Molecular dynamics simulations of self-diffusion in liquid metals using a universal neural network potential

The objective of this study is to establish calculation conditions for self-diffusion coefficients of liquid metals by optimizing their densities through molecular dynamics (MD) simulations with a universal neural network potential. The self-diffusion coefficients of liquid Al, Cu, Zn, Ga, Ag, In, Sn, Sb, and Pb just above the melting points were calculated under various density conditions. The density conditions were prepared either from experimental data or via MD simulations. The calculated self-diffusion coefficients of liquid Al, Cu, Zn, Ga, Ag, In, and Sb agreed with the experimental data, which were obtained under microgravity and on the ground, within the error range of ±20% under the condition of experimental densities. The calculated self-diffusion coefficients in the selected liquid metals, except for Sb and Pb, agreed with the experimental data within the error range of ±22% under the condition of calculated densities. For Al and Cu (category Ⅰ), the densities should be calculated without the dispersion correction, while those for elements heavier than Cu (category Ⅱ) should be calculated by adopting the dispersion correction. For liquid Sb and Pb (category Ⅲ), the calculated interatomic distances deviated from the experimental data, which is assumed to cause the deviation of the self-diffusion coefficients from the experimental data. The present MD simulations can accurately estimate the self-diffusion coefficients by adopting the calculated densities, for liquid metals in the categories Ⅰ and Ⅱ.