Chemical Composition Data‐Driven Machine‐Learning Prediction for Phase Stability and Materials Properties of Inorganic Crystalline Solids

Materials informatics has attracted significant attention toward the efficient discovery and development of new functional materials. Machine-learning regression techniques have often been used to establish a link between properties. In this respect, precise and rational regression is contingent on the choice of descriptors used to represent both composition and structural information. The usage of structure-derived descriptors restricts the prediction range to the registered materials in the crystal structure database owing to the structural information requirements. Conversely, machine-learning regression based only on compositional descriptors is free from this restriction, despite the fact that the prediction performance may diminish. Herein, their prediction performance is improved using compositional descriptors with histograms, and detailed surveys are performed on their ability to extrapolate. The proposed model achieves a high prediction accuracy based on the area under the receiver operating characteristic (ROC) curve (specifically, AUC > 0.9).