A search for the Standard Model Higgs boson produced in association with a pair of top quarks and decaying to a pair of bottom quarks

A search for the Standard Model Higgs boson produced in association with a pair of top-quarks and decaying to a pair of b-quarks ($t\bar{t} H(H\to b\bar{b}$)) is presented. The analysis uses 20.3 $fb^{−1}$ of data taken from proton-proton collision at centre of mass energy $\sqrt{s}$= 8 TeV, collected using the ATLAS detector. This thesis focuses on the search for dilepton events where the two leptons can be either an electron or a muon. A study is presented comparing different techniques (“pairing methods”) for correctly identifying which two b-jets in an event come from the decay of the Higgs boson. The best alternative to the default pairing method used in the Run I analysis, improves the matching efficiency by more than 25%. Individual neural networks are then trained to separate signal from background using input variables based on each pairing method. No significant improvement on the separation obtained with the Run I neural network is achieved when using an alternative pairing method. The neural network performances are then compared with an alternative multi-variate classifier using the same pairing method, namely boosted decision trees. Comparing the two classifiers shows the best performance of the neural network is better than the best performance for the boosted decision tree. Similarly, for each individual pairing method, the neural network outperforms the boosted decision tree trained using the same input variables and pairing method. The inclusion of additional input variables, which provide information on the topology of the event, improves the performance of the alternative classifier (boosted decision tree) by on average ≈ 5%. The best performing BDT, goes from a separation power of 0.1915 ± 0.003 (stat.) to 0.2012±0.003 (stat.). The best performing NN goes from a separation power of 0.1925±0.002 (stat.) to 0.1946 ± 0.003 (stat.). For the best performing classifier, the alternative classifier is seen to improve on the separation achieved by the neural network by 3.4% of the neural networks separation (1.5% stat. error on neural network separation). A calibration of the MV1 b-tagging algorithm using 14.34 $fb^{−1}$ of data collected at 8 TeV using the ATLAS detector is also presented. It is performed on a sample of jets selected from t¯t dilepton events, which ensures a very high-purity, inclusive sample of b-jets. b-tagging scale factors are then derived from two separate samples of jets: those which contain a muon and those which do not. A ratio of the scale factors derived using each jet category is used as a measure of the potential bias. No significant bias is seen in either of the jet pT bins, however due to the size of the uncertainties, the measurement is not sensitive to a bias of less than ≈ 8%.