You are on page 1of 10

Frequency-Domain Finite Element Methods

for Electromagnetic Field Simulation:


Fundamentals, State of the Art,
and Applications to EMI/EMC Analysis
Andreas C. Cangellaris
Center for Electronic PackagingResearch,ECE Department
University of Arizona, Tucson, AZ 85721,U.S.A.
Abstract-This paper provides a critical review of ature has been generated on the application of the finite
frequecy-domain finite element methods and their ap- element method to a variety of electromagnetic scattering
plications to the modeling of electromagnetic interac- and radiation problems. The book by J.M. Jin [l] serves
tions in complex electronic components and systems. both as a tutorial on the procedures for the application of
Emphasis is placed on latest advances in finite element the finite element method to the approximation and solu-
grid generation practices, element interpolation func- tion of electromagnetic boundary value problems, and as
tion selection, and robust, highly absorbing numerical a rather thorough survey of the classesof problems that
grid truncation techniques for modeling electromag- have been tackled. Considering the power of the afore-
netic interactions in unbounded domains. These ad- mentioned attributes, one would have expected that the
vances have helped enhance the robustness and accu- method of finite elements would have gained in popularity
racy of the method. Finally, the advantages of domain among EMC/EMI engineers and would have established
decomposition techniques for the modeling of com- itself as the method of choice in the analysis and predic-
plex geometries are examined. Such domain decom- tion of EM1 and the design of electromagnetically compat-
position techniques are expected to play an important ible systems. Nevertheless, a literature review indicates
role in the continuing effort to extend the applica- that this is not the case. As an example we mention that
tions of frequency-domain finite methods beyond the in a special issue of the IEEE Transactions on Electromag-
subcomponent-level to component and system model- netic Compatibility, dedicated to computational methods
ing for electromagnetic interference and electromag- for EMI/EMC analysis, very few papers on finite elements
netic compatibility analysis and design. appeared, and the applications presented where limited to
rather simple problems of low complexity [2]-[3].
I. INTRODUCTION Before one attempts to search for drawbacks in the
method of finite elements that have prevented its prolifer-
There are two attributes of the method of finite ele- ation as an EMI/EMC prediction tool, one has to keep in
ments that have prompted the rapid growth of its appli- mind that application of electromagnetic CAD for compo-
cation to the modeling of electromagnetic interactions in nent and system EMI/EMC is actually still at its infancy.
electronic systems. One of them is its superior modeling The reason for this is that the complexity of an integrated
versatility where structures of arbitrary shape and compo- electronic component, subsystem or system is such that
sition can be modelled as precisely as the desirable model accurate modeling of source, coupling mechanism, and re-
complexity and available computer resources dictate. The ceiver of electroma.gneticnoise is almost prohibitive using
second, is common to all differential equation-based nu- a single numerical method for solving Maxwells equa-
merical methods, and has to do with the fact that the tions, irrespective of the type of the method used. More
matrix resulting from the discretization of the governing specifically, considering the tremendous variation in fea-
equations is very sparse, which implies savings in com- ture size from chip, to package, to board, to cables, to
puter memory for its storage as well as in CPU time for shields, it becomes clear that the number of elements re-
its inversion. Clearly, these two attributes come at the quired for the discretization of such a system for finite
expense of an increase in the degrees of freedom used in element analysis of electromagnetic interactions is out of
the numerical approximation of the problem since now, the reach of todays most powerful supercomputers.
contrary to integral equation methods, the entire space In view of the above and recognizing that an elec-
surrounding all sources of electromagnetic fields needs be tromagnetic analysis tool will be useful as a CAD tool
incorporated in the numerical model. Nevertheless, be- only if simulation times are in the order of minutes or
cause of the sparsity of the resulting matrix and the sim- at most a few hours, this paper examines the latest ad-
plicity with which complex geometries can be modeled, vances in the method of finite elements that are expected
this increase in the degreesof freedom of the approxima- to help the method establish itself as a reliable candidate
tion is an acceptable penalty. for EMI/EMC problem solving either at the component
Over the past ten years, a significant volume of liter- level or in conjunction with reduced-order electromagnetic
107
O-7803-3207-5/96/$5,00 0 1996lEEE
models of subsystems. As a matter of fact, it is this area of scalar basis functions, &,
where the method of finite elements can have an impor-
tant impact. Indeed, current practices of EMI/EMC anal- E=CEidi, (2)
ysis concentrate on rather simplistic, individual source-to-
victim models, which often suffer from their inability to where Ei denotes the unknown vector field value at node
capture the impact of surrounding conducting, dielectric, i. The relevant weak form, in the spirit of Galerkins
and magnetic material topology on the electromagnetic approximation, is
interaction. The finite element method allows for the de-
velopment of a more precise model that will lead to higher
accuracy in noise prediction and thus facilitate the design
of electromagnetically compatible electronic modules.
(( &V X E) x V&)

- 1
+ (j&E&) =

Finally, the potential of domain decomposition meth- -ii x (V x E)gS& (3)


f jwp
ods for reducing the complexity of the original problem
will be examined. The basic idea behind such methods where ( ) and $ indicate integration over the domain of
is the partitioning of the domain of interest into smaller interest and its boundary, respectively, while ii is the out-
ones and the development of the solution in a piecewise ward unit normal on the boundary. For edge element
manner, one subdomain at a time, using different types expansions, vector basis functions, Ni, are used for the
of both numerical and analytic techniques. The inherent expansion of the field,
parallelism of such approaches combined with the smaller
size of the subdomains makes them extremely well-suited E = C EiNiy
for massively-parallel computation.
where Ei are the unknown coefficients in the expansion.
The relevant weak form is
II. MATHEMATICAL FRAMEWORK FOR FINITE
ELEMENT ANALYSIS

The focus of this paper is on the numerical approxi-


(( LO
3WP
X E) * (V X Ni))
1
+ (j&E * Ni) =

mations of Maxwells equations with time-harmonic field -G X (V X E) . Nids (5)


-f jwp
variation. Therefore, the following discussion pertains to
linear sources and materials. However, time-domain finite For two-dimensional problems, a scalar version of (3)
methods that can handle transient electromagnetic inter- is readily obtained. More specifically, for a transverse
actions in the presence of nonlinear sources and nonlinear magnetic to z (TM,) polarization, the fields, E = bE,
media are possible and are currently the topic of vigorous H = kH, + 9Hy, are independent of z and (3) reduces to
research within the computational electromagnetics com-
munity. As a matter of fact, the finite-element formula-
tion in [2] is such that both transient and time-harmonic
electromagnetic simulations can be effected within a sin-
(C &Vx,,> . Vxyh) + (jwtEq$) =

gle mathematical framework. f jwp


(6. VE)(bjdZ (6)
In order to review the basic steps involved in the fi-
where VW = S/&Z + ?a/ay. For transverse electric to
nite element approximation of electromagnetic boundary-
z (TE,) polarization of the two-dimensional fields, where
value problems, let us consider the double-curl equation
H = i?H, E = S?, + QE,, the weak form is easily found
for the electric field, E, which, in a source-free, isotropic from (6) by duality.
and linear medium with position-dependent magnetic and For static problems (w = 0), a scalar potential, @, is
electric properties has the form often introduced, and the electric or magnetic fields are
obtained in terms of the gradient of the potential. For
Vx (&VxE)+jwiE=O, such cases, it is a weak statement of Laplaces equation
for the scalar potential that is being derived.
The most well-known attribute of Galerkins method,
The time dependence exp(jwt) is assumed (j = G), where the solution is sought in a finite-dimensional sub-
and the complex permittivity, E = c - ja/w, is used to space of the class of admissible functions for the problem
account for any conduction and/or dielectric losses in the of interest using the same set of functions as trial and test
medium. functions, is the symmetry of the resulting stiffness matrix
For the purposes of finite element solutions, a weak form given a symmetric weak formulation. However, another
of (1) is required. For node baaed finite element expan- important merit of Galerkins method is that, if a sym-
sions the unknown vector field is approximated in terms metric weak formulation is used, Galerkins approximate
108
solution exactly conserves energy in the electromagnetic while the elements of the forcing vector are
field despite the fact that it satisfies the vector Helmholtz
equation only approximately over the domain of inter- .fj=-f &G X (V X E) * Njds 01)
est. This is easily shown starting from (5) and using the
complex conjugate of the field, E*, as the test function
M is the number of degrees of freedom in the approxi-
and Faradays law to introduce the magnetic field in the
mation. One of the important attributes of the method
boundary integral
of finite elements is that that the basis functions used

(( $V x E) . (V x E)) + (jw?E . E*) =


have local support, i.e. they are non-zero only over a
set of adjacent elements. This is what leads to the spar-
sity of the resulting system since most of the elements
(fi x H) . Eds (7) Aij are zero. The forcing vector is formed by contribu-
f
tions of the surface integrals over the domain boundaries.
The complex conjugation of (7), use of Faradays law to From the uniqueness theorem, the tangential component
simplify the integrand in the first term on the left, and of the magnetic field, -(l/jwp)G x (V x E), on the do-
use of the constitutive relations B = pH, D = 1E, result main boundaries is all that is needed for a unique solution
in the following equation of Maxwells equations inside the domain. This tangen-
tial magnetic field on the domain boundary is used to
(-jwB . H) + (jwD* . E) = (E x H*) . iids (8) account for all sources exterior to the domain of inter-
f est. The surface term in (5) is used also for enforcing
Clearly, the resulting expression is Poyntings theorem for tangential magnetic field continuity conditions at mate-
time-harmonic fields. Thus energy conservation in the rial interfaces. As far as tangential electric field boundary
electromagnetic field is satisfied exactly by the approxi- conditions are concerned, they are taken into account in
mate solution. the construction of the basis functions. This is discussed
The importance of the aforementioned result is rather in more detail in Section 4.
significant. Considering the various potential sources of
error in the development of a numerical solution to a
III. GRID GENERATION
boundary value problem, it is definitely advantageous to
be able to work with a weak statement that is consistent Numerical grid generation is probably the most critical
with the correct physics of the field we are attempting to step in a finite element analysis of electromagnetic wave
calculate. For the electromagnetic field problems of in- interactions. During the early stages of the application of
terest, the calculated field quantities will be acceptable the finite element method to modeling of electromagnetic
only if they satisfy both energy conservation and elec- interactions, the emphasis was on mathematical model
tric charge conservation. As a matter of fact, the latter and weak statement formulations and their subsequent
has been found to be extremely important in the finite use in the analysis of propagation, radiation and scat-
element solution of three-dimensional vector electromag- tering problems in conjunction with rather simple geome-
netic problems. In later sections, it is pointed out that tries. Consequently, the important issue of automatic gen-
modifications to the weak statement in (3) and careful se- eration of finite element grids appropriate for electromag-
lection of the vector basis functions in (5) are needed to netic propagation and scattering problems received rather
prevent the contamination of the numerical solution from limited attention. Apparently, the assumption was that
spurious fields caused by the lack of enforcement of charge grid generation practices used in low-frequency electro-
conservation in the original weak statement. magnetic field modeling and/or other areas of engineeer-
In order to illustrate the development of the numeri- ing in which the method of finite elements was already
cal approximation of the electromagnetic boundary value being used, could be adopted without significant alter-
problem, let us consider the weak statement in (5). Sub- atioEs. This turned out to be a rather false assumption,
stitution of (4) into (5) and testing with each and every simply because the specific physics of the phenomenon
one of the vector basis functions N.J results in a linear that is being analyzed needs be taken into account be-
system of simultaneous equations fore a discrete model is built for its quantification. The
M famous sampling iheorem of modern communication the-
AijEi=fj, j=l,Z&...,M ory for the sampling of a band-limited signal serves as a
c simple, yet powerful example.
id
Extensive research over the past few years has led to
where the elements, Aij, of the so-called stiffness matrix
the understanding and quantification of the impact of the
are given by discrete approximations of wave equations on the accu-
racy and quality of the simulated wave solutions [5]-[8].
-VxNi).(VxNj))+(jvNi.Nj)(lO) The findings from such analysis are now being used as
I

109
guidelines and/or constraints in the development of pro- new solution is then attempted with the refined grid, and
cedures for automatic grid generation. The following is the process is repeated until the desirable solution accu-
a list of undebatable results that are used to guide the racy has been achieved.
construction of a finite element grid for the modeling of While the manipulation of hexahedral meshes, i.e.,
electromagnetic wave propagation, radiation and scatter- meshes that use the so-called brick elements, is much sim-
ing problems. pler than that of a tetrahedral mesh, tetrahedral meshes
First and foremost, the grid must provide for adequate are the ones that have been adopted for automatic grid
wavelength resolution in order to minimize numerical dis- generation. (In two dimensions, brick elements reduce to
persion (i.e., artificial, discretization induced dependence quadrilaterals while tetrahedra reduce to triangles.) The
of phase velocity on electrical grid size) and thus con- reason for this is that triangles and tetrahedra are sim-
trol phase error. The number of degrees of freedom per plices in two and three dimensions, respectively. Com-
wavelength, required to achieve a fixed accuracy, is de- plete polynomial expansion functions are defined easily
pendent on the degree of the interpolating polynomials on simplex elements. Furthermore, an arbitrary domain,
used in the, constyruction of the basis functions, the elec- no matter how complicated it might be, can always be dis-
trical size of the domain (i.e., the size of the domain in cretized automatically into a set of simplex elements using
wavelengths), and the boundary conditions used for grid Delaunay tesselation [9]. As a matter of fact, grid gener-
truncation. Furthermore, problem geometry and mate- ation software baaed on the Delaunay algorithm and its
rial properties impact the way the finite element grid is variants is now commercially available and is being used
constructed. Electromagnetic fields exhibit singular be- for finite element grid generation with fairly satisfactory
havior at material wedges and corners (especially metal- results. Consequently, most of the recent advances on au-
lic), which cannot be resolved properly by the polynomial tomatic mesh generation and mesh refinement are based
interpolating functions used for the finite element approx- on Delaunay triangulation.
imation. Thus, grid refinement or use of special elements With regards to automatic mesh generation, new pro-
are required in the vicinity of such geometric features to cedures have been developed for preventing the occurence
improve solution accuracy. In addition, abrupt changes of degenerate (sliver) elements, for the redistribution of
in the electric permittivity and the magnetic permeabil- existing nodes and/or the efficient incorporation of new
ity of materials need be accompanied by appropriate grid nodes for improving the quality of the tetrahedra (i.e.
size changes in order to both prevent the occurrence of making them as equilateral as possible), and for preserv-
(nonphysical) numerical reflections and maintain the con- ing prescribed material boundaries during the process of
ditioning of the finite element matrix. Finally, the grid initial mesh generation as well as during subsequent mesh
generation process should be such that element degen- refinement [lO],[ll].
eracy (i.e. triangles degenerating toward line segments, As far as mesh refinement is concerned, a variety of
quadrilaterals approaching triangles, or tetrahedrons de- procedures are currently available. At this point, it is im-
generating to quadrilaterals) never occurs. Such elements
portant to point out that in addition to the so-called h
lead to ill-conditioning and may eventually render the fi- refinement, where new nodes are introduced in the grid
nite element solution useless.
while the order of the polynomial approximation remains
While it is often tempting to adopt the philosophy that the same, solution accuracy can be improved by increasing
the finer the grid the better the quality of the solution, the order of the polynomial interpolation while keeping
one needs to remember that an excessively fine mesh (un- the number of nodes the same. This so called p refinement
less needed) wastes computational resources and thus it tends to be less popular than h refinement, mainly due to
should be avoided. Consequently, it is important that the the desire of maintaining low interpolation function com-
grid generation process is such that adaptive mesh refine- plexity within regions of high material complexity. How-
ment is possible. In other words, for the purposes of com- ever, one should keep in mind that h refinement cannot be
putation efficiency, the analysis begins with the construc- pushed to arbitrarily small element sizes. The reason for
tion of an initial mesh the density of which has been de- this is that as the mesh is progressively refined the eigen-
cided on the basis a specific set of constraints perinent to value spectrum of the stiffness matrix widens; hence the
the material properties, the characteristics of the sources system becomes progressively more ill-conditioned [12].
and anticipated spatial variation of the electromagnetic This progressive ill-conditioning of the stiffness matrix as
fields. A simulation is attempted using this mesh. Next, the mesh is refined results in an increase in the round-
an error estimator is used to examine the quality of the off error which counteracts the decrease in approximation
solution over the entire grid and thus identify areas where error. At some point, approximation error and round-off
mesh refinement is needed to improve solution accuracy. error become equal, and any further mesh refinement will
Such mesh refinement is then effected using appropriate result, in deterioration in the accuracy of the solution. As
mesh refinement schemes that maintain element quality pointed out in [12], the round-off error does not depend
and preserve mesh conformity to material boundaries. A strongly on the degree of the interpolating polynomials.
110
It depends mainly on element size, h, and, for the second-
order problems of interest in electromagnetic modeling,
the dependence is of the form O(he2). Consequently, the
way to maintain accuracy and postpone the deleterious
effect of round-off is to increase the order of the interpo-
lating polynomials (p refinement). The numerical experi-
ments in [13] support this conjecture in a very illustrative
manner.
We conclude this section by pointing out that before
adaptive mesh refinement can be effected, a reliable a pos-
terioti error estimate to be used as a refinement criterion k
is required. A variety of error estimators have been pro-
posed over the past few years [14]-[171. As expected, they Fig. 1. Whitney l-form edge element.
depend on the physical attributes of the fields as well as
the choice of functions and interpolating elements used in ners.
the finite element approximation. The most popular version of these edge elements is the
so-called Whitney l-form. It was long before the method
of finite elements was becoming a popular tool in bound-
IV. THE CHOICE OF ELEMENTS ary value problem solving that Whitney described a fam-
ily of polynomial forms on a simplicial mesh with special
Most of the original applications of the finite element properties that made them attractive for electromagnetic
method to vectorial electromagnetic field modeling in field representations [23]. These polynomials are of, at
three dimensions were based on the so-called nodal ele- most, degree one on tetrahedra. Any two p-forms are
ments, using field representations of the form shown in said to conform on a surface if they take the same val-
(2) and weak statements such as the one in (3). In ues at any given set of p vectors tangent to the surface.
other words, the degrees of freedom were defined to be Finally, p-forms are uniquely determined by integrals on
the three components of the unknown fiield quantity at p-simplices. Let us consider, for example, the popular
the element nodes. It was soon found that such approxi- Whitney l-forms (Fig. 1). They are associated with mesh
mations were plagued by the occurrence of non-physical, edges. Each edge in the tetrahedral mesh contributes an
spurious modes [18]. These spurious solutions manifested independent basis function. In other words, the degrees
themselves as modes with nonzero divergence, and were of freedom of the approximation are associated with the
caused by the inability of the aforementioned choice of element edges; this is the reason they are called edge el-
interpolation and weak statement to enforce Gauss law ements. For an edge e = {i, j} connecting vertices i and
for divergence-free solutions. Consequently, a variety of j the basis function is given by
approaches were proposed for eliminating these spurious
modes within the context of nodal elements [19]-[21]. Ne = We = <iV<j - &V<i (12)
A different approach to eliminating spurious modes that
where <i (i = 1,2,3,4) are the bary-centric (or simplex)
has gained popularity over the past few years is the use
coordinates of the tetrahedron. It is easily shown that the
of the so-called edge elements [22]. Edge elements use
circulation of We is 1 along edge e and 0 along all other
vector basis functions of the form shown in (4). They
edges. With the unknown field quantity interpolated over
have the unique property that the degrees of freedom as-
a tetrahedron as
sociated with these elements are the circulations of the
6
vector field along the edges of the mesh. The impor-
tant implication of this fact is that edge elements im- E=CE,We (13)
e=l
pose the continuity of only the tangential components of
the unknown electric (or magnetic) field across element the aforementioned property implies that the degrees of
boundaries, which is obviously consistent with the phys- freedom of the approximation are the circulations of the
ical constraints for electromagnetic field solution unique- field along the edges of the tetrahedron. Two easily ob-
ness inside a domain. Additional advantages of edge ele- tained properties are the continuity of tangential compo-
ments are: a) Boundary conditions at material interfaces nents across facets (element interfaces) and the zero diver-
are automatically enforced through the natural boundary gence of the basis functions. In view of this last property,
condition in the weak statement (see the surface integral it is clear that by using edge elements the requirement for
term in (5)). b) Dirichlet boundary conditions are eas- divergence-free numerical solutions is built in the basis
ily imposed along element edges. This second property functions.
becomes extremely useful when dealing with conducting Despite their simplicity and their suitability for electro-
boundaries, especially those that exhibit wedges and cor- magnetic field representation, the Whitney l-forms have
111
the drawback that, being first-order polynomials, the in- and/or scatterers. On this mathematical boundary an ap-
terpolation error in the approximated field is only first propriate operator has to be used to effect the reflection-
order. Consequently, a very fine wavelength resolution is less transmission of radiation out of the domain. Such an
required for controling numerical dispersion, which trans- operator can be effected using an integral equation state-
lates to large number of unknowns. To circumvent this ment of Huygens principle. A variety of such global grid
difficulty, higher-order vector finite elements have been truncation operators have been proposed [l]. Such global
proposed by several authors, starting with the pioneering operators have the undesirable property that their numer-
work of Nedelec [24]. These elements are called tangential ical approximation results in a fully populated submatrix
vector finite elements, and continue to be a topic of exten- which, unless managed properly, penalizes the sparsity of
sive research [25]-[28]. However, it should be mentioned the matrix resulting from the finite element approxima-
that the number of degrees of freedom associated with tion inside the domain.
these high-order elements on tetrahedra increases rapidly To overcome this difficulty, significant effort was de-
with the order of the element, and this is an issue that voted over the past few years on the development of local,
needs be weighted properly when trading-off computation differential equation-based operators for grid truncation
complexity and computation efficiency with solution ac- [32]. Such local operators, often called absorbing bound-
curacy. ary conditions (ABCs), are approximate and thus their
However, as mentioned earlier, use of higher-order el- performance is inferior to the one achieved by the global
ements reduces the dimensionality of the approximation, operators. However, they help retain the sparsity of the
improves the conditioning of the finite element system, stiffness matrix.
and leads to increased accuracy in the numerical differen- Despite the successful development and application of
tiation of the finite element solution for post-processing such local operators for the truncation of finite element
purposes. Since higher-order tangential vector elements
grids with either nodal or edge elements, the reflection
are much more expensive than nodal elements, hybrid ap- levels resulting from their applications have not been as
proaches have been proposed as an alternative. The basic
small as needed for applications relevant to EMI/EMC
idea is to use nodal elements over those regions that in- problems. For example, it is well-known that common-
volve either homogeneous media or media with smooth mode currents resulting from imbalances in interconnects
variations in their electromagnetic properties. Edge ele-
and improperly designed current return paths are signifi-
ments, on the other hand, are used at material interfaces
cant contributors to radiated emissions from boards, de-
as well as in the vicinity of conducting wedges, corners spite their very small amplitudes. If the absorbing bound-
and, in general, points where several material interfaces ary condition causes an appreciable level of reflection, the
meet and thus the definition of a vector normal is not spurious reflected fields may alter the common-mode cur-
unique. Mur has shown that the use of such combinations rents in the interconnects in the system and thus lead
of edge and nodal elements in electric or magnetic field- to totally erroneous calculations of radiated emission lev-
based finite element formulations yields optimum compu- els. Clearly, the availability of an almost reflectionless
tational results [29]. In such mixed-element formulations, grid truncation methodology that maintains the sparsity
it is important that the divergence-free character of the of the stiffness matrix is highly desirable for EMI/EMC
fields and the continuity of the flux across interelement finite element modeling.
boundaries are imposed explicitly in a weighted sense.
This is necessary for the elimination of spurious modes, It was only three years ago that a breakthrough was
and helps increase solution accuracy and improve conver- achieved by Berenger in this area of numerical grid trun-
gence [30]. Finally, Boyse and Seidel have used such a cation [33]. His aproach was based on the numerical con-
combination of edge and nodal elements in conjunction struction of anisotropic absorbing media with the prop-
with the finite element approximation of a scalar and vec- erty that the interface between such a medium and a
tor potential formulation of Maxwells equations [31]. homogeneous space is (theoretically) reflectionless for all
frequencies and all angles of incidence (except at gras-
ing). These absorbers were called perfectly matched Zay-
V. GRID TRUNCATION FOR UNBOUNDED PROBLEMS ers PMLs. Even though Berengers implementation of
such PMLs involved a peculiar splitting if the field com-
An important class of problems in EMI/EMC analysis ponents so that the PMLs could be applied in conjunc-
deals with noise radiated out of a component or system tion with transient electromagnetic simulations using the
or radiated noise interacting with a system with part of it finite difference time domain method, several researchers
getting scattered and part coupled to the system compo- explored and continue to explore alternative implemen-
nents. In order to model such electromagnetic radiation tations of PMLs that are more suitable for finite ele-
and scattering problems using finite methods, the domain ment applications [34]-[371. As clearly illustrated in [36],
of computation needs to be truncated by a (non-physical) such PMLs may be effected using a properly constructed
mathematical boundary that encloses all radiation sources anisotropic medium. To illustrate the properties of such
112
a medium, consider the case of a planar interface paral-
lel to the t - y plane in a Cartesian coordinate system.
Let the medium on the left be a homogeneous, isotropic
medium with constant permittivity E and constant per-
meability /J. To effect a reflectionless interface the per-
mittivity and permeability tensors of the medium on the
right need be defined as follows, [E]/E = diag{a, a, u-l},
and [p-J/p = diag{a, a, s-l }. Furthermore, to effect ab-
sorption of the transmitted wave as it propagates inside
the PML, we choose a = 1 - jsZ, s, > 0.
As mentioned earlier the construction of optimum
PMLs is currently the subject of intensive research. It
turns out that the numerical implementation of PMLs is PEC
not reflectionless. The origin of these reflections is the
numerical discretization of the fields as well as the fact Fig. 2. A longitudinal cross section of a complicated structure. Par-
that the thickness of the PML has to be truncated for nu- tition planes PI through P+ are used to decompose the structure in
smaller subdomains.
merical purposes. From a variety of studies performed by
several researchers on the effectiveness of the anisotropic
extension of the aforementioned concept to complicated,
PML in conjunction with finite element grid truncation in
not-necessarily separable domains.
the frequency domain, the current state of the art could be
summarized as follows: a) The effectiveness of the PML The domain decomposition approach presented here is
is strongly dependent on the values assigned to its ma- significantly different from other techniques of the same
terial properties, its thickness and its distance from the name [41],[42]. It is probably most closely related to the
scatterer. Typically PML thicknesses of 0.1X - 0.2X and connection scheme of [43]; however, it is more versatile
distances from the scatterer of 0.2X are expected to lead to and more general in the type and complexity of structures
very good PML performance. b) The possibility exists for it can model. Our presentation will concentrate on the
optimizing the performance of the PML for a given thick- motivation behind such an approach and the fundamental
ness by allowing its material properties to vary smoothly, steps for its implementation. For a more concrete mathe-
in a way similar to that used in conjunction with FDTD matical discussion references [44]-[45] should be consulted.
simulations of transient waves. c) While a choice of the The rationale for the use of domain decomposition is
form a = 1 - js, s > 0, will facilitate the absorption of best explained with the aid of an example structure. Fig.
propagating waves, the absorption of waves exhibiting ei- 2 illustrates a longitudinal cross section of a rather com-
ther only evanescent or both evanescent and propagating plicated shielded environment. Let us assume that the
behavior needs to be assisted by allowing the real part of a objective is the numerical modeling of noise interactions
to assume values greater than 1 [38]-[40]. d) The material within this structure. While the development of a numer-
properties of the PML tend to slow down the convergence ical grid for the entire structure could be attempted, it is
of the iterative solvers used for the solution of the finite rather obvious that the material complexity in the cen-
element system. tral part of the structure will require a much finer grid
To summarize, the use of anisotropic absorbers for finitethan the rather homogeneous end regions to maintain a
element grid truncation is expected to impact significantly desirable numerical solution accuracy. Furthermore, if the
our ability to simulate time-harmonic electromagnetic in- structure length is several wavelengths at the frequencies
teractions in unbounded regions using the method of fi- of interest, the size of the resulting finite element system
nite elements. However, there remains a lot of work to be could become prohibitively large for even the largest su-
done for this grid truncation methodology to reach its full percomputers available.
potential. A domain decomposition approach to the solution of
this problem proceeds as follows. First, the mathematical
partitions Pi through P4 are used to decompose the struc-
VI. DOMAIN DECOMPOSITION ture into five smaller domains Qi through Qs. Clearly, the
choice of the partitions of Fig. 2 was influenced by the
Domain decomposition is a rigorous approach for the variability in material complexity. The power of the do-
solution of linear boundary value problems in complicated main decomposition approach lies on the linearity of the
domains. As a matter of fact, it is routinely applied for governing equations and the uniqueness theorem for the
the development of analytic solutions to boundary value solution to Maxwells equations which is conveniently ex-
problems in physics and engineering when geometries with pressed by the weak statements (3) and (5). For example,
multiple homogeneous regions connecting along separa- from (5) it is clear that E is uniquely specified inside a
ble boundaries are present. What we consider here is an domain if the tangential magnetic field is defined every-
113
where over the domain boundary. On the basis of this and well-conditioned matrices due to better uniformity of
observation, the domain decomposition approach intro- the numerical grid. Consequently, round-off error is re-
duces appropriate expansions for the tangential magnetic duced and solution accuracy is improved.
fields on the partitions. These expansions are in terms of
properly selected known, linearly independent expansion
functions, and the unknown coefficients in these expan- VII. DISCUSSIONAND CONCLUSIONS
sions become the fundamental degrees of freedom of the
problem. For each expansion function on each partition Electromagnetic modeling for EMI/EMC analysis and
a solution for the electric field inside the two subdomains design of components and systems is one of the most chal-
adjacent to the specific partition is generated. The set of lenging areas for numerical electromagnetic simulation
interior solutions thus generated within each subdomain specialists. Topological complexity, material variability
due to all the expansion function excitations on the en- variability, large variability in feature sizes and broadband
closing partitions constitute a basis that can be used for frequency analysis, are the main reasons for the aforemen-
the representation of the unknown field inside the subdo- tioned complexity of EMI/EMC simulations. All of these
main. In other words, using the principle of superposition, factors contribute to a rapid growth in the number of de-
the unknown field inside each subdomain is represented grees of freedom needed for the numerical approximation
as the sum of the generated interior solutions, each one of the problem, as one attempts to increase the sophisti-
weighted by the coefficient in the expansion for the tan- cation and accuracy of the computer model. Very soon
gential magnetic field on the partitions that is associated computer resource requirements exceed availability or, the
with the specific excitation function. Finally, the solution turn-around time for numerical simulation becomes pro-
for the unknown coefficients in the tangential magnetic hibitively long and thus inappropriate for use as a design
field expansions on the partitions is effected by enforcing aid. Inevitably, simplifications need to be made, and a re-
the continuity of the tangential electric field across the duced model is developed with significant effort placed on
partitions and solving the resulting linear system. its ability to encompass the most important contributions
to the process under investigation.
Simply expressed, domain decomposition is a technique So far, applications of frequency-domain finite element
for shifting the unknowns of a large and/or complicated methods have followed such a path. Evaluation of the ef-
boundary value problem from the continuum of points fectiveness of conducting shields with slots or seems have
within the total structure to coefficients on properly se- benefited from finite element modeling [3],[46]. More re-
lected partitions. Its chief strength is the independence cently, commercially available finite element software is
of the subdomains. The first consequence of this inde- being used for multiconductor interconnect transmission-
pendence is computational flexibility. Because each sub- line parameter extraction, extraction of lumped capaci-
domain is modeled independently, different subdomains tive/inductive equivalents for electrical modeling of inter-
may be analyzed with different techniques. Analytic so- connect and package discontinuities; as well as electronic
lutions may be used for separable subdomains, while finite package inductance and capacitance modeling. Such mod-
methods or integral equation techniques may be used for els are subsequently used for crosstalk and switching noise
subdomains of high complexity. The second consequence prediction in packaged electronic systems. The mod-
of the independence of subdomains is modularity. Many eling of electronic circuits, described in terms of their
subdomains may be present in the structure; however, voltages and currents, in conjunction with vectorial field
only some may be unique. Numerical solutions need be solvers based on finite elements has been demonstrated
calculated only for the unique subdomains. The third also [3] ,[47]. Other recent applications include frequency-
major consequence of the independence of subdomains is dependent, scattering parameter characterization of pack-
parallelism. More specifically, two degrees of parallelism age and interconnect discontinuities, investigation of the
are present in the generation of the subdomain interior impact of the heatsink over a chip package on radiated
solutions. The solutions for different subdomains are in- emiseions from the package, as well as the prediction of
dependent as well as the solutions for different excitation resonances inside a shielded box for a given layout of the
functions for a given subdomain. Given enough compu- various components.
tation resources, all these solutions may be generated in The above list is expected to grow substantially as com-
parallel. puting technology continues to advance rapidly, provid-
A number of additional benefits come from modeling ing us with higher computation speeds, larger memory
smaller domains. Numerical grid generation and adaptive resources, parallelism and distributed computing. Con-
refinement is easier in smaller domains. Partitioning al- tinuing research in advancing the state-of-the-art in auto-
lows the separation of disparate mesh regions without the matic grid generation and refinement, improving the per-
need for transition areas between fine meshes and coarse formance and robustness of reflectionless grid truncation,
meshes. Accuracy of the solution can be improved also. and automating the application of domain decomposition
Indeed, domain decomposition results in smaller matrices approaches, will help enhance the power of the finite ele-
114
ment method to solve realistic EMI/EMC problems. PI W.E. Boyse, D.R. Lynch, K.D. Paulsen, and G.N. Minerbo,
Nodal baaed fmite element modeling of Maxwells equations,
IEEE Trans. Antennas PTopagat., vol. 40, no. 6, pp. 642-651,
REFERENCES 1992.
P21 A. Bossavit , Simplicialflnite elements for scattering problems
in electromagnetism, Computer Methods in Applied Mechan-
PI J.M. Jii, The Finite Element Method in Electromaqnetics,
ics and Engineering, vol. 76, pp. 1262-1271,1989
John Wiley & Sons, Inc., New York, 1993.
PI R. Laroussi, G.I. Costache, Finite-element method applied to [231 H. Whitney, Geometric Integration Theory, Princeton U.P.,
EMC problems, IEEE Trans. Electromagnetic Compatibility, 1957.
vol. 35, no. 2, pp. 178-184, May 1993. 1241 J.C. Nedelec, Mixedflnite elements in @,I Numerical Math-
smatics, vol. 35, pp. 315-341, 1980.
131 JR. Brauer and B.S. Brown, Mixed-dimensional finite ele-
ment models of electromagnetic coupling and shielding, IEEE [251 J.C. Nedelec, A new family of mixed finite elements in Es,
Trans. ElectTomaqnetic Compatibility, vol. 35, no. 2, pp. 235- Numerical Mathematics, vol. 50, pp. 57-81. 1986.
241, May 1993. @I S.V. Polstyanko and J.F. Lee, HI (curl) tangential vector fi-
nite element method for modeling anisotropic optical fibers,
[41 D.S. Dixon, M. Obara, and N. Schade, Finite element analysis
(FEA) as an EMC prediction tool, IEEE TTans. Eleckomag- IEEE Journal of L.ightwave technology, in press.
netic Compatibility, vol. 35, no. 2, pp. 241-248, May 1993. [271 I. Bardi et al, Higher order edge elements in electromagnetic
field modeling, Proceedings of 1995 International AP-S Sym-
151 A. Bay&s, C.I. Goldstein, and E. Turkel, On accuracy con-
ditions for the numerical computation of waves, Journal of posium/URSI Meeting, Newport Beach, CA, June 1995.
Computational Physics, vol. 59, pp. 396404,1985. PI T.D. Tsiboukis, private communication, Department of Elec-
trical Engineering, Aristotle University of Thessaloniki, Thes-
PI R. Lee, and A.C. Cangellaris, A study of discretization error salon&i, 54006, Greece.
in the finite element approximation of wave solutions,, IEEE
Transactions on Antennas and Propagation, vol. 40, pp. 542- PI G. Mm, Optimum choice of finite elements for comput-
549,1992. ing three-dimensional electromagnetic fields in inhomogeneous
media, IEEE Trans. on Magnetics, vol. 24, pp. 339-333,198s.
(71 W.R. Scott, Errors due to spatial discretization and numer-
ical precision in the finite-element method, IEEE Transac- [301 G. Mur, Finite-element modeling of three-dimensional time-
tions on Antennas and Propagation, vol. 42, pp. 1565-1570, harmonic electromagnetic fields in inhomogeneous media, Ra-
1994. dio Science, vol. 26, pp. 275-280,199l.
PI G.S. Warren, and W.R. Scott, An investigation of numerical [311 W.E. Boyse and A.A. Seicll, A hybrid ii&e element method
dispersion in the vector finite element method using quadri- for 3-D scattering using nodal and edge elements, IEEE
lateral elements, IEEE Transactions on Antennas and Prop- TTans. Antennas PTopagat., vol. 42, no. 1, pp. 14361442,1994.
agation, vol. 42, pp. 1502-1508,1994. 1321 D. Givoh, Non-reflecting boundary conditions, J. Comput.
Physics, vol. 94, pp. l-29, 1991.
PI J.C. Cavendish, D.A. field, and W.H. Frey, An approach to
automatic three-dimensional finite element generation, Int. [331 J.P. Berenger, A perfectly matched layer for the absorption
J. Numer. Methods Enq., vol. 21, pp. 329-347,198s. of electromagnetic waves, J. Comput. Physics, vol. 114, pp.
185-200, 1994.
ml N.A. Golias and T.D. Tsiboukis, An approach to refln-
ing three-dimensional tetrahedral meshes based on Delaunay 1341 W.C. Chew and W.H. Weedon, A 3-D perfectly matched
transformations, Int. J. Numer. Methods Eng., vol. 37, pp. medium from modified Maxwells equations with stretched co-
793812,1994. ordinates, Microwave and Optical Technology Lett., vol. 7,
no. 13, pp. 599-604,1994.
Pll J.-F. Lee and R. Dyczij-Edlinger, Automatic mesh generation
using a modified Delaunay tesselIation, submitted to IEEE [351 C .M. Rappaport, Perfectly matched absorbing boundary con-
Antennas and Propagation Magazine. ditions based on anisotropic lossy mapping in space, IEEE
Microwave and Guided Wave Lett., vol. 5, no. 3, pp. 90-92,
[=I G. Strang and G.J. Fix, An Analysis of the Finite Element 1995.
Method, Prentice-Hall, Englewood Cliffs, N.J., 1973.
[I31 W.R. Scott, Errors due to spatial discretization and nmneri- [361 Z.S. Sacks, D.M. Kingsland, R. Lee, and J.-F. Lee, A per-
cal precision in the finite-element method, IEEE Trans. An- fectly matched anisotropic absorber for use as an absorbing
tennas PTopaqat., vol. 42, pp. 1565-1570,1994. boundary condition, IEEE Trcms. Antennas Propagat., vol.
43 , no. 12, pp. 1460-1463,1995.
[I41 I. Babuska, O.C. Zienkiewicz, J. Gage, E.R. de Oliveira, Eds.,
Accuracy Estimates and Adaptive Refinement in Finite Ele- [371 L. Zhao and A.C. Cangellaris, A general approach for the
ment Computation, John Wiley & Sons, Inc., New York, 1986. development of unsplit field time domain implementations of
perfectly matched layers for FD-TD grid truncation, IEEE
[I51 P. Fernandes, P. Girdinio, and G. Molinari, Techniques for
h-refinement in adaptive meshing algorithms, COMPEL, vol. Microwave and Guided Wave Lett., (acceptedfor publication),
13, supplement A, pp. 329-334,1994. June 1996.
[=I J.F. Lee, D.K. Sun, and Z. J. Cendes, F&-wave analysis of [331 M.A. Gribbons, W.P. Pineho, and A.C. Cangellaris, A
dielectric waveguides using tangential vector finite elements, stretched coordinate technique for numerical absorption of
IEEE Trans. h!ficTowave Theory Tech., vol. 39, no. 8, pp. 1262- evanescent and propagating waves in planar waveguiding
1271,199l. structures, IEEE TTans. Microwave Theosy Tech., vol. 43,
no. 12, pp. 28852889.1995.
1171 N.A. Gal& and T.D. Tsiboukis, Consitutive inconsistency:
Rigorous solution of Maxwell equations based on a dual ap- [391 Z. Wu and J. Fang. Numerical implementation and per-
proach, IEEE Trans. on Magnetics, vol. 30, no. 5, pp. 3586- formance of perfectly matched layer boundary condition for
3589, Sept. 1994. waveguiding structures, IEEE Trcms. Microwave Theory
Tech., vol. 43, no. 12, pp. 26762683,199s.
ml J.P. Webb, The finite-element method for linding modes of
dielectric loaded cavities, IEEE TTans. MieTowave Theory 1401 J.-F. Lee, ECE Department, Worcester, and R. Lee, EE De-
Tech., vol. 33, no. 7, pp. 635-639,198s. partment, Ohio State University, private communication.
WI A.J. Kobelansky and J.P. Webb, Eliminatingspurious modes [411 G.L. Hennigan, S. Castillo, and E. Hensel, Using domain de-
in finite-element waveguide problems by using divergencefree composition to solve symmetric, positive-definite systems on
fields, Electron. Lett., vol. 22, pp. 569-570,1986. the hypercube computer, Int. J. Num. Meth. Eng., vol. 33,
pp. 1941-1954,1992.
1201K.D. PaulsenandD.R. Lynch, Eliminationof vectorparasites
in finite element Maxwell solutions, IEEE Trans. Microwave [421 G.F. Carey, Parallel subdomain and element by element tech-
Theory Tech., vol. 39, no. 3, pp. 395-404,1991. niques, in Parallel Supercomputing: Methods, Algorithms,
115
and Applications, G.F. Carey, ed., John Wiley & Sons, Inc.,
Cbicester, UK, 1989.
[43] T.-M. Wang and H. Ling, Electromagnetic scattering from
three-dimensional cavities via a connection scheme, IEEE
Trans. Antennas Propagat., vol. 39, no. 10, pp. 1505-1513,
1991.
[44] C.T. Spring and A.C. Cangellaris, Application of domain de-
composition methods for the electromagnetic anaysis of large
linear two-dimensional structures, J. Electromagnetic Waves
Applic., vol. 9, no. l/2, pp. 175-192,1995.
[45] C.T. Spring, Domain decomposition in the rigorous elec-
tromagnetic analysis of large structures, Ph.D. Dissertation,
ECE Department, University of Arizona, Tucson, AZ, 1995.
[46] S.D. Gedney and R. Mittra, Electromagnetic transmission
through inhomogeneously Shed slots in a thick conducting
plane - Arbitrary Incidence, IEEE Trans. Electromagnetic
Compatibility, vol. 34, no. 4, pp. 404415,1992.
[47] J.R. Brauer, B.E. MacNeal, L.A. Larkin, and V.D. Overbye,
New method of modeling electronic circuits coupled with 3D
electromagnetic finite element models, IEEE Trans. Magnet-
its, vol. 27, pp. 4085-4088, 1991.