Neural Networks for Accurate, Scalable and General Core Loss Modeling in Magnetic Components

Abstract

Accurate prediction of magnetic core losses under modern power-electronics excitations is critical for high-efficiency, high-frequency converter design, yet existing approaches often face a trade-off between accuracy, generality to arbitrary waveforms, and scalability across component geometries. This thesis develops a progressive set of hybrid physics–neural modeling strategies that combine physical structure with data-driven learning to achieve accurate, general, scalable, and data-efficient core-loss prediction in magnetic components. First, a Magnetization-Mechanism-Inspired Neural Network (MMINN) decomposes losses into hysteresis and dynamic contributions and is tailored to steady-state, symmetric high-frequency excitations, enabling accurate prediction with compact model size and limited training data. Second, to address transient and quasi-static arbitrary waveforms (including DC bias and minor loops), a History-Dependent Prandtl–Ishlinskii Neural Network (HDPI-NN) augments classical hysteresis operators with neural pathways to capture reversible magnetization and memory effects, improving generalization beyond earlier physics-guided neural models. Finally, for dimension-aware scalability across different core sizes under transient, high-frequency, arbitrary and asymmetric waveforms, a Neural-Enhanced Dynamic Circuit Model (NE-DCM) embeds the HDPI-NN material model into a physics-based mesh dynamic-circuit solver, enabling accurate prediction of terminal current waveforms and associated core losses with minimal calibration effort for new cores. Collectively, the proposed methods demonstrate that synergistically integrating physical modeling with neural networks provides a robust foundation for next-generation magnetic-component design tools, supported by open-source tools and datasets to encourage adoption and further research.