Event vertex reconstruction with deep neural networks for the DarkSide-20k experiment

Abstract

While deep learning techniques are becoming increasingly more popular in high-energy and, since recently, neutrino experiments, they are less confidently used in direct dark matter searches based on dual-phase noble gas TPCs optimized for low-energy signals from particle interactions. In the present study, the application of modern deep learning methods for event vertex reconstruction is demonstrated with an example of the 50-tonne liquid argon DarkSide-20k TPC with 8200 photosensors. The developed methods successfully reconstruct event positions within sub-cm precision and apply to any dual-phase argon or xenon TPC of arbitrary size with any sensor shape and array pattern.

Machine Learning in the Search for New Fundamental Physics

“…This approach is particularly effective at topological signal/background discrimination in Xe time projection chambers (TPCs) for 0νββ [245], including NEXT [246,247], nEXO [248,249], and PandaX-III [250], as well as the KamLAND-Zen 0νββ scintillator experiment [251] and even nuclear emulsion DM searches [252]. 2D detector hit patterns have also been used directly as inputs to CNNs, fully-connected networks, or Bayesian optimization methods [253] for more precise position reconstruction in noble element TPCs, such as EXO-200 [254], XENON1T [255], and the Ar-based DarkSide DM experiments [256,257].…”

“…PICO has observed improved particle discrimination from a simple fully-connected network applied to the first few components of Fourier space data than a CNN in the time domain [242]. EXO-200 sees minimal improvement in discrimination when using a CNN classifier over a BDT with engineered features [238], while DarkSide-20k achieves improved performance using a fully-connected network over a CNN [256]. In contrast, initial training of a convolutional autoencoder followed by training a fully-connected network on its encoded latent space has shown good results in germanium [241], suggesting that separating the tasks of representation and classification may be more robust.…”