Scholarlink

<p>This study introduces a comprehensive data-driven framework for predicting the compressive strength (CS) of Ultra-High-Performance Concrete (UHPC) through the application of three hybrid ensemble machine learning models; Random Forest-Particle Swarm Optimization (RF-PSO), Adaptive -Boosting PSO (AB-PSO), and Gradient Boosting-PSO (GB-PSO). A substantial dataset comprising 700 UHPC mix designs was employed, incorporating eleven critical input variables, including Cement, Slag, Silica Fume, Fly Ash, Limestone Powder, Water, Aggregate, Fiber, Superplasticizer, and Age. The RF-PSO model demonstrated the highest predictive accuracy, with R<sup>2</sup> values of 0.9728 for training and 0.9584 for testing, alongside low error metrics: RMSE = 4.31&#xa0;MPa, MAE = 3.09&#xa0;MPa, and MAPE = 5.81%. In contrast, the GB-PSO model reported R<sup>2</sup> values of 0.9913 for training and 0.9804 for testing, with the lowest RMSE of 3.69&#xa0;MPa, while the AB-PSO model exhibited weaker generalization with R<sup>2</sup> values of 0.9064 for training and 0.9062 for testing. Taylor diagrams validated the RF-PSO model’s alignment with experimental data, evidenced by high correlation coefficients (&gt; 0.95) and optimal standard deviation overlap. SHAP analysis identified Age, Cement, Fiber, and Silica Fume as dominant predictors, with SHAP impacts ranging from − 25 to + 35. The findings provide an interpretable, highly accurate, and scalable UHPC strength prediction methodology. This study introduces a new hybrid modelling framework that integrates global hyperparameter optimization with SHAP based interpretability, offering a reliable and explainable tool for UHPC mix design that minimizes experimental efforts.</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

SHAP-enhanced hybrid PSO-ensemble framework models for interpretable prediction of UHPC compressive strength

  • 期刊论文

  • 发表日期：

    2025年9月25日

查看全文
  • Kamlesh Madurwar,
  • Ali Basem,
  • Anshul Nikhade,
  • Abdul Ateeque Azher,
  • Sandip Khedker,
  • Ahmed Adnan Hadi,
  • Mohammad Amir Khan,
  • Aseel Smerat

摘要

This study introduces a comprehensive data-driven framework for predicting the compressive strength (CS) of Ultra-High-Performance Concrete (UHPC) through the application of three hybrid ensemble machine learning models; Random Forest-Particle Swarm Optimization (RF-PSO), Adaptive -Boosting PSO (AB-PSO), and Gradient Boosting-PSO (GB-PSO). A substantial dataset comprising 700 UHPC mix designs was employed, incorporating eleven critical input variables, including Cement, Slag, Silica Fume, Fly Ash, Limestone Powder, Water, Aggregate, Fiber, Superplasticizer, and Age. The RF-PSO model demonstrated the highest predictive accuracy, with R2 values of 0.9728 for training and 0.9584 for testing, alongside low error metrics: RMSE = 4.31 MPa, MAE = 3.09 MPa, and MAPE = 5.81%. In contrast, the GB-PSO model reported R2 values of 0.9913 for training and 0.9804 for testing, with the lowest RMSE of 3.69 MPa, while the AB-PSO model exhibited weaker generalization with R2 values of 0.9064 for training and 0.9062 for testing. Taylor diagrams validated the RF-PSO model’s alignment with experimental data, evidenced by high correlation coefficients (> 0.95) and optimal standard deviation overlap. SHAP analysis identified Age, Cement, Fiber, and Silica Fume as dominant predictors, with SHAP impacts ranging from − 25 to + 35. The findings provide an interpretable, highly accurate, and scalable UHPC strength prediction methodology. This study introduces a new hybrid modelling framework that integrates global hyperparameter optimization with SHAP based interpretability, offering a reliable and explainable tool for UHPC mix design that minimizes experimental efforts.