Predicting water saturation and oxygen transport resistance in proton exchange membrane fuel cell by artificial intelligence

In the simulation of proton exchange membrane fuel cells, a “3D + 1D” model combines a detailed 3D domain with a simplified 1D domain. The “3D + 1D“ method accurately simulates flooding by considering the resistance to oxygen transfer and water saturation in porous media adjacent to the flow field. This study proposes an efficient model compatible with ”3D + 1D.“ The model should be able to swiftly and accurately predict the resistance and water saturation under various flow fields, GDL porosities, and operating conditions. To this end, a multiphase 3D computational fluid dynamics model generates a data set. The data set comprises anticipated variations. A comparative investigation of six regression methods has been presented. The methods are based on Response Surface Methodology (RSM), Artificial Neural Networks (ANN), and four ensemble machine learning methods, named Random Forest (RF), XGBoost, Lightgbm, and CatBoost. Genetic algorithm optimization tunes hyperparameters. It has been found that every model has proven to be accurate at predicting resistance. With an R2 of 0.9899, CatBoost has the best performance, while RSM presents the simplest regression employed to formulate the resistance. In the context of predicting water saturation, ensemble machine-learning models have demonstrated better performance than RSM and ANN models. CatBoost performs the best with an R2 of 0.983 in predicting water saturation. It is recommended that CatBoost and the “3D + 1D” method be integrated to achieve an accurate prediction of flooding.