Building massing generation using GAN trained on Dutch 3D city models

Abstract

Despite being relatively novel, generative adversarial networks (GAN) have already been appropriated for application to several problems within the field of architectural and urban generative design. However, the preceding GAN based models for building massing generation make use of only simplified and two dimensional representation of the built environment. This work improves upon the existing deep-learning-based methods for generation of building massings and building group layouts, by fusing high accuracy three-dimensional building models with site context derived from cadastral and topographic data, sourced from openly available datasets in the Netherlands. Pix2pixGAN implementation in PyTorch, trained on existing massing data encoded into images as heightmaps, is used to generate building massing geometry. Two methods for geometry extraction from heightmaps are introduced, voxelization and vectorization. The goal for the model is to maximize similarity of morphological traits of configurations generated by the model to the ground truth training data. The effects of multiple proposed training configurations on the resulting massings generated by the model are evaluated, together with visual assessment, using their Spacematrix mappings. Three distinct models with specific goals are presented - parcel infill model, street block infill model, and urban fabric infill model. All three models show a capability to learn spatial traits of existing building configurations and transfer them into new situations not encountered in the training data, which is confirmed by the distribution of Spacematrix mapping of the generated results being similar to the distributions of the ground truth data. The proposed methodology represents a novel approach to generating building massing configurations by autonomously inferring the rules of their composition from existing urban areas. The resulting models could be used to provide initial states in optimization-driven design approaches, or as smart massing suggestion engines, assisting architects and city planners during the early building design process.