Scaling Up Equivariant Potentials: From Fragments Towards Exascale Simulations

Great progress in accuracy and data efficiency has been made in machine learning interatomic potentials (MLIPs) thanks to equivariant models, which leverage the symmetries of the underlying physics to learn rich geometric representations. But while MLIPs have long promised to bring ab-initio accuracy to calculations that could not possibly be conducted with methods like DFT, the need for representative training data from quantum methods has often restricted MLIPs' application too far from the regimes in which DFT is practical. Further, architectural limitations in the most accurate models have impeded their scaling and speed.<br/><br/>In this talk, we present recent developments on how Allegro—a machine learning architecture we recently developed to retain the benefits of equivariance while omitting previous schemes, like message passing, that limit computational performance—can scale to larger length-scales, longer time-scales, larger datasets, and scale up from fragments to complex systems. In particular, I will discuss techniques to improve simulation stability and thus the ability to run long timescale simulations; speed optimizations including variable cutoffs and a simplified multi-path tensor product; and how the Allegro architecture facilitates scaling model capacity for diverse, large data. Indications of its ability to scale up from large databases of fragments to complex systems never seen in the training data will be demonstrated using an example in biochemistry.