CMS Upgrade Workshop - Fermilab19.11.08

Trigger Studies Using Stacked Pixel Layers

Mark Pesaresihttps://twiki.cern.ch/twiki/bin/view/Main/MarkPesaresi

Mark Pesaresi2

Tracking Trigger

Aim is to assess the performance and viability of a stacked pixel layer or layers as part of a L1 tracking trigger by the determination of track pt

Study attempts to simulate the implementation of such a trigger

Generation of trigger primitives using digi information

Performance of the algorithm in finding high pt tracks

Investigate methods of sensor readout and hit correlation for the on-detector implementation

Performance of high pt track reconstruction for a trigger when using two or more stacked layers

Complements previous study reported on last December using a stacked strip layer in the outer tracker

Mark Pesaresi3

Geometry

0.92.142.5

Current Pixel System

Stacked Pixel Layer @ 25cm

Considering a single stacked pixel layer at r=25cm, length=221cm

Current pixel system included in geometry

Outer geometry unnecessary at this point

Using latest version of Strawman B in CMSSW_1_8_4

Mark Pesaresi4

Sensor Geometry

Strawman B parameters modified in pixbar.xml and trackerStructureTopology.xml

Sensor choice: tilted at 23° – to reduce cluster width by minimizing Lorentz drift

100μm thickness

28mm x 72.8cm sensor dimensions

z overlap – to fill gaps in z

100 μm x 2.37mm pixel pitch

256 x 30 pixels per module

Sensor separation varied between 0.5-4mm

Modification made to geometry to aid trigger studies – not yet part of StrawmanB

z offset – to match columns in top and bottom sensors with increasing eta

23°

z overlap

z offset

Mark Pesaresi5

Simulation Overview

adc cut & sorting

Sorted Digis[detId, row, column, adc]

correlation algorithm

Stubs[detIdhigh, rowhigh, columnhigh, adctot, row difference, column difference, simTrackIdhigh, simTrackIdlow]

Stacked Layer Digis[detId, row, column, adc]

adcdigi > 30

sorted by detId into modules with upper and lower sensors

hits between upper and lower sensors are correlated to check for high pt tracks

modifiable search window cuts can be applied

Mark Pesaresi6

Correlation Algorithm

Row difference calculation

Since the sensors are tilted, there is a difference between the position of the higher and lower sensor hits for a high pt track which is also dependent on the position of the incident track on the sensor

The fixed offset as a function of the row number can be applied to calculate the true row difference

Equivalent to an on detector map between the hit position on the higher sensor to a set of positions on the lower sensor

Column difference calculation

Column difference is not symmetrical – dependence on whether hit is in detector +/-z. Can be exploited to maximise rate reduction.

0

256

pixel row 114

pixel row 125

Mark Pesaresi7

Correlation Algorithm

Stub generation

A stub is created when both the row and column difference lie within a given range.

e.g. row offset = 30 ≤ row window ≤ +10 ≤ column window ≤ +1

Upper

Lower

Pass Fail

100μm100μm

Mark Pesaresi8

Algorithm Performance

Separation [mm] Max Efficiency [%] Fake [%] (or average

number/event)

Reduction Factor

0.5 99.05 0.73 (12.22) 8.04

1.0 99.35 4.14 (25.58) 22.26

2.0 97.745 17.83 (18.74) 95.99

3.0 96.00 39.08 (23.76) 210.28

4.0 92.95 47.27 (32.39) 254.35

Max Efficiency: Average maximum efficiency for a high pt track to form a stub. Inefficiencies due to sensor doublet acceptances and algorithmic efficiency (window cuts)

Fake: Average fraction of stubs per event generated by correlating hits from different tracks

Reduction Factor: Average data rate reduction factor per event (NStubs / NDigis) where NDigis is number of hits with charge >adcdigi for the whole stacked layer

Performance of a detector stack at r=25cm for sensors with pitch 100μmx2.37mm. Correlation cuts optimised for high efficiency

Mark Pesaresi9

Algorithm Performance

Separation [mm] Max Efficiency [%] Fake [%] (or average

number/event)

Reduction Factor

0.5 99.05 0.73 (12.22) 8.04

1.0 99.35 4.14 (25.58) 22.26

2.0 97.745 17.83 (18.74) 95.99

3.0 96.00 39.08 (23.76) 210.28

4.0 92.95 47.27 (32.39) 254.35

Performance of a detector stack at r=25cm for sensors with pitch 100μmx2.37mm. Correlation cuts optimised for high efficiency

Max Efficiency calculated using 20,000 single 50GeV Muon/Antimuon events with smearing

Fake/Reduction Factor calculated using 100 MinBias events with an average of 400 interactions per bunch crossing with smearing

Results optimised for high efficiency: Row window = 2 pixelsColumn window = 2 pixels @ 0.5mm

3 pixels @ 1mm, 2mm 4 pixels @ 3mm6 pixels @ 4mm

Mark Pesaresi10

Algorithm Performance

Cuts optimised for high efficiency:Row window = 2 pixelsColumn window = 2 pixels @ 0.5mm; 3 pixels @ 1mm, 2mm;

4 pixels @ 3mm; 6 pixels @ 4mm

pT discriminating performance of a stacked layer at r=25cm for various sensor separations using 10,000 di-muon events with smearing

Sensor separation is again an effective cut on pt – as with the stacked strips

Again, the width of the transition region increases with separation. Due to:

- pixel pitch- sensor thickness- charge sharing- track impact point

Efficiencies decrease with sensor separation due to the larger column window cuts – sensor acceptances and fake containment are issues

Mark Pesaresi11

Implications

In order to reduce Lorentz drift, sensors have been tilted – correlation requires that an offset must be programmed in order to match hits from high p t tracks

- At its most basic, a calibration constant must be uploaded for each pixel row on a sensor

- If technology changes, sensors may not need to be tilted

Instead of the correlator performing a difference analysis on two hits, a programmable map between an address on the upper sensor and multiple addresses on the lower sensor would simplify implementation and reduce rate & fakes. Is this possible?

If layer is part of a multi-stack detector, a high efficiency is preferable to large rate reductions. We only need to remove the majority of low p t tracks. Multiple stacks should remove the fakes if combinatorics are not too high. A 2mm separation at 25cm seems reasonable.

To maintain high efficiencies, the column window cut must be kept wide. Can such a column window cut be implemented on detector?

Mark Pesaresi12

Next Steps

Plenty of work to do still… e.g.

Measure performance of a stacked layer as a function of radius

Measure performance of a stacked layer as a function of pileup

Measure performance of a stacked layer as a function of pixel pitch

Check performance is maintained in different physics events, e.g. jets

Mark Pesaresi13

Double Stack Geometry

0.9

2.142.5

Current Pixel System

Stacked Pixel Layer @ 25cm

Considering now two stacked pixel layers at: r=25cm, length=221cmr=35cm, length=221cm

Current pixel system included in geometry

Outer geometry unnecessary at this point

Using same sensor geometry for each layer

Stacked Pixel Layer @ 35cm 1.8

Mark Pesaresi14

Stack Performance

Cuts optimised for high efficiency:Row window = 2 pixels @ 25cm layer, 3 pixels @ 35cm layerColumn window = 3 pixels @ 2mm; 2 pixels @ 1.45mm

pT discriminating performance of stacked layers at r=25cm and r=35cm for various sensor separations using 10,000 di-muon events with smearing

At a larger radius, a stacked layer with the same sensor separation will effectively cut at a larger pt

Calculations indicate that a sensor separation of 1.45mm at 35cm will produce the same discrimination profile as a 2mm separation at 25cm - simulations agree with this calculation

Having to build modules with different sensor separations may be undesirable for a future tracker

Mark Pesaresi15

Stack Performance

Separation [mm] Max Efficiency [%] Fake [%] (or average

number/event)

Reduction Factor

2.0 (Upper Stack) 99.09 17.86 (10.91) 203.70

2.0 (Lower Stack) 97.745 17.83 (18.74) 95.99

Performance of detector stacks at r=25cm and r=35cm for sensors with pitch 100μmx2.37mm. Correlation cuts optimised for high efficiency

Max Efficiency calculated using 20,000 single 50GeV Muon/Antimuon events with smearing

Fake/Reduction Factor calculated using 100 MinBias events with an average of 400 interactions per bunch crossing with smearing

Results optimised for high efficiency:Row window = 2 pixels @ 25cm layer, 3 pixels @ 35cm layerColumn window = 2 pixels

To simplify the simulations slightly, sensor separations of 2mm have been chosen for both layers

Mark Pesaresi16

Double Stack Simulation Overview

sorting/clustering

correlation algorithm

Tracklets

stubs on each layer can be clustered or processed to remove duplicates

in this case only 1 hit/column is passed – similar to how on detector correlation might work

stubs between upper and lower layers are correlated in eta and phi – performance is trigger hardware dependent

modifiable search window cuts can be applied

Stubs (Lower Stack)[detId, row, column]

Sorted Stubs (Lower Stack)[global coordinates]

Sorted Stubs (Upper Stack)[global coordinates]

Stubs (Lower Stack)[detId, row, column]

Mark Pesaresi17

Double Stack Correlation Algorithm

Correlate stubs in upper sensor with stubs in lower sensor – use upper sensor as seed (fewer stubs, fewer fakes)

Stubs

Upper Stack

Lower Stack

Vertex

Window cut in η applied – wide enough to allow for vertex smearing

Window cut in ϕ applied – wide enough to allow for low pt tracks and scattering

Mark Pesaresi18

Double Stack Correlation Algorithm

Window cut in η applied – wide enough to allow for vertex smearing

Window cut in ϕ applied – wide enough to allow for low pt tracks and scattering

Distributions of Δη and Δϕ between upper and lower stack stubs using 10,000 single 5-50GeV Muon/Antimuon events with smearing

Mark Pesaresi19

Double Stack Algorithm Performance

Using double stack correlation window cuts |Δη| < 0.2, |Δϕ| < 0.015

Tracklet pt resolution vs. track pt and η when using a 3-point pt reconstruction measurement for 10,000 0-30GeV di-muon events with smearing

If the stubs are correlated, we can use the two stubs plus the vertex as r,ϕ points for a 3-point track pt measurement – assumes track originates from (0,0)

Mark Pesaresi20

Double Stack Algorithm Performance

Using double stack correlation window cuts |Δη| < 0.2, |Δϕ| < 0.015

pT discriminating performance using double stacks for 10,000 0-30GeV di-muon events with smearing

With a momentum measurement using two stacks, an effective cut on track pt can be placed

Maximum efficiency is still determined by that of the single stack

A better track pt resolution using the double stack means that the transition region can be reduced

We would like to have better efficiencies at low pt – this would require stacks with smaller sensor separations (or larger windows) increasing the number of stubs per layer and the number of combinatorics for the double stack algorithm

Mark Pesaresi21

Next Steps

Investigate performance at high pileup – measure number of combinatorial fakes, pt resolution, robustness to displaced vertices / secondary interactions

Measure vertex resolution, angular/z resolution at calorimeters

Measure performance of two stacks as a function of radius separation

Measure performance of two stacks as a function of pileup

Check performance is maintained in different physics events

Eric Brownson (Vanderbilt) is currently working on replicating the TDR L1 muon trigger rate plot at LHC luminosity.

plan is to work in collaboration to get the corresponding plot at SLHC luminosity

then use the tracking trigger developed here to measure the effect of combining such a trigger with the L1 muon trigger

Mark Pesaresi22

Summary

Strawman B has been used as the basis to commence trigger studies using a stacked pixel layer at 25cm

Algorithm to correlate digi hits from high pt tracks has been writtenPerformance of algorithm in ideal conditions measured - >95% maximum efficiency of detecting

high pt tracks, ~ x100 reduction in data rate

Stubs from two stacked layers have been correlatedExample layers at 25cm and 35cm demonstrate that high pt tracks can be detected with >90%

maximum efficiencypt of single muons has been measured with ~4% resolutionStill need to check number of combinatorial fakes, robustness of p t measurement in high pileup

events

Still plenty to investigate…Effect of occupancy on performanceEffect of changing layer radiiEffect of changing pixel pitch, short/long pixel stripsPossibility to extend layers to high eta…

Mark Pesaresi23

Backup – Single Stack Performance

Effect of changing window cuts on discrimination curves

Efficiencies are unchanged with larger column windows

Efficiencies are recovered (at larger separations) when row window is increased but also has the effect of decreasing the pt cut

pT discriminating performance of a stacked layer at r=25cm for a sensor separation of 4mm and various algorithm cuts using 10,000 di-muon events with smearing

Mark Pesaresi24

Backup – Single Stack Performance

1 2 3

1 19.05 41.96 42.085

2 44.075 95.585 95.89

3 45.155 97.745 98.07

RowWidth

ColumnWidth

Efficiency of a 50 GeV muon/antimuon generating a stub in the stacked layer [%]

Data rate reduction factor achieved on MinBias events at SLHC pileup

100 MinBias events with an average of 400 interactions per bunch crossing with smearing

20,000 single 50GeV Muon/Antimuon events with smearing

2 3

2 104.6 94.4

3 96.4 86.0

RowWidth

ColumnWidth

Choosing a sensor separation of 2mm, the effect of the window cuts has been determined

Mark Pesaresi25

Backup - Some Numbers

Typical MinBias event at SLHC luminosity:

1455 tracks > 2 GeV

4 tracks > 8 GeV

(in region |eta| < 2.14)

Using a stacked pixel layer at 25cm (|eta| < 2.14) with pixel pitch 100μmx2.37mm and 2mm sensor separation [row window=2, column window=3]

140 stubs • includes 25 fake stubs• includes 20 duplicate stubs

Every event is triggered

A second stacked layer would reduce the number of fakes, the number of tracks (if pt threshold is raised) and allow sufficient resolution for matching to other sub-detectors.

Mark Pesaresi26

Backup - Sensor Readout

A method for reading out stacked sensors for hit correlation is required

- Readout and decision every bunch crossing

- Low power

G.Hall – July 2008

Mark Pesaresi27

Backup - Sensor Readout

Module divided into 64 blocks of 4 rows per column

Requires minimum 10 address lines:

6 bit block address4 bit patterne.g. x000,00x0,0xx0, etc.

Assumes that only 1 block is hit per column – reasonable since <1 pixel hit per column on average

4 x 100μm

2.37mm

Correlator Correlator

Block1

Block0

Mark Pesaresi28

Backup - Sensor Readout

Analysis modified to simulate this method of correlation

Sort data into blocks

Correlate hit blocks

Readout block stubs and

pattern data

Run original

algorithm

Correlate blocks with pattern data

Readout stubs

On-detector Off-detector

Blocks are correlated in a similar way to before with a block (row) difference and a column difference. As before, an offset is required to match the blocks correctly

Cuts can be placed on the window width for both blocks and columns

Investigated how well top method worked and the data rate reductions possible

Mark Pesaresi29

1 2 3

1 49.38 59.64 60.03

2 78.72 95.03 95.59

3 80..57 97.18 97.75

BlockWidth

ColumnWidth

Efficiency of a 50 GeV muon/antimuon generating a stub in the stacked layer [%]

Data rate reduction factor achieved on MinBias events at SLHC pileup

100 MinBias events with an average of 400 interactions per bunch crossing with smearing

20,000 single 50GeV Muon/Antimuon events with smearing

2 3

2 9.20 4.78

3 8.37 4.47

BlockWidth

ColumnWidth

Choosing a sensor separation of 2mm, the effect of block cuts have been determined

Results are for block correlation followed by standard algorithm with [row window=2, column window=3]

Backup - Sensor Readout

Mark Pesaresi30

Backup – Sensor Readout

Require at least a factor 10 reduction in rate to read out detector. Achievable with a block width cut of 2.

For reasonable efficiencies, a column width cut of at least 2 is still required. How can this be performed easily on detector?

Offsets are still needed when applying correlation to blocks – can this be implemented on detector?

A small fraction of columns contain more than one hit per BX (in some cases up to 6 hits). Is this important, can it be reduced or ignored?

Largest cause is due to hits on block boundariese.g. |0000|000x|x000|0000|

Mark Pesaresi31

Backup

Fast Sim gives an average occupancy of 0.05% (up to 0.15% instantaneous) at an average of 400 interactions per event for a layer at 25cm extending to |eta| < 2.14

Assume Full Sim will give x3 occupancy0.15% x 17,448,960 channels = 26,173 hit channels

30k channels require 2813, 2.56Gbps links assuming 12bit address per channel at 20MHzLink Power: 5.6 kW (322uW/ch) – Geoff suggests a budget of 300μW/ch for the pt layers

Cutting the number of channels to readout by x10 using hit correlation brings link power for the pt layers down to reasonable values

Total HitsBlock StubsPixel Stubs

Mark Pesaresi32

Backup

Layer Occupancy(No. digi hits in layer /

total channels in layer)

Module Occupancy(No. digi hits in occupied module /

total channels in module)

Note: Full Sim occupancies estimated at 3x these values

100 MinBias events, ~400 interactions per bx with smearing