280 / 2022-09-29 13:54:25

Deep learning potential for predicting C-S-H/PVA/graphene interfaces

Machine learning potential; Deep neutral network; Graphene; Cementitious materials; Molecular dynamics simulations

A deep learning potential with high precision was developed for predicting the interfaces between C-S-H, PVA and graphene with large scale molecular dynamics (MD) simulations. MD simulations can generate atomic scale details of deformation mechanisms, failure mode and interfacial structures of cementitious materials modified by PVA and graphene. However, accurate MD potential for describing the interactions at the interfaces is lacking, and owing to the complexity of the C-S-H/PVA/graphene interfaces, it is difficult for empirical potentials to predict the interactions between different materials with high precision. Recent development of machine learning potentials such as deep learning potential provide a plausible way for developing MD potential for complicated systems with high accuracy. Herein, we developed a deep learning potential for the  C-S-H/PVA/graphene system using the DeePMD package.The deep learning potential was fitted with its prediction error in energy and force respectively being 12.1 meV/atom and 352 meV/Å, in comparison with first-principles calculations. The interfacial structures between C-S-H, PVA and graphene were than predicted by large scale MD simulations. The developed MD potential allows for further studies on the reinforcing mechanisms of graphene on cementitious materials. The deep learning potential can be easily extended to be used in other systems because of the great scalability of deep neutral network.