Application of machine learning techniques to improve the resolution of particle timing detectors

This thesis investigates the use of machine learning techniques to improve the precision of timing measurements in the Precision Proton Spectrometer (PPS) at CERN's Large Hadron Collider (LHC). PPS is a subsystem of the Compact Muon Solenoid (CMS) detector and is responsible for detecting so-called forward protons. PPS uses diamond-made timing sensors and advanced readout electronics to produce sampled signals of voltage when detecting a particle. These signals are measured with varying amplitudes, which introduces the effect called time walk. The current method of mitigating the time walk effect -- the constant fraction discriminator (CFD) -- has certain drawbacks, such as not taking into account noise, or being unable to handle signal properties that depend on amplitude. Machine learning approaches have the potential to improve the handling of the time walk effect and mitigate the CFD's issues. The practical part of this thesis involves predicting a single timestamp based on 24 values of sampled voltage. Deep neural networks are the focus of the experiments, as they have been shown to be effective in similar applications. Various deep learning architectures were tested using a hyperparameter tuning procedure incorporating the Bayesian optimisation strategy and cross-validation. Ultimately, models based on the UNet convolutional architecture yielded the best results. The study provides a comprehensive analysis of the optimal networks outputted by the tuning procedure. The networks were compared to the current state-of-the-art – CFD – in a series of numerical experiments and consistently demonstrated improved timing precision, despite being trained on a low-quality reference dataset. The improvements ranged from 6% to 18% in the primary numerical experiment, providing strong evidence that neural networks can be effectively used in the timing measurements at PPS.