Evaluating approaches for on-the-fly machine learning interatomic potentials for activated mechanisms sampling with the activation-relaxation technique nouveau

In the last few years, much effort has gone into developing general machine-learning potentials capable of describing interactions for a wide range of structures and phases. Yet, as attention turns to more complex materials, including alloys and disordered and heterogeneous systems, the challenge of providing reliable descriptions for all possible environments becomes ever more costly. In this work, we evaluate the benefits of using specific vs general potentials for the study of activated mechanisms in solid-state materials. More specifically, we test three machine-learning fitting approaches using the moment-tensor potential to reproduce a reference potential when exploring the energy landscape around a vacancy in Stillinger-Weber silicon crystal and silicon-germanium zincblende structures using the activation-relaxation technique nouveau (ARTn). We find that a targeted on-the-fly approach specific to and integrated into ARTn generates the highest precision on the energetics and geometry of activated barriers while remaining cost-effective. This approach expands the types of problems that can be addressed with high-accuracy ML potential.

Automated summary

In recent years, there has been a lot of effort put into developing general machine-learning potentials for describing various structures and interactions. However, as the focus shifts to more complex materials, it becomes increasingly expensive to provide reliable descriptions for all possible environments. In this study, we compare the benefits of using specific potentials versus general potentials for studying activated mechanisms in solid-state materials. We found that a targeted on-the-fly approach integrated into the activation-relaxation technique nouveau (ARTn) generates the highest precision while remaining cost-effective.