A novel DOE surface design method based on reinforcement learning architecture

Abstract The surface profile of the diffractive optical element (DOE) plays a critical role in modulating the intensity and phase of the incident optical field. In this paper, we propose a novel surface profile design method for DOE, based on the Soft Actor-Critic (SAC) deep reinforcement learning algorithm. This approach effectively overcomes the limitations of traditional optimization methods for continuous surface DOE, which are often prone to being trapped in local optima within high-dimensional, non-convex parameter spaces. By leveraging the SAC algorithm’s capability for global optimization in continuous high-dimensional spaces, we construct a multi-objective reward function to simultaneously balance beam splitting performance and other key metrics. The proposed method successfully achieves DOE surface profile designs under uniform beam splitting objectives for 2–5 beams. The corresponding uniformity errors of the output optical field are as low as 0%, 0.21%, 0.29%, and 0.46%. The root-mean-square surface profile errors are 0.005%, 1.38%, 2.48%, and 3.77%, respectively. The average diffraction efficiency exceeds 98.5%. This method offers a solution for the continuous surface design of DOE and demonstrates significant potential for practical applications in fields such as laser processing and optical communication.