Fast and accurate machine-learned interatomic potentials for large-scale simulations of Cu, Al and Ni

TL;DR

This paper introduces low-dimensional, tabulated machine learning interatomic potentials for Cu, Al, and Ni that are significantly faster than traditional ML potentials, enabling large-scale molecular dynamics simulations.

Contribution

The authors develop and validate a new tabGAP method that achieves two orders of magnitude higher efficiency for ML potentials, suitable for multi-million atom simulations.

Findings

Validated potentials accurately reproduce material properties

Enabled large-scale simulations of defects and deformation

Achieved significant computational speedup

Abstract

Machine learning (ML) has become widely used in the development of interatomic potentials for molecular dynamics simulations. However, most ML potentials are still much slower than classical interatomic potentials and are usually trained with near equilibrium simulations in mind. In this work, we develop ML potentials for Cu, Al and Ni using the Gaussian approximation potential (GAP) method. Specifically, we create the low-dimensional tabulated versions (tabGAP) of the potentials, which allow for two orders of magnitude higher computational efficiency than the GAPs, enabling simulations of large multi-million atomic systems. The ML potentials are trained using diverse curated databases of structures and include fixed external repulsive potentials for short-range interactions. The potentials are extensively validated and used to simulate a wide range of fundamental materials properties, such as stacking faults and threshold displacement energies. Furthermore, we use the potentials to simulate single-crystal uniaxial compressive loading in different crystal orientations with both pristine simulation cells and cells containing pre-existing defects.