The SiD Particle Flow Algorithm

plantationscarfAI and Robotics

Nov 25, 2013 (3 years and 10 months ago)

74 views

The
SiD

Particle
Flow Algorithm

List of Contents


Assume Particle Flow needs no introduction


The SiD02 used in the PFA


An Overview of the algorithm


Status at the time of LOI


Introspection of the
SiD
-
PFA


The fix
-
ups and improvements


Where we are


Where to next (short and longer term)


The Detector (SiD02)

3

ECAL: 30+1 layers of (320μm Si + 2.5/5.0mm W), 3.5 x 3.5mm cells.


HCAL: 40 layers of (1.2mm RPC + 2cm steel), 1cm x 1cm cells.

Basic Building Blocks of the (Iowa) PFA



MC hits within 100 ns from IP are digitized



Photon,
Muon

and Electron ID



Track and Seed
Cluster (Directed Tree)



Building Charged
Hadron

Shower



Reconstructing Particles (four
-
vectors)

4

Hits belonging to
Photon,
Muon

and Electron

are
removed from the hit list for clustering algorithm

Next Use
DirectedTree

Clustering
for classifying the remaining hits

into sub
-
cluster types like MIPs, Clumps, Blocks and `leftover’s

Finally, start building
Hadron

Showers for one charged track at a time

Cluster Building



Extrapolate (each) track to the ECAL surface



Find Seed: sub
-
cluster directly connected to extrapolated track



Each track typically has one seed


Special cases: track without seed, or does not reach calorimeter



Now start connecting other sub
-
clusters to the seed of each track



Start with lowest and then progressively higher momentum tracks



Up to ten iterations until all track
-
cluster match satisfy (E


p) within tolerance

5

Scoring :

(a poor man’s) Probability of a link

Connecting Clusters

Based on the sub
-
cluster type and geometric proximity a score between 0 and 1 is
assigned between any two sub
-
clusters starting with the cluster in consideration

The higher the score the higher the probability of a link


A cut
-
off threshold is obtained for an energy by tuning with events

Energy dependence

6

Performance at LOI

Study just how much is contributed because of leakage

Leakage study at 500
GeV

and 1
TeV

Produce data sets a
SiD02
-
like detector
MC
with 6


HCAL


for 1
TeV
, 500
GeV
, 200
GeV



Change Steel for Cu for absorber



Increase to 54 layers from 40 layers in HCAL



1.7


more material in HCAL



No gap between HCAL and
Muon

endcap

(instead of 10 cm)


Compare sid02 with sid02
-
Cu at various energies


Check leakage by observing # hits in
Muon

detector : punch thru; a measure of leakage


Simultaneously

study the corresponding change in Energy resolution




The relative measure from the two gives an approximate semi
-
quantitative measure


of leakage
vs

performance

7

Although substantial leakage is present at 500
GeV


confusion is clearly important

Punch
-
through
muon

hits

SiD02
-
Cu


SiD02

8

Resolution study
(SiD02
-
Cu comparison)

real tracking

SiD02
-
Cu


SiD02

9

Conclusions from Leakage study

10

Although substantial leakage is present at 500
GeV
, algorithm

(confusion) has an important part



At 1
TeV

leakage comparison shows large difference in performance


between
SiD
-
nominal (dashed) and
SiD
-
Cu detectors (solid)



At 500
GeV

leakage comparison shows significant difference in performance


between
SiD
-
nominal (dashed) and
SiD
-
Cu detectors (solid)



Performance of 1
TeV

SiD
-
Cu is similar to 500
GeV

SiD
-
nominal in leakage



At 1
TeV

performance in resolution is worse with


SiD
-
nominal (dashed) and
SiD
-
Cu detectors (solid)



At 500
GeV

performance in resolution is worse with


SiD
-
nominal (dashed) and
SiD
-
Cu detectors (solid)



However : The difference of performance in resolution between 1
TeV


SiD
-
Cu and 500
GeV

SiD
-
nominal is not similar to that in leakage

A 500
GeV

qqbar

event from one side jet

Raw

MC

hits

are

displayed,

each

color

shows

an

individual

shower

Contains a low energy 12
GeV

neutral
hadron

and
several photons in the ECAL;
charged hadrons interacts

reconstructed

The same as before shown without the
isolated and unmatched hits : still no PFA
reconstruction, only with knowledge of MC

Now shown without the isolated hits but
after reconstruction,
alogorithm

of charged
hadron

track
-
cluster match (cone algorithm)

p (orange) = 119
GeV
, E/p match, enough hits (green) = 17
GeV

, algorithm

introduced a cone
-
like path in the
reclustering

to pick up secondary neutrals; but
ended up being too aggressive in stealing pieces from the low
momenta

tracks

has a low energy 12
GeV

neutral
hadron

and several photons present in the
ECAL; interaction of charged
hadron


RefinedCheatCluster

RefinedCluster

-

sharedhits

p (orange) = 119
GeV
,
E/p match, enough

hits (green) = 17
GeV

reconstructed

13

Had introduced a cone
-
like path in the
reclustering

to pick up
secondary neutrals; but ends up being too aggressive

Diagnosis of `A’
problem: an example

DCA

IP

DirAngle

PosAngle

Seed

Cluster

Interaction point

The `Cone’ Algorithm

A detailed Study


All plots show variables defined for links between a seed and a cluster. If the
seed and the cluster belong to the same truth particle, the link is quoted as
“Signal” otherwise it is quoted as “Background”

Top
-
Left Plot:

Scores just before the First cone algorithm runs.

Top
-
Middle Plot:

Scores just after the First cone algorithm runs.

Top
-
Right Plot:

Impact Parameter (IP): Distance between the center of the
seed and the straight line from the center of the cluster extrapolated along the
cluster’s direction

Bottom
-
Left Plot:

Distance of closest approach (DCA) between two straight
lines taken respectively from the center of the seed and the center of the cluster
and along the respective directions.

Bottom
-
Middle Plot:

Angle at the interaction point formed by the positions
of the seed and the cluster.

Bottom
-
Right Plot:

Angular difference between the direction of the seed
and the direction of the cluster.













Left plot:


Scores before the first cone algorithm.

Right plot:



Scores after the cone algorithm.

While Signal/Background discrimination is better after the first cone
algorithm, backgrounds now peak in the Signal region.

Score
disteribution


for links when the first cone algorithm modifies

the score.

Correlated Variables

now zoom on signal region: look at links when the first cone algorithm
gives a high score (>0.8).

Sharing of hits:

Breaking up into smaller clusters

Extending to smaller pieces

Next Steps

Allow flexibility in assignment of hits in clusters from tracks in the vicinity;

Allocate after arbitration


Check where exactly the `cone’ is needed, modify this, dump the rest


Wait for results from ongoing study here….


Faster turn around time


Improved resolution


Next major step : Incorporate the PFA with realistic
SiD

(SiD03) geometry

Now progressing in parallel

Expect to take a step backward: non
-
trivial


Improve sophisticated modifications for special types of clusters, like backscattering,
complex rare occurrences

Particle Flow



Energy of a
hadronic

jet in a calorimeter


E
jet

= E
(

+

0)
+
E
hadronic

(neglecting

’s and leakage etc)


Electromagnetic and
hadronic

components have different responses

Solutions: compensating
calorimetry
, measure
hadronic

and EM separately…..


However:



E
jet

=
E
photons

+
E
neutral
-
hadrons

+
E
charged
-
hadrons



Obtain Charged
hadron

energy (

60

) from tracking

Obtain photon/EM energy (

30

) from ECAL with 19

/

E resolution

Get neutral
hadron

energy (

10

) from E/HCAL with 67

/

E resolution


Therefore the jet energy resolution is



Ejet




Echarged




Ephotons




Eneutral

hadrons





0


19

/

(0.3 x
E
jet
)


67

/

(0.1 x
E
jet
)




20

/

E fantastic !



Particle Flow
contd


The concept depends on ability to measure particles independently


Charged and neutral particle confusion degrades resolution




Ejet




Ephotons




Eneutral

hadrons



confusion


Th
e
confusion should be minimized in a good PFA


Need excellent pattern recognition


(also high granularity and low occupancy)



sid02 :


ECAL: 30+1 layers of (320μm Si + 2.5/5.0mm W), 3.5 x 3.5mm cells.


HCAL: 40 layers of (1.2mm RPC + 2cm steel), 1cm x 1cm cells.



Categorizing:
DirectedTree

Clustering

23

Final Clustering : a flow for each track

24