Transfer Learning Enhanced High-quality Dataset Generation for Axial Flux Machine

It is crucial to generate high-quality datasets for Axial Flux Machines (AFMs) before studying their performance with machine learning. In this paper, a high-quality dataset generation for AFMs is proposed with a combination of surrogate models and FEM solver. The potential samples in the dataset are prescreened using an XGBoost-based classifier, and only high-quality samples will be further analyzed with a FEM solver to improve the efficiency and quality of dataset generation, thereby achieving higher accuracy levels. To enhance the capability of surrogate models for related design tasks, transfer learning is integrated into the dataset generation algorithm, reducing the required samples for surrogate modelling and bypassing the hyperparameter tuning process. The numerical experiments validated the feasibility and efficiency of the proposed dataset generation for AFMs.