PILArNet

PILArNet is an open benchmark simulation dataset designed to support machine learning and reconstruction algorithm development for liquid argon time projection chambers (LArTPCs). It is not associated with a single running experiment but rather provides a clean, model-independent simulation environment usable across the LArTPC community.

The dataset contains 300,000 simulated particle gun events generated with model-independent particle generators (MPV for single particles, MPR for multi-particle events) and full Geant4 detector simulation. Three detector volumes are provided — 192³, 512³, and 768³ voxels — to support studies at different detector scales. Energy depositions are stored in sparse 2D/3D voxel representations using the LArCV data format, with two complementary data products: energy-3D (energy deposition per voxel) and segment-3D (ground-truth particle category label per voxel). A simplified detector response model is applied (no full electronics simulation; spatial smearing only), making the dataset well-controlled for algorithm benchmarking.

Five particle categories are covered: protons, minimum-ionizing particles (MIPs: muons and charged pions), electromagnetic particles (EM: electrons and photons), Michel electrons, and delta rays. Particle multiplicities and kinetic energies are sampled within configurable ranges (50–1000 MeV for most categories, 50–400 MeV for protons).

PILArNet supports a range of core reconstruction tasks: semantic segmentation (per-voxel particle type classification), instance segmentation (grouping voxels into individual particles), kinematic regression (particle energy and direction), vertex reconstruction, and 2D-to-3D reconstruction. Its ground-truth labels and clean simulation make it particularly useful for training and benchmarking supervised deep learning models before applying them to real detector data.