A Multimodal Fusion Framework for NF/RO Performance: Leveraging Molecular Graphs for Superior Predictive Modeling and Mechanistic Explanations

This study tackles two bottlenecks─limited predictive accuracy and opaque mechanistic interpretation─in modeling the rejection of organic micropollutants by nanofiltration/reverse-osmosis (NF/RO) polyamide membranes and introduces a multimodal-fusion machine-learning model within a gradient-boosting neural-network framework. By systematically integrating conventional tabular inputs─membrane structural parameters, operating conditions, and pollutant physicochemical properties─with multimodal molecular representations (graphs, images, and fingerprints), we developed three fusion schemes. The model that fuses tabular features with molecular-graph descriptors (GB+Graph) delivered the best performance (R2 = 0.9014), clearly surpassing the tabular-only baseline (R2 = 0.8494) and the other fusion variants, indicating that molecular topological information is pivotal for improving predictive accuracy. Further interpretability analysis using SHAP showed, in the global feature-importance ranking, the dominant effects of steric-hindrance-related parameters (e.g., molecular radius, molecular weight, and membrane pore size) and quantified the contributions of electrostatic interactions and hydrophobic effects to rejection. Notably, atom-level attribution analysis revealed that the model automatically highlights the central role of key functional groups in molecule–membrane interactions, providing a microstructural mechanistic interpretation beyond traditional descriptors. Overall, this work offers an advanced multimodal-fusion paradigm for accurate prediction of organic-micropollutant rejection while deepening the understanding of the coupled, multimechanistic retention process in polyamide membranes, thereby furnishing a theoretical basis and data-driven pathway for the targeted design of membranes to efficiently remove specific pollutants.