Machine learning is increasingly used to predict the compressive strength of self-compacting concrete with recycled coarse aggregate (SCRCAC), with coefficients of determination of 0.81–0.87 reported in the literature. This paper first shows that the widely used reference dataset of 603 mixtures contains only 504 unique compositions, with 84 groups of identical and 21 contradictory ones, so that identical mixtures leak between the training and test sets. Under an objective, leakage-free evaluation (with the model re-tuned on the deduplicated dataset), the coefficient of determination drops to about 0.73, a correction that applies to all models on this dataset. We then propose an interpretable, hyperparameter-optimized LightGBM model that (i) reaches the level of the best published results under the standard protocol (seed-averaged five-fold cross-validation (CV) R2 = 0.813); (ii) provides a calibrated uncertainty interval for each prediction via split-conformal prediction, achieving an empirical coverage of 0.906 at the 90% nominal level; and (iii) remains fully explainable (SHAP (SHapley Additive exPlanations), partial dependence), with cement as the dominant predictor, followed by water and the mineral admixture. Under a 70/30 protocol averaged over 25 splits, it achieves an R2 = 0.794 ± 0.038 and a root mean squared error (RMSE) = 6.20 ± 0.48 MPa, exceeding all four machine learning models of the reference study. Aspects in which the reference study retains an advantage are also discussed.
Bojana Milošević, Nenad Kojić, Milan Kragović· Journal of Composites Scienc...· 0 citations
Reliable prediction of the compressive strength of self-compacting concrete with recycled coarse aggregate (SCRCAC) from mixture composition supports more rational mix design and fewer experimental tests. Using the benchmark dataset of the reference study (603 mixtures, eight input variables), this work re-examines machine-learning prediction of this property with an emphasis on honest evaluation rather than on a new model. A stacking ensemble of three gradient-boosting models (XGBoost, LightGBM, CatBoost) and an extremely randomized trees model, combined through a ridge meta-learner, is used as a representative model and compared with the four machine-learning models of the reference study (Random Forest, Extra Trees, XGBoost, LightGBM), the recent single-booster model of Abood et al., and the reference artificial neural network. Reported as the mean over 25 repeated 70/30 splits, the ensemble reaches R2 = 0.793 ± 0.038 and RMSE = 6.21 ± 0.48 MPa, above all four reference models (R2 = 0.7249–0.7635) and significantly, though only marginally, above a tuned single XGBoost. The central contribution is the evaluation itself. Because the dataset contains repeated identical compositions, a leakage-free protocol lowers the R2 of every model to between 0.60 and 0.71, showing that the values of about 0.81–0.87 usually reported are inflated by duplicate-composition leakage, and leave-one-source-out evaluation lowers it further to about 0.14. Mutual-information and partial-dependence analyses identify cement as the dominant predictor, with water acting mainly through a nonlinear dependence. Robust, leakage-aware evaluation, rather than model architecture, emerges as the key to credible strength prediction on this benchmark.
This study investigates the application of artificial neural networks (ANNs) for estimating the compressive strength of clay masonry walls based on the mechanical and geometrical properties of their constituent materials. A multilayer perceptron (MLP) neural network was developed using a hybrid dataset derived from Eurocode 6 empirical formulations and representative commercially available masonry units and mortars, enabling systematic generation of realistic input–output relationships. Input parameters included masonry unit dimensions and compressive strength, mortar compressive strength, masonry unit classification group, and mortar type. Different ANN topologies with ReLU, tanh, and logistic activation functions were analyzed, while training was performed using the Adam optimization algorithm. Model performance was evaluated using MSE, MAE, RMSE, R2, and 5-fold cross-validation. The proposed ANN model achieved high prediction accuracy, with R2 values approaching 0.98 for the optimal configuration. Sensitivity, SHAP, and partial dependence analyses confirmed that the constituent material strengths, together with the masonry unit classification and mortar type, are the most influential inputs, in agreement with the Eurocode 6 formulation. The developed model provides a practical tool for preliminary engineering assessment and rapid comparative analysis of masonry wall configurations, reducing reliance on repetitive empirical calculations. The model is based on a Eurocode 6 synthetic dataset and is intended for predictive approximation and engineering support rather than replacement of experimental testing. In this study, the ANN is explicitly framed as a surrogate model of the Eurocode 6 formulation rather than as a replacement for it: its added value lies in providing a fast, continuously differentiable, and interpretable approximation that enables large-scale parameter exploration, interpretability analysis, and deployment as a real-time decision-support web service. To confirm its robustness, the surrogate was benchmarked against Linear Regression, Random Forest, and XGBoost models on the identical dataset.