Optimized machine learning models for predictive performance assessment of fly ash-integrated clay bricks

The growing need for sustainable construction materials has driven interest in integrating coal-derived fly ash into clay bricks, offering a dual solution to industrial waste management and environmental impact reduction. This study investigates the predictive modeling of fly ash-integrated clay bricks using four machine learning models: Extreme Gradient Boosting (XGBoost) with Grey Wolf Optimization(GWO) (XGBoost-GWO), XGBoost Baseline, Particle Swarm Optimized Random Forest (PSO-RF), and Ant Colony Optimized Decision Tree (ACO-DT). A comprehensive dataset comprising 18 input features and 3 output targets, open porosity (OPm), water absorption (WAm), and compressive strength (CSm), was compiled from 26 independent studies, with fly ash content ranging from 1 to 90%, firing temperatures between 750 and 1300 °C, and compressive strength values from 0.64 to 85.1 MPa. Unlike previous single-model studies, this work provides a systematic comparison of metaheuristically optimized versus baseline tree-based models on a heterogeneous multi-study dataset, while critically evaluating generalization via leave-one-study-out validation, an approach rarely applied in construction materials informatics. Among the models, the XGBoost Baseline consistently achieved the highest predictive accuracy, with Mean Absolute Error (MAE) values of 2.94 (WAm), 1.403 (OPm), and 4.150 (CSm), and Pearson correlation coefficients (PCC) above over 0.91 across targets. The hybrid XGBoost‑GWO, despite incorporating Grey Wolf Optimizer, did not surpass the baseline (e.g., MAE for CSm = 4.29), indicating that default XGBoost parameters are near-optimal for this dataset. Feature importance analysis identified firing temperature and SiO₂ content as the most influential parameters. Sensitivity analyses revealed a sigmoidal compressive strength response beyond 1000 °C and optimal fly ash content near 30–40% for balanced strength and porosity. Predictive accuracy degrades substantially under leave-one-study-out validation (e.g., CSm MAE increases from 4.15 to 9.82 MPa), and all models systematically underpredict high-porosity samples (WAm > 22.5%) with a bias of approximately − 1.7. The findings offer actionable insights for optimizing brick formulation and firing conditions, bridging the gap between laboratory research and industrial deployment.