IGeo-LLM : identification of geological features using multi-source data based on LLM during shield tunnelling

Abstract

Geological information is essential to ensure the safety and efficiency of shield tunnelling. However, the concealed excavation environment limits direct observation of geological features and therefore increases construction risk during tunnelling. This study proposes a framework for identification of geological features using multi-source data and large language model (iGeo-LLM) during shield tunnelling. The multi-source data, including shield operation data (SD), muck image data (ID), and vibration data (VD), were analysed by ChatGPT model to extract their features and generate the structured data, thereby facilitating effective understanding by LLMs under a few-shot prompt learning paradigm. The retrieval-augmented generation flow (RAGFlow) was then incorporated to enhance output reliability and improve the factual grounding of LLM inference by introducing project-specific structured knowledge. The external knowledge base in RAGFlow was constructed using different proportions of multi-source data (hereafter called RAG data). Prompt engineering was further employed to constrain evidence-based output generation, and improve the interpretability and reliability. Results show that both RAG enhancement and multi-source data fusion markedly improve the performance of LLM-based geological-feature identification. The proposed iGeo-LLM achieves its best performance under the integrated SD, ID, and VD setting. The main advantage of the proposed framework lies in retrieval-augmented reasoning over heterogeneous tunnelling data, while conventional machine learning models remain highly effective on single-source datasets. A case study in Guangzhou demonstrates the feasibility and effectiveness of the proposed multi-source RAG-enhanced framework for geological interpretation during shield tunnelling.