No. 107 (00230)

Family name : Magacho da Silva
Given name : Paulo Vitor
Affiliation : Universidade Federal do Rio de Janeiro - COPPE/EE
Abbreviation : UFRJ - COPPE/EE
E-mail address : vitor@lps.ufrj.br
Title : Online Neural Trigger for Optimizing Data Acquisition
        DuringParticle 
Beam Calibration Tests with Calorimeters
Authors : Paulo Vitor Magacho da Silva, Jose Manoel de Seixas, Denis
          Oliveira 
Damazio and Bruno Carneiro Ferreira
Abstract : 
The calorimeter system plays an important role in modern high energy
experiments, providing both accurate energy measurements for the
incoming particles and fast and efficient triggering information. For
LHC, the hadronic calorimetry of ATLAS detector is performed by Tilecal,
a scintillating tile calorimeter whose construction is now being
finalized. 
For calibration purposes, a fraction of the Tilecal modules is placed in
particle beam lines at CERN.  Despite beam high quality, experimental
beam contamination is observed and this masks the actual performance of
the calorimeter. For instance, muons can be found within pion beam
samples, and both pions and muons are typically present in electron beam
selections.  Particle contamination affects the measurement efficiency
in beam periods, as data acquisition system has to acquire a significant
number of events that will be discarded by offline analysis as outsider
particles. Moreover, data size for each calibration run also increases
significantly, which makes data access rather difficult to the
geographically dispersed Tilecal colaboration.
For optimizing the calibration task in beam periods, an online neural
particle classifier is being developed for Tilecal. Envisaging a neural
trigger for incoming particles, a neural process runs integrated to the
data acqusition task and performs online training for particle
identification. After network training on a restricted number of events
(500 events suffice for feature extraction), the neural classifier
starts providing a label for each incoming event, which is classified as
electron, pion or muon. The neural label can be recorded for its usage
on the offline analysis, but it can also be used for online rejection,
preventing outsiders  from being recorded. The neural network is
supervised trained using the information of the energy deposited in each
cell of the Tilecal. Despite contamination, the training targets are
made fixed according to the beam selection. Thanks to calorimeter
granularity, the energy depostion profile reveals the class of the
incoming particle and, thus, allows outsiders to be identified in the
data sample. The neural classification performance is evaluated by
correlating the neural response to classical methodology, confirming an
ability for outsider idetification at level as high as 97%. The
classical methods used are based on energy cuts and additional
information from auxiliary detectors positioned in the beam line.
One important characteristic of the neural classifier is that it
basically copes with the event rate of the calibration runs.
Measurements showed that a rate of 0.75 kEv/s is sustainable when the
neural trigger is active, a data acquisition rate of 1.0 kEv/s is
maintained when the neural trigger is off.