Learning Long-Range Representations with Equivariant Messages

Abstract: Machine learning interatomic potentials trained on first-principles reference data are quickly becoming indispensable for computational physics, biology, and chemistry. Equivariant message-passing neural networks, including transformers, are considered state-of-the-art for this task. Since applications require efficient scaling with system size, such models cannot act on fully connected atomistic graphs and thus neglect interactions beyond a certain cutoff, consequently failing to model long-range effects like electrostatics, dispersion, or electron delocalization. While long-range correction schemes based on inverse power laws of interatomic distances have been proposed, they are unable to communicate higher-order geometric information and are thus limited in applicability. To address this shortcoming, we propose the use of equivariant, rather than scalar, charges for long-range interactions, and design a graph neural network architecture, LOREM, around this long-range message passing mechanism. Through tests on a number of long-range datasets, we confirm that equivariant charges enable the learning of orientation-dependent interactions, and that the proposed model is competitive with, or surpasses, other approaches. Moreover, LOREM does not require adapting interaction cutoffs or the number of message passing steps to model long-range interactions, which contributes to its robustness across different systems.