Crystal synthesizability prediction using contrastive positive unlabeled learning

High-throughput screening or generative models rapidly identify crystal structures with the desired properties, but the synthesizable ratio is generally low. Experimentally verifying the synthesizability of individual virtual crystals would entail significant time and resources. Therefore, a method for automatically assessing the synthesizability of virtual crystals is urgently needed. This paper describes an approach that combines contrastive learning and positive unlabeled learning. The resulting contrastive positive unlabeled learning (CPUL) model predicts the crystal-likeness score (CLscore) of virtual materials. The model achieves a true positive (CLscore > 0.5) prediction accuracy of 93.95% on a test set containing 10,000 materials taken from the Materials Project (MP) database. We further validate the model by using all Fe-containing materials from the MP database as the test set, obtaining a true positive rate of 88.89%. This indicates that the CPUL model performs well, even with limited knowledge of the interactions between Fe and the atoms in the crystals. The CPUL model is then used to assess the CLscore of virtual crystals in the MP database and analyze their synthesizability by combining the energy above the hull. Finally, the synthesizability of perovskite materials is predicted using the proposed CPUL model, resulting in seven candidate halide perovskite materials for photovoltaic applications.