Interpretable hybrid data-driven modeling for predicting the compres-sive strength of ready-mixed concrete

Abstract

This research uses a large dataset of ready‑mixed concrete batches collected directly from projects in Vietnam to develop a data-driven modeling framework for compressive‑strength prediction. The dataset comprises 1,157 industrial samples and includes ten input variables describing mixture proportions, aggregate gradation, admixture dosage, and fresh concrete workability. Seven models were evaluated, namely Random Forest (RF), Artificial Neural Network (ANN), Gradient Boosted Regression Trees (GBRT), Extreme Gradient Boosting (XGBoost), Model Tree (M5P-like), Support Vector Regression (SVR), and Gaussian Process Regression (GPR). All models were trained using an 80/20 train–test split, with hyperparameters optimized through Bayesian optimization to ensure robust performance. Among the evaluated models, GPR demonstrated the best overall generalization capability, achieving a coefficient of determination (R²) of 0.9399, a root mean square error (RMSE) of 3.042 MPa, and a mean absolute error (MAE) of 1.883 MPa. Extreme Gradient Boosting (XGBoost) achieved the lowest prediction error in terms of mean absolute error (MAE = 1.756 MPa) and mean absolute percentage error (MAPE = 4.57%), indicating strong predictive accuracy. To enhance model transparency, permutation-based global feature importance analysis and local surrogate interpretation techniques were applied to identify the key factors influencing model predictions. The results reveal that the aggregate-to-cement ratio, cement content, admixture dosage, and water-to-cement ratio are the most influential variables governing compressive strength. Additionally, local prediction behavior is notably affected by curing age, coarse aggregate content, and slump under specific conditions. Overall, the findings confirm that interpretable machine learning models, particularly Gaussian Process Regression and Extreme Gradient Boosting, provide reliable and practical tools for strength prediction, mixture optimization, and production-oriented quality control in real-world construction applications in Vietnam.