Machine Learning- and Feedback-Enhanced Dipole Trap Loading and Cooling

Loading neutral atoms into optical dipole traps has transformed atomic, molecular, and optical physics, driving significant advancements across diverse areas. For example, single atoms trapped in dipole traps have enabled the development of tweezer arrays, a highly promising platform for quantum information processing due to their scalability, arbitrary connectivity, and long coherence times. In the regime of many atoms, dipole trapping has facilitated the creation of novel degenerate phases of matter, such as Bose-Einstein condensates. In this work, we report progress on integrating machine learning to enhance the loading of neutral atoms into dipole traps, which is a highly-dimensional complex space where classical control algorithms are challenging to implement. We demonstrate, for the first time, machine learning-enhanced cooling of a trapped ensemble directly into a Bose-Einstein condensate using polarization gradient cooling alone. Furthermore, we present advancements toward leveraging reinforcement learning-enabled feedback to load atoms directly into tweezers from optical molasses, eliminating the need for high magnetic field gradients required for magneto-optical traps.