Spatial Link Prediction: Learning topological relationships in MEP systems
Christoph Emunds, Jérôme Frisch, Christoph van Treeck · Advanced Engineering Informatics · 2025
In Building Information Modeling (BIM), interactions between building components are defined by parametric relationships. These relationships are critical for analyses and simulations throughout a building’s life-cycle. However, many of these relationships are implicit in BIM models, posing challenges when exchanging models between different BIM authoring tools. Existing query systems and languages for BIM can partially infer or deduce missing information , incorporating semantic, geometric, and topological data to make it more accessible. Nonetheless, they often lack a holistic understanding of the building model, focusing on individual vertices, edges, and faces of components’ 3D representations. In this paper, we present a novel approach for predicting connectivity relationships in mechanical, electrical and plumbing (MEP) systems using a candidate generation procedure followed by an auto-encoder model to identify relevant component connections. Our approach leverages both the topological structure of MEP systems and the geometric attributes of system components, allowing the model to learn effectively and accurately predict potential connections. We conduct extensive experiments across various settings to evaluate the practical applicability of our approach and demonstrate its effectiveness on graphs of real-world MEP systems. In addition, we design a decoder that explicitly considers the direction and distance between components in order to enhance spatial realism in the predicted connections. To train our models for the task of link prediction in MEP systems, we assembled a dataset of 27 BIM models from residential, industrial, and public buildings. Our results indicate that our approach can significantly enhance the prediction of connectivity relationships, offering valuable insights for improving BIM workflows and interoperability. The source code of our models and experiments is available at https://github.com/RWTH-E3D/SpatialLinkPrediction .