Desenvolvimento de um simulador e estudo de mitigação de crosstalk para calorímetros eletromagnéticos

Abstract

Electronic instrumentation for high-energy physics experiments demands sensors with high channel density and precision in signal acquisition. In this context, crosstalk emerges as a critical phenomenon, characterized by electromagnetic interference between adjacent readout channels, which degrades the estimation of energy and time of detected particles. This thesis aims to develop a computational environment for simulating and studying crosstalk mitigation methods in electromagnetic calorimeters, using the ATLAS liquid argon (LAr) calorimeter as a reference. The methodology is based on extending the Lorenzetti Showers framework, in which novel modules were implemented for simulating particle time-of-flight and for modeling crosstalk at the signal level, considering capacitive and inductive couplings. From this environment, datasets were generated consisting of 50 GeV electrons and Z → e −e + decay events, with and without electronic noise, comprising events distributed across different geometric configurations of shower incidence. Preliminary results demonstrate that the implemented time-of-flight calibration satisfactorily removed the temporal bias in cells, evidenced by the convergence of time distributions to values close to zero in the EM1 and EM2 layers. The crosstalk simulation reproduced the expected waveform distortions, resulting in systematic biases in energy and time-of-flight reconstruction. To mitigate these effects, multilayer perceptron (MLP) neural network architectures were proposed and evaluated with optimization strategies based on mean squared error (MSE) and normalized mean squared error (NMSE). Models with one hidden layer and 75 neurons trained with MSE presented the best trade-off between performance and complexity, achieving determination coefficients (R2 ) above 0.95 for reconstructed energy in the central cell, while NMSE proved more effective in recovering peripheral cells with low signal-to-noise ratio. It is concluded that the developed platform constitutes a robust tool for crosstalk investigation and that the proposed machine learning approaches show potential for selective signal correction, guiding the future development of models that integrate direct energy and time estimation with pulse correction.