Compressive strength prediction of carbonated recycled aggregate concrete using regression based machine learning models

Carbonation treatment of recycled concrete aggregates (RCA) enhances aggregate performance while enabling CO₂ sequestration, thereby supporting sustainable construction practices. Despite the growing application of machine learning (ML) in concrete research, predictive modeling of concrete incorporating carbonated RCA remains insufficiently explored. This study investigates regression based ML approaches for predicting the compressive strength of concrete containing carbonated recycled coarse aggregates. A total of 108 experimentally obtained data points were compiled and statistically validated prior to model development. The input variables included water to cement ratio, cement to coarse aggregate ratio, cement to fine aggregate ratio, water absorption and crushing value of natural and recycled coarse aggregates, parent concrete strength, degree of carbonation, and replacement ratio of natural coarse aggregates with recycled aggregates. Composite indices were introduced to simplify predictive formulations and to evaluate the influence of multicollinearity. Six ML algorithms namely, multilinear regression (MLR), ridge regression, polynomial MLR, decision tree, random forest, and LightGBM were developed and optimized using 5-fold cross-validation. Model performance was evaluated using mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²), supported by residual and prediction analyses. The decision tree and LightGBM models achieved the highest predictive accuracy on the test dataset, yielding RMSE of 1.59 MPa, MAE of 1.17 MPa, and R² of 0.991. The random forest model closely followed, with RMSE of 1.64 MPa, MAE of 1.25 MPa, and R² of 0.991. Polynomial MLR demonstrated moderate performance, with RMSE of 2.23 MPa, MAE of 1.89 MPa, and R² of 0.983. In contrast, conventional MLR and ridge regression exhibited lower accuracy, with RMSE values of 3.19 and 3.08 MPa, MAE of 2.65 and 2.56 MPa, and R² values of 0.965 and 0.967, respectively. The tree based ensemble models provided the most accurate and stable predictions, maintaining a maximum error range of approximately ± 13%. Although the decision tree and LightGBM models achieved marginally lower prediction errors, the random forest model was identified as the most reliable overall due to its ability to capture physically meaningful relationships, as evidenced by SHAP analysis, together with its high predictive accuracy and robustness. Composite indexing had a minimal effect on model performance. Sensitivity and SHAP analyses identified the mix proportioning index as the most influential parameter, followed by degree of carbonation and aggregate performance. Overall, the results demonstrate that ML based models, particularly tree based ensembles, effectively capture the experimental behavior of carbonated RCA concrete and enable efficient compressive strength prediction.