Automatic inverse design of second-order differential metasurfaces based on reinforcement learning

Second-order differential metasurfaces, as two-dimensional artificial nanostructures, hold great potential for various applications in optical computing, image processing, microscopy imaging, etc. They have the capability to perform second-order mathematical operations on incident optical fields. In general, the design of second-order differential optical processing metasurfaces needs to consider various constraints, including incident optical fields, material properties, and application requirements. This task is challenging and time-consuming, typically requiring manual intervention. Therefore, a reinforcement learning-based artificial intelligence algorithm is proposed for the automatic design of second-order differential metasurfaces. It can rapidly find the desired result by leveraging the powerful self-learning ability of reinforcement learning. This paper first introduces the basic principles of designing second-order differential optical processing metasurfaces. Then, it presents a new automatic design method for second-order differential metasurfaces based on reinforcement learning. Finally, a design of a second-order differential metasurface is presented. Simulation and experimental results validate the effectiveness of the proposed method. The proposed method establishes a basic framework for applying reinforcement learning to the automatic design of photonic devices and outlines the various constraint factors that need to be coupled and considered during the design process. A surrogate model based on neural networks is presented to fit the transmission calculated by finite difference time domain software. A new reward formula is given to sufficiently consider the transmission efficiency of whole devices. The proposed method effectively leverages the self-learning capability of reinforcement learning, circumventing challenges related to manual intervention, computational complexity, and time-consuming tasks in conventional methods. • Reinforcement learning is applied to metasurfaces inverse design. • Neural network based transmission calculation surrogate model is presented. • New reward formula considering transmission efficiency is given.