Machine Learning Analysis and Monomer Screening of Polyamide Nanofiltration Membranes for Lithium Separation

Abstract

Nanofiltration (NF) membranes are increasingly being used to achieve precise solute-solute separation. These membranes are commonly synthesized using interfacial polymerization, offering great potential to separate lithium from magnesium. In this study, we have developed machine learning models that relate fabrication conditions, membrane properties, and operational conditions of NF membranes to predict water permeability and lithium/magnesium selectivity. Morgan fingerprints (MFs) and molecular descriptors (MDs) are used to represent the chemical and physical properties of the monomers. Explainable artificial intelligence tools such as Shapley additive explanations (SHAP) and partial dependence plots are used to evaluate the effects of the synthesis conditions and membrane properties on membrane performance. Based on the insights obtained from SHAP analysis, we developed a material screening approach to find promising monomers from a list of amines and cation-based ionic liquids. We construct a reference MF using the functional groups that positively contribute to membrane performance and compute a screening score that favors potential candidates with more desirable MDs. Finally, the synthesizability of these monomers is assessed using the synthetic accessibility score to find the most promising candidates. We compared the performance of screened monomers against traditional ones to validate the reliability of our approach. The results of this study provide critical insights into the relationships between synthesis conditions, membrane properties, and performance and establishes a novel, strategic framework for rational screening of monomers for NF membrane synthesis. This approach holds promise to accelerate the discovery of high-performance membranes tailored for specific separation challenges, thereby advancing the field of membrane technology.