towards a full particle flow algorithm (pfa) for lc detector development

Towards a Full Particle Flow Algorithm (PFA)Towards a Full Particle Flow Algorithm (PFA)for LC Detector Developmentfor LC Detector Development

Steve MagillSteve Magill

Argonne National Argonne National LaboratoryLaboratory

Status of PFA Development in the US

Two Density-based Clustering Two Density-based Clustering AlgorithmsAlgorithms

L. Xia (ANL)L. Xia (ANL)

V. Zutshi (NIU)V. Zutshi (NIU)

from Vishnu Zutshi’s talk at Bangalore :

General CommentsGeneral Comments

• Both are calorimeter first approachesBoth are calorimeter first approaches

clustering clustering track match track match fragment…. fragment….

• SiD geometrySiD geometry

Si-W ECAL, RPC or Scintillator HCALSi-W ECAL, RPC or Scintillator HCAL

‘Cheating’ involvedin some steps

Cell Density in ECALZH Events

Density-Weighted ClusteringDensity-Weighted Clustering

Cell X vs. Cell Y in ECALZH Event

Cell Y vs Cell X in ECALOnly cells with Dens.<5

Cell Y vs Cell X in ECAL5 < cellD < 25

Cell Y vs Cell X in ECALcellD > 25

One-time clustering?Optimal multi-passclustering is better

Z-pole Events WW Events

Chargedhadrons

Photons

Neutralhadrons

No. of fragments w/ and w/o cut on fragment size

ECAL

Grad-based

2.70.5 2.7 0.8

2.90.5 2.90.9

No. of fragments w/ and w/o cut on fragment size

Neutralhadrons

Photons

Chargedhadrons

HCAL

Z-pole Events WW EventsGrad-based

After track-cluster matchingAfter track-cluster matching**

Energy of matched clustersEnergy of clusters not matched to any track:neutral candidate

From neutralparticles

From neutralparticles

From chargedparticles From charged

particles(fragments)

On average ~3% came from neutral

Energy from charged particlesis more than real neutral-- need to work on it!

Dist-based

* Perfect Photon ID – hits removed

Fragment identificationFragment identification

Use the three variables to identify fragments:

1. 72% of the energy from fragments is removed2. Only lose 12% of real neutral energy

1 : 1.24 1 : 0.40Eff(neu) ~ 88%

Energy of clusters not matched to any track:neutral candidate

From neutralparticles

From chargedparticles(fragments)

After removingidentified fragments

From chargedparticles(fragments)

From neutralparticles

Dist-based

5 10 15 20 25 30

7.6

7.8

8.0

8.2

8.4

8.6

8.8

9.0

9.2

9.4

9.6

9.8

10.0

proton:#hits per GeV vs Egen

neutron:#hits per GeV vs Egen

neutron-sflin - proton-sflin

5 10 15 20 25 3020

40

60

80

100

120

140

160

180

200

220

240

260

280

300

proton:#hits vs Egen

neutron:#hits vs Egen

neutron-sf - proton-sf

5 10 15 20 25 308.0

8.2

8.4

8.6

8.8

9.0

9.2

9.4

9.6

9.8

10.0

10.2

10.4

10.6

10.8

11.0

K-:#hits per GeV vs Egen

K+:#hits per GeV vs Egen

K0L:#hits per GeV vs Egen

K0L-sflin - K+-sflin - K--sflin

5 10 15 20 25 3020

40

60

80

100

120

140

160

180

200

220

240

260

280

300

320

K-:#hits vs Egen

K+:#hits vs Egen

K0L:#hits vs Egen

K0L-sf - K+-sf - K--sf

Comparison of Charged/Neutral Hadron HitsComparison of Charged/Neutral Hadron Hits

-> linearity of response-> charged hadrons generate slightly more hits than neutral-> calibration (#hits/GeV) different, especially at low energy

Mips before showering – charged hadrons lose ~25 MeV per layer in SSRPC isolated detector. (Normal incidence)Try to correct by weighting N hits (N = # of layers traversed before interacting) by .25

R. Cassell, SLAC

5 10 15 20 25 30

7.6

7.8

8.0

8.2

8.4

8.6

8.8

9.0

9.2

9.4

9.6

9.8

10.0

proton:#hits per GeV vs Egen

neutron:#hits per GeV vs Egen

neutron-sflin - proton-sflin

5 10 15 20 25 3020

40

60

80

100

120

140

160

180

200

220

240

260

280

proton:#hits vs Egen

neutron:#hits vs Egen

neutron-sf - proton-sf

5 10 15 20 25 308.0

8.2

8.4

8.6

8.8

9.0

9.2

9.4

9.6

9.8

10.0

10.2

10.4

10.6

10.8

11.0

K-:#hits per GeV vs Egen

K+:#hits per GeV vs Egen

K0L:#hits per GeV vs Egen

K0L-sflin - K+-sflin - K--sflin

5 10 15 20 25 3020

40

60

80

100

120

140

160

180

200

220

240

260

280

300

K-:#hits vs Egen

K+:#hits vs Egen

K0L:#hits vs Egen

K0L-sf - K+-sf - K--sf

Charged(Mip correction)/Neutral Hadron HitsCharged(Mip correction)/Neutral Hadron Hits

-> account for mip trace properly-> after weighting, #hits charged ~ #hits neutral-> shower calibration (#hits/GeV) now very similar

In PFA, find mips first attached to extrapolated tracks, then can cluster remaining hits with same calibration (#hits/GeV) for charged and neutral hadrons*

* remember, this is simulation!

R. Cassell, SLAC

Nearest-Neighbor Clustering for Charged/Neutral Nearest-Neighbor Clustering for Charged/Neutral Separation – SLAC/ANLSeparation – SLAC/ANL

Photon

Pion

KL0

Piece of KL0 cluster that

wraps around pion - found by nearest-neighbor clusterer and correctly associated

0 5 10 15 20 25 30 35 40

0.86

0.88

0.90

0.92

0.94

0.96

0.98

1.00

1.02

1.04NN2255EMZZ.aida

NN2255EM.aida

gamma: Fraction of particle E in Primary cluster

0 5 10 15 20 25 30 35 40

0.86

0.88

0.90

0.92

0.94

0.96

0.98

1.00

1.02

1.04MSTdefEMZZ.aida

MSTdefEM.aida

gamma: Fraction of particle E in Primary cluster

0 5 10 15 20 25 30 35 40

0.86

0.88

0.90

0.92

0.94

0.96

0.98

1.00

1.02

1.04FCp100EMZZ.aida

FCp100EM.aida

gamma: Fraction of particle E in Primary cluster

0 5 10 15 20 25 30 35 40

0.86

0.88

0.90

0.92

0.94

0.96

0.98

1.00

1.02

1.04DTdefEMZZ.aida

DTdefEM.aida

DTdefEM.aida Entries : 19785 Mean : 1.0667 Rms : 2.1786 SumOfWeights : 30.653

DTdefEMZZ.aida Entries : 8185 Mean : 2.5231 Rms : 7.0324 OutOfRange : 14 SumOfWeights : 75.944

gamma: Fraction of particle E in Primary cluster

Energy efficiency vs generated energy

Photon FindingPhoton Finding R. Cassell, SLAC

0 5 10 15 20 25 30 35 40

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0 NN2255EMZZ.aida

NN2255EM.aida

gamma: Fraction of Primary cluster E from particle

0 5 10 15 20 25 30 35 400.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0 MSTdefEMZZ.aida

MSTdefEM.aida

gamma: Fraction of Primary cluster E from particle

0 5 10 15 20 25 30 35 400.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0 FCp100EMZZ.aida

FCp100EM.aida

gamma: Fraction of Primary cluster E from particle

0 5 10 15 20 25 30 35 400.3

0.4

0.5

0.6

0.7

0.8

0.9

1.0 DTdefEMZZ.aida

DTdefEM.aida

DTdefEM.aida Entries : 19785 Mean : 1.0667 Rms : 2.1786 SumOfWeights : 26.048

DTdefEMZZ.aida Entries : 8185 Mean : 2.5231 Rms : 7.0324 OutOfRange : 14 SumOfWeights : 50.565

gamma: Fraction of Primary cluster E from particle

Energy purity vs generated energy

R. Cassell, SLAC

1st step – Track-linked mip segments (ANL)

-> find mip hits on extrapolated tracks, determine layer of first interaction based solely on cell density (no clustering of hits)

2nd step - Photon Finder (SLAC, Kansas)-> use analytic longitudinal H-matrix fit to layer E profile with ECAL clusters as input

3rd step – Track-linked EM and HAD clusters (ANL, SLAC)-> substitute for Cal objects (mips + ECAL shower clusters + HCAL shower clusters), reconstruct linked mip segments + clusters iterated in E/p-> Analog or digital techniques in HCAL

4th step – Neutral Finder algorithm (SLAC, ANL) -> cluster remaining CAL cells, merge, cut fragments

5th step – Jet algorithm-> tracks + photons + neutral clusters used as input to jet algorithm

Track Extrapolation PFATrack Extrapolation PFA ANL, SLAC, Kansas

Shower reconstruction by track Shower reconstruction by track extrapolationextrapolation

Mip reconstruction :Extrapolate track through CAL layer-by-layerSearch for “Interaction Layer”-> Clean region for photons (ECAL)-> “special” mip clusters matched to tracks

Shower reconstruction :Cluster hits using nearest-neighbor algorithmOptimize matching, iterating in E,HCAL separately (E/p test)

ECAL HCAL

track Shower clusters

Mips one cell wide!

IL Hits in next layer

Photon Cluster Evaluation with (longitudinal) H-MatrixPhoton Cluster Evaluation with (longitudinal) H-Matrix

100 MeV

5 GeV

1 GeV

500 MeV

250 MeV

E (MeV)E (MeV) 100100 250250 500500 10001000 50005000

Effic. (%)Effic. (%) 22 6666 9494 9696 9696

1000 Photons - W/Si ECAL (4mm X 4mm)Nearest-Neighbor Cluster Algorithm candidates

E (MeV)E (MeV) 100100 250250 500500 10001000 50005000

<# hits><# hits> 99** 1212** 2020 3434 116116

Average number of hit cells in photons passing H-Matrix cut

* min of 8 cells required

PFA DemonstrationPFA Demonstration

4.2 GeV K+4.9 GeV p6.9 GeV -3.2 GeV -

6.6 GeV 1.9 GeV 1.6 GeV 3.2 GeV 0.1 GeV 0.9 GeV 0.2 GeV 0.3 GeV 0.7 GeV

8.3 GeV n2.5 GeV KL

0

_

1.9 GeV 3.7 GeV 3.0 GeV 5.5 GeV 1.0 GeV 2.4 GeV 1.3 GeV 0.8 GeV 3.3 GeV 1.5 GeV

1.9 GeV -2.4 GeV -4.0 GeV -5.9 GeV +

1.5 GeV n2.8 GeV n

_

Mip trace/IL Photon Finding

Track-mip-shower Assoc. Neutral Hadrons

Overall Performance : PFA ~33%/E central fit

PFA Module ComparisonsPFA Module Comparisons

Photon E Sum Neutral Hadron E Sum

/mean = 0.05 -> 24%/√E

G4 “feature” (fixed)

No H-Matrix (#hits)

/mean = 0.20 -> 67%/√E

Matches singleparticle fits ofKL

0, n, n mix_

E (GeV)E (GeV)

PFA ResultsPFA Results

SiD Detector ModelSi Strip TrackerW/Si ECAL, IR = 125 cm 4mm X 4mm cellsSS/RPC Digital HCAL 1cm X 1cm cells5 T B field (CAL inside)

Average confusion contribution = 1.9 GeV < Neutral hadron resolution contribution of 2.2 GeV -> PFA goal!*

2.61 GeV 86.5 GeV 59% -> 28%/√E

3.20 GeV 87.0 GeV 59% -> 34%/√E

* other 40% of events!

SiD SS/RPC - 5 T fieldPerfect PFA = 2.6 GeVPFA = 3.2 GeVAverage confusion = 1.9 GeV

SiD SS/RPC - 4 T fieldPerfect PFA = 2.3 GeVPFA = 3.3 GeVAverage confusion = 2.4 GeV

-> Better performance in larger B-field

Vary B-fieldDetector Comparisons with PFAsDetector Comparisons with PFAs

2.25 GeV 86.9 GeV 52% -> 24%/√E

3.26 GeV 87.2 GeV 56% -> 35%/√E

Detector Optimized for PFA?Detector Optimized for PFA?

3.20 GeV 87.0 GeV 59% -> 34%/√E

3.03 GeV 87.3 GeV 53% -> 33%/√E

SiD -> CDC 150ECAL IR increased from 125 cm to 150 cm6 layers of Si Strip trackingHCAL reduced by 22 cm (SS/RPC -> W/Scintillator)Magnet IR only 1 inch bigger!Moves CAL out to improve PFA performance w/o increasing magnet bore

SiD Model CDC Model

Flexible structure for PFA development based on “Hit Collections” (ANL, SLAC, Iowa)

Simulated EMCAL, HCAL Hits (SLAC)DigiSim (NIU) X-talk, Noise, Thresholds, Timing, etc.

EMCAL, HCAL Hit CollectionsTrack-Mip Match Algorithm (ANL)

Modified EMCAL, HCAL Hit CollectionsMST Cluster Algorithm (Iowa)

H-Matrix algorithm (SLAC, Kansas) -> PhotonsModified EMCAL, HCAL Hit Collections

Nearest-Neighbor Cluster Algorithm (SLAC, NIU)Track-Shower Match Algorithm (ANL) -> Tracks

Modified EMCAL, HCAL Hit CollectionsNearest-Neighbor Cluster Algorithm (SLAC, NIU)

Neutral ID Algorithm (SLAC, ANL) -> Neutral hadronsModified EMCAL, HCAL Hit Collections

Post Hit/Cluster ID (leftover hits?)

Tracks, Photons, Neutrals to jet algorithm

Optimized PFA Construction – a Collaborative EffortOptimized PFA Construction – a Collaborative Effort

PFA goal is to use the LC detector optimally –> best measurement of final-state particle properties :

LC detector becomes a precision instrument - even for jetsKey part is separation of charged and neutral hadron showers in the calorimeter – strong influence on calorimeter design

R&D priorities are :PFA development and optimizationDetector design using PFAs to optimize the calorimeter and its parameters – in particular, the design of the HCAL

Approaching PFA performance goal -> confusion < neutral hadrons

Currently, PFAs can be :Made modular to incorporate multiple cluster/analysis algorithmsUsed to optimize detector modelsTuned to optimize detector performance

SummarySummary