Optimizing calcium leaching from cement kiln dust for indirect CO₂ mineralization: Experimental design and machine learning driven insights

The global cement industry produces a significant amount of CO₂ emissions while simultaneously generating large amounts of cement kiln dust (CKD), a highly alkaline byproduct rich in reactive calcium. CO₂ mineralization, a pathway for permanent sequestration through recovered calcium from CKD, has limited efficiency during leaching, owing to many interacting parameters. Traditional statistical models, such as response surface methodology (RSM), are useful to optimize those parameters but may misrepresent the complex, nonlinear interactions thereby limiting prediction accuracy. In this study, calcium leaching from CKD was investigated using a hybrid RSM-machine learning (ML) framework. Preliminary testing showed that stirring speed had little impact, allowing focus on ammonium chloride (NH₄Cl) concentration, temperature, leaching time, and liquid-to-solid (LS^–1) ratio. The RSM model had an adjusted R² of 0.84 (predicted R² of 0.82) providing optimized conditions of 1.5 M NH₄Cl, 40 °C, 30 min, and 10 L kg^–1. Decision Tree ML model achieved higher accuracy (training R² = 0.94, testing R² = 0.87, RMSE = 1.5-2.5), using interpretability tools, identified LS^–1 ratio and NH₄Cl concentration as the most impactful parameters. Experimental validation confirmed the robustness of ML with only 1.9 % deviation compared to the 9.8 % for RSM. Structural and compositional characterization confirmed the selective leaching efficiency of NH₄Cl. Subsequent carbonation of calcium-rich leachate successfully produced calcium carbonate (CaCO₃). Therefore, this integrated RSM-ML framework enhances insights and predictive capability for calcium leaching from CKD, a critical step in indirect CO₂ mineralization.