Machine Learning for Reduced Order Modelling of Nonlinear Magnetic Components

The Finite Element Method (FEM) is a powerful tool for simulating non-linear magnetic devices, but its high computational cost becomes a significant limitation when multiple simulations are required, such as in design optimisation or real-time control. To address this challenge, this paper proposes a non-intrusive surrogate modelling framework designed to significantly reduce computation time with minimal accuracy loss. The approach is based on a low-dimensional parameterisation of the solution space using Proper Orthogonal Decomposition (POD), combined with machine learning-based interpolation in the reduced space. The approach is validated by creating a surrogate model of a nonlinear inductor near an iron piece, reconstructing the field distribution as a function of current, frequency, and relative position between the two. Two surrogate modelling techniques, Gaussian Process Regression (GPR) and Feedforward Neural Networks (FNN), are investigated and compared, with particular attention to their performance under data-scarce conditions, common in engineering workflows due to the high cost of generating training data. Numerical results demonstrate that GPR provides more accurate approximations than FNN, especially when few FEM simulations are available. The findings highlight the potential of POD and GPR as efficient and reliable tools for accelerating the simulation and optimisation of magnetic devices.