Physics-Informed VAE-EVT Framework for Tail-Aware Radio Map Prediction

A recent paper published on arXiv (arXiv:2608.15314) presents a novel variational autoencoder-extreme value theory (VAE-EVT) framework tailored for predicting radio maps, particularly focusing on outage areas. This innovative method overcomes the shortcomings of conventional generative radio map models, which typically reconstruct average signal levels and often overlook critical low signal-to-noise ratio (SNR) values essential for ultra-reliable low-latency communication (URLLC) outage forecasting. The framework features a physics-informed preprocessing step that extracts deterministic elements like line-of-sight, shadowing, and distance based on scene geometry. A dual-latent encoder captures the main SNR distribution, while extreme value theory models the tail distribution, facilitating accurate identification of regions where SNR dips below the outage threshold, which can be as strict as the 0.1% quantile for URLLC.