Speaker
Description
We present a graph neural network-based approach to calorimeter shower reconstruction in liquid xenon, under development for the PIONEER experiment's LXe calorimeter option. The method employs a GravNet backbone operating on a graph of active sensors as nodes, with node features encoding spatial position, sensor type, and a compact temporal representation of photon arrival times via a learned basis decomposition across the characteristic LXe emission timescales: Cherenkov, prompt singlet, and triplet scintillation light. This temporal encoding is motivated by the need to reduce per-event input dimensionality while preserving the physically discriminating timing structure relevant to pile-up rejection and particle identification. The Object Condensation framework is adopted to handle variable-multiplicity shower reconstruction with per-cluster energy, position, and timing estimates. We present the motivation and architecture design, and preliminary results towards improved energy resolution, position reconstruction, and separation of overlapping showers in the high-rate pion decay environment with the LXe calorimeter option.