Completed building information models contain reusable design components, but this knowledge often remains inaccessible since manually maintained component libraries do not scale as archives expand. This work examines how a combined supervised and unsupervised approach can extract reusable components from vendor-neutral Industry Foundation Classes (IFC) data and whether combining geometric, semantic, and graph features enhances component mining and retrieval. An end-to-end pipeline converts raw IFC files into a searchable component library using per-modality encoding and fusion, classification enrichment, direct and deep-embedded clustering, and indexed retrieval. The approach was evaluated through systematic modality ablation with bootstrap confidence intervals on about 42,000 components from 50 models annotated with a three-level taxonomy. In classification, incorporating learned geometric representations improved the ifcclass and component-type classifiers. The best performing subset achieved a F1-score of 97.28 on the IFCNetCore benchmarking for ifcclass. For extraction, semantic-only semi-supervised methods yielded a V-measure of 92.94. Dense semantic retrieval achieved a top-5 precision of 97.13 on the full dataset, and was extended to library-level and free-text queries with optional language-model re-ranking. The results demonstrate that semantic features alone achieve the highest extraction quality under both classical and deep clustering, and that multimodal fusion recovers value only at retrieval.     «

Completed building information models contain reusable design components, but this knowledge often remains inaccessible since manually maintained component libraries do not scale as archives expand. This work examines how a combined supervised and unsupervised approach can extract reusable components from vendor-neutral Industry Foundation Classes (IFC) data and whether combining geometric, semantic, and graph features enhances component mining and retrieval. An end-to-end pipeline converts raw...     »