Structure and dynamics of sulfur vacancies in monolayer MoS2 studied by DFT-based machine learning potentials

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Accurate modeling of activated processes, like defect migration in 2D materials, is crucial.
• Existing machine learning potentials (MLPs) often struggle with out-of-domain (OOD) extrapolation, limiting their reliability.

Purpose of the Study:

• To develop a robust multi-step strategy for training stable and precise MLPs.
• To ensure MLPs perform well in both in-domain interpolation and OOD extrapolation regimes.
• To apply and validate this strategy for vacancy dynamics in 2D materials.

Main Methods:

• Designed a sampling technique using nudged elastic band and constrained molecular dynamics for dataset creation.
• Developed tailor-made metrics focusing on critical atoms for defect migration.
• Benchmarked the MACE MLP model on chalcogen vacancy dynamics in monolayer MoS2, assessing OOD performance.

Main Results:

• A properly trained MACE MLP reliably reproduced energies and forces even in extreme OOD configurations.
• Calculated relaxations, minimum energy paths, and free energy barriers for vacancy transitions in MoS2.
• Demonstrated the utility of the developed training strategy for defect processes.

Conclusions:

• The developed multi-step strategy enables training of stable and precise MLPs for activated processes.
• The approach shows excellent performance for defect dynamics in 2D materials, including OOD generalization.
• Findings are applicable to other equivariant message passing neural network potentials and suggest pathways to quantum Monte Carlo accuracy.