“Identifying Exoplanets with Deep Learning: A Five Planet Resonant Chain around Kepler-80 and an Eighth Planet around Kepler-90” was submitted to arXiv on December 13, 2017 by Christopher J. Shallue and Andrew Vanderburg and accepted for publication in The Astronomical Journal.
The paper trains a deep convolutional neural network to predict whether a given periodic dimming signal in Kepler photometry is a genuine transiting exoplanet or a false positive. Trained on a labeled set of previously vetted Kepler signals, the model ranks plausible planets above false positives 98.8 percent of the time on a held-out test set. Applied to known multi-planet systems, it recovered two new validated planets: Kepler-90i, which made Kepler-90 an eight-planet system, and Kepler-80g, the sixth member of a five-planet resonant chain whose orbital period matched a three-body Laplace prediction.
The work became a widely cited demonstration that supervised deep learning can extract real astrophysical detections from survey data after conventional pipelines have already been run, and it seeded a wave of machine-learning vetting and discovery tools across later transit surveys.