A semantic entropy framework for quantifying and mitigating uncertainty in LLM-assisted geotechnical design

Large language models (LLMs) present new opportunities for decision support in geotechnical engineering but remain prone to uncertainty and domain-specific hallucinations. This paper develops a semantic entropy framework to diagnose and mitigate such uncertainty in geotechnical design. The proposed semantic entropy framework decomposes LLM outputs into atomic factoids, evaluates their entropy through entailment-based clustering of diverse sampled answers, and refines uncertain factoids via targeted reprompting, knowledge-based verification, and calibrated hedging. A reliability index (RI) is also proposed to merge aggregated entropy measures with BERT similarity to empirical sources, enabling task-specific reliability estimation. Experiments in foundation design and soil consolidation testing domains are adopted for validation, which highlights distinct uncertainty distribution profiles linked to domain standardization. Results also show successful remediation of 82%–88% of initially hallucinated responses while preserving 100% of clean outputs. Reprompting emerges as the most effective mitigation strategy, achieving 39%–46% entropy reduction. The RI demonstrates strong predictive capability ( R 2 = 0.88) in identifying reliable LLM responses. Overall, the framework provides a practical decision-support mechanism for integrating uncertainty-aware LLM outputs into geotechnical design practice.