Boosting Superconductor Discovery: CrystalGAT Achieves High-Accuracy Critical Temperature Prediction Through Local-Global Crystal Graph Learning

This study aims to predict the critical temperature (Tc) of superconducting materials and determine their superconducting state using our novel graph neural network model, CrystalGAT. This model skillfully integrates graph attention networks and convolutional neural layers to effectively capture local interactions between atoms and global material properties. Our research is based on the 3DSC_MP dataset, which contains three-dimensional crystal structures of superconductors and their corresponding Tc data. The CrystalGAT model processes atomic features, bond features, angular features, and chemical composition features (MAGPIE features) to learn complex interactions between atoms and bonds, as well as global properties of the entire chemical composition. Experimental results demonstrate that CrystalGAT achieves significant improvements in predictive performance compared to traditional models, exhibiting lower error metrics and higher R² values. This strongly highlights the immense potential of integrating detailed atomic-level and structural characteristics into superconductor prediction models. Our research provides new insights and tools for the discovery and design of high-temperature superconducting materials.