How to validate machine-learned interatomic potentials

Journal article

Meng F.S., Li J.H., Shinzato S., et al. Formation of three-dimensional dislocation networks in α-iron twist grain boundaries: Insights from first-principles neural network interatomic potentials. Computational Materials Science 253, 113812 (2025); https://doi.org/10.1016/j.commatsci.2025.113812.We conducted a systematic analysis of the atomic structure and energy of (001), (110), and (111) twist grain boundaries (TWGBs) in α-iron using a recently developed neural network interatomic potential (NNIP). This study showcases typical dislocation networks within TWGBs that exhibit small twist angles. Notably, we observed a three-dimensional (3D) dislocation network in (111) twist grain boundaries, primarily composed of ½〈111〉 dislocations—structures unattainable using previously proposed empirical potentials, hence unreported in earlier studies. The novel 3D dislocation network was further validated through several approaches, including principal component analysis (PCA), an NNIP ensemble model, and cross-validation with other machine learning interatomic potentials designed for α-iron. This breakthrough offers a new perspective on the properties of twist grain boundaries, potentially impacting our understanding of their strength, toughness, and mobility

Authors: Morrow, JD; Gardner, JLA; Deringer, VL

• Publication status: Published
• Peer review status: Peer reviewed
• Publisher: American Institute of Physics
• Journal: The Journal of Chemical Physics
• Volume: 158
• Issue: 12
• Pages: 121501-121501
• Article number: 121501
• Publication date: 2023-03-02
• DOI: 10.1063/5.0139611
• EISSN: 1089-7690
• ISSN: 0021-9606
• Language: English
• Keywords: Scale (ratio); Computer science; Machine learning; Atomic units; Quantum; Physics