Few-shot learning of geological cross-sections from sparse data using large language model

Artificial intelligence (AI) has achieved remarkable milestones and surpassed human capabilities in many fields, such as Go and image recognition. There have been numerous efforts to leverage AI in promoting digitalization and automation in civil engineering. This is particularly challenging in geotechnical engineering, where geological modelling still heavily depends on subjective prior knowledge and is hindered by data scarcity. Therefore, incorporating valuable prior knowledge for the automatic development of subsurface geological cross-sections from sparse geodata remains an unsolved challenge. To address these issues, this study proposes a few-shot learning strategy based on large language models (LLMs) to generate two-dimensional geological cross-sections from sparse site investigation data. First, geological modelling is framed as an in-context learning problem, where LLMs are customized with geological domain knowledge by learning reasoning logic from a few worked examples. In addition, advanced augmented prompting strategies are employed to minimize prediction errors and improve model reliability. The effectiveness of the proposed strategy is demonstrated through two real-world examples. Results indicate that LLMs can internalize geological reasoning while developing accurate and consistent cross-sections. The findings underscore the significant potential of LLMs in geotechnical site characterization, marking an important step towards digitalization and automation.