Comparative Study of Different Surrogate Models in the Multi-physics Optimization of Synchronous Reluctance Machine

The optimization of synchronous reluctance machine (SynRM) involves multi-physical considerations and high-dimensional input parameters, making it highly time-consuming. To accelerate the design optimization process of SynRM, this study focuses on constructing and comparing four surrogate models in an operating point dependent multi-physics optimization routine. The models compared include Kriging, support vector regression (SVR), fully connected deep neural network (FC-DNN), and convolutional neural network (CNN) models. Through analysis, it is found that the mechanical safety factor, torque ripple and efficiency of SynRM are highly sensitive to geometry variations and difficult to predict. Thus, a hybrid optimization strategy assisted by finite-element-analysis (FEA) is proposed. The optimization results prove that the proposed hybrid optimization strategy is able to achieve a significantly improved Pareto front with only a minimal number of FEA simulations. Additionally, among the four surrogate models, the Kriging model is found to be the most capable model for the parameter optimization in terms of both model construction expenses and optimization results.