(379g) Molecular Simulations and Machine Learning for Multicomponent Adsorption: BTEX Separation with Zeolite Membranes | AIChE

(379g) Molecular Simulations and Machine Learning for Multicomponent Adsorption: BTEX Separation with Zeolite Membranes


Josephson, T. R. - Presenter, University of Minnesota
Sun, Y., University of Minnesota
Fetisov, E., University of Minnesota
Johnson, J., ExxonMobil Research and Engineering
McCool, B. A., ExxonMobil
Daoutidis, P., University of Minnesota-Twin Cities
Siepmann, J., University of Minnesota-Twin Cities
Liu, J., University of Minnesota
Tsapatsis, M., Johns Hopkins University
Separation of p-xylene from mixtures of benzene, toluene, ethylbenzene, and xylenes (BTEX) is a critical industrial process, identified as one of “seven chemical separations to change the world” [1]. Membranes that permit selective transport of p­-xylene, but not its isomers, o-xylene, m-xylene, and ethylbenzene, could dramatically improve the efficiency of p-xylene production. Zeolite membranes are stable under the high temperatures and pressures of the catalytic process; membranes synthesized with the MFI framework enable high selectivity and flux for p-xylene separation when tested at pressures < 1 bar and at temperatures up to 425 K [2]. However, at high temperatures and pressures, adsorption of aromatics into zeolite membranes saturates the pores, shutting down transport.

Process design of a p-xylene-selective membrane reactor requires knowledge of adsorption and transport through the membrane across a range of industrial process conditions. To address the question of multicomponent adsorption into MFI zeolite at elevated temperatures and pressures, we performed high-throughput Monte Carlo simulations in the Gibbs ensemble to predict BTEX adsorption across the expansive “process space” of potential operating conditions. The transferable potentials for phase equilibria (TraPPE) force field was extended to accurately predict vapor-liquid coexistence curves for alkyl-substituted aromatics. Then, unary and multicomponent adsorption isotherms and fluid phase properties were predicted across a multidimensional composition space, and from liquid, vapor, and supercritical phases. These simulations provide individual data points that are used to fit an artificial neural network, which provides a single, self-consistent, continuous, and differentiable hypersurface describing the properties of both the fluid and adsorbed phases across the full range of realistic process conditions.

[1] D.S. Sholl and R.P. Lively, Nature, 532, 2016, 436-437.
[2] M.Y. Jeon, et al. Nature, 543, 2017, 690-694.