Identificação de elétrons baseada em um calorímetro de altas energias finamente segmentado

Abstract

Current applications in several areas may face scenarios involving big data, need for information fusion, high dimensionality and event-rate, rare events and signal pile-up. These features are found in the ATLAS, the largest experiment of the particle accelerator in the current scientific edge (LHC). Electron identification is vital for this purpose, currently subject to ever-increasing pile-up conditions. Detector systems provide discriminant information of distinct nature: calorimetry (energy) and magnetic spectrography (tracking etc.). The identification occurs in an online stage (trigger) followed by an offline computation responsible for the ultimate and benchmark decision. In the trigger, it is employed a decision chain of hybrid methods which counts with the first strategy based on neural networks, proposed by COPPE/UFRJ, operating in experiments similar to ATLAS. This technique (NeuralRinger) is based only on concentric rings of energy and is posteriorly complemented by a decision based on likelihood for the data fusion. This work purposes the NeuralRinger for offline operation and, thereby, inverts the standard logic of development. To allow its operation, other representations of information were added through training specialist neural networks. It was developed a complete framework for the ATLAS, a complex analysis environment. The results, comparing the NeuralRinger with the reference (likelihood), show a fake electron rate reduction from 2.54 % to 1.13 % (negligible uncertainty) when operating at the same detection rate. The NeuralRinger operation in the trigger required the evaluation of its impact in the offline environment. It was purposed a statistical analysis method, where it was observed in 2017 colision data that the distortion in minimal (< 1σ) and, nevertheless, favorable, considering that better quality samples are collected.