The electrochemical conversion of CO2 into value-added chemicals using solid oxide electrolysis cells (SOECs) offers a promising pathway for renewable energy storage and carbon-neutral fuel production, yet its advancement is limited by the lack of predictive design principles for high-performance cathodes. Conventional trial-and-error materials discovery is time-consuming and inefficient. Here, we combine density functional theory calculations and experimental observations to suggest the oxygen 2p-band center (εp) as an activity-relevant descriptor for perovskite cathodes and integrate this insight with interpretable machine learning to establish composition–electronic structure–performance relationships. Regression models trained on 230 perovskites reveal electronegativity mismatch and configurational entropy as primary factors modulating εp. Guided by these physically interpretable descriptors, interpretable machine learning is employed to explore a large chemical space of 9703 A-site high-entropy perovskites and identify candidates with optimized εp positions. As a representative composition, the La1/6Sr1/6Li1/6Ba1/6Ca1/6Sm1/6FeO3−δ cathode delivers a current density of 2.55 A cm−2 at 800 °C and 1.5 V together with stable operation exceeding 1000 h. These findings provide a descriptor-informed strategy for advancing scalable CO2 electrolysis technologies in sustainable energy systems.