End-to-end generation of structural topology for complex architectural layouts with graph neural networks

Current automated structural topology design methods can only deal with limited design spaces or simplified architectural layouts for lack of data or a proper representation of structure topology. To address this, the abundant information of manually designed architectural and structural layouts should be exploited to guide the topology design. To achieve automatic generation of structural topologies according to real-world architectural layouts, this research introduces StrucTopo-generative adversarial network (GAN), an end-to-end generative model with node and edge generation stages based on proper graph representation. Nodes are generated using an image-to-image translation model, and edges are generated with a GAN-based approach. The model is trained and tested on a dataset of 300 complex architectural and structural layouts. Measured against the manually designed topologies, the results indicate that the proposed model can generate reasonable structural topologies, with a recall of 97% and an intersection-over-union of 80% in node generation, with a precision of 92% and a recall of 91% in edge generation. Additionally, the joint generation shows a graph similarity of 72%. The proposed model is the first of its kind to consider complex architectural layout constraints in the generation of structural topology, marking a step forward in applying artificial intelligence to practical structural design.