Machine Learning-Guided Discovery of Polymer Membranes for CO2 Separation

Abstract

Designing polymer membranes with high gas permeability and selectivity remains a grand challenge for energy, the environment, and economic sustainability. Increasing both the selectivity and permeability is a difficult multi-task constrained design problem for polymer membranes due to the trade-off between these two properties. The complexity of chemical composition and morphology of polymers makes this problem especially hard to attack with trial-and-error or intuition-based strategies. In this work, we instead present a machine learning (ML)-driven genetic algorithm to tackle the design problem of polymer membranes for CO2 separation from N2 and O2. Using literature data of permeability for three gases, CO2, N2, and O2, we constructed multiple ML models using different fingerprinting featurization schemes to predict all three gas permeabilities as well as the CO2/N2 and CO2/O2 selectivity values. Then, we employed a genetic algorithm to design new polymers and evaluated their performance with respect to the Robeson upper bounds using our machine learning models. We were able to identify new polymer membranes that are promising for both CO2/N2 and CO2/O2 separations. The top discovered polymers are predicted to have high glass transition temperatures, Tg. Similarly, the pyridine functionality was found in ~ 20% of the predicted polymers. Both of these facts are well in line with currently accepted experimental wisdom for CO2 based separations. The framework developed here can be used to design polymers for any application involving constrained optimization. Finally, we outlined the strengths and limitations of this approach, as well as the imminent challenges and opportunities with using machine learning guided data-driven inverse design of polymers.