Application of ensemble learning, deep neural networks, and explainable AI in predicting compressive and split tensile strength of steel fiber-reinforced recycled aggregate concrete

The depletion of natural aggregates and the increasing accumulation of construction waste have intensified interest in recycled aggregate concrete (RAC) as a sustainable alternative. However, RAC suffers from weaker interfacial transition zones and lower mechanical strength due to the residual mortar on recycled coarse aggregates (RCA), which weakens bonding and increases porosity. To address these challenges, steel fibres are incorporated to enhance strength, ductility, and crack resistance. However, the random dispersion of steel fibres introduces variability in mechanical properties, making conventional strength prediction models less reliable. This study employs machine learning (ML) techniques to predict the compressive strength (f cu ) and split tensile strength (f sp ) of Steel Fiber-Reinforced Recycled Aggregate Concrete (SFR-RAC). The primary objective of this study is to develop robust ML models capable of accurately predicting the mechanical properties of SFR-RAC, addressing the limitations of conventional empirical approaches. This research also aims to bridge the gap in current literature by providing a comparative evaluation of multiple ML techniques, supported by interpretability tools such as SHapley Additive exPlanations and Partial Dependence Plots. A dataset of 324 samples (80% training, 20% testing) was analysed using four ML models: Random Forest (RF), Gradient Boosted Regression Trees (GBRT), Adaptive Boosting Regressor, and Back Propagation Artificial Neural Network. Model performance was assessed using scatter plots, Taylor diagrams, k-fold cross-validation and a 20% margin of error. RF was the best predictor for f cu (training: R 2 = 0.970 and MAPE = 4.3% and Testing: R 2 = 0.915 and MAPE = 6.1%). GBRT performed best for f sp (training: R 2 = 0.983 and MAPE = 4.9% while the testing: R 2 = 0.877 and MAPE = 10.8%). Feature importance analysis using SHAP identified the water-to-binder ratio (W/B%) as the most influential factor negatively affecting strength. Partial Dependence Plots were employed to further examine the relationships between key input features and strength outputs. The findings confirm that ML models provide a reliable approach for predicting the mechanical properties of SFR-RAC, effectively addressing its inherent heterogeneity. Future research should explore hybrid ML models that combine deep learning with optimization methods like genetic algorithms. Application of data augmentation techniques such as bootstrapping and noise injection may improve accuracy and make predictions more reliable for practical use.