Physics Inspired Artificial Neural Network Adaptation for SAR Prediction in Bio-EM Problems

Machine learning (ML) technique is nowadays popular for predicting electromagnetic field exposure, specifically as specific absorption rate (SAR) values, in Bio electromagnetic (Bio-EM) area. Considering the material uncertainty of human tissues, efforts to quantify SAR values mostly rely on 3D full wave simulations, rather than realistic measurements. For precise SAR calculations, a high-resolution human model is often desired and expensive computational resources are required. In this work, an artificial neural network (ANN) as a ML method is employed for SAR prediction in the human head at 13.56 MHz under plane wave illumination. The dependency of SAR values on +/- 20% material parameter changes is expected to be linear and separable towards each parameter, inspired by physical inspection of fields. This allows ANN adaptation for effective acceleration of SAR prediction using reduced ANN models. For validation, the proposed ANN modeling uses up to hundreds of full-wave simulations as training data and multiple human head models. This approach achieves a good accuracy of over 95% for SAR prediction. Using the proposed model facilitates fast and accurate predicted results of SAR values for critical situations, which can be compared with standard levels despite the uncertainty in tissue dielectric properties.