Reconstructing ultra-high-energy particle cascades with deep learning methods

Ultra-high-energy cosmic rays (UHECRs) are atomic nuclei from distant galaxies, representing the most energetic phenomena in the universe. Understanding UHECRs can provide insights into particle physics, cosmology, and astrophysics, as they result from extreme cosmic events like AGN and GRBs. The core observables of UHECRs are energy, the location in the sky they arrived from, and critically the mass composition of the arriving particles.

The current leading method for determining cosmic ray mass relies on Xmax, the point in the atmosphere where energy deposition peaks. However, since UHECR showers can be reconstructed from indirect measurements, we propose that deep learning methods can extract more mass-related information from shower profiles.

We have simulated 2 million cosmic ray showers for training convolutional neural networks (CNNs) and transformer-based models. These models demonstrate exceptional performance on noise-free data, outperforming traditional methods by a factor of 7. Ongoing benchmarking assesses their robustness against noisy data, where CNNs still outperform traditional techniques. This study explores effective network architectures, training processes, and benchmarking, with future work focusing on the network's interpretability and its physics insights.