Engineering Machine\nLearning Features to Predict Adsorption\nof Carbon Dioxide and Nitrogen in Metal–Organic Frameworks

Abstract

In this article,\nwe propose simple, interpretable machine learning\n(ML) features aimed at improving the accuracy of ML models in predicting\nadsorption of gases, such as methane, ethane, propane, krypton, and\nxenon, in metal–organic frameworks (MOFs). In particular, we\nintroduce energy-based features that incorporate surface energy histograms\nand radial distribution functions. These features effectively capture\nspatial variations in interaction energy, thereby improving the predictive\ncapabilities of the ML models. We further extend this approach to\npredict the adsorption of carbon dioxide and nitrogen in a diverse\nset of MOFs under ambient conditions by including the Coulombic energy\ncomponent in the proposed energy-based features. The final ML model\ncombines the energy-based features with geometric characteristics\nof MOFs and Henry’s law constants. This integration results\nin significantly increased predictive accuracy of the models, especially\nwithin the moderate pressure regime, when compared to that of previous\nstudies. Using the CoRE MOF database for training and testing, our\nML model achieves a prediction accuracy of <i>R</i>² > 0.96 for methane, ethane, krypton, and xenon at various\nconditions, <i>R</i>² > 0.96 for propane, <i>R</i>² > 0.80 for carbon dioxide, and <i>R</i>² > 0.97 for nitrogen. Using the CRAFTED\ndatabase of simulated\nadsorption isotherms for carbon dioxide and nitrogen, we demonstrate\nthe robustness of the ML model in predicting single-component isotherms\nat a specific temperature. In this test, the model achieves <i>R</i>² > 0.87 for carbon dioxide and <i>R</i>² > 0.90 for nitrogen. The physics-based nature of\nthe\nproposed features allows us to provide some interpretation of why\nthe predictions of the ML model are better for some materials than\nfor others. According to this interpretation, an accurate representation\nof the shape of the potential energy surface plays an important role\nin capturing the process of adsorption. Finally, we acknowledge certain\nlimitations of the current model such as the oversimplification in\nrepresenting the geometry and interactions of molecules with polar\ninteractions.