Computational intelligence applied to crosstalk mitigation in a liquid argon calorimeter

Abstract

ATLAS (A Toroidal LHC ApparatuS) is the largest experiment at the LHC (Large Hadron Collider) located at CERN (Centre Européen pour la Recherche Nucléaire). ATLAS is positioned at one of the collision points in the accelerator tunnel. It is composed of specialized detectors designed to characterize particles produced in proton-proton (pp) collisions with a center-of-mass energy of 13 TeV. One of its specialized detectors is the Liquid Argon Calorimeter (LAr), featuring approximately 187,000 sensor cells to record electromagnetic particle production. Calorimeters are widely used in particle physics experiments to measure and absorb the energy of particles produced during collisions within the experiment. The ATLAS LAr calorimeter has fine granularity and a high density of cells. Combined with high collision rates and the mechanical and electronic structure of the detector's readout system, this setup generates interference effects from neighboring electronic channels (crosstalk -- XT). XT complicates energy and timing estimation for incident particles. The Optimal Filter (OF) is the standard method used in ATLAS for energy and time estimation. However, XT is not accounted for in the coefficient design, leading to errors in the estimated energy and time-of-flight of the particles. This thesis investigated potential solutions to mitigate XT using machine learning techniques and statistical signal processing methods. The proposed solution is based on the development of an artificial neural network (ANN) estimator to mitigate the undesirable effects of XT. Two simulated datasets, based on electromagnetic particle physics models, were used to apply and evaluate the model. The signals obtained from the developed simulators were employed to develop and assess potential machine-learning solutions. Using data produced by the available simulators, the supervised approach proved effective for estimating both the energy and time-of-flight values of particles. The \textit{Multilayer Perceptron} (MLP) architecture was applied for regression tasks to predict the energy and time values of particles without XT influence. The results showed low RMSE (\textit{root mean squared error}) compared to the standard method for estimating particle energy and time-of-flight. The approach achieved a significant reduction in energy estimation errors and a notable reduction in time-of-flight estimation errors. For energy estimation, the best network architecture featured a single hidden layer. For time estimation, a network with three hidden layers produced the best results. The proposed method reduced energy estimation errors by up to 15.5\% for a 10 GeV impact energy and reduced time estimation errors by up to two orders of magnitude at the same impact energy.