2025 AIChE Annual Meeting

ML?Guided Genetic Algorithm for Designing CO? Selective Polymer Membranes

Polymer membranes offer an energy-efficient and scalable alternative to traditional separation processes; however, their development remains constrained by the intrinsic trade-off between permeability and selectivity, as defined by the Robeson upper bound. The Robeson upper bound arises because enhancing gas permeability often requires increasing polymer free volume or pore sizes, which in turn reduces selectivity. Optimization through traditional experiments tend to be slow and resource-intensive, motivating the use of data-driven methods that can capture complex structure/property relationships and explore vast chemical spaces more efficiently. For our research, we introduce a machine learning-driven framework that integrates predictive modeling with an optimized genetic algorithm (GA) to accelerate the discovery of high-performance polymer membranes.

Polymer repeat units are encoded using Morgan fingerprints and employed to train ensemble regression models (e.g., XGBoost, random forest) capable of accurately predicting CO₂, O₂, and N₂ permeabilities in log-space. The trained model is embedded within a GA pipeline, where BRICS fragmentation allows polymer SMILES to undergo crossover, mutation, and migration to generate chemically diverse offspring. The overall fitness function is evaluated dynamically using ML predictions, incorporating permeability, selectivity, predictive uncertainty, and synthetic accessibility (SA) as multi-objective criteria. Additionally, we incorporate uncertainty-weighted scoring to guide exploration toward promising but under-sampled regions of chemical space, balancing exploitation of known high-performing motifs with discovery of novel polymer backbones. This iterative process enables the discovery of balanced polymer candidates predicted to surpass existing Robeson upper bounds for both CO₂/N₂ and CO₂/O₂ separations, demonstrating the framework’s ability to navigate chemical space efficiently and push performance limits.

Building on this foundation, Our future work focuses on integrating graph neural networks (GNNs) to capture molecular topology and local atomic environments beyond what fingerprint-based representations can encode. The GNN will serve as a structural learning module that refines permeability predictions and provides uncertainty-aware guidance for the GA’s search trajectory, and will decrease the time elapsed for training, molecule building and polymer generation portions of the GA pipeline. Future extensions will couple this GA platform with molecular dynamics simulations to include physically informed descriptors such as diffusivity and solubility, and to also validate the predictions given by the GA. Collectively, this framework establishes a transferable strategy for AI-guided membrane discovery, with future applications in ion-separation membranes for water treatment and sustainable process engineering.