Predicting the atomic structure of magnetic layered materials from ab-initio materials simulations and machine learning

A reliable soft chemical method has been developed to synthesize an air-stable layered material H_xCrS₂, which can be exfoliated down to ultrathin layers, providing the promise for synthesizing two-dimensional magnets¹. However, the atomic structure of H_xCrS₂ is still unknown. We used a combination of density functional theory (DFT), ab-initio molecular dynamics (AIMD) and cluster expansion to study the energetics of H_xCrS₂ as a function of Cr vacancies and H impurities. Then, we investigated the stability, electronic and magnetic properties and examined the effect of layering on the investigated properties. We further applied different strains at different temperatures on the structures to generate a dataset of diverse local atomic environments for machine learning (ML). We used Gaussian processes and neural networks to fit a model for the potential energy surface (PES) of H_xCrS₂. After that, we implemented this potential model in classical molecular dynamics (MD) to be able to study more properties for macroscale structures of H_xCrS₂ to benchmark with available experimental data. This study represents a novel method of predicting the atomic structures of macroscale slabs with quantum mechanical accuracy.

¹X. Song et. al, J. Am. Chem. Soc. 141, 15634 (2019)