Domain Knowledge-Driven Machine Learning Model for Performance Prediction and Structural Optimization of Solid Amine CO₂ Adsorbents

Solid amine adsorbents have attracted considerable attention for their potential in CO2 capture; however, conventional approaches often fail to fully reflect the comprehensive influence of material properties on adsorption performance. In this study, we propose a feature-enhanced machine learning model (FEXGB-DNN) that integrates domain knowledge to accurately predict the CO2 adsorption capacity and optimize material structure. The model incorporates multidimensional features, including chemical composition, pore structure of the support, and adsorption conditions, and introduces a novel feature termed “amine efficiency”, inspired by feature importance analysis from conventional machine learning models. The FEXGB-DNN model demonstrates excellent generalization and predictive accuracy (RMSE = 0.40, R2 = 0.93, k = 0.91, b = 0.18). Partial dependence plot analysis quantifies the effects of individual features and reveals significant feature interactions, especially in pore structure parameters. Notably, silica-based supports exhibit optimal performance with a common average pore size of approximately 9 nm, a pore volume within 0–0.5 or 2–3 cm3/g, and an organic amine content in the range of 18–20 wt % nitrogen. Furthermore, the FEXGB-DNN model was applied to guide the performance prediction and structural optimization of a specific material, with experimental validation confirming the effectiveness of the model-driven design. This work not only provides a high-precision predictive tool but also, more importantly, reveals the complex structure–property relationships between the physicochemical characteristics of the supports and the final adsorption performance. It thereby establishes both theoretical and practical frameworks for the design of advanced materials for CO2 capture.