Neural network regression for particle identification with the ALICE TPC detector in Run 3

The gaseous Time Projection Chamber (TPC) of the ALICE experiment at CERN serves some of the most crucial roles in many physics analyses within the collaboration and is responsible for 92.5% of the raw data taken with the experiment. One of its major advantages is extensive particle identification over a wide range of momenta. The basic underlying physics concerns the process of ionization of gas molecules and the associated specific energy loss of traversing particles, described by the Bethe-Bloch formula. In this thesis, a novel method for the tuning of parameters of the ALEPH parameterization of the Bethe-Bloch formula is presented based on the concept of hyperparameter optimization. A novel framework called OPTUNA and custom designed loss functions are investigated and tested against the performance on datasets with known particle identity from Run 2 of the LHC. Besides the parameterization, further corrections to the mean as well as an estimation of the standard deviation of particle data distributions have to be made in high dimensions, which forms the main body of this thesis. Both parts are approximated with fully connected feed-forward neural networks trained on identified daughter particles from weak decays of K^0_S , Λ, Λ ̄ and γ conversions. An average accuracy of around 3‰ for the mean correction based on a neural network ensemble is achieved. This is compared with the results obtained from one-dimensional spline corrections in Run 2 and it is shown that the neural network introduced in this thesis can perform similarly well as the approaches from Run 2 but shows significant improvements in higher dimensions since it does not rely on a factorization approach. The estimation of uncertainty of the distribution for each particle species is performed likewise, and an overall similar performance as the functional parameterization from Run 2 is achieved. However, due to the multidimensional mean corrections by the neural network and the limitations that a parameterized functional shape inherits, the standard deviation is captured better for all particle species by the neural network. In contrast to the Run 2 approach, this method works without additional iterations and consumes overall less time and effort for quality checks.