UC Santa Barbara

Standard model (SM) four-top quark production, where proton-proton collisions produce two top-antitop pairs, is a rare process with great potential to reveal new physics. Measurement of the cross section is not only a direct probe of the top quark Yukawa coupling with the Higgs, but an enhancement of this cross section is predicted by several beyond the standard model (BSM) theories. This process is studied in fully-hadronic proton-proton collision events collected during Run II of the CERN LHC by the CMS detector, which corresponded to an integrated luminosity of 137 fb^−1 and a center of mass energy of 13 TeV. In order to optimize signal sensitivity with respect to significant and challenging backgrounds, several novel machine-learning based tools are applied in a multi-step and data-driven approach. Boosted decision tree (BDT) and deep neural net (DNN) based hadronic top taggers are used to identify hadronically decaying top quark candidates with moderate and high transverse momenta, respectively, in order to suppress backgrounds and categorize events by the multiplicity of reconstructed top tags, and an event-level kinematic BDT distribution is subsequently used to extract the signal. Control regions inspired by the “ABCD” method are used to obtain a data-driven estimate of the background, and data distributions in these control regions are given as inputs to a DNN in order to estimate the event-level BDT discriminant distributions of the major backgrounds. The observed signal strength, defined as the ratio of the observed four-top production rate to the standard model expectation, is measured to be 5.1^{+2.3}_{-2.0}. The corresponding observed (expected) significance and limit times the SM cross section are 2.25 (0.43) standard deviations and 8.39 (4.88) respectively. A combination of this result with multiple final states is in progress as of the writing of this thesis, and future BSM interpretations to investigate this larger-than-expected signal strength are planned.