Electromagnetic Theory

Computational Electromagnetics

Q: When should I use FEM versus MoM versus FDTD for an RF problem?

Match the method to the geometry and the question. The method of moments (MoM) excels for open-region radiators such as wire antennas, arrays, and metallic scatterers because it meshes only conducting surfaces and enforces the radiation condition automatically through the free-space Green's function. The finite element method (FEM) is the strongest choice for enclosed, inhomogeneous, or highly resonant structures such as waveguide filters, dielectric resonators, and connector transitions, where its tetrahedral mesh conforms to curved geometry. Finite-difference time-domain (FDTD) is ideal for wideband responses, transient and pulse problems, and large dispersive or biological volumes because a single time-domain run yields the full frequency sweep via a Fourier transform. Many commercial tools hybridize these methods, for example FEM with a MoM boundary, to combine their strengths.

Q: Why does my CEM simulation produce inaccurate S-parameters at the band edges?

Band-edge inaccuracy is almost always a meshing or convergence problem rather than a physics error. The classic rule is at least 10 to 20 cells or mesh edges per wavelength at the highest analyzed frequency, but resonant structures, sharp metal edges, and thin dielectric layers demand local refinement well beyond that. For FDTD, the Courant condition limits the time step, and insufficient absorbing boundary layers (typically 8 to 12 PML cells) cause reflections that corrupt the band edges. For FEM, run an adaptive mesh refinement that tightens the delta-S convergence target to 0.01 or 0.02 and verify the result is mesh-independent by re-running at higher resolution. Always place ports and lumped elements far enough from discontinuities that evanescent modes have decayed.

Q: How much memory and runtime does a full-wave CEM solve require?

Resource cost scales sharply with electrical size and method. MoM produces a dense matrix, so memory grows as N-squared and direct-solve time as N-cubed, where N is the number of surface unknowns; a moderate array can reach tens of gigabytes, which is why fast methods such as the multilevel fast multipole method (MLFMM) reduce scaling to roughly N log N. FEM and FDTD produce sparse systems with lower per-unknown cost but require volumetric meshing, so memory grows with the cube of the electrical dimension. A typical 5G antenna module solve runs from minutes to a few hours on a workstation, while an electrically large platform such as a full aircraft for radar cross section may need a GPU cluster and many hours. Frequency sweeps amplify cost for MoM and FEM but are nearly free for FDTD.

/kom-pyoo-TAY-shuhn-uhl ih-lek-troh-mag-NET-iks/

Often abbreviated CEM, this discipline solves Maxwell's equations numerically on a discretized geometry so that the fields, currents, and S-parameters of an RF structure can be predicted before any metal is cut. Rather than relying on closed-form formulas that exist only for simple shapes, CEM divides a device into thousands to millions of mesh elements and solves the resulting linear system. The three dominant solver families are the finite element method, the method of moments, and finite-difference time-domain. Modern full-wave tools routinely model antennas, waveguide filters, connector launches, and entire phased-array modules from sub-GHz through millimeter-wave bands, with accuracy limited mostly by mesh density and material data rather than by the physics itself.

Category: Electromagnetic Theory

Core Methods: FEM, MoM, FDTD

Mesh Density: 10 to 20 / λ

From Maxwell's Equations to a Solvable Matrix

Every CEM technique begins with the same physics: the four Maxwell curl and divergence equations together with the constitutive relations of the materials present. Because analytic solutions exist only for canonical geometries such as infinite parallel plates or perfect spheres, real RF hardware demands a numerical approach. The solver discretizes either the volume or the surface of the structure into a mesh, expands the unknown fields or currents in a finite set of basis functions, and applies boundary conditions to reduce the continuous problem to a large but finite system of linear equations. Solving that system yields the field distribution, from which engineers extract S-parameters, radiation patterns, input impedance, and loss.

The choice among methods is fundamentally a choice of where to spend computational effort. Surface methods like the method of moments mesh only conductors and use a Green's function that already satisfies the radiation condition, making them efficient for open antenna problems but producing a dense matrix. Volume methods like FEM and FDTD mesh the whole space including dielectrics and air, producing sparse matrices that handle inhomogeneous, enclosed, and resonant structures gracefully. FDTD is unique in marching the fields forward in time, so a single broadband pulse excitation captures the entire frequency response through a transform.

Accuracy in CEM is governed less by the underlying equations and more by discretization. A mesh that is too coarse aliases short-wavelength features and shifts resonant frequencies; a mesh that is too fine wastes memory and runtime. Adaptive refinement, where the solver iteratively adds elements wherever the field error is largest, has become the standard way to reach a converged, mesh-independent answer without manual tuning.

Governing Relations and Discretization

Source-free vector wave (Helmholtz) equation:
∇ × (∇ × E) − k₀²ε_rE = 0, k₀ = ω√(μ₀ε₀)

Method of moments matrix form:
[Z]·[I] = [V], where Z_mn = ⟨f_m, L(f_n)⟩

FDTD update (Yee scheme, Courant limit):
Δt ≤ 1 / (c√(1/Δx² + 1/Δy² + 1/Δz²))

Meshing rule of thumb: cell size ≈ λ_min / 10 to λ_min / 20

Where E = electric field, k₀ = free-space wavenumber, ε_r = relative permittivity, [Z] = impedance matrix, L = integro-differential operator, c = speed of light, and λ_min is the wavelength at the highest analyzed frequency.

Solver Method Comparison

Method	Domain	Mesh	Matrix	Memory Scaling	Best For
Method of Moments (MoM)	Frequency	Surface	Dense	O(N²), MLFMM O(N·logN)	Wire and aperture antennas, arrays, RCS
Finite Element (FEM)	Frequency	Volume (tetrahedral)	Sparse	~O(N^1.5)	Waveguide filters, connectors, resonators
FDTD	Time	Volume (Yee grid)	Explicit, matrix-free	O(N)	Wideband, transient, dispersive volumes
Transmission-Line Matrix	Time	Volume (structured)	Explicit	O(N)	EMC, shielding, broadband structures
Hybrid FEM-MoM	Frequency	Volume + surface	Mixed	Problem dependent	Antenna-on-platform, radomes

Common Questions

Frequently Asked Questions

When should I use FEM versus MoM versus FDTD for an RF problem?

Match the method to the geometry. MoM meshes only conductors and handles open radiators (wire antennas, arrays, scatterers) efficiently via the free-space Green's function. FEM, with its conforming tetrahedral mesh, is strongest for enclosed, inhomogeneous, resonant structures such as waveguide filters and connector transitions. FDTD suits wideband, transient, and large dispersive problems because one time-domain run yields the full sweep. Commercial tools often hybridize, for example FEM with a MoM boundary.

Why does my CEM simulation produce inaccurate S-parameters at the band edges?

Band-edge error is almost always meshing or convergence, not physics. Use at least 10 to 20 cells per wavelength at the top frequency, with local refinement at sharp metal edges and thin dielectrics. In FDTD, respect the Courant time step and use 8 to 12 PML cells to stop boundary reflections. In FEM, run adaptive refinement to a delta-S target of 0.01 to 0.02 and confirm mesh independence. Keep ports far enough from discontinuities that evanescent modes have decayed.

How much memory and runtime does a full-wave CEM solve require?

Cost scales with electrical size and method. MoM is dense: memory grows as N² and direct solve as N³, so fast methods like MLFMM cut this to roughly N·logN. FEM and FDTD are sparse but volumetric, so memory grows with the cube of the electrical dimension. A 5G antenna module runs from minutes to a few hours on a workstation; an electrically large platform for radar cross section may need a GPU cluster and many hours. Sweeps are costly for MoM and FEM but nearly free for FDTD.

Related Terms

← Computational Dosimetry Compute-and-Forward →