Physics-Informed Neural Operator Predicts Transient Magnetization in Power Magnetics

Magnetic components in modern high-frequency, high-power-density converters often operate under challenging conditions, including non-sinusoidal flux-density waveforms, fast transitions, and temperature variations. Traditional steady-state models and material curves fail to capture the complex transient magnetization responses in these scenarios. This paper proposes the Physics-Informed Hybrid Neural Operator (PI-HNO) to address this limitation. PI-HNO is a compact, material-specific neural model designed for core-loss-oriented transient magnetization prediction. It combines a local recurrent branch to handle boundary-state representation and rate-dependent evolution with a global, Preisach-inspired branch that extracts waveform-level hysteresis context. Crucially, the model incorporates B-H energy-consistency regularization. Evaluated on 14 ferrite materials from the MagNetX database, PI-HNO achieved excellent sequence accuracy and energy consistency (mean error 1.92%, 95th percentile 7.60%) with a remarkably small number of trainable parameters (4777 per model), demonstrating its efficiency and robustness.