PREDICTION OF HYDROGEN YIELD AND METHANE CONVERSION IN METHANE PYROLYSIS USING A CONVOLUTIONAL NEURAL NETWORK

Abstract

Methane pyrolysis has emerged as a promising low-carbon route for hydrogen production because it generates hydrogen without direct carbon dioxide emissions in the main reaction pathway while producing solid carbon as a by-product. However, the prediction of hydrogen yield and methane conversion remains challenging due to the complex nonlinear interactions among operating conditions and catalyst composition variables. In this study, a convolutional neural network (CNN) framework is proposed to simultaneously predict hydrogen yield and methane conversion in methane pyrolysis using structured process and catalyst data. The developed model was trained on methane pyrolysis input variables, including temperature, gas hourly space velocity, CH₄​ concentration, reaction time, calcined temperature, and catalyst composition descriptors. The CNN was designed as a multi-output regression model to capture hidden nonlinear relationships between the input variables and the two target responses. Experimental evaluation demonstrated stable convergence during training, with training loss decreasing from nearly 0.18 to 0.009 and validation loss dropping from about 0.20 to 0.016 over 60 epochs, indicating effective learning and limited overfitting. Comparative analysis showed that the proposed CNN outperformed benchmark models, including Linear Regression, Support Vector Regression, Random Forest, multilayer perceptron, and CatBoost, achieving the highest predictive accuracy with an R² of approximately 0.95 and the lowest RMSE of nearly 4.4. Feature-importance analysis further revealed that temperature, gas hourly space velocity, reaction time, and methane concentration were the dominant predictors, together accounting for nearly 60% of the overall importance. These findings confirm that CNN-based regression provides a robust and accurate framework for methane pyrolysis modeling and offers strong potential for predictive analysis, process screening, and future optimization of hydrogen production systems.