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

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Commercializing carbon capture requires energy-efficient solid acid catalysts (SACs) for CO2 desorption and amine regeneration.
• Current screening methods for SACs are time-consuming and inefficient, necessitating predictive models.
• Correlating catalyst properties with performance is crucial but challenging.

Purpose of the Study:

• To develop a predictive model for the catalytic performance of SACs.
• To integrate machine learning (ML) with virtual data augmentation (VDA) for enhanced catalyst screening.
• To identify key features influencing SAC performance and optimize catalyst design.

Main Methods:

• Four ML algorithms combined with VDA methods were used to predict SAC catalytic performance.
• The models utilized 13 features related to catalyst properties and reaction conditions.
• Permutation importance, SHAP analysis, response surface methodology, and symbolic regression were employed for analysis and optimization.

Main Results:

• VDA methods generally improved the predictive accuracy of ML models.
• XGBoost models demonstrated the best predictive performance.
• Feature analysis identified key drivers of SAC catalytic performance.

Conclusions:

• An integrated VDA-interpretable ML framework was established for rational SAC design.
• The developed models and equations were integrated into GUI software for user-friendly prediction and screening.
• This approach facilitates the development of high-performance SACs for energy-efficient CO2 desorption.