Identification of $b$-jets and $c$-jets using Deep Neural Networks with the ATLAS Detector - The Development and Performance of a Family of DL1 High-level Flavour Tagging Algorithms

In this thesis, a new family of high-level jet flavour tagging algorithms called DL1 is presented. It is now established within the ATLAS collaboration at the Large Hadron Collider at CERN to be applied to Run 2 pp collision data at $\sqrt{s}$ = 13 TeV. DL1 represents the first use of Deep Learning for ATLAS physics object reconstruction as well as the first major application of advanced deep neural networks within the collaboration. The determination of jets originating from heavy flavour quarks is used to probe the particle identity of particles created in the pp collisions. These heavy flavour quarks play a major role in searches for new physics and precision measurements. The potential of Deep Learning in flavour tagging using inputs from lower-level algorithms has been investigated. A systematic grid search over architectures and the training hyperparameter space is presented. In this neural network approach, the training is performed using multiple output nodes, which is a naturally suited method for the task of jet flavour tagging. This also provides highly flexible tagging algorithms. The DL1 studies presented show that the obtained $b$- and $c$-jet tagging algorithms provide good discrimination against jets of other flavours considered in flavour tagging. Their performance for arbitrary background mixtures can be adjusted after the training according to the needs of the physics analysis. The resulting development and structure of DL1 as well as the architectures of the neural networks used in the tagging algorithms are described and a detailed set of performance plots is presented, obtained from simulated $t\overline{t}$ events at $\sqrt{s}$ = 13 TeV and corresponding to the data taking conditions during Run 2 where these tagging algorithms will be applied. Performance comparison plots between predictions from simulation and collision data as well as the final $b$-jet tagging scale factors of the calibration for physics analyses usage are provided and show excellent agreement. The algorithms are not only well optimised but also generalise the learned jet topologies well to other event topologies. A fully fledged family of robust $b$- and $c$-jet tagging algorithms with a reduced amount of required person power is established and recommended within the ATLAS collaboration. Now that DL1 has been established, it is expected to improve a wide range of physics analyses throughout the collaboration