Uporaba arhitekture transformerja za simulacijo pozicijsko občutljivih silicijevih detektorjev

This thesis explores the use of generative machine learning in particle physics, with a particular focus on simulating the detection and reconstruction of muon trajectories in colliders such as the Large Hadron Collider (LHC) at CERN. The goal of the thesis is to develop a generative model based on the Transformer or GPT architecture, capable of producing realistic synthetic data from actual detector signals. This approach can significantly complement existing simulation methods, which are often computationally intensive and time-consuming. The model is trained on simulated detector data and generates new collision event samples that mimic physically plausible particle trajectories. In this work, we examine various architectural configurations, training methods, and strategies for evaluating the quality of the generated data. Validation results show that such a model can successfully replicate key characteristics of real collision events, offering the potential for faster and more cost-effective experimental preparations in particle physics. The results indicate that generative approaches have the potential to improve the efficiency of data analysis in large-scale experiments and to enhance existing simulation tools. Future work could explore the inclusion of spatial information, improved evaluation of physical consistency, and the application of the model in real-time event detection.