Data-driven predictions of complex organic mixture permeation in polymer membranes

Abstract

Membrane-based organic solvent separations are rapidly emerging as a promising class of technologies for enhancing the energy efficiency of existing separation and purification systems. Polymeric membranes have shown promise in the fractionation or splitting of complex mixtures of organic molecules such as crude oil. Determining the separation performance of a polymer membrane when challenged with a complex mixture has thus far occurred in an ad hoc manner, and methods to predict the performance based on mixture composition and polymer chemistry are unavailable. Here, we combine physics-informed machine learning algorithms (ML) and mass transport simulations to create an integrated predictive model for the separation of complex mixtures containing up to 400 components via any arbitrary linear polymer membrane. We experimentally demonstrate the effectiveness of the model by predicting the separation of two crude oils within 6-7% of the measurements. Integration of ML predictors of diffusion and sorption properties of molecules with transport simulators enables for the rapid screening of polymer membranes prior to physical experimentation for the separation of complex liquid mixtures.

Machine Learning-Assisted Design of Thin-Film Composite Membranes for Solvent Recovery

“…In the field of membranes, it has been increasingly used for gas separation, water treatment, biofuel purification, , and other liquid separation . Nevertheless, only a few ML studies were reported on OSN, specifically, to predict solvent permeances in a polydimethylsiloxane membrane and solute rejections in PuraMem S600 and PuraMem 280 membranes, to evaluate the OSN performance of commercial and polyimide membranes, to predict the rejections of more than 400 solutes in three polyamide membranes, to optimize TFC membranes for OSN, to examine solvent impact on solute rejection, and to predict the separation of complex organic mixtures in linear polymer membranes . Most of these ML studies primarily investigated the effects of operating conditions as well as solvent and solute properties on OSN performance in a handful of membranes, and they could not be applied to design new membranes or predict OSN performance of membranes with new structures because the chemical structures of membranes were not considered.…”