Uncertainty Qualification for Metasurface Design with Amendatory Bayesian Network

Abstract Having a prophetic ability to evaluate the uncertainty of deep learning is important to enable the critical reception of the output result. This is especially pronounced in the emerging domain of intelligent metasurfaces, due to the ubiquitous uncertainties from realistic fabrication and network modeling. Despite the great advancements that have mutated the design and working modality of metasurfaces, this enticing ability remains elusive. Here, a new paradigm to quantify the uncertainty in metasurface design is proposed by generalizing the Bayesian neural network. The uncertainty generally originates from the network model part and data part, the latter of which is imitated by the topologically‐distorted encoding method. The conventional Bayesian neural network is revised by embedding physical‐inspired elements to make it exclusive for metasurface design case. Taking a microwave metasurface as an example, such an approach is benchmarked by simultaneously yielding predicted results and specific uncertainty and also providing experimental reliability for different metasurface manufacturers. This work ushers in a fathomable tool to help users make better decisions for deep learning output, meriting other research domains of optics and materials science.