Machine Learning at LHCb Mikhail Hushchyn National Research University Higher School of Economics Moscow Institute of Physics and Technology Yandex School of Data Analysis on behalf of the LHCb collaboration ICPPA 2017, 3-6 October, Moscow
MachineLearningatLHCbMikhailHushchyn
NationalResearchUniversityHigherSchoolofEconomicsMoscowInstituteofPhysicsandTechnology
Yandex SchoolofDataAnalysis
onbehalfoftheLHCb collaboration
ICPPA2017,3-6October,Moscow
TheLHCb Detector• LHCb isasingle-armforwardspectrometer.• ThemaingoalofthedetectoristosearchforindirectevidenceofnewphysicsinCPviolationandraredecaysofbeautyandcharmhadrons.
Int.J.Mod.Phys.A30(2015)1530022
2
3
WhydoweuseML?
40 MHz bunch crossing rate
450 kHzh±
400 kHzµ/µµ
150 kHze/γ
L0 Hardware Trigger : 1 MHz readout, high ET/PT signatures
Software High Level Trigger
12.5 kHz (0.6 GB/s) to storage
Partial event reconstruction, select displaced tracks/vertices and dimuons
Buffer events to disk, perform online detector calibration and alignment
Full offline-like event selection, mixture of inclusive and exclusive triggers
LHCb 2015 Trigger Diagram
TheLHCb detectorgeneratestoomuchdatatokeepitall.MachinelearningisneededforefficientselectionofthemostinterestingeventsinSoftwareHighLevelTrigger.
MLisalsousedin:• Trackpatternrecognition• Faketracksrejection• Particleidentification• Jetsidentification• DQmonitoring• Optimizeuseofstoragecapacity
TrackPatternRecognition
VELO track Downstream track
Long track
Upstream track
T track
VELOTT
T1 T2 T3
• VELOtracking:vertexreconstruction• Longtracks:usedinmajorityofanalyses(B/Ddecays)• Downstreamtracks:daughtersoflonglivedparticles• Averagetrackingefficiency>96%• Momentumresolutionvariesfrom0.5%atlowmomentumto1.0%at
200GeVInt.J.Mod.Phys.A30(2015)1530022
4
LongTracksReconstruction
StartingfromseedsintheVELO,tracksaresearchedinTstations:• SearchwindowinTstationsdefinedbyVELOtrack.• Projectx-hitintoreferenceplane.• Fit4-layer-x-cluster andremoveoutliers.• Addandfittrackwithstereohits.
TwoDeepNeuralNetworks:• Firstofthemistunedforrejectionofbad4-layer-x-clusters.• Secondoneistrainedforcandidatesselectionafterstereofit.
LHCb-PROC-2017-013
5
DownstreamTracksReconstruction
ThealgorithmisseededbytracksreconstructedinTstations.
FindmatchingTThits
Results:3- 5%improvementinfaketrackrejectionandincreaseinsignalefficiency
Rejectionofabout40%offakeT-SeedsusingBosaiBDT[JINST 8 P02013]
VELO track Downstream track
Long track
Upstream track
T track
VELOTT
T1 T2 T3
6
TMVAMLP
FakeTrackRejection
]2 [MeV/cπKm1800 1850 1900
)2ca
ndid
ates
/ (4
MeV
/c
0
500
1000
1500
2000
2500
3000
3500
LHCbpreliminary
π K→0all Dfakes
NNsaretrainedforbackgroundrejectionatgiven(97to99%)efficiency.Faketrack(ghost)probabilitybasedontheDNNoutputallowstoreducefakerate. Results:• Increasedefficiency• Reducedfakerate(22%→ 14%)
efficiency0 0.5 1
fake
trac
k re
ject
ion
0.4
0.5
0.6
0.7
0.8
0.9
1
LHCbpreliminary
ghost probability
/dof2χtrack fit
UsingDNNs
LHCb-PROC-2017-013
7
ParticleIdentification• Problem:identifyparticletypeassociatedwithatrack.• Particletypes:Ghost,Electron,Muon,Pion,Kaon,Proton.• LHCb subdetectors:RICH,ECAL,HCAL,MuonChambersandTrack
observables• Differentparticletypeshasdifferentresponsesinthesubdetectors.• Theproblemcanbeconsideredasmulticlassclassificationproblemin
machinelearning.
LHCb RICH
8
Int.J.Mod.Phys.A30(2015)1530022
ParticleIdentification• ThefirstmachinelearningalgorithmsusedforthePIDinLHCb isone-
hidden-layerneuralnetwork(TMVAMLP).• EachparticletypehasitsownbinaryNNtrainedinone-particle-vs-
restmode.
Σ# → 𝑝𝜇#𝜇& Σ# → 𝑝𝜇#𝜇&
Int.J.Mod.Phys.A30(2015)1530022
9
Plots:usingdatasidebands forbackgroundsandMonteCarlosimulation forthesignal
ParticleIdentification
• Onemodelforallparticletypes.
• ROCAUCs≈ 0.91 − 0.99 fordifferentparticletypes.
6xProbNNstrainedinone-vs-restmodeareconsideredasbaseline.
FurtherPIDperformanceimprovementisdoneusingdifferentmulticlassmodels:deepneuralnetworks(DNN)andBDTs(XGBoost andCatBoost):
Tracking System
ECAL & HCAL
RICH
Muon Chambers
!"#$%Ghost
10
LHCb Simulation,preliminary
ParticleIdentificationSeveralDBTmodelswithflatefficienciesalong𝑷, 𝑷𝒕, 𝜼 and𝒏𝑻𝒓𝒂𝒄𝒌𝒔areprovided.Themodelsaretrainedwithspeciallossfunctiondescribedin[JINST10(2015)T03002].
LHCb Simulation,preliminary
11
𝝅𝟎 − 𝜸 separationSignal:singlephoton𝛾.Background:photonsfrom𝜋= → 𝛾𝛾 decay.Problem: separatesignalandbackgroundclustersintheelectromagneticcalorimeter.
12
𝝅𝟎 − 𝜸 separationBaselinesolution:• Clustersshapeandsymetryaredescribedbysetoffeatures.• 2-layersMLPistrainedtoseparatesignalandbackgroundclusters.
𝐵= → 𝐾∗=𝛾
MLPresponse>0.6𝜀BCD~98%,𝜀HID~55%
LHCb-PUB-2015-016
13
𝝅𝟎 − 𝜸 separationNewapproach:• Responsesin5x5cellclustersforECALandpre-showerdetectorsare
consideredasnewfeatures.• SeveralNNandBDTmodelsaretrainedonthese2x25inputfeatures.• BDTmodelshowsbetterperformance.• Promisingpossibilityofaggressivebackgroundsuppressionis
demonstratedonsimulateddata.
14
JetTaggingProblem:identifybandcjetswithasmallmisidentificationprobabilityoflight-parton jets.Theidentificationof(b,c)jetsisperformedusingSVsfromthedecaysof(b,c)hadrons.
LHCb-PAPER-2015-016
15
JetTaggingTwoBDTmodelsareconsidered:DBT(𝑏𝑐|𝑢𝑑𝑠𝑔)istrainedtoseparate𝑏𝑐 and𝑙𝑖𝑔ℎ𝑡 jets,BDT(𝑏|𝑐)istrainedtoseparate𝑏 and𝑐 jets.BothBDTsaretrainedonsimulatedsamplesofb,candlight-parton jets.10kinematicobservablesofSVsareusedasinputs.
LHCb-PAPER-2015-016
W+jet events
16
JetTagging
The(b,c)-jetefficienciesversusthemistag probabilityoflight-partonjetsobtainedbyincreasingtheDBT(𝑏𝑐|𝑢𝑑𝑠𝑔)cut.
LHCb-PAPER-2015-016
17
TopologicalTrigger• ThegoalofHLT2topologicaltriggerisefficientselectionofanyB(and
D)decaywithatleast2chargeddaughters.• Itisdesignedtohandlethepossibleomissionofchildparticles.• InRun1,asimpleBDTwasusedtodefineinterestingSVs.• InRun2,thealgorithmisreoptimized usingseveralMLmodels.
LHCb-PUB-2011-016J.Phys.:Conf.Ser.664(2015)082025
18
TopologicalTriggerHLT-1trackislookingforonesuperhighPTorhighdisplacedtrack.HLT-1trackMVAclassifierislookingfortwotracksmakingavertex.HLT-2topologicalclassifierusesfullreconstructedeventtolookfor2,3,4andmoretracksmakingavertex.KinematicobservablesofSVsareusedastheclassifiersinputs.
19
J.Phys.:Conf.Ser.664(2015)082025
HLT-1track:90kHZ HLT-1trackMVA:40kHZ
TopologicalTrigger• SeveralMLmodelsareconsideredduringthetriggerreoptimization:
BDTs(MatrixNet),NeuralNetworks,LogisticRegression.• ROCcurveinaregionwithsmallFalsePositiveRateisoptimized.
20
J.Phys.:Conf.Ser.664(2015)082025
LHCb Simulation,preliminary
LHCb Simulation,preliminary
TopologicalTrigger• Mostn-bodyhadronicBdecays(n≥3)areonlytriggeredon
efficientlyinLHCb bythetopologicaltrigger.• Gain50%..80%efficiencyfordifferentchannels.
21
J.Phys.:Conf.Ser.664(2015)082025
𝜀VWX(𝑅𝑢𝑛2)/𝜀VWX(𝑅𝑢𝑛1)
JetTagging&TOPOTrigger• ThetopologicaltriggeralgorithmusesSVsthatsatisfysimilarcriteria
tothoseusedintheSV-taggeralgorithmtobuildtwo-,three- andfour-trackSVs.
• TheSVusedbytheTOPOtotriggerrecordingoftheeventcanalsobeusedtotagabjet.
• TheBDTusedintheTOPOalgorithmusessimilarinputsasjet-taggerBDTmodels.
The“loose”labelfortheTOPOreferstotheBDTrequirementusedinthetrigger forSVsthatcontainmuoncandidates.
22
LHCb-PAPER-2015-016
DQMonitoringRobo-Shifter
• Robo-shifterismachine-learningbasedsystemdesignedtoassiststheDQshifter.
• Givenrundataitcanpredictprobabilityofrunbeinggoodorbad.
• Providespotentialproblemsourcesextractedfromdecisiontrees.
• Thefirstversionofrobo-shifteriscurrentlybeingtestedbytheDQshifters.
23
MachinelearningiseverywhereatLHCb helpingtoimprovethedetectoroperationanddataprocessing.
24