Image-based reinforcement learning for robust mode hop recovery in external-cavity diode lasers

Conventional laser frequency stabilization techniques such as table-lookup or heuristic methods have significant limitations due to intrinsic hysteresis and stochastic mode competition. To overcome this challenge, we introduce a simple hysteresis-aware protocol and a tailored image-based reinforcement learning (RL) framework. This approach enables robust mode hop recovery in external-cavity diode lasers (ECDLs). Using a Rainbow deep Q-network (DQN) trained entirely in simulation, our RL agent achieves 100% recovery success rate across all 163 continuous temperature drifts produced under deliberately extreme perturbations during a three-day stress test. Under challenging discrete test conditions with abrupt current jumps, the success probability reaches 99.9%, with only one failure recorded out of one thousand trials. The controller generalizes well across environmental perturbations, such as temperature fluctuations, without requiring additional calibration. Our approach provides real-time adaptive control with minimal human intervention, marking a step toward fully autonomous operation in quantum experiments and precision optical systems.