Scholarlink

<p>Deformation induced by deep excavation is inevitable, yet highly challenging, for many infrastructure projects. Existing analytical approaches rely on idealistic simplifications, while alternative physics-informed neural networks are hindered by their simplified physical constraints, ignoring the effects of three-dimensional soil behavior. This study proposes a new finite physics-informed deep learning (Finite-PIDL) framework to accurately predict deformations induced by deep excavation and support risk assessment. The method leverages transfer learning by integrating 3D finite element simulation with the adaptability of governing physical patterns, considering nonlinear physical constraints to fully simulate ground environment within sparse field measurement. To ensure critical knowledge transfer, the model training incorporates an ensemble loss function, enforcing consistency with known physical principles while fitting the observational data. The Finite-PIDL is applied for both forward forecasting of ground deformations and inverse estimations of uncertain parameters from monitoring data.&#xa0;The method was validated using field data from Xiamen Rail Transit Line 6 project. Results show the effectiveness of the proposed model in handling physical and mechanical parameters over existing physics-informed methods. Finite-PIDL realized the lowest mean absolute error (0.29) and highest correlation coefficient (93.6%), demonstrating the accuracy&#xa0;of the model in predicting the ground deformations.</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

Finite physics-informed deep learning for predicting deformations and mitigating risks induced by deep excavation

  • 期刊论文

  • 发表日期：

    2026年7月27日

查看全文
  • Khalid Elbaz,
  • Qipeng Cai,
  • Liang Shuang,
  • Lin Dingzong,
  • Xiangyu Guo

摘要

Deformation induced by deep excavation is inevitable, yet highly challenging, for many infrastructure projects. Existing analytical approaches rely on idealistic simplifications, while alternative physics-informed neural networks are hindered by their simplified physical constraints, ignoring the effects of three-dimensional soil behavior. This study proposes a new finite physics-informed deep learning (Finite-PIDL) framework to accurately predict deformations induced by deep excavation and support risk assessment. The method leverages transfer learning by integrating 3D finite element simulation with the adaptability of governing physical patterns, considering nonlinear physical constraints to fully simulate ground environment within sparse field measurement. To ensure critical knowledge transfer, the model training incorporates an ensemble loss function, enforcing consistency with known physical principles while fitting the observational data. The Finite-PIDL is applied for both forward forecasting of ground deformations and inverse estimations of uncertain parameters from monitoring data. The method was validated using field data from Xiamen Rail Transit Line 6 project. Results show the effectiveness of the proposed model in handling physical and mechanical parameters over existing physics-informed methods. Finite-PIDL realized the lowest mean absolute error (0.29) and highest correlation coefficient (93.6%), demonstrating the accuracy of the model in predicting the ground deformations.