Optimisation and performance studies of the ATLAS $b$-tagging algorithms for the 2017-18 LHC run

The optimisation and performance of the ATLAS $b$-tagging algorithms for the 2017-18 data taking at the LHC are described. This note presents the use of additional taggers to further enhance the discrimination between $b$-, $c$- and light-flavour jets, and on new studies for more performant training of the algorithms and for assessing the universality of the training campaign in typical physics processes where flavour tagging plays a crucial role. Particular attention is paid to the inclusion of novel taggers, namely a Soft Muon Tagger, based on the reconstruction of muons from the semileptonic decay of $b$/$c$-hadrons, and a Recurrent Neural Network Impact-Parameter tagger that exploits correlations between tracks within the jet. New variants of the high-level discriminant, based on boosted decision trees and modern deep learning techniques, are also presented. The overlap between the jets tagged by the various $b$-tagging algorithms is studied, and the dependence of the tagging performance on the physics process producing the jets is explored. Comparisons between Monte Carlo simulation and 2016 data for both the input variables and the output $b$-tagging discriminants are also shown.