Optimizing a guided atom interferometer using machine learning

Atom interferometers that are based on Bose-Einstein condensates (BEC) are useful in sensing rotations. We realize our Sagnac atom interferometer by splitting, reflecting, and recombining the condensate with a sequence of standing-wave Bragg pulses inside a waveguide that is transversely translated. Two important factors impacting the performance of such a device are a delta-kick cooling (DKC) procedure and the precise timing between the reflection pulses (mirror-mirror timing). These factors control the collimation of the wave-packets during the interferometer interrogation cycle and the overlap of the wave-packets at recombination, respectively. Manual optimization of the experimental parameters related to these factors is often time-consuming and challenging because the parameters are not independent. Moreover, recent studies have demonstrated that many machine learning (ML) schemes may significantly outperform manual tuning efforts in complex cold atom experiments and they do so in much shorter times. We demonstrate a guided atom interferometer implemented into an automated learning scheme which can be called as a subroutine. Our ML workflow is designed to minimize the number of calls to the experiment for faster optimization. We are applying ML optimization to the various stages of BEC generation as well as the DKC and mirror-mirror timing parameters of our guided atom interferometer.