Estimação de energia no calorímetro hadrônico do ATLAS para o primeiro nível de trigger utilizando redes neurais artificiais

Abstract

This work aims to implement an energy estimation system, which is adequate to the high luminosity conditions of the most powerful collider ever built, the LHC. Scheduled in its update program, these conditions increase the probability of proton-to-proton collisions within its detectors, like ATLAS. The response from the hadronic calorimeter of the ATLAS detector, Tilecal, needs time corresponding to seven crossing particle beams to generate pulse amplitude proportional to the deposited energy. In this sense, increasing the likelihood of collision occurrence also intensifies the signal pile-up. This effect increases the amplitude estimation error of the signals, generated in the readout electronics of Tilecal in this environment. The interpretation of the calorimeter as a communication channel of dispersive character is necessary for the recovery of the input impulse, removing, this way, the pile-up effect and providing a direct measure of the energy deposition in each bunch-crossing. The utilization of linear techniques for this purpose have shown to be efficient when compared to the method currently being employed in the trigger system. However, non-linearities intrinsic to the channel cannot be modulated through these techniques. Therefore, it is proposed in this work a non-linear channel equalizer based on artificial neural networks. Through exhaustive training, the main parameters that compose the architecture of a feedforward neural network have been determined with the purpose of allying high performance with low network complexity. The neural estimator has shown to be robust in environments with high amounts of signal pile-up, as well as meeting the requirements for implementation of dedicated hardware, in which an FPGA was used. In spite of the implementation of non-linear activation functions in optimal precision hardware requiring high computational resources, this limitation was circumvented by utilizing a highly compact look-up table. Therefore, the neural network was quantified at a fixed point, aiming to reduce as much as possible the number of logical elements used without compromising its capacity to compare it with the network implemented in software.