Preview

PDF, English Download (13MB) | Terms of use

Abstract

The LHC produces huge amounts of data in which signs of new physics can be hidden. To take full advantage of existing or future LHC data, it is worth exploring novel techniques such as deep learning methods or global analysis strategies. We first study BNNs, a deep learning method which has the benefit of providing uncertainty estimates while still performing similarly to ordinary deep neural networks. We show in different studies how BNNs can be applied to LHC physics. A BNN is trained for the task of jet calibration and we illustrate how we can disentangle and understand the predicted uncertainty types. Furthermore, we discuss different ideas of how BNNs can be used for collider event generation, introducing error bars which are necessary when replacing existing Monte Carlo simulations with deep learning methods. The focus is then shifted to the global analysis of LHC data. To derive accurate bounds on the space of new physics, it is crucial to have an optimal understanding of all the uncertainties involved. We first discuss the results of matching a specific UV model to the SMEFT. The matching procedure introduces an additional theory uncertainty which has a significant impact on the derived bounds. We then study more generally the results of a global SMEFT analysis for different statistical approaches, namely a Bayesian one and a profile likelihood based one. Both procedures are compared and the impact of different uncertainty treatments is discussed. We encounter that it is crucial to describe correlations between measurements as accurate as possible.