Machine learning–accelerated discovery of novel high‑entropy spinel oxide cathodes for solid oxide fuel cells

Solid oxide fuel cells (SOFCs) offer high efficiency and fuel flexibility for clean energy, but high operating temperatures accelerate cathode degradation and hinder commercialization. Conventional perovskite cathodes face significant challenges including thermal expansion mismatch, element segregation, and poisoning. High-entropy spinel oxides offer an alternative with enhanced structural stability and electrocatalytic activity, yet efficiently identifying viable compositions within the vast design space remains challenging. Here, we develop a machine learning (ML) model to predict single-phase formation in high-entropy spinel oxides using experimental data from solid-state reactions. Ten features—comprising elemental properties, precursor characteristics, and calcination parameters—are encoded, and ADASYN oversampling addressed class imbalance. Three models (XGBoost, Random Forest, Naïve Bayes) are trained and evaluated via accuracy, precision, recall, F1, and ROC–AUC metrics. XGBoost achieved the highest test accuracy (88 %) and AUC (0.974). SHAP analysis identified precursor volume deviation, electronegativity deviation, and density deviation as key predictors, revealing that uniform precursor properties favor single‑phase formation. Experimental synthesis of five ML‑predicted compositions confirms two single‑phase high‑entropy spinel oxides (Mn0.6Fe0.6Co0.6Cu0.6Mg0.6O4 and Mn0.2Co0.2Ni0.2Cu0.2Zn0.2Fe2O4), which exhibit excellent chemical compatibility with yttria-stabilized zirconia (YSZ) and gadolinium-doped ceria (GDC) electrolytes. Among them, Mn0.6Fe0.6Co0.6Cu0.6Mg0.6O4 exhibits superior electrochemical performance, demonstrating that it would be a promising potential cathode material for SOFCs. This work demonstrates a complete ML‑to‑lab pipeline for high‑entropy spinel cathode discovery, significantly accelerating cathode material development for SOFCs and providing interpretable design rules for future compositional screening.