Predicting fracture energy and durability parameters of concrete through hybrid machine learning models

Abstract The incorporation of fracture mechanics into structural engineering has underscored the importance of fracture energy in understanding crack propagation in quasi-brittle concrete. Nevertheless, conventional laboratory testing remains labor-intensive, time-consuming, and costly. This study applies hybrid machine learning (ML) algorithms, including ANN-GA, SVR-GA, XGB-GA, GBR-GA, and a Stacking Ensemble, to efficiently predict key concrete properties such as initial fracture energy (IFEC), modulus of elasticity (ME), tensile strength (TS), water absorption (WA), and abrasion resistance (AR). Among individual models, IFEC-XGB-GA and IFEC-GBR-GA demonstrated the most stable performance, while the Stacking Ensemble improved overall prediction accuracy and robustness by approximately 5–10 compared with single models. For ME prediction, the ensemble approach reduced bias and variance, although GBR performed better under extreme conditions. TS-ANN-GA exhibited overfitting, whereas TS-Stacking Ensemble achieved the lowest testing errors (MAE0.145, RMSE0.232) and the highest explained variance (≈1.014) and correlation (0.973), confirming the advantages of model aggregation. Moreover, GBR-GA effectively predicted water absorption (R 2 0.998) and classified abrasion resistance with about 84 accuracy, highlighting strong potential for durability evaluation. A graphical user interface (GUI) was developed to integrate these models, providing engineers with a practical, data-driven, and physics-informed tool for accurately predicting the fracture energy of concrete.