FDTD Method for High-Frequency EM Analysis

The core difference is how they handle frequency. FEM is typically a frequency-domain method — you solve at each frequency point separately. FDTD steps the electromagnetic field forward in time, directly discretizing Maxwell's equations. The huge advantage is that a single FDTD run gives you the broadband frequency response all at once — say, from 1 GHz to 100 GHz — by Fourier-transforming the time waveforms. That's why it's the workhorse for antenna design, EMC analysis, and radar cross-section (RCS) calculations.

Yes — FDTD uses a Cartesian (orthogonal) grid, so curved surfaces get staircase-approximated. Fine geometric features or thin structures require a finer grid, which can consume enormous memory for large models. FEM handles arbitrary geometries more elegantly. The choice really depends on problem type: FDTD excels at wideband open-region problems; FEM is better for resonant structures with complex geometry. In practice, FDTD dominates for antenna design, packaging, and EMC. FEM dominates for cavity filters, connectors, and waveguide discontinuities.

It depends on the geometry and frequency. For a simple patch antenna at 2.4 GHz with a uniform mesh of Δx = 1 mm, the staircase error is usually acceptable — resonant frequency errors of less than 1%. But for a highly curved body like a human head phantom in SAR (specific absorption rate) studies, or a capacitor with very thin dielectric, staircase approximation can cause significant errors. Techniques like subcell modeling, conformal FDTD (which uses a locally deformed cell to better match curved boundaries), and the Dey-Mittra algorithm address this. CST's FIT (Finite Integration Technique) is essentially FDTD with conformal boundary treatment built in.

FDTD uses the two curl equations of Maxwell:

Yee (1966) proposed a brilliant staggered grid: the electric field $\mathbf{E}$ and magnetic field $\mathbf{H}$ components are offset from each other by half a grid cell in space, and also offset by half a time step. This leapfrog arrangement means E and H update each other alternately, maintaining second-order accuracy in both space and time with almost no overhead. Specifically, $E_x$ components sit at $(i+\frac{1}{2}, j, k)$, while $H_x$ components sit at $(i, j+\frac{1}{2}, k+\frac{1}{2})$ — offset in both y and z directions.

Exactly. The leapfrog in time combined with the staggered spatial arrangement makes both the spatial and temporal derivatives second-order accurate using simple centered finite differences. The update equation for $E_x$ looks like:

This simple structure means the entire FDTD algorithm requires only floating-point additions and multiplications — no matrix assembly, no linear system solve. That simplicity is what makes GPU acceleration so effective for FDTD, with CUDA implementations achieving 50–100× speedup over single-core CPU.

Definitely not free — FDTD is an explicit method, and the Courant-Friedrichs-Lewy (CFL) condition must be satisfied or the simulation will diverge:

For a uniform 3D grid this simplifies to $\Delta t \le \Delta x / (c\sqrt{3})$. When you refine the mesh to capture finer detail, the time step shrinks proportionally — and the total simulation time grows rapidly. This is FDTD's main computational bottleneck for large, fine-featured models. In practice, $\Delta t = 0.99 \times \Delta t_{CFL}$ is commonly used as a safety margin.

The simulation will diverge exponentially — field values grow without bound, typically visible as NaN or Inf in the output after a few hundred time steps. The global time step is constrained by the smallest cell in the entire domain. That's why local mesh refinement (sub-gridding) is so valuable: you'd like to use a fine grid only around a via or trace corner, but a fine local cell forces the global time step down. Commercial codes like CST use local time stepping or sub-gridding schemes to mitigate this. The Dey-Mittra conformal algorithm also helps by allowing slightly larger cells near curved boundaries without staircase artifacts.

In an ideal electromagnetic wave, phase velocity is constant and equal to $c$. In FDTD, the numerical phase velocity depends on the propagation direction and frequency — this is numerical dispersion. At the grid sampling limit (2 cells per wavelength), the error is large. The dispersion relation for 3D FDTD is:

To keep dispersion error below 1%, use at least 10 cells per wavelength at the highest frequency. 20 cells per wavelength is the safer guideline for narrow-band resonance studies. The dispersion error grows quadratically with cell size, so halving the cell size reduces dispersion error by 4×.

That's where the Perfectly Matched Layer (PML) comes in — introduced by Bérenger in 1994. You surround the computational domain with an artificial absorbing layer. The key property is that PML is theoretically reflection-free regardless of incidence angle or frequency. It works by applying complex coordinate stretching to the material parameters:

In practice, $\sigma_\xi$ is ramped up as a quadratic or cubic profile from the domain edge into the PML, gradually absorbing the outgoing wave. A thickness of 8–16 cells is typical for good absorption. The Uniaxial PML (UPML) formulation is now standard in most codes, as it handles anisotropic media and avoids some numerical stability issues of the original split-field formulation.

Three main causes: (1) Discrete PML — the theoretical PML is reflectionless in the continuous case, but the FDTD discretization introduces a small residual reflection that decreases with finer mesh. (2) Evanescent waves — near-field evanescent components have different impedance characteristics and can reflect slightly more from PML. (3) Insufficient separation from radiating structures — PML should be placed at least $\lambda/4$ away from antennas or scatterers. Increasing the PML thickness from 8 to 16 cells and using a cubic rather than quadratic $\sigma$ profile typically reduces reflections from −60 dB to −80 dB.

The main source types are:

• Hard source: Force the E-field to a specified value at the source point. Simple but reflects impinging waves — creates a spurious "hard wall" that corrupts S-parameter measurements.
• Soft source: Add the source field to whatever E-field exists. Properly handles reflections at the source port — the standard choice for S-parameter extraction.
• Total/Scattered field formulation (TF/SF): Splits the domain into a total-field region containing both incident and scattered fields, and a scattered-field region outside. Essential for plane-wave excitation (RCS analysis, EMC susceptibility testing).
• Waveguide port: Excites a specific mode (TE10, TEM, etc.) at the computational boundary — needed for feeding waveguides and microstrip lines.

For antenna S-parameter extraction, use a soft lumped port at the feed point. For plane-wave scattering (like SAR analysis), use TF/SF with a Gaussian pulse.

No — that's the beauty of FDTD. You excite the structure with a broadband pulse (a Gaussian pulse or modulated Gaussian), which contains energy across all frequencies of interest. You record the time-domain waveform at observation points and apply an FFT to get the spectrum. For S-parameters, you compute the ratio of scattered to incident FFTs. To save memory, most modern FDTD codes do an on-the-fly DFT during the time-stepping loop — accumulating the frequency-domain result incrementally without storing the entire time waveform.

The simulation must run until all transients have decayed — specifically, until the total stored energy in the domain has dropped to below some threshold, typically −40 to −60 dB relative to the peak energy. If you stop early, the time window truncates the field, introducing a spectral leakage error in the FFT — analogous to truncating a sine wave mid-period. The result looks like ripple artifacts in the S-parameter response. For high-Q resonant structures like cavity filters, the energy rings for a very long time, so early stopping errors are especially severe. Automatic time-domain windowing (Hanning, exponential) can reduce leakage if you must terminate early, at the cost of some frequency resolution.

Patch antennas are a classic FDTD benchmark. Here's what matters most:

1. Cell size: Target $\lambda/10$ to $\lambda/20$ at the highest frequency. At 2.4 GHz in FR4 ($\varepsilon_r = 4.2$), wavelength is about 62 mm, so $\Delta x \approx 3$–6 mm for the free-space regions and smaller near the patch itself.
2. Substrate meshing: The dielectric shortens the wavelength by $1/\sqrt{\varepsilon_r}$, so you need finer cells inside the dielectric. A common approach: fine uniform mesh throughout, set to $\lambda_{eff}/10$ in the substrate.
3. Air gap before PML: Ensure at least $\lambda/4$ of free space between the antenna and the PML boundary. Use 8–16 PML cells.
4. Excitation source: Place a soft voltage source at the feed point (microstrip feed or coaxial probe). Record the E-field time waveform at that point.
5. Termination criterion: Stop when the total field energy drops below −40 dB of the peak. Cutting the simulation early corrupts the frequency response.
6. Near-to-far-field transform: If you need radiation patterns, add an NF-FF transformation surface surrounding the antenna, ensuring it encloses the complete antenna and is outside the reactive near-field region ($r > 2D^2/\lambda$).

The most common culprit for over-predicted resonant frequency in patch antenna simulations is an inaccurate substrate permittivity. FR4's $\varepsilon_r$ varies with frequency (dispersion), with values ranging from 4.1 to 4.7 depending on the PCB supplier and glass weave style. A second cause is insufficient mesh density — the staircase approximation of the rectangular patch edges effectively shortens the electrical length. Run a mesh convergence study: refine by 2× in each direction and check if the resonant frequency shifts. If it does, your mesh was too coarse. If not, suspect the material data.

Here's an overview of the major options:

FIT (Finite Integration Technique) was proposed by Thomas Weiland in 1977. Instead of discretizing Maxwell's differential equations, FIT works with the integral forms — Faraday's and Ampère's laws applied to each cell face. On a regular Cartesian grid, FIT reduces exactly to classical FDTD. The advantage appears on non-orthogonal grids: CST's Perfect Boundary Approximation (PBA) uses FIT to handle curved geometries without staircase error, achieving the accuracy of a much finer Cartesian mesh. This is why CST tends to give more accurate results on curved structures than codes using simple staircase FDTD at the same cell size.

Exactly — at 5G mmWave bands (28 GHz, 60 GHz), FDTD is critical because FEM frequency-domain solvers become extremely expensive. Here are the key trends:

• GPU-accelerated FDTD: CUDA-based implementations deliver 10–50× speedup over CPU. CST and Remcom both have mature GPU backends; for 60 GHz array antenna analysis with thousands of elements, GPU FDTD is often the only practical approach.
• Sub-gridding: Refine the grid locally around fine features (a via, bond wire, or IC pad), using a finer time step only in that region. Reduces total cell count and memory significantly.
• PINN-FDTD: Physics-informed neural networks correct errors from coarse grids, maintaining accuracy without full refinement — an active research direction for reducing simulation cost in design-space exploration.
• Quantum device simulation: Extending FDTD to analyze electromagnetic fields in quantum dots and photonic crystals for quantum communication hardware.
• FDTD for 3D-IC co-simulation: As chiplet architectures bring RDL (redistribution layers) and through-silicon vias into the signal path, FDTD is being applied to full 3D-IC electromagnetic co-simulation.

Yes — dispersive materials need auxiliary differential equations (ADE) or recursive convolution to implement frequency-dependent permittivity $\varepsilon(\omega)$ in the time domain. The Debye, Drude, and Lorentz models are the most common. For silicon ($n \approx 3.5$, $\kappa \sim 0$ in the THz range), the Sellmeier model is used. For biological tissue (with both permittivity and conductivity frequency-dependent), multi-pole Debye models are standard. Implementing them in FDTD adds auxiliary field variables but keeps the explicit, local update structure intact — a key advantage over implicit methods.

The classic problems, in order of frequency, are:

• Simulation divergence from CFL violation: Check that your chosen time step is below the CFL limit. If you changed cell size, recalculate.
• PML reflections corrupting the response: Increase PML thickness from 8 to 16 cells. Increase the separation between radiating structures and PML.
• Numerical dispersion causing phase errors: Use at least 10–20 cells per wavelength. Run a dispersion analysis to quantify the expected phase velocity error.
• Staircase error on curved boundaries: Refine locally, use conformal FDTD or FIT, or switch to FEM for the relevant structure.
• Early termination artifacts in S-parameters: Always monitor field energy decay; only terminate when energy drops below −40 dB. Apply Prony or ARMA extrapolation methods if rings are too slow to decay.

A resonance that's higher in frequency and sharper than measurement typically suggests two things: underestimated loss and slightly wrong effective permittivity. Check: (1) Is the substrate's $\tan\delta$ (loss tangent) set correctly? FR4's $\tan\delta$ is typically 0.015–0.025 — omitting it makes resonances artificially sharp. (2) Is the copper conductivity set to $5.8 \times 10^7$ S/m? A perfect conductor gives a slightly different resonance from real copper. (3) Is the substrate thickness modeled accurately — within ±5%? These three material parameters together usually account for the discrepancy. Detailed diagnostic steps are in the troubleshooting guide.