Interpretable machine learning for accelerated discovery of air electrodes in proton-conducting solid oxide cells

The global energy transition faces challenges due to the intermittency of renewable sources. Hydrogen has emerged as a promising solution, with proton-conducting solid oxide cells (P-SOCs) offering reversible advantages in hydrogen production and power generation. To accelerate the development of air electrodes for P-SOCs, we propose an interpretable machine learning (ML) strategy for designing and screening perovskite oxides. We build ML models to predict polarization resistance and peak power density. After evaluating ten algorithms using 10-fold cross-validation, the XGBoost model demonstrates superior predictive accuracy for both targets. Feature importance and SHAP analyses reveal that A-site ionic electronegativity, ionization energy, and temperature are dominant descriptors for R _p, while temperature and Lewis acid strength are crucial for power density. Using this validated model, we screen 4738 candidate compositions and identify two promising air electrode materials, Sr_0.6Pr_0.3Cs_0.1Co_0.6Fe_0.3Mo_0.1O_3-δ (SPCCFM) and Ba_0.8Pr_0.8Cs_0.4Co_1.6Ni_0.4O_5+δ (BPCCN). In fuel cell mode, SPCCFM-based and BPCCN-based single cells achieve power densities of 0.93 and 0.85 W cm⁻² at 650 °C, respectively. In electrolysis cell mode, their current densities reach −1.95 and −2.10 A cm⁻². Experimental validation further confirms their high electrochemical performance and stability, demonstrating the effectiveness of our ML-driven approach to discovering advanced air electrodes for P-SOCs.