Robust Forward Surrogates for Free-Form Frequency Selective Surfaces With Uncertainty Quantification by Bayesian Deep Learning

The free-form frequency selective surface (FF-FSS) is modeled with a binary matrix with shape of (128, 128), approximating a continuous designable space. The free-form representation enables the realization of various functionalities across a wide frequency band, including band-pass, band-stop, as well as more diverse responses. To overcome the computational burden of full-wave simulations, an efficient forward surrogate based on Bayesian deep learning is introduced, capable of predicting electromagnetic (EM) responses (e.g., transmission coefficients) and simultaneously realizing uncertainty quantification (UQ). Specifically, the validation of these predictions is demonstrated using both statistical and individual results, and the model’s capability to capture uncertainty is verified through topological deformation and transmission coefficient noise. Furthermore, the proposed framework exhibits exceptional robustness across distinct FSS patterns, i.e., artificial parametric structures and lower-dimensional binary matrix. Additionally, its expanded applications, such as those involving distinct periodicities, oblique incidence and two-layer FF-FSSs, are also investigated. Finally, we present an efficient pre-screening methodology for identifying candidate designs for specific targets, where the forward surrogate yields similar responses with distinguished uncertainty levels. The practical applicability of our approach is experimentally validated through the fabrication and measurement of two prototypes tailored to specific design targets, each exhibiting different prediction uncertainties.