Inverse Design of Nanophotonic Devices based on Reinforcement Learning

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

Zusammenfassung

Photonic integrated circuits are being employed for increasingly complex quantum optics experiments on compact and interferometrically stable chips. The integration of an ever-increasing number of circuit components poses challenging requirements on the footprint and performance of individual nanophotonic devices thus raising the need for sophisticated design algorithms. While various approaches, for instance based on direct search algorithms or analytically calculated gradients, have been demonstrated, they all suffer from drawbacks such as reliance on convex optimization methods in non-convex solution spaces or exponential runtime scaling for a linear increase in user-specified degrees of freedoms. Here we show how reinforcement learning can be applied to the nanophotonic pixel-discrete inverse design problem. Our method is capable of producing highly efficient devices with small footprints and arbitrary functionality. A distributed software architecture allows us to make efficient use of state-of-the-art high performance parallel computing resources. Multiple interfaces to the dataflow of the algorithm enable us to bias the resulting structures for realizing arbitrary design constraints. To demonstrate the broad applicability of our method, we show a wide range of devices optimized in 3D for different material platforms.