Development of machine learning models for intelligent prediction of workability, mechanical, and durability characteristics of cement mortar made of recycled sand
Construction and Building Materials, vol.530, 2026 (SCI-Expanded, Scopus)
- Publication Type: Article / Article
- Volume: 530
- Publication Date: 2026
- Doi Number: 10.1016/j.conbuildmat.2026.146599
- Journal Name: Construction and Building Materials
- Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC
- Keywords: Cement mortar, Concrete waste, Construction and demolition waste, Excavation waste, Machine learning, Recycled sand
- Middle East Technical University Affiliated: Yes
Abstract
The construction sector’s heavy reliance on natural sand motivates the integration of recycled fine aggregates-derived from excavation waste (EW), construction and demolition waste (CDW), and concrete waste (CW)-into sustainable mortars. This study develops a machine-learning framework to predict the fresh, mechanical, and durability properties of mortars incorporating these recycled sands. Using two experimental datasets (n = 114 and n = 48), six mix-design inputs—including water-to-cement (W/C) and sand-to-cement (S/C) ratios-model flow, compressive strength (CS), density, flexural strength (FS), water absorption (WA), and ultrasonic pulse velocity (UPV). Extreme Gradient Boosting (XGBoost), Random Forest (RF), Support Vector Regression (SVR), Kernel Ridge Regression (KRR), and a Multilayer Perceptron (MLP) are benchmarked using a rigorous 80/20 train-test split and cross-validation. The results demonstrate that XGBoost provides superior accuracy for the fresh and primary mechanical properties, achieving minimized testing sMAPE values of 0.0009 for Flow, 0.0265 for CS, and 0.0051 for Density. While SVR emerges as the optimal model for FS with a testing sMAPE of 0.0264, RF provides the most stable generalization for the durability-related parameters of WA and UPV (testing sMAPE = 0.0500 and 0.0092, respectively). These findings highlight the capacity of ensemble and kernel-based learners to model the acoustic attenuation and capillary suction caused by aggregate heterogeneity. SHAP analyses confirm that binder ratios govern workability and strength, while aggregate porosity dominates durability and microstructural responses. Ultimately, this study supports a hybrid, property-specific modeling framework that leverages boosting, kernel-mapping, and ensemble bagging to enhance the predictive reliability and sustainable design of recycled-sand mortars.