The discovery of novel functional materials is a primary goal in materials science. High-throughput screening (HTS) has emerged as a powerful tool, leveraging computational methods and databases. HTS often begins with screening density functional theory (DFT) databases, followed by higher-level DFT refinements and experimental validation. Databases like Materials Project, OQMD, and AFLOW contain numerous "virtual crystals"—ground-state structures predicted computationally but not yet synthesized. While some virtual crystals have been successfully synthesized, assessing synthesizability remains a significant challenge. Thermodynamic metrics like energy above the convex hull are insufficient because they neglect kinetic factors and synthesis conditions. Therefore, a generalized and reliable method for predicting synthesizability is crucial to accelerate materials discovery. Predicting synthesizability is complex because negative data ("unsynthesizable" crystals) are unavailable. Positive-unlabeled (PU) learning methods address this by utilizing data on synthesized crystals (positive) and unlabeled virtual crystals (potentially synthesizable). Previous work used a transductive PU-learning method with crystal graph convolutions to predict a "crystal-likeness" (CL) score. While showing promise, the accuracy for specific domains, particularly perovskites (around 74%), was lower than the overall accuracy. Perovskites are important due to their applications in photovoltaics, light-emitting diodes, magnetic materials, superconductors, and Li-ion conductors, making accurate synthesizability prediction invaluable. Existing perovskite-focused models often rely on heuristics like the Goldschmidt tolerance factor or machine learning techniques such as gradient boosting decision trees or support vector machines. However, these methods are frequently limited to metal oxide perovskites and depend on the Shannon ionic radii database, making them less applicable to perovskites with more covalent bonding or anti-perovskites. A generalized deep learning model trained on a diverse dataset could overcome these limitations. Domain-specific learning can improve model accuracy within a particular material class, suggesting its potential benefits for perovskite synthesizability prediction. Transfer learning provides a strategy for training deep neural networks with limited data by using pre-training on a larger, related dataset and fine-tuning the model with domain-specific data.

The literature review section highlights the existing methods for predicting the synthesizability of materials, specifically focusing on perovskites. Existing methods, such as those based on the Goldschmidt tolerance factor and other rule-based approaches, are discussed and their limitations are identified. The authors also review previous machine learning approaches, including gradient boosting decision trees, support vector machines, and random forest classification, pointing out their limitations in generalizability and applicability to different types of perovskites (e.g., those with covalent bonding or anti-perovskites). The review emphasizes the need for a more accurate and generalized model, particularly for a specific class of materials like perovskites.

The authors developed a synthesizability prediction model for perovskites using a combination of positive-unlabeled learning, domain-specific learning, and transfer learning. The model is a graph convolutional neural network (GCNN). Initially, the model was pre-trained on a large dataset of inorganic crystals from the Materials Project (MP) database, consisting of both experimentally synthesized crystals (positive) and virtual crystals (unlabeled). This pre-training step utilizes an inductive PU learning framework. The model then undergoes transfer learning where a smaller perovskite dataset extracted from MP, OQMD, and AFLOW databases is used to fine-tune the model. The perovskite dataset includes 943 synthesized and 11,964 virtual perovskite crystals. The StructureMatcher function in pymatgen was used to identify and remove duplicate crystals. A similar inductive PU learning framework is used for this transfer learning step. To evaluate the model's performance, 10% of the synthesized crystals were randomly selected and set aside as a test set, ensuring that this data was not used during pre-training. The features used in the model are derived from the crystal structure. Atom features are represented using one-hot encoding of elements, while edge features capture the distance and solid angle between atoms within a 7 Å radius, represented using Gaussian expansion. The model architecture is a GCNN with multiple layers of graph convolutions, dense layers with softplus activation, minimum pooling with sigmoid activation, and a final linear layer that outputs the synthesizability score (CL score). The Adam optimizer was used for training the model.

The trained model achieved a remarkable out-of-sample true positive rate of 0.957 for predicting perovskite synthesizability. This significantly improves upon the approximately 74% accuracy achieved by non-domain-specific models. Out of the 11,964 virtual perovskites, the model predicted 962 as synthesizable. A literature search revealed that 179 of these predicted synthesizable virtual crystals had indeed been synthesized previously, validating the model's accuracy. Conversely, a literature search for the 1000 virtual crystals with the lowest synthesizability scores yielded no synthesized cases. The model's ability to predict synthesizability extends beyond classical ionic perovskites to encompass anti-perovskites, covalent perovskites, halides, and hydrides, demonstrating its broader applicability compared to empirical models based on ionic radii. The authors identified promising synthesizable candidates for two applications: Li-rich anti-perovskites (8 candidates listed in Supplementary Table 2) as potential solid-state electrolytes and metal halide perovskites (12 candidates in Table 1) with band gaps suitable for photovoltaic applications. While some predicted materials show low thermodynamic stability, suggesting potential synthesis challenges despite high synthesizability scores, this highlights the complementary nature of CL scores and thermodynamic metrics.

The high accuracy achieved by the domain-specific transfer learning approach demonstrates its effectiveness in predicting perovskite synthesizability. The model's generalizability across various perovskite types and its superior performance compared to existing methods are key contributions. The identification of novel synthesizable candidates for Li-rich ion conductors and metal halide optical materials opens up avenues for experimental investigation and potentially new discoveries. The findings underscore the potential of combining computational predictions with experimental verification for accelerating materials discovery. While the model exhibits high accuracy, further work could focus on improving precision (reducing false positives). The combination of CL scores and thermodynamic stability metrics shows promise for improving the overall reliability of synthesizability predictions.

This study presents a novel graph neural network model that accurately predicts the synthesizability of perovskites. The model's high accuracy, generalizability, and successful identification of previously synthesized materials highlight the power of domain-specific transfer learning in materials discovery. The predicted synthesizable candidates for Li-rich ion conductors and metal halide optical materials provide valuable targets for future experimental investigations. Future research directions could focus on improving the model's precision and extending the approach to other material families.

The model's reliance on existing databases introduces potential biases based on the data available. The accuracy of synthesizability prediction might be influenced by factors not captured in the model's features, such as specific synthesis conditions. Some predicted synthesizable materials might be challenging to synthesize due to low thermodynamic stability, suggesting that the combined use of CL scores and thermodynamic metrics provides a more comprehensive approach.