Automated building layout generation using deep learning and graph algorithms

• Published in: Automation in construction Vol. 154; p. 105036
• Main Authors: Wang, Lufeng, Liu, Jiepeng, Zeng, Yan, Cheng, Guozhong, Hu, Huifeng, Hu, Jiahao, Huang, Xuesi
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2023
• Subjects: AI-generated content; Building layout; Deep learning; Graph algorithms; Layout plan; Deep learning; Building layout; Layout plan; AI-generated content; Graph algorithms
• ISSN: 0926-5805; 1872-7891
• DOI: 10.1016/j.autcon.2023.105036

Designing architectural layouts is a complex task that has garnered significant attention in the research community. While automated site layout design and flat layout design have been extensively studied, automated building layout design has been relatively overlooked. This paper describes an approach for generating automated building layouts using deep learning and graph algorithms. A unique building layout dataset is created to support the proposed approach. Euclidean distance, Dice coefficient, and a force-directed graph algorithm are employed for layout selection and fine-tuning. The Input-controlled Spatial Attention U-Net model accurately segments the building region, and the resulting layout is refined through image operations, leading to comprehensive BIM models for designers. Through two generative case studies and a comparative experiment with neural networks, this paper demonstrates the effectiveness of the approach that can assist designers during the initial stages of design and enable a rapid generation of complete layouts for individual buildings. •A framework based on deep learning and graph algorithms for automated building layout generation is proposed.•A unique annotated building layout dataset, GeLayout, is created, which is the first of its kind.•Euclidean distance, Dice coefficient and a force-directed graph algorithm are employed for matching and fine-tuning the graph.•The proposed Input-controlled spatial attention U-Net (ICSA-UNet) generates more concise and well-defined segmentation lines compared to those generated by the standard U-Net.•The generated BIM models in both typical and untypical scenarios are effectively meeting the requirements for design assistance and exhibit stability.