Electronegativity Informed Graph Neural Networks for Superconducting Temperature Prediction with Generative Crystal Validation

Accurate prediction of the superconducting critical temperature ( T c ) remains a major challenge in data-driven materials discovery. Here, we develop an electronegativity (EN) informed graph neural network framework and systematically compare modified MEGNet and CGCNN models with node-, global-, and edge-level electronegativity descriptors. Among the tested strategies, nonlinear radial basis function encoding of edge electronegativity differences gives the best performance. After Bayesian hyperparameter optimization, the mCGCNN-EΔEN-rbf-HPO model achieves a test RMSE of 8.02 K and R 2 of 0.824 on the 3DSC data set, outperforming the baseline models. We then combine the optimized predictor with the crystal generative models CDVAE and CrystaLLM to screen 33234 valid generated structures, from which promising high- T c candidates are identified. First-principles electronic-structure analysis further shows that reconstruction of the local Cu–O coordination environment and enhanced electronic states near the Fermi level are associated with higher T c . This work establishes a physically motivated and transferable workflow that connects chemical descriptor design, graph-based prediction, crystal generation, and first-principles validation for superconducting materials discovery.