Small Data Enabled Prediction and Verification of Potential Polymer Membranes for CO2 Separation

Abstract

Carbon dioxide separation from flue gases is beneficial for mitigating climate change. Among the existing candidate techniques, separation by polymer membranes stands out as an approach that has been already successfully applied at industrial scale. For optimizing the membrane design at molecular scale, machine learning has emerged as a computational method with substantial acceleration potential in the materials discovery process. In this contribution, we report of a graph-based molecule generation algorithm which operates on limited data input and provides flexible design constraints. Data preparation routines and physical simulations are also integrated within the end-to-end, computational and optimization workflow. In our study, we have used as an input 1200 homopolymers representing multiple polymer classes and 4 material properties as optimization targets. As output, we have generated 500 molecular candidates for further evaluation by molecular dynamics simulation. Our preliminary simulated numbers demonstrate how the candidate molecular structures affect the efficiency of membranes in the gas separation process. Our end-to-end workflow provides material scientists with an application specific selection of optimized molecular designs in a reliable and efficient manner.