A New Trackster Linking Algorithm Based on Graph Neural Networks for the CMS Experiment at the Large Hadron Collider at CERN

The upcoming High-Luminosity Large Hadron Collider (HL-LHC) upgrade is set to increase the number of particle collisions, which presents a significant challenge to existing reconstruction algorithms. To address the associated rise in data complexity, the Compact Muon Solenoid (CMS) at LHC is developing a new endcap High-Granularity Calorimeter (HGCAL) that can withstand the HL-LHC’s harsher conditions and investigate high-energy collisions. During the particle shower reconstruction phase in HGCAL, 3D graph structures called tracksters are produced, believed to originate from the same physics object. However, due to the detector’s irregular geometry, physics processes, and particle overlaps (pile-up), tracksters are often fragmented, degrading the reconstruction quality. In this thesis, machine learning approaches are investigated, with a particular emphasis on Graph Neural Network (GNN) models, to enhance event reconstruction through improved calorimetric clustering. An end-to-end trainable GNN-based algorithm for accumulating incomplete energy fragments into well-formed tracksters is proposed with this goal. The algorithm is integrated into the CMS Software package as a linking plug-in, and its clustering and physics reconstruction performance is evaluated on simulation data. The model presented in the thesis outperforms the currently used rule-based state-of-the-art benchmark in all metrics and improves ParticleFlow reconstruction even in the challenging environment of HL-LHC.