Ensemble machine learning and Shapley additive explanations for the ability of C-S-H seeds to accelerate cement hydration

Due to the complexity of the reaction mechanism of calcium-silicate-hydrate (C-S-H) seeds in cement, the influence pattern of various factors, especially Ca/Si, on the acceleration ability of C-S-H seeds is controversial. This study explored the use of ensemble machine learning and Shapley additive explanations (SHAP) to identify the potential relations between various factors and the acceleration effect of C-S-H seeds. Based on the study results and a re-examination of the literature, it was found that the inconsistent particle size (D₉₀) of C-S-H seeds is the main reason for the controversial Ca/Si impact pattern. The D₉₀ value of C-S-H seeds should not exceed 500 nm to achieve a good acceleration effect, as indicated by a positive SHAP value below 500 nm and a negative value above 500 nm. Additionally, the content of Al₂O₃ in cement was found to positively affect the acceleration effect of C-S-H seeds, while the SiO₂ content exhibited a negative impact. This suggests that C-S-H seeds have a better acceleration effect on aluminates than on silicates. Furthermore, the feature dependence analysis revealed that the optimal Ca/Si of C-S-H seeds may vary depending on the chemical composition of the cement. Based on these findings, new insights were proposed on the acting mechanism of C-S-H seeds in cement. It is recommended to design the optimal Ca/Si and the chemical structure of C-S-H seeds based on the chemical composition and hydration products of the cement.