Prediction of Unary Adsorption Isotherms in Zeolites Using Neural Networks

Student

Ramanish Singh

Advisor

Ilja Siepmann

Abstract

Zeolites, nanoporous materials having pore sizes of the order of molecular dimensions, are used in storage and separation processes in the chemical industry. Identifying a suitable zeolite, from around 235 known zeolite structures, for a particular application involving a molecule requires the knowledge of adsorption equilibria (molecules adsorbed in the zeolite as a function of temperature and pressure) of the specific molecule in the zeolite over a wide temperature range. Different molecules (hydrogen, methane, perfluoroethane, etc.) can have distinct optimal conditions (pressure and temperature) for a desired application. Determining the optimal operating conditions requires the knowledge of adsorption equilibria data traditionally computed using Gibbs ensemble Monte Carlo (GEMC) simulations. However, the GEMC simulations can become intractable for the combinatorially large design space of zeolites and molecules. Sun et al. developed a neural network model to predict hydrogen adsorption data in all the known zeolite structures by training the model on hydrogen adsorption data for each zeolite. Since different molecules can have similar adsorption behaviors in zeolites, we extend their model to other molecules such as methane and perfluoroethane, thereby increasing the extensibility of the NN model. We generated methane adsorption data in 235 zeolites to be used as training data for the model. The combined hydrogen and methane adsorption data was then used to train the neural network model and test few-shot learning predictions. We compare the neural network predictions with the base case of fitting the data to well known analytical adsorption isotherm function. We observe that both techniques give similar accuracy when the whole data is used to train the model and fit the adsorption isotherm function. However, in the case of few-shot learning, machine learning performs marginally better than fitting an adsorption isotherm function. We are currently working on generating data for perfluorethane, a more complex molecule than hydrogen and methane. Perfluoroethane is bulkier and more polar as compared to the other two molecules. Therefore, including perfluoroethane data in the training set will help to make the model more transferable.