Machine Learning-based Reduced Order Modeling of Nonlinear and Multiphysics Magnetic Devices

Induction heating processes involve complex multiphysical interactions between electromagnetic and thermal models, often characterized by nonlinear materials. The finite element method is a common approach to tackling such problems, but its computational complexity may become a burden for optimization loops or real-time monitoring. This work proposes a surrogate modeling approach that combines proper orthogonal decomposition and Gaussian process regression, suited for nonlinear and temperature-dependent magnetic materials. The approach is applied to reduce the complexity of the electromagnetic model of the testing electromagnetic analysis method (TEAM) problem 36, consisting of a copper coil heating through induction a steel billet. Results show how the non-intrusive machine learning approach can accurately reconstruct the field distribution, offering a viable first step towards fast multiphysical simulations involving nonlinear magnetic materials.