A High-Efficiency and Effectiveness Designing Methodology for Discrete-Coded Energy Selective Surface Based on Machine Learning

This article introduces an innovative methodology for the design of energy selective surfaces (ESSs) using machine learning (ML). The methodology differs from traditional ESS design approaches that rely on equivalent circuit models (ECMs). Instead, it involves the training of a forward prediction neural network for the electromagnetic performance of ESS and utilizing an optimization algorithm to achieve the optimal ESS structure. The optimization objective is to align the transmission pole in the transmitted state and transmission zero in the protection state with the target frequency point, achieving a minimal insertion loss (IL) and a maximal shielding efficiency (SE) at the same time. The designed ESS structure consists of a metal frame, a nonlinear induction region, and a resonant region formed by a randomly distributed 0/1 matrix, whose electromagnetic performance varies with the distribution of the 0/1 matrix. Through multiple simulations, 6000 sets of training data are obtained and used to train a fully connected (FC) neural network. Subsequently, the particle swarm optimization (PSO) algorithm is employed to optimize the objective function. Following these procedures, several ESS structures are successfully obtained. After prototype fabrication and testing, all samples demonstrate consistency with expected results by achieving a low IL and a high SE near the target frequencies.