Speakers
Description
Particle flow reconstruction algorithms are fundamental for physics analysis at collider experiments. Improving these algorithms with deep learning presents a unique chance to enhance experimental sensitivity at the LHC and future facilities. This talk presents HGPflow, a deep learning method using hypergraphs that offers a physics-motivated framework for the energy assignment task in particle reconstruction. We show that HGPflow can reconstruct full proton-proton and electron-positron collisions, yielding benefits in both precision and interpretability over current methods. We also underscore the importance of maintaining locality when training with full collision events and suggest a technique to ensure the model avoids learning global event topologies.
