Machine learning-based prediction and optimization of solid oxide electrolysis cells for green hydrogen production using RF, FFNN, and PSO approaches

This study presents an innovative approach to optimizing hydrogen production in solid oxide electrolyzer cells (SOECs) by integrating advanced machine learning (ML) techniques with particle swarm optimization (PSO). A curated dataset of 257 data points, representing key SOEC operational parameters, was utilized to develop and optimize predictive models. Two ML algorithms, random forest (RF) and feed-forward neural network (FFNN), were applied to predict hydrogen production rates. Recursive feature elimination with cross-validation (RFECV) confirmed the relevance of all 15 input variables, including numerical features such as current density, voltage, temperature, pressure, surface area, humidity, volumetric flowrate, electrolyte thickness, volume fractions of (H2, CO2, and H2O), and Ohmic resistance, as well as categorical variables such as cathode electrode type, electrolyte type, and inlet gas composition. The RF model outperformed the FFNN, achieving a coefficient of determination (R2) of 0.9561 compared to 0.9081 for the FFNN model. While the RF model's R2 was slightly below the best-reported value in the literature for XGBoost (R2 = 0.9677), it significantly outperformed support vector regression (R2 = 0.8062). To further maximize hydrogen production, PSO was integrated with the RF model to optimize the key input variables. This approach achieved a maximum predicted hydrogen production rate of 1.287 L/(h·cm2), representing a notable improvements over previous benchmarks in the literature under optimal conditions: a temperature of 739.65 °C, current density of 2.70 A/cm2, surface area of 6.40 cm2, an operating cell voltage of 1.30 V, electrolyte thickness of 42.62 μm, humidity of 25.39 %, and volumetric flowrate of 104.13 mL/min. SHapley additive exPlanations (SHAP) provided interpretability by identifying the magnitude and direction of impact for key variables such as current density, flow rate, surface area, and voltage on hydrogen production. The integration of SHAP analysis not only confirmed the importance of features identified by RFECV but also offered deeper insights into the models' decision-making processes. Furthermore, this study underscores the novelty of adopting PSO over traditional genetic algorithms (GA), achieving improved hydrogen production under feasible and realistic operating conditions. These results show that the proposed method can be useful for scaling up SOEC systems and helping to make better decisions in operating green hydrogen technologies. Overall, this approach supports the global move toward clean energy by providing a data-driven pathway to improve hydrogen production in high-temperature electrolysis systems.