Can we accurately calculate viscosity in multicomponent metallic melts?

TL;DR

This paper demonstrates that deep neural network potentials can accurately simulate the viscosity of ternary Al-Cu-Ni metallic melts, matching experimental data within 9% and capturing key features like the eutectic point.

Contribution

The study introduces a deep neural network potential approach for simulating viscosity in multicomponent metallic melts, achieving high accuracy and capturing critical concentration-dependent features.

Findings

Deviations from experimental viscosity are within 9%.

Simulations reproduce the viscosity minimum at the eutectic point.

Deep neural network potentials are effective for multicomponent metallic melt simulations.

Abstract

Calculating viscosity in multicompoinent metallic melts is a challenging task for both classical and \textit{ab~initio} molecular dynamics simulations methods. The former may not to provide enough accuracy and the latter is too resources demanding. Machine learning potentials provide optimal balance between accuracy and computational efficiency and so seem very promising to solve this problem. Here we address simulating kinematic viscosity in ternary Al-Cu-Ni melts with using deep neural network potentials (DP) as implemented in the DeePMD-kit. We calculate both concentration and temperature dependencies of kinematic viscosity in Al-Cu-Ni and conclude that the developed potential allows one to simulate viscosity with high accuracy; the deviation from experimental data does not exceed 9\% and is close to the uncertainty interval of experimental data. More importantly, our simulations reproduce minimum on concentration dependency of the viscosity at the eutectic point. Thus, we conclude that DP-based MD simulations is highly promising way to calculate viscosity in multicomponent metallic melts.