Prediction of Electromagnetic Field Exposure at 20–100 GHz for Clothed Human Body Using an Adaptively Reconfigurable Architecture Neural Network With Weight Analysis (RAWA-NN) Framework

Abstract

In the context of forthcoming sixth-generation (6G) wireless communication, the sub-terahertz and terahertz frequency spectrum are anticipated. At such high frequencies, electromagnetic field (EMF) exposure assessment becomes significantly challenging, requiring substantial computational resources. This article is the first to utilize machine learning (ML) to predict EMF exposure levels for the clothed human body at 20–100 GHz, including temperature rises and absorbed power density (APD) at the exposed skin surface. To predict the EMF exposure, a reconfigurable architecture neural network with weight analysis (RAWA-NN) framework is proposed. This framework is based on the deep neural network (DNN) integrating the proposed weights-analyzer module and optimization module. The proposed novel framework streamlines the training process and reduces training time, while simultaneously adaptively optimizes the hyperparameters (hidden layers and hidden sizes) without the necessity for manual intervention during training and optimization. The model was trained using 70% of forearm data, with the remaining data for testing. Data from other body parts, such as the abdomen and quadriceps, was used to validate the model generalization. Compared to conventional dosimetry analysis, relative difference (RD) across various parameters remains below 2.6% across various parameters, for the same body part of the forearm, and below 9.5% for other body parts. There is an approximate four orders of magnitude improvement in assessment speed.