Toward EM Digital Twins: EM-Guided Deep Learning for Accurate Prediction of Wireless Power Transfer in IIoT Environments

Accurately predicting electromagnetic (EM) field propagation is fundamental for the design of wireless power transfer (WPT) and industrial Internet of Things (IIoT) systems, where sensors performance and energy availability strongly depend on the surrounding environment, particularly in harsh or cluttered conditions. Conventional full-wave simulations are precise but computationally prohibitive for large or dynamic scenarios, limiting their use for real-time cases. Models based on ray tracing rely on the far-field approximation that is often violated in industrial and automotive environments. This work presents a data-efficient deep learning framework that integrates physics-based modelling with convolutional neural networks (CNNs) to reconstruct spatial power density distributions in complex environments. A dedicated data generation pipeline, combining active sampling with an integral equation (IE)-based EM solver, enables the construction of informative training datasets from a limited number of simulations. Different optimization objective functions are investigated to improve the accuracy and physical consistency of the predicted fields. Despite being trained on a reduced dataset, the proposed CNNs accurately reproduce both global propagation patterns and localized field variations, achieving strong agreement with experimental measurements. The results demonstrate that CNN-based models can deliver fast and reliable evaluation of EM power distribution allowing optimal sensor placement in industrial environments. These capabilities represent an important first step toward developing data-driven EM digital twins for next-generation IIoT systems.