Machine Learning Models for z-Reconstruction in DarkSide-20k Experiment

Abstract

DarkSide-20k will be a dual-phase Liquid Argon (LAr) Time Projection Chamber (TPC) with

50 t (total mass of UAr) for direct WIMP search under ultra-low background conditions. It provides access to both scintillation (S1) and electron (S2) signals produced by particle interaction in the LAr target. The DarkSide-20k experiment will be the most advanced and largest liquid argon TPC ever built for dark matter searches. Due to the inherently slow drift of ionization electrons in such detectors, overlapping events (pile-up) can occur when multiple interactions happen within the same drift time window, posing significant challenges for event

reconstruction. When multiple events occur close in time, their S2 signals may overlap, making it difficult to identify which S2 corresponds to which S1. By estimating the z-coordinate directly from the S1 light pattern, it becomes possible to predict when the associated S2 should appear, thus enabling correct matching and reducing the impact of pile-up — namely, the risk of rejecting valid events or introducing dead time in the detector readout. This work explores deep network-based algorithms for position reconstruction in this detector, improving models by learning features. Two different approaches are used: Multilayer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs) to incorporate the spatial distribution of sensors. The best result achieved so far is a 32 cm error in the z-coordinate. These are initial steps toward position reconstruction in pile-up scenarios. Ongoing improvements and comparisons with other methods are underway.