Physics-Informed Sparse Neural Network for Permanent Magnet Eddy Current Device Modeling and Analysis

Abstract

The objective is to study the prediction of the electromagnetic (EM) field and the output performance of permanent magnet eddy current devices (PMECDs) based on a physics-informed sparse neural network (PISNN). In order to achieve this goal, a unified physical model is first defined according to different types of PMECDs, which is equivalent to solving a parameterized magnetic quasi-static problem. A soft constraint module and a hard constraint module, composed of physical equations, are constructed. The soft constraints are then integrated into the neural network's objective function, while the hard constraint module is utilized to predict device performance and physical field. Stochastic gradient descent is used to minimize the residual of the physical equations during PISNN training. Subsequently, the structural parameters and operating parameters of the PMECD are modified to verify the generalization ability of the model. Our results indicate that PISNN accurately and efficiently predicts the EM field distribution and the output torque. Furthermore, our prediction results for permanent magnet eddy current devices with different parameters demonstrate the potential of the method for transfer learning.