Machine learning potentials for tobermorite minerals

Abstract

Molecular dynamics (MD) simulation is an important tool to understand the physical and chemical properties of cement hydrates at the atomic level. MD with the machine learning potential (MLP) is considered a promising approach for accurate prediction of material properties. However, the applications of machine learning MD for multicomponent systems with a liquid–solid interface have been limited so far. In this work, we used artificial neural networks (ANNs) to construct MLPs for tobermorite minerals. Two MLPs were produced by optimization using different objective functions: one MLP was optimized for both the forces and energies of density functional theory (DFT) results (MLP-FE), and the other MLP was fitted to only the energies (MLP-E). Accuracy assessments of the MLPs were performed for lattice parameters, elastic constants, and bulk, shear moduli, and vibrational density of states. The results of the assessments showed that the accuracy of the potentials largely depended on the objective functions used in the training of the ANNs, i.e., MLP-FE was comparable to the potential energy surface (PES) of DFT, although MLP-E failed to reproduce the several physical quantities given by DFT. To evaluate the applicability of MLP-FE to extended systems beyond the limits of DFT calculations, a large scale MD simulation of a nanoporous tobermorite model was demonstrated using MLP-FE. The transport and distribution properties of atoms in the liquid part of the porous model were investigated. Slow diffusion of water confined in a nanopore was detected, and the results were consistent with experimental data and previous works using classical force fields. The results obtained in this study suggested that MD with MLP is a practical method for large scale molecular simulations of cement hydrates with DFT accuracy.