The Search for Life: Exoplanet Detection with Deep Learning

Abstract

The discovery of new exoplanets, planets outside of our solar system, is essential for increasing our understanding of the universe. Exoplanets capable of harboring life are particularly of interest. Over 600 GB of data was collected by the Kepler Space Telescope, and about 30 GB is being collected each day by the Transiting Exoplanet Survey Satellite since its launch in 2018. Traditional methods of experts examining this data manually are no longer tractable; automation is necessary to accomplish the task of vetting all of this data to identify planet candidates from astrophysical false positives. Previous state-of-the-art models, Astronet and Exonet, use deep convolutional neural networks (CNNs) with over 8.8 million parameters. In this paper, I experiment with the application of recurrent networks, attentional models, and scaling down Astronet. I have developed a CNN model with 8x fewer trainable parameters than Astronet with the same accuracy and improved precision. I also provide a CNN-LSTM model with 59x fewer parameters just 1\% behind Astronet in accuracy that, with further tuning, may also be a competitive model for particularly resource-constrained uses. All code for this research is available on GitHub.