• Embed Doc
  • Readcast
  • Collections
  • CommentGo Back
Download
 
9
Computational Electromagnetics:The Finite-DifferenceTime-Domain Method
Allen Taflove
Department of Electrical andComputer Engineering,Northwestern University,Chicago, Illinois, USA
Susan C. Hagness
Department of Electrical andComputer Engineering,University of Wisconsin,Madison, Wisconsin, USA
Melinda Piket-May
Department of Electrical andComputer Engineering,University of Colorado,Boulder, Colorado, USA
9.1 Introduction ..........................................................................................629 9.1.1 Background • 9.1.2 Rise of Finite-Difference Time-Domain Methods • 9.1.3 Characteristicsof FDTD and Related Space-Grid Time-Domain Techniques 9.1.4 Classes of Algorithms •9.1.5 Predictive Dynamic Range • 9.1.6 Scaling to Very Large Problem Sizes9.2 Maxwell's Equations ................................................................................633 9.2.1 Three-Dimensional Case • 9.2.2 Reduction to Two Dimensions9.3 The Yee Mgorithm ..................................................................................634 9.3.1 Basic Ideas • 9.3.2 Finite Differences and Notation • 9.3.3 Finite-Difference Expressionsfor Maxwell's Equations in Three Dimensions • 9.3.4 Reduction to the Two-Dimensional TMzand TEz Modes9.4 Numerical Dispersion ..............................................................................638 9.4.1 Introduction • 9.4.2 Two-Dimensional Wave Propagation 9.4.3 Extension to ThreeDimensions: Cartesian Yee Lattice • 9.4.4 Comparison with the Ideal Dispersion Case •9.4.5 Anisotropy of the Numerical Phase Velocity • 9.4.6 Complex-Valued NumericalWave Numbers9.5 Numerical Stability .................................................................................649 9.5.1 Introduction • 9.5.2 Complex-Frequency Analysis • 9.5.3 Examples of CalculationsInvolving Numerical Instability in a 1-D Grid 9.5.4 Example of Calculation Involving NumericalInstability in a Two-Dimensional Grid 9.5.5 Linear Instability When the Normalized CourantFactor Equals 19.6 Perfectly Matched Layer Absorbing Boundary Conditions ..............................655 9.6.1 Introduction • 9.6.2 Brief History • 9.6.3 Berenger's Perfectly Matched Layer •9.6.4 An Anisotropic PML Absorbing Medium 9.6.5 Theoretical Performance of the PML9.7 Examples of FDTD Modeling Applications ..................................................663 9.7.1 Global ELF Propagation in the Earth's Ionosphere Waveguide 9.7.2 High-Speed Electronics • 9.7.3 Microwave Penetration and Coupling •9.7.4 Ultrawideband Microwave Imaging for Early Stage Breast Cancer Detection •9.7.5 Photonic Integrated Circuits 9.7.6 Light Switching: Light in Femtoseconds9.8 Summary and Conclusions .......................................................................669 Bibliography ..........................................................................................669
9.1 Introduction
9.1.1
BackgroundPrior to about 1990, the modeling of electromagnetic engi-neering systems was primarily implemented using solutiontechniques for the sinusoidal steady-state Maxwell's equations.Before about 1960, the principal approaches in this area in-
Copyright © 2005 by Academic Press.All rights of reproduction in any form reserved.
volved closed-form and infinite-series analytical solutions,with numerical results from these analyses obtained usingmechanical calculators. After 1960, the increasing availabilityof programmable electronic digital computers permitted suchfrequency-domain approaches to rise markedly in sophistica-tion. Researchers were able to take advantage of the capabilitiesafforded by powerful new high-level programming languages629
 
630
Allen Taflove and Susan C. Hagness
such as Fortran, rapid random-access storage of large arrays ofnumbers, and computational speeds that were orders of mag-nitude faster than possible with mechanical calculators. In thisperiod, the principal computational approaches for Maxwell'sequations included the high-frequency asymptotic methods ofKeller (1962) as well as Kouyoumjian and Pathak (1974) andthe integral equation techniques of Harrington (1968).However, these frequency-domain techniques have difficul-ties and trade-offs. For example, while asymptotic analyses arewell suited for modeling the scattering properties of electricallylarge complex shapes, such analyses have difficulty treatingnonmetallic material composition and volumetric complexityof a structure. While integral equation methods can deal withmaterial and structural complexity, their need to construct andsolve systems of linear equations limits the electrical size ofpossible models, especially those requiring detailed treatmentof geometric details in a volume as opposed to just the surfaceshape.While significant progress has been made in solving theultra-large systems of equations generated by frequency-domain integral equations (Song and Chew 1998), the capabil-ities of even the latest such technologies are exhausted by manyvolumetrically complex structures of engineering interest. Thisalso holds true for frequency-domain finite element techniquesthat generate sparse rather than dense matrices. Further, thevery difficult incorporation of material and device nonlinea-rities into frequency-domain solutions of Maxwell's equationsposes a significant problem as engineers seek to design ac-tive electromagnetic/electronic and electromagnetic/quantum-optical systems such as high-speed digital circuits, microwaveand millimeter-wave amplifiers, and lasers.
9.1.2 Rise of Finite-Difference Time-DomainMethods
During the 1970s and 1980s, a number of researchers realizedthe limitations of frequency-domain integral equation solu-tions of Maxwell's equations. This led to early explorations of anovel alternative approach: direct time-domain solutions ofMaxwell's differential (curl) equations on spatial grids or lat-tices. The
finite-difference time-domain
(FDTD) method,introduced by Yee (1966), was the first technique in this classand has remained the subject of continuous development(Taflove and Hagness [2000]).There are seven primary reasons for the expansion of inter-est in FDTD and related computational solution approachesfor Maxwell's equations:1. FDTD uses no linear algebra. Being a fully explicitcomputation, FDTD avoids the difficulties with linearalgebra that limit the size of a frequency-domain inte-gral equation and finite-element electromagneticsmodels to generally fewer than 106 field unknowns.FDTD models with as many as 109 field unknownshave been run. There is no intrinsic upper bound tothis number.2. FDTD is accurate and robust. The sources of error inFDTD calculations are well understood and can bebounded to permit accurate models for a very largevariety of electromagnetic wave interaction problems.3. FDTD treats impulsive behavior naturally. Being a time-domain technique, FDTD directly calculates the im-pulse response of an electromagnetic system. Therefore,a single FDTD simulation can provide either ultrawide-band temporal waveforms or the sinusoidal steady-stateresponse at any frequency in the excitation spectrum.4. FDTD treats nonlinear behavior naturally. Being atime-domain technique, FDTD directly calculates thenonlinear response of an electromagnetic system.5. FDTD is a systematic approach. With FDTD, specifyinga new structure to be modeled is reduced to a problemof mesh generation rather than the potentially complexreformulation of an integral equation. For example,FDTD requires no calculation of structure-dependentGreen's functions.6. Computer memory capacities are increasing rapidly.Although this trend positively influences all numericaltechniques, it is of particular advantage to FDTDmethods that are founded on discretizing space over avolume and that inherently require a large randomaccess memory.7. Computer visualization capabilities are increasing rap-idly. Although this trend positively influences all nu-merical techniques, it is of particular advantage toFDTD methods that generate time-marched arrays offield quantities suitable for use in color videos to illus-trate field dynamics.An indication of the expanding level of interest in FDTDMaxwell's equations solvers is the hundreds of papers currentlypublished in this area worldwide each year, as opposed tofewer than ten as recently as 1985 and prior to that year(Shlager and Schneider, 1998). This expansion continues asengineers and scientists in nontraditional electromagnetics-related areas such as digital systems and integrated opticsbecome aware of the power of such direct solution techniquesfor Maxwell's equations.
9.1.3 Characteristics of FDTD and RelatedSpace-Grid Time-Domain Techniques
FDTD and related space-grid time-domain techniques aredirect solution methods for Max'well's curl equations. Thesemethods employ no potentials. Rather, they are based onvolumetric sampling of the unknown electric and magneticfields in and surrounding the structure of interest over a periodof time. The sampling in space is at subwavelength resolutionset by the user to properly sample the highest near-field spatial
 
9 Computational Electromagnetics: The Finite-Difference Time-Domain Method
631frequencies thought to be important in the physics of theproblem. Typically, 10 to 20 samples per wavelength areneeded. The sampling time is selected to ensure numericalstability of the algorithm.Overall, FDTD and related techniques are marching-in-timeprocedures that simulate the continuous actual electromag-netic waves in a finite spatial region by sampled-data numer-ical analogs propagating in a computer data space. Time-stepping continues as the numerical wave analogs propagatein the space lattice to causally connect the physics of themodeled region. For simulations where the modeled regionmust extend to infinity, absorbing boundary conditions(ABCs) are employed at the outer lattice truncation planes,which ideally permit, all outgoing wave analogs to exit theregion with negligible reflection. Phenomena such as induc-tion of surface currents, scattering and multiple scattering,aperture penetration, and cavity excitation are modeled time-step by time-step by the action of the numerical analog to thecurl equations. Self-consistency of these modeled phenomenais generally ensured if their spatial and temporal variations arewell resolved by the space and time sampling process. In fact,the goal is to provide a self-consistent model of the mutualcoupling of all of the electrically small volume cells constitut-ing the structure and its near field, even if the structure spanstens of wavelengths in three dimensions and there are hun-dreds of millions of space cells.Time-stepping is continued until the desired late-time pulseresponse is observed at the field points of interest. For linearwave interaction problems, the sinusoidal response at thesefield points can be obtained over a wide band of frequenciesby discrete Fourier transformation of the computed fieldversus time waveforms at these points. Prolonged "ringing"of the computed field waveforms due to a high Q-factor orlarge electrical size of the structure being modeled requires acombination of extending the computational window in timeand extrapolating the windowed data before Fourier trans-formation.
9.1.4 Classes of Algorithms
Current FDTD and related space-grid time-domain algorithmsare fully explicit solvers employing highly vectorizable andparallel schemes for time-marching the six components ofthe electric and magnetic field vectors at each of the spacecells. The explicit nature of the solvers is usually maintained byemploying a leapfrog time-stepping scheme. Current methodsdiffer primarily in how the space lattice is set up. In fact,gridding methods can be categorized according to the degreeof structure or regularity in the mesh cells.
Almost Completely Structured
In this case, the space lattice is organized so that its unit cellsare congruent wherever possible. The most basic example ofsuch a mesh is the pioneering work of Yee (1966), whoemployed a uniform Cartesian grid having rectangular cells.Staircasing was used to approximate the surface of structuralfeatures not parallel to the grid coordinate axes. Later workshowed that it is possible to modify the size and shape of thespace cells located immediately adjacent to a structural featureto conformally fit its surface (Jurgens
et al.,
1992; Dey andMittra, 1997). This result is accurate and computationallyefficient for large structures because the number of modifiedcells is proportional to the surface area of the structure. Thus,the number of modified cells becomes progressively smallerrelative to the number of regular cells filling the structurevolume as its size increases. As a result, the computer resourcesneeded to implement a fully conformal model approximatethose required for a staircased model. A key disadvantage ofthis technique, however, is that special mesh-generation soft-ware must be constructed.
Surface-Fitted
In this case, the space lattice is globally distorted to fit theshape of the structure of interest. The lattice can be dividedinto multiple zones to accommodate a set of distinct surfacefeatures (Shankar
et al.,
1990). The major advantage of thisapproach is that well-developed mesh-generation software ofthis type is available. The major disadvantage, relative to theYee algorithm, is the substantial added computer burden due
to:
Memory allocations for the position and stretchingfactors of each cellExtra computer operations to implement Maxwell's equa-tions at each cell and to enforce field continuity at theinterfaces of adjacent cellsAnother disadvantage is the possible presence of numericaldissipation in the time-stepping algorithm used for suchmeshes. This can limit the range of electrical size of the structurebeing modeled due to numerical wave-attenuation artifacts.
Completely Unstructured
In an unstructured case, the space containing the structure ofinterest is completely filled with a collection of lattice cells ofvarying sizes and shapes but conforms to the structure surface(Madsen and Ziolkowski, 1990). As for the case of surface-fitted lattices, mesh-generation software is available andcapable of modeling complicated three-dimensional shapespossibly having volumetric inhomogeneities. A key disadvan-tage of this approach is its potential for numerical inaccuracyand instability due to the unwanted generation of highlyskewed space cells at random points in the lattice. A seconddisadvantage is the difficulty in mapping the unstructuredmesh computations onto the architecture of either parallelvector computers or massively parallel machines. Thestructure-specific irregularity of the mesh mandates a robustpreprocessing algorithm that optimally assigns specific meshcells to specific processors.
of 00

Leave a Comment

You must be to leave a comment.
Submit
Characters: ...
You must be to leave a comment.
Submit
Characters: ...