Establishing exhaustive metasurface robustness against fabrication uncertainties through deep learning

Photonic engineered materials have benefitted in recent years from exciting developments in computational electromagnetics and inverse-design tools. However, a commonly encountered issue is that highly performant and structurally complex functional materials found through inverse-design can lose significant performance upon being fabricated. This work introduces a method using deep learning (DL) to exhaustively analyze how structural issues affect the robustness of metasurface supercells, and we show how systems can be designed to guarantee significantly better performance. Moreover, we show that an exhaustive study of structural error is required to make strong guarantees about the performance of engineered materials. The introduction of DL into the inverse-design process makes this problem tractable, enabling optimization runtimes to be measurable in days rather than months and allowing designers to establish exhaustive metasurface robustness guarantees.