[Abstract] With changes in the information environment and data base, artificial intelligence has made breakthroughs in big data, speech and image recognition, and deep learning. Traditional urban design techniques and methods are also facing the importance of upgrading and iteration Opportunity. This article starts with a systematic explanation of the combination of artificial intelligence and urban design technology methods at home and abroad. It takes the block scale that is widely available in the city and can connect the macro-city scale and the micro-architecture scale as the research object. Through evolutionary algorithms, adaptive algorithms and supervised Artificial intelligence methods such as deep learning have built an intelligent design module for the three-dimensional shape of the block, and proposed a relatively complete set of technical processes for the intelligent design of the three-dimensional shape of the city block. On this basis, by combining specific specific sites, the practical exploration of human-computer interaction from base construction, scheme generation to human-computer interaction is studied, and an urban design mode of artificial intelligence and designer interactive is proposed. This paper uses the block scale as a medium to try to solve the current generation of artificial intelligence urban design at different scales and the problems of unclear internal mechanism. It provides references for the iterative transformation of future digital urban design technology methods and the development and research of related design assistance systems. And reference.
Introduction
With the development of information technology, today's cities have entered an era of "Big Data, Intelligence, Mobile Internet, Cloud Computing", and various technologies are closely related. Together, it will not only greatly change the traditional lifestyle of urban residents, but also is expected to set off a new round of industrial transformation (Figure 1). At present, combined with digital related technologies, urban design methods have shown new characteristics that are different from previous development stages. Intelligence makes the city's database and multi-source big data mining closely integrated, and further supports and develops the method application and technological innovation of artificial intelligence in urban design. Through the close integration of the Internet of Things and the mobile Internet, more urban big data has been generated. In-depth analysis and mining of multi-source big data in the urban design process can help urban designers better analyze and judge urban development issues.Guide urban design.
Figure 1 Innovation cycle and technological surge wave
Through the combination of artificial intelligence technology and urban design technology, it is possible to simulate and calculate the possible economic, social and ecological benefits of different urban design schemes more efficiently and quickly benefit. The previous optimization analysis of urban benefits often required designers to build models and further optimize the plan based on the feedback structure. Through the system construction of artificial intelligence neural network structure, the benefit analysis and design optimization can be integrated, which greatly improves the efficiency of the design.
The application of artificial intelligence technology has been widely involved in many aspects such as space, ecology, transportation, and public management. In terms of urban space, the current research has achieved relatively mature results in urban imagery combined with computer vision technology. Through image learning, various types of buildings, urban streetscapes and even larger-scale urban morphological plans are generated. In the early 1980s, foreign related research tried to apply artificial intelligence technology to the field of architectural design. Professor Gerhard Schmitt of the Federal Institute of Technology Zurich (ETH) proposed to combine "computer creativity" with "user interaction". The scale level has triggered many theoretical and technical explorations such as the "digital chain" building generation, which provides ideas and foundations for the research and exploration of artificial intelligence in the field of urban design. In addition, breakthroughs in technological methods in computer fields such as intelligent algorithms for urban image and large-scale programmatic automatic modeling have also provided technical support for the combination of artificial intelligence and urban design.
China’s urbanization process is going through a new stage of “from quantity growth to quality improvement” [In 2017, the Ministry of Housing and Urban-Rural Development issued the "Urban Design Management Measures", proposing that urban design should run through the entire process of urban planning and construction management] In urban design, it is often necessary to consider the various three-dimensional form possibilities of each block when the total development volume of the area is fixed, so as to match the distribution setting and adjustment of the total development volume and indicators among multiple blocks. At the same time, it is necessary to consider the corresponding concrete three-dimensional building group volume under various development intensity distribution settings, in order to seek the control of the urban form within the entire planning range. For a long time, this work from indicator to form generation has been manually completed by a large number of design assistants of the design unit. When the indicator settings of each neighborhood need to be adjusted, the above process needs to be repeated.The workload is huge, which directly affects the enthusiasm for design deliberations and adjustments, and ultimately affects the design quality. In the context of the second opportunity for the combination of artificial intelligence and design, artificial intelligence has been studied in the field of block-scale urban design. Currently, it mainly includes the following four directions. (1) Strong arrangement design for the basic standard scheme of single-function blocks. At present, most of the existing research focuses on the generative design of residential areas, relying on the systematic rule parameters in the residential area design to generate the strong emission plan, and optimize the evaluation in combination with indicators such as sunshine and fire protection. In addition, this type of method has also been studied in the generation of high-rise building groups with clear rule constraints. (2) Design for the texture of historical blocks with a clear modulus. Relying on the characteristic modulus of traditional architectural settlements and street textures, digital generative design tools are established through data mining and machine learning technology to generate the morphological texture of historical blocks. (3) Morphological design of residential building community based on case library study. Through the search technology of the unstructured database, the case model that can be appropriately deformed is found to realize the automatic generation of high-rise residential neighborhoods that meet the professional requirements. (4) Machine-aided design based on feature recognition of neighboring blocks. Identify specific types of urban environments and main streets through the architectural and socio-economic characteristics of nearby blocks to guide urban design.
With the gradual application of artificial intelligence technology in the field of urban design, it has realized from determining the abstract development volume index to automatically generating the concrete three-dimensional building group volume, but there are still the following problems. (1) Existing methods such as image learning are insufficiently applied in the field of urban design. Although computer vision technology from the field of artificial intelligence has made great progress after more than 40 years of development, it is difficult to explain the uncontrollability of the existing image learning and other intelligent generation methods, and it is difficult to realize the block of non-residential land with a simple image learning method. The learning of the internal mechanism of form and building scale has led to artificial intelligence generation schemes only suitable for residential land with relatively simple spatial planning, and it is difficult to meet the specific requirements for space combination under the mixed situation of different land properties. (2) The design thinking logic at the block level is difficult to reflect the multi-scale urban design decision-making requirements. The “bottom-up” design thinking starting from architecture is suitable for small-scale, single-function areas, while urban design pays more attention to the overall control of the urban spatial form and style, and emphasizes the functions, spaces, landscapes, and landscapes of different street sections. Transportation, walking, and other connections also need to adopt a "top-down" design approach, but there is currently a problem that it is difficult to transmit information on design solutions corresponding to different scales.(3) The lack of block-scale urban design methods for different functional plots. The current block-scale artificial intelligence city design is mainly oriented to residential areas and historical blocks with a single function, and the design ideas for the arrangement and combination of buildings in different functional blocks are not the same as the generation methods of residential areas and historical blocks. The design specifications are also different. There are many differences, so it is necessary to find a more inclusive intelligent design idea. (4) The overall "black box" generation model lacks a step-by-step interactive optimization mechanism. Methods such as image learning to generate settlements forcibly arranged plan are more to adopt the direct generation plan against the network, and adopt basic rules and regulations for verification, but the functions of different blocks are not like the settlements that tend to be self-consistent and closed on the road and pedestrian system. The traditional design process integrates roads, walking paths and even buildings in different blocks to carry out multi-round plan design. The current “black box” method of overall generation blocks the opportunity for designers to interact and optimize step by step.
In summary, this research has combed the research context of the application of artificial intelligence in urban planning and design technology at home and abroad, and focused on the combination of block-scale urban design and artificial intelligence-related technical methods. Further expansion on the basis: (1) In terms of research objects, expand the single function of residence to the blocks composed of different functional plots; (2) In terms of the generation logic, extend the "black box" one-key generation method idea to Top-down thinking such as "road-walk-building"; (3) In the generation algorithm, traditional single methods such as image learning and Bayesian parameterization are extended to be combined with different elements such as roads and buildings in the block. Generate algorithm library; (4) In the design interaction, the final result-oriented selective interaction is extended to the procedural interaction that is generated step by step.
1 The overall architecture of the intelligent design of the three-dimensional shape of the block
This article constructs the intelligent design of the three-dimensional shape of the block according to the top-down urban design logic, which mainly includes the following four modules: three-dimensional space of the block-base calculation module , Program specification translation-intelligent control module, overall intelligent design-step-by-step generation module, man-machine coordinated design-step-by-step optimization module (Figure 2).
Figure 2 The overall architecture diagram of the block 3D shape intelligent design module
(1) Block 3D space—base calculation module.Obtain the geographic information data of the target block and its surrounding blocks to construct a three-dimensional space sand table. The surveying and mapping drone collects the geospatial information of the target block and a block with the target block as the center. The image information in the raster format is converted into vector data through the built-in data collection module and entered into the geographic information platform. In the geographic information platform, the geographic spatial vector data of the block is converted to a unified coordinate system for the alignment of spatial geographic coordinates and projection coordinates, and a high-precision three-dimensional spatial sand table is made. The geospatial information includes block roads, block buildings, block public spaces, and block topography.
(2) Program specification translation-intelligent control module. The translation extracts the design conditions in various upper-level planning documents and standards at all levels and performs spatial information registration, and embeds the pre-design conditions in the target block in the form of an attribute table. Collect the special planning, regulatory texts and related norms and standards related to the target block [Relevant statutory codes include the "Store Building Design Code", "Building Design Fire Protection Code" and "Urban Road Intersection Planning Code"], and extract the blocks by loading database components Design conditions, the extraction content of the block design conditions preset in the component. Standardize the collected block design conditions, unify the data format, and perform spatial matching with the three-dimensional sand table, and link to the target block in the form of an attribute table.
(3) Overall intelligent design—generate modules step by step. Based on the extracted design conditions, a three-dimensional block of the block is generated, and a three-dimensional building model is generated based on the machine learning of the building combination, and the building form is optimized. A step-by-step generation plan and a step-by-step optimization route are adopted at the block scale, including the two-dimensional plan of the block, the three-dimensional space volume of the block, the public space of the block, the pedestrian entrance of the block, the tower building and the podium building of the block, and the block building block scheme. At the same time, build a block 3D contour training sample library, generate a block 3D model by loading a machine learning model, and optimize the podium building form to generate multiple block building block schemes.
(4) Human-machine coordinated design—optimize the module level by level. Use the holographic sand table to display the generated block design plan, and build an interactive feedback instruction library, select and identify instructions, instruction processing and result feedback display. By constructing the holographic sand table and virtual reality real-time interactive system of the urban design plan, the designer can realize the immediate, accurate and efficient feedback and adjustment of the urban design plan, and solve the single angle of view, poor interaction, adjustment delay, Long follow-up work cycle and other issues.
2 Four methods for smart block design
Based on the overall architecture of block three-dimensional shape smart design, through evolutionary algorithms, adaptive algorithms, supervised deep learning, step-by-step interactive design and other artificial intelligence methods , To build an intelligent design module for the three-dimensional shape of the block.
2.1 Evolutionary algorithm: inherent law and mechanism of mapping design
Evolutionary algorithm includes genetic algorithm, genetic programming, evolutionary programming and evolutionary strategy, etc. It refers to starting from any initial population, through limited random selection, mutation and recombination Through the process, the group evolves to better and better regions in the search space (Figure 3). Genetic algorithm has been applied in the layout of building communities, and the fitness function in evolutionary algorithm can be put into a large number of urban planning and design specifications, technical standards and empirical value ranges, further restricting the direction of random selection, and can target the linear network Different prototype modes are subject to evolutionary constraints, which are more suitable for the generation of road and walking networks.
Note: The research object takes (0,0) as the origin and grows in an abstract rectangular ideal block. The numerical unit in the coordinates is meters, and a negative value means that the growth direction is opposite.
Figure 3 N-round growth graph of the street walking system based on evolutionary algorithm
2.2 Adaptive algorithm: learn existing experience and knowledge
Adaptive algorithm utilizes existing urban buildings, road networks and other vectors The data builds a block case study library, and combines the design target plot indicators to build a case matching decision tree. Based on field surveys in the central area in recent years, combined with big data related technologies, this research group has collected spatial data results of major cities at home and abroad, including buildings, roads, water systems, green spaces, etc., and targeted the major cities in this technology. Through multiple rounds of cleaning the spatial data of, the spatial data of blocks with fewer buildings and poor quality, such as construction sites and brownfields, were eliminated, and six types of block data with different functions with high representativeness were screened out. Select similar cases corresponding to indicators and specifications in the target plot,And based on different influencing factors such as the nature of block land, form index, block area, etc., construct a parametric adaptive design operation pedigree (rotate, zoom, pull, delete, etc.), and integrate existing urban blocks with different styles and different environments. The plan is adapted to the target site to form a plan database that meets the unique needs of different regions (Figure 4).
Figure 4 Application ideas of adaptive algorithms in building communities
2.3 Supervised deep learning: scheme control and style selection
The supervised learning method is currently a widely studied machine learning method , Such as neural network propagation algorithm, decision tree learning algorithm, etc. have been successfully applied in many fields. Under the fractal perspective of traditional urban design, there have been studies on the quantitative indicators of block morphology subdivided by different functions. Combining the block morphological characteristics indicators of different functions and connecting various control indicators of the existing regulatory detailed planning of the block, the block decision tree can be constructed. . Through the analysis of training data by supervised learning algorithms, the intelligent scheme of urban design allows various labels to be determined, and samples with known control requirements of certain schemes or certain style characteristics are used as training sets to establish a scheme library for certain styles. Use the established scheme library to generate a new urban design scheme (Figure 5).
Figure 5 Optimization and adjustment of the spatial distribution of intelligent generation schemes under supervised learning The machine is highly interactive. Applied in design, it is mainly manifested as virtual interaction, which realizes interaction technology through gestures, clicks and other recognition, and displays virtual devices and products to designers. For example: in the CityScope and other projects developed by the Massachusetts Institute of Technology (MIT) Media Lab (MediaLab), the tangible interaction matrix (TIM) is formed through the collaboration of optically marked Lego object arrays, computer vision, and 3D projection mapping. City designers can Copy and move objects to change the 3D project model on the digital screen,The School of Architecture of Southeast University uses technical tools such as VR interactive sand table and City Cube in scientific research and practice (Figure 6).
Figure 6 Interactive design application based on immersive virtual reality scenarios
3 "Match-Generate-Feedback" step-by-step intelligent design technology process
3.1 Case library intelligent decision-making: more Factorial decision tree matching step by step
Collect vector 2D data of buildings and road information in a certain range around the design plot and 3D spatial information of surrounding buildings oblique photography; obtain the control detailed planning of the design plot in the upper plan The CAD format data of the road of the plot, and the plot control index data inside the design plot, unify the vector data coordinates and establish the connection relationship between the control plan index and the architectural-road space model of the design plot; obtain the current urban design status Build a project case library (Figure 7 left), extract the feature index system of urban design project cases, prioritize feature parameters, and build a feature index system decision tree (Figure 7 right), select similar cases based on the decision tree, and form feature learning Case Library.
Figure 7 Spatial base construction and multi-factor decision tree step-by-step matching diagram
3.2 Integral intelligent design: "Road-walk-building" is generated step by step
The intelligent design method module integrates the more mature generation methods in existing research such as evolutionary algorithms, and focuses on further method exploration for the design of different functional neighborhoods, trying to build a "road-walk-building" step-by-step generation The overall intelligent design process.
(1) Intelligent design of block road network. According to the location of the entrance and exit of each block determined in the control detailed planning, the grid method and the shortest path method are combined to intelligently generate multiple plot road network plans, and combine the road rule inspection model to verify the plan, and output the road network that meets the rules Schemes and characteristic parameters of each scheme.
(2) Intelligent design of block public space. According to the land use properties of each block determined in the regulatory detailed planning and the intelligently generated road network plan, the basic unit parameter control of walking and the evolutionary algorithm are combined to intelligently generate multiple plots of walking system plans, and combined with walking system rules to check the model The plan is verified, and the walking system plan that meets the rules and the characteristic parameters of each plan are output.
(3) Intelligent design of the building community in the block. Construct a sample database of architectural composition, and intelligently match it with the case database of architectural composition according to the characteristic indicators of the architectural composition sample database of the design plot, and further generate building composition schemes with different functions through evolutionary algorithms and adaptive algorithms, and combine them with land parcel control regulations The space parameters and the sunshine interval verify the plan, and output the building combination plan that meets the rules and the characteristic parameters of each plan (Figure 8).
Figure 8 "Road-Walk-Building" step-by-step generation map
3.3 Interactive intelligent design: man-machine collaborative optimization step-by-step feedback
uses virtual reality instant interaction platform and equipment, through holography The sand table displays the three-dimensional model. The designer wears VR glasses and uses data gloves to better integrate into the three-dimensional virtual scene and perceive the expected effect of the project design in advance (Figure 9). In real-time spatial perception, designers can make immediate adjustments to the design plan according to different needs: voice recognition equipment can capture and recognize voice command information, and data gloves can capture and recognize gesture action command information through sensors. According to the designer's instructions, the intelligent three-dimensional sand table immediately adjusts the design plan, and calculates the adjusted design indicators, so that the plan design adjustment is more efficient, and the plan is closer to the real effect (Table 1).
Figure 9 The holographic sand table shows the interaction mechanism of the 3D model
Table 1 Human-machine collaborative optimization real-time adjustment instruction set (part)
4 The practice of intelligent design of the three-dimensional shape of the street profile under human-computer interaction Application
4.1 Foundation construction: the designer’s interpretation of the rules of the site’s current situation
In practical cases,Five target plots were selected as the basic blocks generated by the scheme (Figure 10). Unify the vector data coordinates, load the surrounding buildings of the design plot, the two-dimensional data of road vector, the design plot control and enclosing road, and the three-dimensional tilt photography information of the surrounding buildings of the plot into the smart city design platform, in the DGX1 V100 ultra deep learning workstation Run on. And collect the plot control index data inside the designed plot, including the land attributes of each block, the location of the entrance and exit of the block, the development intensity of the block (building density, building height, floor area ratio), building line rate, building line regression, etc.
Note: Floor area ratio, building density, and building height are all upper control limits.
Figure 10 Selection of five target plots
4.2 Scheme generation: Preliminary exploration of stepwise intelligent generation design scheme
Intelligent design of block road network, combined with "Urban Road Traffic Planning Design Specification 》(GB 50220-95) on the block size and road density, and limit the rules of road generation. A hidden grid is formed between two determined entrances and exits, and the shortest path between the two intersections is automatically searched and connected on the grid, the design plan road centerline is drawn, and the road is automatically widened according to the urban branch road regulations.
On the basis of multiple road schemes, further intelligently generate block public space. The evolutionary algorithm L-System is used to evolve the basic unit of growth from the starting point of the walking system. The basic unit grows according to the preset walking system fitness function. During the growth process, the function restricts the connectivity of the walking system in different street sections, and avoids excessive parallel overlap between the walking system and the road network generated in the previous round. Good walking space for pedestrians and vehicles.
Based on the multiple schemes of the pedestrian system, further intelligently generate the block building community. Intelligent matching of the architectural combination case database: Take the target plot 4 as an example, compare the interaction indicators of each neighborhood with the case database, arrange the sample architectural combinations with a matching degree of 90% according to the degree of matching, and select the top 1000 of the matching degree A case study database is generated from a combination of buildings; the first 1000 building combinations are collected and used as a database for the CVAE-GAN complementary map algorithm and building adaptive algorithm for machine learning.Generate massive plans of building combinations with different functions for each neighborhood (Figure 11), and control the space parameters of the plot (development intensity, density, building height, building line rate, building line retreat) and residential building sunlight spacing, etc. The technical specifications are checked and the schemes that do not meet the specifications are eliminated. The planning index database is connected with the urban design plan database, and the planning index values of each plan generated by the intelligent system are calculated through the planning index calculation formula. Compare the values of various planning indicators in the urban design with the limited range of planning indicator values in the planning indicator database to determine whether the multi-plan planning indicators are within the limited range of planning indicator values. If the result is no, it is determined to be an abandoned plan; if the result is yes, the plan is loaded into the plan library that meets the planning index (Figure 12).
Figure 11 The intelligent generation process of the neighborhood building combination of the target plot 4
4.3 Human-machine step-by-step interaction: Interactive optimization design of immersive virtual reality
In the step-by-step interaction phase of the plan, the design results that meet the planning indicators are imported into the holographic sand table, and the site chassis 3D model is superimposed with the building 3D space model Generate an intelligent interactive platform for urban design. Position the designer’s eye perspective and the relative spatial position of the three-dimensional holographic sand table, and set sensors with detectable distances around the holographic sand table platform, on the VR glasses worn by the designer, and the palm of the data glove to make the holographic sand table, VR glasses and Data gloves can detect each other's position. Build a holographic sand table intelligent command library, including command names, gesture action command information or voice command information, as well as smart three-dimensional sand table modification and adjustment.
In the process of human-computer interaction, input the intelligent modification and adjustment of the holographic sand table corresponding to the instruction into the workstation. The modification and adjustment corresponding to the instruction include the stretching, scaling, rotation, and adjustment of the building layout, etc., which can be calculated immediately after adjustment The planning indicators of the new plan are developed, and the revised and adjusted interactive artificial intelligence city design results are optimized and displayed step by step (Figure 13).
Figure 13 Schematic diagram of the interactive artificial intelligence city design platform architecture
5 Conclusion
Now,With the increasing application of digital technology in the field of urban planning, the combination of urban design and artificial intelligence is also facing a change. Although artificial intelligence methods have been used in architectural design and other fields for some time, they have also provided inspiration and reference for the innovation of urban design-related methods and technologies. However, the field of urban design includes the internal mechanism of multi-scale space and has more diverse humanities and cultures. The "obvious and hidden mutual learning" of social content, so in the process of artificial intelligence city design, not only need to meet the standard verification of building scale, sunshine, etc., but also consider multi-scale, multi-element gradual design and round-by-round optimization. The artificial intelligence city design proposed in this paper is a block-scale interactive design model. As a block-scale artificial intelligence city design exploration, it is hoped that the intelligent generation of single buildings can be extended to blocks, and the intelligent generation of urban forms can be subdivided into block construction. A link provides a new way of thinking and new exploration in the transformation of the combination of artificial intelligence and urban design that blooms in abundance.
Author: Yang Jun feast (corresponding author), Southeast University School of Architecture, Professor; Urban Institute wisdom, Southeast University, vice president. [email protected]
Zhu Xiao , PhD student, School of Architecture, Southeast University. [email protected]
Research on the Types and Characteristics of Urban Intelligent Planning Technologies at Home and Abroad Original
for this subscription account.
