From NURBS to neural networks: Efficient geometry encoding for generative AI in architectural design

The existing 3D representations for AI models, such as meshes, voxels, signed distance functions, and point clouds, are not compatible with architectural design workflows that rely on NURBS geometry, which is mainly used in CAD programs. ​These formats lead to large datasets, high computational costs, and loss of geometric precision, limiting further usability in CAD software. This research introduces a novel methodology for encoding NURBS geometries into compact, tensor-based NumPy data for training generative AI models and vice versa. Our methodology involves the design of comparative experiments, the comparison of NURBS tensor representations with other 3D representations, and the use of reconstruction accuracy as a key metric to evaluate performance. ​Custom components for the Rhinoceros 3D parametric environment Grasshopper were developed enabling bidirectional conversion between NURBS geometry and NumPy tensors. These components are being released as a Grasshopper plugin under the name Wiener Dog as a free download. ​Our approach maintains geometric accuracy, reduces data size, and integrates seamlessly with existing deep learning libraries. ​The proposed methodology was tested on datasets of helicoid surfaces and lofted polysurfaces, demonstrating high reconstruction accuracy and generative potential. The ultimate aim is to build an AI tool that aids in exploring the great variety of geometric forms for architectural design.