A Hybrid Artificial Intelligence Framework Combining Gradient Boosting and Multi-Task Neural Networks for Simultaneous and Interpretable Prediction of Mechanical Properties in Supplementary Cementitious Material-Blended Concrete

Accurate simultaneous prediction of multiple concrete strengths is important for screening sustainable mixtures containing supplementary cementitious materials (SCMs). This study proposes a CatBoost-Guided Multi-Task Neural Network (CG-MTNN), in which CatBoost predicts compressive strength and a masked Multi-gate Mixture-of-Experts network jointly predicts splitting tensile and flexural strengths from incompletely labelled data. The framework was evaluated on 1,456 SCM-blended concrete mix designs from 27 countries. On the test set, CG-MTNN achieved R 2 values of 0.908, 0.964, and 0.963 for compressive strength (n = 146), splitting tensile strength (n = 34), and flexural strength (n = 28), respectively, and showed balanced performance across the three target properties. SHAP analysis was used to quantify model-specific feature attributions, while ensemble disagreement was examined as an indicator of prediction error. Because all records represent 28-day measurements, age-related features were interpreted only as fixed-age transformations and not as evidence of curing-age effects. The framework offers an interpretable decision-support approach for preliminary screening of 28-day SCM-blended concrete mixtures within the represented data domain. It is not intended to replace laboratory testing or code-based design verification.