Machine Learning Driven Genetic Algorithm Optimization for Polymer Membrane Design

Designing polymer membranes with high gas permeability and selectivity is a difficult multi-task constrained problem due to the trade-off between these two properties. In this work, we 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, we constructed multiple ML models with different fingerprinting featurization schemes to predict gas permeabilities. Then, we employed a genetic algorithm to design new polymers and evaluated their performance using our ML models. Genetic algorithm is a powerful optimization tool where the evolutionary process is driven by the fitness function. Ideally, one should consider including all the property functions that are aimed to be optimized in the fitness function. Since the offspring polymers with high fitness function becomes the next generation parent polymers, with every iteration we see an increase in the fitness function. Building blocks that contribute towards this increase are transferred to the next generation and later in the evolutionary process we have a higher chance of incorporating these building blocks in the offspring polymers which results in desired properties. It is therefore of utmost importance to optimize the fitness function to direct the GA towards the targeted design area on the material design space. After multiple optimizations we identified fitness function that includes gas permeability, selectivity, and Synthetic Accessibility Score was able to guide the GA towards the targeted design area. We were able to identify new polymer membranes that are promising for both CO₂/N₂ and CO₂/O₂ separations. These polymers have pyridine, benzoxazole, benzene, and phosphonamidic acid functional groups, which we identified as the most abundant functional groups that are used experimentally. These polymers also include oxygen-, sulfur-, and nitrogen- containing motifs, similar to the experimental polymers that are on the CO₂/N₂ upper bound. Oxygen- and nitrogen- containing motifs are reminiscent of imines and polyethers, which are known to be high performing polymer membranes. One of the main strengths of this framework is the ability to optimize multiple performance metrics at the same time in the fitness function.