Machine learning predictive models of CO₂ adsorption in sustainable waste-derived activated carbon

This study aimed to develop and evaluate machine learning (ML) models to predict CO₂ adsorption capacity using a wide array of activated carbons from diverse waste materials. The main goal was to create robust predictive tools and assess feature sensitivity to identify the key parameters influencing adsorption behavior. Several ML algorithms, such as Random Forest, Gradient Boosting Regressor, XGBoost, Support Vector Machine, and Artificial Neural Networks, were trained and tested to assess their predictive accuracies. A hybrid modeling framework was also designed by integrating the outputs from primary models into secondary models to enhance performance. Furthermore, SHAP and Sobol sensitivity analyses were employed to interpret the model behavior and quantify the influence of each input feature. The results revealed that all individual models demonstrated strong predictive performance, with R² values ranging from 0.942 to 0.948 and test RMSE values between 0.441 and 0.501. The ANN-XGBoost hybrid model exhibited superior accuracy, achieving an R² of 0.97 and the lowest test RMSE of 0.356. SHAP and Sobol analyses consistently identified adsorption temperature and micropore volume (MPV) as the most influential parameters. At 50°C, MPV's Sobol index of the MPV exceeded 0.6, confirming its dominant impact on adsorption, and the total pore volume (TPV) and carbon content showed moderate importance. This study offers a novel approach by integrating advanced ML algorithms and global and local sensitivity to predict CO₂ adsorption. This methodology facilitates material screening and optimization, thereby contributing to the data-driven design of next-generation porous carbon adsorbents for carbon capture applications.