Objective In the process of architectural renovation or expansion, structural design is a crucial aspect of building design. After the initial completion of the scheme design, architects often need to consider the compatibility between the architecture and its structure. Therefore, early intervention and immediate response in the structural scheme design are urgently needed. In this paper, addressing the preliminary design phase of architecture and focusing on situations where parts of the structure have already been determined, we propose a framework for the overall layout of the structural plan based on a Generative Adversarial Network (GAN), termed PF‒structGAN. This framework facilitates the design of the structural framework under the dual constraints of both architectural forms and predetermined structural elements. The core of this method involves constructing a model for the overall layout of the structural plan, which includes three main stages: constructing datasets, training and evaluating the model, and applying the model. Methods In the dataset construction stage, due to the limited number of data samples, and to reduce model training parameters, refine sample features, and improve training outcomes, this paper proposes three information representation methods for architecture, beams, and columns. It utilizes RGB color channels to store information separately: architectural space information is stored in the blue channel (B), beam information in the green channel (G), and column information in the red channel (R), thereby avoiding feature overlap in overlapping regions. The architectural information representation method is used to express architectural features strongly correlated with structural features. The beam information representation method is designed to express beam cross-sectional features in planar graphics. The column information representation method is designed to express column cross-sectional features in planar graphics. These three methods establish correlations between architectural and structural features. To integrate architectural and structural features, the feature maps are superimposed. The architectural feature map, partial beam feature map, and partial column feature map are superimposed to obtain the architectural and partial structural feature map. The beam and column feature maps are superimposed to obtain the structural feature map. The architectural and partial structural feature map, along with the structural feature map, constitute a pair of feature superposition maps. To address the problem that column features occupy too few pixels in the image, and to help the model learn these features more effectively, the paired feature superposition maps are cropped into four parts to increase the column feature ratio. Further augmentation of the original dataset is achieved by rotating it at 0°, 90°, 180°, and 270°, resulting in an expanded dataset. In the model training phase, the architectural and partial structural feature maps are used as constraint conditions, and the real structural feature maps are used as labels. The generator produces structural feature maps under the given constraints. The discriminator determines whether the generated image is real or synthetic. Through adversarial training, the generator and discriminator iteratively improve until reaching a Nash equilibrium. In the model evaluation phase, to assess the model’s design capability more reasonably, in addition to using the intersection over union (IoU) metric, this paper proposes the original column ratio index (γy), the irrationality index (γS), and the comprehensive index (γall) based on practical experience and frame structure design rules. These indicators comprehensively evaluate the model’s capability to produce an overall frame structure layout. γy evaluates the retention of frame columns generated by the model at their original input positions—the higher the ratio, the better the design compliance. γS evaluates the distribution of columns across different building components and spaces—the lower the index, the more reasonable the arrangement. γall integrates the above indicators—the higher the value, the more reasonable the structural layout. The best-performing model is determined based on these four indicators. Once the PF‒structGAN model is trained, the architectural and partial structure feature maps are input into the optimal model to generate a frame structure layout. Results and Discussions A total of 5 120 dataset pairs were created for training the generative model—4 320 for training and 800 for testing. The training set was input into the pix2pixHD framework, and training was stopped once adversarial training reached a Nash equilibrium. Model performance was evaluated using the four indicators. The IoU curve showed a general upward trend as training epochs increased. After the first epoch, γy remained at 1. γS generally trended downward. γall peaked at epoch 26; therefore, the model from the 26th epoch was selected as the best layout model. To verify the model’s structural design capability, an instance analysis was conducted using a teaching building project. The IoU between the model’s design and the engineer’s design was 0.56, indicating high similarity. γy was 1, showing full retention of original column positions. γS was 1.78, indicating a reasonable arrangement of columns. γall reached 0.88, suggesting that the generated structural layout was sound. The model’s generated frame columns met functional requirements, were well-placed, and had appropriate size and density. The layout of columns and beams closely resembled the engineer’s design. Conclusions This method enables intelligent and rapid generation of structural designs that comply with regulatory standards and follow conventional design practices. It offers reference solutions for architects during the structural design process.
法,用于表达与结构特征具有强关联性的建筑特征;2)框架梁信息表达方法,用于在平面图形中表达梁截面特征;3)框架柱信息表达方法,用于在平面图形中表达柱截面特征。进一步通过不同颜色通道(RGB通道)叠加特征图以充分保留不同信息。在此基础上,利用裁剪和增广等手段,构造用于训练生成式算法模型的5 120对数据集。此外,除沿用交并比(intersection over union,IoU)评价指标外,引入结构设计经验以更合理地评价模型的“设计”能力,基于框架结构设计规则提出了原柱率、不合理指数和综合指标。而后,在建筑和结构特征的双重约束条件下训练了CGAN模型,依据4个评级指标确定了最佳的框架结构平面整体布置模型。在使用PF‒structGAN模型时,将局部建筑和部分结构特征图输入最佳模型,输出局部结构图,最后拼接为框架结构平面整体布置图。
ChengGuozhong, ZhouXuhong, LiuJiepeng,et al.Intelligent design method of high-rise shear wall structures based on deep reinforcement learning[J].Journal of Building Structures,2022,43(9):84‒91.
ZhangYu, MuellerC.Shear wall layout optimization for conceptual design of tall buildings[J].Engineering Structures,2017,140:225‒240. doi:10.1016/j.engstruct.2017.02.059
[6]
TafraoutS, BourahlaN, BourahlaY,et al.Automatic structural design of RC wall-slab buildings using a genetic algorithm with application in BIM environment[J].Automation in Construction,2019,106:102901. doi:10.1016/j.autcon.2019.102901
[7]
ZhouXuhong, WangLufeng, LiuJiepeng,et al.Automated structural design of shear wall structures based on modified genetic algorithm and prior knowledge[J].Automation in Construction,2022,139:104318. doi:10.1016/j.autcon.2022.104318
[8]
LouHaopeng, GaoBoqing, JinFengling,et al.Shear wall layout optimization strategy for high-rise buildings based on conceptual design and data-driven tabu search[J].Computers & Structures,2021,250:106546. doi:10.1016/j.compstruc.2021.106546
[9]
LuXinzheng, LiaoWenjie, GuDonglian,et al.Research progress on building structural design methods:From simulation-based to artificial intelligence-based[J].Engineering Mechanics,2025,42(3):1‒17.
TurhanC G, BilgeH S.Recent trends in deep generative models:A review[C]//Proceedings of the 2018 3rd International Conference on Computer Science and Engineering (UBMK).Sarajevo:IEEE,2018:574‒579. doi:10.1109/ubmk.2018.8566353
[12]
AbdollahiB, NasraouiO.Explainable restricted Boltzmann machines for collaborative filtering[EB/OL].(2016‒06‒12)[2025‒04‒10].doi:10.1145/2872518.2889405
[13]
SalakhutdinovR, LarochelleH.Efficient learning of deep Boltzmann machines[C]//Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics.Cagliari:JMLR,2010:693‒700.
[14]
HintonG E, ZemelR.Autoencoders,minimum description length and Helmholtz free energy[C]//NIPS’93: Proceedings of the 7th International Conference on Neural Information Processing Systems.Denver:Morgan Kaufmann Publishers Inc.,1993:3‒10.
van den OordA, KalchbrennerN, VinyalsO,et al.Conditional image generation with PixelCNN decoders[C]//NIPS’16: Proceedings of the 30th International Conference on Neural Information Processing Systems,2016:4797‒4805.
[17]
vanD O A, KalchbrennerN, KavukcuogluK.Pixel recurrent neural networks[C]//International Conference on Machine Learning.New York:PMLR,2016:1747‒1756.
[18]
GoodfellowI, Pouget‒AbadieJ, MirzaM,et al.Generative adversarial networks[J].Communications of the ACM,2020,63(11):139‒144. doi:10.1145/3422622
NauataN, HosseiniS, ChangK H,et al.House‒GAN++:Generative adversarial layout refinement network towards intelligent computational agent for professional architects[C]//Proceedings of the 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR).Nashville:IEEE,2021:13627‒13636. doi:10.1109/cvpr46437.2021.01342
[21]
ChaillouS.ArchiGAN:Artificial intelligence x architecture[EB/OL]//(2020‒09‒03)[2025‒04‒10].
[22]
ZhangHang, BlasettiE.3D architectural form style transfer through machine learning[C]//Proceedings of the CA-ADRIA.Bangkok:CAADRIA,2020:659‒668. doi:10.52842/conf.caadria.2020.2.659
[23]
ChangK H, ChengC Y, LuoJieliang,et al.Building‒GAN:Graph-conditioned architectural volumetric design generation[C]//Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision(ICCV).Montreal:IEEE,2021:11936‒1945. doi:10.1109/iccv48922.2021.01174
[24]
ZhaoChaowang, YangJian, XiongWuyue,et al.Two generative design methods of hospital operating department layouts based on healthcare systematic layout planning and generative adversarial network[J].Journal of Shanghai Jiaotong University(Science),2021,26(1):103‒115. doi:10.1007/s12204-021-2265-9
[25]
LiaoWenjie, LuXinzheng, HuangYuli,et al.Automated structural design of shear wall residential buildings using generative adversarial networks[J].Automation in Construction,2021,132:103931. doi:10.1016/j.autcon.2021.103931
[26]
LiaoWenjie, HuangYuli, ZhengZhe,et al.Intelligent generative structural design method for shear wall building based on “fused-text-image-to-image” generative adversarial networks[J].Expert Systems with Applications,2022,210:118530. doi:10.1016/j.eswa.2022.118530
[27]
LuXinzheng, LiaoWenjie, ZhangYu,et al.Intelligent structural design of shear wall residence using physics-enhanced generative adversarial networks[J].Earthquake Engineering & Structural Dynamics,2022,51(7):1657‒1676. doi:10.1002/eqe.3632
[28]
ZhaoPengju, LiaoWenjie, HuangYuli,et al.Intelligent design of shear wall layout based on attention-enhanced generative adversarial network[J].Engineering Structures,2023,274:115170. doi:10.1016/j.engstruct.2022.115170
WangTingchun, LiuMingyu, ZhuJunyan,et al.High-res- olution image synthesis and semantic manipulation with conditional GANs[C]//Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition.Salt Lake City:IEEE,2018:8798‒8807. doi:10.1109/cvpr.2018.00917