A neural network potential for Magnesium and application to fracture
Magnesium’s low density and high specific strength makes it a very desirable structural material. Widespread adoption is impeded by its inherent brittleness. To understand the atomic scale properties and improve macroscopic ductility a reliable interatomic potential is required that captures a wide range of subtle properties . Further, it has to be extendable to alloys which has been insufficient to date within the MEAM (modified embedded atom method) framework . We present a neural network (NN) interatomic potential based on Behler-Parinello symmetry functions to describe atomic environments . It is fitted with extensive first-principles DFT calculations on metallurgically-relevant properties. The data set is limited to the same data set used to fit MEAM potentials and we study the ability to achieve comparable results for equal inputs. We demonstrate broad success of the NN potential in reproducing DFT properties, and compared to an existing very good MEAM potential. We discuss the scale bridging capabilities of the approach and extensions to important Mg-Y alloys.
 Z. Wu and W.A. Curtin, Nature 526 (2015) 62
 R. Ahmad, Z. Wu, W.A. Curtin, Acta Materialia 183 (2020) 228
 A. Singraber, J. Behler, C. Dellago, J Chem Theory Comput 15 (2019) 1827