A neural-network potential for aluminum

Aluminum and its alloys are most often used as structural materials due to their specific properties, such as low weight, low energy consumption for remelting and the possibility of almost complete processing. This paper utilizes machine learning, specifically the Behler-Parrinello neural network scheme, to develop a powerful potential for studying the underlying mechanisms of deformation, fracture, and defect formation. By surpassing the limitations of first principles calculations, the application of machine-learned potentials (MLP) becomes highly advantageous for describing pure aluminum (Al) in its solid and liquid phases. Specifically, from the generated potential equilibrium, thermodynamic, elastic, and vibrational properties of face-centered cubic (fcc) Al are obtained in very good agreement especially with density functional theory (DFT) results as well as with previous calculations using existing semi-empirical potentials, such as EAM and MEAM, recent machine-learned potentials, and experimental data. Furthermore, our potential proves to accurately reproduce defect formation energies such as previously computed and measured stacking-fault energy curves. Finally, stacking fault profiles as well as key quantities of the liquid phase such as the melting point at ambient pressure, temperature-dependent densities, and radial distribution functions are also calculated in very good agreement with the results from previous theoretical and experimental investigations. Nevertheless, our investigation goes beyond previous studies in proving excellent agreement with experimental data especially of the specific heat and the melting point at very high pressures. The competitive analysis performed in this work thus clearly demonstrates the validity and accuracy of the generated machine-learned potential to describe a wide range of properties of Al at various temperatures and pressures and thereby lays ground for future applications.