Identifying Exoplanets with Deep Learning

Exoplanets are entering the big data era, but the increasing volume of space data presents a processing challenge. Artificial intelligence algorithms have tremendous promise for taking full advantage of the avalanche of data. We present an update on the state-of-the-art in transiting exoplanet detection and vetting using artificial intelligence. In particular, we will present the latest developments of the Astronet-Triage algorithm being used in production to identify TESS planet candidates from the MIT Quick Look Pipeline, as well as a new deep-learning tool to identify the source of a transit signal in TESS images. We will also introduce SWIPES (Sliding Window Inference Pipeline for Exoplanet Search), which is a novel deep learning architecture that for the first time uses pixel-level data in the exoplanet search stage. With SWIPES, we hope to better discriminate real transit signals from instrumental phenomena, reducing the false positive rate, and ultimately enabling the detection of previously hidden planets. Algorithms and technology like this will be critical for dealing with the extreme data volumes from TESS, PLATO, and Roman, and will help prepare for direct imaging missions designed to detect biosignatures.