Event reconstruction and analysis in CMS using Artificial Intelligence

Recent developments in computer hard- ware and deep-learning algorithms, combined with large datasets, lead to impressive progress in artificial intelligence (AI) in the past few years. Although only marginally studied in high-energy particle collisions, deep-learning algorithms already demonstrated the ability to perform particle and event classification, estimation of kinematic variables, and anomaly detection. Those abilities are extremely useful in the analysis of the unprecedented amount of proton- proton collisions expected in the next running phases of the Large Hadron Collider (LHC) at CERN. The CMS detector will undergo major upgrades to deal with the increasing number of additional collisions per LHC bunch crossing (pileup), benefiting from more finely segmented detectors and precise timing information, and one of the central subjects of this thesis is dedicated to the development of versatile data acquisition software for the new MIP Timing Detector. In addition to hardware improvements, the success of these upgrades will heavily depend on fast, robust, and adaptive event processing and analysis techniques. Consequently, the majority of the work performed for this thesis is dedicated to developing and testing new AI-based reconstruction methods for the electromagnetic calorimeter of the CMS experiment. It covers two steps of the full chain of electromagnetic object reconstruction. The first one is the evaluation of the kinematic variables from the energy signatures left by standalone particles in the calorimeter. The second one combines these standalone particles into a unified object known as SuperCluster, which is crucial for accurate particle energy reconstruction. Both of the tasks are developed separately, and for each of them, a dedicated AI model is created and its performance is assessed and compared with the current traditional approach.