Speaker
Description
We have developed a multi-step strategy for training stable and precise machine learning potentials (MLPs) that are able to efficiently and precisely extrapolate to out-of-domain (OOD) cases. Essential part of obtaining well-performing MLP is obtaining balanced and properly sampled dataset. To achieve this, we have developed a sampling technique based on nudged elastic band (NEB) and constrained molecular dynamics. We demonstrate the utility of our approach by calculating properties of (multi-)vacancies in MoS$_{2}$ monolayer such as structural relaxations, minimum energy paths, barriers and temperature-dependent free energy barriers. We evaluate precision of the model to great detail to find it's limitations in various cases. Usually, MLPs are evaluated using metrics such as RMSE on all the data and all the atoms at once, however, this does not show the full picture, since the most interesting configurations, such as those close to the transition states, are usually the ones with the highest error. We performed an extensive benchmarking of the MACE model in the context of vacancy dynamics in monolayer MoS$_{2}$ and checked the behavior of the model in most extreme cases for atoms close to the vacancy and in near-barrier configurations, which are the most difficult for the model to reproduce. Generally, we found that the MACE model is able to sufficiently reproduce energies and forces even in these extreme cases. We believe our conclusions are also valid for other symmetry-based message passing neural network potentials due to their general similarity.
