剪力墙结构是高层建筑中常用的抗侧力体系, 其设计质量直接影响建筑的安全性与经济性。传统设计方法在高层建筑结构设计中存在局限性, 而基于图像的设计方法在空间感知方面展现出优势。该文提出了一种基于生成对抗网络(GAN)和剪力墙-梁联合训练的结构布置智能设计方法, 以提升剪力墙结构的方案设计质量。通过语义化处理200套建筑结构图纸, 构建了包含建筑设计图、剪力墙布置图、梁布置图及空间功能图的数据集。采用典型的GAN模型pix2pixHD, 通过联合训练同时生成剪力墙与梁的布置。采用基于相似性计算的S_SWall和S_Beam指标, 评估模型生成结果的设计质量。结果表明, 随着输入图像分辨率增加, 模型性能显著提升; 在剪力墙布置阶段引入建筑空间功能信息可进一步提高设计质量。相比剪力墙和梁布置独立训练方法, 联合训练方法在设计质量上更优。在1 536×768像素的图像分辨率下, 联合训练方法的S_SWall达0.850, S_Beam达0.770, 优于其他低图像分辨率下的结果。此外, 通过对一座典型高层住宅楼的案例研究, 验证了该方法的有效性, 生成的结构布置图与工程师设计模型在动力特性及层间位移角上基本一致, 差异在工程可接受范围内。

Objective: Shear wall systems are crucial components of high-rise buildings, providing the essential lateral force resistance. The design and layout of shear walls significantly influence the structural safety, serviceability, and cost-effectiveness of a building. Traditional design methods often rely on the experience and intuition of structural engineers, resulting in lengthy design cycles and potential suboptimal solutions, particularly for complex or large-scale projects. To address these challenges, this research proposes a novel intelligent design method based on generative adversarial networks (GANs), incorporating shear wall-beam joint training. This method aims to enhance the quality and coherence of structural layouts, automating the design process and optimizing the integration between shear walls and beams, which are traditionally designed in isolation. Methods: The proposed intelligent design method utilizes a stacked architecture of the pix2pixHD model, a state-of-the-art image-to-image translation network, to jointly train the generation of shear wall and beam layouts. A dataset consisting of 200 sets of architectural and structural drawings was compiled. These drawings were sourced from leading design institutes to ensure real-world applicability and compliance with relevant building standards. These CAD drawings were semantically processed to convert them into four distinct feature images: architectural design plans, shear wall layout diagrams, beam layout diagrams, and spatial function layouts (e.g., room locations and structural cores). The proposed method processes these feature images through a GAN-based framework, where the first generator (G₁) produces the shear wall layout from the architectural plan. The second generator (G₂) then utilizes the architectural and shear wall layouts to generate a beam layout. The GAN framework is optimized through a feedback loop, where the performance of G₂ is integrated into the loss function of G₁. This integration promotes the production of layouts that are architecturally accurate and structurally optimized for the beam system. The quality of the generated layouts can be evaluated using two key similarity indicators: S_SWall for shear wall layouts and S_Beam for beam layouts, which assess the accuracy of the generated designs relative to the ground truth. Results: The experimental results revealed several key findings highlighting the effectiveness of the proposed method. First, the performance of the model was significantly correlated with the resolution of the input images. Increasing the resolution from 256×128 pixels to 1536×768 pixels resulted in notable improvements in the quality of the generated designs. This is because a higher resolution enables the model to capture both the global layout strategies and intricate geometric details of the components. Second, the introduction of spatial function information, such as room types and functional zones (e.g., stairwells and elevator cores), as an additional input channel substantially enhanced the quality of the generated layouts. This input is most effective during the shear wall design phase, as it improves the overall coordination between the shear wall and beam layouts. The results show that the joint training model outperforms traditional independent training methods: the proposed method generates layouts with better structural continuity—that is, the relationship between the shear walls and beams is more logical and cohesive, aligning more closely with professional engineering designs. Under optimal conditions, with an image resolution of 1536×768 pixels and spatial function information included, the proposed model achieved an S_SWall score of 0.850 and an S_Beam score of 0.770, indicating a high level of design accuracy. Conclusions: This study demonstrated that integrating shear wall and beam layouts through joint training using GANs can significantly enhance the quality, coherence, and efficiency of structural layout designs. The proposed method successfully automates the design process and improves the precision and integration of structural components, which are crucial for optimizing overall building performance. The case study of a 30-story, 90-meter-high residential building revealed that the GAN-generated layouts closely matched the engineer's designs in terms of dynamic characteristics, seismic performance, and interstory displacement angles. Most of the discrepancies were within acceptable engineering tolerances. This result indicates that the method can be effectively integrated into the design workflow for high-rise buildings. This intelligent design approach enhances the efficiency of the early design stages. It also offers a practical, automated solution to drastically reduce the time and labor required for these stages. The results validate the viability of this approach for real-world applications, paving the way for the broader adoption of GAN-driven design methodologies in structural engineering.