Physics-Informed Machine Learning Modelling of RF-EMF Exposure in Massive MIMO Systems

Beamforming and massive multiple-input-multiple-output (mMIMO) technologies are key features of base stations (BSs) in the fifth-generation (5G) of mobile networks. This technology is used to focus more radio frequency (RF) energy towards actively connected users to improve their connection/performance, resulting in high variations in the radio frequency electromagnetic fields (RF-EMFs). This paper proposes a new methodology for modelling the RF-EMF exposure for 5G new radio (NR) mMIMO BS by means of a physics-informed machine learning (ML) approach using empirical measurement data. More precisely, the main focus of our work is to develop a suitable traceable RF-EMF exposure prediction tool in the context of 5G mMIMO BSs that can serve multiple mobile users (i.e. multiple-user MIMO (MU-MIMO)) within realistic real-world environments and scenarios. Our RF-EMF prediction tool relies on empirical measurement data acquired via a user-controllable mMIMO beamforming testbed and traceable RF-EMF measurement capability, where both indoor and outdoor RF-EMF measurement campaigns have been carried out. During the measurement campaigns various factors such as number of users, position of users and data duty cycles were considered. Using an ensemble of gradient boosted decision trees, we show that a physics-informed approach can improve predictive performance of RF-EMF compared with a purely data-driven approach, with the ability to extrapolate values of RF-EMF exposure to larger distances. Results show a coefficient of determination value of 0.86 on a 10-fold cross-validated experimental dataset. We also compare the sensitivity of RF-EMF exposure to various factors in the model, and show that model predictions become isotropic for large numbers of beam configurations, simplifying the exposure measurement methodology of 5G systems.