View in EDS

Structural failure risk assessment of shield tunnel using large language model.

Saved in:
Bibliographic Details
Title: Structural failure risk assessment of shield tunnel using large language model.
Authors: Atangana Njock, Pierre Guy1 (AUTHOR) pierre-guy.atangananjock@polyu.edu.hk, Yin, Zhen-Yu1,2 (AUTHOR) zhenyu.yin@polyu.edu.hk, Xu, Hao-Ruo1 (AUTHOR) haoruo.xu@connect.polyu.hk, Zhang, Ning1 (AUTHOR) ning-cee.zhang@polyu.edu.hk
Source: Tunneling & Underground Space Technology. Nov2025, Vol. 165, pN.PAG-N.PAG. 1p.
Subjects: Risk assessment, Language models, Decision support systems, Natural language processing, Information asymmetry, Machine learning, Tunnel design & construction
Abstract: Tunnel engineering grapples with persistent challenges in risk assessment, including unpredictable geological variability, data scarcity, and the inaccessibility of advanced predictive tools for non-specialists. While existing machine learning methods improve risk quantification, their reliance on structured data inputs and opaque expert-dependent interfaces limit real-world adoption. This study presents the first language-driven framework for tunnel risk assessment, which enables non-experts to conduct robust risk predictions through intuitive natural language interactions. The proposed approach integrates the lightweight transformer DistilBERT with attention-guided class balancing to address data imbalance and enhance interpretability. Fine-tuned on 1000 global tunneling scenarios spanning diverse geological and operational conditions, the model achieves high accuracy across all risk levels with metrics (Recall, Precision, F1) scores varying from 0.96 to 1. The minimum area under curve was also found to be 0.99. Validation against global case studies gathered from literature confirm robust performance under data scarcity, with 85 % confidence in high-risk predictions. Comparative analyses demonstrate superiority over context-free and few-shots GPT-4 and DeepSeek in risk classification, while attention visualizations reveal interpretable decision pathways focused on critical parameters like groundwater levels. This seminal study demonstrates the potential of large language model-driven frameworks as decision-support tools in geotechnical risk assessment. [ABSTRACT FROM AUTHOR]
Copyright of Tunneling & Underground Space Technology is the property of Pergamon Press - An Imprint of Elsevier Science and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:Tunnel engineering grapples with persistent challenges in risk assessment, including unpredictable geological variability, data scarcity, and the inaccessibility of advanced predictive tools for non-specialists. While existing machine learning methods improve risk quantification, their reliance on structured data inputs and opaque expert-dependent interfaces limit real-world adoption. This study presents the first language-driven framework for tunnel risk assessment, which enables non-experts to conduct robust risk predictions through intuitive natural language interactions. The proposed approach integrates the lightweight transformer DistilBERT with attention-guided class balancing to address data imbalance and enhance interpretability. Fine-tuned on 1000 global tunneling scenarios spanning diverse geological and operational conditions, the model achieves high accuracy across all risk levels with metrics (Recall, Precision, F1) scores varying from 0.96 to 1. The minimum area under curve was also found to be 0.99. Validation against global case studies gathered from literature confirm robust performance under data scarcity, with 85 % confidence in high-risk predictions. Comparative analyses demonstrate superiority over context-free and few-shots GPT-4 and DeepSeek in risk classification, while attention visualizations reveal interpretable decision pathways focused on critical parameters like groundwater levels. This seminal study demonstrates the potential of large language model-driven frameworks as decision-support tools in geotechnical risk assessment. [ABSTRACT FROM AUTHOR]
ISSN:08867798
DOI:10.1016/j.tust.2025.106882