A Machine Learning-Assisted Framework for Performance Prediction and Optimization of Graphene Oxide Nanofiltration Membranes

Graphene oxide (GO) holds considerable promise for nanofiltration applications, owing to its distinctive two-dimensional layered architecture and tunable surface chemistry. However, the separation performance of GO nanofiltration membranes is influenced by the interplay of various structural and surface parameters, and the structure-performance relationship remains inadequately quantified. To address this, a comprehensive data set encompassing a variety of membrane structural and chemical parameters was constructed in this study. By combining advanced machine learning models with Bayesian optimization algorithms, the water flux and ion rejection rates of GO nanofiltration membranes under different parameter conditions were accurately predicted. The results demonstrated that the established models exhibited excellent predictive performance, with the XGBoost model performing the best and achieving a coefficient of determination (R 2 ) of 0.92. Further interpretability analysis revealed that zeta potential and interlayer spacing were key factors influencing the separation performance of GO membranes. Additionally, by integrating SMILES representation and Morgan fingerprint analysis, it was identified that amino-functionalized cross-linkers have a significant impact on membrane performance. Based on the model predictions, we systematically explored the parameter space and identified 14 optimal parameter combinations that achieved an optimal balance between water flux and ion rejection rate. This study presents an efficient data-driven approach, which provides insights into the quantitative prediction and structural optimization of GO nanofiltration membranes. This approach also offers valuable guidance for the rational design of high-performance nanofiltration membranes.