Machine learning pipeline for identifying tracks of muons and hadrons at GRAPES-3 muon telescope

Abstract

The GRAPES-3 experiment is a ground-based extensive air shower array which consists of approximately $400$ closely packed plastic scintillator detectors and a large area muon telescope. Estimating the number of associated muons created in an air shower is crucial to understand the properties of primary cosmic rays. The GRAPES-3 muon telescope (G3MT) records these secondary muons, however, the punch-through hadrons can introduce background noise. This study aims to develop a machine learning pipeline to distinguish the tracks of secondary muons and hadrons at G3MT. We have used CORSIKA-simulated proton showers having energy in the range 100-158 TeV as an input for a Geant4-based detector simulation to analyze the signatures of both type of particles. Initially, single-particle classification was performed using decision trees, random forests, neural networks, and XGBoost, with XGBoost achieving the highest accuracy of 88.7\%. For multiparticle classification, we modelled Graph Neural Networks (GNNs) where each event was represented as a graph with detector hits as nodes. A GNN with edge convolution layers was developed to classify each node as a muon or hadron hit. Following this, a deep learning regression model using Dynamic Reduction Network was developed to estimate the number of particles and muons striking G3MT simultaneously. Details of the analysis and results of the multiparticle classification task will be presented.