A universal ML model for segregation in W

Abstract

Segregation of solute elements to grain-boundaries (GB) in alloys is a key process controlling material properties. Examples are phase transformations, strength, or nanocrystalline stability. The central quantities to predict GB segregation are the site-specific segregation energies which can be accurately calculated using density functional theory (DFT). To reduce the computational cost, machine learning (ML) models are trained on DFT segregation data to predict the segregation energies. Here, we combine descriptors for the local structure of the segregation site with element-specific parameters for the solute element to train ML models that can predict the site-specific segregation energies for a wide range of elements. We use cross-validation and extrapolation scores to find the optimal set of descriptors for the model. The thus obtained model is then used to predict the segregation energies of solutes that are not in the data set. We apply our approach to segregation of transition metals in W. Both, cross-validation scores and comparison to literature data highlight excellent results of the ML approach. We make the model available by publishing the relevant codes and data.