Scholarlink

<p>Steel fiber reinforced concrete (SFRC) exhibits complex mechanical behavior because its strength depends on matrix composition, fiber geometry, and fiber–matrix interaction, making accurate and interpretable prediction challenging. This study develops an interpretable and uncertainty-aware hybrid machine learning (ML) framework to predict SFRC mechanical properties. Six algorithms, including Adaptive Boosting, Extreme Gradient Boosting, Extremely Randomized Trees, Random Forest, CatBoost, and Neural Networks, were benchmarked with Tree-structured Parzen Estimator (TPE) optimization to improve generalization. Among the evaluated models, CatBoost–TPE achieved the best testing performance, with R² values of 0.901, 0.803, and 0.851 for compressive, tensile, and flexural strengths, respectively, while reducing RMSE and MAE by 8–29%. To move beyond deterministic prediction, Sobol-based stochastic global sensitivity analysis (SGSA) was integrated with SHAP feature attribution to evaluate parameter influence and uncertainty propagation. Results show that binder–water variability mainly governs compressive behavior, whereas stochastic perturbations in the fiber–matrix subsystem have stronger effects on tensile and flexural responses. The optimized framework is further deployed as a web-based decision-support tool to support preliminary SFRC strength prediction and data-driven mix-design exploration within the applicability domain of the trained model. This approach may support preliminary quality control and reduce trial-and-error laboratory testing within the studied data range.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Stochastic and interpretable hybrid machine learning framework for predicting mechanical properties of steel fiber reinforced concrete

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Tran H. T Le,
  • Anh-Thang Le,
  • Dongkyu Lee

摘要

Steel fiber reinforced concrete (SFRC) exhibits complex mechanical behavior because its strength depends on matrix composition, fiber geometry, and fiber–matrix interaction, making accurate and interpretable prediction challenging. This study develops an interpretable and uncertainty-aware hybrid machine learning (ML) framework to predict SFRC mechanical properties. Six algorithms, including Adaptive Boosting, Extreme Gradient Boosting, Extremely Randomized Trees, Random Forest, CatBoost, and Neural Networks, were benchmarked with Tree-structured Parzen Estimator (TPE) optimization to improve generalization. Among the evaluated models, CatBoost–TPE achieved the best testing performance, with R² values of 0.901, 0.803, and 0.851 for compressive, tensile, and flexural strengths, respectively, while reducing RMSE and MAE by 8–29%. To move beyond deterministic prediction, Sobol-based stochastic global sensitivity analysis (SGSA) was integrated with SHAP feature attribution to evaluate parameter influence and uncertainty propagation. Results show that binder–water variability mainly governs compressive behavior, whereas stochastic perturbations in the fiber–matrix subsystem have stronger effects on tensile and flexural responses. The optimized framework is further deployed as a web-based decision-support tool to support preliminary SFRC strength prediction and data-driven mix-design exploration within the applicability domain of the trained model. This approach may support preliminary quality control and reduce trial-and-error laboratory testing within the studied data range.