Novel Deep Learning Techniques for Reconstruction of Paricles Decaying to Merged Photons and Energy Calibration of Charged Hadrons

Abstract

The Standard Model (SM) of particle physics, continued to be validated by measurements from the CMS and ATLAS experiments at the CERN LHC, is an excellent description of known elementary particles. However, its inadequacy is evident in the divergent values obtained for the quantum corrections to the Higgs boson mass calculated in the SM framework and its failure to explain phenomena such as dark matter and matter-antimatter asymmetry. Null results from searches for physics beyond the standard model (BSM) have prompted investigations into more exotic decay topologies. This study focuses on measuring the invariant mass of highly boosted low mass particles which decay to two photons and are reconstructed as a single photon object in the CMS detector. Such signatures can be used to identify unconventional Higgs boson decays, specifically H→aa, where "a" is a hypothetical particle resembling a light axion-like pseudoscalar and it can decay to two merged photons. To measure the invariant mass of these merged a ! gg candidates, graph neural network (GNNs) methods are explored, leveraging their capability to handle irregular geometries of electromagnetic showers produced in the electromagnetic calorimeter of the CMS detector. We use low level input features to train the machine learning model, necessitating an understanding of CMS event reconstruction machinery and access to relevant underlying information. We also studied the calibration of hadron showers using the GNNs and demonstrate that resolution could be significantly improved as compared to the conventional methods currently used by the CMS collaboration.