Efficient EM Optimization Exploiting Parallel Local Sampling Strategy and Bayesian Optimization for Microwave Applications

Abstract

In electromagnetic (EM) optimization of microwave design, a computationally bad starting point usually leads the local optimization process to be stuck into local optimum, which does not satisfy the design specifications. In this situation, global optimization methods can be an alternative to achieve the final optimal solution. However, global optimization methods always suffer from a relatively low convergence rate. To address this problem, we propose an efficient EM optimization technique with a novel parallel local sampling strategy and Bayesian optimization (BO) for microwave applications in this article. We develop a new parallel local sampling strategy to increase the exploitation ability near the potential optimal solution in each optimization iteration and improve the convergence of the entire optimization process. The local sampling range in each iteration is determined based on the derivative information of the current potential optimal solution. While conventional BO only uses the information of potential optimal solutions in each iteration during optimization, we propose to exploit both the generated local samples and the potential optimal solutions together to build a surrogate model and guide the optimization. Therefore, the exploration and exploitation during the proposed EM optimization are well balanced, and the entire EM optimization process is effectively accelerated in comparison to other existing global methods. Examples of EM optimizations of microwave components are used to demonstrate the proposed technique.