Large language model-empowered paradigm for automated geotechnical site planning and geological characterization

Abstract

A sound site investigation scheme must satisfy requirements of various local, regional, or national codes, and it is imperative to have an efficient system for information retrieval, summarization, and reasoning along with a rapid interpretation tool for real-time risk-informed decision-making. Emerging large language models (LLMs) offer a promising solution for automatically processing unstructured natural languages and facilitating collaborative reasoning between humans and machines. This paper develops a customized LLM-based agent named “Geologist” to streamline geotechnical site planning and subsequent geological interpretation. A Multihop-Retrieval-Augmented Generation system is proposed to retrieve site-specific requirements from multiple site investigation design codes. Moreover, a progressive human-machine collaboration scheme is orchestrated for interpretable geological modelling. The efficiency of the proposed LLM-guided paradigm is validated through illustrative examples and real-world case histories. Results show that the proposed workflow facilitates real-time and accurate information retrieval as well as automatic development of subsurface geological cross-sections with quantified stratigraphic uncertainty.