A New Perspective and Progressive Applications of the Large Language Models in Geotechnical Engineering Field-Paper

Paper

Pdf Summary

The paper reviews the potential and limitations of large language models (LLMs) in geotechnical engineering. It explains how transformer-based models can help process technical documents, standards, and project information, while emphasizing that language models generate likely text rather than reliably applying engineering equations.<br /><br />Examples from prior studies show mixed performance: LLMs can assist with soil classification and engineering calculations, but may misapply concepts or produce incorrect results. The paper compares general-purpose models with domain-specific systems such as BloombergGPT, highlighting the value of curated training data, fine-tuning, suitable model and data scaling, and rigorous evaluation. These approaches could inform future geotechnical models, though regional standards, limited data sharing, and location-dependent empirical parameters present challenges.<br /><br />The authors demonstrate prompt engineering through settlement calculations. Clear instructions and supplied assumptions enabled ChatGPT to match a simple immediate-settlement result verified with Settle3. A more complex case produced discrepancies, illustrating the need for engineering judgment and independent verification. One-shot or few-shot examples can improve responses, but adding many examples may offer little additional benefit.<br /><br />A major concern is hallucination: LLM outputs can conflict with verified facts even when prompts are well formed. Techniques such as retrieval and interaction with external tools, factual benchmarks, and answer evaluation may reduce this risk, but cannot eliminate it. The paper concludes that LLMs could become useful engineering assistants, provided practitioners understand their assumptions, validate their outputs, and retain responsibility for safety-critical decisions.

Keywords

large language models

LLMs

geotechnical engineering

transformer-based models

soil classification

engineering calculations

prompt engineering

hallucination

fine-tuning

retrieval