Building-GNN: Exploring a co-design framework for generating controllable 3D building prototypes by graph and recurrent neural networks

Abstract

This paper discusses a novel deep learning (DL)framework named Building-GNN, which combines the Graph Neural Network (GNN) and the Recurrent neural network (RNN) to address the challenge of generating a controllable 3D voxel building model. The aim is to enable architects and AI to jointly explore the shape and internal spatial planning of 3D building models, forming a co-design paradigm. While the 3D results of previous DL methods, such as 3DGAN, are challenging to control in detail and meet the constraints and preferences of architects' inputs, Building-GNN allows for reasoning about the complex constraint relationships between each voxel. In Building-GNN, the GNN simulates and learns the graph structure relationship between 3D voxels, and the RNN captures the complex interplaying constraint relationships between voxels. The training set consists of 4000 rule-based generated 3D voxel models labeled with different degrees of masking. The quality of the 3D results is evaluated using metrics such as IoU, Fid, and constraint satisfaction. The results demonstrate that adding RNN enhances the accuracy of 3D model shape and voxel relationship prediction. Building-GNN can perform multi-step rational reasoning to complete the 3D model layout planning in different scenarios based on the architect's precise control and incomplete input.