Reinforcement Learning for Optimization and Control of Ultracold Quantum Gas Production

Abstract

Machine-learning techniques are emerging as a valuable tool in experimental physics, and among them, reinforcement learning offers the potential to control high-dimensional, multistage processes in the presence of fluctuating environments. In this experimental work, we apply reinforcement learning to the preparation of an ultracold quantum gas to realize a consistent and large number of atoms at microkelvin temperatures. This reinforcement learning agent determines an optimal set of thirty control parameters in a dynamically changing environment that is characterized by thirty sensed parameters. By comparing this method to that of training supervised--learning regression models, as well as to human-driven control schemes, we find that both machine learning approaches accurately predict the number of cooled atoms and both result in occasional superhuman control schemes. However, only the reinforcement learning method achieves consistent outcomes, even in the presence of a dynamic environment. This thesis provides a comprehensive overview of the theoretical groundwork necessary for understanding the experimental sequence and machine learning techniques employed. Technical details of the cooling apparatus and machine learning agents are presented, along with the results of allowing the trained machine learning agents to autonomously control the cooling sequence.