Searches for New Physics at the Large Hadron Collider

The modern particle phenomenologist must be knowledgeable not only in both experimental and theoretical physics, but also in advanced machine learning techniques that have proven successful in recent years. This thesis provides a pedagogical overview of particle physics analyses at the Large Hadron Collider, including both tried-and-true supervised machine learning methods and also cutting edge novel unsupervised machine learning methods that are in development. It also contains three published studies, and another study that is in progress, in which these techniques are applied to improve current understanding of particle physics theory and experimental approaches. In the first study the classic methodology of a manual cutflow is used to assess the discovery/exclusion potential of a charged Higgs boson in the Two Higgs Doublet Model using the process pp → tH± followed by the H± → W±A and A → t¯t. This provides the signal tt¯tW± and we study the final states with three leptons and with two leptons of the same charge. It is found that with minimal data in the early runs of the 14 TeV LHC the charged Higgs can be excluded at 95% confidence for masses up to 1 TeV if the mass splitting of the charged Higgs and pseudo-scalar Higgs is within 100 to 300 GeV. In the second study the possibility of a beyond the Standard Model CP violating top-Higgs coupling is explored using the process pp → thj. The angular variables of the decay products of the top quark are non-trivially effected by the level of CP violation of the top-Higgs coupling and can be used as a powerful probe into this coupling. A boosted decision tree analysis is performed to fully optimize the extraction of the thj signal. It is found that the combined effect of introducing the angular variable to the analysis as well as the usage of the boosted decision tree algorithm leads to a large increase in exclusion potential of CP violation in the top-Higgs coupling compared to previous literature. In the third study the level of quantum scattering interference between signal and background of the process pp → t¯bH− followed by H− → b¯t (and conjugate process) and the irreducible backgrounds with final state pp → t¯tb¯b is investigated. It is found that in some areas of the parameter space - when charged Higgs width to mass ratio is large - that this interference, which is traditionally assumed to be negligible in many analyses, is extremely large. In some instances it can be as large as the signal cross section itself. Finally, in the fourth study we show that a cutting edge technique known as a variational autoencoder, can be used to effectively parametrize composite images of events detected at the planned XENONnT dark matter detector. The variational autoencoder model is trained exclusively on electron recoil background images and builds a profile of the properties of these images. When the model is then exposed to a new dataset which includes a mixture of both electron recoil and simulated dark matter nuclear recoil events the two signals differ at 95% confidence. This acts as a proof of concept that the anomaly detection methodologies rapidly growing in popularity in many areas can find powerful applications in dark matter direct detection.

Author dc:creator

• Patrick, Riley Richard Geoffrey

Advisors dc:contributor.advisor

• Williams, Anthony
• White, Martin

• Beyond the Standard Model
• Higgs Physics
• Large Hadron Collider
• Deep learning
• Anomaly detection

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/2440/135240

OAI identifier oai:identifier

oai:digital.library.adelaide.edu.au:2440/135240