8. Continuous Calibration of ATLAS Flavour-Tagging Classifiers via Optimal Transport Maps

19 May 2026, 15:37
1m
Patio and Auditorium Hall (CICSU)

Patio and Auditorium Hall

CICSU

Centre de conférences internationales - Sorbonne Université 4 Place Jussieu, 75005 Paris

Speaker

Alessandro Montella (Stockholm University (SE))

Description

Algorithms to identify jets from b-hadrons (b-jets) are widely in use in ATLAS and are crucial for measurements and searches targeting processes with top quarks or Higgs bosons in the final state, for instance, b-tagging played a central role in the observation of the Higgs boson decay into bottom quarks. These b-tagging algorithms are trained on simulation, therefore, their performance in data has to be checked and corrections for simulations have to be derived. Traditionally, operating points using intervals of fixed b-tagging efficiencies are defined such that calibrations can be performed in these intervals. However, this limits the flexibility in the application of the algorithms in analyses. This work introduces a continuous calibration of the ATLAS DL1r flavour-tagging algorithm using optimal transport maps. The simulated flavour-tagging classifier outputs (probabilities for b-, c-, and light-flavour jets) are transformed to match the observed data distributions, minimizing alterations to the simulation while achieving closure with data. The calibration is derived as a function of the jet transverse momentum using high-purity samples of b-jets from top-antitop events in ATLAS Run 2 data, employing neural solvers to approximate the optimal transport maps. This approach provides a fully continuous, three dimensional calibration of flavour probabilities, removing the need for discrete operating points and opening new possibilities for high-dimensional corrections in ATLAS analyses, therefore enhancing the precision of measurements and the sensitivity of searches for new physics.

Track Performance and Tools

Author

Alessandro Montella (Stockholm University (SE))

Presentation materials