Mining TESS Data for Anomalous Light Curves

In the context of the exploration-driven approach to astronomical discovery, which attempts to unveil unknown unknowns by finding new ways to dissect the vast, multi-wavelength, time-dependent space of large astronomical catalogs, machine learning has recently acquired a central role. One relevant question is: how do we use machine learning to discover the unexpected when we are presented with a large dataset, i.e., how do we make serendipitous discoveries systematic? In this contribution we discuss state-of-the-art outlier detection methods that use machine learning to detect anomalies in astronomical datasets. To demonstrate the applicability of these methods in cutting-edge astronomical research, we apply them to the publicly available TESS data to find the weirdest variability patterns in stellar light curves, and link this weirdness to specific stellar evolution stages. Upon calibrating the methods using "ground truth" anomalous objects such as KIC 8462852 (Tabby's star), we characterize the concept of weirdness in terms of the physics of variability (intrinsic or extrinsic) of these objects. We do this by combining our results with Gaia photometry and astrometry that allow us locate anomalous objects in a Hertzprung-Russell diagram. We present a list of anomalous light curves that include objects in rare evolutionary stages, such as cataclysmic variables and irregular transits due to, e.g. exocomets. Our results also enable an increase in the census of known but rare variability types, such as eclipsing multiple systems and long-term pulsating stars. We further discuss the modular nature and adaptability of these methods to other time domain surveys, such as the Large Synoptic Survey Telescope (LSST), and provide recommendations in terms of feature engineering for time-domain astronomy.