An Ensemble of Neural Networks for Online Filtering Implemented in the ATLAS Trigger System

João Victor Da Fonseca Pinto ,Atlas Colaboration

doi:10.1088/1742-6596/1162/1/012039

João Victor Da Fonseca Pinto , Atlas Colaboration

Open Access

https://doi.org/10.1088/1742-6596/1162/1/012039

Copy DOI

Abstract

In 2017 the ATLAS experiment implemented an ensemble of neural networks (NeuralRinger algorithm) dedicated to improve the performance of filtering events containing electrons in the high-input rate online environment of the Large Hadron Collider at CERN, Geneva. The ensemble employs a concept of calorimetry rings. The training procedure and final structure of the ensemble are used to minimize fluctuations from detector response, according to the particle energy and position of incidence. A detailed study was carried out to assess profile distortions in crucial offine quantities through the usage of statistical tests and residual analysis. These details and the online performance of this algorithm during the 2017 data-taking will be presented.

Highlights

Proliferation of big data applications is an outcome of the technological breakthrough during the Digital Era
High energy physics experiments were early pioneers of dealing with the big data applications and their legacy is currently being continued by the experiments at the Large Hadron Collider (LHC) at CERN, Geneva, Switzerland
These thresholds are optimized for four predefined electron selections allowing to create the operating chains according to the desired output rate and required efficiency. i.e., the tight operating point prioritizes the purity of electron candidates collected demanding lower output rate, while the loose and the veryloose ones emphasize not losing potential electron observations required by analysis of rare physics process, resulting in higher fake rate than an equivalent tighter chain

Summary

Introduction

Proliferation of big data applications is an outcome of the technological breakthrough during the Digital Era. Electron pattern recognition relies on discriminating information of the ATLAS calorimeter system for energy measurement and its inner tracking Detector for signal patterns through particle tracks, which involves image processing-like algorithms The latter requires higher processing resources, as a way to achieve lower latency for electron triggering, early discrimination evaluates only calorimetry information. One way to achieve this, is to squeeze the beam, which results in higher number of collisions per bunch-crossing (pile-up) This generates higher pressure on the trigger system, where there is more information to be processed, requiring more bandwidth and processing resources, and the decision taking process is harder to perform once signals overlap deteriorating the distinctive patterns used for particle identification. The collection of selection chains dedicated to select physics of interest amounts to over 2000 chains, of which over 350 have electron requirements

Due to the higher occupation in the detectors of the experiment

Ensemble of Neural Networks

Results

Conclusions