ML-based charged isolation tool for Run 3 and a comparative assessment (LHCb)

During Runs 1 and 2, many LHCb analyses have relied on multivariate isolation techniques to inspect extra particles in a collision event that accompany a signal decay, to suppress/control backgrounds and to reconstruct excited states. There are currently no such tools for Run 3, where the isolation problem becomes more complicated due to the increased particle multiplicity per event. The isolation of interesting particles in hadron collisions at LHCb becomes more challenging as the instantaneous luminosity increases in Run III. A gradient boosted decision tree is used to select associated particles. A systematic study of the input variables leads to a reduced correlation between the parameters. The input variables have been chosen to be topological, and independent of the decay kinematics. They are a combination of cone, track and vertex isolation variables. This model presents better background rejection and signal efficiencies than cone and impact parameter based track isolation tools. Similar performance is achieved in Run III conditions compared to Run II. This report is part of a summer student project.