Deep Learning Techniques for Event Reconstruction at ATLAS

The identification and energy calibration of pions are two fundamental tasks in the reconstruction of jets generated by proton-proton collisions at A Toroidal LHC Apparatus (ATLAS). In this poster, I will present novel approaches using deep learning models for these tasks. The results will be based on the recent developments of graph neural networks with data from calorimeter clusters and tracks in the form of point clouds. The discussion includes two main topics: the use of Mixture Density Networks as an additional deep learning layer, and the generalization of existing networks for calibrating ρ and Δ resonance decays. These are critical steps towards a computer vision approach for ATLAS event reconstruction.