Machine learning prediction on adsorption capacities of steam methane reforming off-gas in silica gels

Abstract

Pressure swing adsorption (PSA) is widely used for hydrogen purification from steam methane reforming (SMR) off-gases, but its design and optimization require extensive equilibrium adsorption data. These data are typically obtained from static experiments, which are often costly, time-consuming, and inefficient. This study presents a machine learning-based approach to predict the adsorption capacities of two types of silica gels (SG1 and SG2) for key gas components (CO₂, CH₄, CO, and H₂) in the SMR process, aiming to significantly reduce experimental costs and enhance data acquisition efficiency. Five widely used machine learning models were investigated, including decision tree (DT), random forest (RF), extreme gradient boosting (XGBoost), support vector machine (SVM), and deep neural network (DNN). To improve model performance, hyperparameters were optimized using the Optuna framework, combined with five-fold cross-validation. All five models demonstrated excellent predictive accuracy, with coefficients of determination (R²) exceeding 0.99. Among them, the DNN model outperformed the others, achieving an R² value of 0.999. To validate the model predictions, three temperature-dependent adsorption isotherm models (single-site Langmuir, single-site Langmuir–Freundlich and dual-site Langmuir) were employed to fit the experimental data. The dual-site Langmuir model provided the best fit for CO₂ and CH₄, while the single-site Langmuir–Freundlich model was most suitable for CO and H₂. The adsorption capacities predicted by the DNN model showed strong agreement with those from the optimal isotherm models for all four gases. Furthermore, the DNN model was used to predict CO₂ adsorption capacities under extrapolated temperature and pressure conditions. The DNN predictions closely matched those from the dual-site Langmuir model and were consistent with experimental measurements. These results confirm that the DNN approach can effectively replace conventional static experiments for accurately and efficiently generating equilibrium adsorption data.