Year: 2025
Type: dissertation
Abstract: In light of the rapid advancement of machine learning and its increasing uptake in high energy physics, foundation models present an encouraging outlook for responding to concerns about data scarcity and computational cost limiting the scale of contemporary and future research. This thesis builds on OmniJet-α, a foundation model released in 2024 using a transformer-based backbone architecture. By leveraging the OmniJet-α codebase, including the use of a Vector Quantised-Variational Autoencoder (VQ-VAE) to generate tokens representing the original data, this work aims to extend the model’s functionality to support the prediction of particle jet evolution. For that purpose, the feasibility of adapting its tokenisation strategy and customising the model architecture was investigated, supported by a revised data preprocessing pipeline to transform raw jet data into an optimal machine learningready format. While jet evolution predictability remains uncertain, the study identified critical aspects for future work and successfully demonstrated model inference functionality. The findings highlight the major role of feature engineering in machine learning applications and contribute to the development of foundation models in particle physics.
Source: Theseus (Ammattikorkeakoulujen)
Author: Cordruwisch, Tom
Institution: Universität Hamburg
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6983119207