Zero-Shot Frequency Generalization for Radio Map Prediction via Cross-Attention Physics-Residual Learning

August 15, 2026 · View on GitHub

Code and dataset for the paper Zero-Shot Frequency Generalization for Radio Map Prediction via Cross-Attention Physics-Residual Learning (Submitted to IEEE Wireless Communication Letters, August 2026).

This repository is the entry point for:

the ray-traced cross-frequency radio-map dataset (hosted on Hugging Face),
a standalone data loader and the frozen train/val/test split,
scripts to download, verify, and reproduce the benchmark.

TL;DR

git clone https://github.com/sajjadhussa1n/PRCA-Net.git && cd PRCA-Net
pip install -r requirements.txt

# fetch the dataset from Hugging Face (~285 MB)
python scripts/download_data.py            # -> ./data

# sanity-check the download
python scripts/verify_dataset.py ./data

# load a map and reproduce the frozen split
python scripts/example_load.py ./data

What's here

Path	What it is
`radiomap_dataset/`	Standalone loader (numpy/pandas; torch optional) + filename utilities
`scripts/download_data.py`	Pull the dataset from the Hugging Face Hub
`scripts/verify_dataset.py`	Check an unpacked copy is intact and self-consistent
`scripts/example_load.py`	Minimal load-and-plot example
`scene_split.csv`	Frozen 124/13/13 scene split — use verbatim to compare to the paper
`docs/DATASET.md`	Full dataset description, layout, filename convention
`docs/REPRODUCE.md`	How to reproduce the reported results

The dataset

150 urban scenes (15 cities), 256×256 rasters, 8 transmitters/scene
6 frequencies: 1.8 / 3.5 / 7 / 28 GHz (train) + 10 / 60 GHz (held out)
7,200 path-loss maps, receiver at 1.5 m, isotropic antennas
Generated with Sionna RT 2.0.1 over OpenStreetMap geometry

Hosted on Hugging Face: [HF DATASET LINK TO BE ADDED HERE] Full details in docs/DATASET.md.

Reproducing the paper

Use the frozen split (scene_split.csv) verbatim. See docs/REPRODUCE.md for the evaluation protocol and the model checkpoint link. [checkpoint link here]

License

Data: ODbL v1.0 (derived from OpenStreetMap — attribution required). See LICENSE_DATA.md.
Code: MIT. See LICENSE.

Citation

@article{[Hussainprcanet2026],
  title   = {Zero-Shot Frequency Generalization for Radio Map Prediction via Cross-Attention Physics-Residual Learning},
  author  = {Sajjad Hussain},
  journal = {Submitted to IEEE Wireless Communication Letters},
  year    = {August 2026},
  note    = {Code: https://github.com/sajjadhussa1n/PRCA-Net}
}

Please also cite the dataset (see CITATION.cff).