Predicting Ceramic Membranes Performance in Organic Solvent Nanofiltration: A Physics-guided Machine Learning Approach

Abstract

Organic Solvent Nanofiltration (OSN) is a pressure-driven membrane separation technique that selectively filters molecules dissolved in organic solvents based on their molecular size and affinity for the membrane material—rather than differences in boiling point as in distillation. This makes OSN an energy-efficient, environmentally friendly alternative to traditional thermal separation methods. Widespread industrial implementation, however, is hampered by the complex physicochemical interactions between the membrane, solvent, and solutes. Existing theoretical models, including the Spiegler-Kedem model rooted in irreversible thermodynamics, often fail to accurately predict flux (the rate of solvent transport through the membrane) and retention (the degree to which solute molecules are rejected, expressed as a percentage) or rely on empirically fitting free parameters. This absence of a predictive theoretical model results from a lack of insight into the complex underlying transport mechanisms.

This research explores how these interactions govern flux and retention by systematically varying system parameters. These measurements will not only serve to fit and adjust irreversible thermodynamics transport models but will also form the basis for data-driven methodologies [1,2]. Although deep neural networks are well-suited for recognizing patterns in high-dimensional data, they are susceptible to overfitting and physically inconsistent predictions. To overcome this, this research employs a physics-guided machine learning approach: known physicochemical laws and thermodynamic constraints are directly embedded into the neural network’s architecture, ensuring that predictions remain physically consistent.

By using Explainable AI, such as feature attribution analyses (e.g., SHAP values), the model remains interpretable: rather than acting as an opaque predictor, a so-called black-box model, it enables identification of the dominant physical driving forces behind membrane transport, serving as a source for new hypotheses and theoretical inspiration. This provides fundamental insights into the complex separation mechanism. We aim to create a robust prediction framework that accelerates the economic valorization of OSN in industry. By optimizing operational costs in their application and minimizing the need for time-consuming pilot experiments, the ‘time-to-market’ is shortened and the threshold for industrial implementation is lowered.

[1] Piccard P-J, Borges P, Cleuren B, Buekenhoudt A and Hooyberghs J 2025 J. Membr. Sci. 735 124509 [2] Linsen W, Piccard P-J, Hooyberghs J and Buekenhoudt A 2026 J. Membr. Sci. Lett. 6100111