Fig. 1: Workflow for constructing a training dataset for machine learning of atomic potentials in the Fe-H binary system.

a Initial training dataset is created using DFT (an example of a structure is shown in the figure). b Four deep potentials (DPs) are learned using different initial parameters. c Using one of the DPs, molecular dynamics calculations are performed under multiple conditions (temperature, pressure, etc.) for atomic structures for exploration, and many atomic structures can be created. Evaluate the force error between the four DPs for these atomic structures and select structures whose error exceeds a certain threshold. d Perform DFT calculations on the selected atomic structures and obtain additional learning data. Perform learning again, including the additionally obtained learning data. Repeat these cycles until the error after the exploration step reaches an acceptable level. In this study, we comprehensively performed this workflow on a series of atomic structures related to hydrogen embrittlement (examples of structures are shown in the figure, including hydrogen-containing bulk, grain boundaries, surfaces, and generalized stacking faults).