Structural Optimization of a One-Dimensional Freeform Metagrating Deflector via Deep Reinforcement Learning

• DOI: 10.1021/acsphotonics.1c00839
• Publication Date: 2021-12-30T20:41:39Z
• Authors: Dongjin Seo; Daniel Wontae Nam; Juho Park; Chan Y. Park; Min Seok Jang
• Keywords: Robustness; Deep Neural Networks

The increasing demand on a versatile high-performance metasurface requires freeform design method that can handle huge space, which is many orders of magnitude larger than conventional fixed-shape optical structures. In this work, we formulate the designing process one-dimensional Si beam deflectors as reinforcement learning problem to find their optimal structures consistently without requiring any prior data. During training, deep Q-network-based agent stochastically explores device space around learned trajectory optimized for deflection efficiency. devices discovered by agents show overall improvements in maximum efficiency compared ones state-of-the-art baseline methods at various wavelengths and angles. Furthermore, efficiencies generated trained from different neural network initializations have small variance, demonstrating robustness proposed method.