AutoStruct: Intelligent design system for shear wall building structures

Architects and engineers frequently face substantial communication challenges in the building design process. By learning the design experience of structural drawings through neural networks, an automated structural design system can be developed to transfer the structural experience of engineers to architects, effectively reducing communication costs. Currently, research on AI-assisted automated structural design remains in a nascent stage. Existing public models exhibit notable limitations, particularly in their insufficient ability to learn drawing features comprehensively and their high usability thresholds. To overcome these challenges, we present AutoStruct, an intelligent AI-powered system for shear wall structure design. The core innovation of the system lies in its efficient Transformer-Wavelet architecture, which simultaneously captures both global features and local details from structural drawings while enhancing the learning of high-frequency information characteristics, such as wall elements. Specifically, to resolve the common issues of discontinuity and irregular distribution in generated layouts, we develop a computer vision-based post-processing method capable of repairing wall defects across various scales, thereby improving both continuity and surface regularity. Furthermore, our system incorporates a specialized sketch tool customized for architects. This web-based interface enables architects to quickly draft building schematics and input them into the model for structural layout generation, resulting in an easy-to-use and end-to-end automated design process. Finally, through comprehensive experiments on four datasets, we demonstrate that AutoStruct generates layouts that are more consistent with engineers’ designs compared to existing open-source solutions, and also shows its robust generalization capabilities.