rfx-fdtd 1.0.0

JAX-native 3D FDTD electromagnetic simulator for RF engineering

rfx

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

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

v1.0.0 -- Production-stable release with 260+ tests and 0.000% cross-validation agreement vs OpenEMS.

v1.0.0 Highlights

• Non-uniform z mesh via dz_profile for thin substrates
• SBP-SAT subgridding research path with JIT support
• Lumped RLC elements (series / parallel)
• Via and CurvedPatch geometry helpers
• Oblique-incidence TFSF plane-wave source
• Field animation export via save_field_animation
• Auto-configuration from geometry + frequency range
• 260+ tests with strong cavity/waveguide benchmarking and active antenna validation work

At a Glance

GPU-accelerated	~3,000 Mcells/s on RTX 4090 via jax.lax.scan JIT
Differentiable	jax.grad through full time-stepping for inverse design
Benchmarked	Strong agreement vs Meep/OpenEMS on core cavity and waveguide cases
Non-uniform mesh	Graded z-profiles for thin substrates without global refinement
Auto-configuration	auto_configure() derives dx, domain, CPML, timesteps from geometry
Lumped RLC	Series, parallel, and shunt lumped elements in any cell
Via & curved patch	First-class geometry primitives for PCB and antenna design
Animation	Time-domain field animation export (MP4/GIF)

Installation

# From source (recommended)
git clone https://github.com/BK3536/rfx.git
cd rfx
pip install -e ".[dev]"

GPU: Complete the JAX GPU setup for your platform before running large simulations.

Quick Start

from rfx import Simulation, Box, GaussianPulse
import numpy as np

# PEC cavity with dielectric slab — simulate up to 10 GHz
sim = Simulation(freq_max=10e9, domain=(0.048, 0.032, 0.032), boundary="pec")
sim.add_material("slab", eps_r=2.2)
sim.add(Box((0.016, 0, 0), (0.032, 0.032, 0.032)), material="slab")

# Lumped port + probe
sim.add_port(
    (0.006, 0.016, 0.016),
    "ez",
    impedance=50.0,
    waveform=GaussianPulse(f0=5e9, bandwidth=0.8),
)
sim.add_probe((0.042, 0.016, 0.016), "ez")

result = sim.run(num_periods=30)

# S11
s11_dB = 20 * np.log10(np.abs(result.s_params[0, 0, :]) + 1e-12)

Patch Antenna with Non-Uniform Mesh

from rfx import Simulation, Box, GaussianPulse, auto_configure
import numpy as np

# Auto-configure from geometry and frequency range
geometry = [
    (Box((0, 0, 0), (0.06, 0.06, 0.0016)), "fr4"),
    (Box((0, 0, 0), (0.06, 0.06, 0)), "pec"),
    (Box((0.015, 0.011, 0.0016), (0.0444, 0.049, 0.0016)), "pec"),
]
materials = {"fr4": {"eps_r": 4.4, "sigma": 0.025}, "pec": {"eps_r": 1.0, "sigma": 1e10}}
config = auto_configure(geometry, (1e9, 4e9), materials=materials, accuracy="standard")
print(config.summary())  # Shows dx, non-uniform dz, CPML, n_steps

# Build simulation from auto-config
sim = Simulation(**config.to_sim_kwargs())
sim.add_material("fr4", eps_r=4.4, sigma=0.025)
for shape, mat in geometry:
    sim.add(shape, material=mat)
sim.add_source((0.025, 0.03, 0.0008), "ez", waveform=GaussianPulse(f0=2.4e9))
sim.add_probe((0.025, 0.03, 0.0008), "ez")

result = sim.run(n_steps=config.n_steps)
modes = result.find_resonances(freq_range=(1.5e9, 3.5e9))
for m in modes:
    print(f"f={m.freq/1e9:.4f} GHz, Q={m.Q:.0f}")

Inverse Design

from rfx import Simulation, Box, GaussianPulse
from rfx.optimize import DesignRegion, optimize
from rfx.optimize_objectives import minimize_s11

sim = Simulation(freq_max=10e9, domain=(0.048, 0.032, 0.032), boundary="pec")
sim.add_port((0.006, 0.016, 0.016), "ez", impedance=50.0,
             waveform=GaussianPulse(f0=5e9, bandwidth=0.8))

region = DesignRegion(
    corner_lo=(0.016, 0.008, 0.008),
    corner_hi=(0.032, 0.024, 0.024),
    eps_range=(1.0, 4.4),
)

result = optimize(sim, region, minimize_s11, n_iters=40, lr=0.05)
print(f"Final loss: {result.loss_history[-1]:.4f}")

Key Features

Simulation Engine

• 3D/2D Yee solver with CFS-CPML absorbing boundaries
• True PEC mask (component-specific tangential zeroing)
• Non-uniform z mesh for thin substrates
• SBP-SAT subgridding research path
• Periodic boundaries and oblique-incidence TFSF plane-wave source
• Auto-configuration: auto_configure() derives dx, domain, CPML, timesteps, and source recommendation
• Lumped RLC elements (series and parallel topologies)

Geometry

• CSG primitives: Box, Sphere, Cylinder
• CurvedPatch for staircase-approximated curved antenna layouts
• Via for PCB through-hole / plated interconnect modeling

Sources & Ports

• GaussianPulse, ModulatedGaussian, CW, custom waveforms
• Soft sources and polarized sources
• Lumped ports and multi-cell wire ports (conductor-to-conductor)
• Waveguide ports with modal decomposition (eigenmode solver)
• Polarized sources (Jones vector: circular, LHCP, slant45)

Materials

• Dispersive: Debye, Lorentz/Drude
• Magnetic (mu_r validated)
• Thin conductors with subcell correction
• Subpixel smoothing (Farjadpour/Kottke)
• Built-in library: pec, fr4, rogers4003c, copper, alumina, ptfe, water_20c, and others

Analysis

• S-parameters: lumped, wire (JIT-DFT), waveguide (N-port matrix)
• Harminv resonance extraction (Matrix Pencil Method)
• NTFF far-field, radiation patterns, RCS
• Far-field polarization (axial ratio, tilt, sense)
• Touchstone I/O, HDF5 checkpoints, VTK export
• Time-domain field animation (MP4/GIF via save_field_animation)

Optimization

• Design regions with jax.grad through full simulation
• Adam optimizer, gradient checkpointing
• Objective library: minimize_s11, maximize_s21, target_impedance, maximize_bandwidth, maximize_directivity

Benchmark Snapshot

Core benchmark areas currently documented most strongly:

Geometry	rfx vs Meep	rfx vs OpenEMS	rfx vs Analytical
PEC cavity TM110	0.010%	0.000%	0.013%
FR4 cavity (eps=4.4)	0.006%	—	0.013%
Rogers (eps=3.55)	0.007%	—	0.026%
PTFE (eps=2.2)	0.004%	—	0.004%
Alumina (eps=9.8)	0.005%	—	0.038%

Patch/microstrip resonance workflows are also active, but feed/port interpretation should be documented more carefully than the cavity/waveguide benchmarks.

GPU Performance (RTX 4090)

Grid	Steps	Time	Throughput
23^3	200	0.087s	28 Mcells/s
33^3	300	0.086s	125 Mcells/s
43^3	400	0.103s	310 Mcells/s
63^3	500	0.095s	1,310 Mcells/s

Gradient (reverse-mode AD): ~0.31s for all grid sizes.

Documentation

Full documentation: remilab.ai/rfx and repo-local guides under docs/guide/.

• User Guide — Installation, API, materials, sources, probes
• AI Agent Guide — Auto-configuration, prompt templates, design workflows

Citation

@software{kim_rfx_2026,
  author       = {Byungkwan Kim},
  title        = {rfx: JAX-based differentiable 3D FDTD simulator},
  institution  = {REMI Lab, Chungnam National University},
  year         = {2026},
  url          = {https://github.com/BK3536/rfx}
}

License

MIT License. See LICENSE.

Acknowledgments

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

AI-assisted development: Claude (Anthropic), Codex (OpenAI).