Replacing Quantum Chemistry With Machine-Learned Interatomic Potentials: Revolution or Evolution?

A long-standing goal in computational chemistry and materials science has been the development of general-purpose interatomic force fields that define the energy and forces associated with arbitrary sets of atoms. This task can be accomplished with density functional theory (DFT) and other levels of computational quantum chemistry, but the computational cost of these methods strongly constrains the physical problems that can be explored. Rapid advances in machine-learned interatomic potentials (MLIPs) mean that calculations at DFT levels of accuracy will soon be accelerated by factors of up to a million, a situation that will dramatically change the landscape of computational chemistry. In this Outlook, we examine the implications and limitations of MLIPs and describe a research agenda for taking full advantage of these remarkable tools.