Physics-Informed Reinforcement Learning for Nanophotonic Device Design

The increasing demand on a high-performance metasurface requires a freeform design method that can handle a huge design space. Accordingly, various nanophotonic device design schemes have been investigated including the ones based on machine learning. In this presentation, I discuss design approaches based on deep reinforcement learning, which have achieved great success in many different fields spanning video games, the game of Go, protein folding problems, to matrix multiplication algorithms. Reinforcement learning has not yet been actively explored in the field of nanophotonics compared to other machine learning methods such as generative or discriminative models. I will introduce how to apply deep reinforcement learning to design a metasurface beam deflector with large degrees of freedom, and discuss how to increase the sample efficiency by informing the agent of the reinforcement learning using physical constraints that govern the electromagnetic system.