Predicting Connection Types Using Graph Neural Networks

Accurate information on how building components are connected is essential for assessing circularity and planning disassembly processes. Yet such connection-type data are rarely available or standardized in digital building representations und thus need manual enrichment. This study investigates whether graph neural networks (GNNs) can predict connection types and their scores between building components. To this end, the archetypal Belgian TOTEM component database was transformed into a disassembly knowledge graph, and multiple GNN architectures (GraphConv, SAGEConv, HGTConv) were trained using both classification and regression strategies based on connection-type labels and scores. The results demonstrate that GNNs can learn structural and semantic patterns between components, enabling the prediction of connection types with up to 70% accuracy, even under incomplete information, with classification-based training tasks showing higher accuracy compared to 60% of regression-based training. These findings advance data-driven methods for design-for-disassembly and circular construction, supporting semi-automated enrichment of building information models and 2D-drawing data with critical constructability information.