Fast Field Strength Prediction Using Modern Machine Learning in European Cities From Few RF-EMF Measurements: A Neural Tangent Kernel Perspective

Abstract

The rapid growth of 5G and 6G networks, with their dense deployments, millimeter-wave communications, and dynamic beamforming, necessitates scalable simulation tools for performance evaluation, network design, and electromagnetic field (EMF) exposure assessment. EMF simulations model exposure across frequency, space, and time, typically based on base stations, user devices, and deterministic and empirical models. These simulations are essential for network operators and researchers for network planning, assessing human exposure to RF-EMF, and ensuring compliance with safety regulations considering factors like frequency, power levels, antenna configurations, and the environment, but they often require extensive computational resources and large simulation time. To address this, we propose an infinitely wide convolutional neural network approach for fast and accurate EMF exposure estimation. Remarkably, taking the width of a neural network to infinity allows for improved computational performance. We compute a convolutional neural tangent kernel from the infinite-width network to perform matrix imputation for exposure estimation. Proposed method estimates exposure fields using less than 7% of simulation points and outperforms other machine learning models, predicting exposure levels which fall below the safety limit set by ICNIRP in under 3,66×10−3 seconds.