A new approach for primary vertex association of B-mesons using machine learning techniques with the LHCb experiment

During Run 3 of the LHC, the LHCb experiment has an expected number of inelastic visible proton–proton interactions per bunch crossing of $\mu = 5.4$. Each of these interactions is called a primary vertex. Particles with a significant flight distance, such as $B$-mesons, must be associated with one of the reconstructed primary vertices. Currently, the association is performed by finding the primary vertex with the minimal impact parameter, defined as the distance of closest approach between the primary vertex and the  $B$-meson track.  In this project, machine learning methods with additional input features, such as the number of tracks per primary vertex and the kinematics of the $B$-meson, are studied as a new approach for associating primary vertices. These studies are performed on partially and fully reconstructed $B_{(s)}^{(+)}$ decays. Among the studied models, the most promising performance is obtained with a multiclass boosted decision tree that simultaneously considers the three primary vertices with the smallest impact parameters. 
For the partially reconstructed $B$-meson decay $B_s^0 \rightarrow K^- \mu^+ \nu_\mu$, the improvement of the association efficiency is $2.0\,\%$ on average. Even larger improvements are  achieved differentially for example for large number of primary vertices per event. Finally, a tool has been developed for the inclusion in the central LHCb software, which applies the  model to signal candidates.