Machine-learned potential predicts how chromium sulfide layers rearrange during exfoliation

The work

Technical · from the record

The authors fitted a Behler-Parrinello neural network potential for nonstoichiometric chromium sulfides, Cr(1-x)S, using a dataset of 10,593 DFT structures built from cluster-expansion-sampled special quasi-random cells, strained bulk cells and slab cuts. The potential was tested against DFT energies and forces for configurations and ground-state structures at 18 compositions not used in training, and compared with a cluster expansion fitted to the same configurations, with lower energy RMSE and rank correlations above 0.96 reported for the potential. Driving simulated annealing, the potential reproduced the three known Cr-S ground-state phases in the range 0.25 <= x <= 0.5 and predicted that at Cr0.5S half of the Cr atoms sit in the van der Waals gaps, a non-vdW CrS2 arrangement whose DFT formation energy is 33 meV/atom below that of the vdW phase. A vacancy-diffusion Monte Carlo run on a 6,912-atom CrS2 slab under about 3.5% lateral compressive strain ended with roughly 97.8% of Cr atoms on sites characteristic of the vdW phase, with the slab expanding 0.56 nm along z.

How AI was used

A cluster expansion Hamiltonian was first fitted iteratively to special quasi-random cells enumerated on the Cr/vacancy sublattice of the NiAs-type Cr2S2 lattice, serving as a sampler of low-energy cells at 22 compositions whose DFT relaxation trajectories, thinned by farthest point sampling, formed the initial training set for a Behler-Parrinello neural network potential with separate atomic networks for Cr and S and symmetry-function descriptors, trained in n2p2 with an adaptive Kalman filter. The dataset was then grown by an active loop: the potential ran volume-constrained relaxations and NVT molecular dynamics on isotropically strained ground-state cells and on two-unit-cell-thick slab cuts, and a committee of two potentials with differing hyperparameters, seeds and splits flagged extrapolative snapshots for new DFT calculations. The fitted potential was used to rank roughly 130,000 enumerated configurations at 18 unseen compositions, to relax the 300 lowest-energy configurations per composition, and to drive short Monte Carlo swap runs with unconstrained sublattice occupation; a second cluster expansion fitted to static on-lattice DFT energies provided the baseline. The potential then supplied energies for twelve-cycle simulated annealing with Cr/vacancy swaps and cell perturbations followed by full relaxations, and for an NPT vacancy-diffusion Monte Carlo simulation of a 6,912-atom laterally strained CrS2 slab run for 210,000 steps.