Neural Network Atomistic Potential for Pyrophyllite Clay Simulations.

Abstract

In this study, a high-dimensional neural network potential for the smectite pyrophyllite clay has been developed from density functional theory (DFT) data, including correction for dispersion interactions. The data set has been built from the adaptive learning approach, resulting in a diverse and very concise set of selected structures comprising only representative ones. Two neural network potential (NNP) data sets have been constituted from sets of energies and forces computed at two different levels of DFT accuracy. Validation tests show very good accuracy for the computed energies and forces of various systems differing by their size and simulation conditions. The developed potentials are able to reproduce structural parameters with excellent agreement with DFT values as well as experimental data and are the first NNPS able to reproduce clay layers' properties held together via van der Waals interactions. The NNP constructed from data of higher DFT levels shows better results for extreme condition simulations. In addition, elastic properties, exfoliation energies, and vibrational density of state are also well reproduced, showing better performances than standard force fields at a fraction of DFT computation time.