Date of Award
Doctor of Philosophy (PhD)
Monte Carlo-based event generators have been the primary source for simulating particle collision experiments for the study of interesting physics scenarios. Monte Carlo generators rely on theoretical assumptions, which limit their ability to capture the full range of possible correlations between particle’s momenta. In addition, the simulations of the complete pipeline often take minutes to generate a single event even with the help of supercomputers.
In recent years, much attention has been devoted to the development of machine learning event generators. They demonstrate attractive advantages, including fast simulations, data compression, and being agnostic of theoretical assumptions. However, most of the efforts ignore faithful reproductions, and detector effects due to their complexity and rely on theories for detector simulations.
In this work, we present a new machine learning-based event generator framework free of theoretical particle dynamics assumption. We first create a Feature-Augmented and Transformed Generative Adversarial Network (FAT-GAN) that selects a set of transformed features to faithfully reproduce simulated and experimental data. Then, we extend FATGAN by conditioning the component neural networks according to the given reaction energy and develop a Conditional FAT-GAN (cFAT-GAN) that can generate events at unrelated beam energies.
Next, we implement a conditional folding model that learns the correlations between vertex-level and detector-level events and simulates the distortions produced by the detector machines. The folding model is then integrated into a generator to reconstruct vertex-level events. This serves as a practical framework in a real experimental analysis where such effects must be incorporated.
We finally evaluate different Neural Network architectures and use machine learning techniques for model interpretation and evaluation. In addition, we analyze the GANs latent variables to extract physics resonance regions, illustrating the ability of the developed model to distinguish between the underlying physics mechanisms.
This framework has been validated on simulated inclusive deep-inelastic scattering data along with the existing parametrizations for detector simulation. The generated results provide a realistic proof of concept for designing a machine learning-based event generator that will be a valuable tool in nuclear and particle physics programs to facilitate the studies of high-energy scattering reactions and understand different physical mechanisms.
In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).
"Machine Learning-Based Event Generator"
(2022). Doctor of Philosophy (PhD), Dissertation, Computer Science, Old Dominion University, DOI: 10.25777/7prx-gq72