My goal was to develop an improved process to search for habitable small planets (HSPs), rocky planets similar in size to Earth that might be able to sustain surface-level liquid water. Currently, we know of less than 60 such planets, and their small size and long orbital periods make them particularly challenging to detect. However, we really care about finding HSPs because they can tell us about how worlds like Earth might develop, evolve, and “fit in” in the universe. To search for HSPs, I created a neural network model designed and trained to identify dips in a star’s observed brightness from potential transiting HSPs, as well as a method for preparing input data that is much more efficient than existing ones. This neural network model correctly identified known HSPs’ transits and, while analyzing star systems in the Kepler catalog, detected very strong signals, which, upon further examination, could potentially originate from 6 previously-unknown planets. This is big news because if these really are HSPs, a new opportunity to learn about the development, evolution, and properties of habitable environments opens up, which could help shed light on Earth’s place in the universe and the possibility of life occurring elsewhere.
Habitable small planets (HSPs) are a uniquely interesting population of exoplanets. Defined as terrestrial planets roughly of Earth’s size, orbiting in their star’s habitable zone, HSPs provide a window to better understanding the development of planets in general and Earth-like worlds in particular, as well as helping place the solar system within the broader context of the Milky Way, and guiding the search for life outside of Earth. Currently, there are fewer than 60 confirmed HSPs, making more comprehensive population studies difficult. This paper presents a set of new and enhanced methodologies to facilitate more efficient discovery of HSP candidates in targets from the Kepler catalog. They include a windowed fastfolding method that leverages GPU parallel processing to quickly process lightcurve data, an enhanced method for simulating HSP transit signals to rapidly generate a large dataset of positive samples for neural network training, and a deep neural network (DNN) trained with the population of synthetic samples and demonstrated to be able to successfully recover 11 out of 11 known HSPs with periods in the 100-day to 200-day range with high confidence. Focusing on identifying potential small planets orbiting F-, G-, and K-dwarfs with periods between 100 and 200 days, these methodologies were then applied to process and analyze lightcurve data from 148 Kepler targets, resulting in the discovery of 2* new HSP candidates.
*At time of submission to Regeneron, I had found 6 signals that seemed to be from potential candidates, but after further analysis, it was narrowed down to 2 most-likely ones.