Grating magneto-optical trap optimization and drift-mitigation based on Bayesian learning

Abstract

The grating magneto-optical trap (GMOT) is a promising approach for miniaturizing cold-atom systems. We demonstrated a real-time parameter optimization and drift-mitigation method for a GMOT system based on Bayesian learning. In a remarkable short convergence period, optimal numbers of cold atoms reached up to 7.7 × 106, which was nearly four times better than with manual optimization. The parameters included not only physical parameters but also mechanical parameters that can control the critical optical alignment. The results experimentally demonstrate that our work can efficiently optimize multiple parameters for a GMOT system and for the atom-based systems that need fine control. The machine learner employed a dual layer Bayesian learning, which could suppress the cost function drift due to the instability of the experimental parameters and environmental factors of the GMOT. The proposed approach validates the feasibility of Bayesian optimization in multiparameter cold-atom systems and can be applied to rapidly determine optimal parameters and high stability of general cold atom-based physical systems.