Fast Multi-Objective Optimization of Electromagnetic Devices Using Adaptive Neural Network Surrogate Model
Hayaho Sato, Hajime Igarashi
Abstract
This article presents a fast population-based multi-objective optimization of electromagnetic devices using an adaptive neural network (NN) surrogate model. The proposed method does not require any training data or construction of a surrogate model before the optimization phase. Instead, the NN surrogate model is built from the initial population in the optimization process, and then it is sequentially updated with high-ranking individuals. All individuals were evaluated using the surrogate model. Based on this evaluation, high-ranking individuals are reevaluated using high-fidelity electromagnetic field computation. The suppression of the execution of expensive field computations effectively reduces the computing costs. It is shown that the proposed method works two to four times faster, maintaining optimization performance than the original method that does not use surrogate models.