Electric Motor Surrogate Model Combining Subdomain Method and Neural Network

    •  Sakamoto, Y., Xu, Y., Wang, B., Yamamoto, T., Nishimura, Y., "Electric Motor Surrogate Model Combining Subdomain Method and Neural Network", Conference on the Computation of Electromagnetic Fields (COMPUMAG), May 2023.
      BibTeX TR2023-041 PDF
      • @inproceedings{Sakamoto2023may2,
      • author = {Sakamoto, Yusuke and Xu, Yihao and Wang, Bingnan and Yamamoto, Tatsuya and Nishimura, Yuki},
      • title = {Electric Motor Surrogate Model Combining Subdomain Method and Neural Network},
      • booktitle = {Conference on the Computation of Electromagnetic Fields (COMPUMAG)},
      • year = 2023,
      • month = may,
      • url = {}
      • }
  • MERL Contact:
  • Research Areas:

    Electric Systems, Machine Learning, Multi-Physical Modeling


This paper proposes a surrogate model for the rapid evaluation of electric machine designs, based on a neural network combined with a semi-analytical subdomain model. Although both analytical physical-model approaches and data-driven approaches have been proposed to construct surrogate models, which can be significantly faster than numerical finite-element simulations, issues still remain. On one hand, simplifications in analytical approaches often cause inaccuracy, especially in the prediction of highly nonlinear phenomena such as cogging torque of permanent magnet synchronous motors; on the other hand, purely data-driven approaches often require a large amount of training data to achieve high accuracy. In our proposed method, the performance of the electric machine is initially approximated by using a semi-analytical subdomain method, and this initial prediction is used as the input of a neural network, together with other design variables, to obtain the final prediction. We test the method to predict the cogging torque of surface-mounted permanent magnet motors. By combining physical-model and data-driven approaches, the proposed method can predict cogging torque with good accuracy, which cannot be achieved with only physical-model; the prediction accuracy is also much improved compared with conventional neural networks, especially when the size of the training dataset is small.