Using Graph Neural Networks for Multi-Detector Particle ID on DUNE’s Near Detector 2x2 Prototype

The main goal of DUNE (Deep Underground Neutrino Experiment) is precise neutrino oscillation measurements, which relies on its near detectors’ capability to accurately distinguish neutrino flavors in the world’s most intense neutrino beam. Thus, we are leveraging both new technologies and machine learning reconstructions to accomplish this. The Near Detector Liquid Argon (ND-LAr) detector uses a modularized, short-drift Time Projection Chamber structure with 2D pixel readouts leading to high resolution, native 3D images of particle interactions. The machine learning methods take in this information at a pixel level, then apply a series of neural networks to identify which particles were in the detector. But this machine learning is only currently applied on the LArTPC volume, ignoring information from the supplementary muon tagging detectors at reconstruction level. In the near detector 2x2 prototype, these muon taggers are especially vital to distinguish particle signatures that leave the smaller LArTPC testing volume. This work discusses the using inputs from both detector types (LArTPC and muon tagger) to feed into a graph neural network to improve the particle identification on the 2x2 prototype.