Polymeric membrane concentration of lithium-magnesium solution for sustainable resource recovery with machine learning

It is crucial to efficiently recover high value sources from seawater and salt-lake brines for global sustainable development goals. However, conventional membrane-based recovery methods face some significant challenges such as insufficient membrane performance and unclear mechanism. To address this issue, we propose a machine learning-driven framework with various models to systematically optimize nanofiltration membrane performance and then elucidate the underlying mechanism at the structure-process-performance nexus. To understand how the membrane and process variables affect the performance, we integrate shapley additive explanations (SHAP) and partial dependence analysis. Interestingly, we find that steric hindrance effect plays a dominant role in separation performance, which is dependent by transmembrane pressure, molecular weight cut-off, and membrane surface roughness. Specifically, high performance such as water permeance and rejection can be achieved under the transmembrane pressures of 5-10 bar and surface roughness of 3.4-3.5 nm. Through particle swarm optimization, a high water permeance of 15.12 L·m⁻²·h⁻¹ could be achieved while retaining high Li⁺ and Mg²⁺ rejection. This work could hopefully advance membrane-based resource recovery by establishing a machine learning-driven paradigm for system optimization and mechanism understanding, thus offering a reliable path for furnishing some theoretical guidance in the field.