Use of machine learning to maximize loading efficiency of cold cesium atoms into a hollow-core photonic crystal fiber

We report the results of a machine-learned algorithm for optimal loading of cold cesium atoms into a hollow-core photonic crystal fiber. Light-matter interactions using atomic ensembles has remained a popular means for quantum communication and quantum memories. However, interacting with single photons using cold atomic ensembles generally requires a large number of atoms within a tightly-focused beam. We accomplish this by inserting cold atoms, prepared in a magneto-optical trap, into a vertical segment of hollow-core fiber using gravity and an attractive dipole beam. The loading efficiency depends on many parameters which dictate the initial position, size, and temperature of the cloud, as well as the time taken to reach the fiber core. Many of these parameters are coupled to one another and performing a full sweep of the parameter space would be time consuming. Thus, the preference is to intelligently scan through the parameter space to find optimal settings within a reasonable timeframe. From here, the machine-learning algorithm M-LOOP was used in conjunction with our experiment control program to maximize the number of atoms loaded inside the fiber. However, the success of the algorithm depends greatly on the feedback received (cost function) accurately represents the end result we desire with minimal uncertainty. Using optical depth requires scanning many detunings to get a single cost function and is very sensitive to probe coupling through the fiber. Instead, we use a method called bleaching where the atom number is determined by counting photons missing from a resonant pulse sent through the fiber.