A Machine Learning Approach to Discovering Exoplanets Transiting Old Stars
This random forest algorithm categorizes old stars as exoplanet hosts based on transits found in Transiting Exoplanet Survey Satellite photometry.
Most stars have planets orbiting them, and some cross in front of the host star causing a temporary decrease in brightness known as a transit. These transits have measurable depth, period and duration. Only ~6% of known transiting planets orbit older host stars, and the rest orbit stars with ages similar to our sun1. As a result of stellar evolution, transits of old stars are sometimes missed by transit searches of young stars. The goal of this large-scale search is to quickly discover enough exoplanet candidates orbiting old stars to assist population level analysis. A manual search for exoplanet transits creates an unreasonable timeline. We run Box Least Squares (BLS) on 90,000 old stars using photometry from the Transiting Exoplanet Survey Satellite (TESS). BLS is a period search algorithm that fits parameters such as orbital period, depth, duration and signal to noise of planet transits whether or not there is a real planet. TESS is a near all-sky space-based photometric survey with a primary mission of discovering transiting exoplanets. Using BLS from a set of injected planet transits and a set of inverted lightcurves, we are able to train a random forest algorithm to classify old stars into hosts or non-hosts. This allows us to reduce vetting time greatly. We identified 17 previously known exoplanet candidates in our set, which shows that our methods are successful. We then discovered 32 new exoplanet hosts, which we are following up with ground based observations currently.
