A Generalizable Indoor Propagation Model Based on Graph Neural Networks

A surrogate model that “learns” the physics of radio wave propagation is indispensable for the efficient optimization of communication network coverages and comprehensive electromagnetic field (EMF) exposure assessments. The capability of a model to predict reasonable outputs given an input that is beyond the data with which the model is trained, namely, “generalizability,” is a fundamental challenge and a key factor for its practical deployment. In this article, by leveraging the concept of graph neural networks (GNNs), a prediction model for indoor propagation that is “generalized” to not only transmitter (Tx) positions but also new geometries is presented. We demonstrate that a geometry and a Tx antenna can be modeled as a graph with all necessary information being included, and a GNN can acquire the knowledge of propagation physics through “learning” from these graphs. We further show that the model can be generalized to new geometry shapes, beyond the shape (square) for model training. We provide useful information on how to obtain an acceptable accuracy for different scenarios. We also discuss the potential solutions to further improve the model’s capability.