Exploring transformations in elemental sulfur using a machine-learned interatomic potential

We investigate liquid polymorphism in elemental sulfur through atomic-scale simulations, specifically using molecular dynamics, with an emphasis on both the lambda and liquid-liquid transitions. We propose a general DFT-ML workflow designed to study liquid polymorphism, aiming to develop a machine-learned interatomic potential (MLIP) that facilitates an efficient exploration of phase diagram regions.

Our workflow incorporates various MLIP architectures tailored to the complexity of the tasks, alongside active learning protocols. For instance, we use efficient models in the form of Behler-Parrinello neural networks for structure selection through query-by-committee, and equivariant MLIPs for the final regression task. To improve the stability, accuracy, and generalization of the final MLIP, we address several factors: analytically managing van der Waals interactions, calculating potential energies and forces on the training dataset using hybrid DFT, and employing a recursive training strategy.

With our developed model, we achieve system sizes and timescales that are otherwise unattainable with DFT. We explore specific zones of the phase diagram in the surroundings of the transition lines using unbiased molecular dynamics, as well as simulations involving enhanced sampling along specific collective variables. Concerning the liquid-liquid transition, we compare our findings with experimental data, as well as with results obtained at the DFT-LDA level of theory,