Understanding the strength and deformation of materials at the atomic level is crucial for the desirable materials design. First-principle calculation and molecular dynamics method (MD) are widely used to obtain information on the time evolution of atoms or molecules. In first-principle calculations, since the behavior of atoms is investigated based on the state of electrons, accurate analysis of various substances is possible. Thanks to recent improvements in computer performance, first-principle calculations based on density functional theory have been greatly speeded up; however, the size of a realistic model is still of about several hundred atoms. MD uses the interatomic potential, which is a function that parameterizes interatomic interactions. Compared to first-principle calculations, MD can treat larger time scales and tens to millions of atoms. In order to parameterize the potentials, potential fitting is performed using data from first-principle calculations such as bond lengths, angles between atoms, lattice constants, and elastic constants, or other experimental data. By using these potentials, it is possible to reduce the time significantly compared to the first-principle calculations and to obtain accurate analysis results for simple models to some extent. However, for the heterogeneous internal structure and inhomogeneous deformation around the defects in materials, the accuracy is not sufficient. In recent years, potential development methods using machine learning have attracted tremendous attention ^[1][2]. Even if the function shape of the potential is unknown, it still can be fitted to the machine learning model, so it is expected to be highly versatile and maintain accuracy. In this study, we developed a magnesium (Mg) interatomic potential by machine learning method based on the data sets obtained from first-principle calculations. We applied the potential to molecular dynamics calculations and compared the results of molecular dynamics calculations with existing empirical potentials and the results obtained from first-principle calculations to evaluate the newly developed machine learning potential. Our results indicated that the calculation accuracy of our machine learning potential is higher than the existing potential and comparable to the first-principle calculation.

Keywords: Deep learning, Interatomic potential, Magnesium, Molecular dynamics, First-principle calculation

References

[1]    T. Wen, C. Z. Wang, M. J. Kramer, Y. Sun, B. Ye, H. Wang, X. Liu, C. Zhang, F. Zhang, K. M. Ho, N. Wang, “Development of a deep machine learning interatomic potential for metalloid-containing Pd-Si compounds”, Phys. Rev. B, 100 (2019).

[2]    V. L. Deringer, M. A. Caro, G. Cs´anyi, “Machine Learning InteratomicPotentials as Emerging Tools for Materials Science”, Adv. Mater., 31 (2019).