Machine Learning-Driven Comparative Prediction of Compressive Strength in True Geopolymer and Hybrid Alkali-Activated Cement Systems
- 1Qatar University , , P.O. Box , ,
Abstract
Ordinary Portland cement (OPC) production accounts for approximately 7–8% of global CO2 emissions, motivating the development of sustainable binders such as metakaolin-based geopolymers and hybrid alkali-activated cements (AACs). However, accurate prediction of compressive strength in these systems remains challenging because of nonlinear interactions among precursor composition, activator chemistry, water-related parameters, and curing conditions. This study develops a reliability-oriented machine-learning framework using literature-derived datasets comprising 223 metakaolin-based geopolymer mixtures and 262 hybrid AAC mixtures. Five tree-based ensemble algorithms were evaluated using five-fold GroupKFold validation, with literature source used as the grouping variable to reduce the risk of study-level information leakage. The framework further incorporated fold-wise hyperparameter optimization, inverse-frequency sample weighting during model fitting, source-wise sensitivity analysis, and leave-one-source-out evaluation. Under grouped internal validation, CatBoost achieved the best performance for the metakaolin-based geopolymer dataset (R2 = 0.8211 ± 0.1275, RMSE = 5.773 ± 2.333 MPa, MAE = 4.131 ± 1.543 MPa), whereas Gradient Boosting performed best for the hybrid AAC dataset (R2 = 0.8876 ± 0.0708, RMSE = 3.333 ± 1.181 MPa, MAE = 2.366 ± 0.737 MPa). TreeSHAP analysis revealed distinct model-derived feature associations: hybrid AAC predictions were most strongly associated with GGBFS content, followed by activator- and curing-related variables, whereas metakaolin-based geopolymer predictions were most strongly associated with the H2O/Na2O ratio, aggregate proportion, activator chemistry, and curing conditions. These SHAP-derived patterns represent statistical associations learned by the models rather than direct evidence of physical causation or reaction mechanisms. Overall, the proposed framework provides a transparent approach for preliminary compressive-strength screening and experimental prioritization of sustainable alkali-activated binders.
Related
- Prediction model for compressive strength of alkali-activated multi-source solid-waste grouting materials based on generative adversarial networks and machine learning
- Integrated Silica-Encapsulated Lauric Acid Phase Change Material into Acid-Activated Geopolymer for Thermal Energy Storage Applications
- Curing regime controlled performance of basalt-modified alkali activated slag: phase composition, porosity profile, microstructure and compressive strength
- Humidity-dependent formaldehyde adsorption on porous frameworks: A comparative study of MOFs, zeolites, and activated carbon
- Understanding the Relationship Between Synthetic Conditions and H ₂ Uptake Capacity in Oxidatively Activated Turbostratic Carbons through Machine Learning
BibTeX
@article{keloth2026machine,
title = {{Machine Learning-Driven Comparative Prediction of Compressive Strength in True Geopolymer and Hybrid Alkali-Activated Cement Systems}},
author = {Jafla Thekke Keloth and Nasser Al Nuaimi and Mohammed Z. E. B. Elshafie and Ibnelwaleed A. Hussein},
journal = {ACS Omega},
year = {2026},
month = nov,
doi = {10.1021/acsomega.6c07604},
url = {https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c07604/5463999/Machine-Learning-Driven-Comparative-Prediction-of},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Machine Learning-Driven Comparative Prediction of Compressive Strength in True Geopolymer and Hybrid Alkali-Activated Cement Systems
AU - Jafla Thekke Keloth
AU - Nasser Al Nuaimi
AU - Mohammed Z. E. B. Elshafie
AU - Ibnelwaleed A. Hussein
PY - 2026
DA - 2026/11/04
JO - ACS Omega
DO - 10.1021/acsomega.6c07604
UR - https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c07604/5463999/Machine-Learning-Driven-Comparative-Prediction-of
AB - Ordinary Portland cement (OPC) production accounts for approximately 7–8% of global CO2 emissions, motivating the development of sustainable binders such as metakaolin-based geopolymers and hybrid alkali-activated cements (AACs). However, accurate prediction of compressive strength in these systems remains challenging because of nonlinear interactions among precursor composition, activator chemistry, water-related parameters, and curing conditions. This study develops a reliability-oriented machine-learning framework using literature-derived datasets comprising 223 metakaolin-based geopolymer mixtures and 262 hybrid AAC mixtures. Five tree-based ensemble algorithms were evaluated using five-fold GroupKFold validation, with literature source used as the grouping variable to reduce the risk of study-level information leakage. The framework further incorporated fold-wise hyperparameter optimization, inverse-frequency sample weighting during model fitting, source-wise sensitivity analysis, and leave-one-source-out evaluation. Under grouped internal validation, CatBoost achieved the best performance for the metakaolin-based geopolymer dataset (R2 = 0.8211 ± 0.1275, RMSE = 5.773 ± 2.333 MPa, MAE = 4.131 ± 1.543 MPa), whereas Gradient Boosting performed best for the hybrid AAC dataset (R2 = 0.8876 ± 0.0708, RMSE = 3.333 ± 1.181 MPa, MAE = 2.366 ± 0.737 MPa). TreeSHAP analysis revealed distinct model-derived feature associations: hybrid AAC predictions were most strongly associated with GGBFS content, followed by activator- and curing-related variables, whereas metakaolin-based geopolymer predictions were most strongly associated with the H2O/Na2O ratio, aggregate proportion, activator chemistry, and curing conditions. These SHAP-derived patterns represent statistical associations learned by the models rather than direct evidence of physical causation or reaction mechanisms. Overall, the proposed framework provides a transparent approach for preliminary compressive-strength screening and experimental prioritization of sustainable alkali-activated binders.
SN - 2470-1343
ER -