GNN-Based Energy Reconstruction Pipeline for GRAPES-3

For ML practitioners, this work demonstrates how graph-based deep learning can be applied to sparse, irregular detector arrays to improve energy reconstruction accuracy and feature automation, offering transferable techniques for other sensor-network experiments. According to the arXiv entry, the paper "The Deep Learning Cosmic Ray Energy Reconstruction Pipeline for the GRAPES-3 Experiment" was submitted on 8 July 2026 by Sambit Sarkar and one co-author (arXiv:2607.07265). Per the ICRC2025 poster hosted on Indico, the authors implemented a modular, hierarchical dynamic GNN reconstruction pipeline and compared multiple fine-tuning strategies. The arXiv paper and the poster report that the GRAPES-3 array comprises 400 scintillator detectors at 8 m spacing covering 25000 m^2, plus a muon detector of 3712 proportional counters (arXiv; Indico). The authors trained and validated models on logarithmically binned showers for hydrogen, helium, nitrogen, aluminium and iron mass groups and used mutual information and F-statistic feature-selection to improve energy-resolution and mitigate large shower-core-distance effects (arXiv; Indico).