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

Abstract: Knowledge of the mass composition of ultra-high-energy cosmic rays (UHECR) is crucial to understanding their origins; however, current approaches have limited event-by-event resolution due to high intrinsic fluctuations of variables like $X_{\mathrm{max}}$ and $N_\mu$. With fluorescence telescope measurements of $X_{\mathrm{max}}$ in particular, there are opportunities to improve this situation by leveraging more information from the longitudinal shower development profile beyond just the depth at which its maximum occurs. Although there have been ML studies on extracting composition or mass groups from surface signals, parametrized profiles, or from derived parameters (e.g. $X_{\mathrm{max}}$ distributions), to our knowledge, we present the first study of a deep-learning neural-network approach to directly predict a primary's mass ($\ln A$) from the full longitudinal energy-deposit profile of simulated extensive air showers. We train and validate our model on a large suite of 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 $\ln A$ on unseen test showers with a resolution ranging between 1.5 for protons and 1 for iron over a large range of noise conditions, corresponding to a proton-iron merit factor of 2.19, outperforming the predictive power of either $X_{\mathrm{max}}$ or $N_\mu$ alone. This performance is only mildly degraded when predictions are made on simulations using the Sibyll-2.3d hadronic interaction model, which was not used in training, showing these features are robust against model choice. Our results suggest that the full shower profile may contain latent composition-sensitive features which have as much discriminating power as $X_{\mathrm{max}}$.