Automated Equivalent Circuit Modeling of Frequency Selective Surfaces Using Self-Supervised Neural Networks

A self-supervised convolutional neural network (SCNN) is proposed to automatically extract equivalent circuit parameters of frequency selective surfaces (FSSs). The network directly maps the reflection coefficients (RCs) of a free-standing FSS under normal incidence to its equivalent inductance and capacitance, bypassing manual identification of transmission poles and zeros. Leveraging self-supervised learning, the SCNN reconstructs RCs from the extracted parameters and minimizes the mean-square error (MSE) against input RCs, eliminating the need for labeled data. The framework is further extended to incorporate dielectric substrates and oblique incidence, enabling efficient analysis of multilayered FSS via integration with the transmission matrix method (TMM). The proposed approach achieves high accuracy across a broad frequency range, up to the grating lobe regime, with remarkable computational efficiency.