Citation: F.-S. Meng, J.-P. Du, S. Shinzato, H. Mori, P. Yu, K. Matsubara, N. Ishikawa, and S. Ogata (2021), "General-purpose neural network interatomic potential for the 𝛼-iron and hydrogen binary system: Toward atomic-scale understanding of hydrogen embrittlement",
Physical Review Materials,
5(11), 113606. DOI:
10.1103/physrevmaterials.5.113606.
Abstract: To understand the physics of hydrogen embrittlement at the atomic scale, a general-purpose neural network interatomic potential (NNIP) for the 𝛼-iron and hydrogen binary system is presented. It is trained using an extensive reference database produced by density functional theory (DFT) calculations. The NNIP can properly describe the interactions of hydrogen with various defects in 𝛼-iron, such as vacancies, surfaces, grain boundaries, and dislocations; in addition to the basic properties of 𝛼-iron itself, the NNIP also handles the defect properties in 𝛼-iron, hydrogen behavior in 𝛼-iron, and hydrogen-hydrogen interactions in 𝛼-iron and in vacuum, including the hydrogen molecule formation and dissociation at the 𝛼-iron surface. These are superb challenges for the existing empirical interatomic potentials, like the embedded-atom method based potentials, for the 𝛼-iron and hydrogen binary system. In this study, the NNIP was applied to several key phenomena necessary for understanding hydrogen embrittlement, such as hydrogen charging and discharging to 𝛼-iron, hydrogen transportation in defective 𝛼-iron, hydrogen trapping and desorption at the defects, and hydrogen-assisted cracking at the grain boundary. Unlike the existing interatomic potentials, the NNIP simulations quantitatively described the atomistic details of hydrogen behavior in the defective 𝛼-iron system with DFT accuracy.
Notes: Fan-Shun Meng notes that "This potential was designed for the general purpose usage of the 𝛼-Fe-H binary system. Additionally, the potential also can be used for pure 𝛼-Fe. To use the potential in LAMMPS, the pair_style of hdnnp should be adopted, and the package of ML-HDNNP should be compiled(see ML-HDNNP documentation)."