beam loss plane recognition - indico-fnal (indico)
TRANSCRIPT
BeamlossplanerecognitionfortheLHC
GianlucaValentinoUniversityofMalta,Msida,Malta
BelenSalvachuaCERN,Geneva,Switzerland
ThanksalsototheLHCcollimationteamforsupportinmeasurementsandanalysis
ICFAMini-Workshop:MachineLearningApplicationsforParticleAcceleratorsSLAC,MenloPark,CA,USA,1st March2018
Outline
• Backgroundandmotivation:LHCcollimationandlossmaps
• Availabledatasets
• Featureselection
• ClassificationresultsusingNNandGBC
• Conclusions
1BeamlossplanerecognitionfortheLHCGianlucaValentino
Background:LHCcollimation• TheLHCisequippedwithamulti-stagecollimationsystemtoprotectitfromnormalandabnormalbeamlosses.• Normallosses:ensurethatprotonleakagetosuperconductingmagnetsisminimal,preventingquenches
• Abnormallosses:protectionagainstfastfailurescenariossuchasasynchronousbeamdump
• Thecollimationsystemcleansparticleswithlargebetatron andoff-momentumoffsets
2BeamlossplanerecognitionfortheLHCGianlucaValentino
TCL.
5R5
TCL.5L1 TCL.
6R1
TCL.
5R1
TCL.
4R1
TCT
PV.4
L2
T
DI.4
L2
TCTP
H.4L
2
TCTPH.4R2
TCTPV.4R2
TCLIB.6R2
TCLIA.4R2
IP1
IP2
IP3
IP4
IP5
IP6
IP7
IP8
TCTPV.4L5
TCTPH
.4L5 TCDQA.A4R6
TCDQA.B4R6
TCSP.A4R6
TCSG.5L3 TCP.6L3
TCTPH.4L8 TCTPV.4L8
TCTPH
.4L1
TCTPV.4L1 TC
TPH
.4R
1 TC
TPV.
4R1
TCTPV.4R8
TDI.4R8
TCTPH.4R8
TCLIB.6L8 TCLIA.4L8
TCSP.A4L6
TCDQA.B4L6
TCDQA.A4L6
TCTP
V.4R
5 TC
TPH
.4R
5
TCSG.A4R7 TCSG.B5R7 TCSG.D5R7 TCSG.E5R7 TCSG.6R7 TCLA.A6R7 TCLA.B6R7 TCLA.C6R7 TCLA.D6R7 TCLA.A7R7
TCL.4L5
TCL.5L5
TCL.6L5
TCP.6R3 TCSG.5R3
TCSG.4L3 TCSG.A5L3 TCSG.B5L3 TCLA.A5L3 TCLA.B5L3 TCLA.6L3 TCLA.7L3
TCLA.7R3 TCLA.6R3 TCLA.B5R3 TCLA.A5R3 TCSG.B5R3 TCSG.A5R3 TCSG.4R3
TCSG.A4R7 TCSG.B4R7 TCSG.D4R7 TCSG.A5R7 TCSG.B5R7 TCSG.A6R7 TCP.B6R7 TCP.C6R7 TCP.D6R7
B1 B2
TCL.
4R5
TCL.
5R5
TCL.
6R5
TCL.6L1
TCL.5L1
TCL.4L1
CMS
ATLAS
ALICE LHC-b
TCLA.A7L7 TCLA.D6L7 TCLA.C6L7 TCLA.B6L7 TCLA.A6L7 TCSG.6L7 TCSG.E5L7 TCSG.D5L7 TCSG.B5L7 TCSG.A4L7
TCP.D6L7 TCP.C6L7 TCP.B6L7 TCSG.A6L7 TCSG.B5L7 TCSG.A5L7 TCSG.D4L7 TCSG.B4L7 TCSG.A4L7 Momentum
cleaning Betatron cleaning
Background:LHCcollimation
3BeamlossplanerecognitionfortheLHCGianlucaValentino
Adouble-sidedLHCcollimator(canbeinstalledtocleaninthehorizontal,verticalorskewplanes)
Multi-stagehalocleaningprocess
Primarycollimator(TCP) Secondarycollimator(TCSG)
Absorber(TCLA) Tertiarycollimator(TCTP)
Background:beamlosses
• Inordertomonitorbeamlosses,~3600BeamLossMonitoring(BLM)ionizationchambersareplacedaroundtheLHC.
4BeamlossplanerecognitionfortheLHCGianlucaValentino
Ionizationchambers
Background:beamlossmaps• Thebetatron cleaningsystemisqualifiedbyintentionallycreatinghighlossesinthehorizontal(H)orvertical(V)planesusingthetransversedamper.
5BeamlossplanerecognitionfortheLHCGianlucaValentino
• Cold=lossesinsuperconductingmagnets(arcs)• Warm=lossesinnormalmagnets/IRs• Collimator=lossesatBLMatcollimator
• Lossmapsaregeneratedduringtestfillswithlowintensity(fewbunches).
• Withthecollimationsystemproperlysetup,weexpecthighestlossesatthebottleneckinIR7.
~3600BLMreadingsaroundthering
IR3IR7
Background:beamlossmaps• Wecanblow-upindividualbunchesinagivenbeam&plane:
6BeamlossplanerecognitionfortheLHCGianlucaValentino
• AzoomintoIR7 givesabetterviewofthehierarchy:
Problemdefinition&motivation
• Theobjectiveistobeabletoautomaticallyclassifybetweenthefourtypesoflossplanes:• Beam1Horizontal(B1H)• Beam1Vertical(B1V)• Beam2Horizontal(B2H)• Beam2Vertical(B2V)
• Thereforeourproblemhas4outputclasses.
• Understandingthebeamlosscharacteristicsanddynamicsduringnormaloperationiscrucialtocorrectthemandunderstandtheirlong-termimpactse.g.R2Eeffects
7BeamlossplanerecognitionfortheLHCGianlucaValentino
Problemdefinition&motivation
8BeamlossplanerecognitionfortheLHCGianlucaValentino
B1H B1V
B2H B2V
Availabledatasets
• Thedatafromthe~3600LHCBLMswereextractedeachtimethetransversedamperblow-upwasrunningduring2017.
• Thisresultedin5893lossmaps,whichwerethennarroweddownbasedon:
• Beamintensityloss:theintensitylossineachlossmapshouldbe>1E8protonstohavesufficientresolution;• Collimatorsettings:thecollimatorpositionshavetobeidenticalineachlossmap;• Visualchecks:toensureacorrecthierarchywasinplace.
9BeamlossplanerecognitionfortheLHCGianlucaValentino
Availabledatasets
10BeamlossplanerecognitionfortheLHC
Beam&Plane Injection Top EnergyB1H 84 496B1V 132 599B2H 127 383B2V 123 129
GianlucaValentino
• Asthebeamparameters,collimatorsettingsetc aredifferentbetweeninjectionandflattop,separatemodelsweretrainedforbothcases:
Featureselection
• Twofeaturesetswereconsidered:
1. AlltheBLMsintheIR7longitudinalpositionrange(19400– 20600m):261BLMs.2. OnlytheBLMslocatedatcollimatorsinthesamelongitudinalpositionrange:41BLMs.
• TheBLMreadingsineachlossmapwerenormalizedtothesameBLM(TCP.A6L7.B1@19796m),whichgenerallygivesthehighestreadingsforB1lossmaps.
11BeamlossplanerecognitionfortheLHCGianlucaValentino
TrainingProcedure• Twoseparatemodelsweretrainedforthelossmaps:
• atinjectionenergy(450GeV)• attopenergy(6.5TeV)
• Lossmapsineachofthefourdatasets(B1H,B1V,B2H,B2V)weresplitusinga50:50ratiobetweentrainingandtestingdatasets.
• Theallocationofaparticularlossmaptothetrainingortestingdatasetswasdonerandomly.
• Themodelswerethenusetopredictlabelsfortheasyetunseentestingdataset,whichwerecomparedtotheoriginalgroundtruth.
• Thefinalclassificationsuccessratewascalculatedbyaveragingthepredictionperformanceonthetestingdatasetover5tries(test+train)toavoidaluckysplit.
12BeamlossplanerecognitionfortheLHCGianlucaValentino
ResultswithaNeuralNetwork
13BeamlossplanerecognitionfortheLHC
Beam&Plane Injection TopEnergy
B1H 98.1% 99.8%
B1V 99.7% 99.9%
B2H 96.9% 98.5%
B2V 97.7% 96.9%
Beam&Plane Injection TopEnergy
B1H 76.6% 99.4%
B1V 91.8% 98.5%
B2H 96.8% 99.5%
B2V 96.5% 96.6%
UsingonlythecollimatorBLMs
UsingallIR7BLMs
GianlucaValentino
4hiddenlayers:(300,130,75,30)
ResultswithGradientBoostingClassifier
14BeamlossplanerecognitionfortheLHC
Beam&Plane Injection TopEnergy
B1H 97.6% 99.0%
B1V 98.2% 99.5%
B2H 95.3% 99.5%
B2V 97.1% 96.3%
Beam&Plane Injection TopEnergy
B1H 97.6% 100%
B1V 99.1% 99.2%
B2H 95.9% 99.1%
B2V 96.5% 96.2%
UsingonlythecollimatorBLMs
UsingallIR7BLMs
GianlucaValentino
n_estimators =1000,max_depth =20
ClassificationoflossesduringLHCoperation
• Lossmapsareperformedincontrolledconditions(beamandplaneisknown),andMLmodelsweretrainedonthesedata.
• Nextstep:applyingthemodelstrainedonlossmapdatatoactuallossesinoperation.
• Twoscenariosconsidered:• Long-RangeBeam-Beam(LRBB)beamtest• LossesduringthestandardLHCmachinecycle
GianlucaValentino BeamlossplanerecognitionfortheLHC 15
Existingbeamlossdecompositionmethod
• BasedonSVD(seeM.Wyszynski,CERNsummerstudentproject&B.Salvachua etal.,“DecompositionofbeamlossesatLHC”,IPAC’17).
• Worksforbothoff-momentumandbetatron losses.
• Usesacalibrationfactor(obtainedthroughdedicatedcollimatorscrapingmeasurements)toconvertBLMreadingsinGy/stoproton/s.
• Asubsetofonly6BLMsperbeam,atHandVcollimators,isused.ThisvectoristhendecomposedasalinearcombinationoftheindividualB1H/B1V/B2H/B2Vcontributions.
• Itisstaticandnoteasilyadaptabletonewmachineconfigurations(requiresmanualselectionofBLMs).
GianlucaValentino BeamlossplanerecognitionfortheLHC 16
Long-RangeBeam-Beamtest
GianlucaValentino BeamlossplanerecognitionfortheLHC 17
• Beamtestusingwiresinstalledincollimatorstocompensatetheoctupolar termofthebeam-beaminIR5.
• TherewasaninitialB2Hblow-up,followedbyadditionallossesinB2Hasthewireswereswitchedonandoff.
• TheMLalgorithmcorrectlyclassifiedthe3spikesinlosseswhichwere~1e8p.Therewasonemisclassification(thoughherethelossesinB1andB2are~equal&~1e7p).
*classificationresult
B2HB1H
B2HB2H
Lossesduringthesqueezeinstandardoperation
GianlucaValentino BeamlossplanerecognitionfortheLHC 18
SVD
GBC
Futurework:changetomore“hierarchical”classification(i.e.considerthatwecanhavelossessimultaneouslyinB1andB2)
Conclusions• MachinelearningtechniqueswereusedtotrainmodelstoclassifybetweendifferenttypesoflossplanesintheLHC.
• TheGradientBoostingClassifiergavethebestperformance.
• UsingonlythecollimatorBLMsgavesimilarorbetterresultsthanusingallBLMs.
• Futurework:• Explorecross-validationofparametersoncemoredataisavailable• Investigatedifferentfeatureselection&scalingtechniques• moresystematicteststoclassifylossesduringthestandardmachinecycle+comparisonwithSVDtechnique.
19BeamlossplanerecognitionfortheLHCGianlucaValentino