Abstract

This paper presents a graph neural network (GNN) technique for low-level reconstruction of neutrino interactions in a Liquid Argon Time Projection Chamber (LArTPC). GNNs are still a relatively novel technique, and have shown great promise for similar reconstruction tasks in the LHC. In this paper, a multihead attention message passing network is used to classify the relationship between detector hits by labelling graph edges, determining whether hits were produced by the same underlying particle, and if so, the particle type. The trained model is 84% accurate overall, and performs best on the EM shower and muon track classes. The model’s strengths and weaknesses are discussed, and plans for developing this technique further are summarised.

Highlights

  • The generation of neutrino physics experiments heavily utilises Liquid Argon Time Projection Chamber (LArTPC) detectors, which measure particle interactions with precise spatial resolution

  • The wealth of rich information provided by a LArTPC means automated particle reconstruction can prove challenging, and in recent years machine learning techniques have been increasingly adopted to meet this challenge due to their ability to outperform traditional methods [1]

  • In the context of the LHC, the Exa.TrkX collaboration has demonstrated that graph neural network (GNN)-based methods show great promise for reconstructing detector hits into particle tracks [2]

Read more

Summary

Introduction

The generation of neutrino physics experiments heavily utilises LArTPC detectors, which measure particle interactions with precise spatial resolution. In the context of the LHC, the Exa.TrkX collaboration has demonstrated that GNN-based methods show great promise for reconstructing detector hits into particle tracks [2]. Such methods have the advantage of operating on any data structure which can be described by quantised nodes and their relationships, and on heterogeneous node definitions, which makes them much more broadly applicable than CNNs, which require input tensors arranged in a regular grid structure. Since neutrino interactions typically have high sparsity, recent studies have shown considerable efficiency improvements by moving to sparse CNNs [6], but even in these techniques, native detector outputs must be transformed through voxelization in order to produce CNN-compatible inputs. If efficient GNN reconstruction techniques can be developed, they will be able to operate on detector output without any form of downsampling

Graph construction
Label construction
Training results
Findings
Discussion
Full Text
Paper version not known

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call