Abstract
Bridges are critical components of infrastructure networks, ensuring the connectivity and accessibility of modern cities. However, structural analysis of bridges, particularly finite element modeling, remains a highly manual, time-consuming, and labor-intensive process. Recent advances in large language models (LLMs) offer transformative potential to revolutionize how engineers interact with computational tools. Motivated by this shift, this paper develops an LLM-based agent to automate finite element analysis of bridges. The agent leverages in-context learning strategies to guide the reasoning process of LLM, integrating expert-crafted chain-of-thought prompts with retrieval-augmented generation (RAG). Upon receiving natural language input, the agent automatically interprets the problem description, generates executable code, and visualizes analysis results. Evaluations on benchmark problems demonstrate the agent's capability to accurately model geometry, boundary, and loading conditions. Across repeated trials, the agent achieves an average accuracy exceeding 98%, indicating strong reliability and consistency.