Multiobjective Optimization of Electrical Machines Using Probabilistic Surrogate Modeling with Limited Data

In the stage of designing electrical machines, repeated calculations are commonly performed using finite element analysis (FEA), implying a high computational burden. To this end, artificial intelligence-based surrogate models (SMs) are being increasingly applied to significantly shorten the optimization time and obtain high estimation accuracy. However, the generation of datasets for model training still requires FEA, resulting in a massive amount of simulation time for large data. In this article, we propose a new optimization workflow tailored for a limited number of samples utilizing a probabilistic SM called the Gaussian process regression model, which can provide not only mean predictions but also additional prediction uncertainty information. The incorporation of uncertainty enables a better selection of key data from previous optimization solutions. With the help of key data, the SMs are retrained iteratively to achieve better prediction ability. In the end, the total optimization time and the errors between prediction and FEA verification are reduced substantially.