A Machine Learning Approach toward Identifying UHECR Galactic Sources between the Second Knee and the Ankle

Over the past two decades, increasing evidence has suggested that Ultra-High-Energy Cosmic Rays (UHECRs) with energies above 10¹⁸ eV mark a transition from a predominantly Galactic to an extragalactic origin. However, the details of this transition remain poorly understood, as the particle rigidities involved appear too high to be explained by standard supernova shock acceleration models.

In this contribution, we present a Machine Learning approach aimed at identifying the potential locations of the most powerful Galactic accelerators of UHECRs. The analysis is based on an open-source Galactic Atlas constructed by our group, which contains sky images of possible source distributions as observed from Earth. The Atlas was generated by backtracking antiprotons from Earth through several Galactic Magnetic Field (GMF) models at discrete energies between 200 PeV and 10 EeV (see Baiza Mand’s talk).

We explore the use of Convolutional Neural Networks (CNNs) trained on HEALPix maps derived from the Atlas to infer source parameters. We present the CNNs’ performance in predicting source locations under varying signal-to-background ratios, and evaluate their robustness when applied to data simulated with GMF models different from those used during training.