Enhanced hydrogen production from methane pyrolysis using CatBoost with reptile search algorithm optimization

The growing demand for clean energy solutions and the rapid depletion of fossil fuel resources have made it essential to find carbon-free energy options. One promising approach is the production of hydrogen via methane pyrolysis. In this study, the CatBoost machine learning (ML) algorithm was employed to predict hydrogen (H₂) yield and methane (CH₄) conversion during the pyrolysis process. To further enhance predictive performance, the reptile search algorithm (RSA) was integrated to optimize the hyperparameters of the CatBoost model, resulting in a hybrid Cat-RSA model. Model evaluation revealed that the RSA optimization significantly improved the prediction accuracy of both H₂ yield and CH₄ conversion. The Cat-RSA model achieved impressive performance with coefficient of determination ( R² ) values of 0.9422 for H₂ yield and 0.9721 for CH₄ conversion. Furthermore, the model’s precision is underscored by low root mean squared error (RMSE) values of 5.464 for H₂ and 3.864 for CH₄, as well as mean absolute relative error (MARE) values of 0.2682 for H₂ and 0.2137 for CH₄. These results are comparable to those reported in previous studies, such as SVM-ABC (R² = 0.9464) and BR-LMMLP (R² = 0.9530) for H₂ yield, and are slightly lower than XGBoost (R² = 0.9996). Additionally, SHapley Additive exPlanations (SHAP) analysis identified key features, including gas hourly space velocity (GHSV), time, copper wt.% in the catalyst, and temperature, as the significant contributors to the model’s predictions. Furthermore, the Cat-RSA model predicted a maximum hydrogen yield of 91.05 % under the following optimal input conditions: Temperature = 714.9 °C, GHSV = 7111.9 mL/h·g, CH₄ concentration = 18.4 %, time = 14.1 min, and calcined temperature = 476.9 °C. The optimal catalyst composition featured high proportions of Ni (55.25 wt.%) and Co (30.00 wt.%). This predicted hydrogen yield is consistent with experimentally reported values in the range of approximately 77 % to 91.3 %, reinforcing the model’s predictive validity and practical relevance for methane pyrolysis systems. Overall, this work enhances our understanding of how hydrogen is produced from methane pyrolysis and demonstrates the potential of innovative optimization algorithms in improving ML applications within energy systems.