Prediction of Motor Characteristic Maps via Deep Operator Networks for Topology Optimization

Abstract

This study proposes a novel methodology for directly training response surfaces of torque and magnet flux density distributions of interior permanent magnet synchronous motors (IPMSMs) under specified input conditions and predicting the speed-torque characteristics or torque wave. Existing models have low prediction accuracy when simultaneously integrating shape information and current conditions. To overcome this challenge and improve the prediction of motor characteristics, a new deep learning (DL) model that combines deep operator networks and convolutional neural networks is proposed. This advanced approach greatly improves the accuracy of motor characteristic predictions across varying current levels and rotational angles, achieving an 89.6% increase in accuracy compared to comparative methods. The model is successfully applied to parameter and topology optimization (TO), effectively maximizing the speed-torque characteristics of IPMSMs.