Scholarlink

<p>Soil liquefaction under strong earthquakes is the primary cause of damage to foundations and superstructures. Predicting the potential for seismic-induced soil liquefaction is key to preventing related disasters. This study compares four machine learning (ML) models for soil liquefaction potential based on cone penetration test datasets: decision tree, random forest, gradient boosting, and extreme gradient boosting. The database was collected from previously published research and includes information on earthquake moment magnitude, peak ground acceleration, depth of soil layer, total vertical stresses, effective vertical stresses, and cone tip stresses. The predictive capabilities of the developed models were evaluated using overall accuracy, precision, recall, F-measure, and receiver operating characteristic curves. The results showed that the extreme gradient boosting model exhibited the highest efficacy. A subsequent analysis of feature importance demonstrated that cone tip stresses exerted the most significant influence on soil liquefaction potential. In a final comparative assessment with conventional liquefaction discrimination theory methods, the study revealed that the Robertson method yielded a higher success rate for liquefaction cases, the Olsen method was more successful in non-liquefaction cases, and the ML approach manifested superior success rates in both liquefaction and non-liquefaction cases.</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

Adoption of tree-based machine learning algorithms and CPT data for liquefaction potential assessment

  • 期刊论文

  • 发表日期：

    2026年4月25日

查看全文
  • Shuang Tian,
  • Pan Si,
  • Liang Tang,
  • Xianzhang Ling,
  • Yanfang Liu

摘要

Soil liquefaction under strong earthquakes is the primary cause of damage to foundations and superstructures. Predicting the potential for seismic-induced soil liquefaction is key to preventing related disasters. This study compares four machine learning (ML) models for soil liquefaction potential based on cone penetration test datasets: decision tree, random forest, gradient boosting, and extreme gradient boosting. The database was collected from previously published research and includes information on earthquake moment magnitude, peak ground acceleration, depth of soil layer, total vertical stresses, effective vertical stresses, and cone tip stresses. The predictive capabilities of the developed models were evaluated using overall accuracy, precision, recall, F-measure, and receiver operating characteristic curves. The results showed that the extreme gradient boosting model exhibited the highest efficacy. A subsequent analysis of feature importance demonstrated that cone tip stresses exerted the most significant influence on soil liquefaction potential. In a final comparative assessment with conventional liquefaction discrimination theory methods, the study revealed that the Robertson method yielded a higher success rate for liquefaction cases, the Olsen method was more successful in non-liquefaction cases, and the ML approach manifested superior success rates in both liquefaction and non-liquefaction cases.