Comparison of Learning-based Surrogate Models for Electric Motors


Multi-objective optimization is frequently employed in electric motor design, where iterative numerical simulations are required to evaluate a large number of design candidates. A trial-and-error design methodology like this is very time-consuming. In this paper, we propose learning-based surrogate models that use trained deep neural networks (NNs) to accomplish the rapid evaluation of motor designs. A motor design candidate can be described with either a list of geometrical parameters of the motor design, or a colored image of the motor cross-section. Different deep learning models can be constructed with either parameter-based input or image-based inputs. Our analysis reveals that deep convolutional neural networks (CNNs) utilizing image-based inputs exhibit a higher degree of predictive accuracy for more intricate responses, such as cogging torque, in comparison to models employing parameter-based inputs, albeit at the cost of increased training time.