Oxidation-aware hybrid tabular-graph neural network for intercalation voltage prediction across multiple working-ion chemistries.

Accurate prediction of average intercalation voltage is critical for battery performance, as voltage directly determines energy density and device efficiency. Experimental measurements and density functional theory calculations, however, are computationally expensive for large scale cathode screening. Here, an oxidation-aware hybrid tabular-graph neural network (HTGNN) is introduced to predict intercalation voltages directly from crystallographic structures and redox descriptors. A unified crystal graph was constructed from the discharged-state structures of 5355 intercalation electrodes from the materials project battery explorer, and consistent representations were employed across the benchmark architectures GCN, GraphSAGE, GAT, transformer-GNN, CGCNN, and MEGNet, yielding incremental improvements up to a test MAE of 0.43 V. The proposed HTGNN integrates a three-layer TransformerConv graph encoder with a tabular MLP, achieving a test MAE of 0.39 V, an RMSE of 0.54 V, and an R2 of 0.76 on a 70/15/15 data split, outperforming structure-only baselines. Ion-specific analysis demonstrates high accuracy for Li (0.35 V), Na (0.41 V), Ca (0.37 V), Mg (0.42 V), and Zn (0.45 V), with lower performance for sparsely represented K systems. t-SNE clustering highlights ion-dependent electrochemical patterns, while parity plots confirm robust generalization. By explicitly incorporating oxidation-state information alongside structural representations, HTGNN offers a physically grounded, low-cost surrogate model suitable for rapid screening of multivalent cathodes, including Li, Na, K, Mg, Ca and Zn, prior to computationally expensive first-principles calculations.