Machine-learning-based molecular dynamics (ML-MD) has emerged as a promising approach to resolving the trade-off between accuracy and computational cost in atomic simulations. The most conventional method in ML-MD, total energy learning (TEL), trains models by matching the sum of predicted atomic energies to the total energy of the system. However, because calculating total energy requires information from all atoms in the system, TEL poses limitations for flexible data refinement. To address this issue, we have developed atomic energy learning (AEL), which enables atomic-level data selection by directly using each atom’s energy as training data. As a fundamental study, we compare AEL and TEL using datasets based on the Embedded Atom Method (EAM). The results demonstrate the effectiveness of selective data sampling in controlling model accuracy and reveal the advantage of AEL in evaluating atomic energies near defects through targeted data selection.
KATO et al. (Sun,) studied this question.