Graph neural networks for event reconstruction at the Large Hadron Collider

Abstract

The Large Hadron Collider (LHC) serves as a powerful experimental platform for investigating the fundamental constituents of matter under extreme conditions.
However, the reconstruction of boosted objects -- highly energetic particles whose decay products manifest as dense sprays of hadrons or jets -- presents enduring challenges.

Accurate reconstruction of these objects is essential not only for validating the Standard Model but also for probing potential new physics phenomena such as heavy resonances and exotic particles.

To address these challenges, this thesis explores the application of Graph Neural Networks (GNNs) for event reconstruction at the LHC.
GNNs offer a transformative approach by leveraging the relational and geometric structure inherent in collider data.

They are permutation invariant and generalise well to point clouds of variable size, which are ideal properties for collider physics.

Additionally, they can be designed to respect physical properties such as Lorentz equivariance, rotational symmetry, and infrared and collinear safety, making them suitable for the complexities of jet clustering and boosted object reconstruction.
This research explores GNN-based frameworks, incorporating novel methodologies to enhance reconstruction fidelity, and establishing new approaches to node prediction in message passing architectures.

This thesis presents our exploratory work as a kind of A - Z prototype for conducting Graph Machine Learning (ML) studies on simulated collider data.
Three major contributions to the field are provided.

First, we introduce a robust software ecosystem tailored for collider data analysis, enabling seamless data manipulation and model integration.

Second, we propose an innovative clustering algorithm that dispenses with traditional jet definitions, instead incorporating simulation-based information, such as particle ancestry and momentum, to achieve superior clustering granularity.

Finally, we demonstrate the application of IRC safe Interaction Networks with the novel Bright Edge Classification for effective node classification, and Cluster Double Sifting for the reconstruction of boosted Higgs bosons and top quarks, achieving state-of-the-art performance in providing detector-level constituents for the reconstruction of the top quark.

More information

Identifiers

ORCID for Jacan Loic Chaplais: orcid.org/0009-0002-4092-1878

ORCID for Stefano Moretti: orcid.org/0000-0002-8601-7246

ORCID for Srinandan Dasmahapatra: orcid.org/0000-0002-9757-5315

Catalogue record

Export record