Data-Centric Approach for Modeling the Adsorption of Small Gas Molecules on Granular Activated Carbons.

Data-driven models are increasingly employed in gas adsorption studies, optimizing adsorption and elucidating mechanisms. Yet, the importance of high-quality data sets, which is crucial for modeling, is often underexplored. Focusing on small gas molecule adsorption, we showcased novel data-centric methods to improve data set quality for adsorption modeling. First, we for the first time showed that solute descriptors could be used as features for gas molecules to increase data quality by merging smaller data sets into larger ones. This approach enabled the development of satisfactory predictive models for chlorofluorocarbons/hydrochlorofluorocarbons and hydrocarbon gas molecules for the first time. Then, we showed that mostly overlooked experimental measurements (Brunauer-Emmett-Teller, BET adsorption curves) enriched the data set quality by providing more detailed characterizations for adsorbents. New models including these curves for CO2 and CH4 reduced mean-squared errors (MSE) by approximately 18%. We also raised attention to data skewness's impact on model performance. Last, we developed a new method for "actively" building suitable data sets for modeling, which aligned with results by the posterior method but without requiring training models in advance. Overall, these new techniques and findings will greatly contribute to future modeling from a data-centric perspective.