Machine-based learning of predictive models in organic solvent nanofiltration: Pure and mixed solvent flux

Abstract During the last decades, the interest in organic solvent nanofiltration (OSN), both in academia and industry, increased substantially. OSN provides great potential for an energy-efficient separation of complex chemical mixtures with dissolved solutes in the range of 200–1000 Dalton. In contrast to conventional thermal separation processes, the pressure-driven membrane separation operates at mild temperatures without energy intensive phase transition. However, the complex interaction of different phenomena in the mass transfer through the membrane complicate the prediction of membrane performance severely, such that OSN is virtually not considered as an option in conceptual process design. Several attempts have been made to determine predictive models, which allow the determination of at least pure solvent flux through a given membrane. While these models correlate different important physical properties of the solvents and are derived from physical understanding, they provide a limited accuracy and not all of their parameters are identifiable based on available data. In contrast to previous approaches, this work presents a machine learning based approach for the identification of membrane-specific models for the prediction of solvent permeance. The data-driven approach, which is based on genetic programming, generates predictive models that show superior results in terms of accuracy and parameter precision when compared to previously proposed models. Applied to two respective sets of permeation data, the developed models were able to describe the permeance of various solvents with a mean percentage error below 9% and to predict different solvents with a mean percentage error of 15%. Further, the method was applied to solvent mixtures successfully.