Active Learning-Based Multiobjective Optimization of Electric Machines Using Probabilistic Metamodeling

While the data-driven optimization of electric machines shows promise, real-world applications still face challenges, including time-consuming data collection and reduced prediction reliability with limited data. To this end, this article first conducts a comparative study of five probabilistic metamodels (MMs) and one deterministic MM. With the help of predictions and uncertainty information, four iterative optimization strategies based on active learning are proposed. In each iteration, a quantity of key data is selected and added to the initial small dataset. The MMs are continually updated with the augmented dataset to obtain more accurate prediction capabilities. The results show that the proposed strategies can significantly enhance the forecast accuracy of the optimal designs in the case of limited data. Furthermore, using predictive uncertainty greatly reduces the overall computational burden for the entire optimization process.