Identifying gravitational microlensing and exoplanet transits using the RAMjET deep neural network

We present RApid Machine-learnEd Triage (RAMjET), an architecture that automatically identifies and characterizes events in photometric light curves using machine learning techniques. We created this tool to meet the need for an efficient and accurate approach to triage an enormous quantity of light curves for further analysis. Thus far, RAMjET consists of a deep neural network pipeline and has been trained using two separate primary data sets for different detection goals: (1) gravitational microlensing and (2) exoplanetary transits. For (1) detecting gravitational microlensing, we use the nine-year data set consisting of 600,000 previously classified light curves acquired by the Microlensing Observations in Astrophysics (MOA) collaboration. The MOA collaboration has previously reported the planet frequency as a function of planet-to-host star mass ratio and separation (Suzuki et al. 2016) and it is the first high cadence microlensing survey towards the Galactic bulge using the purpose-built 1.8m MOA telescope located at Mount John Observatory, New Zealand. Currently, our neural network is able to automatically detect 94% of the known microlensing events in the MOA data set. Additionally, we are performing archival searches with the network to identify previously undetected events, finding a dozen new candidates. Moreover, by training alternately with known microlensing events and synthetically generated events, we will use RAMjET to determine detection efficiencies within the MOA data set. For (2) detecting exoplanet transits, we train our network using light curves with confirmed planet transits based on the Exoplanet Follow-up Observing Program for the Transiting Exoplanet Survey Satellite (TESS). TESS is an all-sky survey mission expected to discover thousands of exoplanets around nearby bright stars. Similar to the microlensing case, our neural network - trained using the known candidates - will perform an archival search to detect exoplanet transits which may be overlooked by traditional handcrafted methods. Up to this moment, a dozen new exoplanet transit candidates have been found. References: Suzuki, D., et al. "The exoplanet mass-ratio function from the MOA-II survey: Discovery of a break and likely peak at a Neptune mass." The Astrophysical Journal 833.2 (2016): 145.