White-box modeling and reliability assessment of roller-compacted concrete for dam and pavement engineering applications

Roller-compacted concrete (RCC) is widely used in dam and pavement engineering because of its rapid construction, cost-effectiveness, and satisfactory mechanical performance. Reliable prediction of RCC compressive strength is essential for optimizing mixture design and ensuring structural safety. However, most existing predictive models rely on black-box techniques with limited engineering interpretability, while uncertainty and reliability aspects are often overlooked. To address these limitations, this study proposes an integrated white-box and reliability-based framework for predicting the compressive strength of RCC. A database containing 430 experimental records collected from published studies was used to develop and compare three interpretable models: Group Method of Data Handling (GMDH), Gene Expression Programming (GEP), and Response Surface Methodology (RSM). Model performance was evaluated using several statistical indices and five-fold cross-validation. In addition, One-at-a-Time (OAT), Shapley Additive Explanations (SHAP), and Accumulated Local Effects (ALE) analyses were employed to quantify the importance of input variables and their interactions. Monte Carlo simulation was further performed to assess the influence of input uncertainty on the probability of achieving target strength levels. Among the white-box approaches, the RSM model exhibited the best predictive performance, achieving RMSE and MAE values of 3.966 and 3.102 MPa, respectively, whereas the RMSE values of GMDH and GEP were 7.591 and 9.806 MPa. Five-fold cross-validation confirmed the robustness of the RSM model, with a mean R² of 0.839 and an average RMSE of 6.859 MPa under identical validation conditions. Sensitivity analysis identified cement and water contents as the most influential parameters, while cement-water and cement-age interactions showed the strongest combined effects. Reliability analysis revealed that increasing input variability significantly reduced the probability of meeting the target strength, particularly in pavement applications. Compared with conventional black-box approaches, the proposed framework provides explicit analytical relationships together with reliability-based decision support, offering a practical and interpretable tool for risk-informed RCC mix design optimization.