Ultra-fast metalens optimization using conditional physics-augmented deep learning surrogate solvers

Physics-augmented training schemes have enabled the use of ultra-fast deep learning surrogate solvers in adjoint optimization algorithms, accelerating design by many orders of magnitude. However, the utility of these solvers for device design is severely limited by their inability to function outside of fixed simulation parameters. We present a foundational method for conditioning deep learning surrogate solvers on arbitrary parameters, such as the source incidence angle. We then demonstrate the capability of a conditional deep learning model to optimize high-efficiency aperiodic metapixel deflectors that are fashioned to create a large-area metalens.