Reinforcement learning-enabled robust phase control for OAM beam generation in coherent beam combining systems

Orbital angular momentum (OAM) beams play a vital role across diverse scientific and technological frontiers. However, in high-power regimes, intra-cavity generation is constrained by mode competition, while extra-cavity conversion suffers from low damage thresholds. Coherent beam combination (CBC) provides a promising pathway by combining multiple lasers into structured beams, yet achieving intelligent, label-free, and high-precision phase control remains a formidable challenge. Here we present a reinforcement learning-based phase control framework tailored to generate OAM beams in CBC systems, featuring a physically informed reward function and a vector-quantized (VQ) module embedded in both the actor and critic networks. The proposed framework enables rapid and robust single-stage generation of ±1 and ±2 OAM beams in a 12-channel CBC system, achieving mode purities exceeding 0.99 in noisy environments. Remarkably, it surpasses the stochastic parallel gradient descent (SPGD) algorithm in control stability, despite operating at a much lower control frequency. This work advances the development of RL-based phase control in CBC and establishes a scalable and physically grounded paradigm for intelligent optical field modulation.