Robust supervised learning for closed loop adaptive optics predictive control

Great progress has been made applying deep learning methods to adaptive optics (AO) control, focus has largely been on reinforcement learning (RL) methods. While RL is a powerful tool and shows promising results, it requires continual learning while on sky to truly be effective. This makes it difficult to apply optimization techniques, such as kernel compilation, pruning, or – in the most extreme cases – hard coded networks in hardware, which may be necessary for high speed extreme AO control. We present a method and optical bench results for supervised training of AO predictive control networks trained using only simulated data. This can be accomplished by varying both the optical parameters of the AO system as well as the parameters of the simulated atmosphere; teaching the network to generalize for optical as well as atmospheric conditions. Our method also alleviates issues with both online and supervised learning methods trained on saved telemetry which may over-fit to local conditions that can vary from night to night. This training methodology is general enough to be widely applicable among most AO systems and has proven to be effective in our optical bench experiments.