24AI AI-Assisted Polymer Design for Carbon Capture

Carbon capture technologies are an important tool for mitigating climate change. In recent years, polymer membrane separation methods have emerged as a promising technology for separating CO2 and other greenhouse gases. Highly CO2-permeable membranes with moderately high CO2/gas selectivity are necessary for energy and cost-efficient membrane-based separation. Increasing the CO2 permeability and selectivity of membranes is crucial for increasing the overall process throughput, reducing energy costs, and removing significant amounts of CO2. However, a negative relationship exists between selectivity and permeability (known as the Robeson bound), which is why very few polymers are suited for this purpose. Moreover, designing new polymers often relies on a slow experimental process and/or large-scale, costly simulations.

Our objective in this work is to develop an AI-driven approach to simplify the problem of that simplifies the complex problem of designing polymer membranes focused on CO2 separation from air. As N2 and O2 are the dominant components of the atmosphere (comprising more than 99\% of gases), high membrane selectivity for CO2/N2 and CO2/O2 pairs is vital for this task, so we focus on identifying polymers with high selectivity and permeability metrics for these gas pairs. We develop a framework composed of an AI-based prediction engine (PPE) for fast property prediction combined with an RL-based model for predicting molecules.

The property engine is created by training an ensemble multi-task model on experimental data. These models are trained with classical features as well as representations from the latest foundation models. The primary objectice of the PPE is predicting a polymer's permeability for all three gases - CO2, N2, and O2 as well as its CO2/N2 and CO2/O2 selectivity. The generative model is first trained on a large database of monomers and dimers to specialize in generating polymer specific molecules. This generative model is then adapted with an RL-based algorithm, specifically RLHF, to design new polymers. The feedback for the RLHF approach is obtained using the PPE. Effectively, the generative model is fine-tuned to sample from the configuration space of molecules with high predicted permeability and specificity. We then evaluate the performance of the predicted molecules with respect to the Robeson upper bounds.

Using our overall approach, we identify new polymer substrates that show promise for both CO2/N2 and CO2/O2 separations. Dynamics simulations on these candidate polymers support the ML-predicted gas permeabilities predicted by the AI model. Overall result suggests that AI based search can be used to discover new polymers and chemical moteifs optimized for carbon capture.