Enhancing a search for long-lived particles that decay to muons using machine learning at CMS

Here we present the most recent expected limits of our long-lived particle (LLP) search using data collected by the Compact Muon Solenoid (CMS) from 2022 to 2024, equivalent to ~173 fb-1, which include limits for LLPs with masses below 10 GeV for the first time. Our search uses pairs of reconstructed muons to probe theories that predict the existence of long-lived particles (LLPs) that decay to at least one pair of muons. This new iteration of our analysis replaces various cut-based decisions with robust neural networks (NNs), considering all the available features at each stage. The first NN improves our ability to correctly associate muons reconstructed using only the muon detectors in the outermost parts of the CMS detector - used to detect muons from LLPs that decayed outside the tracker volume, and those reconstructed using the tracker and the outer muon detectors. The second NN helps us compare all the possible muon pairings by providing a score that indicates how consistent each muon pair is with an LLP decay. This second NN improves our ability to correctly pair muons when a signal contains four muons in the final state. Lastly, the third NN classifies the dimuon objects into three categories: signal-like, Drell-Yan-like, and QCD-like; the last two categories being the dominant backgrounds. These upgrades provide us with higher signal efficiency and reduced background yields compared to previous versions of this search using data from 2022, equivalent to 36.6 fb-1. Furthermore, the improved background rejection enables us to probe dimuons with masses below 10 GeV.