Scene-Conditioned PINN-GNN for Multipath RF Maps: Cross-Scene Generation and In-Scene Completion

This paper presents a unified framework for constructing multipath radio-frequency maps with a physics-informed neural network (PINN) and a graph neural network (GNN). The PINN embeds electromagnetic propagation constraints when predicting path gain, time of arrival, and angles, while the GNN models spatial correlations among neighboring receivers. The method supports both cross-scene generation and in-scene completion using 2D and 2.5D environmental representations. It also introduces a peak-weighted dynamic time warping metric that jointly measures amplitude error and peak-delay misalignment in channel impulse responses. The authors report consistent improvements over image-based, diffusion-based, and interpolation baselines under sparse observations.