Документ взят из кэша поисковой машины. Адрес оригинального документа : http://acat02.sinp.msu.ru/presentations/seixas/posterandre.pdf Дата изменения: Wed Aug 7 17:11:03 2002 Дата индексирования: Mon Oct 1 20:24:46 2012 Кодировка:

Neural Par ticle Discrimination for Triggering Interesting Physics Channels with Calorimetry Data
ґ ґ Andre Rabello dos ANJOS (Andre.Rabello@ufrj.br) and Jose Manoel de SEIXAS (seixas@lps.ufrj.br) LPS/COPPE/UFRJ




Base Environment
The ATLAS experiment

















System Design (1/2)
Calorimeter Data for LVL2





System Design (2/2)
Building rings...



Results





The ATLAS experiment shall be ready to run in mid-2007. It aims to investigate the structure of matter using the stateof-the-art in particle detection. The detector is composed of many specialized sub-detectors. We draw your attention to the calorimeters of ATLAS, placed in the middle of the structure. The Calorimeters measure the energy of particles that interact with its body. The Calorimeters are again sub-divided: there is an electromagnetic and a hadronic part.

Depending on the LVL1 Trigger, the processing at LVL2 starts by analysing data provided by the calorimeters or by the Muon detectors. At the specific case of interest, the e/ trigger, the processing at LVL2 starts by confirming if the main ROI selected by LVL1 belongs to either an electron or a photon and not to a jet. Jets represent, in that particular case, the background to be eliminated. For each 25,000 "high-pt" electrons tagged by LVL1, only 1 is estimated to be a real electron. The calorimeter data that are available for analysis at LVL2 are derived from superimposed fine-grained layers composed of particle energy samplers. The particle energy is sampled as the particle itself interacts with the calorimeter cells. As the processing at LVL2 is ROI seeded, only a small fraction of the calorimeter data is retrieved for particle identification. For an e/ ROI, the region size correspond to a cone of 0.4в0.4 in the в plane (pseudo-rapidity versus angle) at the ATLAS calorimeter. Although the whole detector is barril-shaped, we straight the drawing in order to simplify the process analysis.

There is more than a 1,000 cells to be analysised per event. In order to make the neural training and processing more performant, we propose to group the cells' energy into rings centered on the peak deposition of energy of each layer. The rings to be formed won't be round in this particular case, but following the natural calorimeter segmentation. A total of 58 values is extracted from the calorimeter data using this strategy. This form of compactation looks reasonable when one takes into consideration the isotropic way particles interact with the calorimeter in an RoI.
layer 1 layer 2 layer 3

The 58 inputs were fed into a neural network for training. The training method selected was backpropagation. The stop criteria is based on the neural network efficiency instead of the MSE. The dataset was composed of 3600 jets and 600 electrons. All elements were approved by the LVL1 trigger strategy. The set was divided into two, for training and testing the neural discriminator performance.
Characteristic Curves 1

0.98

0.96

0.94 Electron Efficiency

0.92

0.9

0.88

0.86

0.84

EM Object

Decaying object

0.82

test 17 Neural(4-quant) Linear(4-quant)
0 1 2 3 4 5 6 Jet background (kHz) 7 8 9 10

0.8

Detector

Illustration of an object iteration with the calorimeters, as a shower of par ticles is produced.
PreSampler Second EM Layer

The ATLAS detector.

The ATLAS Trigger
1

The experiment uses an event trigger to guide data taking. The trigger is composed of the three levels connected one after the other, applying slower, but more complex event analysis to the incomming candidates. The First Level Trigger (LVL1), primarily composed of programable devices like FPGA's, looks into the calorimeters and the muon detectors searching for interesting, but isolated channels that might indicate an interesting event. The Second Level Trigger (LVL2) is the first where the generated event is regarded as a whole. The LVL2 retrieves event data for analysis based on small subsets of the detector (Regions of Interest - ROIs). The Third Level Trigger (LVL3) retrieves the whole event data in order to apply its event selection strategy. Both LVL2 and LVL3 are programmed on commodity computers interconnected by fast ethernet-like switches.
Interaction rate aprox. 1 GHz Bunch crossing rate: 40 MHz LEVEL 1 TRIGGER < 75 (100) kHz Derandomizers Regions of Interest LEVEL 2 TRIGGER aprox. 1 kHz Event Builder Full-event buffers and processor sub-farms Readout drivers (RODs) Readout Buffers (ROBs) Pipeline memories

1 2 2 2 3 4 5 6 9876543 345678 7 8 9

Third EM Layer

First Hadronic Layer

The Region of Convergence (RoC) for several Electron/Jet discriminators based on the same data set. The neural discriminator described in this work is marked with the label "Test 17". The other lines refer to a classical (bidimensional cutting) and a neural discriminator based on quantities extracted by the algorithm being applied nowadays at CERN, for the same purpose. The "winner" neural discriminator had 5 neurons at its hidden layer. Based on the relevance of each ring to the discrimination, one can cut down the number of rings used for this task and try to understand the redundancy in the process. This allows faster discrimination and controlled redundancy insertion to the discrimination task. The relevance to the discrimination is calculated as follows.
1 Ri = N
N j =1

[output(- ) - output(- |xj,i=xi )]2 xj xj

1 2 3 2 3 4 5678

1

Comparison between Neural Discriminators Efficiencies

2

1

CALO

MUON

TRACKING

The ring structure to be extracted from the calorimeter layers. These represent examples. The red cells represent the cell where the peak energy deposition was detected by the compactation algorithm.

Electron Efficiency 0.9

A typical electron interacting with the ATLAS calorimeters (simulation). The layers were put side-byside in this printing. In reality they are stacked from topbottom, left-to-right.

4 98765

3
0.95

58-5 20-5 10-5 8-3 5-3 4-Class. Quantities (neural)
0.85 0 1 2 3 4 Jet Background Rate 5 6 7

EVENT FILTER aprox. 100 Hz

Data recording

A simplified sketch of the ATLAS Trigger.

The RoC for several discriminators based on smaller number of rings extracted from the e/ ROI, based on their relevances to discrimination. Note that still with only 5 (out of 58) rings, we can still be more efficient then the algorithm used at CERN for the same task. The legend also shows the number of input neurons and the number of hidden neurons on that par ticular test.