Akademik Veri Yönetim Sistemi
MATERIALS TODAY COMMUNICATIONS, cilt.49, sa.113937, ss.1-15, 2025 (SCI-Expanded, Scopus)
In this paper, the applicability of base and stacking ensemble models for the data-driven prediction of the
compressive strength (CS) of recycled brick aggregate concrete (RBAC) has been extensively evaluated. In this
sense, a raw database consisting of 374 observations compiled from the literature was cleaned of outliers using
the unsupervised learning algorithm Isolation Forest (IF), yielding a clean database of 324 observations with
seven input features for use in the modeling process. During the modeling phase, a total of fourteen different
prediction models were assessed, including four base ensemble machine learning (ML) models and ten stacking
ensemble models built using various combinations of these models. To improve the interpretability of the
model’s decision mechanism and examine the marginal effect of each input feature on the prediction, Shapley
Additive Explanations (SHAP)-based feature importance analysis and Individual Conditional Expectation (ICE)
analysis were implemented. Lastly, to support the practical use of the developed models, a user-friendly graphical
user interface (GUI) was designed, enabling engineers and field practitioners to make fast and reliable durability
predictions for RBAC mixtures. Considering the comprehensive evaluations, all base learners developed in the
test phase achieved an average R² of 0.825 and RMSE of 3.22. In contrast, the stacking ensemble models
improved these metrics, reaching an average R² of 0.851 and RMSE of 2.98, thereby confirming the effectiveness
of ensemble learning in enhancing prediction accuracy. Notably, the SM-2 model, derived from the binary
combination of CB and GBM, demonstrated the best performance among all base and staking ensemble models
with R = 0.927, R2= 0.857, RMSE = 2.91, and U95 = 4.13 during the test phase. Meanwhile, in general, it is
remarkable that stacking models based on ternary combinations showed more stable and accurate predictions
compared to those based on binary combinations. The SHAP and ICE-based explainability analyses revealed the
marginal effects of each input feature on the predicted CS, providing valuable engineering insights for mixture
optimization strategies. All in all, this study presents a robust and interpretable AI-based framework for the CS
prediction of sustainable and eco-friendly RBAC, contributing significantly to the field of data-driven green
concrete design.