GMDH-based comparative modeling and reliability assessment provide interpretable prediction of compressive strength in natural-zeolite concrete

Concrete production places increasing pressure on natural resources, motivating the use of natural zeolite in cementitious systems. This study developed an interpretable data-driven framework for predicting the compressive strength of zeolite-containing concrete using 324 laboratory specimens. The input variables were water-to-binder ratio (W/B), zeolite content (Z), superplasticizer dosage (SP), total aggregate content (TA), fine aggregate ratio (FA), and curing time (T). Three interpretable models, Group Method of Data Handling (GMDH), Response Surface Methodology (RSM), and Gene Expression Programming (GEP), were compared with Random Forest, Gradient-Boosted Regression Trees, and a Deep Neural Network. All models were evaluated using 10-fold cross-validation. Among the interpretable models, GMDH achieved the best performance, with R² = 0.8957, RMSE = 5.113 MPa, and MAE = 4.058 MPa. RSM ranked second (R² = 0.8828; RMSE = 5.422 MPa), while GEP showed the lowest accuracy (R² = 0.7060; RMSE = 8.585 MPa). Although some benchmark models achieved higher accuracy, GMDH provided an explicit formulation suitable for sensitivity and reliability analyses. SHAP analysis showed that curing time had the greatest influence (0.54), followed by W/B (0.32), while local sensitivity analysis identified SP as the strongest control variable around the reference mixture. At the maximum uncertainty level, the 65 MPa target had a failure probability of 0.0161 and a reliability index of 2.142. The proposed framework supports preliminary mixture design and risk-informed screening within the investigated data domain.