CAAD Futures 2025
July 4, 2025 / Hong Kong
On July 4th, Research Associate Markus Renner presented the paper titled “Augmenting IFC-Based BIM Models for Graph Neural Network Training of Non-Orthogonal Spatial Typologies: A Context-Informed Design Methodology” at the CAAD Futures 2025 conference.
Abstract:
Building Information Modeling (BIM) and Industry Foundation Classes (IFC) data are valuable resources in architecture, offering detailed and structured information on building elements. However, using such resources to develop artificial intelligence (AI) models that are capable of learning and generating spatial and geometrical suggestions remains a challenging problem. This study presents a systematic method for converting real-world BIM/IFC projects into graph representations, suitable for training Graph Neural Networks (GNNs), which overcome the limitations inherent to conventional AI techniques based on pixel and 3D voxels in architecture. First, a custom schema was defined to encode each building element and spatial relationship as triplets, allowing for the generation of heterogeneous graphs that capture geometry, topology and semantic labels. Using models provided by multiple architecture offices, we extracted over 180,000 subgraphs that represent diverse non-orthogonal spatial typologies. Next, each subgraph was annotated with multi-label classifications (element type) and continuous attributes (e.g. dimensions, orientation) to support both classification and regression tasks. Finally, to demonstrate the utility of the dataset, we trained a custom generative neural network (NN) to complete partial spatial graphs, producing non-orthogonal room layouts based on the existing building context. The custom NN suggests that the dataset has the capacity to support advanced 3D generative tasks in architecture and could lay the groundwork for future studies on automated design synthesis.