Solid-state synthesizability predictions using positive-unlabeled learning from human-curated literature data

Abstract

The rate of materials discovery is limited by the experimental validation of promising candidate materials generated from high-throughput calculations. Although data-driven approaches, utilizing text-mined datasets, have shown some success in aiding synthesis planning and synthesizability prediction, they are limited by the quality of the underlying datasets. In this study, synthesis information of 4103 ternary oxides was extracted from the literature, including whether the oxide has been synthesized via solid-state reaction and the associated reaction conditions. This dataset provides an opportunity to supplement existing solid-state reaction models via reliable data and information from articles whose content and formats are challenging to extract automatically. A simple screening using this dataset identified 156 outliers from a subset of a text-mined dataset that contains 4800 entries, of which only 15% of the outliers were extracted correctly. Finally, this dataset was used to train a positive-unlabeled learning model to predict the solid-state synthesizability of new ternary oxides, where we predict 134 out of 4312 hypothetical compositions are likely to be synthesizable.