Design of Reconfigurable Periodic Structures Based on Machine Learning
Xiaoxing Fang, Huangyan Li, Qunsheng Cao
Abstract
The design of periodic structures is a very time-consuming process because it requires many cycles of full-wave simulations to obtain the desired electromagnetic (EM) properties. This challenge is amplified when designing reconfigurable periodic structures. This article proposes an intelligent method for the reconfigurable periodic structure design (RPSD) based on machine learning. First, an improved physical and mathematical modeling approach is proposed for reconfigurable periodic structures, which enables the application of machine learning in RPSD. Thereafter, we set up distributed predictors for different reconfigurable states to avoid the one-to-many mapping problem. In addition, the networks of these predictors are trained using variational autoencoders (VAEs) to reduce the associated learning difficulties. These enable the obtention of accurate predictors with small sample data. Moreover, the training of these predictors can be easily accelerated by parallel computation because of their independence. Finally, reconfigurable periodic structures can be intelligently designed by the combination of the trained predictors and a stochastic optimization technique. Some examples are provided to verify the practicality and accuracy of the proposed method.