JET SUBSTRUCTURE AT THE LHC & BEYOND Simone Marzani Università di Genova & INFN Sezione di Genova Interpreting the LHC Run 2 data and Beyond ICTP Trieste 27 th - 31 st May 2019 1
OUTLINE Jet substructure: where we are Machine-learning for jet physics Precision calculations in jet physics Conclusions and Open Questions 2
LOOKING INSIDE JETS the two major goals of the LHC search for new particles characterise the particles we know jets can be formed by QCD particles but also by the decay of massive particles (if they are sufficiently boosted) how can we distinguish signal jets from background ones? courtesy of G. Soyez 3
SUBSTRUCTURE IN A NUTSHELL the final energy deposition pattern is influenced by the originating splitting hard vs soft translate into 2-prong vs 1-prong structure picture is mudded by many effects (hadronisation, Underlying Event, pileup) different energy deposition pattern two-step procedure: grooming: clean the jets up by removing soft radiation tagging: identify the features of hard decays and cut on them courtesy of G. Soyez 4
A THEORIST’S JOB devise clever ways to project the multi- dimensional parameter space of final-state momenta into suitable lower dimensional (typically 1-D) distributions 250 < p /GeV < 300 GeV, 65 < mass/GeV < 95 250 < p /GeV < 300 GeV, 65 < mass/GeV < 95 250 < p /GeV < 300 GeV, 65 < mass/GeV < 95 T T T s = 13 TeV, Pythia 8 s = 13 TeV, Pythia 8 s = 13 TeV, Pythia 8 Normalized to Unity Normalized to Unity Normalized to Unity W → qq' W → qq' W → qq' QCD dijets QCD dijets QCD dijets 0.015 0.02 0.04 0.01 0.01 0.02 0.005 0 0 0 70 80 90 0 0.2 0.4 0.6 0.8 1 0 0.5 1 1.5 Jet Jet Mass [GeV] τ Δ R between subjets 21 for an introduction see SM, Soyez, Spannowsky courtesy of G. Soyez 5
PERFORMANCE & RESILIENCE truth v. parton first-principle understanding of groomers’ and taggers’ perturbative properties has reached remarkable levels p t >500 GeV optimal line 7 all studied (2) [l ⊗ l/l] ATLAS-like D 2 6 (2) [t ⊗ l/l] resilience measures a tagger’s robustness against non- CMS-like D 2 (2,dichroic) (2) [t ⊗ l/t] D 2 D 2 perturbative effects (hadronisation and UE) (2) [t ⊗ l/l] N 2 5 (2) [t ⊗ l/t] N 2 performance (2) [t ⊗ l/l] τ 21 4 (2) [t ⊗ l/t] τ 21 it is defined in terms of signal/background efficiencies (1) [t ⊗ l/t] D 2 ⌘ with/without non-pert. contributions Looking inside (2) [t ⊗ l/t] M 2 3 (2) [trim] jets M 2 (2) [l ⊗ l/l] M 2 ! � 1 / 2 2 ∆ ✏ 2 + ∆ ✏ 2 S B ⇣ = ε S =0.4 h ✏ i 2 h ✏ i 2 1 65<m<105 GeV S B Pythia8(M13) ∆ ✏ S,B = ✏ S,B � ✏ 0 anti-k t (1.0) S,B , 0 h ✏ i S,B = 1 0 1 2 3 4 5 6 7 � � ✏ S,B + ✏ 0 S,B 2 resilience 6
HARD WORK DOES PAY OFF -1 35.9 fb (13 TeV) • state-of-the art jet Events / 7 GeV 8000 • QCD and EW CMS reconstruction (anti-k t W 450 < p < 1000 GeV T Z corrections to obtain 7000 double-b tagger & particle-flow) t t passing region Z+jets and W+jets Multijet 6000 Total background • b-tagging H(b b ) 5000 Data • Higgs p T spectrum 4000 corrected for finite • soft-drop grooming top mass effects 3000 Z : 5.1 σ • 2-prong jets identified 2000 • inclusion of N 3 LO with energy H : 1.5 σ 1000 normalisation correlation function 0 N 12 t t 10 − • matching NLO-PS multijet 5 Data • decorrelation: σ 0 • state-of-the arts PDFs − N 12 → N 1,DDT2 Data 5 − 40 60 80 100 120 140 160 180 200 m (GeV) Phys.Rev.Lett. 120 (2018) no.7, 071802 SD 7
HARD WORK DOES PAY OFF 3 10 × Events / 5 GeV • QCD and EW SM Higgs ( µ = 5.8) Data ATLAS Preliminary H 25 • state-of-the art jet -1 V+Jets ( = 1.5) µ s = 13 TeV, 80.5 fb QCD Fit corrections to obtain V Signal Region QCD Fit 1 Top ± σ reconstruction (anti-k t 20 Z+jets and W+jets & topoclusters) 15 • Higgs p T spectrum 10 • b-tagging corrected for finite 5 top mass effects 3 10 × • trimming 2 t Data-QCD-t QCD and Top 1 ± σ • inclusion of N 3 LO 0 • 2-prong jets identified normalisation by requiring two track 80 3 100 120 140 160 180 200 220 10 × 1 subjets with variable R -V Signal candidate large- R jet mass [GeV] • matching NLO-PS QCD, Top and V+Jets 1 ± σ t Data-QCD-t 0 • state-of-the arts PDFs 1 − more details in 80 100 120 140 160 180 200 220 Signal candidate large- R jet mass [GeV] the backup ATLAS-CONF-2018-052 8
WHAT’S LEFT TO DO? H → bb is the holy grail of jet substructure, where it all started … embarrassingly it’s not been observed yet! Need more efficient tools? enter machine-learning Tremendous work went into understanding groomers and taggers, what’s the best use of these methods? deep thinking meets deep learning precision measurements using jet substructure 9
DEEP LEARNING a wave of machine learning algorithms has hit jet physics in the past 3/4 years ML algorithms are powerful tools for classification, can we then apply them to our task? if an algorithm can distinguish pictures of cats and dogs, can it also distinguish QCD jets from boosted-objects? number of papers trying to answer this question has recently exploded! very active and fast-developing field 10
JETS AS IMAGES jet images do what they say: project the jet into a nxn pixel image, where intensity is given by energy deposition use convolutional neural network (CNN) to classify right pre-processing is crucial for many reasons: we average over many events and Lorentz symmetry would wash away any pattern Cogan, Kagan, Strauss, Schwartzman (2015) de Olivera, Kagan, Mackey, Nachman, Schwartzman (2016) 250 < p /GeV < 260 GeV, 65 < mass/GeV < 95 T Pythia 8, W' → WZ, s = 13 TeV 3 10 ) [GeV] φ [Translated] Azimuthal Angle ( 2 10 Convolved 1 T Pixel p 10 Feature Layers Convolutions 1 250 < p /GeV < 300 GeV, 65 < mass/GeV < 95 0.5 -1 T 10 s = 13 TeV, Pythia 8 10 -2 1/(Background Efficiency) -3 0 10 150 -4 10 mass+ τ 21 -5 10 mass+ Δ R -0.5 -6 10 + R τ Δ 250 < p /GeV < 260 GeV, 65 < mass/GeV < 95 21 -7 10 T MaxOut Pythia 8, QCD dijets, s = 13 TeV -1 -8 3 10 10 100 ) [GeV] Max-Pooling φ Convnet [Translated] Azimuthal Angle ( -9 2 10 10 1 -1 -0.5 0 0.5 1 Convnet-norm W’ → WZ event T 10 Pixel p [Translated] Pseudorapidity ( η ) Random 1 0.5 Repeat Repeat -1 10 50 -2 10 -3 0 10 -4 10 -5 10 -0.5 -6 10 0 0.2 0.4 0.6 0.8 10 -7 -1 Signal Efficiency -8 10 -9 10 -1 -0.5 0 0.5 1 [Translated] Pseudorapidity ( ) η 11
BEYOND IMAGES: 4-MOMENTA analyses typically have access to more information than energy deposit in the calorimeter: e.g. particle id, tracks, clustering history in a jet, etc. build network that take 4-momenta as inputs: clever N-body phase-space parametrisation to maximise information Datta, Larkoski (2017) recurrent / recursive neural networks to model jet clustering history (using techniques borrowed from language recognition) Louppe, Cho, Cranmer (2017) Guest, Cranmer, Whiteson (2018) 12
DEEP LEARNING MEETS DEEP THINKING: LUND JET PLANE inputs of ML algorithms can be low-level (calorimeter cells/particle 4-momenta) but also higher-level variables physics intuition can lead us to construct better representations of a jet: the Lund jet plane de-cluster the jet following the hard branch and record (kt, Δ ) at each step feed this representation to a log-likelihood or a ML algorithm ln ( k t / GeV ) Primary Lund-plane regions ISR (large � ) h a Perturbative r d - c o l – – – – – – – – – l i n s e o a f r t - Non-perturbative ( c l o a l r l g i n e e z a M ) r P I / U E non-pert. (small k t ) ln ( R/ � ) 13 Dryer, Salam, Soyez (2018)
DEEP LEARNING MEETS DEEP THINKING: ENERGY FLOW NET EFN : F ( z i Φ ( θ i , ϕ i ) ) M Particles Observable ∑ Learned Filters 1 . 0 i =1 Per - Particle Representation Event Representation PFN : F ( Φ ( p i ) ) M 0 . 8 Translated Azimuthal Angle φ Latent Space ∑ Φ 0 . 6 i =1 contour 0 . 4 F Φ 0 . 2 1 . 0 . . . . . . . . . 0 . 0 0 . 8 Translated Azimuthal Angle φ 0 . 0 0 . 2 0 . 4 0 . 6 0 . 8 1 . 0 Translated Rapidity y Φ 0 . 6 0 . 4 Energy / Particle Flow Network 0 . 2 R 0 . 0 0 . 0 0 . 2 0 . 4 0 . 6 0 . 8 1 . 0 Translated Azimuthal Angle φ Observable O Function F Translated Rapidity y Map Φ R/ 2 F ( x µ ) = √ x µ x µ p µ Mass m Multiplicity M 1 F ( x ) = x F ( x µ ) = √ x µ x µ 0 p µ I track Track Mass m track overlay Track Multiplicity M track F ( x ) = x I track Jet Charge [72] Q κ ( p T , Q p κ T ) F ( x, y ) = y/x κ − R/ 2 Eventropy [74] z ln z ( p T , p T ln p T ) F ( x, y ) = y/x − ln x p D ( p T , p 2 p Komiske, Metodiev, Thaler (2018) Momentum Dispersion [93] T ) F ( x, y ) = y/x 2 T 2 x 2 [(Tr Y ) 2 − Tr Y 2 ] 3 C parameter [94] C ( | ~ p | , ~ p ⊗ ~ p/ | ~ p | ) F ( x, Y ) = − R − R − R/ 2 0 R/ 2 R Translated Rapidity y 14
Recommend
More recommend
