Machine learning models for prediction of CO2 adsorption capacity of metal-organic frameworks

Abstract

The current study utilizes advanced machine learning (ML) techniques to predict the non-linear adsorption of CO₂ by metal-organic frameworks (MOFs). To model the non-linear nature of the CO₂ adsorption capacity of the treated and no-addition MOFs, a set of 18 different machine learning and deep learning models were used to see which best fit the dataset. Gathering 1154 data points from the literature on MOF adsorption capacities at various temperatures and pressures, the study focused on developing ML models, including RF, support vector machine (SVM), logistic regression (LR), and MLP. Among these models, Polynomial model exhibited the highest coefficient of determination (R², 99.9 %) and the lowest root mean squared error (0.04) for test datasets at 298 K. To enhance the interpretability of ML outcomes, the feature analysis by gradient boosting regression and permutation importance plot was used to quantitatively elucidate the influence of intrinsic properties of MOFs on CO₂ adsorption capacity, with particular attention to their interdependence. Finally, the study analyzed the synergistic effects of chemical compositions on CO₂ adsorption capacity. This investigation aims to inform the design of next-generation high-performance MOFs for carbon capture by extracting new insights from extensive historical data using explainable ML techniques.