4185936 - Machine learning molecular dynamics simulations of clay minerals

Molecular dynamic (MD) simulation is a popular method for investigating the physical and chemical properties of clay minerals. Two methods are mainly used: classical MD and ab initio molecular dynamics (AIMD). In classical MD, atoms are approximated as point particles, of which interatomic potentials are described by simple functions with empirical parameters. This method is computationally efficient, but the accuracy strongly depends on tuning the empirical parameters. In AIMD, using density functional theory (DFT), the electrons in atoms are described by quantum mechanics, although the nuclei are approximated as point particles. Therefore, the accuracy of AIMD is high, but the computational cost is also high.
Machine learning, particularly artificial neural networks (ANNs), is a technology that has developed at an astonishing rate in recent years. ANN is a fitting method for high-dimensional non-linear functions, and it is being applied to a wide range of uses. Behler and Parrinello proposed the application of ANN to the MD method, and in recent years, its application has been rapidly expanding. In their formulation, the potential energy surface of DFT is learned by ANNs, and the learned potential is referred to as machine learning potential (MLP). MD with MLP is called machine learning molecular dynamics (MLMD).
We made MLPs of kaolinite and evaluated its structural, mechanical, and vibrational properties. The results showed that our MLPs accurately reproduced the DFT results for the structural and mechanical properties. Furthermore, our MLPs accurately reproduced the experimental results of the vibrational property of the hydroxy group in kaolinite. The recent development of MLP for clay minerals will be discussed in the presentation.