Learning Local Equivariant Interatomic Potentials for Large-Scale Atomistic Dynamics

E(3)-Equivariance — the property of a function to transform with the actions of the Euclidean group — has recently emerged as a key design principle in machine learning interatomic potentials. Starting with the high accuracy, sample efficiency, and generalization properties of the NequIP potential [1], a series of other equivariant message passing interatomic potentials has since been proposed. However, all current methods share the message passing paradigm, in which information is iteratively propagated along an atomistic graph. This propagation mechanics, however, severely hinders the time- and length-scales that can be studied with message passing approaches. Here, we introduce Allegro, a strictly local, E(3)-equivariant deep learning interatomic potential that combines the high accuracy of equivariant neural networks with the scalability of local methods. We demonstrate that Allegro not only obtains state-of-the-art accuracy on a variety of benchmarks, but also displays remarkable generalization to out-of-distribution data. We then show how molecular dynamics simulations driven by Allegro predict structure and kinetic of a Lithium phosphate electrolyte with high fidelity. Finally, we demonstrate how the locality of the proposed method enables a simulation of 100 million atoms on reasonable compute resources, enabling a previously impossible combination of accuracy, scale, and computational efficiency.<br/><br/>[1] Batzner et al. (2022). E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13(1), 1-11