Hyperparameter Optimization of Deep Learning Models Using High-Performance Computing Resources

This thesis focuses on the hyperparameter optimization of deep-learning models using high-performance computing (HPC) resources, in order to study cluster counting in drift chambers. Cluster counting is a promising particle-identification technique in particle- physics experiments, where the goal is to reconstruct the number of primary ionization clusters from digitized waveforms. To support this study, realistic simulated samples are first generated using a Garfield based simulation package under conditions as close as possible to the real test-beam data recorded in 2022 (180 GeV) and 2023 (2–10 GeV). The next part of this work focuses on the training, optimization, and selection of best machine-learning models using HPC resources on simulated samples for the cluster counting study in two main steps. In the first step, several jobs with different sets of hyperparameters were submitted simultaneously on the local ReCaS Bari HPC system to train Long Short-Term Memory (LSTM) models on simulated samples matched to the 2022 and 2023 real test-beam data for the peak-finding algorithm. This stage was treated as a classification problem, in which signal peaks were separated from the noise background in the waveform. Different hyperparameter settings, including activation functions, optimizers, number of epochs, batch size, patience, and dropout rate etc, were systematically explored, while computational factors such as CPU usage, memory consumption, and job duration were also taken into account. The best LSTM model was selected according to the highest area under the curve (AUC) among all tested configurations. In the second step, the same model-selection logic was applied to the Convolutional Neural Network (CNN) model, which was used for clusterization. This stage was treated as a regression problem, in which the number of primary ionization clusters was estimated from the peaks identified in the first step. Among all tested hyperparameter configurations, the best CNN model was selected according to the lowest mean squared error (MSE). The optimized LSTM and CNN models were then further evaluated on simulated samples of kaons and pions at 2, 4, 6, 8, and 10 GeV for particle-identification studies. In addition, the performance of the selected models was evaluated on national HPC resources and compared with that on the local ReCaS Bari HPC system in terms of CPU usage, memory consumption, and job duration. The final and most important task is related to the application of optimized machine- learning models to real test-beam data. For this purpose, the simulated samples corre- sponding to the 2022 test-beam data are further tuned in terms of noise and maximum amplitude. After this tuning procedure, the LSTM and CNN models are trained on the tuned simulated samples using HPC resources. Among all tested hyperparameter configurations, the best LSTM model for peak finding is selected according to the highest area under the curve (AUC), while the best CNN model for clusterization is selected according to the lowest mean squared error (MSE). Finally, the best selected CNN model is applied to the real digitized waveforms to reconstruct the number of primary clusters, and its performance is compared with that of the Running Template Algorithm (RTA).