1 Jet flavour tagging for the ATLAS Experiment Jonathan Shlomi on behalf of the ATLAS collaboration August 23rd SUSY2021
1
Jet flavour tagging for the ATLAS Experiment
Jonathan Shlomion behalf of the ATLAS collaboration
August 23rd SUSY2021
2
Outline
• Overview of the ATLAS flavour tagging algorithms
• Specialised taggers and new applications• Soft b-hadron tagging• charm tagging• High pt calibration• Xbb tagging
https://twiki.cern.ch/twiki/bin/view/AtlasPublic/FlavourTaggingPublicResultsCollisionDataEnjoy browsing our public results here:
3
• Overview of the ATLAS flavour tagging algorithms
The input to the flavour tagging algorithm is the tracks associated to particle flow jets, or track jets.
The experimental signature of b/c-hadron decays is the secondary vertices inside the jet [Eur. Phys. J. C (2019) 79:970]
Detector
Primary Vertex
Secondary Vertices
Flavour tagging algorithm
4
• Overview of the ATLAS flavour tagging algorithms
Impact parameter:1. IPxD2. RNNIP3. DIPS
Jet axisPoint of closest distance to primary vertex
Impact parameter is a sign of secondary decays, without having to reconstruct secondary vertices.
5
• Overview of the ATLAS flavour tagging algorithms
Impact parameter:1. IPxD2. RNNIP3. DIPS
d0, z0
d0, z0
d0, z0
d0, z0
Tracks{…
The goal of all impact parameter algorithms is to take a set of track impact parameters, and output a probability for the jet to be a b/c/light jet
pb, pc, puImpact parameter algorithm
6
• Overview of the ATLAS flavour tagging algorithms
Impact parameter:1. IPxD2. RNNIP3. DIPS
d0, z0
d0, z0
d0, z0
d0, z0
Tracks{…
IPxDTreats the tracks individually, the final class probability is the product of all track probabilities
pb, pc, pu
pb, pc, pu
pb, pc, pu
pb, pc, pu
…
pb, pc, pu
7
• Overview of the ATLAS flavour tagging algorithms
Impact parameter:1. IPxD2. RNNIP3. DIPS
d0, z0
d0, z0
d0, z0
d0, z0
Tracks{…
pb, pc, pu
ATL-PHYS-PUB-2017-003
RNN - recurrent neural network.Arranges the tracks in a sequence based on impact parameter significance, then passes them through an RNN. Correlations are taken into account => better performance.
Inputs are the significance
100
101
102
103
104
LLght-IlDvRur jet rejectLRn
ATLAS 6LPulDtLRn 3relLPLnDry0s 13 7e9, t ̄t
R11,3D,36
0.6 0.7 0.8 0.9 1.0b-jet eIILcLency
1.0
1.2
RDtLR tR
R11,3
8
• Overview of the ATLAS flavour tagging algorithms
Impact parameter:1. IPxD2. RNNIP3. DIPS
d0, z0
d0, z0
d0, z0
d0, z0
Tracks{…
pb, pc, pu
ATL-PHYS-PUB-2020-014
DIPSUses a deep set neural network - now the order of the tracks does not matter - better performance than the RNN, and much faster to train.
9
• Overview of the ATLAS flavour tagging algorithms
Detector
Primary Vertex
Secondary Vertices
JetFitter:
Try to reconstruct multiple vertices based on the assumption that they all sit on a straight line from the primary vertex
Secondary Vertex Reconstruction:1. JetFitter2. SV1
10
• Overview of the ATLAS flavour tagging algorithms
Detector
Primary Vertex
Secondary Vertices
SV1:
Try to reconstruct a single secondary vertex containing all displaced tracks in the jet
Secondary Vertex Reconstruction:1. JetFitter2. SV1
11
• Overview of the ATLAS flavour tagging algorithms
DL1:Take the output of all algorithms and produce a class prediction
DL1
pb, pc, pulog ( pb
pc ⋅ fc + (1 − fc)pu ) > cut
“charm fraction”
12
• Soft b-hadron tagging
For jets in the range of , calorimeter jets are no longer a viable option - energy resolution is degraded.
the b-hadron energy is low, the secondary vertices are too close to the primary vertex
pT < 20 GeV
Either use jets built only from tracks,or reconstruct the b-hadron decays without considering jets.
ATLAS-CONF-2019-027
13
• Soft b-hadron tagging
Consider the vertex properties as discriminating variables between b-hadron vertices and fakes
Vertex mass, pt, impact parameter with respect to the primary vertex, angular spread of the tracks…
14
• Soft b-hadron tagging
For b-hadron below 15 GeV, the jet-less vertexing works best, while for track-jets are better.
pTpT > 15
Used in ATLAS -CONF-2020-003ATLAS SUSY-2018-12
T-LVT/TC-LVT Are different vertex reconstruction algorithms.
15
• Charm tagging
log ( pb
pc ⋅ fc + (1 − fc)pu ) > cut
log ( pc
pb ⋅ fb + (1 − fb)pu ) > cut
As the performance increases for the DL1 tagger, it becomes viable as a charm tagger by defining the discriminant slightly differently.
ATLAS-CONF-2021-021
pc ↔ pb
fb ↔ fc
16
• High pt calibration
Extend the data/MC comparison of b-tagging efficiency to jets in the range of 500 GeV < pT < 1000 GeV
ATL-PHYS-PUB-2021-004
17
• High pt calibration
Estimate the uncertainty on data/MC efficiency scale factors for jets in the range500 GeV < pT < 3000 GeV
ATL-PHYS-PUB-2021-003
Using simulation based modification to the tagging efficiency
- tracking efficiency/resolution- Jet energy resolution- Interaction of b hadrons with the inner
detector
18
• High pt calibration
Estimate the uncertainty on data/MC efficiency scale factors for jets in the range500 GeV < pT < 3000 GeV
ATL-PHYS-PUB-2021-003
Using simulation based modification to the tagging efficiency
- tracking efficiency/resolution- Jet energy resolution- Interaction of b hadrons with the inner
detector [http://atlas.web.cern.ch/Atlas/GROUPS/PHYSICS/PLOTS/FTAG-2020-002/]
19
• Xbb tagging
Discussed in detail in the talk by Changqiao LiPerformance and calibration for the identification of boosted Higgs bosons decaying into beauty quark pairs in ATLAS
ATL-PHYS-PUB-2021-004
Key idea: Tagging boosted Higgs vs. Z jets
Large R jets
Sub-jets
20
• Summary
The ATLAS flavour tagging algorithms are continuously evolving.In addition to increasing the performance of the core algorithm we are branching out to other specialised domains, and those algorithms are also evolving and increasing in their performance.
Thanks for your attention!