Scholarlink

<p>This study proposes a novel three-fidelity physics-informed neural network (TF-PINN) to solve forward and inverse problems in tunneling applications. The proposed model incorporates: (i) low-fidelity data generated by numerical simulations, (ii) mid-fidelity physics-based analytical patterns for closed-form solution, (iii) high-fidelity field data from a tunnel case study, leveraging the underlying physical information to simulate the soil–tunnel interactions within a complete simulation environment. The model hierarchically trains three-fidelity learning with uncertainty quantification and transfer learning to explicitly enhance training efficiency and preserve critical physical knowledge. TF-PINN is applied to analyze the interaction mechanism between the shield and the surrounding strata. Results reveal that the proposed model exhibits inherent advantages in handling physical constraints and data-driven approaches over existing physics-informed methods. TF-PINN provided the highest correlation coefficient (92.5%) and lowest root mean square error (0.052), demonstrating the precision of the testing model in forecasting the ground deformation. The proposed method can be utilized in the early design process to give accurate estimations for soil–tunnel deformations when data are limited.</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

TF-PINN model for solving forward and inverse problems during shield tunneling

  • 期刊论文

  • 发表日期：

    2025年8月18日

查看全文
  • Khalid Elbaz,
  • Qipeng Cai,
  • Shui-Long Shen,
  • Dongxi Xie

摘要

This study proposes a novel three-fidelity physics-informed neural network (TF-PINN) to solve forward and inverse problems in tunneling applications. The proposed model incorporates: (i) low-fidelity data generated by numerical simulations, (ii) mid-fidelity physics-based analytical patterns for closed-form solution, (iii) high-fidelity field data from a tunnel case study, leveraging the underlying physical information to simulate the soil–tunnel interactions within a complete simulation environment. The model hierarchically trains three-fidelity learning with uncertainty quantification and transfer learning to explicitly enhance training efficiency and preserve critical physical knowledge. TF-PINN is applied to analyze the interaction mechanism between the shield and the surrounding strata. Results reveal that the proposed model exhibits inherent advantages in handling physical constraints and data-driven approaches over existing physics-informed methods. TF-PINN provided the highest correlation coefficient (92.5%) and lowest root mean square error (0.052), demonstrating the precision of the testing model in forecasting the ground deformation. The proposed method can be utilized in the early design process to give accurate estimations for soil–tunnel deformations when data are limited.