Analyzing Kepler Data Through Machine Learning: A Study of Neural Networks for Exoplanet Detection

Abstract. As modern space observatories generate increasingly large photometric datasets, the need for scalable, automated tools to interpret transit signals has become critical. To address this scale problem, we examined which combinations of lightcurve input representations are most informative and computationally efficient for a 1D convolutional neural network (CNN) to separate real transits from false-positive signals. Across eight systematically compared input configurations, local multiscale transit views achieved the best performance (validation AUC = 0.942, minimum validation loss = 0.30), outperforming global phase-curve inputs in validation AUC and minimum validation loss. Building on that result, we propose applying the trained CNN to unconfirmed KOI candidates to produce an independent, confidence-tiered ranking that augments NASA’s existing candidate disposition-confidence framework. The same approach could later be extended to archival catalogs from the K2 mission (the repurposed continuation of Kepler) and future transit-survey candidate lists, improving the efficiency with which limited telescope time is allocated.