Explainable machine learning model for predicting compressive strength of CO2-cured concrete

Compared to conventional concrete, the factors to determine the compressive strength of CO2cured concrete are more complex, and thus, predicting its compressive strength becomes more difficult. Herein, an explainable machine learning (ML) model was developed to predict the compressive strength of CO2-cured concrete. A comprehensive database comprising 198 datasets was collected from published experimental investigations, and four ML algorithms were employed, i.e., RF (random forest), SVR (support vector regression), GBRT (gradient boosting regression tree), and XGB (extreme gradient boosting). To enhance model accuracy and efficiency, K-fold cross-validation and grid search techniques were utilized for hyper-parameter tuning. Furthermore, to resolve the black-box issue associated with ML models, the SHAP (SHapley Additive exPlanations) method was applied to explore the underlying relationships among variables. Overall, all ML models (RF, GBRT, XGB) in this study except SVR proved capable of efficiently predicting the compressive strength of CO2-cured concrete, with XGB being chosen for further analysis in combination with SHAP due to its superior generalization ability. The SHAP analysis reveals that adding cement content is the key driver for increasing the compressive strength of CO2-cured concrete. In terms of CO2 curing parameters, prolonging CO2 curing durations moderately could improve the compressive strength, which can be attributed to the enhancement of carbonation degree. However, higher CO2 pressures may decrease the strength due to the increased risk of microcrack propagation caused by the synergistic effects of excessive pressure and considerable heat from the carbonation reaction.