Transferable long-range interactions in machine-learned interatomic potentials

Viktor Zaverkin, Matheus Ferraz, Francesco Alesiani, Henrik Christiansen, Makoto Takamoto, Federico Errica, Mathias Niepert

Keywords: machine-learned interatomic potentials, long-range interactions, dispersion and electrostatics, molecular and materials simulations

TL;DR: We introduce a framework that augments machine-learned interatomic potentials with analytic long-range electrostatics and dispersion, enabling the development of transferable models across molecular and bulk systems.

Abstract: Machine-learned interatomic potentials (MLIPs) enable efficient large-scale atomistic simulations with near first-principles accuracy. However, their limited ability to model long-range dispersion and electrostatics often reduces accuracy in many practical applications. Achieving consistent modeling of these interactions is also crucial for ensuring transferability between molecular and bulk systems. In this work, we present a framework that augments local MLIPs with analytic long-range electrostatic and dispersion terms parameterized through latent partial charges learned directly from total energies and forces. By ensuring a consistent treatment of long-range interactions across molecular and bulk systems, the framework enables the improved transferability of MLIPs. We assess the framework's effectiveness on the synthetic point-charge datasets and further demonstrate the improved accuracy and generalization ability of the resulting models on the SPICE-v2 dataset and in molecular dynamics simulations.

Release To Public: Yes, please release this paper to the public

Submission Number: 21