Machine learning a neural network magnesium potential

Magnesium is a desirable lightweight structural material but lacks the ductility needed for fabrication and performance. To overcome its limitations, the strongly anisotropic and complex dislocation plasticity in Magnesium and its alloys must be understood. This requires interatomic potentials that capture a wide range of subtle properties including small differences in various dislocation energies. While a very good MEAM-type potential exists for pure Mg, its transferability to alloys has been insufficient to date. Here, we thus first pursue development of a machine learning potential for pure Mg. We use a neural network (NN) framework with the Behler-Parinello symmetry functions to describe atomic environments, with training data obtained from extensive first-principles DFT calculations on metallurgically-relevant properties. We first limit the training data to the same set of data used to fit the MEAM potential and thus study the ability of the NN potential to achieve results comparable to the MEAM potential for equal inputs. We demonstrate broad success of the NN potential, as compared to the MEAM potential. Subsequently we include further data in the training and discuss aspects of the NN potential that might be improvable, and discuss extensions of the general approach to the important Mg-Y alloy.