Predictive transport modelling in polymeric gas separation membranes: From additive contributions to machine learning

Membrane-based gas separation is a commercially practiced technology dominated by polymeric materials. Nevertheless, as established through the accumulation of large datasets of various polymers tested for decades, polymeric membranes suffer from an inherent permeability-selectivity trade-off. The natural culmination of such trade-off behavior has been the construction of chemical/molecular structure-transport property relationships, fueling an ongoing search for new and improved polymers. Yet, considering the time and financial costs of experimental research, it seems hard to fully harness the potential of polymer technology for developing membranes unless we switch towards data-driven prediction as a mainstream approach. Particularly the predictive models capable of estimating polymer permeation properties with high accuracy could propel the field. However, the data presence and accessibility issues hamper such a transition. Here, we provide a historical overview of the predictive models, highlighting the main incentive behind: Facilitating advanced membrane research by identifying chemical structures not studied or synthesized yet. To this end, we specifically focus on the gas transport properties of existing polymers and provide insights into their use and further development. Then, we discuss the establishment of predictive methods, which are mainly based on the representation of structural fragments constituting polymers, analysis of existing transport data, and estimation of increments for corresponding fragments. Within these predictive methods, the models based on the concept of additive contributions and machine learning approaches are particularly instrumental for handling extensive polymer databases. Still, since they complement the semi-empirical models, we also briefly touched upon non-equilibrium thermodynamics-based models for glassy polymers in our analysis. Overall, we address the advantages and challenges of using these models as a tool to identify novel polymer structures for designing high-performance membranes. We hope this review will help initiate new collaborations between membrane scientists/technologists and polymer informaticians.