Wing Sheung Chan

Summary 159 Figure S.2.: Examples of processes that are allowed (left) or forbidden (middle and right) in the Standard Model of particle physics. predicting an observable occurring probability of lepton-flavour-violating (LFV) processes. Therefore, by searching for LFV processes in experiments, we can either verify or constrain these theories, and guide ourselves towards finding the right missing pieces of New Physics. The search for Z → τ decays with the ATLAS detector In this thesis, a search for LFV Z → τ decays ( = e or µ ) using data collected by the ATLAS detector at the Large Hadron Collider (LHC) is presented. At the LHC, protons are accelerated to nearly the speed of light and made to collide with each other. These high-energy proton–proton collisions are able to produce heavy particles that are not seen in everyday life. These particles decay very quickly into lighter particles, which are then detected by the ATLAS detector. The Z boson is one of the heavy particles that are produced. During the data-taking periods in 2015–2018, approximately eight billion Z bosons have been produced by the LHC at the ATLAS detector. Such a large sample size allows us to search for LFV Z → τ decays even if they only occur very rarely. The ATLAS detector is a general-purpose particle detector. It allows physicists to reconstruct and identify most of the SM particles (except neutrinos) from the collision data it collects. Using the reconstructed energies and momenta of the identified particles, and by exploiting expected differences between Z → τ decays (signal) and events from SM processes (background), potential signal events are selected and separated from the background events. Events reconstructed with an electron or muon and a τ lepton that decayed into hadrons are selected for analysis. These selected events are then classified by neural network classifiers based on how similar their kinematic properties are to a possible signal event. The selected observed events are then compared with predictions from the background- only or background-plus-signal models. These predictions are based on simulations as well as data of observed background-like events. Maximum-likelihood fits to the observed neural network output distributions are performed to determine the overall yields of the