Quark vs Gluon Jet Tagging Francesco Rubbo APS DPF 2017 1
q/g discrimination hep-ph/1106.3076 • For background (gluon) suppression - e.g. VBF, SUSY, … • For final state reconstruction - e.g hadronic top or W • Requires detailed understanding of jet fragmentation properties Many observables exploiting different fragmentation • Probes generator PS/tunes properties are useful for q/g discrimination 2
q/g tagging with jet images red = transverse momenta of charged particles green = transverse momenta of neutral particles blue = charged particle multiplicity RGB jet images: represent the jet with multiple images, exploiting tracking information, in addition to calorimeter images. hep-ph/1612.01551 3
Outline • Performance and uncertainties for quark/gluon tagging with charged particle multiplicity ATL-PHYS-PUB-2017-009 • First look quark/gluon tagging with ATLAS jet images ATL-PHYS-PUB-2017-017 4
n track • A simple observable: number of tracks ghost-matched to the jet (n track ) • Proxy for charged multiplicity (n charged ) in quark/gluon fragmentation ATL-PHYS-PUB-2017-009 • Particle multiplicity scales with the color charge (C F /C A ) • —> powerful q/g discriminant 5
n track • A simple observable: number of tracks ghost-matched to the jet (n track ) • Proxy for charged multiplicity (n charged ) in quark/gluon fragmentation ATL-PHYS-PUB-2017-009 • Particle multiplicity scales with the color charge (C F /C A ) • —> powerful q/g discriminant Improved gluon rejection at higher p T 6
n charged modeling Eur. Phys. J. C76(6), 1-23 (2016) • The challenge is assessing the modeling of the discrimination performance in MC • Run 1 n charged measurement shows large data/MC discrepancies depending on generator and tuning choices 7
Recasting Run 1 measurement for Run 2 tagger • Fragmentation modeling uncertainties from Run 1 charged multiplicity measurement. • Experimental uncertainties from track reconstruction uncertainties at 13 TeV (Run 2). 8
n charged for quark and gluon jets NNPDF Collaboration • Extract n charged separately for quarks and gluons by exploiting rapidity difference in dijet events (mostly qg—>qg) “valence” quark at high x “sea” gluon at low x p p 9
n charged for quark and gluon jets Eur. Phys. J. C76(6), 1-23 (2016) Solving for <n q >, <n g > more forward jet more central jet 10
n charged for quark and gluon jets ATL-PHYS-PUB-2017-009 Eur. Phys. J. C76(6), 1-23 (2016) Excellent closure in wide kinematic range Contamination from gg and qq events at low and high p T , respectively 11
n charged for quark and gluon jets Eur. Phys. J. C76(6), 1-23 (2016) • Unfolded measurement of ⟨ n charged ⟩ separately for quark and gluon jets as a function of jet p T • Uncertainties from • ME and PDF used to estimate quark/gluon fractions from MC • detector/physics objects (through unfolding) 12
From Run 1 measurement to Run 2 tagger Use Run 1 measurement to assess modeling uncertainty ATL-PHYS-PUB-2017-009 Uncertainties on quark and gluon efficiency @ ~fixed 60% quark efficiency Add Run 2 experimental tracking uncertainties for use at 13 TeV 13
Overall uncertainties • ~5% overall uncertainties across a wide range of p T • Precision limited by statistics and gg(qq) contamination at low and high p T ATL-PHYS-PUB-2017-009 • Caveat: small(*) additional uncertainties for topologies other than dijet (*) to be assessed case-by-case with MC comparisons 14
First look quark/gluon tagging with ATLAS jet images
Jet images for q/g tagging • Boost and rotate jets so that η , ɸ = 0,0 • Different types of constituents: • truth-particles • charged particle tracks • topological calorimeter clusters • calorimeter towers • Build 16x16 pixel image so that the intensity (I) of a pixel is the sum of the p T of the constituents within the pixel • Normalize each image so that Σ pixels I = 1 ATL-PHYS-PUB-2017-017 16
10 5 15 Translated Azimuthal Angle ϕ 0 5 10 15 Translated Pseudorapidity η 0 Tower Constituents −0.15 −0.10 −0.05 0.00 0.05 0.10 0.15 0.20 Pearson Correlation Coefficient The CNN tagger ATL-PHYS-PUB-2017-017 ATLAS Simulation Preliminary convolutional filters Anti-k R=0.4, 150 GeV < p T < 200 GeV anti-k , R = 0.4, 400 < p /GeV < 500 t T 0.4 0.2 η Pixel Intensity Max-pooling Translated Pseudorapidity dense layer 0.2 0.15 quark jet 0 0.1 0.2 0.05 − gluon jet − 0.4 0 0.4 0.2 0 0.2 0.4 − − Translated Azimuthal Angle φ 3x • Convolutional neural network (CNN) learns non-linear representations of the input image with the goal of discriminating between quark and gluon jet images. • Similar network architecture as in Komiske et al., hep-ph/1612.01551 17
0.4 10.0 0.6 0.7 0.8 0.9 0.5 Quark Jet Efficiency 1.0 Gluon Jet Rejection 0.4 s = 13 TeV Anti-k EM+JES R=0.4 CNN Truth Particles CNN Topo Clusters CNN EM Towers CNN Topo Clusters + Tracks CNN EM Towers + Tracks 0.5 1.0 CNN EM Towers + Tracks 10.0 0.6 0.7 0.8 CNN Topo Clusters + Tracks 1.0 Quark Jet Efficiency 1.0 0.9 Gluon Jet Rejection CNN Truth Particle CNN EM Towers s = 13 TeV CNN Topo Clusters Anti-k EM+JES R=0.4 Input for jet images • Comparison of CNN tagger performance with different input images (network retrained for each input type) ATL-PHYS-PUB-2017-017 ATLAS Simulation Preliminary ATLAS Simulation Preliminary |η| < 2.1, 150 GeV < p T < 200 GeV |η| < 2.1, 400 GeV < p T < 500 GeV High p T : Low p T : - Gap wrt truth-particle image. - Calo+tracks tagger performs - Some differences from treatment of as well as truth-particle one. calo information (towers vs clusters). 18
0.4 1.0 0.6 0.7 0.8 0.5 1.0 Quark Jet Efficiency 10.0 0.4 Gluon Jet Rejection s = 13 TeV Anti-k EM+JES R=0.4 CNN EM Towers + Tracks LLH (# Charged Particles + Jet Width) # Charged Particles Jet Width 0.5 0.9 Jet Width 1.0 0.6 0.7 0.8 # Charged Particles 1.0 Quark Jet Efficiency 0.9 10.0 LLH (# Charged Particles + Jet Width) s = 13 TeV Anti-k EM+JES R=0.4 Gluon Jet Rejection CNN EM Towers + Tracks CNN tagger vs high-level observables • Comparison of CNN tagger performance wrt single observables (n track , jet width) and their combination (likelihood ratio). ATL-PHYS-PUB-2017-017 ATLAS Simulation Preliminary ATLAS Simulation Preliminary |η| < 2.1, 150 GeV < p T < 200 GeV |η| < 2.1, 400 GeV < p T < 500 GeV CNN tagger performs at At high p T , CNN tagger least as well as n track +width performs better/worse than n track +width depending on quark efficiency working point 19
0.4 Gluon Jet Rejection CNN Truth, Trained on Herwig, Test on Pythia 0.5 CNN Truth, Trained on Herwig, Test on Herwig CNN Truth, Trained on Pythia, Test on Pythia Anti-k EM+JES R=0.4 s = 13 TeV CNN Truth, Trained on Pythia, Test on Herwig 10.0 1.0 Quark Jet Efficiency 1.0 0.9 0.8 0.7 0.6 MC generators • Comparison of CNN tagger performance vs fragmentation modeling ATL-PHYS-PUB-2017-017 • Similar performance when comparing Sherpa and Pythia ATLAS Simulation Preliminary (for either training or testing) |η| < 2.1, 150 GeV < p T < 200 GeV • Moderate difference in performance between taggers trained on Pythia vs Herwig and tested on Pythia • Large difference in performance when testing on Pythia vs Herwig (irrespective of training) • Difference in fragmentation modeling plays a role, but features learned by network only partially sensitive to it. 20
Summary and outlook • Established strategy for calibrating and deriving uncertainties for q/g tagging observables using dijet events • ~5% uncertainties for tagger based on charged particle multiplicity • Begun exploration of q/g tagging using jet images and convolutional neural networks • Next steps: understand differences between input objects and what is learnt (e.g. compared with high-level observables). 21
References • Measurement of the charged-particle multiplicity inside jets from sqrt(s) = 8 TeV pp collisions with the ATLAS detector Eur. Phys. J. C76(6), 1-23 (2016) • Quark versus Gluon Jet Tagging Using Charged Particle Multiplicity with the ATLAS Detector ATL-PHYS-PUB-2017-009 • Quark versus Gluon Jet Tagging Using Jet Images with the ATLAS Detector ATL-PHYS-PUB-2017-017 22
Additional slides
n track - Run 2 tunes ATL-PHYS-PUB-2017-009 24
n track - uncertainties ATL-PHYS-PUB-2017-009 25
n track - uncertainties ATL-PHYS-PUB-2017-009 26
CNN - track-truth and η range ATL-PHYS-PUB-2017-017 27
CNN - pileup ATL-PHYS-PUB-2017-017 28
CNN - fragmentation modeling ATL-PHYS-PUB-2017-017 29
0.4 Gluon Jet Rejection CNN Truth, Trained on Herwig, Test on Pythia CNN Truth, Trained on Pythia, Test on Herwig CNN Truth, Trained on Herwig, Test on Herwig CNN Truth, Trained on Pythia, Test on Pythia Anti-k EM+JES R=0.4 0.5 s = 13 TeV 10.0 1.0 Quark Jet Efficiency 1.0 0.9 0.8 0.7 0.6 CNN - fragmentation modeling ATL-PHYS-PUB-2017-017 Komiske, Metodiev, Schwartz - JHEP 01 (2017) 110 ATLAS Simulation Preliminary |η| < 2.1, 150 GeV < p T < 200 GeV Pythia with Monash tune Pythia with A14 tune 30
CNN - average convolution differences ATL-PHYS-PUB-2017-017 31
Recommend
More recommend
