Data-Driven Radio Propagation Modeling using Graph Neural Networks

TL;DR

This work tackles the challenge of flexible radio propagation modeling by replacing physics-based solvers with a data-driven graph neural network trained on real-world measurements. It constructs two graphs from environmental imagery to capture diffusion-like spatial interactions and ray-tracing paths, enabling end-to-end learning of coverage maps using a semi-supervised masking approach. The authors introduce a FiLM-conditioned GraphNetwork architecture and an $L_2$ loss framework to predict per-node attenuation, achieving competitive RMSE while offering faster inference than traditional solvers. The results demonstrate that data-driven, structurally informed GNNs can generate accurate coverage maps in urban environments, supporting rapid network design and optimization from point measurements.

Abstract

Modeling radio propagation is essential for wireless network design and performance optimization. Traditional methods rely on physics models of radio propagation, which can be inaccurate or inflexible. In this work, we propose using graph neural networks to learn radio propagation behaviors directly from real-world network data. Our approach converts the radio propagation environment into a graph representation, with nodes corresponding to locations and edges representing spatial and ray-tracing relationships between locations. The graph is generated by converting images of the environment into a graph structure, with specific relationships between nodes. The model is trained on this graph representation, using sensor measurements as target data. We demonstrate that the graph neural network, which learns to predict radio propagation directly from data, achieves competitive performance compared to traditional heuristic models. This data-driven approach outperforms classic numerical solvers in terms of both speed and accuracy. To the best of our knowledge, we are the first to apply graph neural networks to real-world radio propagation data to generate coverage maps, enabling generative models of signal propagation with point measurements only.

Figures (17)

• Figure 1: 2D histogram of the number of measurement points in France.
• Figure 2: A view of all the measurement points in the city of Paris (specificly around Île de la Cité). We can recognize the street structure of the city which indicate the good quality of the GPS information in the dataset. Each blue points is a measurement point.
• Figure 3: This figure provides a visual representation of the measurement points surrounding the antenna. The antenna is located at the center of the figure, and all points that appear in yellow represent areas where no measurements were taken, and thus have a default value of -1. The remaining points in the figure display the RSRP received at the specific location by users, which are influenced by the specific configuration of the antenna diagram.
• Figure 4: A view of the buildings / vegetation heights information around the antenna (at the center of the image).
• Figure 5: A view of the buildings type information around the antenna. There is several classes : 0 is that there is nothing on the ground, 5 is a building, 7 is vegetation and 6 is water / river.