Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper establishes that the Hypergraph Neural Network, utilizing 3-body message passing, serves as a universal approximator for potential energy surfaces through a newly developed multi-layer completeness theory. The authors demonstrate that L layers of message passing on sparse, cutoff-based graphs can match the representational power of full L-hop neighborhood access under certain conditions. This work provides a rigorous foundation for the prevalent use of multi-layer message passing in graph neural networks for interatomic potentials, confirming the effectiveness of architectures like DPA3 and CHGNet in this context.
Multi-layer message passing in graph neural networks can achieve universal approximation even with cutoffs smaller than the physical interaction range, challenging previous assumptions about model architecture.
We prove that the Hypergraph Neural Network, an invariant architecture with 3-body message passing, is a universal approximator for potential energy surfaces. Our main contribution is a multi-layer completeness theory. We show that $L$ layers of message passing on sparse, cutoff-based graphs achieve the same representational power as having access to the full $L$-hop neighborhood, provided the configurations are generic, satisfy an overlap condition and a connectivity condition. This provides the first rigorous justification for the common practice of using multi-layer message passing with a per-layer cutoff smaller than the physical interaction range, the setting used by virtually all practical graph neural network based machine-learned interatomic potentials. As immediate consequences, we show that both DPA3 and CHGNet architectures inherit universal approximation.