Point proposal network for reconstructing 3D particle endpoints with subpixel precision in liquid argon time projection chambers

Liquid Argon Time Projection Chambers (LArTPC) are particle imaging detectors recording 2D or 3D images of trajectories charged particles. Identifying points interest in these images, namely the initial and terminal track-like such as muons protons, electromagnetic shower-like electrons gamma rays, is a crucial step identifying analyzing particles impacts inference physics signals neutrino interaction. The Point Proposal Network designed to discover specific interest. algorithm predicts with sub-voxel precision their spatial location, also determines category identified Using benchmark PILArNet public LArTPC data sample which voxel resolution 3mm/voxel, our successfully predicted 96.8% 97.8% within distance 3 10~voxels from provided true point locations respectively. For voxels closest locations, median found be 0.25 voxels, achieving level precision. In addition, we report analysis mistakes where prediction differs positions by more than 10~voxels. Among 50 visually scanned, 25 were due definition position 15 legitimate physicist cannot disagree algorithm's prediction, 10 genuine that wish improve future. Further, using points, demonstrate simple cluster into individual clustering efficiency, purity, Adjusted Rand Index 96%, 93%, 91%