Magnetic field estimation for electromagnetic devices using physics-informed deep learning : methods, challenges and future directions

Abstract

As performance demands on electromagnetic devices (EMDs) intensify, the need for high-precision magnetic field estimation techniques becomes increasingly critical. Physics-informed neural networks (PINNs), which integrate physical laws with deep learning, have recently emerged as a powerful approach for tackling complex electromagnetic problems, offering superior computational efficiency and generalization. Compared to traditional numerical methods, PINNs can effectively handle heterogeneous media and complex boundaries, while maintain high prediction accuracy even with limited data, thus garnering significant global research interest. However, the application of artificial intelligence to magnetic field estimation is still nascent, and a systematic survey of existing methods, especially those involving the solution of Maxwell’s equations, has not yet been conducted. This paper provides a comprehensive overview of physics-informed deep learning approaches for magnetic field estimation, summarizes general computational procedures, identifies key challenges, and discusses future directions. Using inverters and permanent magnet machines as representative examples, practical magnetic field estimation processes are outlined and how post-processing of predicted fields is analyzed, which can facilitate the design optimization of EMDs.