Hot Spot Driven Physics-informed Neural Network via Special Designed Quantity of Interest applied to Magneto-thermal Analysis

Convolutional neural networks (CNN) have been widely used and achieved some good results in the field of computational electromagnetism. In this work, a hot spot driven physics-informed neural network (PINN) is proposed for magneto-thermal analysis. A loss function is specially designed, which takes the hot spot as the quantity of interest (QOI), for both its location and its temperature value. Besides, the influence of the QOI on the hyperparameter is investigated to ensure the convergence of the PINN. Numerical experiments show that the proposed approach can significantly improve the accuracy of the predicted temperature of the hot spots.