Predicting and understanding the performance of polyamide nanofiltration membrane for Li/Mg selective separation based on machine learning

Nanofiltration technology holds significant promise for the selective separation of monovalent and multivalent ions, such as lithium (Li) and magnesium (Mg), during Li extraction from salt lakes. Nevertheless, optimizing polyamide nanofiltration membranes for selective ion separation remains inherently challenging due to the complex interactions between ions and membrane structural and experimental parameters, making the ion transport mechanism ambiguous. This work employed a machine learning (ML) approach to identify and comprehensively understand the features that influence the membrane permeability and selectivity using a comprehensive dataset, which encompassed fabrication parameters, experimental conditions, membrane properties, and single salt rejection performance. Initially, ML algorithms accurately predicted intrinsic membrane properties using only fabrication parameters but struggled to predict permeation and selectivity when combining fabrication parameters or membrane properties with experimental conditions. To address this limitation, salt rejection performance was incorporated, and various combinations of input variables were systematically compared to identify optimal input configurations for robust ML algorithms capable of accurately predicting membrane permeability and selectivity. Using the Shapley additive explanation (SHAP) method, we found that membrane permeability was mainly determined by fabrication parameters such as substrate type and heat curing temperature. While these parameters also influenced molecular weight cut-off (MWCO) and zeta potential, they did not fully reflect the physicochemical factors governing ion separation. In contrast, MgCl₂ rejection served as a more integrative and informative descriptor, capturing both pore structure and surface electrostatic effects in predicting Li/Mg selectivity. These findings underscore the necessity of a multifaceted approach in modeling membrane performance, integrating both intrinsic properties and external factors to achieve optimal ion selectivity.