Physics-Informed Surrogate Modeling and System-Level Validation of a Hydrogen Ejector for Proton Exchange Membrane Fuel Cell System

Hydrogen ejectors are widely used for passive anode recirculation in proton exchange membrane fuel cell (PEMFC) systems due to their simplicity and low parasitic power consumption. However, accurate prediction of ejector performance under humid, multispecies, and wide operating conditions remains challenging using conventional geometry-based or steady entrainment models. This study develops a physics-informed surrogate model for efficient prediction of hydrogen ejector mass flow behavior in PEMFC applications. A multispecies CFD model is first established to generate flow-field data under realistic humid hydrogen operating conditions and is complemented by experimental measurements from a 130 kW PEMFC test platform. To improve model robustness and interpretability, 17 candidate features are constructed from raw operating variables using dimensionless analysis and gas-dynamic principles. A correlation-based feature selection strategy is then applied to reduce redundancy, resulting in an 11-variable physically meaningful feature set that preserves key pressure-driven, compressibility, and transient characteristics. Four machine learning algorithms, including random forest, gradient boosting, support vector regression, and feedforward neural networks, are systematically evaluated for predicting primary and secondary mass flow rates. Among them, the feedforward neural network achieves the best performance, with R² values of 0.999 for primary flow and 0.913 for secondary flow prediction. Finally, the trained surrogate model is embedded into a dynamic PEMFC system model for system-level validation. The integrated model reproduces transient system behavior accurately under varying load conditions, achieving a mean absolute percentage error of 1.9% in net power prediction compared with experimental data. The results demonstrate that physics-informed feature reduction combined with surrogate modeling provides an efficient and reliable pathway for integrating hydrogen ejector dynamics into PEMFC system-level simulations.