Predicting which materials superconduct, and at what temperature, remains a central challenge in condensed-matter physics. Although machine-learning approaches have shown promise, their broader impact has been limited by the difficulty of incorporating crystal structure in a form that is both predictive and physically interpretable. Here I present a structure-aware, probabilistic machine-learning framework for superconductivity that combines graphlet-based representations of local chemical and structural environments with Gaussian-process learning[1]. Trained on empirical superconductivity databases, the model achieves state-of-the-art performance in T_c prediction while providing calibrated uncertainty estimates. Beyond predictive accuracy, it identifies a compact set of physically meaningful descriptors, with electron-affinity difference distributions emerging as a key predictor alongside interatomic distance. The framework recovers trends across known superconducting families, rediscovered superconductivity in nickelates without training on nickelate superconductors, and predicted superconductivity in stoichiometric PtPb_3Bi, subsequently confirmed experimentally with T_c in close agreement with the prediction. These results show that interpretable, structure-aware machine learning can extract organizing principles from heterogeneous superconductivity data and accelerate the discovery of new superconductors.

[1] O. Lesser et al, Electron affinity difference distributions guide the discovery of the superconductor PtPb3Bi, arXiv:2510.07373.