Scholarlink

<p>Scour is a leading cause of bridge failure, and the hydraulic critical shear stress of soil governs erosion initiation around bridge foundations. A machine learning–based framework was developed to predict the critical shear stress of lean clay soils using experimental and augmented datasets. Soil samples were collected from two bridge sites in Illinois and tested using a portable scour testing device, while soil index properties were obtained through laboratory testing. Due to limited experimental data, a synthetic minority over-sampling technique with Gaussian noise was applied to balance underrepresented critical shear stress ranges. To avoid data leakage, 40% of the experimental data were reserved exclusively for independent model evaluation. An extreme gradient boosting model was trained and evaluated against a ridge regression baseline. The proposed model achieved strong predictive performance on independent experimental data, with coefficient of determination of 0.986 and mean absolute error of 0.33&#xa0;Pa. Model interpretability analyses confirmed physically consistent relationships between critical shear stress and key soil properties, including soil strength, plasticity, and mean particle size. The results demonstrated that the proposed ML framework effectively captured the nonlinear behavior governing erosion resistance in cohesive soils.</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

Machine Learning-based Prediction of Critical Shear Stress of Lean Clay Soil at Two Bridge Sites in Illinois Using Experimental and Augmented Data

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Majid Khan,
  • Abdolreza Osouli,
  • Khurram Shahzad,
  • Heather Shoup

摘要

Scour is a leading cause of bridge failure, and the hydraulic critical shear stress of soil governs erosion initiation around bridge foundations. A machine learning–based framework was developed to predict the critical shear stress of lean clay soils using experimental and augmented datasets. Soil samples were collected from two bridge sites in Illinois and tested using a portable scour testing device, while soil index properties were obtained through laboratory testing. Due to limited experimental data, a synthetic minority over-sampling technique with Gaussian noise was applied to balance underrepresented critical shear stress ranges. To avoid data leakage, 40% of the experimental data were reserved exclusively for independent model evaluation. An extreme gradient boosting model was trained and evaluated against a ridge regression baseline. The proposed model achieved strong predictive performance on independent experimental data, with coefficient of determination of 0.986 and mean absolute error of 0.33 Pa. Model interpretability analyses confirmed physically consistent relationships between critical shear stress and key soil properties, including soil strength, plasticity, and mean particle size. The results demonstrated that the proposed ML framework effectively captured the nonlinear behavior governing erosion resistance in cohesive soils.