Adaptive phase locking in coherent beam combining with deep reinforcement learning

Locking of optical coherence by single-detector electronic-frequency tagging (LOCSET) stands as the main horse for locking the phase among different channels in coherent beam combining (CBC) systems. However, the coherent demodulation in LOCSET is manually synchronized, and vital hyperparameters of LOCSET are manually determined, and cannot adapt to encountered disturbances. Therefore, the upper limit of LOCSET remains unexplored. Here, a deep-reinforcement-learning-enabled adaptive phase-locking strategy for LOCSET in CBC systems is proposed, which consists of two well-trained agents: the Delay-Agent and the Ada-Agent. The Delay-Agent identifies the fitting phase delay configuration for realizing automatic synchronization in coherent demodulation, and the Ada-Agent dynamically optimizes the proportional coefficient and integration time in response to encountered disturbances. The proposed scheme is experimentally validated in a four-channel CBC system, where the Delay-Agent can locate the fitting phase delay combination in one step. Compared to bare LOCSET, the Ada-Agent brings 4.5-fold improvement in recovery speed under both sinusoidal and step-wise phase disturbance in experiments. We believe the proposed adaptive phase-locking strategy can be applied in situ to enhance the phase-locking performance of LOCSET in various CBC systems.