Presentation #402.04 in the session “Mining TESS Data with Machine Learning and Other Advanced Methods”.
Our understanding of variability of extragalactic sources at ~1 day time scale is incomplete, but there is a large discovery potential for the Transiting Exoplanet Survey Satellite (TESS) regarding supernovae and bright active galaxies. Using a convolutional autoencoder and engineered features, we construct a representational feature space of TESS light curves of extragalactic sources for use in unsupervised learning for classification of light curves and searching for anomalous samples. We compile a TESS catalog of extragalactic variability, including early-rise supernovae discovered by the blind pointing of TESS, which is informative about the progenitors of Type Ia supernovae. We validate our methodology by injecting known transient light curves from the Transient Name Server (TNS) into our sample and recovering them with a background rejection of 99% using the Local Outlier Factor algorithm for anomaly detection. Our uniform census of the extragalactic sky using TESS enhances the coverage of ground-based surveys.