Optimization and Stabilization of Cooling Processes of Neutral Atoms with Machine Learning

When trapping and cooling atomic clouds for cold-atom experiments, cycle-to-cycle variations in the number of atoms can lead to unwanted fluctuations in final measurements. External parameter fluctuations, such as in ambient temperature and stray magnetic fields, are leading contributors to these variations, in addition to instabilities within the system itself. To understand the influence of each parameter on the atom number, we seek to measure and record many parameters over many experimental cycles and construct a model of the atom-number response.

This is a high-dimensional problem, to which machine learning lends itself well. A neural network is used to model the atom number at various stages of our atom cooling system as a function of several parameters measured around the laboratory. Standard methods in neural network optimization improve computation time, and accuracy of out-of-sample predictions. Dimensionality reduction, adaptive learning rates when optimizing the model, and regularization by dropping connections of neurons yielded robust and consistent predictions.

Using this approach, we develop a neural-network model that successfully predicts our atom number in experiments, and we are working towards using this information to reduce instabilities in the atom number.