Machine Learning-Driven Prediction of Li+/Mg2+ Separation Performance in Crown Ether-Modified Graphene Oxide Membranes

The efficient separation of Li+ from Mg2+ in salt lake brines is a critical bottleneck for sustainable lithium extraction. Crown ether (CE)-modified graphene oxide (GO) membranes offer enhanced ion selectivity via specific host–guest recognition. However, their performance is governed by multiple structural descriptors, including CE content, interlayer spacing, asymmetric charge, and membrane inclination, which are difficult to optimize using molecular dynamics (MD) simulations alone due to high computational cost and limited quantitative predictive capability. To address this challenge, this study integrates MD simulation with machine learning (ML) to construct a high-accuracy proxy model for predicting the Li+/Mg2+ separation performance of CE-functionalized GO membranes. Using Random Forest (RF) and Extreme Gradient Boosting (XGBoost) regression models, we established quantitative relationships between four key structural descriptors and three performance indicators: water flux, Li+ permeability, and Mg2+ retention rate. RF outperforms XGBoost, achieving high test accuracy. Feature importance reveals distinct mechanisms: water flux is governed by interlayer spacing and membrane inclination, Li+ permeability is co-determined by interlayer spacing and CE number. Mg2+ retention depends mainly on CE grafts and non-uniform charge distribution, reflecting synergy between Donnan effect and specific recognition. Moreover, interactive effects among structural parameters are identified, CE number couples with spacing to enhance Li+ permeability, and with non-uniform charge to boost Mg2+ retention, providing quantitative evidence for the proposed separation mechanisms. Multi-objective optimization yields two membrane schemes, the optimally balanced design achieves both high selectivity and competitive flux, showing strong application potential. This study not only overcomes the limitations of conventional MD simulations but also establishes a data-driven framework for the rational design and efficient optimization of high-performance lithium-selective membranes.