A machine learning based framework to tailor properties of nanofiltration and reverse osmosis membranes for targeted removal of organic micropollutants.

Nanofiltration (NF) and reverse osmosis (RO) membranes play an increasingly important role in the removal of organic micropollutants (OMPs), which puts higher demands on the customization of membranes suitable for OMPs removal based on the rejection mechanisms. Here, the pathways of OMPs-targeted optimization for membranes were constructed by using machine learning (ML) to capture the correlations between OMPs removal efficiency with properties of membranes and OMPs. Through expertise assistance and rigorous modeling methodology, an accurate and robust Extreme Gradient Boosting (XGBoost) model was established, which could well recognize the dominant rejection mechanisms of OMPs (i.e., the size exclusion effect and electrostatic interactions). An exemplary application to another dataset of several high-risk OMPs showed how the optimized model could be used to estimate the overall efficiency of OMPs risk control and, more importantly, to provide quantitative guidance on membrane properties for specific removal targets. The satisfying prediction results demonstrated the good generalization of the ML model and consequently its ability to sensitively define the ideal membrane properties for the targeted removal of these (and any other concerned) OMPs. This study provides a feasible and universal ML-based framework to achieve the tailored selection and design of NF/RO membranes for OMPs risk control.