(283d) Inferring Structure-Entropy Correlations of Zeolite-Adsorbate Interactions Using Monte-Carlo Simulations & Machine Learning | AIChE

(283d) Inferring Structure-Entropy Correlations of Zeolite-Adsorbate Interactions Using Monte-Carlo Simulations & Machine Learning


Rzepa, C. - Presenter, Lehigh University
Rangarajan, S., Lehigh University - Dept of Chem & Biomolecular
Zeolites are porous crystalline alumino-silicates with distinct pore topologies; and much work has been devoted to optimizing their use in catalytic applications. Advancements in ab-initio simulations have made possible to predict the catalytic properties of a zeolite framework on a particular reaction through quantification of the energy landscape, and subsequently the development of microkinetic models. Of the contributions to the free energy, entropic losses within zeolite catalysis have shown to be crucial toward rate inhibition and therefore product selectivity; and have traditionally been related with the degree of confinement associated with adsorption. Nevertheless, the adsorption entropy remains difficult to quantify, or expensive to calculate, but has been shown to follow straightforward correlations with considerable predictive capabilities.

Campbell and Sellers compiled a collection of experimental adsorbate entropies on two-dimensional catalytic surfaces.[1] The key finding of their work was that the ratio of the adsorbed-phase entropy to the gas-phase entropy was approximately two-thirds across all molecules. They proposed a simple explanation, suggesting that the adsorption of a molecule from an unhindered gas phase onto a two-dimensional surface would eliminate a dimension of translational freedom, i.e. the adsorbate behaves as a 2D gas. Dauenhauer and Abdelrhaman[2] expanded this idea to three-dimensional frameworks by compiling experimentally determined adsorption entropies for alkanes adsorbed in aluminosilicate zeolites. They showed that the entropic loss upon adsorption correlates with the molecule’s gas-phase translational and rotational entropies and that the occupiable volume of a zeolite is a useful descriptor in predicting such losses.

Recently, we have implemented TraPPE forcefields[3] and quantified the adsorbate entropy by performing canonical Monte-Carlo integrations using the FEASST[4] molecular simulation package for one-hundred-eighty-five molecule-zeolite combinations.[5] Our results show that these linear correlations persist for our larger and more chemically diverse dataset; and are governed by descriptors of the zeolite’s topology. The ability of our simulations to qualitatively capture the trends exhibited experimentally suggests that our simulations can be used to build predictive models based on structure-property relationships.

Herein we expand our dataset to over five-thousand molecule-zeolite combinations and implement an artificial neural network using over three-hundred framework and molecule descriptors. Our model is capable in predicting the adsorption entropies with a mean absolute error of less than the universal gas constant. Our ultimate goal is to use these highly predictable models to better understand the structure-property relationships in the adsorption entropies whilst exploiting their predictive capabilities. This talk will discuss the results of training an ANN on an expanded dataset, systematic identification of descriptors capturing the underlying correlations, as well as applying alternative, highly interpretable algebraic models.

  1. Campbell, C. T.; Sellers, J. R. V. The entropies of adsorbed molecules. J. Am. Chem. Soc. 2012, 134 (43), 18109−18115.
  2. J. Dauenhauer, O. A. Abdelrahman, ACS Central Science 2018 4 (9), 1235-1243
  3. G. Martin and; J. Ilja Siepmann, J. Physical Chemistry B 1998 102 (14), 2569-2577
  4. Hatch, H. W., Mahynski, N. A., and Shen, V. K. (2018) FEASST: Free Energy and Advanced Sampling Simulation Toolkit. J. Res. Natl Inst Stan, 123, 123004.
  5. Rzepa, C., Siderius, W. D., Hatch, H. W., Shen, V. K., Rangarajan, S., and Mittal, J., J. Physical Chemistry C 2020 124 (30), 16350-16361