An ab-initio deep neural network potential for accurate large-scale simulations of the muscovite mica-water interface

Abstract

Mineral-water interfaces play a critical role in several geochemical processes relevant to the habitability of our planet. These processes are often underpinned by the hydration of the ions covering the mineral surface and the resultant ion exchange with the aqueous environment. Muscovite mica, whose surface is nominally covered by K + ions, has long been considered an ideal system for probing some of these processes. However, despite several decades of both experimental and computational work, there still remains a lack of consensus on ion hydration and exchange at the mica-water interface. To obtain a detailed microscopic picture of these processes, we have developed an ab-initio based deep neural network potential (DP) describing the potential energy surface (PES) of the mica(001)-water interface. Our training dataset consisted of bulk mica, bulk liquid water, the mica(001) surface in vacuum and the explicit mica(001)-water interface, both with different surface K + arrangements. The trained model is able to reproduce the energies and forces derived from Density functional theory (DFT) based on the SCAN exchange correlation functional with good accuracy. For the surface in vacuum, K + ions located in the ditrigonal cavities composed of 2 Al atoms are predicted to be energetically more favourable, while the specific surface arrangement of K + ions does not affect the water density profile at the mica-water interface. The developed model sets the stage for estimating the energetics of the ion-exchange process at affordable computational cost without compromising the accuracy of first-principles methods.