Insights on the mechanical properties and failure mechanisms of calcium silicate hydrates based on deep-learning potential molecular dynamics

Abstract

The molecular-scale mechanical properties of calcium silicate hydrates are crucial to the macro performance of cementitious materials, while achieving coincidence between accuracy and efficiency in computational simulations still remains a challenge. This study utilizes a deep-learning potential, specifically developed for calcium silicate hydrates based on artificial neural network, to achieve molecular dynamics simulations with accuracy comparable to first-principle methods. With this potential, the elastic properties and uniaxial mechanical behaviors are explored, wherein the anisotropy and impact mechanism of calcium ratios are analyzed. The results add to evidence that the deep-learning potential possess a higher accuracy than common force fields. The anisotropy of elastic modulus is mainly attributed to different atomic interactions in various directions, while the anisotropy of strength is additionally affected by the form of failure. This study may advance the accurate molecular-scale simulation and deepen the understanding of the strength source and cohesion mechanism of cement-based materials.