Wing Sheung Chan

122 Conclusion and outlook each trained to discriminate against a particular background contribution, has also been employed. The outputs of these classifiers are combined into a powerful discriminant in a flexible and transparent way. This allows us to optimise the separation between the different background events, which in turn allows the maximum-likelihood fits to more precisely constrain the background models. It is estimated that the use of neural network classification has improved the sensitivity of the search by roughly 50% with respect to a conventional cut-based analysi s † . Another important factor to the sensitivity is the precision of the signal and background predictions. For this, data-driven techniques are utilised to reduce the impact of uncer- tainties related to theories and simulations. For events where quark- or gluon-initiated jets are misidentified as τ had - vis candidates, predictions are made using the data-driven fake-factor method. For the signal and the Z → τ τ and Z → `` backgrounds, corrections to the production cross section and transverse momentum of the Z bosons are derived from data and applied to the simulated events. Furthermore, the misidentification rate of light leptons as τ had - vis candidates in Z → `` events is also corrected using observed data. Owing to these data-driven estimations and corrections, the search is able to make full use of the available data, without being heavily limited by systematic uncertainties. The result of this thesis marks an important point of transition for lepton flavour violation searches. It demonstrated that searches for rare LFV phenomena at the LHC and with the ATLAS detector are not only possible, but also comparable to the best searches in the past. Better yet, it has been shown that the sensitivity of the search is still primarily limited by statistical uncertainties, thanks to the data-driven approach employed. This implies that with even more data that will be collected in the future runs of the LHC, further improvement on the sensitivity can be expected. Besides directly benefiting from the increase of data size, there are also other improve- ments left to explore. An obvious one is the inclusion of leptonically decaying τ leptons ( τ lep ) in the search. In this thesis, only Z → `τ events where the τ lepton subsequently decays into hadrons ( τ had ) are considered. However, by doing so, roughly 35% of all the possible signal events are being omitted. Although less signal events can be expected in the `τ lep channels compared to the `τ had channels, significantly less background events can also be expected, especially for events with misidentified jets. In fact, extensive efforts have already been made within the ATLAS collaboration to explore these potential new channels. Preliminary studies show that the expected sensitivity of the `τ lep channels could be comparable to that of the `τ had channels considered in this thesis. Since the measurements in the `τ lep and `τ had channels are complementary to each other, combining the measurements can significantly improve the sensitivity of the search. To conclude, we are living in exciting times, where the LHC and the ATLAS experiment have opened up new opportunities for lepton flavour violation searches. The result of this thesis is just the first of many more exciting results to come. † The estimation was done at an early stage of the study, where both the neural network classification and the conventional cut-based analysis in the comparison had not been explored to their full extent. However, the result still shows undeniably the superiority of the neural network classification.