(622d) Machine Learning Guided Discovery of Polymer Membranes for Reducing Greenhouse Gas Emissions

Developing polymer membranes which have high gas permeability and selectivity remains a grand challenge for energy, the environment, and economic sustainability. Our main objective with this project is designing polymer membranes with optimal permeability and selectivity by Machine Learning (ML) assisted polymer engineering and high throughput screening. We hypothesized that ML models can accurately predict gas permeability and selectivity of both existing and hypothetical membranes when trained on large and diverse data. First, we created a library of polymers and studied them in detail by calculating their physicochemical properties. Next, we employ unsupervised ML to study the structural similarities of polymers and map their known materials space. Then, we conduct a feature selection protocol to pick the optimal features that correctly represent the similarities and differences of our polymer materials. Once the features are identified, we use them to train an ensemble of different ML models. Using appropriate validation techniques, we select the top-performing ML model that can correctly predict gas diffusivity in polymer membranes. Later, we will use a Genetic Algorithm (GA) – guided by our trained ML model – to effectively sample the polymer materials space and identify high performance polymers for CO2/N2 separation. Finally, we will use experiments to validate the new materials identified by this procedure. This work delivers a robust hypothetical polymers database and a computational framework to further explore the polymer materials space for any gas separation problem of interest.