Searching for Dark Matter and Vector-like top quark production in the ATLAS experiment and AI-driven Anomaly Detection

The Standard Model of particle physics is the theoretical framework that provides the best description of the subatomic world. This Thesis begins with a review of the Standard Model of particle physics from both historical and phenomenological perspectives. The discussion revisits the experimental milestones that culminated in the discovery of the Higgs boson and underlines the gauge structure that successfully describes strong, electromagnetic, and weak interactions. After discussing the main properties and predictions of this theory, some of its most relevant limitations are highlighted. The work presented in this document is mainly related to two open problems in Particle Physics: the Dark Matter and the hierarchy problems. The astronomical and cosmological evidence for non-baryonic matter is synthesized, and the most popular Dark Matter candidates are briefly reviewed. Subsequently, Vector-Like Quarks are presented as an attractive solution to the hierarchy problem that naturally emerges in a variety of theories beyond the Standard Model. In both cases, a special emphasis is placed on the predictive power of a simplified Lagrangian that incorporate the minimal set of free parameters required to capture the essential collider phenomenology. After the theoretical motivation, the experimental setting is detailed. Key accelerator concepts are explained to introduce the operating conditions of the collisions delivered by the Large Hadron Collider in the period between 2015 and 2018. The data analyzed in the two searches presented in this work correspond to proton-proton collisions at a center-of-mass energy of 13 TeV, with an integrated luminosity of 139 $fb^{-1}$. The main parts of the ATLAS experiment are described, starting from the different subdetectors and ending with the data acquisition system that triggers and record the proton-proton collision at a very high rate. The event generation chain is explained in detail, from the initial scattering to the final state of the event, including the full detector response of all particles. The reconstruction of the particle kinematics and its identification from electronic signals are also discussed. These techniques are applied to both the simulations and the real data, for which the efficiencies need to be calibrated to perform data/MC comparisons and apply statistical inference in the posterior analyses. Both the simulation and the reconstruction procedures introduce systematic uncertainties that are propagated to the final results. The first search of this Thesis is motivated by the Dark Matter problem. Since the Dark Matter is not expected to interact with the detector, the searches for Dark Matter require the production of additional objects to trigger the event. This analysis looks for events with a large missing transverse momentum together with a single top quark, referred to as mono-top events. Two simplified Dark Matter scenarios are considered, in which a different hypothetical mediator is produced in each case to couple the Dark Matter to the Standard Model particles. The results are also interpreted in the context of Vector-Like Quarks, in which a singly-produced top partner decays into a top quark and an invisible Z boson (decaying to neutrinos). The second analysis presented in this document consists of the first combination of searches for a singly-produced vector-like top quark. This combination is part of the ambitious VLQ program of the ATLAS experiment, which has already explored extensively both the pair and single production modes of VLQs in different decay channels. The three combined channels are interpreted within the same simplified Lagrangian, allowing the combination to improve sensitivity across a broad parameter space. Despite no significant excess being observed in any of the analyses, small excesses in the individual searches could potentially be enhanced from this statistical combination. The use of advanced Machine Learning techniques have being increasingly used in the last years at the High Energy Physics experiments, allowing to gain sensitivity in the new physics searches. However, the lack of evidences for new phenomena using traditional searches, in which a particular signal is searched for, has led to the development of more generic searches. The Anomaly Detection approach aims to identify deviations from the Standard Model background, without the need of a specific signal hypothesis. The third study of this Thesis, aims to boost the development of Anomaly Detection in collider searches by reporpusing cutting-edge classifiers with minimal tuning. Data from the ATLAS experiment are not used in this study, but the analysis is performed using benchmark simulated datasets produced by the Dark Machines community. These data also correspond to proton- proton collisions at 13 TeV, but with a simplified ATLAS detector simulation. Finally, an Appendix is included to describe a new technique to parameterize theoretical uncertainties with Neural Networks. This novel technique is tested in this study for hadronization uncertainties. Since the hadronization part of the collisions needs to be generated with approximate phenomenological models, this technique aims to capture in a set of weights the impact from using different hadronization models.