Autonomous analysis of excitations in Bose-Einstein Condensates

Solitons are non-dispersive waves that can occur in many systems, at scales ranging from microscopic, to terrestrial and even astronomical. In Bose-Einstein condensates (BECs), the parameters governing the atomic cloud are under strict experimental control and can be manipulated to contain solitonic excitations including conventional solitons, vortices, and many more. Current research looks to understand the dynamics of solitonic excitations by producing images of atomic clouds that need to be analyzed. Recent work [1] has developed machine learning algorithms to determine if solitonic excitations were created in a BEC, and to classify them into physically-motivated subclasses, such as a well defined single soliton, a partial soliton, or solitonic vortex. This presentation will discuss our current work across two fronts. First, improvements have been made in estimating the quality of multiple solitons appearing close together and thus interfering with each other. For such cases the standard quality metric introduced in Ref. [1] tends to rate both solitons as "bad" quality which is often incorrect. Next, we are testing clustering methods to better assess the expected number of physically-motivated subclasses and random decision forests to aid with proper parameter thresholds for each subclass. Our goal is to make the soliton detection software package [2] more adaptable to a wider array of BEC experiments.