Designing polymer membranes with high gas permeability and selectivity presents a challenging multi-objective optimization problem due to the inherent trade-off between these two properties. In this work, we present a machine learning (ML)-driven framework for the data-guided discovery and evaluation of polymer membranes for gas separation. Using literature-reported permeability data for the target gases, we generated Simplified Molecular Input Line Entry System (SMILES) strings for polymers based on their repeating units. Molecular graphs were constructed using RDKit and processed with Molecular Representation Contrastive Distillation (MRCD), a new enhanced pre-trained graph neural network that leverages contrastive learning to generate chemically meaningful embeddings. These embeddings were then used to train a regression model to predict gas permeability. For transfer learning, MRCD-derived embeddings served as input features to a multilayer perceptron (MLP) regressor, which achieved a high coefficient of determination in predicting gas permeability. To interpret the learned molecular space, we applied t-distributed Stochastic Neighbor Embedding (t-SNE), which revealed clustering patterns associated with polymer structure and performance. To broaden the chemical design space, we embedded 10 million molecules from the PubChem database using the enhanced MRCD model and mapped our polymer library into this latent space. We then computed distances between embeddings and used Tanimoto similarity scores to compare the PubChem molecules to our polymer candidates, identifying over 10,000 structurally similar compounds with potential for high gas permeability and selectivity. Top candidates were further evaluated using molecular dynamics (MD) simulations to assess gas diffusivity and Monte Carlo (MC) simulations to estimate gas solubility. This combined MD–MC approach provided molecular-level insights into gas transport properties. High-performing candidates were integrated into an active learning loop to iteratively guide the search for additional promising polymer structures. Through this iterative, ML-guided workflow, we successfully identified novel polymer membranes with superior CO2 separation performance. This integrated ML–simulation strategy provides a robust and scalable approach for rational polymer design, accelerating the development of sustainable and efficient gas separation technologies.
