Classical density functional theory for alkane adsorption in cationic Faujasites: comparison with grand canonical Monte Carlo simulations

Abstract

Adsorption in zeolites is important for separation processes and storage applications. Cationic zeolites have many degrees of freedom that affect their adsorption properties: pore topology, Si/Al ratio, and the type(s) of extraframework cations. Leveraging the tunability of the zeolite pore characteristics, fast and accurate models to predict the adsorption properties are required for computationally expensive studies, i.e. material screening, process modelling, and data generation for machine learning. In this work, we explore the application of classical density functional theory (DFT) based on the PC-SAFT Helmholtz energy functional, a method significantly faster than the conventional grand-canonical Monte Carlo (GCMC) simulations with comparable accuracies to experimental data, to predict the adsorption of pure and binary alkane mixtures in faujasite structures with varying Si/Al ratios. The adsorption isotherms of methane, ethane, propane, and n-butane were calculated using classical DFT and compared to GCMC. The classical DFT model uses the same forcefield for the framework sites as in the GCMC calculations. Excellent agreement was observed for methane and ethane, as well as their binary mixtures, albeit the deviations increased for longer alkane chains because of modelling of adsorbates as averaged interaction sites in classical DFT. Classical DFT on GPU shows good to excellent agreement for Henry adsorption constants and enthalpies of adsorption across various Si/Al ratios of Faujasites at a very significantly reduced computational cost, suggesting that classical DFT is a powerful tool for screening and material optimisation. For the classical DFT calculations, we adopt the implementation that uses backward mode automatic differentiation, parallelised on GPUs, which significantly reduces the computational cost of predicting adsorption. The computational costs in our work using a GPU have shown a speedup of up to a factor of 12 900 compared to state-of-the-art GCMC simulations.