Detection of Stellar Variability and Anomalies in Sky Surveys Aided by Machine Learning

As past and current sky surveys like CRTS, Euclid, Gaia, Pan-STARRs, TESS, ZTF to name a few, have provided immense datasets with a wide range in parameters like cadence, depth, and resolution, machine learning has become essential and enables systematic, population-level studies of stellar variability, as well as the targeted discovery of rare anomalies that push the horizons of our understanding. As Rubin/LSST will vastly expand nightly data volumes at greater depths, and NASA’s Roman Space Telescope will provide deep and high resolution IR data in survey mode, we will have access to populations and anomalies with few spectroscopic follow-up possibilities. This makes machine learning indispensable for classification through transfer learning. We argue to push using machine learning on both these fronts: assembling uniform samples of well-characterized variable stars to probe their statistical distributions, and, identify atypical sources that lie at the edge of population censuses, and we outline how real-time and archival astronomy analyses need to work hand-in-hand to make this possible.