Uploaded image for project: 'ATLAS Flavour Tagging'
  1. ATLAS Flavour Tagging
  2. AFT-659

QT: Flipped taggers for light jet calibration



    • Task
    • Resolution: Unresolved
    • Minor
    • None
    • None
    • None


      The student (tianao) will work on finding suitable modifications to the algorithms, such as the flipped taggers, to enable the light-flavour jet calibration.

      In more details, he will work on the following topics:

      Run the light-flavour jet calibration to perform fits on pseudo-data from MC simulation with modified flavour fractions to learn which level of b-jet contribution from the flipped tagger is still acceptable for a fit which does not fail

      Improve the plotting scripts for diagnosing the different flipping strategies in puma to provide additional diagnostics, such as correlations of the regular tagger and the flipped tagger. For instance, checking the migration of jets when switching from the nominal tagger to the flipped tagger. This will eventually result in a more sophisticated definition of the flipped tagger operating point, which brings smaller nominal-to-flip extrapolation uncertainty. Identify variables which have a strong impact on the flipped tagger being different from the regular tagger.

      Train GN2 without certain variables which might impact the flipped tagger having similar performance for light-flavour jets and b-jets and quantify the performance loss w.r.t. a training with full set of input features. A good candidate is removing the detector-signed impact parameters and only train using lifetime significances. Another option is to add the lifetime-signed d0 and z0 sin (theta) variables to replace the detector-signed impact parameters.

      Finally, he will investigate how much the f_c value can be varied to reduce the light-flavour jet rejection in favour of increasing the c-jet rejection and test if this results in a better flippable tagger.

      If time allows, investigate the origin why the flipping fails, if it can be traced to certain input variables or if the auxiliary tasks (e.g. the track classification) result in the tagger learning to reject light-flavour jets based on information not susceptible to the currently employed flipping strategies.

      All the above tasks will be carried out using the central FTAG workflows with clear documentation. The work will be presented in the Algorithm sub-group meeting and Software meeting. The technical details will be documented in an internal note (R&D study) and the developed software will be documented in git.

      532770 Performance Studies - Flavour Tagging
      556629 Performance Studies for Flavour Tagging

      Local supervisor: Yanwen Liu
      Technical supervisor: Sam Van Stroud
      glance entry




            tianao Tianao Wang
            fdibello Francesco Armando Di Bello
            0 Vote for this issue
            4 Start watching this issue