A Novel Method of Transit Detection Using Parallel Processing and Machine Learning

Abstract

Ultra-short-period (USP) planets are rare Earth-sized planets with the shortest possible orbital periods  of all known planets. The study of this group is important for investigating planet formation and evolution processes. To date, only slightly over 100 USPs have been detected in Kepler  photometry data of nearby FGKM dwarfs. However, traditional methods used in detecting planets and transit-like events are often biased, inefficient and time-consuming. For the first time, we introduce a GPU fast phase folding technique coupled with a Deep Convolutional Neural Network (DCNN) specifically used for searching for USP planets. The DCNN is trained on a set of 2,000,000 synthetic USP samples and performs exceedingly well in identifying both true and false positive transit signals, with a 99.5% validation accuracy over the given training set. With the computational power provided by GPU fast phase folding, our method, compared to the traditional Box Least Squares method, has shown to be ~1000 times faster in searching for transit signals in a photometric light curve with the same or better precision and recall rate. Furthermore, our method can also be applied to other interesting planet populations beyond USP planets. We used this method to search through all available KIDs in the Kepler database and were able to reproduce all existing USP planets with a 100% recovery rate. We also discuss how further adjustments can be made, making this system even more efficient and powerful, as well as how it can be applied to broader planet populations.