A machine-learning workflow predicted resin performance more accurately when laboratory measurements and literature records were combined. Mean test R2 was 0.656 for the combined data, compared with 0.614 for active-learning experimental data alone and 0.414 for literature-only models. The result comes from held-out model tests, not a conventional treatment-control experiment.
That is the central finding from a study designed to tackle a data problem in direct air capture. The researchers aimed to reconstruct the full response surface for resin-based moisture-swing adsorption from fragmented data, using an exploration-oriented strategy to choose measurements where the model was most uncertain.
Choosing what to measure next
The active-learning loop began with 19 randomly selected conditions. It trained a Gaussian Process Regression model, which estimates both a likely result and the uncertainty around that estimate, then selected batches of high-uncertainty candidates for new measurements. After each batch, the model was retrained until the added improvement fell sharply.
Those measurements were part of a unified set of 290 data points: 166 from internal experiments and 124 extracted from the literature. They covered 20 resins and 201 unique experimental conditions, giving the model a broader but uneven view of the material and operating space.
After data cleaning, the experimental portion contained 166 data points from 77 independent conditions. Eighteen conditions came from the active-learning test set, while 59 came from the iterative design cycles that followed the 19-condition starting set.
The target was equilibrium CO2 uptake, or the amount captured after the measurement reached its equilibrium state. The researchers calculated it from the difference in equilibrium CO2 concentration between a resin sample and a blank system, then normalized the result by resin mass and system volume. The blank test was used to account for possible effects from the measurement system itself.
The gain was real, but not unlimited
Performance improved as the active-learning rounds added data, then began to level off. By the fifth iteration, the model reached an R2 of 0.614 on the active-learning test set, while the marginal improvement in that round was only 0.039. In a retrospective comparison using matched-sized samples, active learning performed better than random selection.
Across the broader model comparison, the strongest combined-data results came from XGBoost and LightGBM, with test R2 values of 0.703, with a standard deviation of 0.117, and 0.702, with a standard deviation of 0.108, respectively. LightGBM was chosen for interpretation, and 97.2% of the data points fell within the model's applicability domain, meaning the range of cases judged to have adequate support from the available data.
Those scores describe predictive performance, not a proven mechanism. The leading models fit their training data better than their test data, and the authors said some overfitting could not be ruled out. The dataset was also limited, heterogeneous, and imbalanced, so the figures should not be read as a guarantee that the model will transfer to very different resins, anions, or operating regimes.
What the model highlighted
To see which inputs the selected model relied on most, the study used SHAP analysis. Functional-group type ranked as the leading feature, followed by initial CO2 concentration; humidity and ionic alkalinity were also prominent. Strongly basic functional groups and more alkaline ions received higher model-attributed capacity contributions.
Grouped SHAP analysis found a negative interaction between the anion's pKa, a chemistry measure used in the model, and functional-group type, and a positive interaction between initial CO2 concentration and humidity. Here, interaction means the model did not treat the features as simply additive. The signs describe patterns in the model, not proof that the variables cause chemical synergy or inhibition.
A shortlist, not a finished process
The model was then used to examine conditions from 10% to 90% relative humidity. Carbonate-loaded A400 and A510 were selected for moisture-swing working capacity, while hydroxide-loaded HPR4100 and IRA910 were selected for maximum equilibrium uptake. In prospective tests, measured capacities generally followed the predicted ranges and the predicted humidity trends.
That validation is useful but narrow. The experiments were closed-chamber batch measurements of equilibrium capacity, not a full test of how a direct-air-capture system would operate over time. The study therefore does not establish kinetics, cyclic stability, regeneration performance, water management, mechanical durability, or process-level energy use.
The authors present the workflow as a way to fill blind spots in fragmented adsorption datasets, not as a substitute for larger experiments. Future work would need to test whether the recommendations hold for different materials and in flow-through, cyclic, and larger-scale systems, where practical performance may depend on more than equilibrium capacity.
Paper data and sources
Original title: Active Learning-Guided Optimization of Moisture Swing Adsorption for Direct Air Capture
Authors: Yu P, Zhang K, Shi X et al.
Journal/Repository: Environmental science & technology
Status: Peer-reviewed
First online: 2026-10-19
DOI: Not available
Original paper · Full text