Search for heavy Higgs Bosons and Axion-Like Particles with the CMS Experiment via Deep Neural Networks

Abstract

This interdisciplinary dissertation investigates several key physical and computational challenges inherent in high-energy physics analyses of top quark pair (tt) production. As the heaviest known elementary particle, the top quark occupies a unique position within the Standard Model (SM), serving as a critical probe for new physics. With the advent of the Large Hadron Collider (LHC), physicists are now able to study precision effects in tt production with unprecedented detail. The foundation of this research is a CMS (Compact Muon Solenoid) experiment analysis searching for heavy scalar or pseudoscalar particles that couple to top quark pairs via Yukawa interactions. A central result of this search is the observation of a significant excess at the mtt production threshold. This work explores several interpretations of this excess, among them a reinterpretation of CMS results within the framework of Axion-Like Particles (ALPs) which could serve as part of an explanation for dark matter. Furthermore, the thesis evaluates the potential for interpreting such excesses in models incorporating CP-violation, supported by a sensitivity study conducted via Simulation-Based Inference (SBI). To improve the sensitivity of future tt analyses, this work addresses the requirement for higher mtt resolution and more accurate simulations. The former is discussed with the implementation of a Lorentz-algebra Transformer architecture designed to optimize the reconstruction of the top quark system. The latter is investigated by integration and validation of the ttJ_MiNNLO code within the CMS software framework to study the impact of next-to- next-to-leading-order (NNLO) corrections. The dissertation further addresses the computational and statistical complexities of high- dimensional inference. In scenarios where the assumptions of Wilks’ Theorem are not met, this work explores a novel approximation method for Feldman-Cousins inference. Finally, the thesis tackles a critical bottleneck in applying machine learning to high-energy physics: the instability caused by simulated collision events with a negative weight through interference effects in Monte Carlo generated training datasets. To resolve this, a new algorithm (Pull by Partial Labels (PPL)) is introduced. Utilizing disambiguation-free partial label learning, the PPL algorithm is shown to stabilize training across controlled and experimental scenarios. Remarkably, the utility of the PPL method extends beyond physics; its interdisciplinary application is demonstrated in the field of "Machine Unlearning" (specifically class forgetting), where it performs competitively against established unlearning techniques.