Machine-based learning of predictive models in organic solvent nanofiltration: Solute rejection in pure and mixed solvents

Organic solvent nanofiltration (OSN) offers great potential for the separation of solvent–solute mixtures with dissolved components in the range of 200–1000 Dalton. The pressure driven membrane process allows for a separation at mild temperatures without phase transition. Yet, industrial applications of OSN are still scarce. Besides the demonstration of long-term stability also a lack of conceptual design tools that enable predictions of solvent flux and solute rejection are considered as major obstacles. For the latter, a few phenomenological models have been developed, primarily for ceramic membranes, based on the assumption of a dominating size exclusion mechanism. However, especially for polymeric membranes the mutual interactions between membrane material, solvent and solute, need to be accounted for in order to accurately describe solvent flux and solute rejection. The dominating phenomena may strongly depend on the specific membrane, as well as the considered chemical systems. Building on our previous work on the automatic development of membrane-specific models for pure and mixed solvent flux, this article addresses the extension towards the prediction of solute rejection in pure and mixed solvents. For this purpose, the rejection of a variety of solutes was determined experimentally for six different solvents using a PuraMem S600 membrane and automatically derived model candidates were evaluated with respect to accuracy and parameter precision. Furthermore, the possibility to predict solute rejection in mixed solvent systems was evaluated based on the concept of membrane rejection maps, evaluating previously reported permeation data for a PuraMem 280 membrane. The derived models show excellent accuracy and decent parameter precision, with deviations between measured and predicted solute rejection of less than 10% for most of the experimental data, while predictions for mixed solvents are in a similar range.

Subjects

Organic solvent nanofiltration

Predictive models

Solute rejection

Solvent resistant nanofiltration

MLE@TUHH

Options

Machine-based learning of predictive models in organic solvent nanofiltration: Solute rejection in pure and mixed solvents