Evaluating the applicability of classical and neural network interatomic potentials for modeling body centered cubic polymorph of magnesium

Abstract

Magnesium (Mg) is one of the most abundant metallic elements in nature and presents attractive mechanical properties in the industry. Particularly, it has a low density and relatively high strength/weight and stiffness/weight ratios, which make it one of the most attractive lightweight metals. However, the huge potential of Mg is restricted by its low ductility, associated with its hexagonal close packed (hcp) structure. This problem can be solved if Mg adopts the body centered cubic (bcc) structure, which is stable at high pressure or in confinement with stiff bcc metals like Nb. Molecular dynamics method is a magnificent tool to study material's structure and deformation mechanisms at the atomic level, however, requiring accurate interatomic potentials. The majority of most interatomic potentials available in the literature for Mg have only been fitted to the properties of its stable hcp phase. In the present work, we perform systematic study of applicability of currently available Mg potentials to modeling the properties of metastable bcc polymorph of Mg, taking into account cohesive energy curves, elastic constants, stacking fault energies, and phonon dispersion curves. We conclude that the modified embedded atom method (MEAM) potentials are the most suitable for investigating bcc Mg in Mg/Nb nano-composites, while the properties of high-pressure bcc Mg would be better modeled by neural network interatomic potentials after different local atomic environments corresponding to bcc Mg being included into the fitting database.

Systematic Atomic Structure Datasets for Machine Learning Potentials: Application to Defects in Magnesium

“…Based on these findings the authors concluded that these properties could be improved by specifically targeting bcc structures in the training set. [11] The fact that our potentials reproduce the dynamical as well as elastic properties so well without specifically targeting these values during the training set generation nicely illustrates the ability of our training approach to produce transferable potentials.…”

“…Our validation results also indicate a significantly better description of bcc Mg, both in the compressed high pressure state as well as at the equilibrium volume, as compared to other recently reviewed Mg potentials. The band structure and density of states for bcc Mg are shown in Appendix C. Troncoso et al [11] review this topic and find MEAM potentials are the best so far to study the dynamical behaviour of bcc Mg, but also report that the same potentials are deficient in their elastic properties. [28] Neural Network Potentials [2,39] reviewed in the same paper are found to be better for this application, but still predict wrong dynamical instabilities or predict them in the wrong part of the band structure.…”