Unsupervised sifting of XMM-Newton EPIC observations using Variational Autoencoders

Finding specific events of interest within the XMM-Newton Science Archive (XSA) is a significant challenge due to the sheer volume of data, which contains over 300,000 source light curves. While the mission provides standardized multi-band products, the vast majority of these time series remain unexplored, and it is often unknown a priori whether they contain intrinsic variability, transients, or flares. Traditional screening is further complicated by non-stationary instrumental backgrounds that often overlap with physical signals. This paper presents a novel unsupervised framework based on Variational Autoencoders (VAEs) to automatically sift through the XSA. Unlike previous methods that rely on hand-crafted statistical features or multi-epoch historical data, our approach operates directly on (pipeline-processed) contemporaneous multi-band light curves. By learning patterns of variability shared across the five contemporaneous energy-band light curves, the VAE maps high-dimensional observations into a low-dimensional latent space. This allows for the discovery of complex variability patterns and the identification of anomalies based on their intrinsic morphological structure. Ultimately, the proposed framework provides a scalable screening tool for multi-band light curves generated by the EPIC-pn Pipeline Processing System (PPS), ranking candidate anomalies using model-based scores, significantly reducing the manual burden of exploratory searches, and prioritizing potentially interesting sources for follow-up analysis. The validity of the method has been assessed both qualitatively and quantitatively.