Multivariate classification of charged particle tracks for an improved b-tagging performance with the ATLAS detector

The identification of jets containing b-hadrons, called b-tagging, is a key element for many precision measurements and searches for new physics. Heavy flavour tagging algorithms used in ATLAS are based on modern machine learning techniques that exploit characteristic features of tracks and displaced (secondary) vertices to distinguish b-hadron jets from c-hadron or light-flavour jets. In particular, the ability to reconstruct secondary vertices dominates the performance of flavour tagging algorithms. However, for high transverse momenta, above 400 GeV, the efficiency to reconstruct a secondary vertex starts to decline significantly. One of the reasons for this degradation is the increase of track multiplicity inside a jet for high jet energies, leading to a degradation of vertex finding algorithms. It has been previously shown, that multivariate analysis techniques can be used to reduce the amount of tracks that do not carry any information from the b-hadron decay and thus help to recover the performance of the vertex finding at high p${}_T$. The studies presented in this talk aim for an optimisation of these track classification taggers. In addition, the performance of such a technique and its input quantities are compared for various parton shower generators in order to gain a better understanding of model dependencies.