Identifying Exoplanets Using Deep Learning and Predicting Their Likelihood of Habitability

Abstract

In the last decade, exoplanet study has taken a huge leap in its ability to not only discover exoplanets but also study the potential of habitability on these planets. Much of this credit is due to the launch of the Kepler Space Telescope that was launched on March 7, 2009, by NASA. Humanity’s century-old quest to find extra-terrestrial life was reinvigorated at the launch of Kepler Space Mission. Primary scientific objective of Kepler Space Mission was to explore the structure and diversity of planetary systems. However, this study soon proved to be more challenging than previously expected. Many of these planets were at the mission’s detection sensitivity, and to accurately determine the occurrence rate of these planets, human intervention was required, and classification was a slow process. This paper presents a method to automatically classify transit signals into exoplanets or non-exoplanets based on different methods of signal preprocessing and a trained deep learning model. For signal preprocessing, this paper compares performance metrics of filtering techniques such as Savitsky Golay filters and Gaussian filters. Use of Gaussian filter is coupled with normalization and standardization steps. We then train a convolutional neural network model to predict whether a given signal is an exoplanet or caused by astrophysical or instrumental phenomena. And finally, in order to comment on the habitability of the identified planets, we study the planetary characteristics using Naïve Bayes algorithm.

Keywords