Nondestructive Characterization of Laser-Cooled Atoms with Machine Learning

The light emitted from fluorescing clouds of laser-cooled atoms provides structural information with known techniques to extract an estimate of atom number, but other ensemble properties, namely temperature, remain inaccessible until destructive techniques such as time-of-flight imaging measure the cloud's expansion at the cost of the trap's release. We demonstrate the use of machine learning models to predict the atom number and temperature of vapor-fed potassium-39 magneto-optical traps from only in-situ fluorescence images. A dataset spanning a wide range of trap loading conditions was collected, capturing dense atomic ensembles to no detectable cloud signal, while destructive techniques were solely used for label generation. Several machine learning models with increasing complexity are employed to infer the number and temperature, along with respective signal-to-noise ratios for quantifying prediction confidence; the trained models have no access to the time-of-flight data. Models range from simple linear regression to convolutional neural networks, with the best models achieving fractional uncertainties of 0.1 in atom number and 0.2 in temperature. The results of this work mark a step towards real-time, nondestructive feedback and characterization of cold atomic ensembles.