Mechanism‐embedded machine learning unravels lithium‐controlled selectivity in nanofiltration for Li ⁺ /Mg ²⁺ separation

Abstract

Extraction of lithium from magnesium‐rich salt‐lake brines using polyamide nanofiltration (NF) membranes is hindered by incomplete mechanistic understanding of ion separation in mixed‐salt environments. Conventional models struggle to decouple the intertwined effects of steric hindrance, Donnan exclusion, and multi‐ion competition, yielding contradictory design principles. Here, we develop a mechanism‐embedded machine‐learning framework that integrates empirical membrane descriptors with physics‐informed steric ( φ _S ) and Donnan ( φ _D ) partitioning factors derived from Donnan–steric pore model (DSPM). The framework reveals a paradigm shift in mixed‐salt systems: Li ⁺ /Mg ²⁺ selectivity is governed by Li ⁺ ‐controlled transport accessibility, not by the exclusion strength of Mg ²⁺ . Bivariate partial dependence plots (PDPs) identify a narrow high selectivity window, constrained to φ _D ‐Li ⁺  ≈ 2–2.35 and φ _S ‐Li ⁺  < 0.08. This work offers a quantitatively interpretable framework that reconciles single‐ and mixed‐salt separation mechanisms and demonstrates that robust Li ⁺ /Mg ²⁺ separation requires balancing between Li ⁺ permeability and Mg ²⁺ rejection, not extremizing individual membrane properties.