Machine-Learning-Based Screening of Perovskite Cathodes for Low-Temperature Solid Oxide Fuel Cell Operation

The discovery of cathode materials that simultaneously exhibit high oxygen-reduction activity, robust stability, and low cost is pivotal to moving solid oxide fuel cells (SOFCs) from the laboratory into commercial deployment. To address this challenge, we compile the largest perovskite dataset to date parameterized by the oxygen tracer surface exchange coefficient (k*). Using only readily obtainable elemental and structural descriptors, we develop machine-learning models that surpass existing approaches in both accuracy and computational efficiency. Specifically, by integrating Mahalanobis-distance-based applicability-domain analysis with random forest-enhanced property descriptors and support vector regression, we high-throughput-screen 1.3 million ABO3 compositions and curate a candidate list that balances thermodynamic stability, cost, and oxygen-reduction activity. Beyond prediction accuracy, SHAP interpretation reveals strong physical correlations between the enhanced descriptors and k*, highlighting the coefficient of thermal expansion, O p-band center, and A-site ionic radius as the dominant factors governing oxygen exchange kinetics. Finally, we identify 209 promising perovskite cathodes predicted to outperform LSC in the low-temperature regime, offering promising directions for experimental realization of practical low-temperature SOFCs.