From Small Data to High Capacity: Symbolic Learning of Sodium-Ion Cathode Materials

Data-driven discovery of sodium-ion cathodes is often limited by small data sets and the poor interpretability of graph neural networks (GNNs). Here, we developed Symbol_ETR, an interpretable framework combining symbolic regression with ensemble learning. By constructing explicit nonlinear descriptors, symbolic regression extra trees regression (Symbol_ETR) achieves a test R2 of 0.90 for average voltage prediction and outperforms representative GNN models in the small-data regime. High-throughput screening of the Materials Project database identified Na5Co2S5 as a promising cathode candidate with a theoretical capacity of 340.86 mAh/g. First-principles calculations indicate a metallic electronic character and reveal weakly hybridized S 3p states near the Fermi level at Na-rich sulfur sites. Their evolution during desodiation supports sulfur-dominated charge compensation and anion redox activity. This work establishes an accurate and interpretable strategy for small-data material screening and the discovery of high-capacity sodium-ion cathodes.