A Novel Stacking Ensemble Framework for Predicting Workability of Cement-Superplasticizer Systems With SHAP and LIME Interpretability
IEEE Access, cilt.14, ss.117891-117915, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3719449
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.117891-117915
- Anahtar Kelimeler: Cement paste, explainable artificial intelligence, LIME, machine learning, polycarboxylate ether, SHAP, stacking ensemble, workability
- Bursa Uludağ Üniversitesi Adresli: Evet
Özet
The fresh-state behavior of cementitious systems is governed by complex interactions between cement mineralogy and the molecular architecture of polycarboxylate ether-based superplasticizers (PCEs). Predicting workability indicators such as mini-slump and flow time across varying cement C3A contents and PCE types remains challenging, and existing machine learning approaches rarely combine structurally diverse learning principles with explainable decision mechanisms. This study introduces XRES-GB, a novel stacking ensemble classifier that combines XGBoost, Random Forest, Extra Trees, and Support Vector Machine as base learners, with Gradient Boosting serving as the meta-model. The framework was applied to an experimental dataset comprising 171 observations spanning 22 input variables, including cement mineralogical composition, physical characteristics, and PCE molecular parameters. Mini-slump and flow time values were converted into binary workability classes using threshold values selected according to practical workability limits, and the resulting class imbalance was addressed via a model-specific strategy combining Random Oversampling and cost-sensitive class weighting. Model performance was evaluated through 5-fold stratified group cross-validation and benchmarked against eleven conventional classifiers. XRES-GB achieved mean test accuracies of 0.91 and 0.92 for mini-slump and flow time classification, respectively, outperforming all competing models in both accuracy and fold-to-fold consistency. SHAP analyses identified Water Reducer Admixture dosage, cement content, and density as the most influential explanatory variables, although their relative importance varied between the mini-slump and flow time classification tasks. LIME explanations further confirmed that individual predictions align with established cement-admixture interaction mechanisms. The proposed framework delivers high-fidelity predictions while maintaining interpretability, offering a data-driven tool for optimizing cement-PCE compatibility in mixture design.