Deep Learning Methods Applied to Higgs Physics at the LHC

The impact that machine learning (ML) has had on research in high-energy physics (HEP) is undeniable; the use of ML-based classifiers in many analyses is now the norm, and they have a long history of being used within reconstruction algorithms (e.g. $b$-tagging at LEP, 1995). The combined effect of advances in both hardware and algorithms, and the vast datasets now available (both collected and simulated), has facilitated the paradigm shift towards deep learning (DL). The true effect of this is only beginning to be felt in the field of HEP, but it is something that can be expected to continue into the future; overturning several "foundational" principles, and challenging more traditional thinking. Over the course of my PhD I have developed and applied a range of advanced ML and DL algorithms to the study of the fundamental interactions of matter. This thesis documents over four and a half years of work; beginning from preliminary feasibility studies, through algorithm optimisation studies and dedicated software development, and concluding with a full study leveraging the refined methods. From a pragmatic perspective, no matter how beautiful a new method is, it is only applicable if either it provides better performance than alternative methods, or it savesthe applier time and effort. Preferably, the new method should meet both of these criteria and throughout this thesis I test DL methods against them. As a context for studyingthese new methods, I pick the sub-field of Higgs physics. With the discovery of the Higgs boson in 2012, this area has now moved into the regime of precision measurements and experimental confirmation of theoretically predicted channels of production and decay; any deviation from the theoretical model (the Standard Model (SM)) could be a sign of new physics, and open up further avenues for progressing humanity's understanding of Nature. The importance of these studies, therefore, goes far beyond demonstrating the applicability of new methods. One key focus of my research is the search for the simultaneous production of two Higgs bosons. One property of the Higgs boson that has yet to be measured is the strength with which they couple to themselves (the Higgs interacts with particles that have mass, and therefore is expected to interact with itself). Beyond academic interest, the value of this parameter can allow us to more precisely estimate the stability of the Universe and determine whether we are at risk of suddenly being snuffed from existence (as we know it) by spontaneous vacuum decay. The production of such di-Higgs processes, however, is extremely rare, meaning that whilst we are unlikely to discover it at the LHC (barring the effects of new physics), it is a perfect testing ground for DL methods; even moderate improvements in sensitivity, through whatever means, can correspond to months or years of less data-taking required to achieve discovery. At the time of writing, the Large Hadron Collider (LHC) has finished its Run-II datataking and three years worth of $\sqrt{s}$ = 13 TeV proton-proton collision-data (137.19 fb$^{-1}$ of integrated luminosity) stands ready to be scrutinised in minute detail for any hints of new phenomena, such as super-SM rates of production of di-Higgs. The concluding chapter of this thesis documents my contributions to the examination of this extremely large dataset.