Prediction of Adsorptive Activities of MOFs for Pollutants in Aqueous Phase Based on Machine Learning

Metal–organic frameworks (MOFs) have gained significant attention in the field of pollutant removal due to their rich pore structures and large specific surface areas. As the number of MOF structures continues to increase, machine learning methods have become a powerful tool for prediction of adsorptive activities of MOFs for pollutants. In this study, 16 models were constructed using published adsorption data, which included 28 MOFs and 30 pollutants, resulting in a dataset of 836 data points. The XGBoost model was determined to be the most effective model, achieving an average R2 of 0.953 during the 5-fold cross-validation. The model's performance was influenced by a combination of MOF features, pollutant features, and adsorption conditions. Key parameters for the XGBoost model's performance included the pollutant concentration, pH, solid–liquid ratio, and temperature. Different types of MOFs, including Zr-MOFs, Cr-MOFs, Al-MOFs, and Fe-MOFs, were observed to display distinct adsorption mechanisms through the machine learning model. These mechanisms included electrostatic interactions, π–π interactions, hydrogen bonding, and van der Waals force. The model's predictions regarding the optimal MOFs and adsorption conditions for the 30 pollutants were partially validated through experimental data, demonstrating the feasibility of the model's predictions. This study provides technical and theoretical support for the prediction and selection of optimal MOFs for pollutant removal in the aqueous phase.