Exploring DHCAL design and performance with Graph Neural Networks

TL;DR

This paper investigates the use of Graph Neural Networks for hadron energy reconstruction and particle identification in a Digital Hadronic Calorimeter, demonstrating improved performance and cost-effectiveness over traditional methods.

Contribution

It introduces GNN-based methods for DHCAL analysis, showing enhanced particle identification and energy resolution, even with coarser detector granularity.

Findings

Achieved over 50% classification efficiency for neutrons and pions.

Protons reached 77% efficiency, highest among studied particles.

Energy resolution improved across 1-50 GeV range, with robustness to detector granularity.

Abstract

In the context of a gas-sampling Digital Hadronic Calorimeter (DHCAL), we explore the potential of using Graph Neural Networks (GNN) for hadron energy reconstruction and Particle Identification (PID) in future collider experiments. For PID, we achieved classification efficiencies exceeding 50% for neutrons and pions, with notably higher efficiencies for kaons and protons. Protons exhibited the highest efficiency of 77%, followed by neutral kaons. The energy resolution for these hadrons is studied in the energy range of 1 -- 50 GeV, with a further investigation into the resolution as a function of the incoming particle's angle and readout granularity, focusing on charged pions. Compared to traditional analysis methods, our results indicate that improved performance can be achieved even with coarser detector granularity, potentially making future DHCAL systems more cost-effective.