A Deep Neural Network Potential for Water Confined in Graphene Nanocapillaries
Zhao Wen, Hu Qiu, Wanlin Guo
Abstract
Water under nanoconfinement exhibits structural and dynamical properties remarkably different from those of bulk water, and empirical potential-based molecular dynamics has played an important role in furthering our understanding of the behavior of confined water. However, existing potentials for water were commonly parametrized based on the properties of the bulk water and may be unreliable for describing nanoconfined water. Here, we develop a machine learning potential for water confined in graphene nanocapillaries using deep neural networks trained on quantum mechanical density-functional theory (DFT) calculations. This deep neural network potential offers near-DFT accuracy in terms of potential energy and atomic forces but at a computational cost much lower than DFT-based ab initio molecular dynamics methods. In particular, this potential reproduces well the DFT reference for a wide range of properties, including O–H bond length distribution, density distribution, radial distribution functions, hydrogen bonding, etc. The developed deep neural network potential opens the door to simulations of nanoconfined water with large system sizes and time scales at near-DFT accuracy.