Electric Machine Inverse Design with Variational Auto-Encoder (VAE)

Abstract

Machine learning and deep learning techniques have been proposed to facilitate the design optimization of electric machines. Most of the existing research focuses on the development of surrogate models, while iterative optimization is still needed. Inverse design approach, on the other hand, can directly provide design candidates with trained deep learning model without iteration. One major challenge in deep learning based inverse design is the so-called one-to-many mapping problem. In this paper, we propose an intelligent inverse design approach for electric machines based on a variational autoencoder (VAE), which can effectively address the problem and provide desired motor design candidates for multiple design targets at the same time. We demonstrate the feasibility of the proposed strategy with multi-objective design task of a surface-mount permanent magnet motor, and show that it is generally applicable for different types of electric motors.