Large scale simulations of soft materials with equivariant deep learning potentials

Access to an accurate and computationally efficient description of the energy and atomic forces of many-atom systems is a long-standing goal in the natural sciences. Neural message passing potentials have emerged as the leading paradigm toward this goal over the past years. However, their propagation mechanics makes parallel computation difficult and inherently limits the length scales they can model. Strictly local descriptor-based methods have been scaled to massive systems, but currently lack the accuracy observed 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. Allegro obtains state-of-the-art accuracy on a series of benchmarks and has been demonstrated to recover structural and kinetic properties of an amorphous phosphate electrolyte in great agreement with AIMD. Exploiting the strict locality, the method finally has been scaled to a simulation of >100 million atoms. Here, we will show how we can leverage the unique combination of scale and accuracy of Allegro to study complex, soft materials using large-scale Molecular Dynamics simulations.