forghani (1993) showed that the performance of the conjugate gradient method degrades
severely as the frequency is lowered. A more reliable edge-element method is one using
the tree gauge, below.
Coulomb gauge. Bir6 et a1 (1993) used nodal elements for T and imposed a Coulomb
gauge, V T = 0,by adding -VpV T to (33), where p is the resistivity. Their method
is well conditioned and seems to be more efficient overall than the previous method.

Tree gauge. The uncertainty in T can be removed by means of a tree gauge, exactly as for
A (Albanese and Rubinacci 1988). This gives what is essentialy a two-component T,and
a scalar a-three variable in all, the same as for the magnetic field method. Interestingly,
if Whitney tetrahedra are used for both methods, the computed H will be exactly the same;
however, the tree gauge has no difficulty with low frequencies (Webb and Forghani 1993).
To date, there have been no reports of ill-conditioning arising from some choices of spanning
tree, as it does in the magnetostatic case when A is used.
Figure 6 shows a helical coil carrying current at 200 Hz, being used to heat a solid
cylinder of conducting material. The diameter of the cylinder is approximately five skindepths. This 3D time-harmonic problem was solved by a commercial finite-element package
(Infolyticas MagNetS) which uses the T-Q method with a tree gauge. The primary quantity
of interest was the power distribution in the cylinder.

Figure 6. An induction heater.

Which are the best potentials to use? In magnetostatics, and for the non-conducting regions
of eddy current problems, it is fairly clear that the most efficient methods are those based
on the magnetic scalar potential, rather than the magnetic vector potential.
In the conducting materials of eddy current problems, it is not possible to represent the
electromagentic field with a single scalar. In fact, the best that has been managed so far is
three components: either a three-component vector, or a two-component vector and a scalar.
Of the methods based on A, the Lorentz gauge seems to be the best; the Coulomb gauge
involves four components and the modified vector potential has difficulties when a/& is
near zero. Of the methods based on T, the tree gauge is better than the use of the magnetic
field directly, and gives a solution in three components. Again, the Coulomb gauge would
seem to be less attractive because it retains all four components.

1692

J P Webb

The potentials for non-conducting and conducting regions of an eddy current problem
have to be stitched together by means of appropriate interface conditions. If Q is used for
the non-conducting region, then it is rather more simple and natural to use a T-C2 method
for the conductors than to switch to a magnetic vector potential. For example, if the tree
gauge is used for T , the tangential components of T can easily be made to vanish on the
surface of conductors (by setting to zero unknowns associated with edges on the surface).
Continuity is then ensured by simply using a magnetic scalar potential that is continuous
throughout the problem domain, from conductors to non-conductors. There is no interface
condition to impose.

Potentials in zD. There are two commonly used 2D models in magnetics: rectangular and
cylindrical. The rectangular model applies to problems in which there is translational
symmetry, i.e. no variation of geometry along the z-axis of a Cartesian coordinate system,
and for which the currents are entirely z-directed. There are a surprising number of devices
for which these assumptions are nearly enough satisfied to give reasonable ZD results. For
example, the currents in motors and generators are mainly axial, and the cross section often
does not vary along the axis (until the ends, of course, but unless the device is short, the
end effects are often neglected).
Under the assumptions of the rectangular model, it can be shown that the magnetic
field can be represented by just the z-component of magnetic vator potential, which in
addition does not vary with z. Such a vector potential automatically satisfies the Coulomb
gauge, so gauging is not a problem in ZD. Both magnetostatic and eddy-current problems
may be analysed with this single-component A without any difficulty. There are no serious
contending potentials. The magnetic scalar-potential can be use for magnetostatics, of
course, but the need for a source field to handle currents adds extra complexity and there
does not seem to be any clear advantage.
The cylindrical model applies to problems in which there is axisymmerry, i.e. no variation
of geometry along the 6 axis of a cylindrical coordinate system, and for which the currents
are entirely &directed. Again, there are a large number of devices for which the assumptions
are nearly satisfied, and to which the cylindrical model may be applied.
In this case, only the &component of A is needed to represent the magnetic field,
and it does not vary with q4, As before, the Coulomb gauge is automatically satisfied,
and both magnetostatic and eddy-current problems may be analysed with this potential.
However, there has been some debate over the exact form of the potential that is best suited
to finite element analysis. Using A+ itself leads to terms involving inverse powers of r ,
the distance to the axis. To compute the finite element coefficient matrix these terms have
to be integrated, which is difficult when the problem domain includes the axis ( r = 0).
Using the potential A,/r instead eliminates the singular terms and makes the integration
straightforward. However, both of these choices can lead to inaccurate results for seemingly
simple problems (Melissen and Simkin 1990). Consider, for example, a thin cylinder of
iron centred on the axis. Let us assume that flux is driven along the cylinder, parallel to
the axis, and that the flux density outside the iron is essentially zero. Figure 7 shows the
disrribution of axial flux density, Bz, with radius r . This is a fairly common situation, and
in the translational case could be handled with a small number of elements, because the
piecewise constant flux density would lead to a piecewise linear potential. However, in
the axisymmeVic case the corresponding distribution of A+ is not so simple (figure 8). In
particular, it falls off as l / r outside the iron core, and this region will give significant errors
unless it is carefully refined. The potential A4/r is even worse, because it falls off as l/rz

Application of the finite-element method

Flgure 7. Magnetic flux density in an axisymmeuic


problem with a cylinder of iron on the axis.

1693

Figurt 8. Magnetic vector potential A+ corresponding


to the flux density in figure 6.

outside the iron.


A solution to this problem is to use A+r as the potential. In the air the potential is
then constant; in the iron it is quadratic and can be modelled exactly with second-order
elements. Of course, the integration problem remains, but is not insurmountable. It can, for
example, be overcome by transforming the whole problem to the s-z plane, where s = r 2
(Melissen and Simkin 1990).
3.2. Connecting to circuits
Many magnetic devices are operated by connecting them to electrical circuits. The magnetic
field is required when a specified voltage is applied (the voltage-driven case), or a specified
current flows through a conductor (the current-driven case). Sometimes neither the voltage
nor current is known, and the electrical circuit equations must be solved simultaneously
with the magnetic field equations. How are these circuit connections made?
Consider first a solid conductor which has a potential difference Vo between its two
ends, designated and - (figure 9). and a current 10 flowing from the end to the - end.
Suppose that the potentials used to solve the problem are A and @. The voltage-driven case
can be handled very simply: the electric potential @ is set to zero over the cross section at
the - end of the conductor and set to the constant value Vo over the cross section at the
end. In the current-driven case, VO is not known. It is used as an additional unknown

Figure 9. 'Terminal voltage and c m n t for a solid


conductor.

I694

J P Webb

N-turn coil

Figure 10.
conducror.

Terminal voltage and cumnl for a stranded

of the problem, which can therefore only be solved if there is an additional equation. This
equation is the one that links 10 to the electromagntic field. It says that IO is the integral of
the current density over the conductor cross section, S:

For formulations based on T and S2, the current 20 can be prescribed without any
additional variables or equations. A fictitious stranded coil is built that follows the path
of the solid conductor, and lies wholly within it. The source field TI of this coil is found
(section 3.1) when it carries a current l o . and the total magnetic field is represented as the
sum of T, -VQ, and TS.Provided T is constructed so that its curl is a current density
which carries no net current through the solid conductor, the latter will automatically cany
a net current IO. One way of enforcing this condition on T is to make its tangential
components vanish on the surface of the conductor-as it must anyway, for the tree-gauge
T-f2 method. For the voltage-driven case, lo is used as an additional unknown, together
with an extra equation that says that VO is the path integral of V Q between the - and
ends of the conductor:

The integral can be manipulated into a form involving T , -VQ, and T,.
A stranded conductor (figure 10) with known current 10 is handled easily by both
A 4 and T-S2 methods because the distribution of current is known, so this is just the
magnetostatic case. When the current is unknown, equation (36) is added, remembering
that the path from - to has to go around all the turns of the coil. Note that for an A-@
method it is not possible to simply constrain Q at the ends of the coils as was done for
the solid conductor, because stranded conductors have to be treated as non-conducting to
prevent current redistribution and therefore Q disappears from the differential equation (29).
When neither the current nor the voltage is known, one or more additional equations
must be obtained from a consideration of the external electrical circuit. Tsukerman et a1
(1993) presented a survey of recent methods.

3.3. Time andfrequencydomain solution


The weighted-residual method, when applied to any of the potential formulations, leads to
a set of equations in which the space variables ( x . y , z ) no longer appear, i.e. the variation
in space is eliminated. In the magnetostatic case there is no time variation either, and the
result is the matrix equation

[Sllu) = ( 6 )

(37)

1695

Application of the jinite-element method

where [SI is a real, N, x N, matrix and { U ) is an N,-vector of unknowns. The vector


on the right-hand side represents the sources in the problem, e.g. voltages and currents.
In general [SI depends on [ U } via the material properties (permeability and conductivity),
and (37) is nonlinear in ( U ) , However, when the materials are linear, i.e. when the material
properties are not functions of field strength or direction, [SI is independent of [ U ) and (37)
is a linear matrix equation, most often solved by ICCG (section 2.3). The nonlinear problem
is discussed further in section 3.4.
On the other hand, if the problem is time varying, after applying the weighted residual
method, [U} still depends on time. The result is a set of coupled ordinary differential
equations in time, that can be written in matrix form:
d
(38)
[AI@) + [ B I z b } = 1 4
where [ A ] and [ E ] are real, N f x N , matrices. The solution of (38) in the time domain
is most often accomplished by an implicit, Crank-Nicolson time-stepping scheme. The
potential at time t At is calculated from the potential at time t by solving

- d(U) = 0 .
[AHi)+ [BIT

(39)

The overbar denotes evaluation at a point in time halfway between t and f


At.
Approximating the variation of { U )over the time step as linear in t , this equation reduces
to
[S](K(t

+At))= [ b ) .

(40)

The matrix [SI and the vector {b) depend on the potential at time t , which is presumed
known. In addition, if there are nonlinear materials, [SI will also depend on ( u ( t At)),
so the equation is nonlinear in the unknown potential. It is similar to that obtained in the
magnetostatic case. Of course, an equation like this must be solved at each time step. so
time-domain problems can be extremely expensive.
Tsukerman et a1 (1993) surveyed other ways of solving (38).
Another important class of problem is that for which the excitations are. sinusoidal in
time. If the materials are linear, the potentials will also vary sinusoidally in time, and the
problem can be handled using complex phasors. Equation (38) becomes

(14 j w [ B I ) ( u ) = [ c )
(41)
where o is the angular frequency. Apart from the fact that the coefficient matrix (the term
in parentheses) and the column vectors in this equation are complex, it is very similar to
the magneostatic case. It is generally solved by a variant of the conjugate gradient method
that is suited to complex matrices-see section 2.3.
When the excitations are sinusoidal, but the materials are nonlinear, the resulting field
contains components at more than one frequency and phasor analysis is not possible.
However, since the fields are periodic in time, an alternative is to eliminate the time
dependence by expanding the potentials in a number of harmonics and solving for those.
Albanese er al (1992) showed that this can be much more efficient than a 'brute-force'
solution in time.
3.4. Magnetic material models
One of the complications of magnetic field analysis is the behaviour of magnetic materials.
The relation between the field H and the flux density B can be not only nonlinear, but

1696

J P Webb

multivalued; the value of B in a material may depend not just on the present value of H ,
but on the past history of the material (figure 1l), a phenomenon known as hysteresis. There
is an extensive literature on the development of hysteresis models, but only recently have
any of these been incorporated into a finite-element code. Most often a Preisach model is
chosen. Ossart and Meunier (1990) compared several scalar models, and concluded that
the generalized Preisach model of Mayergoyz (1988). though complex, 'is the only one to
really fit the experimental behaviour of our material'. Philips and Delince (1993) used a
Preisach model in a 2D magnetostatic code, and obtained satisfactory prediction of BH loops
for an iron torus. Park and Hahn (1993) also used a Preisach model in a U) magnetostatic
code to predict the forces between two permanent magnets as they are brought together,
separated, and brought together again. Because of hysteresis in one magnet, the force versus
distance curve for the second approach is distinctly different from the same curve for the
first approach. The finite-element model successfully predicts the measured results.

Figure 11. Hysteresis in magnetic materials: a multivalued


BH relationship.

The scalar models used in these programs assume that the magnetization is parallel to the
magnetic field. There also exist vector Preisach models, which do not make this assumption.
Adly et nl (1993) analysed a magnetic recording process in 3D using an integral-equation
that incorporates a vector heisach model.
For magnetically-soft materials it is possible to neglect hysteresis and assume a singlevalued relationship between B and H (figure IZ),and this is commonly done in finiteelement analysis. To be used in a finite-element program, B has to he represented as a
continuous function of H (or vice versa), starting from the measured data. The construction
of suitable functions is not particularly simple (Lowther and Silvester 1985). Typical BH
curves have a very wide range of slopes-the ratio of highest to lowest slope may be
104. A relatively large amount of measured data is needed to construct the full relationship
between vector B and vector H . Furthermore, the functions must not only faithfully
represent the data, but are also subject to a number of other conditions. They must be
differentiable, and the Jacobian matrix aBj/aHj must be symmetric and positive-definite
(Haas and Schmoellebeck 1992). Silvester and Gupta (1991) suggested modelling the
coenergj density, instead of B, as a function of H . The flux density is then obtained
by differentiating the coenergy density with respect to H . It is simpler to model the scalar
coenergy than the vector B, and the resulting BH relationship is physically sound.
In the isotropic case it is only necessary to store a single B versus H curve (figure 12),
which represents the material in all directions. An alternative is to store the permeability as
a function of H z (or reluctivity as a function of Bz);
this tums out to be quite well suited

Application of the finite-element method

1697

BH curve
Figure 12 A typical singlevalued nonlinear
relationship for an isotropic material.

BH

Figure 13. The Newton-Raphson method used to


linearize a BH c w e .

to finiteelement algorithms. The curves are generally formed from piecewise polynomials
which impose slope continuity, e.g. cubic Hermite polynomials.
How are these material models incorporated' into the finite element analysis? In
section 3.3 it was shown how the magnetic field problem is reduced to a matrix equation
for an unknown column vector {U]:

[Sllul= { b ) .

(42)
When the BH relationship is linear. [SI is independent of [U]and the linear matrix equation
can be solved directly for (U] in a number of standard ways. A general approach to handling
the nonlinear case is by an iteration in which the BH relationship is repeatedly linearized,
i.e. the true relationship is replaced by a linear relationship that approximates it reasonably
well in the region of the latest guess for the solution. The algorithm is as follows.

(0)Make an initial guess for the state of the material (i.e. B and H ) .
(1) Build a linear BH model, i.e. B as a linear function of H ,that is accurate in the
neighbourhood of the latest guess for the state of the material.
(2) Assemble the matrix [SI using the linear model.
(3) Solve the linear problem [S]{U] = {&I.
(4) From (U], calculate H and improve the guess for the state of tthe material.
(5) If the state of the material has not changed significantly, stop. Otherwise, go back to (1).
Since in general the values of B and H change continuously through a material, different
linear BH models must be used in different finite elements-or even, for better accuracy, at
different points within a finite element.
One of the most effective ways of linearizing the BH relationship is the Newton-Raphson
method. This is demonstrated for the isotropic case in figure 13. Not only does the linear
function match the me BH curve exactly at one point (the latest guess), but also it is tangent
to the true curve at the point. The Newton-Raphson method converges quadratically in the
vicinity of the solution, i.e. the error at step k is proportional to the square of the error
at step k - 1. Convergence to acceptable accuracy in seven or eight steps is typical. For
vector-potential formulations (i.e. those requiring H as a functon of B ) the method is robust
and widely used. However, as explained by Neagoe and Ossart (1994), it is significantly
less stable for scalar-potential formulations (i.e. those requiring B as a function of H ) .

1698

J P Webb

Under-relaxation can be used to improve the stability: the new estimate for ( U ] is formed
from a linear combination of the estimate predicted by the Newton-Raphson method, and
the old estimate

(43)
where a is a number between 0 and 1. Nakata et al (1992) suggested a way of determining
a suitable value of a. Their algorithm works for problems where the plain Newton-Raphson
method fails to converge.
( U ]=a(u"R)]+(l

-cY)(u'"'d')

3.5. Forces and torques


Frequently what is required from a finite element analysis is not the magnetic field, but the
force or torque on a component. There are a number of ways of obtaining these quantities
from the computed field, all of which would give exact results if the exact fields were known.
However, when the fields are only known approximately, as they are from finite element
analysis, some of the force calculation methods perform better than others. Specifically,
some techniques have been found to give very inaccurate forces when there are only modest
errors in the field. For this reason, there is now a considerable literature on the subject of
force and torque calculation.
In a non-magnetic material, such as a copper winding, the force developed is the Lorentz
force of interaction between the current density and the flux density J x B.The integration
of this density over the volume of the material gives the net magnetic loading force. In a
similar way, the net torque can be evaluated (T x ( J x B)).
On the whole, this is a simple
and reliable method.
In a magnetic body, an additional force arises from the interaction of the magnetic
dipoles and the applied field. Although it is possible to express this in terms of a density that
can be integrated over the body to give a net loading force, this often does not give accurate
answers. Expressions for the density involve derivatives of the magnetization, which are
difficult to obtain accurately. A number of other techniques have been developed.
An established result in electromagnetics is that the net force on a body can be obtained
by integration over a surface S which passes through the air around the body, and completely
encloses it (see, for example, chapter IO of Panofsky and Philips (1962)). The expression
is
1
force =
((B n)B (44)
2
fio s
The vector n is a unit vector pointing perpendicularly outwards from S;
is the
permeability of free space. The vector integrand can be factored into n and a tensor,
called the Manoell stress tensor. The simplicity of this expression has made it a popular
force-calculation method. However, it has been found that the results obtained from it are
somewhat variable. They depend on the choice of surface S and in particular are likely
to be quite erroneous when S is too close to sharp corners and edges on the body, as it
sometimes must.
Another classical method that has been applied to finite-element solutions is virtual
work. When a magnetic loading force is allowed to move a body, there is a change in
the stored energy of the system which can be directly related to the force-at
least, in the
magnetostatic case. Specifically, if the currents are held constant during a movement of the
body in the x-direction, the component of force in that direction is just the rate of increase
of coenergy with x (Panofsky and Philips 1962). In magnetostatic finite element analysis,
the coenergy is likely to be known more accurately than any other quantity, because the

-f

Application of the finite-element method

1699

method in effect selects from the available trial fields according to coenergy; the solution
is the field whose coenergy is closest to the true coenergy.
Consequenuy, one way of calculating the force is to analyse the problem with the body
in two or more positions; obtain a system coenergy from each analysis; and use these to
estimate the derivative of the coenergy with respect to position, i.e. force. Torque can be
handled in a similar way by rotating the body and differentiating with respect to angle.
In general, the virtual work method leads to accurate results. The only drawbacks are
that it is not applicable to transient and time-harmonic problems, and requires at least two
finite-element analyses to find the force.
This last drawback prompted the development of a single-solution virtual work method,
described by Coulomb (1983). The rate of change of coenergy is obtained from a single
solution by considering an infinitesimal distortion of the finiteelement mesh in a layer of
air around the body. In fact, McFee er a1 (1988) showed this to be a special case of a more
general force formula

In this formula, V is not the volume of the body, but the volume of a layer of air surrounding
the body. The scalar function g must take the value 0 on the outer surface of the layer
(i.e. away from the body) and the value 1 on the inner surface of the layer, and must be
differentiable in between, but otherwise it is completely arbitrary. By appropriate choice
of V and g, both the Maxwell-stress and the single-solution virtual work methods can
be obtained. Moreover, g can be tuned so that it avoids regions where the field error is
estimated to be high, thus increasing the accuracy of the computed force. Note that the
formula is applicable to any type of magnetic field analysis: it gives the net force on the
body in the steady state, or instantaneously. An analogous formula exists for the net torque.
4. Electric fields

For certain problems the induction of electric fields by time varing magnetic flux is
negligible, and the term aE/at can be dropped from Maxwell's equations. This
considerably simplifies the solution. To begin with, even if electric currents flow and there
is an associated magnetic field, the magnetic field can have no influence on the electric
field without the aE/at term (or motion). The solution, therefore, is for the electric field
alone-this type of problem is known as electro-quasistatic. Further, the electric field under
this assumption is entirely curl-free, and can everywhere be written as -VQ, where Q is the
electric scalar potential. There is no need for complicated potentials of the kind developed
for magneto-quasistatics.
In fact, the simplest electric field problems are very simple indeed. Consider the problem
of finding the electrostatic field produced by a combination of a known stationary charge
distribution, p, and conducting electrodes held at known potentials. The governing equation
is Gauss' Law
V - E E = ~

(46)

which, on substitution E = -VQ, becomes the Poisson equation. Here E is the permittivity
tensor representing inhomogeneous anisotropic materials; p is a volume charge density,
assumed known. Solving (46) for @, subject to boundary conditions on the electrodes, is
straightforward-see section 2. The only complication is that often such problems have no
closed boundaries around them, and one of the methods described in section 2.4 must be

1700

J P Webb

used. The computed electric field may be of interest in itself, as in the design of high-voltage
components where breakdown is important, or merely an intermediate quantity from which
forces, torques, charges and capacitances are calculated. The discussion of force and torque
calculation in section 3.5 applies equally to electric fields.
As an example, consider the electrostatic micromotor shown in figure 14. The outer
diameter of the device is ahout 1 mm. The central (dark) region is the rotor, made of
conducting material. Surrounding it are a number of electrodes, constituting the stator. By
applying voltages to the stator electrodes, the rotor is made to turn. A 3D electrostatic
analysis was used to determine the torque developed for given applied voltages.

Figure 14.
micromotor.

An elecmstalic

radial-field

In general, charge in an electric field moves, and the charge density p is not known in
advance. Moving charge constitutes a current density, J , which is related to p through the
continuity equation (conservation of charge)
ap
v.J+ =o.
at

(47)

Since J is also related to the electric field via the conductivity ( J = oE),equations (46)
and (47) are coupled and must be solved simultaneously for Q and p. There are. four cases
to consider:
4.1. Conductivity constant: Ohmic materials
In Ohmic materials the current density is directly proportional to the electric field at each
point. In other words, the conductivity is a fixed known quantity, though it may vary from
point to point. If further, the currents and charges are steady, equations (46) and (47)
decouple. Equation (47) no longer involves the charge density and may be solved on its
own for Q, subject to boundary conditions on electrodes. Once Q is known, the electric
field, current flow and bulk resistance can be calculated ( p can also be found from (46), if
necessary). Finite element analysis of this sort is used to provide DC resistances of known
suuctures. It can also be used in the reverse process-the determination of conductivity
distribution in a volume from discrete measurements in the volume or on the surface (Daily
and Owen 1991).
In the time-harmonic case at angular frequency w , phason may be introduced for CD
and p. since equations (46) and (47) are linear. The equations can then be combined into

Application of the finite-element method

1701

one, complex, Laplace equation

v . (U + j W & ) V Q =o.
(48)
This equation governs the distribution of potential in a lossy dielectric. Castillo et a1
(1990) solved it by the finite-element method to find the per-unit-length conductance and
capacitance matrices of multiconductor printed-circuit transmission lines.
4.2. Conductivity dependent on electric field strength

When a dielectric breaks down, its conductivity increases sharply with increasing field
strength. The charge density can be eliminated from (46) and (47), and a single nonlinear
equation for @ obtained. Enokizono and Tsutsumi (1994) analysed transient dielectric
breakdown by solving this equation. Their analysis includes a probabilistic component
that accounts for the complete breakdown of the dielectric in a finite element, and predicts
fractal patterns of discharged (highly conducting) material stretching from one electrode to
the other.
4.3. Conductivity dependent on charge density

Current density depends not just on the strength of the electric field but also on the amount
of movable charge available. In some cases, the density of movable charge is the same as
the net charge density, p. The conductivity is, therefore, not known in advance but depends
on p. Usually it is assumed to be directly proportional to p . the constant of proportionality
being the mobility, p. Under these circumstances equation (47) is nonlinear, and the solution
is harder to find.
An example is the problem of finding the electric field between two high voltage
conductors when there is a DC corona discharge and ions move from one conductor to
the other. Jones and Davies (1992) solved this problem by applying finite elements to (46)
and (47), in an axisymmetric geometry.
In fact, charge can move in the absence of an electric field by the process of diesion,
from a region of high charge density to a region of low charge density. The component of
current density associated with diffusion takes the form -DVp, where D is the diffusion
coefficient. Whereas Jones and Davies neglected diffusion in their analysis of corona
discharge, Abdel-Salam and Al-Hamouz (1993) took it into account in their finite element
method for a similar problem.
In semiconductor devices, there are two charge caniers-electrons and h o l e s a n d the
current density of each is governed by a continuity equation similar to (47) but including
generation and recombination terms. As with the corona discharge, the conductivities for
electrons and holes are proportional to the densities of the respective charge carriers, so the
problem is nonlinear. Diffusion terms must also be included.
4.4. Conductivity not defined

There is an important class of devices in which a heated electron gun ejects electrons, which
are then accelerated, focused and deflected by electric and magnetic fields. Examples are
cathode ray tubes and spectrometers.
A charge in such a device does not reach a steady drift velocity in an applied electric
field, but continues to accelerate. It is therefore not possible to define a conductivity.
Instead, to find the charge distribution, individual charges must be tracked through the
electric and magnetic fields by integrating their equations of motion. Often the charges

1702

J P Webb

themselves contribute signficantly to the electric field-this is the so-called space-charge


effect. It is then necessary to solve the equations of motion and equation (46) alternately,
until the unknowns Q and p have converged. Girdinio er al (1994) described some of the
difficulties in implementing such a scheme. There is the problem of the emission from
the electron gun, which may be spacecharge limited. There is the computational cost of
tracking a sufficiently large number of particles at each iteration, a cost that can be reduced
by tracking relatively few carefully chosen particles and interpolating between them in phase.
space (Polak ef a/ 1994). Finally, there is the difficulty of assigning smooth charge densities
to the finite element mesh, based on the computed particle trajectories.
5. Waves

The majority of microwave problems involve structures which range in size from a tenth
of a wavelength to several tens of wavelengths. They are too large, electrically, to permit
a static or quasistatic analysis; on the other hand, they are too small for asymptotic (ray)
methods to work reliably. Over the past decade or so, finite-element methods have been
developed for this type of problem.
The usual assumptions are that the sources are time harmonic and the materials linear
and time invariant, so that a phasor analysis is possible. Generally there is no volume
current or charge densities, these sources being restricted to the surface of conductors.
Then Maxwell's equations for the complex electric and magnetic field are
VxE=-jwp.H
Vx=jw~.H.
(49)
Here w is the angular frequency and p , E are the permeability and permittivity, complex
tensors that can represent lossy, anisotropic materials. The solution is most often determined
via the fields E or H directly. However, sometimes potentials are used (section 5.1).
It is useful to distinguish two classes of problem. Fir% there are resonance problems
for which it is necessary to solve the above equations with only zero boundary conditions
and find the allowed values of w (section 5.2). Resonance plays a part in the operation
of many microwave components and its accurate prediction is very important. Further, the
modes of transmission lines and waveguides are also resonances. Even optical waveguides
can be analysed though most optical structures are hundreds or thousands of wavelengths
long, the cross section of a typical optical waveguide is of the order of a wavelength, and
therefore the modes of the waveguide may be found by finite-element analysis.
Secondly, there are driven, or deterministic, problems (section 5.3). A wave of some
sort is incident, either along a waveguide or through free space, and the response of the
system to this excitation is required. The finite-element method has been used to find the
scattering parameters of microwave components, and radar cross sections of conducting or
dielectric bodies in free space.

5.1. Direct solution and solution via potentials


Direct solution. Eliminating the magnetic field from (49) gives a partial differential
equation in E alone:

v x p;'v

x E - ~ , E , E= O
(50)
where ko is the free-space wavenumber, equal to w divided by the speed of light in a
vacuum. Perhaps the commonest approach to microwave problems has been to solve this
equation directly for E (or to solve the dual equation in H , obtained by eliminating the
electric field from (49)). The boundary conditions that may arise are:

Application of the finite-element method

1703

(a) at microwave frequencies the skin depth of a conductor can usually be taken as zero,
and so on the surface of a conductor the tangential part of E vanishes (the electric wall
condition); the same condition sometimes holds on a plane of symmetry;
(b) on a plane of excitation, e.g. the cross section of a waveguide, the tangential part of E
is constrained to a specified value;
(c) on a symmetry plane the tangential part of the magnetic field may vanish (the magnetic
wall condition)

0
where subscript tang denotes'the tangential part;
(d) generally, there may be a boundary condition of this kind:
Hcaog

H w g

=W E )

(51)

(52)

where P is a given function of the electric field over the surface.


Applying the weighted residual method to (50),and to the boundary conditions (51) and (52),
leads to

.V x E - kiw

-E,.

E ) d V + j u s w x P ( E ) .ds = O .

(53)

s,

Here w is a vector weighting function and S, is that part of the surface bounding the volume
52 on which the general boundary conditon (52) is applied. The boundary conditions (a)
and (b) must be imposed explicitly on the trial field E .
Konrad (1976) was the first to solve (50) using finite elements, in ZD; the 3D solution
appeared in 1978 (Ferrari and Maile). However, a difficulty that arises with such a
formulation is that the electric field is not necessarily continuous. At an abrupt dielectric
interface, the normal component of electric flux density is continuous, which means that
the corresponding component of electric field must be discontinuous. Since nodal finite
elements impose complete continuity of the field, they cannot be easilj. used to find the
electric field in the presence of dielectric interfaces (but see the comments of Yuan et a1
1991). If there are no magnetic materials, a common solution is to use the magnetic field
as the unknown instead, because it is continuous in all its components. However, a better
solution, and one that can handle both dielectric and magnetic materials simultaneously,
is to use edge elements (section 2.2). Edge elements impose tangential continuity on the
electric field; the normal continuity of electric flux density arises as a natural continuity
condition of the weighted residual method, i.e. it is satisfied approximately by the solution
to the weighted residual equations.
As will be explained below (section 5.2), with nodal elements additional measures have
to be taken to avoid the possibility of spurious modes or spurious corruptions. Edge elements
are immune from these. They are also better able to model the variations of field near sharp
conducting points and edges, about which the field changes direction infinitely fast-see,
for example, the discussion by Bardi et a1 (1994).
Despite the several advantages of edge elements, D'Angelo and Mayergoyz (1993) have
argued for the use of nodal elements in combination with an alternate equation for the electric
field. The altemate equation decouples in homogeneous regions into three scalar Helmholtz
equations-one for each Cartesian compontent of E-and thereby provides much greater
sparsity in the system matrix, leading to faster solution times. This approach requires that
normal continuity is imposed explicitly, so nodal elements are essential. Although offering
greater efficiency, it is not yet clear how the method will cope with sharp edges and material
interfaces.

'

1704

J P Webb

The two-dimensional case, in which there is negligible variation of the fields along the
z-axis of a Cartesian coordinate system, reduces to a scalar problem in terms of either EL
or Hz. Once Ez is found, the magnetic field components H, and H, can be derived Gom it.
Once Hz is found, the electric field components E, and E,. can be derived from it. When
both Ez and Hz exist, they are completely decoupled and can be found separately, i.e. there
is no need to solve for more than one variable at a time.
Solution viapotentials. The difficulty of solving directly for the electric or magnetic field
with nodal elements prompted a number of researchers to use potentials. The electric
field can be represented directly in terms of a magnetic vector potential A and an electric
scalar potential e,both of which are continuous across interfaces. The development of
such methods is analogous to their development in the quasistatic case, see section 3.1. It
is necessary to gauge the vector potential, i.e. to impose an extra condition on it to fix it
uniquely. Bardi etal (1992) chose the Coulomb gague, V-A = 0, and found the resonances
of dielectric-loaded cavities. They compared the method to an edgeelement method that
solves directly for E or H ,and concluded that 'almost the same or better approximation
can be obtained using edge elements whereas the number of degrees of freedom and the
number of non-zero elements in the matrices are lower', i.e. the edge element method is
better.
Boyse et al(1993) chose a Lorentz gauge, V . A = -jcvK&Q, and applied their method to
a range of deterministic problems involving inhomogeneous dielectrics. Bryant et a1 (1992)
used the same gauge to eliminate
from the equations, thereby reducing the number of
unknowns per node from four to three.
All of these potential methods are node-based, and while none of them suffers from
spurious modes or spurious corruptions, it is difficult to see how they could cope well with
sharp edges and points, where the vector potential must change direction infinitely fast.

5.2. Eigenvalue problems

Cavity resonance and spurious modes. To find the resonances of a 3D microwave cavity,
equation (50) has to be solved inside the cavity for both the electric field and the resonant
wavenumber ko. There are no excitations-each piece of the boundary is either a conductor
or a symmetry plane. Applying the finite-element method reduces the differential equation
to the matrix eigenvalue equation

[SlluI= ka2ITlb4

(54)

where [SIand [TIare square matrices and { U ] is the column vector of unknowns representing
the electric field. This matrix equation occurs commonly in numerical analysis and there
are a large number of methods for its solution, including many which successfully exploit
the sparsity of [SIand [TIto reduce the computation time (see, for example, Parlett 1980).
Each solution of (54) consists of an eigenvalue, k$ and an eigenvector, (U).
When nodal elements are used. not all of the solutions are physically correct. Some
are spurious modes, non-physical solutions whose eigenvalues lie in between the m e
eigenvalues. Refinement of the finite-element mesh generally reduces the errors in the
approximations to the true eigenvalues, but does nothing to remove the spurious modes,
which change their values but do not disappear. The presence of spurious modes makes
the method almost useless, because it is not easy to distinguish between true and spurious
modes. Further, if a finite-element method gives spurious modes in the eigenvalue problem,

Application of the fnite-element method

1705

the same method applied to a drive problem can give solutions that are affected by spurious
corruptionr.
There has been a great deal of publication on the subject of spurious modes and the
reason for them is now quite well understood. Equation (50) has, in addition to its physical
solutions at non-zero frequencies, an infinite number of solutions at ko = 0-what might
be called static solutions. These are not physical solutions to Maxwells equations because
they, are not divergence-free. When (50) is solved with nodal elements, approximations
to some of these static solutions appear in the computed spectrum. Because they are
only approximations, their frequencies are non-zero, i.e. they appear as spurious modes.
Refinement does not help, because for every static mode which is well enough approximated
to have a computed frequency close to zero, and therefore out of harms way, there are
infinitely many more waiting in the wings to be badly approximated, even by the better
refinement.
There are two basic approaches to the problem of eliminating spurious modes. First,
impose the condition V E ~ =E 0 on the solutions. Doing this exactly makes the static
solutions physical, and reduces their number to a very few, perhaps none. Rahman
and Davies (1984) used a penalty method to impose the divergence-free condition in an
approximate way-the term s V V &J3 is added to the differential equation (50). As the
penalty parameter s is increased, the spurious modes increase in frequency and can be
pushed up out of the range of interest. Unfortunately, there is no reliable way of fixing s
in advance, and increasing it too much will degrade.the accuracy of the physical solutions.
An alternative is to obtain a divergence-free basis by a preliminary finite-element analysis,
and then use the divergence-free basis to solve (50) (Kobelansky and Webb 1986). This
has the disadvantage of destroying the sparsity of the system matrix.
The second way of eliminating spurious modes is to use a different kind of finite element
(Bossavit 1990). Most edge elements have the property that they provide a significantly
better modelling of the static solutions than the same number of nodal elements-so much so
that all the spurious modes have frequencies that are exactly zero, and cause no trouble.
Wang and Ida (1991) used Whitney tetrahedra, and also various higher order tetrahedra
(1993); Webb and Miniowitz (1991) used a curvilinear brick with 54 degrees of freedom;
Bardi et al (1992) used a different curvilinear brick with 36 degrees of freedom-in all
cases spurious modes occur only at ko = 0. The common feature of these elements is that
they are edge elements, i.e. they impose only the tangential continuity of the electric field,
allowing the normal component to be discontinuous. The static solutions are curl-free, and
can therefore be written as gradients of a scalar potential. Such gradients are necessarily
tangentially continuous, but may not be normally continuous. If normal continuity is
imposed, it considerably reduces the number of gradients, and thereby reduces the ability
to model static solutions. By allowing the normal component to be discontinuous, edge
elements provide a relatively large number of gradients and a better modelling of static
solutions.
However, not all edge elements eliminate spurious modes. Crowley eta! (1988) showed
that when using quadrilateral edge elements that are fully second-order in each direction,
spurious corruptions can arise in driven problems. On the other hand, by mixing the orders,
so that the electric field is represented with first-order polynomials in the direction of the
vector and second-order polynomials in the orthogonal direction, the spurious corruptions
(and spurious modes) disappear. Further investigations by Dillon et al (1994) support
the idea that the mixed-order property is essential in quadrilateral elements. There is no
corresponding requirement for triangles.

1706

J P Webb

Modes of uniform waveguides. A mode. of a uniform waveguide or transmission line is


defined as a phasor solution to Maxwells equations that varies as exp(-yz) where the
z-axis is the axis of the waveguide and y is a complex propagation constant. Application
of the finite-element method leads to an algebraic eigenvalue problem, with either y or
as the eigenvalue.
Since a uniform waveguide has an unchanging geometry in the :-direction, and the
variation with z is prescribed, the problem is two dimensional. It is only necessary to
analyse a cross section of the waveguide, which is subdivided into triangular or quadrilateral
elements.
One might expect that waveguide analysis is simple, being just a special two-dimensional
case of the cavity resonance problem discussed in the previous section. However, a large
number of methods for solving the waveguide problem have been invented, and in fact the
topic is by no means simple. A recent review article by Dillon and Webb (1994) contains
considerably more details than space permits here.
For most waveguides it is possible to formulate the problem in terms of the two axial
components, E, and H,. Unfortunately, this formulation gives rise to spurious modes,
which so far nobody has found a way to eliminate. If the waveguide is empty (air-filled),
the two components decouple and the analysis uses H, to get the TE modes, and EL to get
the TM modes. This is a satisfactory and stable approach, which gives no spurious modes.
Unfortunately, many waveguides of interest are not empty.
Konrad (1976) used the full vector electric field and solved (SO) with nodal triangular
elements, assuming a known y . Although this too generates spurious modes, it is possible
to apply any of the remedies discussed above for the cavity resonance problem. The penalty
method is widely used, and so too are edge elements. A remedy that is not possible in 3D
was proposed by Hayata et al (1986): use the equations V c E = 0 to eliminate ELvariables from the algebraic equation (54). Not only does this impose the divergence-free
condition and get rid of spurious modes, but it also reduces the number of variables in the
problem. Unfortunately, the resulting matrices are dense and the solution is consequently
more expensive.
It is better to impose the divergence condition earlier, at the level of the differential
equations. In the waveguide case the vector equation (SO) can be separated into two
equations, one for the transverse directions and one for the axial direction. Femandez
and Lu (1990) threw away the axial equation and used V &,E = 0 to eliminate E,
from the transverse equation. The result is a differential equation for just the transverse
components, and the global matrices from this are sparse. The only drawback is that the
method requires the use of nodal elements (i.e. full field continuity) and therefore does not
work well in the presence of sharp conducting edges. However, it is an excellent method for
optical problems, in which there are no conductors. An additional feature of the method is
that not only does it permit ko to be found for a given y , but also, if ko is specified, it leads
to an eigenproblem linear in y 2 . i.e. like (54) but k i replaced by y 2 . This is particularly
important for the analysis of lossy waveguides, for which y is complex. In such cases
fixing y will lead to a complex (and therefore unrealistic) frequency.
If ko is specified in (50) and y left as an unknown, the resulting eigenproblem is
quadratic in y and significantly more expensive to solve than the linear eigenproblem (54).
However, Lee et al (1991) showed that by changing the variable from Et to E; = y E t ,
where the subscript t denotes the transverse part of the vector, an eigenproblem linear in y 2
can be obtained. Provided edge elements are used there are no spurious modes and sharp
edges can be handled. The matrices [SIand [TIare sparse, but indefinite. A sparse method
for solving the algebraic problem is presented. The application to lossy waveguides was

Application of the finite-element method

1707

described by Lee (1994). Those who prefer smaller, denser matrices can try the variant
of Koshiba and Inoue (1992). in which an eigenproblem linear in y z is obtained by using
the axial part of (50) to eliminate EL from the transverse part. Again, edge elements are
necessary.
S.3. Deterministic problems
Scatrering parameters. A microwave component can be viewed as an N-port device, i.e.
a junction between N uniform waveguides (or transmission lines) (figure 15). The finiteelement method can be used to solve (50) for a given frequency and with given excitations,
and hence to find the electromagnetic fields in the junction, but the fields are often only of
secondary interest. What is most often needed is a characterization of the device as seen
from the N ports, so that the device can be used as a black box in a circuit-based analysis of
the larger system containing it. How can this characterization be obtained by finite-element
analysis?

Port 1
I

/-

Commonly, at the frequency of operation each waveguide supports a single propagating


mode, its dominant mode. Let e and h be the transverse elechic and magnetic fields of
the dominant mode in a certain waveguide, when unit power flows along the guide in one
direction and there is no wave in the reverse direction. These modal fields form the basis
for a definition of a generalized voltage and current. Suppose the total transverse electric
field at the port to which the waveguide is connected is Et,
and suppose further that at the
port only the dominant mode is present, i.e. other modes that might be excited by interaction
with the junction are evanescent and have decayed to a negligible amplitude at the port.
Then E, is proportional to the modal field e, and the ratio of the two is by definition just
the generalized voltage, V. Similarly, if Ht is the transverse magnetic field at the port, the
generalized current is defined to be the ratio of this to h.
Voltage and current defined in this way apply to any type of waveguide or transmission
line, even when the more natural low frequency definitions of voltage (path integral of
elechic field) and current (a Row of charge) do not work because they are ambiguous.
Now it is possible to characterize a microwave N-port device by a complex, N x N
admittance matrix [ Y ] that relates an N-vector of port currents to an N-vector of port
voltages: ( I ) = [ Y ] {V ) . To compute the matrix [ Y ] by the finite-element method, it is only
necessary to simulate what might be done experimentally. Unit voltage is applied to port j ,
all the other ports are short circuited and the current at each port is determined. Under these

1708

J P Webb

circumstances, the current at port i is just Yi,. Using the finite-element method, applying the
voltages at the ports is simple becasue it amounts to imposing a known tangential electric
field, i.e. boundary condition (b) in section 5.1. The electric field inside the device is then
found by solving (50). subject to the boundary conditions. Finally, the magnetic field H, at
the ports can be calculated from the electric field, and from H I the currents at the ports can
be found. Note that to obtain the full N x N adminance matrix requires N field solutions,
though this number can sometimes be reduced by making use of symmetry.
Webb and Parihar (1986) described a better way of extracting the port currents from the
electric field. When port j is excited, the current at port i can be found by integrating over
the volume of the problem an expression involving the electric field due to the excitation
of port j and the electric field due to the excitation of port i. The integral is related to a
variational formulation of the problem and is a stationary with respect to small perturbations
about the true fields.
Since it is easier at microwave frequencies to memure incident and reflected waves
than it is to measure voltage and current, a more usual characterization of a device is the
scattering matrix, [SI.The voltage at each port is the sum of conhibutions from the wave
entering the device and from the wave leaving the device. The two contributions are called
the wave amplitudes, a and b, respectively, and the scattering matrix relates N-vectors of
these quantities: ( b ) = [S][n}.There is a simple relationship between the two matrices
[Y] and [SI, so that once [Y] is known [SI follows trivially. Alternatively, it is possible
to excite the finite element problem with an incident wave at port j only and calculate the
outgoing waves at port i , thereby obtaining Sij directly. The transverse magnetic field at a
port where the incident wave amplitude is a is related to the transverse electric field by

H, = 2a -

(J,E, .edS)/(Lp

e . e dS)

where S, is the area of the port. This is a boundary condition of the type given generally
by (52), which can readily be incorporated into the finite-element method. There are a
number of variants of this general approach (e.g. Koshiba and Suzuki 1986). One of the
commoner ideas is to include in the relationship (55) not just the dominant mode, but also
several higher modes. If this is done, the ports can be placed close to the junction, i.e. there
is no longer a need to model a length of uniform waveguide to ensure that the evanescent
modes have a negligible amplitude at the ports. The trade-off is a loss of sparsity, since (55)
couples together all the degrees of freedom on the port and creates a dense submatrix of
the global system matrix.
Lawrence and Cambrell (1993) discussed and compared various ways of extracting
scattering parameters from 3D field solutions. Examples of 3D N-port devices that have
been analysed with finite elements are:

(i) a coax to waveguide connector and a microstrip low-pass filter (Lee 1990);
(ii) dielectric obstacles in rectangular waveguide (Ise er a1 1991);
(iii) dielectric resonator filters (Cousty et ol 1992); and
(iv) inductive strips in finlines (Foo and Silvester 1993).
The coax to waveguide connector of Lee (1990) is shown in figure 16. Ideally, such a device
should reflect none of the power incident on it along the coaxial line-all of it should be
transferred to the waveguide. A 3D finiteelement analsysis confirmed that, for the design
shown, the reflection is negligible.

Application of the finite-element method

1709

Figure 16. A coax to waveguide connector.

Free-space radiation problems. An antenna or other source of electromagnetic radiation


generates a field that extends, theoretically, to infinity. To predict the field around a radiator
with the finite-element method, some means must be found of dealing with the unbounded
nature of the problem. A general solution is to define a finite volume containing the radiator
and fill that with finite elements; on the boundary of the volume, a boundary condition of
the form (52) is imposed. Ideally, the function P should express the exact relationship
between the electric and magnetic field on the boundary, a relationship that expresses the
fact that outside the volume is nothing but infinite free space, and that any waves incident
on the boundary from within are simply transmitted without reflection. Practically, P is
determined from some numerical approximation, and a number of different choices are
available: see section 2.4.
Free-space scattering problems. When a known electromagnetic wave is incident on a
dielectric or metallic structure in free space, a scattered wave is produced. The finiteelement method can be used to predict the field around the scatterer.
The problem is, in fact, very similar to that of a radiator in free space. The scatterer is
enclosed in a volume that is filled with finite elements, and on the surface of the volume
an appropriate boundary condition is imposed, of the form (52). Let us denote by P, the
function P that was used for the radiator. Clearly, P, is not a correct P for the scatterer
because it does not take into account the source that exists outside the volume and that
generates the incident wave. However, P, does express the relation between the scattered
electric and magnetic fields. the scattered field is just the difference between the total field
and the field of the incident wave, and it is generated by sources confined to the scatterer.
Making use of this, we can derive a relationship between the total magnetic and electric
fields on the surface

Hmg= P,(E) + H& - P,(E'")

(56)
where superscript inc denotes the incident field, which is known. With this boundary
condition, the electric field around the scatterer can be computed directly.
An alternative is to reformulate the problem in terms of the scattered field
V x p;'V x EJcl- k,Z&,ESc'= f(EinC)
(57)
where superscript scat denotes the scattered field, and f is a known function (which vanishes
outside the scatterer). The finite-element method can be used to solve (57) with the same
boundary condition on the outer boundary as would be used in a radiation problem.

J P Webb

1710

Yuan (1990) analysed scattering from a dielectric sphere with a hybrid approach
(finite elements and boundary elements). The hybrid approach was used also by Jin
and Volakis (1991) to predict scattering and radiation from microstrip patch antennas.
Chatterjee er al (1993) found the radar cross section of a variety of 3D dielectric and metallic
objects, using absorbing boundary conditions. Clegg et a1 (1994) and Kanellopoulos and
Webb (1995) also used absorbing boundary conditions to find the fields in and around lossy
dielectric spheres illuminated by a plane wave.
One interesting class of problem is scattering from a periodic structure such as an optical
grating or a helical antenna, in which a basic cell is repeated many times. It is often sufficient
to assume an infinitely repeating structure; in this case the analysis domain can be reduced
to a single cell by the application of binary constraints on opposite surfaces of the cell, i.e.
equations that relate the value of the tangential field on one surface to the tangential field on
the opposite surface. By Floquet's theorem, the fields on opposite surfaces are related by a
constant phase term, determined by the wavenumber of the incident wave. A 3D analysis
of infinite non-orthogonal doubly periodic structures is reported by Mias et a1 1994.
6. Conclusion

As this survey has tried to show, it is not always easy to develop robust and efficient finiteelement methods for the solution of Maxwell's equations. Yet finite-element analysis is now
routinely used in every branch of electromagnetics. As computers become more powerful
and software more intelligent, it is likely that finite-element solvers will become standard
tools for all who work with electric and magnetic fields.
Acknowledgment
The author is grateful to Professor Belmans and his group at the Katholieke Universiteit
Leuven, Belgium, for figures 6 and 14; and to Professor Lee for figure 16.

References
Abdel-Salam M and AI-Hamouz Z 1993 J. Phys. 0:Appl. Phys. 26 2202-1 I
AdlV A A. Mavernovz
- . I D.. Gomez R D and Burke E R 1993 IEEE Tranr. Maan.
- &I-292380-2
Ahmed S 1968 Elecrron. Lclr. 4 387-9
Alb-c
R, Coxorese E, Martone R, Miano G and Rubinxci G 1992 IEEE Trans. Maan.
- M.28 1118-21
Albane.se R and Rubinacci G 1988 IEEE Trans. Magn. M-24 98-101.
-1990
Ini. J. Num. Mer. Engng. 29 515-32
-1992
IEEE Trons. Magn. M-28 1220-31
Bardi I, Bir6 0, Dyuij-Edlinger R. Preis K and Richter K R 1994 IEEE Trans. Magn. M.30 3108-1 I
Bardi I, B i d 0, Preis K. Vrisk G and Richter K R 1992 IEEE Trans. Magn. M-28 I 1 4 2 4 5
Bayliss A, Gunzbvrger M and Turkel E 1982 SLAM J. Appl. Mark 42 430-51
Biddlecombe C S,Heighway E A. Simkin J and Trowbridge C W 1982 IEEE Trons. Magn. M-18492-7
BIr6 0 and Preis K 1989 IEEE Trans. hlagn M-253145-59.
Bir6 0, Preis K, Renhm W.Vrisk G and Richter K R 1993 lEEE Trans. Magn. M-291325-8
Bossavit A 1990 IEEE Trans. Mogn. M-26702-5
Bosravit A and V6rit6 J C 1982 IEEE Trans. Mogn. M-18431-5
Boyse W E. Minerbo G N. Paulsen K D and Lynch D R 1993 IEEE Tmm. hfagn. 01-29 133%
BNssino G and Sonnad V 1989 Inf. 1. Numerical MrrhodF Eng. 28 801-15
Bryant C F, Dillon B, Si&n I and Trowbridge C W 1992 IEEE Trans. Mogn 28 1182-5
Bryant C F,Emon C R I and Trowbridge C W 1990 IEEE Tram. Mag". M-262373-5
Castillo S, Panfic Z and M i m R 1990 Ini. J, Numer. Melhods Eng. 29 103347
Chai M K V 1973 IEEE Trans. Poner Appar. Sysr. PAS-92 62-72
~

Application of the finite-element method

1711

Chatterjee A, lin J M and Volakis J L 1993 IEEE Trans. Antennas Propog. AP.41
Chatterjee A and Volakis J L 1993 Microwave Opt. Tech Len. 6 8 8 6 9
Clegg S T. Murphy K A, loines W T. Rine G and Samulsld T V 1994 IEEE Tranr. Miemwove Theory Techniques
MTT-42 1984-91
Coulomb J L 1983 IEEE Trans. Mag". M-192514-9
Cowsty J-P. Verdeyme S, Aubourg M and Guillon P 1992 IEEE Trans. Microwave Theory Techniques MTT-40
925-32
Crowley C W, Silvester P P. Hurwie H 1988 IEEE Trans. Magn. M-24 3974M)
Daily W and Owen E 1992 Ceosci. Conado 56 1228-35
Davies J B, Femandez F A and Philippou G Y 1982 lEEE Trans. Micmwove Theory Teclniquer hlTT.30 191540
DiUon B M and Webb J P 1994 IEEE Trans. Microwave Theory Techniqlres 24 308-16
Dillon B M Liu P T S and Webb J P 1994 COMPEL Inf J. Comput. Math Elec. Electron Eng. A 13 311-5
Duff I S. Erisman A M and Reid I K 1986 Direct Methods fop Spnrse Matrices (Oxford Clarendon)
Emson C R I and Simkin J 1983 IEEE Trans. Magn. M-19 2450-2
Enokizono M and Tsutsumui H 1994 IEEE Trans. Mngn M-30 2936-9
Femandez F A and Lu Y 1990 Electron. Len. 26 2 1 5 - 6
Ferrari and Maile 1978 Elemon. L e n
Fletcher R 1976 Pme. Dundee CmJ on Numerical Analysis 1975 (Berlin: Springer)
Foo S L and Silvester P P 1993 IEEE Trans. Micmwave Theov Technique3 41 298-304
Freeman E M and Lowther D A 1988 IEEE Trans. Mogn. M-24 19344
Girdinio P, Regetto M and Simkin J 1994 IEEE Trans. Magn. M-30 932-35
Gyimesi M. Lavers D. Pawlak T and Ostergaard D 1993 IEEE Trm. Magn. M-29 1345-1
Haas H and Schmoellebeck F 1992 IEEE Trms. Magn. M-28 1255-8
Hayata. K. Koshiba M, Eguchi M and Suzuki M 1986 IEEE Trans. Microwave Theoq Techniques MTT-34 1120-24
Hoole S R H 1989 Computer-aided Anolysis M d Design of Electromagnetic Devices (New York: Elsevier)
Ise K. lnoue K and Koshiba M 1991 IEEE Tram. Microwave Theory Techniques MTT-39 1289-95
Jacobs D A H 1986 IMA J, Numer. Ami. 6 441-52
Jin J 1993 The Finite Element Method in Eiectromagnelics (New York: Wiley)
Jin J and Volakis J L 1991 IEEE Trans. Antennas Propag. AP-39 1598-504
Jones E and Davies M 1992 J. Phys. D:Appi. Phys. 25 1749-59
Kamexi A 1990 IEEE Tram. Magn. M.26 466-69
Kanellopoulos V and Webb J P 1995 IEEE Trans. Microwave Theory Techniquer MTT-43
Kershaw D S 1918 J. Compur Phys. 26 43-65
Kobelansky A J and Webb J P 1986 Electron. Lett. 22 569-70
Konrad A 1976 IEEE Tram, Microwowe Theov Techniques MTT.24 553-9
Koshiba M and Suzuki M 1986 IEEE Trans. Microwave Theory Techniques 34 103-9
Koshiba M and lnoue K 1992 IEEE Trans. Micmwove Theory Techniques 40 371-7
Lawrence B G and Cambrell G K 1993 Radio Sei. 28 1195-202
Lee 1 F 1994 IEEEE Tmns. Microwave Theory Techniques MTT.42 1025-31
Lee J F, Sun D K and Cendes Z J 1991 IEEE Trans. Microwave Theory Techniques MTT-39 1262-71
Lee J F 1990 Inr. J. Numer. Model 3 235-46
Li H, Saigal S, Ali A and Pawlak T P 1994 Inr J. Num. M e t Eng. 37 ?4>56
Lowther D and Silvester P 1985 CAD in Magnetics (Berlin. Springer)
Lynch D R, Paulsen K D and Suohbehn J W 1986 h t . I,Nvm. Met. Eng. 37 343-56
Mayergoyz I D and Friedman G 1988 IEEE Trans. Magn. M-24212-7
Mayergoyz I D, Chxi M V K and D'Angelo 1 1981 IEEE Trans. Mngn. M-23 3889-94
McFee S, Webb J P and Lawther D A 1988 IEEE Trans. Magn. M-24 439-42
Mei K K 1914 IEEE Trans. AntenMS Propag. AP-22 1 6 M
Mei K K, Pous R, Chan Z and Liu Y W 1992 Proc. IEEE Antennas Propag. Soc. Intl. Symp. (Chicago) pp 2047-50
Melissen 1 B M and Simkin I 1990 IEEE Trans. Magn. M.26 3914
Mias C. Webb I P and Ferrari R L 1994 COMPEL h t . J. Comput. Moth. Elm. Eleermn. Eng. A U 393-8
Mittra R, Ramahi 0.Khebir A. Gordon R and KouM A 1989 IEEE Trans. Magn. M-25 3034-9
Morgan M A, Chen C H, Hill S C and Barber P W 1984 Wave Morion 6 91-103
Morgan M A and Mei K K 1979 IEEE Trans. Antennas Pmpag. AP-2 202-14
Nakata T, Takahashi N, Fujiwara F. Okamoto N and Muramtsu K 1992 IEEE Trans. Mogn. M-28 1048-51
Neagoe C and Ossan F 1994 IEEE Trans Magn. M-30 2865-8
Nedelec J C 1980 Numer. M a t h 35 31541
Ossan F and Meunier G 1990 IEEE Trans Magn. M-26 2837-9

1712

J P Webb

Panofsky W K H and Philips M 1962 Electmmogneiirm 2nd edn (Reading, MA: Addison-Wesley)
Park G-Sand Hahn S-Y 1993 IEEE Trans. M a p . M-29 1542-5
Parlen B N 1980 The Symmetric Eigenwlue Pmblem (Eoglewood Cliffs, NI: Prentice-Hall)
Peterson A F 1988 Microwave and Optical Technolog).Len 1 62-54
Philips D A and Delinc6 F 1993 IEEE Trans. Magn M-29 2383-85
Pol& S. Repetto M and Sabbi G L 1994 IEEE Trans, Mug& M-302928-31
F'reis K. Bardi 1. Bir6 0. Magele C. Vrisk G and Richter K R 1992 IEEE Trans. Magn. M-28 1056-9
Preis K, Badi I, Bir6 0, Magele C. Renkur W, Richter K R and Vrisk G 1991 IEEE Trans. Mogn. M-273798-803
Rahman B M A and Davies 1 B 1984 IEEE Trans. Microu'aw Theory Techniques MTT-32 922-28
Salon S 1 and D'Angelo I 1988 IEEE Trans. Mogn. M-2A 80-5
Silvester P P 1969 IN. J. fig. Sci. 9 849-61
Silvester P P and Chari M K V 1970 IEEE Trans. Power Apparatus and System PAS-89 1642-51
Silvesler P P and Fermi R L 1990 Finite Elemenis for Electrical Engineers (Cambridge: Cambridge University
Press)
Sivester P P and Gupta R P 1991 IEEE Trans. Magn. M-27 3804-7
Simkin J and Trowbridge C W 1979 Int. 3. Num Met. Ens, 14 4 2 3 4 0
Sreele C W 1987 Numerical Computation of Electric ami Magnetic Eel& (New York Van N w m d )
Stochniol A 1992 IEEE Tram. Magn. M-28 1679-81
Towers M S. McCowen A and hlncnab J A R 1993 Tram Antemm Propag, AP41 770-1
Tsukerman 1 A, Kondrad A, Meunier G and Sabonnadihre J C 1993 IEEE Tmm. Magn. M-29 1701-4
'Tsukerman I A, KO&
A. Bedrosian G and Chari M V K 1993 IEEE Tram. Mogn M-20 1711-6
Wang J and Ida N 1991 IEEE Tram. Mogn. M.27 3978-81
Wang J S and Ida N 1993 IEEE Trans. Magn. M-2914914
Webb J P 1993 IEEE Trans, Magn. M-29 1460-5
Webb J P and Forghani B 1993 IEEE Trans. Mags M.29 2461-3
-1989
IEEE Tmnr. Mag". M-25 4126-8
Webb J P and Kanellopoulos V N 1989 Microwave and Optical Technology Lett. 2 370-2
Webb J P and Miniowitz R 1991 Trans. Micmwwe Theory Techniques 39 1895-9
Webb J P and M a r S 1986 IEE Pme. H 133 9 1 4
Wright D B and Cangellaris A C 1994 Radio Sci. 29 907-21
Yuan X 1990 IEEE Tmm, Micmwove Theory Techniques MTT-38 10534
Yuan X. Lynch D R and Paulsen K 1991 IEEE Tmrans. Microwove Theory Techniques MTF-39 638-41
Zienkiewicz 0 C, Bando K, Beness P. Emson C and Chiam T C 1985 Int. J, Num. Me!. Eng. 21 1229-5 I

Anda mungkin juga menyukai