Flavor Tagging Efficiency Parametrisations with Graph Neural Networks

The identification of jets containing $b$-hadrons is obtained through dedicated flavour-tagging algorithms and is crucial for the physics program of the ATLAS experiment. The performance of the flavour-tagging algorithm is such that the statistical precision of the simulated samples is reduced when flavour tagging is applied, in particular when requiring many tagged jets per event. The truth-flavour tagging approach aims at increasing the statistical power of the simulated samples after the event selection. The method is based on a per-event weighting, computed according to the probability for the given event to contain tagged jets. This note describes truth-flavour tagging based on efficiency maps and a novel implementation based on Graph Neural Networks. The second approach is demonstrated to also capture correlations among jets in the same event, improving the overall performance of the truth-flavour tagging method.