Hamiltonian Graph Inference Networks jointly discover structure and dynamics

Researchers have developed the Hamiltonian Graph Inference Network (HGIN), a novel method for simultaneously discovering the interaction structure and predicting the dynamics of lattice Hamiltonian systems from trajectory data. HGIN addresses limitations of previous graph-based approaches by handling both separable and non-separable Hamiltonians, as well as heterogeneous node dynamics. The system couples a structure-learning module with a trajectory-prediction module, achieving significant reductions in prediction errors on benchmark systems. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a new method for learning complex physical system dynamics and structures, potentially improving scientific modeling.

RANK_REASON This is a research paper introducing a new method for learning dynamics and structure in physical systems.

Read on arXiv cs.LG →

• Hamiltonian Graph Inference Network
• HGIN
• Klein-Gordon lattice
• discrete nonlinear Schrödinger lattices
• GNN

• paper
• other