Deep Learning for Nanoscale Arbitrary Meta-element Robustness

Optical metasurface designs often suffer performance degradations originating from the fabrication process due to manufacturing uncertainties. Therefore, a critical need exists for designs which incorporate robustness, but this is typically intractable during optimization due to the large number of function evaluations required. Nevertheless, by augmenting the process with a deep learning intermediate step, the speed of function evaluations can be greatly accelerated, enabling robustness to be directly optimized during the inverse design process rather than as an a posteriori step applied after the fact. A deep learning assisted robustness framework is introduced for application to nanophotonic meta-element design.