Doctoral Thesis Defense Session No. 54 of PPGEEC - COMPUTATIONAL INTELLIGENCE APPLIED FOR CROSSTALK REDUCTION IN A LIQUID ARGON CALORIMETER

STUDENT: MARTON SANDES DOS SANTOS

DATE: 12/11/2024

TIME: 2:00 PM

LOCATION: Congregation Room - Polytechnic School of UFBA, 5th floor

TITLE: COMPUTATIONAL INTELLIGENCE APPLIED FOR CROSSTALK REDUCTION IN A LIQUID ARGON CALORIMETER

KEYWORDS: ATLAS; Analytical Simulator; Crosstalk; Statistical Signal Processing; Machine Learning; Loernzetti Framework;

ABSTRACT: ATLAS (A Toroidal LHC ApparatuS) is the largest experiment at the Large Hadron Collider (LHC) accelerator located at CERN (Centre Européen pour la Recherche Nucléaire). ATLAS is located at one of the collision points in the accelerator tunnel. It consists of specialized detectors to characterize the particles produced by proton-proton (pp) collisions with a center-of-mass energy of 13 TeV. One of its specialized detectors is the Liquid Argon (LAr) calorimeter, with about 187,000 sensor cells to record the production of electromagnetic particles. Calorimeters are detectors widely used in particle physics experiments, which have the function of recording and absorbing the energy of the particles produced in the collisions inside the experiments. The ATLAS LAr calorimeter has a fine granularity and high cell density, which, in association with the high collision rates and the mechanical and electronic structure of the detector readout system, produce interference effects from neighboring electronic channels (crosstalk – XT). XT increases the difficulty of estimating the energy and propagation time of the incident particle. The Optimal Filter is the standard method used in ATLAS to estimate the energy and time, but in its coefficient design, XT is not taken into account, which brings an error in the estimation of the energy and flight time of the particles. This thesis investigated possible solutions for the mitigation of XT using machine learning techniques and statistical signal processing methods. The proposed solution is based on the development of an estimator based on artificial neural networks to mitigate the undesirable effects of XT. Two simulated databases based on electromagnetic particle physics models were used in the application and evaluation of the model. The signals obtained from the developed simulators are used to develop and evaluate possible machine learning solutions. Using the data produced in the available simulators, the approach using supervised methods proved to be efficient for both estimating the energy and flight time of particles using the Multilayer Perceptron (MLP) structure in a regression task of the energy value and time of particles without the influence of XT. The results obtained produced regression performance with low RMSE (root mean squared error) in relation to the standard method used to estimate the energy and flight time of particles. Presenting a reduction in the error in the energy estimation, a significant reduction in the error in the flight time estimation. For the energy value, the best network has one hidden layer, while for time, a network with three hidden layers produced the best result. The proposed method reduced the error in the energy estimation by up to 15.5% for an impact energy of 10 GeV, and reduced the error for the time estimation by up to 2 orders of magnitude at the same impact energy.

MEMBERS OF THE BOARD:

EDUARDO FURTADO DE SIMAS FILHO (ADVISOR)     UFBA

ANTONIO CARLOS LOPES FERNANDES JUNIOR         UFBA

EDUARDO DA SILVA ALMEIDA                                 UFBA

BERNARDO SOTTO-MAIOR PERALVA                        UERJ

BERTRAND LAFORGE                                             Sorbonne University

Em 06/12/2024