Scholarlink

<p>This research introduces a robust machine learning framework for estimating the compressive strength of concrete, utilizing a Light Gradient Boosting Machine (LightGBM) regression algorithm. The model was developed using a diverse dataset that included different mix proportions of fly ash, silica fume, cement, fine and coarse aggregates, along with varying curing durations. After a thorough hyperparameter optimization process, the final model incorporated a learning rate of 0.1, 200 boosting iterations, an unrestricted tree depth, and 31 maximum leaf nodes. The model demonstrated strong predictive accuracy, achieving an R² value of 0.99 on the training set and 0.97 on the testing set, with corresponding Mean Absolute Errors (MAE) of 0.70&#xa0;MPa and 1.35&#xa0;MPa. Feature importance derived from SHAP values highlighted curing duration, silica fume percentage, and cement content as primary contributors to strength outcomes. Additional interpretation through partial dependence plots and monotonicity analysis showed that the model’s predictions aligned with expected trends in concrete behavior. Sensitivity testing indicated that changes in silica fume content and coarse aggregate proportion produced the most significant fluctuations in predicted strength.</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 LightGBM model for predicting compressive strength of silica fume modified high-volume fly ash concrete

  • 期刊论文

  • 发表日期：

    2025年6月23日

查看全文
  • Anish Kumar,
  • Sameer Sen,
  • Sanjeev Sinha,
  • Bimal Kumar,
  • Chaitanya Nidhi

摘要

This research introduces a robust machine learning framework for estimating the compressive strength of concrete, utilizing a Light Gradient Boosting Machine (LightGBM) regression algorithm. The model was developed using a diverse dataset that included different mix proportions of fly ash, silica fume, cement, fine and coarse aggregates, along with varying curing durations. After a thorough hyperparameter optimization process, the final model incorporated a learning rate of 0.1, 200 boosting iterations, an unrestricted tree depth, and 31 maximum leaf nodes. The model demonstrated strong predictive accuracy, achieving an R² value of 0.99 on the training set and 0.97 on the testing set, with corresponding Mean Absolute Errors (MAE) of 0.70 MPa and 1.35 MPa. Feature importance derived from SHAP values highlighted curing duration, silica fume percentage, and cement content as primary contributors to strength outcomes. Additional interpretation through partial dependence plots and monotonicity analysis showed that the model’s predictions aligned with expected trends in concrete behavior. Sensitivity testing indicated that changes in silica fume content and coarse aggregate proportion produced the most significant fluctuations in predicted strength.