Neural-Enhanced Physics: A New Paradigm for Accurate and Dimension-Aware Magnetic Core Loss Modeling
- 1 July 2026
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Power Electronics
- Vol. 41 (7) , 10987-11002
Abstract
The accurate and dimension-aware modeling of magnetic core loss is fundamentally limited by a trade-off: model-driven approaches are inherently dimension-aware but are often inaccurate; purely data-driven methods are often accurate but lack dimension-transferability. This paper resolves this dilemma by introducing a novel neural-enhanced physics (NEP) paradigm, which enables highly data-efficient adaptation to different physical core dimensions while maintaining high fidelity. We demonstrate this paradigm with the neural-enhanced dynamic circuit model (NE-DCM), where a physics-based circuit handles dimension-aware electromagnetic dynamics while a data-efficient neural network accurately captures complex material hysteresis. This architecture, based on a principled decoupling of physics and material behavior, yields a key practical advantage: a highly data-efficient workflow. The neural network is trained only once using quasi-static data for a given material, while the physics-based model requires only minimal calibration for new core dimensions. Validated on TDK N87 toroidal cores of different sizes under diverse excitations, the NE-DCM demonstrates superior accuracy, waveform generality, and—most critically—data-efficient transferability across core dimensions, outperforming both purely model-driven and purely data-driven models. The extensibility of the paradigm is further confirmed through successful validation on a second, distinctly different ferrite material (TDK T65). This work establishes the NEP paradigm as a robust and extensible foundation for the next generation of magnetic component design tools.Keywords
Funding Information
- National Natural Science Foundation of China (52407019)
This publication has 35 references indexed in Scilit:
- From model-driven to data-driven: A review of hysteresis modeling in structural and mechanical systemsMechanical Systems and Signal Processing, 2023
- Physics-constrained 3D convolutional neural networks for electrodynamicsAPL Machine Learning, 2023
- Artificial Neural Network Hysteresis Operators for the Identification of Hammerstein Hysteretic SystemsIFAC-PapersOnLine, 2021
- Adaptive Modeling of Highly Nonlinear Hysteresis Using Preisach Neural NetworksJournal of Engineering Mechanics, 2014
- Using neural networks in the identification of Preisach-type hysteresis modelsIEEE Transactions on Magnetics, 1998
- Magnetic hysteresis modeling via feed-forward neural networksIEEE Transactions on Magnetics, 1998
- Modelling the effects of eddy current losses on frequency dependent hysteresis in electrically conducting mediaIEEE Transactions on Magnetics, 1994
- Discrete modelling of magnetic cores including hysteresis eddy current and anomalous lossesIEE Proceedings A Science, Measurement and Technology, 1993
- General properties of power losses in soft ferromagnetic materialsIEEE Transactions on Magnetics, 1988
- Ein Gedankenmodell zur kinetischen Theorie der festen KörperZAMM - Journal of Applied Mathematics and Mechanics / Zeitschrift für Angewandte Mathematik und Mechanik, 1928