Classificação e reconstrução de chuveiros atmosféricos via deep learning estereoscópico multitarefa para os telescópios de grande porte do CTAO

Abstract

The Cherenkov Telescope Array Observatory (CTAO) represents the next generation of groundbased gamma-ray observatories, designed to improve sensitivity by an order of magnitude over current instruments. A major scientific challenge for CTAO lies in the low-energy range (20-150 GeV), covered by the Large-Sized Telescopes (LSTs), where the scarcity of Cherenkov photons and hadronic background contamination limit the traditional reconstructions methods based on Hillas parametrization. This work proposes and validates a deep convolutional siamese neural network designed for full stereoscopic event reconstruction in LSTs. The architecture operates under the Multi-Task Learning paradigm, simultaneously performing gamma-hadron classification and energy and arrival direction regression. A Curriculum Learning strategy was implemented, evolving the neural network from basic morphological learning to complex geometric correlation. Using Monte Carlo simulations (CORSIKA+sim_telarray), we demonstrate that stereoscopic fusion combined with attention mechanisms permit achieving, in a total stereoscopic operation, an energy resolution under 20% and an angular resolution 𝜃68 of 0,2°. The gamma-hadron classification yielded an Area Under the Curve (AUC) exceeding 0,97. The network maintains high energy fidelity and classification capacity for diffuse sources, which are essential for mapping extended structures and conducting surveys. These results consolidate the proposed network as a promising tool for maximizing the scientific potential of the next generation of Cherenkov telescopes.