Evaluating large language models (LLMs) for semantic interpretation of IFC-based BIM data

초록

Building Information Modeling (BIM) produces structured datasets, commonly exchanged through Industry Foundation Classes (IFC) format, that capture information about building geometry, object properties, and spatial relationships. Although these datasets are rich in detail, they usually require proprietary software and expert knowledge to interpret, which limits their broader usability. Recent advances in large language models (LLMs) raise the question of whether they can serve as independent interpreters of BIM data when provided in text-based formats. This paper investigates that possibility by restructuring IFC files into JSON and testing LLMs across two tasks: (1) exploring natural language queries on a BIM model and (2) detecting modifications between paired models. Results show that LLMs can extract and summarize information with reasonable accuracy when data is explicitly structured, though spatial reasoning and subtle modifications posed consistent challenges. These findings provide evidence for a text-driven approach to BIM interpretation and highlight opportunities for platform-independent applications.

키워드