open access

https://hdl.handle.net/2115/99254

Physics-Aware Machine Learning for Construction of Surrogate Models with Small Training Data

英語

wireless power transfer

2025-10-16

Institute of Electrical and Electronics Engineers (IEEE)

IEEE Transactions on Magnetics

1-7

7

This paper shows that physics-aware machine learning can construct accurate surrogate models for optimal design, uncertainty quantification, and data assimilation, even when limited training data is available. In this method, the surrogate model is trained to predict the ratio of the ground truth—provided by, e.g., field computations or measurements—to the prediction of a physics-based model. In this multiplicative correction, the surrogate model provides the compensation factor to the prediction of the physics-based model, for example, an analytical or empirical formula. After training, accurate predictions can be obtained using the physics-based model and trained machine learning system, both of which have a light computing burden. Additive correction is also tested, in which the difference between the ground truth and the prediction of a physics-based model is used rather than the ratio. Numerical experiments reveal that multiplicative correction based on the ratio is more accurate than the additive correction based on the difference. It is shown that the response surface of the ratio tends to be so flat that it can be represented by the small training data, while this is not the case for the difference-based method and direct prediction. The present method provides an explicit approximation formula for the ratio based on linear regression, allowing physical interpretation. The potential applications of this method, including data assimilation, are suggested.

学術雑誌論文

Accepted Manuscript

© 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.

0018-9464

AA00667933