Presentation #402.07 in the session “Mining TESS Data with Machine Learning and Other Advanced Methods”.
The full-frame images produced by the Transiting Exoplanet Survey Satellite (TESS) are invaluable tools to detect periodic and transient astrophysical events, such as variable stars, supernovae, tidal disruption events, and gamma-ray burst afterglows.
Our pixel-level method separates the time-series FFIs into smaller groups and then calculates the per-pixel variance over the duration of each group. These variance maps are then combined across adjacent groups to produce a “variability” image in which systematics, like those caused by scattered light and differential velocity aberration, are mitigated. All point sources in the variability image are extracted and classified using a convolutional neural network (CNN) that has been trained on known transients from TESS’s Prime Mission. Our CNN has 90% accuracy on test data, which will increase as newly-identified transients are added to the training set.
Sources with the highest probabilities (assigned through classification) are filtered by matching with catalogs like SIMBAD and TNS, and photometry is done on the set of filtered sources. The generated light curves are then analyzed using a variety of statistical techniques to prioritize those that are most important for follow-up. We are thus able to filter the tens to hundreds of thousands of point sources in each variability image to a manageable level (100s per sector) that can then be visually inspected prior to follow-up. Our technique has also shed light on poorly-characterized detector systematics that must be mitigated to increase our pipeline’s SNR.
Future work will involve exploiting the novel 10-minute FFIs to evaluate transient light curves in finer detail. Additionally, we hope to identify TESS-band counterparts to high-energy transient events, such as compact object mergers and FRBs, that have large error ellipses.