Interpretable machine learning for compressive strength prediction and parameter interaction analysis in carbonatable binders under small-data conditions

Accelerated carbonation of carbonatable binders offers a promising route to reduce cement industry CO2 emissions, but performance optimization is hindered by complex parameter interactions and limited experimental data. In this study, an interpretable machine learning framework integrating exhaustive feature selection, Bayesian optimization, and SHAP analysis was developed for compressive strength prediction and parameter interaction analysis of carbonatable binders. An optimized 11-feature XGBoost model achieved an RMSE of 11.68 MPa and R2 of 0.87. SHAP identified curing time, water-to-solid ratio, and CO2 pressure as dominant predictors and indicated that the modeled contribution of CO2 pressure varied with curing time. Targeted experiments on γ-C2S showed a non-monotonic pressure–strength response at 1 and 24 h: the highest compressive strength was obtained at 0.3 MPa, whereas increasing the pressure to 0.4 MPa resulted in lower strength. This study establishes a data-driven framework for accelerating the development of high-performance carbonatable binders under small-data conditions.