Search for non-resonant Higgs boson pair production in the bbℓνℓν final state with the ATLAS detector in pp collisions at √s =13 TeV

(1)

Contents lists available atScienceDirect

Physics Letters B

www.elsevier.com/locate/physletb

Search for non-resonant Higgs boson pair production in the bb ν ν

ﬁnal state with the ATLAS detector in pp collisions at √

s = ^{13 TeV}

.TheATLASCollaboration

a r t i c l e i n f o a b s t ra c t

Articlehistory:

Received19August2019

Receivedinrevisedform4December2019 Accepted5December2019

Availableonline13December2019 Editor:M.Doser

Asearchfornon-resonantHiggsbosonpairproduction,aspredictedbytheStandardModel,ispresented, where one ofthe Higgs bosonsdecays via the H→^{bb channel} ând ^the ôther^via ône ôf ^the ^H→ W W^∗/Z Z^∗/τ τ channels.Theanalysisselection requireseventstohaveatleasttwob-taggedjetsand exactlytwoleptons(electronsormuons)withoppositeelectricchargeinthefinalstate.Candidateevents consistentwithHiggsbosonpairproductionareselectedusingamulti-classneuralnetworkdiscriminant.

Theanalysis uses139 fb⁻¹ ofpp collisiondata recordedatacentre-of-massenergyof13 TeVbythe ATLASdetectorattheLargeHadronCollider.Anobserved(expected)upperlimitof1.2(0.9⁺₋⁰₀^._.⁴₃) pbis set on thenon-resonant Higgsboson pairproductioncross-section at95% conﬁdence level,which is equivalentto40(29⁺₋¹⁴₉)timesthevaluepredictedintheStandardModel.

©²⁰¹⁹^TheÂuthor(s).^Published^byÊlsevier^B.V.^Thisîsânôpenâccessârticleûnder^the^CC^BY^license (http://creativecommons.org/licenses/by/4.0/).FundedbySCOAP³.

1. Introduction

In2012,theATLASandCMSCollaborationsreportedtheobser- vationofanewparticleinthesearchfortheStandardModel(SM) Higgsboson (H )[1,2]. Sofar,measurementsofthespin andcou- plingsofthe newparticleare consistentwiththose predictedby theBrout–Englert–Higgs(BEH)mechanismofthe SM[3–12]. The SM predicts non-resonant productionof Higgs boson pairs (H H ) in proton–proton(pp) collisions, referred to as non-resonantH H production,withthedominantproductionmodesattheLHCpro- ceedingviathegluon–gluonfusion(ggF)process.TheggFprocess hastwo leading order contributions: theﬁrst corresponds to the so-called‘trianglediagram’,which includesthe Higgsboson self- coupling,andthesecond istheso-called‘boxdiagram’,whichin- cludesaheavy-quarkloopwithtwofermion–fermion–Higgs( f f H ) vertices.Thesetwoamplitudesinterferedestructively,resultingin alow cross-section of only31.05±¹.90 fb forthe ggF H H pro- ductionmode,computedatnext-to-next-to-leading order(NNLO) andincludingﬁnitetop-quarkmasseffects[13–20].Feynmandia- gramsillustratingthesetwocontributionsareshowninFig.1.The measurementofnon-resonant H H productionattheLHCstandsas animportanttestoftheBEHmechanism.Inmanybeyond-the-SM (BSM)theories,H H productioncanbeenhancedbymodifyingthe Higgs boson self-coupling, λH H H, or the top-quark Yukawa coupling, yt,and/orbyintroducingnewcontactinteractions between

E-mailaddress:atlas.publications@cern.ch.

Fig. 1. FeynmandiagramsforleadingorderggFproductionofHiggsbosonpairs:the

‘trianglediagram’sensitivetotheHiggsbosonself-couplingontheleftandthe‘box diagram’ontheright.

two top-quarks or gluons and two Higgs bosons or introducing productionmechanismsviaintermediateBSMparticles[21–23].

The ATLAS and CMS Collaborations have performed searches for non-resonant H H production in a variety of ﬁnal states at 13 TeV [24–33]. No signiﬁcant excess of events beyond SM ex- pectationsisobservedinthesesearches,withtheATLAS andCMS data-analyses setting observed(expected) limitson non-resonant H H production to be no larger than 6.9 (10.0) and 22.2 (12.8) timesthepredictedrateintheSM,respectively[34,35].

This Letter describes a search for non-resonant H H produc- tion inthe bbνν final state, where refers to a lepton(either an electron ora muon), using13 TeV pp collision datacollected with the ATLAS detector during 2015–2018 and corresponding to a total integrated luminosity of 139 fb⁻¹. The analysis uses machine-learningtechniquesbasedonfeedforwardneuralnetwork architectures[36] toconstruct an event-level classifier trainedto distinguish between the H H signal and SM backgrounds. Analy- ses searching for non-resonant H H production via similar decay channelswereperformedpreviouslyinthesingle-leptonfinalstate https://doi.org/10.1016/j.physletb.2019.135145

0370-2693/©²⁰¹⁹^TheÂuthor(s).^Published^byÊlsevier^B.V.^Thisîsânôpenâccessârticleûnder^the^CC^BY^license(http://creativecommons.org/licenses/by/4.0/).Fundedby SCOAP³.

(2)

byATLASinsearchesforH H→^{bbW W}^∗ ^{[28] and}ⁱⁿ^the^dilepton channelbyCMSinsearchesforH H→^{bbW W}^∗/bb Z Z^∗ [31].

2. ATLASdetector

The ATLAS detector [37–39] is a general-purpose particle de- tectorwithforward–backwardsymmetriccylindricalgeometry.¹ It includesaninnertrackingdetector(ID),immersedinanaxialmag- netic ﬁeld,which providesprecision tracking ofchargedparticles over the rangeof |η|<2.5.Calorimeter systems with eitherliq- uidargonorscintillatortilesastheactivemediumprovideenergy measurements overthe rangeof|η_|<4.9.The muon spectrome- ter(MS)ispositioned outsidethecalorimetersandincludesthree air-coretoroidalmagnets.TheMSiscomposedofseveraltypesof muondetectorswhichprovidetriggerandhigh-precisiontracking capabilitiesfor |η|<2.4 and |η|<2.7,respectively. A hardware- based trigger followed by a software-based trigger reduce the recordedeventratetoanaverageof1 kHz[40].

3. Datasetandsimulatedevents

Thedatausedforthissearchwerecollectedin pp collisionsat theLHCwithacentre-of-massenergyof13 TeV.Onlythose data collected during stableLHC beam conditions andwith all ATLAS detectorsubsystems fullyoperationalare used,andcorrespondto an integrated luminosity of 139 fb⁻¹. The selection of candidate eventswithoppositelychargedleptons isbasedonacombination of single-lepton and dilepton triggers.² The use of a given trig- gerdepends on the flavour and the transverse momenta (pT) of thetwo(pT-ordered)leptonsintheevent,andonthedata-taking period.Single-leptontriggerswith p_T thresholds between22 and 28 GeVaregivenpriorityoverdileptontriggers.Thecriteriaofthe dileptontriggersarecheckedonlyifnosingle-leptontriggercrite- riaare metandhavepT thresholdsaslowas19(10) GeVforthe leading (subleading)lepton.At leastone reconstructed lepton(or leptonpair)hastomatchacorrespondingtriggerobject,inwhich casetheir offline pT mustbehigherthanthetriggerthresholdby atleast2 GeV,inordertobeontheefficiencyplateauofthecor- respondingtrigger.

MonteCarlo(MC) simulation[41] isused tomodel thesignal processes and in the estimation of SM background processes. A GEANT4 [42] simulation ofthe ATLAS detectorwas used forthe backgroundprocesses.ThesignalMCsampleswereprocessedwith afastsimulationthatreliesonaparameterisationofthecalorime- terresponse[41] andon GEANT4 forthetrackingdetectors.Simu- latedeventsare reconstructedusingthesamealgorithms asused fordataandincludetheeffectsofmultiple pp interactionsinthe sameor neighbouringbunch crossings, collectively referred to as pile-up.The simulation of pile-up collisions was performed with Pythia 8.186 [43] using the ATLAS A3 set of tuned parameters [44] and the NNPDF2.3LO parton distribution function (PDF) set [45].Simulatedeventswere reweightedtomatchthedistribution ofpile-up interactions indata. Theaverage amount ofpile-up in thedatacollectedduring2015–2018was33.7.

Thesignal processes withggF-initiated non-resonant H H pro- ductioninthebbννﬁnalstateweregeneratedwithaneffective

1 ATLASusesaright-handed coordinatesystemwith itsoriginat thenominal interactionpoint(IP)inthecentreofthedetectorandthez-axisalongthebeam pipe.Thex-axispointsfromtheIPtothecentreoftheLHCring,andthe y-axis pointsupwards.Cylindricalcoordinates(r,φ)areusedinthetransverseplane,φ beingtheazimuthalanglearoundthez-axis.Thepseudorapidityisdeﬁnedinterms ofthepolarangleθasη= −ln tan(θ/2).Theangulardistanceismeasuredinunits ofR≡

(η)²+ (φ)².

2 Distinctsetsofsingle-leptontriggersareusedforelectronsandmuons.Dilepton triggersrequireeithertwoelectrons,twomuons,oroneelectronandonemuon.

Lagrangianintheinﬁnitetop-quarkmassapproximation.Thegen- erated signal eventswerereweighted withformfactors thattake into account the ﬁnite mass of the top-quark [46,47]. SM back- groundprocessesweresimulatedusingdifferentMCeventgener- ators.TheMCmatrixelement(ME)eventgeneratorsandPDFsets, the parton showering (PS) and the underlying event (UE) mod- elling, UE tuned parameters (tune), andthe accuracy ofthe theoretical cross-sections usedto normalise thesimulated processes are summarised in Table 1.Each SM backgroundprocess is nor- malised to the best available respective theoretical cross-section.

ThemassoftheHiggsbosonwassetto125 GeVforallsignaland backgroundprocesses.The H H branchingfractions(BF)predicted bytheSM [13] areusedforallHiggsbosondecays. MadSpin [48]

was used to model top-quark spin correlations and EvtGen [49]

was usedtomodelpropertiesofb- andc-hadrondecays forpro- cessesusing Pythia andforthesignalprocesses.

SM top-quarkpair production (tt)¯ andthe production of sin- gletop-quarksinassociationwithW bosons(W t)contributewith signiﬁcantbackgroundcontaminationinthebbνν ﬁnalstate.At next-to-leading-order (NLO) accuracy, there exists non-trivial in- terferencebetweenthesetwoprocessesthatmaybe enhancedin phase-spaceregionswhereintherearehighfractionsofW t events [50]. Twoschemes are typically used to remove the overlap between these two processes: the so-called diagram removal (DR) and diagram subtraction (DS) schemes [51]; the former is used inthe presentanalysisto removetheoverlapping eventsandthe latterisusedtoevaluatethesystematicuncertaintyincorrespond- ing background event yields. Because of these effects, the sum of the simulated t¯t and W t processes is considered as a single background process and referred to asthe ‘Top’ process in what follows.

4. Eventselectionandobjectdeﬁnitions

Selectedeventsarerequiredtohaveatleastone pp interaction vertexreconstructedfromatleasttwoIDtrackswithp_T>0.4 GeV.

Theprimaryvertexforeacheventisdeﬁnedasthevertexwiththe highest

(pT)² ofassociatedID tracks[102].Eventsthat contain atleastonejetarisingfromnon-collisionsourcesordetectornoise arerejectedbyasetofqualitycriteria[103].

Loose and signal criteria are defined in orderto selectrecon- structed lepton and jet candidates, where the latter is a subset of the former. Compared to the loose objects, the signal objects arerequiredtosatisfytighteridentificationorqualitycriteriathat are designedtosuppressbackgroundcontributions.Reconstructed loose(signal) electrons arerequiredto satisfy the‘Loose’(‘Tight’) likelihoodidentificationcriteria[104].Looseelectronsarerequired tohave pT>10 GeV andtobewithin|η_|<2.47.Inaddition,signal electrons are required to be outside the range 1.37<|η|<

1.52,whichcorrespondstothetransitionregionsbetweenthebar- rel and endcaps of the electromagnetic calorimeters. In order to reduce background contributions fromjets misidentiﬁed as electrons,signal electronsarerequiredtobeisolated accordingtothe

‘Gradient’ selection criteria [104]. Reconstructed loose andsignal muon candidatesare requiredtohave p_T>10 GeV,to be within

|η|<2.4,andtosatisfy the‘Medium’identiﬁcationcriteria[105].

Additionally,signalmuonsarerequiredtobeisolatedaccordingto the‘FixedCutLoose’selectioncriteria[105].Signalelectron(muon) candidates are required to originate from the primary vertex by demanding that the significanceof thetransverse impact parameter, definedasthe absolutevalue ofthetracktransverse impact parameter,d₀,measuredrelativetotheprimaryvertex,dividedby its uncertainty, σd0, satisfy |^d0|/σ_d₀ <5 (3). The difference z0 betweenthevalueofthez coordinateofthepointonthetrackat which d₀ is definedandthe longitudinalpositionof theprimary

(3)

Table 1

ListoftheMEgeneratorsandPS/UEmodellingalgorithmsusedinthesimulation.AlternativegeneratorsandPS/UEmodels,usedtoestimatesystematicuncertainties,are showninparentheses.ThePDFsets,tunes,andtheperturbativeQCDhighest-orderaccuracy(leading-order,LO;next-to-leading-order,NLO;next-to-next-to-leading-order, NNLO;next-to-next-to-leading-logarithm,NNLL)usedforthenormalisationofthesamplesarealsoincluded.Thetop-quarkmassissetto172.5 GeV.

Process MEgenerator

(alternative)

ME PDF PS/UEmodel

(alternative)

UE tune Predictionorderfor totalcross-section

tt [52,53]¯ Powheg-Box v2[54,55] NNPDF3.0NLO [56] Pythia8.230 [57] A14 [58] NNLO + NNLL [59–65]

(MadGraph 5_aMC@NLO) (Herwig 7.0.4) (H7-MMHT14)

Single-tops-channel,W t [52,66,67]

Powheg-Box NNPDF3.0NLO Pythia8.230 A14 NLO + NNLL [68,69]

Single-topt-channel[52,66] Powheg-Box, MadSpin [48] NNPDF3.04fNLO Pythia8.230 A14 NLO + NNLL [70]

W,Z/γ^∗+jets [71] Sherpa2.2.1 [72,73] NNPDF3.0NNLO Sherpa2.2.1 Sherpadefault NLO(LO)≤2(4) partons [74–78]

(Z/γ^∗+^jets) (MadGraph 5_aMC@NLO) (Pythia 8.230) (A14)

Diboson(W W,W Z,Z Z ) [79]

Sherpa2.2.2 NNPDF3.0NNLO Sherpa2.2.2 Sherpadefault NLO(LO)≤1(3) partons [75–78]

tt W ,¯ t¯t Z [80] MadGraph5_aMC@NLO [81] NNPDF3.0NLO Pythia8.210 A14 NLO [82,83]

tt H [80]¯ MadGraph5_aMC@NLO NNPDF3.0NLO Pythia8.210 A14 NLO [84,85]

W H,Z H [86] Pythia8.186 [43] NNPDF2.3LO [45] Pythia8.186 A14 NNLO QCD + NLO EW [87–93]

ggFH [94] Powheg-Box v2NNLOPS [95] CT10 [96] Pythia8.212 AZNLO [97] NNNLO QCD + NLO EW [98]

SMH H→^bbνν[99] MadGraph5_aMC@NLO 2.6.2 CT10 Herwig7.0.4 [100] H7-MMHT14 [101] NNLO [14–20]

vertexisrequiredtosatisfy|^z0×^sinθ|<0.5 mm,whereθ isthe polarangleofthetrackwithrespecttothez-axis.

Jetsare reconstructed fromtopological clusters of energy de- positsinthe calorimeters[106] usingthe anti-kt algorithm[107, 108] with a radius parameter of R=⁰.4 and calibrated as described in Ref. [109]. Candidate loose jets are required to have pT>20 GeV. Signaljetsare requiredto have|η_|<2.8 andmust satisfypile-upsuppressionrequirementsbasedontheoutputofa multivariate classiﬁer [110], which identiﬁes jetsconsistent with a primary vertex in the region |η_|<2.4 and pT <120 GeV.

The MV2c10multivariate algorithm [111] isused to identify jets containing b-hadrons (b-tagged jets). An MV2c10 working point with a b-tagging eﬃciency of 70%, estimated from simulated tt¯ events[112], is used. The b-tagged jetsmust have pT>20 GeV and |η_|<2.5. The momentum of b-tagged jets is adjusted using the muon-in-jet correction, as described in Ref. [6], by ac- countingformomentumlossesduetomuonsoriginatingfromin- ﬂightsemileptonicb-hadrondecaysoccurringwithintheb-tagged jet.

The missing transverse momentum p^miss_T , the magnitude of whichis denotedby E^miss_T ,isconstructed fromthe negative vec- torial sumofthe transverse momenta ofcalibrated looseobjects intheevent.Anadditionaltermisincludedtoaccountfortheen- ergy ofID tracks that are matched to the primary vertexin the eventbutnottoanyoftheselectedlooseobjects[113].

Toavoiddouble-counting,looseobjectsaresubjecttotheover- lapremovalprocedure deﬁnedasfollows.Ifareconstructedelec- tron andmuon share a trackin theID, the electron is removed.

However, if the muon sharing the track with the electron is calorimeter-tagged,³thenthemuonisremovedinsteadoftheelec- tron.IfajetandanelectronarereconstructedwithinR=⁰.2 of eachother,thenthejetisremoved.Ifajetandamuonarewithin R=⁰.2 of eachother,andthe jethaslessthan threetracksor carrieslessthan50% ofthemuonp_T,thenthejetisremoved;oth- erwise,themuon isremoved.Electronsormuonsseparatedfrom theremainingjetsbyR<0.4 areremoved.

The analysis selects candidate events with exactly two oppo- sitelychargedsignalleptons,electronsormuons,andatleasttwo signal b-tagged jets. To enhance sensitivity to the signal process

3 Acalorimeter-taggedmuonhasonlyareconstructedtrackintheIDmatched toenergydepositsinthecalorimetercompatiblewithaminimumionisingparticle, butnocorrespondingtracksegmentintheMS.

and to maximise rejection of the expected SM backgrounds, the analysisusesamultivariateapproachtoselectsignalevents.

5. Analysisstrategy

Theanalysisreliesontheuseofamultivariatediscriminantde- signedtoselectcandidateeventsconsistentwithnon-resonantH H production.Section 5.1describesthearchitectureandthetraining ofthe deep neural network(DNN) classiﬁer fromwhich thedis- criminant is constructed. Section 5.2 describes the signal region selection criteria. Section 5.3describesthe ﬁnal backgroundesti- mationprocedure.

5.1. DeeplearningapproachtotargetH H

The discriminant uses the outputs of a DNN classiﬁer that is builtusingthe Keras librarywith Tensorﬂow asabackend[114, 115] andusesthe lwtnn library[116] tointerfacewiththeanaly- sissoftwareinfrastructureoftheATLASexperiment.Thesampleof eventsused fortraining iscomposedof equalnumbersof events fromthe signalandeach ofthe dominantbackgroundprocesses:

Top(asdeﬁnedinSection 3), Z/γ^∗_→(Z-),and Z/γ^∗_→τ τ

(Z-τ τ) production.The signal sample used inthe training ofthe classifier containsonly the H H→^{bbW W}^∗ ^component^due^toîts larger BF relative to the H H→bbτ τ and H H→bb Z Z^∗ compo- nents. However,the sumofall threesignal componentsisevalu- ated asthesignal when performing thestatisticalanalysis. Addi- tionally, all processes that make up the training sample (H H→ bbW W^∗, Top,Z-,and Z-τ τ) havethe same weightduring the trainingoftheclassifier.Thetrainingsampleiscomposedofsimu- latedcandidateeventswithm>20 GeV andhavingoneormore b-taggedjets, whereeventswithexactly oneb-tagged jetare in- cludedtoincreasethenumberofeventsavailablefortraining.For thetrainingeventswithexactlyoneb-taggedjet,eachobservable thatrequiresatleasttwob-taggedjetsissettoitsmeanvalueas computedwiththefullsetoftraining eventsthatcontainatleast two b-taggedjets. Observablesthatrequiretwo b-taggedjetsare definedusingtheleadingtwob-taggedjets.Theclassifiercontains two fullyconnectedhiddenlayers each with250nodes.Rectified linearunit(ReLU)activations areusedforeachlayer[117].Inor- dertoimprovetherobustnessofthetrainingandtoreduceeffects duetoovertraining, thereisa dropoutlayer thatrandomly drops 50% ofthe nodesbetween thetwo fullyconnected layers during training[118].Theclassifierproducesfouroutputsthatarepassed through asoftmaxactivation, constraining their sumto one[36].

Search for non-resonant Higgs boson pair production in the bbℓνℓν final state with the ATLAS detector in pp collisions at √s =13 TeV

Physics Letters B

Search for non-resonant Higgs boson pair production in the bb ν ν

ﬁnal state with the ATLAS detector in pp collisions at √

s = 13 TeV

s = ^{13 TeV}