UC San Diego

In high energy physics (HEP), there has been persistent interest in leveraging generative machine learning to model the structure of jets: the collection of particles generated from particle collisions. Being able to model the distribution of jet data enables downstream tasks such as anomaly detection, improving our search methodologies for rare and new physics. Traditionally, jet modeling has been performed on 2D jet-image representations; however, extending 3D point clouds to jet data has led to the much more natural “particle cloud” representation, where jets are modeled as a set of particles in momentum-space. In this thesis, I present two generative machine learning methods for modeling jets by their particle cloud representations, one using a graph-based autoencoder model and one using diffusion models. In addition, this thesis will demonstrate how these two models can be applied for anomaly detection for new physics once trained.