This work benchmarks nine different machine learning algorithms for optimizing a cold atom experiment, specifically a rubidium molasses system with 10 and 18 adjustable parameters, using the post-cooling atom number as the optimization target. The JCMsuite's Bayesian optimization tool was implemented and extended with a noise-aware strategy (Noisy Expected Improvement) to efficiently handle the inherently noisy experimental data. The benchmarking demonstrated that this enhanced appraoch performed best in terms of speed and final atom number, particularly in high-noise and high-dimensional scenarios, showcasing its utility for automating complex experimental tuning.
O. Anton, et al. Review and experimental benchmarking of machine learning algorithms for efficient optimization of cold atom experiments. Mach. Learn.: Sci. Technol. 5, 025022 (2024).