Scholarlink

<p>This study investigates the effects of incorporating fly ash (FA) and silica fume (SF) into concrete and evaluates the predictive accuracy of machine learning models such as Backpropagation Neural Network (BPNN), Random Forest Regressor (RFR), and Gradient Boosting Regressor (GBR), on compressive strength. Optimal performance was achieved with 50–60% FA and 8–10% SF, reaching strengths above 76&#xa0;MPa at 90 days, while 100% FA with 10% SF reached 71.13&#xa0;MPa at 90 days versus 27.6&#xa0;MPa at 14 days. Among all models, GBR showed the best accuracy (R² = 0.996, MSE = 0.578, MAPE = 0.941%), with SHAP and Partial Dependence analyses confirming curing time as the most influential factor, followed by %SF and %FA. Perturbation analysis confirmed GBR’s robustness to input variation, and monotonicity analysis revealed a strong positive trend between curing time and strength (Spearman correlation = 0.9245), confirming GBR’s suitability for strength prediction and mix optimization.</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

Explainable machine learning models for predicting compressive strength of high-volume fly ash concrete

  • 期刊论文

  • 发表日期：

    2025年7月21日

查看全文
  • Anish Kumar,
  • Sameer Sen,
  • Manish Pratap Singh,
  • Sanjeev Sinha,
  • Bimal Kumar

摘要

This study investigates the effects of incorporating fly ash (FA) and silica fume (SF) into concrete and evaluates the predictive accuracy of machine learning models such as Backpropagation Neural Network (BPNN), Random Forest Regressor (RFR), and Gradient Boosting Regressor (GBR), on compressive strength. Optimal performance was achieved with 50–60% FA and 8–10% SF, reaching strengths above 76 MPa at 90 days, while 100% FA with 10% SF reached 71.13 MPa at 90 days versus 27.6 MPa at 14 days. Among all models, GBR showed the best accuracy (R² = 0.996, MSE = 0.578, MAPE = 0.941%), with SHAP and Partial Dependence analyses confirming curing time as the most influential factor, followed by %SF and %FA. Perturbation analysis confirmed GBR’s robustness to input variation, and monotonicity analysis revealed a strong positive trend between curing time and strength (Spearman correlation = 0.9245), confirming GBR’s suitability for strength prediction and mix optimization.