Identifying Gravitational Microlensing Events in Photometric Light Curves using a Deep Neural Network

We present an evaluation of a neural network pipeline applied to gravitational microlensing detection and event classification, with a focus on planetary microlensing. Our generalized pipeline is designed to automatically identify and characterize various types of transient variabilities in photometric light curves. In a previous work, we applied it to identify new 181 exoplanet transit candidates using light curve data from the Transiting Exoplanet Survey Satellite (Olmschenk, Ishitani Silva, Rau, et al. 2021). Here, we present the results of our pipeline applied to 549,447 previously labeled light curves acquired from 2006 to 2014 by the Microlensing Observations in Astrophysics (MOA) collaboration. The MOA collaboration presented a sample of gravitational microlensing events and reported the planet frequency as a function of planet-to-star mass ratio (Suzuki et al. 2016). The detection of the events of this sample relies on the MOA alert system event identification, which favors events resembling single-lens events. Our neural network approach does not have any intrinsic biases toward single-lens events, which provides the potential to identify additional planetary microlensing events. We expect to use our pipeline to determine the detection efficiencies within the MOA data set for all mass ratios, which is required for the statistical understanding of the exoplanet distribution. We evaluate this pipeline as an alternative planetary microlensing detection method. To make a prediction for a given light curve, our network requires no prior microlensing parameters identified using other methods. Additionally, it performs inference on a MOA light curve in a few milliseconds on a single GPU, enabling large-scale archival searches.