Identification of $b\bar{b}$-Jets Using a Deep-Sets-Based Flavour-Tagging Algorithm at the ATLAS Experiment

Flavour tagging is a crucial tool in particle physics since it allows to identify the flavour of heavy hadrons inside hadronic jets. This is especially important for challenging analyses like the search for the $t\bar{t}H(\to b\bar{b})$ signal. This search suffers from a large irreducible $t\bar{t}+b\bar{b}$ background which contains the same particles in the final state as the signal. In the background process, a radiated gluon can split into a $b\bar{b}$ pair, which can lead to a single jet that contains two $b$-hadrons ($bb$-jet). A successful identification of $bb$-jets could allow for a better rejection of this background. However, this is not possible with common flavour-tagging algorithms. In order to enable $bb$-jet identification, an extended flavour-tagging is presented in this thesis. The design of this algorithm is built on the deep-learning-based flavour-tagging algorithm DL1d, which combines the output of multiple low-level algorithms and is used in the ATLAS experiment in Run III. One of those low-level algorithms is the DIPS algorithm, which is based on the machine-learning concept of Deep Sets. The DL1d tagger receives an additional output category for $bb$-jets, which is done in two stages. First, the DIPS algorithm itself is extended ($bb$-DIPS), which exploits the information of the tracks that are associated to the jet. Afterwards, an extended version of DL1d ($bb$-DL1d) is implemented, which combines the output of $bb$-DIPS with the output of two other low-level algorithms. Furthermore, optimisation studies regarding the definition of the tagger discriminant are performed. The final $bb$-DL1d tagger is able to successfully identify $bb$-jets while still performing well in single-$b$-tagging. At a 70% single-$b$-jet identification efficiency, $bb$-DL1d achieves a $bb$-jet rejection of $8.7\pm0.3$. Moreover, the same algorithm outperforms the ATLAS DL1r tagger used in Run II, by reaching a 27% larger $c$-jet rejection while performing equally well in light-flavour-jet rejection at a 70% inclusive $b$-jet efficiency. Finally, first steps regarding a cut in a two-dimensional discriminant plane are presented, providing an outlook for future applications of $bb$-DL1d.