Prospects for Deep-Learning-Based Mass Reconstruction of Ultra-High-Energy Cosmic Rays using Simulated Air-Shower Profiles

TL;DR

This study demonstrates that deep learning models can accurately predict the mass composition of ultra-high-energy cosmic rays from simulated air-shower profiles, outperforming traditional methods and showing robustness across models and noise conditions.

Contribution

First application of deep learning to directly estimate cosmic ray mass from longitudinal shower profiles, surpassing previous shape-based ML approaches.

Findings

Deep learning achieves bias < 0.4 in lnA and resolution ~1.5 for protons and ~1 for iron.

ML models outperform existing benchmarks using profile-shape parameters.

CNN maintains performance across different simulation models and noise levels.

Abstract

Knowledge of the mass composition of ultra-high-energy cosmic rays is crucial to understanding their origins; however, current approaches have limited event-by-event resolution. With fluorescence telescope measurements of the longitudinal shower profile, there are opportunities to improve this situation by applying Machine Learning (ML) to leverage more information beyond $X_{max}$ alone. To our knowledge, we present the first study of a deep-learning neural-network approach to predict a primary's mass ($\ln{A}$) directly from the longitudinal energy-deposit profile of simulated extensive air showers. We train and validate our model on simulated showers, generated with CONEX and EPOS-LHC, covering nuclei from $A = 1$ to 61, sampled uniformly in $\ln{A}$. After rescaling, our network achieves a maximum bias better than 0.4 in $\ln{A}$ with a resolution between 1.5 for protons and 1 for iron, corresponding to a proton-iron Merit Factor of 2.19 (AUC = 0.976). We benchmark this against simpler ML models trained on profile-shape parameters ($X_{ max}$, $E_{cal}$, $R$, and $L$) extracted from the same data. We find that even simple models can substantially exceed published benchmarks for combinations of these observables, demonstrating that ML methods applied even to standard profile-shape parameters can significantly improve available mass sensitivity. The CNN outperforms this strong baseline, and this performance is only mildly degraded when cross-predicting on simulations made with the Sibyll-2.3d hadronic interaction model, showing robustness against model choice. The network also maintains its performance across a wide range of noise conditions. An ablation study further demonstrates that the full profile contains composition-sensitive structure not captured by the GH parameterization, while the strong performance of the CNN suggests this information should be resolvable in real events.