New avenues for multi-messenger discoveries in high energy astrophysics using machine learning

We have entered a new era in observational X-ray astronomy: time-dependent flux variability of astronomical sources is revealing the nature of new, exotic types of explosions, such as pair-instability supernovae and neutron star mergers, lensing events, exotic planetary transits, as well as other, yet to be understood phenomena. These phenomena are typically observed over a broad range of wavelengths. Yet, while time-domain surveys at optical wavelengths have taken the lead to profit from vast observational programs (e.g. Vera Rubin), and from an even vaster set of Data Science tools to pick up the exotic sources among millions of regular objects, the X-ray community is only now starting to delve into the synergies between data science and large data astronomical datasets for discovery. In this contribution we attempt to provide new avenues of inquiry in observational X-ray astronomy, by providing a framework for the treatment of large X-ray datasets that enables the discovery of relevant high-energy transients. We present novel representations for the X-ray events that allow for the seamless discovery of transients and introduce methods for regression and anomaly detection in large astronomical datasets, including graph neural networks and embedded representations that account for the varying length of the event files. We apply these methods to the Chandra Source Catalog, and present an application in the identification of fast X-ray transients (which might signal neutron star mergers), and a prototype for an automatic alert system for new X-ray transients. We argue that novel data representations are needed in X-ray datasets prior to the application of machine learning techniques in order to maximize scientific discovery