Predictive Modeling of Photocatalytic Hydrogen Production: Integrating Experimental Insights with Machine Learning on Fe/g-C3N4 Catalysts

Abstract

Hydrogen emerges as a promising alternative to fossil fuels with its pollutant-free emissions, high energy density, versatility, and efficiency in generating power. In this study, photocatalytic hydrogen production from using 1000 ppm of model solution prepared with sucrose was investigated in the presence of Fe/g-C₃N₄ photocatalysts over Box–Behnken experimental design developed using the Minitab statistical software. The amount of hydrogen produced was optimized at different pH environments (3, 5, and 7) for 2 h reaction time with different amounts of metal loaded (10, 20, and 30 wt %), Fe/g-C₃N₄ (0.1, 0.2, and 0.3 g/L), and oxidant (H₂O₂; 0, 10, and 20 mM) concentrations. SEM, BET, XRD, FTIR, and PL analyses were employed for the characterization of synthesized photocatalysts. According to the response optimization, using Fe/g-C₃N₄, the optimal conditions for hydrogen production were found as 0.3 g/L catalyst loading, 18.8 mM H₂O₂, and 26.6% Fe loading by mass when the pH was 3 for the reaction medium. Furthermore, machine learning algorithms were employed to predict hydrogen evolution based on experimental parameters. Notably, ensemble models such as Voting Regressor combining the Bagging Regressor, Random Forest Regressor, LGBM Regressor, Extra Trees Regressor, XGB Regressor, and Gradient Boosting Regressor achieved superior performance with a mean squared error of 0.0068 and R-squared (R²) of 0.9895. This integrated approach demonstrates the efficacy of machine learning in optimizing photocatalytic hydrogen generation processes.