首页|嵌入式机器学习方法在街区形态生成设计中的应用探索

嵌入式机器学习方法在街区形态生成设计中的应用探索

扫码查看
城市街区的形态与空间品质关系密切.在城市设计中借助智能工具实现街区形态的生成是辅助设计决策的一项典型应用.既有的生成式方法主要采用参数化建模、形流推理以及图像生成算法实现,缺少对街区尺度形态规律学习的过程因而难以处理复杂、精细的空间形态关系,生成令人满意的形态结果.因此,本研究在既有生成式方法基础上借鉴嵌入式系统理论,提出一种通过嵌入机器学习训练模块优化街区形态生成结果的综合性方法(简称"嵌入式机器学习方法").该方法包含高质量的训练数据、平衡的模型复杂度与性能、特征工程与模型优化能力、独立运行的学习模块以及支持开放接口与部署五个要素,实现在传统生成规则中增加弹性并且可控的人工神经网络,可根据设计需求对训练样本及模型进行自由组合,改善了既有算法的生成效果.该方法也增加了设计流程的灵活性,设计师可根据场地需要配置合适的训练样本,并通过预训练的方法将结果嵌入生成式系统中,从而实现快速、便捷的三维形态生成,辅助街区形态设计过程.文章阐述了一个探索性的实践案例,从模型目标、数据、训练过程等方面进行阐述,并结合生成结果对比探讨了该方法对传统生成方法的改进,结果表明采用嵌入式机器学习方法可以有效改善三维形态的特征细节,生成更契合场地特质的设计方案,为设计师带来有益启示.最后文章提出了嵌入式机器学习方法的未来开发潜力,为城市街区设计中的人工智能技术应用提出了新的思路和可能性.
Practical study on the application of embedded machine learning method in generative design at block level
China's urban development is transformed from rapid growth to optimization of built area,paying more attention to the improvement of spatial quality.Accordingly,urban design work also needs to be more refined and high-quality.The form of urban blocks is closely related to the spatial quality.Block morphological design is a process of creating urban space at block scale.Compared with macro-scale studies,block-scale morphological studies have the characteristics of scale,diverse and discrete,which not only highlight the subtle relationship of urban spatial forms,but also reflect the characteristics and styles of design,which brings challenges to generative models based on preset rules.In the traditional design method,the design of urban form is realized by the designer according to the aesthetics rules and experiences.On the one hand,it is limited by subjective cognition,and on the other hand,it is difficult to define the subtle relationship of spatial form.The development of new data environment and artificial intelligence technology has brought a new vision for the form generation of block scale.In urban design,the generation of block form with the help of intelligent tools is a typical application to assist design decisions,which provides a basis for creating reasonable space forms and enables designers to propose more effective design strategies.The existing intelligent generation methods mainly have three approaches:1)parametric modeling,generating results according to morphological rules;2)form-flow reasoning,through the establishment of form-flow mechanism,according to the flow elements to deduce specific forms;3)image generation algorithm,which directly generates design schemes based on image samples through deep learning.However,due to the lack of learning process of block-scale morphological law,it is difficult to deal with complex and precise spatial morphological relations and generate satisfactory morphological results.On the basis of the existing generative methods,this study draws on the embedded system theory,and proposes a comprehensive method to optimize the block form generation results by embedding machine learning training modules(referred to as"embedded machine learning method").The proposed method is based on the demand analysis for block form design,and is explained from three aspects:morphological knowledge extraction,gain innovation and elastic generation.It is deduced that multiple optimized machine learning modules need to be integrated in the embedded form on the existing generation system,so that the system can meet the requirements of the generative algorithm for block form design.Increase the flexibility and adaptability of the model and promote the man-machine collaborative design process.This method includes five elements:high-quality training data,balanced model with complexity and performance,feature engineering and model optimization capabilities,independently running learning modules,and support for open interface and deployment,to achieve an artificial neural network with increased flexibility and controllability in traditional generation rules.Training samples and models can be freely combined according to design requirements.The generation effect of the existing algorithm is improved.This method also increases the flexibility of the design process.Designers can configure appropriate training datasets according to the needs of the site,and embed the results into the generative system through pre-training method,so as to realize fast and convenient three-dimensional form generation and assist the block form design process.In this paper,an exploratory practice case is presented,which integrates the learning module of the architectural texture characteristics of the riverfront area in the city into the block form generation model.The research process of this case is described from the aspects of model objectives,data and training process.The morphological data of waterfront urban areas of the same type are taken as samples,and the machine learning and image transfer learning models are embedded into the generative algorithm of block morphology.By comparing the generated results,the improvement of this method over the traditional generative method is discussed.The results show that the embedded machine learning method can effectively improve the feature details of the three-dimensional form,generate a design scheme that is more suitable for the characteristics of the site,and has excellent optimization performance in the view of the riverside landscape,which brings beneficial enlightenment for designers.The paper presents the future development potential of embedded machine learning,including multi-modal data fusion machine learning,flexible and diverse combination of algorithm modules,and higher quality man-machine collaborative design processes.The author believes that combining the advantages of artificial intelligence with the classical urban model to create a new flexible integrated model to cope with the complex and diverse design environment can provide new ideas and possibilities for the problem of form creation in block or urban design,and hopes to promote more discussions and practices on embedded machine learning methods and enrich the theory and technology in urban design practice.

block morphological designembedded methodsgenerative modelsmachine learningintelligent design

甘惟、王元楷、李翔

展开 >

同济大学建筑与城市规划学院

同济大学自然资源部国土空间智能规划技术重点实验室

同济大学建筑设计研究院(集团)有限公司

浙江工业大学设计与建筑学院

展开 >

街区形态设计 嵌入式方法 生成式模型 机器学习 智能设计

国家重点研发计划(十四五)

2022YFC3800205

2024

西部人居环境学刊
重庆大学

西部人居环境学刊

CSTPCD北大核心
影响因子:0.698
ISSN:2095-6304
年,卷(期):2024.39(3)
  • 26