Thermodynamic Labeling Enables CO2 Prediction in Polymers of Intrinsic Microporosity under Small Data Constraints.

Porous polymers, particularly polymers of intrinsic microporosity (PIMs), combine high surface areas with tunable functionalities, positioning them as promising materials for post-combustion CO2 capture technologies. A key physical parameter of these materials is the isosteric heat of adsorption (Qst), which quantifies the interaction strength between CO2 molecules and the polymer framework. In this work, we demonstrate that data-driven models constructed from computationally determined physicochemical descriptors can accurately predict Qst at room temperature using a modestly sized dataset of 75 PIMs. Among the multitude of machine learning models evaluated, Kernel Ridge Regression (KRR) utilizing the radial basis function (RBF) yielded notable training and testing coefficients of determination (R2 scores) of 0.9854 and 0.9653, respectively. Additionally, an ensemble model was constructed using the KRR RBF, Lasso Regression, and XGBoost, achieving an average training R2 score of 0.9844 and a testing score of 0.9651. By optimizing model parameters through Bayesian methods and interpreting feature importance using SHAP analysis, we identified the molecular characteristics that most strongly influence the prediction of CO2-PIMs isosteric heats of adsorption. Using a Lasso model, we screened a synthetic PIM parameter space by varying six thermochemical descriptors, revealing density as the most influential factor and identifying optimal combinations to enhance predicted isosteric heats of adsorption. These results demonstrate that accurate predictive workflows can be developed using relatively small datasets for designing polymeric adsorbents by identifying key molecular descriptors that correlate with the isosteric heat of adsorption. Overall, this adaptable methodology can be extended to future gas-separation challenges, helping pave the way for faster discovery of advanced polymeric membranes for capturing CO2.