Fast and accurate Neural-Network-based Ferromagnetic Laminated Stack Model for Electrical Machine Simulations in Periodic Regime

Electromagnetic fields and eddy currents in thin electrical steel laminations are governed by the laws of magnetodynamics with hysteresis. Conventional homogenization techniques are however complex and very time-consuming. In consequence, hysteresis and eddy currents in ferromagnetic laminated cores are usually outright disregarded in finite element simulations, considering only saturation, and magnetic losses are only evaluated a posteriori, by means of a Steinmetz-Bertotti like empirical formula. This model simplification yields however potentially inaccurate results in the presence of non-sinusoidal B-fields, common in modern electrical devices. Assuming a time-periodic excitation of the system, a more accurate and fast approach, based on homogenization and neural networks (NN), is presented. A parametric homogenized material law is used in the macroscopic model, whose parameters are given element-wise by a NN according to the actual local waveform of the magnetic field. It is shown that, with an appropriately trained NN, this NN-based material law allows computing fields and losses inside ferromagnetic laminated stacks efficiently and accurately.