Studying rare astronomical events with explainable artificial intelligence

Sammanfattning

This thesis presents artificial intelligence algorithms designed to accelerate the discovery and analysis of rare high-energy astrophysical phenomena. The research consists of two primary contributions. First, this thesis addresses the "needle in the haystack" problem of identifying pulsating ultraluminous X-ray sources (PULXs) within the broader ULX population. By applying an unsupervised Gaussian Mixture Model (GMM) clustering algorithm to an XMM-Newton database of 640 sources, 85 unique candidate PULXs are successfully isolated. These candidates share high-dimensional spectral and temporal similarities with known neutron star-powered systems, despite a current lack of detected pulsations in their archival light curves. The results are then explained using Decision Trees (DTs), which extract the most relevant patterns from the input data. Then, this thesis proposes DeepRed, a deep learning pipeline for redshift estimation. By adapting modern computer vision architectures, including EfficientNet and Swin Transformers, state-of-the-art performance is demonstrated in estimating redshifts from images of galaxies and gravitationally-lensed transients. The reliability of these predictions is further validated through SHapley Additive exPlanations (SHAP), ensuring that the models focus on physically relevant astronomical structures. Together, these contributions are unified by three shared principles: a focus on statistical outliers that are scientifically valuable precisely because of their rarity, the systematic use of Explainable Artificial Intelligence (XAI) to ensure that model outputs are physically interpretable rather than merely accurate, and an architecture designed for scalability to the petabyte-scale surveys of the coming decade. In doing so, they provide robust and complementary frameworks for the automated discovery and characterisation of rare astrophysical phenomena across observational regimes.