Preparation and performance prediction of PES/TiO₂ ultrafiltration membranes using machine learning

In this study, polyethersulfone (PES) and titanium dioxide (TiO₂) were used as membrane materials and PVP-K30 was used as a pore-forming agent. PES/TiO₂ ultrafiltration membranes were prepared using the nonsolvent-induced phase separation (NIPS) method to investigate the effects of PES concentration, TiO₂ concentration, temperature and PVP-K30 content on membrane performance. The purpose of this study is to use machine learning to conduct an in-depth analysis of the relationship between membrane fabrication process conditions and membrane performance, thereby providing a data-driven basis for ultrafiltration membrane design and process optimization.

This study first completed the preparation of PES/TiO₂ ultrafiltration membranes, measured and conducted a basic analysis of their performance. Given the complex relationship between membrane preparation conditions and performance, traditional methods struggle to fully reveal their underlying patterns. Based on this, this study introduced machine learning (ML) analysis methods and constructed three machine learning models, including gradient boosted decision tree (GBDT), random forest and extreme random tree. The authors conducted an in-depth investigation into the relationship between preparation conditions (including PES concentration, TiO₂ concentration, PVP-K30 content and temperature) and membrane performance (membrane flux and rejection rate). In addition, the authors used the shapley additive explanation (SHAP) method to provide a global interpretation of the ML models and evaluate interaction effects.

The R² values for these three machine learning models were all greater than 0.75 on both the training and test sets, indicating that the models are highly reliable. Among them, the GBDT model performed the best. The study found that, for membrane flux, temperature and TiO₂ generally had a positive effect, while PES and PVP-K30 generally had a negative effect. For rejection rate, temperature and TiO₂ generally had a negative effect, while PES and PVP-K30 generally had a positive effect. Furthermore, the study found that the optimal concentration range for PES is 18 wt%–18.2 wt% and the optimal concentration range for TiO₂ is 0.8 wt%–1 wt%.

To the best of the authors’ knowledge, this study is the first to combine the preparation of PES/TiO₂ ultrafiltration membranes with interpretable machine learning (ML), constructing three machine learning models. This approach provides a new perspective on traditional research into ultrafiltration membrane preparation and performance. Furthermore, this method was used to analyze the direction of the effects of PES/ TiO₂ ultrafiltration membrane preparation conditions on membrane performance, identifying the optimal concentration ranges for PES (18.0 wt%–18.2 wt%) and TiO₂ (0.8 wt%–1.0 wt%), thereby providing a data-driven reference for optimizing membrane formulations.