A physics-informed neural network (PINN) predicting the performance of air electrodes for solid oxide cells: a pilot-scale demonstration on four microstructures

Solid oxide cells are a promising technology for clean energy conversion, with performance strongly impacted by their electrode microstructure. The accurate prediction of the cell performance based on the electrode properties is crucial for optimizing design and, consequently, operation. This study employs Physics-Informed Neural Networks (PINNs) to predict the performance of SmBa0.8Ca0.2Co2O5+δ-based air electrodes featuring different microstructural properties, which impact on the effective transport properties and reaction rate. The proposed phenomenological equation of the kinetic rate for the oxygen redox reactions, based on the Butler-Volmer formalism, is included into the model loss function to enhance its physical accuracy. The developed model is tested on four electrode microstructures, and their electrochemical properties were simulated by combining the PINN framework and a physics-based continuum model, validated for SmBa0.8Ca0.2Co2O5+δ-based air electrodes and used to collect training and test datasets. The PINN framework was trained with a very limited amount of data (<40) and returned mean squared error (MAE) values of ≈10−5. The obtained results offered a pilot-scale demonstration of the PINNs efficacy in prediction accuracy with minimal data and capturing the complex interplay between diffusion coefficients and fuel cell performance, thus standing as potential robust tools to engineer the cells design.