Graph Neural Network‐Guided and Interpretable Discovery of Supercapacitor Electrode Materials

Abstract

Discovering electrode materials that combine high capacitance, stability, conductivity, and practical synthesis remains a significant challenge for advancing supercapacitors beyond the energy‐density limits of traditional carbon‐based systems. Here, we introduce an AI‐driven framework integrating graph neural network‐based virtual screening with interpretable capacitance modeling to identify promising inorganic electrode candidates. Utilizing MEGNet and an attention‐augmented MAGNET architecture, we screened over 22 000 crystal structures derived from density functional theory, employing formation energy and energy above the convex hull as descriptors of thermodynamic stability and synthesizability. Electronic conductivity was assessed using a conservative bandgap filter based on database data, avoiding the low‐accuracy GNN band‐gap predictor. Candidate ranking was refined via Pareto scoring, Monte Carlo Dropout uncertainty analysis, and capacitance classifiers trained for MXenes, metal oxides, and carbon‐based electrodes. The approach yielded strong cross‐project validation: 45 of top 50 GNN‐selected candidates were confirmed as high‐capacitance materials, aligning with independent models at 90%. Notably, the framework recaptured known high‐performance MXenes and identified CeMoO 4 F as a novel pseudocapacitive candidate with favorable stability and potential multi‐redox charge storage. This integrated, reproducible workflow demonstrates the power of AI for accelerated supercapacitor materials discovery, linking high‐throughput screening with uncertainty quantification and interpretability.