Machine learning estimation of CO2 sequestration for mineralized artificial aggregates derived from industrial solid wastes.

CO2 mineralization of artificial aggregates derived from industrial solid wastes serves as one of the most promising strategies for simultaneous solid waste valorization and carbon sequestration. However, accurate prediction of CO2 sequestration is challenging due to the complex nonlinear interactions between raw material properties and process parameters. In this study, a data-driven machine learning (ML) framework based on experimental datasets from reviewed studies was developed. Furthermore, six ML models including Gradient Boosting Decision Tree (GBDT), Light Gradient Boosting Machine (LightGBM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and Multi-Layer Perceptron (MLP) were evaluated with data augmentation and Bayesian hyperparameter optimization. The tree-based ensemble models achieved outstanding prediction performance, with test-set R2 values of 0.92-0.93 and small deviations from the cross-validation results, demonstrating high accuracy and stable generalization ability. SHapley Additive exPlanations (SHAP) analysis revealed that calcium oxide content, silica content, particle size distribution, and CO2 concentration were the dominant features affecting CO2 sequestration. Silica had an inhibitory effect on CO2 sequestration by forming calcium silicate hydrate (C-S-H) at a low Ca/Si ratio during hydration. Finally, a graphical user interface integrated with the optimal models was constructed to quantitatively evaluate CO2 sequestration in engineering applications. This study provides a transferable and interpretable framework for the design of low carbon artificial aggregates, supporting the synergistic valorization of industrial solid wastes and CO2 sequestration.