CAHR-Net Models Magnetic Core Loss with High Accuracy, Interpretability.

Magnetic core loss, a critical factor in power electronics, arises from the hysteresis loop, where energy dissipation per cycle is determined by the loop area. Factors like frequency, temperature, and waveform shape significantly influence this loss by altering the loop's geometry. Most existing models, however, only apply these conditions to a final scalar output, losing the intermediate, interpretable representation of hysteresis. This paper introduces CAHR-Net (Condition-Adaptive Hysteresis Reconstruction Network), a novel approach that injects operating conditions precisely where they physically act within the model. It preserves the clear chain from flux density waveform to magnetic field reconstruction, loop-area integration, and power loss estimation, making the model highly interpretable. Feature-wise linear modulation is used to integrate frequency, temperature, and waveform statistics into the intermediate reconstruction. Evaluated on the MagNet material protocol, CAHR-Net achieves an average p95 relative error of 6.89% with only 1874 parameters, outperforming all compared methods. It also surpasses the strongest black-box solutions in accuracy for the most difficult materials, using approximately 48 times fewer parameters. Ablation studies confirm that the combination of physical loop reconstruction, structured condition modulation, and a specialized training protocol drives these improvements.