rfx-fdtd 1.6.1

JAX-native 3D FDTD electromagnetic simulator for RF engineering

rfx

██████╗ ███████╗██╗  ██╗
██╔══██╗██╔════╝╚██╗██╔╝
██████╔╝█████╗   ╚███╔╝
██╔══██╗██╔══╝   ██╔██╗
██║  ██║██║     ██╔╝ ██╗
╚═╝  ╚═╝╚═╝     ╚═╝  ╚═╝

Differentiable 3D FDTD electromagnetic simulator for RF and microwave engineering — powered by JAX.

v1.5.0 package line + current main docs refresh — 500+ tests, GPU-benchmarked (7,309 Mcells/s on RTX 4090), published RF benchmarks, and clearer routing for practical nonuniform thin-substrate workflows.

Project status (April 2026): rfx is still in active validation and early conceptualization. Treat the current main branch as an initial-stage release rather than a finalized, fully qualified simulator; the support surface, validation evidence, and higher-level workflows will continue to be tightened and expanded in upcoming iterations.

At a Glance

Current correctness-bearing support is centered on the uniform Cartesian Yee reference lane. Non-uniform graded-z, distributed execution, and Floquet/Bloch workflows should be treated according to docs/guides/support_matrix.md, not as blanket guarantees.

GPU-accelerated	7,309 Mcells/s on RTX 4090, 5,249 on A6000 via jax.lax.scan JIT
Differentiable	jax.grad through full time-stepping for inverse design
Topology optimization	Density-based with filtering, projection, and beta continuation
Conformal PEC	Dey-Mittra method for 2nd-order accuracy on curved conductors
Multi-GPU	Single-host multi-GPU distributed FDTD with 1D slab decomposition (experimental lane)
Multi-mode ports	Analytical TE/TM eigenmodes for waveguide S-matrix
Floquet ports	Phased-array unit-cell analysis with Bloch periodic BC (experimental lane)
Non-uniform x/y/z profiles	Practical thin-substrate shadow lane with graded dz and per-cell dx/dy profiles
Accuracy validated	5-case benchmark against Balanis/Pozar (patch 1.97%, cavity 0.016%)
500+ tests	CI with ruff lint + pytest, 0 xfails

Current main highlights

• Nonuniform is now documented as a usable shadow lane, not as a vague half-supported footnote.
• Current practical anchors: examples/crossval/05_patch_antenna.py for the patch cross-check and examples/nonuniform_patch_demo.py for a quick repo-local thin-substrate demo.
• Validation owns claim strength: use the public validation pages for quantitative evidence and lane labels, then use examples pages for runnable entry points.

Installation

pip install rfx-fdtd

GPU support (JAX + CUDA):

pip install "jax[cuda12]" rfx-fdtd

For development:

git clone https://github.com/bk-squared/rfx.git
cd rfx && pip install -e ".[all]"

Quick Start

from rfx import Simulation, Box, GaussianPulse

# 2.4 GHz patch antenna on FR4
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  Q={modes[0].Q:.0f}")

GPU Performance (Measured)

GPU	VRAM	Peak Mcells/s (200^3)
RTX 4090	24 GB	7,309
RTX 3090	24 GB	5,847
RTX A6000	48 GB	5,249

Grid	RTX 4090	RTX 3090	A6000
50^3 (125K)	751	455	483
100^3 (1M)	5,332	2,502	2,410
150^3 (3.4M)	6,258	4,797	4,158
200^3 (8M)	7,309	5,847	5,249

Gradient (reverse-mode AD): ~3-4x forward pass with jax.checkpoint.

Accuracy Validation

Benchmarked against Balanis "Antenna Theory" and Pozar "Microwave Engineering":

Structure	Reference	Error
Patch antenna resonance	Balanis Ch 14	1.97%
WR-90 TE10 cutoff	Analytical	0.60%
Dielectric cavity TM110	Analytical	0.016%
Microstrip Z0	Hammerstad-Jensen	0.47%
Coupled-line filter	Pozar Ch 8	22.5% (formula limitation)

Cross-validation vs Meep/OpenEMS: 0.000-0.007% on cavities and waveguides. For the current practical patch workflow on the nonuniform shadow lane, use examples/crossval/05_patch_antenna.py together with the validation pages rather than treating the example itself as the top-level claim source.

Key Features

Core Simulator

• 3D/2D Yee FDTD with CFS-CPML (kappa=5.0, < -40 dB reflectivity)
• Conformal PEC via Dey-Mittra (2nd-order on curved surfaces)
• Non-uniform z-mesh for thin substrates (shadow qualification lane)
• Multi-GPU via jax.pmap (experimental distributed lane)
• Mixed precision (float16 fields, 2x memory reduction)
• Auto-configuration from geometry + frequency range

Sources & Ports

• GaussianPulse, ModulatedGaussian, CW, custom waveforms
• Lumped/wire ports, lumped RLC (series/parallel ADE)
• Multi-mode waveguide ports (analytical TE/TM eigenmodes)
• Floquet ports with Bloch periodic BC (experimental lane)
• Oblique TFSF (2D TMz + TEz auxiliary grids)

Materials

• Debye/Lorentz/Drude dispersive, Kerr nonlinear (chi3)
• Subpixel smoothing, thin conductor correction
• Material fitting: CSV import, Debye/Lorentz pole fitting
• Differentiable material fitting (jax.grad through FDTD)
• Library: pec, fr4, rogers4003c, copper, alumina, water_20c, ...

Analysis & Optimization

• S-parameters (lumped, wire, waveguide N-port)
• Harminv resonance extraction (MPM)
• Far-field, RCS, radiation patterns, polarization
• Antenna metrics: gain, efficiency, HPBW, F/B ratio, bandwidth
• Topology optimization (density filter + projection + beta schedule)
• Parametric sweep + jax.vmap batch evaluation
• Smith chart, de-embedding, Touchstone I/O (.s2p/.s4p/.snp)
• Auto convergence study with Richardson extrapolation

Geometry & Workflow

• Box, Sphere, Cylinder (CSG), Via, CurvedPatch
• PCB stackup builder (2-layer, 4-layer presets)
• RIS unit cell workflow
• Pre-AMR error indicator + neural surrogate export
• Field animation (GIF/MP4)
• Streamlit web dashboard

Documentation

Full documentation: remilab.ai/rfx

Repo-level support and reference-lane contract artifacts:

• docs/guides/support_matrix.md
• docs/guides/reference_lane_contract.md

Canonical public-doc sources in this repo:

• docs/public/index.mdx — public /rfx/ landing page
• docs/public/guide/ — public guide pages
• docs/public/examples/ — runnable example hubs
• docs/public/validation/ — quantitative evidence and lane-label hubs
• docs/agent/ — public AI-agent pages
• docs/guides/public_docs_architecture.md — ownership, sync, and deploy rules

Public docs maintenance workflow

1. Edit the source pages in docs/public/index.mdx, docs/public/guide/, or docs/agent/.
2. Validate the source tree:

   python scripts/check_public_docs_manifest.py
   

3. Export the updated snapshot to gitops:

   python scripts/export_public_docs_to_gitops.py
   

4. Keep the source repo CI in sync with .github/workflows/public-docs-source.yml.

Source-side CI for this flow lives in:

• .github/workflows/public-docs-source.yml

Gitops-side snapshot/build CI lives in the deploy repo:

• remilab-sites-gitops/.github/workflows/rfx-public-docs-sync.yml

Start here

• Public landing page
• Validation hub — quantitative evidence and lane labels
• Examples hub — current runnable entry points
• Examples showcase — bounded visual tour of the example surface
• Non-Uniform Mesh guide — practical thin-substrate shadow-lane workflows

Tutorials

• Patch Antenna Design — practical patch workflow from local resonance run to external cross-check
• Microstrip Filter — Coupled-line BPF from Pozar
• Convergence Study — Mesh independence methodology

Guides

• Migration from Meep/OpenEMS
• Changelog
• Contributing

Citation

@software{kim_rfx_2026,
  author       = {Byungkwan Kim},
  title        = {rfx: JAX-based differentiable 3D FDTD simulator for RF engineering},
  institution  = {REMI Lab, Chungnam National University},
  year         = {2026},
  version      = {1.5.0},
  url          = {https://github.com/bk-squared/rfx}
}

License

MIT License. See LICENSE.

Acknowledgments

Developed by Byungkwan Kim at the Radar & ElectroMagnetic Intelligence (REMI) Laboratory, Chungnam National University.

AI-Assisted Development

This project was developed with AI coding assistants orchestrated via oh-my-claudecode:

• Claude (Anthropic) — Architecture design, physics validation, cross-validation, code review, documentation
• Codex (OpenAI) — Feature implementation, test generation, review, debugging