Our QM/MM implementation is used to learn grain-boundary segregation from first principles.

Recent Physical Review Letters paper published by Malik Wagih and Christopher A. Schuh describes a novel approach to build a machine learning model of grain boundary segregation energies of a solute atom in a metal polycrystal. The framework learns directly on the ab initio data, thus, bypassing the need for accurate interatomic potentials. To showcase the power of the approach the results of 36 solutes in Al are presented.

QM/MM method is used to provide ab initio accurate data on solute segregation energy for a given grain boundary, which is otherwise inaccessible by standard DFT methods due to system size limitations. Our force mixing scheme is used for relaxation of the system while final segregation energy is obtained from energy mixing scheme .

These are very exciting results that show the power of combination of the QM/MM and machine learning to build ab initio accurate models of solute interaction with extended defects and microstructural features. It opens vast opportunities to study fundamental properties of metals and alloys. We employ a similar philosophy in our QM/ML project to study interaction of point defects in dislocations.