Author(s): , ,
Institution(s): 1. Max Planck Institute for Intelligent Systems, 2. New York University
The planet discoveries made using data from the Kepler mission have revolutionized the field of exoplanet statistics. Thousands of planets have been discovered with orbital periods ranging from hours to two years. Some of the most dynamically interesting planets (Jupiter analogs, for example) only show a single transit in the four-year baseline of the Kepler mission and, as a result, they have not yet been found. Upcoming transit surveys like K2, TESS, and PLATO all have shorter contiguous observation baselines. It is therefore crucial to develop robust techniques for the discovery of single transit events. We present a search procedure designed to find single transits using a supervised classification model. To search for a transit signal in a given month-long light curve from Kepler, we train a random forest classifier on tens of thousands of simulated transit signals injected into the light curve of the same star at other times. Using a different set of simulations, we tune the model to maximize the precision of the recovered signals---minimizing the false alarm rate. With this model, we classify each section of the test light curve into two categories: transit and no transit. We demonstrate that this model is robust to systematic false positives and present an automated catalog of convincing single transit candidates.