Systematic assessment of various universal machine-learning interatomic potentials

TL;DR

This paper systematically evaluates five universal machine-learning interatomic potentials based on graph neural networks, assessing their transferability and guiding materials scientists in selecting suitable models for various research applications.

Contribution

It provides a comprehensive comparison of universal MLIPs, offering insights and recommendations for their selection, optimization, and areas for future improvement.

Findings

All models demonstrate transferability across different chemical systems.

Evaluation highlights strengths and limitations of each MLIP.

Guidelines for model selection in materials science are proposed.

Abstract

Machine-learning interatomic potentials have revolutionized materials modeling at the atomic scale. Thanks to these, it is now indeed possible to perform simulations of \abinitio quality over very large time and length scales. More recently, various universal machine-learning models have been proposed as an out-of-box approach avoiding the need to train and validate specific potentials for each particular material of interest. In this paper, we review and evaluate five different universal machine-learning interatomic potentials (uMLIPs), all based on graph neural network architectures which have demonstrated transferability from one chemical system to another. The evaluation procedure relies on data both from a recent verification study of density-functional-theory implementations and from the Materials Project. Through this comprehensive evaluation, we aim to provide guidance to materials scientists in selecting suitable models for their specific research problems, offer recommendations for model selection and optimization, and stimulate discussion on potential areas for improvement in current machine-learning methodologies in materials science.