A machine learning potential for simulation the dislocation behavior of magnesium

• A broadly applicable and stable deep learning potential for research the dislocation behavior of Mg was developed by combination of 468 density functional theory calculated data and machine learning method. • The developed deep learning potential accurately reproduces the SFE, and gives lattice constants, elastic constants, surface energies reasonable agreement with experimental or density functional theory calculation data. • The deep learning potential predicted basal, prismatic, pyramidal dislocations all agree well with DFT results on dissociation distance and core structures. Accurate predictions of the dislocation behavior of magnesium (Mg) by molecular dynamics (MD) simulations are essential for studying the fundamental mechanisms of deformation and designing high plasticity Mg alloys. However, existing atomic potentials in MD simulation for Mg are not sufficiently quantitative for many dislocations-associated phenomena, such as stacking fault energy (SFE) and dislocation core structures. Here, by combining 468 density functional theory (DFT) calculated data points and a machine learning method, we create a broadly applicable deep learning potential (DLP) to study the dislocation behavior of Mg. We demonstrate that our DLP reproduces the SFE, lattice constants, elastic constants, and surface energies in reasonable agreement with experimental or DFT data. Furthermore, the DLP predicted basal 〈 a 〉, prismatic 〈 a 〉, pyramidal 〈 c + a 〉 dislocations all agree well with DFT results on dissociation distance and core structures. Importantly, the DLP has a superior performance on distinguishing the pyramidal I and II 〈 c + a 〉 screw dislocation core structures. Our results show that the DLP is suitable for investigating the dislocation behavior of Mg, making it valuable for future realistic atomistic studies of general deformation problems.