GNN Semantic Segmentation of neutrino interactions in DUNE's TMS

The Deep Underground Neutrino Experiment (DUNE) seeks to measure the ∆m²₃₂, sin²θ₃₂, and δ_CP neutrino oscillation parameters to world-leading precision. In order to complete a successful three-flavor neutrino oscillation analysis, DUNE’s Near Detector (ND) requires sensitivity to neutrinos with E_v > 1.5 GeV, whose interactions produce muons that range out of the primary Near Detector Liquid Argon Time Projection Chamber (ND-LAr TPC). The muon spectrometer (TMS) is designed to reconstruct the energy of escaping muons by their range in the detector, provided it can identify individual muon track instances in the high-pileup data stream. This contribution details a novel reconstruction pipeline developed to extract track instances in the TMS. The framework employs Kernel Density Estimate (KDE) time segmentation and DBSCAN spatial clustering, as well as a Graph Neural Network (GNN) for candidate track semantic segmentation. The pipeline performed well when tested against realistic Monte Carlo (MC) simulations of neutrino interactions in ND – serving as a promising proof of concept for deep-learning aided track reconstruction in the TMS.