Towards semantic enrichment of IFC models through language modeling

Semantic enrichment of Building Information Models (BIM) is critical for improving model quality and supporting downstream applications, such as compliance checking and design automation. This study explores continued pre-training of large language models on building information models in the Industry Foundation Classes (IFC) format to infer missing information required for Automated Compliance Checking (ACC). By evaluating encoder-only and decoder-only transformer architectures across different test cases, we demonstrate how much information can be inferred automatically and identify which model types and data representations are most suitable for the task. A modified ifcJSON-5a* representation was found to improve performance over traditional EXPRESS formats. The findings suggest that domain-adapted language models are a promising approach for advancing semantic enrichment in BIM and open new directions for integrating AI more deeply into the building design process.