A Spatiotemporal Data Cube Approach to Classification of Variable Stars: A Catalog of Candidate Variable Stars from the TESS Full-frame Image Raw Data

Abstract We present a novel technique to identify variable stars in large surveys that performs classification not on the 1D light curves but on the spatiotemporal data cube. Such an approach can allow the model to better discriminate against the systematics in the background flux and consider the variability across the entire point-spread function, which can be advantageous to large surveys such as Transiting Exoplanet Survey Satellite (TESS). We perform a search for eclipsing, pulsating, and rotating variables in TESS full-frame images. This classification was done using a neural network that has been trained on variable stars identified in other surveys. In total we identify 3.1 million candidate variable sources. Although these sources should undergo significant vetting to confirm them using more traditional techniques, through assembling the catalog of the most promising candidates, it should be possible to significantly narrow down the list of light curves of sources that would need to be constructed for such an analysis. With an increasing number of surveys aiming to do photometric monitoring coming online in the future (e.g., LSST), a similar analysis with an improved training set could be advantageous in classifying variable stars.