Towards a Wireless Network Digital Twin Model: A Heterogeneous Graph Neural Network Approach

A digital twin of wireless networks can enable the ability to safely and rapidly recreate what-if scenarios of mobile networks for more efficient and intelligent network optimization and planning. In this work, we present a novel digital twin model of wireless networks based on Heterogeneous Message Passing Graph Neural Networks (HMPGNNs). Our digital twin model represents wireless network nodes and the underlying wireless phenomena between them as nodes and edges of different type in a heterogeneous graph. Heterogeneous graphs are fed as samples into the HMPGNN model so that the model learns to simulate the underlying wireless phenomena. Results using system-level simulations to train and evaluate our proposal, show that our approach accurately and efficiently captures the dynamics of wireless networks to produce accurate reconstruction of downlink data rates of all users while also generalizing to network deployments unseen during training.