Role of descriptors in predicting the dissolution energy of embedded oxides and the bulk modulus of oxide-embedded iron

Abstract

Oxide-embedded bulk iron is investigated in terms of first principles calculations and data mining. Twenty-nine oxides are embedded into a vacancy site of iron where first principles calculations are performed and the resulting calculations are stored as a data set. A prediction of the dissolution energy of oxides within iron and the bulk modulus of oxide-embedded iron is performed using machine learning. In particular, support vector machine (SVM) and linear regression (LR) are implemented where descriptors for determining the dissolution energy and bulk modulus are revealed. With trained SVM and LR, the prediction of the dissolution energy for different oxides in iron and the inverse problem---deriving the corresponding descriptor variables from a desired bulk modulus---are achieved. The physical origin behind the chosen descriptors is also revealed where manipulating each individual descriptor within a multidimensional space allows for the prediction of the dissolution energy and bulk modulus. Thus, predictions of physical phenomena are, in principle, achievable if the appropriate descriptors are determined.