Data-driven prediction of chemically relevant compositions in multi-component systems using tensor embeddings.

The discovery of novel materials is crucial for developing new functional materials. This study introduces a predictive model designed to forecast complex multi-component oxide compositions, leveraging data derived from simpler pseudo-binary systems. By applying tensor decomposition and machine learning techniques, we transformed pseudo-binary oxide compositions from the Inorganic Crystal Structure Database (ICSD) into tensor representations, capturing key chemical trends such as oxidation states and periodic positions. Tucker decomposition was utilized to extract tensor embeddings, which were used to train a Random Forest classifier. The model successfully predicted the existence probabilities of pseudo-ternary and quaternary oxides, with 84% and 52% of ICSD-registered compositions, respectively, achieving high scores. Our approach highlights the potential of leveraging simpler oxide data to predict more complex compositions, suggesting broader applicability to other material systems such as sulfides and nitrides.