Large Scale Simulations of Soft Materials with Equivariant Deep Learning Potentials

Accurate computationally efficient predictions of the energy and atomic forces of many-atom systems are a long-standing goal in the natural sciences. Interatomic potentials based on Message-Passing Neural Networks have emerged as the leading paradigm toward this goal over the past years. Their propagation mechanics, however, make parallel computation challenging and limit the length scales that can be studied. Strictly local descriptor-based methods on the other hand, have been scaled to massive systems, however, they currently fall short of the accuracy seen in message passing approaches. We have recently introduced the Allegro model, a fully local equivariant deep learning interatomic potential that is massively parallelizable while retaining the high accuracy of equivariant message passing potentials. Molecular dynamics simulations using the Allegro potential recover structural and kinetic properties of an amorphous phosphate electrolyte in great agreement with first principles calculations and can be scaled to 100 million atoms. Here, we demonstrate how Allegro can be used to study complex, soft materials using large-scale Molecular Dynamics simulations.