Searching for exotic particles in 4 bottom quarks-antiquarks final states with machine learning techniques at the LHC

Abstract

The standard model (SM) has been a highly successful theory in explaining fundamental particles and their interactions among themselves. However, the SM has not yet explained several phenomena, and many beyond the standard model (BSM) have been introduced to solve these unexplained phenomena. One example is the bulk Randall-Sundrum (RS) model, which proposed a new higher dimensional mechanism for solving the hierarchy problem and predicted the existence of a hypothetical particle, bulk graviton. In this study, we investigate supervised machine learning methods to search for the bulk graviton decays into a pair of the SM Higgs bosons, and each Higgs boson decays into a pair of bottom anti-bottom quarks (${\mathrm{G}}_{\mathrm{KK}}^{*}\to \mathrm{hh}\to \mathrm{b}\overline{\mathrm{b}}\mathrm{b}\overline{\mathrm{b}}$). We train machine learning models to classify events between ${\mathrm{G}}_{\mathrm{KK}}^{*}\to \mathrm{hh}\to \mathrm{b}\overline{\mathrm{b}}\mathrm{b}\overline{\mathrm{b}}$ (signal) and QCD 4b multi-jet (background) processes. The evaluation metrics are calculated in the highest score to compare the classification efficiency between Adaptive Boosting and Neural Networks algorithms after performing feature importance and hyperparameter tuning techniques to optimize the models. The results show that the Neural Networks better classify our signal versus background events with the AUC score of 0.9836, compared to the Adaptive Boosting model of 0.9741. Furthermore, the signal significance is also predicted and scaled to the integrated luminosities of Run 2, Run 3 and HL-LHC, data-taking periods of the LHC. The predictions are obtained at 1.952, 2.858 and 9.037 for the Neural Networks and at 1.968, 2.881 and 9.111 for the Adaptive Boosting.