Ab Initio Characterization of the CO2-Water Interface Using Unsupervised Machine Learning for Dimensionality Reduction.

Abstract

Precise characterization of the supercritical CO2-water interface under high pressure and temperature conditions is crucial for the geological storage of carbon dioxide (CO2) in deep saline aquifers. Molecular dynamics (MD) simulations offer a valuable approach to gaining insight into the CO2-water interface at a microscopic level. However, no attempt has been made to characterize the CO2-water interface with the accuracy afforded by ab initio calculations. In this study, we performed ab initio MD (AIMD) simulations to investigate the structural and dynamical properties of the CO2-water interface, comparing the results with those obtained from classical force-field MD (FF-MD) simulations. Molecular orientation at the interface was well reproduced in both AIMD and FF-MD simulations. Characteristic structural fluctuations of water at the interface were unveiled by applying multidimensional scaling and time-dependent principal component analysis to the AIMD trajectories; however, they were not prominent in the FF-MD simulations. Furthermore, our study demonstrated a marked difference in the residence time of molecules in the interface region between AIMD and FF-MD simulations, indicating that time-dependent properties of the CO2-water interface strongly depend on the description of the intermolecular forces.