GRAPES-3 Applies ML to Classify Muon and Hadron Tracks

ML pipelines that combine per-hit classifiers with graph-based event models can improve particle identification in dense detector arrays, which matters for label-scarce, high-throughput instrumentation data. Per the arXiv abstract (2607.07263) and the PoS(ICRC2025) conference paper, researchers from the GRAPES-3 collaboration (Tata Institute of Fundamental Research) present a machine learning pipeline to distinguish secondary muon and punch-through hadron tracks in the GRAPES-3 muon telescope. The study uses CORSIKA proton-shower simulations (energy 100-158 TeV) fed into a Geant4 detector simulation. Single-particle classifiers tested include decision trees, random forests, neural networks, and XGBoost, with XGBoost achieving the highest reported accuracy of 88.7%, per the authors. For multiparticle events the team models events as graphs and trains a Graph Neural Network with edge convolution layers for per-hit classification, followed by a regression using Dynamic Reduction Network to estimate particle counts, as described in the PoS paper.