QCD Jets: Rise of the Machines

Transcription

1 QCD Jets: Rise of the Machines Universität Heidelberg Budapest 7/217

2 Fat jets Boosted particles at the LHC 1994 boosted W 2 jets from heavy Higgs [Seymour] 1994 boosted t 3 jets [Seymour] 26 boosted t 3 jets from resonances [Agashe, Belyaev, Krupovnickas, Perez, Virzi] 28 boosted H b b [BDRS Higgs tagger: Butterworth, Davison, Rubin, Salam] 28 boosted t 3 jets from resonances [JH/CMS tagger: Kaplan, Rehermann, Schwartz, Tweedie] 29 boosted t 3 jets in Higgs production [: TP, Salam, Spannowsky] 29 boosted t 3 jets from resonances [Template Tagger: Almeida, Lee, Perez, Sterman, Sung, Virzi] 211 N-Subjettiness [Thaler, van Tilburg] 211 Shower Deconstruction [Soper, Spannowsky] 214 image recognition W -tagger [Cogan, Kagan, Strass, Schwartzman] 217 image recognition top tagger [Kasieczka, Plehn, Russell, Schell] 217 language recognition W -tagger [Louppe, Cho, Becot, Cranmer] 2?? subjet recognition top tagger [Butter, Kasieczka, Plehn, Russel]

3 Motivation: Z t t Exmple LHC signature Z t t purely leptonic decays rate limited semi-leptonic decays with approximate reconstruction purely hadronic decays deemed not useful challenge to analysis objects

4 Motivation: Z t t Exmple LHC signature Z t t purely leptonic decays rate limited semi-leptonic decays with approximate reconstruction purely hadronic decays deemed not useful challenge to analysis objects R bjj Biggest success in sub-jet tools top tagger: hadronic identification and reconstruction hadronic decays vs QCD splittings tagging easy for higher boost, p T,t > 6 GeV SM sample: semileptonic t t events t-tagging the new b-tagging? R bjj P T [GeV] P T [GeV] 1

5 Deterministic Mass drop algorithm [TP, Salam, Spannowsky, Takeuchi] 1 C/A fat jet, R = 1.5 and p T > 2 GeV [FastJet limitation] 2 mass drop, cutoff m sub > 3 GeV 3 filtering leading to hard substructure triple 4 top mass window m 123 = [15, 2] GeV 5 A-shaped mass plane cuts as function of m W /m t 6 consistency condition p (tag) T > 2 GeV m 23 /m 123 m 23 /m m12 =m W m 23 =m W m 13 =mw m arctan m 13 /m =m W m 13 =mw m 23 =m W.2 m 23 /m m arctan m 13 /m =m W m 13 =mw m 23 =m W arctan m 13 /m 12

6 Deterministic Mass drop algorithm [TP, Salam, Spannowsky, Takeuchi] 1 C/A fat jet, R = 1.5 and p T > 2 GeV [FastJet limitation] 2 mass drop, cutoff m sub > 3 GeV 3 filtering leading to hard substructure triple 4 top mass window m 123 = [15, 2] GeV 5 A-shaped mass plane cuts as function of m W /m t 6 consistency condition p (tag) T Top reconstruction > 2 GeV direction less critical energy requiring calibration standard experimental tool [unfortunately, I m an obsolete design] Number of tops rec p/p

7 Multivariate 2 OptimalR [Kasieczka, TP, Salam, Schell, Strebler] optimal fat jet size R opt [large to include decay jets, small to avoid combinatorics] reduce R until we leave jet mass plateau m 123 m (Rmax) 123 <.2 m (Rmax) 123 R opt estimate R (calc) opt from kinematics {m 123, f W, R opt R (calc) opt } [f W = opt m ij /m 123 m W /m t ]

8 Multivariate 2 OptimalR [Kasieczka, TP, Salam, Schell, Strebler] optimal fat jet size R opt [large to include decay jets, small to avoid combinatorics] reduce R until we leave jet mass plateau m 123 m (Rmax) 123 <.2 m (Rmax) 123 R opt estimate R (calc) opt from kinematics {m 123, f W, R opt R (calc) opt } [f W = opt m ij /m 123 m W /m t ] N-Subjettiness [Thaler, van Tilburg] also include rejected events [R filt =.2, N filt = 5] {m 123, f W, R opt R (calc) opt, τ j, τ (filt) } j

9 Multivariate 2 OptimalR [Kasieczka, TP, Salam, Schell, Strebler] optimal fat jet size R opt [large to include decay jets, small to avoid combinatorics] reduce R until we leave jet mass plateau m 123 m (Rmax) 123 <.2 m (Rmax) 123 R opt estimate R (calc) opt from kinematics {m 123, f W, R opt R (calc) opt } [f W = opt m ij /m 123 m W /m t ] N-Subjettiness [Thaler, van Tilburg] also include rejected events [R filt =.2, N filt = 5] {m 123, f W, R opt R (calc) opt, τ j, τ (filt) } j Qjets [Ellis, Hornig, Roy, Krohn, Schwartz] more than one clustering history, weighted by [ ω ij = exp α y ij y min ] ij yij min then using distributions like m 2 m 2 {m 123, f W, R opt R (calc) opt, τ j, τ (filt), {m Qjets j 123 }}

10 Multivariate 2 Top resonance search [Kasieczka, TP, Salam, Schell, Strebler] Z search without ISR/FSR FSR major problem Z W b W b

11 Multivariate 2 Top resonance search Z search without ISR/FSR FSR major problem [Kasieczka, TP, Salam, Schell, Strebler] using reconstructed 4-momentum of tagged top tail in m (rec) tt add 4-momentum of fat jet: from tagged momentum and FSR {m tt, p T,t, m (filt), p (filt) jj T,j }.1 dσ σ dm tt 1 GeV Z' tt QCD.15.1 dσ σ dm ff 1 GeV Filtering: R=.3, N=5 Z' tt QCD m [GeV] tt m [GeV] ff

12 Multivariate 2 Top resonance search Z search without ISR/FSR FSR major problem [Kasieczka, TP, Salam, Schell, Strebler] using reconstructed 4-momentum of tagged top tail in m (rec) tt add 4-momentum of fat jet: from tagged momentum and FSR {m tt, p T,t, m (filt), p (filt) jj T,j } Fully multi-variate 2 accepting Spanno s SD challenge Z to two tagged tops tagging highly complex deterministic QCD-based taggers terminated 1 / ε B ED[PRD89] HTT[JHEP11] filtered fat jets (2.3) variable masses (2.4) optimalr (3.2) N-subjettiness (3.4) Qjets (3.7,.1x.1 cells) s = 14 TeV ε S

13 Applied image regognition wavelet transformation [Rentala, Shepherd, Tait; Monk] W -tagging with image recognition [Cogan etal, Oliveira etal, Baldi etal] top-tagging attempt [Almeida, Backovic, Cliche, Lee, Perelstein] QCD and shower study [Barnard etal] quark-gluon discrimination including tracks [Komiske etal] new avenue in jet physics

14 Applied image regognition wavelet transformation [Rentala, Shepherd, Tait; Monk] W -tagging with image recognition [Cogan etal, Oliveira etal, Baldi etal] top-tagging attempt [Almeida, Backovic, Cliche, Lee, Perelstein] QCD and shower study [Barnard etal] quark-gluon discrimination including tracks [Komiske etal] new avenue in jet physics Questions waiting to be answered what is the best input format? how can be use the tracker resolution? what is the training data? what does the neutal net learn? how much of this is soft activity? how does it compare to QCD tools? [judgement day is inevitable] are there better/alternative approaches? [the future has not been written]

15 Start with anti-k T fat jet [p T = GeV, R = 1.5] shift move image to center the global maximum rotation rotate the second maximum to 12 o clock flip ensure third maximum is in the right half-plane crop crop the image to 4 4 pixels decide on E vs E T for rapidities η 2 pre-processing strictly necessary?

16 Start with anti-k T fat jet [p T = GeV, R = 1.5] shift move image to center the global maximum rotation rotate the second maximum to 12 o clock flip ensure third maximum is in the right half-plane crop crop the image to 4 4 pixels decide on E vs E T for rapidities η 2 pre-processing strictly necessary? Set up network [Kasieczka, TP, Russell, Schell] run on 2-D jet images binning through calorimeter resolution [ η =.1 vs φ = 5 ] 15k events for training 12 analyze geometric patterns [CNN] 1 output: signal or background Arbitrary units Background Signal Signal probability

17 Start with anti-k T fat jet [p T = GeV, R = 1.5] shift move image to center the global maximum rotation rotate the second maximum to 12 o clock flip ensure third maximum is in the right half-plane crop crop the image to 4 4 pixels decide on E vs E T for rapidities η 2 pre-processing strictly necessary? Set up network [Kasieczka, TP, Russell, Schell] run on 2-D jet images binning through calorimeter resolution [ η =.1 vs φ = 5 ] 15k events for training analyze geometric patterns [CNN] output: signal or background Background rejection 1/ɛB nc-kernel = 1 nc-kernel = 6 nc-size = 2 nc-size = 6 nc-size = 8 nc-block = 1 nc-block = 3 nc-block = 4 nc-layer = 3 nc-layer = 4 nc-layer = 5 nd-layer = 2 nd-layer = 4 nd-node = 128 nd-node = 32 default Signal efficiency ɛs

18 Convolutional network Getting binned calorimeter images ready [nη n φ energy bins] convolutional layers to transform into optimal basis with kernel matrix W fully connected DNN weight function linking input and output n 2 y i = max, W ij x j + b i j=1

19 Convolutional network Getting binned calorimeter images ready [nη n φ energy bins] convolutional layers to transform into optimal basis with kernel matrix W fully connected DNN weight function linking input and output n 2 y i = max, W ij x j + b i j=1 4 convolutional layers with 8 feature maps [averaged S-B]

20 Convolutional network Getting binned calorimeter images ready [nη n φ energy bins] convolutional layers to transform into optimal basis with kernel matrix W fully connected DNN weight function linking input and output n 2 y i = max, W ij x j + b i j=1 4 convolutional layers with 8 feature maps [averaged S-B] 3 linear DNN layers [averaged S-B]

21 Convolutional network Getting binned calorimeter images ready [nη n φ energy bins] convolutional layers to transform into optimal basis with kernel matrix W fully connected DNN weight function linking input and output n 2 y i = max, W ij x j + b i j=1 4 convolutional layers with 8 feature maps [averaged S-B] 3 linear DNN layers [averaged S-B] Pearson input-output correlation [pixel x vs label y] ( ) images xij x ij (y ȳ) r ij = ( ) images xij x 2 ij images (y ȳ)2 not a perfect black box η pixels φ pixels Pearson correlation coefficient r

22 tagger QCD-inspired references SoftDrop [separating soft from collinear; CMS default] min(p T 1, p T 2 ) R β > z 12 SoftDrop+N-subjettiness {m sd, m fat, τ 2, τ 3, τ2 sd, τ 3 sd} MotherOfTaggers {m sd, m fat, m rec, f rec, R opt, τ 2, τ 3, τ2 sd, τ 3 sd} with and without pre-processing slight performance gain through CNN Background rejection 1/ɛB SoftDrop+N-subjettiness MotherOfTaggers full minimal Signal efficiency ɛs

23 tagger QCD-inspired references SoftDrop [separating soft from collinear; CMS default] min(p T 1, p T 2 ) R β > z 12 SoftDrop+N-subjettiness {m sd, m fat, τ 2, τ 3, τ sd 2, τ sd 3 } MotherOfTaggers {m sd, m fat, m rec, f rec, R opt, τ 2, τ 3, τ sd 2, τ sd 3 } with and without pre-processing slight performance gain through CNN Detector realities [3% signal efficiency] soft cutoff per calorimeter pixel 11 1 Background rejection Pixel ET cutoff [GeV]

24 tagger QCD-inspired references SoftDrop [separating soft from collinear; CMS default] min(p T 1, p T 2 ) R β > z 12 SoftDrop+N-subjettiness {m sd, m fat, τ 2, τ 3, τ sd 2, τ sd 3 } MotherOfTaggers {m sd, m fat, m rec, f rec, R opt, τ 2, τ 3, τ sd 2, τ sd 3 } with and without pre-processing slight performance gain through CNN Detector realities [3% signal efficiency] soft cutoff per calorimeter pixel (sub-) jet energy scale ±25% CNN surprisingly stable Relative background rejection Hardened Default Relative Jet Energy Scale

25 Checking on the Machine Typical reaction: fuck you, you fucking machine in principle, full control for fully supervised learning lots of events in the grey zone but checks possible for correctly identified signal/background events compare truth vs MotherOfTaggers vs 1- fat jet mass and N-subjettiness Signal MotherOfTaggers Background.8.4 truth mfat [GeV] Signal 3.5 MotherOfTaggers truth τ3/τ mfat [GeV] Background τ3/τ2

26 Checking on the Machine Typical reaction: fuck you, you fucking machine in principle, full control for fully supervised learning lots of events in the grey zone but checks possible for correctly identified signal/background events compare truth vs MotherOfTaggers vs 1- fat jet mass and N-subjettiness 2- soft drop mass and R opt Signal MotherOfTaggers truth Background msd [GeV] Signal msd [GeV] Background MotherOfTaggers truth Ropt Ropt

27 Checking on the Machine Typical reaction: fuck you, you fucking machine in principle, full control for fully supervised learning lots of events in the grey zone but checks possible for correctly identified signal/background events compare truth vs MotherOfTaggers vs 1- fat jet mass and N-subjettiness 2- soft drop mass and R opt 3- transverse momenta Signal MotherOfTaggers truth Background pt,j [GeV] pt,j [GeV].16 Signal.16 Background.12 MotherOfTaggers.12.8 truth pt,sd [GeV] pt,sd [GeV]

28 Checking on the Machine Typical reaction: fuck you, you fucking machine in principle, full control for fully supervised learning lots of events in the grey zone but checks possible for correctly identified signal/background events compare truth vs MotherOfTaggers vs 1- fat jet mass and N-subjettiness 2- soft drop mass and R opt 3- transverse momenta 4- so what do you want checked? HEPTop vs : just aesthetic question...

29 Learning Minkowski Back to QCD [Butter, Kasieczka, TP, Russell; see also Louppe etal, Pearkes etal] physicists are hackers, not users start with table of 4-momenta DNN layer combining momenta [CoLa] DNN layer introducing Lorentz-invariant Minkowski metric [LoLa] search for decays vs QCD

30 Learning Minkowski Back to QCD [Butter, Kasieczka, TP, Russell; see also Louppe etal, Pearkes etal] physicists are hackers, not users start with table of 4-momenta DNN layer combining momenta [CoLa] DNN layer introducing Lorentz-invariant Minkowski metric [LoLa] search for decays vs QCD first, preliminary results promising

31 Outlook Times are moving fast......deterministic taggers estalished/old/boring...information beyond clustering history helps...imagine recognition for subjets is cool... is not a black box...maybe a machine can learn constituents and Minkowski metric Background rejection 1/ɛB SoftDrop+N-subjettiness MotherOfTaggers full minimal Signal efficiency ɛs