Neural network potentials for modeling nonstoichiometric materials: a case of Chromium Sulfides Cr(₁−ₓ)S

Abstract

Deviation from stoichiometry can yield a diverse range of stable phases with distinct physical and chemical properties. To comprehensively explore nonstoichiometric materials, it is crucial to investigate their compositional and structural domains with precision and cost-effectiveness. However, the extensive diversity in these domains render first-principles methods, such as density functional theory (DFT), inappropriate for such endeavors. In this study, we propose a generic framework that utilizes neural network potentials (NNPs) to model nonstoichiometric materials with chemical accuracy at realistic length and time scales. We apply our framework to analyze nonstoichiometric Cr(₁−ₓ)S materials, a compelling material category with significant potential in the field of two-dimensional (2D) magnetic materials applications. The efficacy of the NNP model is shown to outperform the conventional cluster expansion (CE) model, exhibiting near-DFT accuracy and robust transferability to unexplored crystal structures and compositions. Furthermore, we employ the NNP model in simulated annealing (SA) optimizations to predict the low-energy Cr(₁−ₓ)S structures across diverse compositions. A notable structural transition is discerned at the Cr₀.₅S composition, characterized by a preferential migration of half of the Cr atoms to the van der Waals (vdW) gaps. This highlights the experimentally observed non-vdW nature of CrS₂ and emphasizes the pivotal role of excess Cr atoms beyond the composition ratio of Cr/S = 1/2 in stabilizing the vdW gaps. Additionally, we employ the NNP model in a large-scale vacancy diffusion Monte Carlo (MC) simulation to emphasize the impact of lateral compressive strains in catalyzing the formation of vdW gaps within 2D CrS₂ slabs. This provides a direct pathway for more facile exfoliation of ultrathin CrS₂ nanosheets through strain engineering.