Enhancing hadron jet reconstruction in the CMS Level-1 trigger using machine learning.

Level–1 Trigger (L1T) algorithms used in the Compact Muon Solenoid (CMS) experiment for detecting different physics objects must be optimized to ensure that CMS continues to collect the most interesting proton–proton collision events for analysis. In this thesis, a new machine learning based approach using boosted decision trees (BDTs) is presented, which improves the jet detection performance in the L1T. In the first step, a BDT is trained using 12 features of L1T jets to generate an importance ranking of the features. The results indicate that a new algorithm for mitigating the effect of simultaneous collisions (‘pileup’) called the ‘phi–ring’ algorithm could be better at detecting L1T jets than the current ‘chunky donut’ algorithm. New BDTs are then trained separately using phi–ring and chunky donut energies as input, to confirm the previous finding. Outputs of the BDTs that use phi–ring energies as input are found to be more stable in energy scale under varying pilepup conditions, with resolution similar to the current jet detection algorithm. Hence, we propose to use the phi–ring algorithm to calibrate jet energies and improve jet detection in the CMS L1T in Run 3 (2022–2025).