rfx

A differentiable 3-D electromagnetic simulator for RF and microwave engineering, built in JAX.

See waves propagate, measure a component's response, and differentiate a design objective. Start with the visual manual and examples. The website identifies its source version; select the release matching your installation.

Release status: 2.0

rfx 2.1.0 is the current release. What 2.0 changes for a user:

• An input a solver path does not implement is refused, never dropped. If run(), forward() or an S-parameter calculator routes your model to a path that does not solve a declared input (a conformal wall, subpixel smoothing, a lumped element, a port, a dispersive material, a solver or precision request), it raises before the first time step and names the input. It does not run a different model silently.
• Validation cases are judged on criteria stated in advance, on a converged mesh: magnitudes within 2 dB, resonances and cutoffs within 1 %, group delay within 1 % of the reference. These are the criteria the project tests against, not a guarantee for every model. What is supported, and within which limits, is in the support matrix; what is currently wrong is in known limitations.
• Python 3.11 or newer, with JAX and jaxlib from 0.10.2 up to, not including, 0.11.2: pip installs 0.11.1 on Python 3.12 or newer and 0.10.2 on Python 3.11 (jaxlib 0.11.2 and 0.12.0 run the CPU time step 2.3–3× slower). Python 3.10 users stay on 1.x (1.7.0 on PyPI, 1.8.0 on the GitHub releases).
• Multi-device runs are supported within the limits in the support matrix. run(devices=...) and forward(distributed=True) cut the grid into equal slabs along x and are compared with a one-device run of the same model. Lumped RLC elements and surface-impedance sheets are refused on multi-device runs in 2.0 (planned for 2.1), and the support matrix lists the other inputs these runs refuse by name.

Install

pip install "rfx-fdtd>=2.0"

For GPU execution, use a compatible JAX/CUDA installation:

pip install "jax[cuda12]" "rfx-fdtd>=2.0"

For the development version, clone this repository and run pip install -e ".[dev]". The installation guide covers optional packages.

First run

A pulse in a small closed PEC box checks your installation. It records one field probe; it does not measure a device's S-parameters, resonance accuracy, or Q. This code is generated from hello_world.py, the same runnable example tested by CI.

from rfx import GaussianPulse, Simulation

sim = Simulation(
    freq_max=10e9,
    domain=(0.02, 0.02, 0.02),
    dx=2e-3,
    boundary="pec",
)
sim.add_source(
    (0.01, 0.01, 0.01),
    "ez",
    waveform=GaussianPulse(f0=5e9, bandwidth=0.8),
    amplitude_kind="current",
)
sim.add_probe((0.014, 0.01, 0.01), "ez")

preflight = sim.preflight()
print(preflight.format())
preflight.raise_for_failure()
n_steps = 120
result = sim.run(n_steps=n_steps, compute_s_params=False)
print(result.time_series.shape)

All lengths are metres and frequencies are hertz. Read the preflight findings. Continue with building and measuring a model.

Accuracy and development

Support depends on the structure, mesh, boundary and observable. A successful run or an agreeing automatic/finite-difference gradient is not a convergence study. The accuracy guide explains the checks; benchmarks link results to their scope.

To change rfx itself, read Contributing to rfx. Release changes are in CHANGELOG.md.

Citation

@article{kim_rfx_tmtt_2026,
  author = {Byungkwan Kim},
  title = {rfx: An End-to-End Differentiable 3-D FDTD Simulator for RF and Microwave Engineering},
  journal = {IEEE Transactions on Microwave Theory and Techniques},
  year = {2026},
  pages = {1--15},
  doi = {10.1109/TMTT.2026.3729863}
}

The simulation time S-parameter gradients need, and tail extrapolation carried through the gradient:

@misc{kim_sensitivity_convergence_2026,
  author = {Byungkwan Kim},
  title = {Delayed Convergence of S-Parameter Sensitivities to Design Parameters in FDTD Simulations},
  year = {2026},
  eprint = {2610.28496},
  archivePrefix = {arXiv},
  primaryClass = {eess.SP},
  url = {https://arxiv.org/abs/2610.28496}
}

The software itself:

@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},
  url = {https://github.com/bk-squared/rfx}
}

MIT License — see LICENSE. Developed at the Radar & ElectroMagnetic Intelligence Laboratory.