Machine learning–driven inverse design of nanofiltration membranes for selective Mg2+/Na+ separation

Desalination brine, a byproduct of seawater desalination, contains valuable resources, including magnesium. Recovering magnesium requires selectively retaining Mg 2+ while allowing the dominant monovalent cation, Na + , to permeate. Nanofiltration (NF) membranes fabricated by interfacial polymerization are promising for this Mg 2+ /Na + separation, but their performance depends on coupled formulation and processing variables that are inefficient to optimize through trial and error. Here, we demonstrate a workflow integrating dataset curation, forward prediction, and inverse design of thin-film composite NF membranes. A dataset of 1,253 entries, comprising literature and in-house records, was cleaned, imputed, and augmented through chemically constrained permutation of interchangeable polymer, solvent, and salt variables in the training split only. Optimized extremely randomized tree regressors predicted pure water permeability, solute rejection, and intermediate membrane properties, achieving test-set coefficients of determination above 0.79. Shapley analysis indicated that permeability was governed primarily by formulation and processing variables, whereas rejection depended mainly on solute identity and charge, as well as membrane tightness. Using Bayesian optimization with a scalarized objective, two membrane recipes with favorable Mg 2+ /Na + selectivity were identified and experimentally validated. The inversely designed membranes exhibited the required Mg 2+ /Na + selectivity, demonstrating the capability of this workflow to design Mg 2+ -selective NF membranes for resource recovery from desalination brine.