GRAPES-3 Applies ML to Classify Muon and Hadron Tracks
arXiv:2607.07263, submitted on July 8, 2026, presents a GRAPES-3 machine-learning pipeline for distinguishing secondary muon and punch-through hadron tracks. The conference paper uses CORSIKA proton-shower simulations in the 100-158 TeV range with a Geant4 detector simulation, then compares decision trees, random forests, neural networks, and XGBoost. The authors report 88.7% accuracy for XGBoost on single-particle classification and describe graph neural networks with edge convolution for multiparticle events. For practitioners, it is a useful instrumentation pattern for combining classical classifiers with graph-based event modeling.
The practical value of this GRAPES-3 paper is the pipeline pattern: use simulated detector data for labels, keep strong classical baselines for local classification, then move to graph models when the event structure becomes spatially dense and multiparticle.
What happened
The arXiv record arXiv:2607.07263, submitted July 8, 2026, describes a machine-learning pipeline for identifying tracks of muons and hadrons at the GRAPES-3 muon telescope. The work was presented at the 39th International Cosmic Ray Conference and has the journal reference PoS(ICRC2025)410.
Technical context
According to the abstract, the team used CORSIKA-simulated proton showers in the 100-158 TeV range as input to a Geant4 detector simulation. For single-particle classification, it compared decision trees, random forests, neural networks, and XGBoost, with XGBoost achieving the highest reported accuracy of 88.7%. For multiparticle events, the authors represented detector hits as graph nodes and used graph neural networks with edge convolution, followed by a Dynamic Reduction Network regression model to estimate particle counts.
For practitioners
The story is useful for instrument-data teams because it combines classic ML, simulation-based labels, and graph learning in one detector workflow. The important caveat is transfer: simulation-trained models still need validation against real operating data and detector drift.
What to watch
Watch for real-run validation, uncertainty estimates, and comparison against the older GRAPES-3 muon-hadron separation work. Those steps will determine whether the method is operationally useful beyond the conference paper.
Key Points
- 1Combining per-hit classifiers and graph neural networks is practical for disambiguating overlapping signals in dense detector arrays.
- 2Using CORSIKA and Geant4 enables supervised training where real labelled data are scarce, but transfer remains critical.
- 3Classical ensembles like XGBoost remain strong baselines before heavier graph models are used for event-level inference.
Scoring Rationale
This is a solid applied-ML instrumentation paper with clear relevance for detector data and graph-based event modeling. The impact is narrower than general AI research because it is a domain-specific conference paper and still needs operational validation.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems