Scaling Equivariant Machine Learning for Atomic-Scale Simulations
Albert Musaelian, Harvard University
Simulations at the atomic scale can deliver an otherwise inaccessible understanding of the behavior of chemical and material systems. Many of the most powerful and popular simulation techniques require the calculation of the potential energy of a system of atoms given the configuration of those atoms in space. Traditionally, empirical hand-designed approximations or expensive quantum mechanical calculations have been used to compute the potential energy, severely limiting the systems and phenomena that could be studied.
Machine learning interatomic potentials have since emerged and in many cases overcome these limitations. In particular, equivariant neural network potentials, which directly leverage the symmetry structure of the machine learning problem, have shown leading accuracy, data efficiency, and robustness. In this talk I will discuss Allegro, an equivariant neural network potential architecture we developed in order to scale these advantages of equivariant methods to highly parallel computers, and using them, to larger length- and time-scale simulations. In particular, I will discuss the Allegro architecture, demonstrations of and improvements to its computational scalability, and examples of its application to various scientific problems.