An Online Calorimeter Trigger for Removing Outsiders from Particle Beam CalibrationTests

1
An Online Calorimeter Trigger for Removing
Outsiders from Particle Beam CalibrationTests
Denis O. Damazio José Manoel de Seixas Signal
Processing Lab LPS COPPE-EE damazio,seixas_at_lps.
ufrj.br
2
Outline

• Introduction
• Outsiders
• Results
• Conclusions

3
Introduction

• Attempting to search deeper in the matter, CERN
  is now preparing a new proton to proton collider,
  the LHC.
• The LHC will be colliding bunches of particles at
  14 TeV.
• For operating at LHC conditions, the ATLAS
  detector is presently being built.

4
Introduction

• The ATLAS detector relies very much on the
  calorimeter system, which comprises hadronic
  (Tilecal) and e.m. (Liquid Argon) sections.
• The Tilecal prototyping is finished and the
  detector modules are being constructed.

5
Introduction

• Tilecal is split into a central section (Barrel)
  and two lateral sections (Extended Barrels).
• Tilecal is made of iron (absorber) and
  scintillating tile (active). Detector
  segmentation comprises 3 sampling layers, which
  produce 92(Barrel)/56(EB) signals.

6
Introduction

• A fraction of the modules is calibrated using
  particle beams.
• Despite beam quality, contamination is
  unavoidable.
• pions and muons for electron beam selection.
• muons in pion beam selection

7
Introduction

• Classically, contamination (outsiders) is removed
  offline, using both calorimeter and auxiliary
  detectors information.
• In terms of beam period efficiency, it would be
  attractive to remove outsiders online (shorter
  acquisition time periods).

8
Introduction

• Neural networks may use the detailed energy
  deposition profiles furnished by Tilecal to
  accomplish this online task.
• Online training
• Neural networks
• Efficient for pattern recognition problems.
• Easy to implement digitally.
• high-speed processing

9
Introduction

• The online neural system used in the September/
  2001 testbeam was running in the Read-Out Driver
  Crate.
• Data are fetched from the Rod, normalized (by
  the total energy) and feed the NN.
• The NN response is added to the event data
  structure (as Status Word).

10
Introduction

• The neural network was a feed-forward fully
  conected network and was trained with the
  supervised backpropagation algorithm.
• Data coming from the beam line, was kept in a
  circular buffer to be used for training. New
  events substitute older ones.
• Using multithread processing, one thread trained
  the network, while the other just answered to
  incoming events. This assures fast response and
  fast training.

11
Introduction

• The methodology was divided in three steps
• Muon events are acquired to form the profile
  pattern for this particle.
• Pion events (with outsider muons) begin to be
  acquired. A network to descriminate between these
  two particle is trained.
• Electron events (with outsider pions and muons)
  begin to be acquired. Another network is trained
  to discriminate between these three particle
  types.

12
Online Results

• The online test was perfomed in the second phase
  (pion/muon) of the methodology, with 180 GeV
  pions.
• Most of the identified pions were close to their
  target (1).
• Outsiders are in the muon target (-1).

13
Online Results
The correlation between NN and energy cut shows
that the technique works, although some pion
events seem to be mixed with muons.
14
Online Results
15
Online Results
16
Online Results
17
Online Results
18
Online Results
The pions classified as muons have a muon like
profile. Only an energy reference may allow
discrimination. Thus, being independent in energy
(normalizing by the total energy) produces this
bias (pions penetrating deeply in the calorimeter
may not be detected.
19
Online Results

• Alternate approach
• Ein Ein
• SQRT( Et )
• Energy dependence is introduced to eliminate the
  bias.

20
Online Results
Is there any bias in the data now?!
21
Online Results - Barrel
Comparison between the normalization by the total
energy (left) and square root (right). 20 GeV.
22
Online Results - Barrel
Comparison between the normalization by the total
energy (left) and square root (right). 100 GeV.
23
Online Results - Barrel
Comparison between the normalization by the total
energy (left) and square root (right). 180 GeV.
24
Global Parameters StudyBarrel
Bellow is the evaluation of the parameters of the
distributions for the different methodologies.
25
Cherenkov Counter

• The Cherenkov counter can be used to help
  discriminating between electrons and pion
• This helps to validate the NN system when
  electron-pion-muon separation is of concern.

26
Electron-pion-muon separation - 20 GeV
Agreement with both energy cut and Cherenkov
counter. NN trained with data normalized by the
sum of energy (left) and square root (right).
27
Electron-pion separation
28
Preliminary Results
Electrons x Muons analysis
29
Preliminary Results
30
Preliminary Results
31
Preliminary Results
Electrons x Pions analysis
32
Preliminary Results
33
Conclusions

• An online neural network trigger was tested
  during Tilecal testbeam calibration period. The
  system was running the pion-muon discrimination.
• Due to the normalization applied (energy
  independent), the system was introducing some
  bias in the pion data. Analysis suggested the
  usage of SQRT(Et) as normalization factor to
  eliminate such bias. This introduces energy
  dependency.
• Global calorimeter performance was insensitive to
  NN cut.
• Preliminary results for electron-pion and
  electron-muon discrimination showed alternative
  ways for online trigger.
• The event rate was around 2200 events/spill,
  which meets the speed requirements. A peak on
  5502 was registered.