A hybrid algorithm of differential evolution and machine learning for electromagnetic structure optimization

Abstract

Various electromagnetic (EM) structures become more complex and often have increasing degrees of design freedom. Classical optimization methods require numerous simulation trials of different parameter combinations, which leads to a low design efficiency. To address this problem, an efficient EM structure optimization algorithm which combines differential evolution (DE) with machine learning technology is proposed in this paper. By partly substituting electromagnetic (EM) solver, Kriging model predicts the responses and uncertainties of each individual after differential evolution. The exploration and exploitation of the searching can be adjusted by the constitution and prescreening of the population before and after evolution. This algorithm is applied to optimize the resonant frequencies of an E-shaped antenna with 6 variables. Comparing with the other optimization methods, the number of EM simulations needed is reduced by about 80%.