Machine Learning-Driven Insights and Predictions for CO₂ Adsorption in Metal-Organic Frameworks

It is becoming urgent to design MOFs with a high capacity of CO₂ adsorption due to climate change. Machine learning opens up new possibilities for building MOFs with high CO₂ adsorption capacity and understanding the CO₂ adsorption mechanism. In this study, we collected 608 data points from several reports, characterized with several key variables such as BET surface area (SBET), total pore volume (VT), pressure (P), temperature (T), Langmuir surface area (SL). Eight engineered features were also used: electronegativity (E), atomic mass (AM), density (D), first ionization (FI), specific heat (SH), atomic radius (AR) and melting point (MP) to predict CO₂ uptake. A panoply of machine learning models was built to predict the CO₂ adsorption capacity of MOFs. The results indicated that Gradient Boosting Machine (GBM) had the best prediction performance (R2 training = 0.983, R2 testing = 0.945, RMSE training = 1.017, RMSE testing = 1.646). Feature importance study demonstrated that beside pressure and temperature, CO₂ adsorption parameters (75%), textures (20%), and MOF electronegativity or some chemical features (5%) are all important in the CO₂ adsorption process and that higher value of those parameters tend to negatively affect the adsorption intake.