Application of Geometric Deep Learning for Tracking of Hyperons in a Straw Tube Detector

Abstract: We present track reconstruction algorithms based on deep learning, tailored to overcome specific central challenges in the field of hadron physics. Two approaches are used: (i) deep learning (DL) model known as fully-connected neural networks (FCNs), and (ii) a geometric deep learning (GDL) model known as graph neural networks (GNNs). The models have been implemented to reconstruct signals in a non-Euclidean detector geometry of the future antiproton experiment PANDA. In particular, the GDL model shows promising results for cases where other, more conventional track-finders fall short: (i) tracks from low-momentum particles that frequently occur in hadron physics experiments and (ii) tracks from long-lived particles such as hyperons, hence originating far from the beam-target interaction point. Benchmark studies using Monte Carlo simulated data from PANDA yield an average technical reconstruction efficiency of 92.6% for high-multiplicity muon events, and 97.1% for the $\Lambda$ daughter particles in the reaction $\bar{p}p \to \bar{\Lambda}\Lambda \to \bar{p}\pi^+ p\pi^-$. Furthermore, the technical tracking efficiency is found to be larger than 70% even for particles with transverse momenta $p_T$ below 100 MeV/c. For the long-lived $\Lambda$ hyperons, the track reconstruction efficiency is fairly independent of the distance between the beam-target interaction point and the $\Lambda$ decay vertex. This underlines the potential of machine-learning-based tracking, also for experiments at low- and intermediate-beam energies.