Point Cloud Deep Learning Methods for Particle Shower Reconstruction in the DHCAL

TL;DR

This paper investigates the use of graph neural networks and attention mechanisms to improve particle shower reconstruction in digital calorimeters, addressing challenges like shower shape variability and detector granularity.

Contribution

It introduces a novel application of GNNs, GATs, and DeepSets for calorimeter data analysis, enhancing particle identification and energy resolution over traditional methods.

Findings

GNN-based methods outperform baseline techniques in particle ID and energy measurement.

Performance depends on particle incident angle and detector granularity.

Implementation challenges include shower shape diversity and hyper-parameter tuning.

Abstract

Precision measurement of hadronic final states presents complex experimental challenges. The study explores the concept of a gaseous Digital Hadronic Calorimeter (DHCAL) and discusses the potential benefits of employing Graph Neural Network (GNN) methods for future collider experiments. In particular, we use GNN to describe calorimeter clusters as point clouds or a collection of data points representing a three-dimensional object in space. Combined with Graph Attention Transformers (GATs) and DeepSets algorithms, this results in an improvement over existing baseline techniques for particle identification and energy resolution. We discuss the challenges encountered in implementing GNN methods for energy measurement in digital calorimeters, e.g., the large variety of hadronic shower shapes and the hyper-parameter optimization. We also discuss the dependency of the measured performance on the angle of the incoming particle and on the detector granularity. Finally, we highlight potential future directions and applications of these techniques.