[3E2-GS-2g] Machine learning

12:15 PM - 12:30 PM JST(3:15 AM - 3:30 AM UTC)

[3E2-GS-2g-06] A Study on the Convergence Characteristics of Physics-Informed Neural Networks in Magnetic Field Analysis under Induction Heating Conditions

〇Tomonori Suzuki¹, Hiroshi Koizumi², Kotaro Nakamura², Yoshinobu Yamade², Shinichi Mizuhara², Hiromichi Uetake² (1. AISIN DIGITAL ENGINEERING Corporation, 2. Mizuho Research & Technologies, Ltd.)

Keywords:

PINNs,Magnetic Field Analysis,Induction Heating,Convergence Characteristics

This study investigates the convergence characteristics of Physics-Informed Neural Networks (PINNs) applied to AC magnetic field analysis in induction heating. Under high-frequency and high-relative-permeability conditions, it was confirmed that the extreme reduction of skin depth and the handling of internal boundary conditions significantly hinder stable learning. In this work, convergence was improved through non-dimensionalization of the governing equations, optimization of sampling points, explicit incorporation of internal boundary conditions, and the use of domain-decomposed networks. As a result, we demonstrate that even under strongly multiscale and multidomain conditions with a relative permeability of 5000, PINNs are capable of reproducing physically reasonable field distributions consistent with FEM results.