Just another WordPress site - Ruhr-Universität Bochum

Abstract

The prediction and extrapolation of material properties of yet-unknown materials is a significant challenge in materials discovery. We present three models fitted with two ternary materials property datasets to predict performance within their shared quaternary system. The first uses standard Gaussian Process Regression (GPR) based on elemental compositions. The second is also a GPR, however, not based on elemental composition but a representation of ‘composition’ in form of word embedding-derived materials vectors.  The third model advances the methodology by employing the ‘standard vector method’, which synthesizes weighted word embedding-based vector representations of material properties to establish a benchmark vector, facilitating the prediction of material performance in the quaternary system by assessing similarity to this idealized reference. With this we demonstrate the effectiveness to integrate meaningful word embedding-based representations of materials when materials data is scarce. Such robust predictions of higher-dimensional compositions spaces allows effective screening of the materials space based on available lower-dimensional composition-property spaces.