Speaker
Description
At the Large Hadron Collider, the kinematic reconstruction of heavy, short-lived particles is crucial for precision measurements of the Standard Model and searches for new physics. Performing kinematic reconstruction in events with a high multiplicity of final-state objects is especially challenging due to the extensive potential combinatoric assignments. To address this, we present HyPER, a graph Neural Network that utilizes hypergraph representation learning for reconstructing the origin states from among the detected final states. HyPER discovers relational information using message-passing on a conventional graph structure, and studies higher-order correlations through the introduction of a hypergraph structure, to identify probable parent particles. HyPER is tested on simulated all-hadronic $t\bar{t}$ events and shown to perform favorably compared to existing state-of-the-art reconstruction techniques, while demonstrating superior parameter efficiency. The novel hypergraph approach allows the method to be applied to particle reconstruction in a multitude of different physics processes.
