Explainable stacked ensemble learning and multi-objective optimization for the intelligent design of green and low-carbon geopolymer concrete

Katlav M., Türk K.

EXPERT SYSTEMS WITH APPLICATIONS, cilt.332, ss.1-23, 2027 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 332
  • Basım Tarihi: 2027
  • Doi Numarası: 10.1016/j.eswa.2026.133523
  • Dergi Adı: EXPERT SYSTEMS WITH APPLICATIONS
  • Derginin Tarandığı İndeksler: Applied Science & Technology Source, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Scopus, Technology Collection (ProQuest), Aerospace Database, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC, Public Affairs Index
  • Sayfa Sayıları: ss.1-23
  • İnönü Üniversitesi Adresli: Evet

Özet

This paper presents an explainable data-driven framework for reliable prediction and multi-objective optimization

design of compressive strength (CS) of green and low-carbon geopolymer concrete (GPC). A comprehensive

literature-based database containing 2281 instances with 16 defined input features was employed to

develop both individual and stacked ensemble learning models. All stacked ensemble models demonstrated

strong predictive capability; meanwhile, stacked learning consistently improved generalization performance. The

best stacked ensemble model, SCM-17 (ETR + GBM + RF), achieved a testing accuracy of R2 = 0.967, RMSE =

5.83 MPa, MAE = 3.76 MPa, and MAPE = 9.87%, confirming the robustness of the proposed stacked strategy for

CS prediction. Further, to enhance interpretability and extract engineering insight, SHAP-based sensitivity

analysis was conducted. The SHAP analysis indicates that precursor-related variables, particularly ground

granulated blast furnace slag and fly ash contents, together with curing age and sodium silicate dosage, exhibit

the strongest influence on the model-predicted CS. Beyond prediction, a target-strength-oriented multi-objective

optimization framework was implemented to identify optimal GPC mixture proportions while simultaneously

minimizing carbon emissions, cost, and energy consumption. The optimized solutions demonstrated balanced

sustainability–performance trade-offs and confirmed that feasible high-strength mixtures can be obtained

without extreme material configurations. Furthermore, selected optimized mixtures were experimentally validated,

and the measured CSs showed good agreement with the corresponding target and predicted values,

thereby providing independent verification of the proposed inverse-design framework. Lastly, a graphical user

interface was developed to enable rapid CS prediction together with sustainability assessment within validated

parameter limits. All in all, the proposed integrated prediction–interpretation–optimization workflow provides a

reliable decision-support framework for data-driven and sustainability-oriented GPC design.