Abstract
Graph neural networks for crystal property prediction typically require precise atomic positions and types, limiting their applicability for novel materials with unknown structures. To address this limitation, we introduce BatteryFormer, a versatile machine learning model that employs average interatomic radius distance instead of precise bond lengths as edge embedding, enabling rapid, high-throughput material screening based solely on composition and structural prototypes. BatteryFormer demonstrates robust predictive performance across a wide range of intervals. It accurately predicted high redox potentials for four distinct cathode materials: layered oxides, fluorophosphate salts, vanadium fluorophosphate salts, and ferric pyrophosphate salts. Notably, it also correctly predicted the low redox potential (1.56 V) of the recently reported cathode material Na(6)CoS(4), highlighting its reliability in diverse chemical spaces. Beyond numerical accuracy, BatteryFormer captures crucial local structural features, such as the linear Na-O-Li configuration in layered transition metal oxide cathodes, essential for enhancing redox potentials. The model also maintains high predictive accuracy for a variety of lithium-ion battery cathode materials, further validating its strong generalization capability. Integrating knowledge graphs and knowledge inference, this work provides a visual mapping of relationships among material types, doping element combinations, doping ratios, redox potentials, capacities, and energy densities. This integration offers practical guidance for synthesizing high-entropy sodium-ion battery cathodes with enhanced cycling stability and energy density. The proposed data-driven approach provides a robust framework for accelerating materials discovery and transitioning from empirical materials design strategies.