Witwin Channel

Warning This is an experimental early release. APIs may change frequently, bugs are expected, and the project will be updated aggressively.

witwin-channel is a differentiable, geometric differentiable, GPU-first wireless channel simulation package for the Witwin platform. It models radio propagation through declarative scenes and solver-specific entrypoints for deterministic radiomaps, Monte Carlo radiomaps, and receiver-level path export.

The stable public contract is:

Scene + solver.solve(scene, config) + Result

Solvers live under witwin.channel.deterministic, witwin.channel.montecarlo, and witwin.channel.path. There is no shared public tracer object.

Capabilities

• Differentiability-first channel simulation for inverse rendering, optimization, and gradient-based RF digital twin workflows.
• Geometric differentiability through scene geometry, material parameters, endpoints, and solver outputs.
• Declarative scene construction with witwin.core.Structure, Material, and geometry objects.
• Scene-owned Transmitter, Receiver, and ReceiverGrid endpoints selected by name at solve time.
• Line-of-sight, multi-bounce reflection, and multi-order UTD-style diffraction support.
• Deterministic radiomap solving for structured grid outputs.
• Monte Carlo radiomap solving with basic and BDPT integrators.
• Path-level solving for CIR/CFR, delay, angle, interaction, and optional geometry payloads.
• Dr.Jit-native runtime internals with CUDA/native extension acceleration paths.
• Pytest validation, GPU acceptance tests, benchmark scripts, and Sionna RT comparison workflows.

Requirements

Core solver workflows expect:

• Windows or Linux with an NVIDIA GPU.
• CUDA-capable Python environment.
• Python 3.10 or newer.
• Dr.Jit 1.3.1.
• PyTorch with CUDA support.
• RayD and native extension build dependencies where acceleration paths are required.

Repository development uses the witwin conda environment.

Installation

Create and activate a local conda environment:

conda create -n witwin python=3.11 -y
conda activate witwin

Install the CUDA-enabled PyTorch build that matches your driver and platform, then install this package. For example:

python -m pip install torch --index-url https://download.pytorch.org/whl/cu121

From the channel subproject root:

conda activate witwin
python -m pip install -e . --no-build-isolation --no-deps

For non-editable local installs, keep dependency resolution disabled:

conda activate witwin
python -m pip install . --no-deps

Native C++/CUDA iteration should use the existing CMake build directory when possible instead of repeated editable reinstalls:

cmake --build build\cp311-cp311-win_amd64 --config Release --target <native_target>
cmake --install build\cp311-cp311-win_amd64 --config Release --prefix <witwin>\Lib\site-packages

On Windows, close Python and Jupyter processes that have loaded the extension before installing .pyd files, because loaded extension modules are locked.

Quick Start

import numpy as np
import witwin.channel as wc

scene = wc.Scene(
    structures=[
        wc.Structure(
            name="wall",
            geometry=wc.Box(
                position=(0.0, 0.0, 1.5),
                size=(0.25, 4.0, 3.0),
                device="cuda",
            ),
            material=wc.Material(eps_r=4.0, sigma_e=0.0),
        ),
    ],
    transmitters=[
        wc.Transmitter("tx", (-2.0, 0.0, 1.5)),
    ],
    receivers=[
        wc.ReceiverGrid(
            "rm",
            axis="z",
            position=1.5,
            bounds=((-3.0, 3.0), (-3.0, 3.0)),
            grid_shape=(32, 32),
        ),
    ],
    frequency=3.5e9,
    device="cuda",
)

result = wc.deterministic.solve(
    scene=scene,
    transmitter="tx",
    receiver="rm",
    config=wc.deterministic.Config(
        num_samples=256,
        max_bounces=1,
        max_diffraction_order=0,
        edge_policy=wc.EdgePolicy(edge_selection_mode="all_edges"),
    ),
)

path_gain = np.asarray(result.path_gain, dtype=np.float32)
print(path_gain.shape, float(path_gain.max()))

Solver Entry Points

Deterministic Radiomap

Use wc.deterministic.solve(...) when you need a repeatable radiomap over a scene-owned ReceiverGrid.

config = wc.deterministic.Config(
    num_samples=512,
    max_bounces=2,
    max_diffraction_order=1,
    edge_policy=wc.EdgePolicy(edge_selection_mode="all_edges"),
    tuning=wc.deterministic.Tuning(enable_rd_diffraction=True),
)

result = wc.deterministic.solve(
    scene=scene,
    transmitter="tx",
    receiver="rm",
    config=config,
)

The result is a shared wc.RadioMapResult with path_gain, component maps, coordinates, transmitter-axis helpers, and metadata.

Monte Carlo Radiomap

Use wc.montecarlo.solve(...) for transmitter-driven radiomap sampling and Monte Carlo integrators.

config = wc.montecarlo.Config(
    num_samples=128,
    max_bounces=1,
    max_diffraction_order=0,
    integrator_options=wc.montecarlo.IntegratorOptions(
        integrator="basic",
        samples_per_tx=4096,
        accumulation_backend="auto",
        seed=7,
    ),
)

result = wc.montecarlo.solve(
    scene=scene,
    transmitter="tx",
    receiver="rm",
    config=config,
)

The Monte Carlo package also exposes IntegratorOptions(integrator="bdpt") for BDPT workflows.

Path Solver

Use wc.path.solve(...) when you need discrete paths for named receivers. The scene must contain the selected Receiver endpoints.

path_result = wc.path.solve(
    scene=scene,
    transmitter="tx",
    receiver=["rx0", "rx1"],
    config=wc.path.Config(
        num_samples=256,
        max_bounces=1,
        max_diffraction_order=0,
        max_num_paths=8,
        return_geometry=True,
        edge_policy=wc.EdgePolicy(edge_selection_mode="all_edges"),
    ),
)

coeff, delay = path_result.cir()
response = path_result.cfr(subcarriers_hz)

PathResult includes path coefficients, delays, AoD/AoA metadata, interaction types, optional geometry, and torch-backed CIR/CFR helpers.

Public API Overview

The recommended user-facing import is:

import witwin.channel as wc

The umbrella namespace exports:

• Scene and endpoint types: Scene, Transmitter, Receiver, ReceiverGrid.
• Core geometry and materials: Box, Material, Structure.
• Antenna arrays: AntennaArray, PlanarArray, ULA, UPA.
• Solver namespaces: deterministic, montecarlo, path.
• Shared result and Dr.Jit aliases such as RadioMapResult, Point3f, and Vector3f.

Scene frequency, materials, endpoint positions, endpoint power, polarization, and arrays are owned by Scene and its endpoint/material objects. Solver configs should describe algorithm controls such as sample budgets, bounce limits, diffraction order, edge policy, integrator settings, and advanced tuning.

Tutorials

Introductory notebook workflows live under tutorials/.

Testing

Run tests from this directory with the witwin environment active:

conda activate witwin
python -m pytest tests

Common targeted runs:

python -m pytest tests/scene
python -m pytest tests/deterministic
python -m pytest tests/montecarlo
python -m pytest tests/path
python -m pytest tests --gpu
python -m pytest tests --gpu --acceptance

Manual validation and benchmark entrypoints live under tests/support/bin/. Useful maintained commands include:

python -m tests.support.bin.benchmark_monte_carlo_radiomap_package --mode forward --integrator basic --json
python -m tests.support.bin.benchmark_mc_basic_munich_vs_sionna --json
python -m tests.support.bin.validate_path_solver_munich --json

Repository Layout

Path	Role
witwin/channel/core/	Shared geometry, numerics, physics, runtime, scene, kernels, and result helpers
witwin/channel/core/scene/	Declarative Scene, mesh adaptors, endpoints, arrays, and edge policy
witwin/channel/deterministic/	Deterministic radiomap solver package
witwin/channel/montecarlo/	Monte Carlo radiomap solver package
witwin/channel/path/	Path-level solver package
witwin/channel/_native/	Native extension loaders and package-level native bindings
tests/	Regression, GPU, acceptance, and support-script coverage
tutorials/	Introductory notebook workflows

The package layering is:

channel/core/ -> channel/core/scene/ -> {channel/deterministic, channel/montecarlo, channel/path}/

Solver packages must not import from each other. Shared solver-neutral behavior belongs in witwin/channel/core/.

Citation

If you use this work in academic research, please cite:

@inproceedings{chen2026rfdt,
  title     = {Physically Accurate Differentiable Inverse Rendering
               for Radio Frequency Digital Twin},
  author    = {Chen, Xingyu and Zhang, Xinyu and Zheng, Kai and
               Fang, Xinmin and Li, Tzu-Mao and Lu, Chris Xiaoxuan
               and Li, Zhengxiong},
  booktitle = {Proceedings of the 32nd Annual International Conference
               on Mobile Computing and Networking (MobiCom)},
  year      = {2026},
  doi       = {10.1145/3795866.3796686},
  publisher = {ACM},
  address   = {Austin, TX, USA},
}

License

All rights reserved. This repository is provided for research use only. For commercial use, redistribution, or sublicensing, please contact xic063@ucsd.edu.