Scholarlink

<p>This study presents a framework enhanced by AI for high-resolution liquefaction hazard mapping. This framework leverages stacked ensemble machine learning and large-scale borehole databases. The proposed approach uses random forest, support vector regression, and gradient boosting decision trees to predict subsurface parameters at unsampled locations and generate potential liquefaction (PL) maps on 200&#xa0;m by 200&#xa0;m grids. Using 13,926 geotechnical borehole records from a seismically active metropolitan area, the optimized ensemble achieves coefficients of determination greater than 0.70 at shallow critical depths. This represents a fourfold improvement in spatial resolution compared to conventional 1&#xa0;km grid hazard maps. Model performance analysis reveals a strong dependence on local data density. This indicates that approximately 350 boreholes within a 5,000&#xa0;m radius are required to ensure robust predictions. Contributions from data beyond 10,000&#xa0;m become negligible. The resulting high-resolution PL maps delineate extensive moderate to high liquefaction hazard zones, providing actionable information for targeted retrofitting, land-use regulation, and resilient infrastructure planning. This scalable, transferable, AI-driven methodology offers a framework for evaluating liquefaction risk in earthquake-prone megacities. It supports sustainable urban planning and disaster risk reduction under conditions of sparse but expanding geotechnical data.</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

AI-Enhanced High-Resolution Liquefaction Hazard Mapping Using Ensemble Machine Learning and Borehole Big Data

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • Shinya Inazumi,
  • Kuo Chieh Chao

摘要

This study presents a framework enhanced by AI for high-resolution liquefaction hazard mapping. This framework leverages stacked ensemble machine learning and large-scale borehole databases. The proposed approach uses random forest, support vector regression, and gradient boosting decision trees to predict subsurface parameters at unsampled locations and generate potential liquefaction (PL) maps on 200 m by 200 m grids. Using 13,926 geotechnical borehole records from a seismically active metropolitan area, the optimized ensemble achieves coefficients of determination greater than 0.70 at shallow critical depths. This represents a fourfold improvement in spatial resolution compared to conventional 1 km grid hazard maps. Model performance analysis reveals a strong dependence on local data density. This indicates that approximately 350 boreholes within a 5,000 m radius are required to ensure robust predictions. Contributions from data beyond 10,000 m become negligible. The resulting high-resolution PL maps delineate extensive moderate to high liquefaction hazard zones, providing actionable information for targeted retrofitting, land-use regulation, and resilient infrastructure planning. This scalable, transferable, AI-driven methodology offers a framework for evaluating liquefaction risk in earthquake-prone megacities. It supports sustainable urban planning and disaster risk reduction under conditions of sparse but expanding geotechnical data.