Machine-Learning Optimization of Laser Polarization in a Rubidium Magneto-Optical Trap

Magneto-optical traps (MOTs) are a foundational tool in atomic, molecular, and optical physics with applications for high-precision atomic clocks, quantum information processing, and quantum gas studies. Their performance is sensitive to laser polarization entering the trapping region, particularly in unconventional polarization regimes. We present an undergraduate-driven project to systematically investigate and optimize the performance of an ⁸⁵Rb MOT by controlling and automating laser polarization using computer-operated, motorized quarter waveplates and machine-learning-based optimization. We construct a 3D MOT apparatus with frequency-stabilized laser light, fluorescence detection, and programmable polarization control. Here, we will focus on optimizing the number of atoms trapped, which is proportional to the fluorescence of the atom cloud, as a function of the waveplate configuration. Bayesian optimization, using Gaussian process regression implemented via the Python package M-LOOP, will be used to explore the polarization landscape and predict fluorescence of the MOT. The results will establish polarization-optimization protocols applicable to both conventional and unconventional MOT configurations, improving the robustness and versatility of ultracold atom experiments.