Statistical analysis and Partial Least Square regression as new tools for modelling and understanding the adsorption properties of zeolites

Statistical data analysis was applied to examine previously published adsorption enthalpies of linear and branched alkanes on zeolites Na-Y, Na-USY, Mordenite, Beta, ZSM-5 and ZSM-22. The objective was to establish predictive models that correlate the adsorption enthalpies for the six zeolites to simple structural properties or descriptors of the adsorbates. Two sets of descriptors were tested; a first set was chosen from a library of topological descriptors (the Wiener index, the Kier shape indices and the Kier + Hall connectivity indices), the other set was custom-made, by directly parametrizing the number of carbon atoms as well as position and degree of branching. From the set of 17 available adsorbates, training sets of 10 molecules were selected by a clustering analysis and this training set was subjected to a Partial Least Squares (PLS) regression in order to establish a linear relationship between the adsorption enthalpy and the set of descriptors. The model was then applied to predict the adsorption enthalpies of the other molecules. In all cases, predicted and experimental values were in good agreement. The model based on the custom-made descriptors that code the branching of the molecules was less precise, but allowed a more straightforward interpretation of the coefficients of the model, which provides very interesting insights into the adsorption behaviour of the linear and branched alkanes on large and medium pore zeolites.