rfx

Differentiable FDTD for RF/Microwave — powered by JAX

rfx is a 3-D finite-difference time-domain (FDTD) electromagnetic simulator built in JAX. These docs focus on maintained user workflows with explicit support boundaries.

Current public status

Use the uniform Cartesian Yee RF/FDTD lane first. The public docs highlight maintained workflows and explicitly bounded claim envelopes:

Surface	Public role
Uniform Yee RF workflows	recommended default for tutorials, examples, probes, Harminv, selected S-parameter workflows, and benchmarked far-field workflows
Rectangular waveguide S-matrices	validated inside the documented rectangular-guide envelope
Lumped, wire, and microstrip-line port workflows	available through their matching calculators; validation scope follows the support matrix
Coaxial line reflection	bounded one-port transmission-line workflow only; not a general coaxial S-matrix promise
Differentiable design loops	use proxy objectives, gradient checks, and documented validation before treating a design result as evidence

If a capability is not listed here or in the support-boundary page, treat it as outside the documented public support scope until the support matrix lists it.

Why rfx

Feature	Detail
GPU acceleration	JAX/JIT execution for large 3-D grids
Differentiable simulation	`jax.grad` through time-domain workflows for inverse design
RF workflow tools	materials, sources, probes, ports, S-parameter helpers, Harminv, far-field utilities
Recommended examples	current runnable scripts under `examples/crossval/` and selected inverse-design examples
Support-boundary discipline	public claims are tied to explicit guide/support-matrix envelopes, not to every importable symbol

Quick install

pip install rfx-fdtd

GPU support depends on your JAX/CUDA environment:

pip install "jax[cuda12]" rfx-fdtd

Ten-line example

from rfx import Simulation, Box, GaussianPulse

sim = Simulation(freq_max=4e9, domain=(0.08, 0.06, 0.025), boundary="cpml")
sim.add(Box((0.0, 0.0, 0.0), (0.08, 0.06, 0.0016)), material="fr4")
sim.add(Box((0.02, 0.01, 0.0016), (0.049, 0.049, 0.0016)), material="pec")
sim.add(Box((0.0, 0.0, 0.0), (0.08, 0.06, 0.0)), material="pec")
sim.add_source((0.029, 0.03, 0.0008), "ez",
               waveform=GaussianPulse(f0=2.4e9, bandwidth=0.8))
sim.add_probe((0.029, 0.03, 0.0008), "ez")

result = sim.run(n_steps=8000)
modes = result.find_resonances(freq_range=(1.5e9, 3.5e9))
print(f"Resonance: {modes[0].freq/1e9:.4f} GHz")

Documentation by topic

Getting Started

Installation — Python/JAX install, GPU notes, dev setup
Quick Start — first simulation with the current high-level API
Your First Patch Antenna — current-source resonance workflow

Modeling & Setup

Simulation API — Simulation, Result, auto_configure, ports, NTFF, and design helpers
Materials & Geometry — library materials, Debye/Lorentz, CSG shapes, PCB stackup
Sources & Ports — point sources, lumped/wire ports, microstrip-line ports, waveguide ports, and bounded coaxial-line reflection
Probes & S-Parameters — DFT probes, S-matrix helpers, Harminv, de-embedding, exports
Memory Reduction — reduce FDTD/AD memory while preserving validation boundaries
Waveguide Ports — rectangular waveguide modal workflows

Analysis & Validation

Cross-Validation & Accuracy — short public validation overview
Benchmarks — quantitative benchmark summary
Convergence Study — mesh refinement workflow
Far-Field & RCS — NTFF radiation patterns and scattering workflows
Visualization & Analysis — plotting, exports, post-processing

Design & Optimization

Autodiff and Adjoint Background — Meep-informed gradient concepts for microwave engineers
Inverse Design — autodiff-driven optimization with documented proxy objectives
Gradient Behavior — where gradients are reliable vs noisy
Parametric Sweeps — sequential sweep and jax.vmap design-space exploration
Patch Antenna Design — rectangular patch workflow

Secondary hubs

Examples — recommended public runnable paths
Validation — public support and validation overview
API — curated public API contract