Predicting CO2 adsorption in metal-organic frameworks: Integrating machine learning with virtual sample generation

This study introduces a machine learning (ML) strategy that uses kernel density estimation (KDE) to generate virtual samples, significantly enhancing the prediction of CO 2 adsorption capacities in metal-organic frameworks (MOFs). Utilizing a dataset of 339 entries with features such as surface areas (S BET and S L ), pore volume (V T ), pressure (P), and temperature (T), we compared four ML models: Random Forest (RF), Gradient Boosting Decision Trees (GBDTs), Light Gradient Boosting Machines (LGBMs), and eXtreme Gradient Boost (XGBOOST). The incorporation of virtual samples led to notable improvements in model precision. For instance, correlation coefficients (R 2 ) improved, with RF increasing from 0.988 to 0.997, and root mean square error (RMSE) values decreased from 0.642 to 0.360. To demonstrate the model's generalizability, we applied it to the MOF-74 series, a new set of MOF materials not included in the training dataset. The predictions closely matched experimental and simulation results, validating our methodology as a reliable and efficient tool for evaluating and developing new MOF materials. This approach effectively bridges the gap between theoretical and experimental research in predicting CO 2 uptake.