Novel Data Analysis Techniques for BSM Physics Applications

Abstract

Since the discovery of the Higgs boson in 2012, the exact nature of new physics beyond the Standard Model (BSM) remains unknown. Modern experiments work to optimise analyses on specific regions of the parameter space where new physics is considered likely to exist. This thesis aims to identify issues with modern experimental techniques, and proposes solutions using a variety of novel data analysis techniques. Throughout this thesis, a particular emphasis is placed on the BSM theory known as supersymmetry which introduces superpartners for every particle in the standard model. This thesis is broadly split into four parts. The first part is an overview of modern particle physics, including the standard model, supersymmetry, and high energy collider experiments. Additionally, an in depth introduction to machine learning is presented. The following part concerns unsupervised anomaly detection in the context of high energy collider experiments. In a typical supervised experiment, one must specify a number of assumptions about the nature of the BSM signal being searched for. I show that by using unsupervised machine learning techniques, one is able to construct a quantity that is able to improve the performance of a typical analysis in a signal agnostic fashion. These techniques are tested on a wide array of BSM signals from various theories including supersymmetry. The next part explores dimensional reduction of a supersymmetric theory. Typically supervised analyses are done on a small subset of the original parameter space, fixing the other parameters at arbitrary values. This shields the rich phenomenology of the BSM theory from the analysis, allowing many spectra to go undetected. By performing a dimensional reduction using machine learning, I am able to construct a 2-D space which captures the full phenomenology of the original high dimensional parameter space. Using this dimensionally reduced representation, I identify interesting regions of the parameter space, and exclude a number of previously unexcluded models. The final part examines a optimisation algorithms in high dimensional spaces. Once a particular BSM model has been chosen, it is important to consider which parameter values yield the closest match with current experimental data. This can be considered as an optimisation problem in a high dimensional space. By comparing the performance of a number of optimisation algorithms on analytic test functions, as well as a likelihood function based on a global fit of a supersymmetric theory performed by the GAMBIT collaboration, the strengths and weaknesses of each algorithm are identified. Ultimately I am able to draw conclusions on which algorithms are suitable for high dimensional particle astrophysics problems in general.