Simultaneous Inverse Design and Uncertainty Quantification for Frequency-Selective Rasorber With Tunable and Switchable Abilities by Bayesian Deep Learning

Abstract

A Bayesian deep learning scheme is proposed for simultaneous inverse design and uncertainty qualification (UQ) for frequency selective rasorber (FSR) with switchable and tunable (S/T) abilities. The inversely designed FSR could work in single/two passband modes with bilateral absorption bands, where the tunable passband is controlled by varactor diodes, and the number of passbands is switched by elaborately designed bias lines. Further, the constraints of resonance are embedded into the inverse-design process based on the equivalent circuit model (ECM). In the uncertainty quantification process, both data uncertainty and model uncertainty of predicted S-parameters are modelled by the Bayesian neural network (BNN), whose effectiveness is verified by the correlation coefficient between true error and predicted uncertainty. At last, the inversely designed FSR sample is manufactured and measured, where the electromagnetic (EM) responses including S-parameters and absorption bands verify the accuracy and efficiency of the proposed method.