Using Deep Neural Networks to Identify Exoplanets

Overview

In 2018, NASA initiated the Transiting Exoplanet Survey Satellite (TESS) to search for planets outside of our solar system. TESS has been gathering large sets of star light data, but the quantity of data is too large for scientists to analyze. Recently, machine learning has been applied to the TESS star light data sets. One example is the Nigraha Pipeline, an open-source convolutional neural network which is designed to identify unknown exoplanets from the light of stars. The complexity of the Nigraha Pipeline requires a large training data set and long run times. This work will utilize Deep Feed Forward (DFF) Neural Networks to identify potential exoplanets using a smaller training data set and maintaining high accuracy. Multiple DFF networks were created using varying number of hidden layers and activation functions. Using only one-third of the training data used in the Nigraha Pipeline, accuracies approaching 90% were achieved as compared to 88% for the Nigraha Pipeline. In addition, a significant decrease in computation time was observed with the DFF network taking minutes as compared to the Nigraha Pipeline taking hours. This achievement will allow for more exoplanet data to be accurately classified and progression for a deeper understanding of surrounding solar systems.