IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, VOL. 48, NO. 2, FEBRUARY 2000 231
Numerical Stability of NonorthogonalFDTD Methods
Stephen D. Gedney, Senior Member, IEEE, and J. Alan Roden, Member, IEEE
Abstract—In this paper, a sufficient test for the numerical
stability of generalized grid finite-difference time-domain (FDTD)
schemes is presented. It is shown that the projection operators of
such schemes must be symmetric positive definite. Without this
property, such schemes can exhibit late-time instabilities. The
origin and the characteristics of these late-time instabilities are
also uncovered. Based on this study, nonorthogonal grid FDTD
schemes (NFDTD) and the generalized Yee (GY) methods are
proposed that are numerically stable in the late time for quadri-
lateral prism elements, allowing these methods to be extended
to problems requiring very long-time simulations. The study of
numerical stability that is presented is very general and can be
applied to most solutions of Maxwell’s equations based on explicit
time-domain schemes.
Index Terms—FDTD methods, numerical stability.
I. INTRODUCTION
T
HE finite-difference time-domain (FDTD) method has
been highly successful for the analysis of a plethora of
electromagnetic interaction problems [1]. Yet one principal
limitation of the classical FDTD method is the restriction to
orthogonal grids. A number of techniques have been proposed
to develop FDTD methods based on conformal meshing
such as the contour path FDTD (CPFDTD) method [2], the
nonorthogonal FDTD (NFDTD) method [3], and the discrete
integral equation (DSI) [4] and generalized Yee (GY) methods
[5]. The advantage of these techniques is that through the
introduction of more generalized discretizations, error due
to boundary discretization can be alleviated. Unfortunately,
these and similar methods can sometimes suffer from late-time
instabilities [6].
It was demonstrated by Craddock et al. that the source of in-
stability of the CPFDTD method was due to the nonreciprocal
nature of the original algorithm [7]. By analyzing the FDTD
method as a passive circuit, an alternative stable and accurate
solution was proposed. Such an extension is not directly appli-
cable to the NFDTD, DSI, and GY methods because of the com-
plexity of the projection operations required to project the fields
normal to the grid faces onto the dual edges passing through the
faces.
In this paper, it is demonstrated that the source of late-time
instabilities in the NFDTD and DSI/GY methods is due to the
basic definitions for the projection operators. Consequently,
Manuscript received March 25, 1998; revised November 15, 1999. This
work was supported in part by the National Science Foundation under Grants
ECS-9624628 and CDA-9502645 and the Army Research Office under Grant
DAAH04-94-G-0243.
S. D. Gedney is with the Department of Electrical Engineering, University of
Kentucky, Lexington, KY 40506-0046 USA.
J. A. Roden is with IBM Personal Systems Group, IBM Corporation, Re-
search Triangle Park, NC 27709 USA.
Publisher Item Identifier S 0018-926X(00)01650-1.
these methods can lead to formulations that are unstable in the
late time independent of the time step. Alternative methods
based on the NFDTD and the DSI/GY algorithms are proposed
that are accurate and stable when employing quadrilateral prism
elements (i.e., elements that are orthogonal in the vertical di-
rection and irregular quadrilaterals in the horizontal direction).
II. GENERAL FORMULATION
The FDTD methods described herein assume a dual-stag-
gered grid with an edge-based discretization of Maxwell’s equa-
tions. Specifically, the electric field intensity is projected onto
the edges of a primary grid and the magnetic field intensity is
projected onto the edges of the secondary grid. Implementing
Maxwell’s curl equations in their integral form, the flux den-
sities normal to the primary and secondary grid faces are nat-
urally updated given the circulation of the dual field about the
faces. Before performing the field update, the normal flux den-
sity vector must be projected onto the dual edge passing through
the face. This is accomplished in the NFDTD algorithm [3] via
the local Jacobian tensor and local interpolation and a local in-
terpolation scheme in the DSI and GY methods [5], [8]. The de-
tails of these algorithms are not repeated here. Rather, the reader
is referred to the original articles as well as [9] for a detailed
summary of the algorithms.
Both the NFDTD and the DSI/GY methods result in explicit
update schemes that can be expressed in general form as a cou-
pled set of first-order difference equations
(1)
(2)
where and are vectors of the discrete vector flux densities, the
superscripts refer to discrete time, and represent the dis-
crete contour integrals of the electric and magnetic fields about
primary and secondary cell faces, respectfully, is a diagonal
matrix with entries representing the inverse of the relative per-
mittivity, and and are the projection matrices. Note that
for simplicity the domain is assumed to be lossless and nonmag-
netic. However, the following analyses are not limited by these
simplifications.
III. STABILITY ANALYSIS
The coupled difference equations in (1) and (2) are explicit
in nature and are conditionally stable. To derive a sufficient sta-
bility condition, the coupled difference equations are reposed as
a single first-order difference equation. To this end, first substi-
tute (1) into (2). Then, introducing the vector
(3)
0018–926X/00$10.00 © 2000 IEEE