Comparison of algorithm performance
Several instances of the Dips flavour tagging algorithm have already been trained using particle-flow ("Pflow") jets, which are reconstructed from calorimeter and tracking detector information.
It is possible to re-scale the input features for the Dips algorithm in a way, so that the performance of the algorithm trained on particle-flow jets is also performing well on VR track-jets.
The Dips flavour tagging algorithm has been trained using a hybrid sample composed of events from top quark pair production and decays of hypothetical Z' bosons. Doing so, allows the training to capture both collision events with low-pt jets as they are present in top quark decays, as well as the high-pt jets originating from the very heavy Z' bosons.
An interesting study is to evaluate the performance of these algorithms on various samples, such as top quarks, hypothetical Z' bosons decaying into quarks and also hypothetical spin-2 gravitons which decay into quarks and gluons.
You can study the performance of the already trained Dips flavour tagging algorithms applied on VR track jets for three different processes.
There are four different versions of the Dips algorithm, all trained last year in 2021. They differ in the selection of tracks used for training and evaluating the algorithm. The default algorithm employs a stricter track selection, while Dips (loose) allows for tracks also with larger impact parameters (d0 and z0) and lower transverse momentum. In addition to the distinction between Dips and Dips (loose), there are also two versions in which the algorithm trained on particle-flow jets has been modified by scaling the input features such that a better performance on VR track jets is expected (hence the ptfrac in the name).
Below are tables with location of h5 files which are created from the simulated collision events and contain the already evaluated tagger scores per jet. The names of the scores are in the second table, while the paths to the h5 files and brief descriptions of the samples are in the first table.
Overview of simulated MC samples
ttbar
/nfs/dust/atlas/user/pgadow/summie2022/data/vr_ftag1/user.pgadow.410470.e6337_s3681_r13144_p5169.tdd.TrackJets.22_2_82.22-08-01_vr_ftag1_00_output.h5/user.pgadow.29908485._000002.output.h5
Powheg+Pythia8 top quark pairs with semi-leptonic decays, default benchmark sample, can be used for algorithm training and performance studies
Zprime
/nfs/dust/atlas/user/pgadow/summie2022/data/vr_ftag1/user.pgadow.427080.e5362_s3681_r13144_p5169.tdd.TrackJets.22_2_82.22-08-01_vr_ftag1_00_output.h5/user.pgadow.29908483._000001.output.h5
Pythia8 Z' (with mass of 4 TeV and flat pt spectrum) with decays to b-quarks, c-quarks and light-flavour quarks, can be used for algorithm training and performance studies.
graviton
/nfs/dust/atlas/user/pgadow/summie2022/data/vr_ftag1/user.pgadow.504648.e8418_s3681_r13144_p5169.tdd.TrackJets.22_2_82.22-08-01_vr_ftag1_00_output.h5/user.pgadow.29908484._000005.output.h5
MG5+Pythia8 KK spin-2 graviton (with mass of 3 TeV, 50% width) with decays to b-quarks, c-quarks and light-flavour quarks, as well as decays to tau leptons. Has not yet been used for algorithm training or evaluation.
Overview of already trained Dips algorithms
Dips
dips20210729_pu
training from June 2021, trained with PFlow jets, default track selection
dips20210729_pc
dips20210729_pb
Dips (re-scaled)
dips20210729_ptfrac_pu
training from June 2021, trained with PFlow jets, default track selection, re-scaled for better performance on VR track jets
dips20210729_ptfrac_pc
dips20210729_ptfrac_pb
Dips loose
dipsLoose20210729_pu
training from June 2021, trained with PFlow jets, loose track selection
dipsLoose20210729_pc
dipsLoose20210729_pb
Dips loose (re-scaled)
dipsLoose20210729_ptfrac_pu
training from June 2021, trained with PFlow jets, loose track selection, re-scaled for better performance on VR track jets
dipsLoose20210729_ptfrac_pc
dipsLoose20210729_ptfrac_pb
Last updated