Otimização via machine learning de armadilha magneto-óptica

Abstract

Laser-cooled samples have been around for decades and have helped to produce great science contributions. More recently, machine learning (ML) has evolved a lot and has become an efficient means to investigate empirical models of complex systems. Typically, the complex dynamics presented by many-body interaction systems preclude precise analytical optimization of cooling mechanisms and capture. In this project, fundamental machine learning methods will be applied to optimize the preparation of neutral Rubidium-87 atoms, as it is one of the fundamental stages concerning the experimental realization of Bose-Einstein condensates. More specifically, the aim is to increase the number of trapped atoms in a magneto-optical trap (MOT) by establishing a connection between the physical model that describes the MOT operation and the function related to the ML algorithms’ performance during the optimization (i. e. the cost function). For this purpose, three widely known ML models will be investigated: Differential Evolution, Gaussian Process and Artificial Neural Networks. The solutions found by these algorithms tend to be radically different from the adiabatic, analytic solutions currently used by researchers and will be studied and compared. Despite this, the new solutions are expected to overcome the combination of previously known parameters optimizing the preparation of trapped atomic samples.