Master's Thesis Award Winners 2023

Learning and Generating Spatial Concepts of Modernist Architecture via Graph Machine Learning

Erik Bauscher
Supervisor: Prof. Thomas Wortmann & Prof. Stephan Trüby
Advisors: Anni Dai & Diellza Elshani

With the world of architecture starting to explore the potential of generative AI, this project showcases a novel approach: generating three-dimensional spatial configurations of architectural elements by sampling the latent space of a graph variational autoencoder. Conventional architectural design works with two-dimensional plans and sections as well as physical three-dimensional models, which all contain compressed and partial information of a complete building design. While this conventional approach works well in an environment where the human architect can mentally unroll this compressed information, in the world of computers this step of compressing is redundant (Carpo, 2017) and rather unhelpful for neural networks to understand complex content. Using graphs as inputs to a graph neural network reduces the information loss from the original building compared to most other research done in the field of generative architecture, which uses image-based generation. This reduced information loss is due to the fact that graphs do not only represent information about elements (nodes, pixels, etc.) but also the relationships between nodes (Hamilton, 2020). This representation of relationships seems especially helpful in architecture, where space is defined by an assemblage of physical elements which are connected (e.g., walls that are touching floors). We construct our graph-based generative model in four steps: (1) Creating a dataset: Since most architectural datasets are made of two-dimensional images, we create our own 3D dataset consisting of four 3D models of well-known modernist buildings. We augment this data set by generating 500 parametric variants of each 3D model. Although this method can be applied to any architectural style, we chose the modernist style due to its simplicity of elements and their connections (orthogonality, rectangularity, etc.). (2) Converting 3D buildings into graphs: All surfaces (floors, walls, windows, etc.) are nodes in a graph, which are connected via an edge when touching. The details of this conversion process play a crucial role in the generated geometry and can be considered one of the main design choices of the method. (3) Training and tuning a graph variational autoencoder: While the encoder architecture is a common GNN, the decoder needs to reconstruct the geometry from a low-dimensional embedding. Accordingly, its design is decisive for the resulting geometry. (4) Interpreting the output and the method: While the designer of the tool has a large degree of control over the outcome of the generation process, the method generates valuable, logical and original geometries that represent the architectural style chosen in the training data. These geometries are far more coherent than those from an image-based generation process and demonstrate the importance of graph-based 3D geometry generation of architecture via GNNs. The method introduces a novel conversion process from architecture to graph and an adapted decoder architecture, which make generative machine learning more applicable and relevant to real-world scenarios in building design.