Accelerated Discovery of High-Performance Polyamines for Solid-State Direct CO$_2$ Capture via Efficient Simulations and Bayesian Optimization

Solid amine-based sorbents are a leading approach for direct air capture (DAC) of CO$_2$, owing to their energy efficiency and scalability. To enable data-driven discovery of improved sorbents, we developed a computational framework that integrates fragment-based polymer generation with Density Functional Theory (DFT), molecular dynamics (MD) relaxations, and grand canonical Monte Carlo (GCMC) sampling. This workflow provides accurate yet efficient estimates of CO$_2$ uptake while capturing key structure-property relationships across a diverse library of polymers assembled from well-characterized polyamines for DAC. Leveraging such adsorption data, we investigated the application of the Bayesian optimization (BO) strategy in accelerating the discovery process of high-performing polymer candidates with our developed simulation workflow. Computational experimental results demonstrated the sensitivity of this discovery process to the choice of molecular representation in the surrogate models of BO, especially in a small computational budget scenario, where polymer-specific pre-training provided an early advantage over models trained for general chemical space.