A Novel Approach to Nanophotonic Black-Box Optimization Through Reinforcement Learning

Butz, Marco; Leifhelm, Alexander; Becker, Marlon; Risse, Benjamin; Schuck, Carsten

Abstract

After the use of Photonic integrated circuits (PICs) has led to a significant increase in the performance of devices employed in classical telecommunication schemes in the last years, complex quantum optics experiments have recently undergone a similar transition from free space setups to PICs. This development poses challenging requirements on the PICs' individual components in both footprint and performance and even raises the need for novel functionalities that are not accessible by conventional design methods. Recently, various design algorithms addressing this problem have been demonstrated. However, they all suffer from various drawbacks such as reliance on convex optimization methods in non-convex environments or the presence of gradient fields, which cannot always be accessed easily. Here, we show a novel inverse-design method based on reinforcement learning capable of producing pixel-discrete nanophotonic devices with arbitrary functionality and small footprints. Freely configurable design constraints can be realized through multiple interfaces enabling manipulation of the internal data flow. To demonstrate the capabilities of our method we show the fully automated design of a silicon-on-insulator waveguide-mode converter with > 95% conversion efficiency from scratch.