Physics-Guided Neural Networks to Predict Ceramic OSN Membrane Performance

Abstract

Industrial acceptance of organic solvent nanofiltration (OSN) as a separation technique is hampered by the slow, trial-and-error membrane screening process for each solute-solvent pair [1, 2]. Such extensive experimental screening is still necessary due to limitations of predictive models, which are generally based on separations in water [2, 3]. The complexity and variety of competing interactions of all solute-solvent-membrane affinities challenge our understanding and accuracy of predictions of the underlying separation mechanism. Recently, data-driven techniques showed their potential to enhance predictability in the field of OSN [3]. We explore the use of data-driven techniques specifically for ceramic membranes, both unmodified and functionalized. The absence of swelling and compaction in ceramic membranes eliminates complexities induced by solvent-dependent structural changes. Hence, the separation in ceramic membranes is expected to be simplified, making them ideal candidates to model and study to gain a deeper understanding of their transport. On the one hand, we employ statistical methods on diverse datasets to identify the true drivers of the separation mechanism, and to identify and understand relations among system properties and membrane performance [4, 5]. Additionally, we assess a decrease in solute-membrane interactions when methyl-grafting titania membranes, reveal a consistent relationship between retention and flux (effectively modelled by Spiegler-Kedem theory) [4], and identify [4] and later decouple [5] spurious relationships between solvent Hansen solubility and molecular size (which was detected as the true driver) present in common OSN solvents. On the other hand, we employ machine learning models to predict the retention and flux of ceramic membranes, and we investigate hybrid models that guide neural networks with physical laws and knowledge obtained from the foregoing statistical analysis. Combining physical knowledge of the system with machine learning provided both explainability and improved performance [6]. Understanding and predictability of OSN can lead to well-understood designs and leverage the acceptance of the technology and its contribution to sustainable chemistry. We believe data-driven models can provide the predictive power sought for in OSN and help unravel its complex transport process.

References: [1] Dangayach et al. (2024), Environmental Science & Technology, DOI: 10.1021/acs.est.4c08298. [2] Galizia & Bye (2018), Frontiers in chemistry, DOI:10.3389/fchem.2018.00511 [3] Piccard et al. (2023), Separations, DOI:10.3390/separations10090516. [4] Piccard et al. (2025), Journal of Membrane Science, DOI:10.1016/j.memsci.2025.124509. [5] Linsen & Piccard et al. (2026), Journal of Membrane Science Letters, DOI:10.1016/j.memlet.2026.100111. [6] Piccard et al. (2026) Manuscript in preparation.