A predictive model of impurity diffusion coefficients in face-centered-cubic metallic systems based on machine-learning

This paper develops models for diffusion coefficient prediction to provide parameters for atomic mobility databases and to assist material design in a multi-scale simulation framework for face-centered-cubic (fcc) alloys. Models of impurity-diffusion activation energy (Q_I) and self-diffusion activation energy (Q_s) are trained using machine-learning with experimental diffusion data and basic physical properties. The values of Q_s in body-centered cubic (bcc), fcc and hexagonal close-packed (hcp) can be well-predicted using melting temperature, electronic configuration, atomic properties and elasticity parameters. Estimates of Q_I in fcc metallic systems calculated using a model with six features agreed well with experimental data. Compared with previous models of Q_s and Q_I, the newly developed models exhibit higher coefficients of determination (R²) and significantly lower mean absolute errors. The self- and impurity-diffusion coefficients in fcc metallic systems can be simulated by these models. The models are also successfully applied during the assessment process of the Ni–Ti binary atomic mobility database. Thus, the developed models provide an easy and reliable method for estimating the self- or impurity-diffusion coefficients of fcc alloys when they are unavailable.