A high-dimensional unbinned measurement of 24 kinematic observables using a machine learning-based analysis of data recorded by the ATLAS detector

In this thesis, a precision measurement of high momentum $Z\rightarrow\mu\mu$ events in proton-proton collision data recorded during 2015–2018 by the ATLAS detector is performed. This measurement is made using a new method, MultiFold, that leverages machine learning to produce a result that is not only corrected for detector effects, but is also high-dimensional and unbinned. This marks the first measurement of its kind based on LHC data and includes full uncertainty covariance. A result in this form greatly increases its utility: it allows the user to construct distributions with any reasonable binning, craft multi-dimensional distributions, and construct new observables that are a function of those included in the result. The measurement is presented in the form of a large event dataset with 24 observables and a variety of weights corresponding to the nominal result and the associated uncertainties. It will be publicly available with an associated user guide to provide examples and give instructions for appropriate use. The $pp\rightarrow Z$+jets cross section is measured as a function of 16 observables that probe the kinematics of the Z boson, the produced muons, and the two highest momentum charged particle jets. A further eight observables probe the internal structure of the two jets. In order to demonstrate the flexibility of the results obtained with the MultiFold method, cross sections are also presented as a function of three additional observables that are calculated from the resulting event sample. A series of validation tests are performed in order to ensure the measurement is accurate. These tests make use of a set of mock data with known differential cross section distributions, and examine subsets of the dataset and the closure of multidimensional distributions to ensure that agreement within uncertainties is maintained between the measured result and the true values. Based on these tests, a set of recommendations will be provided in the user guide regarding the usage of the results.