Detection of Exoplanets using Machine Learning

Abstract

Three methods for using machine learning to decide if a star has an exoplanet from transit survey data are discussed in this paper and are also used to determine which approach performs better on a labeled data set containing light intensity time series from extrasolar stars. Convolutional Neural Network (C.N.N.), a C.N.N. autoencoder, and Support Vector Machines ( S.V.M.) are among the three models. We trained the models using data from the Kepler Space Telescope; because there were very few confirmed exoplanet in the data, some preprocessing is performed before implementing the data set methods. SMOTE (Synthetic Minority Oversampling Technique) was used to over-sample the exoplanet class to further boost the performance. For the supervised learning approaches, CNN and SVM, the outcomes were promising. Therefore, CNN autoencoder, the unsupervised method, did not generate meaningful results.