A data-driven approach to interfacial polymerization exploiting machine learning for predicting thin-film composite membrane formation

Abstract

Polymeric thin-film membranes prepared by interfacial polymerization are the cornerstone of liquid separation, with the potential to reduce industrial waste and energy consumption. However, the limited diversity of monomers may hinder further development by restricting the accessible chemical space. To address this, we propose a divide-and-conquer approach for the interfacial polymerization membrane development pipeline. We constructed a dataset using 18 organic and 73 aqueous phase monomers, conducting 1,246 interfacial reactions and analyzing membranes via AFM and optical microscopy. This unprecedentedly large and open access dataset marks a considerable step toward data-driven thin-film membrane development. We trained five machine learning models on molecular structures and density functional theory calculations to study film formation parameters and their binary outcomes. The results indicate that film formation can be predicted directly from monomers, facilitating the potential of data-driven membrane development. Our work shifts the focus from performance prediction to the fundamental step of thin-film formation, offering a new perspective in data-driven membrane research.