Using Large Language Models for Systematic Failure Model Identification in Levee Infrastructure

Levees are essential for flood protection; however, failures occur frequently due to complex geotechnical conditions exacerbated by shifting climate patterns. Traditional methods for evaluating levee vulnerability rely heavily on manual interpretation of heterogeneous inspection reports, which limits accuracy and scalability. We introduce a framework that utilizes large language models (LLMs) to automatically extract and classify levee failure mechanisms (e.g., seepage, erosion) from unstructured textual evaluations. Our approach employs zero-shot learning to convert diverse assessment data into structured semantic representations, followed by fine-tuning domain-adapted LLMs for classification. Experimental results show that domain-adapted LLMs outperform baseline classifiers by 21% based on the Matthews correlation coefficient, significantly improving failure-mode prediction accuracy. This automated, data-driven approach provides a robust foundation for accurate identification of levee vulnerabilities, allowing for proactive risk mitigation in response to increasing climate-related threats. Our findings demonstrate how advancements in natural language processing can improve resilience assessments in critical infrastructure management.