A physics-inspired memory-augmented deep learning framework for magnetic core loss prediction

Accurate prediction of magnetic core loss is a key challenge for improving the efficiency and reliability of power electronic systems. Traditional empirical models such as the Steinmetz equation are only applicable to sinusoidal steady-state conditions and struggle with the complex non-sinusoidal waveforms and variable operating conditions in modern power electronics. While existing deep learning methods have shown improvements, they still face fundamental limitations in handling the nonlinear mismatch between B(t) and H(t) waveforms, coupling of multi-scale loss mechanisms, and generalization under extreme operating conditions. This paper proposes an Enhanced Memory Augmented Mamba (EMA-Mamba) model that achieves breakthrough progress in magnetic core loss prediction. It utilizes a state-space memory augmentation mechanism that stores and retrieves typical magnetization patterns through a trainable external memory matrix, endowing the model with a capability similar to the "magnetic memory" of magnetic materials, effectively solving the gradient vanishing problem in long sequence modeling. Combined with an attention-guided intelligent feature selection mechanism, it adaptively identifies critical turning points in hysteresis curves through a Top-K strategy, fundamentally solving the temporal mismatch problem between B(t) and H(t) waveforms. Finally, through a physics-constrained multi-objective optimization framework, it achieves decoupled modeling of hysteresis loss, eddy current loss, and residual loss through loss function combination, overcoming the optimization difficulties caused by data spanning six orders of magnitude. Experiments on the MagNet dataset containing 10 materials and over 150,000 data points show that EMA-Mamba achieves an average prediction error of 4.50% and a coefficient of determination of 99.9947%, reducing error by 34.2% compared to state-of-the-art baseline methods, with a 36.2% reduction in 95th percentile error under extreme conditions. The model demonstrates excellent temperature robustness and cross-material generalization capability, providing a reliable theoretical tool for intelligent design and optimization of magnetic components.