Revealing nanoscale local chemical environments governing diffusion within binary concentrated alloys through machine learning

Bulk diffusion in binary and multi-component alloys is governed by correlated nanoscale mechanisms that strongly depend on the configuration of alloying elements near vacancies. Here, we present a numerical framework wherein a deep-learning algorithm supervised by atomistic-scale simulations is used to explore the immense configurational breadth of alloys near vacancies. We apply this framework to predict associated energies for the exchange of atomic sites with vacancies. For a model Ni_1-xAl_x system, our approach has less than 1 meV error in prediction of the formation energy, while being trained by data representing only 10% of the total compositional space. We reveal that in a Ni matrix, exchanging sites of vacancies with Al atoms has a higher average energy barrier and formation energy than for Ni atoms, and this governs the mobility and interdiffusion coefficient of Al in Ni. Moreover, the propensity of Al to form short-range-order near vacancies correlates with the generalized stacking fault energy of configurations with mobile Ni atoms. Future applications of this framework include studying the diffusivity of multi-component alloys as well as crystals under different mechanical and thermal boundary conditions.