Akademik Veri Yönetim Sistemi
MULTISCALE AND MULTIDISCIPLINARY MODELING EXPERIMENTS AND DESIGN, cilt.8, sa.8, 2025 (ESCI)
Predicting Ultra-High-Performance Concrete (UHPC) compressive strength using advanced machine learning (ML) techniques represents a paradigm shift in material science. This study investigates the application of nine ML models, including CatBoost, Extra Trees Regressor, and XGBoost, optimized through the Hyperopt framework to enhance predictive accuracy. A comprehensive dataset comprising 810 observations was employed, and critical performance metrics such as R2, RMSE, a20-index, Nash-Sutcliffe Efficiency (NSE), and sMAPE(Symmetric Mean Absolute Percentage Error) were analyzed to assess model efficacy. The CatBoost model, optimized using the Hyperopt framework, demonstrated superior predictive performance, achieving an R2 of 0.9742, RMSE of 6.20 MPa, a20-index of 1.0, NSE of 0.9742, and sMAPE of 3.6445 on the test set. Feature interpretability was further advanced using SHAP and Partial Dependence Plots (PDP), which identified Age, Fiber Content (Fi), and Cement-to-Water Ratio (C/W) as the most influential predictors. These tools elucidated complex nonlinear relationships and provided actionable insights for optimizing UHPC mix designs. This research highlights the transformative potential of ML in reducing experimental overhead, promoting sustainable construction practices, and enhancing material performance. While the study focused on steel-fiber-reinforced UHPC under controlled conditions, future research should explore diverse fiber types, environmental factors, and dynamic material behaviors. By advancing the integration of ML with experimental methodologies, this study paves the way for innovative solutions in high-performance concrete applications.