Machine Learning for the Categorization and Discovery of TESS FFI Light Curves

Adam Friedman, Brian P. Powell

Project Description

This project utilized machine learning to classify Full Frame Image (FFI) light curves from the Transiting Exoplanet Survey Satellite (TESS) based on light curve morphology. In order to effectively classify light curves, a substantial amount of training data is required. An autoencoder was used to force self-organization of similar light curves in a latent representation, which constituted the basis for several classes such as sinusoidal variability, oscillation, and eclipsing binaries. Through a process of iterative training data collection and model strengthening, enough training data was acquired to reliably extract examples of each category of light curve from batches of test data. The results of the project allowed not only the identification of thousands of light curves from previous TESS sectors, but also from the newest release of sector 25 within a week of publication. Furthermore, the classification of eclipsing binaries resulted in the discovery of 6 new quadruple star system candidates (pending publication) as well as a promising candidate for a circumbinary planet (awaiting further data to confirm). Overall, this project has demonstrated the capability of machine learning immediately process and distribute relevant light curves to astronomers that they were previously unable to obtain, let alone analyze. The neural network created as a part of this project will continue to be improved and used for the duration of the TESS mission.