基于贝叶斯优化的磁光阱多参数自主优化系统

磁光阱是一种冷却陷俘原子的装置，磁光阱实验参数的优化是冷原子实验中基础且重要的工作，人工手动优化参数需耗费大量时间，且很难确保最终参数是全局最优的。基于贝叶斯优化的机器学习方法是一种对目标表达式未知、非凸、多峰的量子物理系统进行参数优化的有效方案，该过程通常远快于人工手动调节，且有更大概率找到全局最优值。提出了一种基于贝叶斯优化方法的冷原子多参数自主实时优化实验方案，该方案通过成本函数构造、控制程序编写、贝叶斯算法优化等形成一个可自主优化的闭环系统。实验结果表明，经过约30 min的迭代优化，所提方案可有效完成磁光阱系统的多参数优化，并得到最优的实验结果；所提方案验证了贝叶斯优化方法在多参数物理系统中应用的可行性，通过改进成本函数，还可应用于其他的复杂多参数实验物理系统最优参数快速确定。

Multi-parameter autonomous optimization system of magneto-optical trap based on Bayesian optimization

Magneto-optical trap is a device for cooling and trapped atoms, and the optimization of magnetic optical trap experimental parameters is a basic and important work in cold atom experiments, however, manual optimization of parameters takes a lot of time, and it is difficult to ensure that the final parameters are globally optimal. The machine learning method based on Bayesian optimization is an effective scheme for parameter optimization of quantum physics systems with unknown target expressions, non-convex, and multi-peaks, which is usually much faster than manual adjustment, and has a greater probability of finding the global optimal value. In this paper, a cold atom multi-parameter autonomous real-time optimization experimental scheme based on Bayesian optimization method is proposed, which forms a closed-loop system that can be independently optimized through cost function construction, control program writing, Bayesian algorithm optimization, etc. The experimental results show that after about 30 minutes of iterative optimization, the proposed scheme can effectively complete the multi-parameter optimization of the magneto-optical trap system and obtain the optimal experimental results. The proposed scheme verifies the feasibility of applying Bayesian optimization method in multi-parameter physical systems, and can be applied to other complex multi-parameter experimental physics systems to quickly determine the optimal parameters by improving the cost function.