Neural network interatomic potentials for kaolin minerals

Abstract

This thesis presents a theoretical study of the kaolin minerals carried out using previously unattainable levels of physics, enabled using machine learned interatomic potentials (MLIPs). The kaolin minerals represent systems which have evaded extensive computational study, due largely in part to the complex interlayer physics and large system sizes necessary to adequately compute properties of interest. Using high-dimensional neural networks to create DFT level interatomic potentials (which all share the moniker of ClayNN), the consideration of the necessary physics can be achieved in a way that minimizes the associated computational expense. This allows for longer, more extensive simulations to be performed, ultimately affording a more detailed description of the minerals using the studied potentials. The three predominant kaolin minerals, namely kaolinite, dickite, and nacrite are studied using three different MLIPs, each created using a different DFT functional, which includes its own description of dispersion interactions. A characterization of each mineral, provided through the computation of a number of key structural and dynamical properties are computed from molecular dynamics simulations using each MLIP and compared to experimental values. These results show the accuracy of the created MLIPs to the reference DFT potential, allowing for them to be employed as surrogate DFT interatomic potentials for future work. Through comparison of the computed properties, the need for an accurate description of dispersion interactions when considering these minerals is reinforced. It is also found that, depending on the level of theory desired when studying these systems, different dispersion interactions are better suited to compute specific structural and dynamical properties of the minerals. In addition, as hydrogen bonding is the predominant interlayer interaction within kaolin minerals, quantum mechanics in the form of nuclear quantum effects (NQEs) are included through the use of path integral molecular dynamics (PIMD) simulations of the main polymorph, kaolinite. These results show the small, but nonnegligible effects of NQEs when considering properties that involve the interlayer 3region of these minerals. In all presented cases, NQEs are of less importance than an accurate description of dispersion interactions on the structural properties of the mineral, but do weaken the interlayer hydrogen bonds and lower the cohesion of the individual mineral layers to one another. Dynamical properties are much more acutely affected by the explicit consideration of NQEs as observed when considering the O–H regions of computed vibrational spectra to be substantially shifted.

Date of Award	Dec 2024
Original language	English
Awarding Institution	Queen's University Belfast
Sponsors	Northern Ireland Department for the Economy
Supervisor	David Wilkins (Supervisor) & Gareth Tribello (Supervisor)