Neural Network Potentials for Nonstoichiometric Materials: a case study for chromium sulfides

Controlling the properties of nonstoichiometric semiconductors through modifying their structural vacancy compositions can lead to a plethora of applications in many disciplines of materials science. These materials can give rise to a lot of stable phases over a wide range of compositions, where the crystal structure is often unknown. The large size of the composition space and the associated structure space that need to be explored for crystal structure prediction makes first-principles methods ill-suited for this task, especially for compositions that require large numbers of atoms per unit cell. We here fit a neural network potential (NNP), trained on data from density functional theory (DFT), to generalize to the special quasi-random structures (SQS) of nonstoichiometric chromium sulfides, Cr_(1−x)S, over the full range of Cr vacancy concentrations. We indicate that NNP is able to accurately rank the SQS cells available at each composition and identify the ground-state structures, providing a superior performance compared to the conventional cluster expansion Hamiltonian. Furthermore, we show the ability of NNP to reproduce the DFT properties for the identified ground-state structures such as the vibrational properties, phonon dispersion relations, and equations of state. Finally, we use NNP to predict the exfoliation energy of multilayered Cr_(1−x)S slabs as a function of thickness and composition, to indicate the capability of exfoliating nanosheets from the bulk phases.