Astroparticle physics, KM3NeT experiment. Muon reconstruction in the KM3NeT experiment using neural network models with transformer architecture

Author: Davit Janezashvili
Keywords: KM3NeT, Transformers, Machine Lerning, Neutrino
Annotation:

KM3NeT (Cubic Kilometre Neutrino Telescope) is a new research infrastructure that includes the next generation of neutrino telescopes. The telescope is located in the depths of the Mediterranean Sea and consists of two telescopes: KM3NeT/ARCA and KM3NeT/ORCA. Upon completion of installation, the telescope will include 12,000 optical detectors distributed across 345 vertical strings (230 in the ARCA detector and 115 in the ORCA detector). Cherenkov light emitted by charged particles propagating in water is detected by photomultipliers in the DOMs (Digital Optical Modules). KM3NeT/ORCA aims to study atmospheric neutrino oscillations through the Earth to determine the neutrino mass ordering and provide precise measurements of oscillation parameters. These measurements depend on the accurate reconstruction of neutrino events in the telescope. Current deep learning models used in KM3NeT are graph neural networks. This work presents a new approach to deep learning models using transformer architecture. The model is trained on simulations conducted for the KM3NeT/ORCA detector, prepared in a special dataset for comparative analysis. It is evaluated for various event reconstruction tasks in KM3NeT, including background rejection, event topology classification, and energy and direction reconstruction. Comparisons with other reconstruction methods, practical aspects of building this architecture, and its future development will also be discussed.