Applied Machine Learning for Prediction of Energy-Efficient CO₂ Desorption on Solid Acid Catalysts

The development of solid acid catalysts (SACs) for energy-efficient CO₂ desorption and amine regeneration is critical to carbon capture commercialization. To avoid the time-consuming and ineffective screening process, a predictive model correlating the physicochemical properties of SACs with catalytic performance is desired, but it remains a challenging task. Herein, four machine learning (ML) algorithms were integrated with virtual data augmentation (VDA) methods to develop the predictive model of catalytic performance of SACs based on 13 features associated with catalyst properties and reaction conditions. The results showed that VDA methods could generally improve the predictive accuracy of ML models, with the XGBoost models achieving the best predictive performances. Permutation importance and SHAP analysis revealed the features' impact on the catalytic performance of SACs from complementary perspectives. Based on insights gained from ML models, response surface methodology was implemented to delineate potential catalyst optimization pathways, with symbolic regression enabling the formulation of predictive equations. Both the equations and the ML models were subsequently integrated into graphical user interface (GUI) software to develop a user-friendly tool for rapidly predicting and screening high-performance SACs. This study establishes an integrated VDA-interpretable ML framework for rational SACs design in energy-efficient CO₂ desorption.