Data-driven prediction of pure mode I fracture toughness in cementitious geomaterials using hybrid ANN–metaheuristic models

Accurate prediction of fracture toughness in cementitious geomaterials is critical for ensuring the structural integrity and durability of masonry constructions, yet traditional experimental determination remains costly, time-intensive, and susceptible to variability arising from specimen geometry, material heterogeneity, and testing conditions. This study addresses these challenges by developing five hybrid artificial neural network (ANN) models optimized with metaheuristic algorithms to predict pure mode I fracture toughness of cementitious geomaterials. The dataset integrated three input categories: mechanical parameters, physical properties, and mix design variables. Five hybrid models were developed and rigorously evaluated: ANN-Whale Optimization Algorithm (WOA), ANN-Harris Hawks Optimization (HHO), ANN-Reptile Search Optimization Algorithm (RSO), ANN-Arithmetic Optimization Algorithm (AOA), and ANN-Mayfly Optimization Algorithm (MOA). The ANN-MOA model demonstrated superior performance, achieving the highest R 2 values of 0.9536 and 0.9157 during training and testing phases, respectively, with the lowest RMSE (0.0166 and 0.0215) and most balanced MBE (0.0029 and 0.0033), indicating minimal systematic bias. Residual analysis confirmed ANN-MOA's compact error distribution with the narrowest interquartile range and fewest extreme outliers, while Taylor diagrams positioned it closest to the reference point among all models. The ANN-RSO exhibited exceptional generalization capability with a testing R 2 of 0.9375, slightly outperforming other models on unseen data. Sensitivity analysis revealed that the mechanical properties were the most influential predictors of fracture toughness. This research provides a robust, non-destructive, and cost-effective data-driven framework for estimating fracture toughness of cementitious geomaterials, with direct implications for mixture design optimization, quality control in precast concrete production, and structural health monitoring applications. The proposed methodology significantly reduces reliance on extensive experimental testing while maintaining high prediction accuracy, advancing the state-of-the-art in computational modeling of quasi-brittle fracture behavior.