Machine learning prediction and explainability analysis of high strength glass powder concrete using SHAP PDP and ICE

Abstract Achieving high-strength concrete (HSC) with sustainable supplementary cementitious materials (SCMs) remains a significant challenge in the construction industry. Although glass powder has shown promise as a partial cement substitute, its specific impact on HSC growth is still unclear. This study aims to evaluate the compressive strength (CS) of high strength glass-powder concrete (HSGPC) using machine learning (ML) models and enhance predictive accuracy through hybrid optimization techniques. A dataset comprising 598 points was compiled, considering cement, glass powder, aggregates, water, superplasticizer, and curing days as key input parameters. Three standalone ML models—K-Nearest Neighbors (KNN), Random Forest (RF), and Extreme Gradient Boosting (XGB)—were trained, with RF achieving R² = 0.963 and XGB achieving R² = 0.946 on the test set. To further enhance performance, XGB was optimized using Particle Swarm Optimization (PSO), Firefly Algorithm (FA), and Grey Wolf Optimizer (GWO). Among these, XGB-GWO demonstrated the highest accuracy, with R² improving to 0.991 and MSE decreasing significantly from 83.95 to 14.42, resulting in an 82.82% error reduction. SHAP, PDP, and ICE analyses identified superplasticizer dosage, curing days, and coarse aggregate as the most influential parameters affecting compressive strength (CS). PDP and ICE validated these findings, showing reduced strength gains beyond 600 kg/m³ of cement and a decline beyond 800 kg/m³ of coarse aggregate. This study highlights the potential of ML-driven optimization for sustainable concrete design, offering an efficient, data-driven approach to optimizing material proportions for high-strength, eco-friendly concrete.