Semiparametric graph neural networks for energy regression of hadron showers in the CMS High Granularity Calorimeter

Abstract

The high luminosity phase of the LHC (HL-LHC) poses significant challenges of radiation damage to the components of the CMS detector expected from the increased integrated luminosity and high event pileup. In view of HL-LHC phase, the CMS collaboration has opted for the High Granularity Calorimeter (HGCAL) to replace the current electromagnetic and hadronic calorimeters. The HGCAL features high radiation tolerance, unprecedented transverse and longitudinal segmentation for both electromagnetic and hadronic compartments, and high-precision timing capabilities, facilitating efficient particle flow reconstruction, energy rejection from the pileup, and particle identification. To validate the proposed design of the HGCAL, a series of beam tests have been carried out using silicon and scintillator based sampling calorimeter prototype. The prototype was exposed to beams of high-energy e+ and pi- of momentum ranging from 20 to 300 GeV/c, and µ- of momentum 200 GeV/c. This thesis focuses on the energy regression of charged pions in the HGCAL test beam prototype using semiparametric graph neural networks. Given the complexity of the pion showers, advanced machine learning models like graph neural networks can fully utilize the reconstructed hit information, i.e., the spatial coordinates and the energy information from the fine lateral and longitudinal granularity of the HGCAL, to reconstruct pion energy efficiently.