Equivariance is essential, local representation is a need: A comprehensive and critical study of machine learning potentials for tobermorite phases

Abstract

Machine learning potentials (MLPs), with their combination of high efficiency and accuracy, are increasingly replacing ab initio simulations by directly predicting system energies and forces from atomic structures. However, the validation and application of these MLPs are often limited to small organic molecules. Our aim is to fill the gap in the development and validation of MLPs for Tobermorite 9, 11, and 14 Å by establishing benchmarks that include evaluation metrics corresponding to scientific objectives. Furthermore, we outline the respective strengths and weaknesses of various advanced MLPs when applied to tobermorites and suggest some directions for improvement. This approach provides a pathway to developing dedicated MLPs for high efficiency molecular dynamics simulations at long time scales, suitable for cement-based materials in future studies.