Automated structural analysis model generation for PSC-I girder bridges using an ontology-grounded GraphRAG and prompt engineering framework

: Structural analysis modeling of bridges requires defining geometry, materials, connectivity, and boundary conditions, and remains heavily reliant on expert manual effort. Although large language models can generate code, their direct application introduces errors, including hallucinated properties, unit inconsistencies, and missing boundary conditions. This study proposes a three-layer knowledge architecture for PSC-I girder bridges integrated via GraphRAG, comprising an OWL ontology, a Neo4j knowledge graph, and structured prompts encoding OpenSeesPy modeling rules. An ablation study on a single-span bridge showed that the dataset-only approach produced critical errors, whereas adding prompts resolved structural errors but not data-interpretation failures. The full framework generated models matching the references in displacement, shear force, reactions, and bending moment. In ten repeated generation runs per bridge, all codes executed successfully, member-bound geometry and layout information were reproduced identically, and residual errors remained at the code-implementation level, confirming that hierarchical knowledge separation enables reliable automated model generation.