Ensemble Machine Learning-Based Techniques for Prediction of Mechanical Properties of Construction and Demolition Waste-Based Geopolymer Materials

Mahmoodi O., Siad H., Lachemi M., ŞAHMARAN M.

Canadian Society of Civil Engineering Annual Conference, CSCE 2024, Niagara Falls, Kanada, 5 - 07 Haziran 2024, cilt.717 LNCE, ss.237-249, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Cilt numarası: 717 LNCE
  • Doi Numarası: 10.1007/978-3-031-98775-5_21
  • Basıldığı Şehir: Niagara Falls
  • Basıldığı Ülke: Kanada
  • Sayfa Sayıları: ss.237-249
  • Anahtar Kelimeler: Compressive strength, Construction and demolition waste, Ensemble models, Geopolymer, Machine learning
  • Hacettepe Üniversitesi Adresli: Evet

Özet

This study employed machine learning (ML) algorithms including decision tree (DT)-based random forest (RF), extreme gradient boosting (XGB), and extra trees regressor (ET) as individual ML models and ensembled to predict the compressive strength of geopolymer mortars (GPMs) prepared from construction and demolition waste (CDW)-based precursors and aggregates. These models were developed using 13 different input features, including SiO2/Al2O3 (Si/Al), Na2O/SiO2 (Na/Si), curing temperature (CT), sodium silicate, sodium hydroxide and superplasticizer concentrations, aggregate type and water-to-binder (W/B) ratio. The ML models were constructed based on the experimental results gathered from the literature, where a total of 537 data points were utilized for modeling, of which 70% were allocated for training, and the remaining 30% for testing. The performance and generalization ability of the ML models was evaluated through ten-fold cross-validations. The coefficient of determination (R2) and statistical error metrics of MAE and MSE were employed to further assess the predictive accuracy of ML models. The tuning of ML models was performed through Scikit-Optimize. The results demonstrated robust performance of tuned ensemble models with an improved R2 of 0.942 and lower statistical errors. The SHAP analysis results identified Si/Al, CT, and W/B ratio as the most influential parameters on the compressive strength development of GPMs.