Influence of exchange–correlation functional choices on machine learning potential accuracy in the coupled PWDFT-DeePMD framework

Abstract

Machine learning potentials offer a promising approach for large-scale first-principles calculations. However, the accuracy of models derived from different Jacob’s ladder levels significantly affects their predictive performance, as the quality of the training dataset plays a crucial role in model effectiveness. Therefore, generating a sufficiently large and diverse dataset for training machine learning potentials remains a major challenge. In this work, we couple plane-wave density functional theory (PWDFT) with deep potential molecular dynamics (DeePMD), utilizing the rapid and accurate hybrid functional calculations within PWDFT to generate diverse training sets. This coupling enables us to systematically assess the impact of different functional-based training sets on machine learning potentials within the plane-wave basis set, thus improving computational efficiency and model robustness. We find that local and semi-local functionals are more suitable for solid systems, while hybrid functionals perform better for complex systems like molecules. This observation underscores the importance of selecting appropriate functionals for specific systems to enhance the accuracy and reliability of model predictions.