Preview

PDF, English Download (83MB) | Terms of use

Abstract

The ALICE Time Projection Chamber (TPC) is one of the main tracking and particle identification detectors of the ALICE experiment at CERN, designed for the study of Quantum Chromodynamics (QCD) and the Quark–Gluon Plasma (QGP). The upgrade of the TPC readout from Multiwire Proportional Chambers to Gas Electron Multipliers enables continuous readout at interaction rates up to 50 times higher than previously achievable. While being essential for the physics program of Run 3 and beyond, this upgrade introduces new challenges for both reconstruction algorithms and the online computing infrastructure. The TPC reconstruction chain benefits significantly from GPU-based hardware acceleration, which also provides an ideal environment for modern machine learning techniques such as neural networks. This thesis develops a novel neural-network-based cluster reconstruction algorithm designed for online deployment in Run 4 and beyond. Two neural networks are trained: a classification network aimed at data-size reduction and improved performance in high-density tracking environments, and a regression network for charge deconvolution and the precise determination of cluster properties. For the training and benchmarking of these algorithms, an ideal cluster finder, based on simulated data, is developed. The performance of the neural-network-based reconstruction is evaluated in Monte Carlo studies covering a wide range of track densities and detector occupancies, with continuous emphasis on the balance between physics performance and computational feasibility for online operation. The results are compared to the current reconstruction framework throughout the analysis. The algorithm is subsequently validated on reconstructed real data from Pb--Pb collisions at interaction rates of 30–38 kHz. A cluster reduction of up to 18\% is achieved while improving the track-fit quality ($\chi^2/\text{NDF}$) and the \dedx separation power, with maintained particle identification performance. Finally, the scope of this work exceeds the original deployment timeline with a first successful commissioning in online proton–proton data taking at 507 kHz, where the data-size reduction is confirmed and shared cluster contributions to tracks are reduced.