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

Area of Science:

• Electrical Engineering
• Materials Science
• Artificial Intelligence

Background:

• Magnetic core loss prediction is crucial for power electronic system efficiency.
• Traditional models fail with non-sinusoidal waveforms; deep learning methods have limitations.
• Existing models struggle with nonlinear B(t)/H(t) mismatch and multi-scale loss mechanisms.

Purpose of the Study:

• To develop an advanced deep learning model for accurate magnetic core loss prediction.
• To address limitations of existing models in handling complex magnetic material behavior.
• To improve the reliability and efficiency of power electronic systems through better loss prediction.

Main Methods:

• Proposed an Enhanced Memory Augmented Mamba (EMA-Mamba) model.
• Utilized state-space memory augmentation for magnetization pattern storage and retrieval.
• Implemented attention-guided feature selection and physics-constrained multi-objective optimization.

Main Results:

• Achieved an average prediction error of 4.50% and R² of 99.9947% on the MagNet dataset.
• Reduced prediction error by 34.2% compared to state-of-the-art methods.
• Demonstrated excellent temperature robustness and cross-material generalization.

Conclusions:

• EMA-Mamba offers breakthrough performance in magnetic core loss prediction.
• The model effectively handles nonlinearities and complex loss mechanisms.
• Provides a reliable tool for intelligent magnetic component design and optimization.