A deep learning potential applied in tobermorite phases and extended to calcium silicate hydrates

Abstract

Based on the datasets of first-principle simulations, machine learning potentials (MLPs) can achieve both accurate predictions of complex material systems and high-efficiency molecular dynamics (MD) computations. In this work, three MLPs of tobermorite (9, 11, and 14 Å) are respectively constructed by Deep Potential (DP) method. Accuracy assessments of each MLP are performed via MD computations of lattice constants, radial and angular distribution function, mean square displacement and elastic properties. Results using MLPs are highly consistent with those obtained from ab initio MD simulations, while with respect to efficiency, MLPs are 2 to 3 orders of magnitude faster. Furthermore, three MLPs are integrated and extended to the calcium silica hydrates system, and the calculated structure and mechanical properties show good agreement with experimental results. It proves the introduction of MLPs is feasible for larger scale MD simulations on cement-based materials with higher accuracy.