Machine learning-driven optimization of alkali-activated sustainable pavement concrete incorporating agro-industrial wastes

Abstract The integration of agro-industrial wastes into alkali-activated concrete (AAC) offers a sustainable alternative to traditional concrete pavement material, yet the synergistic effects of waste foundry sand (WFS) as a fine aggregate replacement and rice husk ash (RHA) as a partial binder replacement in ground granulated blast furnace slag (GGBS)-based AAC designed for rigid pavement applications remain under-explored. This study investigates the machine learning (ML)-driven optimization of AAC, addressing a critical research gap in material synergy and predictive modeling. Experimental results indicate that increasing WFS (0–30%) and RHA (0–20%) reduces compressive strength (CS) to some extent; however, an optimized mix containing 20% WFS and 15% RHA achieved a 28-day CS satisfying pavement-quality concrete requirements while maintaining workability within acceptable limits. A comparative ML framework comprising six algorithms; Multiple Linear Regression, Decision Tree, Random Forest, AdaBoost, Support Vector Regression, and Gradient Boosting was developed and benchmarked to predict the CS of 330 experimental samples. The Random Forest model achieved the highest predictive accuracy (R 2 = 0.9176, RMSE = 2.99 MPa, MAE = 2.62 MPa), with performance statistically comparable to Gradient Boosting and AdaBoost. Feature importance and interpretability analysis revealed that GGBS and fine aggregate content significantly influence CS, while higher WFS and RHA levels adversely affect strength. This research provides a scalable framework for designing low-carbon pavement materials by bridging the gap between experimental mix design and advanced computational optimization that substantially reduce dependence on resource-intensive experimental trials for AAC incorporating agro-industrial wastes