Microstructure-informed performance boost in solid oxide fuel cells through multiphysical modeling and machine learning

The optimization of the macro- and microstructures of traditional solid oxide fuel cells (SOFCs) faces the dual challenges of time-consuming experimental iterations and insufficient exploration of parameter space. This study proposes an anode-supported SOFC optimization approach based on multiphysical modeling and machine learning, aiming to achieve the coordinated optimization design of its macro- and microstructures, thereby ensuring the improvement of power density and the reduction of failure probability. The study first constructed a database of maximum power density and failure probability based on multiphysical modeling, and then screened out 10 key features that affect the above two target parameters through feature engineering. On this basis, 15 machine learning predictive models were constructed, among which the random forest (RF) regression model showed excellent prediction performance, and the determination coefficients (R²) of the maximum power density and failure probability predictive models reached 0.99 and 0.95 respectively. The cooperation of the genetic algorithm and RF obtained the optimal combination of key parameters, ensuring that the cell achieved the highest power output within the failure probability range of 0.632. The SOFC button cell prepared based on the cathode optimization results was experimentally verified, and its maximum power density reached 1.43 W cm⁻², which was 29% higher than the initial sample, verifying the effectiveness of the proposed optimization approach. In addition, Shapley additive explanations (SHAP) were introduced to improve the interpretability of the model. The results show that most key features have opposite effects on the two target quantities, demonstrating the necessity of considering the failure probability.