This thesis explores how to use graph neural networks to identify important connections between architectural and structural building models in a cloud-based Building Information Modelling (BIM) setup. The work is motivated by challenges faced when digital BIM models from different disciplines are linked together, in particular the lack of important information such as explicit semantic relationships between elements. While links within one discipline are usually defined clearly, cross-disciplinary links, such as host or corresponds_to, are often missing or only implicitly assumed, which easily leads to confusion during coordination. To address this, a hybrid approach is developed that combines a rule-based method with Graph Neural Networks (GNNs) for link prediction. A custom pipeline is implemented to extract geometry from IFC files, generate candidate element pairs, and label them manually using a custom Revit-based tool. Several GNN architectures, including GCN, GraphSAGE, GAT, and GIN, are tested and compared with the rule-based baseline using precision, recall, F1-score, and confusion matrices on both seen and previously unseen projects. The experiments show that the rule-based method performs well when the geometry is clean and the situation can be described easily using rules. As cases become more ambiguous, the learning-based models handle them better and generalize more consistently across projects. The results suggest that GNN-based link prediction can complement rule-based methods and support the automated inference of semantic links in federated BIM environments.     «

This thesis explores how to use graph neural networks to identify important connections between architectural and structural building models in a cloud-based Building Information Modelling (BIM) setup. The work is motivated by challenges faced when digital BIM models from different disciplines are linked together, in particular the lack of important information such as explicit semantic relationships between elements. While links within one discipline are usually defined clearly, cross-disciplin...     »