Accelerating inverse design of thin-film composite polyamide membranes for water purification via fragment-based machine learning

Thin-film composite polyamide (TFC PA) membranes are renowned for their excellent liquid separation capabilities. Developing novel monomers through the relationship between the chemical structures and performance will help accelerate the development of TFC membranes. Here, machine learning (ML) is employed to guide the rational design of TFC PA membranes. A comprehensive dataset integrating monomer fragments, synthetic and operation conditions, and properties of both membranes and solutes was used to train ML models. Fragment-based Catboost models exhibit robust predictive capability for both permeance and solute rejection. The feature contribution analysis revealed that the frgament of monomers such as amine and methylene groups, as well as pore size and anion radius, are paramount determinants for performance. Leveraging the well-trained model for multi-objective screening, top-performing membranes are efficiently identified from a pool of 133864 candidates. Experimental validation confirmed that two ML-identified TFC PA membranes exhibited enhanced permeance while sustaining high Na₂SO₄ rejection. This work presents a novel approach that shifts the paradigm in the inverse design of TFC membranes and can be extended to other material systems for accelerated materials discovery.