Elucidating Thermodynamically Driven Structure-Property Relations for Zeolite Adsorption Using Neural Networks.

Understanding the origin and effect of the confinement of molecules and transition states within the micropores of a zeolite can enable targeted design of such materials for catalysis, gas storage, and membrane-based separations. Linear correlations of the thermodynamic parameters of molecular adsorption in zeolites have been proposed; however, their generalizability across diverse molecular classes and zeolite structures has not been established. Here, using molecular simulations of >3500 combinations of adsorbates and zeolites, we show that linear trends hold in many cases; however, they collapse for highly confined systems. Further, there are no simple predictors of the slope of the linear correlations, thereby indicating that there are no universal linear models relating molecule and zeolite pore structures with adsorption properties. We show that nonlinear models, in particular bootstrapped neural networks, that only use geometric and physical descriptors of the adsorbate and zeolite as features can predict the entropy of adsorption, isosteric heat, and Henry's constant (log( )) to within 4.71 [J/mol/K], 3.14 [kJ/mol], 1.15 [mg/(g-cat·atm)], respectively. A SHAP analysis that deconvolutes the effect of correlated features to compute their independent additive contributions showed that framework, rather than adsorbate features, was significantly more important for predicting the entropy of adsorption but equivalently important for predicting the Henry's constant. The largest pore diameter along a free sphere path was identified as the most critical framework feature, while the van der Waals volume (which captures the trend in electronegativity) was the most important adsorbate feature toward predicting the entropy of adsorption.