From Noise to Pattern: Inverse Design of FSS Using Variational Autoencoder,

Abstract

Frequency Selective Surfaces (FSSs) are critical for modern electromagnetic filtering, but their design traditionally relies on costly trial-and-error simulations. We present a generative machine learning framework for the inverse design of FSS unit cell patterns. By utilizing a conditional variational autoencoder (cVAE), the proposed method directly maps desired electromagnetic scattering parameters (S-parameters) to FSS patterns, thereby circumventing the traditional energy- and time-expensive trial-and-error approach inherent in FSS design. A dataset of 10,000 simulated (pat tern, S-parameter) samples was generated using Ansys HFSS over the 2 to 8 GHz frequency range to train both a surrogate neural network, which accurately predicts the S-parameters from a given pattern, and a cVAE-based generator that synthe sizes novel pattern designs conditioned on target frequency responses. The integrated framework employs a gradient-based optimization strategy in the la tent space to minimize the deviation between the predicted and desired S-parameter responses, with particular emphasis on preserving the resonant frequency. Bench marking on the test dataset demonstrates that the surrogate model achieves mean absolute errors from 0.5 dB at 2 GHz to 1.9 dB at 8 GHz, while the optimization loop refines designs to yield deviations as low as 0.0-0.2 GHz at the resonant fre quency for half of the samples. These results underscore the promising potential of generative machine learning for rapid FSS inverse design.