Large-scale atomistic simulation at near-quantum accuracy with equivariant machine learning

Abstract

Since the first digital computers were built, scientists have programmed them to make concrete predictions about real-world systems by numerically solving the relevant physical laws that govern the behavior of those systems. Many of these efforts have focused on simulating the behavior of the atoms that make up materials and chemicals.

As computational power has grown, these simulations have shifted from using simple qualitative models of idealized interatomic interactions to sophisticated techniques for solving the quantum physics of realistic systems of electrons and atomic nuclei. These first-principles models offer a remarkable degree of accuracy and predictive power, but their computational cost also imposes severe limitations on the kinds of length- and time-scales that can be simulated. Machine learning models of interatomic interactions---called machine learning interatomic potentials (MLIPs)---can be used as fast approximations of these accurate but costly quantum calculations. In principle, MLIPs can enable much longer and larger simulations to be run at near-quantum accuracy.

This thesis presents algorithms and implementations that can realize this promise through "equivariant" machine learning techniques that leverage the symmetries of the underlying physics. The first half discusses the equivariant MLIP NequIP, with its significantly improved data efficiency, accuracy, generalization, and robustness. The second half presents the development of the novel Allegro architecture. Allegro makes it possible to scale up equivariant machine learning: this ability is demonstrated through comprehensive performance experiments that apply a powerful Allegro model trained on a large dataset of more than 1M quantum calculations to biomolecular systems of realistic sizes of up to 44M atoms while taking advantage of up to 5120 GPUs.