Direct Vertex Reconstruction of $\Lambda$ Baryons from Hits in CLAS12 using Graph Neural Networks

TL;DR

This paper demonstrates that Graph Neural Networks can directly reconstruct $ ext{Lambda}$ hyperon decay vertices from detector hits, outperforming traditional track-based methods in simulations, offering a promising new approach for particle physics data analysis.

Contribution

It introduces a novel GNN-based method for direct vertex reconstruction from hits, bypassing traditional track fitting in complex magnetic fields.

Findings

GNN improves vertex reconstruction accuracy in simulation.

Potential to replace track fitting with direct hit-to-vertex neural network mapping.

Highlights the feasibility of GNNs for complex particle tracking tasks.

Abstract

Machine learning techniques, including Graph Neural Networks (GNNs), have been used extensively for data analysis in high energy and nuclear physics. Here we report on the use of a GNN to reconstruct decay vertices of $\Lambda$ hyperons directly from hits in the tracking detector at the CLAS12 experiment at Jefferson Laboratory (JLab). We show that we can improve the vertex reconstruction in simulation compared to the standard, track based, algorithm. We believe this warrants further study. The current study is limited by available training resources but points to an interesting possibility to forgo vertex reconstruction by track fitting in a complicated magnetic field for a more direct approach where the hit to vertex mapping is encoded in a neural network.