Optimization of SOFC Electrode Microstructures with Long Term Performance Modeling and Machine Learning

Abstract

Performance degradation over time is a major barrier to the commercialization of solid oxide fuel cells (SOFCs). A major driver of this degradation is grain coarsening in the 3D microstructure of the porous, composite electrodes. The rate and extent of this microstructural degradation depends on the initial microstructure itself, raising the question of whether an optimal electrode microstructure can be determined for long-term performance. The high dimensionality of parameters that determine the initial microstructure makes parametric optimization challenging. In this work, we present a fully resolved 4D model for simulating microstructural and performance changes over 1,000 hours of cell operation by combining a phase-field coarsening model with spatially resolved microstructural analysis and continuum-level multiphysics performance modeling. The integrated model framework was run in a highly parallelized fashion to simulate the long-term performance of hundreds of LSM-YSZ cathodes and Ni-YSZ anodes. These electrode microstructures were synthetically generated to intentionally span 11 independent initial microstructural parameters. The results of the long-term performance model included electrochemical performance every 100 hours throughout the entire 1,000 hours of operation. Together with a novel figure of merit that accounts for both the initial and long-term performance, this high-dimensional bank of results was used to train a machine learning regression model. The machine learning model was trained to predict the long term performance based on the 11 independent initial microstructural parameters, allowing interpolation to values not included in the fully resolved models, as well as a more robust analysis of each parameter’s influence in this highly dimensional parameter space. A SHAP analysis [1] was performed on the trained machine learning model to determine the relative impact of each of the 11 independent initial microstructural parameters. This demonstrated that, for example, the LSM/YSZ ratio was the most impactful parameter in the cathode microstructure, with lower LSM/YSZ ratios generally leading to more energy produced over the cell’s lifetime. The fully trained model and SHAP analysis are also able to make specific recommendations for changes that would improve existing electrodes, providing a valuable tool for SOFC electrode developers and manufacturers.[1] S. M. Lundberg, S-I. Lee. A Unified Approach to Interpreting Model Predictions. NIPS 2017. Figure 1