Calculation of dislocation binding to helium-vacancy defects in tungsten using hybrid ab initio-machine learning methods

Calculations of dislocation-defect interactions are essential to model metallic strength, but the required system sizes are at or beyond ab initio limits. Current estimates thus have extrapolation or finite size errors that are very challenging to quantify. Hybrid methods offer a solution, embedding small ab initio simulations in an empirical medium. However, current implementations can only match mild elastic deformations at the ab initio boundary. We describe a robust method to employ linear-in-descriptor machine learning potentials as a highly flexible embedding medium, precisely matching dislocation migration pathways whilst keeping at least the elastic properties constant. This advanced coupling allows dislocations to cross the ab initio boundary in fully three dimensional defect geometries. Investigating helium and vacancy segregation to edge and screw dislocations in tungsten, we find long-range relaxations qualitatively change impurity-induced core reconstructions compared to those in short periodic supercells, even when multiple helium atoms are present. We also show that heliumvacancy complexes, considered to be the dominant configuration at low temperatures, have only a very weak binding to screw dislocations. These results are discussed in the context of recent experimental and theoretical studies. More generally, our approach opens a vast range of mechanisms to ab initio investigation and provides new reference data to both validate and improve interatomic potentials.

Automated summary

Calculations of dislocation-defect interactions are difficult due to the limitations of ab initio simulations. Hybrid methods, such as linear-in-descriptor machine learning potentials, offer a solution by embedding ab initio simulations in an empirical medium. This allows for more accurate modeling of dislocation migration pathways and defect geometries.