Machine-learning-assisted prediction of highly active and stable PrBaCo2O5+δ-based oxygen electrodes for solid oxide cells

Abstract

Highly active and stable oxygen electrodes are crucial for enhancing the power density and durability of solid oxide cells (SOCs). The recently developed machine learning (ML) method for material exploration shows tremendous advantages over the traditional time-consuming and costly trial-and-error process; however, the study for SOC electrode materials is limited by a lack of sufficient and reliable datasets. Herein, 28 sets of in-lab data were used as high-quality datasets for ML training to screen PrBaCo₂O_5+δ-based oxygen electrodes with high activity and strong resistance to CO₂ poisoning for SOCs. The result shows that the random forest model outperforms other models in predicting both catalytic activity and stability. Specifically, the descriptors “radius difference between A′ and A site ions” and “ionic absolute hardness” are identified as significant factors for predicting activity and stability, respectively. Based on the well-trained model, three oxides, screened from 48 distinct PBC-based oxides, are successfully synthesized and confirmed to have excellent catalytic activity and robust resistance to CO₂ poisoning, demonstrating the high reliability of the ML model. This work establishes a new avenue for machine learning-based development of highly active and stable oxygen electrodes for SOCs.