EXTENDED ABSTRACT: Superconductors have a wide range of applications; however, currently discovered superconductors can only
achieve superconductivity at extremely low temperatures. The discovery of room-temperature superconductors has long been a dream for physicists and material scientists. Due to the unclear mechanisms of high-temperature superconductivity, guiding the discovery of new materials has been challenging. Experimental methods for finding high-temperature superconductors are time-consuming and costly, severely hindering the development of superconducting materials. Fortunately, after over a century of research, there is now more than 20,000 entries of superconducting material data available. Machine learning can utilize this existing data to search for new high-temperature superconductors in a shorter time and at a lower cost, providing material scientists with a new research approach.Currently, there is a wealth of work in superconducting machine learning that has achieved some success; however, most models are based on material composition information for predicting new superconductors. The factors that determine or influence the superconducting transition temperature (Tc) are numerous (e.g., structure, internal atomic interactions), making it difficult for Tc prediction models that consider only composition information to support the discovery of new high-temperature superconductors.To address the insufficiency of input information in current superconducting machine learning, this paper develops a novel graph neural network—GraphTTS—that can handle crystal structures and internal atomic interactions. Based on GraphTTS, the relationship between the internal interaction information of superconducting materials and Tc is explored, and new high-temperature superconductors are sought through the ICSD database.First, the GraphTTS model consists of three modules: Crystal Graph Representation (CGR), Complex Message Passing (CMP), and Attention. The CGR module represents crystallographic structures as graph structures that can be processed by machine learning. The CMP module simulates complex physical and chemical interactions between atoms. The Attention module captures key features related to superconducting properties.To improve the accuracy of superconducting predictions, a multi-level screening strategy is adopted, training three binary classification models for metal/nonmetal, superconductor/ non-superconductor, and high-temperature superconductor/low-temperature superconductor. These are used for progressive auxiliary searches for metals, superconductors, and high-temperature superconductors, respectively. A regression model is then employed to predict the Tc of high-temperature superconductors retained after classification. The model is validated using superconducting data discovered in the past five years (unknown data), achieving an AUC of 0.99 for classification models and an R² of 0.82 for the regression model, proving reliability of the model.The model interpretation is conducted by constructing "hypothetical crystals" with different crystallographic parameters. Analysis of the predicted Tc variations with bond length, space group, and elements reveals: 1. Specific atomic interactions primarily determine the Tc values of superconductors, such as Ca-O in HgBa2Ca2Cu3Oy and H-H in LaH10. 2. The predicted Tc of superconductors increases with enhanced internal interactions (shorter bond lengths). 3. "Hypothetical crystals" containing F, Ca, H, and Cl often show higher predicted Tc values, indicating that materials containing these elements may have high-temperature superconducting potential.Finally, based on the GraphTTS model and the ICSD database, a search for new hightemperature superconductors was conducted, discovering several potential candidates, including NaB0.08Cl0.92H0.32, which has a predicted Tc of 227 K. The findings of this study demonstrate that the GraphTTS machine learning model, which integrates crystallographic structures and internal atomic interaction information based on graph neural network technology, possesses superior capabilities for uncovering superconducting mechanisms and predicting superconducting materials. Additionally, GraphTTS has a universal architecture that can be applied to the development of new materials and mechanism research in other material systems.
Keywords: graph neural network, machine learning, superconductivity, superconductors, transition temperature