MLP–CNN Integrated Active Learning Framework for Global Optimization of Biochar Pore Structure toward Efficient CO ₂ /N ₂ Separation

The pore structure exerts a critical influence on the CO2 adsorption and separation performance of biochar, and gaining a deep understanding of its underlying mechanisms is essential for developing high-performance CO2 adsorbents. However, most existing studies perform simplified characterization of pore structures using specific parameters, making it difficult to reveal the synergistic effect among pores of different sizes on CO2 adsorption performance and limiting the scientific guidance for the preparation process of adsorbents under varying operating conditions. To this end, this study employed 77 K N2 adsorption isotherms as descriptors to characterize the continuous distribution of pore structures in carbon-based adsorbents and established a global active learning optimization framework linking “preparation conditions–continuous distribution characteristics of pore structure–CO2 adsorption and separation performance.” Specifically, based on experimental data from 60 datasets of lignin-derived biochars, a Convolutional Neural Network (CNN) was trained to uncover the global linkage between continuous distribution characteristics of pore structures and CO2 adsorption performance, highlighting the cooperative role of narrow micropores and mesopores in promoting efficient CO2 capture. Furthermore, a Multilayer Perceptron (MLP) model was employed to establish the correlation between preparation conditions and pore structures, elucidating the evolution patterns of the pore structure. The model was further integrated with the Limited-memory Broyden–Fletcher–Goldfarb–Shanno with Bounds (L-BFGS-B) optimization algorithm to achieve global regulation of pore structure distributions and to conduct multiple rounds of model-guided experiments. After three rounds of active learning iterations, the CO2/N2 separation capacity of the produced biochar increased by 130.7%, with a dynamic CO2/N2 separation coefficient of 33.14. Finally, the established MLP–CNN integrated prediction framework was deployed as an interactive interface, enabling rapid mapping among the preparation parameters, pore structures, and CO2 separation performance. This study provides a novel methodological foundation for the rational design and efficient application of biochar-based CO2 adsorbents.