Resonant Cavity Design Using the Finite Element Method

D. Walsh, C.R.I. Emson and C.P. Riley, Vector Fields Ltd., 24 Bankside, Oxford OX5 1JE, U.K.

Abstract vx x
where ( ) and ( ) are arbitrary vector and scalar fields
respectively. The second term is familiar from electrostatics
This paper presents a finite element method for calculating
where it is common to solve the potential problem,
resonant frequencies for an arbitrary three dimensional cav-
ity. Firstly traditional nodal methods are considered and the r
"r(x) = (x) (5)
reason for their failure in modelling high frequency fields.
An edge element formulation and its solution of the prob- and from  determine E . Any electrostatic solution may
lems of spurious modes and modeling of reentrant corners be formally considered to be a time harmonic solution with
is presented. With the aid of two realistic cavity designs the zero resonant frequency. Indeed the solutions to (3) fall into
strengths of the finite element approach is illustrated. two categories,

1 TRADITIONAL METHODS E(x) = rEphysical
source;
; !n 6= 0
!n = 0 (6)
The use of finite element analysis is widespread throughout
the spectrum of engineering disciplines. Indeed its use in The dynamic relation (3) contains a subset of zero eigen-
the design of electrical equipment goes back to the very in- frequency solutions corresponding to electrostatic problems
ception of the method. However, until recently, the analy- with arbitrary sources. These are the “spurious” modes.
sis of high frequency devices such as resonant cavities has The traditional method for representing a vector field in
finite elements is to expand each of its components in terms
been beyond their reach. Traditional approaches have been
plagued by “spurious”, i.e. nonphysical, modes and the in- of nodal scalar shape functions Ni ( ), x
X
ability to model singularities adequately. The background
theory and the reasons for these difficulties are discussed be- E(x) = ep Np (x) (7)
nodes p
low.
e
where p are the vector unknowns at each node. The
1.1 Background Theory Galerkin method then leads to the following functional
Ex
equation for ( ),
For simplicity it can be assumed that the cavity contains
Z
r^ W(x)
( 1 r^ E(x)) "!n2 W(x):E(x)d
= 0:
no charges or currents and the walls are made from per-
fectly electrically conducting (PEC) material. The funda-

0
mental equation describing the electric field is the vector (8)
wave equation: where W is a trial vector field of form (7). Performing the

@ 2 ("E(x))
integrals leaves a sparse generalised eigenvalue problem,
r^ ( 1 r^ E(x)) = @t2
(1)
A e = !2 B e (9)
n

with the following condition holding PEC boundaries for which standard methods of solution exist [1]. Unfortu-
nately the resulting spectrum of solutions obtained is a tan-
E
t = 0 on PEC (2) gle of spurious and physical modes. Although examination
of the modes calculated may be made, discarding those with
Further assuming all materials within the cavity are linear,
(1) reduces to the vector Helmholtz equation for each indi-
E
“significant” values of r
, this is hardly satisfactory.
This discrepancy between the continuum prediction of (6)
vidual mode n, and the discrete implementation is due to an inappropriate

r^ ( 1 r^ E(x))
choice of basis functions. The problem is that nodal basis
"!n2 E(x) = 0 (3) vectors cannot exactly represent fields of the form r. This
is evident from the discontinuity of r crossing element
In a local volume of space it is always possible to decompose Wx
boundaries whereas ( ) is continuous everywhere. As a
a vector field into rotational and irrotational components, result the zero frequency continuum solutions aquire mesh
and problem dependent frequencies, typically comparable
E(x) = r^ v(x) + r(x) (4) to the physical modes of the device.
 E
The use of as the state variable in traditional electro- where the nodes i, j , k and l form the terathedra enclosing
magnetic finite elements has always been problematic be- x
the point . This is the property that resolves the problem
cause it is discontinuous between media of differing dielec- associated with “spurious” modes found in traditional meth-
tric properties. Alternatively the problem may be described ods. Modeling a scalar field using the standard nodal shape
A
in terms of the underlying potentials and  which are con- functions, , the gradient of this field can be exactly rep-
tinuous. Choosing the Lorentz gauge the problem may be w
resented as a linear combination of edge vectors fg . The
reformulated as, “spurious” modes still exist if edge variables are used but

r2 A(x) "!2 A(x) = 0

they can be exactly represented in terms of edge vectors and
(10) therefore do not acquire non-zero eigenfrequencies. Conse-
A
with the boundary condition for analogous to (2). The
quently physical and spurious spectra are completely sepa-
rated.
“spurious” modes in this approach are then removed by the
Another advantage is that the boundary condition (2) is
boundary conditions.
completely natural. The coefficients efij g for edges lying
Unfortunately for problems with reentrant corners, and
in the PEC wall are simply set to zero. The values for edges
therefore singular fields, methods of this type are found to be
touching a reentrant corner are no different from any other
significantly in error [2]. Even for models where (2) is self
and their values are determined by the solution. Also being
evident they do not even converge to the correct solution. It
related to the emf jump their values are strictly finite.
is only by using singular expansion functions around reen-
trant corners that the correct results are recovered [3]. For
this approach to be generally applicable it would be neces- 3 CAVITY EXAMPLES
sary to determine the asymptotic solution of the field around To test this method a simple rotationally symmetric cavity,
arbitrary junctions of this type. Thankfully there is a more see Figure 1, has been analysed and compared to the 2D fi-
elegant solution to this problem which is presented in the nite difference program URMEL-T and experiment at the
following section. Daresbury laboratory [6]. The harmonics of this system and

2 EDGE ELEMENTS
As shown, the combination of discrete meshes and vector
components is not a comfortable match. Edge elements side
step the whole question of components by only dealing with
well defined scalar quantities. In this approach the degrees
E
of freedom associated with are not its components but the
emf jump between connected nodes. The emf efij g is de-
fined as, Z
efij g =
E dl (11)
f g l ij

where i and j are two nodes in the mesh and lfijg is the di-
rected edge connecting them. This is the origin of the term
“edge” in edge elements. In some respects this approach
is very similar to the method of Yee [4] in Finite Differ-
ence Time Domain, although this method is more rooted in
the abstract field of differential geometry [5]. The electric
field is represented by an expansion in vector shape func-
w
tions fijg associated with each edge in the mesh,
X OPERA-3d

E(x) = e fijg wfijg

Post-processor 2.605

(12)
edges Figure 1: Quarter section of RF cavity given in [6].
The most well known edge element is the Whitney tetra-
coupling to the main beam line are required. Contrary to fi-
hedral element. The vector shape function associated with
nite difference methods the mesh density may be varied ar-
edge lfijg is,
bitrarily in three dimensions and in this example it is quite
wf g (x) =  (x)r (x)  (x)r (x) (13)
ij i j j i
important to do so. There are two considerations when do-
ing this:
where the  (x) are the barycentric coordinates of x. These
i
basis vectors have many special properties but one of the  Accurate field values in the beam tube to calculate the
most important is the following: coupling of each harmonic to the beam
 Accurately modelling the field variation, possibly sin-
r (x) = wf g (x) + wf g (x) + wf g (x)
i ij ik il (14) gular, of the harmonics themselves.
Therefore the mesh in the beam tube and reentrant sections Mode Freq. (MHz) Q g (
)
have been graded finely and then progressively coarsened No Ports 361.40 36876 147.2
into the homogeneous chamber volumes. Zero 356.35 41199 150.5
For devices operating at these frequencies the skin depth -1.4% +12% +2.2%
of the real conducting walls is so small that they may be re-  352.0 35677 155.0
placed with PEC boundaries. For this model the resonant -2.6% -3.3% +5.3%
frequencies in MHz are: Expected 1-2% 5-20% 2-5%
The final row contains the perturbations expected by CEA.
Mode Experimental SopranoEV URMEL-T
Zero 125.12 125.54 124.66
L 541.79 546.82 544.31 4 CONCLUSIONS
L 728.48 730.9 725.24 The use of Edge elements has removed the problems as-
sociated with “spurious” modes and reentrant corners, that
The column labelled SopranoEV is the implementation of rendered traditional methods unviable. This new approach
the edge method and URMEL-T is the standard 2D finite brings with it all the flexibility associated with finite ele-
difference code. These are all longitudinal (L) modes. ments and has shown consistently good results for real prob-
A true three dimensional RF cavity, a section of which is lems. The extension of this method to lossy material and
shown in Figure 2, has also been modelled [7]. It comprises complex eigenfrequency determination is straightforward
and ongoing.

5 REFERENCES
[1] R.G. Grimes, J.G. Lewis and H.D. Simon: ‘A shifted block
Lanczos algorithm for solving sparse symmetric generalized
eigenproblems’, SIAM J. Matrix Anal. Appl. Vol. 15 No. 1 pp.
228-272 1994
[2] I. Bardi, O. Biro, R. Dyczij-Edlinger, K. Preis and K.R.
Righter: ‘On the treatment of sharp corners in the FEM analy-
sis of high frequency problems.’ Record of 9th COMPUMAG
Conference pp. 372-373, 1993
[3] J.P. Webb: ‘Finite element analysis of dispersion in waveg-
uides with sharp metal edges’, IEEE Trans. on Microwave
Theory and Techniques MTT-39 no. 3, pp.501-505 (1991)
[4] ‘Computational Electrodynamics’, by A. Taflove, published
by Artech House Inc., 1995.
[5] A. Bossavit: ‘Whitney forms: a class of finite elements for
three dimensional computations in electromagnetism’, IEE
Proc. Vol. 135, Pt. A No. 8 pp. 493-499
[6] P.A. MacIntosh: ‘An Investigation of model performance of
low frequency RF cavities’, EPAC 94
[7] Pascal Balleyguier, CEA/PTN, Private Communication.

OPERA-3d
Post-processor 2.605

Figure 2: One sixteenth section of CEA RF cavity.

oval chambers linked by the beam pipe and four symmetri-

cally placed ports. A similar discretisation stategy was used.
The aim of this exercise was to determine the effect of these
ports. Below are presented the main quantities of interest,
including the geometric shunt impedance g: