基于物理驱动神经网络的低频电磁场快速算法研究综述OA

Numerical computation of low-frequency electromagnetic fields is a core foundational technology for the electromagnetic characteristic analysis of electrical equipment.Faced with the prevalent multi-scale structural features,strong material nonlinearity,and the need for iterative solutions under various operating conditions in electrical equipment,traditional numerical algorithms face challenges in improving computational efficiency,solution accuracy,and cross-scenario generalization ability.With the development of deep learning,physics-driven neural networks,by integrat-ing physical constraints such as control equations and initial/boundary conditions,can solve physical fields with few or no data,and become an effective method to improve the computational efficiency,to realize the model generalization,and to ensure the physical reliability of results.This paper systematically reviews the research progress of physics-driven neural networks in the rapid computation of low-frequency electromagnetic fields.First,this paper elucidates the general solu-tion framework and basic process of this type of algorithm.Then,this paper classifies and compares mainstream algorithms according to the mathematical form of the loss function used to constrain the control equations,and focuses on reviewing application cases of various algorithms in typical electromagnetic field problems.Based on this,this paper ex-plores the extended applications of this type of algorithm in three-dimensional electromagnetic models,and provides an in-depth analysis and discussion of its computational efficiency,engineering applicability,and related computational framework.Finally,the future development trend of this field is envisioned.It is pointed out that,through systematic ad-vancements in following four directions:complex 3D modeling,nonlinear material characterization,deep migration mechanisms,and data-physics fusion-driven architecture,it is hoped that a novel computational method with high effi-ciency,high accuracy,and strong generalization ability can be provided for low-frequency electromagnetic field analysis.