Searching for Exoplanets in Kepler Data with a Convolutional Neural Network and GPU Fast-Folding

The NASA Kepler mission has identified over 4000 new planets and candidates. Its released data still offer new opportunities in developing cutting-edge technology and methods for vetting new exoplanets. We took advantage of this and developed an innovative approach with deep learning and GPU fast-folding techniques to produce some of the fastest, most accurate, and most sensitive exoplanet detections to date. Our search pipeline was developed to look for exoplanet candidates directly from Kepler data by producing and vetting hundreds of thousands of folds per lightcurve. GPU fast-folding greatly reduced computational necessity, and a convolutional neural network trained on artificial transit data was used to efficiently analyze the folds. The model was able to independently recover 99% of 2355 confirmed exoplanets. This methodology successfully identified two previously undiscovered exoplanet candidates. After thorough data analysis, including an independent verification with the Box Least Squares method and transit curve fitting, candidacy was confirmed, and relevant parameters were derived. New discoveries and our design of neural network will be reported.