Strategy to consider element distribution when constructing training datasets for developing machine learning potentials of alloys based on a Monte-Carlo-like method

Molecular dynamics simulations can explore the characteristics and evolution of microstructures in alloys outside of experiments, with reliability and accuracy guaranteed by the interatomic potentials employed. Machine learning potential (MLP) is widely used for its accuracy close to first-principles calculations. When developing an MLP, the construction of the training dataset is crucial, determining the accuracy and generalization of the MLP. In this work, a Monte-Carlo-like (MCL) strategy is proposed to construct training datasets for developing MLPs of alloys, which is characterized by the efficient consideration of element distributions in alloys. As an example, a training dataset for the equimolar NbTiZrHf alloy is constructed based on the MCL strategy, and the corresponding MLP is developed subsequently. By comparing with two traditional strategies, it is found that the training dataset constructed based on the MCL strategy has greater dispersion, and the corresponding MLP has better prediction performance. In addition, a hybrid molecular statics and Monte Carlo simulation with the MCL-based MLP is performed to optimize the element distribution of the equimolar NbTiZrHf alloy, and segregation and short-range ordered structures are observed in the final configuration, which is consistent with the experimental results reported in the literature. The MCL strategy proposed in this work can provide a fast solution for considering the element distribution when constructing training datasets for developing MLPs of alloys.