A cross-scale normative encoding representation method for 3D building models suitable for graph neural networks

Artificial intelligence has brought significant efficiency improvements to computer-aided architectural design (CAAD), especially in the field of building simulation. Machine learning methods, particularly neural networks, have been widely applied to achieve rapid performance predictions for energy consumption, lighting, and wind environments. However, traditional neural network methods, such as CNNs, often require data input in the form of single 2D drawings or preprocessed data, resulting in the loss of spatial information inherent to the building. These methods are also challenging to adapt for different types of performance predictions. Although proxy models trained with neural networks can improve prediction speed, the inconvenience in the encoding process reduces the usability of workflow.

This study aims to address this issue by developing a normative and scalable encoding method. A Graph Neural Network (GNN) is used to represent the inherent complex relationships and hierarchical structures within 3D building models, encoding them to preserve their geometric and topological attributes across different scales.

The proposed method implements a three-stage process for spatial encoding of 3D models. First, a graph segmentation algorithm suitable for building spaces is introduced, which segments complex non-convex spaces such as corridors and atriums into convex space shapes, automatically adding air walls at the segmentation interfaces to simplify the spatial topology. Second, a graph topology structure is proposed that simultaneously represents building components like walls, windows and roofs, as well as building spaces and their interrelationships. Finally, an encoding method suitable for GNN is constructed, encapsulating geometric objects as node attributes in the form of oriented bounding boxes (OBB) and converting the connection relationships between faces and spaces into multi-dimensional arrays.

Preliminary results indicate that our method significantly enhances the encoding robustness of building models, achieving a conversion success rate of over 95% for more than 2,000 cases, and greatly reduces the loss of model and spatial information. The encoded data can be unfolded into simulation file formats for performance software like EnergyPlus, or directly converted into parametric building models for optimization design. More importantly, this method achieves generalized encoding of building models, providing a technical and algorithmic foundation for large-scale neural network training and generation based on this specific type of 3D model. Future work will focus on optimizing the method to improve encoding precision and robustness and exploring its diverse applications in architectural and urban planning scenarios.