In recent decades, atomic-resolution simulation and materials modeling have become integral tools for tackling challenges such as climate change and power inefficiencies by reducing experimental trial-and-error. Recently, researchers have turned to machine-learned interatomic potentials (ML-IAPs) to accelerate these endeavors, since they can serve as an efficient proxy for quantum-based methods capable of materials simulations at orders of magnitude greater spatiotemporal scales.
While powerful, ML-IAPs remain less efficient than classical forcefield based approaches, and these models tend to be most expensive when applied to molecular and reacting systems due to the separation of energies scales associated with bonded and non-bonded interactions. In this study, we developed multilayered ML-IAP to meet these challenges. This strategy works by overlaying two sets of basis functions, separately describing short- and long-range interactions. We demonstrate that this approach drives down computational costs of ML-IAM-based molecular dynamics simulation by up to 93% and speeds up simulation by an order of magnitude, enabling “quantum accuracy” simulation on massive spatiotemporal scales.
