Predicting calcium carbonate yield from wet carbonation of recycled cement paste using interpretable ensemble machine learning

Wet carbonation is a promising method for enhancing the performance of recycled cement paste as a low carbon supplementary cementitious material, however, its optimisation for yield is hindered by an incomplete understanding of its mechanisms and uncertainty about the relative importance of various influencing factors. This study applies advanced machine learning techniques to optimise the wet carbonation of recycled cement paste to maximise calcium carbonate yield, utilising a dataset of 227 experimental records covering diverse process conditions. Six machine learning models—support vector regression, decision tree, random forest, gradient boosting regressor, extreme gradient boosting, and stacking—were trained and evaluated using R 2 , MAE, and RMSE as performance metrics. Model evaluation was performed using 10-fold cross-validation to enhance robustness and ensure generalisation across different data subsets. The stacking ensemble model achieved the highest predictive accuracy, with R 2 values of 0.997 (training) and 0.948 (testing), improving prediction accuracy by 19.85 % compared to the least accurate model (decision tree testing). Additionally, the stacking model achieved MAE values of 0.0101 (training) and 0.0393 (testing), and RMSE values of 0.0125 (training) and 0.0536 (testing), demonstrating its superior predictive capability. This model effectively addressed the challenge of limited data, a common issue in applying machine learning in materials science. Shapley additive explanations identified the fineness of cement paste particles and carbonation time as critical factors affecting the calcium carbonate yield variability, contributing 0.14 and 0.11, respectively. Therefore, these results show that finer particle size, extended carbonation time, and the addition of extra OH- ions in the reactor can be applied to develop more efficient wet carbonation processes, enhancing the production of CaCO 3 and CO 2 uptake. This research highlights the potential of ensemble learning to optimise concrete recycling techniques, offering a sustainable solution for the construction industry. • Six machine learning models are tested to predict calcium carbonate yield from wet carbonation of recycled cement paste. • A novel stacking ensemble model achieved R² = 0.997, outperforming other ML models in predicting carbonation yield. • SHAP is applied to analyse the influence of key input variables on calcium carbonate yield. • Cement fineness, carbonation time, and concentration of added OH⁻ ion in reactor were key factors influencing yield. • The study offers insights to improve wet carbonation efficiency, aiding sustainable cement production and CO₂ reduction.