Factorised Neural Relational Inference for Particle Decay Reconstruction

Investigating the interactions among particles in high-energy physics is essential for various tasks, such as reconstructing particle decays. The neural relational inference encoder model, capable of capturing inter-entity dependencies and system dynamics, has been recently introduced for this purpose. However, it faces limitations in accurately capturing diverse relations between complex particles. To address this, we propose the factorised neural relational inference (fNRI) encoder model, which factorises the inferred interaction graph into a multigraph, with each layer encoding a distinct relation. Additionally, we present a simplified version of our model that offers improved efficiency without compromising performance. Moreover, we explore three different inference strategies to leverage the learned relation representations effectively. We demonstrate on a simulated particle physics decay dataset that fNRI improves the reconstruction of the decay tree structure.