3D molecular feature-based hybrid machine learning for elucidating membrane rejection of organic micropollutants

Efficient removal of organic micropollutants (OMPs) by membrane separation processes, including reverse osmosis (RO) and nanofiltration (NF), is essential for protecting aquatic environments and human health. However, the pronounced diversity in OMP molecular size, conformation, and physicochemical properties hampers the ability of existing predictive models to robustly associate molecular structure with membrane separation behavior. Here, we introduce 3D-SFDFNet, a dual-stream framework that integrates three-dimensional (3D) molecular features with structured physicochemical descriptors to predict OMP rejection by RO and NF membranes. Compared with one-dimensional (1D) and two-dimensional (2D) molecular representations, 3D molecular features more explicitly encode atomic spatial distributions and molecular geometry, and the proposed model achieved an R² of 0.911 under five-fold cross-validation. According to established membrane separation theory, RO and NF rejection are commonly interpreted using the solution-diffusion mechanism and the coupled effects of size sieving and Donnan exclusion, respectively. In this work, SHAP analysis further suggested model-level associations consistent with these theoretical interpretations: molecular size, 3D geometry, and flexibility-related descriptors contributed strongly to RO rejection prediction, whereas molecular size and charge-related descriptors showed greater contributions to NF rejection prediction. Overall, this work provides a data-driven framework for integrating 3D molecular information with membrane and operating descriptors to predict OMP rejection, offering a useful reference for preliminary membrane screening and hypothesis generation toward more effective OMP removal.