Active Learning-Guided Optimization of Moisture Swing Adsorption for Direct Air Capture

Moisture swing adsorption (MSA) offers an energy-efficient route for direct air capture of CO2 by exploiting humidity-driven sorbent regeneration, yet its performance remains difficult to predict due to the coupled effects of sorbent chemistry, counterion properties, and operating conditions. Here, we present a data-efficient experimental and modeling strategy that systematically maps these interactions across a high-dimensional design space. Using active learning to target regions of maximal uncertainty, we expand experimental coverage from 19 to 77 conditions and, together with curated literature data, assemble a unified dataset of 290 data points (201 distinct experimental conditions) spanning 20 commercial ion-exchange resins. Interpretable machine-learning models trained on this dataset predict CO2 capture capacity (R2 = 0.702, RMSE = 0.164) and identify functional groups, anion basicity, humidity, and CO2 concentration as dominant performance determinants. Guided by these insights, we experimentally validate model-selected resin-anion combinations for two practical scenarios─maximum CO2 capture capacity and moisture-swing efficiency. This work converts fragmented experimental observations into a predictive structure-property framework and demonstrates a generalizable approach for understanding and optimizing complex adsorption systems under data and experimental constraints.