Zero-Shot Frequency Generalization for Radio Map Prediction via Cross-Attention Physics-Residual Learning
Year: 2026
Type: preprint
Abstract: Deep learning models for radio map prediction are trained and tested at the same carriers and cannot serve unseen frequencies. We propose a two-stream network learning a residual over an analytic prior, fusing environment and physics streams by cross-attention. Evaluated at the query frequency, this free-space and knife-edge prior absorbs dominant frequency scaling of pathloss, leaving a residual that varies little across bands. Across 150 ray-traced scenes and four training carriers (1.8--28 GHz), the method reduces RMSE by 35.3% zero-shot at unseen carriers and scenes, and by 37.6% at extrapolated 60 GHz, outperforming interpolated 10 GHz.
Source: arXiv (Cornell University)
Author: Hussain, Sajjad
Institution: National University of Sciences and Technology
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203707192
Cites:
Cited by:
Related to: