Presentation #111.25 in the session “Time Domain Astrophysics (Poster)”.
The era of time-domain astronomy has brought new challenges and opportunity to X-ray observational astronomy. Recent discoveries indicate that fast X-ray transients, X-ray detected extrasolar planet transits and several other exotic events remain hidden in the rich data sets that X-ray telescopes such as Chandra, XMM-Newton, and eROSITA have produced. How to harvest these high energy data sets in order to find the most unique and compelling sources has become a big question in data-driven high energy astronomy, as it has done it already in other wavelength regimes. One relevant question is: can we find those that remain hidden in archival data, and build an automatic alert system that will identify similar transients in new X-ray datasets, starting from the raw arrival times an energies of every single photon detected? Here we present an approach based on DeepSets that is used to predict transient events and other anomalous objects, when Chandra event files are fed into the algorithm. Event files are ordered time series of variable length, which poses a problem for designing deep learning and statistical techniques, especially in the context of anomaly detection. We present a framework that re-casts raw event files into sets of order-independent elements, namely differences in time, energy, and position. DeepSets are machine learning architectures that operate on order-equivariant set data (undirected, non-connected graphs) that can be used to learn global properties, such as class label or variability, of the (potentially) variably-sized sets. This format frees us from order-dependent event files in the context of regression and event file classification for anomaly detection, and offers a way forward for managing massive new Chandra data releases. We apply this method to sources in the Chandra Source Catalog and demonstrate its ability to automatically detect interesting variability patterns and flag anomalies in sets of events.