Scholarlink

<p>Traditional analytical solutions and finite element modelling for the settlement of cohesionless soil under strip-footing loads are constrained by simplifying assumptions and high computational cost. This study developed and compared four ensemble learning models, namely Random Forest, XGBoost, LightGBM, and CatBoost, to reconstruct continuous settlement–load curves rather than a single settlement value. A dataset of 418 samples was generated from 19 finite element simulation scenarios, each described by ten input variables covering the applied loading and the physical and mechanical properties of the soil. The hyperparameters of the three boosting models were tuned with a Tree-structured Parzen Estimator using 60–80 trials and a fixed random seed. CatBoost achieved the highest accuracy on the test set, with a coefficient of determination of 0.9996, a root-mean-square error (RMSE) of 6.28&#xa0;mm, and a mean absolute error of 3.80&#xa0;mm, which reduced the RMSE by 89.3% relative to the Random Forest baseline. A Shapley additive explanations analysis identified the applied load and the internal friction angle as the dominant variables, in agreement with the Mohr–Coulomb shear-strength criterion. The optimised CatBoost model reproduced the finite element settlement curves with relative errors below 1.3%.</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

Bayesian-Optimized Ensemble Machine Learning for Predicting Settlement of Cohesionless Soil Under Loads

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Luan Vo Nhat,
  • Tuan Nguyen Anh,
  • Hoa Tran Vu Van

摘要

Traditional analytical solutions and finite element modelling for the settlement of cohesionless soil under strip-footing loads are constrained by simplifying assumptions and high computational cost. This study developed and compared four ensemble learning models, namely Random Forest, XGBoost, LightGBM, and CatBoost, to reconstruct continuous settlement–load curves rather than a single settlement value. A dataset of 418 samples was generated from 19 finite element simulation scenarios, each described by ten input variables covering the applied loading and the physical and mechanical properties of the soil. The hyperparameters of the three boosting models were tuned with a Tree-structured Parzen Estimator using 60–80 trials and a fixed random seed. CatBoost achieved the highest accuracy on the test set, with a coefficient of determination of 0.9996, a root-mean-square error (RMSE) of 6.28 mm, and a mean absolute error of 3.80 mm, which reduced the RMSE by 89.3% relative to the Random Forest baseline. A Shapley additive explanations analysis identified the applied load and the internal friction angle as the dominant variables, in agreement with the Mohr–Coulomb shear-strength criterion. The optimised CatBoost model reproduced the finite element settlement curves with relative errors below 1.3%.