Graph Neural Networks for Photon Searches with the Underground Muon Detector of the Pierre Auger Observatory

TL;DR

This paper introduces a deep learning-based method using graph neural networks to distinguish ultra-high-energy photons from hadronic cosmic rays by analyzing data from the Pierre Auger Observatory's detectors, aiming to improve indirect photon detection.

Contribution

The paper presents a novel graph attention network approach for photon-hadron discrimination utilizing combined surface and underground muon detector data, enhancing sensitivity to diffuse photon fluxes.

Findings

Effective discrimination between photon and hadron-induced air showers.

Potential to improve detection sensitivity for ultra-high-energy photons.

Framework adaptable to increasing data volumes in future observations.

Abstract

Ultra-high-energy photons have long been sought as tracers of the most energetic processes in the Universe. Several sources can contribute to a diffuse photon flux, including interactions of cosmic rays with Galactic matter and radiation fields, as well as more exotic scenarios such as the decay of super-heavy dark matter. Regardless of their origin, the expected flux is extremely low, making direct detection impractical and thereby requiring indirect detection by extensive ground-based detector arrays. In this contribution, we present a novel method for photon-hadron discrimination in the energy range of $50$ to $300\,\text{PeV}$ based on deep learning algorithms. Our approach relies on information from both the Surface Detector (SD) and the Underground Muon Detector (UMD) of the Pierre Auger Observatory. The SD consists of an array of water-Cherenkov detectors. It is used to measure the electromagnetic and muonic components of extensive air showers at ground level. Meanwhile, the UMD is composed of buried scintillator modules. It is sensitive to air-shower muons with energies above ${\sim}1\,\text{GeV}$, enhancing the identification of muon-poor air showers as initiated by photon primaries. Our method represents air-shower events as graphs, and consequently, the network architecture is composed of graph attention layers. We assess the performance of the method on a data subset and discuss the implications of unblinding the full current dataset, as well as the prospects of the increasing data volume expected in the coming years, particularly in terms of sensitivity to various diffuse fluxes from theoretical predictions.