Graph neural networks and machine learning interatomic potentials for voltage prediction and screening high energy density sodium-ion cathode materials

Sodium-ion batteries are gaining increasing attention, driven by sodium’s natural abundance and promising performance for large-scale energy storage. In this work, we evaluate the performance of machine-learning interatomic potentials and Materials Project-trained models in predicting average voltages of electrode materials, demonstrating their potential for screening high energy density cathode materials for sodium-ion batteries. Initially, we tested dual-branch machine learning models for predicting average voltage of electrodes using the Materials Project dataset, spanning tree-based models, deep neural networks, a domain-adapted large language model, and graph neural networks. The results showed that the graph transformer model achieved the highest R 2 in the MP dataset. Density Functional Theory calculations were carried out for further validation with sodium layered oxide cathodes. Two machine learning interatomic potential models, namely, universal model for atoms (UMA) and message passing atomic cluster expansion (MACE) were also compared along with the density functional theory results and the graph transformer. The evaluations revealed that the UMA model outperformed the graph transformer and MACE, reaching a mean absolute error of 0.11 V. Motivated by this performance advantage, we evaluated the average voltage and corresponding energy density of 35460 sodium layered oxides using the UMA model.