We present results from a NASA Frontier Development Lab (FDL) project to automatically classify candidate transit signals identified by the Kepler mission and the Transiting Exoplanet Survey Satellite (TESS) using deep learning techniques applied with compute resources provided by the Google Cloud Platform. NASA FDL is an applied artificial intelligence research accelerator aimed at implementing cutting-edge machine learning techniques to challenges in the space sciences. The Kepler and TESS missions produce large datasets that need to be analyzed efficiently and systematically in order to yield accurate exoplanet statistics as well as reliably identify small, Earth-sized planets at the edge of detectability. Thus we have developed a deep neural network classification system to rapidly and reliably identify real planet transits and flag false positives. We build on the recent work of Shallue & Vanderburg (2018) by adding "scientific domain knowledge" to their deep learning model architecture and input representations to significantly increase model performance on Kepler data, in particular for the lowest signal-to-noise transits that can represent the most interesting cases of rocky planets in the habitable zone. These improvements also allowed us to drastically reduce the size of the deep learning model, while still maintaining improved performance; smaller models are better for generalization, for example from Kepler to TESS data. This classification tool will be especially useful for the next generation of space-based photometry missions focused on finding small planets, such as TESS and PLATO.
Link to PDF (may not be available yet): O1-1.pdf